Fix #9447 bad memory usage

A fix for #9447

Before this change, for the workload in the ticket, we saw a memory stay at around 4 GB an hour after the workload ended.  after this change we see both a lower memory high watermark, and also see memory come back to baseline at the end as we expect (after the `HASURA_GRAPHQL_PG_CONN_LIFETIME` and a subsequent `_idleStaleReaperThread` interval elapses).  But note the numbers given on any given machine may be slightly different, since the behavior depends on the version of glibc and processor speed (Since the test case depends on the server being overloaded).

This might also help some users commenting in https://github.com/hasura/graphql-engine/issues/9592

PR-URL: https://github.com/hasura/graphql-engine-mono/pull/9823
GitOrigin-RevId: 5650aa42d10d46c418c21686983a982d69011884

cabal.project

@@ -67,6 +67,13 @@ package *
   -- For tooling, e.g. 'weeder', and IDE-like stuff:
   ghc-options: -fwrite-ide-info
 
+  -- we statically link malloc from mimalloc. Out of an abundance of caution, 
+  -- disable special treatment of these in all the foreign code we build. The
+  -- only risk is potential for some missed optimizations.
+  -- See: https://github.com/microsoft/mimalloc/issues/785
+  -- NOTE: cc-options is not recognized here for some reason, so we use ghc-options+optc:
+  ghc-options: -optc-fno-builtin-malloc -optc-fno-builtin-calloc -optc-fno-builtin-realloc -optc-fno-builtin-free
+
   haddock-html: true
   haddock-hoogle: true
   haddock-hyperlink-source: true

preload-mimalloc/README.md

@@ -0,0 +1,40 @@
+# Vendored mimalloc
+
+
+This is [mimalloc](https://github.com/microsoft/mimalloc) vendored for use in
+our server code. You can upgrade mimalloc by editing `VERSION` in
+`update-sources.sh` and re-running it.
+
+Usage sites in haskell projects look like:
+
+    -- See 'preload-mimalloc/README.md'
+    -- We want this in all server binary executable sections
+    common preload-mimalloc
+        c-sources:       ../preload-mimalloc/mimalloc/src/static.c
+        include-dirs:    ../preload-mimalloc/mimalloc/include
+        cc-options:      -DMI_MALLOC_OVERRIDE
+
+## Usage in haskell projects
+
+In [#9447](https://github.com/hasura/graphql-engine-mono/pull/9447) we see
+overloading the server with large and slow queries results in high memory usage
+and what looks like a space leak, involving foreign data from libpq.
+
+See that ticket for details, but mimalloc helps (relative to my glibc) by:
+
+- better avoiding fragmentation, resulting in lower overall memory usage (4GB peak 2.8 aftervs.
+- eagerly decomitting memory so that OS-reported memory usage is a useful
+  metric (we do the same in the haskell RTS by specifying
+  `--disable-delayed-os-memory-return`)
+
+We also benefit by standardizing this, rather than relying on users' glibc
+version, which might vary in behavior, making debugging difficult.
+
+Other implementations such as tcmalloc or jemalloc were not tested but might
+perform better or be more tunable.
+
+### Why not LD_PRELOAD?
+
+It's most common and simplest to use the dynamic linker to override malloc.
+This would work in docker containers we ship, but we want all users to have the
+same (better) behavior out of the box.

preload-mimalloc/mimalloc/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2018-2021 Microsoft Corporation, Daan Leijen
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

preload-mimalloc/mimalloc/include/mimalloc-new-delete.h

@@ -0,0 +1,66 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2020 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_NEW_DELETE_H
+#define MIMALLOC_NEW_DELETE_H
+
+// ----------------------------------------------------------------------------
+// This header provides convenient overrides for the new and
+// delete operations in C++.
+//
+// This header should be included in only one source file!
+//
+// On Windows, or when linking dynamically with mimalloc, these
+// can be more performant than the standard new-delete operations.
+// See <https://en.cppreference.com/w/cpp/memory/new/operator_new>
+// ---------------------------------------------------------------------------
+#if defined(__cplusplus)
+  #include <new>
+  #include <mimalloc.h>
+
+  #if defined(_MSC_VER) && defined(_Ret_notnull_) && defined(_Post_writable_byte_size_)
+  // stay consistent with VCRT definitions
+  #define mi_decl_new(n)          mi_decl_nodiscard mi_decl_restrict _Ret_notnull_ _Post_writable_byte_size_(n)
+  #define mi_decl_new_nothrow(n)  mi_decl_nodiscard mi_decl_restrict _Ret_maybenull_ _Success_(return != NULL) _Post_writable_byte_size_(n)
+  #else
+  #define mi_decl_new(n)          mi_decl_nodiscard mi_decl_restrict
+  #define mi_decl_new_nothrow(n)  mi_decl_nodiscard mi_decl_restrict
+  #endif
+
+  void operator delete(void* p) noexcept              { mi_free(p); };
+  void operator delete[](void* p) noexcept            { mi_free(p); };
+
+  void operator delete  (void* p, const std::nothrow_t&) noexcept { mi_free(p); }
+  void operator delete[](void* p, const std::nothrow_t&) noexcept { mi_free(p); }
+
+  mi_decl_new(n) void* operator new(std::size_t n) noexcept(false) { return mi_new(n); }
+  mi_decl_new(n) void* operator new[](std::size_t n) noexcept(false) { return mi_new(n); }
+
+  mi_decl_new_nothrow(n) void* operator new  (std::size_t n, const std::nothrow_t& tag) noexcept { (void)(tag); return mi_new_nothrow(n); }
+  mi_decl_new_nothrow(n) void* operator new[](std::size_t n, const std::nothrow_t& tag) noexcept { (void)(tag); return mi_new_nothrow(n); }
+
+  #if (__cplusplus >= 201402L || _MSC_VER >= 1916)
+  void operator delete  (void* p, std::size_t n) noexcept { mi_free_size(p,n); };
+  void operator delete[](void* p, std::size_t n) noexcept { mi_free_size(p,n); };
+  #endif
+
+  #if (__cplusplus > 201402L || defined(__cpp_aligned_new))
+  void operator delete  (void* p, std::align_val_t al) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+  void operator delete[](void* p, std::align_val_t al) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+  void operator delete  (void* p, std::size_t n, std::align_val_t al) noexcept { mi_free_size_aligned(p, n, static_cast<size_t>(al)); };
+  void operator delete[](void* p, std::size_t n, std::align_val_t al) noexcept { mi_free_size_aligned(p, n, static_cast<size_t>(al)); };
+  void operator delete  (void* p, std::align_val_t al, const std::nothrow_t&) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+  void operator delete[](void* p, std::align_val_t al, const std::nothrow_t&) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+
+  void* operator new  (std::size_t n, std::align_val_t al) noexcept(false) { return mi_new_aligned(n, static_cast<size_t>(al)); }
+  void* operator new[](std::size_t n, std::align_val_t al) noexcept(false) { return mi_new_aligned(n, static_cast<size_t>(al)); }
+  void* operator new  (std::size_t n, std::align_val_t al, const std::nothrow_t&) noexcept { return mi_new_aligned_nothrow(n, static_cast<size_t>(al)); }
+  void* operator new[](std::size_t n, std::align_val_t al, const std::nothrow_t&) noexcept { return mi_new_aligned_nothrow(n, static_cast<size_t>(al)); }
+  #endif
+#endif
+
+#endif // MIMALLOC_NEW_DELETE_H

preload-mimalloc/mimalloc/include/mimalloc-override.h

@@ -0,0 +1,67 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2020 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_OVERRIDE_H
+#define MIMALLOC_OVERRIDE_H
+
+/* ----------------------------------------------------------------------------
+This header can be used to statically redirect malloc/free and new/delete
+to the mimalloc variants. This can be useful if one can include this file on
+each source file in a project (but be careful when using external code to
+not accidentally mix pointers from different allocators).
+-----------------------------------------------------------------------------*/
+
+#include <mimalloc.h>
+
+// Standard C allocation
+#define malloc(n)               mi_malloc(n)
+#define calloc(n,c)             mi_calloc(n,c)
+#define realloc(p,n)            mi_realloc(p,n)
+#define free(p)                 mi_free(p)
+
+#define strdup(s)               mi_strdup(s)
+#define strndup(s,n)              mi_strndup(s,n)
+#define realpath(f,n)           mi_realpath(f,n)
+
+// Microsoft extensions
+#define _expand(p,n)            mi_expand(p,n)
+#define _msize(p)               mi_usable_size(p)
+#define _recalloc(p,n,c)        mi_recalloc(p,n,c)
+
+#define _strdup(s)              mi_strdup(s)
+#define _strndup(s,n)           mi_strndup(s,n)
+#define _wcsdup(s)              (wchar_t*)mi_wcsdup((const unsigned short*)(s))
+#define _mbsdup(s)              mi_mbsdup(s)
+#define _dupenv_s(b,n,v)        mi_dupenv_s(b,n,v)
+#define _wdupenv_s(b,n,v)       mi_wdupenv_s((unsigned short*)(b),n,(const unsigned short*)(v))
+
+// Various Posix and Unix variants
+#define reallocf(p,n)           mi_reallocf(p,n)
+#define malloc_size(p)          mi_usable_size(p)
+#define malloc_usable_size(p)   mi_usable_size(p)
+#define cfree(p)                mi_free(p)
+
+#define valloc(n)               mi_valloc(n)
+#define pvalloc(n)              mi_pvalloc(n)
+#define reallocarray(p,s,n)     mi_reallocarray(p,s,n)
+#define reallocarr(p,s,n)       mi_reallocarr(p,s,n)
+#define memalign(a,n)           mi_memalign(a,n)
+#define aligned_alloc(a,n)      mi_aligned_alloc(a,n)
+#define posix_memalign(p,a,n)   mi_posix_memalign(p,a,n)
+#define _posix_memalign(p,a,n)  mi_posix_memalign(p,a,n)
+
+// Microsoft aligned variants
+#define _aligned_malloc(n,a)                  mi_malloc_aligned(n,a)
+#define _aligned_realloc(p,n,a)               mi_realloc_aligned(p,n,a)
+#define _aligned_recalloc(p,s,n,a)            mi_aligned_recalloc(p,s,n,a)
+#define _aligned_msize(p,a,o)                 mi_usable_size(p)
+#define _aligned_free(p)                      mi_free(p)
+#define _aligned_offset_malloc(n,a,o)         mi_malloc_aligned_at(n,a,o)
+#define _aligned_offset_realloc(p,n,a,o)      mi_realloc_aligned_at(p,n,a,o)
+#define _aligned_offset_recalloc(p,s,n,a,o)   mi_recalloc_aligned_at(p,s,n,a,o)
+
+#endif // MIMALLOC_OVERRIDE_H

preload-mimalloc/mimalloc/include/mimalloc.h

@@ -0,0 +1,559 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_H
+#define MIMALLOC_H
+
+#define MI_MALLOC_VERSION 211   // major + 2 digits minor
+
+// ------------------------------------------------------
+// Compiler specific attributes
+// ------------------------------------------------------
+
+#ifdef __cplusplus
+  #if (__cplusplus >= 201103L) || (_MSC_VER > 1900)  // C++11
+    #define mi_attr_noexcept   noexcept
+  #else
+    #define mi_attr_noexcept   throw()
+  #endif
+#else
+  #define mi_attr_noexcept
+#endif
+
+#if defined(__cplusplus) && (__cplusplus >= 201703)
+  #define mi_decl_nodiscard    [[nodiscard]]
+#elif (defined(__GNUC__) && (__GNUC__ >= 4)) || defined(__clang__)  // includes clang, icc, and clang-cl
+  #define mi_decl_nodiscard    __attribute__((warn_unused_result))
+#elif defined(_HAS_NODISCARD)
+  #define mi_decl_nodiscard    _NODISCARD
+#elif (_MSC_VER >= 1700)
+  #define mi_decl_nodiscard    _Check_return_
+#else
+  #define mi_decl_nodiscard
+#endif
+
+#if defined(_MSC_VER) || defined(__MINGW32__)
+  #if !defined(MI_SHARED_LIB)
+    #define mi_decl_export
+  #elif defined(MI_SHARED_LIB_EXPORT)
+    #define mi_decl_export              __declspec(dllexport)
+  #else
+    #define mi_decl_export              __declspec(dllimport)
+  #endif
+  #if defined(__MINGW32__)
+    #define mi_decl_restrict
+    #define mi_attr_malloc              __attribute__((malloc))
+  #else
+    #if (_MSC_VER >= 1900) && !defined(__EDG__)
+      #define mi_decl_restrict          __declspec(allocator) __declspec(restrict)
+    #else
+      #define mi_decl_restrict          __declspec(restrict)
+    #endif
+    #define mi_attr_malloc
+  #endif
+  #define mi_cdecl                      __cdecl
+  #define mi_attr_alloc_size(s)
+  #define mi_attr_alloc_size2(s1,s2)
+  #define mi_attr_alloc_align(p)
+#elif defined(__GNUC__)                 // includes clang and icc
+  #if defined(MI_SHARED_LIB) && defined(MI_SHARED_LIB_EXPORT)
+    #define mi_decl_export              __attribute__((visibility("default")))
+  #else
+    #define mi_decl_export
+  #endif
+  #define mi_cdecl                      // leads to warnings... __attribute__((cdecl))
+  #define mi_decl_restrict
+  #define mi_attr_malloc                __attribute__((malloc))
+  #if (defined(__clang_major__) && (__clang_major__ < 4)) || (__GNUC__ < 5)
+    #define mi_attr_alloc_size(s)
+    #define mi_attr_alloc_size2(s1,s2)
+    #define mi_attr_alloc_align(p)
+  #elif defined(__INTEL_COMPILER)
+    #define mi_attr_alloc_size(s)       __attribute__((alloc_size(s)))
+    #define mi_attr_alloc_size2(s1,s2)  __attribute__((alloc_size(s1,s2)))
+    #define mi_attr_alloc_align(p)
+  #else
+    #define mi_attr_alloc_size(s)       __attribute__((alloc_size(s)))
+    #define mi_attr_alloc_size2(s1,s2)  __attribute__((alloc_size(s1,s2)))
+    #define mi_attr_alloc_align(p)      __attribute__((alloc_align(p)))
+  #endif
+#else
+  #define mi_cdecl
+  #define mi_decl_export
+  #define mi_decl_restrict
+  #define mi_attr_malloc
+  #define mi_attr_alloc_size(s)
+  #define mi_attr_alloc_size2(s1,s2)
+  #define mi_attr_alloc_align(p)
+#endif
+
+// ------------------------------------------------------
+// Includes
+// ------------------------------------------------------
+
+#include <stddef.h>     // size_t
+#include <stdbool.h>    // bool
+#include <stdint.h>     // INTPTR_MAX
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// ------------------------------------------------------
+// Standard malloc interface
+// ------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc(size_t size)  mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_calloc(size_t count, size_t size)  mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2);
+mi_decl_nodiscard mi_decl_export void* mi_realloc(void* p, size_t newsize)      mi_attr_noexcept mi_attr_alloc_size(2);
+mi_decl_export void* mi_expand(void* p, size_t newsize)                         mi_attr_noexcept mi_attr_alloc_size(2);
+
+mi_decl_export void mi_free(void* p) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_strdup(const char* s) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_strndup(const char* s, size_t n) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_realpath(const char* fname, char* resolved_name) mi_attr_noexcept mi_attr_malloc;
+
+// ------------------------------------------------------
+// Extended functionality
+// ------------------------------------------------------
+#define MI_SMALL_WSIZE_MAX  (128)
+#define MI_SMALL_SIZE_MAX   (MI_SMALL_WSIZE_MAX*sizeof(void*))
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc_small(size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc_small(size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc(size_t size)       mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_mallocn(size_t count, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2);
+mi_decl_nodiscard mi_decl_export void* mi_reallocn(void* p, size_t count, size_t size)        mi_attr_noexcept mi_attr_alloc_size2(2,3);
+mi_decl_nodiscard mi_decl_export void* mi_reallocf(void* p, size_t newsize)                   mi_attr_noexcept mi_attr_alloc_size(2);
+
+mi_decl_nodiscard mi_decl_export size_t mi_usable_size(const void* p) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_good_size(size_t size)     mi_attr_noexcept;
+
+
+// ------------------------------------------------------
+// Internals
+// ------------------------------------------------------
+
+typedef void (mi_cdecl mi_deferred_free_fun)(bool force, unsigned long long heartbeat, void* arg);
+mi_decl_export void mi_register_deferred_free(mi_deferred_free_fun* deferred_free, void* arg) mi_attr_noexcept;
+
+typedef void (mi_cdecl mi_output_fun)(const char* msg, void* arg);
+mi_decl_export void mi_register_output(mi_output_fun* out, void* arg) mi_attr_noexcept;
+
+typedef void (mi_cdecl mi_error_fun)(int err, void* arg);
+mi_decl_export void mi_register_error(mi_error_fun* fun, void* arg);
+
+mi_decl_export void mi_collect(bool force)    mi_attr_noexcept;
+mi_decl_export int  mi_version(void)          mi_attr_noexcept;
+mi_decl_export void mi_stats_reset(void)      mi_attr_noexcept;
+mi_decl_export void mi_stats_merge(void)      mi_attr_noexcept;
+mi_decl_export void mi_stats_print(void* out) mi_attr_noexcept;  // backward compatibility: `out` is ignored and should be NULL
+mi_decl_export void mi_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept;
+
+mi_decl_export void mi_process_init(void)     mi_attr_noexcept;
+mi_decl_export void mi_thread_init(void)      mi_attr_noexcept;
+mi_decl_export void mi_thread_done(void)      mi_attr_noexcept;
+mi_decl_export void mi_thread_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept;
+
+mi_decl_export void mi_process_info(size_t* elapsed_msecs, size_t* user_msecs, size_t* system_msecs,
+                                    size_t* current_rss, size_t* peak_rss,
+                                    size_t* current_commit, size_t* peak_commit, size_t* page_faults) mi_attr_noexcept;
+
+// -------------------------------------------------------------------------------------
+// Aligned allocation
+// Note that `alignment` always follows `size` for consistency with unaligned
+// allocation, but unfortunately this differs from `posix_memalign` and `aligned_alloc`.
+// -------------------------------------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc_aligned(size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_malloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc_aligned(size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_zalloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_calloc_aligned(size_t count, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_calloc_aligned_at(size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(1,2);
+mi_decl_nodiscard mi_decl_export void* mi_realloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export void* mi_realloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(2);
+
+
+// -------------------------------------------------------------------------------------
+// Heaps: first-class, but can only allocate from the same thread that created it.
+// -------------------------------------------------------------------------------------
+
+struct mi_heap_s;
+typedef struct mi_heap_s mi_heap_t;
+
+mi_decl_nodiscard mi_decl_export mi_heap_t* mi_heap_new(void);
+mi_decl_export void       mi_heap_delete(mi_heap_t* heap);
+mi_decl_export void       mi_heap_destroy(mi_heap_t* heap);
+mi_decl_export mi_heap_t* mi_heap_set_default(mi_heap_t* heap);
+mi_decl_export mi_heap_t* mi_heap_get_default(void);
+mi_decl_export mi_heap_t* mi_heap_get_backing(void);
+mi_decl_export void       mi_heap_collect(mi_heap_t* heap, bool force) mi_attr_noexcept;
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc(mi_heap_t* heap, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_zalloc(mi_heap_t* heap, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_calloc(mi_heap_t* heap, size_t count, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_mallocn(mi_heap_t* heap, size_t count, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc_small(mi_heap_t* heap, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+
+mi_decl_nodiscard mi_decl_export void* mi_heap_realloc(mi_heap_t* heap, void* p, size_t newsize)              mi_attr_noexcept mi_attr_alloc_size(3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_reallocn(mi_heap_t* heap, void* p, size_t count, size_t size)  mi_attr_noexcept mi_attr_alloc_size2(3,4);
+mi_decl_nodiscard mi_decl_export void* mi_heap_reallocf(mi_heap_t* heap, void* p, size_t newsize)             mi_attr_noexcept mi_attr_alloc_size(3);
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_heap_strdup(mi_heap_t* heap, const char* s)            mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_heap_strndup(mi_heap_t* heap, const char* s, size_t n) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict char* mi_heap_realpath(mi_heap_t* heap, const char* fname, char* resolved_name) mi_attr_noexcept mi_attr_malloc;
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_malloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_zalloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_zalloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_calloc_aligned(mi_heap_t* heap, size_t count, size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_calloc_aligned_at(mi_heap_t* heap, size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size2(2, 3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_realloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export void* mi_heap_realloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(3);
+
+
+// --------------------------------------------------------------------------------
+// Zero initialized re-allocation.
+// Only valid on memory that was originally allocated with zero initialization too.
+// e.g. `mi_calloc`, `mi_zalloc`, `mi_zalloc_aligned` etc.
+// see <https://github.com/microsoft/mimalloc/issues/63#issuecomment-508272992>
+// --------------------------------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_export void* mi_rezalloc(void* p, size_t newsize)                mi_attr_noexcept mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export void* mi_recalloc(void* p, size_t newcount, size_t size)  mi_attr_noexcept mi_attr_alloc_size2(2,3);
+
+mi_decl_nodiscard mi_decl_export void* mi_rezalloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(2) mi_attr_alloc_align(3);
+mi_decl_nodiscard mi_decl_export void* mi_rezalloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export void* mi_recalloc_aligned(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept mi_attr_alloc_size2(2,3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export void* mi_recalloc_aligned_at(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size2(2,3);
+
+mi_decl_nodiscard mi_decl_export void* mi_heap_rezalloc(mi_heap_t* heap, void* p, size_t newsize)                mi_attr_noexcept mi_attr_alloc_size(3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_recalloc(mi_heap_t* heap, void* p, size_t newcount, size_t size)  mi_attr_noexcept mi_attr_alloc_size2(3,4);
+
+mi_decl_nodiscard mi_decl_export void* mi_heap_rezalloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept mi_attr_alloc_size(3) mi_attr_alloc_align(4);
+mi_decl_nodiscard mi_decl_export void* mi_heap_rezalloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size(3);
+mi_decl_nodiscard mi_decl_export void* mi_heap_recalloc_aligned(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept mi_attr_alloc_size2(3,4) mi_attr_alloc_align(5);
+mi_decl_nodiscard mi_decl_export void* mi_heap_recalloc_aligned_at(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept mi_attr_alloc_size2(3,4);
+
+
+// ------------------------------------------------------
+// Analysis
+// ------------------------------------------------------
+
+mi_decl_export bool mi_heap_contains_block(mi_heap_t* heap, const void* p);
+mi_decl_export bool mi_heap_check_owned(mi_heap_t* heap, const void* p);
+mi_decl_export bool mi_check_owned(const void* p);
+
+// An area of heap space contains blocks of a single size.
+typedef struct mi_heap_area_s {
+  void*  blocks;      // start of the area containing heap blocks
+  size_t reserved;    // bytes reserved for this area (virtual)
+  size_t committed;   // current available bytes for this area
+  size_t used;        // number of allocated blocks
+  size_t block_size;  // size in bytes of each block
+  size_t full_block_size; // size in bytes of a full block including padding and metadata.
+} mi_heap_area_t;
+
+typedef bool (mi_cdecl mi_block_visit_fun)(const mi_heap_t* heap, const mi_heap_area_t* area, void* block, size_t block_size, void* arg);
+
+mi_decl_export bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_all_blocks, mi_block_visit_fun* visitor, void* arg);
+
+// Experimental
+mi_decl_nodiscard mi_decl_export bool mi_is_in_heap_region(const void* p) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export bool mi_is_redirected(void) mi_attr_noexcept;
+
+mi_decl_export int mi_reserve_huge_os_pages_interleave(size_t pages, size_t numa_nodes, size_t timeout_msecs) mi_attr_noexcept;
+mi_decl_export int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size_t timeout_msecs) mi_attr_noexcept;
+
+mi_decl_export int  mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noexcept;
+mi_decl_export bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node) mi_attr_noexcept;
+
+mi_decl_export void mi_debug_show_arenas(void) mi_attr_noexcept;
+
+// Experimental: heaps associated with specific memory arena's
+typedef int mi_arena_id_t;
+mi_decl_export void* mi_arena_area(mi_arena_id_t arena_id, size_t* size);
+mi_decl_export int   mi_reserve_huge_os_pages_at_ex(size_t pages, int numa_node, size_t timeout_msecs, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept;
+mi_decl_export int   mi_reserve_os_memory_ex(size_t size, bool commit, bool allow_large, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept;
+mi_decl_export bool  mi_manage_os_memory_ex(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept;
+
+#if MI_MALLOC_VERSION >= 200
+// Create a heap that only allocates in the specified arena
+mi_decl_nodiscard mi_decl_export mi_heap_t* mi_heap_new_in_arena(mi_arena_id_t arena_id);
+#endif
+
+// deprecated
+mi_decl_export int  mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept;
+
+
+// ------------------------------------------------------
+// Convenience
+// ------------------------------------------------------
+
+#define mi_malloc_tp(tp)                ((tp*)mi_malloc(sizeof(tp)))
+#define mi_zalloc_tp(tp)                ((tp*)mi_zalloc(sizeof(tp)))
+#define mi_calloc_tp(tp,n)              ((tp*)mi_calloc(n,sizeof(tp)))
+#define mi_mallocn_tp(tp,n)             ((tp*)mi_mallocn(n,sizeof(tp)))
+#define mi_reallocn_tp(p,tp,n)          ((tp*)mi_reallocn(p,n,sizeof(tp)))
+#define mi_recalloc_tp(p,tp,n)          ((tp*)mi_recalloc(p,n,sizeof(tp)))
+
+#define mi_heap_malloc_tp(hp,tp)        ((tp*)mi_heap_malloc(hp,sizeof(tp)))
+#define mi_heap_zalloc_tp(hp,tp)        ((tp*)mi_heap_zalloc(hp,sizeof(tp)))
+#define mi_heap_calloc_tp(hp,tp,n)      ((tp*)mi_heap_calloc(hp,n,sizeof(tp)))
+#define mi_heap_mallocn_tp(hp,tp,n)     ((tp*)mi_heap_mallocn(hp,n,sizeof(tp)))
+#define mi_heap_reallocn_tp(hp,p,tp,n)  ((tp*)mi_heap_reallocn(hp,p,n,sizeof(tp)))
+#define mi_heap_recalloc_tp(hp,p,tp,n)  ((tp*)mi_heap_recalloc(hp,p,n,sizeof(tp)))
+
+
+// ------------------------------------------------------
+// Options
+// ------------------------------------------------------
+
+typedef enum mi_option_e {
+  // stable options
+  mi_option_show_errors,
+  mi_option_show_stats,
+  mi_option_verbose,
+  // some of the following options are experimental
+  // (deprecated options are kept for binary backward compatibility with v1.x versions)
+  mi_option_eager_commit,
+  mi_option_deprecated_eager_region_commit,
+  mi_option_deprecated_reset_decommits,
+  mi_option_large_os_pages,           // use large (2MiB) OS pages, implies eager commit
+  mi_option_reserve_huge_os_pages,    // reserve N huge OS pages (1GiB) at startup
+  mi_option_reserve_huge_os_pages_at, // reserve huge OS pages at a specific NUMA node
+  mi_option_reserve_os_memory,        // reserve specified amount of OS memory at startup
+  mi_option_deprecated_segment_cache,
+  mi_option_page_reset,
+  mi_option_abandoned_page_decommit,
+  mi_option_deprecated_segment_reset,
+  mi_option_eager_commit_delay,
+  mi_option_decommit_delay,
+  mi_option_use_numa_nodes,           // 0 = use available numa nodes, otherwise use at most N nodes.
+  mi_option_limit_os_alloc,           // 1 = do not use OS memory for allocation (but only reserved arenas)
+  mi_option_os_tag,
+  mi_option_max_errors,
+  mi_option_max_warnings,
+  mi_option_max_segment_reclaim,
+  mi_option_allow_decommit,
+  mi_option_segment_decommit_delay,  
+  mi_option_decommit_extend_delay,
+  mi_option_destroy_on_exit,          
+  _mi_option_last
+} mi_option_t;
+
+
+mi_decl_nodiscard mi_decl_export bool mi_option_is_enabled(mi_option_t option);
+mi_decl_export void mi_option_enable(mi_option_t option);
+mi_decl_export void mi_option_disable(mi_option_t option);
+mi_decl_export void mi_option_set_enabled(mi_option_t option, bool enable);
+mi_decl_export void mi_option_set_enabled_default(mi_option_t option, bool enable);
+
+mi_decl_nodiscard mi_decl_export long mi_option_get(mi_option_t option);
+mi_decl_nodiscard mi_decl_export long mi_option_get_clamp(mi_option_t option, long min, long max);
+mi_decl_export void mi_option_set(mi_option_t option, long value);
+mi_decl_export void mi_option_set_default(mi_option_t option, long value);
+
+
+// -------------------------------------------------------------------------------------------------------
+// "mi" prefixed implementations of various posix, Unix, Windows, and C++ allocation functions.
+// (This can be convenient when providing overrides of these functions as done in `mimalloc-override.h`.)
+// note: we use `mi_cfree` as "checked free" and it checks if the pointer is in our heap before free-ing.
+// -------------------------------------------------------------------------------------------------------
+
+mi_decl_export void  mi_cfree(void* p) mi_attr_noexcept;
+mi_decl_export void* mi__expand(void* p, size_t newsize) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_malloc_size(const void* p)        mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_malloc_good_size(size_t size)     mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export size_t mi_malloc_usable_size(const void *p) mi_attr_noexcept;
+
+mi_decl_export int mi_posix_memalign(void** p, size_t alignment, size_t size)   mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_memalign(size_t alignment, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_valloc(size_t size)  mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_pvalloc(size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_aligned_alloc(size_t alignment, size_t size) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(2) mi_attr_alloc_align(1);
+
+mi_decl_nodiscard mi_decl_export void* mi_reallocarray(void* p, size_t count, size_t size) mi_attr_noexcept mi_attr_alloc_size2(2,3);
+mi_decl_nodiscard mi_decl_export int   mi_reallocarr(void* p, size_t count, size_t size) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export void* mi_aligned_recalloc(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept;
+mi_decl_nodiscard mi_decl_export void* mi_aligned_offset_recalloc(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept;
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict unsigned short* mi_wcsdup(const unsigned short* s) mi_attr_noexcept mi_attr_malloc;
+mi_decl_nodiscard mi_decl_export mi_decl_restrict unsigned char*  mi_mbsdup(const unsigned char* s)  mi_attr_noexcept mi_attr_malloc;
+mi_decl_export int mi_dupenv_s(char** buf, size_t* size, const char* name)                      mi_attr_noexcept;
+mi_decl_export int mi_wdupenv_s(unsigned short** buf, size_t* size, const unsigned short* name) mi_attr_noexcept;
+
+mi_decl_export void mi_free_size(void* p, size_t size)                           mi_attr_noexcept;
+mi_decl_export void mi_free_size_aligned(void* p, size_t size, size_t alignment) mi_attr_noexcept;
+mi_decl_export void mi_free_aligned(void* p, size_t alignment)                   mi_attr_noexcept;
+
+// The `mi_new` wrappers implement C++ semantics on out-of-memory instead of directly returning `NULL`.
+// (and call `std::get_new_handler` and potentially raise a `std::bad_alloc` exception).
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new(size_t size)                   mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_aligned(size_t size, size_t alignment) mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_nothrow(size_t size)           mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_aligned_nothrow(size_t size, size_t alignment) mi_attr_noexcept mi_attr_malloc mi_attr_alloc_size(1) mi_attr_alloc_align(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_new_n(size_t count, size_t size)   mi_attr_malloc mi_attr_alloc_size2(1, 2);
+mi_decl_nodiscard mi_decl_export void* mi_new_realloc(void* p, size_t newsize)                mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export void* mi_new_reallocn(void* p, size_t newcount, size_t size) mi_attr_alloc_size2(2, 3);
+
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_alloc_new(mi_heap_t* heap, size_t size)                mi_attr_malloc mi_attr_alloc_size(2);
+mi_decl_nodiscard mi_decl_export mi_decl_restrict void* mi_heap_alloc_new_n(mi_heap_t* heap, size_t count, size_t size) mi_attr_malloc mi_attr_alloc_size2(2, 3);
+
+#ifdef __cplusplus
+}
+#endif
+
+// ---------------------------------------------------------------------------------------------
+// Implement the C++ std::allocator interface for use in STL containers.
+// (note: see `mimalloc-new-delete.h` for overriding the new/delete operators globally)
+// ---------------------------------------------------------------------------------------------
+#ifdef __cplusplus
+
+#include <cstddef>     // std::size_t
+#include <cstdint>     // PTRDIFF_MAX
+#if (__cplusplus >= 201103L) || (_MSC_VER > 1900)  // C++11
+#include <type_traits> // std::true_type
+#include <utility>     // std::forward
+#endif
+
+template<class T> struct _mi_stl_allocator_common {
+  typedef T                 value_type;
+  typedef std::size_t       size_type;
+  typedef std::ptrdiff_t    difference_type;
+  typedef value_type&       reference;
+  typedef value_type const& const_reference;
+  typedef value_type*       pointer;
+  typedef value_type const* const_pointer;
+
+  #if ((__cplusplus >= 201103L) || (_MSC_VER > 1900))  // C++11
+  using propagate_on_container_copy_assignment = std::true_type;
+  using propagate_on_container_move_assignment = std::true_type;
+  using propagate_on_container_swap            = std::true_type;
+  template <class U, class ...Args> void construct(U* p, Args&& ...args) { ::new(p) U(std::forward<Args>(args)...); }
+  template <class U> void destroy(U* p) mi_attr_noexcept { p->~U(); }
+  #else
+  void construct(pointer p, value_type const& val) { ::new(p) value_type(val); }
+  void destroy(pointer p) { p->~value_type(); }
+  #endif
+
+  size_type     max_size() const mi_attr_noexcept { return (PTRDIFF_MAX/sizeof(value_type)); }
+  pointer       address(reference x) const        { return &x; }
+  const_pointer address(const_reference x) const  { return &x; }
+};
+
+template<class T> struct mi_stl_allocator : public _mi_stl_allocator_common<T> {
+  using typename _mi_stl_allocator_common<T>::size_type;
+  using typename _mi_stl_allocator_common<T>::value_type;
+  using typename _mi_stl_allocator_common<T>::pointer;
+  template <class U> struct rebind { typedef mi_stl_allocator<U> other; };
+
+  mi_stl_allocator()                                             mi_attr_noexcept = default;
+  mi_stl_allocator(const mi_stl_allocator&)                      mi_attr_noexcept = default;
+  template<class U> mi_stl_allocator(const mi_stl_allocator<U>&) mi_attr_noexcept { }
+  mi_stl_allocator  select_on_container_copy_construction() const { return *this; }
+  void              deallocate(T* p, size_type) { mi_free(p); }
+
+  #if (__cplusplus >= 201703L)  // C++17
+  mi_decl_nodiscard T* allocate(size_type count) { return static_cast<T*>(mi_new_n(count, sizeof(T))); }
+  mi_decl_nodiscard T* allocate(size_type count, const void*) { return allocate(count); }
+  #else
+  mi_decl_nodiscard pointer allocate(size_type count, const void* = 0) { return static_cast<pointer>(mi_new_n(count, sizeof(value_type))); }
+  #endif
+
+  #if ((__cplusplus >= 201103L) || (_MSC_VER > 1900))  // C++11
+  using is_always_equal = std::true_type;
+  #endif
+};
+
+template<class T1,class T2> bool operator==(const mi_stl_allocator<T1>& , const mi_stl_allocator<T2>& ) mi_attr_noexcept { return true; }
+template<class T1,class T2> bool operator!=(const mi_stl_allocator<T1>& , const mi_stl_allocator<T2>& ) mi_attr_noexcept { return false; }
+
+
+#if (__cplusplus >= 201103L) || (_MSC_VER >= 1900)  // C++11
+#define MI_HAS_HEAP_STL_ALLOCATOR 1
+
+#include <memory>      // std::shared_ptr
+
+// Common base class for STL allocators in a specific heap
+template<class T, bool _mi_destroy> struct _mi_heap_stl_allocator_common : public _mi_stl_allocator_common<T> {
+  using typename _mi_stl_allocator_common<T>::size_type;
+  using typename _mi_stl_allocator_common<T>::value_type;
+  using typename _mi_stl_allocator_common<T>::pointer;
+
+  _mi_heap_stl_allocator_common(mi_heap_t* hp) : heap(hp) { }    /* will not delete nor destroy the passed in heap */
+
+  #if (__cplusplus >= 201703L)  // C++17
+  mi_decl_nodiscard T* allocate(size_type count) { return static_cast<T*>(mi_heap_alloc_new_n(this->heap.get(), count, sizeof(T))); }
+  mi_decl_nodiscard T* allocate(size_type count, const void*) { return allocate(count); }
+  #else
+  mi_decl_nodiscard pointer allocate(size_type count, const void* = 0) { return static_cast<pointer>(mi_heap_alloc_new_n(this->heap.get(), count, sizeof(value_type))); }
+  #endif
+
+  #if ((__cplusplus >= 201103L) || (_MSC_VER > 1900))  // C++11
+  using is_always_equal = std::false_type;
+  #endif
+
+  void collect(bool force) { mi_heap_collect(this->heap.get(), force); }
+  template<class U> bool is_equal(const _mi_heap_stl_allocator_common<U, _mi_destroy>& x) const { return (this->heap == x.heap); }
+
+protected:
+  std::shared_ptr<mi_heap_t> heap;
+  template<class U, bool D> friend struct _mi_heap_stl_allocator_common;
+  
+  _mi_heap_stl_allocator_common() {
+    mi_heap_t* hp = mi_heap_new();
+    this->heap.reset(hp, (_mi_destroy ? &heap_destroy : &heap_delete));  /* calls heap_delete/destroy when the refcount drops to zero */
+  }
+  _mi_heap_stl_allocator_common(const _mi_heap_stl_allocator_common& x) mi_attr_noexcept : heap(x.heap) { }
+  template<class U> _mi_heap_stl_allocator_common(const _mi_heap_stl_allocator_common<U, _mi_destroy>& x) mi_attr_noexcept : heap(x.heap) { }
+
+private:
+  static void heap_delete(mi_heap_t* hp)  { if (hp != NULL) { mi_heap_delete(hp); } }
+  static void heap_destroy(mi_heap_t* hp) { if (hp != NULL) { mi_heap_destroy(hp); } }
+};
+
+// STL allocator allocation in a specific heap
+template<class T> struct mi_heap_stl_allocator : public _mi_heap_stl_allocator_common<T, false> {
+  using typename _mi_heap_stl_allocator_common<T, false>::size_type;
+  mi_heap_stl_allocator() : _mi_heap_stl_allocator_common<T, false>() { } // creates fresh heap that is deleted when the destructor is called
+  mi_heap_stl_allocator(mi_heap_t* hp) : _mi_heap_stl_allocator_common<T, false>(hp) { }  // no delete nor destroy on the passed in heap 
+  template<class U> mi_heap_stl_allocator(const mi_heap_stl_allocator<U>& x) mi_attr_noexcept : _mi_heap_stl_allocator_common<T, false>(x) { }
+
+  mi_heap_stl_allocator select_on_container_copy_construction() const { return *this; }
+  void deallocate(T* p, size_type) { mi_free(p); }
+  template<class U> struct rebind { typedef mi_heap_stl_allocator<U> other; };
+};
+
+template<class T1, class T2> bool operator==(const mi_heap_stl_allocator<T1>& x, const mi_heap_stl_allocator<T2>& y) mi_attr_noexcept { return (x.is_equal(y)); }
+template<class T1, class T2> bool operator!=(const mi_heap_stl_allocator<T1>& x, const mi_heap_stl_allocator<T2>& y) mi_attr_noexcept { return (!x.is_equal(y)); }
+
+
+// STL allocator allocation in a specific heap, where `free` does nothing and
+// the heap is destroyed in one go on destruction -- use with care!
+template<class T> struct mi_heap_destroy_stl_allocator : public _mi_heap_stl_allocator_common<T, true> {
+  using typename _mi_heap_stl_allocator_common<T, true>::size_type;
+  mi_heap_destroy_stl_allocator() : _mi_heap_stl_allocator_common<T, true>() { } // creates fresh heap that is destroyed when the destructor is called
+  mi_heap_destroy_stl_allocator(mi_heap_t* hp) : _mi_heap_stl_allocator_common<T, true>(hp) { }  // no delete nor destroy on the passed in heap 
+  template<class U> mi_heap_destroy_stl_allocator(const mi_heap_destroy_stl_allocator<U>& x) mi_attr_noexcept : _mi_heap_stl_allocator_common<T, true>(x) { }
+
+  mi_heap_destroy_stl_allocator select_on_container_copy_construction() const { return *this; }
+  void deallocate(T*, size_type) { /* do nothing as we destroy the heap on destruct. */ }
+  template<class U> struct rebind { typedef mi_heap_destroy_stl_allocator<U> other; };
+};
+
+template<class T1, class T2> bool operator==(const mi_heap_destroy_stl_allocator<T1>& x, const mi_heap_destroy_stl_allocator<T2>& y) mi_attr_noexcept { return (x.is_equal(y)); }
+template<class T1, class T2> bool operator!=(const mi_heap_destroy_stl_allocator<T1>& x, const mi_heap_destroy_stl_allocator<T2>& y) mi_attr_noexcept { return (!x.is_equal(y)); }
+
+#endif // C++11
+
+#endif // __cplusplus
+
+#endif

preload-mimalloc/mimalloc/include/mimalloc/atomic.h

@@ -0,0 +1,347 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_ATOMIC_H
+#define MIMALLOC_ATOMIC_H
+
+// --------------------------------------------------------------------------------------------
+// Atomics
+// We need to be portable between C, C++, and MSVC.
+// We base the primitives on the C/C++ atomics and create a mimimal wrapper for MSVC in C compilation mode.
+// This is why we try to use only `uintptr_t` and `<type>*` as atomic types.
+// To gain better insight in the range of used atomics, we use explicitly named memory order operations
+// instead of passing the memory order as a parameter.
+// -----------------------------------------------------------------------------------------------
+
+#if defined(__cplusplus)
+// Use C++ atomics
+#include <atomic>
+#define  _Atomic(tp)            std::atomic<tp>
+#define  mi_atomic(name)        std::atomic_##name
+#define  mi_memory_order(name)  std::memory_order_##name
+#if !defined(ATOMIC_VAR_INIT) || (__cplusplus >= 202002L) // c++20, see issue #571
+ #define MI_ATOMIC_VAR_INIT(x)  x
+#else
+ #define MI_ATOMIC_VAR_INIT(x)  ATOMIC_VAR_INIT(x)
+#endif
+#elif defined(_MSC_VER)
+// Use MSVC C wrapper for C11 atomics
+#define  _Atomic(tp)            tp
+#define  MI_ATOMIC_VAR_INIT(x)  x
+#define  mi_atomic(name)        mi_atomic_##name
+#define  mi_memory_order(name)  mi_memory_order_##name
+#else
+// Use C11 atomics
+#include <stdatomic.h>
+#define  mi_atomic(name)        atomic_##name
+#define  mi_memory_order(name)  memory_order_##name
+#define  MI_ATOMIC_VAR_INIT(x)  ATOMIC_VAR_INIT(x)
+#endif
+
+// Various defines for all used memory orders in mimalloc
+#define mi_atomic_cas_weak(p,expected,desired,mem_success,mem_fail)  \
+  mi_atomic(compare_exchange_weak_explicit)(p,expected,desired,mem_success,mem_fail)
+
+#define mi_atomic_cas_strong(p,expected,desired,mem_success,mem_fail)  \
+  mi_atomic(compare_exchange_strong_explicit)(p,expected,desired,mem_success,mem_fail)
+
+#define mi_atomic_load_acquire(p)                mi_atomic(load_explicit)(p,mi_memory_order(acquire))
+#define mi_atomic_load_relaxed(p)                mi_atomic(load_explicit)(p,mi_memory_order(relaxed))
+#define mi_atomic_store_release(p,x)             mi_atomic(store_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_store_relaxed(p,x)             mi_atomic(store_explicit)(p,x,mi_memory_order(relaxed))
+#define mi_atomic_exchange_release(p,x)          mi_atomic(exchange_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_exchange_acq_rel(p,x)          mi_atomic(exchange_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_cas_weak_release(p,exp,des)    mi_atomic_cas_weak(p,exp,des,mi_memory_order(release),mi_memory_order(relaxed))
+#define mi_atomic_cas_weak_acq_rel(p,exp,des)    mi_atomic_cas_weak(p,exp,des,mi_memory_order(acq_rel),mi_memory_order(acquire))
+#define mi_atomic_cas_strong_release(p,exp,des)  mi_atomic_cas_strong(p,exp,des,mi_memory_order(release),mi_memory_order(relaxed))
+#define mi_atomic_cas_strong_acq_rel(p,exp,des)  mi_atomic_cas_strong(p,exp,des,mi_memory_order(acq_rel),mi_memory_order(acquire))
+
+#define mi_atomic_add_relaxed(p,x)               mi_atomic(fetch_add_explicit)(p,x,mi_memory_order(relaxed))
+#define mi_atomic_sub_relaxed(p,x)               mi_atomic(fetch_sub_explicit)(p,x,mi_memory_order(relaxed))
+#define mi_atomic_add_acq_rel(p,x)               mi_atomic(fetch_add_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_sub_acq_rel(p,x)               mi_atomic(fetch_sub_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_and_acq_rel(p,x)               mi_atomic(fetch_and_explicit)(p,x,mi_memory_order(acq_rel))
+#define mi_atomic_or_acq_rel(p,x)                mi_atomic(fetch_or_explicit)(p,x,mi_memory_order(acq_rel))
+
+#define mi_atomic_increment_relaxed(p)           mi_atomic_add_relaxed(p,(uintptr_t)1)
+#define mi_atomic_decrement_relaxed(p)           mi_atomic_sub_relaxed(p,(uintptr_t)1)
+#define mi_atomic_increment_acq_rel(p)           mi_atomic_add_acq_rel(p,(uintptr_t)1)
+#define mi_atomic_decrement_acq_rel(p)           mi_atomic_sub_acq_rel(p,(uintptr_t)1)
+
+static inline void mi_atomic_yield(void);
+static inline intptr_t mi_atomic_addi(_Atomic(intptr_t)*p, intptr_t add);
+static inline intptr_t mi_atomic_subi(_Atomic(intptr_t)*p, intptr_t sub);
+
+
+#if defined(__cplusplus) || !defined(_MSC_VER)
+
+// In C++/C11 atomics we have polymorphic atomics so can use the typed `ptr` variants (where `tp` is the type of atomic value)
+// We use these macros so we can provide a typed wrapper in MSVC in C compilation mode as well
+#define mi_atomic_load_ptr_acquire(tp,p)                mi_atomic_load_acquire(p)
+#define mi_atomic_load_ptr_relaxed(tp,p)                mi_atomic_load_relaxed(p)
+
+// In C++ we need to add casts to help resolve templates if NULL is passed
+#if defined(__cplusplus)
+#define mi_atomic_store_ptr_release(tp,p,x)             mi_atomic_store_release(p,(tp*)x)
+#define mi_atomic_store_ptr_relaxed(tp,p,x)             mi_atomic_store_relaxed(p,(tp*)x)
+#define mi_atomic_cas_ptr_weak_release(tp,p,exp,des)    mi_atomic_cas_weak_release(p,exp,(tp*)des)
+#define mi_atomic_cas_ptr_weak_acq_rel(tp,p,exp,des)    mi_atomic_cas_weak_acq_rel(p,exp,(tp*)des)
+#define mi_atomic_cas_ptr_strong_release(tp,p,exp,des)  mi_atomic_cas_strong_release(p,exp,(tp*)des)
+#define mi_atomic_exchange_ptr_release(tp,p,x)          mi_atomic_exchange_release(p,(tp*)x)
+#define mi_atomic_exchange_ptr_acq_rel(tp,p,x)          mi_atomic_exchange_acq_rel(p,(tp*)x)
+#else
+#define mi_atomic_store_ptr_release(tp,p,x)             mi_atomic_store_release(p,x)
+#define mi_atomic_store_ptr_relaxed(tp,p,x)             mi_atomic_store_relaxed(p,x)
+#define mi_atomic_cas_ptr_weak_release(tp,p,exp,des)    mi_atomic_cas_weak_release(p,exp,des)
+#define mi_atomic_cas_ptr_weak_acq_rel(tp,p,exp,des)    mi_atomic_cas_weak_acq_rel(p,exp,des)
+#define mi_atomic_cas_ptr_strong_release(tp,p,exp,des)  mi_atomic_cas_strong_release(p,exp,des)
+#define mi_atomic_exchange_ptr_release(tp,p,x)          mi_atomic_exchange_release(p,x)
+#define mi_atomic_exchange_ptr_acq_rel(tp,p,x)          mi_atomic_exchange_acq_rel(p,x)
+#endif
+
+// These are used by the statistics
+static inline int64_t mi_atomic_addi64_relaxed(volatile int64_t* p, int64_t add) {
+  return mi_atomic(fetch_add_explicit)((_Atomic(int64_t)*)p, add, mi_memory_order(relaxed));
+}
+static inline void mi_atomic_maxi64_relaxed(volatile int64_t* p, int64_t x) {
+  int64_t current = mi_atomic_load_relaxed((_Atomic(int64_t)*)p);
+  while (current < x && !mi_atomic_cas_weak_release((_Atomic(int64_t)*)p, &current, x)) { /* nothing */ };
+}
+
+// Used by timers
+#define mi_atomic_loadi64_acquire(p)    mi_atomic(load_explicit)(p,mi_memory_order(acquire))
+#define mi_atomic_loadi64_relaxed(p)    mi_atomic(load_explicit)(p,mi_memory_order(relaxed))
+#define mi_atomic_storei64_release(p,x) mi_atomic(store_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_storei64_relaxed(p,x) mi_atomic(store_explicit)(p,x,mi_memory_order(relaxed))
+
+
+
+#elif defined(_MSC_VER)
+
+// MSVC C compilation wrapper that uses Interlocked operations to model C11 atomics.
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include <intrin.h>
+#ifdef _WIN64
+typedef LONG64   msc_intptr_t;
+#define MI_64(f) f##64
+#else
+typedef LONG     msc_intptr_t;
+#define MI_64(f) f
+#endif
+
+typedef enum mi_memory_order_e {
+  mi_memory_order_relaxed,
+  mi_memory_order_consume,
+  mi_memory_order_acquire,
+  mi_memory_order_release,
+  mi_memory_order_acq_rel,
+  mi_memory_order_seq_cst
+} mi_memory_order;
+
+static inline uintptr_t mi_atomic_fetch_add_explicit(_Atomic(uintptr_t)*p, uintptr_t add, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedExchangeAdd)((volatile msc_intptr_t*)p, (msc_intptr_t)add);
+}
+static inline uintptr_t mi_atomic_fetch_sub_explicit(_Atomic(uintptr_t)*p, uintptr_t sub, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedExchangeAdd)((volatile msc_intptr_t*)p, -((msc_intptr_t)sub));
+}
+static inline uintptr_t mi_atomic_fetch_and_explicit(_Atomic(uintptr_t)*p, uintptr_t x, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedAnd)((volatile msc_intptr_t*)p, (msc_intptr_t)x);
+}
+static inline uintptr_t mi_atomic_fetch_or_explicit(_Atomic(uintptr_t)*p, uintptr_t x, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedOr)((volatile msc_intptr_t*)p, (msc_intptr_t)x);
+}
+static inline bool mi_atomic_compare_exchange_strong_explicit(_Atomic(uintptr_t)*p, uintptr_t* expected, uintptr_t desired, mi_memory_order mo1, mi_memory_order mo2) {
+  (void)(mo1); (void)(mo2);
+  uintptr_t read = (uintptr_t)MI_64(_InterlockedCompareExchange)((volatile msc_intptr_t*)p, (msc_intptr_t)desired, (msc_intptr_t)(*expected));
+  if (read == *expected) {
+    return true;
+  }
+  else {
+    *expected = read;
+    return false;
+  }
+}
+static inline bool mi_atomic_compare_exchange_weak_explicit(_Atomic(uintptr_t)*p, uintptr_t* expected, uintptr_t desired, mi_memory_order mo1, mi_memory_order mo2) {
+  return mi_atomic_compare_exchange_strong_explicit(p, expected, desired, mo1, mo2);
+}
+static inline uintptr_t mi_atomic_exchange_explicit(_Atomic(uintptr_t)*p, uintptr_t exchange, mi_memory_order mo) {
+  (void)(mo);
+  return (uintptr_t)MI_64(_InterlockedExchange)((volatile msc_intptr_t*)p, (msc_intptr_t)exchange);
+}
+static inline void mi_atomic_thread_fence(mi_memory_order mo) {
+  (void)(mo);
+  _Atomic(uintptr_t) x = 0;
+  mi_atomic_exchange_explicit(&x, 1, mo);
+}
+static inline uintptr_t mi_atomic_load_explicit(_Atomic(uintptr_t) const* p, mi_memory_order mo) {
+  (void)(mo);
+#if defined(_M_IX86) || defined(_M_X64)
+  return *p;
+#else
+  uintptr_t x = *p;
+  if (mo > mi_memory_order_relaxed) {
+    while (!mi_atomic_compare_exchange_weak_explicit(p, &x, x, mo, mi_memory_order_relaxed)) { /* nothing */ };
+  }
+  return x;
+#endif
+}
+static inline void mi_atomic_store_explicit(_Atomic(uintptr_t)*p, uintptr_t x, mi_memory_order mo) {
+  (void)(mo);
+#if defined(_M_IX86) || defined(_M_X64)
+  *p = x;
+#else
+  mi_atomic_exchange_explicit(p, x, mo);
+#endif
+}
+static inline int64_t mi_atomic_loadi64_explicit(_Atomic(int64_t)*p, mi_memory_order mo) {
+  (void)(mo);
+#if defined(_M_X64)
+  return *p;
+#else
+  int64_t old = *p;
+  int64_t x = old;
+  while ((old = InterlockedCompareExchange64(p, x, old)) != x) {
+    x = old;
+  }
+  return x;
+#endif
+}
+static inline void mi_atomic_storei64_explicit(_Atomic(int64_t)*p, int64_t x, mi_memory_order mo) {
+  (void)(mo);
+#if defined(x_M_IX86) || defined(_M_X64)
+  *p = x;
+#else
+  InterlockedExchange64(p, x);
+#endif
+}
+
+// These are used by the statistics
+static inline int64_t mi_atomic_addi64_relaxed(volatile _Atomic(int64_t)*p, int64_t add) {
+#ifdef _WIN64
+  return (int64_t)mi_atomic_addi((int64_t*)p, add);
+#else
+  int64_t current;
+  int64_t sum;
+  do {
+    current = *p;
+    sum = current + add;
+  } while (_InterlockedCompareExchange64(p, sum, current) != current);
+  return current;
+#endif
+}
+static inline void mi_atomic_maxi64_relaxed(volatile _Atomic(int64_t)*p, int64_t x) {
+  int64_t current;
+  do {
+    current = *p;
+  } while (current < x && _InterlockedCompareExchange64(p, x, current) != current);
+}
+
+// The pointer macros cast to `uintptr_t`.
+#define mi_atomic_load_ptr_acquire(tp,p)                (tp*)mi_atomic_load_acquire((_Atomic(uintptr_t)*)(p))
+#define mi_atomic_load_ptr_relaxed(tp,p)                (tp*)mi_atomic_load_relaxed((_Atomic(uintptr_t)*)(p))
+#define mi_atomic_store_ptr_release(tp,p,x)             mi_atomic_store_release((_Atomic(uintptr_t)*)(p),(uintptr_t)(x))
+#define mi_atomic_store_ptr_relaxed(tp,p,x)             mi_atomic_store_relaxed((_Atomic(uintptr_t)*)(p),(uintptr_t)(x))
+#define mi_atomic_cas_ptr_weak_release(tp,p,exp,des)    mi_atomic_cas_weak_release((_Atomic(uintptr_t)*)(p),(uintptr_t*)exp,(uintptr_t)des)
+#define mi_atomic_cas_ptr_weak_acq_rel(tp,p,exp,des)    mi_atomic_cas_weak_acq_rel((_Atomic(uintptr_t)*)(p),(uintptr_t*)exp,(uintptr_t)des)
+#define mi_atomic_cas_ptr_strong_release(tp,p,exp,des)  mi_atomic_cas_strong_release((_Atomic(uintptr_t)*)(p),(uintptr_t*)exp,(uintptr_t)des)
+#define mi_atomic_exchange_ptr_release(tp,p,x)          (tp*)mi_atomic_exchange_release((_Atomic(uintptr_t)*)(p),(uintptr_t)x)
+#define mi_atomic_exchange_ptr_acq_rel(tp,p,x)          (tp*)mi_atomic_exchange_acq_rel((_Atomic(uintptr_t)*)(p),(uintptr_t)x)
+
+#define mi_atomic_loadi64_acquire(p)    mi_atomic(loadi64_explicit)(p,mi_memory_order(acquire))
+#define mi_atomic_loadi64_relaxed(p)    mi_atomic(loadi64_explicit)(p,mi_memory_order(relaxed))
+#define mi_atomic_storei64_release(p,x) mi_atomic(storei64_explicit)(p,x,mi_memory_order(release))
+#define mi_atomic_storei64_relaxed(p,x) mi_atomic(storei64_explicit)(p,x,mi_memory_order(relaxed))
+
+
+#endif
+
+
+// Atomically add a signed value; returns the previous value.
+static inline intptr_t mi_atomic_addi(_Atomic(intptr_t)*p, intptr_t add) {
+  return (intptr_t)mi_atomic_add_acq_rel((_Atomic(uintptr_t)*)p, (uintptr_t)add);
+}
+
+// Atomically subtract a signed value; returns the previous value.
+static inline intptr_t mi_atomic_subi(_Atomic(intptr_t)*p, intptr_t sub) {
+  return (intptr_t)mi_atomic_addi(p, -sub);
+}
+
+typedef _Atomic(uintptr_t) mi_atomic_once_t;
+
+// Returns true only on the first invocation
+static inline bool mi_atomic_once( mi_atomic_once_t* once ) {
+  if (mi_atomic_load_relaxed(once) != 0) return false;     // quick test 
+  uintptr_t expected = 0;
+  return mi_atomic_cas_strong_acq_rel(once, &expected, 1); // try to set to 1
+}
+
+// Yield
+#if defined(__cplusplus)
+#include <thread>
+static inline void mi_atomic_yield(void) {
+  std::this_thread::yield();
+}
+#elif defined(_WIN32)
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+static inline void mi_atomic_yield(void) {
+  YieldProcessor();
+}
+#elif defined(__SSE2__)
+#include <emmintrin.h>
+static inline void mi_atomic_yield(void) {
+  _mm_pause();
+}
+#elif (defined(__GNUC__) || defined(__clang__)) && \
+      (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__armel__) || defined(__ARMEL__) || \
+       defined(__aarch64__) || defined(__powerpc__) || defined(__ppc__) || defined(__PPC__))
+#if defined(__x86_64__) || defined(__i386__)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile ("pause" ::: "memory");
+}
+#elif defined(__aarch64__)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile("wfe");
+}
+#elif (defined(__arm__) && __ARM_ARCH__ >= 7)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile("yield" ::: "memory");
+}
+#elif defined(__powerpc__) || defined(__ppc__) || defined(__PPC__)
+static inline void mi_atomic_yield(void) {
+  __asm__ __volatile__ ("or 27,27,27" ::: "memory");
+}
+#elif defined(__armel__) || defined(__ARMEL__)
+static inline void mi_atomic_yield(void) {
+  __asm__ volatile ("nop" ::: "memory");
+}
+#endif
+#elif defined(__sun)
+// Fallback for other archs
+#include <synch.h>
+static inline void mi_atomic_yield(void) {
+  smt_pause();
+}
+#elif defined(__wasi__)
+#include <sched.h>
+static inline void mi_atomic_yield(void) {
+  sched_yield();
+}
+#else
+#include <unistd.h>
+static inline void mi_atomic_yield(void) {
+  sleep(0);
+}
+#endif
+
+
+#endif // __MIMALLOC_ATOMIC_H

preload-mimalloc/mimalloc/include/mimalloc/internal.h

@@ -0,0 +1,953 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_INTERNAL_H
+#define MIMALLOC_INTERNAL_H
+
+
+// --------------------------------------------------------------------------
+// This file contains the interal API's of mimalloc and various utility
+// functions and macros.
+// --------------------------------------------------------------------------
+
+#include "mimalloc/types.h"
+#include "mimalloc/track.h"
+
+#if (MI_DEBUG>0)
+#define mi_trace_message(...)  _mi_trace_message(__VA_ARGS__)
+#else
+#define mi_trace_message(...)
+#endif
+
+#define MI_CACHE_LINE          64
+#if defined(_MSC_VER)
+#pragma warning(disable:4127)   // suppress constant conditional warning (due to MI_SECURE paths)
+#pragma warning(disable:26812)  // unscoped enum warning
+#define mi_decl_noinline        __declspec(noinline)
+#define mi_decl_thread          __declspec(thread)
+#define mi_decl_cache_align     __declspec(align(MI_CACHE_LINE))
+#elif (defined(__GNUC__) && (__GNUC__ >= 3)) || defined(__clang__) // includes clang and icc
+#define mi_decl_noinline        __attribute__((noinline))
+#define mi_decl_thread          __thread
+#define mi_decl_cache_align     __attribute__((aligned(MI_CACHE_LINE)))
+#else
+#define mi_decl_noinline
+#define mi_decl_thread          __thread        // hope for the best :-)
+#define mi_decl_cache_align
+#endif
+
+#if defined(__EMSCRIPTEN__) && !defined(__wasi__)
+#define __wasi__
+#endif
+
+#if defined(__cplusplus)
+#define mi_decl_externc       extern "C"
+#else
+#define mi_decl_externc
+#endif
+
+// pthreads
+#if !defined(_WIN32) && !defined(__wasi__)
+#define  MI_USE_PTHREADS
+#include <pthread.h>
+#endif
+
+// "options.c"
+void       _mi_fputs(mi_output_fun* out, void* arg, const char* prefix, const char* message);
+void       _mi_fprintf(mi_output_fun* out, void* arg, const char* fmt, ...);
+void       _mi_warning_message(const char* fmt, ...);
+void       _mi_verbose_message(const char* fmt, ...);
+void       _mi_trace_message(const char* fmt, ...);
+void       _mi_options_init(void);
+void       _mi_error_message(int err, const char* fmt, ...);
+
+// random.c
+void       _mi_random_init(mi_random_ctx_t* ctx);
+void       _mi_random_init_weak(mi_random_ctx_t* ctx);
+void       _mi_random_reinit_if_weak(mi_random_ctx_t * ctx);
+void       _mi_random_split(mi_random_ctx_t* ctx, mi_random_ctx_t* new_ctx);
+uintptr_t  _mi_random_next(mi_random_ctx_t* ctx);
+uintptr_t  _mi_heap_random_next(mi_heap_t* heap);
+uintptr_t  _mi_os_random_weak(uintptr_t extra_seed);
+static inline uintptr_t _mi_random_shuffle(uintptr_t x);
+
+// init.c
+extern mi_decl_cache_align mi_stats_t       _mi_stats_main;
+extern mi_decl_cache_align const mi_page_t  _mi_page_empty;
+bool       _mi_is_main_thread(void);
+size_t     _mi_current_thread_count(void);
+bool       _mi_preloading(void);        // true while the C runtime is not ready
+mi_threadid_t _mi_thread_id(void) mi_attr_noexcept;
+mi_heap_t* _mi_heap_main_get(void);     // statically allocated main backing heap
+void       _mi_thread_done(mi_heap_t* heap);
+
+// os.c
+void       _mi_os_init(void);                                      // called from process init
+void*      _mi_os_alloc(size_t size, mi_stats_t* stats);           // to allocate thread local data
+void       _mi_os_free(void* p, size_t size, mi_stats_t* stats);   // to free thread local data
+size_t     _mi_os_page_size(void);
+size_t     _mi_os_good_alloc_size(size_t size);
+bool       _mi_os_has_overcommit(void);
+
+bool       _mi_os_reset(void* addr, size_t size, mi_stats_t* tld_stats);
+bool       _mi_os_commit(void* p, size_t size, bool* is_zero, mi_stats_t* stats);
+bool       _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats);
+bool       _mi_os_protect(void* addr, size_t size);
+bool       _mi_os_unprotect(void* addr, size_t size);
+
+void*      _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* large, mi_stats_t* stats);
+void*      _mi_os_alloc_aligned_offset(size_t size, size_t alignment, size_t align_offset, bool commit, bool* large, mi_stats_t* tld_stats);
+void       _mi_os_free_aligned(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats);
+void*      _mi_os_get_aligned_hint(size_t try_alignment, size_t size);
+bool       _mi_os_use_large_page(size_t size, size_t alignment);
+size_t     _mi_os_large_page_size(void);
+
+void       _mi_os_free_ex(void* p, size_t size, bool was_committed, mi_stats_t* stats);
+void*      _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_secs, size_t* pages_reserved, size_t* psize);
+void       _mi_os_free_huge_pages(void* p, size_t size, mi_stats_t* stats);
+
+// arena.c
+mi_arena_id_t _mi_arena_id_none(void);
+void       _mi_arena_free(void* p, size_t size, size_t alignment, size_t align_offset, size_t memid, bool all_committed, mi_stats_t* stats);
+void*      _mi_arena_alloc(size_t size, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
+void*      _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
+bool       _mi_arena_memid_is_suitable(size_t arena_memid, mi_arena_id_t request_arena_id);
+bool       _mi_arena_is_os_allocated(size_t arena_memid);
+
+// "segment-cache.c"
+void*      _mi_segment_cache_pop(size_t size, mi_commit_mask_t* commit_mask, mi_commit_mask_t* decommit_mask, bool large_allowed, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld);
+bool       _mi_segment_cache_push(void* start, size_t size, size_t memid, const mi_commit_mask_t* commit_mask, const mi_commit_mask_t* decommit_mask, bool is_large, bool is_pinned, mi_os_tld_t* tld);
+void       _mi_segment_cache_collect(bool force, mi_os_tld_t* tld);
+void       _mi_segment_cache_free_all(mi_os_tld_t* tld);
+void       _mi_segment_map_allocated_at(const mi_segment_t* segment);
+void       _mi_segment_map_freed_at(const mi_segment_t* segment);
+
+// "segment.c"
+mi_page_t* _mi_segment_page_alloc(mi_heap_t* heap, size_t block_size, size_t page_alignment, mi_segments_tld_t* tld, mi_os_tld_t* os_tld);
+void       _mi_segment_page_free(mi_page_t* page, bool force, mi_segments_tld_t* tld);
+void       _mi_segment_page_abandon(mi_page_t* page, mi_segments_tld_t* tld);
+bool       _mi_segment_try_reclaim_abandoned( mi_heap_t* heap, bool try_all, mi_segments_tld_t* tld);
+void       _mi_segment_thread_collect(mi_segments_tld_t* tld);
+
+#if MI_HUGE_PAGE_ABANDON
+void       _mi_segment_huge_page_free(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
+#else
+void       _mi_segment_huge_page_reset(mi_segment_t* segment, mi_page_t* page, mi_block_t* block);
+#endif
+
+uint8_t*   _mi_segment_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size); // page start for any page
+void       _mi_abandoned_reclaim_all(mi_heap_t* heap, mi_segments_tld_t* tld);
+void       _mi_abandoned_await_readers(void);
+void       _mi_abandoned_collect(mi_heap_t* heap, bool force, mi_segments_tld_t* tld);
+
+// "page.c"
+void*      _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment)  mi_attr_noexcept mi_attr_malloc;
+
+void       _mi_page_retire(mi_page_t* page) mi_attr_noexcept;                  // free the page if there are no other pages with many free blocks
+void       _mi_page_unfull(mi_page_t* page);
+void       _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force);   // free the page
+void       _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq);            // abandon the page, to be picked up by another thread...
+void       _mi_heap_delayed_free_all(mi_heap_t* heap);
+bool       _mi_heap_delayed_free_partial(mi_heap_t* heap);
+void       _mi_heap_collect_retired(mi_heap_t* heap, bool force);
+
+void       _mi_page_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never);
+bool       _mi_page_try_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never);
+size_t     _mi_page_queue_append(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_queue_t* append);
+void       _mi_deferred_free(mi_heap_t* heap, bool force);
+
+void       _mi_page_free_collect(mi_page_t* page,bool force);
+void       _mi_page_reclaim(mi_heap_t* heap, mi_page_t* page);   // callback from segments
+
+size_t     _mi_bin_size(uint8_t bin);           // for stats
+uint8_t    _mi_bin(size_t size);                // for stats
+
+// "heap.c"
+void       _mi_heap_destroy_pages(mi_heap_t* heap);
+void       _mi_heap_collect_abandon(mi_heap_t* heap);
+void       _mi_heap_set_default_direct(mi_heap_t* heap);
+bool       _mi_heap_memid_is_suitable(mi_heap_t* heap, size_t memid);
+void       _mi_heap_destroy_all(void);
+
+// "stats.c"
+void       _mi_stats_done(mi_stats_t* stats);
+mi_msecs_t  _mi_clock_now(void);
+mi_msecs_t  _mi_clock_end(mi_msecs_t start);
+mi_msecs_t  _mi_clock_start(void);
+
+// "alloc.c"
+void*       _mi_page_malloc(mi_heap_t* heap, mi_page_t* page, size_t size, bool zero) mi_attr_noexcept;  // called from `_mi_malloc_generic`
+void*       _mi_heap_malloc_zero(mi_heap_t* heap, size_t size, bool zero) mi_attr_noexcept;
+void*       _mi_heap_malloc_zero_ex(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept;     // called from `_mi_heap_malloc_aligned`
+void*       _mi_heap_realloc_zero(mi_heap_t* heap, void* p, size_t newsize, bool zero) mi_attr_noexcept;
+mi_block_t* _mi_page_ptr_unalign(const mi_segment_t* segment, const mi_page_t* page, const void* p);
+bool        _mi_free_delayed_block(mi_block_t* block);
+void        _mi_free_generic(const mi_segment_t* segment, mi_page_t* page, bool is_local, void* p) mi_attr_noexcept;  // for runtime integration
+void        _mi_padding_shrink(const mi_page_t* page, const mi_block_t* block, const size_t min_size);
+
+// option.c, c primitives
+char        _mi_toupper(char c);
+int         _mi_strnicmp(const char* s, const char* t, size_t n);
+void        _mi_strlcpy(char* dest, const char* src, size_t dest_size);
+void        _mi_strlcat(char* dest, const char* src, size_t dest_size);
+size_t      _mi_strlen(const char* s);
+size_t      _mi_strnlen(const char* s, size_t max_len);
+
+
+#if MI_DEBUG>1
+bool        _mi_page_is_valid(mi_page_t* page);
+#endif
+
+
+// ------------------------------------------------------
+// Branches
+// ------------------------------------------------------
+
+#if defined(__GNUC__) || defined(__clang__)
+#define mi_unlikely(x)     (__builtin_expect(!!(x),false))
+#define mi_likely(x)       (__builtin_expect(!!(x),true))
+#elif (defined(__cplusplus) && (__cplusplus >= 202002L)) || (defined(_MSVC_LANG) && _MSVC_LANG >= 202002L)
+#define mi_unlikely(x)     (x) [[unlikely]]
+#define mi_likely(x)       (x) [[likely]]
+#else
+#define mi_unlikely(x)     (x)
+#define mi_likely(x)       (x)
+#endif
+
+#ifndef __has_builtin
+#define __has_builtin(x)  0
+#endif
+
+
+/* -----------------------------------------------------------
+  Error codes passed to `_mi_fatal_error`
+  All are recoverable but EFAULT is a serious error and aborts by default in secure mode.
+  For portability define undefined error codes using common Unix codes:
+  <https://www-numi.fnal.gov/offline_software/srt_public_context/WebDocs/Errors/unix_system_errors.html>
+----------------------------------------------------------- */
+#include <errno.h>
+#ifndef EAGAIN         // double free
+#define EAGAIN (11)
+#endif
+#ifndef ENOMEM         // out of memory
+#define ENOMEM (12)
+#endif
+#ifndef EFAULT         // corrupted free-list or meta-data
+#define EFAULT (14)
+#endif
+#ifndef EINVAL         // trying to free an invalid pointer
+#define EINVAL (22)
+#endif
+#ifndef EOVERFLOW      // count*size overflow
+#define EOVERFLOW (75)
+#endif
+
+
+/* -----------------------------------------------------------
+  Inlined definitions
+----------------------------------------------------------- */
+#define MI_UNUSED(x)     (void)(x)
+#if (MI_DEBUG>0)
+#define MI_UNUSED_RELEASE(x)
+#else
+#define MI_UNUSED_RELEASE(x)  MI_UNUSED(x)
+#endif
+
+#define MI_INIT4(x)   x(),x(),x(),x()
+#define MI_INIT8(x)   MI_INIT4(x),MI_INIT4(x)
+#define MI_INIT16(x)  MI_INIT8(x),MI_INIT8(x)
+#define MI_INIT32(x)  MI_INIT16(x),MI_INIT16(x)
+#define MI_INIT64(x)  MI_INIT32(x),MI_INIT32(x)
+#define MI_INIT128(x) MI_INIT64(x),MI_INIT64(x)
+#define MI_INIT256(x) MI_INIT128(x),MI_INIT128(x)
+
+
+// Is `x` a power of two? (0 is considered a power of two)
+static inline bool _mi_is_power_of_two(uintptr_t x) {
+  return ((x & (x - 1)) == 0);
+}
+
+// Is a pointer aligned?
+static inline bool _mi_is_aligned(void* p, size_t alignment) {
+  mi_assert_internal(alignment != 0);
+  return (((uintptr_t)p % alignment) == 0);
+}
+
+// Align upwards
+static inline uintptr_t _mi_align_up(uintptr_t sz, size_t alignment) {
+  mi_assert_internal(alignment != 0);
+  uintptr_t mask = alignment - 1;
+  if ((alignment & mask) == 0) {  // power of two?
+    return ((sz + mask) & ~mask);
+  }
+  else {
+    return (((sz + mask)/alignment)*alignment);
+  }
+}
+
+// Align downwards
+static inline uintptr_t _mi_align_down(uintptr_t sz, size_t alignment) {
+  mi_assert_internal(alignment != 0);
+  uintptr_t mask = alignment - 1;
+  if ((alignment & mask) == 0) { // power of two?
+    return (sz & ~mask);
+  }
+  else {
+    return ((sz / alignment) * alignment);
+  }
+}
+
+// Divide upwards: `s <= _mi_divide_up(s,d)*d < s+d`.
+static inline uintptr_t _mi_divide_up(uintptr_t size, size_t divider) {
+  mi_assert_internal(divider != 0);
+  return (divider == 0 ? size : ((size + divider - 1) / divider));
+}
+
+// Is memory zero initialized?
+static inline bool mi_mem_is_zero(void* p, size_t size) {
+  for (size_t i = 0; i < size; i++) {
+    if (((uint8_t*)p)[i] != 0) return false;
+  }
+  return true;
+}
+
+
+// Align a byte size to a size in _machine words_,
+// i.e. byte size == `wsize*sizeof(void*)`.
+static inline size_t _mi_wsize_from_size(size_t size) {
+  mi_assert_internal(size <= SIZE_MAX - sizeof(uintptr_t));
+  return (size + sizeof(uintptr_t) - 1) / sizeof(uintptr_t);
+}
+
+// Overflow detecting multiply
+#if __has_builtin(__builtin_umul_overflow) || (defined(__GNUC__) && (__GNUC__ >= 5))
+#include <limits.h>      // UINT_MAX, ULONG_MAX
+#if defined(_CLOCK_T)    // for Illumos
+#undef _CLOCK_T
+#endif
+static inline bool mi_mul_overflow(size_t count, size_t size, size_t* total) {
+  #if (SIZE_MAX == ULONG_MAX)
+    return __builtin_umull_overflow(count, size, (unsigned long *)total);
+  #elif (SIZE_MAX == UINT_MAX)
+    return __builtin_umul_overflow(count, size, (unsigned int *)total);
+  #else
+    return __builtin_umulll_overflow(count, size, (unsigned long long *)total);
+  #endif
+}
+#else /* __builtin_umul_overflow is unavailable */
+static inline bool mi_mul_overflow(size_t count, size_t size, size_t* total) {
+  #define MI_MUL_NO_OVERFLOW ((size_t)1 << (4*sizeof(size_t)))  // sqrt(SIZE_MAX)
+  *total = count * size;
+  // note: gcc/clang optimize this to directly check the overflow flag
+  return ((size >= MI_MUL_NO_OVERFLOW || count >= MI_MUL_NO_OVERFLOW) && size > 0 && (SIZE_MAX / size) < count);
+}
+#endif
+
+// Safe multiply `count*size` into `total`; return `true` on overflow.
+static inline bool mi_count_size_overflow(size_t count, size_t size, size_t* total) {
+  if (count==1) {  // quick check for the case where count is one (common for C++ allocators)
+    *total = size;
+    return false;
+  }
+  else if mi_unlikely(mi_mul_overflow(count, size, total)) {
+    #if MI_DEBUG > 0
+    _mi_error_message(EOVERFLOW, "allocation request is too large (%zu * %zu bytes)\n", count, size);
+    #endif
+    *total = SIZE_MAX;
+    return true;
+  }
+  else return false;
+}
+
+
+/*----------------------------------------------------------------------------------------
+  Heap functions
+------------------------------------------------------------------------------------------- */
+
+extern const mi_heap_t _mi_heap_empty;  // read-only empty heap, initial value of the thread local default heap
+
+static inline bool mi_heap_is_backing(const mi_heap_t* heap) {
+  return (heap->tld->heap_backing == heap);
+}
+
+static inline bool mi_heap_is_initialized(mi_heap_t* heap) {
+  mi_assert_internal(heap != NULL);
+  return (heap != &_mi_heap_empty);
+}
+
+static inline uintptr_t _mi_ptr_cookie(const void* p) {
+  extern mi_heap_t _mi_heap_main;
+  mi_assert_internal(_mi_heap_main.cookie != 0);
+  return ((uintptr_t)p ^ _mi_heap_main.cookie);
+}
+
+/* -----------------------------------------------------------
+  Pages
+----------------------------------------------------------- */
+
+static inline mi_page_t* _mi_heap_get_free_small_page(mi_heap_t* heap, size_t size) {
+  mi_assert_internal(size <= (MI_SMALL_SIZE_MAX + MI_PADDING_SIZE));
+  const size_t idx = _mi_wsize_from_size(size);
+  mi_assert_internal(idx < MI_PAGES_DIRECT);
+  return heap->pages_free_direct[idx];
+}
+
+// Segment that contains the pointer
+// Large aligned blocks may be aligned at N*MI_SEGMENT_SIZE (inside a huge segment > MI_SEGMENT_SIZE),
+// and we need align "down" to the segment info which is `MI_SEGMENT_SIZE` bytes before it;
+// therefore we align one byte before `p`.
+static inline mi_segment_t* _mi_ptr_segment(const void* p) {
+  mi_assert_internal(p != NULL);
+  return (mi_segment_t*)(((uintptr_t)p - 1) & ~MI_SEGMENT_MASK);
+}
+
+static inline mi_page_t* mi_slice_to_page(mi_slice_t* s) {
+  mi_assert_internal(s->slice_offset== 0 && s->slice_count > 0);
+  return (mi_page_t*)(s);
+}
+
+static inline mi_slice_t* mi_page_to_slice(mi_page_t* p) {
+  mi_assert_internal(p->slice_offset== 0 && p->slice_count > 0);
+  return (mi_slice_t*)(p);
+}
+
+// Segment belonging to a page
+static inline mi_segment_t* _mi_page_segment(const mi_page_t* page) {
+  mi_segment_t* segment = _mi_ptr_segment(page); 
+  mi_assert_internal(segment == NULL || ((mi_slice_t*)page >= segment->slices && (mi_slice_t*)page < segment->slices + segment->slice_entries));
+  return segment;
+}
+
+static inline mi_slice_t* mi_slice_first(const mi_slice_t* slice) {
+  mi_slice_t* start = (mi_slice_t*)((uint8_t*)slice - slice->slice_offset);
+  mi_assert_internal(start >= _mi_ptr_segment(slice)->slices);
+  mi_assert_internal(start->slice_offset == 0);
+  mi_assert_internal(start + start->slice_count > slice);
+  return start;
+}
+
+// Get the page containing the pointer (performance critical as it is called in mi_free)
+static inline mi_page_t* _mi_segment_page_of(const mi_segment_t* segment, const void* p) {
+  mi_assert_internal(p > (void*)segment);
+  ptrdiff_t diff = (uint8_t*)p - (uint8_t*)segment;
+  mi_assert_internal(diff > 0 && diff <= (ptrdiff_t)MI_SEGMENT_SIZE);
+  size_t idx = (size_t)diff >> MI_SEGMENT_SLICE_SHIFT;
+  mi_assert_internal(idx <= segment->slice_entries);
+  mi_slice_t* slice0 = (mi_slice_t*)&segment->slices[idx];
+  mi_slice_t* slice = mi_slice_first(slice0);  // adjust to the block that holds the page data
+  mi_assert_internal(slice->slice_offset == 0);
+  mi_assert_internal(slice >= segment->slices && slice < segment->slices + segment->slice_entries);
+  return mi_slice_to_page(slice);
+}
+
+// Quick page start for initialized pages
+static inline uint8_t* _mi_page_start(const mi_segment_t* segment, const mi_page_t* page, size_t* page_size) {
+  return _mi_segment_page_start(segment, page, page_size);
+}
+
+// Get the page containing the pointer
+static inline mi_page_t* _mi_ptr_page(void* p) {
+  return _mi_segment_page_of(_mi_ptr_segment(p), p);
+}
+
+// Get the block size of a page (special case for huge objects)
+static inline size_t mi_page_block_size(const mi_page_t* page) {
+  const size_t bsize = page->xblock_size;
+  mi_assert_internal(bsize > 0);
+  if mi_likely(bsize < MI_HUGE_BLOCK_SIZE) {
+    return bsize;
+  }
+  else {
+    size_t psize;
+    _mi_segment_page_start(_mi_page_segment(page), page, &psize);
+    return psize;
+  }
+}
+
+static inline bool mi_page_is_huge(const mi_page_t* page) {
+  return (_mi_page_segment(page)->kind == MI_SEGMENT_HUGE);
+}
+
+// Get the usable block size of a page without fixed padding.
+// This may still include internal padding due to alignment and rounding up size classes.
+static inline size_t mi_page_usable_block_size(const mi_page_t* page) {
+  return mi_page_block_size(page) - MI_PADDING_SIZE;
+}
+
+// size of a segment
+static inline size_t mi_segment_size(mi_segment_t* segment) {
+  return segment->segment_slices * MI_SEGMENT_SLICE_SIZE;
+}
+
+static inline uint8_t* mi_segment_end(mi_segment_t* segment) {
+  return (uint8_t*)segment + mi_segment_size(segment);
+}
+
+// Thread free access
+static inline mi_block_t* mi_page_thread_free(const mi_page_t* page) {
+  return (mi_block_t*)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xthread_free) & ~3);
+}
+
+static inline mi_delayed_t mi_page_thread_free_flag(const mi_page_t* page) {
+  return (mi_delayed_t)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xthread_free) & 3);
+}
+
+// Heap access
+static inline mi_heap_t* mi_page_heap(const mi_page_t* page) {
+  return (mi_heap_t*)(mi_atomic_load_relaxed(&((mi_page_t*)page)->xheap));
+}
+
+static inline void mi_page_set_heap(mi_page_t* page, mi_heap_t* heap) {
+  mi_assert_internal(mi_page_thread_free_flag(page) != MI_DELAYED_FREEING);
+  mi_atomic_store_release(&page->xheap,(uintptr_t)heap);
+}
+
+// Thread free flag helpers
+static inline mi_block_t* mi_tf_block(mi_thread_free_t tf) {
+  return (mi_block_t*)(tf & ~0x03);
+}
+static inline mi_delayed_t mi_tf_delayed(mi_thread_free_t tf) {
+  return (mi_delayed_t)(tf & 0x03);
+}
+static inline mi_thread_free_t mi_tf_make(mi_block_t* block, mi_delayed_t delayed) {
+  return (mi_thread_free_t)((uintptr_t)block | (uintptr_t)delayed);
+}
+static inline mi_thread_free_t mi_tf_set_delayed(mi_thread_free_t tf, mi_delayed_t delayed) {
+  return mi_tf_make(mi_tf_block(tf),delayed);
+}
+static inline mi_thread_free_t mi_tf_set_block(mi_thread_free_t tf, mi_block_t* block) {
+  return mi_tf_make(block, mi_tf_delayed(tf));
+}
+
+// are all blocks in a page freed?
+// note: needs up-to-date used count, (as the `xthread_free` list may not be empty). see `_mi_page_collect_free`.
+static inline bool mi_page_all_free(const mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  return (page->used == 0);
+}
+
+// are there any available blocks?
+static inline bool mi_page_has_any_available(const mi_page_t* page) {
+  mi_assert_internal(page != NULL && page->reserved > 0);
+  return (page->used < page->reserved || (mi_page_thread_free(page) != NULL));
+}
+
+// are there immediately available blocks, i.e. blocks available on the free list.
+static inline bool mi_page_immediate_available(const mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  return (page->free != NULL);
+}
+
+// is more than 7/8th of a page in use?
+static inline bool mi_page_mostly_used(const mi_page_t* page) {
+  if (page==NULL) return true;
+  uint16_t frac = page->reserved / 8U;
+  return (page->reserved - page->used <= frac);
+}
+
+static inline mi_page_queue_t* mi_page_queue(const mi_heap_t* heap, size_t size) {
+  return &((mi_heap_t*)heap)->pages[_mi_bin(size)];
+}
+
+
+
+//-----------------------------------------------------------
+// Page flags
+//-----------------------------------------------------------
+static inline bool mi_page_is_in_full(const mi_page_t* page) {
+  return page->flags.x.in_full;
+}
+
+static inline void mi_page_set_in_full(mi_page_t* page, bool in_full) {
+  page->flags.x.in_full = in_full;
+}
+
+static inline bool mi_page_has_aligned(const mi_page_t* page) {
+  return page->flags.x.has_aligned;
+}
+
+static inline void mi_page_set_has_aligned(mi_page_t* page, bool has_aligned) {
+  page->flags.x.has_aligned = has_aligned;
+}
+
+
+/* -------------------------------------------------------------------
+Encoding/Decoding the free list next pointers
+
+This is to protect against buffer overflow exploits where the
+free list is mutated. Many hardened allocators xor the next pointer `p`
+with a secret key `k1`, as `p^k1`. This prevents overwriting with known
+values but might be still too weak: if the attacker can guess
+the pointer `p` this  can reveal `k1` (since `p^k1^p == k1`).
+Moreover, if multiple blocks can be read as well, the attacker can
+xor both as `(p1^k1) ^ (p2^k1) == p1^p2` which may reveal a lot
+about the pointers (and subsequently `k1`).
+
+Instead mimalloc uses an extra key `k2` and encodes as `((p^k2)<<<k1)+k1`.
+Since these operations are not associative, the above approaches do not
+work so well any more even if the `p` can be guesstimated. For example,
+for the read case we can subtract two entries to discard the `+k1` term,
+but that leads to `((p1^k2)<<<k1) - ((p2^k2)<<<k1)` at best.
+We include the left-rotation since xor and addition are otherwise linear
+in the lowest bit. Finally, both keys are unique per page which reduces
+the re-use of keys by a large factor.
+
+We also pass a separate `null` value to be used as `NULL` or otherwise
+`(k2<<<k1)+k1` would appear (too) often as a sentinel value.
+------------------------------------------------------------------- */
+
+static inline bool mi_is_in_same_segment(const void* p, const void* q) {
+  return (_mi_ptr_segment(p) == _mi_ptr_segment(q));
+}
+
+static inline bool mi_is_in_same_page(const void* p, const void* q) {
+  mi_segment_t* segment = _mi_ptr_segment(p);
+  if (_mi_ptr_segment(q) != segment) return false;
+  // assume q may be invalid // return (_mi_segment_page_of(segment, p) == _mi_segment_page_of(segment, q));
+  mi_page_t* page = _mi_segment_page_of(segment, p);
+  size_t psize;
+  uint8_t* start = _mi_segment_page_start(segment, page, &psize);
+  return (start <= (uint8_t*)q && (uint8_t*)q < start + psize);
+}
+
+static inline uintptr_t mi_rotl(uintptr_t x, uintptr_t shift) {
+  shift %= MI_INTPTR_BITS;
+  return (shift==0 ? x : ((x << shift) | (x >> (MI_INTPTR_BITS - shift))));
+}
+static inline uintptr_t mi_rotr(uintptr_t x, uintptr_t shift) {
+  shift %= MI_INTPTR_BITS;
+  return (shift==0 ? x : ((x >> shift) | (x << (MI_INTPTR_BITS - shift))));
+}
+
+static inline void* mi_ptr_decode(const void* null, const mi_encoded_t x, const uintptr_t* keys) {
+  void* p = (void*)(mi_rotr(x - keys[0], keys[0]) ^ keys[1]);
+  return (p==null ? NULL : p);
+}
+
+static inline mi_encoded_t mi_ptr_encode(const void* null, const void* p, const uintptr_t* keys) {
+  uintptr_t x = (uintptr_t)(p==NULL ? null : p);
+  return mi_rotl(x ^ keys[1], keys[0]) + keys[0];
+}
+
+static inline mi_block_t* mi_block_nextx( const void* null, const mi_block_t* block, const uintptr_t* keys ) {
+  mi_track_mem_defined(block,sizeof(mi_block_t));
+  mi_block_t* next;
+  #ifdef MI_ENCODE_FREELIST
+  next = (mi_block_t*)mi_ptr_decode(null, block->next, keys);
+  #else
+  MI_UNUSED(keys); MI_UNUSED(null);
+  next = (mi_block_t*)block->next;
+  #endif
+  mi_track_mem_noaccess(block,sizeof(mi_block_t));
+  return next;
+}
+
+static inline void mi_block_set_nextx(const void* null, mi_block_t* block, const mi_block_t* next, const uintptr_t* keys) {
+  mi_track_mem_undefined(block,sizeof(mi_block_t));
+  #ifdef MI_ENCODE_FREELIST
+  block->next = mi_ptr_encode(null, next, keys);
+  #else
+  MI_UNUSED(keys); MI_UNUSED(null);
+  block->next = (mi_encoded_t)next;
+  #endif
+  mi_track_mem_noaccess(block,sizeof(mi_block_t));
+}
+
+static inline mi_block_t* mi_block_next(const mi_page_t* page, const mi_block_t* block) {
+  #ifdef MI_ENCODE_FREELIST
+  mi_block_t* next = mi_block_nextx(page,block,page->keys);
+  // check for free list corruption: is `next` at least in the same page?
+  // TODO: check if `next` is `page->block_size` aligned?
+  if mi_unlikely(next!=NULL && !mi_is_in_same_page(block, next)) {
+    _mi_error_message(EFAULT, "corrupted free list entry of size %zub at %p: value 0x%zx\n", mi_page_block_size(page), block, (uintptr_t)next);
+    next = NULL;
+  }
+  return next;
+  #else
+  MI_UNUSED(page);
+  return mi_block_nextx(page,block,NULL);
+  #endif
+}
+
+static inline void mi_block_set_next(const mi_page_t* page, mi_block_t* block, const mi_block_t* next) {
+  #ifdef MI_ENCODE_FREELIST
+  mi_block_set_nextx(page,block,next, page->keys);
+  #else
+  MI_UNUSED(page);
+  mi_block_set_nextx(page,block,next,NULL);
+  #endif
+}
+
+
+// -------------------------------------------------------------------
+// commit mask
+// -------------------------------------------------------------------
+
+static inline void mi_commit_mask_create_empty(mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    cm->mask[i] = 0;
+  }
+}
+
+static inline void mi_commit_mask_create_full(mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    cm->mask[i] = ~((size_t)0);
+  }
+}
+
+static inline bool mi_commit_mask_is_empty(const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    if (cm->mask[i] != 0) return false;
+  }
+  return true;
+}
+
+static inline bool mi_commit_mask_is_full(const mi_commit_mask_t* cm) {
+  for (size_t i = 0; i < MI_COMMIT_MASK_FIELD_COUNT; i++) {
+    if (cm->mask[i] != ~((size_t)0)) return false;
+  }
+  return true;
+}
+
+// defined in `segment.c`:
+size_t _mi_commit_mask_committed_size(const mi_commit_mask_t* cm, size_t total);
+size_t _mi_commit_mask_next_run(const mi_commit_mask_t* cm, size_t* idx);
+
+#define mi_commit_mask_foreach(cm,idx,count) \
+  idx = 0; \
+  while ((count = _mi_commit_mask_next_run(cm,&idx)) > 0) { 
+        
+#define mi_commit_mask_foreach_end() \
+    idx += count; \
+  }
+      
+
+
+
+// -------------------------------------------------------------------
+// Fast "random" shuffle
+// -------------------------------------------------------------------
+
+static inline uintptr_t _mi_random_shuffle(uintptr_t x) {
+  if (x==0) { x = 17; }   // ensure we don't get stuck in generating zeros
+#if (MI_INTPTR_SIZE==8)
+  // by Sebastiano Vigna, see: <http://xoshiro.di.unimi.it/splitmix64.c>
+  x ^= x >> 30;
+  x *= 0xbf58476d1ce4e5b9UL;
+  x ^= x >> 27;
+  x *= 0x94d049bb133111ebUL;
+  x ^= x >> 31;
+#elif (MI_INTPTR_SIZE==4)
+  // by Chris Wellons, see: <https://nullprogram.com/blog/2018/07/31/>
+  x ^= x >> 16;
+  x *= 0x7feb352dUL;
+  x ^= x >> 15;
+  x *= 0x846ca68bUL;
+  x ^= x >> 16;
+#endif
+  return x;
+}
+
+// -------------------------------------------------------------------
+// Optimize numa node access for the common case (= one node)
+// -------------------------------------------------------------------
+
+int    _mi_os_numa_node_get(mi_os_tld_t* tld);
+size_t _mi_os_numa_node_count_get(void);
+
+extern _Atomic(size_t) _mi_numa_node_count;
+static inline int _mi_os_numa_node(mi_os_tld_t* tld) {
+  if mi_likely(mi_atomic_load_relaxed(&_mi_numa_node_count) == 1) { return 0; }
+  else return _mi_os_numa_node_get(tld);
+}
+static inline size_t _mi_os_numa_node_count(void) {
+  const size_t count = mi_atomic_load_relaxed(&_mi_numa_node_count);
+  if mi_likely(count > 0) { return count; }
+  else return _mi_os_numa_node_count_get();
+}
+
+
+
+// -----------------------------------------------------------------------
+// Count bits: trailing or leading zeros (with MI_INTPTR_BITS on all zero)
+// -----------------------------------------------------------------------
+
+#if defined(__GNUC__)
+
+#include <limits.h>       // LONG_MAX
+#define MI_HAVE_FAST_BITSCAN
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (INTPTR_MAX == LONG_MAX)
+  return __builtin_clzl(x);
+#else
+  return __builtin_clzll(x);
+#endif
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (INTPTR_MAX == LONG_MAX)
+  return __builtin_ctzl(x);
+#else
+  return __builtin_ctzll(x);
+#endif
+}
+
+#elif defined(_MSC_VER)
+
+#include <limits.h>       // LONG_MAX
+#include <intrin.h>       // BitScanReverse64
+#define MI_HAVE_FAST_BITSCAN
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+  unsigned long idx;
+#if (INTPTR_MAX == LONG_MAX)
+  _BitScanReverse(&idx, x);
+#else
+  _BitScanReverse64(&idx, x);
+#endif
+  return ((MI_INTPTR_BITS - 1) - idx);
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+  unsigned long idx;
+#if (INTPTR_MAX == LONG_MAX)
+  _BitScanForward(&idx, x);
+#else
+  _BitScanForward64(&idx, x);
+#endif
+  return idx;
+}
+
+#else
+static inline size_t mi_ctz32(uint32_t x) {
+  // de Bruijn multiplication, see <http://supertech.csail.mit.edu/papers/debruijn.pdf>
+  static const unsigned char debruijn[32] = {
+    0, 1, 28, 2, 29, 14, 24, 3, 30, 22, 20, 15, 25, 17, 4, 8,
+    31, 27, 13, 23, 21, 19, 16, 7, 26, 12, 18, 6, 11, 5, 10, 9
+  };
+  if (x==0) return 32;
+  return debruijn[((x & -(int32_t)x) * 0x077CB531UL) >> 27];
+}
+static inline size_t mi_clz32(uint32_t x) {
+  // de Bruijn multiplication, see <http://supertech.csail.mit.edu/papers/debruijn.pdf>
+  static const uint8_t debruijn[32] = {
+    31, 22, 30, 21, 18, 10, 29, 2, 20, 17, 15, 13, 9, 6, 28, 1,
+    23, 19, 11, 3, 16, 14, 7, 24, 12, 4, 8, 25, 5, 26, 27, 0
+  };
+  if (x==0) return 32;
+  x |= x >> 1;
+  x |= x >> 2;
+  x |= x >> 4;
+  x |= x >> 8;
+  x |= x >> 16;
+  return debruijn[(uint32_t)(x * 0x07C4ACDDUL) >> 27];
+}
+
+static inline size_t mi_clz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (MI_INTPTR_BITS <= 32)
+  return mi_clz32((uint32_t)x);
+#else
+  size_t count = mi_clz32((uint32_t)(x >> 32));
+  if (count < 32) return count;
+  return (32 + mi_clz32((uint32_t)x));
+#endif
+}
+static inline size_t mi_ctz(uintptr_t x) {
+  if (x==0) return MI_INTPTR_BITS;
+#if (MI_INTPTR_BITS <= 32)
+  return mi_ctz32((uint32_t)x);
+#else
+  size_t count = mi_ctz32((uint32_t)x);
+  if (count < 32) return count;
+  return (32 + mi_ctz32((uint32_t)(x>>32)));
+#endif
+}
+
+#endif
+
+// "bit scan reverse": Return index of the highest bit (or MI_INTPTR_BITS if `x` is zero)
+static inline size_t mi_bsr(uintptr_t x) {
+  return (x==0 ? MI_INTPTR_BITS : MI_INTPTR_BITS - 1 - mi_clz(x));
+}
+
+
+// ---------------------------------------------------------------------------------
+// Provide our own `_mi_memcpy` for potential performance optimizations.
+//
+// For now, only on Windows with msvc/clang-cl we optimize to `rep movsb` if
+// we happen to run on x86/x64 cpu's that have "fast short rep movsb" (FSRM) support
+// (AMD Zen3+ (~2020) or Intel Ice Lake+ (~2017). See also issue #201 and pr #253.
+// ---------------------------------------------------------------------------------
+
+#if !MI_TRACK_ENABLED && defined(_WIN32) && (defined(_M_IX86) || defined(_M_X64))
+#include <intrin.h>
+#include <string.h>
+extern bool _mi_cpu_has_fsrm;
+static inline void _mi_memcpy(void* dst, const void* src, size_t n) {
+  if (_mi_cpu_has_fsrm) {
+    __movsb((unsigned char*)dst, (const unsigned char*)src, n);
+  }
+  else {
+    memcpy(dst, src, n);
+  }
+}
+static inline void _mi_memzero(void* dst, size_t n) {
+  if (_mi_cpu_has_fsrm) {
+    __stosb((unsigned char*)dst, 0, n);
+  }
+  else {
+    memset(dst, 0, n);
+  }
+}
+#else
+#include <string.h>
+static inline void _mi_memcpy(void* dst, const void* src, size_t n) {
+  memcpy(dst, src, n);
+}
+static inline void _mi_memzero(void* dst, size_t n) {
+  memset(dst, 0, n);
+}
+#endif
+
+
+// -------------------------------------------------------------------------------
+// The `_mi_memcpy_aligned` can be used if the pointers are machine-word aligned
+// This is used for example in `mi_realloc`.
+// -------------------------------------------------------------------------------
+
+#if (defined(__GNUC__) && (__GNUC__ >= 4)) || defined(__clang__)
+// On GCC/CLang we provide a hint that the pointers are word aligned.
+#include <string.h>
+static inline void _mi_memcpy_aligned(void* dst, const void* src, size_t n) {
+  mi_assert_internal(((uintptr_t)dst % MI_INTPTR_SIZE == 0) && ((uintptr_t)src % MI_INTPTR_SIZE == 0));
+  void* adst = __builtin_assume_aligned(dst, MI_INTPTR_SIZE);
+  const void* asrc = __builtin_assume_aligned(src, MI_INTPTR_SIZE);
+  _mi_memcpy(adst, asrc, n);
+}
+
+static inline void _mi_memzero_aligned(void* dst, size_t n) {
+  mi_assert_internal((uintptr_t)dst % MI_INTPTR_SIZE == 0);
+  void* adst = __builtin_assume_aligned(dst, MI_INTPTR_SIZE);
+  _mi_memzero(adst, n);
+}
+#else
+// Default fallback on `_mi_memcpy`
+static inline void _mi_memcpy_aligned(void* dst, const void* src, size_t n) {
+  mi_assert_internal(((uintptr_t)dst % MI_INTPTR_SIZE == 0) && ((uintptr_t)src % MI_INTPTR_SIZE == 0));
+  _mi_memcpy(dst, src, n);
+}
+
+static inline void _mi_memzero_aligned(void* dst, size_t n) {
+  mi_assert_internal((uintptr_t)dst % MI_INTPTR_SIZE == 0);
+  _mi_memzero(dst, n);
+}
+#endif
+
+
+#endif

