[NFCI][LTO][lld] Optimize away symbol copies within LTO global resolution in ELF (#106193)

`StringMap<T>` creates a [copy of the
string](d4c519e7b2/llvm/include/llvm/ADT/StringMapEntry.h (L55-L58))
for entry insertions and intentionally keep copies [since the
implementation optimizes string memory
usage](d4c519e7b2/llvm/include/llvm/ADT/StringMap.h (L124)).
On the other hand, linker keeps copies of symbol names [1] in
`lld:🧝:parseFiles` [2] before invoking `compileBitcodeFiles` [3].

This change proposes to optimize away string copies inside
[LTO::GlobalResolutions](24e791b416/llvm/include/llvm/LTO/LTO.h (L409)),
which will make LTO indexing more memory efficient for ELF. There are
similar opportunities for other (COFF, wasm, MachO) formats.

The optimization takes place for lld (ELF) only. For the rest of use
cases (gold plugin, `llvm-lto2`, etc), LTO owns a string saver to keep
copies and use global resolution key for de-duplication.

Together with @kazutakahirata's work to make `ComputeCrossModuleImport`
more memory efficient, we see a ~20% peak memory usage reduction in a
binary where peak memory usage needs to go down. Thanks to the
optimization in
329ba523cc,
the max (as opposed to the sum) of `ComputeCrossModuleImport` or
`GlobalResolution` shows up in peak memory usage.
* Regarding correctness, the set of
[resolved](80c47ad3ae/llvm/lib/LTO/LTO.cpp (L739))
[per-module
symbols](80c47ad3ae/llvm/include/llvm/LTO/LTO.h (L188-L191))
is a subset of
[llvm::lto::InputFile::Symbols](80c47ad3ae/llvm/include/llvm/LTO/LTO.h (L120)).
And bitcode symbol parsing saves symbol name when iterating
`obj->symbols` in `BitcodeFile::parse` already. This change updates
`BitcodeFile::parseLazy` to keep copies of per-module undefined symbols.
* Presumably the undefined symbols in a LTO unit (copied in this patch
in linker unique saver) is a small set compared with the set of symbols
in global-resolution (copied before this patch), making this a
worthwhile trade-off. Benchmarking this change alone shows measurable
memory savings across various benchmarks.

[1] ELF
1cea5c2138/lld/ELF/InputFiles.cpp (L1748)
[2]
ef7b18a53c/lld/ELF/Driver.cpp (L2863)
[3]
ef7b18a53c/lld/ELF/Driver.cpp (L2995)

lld/ELF/InputFiles.cpp

@@ -1804,6 +1804,10 @@ void BitcodeFile::parseLazy() {
       auto *sym = symtab.insert(unique_saver().save(irSym.getName()));
       sym->resolve(LazySymbol{*this});
       symbols[i] = sym;
+    } else {
+      // Keep copies of per-module undefined symbols for LTO::GlobalResolutions
+      // usage.
+      unique_saver().save(irSym.getName());
     }
 }
 

lld/ELF/LTO.cpp

@@ -135,6 +135,7 @@ static lto::Config createConfig() {
       config->ltoValidateAllVtablesHaveTypeInfos;
   c.AllVtablesHaveTypeInfos = ctx.ltoAllVtablesHaveTypeInfos;
   c.AlwaysEmitRegularLTOObj = !config->ltoObjPath.empty();
+  c.KeepSymbolNameCopies = false;
 
   for (const llvm::StringRef &name : config->thinLTOModulesToCompile)
     c.ThinLTOModulesToCompile.emplace_back(name);

llvm/include/llvm/LTO/Config.h

@@ -88,6 +88,11 @@ struct Config {
   /// want to know a priori all possible output files.
   bool AlwaysEmitRegularLTOObj = false;
 
+  /// If true, the LTO instance creates copies of the symbol names for LTO::run.
+  /// The lld linker uses string saver to keep symbol names alive and doesn't
+  /// need to create copies, so it can set this field to false.
+  bool KeepSymbolNameCopies = true;
+
   /// Allows non-imported definitions to get the potentially more constraining
   /// visibility from the prevailing definition. FromPrevailing is the default
   /// because it works for many binary formats. ELF can use the more optimized

llvm/include/llvm/LTO/LTO.h

@@ -15,6 +15,9 @@
 #ifndef LLVM_LTO_LTO_H
 #define LLVM_LTO_LTO_H
 
+#include <memory>
+
+#include "llvm/ADT/DenseMap.h"
 #include "llvm/ADT/MapVector.h"
 #include "llvm/ADT/StringMap.h"
 #include "llvm/Bitcode/BitcodeReader.h"
@@ -23,6 +26,7 @@
 #include "llvm/Object/IRSymtab.h"
 #include "llvm/Support/Caching.h"
 #include "llvm/Support/Error.h"
+#include "llvm/Support/StringSaver.h"
 #include "llvm/Support/thread.h"
 #include "llvm/Transforms/IPO/FunctionAttrs.h"
 #include "llvm/Transforms/IPO/FunctionImport.h"
@@ -403,10 +407,19 @@ private:
     };
   };
 
