[llvm-branch-commits] [llvm] [compiler-rt] [ThinLTO] (PR #79381)
Mingming Liu via llvm-branch-commits
llvm-branch-commits at lists.llvm.org
Thu Jan 25 13:59:24 PST 2024
https://github.com/minglotus-6 updated https://github.com/llvm/llvm-project/pull/79381
>From d4caa0997799b712edb11d90c5be79d0aab3c312 Mon Sep 17 00:00:00 2001
From: mingmingl <mingmingl at google.com>
Date: Thu, 25 Jan 2024 13:59:03 -0800
Subject: [PATCH] Introduce an opton to control the number of vtables to
annotate and use it when generating function summaries.
Created using spr 1.3.4
---
.../IndirectCallPromotionAnalysis.cpp | 3 +++
llvm/lib/Analysis/ModuleSummaryAnalysis.cpp | 12 ++++-----
.../thinlto-func-summary-vtableref-pgo.ll | 25 ++++++++++++-------
3 files changed, 24 insertions(+), 16 deletions(-)
diff --git a/llvm/lib/Analysis/IndirectCallPromotionAnalysis.cpp b/llvm/lib/Analysis/IndirectCallPromotionAnalysis.cpp
index ebfa1c8fc08e1c6..18cb6a220e3bd00 100644
--- a/llvm/lib/Analysis/IndirectCallPromotionAnalysis.cpp
+++ b/llvm/lib/Analysis/IndirectCallPromotionAnalysis.cpp
@@ -45,6 +45,9 @@ static cl::opt<unsigned>
cl::desc("Max number of promotions for a single indirect "
"call callsite"));
+cl::opt<unsigned> MaxNumVTableAnnotations("icp-max-num-vtables", cl::init(6), cl::Hidden,
+ cl::desc("Max number of vtables annotated for a vtable load instruction."));
+
ICallPromotionAnalysis::ICallPromotionAnalysis() {
ValueDataArray = std::make_unique<InstrProfValueData[]>(MaxNumPromotions);
}
diff --git a/llvm/lib/Analysis/ModuleSummaryAnalysis.cpp b/llvm/lib/Analysis/ModuleSummaryAnalysis.cpp
index fc8c31de0f4501f..0f0085025cc56b1 100644
--- a/llvm/lib/Analysis/ModuleSummaryAnalysis.cpp
+++ b/llvm/lib/Analysis/ModuleSummaryAnalysis.cpp
@@ -82,6 +82,8 @@ static cl::opt<std::string> ModuleSummaryDotFile(
extern cl::opt<bool> ScalePartialSampleProfileWorkingSetSize;
+extern cl::opt<unsigned> MaxNumVTableAnnotations;
+
// Walk through the operands of a given User via worklist iteration and populate
// the set of GlobalValue references encountered. Invoked either on an
// Instruction or a GlobalVariable (which walks its initializer).
@@ -129,14 +131,10 @@ static bool findRefEdges(ModuleSummaryIndex &Index, const User *CurUser,
if (I) {
uint32_t ActualNumValueData = 0;
uint64_t TotalCount = 0;
- // 24 is the maximum number of values preserved for one instrumented site,
- // defined by INSTR_PROF_DEFAULT_NUM_VAL_PER_SITE in
- // compiler-rt/lib/profile/InstrProfilingValue.c; passing 24 as
- // `MaxNumValueData` controls the max number of elements in the returned
- // array. The actual number of values is gated by the number of ops in !prof
- // metadata.
+ // MaxNumVTableAnnotations is the maximum number of vtables annotated on
+ // the instruction.
auto ValueDataArray = getValueProfDataFromInst(
- *I, IPVK_VTableTarget, 24 /* MaxNumValueData */, ActualNumValueData,
+ *I, IPVK_VTableTarget, MaxNumVTableAnnotations /* MaxNumValueData */, ActualNumValueData,
TotalCount);
if (ValueDataArray.get()) {
diff --git a/llvm/test/Bitcode/thinlto-func-summary-vtableref-pgo.ll b/llvm/test/Bitcode/thinlto-func-summary-vtableref-pgo.ll
index 28e4b1d19aef72c..ba3ce9a75ee8321 100644
--- a/llvm/test/Bitcode/thinlto-func-summary-vtableref-pgo.ll
+++ b/llvm/test/Bitcode/thinlto-func-summary-vtableref-pgo.ll
@@ -1,4 +1,8 @@
-; RUN: opt -module-summary %s -o %t.o
+; Promote at most one function and annotate at most one vtable.