preload-mimalloc/mimalloc/include/mimalloc/prim.h

@@ -0,0 +1,311 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_PRIM_H
+#define MIMALLOC_PRIM_H
+
+
+// --------------------------------------------------------------------------
+// This file specifies the primitive portability API.
+// Each OS/host needs to implement these primitives, see `src/prim`
+// for implementations on Window, macOS, WASI, and Linux/Unix.
+//
+// note: on all primitive functions, we always get:
+//  addr != NULL and page aligned
+//  size > 0     and page aligned
+//  return value is an error code an int where 0 is success.
+// --------------------------------------------------------------------------
+
+// OS memory configuration
+typedef struct mi_os_mem_config_s {
+  size_t  page_size;          // 4KiB
+  size_t  large_page_size;    // 2MiB
+  size_t  alloc_granularity;  // smallest allocation size (on Windows 64KiB)
+  bool    has_overcommit;     // can we reserve more memory than can be actually committed?
+  bool    must_free_whole;    // must allocated blocks free as a whole (false for mmap, true for VirtualAlloc)
+} mi_os_mem_config_t;
+
+// Initialize
+void _mi_prim_mem_init( mi_os_mem_config_t* config );
+
+// Free OS memory
+int _mi_prim_free(void* addr, size_t size );
+  
+// Allocate OS memory. Return NULL on error.
+// The `try_alignment` is just a hint and the returned pointer does not have to be aligned.
+// pre: !commit => !allow_large
+//      try_alignment >= _mi_os_page_size() and a power of 2
+int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, void** addr);
+
+// Commit memory. Returns error code or 0 on success.
+int _mi_prim_commit(void* addr, size_t size, bool commit);
+
+// Reset memory. The range keeps being accessible but the content might be reset.
+// Returns error code or 0 on success.
+int _mi_prim_reset(void* addr, size_t size);
+
+// Protect memory. Returns error code or 0 on success.
+int _mi_prim_protect(void* addr, size_t size, bool protect);
+
+// Allocate huge (1GiB) pages possibly associated with a NUMA node.
+// pre: size > 0  and a multiple of 1GiB.
+//      addr is either NULL or an address hint.
+//      numa_node is either negative (don't care), or a numa node number.
+int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, void** addr);
+
+// Return the current NUMA node
+size_t _mi_prim_numa_node(void);
+
+// Return the number of logical NUMA nodes
+size_t _mi_prim_numa_node_count(void);
+
+// Clock ticks
+mi_msecs_t _mi_prim_clock_now(void);
+
+// Return process information (only for statistics)
+typedef struct mi_process_info_s {
+  mi_msecs_t  elapsed;
+  mi_msecs_t  utime;
+  mi_msecs_t  stime; 
+  size_t      current_rss; 
+  size_t      peak_rss;  
+  size_t      current_commit;
+  size_t      peak_commit; 
+  size_t      page_faults;
+} mi_process_info_t;
+
+void _mi_prim_process_info(mi_process_info_t* pinfo);
+
+// Default stderr output. (only for warnings etc. with verbose enabled)
+// msg != NULL && _mi_strlen(msg) > 0
+void _mi_prim_out_stderr( const char* msg );
+
+// Get an environment variable. (only for options)
+// name != NULL, result != NULL, result_size >= 64
+bool _mi_prim_getenv(const char* name, char* result, size_t result_size);
+
+
+// Fill a buffer with strong randomness; return `false` on error or if
+// there is no strong randomization available.
+bool _mi_prim_random_buf(void* buf, size_t buf_len);
+
+// Called on the first thread start, and should ensure `_mi_thread_done` is called on thread termination.
+void _mi_prim_thread_init_auto_done(void);
+
+// Called on process exit and may take action to clean up resources associated with the thread auto done.
+void _mi_prim_thread_done_auto_done(void);
+
+// Called when the default heap for a thread changes
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap);
+
+
+//-------------------------------------------------------------------
+// Thread id: `_mi_prim_thread_id()`
+// 
+// Getting the thread id should be performant as it is called in the
+// fast path of `_mi_free` and we specialize for various platforms as
+// inlined definitions. Regular code should call `init.c:_mi_thread_id()`.
+// We only require _mi_prim_thread_id() to return a unique id
+// for each thread (unequal to zero).
+//-------------------------------------------------------------------
+
+// defined in `init.c`; do not use these directly
+extern mi_decl_thread mi_heap_t* _mi_heap_default;  // default heap to allocate from
+extern bool _mi_process_is_initialized;             // has mi_process_init been called?
+
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept;
+
+#if defined(_WIN32)
+
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
+  // Windows: works on Intel and ARM in both 32- and 64-bit
+  return (uintptr_t)NtCurrentTeb();
+}
+
+// We use assembly for a fast thread id on the main platforms. The TLS layout depends on
+// both the OS and libc implementation so we use specific tests for each main platform.
+// If you test on another platform and it works please send a PR :-)
+// see also https://akkadia.org/drepper/tls.pdf for more info on the TLS register.
+#elif defined(__GNUC__) && ( \
+           (defined(__GLIBC__)   && (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__))) \
+        || (defined(__APPLE__)   && (defined(__x86_64__) || defined(__aarch64__))) \
+        || (defined(__BIONIC__)  && (defined(__x86_64__) || defined(__i386__) || defined(__arm__) || defined(__aarch64__))) \
+        || (defined(__FreeBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
+        || (defined(__OpenBSD__) && (defined(__x86_64__) || defined(__i386__) || defined(__aarch64__))) \
+      )
+
+static inline void* mi_prim_tls_slot(size_t slot) mi_attr_noexcept {
+  void* res;
+  const size_t ofs = (slot*sizeof(void*));
+  #if defined(__i386__)
+    __asm__("movl %%gs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x86 32-bit always uses GS
+  #elif defined(__APPLE__) && defined(__x86_64__)
+    __asm__("movq %%gs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x86_64 macOSX uses GS
+  #elif defined(__x86_64__) && (MI_INTPTR_SIZE==4)
+    __asm__("movl %%fs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x32 ABI
+  #elif defined(__x86_64__)
+    __asm__("movq %%fs:%1, %0" : "=r" (res) : "m" (*((void**)ofs)) : );  // x86_64 Linux, BSD uses FS
+  #elif defined(__arm__)
+    void** tcb; MI_UNUSED(ofs);
+    __asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
+    res = tcb[slot];
+  #elif defined(__aarch64__)
+    void** tcb; MI_UNUSED(ofs);
+    #if defined(__APPLE__) // M1, issue #343
+    __asm__ volatile ("mrs %0, tpidrro_el0\nbic %0, %0, #7" : "=r" (tcb));
+    #else
+    __asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
+    #endif
+    res = tcb[slot];
+  #endif
+  return res;
+}
+
+// setting a tls slot is only used on macOS for now
+static inline void mi_prim_tls_slot_set(size_t slot, void* value) mi_attr_noexcept {
+  const size_t ofs = (slot*sizeof(void*));
+  #if defined(__i386__)
+    __asm__("movl %1,%%gs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // 32-bit always uses GS
+  #elif defined(__APPLE__) && defined(__x86_64__)
+    __asm__("movq %1,%%gs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // x86_64 macOS uses GS
+  #elif defined(__x86_64__) && (MI_INTPTR_SIZE==4)
+    __asm__("movl %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // x32 ABI
+  #elif defined(__x86_64__)
+    __asm__("movq %1,%%fs:%0" : "=m" (*((void**)ofs)) : "rn" (value) : );  // x86_64 Linux, BSD uses FS
+  #elif defined(__arm__)
+    void** tcb; MI_UNUSED(ofs);
+    __asm__ volatile ("mrc p15, 0, %0, c13, c0, 3\nbic %0, %0, #3" : "=r" (tcb));
+    tcb[slot] = value;
+  #elif defined(__aarch64__)
+    void** tcb; MI_UNUSED(ofs);
+    #if defined(__APPLE__) // M1, issue #343
+    __asm__ volatile ("mrs %0, tpidrro_el0\nbic %0, %0, #7" : "=r" (tcb));
+    #else
+    __asm__ volatile ("mrs %0, tpidr_el0" : "=r" (tcb));
+    #endif
+    tcb[slot] = value;
+  #endif
+}
+
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
+  #if defined(__BIONIC__)
+    // issue #384, #495: on the Bionic libc (Android), slot 1 is the thread id
+    // see: https://github.com/aosp-mirror/platform_bionic/blob/c44b1d0676ded732df4b3b21c5f798eacae93228/libc/platform/bionic/tls_defines.h#L86
+    return (uintptr_t)mi_prim_tls_slot(1);
+  #else
+    // in all our other targets, slot 0 is the thread id
+    // glibc: https://sourceware.org/git/?p=glibc.git;a=blob_plain;f=sysdeps/x86_64/nptl/tls.h
+    // apple: https://github.com/apple/darwin-xnu/blob/main/libsyscall/os/tsd.h#L36
+    return (uintptr_t)mi_prim_tls_slot(0);
+  #endif
+}
+
+#else
+
+// otherwise use portable C, taking the address of a thread local variable (this is still very fast on most platforms).
+static inline mi_threadid_t _mi_prim_thread_id(void) mi_attr_noexcept {
+  return (uintptr_t)&_mi_heap_default;
+}
+
+#endif
+
+
+
+/* ----------------------------------------------------------------------------------------
+The thread local default heap: `_mi_prim_get_default_heap()`
+This is inlined here as it is on the fast path for allocation functions.
+
+On most platforms (Windows, Linux, FreeBSD, NetBSD, etc), this just returns a
+__thread local variable (`_mi_heap_default`). With the initial-exec TLS model this ensures
+that the storage will always be available (allocated on the thread stacks).
+
+On some platforms though we cannot use that when overriding `malloc` since the underlying
+TLS implementation (or the loader) will call itself `malloc` on a first access and recurse.
+We try to circumvent this in an efficient way:
+- macOSX : we use an unused TLS slot from the OS allocated slots (MI_TLS_SLOT). On OSX, the
+           loader itself calls `malloc` even before the modules are initialized.
+- OpenBSD: we use an unused slot from the pthread block (MI_TLS_PTHREAD_SLOT_OFS).
+- DragonFly: defaults are working but seem slow compared to freeBSD (see PR #323)
+------------------------------------------------------------------------------------------- */
+
+static inline mi_heap_t* mi_prim_get_default_heap(void);
+
+#if defined(MI_MALLOC_OVERRIDE)
+#if defined(__APPLE__) // macOS
+  #define MI_TLS_SLOT               89  // seems unused?
+  // #define MI_TLS_RECURSE_GUARD 1
+  // other possible unused ones are 9, 29, __PTK_FRAMEWORK_JAVASCRIPTCORE_KEY4 (94), __PTK_FRAMEWORK_GC_KEY9 (112) and __PTK_FRAMEWORK_OLDGC_KEY9 (89)
+  // see <https://github.com/rweichler/substrate/blob/master/include/pthread_machdep.h>
+#elif defined(__OpenBSD__)
+  // use end bytes of a name; goes wrong if anyone uses names > 23 characters (ptrhread specifies 16)
+  // see <https://github.com/openbsd/src/blob/master/lib/libc/include/thread_private.h#L371>
+  #define MI_TLS_PTHREAD_SLOT_OFS   (6*sizeof(int) + 4*sizeof(void*) + 24)
+  // #elif defined(__DragonFly__)
+  // #warning "mimalloc is not working correctly on DragonFly yet."
+  // #define MI_TLS_PTHREAD_SLOT_OFS   (4 + 1*sizeof(void*))  // offset `uniqueid` (also used by gdb?) <https://github.com/DragonFlyBSD/DragonFlyBSD/blob/master/lib/libthread_xu/thread/thr_private.h#L458>
+#elif defined(__ANDROID__)
+  // See issue #381
+  #define MI_TLS_PTHREAD
+#endif
+#endif
+
+
+#if defined(MI_TLS_SLOT)
+
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  mi_heap_t* heap = (mi_heap_t*)mi_prim_tls_slot(MI_TLS_SLOT);
+  if mi_unlikely(heap == NULL) {
+    #ifdef __GNUC__
+    __asm(""); // prevent conditional load of the address of _mi_heap_empty
+    #endif
+    heap = (mi_heap_t*)&_mi_heap_empty;
+  }
+  return heap;
+}
+
+#elif defined(MI_TLS_PTHREAD_SLOT_OFS)
+
+static inline mi_heap_t** mi_prim_tls_pthread_heap_slot(void) {
+  pthread_t self = pthread_self();
+  #if defined(__DragonFly__)
+  if (self==NULL) return NULL;
+  #endif
+  return (mi_heap_t**)((uint8_t*)self + MI_TLS_PTHREAD_SLOT_OFS);
+}
+
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  mi_heap_t** pheap = mi_prim_tls_pthread_heap_slot();
+  if mi_unlikely(pheap == NULL) return _mi_heap_main_get();
+  mi_heap_t* heap = *pheap;
+  if mi_unlikely(heap == NULL) return (mi_heap_t*)&_mi_heap_empty;
+  return heap;
+}
+
+#elif defined(MI_TLS_PTHREAD)
+
+extern pthread_key_t _mi_heap_default_key;
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  mi_heap_t* heap = (mi_unlikely(_mi_heap_default_key == (pthread_key_t)(-1)) ? _mi_heap_main_get() : (mi_heap_t*)pthread_getspecific(_mi_heap_default_key));
+  return (mi_unlikely(heap == NULL) ? (mi_heap_t*)&_mi_heap_empty : heap);
+}
+
+#else // default using a thread local variable; used on most platforms.
+
+static inline mi_heap_t* mi_prim_get_default_heap(void) {
+  #if defined(MI_TLS_RECURSE_GUARD)
+  if (mi_unlikely(!_mi_process_is_initialized)) return _mi_heap_main_get();
+  #endif
+  return _mi_heap_default;
+}
+
+#endif  // mi_prim_get_default_heap()
+
+
+
+#endif  // MIMALLOC_PRIM_H

preload-mimalloc/mimalloc/include/mimalloc/track.h

@@ -0,0 +1,147 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_TRACK_H
+#define MIMALLOC_TRACK_H
+
+/* ------------------------------------------------------------------------------------------------------
+Track memory ranges with macros for tools like Valgrind address sanitizer, or other memory checkers.
+These can be defined for tracking allocation:
+
+  #define mi_track_malloc_size(p,reqsize,size,zero)
+  #define mi_track_free_size(p,_size)
+
+The macros are set up such that the size passed to `mi_track_free_size`
+always matches the size of `mi_track_malloc_size`. (currently, `size == mi_usable_size(p)`).
+The `reqsize` is what the user requested, and `size >= reqsize`.
+The `size` is either byte precise (and `size==reqsize`) if `MI_PADDING` is enabled,
+or otherwise it is the usable block size which may be larger than the original request.
+Use `_mi_block_size_of(void* p)` to get the full block size that was allocated (including padding etc).
+The `zero` parameter is `true` if the allocated block is zero initialized.
+
+Optional:
+
+  #define mi_track_align(p,alignedp,offset,size)
+  #define mi_track_resize(p,oldsize,newsize)
+  #define mi_track_init()
+
+The `mi_track_align` is called right after a `mi_track_malloc` for aligned pointers in a block.
+The corresponding `mi_track_free` still uses the block start pointer and original size (corresponding to the `mi_track_malloc`).
+The `mi_track_resize` is currently unused but could be called on reallocations within a block.
+`mi_track_init` is called at program start.
+
+The following macros are for tools like asan and valgrind to track whether memory is 
+defined, undefined, or not accessible at all:
+
+  #define mi_track_mem_defined(p,size)
+  #define mi_track_mem_undefined(p,size)
+  #define mi_track_mem_noaccess(p,size)
+
+-------------------------------------------------------------------------------------------------------*/
+
+#if MI_TRACK_VALGRIND
+// valgrind tool
+
+#define MI_TRACK_ENABLED      1
+#define MI_TRACK_HEAP_DESTROY 1           // track free of individual blocks on heap_destroy
+#define MI_TRACK_TOOL         "valgrind"
+
+#include <valgrind/valgrind.h>
+#include <valgrind/memcheck.h>
+
+#define mi_track_malloc_size(p,reqsize,size,zero) VALGRIND_MALLOCLIKE_BLOCK(p,size,MI_PADDING_SIZE /*red zone*/,zero)
+#define mi_track_free_size(p,_size)               VALGRIND_FREELIKE_BLOCK(p,MI_PADDING_SIZE /*red zone*/)
+#define mi_track_resize(p,oldsize,newsize)        VALGRIND_RESIZEINPLACE_BLOCK(p,oldsize,newsize,MI_PADDING_SIZE /*red zone*/)
+#define mi_track_mem_defined(p,size)              VALGRIND_MAKE_MEM_DEFINED(p,size)
+#define mi_track_mem_undefined(p,size)            VALGRIND_MAKE_MEM_UNDEFINED(p,size)
+#define mi_track_mem_noaccess(p,size)             VALGRIND_MAKE_MEM_NOACCESS(p,size)
+
+#elif MI_TRACK_ASAN
+// address sanitizer
+
+#define MI_TRACK_ENABLED      1
+#define MI_TRACK_HEAP_DESTROY 0
+#define MI_TRACK_TOOL         "asan"
+
+#include <sanitizer/asan_interface.h>
+
+#define mi_track_malloc_size(p,reqsize,size,zero) ASAN_UNPOISON_MEMORY_REGION(p,size)
+#define mi_track_free_size(p,size)                ASAN_POISON_MEMORY_REGION(p,size)
+#define mi_track_mem_defined(p,size)              ASAN_UNPOISON_MEMORY_REGION(p,size)
+#define mi_track_mem_undefined(p,size)            ASAN_UNPOISON_MEMORY_REGION(p,size)
+#define mi_track_mem_noaccess(p,size)             ASAN_POISON_MEMORY_REGION(p,size)
+
+#elif MI_TRACK_ETW
+// windows event tracing
+
+#define MI_TRACK_ENABLED      1
+#define MI_TRACK_HEAP_DESTROY 0
+#define MI_TRACK_TOOL         "ETW"
+
+#define WIN32_LEAN_AND_MEAN
+#include <windows.h>
+#include "../src/prim/windows/etw.h"
+
+#define mi_track_init()                           EventRegistermicrosoft_windows_mimalloc();
+#define mi_track_malloc_size(p,reqsize,size,zero) EventWriteETW_MI_ALLOC((UINT64)(p), size)
+#define mi_track_free_size(p,size)                EventWriteETW_MI_FREE((UINT64)(p), size)
+
+#else
+// no tracking
+
+#define MI_TRACK_ENABLED      0
+#define MI_TRACK_HEAP_DESTROY 0 
+#define MI_TRACK_TOOL         "none"
+
+#define mi_track_malloc_size(p,reqsize,size,zero)
+#define mi_track_free_size(p,_size)
+
+#endif
+
+// -------------------
+// Utility definitions
+
+#ifndef mi_track_resize
+#define mi_track_resize(p,oldsize,newsize)      mi_track_free_size(p,oldsize); mi_track_malloc(p,newsize,false)
+#endif
+
+#ifndef mi_track_align
+#define mi_track_align(p,alignedp,offset,size)  mi_track_mem_noaccess(p,offset)
+#endif
+
+#ifndef mi_track_init
+#define mi_track_init()
+#endif
+
+#ifndef mi_track_mem_defined
+#define mi_track_mem_defined(p,size)
+#endif
+
+#ifndef mi_track_mem_undefined
+#define mi_track_mem_undefined(p,size)
+#endif
+
+#ifndef mi_track_mem_noaccess
+#define mi_track_mem_noaccess(p,size)
+#endif
+
+
+#if MI_PADDING
+#define mi_track_malloc(p,reqsize,zero) \
+  if ((p)!=NULL) { \
+    mi_assert_internal(mi_usable_size(p)==(reqsize)); \
+    mi_track_malloc_size(p,reqsize,reqsize,zero); \
+  }
+#else
+#define mi_track_malloc(p,reqsize,zero) \
+  if ((p)!=NULL) { \
+    mi_assert_internal(mi_usable_size(p)>=(reqsize)); \
+    mi_track_malloc_size(p,reqsize,mi_usable_size(p),zero); \
+  }
+#endif
+
+#endif

preload-mimalloc/mimalloc/include/mimalloc/types.h

@@ -0,0 +1,627 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#pragma once
+#ifndef MIMALLOC_TYPES_H
+#define MIMALLOC_TYPES_H
+
+// --------------------------------------------------------------------------
+// This file contains the main type definitions for mimalloc:
+// mi_heap_t      : all data for a thread-local heap, contains
+//                  lists of all managed heap pages.
+// mi_segment_t   : a larger chunk of memory (32GiB) from where pages
+//                  are allocated.
+// mi_page_t      : a mimalloc page (usually 64KiB or 512KiB) from
+//                  where objects are allocated.
+// --------------------------------------------------------------------------
+
+
+#include <stddef.h>   // ptrdiff_t
+#include <stdint.h>   // uintptr_t, uint16_t, etc
+#include "mimalloc/atomic.h"  // _Atomic
+
+#ifdef _MSC_VER
+#pragma warning(disable:4214) // bitfield is not int
+#endif
+
+// Minimal alignment necessary. On most platforms 16 bytes are needed
+// due to SSE registers for example. This must be at least `sizeof(void*)`
+#ifndef MI_MAX_ALIGN_SIZE
+#define MI_MAX_ALIGN_SIZE  16   // sizeof(max_align_t)
+#endif
+
+// ------------------------------------------------------
+// Variants
+// ------------------------------------------------------
+
+// Define NDEBUG in the release version to disable assertions.
+// #define NDEBUG
+
+// Define MI_TRACK_<tool> to enable tracking support
+// #define MI_TRACK_VALGRIND 1
+// #define MI_TRACK_ASAN     1
+// #define MI_TRACK_ETW      1
+
+// Define MI_STAT as 1 to maintain statistics; set it to 2 to have detailed statistics (but costs some performance).
+// #define MI_STAT 1
+
+// Define MI_SECURE to enable security mitigations
+// #define MI_SECURE 1  // guard page around metadata
+// #define MI_SECURE 2  // guard page around each mimalloc page
+// #define MI_SECURE 3  // encode free lists (detect corrupted free list (buffer overflow), and invalid pointer free)
+// #define MI_SECURE 4  // checks for double free. (may be more expensive)
+
+#if !defined(MI_SECURE)
+#define MI_SECURE 0
+#endif
+
+// Define MI_DEBUG for debug mode
+// #define MI_DEBUG 1  // basic assertion checks and statistics, check double free, corrupted free list, and invalid pointer free.
+// #define MI_DEBUG 2  // + internal assertion checks
+// #define MI_DEBUG 3  // + extensive internal invariant checking (cmake -DMI_DEBUG_FULL=ON)
+#if !defined(MI_DEBUG)
+#if !defined(NDEBUG) || defined(_DEBUG)
+#define MI_DEBUG 2
+#else
+#define MI_DEBUG 0
+#endif
+#endif
+
+// Reserve extra padding at the end of each block to be more resilient against heap block overflows.
+// The padding can detect buffer overflow on free.
+#if !defined(MI_PADDING) && (MI_SECURE>=3 || MI_DEBUG>=1 || (MI_TRACK_VALGRIND || MI_TRACK_ASAN || MI_TRACK_ETW))
+#define MI_PADDING  1
+#endif
+
+// Check padding bytes; allows byte-precise buffer overflow detection
+#if !defined(MI_PADDING_CHECK) && MI_PADDING && (MI_SECURE>=3 || MI_DEBUG>=1)
+#define MI_PADDING_CHECK 1
+#endif
+
+
+// Encoded free lists allow detection of corrupted free lists
+// and can detect buffer overflows, modify after free, and double `free`s.
+#if (MI_SECURE>=3 || MI_DEBUG>=1)
+#define MI_ENCODE_FREELIST  1
+#endif
+
+
+// We used to abandon huge pages but to eagerly deallocate if freed from another thread,
+// but that makes it not possible to visit them during a heap walk or include them in a
+// `mi_heap_destroy`. We therefore instead reset/decommit the huge blocks if freed from
+// another thread so most memory is available until it gets properly freed by the owning thread.
+// #define MI_HUGE_PAGE_ABANDON 1
+
+
+// ------------------------------------------------------
+// Platform specific values
+// ------------------------------------------------------
+
+// ------------------------------------------------------
+// Size of a pointer.
+// We assume that `sizeof(void*)==sizeof(intptr_t)`
+// and it holds for all platforms we know of.
+//
+// However, the C standard only requires that:
+//  p == (void*)((intptr_t)p))
+// but we also need:
+//  i == (intptr_t)((void*)i)
+// or otherwise one might define an intptr_t type that is larger than a pointer...
+// ------------------------------------------------------
+
+#if INTPTR_MAX > INT64_MAX
+# define MI_INTPTR_SHIFT (4)  // assume 128-bit  (as on arm CHERI for example)
+#elif INTPTR_MAX == INT64_MAX
+# define MI_INTPTR_SHIFT (3)
+#elif INTPTR_MAX == INT32_MAX
+# define MI_INTPTR_SHIFT (2)
+#else
+#error platform pointers must be 32, 64, or 128 bits
+#endif
+
+#if SIZE_MAX == UINT64_MAX
+# define MI_SIZE_SHIFT (3)
+typedef int64_t  mi_ssize_t;
+#elif SIZE_MAX == UINT32_MAX
+# define MI_SIZE_SHIFT (2)
+typedef int32_t  mi_ssize_t;
+#else
+#error platform objects must be 32 or 64 bits
+#endif
+
+#if (SIZE_MAX/2) > LONG_MAX
+# define MI_ZU(x)  x##ULL
+# define MI_ZI(x)  x##LL
+#else
+# define MI_ZU(x)  x##UL
+# define MI_ZI(x)  x##L
+#endif
+
+#define MI_INTPTR_SIZE  (1<<MI_INTPTR_SHIFT)
+#define MI_INTPTR_BITS  (MI_INTPTR_SIZE*8)
+
+#define MI_SIZE_SIZE  (1<<MI_SIZE_SHIFT)
+#define MI_SIZE_BITS  (MI_SIZE_SIZE*8)
+
+#define MI_KiB     (MI_ZU(1024))
+#define MI_MiB     (MI_KiB*MI_KiB)
+#define MI_GiB     (MI_MiB*MI_KiB)
+
+
+// ------------------------------------------------------
+// Main internal data-structures
+// ------------------------------------------------------
+
+// Main tuning parameters for segment and page sizes
+// Sizes for 64-bit (usually divide by two for 32-bit)
+#define MI_SEGMENT_SLICE_SHIFT            (13 + MI_INTPTR_SHIFT)         // 64KiB  (32KiB on 32-bit)
+
+#if MI_INTPTR_SIZE > 4
+#define MI_SEGMENT_SHIFT                  ( 9 + MI_SEGMENT_SLICE_SHIFT)  // 32MiB
+#else
+#define MI_SEGMENT_SHIFT                  ( 7 + MI_SEGMENT_SLICE_SHIFT)  // 4MiB on 32-bit
+#endif
+
+#define MI_SMALL_PAGE_SHIFT               (MI_SEGMENT_SLICE_SHIFT)       // 64KiB
+#define MI_MEDIUM_PAGE_SHIFT              ( 3 + MI_SMALL_PAGE_SHIFT)     // 512KiB
+
+
+// Derived constants
+#define MI_SEGMENT_SIZE                   (MI_ZU(1)<<MI_SEGMENT_SHIFT)
+#define MI_SEGMENT_ALIGN                  MI_SEGMENT_SIZE
+#define MI_SEGMENT_MASK                   (MI_SEGMENT_ALIGN - 1)
+#define MI_SEGMENT_SLICE_SIZE             (MI_ZU(1)<< MI_SEGMENT_SLICE_SHIFT)
+#define MI_SLICES_PER_SEGMENT             (MI_SEGMENT_SIZE / MI_SEGMENT_SLICE_SIZE) // 1024
+
+#define MI_SMALL_PAGE_SIZE                (MI_ZU(1)<<MI_SMALL_PAGE_SHIFT)
+#define MI_MEDIUM_PAGE_SIZE               (MI_ZU(1)<<MI_MEDIUM_PAGE_SHIFT)
+
+#define MI_SMALL_OBJ_SIZE_MAX             (MI_SMALL_PAGE_SIZE/4)   // 8KiB on 64-bit
+#define MI_MEDIUM_OBJ_SIZE_MAX            (MI_MEDIUM_PAGE_SIZE/4)  // 128KiB on 64-bit
+#define MI_MEDIUM_OBJ_WSIZE_MAX           (MI_MEDIUM_OBJ_SIZE_MAX/MI_INTPTR_SIZE)   
+#define MI_LARGE_OBJ_SIZE_MAX             (MI_SEGMENT_SIZE/2)      // 32MiB on 64-bit
+#define MI_LARGE_OBJ_WSIZE_MAX            (MI_LARGE_OBJ_SIZE_MAX/MI_INTPTR_SIZE)
+
+// Maximum number of size classes. (spaced exponentially in 12.5% increments)
+#define MI_BIN_HUGE  (73U)
+
+#if (MI_MEDIUM_OBJ_WSIZE_MAX >= 655360)
+#error "mimalloc internal: define more bins"
+#endif
+
+// Maximum slice offset (15)
+#define MI_MAX_SLICE_OFFSET               ((MI_ALIGNMENT_MAX / MI_SEGMENT_SLICE_SIZE) - 1)
+
+// Used as a special value to encode block sizes in 32 bits.
+#define MI_HUGE_BLOCK_SIZE                ((uint32_t)(2*MI_GiB))
+
+// blocks up to this size are always allocated aligned
+#define MI_MAX_ALIGN_GUARANTEE            (8*MI_MAX_ALIGN_SIZE)  
+
+// Alignments over MI_ALIGNMENT_MAX are allocated in dedicated huge page segments 
+#define MI_ALIGNMENT_MAX                  (MI_SEGMENT_SIZE >> 1)  
+
+
+// ------------------------------------------------------
+// Mimalloc pages contain allocated blocks
+// ------------------------------------------------------
+
+// The free lists use encoded next fields
+// (Only actually encodes when MI_ENCODED_FREELIST is defined.)
+typedef uintptr_t  mi_encoded_t;
+
+// thread id's
+typedef size_t     mi_threadid_t;
+
+// free lists contain blocks
+typedef struct mi_block_s {
+  mi_encoded_t next;
+} mi_block_t;
+
+
+// The delayed flags are used for efficient multi-threaded free-ing
+typedef enum mi_delayed_e {
+  MI_USE_DELAYED_FREE   = 0, // push on the owning heap thread delayed list
+  MI_DELAYED_FREEING    = 1, // temporary: another thread is accessing the owning heap
+  MI_NO_DELAYED_FREE    = 2, // optimize: push on page local thread free queue if another block is already in the heap thread delayed free list
+  MI_NEVER_DELAYED_FREE = 3  // sticky, only resets on page reclaim
+} mi_delayed_t;
+
+
+// The `in_full` and `has_aligned` page flags are put in a union to efficiently
+// test if both are false (`full_aligned == 0`) in the `mi_free` routine.
+#if !MI_TSAN
+typedef union mi_page_flags_s {
+  uint8_t full_aligned;
+  struct {
+    uint8_t in_full : 1;
+    uint8_t has_aligned : 1;
+  } x;
+} mi_page_flags_t;
+#else
+// under thread sanitizer, use a byte for each flag to suppress warning, issue #130
+typedef union mi_page_flags_s {
+  uint16_t full_aligned;
+  struct {
+    uint8_t in_full;
+    uint8_t has_aligned;
+  } x;
+} mi_page_flags_t;
+#endif
+
+// Thread free list.
+// We use the bottom 2 bits of the pointer for mi_delayed_t flags
+typedef uintptr_t mi_thread_free_t;
+
+// A page contains blocks of one specific size (`block_size`).
+// Each page has three list of free blocks:
+// `free` for blocks that can be allocated,
+// `local_free` for freed blocks that are not yet available to `mi_malloc`
+// `thread_free` for freed blocks by other threads
+// The `local_free` and `thread_free` lists are migrated to the `free` list
+// when it is exhausted. The separate `local_free` list is necessary to
+// implement a monotonic heartbeat. The `thread_free` list is needed for
+// avoiding atomic operations in the common case.
+//
+//
+// `used - |thread_free|` == actual blocks that are in use (alive)
+// `used - |thread_free| + |free| + |local_free| == capacity`
+//
+// We don't count `freed` (as |free|) but use `used` to reduce
+// the number of memory accesses in the `mi_page_all_free` function(s).
+//
+// Notes:
+// - Access is optimized for `mi_free` and `mi_page_alloc` (in `alloc.c`)
+// - Using `uint16_t` does not seem to slow things down
+// - The size is 8 words on 64-bit which helps the page index calculations
+//   (and 10 words on 32-bit, and encoded free lists add 2 words. Sizes 10
+//    and 12 are still good for address calculation)
+// - To limit the structure size, the `xblock_size` is 32-bits only; for
+//   blocks > MI_HUGE_BLOCK_SIZE the size is determined from the segment page size
+// - `thread_free` uses the bottom bits as a delayed-free flags to optimize
+//   concurrent frees where only the first concurrent free adds to the owning
+//   heap `thread_delayed_free` list (see `alloc.c:mi_free_block_mt`).
+//   The invariant is that no-delayed-free is only set if there is
+//   at least one block that will be added, or as already been added, to
+//   the owning heap `thread_delayed_free` list. This guarantees that pages
+//   will be freed correctly even if only other threads free blocks.
+typedef struct mi_page_s {
+  // "owned" by the segment
+  uint32_t              slice_count;       // slices in this page (0 if not a page)
+  uint32_t              slice_offset;      // distance from the actual page data slice (0 if a page)
+  uint8_t               is_reset : 1;      // `true` if the page memory was reset
+  uint8_t               is_committed : 1;  // `true` if the page virtual memory is committed
+  uint8_t               is_zero_init : 1;  // `true` if the page was zero initialized
+
+  // layout like this to optimize access in `mi_malloc` and `mi_free`
+  uint16_t              capacity;          // number of blocks committed, must be the first field, see `segment.c:page_clear`
+  uint16_t              reserved;          // number of blocks reserved in memory
+  mi_page_flags_t       flags;             // `in_full` and `has_aligned` flags (8 bits)
+  uint8_t               is_zero : 1;       // `true` if the blocks in the free list are zero initialized
+  uint8_t               retire_expire : 7; // expiration count for retired blocks
+
+  mi_block_t*           free;              // list of available free blocks (`malloc` allocates from this list)
+  uint32_t              used;              // number of blocks in use (including blocks in `local_free` and `thread_free`)
+  uint32_t              xblock_size;       // size available in each block (always `>0`)
+  mi_block_t*           local_free;        // list of deferred free blocks by this thread (migrates to `free`)
+
+  #if (MI_ENCODE_FREELIST || MI_PADDING)
+  uintptr_t             keys[2];           // two random keys to encode the free lists (see `_mi_block_next`) or padding canary
+  #endif
+
+  _Atomic(mi_thread_free_t) xthread_free;  // list of deferred free blocks freed by other threads
+  _Atomic(uintptr_t)        xheap;
+
+  struct mi_page_s*     next;              // next page owned by this thread with the same `block_size`
+  struct mi_page_s*     prev;              // previous page owned by this thread with the same `block_size`
+
+  // 64-bit 9 words, 32-bit 12 words, (+2 for secure)
+  #if MI_INTPTR_SIZE==8
+  uintptr_t padding[1];
+  #endif
+} mi_page_t;
+
+
+
+typedef enum mi_page_kind_e {
+  MI_PAGE_SMALL,    // small blocks go into 64KiB pages inside a segment
+  MI_PAGE_MEDIUM,   // medium blocks go into medium pages inside a segment
+  MI_PAGE_LARGE,    // larger blocks go into a page of just one block
+  MI_PAGE_HUGE,     // huge blocks (> 16 MiB) are put into a single page in a single segment.
+} mi_page_kind_t;
+
+typedef enum mi_segment_kind_e {
+  MI_SEGMENT_NORMAL, // MI_SEGMENT_SIZE size with pages inside.
+  MI_SEGMENT_HUGE,   // > MI_LARGE_SIZE_MAX segment with just one huge page inside.
+} mi_segment_kind_t;
+
+// ------------------------------------------------------
+// A segment holds a commit mask where a bit is set if
+// the corresponding MI_COMMIT_SIZE area is committed.
+// The MI_COMMIT_SIZE must be a multiple of the slice
+// size. If it is equal we have the most fine grained 
+// decommit (but setting it higher can be more efficient).
+// The MI_MINIMAL_COMMIT_SIZE is the minimal amount that will
+// be committed in one go which can be set higher than
+// MI_COMMIT_SIZE for efficiency (while the decommit mask
+// is still tracked in fine-grained MI_COMMIT_SIZE chunks)
+// ------------------------------------------------------
+
+#define MI_MINIMAL_COMMIT_SIZE      (16*MI_SEGMENT_SLICE_SIZE)           // 1MiB
+#define MI_COMMIT_SIZE              (MI_SEGMENT_SLICE_SIZE)              // 64KiB
+#define MI_COMMIT_MASK_BITS         (MI_SEGMENT_SIZE / MI_COMMIT_SIZE)  
+#define MI_COMMIT_MASK_FIELD_BITS    MI_SIZE_BITS
+#define MI_COMMIT_MASK_FIELD_COUNT  (MI_COMMIT_MASK_BITS / MI_COMMIT_MASK_FIELD_BITS)
+
+#if (MI_COMMIT_MASK_BITS != (MI_COMMIT_MASK_FIELD_COUNT * MI_COMMIT_MASK_FIELD_BITS))
+#error "the segment size must be exactly divisible by the (commit size * size_t bits)"
+#endif
+
+typedef struct mi_commit_mask_s {
+  size_t mask[MI_COMMIT_MASK_FIELD_COUNT];
+} mi_commit_mask_t;
+
+typedef mi_page_t  mi_slice_t;
+typedef int64_t    mi_msecs_t;
+
+
+// Segments are large allocated memory blocks (8mb on 64 bit) from
+// the OS. Inside segments we allocated fixed size _pages_ that
+// contain blocks.
+typedef struct mi_segment_s {
+  size_t            memid;              // memory id for arena allocation
+  bool              mem_is_pinned;      // `true` if we cannot decommit/reset/protect in this memory (i.e. when allocated using large OS pages)    
+  bool              mem_is_large;       // in large/huge os pages?
+  bool              mem_is_committed;   // `true` if the whole segment is eagerly committed
+  size_t            mem_alignment;      // page alignment for huge pages (only used for alignment > MI_ALIGNMENT_MAX)
+  size_t            mem_align_offset;   // offset for huge page alignment (only used for alignment > MI_ALIGNMENT_MAX)
+
+  bool              allow_decommit;     
+  mi_msecs_t        decommit_expire;
+  mi_commit_mask_t  decommit_mask;
+  mi_commit_mask_t  commit_mask;
+
+  _Atomic(struct mi_segment_s*) abandoned_next;
+
+  // from here is zero initialized
+  struct mi_segment_s* next;            // the list of freed segments in the cache (must be first field, see `segment.c:mi_segment_init`)
+  
+  size_t            abandoned;          // abandoned pages (i.e. the original owning thread stopped) (`abandoned <= used`)
+  size_t            abandoned_visits;   // count how often this segment is visited in the abandoned list (to force reclaim it it is too long)
+  size_t            used;               // count of pages in use
+  uintptr_t         cookie;             // verify addresses in debug mode: `mi_ptr_cookie(segment) == segment->cookie`  
+
+  size_t            segment_slices;      // for huge segments this may be different from `MI_SLICES_PER_SEGMENT`
+  size_t            segment_info_slices; // initial slices we are using segment info and possible guard pages.
+
+  // layout like this to optimize access in `mi_free`
+  mi_segment_kind_t kind;
+  size_t            slice_entries;       // entries in the `slices` array, at most `MI_SLICES_PER_SEGMENT`
+  _Atomic(mi_threadid_t) thread_id;      // unique id of the thread owning this segment
+
+  mi_slice_t        slices[MI_SLICES_PER_SEGMENT+1];  // one more for huge blocks with large alignment
+} mi_segment_t;
+
+
+// ------------------------------------------------------
+// Heaps
+// Provide first-class heaps to allocate from.
+// A heap just owns a set of pages for allocation and
+// can only be allocate/reallocate from the thread that created it.
+// Freeing blocks can be done from any thread though.
+// Per thread, the segments are shared among its heaps.
+// Per thread, there is always a default heap that is
+// used for allocation; it is initialized to statically
+// point to an empty heap to avoid initialization checks
+// in the fast path.
+// ------------------------------------------------------
+
+// Thread local data
+typedef struct mi_tld_s mi_tld_t;
+
+// Pages of a certain block size are held in a queue.
+typedef struct mi_page_queue_s {
+  mi_page_t* first;
+  mi_page_t* last;
+  size_t     block_size;
+} mi_page_queue_t;
+
+#define MI_BIN_FULL  (MI_BIN_HUGE+1)
+
+// Random context
+typedef struct mi_random_cxt_s {
+  uint32_t input[16];
+  uint32_t output[16];
+  int      output_available;
+  bool     weak;
+} mi_random_ctx_t;
+
+
+// In debug mode there is a padding structure at the end of the blocks to check for buffer overflows
+#if (MI_PADDING)
+typedef struct mi_padding_s {
+  uint32_t canary; // encoded block value to check validity of the padding (in case of overflow)
+  uint32_t delta;  // padding bytes before the block. (mi_usable_size(p) - delta == exact allocated bytes)
+} mi_padding_t;
+#define MI_PADDING_SIZE   (sizeof(mi_padding_t))
+#define MI_PADDING_WSIZE  ((MI_PADDING_SIZE + MI_INTPTR_SIZE - 1) / MI_INTPTR_SIZE)
+#else
+#define MI_PADDING_SIZE   0
+#define MI_PADDING_WSIZE  0
+#endif
+
+#define MI_PAGES_DIRECT   (MI_SMALL_WSIZE_MAX + MI_PADDING_WSIZE + 1)
+
+
+// A heap owns a set of pages.
+struct mi_heap_s {
+  mi_tld_t*             tld;
+  mi_page_t*            pages_free_direct[MI_PAGES_DIRECT];  // optimize: array where every entry points a page with possibly free blocks in the corresponding queue for that size.
+  mi_page_queue_t       pages[MI_BIN_FULL + 1];              // queue of pages for each size class (or "bin")
+  _Atomic(mi_block_t*)  thread_delayed_free;
+  mi_threadid_t         thread_id;                           // thread this heap belongs too
+  mi_arena_id_t         arena_id;                            // arena id if the heap belongs to a specific arena (or 0)  
+  uintptr_t             cookie;                              // random cookie to verify pointers (see `_mi_ptr_cookie`)
+  uintptr_t             keys[2];                             // two random keys used to encode the `thread_delayed_free` list
+  mi_random_ctx_t       random;                              // random number context used for secure allocation
+  size_t                page_count;                          // total number of pages in the `pages` queues.
+  size_t                page_retired_min;                    // smallest retired index (retired pages are fully free, but still in the page queues)
+  size_t                page_retired_max;                    // largest retired index into the `pages` array.
+  mi_heap_t*            next;                                // list of heaps per thread
+  bool                  no_reclaim;                          // `true` if this heap should not reclaim abandoned pages
+};
+
+
+
+// ------------------------------------------------------
+// Debug
+// ------------------------------------------------------
+
+#if !defined(MI_DEBUG_UNINIT)
+#define MI_DEBUG_UNINIT     (0xD0)
+#endif
+#if !defined(MI_DEBUG_FREED)
+#define MI_DEBUG_FREED      (0xDF)
+#endif
+#if !defined(MI_DEBUG_PADDING)
+#define MI_DEBUG_PADDING    (0xDE)
+#endif
+
+#if (MI_DEBUG)
+// use our own assertion to print without memory allocation
+void _mi_assert_fail(const char* assertion, const char* fname, unsigned int line, const char* func );
+#define mi_assert(expr)     ((expr) ? (void)0 : _mi_assert_fail(#expr,__FILE__,__LINE__,__func__))
+#else
+#define mi_assert(x)
+#endif
+
+#if (MI_DEBUG>1)
+#define mi_assert_internal    mi_assert
+#else
+#define mi_assert_internal(x)
+#endif
+
+#if (MI_DEBUG>2)
+#define mi_assert_expensive   mi_assert
+#else
+#define mi_assert_expensive(x)
+#endif
+
+// ------------------------------------------------------
+// Statistics
+// ------------------------------------------------------
+
+#ifndef MI_STAT
+#if (MI_DEBUG>0)
+#define MI_STAT 2
+#else
+#define MI_STAT 0
+#endif
+#endif
+
+typedef struct mi_stat_count_s {
+  int64_t allocated;
+  int64_t freed;
+  int64_t peak;
+  int64_t current;
+} mi_stat_count_t;
+
+typedef struct mi_stat_counter_s {
+  int64_t total;
+  int64_t count;
+} mi_stat_counter_t;
+
+typedef struct mi_stats_s {
+  mi_stat_count_t segments;
+  mi_stat_count_t pages;
+  mi_stat_count_t reserved;
+  mi_stat_count_t committed;
+  mi_stat_count_t reset;
+  mi_stat_count_t page_committed;
+  mi_stat_count_t segments_abandoned;
+  mi_stat_count_t pages_abandoned;
+  mi_stat_count_t threads;
+  mi_stat_count_t normal;
+  mi_stat_count_t huge;
+  mi_stat_count_t large;
+  mi_stat_count_t malloc;
+  mi_stat_count_t segments_cache;
+  mi_stat_counter_t pages_extended;
+  mi_stat_counter_t mmap_calls;
+  mi_stat_counter_t commit_calls;
+  mi_stat_counter_t page_no_retire;
+  mi_stat_counter_t searches;
+  mi_stat_counter_t normal_count;
+  mi_stat_counter_t huge_count;
+  mi_stat_counter_t large_count;
+#if MI_STAT>1
+  mi_stat_count_t normal_bins[MI_BIN_HUGE+1];
+#endif
+} mi_stats_t;
+
+
+void _mi_stat_increase(mi_stat_count_t* stat, size_t amount);
+void _mi_stat_decrease(mi_stat_count_t* stat, size_t amount);
+void _mi_stat_counter_increase(mi_stat_counter_t* stat, size_t amount);
+
+#if (MI_STAT)
+#define mi_stat_increase(stat,amount)         _mi_stat_increase( &(stat), amount)
+#define mi_stat_decrease(stat,amount)         _mi_stat_decrease( &(stat), amount)
+#define mi_stat_counter_increase(stat,amount) _mi_stat_counter_increase( &(stat), amount)
+#else
+#define mi_stat_increase(stat,amount)         (void)0
+#define mi_stat_decrease(stat,amount)         (void)0
+#define mi_stat_counter_increase(stat,amount) (void)0
+#endif
+
+#define mi_heap_stat_counter_increase(heap,stat,amount)  mi_stat_counter_increase( (heap)->tld->stats.stat, amount)
+#define mi_heap_stat_increase(heap,stat,amount)  mi_stat_increase( (heap)->tld->stats.stat, amount)
+#define mi_heap_stat_decrease(heap,stat,amount)  mi_stat_decrease( (heap)->tld->stats.stat, amount)
+
+// ------------------------------------------------------
+// Thread Local data
+// ------------------------------------------------------
+
+// A "span" is is an available range of slices. The span queues keep
+// track of slice spans of at most the given `slice_count` (but more than the previous size class).
+typedef struct mi_span_queue_s {
+  mi_slice_t* first;
+  mi_slice_t* last;
+  size_t      slice_count;
+} mi_span_queue_t;
+
+#define MI_SEGMENT_BIN_MAX (35)     // 35 == mi_segment_bin(MI_SLICES_PER_SEGMENT)
+
+// OS thread local data
+typedef struct mi_os_tld_s {
+  size_t                region_idx;   // start point for next allocation
+  mi_stats_t*           stats;        // points to tld stats
+} mi_os_tld_t;
+
+
+// Segments thread local data
+typedef struct mi_segments_tld_s {
+  mi_span_queue_t     spans[MI_SEGMENT_BIN_MAX+1];  // free slice spans inside segments
+  size_t              count;        // current number of segments;
+  size_t              peak_count;   // peak number of segments
+  size_t              current_size; // current size of all segments
+  size_t              peak_size;    // peak size of all segments
+  mi_stats_t*         stats;        // points to tld stats
+  mi_os_tld_t*        os;           // points to os stats
+} mi_segments_tld_t;
+
+// Thread local data
+struct mi_tld_s {
+  unsigned long long  heartbeat;     // monotonic heartbeat count
+  bool                recurse;       // true if deferred was called; used to prevent infinite recursion.
+  mi_heap_t*          heap_backing;  // backing heap of this thread (cannot be deleted)
+  mi_heap_t*          heaps;         // list of heaps in this thread (so we can abandon all when the thread terminates)
+  mi_segments_tld_t   segments;      // segment tld
+  mi_os_tld_t         os;            // os tld
+  mi_stats_t          stats;         // statistics
+};
+
+#endif

