[CSSPGO] Use merged base profile for hot threshold calculation

[wenlei]

Summary

Context-sensitive profile effectively split a function profile into many copies each representing the CFG profile of a particular calling context. That makes the count distribution looks more flat as we now have more function profiles each with lower counts, which in turn leads to lower hot thresholds. Now we tells threshold computation to merge context profile first before calculating percentile based cutoffs to compensate for seemingly flat context profile. This can be controlled by swtich sample-profile-contextless-threshold.

Earlier measurement showed ~0.4% perf boost with this tuning on spec2k6 for CSSPGO (with pseudo-probe and new inliner).

[wmi]

For debug info based CSSPGO, I assume the threshold should be the same, correct?

llvm/lib/ProfileData/SampleProfWriter.cpp

766

ProfileMap --> *ProfilesToUse here?

Actually the computeSummary in the Reader and the Writer are mostly the same. It is better to extract them and make it a shared function somewhere, for example in SampleProfileBuilder.

[wenlei]

In D95980#2545546, @wmi wrote:

For debug info based CSSPGO, I assume the threshold should be the same, correct?

Yes, this applies to both. In fact the issue was originally uncovered by experimenting with debug info based CSSPGO.

llvm/lib/ProfileData/SampleProfWriter.cpp

766

Good catch, thanks! I moved everything into ProfileSummaryBuilder, including the switch.

[wenlei]

Address Wei's comment, rebase.

[wmi]

LGTM.

llvm/include/llvm/IR/ProfileSummary.h

 //===- ProfileSummary.h - Profile summary data structure. -------*- 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 defines the profile summary data structure.
 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_IR_PROFILESUMMARY_H
 #define LLVM_IR_PROFILESUMMARY_H
 
+#include "llvm/Support/CommandLine.h"
 #include <algorithm>
 #include <cassert>
 #include <cstdint>
 #include <vector>
 
 namespace llvm {
 
 class LLVMContext;

llvm/lib/IR/ProfileSummary.cpp

 #include "llvm/IR/Function.h"
 #include "llvm/IR/Metadata.h"
 #include "llvm/IR/Type.h"
 #include "llvm/Support/Casting.h"
 #include "llvm/Support/Format.h"
 
 using namespace llvm;
 
+cl::opt<bool> UseContextLessSummary(
+    "profile-summary-contextless", cl::Hidden, cl::init(false), cl::ZeroOrMore,
+    cl::desc("Merge context profiles before calculating thresholds."));
+
 // Return an MDTuple with two elements. The first element is a string Key and
 // the second is a uint64_t Value.
 static Metadata *getKeyValMD(LLVMContext &Context, const char *Key,
                              uint64_t Val) {
   Type *Int64Ty = Type::getInt64Ty(Context);
   Metadata *Ops[2] = {MDString::get(Context, Key),
                       ConstantAsMetadata::get(ConstantInt::get(Int64Ty, Val))};
   return MDTuple::get(Context, Ops);

llvm/lib/ProfileData/SampleProfReader.cpp

 #include <limits>
 #include <memory>
 #include <system_error>
 #include <vector>
 
 using namespace llvm;
 using namespace sampleprof;
 
+extern cl::opt<bool> UseContextLessSummary;
+
 /// Dump the function profile for \p FName.
 ///
 /// \param FName Name of the function to print.
 /// \param OS Stream to emit the output to.
 void SampleProfileReader::dumpFunctionProfile(StringRef FName,
                                               raw_ostream &OS) {
   OS << "Function: " << FName << ": " << Profiles[FName];
 }
SampleProfileReader::create(std::unique_ptr<MemoryBuffer> &B, LLVMContext &C,
   }
 
   return std::move(Reader);
 }
 
 // For text and GCC file formats, we compute the summary after reading the
 // profile. Binary format has the profile summary in its header.
 void SampleProfileReader::computeSummary() {
-  SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
+  StringMap<FunctionSamples> ContextLessProfiles;
+  const StringMap<FunctionSamples> *ProfilesToUse = &Profiles;
+  if (UseContextLessSummary ||
+      (ProfileIsCS && !UseContextLessSummary.getNumOccurrences())) {
     for (const auto &I : Profiles) {
+      ContextLessProfiles[I.second.getName()].merge(I.second);
+    }
+    ProfilesToUse = &ContextLessProfiles;
+  }
+  SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
+  for (const auto &I : *ProfilesToUse) {
     const FunctionSamples &Profile = I.second;
     Builder.addRecord(Profile);
   }
   Summary = Builder.getSummary();
 }

llvm/lib/ProfileData/SampleProfWriter.cpp

 #include <set>
 #include <system_error>
 #include <utility>
 #include <vector>
 
 using namespace llvm;
 using namespace sampleprof;
 
