Store GUIDs in metadata

orodley · orodley · commit 9d34fae78840 · 2025-04-14T11:03:11.000+10:00
This takes the existing AssignGUID pass from CtxProfAnalysis, and runs
it by default, at the appropriate stages of the LTO pipeline. It also
changes GlobalValue::getGUID() to retrieve the GUID from the metadata
instead of computing it.

We don't yet have the supporting downstream changes to make a dedicated
GUID table in bitcode, nor do we use the metadata as part of ThinLTO --
it retains its existing mechanisms of recomputing GUIDs from separately
saved data. That will be changed later.
diff --git a/llvm/include/llvm/Analysis/CtxProfAnalysis.h b/llvm/include/llvm/Analysis/CtxProfAnalysis.h
@@ -46,9 +46,6 @@ class PGOContextualProfile {
   // we'll need when we maintain the profiles during IPO transformations.
   std::map<GlobalValue::GUID, FunctionInfo> FuncInfo;
 
-  /// Get the GUID of this Function if it's defined in this module.
-  GlobalValue::GUID getDefinedFunctionGUID(const Function &F) const;
-
   // This is meant to be constructed from CtxProfAnalysis, which will also set
   // its state piecemeal.
   PGOContextualProfile() = default;
@@ -67,9 +64,9 @@ class PGOContextualProfile {
 
   bool isInSpecializedModule() const;
 
-  bool isFunctionKnown(const Function &F) const {
-    return getDefinedFunctionGUID(F) != 0;
-  }
+  bool isInSpecializedModule() const;
+
+  bool isFunctionKnown(const Function &F) const { return F.getGUID() != 0; }
 
   StringRef getFunctionName(GlobalValue::GUID GUID) const {
     auto It = FuncInfo.find(GUID);
@@ -80,22 +77,22 @@ class PGOContextualProfile {
 
   uint32_t getNumCounters(const Function &F) const {
     assert(isFunctionKnown(F));
-    return FuncInfo.find(getDefinedFunctionGUID(F))->second.NextCounterIndex;
+    return FuncInfo.find(F.getGUID())->second.NextCounterIndex;
   }
 
   uint32_t getNumCallsites(const Function &F) const {
     assert(isFunctionKnown(F));
-    return FuncInfo.find(getDefinedFunctionGUID(F))->second.NextCallsiteIndex;
+    return FuncInfo.find(F.getGUID())->second.NextCallsiteIndex;
   }
 
   uint32_t allocateNextCounterIndex(const Function &F) {
     assert(isFunctionKnown(F));
-    return FuncInfo.find(getDefinedFunctionGUID(F))->second.NextCounterIndex++;
+    return FuncInfo.find(F.getGUID())->second.NextCounterIndex++;
   }
 
   uint32_t allocateNextCallsiteIndex(const Function &F) {
     assert(isFunctionKnown(F));
-    return FuncInfo.find(getDefinedFunctionGUID(F))->second.NextCallsiteIndex++;
+    return FuncInfo.find(F.getGUID())->second.NextCallsiteIndex++;
   }
 
   using ConstVisitor = function_ref<void(const PGOCtxProfContext &)>;
@@ -157,26 +154,5 @@ class CtxProfAnalysisPrinterPass
   const PrintMode Mode;
 };
 
