This is an archive of the discontinued LLVM Phabricator instance.

Differential D106482

[OpenMP] Strip NoInline from known OpenMP runtime functions
ClosedPublic

Authored by jhuber6 on Jul 21 2021, 12:53 PM.

Details

Reviewers
jdoerfert
josemonsalve2
Commits
rG16206d17cd9e: [OpenMP] Strip NoInline from known OpenMP runtime functions
Summary

This patch strips the NoInline attribute from known OpenMP runtime functions.
This is done so that we can denote certain runtime functions as NoInline to
ensure their call sites are intact so they can be checked by OpenMPOpt. We
don't wan't this noinline attribute to remain for any functions after OpenMPOpt
has been run however.

Diff Detail

Event Timeline

jhuber6 created this revision.Jul 21 2021, 12:53 PM
Herald added subscribers: ormris, guansong, hiraditya, yaxunl. · View Herald TranscriptJul 21 2021, 12:53 PM
jhuber6 requested review of this revision.Jul 21 2021, 12:53 PM
Herald added a project: Restricted Project. · View Herald TranscriptJul 21 2021, 12:53 PM
Herald added subscribers: llvm-commits, sstefan1. · View Herald Transcript
jdoerfert accepted this revision.Jul 21 2021, 1:05 PM

LG

This revision is now accepted and ready to land.Jul 21 2021, 1:05 PM
Harbormaster completed remote builds in B115396: Diff 360565.Jul 21 2021, 3:08 PM
This revision was landed with ongoing or failed builds.Jul 21 2021, 6:18 PM
Closed by commit rG16206d17cd9e: [OpenMP] Strip NoInline from known OpenMP runtime functions (authored by jhuber6). · Explain Why
This revision was automatically updated to reflect the committed changes.
jhuber6 added a commit: rG16206d17cd9e: [OpenMP] Strip NoInline from known OpenMP runtime functions.
jhuber6 mentioned this in D106033: [OpenMP] Folding threadLimit and numThreads when single value in kernels.Jul 26 2021, 11:44 AM
JonChesterfield added a subscriber: JonChesterfield.Jul 26 2021, 11:51 AM

Revision Contents

PathSize
llvm/
lib/
Transforms/
IPO/
1 line
test/
Transforms/
OpenMP/
8 lines

Diff 360669

llvm/lib/Transforms/IPO/OpenMPOpt.cpp

Show First 20 Lines • Show All 408 Lines • ▼ Show 20 Lines
#define OMP_RTL(_Enum, _Name, _IsVarArg, _ReturnType, ...) \ #define OMP_RTL(_Enum, _Name, _IsVarArg, _ReturnType, ...) \
{ \ { \
SmallVector<Type *, 8> ArgsTypes({__VA_ARGS__}); \ SmallVector<Type *, 8> ArgsTypes({__VA_ARGS__}); \
Function *F = M.getFunction(_Name); \ Function *F = M.getFunction(_Name); \
RTLFunctions.insert(F); \ RTLFunctions.insert(F); \
if (declMatchesRTFTypes(F, OMPBuilder._ReturnType, ArgsTypes)) { \ if (declMatchesRTFTypes(F, OMPBuilder._ReturnType, ArgsTypes)) { \
RuntimeFunctionIDMap[F] = _Enum; \ RuntimeFunctionIDMap[F] = _Enum; \
F->removeFnAttr(Attribute::NoInline); \
auto &RFI = RFIs[_Enum]; \ auto &RFI = RFIs[_Enum]; \
RFI.Kind = _Enum; \ RFI.Kind = _Enum; \
RFI.Name = _Name; \ RFI.Name = _Name; \
RFI.IsVarArg = _IsVarArg; \ RFI.IsVarArg = _IsVarArg; \
RFI.ReturnType = OMPBuilder._ReturnType; \ RFI.ReturnType = OMPBuilder._ReturnType; \
RFI.ArgumentTypes = std::move(ArgsTypes); \ RFI.ArgumentTypes = std::move(ArgsTypes); \
RFI.Declaration = F; \ RFI.Declaration = F; \
unsigned NumUses = collectUses(RFI); \ unsigned NumUses = collectUses(RFI); \
▲ Show 20 Lines • Show All 3,665 Lines • Show Last 20 Lines

