Introduce UnpredictableProfileLoader for PMU branch-miss profiles

llvm/include/llvm/Transforms/IPO/UnpredictableProfileLoader.h

@@ -0,0 +1,36 @@
+//===-- UnpredictableProfileLoader.h - Unpredictable Profile Loader -------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_TRANSFORMS_IPO_UNPREDICTABLEPROFILELOADER_H
+#define LLVM_TRANSFORMS_IPO_UNPREDICTABLEPROFILELOADER_H
+
+#include "llvm/IR/PassManager.h"
+#include "llvm/ProfileData/SampleProfReader.h"
+
+namespace llvm {
+
+class Module;
+
+struct UnpredictableProfileLoaderPass
+    : PassInfoMixin<UnpredictableProfileLoaderPass> {
+  UnpredictableProfileLoaderPass(StringRef FrequencyProfileFile);
+  UnpredictableProfileLoaderPass();
+  PreservedAnalyses run(Module &M, ModuleAnalysisManager &);
+  std::unique_ptr<SampleProfileReader> FreqReader, MispReader;
+  bool loadSampleProfile(Module &M);
+  bool addUpredictableMetadata(Module &F);
+  bool addUpredictableMetadata(Function &F);
+  ErrorOr<double> getMispredictRatio(const FunctionSamples *FreqSamples,
+                                     const FunctionSamples *MispSamples,
+                                     const Instruction *I);
+  const std::string FrequencyProfileFile;
+};
+
+} // end namespace llvm
+
+#endif // LLVM_TRANSFORMS_IPO_UNPREDICTABLEPROFILELOADER_H

llvm/lib/Passes/PassBuilder.cpp

@@ -176,6 +176,7 @@
 #include "llvm/Transforms/IPO/StripDeadPrototypes.h"
 #include "llvm/Transforms/IPO/StripSymbols.h"
 #include "llvm/Transforms/IPO/SyntheticCountsPropagation.h"
+#include "llvm/Transforms/IPO/UnpredictableProfileLoader.h"
 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
 #include "llvm/Transforms/InstCombine/InstCombine.h"
 #include "llvm/Transforms/Instrumentation.h"

llvm/lib/Passes/PassBuilderPipelines.cpp

@@ -67,6 +67,7 @@
 #include "llvm/Transforms/IPO/SampleProfile.h"
 #include "llvm/Transforms/IPO/SampleProfileProbe.h"
 #include "llvm/Transforms/IPO/SyntheticCountsPropagation.h"
+#include "llvm/Transforms/IPO/UnpredictableProfileLoader.h"
 #include "llvm/Transforms/IPO/WholeProgramDevirt.h"
 #include "llvm/Transforms/InstCombine/InstCombine.h"
 #include "llvm/Transforms/Instrumentation/CGProfile.h"
@@ -1092,6 +1093,9 @@ PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
     // Cache ProfileSummaryAnalysis once to avoid the potential need to insert
     // RequireAnalysisPass for PSI before subsequent non-module passes.
     MPM.addPass(RequireAnalysisPass<ProfileSummaryAnalysis, Module>());
+    // Run after inlining decisions made by SampleProfileLoader. This can apply
+    // mispredict metadata to specific inlined callees.
+    MPM.addPass(UnpredictableProfileLoaderPass(PGOOpt->ProfileFile));
     // Do not invoke ICP in the LTOPrelink phase as it makes it hard
     // for the profile annotation to be accurate in the LTO backend.
     if (!isLTOPreLink(Phase))

llvm/lib/Passes/PassRegistry.def

@@ -140,6 +140,7 @@ MODULE_PASS("strip-nonlinetable-debuginfo", StripNonLineTableDebugInfoPass())
 MODULE_PASS("synthetic-counts-propagation", SyntheticCountsPropagation())
 MODULE_PASS("trigger-crash-module", TriggerCrashModulePass())
 MODULE_PASS("trigger-verifier-error", TriggerVerifierErrorPass())
+MODULE_PASS("unpredictable-profile-loader", UnpredictableProfileLoaderPass())
 MODULE_PASS("tsan-module", ModuleThreadSanitizerPass())
 MODULE_PASS("verify", VerifierPass())
 MODULE_PASS("view-callgraph", CallGraphViewerPass())

llvm/lib/Transforms/IPO/CMakeLists.txt

@@ -43,6 +43,7 @@ add_llvm_component_library(LLVMipo
   StripSymbols.cpp
   SyntheticCountsPropagation.cpp
   ThinLTOBitcodeWriter.cpp
+  UnpredictableProfileLoader.cpp
   WholeProgramDevirt.cpp
 