-/// Assign a GUID to functions as metadata. GUID calculation takes linkage into
-/// account, which may change especially through and after thinlto. By
-/// pre-computing and assigning as metadata, this mechanism is resilient to such
-/// changes (as well as name changes e.g. suffix ".llvm." additions).
-
-// FIXME(mtrofin): we can generalize this mechanism to calculate a GUID early in
-// the pass pipeline, associate it with any Global Value, and then use it for
-// PGO and ThinLTO.
-// At that point, this should be moved elsewhere.
-class AssignGUIDPass : public PassInfoMixin<AssignGUIDPass> {
-public:
-  explicit AssignGUIDPass() = default;
-
-  /// Assign a GUID *if* one is not already assign, as a function metadata named
-  /// `GUIDMetadataName`.
-  PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
-  static const char *GUIDMetadataName;
-  // This should become GlobalValue::getGUID
-  static uint64_t getGUID(const Function &F);
-};
-
 } // namespace llvm
 #endif // LLVM_ANALYSIS_CTXPROFANALYSIS_H
diff --git a/llvm/include/llvm/IR/FixedMetadataKinds.def b/llvm/include/llvm/IR/FixedMetadataKinds.def
@@ -53,3 +53,4 @@ LLVM_FIXED_MD_KIND(MD_DIAssignID, "DIAssignID", 38)
 LLVM_FIXED_MD_KIND(MD_coro_outside_frame, "coro.outside.frame", 39)
 LLVM_FIXED_MD_KIND(MD_mmra, "mmra", 40)
 LLVM_FIXED_MD_KIND(MD_noalias_addrspace, "noalias.addrspace", 41)
+LLVM_FIXED_MD_KIND(MD_unique_id, "unique_id", 42)
diff --git a/llvm/include/llvm/IR/GlobalValue.h b/llvm/include/llvm/IR/GlobalValue.h
@@ -594,11 +594,11 @@ class GlobalValue : public Constant {
   /// the GUID computation will assume that the global has external linkage.
   static GUID getGUIDAssumingExternalLinkage(StringRef GlobalName);
 
+  bool assignGUID();
+
   /// Return a 64-bit global unique ID constructed from global value name
   /// (i.e. returned by getGlobalIdentifier()).
-  GUID getGUID() const {
-    return getGUIDAssumingExternalLinkage(getGlobalIdentifier());
-  }
+  GUID getGUID() const;
 
   /// @name Materialization
   /// Materialization is used to construct functions only as they're needed.
diff --git a/llvm/include/llvm/Transforms/Utils/AssignGUID.h b/llvm/include/llvm/Transforms/Utils/AssignGUID.h
@@ -0,0 +1,34 @@
+//===-- AssignGUID.h - Unique identifier assignment pass --------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file provides a pass which assigns a a GUID (globally unique identifier)
+// to every GlobalValue in the module, according to its current name, linkage,
+// and originating file.
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_TRANSFORMS_UTILS_ASSIGNGUID_H
+#define LLVM_TRANSFORMS_UTILS_ASSIGNGUID_H
+
+#include "llvm/IR/Module.h"
+#include "llvm/IR/PassManager.h"
+
+namespace llvm {
+
+class AssignGUIDPass : public PassInfoMixin<AssignGUIDPass> {
+public:
+  AssignGUIDPass() = default;
+
+  PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM);
+
+  static bool isRequired() { return true; }
+};
+
+} // end namespace llvm
+
+#endif // LLVM_TRANSFORMS_UTILS_ASSIGNGUID_H
diff --git a/llvm/lib/Analysis/CtxProfAnalysis.cpp b/llvm/lib/Analysis/CtxProfAnalysis.cpp
@@ -44,35 +44,6 @@ static cl::opt<bool> ForceIsInSpecializedModule(
     cl::desc("Treat the given module as-if it were containing the "
              "post-thinlink module containing the root"));
 
-const char *AssignGUIDPass::GUIDMetadataName = "guid";
-
-PreservedAnalyses AssignGUIDPass::run(Module &M, ModuleAnalysisManager &MAM) {
-  for (auto &F : M.functions()) {
-    if (F.isDeclaration())
-      continue;
-    if (F.getMetadata(GUIDMetadataName))
-      continue;
-    const GlobalValue::GUID GUID = F.getGUID();
-    F.setMetadata(GUIDMetadataName,
-                  MDNode::get(M.getContext(),
-                              {ConstantAsMetadata::get(ConstantInt::get(
-                                  Type::getInt64Ty(M.getContext()), GUID))}));
-  }
-  return PreservedAnalyses::none();
-}
-
-GlobalValue::GUID AssignGUIDPass::getGUID(const Function &F) {
-  if (F.isDeclaration()) {
-    assert(GlobalValue::isExternalLinkage(F.getLinkage()));
-    return F.getGUID();
-  }
-  auto *MD = F.getMetadata(GUIDMetadataName);
-  assert(MD && "guid not found for defined function");
-  return cast<ConstantInt>(cast<ConstantAsMetadata>(MD->getOperand(0))
-                               ->getValue()
-                               ->stripPointerCasts())
-      ->getZExtValue();
-}
 AnalysisKey CtxProfAnalysis::Key;
 
