[clang] [lld] [llvm] [LTO][LLD] Prevent invalid LTO libfunc transforms (PR #164916)
Daniel Thornburgh via cfe-commits
cfe-commits at lists.llvm.org
Fri Mar 13 11:54:13 PDT 2026
https://github.com/mysterymath updated https://github.com/llvm/llvm-project/pull/164916
>From 44c3ca2828ce68ca8a05e06aed0a1909a45cd32f Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 7 Aug 2025 16:05:09 -0700
Subject: [PATCH 01/22] [LTO][LLD] Prevent invalid LTO libfunc transforms
This patch ensures that:
1) New bitcode is not extracted for libfuncs after LTO occurs, and
2) Extracted bitcode for libfuncs is considered external, since new
calls to it may be emitted.
---
clang/lib/CodeGen/BackendUtil.cpp | 10 +--
lld/COFF/InputFiles.cpp | 4 +-
lld/ELF/Driver.cpp | 19 ++++-
lld/ELF/LTO.cpp | 5 +-
lld/ELF/LTO.h | 3 +-
lld/test/ELF/lto/libcall-archive-bitcode.test | 41 +++++++++++
llvm/include/llvm/LTO/LTO.h | 27 +++++--
llvm/include/llvm/LTO/LTOBackend.h | 7 +-
llvm/lib/LTO/LTO.cpp | 70 ++++++++++++++-----
llvm/lib/LTO/LTOBackend.cpp | 43 ++++++++++--
llvm/lib/LTO/LTOCodeGenerator.cpp | 4 +-
llvm/lib/LTO/ThinLTOCodeGenerator.cpp | 4 +-
llvm/lib/Object/CMakeLists.txt | 1 +
.../X86/libcall-external-bitcode.ll | 20 ++++++
.../X86/libcall-external-not-bitcode.ll | 20 ++++++
llvm/test/LTO/Resolution/X86/libcall-in-tu.ll | 34 +++++++++
llvm/tools/llvm-lto2/llvm-lto2.cpp | 7 ++
17 files changed, 274 insertions(+), 45 deletions(-)
create mode 100644 lld/test/ELF/lto/libcall-archive-bitcode.test
create mode 100644 llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll
create mode 100644 llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll
create mode 100644 llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
diff --git a/clang/lib/CodeGen/BackendUtil.cpp b/clang/lib/CodeGen/BackendUtil.cpp
index 4b4eaadcf1527..7ed7cece9d2d8 100644
--- a/clang/lib/CodeGen/BackendUtil.cpp
+++ b/clang/lib/CodeGen/BackendUtil.cpp
@@ -1410,11 +1410,11 @@ runThinLTOBackend(CompilerInstance &CI, ModuleSummaryIndex *CombinedIndex,
// FIXME: Both ExecuteAction and thinBackend set up optimization remarks for
// the same context.
finalizeLLVMOptimizationRemarks(M->getContext());
- if (Error E =
- thinBackend(Conf, -1, AddStream, *M, *CombinedIndex, ImportList,
- ModuleToDefinedGVSummaries[M->getModuleIdentifier()],
- /*ModuleMap=*/nullptr, Conf.CodeGenOnly,
- /*IRAddStream=*/nullptr, CGOpts.CmdArgs)) {
+ if (Error E = thinBackend(
+ Conf, -1, AddStream, *M, *CombinedIndex, ImportList,
+ ModuleToDefinedGVSummaries[M->getModuleIdentifier()],
+ /*ModuleMap=*/nullptr, Conf.CodeGenOnly, /*BitcodeLibFuncs=*/{},
+ /*IRAddStream=*/nullptr, CGOpts.CmdArgs)) {
handleAllErrors(std::move(E), [&](ErrorInfoBase &EIB) {
errs() << "Error running ThinLTO backend: " << EIB.message() << '\n';
});
diff --git a/lld/COFF/InputFiles.cpp b/lld/COFF/InputFiles.cpp
index 492b2ad80166d..0a40309174a67 100644
--- a/lld/COFF/InputFiles.cpp
+++ b/lld/COFF/InputFiles.cpp
@@ -1397,6 +1397,8 @@ void BitcodeFile::parse() {
comdat[i] =
symtab.addComdat(this, saver.save(obj->getComdatTable()[i].first));
Triple tt(obj->getTargetTriple());
+ TargetLibraryInfoImpl tlii(tt);
+ TargetLibraryInfo tli(tlii);
RTLIB::RuntimeLibcallsInfo libcalls(tt);
for (const lto::InputFile::Symbol &objSym : obj->symbols()) {
StringRef symName = saver.save(objSym.getName());
@@ -1447,7 +1449,7 @@ void BitcodeFile::parse() {
symtab.addRegular(this, symName, nullptr, fakeSC, 0, objSym.isWeak());
}
symbols.push_back(sym);
- if (objSym.isUsed() || objSym.isLibcall(libcalls))
+ if (objSym.isUsed() || objSym.isLibcall(tli, libcalls))
symtab.ctx.config.gcroot.push_back(sym);
}
directives = saver.save(obj->getCOFFLinkerOpts());
diff --git a/lld/ELF/Driver.cpp b/lld/ELF/Driver.cpp
index fac8fa2db04d6..4d283ddd6e3c4 100644
--- a/lld/ELF/Driver.cpp
+++ b/lld/ELF/Driver.cpp
@@ -2703,15 +2703,30 @@ static void markBuffersAsDontNeed(Ctx &ctx, bool skipLinkedOutput) {
template <class ELFT>
void LinkerDriver::compileBitcodeFiles(bool skipLinkedOutput) {
llvm::TimeTraceScope timeScope("LTO");
+ // Capture the triple before moving the bitcode into the bitcode compiler.
+ std::optional<llvm::Triple> tt;
+ if (!ctx.bitcodeFiles.empty())
+ tt = llvm::Triple(ctx.bitcodeFiles.front()->obj->getTargetTriple());
// Compile bitcode files and replace bitcode symbols.
lto.reset(new BitcodeCompiler(ctx));
for (BitcodeFile *file : ctx.bitcodeFiles)
lto->add(*file);
- if (!ctx.bitcodeFiles.empty())
+ llvm::BumpPtrAllocator alloc;
+ llvm::StringSaver saver(alloc);
+ SmallVector<StringRef> bitcodeLibFuncs;
+ if (!ctx.bitcodeFiles.empty()) {
markBuffersAsDontNeed(ctx, skipLinkedOutput);
+ for (StringRef libFunc : lto::LTO::getLibFuncSymbols(*tt, saver)) {
+ Symbol *sym = ctx.symtab->find(libFunc);
+ if (!sym)
+ continue;
+ if (isa<BitcodeFile>(sym->file))
+ bitcodeLibFuncs.push_back(libFunc);
+ }
+ }
- ltoObjectFiles = lto->compile();
+ ltoObjectFiles = lto->compile(bitcodeLibFuncs);
for (auto &file : ltoObjectFiles) {
auto *obj = cast<ObjFile<ELFT>>(file.get());
obj->parse(/*ignoreComdats=*/true);
diff --git a/lld/ELF/LTO.cpp b/lld/ELF/LTO.cpp
index ff8d44e0d7f64..2401e49e4cd30 100644
--- a/lld/ELF/LTO.cpp
+++ b/lld/ELF/LTO.cpp
@@ -317,7 +317,10 @@ static void thinLTOCreateEmptyIndexFiles(Ctx &ctx) {
// Merge all the bitcode files we have seen, codegen the result
// and return the resulting ObjectFile(s).
-SmallVector<std::unique_ptr<InputFile>, 0> BitcodeCompiler::compile() {
+SmallVector<std::unique_ptr<InputFile>, 0>
+BitcodeCompiler::compile(const SmallVector<StringRef> &bitcodeLibFuncs) {
+ ltoObj->setBitcodeLibFuncs(bitcodeLibFuncs);
+
unsigned maxTasks = ltoObj->getMaxTasks();
buf.resize(maxTasks);
files.resize(maxTasks);
diff --git a/lld/ELF/LTO.h b/lld/ELF/LTO.h
index acf3bcff7f2f1..8207e91460785 100644
--- a/lld/ELF/LTO.h
+++ b/lld/ELF/LTO.h
@@ -42,7 +42,8 @@ class BitcodeCompiler {
~BitcodeCompiler();
void add(BitcodeFile &f);
- SmallVector<std::unique_ptr<InputFile>, 0> compile();
+ SmallVector<std::unique_ptr<InputFile>, 0>
+ compile(const SmallVector<StringRef> &bitcodeLibFuncs);
private:
Ctx &ctx;
diff --git a/lld/test/ELF/lto/libcall-archive-bitcode.test b/lld/test/ELF/lto/libcall-archive-bitcode.test
new file mode 100644
index 0000000000000..20735b5c89c99
--- /dev/null
+++ b/lld/test/ELF/lto/libcall-archive-bitcode.test
@@ -0,0 +1,41 @@
+; REQUIRES: x86
+
+; RUN: rm -rf %t && split-file %s %t && cd %t
+; RUN: llvm-as main.ll -o main.o
+; RUN: llvm-as bcmp.ll -o bcmp.o
+; RUN: llvm-mc -filetype=obj -triple=x86_64-unknown-linux-gnu memcmp.s -o memcmp.o
+; RUN: llvm-ar rc libc.a bcmp.o memcmp.o
+
+;; Ensure that no memcmp->bcmp translation occurs during LTO because bcmp is in
+;; bitcode, but was not brought into the link. This would fail the link by
+;; extracting bitcode after LTO.
+; RUN: ld.lld -o out main.o -L. -lc
+; RUN: llvm-nm out | FileCheck %s
+
+;--- bcmp.ll
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i32 @bcmp(ptr %0, ptr %1, i64 %2) {
+ ret i32 0
+}
+
+;--- memcmp.s
+.globl memcmp
+memcmp:
+ ret
+
+;--- main.ll
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i1 @_start(ptr %0, ptr %1, i64 %2) {
+ %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
+ %eq = icmp eq i32 %cmp, 0
+ ret i1 %eq
+}
+
+; CHECK-NOT: bcmp
+; CHECK: memcmp
+declare i32 @memcmp(ptr, ptr, i64)
+
diff --git a/llvm/include/llvm/LTO/LTO.h b/llvm/include/llvm/LTO/LTO.h
index 38349b604b6b3..6cb063f68071f 100644
--- a/llvm/include/llvm/LTO/LTO.h
+++ b/llvm/include/llvm/LTO/LTO.h
@@ -179,7 +179,8 @@ class InputFile {
// may emit references to. Such symbols must be considered external, as
// removing them or modifying their interfaces would invalidate the code
// generator's knowledge about them.
- bool isLibcall(const RTLIB::RuntimeLibcallsInfo &Libcalls) const;
+ bool isLibcall(const TargetLibraryInfo &TLI,
+ const RTLIB::RuntimeLibcallsInfo &Libcalls) const;
};
/// A range over the symbols in this InputFile.
@@ -308,7 +309,8 @@ class ThinBackendProc {
using ThinBackendFunction = std::function<std::unique_ptr<ThinBackendProc>(
const Config &C, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
- AddStreamFn AddStream, FileCache Cache)>;
+ AddStreamFn AddStream, FileCache Cache,
+ const SmallVector<StringRef> &BitcodeLibFuncs)>;
/// This type defines the behavior following the thin-link phase during ThinLTO.
/// It encapsulates a backend function and a strategy for thread pool
@@ -323,10 +325,11 @@ struct ThinBackend {
std::unique_ptr<ThinBackendProc> operator()(
const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
- AddStreamFn AddStream, FileCache Cache) {
+ AddStreamFn AddStream, FileCache Cache,
+ const SmallVector<StringRef> &BitcodeLibFuncs) {
assert(isValid() && "Invalid backend function");
return Func(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
- std::move(AddStream), std::move(Cache));
+ std::move(AddStream), std::move(Cache), BitcodeLibFuncs);
}
ThreadPoolStrategy getParallelism() const { return Parallelism; }
bool isValid() const { return static_cast<bool>(Func); }
@@ -444,6 +447,12 @@ class LTO {
LLVM_ABI Error add(std::unique_ptr<InputFile> Obj,
ArrayRef<SymbolResolution> Res);
+ /// Set the list of functions implemented in bitcode across the link, whether
+ /// extracted or not. Such functions may not be referenced if they were not
+ /// extracted by the time LTO occurs.
+ LLVM_ABI void
+ setBitcodeLibFuncs(const SmallVector<StringRef> &BitcodeLibFuncs);
+
/// Returns an upper bound on the number of tasks that the client may expect.
/// This may only be called after all IR object files have been added. For a
/// full description of tasks see LTOBackend.h.
@@ -464,6 +473,14 @@ class LTO {
LLVM_ABI static SmallVector<const char *>
getRuntimeLibcallSymbols(const Triple &TT);
+ /// Static method that returns a list of library function symbols that can be
+ /// generated by LTO but might not be visible from bitcode symbol table.
+ /// Unlike the runtime libcalls, the linker can report to the code generator
+ /// which of these are actually available in the link, and the code generator
+ /// can then only reference that set of symbols.
+ LLVM_ABI static SmallVector<StringRef>
+ getLibFuncSymbols(const Triple &TT, llvm::StringSaver &Saver);
+
protected:
// Called at the start of run().
virtual Error serializeInputsForDistribution() { return Error::success(); }
@@ -655,6 +672,8 @@ class LTO {
// Setup optimization remarks according to the provided configuration.
