[ctx_prof] Add Inlining support (llvm#106154)

Add an overload of `InlineFunction` that updates the contextual profile. If there is no contextual profile, this overload is equivalent to the non-contextual profile variant.

Post-inlining, the update mainly consists of:
- making the PGO instrumentation of the callee "the caller's": the owner function (the "name" parameter of the instrumentation instructions) becomes the caller, and new index values are allocated for each of the callee's indices (this happens for both increment and callsite instrumentation instructions)
- in the contextual profile:
   - each context corresponding to the caller has its counters updated to incorporate the counters inherited from the callee at the inlined callsite. Counter values are copied as-is because no scaling is required since the profile is contextual.
   - the contexts of the callee (at the inlined callsite) are moved to the caller.
   - the callee context at the inlined callsite is deleted.

Change-Id: If4814bc9aeea2a55e597b93cb0ecb33aa73be01f

Author: mtrofin

llvm/include/llvm/Analysis/CtxProfAnalysis.h

@@ -15,6 +15,7 @@
 #include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/PassManager.h"
 #include "llvm/ProfileData/PGOCtxProfReader.h"
+#include <optional>
 
 namespace llvm {
 
@@ -63,6 +64,16 @@ class PGOContextualProfile {
     return getDefinedFunctionGUID(F) != 0;
   }
 
+  uint32_t getNumCounters(const Function &F) const {
+    assert(isFunctionKnown(F));
+    return FuncInfo.find(getDefinedFunctionGUID(F))->second.NextCounterIndex;
+  }
+
+  uint32_t getNumCallsites(const Function &F) const {
+    assert(isFunctionKnown(F));
+    return FuncInfo.find(getDefinedFunctionGUID(F))->second.NextCallsiteIndex;
+  }
+
   uint32_t allocateNextCounterIndex(const Function &F) {
     assert(isFunctionKnown(F));
     return FuncInfo.find(getDefinedFunctionGUID(F))->second.NextCounterIndex++;
@@ -91,11 +102,11 @@ class PGOContextualProfile {
 };
 
 class CtxProfAnalysis : public AnalysisInfoMixin<CtxProfAnalysis> {
-  StringRef Profile;
+  const std::optional<StringRef> Profile;
 
 public:
   static AnalysisKey Key;
-  explicit CtxProfAnalysis(StringRef Profile = "");
+  explicit CtxProfAnalysis(std::optional<StringRef> Profile = std::nullopt);
 
   using Result = PGOContextualProfile;
 
@@ -113,9 +124,7 @@ class CtxProfAnalysisPrinterPass
     : public PassInfoMixin<CtxProfAnalysisPrinterPass> {
 public:
   enum class PrintMode { Everything, JSON };
-  explicit CtxProfAnalysisPrinterPass(raw_ostream &OS,
-                                      PrintMode Mode = PrintMode::Everything)
-      : OS(OS), Mode(Mode) {}
+  explicit CtxProfAnalysisPrinterPass(raw_ostream &OS);
 
   PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
   static bool isRequired() { return true; }

llvm/include/llvm/IR/IntrinsicInst.h

@@ -1578,6 +1578,8 @@ class InstrProfInstBase : public IntrinsicInst {
     return const_cast<Value *>(getArgOperand(0))->stripPointerCasts();
   }
 
+  void setNameValue(Value *V) { setArgOperand(0, V); }
+
   // The hash of the CFG for the instrumented function.
   ConstantInt *getHash() const {
     return cast<ConstantInt>(const_cast<Value *>(getArgOperand(1)));

llvm/include/llvm/ProfileData/PGOCtxProfReader.h

@@ -74,6 +74,13 @@ class PGOCtxProfContext final {
     Iter->second.emplace(Other.guid(), std::move(Other));
   }
 
+  void ingestAllContexts(uint32_t CSId, CallTargetMapTy &&Other) {
+    auto [_, Inserted] = callsites().try_emplace(CSId, std::move(Other));
+    (void)Inserted;
+    assert(Inserted &&
+           "CSId was expected to be newly created as result of e.g. inlining");
+  }
+
   void resizeCounters(uint32_t Size) { Counters.resize(Size); }
 
   bool hasCallsite(uint32_t I) const {

llvm/include/llvm/Transforms/Utils/Cloning.h

@@ -20,6 +20,7 @@
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/Twine.h"
 #include "llvm/Analysis/AssumptionCache.h"
+#include "llvm/Analysis/CtxProfAnalysis.h"
 #include "llvm/Analysis/InlineCost.h"
 #include "llvm/IR/BasicBlock.h"
 #include "llvm/IR/ValueHandle.h"
@@ -271,6 +272,17 @@ InlineResult InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
                             bool InsertLifetime = true,
                             Function *ForwardVarArgsTo = nullptr);
 