 CtxProfAnalysis::CtxProfAnalysis(std::optional<StringRef> Profile)
@@ -141,7 +112,7 @@ PGOContextualProfile CtxProfAnalysis::run(Module &M,
   for (const auto &F : M) {
     if (F.isDeclaration())
       continue;
-    auto GUID = AssignGUIDPass::getGUID(F);
+    auto GUID = F.getGUID();
     assert(GUID && "guid not found for defined function");
     const auto &Entry = F.begin();
     uint32_t MaxCounters = 0; // we expect at least a counter.
@@ -175,13 +146,6 @@ PGOContextualProfile CtxProfAnalysis::run(Module &M,
   return Result;
 }
 
-GlobalValue::GUID
-PGOContextualProfile::getDefinedFunctionGUID(const Function &F) const {
-  if (auto It = FuncInfo.find(AssignGUIDPass::getGUID(F)); It != FuncInfo.end())
-    return It->first;
-  return 0;
-}
-
 CtxProfAnalysisPrinterPass::CtxProfAnalysisPrinterPass(raw_ostream &OS)
     : OS(OS), Mode(PrintLevel) {}
 
@@ -291,7 +255,7 @@ bool PGOContextualProfile::isInSpecializedModule() const {
 
 void PGOContextualProfile::update(Visitor V, const Function &F) {
   assert(isFunctionKnown(F));
-  GlobalValue::GUID G = getDefinedFunctionGUID(F);
+  GlobalValue::GUID G = F.getGUID();
   for (auto *Node = FuncInfo.find(G)->second.Index.Next; Node;
        Node = Node->Next)
     V(*reinterpret_cast<PGOCtxProfContext *>(Node));
@@ -302,7 +266,7 @@ void PGOContextualProfile::visit(ConstVisitor V, const Function *F) const {
     return preorderVisit<const PGOCtxProfContext::CallTargetMapTy,
                          const PGOCtxProfContext>(Profiles.Contexts, V);
   assert(isFunctionKnown(*F));
-  GlobalValue::GUID G = getDefinedFunctionGUID(*F);
+  GlobalValue::GUID G = F->getGUID();
   for (const auto *Node = FuncInfo.find(G)->second.Index.Next; Node;
        Node = Node->Next)
     V(*reinterpret_cast<const PGOCtxProfContext *>(Node));
diff --git a/llvm/lib/IR/Globals.cpp b/llvm/lib/IR/Globals.cpp
@@ -78,6 +78,40 @@ GlobalValue::getGUIDAssumingExternalLinkage(StringRef GlobalIdentifier) {
   return MD5Hash(GlobalIdentifier);
 }
 
+bool GlobalValue::assignGUID() {
+  if (isDeclaration())
+    return false;
+  if (getMetadata(LLVMContext::MD_unique_id) != nullptr)
+    return false;
+
+  const GUID G =
+      GlobalValue::getGUIDAssumingExternalLinkage(getGlobalIdentifier());
+  setMetadata(
+      LLVMContext::MD_unique_id,
+      MDNode::get(getContext(), {ConstantAsMetadata::get(ConstantInt::get(
+                                    Type::getInt64Ty(getContext()), G))}));
+  return true;
+}
+
+GlobalValue::GUID GlobalValue::getGUID() const {
+  // TODO: do we need a special case for GlobalAlias?
+
+  if (isDeclaration()) {
+    // Declarations don't get an assigned GUID. Their definition is in another
+    // module, and it may have been promoted from local to external linkage
+    // during LTO. Because of this, we can only determine their GUID once the
+    // definition is available.
+    return 0;
+  }
+
+  auto *MD = getMetadata(LLVMContext::MD_unique_id);
+  assert(MD != nullptr && "GUID was not assigned before calling GetGUID()");
+  return cast<ConstantInt>(cast<ConstantAsMetadata>(MD->getOperand(0))
+                               ->getValue()
+                               ->stripPointerCasts())
+      ->getZExtValue();
+}
+
 void GlobalValue::removeFromParent() {
   switch (getValueID()) {
 #define HANDLE_GLOBAL_VALUE(NAME)                                              \
diff --git a/llvm/lib/Passes/PassBuilder.cpp b/llvm/lib/Passes/PassBuilder.cpp
@@ -329,6 +329,7 @@
 #include "llvm/Transforms/Scalar/TailRecursionElimination.h"
 #include "llvm/Transforms/Scalar/WarnMissedTransforms.h"
 #include "llvm/Transforms/Utils/AddDiscriminators.h"
+#include "llvm/Transforms/Utils/AssignGUID.h"
 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
 #include "llvm/Transforms/Utils/BreakCriticalEdges.h"
 #include "llvm/Transforms/Utils/CanonicalizeAliases.h"
diff --git a/llvm/lib/Passes/PassBuilderPipelines.cpp b/llvm/lib/Passes/PassBuilderPipelines.cpp
@@ -130,6 +130,7 @@
 #include "llvm/Transforms/Scalar/TailRecursionElimination.h"
 #include "llvm/Transforms/Scalar/WarnMissedTransforms.h"
 #include "llvm/Transforms/Utils/AddDiscriminators.h"
+#include "llvm/Transforms/Utils/AssignGUID.h"
 #include "llvm/Transforms/Utils/AssumeBundleBuilder.h"
 #include "llvm/Transforms/Utils/CanonicalizeAliases.h"
 #include "llvm/Transforms/Utils/CountVisits.h"
@@ -791,6 +792,7 @@ PassBuilder::buildFunctionSimplificationPipeline(OptimizationLevel Level,
 void PassBuilder::addRequiredLTOPreLinkPasses(ModulePassManager &MPM) {
   MPM.addPass(CanonicalizeAliasesPass());
   MPM.addPass(NameAnonGlobalPass());
+  MPM.addPass(AssignGUIDPass());
 }
 