Error setupOptimizationRemarks();
+ SmallVector<StringRef> BitcodeLibFuncs;
+
public:
/// Helper to emit an optimization remark during the LTO link when outside of
/// the standard optimization pass pipeline.
diff --git a/llvm/include/llvm/LTO/LTOBackend.h b/llvm/include/llvm/LTO/LTOBackend.h
index 48ad5aa64f61f..6a7d7e0d87ac9 100644
--- a/llvm/include/llvm/LTO/LTOBackend.h
+++ b/llvm/include/llvm/LTO/LTOBackend.h
@@ -39,13 +39,15 @@ LLVM_ABI bool opt(const Config &Conf, TargetMachine *TM, unsigned Task,
Module &Mod, bool IsThinLTO,
ModuleSummaryIndex *ExportSummary,
const ModuleSummaryIndex *ImportSummary,
- const std::vector<uint8_t> &CmdArgs);
+ const std::vector<uint8_t> &CmdArgs,
+ const SmallVector<StringRef> &BitcodeLibFuncs);
/// Runs a regular LTO backend. The regular LTO backend can also act as the
/// regular LTO phase of ThinLTO, which may need to access the combined index.
LLVM_ABI Error backend(const Config &C, AddStreamFn AddStream,
unsigned ParallelCodeGenParallelismLevel, Module &M,
- ModuleSummaryIndex &CombinedIndex);
+ ModuleSummaryIndex &CombinedIndex,
+ const SmallVector<StringRef> &BitcodeLibFuncs);
/// Runs a ThinLTO backend.
/// If \p ModuleMap is not nullptr, all the module files to be imported have
@@ -62,6 +64,7 @@ thinBackend(const Config &C, unsigned Task, AddStreamFn AddStream, Module &M,
const FunctionImporter::ImportMapTy &ImportList,
const GVSummaryMapTy &DefinedGlobals,
MapVector<StringRef, BitcodeModule> *ModuleMap, bool CodeGenOnly,
+ const SmallVector<StringRef> &BitcodeLibFuncs,
AddStreamFn IRAddStream = nullptr,
const std::vector<uint8_t> &CmdArgs = std::vector<uint8_t>());
diff --git a/llvm/lib/LTO/LTO.cpp b/llvm/lib/LTO/LTO.cpp
index 36360081fa4a0..0f3f3c344fb0a 100644
--- a/llvm/lib/LTO/LTO.cpp
+++ b/llvm/lib/LTO/LTO.cpp
@@ -653,7 +653,11 @@ Expected<std::unique_ptr<InputFile>> InputFile::create(MemoryBufferRef Object) {
}
bool InputFile::Symbol::isLibcall(
+ const TargetLibraryInfo &TLI,
const RTLIB::RuntimeLibcallsInfo &Libcalls) const {
+ LibFunc F;
+ if (TLI.getLibFunc(IRName, F) && TLI.has(F))
+ return true;
return Libcalls.getSupportedLibcallImpl(IRName) != RTLIB::Unsupported;
}
@@ -715,6 +719,8 @@ void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
auto *ResE = Res.end();
(void)ResE;
RTLIB::RuntimeLibcallsInfo Libcalls(TT);
+ TargetLibraryInfoImpl TLII(TT);
+ TargetLibraryInfo TLI(TLII);
for (const InputFile::Symbol &Sym : Syms) {
assert(ResI != ResE);
SymbolResolution Res = *ResI++;
@@ -757,7 +763,7 @@ void LTO::addModuleToGlobalRes(ArrayRef<InputFile::Symbol> Syms,
GlobalRes.VisibleOutsideSummary = true;
}
- bool IsLibcall = Sym.isLibcall(Libcalls);
+ bool IsLibcall = Sym.isLibcall(TLI, Libcalls);
// Set the partition to external if we know it is re-defined by the linker
// with -defsym or -wrap options, used elsewhere, e.g. it is visible to a
@@ -844,6 +850,10 @@ Error LTO::add(std::unique_ptr<InputFile> InputPtr,
return Error::success();
}
+void LTO::setBitcodeLibFuncs(const SmallVector<StringRef> &BitcodeLibFuncs) {
+ this->BitcodeLibFuncs = BitcodeLibFuncs;
+}
+
Expected<ArrayRef<SymbolResolution>>
LTO::addModule(InputFile &Input, ArrayRef<SymbolResolution> InputRes,
unsigned ModI, ArrayRef<SymbolResolution> Res) {
@@ -1469,9 +1479,9 @@ Error LTO::runRegularLTO(AddStreamFn AddStream) {
}
if (!RegularLTO.EmptyCombinedModule || Conf.AlwaysEmitRegularLTOObj) {
- if (Error Err =
- backend(Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel,
- *RegularLTO.CombinedModule, ThinLTO.CombinedIndex))
+ if (Error Err = backend(
+ Conf, AddStream, RegularLTO.ParallelCodeGenParallelismLevel,
+ *RegularLTO.CombinedModule, ThinLTO.CombinedIndex, BitcodeLibFuncs))
return Err;
}
@@ -1491,6 +1501,21 @@ SmallVector<const char *> LTO::getRuntimeLibcallSymbols(const Triple &TT) {
return LibcallSymbols;
}
+SmallVector<StringRef> LTO::getLibFuncSymbols(const Triple &TT,
+ StringSaver &Saver) {
+ auto TLII = std::make_unique<TargetLibraryInfoImpl>(TT);
+ TargetLibraryInfo TLI(*TLII);
+ SmallVector<StringRef> LibFuncSymbols;
+ LibFuncSymbols.reserve(LibFunc::NumLibFuncs);
+ for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs); I != E;
+ ++I) {
+ LibFunc F = static_cast<LibFunc>(I);
+ if (TLI.has(F))
+ LibFuncSymbols.push_back(Saver.save(TLI.getName(F)).data());
+ }
+ return LibFuncSymbols;
+}
+
Error ThinBackendProc::emitFiles(
const FunctionImporter::ImportMapTy &ImportList, llvm::StringRef ModulePath,
const std::string &NewModulePath) const {
@@ -1568,6 +1593,7 @@ class CGThinBackend : public ThinBackendProc {
class InProcessThinBackend : public CGThinBackend {
protected:
FileCache Cache;
+ const SmallVector<StringRef> &BitcodeLibFuncs;
public:
InProcessThinBackend(
@@ -1575,11 +1601,12 @@ class InProcessThinBackend : public CGThinBackend {
ThreadPoolStrategy ThinLTOParallelism,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache, lto::IndexWriteCallback OnWrite,
- bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles)
+ bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles,
+ const SmallVector<StringRef> &BitcodeLibFuncs)
: CGThinBackend(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
AddStream, OnWrite, ShouldEmitIndexFiles,
ShouldEmitImportsFiles, ThinLTOParallelism),
- Cache(std::move(Cache)) {}
+ Cache(std::move(Cache)), BitcodeLibFuncs(BitcodeLibFuncs) {}
virtual Error runThinLTOBackendThread(
AddStreamFn AddStream, FileCache Cache, unsigned Task, BitcodeModule BM,
@@ -1600,7 +1627,7 @@ class InProcessThinBackend : public CGThinBackend {
return thinBackend(Conf, Task, AddStream, **MOrErr, CombinedIndex,
ImportList, DefinedGlobals, &ModuleMap,
- Conf.CodeGenOnly);
+ Conf.CodeGenOnly, BitcodeLibFuncs);
};
if (ShouldEmitIndexFiles) {
if (auto E = emitFiles(ImportList, ModuleID, ModuleID.str()))
@@ -1685,13 +1712,14 @@ class FirstRoundThinBackend : public InProcessThinBackend {
const Config &Conf, ModuleSummaryIndex &CombinedIndex,
ThreadPoolStrategy ThinLTOParallelism,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
- AddStreamFn CGAddStream, FileCache CGCache, AddStreamFn IRAddStream,
+ AddStreamFn CGAddStream, FileCache CGCache,
+ const SmallVector<StringRef> &BitcodeLibFuncs, AddStreamFn IRAddStream,
FileCache IRCache)
: InProcessThinBackend(Conf, CombinedIndex, ThinLTOParallelism,
ModuleToDefinedGVSummaries, std::move(CGAddStream),
std::move(CGCache), /*OnWrite=*/nullptr,
/*ShouldEmitIndexFiles=*/false,
- /*ShouldEmitImportsFiles=*/false),
+ /*ShouldEmitImportsFiles=*/false, BitcodeLibFuncs),
IRAddStream(std::move(IRAddStream)), IRCache(std::move(IRCache)) {}
Error runThinLTOBackendThread(
@@ -1714,7 +1742,7 @@ class FirstRoundThinBackend : public InProcessThinBackend {
return thinBackend(Conf, Task, CGAddStream, **MOrErr, CombinedIndex,
ImportList, DefinedGlobals, &ModuleMap,
- Conf.CodeGenOnly, IRAddStream);
+ Conf.CodeGenOnly, BitcodeLibFuncs, IRAddStream);
};
// Like InProcessThinBackend, we produce index files as needed for
// FirstRoundThinBackend. However, these files are not generated for
@@ -1781,6 +1809,7 @@ class SecondRoundThinBackend : public InProcessThinBackend {
ThreadPoolStrategy ThinLTOParallelism,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache,
+ const SmallVector<StringRef> &BitcodeLibFuncs,
std::unique_ptr<SmallVector<StringRef>> IRFiles,
stable_hash CombinedCGDataHash)
: InProcessThinBackend(Conf, CombinedIndex, ThinLTOParallelism,
@@ -1788,7 +1817,7 @@ class SecondRoundThinBackend : public InProcessThinBackend {
std::move(Cache),
/*OnWrite=*/nullptr,
/*ShouldEmitIndexFiles=*/false,
- /*ShouldEmitImportsFiles=*/false),
+ /*ShouldEmitImportsFiles=*/false, BitcodeLibFuncs),
IRFiles(std::move(IRFiles)), CombinedCGDataHash(CombinedCGDataHash) {}
Error runThinLTOBackendThread(
@@ -1809,7 +1838,7 @@ class SecondRoundThinBackend : public InProcessThinBackend {
return thinBackend(Conf, Task, AddStream, *LoadedModule, CombinedIndex,
ImportList, DefinedGlobals, &ModuleMap,
- /*CodeGenOnly=*/true);
+ /*CodeGenOnly=*/true, BitcodeLibFuncs);
};
if (!Cache.isValid() || !CombinedIndex.modulePaths().count(ModuleID) ||
all_of(CombinedIndex.getModuleHash(ModuleID),
@@ -1848,11 +1877,12 @@ ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism,
auto Func =
[=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
- AddStreamFn AddStream, FileCache Cache) {
+ AddStreamFn AddStream, FileCache Cache,
+ const SmallVector<StringRef> &BitcodeLibFuncs) {
return std::make_unique<InProcessThinBackend>(
Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
AddStream, Cache, OnWrite, ShouldEmitIndexFiles,
- ShouldEmitImportsFiles);
+ ShouldEmitImportsFiles, BitcodeLibFuncs);
};
return ThinBackend(Func, Parallelism);
}
@@ -1969,7 +1999,8 @@ ThinBackend lto::createWriteIndexesThinBackend(
auto Func =
[=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
- AddStreamFn AddStream, FileCache Cache) {
+ AddStreamFn AddStream, FileCache Cache,
+ const SmallVector<StringRef> &BitcodeLibFuncs) {
return std::make_unique<WriteIndexesThinBackend>(
Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
OldPrefix, NewPrefix, NativeObjectPrefix, ShouldEmitImportsFiles,
@@ -2222,7 +2253,7 @@ Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache,
if (!CodeGenDataThinLTOTwoRounds) {
std::unique_ptr<ThinBackendProc> BackendProc =
ThinLTO.Backend(Conf, ThinLTO.CombinedIndex, ModuleToDefinedGVSummaries,
- AddStream, Cache);
+ AddStream, Cache, BitcodeLibFuncs);
return RunBackends(BackendProc.get());
}
@@ -2245,7 +2276,7 @@ Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache,
LLVM_DEBUG(dbgs() << "[TwoRounds] Running the first round of codegen\n");
auto FirstRoundLTO = std::make_unique<FirstRoundThinBackend>(
Conf, ThinLTO.CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
- CG.AddStream, CG.Cache, IR.AddStream, IR.Cache);
+ CG.AddStream, CG.Cache, BitcodeLibFuncs, IR.AddStream, IR.Cache);
if (Error E = RunBackends(FirstRoundLTO.get()))
return E;
@@ -2261,7 +2292,7 @@ Error LTO::runThinLTO(AddStreamFn AddStream, FileCache Cache,
LLVM_DEBUG(dbgs() << "[TwoRounds] Running the second round of codegen\n");
auto SecondRoundLTO = std::make_unique<SecondRoundThinBackend>(
Conf, ThinLTO.CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
- AddStream, Cache, IR.getResult(), CombinedHash);
+ AddStream, Cache, BitcodeLibFuncs, IR.getResult(), CombinedHash);
return RunBackends(SecondRoundLTO.get());
}
@@ -2768,7 +2799,8 @@ ThinBackend lto::createOutOfProcessThinBackend(
auto Func =
[=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
- AddStreamFn AddStream, FileCache Cache) {
+ AddStreamFn AddStream, FileCache Cache,
+ const SmallVector<StringRef> &BitcodeLibFuncs) {
return std::make_unique<OutOfProcessThinBackend>(
Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
AddStream, Cache, OnWrite, ShouldEmitIndexFiles,
diff --git a/llvm/lib/LTO/LTOBackend.cpp b/llvm/lib/LTO/LTOBackend.cpp
index 7ee2557a68bd5..8cd8a75c52cab 100644
--- a/llvm/lib/LTO/LTOBackend.cpp
+++ b/llvm/lib/LTO/LTOBackend.cpp
@@ -259,7 +259,8 @@ createTargetMachine(const Config &Conf, const Target *TheTarget, Module &M) {
static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
unsigned OptLevel, bool IsThinLTO,
ModuleSummaryIndex *ExportSummary,
- const ModuleSummaryIndex *ImportSummary) {
+ const ModuleSummaryIndex *ImportSummary,
+ const DenseSet<StringRef> &BitcodeLibFuncs) {
std::optional<PGOOptions> PGOOpt;
if (!Conf.SampleProfile.empty())
PGOOpt = PGOOptions(Conf.SampleProfile, "", Conf.ProfileRemapping,
@@ -301,6 +302,28 @@ static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
new TargetLibraryInfoImpl(TM->getTargetTriple(), TM->Options.VecLib));
if (Conf.Freestanding)
TLII->disableAllFunctions();
+
+ TargetLibraryInfo TLI(*TLII);
+ for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs); I != E;
+ ++I) {
+ LibFunc F = static_cast<LibFunc>(I);
+ StringRef Name = TLI.getName(F);
+ GlobalValue *Val = Mod.getNamedValue(Name);
+
+ // LibFuncs present in the current TU can always be referenced.
