[llvm] [clang] [clang-tools-extra] [PGO] Sampled instrumentation in PGO to speed up instrumentation binary (PR #69535)

via cfe-commits cfe-commits at lists.llvm.org
Tue Jan 23 16:51:56 PST 2024


https://github.com/xur-llvm updated https://github.com/llvm/llvm-project/pull/69535

>From 62728756fd56e3427376268c4178765950a27636 Mon Sep 17 00:00:00 2001
From: Rong Xu <xur at google.com>
Date: Wed, 18 Oct 2023 15:13:15 -0700
Subject: [PATCH 1/2] [PGO] Sampled instrumentation in PGO to speed up
 instrumentation binary

PGO instrumentation binary can be very slow comparing to the
non-instrumented binary. It's not uncommon to see 10x slowdown
for highly threaded programs, due to data race of false sharing
to the counters.

This patch uses sampling in PGO instrumentation to speed up
instrumentation binary. The basic idea is the same as one:
here: https://reviews.llvm.org/D63949

This patch makes some improvements so that we only use one
condition. We now fix the WholeDuring at 65536 and use the
wraparound of unsigned short.

With this sampled instrumentation, the binary runs much
faster. We measure 5x speedup using the default duration.
We now only see about 20% to 30% slow down (comparing to
8 to 10x slowdown without sampling).

The profile quality is pretty good with sampling: the edge
counts usually report >90% overlap.

For the apps that program behaviors change due to binary
speed, sampling instrumentation can improve the performance.
We have observed some apps getting up ~2% improvement in PGO.

One potential issue of this patch is the increased binary
size and compilation time.
---
 .../llvm/ProfileData/InstrProfData.inc        |   1 +
 .../include/llvm/Transforms/Instrumentation.h |   6 +
 .../Instrumentation/InstrProfiling.h          |   6 +
 .../Instrumentation/PGOInstrumentation.h      |   6 +-
 llvm/lib/Passes/PassBuilderPipelines.cpp      |  10 +-
 .../Instrumentation/InstrProfiling.cpp        | 131 ++++++++++++++++--
 .../Instrumentation/PGOInstrumentation.cpp    |   2 +
 .../PGOProfile/Inputs/cspgo_bar_sample.ll     |  82 +++++++++++
 .../PGOProfile/counter_promo_sampling.ll      |  78 +++++++++++
 .../Transforms/PGOProfile/cspgo_sample.ll     | 112 +++++++++++++++
 .../Transforms/PGOProfile/instrprof_sample.ll |  47 +++++++
 11 files changed, 463 insertions(+), 18 deletions(-)
 create mode 100644 llvm/test/Transforms/PGOProfile/Inputs/cspgo_bar_sample.ll
 create mode 100644 llvm/test/Transforms/PGOProfile/counter_promo_sampling.ll
 create mode 100644 llvm/test/Transforms/PGOProfile/cspgo_sample.ll
 create mode 100644 llvm/test/Transforms/PGOProfile/instrprof_sample.ll

diff --git a/llvm/include/llvm/ProfileData/InstrProfData.inc b/llvm/include/llvm/ProfileData/InstrProfData.inc
index 13be2753e514ef..6294505ac39685 100644
--- a/llvm/include/llvm/ProfileData/InstrProfData.inc
+++ b/llvm/include/llvm/ProfileData/InstrProfData.inc
@@ -676,6 +676,7 @@ 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. */
diff --git a/llvm/include/llvm/Transforms/Instrumentation.h b/llvm/include/llvm/Transforms/Instrumentation.h
index 392983a1984445..76d4e1de75154f 100644
--- a/llvm/include/llvm/Transforms/Instrumentation.h
+++ b/llvm/include/llvm/Transforms/Instrumentation.h
@@ -116,12 +116,18 @@ 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 {
diff --git a/llvm/include/llvm/Transforms/Instrumentation/InstrProfiling.h b/llvm/include/llvm/Transforms/Instrumentation/InstrProfiling.h
index cb0c055dcb74ae..d0581ff72a1586 100644
--- a/llvm/include/llvm/Transforms/Instrumentation/InstrProfiling.h
+++ b/llvm/include/llvm/Transforms/Instrumentation/InstrProfiling.h
@@ -86,6 +86,9 @@ class InstrProfiling : public PassInfoMixin<InstrProfiling> {
   /// 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);
 
