diff --git a/CMakeLists.txt b/CMakeLists.txt
index ae53f76b..1a1b6b02 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -346,6 +346,7 @@ target_sources(mold PRIVATE
   common/filepath.cc
   common/glob.cc
   common/hyperloglog.cc
+  common/malloc.cc
   common/multi-glob.cc
   common/perf.cc
   common/tar.cc
@@ -362,6 +363,7 @@ target_sources(mold PRIVATE
   elf/arch-sh4.cc
   elf/arch-sparc64.cc
   elf/arch-x86-64.cc
+  elf/config.cc
   elf/elf.cc
   git-hash.cc
   third-party/rust-demangle/rust-demangle.c
diff --git a/common/common.h b/common/common.h
index af3477aa..0968e879 100644
--- a/common/common.h
+++ b/common/common.h
@@ -401,13 +401,11 @@ inline std::vector<T> flatten(std::vector<std::vector<T>> &vec) {
   return ret;
 }
 
-template <typename T>
-inline void sort(T &vec) {
+inline void sort(auto &vec) {
   std::stable_sort(vec.begin(), vec.end());
 }
 
-template <typename T, typename U>
-inline void sort(T &vec, U less) {
+inline void sort(auto &vec, auto less) {
   std::stable_sort(vec.begin(), vec.end(), less);
 }
 
@@ -535,6 +533,9 @@ inline bool remove_prefix(std::string_view &s, std::string_view prefix) {
 // So you need to give a correct estimation of the final size before
 // using it. We use this hash map to uniquify pieces of data in
 // mergeable sections.
+//
+// We've implemented this ourselves because the performance of
+// conrurent hash map is critical for our linker.
 template <typename T>
 class ConcurrentMap {
 public:
@@ -840,8 +841,7 @@ std::array<u8, 16> get_uuid_v4();
 // filepath.cc
 //
 
-template <typename T>
-std::filesystem::path filepath(const T &path) {
+std::filesystem::path filepath(const auto &path) {
   return {path, std::filesystem::path::format::generic_format};
 }
 
diff --git a/common/malloc.cc b/common/malloc.cc
new file mode 100644
index 00000000..fc2f0afc
--- /dev/null
+++ b/common/malloc.cc
@@ -0,0 +1,5 @@
+#include "config.h"
+
+#ifdef USE_SYSTEM_MIMALLOC
+# include <mimalloc-new-delete.h>
+#endif
diff --git a/elf/config.cc b/elf/config.cc
new file mode 100644
index 00000000..55db9603
--- /dev/null
+++ b/elf/config.cc
@@ -0,0 +1,13 @@
+#include "mold.h"
+#include "config.h"
+
+namespace mold::elf {
+
+std::string get_mold_version() {
+  if (mold_git_hash.empty())
+    return "mold "s + MOLD_VERSION + " (compatible with GNU ld)";
+  return "mold "s + MOLD_VERSION + " (" + mold_git_hash +
+         "; compatible with GNU ld)";
+}
+
+} // namespace mold::elf
diff --git a/elf/elf.cc b/elf/elf.cc
index ab5fa5cc..2ce2ec47 100644
--- a/elf/elf.cc
+++ b/elf/elf.cc
@@ -1,15 +1,7 @@
 #include "mold.h"
-#include "config.h"
 
 namespace mold::elf {
 
-std::string get_mold_version() {
-  if (mold_git_hash.empty())
-    return "mold "s + MOLD_VERSION + " (compatible with GNU ld)";
-  return "mold "s + MOLD_VERSION + " (" + mold_git_hash +
-         "; compatible with GNU ld)";
-}
-
 static std::string unknown_type(u32 r_type) {
   char buf[50];
   snprintf(buf, sizeof(buf), "unknown (0x%x)", r_type);
diff --git a/elf/input-files.cc b/elf/input-files.cc
index 68a5e8b1..d1f71c3d 100644
--- a/elf/input-files.cc
+++ b/elf/input-files.cc
@@ -98,22 +98,6 @@ ElfShdr<E> *InputFile<E>::find_section(i64 type) {
   return nullptr;
 }
 