+ if (Val && !Val->isDeclaration())
+ continue;
+
+ // LibFuncs not implemented in bitcode can always be referenced.
+ if (!BitcodeLibFuncs.contains(Name))
+ continue;
+
+ // FIXME: Functions that are somewhere in a ThinLTO link (just not imported
+ // in this module) should not be disabled, as they have already been
+ // extracted.
+ TLII->setUnavailable(F);
+ }
+
FAM.registerPass([&] { return TargetLibraryAnalysis(*TLII); });
// Parse a custom AA pipeline if asked to.
@@ -384,7 +407,8 @@ static bool isEmptyModule(const Module &Mod) {
bool lto::opt(const Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
const ModuleSummaryIndex *ImportSummary,
- const std::vector<uint8_t> &CmdArgs) {
+ const std::vector<uint8_t> &CmdArgs,
+ const SmallVector<StringRef> &BitcodeLibFuncs) {
llvm::TimeTraceScope timeScope("opt");
if (EmbedBitcode == LTOBitcodeEmbedding::EmbedPostMergePreOptimized) {
// FIXME: the motivation for capturing post-merge bitcode and command line
@@ -409,9 +433,11 @@ bool lto::opt(const Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
// analysis in the case of a ThinLTO build where this might be an empty
// regular LTO combined module, with a large combined index from ThinLTO.
if (!isEmptyModule(Mod)) {
+ DenseSet<StringRef> BitcodeLibFuncsSet(BitcodeLibFuncs.begin(),
+ BitcodeLibFuncs.end());
// FIXME: Plumb the combined index into the new pass manager.
runNewPMPasses(Conf, Mod, TM, Conf.OptLevel, IsThinLTO, ExportSummary,
- ImportSummary);
+ ImportSummary, BitcodeLibFuncsSet);
}
return !Conf.PostOptModuleHook || Conf.PostOptModuleHook(Task, Mod);
}
@@ -577,7 +603,8 @@ Error lto::finalizeOptimizationRemarks(LLVMRemarkFileHandle DiagOutputFile) {
Error lto::backend(const Config &C, AddStreamFn AddStream,
unsigned ParallelCodeGenParallelismLevel, Module &Mod,
- ModuleSummaryIndex &CombinedIndex) {
+ ModuleSummaryIndex &CombinedIndex,
+ const SmallVector<StringRef> &BitcodeLibFuncs) {
llvm::TimeTraceScope timeScope("LTO backend");
Expected<const Target *> TOrErr = initAndLookupTarget(C, Mod);
if (!TOrErr)
@@ -589,7 +616,7 @@ Error lto::backend(const Config &C, AddStreamFn AddStream,
if (!C.CodeGenOnly) {
if (!opt(C, TM.get(), 0, Mod, /*IsThinLTO=*/false,
/*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr,
- /*CmdArgs*/ std::vector<uint8_t>()))
+ /*CmdArgs*/ std::vector<uint8_t>(), BitcodeLibFuncs))
return Error::success();
}
@@ -629,7 +656,9 @@ Error lto::thinBackend(const Config &Conf, unsigned Task, AddStreamFn AddStream,
const FunctionImporter::ImportMapTy &ImportList,
const GVSummaryMapTy &DefinedGlobals,
MapVector<StringRef, BitcodeModule> *ModuleMap,
- bool CodeGenOnly, AddStreamFn IRAddStream,
+ bool CodeGenOnly,
+ const SmallVector<StringRef> &BitcodeLibFuncs,
+ AddStreamFn IRAddStream,
const std::vector<uint8_t> &CmdArgs) {
llvm::TimeTraceScope timeScope("Thin backend", Mod.getModuleIdentifier());
Expected<const Target *> TOrErr = initAndLookupTarget(Conf, Mod);
@@ -668,7 +697,7 @@ Error lto::thinBackend(const Config &Conf, unsigned Task, AddStreamFn AddStream,
// Perform optimization and code generation for ThinLTO.
if (!opt(Conf, TM, Task, Mod, /*IsThinLTO=*/true,
/*ExportSummary=*/nullptr, /*ImportSummary=*/&CombinedIndex,
- CmdArgs))
+ CmdArgs, BitcodeLibFuncs))
return finalizeOptimizationRemarks(std::move(DiagnosticOutputFile));
// Save the current module before the first codegen round.
diff --git a/llvm/lib/LTO/LTOCodeGenerator.cpp b/llvm/lib/LTO/LTOCodeGenerator.cpp
index 46be71da5a092..fda0d36458f6d 100644
--- a/llvm/lib/LTO/LTOCodeGenerator.cpp
+++ b/llvm/lib/LTO/LTOCodeGenerator.cpp
@@ -614,7 +614,7 @@ bool LTOCodeGenerator::optimize() {
TargetMach = createTargetMachine();
if (!opt(Config, TargetMach.get(), 0, *MergedModule, /*IsThinLTO=*/false,
/*ExportSummary=*/&CombinedIndex, /*ImportSummary=*/nullptr,
- /*CmdArgs*/ std::vector<uint8_t>())) {
+ /*CmdArgs*/ std::vector<uint8_t>(), {})) {
emitError("LTO middle-end optimizations failed");
return false;
}
@@ -639,7 +639,7 @@ bool LTOCodeGenerator::compileOptimized(AddStreamFn AddStream,
Config.CodeGenOnly = true;
Error Err = backend(Config, AddStream, ParallelismLevel, *MergedModule,
- CombinedIndex);
+ CombinedIndex, {});
assert(!Err && "unexpected code-generation failure");
(void)Err;
diff --git a/llvm/lib/LTO/ThinLTOCodeGenerator.cpp b/llvm/lib/LTO/ThinLTOCodeGenerator.cpp
index 93b672ae7840c..6397fdffed687 100644
--- a/llvm/lib/LTO/ThinLTOCodeGenerator.cpp
+++ b/llvm/lib/LTO/ThinLTOCodeGenerator.cpp
@@ -292,8 +292,10 @@ addUsedSymbolToPreservedGUID(const lto::InputFile &File,
DenseSet<GlobalValue::GUID> &PreservedGUID) {
Triple TT(File.getTargetTriple());
RTLIB::RuntimeLibcallsInfo Libcalls(TT);
+ TargetLibraryInfoImpl TLII(TT);
+ TargetLibraryInfo TLI(TLII);
for (const auto &Sym : File.symbols())
- if (Sym.isUsed() || Sym.isLibcall(Libcalls))
+ if (Sym.isUsed() || Sym.isLibcall(TLI, Libcalls))
PreservedGUID.insert(
GlobalValue::getGUIDAssumingExternalLinkage(Sym.getIRName()));
}
diff --git a/llvm/lib/Object/CMakeLists.txt b/llvm/lib/Object/CMakeLists.txt
index 0f6d2f7c59a5c..c48d251249488 100644
--- a/llvm/lib/Object/CMakeLists.txt
+++ b/llvm/lib/Object/CMakeLists.txt
@@ -51,6 +51,7 @@ add_llvm_component_library(LLVMObject
BinaryFormat
MCParser
Support
+ Target
TargetParser
TextAPI
)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll b/llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll
new file mode 100644
index 0000000000000..95a599fe75e8b
--- /dev/null
+++ b/llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll
@@ -0,0 +1,20 @@
+; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
+; RUN: llvm-lto2 run -o %t2 \
+; RUN: -r %t,foo,plx \
+; RUN: -r %t,memcmp,x \
+; RUN: -r %t,bcmp,pl --bitcode-libfuncs=bcmp %t -save-temps
+; RUN: llvm-dis %t2.1.4.opt.bc -o - | FileCheck %s
+
+define i1 @foo(ptr %0, ptr %1, i64 %2) {
+ ; CHECK-LABEL: define{{.*}}i1 @foo
+ ; CHECK-NEXT: %cmp = {{.*}}call i32 @memcmp
+ ; CHECK-NEXT: %eq = icmp eq i32 %cmp, 0
+ ; CHECK-NEXT: ret i1 %eq
+
+ %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
+ %eq = icmp eq i32 %cmp, 0
+ ret i1 %eq
+}
+
+declare i32 @memcmp(ptr, ptr, i64)
+declare i32 @bcmp(ptr, ptr, i64)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll b/llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll
new file mode 100644
index 0000000000000..2e6cc798d22cd
--- /dev/null
+++ b/llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll
@@ -0,0 +1,20 @@
+; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
+; RUN: llvm-lto2 run -o %t2 \
+; RUN: -r %t,foo,plx \
+; RUN: -r %t,memcmp,x \
+; RUN: -r %t,bcmp,pl %t -save-temps
+; RUN: llvm-dis %t2.1.4.opt.bc -o - | FileCheck %s
+
+define i1 @foo(ptr %0, ptr %1, i64 %2) {
+ ; CHECK-LABEL: define{{.*}}i1 @foo
+ ; CHECK-NEXT: %bcmp = {{.*}}call i32 @bcmp
+ ; CHECK-NEXT: %eq = icmp eq i32 %bcmp, 0
+ ; CHECK-NEXT: ret i1 %eq
+
+ %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
+ %eq = icmp eq i32 %cmp, 0
+ ret i1 %eq
+}
+
+declare i32 @memcmp(ptr, ptr, i64)
+declare i32 @bcmp(ptr, ptr, i64)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll b/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
new file mode 100644
index 0000000000000..948f21a6536ca
--- /dev/null
+++ b/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
@@ -0,0 +1,34 @@
+;; This test comes from a real world scenario in LTO, where the definition of
+;; bcmp was deleted because it has no uses, but later instcombine re-introduced
+;; a call to bcmp() as part of SimplifyLibCalls. Such deletions must not be
+;; allowed.
+
+; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
+; RUN: llvm-lto2 run -o %t2 \
+; RUN: -r %t,foo,plx \
+; RUN: -r %t,memcmp,x \
+; RUN: -r %t,bcmp,pl \
+; RUN: -r %t,bcmp_impl,x %t -save-temps
+; RUN: llvm-dis %t2.1.4.opt.bc -o - | FileCheck %s
+
+define i1 @foo(ptr %0, ptr %1, i64 %2) {
+ ; CHECK-LABEL: define{{.*}}i1 @foo
+ ; CHECK-NEXT: %bcmp = {{.*}}call i32 @bcmp
+ ; CHECK-NEXT: %eq = icmp eq i32 %bcmp, 0
+ ; CHECK-NEXT: ret i1 %eq
+
+ %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
+ %eq = icmp eq i32 %cmp, 0
+ ret i1 %eq
+}
+
+declare i32 @memcmp(ptr, ptr, i64)
+declare i32 @bcmp_impl(ptr, ptr, i64)
+
+;; Ensure bcmp is not removed from module because it is external.
+; CHECK: define dso_local i32 @bcmp
+define i32 @bcmp(ptr %0, ptr %1, i64 %2) noinline {
+ %r = call i32 @bcmp_impl(ptr %0, ptr %1, i64 %2)
+ ret i32 %r
+}
+
diff --git a/llvm/tools/llvm-lto2/llvm-lto2.cpp b/llvm/tools/llvm-lto2/llvm-lto2.cpp
index 95e8053dab32d..66fa5789feb64 100644
--- a/llvm/tools/llvm-lto2/llvm-lto2.cpp
+++ b/llvm/tools/llvm-lto2/llvm-lto2.cpp
@@ -234,6 +234,10 @@ static cl::opt<bool>
AllVtablesHaveTypeInfos("all-vtables-have-type-infos", cl::Hidden,
cl::desc("All vtables have type infos"));
+static cl::list<std::string>
+ BitcodeLibFuncs("bitcode-libfuncs", cl::Hidden,
+ cl::desc("set of libfuncs implemented in bitcode"));
+
static cl::opt<bool> TimeTrace("time-trace", cl::desc("Record time trace"));
static cl::opt<unsigned> TimeTraceGranularity(
@@ -496,6 +500,9 @@ static int run(int argc, char **argv) {
if (HasErrors)
return 1;
+ Lto.setBitcodeLibFuncs(
+ SmallVector<StringRef>(BitcodeLibFuncs.begin(), BitcodeLibFuncs.end()));
+
auto AddStream =
[&](size_t Task,
const Twine &ModuleName) -> std::unique_ptr<CachedFileStream> {
>From c516b9070bf477bcdc32cc4b1b573b468a40d7d0 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 11 Dec 2025 15:33:23 -0800
Subject: [PATCH 02/22] Move bitcodeLibFuncs from compile() to setter
---
lld/ELF/LTO.cpp | 10 ++++++----
lld/ELF/LTO.h | 4 ++--
2 files changed, 8 insertions(+), 6 deletions(-)
diff --git a/lld/ELF/LTO.cpp b/lld/ELF/LTO.cpp
index 2401e49e4cd30..23cf9a20c21c4 100644
--- a/lld/ELF/LTO.cpp
+++ b/lld/ELF/LTO.cpp
@@ -317,10 +317,7 @@ static void thinLTOCreateEmptyIndexFiles(Ctx &ctx) {
// Merge all the bitcode files we have seen, codegen the result
// and return the resulting ObjectFile(s).