+extern cl::opt<bool> UseContextLessSummary;
+
 std::error_code SampleProfileWriter::writeFuncProfiles(
     const StringMap<FunctionSamples> &ProfileMap) {
   // Sort the ProfileMap by total samples.
   typedef std::pair<StringRef, const FunctionSamples *> NameFunctionSamples;
   std::vector<NameFunctionSamples> V;
   for (const auto &I : ProfileMap)
     V.push_back(std::make_pair(I.getKey(), &I.second));
 
if (EC)
     return EC;
 
   Writer->Format = Format;
   return std::move(Writer);
 }
 
 void SampleProfileWriter::computeSummary(
     const StringMap<FunctionSamples> &ProfileMap) {
+  StringMap<FunctionSamples> ContextLessProfiles;
+  const StringMap<FunctionSamples> *ProfilesToUse = &ProfileMap;
+  if (UseContextLessSummary || (FunctionSamples::ProfileIsCS &&
+                                !UseContextLessSummary.getNumOccurrences())) {
+    for (const auto &I : ProfileMap) {
+      ContextLessProfiles[I.second.getName()].merge(I.second);
+    }
+    ProfilesToUse = &ContextLessProfiles;
+  }
   SampleProfileSummaryBuilder Builder(ProfileSummaryBuilder::DefaultCutoffs);
   for (const auto &I : ProfileMap) {
     const FunctionSamples &Profile = I.second;
     Builder.addRecord(Profile);
   }
   Summary = Builder.getSummary();
 }

llvm/test/Transforms/SampleProfile/csspgo-inline.ll

 ; INLINE-BASE: remark: merged.cpp:27:11: _Z8funcLeafi inlined into main to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcAi:1:11 @ main:3:10
 ; INLINE-BASE: remark: merged.cpp:33:11: _Z8funcLeafi inlined into _Z5funcBi to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcBi:1:11
 
 ; INLINE-NEW: remark: merged.cpp:14:10: _Z5funcAi inlined into main to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite main:3:10
 ; INLINE-NEW-NOT: remark
 
 ; INLINE-NEW-LIMIT1-NOT: remark
 
-; INLINE-NEW-LIMIT2: remark: merged.cpp:27:11: _Z8funcLeafi inlined into _Z5funcAi to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcAi:1:11
 ; INLINE-NEW-LIMIT2: remark: merged.cpp:33:11: _Z8funcLeafi inlined into _Z5funcBi to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcBi:1:11
 ; INLINE-NEW-LIMIT2-NOT: remark
 
 @factor = dso_local global i32 3, align 4, !dbg !0
 
 define dso_local i32 @main() local_unnamed_addr #0 !dbg !18 {
 entry:
   br label %for.body, !dbg !25

llvm/test/Transforms/SampleProfile/csspgo-summary.ll

-; Test for CSSPGO's new early inliner using priority queue
+; Test for CSSPGO's profile summary computation with and without pre-merging context profiles
+
+; RUN: opt < %s -passes=sample-profile,print-profile-summary -sample-profile-file=%S/Inputs/profile-context-tracker.prof -profile-summary-cutoff-hot=999900 -profile-sample-accurate -profile-summary-contextless=0 -S -o /dev/null 2>&1 | FileCheck %s --check-prefix=SUMMARY-UNMERGED
+; RUN: opt < %s -passes=sample-profile,print-profile-summary -sample-profile-file=%S/Inputs/profile-context-tracker.prof -profile-summary-cutoff-hot=999900 -profile-sample-accurate -profile-summary-contextless=1 -S -o /dev/null 2>&1 | FileCheck %s --check-prefix=SUMMARY-MERGED
+
+; SUMMARY-UNMERGED: main :hot entry
+; SUMMARY-MERGED-NOT: main :hot entry
 