llvm/test/Transforms/OpenMP/add_attributes.ll

Show First 20 Lines • Show All 660 Lines • ▼ Show 20 Lines
declare i8* @__kmpc_task_reduction_get_th_data(i32, i8*, i8*) declare i8* @__kmpc_task_reduction_get_th_data(i32, i8*, i8*)
declare i8* @__kmpc_task_reduction_init(i32, i32, i8*) declare i8* @__kmpc_task_reduction_init(i32, i32, i8*)
declare i8* @__kmpc_task_reduction_modifier_init(i8*, i32, i32, i32, i8*) declare i8* @__kmpc_task_reduction_modifier_init(i8*, i32, i32, i32, i8*)
declare void @__kmpc_proxy_task_completed_ooo(i8*) declare void @__kmpc_proxy_task_completed_ooo(i8*)
; Function Attrs: noinline ; Function Attrs: noinline cold
declare void @__kmpc_barrier_simple_spmd(%struct.ident_t* nocapture nofree readonly, i32) #0 declare void @__kmpc_barrier_simple_spmd(%struct.ident_t* nocapture nofree readonly, i32) #0
attributes #0 = { noinline } attributes #0 = { noinline cold }
; CHECK: ; Function Attrs: nounwind ; CHECK: ; Function Attrs: nounwind
; CHECK-NEXT: declare dso_local void @omp_set_num_threads(i32) ; CHECK-NEXT: declare dso_local void @omp_set_num_threads(i32)
; CHECK: ; Function Attrs: nounwind ; CHECK: ; Function Attrs: nounwind
; CHECK-NEXT: declare dso_local void @omp_set_dynamic(i32) ; CHECK-NEXT: declare dso_local void @omp_set_dynamic(i32)
; CHECK: ; Function Attrs: nounwind ; CHECK: ; Function Attrs: nounwind
▲ Show 20 Lines • Show All 522 Lines • ▼ Show 20 Lines
; CHECK-NEXT: declare i8* @__kmpc_task_reduction_init(i32, i32, i8*) ; CHECK-NEXT: declare i8* @__kmpc_task_reduction_init(i32, i32, i8*)
; CHECK: ; Function Attrs: nounwind ; CHECK: ; Function Attrs: nounwind
; CHECK-NEXT: declare i8* @__kmpc_task_reduction_modifier_init(i8*, i32, i32, i32, i8*) ; CHECK-NEXT: declare i8* @__kmpc_task_reduction_modifier_init(i8*, i32, i32, i32, i8*)
; CHECK: ; Function Attrs: nounwind ; CHECK: ; Function Attrs: nounwind
; CHECK-NEXT: declare void @__kmpc_proxy_task_completed_ooo(i8*) ; CHECK-NEXT: declare void @__kmpc_proxy_task_completed_ooo(i8*)
; CHECK: ; Function Attrs: convergent noinline nounwind ; CHECK: ; Function Attrs: cold convergent nounwind
; CHECK-NEXT: declare void @__kmpc_barrier_simple_spmd(%struct.ident_t* nocapture nofree readonly, i32) ; CHECK-NEXT: declare void @__kmpc_barrier_simple_spmd(%struct.ident_t* nocapture nofree readonly, i32)
; OPTIMISTIC: ; Function Attrs: inaccessiblememonly nofree nosync nounwind willreturn writeonly ; OPTIMISTIC: ; Function Attrs: inaccessiblememonly nofree nosync nounwind willreturn writeonly
; OPTIMISTIC-NEXT: declare dso_local void @omp_set_num_threads(i32) ; OPTIMISTIC-NEXT: declare dso_local void @omp_set_num_threads(i32)
; OPTIMISTIC: ; Function Attrs: inaccessiblememonly nofree nosync nounwind willreturn writeonly ; OPTIMISTIC: ; Function Attrs: inaccessiblememonly nofree nosync nounwind willreturn writeonly
; OPTIMISTIC-NEXT: declare dso_local void @omp_set_dynamic(i32) ; OPTIMISTIC-NEXT: declare dso_local void @omp_set_dynamic(i32)
▲ Show 20 Lines • Show All 511 Lines • ▼ Show 20 Lines
; OPTIMISTIC-NEXT: declare noalias i8* @__kmpc_task_reduction_init(i32, i32, i8*) ; OPTIMISTIC-NEXT: declare noalias i8* @__kmpc_task_reduction_init(i32, i32, i8*)
; OPTIMISTIC: ; Function Attrs: nofree nosync nounwind willreturn ; OPTIMISTIC: ; Function Attrs: nofree nosync nounwind willreturn
; OPTIMISTIC-NEXT: declare noalias i8* @__kmpc_task_reduction_modifier_init(i8*, i32, i32, i32, i8*) ; OPTIMISTIC-NEXT: declare noalias i8* @__kmpc_task_reduction_modifier_init(i8*, i32, i32, i32, i8*)
; OPTIMISTIC: ; Function Attrs: nofree nosync nounwind willreturn ; OPTIMISTIC: ; Function Attrs: nofree nosync nounwind willreturn
; OPTIMISTIC-NEXT: declare void @__kmpc_proxy_task_completed_ooo(i8*) ; OPTIMISTIC-NEXT: declare void @__kmpc_proxy_task_completed_ooo(i8*)
; OPTIMISTIC: ; Function Attrs: convergent noinline nounwind ; OPTIMISTIC: ; Function Attrs: cold convergent nounwind
; OPTIMISTIC-NEXT: declare void @__kmpc_barrier_simple_spmd(%struct.ident_t* nocapture nofree readonly, i32) ; OPTIMISTIC-NEXT: declare void @__kmpc_barrier_simple_spmd(%struct.ident_t* nocapture nofree readonly, i32)
!llvm.module.flags = !{!0} !llvm.module.flags = !{!0}
!0 = !{i32 7, !"openmp", i32 50} !0 = !{i32 7, !"openmp", i32 50}