@@ -109,6 +112,9 @@ class InstrProfiling : public PassInfoMixin<InstrProfiling> {
   /// acts on.
   Value *getCounterAddress(InstrProfInstBase *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
diff --git a/llvm/include/llvm/Transforms/Instrumentation/PGOInstrumentation.h b/llvm/include/llvm/Transforms/Instrumentation/PGOInstrumentation.h
index 5b1977b7de9a2a..7199f27dbc991a 100644
--- a/llvm/include/llvm/Transforms/Instrumentation/PGOInstrumentation.h
+++ b/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.
diff --git a/llvm/lib/Passes/PassBuilderPipelines.cpp b/llvm/lib/Passes/PassBuilderPipelines.cpp
index 600f8d43caaf21..5595f92e24aa86 100644
--- a/llvm/lib/Passes/PassBuilderPipelines.cpp
+++ b/llvm/lib/Passes/PassBuilderPipelines.cpp
@@ -273,6 +273,9 @@ static cl::opt<AttributorRunOption> AttributorRun(
                clEnumValN(AttributorRunOption::NONE, "none",
                           "disable attributor runs")));
 
+static cl::opt<bool> EnableSampledInstr(
+    "enable-sampled-instr", 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"));
@@ -805,6 +808,10 @@ void PassBuilder::addPGOInstrPasses(ModulePassManager &MPM,
   // Do counter promotion at Level greater than O0.
   Options.DoCounterPromotion = true;
   Options.UseBFIInPromotion = IsCS;
+  if (EnableSampledInstr) {
+    Options.Sampling = true;
+    Options.DoCounterPromotion = false;
+  }
   Options.Atomic = AtomicCounterUpdate;
   MPM.addPass(InstrProfiling(Options, IsCS));
 }
@@ -1117,7 +1124,8 @@ PassBuilder::buildModuleSimplificationPipeline(OptimizationLevel Level,
   }
   if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
       PGOOpt->CSAction == PGOOptions::CSIRInstr)
-    MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile));
+    MPM.addPass(PGOInstrumentationGenCreateVar(PGOOpt->CSProfileGenFile,
+                                               EnableSampledInstr));
 
   if (PGOOpt && Phase != ThinOrFullLTOPhase::ThinLTOPostLink &&
       !PGOOpt->MemoryProfile.empty())
diff --git a/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp b/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp
index 57fcfd53836911..89e8e152fcee7e 100644
--- a/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp
+++ b/llvm/lib/Transforms/Instrumentation/InstrProfiling.cpp
@@ -36,6 +36,7 @@
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/IntrinsicInst.h"
+#include "llvm/IR/MDBuilder.h"
 #include "llvm/IR/Module.h"
 #include "llvm/IR/Type.h"
 #include "llvm/InitializePasses.h"
@@ -48,6 +49,7 @@
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/TargetParser/Triple.h"
 #include "llvm/Transforms/Instrumentation/PGOInstrumentation.h"
+#include "llvm/Transforms/Utils/BasicBlockUtils.h"
 #include "llvm/Transforms/Utils/ModuleUtils.h"
 #include "llvm/Transforms/Utils/SSAUpdater.h"
 #include <algorithm>
@@ -148,6 +150,16 @@ 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>
+    SampledInstrument("sampled-instr", cl::ZeroOrMore, cl::init(false),
+                      cl::desc("Do PGO instrumentation sampling"));
+
+static cl::opt<unsigned> SampledInstrumentDuration(
+    "sampled-instr-duration",
+    cl::desc("Set the sample rate for profile instrumentation, with a value "
+             "range 0 to 65535. We will record this number of samples for "
+             "every 65536 count updates"),
+    cl::init(200));
 ///
 /// A helper class to promote one counter RMW operation in the loop
 /// into register update.
@@ -412,30 +424,91 @@ PreservedAnalyses InstrProfiling::run(Module &M, ModuleAnalysisManager &AM) {
   return PreservedAnalyses::none();
 }
 
