[OpenMP] Adding InaccessibleMemOnly and InaccessibleMemOrArgMemOnly for runtime calls.

[sstefan1]

Summary

Attempt to add more attributes for runtime calls.

[jdoerfert]

Thx! I'll ask more people to look at the barrier/flush stuff to make sure my suggestion wrt. inacc+argmemonly was correct.
The ones I would also love to see are the worksharing loop functions (end with _init_4, _fini etc.). Maybe we should do those and the getters/setters separate to speed things up.

llvm/include/llvm/Frontend/OpenMP/OMPKinds.def

277

I will have to think about the barrier and flush ones.

278

Where did this one go?

[sstefan1]

In D75010#1887984, @jdoerfert wrote:

Thx! I'll ask more people to look at the barrier/flush stuff to make sure my suggestion wrt. inacc+argmemonly was correct.
The ones I would also love to see are the worksharing loop functions (end with _init_4, _fini etc.). Maybe we should do those and the getters/setters separate to speed things up.

Sounds good.

llvm/include/llvm/Frontend/OpenMP/OMPKinds.def

278

287

[jdoerfert]

LGTM. We also need nounwind on all, willreturn on most, and nosync on some of them. Maybe define new __OMP_ATTRS_SET, e.g., for "synchronizing runtime calls" for "non-synchronizing runtime calls" ...

[lebedev.ri]

Tests?

[sstefan1]

In D75010#1903773, @lebedev.ri wrote:

Tests?

I didn't see any that tested this before. I guess I'll add some.

[sstefan1]

tests

[sstefan1]

fix clang test

[jdoerfert]

I always thought there was a test for this, turns out I never committed it.
Since I did now (1c9c23d60ea7656174ad3d76293c2a90dd25e24f) you need to rebase and adjust it, sorry.

llvm/test/Transforms/OpenMP/parallel_deletion.ll

3

this test was created with utils/update_test_checks.py --function-signatures before the script was smart enough to encode that into the file directly. Can you rerun with the option please.

204

Use UTC_ARGS: -disable here and -enable below to keep the check lines even when the update script is run :)

[sstefan1]

more tests

couldn't get update_test_checks.py to update function signatures, even with the flag.

[jdoerfert]

LGTM if the test all pass.

clang/test/OpenMP/barrier_codegen.cpp

#pragma omp barrier
   // CHECK: call {{.+}} [[TMAIN_CHAR:@.+]](i{{[0-9]}}
   return tmain(argc) + tmain(argv[0][0]) + a;
 }
 
 // CLANGCG:            declare i32 @__kmpc_global_thread_num(%struct.ident_t*)
 // CLANGCG-NOT:        #
 // IRBUILDER:          ; Function Attrs: nounwind
 // IRBUILDER-NEXT:     declare i32 @__kmpc_global_thread_num(%struct.ident_t*) #
-// IRBUILDER_OPT:      ; Function Attrs: nofree nosync nounwind readonly
+// IRBUILDER_OPT:      ; Function Attrs: inaccessiblememonly nofree nosync nounwind readonly
 // IRBUILDER_OPT-NEXT: declare i32 @__kmpc_global_thread_num(%struct.ident_t*) #
 
 // CHECK: define {{.+}} [[TMAIN_INT]](
 // CHECK: [[GTID:%.+]] = call i32 @__kmpc_global_thread_num([[IDENT_T]]* [[LOC]])
 // CHECK: call void @__kmpc_barrier([[IDENT_T]]* [[EXPLICIT_BARRIER_LOC]], i32 [[GTID]])
 
 // CHECK: define {{.+}} [[TMAIN_CHAR]](
 // CHECK: [[GTID:%.+]] = call i32 @__kmpc_global_thread_num([[IDENT_T]]* [[LOC]])
 // CHECK: call void @__kmpc_barrier([[IDENT_T]]* [[EXPLICIT_BARRIER_LOC]], i32 [[GTID]])
 
 #endif

llvm/include/llvm/Frontend/OpenMP/OMPKinds.def

 #define OMP_ATTRS_SET(VarName, AttrSet)
 #endif
 
 #define __OMP_ATTRS_SET(VarName, AttrSet) OMP_ATTRS_SET(VarName, AttrSet)
 