-template <typename E>
-void InputFile<E>::clear_symbols() {
-  for (Symbol<E> *sym : get_global_syms()) {
-    if (__atomic_load_n(&sym->file, __ATOMIC_ACQUIRE) == this) {
-      sym->origin = 0;
-      sym->value = -1;
-      sym->sym_idx = -1;
-      sym->ver_idx = VER_NDX_UNSPECIFIED;
-      sym->is_weak = false;
-      sym->is_imported = false;
-      sym->is_exported = false;
-      __atomic_store_n(&sym->file, nullptr, __ATOMIC_RELEASE);
-    }
-  }
-}
-
 // Find the source filename. It should be listed in symtab as STT_FILE.
 template <typename E>
 std::string_view InputFile<E>::get_source_name() const {
@@ -1512,7 +1496,7 @@ std::span<Symbol<E> *> SharedFile<E>::find_aliases(Symbol<E> *sym) {
     return x->esym().st_value < y->esym().st_value;
   });
 
-  return {&*begin, &*end};
+  return {&*begin, (size_t)(end - begin)};
 }
 
 // Infer an alignment of a DSO symbol. An alignment of a symbol in other
diff --git a/elf/main.cc b/elf/main.cc
index e0793521..68060f3a 100644
--- a/elf/main.cc
+++ b/elf/main.cc
@@ -21,8 +21,10 @@
 # include <unistd.h>
 #endif
 
-#if defined(USE_SYSTEM_MIMALLOC) && defined(MOLD_X86_64)
-# include <mimalloc-new-delete.h>
+#ifdef MOLD_X86_64
+int main(int argc, char **argv) {
+  return mold::elf::elf_main<mold::elf::X86_64>(argc, argv);
+}
 #endif
 
 namespace mold::elf {
@@ -718,9 +720,3 @@ using E = MOLD_TARGET;
 template int elf_main<E>(int, char **);
 
 } // namespace mold::elf
-
-#ifdef MOLD_X86_64
-int main(int argc, char **argv) {
-  return mold::elf::elf_main<mold::elf::X86_64>(argc, argv);
-}
-#endif
diff --git a/elf/mapfile.cc b/elf/mapfile.cc
index 1a894f52..c5805581 100644
--- a/elf/mapfile.cc
+++ b/elf/mapfile.cc
@@ -54,6 +54,8 @@ static Map<E> get_map(Context<E> &ctx) {
 
 template <typename E>
 void print_map(Context<E> &ctx) {
+  Timer t(ctx, "print_map");
+
   std::ostream *out = &std::cout;
   std::unique_ptr<std::ofstream> file;
 
diff --git a/elf/mold.h b/elf/mold.h
index d3039771..3046b4d9 100644
--- a/elf/mold.h
+++ b/elf/mold.h
@@ -1100,7 +1100,6 @@ public:
   ElfShdr<E> *find_section(i64 type);
 
   virtual void resolve_symbols(Context<E> &ctx) = 0;
-  void clear_symbols();
 
   virtual void
   mark_live_objects(Context<E> &ctx,
@@ -2043,7 +2042,7 @@ public:
 
   // A symbol is owned by a file. If two or more files define the
   // same symbol, the one with the strongest definition owns the symbol.
-  // If `file` is null, the symbol is equivalent to nonexistent.
+  // If `file` is null, the symbol is not defined by any input file.
   InputFile<E> *file = nullptr;
 
   // A symbol usually belongs to an input section, but it can belong
diff --git a/elf/output-chunks.cc b/elf/output-chunks.cc
index b36cc8c2..0a22fd9a 100644
--- a/elf/output-chunks.cc
+++ b/elf/output-chunks.cc
@@ -971,7 +971,7 @@ void OutputSection<E>::construct_relr(Context<E> &ctx) {
       return;
 
     for (const ElfRel<E> &r : isec.get_rels(ctx))
-      if (r.r_type == E::R_ABS && (r.r_offset % sizeof(Word<E>)) == 0)
+      if (r.r_type == E::R_ABS && r.r_offset % sizeof(Word<E>) == 0)
         if (Symbol<E> &sym = *isec.file.symbols[r.r_sym];
             !sym.is_absolute() && !sym.is_imported)
           shards[i].push_back(isec.offset + r.r_offset);
@@ -2337,8 +2337,8 @@ void CopyrelSection<E>::copy_buf(Context<E> &ctx) {
                                  this->reldyn_offset);
 