+// Perform instrumentation sampling.
+// We transform:
+//   Increment_Instruction;
+// to:
+//   if (__llvm_profile_sampling__ <= SampleDuration) {
+//     Increment_Instruction;
+//   }
+//   __llvm_profile_sampling__ += 1;
+//
+// "__llvm_profile_sampling__" is a thread-local global shared by all PGO
+// instrumentation variables (value-instrumentation and edge instrumentation).
+// It has a unsigned short type and will wrapper around when overflow.
+//
+// Note that, the code snippet after the transformation can still be
+// counter promoted. But I don't see a reason for that because the
+// counter updated should be sparse. That's the reason we disable
+// counter promotion by default when sampling is enabled.
+// This can be overwritten by the internal option.
+//
+void InstrProfiling::doSampling(Instruction *I) {
+  if (!isSamplingEnabled())
+    return;
+  int SampleDuration = SampledInstrumentDuration.getValue();
+  unsigned WrapToZeroValue = USHRT_MAX + 1;
+  assert(SampleDuration < USHRT_MAX);
+  auto *Int16Ty = Type::getInt16Ty(M->getContext());
+  auto *CountVar =
+      M->getGlobalVariable(INSTR_PROF_QUOTE(INSTR_PROF_PROFILE_SAMPLING_VAR));
+  assert(CountVar && "CountVar not set properly");
+  IRBuilder<> CondBuilder(I);
+  auto *LoadCountVar = CondBuilder.CreateLoad(Int16Ty, CountVar);
+  auto *DurationCond = CondBuilder.CreateICmpULE(
+      LoadCountVar, CondBuilder.getInt16(SampleDuration));
+  MDBuilder MDB(I->getContext());
+  MDNode *BranchWeight =
+      MDB.createBranchWeights(SampleDuration, WrapToZeroValue - SampleDuration);
+  Instruction *ThenTerm = SplitBlockAndInsertIfThen(
+      DurationCond, I, /* Unreacheable */ false, BranchWeight);
+  IRBuilder<> IncBuilder(I);
+  auto *NewVal = IncBuilder.CreateAdd(LoadCountVar, IncBuilder.getInt16(1));
+  IncBuilder.CreateStore(NewVal, CountVar);
+  I->moveBefore(ThenTerm);
+}
+
 bool InstrProfiling::lowerIntrinsics(Function *F) {
   bool MadeChange = false;
   PromotionCandidates.clear();
+  SmallVector<InstrProfInstBase *, 8> InstrProfInsts;
+
   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;
+      if (auto *IP = dyn_cast<InstrProfInstBase>(&Instr)) {
+        InstrProfInsts.push_back(IP);
       }
     }
   }
 
+  for (auto *IP : InstrProfInsts) {
+    if (auto *IPIS = dyn_cast<InstrProfIncrementInstStep>(IP)) {
+      doSampling(IP);
+      lowerIncrement(IPIS);
+      MadeChange = true;
+    } else if (auto *IPI = dyn_cast<InstrProfIncrementInst>(IP)) {
+      doSampling(IP);
+      lowerIncrement(IPI);
+      MadeChange = true;
+    } else if (auto *IPC = dyn_cast<InstrProfTimestampInst>(IP)) {
+      doSampling(IP);
+      lowerTimestamp(IPC);
+      MadeChange = true;
+    } else if (auto *IPC = dyn_cast<InstrProfCoverInst>(IP)) {
+      doSampling(IP);
+      lowerCover(IPC);
+      MadeChange = true;
+    } else if (auto *IPVP = dyn_cast<InstrProfValueProfileInst>(IP)) {
+      doSampling(IP);
+      lowerValueProfileInst(IPVP);
+      MadeChange = true;
+    } else {
+      LLVM_DEBUG(dbgs() << "Invalid InstroProf intrinsic: " << *IP << "\n");
+      // ?? Seeing "call void @llvm.memcpy.p0.p0.i64..." here ??
+      // llvm_unreachable("Invalid InstroProf intrinsic");
+    }
+  }
+
   if (!MadeChange)
     return false;
 
@@ -455,6 +528,12 @@ bool InstrProfiling::isRuntimeCounterRelocationEnabled() const {
   return TT.isOSFuchsia();
 }
 
+bool InstrProfiling::isSamplingEnabled() const {
+  if (SampledInstrument.getNumOccurrences() > 0)
+    return SampledInstrument;
+  return Options.Sampling;
+}
+
 bool InstrProfiling::isCounterPromotionEnabled() const {
   if (DoCounterPromotion.getNumOccurrences() > 0)
     return DoCounterPromotion;
@@ -535,6 +614,9 @@ bool InstrProfiling::run(
   if (NeedsRuntimeHook)
     MadeChange = emitRuntimeHook();
 