preload-mimalloc/mimalloc/readme.md

@@ -0,0 +1,822 @@
+
+<img align="left" width="100" height="100" src="doc/mimalloc-logo.png"/>
+
+[<img align="right" src="https://dev.azure.com/Daan0324/mimalloc/_apis/build/status/microsoft.mimalloc?branchName=dev"/>](https://dev.azure.com/Daan0324/mimalloc/_build?definitionId=1&_a=summary)
+
+# mimalloc
+
+&nbsp;
+
+mimalloc (pronounced "me-malloc")
+is a general purpose allocator with excellent [performance](#performance) characteristics.
+Initially developed by Daan Leijen for the runtime systems of the
+[Koka](https://koka-lang.github.io) and [Lean](https://github.com/leanprover/lean) languages.
+
+Latest release tag: `v2.1.1` (2023-04-03).
+Latest stable  tag: `v1.8.1` (2023-04-03).
+
+mimalloc is a drop-in replacement for `malloc` and can be used in other programs
+without code changes, for example, on dynamically linked ELF-based systems (Linux, BSD, etc.) you can use it as:
+```
+> LD_PRELOAD=/usr/lib/libmimalloc.so  myprogram
+```
+It also includes a robust way to override the default allocator in [Windows](#override_on_windows). Notable aspects of the design include:
+
+- __small and consistent__: the library is about 8k LOC using simple and
+  consistent data structures. This makes it very suitable
+  to integrate and adapt in other projects. For runtime systems it
+  provides hooks for a monotonic _heartbeat_ and deferred freeing (for
+  bounded worst-case times with reference counting).
+  Partly due to its simplicity, mimalloc has been ported to many systems (Windows, macOS,
+  Linux, WASM, various BSD's, Haiku, MUSL, etc) and has excellent support for dynamic overriding.
+- __free list sharding__: instead of one big free list (per size class) we have
+  many smaller lists per "mimalloc page" which reduces fragmentation and
+  increases locality --
+  things that are allocated close in time get allocated close in memory.
+  (A mimalloc page contains blocks of one size class and is usually 64KiB on a 64-bit system).
+- __free list multi-sharding__: the big idea! Not only do we shard the free list
+  per mimalloc page, but for each page we have multiple free lists. In particular, there
+  is one list for thread-local `free` operations, and another one for concurrent `free`
+  operations. Free-ing from another thread can now be a single CAS without needing
+  sophisticated coordination between threads. Since there will be
+  thousands of separate free lists, contention is naturally distributed over the heap,
+  and the chance of contending on a single location will be low -- this is quite
+  similar to randomized algorithms like skip lists where adding
+  a random oracle removes the need for a more complex algorithm.
+- __eager page reset__: when a "page" becomes empty (with increased chance
+  due to free list sharding) the memory is marked to the OS as unused (reset or decommitted)
+  reducing (real) memory pressure and fragmentation, especially in long running
+  programs.
+- __secure__: _mimalloc_ can be built in secure mode, adding guard pages,
+  randomized allocation, encrypted free lists, etc. to protect against various
+  heap vulnerabilities. The performance penalty is usually around 10% on average
+  over our benchmarks.
+- __first-class heaps__: efficiently create and use multiple heaps to allocate across different regions.
+  A heap can be destroyed at once instead of deallocating each object separately.
+- __bounded__: it does not suffer from _blowup_ \[1\], has bounded worst-case allocation
+  times (_wcat_) (upto OS primitives), bounded space overhead (~0.2% meta-data, with low
+  internal fragmentation), and has no internal points of contention using only atomic operations.
+- __fast__: In our benchmarks (see [below](#performance)),
+  _mimalloc_ outperforms other leading allocators (_jemalloc_, _tcmalloc_, _Hoard_, etc),
+  and often uses less memory. A nice property is that it does consistently well over a wide range
+  of benchmarks. There is also good huge OS page support for larger server programs.
+
+The [documentation](https://microsoft.github.io/mimalloc) gives a full overview of the API.
+You can read more on the design of _mimalloc_ in the [technical report](https://www.microsoft.com/en-us/research/publication/mimalloc-free-list-sharding-in-action) which also has detailed benchmark results.
+
+Enjoy!
+
+### Branches
+
+* `master`: latest stable release (based on `dev-slice`).
+* `dev`: development branch for mimalloc v1. Use this branch for submitting PR's.
+* `dev-slice`: development branch for mimalloc v2. This branch is downstream of `dev`.
+
+### Releases
+
+Note: the `v2.x` version has a new algorithm for managing internal mimalloc pages that tends to use reduce memory usage
+  and fragmentation compared to mimalloc `v1.x` (especially for large workloads). Should otherwise have similar performance
+  (see [below](#performance)); please report if you observe any significant performance regression.
+
+* 2023-04-03, `v1.8.1`, `v2.1.1`: Fixes build issues on some platforms.
+
+* 2023-03-29, `v1.8.0`, `v2.1.0`: Improved support dynamic overriding on Windows 11. Improved tracing precision
+  with [asan](#asan) and [Valgrind](#valgrind), and added Windows event tracing [ETW](#ETW) (contributed by Xinglong He). Created an OS
+  abstraction layer to make it easier to port and separate platform dependent code (in `src/prim`). Fixed C++ STL compilation on older Microsoft C++ compilers, and various small bug fixes.
+
+* 2022-12-23, `v1.7.9`, `v2.0.9`: Supports building with [asan](#asan) and improved [Valgrind](#valgrind) support.
+  Support abitrary large alignments (in particular for `std::pmr` pools). 
+  Added C++ STL allocators attached to a specific heap (thanks @vmarkovtsev). 
+  Heap walks now visit all object (including huge objects). Support Windows nano server containers (by Johannes Schindelin,@dscho). 
+  Various small bug fixes.
+
+* 2022-11-03, `v1.7.7`, `v2.0.7`: Initial support for [Valgrind](#valgrind) for leak testing and heap block overflow
+  detection. Initial
+  support for attaching heaps to a speficic memory area (only in v2). Fix `realloc` behavior for zero size blocks, remove restriction to integral multiple of the alignment in `alloc_align`, improved aligned allocation performance, reduced contention with many threads on few processors (thank you @dposluns!), vs2022 support, support `pkg-config`, .
+
+* 2022-04-14, `v1.7.6`, `v2.0.6`: fix fallback path for aligned OS allocation on Windows, improve Windows aligned allocation
+  even when compiling with older SDK's, fix dynamic overriding on macOS Monterey, fix MSVC C++ dynamic overriding, fix
+  warnings under Clang 14, improve performance if many OS threads are created and destroyed, fix statistics for large object
+  allocations, using MIMALLOC_VERBOSE=1 has no maximum on the number of error messages, various small fixes.
+
+* 2022-02-14, `v1.7.5`, `v2.0.5` (alpha): fix malloc override on
+  Windows 11, fix compilation with musl, potentially reduced
+  committed memory, add `bin/minject` for Windows,
+  improved wasm support, faster aligned allocation,
+  various small fixes.
+
+* 2021-11-14, `v1.7.3`, `v2.0.3` (beta): improved WASM support, improved macOS support and performance (including
+  M1), improved performance for v2 for large objects, Python integration improvements, more standard
+  installation directories, various small fixes.
+
+* 2021-06-17, `v1.7.2`, `v2.0.2` (beta): support M1, better installation layout on Linux, fix
+  thread_id on Android, prefer 2-6TiB area for aligned allocation to work better on pre-windows 8, various small fixes.
+
+* 2021-04-06, `v1.7.1`, `v2.0.1` (beta): fix bug in arena allocation for huge pages, improved aslr on large allocations, initial M1 support (still experimental).
+
+* 2021-01-31, `v2.0.0`: beta release 2.0: new slice algorithm for managing internal mimalloc pages.
+
+* 2021-01-31, `v1.7.0`: stable release 1.7: support explicit user provided memory regions, more precise statistics,
+  improve macOS overriding, initial support for Apple M1, improved DragonFly support, faster memcpy on Windows, various small fixes.
+
+* [Older release notes](#older-release-notes)
+
+Special thanks to:
+
+* [David Carlier](https://devnexen.blogspot.com/) (@devnexen) for his many contributions, and making
+  mimalloc work better on many less common operating systems, like Haiku, Dragonfly, etc.
+* Mary Feofanova (@mary3000), Evgeniy Moiseenko, and Manuel Pöter (@mpoeter) for making mimalloc TSAN checkable, and finding
+  memory model bugs using the [genMC] model checker.
+* Weipeng Liu (@pongba), Zhuowei Li, Junhua Wang, and Jakub Szymanski, for their early support of mimalloc and deployment
+  at large scale services, leading to many improvements in the mimalloc algorithms for large workloads.
+* Jason Gibson (@jasongibson) for exhaustive testing on large scale workloads and server environments, and finding complex bugs
+  in (early versions of) `mimalloc`.
+* Manuel Pöter (@mpoeter) and Sam Gross(@colesbury) for finding an ABA concurrency issue in abandoned segment reclamation. Sam also created the [no GIL](https://github.com/colesbury/nogil) Python fork which
+  uses mimalloc internally.
+
+
+[genMC]: https://plv.mpi-sws.org/genmc/
+
+### Usage
+
+mimalloc is used in various large scale low-latency services and programs, for example:
+
+<a href="https://www.bing.com"><img height="50" align="left" src="https://upload.wikimedia.org/wikipedia/commons/e/e9/Bing_logo.svg"></a>
+<a href="https://azure.microsoft.com/"><img height="50" align="left" src="https://upload.wikimedia.org/wikipedia/commons/a/a8/Microsoft_Azure_Logo.svg"></a>
+<a href="https://deathstrandingpc.505games.com"><img height="100" src="doc/ds-logo.png"></a>
+<a href="https://docs.unrealengine.com/4.26/en-US/WhatsNew/Builds/ReleaseNotes/4_25/"><img height="100" src="doc/unreal-logo.svg"></a>
+<a href="https://cab.spbu.ru/software/spades/"><img height="100" src="doc/spades-logo.png"></a>
+
+
+# Building
+
+## Windows
+
+Open `ide/vs2019/mimalloc.sln` in Visual Studio 2019 and build.
+The `mimalloc` project builds a static library (in `out/msvc-x64`), while the
+`mimalloc-override` project builds a DLL for overriding malloc
+in the entire program.
+
+## macOS, Linux, BSD, etc.
+
+We use [`cmake`](https://cmake.org)<sup>1</sup> as the build system:
+
+```
+> mkdir -p out/release
+> cd out/release
+> cmake ../..
+> make
+```
+This builds the library as a shared (dynamic)
+library (`.so` or `.dylib`), a static library (`.a`), and
+as a single object file (`.o`).
+
+`> sudo make install` (install the library and header files in `/usr/local/lib`  and `/usr/local/include`)
+
+You can build the debug version which does many internal checks and
+maintains detailed statistics as:
+
+```
+> mkdir -p out/debug
+> cd out/debug
+> cmake -DCMAKE_BUILD_TYPE=Debug ../..
+> make
+```
+This will name the shared library as `libmimalloc-debug.so`.
+
+Finally, you can build a _secure_ version that uses guard pages, encrypted
+free lists, etc., as:
+```
+> mkdir -p out/secure
+> cd out/secure
+> cmake -DMI_SECURE=ON ../..
+> make
+```
+This will name the shared library as `libmimalloc-secure.so`.
+Use `ccmake`<sup>2</sup> instead of `cmake`
+to see and customize all the available build options.
+
+Notes:
+1. Install CMake: `sudo apt-get install cmake`
+2. Install CCMake: `sudo apt-get install cmake-curses-gui`
+
+
+## Single source
+
+You can also directly build the single `src/static.c` file as part of your project without
+needing `cmake` at all. Make sure to also add the mimalloc `include` directory to the include path.
+
+
+# Using the library
+
+The preferred usage is including `<mimalloc.h>`, linking with
+the shared- or static library, and using the `mi_malloc` API exclusively for allocation. For example,
+```
+> gcc -o myprogram -lmimalloc myfile.c
+```
+
+mimalloc uses only safe OS calls (`mmap` and `VirtualAlloc`) and can co-exist
+with other allocators linked to the same program.
+If you use `cmake`, you can simply use:
+```
+find_package(mimalloc 1.4 REQUIRED)
+```
+in your `CMakeLists.txt` to find a locally installed mimalloc. Then use either:
+```
+target_link_libraries(myapp PUBLIC mimalloc)
+```
+to link with the shared (dynamic) library, or:
+```
+target_link_libraries(myapp PUBLIC mimalloc-static)
+```
+to link with the static library. See `test\CMakeLists.txt` for an example.
+
+For best performance in C++ programs, it is also recommended to override the
+global `new` and `delete` operators. For convience, mimalloc provides
+[`mimalloc-new-delete.h`](https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-new-delete.h) which does this for you -- just include it in a single(!) source file in your project.
+In C++, mimalloc also provides the `mi_stl_allocator` struct which implements the `std::allocator`
+interface.
+
+You can pass environment variables to print verbose messages (`MIMALLOC_VERBOSE=1`)
+and statistics (`MIMALLOC_SHOW_STATS=1`) (in the debug version):
+```
+> env MIMALLOC_SHOW_STATS=1 ./cfrac 175451865205073170563711388363
+
+175451865205073170563711388363 = 374456281610909315237213 * 468551
+
+heap stats:     peak      total      freed       unit
+normal   2:    16.4 kb    17.5 mb    17.5 mb      16 b   ok
+normal   3:    16.3 kb    15.2 mb    15.2 mb      24 b   ok
+normal   4:      64 b      4.6 kb     4.6 kb      32 b   ok
+normal   5:      80 b    118.4 kb   118.4 kb      40 b   ok
+normal   6:      48 b       48 b       48 b       48 b   ok
+normal  17:     960 b      960 b      960 b      320 b   ok
+
+heap stats:     peak      total      freed       unit
+    normal:    33.9 kb    32.8 mb    32.8 mb       1 b   ok
+      huge:       0 b        0 b        0 b        1 b   ok
+     total:    33.9 kb    32.8 mb    32.8 mb       1 b   ok
+malloc requested:         32.8 mb
+
+ committed:    58.2 kb    58.2 kb    58.2 kb       1 b   ok
+  reserved:     2.0 mb     2.0 mb     2.0 mb       1 b   ok
+     reset:       0 b        0 b        0 b        1 b   ok
+  segments:       1          1          1
+-abandoned:       0
+     pages:       6          6          6
+-abandoned:       0
+     mmaps:       3
+ mmap fast:       0
+ mmap slow:       1
+   threads:       0
+   elapsed:     2.022s
+   process: user: 1.781s, system: 0.016s, faults: 756, reclaims: 0, rss: 2.7 mb
+```
+
+The above model of using the `mi_` prefixed API is not always possible
+though in existing programs that already use the standard malloc interface,
+and another option is to override the standard malloc interface
+completely and redirect all calls to the _mimalloc_ library instead .
+
+## Environment Options
+
+You can set further options either programmatically (using [`mi_option_set`](https://microsoft.github.io/mimalloc/group__options.html)),
+or via environment variables:
+
+- `MIMALLOC_SHOW_STATS=1`: show statistics when the program terminates.
+- `MIMALLOC_VERBOSE=1`: show verbose messages.
+- `MIMALLOC_SHOW_ERRORS=1`: show error and warning messages.
+- `MIMALLOC_PAGE_RESET=0`: by default, mimalloc will reset (or purge) OS pages that are not in use, to signal to the OS
+   that the underlying physical memory can be reused. This can reduce memory fragmentation in long running (server)
+   programs. By setting it to `0` this will no longer be done which can improve performance for batch-like programs.
+   As an alternative, the `MIMALLOC_RESET_DELAY=`<msecs> can be set higher (100ms by default) to make the page
+   reset occur less frequently instead of turning it off completely.
+- `MIMALLOC_USE_NUMA_NODES=N`: pretend there are at most `N` NUMA nodes. If not set, the actual NUMA nodes are detected
+   at runtime. Setting `N` to 1 may avoid problems in some virtual environments. Also, setting it to a lower number than
+   the actual NUMA nodes is fine and will only cause threads to potentially allocate more memory across actual NUMA
+   nodes (but this can happen in any case as NUMA local allocation is always a best effort but not guaranteed).
+- `MIMALLOC_LARGE_OS_PAGES=1`: use large OS pages (2MiB) when available; for some workloads this can significantly
+   improve performance. Use `MIMALLOC_VERBOSE` to check if the large OS pages are enabled -- usually one needs
+   to explicitly allow large OS pages (as on [Windows][windows-huge] and [Linux][linux-huge]). However, sometimes
+   the OS is very slow to reserve contiguous physical memory for large OS pages so use with care on systems that
+   can have fragmented memory (for that reason, we generally recommend to use `MIMALLOC_RESERVE_HUGE_OS_PAGES` instead whenever possible).
+   <!--
+   - `MIMALLOC_EAGER_REGION_COMMIT=1`: on Windows, commit large (256MiB) regions eagerly. On Windows, these regions
+   show in the working set even though usually just a small part is committed to physical memory. This is why it
+   turned off by default on Windows as it looks not good in the task manager. However, turning it on has no
+   real drawbacks and may improve performance by a little.
+   -->
+- `MIMALLOC_RESERVE_HUGE_OS_PAGES=N`: where N is the number of 1GiB _huge_ OS pages. This reserves the huge pages at
+   startup and sometimes this can give a large (latency) performance improvement on big workloads.
+   Usually it is better to not use
+   `MIMALLOC_LARGE_OS_PAGES` in combination with this setting. Just like large OS pages, use with care as reserving
+   contiguous physical memory can take a long time when memory is fragmented (but reserving the huge pages is done at
+   startup only once).
+   Note that we usually need to explicitly enable huge OS pages (as on [Windows][windows-huge] and [Linux][linux-huge])).
+   With huge OS pages, it may be beneficial to set the setting
+   `MIMALLOC_EAGER_COMMIT_DELAY=N` (`N` is 1 by default) to delay the initial `N` segments (of 4MiB)
+   of a thread to not allocate in the huge OS pages; this prevents threads that are short lived
+   and allocate just a little to take up space in the huge OS page area (which cannot be reset).
+   The huge pages are usually allocated evenly among NUMA nodes.
+   We can use `MIMALLOC_RESERVE_HUGE_OS_PAGES_AT=N` where `N` is the numa node (starting at 0) to allocate all
+   the huge pages at a specific numa node instead.
+
+Use caution when using `fork` in combination with either large or huge OS pages: on a fork, the OS uses copy-on-write
+for all pages in the original process including the huge OS pages. When any memory is now written in that area, the
+OS will copy the entire 1GiB huge page (or 2MiB large page) which can cause the memory usage to grow in large increments.
+
+[linux-huge]: https://access.redhat.com/documentation/en-us/red_hat_enterprise_linux/5/html/tuning_and_optimizing_red_hat_enterprise_linux_for_oracle_9i_and_10g_databases/sect-oracle_9i_and_10g_tuning_guide-large_memory_optimization_big_pages_and_huge_pages-configuring_huge_pages_in_red_hat_enterprise_linux_4_or_5
+[windows-huge]: https://docs.microsoft.com/en-us/sql/database-engine/configure-windows/enable-the-lock-pages-in-memory-option-windows?view=sql-server-2017
+
+## Secure Mode
+
+_mimalloc_ can be build in secure mode by using the `-DMI_SECURE=ON` flags in `cmake`. This build enables various mitigations
+to make mimalloc more robust against exploits. In particular:
+
+- All internal mimalloc pages are surrounded by guard pages and the heap metadata is behind a guard page as well (so a buffer overflow
+  exploit cannot reach into the metadata).
+- All free list pointers are
+  [encoded](https://github.com/microsoft/mimalloc/blob/783e3377f79ee82af43a0793910a9f2d01ac7863/include/mimalloc-internal.h#L396)
+  with per-page keys which is used both to prevent overwrites with a known pointer, as well as to detect heap corruption.
+- Double free's are detected (and ignored).
+- The free lists are initialized in a random order and allocation randomly chooses between extension and reuse within a page to
+  mitigate against attacks that rely on a predicable allocation order. Similarly, the larger heap blocks allocated by mimalloc
+  from the OS are also address randomized.
+
+As always, evaluate with care as part of an overall security strategy as all of the above are mitigations but not guarantees.
+
+## Debug Mode
+
+When _mimalloc_ is built using debug mode, various checks are done at runtime to catch development errors.
+
+- Statistics are maintained in detail for each object size. They can be shown using `MIMALLOC_SHOW_STATS=1` at runtime.
+- All objects have padding at the end to detect (byte precise) heap block overflows.
+- Double free's, and freeing invalid heap pointers are detected.
+- Corrupted free-lists and some forms of use-after-free are detected.
+
+
+# Overriding Standard Malloc
+
+Overriding the standard `malloc` (and `new`) can be done either _dynamically_ or _statically_.
+
+## Dynamic override
+
+This is the recommended way to override the standard malloc interface.
+
+### Dynamic Override on Linux, BSD
+
+On these ELF-based systems we preload the mimalloc shared
+library so all calls to the standard `malloc` interface are
+resolved to the _mimalloc_ library.
+```
+> env LD_PRELOAD=/usr/lib/libmimalloc.so myprogram
+```
+
+You can set extra environment variables to check that mimalloc is running,
+like:
+```
+> env MIMALLOC_VERBOSE=1 LD_PRELOAD=/usr/lib/libmimalloc.so myprogram
+```
+or run with the debug version to get detailed statistics:
+```
+> env MIMALLOC_SHOW_STATS=1 LD_PRELOAD=/usr/lib/libmimalloc-debug.so myprogram
+```
+
+### Dynamic Override on MacOS
+
+On macOS we can also preload the mimalloc shared
+library so all calls to the standard `malloc` interface are
+resolved to the _mimalloc_ library.
+```
+> env DYLD_INSERT_LIBRARIES=/usr/lib/libmimalloc.dylib myprogram
+```
+
+Note that certain security restrictions may apply when doing this from
+the [shell](https://stackoverflow.com/questions/43941322/dyld-insert-libraries-ignored-when-calling-application-through-bash).
+
+
+### Dynamic Override on Windows
+
+<span id="override_on_windows">Overriding on Windows</span> is robust and has the
+particular advantage to be able to redirect all malloc/free calls that go through
+the (dynamic) C runtime allocator, including those from other DLL's or libraries.
+
+The overriding on Windows requires that you link your program explicitly with
+the mimalloc DLL and use the C-runtime library as a DLL (using the `/MD` or `/MDd` switch).
+Also, the `mimalloc-redirect.dll` (or `mimalloc-redirect32.dll`) must be put
+in the same folder as the main `mimalloc-override.dll` at runtime (as it is a dependency).
+The redirection DLL ensures that all calls to the C runtime malloc API get redirected to
+mimalloc (in `mimalloc-override.dll`).
+
+To ensure the mimalloc DLL is loaded at run-time it is easiest to insert some
+call to the mimalloc API in the `main` function, like `mi_version()`
+(or use the `/INCLUDE:mi_version` switch on the linker). See the `mimalloc-override-test` project
+for an example on how to use this. For best performance on Windows with C++, it
+is also recommended to also override the `new`/`delete` operations (by including
+[`mimalloc-new-delete.h`](https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-new-delete.h) a single(!) source file in your project).
+
+The environment variable `MIMALLOC_DISABLE_REDIRECT=1` can be used to disable dynamic
+overriding at run-time. Use `MIMALLOC_VERBOSE=1` to check if mimalloc was successfully redirected.
+
+(Note: in principle, it is possible to even patch existing executables without any recompilation
+if they are linked with the dynamic C runtime (`ucrtbase.dll`) -- just put the `mimalloc-override.dll`
+into the import table (and put `mimalloc-redirect.dll` in the same folder)
+Such patching can be done for example with [CFF Explorer](https://ntcore.com/?page_id=388)).
+
+
+## Static override
+
+On Unix-like systems, you can also statically link with _mimalloc_ to override the standard
+malloc interface. The recommended way is to link the final program with the
+_mimalloc_ single object file (`mimalloc.o`). We use
+an object file instead of a library file as linkers give preference to
+that over archives to resolve symbols. To ensure that the standard
+malloc interface resolves to the _mimalloc_ library, link it as the first
+object file. For example:
+```
+> gcc -o myprogram mimalloc.o  myfile1.c ...
+```
+
+Another way to override statically that works on all platforms, is to
+link statically to mimalloc (as shown in the introduction) and include a
+header file in each source file that re-defines `malloc` etc. to `mi_malloc`.
+This is provided by [`mimalloc-override.h`](https://github.com/microsoft/mimalloc/blob/master/include/mimalloc-override.h). This only works reliably though if all sources are
+under your control or otherwise mixing of pointers from different heaps may occur!
+
+
+## Tools
+
+Generally, we recommend using the standard allocator with memory tracking tools, but mimalloc
+can also be build to support the [address sanitizer][asan] or the excellent [Valgrind] tool. 
+Moreover, it can be build to support Windows event tracing ([ETW]).
+This has a small performance overhead but does allow detecting memory leaks and byte-precise 
+buffer overflows directly on final executables. See also the `test/test-wrong.c` file to test with various tools.
+
+### Valgrind
+
+To build with [valgrind] support, use the `MI_TRACK_VALGRIND=ON` cmake option:
+
+```
+> cmake ../.. -DMI_TRACK_VALGRIND=ON
+```
+
+This can also be combined with secure mode or debug mode.
+You can then run your programs directly under valgrind:
+
+```
+> valgrind <myprogram>
+```
+
+If you rely on overriding `malloc`/`free` by mimalloc (instead of using the `mi_malloc`/`mi_free` API directly),
+you also need to tell `valgrind` to not intercept those calls itself, and use:
+
+```
+> MIMALLOC_SHOW_STATS=1 valgrind  --soname-synonyms=somalloc=*mimalloc* -- <myprogram>
+```
+
+By setting the `MIMALLOC_SHOW_STATS` environment variable you can check that mimalloc is indeed
+used and not the standard allocator. Even though the [Valgrind option][valgrind-soname]
+is called `--soname-synonyms`, this also
+works when overriding with a static library or object file. Unfortunately, it is not possible to
+dynamically override mimalloc using `LD_PRELOAD` together with `valgrind`.
+See also the `test/test-wrong.c` file to test with `valgrind`.
+
+Valgrind support is in its initial development -- please report any issues.
+
+[Valgrind]: https://valgrind.org/
+[valgrind-soname]: https://valgrind.org/docs/manual/manual-core.html#opt.soname-synonyms
+
+### ASAN
+
+To build with the address sanitizer, use the `-DMI_TRACK_ASAN=ON` cmake option:
+
+```
+> cmake ../.. -DMI_TRACK_ASAN=ON
+```
+
+This can also be combined with secure mode or debug mode. 
+You can then run your programs as:'
+
+```
+> ASAN_OPTIONS=verbosity=1 <myprogram>
+```
+
+When you link a program with an address sanitizer build of mimalloc, you should
+generally compile that program too with the address sanitizer enabled. 
+For example, assuming you build mimalloc in `out/debug`:
+
+```
+clang -g -o test-wrong -Iinclude test/test-wrong.c out/debug/libmimalloc-asan-debug.a -lpthread -fsanitize=address -fsanitize-recover=address
+```
+
+Since the address sanitizer redirects the standard allocation functions, on some platforms (macOSX for example)
+it is required to compile mimalloc with `-DMI_OVERRIDE=OFF`.
+Adress sanitizer support is in its initial development -- please report any issues.
+
+[asan]: https://github.com/google/sanitizers/wiki/AddressSanitizer
+
+### ETW
+
+Event tracing for Windows ([ETW]) provides a high performance way to capture all allocations though
+mimalloc and analyze them later. To build with ETW support, use the `-DMI_TRACE_ETW=ON` cmake option. 
+
+You can then capture an allocation trace using the Windows performance recorder (WPR), using the 
+`src/prim/windows/etw-mimalloc.wprp` profile. In an admin prompt, you can use:
+```
+> wpr -start src\prim\windows\etw-mimalloc.wprp -filemode
+> <my_mimalloc_program>
+> wpr -stop <my_mimalloc_program>.etl
+``` 
+and then open `<my_mimalloc_program>.etl` in the Windows Performance Analyzer (WPA), or 
+use a tool like [TraceControl] that is specialized for analyzing mimalloc traces.
+
+[ETW]: https://learn.microsoft.com/en-us/windows-hardware/test/wpt/event-tracing-for-windows
+[TraceControl]: https://github.com/xinglonghe/TraceControl
+
+
+# Performance
+
+Last update: 2021-01-30
+
+We tested _mimalloc_ against many other top allocators over a wide
+range of benchmarks, ranging from various real world programs to
+synthetic benchmarks that see how the allocator behaves under more
+extreme circumstances. In our benchmark suite, _mimalloc_ outperforms other leading
+allocators (_jemalloc_, _tcmalloc_, _Hoard_, etc), and has a similar memory footprint. A nice property is that it
+does consistently well over the wide range of benchmarks.
+
+General memory allocators are interesting as there exists no algorithm that is
+optimal -- for a given allocator one can usually construct a workload
+where it does not do so well. The goal is thus to find an allocation
+strategy that performs well over a wide range of benchmarks without
+suffering from (too much) underperformance in less common situations.
+
+As always, interpret these results with care since some benchmarks test synthetic
+or uncommon situations that may never apply to your workloads. For example, most
+allocators do not do well on `xmalloc-testN` but that includes even the best
+industrial allocators like _jemalloc_ and _tcmalloc_ that are used in some of
+the world's largest systems (like Chrome or FreeBSD).
+
+Also, the benchmarks here do not measure the behaviour on very large and long-running server workloads,
+or worst-case latencies of allocation. Much work has gone into `mimalloc` to work well on such
+workloads (for example, to reduce virtual memory fragmentation on long-running services)
+but such optimizations are not always reflected in the current benchmark suite.
+
+We show here only an overview -- for
+more specific details and further benchmarks we refer to the
+[technical report](https://www.microsoft.com/en-us/research/publication/mimalloc-free-list-sharding-in-action).
+The benchmark suite is automated and available separately
+as [mimalloc-bench](https://github.com/daanx/mimalloc-bench).
+
+
+## Benchmark Results on a 16-core AMD 5950x (Zen3)
+
+Testing on the 16-core AMD 5950x processor at 3.4Ghz (4.9Ghz boost), with
+with 32GiB memory at 3600Mhz, running	Ubuntu 20.04 with glibc 2.31 and GCC 9.3.0.
+
+We measure three versions of _mimalloc_: the main version `mi` (tag:v1.7.0),
+the new v2.0 beta version as `xmi` (tag:v2.0.0), and the main version in secure mode as `smi` (tag:v1.7.0).
+
+The other allocators are
+Google's [_tcmalloc_](https://github.com/gperftools/gperftools) (`tc`, tag:gperftools-2.8.1) used in Chrome,
+Facebook's [_jemalloc_](https://github.com/jemalloc/jemalloc) (`je`, tag:5.2.1) by Jason Evans used in Firefox and FreeBSD,
+the Intel thread building blocks [allocator](https://github.com/intel/tbb) (`tbb`, tag:v2020.3),
+[rpmalloc](https://github.com/mjansson/rpmalloc) (`rp`,tag:1.4.1) by Mattias Jansson,
+the original scalable [_Hoard_](https://github.com/emeryberger/Hoard) (git:d880f72) allocator by Emery Berger \[1],
+the memory compacting [_Mesh_](https://github.com/plasma-umass/Mesh) (git:67ff31a) allocator by
+Bobby Powers _et al_ \[8],
+and finally the default system allocator (`glibc`, 2.31) (based on _PtMalloc2_).
+
+<img width="90%" src="doc/bench-2021/bench-amd5950x-2021-01-30-a.svg"/>
+<img width="90%" src="doc/bench-2021/bench-amd5950x-2021-01-30-b.svg"/>
+
+Any benchmarks ending in `N` run on all 32 logical cores in parallel.
+Results are averaged over 10 runs and reported relative
+to mimalloc (where 1.2 means it took 1.2&times; longer to run).
+The legend also contains the _overall relative score_ between the
+allocators where 100 points is the maximum if an allocator is fastest on
+all benchmarks.
+
+The single threaded _cfrac_ benchmark by Dave Barrett is an implementation of
+continued fraction factorization which uses many small short-lived allocations.
+All allocators do well on such common usage, where _mimalloc_ is just a tad
+faster than _tcmalloc_ and
+_jemalloc_.
+
+The _leanN_ program is interesting as a large realistic and
+concurrent workload of the [Lean](https://github.com/leanprover/lean)
+theorem prover compiling its own standard library, and there is a 13%
+speedup over _tcmalloc_. This is
+quite significant: if Lean spends 20% of its time in the
+allocator that means that _mimalloc_ is 1.6&times; faster than _tcmalloc_
+here. (This is surprising as that is not measured in a pure
+allocation benchmark like _alloc-test_. We conjecture that we see this
+outsized improvement here because _mimalloc_ has better locality in
+the allocation which improves performance for the *other* computations
+in a program as well).
+
+The single threaded _redis_ benchmark again show that most allocators do well on such workloads.
+
+The _larsonN_ server benchmark by Larson and Krishnan \[2] allocates and frees between threads. They observed this
+behavior (which they call _bleeding_) in actual server applications, and the benchmark simulates this.
+Here, _mimalloc_ is quite a bit faster than _tcmalloc_ and _jemalloc_ probably due to the object migration between different threads.
+
+The _mstressN_ workload performs many allocations and re-allocations,
+and migrates objects between threads (as in _larsonN_). However, it also
+creates  and destroys the _N_ worker threads a few times keeping some objects
+alive beyond the life time of the allocating thread. We observed this
+behavior in many larger server applications.
+
+The [_rptestN_](https://github.com/mjansson/rpmalloc-benchmark) benchmark
+by Mattias Jansson is a allocator test originally designed
+for _rpmalloc_, and tries to simulate realistic allocation patterns over
+multiple threads. Here the differences between allocators become more apparent.
+
+The second benchmark set tests specific aspects of the allocators and
+shows even more extreme differences between them.
+
+The _alloc-test_, by
+[OLogN Technologies AG](http://ithare.com/testing-memory-allocators-ptmalloc2-tcmalloc-hoard-jemalloc-while-trying-to-simulate-real-world-loads/), is a very allocation intensive benchmark doing millions of
+allocations in various size classes. The test is scaled such that when an
+allocator performs almost identically on _alloc-test1_ as _alloc-testN_ it
+means that it scales linearly.
+
+The _sh6bench_ and _sh8bench_ benchmarks are
+developed by [MicroQuill](http://www.microquill.com/) as part of SmartHeap.
+In _sh6bench_ _mimalloc_ does much
+better than the others (more than 2.5&times; faster than _jemalloc_).
+We cannot explain this well but believe it is
+caused in part by the "reverse" free-ing pattern in _sh6bench_.
+The _sh8bench_ is a variation with object migration
+between threads; whereas _tcmalloc_ did well on _sh6bench_, the addition of object migration causes it to be 10&times; slower than before.
+
+The _xmalloc-testN_ benchmark by Lever and Boreham \[5] and Christian Eder, simulates an asymmetric workload where
+some threads only allocate, and others only free -- they observed this pattern in
+larger server applications. Here we see that
+the _mimalloc_ technique of having non-contended sharded thread free
+lists pays off as it outperforms others by a very large margin. Only _rpmalloc_, _tbb_, and _glibc_ also scale well on this benchmark.
+
+The _cache-scratch_ benchmark by Emery Berger \[1], and introduced with
+the Hoard allocator to test for _passive-false_ sharing of cache lines.
+With a single thread they all
+perform the same, but when running with multiple threads the potential allocator
+induced false sharing of the cache lines can cause large run-time differences.
+Crundal \[6] describes in detail why the false cache line sharing occurs in the _tcmalloc_ design, and also discusses how this
+can be avoided with some small implementation changes.
+Only the _tbb_, _rpmalloc_ and _mesh_ allocators also avoid the
+cache line sharing completely, while _Hoard_ and _glibc_ seem to mitigate
+the effects. Kukanov and Voss \[7] describe in detail
+how the design of _tbb_ avoids the false cache line sharing.
+
+
+## On a 36-core Intel Xeon
+
+For completeness, here are the results on a big Amazon
+[c5.18xlarge](https://aws.amazon.com/ec2/instance-types/#Compute_Optimized) instance
+consisting of a 2&times;18-core Intel Xeon (Cascade Lake) at 3.4GHz (boost 3.5GHz)
+with 144GiB ECC memory, running	Ubuntu 20.04 with glibc 2.31, GCC 9.3.0, and
+Clang 10.0.0. This time, the mimalloc allocators (mi, xmi, and smi) were
+compiled with the Clang compiler instead of GCC.
+The results are similar to the AMD results but it is interesting to
+see the differences in the _larsonN_, _mstressN_, and _xmalloc-testN_ benchmarks.
+
+<img width="90%" src="doc/bench-2021/bench-c5-18xlarge-2021-01-30-a.svg"/>
+<img width="90%" src="doc/bench-2021/bench-c5-18xlarge-2021-01-30-b.svg"/>
+
+
+## Peak Working Set
+
+The following figure shows the peak working set (rss) of the allocators
+on the benchmarks (on the c5.18xlarge instance).
+
+<img width="90%" src="doc/bench-2021/bench-c5-18xlarge-2021-01-30-rss-a.svg"/>
+<img width="90%" src="doc/bench-2021/bench-c5-18xlarge-2021-01-30-rss-b.svg"/>
+
+Note that the _xmalloc-testN_ memory usage should be disregarded as it
+allocates more the faster the program runs. Similarly, memory usage of
+_larsonN_, _mstressN_, _rptestN_ and _sh8bench_ can vary depending on scheduling and
+speed. Nevertheless, we hope to improve the memory usage on _mstressN_
+and _rptestN_ (just as _cfrac_, _larsonN_ and _sh8bench_ have a small working set which skews the results).
+
+<!--
+# Previous Benchmarks
+
+Todo: should we create a separate page for this?
+
+## Benchmark Results on 36-core Intel: 2020-01-20
+
+Testing on a big Amazon EC2 compute instance
+([c5.18xlarge](https://aws.amazon.com/ec2/instance-types/#Compute_Optimized))
+consisting of a 72 processor Intel Xeon at 3GHz
+with 144GiB ECC memory, running	Ubuntu 18.04.1 with glibc 2.27 and GCC 7.4.0.
+The measured allocators are _mimalloc_ (xmi, tag:v1.4.0, page reset enabled)
+and its secure build as _smi_,
+Google's [_tcmalloc_](https://github.com/gperftools/gperftools) (tc, tag:gperftools-2.7) used in Chrome,
+Facebook's [_jemalloc_](https://github.com/jemalloc/jemalloc) (je, tag:5.2.1) by Jason Evans used in Firefox and FreeBSD,
+the Intel thread building blocks [allocator](https://github.com/intel/tbb) (tbb, tag:2020),
+[rpmalloc](https://github.com/mjansson/rpmalloc) (rp,tag:1.4.0) by Mattias Jansson,
+the original scalable [_Hoard_](https://github.com/emeryberger/Hoard) (tag:3.13) allocator by Emery Berger \[1],
+the memory compacting [_Mesh_](https://github.com/plasma-umass/Mesh) (git:51222e7) allocator by
+Bobby Powers _et al_ \[8],
+and finally the default system allocator (glibc, 2.27) (based on _PtMalloc2_).
+
+<img width="90%" src="doc/bench-2020/bench-c5-18xlarge-2020-01-20-a.svg"/>
+<img width="90%" src="doc/bench-2020/bench-c5-18xlarge-2020-01-20-b.svg"/>
+
+The following figure shows the peak working set (rss) of the allocators
+on the benchmarks (on the c5.18xlarge instance).
+
+<img width="90%" src="doc/bench-2020/bench-c5-18xlarge-2020-01-20-rss-a.svg"/>
+<img width="90%" src="doc/bench-2020/bench-c5-18xlarge-2020-01-20-rss-b.svg"/>
+
+
+## On 24-core AMD Epyc, 2020-01-16
+
+For completeness, here are the results on a
+[r5a.12xlarge](https://aws.amazon.com/ec2/instance-types/#Memory_Optimized) instance
+having a 48 processor AMD Epyc 7000 at 2.5GHz with 384GiB of memory.
+The results are similar to the Intel results but it is interesting to
+see the differences in the _larsonN_, _mstressN_, and _xmalloc-testN_ benchmarks.
+
+<img width="90%" src="doc/bench-2020/bench-r5a-12xlarge-2020-01-16-a.svg"/>
+<img width="90%" src="doc/bench-2020/bench-r5a-12xlarge-2020-01-16-b.svg"/>
+
+-->
+
+
+# References
+
+- \[1] Emery D. Berger, Kathryn S. McKinley, Robert D. Blumofe, and Paul R. Wilson.
+   _Hoard: A Scalable Memory Allocator for Multithreaded Applications_
+   the Ninth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS-IX). Cambridge, MA, November 2000.
+   [pdf](http://www.cs.utexas.edu/users/mckinley/papers/asplos-2000.pdf)
+
+- \[2] P. Larson and M. Krishnan. _Memory allocation for long-running server applications_.
+  In ISMM, Vancouver, B.C., Canada, 1998. [pdf](http://citeseer.ist.psu.edu/viewdoc/download?doi=10.1.1.45.1947&rep=rep1&type=pdf)
+
+- \[3] D. Grunwald, B. Zorn, and R. Henderson.
+  _Improving the cache locality of memory allocation_. In R. Cartwright, editor,
+  Proceedings of the Conference on Programming Language Design and Implementation, pages 177–186, New York, NY, USA, June 1993. [pdf](http://citeseer.ist.psu.edu/viewdoc/download?doi=10.1.1.43.6621&rep=rep1&type=pdf)
+
+- \[4] J. Barnes and P. Hut. _A hierarchical O(n*log(n)) force-calculation algorithm_. Nature, 324:446-449, 1986.
+
+- \[5] C. Lever, and D. Boreham. _Malloc() Performance in a Multithreaded Linux Environment._
+  In USENIX Annual Technical Conference, Freenix Session. San Diego, CA. Jun. 2000.
+  Available at <https://github.com/kuszmaul/SuperMalloc/tree/master/tests>
+
+- \[6] Timothy Crundal. _Reducing Active-False Sharing in TCMalloc_. 2016. CS16S1 project at the Australian National University. [pdf](http://courses.cecs.anu.edu.au/courses/CSPROJECTS/16S1/Reports/Timothy_Crundal_Report.pdf)
+
+- \[7] Alexey Kukanov, and Michael J Voss.
+   _The Foundations for Scalable Multi-Core Software in Intel Threading Building Blocks._
+   Intel Technology Journal 11 (4). 2007
+
+- \[8] Bobby Powers, David Tench, Emery D. Berger, and Andrew McGregor.
+ _Mesh: Compacting Memory Management for C/C++_
+ In Proceedings of the 40th ACM SIGPLAN Conference on Programming Language Design and Implementation (PLDI'19), June 2019, pages 333-–346.
+
+<!--
+- \[9] Paul Liétar, Theodore Butler, Sylvan Clebsch, Sophia Drossopoulou, Juliana Franco, Matthew J Parkinson,
+  Alex Shamis, Christoph M Wintersteiger, and David Chisnall.
+  _Snmalloc: A Message Passing Allocator._
+  In Proceedings of the 2019 ACM SIGPLAN International Symposium on Memory Management, 122–135. ACM. 2019.
+-->
+
+# Contributing
+
+This project welcomes contributions and suggestions.  Most contributions require you to agree to a
+Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us
+the rights to use your contribution. For details, visit https://cla.microsoft.com.
+
+When you submit a pull request, a CLA-bot will automatically determine whether you need to provide
+a CLA and decorate the PR appropriately (e.g., label, comment). Simply follow the instructions
+provided by the bot. You will only need to do this once across all repos using our CLA.
+
+
+# Older Release Notes
+
+* 2020-09-24, `v1.6.7`: stable release 1.6: using standard C atomics, passing tsan testing, improved
+  handling of failing to commit on Windows, add [`mi_process_info`](https://github.com/microsoft/mimalloc/blob/master/include/mimalloc.h#L156) api call.
+* 2020-08-06, `v1.6.4`: stable release 1.6: improved error recovery in low-memory situations,
+  support for IllumOS and Haiku, NUMA support for Vista/XP, improved NUMA detection for AMD Ryzen, ubsan support.
+* 2020-05-05, `v1.6.3`: stable release 1.6: improved behavior in out-of-memory situations, improved malloc zones on macOS,
+  build PIC static libraries by default, add option to abort on out-of-memory, line buffered statistics.
+* 2020-04-20, `v1.6.2`: stable release 1.6: fix compilation on Android, MingW, Raspberry, and Conda,
+  stability fix for Windows 7, fix multiple mimalloc instances in one executable, fix `strnlen` overload,
+  fix aligned debug padding.
+* 2020-02-17, `v1.6.1`: stable release 1.6: minor updates (build with clang-cl, fix alignment issue for small objects).
+* 2020-02-09, `v1.6.0`: stable release 1.6: fixed potential memory leak, improved overriding
+  and thread local support on FreeBSD, NetBSD, DragonFly, and macOSX. New byte-precise
+  heap block overflow detection in debug mode (besides the double-free detection and free-list
+  corruption detection). Add `nodiscard` attribute to most allocation functions.
+  Enable `MIMALLOC_PAGE_RESET` by default. New reclamation strategy for abandoned heap pages
+  for better memory footprint.
+* 2020-02-09, `v1.5.0`: stable release 1.5: improved free performance, small bug fixes.
+* 2020-01-22, `v1.4.0`: stable release 1.4: improved performance for delayed OS page reset,
+more eager concurrent free, addition of STL allocator, fixed potential memory leak.
+* 2020-01-15, `v1.3.0`: stable release 1.3: bug fixes, improved randomness and [stronger
+free list encoding](https://github.com/microsoft/mimalloc/blob/783e3377f79ee82af43a0793910a9f2d01ac7863/include/mimalloc-internal.h#L396) in secure mode.
+* 2019-12-22, `v1.2.2`: stable release 1.2: minor updates.
+* 2019-11-22, `v1.2.0`: stable release 1.2: bug fixes, improved secure mode (free list corruption checks, double free mitigation). Improved dynamic overriding on Windows.
+* 2019-10-07, `v1.1.0`: stable release 1.1.
+* 2019-09-01, `v1.0.8`: pre-release 8: more robust windows dynamic overriding, initial huge page support.
+* 2019-08-10, `v1.0.6`: pre-release 6: various performance improvements.

preload-mimalloc/mimalloc/src/alloc-aligned.c

@@ -0,0 +1,307 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/prim.h"  // mi_prim_get_default_heap
+
+#include <string.h>     // memset
+
+// ------------------------------------------------------
+// Aligned Allocation
+// ------------------------------------------------------
+
+// Fallback primitive aligned allocation -- split out for better codegen
+static mi_decl_noinline void* mi_heap_malloc_zero_aligned_at_fallback(mi_heap_t* const heap, const size_t size, const size_t alignment, const size_t offset, const bool zero) mi_attr_noexcept
+{
+  mi_assert_internal(size <= PTRDIFF_MAX);
+  mi_assert_internal(alignment != 0 && _mi_is_power_of_two(alignment));
+
+  const uintptr_t align_mask = alignment - 1;  // for any x, `(x & align_mask) == (x % alignment)`
+  const size_t padsize = size + MI_PADDING_SIZE;
+
+  // use regular allocation if it is guaranteed to fit the alignment constraints
+  if (offset==0 && alignment<=padsize && padsize<=MI_MAX_ALIGN_GUARANTEE && (padsize&align_mask)==0) {
+    void* p = _mi_heap_malloc_zero(heap, size, zero);
+    mi_assert_internal(p == NULL || ((uintptr_t)p % alignment) == 0);
+    return p;
+  }
+
+  void* p;
+  size_t oversize;
+  if mi_unlikely(alignment > MI_ALIGNMENT_MAX) {
+    // use OS allocation for very large alignment and allocate inside a huge page (dedicated segment with 1 page)
+    // This can support alignments >= MI_SEGMENT_SIZE by ensuring the object can be aligned at a point in the
+    // first (and single) page such that the segment info is `MI_SEGMENT_SIZE` bytes before it (so it can be found by aligning the pointer down)
+    if mi_unlikely(offset != 0) {
+      // todo: cannot support offset alignment for very large alignments yet
+      #if MI_DEBUG > 0
+      _mi_error_message(EOVERFLOW, "aligned allocation with a very large alignment cannot be used with an alignment offset (size %zu, alignment %zu, offset %zu)\n", size, alignment, offset);
+      #endif
+      return NULL;
+    }
+    oversize = (size <= MI_SMALL_SIZE_MAX ? MI_SMALL_SIZE_MAX + 1 /* ensure we use generic malloc path */ : size);
+    p = _mi_heap_malloc_zero_ex(heap, oversize, false, alignment); // the page block size should be large enough to align in the single huge page block
+    // zero afterwards as only the area from the aligned_p may be committed!
+    if (p == NULL) return NULL;    
+  }
+  else {
+    // otherwise over-allocate
+    oversize = size + alignment - 1;
+    p = _mi_heap_malloc_zero(heap, oversize, zero);
+    if (p == NULL) return NULL;
+  }
+
+  // .. and align within the allocation
+  const uintptr_t poffset = ((uintptr_t)p + offset) & align_mask;
+  const uintptr_t adjust  = (poffset == 0 ? 0 : alignment - poffset);
+  mi_assert_internal(adjust < alignment);
+  void* aligned_p = (void*)((uintptr_t)p + adjust);
+  if (aligned_p != p) {
+    mi_page_t* page = _mi_ptr_page(p);
+    mi_page_set_has_aligned(page, true);
+    _mi_padding_shrink(page, (mi_block_t*)p, adjust + size);
+  }
+  // todo: expand padding if overallocated ?
+
+  mi_assert_internal(mi_page_usable_block_size(_mi_ptr_page(p)) >= adjust + size);
+  mi_assert_internal(p == _mi_page_ptr_unalign(_mi_ptr_segment(aligned_p), _mi_ptr_page(aligned_p), aligned_p));
+  mi_assert_internal(((uintptr_t)aligned_p + offset) % alignment == 0);
+  mi_assert_internal(mi_usable_size(aligned_p)>=size);
+  mi_assert_internal(mi_usable_size(p) == mi_usable_size(aligned_p)+adjust);
+    
+  // now zero the block if needed
+  if (alignment > MI_ALIGNMENT_MAX) {
+    // for the tracker, on huge aligned allocations only from the start of the large block is defined
+    mi_track_mem_undefined(aligned_p, size);
+    if (zero) {
+      _mi_memzero(aligned_p, mi_usable_size(aligned_p));
+    }
+  }
+
+  if (p != aligned_p) {
+    mi_track_align(p,aligned_p,adjust,mi_usable_size(aligned_p));
+  }  
+  return aligned_p;
+}
+
+// Primitive aligned allocation
+static void* mi_heap_malloc_zero_aligned_at(mi_heap_t* const heap, const size_t size, const size_t alignment, const size_t offset, const bool zero) mi_attr_noexcept
+{
+  // note: we don't require `size > offset`, we just guarantee that the address at offset is aligned regardless of the allocated size.
+  mi_assert(alignment > 0);
+  if mi_unlikely(alignment == 0 || !_mi_is_power_of_two(alignment)) { // require power-of-two (see <https://en.cppreference.com/w/c/memory/aligned_alloc>)
+    #if MI_DEBUG > 0
+    _mi_error_message(EOVERFLOW, "aligned allocation requires the alignment to be a power-of-two (size %zu, alignment %zu)\n", size, alignment);
+    #endif
+    return NULL;
+  }
+  /*
+  if mi_unlikely(alignment > MI_ALIGNMENT_MAX) {  // we cannot align at a boundary larger than this (or otherwise we cannot find segment headers)
+    #if MI_DEBUG > 0
+    _mi_error_message(EOVERFLOW, "aligned allocation has a maximum alignment of %zu (size %zu, alignment %zu)\n", MI_ALIGNMENT_MAX, size, alignment);
+    #endif
+    return NULL;
+  }
+  */
+  if mi_unlikely(size > PTRDIFF_MAX) {          // we don't allocate more than PTRDIFF_MAX (see <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
+    #if MI_DEBUG > 0
+    _mi_error_message(EOVERFLOW, "aligned allocation request is too large (size %zu, alignment %zu)\n", size, alignment);
+    #endif
+    return NULL;
+  }
+  const uintptr_t align_mask = alignment-1;       // for any x, `(x & align_mask) == (x % alignment)`
+  const size_t padsize = size + MI_PADDING_SIZE;  // note: cannot overflow due to earlier size > PTRDIFF_MAX check
+
+  // try first if there happens to be a small block available with just the right alignment
+  if mi_likely(padsize <= MI_SMALL_SIZE_MAX && alignment <= padsize) {
+    mi_page_t* page = _mi_heap_get_free_small_page(heap, padsize);
+    const bool is_aligned = (((uintptr_t)page->free+offset) & align_mask)==0;
+    if mi_likely(page->free != NULL && is_aligned)
+    {
+      #if MI_STAT>1
+      mi_heap_stat_increase(heap, malloc, size);
+      #endif
+      void* p = _mi_page_malloc(heap, page, padsize, zero); // TODO: inline _mi_page_malloc
+      mi_assert_internal(p != NULL);
+      mi_assert_internal(((uintptr_t)p + offset) % alignment == 0);
+      mi_track_malloc(p,size,zero);
+      return p;
+    }
+  }
+  // fallback
+  return mi_heap_malloc_zero_aligned_at_fallback(heap, size, alignment, offset, zero);
+}
+
+
+// ------------------------------------------------------
+// Optimized mi_heap_malloc_aligned / mi_malloc_aligned
+// ------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_malloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_malloc_zero_aligned_at(heap, size, alignment, offset, false);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_malloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept {
+  #if !MI_PADDING
+  // without padding, any small sized allocation is naturally aligned (see also `_mi_segment_page_start`)
+  if (!_mi_is_power_of_two(alignment)) return NULL;
+  if mi_likely(_mi_is_power_of_two(size) && size >= alignment && size <= MI_SMALL_SIZE_MAX)
+  #else
+  // with padding, we can only guarantee this for fixed alignments
+  if mi_likely((alignment == sizeof(void*) || (alignment == MI_MAX_ALIGN_SIZE && size > (MI_MAX_ALIGN_SIZE/2)))
+                && size <= MI_SMALL_SIZE_MAX)
+  #endif
+  {
+    // fast path for common alignment and size
+    return mi_heap_malloc_small(heap, size);
+  }
+  else {
+    return mi_heap_malloc_aligned_at(heap, size, alignment, 0);
+  }
+}
+
+// ------------------------------------------------------
+// Aligned Allocation
+// ------------------------------------------------------
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_zalloc_aligned_at(mi_heap_t* heap, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_malloc_zero_aligned_at(heap, size, alignment, offset, true);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_zalloc_aligned(mi_heap_t* heap, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_zalloc_aligned_at(heap, size, alignment, 0);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_calloc_aligned_at(mi_heap_t* heap, size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(count, size, &total)) return NULL;
+  return mi_heap_zalloc_aligned_at(heap, total, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_heap_calloc_aligned(mi_heap_t* heap, size_t count, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_calloc_aligned_at(heap,count,size,alignment,0);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_malloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_malloc_aligned_at(mi_prim_get_default_heap(), size, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_malloc_aligned(size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_malloc_aligned(mi_prim_get_default_heap(), size, alignment);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_zalloc_aligned_at(size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_zalloc_aligned_at(mi_prim_get_default_heap(), size, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_zalloc_aligned(size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_zalloc_aligned(mi_prim_get_default_heap(), size, alignment);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_calloc_aligned_at(size_t count, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_calloc_aligned_at(mi_prim_get_default_heap(), count, size, alignment, offset);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_calloc_aligned(size_t count, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_calloc_aligned(mi_prim_get_default_heap(), count, size, alignment);
+}
+
+
+// ------------------------------------------------------
+// Aligned re-allocation
+// ------------------------------------------------------
+
+static void* mi_heap_realloc_zero_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset, bool zero) mi_attr_noexcept {
+  mi_assert(alignment > 0);
+  if (alignment <= sizeof(uintptr_t)) return _mi_heap_realloc_zero(heap,p,newsize,zero);
+  if (p == NULL) return mi_heap_malloc_zero_aligned_at(heap,newsize,alignment,offset,zero);
+  size_t size = mi_usable_size(p);
+  if (newsize <= size && newsize >= (size - (size / 2))
+      && (((uintptr_t)p + offset) % alignment) == 0) {
+    return p;  // reallocation still fits, is aligned and not more than 50% waste
+  }
+  else {
+    void* newp = mi_heap_malloc_aligned_at(heap,newsize,alignment,offset);
+    if (newp != NULL) {
+      if (zero && newsize > size) {
+        const mi_page_t* page = _mi_ptr_page(newp);
+        if (page->is_zero) {
+          // already zero initialized
+          mi_assert_expensive(mi_mem_is_zero(newp,newsize));
+        }
+        else {
+          // also set last word in the previous allocation to zero to ensure any padding is zero-initialized
+          size_t start = (size >= sizeof(intptr_t) ? size - sizeof(intptr_t) : 0);
+          memset((uint8_t*)newp + start, 0, newsize - start);
+        }
+      }
+      _mi_memcpy_aligned(newp, p, (newsize > size ? size : newsize));
+      mi_free(p); // only free if successful
+    }
+    return newp;
+  }
+}
+
+static void* mi_heap_realloc_zero_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, bool zero) mi_attr_noexcept {
+  mi_assert(alignment > 0);
+  if (alignment <= sizeof(uintptr_t)) return _mi_heap_realloc_zero(heap,p,newsize,zero);
+  size_t offset = ((uintptr_t)p % alignment); // use offset of previous allocation (p can be NULL)
+  return mi_heap_realloc_zero_aligned_at(heap,p,newsize,alignment,offset,zero);
+}
+
+mi_decl_nodiscard void* mi_heap_realloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned_at(heap,p,newsize,alignment,offset,false);
+}
+
+mi_decl_nodiscard void* mi_heap_realloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned(heap,p,newsize,alignment,false);
+}
+
+mi_decl_nodiscard void* mi_heap_rezalloc_aligned_at(mi_heap_t* heap, void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned_at(heap, p, newsize, alignment, offset, true);
+}
+
+mi_decl_nodiscard void* mi_heap_rezalloc_aligned(mi_heap_t* heap, void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_realloc_zero_aligned(heap, p, newsize, alignment, true);
+}
+
+mi_decl_nodiscard void* mi_heap_recalloc_aligned_at(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(newcount, size, &total)) return NULL;
+  return mi_heap_rezalloc_aligned_at(heap, p, total, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_heap_recalloc_aligned(mi_heap_t* heap, void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept {
+  size_t total;
+  if (mi_count_size_overflow(newcount, size, &total)) return NULL;
+  return mi_heap_rezalloc_aligned(heap, p, total, alignment);
+}
+
+mi_decl_nodiscard void* mi_realloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_realloc_aligned_at(mi_prim_get_default_heap(), p, newsize, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_realloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_realloc_aligned(mi_prim_get_default_heap(), p, newsize, alignment);
+}
+
+mi_decl_nodiscard void* mi_rezalloc_aligned_at(void* p, size_t newsize, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_rezalloc_aligned_at(mi_prim_get_default_heap(), p, newsize, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_rezalloc_aligned(void* p, size_t newsize, size_t alignment) mi_attr_noexcept {
+  return mi_heap_rezalloc_aligned(mi_prim_get_default_heap(), p, newsize, alignment);
+}
+
+mi_decl_nodiscard void* mi_recalloc_aligned_at(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept {
+  return mi_heap_recalloc_aligned_at(mi_prim_get_default_heap(), p, newcount, size, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_recalloc_aligned(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept {
+  return mi_heap_recalloc_aligned(mi_prim_get_default_heap(), p, newcount, size, alignment);
+}

preload-mimalloc/mimalloc/src/alloc-override.c

@@ -0,0 +1,295 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+#if !defined(MI_IN_ALLOC_C)
+#error "this file should be included from 'alloc.c' (so aliases can work)"
+#endif
+
+#if defined(MI_MALLOC_OVERRIDE) && defined(_WIN32) && !(defined(MI_SHARED_LIB) && defined(_DLL))
+#error "It is only possible to override "malloc" on Windows when building as a DLL (and linking the C runtime as a DLL)"
+#endif
+
+#if defined(MI_MALLOC_OVERRIDE) && !(defined(_WIN32))
+
+#if defined(__APPLE__)
+#include <AvailabilityMacros.h>
+mi_decl_externc void   vfree(void* p);
+mi_decl_externc size_t malloc_size(const void* p);
+mi_decl_externc size_t malloc_good_size(size_t size);
+#endif
+
+// helper definition for C override of C++ new
+typedef struct mi_nothrow_s { int _tag; } mi_nothrow_t;
+
+// ------------------------------------------------------
+// Override system malloc
+// ------------------------------------------------------
+
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__APPLE__) && !MI_TRACK_ENABLED
+  // gcc, clang: use aliasing to alias the exported function to one of our `mi_` functions
+  #if (defined(__GNUC__) && __GNUC__ >= 9)
+    #pragma GCC diagnostic ignored "-Wattributes"  // or we get warnings that nodiscard is ignored on a forward
+    #define MI_FORWARD(fun)      __attribute__((alias(#fun), used, visibility("default"), copy(fun)));
+  #else
+    #define MI_FORWARD(fun)      __attribute__((alias(#fun), used, visibility("default")));
+  #endif
+  #define MI_FORWARD1(fun,x)      MI_FORWARD(fun)
+  #define MI_FORWARD2(fun,x,y)    MI_FORWARD(fun)
+  #define MI_FORWARD3(fun,x,y,z)  MI_FORWARD(fun)
+  #define MI_FORWARD0(fun,x)      MI_FORWARD(fun)
+  #define MI_FORWARD02(fun,x,y)   MI_FORWARD(fun)
+#else
+  // otherwise use forwarding by calling our `mi_` function
+  #define MI_FORWARD1(fun,x)      { return fun(x); }
+  #define MI_FORWARD2(fun,x,y)    { return fun(x,y); }
+  #define MI_FORWARD3(fun,x,y,z)  { return fun(x,y,z); }
+  #define MI_FORWARD0(fun,x)      { fun(x); }
+  #define MI_FORWARD02(fun,x,y)   { fun(x,y); }
+#endif
+
+
+#if defined(__APPLE__) && defined(MI_SHARED_LIB_EXPORT) && defined(MI_OSX_INTERPOSE)
+  // define MI_OSX_IS_INTERPOSED as we should not provide forwarding definitions for
+  // functions that are interposed (or the interposing does not work)
+  #define MI_OSX_IS_INTERPOSED
+
+  mi_decl_externc size_t mi_malloc_size_checked(void *p) {
+    if (!mi_is_in_heap_region(p)) return 0;
+    return mi_usable_size(p);
+  }
+
+  // use interposing so `DYLD_INSERT_LIBRARIES` works without `DYLD_FORCE_FLAT_NAMESPACE=1`
+  // See: <https://books.google.com/books?id=K8vUkpOXhN4C&pg=PA73>
+  struct mi_interpose_s {
+    const void* replacement;
+    const void* target;
+  };
+  #define MI_INTERPOSE_FUN(oldfun,newfun) { (const void*)&newfun, (const void*)&oldfun }
+  #define MI_INTERPOSE_MI(fun)            MI_INTERPOSE_FUN(fun,mi_##fun)
+
+  __attribute__((used)) static struct mi_interpose_s _mi_interposes[]  __attribute__((section("__DATA, __interpose"))) =
+  {
+    MI_INTERPOSE_MI(malloc),
+    MI_INTERPOSE_MI(calloc),
+    MI_INTERPOSE_MI(realloc),
+    MI_INTERPOSE_MI(strdup),
+    MI_INTERPOSE_MI(strndup),
+    MI_INTERPOSE_MI(realpath),
+    MI_INTERPOSE_MI(posix_memalign),
+    MI_INTERPOSE_MI(reallocf),
+    MI_INTERPOSE_MI(valloc),
+    MI_INTERPOSE_FUN(malloc_size,mi_malloc_size_checked),
+    MI_INTERPOSE_MI(malloc_good_size),
+    #if defined(MAC_OS_X_VERSION_10_15) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_15
+    MI_INTERPOSE_MI(aligned_alloc),
+    #endif
+    #ifdef MI_OSX_ZONE
+    // we interpose malloc_default_zone in alloc-override-osx.c so we can use mi_free safely
+    MI_INTERPOSE_MI(free),
+    MI_INTERPOSE_FUN(vfree,mi_free),
+    #else
+    // sometimes code allocates from default zone but deallocates using plain free :-( (like NxHashResizeToCapacity <https://github.com/nneonneo/osx-10.9-opensource/blob/master/objc4-551.1/runtime/hashtable2.mm>)
+    MI_INTERPOSE_FUN(free,mi_cfree), // use safe free that checks if pointers are from us
+    MI_INTERPOSE_FUN(vfree,mi_cfree),
+    #endif
+  };
+
+  #ifdef __cplusplus
+  extern "C" {
+  #endif
+  void  _ZdlPv(void* p);   // delete
+  void  _ZdaPv(void* p);   // delete[]
+  void  _ZdlPvm(void* p, size_t n);  // delete
+  void  _ZdaPvm(void* p, size_t n);  // delete[]
+  void* _Znwm(size_t n);  // new
+  void* _Znam(size_t n);  // new[]
+  void* _ZnwmRKSt9nothrow_t(size_t n, mi_nothrow_t tag); // new nothrow
+  void* _ZnamRKSt9nothrow_t(size_t n, mi_nothrow_t tag); // new[] nothrow
+  #ifdef __cplusplus
+  }
+  #endif
+  __attribute__((used)) static struct mi_interpose_s _mi_cxx_interposes[]  __attribute__((section("__DATA, __interpose"))) =
+  {
+    MI_INTERPOSE_FUN(_ZdlPv,mi_free),
+    MI_INTERPOSE_FUN(_ZdaPv,mi_free),
+    MI_INTERPOSE_FUN(_ZdlPvm,mi_free_size),
+    MI_INTERPOSE_FUN(_ZdaPvm,mi_free_size),
+    MI_INTERPOSE_FUN(_Znwm,mi_new),
+    MI_INTERPOSE_FUN(_Znam,mi_new),
+    MI_INTERPOSE_FUN(_ZnwmRKSt9nothrow_t,mi_new_nothrow),
+    MI_INTERPOSE_FUN(_ZnamRKSt9nothrow_t,mi_new_nothrow),
+  };
+
+#elif defined(_MSC_VER)
+  // cannot override malloc unless using a dll.
+  // we just override new/delete which does work in a static library.
+#else
+  // On all other systems forward to our API
+  mi_decl_export void* malloc(size_t size)              MI_FORWARD1(mi_malloc, size)
+  mi_decl_export void* calloc(size_t size, size_t n)    MI_FORWARD2(mi_calloc, size, n)
+  mi_decl_export void* realloc(void* p, size_t newsize) MI_FORWARD2(mi_realloc, p, newsize)
+  mi_decl_export void  free(void* p)                    MI_FORWARD0(mi_free, p)
+#endif
+
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__APPLE__)
+#pragma GCC visibility push(default)
+#endif
+
+// ------------------------------------------------------
+// Override new/delete
+// This is not really necessary as they usually call
+// malloc/free anyway, but it improves performance.
+// ------------------------------------------------------
+#ifdef __cplusplus
+  // ------------------------------------------------------
+  // With a C++ compiler we override the new/delete operators.
+  // see <https://en.cppreference.com/w/cpp/memory/new/operator_new>
+  // ------------------------------------------------------
+  #include <new>
+
+  #ifndef MI_OSX_IS_INTERPOSED
+    void operator delete(void* p) noexcept              MI_FORWARD0(mi_free,p)
+    void operator delete[](void* p) noexcept            MI_FORWARD0(mi_free,p)
+
+    void* operator new(std::size_t n) noexcept(false)   MI_FORWARD1(mi_new,n)
+    void* operator new[](std::size_t n) noexcept(false) MI_FORWARD1(mi_new,n)
+
+    void* operator new  (std::size_t n, const std::nothrow_t& tag) noexcept { MI_UNUSED(tag); return mi_new_nothrow(n); }
+    void* operator new[](std::size_t n, const std::nothrow_t& tag) noexcept { MI_UNUSED(tag); return mi_new_nothrow(n); }
+
+    #if (__cplusplus >= 201402L || _MSC_VER >= 1916)
+    void operator delete  (void* p, std::size_t n) noexcept MI_FORWARD02(mi_free_size,p,n)
+    void operator delete[](void* p, std::size_t n) noexcept MI_FORWARD02(mi_free_size,p,n)
+    #endif
+  #endif
+
+  #if (__cplusplus > 201402L && defined(__cpp_aligned_new)) && (!defined(__GNUC__) || (__GNUC__ > 5))
+  void operator delete  (void* p, std::align_val_t al) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+  void operator delete[](void* p, std::align_val_t al) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+  void operator delete  (void* p, std::size_t n, std::align_val_t al) noexcept { mi_free_size_aligned(p, n, static_cast<size_t>(al)); };
+  void operator delete[](void* p, std::size_t n, std::align_val_t al) noexcept { mi_free_size_aligned(p, n, static_cast<size_t>(al)); };
+  void operator delete  (void* p, std::align_val_t al, const std::nothrow_t&) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+  void operator delete[](void* p, std::align_val_t al, const std::nothrow_t&) noexcept { mi_free_aligned(p, static_cast<size_t>(al)); }
+
+  void* operator new( std::size_t n, std::align_val_t al)   noexcept(false) { return mi_new_aligned(n, static_cast<size_t>(al)); }
+  void* operator new[]( std::size_t n, std::align_val_t al) noexcept(false) { return mi_new_aligned(n, static_cast<size_t>(al)); }
+  void* operator new  (std::size_t n, std::align_val_t al, const std::nothrow_t&) noexcept { return mi_new_aligned_nothrow(n, static_cast<size_t>(al)); }
+  void* operator new[](std::size_t n, std::align_val_t al, const std::nothrow_t&) noexcept { return mi_new_aligned_nothrow(n, static_cast<size_t>(al)); }
+  #endif
+
+#elif (defined(__GNUC__) || defined(__clang__))
+  // ------------------------------------------------------
+  // Override by defining the mangled C++ names of the operators (as
+  // used by GCC and CLang).
+  // See <https://itanium-cxx-abi.github.io/cxx-abi/abi.html#mangling>
+  // ------------------------------------------------------
+
+  void _ZdlPv(void* p)            MI_FORWARD0(mi_free,p) // delete
+  void _ZdaPv(void* p)            MI_FORWARD0(mi_free,p) // delete[]
+  void _ZdlPvm(void* p, size_t n) MI_FORWARD02(mi_free_size,p,n)
+  void _ZdaPvm(void* p, size_t n) MI_FORWARD02(mi_free_size,p,n)
+  void _ZdlPvSt11align_val_t(void* p, size_t al)            { mi_free_aligned(p,al); }
+  void _ZdaPvSt11align_val_t(void* p, size_t al)            { mi_free_aligned(p,al); }
+  void _ZdlPvmSt11align_val_t(void* p, size_t n, size_t al) { mi_free_size_aligned(p,n,al); }
+  void _ZdaPvmSt11align_val_t(void* p, size_t n, size_t al) { mi_free_size_aligned(p,n,al); }
+
+  #if (MI_INTPTR_SIZE==8)
+    void* _Znwm(size_t n)                             MI_FORWARD1(mi_new,n)  // new 64-bit
+    void* _Znam(size_t n)                             MI_FORWARD1(mi_new,n)  // new[] 64-bit
+    void* _ZnwmRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
+    void* _ZnamRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
+    void* _ZnwmSt11align_val_t(size_t n, size_t al)   MI_FORWARD2(mi_new_aligned, n, al)
+    void* _ZnamSt11align_val_t(size_t n, size_t al)   MI_FORWARD2(mi_new_aligned, n, al)
+    void* _ZnwmSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
+    void* _ZnamSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
+  #elif (MI_INTPTR_SIZE==4)
+    void* _Znwj(size_t n)                             MI_FORWARD1(mi_new,n)  // new 64-bit
+    void* _Znaj(size_t n)                             MI_FORWARD1(mi_new,n)  // new[] 64-bit
+    void* _ZnwjRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
+    void* _ZnajRKSt9nothrow_t(size_t n, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_nothrow(n); }
+    void* _ZnwjSt11align_val_t(size_t n, size_t al)   MI_FORWARD2(mi_new_aligned, n, al)
+    void* _ZnajSt11align_val_t(size_t n, size_t al)   MI_FORWARD2(mi_new_aligned, n, al)
+    void* _ZnwjSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
+    void* _ZnajSt11align_val_tRKSt9nothrow_t(size_t n, size_t al, mi_nothrow_t tag) { MI_UNUSED(tag); return mi_new_aligned_nothrow(n,al); }
+  #else
+    #error "define overloads for new/delete for this platform (just for performance, can be skipped)"
+  #endif
+#endif // __cplusplus
+
+// ------------------------------------------------------
+// Further Posix & Unix functions definitions
+// ------------------------------------------------------
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#ifndef MI_OSX_IS_INTERPOSED
+  // Forward Posix/Unix calls as well
+  void*  reallocf(void* p, size_t newsize) MI_FORWARD2(mi_reallocf,p,newsize)
+  size_t malloc_size(const void* p)        MI_FORWARD1(mi_usable_size,p)
+  #if !defined(__ANDROID__) && !defined(__FreeBSD__)
+  size_t malloc_usable_size(void *p)       MI_FORWARD1(mi_usable_size,p)
+  #else
+  size_t malloc_usable_size(const void *p) MI_FORWARD1(mi_usable_size,p)
+  #endif
+
+  // No forwarding here due to aliasing/name mangling issues
+  void*  valloc(size_t size)               { return mi_valloc(size); }
+  void   vfree(void* p)                    { mi_free(p); }
+  size_t malloc_good_size(size_t size)     { return mi_malloc_good_size(size); }
+  int    posix_memalign(void** p, size_t alignment, size_t size) { return mi_posix_memalign(p, alignment, size); }
+
+  // `aligned_alloc` is only available when __USE_ISOC11 is defined.
+  // Note: Conda has a custom glibc where `aligned_alloc` is declared `static inline` and we cannot
+  // override it, but both _ISOC11_SOURCE and __USE_ISOC11 are undefined in Conda GCC7 or GCC9.
+  // Fortunately, in the case where `aligned_alloc` is declared as `static inline` it
+  // uses internally `memalign`, `posix_memalign`, or `_aligned_malloc` so we  can avoid overriding it ourselves.
+  #if __USE_ISOC11
+  void* aligned_alloc(size_t alignment, size_t size) { return mi_aligned_alloc(alignment, size); }
+  #endif
+#endif
+
+// no forwarding here due to aliasing/name mangling issues
+void  cfree(void* p)                                    { mi_free(p); }
+void* pvalloc(size_t size)                              { return mi_pvalloc(size); }
+void* reallocarray(void* p, size_t count, size_t size)  { return mi_reallocarray(p, count, size); }
+int   reallocarr(void* p, size_t count, size_t size)    { return mi_reallocarr(p, count, size); }
+void* memalign(size_t alignment, size_t size)           { return mi_memalign(alignment, size); }
+void* _aligned_malloc(size_t alignment, size_t size)    { return mi_aligned_alloc(alignment, size); }
+
+#if defined(__wasi__)
+  // forward __libc interface (see PR #667)
+  void* __libc_malloc(size_t size)                      MI_FORWARD1(mi_malloc, size)
+  void* __libc_calloc(size_t count, size_t size)        MI_FORWARD2(mi_calloc, count, size)
+  void* __libc_realloc(void* p, size_t size)            MI_FORWARD2(mi_realloc, p, size)
+  void  __libc_free(void* p)                            MI_FORWARD0(mi_free, p)
+  void* __libc_memalign(size_t alignment, size_t size)  { return mi_memalign(alignment, size); }
+
+#elif defined(__GLIBC__) && defined(__linux__)
+  // forward __libc interface (needed for glibc-based Linux distributions)
+  void* __libc_malloc(size_t size)                      MI_FORWARD1(mi_malloc,size)
+  void* __libc_calloc(size_t count, size_t size)        MI_FORWARD2(mi_calloc,count,size)
+  void* __libc_realloc(void* p, size_t size)            MI_FORWARD2(mi_realloc,p,size)
+  void  __libc_free(void* p)                            MI_FORWARD0(mi_free,p)
+  void  __libc_cfree(void* p)                           MI_FORWARD0(mi_free,p)
+
+  void* __libc_valloc(size_t size)                      { return mi_valloc(size); }
+  void* __libc_pvalloc(size_t size)                     { return mi_pvalloc(size); }
+  void* __libc_memalign(size_t alignment, size_t size)  { return mi_memalign(alignment,size); }
+  int   __posix_memalign(void** p, size_t alignment, size_t size) { return mi_posix_memalign(p,alignment,size); }
+#endif
+
+#ifdef __cplusplus
+}
+#endif
+
+#if (defined(__GNUC__) || defined(__clang__)) && !defined(__APPLE__)
+#pragma GCC visibility pop
+#endif
+
+#endif // MI_MALLOC_OVERRIDE && !_WIN32

preload-mimalloc/mimalloc/src/alloc-posix.c

@@ -0,0 +1,184 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+// ------------------------------------------------------------------------
+// mi prefixed publi definitions of various Posix, Unix, and C++ functions
+// for convenience and used when overriding these functions.
+// ------------------------------------------------------------------------
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+
+// ------------------------------------------------------
+// Posix & Unix functions definitions
+// ------------------------------------------------------
+
+#include <errno.h>
+#include <string.h>  // memset
+#include <stdlib.h>  // getenv
+
+#ifdef _MSC_VER
+#pragma warning(disable:4996)  // getenv _wgetenv
+#endif
+
+#ifndef EINVAL
+#define EINVAL 22
+#endif
+#ifndef ENOMEM
+#define ENOMEM 12
+#endif
+
+
+mi_decl_nodiscard size_t mi_malloc_size(const void* p) mi_attr_noexcept {
+  // if (!mi_is_in_heap_region(p)) return 0;
+  return mi_usable_size(p);
+}
+
+mi_decl_nodiscard size_t mi_malloc_usable_size(const void *p) mi_attr_noexcept {
+  // if (!mi_is_in_heap_region(p)) return 0;
+  return mi_usable_size(p);
+}
+
+mi_decl_nodiscard size_t mi_malloc_good_size(size_t size) mi_attr_noexcept {
+  return mi_good_size(size);
+}
+
+void mi_cfree(void* p) mi_attr_noexcept {
+  if (mi_is_in_heap_region(p)) {
+    mi_free(p);
+  }
+}
+
+int mi_posix_memalign(void** p, size_t alignment, size_t size) mi_attr_noexcept {
+  // Note: The spec dictates we should not modify `*p` on an error. (issue#27)
+  // <http://man7.org/linux/man-pages/man3/posix_memalign.3.html>
+  if (p == NULL) return EINVAL;
+  if (alignment % sizeof(void*) != 0) return EINVAL;                   // natural alignment
+  if (alignment==0 || !_mi_is_power_of_two(alignment)) return EINVAL;  // not a power of 2
+  void* q = mi_malloc_aligned(size, alignment);
+  if (q==NULL && size != 0) return ENOMEM;
+  mi_assert_internal(((uintptr_t)q % alignment) == 0);
+  *p = q;
+  return 0;
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_memalign(size_t alignment, size_t size) mi_attr_noexcept {
+  void* p = mi_malloc_aligned(size, alignment);
+  mi_assert_internal(((uintptr_t)p % alignment) == 0);
+  return p;
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_valloc(size_t size) mi_attr_noexcept {
+  return mi_memalign( _mi_os_page_size(), size );
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_pvalloc(size_t size) mi_attr_noexcept {
+  size_t psize = _mi_os_page_size();
+  if (size >= SIZE_MAX - psize) return NULL; // overflow
+  size_t asize = _mi_align_up(size, psize);
+  return mi_malloc_aligned(asize, psize);
+}
+
+mi_decl_nodiscard mi_decl_restrict void* mi_aligned_alloc(size_t alignment, size_t size) mi_attr_noexcept {
+  // C11 requires the size to be an integral multiple of the alignment, see <https://en.cppreference.com/w/c/memory/aligned_alloc>.
+  // unfortunately, it turns out quite some programs pass a size that is not an integral multiple so skip this check..
+  /* if mi_unlikely((size & (alignment - 1)) != 0) { // C11 requires alignment>0 && integral multiple, see <https://en.cppreference.com/w/c/memory/aligned_alloc>
+      #if MI_DEBUG > 0
+      _mi_error_message(EOVERFLOW, "(mi_)aligned_alloc requires the size to be an integral multiple of the alignment (size %zu, alignment %zu)\n", size, alignment);
+      #endif
+      return NULL;
+    }
+  */
+  // C11 also requires alignment to be a power-of-two (and > 0) which is checked in mi_malloc_aligned
+  void* p = mi_malloc_aligned(size, alignment);
+  mi_assert_internal(((uintptr_t)p % alignment) == 0);
+  return p;
+}
+
+mi_decl_nodiscard void* mi_reallocarray( void* p, size_t count, size_t size ) mi_attr_noexcept {  // BSD
+  void* newp = mi_reallocn(p,count,size);
+  if (newp==NULL) { errno = ENOMEM; }
+  return newp;
+}
+
+mi_decl_nodiscard int mi_reallocarr( void* p, size_t count, size_t size ) mi_attr_noexcept { // NetBSD
+  mi_assert(p != NULL);
+  if (p == NULL) {
+    errno = EINVAL;
+    return EINVAL;
+  }
+  void** op = (void**)p;
+  void* newp = mi_reallocarray(*op, count, size);
+  if mi_unlikely(newp == NULL) { return errno; }
+  *op = newp;
+  return 0;
+}
+
+void* mi__expand(void* p, size_t newsize) mi_attr_noexcept {  // Microsoft
+  void* res = mi_expand(p, newsize);
+  if (res == NULL) { errno = ENOMEM; }
+  return res;
+}
+
+mi_decl_nodiscard mi_decl_restrict unsigned short* mi_wcsdup(const unsigned short* s) mi_attr_noexcept {
+  if (s==NULL) return NULL;
+  size_t len;
+  for(len = 0; s[len] != 0; len++) { }
+  size_t size = (len+1)*sizeof(unsigned short);
+  unsigned short* p = (unsigned short*)mi_malloc(size);
+  if (p != NULL) {
+    _mi_memcpy(p,s,size);
+  }
+  return p;
+}
+
+mi_decl_nodiscard mi_decl_restrict unsigned char* mi_mbsdup(const unsigned char* s)  mi_attr_noexcept {
+  return (unsigned char*)mi_strdup((const char*)s);
+}
+
+int mi_dupenv_s(char** buf, size_t* size, const char* name) mi_attr_noexcept {
+  if (buf==NULL || name==NULL) return EINVAL;
+  if (size != NULL) *size = 0;
+  char* p = getenv(name);        // mscver warning 4996
+  if (p==NULL) {
+    *buf = NULL;
+  }
+  else {
+    *buf = mi_strdup(p);
+    if (*buf==NULL) return ENOMEM;
+    if (size != NULL) *size = _mi_strlen(p);
+  }
+  return 0;
+}
+
+int mi_wdupenv_s(unsigned short** buf, size_t* size, const unsigned short* name) mi_attr_noexcept {
+  if (buf==NULL || name==NULL) return EINVAL;
+  if (size != NULL) *size = 0;
+#if !defined(_WIN32) || (defined(WINAPI_FAMILY) && (WINAPI_FAMILY != WINAPI_FAMILY_DESKTOP_APP))
+  // not supported
+  *buf = NULL;
+  return EINVAL;
+#else
+  unsigned short* p = (unsigned short*)_wgetenv((const wchar_t*)name);  // msvc warning 4996
+  if (p==NULL) {
+    *buf = NULL;
+  }
+  else {
+    *buf = mi_wcsdup(p);
+    if (*buf==NULL) return ENOMEM;
+    if (size != NULL) *size = wcslen((const wchar_t*)p);
+  }
+  return 0;
+#endif
+}
+
+mi_decl_nodiscard void* mi_aligned_offset_recalloc(void* p, size_t newcount, size_t size, size_t alignment, size_t offset) mi_attr_noexcept { // Microsoft
+  return mi_recalloc_aligned_at(p, newcount, size, alignment, offset);
+}
+
+mi_decl_nodiscard void* mi_aligned_recalloc(void* p, size_t newcount, size_t size, size_t alignment) mi_attr_noexcept { // Microsoft
+  return mi_recalloc_aligned(p, newcount, size, alignment);
+}