-SmallVector<std::unique_ptr<InputFile>, 0>
-BitcodeCompiler::compile(const SmallVector<StringRef> &bitcodeLibFuncs) {
- ltoObj->setBitcodeLibFuncs(bitcodeLibFuncs);
-
+SmallVector<std::unique_ptr<InputFile>, 0> BitcodeCompiler::compile() {
unsigned maxTasks = ltoObj->getMaxTasks();
buf.resize(maxTasks);
files.resize(maxTasks);
@@ -431,3 +428,8 @@ BitcodeCompiler::compile(const SmallVector<StringRef> &bitcodeLibFuncs) {
}
return ret;
}
+
+void BitcodeCompiler::setBitcodeLibFuncs(
+ const SmallVector<StringRef> &bitcodeLibFuncs) {
+ ltoObj->setBitcodeLibFuncs(bitcodeLibFuncs);
+}
diff --git a/lld/ELF/LTO.h b/lld/ELF/LTO.h
index 8207e91460785..b9547202901fd 100644
--- a/lld/ELF/LTO.h
+++ b/lld/ELF/LTO.h
@@ -42,8 +42,8 @@ class BitcodeCompiler {
~BitcodeCompiler();
void add(BitcodeFile &f);
- SmallVector<std::unique_ptr<InputFile>, 0>
- compile(const SmallVector<StringRef> &bitcodeLibFuncs);
+ SmallVector<std::unique_ptr<InputFile>, 0> compile();
+ void setBitcodeLibFuncs(const SmallVector<StringRef> &bitcodeLibFuncs);
private:
Ctx &ctx;
>From 936c87df5da10c69cebea576f29597c8100f647b Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 11 Dec 2025 15:35:22 -0800
Subject: [PATCH 03/22] Densify bitcode file finding
---
lld/ELF/Driver.cpp | 15 +++++++--------
1 file changed, 7 insertions(+), 8 deletions(-)
diff --git a/lld/ELF/Driver.cpp b/lld/ELF/Driver.cpp
index 4d283ddd6e3c4..db28f180557bb 100644
--- a/lld/ELF/Driver.cpp
+++ b/lld/ELF/Driver.cpp
@@ -2714,19 +2714,18 @@ void LinkerDriver::compileBitcodeFiles(bool skipLinkedOutput) {
llvm::BumpPtrAllocator alloc;
llvm::StringSaver saver(alloc);
- SmallVector<StringRef> bitcodeLibFuncs;
if (!ctx.bitcodeFiles.empty()) {
markBuffersAsDontNeed(ctx, skipLinkedOutput);
- for (StringRef libFunc : lto::LTO::getLibFuncSymbols(*tt, saver)) {
- Symbol *sym = ctx.symtab->find(libFunc);
- if (!sym)
- continue;
- if (isa<BitcodeFile>(sym->file))
+
+ SmallVector<StringRef> bitcodeLibFuncs;
+ for (StringRef libFunc : lto::LTO::getLibFuncSymbols(*tt, saver))
+ if (Symbol *sym = ctx.symtab->find(libFunc);
+ sym && isa<BitcodeFile>(sym->file))
bitcodeLibFuncs.push_back(libFunc);
- }
+ lto->setBitcodeLibFuncs(bitcodeLibFuncs);
}
- ltoObjectFiles = lto->compile(bitcodeLibFuncs);
+ ltoObjectFiles = lto->compile();
for (auto &file : ltoObjectFiles) {
auto *obj = cast<ObjFile<ELFT>>(file.get());
obj->parse(/*ignoreComdats=*/true);
>From eb4ed2bccfa75267a7c5240771585eae8b013685 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 11 Dec 2025 15:39:54 -0800
Subject: [PATCH 04/22] Add comment about triple assumption
---
lld/ELF/Driver.cpp | 2 ++
1 file changed, 2 insertions(+)
diff --git a/lld/ELF/Driver.cpp b/lld/ELF/Driver.cpp
index db28f180557bb..c2304b7677f2a 100644
--- a/lld/ELF/Driver.cpp
+++ b/lld/ELF/Driver.cpp
@@ -2704,6 +2704,8 @@ template <class ELFT>
void LinkerDriver::compileBitcodeFiles(bool skipLinkedOutput) {
llvm::TimeTraceScope timeScope("LTO");
// Capture the triple before moving the bitcode into the bitcode compiler.
+ // Note that this assumes that the set of possible libfuncs is roughly
+ // equivalent for all bitcode translation units.
std::optional<llvm::Triple> tt;
if (!ctx.bitcodeFiles.empty())
tt = llvm::Triple(ctx.bitcodeFiles.front()->obj->getTargetTriple());
>From 3f8277fc35a24e637319d9b5845b3ac968a26f5e Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 13 Jan 2026 16:26:00 -0800
Subject: [PATCH 05/22] Remove optional from triple in LLD Driver.cpp
---
lld/ELF/Driver.cpp | 4 ++--
1 file changed, 2 insertions(+), 2 deletions(-)
diff --git a/lld/ELF/Driver.cpp b/lld/ELF/Driver.cpp
index c2304b7677f2a..476ccc9433e3a 100644
--- a/lld/ELF/Driver.cpp
+++ b/lld/ELF/Driver.cpp
@@ -2706,7 +2706,7 @@ void LinkerDriver::compileBitcodeFiles(bool skipLinkedOutput) {
// Capture the triple before moving the bitcode into the bitcode compiler.
// Note that this assumes that the set of possible libfuncs is roughly
// equivalent for all bitcode translation units.
- std::optional<llvm::Triple> tt;
+ llvm::Triple tt;
if (!ctx.bitcodeFiles.empty())
tt = llvm::Triple(ctx.bitcodeFiles.front()->obj->getTargetTriple());
// Compile bitcode files and replace bitcode symbols.
@@ -2720,7 +2720,7 @@ void LinkerDriver::compileBitcodeFiles(bool skipLinkedOutput) {
markBuffersAsDontNeed(ctx, skipLinkedOutput);
SmallVector<StringRef> bitcodeLibFuncs;
- for (StringRef libFunc : lto::LTO::getLibFuncSymbols(*tt, saver))
+ for (StringRef libFunc : lto::LTO::getLibFuncSymbols(tt, saver))
if (Symbol *sym = ctx.symtab->find(libFunc);
sym && isa<BitcodeFile>(sym->file))
bitcodeLibFuncs.push_back(libFunc);
>From 1bea9ff4c498722f65c475b50838519f908f9c5f Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 13 Jan 2026 16:33:46 -0800
Subject: [PATCH 06/22] Replace SmallVector& with ArrayRef
---
lld/ELF/LTO.cpp | 3 +--
lld/ELF/LTO.h | 2 +-
llvm/include/llvm/LTO/LTO.h | 7 +++----
llvm/include/llvm/LTO/LTOBackend.h | 21 ++++++++++-----------
llvm/lib/LTO/LTO.cpp | 19 ++++++++++---------
llvm/lib/LTO/LTOBackend.cpp | 7 +++----
6 files changed, 28 insertions(+), 31 deletions(-)
diff --git a/lld/ELF/LTO.cpp b/lld/ELF/LTO.cpp
index 23cf9a20c21c4..cb81decfba8ce 100644
--- a/lld/ELF/LTO.cpp
+++ b/lld/ELF/LTO.cpp
@@ -429,7 +429,6 @@ SmallVector<std::unique_ptr<InputFile>, 0> BitcodeCompiler::compile() {
return ret;
}
-void BitcodeCompiler::setBitcodeLibFuncs(
- const SmallVector<StringRef> &bitcodeLibFuncs) {
+void BitcodeCompiler::setBitcodeLibFuncs(ArrayRef<StringRef> bitcodeLibFuncs) {
ltoObj->setBitcodeLibFuncs(bitcodeLibFuncs);
}
diff --git a/lld/ELF/LTO.h b/lld/ELF/LTO.h
index b9547202901fd..c8cb2156d90ca 100644
--- a/lld/ELF/LTO.h
+++ b/lld/ELF/LTO.h
@@ -43,7 +43,7 @@ class BitcodeCompiler {
void add(BitcodeFile &f);
SmallVector<std::unique_ptr<InputFile>, 0> compile();
- void setBitcodeLibFuncs(const SmallVector<StringRef> &bitcodeLibFuncs);
+ void setBitcodeLibFuncs(ArrayRef<StringRef> bitcodeLibFuncs);
private:
Ctx &ctx;
diff --git a/llvm/include/llvm/LTO/LTO.h b/llvm/include/llvm/LTO/LTO.h
index 6cb063f68071f..bb5d38da8d9b6 100644
--- a/llvm/include/llvm/LTO/LTO.h
+++ b/llvm/include/llvm/LTO/LTO.h
@@ -310,7 +310,7 @@ using ThinBackendFunction = std::function<std::unique_ptr<ThinBackendProc>(
const Config &C, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache,
- const SmallVector<StringRef> &BitcodeLibFuncs)>;
+ ArrayRef<StringRef> BitcodeLibFuncs)>;
/// This type defines the behavior following the thin-link phase during ThinLTO.
/// It encapsulates a backend function and a strategy for thread pool
@@ -326,7 +326,7 @@ struct ThinBackend {
const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache,
- const SmallVector<StringRef> &BitcodeLibFuncs) {
+ ArrayRef<StringRef> BitcodeLibFuncs) {
assert(isValid() && "Invalid backend function");
return Func(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
std::move(AddStream), std::move(Cache), BitcodeLibFuncs);
@@ -450,8 +450,7 @@ class LTO {
/// Set the list of functions implemented in bitcode across the link, whether
/// extracted or not. Such functions may not be referenced if they were not
/// extracted by the time LTO occurs.
- LLVM_ABI void
- setBitcodeLibFuncs(const SmallVector<StringRef> &BitcodeLibFuncs);
+ LLVM_ABI void setBitcodeLibFuncs(ArrayRef<StringRef> BitcodeLibFuncs);
/// Returns an upper bound on the number of tasks that the client may expect.
/// This may only be called after all IR object files have been added. For a
diff --git a/llvm/include/llvm/LTO/LTOBackend.h b/llvm/include/llvm/LTO/LTOBackend.h
index 6a7d7e0d87ac9..4bb38529ec754 100644
--- a/llvm/include/llvm/LTO/LTOBackend.h
+++ b/llvm/include/llvm/LTO/LTOBackend.h
@@ -40,14 +40,14 @@ LLVM_ABI bool opt(const Config &Conf, TargetMachine *TM, unsigned Task,
ModuleSummaryIndex *ExportSummary,
const ModuleSummaryIndex *ImportSummary,
const std::vector<uint8_t> &CmdArgs,
- const SmallVector<StringRef> &BitcodeLibFuncs);
+ ArrayRef<StringRef> BitcodeLibFuncs);
/// Runs a regular LTO backend. The regular LTO backend can also act as the
/// regular LTO phase of ThinLTO, which may need to access the combined index.
LLVM_ABI Error backend(const Config &C, AddStreamFn AddStream,
unsigned ParallelCodeGenParallelismLevel, Module &M,
ModuleSummaryIndex &CombinedIndex,
- const SmallVector<StringRef> &BitcodeLibFuncs);
+ ArrayRef<StringRef> BitcodeLibFuncs);
/// Runs a ThinLTO backend.
/// If \p ModuleMap is not nullptr, all the module files to be imported have
@@ -58,15 +58,14 @@ LLVM_ABI Error backend(const Config &C, AddStreamFn AddStream,
/// the backend will skip optimization and only perform code generation. If
/// \p IRAddStream is not nullptr, it will be called just before code generation
/// to serialize the optimized IR.