 __OMP_ATTRS_SET(GetterAttrs,
                 OptimisticAttributes
                     ? AttributeSet(EnumAttr(NoUnwind), EnumAttr(ReadOnly),
-                                   EnumAttr(NoSync), EnumAttr(NoFree))
+                                   EnumAttr(NoSync), EnumAttr(NoFree), EnumAttr(InaccessibleMemOnly))
                     : AttributeSet(EnumAttr(NoUnwind)))
 __OMP_ATTRS_SET(GetterArgWriteAttrs,
                 OptimisticAttributes
                     ? AttributeSet(EnumAttr(NoUnwind), EnumAttr(NoSync),
-                                   EnumAttr(NoFree))
+                                   EnumAttr(NoFree), EnumAttr(InaccessibleMemOrArgMemOnly))
                     : AttributeSet(EnumAttr(NoUnwind)))
 __OMP_ATTRS_SET(SetterAttrs,
                 OptimisticAttributes
                     ? AttributeSet(EnumAttr(NoUnwind), EnumAttr(WriteOnly),
-                                   EnumAttr(NoSync), EnumAttr(NoFree))
+                                   EnumAttr(NoSync), EnumAttr(NoFree), EnumAttr(InaccessibleMemOnly))
                     : AttributeSet(EnumAttr(NoUnwind)))
 
 #undef __OMP_ATTRS_SET
 #undef OMP_ATTRS_SET
 
 #ifndef OMP_RTL_ATTRS
 #define OMP_RTL_ATTRS(Enum, FnAttrSet, RetAttrSet, ArgAttrSets)
 #endif
 
 #define __OMP_RTL_ATTRS(Name, FnAttrSet, RetAttrSet, ArgAttrSets)              \
   OMP_RTL_ATTRS(OMPRTL_##Name, FnAttrSet, RetAttrSet, ArgAttrSets)
 
+__OMP_RTL_ATTRS(__kmpc_barrier,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_cancel,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_cancel_barrier,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_flush,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_global_thread_num, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(__kmpc_fork_call, AttributeSet(EnumAttr(NoUnwind)),
                 AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_omp_taskwait,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_omp_taskyield,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_push_num_threads,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_push_proc_bind,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_serialized_parallel,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_end_serialized_parallel,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
 
-__OMP_RTL_ATTRS(__kmpc_global_thread_num, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_get_thread_num, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_get_num_threads, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_get_max_threads, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_in_parallel, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_get_dynamic, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_get_cancellation, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_get_nested, GetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_get_schedule, GetterArgWriteAttrs, AttributeSet(),

 __OMP_RTL_ATTRS(omp_get_partition_place_nums, GetterAttrs, AttributeSet(), {})
 
 __OMP_RTL_ATTRS(omp_set_num_threads, SetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_set_dynamic, SetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_set_nested, SetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_set_schedule, SetterAttrs, AttributeSet(), {})
 __OMP_RTL_ATTRS(omp_set_max_active_levels, SetterAttrs, AttributeSet(), {})
 
+__OMP_RTL_ATTRS(__kmpc_master,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_end_master,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_critical,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_critical_with_hint,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+__OMP_RTL_ATTRS(__kmpc_end_critical,
+                AttributeSet(EnumAttr(InaccessibleMemOrArgMemOnly)),
+                AttributeSet(), {})
+
 #undef __OMP_RTL_ATTRS
 #undef OMP_RTL_ATTRS
 #undef AttributeSet
 #undef EnumAttr
 
 ///}
 
 /// KMP ident_t bit flags

llvm/test/Transforms/OpenMP/gtid.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature
-; RUN: opt -openmpopt -S < %s | FileCheck %s
-; RUN: opt -passes=openmpopt -S < %s | FileCheck %s
+; RUN: opt -openmpopt -openmp-ir-builder-optimistic-attributes=true -S < %s | FileCheck %s
+; RUN: opt -passes=openmpopt -openmp-ir-builder-optimistic-attributes=true -S < %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"
 
 %struct.ident_t = type { i32, i32, i32, i32, i8* }
 
 @0 = private unnamed_addr global %struct.ident_t { i32 0, i32 34, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @.str, i32 0, i32 0) }, align 8
 @.str = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
 