preload-mimalloc/mimalloc/src/arena.c

@@ -0,0 +1,526 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2022, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+"Arenas" are fixed area's of OS memory from which we can allocate
+large blocks (>= MI_ARENA_MIN_BLOCK_SIZE, 4MiB).
+In contrast to the rest of mimalloc, the arenas are shared between
+threads and need to be accessed using atomic operations.
+
+Arenas are used to for huge OS page (1GiB) reservations or for reserving
+OS memory upfront which can be improve performance or is sometimes needed
+on embedded devices. We can also employ this with WASI or `sbrk` systems 
+to reserve large arenas upfront and be able to reuse the memory more effectively.
+
+The arena allocation needs to be thread safe and we use an atomic bitmap to allocate.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+#include <string.h>  // memset
+#include <errno.h> // ENOMEM
+
+#include "bitmap.h"  // atomic bitmap
+
+/* -----------------------------------------------------------
+  Arena allocation
+----------------------------------------------------------- */
+
+// Block info: bit 0 contains the `in_use` bit, the upper bits the
+// size in count of arena blocks.
+typedef uintptr_t mi_block_info_t;
+#define MI_ARENA_BLOCK_SIZE   (MI_SEGMENT_SIZE)        // 64MiB  (must be at least MI_SEGMENT_ALIGN)
+#define MI_ARENA_MIN_OBJ_SIZE (MI_ARENA_BLOCK_SIZE/2)  // 32MiB
+#define MI_MAX_ARENAS         (64)                     // not more than 126 (since we use 7 bits in the memid and an arena index + 1)
+
+// A memory arena descriptor
+typedef struct mi_arena_s {
+  mi_arena_id_t id;                       // arena id; 0 for non-specific
+  bool     exclusive;                     // only allow allocations if specifically for this arena
+  _Atomic(uint8_t*) start;                // the start of the memory area
+  size_t   block_count;                   // size of the area in arena blocks (of `MI_ARENA_BLOCK_SIZE`)
+  size_t   field_count;                   // number of bitmap fields (where `field_count * MI_BITMAP_FIELD_BITS >= block_count`)
+  int      numa_node;                     // associated NUMA node
+  bool     is_zero_init;                  // is the arena zero initialized?
+  bool     allow_decommit;                // is decommit allowed? if true, is_large should be false and blocks_committed != NULL
+  bool     is_large;                      // large- or huge OS pages (always committed)
+  _Atomic(size_t) search_idx;             // optimization to start the search for free blocks
+  mi_bitmap_field_t* blocks_dirty;        // are the blocks potentially non-zero?
+  mi_bitmap_field_t* blocks_committed;    // are the blocks committed? (can be NULL for memory that cannot be decommitted)
+  mi_bitmap_field_t  blocks_inuse[1];     // in-place bitmap of in-use blocks (of size `field_count`)
+} mi_arena_t;
+
+
+// The available arenas
+static mi_decl_cache_align _Atomic(mi_arena_t*) mi_arenas[MI_MAX_ARENAS];
+static mi_decl_cache_align _Atomic(size_t)      mi_arena_count; // = 0
+
+
+/* -----------------------------------------------------------
+  Arena id's
+  0 is used for non-arena's (like OS memory)
+  id = arena_index + 1
+----------------------------------------------------------- */
+
+static size_t mi_arena_id_index(mi_arena_id_t id) {
+  return (size_t)(id <= 0 ? MI_MAX_ARENAS : id - 1);
+}
+
+static mi_arena_id_t mi_arena_id_create(size_t arena_index) {
+  mi_assert_internal(arena_index < MI_MAX_ARENAS);
+  mi_assert_internal(MI_MAX_ARENAS <= 126);
+  int id = (int)arena_index + 1;
+  mi_assert_internal(id >= 1 && id <= 127);
+  return id;
+}
+
+mi_arena_id_t _mi_arena_id_none(void) {
+  return 0;
+}
+
+static bool mi_arena_id_is_suitable(mi_arena_id_t arena_id, bool arena_is_exclusive, mi_arena_id_t req_arena_id) {
+  return ((!arena_is_exclusive && req_arena_id == _mi_arena_id_none()) ||
+          (arena_id == req_arena_id));
+}
+
+
+/* -----------------------------------------------------------
+  Arena allocations get a memory id where the lower 8 bits are
+  the arena id, and the upper bits the block index.
+----------------------------------------------------------- */
+
+// Use `0` as a special id for direct OS allocated memory.
+#define MI_MEMID_OS   0
+
+static size_t mi_arena_memid_create(mi_arena_id_t id, bool exclusive, mi_bitmap_index_t bitmap_index) {
+  mi_assert_internal(((bitmap_index << 8) >> 8) == bitmap_index); // no overflow?
+  mi_assert_internal(id >= 0 && id <= 0x7F);
+  return ((bitmap_index << 8) | ((uint8_t)id & 0x7F) | (exclusive ? 0x80 : 0));
+}
+
+static bool mi_arena_memid_indices(size_t arena_memid, size_t* arena_index, mi_bitmap_index_t* bitmap_index) {
+  *bitmap_index = (arena_memid >> 8);
+  mi_arena_id_t id = (int)(arena_memid & 0x7F);
+  *arena_index = mi_arena_id_index(id);
+  return ((arena_memid & 0x80) != 0);
+}
+
+bool _mi_arena_memid_is_suitable(size_t arena_memid, mi_arena_id_t request_arena_id) {
+  mi_arena_id_t id = (int)(arena_memid & 0x7F);
+  bool exclusive = ((arena_memid & 0x80) != 0);
+  return mi_arena_id_is_suitable(id, exclusive, request_arena_id);
+}
+
+bool _mi_arena_is_os_allocated(size_t arena_memid) {
+  return (arena_memid == MI_MEMID_OS);
+}
+
+static size_t mi_block_count_of_size(size_t size) {
+  return _mi_divide_up(size, MI_ARENA_BLOCK_SIZE);
+}
+
+/* -----------------------------------------------------------
+  Thread safe allocation in an arena
+----------------------------------------------------------- */
+static bool mi_arena_alloc(mi_arena_t* arena, size_t blocks, mi_bitmap_index_t* bitmap_idx)
+{
+  size_t idx = 0; // mi_atomic_load_relaxed(&arena->search_idx);  // start from last search; ok to be relaxed as the exact start does not matter
+  if (_mi_bitmap_try_find_from_claim_across(arena->blocks_inuse, arena->field_count, idx, blocks, bitmap_idx)) {
+    mi_atomic_store_relaxed(&arena->search_idx, mi_bitmap_index_field(*bitmap_idx));  // start search from found location next time around
+    return true;
+  };
+  return false;
+}
+
+
+/* -----------------------------------------------------------
+  Arena Allocation
+----------------------------------------------------------- */
+
+static mi_decl_noinline void* mi_arena_alloc_from(mi_arena_t* arena, size_t arena_index, size_t needed_bcount,
+                                                   bool* commit, bool* large, bool* is_pinned, bool* is_zero, 
+                                                   mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld)
+{
+  MI_UNUSED(arena_index);
+  mi_assert_internal(mi_arena_id_index(arena->id) == arena_index);
+  if (!mi_arena_id_is_suitable(arena->id, arena->exclusive, req_arena_id)) return NULL;
+
+  mi_bitmap_index_t bitmap_index;
+  if (!mi_arena_alloc(arena, needed_bcount, &bitmap_index)) return NULL;
+
+  // claimed it! set the dirty bits (todo: no need for an atomic op here?)
+  void* p    = arena->start + (mi_bitmap_index_bit(bitmap_index)*MI_ARENA_BLOCK_SIZE);
+  *memid     = mi_arena_memid_create(arena->id, arena->exclusive, bitmap_index);
+  *is_zero   = _mi_bitmap_claim_across(arena->blocks_dirty, arena->field_count, needed_bcount, bitmap_index, NULL);
+  *large     = arena->is_large;
+  *is_pinned = (arena->is_large || !arena->allow_decommit);
+  if (arena->blocks_committed == NULL) {
+    // always committed
+    *commit = true;
+  }
+  else if (*commit) {
+    // arena not committed as a whole, but commit requested: ensure commit now
+    bool any_uncommitted;
+    _mi_bitmap_claim_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index, &any_uncommitted);
+    if (any_uncommitted) {
+      bool commit_zero;
+      _mi_os_commit(p, needed_bcount * MI_ARENA_BLOCK_SIZE, &commit_zero, tld->stats);
+      if (commit_zero) *is_zero = true;
+    }
+  }
+  else {
+    // no need to commit, but check if already fully committed
+    *commit = _mi_bitmap_is_claimed_across(arena->blocks_committed, arena->field_count, needed_bcount, bitmap_index);
+  }
+  return p;
+}
+
+// allocate from an arena with fallback to the OS
+static mi_decl_noinline void* mi_arena_allocate(int numa_node, size_t size, size_t alignment, bool* commit, bool* large,
+                                                bool* is_pinned, bool* is_zero,
+                                                mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld )
+{
+  MI_UNUSED(alignment);
+  mi_assert_internal(alignment <= MI_SEGMENT_ALIGN);
+  const size_t max_arena = mi_atomic_load_relaxed(&mi_arena_count);
+  const size_t bcount = mi_block_count_of_size(size);
+  if mi_likely(max_arena == 0) return NULL;
+  mi_assert_internal(size <= bcount * MI_ARENA_BLOCK_SIZE);
+
+  size_t arena_index = mi_arena_id_index(req_arena_id);
+  if (arena_index < MI_MAX_ARENAS) {
+    // try a specific arena if requested
+    mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t, &mi_arenas[arena_index]);
+    if ((arena != NULL) &&
+        (arena->numa_node < 0 || arena->numa_node == numa_node) && // numa local?
+        (*large || !arena->is_large)) // large OS pages allowed, or arena is not large OS pages
+    {
+      void* p = mi_arena_alloc_from(arena, arena_index, bcount, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
+      mi_assert_internal((uintptr_t)p % alignment == 0);
+      if (p != NULL) return p;
+    }
+  }
+  else {
+    // try numa affine allocation
+    for (size_t i = 0; i < max_arena; i++) {
+      mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t, &mi_arenas[i]);
+      if (arena == NULL) break; // end reached
+      if ((arena->numa_node < 0 || arena->numa_node == numa_node) && // numa local?
+          (*large || !arena->is_large)) // large OS pages allowed, or arena is not large OS pages
+      {
+        void* p = mi_arena_alloc_from(arena, i, bcount, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
+        mi_assert_internal((uintptr_t)p % alignment == 0);
+        if (p != NULL) return p;
+      }
+    }
+
+    // try from another numa node instead..
+    for (size_t i = 0; i < max_arena; i++) {
+      mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t, &mi_arenas[i]);
+      if (arena == NULL) break; // end reached
+      if ((arena->numa_node >= 0 && arena->numa_node != numa_node) && // not numa local!
+          (*large || !arena->is_large)) // large OS pages allowed, or arena is not large OS pages
+      {
+        void* p = mi_arena_alloc_from(arena, i, bcount, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
+        mi_assert_internal((uintptr_t)p % alignment == 0);
+        if (p != NULL) return p;
+      }
+    }
+  }
+  return NULL;
+}
+
+void* _mi_arena_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool* commit, bool* large, bool* is_pinned, bool* is_zero,
+                              mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld)
+{
+  mi_assert_internal(commit != NULL && is_pinned != NULL && is_zero != NULL && memid != NULL && tld != NULL);
+  mi_assert_internal(size > 0);
+  *memid   = MI_MEMID_OS;
+  *is_zero = false;
+  *is_pinned = false;
+
+  bool default_large = false;
+  if (large == NULL) large = &default_large;   // ensure `large != NULL`
+  const int numa_node = _mi_os_numa_node(tld); // current numa node
+
+  // try to allocate in an arena if the alignment is small enough and the object is not too small (as for heap meta data)
+  if (size >= MI_ARENA_MIN_OBJ_SIZE && alignment <= MI_SEGMENT_ALIGN && align_offset == 0) {
+    void* p = mi_arena_allocate(numa_node, size, alignment, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
+    if (p != NULL) return p;
+  }
+
+  // finally, fall back to the OS
+  if (mi_option_is_enabled(mi_option_limit_os_alloc) || req_arena_id != _mi_arena_id_none()) {
+    errno = ENOMEM;
+    return NULL;
+  }
+  *is_zero = true;
+  *memid   = MI_MEMID_OS;
+  void* p = _mi_os_alloc_aligned_offset(size, alignment, align_offset, *commit, large, tld->stats);
+  if (p != NULL) { *is_pinned = *large; }
+  return p;
+}
+
+void* _mi_arena_alloc(size_t size, bool* commit, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t req_arena_id, size_t* memid, mi_os_tld_t* tld)
+{
+  return _mi_arena_alloc_aligned(size, MI_ARENA_BLOCK_SIZE, 0, commit, large, is_pinned, is_zero, req_arena_id, memid, tld);
+}
+
+void* mi_arena_area(mi_arena_id_t arena_id, size_t* size) {
+  if (size != NULL) *size = 0;
+  size_t arena_index = mi_arena_id_index(arena_id);
+  if (arena_index >= MI_MAX_ARENAS) return NULL;
+  mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t, &mi_arenas[arena_index]);
+  if (arena == NULL) return NULL;
+  if (size != NULL) *size = arena->block_count * MI_ARENA_BLOCK_SIZE;
+  return arena->start;
+}
+
+/* -----------------------------------------------------------
+  Arena free
+----------------------------------------------------------- */
+
+void _mi_arena_free(void* p, size_t size, size_t alignment, size_t align_offset, size_t memid, bool all_committed, mi_stats_t* stats) {
+  mi_assert_internal(size > 0 && stats != NULL);
+  if (p==NULL) return;
+  if (size==0) return;
+
+  if (memid == MI_MEMID_OS) {
+    // was a direct OS allocation, pass through
+    _mi_os_free_aligned(p, size, alignment, align_offset, all_committed, stats);
+  }
+  else {
+    // allocated in an arena
+    mi_assert_internal(align_offset == 0);
+    size_t arena_idx;
+    size_t bitmap_idx;
+    mi_arena_memid_indices(memid, &arena_idx, &bitmap_idx);
+    mi_assert_internal(arena_idx < MI_MAX_ARENAS);
+    mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t,&mi_arenas[arena_idx]);
+    mi_assert_internal(arena != NULL);
+    const size_t blocks = mi_block_count_of_size(size);
+    // checks
+    if (arena == NULL) {
+      _mi_error_message(EINVAL, "trying to free from non-existent arena: %p, size %zu, memid: 0x%zx\n", p, size, memid);
+      return;
+    }
+    mi_assert_internal(arena->field_count > mi_bitmap_index_field(bitmap_idx));
+    if (arena->field_count <= mi_bitmap_index_field(bitmap_idx)) {
+      _mi_error_message(EINVAL, "trying to free from non-existent arena block: %p, size %zu, memid: 0x%zx\n", p, size, memid);
+      return;
+    }
+    // potentially decommit
+    if (!arena->allow_decommit || arena->blocks_committed == NULL) {
+      mi_assert_internal(all_committed); // note: may be not true as we may "pretend" to be not committed (in segment.c)
+    }
+    else {
+      mi_assert_internal(arena->blocks_committed != NULL);
+      _mi_os_decommit(p, blocks * MI_ARENA_BLOCK_SIZE, stats); // ok if this fails
+      _mi_bitmap_unclaim_across(arena->blocks_committed, arena->field_count, blocks, bitmap_idx);
+    }
+    // and make it available to others again
+    bool all_inuse = _mi_bitmap_unclaim_across(arena->blocks_inuse, arena->field_count, blocks, bitmap_idx);
+    if (!all_inuse) {
+      _mi_error_message(EAGAIN, "trying to free an already freed block: %p, size %zu\n", p, size);
+      return;
+    };
+  }
+}
+
+/* -----------------------------------------------------------
+  Add an arena.
+----------------------------------------------------------- */
+
+static bool mi_arena_add(mi_arena_t* arena, mi_arena_id_t* arena_id) {
+  mi_assert_internal(arena != NULL);
+  mi_assert_internal((uintptr_t)mi_atomic_load_ptr_relaxed(uint8_t,&arena->start) % MI_SEGMENT_ALIGN == 0);
+  mi_assert_internal(arena->block_count > 0);
+  if (arena_id != NULL) *arena_id = -1;
+
+  size_t i = mi_atomic_increment_acq_rel(&mi_arena_count);
+  if (i >= MI_MAX_ARENAS) {
+    mi_atomic_decrement_acq_rel(&mi_arena_count);
+    return false;
+  }
+  mi_atomic_store_ptr_release(mi_arena_t,&mi_arenas[i], arena);
+  arena->id = mi_arena_id_create(i);
+  if (arena_id != NULL) *arena_id = arena->id;
+  return true;
+}
+
+bool mi_manage_os_memory_ex(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept
+{
+  if (arena_id != NULL) *arena_id = _mi_arena_id_none();
+  if (size < MI_ARENA_BLOCK_SIZE) return false;
+
+  if (is_large) {
+    mi_assert_internal(is_committed);
+    is_committed = true;
+  }
+
+  const size_t bcount = size / MI_ARENA_BLOCK_SIZE;
+  const size_t fields = _mi_divide_up(bcount, MI_BITMAP_FIELD_BITS);
+  const size_t bitmaps = (is_committed ? 2 : 3);
+  const size_t asize  = sizeof(mi_arena_t) + (bitmaps*fields*sizeof(mi_bitmap_field_t));
+  mi_arena_t* arena   = (mi_arena_t*)_mi_os_alloc(asize, &_mi_stats_main); // TODO: can we avoid allocating from the OS?
+  if (arena == NULL) return false;
+
+  arena->id = _mi_arena_id_none();
+  arena->exclusive = exclusive;
+  arena->block_count = bcount;
+  arena->field_count = fields;
+  arena->start = (uint8_t*)start;
+  arena->numa_node    = numa_node; // TODO: or get the current numa node if -1? (now it allows anyone to allocate on -1)
+  arena->is_large     = is_large;
+  arena->is_zero_init = is_zero;
+  arena->allow_decommit = !is_large && !is_committed; // only allow decommit for initially uncommitted memory
+  arena->search_idx   = 0;
+  arena->blocks_dirty = &arena->blocks_inuse[fields]; // just after inuse bitmap
+  arena->blocks_committed = (!arena->allow_decommit ? NULL : &arena->blocks_inuse[2*fields]); // just after dirty bitmap
+  // the bitmaps are already zero initialized due to os_alloc
+  // initialize committed bitmap?
+  if (arena->blocks_committed != NULL && is_committed) {
+    memset((void*)arena->blocks_committed, 0xFF, fields*sizeof(mi_bitmap_field_t)); // cast to void* to avoid atomic warning
+  }
+  // and claim leftover blocks if needed (so we never allocate there)
+  ptrdiff_t post = (fields * MI_BITMAP_FIELD_BITS) - bcount;
+  mi_assert_internal(post >= 0);
+  if (post > 0) {
+    // don't use leftover bits at the end
+    mi_bitmap_index_t postidx = mi_bitmap_index_create(fields - 1, MI_BITMAP_FIELD_BITS - post);
+    _mi_bitmap_claim(arena->blocks_inuse, fields, post, postidx, NULL);
+  }
+
+  return mi_arena_add(arena, arena_id);
+
+}
+
+// Reserve a range of regular OS memory
+int mi_reserve_os_memory_ex(size_t size, bool commit, bool allow_large, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept
+{
+  if (arena_id != NULL) *arena_id = _mi_arena_id_none();
+  size = _mi_align_up(size, MI_ARENA_BLOCK_SIZE); // at least one block
+  bool large = allow_large;
+  void* start = _mi_os_alloc_aligned(size, MI_SEGMENT_ALIGN, commit, &large, &_mi_stats_main);
+  if (start==NULL) return ENOMEM;
+  if (!mi_manage_os_memory_ex(start, size, (large || commit), large, true, -1, exclusive, arena_id)) {
+    _mi_os_free_ex(start, size, commit, &_mi_stats_main);
+    _mi_verbose_message("failed to reserve %zu k memory\n", _mi_divide_up(size,1024));
+    return ENOMEM;
+  }
+  _mi_verbose_message("reserved %zu KiB memory%s\n", _mi_divide_up(size,1024), large ? " (in large os pages)" : "");
+  return 0;
+}
+
+bool mi_manage_os_memory(void* start, size_t size, bool is_committed, bool is_large, bool is_zero, int numa_node) mi_attr_noexcept {
+  return mi_manage_os_memory_ex(start, size, is_committed, is_large, is_zero, numa_node, false, NULL);
+}
+
+int mi_reserve_os_memory(size_t size, bool commit, bool allow_large) mi_attr_noexcept {
+  return mi_reserve_os_memory_ex(size, commit, allow_large, false, NULL);
+}
+
+
+/* -----------------------------------------------------------
+  Debugging
+----------------------------------------------------------- */
+
+static size_t mi_debug_show_bitmap(const char* prefix, mi_bitmap_field_t* fields, size_t field_count ) {
+  size_t inuse_count = 0;
+  for (size_t i = 0; i < field_count; i++) {
+    char buf[MI_BITMAP_FIELD_BITS + 1];
+    uintptr_t field = mi_atomic_load_relaxed(&fields[i]);
+    for (size_t bit = 0; bit < MI_BITMAP_FIELD_BITS; bit++) {
+      bool inuse = ((((uintptr_t)1 << bit) & field) != 0);
+      if (inuse) inuse_count++;
+      buf[MI_BITMAP_FIELD_BITS - 1 - bit] = (inuse ? 'x' : '.');
+    }
+    buf[MI_BITMAP_FIELD_BITS] = 0;
+    _mi_verbose_message("%s%s\n", prefix, buf);
+  }
+  return inuse_count;
+}
+
+void mi_debug_show_arenas(void) mi_attr_noexcept {
+  size_t max_arenas = mi_atomic_load_relaxed(&mi_arena_count);
+  for (size_t i = 0; i < max_arenas; i++) {
+    mi_arena_t* arena = mi_atomic_load_ptr_relaxed(mi_arena_t, &mi_arenas[i]);
+    if (arena == NULL) break;
+    size_t inuse_count = 0;
+    _mi_verbose_message("arena %zu: %zu blocks with %zu fields\n", i, arena->block_count, arena->field_count);
+    inuse_count += mi_debug_show_bitmap("  ", arena->blocks_inuse, arena->field_count);
+    _mi_verbose_message("  blocks in use ('x'): %zu\n", inuse_count);
+  }
+}
+
+
+/* -----------------------------------------------------------
+  Reserve a huge page arena.
+----------------------------------------------------------- */
+// reserve at a specific numa node
+int mi_reserve_huge_os_pages_at_ex(size_t pages, int numa_node, size_t timeout_msecs, bool exclusive, mi_arena_id_t* arena_id) mi_attr_noexcept {
+  if (arena_id != NULL) *arena_id = -1;
+  if (pages==0) return 0;
+  if (numa_node < -1) numa_node = -1;
+  if (numa_node >= 0) numa_node = numa_node % _mi_os_numa_node_count();
+  size_t hsize = 0;
+  size_t pages_reserved = 0;
+  void* p = _mi_os_alloc_huge_os_pages(pages, numa_node, timeout_msecs, &pages_reserved, &hsize);
+  if (p==NULL || pages_reserved==0) {
+    _mi_warning_message("failed to reserve %zu GiB huge pages\n", pages);
+    return ENOMEM;
+  }
+  _mi_verbose_message("numa node %i: reserved %zu GiB huge pages (of the %zu GiB requested)\n", numa_node, pages_reserved, pages);
+
+  if (!mi_manage_os_memory_ex(p, hsize, true, true, true, numa_node, exclusive, arena_id)) {
+    _mi_os_free_huge_pages(p, hsize, &_mi_stats_main);
+    return ENOMEM;
+  }
+  return 0;
+}
+
+int mi_reserve_huge_os_pages_at(size_t pages, int numa_node, size_t timeout_msecs) mi_attr_noexcept {
+  return mi_reserve_huge_os_pages_at_ex(pages, numa_node, timeout_msecs, false, NULL);
+}
+
+// reserve huge pages evenly among the given number of numa nodes (or use the available ones as detected)
+int mi_reserve_huge_os_pages_interleave(size_t pages, size_t numa_nodes, size_t timeout_msecs) mi_attr_noexcept {
+  if (pages == 0) return 0;
+
+  // pages per numa node
+  size_t numa_count = (numa_nodes > 0 ? numa_nodes : _mi_os_numa_node_count());
+  if (numa_count <= 0) numa_count = 1;
+  const size_t pages_per = pages / numa_count;
+  const size_t pages_mod = pages % numa_count;
+  const size_t timeout_per = (timeout_msecs==0 ? 0 : (timeout_msecs / numa_count) + 50);
+
+  // reserve evenly among numa nodes
+  for (size_t numa_node = 0; numa_node < numa_count && pages > 0; numa_node++) {
+    size_t node_pages = pages_per;  // can be 0
+    if (numa_node < pages_mod) node_pages++;
+    int err = mi_reserve_huge_os_pages_at(node_pages, (int)numa_node, timeout_per);
+    if (err) return err;
+    if (pages < node_pages) {
+      pages = 0;
+    }
+    else {
+      pages -= node_pages;
+    }
+  }
+
+  return 0;
+}
+
+int mi_reserve_huge_os_pages(size_t pages, double max_secs, size_t* pages_reserved) mi_attr_noexcept {
+  MI_UNUSED(max_secs);
+  _mi_warning_message("mi_reserve_huge_os_pages is deprecated: use mi_reserve_huge_os_pages_interleave/at instead\n");
+  if (pages_reserved != NULL) *pages_reserved = 0;
+  int err = mi_reserve_huge_os_pages_interleave(pages, 0, (size_t)(max_secs * 1000.0));
+  if (err==0 && pages_reserved!=NULL) *pages_reserved = pages;
+  return err;
+}

preload-mimalloc/mimalloc/src/bitmap.c

@@ -0,0 +1,414 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2021 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+Concurrent bitmap that can set/reset sequences of bits atomically,
+represeted as an array of fields where each field is a machine word (`size_t`)
+
+There are two api's; the standard one cannot have sequences that cross
+between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
+(this is used in region allocation)
+
+The `_across` postfixed functions do allow sequences that can cross over
+between the fields. (This is used in arena allocation)
+---------------------------------------------------------------------------- */
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "bitmap.h"
+
+/* -----------------------------------------------------------
+  Bitmap definition
+----------------------------------------------------------- */
+
+// The bit mask for a given number of blocks at a specified bit index.
+static inline size_t mi_bitmap_mask_(size_t count, size_t bitidx) {
+  mi_assert_internal(count + bitidx <= MI_BITMAP_FIELD_BITS);
+  mi_assert_internal(count > 0);
+  if (count >= MI_BITMAP_FIELD_BITS) return MI_BITMAP_FIELD_FULL;
+  if (count == 0) return 0;
+  return ((((size_t)1 << count) - 1) << bitidx);
+}
+
+
+/* -----------------------------------------------------------
+  Claim a bit sequence atomically
+----------------------------------------------------------- */
+
+// Try to atomically claim a sequence of `count` bits in a single
+// field at `idx` in `bitmap`. Returns `true` on success.
+inline bool _mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_t count, mi_bitmap_index_t* bitmap_idx)
+{
+  mi_assert_internal(bitmap_idx != NULL);
+  mi_assert_internal(count <= MI_BITMAP_FIELD_BITS);
+  mi_assert_internal(count > 0);
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t map  = mi_atomic_load_relaxed(field);
+  if (map==MI_BITMAP_FIELD_FULL) return false; // short cut
+
+  // search for 0-bit sequence of length count
+  const size_t mask = mi_bitmap_mask_(count, 0);
+  const size_t bitidx_max = MI_BITMAP_FIELD_BITS - count;
+
+#ifdef MI_HAVE_FAST_BITSCAN
+  size_t bitidx = mi_ctz(~map);    // quickly find the first zero bit if possible
+#else
+  size_t bitidx = 0;               // otherwise start at 0
+#endif
+  size_t m = (mask << bitidx);     // invariant: m == mask shifted by bitidx
+
+  // scan linearly for a free range of zero bits
+  while (bitidx <= bitidx_max) {
+    const size_t mapm = map & m;
+    if (mapm == 0) {  // are the mask bits free at bitidx?
+      mi_assert_internal((m >> bitidx) == mask); // no overflow?
+      const size_t newmap = map | m;
+      mi_assert_internal((newmap^map) >> bitidx == mask);
+      if (!mi_atomic_cas_weak_acq_rel(field, &map, newmap)) {  // TODO: use strong cas here?
+        // no success, another thread claimed concurrently.. keep going (with updated `map`)
+        continue;
+      }
+      else {
+        // success, we claimed the bits!
+        *bitmap_idx = mi_bitmap_index_create(idx, bitidx);
+        return true;
+      }
+    }
+    else {
+      // on to the next bit range
+#ifdef MI_HAVE_FAST_BITSCAN
+      const size_t shift = (count == 1 ? 1 : mi_bsr(mapm) - bitidx + 1);
+      mi_assert_internal(shift > 0 && shift <= count);
+#else
+      const size_t shift = 1;
+#endif
+      bitidx += shift;
+      m <<= shift;
+    }
+  }
+  // no bits found
+  return false;
+}
+
+// Find `count` bits of 0 and set them to 1 atomically; returns `true` on success.
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+// `count` can be at most MI_BITMAP_FIELD_BITS and will never cross fields.
+bool _mi_bitmap_try_find_from_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+  size_t idx = start_field_idx;
+  for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
+    if (idx >= bitmap_fields) idx = 0; // wrap
+    if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Like _mi_bitmap_try_find_from_claim but with an extra predicate that must be fullfilled
+bool _mi_bitmap_try_find_from_claim_pred(mi_bitmap_t bitmap, const size_t bitmap_fields, 
+            const size_t start_field_idx, const size_t count, 
+            mi_bitmap_pred_fun_t pred_fun, void* pred_arg,            
+            mi_bitmap_index_t* bitmap_idx) {
+  size_t idx = start_field_idx;
+  for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
+    if (idx >= bitmap_fields) idx = 0; // wrap
+    if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+      if (pred_fun == NULL || pred_fun(*bitmap_idx, pred_arg)) { 
+        return true;
+      }
+      // predicate returned false, unclaim and look further
+      _mi_bitmap_unclaim(bitmap, bitmap_fields, count, *bitmap_idx);
+    }
+  }
+  return false;
+}
+
+/*
+// Find `count` bits of 0 and set them to 1 atomically; returns `true` on success.
+// For now, `count` can be at most MI_BITMAP_FIELD_BITS and will never span fields.
+bool _mi_bitmap_try_find_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+  return _mi_bitmap_try_find_from_claim(bitmap, bitmap_fields, 0, count, bitmap_idx);
+}
+*/
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  const size_t idx = mi_bitmap_index_field(bitmap_idx);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  const size_t mask = mi_bitmap_mask_(count, bitidx);
+  mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
+  // mi_assert_internal((bitmap[idx] & mask) == mask);
+  size_t prev = mi_atomic_and_acq_rel(&bitmap[idx], ~mask);
+  return ((prev & mask) == mask);
+}
+
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero) {
+  const size_t idx = mi_bitmap_index_field(bitmap_idx);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  const size_t mask = mi_bitmap_mask_(count, bitidx);
+  mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
+  //mi_assert_internal(any_zero != NULL || (bitmap[idx] & mask) == 0);
+  size_t prev = mi_atomic_or_acq_rel(&bitmap[idx], mask);
+  if (any_zero != NULL) *any_zero = ((prev & mask) != mask);
+  return ((prev & mask) == 0);
+}
+
+// Returns `true` if all `count` bits were 1. `any_ones` is `true` if there was at least one bit set to one.
+static bool mi_bitmap_is_claimedx(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_ones) {
+  const size_t idx = mi_bitmap_index_field(bitmap_idx);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  const size_t mask = mi_bitmap_mask_(count, bitidx);
+  mi_assert_internal(bitmap_fields > idx); MI_UNUSED(bitmap_fields);
+  size_t field = mi_atomic_load_relaxed(&bitmap[idx]);
+  if (any_ones != NULL) *any_ones = ((field & mask) != 0);
+  return ((field & mask) == mask);
+}
+
+bool _mi_bitmap_is_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  return mi_bitmap_is_claimedx(bitmap, bitmap_fields, count, bitmap_idx, NULL);
+}
+
+bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  bool any_ones;
+  mi_bitmap_is_claimedx(bitmap, bitmap_fields, count, bitmap_idx, &any_ones);
+  return any_ones;
+}
+
+
+//--------------------------------------------------------------------------
+// the `_across` functions work on bitmaps where sequences can cross over
+// between the fields. This is used in arena allocation
+//--------------------------------------------------------------------------
+
+// Try to atomically claim a sequence of `count` bits starting from the field
+// at `idx` in `bitmap` and crossing into subsequent fields. Returns `true` on success.
+static bool mi_bitmap_try_find_claim_field_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t idx, const size_t count, const size_t retries, mi_bitmap_index_t* bitmap_idx)
+{
+  mi_assert_internal(bitmap_idx != NULL);
+
+  // check initial trailing zeros
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t map = mi_atomic_load_relaxed(field);
+  const size_t initial = mi_clz(map);  // count of initial zeros starting at idx
+  mi_assert_internal(initial <= MI_BITMAP_FIELD_BITS);
+  if (initial == 0)     return false;
+  if (initial >= count) return _mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx);     // no need to cross fields
+  if (_mi_divide_up(count - initial, MI_BITMAP_FIELD_BITS) >= (bitmap_fields - idx)) return false; // not enough entries
+
+  // scan ahead
+  size_t found = initial;
+  size_t mask = 0;     // mask bits for the final field
+  while(found < count) {
+    field++;
+    map = mi_atomic_load_relaxed(field);
+    const size_t mask_bits = (found + MI_BITMAP_FIELD_BITS <= count ? MI_BITMAP_FIELD_BITS : (count - found));
+    mask = mi_bitmap_mask_(mask_bits, 0);
+    if ((map & mask) != 0) return false;
+    found += mask_bits;
+  }
+  mi_assert_internal(field < &bitmap[bitmap_fields]);
+
+  // found range of zeros up to the final field; mask contains mask in the final field
+  // now claim it atomically
+  mi_bitmap_field_t* const final_field = field;
+  const size_t final_mask = mask;
+  mi_bitmap_field_t* const initial_field = &bitmap[idx];
+  const size_t initial_mask = mi_bitmap_mask_(initial, MI_BITMAP_FIELD_BITS - initial);
+
+  // initial field
+  size_t newmap;
+  field = initial_field;
+  map = mi_atomic_load_relaxed(field);
+  do {
+    newmap = map | initial_mask;
+    if ((map & initial_mask) != 0) { goto rollback; };
+  } while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
+
+  // intermediate fields
+  while (++field < final_field) {
+    newmap = MI_BITMAP_FIELD_FULL;
+    map = 0;
+    if (!mi_atomic_cas_strong_acq_rel(field, &map, newmap)) { goto rollback; }
+  }
+
+  // final field
+  mi_assert_internal(field == final_field);
+  map = mi_atomic_load_relaxed(field);
+  do {
+    newmap = map | final_mask;
+    if ((map & final_mask) != 0) { goto rollback; }
+  } while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
+
+  // claimed!
+  *bitmap_idx = mi_bitmap_index_create(idx, MI_BITMAP_FIELD_BITS - initial);
+  return true;
+
+rollback:
+  // roll back intermediate fields
+  while (--field > initial_field) {
+    newmap = 0;
+    map = MI_BITMAP_FIELD_FULL;
+    mi_assert_internal(mi_atomic_load_relaxed(field) == map);
+    mi_atomic_store_release(field, newmap);
+  }
+  if (field == initial_field) {
+    map = mi_atomic_load_relaxed(field);
+    do {
+      mi_assert_internal((map & initial_mask) == initial_mask);
+      newmap = map & ~initial_mask;
+    } while (!mi_atomic_cas_strong_acq_rel(field, &map, newmap));
+  }
+  // retry? (we make a recursive call instead of goto to be able to use const declarations)
+  if (retries < 4) {
+    return mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, retries+1, bitmap_idx);
+  }
+  else {
+    return false;
+  }
+}
+
+
+// Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx) {
+  mi_assert_internal(count > 0);
+  if (count==1) return _mi_bitmap_try_find_from_claim(bitmap, bitmap_fields, start_field_idx, count, bitmap_idx);
+  size_t idx = start_field_idx;
+  for (size_t visited = 0; visited < bitmap_fields; visited++, idx++) {
+    if (idx >= bitmap_fields) idx = 0; // wrap
+    // try to claim inside the field
+    if (count <= MI_BITMAP_FIELD_BITS) {
+      if (_mi_bitmap_try_find_claim_field(bitmap, idx, count, bitmap_idx)) {
+        return true;
+      }
+    }
+    // try to claim across fields
+    if (mi_bitmap_try_find_claim_field_across(bitmap, bitmap_fields, idx, count, 0, bitmap_idx)) {
+      return true;
+    }
+  }
+  return false;
+}
+
+// Helper for masks across fields; returns the mid count, post_mask may be 0
+static size_t mi_bitmap_mask_across(mi_bitmap_index_t bitmap_idx, size_t bitmap_fields, size_t count, size_t* pre_mask, size_t* mid_mask, size_t* post_mask) {
+  MI_UNUSED(bitmap_fields);
+  const size_t bitidx = mi_bitmap_index_bit_in_field(bitmap_idx);
+  if mi_likely(bitidx + count <= MI_BITMAP_FIELD_BITS) {
+    *pre_mask = mi_bitmap_mask_(count, bitidx);
+    *mid_mask = 0;
+    *post_mask = 0;
+    mi_assert_internal(mi_bitmap_index_field(bitmap_idx) < bitmap_fields);
+    return 0;
+  }
+  else {
+    const size_t pre_bits = MI_BITMAP_FIELD_BITS - bitidx;
+    mi_assert_internal(pre_bits < count);
+    *pre_mask = mi_bitmap_mask_(pre_bits, bitidx);
+    count -= pre_bits;
+    const size_t mid_count = (count / MI_BITMAP_FIELD_BITS);
+    *mid_mask = MI_BITMAP_FIELD_FULL;
+    count %= MI_BITMAP_FIELD_BITS;
+    *post_mask = (count==0 ? 0 : mi_bitmap_mask_(count, 0));
+    mi_assert_internal(mi_bitmap_index_field(bitmap_idx) + mid_count + (count==0 ? 0 : 1) < bitmap_fields);
+    return mid_count;
+  }
+}
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  size_t idx = mi_bitmap_index_field(bitmap_idx);
+  size_t pre_mask;
+  size_t mid_mask;
+  size_t post_mask;
+  size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
+  bool all_one = true;
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t prev = mi_atomic_and_acq_rel(field++, ~pre_mask);
+  if ((prev & pre_mask) != pre_mask) all_one = false;
+  while(mid_count-- > 0) {
+    prev = mi_atomic_and_acq_rel(field++, ~mid_mask);
+    if ((prev & mid_mask) != mid_mask) all_one = false;
+  }
+  if (post_mask!=0) {
+    prev = mi_atomic_and_acq_rel(field, ~post_mask);
+    if ((prev & post_mask) != post_mask) all_one = false;
+  }
+  return all_one;
+}
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero) {
+  size_t idx = mi_bitmap_index_field(bitmap_idx);
+  size_t pre_mask;
+  size_t mid_mask;
+  size_t post_mask;
+  size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
+  bool all_zero = true;
+  bool any_zero = false;
+  _Atomic(size_t)*field = &bitmap[idx];
+  size_t prev = mi_atomic_or_acq_rel(field++, pre_mask);
+  if ((prev & pre_mask) != 0) all_zero = false;
+  if ((prev & pre_mask) != pre_mask) any_zero = true;
+  while (mid_count-- > 0) {
+    prev = mi_atomic_or_acq_rel(field++, mid_mask);
+    if ((prev & mid_mask) != 0) all_zero = false;
+    if ((prev & mid_mask) != mid_mask) any_zero = true;
+  }
+  if (post_mask!=0) {
+    prev = mi_atomic_or_acq_rel(field, post_mask);
+    if ((prev & post_mask) != 0) all_zero = false;
+    if ((prev & post_mask) != post_mask) any_zero = true;
+  }
+  if (pany_zero != NULL) *pany_zero = any_zero;
+  return all_zero;
+}
+
+
+// Returns `true` if all `count` bits were 1.
+// `any_ones` is `true` if there was at least one bit set to one.
+static bool mi_bitmap_is_claimedx_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_ones) {
+  size_t idx = mi_bitmap_index_field(bitmap_idx);
+  size_t pre_mask;
+  size_t mid_mask;
+  size_t post_mask;
+  size_t mid_count = mi_bitmap_mask_across(bitmap_idx, bitmap_fields, count, &pre_mask, &mid_mask, &post_mask);
+  bool all_ones = true;
+  bool any_ones = false;
+  mi_bitmap_field_t* field = &bitmap[idx];
+  size_t prev = mi_atomic_load_relaxed(field++);
+  if ((prev & pre_mask) != pre_mask) all_ones = false;
+  if ((prev & pre_mask) != 0) any_ones = true;
+  while (mid_count-- > 0) {
+    prev = mi_atomic_load_relaxed(field++);
+    if ((prev & mid_mask) != mid_mask) all_ones = false;
+    if ((prev & mid_mask) != 0) any_ones = true;
+  }
+  if (post_mask!=0) {
+    prev = mi_atomic_load_relaxed(field);
+    if ((prev & post_mask) != post_mask) all_ones = false;
+    if ((prev & post_mask) != 0) any_ones = true;
+  }
+  if (pany_ones != NULL) *pany_ones = any_ones;
+  return all_ones;
+}
+
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  return mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, NULL);
+}
+
+bool _mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx) {
+  bool any_ones;
+  mi_bitmap_is_claimedx_across(bitmap, bitmap_fields, count, bitmap_idx, &any_ones);
+  return any_ones;
+}

preload-mimalloc/mimalloc/src/bitmap.h

@@ -0,0 +1,111 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2023 Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+Concurrent bitmap that can set/reset sequences of bits atomically,
+represeted as an array of fields where each field is a machine word (`size_t`)
+
+There are two api's; the standard one cannot have sequences that cross
+between the bitmap fields (and a sequence must be <= MI_BITMAP_FIELD_BITS).
+(this is used in region allocation)
+
+The `_across` postfixed functions do allow sequences that can cross over
+between the fields. (This is used in arena allocation)
+---------------------------------------------------------------------------- */
+#pragma once
+#ifndef MI_BITMAP_H
+#define MI_BITMAP_H
+
+/* -----------------------------------------------------------
+  Bitmap definition
+----------------------------------------------------------- */
+
+#define MI_BITMAP_FIELD_BITS   (8*MI_SIZE_SIZE)
+#define MI_BITMAP_FIELD_FULL   (~((size_t)0))   // all bits set
+
+// An atomic bitmap of `size_t` fields
+typedef _Atomic(size_t)  mi_bitmap_field_t;
+typedef mi_bitmap_field_t*  mi_bitmap_t;
+
+// A bitmap index is the index of the bit in a bitmap.
+typedef size_t mi_bitmap_index_t;
+
+// Create a bit index.
+static inline mi_bitmap_index_t mi_bitmap_index_create(size_t idx, size_t bitidx) {
+  mi_assert_internal(bitidx < MI_BITMAP_FIELD_BITS);
+  return (idx*MI_BITMAP_FIELD_BITS) + bitidx;
+}
+
+// Create a bit index.
+static inline mi_bitmap_index_t mi_bitmap_index_create_from_bit(size_t full_bitidx) {  
+  return mi_bitmap_index_create(full_bitidx / MI_BITMAP_FIELD_BITS, full_bitidx % MI_BITMAP_FIELD_BITS);
+}
+
+// Get the field index from a bit index.
+static inline size_t mi_bitmap_index_field(mi_bitmap_index_t bitmap_idx) {
+  return (bitmap_idx / MI_BITMAP_FIELD_BITS);
+}
+
+// Get the bit index in a bitmap field
+static inline size_t mi_bitmap_index_bit_in_field(mi_bitmap_index_t bitmap_idx) {
+  return (bitmap_idx % MI_BITMAP_FIELD_BITS);
+}
+
+// Get the full bit index
+static inline size_t mi_bitmap_index_bit(mi_bitmap_index_t bitmap_idx) {
+  return bitmap_idx;
+}
+
+/* -----------------------------------------------------------
+  Claim a bit sequence atomically
+----------------------------------------------------------- */
+
+// Try to atomically claim a sequence of `count` bits in a single
+// field at `idx` in `bitmap`. Returns `true` on success.
+bool _mi_bitmap_try_find_claim_field(mi_bitmap_t bitmap, size_t idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+// For now, `count` can be at most MI_BITMAP_FIELD_BITS and will never cross fields.
+bool _mi_bitmap_try_find_from_claim(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Like _mi_bitmap_try_find_from_claim but with an extra predicate that must be fullfilled
+typedef bool (mi_cdecl *mi_bitmap_pred_fun_t)(mi_bitmap_index_t bitmap_idx, void* pred_arg);
+bool _mi_bitmap_try_find_from_claim_pred(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_pred_fun_t pred_fun, void* pred_arg, mi_bitmap_index_t* bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* any_zero);
+
+bool _mi_bitmap_is_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+bool _mi_bitmap_is_any_claimed(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+
+//--------------------------------------------------------------------------
+// the `_across` functions work on bitmaps where sequences can cross over
+// between the fields. This is used in arena allocation
+//--------------------------------------------------------------------------
+
+// Find `count` bits of zeros and set them to 1 atomically; returns `true` on success.
+// Starts at idx, and wraps around to search in all `bitmap_fields` fields.
+bool _mi_bitmap_try_find_from_claim_across(mi_bitmap_t bitmap, const size_t bitmap_fields, const size_t start_field_idx, const size_t count, mi_bitmap_index_t* bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 0 atomically
+// Returns `true` if all `count` bits were 1 previously.
+bool _mi_bitmap_unclaim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+// Set `count` bits at `bitmap_idx` to 1 atomically
+// Returns `true` if all `count` bits were 0 previously. `any_zero` is `true` if there was at least one zero bit.
+bool _mi_bitmap_claim_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx, bool* pany_zero);
+
+bool _mi_bitmap_is_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+bool _mi_bitmap_is_any_claimed_across(mi_bitmap_t bitmap, size_t bitmap_fields, size_t count, mi_bitmap_index_t bitmap_idx);
+
+#endif

preload-mimalloc/mimalloc/src/heap.c

@@ -0,0 +1,628 @@
+/*----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"  // mi_prim_get_default_heap
+
+#include <string.h>  // memset, memcpy
+
+#if defined(_MSC_VER) && (_MSC_VER < 1920)
+#pragma warning(disable:4204)  // non-constant aggregate initializer
+#endif
+
+/* -----------------------------------------------------------
+  Helpers
+----------------------------------------------------------- */
+
+// return `true` if ok, `false` to break
+typedef bool (heap_page_visitor_fun)(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2);
+
+// Visit all pages in a heap; returns `false` if break was called.
+static bool mi_heap_visit_pages(mi_heap_t* heap, heap_page_visitor_fun* fn, void* arg1, void* arg2)
+{
+  if (heap==NULL || heap->page_count==0) return 0;
+
+  // visit all pages
+  #if MI_DEBUG>1
+  size_t total = heap->page_count;
+  size_t count = 0;
+  #endif  
+
+  for (size_t i = 0; i <= MI_BIN_FULL; i++) {
+    mi_page_queue_t* pq = &heap->pages[i];
+    mi_page_t* page = pq->first;
+    while(page != NULL) {
+      mi_page_t* next = page->next; // save next in case the page gets removed from the queue
+      mi_assert_internal(mi_page_heap(page) == heap);
+      #if MI_DEBUG>1
+      count++;
+      #endif
+      if (!fn(heap, pq, page, arg1, arg2)) return false;
+      page = next; // and continue
+    }
+  }
+  mi_assert_internal(count == total);
+  return true;
+}
+
+
+#if MI_DEBUG>=2
+static bool mi_heap_page_is_valid(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
+  MI_UNUSED(arg1);
+  MI_UNUSED(arg2);
+  MI_UNUSED(pq);
+  mi_assert_internal(mi_page_heap(page) == heap);
+  mi_segment_t* segment = _mi_page_segment(page);
+  mi_assert_internal(segment->thread_id == heap->thread_id);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  return true;
+}
+#endif
+#if MI_DEBUG>=3
+static bool mi_heap_is_valid(mi_heap_t* heap) {
+  mi_assert_internal(heap!=NULL);
+  mi_heap_visit_pages(heap, &mi_heap_page_is_valid, NULL, NULL);
+  return true;
+}
+#endif
+
+
+
+
+/* -----------------------------------------------------------
+  "Collect" pages by migrating `local_free` and `thread_free`
+  lists and freeing empty pages. This is done when a thread
+  stops (and in that case abandons pages if there are still
+  blocks alive)
+----------------------------------------------------------- */
+
+typedef enum mi_collect_e {
+  MI_NORMAL,
+  MI_FORCE,
+  MI_ABANDON
+} mi_collect_t;
+
+
+static bool mi_heap_page_collect(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg_collect, void* arg2 ) {
+  MI_UNUSED(arg2);
+  MI_UNUSED(heap);
+  mi_assert_internal(mi_heap_page_is_valid(heap, pq, page, NULL, NULL));
+  mi_collect_t collect = *((mi_collect_t*)arg_collect);
+  _mi_page_free_collect(page, collect >= MI_FORCE);
+  if (mi_page_all_free(page)) {
+    // no more used blocks, free the page.
+    // note: this will free retired pages as well.
+    _mi_page_free(page, pq, collect >= MI_FORCE);
+  }
+  else if (collect == MI_ABANDON) {
+    // still used blocks but the thread is done; abandon the page
+    _mi_page_abandon(page, pq);
+  }
+  return true; // don't break
+}
+
+static bool mi_heap_page_never_delayed_free(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
+  MI_UNUSED(arg1);
+  MI_UNUSED(arg2);
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+  _mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE, false);
+  return true; // don't break
+}
+
+static void mi_heap_collect_ex(mi_heap_t* heap, mi_collect_t collect)
+{
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+
+  const bool force = collect >= MI_FORCE;  
+  _mi_deferred_free(heap, force);
+
+  // note: never reclaim on collect but leave it to threads that need storage to reclaim 
+  const bool force_main = 
+    #ifdef NDEBUG
+      collect == MI_FORCE
+    #else
+      collect >= MI_FORCE
+    #endif
+      && _mi_is_main_thread() && mi_heap_is_backing(heap) && !heap->no_reclaim;
+
+  if (force_main) {
+    // the main thread is abandoned (end-of-program), try to reclaim all abandoned segments.
+    // if all memory is freed by now, all segments should be freed.
+    _mi_abandoned_reclaim_all(heap, &heap->tld->segments);
+  }
+
+  // if abandoning, mark all pages to no longer add to delayed_free
+  if (collect == MI_ABANDON) {
+    mi_heap_visit_pages(heap, &mi_heap_page_never_delayed_free, NULL, NULL);
+  }
+
+  // free all current thread delayed blocks.
+  // (if abandoning, after this there are no more thread-delayed references into the pages.)
+  _mi_heap_delayed_free_all(heap);
+
+  // collect retired pages
+  _mi_heap_collect_retired(heap, force);
+
+  // collect all pages owned by this thread
+  mi_heap_visit_pages(heap, &mi_heap_page_collect, &collect, NULL);
+  mi_assert_internal( collect != MI_ABANDON || mi_atomic_load_ptr_acquire(mi_block_t,&heap->thread_delayed_free) == NULL );
+
+  // collect abandoned segments (in particular, decommit expired parts of segments in the abandoned segment list)
+  // note: forced decommit can be quite expensive if many threads are created/destroyed so we do not force on abandonment
+  _mi_abandoned_collect(heap, collect == MI_FORCE /* force? */, &heap->tld->segments);
+
+  // collect segment local caches
+  if (force) {
+    _mi_segment_thread_collect(&heap->tld->segments);
+  }
+
+  // decommit in global segment caches
+  // note: forced decommit can be quite expensive if many threads are created/destroyed so we do not force on abandonment
+  _mi_segment_cache_collect( collect == MI_FORCE, &heap->tld->os);  
+
+  // collect regions on program-exit (or shared library unload)
+  if (force && _mi_is_main_thread() && mi_heap_is_backing(heap)) {
+    //_mi_mem_collect(&heap->tld->os);
+  }
+}
+
+void _mi_heap_collect_abandon(mi_heap_t* heap) {
+  mi_heap_collect_ex(heap, MI_ABANDON);
+}
+
+void mi_heap_collect(mi_heap_t* heap, bool force) mi_attr_noexcept {
+  mi_heap_collect_ex(heap, (force ? MI_FORCE : MI_NORMAL));
+}
+
+void mi_collect(bool force) mi_attr_noexcept {
+  mi_heap_collect(mi_prim_get_default_heap(), force);
+}
+
+
+/* -----------------------------------------------------------
+  Heap new
+----------------------------------------------------------- */
+
+mi_heap_t* mi_heap_get_default(void) {
+  mi_thread_init();
+  return mi_prim_get_default_heap();
+}
+
+static bool mi_heap_is_default(const mi_heap_t* heap) {
+  return (heap == mi_prim_get_default_heap());
+}
+
+
+mi_heap_t* mi_heap_get_backing(void) {
+  mi_heap_t* heap = mi_heap_get_default();
+  mi_assert_internal(heap!=NULL);
+  mi_heap_t* bheap = heap->tld->heap_backing;
+  mi_assert_internal(bheap!=NULL);
+  mi_assert_internal(bheap->thread_id == _mi_thread_id());
+  return bheap;
+}
+
+mi_decl_nodiscard mi_heap_t* mi_heap_new_in_arena( mi_arena_id_t arena_id ) {
+  mi_heap_t* bheap = mi_heap_get_backing();
+  mi_heap_t* heap = mi_heap_malloc_tp(bheap, mi_heap_t);  // todo: OS allocate in secure mode?
+  if (heap==NULL) return NULL;
+  _mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(mi_heap_t));
+  heap->tld = bheap->tld;
+  heap->thread_id = _mi_thread_id();
+  heap->arena_id = arena_id;
+  _mi_random_split(&bheap->random, &heap->random);
+  heap->cookie  = _mi_heap_random_next(heap) | 1;
+  heap->keys[0] = _mi_heap_random_next(heap);
+  heap->keys[1] = _mi_heap_random_next(heap);
+  heap->no_reclaim = true;  // don't reclaim abandoned pages or otherwise destroy is unsafe
+  // push on the thread local heaps list
+  heap->next = heap->tld->heaps;
+  heap->tld->heaps = heap;
+  return heap;
+}
+
+mi_decl_nodiscard mi_heap_t* mi_heap_new(void) {
+  return mi_heap_new_in_arena(_mi_arena_id_none());
+}
+
+bool _mi_heap_memid_is_suitable(mi_heap_t* heap, size_t memid) {
+  return _mi_arena_memid_is_suitable(memid, heap->arena_id);
+}
+
+uintptr_t _mi_heap_random_next(mi_heap_t* heap) {
+  return _mi_random_next(&heap->random);
+}
+
+// zero out the page queues
+static void mi_heap_reset_pages(mi_heap_t* heap) {
+  mi_assert_internal(heap != NULL);
+  mi_assert_internal(mi_heap_is_initialized(heap));
+  // TODO: copy full empty heap instead?
+  memset(&heap->pages_free_direct, 0, sizeof(heap->pages_free_direct));
+  _mi_memcpy_aligned(&heap->pages, &_mi_heap_empty.pages, sizeof(heap->pages));
+  heap->thread_delayed_free = NULL;
+  heap->page_count = 0;
+}
+
+// called from `mi_heap_destroy` and `mi_heap_delete` to free the internal heap resources.
+static void mi_heap_free(mi_heap_t* heap) {
+  mi_assert(heap != NULL);
+  mi_assert_internal(mi_heap_is_initialized(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+  if (mi_heap_is_backing(heap)) return; // dont free the backing heap
+
+  // reset default
+  if (mi_heap_is_default(heap)) {
+    _mi_heap_set_default_direct(heap->tld->heap_backing);
+  }
+
+  // remove ourselves from the thread local heaps list
+  // linear search but we expect the number of heaps to be relatively small
+  mi_heap_t* prev = NULL;
+  mi_heap_t* curr = heap->tld->heaps;
+  while (curr != heap && curr != NULL) {
+    prev = curr;
+    curr = curr->next;
+  }
+  mi_assert_internal(curr == heap);
+  if (curr == heap) {
+    if (prev != NULL) { prev->next = heap->next; }
+                 else { heap->tld->heaps = heap->next; }
+  }
+  mi_assert_internal(heap->tld->heaps != NULL);
+
+  // and free the used memory
+  mi_free(heap);
+}
+
+
+/* -----------------------------------------------------------
+  Heap destroy
+----------------------------------------------------------- */
+
+static bool _mi_heap_page_destroy(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* arg1, void* arg2) {
+  MI_UNUSED(arg1);
+  MI_UNUSED(arg2);
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+
+  // ensure no more thread_delayed_free will be added
+  _mi_page_use_delayed_free(page, MI_NEVER_DELAYED_FREE, false);
+
+  // stats
+  const size_t bsize = mi_page_block_size(page);
+  if (bsize > MI_MEDIUM_OBJ_SIZE_MAX) {
+    if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+      mi_heap_stat_decrease(heap, large, bsize);
+    }
+    else {
+      mi_heap_stat_decrease(heap, huge, bsize);
+    }
+  }
+#if (MI_STAT)
+  _mi_page_free_collect(page, false);  // update used count
+  const size_t inuse = page->used;
+  if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+    mi_heap_stat_decrease(heap, normal, bsize * inuse);
+#if (MI_STAT>1)
+    mi_heap_stat_decrease(heap, normal_bins[_mi_bin(bsize)], inuse);
+#endif
+  }
+  mi_heap_stat_decrease(heap, malloc, bsize * inuse);  // todo: off for aligned blocks...
+#endif
+
+  /// pretend it is all free now
+  mi_assert_internal(mi_page_thread_free(page) == NULL);
+  page->used = 0;
+
+  // and free the page
+  // mi_page_free(page,false);
+  page->next = NULL;
+  page->prev = NULL;
+  _mi_segment_page_free(page,false /* no force? */, &heap->tld->segments);
+
+  return true; // keep going
+}
+
+void _mi_heap_destroy_pages(mi_heap_t* heap) {
+  mi_heap_visit_pages(heap, &_mi_heap_page_destroy, NULL, NULL);
+  mi_heap_reset_pages(heap);
+}
+
+#if MI_TRACK_HEAP_DESTROY
+static bool mi_cdecl mi_heap_track_block_free(const mi_heap_t* heap, const mi_heap_area_t* area, void* block, size_t block_size, void* arg) {
+  MI_UNUSED(heap); MI_UNUSED(area);  MI_UNUSED(arg); MI_UNUSED(block_size);
+  mi_track_free_size(block,mi_usable_size(block));
+  return true;
+}
+#endif
+
+void mi_heap_destroy(mi_heap_t* heap) {
+  mi_assert(heap != NULL);
+  mi_assert(mi_heap_is_initialized(heap));
+  mi_assert(heap->no_reclaim);
+  mi_assert_expensive(mi_heap_is_valid(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+  if (!heap->no_reclaim) {
+    // don't free in case it may contain reclaimed pages
+    mi_heap_delete(heap);
+  }
+  else {
+    // track all blocks as freed
+    #if MI_TRACK_HEAP_DESTROY
+    mi_heap_visit_blocks(heap, true, mi_heap_track_block_free, NULL);
+    #endif
+    // free all pages
+    _mi_heap_destroy_pages(heap);
+    mi_heap_free(heap);
+  }
+}
+
+void _mi_heap_destroy_all(void) {
+  mi_heap_t* bheap = mi_heap_get_backing();
+  mi_heap_t* curr = bheap->tld->heaps;
+  while (curr != NULL) {
+    mi_heap_t* next = curr->next;
+    if (curr->no_reclaim) {
+      mi_heap_destroy(curr);
+    }
+    else {
+      _mi_heap_destroy_pages(curr);
+    }
+    curr = next;
+  }
+}
+
+/* -----------------------------------------------------------
+  Safe Heap delete
+----------------------------------------------------------- */
+
+// Transfer the pages from one heap to the other
+static void mi_heap_absorb(mi_heap_t* heap, mi_heap_t* from) {
+  mi_assert_internal(heap!=NULL);
+  if (from==NULL || from->page_count == 0) return;
+
+  // reduce the size of the delayed frees
+  _mi_heap_delayed_free_partial(from);
+
+  // transfer all pages by appending the queues; this will set a new heap field
+  // so threads may do delayed frees in either heap for a while.
+  // note: appending waits for each page to not be in the `MI_DELAYED_FREEING` state
+  // so after this only the new heap will get delayed frees
+  for (size_t i = 0; i <= MI_BIN_FULL; i++) {
+    mi_page_queue_t* pq = &heap->pages[i];
+    mi_page_queue_t* append = &from->pages[i];
+    size_t pcount = _mi_page_queue_append(heap, pq, append);
+    heap->page_count += pcount;
+    from->page_count -= pcount;
+  }
+  mi_assert_internal(from->page_count == 0);
+
+  // and do outstanding delayed frees in the `from` heap
+  // note: be careful here as the `heap` field in all those pages no longer point to `from`,
+  // turns out to be ok as `_mi_heap_delayed_free` only visits the list and calls a
+  // the regular `_mi_free_delayed_block` which is safe.
+  _mi_heap_delayed_free_all(from);
+  #if !defined(_MSC_VER) || (_MSC_VER > 1900) // somehow the following line gives an error in VS2015, issue #353
+  mi_assert_internal(mi_atomic_load_ptr_relaxed(mi_block_t,&from->thread_delayed_free) == NULL);
+  #endif
+
+  // and reset the `from` heap
+  mi_heap_reset_pages(from);
+}
+
+// Safe delete a heap without freeing any still allocated blocks in that heap.
+void mi_heap_delete(mi_heap_t* heap)
+{
+  mi_assert(heap != NULL);
+  mi_assert(mi_heap_is_initialized(heap));
+  mi_assert_expensive(mi_heap_is_valid(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return;
+
+  if (!mi_heap_is_backing(heap)) {
+    // tranfer still used pages to the backing heap
+    mi_heap_absorb(heap->tld->heap_backing, heap);
+  }
+  else {
+    // the backing heap abandons its pages
+    _mi_heap_collect_abandon(heap);
+  }
+  mi_assert_internal(heap->page_count==0);
+  mi_heap_free(heap);
+}
+
+mi_heap_t* mi_heap_set_default(mi_heap_t* heap) {
+  mi_assert(heap != NULL);
+  mi_assert(mi_heap_is_initialized(heap));
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return NULL;
+  mi_assert_expensive(mi_heap_is_valid(heap));
+  mi_heap_t* old = mi_prim_get_default_heap();
+  _mi_heap_set_default_direct(heap);
+  return old;
+}
+
+
+
+
+/* -----------------------------------------------------------
+  Analysis
+----------------------------------------------------------- */
+
+// static since it is not thread safe to access heaps from other threads.
+static mi_heap_t* mi_heap_of_block(const void* p) {
+  if (p == NULL) return NULL;
+  mi_segment_t* segment = _mi_ptr_segment(p);
+  bool valid = (_mi_ptr_cookie(segment) == segment->cookie);
+  mi_assert_internal(valid);
+  if mi_unlikely(!valid) return NULL;
+  return mi_page_heap(_mi_segment_page_of(segment,p));
+}
+
+bool mi_heap_contains_block(mi_heap_t* heap, const void* p) {
+  mi_assert(heap != NULL);
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return false;
+  return (heap == mi_heap_of_block(p));
+}
+
+
+static bool mi_heap_page_check_owned(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* p, void* vfound) {
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+  bool* found = (bool*)vfound;
+  mi_segment_t* segment = _mi_page_segment(page);
+  void* start = _mi_page_start(segment, page, NULL);
+  void* end   = (uint8_t*)start + (page->capacity * mi_page_block_size(page));
+  *found = (p >= start && p < end);
+  return (!*found); // continue if not found
+}
+
+bool mi_heap_check_owned(mi_heap_t* heap, const void* p) {
+  mi_assert(heap != NULL);
+  if (heap==NULL || !mi_heap_is_initialized(heap)) return false;
+  if (((uintptr_t)p & (MI_INTPTR_SIZE - 1)) != 0) return false;  // only aligned pointers
+  bool found = false;
+  mi_heap_visit_pages(heap, &mi_heap_page_check_owned, (void*)p, &found);
+  return found;
+}
+
+bool mi_check_owned(const void* p) {
+  return mi_heap_check_owned(mi_prim_get_default_heap(), p);
+}
+
+/* -----------------------------------------------------------
+  Visit all heap blocks and areas
+  Todo: enable visiting abandoned pages, and
+        enable visiting all blocks of all heaps across threads
+----------------------------------------------------------- */
+
+// Separate struct to keep `mi_page_t` out of the public interface
+typedef struct mi_heap_area_ex_s {
+  mi_heap_area_t area;
+  mi_page_t*     page;
+} mi_heap_area_ex_t;
+
+static bool mi_heap_area_visit_blocks(const mi_heap_area_ex_t* xarea, mi_block_visit_fun* visitor, void* arg) {
+  mi_assert(xarea != NULL);
+  if (xarea==NULL) return true;
+  const mi_heap_area_t* area = &xarea->area;
+  mi_page_t* page = xarea->page;
+  mi_assert(page != NULL);
+  if (page == NULL) return true;
+
+  _mi_page_free_collect(page,true);
+  mi_assert_internal(page->local_free == NULL);
+  if (page->used == 0) return true;
+
+  const size_t bsize = mi_page_block_size(page);
+  const size_t ubsize = mi_page_usable_block_size(page); // without padding
+  size_t   psize;
+  uint8_t* pstart = _mi_page_start(_mi_page_segment(page), page, &psize);
+
+  if (page->capacity == 1) {
+    // optimize page with one block
+    mi_assert_internal(page->used == 1 && page->free == NULL);
+    return visitor(mi_page_heap(page), area, pstart, ubsize, arg);
+  }
+
+  // create a bitmap of free blocks.
+  #define MI_MAX_BLOCKS   (MI_SMALL_PAGE_SIZE / sizeof(void*))
+  uintptr_t free_map[MI_MAX_BLOCKS / sizeof(uintptr_t)];
+  memset(free_map, 0, sizeof(free_map));
+
+  #if MI_DEBUG>1
+  size_t free_count = 0;
+  #endif
+  for (mi_block_t* block = page->free; block != NULL; block = mi_block_next(page,block)) {
+    #if MI_DEBUG>1
+    free_count++;
+    #endif
+    mi_assert_internal((uint8_t*)block >= pstart && (uint8_t*)block < (pstart + psize));
+    size_t offset = (uint8_t*)block - pstart;
+    mi_assert_internal(offset % bsize == 0);
+    size_t blockidx = offset / bsize;  // Todo: avoid division?
+    mi_assert_internal( blockidx < MI_MAX_BLOCKS);
+    size_t bitidx = (blockidx / sizeof(uintptr_t));
+    size_t bit = blockidx - (bitidx * sizeof(uintptr_t));
+    free_map[bitidx] |= ((uintptr_t)1 << bit);
+  }
+  mi_assert_internal(page->capacity == (free_count + page->used));
+
+  // walk through all blocks skipping the free ones
+  #if MI_DEBUG>1
+  size_t used_count = 0;
+  #endif
+  for (size_t i = 0; i < page->capacity; i++) {
+    size_t bitidx = (i / sizeof(uintptr_t));
+    size_t bit = i - (bitidx * sizeof(uintptr_t));
+    uintptr_t m = free_map[bitidx];
+    if (bit == 0 && m == UINTPTR_MAX) {
+      i += (sizeof(uintptr_t) - 1); // skip a run of free blocks
+    }
+    else if ((m & ((uintptr_t)1 << bit)) == 0) {
+      #if MI_DEBUG>1
+      used_count++;
+      #endif
+      uint8_t* block = pstart + (i * bsize);
+      if (!visitor(mi_page_heap(page), area, block, ubsize, arg)) return false;
+    }
+  }
+  mi_assert_internal(page->used == used_count);
+  return true;
+}
+
+typedef bool (mi_heap_area_visit_fun)(const mi_heap_t* heap, const mi_heap_area_ex_t* area, void* arg);
+
+
+static bool mi_heap_visit_areas_page(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_t* page, void* vfun, void* arg) {
+  MI_UNUSED(heap);
+  MI_UNUSED(pq);
+  mi_heap_area_visit_fun* fun = (mi_heap_area_visit_fun*)vfun;
+  mi_heap_area_ex_t xarea;
+  const size_t bsize = mi_page_block_size(page);
+  const size_t ubsize = mi_page_usable_block_size(page);
+  xarea.page = page;
+  xarea.area.reserved = page->reserved * bsize;
+  xarea.area.committed = page->capacity * bsize;
+  xarea.area.blocks = _mi_page_start(_mi_page_segment(page), page, NULL);
+  xarea.area.used = page->used;   // number of blocks in use (#553)
+  xarea.area.block_size = ubsize;
+  xarea.area.full_block_size = bsize;
+  return fun(heap, &xarea, arg);
+}
+
+// Visit all heap pages as areas
+static bool mi_heap_visit_areas(const mi_heap_t* heap, mi_heap_area_visit_fun* visitor, void* arg) {
+  if (visitor == NULL) return false;
+  return mi_heap_visit_pages((mi_heap_t*)heap, &mi_heap_visit_areas_page, (void*)(visitor), arg); // note: function pointer to void* :-{
+}
+
+// Just to pass arguments
+typedef struct mi_visit_blocks_args_s {
+  bool  visit_blocks;
+  mi_block_visit_fun* visitor;
+  void* arg;
+} mi_visit_blocks_args_t;
+
+static bool mi_heap_area_visitor(const mi_heap_t* heap, const mi_heap_area_ex_t* xarea, void* arg) {
+  mi_visit_blocks_args_t* args = (mi_visit_blocks_args_t*)arg;
+  if (!args->visitor(heap, &xarea->area, NULL, xarea->area.block_size, args->arg)) return false;
+  if (args->visit_blocks) {
+    return mi_heap_area_visit_blocks(xarea, args->visitor, args->arg);
+  }
+  else {
+    return true;
+  }
+}
+
+// Visit all blocks in a heap
+bool mi_heap_visit_blocks(const mi_heap_t* heap, bool visit_blocks, mi_block_visit_fun* visitor, void* arg) {
+  mi_visit_blocks_args_t args = { visit_blocks, visitor, arg };
+  return mi_heap_visit_areas(heap, &mi_heap_area_visitor, &args);
+}

preload-mimalloc/mimalloc/src/init.c

@@ -0,0 +1,688 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2022, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/prim.h"
+
+#include <string.h>  // memcpy, memset
+#include <stdlib.h>  // atexit
+
+
+// Empty page used to initialize the small free pages array
+const mi_page_t _mi_page_empty = {
+  0, false, false, false, false,
+  0,       // capacity
+  0,       // reserved capacity
+  { 0 },   // flags
+  false,   // is_zero
+  0,       // retire_expire
+  NULL,    // free
+  0,       // used
+  0,       // xblock_size
+  NULL,    // local_free
+  #if (MI_PADDING || MI_ENCODE_FREELIST)
+  { 0, 0 },
+  #endif
+  MI_ATOMIC_VAR_INIT(0), // xthread_free
+  MI_ATOMIC_VAR_INIT(0), // xheap
+  NULL, NULL
+  #if MI_INTPTR_SIZE==8
+  , { 0 }  // padding
+  #endif
+};
+
+#define MI_PAGE_EMPTY() ((mi_page_t*)&_mi_page_empty)
+
+#if (MI_PADDING>0) && (MI_INTPTR_SIZE >= 8)
+#define MI_SMALL_PAGES_EMPTY  { MI_INIT128(MI_PAGE_EMPTY), MI_PAGE_EMPTY(), MI_PAGE_EMPTY() }
+#elif (MI_PADDING>0)
+#define MI_SMALL_PAGES_EMPTY  { MI_INIT128(MI_PAGE_EMPTY), MI_PAGE_EMPTY(), MI_PAGE_EMPTY(), MI_PAGE_EMPTY() }
+#else
+#define MI_SMALL_PAGES_EMPTY  { MI_INIT128(MI_PAGE_EMPTY), MI_PAGE_EMPTY() }
+#endif
+
+
+// Empty page queues for every bin
+#define QNULL(sz)  { NULL, NULL, (sz)*sizeof(uintptr_t) }
+#define MI_PAGE_QUEUES_EMPTY \
+  { QNULL(1), \
+    QNULL(     1), QNULL(     2), QNULL(     3), QNULL(     4), QNULL(     5), QNULL(     6), QNULL(     7), QNULL(     8), /* 8 */ \
+    QNULL(    10), QNULL(    12), QNULL(    14), QNULL(    16), QNULL(    20), QNULL(    24), QNULL(    28), QNULL(    32), /* 16 */ \
+    QNULL(    40), QNULL(    48), QNULL(    56), QNULL(    64), QNULL(    80), QNULL(    96), QNULL(   112), QNULL(   128), /* 24 */ \
+    QNULL(   160), QNULL(   192), QNULL(   224), QNULL(   256), QNULL(   320), QNULL(   384), QNULL(   448), QNULL(   512), /* 32 */ \
+    QNULL(   640), QNULL(   768), QNULL(   896), QNULL(  1024), QNULL(  1280), QNULL(  1536), QNULL(  1792), QNULL(  2048), /* 40 */ \
+    QNULL(  2560), QNULL(  3072), QNULL(  3584), QNULL(  4096), QNULL(  5120), QNULL(  6144), QNULL(  7168), QNULL(  8192), /* 48 */ \
+    QNULL( 10240), QNULL( 12288), QNULL( 14336), QNULL( 16384), QNULL( 20480), QNULL( 24576), QNULL( 28672), QNULL( 32768), /* 56 */ \
+    QNULL( 40960), QNULL( 49152), QNULL( 57344), QNULL( 65536), QNULL( 81920), QNULL( 98304), QNULL(114688), QNULL(131072), /* 64 */ \
+    QNULL(163840), QNULL(196608), QNULL(229376), QNULL(262144), QNULL(327680), QNULL(393216), QNULL(458752), QNULL(524288), /* 72 */ \
+    QNULL(MI_MEDIUM_OBJ_WSIZE_MAX + 1  /* 655360, Huge queue */), \
+    QNULL(MI_MEDIUM_OBJ_WSIZE_MAX + 2) /* Full queue */ }
+
+#define MI_STAT_COUNT_NULL()  {0,0,0,0}
+
+// Empty statistics
+#if MI_STAT>1
+#define MI_STAT_COUNT_END_NULL()  , { MI_STAT_COUNT_NULL(), MI_INIT32(MI_STAT_COUNT_NULL) }
+#else
+#define MI_STAT_COUNT_END_NULL()
+#endif
+
+#define MI_STATS_NULL  \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  MI_STAT_COUNT_NULL(), MI_STAT_COUNT_NULL(), \
+  { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 },     \
+  { 0, 0 }, { 0, 0 }, { 0, 0 }, { 0, 0 } \
+  MI_STAT_COUNT_END_NULL()
+
+
+// Empty slice span queues for every bin
+#define SQNULL(sz)  { NULL, NULL, sz }
+#define MI_SEGMENT_SPAN_QUEUES_EMPTY \
+  { SQNULL(1), \
+    SQNULL(     1), SQNULL(     2), SQNULL(     3), SQNULL(     4), SQNULL(     5), SQNULL(     6), SQNULL(     7), SQNULL(    10), /*  8 */ \
+    SQNULL(    12), SQNULL(    14), SQNULL(    16), SQNULL(    20), SQNULL(    24), SQNULL(    28), SQNULL(    32), SQNULL(    40), /* 16 */ \
+    SQNULL(    48), SQNULL(    56), SQNULL(    64), SQNULL(    80), SQNULL(    96), SQNULL(   112), SQNULL(   128), SQNULL(   160), /* 24 */ \
+    SQNULL(   192), SQNULL(   224), SQNULL(   256), SQNULL(   320), SQNULL(   384), SQNULL(   448), SQNULL(   512), SQNULL(   640), /* 32 */ \
+    SQNULL(   768), SQNULL(   896), SQNULL(  1024) /* 35 */ }
+
+
+// --------------------------------------------------------
+// Statically allocate an empty heap as the initial
+// thread local value for the default heap,
+// and statically allocate the backing heap for the main
+// thread so it can function without doing any allocation
+// itself (as accessing a thread local for the first time
+// may lead to allocation itself on some platforms)
+// --------------------------------------------------------
+
+mi_decl_cache_align const mi_heap_t _mi_heap_empty = {
+  NULL,
+  MI_SMALL_PAGES_EMPTY,
+  MI_PAGE_QUEUES_EMPTY,
+  MI_ATOMIC_VAR_INIT(NULL),
+  0,                // tid
+  0,                // cookie
+  0,                // arena id
+  { 0, 0 },         // keys
+  { {0}, {0}, 0, true }, // random
+  0,                // page count
+  MI_BIN_FULL, 0,   // page retired min/max
+  NULL,             // next
+  false
+};
+
+#define tld_empty_stats  ((mi_stats_t*)((uint8_t*)&tld_empty + offsetof(mi_tld_t,stats)))
+#define tld_empty_os     ((mi_os_tld_t*)((uint8_t*)&tld_empty + offsetof(mi_tld_t,os)))
+
+mi_decl_cache_align static const mi_tld_t tld_empty = {
+  0,
+  false,
+  NULL, NULL,
+  { MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, tld_empty_stats, tld_empty_os }, // segments
+  { 0, tld_empty_stats }, // os
+  { MI_STATS_NULL }       // stats
+};
+
+mi_threadid_t _mi_thread_id(void) mi_attr_noexcept {
+  return _mi_prim_thread_id();
+}
+
+// the thread-local default heap for allocation
+mi_decl_thread mi_heap_t* _mi_heap_default = (mi_heap_t*)&_mi_heap_empty;
+
+extern mi_heap_t _mi_heap_main;
+
+static mi_tld_t tld_main = {
+  0, false,
+  &_mi_heap_main, & _mi_heap_main,
+  { MI_SEGMENT_SPAN_QUEUES_EMPTY, 0, 0, 0, 0, &tld_main.stats, &tld_main.os }, // segments
+  { 0, &tld_main.stats },  // os
+  { MI_STATS_NULL }       // stats
+};
+
+mi_heap_t _mi_heap_main = {
+  &tld_main,
+  MI_SMALL_PAGES_EMPTY,
+  MI_PAGE_QUEUES_EMPTY,
+  MI_ATOMIC_VAR_INIT(NULL),
+  0,                // thread id
+  0,                // initial cookie
+  0,                // arena id
+  { 0, 0 },         // the key of the main heap can be fixed (unlike page keys that need to be secure!)
+  { {0x846ca68b}, {0}, 0, true },  // random
+  0,                // page count
+  MI_BIN_FULL, 0,   // page retired min/max
+  NULL,             // next heap
+  false             // can reclaim
+};
+
+bool _mi_process_is_initialized = false;  // set to `true` in `mi_process_init`.
+
+mi_stats_t _mi_stats_main = { MI_STATS_NULL };
+
+
+static void mi_heap_main_init(void) {
+  if (_mi_heap_main.cookie == 0) {
+    _mi_heap_main.thread_id = _mi_thread_id();
+    _mi_heap_main.cookie = 1;
+    #if defined(_WIN32) && !defined(MI_SHARED_LIB)
+      _mi_random_init_weak(&_mi_heap_main.random);    // prevent allocation failure during bcrypt dll initialization with static linking
+    #else
+      _mi_random_init(&_mi_heap_main.random);
+    #endif
+    _mi_heap_main.cookie  = _mi_heap_random_next(&_mi_heap_main);
+    _mi_heap_main.keys[0] = _mi_heap_random_next(&_mi_heap_main);
+    _mi_heap_main.keys[1] = _mi_heap_random_next(&_mi_heap_main);
+  }
+}
+
+mi_heap_t* _mi_heap_main_get(void) {
+  mi_heap_main_init();
+  return &_mi_heap_main;
+}
+
+
+/* -----------------------------------------------------------
+  Initialization and freeing of the thread local heaps
+----------------------------------------------------------- */
+
+// note: in x64 in release build `sizeof(mi_thread_data_t)` is under 4KiB (= OS page size).
+typedef struct mi_thread_data_s {
+  mi_heap_t  heap;  // must come first due to cast in `_mi_heap_done`
+  mi_tld_t   tld;
+} mi_thread_data_t;
+
+
+// Thread meta-data is allocated directly from the OS. For
+// some programs that do not use thread pools and allocate and
+// destroy many OS threads, this may causes too much overhead
+// per thread so we maintain a small cache of recently freed metadata.
+
+#define TD_CACHE_SIZE (8)
+static _Atomic(mi_thread_data_t*) td_cache[TD_CACHE_SIZE];
+
+static mi_thread_data_t* mi_thread_data_alloc(void) {
+  // try to find thread metadata in the cache
+  mi_thread_data_t* td;
+  for (int i = 0; i < TD_CACHE_SIZE; i++) {
+    td = mi_atomic_load_ptr_relaxed(mi_thread_data_t, &td_cache[i]);
+    if (td != NULL) {
+      td = mi_atomic_exchange_ptr_acq_rel(mi_thread_data_t, &td_cache[i], NULL);
+      if (td != NULL) {
+        return td;
+      }
+    }
+  }
+  // if that fails, allocate directly from the OS
+  td = (mi_thread_data_t*)_mi_os_alloc(sizeof(mi_thread_data_t), &_mi_stats_main);
+  if (td == NULL) {
+    // if this fails, try once more. (issue #257)
+    td = (mi_thread_data_t*)_mi_os_alloc(sizeof(mi_thread_data_t), &_mi_stats_main);
+    if (td == NULL) {
+      // really out of memory
+      _mi_error_message(ENOMEM, "unable to allocate thread local heap metadata (%zu bytes)\n", sizeof(mi_thread_data_t));
+    }
+  }
+  return td;
+}
+
+static void mi_thread_data_free( mi_thread_data_t* tdfree ) {
+  // try to add the thread metadata to the cache
+  for (int i = 0; i < TD_CACHE_SIZE; i++) {
+    mi_thread_data_t* td = mi_atomic_load_ptr_relaxed(mi_thread_data_t, &td_cache[i]);
+    if (td == NULL) {
+      mi_thread_data_t* expected = NULL;
+      if (mi_atomic_cas_ptr_weak_acq_rel(mi_thread_data_t, &td_cache[i], &expected, tdfree)) {
+        return;
+      }
+    }
+  }
+  // if that fails, just free it directly
+  _mi_os_free(tdfree, sizeof(mi_thread_data_t), &_mi_stats_main);
+}
+
+static void mi_thread_data_collect(void) {
+  // free all thread metadata from the cache
+  for (int i = 0; i < TD_CACHE_SIZE; i++) {
+    mi_thread_data_t* td = mi_atomic_load_ptr_relaxed(mi_thread_data_t, &td_cache[i]);
+    if (td != NULL) {
+      td = mi_atomic_exchange_ptr_acq_rel(mi_thread_data_t, &td_cache[i], NULL);
+      if (td != NULL) {
+        _mi_os_free( td, sizeof(mi_thread_data_t), &_mi_stats_main );
+      }
+    }
+  }
+}
+
+// Initialize the thread local default heap, called from `mi_thread_init`
+static bool _mi_heap_init(void) {
+  if (mi_heap_is_initialized(mi_prim_get_default_heap())) return true;
+  if (_mi_is_main_thread()) {
+    // mi_assert_internal(_mi_heap_main.thread_id != 0);  // can happen on freeBSD where alloc is called before any initialization
+    // the main heap is statically allocated
+    mi_heap_main_init();
+    _mi_heap_set_default_direct(&_mi_heap_main);
+    //mi_assert_internal(_mi_heap_default->tld->heap_backing == mi_prim_get_default_heap());
+  }
+  else {
+    // use `_mi_os_alloc` to allocate directly from the OS
+    mi_thread_data_t* td = mi_thread_data_alloc();
+    if (td == NULL) return false;
+
+    // OS allocated so already zero initialized
+    mi_tld_t*  tld = &td->tld;
+    mi_heap_t* heap = &td->heap;
+    _mi_memcpy_aligned(tld, &tld_empty, sizeof(*tld));
+    _mi_memcpy_aligned(heap, &_mi_heap_empty, sizeof(*heap));
+    heap->thread_id = _mi_thread_id();
+    _mi_random_init(&heap->random);
+    heap->cookie  = _mi_heap_random_next(heap) | 1;
+    heap->keys[0] = _mi_heap_random_next(heap);
+    heap->keys[1] = _mi_heap_random_next(heap);
+    heap->tld = tld;
+    tld->heap_backing = heap;
+    tld->heaps = heap;
+    tld->segments.stats = &tld->stats;
+    tld->segments.os = &tld->os;
+    tld->os.stats = &tld->stats;
+    _mi_heap_set_default_direct(heap);
+  }
+  return false;
+}
+
+// Free the thread local default heap (called from `mi_thread_done`)
+static bool _mi_heap_done(mi_heap_t* heap) {
+  if (!mi_heap_is_initialized(heap)) return true;
+
+  // reset default heap
+  _mi_heap_set_default_direct(_mi_is_main_thread() ? &_mi_heap_main : (mi_heap_t*)&_mi_heap_empty);
+
+  // switch to backing heap
+  heap = heap->tld->heap_backing;
+  if (!mi_heap_is_initialized(heap)) return false;
+
+  // delete all non-backing heaps in this thread
+  mi_heap_t* curr = heap->tld->heaps;
+  while (curr != NULL) {
+    mi_heap_t* next = curr->next; // save `next` as `curr` will be freed
+    if (curr != heap) {
+      mi_assert_internal(!mi_heap_is_backing(curr));
+      mi_heap_delete(curr);
+    }
+    curr = next;
+  }
+  mi_assert_internal(heap->tld->heaps == heap && heap->next == NULL);
+  mi_assert_internal(mi_heap_is_backing(heap));
+
+  // collect if not the main thread
+  if (heap != &_mi_heap_main) {
+    _mi_heap_collect_abandon(heap);
+  }
+
+  // merge stats
+  _mi_stats_done(&heap->tld->stats);
+
+  // free if not the main thread
+  if (heap != &_mi_heap_main) {
+    // the following assertion does not always hold for huge segments as those are always treated
+    // as abondened: one may allocate it in one thread, but deallocate in another in which case
+    // the count can be too large or negative. todo: perhaps not count huge segments? see issue #363
+    // mi_assert_internal(heap->tld->segments.count == 0 || heap->thread_id != _mi_thread_id());
+    mi_thread_data_free((mi_thread_data_t*)heap);
+  }
+  else {
+    mi_thread_data_collect(); // free cached thread metadata
+    #if 0
+    // never free the main thread even in debug mode; if a dll is linked statically with mimalloc,
+    // there may still be delete/free calls after the mi_fls_done is called. Issue #207
+    _mi_heap_destroy_pages(heap);
+    mi_assert_internal(heap->tld->heap_backing == &_mi_heap_main);
+    #endif
+  }
+  return false;
+}
+
+
+
+// --------------------------------------------------------
+// Try to run `mi_thread_done()` automatically so any memory
+// owned by the thread but not yet released can be abandoned
+// and re-owned by another thread.
+//
+// 1. windows dynamic library:
+//     call from DllMain on DLL_THREAD_DETACH
+// 2. windows static library:
+//     use `FlsAlloc` to call a destructor when the thread is done
+// 3. unix, pthreads:
+//     use a pthread key to call a destructor when a pthread is done
+//
+// In the last two cases we also need to call `mi_process_init`
+// to set up the thread local keys.
+// --------------------------------------------------------
+
+// Set up handlers so `mi_thread_done` is called automatically
+static void mi_process_setup_auto_thread_done(void) {
+  static bool tls_initialized = false; // fine if it races
+  if (tls_initialized) return;
+  tls_initialized = true;
+  _mi_prim_thread_init_auto_done();
+  _mi_heap_set_default_direct(&_mi_heap_main);
+}
+
+
+bool _mi_is_main_thread(void) {
+  return (_mi_heap_main.thread_id==0 || _mi_heap_main.thread_id == _mi_thread_id());
+}
+
+static _Atomic(size_t) thread_count = MI_ATOMIC_VAR_INIT(1);
+
+size_t  _mi_current_thread_count(void) {
+  return mi_atomic_load_relaxed(&thread_count);
+}
+
+// This is called from the `mi_malloc_generic`
+void mi_thread_init(void) mi_attr_noexcept
+{
+  // ensure our process has started already
+  mi_process_init();
+
+  // initialize the thread local default heap
+  // (this will call `_mi_heap_set_default_direct` and thus set the
+  //  fiber/pthread key to a non-zero value, ensuring `_mi_thread_done` is called)
+  if (_mi_heap_init()) return;  // returns true if already initialized
+
+  _mi_stat_increase(&_mi_stats_main.threads, 1);
+  mi_atomic_increment_relaxed(&thread_count);
+  //_mi_verbose_message("thread init: 0x%zx\n", _mi_thread_id());
+}
+
+void mi_thread_done(void) mi_attr_noexcept {
+  _mi_thread_done(NULL);
+}
+
+void _mi_thread_done(mi_heap_t* heap) 
+{
+  // calling with NULL implies using the default heap
+  if (heap == NULL) { 
+    heap = mi_prim_get_default_heap(); 
+    if (heap == NULL) return;
+  }
+
+  // prevent re-entrancy through heap_done/heap_set_default_direct (issue #699)
+  if (!mi_heap_is_initialized(heap)) {
+    return; 
+  }
+
+  // adjust stats
+  mi_atomic_decrement_relaxed(&thread_count);
+  _mi_stat_decrease(&_mi_stats_main.threads, 1);
+  
+  // check thread-id as on Windows shutdown with FLS the main (exit) thread may call this on thread-local heaps...
+  if (heap->thread_id != _mi_thread_id()) return;
+
+  // abandon the thread local heap
+  if (_mi_heap_done(heap)) return;  // returns true if already ran
+}
+
+void _mi_heap_set_default_direct(mi_heap_t* heap)  {
+  mi_assert_internal(heap != NULL);
+  #if defined(MI_TLS_SLOT)
+  mi_prim_tls_slot_set(MI_TLS_SLOT,heap);
+  #elif defined(MI_TLS_PTHREAD_SLOT_OFS)
+  *mi_tls_pthread_heap_slot() = heap;
+  #elif defined(MI_TLS_PTHREAD)
+  // we use _mi_heap_default_key
+  #else
+  _mi_heap_default = heap;
+  #endif
+
+  // ensure the default heap is passed to `_mi_thread_done`
+  // setting to a non-NULL value also ensures `mi_thread_done` is called.
+  _mi_prim_thread_associate_default_heap(heap);    
+}
+
+
+// --------------------------------------------------------
+// Run functions on process init/done, and thread init/done
+// --------------------------------------------------------
+static void mi_cdecl mi_process_done(void);
+
+static bool os_preloading = true;    // true until this module is initialized
+static bool mi_redirected = false;   // true if malloc redirects to mi_malloc
+
+// Returns true if this module has not been initialized; Don't use C runtime routines until it returns false.
+bool mi_decl_noinline _mi_preloading(void) {
+  return os_preloading;
+}
+
+mi_decl_nodiscard bool mi_is_redirected(void) mi_attr_noexcept {
+  return mi_redirected;
+}
+
+// Communicate with the redirection module on Windows
+#if defined(_WIN32) && defined(MI_SHARED_LIB) && !defined(MI_WIN_NOREDIRECT)
+#ifdef __cplusplus
+extern "C" {
+#endif
+mi_decl_export void _mi_redirect_entry(DWORD reason) {
+  // called on redirection; careful as this may be called before DllMain
+  if (reason == DLL_PROCESS_ATTACH) {
+    mi_redirected = true;
+  }
+  else if (reason == DLL_PROCESS_DETACH) {
+    mi_redirected = false;
+  }
+  else if (reason == DLL_THREAD_DETACH) {
+    mi_thread_done();
+  }
+}
+__declspec(dllimport) bool mi_cdecl mi_allocator_init(const char** message);
+__declspec(dllimport) void mi_cdecl mi_allocator_done(void);
+#ifdef __cplusplus
+}
+#endif
+#else
+static bool mi_allocator_init(const char** message) {
+  if (message != NULL) *message = NULL;
+  return true;
+}
+static void mi_allocator_done(void) {
+  // nothing to do
+}
+#endif
+
+// Called once by the process loader
+static void mi_process_load(void) {
+  mi_heap_main_init();
+  #if defined(__APPLE__) || defined(MI_TLS_RECURSE_GUARD)
+  volatile mi_heap_t* dummy = _mi_heap_default; // access TLS to allocate it before setting tls_initialized to true;
+  if (dummy == NULL) return;                    // use dummy or otherwise the access may get optimized away (issue #697)
+  #endif
+  os_preloading = false;
+  mi_assert_internal(_mi_is_main_thread());
+  #if !(defined(_WIN32) && defined(MI_SHARED_LIB))  // use Dll process detach (see below) instead of atexit (issue #521)
+  atexit(&mi_process_done);
+  #endif
+  _mi_options_init();
+  mi_process_setup_auto_thread_done();
+  mi_process_init();
+  if (mi_redirected) _mi_verbose_message("malloc is redirected.\n");
+
+  // show message from the redirector (if present)
+  const char* msg = NULL;
+  mi_allocator_init(&msg);
+  if (msg != NULL && (mi_option_is_enabled(mi_option_verbose) || mi_option_is_enabled(mi_option_show_errors))) {
+    _mi_fputs(NULL,NULL,NULL,msg);
+  }
+
+  // reseed random
+  _mi_random_reinit_if_weak(&_mi_heap_main.random);
+}
+
+#if defined(_WIN32) && (defined(_M_IX86) || defined(_M_X64))
+#include <intrin.h>
+mi_decl_cache_align bool _mi_cpu_has_fsrm = false;
+
+static void mi_detect_cpu_features(void) {
+  // FSRM for fast rep movsb support (AMD Zen3+ (~2020) or Intel Ice Lake+ (~2017))
+  int32_t cpu_info[4];
+  __cpuid(cpu_info, 7);
+  _mi_cpu_has_fsrm = ((cpu_info[3] & (1 << 4)) != 0); // bit 4 of EDX : see <https://en.wikipedia.org/wiki/CPUID#EAX=7,_ECX=0:_Extended_Features>
+}
+#else
+static void mi_detect_cpu_features(void) {
+  // nothing
+}
+#endif
+
+// Initialize the process; called by thread_init or the process loader
+void mi_process_init(void) mi_attr_noexcept {
+  // ensure we are called once
+  static mi_atomic_once_t process_init;
+  if (!mi_atomic_once(&process_init)) return;
+  _mi_process_is_initialized = true;
+  _mi_verbose_message("process init: 0x%zx\n", _mi_thread_id());
+  mi_process_setup_auto_thread_done();
+
+  mi_detect_cpu_features();
+  _mi_os_init();
+  mi_heap_main_init();
+  #if MI_DEBUG
+  _mi_verbose_message("debug level : %d\n", MI_DEBUG);
+  #endif
+  _mi_verbose_message("secure level: %d\n", MI_SECURE);
+  _mi_verbose_message("mem tracking: %s\n", MI_TRACK_TOOL);
+  #if MI_TSAN
+  _mi_verbose_message("thread santizer enabled\n");
+  #endif
+  mi_thread_init();
+
+  #if defined(_WIN32)
+  // On windows, when building as a static lib the FLS cleanup happens to early for the main thread.
+  // To avoid this, set the FLS value for the main thread to NULL so the fls cleanup
+  // will not call _mi_thread_done on the (still executing) main thread. See issue #508.
+  _mi_prim_thread_associate_default_heap(NULL);
+  #endif
+
+  mi_stats_reset();  // only call stat reset *after* thread init (or the heap tld == NULL)
+  mi_track_init();
+
+  if (mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
+    size_t pages = mi_option_get_clamp(mi_option_reserve_huge_os_pages, 0, 128*1024);
+    long reserve_at = mi_option_get(mi_option_reserve_huge_os_pages_at);
+    if (reserve_at != -1) {
+      mi_reserve_huge_os_pages_at(pages, reserve_at, pages*500);
+    } else {
+      mi_reserve_huge_os_pages_interleave(pages, 0, pages*500);
+    }
+  }
+  if (mi_option_is_enabled(mi_option_reserve_os_memory)) {
+    long ksize = mi_option_get(mi_option_reserve_os_memory);
+    if (ksize > 0) {
+      mi_reserve_os_memory((size_t)ksize*MI_KiB, true /* commit? */, true /* allow large pages? */);
+    }
+  }
+}
+
+// Called when the process is done (through `at_exit`)
+static void mi_cdecl mi_process_done(void) {
+  // only shutdown if we were initialized
+  if (!_mi_process_is_initialized) return;
+  // ensure we are called once
+  static bool process_done = false;
+  if (process_done) return;
+  process_done = true;
+
+  // release any thread specific resources and ensure _mi_thread_done is called on all but the main thread
+  _mi_prim_thread_done_auto_done();
+  
+  #ifndef MI_SKIP_COLLECT_ON_EXIT
+    #if (MI_DEBUG != 0) || !defined(MI_SHARED_LIB)
+    // free all memory if possible on process exit. This is not needed for a stand-alone process
+    // but should be done if mimalloc is statically linked into another shared library which
+    // is repeatedly loaded/unloaded, see issue #281.
+    mi_collect(true /* force */ );
+    #endif
+  #endif
+
+  // Forcefully release all retained memory; this can be dangerous in general if overriding regular malloc/free
+  // since after process_done there might still be other code running that calls `free` (like at_exit routines,
+  // or C-runtime termination code.
+  if (mi_option_is_enabled(mi_option_destroy_on_exit)) {
+    _mi_heap_destroy_all();                          // forcefully release all memory held by all heaps (of this thread only!)
+    _mi_segment_cache_free_all(&_mi_heap_main_get()->tld->os);  // release all cached segments
+  }
+
+  if (mi_option_is_enabled(mi_option_show_stats) || mi_option_is_enabled(mi_option_verbose)) {
+    mi_stats_print(NULL);
+  }
+  mi_allocator_done();
+  _mi_verbose_message("process done: 0x%zx\n", _mi_heap_main.thread_id);
+  os_preloading = true; // don't call the C runtime anymore
+}
+
+
+
+#if defined(_WIN32) && defined(MI_SHARED_LIB)
+  // Windows DLL: easy to hook into process_init and thread_done
+  __declspec(dllexport) BOOL WINAPI DllMain(HINSTANCE inst, DWORD reason, LPVOID reserved) {
+    MI_UNUSED(reserved);
+    MI_UNUSED(inst);
+    if (reason==DLL_PROCESS_ATTACH) {
+      mi_process_load();
+    }
+    else if (reason==DLL_PROCESS_DETACH) {
+      mi_process_done();
+    }
+    else if (reason==DLL_THREAD_DETACH) {
+      if (!mi_is_redirected()) {
+        mi_thread_done();
+      }
+    }
+    return TRUE;
+  }
+
+#elif defined(_MSC_VER)
+  // MSVC: use data section magic for static libraries
+  // See <https://www.codeguru.com/cpp/misc/misc/applicationcontrol/article.php/c6945/Running-Code-Before-and-After-Main.htm>
+  static int _mi_process_init(void) {
+    mi_process_load();
+    return 0;
+  }
+  typedef int(*_mi_crt_callback_t)(void);
+  #if defined(_M_X64) || defined(_M_ARM64)
+    __pragma(comment(linker, "/include:" "_mi_msvc_initu"))
+    #pragma section(".CRT$XIU", long, read)
+  #else
+    __pragma(comment(linker, "/include:" "__mi_msvc_initu"))
+  #endif
+  #pragma data_seg(".CRT$XIU")
+  mi_decl_externc _mi_crt_callback_t _mi_msvc_initu[] = { &_mi_process_init };
+  #pragma data_seg()
+
+#elif defined(__cplusplus)
+  // C++: use static initialization to detect process start
+  static bool _mi_process_init(void) {
+    mi_process_load();
+    return (_mi_heap_main.thread_id != 0);
+  }
+  static bool mi_initialized = _mi_process_init();
+
+#elif defined(__GNUC__) || defined(__clang__)
+  // GCC,Clang: use the constructor attribute
+  static void __attribute__((constructor)) _mi_process_init(void) {
+    mi_process_load();
+  }
+
+#else
+#pragma message("define a way to call mi_process_load on your platform")
+#endif