+/// Same as above, but it will update the contextual profile. If the contextual
+/// profile is invalid (i.e. not loaded because it is not present), it defaults
+/// to the behavior of the non-contextual profile updating variant above. This
+/// makes it easy to drop-in replace uses of the non-contextual overload.
+InlineResult InlineFunction(CallBase &CB, InlineFunctionInfo &IFI,
+                            CtxProfAnalysis::Result &CtxProf,
+                            bool MergeAttributes = false,
+                            AAResults *CalleeAAR = nullptr,
+                            bool InsertLifetime = true,
+                            Function *ForwardVarArgsTo = nullptr);
+
 /// Clones a loop \p OrigLoop.  Returns the loop and the blocks in \p
 /// Blocks.
 ///

llvm/lib/Analysis/CtxProfAnalysis.cpp

@@ -29,6 +29,14 @@ cl::opt<std::string>
     UseCtxProfile("use-ctx-profile", cl::init(""), cl::Hidden,
                   cl::desc("Use the specified contextual profile file"));
 
+static cl::opt<CtxProfAnalysisPrinterPass::PrintMode> PrintLevel(
+    "ctx-profile-printer-level",
+    cl::init(CtxProfAnalysisPrinterPass::PrintMode::JSON), cl::Hidden,
+    cl::values(clEnumValN(CtxProfAnalysisPrinterPass::PrintMode::Everything,
+                          "everything", "print everything - most verbose"),
+               clEnumValN(CtxProfAnalysisPrinterPass::PrintMode::JSON, "json",
+                          "just the json representation of the profile")),
+    cl::desc("Verbosity level of the contextual profile printer pass."));
 
 namespace llvm {
 namespace json {
@@ -99,12 +107,20 @@ GlobalValue::GUID AssignGUIDPass::getGUID(const Function &F) {
 }
 AnalysisKey CtxProfAnalysis::Key;
 
-CtxProfAnalysis::CtxProfAnalysis(StringRef Profile)
-    : Profile(Profile.empty() ? UseCtxProfile : Profile) {}
+CtxProfAnalysis::CtxProfAnalysis(std::optional<StringRef> Profile)
+    : Profile([&]() -> std::optional<StringRef> {
+        if (Profile)
+          return *Profile;
+        if (UseCtxProfile.getNumOccurrences())
+          return UseCtxProfile;
+        return std::nullopt;
+      }()) {}
 
 PGOContextualProfile CtxProfAnalysis::run(Module &M,
                                           ModuleAnalysisManager &MAM) {
-  ErrorOr<std::unique_ptr<MemoryBuffer>> MB = MemoryBuffer::getFile(Profile);
+  if (!Profile)
+    return {};
+  ErrorOr<std::unique_ptr<MemoryBuffer>> MB = MemoryBuffer::getFile(*Profile);
   if (auto EC = MB.getError()) {
     M.getContext().emitError("could not open contextual profile file: " +
                              EC.message());
@@ -153,7 +169,6 @@ PGOContextualProfile CtxProfAnalysis::run(Module &M,
   // If we made it this far, the Result is valid - which we mark by setting
   // .Profiles.
   // Trim first the roots that aren't in this module.
-  DenseSet<GlobalValue::GUID> ProfiledGUIDs;
   for (auto &[RootGuid, _] : llvm::make_early_inc_range(*MaybeCtx))
     if (!Result.FuncInfo.contains(RootGuid))
       MaybeCtx->erase(RootGuid);
@@ -168,11 +183,14 @@ PGOContextualProfile::getDefinedFunctionGUID(const Function &F) const {
   return 0;
 }
 
+CtxProfAnalysisPrinterPass::CtxProfAnalysisPrinterPass(raw_ostream &OS)
+    : OS(OS), Mode(PrintLevel) {}
+
 PreservedAnalyses CtxProfAnalysisPrinterPass::run(Module &M,
                                                   ModuleAnalysisManager &MAM) {
   CtxProfAnalysis::Result &C = MAM.getResult<CtxProfAnalysis>(M);
   if (!C) {
-    M.getContext().emitError("Invalid CtxProfAnalysis");
+    OS << "No contextual profile was provided.\n";
     return PreservedAnalyses::all();
   }
 

llvm/lib/Transforms/IPO/ModuleInliner.cpp

@@ -20,6 +20,7 @@
 #include "llvm/Analysis/AliasAnalysis.h"
 #include "llvm/Analysis/AssumptionCache.h"
 #include "llvm/Analysis/BlockFrequencyInfo.h"
+#include "llvm/Analysis/CtxProfAnalysis.h"
 #include "llvm/Analysis/InlineAdvisor.h"
 #include "llvm/Analysis/InlineCost.h"
 #include "llvm/Analysis/InlineOrder.h"
@@ -113,6 +114,8 @@ PreservedAnalyses ModuleInlinerPass::run(Module &M,
     return PreservedAnalyses::all();
   }
 
+  auto &CtxProf = MAM.getResult<CtxProfAnalysis>(M);
+
   bool Changed = false;
 
   ProfileSummaryInfo *PSI = MAM.getCachedResult<ProfileSummaryAnalysis>(M);
@@ -213,7 +216,7 @@ PreservedAnalyses ModuleInlinerPass::run(Module &M,
         &FAM.getResult<BlockFrequencyAnalysis>(Callee));
 
     InlineResult IR =
-        InlineFunction(*CB, IFI, /*MergeAttributes=*/true,
+        InlineFunction(*CB, IFI, CtxProf, /*MergeAttributes=*/true,
                        &FAM.getResult<AAManager>(*CB->getCaller()));
     if (!IR.isSuccess()) {
       Advice->recordUnsuccessfulInlining(IR);