[StaticDataLayout] Reconcile data hotness based on PGO counters and data access profiles (#163325)

This PR enhances the `StaticDataProfileInfo::getConstantSectionPrefix`
pass to reconcile data hotness information from both PGO counters and
data access profiles. When both profiles are available for a global
variable, the pass will now use the "hotter" of the two to determine the
variable's section placement.

This is a follow-up patch of
#162388

Author: mingmingl-llvm

llvm/include/llvm/Analysis/StaticDataProfileInfo.h

@@ -60,11 +60,18 @@ class StaticDataProfileInfo {
   LLVM_ABI StaticDataHotness getConstantHotnessUsingProfileCount(
       const Constant *C, const ProfileSummaryInfo *PSI, uint64_t Count) const;
 
+  /// Return the hotness based on section prefix \p SectionPrefix.
+  LLVM_ABI StaticDataHotness getSectionHotnessUsingDataAccessProfile(
+      std::optional<StringRef> SectionPrefix) const;
+
   /// Return the string representation of the hotness enum \p Hotness.
   LLVM_ABI StringRef hotnessToStr(StaticDataHotness Hotness) const;
 
+  bool EnableDataAccessProf = false;
+
 public:
-  StaticDataProfileInfo() = default;
+  StaticDataProfileInfo(bool EnableDataAccessProf)
+      : EnableDataAccessProf(EnableDataAccessProf) {}
 
   /// If \p Count is not nullopt, add it to the profile count of the constant \p
   /// C in a saturating way, and clamp the count to \p getInstrMaxCountValue if
@@ -73,14 +80,10 @@ class StaticDataProfileInfo {
   LLVM_ABI void addConstantProfileCount(const Constant *C,
                                         std::optional<uint64_t> Count);
 
-  /// Return a section prefix for the constant \p C based on its profile count.
-  /// - If a constant doesn't have a counter, return an empty string.
-  /// - Otherwise,
-  ///   - If it has a hot count, return "hot".
-  ///   - If it is seen by unprofiled function, return an empty string.
-  ///   - If it has a cold count, return "unlikely".
-  ///   - Otherwise (e.g. it's used by lukewarm functions), return an empty
-  ///     string.
+  /// Given a constant \p C, returns a section prefix.
+  /// If \p C is a global variable, the section prefix is the bigger one
+  /// between its existing section prefix and its use profile count. Otherwise,
+  /// the section prefix is based on its use profile count.
   LLVM_ABI StringRef getConstantSectionPrefix(
       const Constant *C, const ProfileSummaryInfo *PSI) const;
 };

llvm/lib/Analysis/StaticDataProfileInfo.cpp

@@ -1,10 +1,14 @@
 #include "llvm/Analysis/StaticDataProfileInfo.h"
 #include "llvm/Analysis/ProfileSummaryInfo.h"
 #include "llvm/IR/Constant.h"
+#include "llvm/IR/Constants.h"
 #include "llvm/IR/GlobalVariable.h"
+#include "llvm/IR/Module.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/ProfileData/InstrProf.h"
 
+#define DEBUG_TYPE "static-data-profile-info"
+
 using namespace llvm;
 
 namespace llvm {
@@ -79,6 +83,17 @@ StaticDataProfileInfo::getConstantHotnessUsingProfileCount(
   return StaticDataHotness::LukewarmOrUnknown;
 }
 
