[PGO] Sampled instrumentation in PGO to speed up instrumentation bina…

…ry (#69535)

Summary:
In comparison to non-instrumented binaries, PGO instrumentation binaries
can be significantly slower. For highly threaded programs, this slowdown
can
reach 10x due to data races or false sharing within counters.

This patch incorporates sampling into the PGO instrumentation process to
enhance the speed of instrumentation binaries. The fundamental concept
is similar to the one proposed in https://reviews.llvm.org/D63949.

Three sampling modes are introduced:
1. Simple Sampling: When '-sampled-instr-bust-duration' is set to 1.
2. Fast Burst Sampling: When not using simple sampling, and 
'-sampled-instr-period' is set to 65535. This is the default mode of
sampling.
3. Full Burst Sampling: When neither simple nor fast burst sampling is
used.

Utilizing this sampled instrumentation significantly improves the
binary's
execution speed. Measurements show up to 5x speedup with default
settings. Fast burst sampling now results in only around 20% to 30%
slowdown (compared to 8 to 10x slowdown without sampling).

Out tests show that profile quality remains good with sampling,
with edge counts typically showing more than 90% overlap.
For applications whose behavior changes due to binary speed,
sampling instrumentation can enhance performance.
Observations have shown some apps experiencing up to
a ~2% improvement in PGO.

A potential drawback of this patch is the increased binary size
and compilation time. The Sampling method in this patch does
not improve single threaded program instrumentation binary
speed.

Test Plan: 

Reviewers: 

Subscribers: 

Tasks: 

Tags: 


Differential Revision: https://phabricator.intern.facebook.com/D60251137

llvm/include/llvm/ProfileData/InstrProfData.inc

@@ -739,6 +739,7 @@ serializeValueProfDataFrom(ValueProfRecordClosure *Closure,
 #define INSTR_PROF_PROFILE_RUNTIME_VAR __llvm_profile_runtime
 #define INSTR_PROF_PROFILE_COUNTER_BIAS_VAR __llvm_profile_counter_bias
 #define INSTR_PROF_PROFILE_SET_TIMESTAMP __llvm_profile_set_timestamp
+#define INSTR_PROF_PROFILE_SAMPLING_VAR __llvm_profile_sampling
 
 /* The variable that holds the name of the profile data
  * specified via command line. */

llvm/include/llvm/Transforms/Instrumentation.h

@@ -121,12 +121,18 @@ struct InstrProfOptions {
   // Use BFI to guide register promotion
   bool UseBFIInPromotion = false;
 
+  // Use sampling to reduce the profile instrumentation runtime overhead.
+  bool Sampling = false;
+
   // Name of the profile file to use as output
   std::string InstrProfileOutput;
 
   InstrProfOptions() = default;
 };
 
+// Create the variable for profile sampling.
+void createProfileSamplingVar(Module &M);
+
 // Options for sanitizer coverage instrumentation.
 struct SanitizerCoverageOptions {
   enum Type {

llvm/include/llvm/Transforms/Instrumentation/PGOInstrumentation.h

@@ -43,12 +43,14 @@ class FileSystem;
 class PGOInstrumentationGenCreateVar
     : public PassInfoMixin<PGOInstrumentationGenCreateVar> {
 public:
-  PGOInstrumentationGenCreateVar(std::string CSInstrName = "")
-      : CSInstrName(CSInstrName) {}
+  PGOInstrumentationGenCreateVar(std::string CSInstrName = "",
+                                 bool Sampling = false)
+      : CSInstrName(CSInstrName), ProfileSampling(Sampling) {}
   PreservedAnalyses run(Module &M, ModuleAnalysisManager &MAM);
 
 private:
   std::string CSInstrName;
+  bool ProfileSampling;
 };
 
 /// The instrumentation (profile-instr-gen) pass for IR based PGO.

llvm/lib/Passes/PassBuilderPipelines.cpp

@@ -296,6 +296,9 @@ static cl::opt<AttributorRunOption> AttributorRun(
                clEnumValN(AttributorRunOption::NONE, "none",
                           "disable attributor runs")));
 
