Store GUIDs in metadata

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.

Author: orodley

llvm/include/llvm/Analysis/CtxProfAnalysis.h

@@ -37,9 +37,6 @@ 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;
@@ -56,9 +53,7 @@ class PGOContextualProfile {
 
   const PGOCtxProfile &profiles() const { return Profiles; }
 
-  bool isFunctionKnown(const Function &F) const {
-    return getDefinedFunctionGUID(F) != 0;
-  }
+  bool isFunctionKnown(const Function &F) const { return F.getGUID() != 0; }
 
   StringRef getFunctionName(GlobalValue::GUID GUID) const {
     auto It = FuncInfo.find(GUID);
@@ -69,22 +64,22 @@ 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 &)>;
@@ -145,26 +140,5 @@ 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

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)

llvm/include/llvm/IR/GlobalValue.h

@@ -601,9 +601,11 @@ class GlobalValue : public Constant {
         getGlobalIdentifier(GlobalName, GlobalValue::ExternalLinkage, ""));
   }
 
+  bool assignGUID();
+
   /// Return a 64-bit global unique ID constructed from global value name
   /// (i.e. returned by getGlobalIdentifier()).
-  GUID getGUID() const { return getGUID(getGlobalIdentifier()); }
+  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 +1,34 @@
+//===-- 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

llvm/lib/Analysis/CtxProfAnalysis.cpp

@@ -37,35 +37,6 @@ static cl::opt<CtxProfAnalysisPrinterPass::PrintMode> PrintLevel(
                           "just the yaml representation of the profile")),
     cl::desc("Verbosity level of the contextual profile printer pass."));
 
-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)
@@ -98,7 +69,7 @@ PGOContextualProfile CtxProfAnalysis::run(Module &M,
   DenseSet<GlobalValue::GUID> ProfileRootsInModule;
   for (const auto &F : M)
     if (!F.isDeclaration())
-      if (auto GUID = AssignGUIDPass::getGUID(F);
+      if (auto GUID = F.getGUID();
           MaybeProfiles->Contexts.find(GUID) != MaybeProfiles->Contexts.end())
         ProfileRootsInModule.insert(GUID);
 
@@ -118,7 +89,7 @@ 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.
@@ -152,13 +123,6 @@ 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) {}
 
@@ -262,7 +226,7 @@ void PGOContextualProfile::initIndex() {
 
 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));
@@ -273,7 +237,7 @@ 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

@@ -77,6 +77,39 @@ GlobalValue::GUID GlobalValue::getGUID(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::getGUID(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)                                              \

llvm/lib/Passes/PassBuilder.cpp

@@ -324,6 +324,7 @@
 #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"
@@ -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)
       return MPM;

llvm/lib/Transforms/Instrumentation/PGOCtxProfFlattening.cpp

@@ -451,7 +451,7 @@ 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

@@ -235,8 +235,7 @@ 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.