-LLVM_ABI Error
-thinBackend(const Config &C, unsigned Task, AddStreamFn AddStream, Module &M,
- const ModuleSummaryIndex &CombinedIndex,
- const FunctionImporter::ImportMapTy &ImportList,
- const GVSummaryMapTy &DefinedGlobals,
- MapVector<StringRef, BitcodeModule> *ModuleMap, bool CodeGenOnly,
- const SmallVector<StringRef> &BitcodeLibFuncs,
- AddStreamFn IRAddStream = nullptr,
- const std::vector<uint8_t> &CmdArgs = std::vector<uint8_t>());
+LLVM_ABI Error thinBackend(
+ const Config &C, unsigned Task, AddStreamFn AddStream, Module &M,
+ const ModuleSummaryIndex &CombinedIndex,
+ const FunctionImporter::ImportMapTy &ImportList,
+ const GVSummaryMapTy &DefinedGlobals,
+ MapVector<StringRef, BitcodeModule> *ModuleMap, bool CodeGenOnly,
+ ArrayRef<StringRef> BitcodeLibFuncs, AddStreamFn IRAddStream = nullptr,
+ const std::vector<uint8_t> &CmdArgs = std::vector<uint8_t>());
LLVM_ABI Error finalizeOptimizationRemarks(LLVMRemarkFileHandle DiagOutputFile);
diff --git a/llvm/lib/LTO/LTO.cpp b/llvm/lib/LTO/LTO.cpp
index 0f3f3c344fb0a..35e5daa8a6cef 100644
--- a/llvm/lib/LTO/LTO.cpp
+++ b/llvm/lib/LTO/LTO.cpp
@@ -850,8 +850,9 @@ Error LTO::add(std::unique_ptr<InputFile> InputPtr,
return Error::success();
}
-void LTO::setBitcodeLibFuncs(const SmallVector<StringRef> &BitcodeLibFuncs) {
- this->BitcodeLibFuncs = BitcodeLibFuncs;
+void LTO::setBitcodeLibFuncs(ArrayRef<StringRef> BitcodeLibFuncs) {
+ this->BitcodeLibFuncs.clear();
+ this->BitcodeLibFuncs.append(BitcodeLibFuncs.begin(), BitcodeLibFuncs.end());
}
Expected<ArrayRef<SymbolResolution>>
@@ -1593,7 +1594,7 @@ class CGThinBackend : public ThinBackendProc {
class InProcessThinBackend : public CGThinBackend {
protected:
FileCache Cache;
- const SmallVector<StringRef> &BitcodeLibFuncs;
+ ArrayRef<StringRef> BitcodeLibFuncs;
public:
InProcessThinBackend(
@@ -1602,7 +1603,7 @@ class InProcessThinBackend : public CGThinBackend {
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache, lto::IndexWriteCallback OnWrite,
bool ShouldEmitIndexFiles, bool ShouldEmitImportsFiles,
- const SmallVector<StringRef> &BitcodeLibFuncs)
+ ArrayRef<StringRef> BitcodeLibFuncs)
: CGThinBackend(Conf, CombinedIndex, ModuleToDefinedGVSummaries,
AddStream, OnWrite, ShouldEmitIndexFiles,
ShouldEmitImportsFiles, ThinLTOParallelism),
@@ -1713,7 +1714,7 @@ class FirstRoundThinBackend : public InProcessThinBackend {
ThreadPoolStrategy ThinLTOParallelism,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn CGAddStream, FileCache CGCache,
- const SmallVector<StringRef> &BitcodeLibFuncs, AddStreamFn IRAddStream,
+ ArrayRef<StringRef> BitcodeLibFuncs, AddStreamFn IRAddStream,
FileCache IRCache)
: InProcessThinBackend(Conf, CombinedIndex, ThinLTOParallelism,
ModuleToDefinedGVSummaries, std::move(CGAddStream),
@@ -1809,7 +1810,7 @@ class SecondRoundThinBackend : public InProcessThinBackend {
ThreadPoolStrategy ThinLTOParallelism,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache,
- const SmallVector<StringRef> &BitcodeLibFuncs,
+ ArrayRef<StringRef> BitcodeLibFuncs,
std::unique_ptr<SmallVector<StringRef>> IRFiles,
stable_hash CombinedCGDataHash)
: InProcessThinBackend(Conf, CombinedIndex, ThinLTOParallelism,
@@ -1878,7 +1879,7 @@ ThinBackend lto::createInProcessThinBackend(ThreadPoolStrategy Parallelism,
[=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache,
- const SmallVector<StringRef> &BitcodeLibFuncs) {
+ ArrayRef<StringRef> BitcodeLibFuncs) {
return std::make_unique<InProcessThinBackend>(
Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
AddStream, Cache, OnWrite, ShouldEmitIndexFiles,
@@ -2000,7 +2001,7 @@ ThinBackend lto::createWriteIndexesThinBackend(
[=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache,
- const SmallVector<StringRef> &BitcodeLibFuncs) {
+ ArrayRef<StringRef> BitcodeLibFuncs) {
return std::make_unique<WriteIndexesThinBackend>(
Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
OldPrefix, NewPrefix, NativeObjectPrefix, ShouldEmitImportsFiles,
@@ -2800,7 +2801,7 @@ ThinBackend lto::createOutOfProcessThinBackend(
[=](const Config &Conf, ModuleSummaryIndex &CombinedIndex,
const DenseMap<StringRef, GVSummaryMapTy> &ModuleToDefinedGVSummaries,
AddStreamFn AddStream, FileCache Cache,
- const SmallVector<StringRef> &BitcodeLibFuncs) {
+ ArrayRef<StringRef> BitcodeLibFuncs) {
return std::make_unique<OutOfProcessThinBackend>(
Conf, CombinedIndex, Parallelism, ModuleToDefinedGVSummaries,
AddStream, Cache, OnWrite, ShouldEmitIndexFiles,
diff --git a/llvm/lib/LTO/LTOBackend.cpp b/llvm/lib/LTO/LTOBackend.cpp
index 8cd8a75c52cab..826313055cb82 100644
--- a/llvm/lib/LTO/LTOBackend.cpp
+++ b/llvm/lib/LTO/LTOBackend.cpp
@@ -408,7 +408,7 @@ bool lto::opt(const Config &Conf, TargetMachine *TM, unsigned Task, Module &Mod,
bool IsThinLTO, ModuleSummaryIndex *ExportSummary,
const ModuleSummaryIndex *ImportSummary,
const std::vector<uint8_t> &CmdArgs,
- const SmallVector<StringRef> &BitcodeLibFuncs) {
+ ArrayRef<StringRef> BitcodeLibFuncs) {
llvm::TimeTraceScope timeScope("opt");
if (EmbedBitcode == LTOBitcodeEmbedding::EmbedPostMergePreOptimized) {
// FIXME: the motivation for capturing post-merge bitcode and command line
@@ -604,7 +604,7 @@ Error lto::finalizeOptimizationRemarks(LLVMRemarkFileHandle DiagOutputFile) {
Error lto::backend(const Config &C, AddStreamFn AddStream,
unsigned ParallelCodeGenParallelismLevel, Module &Mod,
ModuleSummaryIndex &CombinedIndex,
- const SmallVector<StringRef> &BitcodeLibFuncs) {
+ ArrayRef<StringRef> BitcodeLibFuncs) {
llvm::TimeTraceScope timeScope("LTO backend");
Expected<const Target *> TOrErr = initAndLookupTarget(C, Mod);
if (!TOrErr)
@@ -656,8 +656,7 @@ Error lto::thinBackend(const Config &Conf, unsigned Task, AddStreamFn AddStream,
const FunctionImporter::ImportMapTy &ImportList,
const GVSummaryMapTy &DefinedGlobals,
MapVector<StringRef, BitcodeModule> *ModuleMap,
- bool CodeGenOnly,
- const SmallVector<StringRef> &BitcodeLibFuncs,
+ bool CodeGenOnly, ArrayRef<StringRef> BitcodeLibFuncs,
AddStreamFn IRAddStream,
const std::vector<uint8_t> &CmdArgs) {
llvm::TimeTraceScope timeScope("Thin backend", Mod.getModuleIdentifier());
>From fb7f0c8d1f8c5c73e7af81fda6fa12364434286e Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 15 Jan 2026 11:17:28 -0800
Subject: [PATCH 07/22] More idiomatic LibFunc iteration
---
llvm/lib/LTO/LTO.cpp | 3 +--
1 file changed, 1 insertion(+), 2 deletions(-)
diff --git a/llvm/lib/LTO/LTO.cpp b/llvm/lib/LTO/LTO.cpp
index 35e5daa8a6cef..462873a4b484b 100644
--- a/llvm/lib/LTO/LTO.cpp
+++ b/llvm/lib/LTO/LTO.cpp
@@ -1508,8 +1508,7 @@ SmallVector<StringRef> LTO::getLibFuncSymbols(const Triple &TT,
TargetLibraryInfo TLI(*TLII);
SmallVector<StringRef> LibFuncSymbols;
LibFuncSymbols.reserve(LibFunc::NumLibFuncs);
- for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs); I != E;
- ++I) {
+ for (unsigned I = LibFunc::Begin_LibFunc; I != LibFunc::End_LibFunc; ++I) {
LibFunc F = static_cast<LibFunc>(I);
if (TLI.has(F))
LibFuncSymbols.push_back(Saver.save(TLI.getName(F)).data());
>From da7fdf5d37f584e8e05e982234614f97ffd5fcc0 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 20 Jan 2026 14:54:57 -0800
Subject: [PATCH 08/22] Add release note
---
llvm/docs/ReleaseNotes.md | 6 ++++++
1 file changed, 6 insertions(+)
diff --git a/llvm/docs/ReleaseNotes.md b/llvm/docs/ReleaseNotes.md
index 2053e087ba23e..0d709fec08dc3 100644
--- a/llvm/docs/ReleaseNotes.md
+++ b/llvm/docs/ReleaseNotes.md
@@ -86,6 +86,12 @@ Changes to LLVM infrastructure
* The ``Br`` opcode was split into two opcodes separating unconditional
(``UncondBr``) and conditional (``CondBr``) branches.
+* Bitcode libraries can now implement compiler-managed library functions
+ (libcalls) without causing incorrect API manipulation or undefined references
+ ([#177046](https://github.com/llvm/llvm-project/pull/125687)). Note that
+ there are still issues with invalid compiler reasoning about some functions
+ in bitcode, e.g. `malloc`. Not yet supported on MachO or when using DTLTO.
+
Changes to building LLVM
------------------------
>From f140e65f33519ad250f56353b98e8f6ca2a3c48d Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 22 Jan 2026 11:59:12 -0800
Subject: [PATCH 09/22] Add FIXME for -bitcode-libfuncs clunkiness
---
llvm/tools/llvm-lto2/llvm-lto2.cpp | 5 +++++
1 file changed, 5 insertions(+)
diff --git a/llvm/tools/llvm-lto2/llvm-lto2.cpp b/llvm/tools/llvm-lto2/llvm-lto2.cpp
index 66fa5789feb64..49196a94bcb96 100644
--- a/llvm/tools/llvm-lto2/llvm-lto2.cpp
+++ b/llvm/tools/llvm-lto2/llvm-lto2.cpp
@@ -234,6 +234,11 @@ static cl::opt<bool>
AllVtablesHaveTypeInfos("all-vtables-have-type-infos", cl::Hidden,
cl::desc("All vtables have type infos"));
+// FIXME: Listing all bitcode libfunc symbols here is clunky. A higher-level way
+// to indicate which TUs made it into the link might be better, but this would
+// required more detailed tracking of the sources of constructs in the IR.
+// Alternatively, there may be some other data structure that could hold this
+// information.
static cl::list<std::string>
BitcodeLibFuncs("bitcode-libfuncs", cl::Hidden,
cl::desc("set of libfuncs implemented in bitcode"));
>From 6e75c8fa968c81f77518dd8cbfdcd4f79556127c Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Mon, 2 Feb 2026 12:15:04 -0800
Subject: [PATCH 10/22] Remove added Target dep; missed.
---
llvm/lib/Object/CMakeLists.txt | 1 -
1 file changed, 1 deletion(-)
diff --git a/llvm/lib/Object/CMakeLists.txt b/llvm/lib/Object/CMakeLists.txt
index c48d251249488..0f6d2f7c59a5c 100644
--- a/llvm/lib/Object/CMakeLists.txt
+++ b/llvm/lib/Object/CMakeLists.txt
@@ -51,7 +51,6 @@ add_llvm_component_library(LLVMObject
BinaryFormat
MCParser
Support
- Target
TargetParser
TextAPI
)
>From 1ce9fceb040cbb267f1156a00f6293d764508720 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 10 Feb 2026 13:55:18 -0800
Subject: [PATCH 11/22] Don't make libcalls GC roots in COFF; only @llvm.used
---
lld/COFF/InputFiles.cpp | 2 +-
1 file changed, 1 insertion(+), 1 deletion(-)
diff --git a/lld/COFF/InputFiles.cpp b/lld/COFF/InputFiles.cpp
index 0a40309174a67..3041e1ce6b445 100644
--- a/lld/COFF/InputFiles.cpp
+++ b/lld/COFF/InputFiles.cpp
@@ -1449,7 +1449,7 @@ void BitcodeFile::parse() {
symtab.addRegular(this, symName, nullptr, fakeSC, 0, objSym.isWeak());
}
symbols.push_back(sym);
- if (objSym.isUsed() || objSym.isLibcall(tli, libcalls))
+ if (objSym.isUsed())
symtab.ctx.config.gcroot.push_back(sym);
}
directives = saver.save(obj->getCOFFLinkerOpts());
>From 13035b6834ced3b5ed780dc2afc969a8da832e6e Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 10 Feb 2026 14:52:31 -0800
Subject: [PATCH 12/22] Merge two libcall-external-* tests into
libcall-external.ll
---
.../X86/libcall-external-bitcode.ll | 20 -------------
.../X86/libcall-external-not-bitcode.ll | 20 -------------
.../LTO/Resolution/X86/libcall-external.ll | 29 +++++++++++++++++++
3 files changed, 29 insertions(+), 40 deletions(-)
delete mode 100644 llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll
delete mode 100644 llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll
create mode 100644 llvm/test/LTO/Resolution/X86/libcall-external.ll
diff --git a/llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll b/llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll
deleted file mode 100644
index 95a599fe75e8b..0000000000000
--- a/llvm/test/LTO/Resolution/X86/libcall-external-bitcode.ll
+++ /dev/null
@@ -1,20 +0,0 @@
-; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
-; RUN: llvm-lto2 run -o %t2 \
-; RUN: -r %t,foo,plx \
-; RUN: -r %t,memcmp,x \
-; RUN: -r %t,bcmp,pl --bitcode-libfuncs=bcmp %t -save-temps
-; RUN: llvm-dis %t2.1.4.opt.bc -o - | FileCheck %s
-
-define i1 @foo(ptr %0, ptr %1, i64 %2) {
- ; CHECK-LABEL: define{{.*}}i1 @foo
- ; CHECK-NEXT: %cmp = {{.*}}call i32 @memcmp
- ; CHECK-NEXT: %eq = icmp eq i32 %cmp, 0
- ; CHECK-NEXT: ret i1 %eq
-
- %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
- %eq = icmp eq i32 %cmp, 0
- ret i1 %eq
-}
-
-declare i32 @memcmp(ptr, ptr, i64)
-declare i32 @bcmp(ptr, ptr, i64)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll b/llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll
deleted file mode 100644
index 2e6cc798d22cd..0000000000000
--- a/llvm/test/LTO/Resolution/X86/libcall-external-not-bitcode.ll
+++ /dev/null
@@ -1,20 +0,0 @@
-; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
-; RUN: llvm-lto2 run -o %t2 \
-; RUN: -r %t,foo,plx \
-; RUN: -r %t,memcmp,x \
-; RUN: -r %t,bcmp,pl %t -save-temps
-; RUN: llvm-dis %t2.1.4.opt.bc -o - | FileCheck %s
-
-define i1 @foo(ptr %0, ptr %1, i64 %2) {
- ; CHECK-LABEL: define{{.*}}i1 @foo
- ; CHECK-NEXT: %bcmp = {{.*}}call i32 @bcmp
- ; CHECK-NEXT: %eq = icmp eq i32 %bcmp, 0
- ; CHECK-NEXT: ret i1 %eq
-
- %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
- %eq = icmp eq i32 %cmp, 0
- ret i1 %eq
-}
-
-declare i32 @memcmp(ptr, ptr, i64)
-declare i32 @bcmp(ptr, ptr, i64)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-external.ll b/llvm/test/LTO/Resolution/X86/libcall-external.ll
new file mode 100644
index 0000000000000..8e466616f01c1
--- /dev/null
+++ b/llvm/test/LTO/Resolution/X86/libcall-external.ll
@@ -0,0 +1,29 @@
+;; When a libcall was not brought into the link, it can be used iff it is
+;; defined in native code, not bitcode.
+; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
+; RUN: llvm-lto2 run -o %t.bitcode \
+; RUN: -r %t,foo,plx \
+; RUN: -r %t,memcmp,x \
+; RUN: -r %t,bcmp,pl --bitcode-libfuncs=bcmp %t -save-temps
+; RUN: llvm-dis %t.bitcode.1.4.opt.bc -o - | FileCheck --check-prefixes=CHECK,BITCODE %s
+; RUN: llvm-lto2 run -o %t.native \
+; RUN: -r %t,foo,plx \
+; RUN: -r %t,memcmp,x \
+; RUN: -r %t,bcmp,pl %t -save-temps
+; RUN: llvm-dis %t.native.1.4.opt.bc -o - | FileCheck --check-prefixes=CHECK,NATIVE %s
+define i1 @foo(ptr %0, ptr %1, i64 %2) {
+ ; CHECK-LABEL: define{{.*}}i1 @foo
+ ; BITCODE-NEXT: %cmp = {{.*}}call i32 @memcmp
+ ; BITCODE-NEXT: %eq = icmp eq i32 %cmp, 0
+ ; NATIVE-NEXT: %bcmp = {{.*}}call i32 @bcmp
+ ; NATIVE-NEXT: %eq = icmp eq i32 %bcmp, 0
+ ; CHECK-NEXT: ret i1 %eq
+
+
+ %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
+ %eq = icmp eq i32 %cmp, 0
+ ret i1 %eq
+}
+
+declare i32 @memcmp(ptr, ptr, i64)
+declare i32 @bcmp(ptr, ptr, i64)
>From e51dbccee20ca7bd55a28223614677caf268c998 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 10 Feb 2026 15:34:58 -0800
Subject: [PATCH 13/22] Document the semantics of the new flag and data
structure
---
llvm/include/llvm/LTO/LTO.h | 4 ++++
llvm/tools/llvm-lto2/llvm-lto2.cpp | 5 +++++
2 files changed, 9 insertions(+)
diff --git a/llvm/include/llvm/LTO/LTO.h b/llvm/include/llvm/LTO/LTO.h
index bb5d38da8d9b6..d78a9dfcfd1bf 100644
--- a/llvm/include/llvm/LTO/LTO.h
+++ b/llvm/include/llvm/LTO/LTO.h
@@ -671,6 +671,10 @@ class LTO {
// Setup optimization remarks according to the provided configuration.
Error setupOptimizationRemarks();
+ // LibFuncs that could have been part of the LTO unit, but was not typically
+ // because they weren't extracted from their libraries. Such functions cannot
+ // safely be called, since they have already lost their only opportunity to be
+ // defined.
SmallVector<StringRef> BitcodeLibFuncs;
public:
diff --git a/llvm/tools/llvm-lto2/llvm-lto2.cpp b/llvm/tools/llvm-lto2/llvm-lto2.cpp
index 49196a94bcb96..ede47379672f8 100644
--- a/llvm/tools/llvm-lto2/llvm-lto2.cpp
+++ b/llvm/tools/llvm-lto2/llvm-lto2.cpp
@@ -234,6 +234,11 @@ static cl::opt<bool>
AllVtablesHaveTypeInfos("all-vtables-have-type-infos", cl::Hidden,
cl::desc("All vtables have type infos"));
+// Specifying a function here states that it is a library function that could
+// have been part of the LTO unit, but was not, typically because it wasn't
+// extracted from a library. Such functions cannot safely be called, since they
+// have already lost their only opportunity to be defined.
+//
// FIXME: Listing all bitcode libfunc symbols here is clunky. A higher-level way
// to indicate which TUs made it into the link might be better, but this would
// required more detailed tracking of the sources of constructs in the IR.
>From 221d338938b01e7d089de0a1e77e10b11147c622 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Tue, 10 Feb 2026 16:01:24 -0800
Subject: [PATCH 14/22] Document the LTOBackend implementation decisionmaking
---
llvm/lib/LTO/LTOBackend.cpp | 8 +++++++-
1 file changed, 7 insertions(+), 1 deletion(-)
diff --git a/llvm/lib/LTO/LTOBackend.cpp b/llvm/lib/LTO/LTOBackend.cpp
index 826313055cb82..3d8d8ec5e963f 100644
--- a/llvm/lib/LTO/LTOBackend.cpp
+++ b/llvm/lib/LTO/LTOBackend.cpp
@@ -303,6 +303,10 @@ static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
if (Conf.Freestanding)
TLII->disableAllFunctions();
+ // Determine whether or not its safe to emit calls to each libfunc. Libfuncs
+ // that might have been present in the current LTO unit, but are not, have
+ // lost their only opportunity to be defined, and calls must not be emitted to
+ // them.
TargetLibraryInfo TLI(*TLII);
for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs); I != E;
++I) {
@@ -314,7 +318,9 @@ static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
if (Val && !Val->isDeclaration())
continue;
- // LibFuncs not implemented in bitcode can always be referenced.
+ // LibFuncs not implemented in bitcode can always be referenced, since they
+ // can safely be extracted from whatever library they reside in after LTO
+ // without changing the linker symbol table that LTO depends on.
if (!BitcodeLibFuncs.contains(Name))
continue;
>From 53756408cd73f82e66b00805ee8393566fff3956 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 19 Feb 2026 14:18:11 -0800
Subject: [PATCH 15/22] Correct semantics of libcall tests
---
.../LTO/Resolution/X86/libcall-external.ll | 20 ++++++++-----------
llvm/test/LTO/Resolution/X86/libcall-in-tu.ll | 14 ++++++-------
2 files changed, 15 insertions(+), 19 deletions(-)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-external.ll b/llvm/test/LTO/Resolution/X86/libcall-external.ll
index 8e466616f01c1..46756b387668c 100644
--- a/llvm/test/LTO/Resolution/X86/libcall-external.ll
+++ b/llvm/test/LTO/Resolution/X86/libcall-external.ll
@@ -1,16 +1,13 @@
;; When a libcall was not brought into the link, it can be used iff it is
;; defined in native code, not bitcode.
-; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
-; RUN: llvm-lto2 run -o %t.bitcode \
-; RUN: -r %t,foo,plx \
-; RUN: -r %t,memcmp,x \
-; RUN: -r %t,bcmp,pl --bitcode-libfuncs=bcmp %t -save-temps
-; RUN: llvm-dis %t.bitcode.1.4.opt.bc -o - | FileCheck --check-prefixes=CHECK,BITCODE %s
-; RUN: llvm-lto2 run -o %t.native \
-; RUN: -r %t,foo,plx \
-; RUN: -r %t,memcmp,x \
-; RUN: -r %t,bcmp,pl %t -save-temps
-; RUN: llvm-dis %t.native.1.4.opt.bc -o - | FileCheck --check-prefixes=CHECK,NATIVE %s
+; RUN: opt %s -o %t.o -mtriple x86_64-unknown-linux-musl
+; RUN: llvm-lto2 run -o %t.bitcode.o \
+; RUN: -r %t.o,foo,plx -r %t.o,memcmp,x -save-temps %t.o \
+; RUN: --bitcode-libfuncs=bcmp
+; RUN: llvm-dis %t.bitcode.o.0.4.opt.bc -o - | FileCheck --check-prefixes=CHECK,BITCODE %s
+; RUN: llvm-lto2 run -o %t.native.o \
+; RUN: -r %t.o,foo,plx -r %t.o,memcmp,x -save-temps %t.o
+; RUN: llvm-dis %t.native.o.0.4.opt.bc -o - | FileCheck --check-prefixes=CHECK,NATIVE %s
define i1 @foo(ptr %0, ptr %1, i64 %2) {
; CHECK-LABEL: define{{.*}}i1 @foo
; BITCODE-NEXT: %cmp = {{.*}}call i32 @memcmp
@@ -26,4 +23,3 @@ define i1 @foo(ptr %0, ptr %1, i64 %2) {
}
declare i32 @memcmp(ptr, ptr, i64)
-declare i32 @bcmp(ptr, ptr, i64)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll b/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
index 948f21a6536ca..33d1dc47ff0d1 100644
--- a/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
+++ b/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
@@ -3,13 +3,13 @@
;; a call to bcmp() as part of SimplifyLibCalls. Such deletions must not be
;; allowed.
-; RUN: opt %s -o %t -module-summary -mtriple x86_64-unknown-linux-musl
-; RUN: llvm-lto2 run -o %t2 \
-; RUN: -r %t,foo,plx \
-; RUN: -r %t,memcmp,x \
-; RUN: -r %t,bcmp,pl \
-; RUN: -r %t,bcmp_impl,x %t -save-temps
-; RUN: llvm-dis %t2.1.4.opt.bc -o - | FileCheck %s
+; RUN: opt %s -o %t.o -mtriple x86_64-unknown-linux-musl
+; RUN: llvm-lto2 run -o %t.lto.o \
+; RUN: -r %t.o,foo,plx \
+; RUN: -r %t.o,memcmp,x \
+; RUN: -r %t.o,bcmp,pl \
+; RUN: -r %t.o,bcmp_impl,x %t.o --bitcode-libfuncs=bcmp -save-temps
+; RUN: llvm-dis %t.lto.o.0.4.opt.bc -o - | FileCheck %s
define i1 @foo(ptr %0, ptr %1, i64 %2) {
; CHECK-LABEL: define{{.*}}i1 @foo
>From 3b8c9af9f993720cac4d5aeb5a2d62155812ea26 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 19 Feb 2026 15:36:46 -0800
Subject: [PATCH 16/22] Correct description of --bitcode-libfuncs flag.
---
llvm/tools/llvm-lto2/llvm-lto2.cpp | 8 ++++----
1 file changed, 4 insertions(+), 4 deletions(-)
diff --git a/llvm/tools/llvm-lto2/llvm-lto2.cpp b/llvm/tools/llvm-lto2/llvm-lto2.cpp
index ede47379672f8..054369b289bb2 100644
--- a/llvm/tools/llvm-lto2/llvm-lto2.cpp
+++ b/llvm/tools/llvm-lto2/llvm-lto2.cpp
@@ -234,10 +234,10 @@ static cl::opt<bool>
AllVtablesHaveTypeInfos("all-vtables-have-type-infos", cl::Hidden,
cl::desc("All vtables have type infos"));
-// Specifying a function here states that it is a library function that could
-// have been part of the LTO unit, but was not, typically because it wasn't
-// extracted from a library. Such functions cannot safely be called, since they
-// have already lost their only opportunity to be defined.
+// Specifying a symbol here states that it is a library symbol with a prevailing
+// definition in bitcode. If such symbols are not defined somewhere in the LTO
+// unit, (e.g., they were not extracted), they cannot safely be referenced,
+// since they have already lost their only opportunity to be defined.
//
// FIXME: Listing all bitcode libfunc symbols here is clunky. A higher-level way
// to indicate which TUs made it into the link might be better, but this would
>From 853bb38c338fa756059ffbb1db0b47765c9fd3f5 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 19 Feb 2026 14:22:18 -0800
Subject: [PATCH 17/22] Fix ThinLTO by filtering BitcodeLibfuncs in run()
---
llvm/include/llvm/LTO/LTO.h | 3 +-
llvm/lib/LTO/LTO.cpp | 8 +++++
llvm/lib/LTO/LTOBackend.cpp | 19 ++--------
.../Resolution/X86/libcall-in-thin-link.ll | 35 +++++++++++++++++++
4 files changed, 47 insertions(+), 18 deletions(-)
create mode 100644 llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll
diff --git a/llvm/include/llvm/LTO/LTO.h b/llvm/include/llvm/LTO/LTO.h
index d78a9dfcfd1bf..72e1f45965f51 100644
--- a/llvm/include/llvm/LTO/LTO.h
+++ b/llvm/include/llvm/LTO/LTO.h
@@ -674,7 +674,8 @@ class LTO {
// LibFuncs that could have been part of the LTO unit, but was not typically
// because they weren't extracted from their libraries. Such functions cannot
// safely be called, since they have already lost their only opportunity to be
- // defined.
+ // defined. Before run(), this contains all libfuncs defined anywhere in
+ // bitcode; during run(), the ones defined in the LTO unit are filtered out.