+  // GlobalResolutionSymbolSaver allocator.
+  std::unique_ptr<llvm::BumpPtrAllocator> Alloc;
+
+  // Symbol saver for global resolution map.
+  std::unique_ptr<llvm::StringSaver> GlobalResolutionSymbolSaver;
+
   // Global mapping from mangled symbol names to resolutions.
-  // Make this an optional to guard against accessing after it has been reset
+  // Make this an unique_ptr to guard against accessing after it has been reset
   // (to reduce memory after we're done with it).
-  std::optional<StringMap<GlobalResolution>> GlobalResolutions;
+  std::unique_ptr<llvm::DenseMap<StringRef, GlobalResolution>>
+      GlobalResolutions;
+
+  void releaseGlobalResolutionsMemory();
 
   void addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
                             ArrayRef<SymbolResolution> Res, unsigned Partition,

llvm/lib/LTO/LTO.cpp

@@ -77,6 +77,10 @@ cl::opt<bool> EnableLTOInternalization(
     "enable-lto-internalization", cl::init(true), cl::Hidden,
     cl::desc("Enable global value internalization in LTO"));
 
+static cl::opt<bool>
+    LTOKeepSymbolCopies("lto-keep-symbol-copies", cl::init(false), cl::Hidden,
+                        cl::desc("Keep copies of symbols in LTO indexing"));
+
 /// Indicate we are linking with an allocator that supports hot/cold operator
 /// new interfaces.
 extern cl::opt<bool> SupportsHotColdNew;
@@ -587,8 +591,14 @@ LTO::LTO(Config Conf, ThinBackend Backend,
     : Conf(std::move(Conf)),
       RegularLTO(ParallelCodeGenParallelismLevel, this->Conf),
       ThinLTO(std::move(Backend)),
-      GlobalResolutions(std::make_optional<StringMap<GlobalResolution>>()),
-      LTOMode(LTOMode) {}
+      GlobalResolutions(
+          std::make_unique<DenseMap<StringRef, GlobalResolution>>()),
+      LTOMode(LTOMode) {
+  if (Conf.KeepSymbolNameCopies || LTOKeepSymbolCopies) {
+    Alloc = std::make_unique<BumpPtrAllocator>();
+    GlobalResolutionSymbolSaver = std::make_unique<llvm::StringSaver>(*Alloc);
+  }
+}
 
 // Requires a destructor for MapVector<BitcodeModule>.
 LTO::~LTO() = default;
@@ -606,7 +616,12 @@ void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
     assert(ResI != ResE);
     SymbolResolution Res = *ResI++;
 
-    auto &GlobalRes = (*GlobalResolutions)[Sym.getName()];
+    StringRef SymbolName = Sym.getName();
+    // Keep copies of symbols if the client of LTO says so.
+    if (GlobalResolutionSymbolSaver && !GlobalResolutions->contains(SymbolName))
+      SymbolName = GlobalResolutionSymbolSaver->save(SymbolName);
+
+    auto &GlobalRes = (*GlobalResolutions)[SymbolName];
     GlobalRes.UnnamedAddr &= Sym.isUnnamedAddr();
     if (Res.Prevailing) {
       assert(!GlobalRes.Prevailing &&
@@ -660,6 +675,14 @@ void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
   }
 }
 
+void LTO::releaseGlobalResolutionsMemory() {
+  // Release GlobalResolutions dense-map itself.
+  GlobalResolutions.reset();
+  // Release the string saver memory.
+  GlobalResolutionSymbolSaver.reset();
+  Alloc.reset();
+}
+
 static void writeToResolutionFile(raw_ostream &OS, InputFile *Input,
                                   ArrayRef<SymbolResolution> Res) {
   StringRef Path = Input->getName();
@@ -1771,7 +1794,7 @@ Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache,
   // are no further accesses. We specifically want to do this before computing
   // cross module importing, which adds to peak memory via the computed import
   // and export lists.
-  GlobalResolutions.reset();
+  releaseGlobalResolutionsMemory();
 
   if (Conf.OptLevel > 0)
     ComputeCrossModuleImport(ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,