 void PassBuilder::addPreInlinerPasses(ModulePassManager &MPM,
@@ -1076,6 +1078,11 @@ PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
 
   ModulePassManager MPM;
 
+  if (Phase != ThinOrFullLTOPhase::ThinLTOPostLink) {
+    MPM.addPass(NameAnonGlobalPass());
+    MPM.addPass(AssignGUIDPass());
+  }
+
   // Place pseudo probe instrumentation as the first pass of the pipeline to
   // minimize the impact of optimization changes.
   if (PGOOpt && PGOOpt->PseudoProbeForProfiling &&
@@ -1239,8 +1246,6 @@ PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
     // In pre-link, we just want the instrumented IR. We use the contextual
     // profile in the post-thinlink phase.
     // The instrumentation will be removed in post-thinlink after IPO.
-    // FIXME(mtrofin): move AssignGUIDPass if there is agreement to use this
-    // mechanism for GUIDs.
     MPM.addPass(AssignGUIDPass());
     if (IsCtxProfUse) {
       MPM.addPass(PGOCtxProfFlatteningPass(/*IsPreThinlink=*/true));
diff --git a/llvm/lib/Transforms/Instrumentation/PGOCtxProfFlattening.cpp b/llvm/lib/Transforms/Instrumentation/PGOCtxProfFlattening.cpp
@@ -505,7 +505,7 @@ PreservedAnalyses PGOCtxProfFlatteningPass::run(Module &M,
            "Function has unreacheable basic blocks. The expectation was that "
            "DCE was run before.");
 
-    auto It = FlattenedProfile.find(AssignGUIDPass::getGUID(F));
+    auto It = FlattenedProfile.find(F.getGUID());
     // If this function didn't appear in the contextual profile, it's cold.
     if (It == FlattenedProfile.end())
       clearColdFunctionProfile(F);
diff --git a/llvm/lib/Transforms/Instrumentation/PGOCtxProfLowering.cpp b/llvm/lib/Transforms/Instrumentation/PGOCtxProfLowering.cpp
@@ -247,8 +247,7 @@ bool CtxInstrumentationLowerer::lowerFunction(Function &F) {
       assert(Mark->getIndex()->isZero());
 
