forked from OSchip/llvm-project
[AFDO] Merge function attributes after inlining
This patch teaches the sample profile loader to merge function attributes after inlining functions. Without this patch, the compiler could inline a function requiring the 512-bit vector width into its caller without merging function attributes, triggering a failure during instruction selection. Differential Revision: https://reviews.llvm.org/D105729
This commit is contained in:
parent
5f306feb4d
commit
49d66d9f9f
|
@ -1176,6 +1176,10 @@ bool SampleProfileLoader::tryInlineCandidate(
|
|||
InlineFunctionInfo IFI(nullptr, GetAC);
|
||||
IFI.UpdateProfile = false;
|
||||
if (InlineFunction(CB, IFI).isSuccess()) {
|
||||
// Merge the attributes based on the inlining.
|
||||
AttributeFuncs::mergeAttributesForInlining(*BB->getParent(),
|
||||
*CalledFunction);
|
||||
|
||||
// The call to InlineFunction erases I, so we can't pass it here.
|
||||
emitInlinedInto(*ORE, DLoc, BB, *CalledFunction, *BB->getParent(), Cost,
|
||||
true, CSINLINE_DEBUG);
|
||||
|
|
|
@ -0,0 +1,2 @@
|
|||
xxx:82695217:2232
|
||||
67: yyy:5566623
|
|
@ -0,0 +1,38 @@
|
|||
; RUN: opt < %s -passes=sample-profile -sample-profile-file=%S/Inputs/merge-function-attributes.afdo -S | FileCheck %s
|
||||
|
||||
target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
|
||||
target triple = "x86_64-grtev4-linux-gnu"
|
||||
|
||||
; Verify that yyy is inlined into xxx with the function attibutes properly merged.
|
||||
; CHECK: define <8 x double> @xxx(){{.*}} #[[ATTRNO:[0-9]+]]
|
||||
; CHECK-NEXT: call <8 x double> @llvm.x86.avx512.mask.rndscale.pd.512
|
||||
define <8 x double> @xxx() #0 !dbg !5 {
|
||||
%x = call <8 x double> @yyy(), !dbg !7
|
||||
ret <8 x double> %x
|
||||
}
|
||||
|
||||
define available_externally <8 x double> @yyy() #1 !dbg !8 {
|
||||
%y = call <8 x double> @llvm.x86.avx512.mask.rndscale.pd.512(<8 x double> zeroinitializer, i32 9, <8 x double> zeroinitializer, i8 -1, i32 4)
|
||||
ret <8 x double> %y
|
||||
}
|
||||
|
||||
; Function Attrs: nounwind readnone
|
||||
declare <8 x double> @llvm.x86.avx512.mask.rndscale.pd.512(<8 x double>, i32 immarg, <8 x double>, i8, i32 immarg) #2
|
||||
|
||||
; CHECK: attributes #[[ATTRNO]] = { "min-legal-vector-width"="512"
|
||||
attributes #0 = { "min-legal-vector-width"="128" "prefer-vector-width"="128" "target-features"="+avx512vl" "use-sample-profile" }
|
||||
attributes #1 = { "min-legal-vector-width"="512" "use-sample-profile" }
|
||||
attributes #2 = { nounwind readnone }
|
||||
|
||||
!llvm.dbg.cu = !{!0, !3}
|
||||
!llvm.module.flags = !{!4}
|
||||
|
||||
!0 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !1, producer: "clang", isOptimized: true, runtimeVersion: 0, emissionKind: NoDebug, enums: !2, splitDebugInlining: false, debugInfoForProfiling: true, nameTableKind: None)
|
||||
!1 = !DIFile(filename: "test.cc", directory: "/proc/self/cwd")
|
||||
!2 = !{}
|
||||
!3 = distinct !DICompileUnit(language: DW_LANG_C_plus_plus_14, file: !1, producer: "clang", isOptimized: true, runtimeVersion: 0, emissionKind: NoDebug, splitDebugInlining: false, debugInfoForProfiling: true, nameTableKind: None)
|
||||
!4 = !{i32 2, !"Debug Info Version", i32 3}
|
||||
!5 = distinct !DISubprogram(name: "xxx", linkageName: "xxx", scope: !1, file: !1, line: 11, type: !6, scopeLine: 14, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !0, retainedNodes: !2)
|
||||
!6 = !DISubroutineType(types: !2)
|
||||
!7 = !DILocation(line: 78, column: 10, scope: !5)
|
||||
!8 = distinct !DISubprogram(name: "yyy", linkageName: "yyy", scope: !1, file: !1, line: 270, type: !6, scopeLine: 273, flags: DIFlagPrototyped, spFlags: DISPFlagDefinition | DISPFlagOptimized, unit: !3, retainedNodes: !2)
|
Loading…
Reference in New Issue