   for (Symbol<E> *sym : symbols)
-    *rel++ = ElfRel<E>(sym->get_addr(ctx), E::R_COPY, sym->get_dynsym_idx(ctx),
-                       0);
+    *rel++ = ElfRel<E>(sym->get_addr(ctx), E::R_COPY,
+                       sym->get_dynsym_idx(ctx), 0);
 }
 
 template <typename E>
diff --git a/elf/passes.cc b/elf/passes.cc
index 86e06594..75ae7242 100644
--- a/elf/passes.cc
+++ b/elf/passes.cc
@@ -229,6 +229,28 @@ static void mark_live_objects(Context<E> &ctx) {
   });
 }
 
+template <typename E>
+static void clear_symbols(Context<E> &ctx) {
+  std::vector<InputFile<E> *> files;
+  append(files, ctx.objs);
+  append(files, ctx.dsos);
+
+  tbb::parallel_for_each(files, [](InputFile<E> *file) {
+    for (Symbol<E> *sym : file->get_global_syms()) {
+      if (__atomic_load_n(&sym->file, __ATOMIC_ACQUIRE) == file) {
+        sym->origin = 0;
+        sym->value = -1;
+        sym->sym_idx = -1;
+        sym->ver_idx = VER_NDX_UNSPECIFIED;
+        sym->is_weak = false;
+        sym->is_imported = false;
+        sym->is_exported = false;
+        __atomic_store_n(&sym->file, nullptr, __ATOMIC_RELEASE);
+      }
+    }
+  });
+}
+
 template <typename E>
 void do_resolve_symbols(Context<E> &ctx) {
   std::vector<InputFile<E> *> files;
@@ -269,9 +291,7 @@ void do_resolve_symbols(Context<E> &ctx) {
     // Cleanup. The rule used for archive extraction isn't accurate for the
     // general case of symbol extraction, so reset the resolution to be redone
     // later.
-    tbb::parallel_for_each(files, [&](InputFile<E> *file) {
-      file->clear_symbols();
-    });
+    clear_symbols(ctx);
 
     // Now that the symbol references are gone, remove the eliminated files from
     // the file list.
@@ -349,13 +369,7 @@ void resolve_symbols(Context<E> &ctx) {
     append(ctx.objs, lto_objs);
 
     // Redo name resolution from scratch.
-    tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
-      file->clear_symbols();
-    });
-
-    tbb::parallel_for_each(ctx.dsos, [&](SharedFile<E> *file) {
-      file->clear_symbols();
-    });
+    clear_symbols(ctx);
 
     // Remove IR object files.
     for (ObjectFile<E> *file : ctx.objs)
@@ -864,8 +878,8 @@ void add_synthetic_symbols(Context<E> &ctx) {
   }
 
   if constexpr (is_ppc64v2<E>)
-    for (auto [label, insn] : ppc64_save_restore_insns)
-      if (!label.empty())
+    for (std::pair<std::string_view, u32> p : ppc64_save_restore_insns)
+      if (std::string_view label = p.first; !label.empty())
         add(label);
 
   obj.elf_syms = ctx.internal_esyms;
@@ -2855,7 +2869,8 @@ void fix_synthetic_symbols(Context<E> &ctx) {
   // PPC64's _{save,rest}gpr{0,1}_{14,15,16,...,31} symbols
   if constexpr (is_ppc64v2<E>) {
     i64 offset = 0;
-    for (auto [label, insn] : ppc64_save_restore_insns) {
+    for (std::pair<std::string_view, u32> p : ppc64_save_restore_insns) {
+      std::string_view label = p.first;
       if (!label.empty())
         if (Symbol<E> *sym = get_symbol(ctx, label);
             sym->file == ctx.internal_obj)
diff --git a/test/elf/aarch64_variant-pcs.sh b/test/elf/aarch64_variant-pcs.sh
index 6b5e4e8a..12361b79 100755
--- a/test/elf/aarch64_variant-pcs.sh
+++ b/test/elf/aarch64_variant-pcs.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = aarch64 ] || skip
-
 cat <<EOF | $CC -c -o $t/a.o -fPIC -xassembler - 2> /dev/null || skip
 .global foo
 .type foo, %function
diff --git a/test/elf/arm_range-extension-thunk-disassembly.sh b/test/elf/arm_range-extension-thunk-disassembly.sh
index 9f3dcc59..b1f28d79 100755
--- a/test/elf/arm_range-extension-thunk-disassembly.sh
+++ b/test/elf/arm_range-extension-thunk-disassembly.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = arm ] || skip
-
 cat <<EOF | $CC -c -o $t/a.o -fPIC -xc -
 #include <stdio.h>
 