+  if (!IsCS && isSamplingEnabled())
+    createProfileSamplingVar(M);
+
   bool ContainsProfiling = containsProfilingIntrinsics(M);
   GlobalVariable *CoverageNamesVar =
       M.getNamedGlobal(getCoverageUnusedNamesVarName());
@@ -1372,3 +1454,22 @@ void InstrProfiling::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));
+  Type *IntTy16 = Type::getInt16Ty(M.getContext());
+  auto SamplingVar = new GlobalVariable(
+      M, IntTy16, false, GlobalValue::WeakAnyLinkage,
+      Constant::getIntegerValue(IntTy16, APInt(16, 0)), 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
diff --git a/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp b/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
index 7ad1c9bc54f378..0ea6398fbdedc1 100644
--- a/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
+++ b/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
@@ -1820,6 +1820,8 @@ 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>>();
diff --git a/llvm/test/Transforms/PGOProfile/Inputs/cspgo_bar_sample.ll b/llvm/test/Transforms/PGOProfile/Inputs/cspgo_bar_sample.ll
new file mode 100644
index 00000000000000..1c8be82715f253
--- /dev/null
+++ b/llvm/test/Transforms/PGOProfile/Inputs/cspgo_bar_sample.ll
@@ -0,0 +1,82 @@
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+$__llvm_profile_filename = comdat any
+$__llvm_profile_raw_version = comdat any
+$__llvm_profile_sampling = comdat any
+
+ at odd = common dso_local local_unnamed_addr global i32 0, align 4
+ at even = common dso_local local_unnamed_addr global i32 0, align 4
+ at __llvm_profile_filename = local_unnamed_addr constant [25 x i8] c"pass2/default_%m.profraw\00", comdat
+ at __llvm_profile_raw_version = local_unnamed_addr constant i64 216172782113783812, comdat
+ at __llvm_profile_sampling = thread_local global i16 0, comdat
+ at llvm.used = appending global [1 x i8*] [i8* bitcast (i64* @__llvm_profile_sampling to i8*)], section "llvm.metadata"
+
+define dso_local void @bar(i32 %n) !prof !30 {
+entry:
+  %call = tail call fastcc i32 @cond(i32 %n)
+  %tobool = icmp eq i32 %call, 0
+  br i1 %tobool, label %if.else, label %if.then, !prof !31
+
+if.then:
+  %0 = load i32, i32* @odd, align 4, !tbaa !32
+  %inc = add i32 %0, 1
+  store i32 %inc, i32* @odd, align 4, !tbaa !32
+  br label %if.end
+
+if.else:
+  %1 = load i32, i32* @even, align 4, !tbaa !32
+  %inc1 = add i32 %1, 1
+  store i32 %inc1, i32* @even, align 4, !tbaa !32
+  br label %if.end
+
+if.end:
+  ret void
+}
+
+define internal fastcc i32 @cond(i32 %i) #1 !prof !30 !PGOFuncName !36 {
+entry:
+  %rem = srem i32 %i, 2
+  ret i32 %rem
+}
+
+attributes #1 = { inlinehint noinline }
+
+!llvm.module.flags = !{!0, !1, !2}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"EnableSplitLTOUnit", i32 0}
+!2 = !{i32 1, !"ProfileSummary", !3}
+!3 = !{!4, !5, !6, !7, !8, !9, !10, !11}
+!4 = !{!"ProfileFormat", !"InstrProf"}
+!5 = !{!"TotalCount", i64 500002}
+!6 = !{!"MaxCount", i64 200000}
+!7 = !{!"MaxInternalCount", i64 100000}
+!8 = !{!"MaxFunctionCount", i64 200000}
+!9 = !{!"NumCounts", i64 6}
+!10 = !{!"NumFunctions", i64 4}
+!11 = !{!"DetailedSummary", !12}
+!12 = !{!13, !14, !15, !16, !17, !18, !19, !20, !21, !22, !23, !24, !25, !26, !27, !28}
+!13 = !{i32 10000, i64 200000, i32 1}
+!14 = !{i32 100000, i64 200000, i32 1}
+!15 = !{i32 200000, i64 200000, i32 1}
+!16 = !{i32 300000, i64 200000, i32 1}
+!17 = !{i32 400000, i64 200000, i32 1}
+!18 = !{i32 500000, i64 100000, i32 4}
+!19 = !{i32 600000, i64 100000, i32 4}
+!20 = !{i32 700000, i64 100000, i32 4}
+!21 = !{i32 800000, i64 100000, i32 4}
+!22 = !{i32 900000, i64 100000, i32 4}