+StaticDataProfileInfo::StaticDataHotness
+StaticDataProfileInfo::getSectionHotnessUsingDataAccessProfile(
+    std::optional<StringRef> MaybeSectionPrefix) const {
+  if (!MaybeSectionPrefix)
+    return StaticDataHotness::LukewarmOrUnknown;
+  StringRef Prefix = *MaybeSectionPrefix;
+  assert((Prefix == "hot" || Prefix == "unlikely") &&
+         "Expect section_prefix to be one of hot or unlikely");
+  return Prefix == "hot" ? StaticDataHotness::Hot : StaticDataHotness::Cold;
+}
+
 StringRef StaticDataProfileInfo::hotnessToStr(StaticDataHotness Hotness) const {
   switch (Hotness) {
   case StaticDataHotness::Cold:
@@ -101,13 +116,66 @@ StaticDataProfileInfo::getConstantProfileCount(const Constant *C) const {
 StringRef StaticDataProfileInfo::getConstantSectionPrefix(
     const Constant *C, const ProfileSummaryInfo *PSI) const {
   std::optional<uint64_t> Count = getConstantProfileCount(C);
+
+#ifndef NDEBUG
+  auto DbgPrintPrefix = [](StringRef Prefix) {
+    return Prefix.empty() ? "<empty>" : Prefix;
+  };
+#endif
+
+  if (EnableDataAccessProf) {
+    // Module flag `HasDataAccessProf` is 1 -> empty section prefix means
+    // unknown hotness except for string literals.
+    if (const GlobalVariable *GV = dyn_cast<GlobalVariable>(C);
+        GV && llvm::memprof::IsAnnotationOK(*GV) &&
+        !GV->getName().starts_with(".str")) {
+      auto HotnessFromDataAccessProf =
+          getSectionHotnessUsingDataAccessProfile(GV->getSectionPrefix());
+
+      if (!Count) {
+        StringRef Prefix = hotnessToStr(HotnessFromDataAccessProf);
+        LLVM_DEBUG(dbgs() << GV->getName() << " has section prefix "
+                          << DbgPrintPrefix(Prefix)
+                          << ", solely from data access profiles\n");
+        return Prefix;
+      }
+
+      // Both data access profiles and PGO counters are available. Use the
+      // hotter one.
+      auto HotnessFromPGO = getConstantHotnessUsingProfileCount(C, PSI, *Count);
+      StaticDataHotness GlobalVarHotness = StaticDataHotness::LukewarmOrUnknown;
+      if (HotnessFromDataAccessProf == StaticDataHotness::Hot ||
+          HotnessFromPGO == StaticDataHotness::Hot) {
+        GlobalVarHotness = StaticDataHotness::Hot;
+      } else if (HotnessFromDataAccessProf ==
+                     StaticDataHotness::LukewarmOrUnknown ||
+                 HotnessFromPGO == StaticDataHotness::LukewarmOrUnknown) {
+        GlobalVarHotness = StaticDataHotness::LukewarmOrUnknown;
+      } else {
+        GlobalVarHotness = StaticDataHotness::Cold;
+      }
+      StringRef Prefix = hotnessToStr(GlobalVarHotness);
+      LLVM_DEBUG(
+          dbgs() << GV->getName() << " has section prefix "
+                 << DbgPrintPrefix(Prefix)
+                 << ", the max from data access profiles as "
+                 << DbgPrintPrefix(hotnessToStr(HotnessFromDataAccessProf))
+                 << " and PGO counters as "
+                 << DbgPrintPrefix(hotnessToStr(HotnessFromPGO)) << "\n");
+      return Prefix;
+    }
+  }
   if (!Count)
     return "";
   return hotnessToStr(getConstantHotnessUsingProfileCount(C, PSI, *Count));
 }
 
 bool StaticDataProfileInfoWrapperPass::doInitialization(Module &M) {
-  Info.reset(new StaticDataProfileInfo());
+  bool EnableDataAccessProf = false;
+  if (auto *MD = mdconst::extract_or_null<ConstantInt>(
+          M.getModuleFlag("EnableDataAccessProf")))
+    EnableDataAccessProf = MD->getZExtValue();
+  Info.reset(new StaticDataProfileInfo(EnableDataAccessProf));
   return false;
 }
 

llvm/test/CodeGen/X86/global-variable-partition-with-dap.ll

@@ -1,17 +1,101 @@
 target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-unknown-linux-gnu"
 