preload-mimalloc/mimalloc/src/options.c

@@ -0,0 +1,563 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"  // mi_prim_out_stderr
+
+#include <stdio.h>      // FILE
+#include <stdlib.h>     // abort
+#include <stdarg.h>
+
+
+static long mi_max_error_count   = 16; // stop outputting errors after this (use < 0 for no limit)
+static long mi_max_warning_count = 16; // stop outputting warnings after this (use < 0 for no limit)
+
+static void mi_add_stderr_output(void);
+
+int mi_version(void) mi_attr_noexcept {
+  return MI_MALLOC_VERSION;
+}
+
+
+// --------------------------------------------------------
+// Options
+// These can be accessed by multiple threads and may be
+// concurrently initialized, but an initializing data race
+// is ok since they resolve to the same value.
+// --------------------------------------------------------
+typedef enum mi_init_e {
+  UNINIT,       // not yet initialized
+  DEFAULTED,    // not found in the environment, use default value
+  INITIALIZED   // found in environment or set explicitly
+} mi_init_t;
+
+typedef struct mi_option_desc_s {
+  long        value;  // the value
+  mi_init_t   init;   // is it initialized yet? (from the environment)
+  mi_option_t option; // for debugging: the option index should match the option
+  const char* name;   // option name without `mimalloc_` prefix
+  const char* legacy_name; // potential legacy v1.x option name
+} mi_option_desc_t;
+
+#define MI_OPTION(opt)                  mi_option_##opt, #opt, NULL
+#define MI_OPTION_LEGACY(opt,legacy)    mi_option_##opt, #opt, #legacy
+
+static mi_option_desc_t options[_mi_option_last] =
+{
+  // stable options
+  #if MI_DEBUG || defined(MI_SHOW_ERRORS)
+  { 1, UNINIT, MI_OPTION(show_errors) },
+  #else
+  { 0, UNINIT, MI_OPTION(show_errors) },
+  #endif
+  { 0, UNINIT, MI_OPTION(show_stats) },
+  { 0, UNINIT, MI_OPTION(verbose) },
+
+  // Some of the following options are experimental and not all combinations are valid. Use with care.
+  { 1, UNINIT, MI_OPTION(eager_commit) },        // commit per segment directly (8MiB)  (but see also `eager_commit_delay`)
+  { 0, UNINIT, MI_OPTION(deprecated_eager_region_commit) },
+  { 0, UNINIT, MI_OPTION(deprecated_reset_decommits) },
+  { 0, UNINIT, MI_OPTION(large_os_pages) },      // use large OS pages, use only with eager commit to prevent fragmentation of VMA's
+  { 0, UNINIT, MI_OPTION(reserve_huge_os_pages) },  // per 1GiB huge pages
+  { -1, UNINIT, MI_OPTION(reserve_huge_os_pages_at) }, // reserve huge pages at node N
+  { 0, UNINIT, MI_OPTION(reserve_os_memory)     },
+  { 0, UNINIT, MI_OPTION(deprecated_segment_cache) },  // cache N segments per thread
+  { 0, UNINIT, MI_OPTION(page_reset) },          // reset page memory on free
+  { 0, UNINIT, MI_OPTION_LEGACY(abandoned_page_decommit, abandoned_page_reset) },// decommit free page memory when a thread terminates  
+  { 0, UNINIT, MI_OPTION(deprecated_segment_reset) },
+  #if defined(__NetBSD__)
+  { 0, UNINIT, MI_OPTION(eager_commit_delay) },  // the first N segments per thread are not eagerly committed
+  #elif defined(_WIN32)
+  { 4, UNINIT, MI_OPTION(eager_commit_delay) },  // the first N segments per thread are not eagerly committed (but per page in the segment on demand)
+  #else
+  { 1, UNINIT, MI_OPTION(eager_commit_delay) },  // the first N segments per thread are not eagerly committed (but per page in the segment on demand)
+  #endif
+  { 25,   UNINIT, MI_OPTION_LEGACY(decommit_delay, reset_delay) }, // page decommit delay in milli-seconds
+  { 0,    UNINIT, MI_OPTION(use_numa_nodes) },    // 0 = use available numa nodes, otherwise use at most N nodes. 
+  { 0,    UNINIT, MI_OPTION(limit_os_alloc) },    // 1 = do not use OS memory for allocation (but only reserved arenas)
+  { 100,  UNINIT, MI_OPTION(os_tag) },            // only apple specific for now but might serve more or less related purpose
+  { 16,   UNINIT, MI_OPTION(max_errors) },        // maximum errors that are output
+  { 16,   UNINIT, MI_OPTION(max_warnings) },      // maximum warnings that are output
+  { 8,    UNINIT, MI_OPTION(max_segment_reclaim)},// max. number of segment reclaims from the abandoned segments per try.  
+  { 1,    UNINIT, MI_OPTION(allow_decommit) },    // decommit slices when no longer used (after decommit_delay milli-seconds)
+  { 500,  UNINIT, MI_OPTION(segment_decommit_delay) }, // decommit delay in milli-seconds for freed segments
+  { 1,    UNINIT, MI_OPTION(decommit_extend_delay) },
+  { 0,    UNINIT, MI_OPTION(destroy_on_exit)}     // release all OS memory on process exit; careful with dangling pointer or after-exit frees!
+};
+
+static void mi_option_init(mi_option_desc_t* desc);
+
+void _mi_options_init(void) {
+  // called on process load; should not be called before the CRT is initialized!
+  // (e.g. do not call this from process_init as that may run before CRT initialization)
+  mi_add_stderr_output(); // now it safe to use stderr for output
+  for(int i = 0; i < _mi_option_last; i++ ) {
+    mi_option_t option = (mi_option_t)i;
+    long l = mi_option_get(option); MI_UNUSED(l); // initialize
+    // if (option != mi_option_verbose)
+    {
+      mi_option_desc_t* desc = &options[option];
+      _mi_verbose_message("option '%s': %ld\n", desc->name, desc->value);
+    }
+  }
+  mi_max_error_count = mi_option_get(mi_option_max_errors);
+  mi_max_warning_count = mi_option_get(mi_option_max_warnings);
+}
+
+mi_decl_nodiscard long mi_option_get(mi_option_t option) {
+  mi_assert(option >= 0 && option < _mi_option_last);
+  if (option < 0 || option >= _mi_option_last) return 0;
+  mi_option_desc_t* desc = &options[option];
+  mi_assert(desc->option == option);  // index should match the option
+  if mi_unlikely(desc->init == UNINIT) {
+    mi_option_init(desc);
+  }
+  return desc->value;
+}
+
+mi_decl_nodiscard long mi_option_get_clamp(mi_option_t option, long min, long max) {
+  long x = mi_option_get(option);
+  return (x < min ? min : (x > max ? max : x));
+}
+
+void mi_option_set(mi_option_t option, long value) {
+  mi_assert(option >= 0 && option < _mi_option_last);
+  if (option < 0 || option >= _mi_option_last) return;
+  mi_option_desc_t* desc = &options[option];
+  mi_assert(desc->option == option);  // index should match the option
+  desc->value = value;
+  desc->init = INITIALIZED;
+}
+
+void mi_option_set_default(mi_option_t option, long value) {
+  mi_assert(option >= 0 && option < _mi_option_last);
+  if (option < 0 || option >= _mi_option_last) return;
+  mi_option_desc_t* desc = &options[option];
+  if (desc->init != INITIALIZED) {
+    desc->value = value;
+  }
+}
+
+mi_decl_nodiscard bool mi_option_is_enabled(mi_option_t option) {
+  return (mi_option_get(option) != 0);
+}
+
+void mi_option_set_enabled(mi_option_t option, bool enable) {
+  mi_option_set(option, (enable ? 1 : 0));
+}
+
+void mi_option_set_enabled_default(mi_option_t option, bool enable) {
+  mi_option_set_default(option, (enable ? 1 : 0));
+}
+
+void mi_option_enable(mi_option_t option) {
+  mi_option_set_enabled(option,true);
+}
+
+void mi_option_disable(mi_option_t option) {
+  mi_option_set_enabled(option,false);
+}
+
+static void mi_cdecl mi_out_stderr(const char* msg, void* arg) {
+  MI_UNUSED(arg);
+  if (msg != NULL && msg[0] != 0) {
+    _mi_prim_out_stderr(msg);
+  }
+}
+
+// Since an output function can be registered earliest in the `main`
+// function we also buffer output that happens earlier. When
+// an output function is registered it is called immediately with
+// the output up to that point.
+#ifndef MI_MAX_DELAY_OUTPUT
+#define MI_MAX_DELAY_OUTPUT ((size_t)(32*1024))
+#endif
+static char out_buf[MI_MAX_DELAY_OUTPUT+1];
+static _Atomic(size_t) out_len;
+
+static void mi_cdecl mi_out_buf(const char* msg, void* arg) {
+  MI_UNUSED(arg);
+  if (msg==NULL) return;
+  if (mi_atomic_load_relaxed(&out_len)>=MI_MAX_DELAY_OUTPUT) return;
+  size_t n = _mi_strlen(msg);
+  if (n==0) return;
+  // claim space
+  size_t start = mi_atomic_add_acq_rel(&out_len, n);
+  if (start >= MI_MAX_DELAY_OUTPUT) return;
+  // check bound
+  if (start+n >= MI_MAX_DELAY_OUTPUT) {
+    n = MI_MAX_DELAY_OUTPUT-start-1;
+  }
+  _mi_memcpy(&out_buf[start], msg, n);
+}
+
+static void mi_out_buf_flush(mi_output_fun* out, bool no_more_buf, void* arg) {
+  if (out==NULL) return;
+  // claim (if `no_more_buf == true`, no more output will be added after this point)
+  size_t count = mi_atomic_add_acq_rel(&out_len, (no_more_buf ? MI_MAX_DELAY_OUTPUT : 1));
+  // and output the current contents
+  if (count>MI_MAX_DELAY_OUTPUT) count = MI_MAX_DELAY_OUTPUT;
+  out_buf[count] = 0;
+  out(out_buf,arg);
+  if (!no_more_buf) {
+    out_buf[count] = '\n'; // if continue with the buffer, insert a newline
+  }
+}
+
+
+// Once this module is loaded, switch to this routine
+// which outputs to stderr and the delayed output buffer.
+static void mi_cdecl mi_out_buf_stderr(const char* msg, void* arg) {
+  mi_out_stderr(msg,arg);
+  mi_out_buf(msg,arg);
+}
+
+
+
+// --------------------------------------------------------
+// Default output handler
+// --------------------------------------------------------
+
+// Should be atomic but gives errors on many platforms as generally we cannot cast a function pointer to a uintptr_t.
+// For now, don't register output from multiple threads.
+static mi_output_fun* volatile mi_out_default; // = NULL
+static _Atomic(void*) mi_out_arg; // = NULL
+
+static mi_output_fun* mi_out_get_default(void** parg) {
+  if (parg != NULL) { *parg = mi_atomic_load_ptr_acquire(void,&mi_out_arg); }
+  mi_output_fun* out = mi_out_default;
+  return (out == NULL ? &mi_out_buf : out);
+}
+
+void mi_register_output(mi_output_fun* out, void* arg) mi_attr_noexcept {
+  mi_out_default = (out == NULL ? &mi_out_stderr : out); // stop using the delayed output buffer
+  mi_atomic_store_ptr_release(void,&mi_out_arg, arg);
+  if (out!=NULL) mi_out_buf_flush(out,true,arg);         // output all the delayed output now
+}
+
+// add stderr to the delayed output after the module is loaded
+static void mi_add_stderr_output() {
+  mi_assert_internal(mi_out_default == NULL);
+  mi_out_buf_flush(&mi_out_stderr, false, NULL); // flush current contents to stderr
+  mi_out_default = &mi_out_buf_stderr;           // and add stderr to the delayed output
+}
+
+// --------------------------------------------------------
+// Messages, all end up calling `_mi_fputs`.
+// --------------------------------------------------------
+static _Atomic(size_t) error_count;   // = 0;  // when >= max_error_count stop emitting errors
+static _Atomic(size_t) warning_count; // = 0;  // when >= max_warning_count stop emitting warnings
+
+// When overriding malloc, we may recurse into mi_vfprintf if an allocation
+// inside the C runtime causes another message.
+// In some cases (like on macOS) the loader already allocates which
+// calls into mimalloc; if we then access thread locals (like `recurse`)
+// this may crash as the access may call _tlv_bootstrap that tries to
+// (recursively) invoke malloc again to allocate space for the thread local
+// variables on demand. This is why we use a _mi_preloading test on such
+// platforms. However, C code generator may move the initial thread local address
+// load before the `if` and we therefore split it out in a separate funcion.
+static mi_decl_thread bool recurse = false;
+
+static mi_decl_noinline bool mi_recurse_enter_prim(void) {
+  if (recurse) return false;
+  recurse = true;
+  return true;
+}
+
+static mi_decl_noinline void mi_recurse_exit_prim(void) {
+  recurse = false;
+}
+
+static bool mi_recurse_enter(void) {
+  #if defined(__APPLE__) || defined(MI_TLS_RECURSE_GUARD)
+  if (_mi_preloading()) return false;
+  #endif
+  return mi_recurse_enter_prim();
+}
+
+static void mi_recurse_exit(void) {
+  #if defined(__APPLE__) || defined(MI_TLS_RECURSE_GUARD)
+  if (_mi_preloading()) return;
+  #endif
+  mi_recurse_exit_prim();
+}
+
+void _mi_fputs(mi_output_fun* out, void* arg, const char* prefix, const char* message) {
+  if (out==NULL || (void*)out==(void*)stdout || (void*)out==(void*)stderr) { // TODO: use mi_out_stderr for stderr?
+    if (!mi_recurse_enter()) return;
+    out = mi_out_get_default(&arg);
+    if (prefix != NULL) out(prefix, arg);
+    out(message, arg);
+    mi_recurse_exit();
+  }
+  else {
+    if (prefix != NULL) out(prefix, arg);
+    out(message, arg);
+  }
+}
+
+// Define our own limited `fprintf` that avoids memory allocation.
+// We do this using `snprintf` with a limited buffer.
+static void mi_vfprintf( mi_output_fun* out, void* arg, const char* prefix, const char* fmt, va_list args ) {
+  char buf[512];
+  if (fmt==NULL) return;
+  if (!mi_recurse_enter()) return;
+  vsnprintf(buf,sizeof(buf)-1,fmt,args);
+  mi_recurse_exit();
+  _mi_fputs(out,arg,prefix,buf);
+}
+
+void _mi_fprintf( mi_output_fun* out, void* arg, const char* fmt, ... ) {
+  va_list args;
+  va_start(args,fmt);
+  mi_vfprintf(out,arg,NULL,fmt,args);
+  va_end(args);
+}
+
+static void mi_vfprintf_thread(mi_output_fun* out, void* arg, const char* prefix, const char* fmt, va_list args) {
+  if (prefix != NULL && _mi_strnlen(prefix,33) <= 32 && !_mi_is_main_thread()) {
+    char tprefix[64];
+    snprintf(tprefix, sizeof(tprefix), "%sthread 0x%llx: ", prefix, (unsigned long long)_mi_thread_id());
+    mi_vfprintf(out, arg, tprefix, fmt, args);
+  }
+  else {
+    mi_vfprintf(out, arg, prefix, fmt, args);
+  }
+}
+
+void _mi_trace_message(const char* fmt, ...) {
+  if (mi_option_get(mi_option_verbose) <= 1) return;  // only with verbose level 2 or higher
+  va_list args;
+  va_start(args, fmt);
+  mi_vfprintf_thread(NULL, NULL, "mimalloc: ", fmt, args);
+  va_end(args);
+}
+
+void _mi_verbose_message(const char* fmt, ...) {
+  if (!mi_option_is_enabled(mi_option_verbose)) return;
+  va_list args;
+  va_start(args,fmt);
+  mi_vfprintf(NULL, NULL, "mimalloc: ", fmt, args);
+  va_end(args);
+}
+
+static void mi_show_error_message(const char* fmt, va_list args) {
+  if (!mi_option_is_enabled(mi_option_verbose)) {
+    if (!mi_option_is_enabled(mi_option_show_errors)) return;
+    if (mi_max_error_count >= 0 && (long)mi_atomic_increment_acq_rel(&error_count) > mi_max_error_count) return;
+  }
+  mi_vfprintf_thread(NULL, NULL, "mimalloc: error: ", fmt, args);
+}
+
+void _mi_warning_message(const char* fmt, ...) {
+  if (!mi_option_is_enabled(mi_option_verbose)) {
+    if (!mi_option_is_enabled(mi_option_show_errors)) return;
+    if (mi_max_warning_count >= 0 && (long)mi_atomic_increment_acq_rel(&warning_count) > mi_max_warning_count) return;
+  }
+  va_list args;
+  va_start(args,fmt);
+  mi_vfprintf_thread(NULL, NULL, "mimalloc: warning: ", fmt, args);
+  va_end(args);
+}
+
+
+#if MI_DEBUG
+void _mi_assert_fail(const char* assertion, const char* fname, unsigned line, const char* func ) {
+  _mi_fprintf(NULL, NULL, "mimalloc: assertion failed: at \"%s\":%u, %s\n  assertion: \"%s\"\n", fname, line, (func==NULL?"":func), assertion);
+  abort();
+}
+#endif
+
+// --------------------------------------------------------
+// Errors
+// --------------------------------------------------------
+
+static mi_error_fun* volatile  mi_error_handler; // = NULL
+static _Atomic(void*) mi_error_arg;     // = NULL
+
+static void mi_error_default(int err) {
+  MI_UNUSED(err);
+#if (MI_DEBUG>0)
+  if (err==EFAULT) {
+    #ifdef _MSC_VER
+    __debugbreak();
+    #endif
+    abort();
+  }
+#endif
+#if (MI_SECURE>0)
+  if (err==EFAULT) {  // abort on serious errors in secure mode (corrupted meta-data)
+    abort();
+  }
+#endif
+#if defined(MI_XMALLOC)
+  if (err==ENOMEM || err==EOVERFLOW) { // abort on memory allocation fails in xmalloc mode
+    abort();
+  }
+#endif
+}
+
+void mi_register_error(mi_error_fun* fun, void* arg) {
+  mi_error_handler = fun;  // can be NULL
+  mi_atomic_store_ptr_release(void,&mi_error_arg, arg);
+}
+
+void _mi_error_message(int err, const char* fmt, ...) {
+  // show detailed error message
+  va_list args;
+  va_start(args, fmt);
+  mi_show_error_message(fmt, args);
+  va_end(args);
+  // and call the error handler which may abort (or return normally)
+  if (mi_error_handler != NULL) {
+    mi_error_handler(err, mi_atomic_load_ptr_acquire(void,&mi_error_arg));
+  }
+  else {
+    mi_error_default(err);
+  }
+}
+
+// --------------------------------------------------------
+// Initialize options by checking the environment
+// --------------------------------------------------------
+char _mi_toupper(char c) {
+  if (c >= 'a' && c <= 'z') return (c - 'a' + 'A');
+                       else return c;
+}
+
+int _mi_strnicmp(const char* s, const char* t, size_t n) {
+  if (n == 0) return 0;
+  for (; *s != 0 && *t != 0 && n > 0; s++, t++, n--) {
+    if (_mi_toupper(*s) != _mi_toupper(*t)) break;
+  }
+  return (n == 0 ? 0 : *s - *t);
+}
+
+void _mi_strlcpy(char* dest, const char* src, size_t dest_size) {
+  if (dest==NULL || src==NULL || dest_size == 0) return;
+  // copy until end of src, or when dest is (almost) full
+  while (*src != 0 && dest_size > 1) {
+    *dest++ = *src++;
+    dest_size--;
+  }
+  // always zero terminate
+  *dest = 0;
+}
+
+void _mi_strlcat(char* dest, const char* src, size_t dest_size) {
+  if (dest==NULL || src==NULL || dest_size == 0) return;
+  // find end of string in the dest buffer
+  while (*dest != 0 && dest_size > 1) {
+    dest++;
+    dest_size--;
+  }
+  // and catenate
+  _mi_strlcpy(dest, src, dest_size);
+}
+
+size_t _mi_strlen(const char* s) {
+  if (s==NULL) return 0;
+  size_t len = 0;
+  while(s[len] != 0) { len++; }
+  return len;
+}
+
+size_t _mi_strnlen(const char* s, size_t max_len) {
+  if (s==NULL) return 0;
+  size_t len = 0;
+  while(s[len] != 0 && len < max_len) { len++; }
+  return len;
+}
+
+#ifdef MI_NO_GETENV
+static bool mi_getenv(const char* name, char* result, size_t result_size) {
+  MI_UNUSED(name);
+  MI_UNUSED(result);
+  MI_UNUSED(result_size);
+  return false;
+}
+#else
+static bool mi_getenv(const char* name, char* result, size_t result_size) {
+  if (name==NULL || result == NULL || result_size < 64) return false;
+  return _mi_prim_getenv(name,result,result_size);
+}
+#endif
+
+// TODO: implement ourselves to reduce dependencies on the C runtime
+#include <stdlib.h> // strtol
+#include <string.h> // strstr
+
+
+static void mi_option_init(mi_option_desc_t* desc) {
+  // Read option value from the environment
+  char s[64+1];
+  char buf[64+1];
+  _mi_strlcpy(buf, "mimalloc_", sizeof(buf));
+  _mi_strlcat(buf, desc->name, sizeof(buf));
+  bool found = mi_getenv(buf,s,sizeof(s));
+  if (!found && desc->legacy_name != NULL) {
+    _mi_strlcpy(buf, "mimalloc_", sizeof(buf));
+    _mi_strlcat(buf, desc->legacy_name, sizeof(buf));
+    found = mi_getenv(buf,s,sizeof(s));
+    if (found) {
+      _mi_warning_message("environment option \"mimalloc_%s\" is deprecated -- use \"mimalloc_%s\" instead.\n", desc->legacy_name, desc->name );
+    }    
+  }
+
+  if (found) {
+    size_t len = _mi_strnlen(s,sizeof(buf)-1);
+    for (size_t i = 0; i < len; i++) {
+      buf[i] = _mi_toupper(s[i]);
+    }
+    buf[len] = 0;
+    if (buf[0]==0 || strstr("1;TRUE;YES;ON", buf) != NULL) {
+      desc->value = 1;
+      desc->init = INITIALIZED;
+    }
+    else if (strstr("0;FALSE;NO;OFF", buf) != NULL) {
+      desc->value = 0;
+      desc->init = INITIALIZED;
+    }
+    else {
+      char* end = buf;
+      long value = strtol(buf, &end, 10);
+      if (desc->option == mi_option_reserve_os_memory) {
+        // this option is interpreted in KiB to prevent overflow of `long`
+        if (*end == 'K') { end++; }
+        else if (*end == 'M') { value *= MI_KiB; end++; }
+        else if (*end == 'G') { value *= MI_MiB; end++; }
+        else { value = (value + MI_KiB - 1) / MI_KiB; }
+        if (end[0] == 'I' && end[1] == 'B') { end += 2; }
+        else if (*end == 'B') { end++; }
+      }
+      if (*end == 0) {
+        desc->value = value;
+        desc->init = INITIALIZED;
+      }
+      else {
+        // set `init` first to avoid recursion through _mi_warning_message on mimalloc_verbose.
+        desc->init = DEFAULTED;
+        if (desc->option == mi_option_verbose && desc->value == 0) {
+          // if the 'mimalloc_verbose' env var has a bogus value we'd never know
+          // (since the value defaults to 'off') so in that case briefly enable verbose
+          desc->value = 1;
+          _mi_warning_message("environment option mimalloc_%s has an invalid value.\n", desc->name );
+          desc->value = 0;
+        }
+        else {
+          _mi_warning_message("environment option mimalloc_%s has an invalid value.\n", desc->name );
+        }
+      }
+    }
+    mi_assert_internal(desc->init != UNINIT);
+  }
+  else if (!_mi_preloading()) {
+    desc->init = DEFAULTED;
+  }
+}

preload-mimalloc/mimalloc/src/os.c

@@ -0,0 +1,610 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+
+
+/* -----------------------------------------------------------
+  Initialization.
+  On windows initializes support for aligned allocation and
+  large OS pages (if MIMALLOC_LARGE_OS_PAGES is true).
+----------------------------------------------------------- */
+
+static mi_os_mem_config_t mi_os_mem_config = {
+  4096,   // page size
+  0,      // large page size (usually 2MiB)
+  4096,   // allocation granularity
+  true,   // has overcommit?  (if true we use MAP_NORESERVE on mmap systems)
+  false   // must free whole? (on mmap systems we can free anywhere in a mapped range, but on Windows we must free the entire span)
+};
+
+bool _mi_os_has_overcommit(void) {
+  return mi_os_mem_config.has_overcommit;
+}
+
+// OS (small) page size
+size_t _mi_os_page_size(void) {
+  return mi_os_mem_config.page_size;
+}
+
+// if large OS pages are supported (2 or 4MiB), then return the size, otherwise return the small page size (4KiB)
+size_t _mi_os_large_page_size(void) {
+  return (mi_os_mem_config.large_page_size != 0 ? mi_os_mem_config.large_page_size : _mi_os_page_size());
+}
+
+bool _mi_os_use_large_page(size_t size, size_t alignment) {
+  // if we have access, check the size and alignment requirements
+  if (mi_os_mem_config.large_page_size == 0 || !mi_option_is_enabled(mi_option_large_os_pages)) return false;
+  return ((size % mi_os_mem_config.large_page_size) == 0 && (alignment % mi_os_mem_config.large_page_size) == 0);
+}
+
+// round to a good OS allocation size (bounded by max 12.5% waste)
+size_t _mi_os_good_alloc_size(size_t size) {
+  size_t align_size;
+  if (size < 512*MI_KiB) align_size = _mi_os_page_size();
+  else if (size < 2*MI_MiB) align_size = 64*MI_KiB;
+  else if (size < 8*MI_MiB) align_size = 256*MI_KiB;
+  else if (size < 32*MI_MiB) align_size = 1*MI_MiB;
+  else align_size = 4*MI_MiB;
+  if mi_unlikely(size >= (SIZE_MAX - align_size)) return size; // possible overflow?
+  return _mi_align_up(size, align_size);
+}
+
+void _mi_os_init(void) {
+  _mi_prim_mem_init(&mi_os_mem_config);
+}
+
+
+/* -----------------------------------------------------------
+  Util
+-------------------------------------------------------------- */
+bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* stats);
+bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats);
+
+static void* mi_align_up_ptr(void* p, size_t alignment) {
+  return (void*)_mi_align_up((uintptr_t)p, alignment);
+}
+
+static void* mi_align_down_ptr(void* p, size_t alignment) {
+  return (void*)_mi_align_down((uintptr_t)p, alignment);
+}
+
+
+/* -----------------------------------------------------------
+  aligned hinting
+-------------------------------------------------------------- */
+
+// On 64-bit systems, we can do efficient aligned allocation by using
+// the 2TiB to 30TiB area to allocate those.
+#if (MI_INTPTR_SIZE >= 8)
+static mi_decl_cache_align _Atomic(uintptr_t)aligned_base;
+
+// Return a MI_SEGMENT_SIZE aligned address that is probably available.
+// If this returns NULL, the OS will determine the address but on some OS's that may not be
+// properly aligned which can be more costly as it needs to be adjusted afterwards.
+// For a size > 1GiB this always returns NULL in order to guarantee good ASLR randomization;
+// (otherwise an initial large allocation of say 2TiB has a 50% chance to include (known) addresses
+//  in the middle of the 2TiB - 6TiB address range (see issue #372))
+
+#define MI_HINT_BASE ((uintptr_t)2 << 40)  // 2TiB start
+#define MI_HINT_AREA ((uintptr_t)4 << 40)  // upto 6TiB   (since before win8 there is "only" 8TiB available to processes)
+#define MI_HINT_MAX  ((uintptr_t)30 << 40) // wrap after 30TiB (area after 32TiB is used for huge OS pages)
+
+void* _mi_os_get_aligned_hint(size_t try_alignment, size_t size)
+{
+  if (try_alignment <= 1 || try_alignment > MI_SEGMENT_SIZE) return NULL;
+  size = _mi_align_up(size, MI_SEGMENT_SIZE);
+  if (size > 1*MI_GiB) return NULL;  // guarantee the chance of fixed valid address is at most 1/(MI_HINT_AREA / 1<<30) = 1/4096.
+  #if (MI_SECURE>0)
+  size += MI_SEGMENT_SIZE;        // put in `MI_SEGMENT_SIZE` virtual gaps between hinted blocks; this splits VLA's but increases guarded areas.
+  #endif
+
+  uintptr_t hint = mi_atomic_add_acq_rel(&aligned_base, size);
+  if (hint == 0 || hint > MI_HINT_MAX) {   // wrap or initialize
+    uintptr_t init = MI_HINT_BASE;
+    #if (MI_SECURE>0 || MI_DEBUG==0)       // security: randomize start of aligned allocations unless in debug mode
+    uintptr_t r = _mi_heap_random_next(mi_prim_get_default_heap());
+    init = init + ((MI_SEGMENT_SIZE * ((r>>17) & 0xFFFFF)) % MI_HINT_AREA);  // (randomly 20 bits)*4MiB == 0 to 4TiB
+    #endif
+    uintptr_t expected = hint + size;
+    mi_atomic_cas_strong_acq_rel(&aligned_base, &expected, init);
+    hint = mi_atomic_add_acq_rel(&aligned_base, size); // this may still give 0 or > MI_HINT_MAX but that is ok, it is a hint after all
+  }
+  if (hint%try_alignment != 0) return NULL;
+  return (void*)hint;
+}
+#else
+void* _mi_os_get_aligned_hint(size_t try_alignment, size_t size) {
+  MI_UNUSED(try_alignment); MI_UNUSED(size);
+  return NULL;
+}
+#endif
+
+
+/* -----------------------------------------------------------
+  Free memory
+-------------------------------------------------------------- */
+
+static void mi_os_mem_free(void* addr, size_t size, bool was_committed, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_assert_internal((size % _mi_os_page_size()) == 0);
+  if (addr == NULL || size == 0) return; // || _mi_os_is_huge_reserved(addr)
+  int err = _mi_prim_free(addr, size);
+  if (err != 0) {
+    _mi_warning_message("unable to free OS memory (error: %d (0x%x), size: 0x%zx bytes, address: %p)\n", err, err, size, addr);
+  }
+  mi_stats_t* stats = &_mi_stats_main;
+  if (was_committed) { _mi_stat_decrease(&stats->committed, size); }
+  _mi_stat_decrease(&stats->reserved, size);
+}
+
+
+void _mi_os_free_ex(void* addr, size_t size, bool was_committed, mi_stats_t* tld_stats) {
+  const size_t csize = _mi_os_good_alloc_size(size);
+  mi_os_mem_free(addr,csize,was_committed,tld_stats);
+}
+
+void  _mi_os_free(void* p, size_t size, mi_stats_t* tld_stats) {
+  _mi_os_free_ex(p, size, true, tld_stats);
+}
+
+
+/* -----------------------------------------------------------
+   Primitive allocation from the OS.
+-------------------------------------------------------------- */
+
+// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
+static void* mi_os_mem_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, mi_stats_t* stats) {
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  if (size == 0) return NULL;
+  if (!commit) allow_large = false;
+  if (try_alignment == 0) try_alignment = 1; // avoid 0 to ensure there will be no divide by zero when aligning
+
+  void* p = NULL; 
+  int err = _mi_prim_alloc(size, try_alignment, commit, allow_large, is_large, &p);
+  if (err != 0) {
+    _mi_warning_message("unable to allocate OS memory (error: %d (0x%x), size: 0x%zx bytes, align: 0x%zx, commit: %d, allow large: %d)\n", err, err, size, try_alignment, commit, allow_large);
+  }
+  /*
+  if (commit && allow_large) {
+    p = _mi_os_try_alloc_from_huge_reserved(size, try_alignment);
+    if (p != NULL) {
+      *is_large = true;
+      return p;
+    }
+  }
+  */
+
+  mi_stat_counter_increase(stats->mmap_calls, 1);
+  if (p != NULL) {
+    _mi_stat_increase(&stats->reserved, size);
+    if (commit) { _mi_stat_increase(&stats->committed, size); }
+  }
+  return p;
+}
+
+
+// Primitive aligned allocation from the OS.
+// This function guarantees the allocated memory is aligned.
+static void* mi_os_mem_alloc_aligned(size_t size, size_t alignment, bool commit, bool allow_large, bool* is_large, mi_stats_t* stats) {
+  mi_assert_internal(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0));
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  mi_assert_internal(is_large != NULL);
+  if (!commit) allow_large = false;
+  if (!(alignment >= _mi_os_page_size() && ((alignment & (alignment - 1)) == 0))) return NULL;
+  size = _mi_align_up(size, _mi_os_page_size());
+
+  // try first with a hint (this will be aligned directly on Win 10+ or BSD)
+  void* p = mi_os_mem_alloc(size, alignment, commit, allow_large, is_large, stats);
+  if (p == NULL) return NULL;
+
+  // if not aligned, free it, overallocate, and unmap around it
+  if (((uintptr_t)p % alignment != 0)) {
+    mi_os_mem_free(p, size, commit, stats);
+    _mi_warning_message("unable to allocate aligned OS memory directly, fall back to over-allocation (size: 0x%zx bytes, address: %p, alignment: 0x%zx, commit: %d)\n", size, p, alignment, commit);
+    if (size >= (SIZE_MAX - alignment)) return NULL; // overflow
+    const size_t over_size = size + alignment;
+
+    if (mi_os_mem_config.must_free_whole) {  // win32 virtualAlloc cannot free parts of an allocate block
+      // over-allocate uncommitted (virtual) memory
+      p = mi_os_mem_alloc(over_size, 0 /*alignment*/, false /* commit? */, false /* allow_large */, is_large, stats);
+      if (p == NULL) return NULL;
+
+      // set p to the aligned part in the full region
+      // note: this is dangerous on Windows as VirtualFree needs the actual region pointer
+      // but in mi_os_mem_free we handle this (hopefully exceptional) situation.
+      p = mi_align_up_ptr(p, alignment);
+
+      // explicitly commit only the aligned part
+      if (commit) {
+        _mi_os_commit(p, size, NULL, stats);
+      }
+    }
+    else  { // mmap can free inside an allocation
+      // overallocate...
+      p = mi_os_mem_alloc(over_size, 1, commit, false, is_large, stats);
+      if (p == NULL) return NULL;
+      // and selectively unmap parts around the over-allocated area. (noop on sbrk)
+      void* aligned_p = mi_align_up_ptr(p, alignment);
+      size_t pre_size = (uint8_t*)aligned_p - (uint8_t*)p;
+      size_t mid_size = _mi_align_up(size, _mi_os_page_size());
+      size_t post_size = over_size - pre_size - mid_size;
+      mi_assert_internal(pre_size < over_size&& post_size < over_size&& mid_size >= size);
+      if (pre_size > 0)  mi_os_mem_free(p, pre_size, commit, stats);
+      if (post_size > 0) mi_os_mem_free((uint8_t*)aligned_p + mid_size, post_size, commit, stats);
+      // we can return the aligned pointer on `mmap` (and sbrk) systems
+      p = aligned_p;
+    }
+  }
+
+  mi_assert_internal(p == NULL || (p != NULL && ((uintptr_t)p % alignment) == 0));
+  return p;
+}
+
+
+/* -----------------------------------------------------------
+  OS API: alloc and alloc_aligned
+----------------------------------------------------------- */
+
+void* _mi_os_alloc(size_t size, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_stats_t* stats = &_mi_stats_main;
+  if (size == 0) return NULL;
+  size = _mi_os_good_alloc_size(size);
+  bool is_large = false;
+  return mi_os_mem_alloc(size, 0, true, false, &is_large, stats);
+}
+
+void* _mi_os_alloc_aligned(size_t size, size_t alignment, bool commit, bool* large, mi_stats_t* tld_stats)
+{
+  MI_UNUSED(&_mi_os_get_aligned_hint); // suppress unused warnings
+  MI_UNUSED(tld_stats);
+  if (size == 0) return NULL;
+  size = _mi_os_good_alloc_size(size);
+  alignment = _mi_align_up(alignment, _mi_os_page_size());
+  bool allow_large = false;
+  if (large != NULL) {
+    allow_large = *large;
+    *large = false;
+  }
+  return mi_os_mem_alloc_aligned(size, alignment, commit, allow_large, (large!=NULL?large:&allow_large), &_mi_stats_main /*tld->stats*/ );
+}
+
+/* -----------------------------------------------------------
+  OS aligned allocation with an offset. This is used
+  for large alignments > MI_ALIGNMENT_MAX. We use a large mimalloc
+  page where the object can be aligned at an offset from the start of the segment.
+  As we may need to overallocate, we need to free such pointers using `mi_free_aligned`
+  to use the actual start of the memory region.
+----------------------------------------------------------- */
+
+void* _mi_os_alloc_aligned_offset(size_t size, size_t alignment, size_t offset, bool commit, bool* large, mi_stats_t* tld_stats) {
+  mi_assert(offset <= MI_SEGMENT_SIZE);
+  mi_assert(offset <= size);
+  mi_assert((alignment % _mi_os_page_size()) == 0);
+  if (offset > MI_SEGMENT_SIZE) return NULL;
+  if (offset == 0) {
+    // regular aligned allocation
+    return _mi_os_alloc_aligned(size, alignment, commit, large, tld_stats);
+  }
+  else {
+    // overallocate to align at an offset
+    const size_t extra = _mi_align_up(offset, alignment) - offset;
+    const size_t oversize = size + extra;
+    void* start = _mi_os_alloc_aligned(oversize, alignment, commit, large, tld_stats);
+    if (start == NULL) return NULL;
+    void* p = (uint8_t*)start + extra;
+    mi_assert(_mi_is_aligned((uint8_t*)p + offset, alignment));
+    // decommit the overallocation at the start
+    if (commit && extra > _mi_os_page_size()) {
+      _mi_os_decommit(start, extra, tld_stats);
+    }
+    return p;
+  }
+}
+
+void _mi_os_free_aligned(void* p, size_t size, size_t alignment, size_t align_offset, bool was_committed, mi_stats_t* tld_stats) {
+  mi_assert(align_offset <= MI_SEGMENT_SIZE);
+  const size_t extra = _mi_align_up(align_offset, alignment) - align_offset;
+  void* start = (uint8_t*)p - extra;
+  _mi_os_free_ex(start, size + extra, was_committed, tld_stats);
+}
+
+
+/* -----------------------------------------------------------
+  OS memory API: reset, commit, decommit, protect, unprotect.
+----------------------------------------------------------- */
+
+// OS page align within a given area, either conservative (pages inside the area only),
+// or not (straddling pages outside the area is possible)
+static void* mi_os_page_align_areax(bool conservative, void* addr, size_t size, size_t* newsize) {
+  mi_assert(addr != NULL && size > 0);
+  if (newsize != NULL) *newsize = 0;
+  if (size == 0 || addr == NULL) return NULL;
+
+  // page align conservatively within the range
+  void* start = (conservative ? mi_align_up_ptr(addr, _mi_os_page_size())
+    : mi_align_down_ptr(addr, _mi_os_page_size()));
+  void* end = (conservative ? mi_align_down_ptr((uint8_t*)addr + size, _mi_os_page_size())
+    : mi_align_up_ptr((uint8_t*)addr + size, _mi_os_page_size()));
+  ptrdiff_t diff = (uint8_t*)end - (uint8_t*)start;
+  if (diff <= 0) return NULL;
+
+  mi_assert_internal((conservative && (size_t)diff <= size) || (!conservative && (size_t)diff >= size));
+  if (newsize != NULL) *newsize = (size_t)diff;
+  return start;
+}
+
+static void* mi_os_page_align_area_conservative(void* addr, size_t size, size_t* newsize) {
+  return mi_os_page_align_areax(true, addr, size, newsize);
+}
+
+// Commit/Decommit memory.
+// Usually commit is aligned liberal, while decommit is aligned conservative.
+// (but not for the reset version where we want commit to be conservative as well)
+static bool mi_os_commitx(void* addr, size_t size, bool commit, bool conservative, bool* is_zero, mi_stats_t* stats) {
+  // page align in the range, commit liberally, decommit conservative
+  if (is_zero != NULL) { *is_zero = false; }
+  size_t csize;
+  void* start = mi_os_page_align_areax(conservative, addr, size, &csize);
+  if (csize == 0) return true;  // || _mi_os_is_huge_reserved(addr))
+  if (commit) {
+    _mi_stat_increase(&stats->committed, size);  // use size for precise commit vs. decommit
+    _mi_stat_counter_increase(&stats->commit_calls, 1);
+  }
+  else {
+    _mi_stat_decrease(&stats->committed, size);
+  }
+
+  int err = _mi_prim_commit(start, csize, commit);  
+  if (err != 0) {
+    _mi_warning_message("cannot %s OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", commit ? "commit" : "decommit", err, err, start, csize);
+  }
+  mi_assert_internal(err == 0);
+  return (err == 0);
+}
+
+bool _mi_os_commit(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_stats_t* stats = &_mi_stats_main;
+  return mi_os_commitx(addr, size, true, false /* liberal */, is_zero, stats);
+}
+
+bool _mi_os_decommit(void* addr, size_t size, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_stats_t* stats = &_mi_stats_main;
+  bool is_zero;
+  return mi_os_commitx(addr, size, false, true /* conservative */, &is_zero, stats);
+}
+
+/*
+static bool mi_os_commit_unreset(void* addr, size_t size, bool* is_zero, mi_stats_t* stats) {  
+  return mi_os_commitx(addr, size, true, true // conservative
+                      , is_zero, stats);
+}
+*/
+
+// Signal to the OS that the address range is no longer in use
+// but may be used later again. This will release physical memory
+// pages and reduce swapping while keeping the memory committed.
+// We page align to a conservative area inside the range to reset.
+static bool mi_os_resetx(void* addr, size_t size, bool reset, mi_stats_t* stats) {
+  // page align conservatively within the range
+  size_t csize;
+  void* start = mi_os_page_align_area_conservative(addr, size, &csize);
+  if (csize == 0) return true;  // || _mi_os_is_huge_reserved(addr)
+  if (reset) _mi_stat_increase(&stats->reset, csize);
+        else _mi_stat_decrease(&stats->reset, csize);
+  if (!reset) return true; // nothing to do on unreset!
+
+  #if (MI_DEBUG>1) && !MI_SECURE && !MI_TRACK_ENABLED // && !MI_TSAN
+  memset(start, 0, csize); // pretend it is eagerly reset
+  #endif
+
+  int err = _mi_prim_reset(start, csize);
+  if (err != 0) {
+    _mi_warning_message("cannot reset OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", err, err, start, csize);
+  }
+  return (err == 0);
+}
+
+// Signal to the OS that the address range is no longer in use
+// but may be used later again. This will release physical memory
+// pages and reduce swapping while keeping the memory committed.
+// We page align to a conservative area inside the range to reset.
+bool _mi_os_reset(void* addr, size_t size, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_stats_t* stats = &_mi_stats_main;
+  return mi_os_resetx(addr, size, true, stats);
+}
+
+/*
+bool _mi_os_unreset(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats) {
+  MI_UNUSED(tld_stats);
+  mi_stats_t* stats = &_mi_stats_main;
+  *is_zero = false;
+  return mi_os_resetx(addr, size, false, stats);
+}
+*/
+
+// Protect a region in memory to be not accessible.
+static  bool mi_os_protectx(void* addr, size_t size, bool protect) {
+  // page align conservatively within the range
+  size_t csize = 0;
+  void* start = mi_os_page_align_area_conservative(addr, size, &csize);
+  if (csize == 0) return false;
+  /*
+  if (_mi_os_is_huge_reserved(addr)) {
+	  _mi_warning_message("cannot mprotect memory allocated in huge OS pages\n");
+  }
+  */
+  int err = _mi_prim_protect(start,csize,protect);
+  if (err != 0) {
+    _mi_warning_message("cannot %s OS memory (error: %d (0x%x), address: %p, size: 0x%zx bytes)\n", (protect ? "protect" : "unprotect"), err, err, start, csize);
+  }
+  return (err == 0);
+}
+
+bool _mi_os_protect(void* addr, size_t size) {
+  return mi_os_protectx(addr, size, true);
+}
+
+bool _mi_os_unprotect(void* addr, size_t size) {
+  return mi_os_protectx(addr, size, false);
+}
+
+
+
+/* ----------------------------------------------------------------------------
+Support for allocating huge OS pages (1Gib) that are reserved up-front
+and possibly associated with a specific NUMA node. (use `numa_node>=0`)
+-----------------------------------------------------------------------------*/
+#define MI_HUGE_OS_PAGE_SIZE  (MI_GiB)
+
+
+#if (MI_INTPTR_SIZE >= 8)
+// To ensure proper alignment, use our own area for huge OS pages
+static mi_decl_cache_align _Atomic(uintptr_t)  mi_huge_start; // = 0
+
+// Claim an aligned address range for huge pages
+static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
+  if (total_size != NULL) *total_size = 0;
+  const size_t size = pages * MI_HUGE_OS_PAGE_SIZE;
+
+  uintptr_t start = 0;
+  uintptr_t end = 0;
+  uintptr_t huge_start = mi_atomic_load_relaxed(&mi_huge_start);
+  do {
+    start = huge_start;
+    if (start == 0) {
+      // Initialize the start address after the 32TiB area
+      start = ((uintptr_t)32 << 40);  // 32TiB virtual start address
+    #if (MI_SECURE>0 || MI_DEBUG==0)      // security: randomize start of huge pages unless in debug mode
+      uintptr_t r = _mi_heap_random_next(mi_prim_get_default_heap());
+      start = start + ((uintptr_t)MI_HUGE_OS_PAGE_SIZE * ((r>>17) & 0x0FFF));  // (randomly 12bits)*1GiB == between 0 to 4TiB
+    #endif
+    }
+    end = start + size;
+    mi_assert_internal(end % MI_SEGMENT_SIZE == 0);
+  } while (!mi_atomic_cas_strong_acq_rel(&mi_huge_start, &huge_start, end));
+
+  if (total_size != NULL) *total_size = size;
+  return (uint8_t*)start;
+}
+#else
+static uint8_t* mi_os_claim_huge_pages(size_t pages, size_t* total_size) {
+  MI_UNUSED(pages);
+  if (total_size != NULL) *total_size = 0;
+  return NULL;
+}
+#endif
+
+// Allocate MI_SEGMENT_SIZE aligned huge pages
+void* _mi_os_alloc_huge_os_pages(size_t pages, int numa_node, mi_msecs_t max_msecs, size_t* pages_reserved, size_t* psize) {
+  if (psize != NULL) *psize = 0;
+  if (pages_reserved != NULL) *pages_reserved = 0;
+  size_t size = 0;
+  uint8_t* start = mi_os_claim_huge_pages(pages, &size);
+  if (start == NULL) return NULL; // or 32-bit systems
+
+  // Allocate one page at the time but try to place them contiguously
+  // We allocate one page at the time to be able to abort if it takes too long
+  // or to at least allocate as many as available on the system.
+  mi_msecs_t start_t = _mi_clock_start();
+  size_t page = 0;
+  while (page < pages) {
+    // allocate a page
+    void* addr = start + (page * MI_HUGE_OS_PAGE_SIZE);
+    void* p = NULL;
+    int err = _mi_prim_alloc_huge_os_pages(addr, MI_HUGE_OS_PAGE_SIZE, numa_node, &p);
+    if (err != 0) {
+      _mi_warning_message("unable to allocate huge OS page (error: %d (0x%x), address: %p, size: %zx bytes)\n", err, err, addr, MI_HUGE_OS_PAGE_SIZE);
+      break;
+    }
+
+    // Did we succeed at a contiguous address?
+    if (p != addr) {
+      // no success, issue a warning and break
+      if (p != NULL) {
+        _mi_warning_message("could not allocate contiguous huge OS page %zu at %p\n", page, addr);
+        _mi_os_free(p, MI_HUGE_OS_PAGE_SIZE, &_mi_stats_main);
+      }
+      break;
+    }
+
+    // success, record it
+    page++;  // increase before timeout check (see issue #711)
+    _mi_stat_increase(&_mi_stats_main.committed, MI_HUGE_OS_PAGE_SIZE);
+    _mi_stat_increase(&_mi_stats_main.reserved, MI_HUGE_OS_PAGE_SIZE);
+
+    // check for timeout
+    if (max_msecs > 0) {
+      mi_msecs_t elapsed = _mi_clock_end(start_t);
+      if (page >= 1) {
+        mi_msecs_t estimate = ((elapsed / (page+1)) * pages);
+        if (estimate > 2*max_msecs) { // seems like we are going to timeout, break
+          elapsed = max_msecs + 1;
+        }
+      }
+      if (elapsed > max_msecs) {
+        _mi_warning_message("huge OS page allocation timed out (after allocating %zu page(s))\n", page);
+        break;
+      }
+    }
+  }
+  mi_assert_internal(page*MI_HUGE_OS_PAGE_SIZE <= size);
+  if (pages_reserved != NULL) { *pages_reserved = page; }
+  if (psize != NULL) { *psize = page * MI_HUGE_OS_PAGE_SIZE; }
+  return (page == 0 ? NULL : start);
+}
+
+// free every huge page in a range individually (as we allocated per page)
+// note: needed with VirtualAlloc but could potentially be done in one go on mmap'd systems.
+void _mi_os_free_huge_pages(void* p, size_t size, mi_stats_t* stats) {
+  if (p==NULL || size==0) return;
+  uint8_t* base = (uint8_t*)p;
+  while (size >= MI_HUGE_OS_PAGE_SIZE) {
+    _mi_os_free(base, MI_HUGE_OS_PAGE_SIZE, stats);
+    size -= MI_HUGE_OS_PAGE_SIZE;
+    base += MI_HUGE_OS_PAGE_SIZE;
+  }
+}
+
+/* ----------------------------------------------------------------------------
+Support NUMA aware allocation
+-----------------------------------------------------------------------------*/
+
+_Atomic(size_t)  _mi_numa_node_count; // = 0   // cache the node count
+
+size_t _mi_os_numa_node_count_get(void) {
+  size_t count = mi_atomic_load_acquire(&_mi_numa_node_count);
+  if (count <= 0) {
+    long ncount = mi_option_get(mi_option_use_numa_nodes); // given explicitly?
+    if (ncount > 0) {
+      count = (size_t)ncount;
+    }
+    else {
+      count = _mi_prim_numa_node_count(); // or detect dynamically
+      if (count == 0) count = 1;
+    }
+    mi_atomic_store_release(&_mi_numa_node_count, count); // save it
+    _mi_verbose_message("using %zd numa regions\n", count);
+  }
+  return count;
+}
+
+int _mi_os_numa_node_get(mi_os_tld_t* tld) {
+  MI_UNUSED(tld);
+  size_t numa_count = _mi_os_numa_node_count();
+  if (numa_count<=1) return 0; // optimize on single numa node systems: always node 0
+  // never more than the node count and >= 0
+  size_t numa_node = _mi_prim_numa_node();
+  if (numa_node >= numa_count) { numa_node = numa_node % numa_count; }
+  return (int)numa_node;
+}

preload-mimalloc/mimalloc/src/page-queue.c

@@ -0,0 +1,332 @@
+/*----------------------------------------------------------------------------
+Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* -----------------------------------------------------------
+  Definition of page queues for each block size
+----------------------------------------------------------- */
+
+#ifndef MI_IN_PAGE_C
+#error "this file should be included from 'page.c'"
+#endif
+
+/* -----------------------------------------------------------
+  Minimal alignment in machine words (i.e. `sizeof(void*)`)
+----------------------------------------------------------- */
+
+#if (MI_MAX_ALIGN_SIZE > 4*MI_INTPTR_SIZE)
+  #error "define alignment for more than 4x word size for this platform"
+#elif (MI_MAX_ALIGN_SIZE > 2*MI_INTPTR_SIZE)
+  #define MI_ALIGN4W   // 4 machine words minimal alignment
+#elif (MI_MAX_ALIGN_SIZE > MI_INTPTR_SIZE)
+  #define MI_ALIGN2W   // 2 machine words minimal alignment
+#else
+  // ok, default alignment is 1 word
+#endif
+
+
+/* -----------------------------------------------------------
+  Queue query
+----------------------------------------------------------- */
+
+
+static inline bool mi_page_queue_is_huge(const mi_page_queue_t* pq) {
+  return (pq->block_size == (MI_MEDIUM_OBJ_SIZE_MAX+sizeof(uintptr_t)));
+}
+
+static inline bool mi_page_queue_is_full(const mi_page_queue_t* pq) {
+  return (pq->block_size == (MI_MEDIUM_OBJ_SIZE_MAX+(2*sizeof(uintptr_t))));
+}
+
+static inline bool mi_page_queue_is_special(const mi_page_queue_t* pq) {
+  return (pq->block_size > MI_MEDIUM_OBJ_SIZE_MAX);
+}
+
+/* -----------------------------------------------------------
+  Bins
+----------------------------------------------------------- */
+
+// Return the bin for a given field size.
+// Returns MI_BIN_HUGE if the size is too large.
+// We use `wsize` for the size in "machine word sizes",
+// i.e. byte size == `wsize*sizeof(void*)`.
+static inline uint8_t mi_bin(size_t size) {
+  size_t wsize = _mi_wsize_from_size(size);
+  uint8_t bin;
+  if (wsize <= 1) {
+    bin = 1;
+  }
+  #if defined(MI_ALIGN4W)
+  else if (wsize <= 4) {
+    bin = (uint8_t)((wsize+1)&~1); // round to double word sizes
+  }
+  #elif defined(MI_ALIGN2W)
+  else if (wsize <= 8) {
+    bin = (uint8_t)((wsize+1)&~1); // round to double word sizes
+  }
+  #else
+  else if (wsize <= 8) {
+    bin = (uint8_t)wsize;
+  }
+  #endif
+  else if (wsize > MI_MEDIUM_OBJ_WSIZE_MAX) {
+    bin = MI_BIN_HUGE;
+  }
+  else {
+    #if defined(MI_ALIGN4W)
+    if (wsize <= 16) { wsize = (wsize+3)&~3; } // round to 4x word sizes
+    #endif
+    wsize--;
+    // find the highest bit
+    uint8_t b = (uint8_t)mi_bsr(wsize);  // note: wsize != 0
+    // and use the top 3 bits to determine the bin (~12.5% worst internal fragmentation).
+    // - adjust with 3 because we use do not round the first 8 sizes
+    //   which each get an exact bin
+    bin = ((b << 2) + (uint8_t)((wsize >> (b - 2)) & 0x03)) - 3;
+    mi_assert_internal(bin < MI_BIN_HUGE);
+  }
+  mi_assert_internal(bin > 0 && bin <= MI_BIN_HUGE);
+  return bin;
+}
+
+
+
+/* -----------------------------------------------------------
+  Queue of pages with free blocks
+----------------------------------------------------------- */
+
+uint8_t _mi_bin(size_t size) {
+  return mi_bin(size);
+}
+
+size_t _mi_bin_size(uint8_t bin) {
+  return _mi_heap_empty.pages[bin].block_size;
+}
+
+// Good size for allocation
+size_t mi_good_size(size_t size) mi_attr_noexcept {
+  if (size <= MI_MEDIUM_OBJ_SIZE_MAX) {
+    return _mi_bin_size(mi_bin(size));
+  }
+  else {
+    return _mi_align_up(size,_mi_os_page_size());
+  }
+}
+
+#if (MI_DEBUG>1)
+static bool mi_page_queue_contains(mi_page_queue_t* queue, const mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_page_t* list = queue->first;
+  while (list != NULL) {
+    mi_assert_internal(list->next == NULL || list->next->prev == list);
+    mi_assert_internal(list->prev == NULL || list->prev->next == list);
+    if (list == page) break;
+    list = list->next;
+  }
+  return (list == page);
+}
+
+#endif
+
+#if (MI_DEBUG>1)
+static bool mi_heap_contains_queue(const mi_heap_t* heap, const mi_page_queue_t* pq) {
+  return (pq >= &heap->pages[0] && pq <= &heap->pages[MI_BIN_FULL]);
+}
+#endif
+
+static mi_page_queue_t* mi_page_queue_of(const mi_page_t* page) {
+  uint8_t bin = (mi_page_is_in_full(page) ? MI_BIN_FULL : mi_bin(page->xblock_size));
+  mi_heap_t* heap = mi_page_heap(page);
+  mi_assert_internal(heap != NULL && bin <= MI_BIN_FULL);
+  mi_page_queue_t* pq = &heap->pages[bin];
+  mi_assert_internal(bin >= MI_BIN_HUGE || page->xblock_size == pq->block_size);
+  mi_assert_expensive(mi_page_queue_contains(pq, page));
+  return pq;
+}
+
+static mi_page_queue_t* mi_heap_page_queue_of(mi_heap_t* heap, const mi_page_t* page) {
+  uint8_t bin = (mi_page_is_in_full(page) ? MI_BIN_FULL : mi_bin(page->xblock_size));
+  mi_assert_internal(bin <= MI_BIN_FULL);
+  mi_page_queue_t* pq = &heap->pages[bin];
+  mi_assert_internal(mi_page_is_in_full(page) || page->xblock_size == pq->block_size);
+  return pq;
+}
+
+// The current small page array is for efficiency and for each
+// small size (up to 256) it points directly to the page for that
+// size without having to compute the bin. This means when the
+// current free page queue is updated for a small bin, we need to update a
+// range of entries in `_mi_page_small_free`.
+static inline void mi_heap_queue_first_update(mi_heap_t* heap, const mi_page_queue_t* pq) {
+  mi_assert_internal(mi_heap_contains_queue(heap,pq));
+  size_t size = pq->block_size;
+  if (size > MI_SMALL_SIZE_MAX) return;
+
+  mi_page_t* page = pq->first;
+  if (pq->first == NULL) page = (mi_page_t*)&_mi_page_empty;
+
+  // find index in the right direct page array
+  size_t start;
+  size_t idx = _mi_wsize_from_size(size);
+  mi_page_t** pages_free = heap->pages_free_direct;
+
+  if (pages_free[idx] == page) return;  // already set
+
+  // find start slot
+  if (idx<=1) {
+    start = 0;
+  }
+  else {
+    // find previous size; due to minimal alignment upto 3 previous bins may need to be skipped
+    uint8_t bin = mi_bin(size);
+    const mi_page_queue_t* prev = pq - 1;
+    while( bin == mi_bin(prev->block_size) && prev > &heap->pages[0]) {
+      prev--;
+    }
+    start = 1 + _mi_wsize_from_size(prev->block_size);
+    if (start > idx) start = idx;
+  }
+
+  // set size range to the right page
+  mi_assert(start <= idx);
+  for (size_t sz = start; sz <= idx; sz++) {
+    pages_free[sz] = page;
+  }
+}
+
+/*
+static bool mi_page_queue_is_empty(mi_page_queue_t* queue) {
+  return (queue->first == NULL);
+}
+*/
+
+static void mi_page_queue_remove(mi_page_queue_t* queue, mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(mi_page_queue_contains(queue, page));
+  mi_assert_internal(page->xblock_size == queue->block_size || (page->xblock_size > MI_MEDIUM_OBJ_SIZE_MAX && mi_page_queue_is_huge(queue))  || (mi_page_is_in_full(page) && mi_page_queue_is_full(queue)));
+  mi_heap_t* heap = mi_page_heap(page);
+
+  if (page->prev != NULL) page->prev->next = page->next;
+  if (page->next != NULL) page->next->prev = page->prev;
+  if (page == queue->last)  queue->last = page->prev;
+  if (page == queue->first) {
+    queue->first = page->next;
+    // update first
+    mi_assert_internal(mi_heap_contains_queue(heap, queue));
+    mi_heap_queue_first_update(heap,queue);
+  }
+  heap->page_count--;
+  page->next = NULL;
+  page->prev = NULL;
+  // mi_atomic_store_ptr_release(mi_atomic_cast(void*, &page->heap), NULL);
+  mi_page_set_in_full(page,false);
+}
+
+
+static void mi_page_queue_push(mi_heap_t* heap, mi_page_queue_t* queue, mi_page_t* page) {
+  mi_assert_internal(mi_page_heap(page) == heap);
+  mi_assert_internal(!mi_page_queue_contains(queue, page));
+  #if MI_HUGE_PAGE_ABANDON
+  mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+  #endif
+  mi_assert_internal(page->xblock_size == queue->block_size ||
+                      (page->xblock_size > MI_MEDIUM_OBJ_SIZE_MAX) ||
+                        (mi_page_is_in_full(page) && mi_page_queue_is_full(queue)));
+
+  mi_page_set_in_full(page, mi_page_queue_is_full(queue));
+  // mi_atomic_store_ptr_release(mi_atomic_cast(void*, &page->heap), heap);
+  page->next = queue->first;
+  page->prev = NULL;
+  if (queue->first != NULL) {
+    mi_assert_internal(queue->first->prev == NULL);
+    queue->first->prev = page;
+    queue->first = page;
+  }
+  else {
+    queue->first = queue->last = page;
+  }
+
+  // update direct
+  mi_heap_queue_first_update(heap, queue);
+  heap->page_count++;
+}
+
+
+static void mi_page_queue_enqueue_from(mi_page_queue_t* to, mi_page_queue_t* from, mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(mi_page_queue_contains(from, page));
+  mi_assert_expensive(!mi_page_queue_contains(to, page));
+
+  mi_assert_internal((page->xblock_size == to->block_size && page->xblock_size == from->block_size) ||
+                     (page->xblock_size == to->block_size && mi_page_queue_is_full(from)) ||
+                     (page->xblock_size == from->block_size && mi_page_queue_is_full(to)) ||
+                     (page->xblock_size > MI_LARGE_OBJ_SIZE_MAX && mi_page_queue_is_huge(to)) ||
+                     (page->xblock_size > MI_LARGE_OBJ_SIZE_MAX && mi_page_queue_is_full(to)));
+
+  mi_heap_t* heap = mi_page_heap(page);
+  if (page->prev != NULL) page->prev->next = page->next;
+  if (page->next != NULL) page->next->prev = page->prev;
+  if (page == from->last)  from->last = page->prev;
+  if (page == from->first) {
+    from->first = page->next;
+    // update first
+    mi_assert_internal(mi_heap_contains_queue(heap, from));
+    mi_heap_queue_first_update(heap, from);
+  }
+
+  page->prev = to->last;
+  page->next = NULL;
+  if (to->last != NULL) {
+    mi_assert_internal(heap == mi_page_heap(to->last));
+    to->last->next = page;
+    to->last = page;
+  }
+  else {
+    to->first = page;
+    to->last = page;
+    mi_heap_queue_first_update(heap, to);
+  }
+
+  mi_page_set_in_full(page, mi_page_queue_is_full(to));
+}
+
+// Only called from `mi_heap_absorb`.
+size_t _mi_page_queue_append(mi_heap_t* heap, mi_page_queue_t* pq, mi_page_queue_t* append) {
+  mi_assert_internal(mi_heap_contains_queue(heap,pq));
+  mi_assert_internal(pq->block_size == append->block_size);
+
+  if (append->first==NULL) return 0;
+
+  // set append pages to new heap and count
+  size_t count = 0;
+  for (mi_page_t* page = append->first; page != NULL; page = page->next) {
+    // inline `mi_page_set_heap` to avoid wrong assertion during absorption;
+    // in this case it is ok to be delayed freeing since both "to" and "from" heap are still alive.
+    mi_atomic_store_release(&page->xheap, (uintptr_t)heap);
+    // set the flag to delayed free (not overriding NEVER_DELAYED_FREE) which has as a
+    // side effect that it spins until any DELAYED_FREEING is finished. This ensures
+    // that after appending only the new heap will be used for delayed free operations.
+    _mi_page_use_delayed_free(page, MI_USE_DELAYED_FREE, false);
+    count++;
+  }
+
+  if (pq->last==NULL) {
+    // take over afresh
+    mi_assert_internal(pq->first==NULL);
+    pq->first = append->first;
+    pq->last = append->last;
+    mi_heap_queue_first_update(heap, pq);
+  }
+  else {
+    // append to end
+    mi_assert_internal(pq->last!=NULL);
+    mi_assert_internal(append->first!=NULL);
+    pq->last->next = append->first;
+    append->first->prev = pq->last;
+    pq->last = append->last;
+  }
+  return count;
+}

