Optimize

common/common.h

@@ -535,12 +535,22 @@ inline bool remove_prefix(std::string_view &s, std::string_view prefix) {
 template <typename T>
 class ConcurrentMap {
 public:
-  ConcurrentMap() {}
+  ConcurrentMap() = default;
 
   ConcurrentMap(i64 nbuckets) {
     resize(nbuckets);
   }
 
+  ~ConcurrentMap() {
+    if (entries) {
+#ifdef _WIN32
+      _aligned_free(entries);
+#else
+      munmap(entries, sizeof(Entry) * nbuckets);
+#endif
+    }
+  }
+
   // In order to avoid unnecessary cache-line false sharing, we want
   // to make this object to be aligned to a reasonably large
   // power-of-two address.
@@ -551,14 +561,19 @@ public:
   };
 
   void resize(i64 nbuckets) {
+    assert(!entries);
     this->nbuckets = std::max<i64>(MIN_NBUCKETS, bit_ceil(nbuckets));
+    i64 bufsize = sizeof(Entry) * this->nbuckets;
 
-    // Even though std::aligned_alloc is defined in C++17, MSVC doesn't
-    // seem to provide that function. C11's aligned_alloc may not always be
-    // available. Therefore, we'll align the buffer ourselves.
-    entries_buf.clear();
-    entries_buf.resize(sizeof(Entry) * this->nbuckets + alignof(Entry) - 1);
-    entries = (Entry *)align_to((uintptr_t)&entries_buf[0], alignof(Entry));
+    // Allocate a zero-initialized buffer. We use mmap() if available
+    // because it's faster than malloc() and memset().
+#ifdef _WIN32
+    entries = (Entry *)_aligned_malloc(bufsize, alignof(Entry));
+    memset((void *)entries, 0, bufsize);
+#else
+    entries = (Entry *)mmap(nullptr, bufsize, PROT_READ | PROT_WRITE,
+                            MAP_ANONYMOUS | MAP_PRIVATE, -1, 0);
+#endif
   }
 
   std::pair<T *, bool> insert(std::string_view key, u64 hash, const T &val) {
@@ -669,7 +684,6 @@ public:
   static constexpr i64 NUM_SHARDS = 16;
   static constexpr i64 MAX_RETRY = 128;
 
-  std::vector<u8> entries_buf;
   Entry *entries = nullptr;
   i64 nbuckets = 0;
 

elf/mold.h

@@ -798,14 +798,13 @@ public:
   void write_to(Context<E> &ctx, u8 *buf) override;
   void print_stats(Context<E> &ctx);
 
+  ConcurrentMap<SectionFragment<E>> map;
   HyperLogLog estimator;
 
 private:
   MergedSection(std::string_view name, i64 flags, i64 type, i64 entsize);
 
-  ConcurrentMap<SectionFragment<E>> map;
   std::vector<i64> shard_offsets;
-  std::once_flag once_flag;
 };
 
 template <typename E>

elf/output-chunks.cc

@@ -1933,11 +1933,6 @@ template <typename E>
 SectionFragment<E> *
 MergedSection<E>::insert(Context<E> &ctx, std::string_view data, u64 hash,
                          i64 p2align) {
-  std::call_once(once_flag, [&] {
-    // We aim 2/3 occupation ratio
-    map.resize(estimator.get_cardinality() * 3 / 2);
-  });
-
   // Even if GC is enabled, we garbage-collect only memory-mapped strings.
   // Non-memory-allocated strings are typically identifiers used by debug info.
   // To remove such strings, use the `strip` command.

elf/passes.cc

@@ -412,6 +412,10 @@ template <typename E>
 void resolve_section_pieces(Context<E> &ctx) {
   Timer t(ctx, "resolve_section_pieces");
 
+  // We aim 2/3 occupation ratio
+  for (std::unique_ptr<MergedSection<E>> &sec : ctx.merged_sections)
+    sec->map.resize(sec->estimator.get_cardinality() * 3 / 2);
+
   tbb::parallel_for_each(ctx.objs, [&](ObjectFile<E> *file) {
     file->resolve_section_pieces(ctx);
   });
@@ -441,6 +445,9 @@ void add_comment_string(Context<E> &ctx, std::string str) {
     MergedSection<E>::get_instance(ctx, ".comment", SHT_PROGBITS,
                                    SHF_MERGE | SHF_STRINGS, 1, 1);
 
+  if (sec->map.nbuckets == 0)
+    sec->map.resize(4096);
+
   std::string_view buf = save_string(ctx, str);
   std::string_view data(buf.data(), buf.size() + 1);
   sec->insert(ctx, data, hash_string(data), 0);