+; As a result, only one value (of each relevant kind) shows up in the function
+; summary.
+
+; RUN: opt -module-summary -icp-max-num-vtables=1 -icp-max-prom=1 %s -o %t.o
; RUN: llvm-bcanalyzer -dump %t.o | FileCheck %s
@@ -11,15 +15,17 @@
; CHECK-NEXT: <FLAGS op0=0/>
; The `VALUE_GUID` below represents the "_ZTV4Base" referenced by the instruction
; that loads vtable pointers.
-; CHECK-NEXT: <VALUE_GUID op0=18 op1=1960855528937986108/>
+; CHECK-NEXT: <VALUE_GUID op0=21 op1=1960855528937986108/>
; The `VALUE_GUID` below represents the "_ZN4Base4funcEv" referenced by the
; indirect call instruction.
-; CHECK-NEXT: <VALUE_GUID op0=17 op1=5459407273543877811/>
+; CHECK-NEXT: <VALUE_GUID op0=20 op1=5459407273543877811/>
+; NOTE vtables and functions from Derived class is dropped because
+; `-icp-max-num-vtables` and `-icp-max-prom` are both set to one.
; <PERMODULE_PROFILE> has the format [valueid, flags, instcount, funcflags,
; numrefs, rorefcnt, worefcnt,
; m x valueid,
; n x (valueid, hotness+tailcall)]
-; CHECK-NEXT: <PERMODULE_PROFILE abbrevid=4 op0=0 op1=0 op2=4 op3=256 op4=1 op5=1 op6=0 op7=18 op8=17 op9=3/>
+; CHECK-NEXT: <PERMODULE_PROFILE abbrevid=4 op0=0 op1=0 op2=4 op3=256 op4=1 op5=1 op6=0 op7=21 op8=20 op9=3/>
; CHECK-NEXT: </GLOBALVAL_SUMMARY_BLOCK>
target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
@@ -36,7 +42,6 @@ define i32 @_Z4testP4Base(ptr %0) !prof !15 {
!llvm.module.flags = !{!1}
-
!1 = !{i32 1, !"ProfileSummary", !2}
!2 = !{!3, !4, !5, !6, !7, !8, !9, !10}
!3 = !{!"ProfileFormat", !"InstrProf"}
@@ -53,10 +58,12 @@ define i32 @_Z4testP4Base(ptr %0) !prof !15 {
!14 = !{i32 999999, i64 1, i32 2}
!15 = !{!"function_entry_count", i32 150}
-; 1960855528937986108 is the MD5 hash of _ZTV4Base
-!16 = !{!"VP", i32 2, i64 1600, i64 1960855528937986108, i64 1600}
-; 5459407273543877811 is the MD5 hash of _ZN4Base4funcEv
-!17 = !{!"VP", i32 0, i64 1600, i64 5459407273543877811, i64 1600}
+; 1960855528937986108 is the MD5 hash of _ZTV4Base, and
+; 13870436605473471591 is the MD5 hash of _ZTV7Derived
+!16 = !{!"VP", i32 2, i64 150, i64 1960855528937986108, i64 100, i64 13870436605473471591, i64 50}
+; 5459407273543877811 is the MD5 hash of _ZN4Base4funcEv, and
+; 6174874150489409711 is the MD5 hash of _ZN7Derived4funcEv
+!17 = !{!"VP", i32 0, i64 150, i64 5459407273543877811, i64 100, i64 6174874150489409711, i64 50}
; ModuleSummaryIndex stores <guid, global-value summary> map in std::map; so
; global value summares are printed out in the order that gv's guid increases.
More information about the llvm-branch-commits
mailing list