SmallVector<StringRef> BitcodeLibFuncs;
public:
diff --git a/llvm/lib/LTO/LTO.cpp b/llvm/lib/LTO/LTO.cpp
index 462873a4b484b..859d8c69e5401 100644
--- a/llvm/lib/LTO/LTO.cpp
+++ b/llvm/lib/LTO/LTO.cpp
@@ -1334,6 +1334,14 @@ Error LTO::run(AddStreamFn AddStream, FileCache Cache) {
computeDeadSymbolsWithConstProp(ThinLTO.CombinedIndex, GUIDPreservedSymbols,
isPrevailing, Conf.OptLevel > 0);
+ // Remove any prevailing definitions from BitcodeLibfuncs, as these are safe
+ // to call.
+ llvm::erase_if(BitcodeLibFuncs, [&](StringRef Name) {
+ return isPrevailing(GlobalValue::getGUIDAssumingExternalLinkage(
+ GlobalValue::dropLLVMManglingEscape(Name))) ==
+ PrevailingType::Yes;
+ });
+
// Setup output file to emit statistics.
auto StatsFileOrErr = setupStatsFile(Conf.StatsFile);
if (!StatsFileOrErr)
diff --git a/llvm/lib/LTO/LTOBackend.cpp b/llvm/lib/LTO/LTOBackend.cpp
index 3d8d8ec5e963f..0b0257163dcd2 100644
--- a/llvm/lib/LTO/LTOBackend.cpp
+++ b/llvm/lib/LTO/LTOBackend.cpp
@@ -311,23 +311,8 @@ static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs); I != E;
++I) {
LibFunc F = static_cast<LibFunc>(I);
- StringRef Name = TLI.getName(F);
- GlobalValue *Val = Mod.getNamedValue(Name);
-
- // LibFuncs present in the current TU can always be referenced.
- if (Val && !Val->isDeclaration())
- continue;
-
- // LibFuncs not implemented in bitcode can always be referenced, since they
- // can safely be extracted from whatever library they reside in after LTO
- // without changing the linker symbol table that LTO depends on.
- if (!BitcodeLibFuncs.contains(Name))
- continue;
-
- // FIXME: Functions that are somewhere in a ThinLTO link (just not imported
- // in this module) should not be disabled, as they have already been
- // extracted.
- TLII->setUnavailable(F);
+ if (BitcodeLibFuncs.contains(TLI.getName(F)))
+ TLII->setUnavailable(F);
}
FAM.registerPass([&] { return TargetLibraryAnalysis(*TLII); });
diff --git a/llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll b/llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll
new file mode 100644
index 0000000000000..962088094c438
--- /dev/null
+++ b/llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll
@@ -0,0 +1,35 @@
+;; If a libcall was extracted in a thin link, it can be used even if not
+;; present in the current TU.
+
+; RUN: rm -rf %t && split-file %s %t && cd %t
+; RUN: opt foo.ll -o foo.o -module-summary -mtriple x86_64-unknown-linux-musl
+; RUN: opt bcmp.ll -o bcmp.o -module-summary -mtriple x86_64-unknown-linux-musl
+; RUN: llvm-lto2 run -o lto.o \
+; RUN: -r foo.o,foo,plx \
+; RUN: -r foo.o,memcmp,x \
+; RUN: -r bcmp.o,bcmp,pl \
+; RUN: -r bcmp.o,bcmp_impl,x --bitcode-libfuncs=bcmp foo.o bcmp.o -save-temps
+; RUN: llvm-dis lto.o.1.4.opt.bc -o - | FileCheck %s
+
+;--- foo.ll
+define i1 @foo(ptr %0, ptr %1, i64 %2) {
+ ; CHECK-LABEL: define{{.*}}i1 @foo
+ ; CHECK-NEXT: %bcmp = {{.*}}call i32 @bcmp
+ ; CHECK-NEXT: %eq = icmp eq i32 %bcmp, 0
+ ; CHECK-NEXT: ret i1 %eq
+
+ %cmp = call i32 @memcmp(ptr %0, ptr %1, i64 %2)
+ %eq = icmp eq i32 %cmp, 0
+ ret i1 %eq
+}
+
+declare i32 @memcmp(ptr, ptr, i64)
+
+;--- bcmp.ll
+define i32 @bcmp(ptr %0, ptr %1, i64 %2) noinline {
+ %r = call i32 @bcmp_impl(ptr %0, ptr %1, i64 %2)
+ ret i32 %r
+}
+
+declare i32 @bcmp_impl(ptr, ptr, i64)
+
>From d80373b32635f3856ff56ff4c4a2d4fd4c57eec3 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 11 Mar 2026 13:40:24 -0700
Subject: [PATCH 18/22] Change BitcodeLibFuncs semantics to only track
unextracted symbols
---
lld/ELF/Driver.cpp | 2 +-
llvm/include/llvm/LTO/LTO.h | 14 ++++++--------
llvm/lib/LTO/LTO.cpp | 8 --------
.../LTO/Resolution/X86/libcall-in-thin-link.ll | 2 +-
llvm/test/LTO/Resolution/X86/libcall-in-tu.ll | 2 +-
llvm/tools/llvm-lto2/llvm-lto2.cpp | 15 +++++++--------
6 files changed, 16 insertions(+), 27 deletions(-)
diff --git a/lld/ELF/Driver.cpp b/lld/ELF/Driver.cpp
index 476ccc9433e3a..b536d895bf657 100644
--- a/lld/ELF/Driver.cpp
+++ b/lld/ELF/Driver.cpp
@@ -2722,7 +2722,7 @@ void LinkerDriver::compileBitcodeFiles(bool skipLinkedOutput) {
SmallVector<StringRef> bitcodeLibFuncs;
for (StringRef libFunc : lto::LTO::getLibFuncSymbols(tt, saver))
if (Symbol *sym = ctx.symtab->find(libFunc);
- sym && isa<BitcodeFile>(sym->file))
+ sym && sym->isLazy() && isa<BitcodeFile>(sym->file))
bitcodeLibFuncs.push_back(libFunc);
lto->setBitcodeLibFuncs(bitcodeLibFuncs);
}
diff --git a/llvm/include/llvm/LTO/LTO.h b/llvm/include/llvm/LTO/LTO.h
index 72e1f45965f51..b9ae6fb9a3de0 100644
--- a/llvm/include/llvm/LTO/LTO.h
+++ b/llvm/include/llvm/LTO/LTO.h
@@ -447,9 +447,9 @@ class LTO {
LLVM_ABI Error add(std::unique_ptr<InputFile> Obj,
ArrayRef<SymbolResolution> Res);
- /// Set the list of functions implemented in bitcode across the link, whether
- /// extracted or not. Such functions may not be referenced if they were not
- /// extracted by the time LTO occurs.
+ /// Set the list of functions implemented in bitcode that were not extracted
+ /// from an archive. Such functions may not be referenced, as they have
+ /// lost their opportunity to be defined.
LLVM_ABI void setBitcodeLibFuncs(ArrayRef<StringRef> BitcodeLibFuncs);
/// Returns an upper bound on the number of tasks that the client may expect.
@@ -671,11 +671,9 @@ class LTO {
// Setup optimization remarks according to the provided configuration.
Error setupOptimizationRemarks();
- // LibFuncs that could have been part of the LTO unit, but was not typically
- // because they weren't extracted from their libraries. Such functions cannot
- // safely be called, since they have already lost their only opportunity to be
- // defined. Before run(), this contains all libfuncs defined anywhere in
- // bitcode; during run(), the ones defined in the LTO unit are filtered out.
+ // LibFuncs that were implemented in bitcode but were not extracted
+ // from their libraries. Such functions cannot safely be called, since
+ // they have lost their opportunity to be defined.
SmallVector<StringRef> BitcodeLibFuncs;
public:
diff --git a/llvm/lib/LTO/LTO.cpp b/llvm/lib/LTO/LTO.cpp
index 859d8c69e5401..462873a4b484b 100644
--- a/llvm/lib/LTO/LTO.cpp
+++ b/llvm/lib/LTO/LTO.cpp
@@ -1334,14 +1334,6 @@ Error LTO::run(AddStreamFn AddStream, FileCache Cache) {
computeDeadSymbolsWithConstProp(ThinLTO.CombinedIndex, GUIDPreservedSymbols,
isPrevailing, Conf.OptLevel > 0);
- // Remove any prevailing definitions from BitcodeLibfuncs, as these are safe
- // to call.
- llvm::erase_if(BitcodeLibFuncs, [&](StringRef Name) {
- return isPrevailing(GlobalValue::getGUIDAssumingExternalLinkage(
- GlobalValue::dropLLVMManglingEscape(Name))) ==
- PrevailingType::Yes;
- });
-
// Setup output file to emit statistics.
auto StatsFileOrErr = setupStatsFile(Conf.StatsFile);
if (!StatsFileOrErr)
diff --git a/llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll b/llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll
index 962088094c438..fe8e492580d42 100644
--- a/llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll
+++ b/llvm/test/LTO/Resolution/X86/libcall-in-thin-link.ll
@@ -8,7 +8,7 @@
; RUN: -r foo.o,foo,plx \
; RUN: -r foo.o,memcmp,x \
; RUN: -r bcmp.o,bcmp,pl \
-; RUN: -r bcmp.o,bcmp_impl,x --bitcode-libfuncs=bcmp foo.o bcmp.o -save-temps
+; RUN: -r bcmp.o,bcmp_impl,x foo.o bcmp.o -save-temps
; RUN: llvm-dis lto.o.1.4.opt.bc -o - | FileCheck %s
;--- foo.ll
diff --git a/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll b/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
index 33d1dc47ff0d1..cddf5d30697d0 100644
--- a/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
+++ b/llvm/test/LTO/Resolution/X86/libcall-in-tu.ll
@@ -8,7 +8,7 @@
; RUN: -r %t.o,foo,plx \
; RUN: -r %t.o,memcmp,x \
; RUN: -r %t.o,bcmp,pl \
-; RUN: -r %t.o,bcmp_impl,x %t.o --bitcode-libfuncs=bcmp -save-temps
+; RUN: -r %t.o,bcmp_impl,x %t.o -save-temps
; RUN: llvm-dis %t.lto.o.0.4.opt.bc -o - | FileCheck %s
define i1 @foo(ptr %0, ptr %1, i64 %2) {
diff --git a/llvm/tools/llvm-lto2/llvm-lto2.cpp b/llvm/tools/llvm-lto2/llvm-lto2.cpp
index 054369b289bb2..2b89b20a87757 100644
--- a/llvm/tools/llvm-lto2/llvm-lto2.cpp
+++ b/llvm/tools/llvm-lto2/llvm-lto2.cpp
@@ -234,19 +234,18 @@ static cl::opt<bool>
AllVtablesHaveTypeInfos("all-vtables-have-type-infos", cl::Hidden,
cl::desc("All vtables have type infos"));
-// Specifying a symbol here states that it is a library symbol with a prevailing
-// definition in bitcode. If such symbols are not defined somewhere in the LTO
-// unit, (e.g., they were not extracted), they cannot safely be referenced,
-// since they have already lost their only opportunity to be defined.
+// Specifying a symbol here states that it is a library symbol that had a
+// definition in bitcode, but was not extracted. Such symbols cannot safely
+// be referenced, since they have already lost their opportunity to be defined.
//
// FIXME: Listing all bitcode libfunc symbols here is clunky. A higher-level way
// to indicate which TUs made it into the link might be better, but this would
-// required more detailed tracking of the sources of constructs in the IR.
+// require more detailed tracking of the sources of constructs in the IR.
// Alternatively, there may be some other data structure that could hold this
// information.
-static cl::list<std::string>
- BitcodeLibFuncs("bitcode-libfuncs", cl::Hidden,
- cl::desc("set of libfuncs implemented in bitcode"));
+static cl::list<std::string> BitcodeLibFuncs(
+ "bitcode-libfuncs", cl::Hidden,
+ cl::desc("set of unextracted libfuncs implemented in bitcode"));
static cl::opt<bool> TimeTrace("time-trace", cl::desc("Record time trace"));
>From 155f6a01062a077303c12e65f88b302f3dab6468 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Mon, 9 Mar 2026 20:54:40 -0700
Subject: [PATCH 19/22] Port to COFF
---
lld/COFF/LTO.cpp | 4 ++
lld/COFF/LTO.h | 1 +
lld/COFF/SymbolTable.cpp | 21 ++++++++
.../COFF/lto-libcall-archive-bitcode.test | 51 +++++++++++++++++++
4 files changed, 77 insertions(+)
create mode 100644 lld/test/COFF/lto-libcall-archive-bitcode.test
diff --git a/lld/COFF/LTO.cpp b/lld/COFF/LTO.cpp
index d55f95493a85f..d5f643782f615 100644
--- a/lld/COFF/LTO.cpp
+++ b/lld/COFF/LTO.cpp
@@ -280,3 +280,7 @@ std::vector<InputFile *> BitcodeCompiler::compile() {
return ret;
}
+
+void BitcodeCompiler::setBitcodeLibFuncs(ArrayRef<StringRef> bitcodeLibFuncs) {
+ ltoObj->setBitcodeLibFuncs(bitcodeLibFuncs);
+}
diff --git a/lld/COFF/LTO.h b/lld/COFF/LTO.h
index 6826251b5ffa7..73e855e567b09 100644
--- a/lld/COFF/LTO.h
+++ b/lld/COFF/LTO.h
@@ -45,6 +45,7 @@ class BitcodeCompiler {
void add(BitcodeFile &f);
std::vector<InputFile *> compile();
+ void setBitcodeLibFuncs(ArrayRef<StringRef> bitcodeLibFuncs);
private:
std::unique_ptr<llvm::lto::LTO> ltoObj;
diff --git a/lld/COFF/SymbolTable.cpp b/lld/COFF/SymbolTable.cpp
index 38a43390c15ab..4c381235de48b 100644
--- a/lld/COFF/SymbolTable.cpp
+++ b/lld/COFF/SymbolTable.cpp
@@ -1437,6 +1437,10 @@ void SymbolTable::compileBitcodeFiles() {
if (bitcodeFileInstances.empty())
return;
+ // Note that this assumes that the set of possible libfuncs is roughly
+ // equivalent for all bitcode translation units.
+ llvm::Triple tt =
+ llvm::Triple(bitcodeFileInstances.front()->obj->getTargetTriple());
ScopedTimer t(ctx.ltoTimer);
lto.reset(new BitcodeCompiler(ctx));
{
@@ -1444,6 +1448,23 @@ void SymbolTable::compileBitcodeFiles() {
for (BitcodeFile *f : bitcodeFileInstances)
lto->add(*f);
}
+
+ llvm::BumpPtrAllocator alloc;
+ llvm::StringSaver saver(alloc);
+ SmallVector<StringRef> bitcodeLibFuncs;
+ for (StringRef libFunc : lto::LTO::getLibFuncSymbols(tt, saver)) {
+ if (Symbol *sym = find(libFunc)) {
+ if (auto *l = dyn_cast<LazyArchive>(sym)) {
+ if (isBitcode(l->getMemberBuffer()))
+ bitcodeLibFuncs.push_back(libFunc);
+ } else if (auto *o = dyn_cast<LazyObject>(sym)) {
+ if (isBitcode(o->file->mb))
+ bitcodeLibFuncs.push_back(libFunc);
+ }
+ }
+ }
+ lto->setBitcodeLibFuncs(bitcodeLibFuncs);
+
for (InputFile *newObj : lto->compile()) {
ObjFile *obj = cast<ObjFile>(newObj);
obj->parse();
diff --git a/lld/test/COFF/lto-libcall-archive-bitcode.test b/lld/test/COFF/lto-libcall-archive-bitcode.test
new file mode 100644
index 0000000000000..bd172f1ff3955
--- /dev/null
+++ b/lld/test/COFF/lto-libcall-archive-bitcode.test
@@ -0,0 +1,51 @@
+; REQUIRES: x86
+
+; RUN: rm -rf %t && split-file %s %t && cd %t
+; RUN: llvm-as main.ll -o main.obj
+; RUN: llvm-as puts.ll -o puts.obj
+; RUN: llvm-mc -filetype=obj -triple=x86_64-pc-windows-msvc printf.s -o printf.obj
+; RUN: llvm-ar rcs libc.lib puts.obj printf.obj
+
+;; Ensure that no printf->puts translation occurs during LTO because puts is in
+;; bitcode, but was not brought into the link. This would fail the link by
+;; extracting bitcode after LTO.
+; RUN: lld-link -out:out.exe -entry:main -subsystem:console -lldmap:- -nodefaultlib main.obj libc.lib | FileCheck %s
+
+;; Test the same behavior with lazy objects.
+; RUN: lld-link -out:out-lazy.exe -entry:main -subsystem:console -lldmap:- -nodefaultlib main.obj /start-lib puts.obj /end-lib printf.obj | FileCheck %s
+
+;; Test that translation DOES occur when puts is extracted and brought into the link.
+; RUN: lld-link -out:out-extracted.exe -entry:main -subsystem:console -lldmap:- -nodefaultlib main.obj puts.obj printf.obj | FileCheck %s --check-prefix=EXTRACTED
+
+; CHECK-NOT: puts
+; CHECK: printf
+
+; EXTRACTED: printf
+; EXTRACTED: puts
+
+;--- puts.ll
+target datalayout = "e-m:w-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-pc-windows-msvc"
+
+define i32 @puts(ptr nocapture readonly %0) noinline {
+ call void asm sideeffect "", ""()
+ ret i32 0
+}
+
+;--- printf.s
+.globl printf
+printf:
+ ret
+
+;--- main.ll
+target datalayout = "e-m:w-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-pc-windows-msvc"
+
+ at str = constant [5 x i8] c"foo\0A\00"
+
+define i32 @main() {
+ %call = call i32 (ptr, ...) @printf(ptr @str)
+ ret i32 0
+}
+
+declare i32 @printf(ptr, ...)
>From 3b1ffb6514c6c647819d1744eaa89cf4bae3ee8f Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Wed, 11 Mar 2026 15:44:16 -0700
Subject: [PATCH 20/22] Port to WASM
---
lld/test/wasm/lto/libcall-archive-bitcode.ll | 56 ++++++++++++++++++++
lld/wasm/LTO.cpp | 4 ++
lld/wasm/LTO.h | 1 +
lld/wasm/SymbolTable.cpp | 20 +++++++
4 files changed, 81 insertions(+)
create mode 100644 lld/test/wasm/lto/libcall-archive-bitcode.ll
diff --git a/lld/test/wasm/lto/libcall-archive-bitcode.ll b/lld/test/wasm/lto/libcall-archive-bitcode.ll
new file mode 100644
index 0000000000000..3b640efc1091f
--- /dev/null
+++ b/lld/test/wasm/lto/libcall-archive-bitcode.ll
@@ -0,0 +1,56 @@
+; REQUIRES: x86
+
+; RUN: rm -rf %t && split-file %s %t && cd %t
+; RUN: llvm-as main.ll -o main.o
+; RUN: llvm-as puts.ll -o puts.o
+; RUN: llvm-mc -filetype=obj -triple=wasm32-unknown-unknown printf.s -o printf.o
+; RUN: llvm-ar rcs libc.a puts.o printf.o
+
+;; Ensure that no printf->puts translation occurs during LTO because puts is in
+;; bitcode, but was not brought into the link. This would fail the link by
+;; extracting bitcode after LTO.
+; RUN: wasm-ld -o out.wasm main.o libc.a
+; RUN: obj2yaml out.wasm | FileCheck %s
+
+;; Test the same behavior with lazy objects.
+; RUN: wasm-ld -o out-lazy.wasm main.o --start-lib puts.o --end-lib printf.o
+; RUN: obj2yaml out-lazy.wasm | FileCheck %s
+
+;; Test that translation DOES occur when puts is extracted and brought into the link.
+; RUN: wasm-ld -o out-extracted.wasm main.o puts.o printf.o
+; RUN: obj2yaml out-extracted.wasm | FileCheck %s --check-prefix=EXTRACTED
+
+; CHECK-NOT: Name: puts
+; CHECK: Name: printf
+
+; EXTRACTED: Name: puts
+; EXTRACTED-NOT: Name: printf
+
+;--- puts.ll
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target triple = "wasm32-unknown-unknown"
+
+define i32 @puts(ptr nocapture readonly %0) noinline {
+ call void asm sideeffect "", ""()
+ ret i32 0
+}
+
+;--- printf.s
+.globl printf
+printf:
+ .functype printf (i32, i32) -> (i32)
+ i32.const 0
+ end_function
+
+;--- main.ll
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target triple = "wasm32-unknown-unknown"
+
+ at str = constant [5 x i8] c"foo\0A\00"
+
+define i32 @_start() {
+ %call = call i32 (ptr, ...) @printf(ptr @str)
+ ret i32 0
+}
+
+declare i32 @printf(ptr, ...)
\ No newline at end of file
diff --git a/lld/wasm/LTO.cpp b/lld/wasm/LTO.cpp
index 668cdf21ea3ed..2cc1a774a254c 100644
--- a/lld/wasm/LTO.cpp
+++ b/lld/wasm/LTO.cpp
@@ -184,6 +184,10 @@ static void thinLTOCreateEmptyIndexFiles() {
// Merge all the bitcode files we have seen, codegen the result
// and return the resulting objects.
+void BitcodeCompiler::setBitcodeLibFuncs(ArrayRef<StringRef> bitcodeLibFuncs) {
+ ltoObj->setBitcodeLibFuncs(bitcodeLibFuncs);
+}
+
SmallVector<InputFile *, 0> BitcodeCompiler::compile() {
unsigned maxTasks = ltoObj->getMaxTasks();
buf.resize(maxTasks);
diff --git a/lld/wasm/LTO.h b/lld/wasm/LTO.h
index 21b1d59024663..6fa110d6099bf 100644
--- a/lld/wasm/LTO.h
+++ b/lld/wasm/LTO.h
@@ -46,6 +46,7 @@ class BitcodeCompiler {
void add(BitcodeFile &f);
SmallVector<InputFile *, 0> compile();
+ void setBitcodeLibFuncs(ArrayRef<StringRef> bitcodeLibFuncs);
private:
std::unique_ptr<llvm::lto::LTO> ltoObj;
diff --git a/lld/wasm/SymbolTable.cpp b/lld/wasm/SymbolTable.cpp
index 9bd93f317c3c5..9ee91e57a4f4d 100644
--- a/lld/wasm/SymbolTable.cpp
+++ b/lld/wasm/SymbolTable.cpp
@@ -82,8 +82,28 @@ void SymbolTable::compileBitcodeFiles() {
// Prevent further LTO objects being included
BitcodeFile::doneLTO = true;
+ llvm::Triple tt;
+ llvm::BumpPtrAllocator alloc;
+ llvm::StringSaver saver(alloc);
+ SmallVector<StringRef> bitcodeLibFuncs;
+
+ if (!ctx.bitcodeFiles.empty()) {
+ tt = llvm::Triple(ctx.bitcodeFiles.front()->obj->getTargetTriple());
+ for (StringRef libFunc : llvm::lto::LTO::getLibFuncSymbols(tt, saver)) {
+ if (Symbol *sym = find(libFunc)) {
+ if (auto *lazy = dyn_cast<LazySymbol>(sym)) {
+ if (isa<BitcodeFile>(lazy->getFile()))
+ bitcodeLibFuncs.push_back(libFunc);
+ }
+ }
+ }
+ }
+
// Compile bitcode files and replace bitcode symbols.
lto.reset(new BitcodeCompiler);
+ if (!ctx.bitcodeFiles.empty())
+ lto->setBitcodeLibFuncs(bitcodeLibFuncs);
+
for (BitcodeFile *f : ctx.bitcodeFiles)
lto->add(*f);
>From 07ff66e6798eb60a79779c00999ca2ba4adb1631 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 12 Mar 2026 16:10:00 -0700
Subject: [PATCH 21/22] Remove vestigial libcall info from COFF InputFiles
---
lld/COFF/InputFiles.cpp | 5 -----
1 file changed, 5 deletions(-)
diff --git a/lld/COFF/InputFiles.cpp b/lld/COFF/InputFiles.cpp
index 3041e1ce6b445..d415955b6093b 100644
--- a/lld/COFF/InputFiles.cpp
+++ b/lld/COFF/InputFiles.cpp
@@ -25,7 +25,6 @@
#include "llvm/DebugInfo/PDB/Native/NativeSession.h"
#include "llvm/DebugInfo/PDB/Native/PDBFile.h"
#include "llvm/IR/Mangler.h"
-#include "llvm/IR/RuntimeLibcalls.h"
#include "llvm/LTO/LTO.h"
#include "llvm/Object/Binary.h"
#include "llvm/Object/COFF.h"
@@ -1396,10 +1395,6 @@ void BitcodeFile::parse() {
// FIXME: Check nodeduplicate
comdat[i] =
symtab.addComdat(this, saver.save(obj->getComdatTable()[i].first));
- Triple tt(obj->getTargetTriple());
- TargetLibraryInfoImpl tlii(tt);
- TargetLibraryInfo tli(tlii);
- RTLIB::RuntimeLibcallsInfo libcalls(tt);
for (const lto::InputFile::Symbol &objSym : obj->symbols()) {
StringRef symName = saver.save(objSym.getName());
int comdatIndex = objSym.getComdatIndex();
>From c0baeb9cac0fb6752928e83bc39151ea2065d8a3 Mon Sep 17 00:00:00 2001
From: Daniel Thornburgh <dthorn at google.com>
Date: Thu, 12 Mar 2026 16:19:38 -0700
Subject: [PATCH 22/22] Add FIXMEs for DTLTO
---
clang/lib/CodeGen/BackendUtil.cpp | 2 ++
llvm/lib/LTO/LTOBackend.cpp | 1 +
2 files changed, 3 insertions(+)
diff --git a/clang/lib/CodeGen/BackendUtil.cpp b/clang/lib/CodeGen/BackendUtil.cpp
index 7ed7cece9d2d8..a5517a09b0489 100644
--- a/clang/lib/CodeGen/BackendUtil.cpp
+++ b/clang/lib/CodeGen/BackendUtil.cpp
@@ -1409,6 +1409,8 @@ runThinLTOBackend(CompilerInstance &CI, ModuleSummaryIndex *CombinedIndex,
// FIXME: Both ExecuteAction and thinBackend set up optimization remarks for
// the same context.
+ // FIXME: This does not yet set the list of bitcode libfuncs that it isn't
+ // safe to call. This precludes bitcode libc in DTLTO.
finalizeLLVMOptimizationRemarks(M->getContext());
if (Error E = thinBackend(
Conf, -1, AddStream, *M, *CombinedIndex, ImportList,
diff --git a/llvm/lib/LTO/LTOBackend.cpp b/llvm/lib/LTO/LTOBackend.cpp
index 0b0257163dcd2..83f0a1de35040 100644
--- a/llvm/lib/LTO/LTOBackend.cpp
+++ b/llvm/lib/LTO/LTOBackend.cpp
@@ -307,6 +307,7 @@ static void runNewPMPasses(const Config &Conf, Module &Mod, TargetMachine *TM,
// that might have been present in the current LTO unit, but are not, have
// lost their only opportunity to be defined, and calls must not be emitted to
// them.
+ // FIXME: BitcodeLibFuncs isn't yet set for DTLTO.
TargetLibraryInfo TLI(*TLII);
for (unsigned I = 0, E = static_cast<unsigned>(LibFunc::NumLibFuncs); I != E;
++I) {
More information about the cfe-commits
mailing list