preload-mimalloc/mimalloc/src/page.c

@@ -0,0 +1,935 @@
+/*----------------------------------------------------------------------------
+Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* -----------------------------------------------------------
+  The core of the allocator. Every segment contains
+  pages of a certain block size. The main function
+  exported is `mi_malloc_generic`.
+----------------------------------------------------------- */
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+/* -----------------------------------------------------------
+  Definition of page queues for each block size
+----------------------------------------------------------- */
+
+#define MI_IN_PAGE_C
+#include "page-queue.c"
+#undef MI_IN_PAGE_C
+
+
+/* -----------------------------------------------------------
+  Page helpers
+----------------------------------------------------------- */
+
+// Index a block in a page
+static inline mi_block_t* mi_page_block_at(const mi_page_t* page, void* page_start, size_t block_size, size_t i) {
+  MI_UNUSED(page);
+  mi_assert_internal(page != NULL);
+  mi_assert_internal(i <= page->reserved);
+  return (mi_block_t*)((uint8_t*)page_start + (i * block_size));
+}
+
+static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t size, mi_tld_t* tld);
+static void mi_page_extend_free(mi_heap_t* heap, mi_page_t* page, mi_tld_t* tld);
+
+#if (MI_DEBUG>=3)
+static size_t mi_page_list_count(mi_page_t* page, mi_block_t* head) {
+  size_t count = 0;
+  while (head != NULL) {
+    mi_assert_internal(page == _mi_ptr_page(head));
+    count++;
+    head = mi_block_next(page, head);
+  }
+  return count;
+}
+
+/*
+// Start of the page available memory
+static inline uint8_t* mi_page_area(const mi_page_t* page) {
+  return _mi_page_start(_mi_page_segment(page), page, NULL);
+}
+*/
+
+static bool mi_page_list_is_valid(mi_page_t* page, mi_block_t* p) {
+  size_t psize;
+  uint8_t* page_area = _mi_page_start(_mi_page_segment(page), page, &psize);
+  mi_block_t* start = (mi_block_t*)page_area;
+  mi_block_t* end   = (mi_block_t*)(page_area + psize);
+  while(p != NULL) {
+    if (p < start || p >= end) return false;
+    p = mi_block_next(page, p);
+  }
+  return true;
+}
+
+static bool mi_page_is_valid_init(mi_page_t* page) {
+  mi_assert_internal(page->xblock_size > 0);
+  mi_assert_internal(page->used <= page->capacity);
+  mi_assert_internal(page->capacity <= page->reserved);
+
+  mi_segment_t* segment = _mi_page_segment(page);
+  uint8_t* start = _mi_page_start(segment,page,NULL);
+  mi_assert_internal(start == _mi_segment_page_start(segment,page,NULL));
+  //const size_t bsize = mi_page_block_size(page);
+  //mi_assert_internal(start + page->capacity*page->block_size == page->top);
+
+  mi_assert_internal(mi_page_list_is_valid(page,page->free));
+  mi_assert_internal(mi_page_list_is_valid(page,page->local_free));
+
+  #if MI_DEBUG>3 // generally too expensive to check this
+  if (page->is_zero) {
+    const size_t ubsize = mi_page_usable_block_size(page);
+    for(mi_block_t* block = page->free; block != NULL; block = mi_block_next(page,block)) {
+      mi_assert_expensive(mi_mem_is_zero(block + 1, ubsize - sizeof(mi_block_t)));
+    }
+  }
+  #endif
+
+  #if !MI_TRACK_ENABLED && !MI_TSAN
+  mi_block_t* tfree = mi_page_thread_free(page);
+  mi_assert_internal(mi_page_list_is_valid(page, tfree));
+  //size_t tfree_count = mi_page_list_count(page, tfree);
+  //mi_assert_internal(tfree_count <= page->thread_freed + 1);
+  #endif
+
+  size_t free_count = mi_page_list_count(page, page->free) + mi_page_list_count(page, page->local_free);
+  mi_assert_internal(page->used + free_count == page->capacity);
+
+  return true;
+}
+
+extern bool _mi_process_is_initialized;             // has mi_process_init been called?
+
+bool _mi_page_is_valid(mi_page_t* page) {
+  mi_assert_internal(mi_page_is_valid_init(page));
+  #if MI_SECURE
+  mi_assert_internal(page->keys[0] != 0);
+  #endif
+  if (mi_page_heap(page)!=NULL) {
+    mi_segment_t* segment = _mi_page_segment(page);
+
+    mi_assert_internal(!_mi_process_is_initialized || segment->thread_id==0 || segment->thread_id == mi_page_heap(page)->thread_id);
+    #if MI_HUGE_PAGE_ABANDON
+    if (segment->kind != MI_SEGMENT_HUGE) 
+    #endif
+    {    
+      mi_page_queue_t* pq = mi_page_queue_of(page);
+      mi_assert_internal(mi_page_queue_contains(pq, page));
+      mi_assert_internal(pq->block_size==mi_page_block_size(page) || mi_page_block_size(page) > MI_MEDIUM_OBJ_SIZE_MAX || mi_page_is_in_full(page));
+      mi_assert_internal(mi_heap_contains_queue(mi_page_heap(page),pq));
+    }
+  }
+  return true;
+}
+#endif
+
+void _mi_page_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never) {
+  while (!_mi_page_try_use_delayed_free(page, delay, override_never)) {
+    mi_atomic_yield();
+  }
+}
+
+bool _mi_page_try_use_delayed_free(mi_page_t* page, mi_delayed_t delay, bool override_never) {
+  mi_thread_free_t tfreex;
+  mi_delayed_t     old_delay;
+  mi_thread_free_t tfree;
+  size_t yield_count = 0;
+  do {
+    tfree = mi_atomic_load_acquire(&page->xthread_free); // note: must acquire as we can break/repeat this loop and not do a CAS;
+    tfreex = mi_tf_set_delayed(tfree, delay);
+    old_delay = mi_tf_delayed(tfree);
+    if mi_unlikely(old_delay == MI_DELAYED_FREEING) {
+      if (yield_count >= 4) return false;  // give up after 4 tries
+      yield_count++;
+      mi_atomic_yield(); // delay until outstanding MI_DELAYED_FREEING are done.
+      // tfree = mi_tf_set_delayed(tfree, MI_NO_DELAYED_FREE); // will cause CAS to busy fail
+    }
+    else if (delay == old_delay) {
+      break; // avoid atomic operation if already equal
+    }
+    else if (!override_never && old_delay == MI_NEVER_DELAYED_FREE) {
+      break; // leave never-delayed flag set
+    }
+  } while ((old_delay == MI_DELAYED_FREEING) ||
+           !mi_atomic_cas_weak_release(&page->xthread_free, &tfree, tfreex));
+
+  return true; // success
+}
+
+/* -----------------------------------------------------------
+  Page collect the `local_free` and `thread_free` lists
+----------------------------------------------------------- */
+
+// Collect the local `thread_free` list using an atomic exchange.
+// Note: The exchange must be done atomically as this is used right after
+// moving to the full list in `mi_page_collect_ex` and we need to
+// ensure that there was no race where the page became unfull just before the move.
+static void _mi_page_thread_free_collect(mi_page_t* page)
+{
+  mi_block_t* head;
+  mi_thread_free_t tfreex;
+  mi_thread_free_t tfree = mi_atomic_load_relaxed(&page->xthread_free);
+  do {
+    head = mi_tf_block(tfree);
+    tfreex = mi_tf_set_block(tfree,NULL);
+  } while (!mi_atomic_cas_weak_acq_rel(&page->xthread_free, &tfree, tfreex));
+
+  // return if the list is empty
+  if (head == NULL) return;
+
+  // find the tail -- also to get a proper count (without data races)
+  uint32_t max_count = page->capacity; // cannot collect more than capacity
+  uint32_t count = 1;
+  mi_block_t* tail = head;
+  mi_block_t* next;
+  while ((next = mi_block_next(page,tail)) != NULL && count <= max_count) {
+    count++;
+    tail = next;
+  }
+  // if `count > max_count` there was a memory corruption (possibly infinite list due to double multi-threaded free)
+  if (count > max_count) {
+    _mi_error_message(EFAULT, "corrupted thread-free list\n");
+    return; // the thread-free items cannot be freed
+  }
+
+  // and append the current local free list
+  mi_block_set_next(page,tail, page->local_free);
+  page->local_free = head;
+
+  // update counts now
+  page->used -= count;
+}
+
+void _mi_page_free_collect(mi_page_t* page, bool force) {
+  mi_assert_internal(page!=NULL);
+
+  // collect the thread free list
+  if (force || mi_page_thread_free(page) != NULL) {  // quick test to avoid an atomic operation
+    _mi_page_thread_free_collect(page);
+  }
+
+  // and the local free list
+  if (page->local_free != NULL) {
+    if mi_likely(page->free == NULL) {
+      // usual case
+      page->free = page->local_free;
+      page->local_free = NULL;
+      page->is_zero = false;
+    }
+    else if (force) {
+      // append -- only on shutdown (force) as this is a linear operation
+      mi_block_t* tail = page->local_free;
+      mi_block_t* next;
+      while ((next = mi_block_next(page, tail)) != NULL) {
+        tail = next;
+      }
+      mi_block_set_next(page, tail, page->free);
+      page->free = page->local_free;
+      page->local_free = NULL;
+      page->is_zero = false;
+    }
+  }
+
+  mi_assert_internal(!force || page->local_free == NULL);
+}
+
+
+
+/* -----------------------------------------------------------
+  Page fresh and retire
+----------------------------------------------------------- */
+
+// called from segments when reclaiming abandoned pages
+void _mi_page_reclaim(mi_heap_t* heap, mi_page_t* page) {
+  mi_assert_expensive(mi_page_is_valid_init(page));
+
+  mi_assert_internal(mi_page_heap(page) == heap);
+  mi_assert_internal(mi_page_thread_free_flag(page) != MI_NEVER_DELAYED_FREE);
+  #if MI_HUGE_PAGE_ABANDON
+  mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+  #endif
+  mi_assert_internal(!page->is_reset);
+  // TODO: push on full queue immediately if it is full?
+  mi_page_queue_t* pq = mi_page_queue(heap, mi_page_block_size(page));
+  mi_page_queue_push(heap, pq, page);
+  mi_assert_expensive(_mi_page_is_valid(page));
+}
+
+// allocate a fresh page from a segment
+static mi_page_t* mi_page_fresh_alloc(mi_heap_t* heap, mi_page_queue_t* pq, size_t block_size, size_t page_alignment) {
+  #if !MI_HUGE_PAGE_ABANDON
+  mi_assert_internal(pq != NULL);
+  mi_assert_internal(mi_heap_contains_queue(heap, pq));
+  mi_assert_internal(page_alignment > 0 || block_size > MI_MEDIUM_OBJ_SIZE_MAX || block_size == pq->block_size);
+  #endif
+  mi_page_t* page = _mi_segment_page_alloc(heap, block_size, page_alignment, &heap->tld->segments, &heap->tld->os);
+  if (page == NULL) {
+    // this may be out-of-memory, or an abandoned page was reclaimed (and in our queue)
+    return NULL;
+  }
+  mi_assert_internal(page_alignment >0 || block_size > MI_MEDIUM_OBJ_SIZE_MAX || _mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+  mi_assert_internal(pq!=NULL || page->xblock_size != 0);
+  mi_assert_internal(pq!=NULL || mi_page_block_size(page) >= block_size);
+  // a fresh page was found, initialize it
+  const size_t full_block_size = ((pq == NULL || mi_page_queue_is_huge(pq)) ? mi_page_block_size(page) : block_size); // see also: mi_segment_huge_page_alloc
+  mi_assert_internal(full_block_size >= block_size);
+  mi_page_init(heap, page, full_block_size, heap->tld);
+  mi_heap_stat_increase(heap, pages, 1);
+  if (pq != NULL) { mi_page_queue_push(heap, pq, page); }
+  mi_assert_expensive(_mi_page_is_valid(page));
+  return page;
+}
+
+// Get a fresh page to use
+static mi_page_t* mi_page_fresh(mi_heap_t* heap, mi_page_queue_t* pq) {
+  mi_assert_internal(mi_heap_contains_queue(heap, pq));
+  mi_page_t* page = mi_page_fresh_alloc(heap, pq, pq->block_size, 0);
+  if (page==NULL) return NULL;
+  mi_assert_internal(pq->block_size==mi_page_block_size(page));
+  mi_assert_internal(pq==mi_page_queue(heap, mi_page_block_size(page)));
+  return page;
+}
+
+/* -----------------------------------------------------------
+   Do any delayed frees
+   (put there by other threads if they deallocated in a full page)
+----------------------------------------------------------- */
+void _mi_heap_delayed_free_all(mi_heap_t* heap) {
+  while (!_mi_heap_delayed_free_partial(heap)) {
+    mi_atomic_yield();
+  }
+}
+
+// returns true if all delayed frees were processed
+bool _mi_heap_delayed_free_partial(mi_heap_t* heap) {
+  // take over the list (note: no atomic exchange since it is often NULL)
+  mi_block_t* block = mi_atomic_load_ptr_relaxed(mi_block_t, &heap->thread_delayed_free);
+  while (block != NULL && !mi_atomic_cas_ptr_weak_acq_rel(mi_block_t, &heap->thread_delayed_free, &block, NULL)) { /* nothing */ };
+  bool all_freed = true;
+
+  // and free them all
+  while(block != NULL) {
+    mi_block_t* next = mi_block_nextx(heap,block, heap->keys);
+    // use internal free instead of regular one to keep stats etc correct
+    if (!_mi_free_delayed_block(block)) {
+      // we might already start delayed freeing while another thread has not yet
+      // reset the delayed_freeing flag; in that case delay it further by reinserting the current block
+      // into the delayed free list
+      all_freed = false;
+      mi_block_t* dfree = mi_atomic_load_ptr_relaxed(mi_block_t, &heap->thread_delayed_free);
+      do {
+        mi_block_set_nextx(heap, block, dfree, heap->keys);
+      } while (!mi_atomic_cas_ptr_weak_release(mi_block_t,&heap->thread_delayed_free, &dfree, block));
+    }
+    block = next;
+  }
+  return all_freed;
+}
+
+/* -----------------------------------------------------------
+  Unfull, abandon, free and retire
+----------------------------------------------------------- */
+
+// Move a page from the full list back to a regular list
+void _mi_page_unfull(mi_page_t* page) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(mi_page_is_in_full(page));
+  if (!mi_page_is_in_full(page)) return;
+
+  mi_heap_t* heap = mi_page_heap(page);
+  mi_page_queue_t* pqfull = &heap->pages[MI_BIN_FULL];
+  mi_page_set_in_full(page, false); // to get the right queue
+  mi_page_queue_t* pq = mi_heap_page_queue_of(heap, page);
+  mi_page_set_in_full(page, true);
+  mi_page_queue_enqueue_from(pq, pqfull, page);
+}
+
+static void mi_page_to_full(mi_page_t* page, mi_page_queue_t* pq) {
+  mi_assert_internal(pq == mi_page_queue_of(page));
+  mi_assert_internal(!mi_page_immediate_available(page));
+  mi_assert_internal(!mi_page_is_in_full(page));
+
+  if (mi_page_is_in_full(page)) return;
+  mi_page_queue_enqueue_from(&mi_page_heap(page)->pages[MI_BIN_FULL], pq, page);
+  _mi_page_free_collect(page,false);  // try to collect right away in case another thread freed just before MI_USE_DELAYED_FREE was set
+}
+
+
+// Abandon a page with used blocks at the end of a thread.
+// Note: only call if it is ensured that no references exist from
+// the `page->heap->thread_delayed_free` into this page.
+// Currently only called through `mi_heap_collect_ex` which ensures this.
+void _mi_page_abandon(mi_page_t* page, mi_page_queue_t* pq) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(pq == mi_page_queue_of(page));
+  mi_assert_internal(mi_page_heap(page) != NULL);
+
+  mi_heap_t* pheap = mi_page_heap(page);
+
+  // remove from our page list
+  mi_segments_tld_t* segments_tld = &pheap->tld->segments;
+  mi_page_queue_remove(pq, page);
+
+  // page is no longer associated with our heap
+  mi_assert_internal(mi_page_thread_free_flag(page)==MI_NEVER_DELAYED_FREE);
+  mi_page_set_heap(page, NULL);
+
+#if (MI_DEBUG>1) && !MI_TRACK_ENABLED
+  // check there are no references left..
+  for (mi_block_t* block = (mi_block_t*)pheap->thread_delayed_free; block != NULL; block = mi_block_nextx(pheap, block, pheap->keys)) {
+    mi_assert_internal(_mi_ptr_page(block) != page);
+  }
+#endif
+
+  // and abandon it
+  mi_assert_internal(mi_page_heap(page) == NULL);
+  _mi_segment_page_abandon(page,segments_tld);
+}
+
+
+// Free a page with no more free blocks
+void _mi_page_free(mi_page_t* page, mi_page_queue_t* pq, bool force) {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(pq == mi_page_queue_of(page));
+  mi_assert_internal(mi_page_all_free(page));
+  mi_assert_internal(mi_page_thread_free_flag(page)!=MI_DELAYED_FREEING);
+
+  // no more aligned blocks in here
+  mi_page_set_has_aligned(page, false);
+
+  mi_heap_t* heap = mi_page_heap(page);
+
+  // remove from the page list
+  // (no need to do _mi_heap_delayed_free first as all blocks are already free)
+  mi_segments_tld_t* segments_tld = &heap->tld->segments;
+  mi_page_queue_remove(pq, page);
+
+  // and free it
+  mi_page_set_heap(page,NULL);
+  _mi_segment_page_free(page, force, segments_tld);
+}
+
+// Retire parameters
+#define MI_MAX_RETIRE_SIZE    (MI_MEDIUM_OBJ_SIZE_MAX)
+#define MI_RETIRE_CYCLES      (8)
+
+// Retire a page with no more used blocks
+// Important to not retire too quickly though as new
+// allocations might coming.
+// Note: called from `mi_free` and benchmarks often
+// trigger this due to freeing everything and then
+// allocating again so careful when changing this.
+void _mi_page_retire(mi_page_t* page) mi_attr_noexcept {
+  mi_assert_internal(page != NULL);
+  mi_assert_expensive(_mi_page_is_valid(page));
+  mi_assert_internal(mi_page_all_free(page));
+  
+  mi_page_set_has_aligned(page, false);
+
+  // don't retire too often..
+  // (or we end up retiring and re-allocating most of the time)
+  // NOTE: refine this more: we should not retire if this
+  // is the only page left with free blocks. It is not clear
+  // how to check this efficiently though...
+  // for now, we don't retire if it is the only page left of this size class.
+  mi_page_queue_t* pq = mi_page_queue_of(page);
+  if mi_likely(page->xblock_size <= MI_MAX_RETIRE_SIZE && !mi_page_queue_is_special(pq)) {  // not too large && not full or huge queue?
+    if (pq->last==page && pq->first==page) { // the only page in the queue?
+      mi_stat_counter_increase(_mi_stats_main.page_no_retire,1);
+      page->retire_expire = 1 + (page->xblock_size <= MI_SMALL_OBJ_SIZE_MAX ? MI_RETIRE_CYCLES : MI_RETIRE_CYCLES/4);      
+      mi_heap_t* heap = mi_page_heap(page);
+      mi_assert_internal(pq >= heap->pages);
+      const size_t index = pq - heap->pages;
+      mi_assert_internal(index < MI_BIN_FULL && index < MI_BIN_HUGE);
+      if (index < heap->page_retired_min) heap->page_retired_min = index;
+      if (index > heap->page_retired_max) heap->page_retired_max = index;
+      mi_assert_internal(mi_page_all_free(page));
+      return; // dont't free after all
+    }
+  }
+  _mi_page_free(page, pq, false);
+}
+
+// free retired pages: we don't need to look at the entire queues
+// since we only retire pages that are at the head position in a queue.
+void _mi_heap_collect_retired(mi_heap_t* heap, bool force) {
+  size_t min = MI_BIN_FULL;
+  size_t max = 0;
+  for(size_t bin = heap->page_retired_min; bin <= heap->page_retired_max; bin++) {
+    mi_page_queue_t* pq   = &heap->pages[bin];
+    mi_page_t*       page = pq->first;
+    if (page != NULL && page->retire_expire != 0) {
+      if (mi_page_all_free(page)) {
+        page->retire_expire--;
+        if (force || page->retire_expire == 0) {
+          _mi_page_free(pq->first, pq, force);
+        }
+        else {
+          // keep retired, update min/max
+          if (bin < min) min = bin;
+          if (bin > max) max = bin;
+        }
+      }
+      else {
+        page->retire_expire = 0;
+      }
+    }
+  }
+  heap->page_retired_min = min;
+  heap->page_retired_max = max;
+}
+
+
+/* -----------------------------------------------------------
+  Initialize the initial free list in a page.
+  In secure mode we initialize a randomized list by
+  alternating between slices.
+----------------------------------------------------------- */
+
+#define MI_MAX_SLICE_SHIFT  (6)   // at most 64 slices
+#define MI_MAX_SLICES       (1UL << MI_MAX_SLICE_SHIFT)
+#define MI_MIN_SLICES       (2)
+
+static void mi_page_free_list_extend_secure(mi_heap_t* const heap, mi_page_t* const page, const size_t bsize, const size_t extend, mi_stats_t* const stats) {
+  MI_UNUSED(stats);
+  #if (MI_SECURE<=2)
+  mi_assert_internal(page->free == NULL);
+  mi_assert_internal(page->local_free == NULL);
+  #endif
+  mi_assert_internal(page->capacity + extend <= page->reserved);
+  mi_assert_internal(bsize == mi_page_block_size(page));
+  void* const page_area = _mi_page_start(_mi_page_segment(page), page, NULL);
+
+  // initialize a randomized free list
+  // set up `slice_count` slices to alternate between
+  size_t shift = MI_MAX_SLICE_SHIFT;
+  while ((extend >> shift) == 0) {
+    shift--;
+  }
+  const size_t slice_count = (size_t)1U << shift;
+  const size_t slice_extend = extend / slice_count;
+  mi_assert_internal(slice_extend >= 1);
+  mi_block_t* blocks[MI_MAX_SLICES];   // current start of the slice
+  size_t      counts[MI_MAX_SLICES];   // available objects in the slice
+  for (size_t i = 0; i < slice_count; i++) {
+    blocks[i] = mi_page_block_at(page, page_area, bsize, page->capacity + i*slice_extend);
+    counts[i] = slice_extend;
+  }
+  counts[slice_count-1] += (extend % slice_count);  // final slice holds the modulus too (todo: distribute evenly?)
+
+  // and initialize the free list by randomly threading through them
+  // set up first element
+  const uintptr_t r = _mi_heap_random_next(heap);
+  size_t current = r % slice_count;
+  counts[current]--;
+  mi_block_t* const free_start = blocks[current];
+  // and iterate through the rest; use `random_shuffle` for performance
+  uintptr_t rnd = _mi_random_shuffle(r|1); // ensure not 0
+  for (size_t i = 1; i < extend; i++) {
+    // call random_shuffle only every INTPTR_SIZE rounds
+    const size_t round = i%MI_INTPTR_SIZE;
+    if (round == 0) rnd = _mi_random_shuffle(rnd);
+    // select a random next slice index
+    size_t next = ((rnd >> 8*round) & (slice_count-1));
+    while (counts[next]==0) {                            // ensure it still has space
+      next++;
+      if (next==slice_count) next = 0;
+    }
+    // and link the current block to it
+    counts[next]--;
+    mi_block_t* const block = blocks[current];
+    blocks[current] = (mi_block_t*)((uint8_t*)block + bsize);  // bump to the following block
+    mi_block_set_next(page, block, blocks[next]);   // and set next; note: we may have `current == next`
+    current = next;
+  }
+  // prepend to the free list (usually NULL)
+  mi_block_set_next(page, blocks[current], page->free);  // end of the list
+  page->free = free_start;
+}
+
+static mi_decl_noinline void mi_page_free_list_extend( mi_page_t* const page, const size_t bsize, const size_t extend, mi_stats_t* const stats)
+{
+  MI_UNUSED(stats);
+  #if (MI_SECURE <= 2)
+  mi_assert_internal(page->free == NULL);
+  mi_assert_internal(page->local_free == NULL);
+  #endif
+  mi_assert_internal(page->capacity + extend <= page->reserved);
+  mi_assert_internal(bsize == mi_page_block_size(page));
+  void* const page_area = _mi_page_start(_mi_page_segment(page), page, NULL );
+
+  mi_block_t* const start = mi_page_block_at(page, page_area, bsize, page->capacity);
+
+  // initialize a sequential free list
+  mi_block_t* const last = mi_page_block_at(page, page_area, bsize, page->capacity + extend - 1);
+  mi_block_t* block = start;
+  while(block <= last) {
+    mi_block_t* next = (mi_block_t*)((uint8_t*)block + bsize);
+    mi_block_set_next(page,block,next);
+    block = next;
+  }
+  // prepend to free list (usually `NULL`)
+  mi_block_set_next(page, last, page->free);
+  page->free = start;
+}
+
+/* -----------------------------------------------------------
+  Page initialize and extend the capacity
+----------------------------------------------------------- */
+
+#define MI_MAX_EXTEND_SIZE    (4*1024)      // heuristic, one OS page seems to work well.
+#if (MI_SECURE>0)
+#define MI_MIN_EXTEND         (8*MI_SECURE) // extend at least by this many
+#else
+#define MI_MIN_EXTEND         (4)
+#endif
+
+// Extend the capacity (up to reserved) by initializing a free list
+// We do at most `MI_MAX_EXTEND` to avoid touching too much memory
+// Note: we also experimented with "bump" allocation on the first
+// allocations but this did not speed up any benchmark (due to an
+// extra test in malloc? or cache effects?)
+static void mi_page_extend_free(mi_heap_t* heap, mi_page_t* page, mi_tld_t* tld) {
+  MI_UNUSED(tld); 
+  mi_assert_expensive(mi_page_is_valid_init(page));
+  #if (MI_SECURE<=2)
+  mi_assert(page->free == NULL);
+  mi_assert(page->local_free == NULL);
+  if (page->free != NULL) return;
+  #endif
+  if (page->capacity >= page->reserved) return;
+
+  size_t page_size;
+  _mi_page_start(_mi_page_segment(page), page, &page_size);
+  mi_stat_counter_increase(tld->stats.pages_extended, 1);
+
+  // calculate the extend count
+  const size_t bsize = (page->xblock_size < MI_HUGE_BLOCK_SIZE ? page->xblock_size : page_size);
+  size_t extend = page->reserved - page->capacity;
+  mi_assert_internal(extend > 0);
+
+  size_t max_extend = (bsize >= MI_MAX_EXTEND_SIZE ? MI_MIN_EXTEND : MI_MAX_EXTEND_SIZE/(uint32_t)bsize);
+  if (max_extend < MI_MIN_EXTEND) { max_extend = MI_MIN_EXTEND; }
+  mi_assert_internal(max_extend > 0);
+
+  if (extend > max_extend) {
+    // ensure we don't touch memory beyond the page to reduce page commit.
+    // the `lean` benchmark tests this. Going from 1 to 8 increases rss by 50%.
+    extend = max_extend;
+  }
+
+  mi_assert_internal(extend > 0 && extend + page->capacity <= page->reserved);
+  mi_assert_internal(extend < (1UL<<16));
+
+  // and append the extend the free list
+  if (extend < MI_MIN_SLICES || MI_SECURE==0) { //!mi_option_is_enabled(mi_option_secure)) {
+    mi_page_free_list_extend(page, bsize, extend, &tld->stats );
+  }
+  else {
+    mi_page_free_list_extend_secure(heap, page, bsize, extend, &tld->stats);
+  }
+  // enable the new free list
+  page->capacity += (uint16_t)extend;
+  mi_stat_increase(tld->stats.page_committed, extend * bsize);
+
+  // extension into zero initialized memory preserves the zero'd free list
+  if (!page->is_zero_init) {
+    page->is_zero = false;
+  }
+  mi_assert_expensive(mi_page_is_valid_init(page));
+}
+
+// Initialize a fresh page
+static void mi_page_init(mi_heap_t* heap, mi_page_t* page, size_t block_size, mi_tld_t* tld) {
+  mi_assert(page != NULL);
+  mi_segment_t* segment = _mi_page_segment(page);
+  mi_assert(segment != NULL);
+  mi_assert_internal(block_size > 0);
+  // set fields
+  mi_page_set_heap(page, heap);
+  page->xblock_size = (block_size < MI_HUGE_BLOCK_SIZE ? (uint32_t)block_size : MI_HUGE_BLOCK_SIZE); // initialize before _mi_segment_page_start
+  size_t page_size;
+  const void* page_start = _mi_segment_page_start(segment, page, &page_size);
+  MI_UNUSED(page_start);
+  mi_track_mem_noaccess(page_start,page_size);
+  mi_assert_internal(mi_page_block_size(page) <= page_size);
+  mi_assert_internal(page_size <= page->slice_count*MI_SEGMENT_SLICE_SIZE);
+  mi_assert_internal(page_size / block_size < (1L<<16));
+  page->reserved = (uint16_t)(page_size / block_size);
+  mi_assert_internal(page->reserved > 0);
+  #if (MI_PADDING || MI_ENCODE_FREELIST)
+  page->keys[0] = _mi_heap_random_next(heap);
+  page->keys[1] = _mi_heap_random_next(heap);
+  #endif
+  #if MI_DEBUG > 0
+  page->is_zero = false; // ensure in debug mode we initialize with MI_DEBUG_UNINIT, see issue #501
+  #else
+  page->is_zero = page->is_zero_init;
+  #endif
+
+  mi_assert_internal(page->is_committed);
+  mi_assert_internal(!page->is_reset);
+  mi_assert_internal(page->capacity == 0);
+  mi_assert_internal(page->free == NULL);
+  mi_assert_internal(page->used == 0);
+  mi_assert_internal(page->xthread_free == 0);
+  mi_assert_internal(page->next == NULL);
+  mi_assert_internal(page->prev == NULL);
+  mi_assert_internal(page->retire_expire == 0);
+  mi_assert_internal(!mi_page_has_aligned(page));
+  #if (MI_PADDING || MI_ENCODE_FREELIST)
+  mi_assert_internal(page->keys[0] != 0);
+  mi_assert_internal(page->keys[1] != 0);
+  #endif
+  mi_assert_expensive(mi_page_is_valid_init(page));
+
+  // initialize an initial free list
+  mi_page_extend_free(heap,page,tld);
+  mi_assert(mi_page_immediate_available(page));
+}
+
+
+/* -----------------------------------------------------------
+  Find pages with free blocks
+-------------------------------------------------------------*/
+
+// Find a page with free blocks of `page->block_size`.
+static mi_page_t* mi_page_queue_find_free_ex(mi_heap_t* heap, mi_page_queue_t* pq, bool first_try)
+{
+  // search through the pages in "next fit" order
+  #if MI_STAT
+  size_t count = 0;
+  #endif
+  mi_page_t* page = pq->first;
+  while (page != NULL)
+  {
+    mi_page_t* next = page->next; // remember next
+    #if MI_STAT    
+    count++;
+    #endif
+
+    // 0. collect freed blocks by us and other threads
+    _mi_page_free_collect(page, false);
+
+    // 1. if the page contains free blocks, we are done
+    if (mi_page_immediate_available(page)) {
+      break;  // pick this one
+    }
+
+    // 2. Try to extend
+    if (page->capacity < page->reserved) {
+      mi_page_extend_free(heap, page, heap->tld);
+      mi_assert_internal(mi_page_immediate_available(page));
+      break;
+    }
+
+    // 3. If the page is completely full, move it to the `mi_pages_full`
+    // queue so we don't visit long-lived pages too often.
+    mi_assert_internal(!mi_page_is_in_full(page) && !mi_page_immediate_available(page));
+    mi_page_to_full(page, pq);
+
+    page = next;
+  } // for each page
+
+  mi_heap_stat_counter_increase(heap, searches, count);
+
+  if (page == NULL) {
+    _mi_heap_collect_retired(heap, false); // perhaps make a page available?
+    page = mi_page_fresh(heap, pq);
+    if (page == NULL && first_try) {
+      // out-of-memory _or_ an abandoned page with free blocks was reclaimed, try once again
+      page = mi_page_queue_find_free_ex(heap, pq, false);
+    }
+  }
+  else {
+    mi_assert(pq->first == page);
+    page->retire_expire = 0;
+  }
+  mi_assert_internal(page == NULL || mi_page_immediate_available(page));
+  return page;
+}
+
+
+
+// Find a page with free blocks of `size`.
+static inline mi_page_t* mi_find_free_page(mi_heap_t* heap, size_t size) {
+  mi_page_queue_t* pq = mi_page_queue(heap,size);
+  mi_page_t* page = pq->first;
+  if (page != NULL) {
+   #if (MI_SECURE>=3) // in secure mode, we extend half the time to increase randomness
+    if (page->capacity < page->reserved && ((_mi_heap_random_next(heap) & 1) == 1)) {
+      mi_page_extend_free(heap, page, heap->tld);
+      mi_assert_internal(mi_page_immediate_available(page));
+    }
+    else
+   #endif
+    {
+      _mi_page_free_collect(page,false);
+    }
+
+    if (mi_page_immediate_available(page)) {
+      page->retire_expire = 0;
+      return page; // fast path
+    }
+  }
+  return mi_page_queue_find_free_ex(heap, pq, true);
+}
+
+
+/* -----------------------------------------------------------
+  Users can register a deferred free function called
+  when the `free` list is empty. Since the `local_free`
+  is separate this is deterministically called after
+  a certain number of allocations.
+----------------------------------------------------------- */
+
+static mi_deferred_free_fun* volatile deferred_free = NULL;
+static _Atomic(void*) deferred_arg; // = NULL
+
+void _mi_deferred_free(mi_heap_t* heap, bool force) {
+  heap->tld->heartbeat++;
+  if (deferred_free != NULL && !heap->tld->recurse) {
+    heap->tld->recurse = true;
+    deferred_free(force, heap->tld->heartbeat, mi_atomic_load_ptr_relaxed(void,&deferred_arg));
+    heap->tld->recurse = false;
+  }
+}
+
+void mi_register_deferred_free(mi_deferred_free_fun* fn, void* arg) mi_attr_noexcept {
+  deferred_free = fn;
+  mi_atomic_store_ptr_release(void,&deferred_arg, arg);
+}
+
+
+/* -----------------------------------------------------------
+  General allocation
+----------------------------------------------------------- */
+
+// Large and huge page allocation.
+// Huge pages are allocated directly without being in a queue.
+// Because huge pages contain just one block, and the segment contains
+// just that page, we always treat them as abandoned and any thread
+// that frees the block can free the whole page and segment directly.
+// Huge pages are also use if the requested alignment is very large (> MI_ALIGNMENT_MAX).
+static mi_page_t* mi_large_huge_page_alloc(mi_heap_t* heap, size_t size, size_t page_alignment) {
+  size_t block_size = _mi_os_good_alloc_size(size);
+  mi_assert_internal(mi_bin(block_size) == MI_BIN_HUGE || page_alignment > 0);
+  bool is_huge = (block_size > MI_LARGE_OBJ_SIZE_MAX || page_alignment > 0);
+  #if MI_HUGE_PAGE_ABANDON
+  mi_page_queue_t* pq = (is_huge ? NULL : mi_page_queue(heap, block_size));
+  #else
+  mi_page_queue_t* pq = mi_page_queue(heap, is_huge ? MI_HUGE_BLOCK_SIZE : block_size); // not block_size as that can be low if the page_alignment > 0
+  mi_assert_internal(!is_huge || mi_page_queue_is_huge(pq));
+  #endif
+  mi_page_t* page = mi_page_fresh_alloc(heap, pq, block_size, page_alignment);
+  if (page != NULL) {
+    mi_assert_internal(mi_page_immediate_available(page));
+    
+    if (is_huge) {
+      mi_assert_internal(_mi_page_segment(page)->kind == MI_SEGMENT_HUGE);
+      mi_assert_internal(_mi_page_segment(page)->used==1);
+      #if MI_HUGE_PAGE_ABANDON
+      mi_assert_internal(_mi_page_segment(page)->thread_id==0); // abandoned, not in the huge queue
+      mi_page_set_heap(page, NULL);
+      #endif      
+    }
+    else {
+      mi_assert_internal(_mi_page_segment(page)->kind != MI_SEGMENT_HUGE);
+    }
+    
+    const size_t bsize = mi_page_usable_block_size(page);  // note: not `mi_page_block_size` to account for padding
+    if (bsize <= MI_LARGE_OBJ_SIZE_MAX) {
+      mi_heap_stat_increase(heap, large, bsize);
+      mi_heap_stat_counter_increase(heap, large_count, 1);
+    }
+    else {
+      mi_heap_stat_increase(heap, huge, bsize);
+      mi_heap_stat_counter_increase(heap, huge_count, 1);
+    }
+  }
+  return page;
+}
+
+
+// Allocate a page
+// Note: in debug mode the size includes MI_PADDING_SIZE and might have overflowed.
+static mi_page_t* mi_find_page(mi_heap_t* heap, size_t size, size_t huge_alignment) mi_attr_noexcept {
+  // huge allocation?
+  const size_t req_size = size - MI_PADDING_SIZE;  // correct for padding_size in case of an overflow on `size`  
+  if mi_unlikely(req_size > (MI_MEDIUM_OBJ_SIZE_MAX - MI_PADDING_SIZE) || huge_alignment > 0) {
+    if mi_unlikely(req_size > PTRDIFF_MAX) {  // we don't allocate more than PTRDIFF_MAX (see <https://sourceware.org/ml/libc-announce/2019/msg00001.html>)
+      _mi_error_message(EOVERFLOW, "allocation request is too large (%zu bytes)\n", req_size);
+      return NULL;
+    }
+    else {
+      return mi_large_huge_page_alloc(heap,size,huge_alignment);
+    }
+  }
+  else {
+    // otherwise find a page with free blocks in our size segregated queues
+    #if MI_PADDING
+    mi_assert_internal(size >= MI_PADDING_SIZE); 
+    #endif
+    return mi_find_free_page(heap, size);
+  }
+}
+
+// Generic allocation routine if the fast path (`alloc.c:mi_page_malloc`) does not succeed.
+// Note: in debug mode the size includes MI_PADDING_SIZE and might have overflowed.
+// The `huge_alignment` is normally 0 but is set to a multiple of MI_SEGMENT_SIZE for
+// very large requested alignments in which case we use a huge segment.
+void* _mi_malloc_generic(mi_heap_t* heap, size_t size, bool zero, size_t huge_alignment) mi_attr_noexcept
+{
+  mi_assert_internal(heap != NULL);
+
+  // initialize if necessary
+  if mi_unlikely(!mi_heap_is_initialized(heap)) {
+    heap = mi_heap_get_default(); // calls mi_thread_init 
+    if mi_unlikely(!mi_heap_is_initialized(heap)) { return NULL; }
+  }
+  mi_assert_internal(mi_heap_is_initialized(heap));
+
+  // call potential deferred free routines
+  _mi_deferred_free(heap, false);
+
+  // free delayed frees from other threads (but skip contended ones)
+  _mi_heap_delayed_free_partial(heap);
+
+  // find (or allocate) a page of the right size
+  mi_page_t* page = mi_find_page(heap, size, huge_alignment);
+  if mi_unlikely(page == NULL) { // first time out of memory, try to collect and retry the allocation once more
+    mi_heap_collect(heap, true /* force */);
+    page = mi_find_page(heap, size, huge_alignment);
+  }
+
+  if mi_unlikely(page == NULL) { // out of memory
+    const size_t req_size = size - MI_PADDING_SIZE;  // correct for padding_size in case of an overflow on `size`
+    _mi_error_message(ENOMEM, "unable to allocate memory (%zu bytes)\n", req_size);
+    return NULL;
+  }
+
+  mi_assert_internal(mi_page_immediate_available(page));
+  mi_assert_internal(mi_page_block_size(page) >= size);
+
+  // and try again, this time succeeding! (i.e. this should never recurse through _mi_page_malloc)
+  if mi_unlikely(zero && page->xblock_size == 0) {
+    // note: we cannot call _mi_page_malloc with zeroing for huge blocks; we zero it afterwards in that case.
+    void* p = _mi_page_malloc(heap, page, size, false);
+    mi_assert_internal(p != NULL);
+    _mi_memzero_aligned(p, mi_page_usable_block_size(page));
+    return p;
+  }
+  else {
+    return _mi_page_malloc(heap, page, size, zero);
+  }
+}

preload-mimalloc/mimalloc/src/prim/osx/alloc-override-zone.c

@@ -0,0 +1,458 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2022, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+
+#if defined(MI_MALLOC_OVERRIDE)
+
+#if !defined(__APPLE__)
+#error "this file should only be included on macOS"
+#endif
+
+/* ------------------------------------------------------
+   Override system malloc on macOS
+   This is done through the malloc zone interface.
+   It seems to be most robust in combination with interposing
+   though or otherwise we may get zone errors as there are could
+   be allocations done by the time we take over the
+   zone.
+------------------------------------------------------ */
+
+#include <AvailabilityMacros.h>
+#include <malloc/malloc.h>
+#include <string.h>  // memset
+#include <stdlib.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#if defined(MAC_OS_X_VERSION_10_6) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6)
+// only available from OSX 10.6
+extern malloc_zone_t* malloc_default_purgeable_zone(void) __attribute__((weak_import));
+#endif
+
+/* ------------------------------------------------------
+   malloc zone members
+------------------------------------------------------ */
+
+static size_t zone_size(malloc_zone_t* zone, const void* p) {
+  MI_UNUSED(zone);
+  if (!mi_is_in_heap_region(p)){ return 0; } // not our pointer, bail out
+  return mi_usable_size(p);
+}
+
+static void* zone_malloc(malloc_zone_t* zone, size_t size) {
+  MI_UNUSED(zone);
+  return mi_malloc(size);
+}
+
+static void* zone_calloc(malloc_zone_t* zone, size_t count, size_t size) {
+  MI_UNUSED(zone);
+  return mi_calloc(count, size);
+}
+
+static void* zone_valloc(malloc_zone_t* zone, size_t size) {
+  MI_UNUSED(zone);
+  return mi_malloc_aligned(size, _mi_os_page_size());
+}
+
+static void zone_free(malloc_zone_t* zone, void* p) {
+  MI_UNUSED(zone);
+  mi_cfree(p);
+}
+
+static void* zone_realloc(malloc_zone_t* zone, void* p, size_t newsize) {
+  MI_UNUSED(zone);
+  return mi_realloc(p, newsize);
+}
+
+static void* zone_memalign(malloc_zone_t* zone, size_t alignment, size_t size) {
+  MI_UNUSED(zone);
+  return mi_malloc_aligned(size,alignment);
+}
+
+static void zone_destroy(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+  // todo: ignore for now?
+}
+
+static unsigned zone_batch_malloc(malloc_zone_t* zone, size_t size, void** ps, unsigned count) {
+  size_t i;
+  for (i = 0; i < count; i++) {
+    ps[i] = zone_malloc(zone, size);
+    if (ps[i] == NULL) break;
+  }
+  return i;
+}
+
+static void zone_batch_free(malloc_zone_t* zone, void** ps, unsigned count) {
+  for(size_t i = 0; i < count; i++) {
+    zone_free(zone, ps[i]);
+    ps[i] = NULL;
+  }
+}
+
+static size_t zone_pressure_relief(malloc_zone_t* zone, size_t size) {
+  MI_UNUSED(zone); MI_UNUSED(size);
+  mi_collect(false);
+  return 0;
+}
+
+static void zone_free_definite_size(malloc_zone_t* zone, void* p, size_t size) {
+  MI_UNUSED(size);
+  zone_free(zone,p);
+}
+
+static boolean_t zone_claimed_address(malloc_zone_t* zone, void* p) {
+  MI_UNUSED(zone);
+  return mi_is_in_heap_region(p);
+}
+
+
+/* ------------------------------------------------------
+   Introspection members
+------------------------------------------------------ */
+
+static kern_return_t intro_enumerator(task_t task, void* p,
+                            unsigned type_mask, vm_address_t zone_address,
+                            memory_reader_t reader,
+                            vm_range_recorder_t recorder)
+{
+  // todo: enumerate all memory
+  MI_UNUSED(task); MI_UNUSED(p); MI_UNUSED(type_mask); MI_UNUSED(zone_address);
+  MI_UNUSED(reader); MI_UNUSED(recorder);
+  return KERN_SUCCESS;
+}
+
+static size_t intro_good_size(malloc_zone_t* zone, size_t size) {
+  MI_UNUSED(zone);
+  return mi_good_size(size);
+}
+
+static boolean_t intro_check(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+  return true;
+}
+
+static void intro_print(malloc_zone_t* zone, boolean_t verbose) {
+  MI_UNUSED(zone); MI_UNUSED(verbose);
+  mi_stats_print(NULL);
+}
+
+static void intro_log(malloc_zone_t* zone, void* p) {
+  MI_UNUSED(zone); MI_UNUSED(p);
+  // todo?
+}
+
+static void intro_force_lock(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+  // todo?
+}
+
+static void intro_force_unlock(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+  // todo?
+}
+
+static void intro_statistics(malloc_zone_t* zone, malloc_statistics_t* stats) {
+  MI_UNUSED(zone);
+  // todo...
+  stats->blocks_in_use = 0;
+  stats->size_in_use = 0;
+  stats->max_size_in_use = 0;
+  stats->size_allocated = 0;
+}
+
+static boolean_t intro_zone_locked(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+  return false;
+}
+
+
+/* ------------------------------------------------------
+  At process start, override the default allocator
+------------------------------------------------------ */
+
+#if defined(__GNUC__) && !defined(__clang__)
+#pragma GCC diagnostic ignored "-Wmissing-field-initializers"
+#endif
+
+#if defined(__clang__)
+#pragma clang diagnostic ignored "-Wc99-extensions"
+#endif
+
+static malloc_introspection_t mi_introspect = {
+  .enumerator = &intro_enumerator,
+  .good_size = &intro_good_size,
+  .check = &intro_check,
+  .print = &intro_print,
+  .log = &intro_log,
+  .force_lock = &intro_force_lock,
+  .force_unlock = &intro_force_unlock,
+#if defined(MAC_OS_X_VERSION_10_6) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6)
+  .statistics = &intro_statistics,
+  .zone_locked = &intro_zone_locked,
+#endif
+};
+
+static malloc_zone_t mi_malloc_zone = {
+  // note: even with designators, the order is important for C++ compilation
+  //.reserved1 = NULL,
+  //.reserved2 = NULL,
+  .size = &zone_size,
+  .malloc = &zone_malloc,
+  .calloc = &zone_calloc,
+  .valloc = &zone_valloc,
+  .free = &zone_free,
+  .realloc = &zone_realloc,
+  .destroy = &zone_destroy,
+  .zone_name = "mimalloc",
+  .batch_malloc = &zone_batch_malloc,
+  .batch_free = &zone_batch_free,
+  .introspect = &mi_introspect,
+#if defined(MAC_OS_X_VERSION_10_6) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6)
+  #if defined(MAC_OS_X_VERSION_10_14) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_14)
+  .version = 10,
+  #else
+  .version = 9,
+  #endif
+  // switch to version 9+ on OSX 10.6 to support memalign.
+  .memalign = &zone_memalign,
+  .free_definite_size = &zone_free_definite_size,
+  .pressure_relief = &zone_pressure_relief,
+  #if defined(MAC_OS_X_VERSION_10_14) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_14)
+  .claimed_address = &zone_claimed_address,
+  #endif
+#else
+  .version = 4,
+#endif
+};
+
+#ifdef __cplusplus
+}
+#endif
+
+
+#if defined(MI_OSX_INTERPOSE) && defined(MI_SHARED_LIB_EXPORT)
+
+// ------------------------------------------------------
+// Override malloc_xxx and malloc_zone_xxx api's to use only
+// our mimalloc zone. Since even the loader uses malloc
+// on macOS, this ensures that all allocations go through
+// mimalloc (as all calls are interposed).
+// The main `malloc`, `free`, etc calls are interposed in `alloc-override.c`,
+// Here, we also override macOS specific API's like
+// `malloc_zone_calloc` etc. see <https://github.com/aosm/libmalloc/blob/master/man/malloc_zone_malloc.3>
+// ------------------------------------------------------
+
+static inline malloc_zone_t* mi_get_default_zone(void)
+{
+  static bool init;
+  if mi_unlikely(!init) {
+    init = true;
+    malloc_zone_register(&mi_malloc_zone);  // by calling register we avoid a zone error on free (see <http://eatmyrandom.blogspot.com/2010/03/mallocfree-interception-on-mac-os-x.html>)
+  }
+  return &mi_malloc_zone;
+}
+
+mi_decl_externc int  malloc_jumpstart(uintptr_t cookie);
+mi_decl_externc void _malloc_fork_prepare(void);
+mi_decl_externc void _malloc_fork_parent(void);
+mi_decl_externc void _malloc_fork_child(void);
+
+
+static malloc_zone_t* mi_malloc_create_zone(vm_size_t size, unsigned flags) {
+  MI_UNUSED(size); MI_UNUSED(flags);
+  return mi_get_default_zone();
+}
+
+static malloc_zone_t* mi_malloc_default_zone (void) {
+  return mi_get_default_zone();
+}
+
+static malloc_zone_t* mi_malloc_default_purgeable_zone(void) {
+  return mi_get_default_zone();
+}
+
+static void mi_malloc_destroy_zone(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+  // nothing.
+}
+
+static kern_return_t mi_malloc_get_all_zones (task_t task, memory_reader_t mr, vm_address_t** addresses, unsigned* count) {
+  MI_UNUSED(task); MI_UNUSED(mr);
+  if (addresses != NULL) *addresses = NULL;
+  if (count != NULL) *count = 0;
+  return KERN_SUCCESS;
+}
+
+static const char* mi_malloc_get_zone_name(malloc_zone_t* zone) {
+  return (zone == NULL ? mi_malloc_zone.zone_name : zone->zone_name);
+}
+
+static void mi_malloc_set_zone_name(malloc_zone_t* zone, const char* name) {
+  MI_UNUSED(zone); MI_UNUSED(name);
+}
+
+static int mi_malloc_jumpstart(uintptr_t cookie) {
+  MI_UNUSED(cookie);
+  return 1; // or 0 for no error?
+}
+
+static void mi__malloc_fork_prepare(void) {
+  // nothing
+}
+static void mi__malloc_fork_parent(void) {
+  // nothing
+}
+static void mi__malloc_fork_child(void) {
+  // nothing
+}
+
+static void mi_malloc_printf(const char* fmt, ...) {
+  MI_UNUSED(fmt);
+}
+
+static bool zone_check(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+  return true;
+}
+
+static malloc_zone_t* zone_from_ptr(const void* p) {
+  MI_UNUSED(p);
+  return mi_get_default_zone();
+}
+
+static void zone_log(malloc_zone_t* zone, void* p) {
+  MI_UNUSED(zone); MI_UNUSED(p);
+}
+
+static void zone_print(malloc_zone_t* zone, bool b) {
+  MI_UNUSED(zone); MI_UNUSED(b);
+}
+
+static void zone_print_ptr_info(void* p) {
+  MI_UNUSED(p);
+}
+
+static void zone_register(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+}
+
+static void zone_unregister(malloc_zone_t* zone) {
+  MI_UNUSED(zone);
+}
+
+// use interposing so `DYLD_INSERT_LIBRARIES` works without `DYLD_FORCE_FLAT_NAMESPACE=1`
+// See: <https://books.google.com/books?id=K8vUkpOXhN4C&pg=PA73>
+struct mi_interpose_s {
+  const void* replacement;
+  const void* target;
+};
+#define MI_INTERPOSE_FUN(oldfun,newfun) { (const void*)&newfun, (const void*)&oldfun }
+#define MI_INTERPOSE_MI(fun)            MI_INTERPOSE_FUN(fun,mi_##fun)
+#define MI_INTERPOSE_ZONE(fun)          MI_INTERPOSE_FUN(malloc_##fun,fun)
+__attribute__((used)) static const struct mi_interpose_s _mi_zone_interposes[]  __attribute__((section("__DATA, __interpose"))) =
+{
+
+  MI_INTERPOSE_MI(malloc_create_zone),
+  MI_INTERPOSE_MI(malloc_default_purgeable_zone),
+  MI_INTERPOSE_MI(malloc_default_zone),
+  MI_INTERPOSE_MI(malloc_destroy_zone),
+  MI_INTERPOSE_MI(malloc_get_all_zones),
+  MI_INTERPOSE_MI(malloc_get_zone_name),
+  MI_INTERPOSE_MI(malloc_jumpstart),
+  MI_INTERPOSE_MI(malloc_printf),
+  MI_INTERPOSE_MI(malloc_set_zone_name),
+  MI_INTERPOSE_MI(_malloc_fork_child),
+  MI_INTERPOSE_MI(_malloc_fork_parent),
+  MI_INTERPOSE_MI(_malloc_fork_prepare),
+
+  MI_INTERPOSE_ZONE(zone_batch_free),
+  MI_INTERPOSE_ZONE(zone_batch_malloc),
+  MI_INTERPOSE_ZONE(zone_calloc),
+  MI_INTERPOSE_ZONE(zone_check),
+  MI_INTERPOSE_ZONE(zone_free),
+  MI_INTERPOSE_ZONE(zone_from_ptr),
+  MI_INTERPOSE_ZONE(zone_log),
+  MI_INTERPOSE_ZONE(zone_malloc),
+  MI_INTERPOSE_ZONE(zone_memalign),
+  MI_INTERPOSE_ZONE(zone_print),
+  MI_INTERPOSE_ZONE(zone_print_ptr_info),
+  MI_INTERPOSE_ZONE(zone_realloc),
+  MI_INTERPOSE_ZONE(zone_register),
+  MI_INTERPOSE_ZONE(zone_unregister),
+  MI_INTERPOSE_ZONE(zone_valloc)
+};
+
+
+#else
+
+// ------------------------------------------------------
+// hook into the zone api's without interposing
+// This is the official way of adding an allocator but
+// it seems less robust than using interpose.
+// ------------------------------------------------------
+
+static inline malloc_zone_t* mi_get_default_zone(void)
+{
+  // The first returned zone is the real default
+  malloc_zone_t** zones = NULL;
+  unsigned count = 0;
+  kern_return_t ret = malloc_get_all_zones(0, NULL, (vm_address_t**)&zones, &count);
+  if (ret == KERN_SUCCESS && count > 0) {
+    return zones[0];
+  }
+  else {
+    // fallback
+    return malloc_default_zone();
+  }
+}
+
+#if defined(__clang__)
+__attribute__((constructor(0)))
+#else
+__attribute__((constructor))      // seems not supported by g++-11 on the M1
+#endif
+static void _mi_macos_override_malloc(void) {
+  malloc_zone_t* purgeable_zone = NULL;
+
+  #if defined(MAC_OS_X_VERSION_10_6) && (MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_6)
+  // force the purgeable zone to exist to avoid strange bugs
+  if (malloc_default_purgeable_zone) {
+    purgeable_zone = malloc_default_purgeable_zone();
+  }
+  #endif
+
+  // Register our zone.
+  // thomcc: I think this is still needed to put us in the zone list.
+  malloc_zone_register(&mi_malloc_zone);
+  // Unregister the default zone, this makes our zone the new default
+  // as that was the last registered.
+  malloc_zone_t *default_zone = mi_get_default_zone();
+  // thomcc: Unsure if the next test is *always* false or just false in the
+  // cases I've tried. I'm also unsure if the code inside is needed. at all
+  if (default_zone != &mi_malloc_zone) {
+    malloc_zone_unregister(default_zone);
+
+    // Reregister the default zone so free and realloc in that zone keep working.
+    malloc_zone_register(default_zone);
+  }
+
+  // Unregister, and re-register the purgeable_zone to avoid bugs if it occurs
+  // earlier than the default zone.
+  if (purgeable_zone != NULL) {
+    malloc_zone_unregister(purgeable_zone);
+    malloc_zone_register(purgeable_zone);
+  }
+
+}
+#endif  // MI_OSX_INTERPOSE
+
+#endif // MI_MALLOC_OVERRIDE

preload-mimalloc/mimalloc/src/prim/osx/prim.c

@@ -0,0 +1,9 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+// We use the unix/prim.c with the mmap API on macOSX
+#include "../unix/prim.c"

preload-mimalloc/mimalloc/src/prim/prim.c

@@ -0,0 +1,24 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+// Select the implementation of the primitives
+// depending on the OS.
+
+#if defined(_WIN32)
+#include "windows/prim.c"  // VirtualAlloc (Windows)
+
+#elif defined(__APPLE__)
+#include "osx/prim.c"      // macOSX (actually defers to mmap in unix/prim.c)
+
+#elif defined(__wasi__)
+#define MI_USE_SBRK
+#include "wasi/prim.c"     // memory-grow or sbrk (Wasm)
+
+#else
+#include "unix/prim.c"     // mmap() (Linux, macOSX, BSD, Illumnos, Haiku, DragonFly, etc.)
+
+#endif

preload-mimalloc/mimalloc/src/prim/readme.md

@@ -0,0 +1,9 @@
+## Portability Primitives
+
+This is the portability layer where all primitives needed from the OS are defined.
+
+- `include/mimalloc/prim.h`: primitive portability API definition.
+- `prim.c`: Selects one of `unix/prim.c`, `wasi/prim.c`, or `windows/prim.c` depending on the host platform
+            (and on macOS, `osx/prim.c` defers to `unix/prim.c`).
+
+Note: still work in progress, there may still be places in the sources that still depend on OS ifdef's.

preload-mimalloc/mimalloc/src/prim/unix/prim.c

@@ -0,0 +1,838 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+// This file is included in `src/prim/prim.c`
+
+#ifndef _DEFAULT_SOURCE
+#define _DEFAULT_SOURCE   // ensure mmap flags and syscall are defined
+#endif
+
+#if defined(__sun)
+// illumos provides new mman.h api when any of these are defined
+// otherwise the old api based on caddr_t which predates the void pointers one.
+// stock solaris provides only the former, chose to atomically to discard those
+// flags only here rather than project wide tough.
+#undef _XOPEN_SOURCE
+#undef _POSIX_C_SOURCE
+#endif
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+
+#include <sys/mman.h>  // mmap
+#include <unistd.h>    // sysconf
+
+#if defined(__linux__)
+  #include <features.h>
+  #include <fcntl.h>
+  #if defined(__GLIBC__)
+  #include <linux/mman.h> // linux mmap flags
+  #else
+  #include <sys/mman.h>
+  #endif
+#elif defined(__APPLE__)
+  #include <TargetConditionals.h>
+  #if !TARGET_IOS_IPHONE && !TARGET_IOS_SIMULATOR
+  #include <mach/vm_statistics.h>
+  #endif
+#elif defined(__FreeBSD__) || defined(__DragonFly__)
+  #include <sys/param.h>
+  #if __FreeBSD_version >= 1200000
+  #include <sys/cpuset.h>
+  #include <sys/domainset.h>
+  #endif
+  #include <sys/sysctl.h>
+#endif
+
+#if !defined(__HAIKU__) && !defined(__APPLE__) && !defined(__CYGWIN__)
+  #define MI_HAS_SYSCALL_H
+  #include <sys/syscall.h>
+#endif
+
+//------------------------------------------------------------------------------------
+// Use syscalls for some primitives to allow for libraries that override open/read/close etc.
+// and do allocation themselves; using syscalls prevents recursion when mimalloc is 
+// still initializing (issue #713)
+//------------------------------------------------------------------------------------
+
+#if defined(MI_HAS_SYSCALL_H) && defined(SYS_open) && defined(SYS_close) && defined(SYS_read) && defined(SYS_access)
+
+static int mi_prim_open(const char* fpath, int open_flags) {
+  return syscall(SYS_open,fpath,open_flags,0);
+}
+static ssize_t mi_prim_read(int fd, void* buf, size_t bufsize) {
+  return syscall(SYS_read,fd,buf,bufsize);
+}
+static int mi_prim_close(int fd) {
+  return syscall(SYS_close,fd);
+}
+static int mi_prim_access(const char *fpath, int mode) {
+  return syscall(SYS_access,fpath,mode);
+}
+
+#elif !defined(__APPLE__)  // avoid unused warnings
+
+static int mi_prim_open(const char* fpath, int open_flags) {
+  return open(fpath,open_flags);
+}
+static ssize_t mi_prim_read(int fd, void* buf, size_t bufsize) {
+  return read(fd,buf,bufsize);
+}
+static int mi_prim_close(int fd) {
+  return close(fd);
+}
+static int mi_prim_access(const char *fpath, int mode) {
+  return access(fpath,mode);
+}
+
+#endif
+
+
+
+//---------------------------------------------
+// init
+//---------------------------------------------
+
+static bool unix_detect_overcommit(void) {
+  bool os_overcommit = true;
+#if defined(__linux__)
+  int fd = mi_prim_open("/proc/sys/vm/overcommit_memory", O_RDONLY);
+	if (fd >= 0) {
+    char buf[32];
+    ssize_t nread = mi_prim_read(fd, &buf, sizeof(buf));
+    mi_prim_close(fd);
+    // <https://www.kernel.org/doc/Documentation/vm/overcommit-accounting>
+    // 0: heuristic overcommit, 1: always overcommit, 2: never overcommit (ignore NORESERVE)
+    if (nread >= 1) {
+      os_overcommit = (buf[0] == '0' || buf[0] == '1');
+    }
+  }
+#elif defined(__FreeBSD__)
+  int val = 0;
+  size_t olen = sizeof(val);
+  if (sysctlbyname("vm.overcommit", &val, &olen, NULL, 0) == 0) {
+    os_overcommit = (val != 0);
+  }
+#else
+  // default: overcommit is true  
+#endif
+  return os_overcommit;
+}
+
+void _mi_prim_mem_init( mi_os_mem_config_t* config ) {
+  long psize = sysconf(_SC_PAGESIZE);
+  if (psize > 0) {
+    config->page_size = (size_t)psize;
+    config->alloc_granularity = (size_t)psize;
+  }
+  config->large_page_size = 2*MI_MiB; // TODO: can we query the OS for this?
+  config->has_overcommit = unix_detect_overcommit();
+  config->must_free_whole = false;    // mmap can free in parts
+}
+
+
+//---------------------------------------------
+// free
+//---------------------------------------------
+
+int _mi_prim_free(void* addr, size_t size ) {
+  bool err = (munmap(addr, size) == -1);
+  return (err ? errno : 0);
+}
+
+
+//---------------------------------------------
+// mmap
+//---------------------------------------------
+
+static int unix_madvise(void* addr, size_t size, int advice) {
+  #if defined(__sun)
+  return madvise((caddr_t)addr, size, advice);  // Solaris needs cast (issue #520)
+  #else
+  return madvise(addr, size, advice);
+  #endif
+}
+
+static void* unix_mmap_prim(void* addr, size_t size, size_t try_alignment, int protect_flags, int flags, int fd) {
+  MI_UNUSED(try_alignment);
+  void* p = NULL;
+  #if defined(MAP_ALIGNED)  // BSD
+  if (addr == NULL && try_alignment > 1 && (try_alignment % _mi_os_page_size()) == 0) {
+    size_t n = mi_bsr(try_alignment);
+    if (((size_t)1 << n) == try_alignment && n >= 12 && n <= 30) {  // alignment is a power of 2 and 4096 <= alignment <= 1GiB
+      p = mmap(addr, size, protect_flags, flags | MAP_ALIGNED(n), fd, 0);
+      if (p==MAP_FAILED || !_mi_is_aligned(p,try_alignment)) { 
+        int err = errno;
+        _mi_warning_message("unable to directly request aligned OS memory (error: %d (0x%x), size: 0x%zx bytes, alignment: 0x%zx, hint address: %p)\n", err, err, size, try_alignment, hint);
+      }
+      if (p!=MAP_FAILED) return p;
+      // fall back to regular mmap      
+    }
+  }
+  #elif defined(MAP_ALIGN)  // Solaris
+  if (addr == NULL && try_alignment > 1 && (try_alignment % _mi_os_page_size()) == 0) {
+    p = mmap((void*)try_alignment, size, protect_flags, flags | MAP_ALIGN, fd, 0);  // addr parameter is the required alignment
+    if (p!=MAP_FAILED) return p;
+    // fall back to regular mmap
+  }
+  #endif
+  #if (MI_INTPTR_SIZE >= 8) && !defined(MAP_ALIGNED)
+  // on 64-bit systems, use the virtual address area after 2TiB for 4MiB aligned allocations
+  if (addr == NULL) {
+    void* hint = _mi_os_get_aligned_hint(try_alignment, size);
+    if (hint != NULL) {
+      p = mmap(hint, size, protect_flags, flags, fd, 0);
+      if (p==MAP_FAILED || !_mi_is_aligned(p,try_alignment)) { 
+        int err = errno;
+        _mi_warning_message("unable to directly request hinted aligned OS memory (error: %d (0x%x), size: 0x%zx bytes, alignment: 0x%zx, hint address: %p)\n", err, err, size, try_alignment, hint);
+      }
+      if (p!=MAP_FAILED) return p;
+      // fall back to regular mmap      
+    }
+  }
+  #endif
+  // regular mmap
+  p = mmap(addr, size, protect_flags, flags, fd, 0);
+  if (p!=MAP_FAILED) return p;
+  // failed to allocate
+  return NULL;
+}
+
+static void* unix_mmap(void* addr, size_t size, size_t try_alignment, int protect_flags, bool large_only, bool allow_large, bool* is_large) {
+  void* p = NULL;
+  #if !defined(MAP_ANONYMOUS)
+  #define MAP_ANONYMOUS  MAP_ANON
+  #endif
+  #if !defined(MAP_NORESERVE)
+  #define MAP_NORESERVE  0
+  #endif
+  int flags = MAP_PRIVATE | MAP_ANONYMOUS;
+  int fd = -1;
+  if (_mi_os_has_overcommit()) {
+    flags |= MAP_NORESERVE;
+  }
+  #if defined(PROT_MAX)
+  protect_flags |= PROT_MAX(PROT_READ | PROT_WRITE); // BSD
+  #endif
+  #if defined(VM_MAKE_TAG)
+  // macOS: tracking anonymous page with a specific ID. (All up to 98 are taken officially but LLVM sanitizers had taken 99)
+  int os_tag = (int)mi_option_get(mi_option_os_tag);
+  if (os_tag < 100 || os_tag > 255) { os_tag = 100; }
+  fd = VM_MAKE_TAG(os_tag);
+  #endif
+  // huge page allocation
+  if ((large_only || _mi_os_use_large_page(size, try_alignment)) && allow_large) {
+    static _Atomic(size_t) large_page_try_ok; // = 0;
+    size_t try_ok = mi_atomic_load_acquire(&large_page_try_ok);
+    if (!large_only && try_ok > 0) {
+      // If the OS is not configured for large OS pages, or the user does not have
+      // enough permission, the `mmap` will always fail (but it might also fail for other reasons).
+      // Therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times
+      // to avoid too many failing calls to mmap.
+      mi_atomic_cas_strong_acq_rel(&large_page_try_ok, &try_ok, try_ok - 1);
+    }
+    else {
+      int lflags = flags & ~MAP_NORESERVE;  // using NORESERVE on huge pages seems to fail on Linux
+      int lfd = fd;
+      #ifdef MAP_ALIGNED_SUPER
+      lflags |= MAP_ALIGNED_SUPER;
+      #endif
+      #ifdef MAP_HUGETLB
+      lflags |= MAP_HUGETLB;
+      #endif
+      #ifdef MAP_HUGE_1GB
+      static bool mi_huge_pages_available = true;
+      if ((size % MI_GiB) == 0 && mi_huge_pages_available) {
+        lflags |= MAP_HUGE_1GB;
+      }
+      else
+      #endif
+      {
+        #ifdef MAP_HUGE_2MB
+        lflags |= MAP_HUGE_2MB;
+        #endif
+      }
+      #ifdef VM_FLAGS_SUPERPAGE_SIZE_2MB
+      lfd |= VM_FLAGS_SUPERPAGE_SIZE_2MB;
+      #endif
+      if (large_only || lflags != flags) {
+        // try large OS page allocation
+        *is_large = true;
+        p = unix_mmap_prim(addr, size, try_alignment, protect_flags, lflags, lfd);
+        #ifdef MAP_HUGE_1GB
+        if (p == NULL && (lflags & MAP_HUGE_1GB) != 0) {
+          mi_huge_pages_available = false; // don't try huge 1GiB pages again
+          _mi_warning_message("unable to allocate huge (1GiB) page, trying large (2MiB) pages instead (errno: %i)\n", errno);
+          lflags = ((lflags & ~MAP_HUGE_1GB) | MAP_HUGE_2MB);
+          p = unix_mmap_prim(addr, size, try_alignment, protect_flags, lflags, lfd);
+        }
+        #endif
+        if (large_only) return p;
+        if (p == NULL) {
+          mi_atomic_store_release(&large_page_try_ok, (size_t)8);  // on error, don't try again for the next N allocations
+        }
+      }
+    }
+  }
+  // regular allocation
+  if (p == NULL) {
+    *is_large = false;
+    p = unix_mmap_prim(addr, size, try_alignment, protect_flags, flags, fd);
+    if (p != NULL) {
+      #if defined(MADV_HUGEPAGE)
+      // Many Linux systems don't allow MAP_HUGETLB but they support instead
+      // transparent huge pages (THP). Generally, it is not required to call `madvise` with MADV_HUGE
+      // though since properly aligned allocations will already use large pages if available
+      // in that case -- in particular for our large regions (in `memory.c`).
+      // However, some systems only allow THP if called with explicit `madvise`, so
+      // when large OS pages are enabled for mimalloc, we call `madvise` anyways.
+      if (allow_large && _mi_os_use_large_page(size, try_alignment)) {
+        if (unix_madvise(p, size, MADV_HUGEPAGE) == 0) {
+          *is_large = true; // possibly
+        };
+      }
+      #elif defined(__sun)
+      if (allow_large && _mi_os_use_large_page(size, try_alignment)) {
+        struct memcntl_mha cmd = {0};
+        cmd.mha_pagesize = large_os_page_size;
+        cmd.mha_cmd = MHA_MAPSIZE_VA;
+        if (memcntl((caddr_t)p, size, MC_HAT_ADVISE, (caddr_t)&cmd, 0, 0) == 0) {
+          *is_large = true;
+        }
+      }
+      #endif
+    }
+  }
+  return p;
+}
+
+// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
+int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, void** addr) {
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  mi_assert_internal(commit || !allow_large);
+  mi_assert_internal(try_alignment > 0);
+  
+  int protect_flags = (commit ? (PROT_WRITE | PROT_READ) : PROT_NONE);
+  *addr = unix_mmap(NULL, size, try_alignment, protect_flags, false, allow_large, is_large);
+  return (*addr != NULL ? 0 : errno);
+}
+
+
+//---------------------------------------------
+// Commit/Reset
+//---------------------------------------------
+
+static void unix_mprotect_hint(int err) {
+  #if defined(__linux__) && (MI_SECURE>=2) // guard page around every mimalloc page
+  if (err == ENOMEM) {
+    _mi_warning_message("The next warning may be caused by a low memory map limit.\n"
+                        "  On Linux this is controlled by the vm.max_map_count -- maybe increase it?\n"
+                        "  For example: sudo sysctl -w vm.max_map_count=262144\n");
+  }
+  #else
+  MI_UNUSED(err);
+  #endif
+}
+
+
+int _mi_prim_commit(void* start, size_t size, bool commit) {
+  /*
+  #if 0 && defined(MAP_FIXED) && !defined(__APPLE__)
+  // Linux: disabled for now as mmap fixed seems much more expensive than MADV_DONTNEED (and splits VMA's?)
+  if (commit) {
+    // commit: just change the protection
+    err = mprotect(start, csize, (PROT_READ | PROT_WRITE));
+    if (err != 0) { err = errno; }
+  }
+  else {
+    // decommit: use mmap with MAP_FIXED to discard the existing memory (and reduce rss)
+    const int fd = mi_unix_mmap_fd();
+    void* p = mmap(start, csize, PROT_NONE, (MAP_FIXED | MAP_PRIVATE | MAP_ANONYMOUS | MAP_NORESERVE), fd, 0);
+    if (p != start) { err = errno; }
+  }
+  #else
+  */
+  int err = 0;
+  if (commit) {
+    // commit: ensure we can access the area
+    err = mprotect(start, size, (PROT_READ | PROT_WRITE));
+    if (err != 0) { err = errno; }
+  }
+  else {
+    #if defined(MADV_DONTNEED) && MI_DEBUG == 0 && MI_SECURE == 0
+    // decommit: use MADV_DONTNEED as it decreases rss immediately (unlike MADV_FREE)
+    // (on the other hand, MADV_FREE would be good enough.. it is just not reflected in the stats :-( )
+    err = unix_madvise(start, size, MADV_DONTNEED);
+    #else
+    // decommit: just disable access (also used in debug and secure mode to trap on illegal access)
+    err = mprotect(start, size, PROT_NONE);
+    if (err != 0) { err = errno; }
+    #endif    
+  }
+  unix_mprotect_hint(err);
+  return err;
+}
+
+int _mi_prim_reset(void* start, size_t size) {
+  #if defined(MADV_FREE)
+  static _Atomic(size_t) advice = MI_ATOMIC_VAR_INIT(MADV_FREE);
+  int oadvice = (int)mi_atomic_load_relaxed(&advice);
+  int err;
+  while ((err = unix_madvise(start, size, oadvice)) != 0 && errno == EAGAIN) { errno = 0;  };
+  if (err != 0 && errno == EINVAL && oadvice == MADV_FREE) {
+    // if MADV_FREE is not supported, fall back to MADV_DONTNEED from now on
+    mi_atomic_store_release(&advice, (size_t)MADV_DONTNEED);
+    err = unix_madvise(start, size, MADV_DONTNEED);
+  }
+  #else
+  int err = unix_madvise(start, size, MADV_DONTNEED);
+  #endif
+  return err;
+}
+
+int _mi_prim_protect(void* start, size_t size, bool protect) {
+  int err = mprotect(start, size, protect ? PROT_NONE : (PROT_READ | PROT_WRITE));
+  if (err != 0) { err = errno; }  
+  unix_mprotect_hint(err);
+  return err;
+}
+
+
+
+//---------------------------------------------
+// Huge page allocation
+//---------------------------------------------
+
+#if (MI_INTPTR_SIZE >= 8) && !defined(__HAIKU__) && !defined(__CYGWIN__)
+
+#ifndef MPOL_PREFERRED
+#define MPOL_PREFERRED 1
+#endif
+
+#if defined(MI_HAS_SYSCALL_H) && defined(SYS_mbind)
+static long mi_prim_mbind(void* start, unsigned long len, unsigned long mode, const unsigned long* nmask, unsigned long maxnode, unsigned flags) {
+  return syscall(SYS_mbind, start, len, mode, nmask, maxnode, flags);
+}
+#else
+static long mi_prim_mbind(void* start, unsigned long len, unsigned long mode, const unsigned long* nmask, unsigned long maxnode, unsigned flags) {
+  MI_UNUSED(start); MI_UNUSED(len); MI_UNUSED(mode); MI_UNUSED(nmask); MI_UNUSED(maxnode); MI_UNUSED(flags);
+  return 0;
+}
+#endif
+
+int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, void** addr) {
+  bool is_large = true;
+  *addr = unix_mmap(hint_addr, size, MI_SEGMENT_SIZE, PROT_READ | PROT_WRITE, true, true, &is_large);
+  if (*addr != NULL && numa_node >= 0 && numa_node < 8*MI_INTPTR_SIZE) { // at most 64 nodes
+    unsigned long numa_mask = (1UL << numa_node);
+    // TODO: does `mbind` work correctly for huge OS pages? should we
+    // use `set_mempolicy` before calling mmap instead?
+    // see: <https://lkml.org/lkml/2017/2/9/875>
+    long err = mi_prim_mbind(*addr, size, MPOL_PREFERRED, &numa_mask, 8*MI_INTPTR_SIZE, 0);
+    if (err != 0) {
+      err = errno;
+      _mi_warning_message("failed to bind huge (1GiB) pages to numa node %d (error: %d (0x%x))\n", numa_node, err, err);
+    }    
+  }
+  return (*addr != NULL ? 0 : errno);
+}
+
+#else
+
+int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, void** addr) {
+  MI_UNUSED(hint_addr); MI_UNUSED(size); MI_UNUSED(numa_node);
+  *addr = NULL;
+  return ENOMEM;
+}
+
+#endif
+
+//---------------------------------------------
+// NUMA nodes
+//---------------------------------------------
+
+#if defined(__linux__)
+
+#include <stdio.h>    // snprintf
+
+size_t _mi_prim_numa_node(void) {
+  #if defined(MI_HAS_SYSCALL_H) && defined(SYS_getcpu)
+    unsigned long node = 0;
+    unsigned long ncpu = 0;
+    long err = syscall(SYS_getcpu, &ncpu, &node, NULL);
+    if (err != 0) return 0;
+    return node;
+  #else
+    return 0;
+  #endif
+}
+
+size_t _mi_prim_numa_node_count(void) {
+  char buf[128];
+  unsigned node = 0;
+  for(node = 0; node < 256; node++) {
+    // enumerate node entries -- todo: it there a more efficient way to do this? (but ensure there is no allocation)
+    snprintf(buf, 127, "/sys/devices/system/node/node%u", node + 1);
+    if (mi_prim_access(buf,R_OK) != 0) break;
+  }
+  return (node+1);
+}
+
+#elif defined(__FreeBSD__) && __FreeBSD_version >= 1200000
+
+size_t _mi_prim_numa_node(void) {
+  domainset_t dom;
+  size_t node;
+  int policy;
+  if (cpuset_getdomain(CPU_LEVEL_CPUSET, CPU_WHICH_PID, -1, sizeof(dom), &dom, &policy) == -1) return 0ul;
+  for (node = 0; node < MAXMEMDOM; node++) {
+    if (DOMAINSET_ISSET(node, &dom)) return node;
+  }
+  return 0ul;
+}
+
+size_t _mi_prim_numa_node_count(void) {
+  size_t ndomains = 0;
+  size_t len = sizeof(ndomains);
+  if (sysctlbyname("vm.ndomains", &ndomains, &len, NULL, 0) == -1) return 0ul;
+  return ndomains;
+}
+
+#elif defined(__DragonFly__)
+
+size_t _mi_prim_numa_node(void) {
+  // TODO: DragonFly does not seem to provide any userland means to get this information.
+  return 0ul;
+}
+
+size_t _mi_prim_numa_node_count(void) {
+  size_t ncpus = 0, nvirtcoresperphys = 0;
+  size_t len = sizeof(size_t);
+  if (sysctlbyname("hw.ncpu", &ncpus, &len, NULL, 0) == -1) return 0ul;
+  if (sysctlbyname("hw.cpu_topology_ht_ids", &nvirtcoresperphys, &len, NULL, 0) == -1) return 0ul;
+  return nvirtcoresperphys * ncpus;
+}
+
+#else
+
+size_t _mi_prim_numa_node(void) {
+  return 0;
+}
+
+size_t _mi_prim_numa_node_count(void) {
+  return 1;
+}
+
+#endif
+
+// ----------------------------------------------------------------
+// Clock
+// ----------------------------------------------------------------
+
+#include <time.h>
+
+#if defined(CLOCK_REALTIME) || defined(CLOCK_MONOTONIC)
+
+mi_msecs_t _mi_prim_clock_now(void) {
+  struct timespec t;
+  #ifdef CLOCK_MONOTONIC
+  clock_gettime(CLOCK_MONOTONIC, &t);
+  #else
+  clock_gettime(CLOCK_REALTIME, &t);
+  #endif
+  return ((mi_msecs_t)t.tv_sec * 1000) + ((mi_msecs_t)t.tv_nsec / 1000000);
+}
+
+#else
+
+// low resolution timer
+mi_msecs_t _mi_prim_clock_now(void) {
+  #if !defined(CLOCKS_PER_SEC) || (CLOCKS_PER_SEC == 1000) || (CLOCKS_PER_SEC == 0)
+  return (mi_msecs_t)clock();  
+  #elif (CLOCKS_PER_SEC < 1000)
+  return (mi_msecs_t)clock() * (1000 / (mi_msecs_t)CLOCKS_PER_SEC);  
+  #else
+  return (mi_msecs_t)clock() / ((mi_msecs_t)CLOCKS_PER_SEC / 1000);
+  #endif
+}
+
+#endif
+
+
+
+
+//----------------------------------------------------------------
+// Process info
+//----------------------------------------------------------------
+
+#if defined(__unix__) || defined(__unix) || defined(unix) || defined(__APPLE__) || defined(__HAIKU__)
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/resource.h>
+
+#if defined(__APPLE__)
+#include <mach/mach.h>
+#endif
+
+#if defined(__HAIKU__)
+#include <kernel/OS.h>
+#endif
+
+static mi_msecs_t timeval_secs(const struct timeval* tv) {
+  return ((mi_msecs_t)tv->tv_sec * 1000L) + ((mi_msecs_t)tv->tv_usec / 1000L);
+}
+
+void _mi_prim_process_info(mi_process_info_t* pinfo)
+{
+  struct rusage rusage;
+  getrusage(RUSAGE_SELF, &rusage);
+  pinfo->utime = timeval_secs(&rusage.ru_utime);
+  pinfo->stime = timeval_secs(&rusage.ru_stime);
+#if !defined(__HAIKU__)
+  pinfo->page_faults = rusage.ru_majflt;
+#endif  
+#if defined(__HAIKU__)
+  // Haiku does not have (yet?) a way to
+  // get these stats per process
+  thread_info tid;
+  area_info mem;
+  ssize_t c;
+  get_thread_info(find_thread(0), &tid);
+  while (get_next_area_info(tid.team, &c, &mem) == B_OK) {
+    pinfo->peak_rss += mem.ram_size;
+  }
+  pinfo->page_faults = 0;
+#elif defined(__APPLE__)
+  pinfo->peak_rss = rusage.ru_maxrss;         // macos reports in bytes
+  struct mach_task_basic_info info;
+  mach_msg_type_number_t infoCount = MACH_TASK_BASIC_INFO_COUNT;
+  if (task_info(mach_task_self(), MACH_TASK_BASIC_INFO, (task_info_t)&info, &infoCount) == KERN_SUCCESS) {
+    pinfo->current_rss = (size_t)info.resident_size;
+  }
+#else
+  pinfo->peak_rss = rusage.ru_maxrss * 1024;  // Linux/BSD report in KiB
+#endif
+  // use defaults for commit
+}
+
+#else
+
+#ifndef __wasi__
+// WebAssembly instances are not processes
+#pragma message("define a way to get process info")
+#endif
+
+void _mi_prim_process_info(mi_process_info_t* pinfo)
+{
+  // use defaults
+  MI_UNUSED(pinfo);
+}
+
+#endif
+
+
+//----------------------------------------------------------------
+// Output
+//----------------------------------------------------------------
+
+void _mi_prim_out_stderr( const char* msg ) {
+  fputs(msg,stderr);
+}
+
+
+//----------------------------------------------------------------
+// Environment
+//----------------------------------------------------------------
+
+#if !defined(MI_USE_ENVIRON) || (MI_USE_ENVIRON!=0)
+// On Posix systemsr use `environ` to access environment variables
+// even before the C runtime is initialized.
+#if defined(__APPLE__) && defined(__has_include) && __has_include(<crt_externs.h>)
+#include <crt_externs.h>
+static char** mi_get_environ(void) {
+  return (*_NSGetEnviron());
+}
+#else
+extern char** environ;
+static char** mi_get_environ(void) {
+  return environ;
+}
+#endif
+bool _mi_prim_getenv(const char* name, char* result, size_t result_size) {
+  if (name==NULL) return false;
+  const size_t len = _mi_strlen(name);
+  if (len == 0) return false;
+  char** env = mi_get_environ();
+  if (env == NULL) return false;
+  // compare up to 10000 entries
+  for (int i = 0; i < 10000 && env[i] != NULL; i++) {
+    const char* s = env[i];
+    if (_mi_strnicmp(name, s, len) == 0 && s[len] == '=') { // case insensitive
+      // found it
+      _mi_strlcpy(result, s + len + 1, result_size);
+      return true;
+    }
+  }
+  return false;
+}
+#else
+// fallback: use standard C `getenv` but this cannot be used while initializing the C runtime
+bool _mi_prim_getenv(const char* name, char* result, size_t result_size) {
+  // cannot call getenv() when still initializing the C runtime.
+  if (_mi_preloading()) return false;
+  const char* s = getenv(name);
+  if (s == NULL) {
+    // we check the upper case name too.
+    char buf[64+1];
+    size_t len = _mi_strnlen(name,sizeof(buf)-1);
+    for (size_t i = 0; i < len; i++) {
+      buf[i] = _mi_toupper(name[i]);
+    }
+    buf[len] = 0;
+    s = getenv(buf);
+  }
+  if (s == NULL || _mi_strnlen(s,result_size) >= result_size)  return false;
+  _mi_strlcpy(result, s, result_size);
+  return true;
+}
+#endif  // !MI_USE_ENVIRON
+
+
+//----------------------------------------------------------------
+// Random
+//----------------------------------------------------------------
+
+#if defined(__APPLE__)
+
+#include <AvailabilityMacros.h>
+#if defined(MAC_OS_X_VERSION_10_10) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_10
+#include <CommonCrypto/CommonCryptoError.h>
+#include <CommonCrypto/CommonRandom.h>
+#endif
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  #if defined(MAC_OS_X_VERSION_10_15) && MAC_OS_X_VERSION_MAX_ALLOWED >= MAC_OS_X_VERSION_10_15
+    // We prefere CCRandomGenerateBytes as it returns an error code while arc4random_buf
+    // may fail silently on macOS. See PR #390, and <https://opensource.apple.com/source/Libc/Libc-1439.40.11/gen/FreeBSD/arc4random.c.auto.html>
+    return (CCRandomGenerateBytes(buf, buf_len) == kCCSuccess);
+  #else
+    // fall back on older macOS
+    arc4random_buf(buf, buf_len);
+    return true;
+  #endif
+}
+
+#elif defined(__ANDROID__) || defined(__DragonFly__) || \
+      defined(__FreeBSD__) || defined(__NetBSD__) || defined(__OpenBSD__) || \
+      defined(__sun) 
+
+#include <stdlib.h>
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  arc4random_buf(buf, buf_len);
+  return true;
+}
+
+#elif defined(__linux__) || defined(__HAIKU__)
+
+#include <sys/types.h>
+#include <sys/stat.h>
+#include <fcntl.h>
+#include <errno.h>
+
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  // Modern Linux provides `getrandom` but different distributions either use `sys/random.h` or `linux/random.h`
+  // and for the latter the actual `getrandom` call is not always defined.
+  // (see <https://stackoverflow.com/questions/45237324/why-doesnt-getrandom-compile>)
+  // We therefore use a syscall directly and fall back dynamically to /dev/urandom when needed.
+  #if defined(MI_HAS_SYSCALL_H) && defined(SYS_getrandom)
+    #ifndef GRND_NONBLOCK
+    #define GRND_NONBLOCK (1)
+    #endif
+    static _Atomic(uintptr_t) no_getrandom; // = 0
+    if (mi_atomic_load_acquire(&no_getrandom)==0) {
+      ssize_t ret = syscall(SYS_getrandom, buf, buf_len, GRND_NONBLOCK);
+      if (ret >= 0) return (buf_len == (size_t)ret);
+      if (errno != ENOSYS) return false;
+      mi_atomic_store_release(&no_getrandom, (uintptr_t)1); // don't call again, and fall back to /dev/urandom
+    }
+  #endif
+  int flags = O_RDONLY;
+  #if defined(O_CLOEXEC)
+  flags |= O_CLOEXEC;
+  #endif
+  int fd = mi_prim_open("/dev/urandom", flags);
+  if (fd < 0) return false;
+  size_t count = 0;
+  while(count < buf_len) {
+    ssize_t ret = mi_prim_read(fd, (char*)buf + count, buf_len - count);
+    if (ret<=0) {
+      if (errno!=EAGAIN && errno!=EINTR) break;
+    }
+    else {
+      count += ret;
+    }
+  }
+  mi_prim_close(fd);
+  return (count==buf_len);
+}
+
+#else
+
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  return false;
+}
+
+#endif
+
+
+//----------------------------------------------------------------
+// Thread init/done
+//----------------------------------------------------------------
+
+#if defined(MI_USE_PTHREADS)
+
+// use pthread local storage keys to detect thread ending
+// (and used with MI_TLS_PTHREADS for the default heap)
+pthread_key_t _mi_heap_default_key = (pthread_key_t)(-1);
+
+static void mi_pthread_done(void* value) {
+  if (value!=NULL) {
+    _mi_thread_done((mi_heap_t*)value);
+  }
+}
+
+void _mi_prim_thread_init_auto_done(void) {
+  mi_assert_internal(_mi_heap_default_key == (pthread_key_t)(-1));
+  pthread_key_create(&_mi_heap_default_key, &mi_pthread_done);
+}
+
+void _mi_prim_thread_done_auto_done(void) {
+  // nothing to do
+}
+
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
+  if (_mi_heap_default_key != (pthread_key_t)(-1)) {  // can happen during recursive invocation on freeBSD
+    pthread_setspecific(_mi_heap_default_key, heap);
+  }
+}
+
+#else 
+
+void _mi_prim_thread_init_auto_done(void) {
+  // nothing
+}
+
+void _mi_prim_thread_done_auto_done(void) {
+  // nothing
+}
+
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
+  MI_UNUSED(heap);
+}
+
+#endif

preload-mimalloc/mimalloc/src/prim/wasi/prim.c

@@ -0,0 +1,265 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+// This file is included in `src/prim/prim.c`
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+
+//---------------------------------------------
+// Initialize
+//---------------------------------------------
+
+void _mi_prim_mem_init( mi_os_mem_config_t* config ) {
+  config->page_size = 64*MI_KiB; // WebAssembly has a fixed page size: 64KiB
+  config->alloc_granularity = 16;
+  config->has_overcommit = false;  
+  config->must_free_whole = true;
+}
+
+//---------------------------------------------
+// Free
+//---------------------------------------------
+
+int _mi_prim_free(void* addr, size_t size ) {
+  MI_UNUSED(addr); MI_UNUSED(size);
+  // wasi heap cannot be shrunk
+  return 0;
+}
+
+
+//---------------------------------------------
+// Allocation: sbrk or memory_grow
+//---------------------------------------------
+
+#if defined(MI_USE_SBRK)
+  static void* mi_memory_grow( size_t size ) {
+    void* p = sbrk(size);
+    if (p == (void*)(-1)) return NULL;
+    #if !defined(__wasi__) // on wasi this is always zero initialized already (?)
+    memset(p,0,size);
+    #endif
+    return p;
+  }
+#elif defined(__wasi__)
+  static void* mi_memory_grow( size_t size ) {
+    size_t base = (size > 0 ? __builtin_wasm_memory_grow(0,_mi_divide_up(size, _mi_os_page_size()))
+                            : __builtin_wasm_memory_size(0));
+    if (base == SIZE_MAX) return NULL;
+    return (void*)(base * _mi_os_page_size());
+  }
+#endif
+
+#if defined(MI_USE_PTHREADS)
+static pthread_mutex_t mi_heap_grow_mutex = PTHREAD_MUTEX_INITIALIZER;
+#endif
+
+static void* mi_prim_mem_grow(size_t size, size_t try_alignment) {
+  void* p = NULL;
+  if (try_alignment <= 1) {
+    // `sbrk` is not thread safe in general so try to protect it (we could skip this on WASM but leave it in for now)
+    #if defined(MI_USE_PTHREADS)
+    pthread_mutex_lock(&mi_heap_grow_mutex);
+    #endif
+    p = mi_memory_grow(size);
+    #if defined(MI_USE_PTHREADS)
+    pthread_mutex_unlock(&mi_heap_grow_mutex);
+    #endif
+  }
+  else {
+    void* base = NULL;
+    size_t alloc_size = 0;
+    // to allocate aligned use a lock to try to avoid thread interaction
+    // between getting the current size and actual allocation
+    // (also, `sbrk` is not thread safe in general)
+    #if defined(MI_USE_PTHREADS)
+    pthread_mutex_lock(&mi_heap_grow_mutex);
+    #endif
+    {
+      void* current = mi_memory_grow(0);  // get current size
+      if (current != NULL) {
+        void* aligned_current = mi_align_up_ptr(current, try_alignment);  // and align from there to minimize wasted space
+        alloc_size = _mi_align_up( ((uint8_t*)aligned_current - (uint8_t*)current) + size, _mi_os_page_size());
+        base = mi_memory_grow(alloc_size);
+      }
+    }
+    #if defined(MI_USE_PTHREADS)
+    pthread_mutex_unlock(&mi_heap_grow_mutex);
+    #endif
+    if (base != NULL) {
+      p = mi_align_up_ptr(base, try_alignment);
+      if ((uint8_t*)p + size > (uint8_t*)base + alloc_size) {
+        // another thread used wasm_memory_grow/sbrk in-between and we do not have enough
+        // space after alignment. Give up (and waste the space as we cannot shrink :-( )
+        // (in `mi_os_mem_alloc_aligned` this will fall back to overallocation to align)
+        p = NULL;
+      }
+    }
+  }
+  /*
+  if (p == NULL) {
+    _mi_warning_message("unable to allocate sbrk/wasm_memory_grow OS memory (%zu bytes, %zu alignment)\n", size, try_alignment);
+    errno = ENOMEM;
+    return NULL;
+  }
+  */
+  mi_assert_internal( p == NULL || try_alignment == 0 || (uintptr_t)p % try_alignment == 0 );
+  return p;
+}
+
+// Note: the `try_alignment` is just a hint and the returned pointer is not guaranteed to be aligned.
+int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, void** addr) {
+  MI_UNUSED(allow_large); MI_UNUSED(commit);
+  *is_large = false;
+  *addr = mi_prim_mem_grow(size, try_alignment);
+  return (*addr != NULL ? 0 : ENOMEM);
+}
+
+
+//---------------------------------------------
+// Commit/Reset/Protect
+//---------------------------------------------
+
+int _mi_prim_commit(void* addr, size_t size, bool commit) {
+  MI_UNUSED(addr); MI_UNUSED(size); MI_UNUSED(commit);
+  return 0;
+}
+
+int _mi_prim_reset(void* addr, size_t size) {
+  MI_UNUSED(addr); MI_UNUSED(size);
+  return 0;
+}
+
+int _mi_prim_protect(void* addr, size_t size, bool protect) {
+  MI_UNUSED(addr); MI_UNUSED(size); MI_UNUSED(protect);
+  return 0;
+}
+
+
+//---------------------------------------------
+// Huge pages and NUMA nodes
+//---------------------------------------------
+
+int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, void** addr) {
+  MI_UNUSED(hint_addr); MI_UNUSED(size); MI_UNUSED(numa_node);
+  *addr = NULL;
+  return ENOSYS;
+}
+
+size_t _mi_prim_numa_node(void) {
+  return 0;
+}
+
+size_t _mi_prim_numa_node_count(void) {
+  return 1;
+}
+
+
+//----------------------------------------------------------------
+// Clock
+//----------------------------------------------------------------
+
+#include <time.h>
+
+#if defined(CLOCK_REALTIME) || defined(CLOCK_MONOTONIC)
+
+mi_msecs_t _mi_prim_clock_now(void) {
+  struct timespec t;
+  #ifdef CLOCK_MONOTONIC
+  clock_gettime(CLOCK_MONOTONIC, &t);
+  #else
+  clock_gettime(CLOCK_REALTIME, &t);
+  #endif
+  return ((mi_msecs_t)t.tv_sec * 1000) + ((mi_msecs_t)t.tv_nsec / 1000000);
+}
+
+#else
+
+// low resolution timer
+mi_msecs_t _mi_prim_clock_now(void) {
+  #if !defined(CLOCKS_PER_SEC) || (CLOCKS_PER_SEC == 1000) || (CLOCKS_PER_SEC == 0)
+  return (mi_msecs_t)clock();  
+  #elif (CLOCKS_PER_SEC < 1000)
+  return (mi_msecs_t)clock() * (1000 / (mi_msecs_t)CLOCKS_PER_SEC);  
+  #else
+  return (mi_msecs_t)clock() / ((mi_msecs_t)CLOCKS_PER_SEC / 1000);
+  #endif
+}
+
+#endif
+
+
+//----------------------------------------------------------------
+// Process info
+//----------------------------------------------------------------
+
+void _mi_prim_process_info(mi_process_info_t* pinfo)
+{
+  // use defaults
+  MI_UNUSED(pinfo);
+}
+
+
+//----------------------------------------------------------------
+// Output
+//----------------------------------------------------------------
+
+void _mi_prim_out_stderr( const char* msg ) {
+  fputs(msg,stderr);
+}
+
+
+//----------------------------------------------------------------
+// Environment
+//----------------------------------------------------------------
+
+bool _mi_prim_getenv(const char* name, char* result, size_t result_size) {
+  // cannot call getenv() when still initializing the C runtime.
+  if (_mi_preloading()) return false;
+  const char* s = getenv(name);
+  if (s == NULL) {
+    // we check the upper case name too.
+    char buf[64+1];
+    size_t len = _mi_strnlen(name,sizeof(buf)-1);
+    for (size_t i = 0; i < len; i++) {
+      buf[i] = _mi_toupper(name[i]);
+    }
+    buf[len] = 0;
+    s = getenv(buf);
+  }
+  if (s == NULL || _mi_strnlen(s,result_size) >= result_size)  return false;
+  _mi_strlcpy(result, s, result_size);
+  return true;
+}
+
+
+//----------------------------------------------------------------
+// Random
+//----------------------------------------------------------------
+
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  return false;
+}
+
+
+//----------------------------------------------------------------
+// Thread init/done
+//----------------------------------------------------------------
+
+void _mi_prim_thread_init_auto_done(void) {
+  // nothing
+}
+
+void _mi_prim_thread_done_auto_done(void) {
+  // nothing
+}
+
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
+  MI_UNUSED(heap);
+}

