From 16921c78eaacb259bd9a3f20123db55b71a7b0ff Mon Sep 17 00:00:00 2001
From: Rui Ueyama <ruiu@cs.stanford.edu>
Date: Thu, 2 Dec 2021 11:35:48 +0900
Subject: [PATCH] [Mach-O] Add ARM64 stub

---
 macho/arch-arm64.cc     |  22 ++++++
 macho/arch-x86-64.cc    | 165 ++++++++++++++++++++++++++++++++++++++
 macho/cmdline.cc        |   7 +-
 macho/dead-strip.cc     |   1 +
 macho/input-sections.cc | 170 +---------------------------------------
 macho/macho.h           |  22 +++++-
 macho/main.cc           |   7 +-
 macho/mapfile.cc        |   1 +
 macho/mold.h            |  12 +++
 macho/object-file.cc    |   1 +
 macho/output-chunks.cc  |   1 +
 macho/output-file.cc    |   1 +
 macho/tapi.cc           |   1 +
 13 files changed, 236 insertions(+), 175 deletions(-)
 create mode 100644 macho/arch-arm64.cc
 create mode 100644 macho/arch-x86-64.cc

diff --git a/macho/arch-arm64.cc b/macho/arch-arm64.cc
new file mode 100644
index 00000000..c41ccb3e
--- /dev/null
+++ b/macho/arch-arm64.cc
@@ -0,0 +1,22 @@
+#include "mold.h"
+
+#include <algorithm>
+
+namespace mold::macho {
+
+template <>
+Relocation<ARM64>
+read_reloc(Context<ARM64> &ctx, ObjectFile<ARM64> &file,
+           const MachSection &hdr, MachRel r) {
+  return {};
+}
+
+template <>
+void Subsection<ARM64>::scan_relocations(Context<ARM64> &ctx) {
+}
+
+template <>
+void Subsection<ARM64>::apply_reloc(Context<ARM64> &ctx, u8 *buf) {
+}
+
+} // namespace mold::macho
diff --git a/macho/arch-x86-64.cc b/macho/arch-x86-64.cc
new file mode 100644
index 00000000..08cf2b1f
--- /dev/null
+++ b/macho/arch-x86-64.cc
@@ -0,0 +1,165 @@
+#include "mold.h"
+
+#include <algorithm>
+
+namespace mold::macho {
+
+static i64 get_reloc_addend(u32 type) {
+  switch (type) {
+  case X86_64_RELOC_SIGNED_1:
+    return 1;
+  case X86_64_RELOC_SIGNED_2:
+    return 2;
+  case X86_64_RELOC_SIGNED_4:
+    return 4;
+  default:
+    return 0;
+  }
+}
+
+static i64 read_addend(u8 *buf, const MachRel &r) {
+  switch (r.p2size) {
+  case 2:
+    return *(i32 *)(buf + r.offset) + get_reloc_addend(r.type);
+  case 3:
+    return *(i64 *)(buf + r.offset) + get_reloc_addend(r.type);
+  default:
+    unreachable();
+  }
+}
+
+template <>
+Relocation<X86_64>
+read_reloc(Context<X86_64> &ctx, ObjectFile<X86_64> &file,
+           const MachSection &hdr, MachRel r) {
+  if (r.p2size != 2 && r.p2size != 3)
+    Fatal(ctx) << file << ": invalid r.p2size: " << (u32)r.p2size;
+
+  if (r.is_pcrel) {
+    if (r.p2size != 2)
+      Fatal(ctx) << file << ": invalid PC-relative reloc: " << r.offset;
+  } else {
+    if (r.p2size != 3)
+      Fatal(ctx) << file << ": invalid non-PC-relative reloc: " << r.offset;
+  }
+
+  u8 *buf = (u8 *)file.mf->data + hdr.offset;
+  Relocation<X86_64> rel{r.offset, (u8)r.type, (u8)r.p2size, (bool)r.is_pcrel};
+  i64 addend = read_addend(buf, r);
+
+  if (r.is_extern) {
+    rel.sym = file.syms[r.idx];
+    rel.addend = addend;
+    return rel;
+  }
+
+  u32 addr;
+  if (r.is_pcrel)
+    addr = hdr.addr + r.offset + 4 + addend;
+  else
+    addr = addend;
+
+  Subsection<X86_64> *target = file.find_subsection(ctx, addr);
+  if (!target)
+    Fatal(ctx) << file << ": bad relocation: " << r.offset;
+
+  rel.subsec = target;
+  rel.addend = addr - target->input_addr;
+  return rel;
+}
+
+template <>
+void Subsection<X86_64>::scan_relocations(Context<X86_64> &ctx) {
+  for (Relocation<X86_64> &rel : get_rels()) {
+    Symbol<X86_64> *sym = rel.sym;
+    if (!sym)
+      continue;
+
+    switch (rel.type) {
+    case X86_64_RELOC_GOT_LOAD:
+      if (sym->file && sym->file->is_dylib)
+        sym->flags |= NEEDS_GOT;
+      break;
+    case X86_64_RELOC_GOT:
+      sym->flags |= NEEDS_GOT;
+      break;
+    case X86_64_RELOC_TLV:
+      if (sym->file && sym->file->is_dylib)
+        sym->flags |= NEEDS_THREAD_PTR;
+      break;
+    }
+
+    if (sym->file && sym->file->is_dylib) {
+      sym->flags |= NEEDS_STUB;
+      ((DylibFile<X86_64> *)sym->file)->is_needed = true;
+    }
+  }
+}
+
+template <>
+void Subsection<X86_64>::apply_reloc(Context<X86_64> &ctx, u8 *buf) {
+  for (const Relocation<X86_64> &rel : get_rels()) {
+    if (rel.sym && !rel.sym->file) {
+      Error(ctx) << "undefined symbol: " << isec.file << ": " << *rel.sym;
+      continue;
+    }
+
+    u64 val = 0;
+
+    switch (rel.type) {
+    case X86_64_RELOC_UNSIGNED:
+    case X86_64_RELOC_SIGNED:
+    case X86_64_RELOC_BRANCH:
+    case X86_64_RELOC_SIGNED_1:
+    case X86_64_RELOC_SIGNED_2:
+    case X86_64_RELOC_SIGNED_4:
+      val = rel.sym ? rel.sym->get_addr(ctx) : rel.subsec->get_addr(ctx);
+      break;
+    case X86_64_RELOC_GOT_LOAD:
+      if (rel.sym->got_idx != -1) {
+        val = rel.sym->get_got_addr(ctx);
+      } else {
+        // Relax MOVQ into LEAQ
+        if (buf[rel.offset - 2] != 0x8b)
+          Error(ctx) << isec << ": invalid GOT_LOAD relocation";
+        buf[rel.offset - 2] = 0x8d;
+        val = rel.sym->get_addr(ctx);
+      }
+      break;
+    case X86_64_RELOC_GOT:
+      val = rel.sym->get_got_addr(ctx);
+      break;
+    case X86_64_RELOC_TLV:
+      if (rel.sym->tlv_idx != -1) {
+        val = rel.sym->get_tlv_addr(ctx);
+      } else {
+        // Relax MOVQ into LEAQ
+        if (buf[rel.offset - 2] != 0x8b)
+          Error(ctx) << isec << ": invalid TLV relocation";
+        buf[rel.offset - 2] = 0x8d;
+        val = rel.sym->get_addr(ctx);
+      }
+      break;
+    default:
+      Fatal(ctx) << isec << ": unknown reloc: " << (int)rel.type;
+    }
+
+    val += rel.addend;
+
+    if (rel.is_pcrel)
+      val -= get_addr(ctx) + rel.offset + 4 + get_reloc_addend(rel.type);
+
+    switch (rel.p2size) {
+    case 2:
+      *(u32 *)(buf + rel.offset) = val;
+      break;
+    case 3:
+      *(u64 *)(buf + rel.offset) = val;
+      break;
+    default:
+      unreachable();
+    };
+  }
+}
+
+} // namespace mold::macho
diff --git a/macho/cmdline.cc b/macho/cmdline.cc
index 52dc6298..c96d9ad2 100644
--- a/macho/cmdline.cc
+++ b/macho/cmdline.cc
@@ -179,7 +179,11 @@ void parse_nonpositional_args(Context<E> &ctx,
     } else if (read_flag("-no_adhoc_codesign")) {
       ctx.arg.adhoc_codesign = false;
     } else if (read_arg("-arch")) {
-      if (arg != "x86_64")
+      if (arg == "x86_64")
+        ctx.arg.arch = CPU_TYPE_X86_64;
+      else if (arg == "arm64")
+        ctx.arg.arch = CPU_TYPE_ARM64;
+      else
         Fatal(ctx) << "unknown -arch: " << arg;
     } else if (read_flag("-color-diagnostics") ||
                read_flag("--color-diagnostics")) {
@@ -297,6 +301,7 @@ void parse_nonpositional_args(Context<E> &ctx,
 #define INSTANTIATE(E) \
   template void parse_nonpositional_args(Context<E> &, std::vector<std::string> &)
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho
diff --git a/macho/dead-strip.cc b/macho/dead-strip.cc
index b1462480..19aed4d3 100644
--- a/macho/dead-strip.cc
+++ b/macho/dead-strip.cc
@@ -105,6 +105,7 @@ void dead_strip(Context<E> &ctx) {
 #define INSTANTIATE(E)                          \
   template void dead_strip(Context<E> &)
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho
diff --git a/macho/input-sections.cc b/macho/input-sections.cc
index b2314d81..d39ee19c 100644
--- a/macho/input-sections.cc
+++ b/macho/input-sections.cc
@@ -1,16 +1,7 @@
 #include "mold.h"
 
-#include <algorithm>
-
 namespace mold::macho {
 
-template <typename E>
-std::ostream &operator<<(std::ostream &out, const InputSection<E> &sec) {
-  out << sec.file << "(" << sec.hdr.get_segname() << ","
-      << sec.hdr.get_sectname() << ")";
-  return out;
-}
-
 template <typename E>
 InputSection<E>::InputSection(Context<E> &ctx, ObjectFile<E> &file,
                               const MachSection &hdr)
@@ -21,69 +12,6 @@ InputSection<E>::InputSection(Context<E> &ctx, ObjectFile<E> &file,
     contents = file.mf->get_contents().substr(hdr.offset, hdr.size);
 }
 
-static i64 get_reloc_addend(u32 type) {
-  switch (type) {
-  case X86_64_RELOC_SIGNED_1:
-    return 1;
-  case X86_64_RELOC_SIGNED_2:
-    return 2;
-  case X86_64_RELOC_SIGNED_4:
-    return 4;
-  default:
-    return 0;
-  }
-}
-
-static i64 read_addend(u8 *buf, const MachRel &r) {
-  switch (r.p2size) {
-  case 2:
-    return *(i32 *)(buf + r.offset) + get_reloc_addend(r.type);
-  case 3:
-    return *(i64 *)(buf + r.offset) + get_reloc_addend(r.type);
-  default:
-    unreachable();
-  }
-}
-
-template <typename E>
-static Relocation<E> read_reloc(Context<E> &ctx, ObjectFile<E> &file,
-                                const MachSection &hdr, MachRel r) {
-  if (r.p2size != 2 && r.p2size != 3)
-    Fatal(ctx) << file << ": invalid r.p2size: " << (u32)r.p2size;
-
-  if (r.is_pcrel) {
-    if (r.p2size != 2)
-      Fatal(ctx) << file << ": invalid PC-relative reloc: " << r.offset;
-  } else {
-    if (r.p2size != 3)
-      Fatal(ctx) << file << ": invalid non-PC-relative reloc: " << r.offset;
-  }
-
-  u8 *buf = (u8 *)file.mf->data + hdr.offset;
-  Relocation<E> rel{r.offset, (u8)r.type, (u8)r.p2size, (bool)r.is_pcrel};
-  i64 addend = read_addend(buf, r);
-
-  if (r.is_extern) {
-    rel.sym = file.syms[r.idx];
-    rel.addend = addend;
-    return rel;
-  }
-
-  u32 addr;
-  if (r.is_pcrel)
-    addr = hdr.addr + r.offset + 4 + addend;
-  else
-    addr = addend;
-
-  Subsection<E> *target = file.find_subsection(ctx, addr);
-  if (!target)
-    Fatal(ctx) << file << ": bad relocation: " << r.offset;
-
-  rel.subsec = target;
-  rel.addend = addr - target->input_addr;
-  return rel;
-}
-
 template <typename E>
 void InputSection<E>::parse_relocations(Context<E> &ctx) {
   rels.reserve(hdr.nreloc);
@@ -125,104 +53,10 @@ void InputSection<E>::parse_relocations(Context<E> &ctx) {
   }
 }
 
-template <typename E>
-void Subsection<E>::scan_relocations(Context<E> &ctx) {
-  for (Relocation<E> &rel : get_rels()) {
-    Symbol<E> *sym = rel.sym;
-    if (!sym)
-      continue;
-
-    switch (rel.type) {
-    case X86_64_RELOC_GOT_LOAD:
-      if (sym->file && sym->file->is_dylib)
-        sym->flags |= NEEDS_GOT;
-      break;
-    case X86_64_RELOC_GOT:
-      sym->flags |= NEEDS_GOT;
-      break;
-    case X86_64_RELOC_TLV:
-      if (sym->file && sym->file->is_dylib)
-        sym->flags |= NEEDS_THREAD_PTR;
-      break;
-    }
-
-    if (sym->file && sym->file->is_dylib) {
-      sym->flags |= NEEDS_STUB;
-      ((DylibFile<E> *)sym->file)->is_needed = true;
-    }
-  }
-}
-
-template <typename E>
-void Subsection<E>::apply_reloc(Context<E> &ctx, u8 *buf) {
-  for (const Relocation<E> &rel : get_rels()) {
-    if (rel.sym && !rel.sym->file) {
-      Error(ctx) << "undefined symbol: " << isec.file << ": " << *rel.sym;
-      continue;
-    }
-
-    u64 val = 0;
-
-    switch (rel.type) {
-    case X86_64_RELOC_UNSIGNED:
-    case X86_64_RELOC_SIGNED:
-    case X86_64_RELOC_BRANCH:
-    case X86_64_RELOC_SIGNED_1:
-    case X86_64_RELOC_SIGNED_2:
-    case X86_64_RELOC_SIGNED_4:
-      val = rel.sym ? rel.sym->get_addr(ctx) : rel.subsec->get_addr(ctx);
-      break;
-    case X86_64_RELOC_GOT_LOAD:
-      if (rel.sym->got_idx != -1) {
-        val = rel.sym->get_got_addr(ctx);
-      } else {
-        // Relax MOVQ into LEAQ
-        if (buf[rel.offset - 2] != 0x8b)
-          Error(ctx) << isec << ": invalid GOT_LOAD relocation";
-        buf[rel.offset - 2] = 0x8d;
-        val = rel.sym->get_addr(ctx);
-      }
-      break;
-    case X86_64_RELOC_GOT:
-      val = rel.sym->get_got_addr(ctx);
-      break;
-    case X86_64_RELOC_TLV:
-      if (rel.sym->tlv_idx != -1) {
-        val = rel.sym->get_tlv_addr(ctx);
-      } else {
-        // Relax MOVQ into LEAQ
-        if (buf[rel.offset - 2] != 0x8b)
-          Error(ctx) << isec << ": invalid TLV relocation";
-        buf[rel.offset - 2] = 0x8d;
-        val = rel.sym->get_addr(ctx);
-      }
-      break;
-    default:
-      Fatal(ctx) << isec << ": unknown reloc: " << (int)rel.type;
-    }
-
-    val += rel.addend;
-
-    if (rel.is_pcrel)
-      val -= get_addr(ctx) + rel.offset + 4 + get_reloc_addend(rel.type);
-
-    switch (rel.p2size) {
-    case 2:
-      *(u32 *)(buf + rel.offset) = val;
-      break;
-    case 3:
-      *(u64 *)(buf + rel.offset) = val;
-      break;
-    default:
-      unreachable();
-    };
-  }
-}
-
 #define INSTANTIATE(E)                          \
-  template class InputSection<E>;               \
-  template class Subsection<E>
+  template class InputSection<E>
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho
diff --git a/macho/macho.h b/macho/macho.h
index 6fc33c83..f1adce42 100644
--- a/macho/macho.h
+++ b/macho/macho.h
@@ -168,7 +168,7 @@ static constexpr u32 S_ATTR_NO_TOC = 0x400000;
 static constexpr u32 S_ATTR_PURE_INSTRUCTIONS = 0x800000;
 
 static constexpr u32 CPU_TYPE_X86_64 = 0x1000007;
-static constexpr u32 CPU_TYPE_ARM64 = 0x1000012;
+static constexpr u32 CPU_TYPE_ARM64 = 0x100000c;
 
 static constexpr u32 CPU_SUBTYPE_X86_64_ALL = 3;
 static constexpr u32 CPU_SUBTYPE_ARM64_ALL = 0;
@@ -306,6 +306,18 @@ static constexpr u32 TOOL_CLANG = 1;
 static constexpr u32 TOOL_SWIFT = 2;
 static constexpr u32 TOOL_LD = 3;
 
+static constexpr u32 ARM64_RELOC_UNSIGNED = 0;
+static constexpr u32 ARM64_RELOC_SUBTRACTOR = 1;
+static constexpr u32 ARM64_RELOC_BRANCH26 = 2;
+static constexpr u32 ARM64_RELOC_PAGE21 = 3;
+static constexpr u32 ARM64_RELOC_PAGEOFF12 = 4;
+static constexpr u32 ARM64_RELOC_GOT_LOAD_PAGE21 = 5;
+static constexpr u32 ARM64_RELOC_GOT_LOAD_PAGEOFF12 = 6;
+static constexpr u32 ARM64_RELOC_POINTER_TO_GOT = 7;
+static constexpr u32 ARM64_RELOC_TLVP_LOAD_PAGE21 = 8;
+static constexpr u32 ARM64_RELOC_TLVP_LOAD_PAGEOFF12 = 9;
+static constexpr u32 ARM64_RELOC_ADDEND = 10;
+
 static constexpr u32 X86_64_RELOC_UNSIGNED = 0;
 static constexpr u32 X86_64_RELOC_SIGNED = 1;
 static constexpr u32 X86_64_RELOC_BRANCH = 2;
@@ -657,12 +669,14 @@ struct CodeSignatureDirectory {
   ubig64 exec_seg_flags;
 };
 
+struct ARM64 {
+  static constexpr u32 cputype = CPU_TYPE_ARM64;
+  static constexpr u32 cpusubtype = CPU_SUBTYPE_ARM64_ALL;
+};
+
 struct X86_64 {
   static constexpr u32 cputype = CPU_TYPE_X86_64;
   static constexpr u32 cpusubtype = CPU_SUBTYPE_X86_64_ALL;
 };
 
-struct AARCH64 {
-};
-
 } // namespace mold::macho
diff --git a/macho/main.cc b/macho/main.cc
index d2d95413..cb173705 100644
--- a/macho/main.cc
+++ b/macho/main.cc
@@ -254,7 +254,7 @@ strip_universal_header(Context<E> &ctx, MappedFile<Context<E>> *mf) {
 
   FatArch *arch = (FatArch *)(mf->data + sizeof(hdr));
   for (i64 i = 0; i < hdr.nfat_arch; i++)
-    if (arch[i].cputype == CPU_TYPE_X86_64)
+    if (arch[i].cputype == E::cputype)
       return mf->slice(ctx, mf->name, arch[i].offset, arch[i].size);
   Fatal(ctx) << mf->name << ": fat file contains no matching file";
 }
@@ -358,6 +358,9 @@ static int do_main(int argc, char **argv) {
   std::vector<std::string> file_args;
   parse_nonpositional_args(ctx, file_args);
 