+!23 = !{i32 950000, i64 100000, i32 4}
+!24 = !{i32 990000, i64 100000, i32 4}
+!25 = !{i32 999000, i64 100000, i32 4}
+!26 = !{i32 999900, i64 100000, i32 4}
+!27 = !{i32 999990, i64 100000, i32 4}
+!28 = !{i32 999999, i64 1, i32 6}
+!30 = !{!"function_entry_count", i64 200000}
+!31 = !{!"branch_weights", i32 100000, i32 100000}
+!32 = !{!33, !33, i64 0}
+!33 = !{!"int", !34, i64 0}
+!34 = !{!"omnipotent char", !35, i64 0}
+!35 = !{!"Simple C/C++ TBAA"}
+!36 = !{!"cspgo_bar.c:cond"}
diff --git a/llvm/test/Transforms/PGOProfile/counter_promo_sampling.ll b/llvm/test/Transforms/PGOProfile/counter_promo_sampling.ll
new file mode 100644
index 00000000000000..6f13196a724994
--- /dev/null
+++ b/llvm/test/Transforms/PGOProfile/counter_promo_sampling.ll
@@ -0,0 +1,78 @@
+; RUN: opt < %s --passes=pgo-instr-gen,instrprof -do-counter-promotion=true -sampled-instr=true -skip-ret-exit-block=0 -S | FileCheck --check-prefixes=SAMPLING,PROMO %s
+
+; SAMPLING: $__llvm_profile_sampling = comdat any
+; SAMPLING: @__llvm_profile_sampling = thread_local global i16 0, comdat
+
+define void @foo(i32 %n, i32 %N) {
+; SAMPLING-LABEL: @foo
+; SAMPLING:  %[[VV0:[0-9]+]] = load i16, ptr @__llvm_profile_sampling, align 2
+; SAMPLING:  %[[VV1:[0-9]+]] = icmp ule i16 %[[VV0]], 200
+; SAMPLING:  br i1 %[[VV1]], label {{.*}}, label {{.*}}, !prof !0
+; SAMPLING: {{.*}} = load {{.*}} @__profc_foo{{.*}} 3)
+; SAMPLING-NEXT: add
+; SAMPLING-NEXT: store {{.*}}@__profc_foo{{.*}}3)
+bb:
+  %tmp = add nsw i32 %n, 1
+  %tmp1 = add nsw i32 %n, -1
+  br label %bb2
+
+bb2:
+; PROMO: phi {{.*}}
+; PROMO-NEXT: phi {{.*}}
+; PROMO-NEXT: phi {{.*}}
+; PROMO-NEXT: phi {{.*}}
+  %i.0 = phi i32 [ 0, %bb ], [ %tmp10, %bb9 ]
+  %tmp3 = icmp slt i32 %i.0, %tmp
+  br i1 %tmp3, label %bb4, label %bb5
+
+bb4:
+  tail call void @bar(i32 1)
+  br label %bb9
+
+bb5:
+  %tmp6 = icmp slt i32 %i.0, %tmp1
+  br i1 %tmp6, label %bb7, label %bb8
+
+bb7:
+  tail call void @bar(i32 2)
+  br label %bb9
+
+bb8:
+  tail call void @bar(i32 3)
+  br label %bb9
+
+bb9:
+; SAMPLING:       phi {{.*}}
+; SAMPLING-NEXT:  %[[V1:[0-9]+]] = add i16 {{.*}}, 1
+; SAMPLING-NEXT:  store i16 %[[V1]], ptr @__llvm_profile_sampling, align 2
+; SAMPLING:       phi {{.*}}
+; SAMPLING-NEXT:  %[[V2:[0-9]+]] = add i16 {{.*}}, 1
+; SAMPLING-NEXT:  store i16 %[[V2]], ptr @__llvm_profile_sampling, align 2
+; SAMPLING:       phi {{.*}}
+; SAMPLING-NEXT:  %[[V3:[0-9]+]] = add i16 {{.*}}, 1
+; SAMPLING-NEXT:  store i16 %[[V3]], ptr @__llvm_profile_sampling, align 2
+; PROMO: %[[LIVEOUT3:[a-z0-9]+]] = phi {{.*}}
+; PROMO-NEXT: %[[LIVEOUT2:[a-z0-9]+]] = phi {{.*}}
+; PROMO-NEXT: %[[LIVEOUT1:[a-z0-9]+]] = phi {{.*}}
+  %tmp10 = add nsw i32 %i.0, 1
+  %tmp11 = icmp slt i32 %tmp10, %N
+  br i1 %tmp11, label %bb2, label %bb12
+
+bb12:
+  ret void
+; PROMO: %[[CHECK1:[a-z0-9.]+]] = load {{.*}} @__profc_foo{{.*}}
+; PROMO-NEXT: add {{.*}} %[[CHECK1]], %[[LIVEOUT1]]
+; PROMO-NEXT: store {{.*}}@__profc_foo{{.*}}
+; PROMO-NEXT: %[[CHECK2:[a-z0-9.]+]] = load {{.*}} @__profc_foo{{.*}} 1)
+; PROMO-NEXT: add {{.*}} %[[CHECK2]], %[[LIVEOUT2]]
+; PROMO-NEXT: store {{.*}}@__profc_foo{{.*}}1)
+; PROMO-NEXT: %[[CHECK3:[a-z0-9.]+]] = load {{.*}} @__profc_foo{{.*}} 2)
+; PROMO-NEXT: add {{.*}} %[[CHECK3]], %[[LIVEOUT3]]
+; PROMO-NEXT: store {{.*}}@__profc_foo{{.*}}2)
+; PROMO-NOT: @__profc_foo{{.*}})
+
+}
+
+declare void @bar(i32)
+
+; SAMPLING: !0 = !{!"branch_weights", i32 200, i32 65336}
diff --git a/llvm/test/Transforms/PGOProfile/cspgo_sample.ll b/llvm/test/Transforms/PGOProfile/cspgo_sample.ll
new file mode 100644
index 00000000000000..6683cae4e64c10
--- /dev/null