+static cl::opt<bool> EnableSampledInstr(
+    "enable-sampled-instrumentation", cl::init(false), cl::Hidden,
+    cl::desc("Enable profile instrumentation sampling (default = off)"));
 static cl::opt<bool> UseLoopVersioningLICM(
     "enable-loop-versioning-licm", cl::init(false), cl::Hidden,
     cl::desc("Enable the experimental Loop Versioning LICM pass"));
@@ -847,6 +850,12 @@ void PassBuilder::addPGOInstrPasses(ModulePassManager &MPM,
   // Do counter promotion at Level greater than O0.
   Options.DoCounterPromotion = true;
   Options.UseBFIInPromotion = IsCS;
+  if (EnableSampledInstr) {
+    Options.Sampling = true;
+    // With sampling, there is little beneifit to enable counter promotion.
+    // But note that sampling does work with counter promotion.
+    Options.DoCounterPromotion = false;
+  }
   Options.Atomic = AtomicCounterUpdate;
   MPM.addPass(InstrProfilingLoweringPass(Options, IsCS));
 }
@@ -1185,7 +1194,8 @@ PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
     MPM.addPass(PGOIndirectCallPromotion(false, false));
 
   if (IsPGOPreLink && PGOOpt->CSAction == PGOOptions::CSIRInstr)
-    MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile));
+    MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile,
+                                               EnableSampledInstr));
 
   if (IsMemprofUse)
     MPM.addPass(MemProfUsePass(PGOOpt->MemoryProfile, PGOOpt->FS));

llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp

@@ -170,6 +170,30 @@ cl::opt<bool> SkipRetExitBlock(
     "skip-ret-exit-block", cl::init(true),
     cl::desc("Suppress counter promotion if exit blocks contain ret."));
 
+static cl::opt<bool> SampledInstr("sampled-instrumentation", cl::ZeroOrMore,
+                                  cl::init(false),
+                                  cl::desc("Do PGO instrumentation sampling"));
+
+static cl::opt<unsigned> SampledInstrPeriod(
+    "sampled-instr-period",
+    cl::desc("Set the profile instrumentation sample period. For each sample "
+             "period, a fixed number of consecutive samples will be recorded. "
+             "The number is controlled by 'sampled-instr-burst-duration' flag. "
+             "The default sample period of 65535 is optimized for generating "
+             "efficient code that leverages unsigned integer wrapping in "
+             "overflow."),
+    cl::init(65535));
+
+static cl::opt<unsigned> SampledInstrBurstDuration(
+    "sampled-instr-burst-duration",
+    cl::desc("Set the profile instrumentation burst duration, which can range "
+             "from 0 to one less than the value of 'sampled-instr-period'. "
+             "This number of samples will be recorded for each "
+             "'sampled-instr-period' count update. Setting to 1 enables "
+             "simple sampling, in which case it is recommended to set "
+             "'sampled-instr-period' to a prime number."),
+    cl::init(200));
+
 using LoadStorePair = std::pair<Instruction *, Instruction *>;
 
 static uint64_t getIntModuleFlagOrZero(const Module &M, StringRef Flag) {
@@ -260,6 +284,9 @@ class InstrLowerer final {
   /// Returns true if profile counter update register promotion is enabled.
   bool isCounterPromotionEnabled() const;
 
+  /// Return true if profile sampling is enabled.
+  bool isSamplingEnabled() const;
+
   /// Count the number of instrumented value sites for the function.
   void computeNumValueSiteCounts(InstrProfValueProfileInst *Ins);
 
@@ -291,6 +318,9 @@ class InstrLowerer final {
   /// acts on.
   Value *getCounterAddress(InstrProfCntrInstBase *I);
 
+  /// Lower the incremental instructions under profile sampling predicates.
+  void doSampling(Instruction *I);
+
   /// Get the region counters for an increment, creating them if necessary.
   ///
   /// If the counter array doesn't yet exist, the profile data variables
@@ -635,33 +665,169 @@ PreservedAnalyses InstrProfilingLoweringPass::run(Module &M,
   return PreservedAnalyses::none();
 }
 