   ADDITIONAL_HEADER_DIRS

llvm/lib/Transforms/IPO/UnpredictableProfileLoader.cpp

@@ -0,0 +1,226 @@
+//=== UnpredictableProfileLoader.cpp - Unpredictable Profile Loader -------===//
+//
+// 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 pass reads a sample profile containing mispredict counts and a sample
+// profile containing execution counts and computes branch mispredict ratios for
+// each conditional instruction. If a sufficiently high mispredict ratio is
+// found !unpredictable metadata is added.
+//
+// Note that this requires that the mispredict and frequency profiles have
+// comparable magnitudes.
+//
+//===----------------------------------------------------------------------===//
+
+#include "llvm/Transforms/IPO/UnpredictableProfileLoader.h"
+#include "llvm/IR/DebugInfoMetadata.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Instructions.h"
+#include "llvm/IR/Module.h"
+#include "llvm/ProfileData/SampleProf.h"
+#include "llvm/ProfileData/SampleProfReader.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/VirtualFileSystem.h"
+#include "llvm/Transforms/IPO.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "unpredictable-profile-loader"
+
+static cl::opt<std::string> UnpredictableHintsFile(
+    "unpredictable-hints-file",
+    cl::desc("Path to the unpredictability hints profile"), cl::Hidden);
+
+// Typically this file will be provided via PGOOpt. This option is provided
+// primarily for debugging and testing.
+static cl::opt<std::string>
+    FrequencyProfileOption("unpredictable-hints-frequency-profile",
+                           cl::desc("Path to an execution frequency profile to "
+                                    "use as a baseline for unpredictability"),
+                           cl::Hidden);
+
+// This determines the minimum apparent mispredict ratio which should earn a
+// mispredict metadata annotation.
+static cl::opt<double> MinimumRatio(
+    "unpredictable-hints-min-ratio",
+    cl::desc(
+        "Absolute minimum branch miss ratio to apply MD_unpredictable from"),
+    cl::init(0.2), cl::Hidden);
+
+// This option is useful for dealing with two different sampling frequencies.
+static cl::opt<double>
+    RatioFactor("unpredictable-hints-factor",
+                cl::desc("Multiply all ratios by this factor"), cl::init(1.0),
+                cl::ReallyHidden);
+
+// Lookup execution frequency and mispredict samples for an Instruction's
+// corresponding location in a the two FunctionSamples profiles and compute an
+// effective branch mispredict ratio. The counts used to compute the ratio are
+// uint64s read directly from the profile files.
+ErrorOr<double> UnpredictableProfileLoaderPass::getMispredictRatio(
+    const FunctionSamples *FuncFreqSamples,
+    const FunctionSamples *FuncMispSamples, const Instruction *I) {
+
+  const auto &Loc = I->getDebugLoc();
+  if (!Loc)
+    return std::error_code();
+
+  const FunctionSamples *FreqSamples =
+      FuncFreqSamples->findFunctionSamples(Loc, FreqReader->getRemapper());
+  if (!FreqSamples)
+    return std::error_code();
+  const ErrorOr<uint64_t> FreqCount = FreqSamples->findSamplesAt(
+      FunctionSamples::getOffset(Loc), Loc->getBaseDiscriminator());
+  if (!FreqCount)
+    return std::error_code();
+
+  const FunctionSamples *MispSamples =
+      FuncMispSamples->findFunctionSamples(Loc, MispReader->getRemapper());
+  if (!MispSamples)
+    return std::error_code();
+  const ErrorOr<uint64_t> MispCount = MispSamples->findSamplesAt(
+      FunctionSamples::getOffset(Loc), Loc->getBaseDiscriminator());
+  if (!MispCount)
+    return std::error_code();
+
+  const double Freq = FreqCount.get();
+  if (!Freq)
+    return std::error_code();
+
+  const double Misp = MispCount.get();
+  const double MissRatio = (Misp * RatioFactor) / Freq;
+
+  LLVM_DEBUG(dbgs() << "Computing mispredict ratio of " << format("%0.2f", Misp)
+                    << "/" << format("%0.2f", Freq) << " * "
+                    << format("%0.2f", RatioFactor.getValue()) << " = "
+                    << format("%0.2f", MissRatio) << " for instruction\n"