preload-mimalloc/mimalloc/src/prim/windows/etw-mimalloc.wprp

@@ -0,0 +1,61 @@
+<WindowsPerformanceRecorder Version="1.0">
+  <Profiles>
+    <SystemCollector Id="WPR_initiated_WprApp_WPR_System_Collector" Name="WPR_initiated_WprApp_WPR System Collector">
+      <BufferSize Value="1024" />
+      <Buffers Value="100" />
+    </SystemCollector>
+    <EventCollector Id="Mimalloc_Collector" Name="Mimalloc Collector">
+      <BufferSize Value="1024" />
+      <Buffers Value="100" />
+    </EventCollector>
+    <SystemProvider Id="WPR_initiated_WprApp_WPR_System_Collector_Provider">
+      <Keywords>
+        <Keyword Value="Loader" />
+      </Keywords>
+    </SystemProvider>
+    <EventProvider Id="MimallocEventProvider" Name="138f4dbb-ee04-4899-aa0a-572ad4475779" NonPagedMemory="true" Stack="true">
+      <EventFilters FilterIn="true">
+        <EventId Value="100" />
+        <EventId Value="101" />
+      </EventFilters>
+    </EventProvider>
+    <Profile Id="CustomHeap.Verbose.File" Name="CustomHeap" Description="RunningProfile:CustomHeap.Verbose.File" LoggingMode="File" DetailLevel="Verbose">
+      <ProblemCategories>
+        <ProblemCategory Value="Resource Analysis" />
+      </ProblemCategories>
+      <Collectors>
+        <SystemCollectorId Value="WPR_initiated_WprApp_WPR_System_Collector">
+          <SystemProviderId Value="WPR_initiated_WprApp_WPR_System_Collector_Provider" />
+        </SystemCollectorId>
+        <EventCollectorId Value="Mimalloc_Collector">
+          <EventProviders>
+            <EventProviderId Value="MimallocEventProvider" >
+              <Keywords>
+                <Keyword Value="100"/>
+                <Keyword Value="101"/>
+              </Keywords>
+            </EventProviderId>
+          </EventProviders>
+        </EventCollectorId>
+      </Collectors>
+      <TraceMergeProperties>
+        <TraceMergeProperty Id="BaseVerboseTraceMergeProperties" Name="BaseTraceMergeProperties">
+          <DeletePreMergedTraceFiles Value="true" />
+          <FileCompression Value="false" />
+          <InjectOnly Value="false" />
+          <SkipMerge Value="false" />
+          <CustomEvents>
+            <CustomEvent Value="ImageId" />
+            <CustomEvent Value="BuildInfo" />
+            <CustomEvent Value="VolumeMapping" />
+            <CustomEvent Value="EventMetadata" />
+            <CustomEvent Value="PerfTrackMetadata" />
+            <CustomEvent Value="WinSAT" />
+            <CustomEvent Value="NetworkInterface" />
+          </CustomEvents>
+        </TraceMergeProperty>
+      </TraceMergeProperties>
+    </Profile>
+  </Profiles>
+</WindowsPerformanceRecorder>
+

preload-mimalloc/mimalloc/src/prim/windows/etw.h

@@ -0,0 +1,905 @@
+//**********************************************************************`
+//* This is an include file generated by Message Compiler.             *`
+//*                                                                    *`
+//* Copyright (c) Microsoft Corporation. All Rights Reserved.          *`
+//**********************************************************************`
+#pragma once
+
+//*****************************************************************************
+//
+// Notes on the ETW event code generated by MC:
+//
+// - Structures and arrays of structures are treated as an opaque binary blob.
+//   The caller is responsible for packing the data for the structure into a
+//   single region of memory, with no padding between values. The macro will
+//   have an extra parameter for the length of the blob.
+// - Arrays of nul-terminated strings must be packed by the caller into a
+//   single binary blob containing the correct number of strings, with a nul
+//   after each string. The size of the blob is specified in characters, and
+//   includes the final nul.
+// - Arrays of SID are treated as a single binary blob. The caller is
+//   responsible for packing the SID values into a single region of memory with
+//   no padding.
+// - The length attribute on the data element in the manifest is significant
+//   for values with intype win:UnicodeString, win:AnsiString, or win:Binary.
+//   The length attribute must be specified for win:Binary, and is optional for
+//   win:UnicodeString and win:AnsiString (if no length is given, the strings
+//   are assumed to be nul-terminated). For win:UnicodeString, the length is
+//   measured in characters, not bytes.
+// - For an array of win:UnicodeString, win:AnsiString, or win:Binary, the
+//   length attribute applies to every value in the array, so every value in
+//   the array must have the same length. The values in the array are provided
+//   to the macro via a single pointer -- the caller is responsible for packing
+//   all of the values into a single region of memory with no padding between
+//   values.
+// - Values of type win:CountedUnicodeString, win:CountedAnsiString, and
+//   win:CountedBinary can be generated and collected on Vista or later.
+//   However, they may not decode properly without the Windows 10 2018 Fall
+//   Update.
+// - Arrays of type win:CountedUnicodeString, win:CountedAnsiString, and
+//   win:CountedBinary must be packed by the caller into a single region of
+//   memory. The format for each item is a UINT16 byte-count followed by that
+//   many bytes of data. When providing the array to the generated macro, you
+//   must provide the total size of the packed array data, including the UINT16
+//   sizes for each item. In the case of win:CountedUnicodeString, the data
+//   size is specified in WCHAR (16-bit) units. In the case of
+//   win:CountedAnsiString and win:CountedBinary, the data size is specified in
+//   bytes.
+//
+//*****************************************************************************
+
+#include <wmistr.h>
+#include <evntrace.h>
+#include <evntprov.h>
+
+#ifndef ETW_INLINE
+  #ifdef _ETW_KM_
+    // In kernel mode, save stack space by never inlining templates.
+    #define ETW_INLINE DECLSPEC_NOINLINE __inline
+  #else
+    // In user mode, save code size by inlining templates as appropriate.
+    #define ETW_INLINE __inline
+  #endif
+#endif // ETW_INLINE
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+//
+// MCGEN_DISABLE_PROVIDER_CODE_GENERATION macro:
+// Define this macro to have the compiler skip the generated functions in this
+// header.
+//
+#ifndef MCGEN_DISABLE_PROVIDER_CODE_GENERATION
+
+//
+// MCGEN_USE_KERNEL_MODE_APIS macro:
+// Controls whether the generated code uses kernel-mode or user-mode APIs.
+// - Set to 0 to use Windows user-mode APIs such as EventRegister.
+// - Set to 1 to use Windows kernel-mode APIs such as EtwRegister.
+// Default is based on whether the _ETW_KM_ macro is defined (i.e. by wdm.h).
+// Note that the APIs can also be overridden directly, e.g. by setting the
+// MCGEN_EVENTWRITETRANSFER or MCGEN_EVENTREGISTER macros.
+//
+#ifndef MCGEN_USE_KERNEL_MODE_APIS
+  #ifdef _ETW_KM_
+    #define MCGEN_USE_KERNEL_MODE_APIS 1
+  #else
+    #define MCGEN_USE_KERNEL_MODE_APIS 0
+  #endif
+#endif // MCGEN_USE_KERNEL_MODE_APIS
+
+//
+// MCGEN_HAVE_EVENTSETINFORMATION macro:
+// Controls how McGenEventSetInformation uses the EventSetInformation API.
+// - Set to 0 to disable the use of EventSetInformation
+//   (McGenEventSetInformation will always return an error).
+// - Set to 1 to directly invoke MCGEN_EVENTSETINFORMATION.
+// - Set to 2 to to locate EventSetInformation at runtime via GetProcAddress
+//   (user-mode) or MmGetSystemRoutineAddress (kernel-mode).
+// Default is determined as follows:
+// - If MCGEN_EVENTSETINFORMATION has been customized, set to 1
+//   (i.e. use MCGEN_EVENTSETINFORMATION).
+// - Else if the target OS version has EventSetInformation, set to 1
+//   (i.e. use MCGEN_EVENTSETINFORMATION).
+// - Else set to 2 (i.e. try to dynamically locate EventSetInformation).
+// Note that an McGenEventSetInformation function will only be generated if one
+// or more provider in a manifest has provider traits.
+//
+#ifndef MCGEN_HAVE_EVENTSETINFORMATION
+  #ifdef MCGEN_EVENTSETINFORMATION             // if MCGEN_EVENTSETINFORMATION has been customized,
+    #define MCGEN_HAVE_EVENTSETINFORMATION   1 //   directly invoke MCGEN_EVENTSETINFORMATION(...).
+  #elif MCGEN_USE_KERNEL_MODE_APIS             // else if using kernel-mode APIs,
+    #if NTDDI_VERSION >= 0x06040000            //   if target OS is Windows 10 or later,
+      #define MCGEN_HAVE_EVENTSETINFORMATION 1 //     directly invoke MCGEN_EVENTSETINFORMATION(...).
+    #else                                      //   else
+      #define MCGEN_HAVE_EVENTSETINFORMATION 2 //     find "EtwSetInformation" via MmGetSystemRoutineAddress.
+    #endif                                     // else (using user-mode APIs)
+  #else                                        //   if target OS and SDK is Windows 8 or later,
+    #if WINVER >= 0x0602 && defined(EVENT_FILTER_TYPE_SCHEMATIZED)
+      #define MCGEN_HAVE_EVENTSETINFORMATION 1 //     directly invoke MCGEN_EVENTSETINFORMATION(...).
+    #else                                      //   else
+      #define MCGEN_HAVE_EVENTSETINFORMATION 2 //     find "EventSetInformation" via GetModuleHandleExW/GetProcAddress.
+    #endif
+  #endif
+#endif // MCGEN_HAVE_EVENTSETINFORMATION
+
+//
+// MCGEN Override Macros
+//
+// The following override macros may be defined before including this header
+// to control the APIs used by this header:
+//
+// - MCGEN_EVENTREGISTER
+// - MCGEN_EVENTUNREGISTER
+// - MCGEN_EVENTSETINFORMATION
+// - MCGEN_EVENTWRITETRANSFER
+//
+// If the the macro is undefined, the MC implementation will default to the
+// corresponding ETW APIs. For example, if the MCGEN_EVENTREGISTER macro is
+// undefined, the EventRegister[MyProviderName] macro will use EventRegister
+// in user mode and will use EtwRegister in kernel mode.
+//
+// To prevent issues from conflicting definitions of these macros, the value
+// of the override macro will be used as a suffix in certain internal function
+// names. Because of this, the override macros must follow certain rules:
+//
+// - The macro must be defined before any MC-generated header is included and
+//   must not be undefined or redefined after any MC-generated header is
+//   included. Different translation units (i.e. different .c or .cpp files)
+//   may set the macros to different values, but within a translation unit
+//   (within a single .c or .cpp file), the macro must be set once and not
+//   changed.
+// - The override must be an object-like macro, not a function-like macro
+//   (i.e. the override macro must not have a parameter list).
+// - The override macro's value must be a simple identifier, i.e. must be
+//   something that starts with a letter or '_' and contains only letters,
+//   numbers, and '_' characters.
+// - If the override macro's value is the name of a second object-like macro,
+//   the second object-like macro must follow the same rules. (The override
+//   macro's value can also be the name of a function-like macro, in which
+//   case the function-like macro does not need to follow the same rules.)
+//
+// For example, the following will cause compile errors:
+//
+//   #define MCGEN_EVENTWRITETRANSFER MyNamespace::MyClass::MyFunction // Value has non-identifier characters (colon).
+//   #define MCGEN_EVENTWRITETRANSFER GetEventWriteFunctionPointer(7)  // Value has non-identifier characters (parentheses).
+//   #define MCGEN_EVENTWRITETRANSFER(h,e,a,r,c,d) EventWrite(h,e,c,d) // Override is defined as a function-like macro.
+//   #define MY_OBJECT_LIKE_MACRO     MyNamespace::MyClass::MyEventWriteFunction
+//   #define MCGEN_EVENTWRITETRANSFER MY_OBJECT_LIKE_MACRO // Evaluates to something with non-identifier characters (colon).
+//
+// The following would be ok:
+//
+//   #define MCGEN_EVENTWRITETRANSFER  MyEventWriteFunction1  // OK, suffix will be "MyEventWriteFunction1".
+//   #define MY_OBJECT_LIKE_MACRO      MyEventWriteFunction2
+//   #define MCGEN_EVENTWRITETRANSFER  MY_OBJECT_LIKE_MACRO   // OK, suffix will be "MyEventWriteFunction2".
+//   #define MY_FUNCTION_LIKE_MACRO(h,e,a,r,c,d) MyNamespace::MyClass::MyEventWriteFunction3(h,e,c,d)
+//   #define MCGEN_EVENTWRITETRANSFER  MY_FUNCTION_LIKE_MACRO // OK, suffix will be "MY_FUNCTION_LIKE_MACRO".
+//
+#ifndef MCGEN_EVENTREGISTER
+  #if MCGEN_USE_KERNEL_MODE_APIS
+    #define MCGEN_EVENTREGISTER        EtwRegister
+  #else
+    #define MCGEN_EVENTREGISTER        EventRegister
+  #endif
+#endif // MCGEN_EVENTREGISTER
+#ifndef MCGEN_EVENTUNREGISTER
+  #if MCGEN_USE_KERNEL_MODE_APIS
+    #define MCGEN_EVENTUNREGISTER      EtwUnregister
+  #else
+    #define MCGEN_EVENTUNREGISTER      EventUnregister
+  #endif
+#endif // MCGEN_EVENTUNREGISTER
+#ifndef MCGEN_EVENTSETINFORMATION
+  #if MCGEN_USE_KERNEL_MODE_APIS
+    #define MCGEN_EVENTSETINFORMATION  EtwSetInformation
+  #else
+    #define MCGEN_EVENTSETINFORMATION  EventSetInformation
+  #endif
+#endif // MCGEN_EVENTSETINFORMATION
+#ifndef MCGEN_EVENTWRITETRANSFER
+  #if MCGEN_USE_KERNEL_MODE_APIS
+    #define MCGEN_EVENTWRITETRANSFER   EtwWriteTransfer
+  #else
+    #define MCGEN_EVENTWRITETRANSFER   EventWriteTransfer
+  #endif
+#endif // MCGEN_EVENTWRITETRANSFER
+
+//
+// MCGEN_EVENT_ENABLED macro:
+// Override to control how the EventWrite[EventName] macros determine whether
+// an event is enabled. The default behavior is for EventWrite[EventName] to
+// use the EventEnabled[EventName] macros.
+//
+#ifndef MCGEN_EVENT_ENABLED
+#define MCGEN_EVENT_ENABLED(EventName) EventEnabled##EventName()
+#endif
+
+//
+// MCGEN_EVENT_ENABLED_FORCONTEXT macro:
+// Override to control how the EventWrite[EventName]_ForContext macros
+// determine whether an event is enabled. The default behavior is for
+// EventWrite[EventName]_ForContext to use the
+// EventEnabled[EventName]_ForContext macros.
+//
+#ifndef MCGEN_EVENT_ENABLED_FORCONTEXT
+#define MCGEN_EVENT_ENABLED_FORCONTEXT(pContext, EventName) EventEnabled##EventName##_ForContext(pContext)
+#endif
+
+//
+// MCGEN_ENABLE_CHECK macro:
+// Determines whether the specified event would be considered as enabled
+// based on the state of the specified context. Slightly faster than calling
+// McGenEventEnabled directly.
+//
+#ifndef MCGEN_ENABLE_CHECK
+#define MCGEN_ENABLE_CHECK(Context, Descriptor) (Context.IsEnabled && McGenEventEnabled(&Context, &Descriptor))
+#endif
+
+#if !defined(MCGEN_TRACE_CONTEXT_DEF)
+#define MCGEN_TRACE_CONTEXT_DEF
+// This structure is for use by MC-generated code and should not be used directly.
+typedef struct _MCGEN_TRACE_CONTEXT
+{
+    TRACEHANDLE            RegistrationHandle;
+    TRACEHANDLE            Logger;      // Used as pointer to provider traits.
+    ULONGLONG              MatchAnyKeyword;
+    ULONGLONG              MatchAllKeyword;
+    ULONG                  Flags;
+    ULONG                  IsEnabled;
+    UCHAR                  Level;
+    UCHAR                  Reserve;
+    USHORT                 EnableBitsCount;
+    PULONG                 EnableBitMask;
+    const ULONGLONG*       EnableKeyWords;
+    const UCHAR*           EnableLevel;
+} MCGEN_TRACE_CONTEXT, *PMCGEN_TRACE_CONTEXT;
+#endif // MCGEN_TRACE_CONTEXT_DEF
+
+#if !defined(MCGEN_LEVEL_KEYWORD_ENABLED_DEF)
+#define MCGEN_LEVEL_KEYWORD_ENABLED_DEF
+//
+// Determines whether an event with a given Level and Keyword would be
+// considered as enabled based on the state of the specified context.
+// Note that you may want to use MCGEN_ENABLE_CHECK instead of calling this
+// function directly.
+//
+FORCEINLINE
+BOOLEAN
+McGenLevelKeywordEnabled(
+    _In_ PMCGEN_TRACE_CONTEXT EnableInfo,
+    _In_ UCHAR Level,
+    _In_ ULONGLONG Keyword
+    )
+{
+    //
+    // Check if the event Level is lower than the level at which
+    // the channel is enabled.
+    // If the event Level is 0 or the channel is enabled at level 0,
+    // all levels are enabled.
+    //
+
+    if ((Level <= EnableInfo->Level) || // This also covers the case of Level == 0.
+        (EnableInfo->Level == 0)) {
+
+        //
+        // Check if Keyword is enabled
+        //
+
+        if ((Keyword == (ULONGLONG)0) ||
+            ((Keyword & EnableInfo->MatchAnyKeyword) &&
+             ((Keyword & EnableInfo->MatchAllKeyword) == EnableInfo->MatchAllKeyword))) {
+            return TRUE;
+        }
+    }
+
+    return FALSE;
+}
+#endif // MCGEN_LEVEL_KEYWORD_ENABLED_DEF
+
+#if !defined(MCGEN_EVENT_ENABLED_DEF)
+#define MCGEN_EVENT_ENABLED_DEF
+//
+// Determines whether the specified event would be considered as enabled based
+// on the state of the specified context. Note that you may want to use
+// MCGEN_ENABLE_CHECK instead of calling this function directly.
+//
+FORCEINLINE
+BOOLEAN
+McGenEventEnabled(
+    _In_ PMCGEN_TRACE_CONTEXT EnableInfo,
+    _In_ PCEVENT_DESCRIPTOR EventDescriptor
+    )
+{
+    return McGenLevelKeywordEnabled(EnableInfo, EventDescriptor->Level, EventDescriptor->Keyword);
+}
+#endif // MCGEN_EVENT_ENABLED_DEF
+
+#if !defined(MCGEN_CONTROL_CALLBACK)
+#define MCGEN_CONTROL_CALLBACK
+
+// This function is for use by MC-generated code and should not be used directly.
+DECLSPEC_NOINLINE __inline
+VOID
+__stdcall
+McGenControlCallbackV2(
+    _In_ LPCGUID SourceId,
+    _In_ ULONG ControlCode,
+    _In_ UCHAR Level,
+    _In_ ULONGLONG MatchAnyKeyword,
+    _In_ ULONGLONG MatchAllKeyword,
+    _In_opt_ PEVENT_FILTER_DESCRIPTOR FilterData,
+    _Inout_opt_ PVOID CallbackContext
+    )
+/*++
+
+Routine Description:
+
+    This is the notification callback for Windows Vista and later.
+
+Arguments:
+
+    SourceId - The GUID that identifies the session that enabled the provider.
+
+    ControlCode - The parameter indicates whether the provider
+                  is being enabled or disabled.
+
+    Level - The level at which the event is enabled.
+
+    MatchAnyKeyword - The bitmask of keywords that the provider uses to
+                      determine the category of events that it writes.
+
+    MatchAllKeyword - This bitmask additionally restricts the category
+                      of events that the provider writes.
+
+    FilterData - The provider-defined data.
+
+    CallbackContext - The context of the callback that is defined when the provider
+                      called EtwRegister to register itself.
+
+Remarks:
+
+    ETW calls this function to notify provider of enable/disable
+
+--*/
+{
+    PMCGEN_TRACE_CONTEXT Ctx = (PMCGEN_TRACE_CONTEXT)CallbackContext;
+    ULONG Ix;
+#ifndef MCGEN_PRIVATE_ENABLE_CALLBACK_V2
+    UNREFERENCED_PARAMETER(SourceId);
+    UNREFERENCED_PARAMETER(FilterData);
+#endif
+
+    if (Ctx == NULL) {
+        return;
+    }
+
+    switch (ControlCode) {
+
+        case EVENT_CONTROL_CODE_ENABLE_PROVIDER:
+            Ctx->Level = Level;
+            Ctx->MatchAnyKeyword = MatchAnyKeyword;
+            Ctx->MatchAllKeyword = MatchAllKeyword;
+            Ctx->IsEnabled = EVENT_CONTROL_CODE_ENABLE_PROVIDER;
+
+            for (Ix = 0; Ix < Ctx->EnableBitsCount; Ix += 1) {
+                if (McGenLevelKeywordEnabled(Ctx, Ctx->EnableLevel[Ix], Ctx->EnableKeyWords[Ix]) != FALSE) {
+                    Ctx->EnableBitMask[Ix >> 5] |= (1 << (Ix % 32));
+                } else {
+                    Ctx->EnableBitMask[Ix >> 5] &= ~(1 << (Ix % 32));
+                }
+            }
+            break;
+
+        case EVENT_CONTROL_CODE_DISABLE_PROVIDER:
+            Ctx->IsEnabled = EVENT_CONTROL_CODE_DISABLE_PROVIDER;
+            Ctx->Level = 0;
+            Ctx->MatchAnyKeyword = 0;
+            Ctx->MatchAllKeyword = 0;
+            if (Ctx->EnableBitsCount > 0) {
+#pragma warning(suppress: 26451) // Arithmetic overflow cannot occur, no matter the value of EnableBitCount
+                RtlZeroMemory(Ctx->EnableBitMask, (((Ctx->EnableBitsCount - 1) / 32) + 1) * sizeof(ULONG));
+            }
+            break;
+
+        default:
+            break;
+    }
+
+#ifdef MCGEN_PRIVATE_ENABLE_CALLBACK_V2
+    //
+    // Call user defined callback
+    //
+    MCGEN_PRIVATE_ENABLE_CALLBACK_V2(
+        SourceId,
+        ControlCode,
+        Level,
+        MatchAnyKeyword,
+        MatchAllKeyword,
+        FilterData,
+        CallbackContext
+        );
+#endif // MCGEN_PRIVATE_ENABLE_CALLBACK_V2
+
+    return;
+}
+
+#endif // MCGEN_CONTROL_CALLBACK
+
+#ifndef _mcgen_PENABLECALLBACK
+  #if MCGEN_USE_KERNEL_MODE_APIS
+    #define _mcgen_PENABLECALLBACK      PETWENABLECALLBACK
+  #else
+    #define _mcgen_PENABLECALLBACK      PENABLECALLBACK
+  #endif
+#endif // _mcgen_PENABLECALLBACK
+
+#if !defined(_mcgen_PASTE2)
+// This macro is for use by MC-generated code and should not be used directly.
+#define _mcgen_PASTE2(a, b) _mcgen_PASTE2_imp(a, b)
+#define _mcgen_PASTE2_imp(a, b) a##b
+#endif // _mcgen_PASTE2
+
+#if !defined(_mcgen_PASTE3)
+// This macro is for use by MC-generated code and should not be used directly.
+#define _mcgen_PASTE3(a, b, c) _mcgen_PASTE3_imp(a, b, c)
+#define _mcgen_PASTE3_imp(a, b, c) a##b##_##c
+#endif // _mcgen_PASTE3
+
+//
+// Macro validation
+//
+
+// Validate MCGEN_EVENTREGISTER:
+
+// Trigger an error if MCGEN_EVENTREGISTER is not an unqualified (simple) identifier:
+struct _mcgen_PASTE2(MCGEN_EVENTREGISTER_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTREGISTER);
+
+// Trigger an error if MCGEN_EVENTREGISTER is redefined:
+typedef struct _mcgen_PASTE2(MCGEN_EVENTREGISTER_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTREGISTER)
+    MCGEN_EVENTREGISTER_must_not_be_redefined_between_headers;
+
+// Trigger an error if MCGEN_EVENTREGISTER is defined as a function-like macro:
+typedef void MCGEN_EVENTREGISTER_must_not_be_a_functionLike_macro_MCGEN_EVENTREGISTER;
+typedef int _mcgen_PASTE2(MCGEN_EVENTREGISTER_must_not_be_a_functionLike_macro_, MCGEN_EVENTREGISTER);
+
+// Validate MCGEN_EVENTUNREGISTER:
+
+// Trigger an error if MCGEN_EVENTUNREGISTER is not an unqualified (simple) identifier:
+struct _mcgen_PASTE2(MCGEN_EVENTUNREGISTER_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTUNREGISTER);
+
+// Trigger an error if MCGEN_EVENTUNREGISTER is redefined:
+typedef struct _mcgen_PASTE2(MCGEN_EVENTUNREGISTER_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTUNREGISTER)
+    MCGEN_EVENTUNREGISTER_must_not_be_redefined_between_headers;
+
+// Trigger an error if MCGEN_EVENTUNREGISTER is defined as a function-like macro:
+typedef void MCGEN_EVENTUNREGISTER_must_not_be_a_functionLike_macro_MCGEN_EVENTUNREGISTER;
+typedef int _mcgen_PASTE2(MCGEN_EVENTUNREGISTER_must_not_be_a_functionLike_macro_, MCGEN_EVENTUNREGISTER);
+
+// Validate MCGEN_EVENTSETINFORMATION:
+
+// Trigger an error if MCGEN_EVENTSETINFORMATION is not an unqualified (simple) identifier:
+struct _mcgen_PASTE2(MCGEN_EVENTSETINFORMATION_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTSETINFORMATION);
+
+// Trigger an error if MCGEN_EVENTSETINFORMATION is redefined:
+typedef struct _mcgen_PASTE2(MCGEN_EVENTSETINFORMATION_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTSETINFORMATION)
+    MCGEN_EVENTSETINFORMATION_must_not_be_redefined_between_headers;
+
+// Trigger an error if MCGEN_EVENTSETINFORMATION is defined as a function-like macro:
+typedef void MCGEN_EVENTSETINFORMATION_must_not_be_a_functionLike_macro_MCGEN_EVENTSETINFORMATION;
+typedef int _mcgen_PASTE2(MCGEN_EVENTSETINFORMATION_must_not_be_a_functionLike_macro_, MCGEN_EVENTSETINFORMATION);
+
+// Validate MCGEN_EVENTWRITETRANSFER:
+
+// Trigger an error if MCGEN_EVENTWRITETRANSFER is not an unqualified (simple) identifier:
+struct _mcgen_PASTE2(MCGEN_EVENTWRITETRANSFER_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTWRITETRANSFER);
+
+// Trigger an error if MCGEN_EVENTWRITETRANSFER is redefined:
+typedef struct _mcgen_PASTE2(MCGEN_EVENTWRITETRANSFER_definition_must_be_an_unqualified_identifier_, MCGEN_EVENTWRITETRANSFER)
+    MCGEN_EVENTWRITETRANSFER_must_not_be_redefined_between_headers;;
+
+// Trigger an error if MCGEN_EVENTWRITETRANSFER is defined as a function-like macro:
+typedef void MCGEN_EVENTWRITETRANSFER_must_not_be_a_functionLike_macro_MCGEN_EVENTWRITETRANSFER;
+typedef int _mcgen_PASTE2(MCGEN_EVENTWRITETRANSFER_must_not_be_a_functionLike_macro_, MCGEN_EVENTWRITETRANSFER);
+
+#ifndef McGenEventWrite_def
+#define McGenEventWrite_def
+
+// This macro is for use by MC-generated code and should not be used directly.
+#define McGenEventWrite _mcgen_PASTE2(McGenEventWrite_, MCGEN_EVENTWRITETRANSFER)
+
+// This function is for use by MC-generated code and should not be used directly.
+DECLSPEC_NOINLINE __inline
+ULONG __stdcall
+McGenEventWrite(
+    _In_ PMCGEN_TRACE_CONTEXT Context,
+    _In_ PCEVENT_DESCRIPTOR Descriptor,
+    _In_opt_ LPCGUID ActivityId,
+    _In_range_(1, 128) ULONG EventDataCount,
+    _Pre_cap_(EventDataCount) EVENT_DATA_DESCRIPTOR* EventData
+    )
+{
+    const USHORT UNALIGNED* Traits;
+
+    // Some customized MCGEN_EVENTWRITETRANSFER macros might ignore ActivityId.
+    UNREFERENCED_PARAMETER(ActivityId);
+
+    Traits = (const USHORT UNALIGNED*)(UINT_PTR)Context->Logger;
+
+    if (Traits == NULL) {
+        EventData[0].Ptr = 0;
+        EventData[0].Size = 0;
+        EventData[0].Reserved = 0;
+    } else {
+        EventData[0].Ptr = (ULONG_PTR)Traits;
+        EventData[0].Size = *Traits;
+        EventData[0].Reserved = 2; // EVENT_DATA_DESCRIPTOR_TYPE_PROVIDER_METADATA
+    }
+
+    return MCGEN_EVENTWRITETRANSFER(
+        Context->RegistrationHandle,
+        Descriptor,
+        ActivityId,
+        NULL,
+        EventDataCount,
+        EventData);
+}
+#endif // McGenEventWrite_def
+
+#if !defined(McGenEventRegisterUnregister)
+#define McGenEventRegisterUnregister
+
+// This macro is for use by MC-generated code and should not be used directly.
+#define McGenEventRegister _mcgen_PASTE2(McGenEventRegister_, MCGEN_EVENTREGISTER)
+
+#pragma warning(push)
+#pragma warning(disable:6103)
+// This function is for use by MC-generated code and should not be used directly.
+DECLSPEC_NOINLINE __inline
+ULONG __stdcall
+McGenEventRegister(
+    _In_ LPCGUID ProviderId,
+    _In_opt_ _mcgen_PENABLECALLBACK EnableCallback,
+    _In_opt_ PVOID CallbackContext,
+    _Inout_ PREGHANDLE RegHandle
+    )
+/*++
+
+Routine Description:
+
+    This function registers the provider with ETW.
+
+Arguments:
+
+    ProviderId - Provider ID to register with ETW.
+
+    EnableCallback - Callback to be used.
+
+    CallbackContext - Context for the callback.
+
+    RegHandle - Pointer to registration handle.
+
+Remarks:
+
+    Should not be called if the provider is already registered (i.e. should not
+    be called if *RegHandle != 0). Repeatedly registering a provider is a bug
+    and may indicate a race condition. However, for compatibility with previous
+    behavior, this function will return SUCCESS in this case.
+
+--*/
+{
+    ULONG Error;
+
+    if (*RegHandle != 0)
+    {
+        Error = 0; // ERROR_SUCCESS
+    }
+    else
+    {
+        Error = MCGEN_EVENTREGISTER(ProviderId, EnableCallback, CallbackContext, RegHandle);
+    }
+
+    return Error;
+}
+#pragma warning(pop)
+
+// This macro is for use by MC-generated code and should not be used directly.
+#define McGenEventUnregister _mcgen_PASTE2(McGenEventUnregister_, MCGEN_EVENTUNREGISTER)
+
+// This function is for use by MC-generated code and should not be used directly.
+DECLSPEC_NOINLINE __inline
+ULONG __stdcall
+McGenEventUnregister(_Inout_ PREGHANDLE RegHandle)
+/*++
+
+Routine Description:
+
+    Unregister from ETW and set *RegHandle = 0.
+
+Arguments:
+
+    RegHandle - the pointer to the provider registration handle
+
+Remarks:
+
+    If provider has not been registered (i.e. if *RegHandle == 0),
+    return SUCCESS. It is safe to call McGenEventUnregister even if the
+    call to McGenEventRegister returned an error.
+
+--*/
+{
+    ULONG Error;
+
+    if(*RegHandle == 0)
+    {
+        Error = 0; // ERROR_SUCCESS
+    }
+    else
+    {
+        Error = MCGEN_EVENTUNREGISTER(*RegHandle);
+        *RegHandle = (REGHANDLE)0;
+    }
+
+    return Error;
+}
+
+#endif // McGenEventRegisterUnregister
+
+#ifndef _mcgen_EVENT_BIT_SET
+  #if defined(_M_IX86) || defined(_M_X64)
+    // This macro is for use by MC-generated code and should not be used directly.
+    #define _mcgen_EVENT_BIT_SET(EnableBits, BitPosition) ((((const unsigned char*)EnableBits)[BitPosition >> 3] & (1u << (BitPosition & 7))) != 0)
+  #else // CPU type
+    // This macro is for use by MC-generated code and should not be used directly.
+    #define _mcgen_EVENT_BIT_SET(EnableBits, BitPosition) ((EnableBits[BitPosition >> 5] & (1u << (BitPosition & 31))) != 0)
+  #endif // CPU type
+#endif // _mcgen_EVENT_BIT_SET
+
+#endif // MCGEN_DISABLE_PROVIDER_CODE_GENERATION
+
+//+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+// Provider "microsoft-windows-mimalloc" event count 2
+//+++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++
+
+// Provider GUID = 138f4dbb-ee04-4899-aa0a-572ad4475779
+EXTERN_C __declspec(selectany) const GUID ETW_MI_Provider = {0x138f4dbb, 0xee04, 0x4899, {0xaa, 0x0a, 0x57, 0x2a, 0xd4, 0x47, 0x57, 0x79}};
+
+#ifndef ETW_MI_Provider_Traits
+#define ETW_MI_Provider_Traits NULL
+#endif // ETW_MI_Provider_Traits
+
+//
+// Event Descriptors
+//
+EXTERN_C __declspec(selectany) const EVENT_DESCRIPTOR ETW_MI_ALLOC = {0x64, 0x1, 0x0, 0x4, 0x0, 0x0, 0x0};
+#define ETW_MI_ALLOC_value 0x64
+EXTERN_C __declspec(selectany) const EVENT_DESCRIPTOR ETW_MI_FREE = {0x65, 0x1, 0x0, 0x4, 0x0, 0x0, 0x0};
+#define ETW_MI_FREE_value 0x65
+
+//
+// MCGEN_DISABLE_PROVIDER_CODE_GENERATION macro:
+// Define this macro to have the compiler skip the generated functions in this
+// header.
+//
+#ifndef MCGEN_DISABLE_PROVIDER_CODE_GENERATION
+
+//
+// Event Enablement Bits
+// These variables are for use by MC-generated code and should not be used directly.
+//
+EXTERN_C __declspec(selectany) DECLSPEC_CACHEALIGN ULONG microsoft_windows_mimallocEnableBits[1];
+EXTERN_C __declspec(selectany) const ULONGLONG microsoft_windows_mimallocKeywords[1] = {0x0};
+EXTERN_C __declspec(selectany) const unsigned char microsoft_windows_mimallocLevels[1] = {4};
+
+//
+// Provider context
+//
+EXTERN_C __declspec(selectany) MCGEN_TRACE_CONTEXT ETW_MI_Provider_Context = {0, (ULONG_PTR)ETW_MI_Provider_Traits, 0, 0, 0, 0, 0, 0, 1, microsoft_windows_mimallocEnableBits, microsoft_windows_mimallocKeywords, microsoft_windows_mimallocLevels};
+
+//
+// Provider REGHANDLE
+//
+#define microsoft_windows_mimallocHandle (ETW_MI_Provider_Context.RegistrationHandle)
+
+//
+// This macro is set to 0, indicating that the EventWrite[Name] macros do not
+// have an Activity parameter. This is controlled by the -km and -um options.
+//
+#define ETW_MI_Provider_EventWriteActivity 0
+
+//
+// Register with ETW using the control GUID specified in the manifest.
+// Invoke this macro during module initialization (i.e. program startup,
+// DLL process attach, or driver load) to initialize the provider.
+// Note that if this function returns an error, the error means that
+// will not work, but no action needs to be taken -- even if EventRegister
+// returns an error, it is generally safe to use EventWrite and
+// EventUnregister macros (they will be no-ops if EventRegister failed).
+//
+#ifndef EventRegistermicrosoft_windows_mimalloc
+#define EventRegistermicrosoft_windows_mimalloc() McGenEventRegister(&ETW_MI_Provider, McGenControlCallbackV2, &ETW_MI_Provider_Context, &microsoft_windows_mimallocHandle)
+#endif
+
+//
+// Register with ETW using a specific control GUID (i.e. a GUID other than what
+// is specified in the manifest). Advanced scenarios only.
+//
+#ifndef EventRegisterByGuidmicrosoft_windows_mimalloc
+#define EventRegisterByGuidmicrosoft_windows_mimalloc(Guid) McGenEventRegister(&(Guid), McGenControlCallbackV2, &ETW_MI_Provider_Context, &microsoft_windows_mimallocHandle)
+#endif
+
+//
+// Unregister with ETW and close the provider.
+// Invoke this macro during module shutdown (i.e. program exit, DLL process
+// detach, or driver unload) to unregister the provider.
+// Note that you MUST call EventUnregister before DLL or driver unload
+// (not optional): failure to unregister a provider before DLL or driver unload
+// will result in crashes.
+//
+#ifndef EventUnregistermicrosoft_windows_mimalloc
+#define EventUnregistermicrosoft_windows_mimalloc() McGenEventUnregister(&microsoft_windows_mimallocHandle)
+#endif
+
+//
+// MCGEN_ENABLE_FORCONTEXT_CODE_GENERATION macro:
+// Define this macro to enable support for caller-allocated provider context.
+//
+#ifdef MCGEN_ENABLE_FORCONTEXT_CODE_GENERATION
+
+//
+// Advanced scenarios: Caller-allocated provider context.
+// Use when multiple differently-configured provider handles are needed,
+// e.g. for container-aware drivers, one context per container.
+//
+// Usage:
+//
+// - Caller enables the feature before including this header, e.g.
+//   #define MCGEN_ENABLE_FORCONTEXT_CODE_GENERATION 1
+// - Caller allocates memory, e.g. pContext = malloc(sizeof(McGenContext_microsoft_windows_mimalloc));
+// - Caller registers the provider, e.g. EventRegistermicrosoft_windows_mimalloc_ForContext(pContext);
+// - Caller writes events, e.g. EventWriteMyEvent_ForContext(pContext, ...);
+// - Caller unregisters, e.g. EventUnregistermicrosoft_windows_mimalloc_ForContext(pContext);
+// - Caller frees memory, e.g. free(pContext);
+//
+
+typedef struct tagMcGenContext_microsoft_windows_mimalloc {
+    // The fields of this structure are subject to change and should
+    // not be accessed directly. To access the provider's REGHANDLE,
+    // use microsoft_windows_mimallocHandle_ForContext(pContext).
+    MCGEN_TRACE_CONTEXT Context;
+    ULONG EnableBits[1];
+} McGenContext_microsoft_windows_mimalloc;
+
+#define EventRegistermicrosoft_windows_mimalloc_ForContext(pContext)             _mcgen_PASTE2(_mcgen_RegisterForContext_microsoft_windows_mimalloc_, MCGEN_EVENTREGISTER)(&ETW_MI_Provider, pContext)
+#define EventRegisterByGuidmicrosoft_windows_mimalloc_ForContext(Guid, pContext) _mcgen_PASTE2(_mcgen_RegisterForContext_microsoft_windows_mimalloc_, MCGEN_EVENTREGISTER)(&(Guid), pContext)
+#define EventUnregistermicrosoft_windows_mimalloc_ForContext(pContext)           McGenEventUnregister(&(pContext)->Context.RegistrationHandle)
+
+//
+// Provider REGHANDLE for caller-allocated context.
+//
+#define microsoft_windows_mimallocHandle_ForContext(pContext) ((pContext)->Context.RegistrationHandle)
+
+// This function is for use by MC-generated code and should not be used directly.
+// Initialize and register the caller-allocated context.
+__inline
+ULONG __stdcall
+_mcgen_PASTE2(_mcgen_RegisterForContext_microsoft_windows_mimalloc_, MCGEN_EVENTREGISTER)(
+    _In_ LPCGUID pProviderId,
+    _Out_ McGenContext_microsoft_windows_mimalloc* pContext)
+{
+    RtlZeroMemory(pContext, sizeof(*pContext));
+    pContext->Context.Logger = (ULONG_PTR)ETW_MI_Provider_Traits;
+    pContext->Context.EnableBitsCount = 1;
+    pContext->Context.EnableBitMask = pContext->EnableBits;
+    pContext->Context.EnableKeyWords = microsoft_windows_mimallocKeywords;
+    pContext->Context.EnableLevel = microsoft_windows_mimallocLevels;
+    return McGenEventRegister(
+        pProviderId,
+        McGenControlCallbackV2,
+        &pContext->Context,
+        &pContext->Context.RegistrationHandle);
+}
+
+// This function is for use by MC-generated code and should not be used directly.
+// Trigger a compile error if called with the wrong parameter type.
+FORCEINLINE
+_Ret_ McGenContext_microsoft_windows_mimalloc*
+_mcgen_CheckContextType_microsoft_windows_mimalloc(_In_ McGenContext_microsoft_windows_mimalloc* pContext)
+{
+    return pContext;
+}
+
+#endif // MCGEN_ENABLE_FORCONTEXT_CODE_GENERATION
+
+//
+// Enablement check macro for event "ETW_MI_ALLOC"
+//
+#define EventEnabledETW_MI_ALLOC() _mcgen_EVENT_BIT_SET(microsoft_windows_mimallocEnableBits, 0)
+#define EventEnabledETW_MI_ALLOC_ForContext(pContext) _mcgen_EVENT_BIT_SET(_mcgen_CheckContextType_microsoft_windows_mimalloc(pContext)->EnableBits, 0)
+
+//
+// Event write macros for event "ETW_MI_ALLOC"
+//
+#define EventWriteETW_MI_ALLOC(Address, Size) \
+        MCGEN_EVENT_ENABLED(ETW_MI_ALLOC) \
+        ? _mcgen_TEMPLATE_FOR_ETW_MI_ALLOC(&ETW_MI_Provider_Context, &ETW_MI_ALLOC, Address, Size) : 0
+#define EventWriteETW_MI_ALLOC_AssumeEnabled(Address, Size) \
+        _mcgen_TEMPLATE_FOR_ETW_MI_ALLOC(&ETW_MI_Provider_Context, &ETW_MI_ALLOC, Address, Size)
+#define EventWriteETW_MI_ALLOC_ForContext(pContext, Address, Size) \
+        MCGEN_EVENT_ENABLED_FORCONTEXT(pContext, ETW_MI_ALLOC) \
+        ? _mcgen_TEMPLATE_FOR_ETW_MI_ALLOC(&(pContext)->Context, &ETW_MI_ALLOC, Address, Size) : 0
+#define EventWriteETW_MI_ALLOC_ForContextAssumeEnabled(pContext, Address, Size) \
+        _mcgen_TEMPLATE_FOR_ETW_MI_ALLOC(&_mcgen_CheckContextType_microsoft_windows_mimalloc(pContext)->Context, &ETW_MI_ALLOC, Address, Size)
+
+// This macro is for use by MC-generated code and should not be used directly.
+#define _mcgen_TEMPLATE_FOR_ETW_MI_ALLOC _mcgen_PASTE2(McTemplateU0xx_, MCGEN_EVENTWRITETRANSFER)
+
+//
+// Enablement check macro for event "ETW_MI_FREE"
+//
+#define EventEnabledETW_MI_FREE() _mcgen_EVENT_BIT_SET(microsoft_windows_mimallocEnableBits, 0)
+#define EventEnabledETW_MI_FREE_ForContext(pContext) _mcgen_EVENT_BIT_SET(_mcgen_CheckContextType_microsoft_windows_mimalloc(pContext)->EnableBits, 0)
+
+//
+// Event write macros for event "ETW_MI_FREE"
+//
+#define EventWriteETW_MI_FREE(Address, Size) \
+        MCGEN_EVENT_ENABLED(ETW_MI_FREE) \
+        ? _mcgen_TEMPLATE_FOR_ETW_MI_FREE(&ETW_MI_Provider_Context, &ETW_MI_FREE, Address, Size) : 0
+#define EventWriteETW_MI_FREE_AssumeEnabled(Address, Size) \
+        _mcgen_TEMPLATE_FOR_ETW_MI_FREE(&ETW_MI_Provider_Context, &ETW_MI_FREE, Address, Size)
+#define EventWriteETW_MI_FREE_ForContext(pContext, Address, Size) \
+        MCGEN_EVENT_ENABLED_FORCONTEXT(pContext, ETW_MI_FREE) \
+        ? _mcgen_TEMPLATE_FOR_ETW_MI_FREE(&(pContext)->Context, &ETW_MI_FREE, Address, Size) : 0
+#define EventWriteETW_MI_FREE_ForContextAssumeEnabled(pContext, Address, Size) \
+        _mcgen_TEMPLATE_FOR_ETW_MI_FREE(&_mcgen_CheckContextType_microsoft_windows_mimalloc(pContext)->Context, &ETW_MI_FREE, Address, Size)
+
+// This macro is for use by MC-generated code and should not be used directly.
+#define _mcgen_TEMPLATE_FOR_ETW_MI_FREE _mcgen_PASTE2(McTemplateU0xx_, MCGEN_EVENTWRITETRANSFER)
+
+#endif // MCGEN_DISABLE_PROVIDER_CODE_GENERATION
+
+//
+// MCGEN_DISABLE_PROVIDER_CODE_GENERATION macro:
+// Define this macro to have the compiler skip the generated functions in this
+// header.
+//
+#ifndef MCGEN_DISABLE_PROVIDER_CODE_GENERATION
+
+//
+// Template Functions
+//
+
+//
+// Function for template "ETW_CUSTOM_HEAP_ALLOC_DATA" (and possibly others).
+// This function is for use by MC-generated code and should not be used directly.
+//
+#ifndef McTemplateU0xx_def
+#define McTemplateU0xx_def
+ETW_INLINE
+ULONG
+_mcgen_PASTE2(McTemplateU0xx_, MCGEN_EVENTWRITETRANSFER)(
+    _In_ PMCGEN_TRACE_CONTEXT Context,
+    _In_ PCEVENT_DESCRIPTOR Descriptor,
+    _In_ const unsigned __int64  _Arg0,
+    _In_ const unsigned __int64  _Arg1
+    )
+{
+#define McTemplateU0xx_ARGCOUNT 2
+
+    EVENT_DATA_DESCRIPTOR EventData[McTemplateU0xx_ARGCOUNT + 1];
+
+    EventDataDescCreate(&EventData[1],&_Arg0, sizeof(const unsigned __int64)  );
+
+    EventDataDescCreate(&EventData[2],&_Arg1, sizeof(const unsigned __int64)  );
+
+    return McGenEventWrite(Context, Descriptor, NULL, McTemplateU0xx_ARGCOUNT + 1, EventData);
+}
+#endif // McTemplateU0xx_def
+
+#endif // MCGEN_DISABLE_PROVIDER_CODE_GENERATION
+
+#if defined(__cplusplus)
+}
+#endif

preload-mimalloc/mimalloc/src/prim/windows/prim.c

@@ -0,0 +1,607 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2023, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+// This file is included in `src/prim/prim.c`
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+#include <stdio.h>   // fputs, stderr
+
+
+//---------------------------------------------
+// Dynamically bind Windows API points for portability
+//---------------------------------------------
+
+// We use VirtualAlloc2 for aligned allocation, but it is only supported on Windows 10 and Windows Server 2016.
+// So, we need to look it up dynamically to run on older systems. (use __stdcall for 32-bit compatibility)
+// NtAllocateVirtualAllocEx is used for huge OS page allocation (1GiB)
+// We define a minimal MEM_EXTENDED_PARAMETER ourselves in order to be able to compile with older SDK's.
+typedef enum MI_MEM_EXTENDED_PARAMETER_TYPE_E {
+  MiMemExtendedParameterInvalidType = 0,
+  MiMemExtendedParameterAddressRequirements,
+  MiMemExtendedParameterNumaNode,
+  MiMemExtendedParameterPartitionHandle,
+  MiMemExtendedParameterUserPhysicalHandle,
+  MiMemExtendedParameterAttributeFlags,
+  MiMemExtendedParameterMax
+} MI_MEM_EXTENDED_PARAMETER_TYPE;
+
+typedef struct DECLSPEC_ALIGN(8) MI_MEM_EXTENDED_PARAMETER_S {
+  struct { DWORD64 Type : 8; DWORD64 Reserved : 56; } Type;
+  union  { DWORD64 ULong64; PVOID Pointer; SIZE_T Size; HANDLE Handle; DWORD ULong; } Arg;
+} MI_MEM_EXTENDED_PARAMETER;
+
+typedef struct MI_MEM_ADDRESS_REQUIREMENTS_S {
+  PVOID  LowestStartingAddress;
+  PVOID  HighestEndingAddress;
+  SIZE_T Alignment;
+} MI_MEM_ADDRESS_REQUIREMENTS;
+
+#define MI_MEM_EXTENDED_PARAMETER_NONPAGED_HUGE   0x00000010
+
+#include <winternl.h>
+typedef PVOID    (__stdcall *PVirtualAlloc2)(HANDLE, PVOID, SIZE_T, ULONG, ULONG, MI_MEM_EXTENDED_PARAMETER*, ULONG);
+typedef NTSTATUS (__stdcall *PNtAllocateVirtualMemoryEx)(HANDLE, PVOID*, SIZE_T*, ULONG, ULONG, MI_MEM_EXTENDED_PARAMETER*, ULONG);
+static PVirtualAlloc2 pVirtualAlloc2 = NULL;
+static PNtAllocateVirtualMemoryEx pNtAllocateVirtualMemoryEx = NULL;
+
+// Similarly, GetNumaProcesorNodeEx is only supported since Windows 7
+typedef struct MI_PROCESSOR_NUMBER_S { WORD Group; BYTE Number; BYTE Reserved; } MI_PROCESSOR_NUMBER;
+
+typedef VOID (__stdcall *PGetCurrentProcessorNumberEx)(MI_PROCESSOR_NUMBER* ProcNumber);
+typedef BOOL (__stdcall *PGetNumaProcessorNodeEx)(MI_PROCESSOR_NUMBER* Processor, PUSHORT NodeNumber);
+typedef BOOL (__stdcall* PGetNumaNodeProcessorMaskEx)(USHORT Node, PGROUP_AFFINITY ProcessorMask);
+typedef BOOL (__stdcall *PGetNumaProcessorNode)(UCHAR Processor, PUCHAR NodeNumber);
+static PGetCurrentProcessorNumberEx pGetCurrentProcessorNumberEx = NULL;
+static PGetNumaProcessorNodeEx      pGetNumaProcessorNodeEx = NULL;
+static PGetNumaNodeProcessorMaskEx  pGetNumaNodeProcessorMaskEx = NULL;
+static PGetNumaProcessorNode        pGetNumaProcessorNode = NULL;
+
+//---------------------------------------------
+// Enable large page support dynamically (if possible)
+//---------------------------------------------
+
+static bool win_enable_large_os_pages(size_t* large_page_size)
+{
+  static bool large_initialized = false;
+  if (large_initialized) return (_mi_os_large_page_size() > 0);
+  large_initialized = true;
+
+  // Try to see if large OS pages are supported
+  // To use large pages on Windows, we first need access permission
+  // Set "Lock pages in memory" permission in the group policy editor
+  // <https://devblogs.microsoft.com/oldnewthing/20110128-00/?p=11643>
+  unsigned long err = 0;
+  HANDLE token = NULL;
+  BOOL ok = OpenProcessToken(GetCurrentProcess(), TOKEN_ADJUST_PRIVILEGES | TOKEN_QUERY, &token);
+  if (ok) {
+    TOKEN_PRIVILEGES tp;
+    ok = LookupPrivilegeValue(NULL, TEXT("SeLockMemoryPrivilege"), &tp.Privileges[0].Luid);
+    if (ok) {
+      tp.PrivilegeCount = 1;
+      tp.Privileges[0].Attributes = SE_PRIVILEGE_ENABLED;
+      ok = AdjustTokenPrivileges(token, FALSE, &tp, 0, (PTOKEN_PRIVILEGES)NULL, 0);
+      if (ok) {
+        err = GetLastError();
+        ok = (err == ERROR_SUCCESS);
+        if (ok && large_page_size != NULL) {
+          *large_page_size = GetLargePageMinimum();
+        }
+      }
+    }
+    CloseHandle(token);
+  }
+  if (!ok) {
+    if (err == 0) err = GetLastError();
+    _mi_warning_message("cannot enable large OS page support, error %lu\n", err);
+  }
+  return (ok!=0);
+}
+
+
+//---------------------------------------------
+// Initialize
+//---------------------------------------------
+
+void _mi_prim_mem_init( mi_os_mem_config_t* config )
+{
+  config->has_overcommit = false;
+  config->must_free_whole = true;
+  // get the page size
+  SYSTEM_INFO si;
+  GetSystemInfo(&si);
+  if (si.dwPageSize > 0) { config->page_size = si.dwPageSize; }
+  if (si.dwAllocationGranularity > 0) { config->alloc_granularity = si.dwAllocationGranularity; }
+  // get the VirtualAlloc2 function
+  HINSTANCE  hDll;
+  hDll = LoadLibrary(TEXT("kernelbase.dll"));
+  if (hDll != NULL) {
+    // use VirtualAlloc2FromApp if possible as it is available to Windows store apps
+    pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2FromApp");
+    if (pVirtualAlloc2==NULL) pVirtualAlloc2 = (PVirtualAlloc2)(void (*)(void))GetProcAddress(hDll, "VirtualAlloc2");
+    FreeLibrary(hDll);
+  }
+  // NtAllocateVirtualMemoryEx is used for huge page allocation
+  hDll = LoadLibrary(TEXT("ntdll.dll"));
+  if (hDll != NULL) {
+    pNtAllocateVirtualMemoryEx = (PNtAllocateVirtualMemoryEx)(void (*)(void))GetProcAddress(hDll, "NtAllocateVirtualMemoryEx");
+    FreeLibrary(hDll);
+  }
+  // Try to use Win7+ numa API
+  hDll = LoadLibrary(TEXT("kernel32.dll"));
+  if (hDll != NULL) {
+    pGetCurrentProcessorNumberEx = (PGetCurrentProcessorNumberEx)(void (*)(void))GetProcAddress(hDll, "GetCurrentProcessorNumberEx");
+    pGetNumaProcessorNodeEx = (PGetNumaProcessorNodeEx)(void (*)(void))GetProcAddress(hDll, "GetNumaProcessorNodeEx");
+    pGetNumaNodeProcessorMaskEx = (PGetNumaNodeProcessorMaskEx)(void (*)(void))GetProcAddress(hDll, "GetNumaNodeProcessorMaskEx");
+    pGetNumaProcessorNode = (PGetNumaProcessorNode)(void (*)(void))GetProcAddress(hDll, "GetNumaProcessorNode");
+    FreeLibrary(hDll);
+  }
+  if (mi_option_is_enabled(mi_option_large_os_pages) || mi_option_is_enabled(mi_option_reserve_huge_os_pages)) {
+    win_enable_large_os_pages(&config->large_page_size);
+  }
+}
+
+
+//---------------------------------------------
+// Free
+//---------------------------------------------
+
+int _mi_prim_free(void* addr, size_t size ) {
+  MI_UNUSED(size);
+  DWORD errcode = 0;
+  bool err = (VirtualFree(addr, 0, MEM_RELEASE) == 0);
+  if (err) { errcode = GetLastError(); }
+  if (errcode == ERROR_INVALID_ADDRESS) {
+    // In mi_os_mem_alloc_aligned the fallback path may have returned a pointer inside
+    // the memory region returned by VirtualAlloc; in that case we need to free using
+    // the start of the region.
+    MEMORY_BASIC_INFORMATION info = { 0 };
+    VirtualQuery(addr, &info, sizeof(info));
+    if (info.AllocationBase < addr && ((uint8_t*)addr - (uint8_t*)info.AllocationBase) < (ptrdiff_t)MI_SEGMENT_SIZE) {
+      errcode = 0;
+      err = (VirtualFree(info.AllocationBase, 0, MEM_RELEASE) == 0);
+      if (err) { errcode = GetLastError(); }
+    }
+  }
+  return (int)errcode;
+}
+
+
+//---------------------------------------------
+// VirtualAlloc
+//---------------------------------------------
+
+static void* win_virtual_alloc_prim(void* addr, size_t size, size_t try_alignment, DWORD flags) {
+  #if (MI_INTPTR_SIZE >= 8)
+  // on 64-bit systems, try to use the virtual address area after 2TiB for 4MiB aligned allocations
+  if (addr == NULL) {
+    void* hint = _mi_os_get_aligned_hint(try_alignment,size);
+    if (hint != NULL) {
+      void* p = VirtualAlloc(hint, size, flags, PAGE_READWRITE);
+      if (p != NULL) return p;
+      _mi_verbose_message("warning: unable to allocate hinted aligned OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x)\n", size, GetLastError(), hint, try_alignment, flags);
+      // fall through on error
+    }
+  }
+  #endif
+  // on modern Windows try use VirtualAlloc2 for aligned allocation
+  if (try_alignment > 1 && (try_alignment % _mi_os_page_size()) == 0 && pVirtualAlloc2 != NULL) {
+    MI_MEM_ADDRESS_REQUIREMENTS reqs = { 0, 0, 0 };
+    reqs.Alignment = try_alignment;
+    MI_MEM_EXTENDED_PARAMETER param = { {0, 0}, {0} };
+    param.Type.Type = MiMemExtendedParameterAddressRequirements;
+    param.Arg.Pointer = &reqs;
+    void* p = (*pVirtualAlloc2)(GetCurrentProcess(), addr, size, flags, PAGE_READWRITE, &param, 1);
+    if (p != NULL) return p;
+    _mi_warning_message("unable to allocate aligned OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x)\n", size, GetLastError(), addr, try_alignment, flags);
+    // fall through on error
+  }
+  // last resort
+  return VirtualAlloc(addr, size, flags, PAGE_READWRITE);
+}
+
+static void* win_virtual_alloc(void* addr, size_t size, size_t try_alignment, DWORD flags, bool large_only, bool allow_large, bool* is_large) {
+  mi_assert_internal(!(large_only && !allow_large));
+  static _Atomic(size_t) large_page_try_ok; // = 0;
+  void* p = NULL;
+  // Try to allocate large OS pages (2MiB) if allowed or required.
+  if ((large_only || _mi_os_use_large_page(size, try_alignment))
+      && allow_large && (flags&MEM_COMMIT)!=0 && (flags&MEM_RESERVE)!=0) {
+    size_t try_ok = mi_atomic_load_acquire(&large_page_try_ok);
+    if (!large_only && try_ok > 0) {
+      // if a large page allocation fails, it seems the calls to VirtualAlloc get very expensive.
+      // therefore, once a large page allocation failed, we don't try again for `large_page_try_ok` times.
+      mi_atomic_cas_strong_acq_rel(&large_page_try_ok, &try_ok, try_ok - 1);
+    }
+    else {
+      // large OS pages must always reserve and commit.
+      *is_large = true;
+      p = win_virtual_alloc_prim(addr, size, try_alignment, flags | MEM_LARGE_PAGES);
+      if (large_only) return p;
+      // fall back to non-large page allocation on error (`p == NULL`).
+      if (p == NULL) {
+        mi_atomic_store_release(&large_page_try_ok,10UL);  // on error, don't try again for the next N allocations
+      }
+    }
+  }
+  // Fall back to regular page allocation
+  if (p == NULL) {
+    *is_large = ((flags&MEM_LARGE_PAGES) != 0);
+    p = win_virtual_alloc_prim(addr, size, try_alignment, flags);
+  }
+  //if (p == NULL) { _mi_warning_message("unable to allocate OS memory (%zu bytes, error code: 0x%x, address: %p, alignment: %zu, flags: 0x%x, large only: %d, allow large: %d)\n", size, GetLastError(), addr, try_alignment, flags, large_only, allow_large); }
+  return p;
+}
+
+int _mi_prim_alloc(size_t size, size_t try_alignment, bool commit, bool allow_large, bool* is_large, void** addr) {
+  mi_assert_internal(size > 0 && (size % _mi_os_page_size()) == 0);
+  mi_assert_internal(commit || !allow_large);
+  mi_assert_internal(try_alignment > 0);
+  int flags = MEM_RESERVE;
+  if (commit) { flags |= MEM_COMMIT; }
+  *addr = win_virtual_alloc(NULL, size, try_alignment, flags, false, allow_large, is_large);
+  return (*addr != NULL ? 0 : (int)GetLastError());
+}
+
+
+//---------------------------------------------
+// Commit/Reset/Protect
+//---------------------------------------------
+#ifdef _MSC_VER
+#pragma warning(disable:6250)   // suppress warning calling VirtualFree without MEM_RELEASE (for decommit)
+#endif
+
+int _mi_prim_commit(void* addr, size_t size, bool commit) {
+  if (commit) {
+    void* p = VirtualAlloc(addr, size, MEM_COMMIT, PAGE_READWRITE);
+    return (p == addr ? 0 : (int)GetLastError());
+  }
+  else {
+    BOOL ok = VirtualFree(addr, size, MEM_DECOMMIT);
+    return (ok ? 0 : (int)GetLastError());  
+  }
+}
+
+int _mi_prim_reset(void* addr, size_t size) {
+  void* p = VirtualAlloc(addr, size, MEM_RESET, PAGE_READWRITE);
+  mi_assert_internal(p == addr);
+  #if 1
+  if (p == addr && addr != NULL) {
+    VirtualUnlock(addr,size); // VirtualUnlock after MEM_RESET removes the memory from the working set
+  }
+  #endif
+  return (p == addr ? 0 : (int)GetLastError());
+}
+
+int _mi_prim_protect(void* addr, size_t size, bool protect) {
+  DWORD oldprotect = 0;
+  BOOL ok = VirtualProtect(addr, size, protect ? PAGE_NOACCESS : PAGE_READWRITE, &oldprotect);
+  return (ok ? 0 : (int)GetLastError());
+}
+
+
+//---------------------------------------------
+// Huge page allocation
+//---------------------------------------------
+
+static void* _mi_prim_alloc_huge_os_pagesx(void* hint_addr, size_t size, int numa_node)
+{
+  const DWORD flags = MEM_LARGE_PAGES | MEM_COMMIT | MEM_RESERVE;
+
+  win_enable_large_os_pages(NULL);
+
+  MI_MEM_EXTENDED_PARAMETER params[3] = { {{0,0},{0}},{{0,0},{0}},{{0,0},{0}} };
+  // on modern Windows try use NtAllocateVirtualMemoryEx for 1GiB huge pages
+  static bool mi_huge_pages_available = true;
+  if (pNtAllocateVirtualMemoryEx != NULL && mi_huge_pages_available) {
+    params[0].Type.Type = MiMemExtendedParameterAttributeFlags;
+    params[0].Arg.ULong64 = MI_MEM_EXTENDED_PARAMETER_NONPAGED_HUGE;
+    ULONG param_count = 1;
+    if (numa_node >= 0) {
+      param_count++;
+      params[1].Type.Type = MiMemExtendedParameterNumaNode;
+      params[1].Arg.ULong = (unsigned)numa_node;
+    }
+    SIZE_T psize = size;
+    void* base = hint_addr;
+    NTSTATUS err = (*pNtAllocateVirtualMemoryEx)(GetCurrentProcess(), &base, &psize, flags, PAGE_READWRITE, params, param_count);
+    if (err == 0 && base != NULL) {
+      return base;
+    }
+    else {
+      // fall back to regular large pages
+      mi_huge_pages_available = false; // don't try further huge pages
+      _mi_warning_message("unable to allocate using huge (1GiB) pages, trying large (2MiB) pages instead (status 0x%lx)\n", err);
+    }
+  }
+  // on modern Windows try use VirtualAlloc2 for numa aware large OS page allocation
+  if (pVirtualAlloc2 != NULL && numa_node >= 0) {
+    params[0].Type.Type = MiMemExtendedParameterNumaNode;
+    params[0].Arg.ULong = (unsigned)numa_node;
+    return (*pVirtualAlloc2)(GetCurrentProcess(), hint_addr, size, flags, PAGE_READWRITE, params, 1);
+  }
+
+  // otherwise use regular virtual alloc on older windows
+  return VirtualAlloc(hint_addr, size, flags, PAGE_READWRITE);
+}
+
+int _mi_prim_alloc_huge_os_pages(void* hint_addr, size_t size, int numa_node, void** addr) {
+  *addr = _mi_prim_alloc_huge_os_pagesx(hint_addr,size,numa_node);
+  return (*addr != NULL ? 0 : (int)GetLastError());
+}
+
+
+//---------------------------------------------
+// Numa nodes
+//---------------------------------------------
+
+size_t _mi_prim_numa_node(void) {
+  USHORT numa_node = 0;
+  if (pGetCurrentProcessorNumberEx != NULL && pGetNumaProcessorNodeEx != NULL) {
+    // Extended API is supported
+    MI_PROCESSOR_NUMBER pnum;
+    (*pGetCurrentProcessorNumberEx)(&pnum);
+    USHORT nnode = 0;
+    BOOL ok = (*pGetNumaProcessorNodeEx)(&pnum, &nnode);
+    if (ok) { numa_node = nnode; }
+  }
+  else if (pGetNumaProcessorNode != NULL) {
+    // Vista or earlier, use older API that is limited to 64 processors. Issue #277
+    DWORD pnum = GetCurrentProcessorNumber();
+    UCHAR nnode = 0;
+    BOOL ok = pGetNumaProcessorNode((UCHAR)pnum, &nnode);
+    if (ok) { numa_node = nnode; }
+  }
+  return numa_node;
+}
+
+size_t _mi_prim_numa_node_count(void) {
+  ULONG numa_max = 0;
+  GetNumaHighestNodeNumber(&numa_max);
+  // find the highest node number that has actual processors assigned to it. Issue #282
+  while(numa_max > 0) {
+    if (pGetNumaNodeProcessorMaskEx != NULL) {
+      // Extended API is supported
+      GROUP_AFFINITY affinity;
+      if ((*pGetNumaNodeProcessorMaskEx)((USHORT)numa_max, &affinity)) {
+        if (affinity.Mask != 0) break;  // found the maximum non-empty node
+      }
+    }
+    else {
+      // Vista or earlier, use older API that is limited to 64 processors.
+      ULONGLONG mask;
+      if (GetNumaNodeProcessorMask((UCHAR)numa_max, &mask)) {
+        if (mask != 0) break; // found the maximum non-empty node
+      };
+    }
+    // max node was invalid or had no processor assigned, try again
+    numa_max--;
+  }
+  return ((size_t)numa_max + 1);
+}
+
+
+//----------------------------------------------------------------
+// Clock
+//----------------------------------------------------------------
+
+static mi_msecs_t mi_to_msecs(LARGE_INTEGER t) {
+  static LARGE_INTEGER mfreq; // = 0
+  if (mfreq.QuadPart == 0LL) {
+    LARGE_INTEGER f;
+    QueryPerformanceFrequency(&f);
+    mfreq.QuadPart = f.QuadPart/1000LL;
+    if (mfreq.QuadPart == 0) mfreq.QuadPart = 1;
+  }
+  return (mi_msecs_t)(t.QuadPart / mfreq.QuadPart);
+}
+
+mi_msecs_t _mi_prim_clock_now(void) {
+  LARGE_INTEGER t;
+  QueryPerformanceCounter(&t);
+  return mi_to_msecs(t);
+}
+
+
+//----------------------------------------------------------------
+// Process Info
+//----------------------------------------------------------------
+
+#include <windows.h>
+#include <psapi.h>
+
+static mi_msecs_t filetime_msecs(const FILETIME* ftime) {
+  ULARGE_INTEGER i;
+  i.LowPart = ftime->dwLowDateTime;
+  i.HighPart = ftime->dwHighDateTime;
+  mi_msecs_t msecs = (i.QuadPart / 10000); // FILETIME is in 100 nano seconds
+  return msecs;
+}
+
+typedef BOOL (WINAPI *PGetProcessMemoryInfo)(HANDLE, PPROCESS_MEMORY_COUNTERS, DWORD);
+static PGetProcessMemoryInfo pGetProcessMemoryInfo = NULL;
+
+void _mi_prim_process_info(mi_process_info_t* pinfo)
+{
+  FILETIME ct;
+  FILETIME ut;
+  FILETIME st;
+  FILETIME et;
+  GetProcessTimes(GetCurrentProcess(), &ct, &et, &st, &ut);
+  pinfo->utime = filetime_msecs(&ut);
+  pinfo->stime = filetime_msecs(&st);
+  
+  // load psapi on demand
+  if (pGetProcessMemoryInfo == NULL) {
+    HINSTANCE hDll = LoadLibrary(TEXT("psapi.dll"));
+    if (hDll != NULL) {
+      pGetProcessMemoryInfo = (PGetProcessMemoryInfo)(void (*)(void))GetProcAddress(hDll, "GetProcessMemoryInfo");
+    }
+  }
+
+  // get process info
+  PROCESS_MEMORY_COUNTERS info;
+  memset(&info, 0, sizeof(info));
+  if (pGetProcessMemoryInfo != NULL) {
+    pGetProcessMemoryInfo(GetCurrentProcess(), &info, sizeof(info));
+  } 
+  pinfo->current_rss    = (size_t)info.WorkingSetSize;
+  pinfo->peak_rss       = (size_t)info.PeakWorkingSetSize;
+  pinfo->current_commit = (size_t)info.PagefileUsage;
+  pinfo->peak_commit    = (size_t)info.PeakPagefileUsage;
+  pinfo->page_faults    = (size_t)info.PageFaultCount;
+}
+
+//----------------------------------------------------------------
+// Output
+//----------------------------------------------------------------
+
+void _mi_prim_out_stderr( const char* msg ) 
+{
+  // on windows with redirection, the C runtime cannot handle locale dependent output
+  // after the main thread closes so we use direct console output.
+  if (!_mi_preloading()) {
+    // _cputs(msg);  // _cputs cannot be used at is aborts if it fails to lock the console
+    static HANDLE hcon = INVALID_HANDLE_VALUE;
+    static bool hconIsConsole;
+    if (hcon == INVALID_HANDLE_VALUE) {
+      CONSOLE_SCREEN_BUFFER_INFO sbi;
+      hcon = GetStdHandle(STD_ERROR_HANDLE);
+      hconIsConsole = ((hcon != INVALID_HANDLE_VALUE) && GetConsoleScreenBufferInfo(hcon, &sbi));
+    }
+    const size_t len = _mi_strlen(msg);
+    if (len > 0 && len < UINT32_MAX) {
+      DWORD written = 0;
+      if (hconIsConsole) {
+        WriteConsoleA(hcon, msg, (DWORD)len, &written, NULL);
+      }
+      else if (hcon != INVALID_HANDLE_VALUE) {
+        // use direct write if stderr was redirected
+        WriteFile(hcon, msg, (DWORD)len, &written, NULL);
+      }
+      else {
+        // finally fall back to fputs after all
+        fputs(msg, stderr);
+      }
+    }
+  }
+}
+
+
+//----------------------------------------------------------------
+// Environment
+//----------------------------------------------------------------
+
+// On Windows use GetEnvironmentVariable instead of getenv to work
+// reliably even when this is invoked before the C runtime is initialized.
+// i.e. when `_mi_preloading() == true`.
+// Note: on windows, environment names are not case sensitive.
+bool _mi_prim_getenv(const char* name, char* result, size_t result_size) {
+  result[0] = 0;
+  size_t len = GetEnvironmentVariableA(name, result, (DWORD)result_size);
+  return (len > 0 && len < result_size);
+}
+
+
+
+//----------------------------------------------------------------
+// Random
+//----------------------------------------------------------------
+
+#if defined(MI_USE_RTLGENRANDOM) // || defined(__cplusplus)
+// We prefer to use BCryptGenRandom instead of (the unofficial) RtlGenRandom but when using
+// dynamic overriding, we observed it can raise an exception when compiled with C++, and
+// sometimes deadlocks when also running under the VS debugger.
+// In contrast, issue #623 implies that on Windows Server 2019 we need to use BCryptGenRandom.
+// To be continued..
+#pragma comment (lib,"advapi32.lib")
+#define RtlGenRandom  SystemFunction036
+mi_decl_externc BOOLEAN NTAPI RtlGenRandom(PVOID RandomBuffer, ULONG RandomBufferLength);
+
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  return (RtlGenRandom(buf, (ULONG)buf_len) != 0);
+}
+
+#else
+
+#ifndef BCRYPT_USE_SYSTEM_PREFERRED_RNG
+#define BCRYPT_USE_SYSTEM_PREFERRED_RNG 0x00000002
+#endif
+
+typedef LONG (NTAPI *PBCryptGenRandom)(HANDLE, PUCHAR, ULONG, ULONG);
+static  PBCryptGenRandom pBCryptGenRandom = NULL;
+
+bool _mi_prim_random_buf(void* buf, size_t buf_len) {
+  if (pBCryptGenRandom == NULL) {
+    HINSTANCE hDll = LoadLibrary(TEXT("bcrypt.dll"));
+    if (hDll != NULL) {
+      pBCryptGenRandom = (PBCryptGenRandom)(void (*)(void))GetProcAddress(hDll, "BCryptGenRandom");
+    }
+    if (pBCryptGenRandom == NULL) return false;
+  }
+  return (pBCryptGenRandom(NULL, (PUCHAR)buf, (ULONG)buf_len, BCRYPT_USE_SYSTEM_PREFERRED_RNG) >= 0);  
+}
+
+#endif  // MI_USE_RTLGENRANDOM
+
+//----------------------------------------------------------------
+// Thread init/done
+//----------------------------------------------------------------
+
+#if !defined(MI_SHARED_LIB)
+
+// use thread local storage keys to detect thread ending
+#include <fibersapi.h>
+#if (_WIN32_WINNT < 0x600)  // before Windows Vista
+WINBASEAPI DWORD WINAPI FlsAlloc( _In_opt_ PFLS_CALLBACK_FUNCTION lpCallback );
+WINBASEAPI PVOID WINAPI FlsGetValue( _In_ DWORD dwFlsIndex );
+WINBASEAPI BOOL  WINAPI FlsSetValue( _In_ DWORD dwFlsIndex, _In_opt_ PVOID lpFlsData );
+WINBASEAPI BOOL  WINAPI FlsFree(_In_ DWORD dwFlsIndex);
+#endif
+
+static DWORD mi_fls_key = (DWORD)(-1);
+
+static void NTAPI mi_fls_done(PVOID value) {
+  mi_heap_t* heap = (mi_heap_t*)value;
+  if (heap != NULL) {
+    _mi_thread_done(heap);
+    FlsSetValue(mi_fls_key, NULL);  // prevent recursion as _mi_thread_done may set it back to the main heap, issue #672
+  }
+}
+
+void _mi_prim_thread_init_auto_done(void) {
+  mi_fls_key = FlsAlloc(&mi_fls_done);
+}
+
+void _mi_prim_thread_done_auto_done(void) {
+  // call thread-done on all threads (except the main thread) to prevent 
+  // dangling callback pointer if statically linked with a DLL; Issue #208
+  FlsFree(mi_fls_key);  
+}
+
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
+  mi_assert_internal(mi_fls_key != (DWORD)(-1));
+  FlsSetValue(mi_fls_key, heap);
+}
+
+#else
+
+// Dll; nothing to do as in that case thread_done is handled through the DLL_THREAD_DETACH event.
+
+void _mi_prim_thread_init_auto_done(void) {
+}
+
+void _mi_prim_thread_done_auto_done(void) {
+}
+
+void _mi_prim_thread_associate_default_heap(mi_heap_t* heap) {
+  MI_UNUSED(heap);
+}
+
+#endif

preload-mimalloc/mimalloc/src/prim/windows/readme.md

@@ -0,0 +1,17 @@
+## Primitives:
+
+- `prim.c` contains Windows primitives for OS allocation.
+
+## Event Tracing for Windows (ETW)
+
+- `etw.h` is generated from `etw.man` which contains the manifest for mimalloc events.
+  (100 is an allocation, 101 is for a free)
+
+- `etw-mimalloc.wprp` is a profile for the Windows Performance Recorder (WPR).
+  In an admin prompt, you can use:
+  ```
+  > wpr -start src\prim\windows\etw-mimalloc.wprp -filemode
+  > <my mimalloc program>
+  > wpr -stop test.etl
+  ``` 
+  and then open `test.etl` in the Windows Performance Analyzer (WPA).

