temporary

Makefile

@@ -7,7 +7,7 @@ LLVM_LIBS=$(wildcard llvm-project/build/lib/libLLVM*.a)
 CURRENT_DIR=$(shell pwd)
 TBB_LIBDIR=$(wildcard $(CURRENT_DIR)/oneTBB/build/linux_intel64_*_release/)
 
-CPPFLAGS=-g $(shell $(LLVM_CONFIG) --cxxflags) -IoneTBB/include -pthread -std=c++17
+CPPFLAGS=-g $(shell $(LLVM_CONFIG) --cxxflags) -IoneTBB/include -pthread -std=c++17 -O2
 LDFLAGS=$(shell $(LLVM_CONFIG) --ldflags) -L$(TBB_LIBDIR) -Wl,-rpath=$(TBB_LIBDIR) -fuse-ld=lld
 LIBS=-pthread -ltbb -lcurses -Wl,--start-group $(LLVM_LIBS) -Wl,--end-group
 OBJS=main.o object_file.o input_sections.o output_chunks.o mapfile.o perf.o

input_sections.cc

@@ -24,7 +24,9 @@ void InputSection::copy_to(u8 *buf) {
   memcpy(buf, &data[0], data.size());
 }
 
-void InputSection::scan_relocations() {
+ScanRelResult InputSection::scan_relocations() {
+  ScanRelResult res;
+
   for (const ELF64LE::Rela &rel : rels) {
     Symbol *sym = file->symbols[rel.getSymbol(false)];
     if (!sym->file || !sym->file->is_alive)
@@ -43,7 +45,7 @@ void InputSection::scan_relocations() {
       std::lock_guard lock(sym->mu);
       if (!sym->needs_got) {
         sym->needs_got = true;
-        file->num_got++;
+        res.num_got++;
       }
       break;
     }
@@ -51,7 +53,7 @@ void InputSection::scan_relocations() {
       std::lock_guard lock(sym->mu);
       if (!sym->needs_gottp) {
         sym->needs_gottp = true;
-        file->num_got++;
+        res.num_got++;
       }
       break;
     }
@@ -62,14 +64,15 @@ void InputSection::scan_relocations() {
       std::lock_guard lock(sym->mu);
       if (!sym->needs_plt) {
         sym->needs_plt = true;
-        file->num_plt++;
-        file->num_gotplt++;
-        file->num_relplt++;
+        res.num_plt++;
+        res.num_gotplt++;
+        res.num_relplt++;
       }
       break;
     }
     }
   }
+  return res;
 }
 
 void InputSection::relocate(u8 *buf) {
@@ -146,7 +149,7 @@ void InputSection::relocate(u8 *buf) {
       error(toString(this) + ": unknown relocation: " +
             std::to_string(rel.getType(false)));
     }
-    
+
     static Counter counter("relocs");
     counter.inc();
   }

main.cc

@@ -232,26 +232,22 @@ static void set_isec_offsets() {
 }
 
 static void scan_rels(ArrayRef<ObjectFile *> files) {
-  std::atomic_int32_t num_got = 0;
-  std::atomic_int32_t num_gotplt = 0;
-  std::atomic_int32_t num_plt = 0;
-  std::atomic_int32_t num_relplt = 0;
+  ScanRelResult res =
+    tbb::parallel_reduce(tbb::blocked_range<int>(0, files.size()),
+                         ScanRelResult(),
+                         [&](const tbb::blocked_range<int> &r, ScanRelResult res) {
+                           for (int i = r.begin(); i != r.end(); i++)
+                             res = res.add(files[i]->scan_relocations());
+                           return res;
+                         },
+                         [](const ScanRelResult &x, const ScanRelResult &y) {
+                           return x.add(y);
+                         });
 
-  for_each(files, [&](ObjectFile *file) {
-                    for (InputSection *isec : file->sections)
-                      if (isec)
-                        isec->scan_relocations();
-
-                    num_got += file->num_got;
-                    num_gotplt += file->num_gotplt;
-                    num_plt += file->num_plt;
-                    num_relplt += file->num_relplt;
-                  });
-
-  out::got->shdr.sh_size = num_got * 8;
-  out::gotplt->shdr.sh_size = num_gotplt * 8;
-  out::plt->shdr.sh_size = num_plt * 16;
-  out::relplt->shdr.sh_size = num_relplt * sizeof(ELF64LE::Rela);
+  out::got->shdr.sh_size = res.num_got * 8;
+  out::gotplt->shdr.sh_size = res.num_gotplt * 8;
+  out::plt->shdr.sh_size = res.num_plt * 16;
+  out::relplt->shdr.sh_size = res.num_relplt * sizeof(ELF64LE::Rela);
 }
 
 static void assign_got_offsets(u8 *buf, ArrayRef<ObjectFile *> files) {

mold.h

@@ -12,6 +12,7 @@
 #include "tbb/concurrent_hash_map.h"
 #include "tbb/global_control.h"
 #include "tbb/parallel_for_each.h"
+#include "tbb/parallel_reduce.h"
 #include "tbb/spin_mutex.h"
 #include "tbb/task_group.h"
 
@@ -193,16 +194,30 @@ inline std::string toString(Symbol sym) {
 }
 
 //
-// input_chunks.cc
+// input_sections.cc
 //
 
+struct ScanRelResult {
+  ScanRelResult add(const ScanRelResult &other) const {
+    return {num_got + other.num_got,
+            num_gotplt + other.num_gotplt,
+            num_plt + other.num_plt,
+            num_relplt + other.num_relplt};
+  }
+
+  u32 num_got = 0;
+  u32 num_gotplt = 0;
+  u32 num_plt = 0;
+  u32 num_relplt = 0;
+};
+
 class InputSection {
 public:
   InputSection(ObjectFile *file, const ELF64LE::Shdr &shdr, StringRef name);
 
   void copy_to(u8 *buf);
   void relocate(u8 *buf);
-  void scan_relocations();
+  ScanRelResult scan_relocations();
 
   ObjectFile *file;
   OutputSection *output_section;
@@ -498,6 +513,7 @@ public:
   void hanlde_undefined_weak_symbols();
   void eliminate_duplicate_comdat_groups();
   void convert_common_symbols();
+  ScanRelResult scan_relocations();
   void fix_sym_addrs();
   void compute_symtab();
 
@@ -522,13 +538,6 @@ public:
   u64 global_symtab_size = 0;
   u64 global_strtab_size = 0;
 
-  // Number of .got, .got.plt, .plt and .rel.plt entries
-  // needed for this file.
-  i32 num_got = 0;
-  i32 num_gotplt = 0;
-  i32 num_plt = 0;
-  i32 num_relplt = 0;
-
 private:
   void initialize_sections();
   void initialize_symbols();

object_file.cc

@@ -401,6 +401,14 @@ void ObjectFile::convert_common_symbols() {
   }
 }
 
+ScanRelResult ObjectFile::scan_relocations() {
+  ScanRelResult res;
+  for (InputSection *isec : sections)
+    if (isec)
+      res = res.add(isec->scan_relocations());
+  return res;
+}
+
 void ObjectFile::fix_sym_addrs() {
   for (Symbol *sym : symbols) {
     if (sym->file != this)