+//
+// Perform instrumentation sampling.
+//
+// There are 3 favors of sampling:
+// (1) Full burst sampling: We transform:
+//   Increment_Instruction;
+// to:
+//   if (__llvm_profile_sampling__ < SampledInstrBurstDuration) {
+//     Increment_Instruction;
+//   }
+//   __llvm_profile_sampling__ += 1;
+//   if (__llvm_profile_sampling__ >= SampledInstrPeriod) {
+//     __llvm_profile_sampling__ = 0;
+//   }
+//
+// "__llvm_profile_sampling__" is a thread-local global shared by all PGO
+// counters (value-instrumentation and edge instrumentation).
+//
+// (2) Fast burst sampling:
+// "__llvm_profile_sampling__" variable is an unsigned type, meaning it will
+// wrap around to zero when overflows. In this case, the second check is
+// unnecessary, so we won't generate check2 when the SampledInstrPeriod is
+// set to 65535 (64K - 1). The code after:
+//   if (__llvm_profile_sampling__ < SampledInstrBurstDuration) {
+//     Increment_Instruction;
+//   }
+//   __llvm_profile_sampling__ += 1;
+//
+// (3) Simple sampling:
+// When SampledInstrBurstDuration sets to 1, we do a simple sampling:
+//   __llvm_profile_sampling__ += 1;
+//   if (__llvm_profile_sampling__ >= SampledInstrPeriod) {
+//     __llvm_profile_sampling__ = 0;
+//     Increment_Instruction;
+//   }
+//
+// Note that, the code snippet after the transformation can still be counter
+// promoted. However, with sampling enabled, counter updates are expected to
+// be infrequent, making the benefits of counter promotion negligible.
+// Moreover, counter promotion can potentially cause issues in server
+// applications, particularly when the counters are dumped without a clean
+// exit. To mitigate this risk, counter promotion is disabled by default when
+// sampling is enabled. This behavior can be overridden using the internal
+// option.
+void InstrLowerer::doSampling(Instruction *I) {
+  if (!isSamplingEnabled())
+    return;
+
+  unsigned SampledBurstDuration = SampledInstrBurstDuration.getValue();
+  unsigned SampledPeriod = SampledInstrPeriod.getValue();
+  if (SampledBurstDuration >= SampledPeriod) {
+    report_fatal_error(
+        "SampledPeriod needs to be greater than SampledBurstDuration");
+  }
+  bool UseShort = (SampledPeriod <= USHRT_MAX);
+  bool IsSimpleSampling = (SampledBurstDuration == 1);
+  // If (SampledBurstDuration == 1 && SampledPeriod == 65535), generate
+  // the simple sampling style code.
+  bool IsFastSampling = (!IsSimpleSampling && SampledPeriod == 65535);
+
+  auto GetConstant = [UseShort](IRBuilder<> &Builder, uint32_t C) {
+    if (UseShort)
+      return Builder.getInt16(C);
+    else
+      return Builder.getInt32(C);
+  };
+
+  IntegerType *SamplingVarTy;
+  if (UseShort)
+    SamplingVarTy = Type::getInt16Ty(M.getContext());
+  else
+    SamplingVarTy = Type::getInt32Ty(M.getContext());
+  auto *SamplingVar =
+      M.getGlobalVariable(INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_SAMPLING_VAR));
+  assert(SamplingVar && "SamplingVar not set properly");
+
+  // Create the condition for checking the burst duration.
+  Instruction *SamplingVarIncr;
+  Value *NewSamplingVarVal;
+  MDBuilder MDB(I->getContext());
+  MDNode *BranchWeight;
+  IRBuilder<> CondBuilder(I);
+  auto *LoadSamplingVar = CondBuilder.CreateLoad(SamplingVarTy, SamplingVar);
+  if (IsSimpleSampling) {
+    // For the simple sampling, just create the load and increments.
+    IRBuilder<> IncBuilder(I);
+    NewSamplingVarVal =
+        IncBuilder.CreateAdd(LoadSamplingVar, GetConstant(IncBuilder, 1));
+    SamplingVarIncr = IncBuilder.CreateStore(NewSamplingVarVal, SamplingVar);
+  } else {
+    // For the bust-sampling, create the conditonal update.
+    auto *DurationCond = CondBuilder.CreateICmpULE(
+        LoadSamplingVar, GetConstant(CondBuilder, SampledBurstDuration));
