[lld] [LLD][COFF] Deduplicate common chunks when linking COFF files. (PR #162553)

[Joshua Cranmer via llvm-commits]

https://github.com/jcranmer-intel created https://github.com/llvm/llvm-project/pull/162553

This fixes issue 162148.

Common symbols are intended to have only a single version of the data present in the final executable. The MSVC linker is able to successfully deduplicate these chunks. If you have an application with a large number of translation units with a large block of common data (this is possible, for example, with Fortran code), then failing to deduplicate these chunks can make the data size so large that the resulting executable fails to load.

The logic in this patch doesn't catch all of the potential cases for deduplication, but it should catch the most common ones.

>From 11e93ea411b33ff20d0eb8cc8ba798e9f1a11f87 Mon Sep 17 00:00:00 2001
From: Joshua Cranmer <joshua.cranmer at intel.com>
Date: Mon, 6 Oct 2025 13:42:59 -0700
Subject: [PATCH] [LLD][COFF] Deduplicate common chunks when linking COFF
 files.

This fixes issue 162148.

Common symbols are intended to have only a single version of the data
present in the final executable. The MSVC linker is able to successfully
deduplicate these chunks. If you have an application with a large number
of translation units with a large block of common data (this is
possible, for example, with Fortran code), then failing to deduplicate
these chunks can make the data size so large that the resulting
executable fails to load.

The logic in this patch doesn't catch all of the potential cases for
deduplication, but it should catch the most common ones.
---
 lld/COFF/Chunks.cpp | 2 +-
 lld/COFF/Chunks.h   | 2 ++
 lld/COFF/Symbols.h  | 2 ++
 lld/COFF/Writer.cpp | 4 ++++
 4 files changed, 9 insertions(+), 1 deletion(-)

diff --git a/lld/COFF/Chunks.cpp b/lld/COFF/Chunks.cpp
index ff3c89884c24d..d752a5bef7594 100644
--- a/lld/COFF/Chunks.cpp
+++ b/lld/COFF/Chunks.cpp
@@ -773,7 +773,7 @@ uint32_t SectionChunk::getSectionNumber() const {
   return s.getIndex() + 1;
 }
 
-CommonChunk::CommonChunk(const COFFSymbolRef s) : sym(s) {
+CommonChunk::CommonChunk(const COFFSymbolRef s) : active(false), sym(s) {
   // The value of a common symbol is its size. Align all common symbols smaller
   // than 32 bytes naturally, i.e. round the size up to the next power of two.
   // This is what MSVC link.exe does.
diff --git a/lld/COFF/Chunks.h b/lld/COFF/Chunks.h
index 7ba58e336451f..bf05d547f9c88 100644
--- a/lld/COFF/Chunks.h
+++ b/lld/COFF/Chunks.h
@@ -520,6 +520,8 @@ class CommonChunk : public NonSectionChunk {
   uint32_t getOutputCharacteristics() const override;
   StringRef getSectionName() const override { return ".bss"; }
 
+  bool active;
+
 private:
   const COFFSymbolRef sym;
 };
diff --git a/lld/COFF/Symbols.h b/lld/COFF/Symbols.h
index 465d4df52c630..e166329a66bdf 100644
--- a/lld/COFF/Symbols.h
+++ b/lld/COFF/Symbols.h
@@ -233,6 +233,8 @@ class DefinedCommon : public DefinedCOFF {
                 CommonChunk *c = nullptr)
       : DefinedCOFF(DefinedCommonKind, f, n, s), data(c), size(size) {
     this->isExternal = true;
+    if (c)
+      c->active = true;
   }
 
   static bool classof(const Symbol *s) {
diff --git a/lld/COFF/Writer.cpp b/lld/COFF/Writer.cpp
index 3d95d219a493c..e365eb140f52b 100644
--- a/lld/COFF/Writer.cpp
+++ b/lld/COFF/Writer.cpp
@@ -1093,6 +1093,10 @@ void Writer::createSections() {
         sc->printDiscardedMessage();
       continue;
     }
+    if (auto *cc = dyn_cast<CommonChunk>(c)) {
+      if (!cc->active)
+        continue;
+    }
     StringRef name = c->getSectionName();
     if (shouldStripSectionSuffix(sc, name, ctx.config.mingw))
       name = name.split('$').first;