-; Note that we need new pass manager to enable top-down processing for sample profile loader
-; Test we inlined the following in top-down order with old inliner
-;   main:3 @ _Z5funcAi
-;   main:3 @ _Z5funcAi:1 @ _Z8funcLeafi
-;   _Z5funcBi:1 @ _Z8funcLeafi
-; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/profile-context-tracker.prof -sample-profile-inline-size -sample-profile-prioritized-inline=0 -profile-sample-accurate -S -pass-remarks=inline -o /dev/null 2>&1 | FileCheck %s --check-prefix=INLINE-BASE
-;
-; With new FDO early inliner, callee entry count is used to drive inlining instead of callee total samples, so we get less inlining for given profile
-; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/profile-context-tracker.prof -sample-profile-inline-size -profile-sample-accurate -S -pass-remarks=inline -o /dev/null 2>&1 | FileCheck %s --check-prefix=INLINE-NEW
-;
-; With new FDO early inliner, callee entry count is used to drive inlining instead of callee total samples, tuning hot cutoff can get us the same inlining
-; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/profile-context-tracker.prof -sample-profile-inline-size -profile-summary-cutoff-hot=999900 -profile-sample-accurate -S -pass-remarks=inline -o /dev/null 2>&1 | FileCheck %s --check-prefix=INLINE-BASE
-;
-; With new FDO early inliner, callee entry count is used to drive inlining instead of callee total samples, tuning cold sample profile inline threshold can get us the same inlining
-; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/profile-context-tracker.prof -sample-profile-inline-size -sample-profile-cold-inline-threshold=200 -profile-sample-accurate -S -pass-remarks=inline -o /dev/null 2>&1 | FileCheck %s --check-prefix=INLINE-BASE
-;
-; With new FDO early inliner and tuned cutoff, we can control inlining through size growth tuning knob.
-; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/profile-context-tracker.prof -sample-profile-inline-size -profile-summary-cutoff-hot=999900 -sample-profile-inline-limit-min=0 -sample-profile-inline-growth-limit=1 -profile-sample-accurate -S -pass-remarks=inline -o /dev/null 2>&1 | FileCheck %s --allow-empty --check-prefix=INLINE-NEW-LIMIT1
-; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/profile-context-tracker.prof -sample-profile-inline-size -profile-summary-cutoff-hot=999900 -sample-profile-inline-limit-min=10 -sample-profile-inline-growth-limit=1 -profile-sample-accurate -S -pass-remarks=inline -o /dev/null 2>&1 | FileCheck %s --check-prefix=INLINE-NEW-LIMIT2
-
-
-; INLINE-BASE: remark: merged.cpp:14:10: _Z5funcAi inlined into main to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite main:3:10
-; INLINE-BASE: remark: merged.cpp:27:11: _Z8funcLeafi inlined into main to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcAi:1:11 @ main:3:10
-; INLINE-BASE: remark: merged.cpp:33:11: _Z8funcLeafi inlined into _Z5funcBi to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcBi:1:11
-
-; INLINE-NEW: remark: merged.cpp:14:10: _Z5funcAi inlined into main to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite main:3:10
-; INLINE-NEW-NOT: remark
-
-; INLINE-NEW-LIMIT1-NOT: remark
-
-; INLINE-NEW-LIMIT2: remark: merged.cpp:27:11: _Z8funcLeafi inlined into _Z5funcAi to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcAi:1:11
-; INLINE-NEW-LIMIT2: remark: merged.cpp:33:11: _Z8funcLeafi inlined into _Z5funcBi to match profiling context with (cost={{[0-9]+}}, threshold={{[0-9]+}}) at callsite _Z5funcBi:1:11
-; INLINE-NEW-LIMIT2-NOT: remark
 
 @factor = dso_local global i32 3, align 4, !dbg !0
 
 define dso_local i32 @main() local_unnamed_addr #0 !dbg !18 {
 entry:
   br label %for.body, !dbg !25
 
 for.cond.cleanup:                                 ; preds = %for.body