Store GUIDs in metadata

llvm/include/llvm/Analysis/CtxProfAnalysis.h

@@ -47,9 +47,6 @@ 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.
-  LLVM_ABI GlobalValue::GUID getDefinedFunctionGUID(const Function &F) const;
-
   // This is meant to be constructed from CtxProfAnalysis, which will also set
   // its state piecemeal.
   PGOContextualProfile() = default;
@@ -68,9 +65,7 @@ class PGOContextualProfile {
 
   LLVM_ABI bool isInSpecializedModule() const;
 
-  bool isFunctionKnown(const Function &F) const {
-    return getDefinedFunctionGUID(F) != 0;
-  }
+  bool isFunctionKnown(const Function &F) const { return !F.isDeclaration(); }
 
   StringRef getFunctionName(GlobalValue::GUID GUID) const {
     auto It = FuncInfo.find(GUID);
@@ -81,22 +76,22 @@ 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 &)>;
@@ -188,26 +183,5 @@ class ProfileAnnotator {
   LLVM_ABI ~ProfileAnnotator();
 };
 
-/// 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`.
-  LLVM_ABI PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
-  LLVM_ABI static const char *GUIDMetadataName;
-  // This should become GlobalValue::getGUID
-  LLVM_ABI static uint64_t getGUID(const Function &F);
-};
-
 } // namespace llvm
 #endif // LLVM_ANALYSIS_CTXPROFANALYSIS_H

llvm/include/llvm/IR/FixedMetadataKinds.def

@@ -54,3 +54,4 @@ 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_callee_type, "callee_type", 42)
+LLVM_FIXED_MD_KIND(MD_unique_id, "guid", 43)

llvm/include/llvm/IR/GlobalValue.h

@@ -589,17 +589,33 @@ class GlobalValue : public Constant {
   /// used as the key for a global lookup (e.g. profile or ThinLTO).
   LLVM_ABI std::string getGlobalIdentifier() const;
 
+  /// Assign a GUID to this value based on its current name and linkage.
+  /// This GUID will remain the same even if those change. This method is
+  /// idempotent -- if a GUID has already been assigned, calling it again
+  /// will do nothing.
+  ///
+  /// This is private (exposed only to \c AssignGUIDPass), as users don't need
+  /// to call it. GUIDs are assigned only by \c AssignGUIDPass. The pass
+  /// pipeline should be set up such that GUIDs are always available when
+  /// needed. If not, the GUID assignment pass should be moved (or run again)
+  /// such that they are.
+  void assignGUID();
+
+  // assignGUID needs to be accessible from AssignGUIDPass, which is called
+  // early in the pipeline to make GUIDs available to later passes. But we'd
+  // rather not expose it publicly, as no-one else should call it.
+  friend class AssignGUIDPass;
+
 public:
   /// Return a 64-bit global unique ID constructed from the name of a global
   /// symbol. Since this call doesn't supply the linkage or defining filename,
   /// the GUID computation will assume that the global has external linkage.
   LLVM_ABI static GUID getGUIDAssumingExternalLinkage(StringRef GlobalName);
 
-  /// Return a 64-bit global unique ID constructed from global value name
-  /// (i.e. returned by getGlobalIdentifier()).
-  GUID getGUID() const {
-    return getGUIDAssumingExternalLinkage(getGlobalIdentifier());
-  }
+  /// Return a 64-bit global unique ID for this value. It is based on the
+  /// "original" name and linkage of this value (i.e. whenever its GUID was
+  /// assigned). This might not match the current name and linkage.
+  GUID getGUID() const;
 
   /// @name Materialization
   /// Materialization is used to construct functions only as they're needed.

llvm/include/llvm/Transforms/Utils/AssignGUID.h

@@ -0,0 +1,40 @@
+//===-- 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"
+#include "llvm/Support/Debug.h"
+
+namespace llvm {
+
+class AssignGUIDPass : public PassInfoMixin<AssignGUIDPass> {
+public:
+  AssignGUIDPass() = default;
+
+  static void runOnModule(Module &M);
+
+  PreservedAnalyses run(Module &M, ModuleAnalysisManager &AM) {
+    AssignGUIDPass::runOnModule(M);
+    return PreservedAnalyses::none();
+  }
+
+  static bool isRequired() { return true; }
+};
+
+} // end namespace llvm
+
+#endif // LLVM_TRANSFORMS_UTILS_ASSIGNGUID_H

llvm/lib/Analysis/CtxProfAnalysis.cpp

@@ -48,8 +48,6 @@ 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";
-
 namespace llvm {
 class ProfileAnnotatorImpl final {
   friend class ProfileAnnotator;
@@ -418,33 +416,6 @@ bool ProfileAnnotator::getOutgoingBranchWeights(
   return MaxCount > 0;
 }
 
-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)
@@ -513,7 +484,7 @@ 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.
@@ -547,13 +518,6 @@ 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) {}
 