+++ b/llvm/test/Transforms/PGOProfile/cspgo_sample.ll
@@ -0,0 +1,112 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; REQUIRES: x86-registered-target
+
+; RUN: opt -module-summary %s -o %t1.bc
+; RUN: opt -module-summary %S/Inputs/cspgo_bar_sample.ll -o %t2.bc
+; RUN: llvm-lto2 run -lto-cspgo-profile-file=alloc -enable-sampled-instr -lto-cspgo-gen -save-temps -o %t %t1.bc %t2.bc \
+; RUN:   -r=%t1.bc,foo,pl \
+; RUN:   -r=%t1.bc,bar,l \
+; RUN:   -r=%t1.bc,main,plx \
+; RUN:   -r=%t1.bc,__llvm_profile_filename,plx \
+; RUN:   -r=%t1.bc,__llvm_profile_raw_version,plx \
+; RUN:   -r=%t1.bc,__llvm_profile_sampling,pl \
+; RUN:   -r=%t2.bc,bar,pl \
+; RUN:   -r=%t2.bc,odd,pl \
+; RUN:   -r=%t2.bc,even,pl \
+; RUN:   -r=%t2.bc,__llvm_profile_filename,x \
+; RUN:   -r=%t2.bc,__llvm_profile_raw_version,x \
+; RUN:   -r=%t2.bc,__llvm_profile_sampling,
+; RUN: llvm-dis %t.1.4.opt.bc -o - | FileCheck %s --check-prefix=CSGEN
+
+; CSGEN: @__llvm_profile_sampling = thread_local global i16 0, comdat
+; CSGEN: @__profc_
+; CSGEN: @__profd_
+
+source_filename = "cspgo.c"
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+$__llvm_profile_filename = comdat any
+$__llvm_profile_raw_version = comdat any
+$__llvm_profile_sampling = comdat any
+ at __llvm_profile_filename = local_unnamed_addr constant [25 x i8] c"pass2/default_%m.profraw\00", comdat
+ at __llvm_profile_raw_version = local_unnamed_addr constant i64 216172782113783812, comdat
+ at __llvm_profile_sampling = thread_local global i16 0, comdat
+ at llvm.used = appending global [1 x i8*] [i8* bitcast (i64* @__llvm_profile_sampling to i8*)], section "llvm.metadata"
+
+define dso_local void @foo() #0 !prof !30 {
+entry:
+  br label %for.body
+
+for.body:
+  %i.06 = phi i32 [ 0, %entry ], [ %add1, %for.body ]
+  tail call void @bar(i32 %i.06) #3
+  %add = or i32 %i.06, 1
+  tail call void @bar(i32 %add) #3
+  %add1 = add nuw nsw i32 %i.06, 2
+  %cmp = icmp ult i32 %add1, 200000
+  br i1 %cmp, label %for.body, label %for.end, !prof !31
+
+for.end:
+  ret void
+}
+
+; CSGEN-LABEL: @foo
+; CSGEN:        [[TMP0:%.*]]  = load i16, ptr @__llvm_profile_sampling, align 2
+; CSGEN-NEXT:   [[TMP1:%.*]] = icmp ult i16 [[TMP0]], 201
+; CSGEN-NEXT:   br i1 [[TMP1]], label %{{.*}}, label %{{.*}}, !prof [[PROF:![0-9]+]]
+; CSGEN:        [[TMP2:%.*]] = add i16 {{.*}}, 1
+; CSGEN-NEXT:   store i16 [[TMP2]], ptr @__llvm_profile_sampling, align 2
+
+declare dso_local void @bar(i32)
+
+define dso_local i32 @main() !prof !30 {
+entry:
+  tail call void @foo()
+  ret i32 0
+}
+; CSGEN-LABEL: @main
+; CSGEN:        [[TMP0:%.*]]  = load i16, ptr @__llvm_profile_sampling, align 2
+; CSGEN-NEXT:   [[TMP1:%.*]] = icmp ult i16 [[TMP0]], 201
+; CSGEN-NEXT:   br i1 [[TMP1]], label %{{.*}}, label %{{.*}}, !prof [[PROF:![0-9]+]]
+; CSGEN:        [[TMP2:%.*]] = add i16 {{.*}}, 1
+; CSGEN-NEXT:   store i16 [[TMP2]], ptr @__llvm_profile_sampling, align 2
+
+attributes #0 = { "target-cpu"="x86-64" }
+
+!llvm.module.flags = !{!0, !1, !2}
+
+!0 = !{i32 1, !"wchar_size", i32 4}
+!1 = !{i32 1, !"EnableSplitLTOUnit", i32 0}
+!2 = !{i32 1, !"ProfileSummary", !3}
+!3 = !{!4, !5, !6, !7, !8, !9, !10, !11}
+!4 = !{!"ProfileFormat", !"InstrProf"}
+!5 = !{!"TotalCount", i64 500002}
+!6 = !{!"MaxCount", i64 200000}
+!7 = !{!"MaxInternalCount", i64 100000}
+!8 = !{!"MaxFunctionCount", i64 200000}
+!9 = !{!"NumCounts", i64 6}
+!10 = !{!"NumFunctions", i64 4}
+!11 = !{!"DetailedSummary", !12}
+!12 = !{!13, !14, !15, !16, !17, !18, !19, !20, !21, !22, !23, !24, !25, !26, !27, !28}
+!13 = !{i32 10000, i64 200000, i32 1}
+!14 = !{i32 100000, i64 200000, i32 1}
+!15 = !{i32 200000, i64 200000, i32 1}