       IRBuilder<> Builder(Mark);
-      Guid = Builder.getInt64(
-          AssignGUIDPass::getGUID(cast<Function>(*Mark->getNameValue())));
+      Guid = Builder.getInt64(cast<Function>(*Mark->getNameValue()).getGUID());
       // The type of the context of this function is now knowable since we have
       // NumCallsites and NumCounters. We delcare it here because it's more
       // convenient - we have the Builder.
diff --git a/llvm/lib/Transforms/Utils/AssignGUID.cpp b/llvm/lib/Transforms/Utils/AssignGUID.cpp
@@ -0,0 +1,34 @@
+//===-- AssignGUID.cpp - Unique identifier assignment pass ------*- C++ -*-===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// This file provides a pass which assigns a a GUID (globally unique identifier)
+// to every GlobalValue in the module, according to its current name, linkage,
+// and originating file.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/Utils/AssignGUID.h"
+#include "llvm/Support/Debug.h"
+
+using namespace llvm;
+
+PreservedAnalyses AssignGUIDPass::run(Module &M, ModuleAnalysisManager &AM) {
+  for (auto &GV : M.globals()) {
+    if (GV.isDeclaration())
+      continue;
+    GV.assignGUID();
+    dbgs() << "[Added GUID to GV:] " << GV.getName() << "\n";
+  }
+  for (auto &F : M.functions()) {
+    if (F.isDeclaration())
+      continue;
+    F.assignGUID();
+    dbgs() << "[Added GUID to F:] " << F.getName() << "\n";
+  }
+  return PreservedAnalyses::none();
+}
diff --git a/llvm/lib/Transforms/Utils/CMakeLists.txt b/llvm/lib/Transforms/Utils/CMakeLists.txt
@@ -2,6 +2,7 @@ add_llvm_component_library(LLVMTransformUtils
   AddDiscriminators.cpp
   AMDGPUEmitPrintf.cpp
   ASanStackFrameLayout.cpp
+  AssignGUID.cpp
   AssumeBundleBuilder.cpp
   BasicBlockUtils.cpp
   BreakCriticalEdges.cpp
diff --git a/llvm/lib/Transforms/Utils/CallPromotionUtils.cpp b/llvm/lib/Transforms/Utils/CallPromotionUtils.cpp
@@ -616,11 +616,11 @@ CallBase *llvm::promoteCallWithIfThenElse(CallBase &CB, Function &Callee,
   IndirectBBIns->setIndex(IndirectID);
   IndirectBBIns->insertInto(&IndirectBB, IndirectBB.getFirstInsertionPt());
 
-  const GlobalValue::GUID CalleeGUID = AssignGUIDPass::getGUID(Callee);
+  const GlobalValue::GUID CalleeGUID = Callee.getGUID();
   const uint32_t NewCountersSize = IndirectID + 1;
 
   auto ProfileUpdater = [&](PGOCtxProfContext &Ctx) {
-    assert(Ctx.guid() == AssignGUIDPass::getGUID(Caller));
+    assert(Ctx.guid() == Caller.getGUID());
     assert(NewCountersSize - 2 == Ctx.counters().size());
     // All the ctx-es belonging to a function must have the same size counters.
     Ctx.resizeCounters(NewCountersSize);
@@ -655,7 +655,6 @@ CallBase *llvm::promoteCallWithIfThenElse(CallBase &CB, Function &Callee,
     // times, and the indirect BB, IndirectCount times
     Ctx.counters()[DirectID] = DirectCount;
     Ctx.counters()[IndirectID] = IndirectCount;
-
   };
   CtxProf.update(ProfileUpdater, Caller);
   return &DirectCall;
diff --git a/llvm/lib/Transforms/Utils/InlineFunction.cpp b/llvm/lib/Transforms/Utils/InlineFunction.cpp
@@ -2373,7 +2373,7 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
   // Get some preliminary data about the callsite before it might get inlined.
   // Inlining shouldn't delete the callee, but it's cleaner (and low-cost) to
   // get this data upfront and rely less on InlineFunction's behavior.
-  const auto CalleeGUID = AssignGUIDPass::getGUID(Callee);
+  const auto CalleeGUID = Callee.getGUID();
   auto *CallsiteIDIns = CtxProfAnalysis::getCallsiteInstrumentation(CB);
   const auto CallsiteID =
       static_cast<uint32_t>(CallsiteIDIns->getIndex()->getZExtValue());
@@ -2398,7 +2398,7 @@ llvm::InlineResult llvm::InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
   const uint32_t NewCountersSize = CtxProf.getNumCounters(Caller);
 
   auto Updater = [&](PGOCtxProfContext &Ctx) {
-    assert(Ctx.guid() == AssignGUIDPass::getGUID(Caller));
+    assert(Ctx.guid() == Caller.getGUID());
     const auto &[CalleeCounterMap, CalleeCallsiteMap] = IndicesMaps;
     assert(
         (Ctx.counters().size() +