+    BranchWeight = MDB.createBranchWeights(
+        SampledBurstDuration, SampledPeriod + 1 - SampledBurstDuration);
+    Instruction *ThenTerm = SplitBlockAndInsertIfThen(
+        DurationCond, I, /* Unreachable */ false, BranchWeight);
+    IRBuilder<> IncBuilder(I);
+    NewSamplingVarVal =
+        IncBuilder.CreateAdd(LoadSamplingVar, GetConstant(IncBuilder, 1));
+    SamplingVarIncr = IncBuilder.CreateStore(NewSamplingVarVal, SamplingVar);
+    I->moveBefore(ThenTerm);
+  }
+
+  if (IsFastSampling)
+    return;
+
+  // Create the condtion for checking the period.
+  Instruction *ThenTerm, *ElseTerm;
+  IRBuilder<> PeriodCondBuilder(SamplingVarIncr);
+  auto *PeriodCond = PeriodCondBuilder.CreateICmpUGE(
+      NewSamplingVarVal, GetConstant(PeriodCondBuilder, SampledPeriod));
+  BranchWeight = MDB.createBranchWeights(1, SampledPeriod);
+  SplitBlockAndInsertIfThenElse(PeriodCond, SamplingVarIncr, &ThenTerm,
+                                &ElseTerm, BranchWeight);
+
+  // For the simple sampling, the counter update happens in sampling var reset.
+  if (IsSimpleSampling)
+    I->moveBefore(ThenTerm);
+
+  IRBuilder<> ResetBuilder(ThenTerm);
+  ResetBuilder.CreateStore(GetConstant(ResetBuilder, 0), SamplingVar);
+  SamplingVarIncr->moveBefore(ElseTerm);
+}
+
 bool InstrLowerer::lowerIntrinsics(Function *F) {
   bool MadeChange = false;
   PromotionCandidates.clear();
+  SmallVector<InstrProfInstBase *, 8> InstrProfInsts;
+
+  // To ensure compatibility with sampling, we save the intrinsics into
+  // a buffer to prevent potential breakage of the iterator (as the
+  // intrinsics will be moved to a different BB).
   for (BasicBlock &BB : *F) {
     for (Instruction &Instr : llvm::make_early_inc_range(BB)) {
-      if (auto *IPIS = dyn_cast<InstrProfIncrementInstStep>(&Instr)) {
-        lowerIncrement(IPIS);
-        MadeChange = true;
-      } else if (auto *IPI = dyn_cast<InstrProfIncrementInst>(&Instr)) {
-        lowerIncrement(IPI);
-        MadeChange = true;
-      } else if (auto *IPC = dyn_cast<InstrProfTimestampInst>(&Instr)) {
-        lowerTimestamp(IPC);
-        MadeChange = true;
-      } else if (auto *IPC = dyn_cast<InstrProfCoverInst>(&Instr)) {
-        lowerCover(IPC);
-        MadeChange = true;
-      } else if (auto *IPVP = dyn_cast<InstrProfValueProfileInst>(&Instr)) {
-        lowerValueProfileInst(IPVP);
-        MadeChange = true;
-      } else if (auto *IPMP = dyn_cast<InstrProfMCDCBitmapParameters>(&Instr)) {
-        IPMP->eraseFromParent();
-        MadeChange = true;
-      } else if (auto *IPBU = dyn_cast<InstrProfMCDCTVBitmapUpdate>(&Instr)) {
-        lowerMCDCTestVectorBitmapUpdate(IPBU);
-        MadeChange = true;
-      }
+      if (auto *IP = dyn_cast<InstrProfInstBase>(&Instr))
+        InstrProfInsts.push_back(IP);
+    }
+  }
+
+  for (auto *Instr : InstrProfInsts) {
+    doSampling(Instr);
+    if (auto *IPIS = dyn_cast<InstrProfIncrementInstStep>(Instr)) {
+      lowerIncrement(IPIS);
+      MadeChange = true;
+    } else if (auto *IPI = dyn_cast<InstrProfIncrementInst>(Instr)) {
+      lowerIncrement(IPI);
+      MadeChange = true;
+    } else if (auto *IPC = dyn_cast<InstrProfTimestampInst>(Instr)) {
+      lowerTimestamp(IPC);
+      MadeChange = true;
+    } else if (auto *IPC = dyn_cast<InstrProfCoverInst>(Instr)) {
+      lowerCover(IPC);
+      MadeChange = true;
+    } else if (auto *IPVP = dyn_cast<InstrProfValueProfileInst>(Instr)) {
+      lowerValueProfileInst(IPVP);
+      MadeChange = true;
+    } else if (auto *IPMP = dyn_cast<InstrProfMCDCBitmapParameters>(Instr)) {
+      IPMP->eraseFromParent();
+      MadeChange = true;
+    } else if (auto *IPBU = dyn_cast<InstrProfMCDCTVBitmapUpdate>(Instr)) {
+      lowerMCDCTestVectorBitmapUpdate(IPBU);
+      MadeChange = true;
     }
   }
 