+                    << *I << "\n");
+  return MissRatio;
+}
+
+// Examine all Branch, Select, and SwitchInsts in a function, adding
+// !unpredictable metadata if they appear in the mispredict profile with
+// sufficient weight.
+bool UnpredictableProfileLoaderPass::addUpredictableMetadata(Function &F) {
+
+  const FunctionSamples *FreqSamples = FreqReader->getSamplesFor(F);
+  if (!FreqSamples)
+    return false;
+
+  const FunctionSamples *MispSamples = MispReader->getSamplesFor(F);
+  if (!MispSamples)
+    return false;
+
+  bool MadeChange = false;
+  for (BasicBlock &BB : F) {
+    for (Instruction &I : BB) {
+      if (!isa<BranchInst>(&I) && !isa<SelectInst>(&I) && !isa<SwitchInst>(&I))
+        continue;
+      if (I.hasMetadata(LLVMContext::MD_unpredictable))
+        continue;
+
+      const ErrorOr<double> RatioOrError =
+          getMispredictRatio(FreqSamples, MispSamples, &I);
+      if (!RatioOrError)
+        continue;
+      const double MissRatio = RatioOrError.get();
+
+      if (MissRatio < MinimumRatio) {
+        LLVM_DEBUG(dbgs() << "\tRatio " << format("%0.2f", MissRatio)
+                          << " is below threshold of "
+                          << format("%0.2f", MinimumRatio.getValue())
+                          << "; ignoring.\n");
+        continue;
+      }
+
+      // In the future we probably want to attach more information here, such as
+      // the mispredict count or ratio.
+      MDNode *MD = MDNode::get(I.getContext(), std::nullopt);
+      I.setMetadata(LLVMContext::MD_unpredictable, MD);
+      MadeChange = true;
+    }
+  }
+
+  return MadeChange;
+}
+
+bool UnpredictableProfileLoaderPass::addUpredictableMetadata(Module &M) {
+  bool MadeChange = false;
+
+  for (Function &F : M)
+    MadeChange |= addUpredictableMetadata(F);
+
+  // Return an indication of whether we changed anything or not.
+  return MadeChange;
+}
+
+bool UnpredictableProfileLoaderPass::loadSampleProfile(Module &M) {
+  if (MispReader && FreqReader)
+    return true;
+
+  assert(!MispReader && !FreqReader &&
+         "Expected both or neither profile readers");
+
+  LLVMContext &Ctx = M.getContext();
+  auto FS = vfs::getRealFileSystem();
+
+  auto ReadProfile = [&Ctx,
+                      &FS](const std::string ProfileFile,
+                           std::unique_ptr<SampleProfileReader> &ReaderPtr) {
+    if (ProfileFile.empty())
+      return false;
+
+    ErrorOr<std::unique_ptr<SampleProfileReader>> ReaderOrErr =
+        SampleProfileReader::create(ProfileFile, Ctx, *FS);
+    if (std::error_code EC = ReaderOrErr.getError()) {
+      std::string Msg = "Could not open profile: " + EC.message();
+      Ctx.diagnose(DiagnosticInfoSampleProfile(ProfileFile, Msg,
+                                               DiagnosticSeverity::DS_Warning));
+      return false;
+    }
+
+    ReaderPtr = std::move(ReaderOrErr.get());
+    if (std::error_code EC = ReaderPtr->read()) {
+      std::string Msg = "Profile reading failed: " + EC.message();
+      Ctx.diagnose(DiagnosticInfoSampleProfile(ProfileFile, Msg));
+      return false;
+    }
+
+    return true;
+  };
+
+  if (!ReadProfile(UnpredictableHintsFile, MispReader))
+    return false;
+
+  if (!ReadProfile(FrequencyProfileFile, FreqReader))
+    return false;
+
+  return true;
+}
+
+UnpredictableProfileLoaderPass::UnpredictableProfileLoaderPass()
+    : FrequencyProfileFile(FrequencyProfileOption) {}
+
+UnpredictableProfileLoaderPass::UnpredictableProfileLoaderPass(
+    StringRef PGOProfileFile)
+    : FrequencyProfileFile(FrequencyProfileOption.empty()
+                               ? PGOProfileFile
+                               : FrequencyProfileOption) {}
+
+PreservedAnalyses UnpredictableProfileLoaderPass::run(Module &M,
+                                                      ModuleAnalysisManager &) {
+  if (!loadSampleProfile(M))
+    return PreservedAnalyses::all();
+
+  if (addUpredictableMetadata(M)) {
+    PreservedAnalyses PA;
+    PA.preserveSet<CFGAnalyses>();
+    return PA;
+  }
+
+  return PreservedAnalyses::all();
+}