+!16 = !{i32 300000, i64 200000, i32 1}
+!17 = !{i32 400000, i64 200000, i32 1}
+!18 = !{i32 500000, i64 100000, i32 4}
+!19 = !{i32 600000, i64 100000, i32 4}
+!20 = !{i32 700000, i64 100000, i32 4}
+!21 = !{i32 800000, i64 100000, i32 4}
+!22 = !{i32 900000, i64 100000, i32 4}
+!23 = !{i32 950000, i64 100000, i32 4}
+!24 = !{i32 990000, i64 100000, i32 4}
+!25 = !{i32 999000, i64 100000, i32 4}
+!26 = !{i32 999900, i64 100000, i32 4}
+!27 = !{i32 999990, i64 100000, i32 4}
+!28 = !{i32 999999, i64 1, i32 6}
+!30 = !{!"function_entry_count", i64 1}
+!31 = !{!"branch_weights", i32 100000, i32 1}
+
+; CSGEN: [[PROF]] = !{!"branch_weights", i32 200, i32 65336}
+
diff --git a/llvm/test/Transforms/PGOProfile/instrprof_sample.ll b/llvm/test/Transforms/PGOProfile/instrprof_sample.ll
new file mode 100644
index 00000000000000..1b897b1bbc12d0
--- /dev/null
+++ b/llvm/test/Transforms/PGOProfile/instrprof_sample.ll
@@ -0,0 +1,47 @@
+; RUN: opt < %s -passes=instrprof -sampled-instr -S | FileCheck %s --check-prefixes=SAMPLE-VAR,SAMPLE-CODE,SAMPLE-DURATION
+; RRRRUN: opt < %s -passes=instrprof -sampled-instr -sampled-instr-duration=100 -S| FileCheck %s --check-prefixes=SAMPLE-VAR,SAMPLE-CODE,SAMPLE-DURATION-100
+
+target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+$__llvm_profile_raw_version = comdat any
+
+; SAMPLE-VAR: $__llvm_profile_sampling = comdat any
+
+ at __llvm_profile_raw_version = constant i64 72057594037927940, comdat
+ at __profn_f = private constant [1 x i8] c"f"
+
+; SAMPLE-VAR: @__llvm_profile_sampling = thread_local global i16 0, comdat
+; SAMPLE-VAR: @__profc_f = private global [1 x i64] zeroinitializer, section "__llvm_prf_cnts", comdat, align 8
+; SAMPLE-VAR: @__profd_f = private global { i64, i64, i64, ptr, ptr, i32, [2 x i16] } { i64 -3706093650706652785, i64 12884901887, i64 sub (i64 ptrtoint (ptr @__profc_f to i64), i64 ptrtoint (ptr @__profd_f to i64)), ptr @f.local, ptr null, i32 1, [2 x i16] zeroinitializer }, section "__llvm_prf_data", comdat($__profc_f), align 8
+; SAMPLE-VAR: @__llvm_prf_nm = private constant [11 x i8] c"\01\09x\DAK\03\00\00g\00g", section "__llvm_prf_names", align 1
+; SAMPLE-VAR: @llvm.compiler.used = appending global [2 x ptr] [ptr @__llvm_profile_sampling, ptr @__profd_f], section "llvm.metadata"
+; SAMPLE-VAR: @llvm.used = appending global [1 x ptr] [ptr @__llvm_prf_nm], section "llvm.metadata"
+
+
+define void @f() {
+; SAMPLE-CODE-LABEL: @f(
+; SAMPLE-CODE:  entry:
+; SAMPLE-CODE-NEXT:    [[TMP0:%.*]] = load i16, ptr @__llvm_profile_sampling, align 2
+; SAMPLE-DURATION:         [[TMP1:%.*]] = icmp ule i16 [[TMP0]], 200
+; SAMPLE-DURATION-100:     [[TMP1:%.*]] = icmp ule i16 [[TMP0]], 100
+; SAMPLE-CODE:         br i1 [[TMP1]], label %[[TMP2:.*]], label %[[TMP4:.*]], !prof !0
+; SAMPLE-CODE:       [[TMP2]]:
+; SAMPLE-CODE-NEXT:    [[PGOCOUNT:%.*]] = load i64, ptr @__profc_f
+; SAMPLE-CODE-NEXT:    [[TMP3:%.*]] = add i64 [[PGOCOUNT]], 1
+; SAMPLE-CODE-NEXT:    store i64 [[TMP3]], ptr @__profc_f
+; SAMPLE-CODE-NEXT:    br label %[[TMP4]]
+; SAMPLE-CODE:       [[TMP4]]:
+; SAMPLE-CODE-NEXT:    [[TMP5:%.*]] = add i16 [[TMP0]], 1
+; SAMPLE-CODE-NEXT:    store i16 [[TMP5]],  ptr @__llvm_profile_sampling, align 2
+; SAMPLE-CODE-NEXT:    ret void
+;
+entry:
+  call void @llvm.instrprof.increment(i8* getelementptr inbounds ([1 x i8], [1 x i8]* @__profn_f, i32 0, i32 0), i64 12884901887, i32 1, i32 0)
+  ret void
+}
+
+; SAMPLE-DURATION: !0 = !{!"branch_weights", i32 200, i32 65336}
+; SAMPLE-DURATION-100: !0 = !{!"branch_weights", i32 100, i32 65436}
+
+declare void @llvm.instrprof.increment(i8*, i64, i32, i32)

>From 34184d6be2d5be24d809c29313b2f8585764b780 Mon Sep 17 00:00:00 2001
From: Rong Xu <xur at google.com>
Date: Tue, 23 Jan 2024 16:49:28 -0800
Subject: [PATCH 2/2] [PGO] Instrument after alloc instruction.

---
 .../lib/Transforms/Instrumentation/PGOInstrumentation.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp b/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
index 31252769904083..9de9dfc6fc9db2 100644
--- a/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
+++ b/llvm/lib/Transforms/Instrumentation/PGOInstrumentation.cpp
@@ -770,7 +770,7 @@ BasicBlock *FuncPGOInstrumentation<Edge, BBInfo>::getInstrBB(Edge *E) {
   auto canInstrument = [](BasicBlock *BB) -> BasicBlock * {
     // There are basic blocks (such as catchswitch) cannot be instrumented.
     // If the returned first insertion point is the end of BB, skip this BB.
-    if (BB->getFirstInsertionPt() == BB->end())
+    if (BB->getFirstNonPHIOrDbgOrAlloca() == BB->end())
       return nullptr;
     return BB;
   };
@@ -864,7 +864,7 @@ static void instrumentOneFunc(
                                   FuncInfo.FunctionHash);
   if (PGOFunctionEntryCoverage) {
     auto &EntryBB = F.getEntryBlock();
-    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
+    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstNonPHIOrDbgOrAlloca());
     // llvm.instrprof.cover(i8* <name>, i64 <hash>, i32 <num-counters>,
     //                      i32 <index>)
     Builder.CreateCall(
@@ -882,7 +882,7 @@ static void instrumentOneFunc(
   if (PGOTemporalInstrumentation) {
     NumCounters += PGOBlockCoverage ? 8 : 1;
     auto &EntryBB = F.getEntryBlock();
-    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstInsertionPt());
+    IRBuilder<> Builder(&EntryBB, EntryBB.getFirstNonPHIOrDbgOrAlloca());
     // llvm.instrprof.timestamp(i8* <name>, i64 <hash>, i32 <num-counters>,
     //                          i32 <index>)
     Builder.CreateCall(
@@ -892,7 +892,7 @@ static void instrumentOneFunc(
   }
 
   for (auto *InstrBB : InstrumentBBs) {
-    IRBuilder<> Builder(InstrBB, InstrBB->getFirstInsertionPt());
+    IRBuilder<> Builder(InstrBB, InstrBB->getFirstNonPHIOrDbgOrAlloca());
     assert(Builder.GetInsertPoint() != InstrBB->end() &&
            "Cannot get the Instrumentation point");
     // llvm.instrprof.increment(i8* <name>, i64 <hash>, i32 <num-counters>,



More information about the cfe-commits mailing list