@@ -669,7 +633,7 @@ 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));
@@ -680,7 +644,7 @@ 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));

llvm/lib/IR/Globals.cpp

@@ -78,6 +78,34 @@ GlobalValue::getGUIDAssumingExternalLinkage(StringRef GlobalIdentifier) {
   return MD5Hash(GlobalIdentifier);
 }
 
+void GlobalValue::assignGUID() {
+  if (getMetadata(LLVMContext::MD_unique_id) != nullptr)
+    return;
+
+  const GUID G =
+      GlobalValue::getGUIDAssumingExternalLinkage(getGlobalIdentifier());
+  setMetadata(
+      LLVMContext::MD_unique_id,
+      MDNode::get(getContext(), {ConstantAsMetadata::get(ConstantInt::get(
+                                    Type::getInt64Ty(getContext()), G))}));
+}
+
+GlobalValue::GUID GlobalValue::getGUID() const {
+  if (isDeclaration()) {
+    return GlobalValue::getGUIDAssumingExternalLinkage(getGlobalIdentifier());
+  }
+  if (isa<GlobalAlias>(this)) {
+    return GlobalValue::getGUIDAssumingExternalLinkage(getGlobalIdentifier());
+  }
+
+  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)                                              \

llvm/lib/Passes/PassBuilder.cpp

@@ -338,6 +338,7 @@
 #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"

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"
@@ -786,6 +787,7 @@ PassBuilder::buildFunctionSimplificationPipeline(OptimizationLevel Level,
 void PassBuilder::addRequiredLTOPreLinkPasses(ModulePassManager &MPM) {
   MPM.addPass(CanonicalizeAliasesPass());
   MPM.addPass(NameAnonGlobalPass());
+  MPM.addPass(AssignGUIDPass());
 }
 
 void PassBuilder::addPreInlinerPasses(ModulePassManager &MPM,
@@ -1234,9 +1236,6 @@ 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));
       return MPM;
@@ -1640,6 +1639,11 @@ PassBuilder::buildPerModuleDefaultPipeline(OptimizationLevel Level,
 
   ModulePassManager MPM;
 
+  /*
+  MPM.addPass(NameAnonGlobalPass());
+  MPM.addPass(AssignGUIDPass());
+  */
+
   // Convert @llvm.global.annotations to !annotation metadata.
   MPM.addPass(Annotation2MetadataPass());
 
@@ -2208,6 +2212,11 @@ PassBuilder::buildO0DefaultPipeline(OptimizationLevel Level,
 
   ModulePassManager MPM;
 
+  /*
+  MPM.addPass(NameAnonGlobalPass());
+  MPM.addPass(AssignGUIDPass());
+  */
+
   // Perform pseudo probe instrumentation in O0 mode. This is for the
   // consistency between different build modes. For example, a LTO build can be
   // mixed with an O0 prelink and an O2 postlink. Loading a sample profile in

llvm/lib/Transforms/IPO/WholeProgramDevirt.cpp

@@ -91,6 +91,7 @@
 #include "llvm/TargetParser/Triple.h"
 #include "llvm/Transforms/IPO.h"
 #include "llvm/Transforms/IPO/FunctionAttrs.h"
+#include "llvm/Transforms/Utils/AssignGUID.h"
 #include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/CallPromotionUtils.h"
 #include "llvm/Transforms/Utils/Evaluator.h"
@@ -860,6 +861,12 @@ void llvm::updateVCallVisibilityInModule(
     function_ref<bool(StringRef)> IsVisibleToRegularObj) {
   if (!hasWholeProgramVisibility(WholeProgramVisibilityEnabledInLTO))
     return;
+
+  // Manually assign GUIDs -- updateVCallVisibilityInModule accesses GUIDs, and
+  // there's no way to specify it in the pass pipeline since this runs before
+  // any pass given on the command line.
+  AssignGUIDPass::runOnModule(M);
+
   for (GlobalVariable &GV : M.globals()) {
     // Add linkage unit visibility to any variable with type metadata, which are
     // the vtable definitions. We won't have an existing vcall_visibility

llvm/lib/Transforms/Instrumentation/PGOCtxProfFlattening.cpp

@@ -126,7 +126,7 @@ void annotateIndirectCalls(Module &M, const CtxProfAnalysis::Result &CtxProf) {
   for (auto &F : M) {
     if (F.isDeclaration())
       continue;
-    auto FlatProfIter = FlatIndCalls.find(AssignGUIDPass::getGUID(F));
+    auto FlatProfIter = FlatIndCalls.find(F.getGUID());
     if (FlatProfIter == FlatIndCalls.end())
       continue;
     const auto &FlatProf = FlatProfIter->second;
@@ -179,7 +179,7 @@ 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);

llvm/lib/Transforms/Instrumentation/PGOCtxProfLowering.cpp

@@ -280,8 +280,7 @@ 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.