llvm/test/Other/new-pm-pgo.ll

@@ -25,6 +25,7 @@
 ; SAMPLE_USE_PRE_LINK: Running pass: SROAPass
 ; SAMPLE_USE_PRE_LINK: Running pass: EarlyCSEPass
 ; SAMPLE_USE: Running pass: SampleProfileLoaderPass
+; SAMPLE_USE: Running pass: UnpredictableProfileLoaderPass
 ; SAMPLE_USE_O: Running pass: PGOIndirectCallPromotion
 ; SAMPLE_USE_POST_LINK-NOT: Running pass: GlobalOptPass
 ; SAMPLE_USE_POST_LINK: Running pass: PGOIndirectCallPromotion

llvm/test/Other/new-pm-thinlto-postlink-samplepgo-defaults.ll

@@ -35,6 +35,7 @@
 ; CHECK-O-NEXT: Running analysis: LazyCallGraphAnalysis
 ; CHECK-O-NEXT: Running analysis: TargetLibraryAnalysis
 ; CHECK-O-NEXT: Running pass: RequireAnalysisPass<{{.*}}ProfileSummaryAnalysis
+; CHECK-O-NEXT: Running pass: UnpredictableProfileLoaderPass
 ; CHECK-O-NEXT: Running pass: PGOIndirectCallPromotion
 ; CHECK-O-NEXT: Running analysis: OptimizationRemarkEmitterAnalysis
 ; CHECK-O-NEXT: Running pass: OpenMPOptPass

llvm/test/Other/new-pm-thinlto-prelink-samplepgo-defaults.ll

@@ -47,6 +47,7 @@
 ; CHECK-O-NEXT: Running analysis: ProfileSummaryAnalysis
 ; CHECK-O-NEXT: Running analysis: LazyCallGraphAnalysis
 ; CHECK-O-NEXT: Running pass: RequireAnalysisPass<{{.*}}ProfileSummaryAnalysis
+; CHECK-O-NEXT: Running pass: UnpredictableProfileLoaderPass
 ; CHECK-O-NEXT: Running pass: OpenMPOptPass
 ; CHECK-O-NEXT: Running pass: IPSCCPPass
 ; CHECK-O-NEXT: Running pass: CalledValuePropagationPass

llvm/test/Transforms/UnpredictableProfileLoader/Inputs/frequency.prof

@@ -0,0 +1,3 @@
+# This is a standard SPGO profile indicating basic block execution frequency.
+sel_arr:1:0
+ 11: 4000

llvm/test/Transforms/UnpredictableProfileLoader/Inputs/inline.freq.prof

@@ -0,0 +1,3 @@
+caller:1:0
+ 1: callee:1
+  3: 997

llvm/test/Transforms/UnpredictableProfileLoader/Inputs/inline.misp.prof

@@ -0,0 +1,3 @@
+caller:1:0
+ 1: callee:1
+  3: 400

llvm/test/Transforms/UnpredictableProfileLoader/Inputs/mispredict.prof

@@ -0,0 +1,4 @@
+# This profile indicates 1000 mispredict samples for instructions 11 source
+# lines into in the sel_arr function.
+sel_arr:1:0
+  11: 1000