preload-mimalloc/mimalloc/src/random.c

@@ -0,0 +1,254 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/prim.h"    // _mi_prim_random_buf
+#include <string.h>       // memset
+
+/* ----------------------------------------------------------------------------
+We use our own PRNG to keep predictable performance of random number generation
+and to avoid implementations that use a lock. We only use the OS provided
+random source to initialize the initial seeds. Since we do not need ultimate
+performance but we do rely on the security (for secret cookies in secure mode)
+we use a cryptographically secure generator (chacha20).
+-----------------------------------------------------------------------------*/
+
+#define MI_CHACHA_ROUNDS (20)   // perhaps use 12 for better performance?
+
+
+/* ----------------------------------------------------------------------------
+Chacha20 implementation as the original algorithm with a 64-bit nonce
+and counter: https://en.wikipedia.org/wiki/Salsa20
+The input matrix has sixteen 32-bit values:
+Position  0 to  3: constant key
+Position  4 to 11: the key
+Position 12 to 13: the counter.
+Position 14 to 15: the nonce.
+
+The implementation uses regular C code which compiles very well on modern compilers.
+(gcc x64 has no register spills, and clang 6+ uses SSE instructions)
+-----------------------------------------------------------------------------*/
+
+static inline uint32_t rotl(uint32_t x, uint32_t shift) {
+  return (x << shift) | (x >> (32 - shift));
+}
+
+static inline void qround(uint32_t x[16], size_t a, size_t b, size_t c, size_t d) {
+  x[a] += x[b]; x[d] = rotl(x[d] ^ x[a], 16);
+  x[c] += x[d]; x[b] = rotl(x[b] ^ x[c], 12);
+  x[a] += x[b]; x[d] = rotl(x[d] ^ x[a], 8);
+  x[c] += x[d]; x[b] = rotl(x[b] ^ x[c], 7);
+}
+
+static void chacha_block(mi_random_ctx_t* ctx)
+{
+  // scramble into `x`
+  uint32_t x[16];
+  for (size_t i = 0; i < 16; i++) {
+    x[i] = ctx->input[i];
+  }
+  for (size_t i = 0; i < MI_CHACHA_ROUNDS; i += 2) {
+    qround(x, 0, 4,  8, 12);
+    qround(x, 1, 5,  9, 13);
+    qround(x, 2, 6, 10, 14);
+    qround(x, 3, 7, 11, 15);
+    qround(x, 0, 5, 10, 15);
+    qround(x, 1, 6, 11, 12);
+    qround(x, 2, 7,  8, 13);
+    qround(x, 3, 4,  9, 14);
+  }
+
+  // add scrambled data to the initial state
+  for (size_t i = 0; i < 16; i++) {
+    ctx->output[i] = x[i] + ctx->input[i];
+  }
+  ctx->output_available = 16;
+
+  // increment the counter for the next round
+  ctx->input[12] += 1;
+  if (ctx->input[12] == 0) {
+    ctx->input[13] += 1;
+    if (ctx->input[13] == 0) {  // and keep increasing into the nonce
+      ctx->input[14] += 1;
+    }
+  }
+}
+
+static uint32_t chacha_next32(mi_random_ctx_t* ctx) {
+  if (ctx->output_available <= 0) {
+    chacha_block(ctx);
+    ctx->output_available = 16; // (assign again to suppress static analysis warning)
+  }
+  const uint32_t x = ctx->output[16 - ctx->output_available];
+  ctx->output[16 - ctx->output_available] = 0; // reset once the data is handed out
+  ctx->output_available--;
+  return x;
+}
+
+static inline uint32_t read32(const uint8_t* p, size_t idx32) {
+  const size_t i = 4*idx32;
+  return ((uint32_t)p[i+0] | (uint32_t)p[i+1] << 8 | (uint32_t)p[i+2] << 16 | (uint32_t)p[i+3] << 24);
+}
+
+static void chacha_init(mi_random_ctx_t* ctx, const uint8_t key[32], uint64_t nonce)
+{
+  // since we only use chacha for randomness (and not encryption) we
+  // do not _need_ to read 32-bit values as little endian but we do anyways
+  // just for being compatible :-)
+  memset(ctx, 0, sizeof(*ctx));
+  for (size_t i = 0; i < 4; i++) {
+    const uint8_t* sigma = (uint8_t*)"expand 32-byte k";
+    ctx->input[i] = read32(sigma,i);
+  }
+  for (size_t i = 0; i < 8; i++) {
+    ctx->input[i + 4] = read32(key,i);
+  }
+  ctx->input[12] = 0;
+  ctx->input[13] = 0;
+  ctx->input[14] = (uint32_t)nonce;
+  ctx->input[15] = (uint32_t)(nonce >> 32);
+}
+
+static void chacha_split(mi_random_ctx_t* ctx, uint64_t nonce, mi_random_ctx_t* ctx_new) {
+  memset(ctx_new, 0, sizeof(*ctx_new));
+  _mi_memcpy(ctx_new->input, ctx->input, sizeof(ctx_new->input));
+  ctx_new->input[12] = 0;
+  ctx_new->input[13] = 0;
+  ctx_new->input[14] = (uint32_t)nonce;
+  ctx_new->input[15] = (uint32_t)(nonce >> 32);
+  mi_assert_internal(ctx->input[14] != ctx_new->input[14] || ctx->input[15] != ctx_new->input[15]); // do not reuse nonces!
+  chacha_block(ctx_new);
+}
+
+
+/* ----------------------------------------------------------------------------
+Random interface
+-----------------------------------------------------------------------------*/
+
+#if MI_DEBUG>1
+static bool mi_random_is_initialized(mi_random_ctx_t* ctx) {
+  return (ctx != NULL && ctx->input[0] != 0);
+}
+#endif
+
+void _mi_random_split(mi_random_ctx_t* ctx, mi_random_ctx_t* ctx_new) {
+  mi_assert_internal(mi_random_is_initialized(ctx));
+  mi_assert_internal(ctx != ctx_new);
+  chacha_split(ctx, (uintptr_t)ctx_new /*nonce*/, ctx_new);
+}
+
+uintptr_t _mi_random_next(mi_random_ctx_t* ctx) {
+  mi_assert_internal(mi_random_is_initialized(ctx));
+  #if MI_INTPTR_SIZE <= 4
+    return chacha_next32(ctx);
+  #elif MI_INTPTR_SIZE == 8
+    return (((uintptr_t)chacha_next32(ctx) << 32) | chacha_next32(ctx));
+  #else
+  # error "define mi_random_next for this platform"
+  #endif
+}
+
+
+/* ----------------------------------------------------------------------------
+To initialize a fresh random context.
+If we cannot get good randomness, we fall back to weak randomness based on a timer and ASLR.
+-----------------------------------------------------------------------------*/
+
+uintptr_t _mi_os_random_weak(uintptr_t extra_seed) {
+  uintptr_t x = (uintptr_t)&_mi_os_random_weak ^ extra_seed; // ASLR makes the address random
+  x ^= _mi_prim_clock_now();  
+  // and do a few randomization steps
+  uintptr_t max = ((x ^ (x >> 17)) & 0x0F) + 1;
+  for (uintptr_t i = 0; i < max; i++) {
+    x = _mi_random_shuffle(x);
+  }
+  mi_assert_internal(x != 0);
+  return x;
+}
+
+static void mi_random_init_ex(mi_random_ctx_t* ctx, bool use_weak) {
+  uint8_t key[32];
+  if (use_weak || !_mi_prim_random_buf(key, sizeof(key))) {
+    // if we fail to get random data from the OS, we fall back to a
+    // weak random source based on the current time
+    #if !defined(__wasi__)
+    if (!use_weak) { _mi_warning_message("unable to use secure randomness\n"); }
+    #endif
+    uintptr_t x = _mi_os_random_weak(0);
+    for (size_t i = 0; i < 8; i++) {  // key is eight 32-bit words.
+      x = _mi_random_shuffle(x);
+      ((uint32_t*)key)[i] = (uint32_t)x;
+    }
+    ctx->weak = true;
+  }
+  else {
+    ctx->weak = false;
+  }
+  chacha_init(ctx, key, (uintptr_t)ctx /*nonce*/ );
+}
+
+void _mi_random_init(mi_random_ctx_t* ctx) {
+  mi_random_init_ex(ctx, false);
+}
+
+void _mi_random_init_weak(mi_random_ctx_t * ctx) {
+  mi_random_init_ex(ctx, true);
+}
+
+void _mi_random_reinit_if_weak(mi_random_ctx_t * ctx) {
+  if (ctx->weak) {
+    _mi_random_init(ctx);
+  }
+}
+
+/* --------------------------------------------------------
+test vectors from <https://tools.ietf.org/html/rfc8439>
+----------------------------------------------------------- */
+/*
+static bool array_equals(uint32_t* x, uint32_t* y, size_t n) {
+  for (size_t i = 0; i < n; i++) {
+    if (x[i] != y[i]) return false;
+  }
+  return true;
+}
+static void chacha_test(void)
+{
+  uint32_t x[4] = { 0x11111111, 0x01020304, 0x9b8d6f43, 0x01234567 };
+  uint32_t x_out[4] = { 0xea2a92f4, 0xcb1cf8ce, 0x4581472e, 0x5881c4bb };
+  qround(x, 0, 1, 2, 3);
+  mi_assert_internal(array_equals(x, x_out, 4));
+
+  uint32_t y[16] = {
+       0x879531e0,  0xc5ecf37d,  0x516461b1,  0xc9a62f8a,
+       0x44c20ef3,  0x3390af7f,  0xd9fc690b,  0x2a5f714c,
+       0x53372767,  0xb00a5631,  0x974c541a,  0x359e9963,
+       0x5c971061,  0x3d631689,  0x2098d9d6,  0x91dbd320 };
+  uint32_t y_out[16] = {
+       0x879531e0,  0xc5ecf37d,  0xbdb886dc,  0xc9a62f8a,
+       0x44c20ef3,  0x3390af7f,  0xd9fc690b,  0xcfacafd2,
+       0xe46bea80,  0xb00a5631,  0x974c541a,  0x359e9963,
+       0x5c971061,  0xccc07c79,  0x2098d9d6,  0x91dbd320 };
+  qround(y, 2, 7, 8, 13);
+  mi_assert_internal(array_equals(y, y_out, 16));
+
+  mi_random_ctx_t r = {
+    { 0x61707865, 0x3320646e, 0x79622d32, 0x6b206574,
+      0x03020100, 0x07060504, 0x0b0a0908, 0x0f0e0d0c,
+      0x13121110, 0x17161514, 0x1b1a1918, 0x1f1e1d1c,
+      0x00000001, 0x09000000, 0x4a000000, 0x00000000 },
+    {0},
+    0
+  };
+  uint32_t r_out[16] = {
+       0xe4e7f110, 0x15593bd1, 0x1fdd0f50, 0xc47120a3,
+       0xc7f4d1c7, 0x0368c033, 0x9aaa2204, 0x4e6cd4c3,
+       0x466482d2, 0x09aa9f07, 0x05d7c214, 0xa2028bd9,
+       0xd19c12b5, 0xb94e16de, 0xe883d0cb, 0x4e3c50a2 };
+  chacha_block(&r);
+  mi_assert_internal(array_equals(r.output, r_out, 16));
+}
+*/

preload-mimalloc/mimalloc/src/region.c

@@ -0,0 +1,501 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2019-2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+This implements a layer between the raw OS memory (VirtualAlloc/mmap/sbrk/..)
+and the segment and huge object allocation by mimalloc. There may be multiple
+implementations of this (one could be the identity going directly to the OS,
+another could be a simple cache etc), but the current one uses large "regions".
+In contrast to the rest of mimalloc, the "regions" are shared between threads and
+need to be accessed using atomic operations.
+We need this memory layer between the raw OS calls because of:
+1. on `sbrk` like systems (like WebAssembly) we need our own memory maps in order
+   to reuse memory effectively.
+2. It turns out that for large objects, between 1MiB and 32MiB (?), the cost of
+   an OS allocation/free is still (much) too expensive relative to the accesses
+   in that object :-( (`malloc-large` tests this). This means we need a cheaper
+   way to reuse memory.
+3. This layer allows for NUMA aware allocation.
+
+Possible issues:
+- (2) can potentially be addressed too with a small cache per thread which is much
+  simpler. Generally though that requires shrinking of huge pages, and may overuse
+  memory per thread. (and is not compatible with `sbrk`).
+- Since the current regions are per-process, we need atomic operations to
+  claim blocks which may be contended
+- In the worst case, we need to search the whole region map (16KiB for 256GiB)
+  linearly. At what point will direct OS calls be faster? Is there a way to
+  do this better without adding too much complexity?
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+#include <string.h>  // memset
+
+#include "bitmap.h"
+
+// os.c
+bool _mi_os_unreset(void* addr, size_t size, bool* is_zero, mi_stats_t* tld_stats);
+
+// Constants
+#if (MI_INTPTR_SIZE==8)
+#define MI_HEAP_REGION_MAX_SIZE    (256 * MI_GiB)  // 64KiB for the region map
+#elif (MI_INTPTR_SIZE==4)
+#define MI_HEAP_REGION_MAX_SIZE    (3 * MI_GiB)    // ~ KiB for the region map
+#else
+#error "define the maximum heap space allowed for regions on this platform"
+#endif
+
+#define MI_REGION_MAX_BLOCKS      MI_BITMAP_FIELD_BITS
+#define MI_REGION_SIZE            (MI_SEGMENT_SIZE * MI_BITMAP_FIELD_BITS)    // 256MiB  (64MiB on 32 bits)
+#define MI_REGION_MAX             (MI_HEAP_REGION_MAX_SIZE / MI_REGION_SIZE)  // 1024  (48 on 32 bits)
+#define MI_REGION_MAX_OBJ_BLOCKS  (MI_REGION_MAX_BLOCKS/4)                    // 64MiB
+#define MI_REGION_MAX_OBJ_SIZE    (MI_REGION_MAX_OBJ_BLOCKS*MI_SEGMENT_SIZE)
+
+// Region info
+typedef union mi_region_info_u {
+  size_t value;
+  struct {
+    bool  valid;        // initialized?
+    bool  is_large:1;   // allocated in fixed large/huge OS pages
+    bool  is_pinned:1;  // pinned memory cannot be decommitted
+    short numa_node;    // the associated NUMA node (where -1 means no associated node)
+  } x;
+} mi_region_info_t;
+
+
+// A region owns a chunk of REGION_SIZE (256MiB) (virtual) memory with
+// a bit map with one bit per MI_SEGMENT_SIZE (4MiB) block.
+typedef struct mem_region_s {
+  _Atomic(size_t)           info;        // mi_region_info_t.value
+  _Atomic(void*)            start;       // start of the memory area
+  mi_bitmap_field_t         in_use;      // bit per in-use block
+  mi_bitmap_field_t         dirty;       // track if non-zero per block
+  mi_bitmap_field_t         commit;      // track if committed per block
+  mi_bitmap_field_t         reset;       // track if reset per block
+  _Atomic(size_t)           arena_memid; // if allocated from a (huge page) arena
+  _Atomic(size_t)           padding;     // round to 8 fields (needs to be atomic for msvc, see issue #508)
+} mem_region_t;
+
+// The region map
+static mem_region_t regions[MI_REGION_MAX];
+
+// Allocated regions
+static _Atomic(size_t) regions_count; // = 0;
+
+
+/* ----------------------------------------------------------------------------
+Utility functions
+-----------------------------------------------------------------------------*/
+
+// Blocks (of 4MiB) needed for the given size.
+static size_t mi_region_block_count(size_t size) {
+  return _mi_divide_up(size, MI_SEGMENT_SIZE);
+}
+
+/*
+// Return a rounded commit/reset size such that we don't fragment large OS pages into small ones.
+static size_t mi_good_commit_size(size_t size) {
+  if (size > (SIZE_MAX - _mi_os_large_page_size())) return size;
+  return _mi_align_up(size, _mi_os_large_page_size());
+}
+*/
+
+// Return if a pointer points into a region reserved by us.
+mi_decl_nodiscard bool mi_is_in_heap_region(const void* p) mi_attr_noexcept {
+  if (p==NULL) return false;
+  size_t count = mi_atomic_load_relaxed(&regions_count);
+  for (size_t i = 0; i < count; i++) {
+    uint8_t* start = (uint8_t*)mi_atomic_load_ptr_relaxed(uint8_t, &regions[i].start);
+    if (start != NULL && (uint8_t*)p >= start && (uint8_t*)p < start + MI_REGION_SIZE) return true;
+  }
+  return false;
+}
+
+
+static void* mi_region_blocks_start(const mem_region_t* region, mi_bitmap_index_t bit_idx) {
+  uint8_t* start = (uint8_t*)mi_atomic_load_ptr_acquire(uint8_t, &((mem_region_t*)region)->start);
+  mi_assert_internal(start != NULL);
+  return (start + (bit_idx * MI_SEGMENT_SIZE));
+}
+
+static size_t mi_memid_create(mem_region_t* region, mi_bitmap_index_t bit_idx) {
+  mi_assert_internal(bit_idx < MI_BITMAP_FIELD_BITS);
+  size_t idx = region - regions;
+  mi_assert_internal(&regions[idx] == region);
+  return (idx*MI_BITMAP_FIELD_BITS + bit_idx)<<1;
+}
+
+static size_t mi_memid_create_from_arena(size_t arena_memid) {
+  return (arena_memid << 1) | 1;
+}
+
+
+static bool mi_memid_is_arena(size_t id, mem_region_t** region, mi_bitmap_index_t* bit_idx, size_t* arena_memid) {
+  if ((id&1)==1) {
+    if (arena_memid != NULL) *arena_memid = (id>>1);
+    return true;
+  }
+  else {
+    size_t idx = (id >> 1) / MI_BITMAP_FIELD_BITS;
+    *bit_idx   = (mi_bitmap_index_t)(id>>1) % MI_BITMAP_FIELD_BITS;
+    *region    = &regions[idx];
+    return false;
+  }
+}
+
+
+/* ----------------------------------------------------------------------------
+  Allocate a region is allocated from the OS (or an arena)
+-----------------------------------------------------------------------------*/
+
+static bool mi_region_try_alloc_os(size_t blocks, bool commit, bool allow_large, mem_region_t** region, mi_bitmap_index_t* bit_idx, mi_os_tld_t* tld)
+{
+  // not out of regions yet?
+  if (mi_atomic_load_relaxed(&regions_count) >= MI_REGION_MAX - 1) return false;
+
+  // try to allocate a fresh region from the OS
+  bool region_commit = (commit && mi_option_is_enabled(mi_option_eager_region_commit));
+  bool region_large = (commit && allow_large);
+  bool is_zero = false;
+  bool is_pinned = false;
+  size_t arena_memid = 0;
+  void* const start = _mi_arena_alloc_aligned(MI_REGION_SIZE, MI_SEGMENT_ALIGN, 0, &region_commit, &region_large, &is_pinned, &is_zero, _mi_arena_id_none(),  & arena_memid, tld);
+  if (start == NULL) return false;
+  mi_assert_internal(!(region_large && !allow_large));
+  mi_assert_internal(!region_large || region_commit);
+
+  // claim a fresh slot
+  const size_t idx = mi_atomic_increment_acq_rel(&regions_count);
+  if (idx >= MI_REGION_MAX) {
+    mi_atomic_decrement_acq_rel(&regions_count);
+    _mi_arena_free(start, MI_REGION_SIZE, MI_SEGMENT_ALIGN, 0, arena_memid, region_commit, tld->stats);
+    _mi_warning_message("maximum regions used: %zu GiB (perhaps recompile with a larger setting for MI_HEAP_REGION_MAX_SIZE)", _mi_divide_up(MI_HEAP_REGION_MAX_SIZE, MI_GiB));
+    return false;
+  }
+
+  // allocated, initialize and claim the initial blocks
+  mem_region_t* r = &regions[idx];
+  r->arena_memid  = arena_memid;
+  mi_atomic_store_release(&r->in_use, (size_t)0);
+  mi_atomic_store_release(&r->dirty, (is_zero ? 0 : MI_BITMAP_FIELD_FULL));
+  mi_atomic_store_release(&r->commit, (region_commit ? MI_BITMAP_FIELD_FULL : 0));
+  mi_atomic_store_release(&r->reset, (size_t)0);
+  *bit_idx = 0;
+  _mi_bitmap_claim(&r->in_use, 1, blocks, *bit_idx, NULL);
+  mi_atomic_store_ptr_release(void,&r->start, start);
+
+  // and share it
+  mi_region_info_t info;
+  info.value = 0;                        // initialize the full union to zero
+  info.x.valid = true;
+  info.x.is_large = region_large;
+  info.x.is_pinned = is_pinned;
+  info.x.numa_node = (short)_mi_os_numa_node(tld);
+  mi_atomic_store_release(&r->info, info.value); // now make it available to others
+  *region = r;
+  return true;
+}
+
+/* ----------------------------------------------------------------------------
+  Try to claim blocks in suitable regions
+-----------------------------------------------------------------------------*/
+
+static bool mi_region_is_suitable(const mem_region_t* region, int numa_node, bool allow_large ) {
+  // initialized at all?
+  mi_region_info_t info;
+  info.value = mi_atomic_load_relaxed(&((mem_region_t*)region)->info);
+  if (info.value==0) return false;
+
+  // numa correct
+  if (numa_node >= 0) {  // use negative numa node to always succeed
+    int rnode = info.x.numa_node;
+    if (rnode >= 0 && rnode != numa_node) return false;
+  }
+
+  // check allow-large
+  if (!allow_large && info.x.is_large) return false;
+
+  return true;
+}
+
+
+static bool mi_region_try_claim(int numa_node, size_t blocks, bool allow_large, mem_region_t** region, mi_bitmap_index_t* bit_idx, mi_os_tld_t* tld)
+{
+  // try all regions for a free slot
+  const size_t count = mi_atomic_load_relaxed(&regions_count); // monotonic, so ok to be relaxed
+  size_t idx = tld->region_idx; // Or start at 0 to reuse low addresses? Starting at 0 seems to increase latency though
+  for (size_t visited = 0; visited < count; visited++, idx++) {
+    if (idx >= count) idx = 0;  // wrap around
+    mem_region_t* r = &regions[idx];
+    // if this region suits our demand (numa node matches, large OS page matches)
+    if (mi_region_is_suitable(r, numa_node, allow_large)) {
+      // then try to atomically claim a segment(s) in this region
+      if (_mi_bitmap_try_find_claim_field(&r->in_use, 0, blocks, bit_idx)) {
+        tld->region_idx = idx;    // remember the last found position
+        *region = r;
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+
+static void* mi_region_try_alloc(size_t blocks, bool* commit, bool* large, bool* is_pinned, bool* is_zero, size_t* memid, mi_os_tld_t* tld)
+{
+  mi_assert_internal(blocks <= MI_BITMAP_FIELD_BITS);
+  mem_region_t* region;
+  mi_bitmap_index_t bit_idx;
+  const int numa_node = (_mi_os_numa_node_count() <= 1 ? -1 : _mi_os_numa_node(tld));
+  // try to claim in existing regions
+  if (!mi_region_try_claim(numa_node, blocks, *large, &region, &bit_idx, tld)) {
+    // otherwise try to allocate a fresh region and claim in there
+    if (!mi_region_try_alloc_os(blocks, *commit, *large, &region, &bit_idx, tld)) {
+      // out of regions or memory
+      return NULL;
+    }
+  }
+
+  // ------------------------------------------------
+  // found a region and claimed `blocks` at `bit_idx`, initialize them now
+  mi_assert_internal(region != NULL);
+  mi_assert_internal(_mi_bitmap_is_claimed(&region->in_use, 1, blocks, bit_idx));
+
+  mi_region_info_t info;
+  info.value = mi_atomic_load_acquire(&region->info);
+  uint8_t* start = (uint8_t*)mi_atomic_load_ptr_acquire(uint8_t,&region->start);
+  mi_assert_internal(!(info.x.is_large && !*large));
+  mi_assert_internal(start != NULL);
+
+  *is_zero   = _mi_bitmap_claim(&region->dirty, 1, blocks, bit_idx, NULL);
+  *large     = info.x.is_large;
+  *is_pinned = info.x.is_pinned;
+  *memid     = mi_memid_create(region, bit_idx);
+  void* p = start + (mi_bitmap_index_bit_in_field(bit_idx) * MI_SEGMENT_SIZE);
+
+  // commit
+  if (*commit) {
+    // ensure commit
+    bool any_uncommitted;
+    _mi_bitmap_claim(&region->commit, 1, blocks, bit_idx, &any_uncommitted);
+    if (any_uncommitted) {
+      mi_assert_internal(!info.x.is_large && !info.x.is_pinned);
+      bool commit_zero = false;
+      if (!_mi_mem_commit(p, blocks * MI_SEGMENT_SIZE, &commit_zero, tld)) {
+        // failed to commit! unclaim and return
+        mi_bitmap_unclaim(&region->in_use, 1, blocks, bit_idx);
+        return NULL;
+      }
+      if (commit_zero) *is_zero = true;
+    }
+  }
+  else {
+    // no need to commit, but check if already fully committed
+    *commit = _mi_bitmap_is_claimed(&region->commit, 1, blocks, bit_idx);
+  }
+  mi_assert_internal(!*commit || _mi_bitmap_is_claimed(&region->commit, 1, blocks, bit_idx));
+
+  // unreset reset blocks
+  if (_mi_bitmap_is_any_claimed(&region->reset, 1, blocks, bit_idx)) {
+    // some blocks are still reset
+    mi_assert_internal(!info.x.is_large && !info.x.is_pinned);
+    mi_assert_internal(!mi_option_is_enabled(mi_option_eager_commit) || *commit || mi_option_get(mi_option_eager_commit_delay) > 0);
+    mi_bitmap_unclaim(&region->reset, 1, blocks, bit_idx);
+    if (*commit || !mi_option_is_enabled(mi_option_reset_decommits)) { // only if needed
+      bool reset_zero = false;
+      _mi_mem_unreset(p, blocks * MI_SEGMENT_SIZE, &reset_zero, tld);
+      if (reset_zero) *is_zero = true;
+    }
+  }
+  mi_assert_internal(!_mi_bitmap_is_any_claimed(&region->reset, 1, blocks, bit_idx));
+
+  #if (MI_DEBUG>=2) && !MI_TRACK_ENABLED  // && !MI_TSAN
+  if (*commit) { ((uint8_t*)p)[0] = 0; }
+  #endif
+
+  // and return the allocation
+  mi_assert_internal(p != NULL);
+  return p;
+}
+
+
+/* ----------------------------------------------------------------------------
+ Allocation
+-----------------------------------------------------------------------------*/
+
+// Allocate `size` memory aligned at `alignment`. Return non NULL on success, with a given memory `id`.
+// (`id` is abstract, but `id = idx*MI_REGION_MAP_BITS + bitidx`)
+void* _mi_mem_alloc_aligned(size_t size, size_t alignment, size_t align_offset, bool* commit, bool* large, bool* is_pinned, bool* is_zero, size_t* memid, mi_os_tld_t* tld)
+{
+  mi_assert_internal(memid != NULL && tld != NULL);
+  mi_assert_internal(size > 0);
+  *memid = 0;
+  *is_zero = false;
+  *is_pinned = false;
+  bool default_large = false;
+  if (large==NULL) large = &default_large;  // ensure `large != NULL`
+  if (size == 0) return NULL;
+  size = _mi_align_up(size, _mi_os_page_size());
+
+  // allocate from regions if possible
+  void* p = NULL;
+  size_t arena_memid;
+  const size_t blocks = mi_region_block_count(size);
+  if (blocks <= MI_REGION_MAX_OBJ_BLOCKS && alignment <= MI_SEGMENT_ALIGN && align_offset == 0) {
+    p = mi_region_try_alloc(blocks, commit, large, is_pinned, is_zero, memid, tld);
+    if (p == NULL) {
+      _mi_warning_message("unable to allocate from region: size %zu\n", size);
+    }
+  }
+  if (p == NULL) {
+    // and otherwise fall back to the OS
+    p = _mi_arena_alloc_aligned(size, alignment, align_offset, commit, large, is_pinned, is_zero, _mi_arena_id_none(),  & arena_memid, tld);
+    *memid = mi_memid_create_from_arena(arena_memid);
+  }
+
+  if (p != NULL) {
+    mi_assert_internal(((uintptr_t)p + align_offset) % alignment == 0);
+    #if (MI_DEBUG>=2) && !MI_TRACK_ENABLED  // && !MI_TSAN
+    if (*commit) { ((uint8_t*)p)[0] = 0; } // ensure the memory is committed
+    #endif
+  }
+  return p;
+}
+
+
+
+/* ----------------------------------------------------------------------------
+Free
+-----------------------------------------------------------------------------*/
+
+// Free previously allocated memory with a given id.
+void _mi_mem_free(void* p, size_t size, size_t alignment, size_t align_offset, size_t id, bool full_commit, bool any_reset, mi_os_tld_t* tld) {
+  mi_assert_internal(size > 0 && tld != NULL);
+  if (p==NULL) return;
+  if (size==0) return;
+  size = _mi_align_up(size, _mi_os_page_size());
+
+  size_t arena_memid = 0;
+  mi_bitmap_index_t bit_idx;
+  mem_region_t* region;
+  if (mi_memid_is_arena(id,&region,&bit_idx,&arena_memid)) {
+   // was a direct arena allocation, pass through
+    _mi_arena_free(p, size, alignment, align_offset, arena_memid, full_commit, tld->stats);
+  }
+  else {
+    // allocated in a region
+    mi_assert_internal(align_offset == 0);
+    mi_assert_internal(size <= MI_REGION_MAX_OBJ_SIZE); if (size > MI_REGION_MAX_OBJ_SIZE) return;
+    const size_t blocks = mi_region_block_count(size);
+    mi_assert_internal(blocks + bit_idx <= MI_BITMAP_FIELD_BITS);
+    mi_region_info_t info;
+    info.value = mi_atomic_load_acquire(&region->info);
+    mi_assert_internal(info.value != 0);
+    void* blocks_start = mi_region_blocks_start(region, bit_idx);
+    mi_assert_internal(blocks_start == p); // not a pointer in our area?
+    mi_assert_internal(bit_idx + blocks <= MI_BITMAP_FIELD_BITS);
+    if (blocks_start != p || bit_idx + blocks > MI_BITMAP_FIELD_BITS) return; // or `abort`?
+
+    // committed?
+    if (full_commit && (size % MI_SEGMENT_SIZE) == 0) {
+      _mi_bitmap_claim(&region->commit, 1, blocks, bit_idx, NULL);
+    }
+
+    if (any_reset) {
+      // set the is_reset bits if any pages were reset
+      _mi_bitmap_claim(&region->reset, 1, blocks, bit_idx, NULL);
+    }
+
+    // reset the blocks to reduce the working set.
+    if (!info.x.is_large && !info.x.is_pinned && mi_option_is_enabled(mi_option_segment_reset)
+       && (mi_option_is_enabled(mi_option_eager_commit) ||
+           mi_option_is_enabled(mi_option_reset_decommits))) // cannot reset halfway committed segments, use only `option_page_reset` instead
+    {
+      bool any_unreset;
+      _mi_bitmap_claim(&region->reset, 1, blocks, bit_idx, &any_unreset);
+      if (any_unreset) {
+        _mi_abandoned_await_readers(); // ensure no more pending write (in case reset = decommit)
+        _mi_mem_reset(p, blocks * MI_SEGMENT_SIZE, tld);
+      }
+    }
+
+    // and unclaim
+    bool all_unclaimed = mi_bitmap_unclaim(&region->in_use, 1, blocks, bit_idx);
+    mi_assert_internal(all_unclaimed); MI_UNUSED(all_unclaimed);
+  }
+}
+
+
+/* ----------------------------------------------------------------------------
+  collection
+-----------------------------------------------------------------------------*/
+void _mi_mem_collect(mi_os_tld_t* tld) {
+  // free every region that has no segments in use.
+  size_t rcount = mi_atomic_load_relaxed(&regions_count);
+  for (size_t i = 0; i < rcount; i++) {
+    mem_region_t* region = &regions[i];
+    if (mi_atomic_load_relaxed(&region->info) != 0) {
+      // if no segments used, try to claim the whole region
+      size_t m = mi_atomic_load_relaxed(&region->in_use);
+      while (m == 0 && !mi_atomic_cas_weak_release(&region->in_use, &m, MI_BITMAP_FIELD_FULL)) { /* nothing */ };
+      if (m == 0) {
+        // on success, free the whole region
+        uint8_t* start = (uint8_t*)mi_atomic_load_ptr_acquire(uint8_t,&regions[i].start);
+        size_t arena_memid = mi_atomic_load_relaxed(&regions[i].arena_memid);
+        size_t commit = mi_atomic_load_relaxed(&regions[i].commit);
+        memset((void*)&regions[i], 0, sizeof(mem_region_t));  // cast to void* to avoid atomic warning
+        // and release the whole region
+        mi_atomic_store_release(&region->info, (size_t)0);
+        if (start != NULL) { // && !_mi_os_is_huge_reserved(start)) {
+          _mi_abandoned_await_readers(); // ensure no pending reads
+          _mi_arena_free(start, MI_REGION_SIZE, MI_SEGMENT_ALIGN, 0, arena_memid, (~commit == 0), tld->stats);
+        }
+      }
+    }
+  }
+}
+
+
+/* ----------------------------------------------------------------------------
+  Other
+-----------------------------------------------------------------------------*/
+
+bool _mi_mem_reset(void* p, size_t size, mi_os_tld_t* tld) {
+  if (mi_option_is_enabled(mi_option_reset_decommits)) {
+    return _mi_os_decommit(p, size, tld->stats);
+  }
+  else {
+    return _mi_os_reset(p, size, tld->stats);
+  }
+}
+
+bool _mi_mem_unreset(void* p, size_t size, bool* is_zero, mi_os_tld_t* tld) {
+  if (mi_option_is_enabled(mi_option_reset_decommits)) {
+    return _mi_os_commit(p, size, is_zero, tld->stats);
+  }
+  else {
+    return _mi_os_unreset(p, size, is_zero, tld->stats);
+  }
+}
+
+bool _mi_mem_commit(void* p, size_t size, bool* is_zero, mi_os_tld_t* tld) {
+  return _mi_os_commit(p, size, is_zero, tld->stats);
+}
+
+bool _mi_mem_decommit(void* p, size_t size, mi_os_tld_t* tld) {
+  return _mi_os_decommit(p, size, tld->stats);
+}
+
+bool _mi_mem_protect(void* p, size_t size) {
+  return _mi_os_protect(p, size);
+}
+
+bool _mi_mem_unprotect(void* p, size_t size) {
+  return _mi_os_unprotect(p, size);
+}

preload-mimalloc/mimalloc/src/segment-cache.c

@@ -0,0 +1,423 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+
+/* ----------------------------------------------------------------------------
+  Implements a cache of segments to avoid expensive OS calls and to reuse
+  the commit_mask to optimize the commit/decommit calls.
+  The full memory map of all segments is also implemented here.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+
+#include "./bitmap.h"  // atomic bitmap
+
+//#define MI_CACHE_DISABLE 1    // define to completely disable the segment cache
+
+#define MI_CACHE_FIELDS     (16)
+#define MI_CACHE_MAX        (MI_BITMAP_FIELD_BITS*MI_CACHE_FIELDS)       // 1024 on 64-bit
+
+#define BITS_SET()          MI_ATOMIC_VAR_INIT(UINTPTR_MAX)
+#define MI_CACHE_BITS_SET   MI_INIT16(BITS_SET)                          // note: update if MI_CACHE_FIELDS changes
+
+typedef struct mi_cache_slot_s {
+  void*               p;
+  size_t              memid;
+  bool                is_pinned;
+  mi_commit_mask_t    commit_mask;
+  mi_commit_mask_t    decommit_mask;
+  _Atomic(mi_msecs_t) expire;
+} mi_cache_slot_t;
+
+static mi_decl_cache_align mi_cache_slot_t cache[MI_CACHE_MAX];    // = 0
+
+static mi_decl_cache_align mi_bitmap_field_t cache_unavailable[MI_CACHE_FIELDS] = { MI_CACHE_BITS_SET };        // zero bit = available!
+static mi_decl_cache_align mi_bitmap_field_t cache_unavailable_large[MI_CACHE_FIELDS] = { MI_CACHE_BITS_SET };
+static mi_decl_cache_align mi_bitmap_field_t cache_inuse[MI_CACHE_FIELDS];   // zero bit = free
+
+static bool mi_cdecl mi_segment_cache_is_suitable(mi_bitmap_index_t bitidx, void* arg) {
+  mi_arena_id_t req_arena_id = *((mi_arena_id_t*)arg);
+  mi_cache_slot_t* slot = &cache[mi_bitmap_index_bit(bitidx)];
+  return _mi_arena_memid_is_suitable(slot->memid, req_arena_id);
+}
+
+mi_decl_noinline static void* mi_segment_cache_pop_ex(
+                              bool all_suitable,
+                              size_t size, mi_commit_mask_t* commit_mask, 
+                              mi_commit_mask_t* decommit_mask, bool large_allowed,
+                              bool* large, bool* is_pinned, bool* is_zero, 
+                              mi_arena_id_t _req_arena_id, size_t* memid, mi_os_tld_t* tld)
+{
+#ifdef MI_CACHE_DISABLE
+  return NULL;
+#else
+
+  // only segment blocks
+  if (size != MI_SEGMENT_SIZE) return NULL;
+
+  // numa node determines start field
+  const int numa_node = _mi_os_numa_node(tld);
+  size_t start_field = 0;
+  if (numa_node > 0) {
+    start_field = (MI_CACHE_FIELDS / _mi_os_numa_node_count())*numa_node;
+    if (start_field >= MI_CACHE_FIELDS) start_field = 0;
+  }
+
+  // find an available slot and make it unavailable
+  mi_bitmap_index_t bitidx = 0;
+  bool claimed = false;
+  mi_arena_id_t req_arena_id = _req_arena_id;
+  mi_bitmap_pred_fun_t pred_fun = (all_suitable ? NULL : &mi_segment_cache_is_suitable);  // cannot pass NULL as the arena may be exclusive itself; todo: do not put exclusive arenas in the cache?
+
+  if (large_allowed) {  // large allowed?
+    claimed = _mi_bitmap_try_find_from_claim_pred(cache_unavailable_large, MI_CACHE_FIELDS, start_field, 1, pred_fun, &req_arena_id, &bitidx);
+    if (claimed) *large = true;
+  }
+  if (!claimed) {
+    claimed = _mi_bitmap_try_find_from_claim_pred (cache_unavailable, MI_CACHE_FIELDS, start_field, 1, pred_fun, &req_arena_id, &bitidx);
+    if (claimed) *large = false;
+  }
+
+  if (!claimed) return NULL;
+
+  // no longer available but still in-use
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx));
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable_large, MI_CACHE_FIELDS, 1, bitidx));
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_inuse, MI_CACHE_FIELDS, 1, bitidx));
+
+  // found a slot
+  mi_cache_slot_t* slot = &cache[mi_bitmap_index_bit(bitidx)];
+  void* p = slot->p;
+  *memid = slot->memid;
+  *is_pinned = slot->is_pinned;
+  *is_zero = false;
+  *commit_mask = slot->commit_mask;     
+  *decommit_mask = slot->decommit_mask;
+  slot->p = NULL;
+  mi_atomic_storei64_release(&slot->expire,(mi_msecs_t)0);
+  
+  // mark the slot as free again
+  _mi_bitmap_unclaim(cache_inuse, MI_CACHE_FIELDS, 1, bitidx);
+  return p;
+#endif
+}
+
+
+mi_decl_noinline void* _mi_segment_cache_pop(size_t size, mi_commit_mask_t* commit_mask, mi_commit_mask_t* decommit_mask, bool large_allowed, bool* large, bool* is_pinned, bool* is_zero, mi_arena_id_t _req_arena_id, size_t* memid, mi_os_tld_t* tld)
+{
+  return mi_segment_cache_pop_ex(false, size, commit_mask, decommit_mask, large_allowed, large, is_pinned, is_zero, _req_arena_id, memid, tld);
+}
+
+static mi_decl_noinline void mi_commit_mask_decommit(mi_commit_mask_t* cmask, void* p, size_t total, mi_stats_t* stats)
+{
+  if (mi_commit_mask_is_empty(cmask)) {
+    // nothing
+  }
+  else if (mi_commit_mask_is_full(cmask)) {
+    // decommit the whole in one call
+    _mi_os_decommit(p, total, stats);
+  }
+  else {
+    // decommit parts
+    mi_assert_internal((total%MI_COMMIT_MASK_BITS)==0);
+    size_t part = total/MI_COMMIT_MASK_BITS;
+    size_t idx;
+    size_t count;    
+    mi_commit_mask_foreach(cmask, idx, count) {
+      void*  start = (uint8_t*)p + (idx*part);
+      size_t size = count*part;
+      _mi_os_decommit(start, size, stats);
+    }
+    mi_commit_mask_foreach_end()
+  }
+  mi_commit_mask_create_empty(cmask);
+}
+
+#define MI_MAX_PURGE_PER_PUSH  (4)
+
+static mi_decl_noinline void mi_segment_cache_purge(bool visit_all, bool force, mi_os_tld_t* tld)
+{
+  MI_UNUSED(tld);
+  if (!mi_option_is_enabled(mi_option_allow_decommit)) return;
+  mi_msecs_t now = _mi_clock_now();
+  size_t purged = 0;
+  const size_t max_visits = (visit_all ? MI_CACHE_MAX /* visit all */ : MI_CACHE_FIELDS /* probe at most N (=16) slots */);
+  size_t idx              = (visit_all ? 0 : _mi_random_shuffle((uintptr_t)now) % MI_CACHE_MAX /* random start */ );
+  for (size_t visited = 0; visited < max_visits; visited++,idx++) {  // visit N slots
+    if (idx >= MI_CACHE_MAX) idx = 0; // wrap
+    mi_cache_slot_t* slot = &cache[idx];
+    mi_msecs_t expire = mi_atomic_loadi64_relaxed(&slot->expire);
+    if (expire != 0 && (force || now >= expire)) {  // racy read
+      // seems expired, first claim it from available
+      purged++;
+      mi_bitmap_index_t bitidx = mi_bitmap_index_create_from_bit(idx);      
+      if (_mi_bitmap_claim(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx, NULL)) {  // no need to check large as those cannot be decommitted anyways
+        // it was available, we claimed it (and made it unavailable)
+        mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx));
+        mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable_large, MI_CACHE_FIELDS, 1, bitidx));
+        // we can now access it safely
+        expire = mi_atomic_loadi64_acquire(&slot->expire);
+        if (expire != 0 && (force || now >= expire)) {  // safe read
+          mi_assert_internal(_mi_bitmap_is_claimed(cache_inuse, MI_CACHE_FIELDS, 1, bitidx));
+          // still expired, decommit it
+          mi_atomic_storei64_relaxed(&slot->expire,(mi_msecs_t)0);
+          mi_assert_internal(!mi_commit_mask_is_empty(&slot->commit_mask));
+          _mi_abandoned_await_readers();  // wait until safe to decommit
+          // decommit committed parts
+          // TODO: instead of decommit, we could also free to the OS?
+          mi_commit_mask_decommit(&slot->commit_mask, slot->p, MI_SEGMENT_SIZE, tld->stats);
+          mi_commit_mask_create_empty(&slot->decommit_mask);
+        }
+        _mi_bitmap_unclaim(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx); // make it available again for a pop
+      }
+      if (!visit_all && purged > MI_MAX_PURGE_PER_PUSH) break;  // bound to no more than N purge tries per push
+    }
+  }
+}
+
+void _mi_segment_cache_collect(bool force, mi_os_tld_t* tld) {
+  if (force) {
+    // called on `mi_collect(true)` but not on thread termination    
+    _mi_segment_cache_free_all(tld);
+  }
+  else {
+    mi_segment_cache_purge(true /* visit all */, false /* don't force unexpired */, tld);
+  }
+}
+
+void _mi_segment_cache_free_all(mi_os_tld_t* tld) {
+  mi_commit_mask_t commit_mask;
+  mi_commit_mask_t decommit_mask;
+  bool is_pinned;
+  bool is_zero;
+  bool is_large;
+  size_t memid;
+  const size_t size = MI_SEGMENT_SIZE;
+  void* p;
+  do {
+    // keep popping and freeing the memory
+    p = mi_segment_cache_pop_ex(true /* all */, size, &commit_mask, &decommit_mask,
+                                true /* allow large */, &is_large, &is_pinned, &is_zero, _mi_arena_id_none(), &memid, tld);
+    if (p != NULL) {
+      size_t csize = _mi_commit_mask_committed_size(&commit_mask, size);
+      if (csize > 0 && !is_pinned) { _mi_stat_decrease(&_mi_stats_main.committed, csize); }
+      _mi_arena_free(p, size, MI_SEGMENT_ALIGN, 0, memid, is_pinned /* pretend not committed to not double count decommits */, tld->stats);
+    }
+  } while (p != NULL);  
+}
+
+mi_decl_noinline bool _mi_segment_cache_push(void* start, size_t size, size_t memid, const mi_commit_mask_t* commit_mask, const mi_commit_mask_t* decommit_mask, bool is_large, bool is_pinned, mi_os_tld_t* tld)
+{
+#ifdef MI_CACHE_DISABLE
+  return false;
+#else
+
+  // purge expired entries
+  mi_segment_cache_purge(false /* limit purges to a constant N */, false /* don't force unexpired */, tld);
+
+  // only cache normal segment blocks
+  if (size != MI_SEGMENT_SIZE || ((uintptr_t)start % MI_SEGMENT_ALIGN) != 0) return false;
+
+  // Also do not cache arena allocated segments that cannot be decommitted. (as arena allocation is fast)
+  // This is a common case with reserved huge OS pages.
+  // 
+  // (note: we could also allow segments that are already fully decommitted but that never happens
+  //  as the first slice is always committed (for the segment metadata))
+  if (!_mi_arena_is_os_allocated(memid) && is_pinned) return false;
+
+  // numa node determines start field
+  int numa_node = _mi_os_numa_node(NULL);
+  size_t start_field = 0;
+  if (numa_node > 0) {
+    start_field = (MI_CACHE_FIELDS / _mi_os_numa_node_count()) * numa_node;
+    if (start_field >= MI_CACHE_FIELDS) start_field = 0;
+  }
+
+  // find an available slot
+  mi_bitmap_index_t bitidx;
+  bool claimed = _mi_bitmap_try_find_from_claim(cache_inuse, MI_CACHE_FIELDS, start_field, 1, &bitidx);
+  if (!claimed) return false;
+
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable, MI_CACHE_FIELDS, 1, bitidx));
+  mi_assert_internal(_mi_bitmap_is_claimed(cache_unavailable_large, MI_CACHE_FIELDS, 1, bitidx));
+#if MI_DEBUG>1
+  if (is_pinned || is_large) {
+    mi_assert_internal(mi_commit_mask_is_full(commit_mask));
+  }
+#endif
+
+  // set the slot
+  mi_cache_slot_t* slot = &cache[mi_bitmap_index_bit(bitidx)];
+  slot->p = start;
+  slot->memid = memid;
+  slot->is_pinned = is_pinned;
+  mi_atomic_storei64_relaxed(&slot->expire,(mi_msecs_t)0);
+  slot->commit_mask = *commit_mask;
+  slot->decommit_mask = *decommit_mask;
+  if (!mi_commit_mask_is_empty(commit_mask) && !is_large && !is_pinned && mi_option_is_enabled(mi_option_allow_decommit)) {
+    long delay = mi_option_get(mi_option_segment_decommit_delay);
+    if (delay == 0) {
+      _mi_abandoned_await_readers(); // wait until safe to decommit
+      mi_commit_mask_decommit(&slot->commit_mask, start, MI_SEGMENT_SIZE, tld->stats);
+      mi_commit_mask_create_empty(&slot->decommit_mask);
+    }
+    else {
+      mi_atomic_storei64_release(&slot->expire, _mi_clock_now() + delay);
+    }
+  }
+
+  // make it available
+  _mi_bitmap_unclaim((is_large ? cache_unavailable_large : cache_unavailable), MI_CACHE_FIELDS, 1, bitidx);
+  return true;
+#endif
+}
+
+
+/* -----------------------------------------------------------
+  The following functions are to reliably find the segment or
+  block that encompasses any pointer p (or NULL if it is not
+  in any of our segments).
+  We maintain a bitmap of all memory with 1 bit per MI_SEGMENT_SIZE (64MiB)
+  set to 1 if it contains the segment meta data.
+----------------------------------------------------------- */
+
+
+#if (MI_INTPTR_SIZE==8)
+#define MI_MAX_ADDRESS    ((size_t)40 << 40)  // 20TB
+#else
+#define MI_MAX_ADDRESS    ((size_t)2 << 30)   // 2Gb
+#endif
+
+#define MI_SEGMENT_MAP_BITS  (MI_MAX_ADDRESS / MI_SEGMENT_SIZE)
+#define MI_SEGMENT_MAP_SIZE  (MI_SEGMENT_MAP_BITS / 8)
+#define MI_SEGMENT_MAP_WSIZE (MI_SEGMENT_MAP_SIZE / MI_INTPTR_SIZE)
+
+static _Atomic(uintptr_t) mi_segment_map[MI_SEGMENT_MAP_WSIZE + 1];  // 2KiB per TB with 64MiB segments
+
+static size_t mi_segment_map_index_of(const mi_segment_t* segment, size_t* bitidx) {
+  mi_assert_internal(_mi_ptr_segment(segment + 1) == segment); // is it aligned on MI_SEGMENT_SIZE?
+  if ((uintptr_t)segment >= MI_MAX_ADDRESS) {
+    *bitidx = 0;
+    return MI_SEGMENT_MAP_WSIZE;
+  }
+  else {
+    const uintptr_t segindex = ((uintptr_t)segment) / MI_SEGMENT_SIZE;
+    *bitidx = segindex % MI_INTPTR_BITS;
+    const size_t mapindex = segindex / MI_INTPTR_BITS;
+    mi_assert_internal(mapindex < MI_SEGMENT_MAP_WSIZE);
+    return mapindex;
+  }
+}
+
+void _mi_segment_map_allocated_at(const mi_segment_t* segment) {
+  size_t bitidx;
+  size_t index = mi_segment_map_index_of(segment, &bitidx);
+  mi_assert_internal(index <= MI_SEGMENT_MAP_WSIZE);
+  if (index==MI_SEGMENT_MAP_WSIZE) return;
+  uintptr_t mask = mi_atomic_load_relaxed(&mi_segment_map[index]);
+  uintptr_t newmask;
+  do {
+    newmask = (mask | ((uintptr_t)1 << bitidx));
+  } while (!mi_atomic_cas_weak_release(&mi_segment_map[index], &mask, newmask));
+}
+
+void _mi_segment_map_freed_at(const mi_segment_t* segment) {
+  size_t bitidx;
+  size_t index = mi_segment_map_index_of(segment, &bitidx);
+  mi_assert_internal(index <= MI_SEGMENT_MAP_WSIZE);
+  if (index == MI_SEGMENT_MAP_WSIZE) return;
+  uintptr_t mask = mi_atomic_load_relaxed(&mi_segment_map[index]);
+  uintptr_t newmask;
+  do {
+    newmask = (mask & ~((uintptr_t)1 << bitidx));
+  } while (!mi_atomic_cas_weak_release(&mi_segment_map[index], &mask, newmask));
+}
+
+// Determine the segment belonging to a pointer or NULL if it is not in a valid segment.
+static mi_segment_t* _mi_segment_of(const void* p) {
+  if (p == NULL) return NULL;
+  mi_segment_t* segment = _mi_ptr_segment(p);
+  mi_assert_internal(segment != NULL);
+  size_t bitidx;
+  size_t index = mi_segment_map_index_of(segment, &bitidx);
+  // fast path: for any pointer to valid small/medium/large object or first MI_SEGMENT_SIZE in huge
+  const uintptr_t mask = mi_atomic_load_relaxed(&mi_segment_map[index]);
+  if mi_likely((mask & ((uintptr_t)1 << bitidx)) != 0) {
+    return segment; // yes, allocated by us
+  }
+  if (index==MI_SEGMENT_MAP_WSIZE) return NULL;
+
+  // TODO: maintain max/min allocated range for efficiency for more efficient rejection of invalid pointers?
+
+  // search downwards for the first segment in case it is an interior pointer
+  // could be slow but searches in MI_INTPTR_SIZE * MI_SEGMENT_SIZE (512MiB) steps trough
+  // valid huge objects
+  // note: we could maintain a lowest index to speed up the path for invalid pointers?
+  size_t lobitidx;
+  size_t loindex;
+  uintptr_t lobits = mask & (((uintptr_t)1 << bitidx) - 1);
+  if (lobits != 0) {
+    loindex = index;
+    lobitidx = mi_bsr(lobits);    // lobits != 0
+  }
+  else if (index == 0) {
+    return NULL;
+  }
+  else {
+    mi_assert_internal(index > 0);
+    uintptr_t lomask = mask;
+    loindex = index;
+    do {
+      loindex--;  
+      lomask = mi_atomic_load_relaxed(&mi_segment_map[loindex]);      
+    } while (lomask != 0 && loindex > 0);
+    if (lomask == 0) return NULL;
+    lobitidx = mi_bsr(lomask);    // lomask != 0
+  }
+  mi_assert_internal(loindex < MI_SEGMENT_MAP_WSIZE);
+  // take difference as the addresses could be larger than the MAX_ADDRESS space.
+  size_t diff = (((index - loindex) * (8*MI_INTPTR_SIZE)) + bitidx - lobitidx) * MI_SEGMENT_SIZE;
+  segment = (mi_segment_t*)((uint8_t*)segment - diff);
+
+  if (segment == NULL) return NULL;
+  mi_assert_internal((void*)segment < p);
+  bool cookie_ok = (_mi_ptr_cookie(segment) == segment->cookie);
+  mi_assert_internal(cookie_ok);
+  if mi_unlikely(!cookie_ok) return NULL;
+  if (((uint8_t*)segment + mi_segment_size(segment)) <= (uint8_t*)p) return NULL; // outside the range
+  mi_assert_internal(p >= (void*)segment && (uint8_t*)p < (uint8_t*)segment + mi_segment_size(segment));
+  return segment;
+}
+
+// Is this a valid pointer in our heap?
+static bool  mi_is_valid_pointer(const void* p) {
+  return (_mi_segment_of(p) != NULL);
+}
+
+mi_decl_nodiscard mi_decl_export bool mi_is_in_heap_region(const void* p) mi_attr_noexcept {
+  return mi_is_valid_pointer(p);
+}
+
+/*
+// Return the full segment range belonging to a pointer
+static void* mi_segment_range_of(const void* p, size_t* size) {
+  mi_segment_t* segment = _mi_segment_of(p);
+  if (segment == NULL) {
+    if (size != NULL) *size = 0;
+    return NULL;
+  }
+  else {
+    if (size != NULL) *size = segment->segment_size;
+    return segment;
+  }
+  mi_assert_expensive(page == NULL || mi_segment_is_valid(_mi_page_segment(page),tld));
+  mi_assert_internal(page == NULL || (mi_segment_page_size(_mi_page_segment(page)) - (MI_SECURE == 0 ? 0 : _mi_os_page_size())) >= block_size);
+  mi_reset_delayed(tld);
+  mi_assert_internal(page == NULL || mi_page_not_in_queue(page, tld));
+  return page;
+}
+*/

preload-mimalloc/mimalloc/src/static.c

@@ -0,0 +1,40 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2020, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#ifndef _DEFAULT_SOURCE
+#define _DEFAULT_SOURCE
+#endif
+#if defined(__sun)
+// same remarks as os.c for the static's context.
+#undef _XOPEN_SOURCE
+#undef _POSIX_C_SOURCE
+#endif
+
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+
+// For a static override we create a single object file
+// containing the whole library. If it is linked first
+// it will override all the standard library allocation
+// functions (on Unix's).
+#include "alloc.c"          // includes alloc-override.c
+#include "alloc-aligned.c"
+#include "alloc-posix.c"
+#include "arena.c"
+#include "bitmap.c"
+#include "heap.c"
+#include "init.c"
+#include "options.c"
+#include "os.c"
+#include "page.c"           // includes page-queue.c
+#include "random.c" 
+#include "segment.c"
+#include "segment-cache.c"
+#include "stats.c"
+#include "prim/prim.c"
+#if MI_OSX_ZONE
+#include "prim/osx/alloc-override-zone.c"
+#endif

preload-mimalloc/mimalloc/src/stats.c

@@ -0,0 +1,454 @@
+/* ----------------------------------------------------------------------------
+Copyright (c) 2018-2021, Microsoft Research, Daan Leijen
+This is free software; you can redistribute it and/or modify it under the
+terms of the MIT license. A copy of the license can be found in the file
+"LICENSE" at the root of this distribution.
+-----------------------------------------------------------------------------*/
+#include "mimalloc.h"
+#include "mimalloc/internal.h"
+#include "mimalloc/atomic.h"
+#include "mimalloc/prim.h"
+
+#include <stdio.h>  // snprintf
+#include <string.h> // memset
+
+#if defined(_MSC_VER) && (_MSC_VER < 1920)
+#pragma warning(disable:4204)  // non-constant aggregate initializer
+#endif
+
+/* -----------------------------------------------------------
+  Statistics operations
+----------------------------------------------------------- */
+
+static bool mi_is_in_main(void* stat) {
+  return ((uint8_t*)stat >= (uint8_t*)&_mi_stats_main
+         && (uint8_t*)stat < ((uint8_t*)&_mi_stats_main + sizeof(mi_stats_t)));
+}
+
+static void mi_stat_update(mi_stat_count_t* stat, int64_t amount) {
+  if (amount == 0) return;
+  if (mi_is_in_main(stat))
+  {
+    // add atomically (for abandoned pages)
+    int64_t current = mi_atomic_addi64_relaxed(&stat->current, amount);
+    mi_atomic_maxi64_relaxed(&stat->peak, current + amount);
+    if (amount > 0) {
+      mi_atomic_addi64_relaxed(&stat->allocated,amount);
+    }
+    else {
+      mi_atomic_addi64_relaxed(&stat->freed, -amount);
+    }
+  }
+  else {
+    // add thread local
+    stat->current += amount;
+    if (stat->current > stat->peak) stat->peak = stat->current;
+    if (amount > 0) {
+      stat->allocated += amount;
+    }
+    else {
+      stat->freed += -amount;
+    }
+  }
+}
+
+void _mi_stat_counter_increase(mi_stat_counter_t* stat, size_t amount) {
+  if (mi_is_in_main(stat)) {
+    mi_atomic_addi64_relaxed( &stat->count, 1 );
+    mi_atomic_addi64_relaxed( &stat->total, (int64_t)amount );
+  }
+  else {
+    stat->count++;
+    stat->total += amount;
+  }
+}
+
+void _mi_stat_increase(mi_stat_count_t* stat, size_t amount) {
+  mi_stat_update(stat, (int64_t)amount);
+}
+
+void _mi_stat_decrease(mi_stat_count_t* stat, size_t amount) {
+  mi_stat_update(stat, -((int64_t)amount));
+}
+
+// must be thread safe as it is called from stats_merge
+static void mi_stat_add(mi_stat_count_t* stat, const mi_stat_count_t* src, int64_t unit) {
+  if (stat==src) return;
+  if (src->allocated==0 && src->freed==0) return;
+  mi_atomic_addi64_relaxed( &stat->allocated, src->allocated * unit);
+  mi_atomic_addi64_relaxed( &stat->current, src->current * unit);
+  mi_atomic_addi64_relaxed( &stat->freed, src->freed * unit);
+  // peak scores do not work across threads..
+  mi_atomic_addi64_relaxed( &stat->peak, src->peak * unit);
+}
+
+static void mi_stat_counter_add(mi_stat_counter_t* stat, const mi_stat_counter_t* src, int64_t unit) {
+  if (stat==src) return;
+  mi_atomic_addi64_relaxed( &stat->total, src->total * unit);
+  mi_atomic_addi64_relaxed( &stat->count, src->count * unit);
+}
+
+// must be thread safe as it is called from stats_merge
+static void mi_stats_add(mi_stats_t* stats, const mi_stats_t* src) {
+  if (stats==src) return;
+  mi_stat_add(&stats->segments, &src->segments,1);
+  mi_stat_add(&stats->pages, &src->pages,1);
+  mi_stat_add(&stats->reserved, &src->reserved, 1);
+  mi_stat_add(&stats->committed, &src->committed, 1);
+  mi_stat_add(&stats->reset, &src->reset, 1);
+  mi_stat_add(&stats->page_committed, &src->page_committed, 1);
+
+  mi_stat_add(&stats->pages_abandoned, &src->pages_abandoned, 1);
+  mi_stat_add(&stats->segments_abandoned, &src->segments_abandoned, 1);
+  mi_stat_add(&stats->threads, &src->threads, 1);
+
+  mi_stat_add(&stats->malloc, &src->malloc, 1);
+  mi_stat_add(&stats->segments_cache, &src->segments_cache, 1);
+  mi_stat_add(&stats->normal, &src->normal, 1);
+  mi_stat_add(&stats->huge, &src->huge, 1);
+  mi_stat_add(&stats->large, &src->large, 1);
+
+  mi_stat_counter_add(&stats->pages_extended, &src->pages_extended, 1);
+  mi_stat_counter_add(&stats->mmap_calls, &src->mmap_calls, 1);
+  mi_stat_counter_add(&stats->commit_calls, &src->commit_calls, 1);
+
+  mi_stat_counter_add(&stats->page_no_retire, &src->page_no_retire, 1);
+  mi_stat_counter_add(&stats->searches, &src->searches, 1);
+  mi_stat_counter_add(&stats->normal_count, &src->normal_count, 1);
+  mi_stat_counter_add(&stats->huge_count, &src->huge_count, 1);
+  mi_stat_counter_add(&stats->large_count, &src->large_count, 1);
+#if MI_STAT>1
+  for (size_t i = 0; i <= MI_BIN_HUGE; i++) {
+    if (src->normal_bins[i].allocated > 0 || src->normal_bins[i].freed > 0) {
+      mi_stat_add(&stats->normal_bins[i], &src->normal_bins[i], 1);
+    }
+  }
+#endif
+}
+
+/* -----------------------------------------------------------
+  Display statistics
+----------------------------------------------------------- */
+
+// unit > 0 : size in binary bytes
+// unit == 0: count as decimal
+// unit < 0 : count in binary
+static void mi_printf_amount(int64_t n, int64_t unit, mi_output_fun* out, void* arg, const char* fmt) {
+  char buf[32]; buf[0] = 0;
+  int  len = 32;
+  const char* suffix = (unit <= 0 ? " " : "B");
+  const int64_t base = (unit == 0 ? 1000 : 1024);
+  if (unit>0) n *= unit;
+
+  const int64_t pos = (n < 0 ? -n : n);
+  if (pos < base) {
+    if (n!=1 || suffix[0] != 'B') {  // skip printing 1 B for the unit column
+      snprintf(buf, len, "%d %-3s", (int)n, (n==0 ? "" : suffix));
+    }
+  }
+  else {
+    int64_t divider = base;
+    const char* magnitude = "K";
+    if (pos >= divider*base) { divider *= base; magnitude = "M"; }
+    if (pos >= divider*base) { divider *= base; magnitude = "G"; }
+    const int64_t tens = (n / (divider/10));
+    const long whole = (long)(tens/10);
+    const long frac1 = (long)(tens%10);
+    char unitdesc[8];
+    snprintf(unitdesc, 8, "%s%s%s", magnitude, (base==1024 ? "i" : ""), suffix);
+    snprintf(buf, len, "%ld.%ld %-3s", whole, (frac1 < 0 ? -frac1 : frac1), unitdesc);
+  }
+  _mi_fprintf(out, arg, (fmt==NULL ? "%11s" : fmt), buf);
+}
+
+
+static void mi_print_amount(int64_t n, int64_t unit, mi_output_fun* out, void* arg) {
+  mi_printf_amount(n,unit,out,arg,NULL);
+}
+
+static void mi_print_count(int64_t n, int64_t unit, mi_output_fun* out, void* arg) {
+  if (unit==1) _mi_fprintf(out, arg, "%11s"," ");
+          else mi_print_amount(n,0,out,arg);
+}
+
+static void mi_stat_print_ex(const mi_stat_count_t* stat, const char* msg, int64_t unit, mi_output_fun* out, void* arg, const char* notok ) {
+  _mi_fprintf(out, arg,"%10s:", msg);
+  if (unit > 0) {
+    mi_print_amount(stat->peak, unit, out, arg);
+    mi_print_amount(stat->allocated, unit, out, arg);
+    mi_print_amount(stat->freed, unit, out, arg);
+    mi_print_amount(stat->current, unit, out, arg);
+    mi_print_amount(unit, 1, out, arg);
+    mi_print_count(stat->allocated, unit, out, arg);
+    if (stat->allocated > stat->freed) {
+      _mi_fprintf(out, arg, "  ");
+      _mi_fprintf(out, arg, (notok == NULL ? "not all freed!" : notok));
+      _mi_fprintf(out, arg, "\n");
+    }
+    else {
+      _mi_fprintf(out, arg, "  ok\n");
+    }
+  }
+  else if (unit<0) {
+    mi_print_amount(stat->peak, -1, out, arg);
+    mi_print_amount(stat->allocated, -1, out, arg);
+    mi_print_amount(stat->freed, -1, out, arg);
+    mi_print_amount(stat->current, -1, out, arg);
+    if (unit==-1) {
+      _mi_fprintf(out, arg, "%22s", "");
+    }
+    else {
+      mi_print_amount(-unit, 1, out, arg);
+      mi_print_count((stat->allocated / -unit), 0, out, arg);
+    }
+    if (stat->allocated > stat->freed)
+      _mi_fprintf(out, arg, "  not all freed!\n");
+    else
+      _mi_fprintf(out, arg, "  ok\n");
+  }
+  else {
+    mi_print_amount(stat->peak, 1, out, arg);
+    mi_print_amount(stat->allocated, 1, out, arg);
+    _mi_fprintf(out, arg, "%11s", " ");  // no freed
+    mi_print_amount(stat->current, 1, out, arg);
+    _mi_fprintf(out, arg, "\n");
+  }
+}
+
+static void mi_stat_print(const mi_stat_count_t* stat, const char* msg, int64_t unit, mi_output_fun* out, void* arg) {
+  mi_stat_print_ex(stat, msg, unit, out, arg, NULL);
+}
+
+static void mi_stat_counter_print(const mi_stat_counter_t* stat, const char* msg, mi_output_fun* out, void* arg ) {
+  _mi_fprintf(out, arg, "%10s:", msg);
+  mi_print_amount(stat->total, -1, out, arg);
+  _mi_fprintf(out, arg, "\n");
+}
+
+static void mi_stat_counter_print_avg(const mi_stat_counter_t* stat, const char* msg, mi_output_fun* out, void* arg) {
+  const int64_t avg_tens = (stat->count == 0 ? 0 : (stat->total*10 / stat->count));
+  const long avg_whole = (long)(avg_tens/10);
+  const long avg_frac1 = (long)(avg_tens%10);
+  _mi_fprintf(out, arg, "%10s: %5ld.%ld avg\n", msg, avg_whole, avg_frac1);
+}
+
+
+static void mi_print_header(mi_output_fun* out, void* arg ) {
+  _mi_fprintf(out, arg, "%10s: %10s %10s %10s %10s %10s %10s\n", "heap stats", "peak   ", "total   ", "freed   ", "current   ", "unit   ", "count   ");
+}
+
+#if MI_STAT>1
+static void mi_stats_print_bins(const mi_stat_count_t* bins, size_t max, const char* fmt, mi_output_fun* out, void* arg) {
+  bool found = false;
+  char buf[64];
+  for (size_t i = 0; i <= max; i++) {
+    if (bins[i].allocated > 0) {
+      found = true;
+      int64_t unit = _mi_bin_size((uint8_t)i);
+      snprintf(buf, 64, "%s %3lu", fmt, (long)i);
+      mi_stat_print(&bins[i], buf, unit, out, arg);
+    }
+  }
+  if (found) {
+    _mi_fprintf(out, arg, "\n");
+    mi_print_header(out, arg);
+  }
+}
+#endif
+
+
+
+//------------------------------------------------------------
+// Use an output wrapper for line-buffered output
+// (which is nice when using loggers etc.)
+//------------------------------------------------------------
+typedef struct buffered_s {
+  mi_output_fun* out;   // original output function
+  void*          arg;   // and state
+  char*          buf;   // local buffer of at least size `count+1`
+  size_t         used;  // currently used chars `used <= count`
+  size_t         count; // total chars available for output
+} buffered_t;
+
+static void mi_buffered_flush(buffered_t* buf) {
+  buf->buf[buf->used] = 0;
+  _mi_fputs(buf->out, buf->arg, NULL, buf->buf);
+  buf->used = 0;
+}
+
+static void mi_cdecl mi_buffered_out(const char* msg, void* arg) {
+  buffered_t* buf = (buffered_t*)arg;
+  if (msg==NULL || buf==NULL) return;
+  for (const char* src = msg; *src != 0; src++) {
+    char c = *src;
+    if (buf->used >= buf->count) mi_buffered_flush(buf);
+    mi_assert_internal(buf->used < buf->count);
+    buf->buf[buf->used++] = c;
+    if (c == '\n') mi_buffered_flush(buf);
+  }
+}
+
+//------------------------------------------------------------
+// Print statistics
+//------------------------------------------------------------
+
+static void _mi_stats_print(mi_stats_t* stats, mi_output_fun* out0, void* arg0) mi_attr_noexcept {
+  // wrap the output function to be line buffered
+  char buf[256];
+  buffered_t buffer = { out0, arg0, NULL, 0, 255 };
+  buffer.buf = buf;
+  mi_output_fun* out = &mi_buffered_out;
+  void* arg = &buffer;
+
+  // and print using that
+  mi_print_header(out,arg);
+  #if MI_STAT>1
+  mi_stats_print_bins(stats->normal_bins, MI_BIN_HUGE, "normal",out,arg);
+  #endif
+  #if MI_STAT
+  mi_stat_print(&stats->normal, "normal", (stats->normal_count.count == 0 ? 1 : -(stats->normal.allocated / stats->normal_count.count)), out, arg);
+  mi_stat_print(&stats->large, "large", (stats->large_count.count == 0 ? 1 : -(stats->large.allocated / stats->large_count.count)), out, arg);
+  mi_stat_print(&stats->huge, "huge", (stats->huge_count.count == 0 ? 1 : -(stats->huge.allocated / stats->huge_count.count)), out, arg);
+  mi_stat_count_t total = { 0,0,0,0 };
+  mi_stat_add(&total, &stats->normal, 1);
+  mi_stat_add(&total, &stats->large, 1);
+  mi_stat_add(&total, &stats->huge, 1);
+  mi_stat_print(&total, "total", 1, out, arg);
+  #endif
+  #if MI_STAT>1
+  mi_stat_print(&stats->malloc, "malloc req", 1, out, arg);
+  _mi_fprintf(out, arg, "\n");
+  #endif
+  mi_stat_print_ex(&stats->reserved, "reserved", 1, out, arg, "");
+  mi_stat_print_ex(&stats->committed, "committed", 1, out, arg, "");
+  mi_stat_print(&stats->reset, "reset", 1, out, arg);
+  mi_stat_print(&stats->page_committed, "touched", 1, out, arg);
+  mi_stat_print(&stats->segments, "segments", -1, out, arg);
+  mi_stat_print(&stats->segments_abandoned, "-abandoned", -1, out, arg);
+  mi_stat_print(&stats->segments_cache, "-cached", -1, out, arg);
+  mi_stat_print(&stats->pages, "pages", -1, out, arg);
+  mi_stat_print(&stats->pages_abandoned, "-abandoned", -1, out, arg);
+  mi_stat_counter_print(&stats->pages_extended, "-extended", out, arg);
+  mi_stat_counter_print(&stats->page_no_retire, "-noretire", out, arg);
+  mi_stat_counter_print(&stats->mmap_calls, "mmaps", out, arg);
+  mi_stat_counter_print(&stats->commit_calls, "commits", out, arg);
+  mi_stat_print(&stats->threads, "threads", -1, out, arg);
+  mi_stat_counter_print_avg(&stats->searches, "searches", out, arg);
+  _mi_fprintf(out, arg, "%10s: %7zu\n", "numa nodes", _mi_os_numa_node_count());
+
+  size_t elapsed;
+  size_t user_time;
+  size_t sys_time;
+  size_t current_rss;
+  size_t peak_rss;
+  size_t current_commit;
+  size_t peak_commit;
+  size_t page_faults;
+  mi_process_info(&elapsed, &user_time, &sys_time, &current_rss, &peak_rss, &current_commit, &peak_commit, &page_faults);
+  _mi_fprintf(out, arg, "%10s: %7ld.%03ld s\n", "elapsed", elapsed/1000, elapsed%1000);
+  _mi_fprintf(out, arg, "%10s: user: %ld.%03ld s, system: %ld.%03ld s, faults: %lu, rss: ", "process",
+              user_time/1000, user_time%1000, sys_time/1000, sys_time%1000, (unsigned long)page_faults );
+  mi_printf_amount((int64_t)peak_rss, 1, out, arg, "%s");
+  if (peak_commit > 0) {
+    _mi_fprintf(out, arg, ", commit: ");
+    mi_printf_amount((int64_t)peak_commit, 1, out, arg, "%s");
+  }
+  _mi_fprintf(out, arg, "\n");
+}
+
+static mi_msecs_t mi_process_start; // = 0
+
+static mi_stats_t* mi_stats_get_default(void) {
+  mi_heap_t* heap = mi_heap_get_default();
+  return &heap->tld->stats;
+}
+
+static void mi_stats_merge_from(mi_stats_t* stats) {
+  if (stats != &_mi_stats_main) {
+    mi_stats_add(&_mi_stats_main, stats);
+    memset(stats, 0, sizeof(mi_stats_t));
+  }
+}
+
+void mi_stats_reset(void) mi_attr_noexcept {
+  mi_stats_t* stats = mi_stats_get_default();
+  if (stats != &_mi_stats_main) { memset(stats, 0, sizeof(mi_stats_t)); }
+  memset(&_mi_stats_main, 0, sizeof(mi_stats_t));
+  if (mi_process_start == 0) { mi_process_start = _mi_clock_start(); };
+}
+
+void mi_stats_merge(void) mi_attr_noexcept {
+  mi_stats_merge_from( mi_stats_get_default() );
+}
+
+void _mi_stats_done(mi_stats_t* stats) {  // called from `mi_thread_done`
+  mi_stats_merge_from(stats);
+}
+
+void mi_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept {
+  mi_stats_merge_from(mi_stats_get_default());
+  _mi_stats_print(&_mi_stats_main, out, arg);
+}
+
+void mi_stats_print(void* out) mi_attr_noexcept {
+  // for compatibility there is an `out` parameter (which can be `stdout` or `stderr`)
+  mi_stats_print_out((mi_output_fun*)out, NULL);
+}
+
+void mi_thread_stats_print_out(mi_output_fun* out, void* arg) mi_attr_noexcept {
+  _mi_stats_print(mi_stats_get_default(), out, arg);
+}
+
+
+// ----------------------------------------------------------------
+// Basic timer for convenience; use milli-seconds to avoid doubles
+// ----------------------------------------------------------------
+
+static mi_msecs_t mi_clock_diff;
+
+mi_msecs_t _mi_clock_now(void) {
+  return _mi_prim_clock_now();
+}
+
+mi_msecs_t _mi_clock_start(void) {
+  if (mi_clock_diff == 0.0) {
+    mi_msecs_t t0 = _mi_clock_now();
+    mi_clock_diff = _mi_clock_now() - t0;
+  }
+  return _mi_clock_now();
+}
+
+mi_msecs_t _mi_clock_end(mi_msecs_t start) {
+  mi_msecs_t end = _mi_clock_now();
+  return (end - start - mi_clock_diff);
+}
+
+
+// --------------------------------------------------------
+// Basic process statistics
+// --------------------------------------------------------
+
+mi_decl_export void mi_process_info(size_t* elapsed_msecs, size_t* user_msecs, size_t* system_msecs, size_t* current_rss, size_t* peak_rss, size_t* current_commit, size_t* peak_commit, size_t* page_faults) mi_attr_noexcept
+{
+  mi_process_info_t pinfo;
+  _mi_memzero(&pinfo,sizeof(pinfo));
+  pinfo.elapsed        = _mi_clock_end(mi_process_start);
+  pinfo.current_commit = (size_t)(mi_atomic_loadi64_relaxed((_Atomic(int64_t)*)&_mi_stats_main.committed.current));
+  pinfo.peak_commit    = (size_t)(mi_atomic_loadi64_relaxed((_Atomic(int64_t)*)&_mi_stats_main.committed.peak));
+  pinfo.current_rss    = pinfo.current_commit;
+  pinfo.peak_rss       = pinfo.peak_commit;
+  pinfo.utime          = 0;
+  pinfo.stime          = 0;
+  pinfo.page_faults    = 0;
+
+  _mi_prim_process_info(&pinfo);
+  
+  if (elapsed_msecs!=NULL)  *elapsed_msecs  = (pinfo.elapsed < 0 ? 0 : (pinfo.elapsed < (mi_msecs_t)PTRDIFF_MAX ? (size_t)pinfo.elapsed : PTRDIFF_MAX));
+  if (user_msecs!=NULL)     *user_msecs     = (pinfo.utime < 0 ? 0 : (pinfo.utime < (mi_msecs_t)PTRDIFF_MAX ? (size_t)pinfo.utime : PTRDIFF_MAX));
+  if (system_msecs!=NULL)   *system_msecs   = (pinfo.stime < 0 ? 0 : (pinfo.stime < (mi_msecs_t)PTRDIFF_MAX ? (size_t)pinfo.stime : PTRDIFF_MAX));
+  if (current_rss!=NULL)    *current_rss    = pinfo.current_rss;
+  if (peak_rss!=NULL)       *peak_rss       = pinfo.peak_rss;
+  if (current_commit!=NULL) *current_commit = pinfo.current_commit;
+  if (peak_commit!=NULL)    *peak_commit    = pinfo.peak_commit;
+  if (page_faults!=NULL)    *page_faults    = pinfo.page_faults;
+}

preload-mimalloc/update-sources.sh

@@ -0,0 +1,24 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# You can use this to update the mimalloc source code vendored here, if you
+# wish to upgrade the version for instance. This should be idempotent.
+#
+# NOTE!: v2.1.2 regresses from our point of view. See:
+# https://github.com/microsoft/mimalloc/issues/776
+VERSION=v2.1.1
+ 
+
+THIS_DIR="$( cd "$( dirname "${BASH_SOURCE[0]}" )" >/dev/null 2>&1 && pwd )"   # ... https://stackoverflow.com/a/246128/176841
+cd "$THIS_DIR"
+
+rm -rf mimalloc
+git clone git@github.com:microsoft/mimalloc.git 2>/dev/null
+(
+cd mimalloc
+git checkout "$VERSION" 2>/dev/null
+# stuff we don't need:
+rm -rf .git docs doc cmake bin azure-pipelines.yml CMakeLists.txt .gitattributes .gitignore ide mimalloc.pc.in test
+) # in mimalloc/
+
+echo "Done. $VERSION vendored in 'mimalloc/'"

server/benchmarks/benchmark_sets/9447_repro/README.md

@@ -0,0 +1,13 @@
+This is a repro case for the bad memory usage behavior scenario described in
+https://github.com/hasura/graphql-engine-mono/pull/9447 .  see that thread for
+details.
+
+This is just here as a runnable example of the workload; the numbers we get
+back from the benchmark are not really useful, since e.g.  many of the queries
+actually timeout.
+
+One thing that wasn't fixed in #9823, and for which this is a good repro case:
+The server has no ability to detect when a client has disconnected in the
+middle of an ongoing query, for example because the client library timed out,
+and likewise does not cancel the corresponding long-running database query. 
+

server/benchmarks/benchmark_sets/9447_repro/config.query.yaml

@@ -0,0 +1,101 @@
+### See README.md
+#
+# This tells graphql-bench that it's testing a hasura instance and should
+# collect some additional metrics:
+extended_hasura_checks: true
+headers:
+  X-Hasura-Admin-Secret: my-secret
+  X-Hasura-Role: employee
+  X-Hasura-User-Id: 4
+  # On cloud about 95% of requests incoming Accept gzip (which as of writing
+  # means we unconditionally compress, which has significant performance
+  # impact), so it's important that we include it in all benchmarks unless we
+  # have an intentional reason not to.
+  Accept-Encoding: gzip
+
+# Anchors to help us DRY below; settings here may be overridden selectively
+constants:
+  scalars:
+    # We'll measure at just two consistent load levels, which makes comparing
+    # benchmarks within the same run useful.
+    #
+    # NOTE: a load of 500 may cause hasura to fall over on a laptop. On our
+    # beefy CI benchmark runner we cannot sustain 1,000 RPS for the
+    # "large_result" queries.
+    - &low_load 20
+    - &high_load 500
+
+  k6_custom: &k6_custom
+    tools: [k6]
+    execution_strategy: CUSTOM
+
+  settings: &settings
+    # This is equivalent to wrk2's approach:
+    executor: 'constant-arrival-rate'
+    timeUnit: '1s'
+    maxVUs: 500 # NOTE: required, else defaults to `preAllocatedVUs`
+    # NOTE: ideally we'd test all of the queries with the same *number of requests*
+    # but that would mean running the "low_load" queries for much longer than
+    # is acceptable.
+    duration: '60s'
+
+queries:
+  - name: very_slow_query
+    <<: *k6_custom
+    options:
+      k6:
+        # NOTE: setting this to true will ignore graphql-layer errors, which
+        # still return a 200 HTTP status code.  Unfortunately we need to set to
+        # `true` so that we aren't measuring e.g. decompression on the k6 side
+        # (see: https://github.com/grafana/k6/issues/2685)
+        discardResponseBodies: true
+        scenarios:
+          main:
+            <<: *settings
+            rate: *low_load
+            # tune this so it's just high enough that we can expect to not need
+            # to allocate during the test:
+            preAllocatedVUs: 10
+    query:
+       query LongQuery {
+         Track {
+           Name
+           Milliseconds
+           Album {
+             ArtistId
+             AlbumId
+             Title
+             Artist {
+               ArtistId
+               Name
+               Albums {
+                 AlbumId
+                 Title
+                 Tracks {
+                   AlbumId
+                   Bytes
+                   Composer
+                 }
+               }
+             }
+             Tracks {
+               InvoiceLines {
+                 Quantity
+                 InvoiceId
+               }
+               Genre {
+                 Name
+                 GenreId
+               }
+               PlaylistTracks {
+                 TrackId
+                 PlaylistId
+                 Playlist {
+                   Name
+                   PlaylistId
+                 }
+               }
+             }
+           }
+         }
+       }

server/benchmarks/benchmark_sets/9447_repro/dump.sql.gz

@@ -0,0 +1 @@
+../chinook/dump.sql.gz

server/benchmarks/benchmark_sets/9447_repro/replace_metadata.json

@@ -0,0 +1 @@
+../chinook/replace_metadata.json

server/graphql-engine.cabal

@@ -258,6 +258,16 @@ common common-exe
     -- workaround for limitations in monitoring tools than anything...
     "-with-rtsopts=-N -I0 -T -kc8K --disable-delayed-os-memory-return"
 
+
+-- See 'preload-mimalloc/README.md'
+-- We want this in all server binary executable sections
+common preload-mimalloc
+    c-sources:       ../preload-mimalloc/mimalloc/src/static.c
+    include-dirs:    ../preload-mimalloc/mimalloc/include
+    cc-options:      -DMI_MALLOC_OVERRIDE
+    cc-options: -DMI_STATIC_LIB
+    cc-options:  -O3 -DNDEBUG -fPIC   -Wall -Wextra -Wno-unknown-pragmas -fvisibility=hidden -Wstrict-prototypes -ftls-model=initial-exec -std=gnu11
+
 common lib-depends
   build-depends:       Spock-core
                      , aeson
@@ -1029,6 +1039,7 @@ library
 
 executable graphql-engine
   import: common-all, common-exe
+  import: preload-mimalloc
 
   if flag(optimize-hasura)
     ghc-options: