[llvm_stats] Do not import llvm.stats metadata for thinlto

[wlei]

Summary

The stats are computed per module and will all be merged in the binary, importing the metadata will cause duplication of the stats.

[hoy]

llvm/include/llvm/IR/Metadata.h

60

I'm not sure this is a good place to define the constant and I don't see any other good place for it either. Maybe just leave "llvm.stats" in place? "llvm.dbg.cu" is an example.

[wlei]

llvm/include/llvm/IR/Metadata.h

60

Sounds good!

[wlei]

Updating D138833: [llvm_stats] Do not import llvm.stats metadata for thinlto

[hoy]

lgtm, thanks.

[wenlei]

thanks for the fix!

llvm/include/llvm/IR/Metadata.h

 enum LLVMConstants : uint32_t {
   DEBUG_METADATA_VERSION = 3 // Current debug info version number.
 };
 
 /// Magic number in the value profile metadata showing a target has been
 /// promoted for the instruction and shouldn't be promoted again.
 const uint64_t NOMORE_ICP_MAGICNUM = -1;
 
+/// Special metadata node name for llvm statistic.
+constexpr const char *LLVMStatsMetadataName = "llvm.stats";
+
 /// Root of the metadata hierarchy.
 ///
 /// This is a root class for typeless data in the IR.
 class Metadata {
   friend class ReplaceableMetadataImpl;
 
   /// RTTI.
   const unsigned char SubclassID;
public:
 
   LLVMContext &getContext() const { return Context; }
 
   /// Replace all uses of this with MD.
   ///
   /// Replace all uses of this with \c MD, which is allowed to be null.
   void replaceAllUsesWith(Metadata *MD);
    /// Replace all uses of the constant with Undef in debug info metadata
-  static void SalvageDebugInfo(const Constant &C); 
+  static void SalvageDebugInfo(const Constant &C);
   /// Returns the list of all DIArgList users of this.
   SmallVector<Metadata *> getAllArgListUsers();
 
   /// Resolve all uses of this.
   ///
   /// Resolve all uses of this, turning off RAUW permanently.  If \c
   /// ResolveUsers, call \a MDNode::resolve() on any users whose last operand
   /// is resolved.

llvm/lib/CodeGen/TargetLoweringObjectFileImpl.cpp

for (const auto *Operand : FuncInfo->operands()) {
       Streamer.switchSection(S);
       Streamer.emitInt64(GUID->getZExtValue());
       Streamer.emitInt64(Hash->getZExtValue());
       Streamer.emitULEB128IntValue(Name->getString().size());
       Streamer.emitBytes(Name->getString());
     }
   }
 
-  if (NamedMDNode *LLVMStats = M.getNamedMetadata("llvm.stats")) {
+  if (NamedMDNode *LLVMStats = M.getNamedMetadata(LLVMStatsMetadataName)) {
     // Emit the metadata for llvm statistics into .llvm_stats section, which is
     // formatted as a list of key/value pair, the value is base64 encoded.
     auto *S = C.getObjectFileInfo()->getLLVMStatsSection();
     Streamer.switchSection(S);
     for (const auto *Operand : LLVMStats->operands()) {
       const auto *MD = cast<MDNode>(Operand);
       assert(MD->getNumOperands() % 2 == 0 &&
              ("Operand num should be even for a list of key/value pair"));

llvm/lib/Linker/IRMover.cpp

if (IsPerformingImport && NMD.getName() == PseudoProbeDescMetadataName) {
       if (!DstM.getNamedMetadata(NMD.getName()))
         emitWarning("Pseudo-probe ignored: source module '" +
                     SrcM->getModuleIdentifier() +
                     "' is compiled with -fpseudo-probe-for-profiling while "
                     "destination module '" +
                     DstM.getModuleIdentifier() + "' is not\n");
       continue;
     }
+    // The stats are computed per module and will all be merged in the binary.
+    // Importing the metadata will cause duplication of the stats.
+    if (IsPerformingImport && NMD.getName() == LLVMStatsMetadataName)
+      continue;
+
     NamedMDNode *DestNMD = DstM.getOrInsertNamedMetadata(NMD.getName());
     // Add Src elements into Dest node.
     for (const MDNode *Op : NMD.operands())
       DestNMD->addOperand(Mapper.mapMDNode(*Op));
   }
 }
 
 /// Merge the linker flags in Src into the Dest module.

llvm/lib/Transforms/IPO/SampleProfile.cpp

if (FunctionSamples::ProfileIsProbeBased) {
                                 MismatchedFuncHashSamples);
       ProfStatsVec.emplace_back("TotalFuncHashSamples", TotalFuncHashSamples);
     }
     ProfStatsVec.emplace_back("MismatchedCallsiteSamples",
                               MismatchedCallsiteSamples);
     ProfStatsVec.emplace_back("TotalCallsiteSamples", TotalCallsiteSamples);
 
     auto *MD = MDB.createLLVMStats(ProfStatsVec);
-    auto *NMD = M.getOrInsertNamedMetadata("llvm.stats");
+    auto *NMD = M.getOrInsertNamedMetadata(LLVMStatsMetadataName);
     NMD->addOperand(MD);
   }
 }
 
 bool SampleProfileLoader::runOnModule(Module &M, ModuleAnalysisManager *AM,
                                       ProfileSummaryInfo *_PSI, CallGraph *CG) {
   GUIDToFuncNameMapper Mapper(M, *Reader, GUIDToFuncNameMap);
 

llvm/test/ThinLTO/X86/Inputs/llvm-stats-import.ll

+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define void @foo() {
+entry:
+  ret void
+}
+
+attributes #0 = { inaccessiblememonly nounwind willreturn }
+
+!llvm.stats = !{!0}
+!0 = !{!"foo", i64 456}

llvm/test/ThinLTO/X86/llvm-stats-import.ll

+; RUN: opt -module-summary %s -o %t1.bc
+; RUN: opt -module-summary %p/Inputs/llvm-stats-import.ll -o %t2.bc
+; RUN: llvm-lto -thinlto-action=thinlink -o %t.index.bc %t1.bc %t2.bc
+
+; RUN: llvm-lto -thinlto-action=import %t1.bc -thinlto-index=%t.index.bc -o - | llvm-dis -o - | FileCheck %s --check-prefix=CHECK-MAIN
+; RUN: llvm-lto -thinlto-action=import %t2.bc -thinlto-index=%t.index.bc -o - | llvm-dis -o - | FileCheck %s --check-prefix=CHECK-FOO
+
+; CHECK-MAIN: !0 = !{!"main", i64 123}
+; CHECK-MAIN-NOT: !{!"foo", i64 456}
+
+; CHECK-FOO: !0 = !{!"foo", i64 456}
+; CHECK-FOO-NOT: !{!"main", i64 123}
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+define i32 @main() {
+entry:
+  call void (...) @foo()
+  ret i32 0
+}
+
+declare void @foo(...)
+
+attributes #0 = { inaccessiblememonly nounwind willreturn }
+
+!llvm.stats = !{!0}
+!0 = !{!"main", i64 123}