@@ -684,6 +850,12 @@ bool InstrLowerer::isRuntimeCounterRelocationEnabled() const {
   return TT.isOSFuchsia();
 }
 
+bool InstrLowerer::isSamplingEnabled() const {
+  if (SampledInstr.getNumOccurrences() > 0)
+    return SampledInstr;
+  return Options.Sampling;
+}
+
 bool InstrLowerer::isCounterPromotionEnabled() const {
   if (DoCounterPromotion.getNumOccurrences() > 0)
     return DoCounterPromotion;
@@ -754,6 +926,9 @@ bool InstrLowerer::lower() {
   if (NeedsRuntimeHook)
     MadeChange = emitRuntimeHook();
 
+  if (!IsCS && isSamplingEnabled())
+    createProfileSamplingVar(M);
+
   bool ContainsProfiling = containsProfilingIntrinsics(M);
   GlobalVariable *CoverageNamesVar =
       M.getNamedGlobal(getCoverageUnusedNamesVarName());
@@ -1955,3 +2130,29 @@ void InstrLowerer::emitInitialization() {
 
   appendToGlobalCtors(M, F, 0);
 }
+
+namespace llvm {
+// Create the variable for profile sampling.
+void createProfileSamplingVar(Module &M) {
+  const StringRef VarName(INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_SAMPLING_VAR));
+  IntegerType *SamplingVarTy;
+  Constant *ValueZero;
+  if (SampledInstrPeriod.getValue() <= USHRT_MAX) {
+    SamplingVarTy = Type::getInt16Ty(M.getContext());
+    ValueZero = Constant::getIntegerValue(SamplingVarTy, APInt(16, 0));
+  } else {
+    SamplingVarTy = Type::getInt32Ty(M.getContext());
+    ValueZero = Constant::getIntegerValue(SamplingVarTy, APInt(32, 0));
+  }
+  auto SamplingVar = new GlobalVariable(
+      M, SamplingVarTy, false, GlobalValue::WeakAnyLinkage, ValueZero, VarName);
+  SamplingVar->setVisibility(GlobalValue::DefaultVisibility);
+  SamplingVar->setThreadLocal(true);
+  Triple TT(M.getTargetTriple());
+  if (TT.supportsCOMDAT()) {
+    SamplingVar->setLinkage(GlobalValue::ExternalLinkage);
+    SamplingVar->setComdat(M.getOrInsertComdat(VarName));
+  }
+  appendToCompilerUsed(M, SamplingVar);
+}
+} // namespace llvm

llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp

@@ -1875,6 +1875,8 @@ PGOInstrumentationGenCreateVar::run(Module &M, ModuleAnalysisManager &MAM) {
   // The variable in a comdat may be discarded by LTO. Ensure the declaration
   // will be retained.
   appendToCompilerUsed(M, createIRLevelProfileFlagVar(M, /*IsCS=*/true));
+  if (ProfileSampling)
+    createProfileSamplingVar(M);
   PreservedAnalyses PA;
   PA.preserve<FunctionAnalysisManagerModuleProxy>();
   PA.preserveSet<AllAnalysesOn<Function>>();