-;; A minimal test case. Subsequent PRs will expand on this test case
-;; (e.g., with more functions, variables and profiles) and test the hotness
-;; reconcillation implementation.
+;; Requires asserts for -debug-only.
+; REQUIRES: asserts
+
+; RUN: rm -rf %t && split-file %s %t && cd %t
+
+; RUN: llc -mtriple=x86_64-unknown-linux-gnu -relocation-model=pic \
+; RUN:     -partition-static-data-sections=true \
+; RUN:     -debug-only=static-data-profile-info \
+; RUN:     -data-sections=true  -unique-section-names=false \
+; RUN:     input-with-data-access-prof-on.ll -o - 2>&1 | FileCheck %s --check-prefixes=LOG,IR
+
 ; RUN: llc -mtriple=x86_64-unknown-linux-gnu -relocation-model=pic \
 ; RUN:     -partition-static-data-sections=true \
+; RUN:     -debug-only=static-data-profile-info \
 ; RUN:     -data-sections=true  -unique-section-names=false \
-; RUN:     %s -o - 2>&1 | FileCheck %s --check-prefix=IR
+; RUN:     input-with-data-access-prof-off.ll -o - 2>&1 | FileCheck %s --check-prefixes=OFF
+
+; LOG: hot_bss has section prefix hot, the max from data access profiles as hot and PGO counters as hot
+; LOG: data_unknown_hotness has section prefix <empty>, the max from data access profiles as <empty> and PGO counters as unlikely
+; LOG: external_relro_array has section prefix unlikely, solely from data access profiles
+
+; IR:          .type   hot_bss,@object
+; IR-NEXT:     .section .bss.hot.,"aw"
+; IR:          .type   data_unknown_hotness,@object
+; IR-NEXT:    .section .data,"aw"
+; IR:          .type   external_relro_array,@object
+; IR-NEXT:     .section        .data.rel.ro.unlikely.,"aw"
+
+
+; OFF:        .type   hot_bss,@object
+; OFF-NEXT:   .section        .bss.hot.,"aw"
+; OFF:        .type   data_unknown_hotness,@object
+; OFF-NEXT:   .section        .data.unlikely.,"aw"
+;; Global variable section prefix metadata is not used when
+;; module flag `EnableDataAccessProf` is 0, and @external_relro_array has
+;; external linkage, so analysis based on PGO counters doesn't apply. 
+; OFF:        .type   external_relro_array,@object    # @external_relro_array
+; OFF-NEXT:   .section        .data.rel.ro,"aw"
+
+;--- input-with-data-access-prof-on.ll
+; Internal vars
+@hot_bss = internal global i32 0, !section_prefix !17
+@data_unknown_hotness = internal global i32 1
+; External vars
+@external_relro_array = constant [2 x ptr] [ptr @hot_bss, ptr @data_unknown_hotness], !section_prefix !18
+
+define void @cold_func() !prof !15 {
+  %9 = load i32, ptr @data_unknown_hotness
+  %11 = call i32 (...) @func_taking_arbitrary_param(i32 %9)
+  ret void
+}
+
+define void @hot_func() !prof !14 {
+  %9 = load i32, ptr @hot_bss
+  %11 = call i32 (...) @func_taking_arbitrary_param(i32 %9)
+  ret void
+}
+
+declare i32 @func_taking_arbitrary_param(...)
 
-; IR: .section .bss.hot.,"aw"
+!llvm.module.flags = !{!0, !1}
 
+!0 = !{i32 2, !"EnableDataAccessProf", i32 1}
+!1 = !{i32 1, !"ProfileSummary", !2}
+!2 = !{!3, !4, !5, !6, !7, !8, !9, !10}
+!3 = !{!"ProfileFormat", !"InstrProf"}
+!4 = !{!"TotalCount", i64 1460183}
+!5 = !{!"MaxCount", i64 849024}
+!6 = !{!"MaxInternalCount", i64 32769}
+!7 = !{!"MaxFunctionCount", i64 849024}
+!8 = !{!"NumCounts", i64 23627}
+!9 = !{!"NumFunctions", i64 3271}
+!10 = !{!"DetailedSummary", !11}
+!11 = !{!12, !13}
+!12 = !{i32 990000, i64 166, i32 73}
+!13 = !{i32 999999, i64 3, i32 1443}
+!14 = !{!"function_entry_count", i64 100000}
+!15 = !{!"function_entry_count", i64 1}
+!16 = !{!"branch_weights", i32 1, i32 99999}
+!17 = !{!"section_prefix", !"hot"}
+!18 = !{!"section_prefix", !"unlikely"}
+
+;--- input-with-data-access-prof-off.ll
+; Same as file above except that module flag `EnableDataAccessProf` has value 0.
+; Internal vars
 @hot_bss = internal global i32 0, !section_prefix !17
+@data_unknown_hotness = internal global i32 1
+; External vars
+@external_relro_array = constant [2 x ptr] [ptr @hot_bss, ptr @data_unknown_hotness], !section_prefix !18
+
+define void @cold_func() !prof !15 {
+  %9 = load i32, ptr @data_unknown_hotness
+  %11 = call i32 (...) @func_taking_arbitrary_param(i32 %9)
+  ret void
+}
 
 define void @hot_func() !prof !14 {
   %9 = load i32, ptr @hot_bss
@@ -21,8 +105,9 @@ define void @hot_func() !prof !14 {
 
 declare i32 @func_taking_arbitrary_param(...)
 
-!llvm.module.flags = !{!1}
+!llvm.module.flags = !{!0, !1}
 
+!0 = !{i32 2, !"EnableDataAccessProf", i32 0}
 !1 = !{i32 1, !"ProfileSummary", !2}
 !2 = !{!3, !4, !5, !6, !7, !8, !9, !10}
 !3 = !{!"ProfileFormat", !"InstrProf"}
@@ -40,3 +125,4 @@ declare i32 @func_taking_arbitrary_param(...)
 !15 = !{!"function_entry_count", i64 1}
 !16 = !{!"branch_weights", i32 1, i32 99999}
 !17 = !{!"section_prefix", !"hot"}
+!18 = !{!"section_prefix", !"unlikely"}