diff --git a/test/elf/arm_range-extension-thunk.sh b/test/elf/arm_range-extension-thunk.sh
index ddf01313..191b1c51 100755
--- a/test/elf/arm_range-extension-thunk.sh
+++ b/test/elf/arm_range-extension-thunk.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = arm ] || skip
-
 echo 'int main() {}' | $CC -c -o /dev/null -xc - -O0 -mthumb >& /dev/null \
   || skip
 
diff --git a/test/elf/arm_thumb-interwork.sh b/test/elf/arm_thumb-interwork.sh
index baf8e04e..aed7b236 100755
--- a/test/elf/arm_thumb-interwork.sh
+++ b/test/elf/arm_thumb-interwork.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[[ $MACHINE == arm* ]] || skip
-
 echo 'int foo() { return 0; }' | $CC -o /dev/null -c -xc - -mthumb 2> /dev/null || skip
 
 cat <<EOF | $CC -o $t/a.o -c -xc - -mthumb
diff --git a/test/elf/arm_tlsdesc.sh b/test/elf/arm_tlsdesc.sh
index af6de79c..f8a2e5a3 100755
--- a/test/elf/arm_tlsdesc.sh
+++ b/test/elf/arm_tlsdesc.sh
@@ -1,7 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = arm ] || skip
 is_musl && skip
 
 echo 'int main() {}' | $GCC -c -o /dev/null -xc - -O0 -mthumb >& /dev/null \
diff --git a/test/elf/i686_tls-module-base.sh b/test/elf/i686_tls-module-base.sh
index f0bc4fea..2e906ca7 100755
--- a/test/elf/i686_tls-module-base.sh
+++ b/test/elf/i686_tls-module-base.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = i686 ] || skip
-
 cat <<'EOF' | $CC -fPIC -o $t/a.o -c -xassembler -
 .globl get_foo
 .type get_foo, @function
diff --git a/test/elf/riscv64_norvc.sh b/test/elf/riscv64_norvc.sh
index c679f64b..1383d4e0 100755
--- a/test/elf/riscv64_norvc.sh
+++ b/test/elf/riscv64_norvc.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = riscv64 -o $MACHINE = riscv32 ] || skip
-
 # Disable C extension
 if [ $MACHINE = riscv32 ]; then
     ISA=rv32g
diff --git a/test/elf/riscv64_relax-got.sh b/test/elf/riscv64_relax-got.sh
index 0acc11a6..14e33355 100755
--- a/test/elf/riscv64_relax-got.sh
+++ b/test/elf/riscv64_relax-got.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[[ $MACHINE = riscv* ]] || skip
-
 cat <<EOF | $CC -o $t/a.o -c -xassembler -
 .globl get_sym1, get_sym2, get_sym3, get_sym4, get_sym5
 .option norvc
diff --git a/test/elf/riscv64_relax-hi20.sh b/test/elf/riscv64_relax-hi20.sh
index 729ec063..a0befcda 100755
--- a/test/elf/riscv64_relax-hi20.sh
+++ b/test/elf/riscv64_relax-hi20.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = riscv64 -o $MACHINE = riscv32 ] || skip
-
 cat <<EOF | $CC -o $t/a.o -c -xassembler -
 .globl get_foo, get_foo2, get_bar, get_baz
 get_foo:
diff --git a/test/elf/s390x_got.sh b/test/elf/s390x_got.sh
index f50d7356..60234e42 100755
--- a/test/elf/s390x_got.sh
+++ b/test/elf/s390x_got.sh
@@ -1,8 +1,6 @@
 #!/bin/bash
 . $(dirname $0)/common.inc
 
-[ $MACHINE = s390x ] || skip
-
 # GOT[0] must be set to the link-time address of .dynamic on s390x.
 
 cat <<EOF | $CC -c -fPIC -o $t/a.o -xc -