+  if (ctx.arg.arch == CPU_TYPE_X86_64)
+    return do_main<X86_64>(argc, argv);
+
   read_input_files(ctx, file_args);
 
   i64 priority = 1;
@@ -453,7 +456,7 @@ static int do_main(int argc, char **argv) {
 }
 
 int main(int argc, char **argv) {
-  return do_main<X86_64>(argc, argv);
+  return do_main<ARM64>(argc, argv);
 }
 
 }
diff --git a/macho/mapfile.cc b/macho/mapfile.cc
index 9abed894..ba7e6416 100644
--- a/macho/mapfile.cc
+++ b/macho/mapfile.cc
@@ -74,6 +74,7 @@ void print_map(Context<E> &ctx) {
 #define INSTANTIATE(E)                          \
   template void print_map(Context<E> &)
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho
diff --git a/macho/mold.h b/macho/mold.h
index 271f85d5..cf7a1886 100644
--- a/macho/mold.h
+++ b/macho/mold.h
@@ -191,6 +191,10 @@ public:
   std::atomic_bool is_alive = false;
 };
 
+template <typename E>
+Relocation<E> read_reloc(Context<E> &ctx, ObjectFile<E> &file,
+                         const MachSection &hdr, MachRel r);
+
 //
 // Symbol
 //
@@ -781,6 +785,7 @@ struct Context {
     bool dynamic = true;
     bool fatal_warnings = false;
     bool trace = false;
+    i64 arch = CPU_TYPE_ARM64;
     i64 headerpad = 256;
     i64 pagezero_size = 0;
     i64 platform = PLATFORM_MACOS;
@@ -857,6 +862,13 @@ int main(int argc, char **argv);
 // Inline functions
 //
 
+template <typename E>
+std::ostream &operator<<(std::ostream &out, const InputSection<E> &sec) {
+  out << sec.file << "(" << sec.hdr.get_segname() << ","
+      << sec.hdr.get_sectname() << ")";
+  return out;
+}
+
 template <typename E>
 u64 Subsection<E>::get_addr(Context<E> &ctx) const {
   return ctx.arg.pagezero_size + raddr;
diff --git a/macho/object-file.cc b/macho/object-file.cc
index d430444a..1267ea56 100644
--- a/macho/object-file.cc
+++ b/macho/object-file.cc
@@ -602,6 +602,7 @@ void DylibFile<E>::resolve_symbols(Context<E> &ctx) {
   template class ObjectFile<E>;                                 \
   template class DylibFile<E>
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho
diff --git a/macho/output-chunks.cc b/macho/output-chunks.cc
index 0d78cc1e..20a2fe36 100644
--- a/macho/output-chunks.cc
+++ b/macho/output-chunks.cc
@@ -1342,6 +1342,7 @@ void ThreadPtrsSection<E>::copy_buf(Context<E> &ctx) {
   template class LazySymbolPtrSection<E>;               \
   template class ThreadPtrsSection<E>
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho
diff --git a/macho/output-file.cc b/macho/output-file.cc
index 051644a9..18ee76bb 100644
--- a/macho/output-file.cc
+++ b/macho/output-file.cc
@@ -114,6 +114,7 @@ OutputFile<E>::open(Context<E> &ctx, std::string path, i64 filesize, i64 perm) {
 #define INSTANTIATE(E)                          \
   template class OutputFile<E>
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho
diff --git a/macho/tapi.cc b/macho/tapi.cc
index 306e26e4..42adf6b2 100644
--- a/macho/tapi.cc
+++ b/macho/tapi.cc
@@ -131,6 +131,7 @@ TextDylib parse_tbd(Context<E> &ctx, MappedFile<Context<E>> *mf) {
 #define INSTANTIATE(E)                                                  \
   template TextDylib parse_tbd(Context<E> &, MappedFile<Context<E>> *);
 
+INSTANTIATE(ARM64);
 INSTANTIATE(X86_64);
 
 } // namespace mold::macho