+; CHECK: Function Attrs: inaccessiblememonly nofree nosync nounwind readonly
+; CHECK-NEXT: declare i32 @__kmpc_global_thread_num
 declare i32 @__kmpc_global_thread_num(%struct.ident_t*)
 declare void @useI32(i32)
 
 define void @external(i1 %c) {
 ; CHECK-LABEL: define {{[^@]+}}@external
 ; CHECK-SAME: (i1 [[C:%.*]])
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[C2:%.*]] = tail call i32 @__kmpc_global_thread_num(%struct.ident_t* nonnull @0)

llvm/test/Transforms/OpenMP/parallel_deletion.ll

 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
-; RUN: opt -S -attributor -openmpopt -attributor-disable=false < %s | FileCheck %s
-; RUN: opt -S -passes='attributor,cgscc(openmpopt)' -attributor-disable=false < %s | FileCheck %s
+; RUN: opt -S -attributor -openmpopt -attributor-disable=false -openmp-ir-builder-optimistic-attributes=true < %s | FileCheck %s
+; RUN: opt -S -passes='attributor,cgscc(openmpopt)' -attributor-disable=false -openmp-ir-builder-optimistic-attributes=true < %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"
 
 %struct.ident_t = type { i32, i32, i32, i32, i8* }
 
 @.str = private unnamed_addr constant [23 x i8] c";unknown;unknown;0;0;;\00", align 1
 @0 = private unnamed_addr global %struct.ident_t { i32 0, i32 2, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @.str, i32 0, i32 0) }, align 8
 @1 = private unnamed_addr global %struct.ident_t { i32 0, i32 322, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @.str, i32 0, i32 0) }, align 8
 @.gomp_critical_user_.reduction.var = common global [8 x i32] zeroinitializer
 @2 = private unnamed_addr global %struct.ident_t { i32 0, i32 18, i32 0, i32 0, i8* getelementptr inbounds ([23 x i8], [23 x i8]* @.str, i32 0, i32 0) }, align 8
 
 ;    void delete_parallel_0(void) {
 ;    #pragma omp parallel
 ;      { unknown_willreturn(); }
 ;    #pragma omp parallel
 ;      { readonly_willreturn(); }
 ;    #pragma omp parallel
 ;      { readnone_willreturn(); }
 ;    #pragma omp parallel
 ;      {}
 ;    }
 ;
 ; We delete all but the first of the parallel regions in this test.
 define void @delete_parallel_0() {
-; CHECK-LABEL: define {{[^@]+}}@delete_parallel_0()
+; CHECK-LABEL: @delete_parallel_0(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 0, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*)* @.omp_outlined.willreturn to void (i32*, i32*, ...)*))
 ; CHECK-NEXT:    ret void
 ;
 entry:
   call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull @0, i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined.willreturn to void (i32*, i32*, ...)*))
   call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull @0, i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined.willreturn.0 to void (i32*, i32*, ...)*))
   call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull @0, i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined.willreturn.1 to void (i32*, i32*, ...)*))

 ;    #pragma omp parallel
 ;      { readnone(); }
 ;    #pragma omp parallel
 ;      {}
 ;    }
 ;
 ; We delete only the last parallel regions in this test because the others might not return.
 define void @delete_parallel_1() {
-; CHECK-LABEL: define {{[^@]+}}@delete_parallel_1()
+; CHECK-LABEL: @delete_parallel_1(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 0, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))
 ; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 0, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*)* @.omp_outlined..0 to void (i32*, i32*, ...)*))
 ; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 0, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*)* @.omp_outlined..1 to void (i32*, i32*, ...)*))
 ; CHECK-NEXT:    ret void
 ;
 entry:
   call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull @0, i32 0, void (i32*, i32*, ...)* bitcast (void (i32*, i32*)* @.omp_outlined. to void (i32*, i32*, ...)*))

 ;      {
 ;        ++a;
 ;      }
 ;    }
 ;
 ; FIXME: We do not realize that `a` is dead and all accesses to it can be removed
 ;        making the parallel regions readonly and deletable.
 define void @delete_parallel_2() {
-; CHECK-LABEL: define {{[^@]+}}@delete_parallel_2()
+; CHECK-LABEL: @delete_parallel_2(
 ; CHECK-NEXT:  entry:
 ; CHECK-NEXT:    [[A:%.*]] = alloca i32, align 4
 ; CHECK-NEXT:    [[TMP:%.*]] = bitcast i32* [[A]] to i8*
-; CHECK-NEXT:    call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull align 4 dereferenceable(4) [[TMP]])
+; CHECK-NEXT:    call void @llvm.lifetime.start.p0i8(i64 4, i8* nonnull align 4 dereferenceable(4) [[TMP]]) #0
 ; CHECK-NEXT:    store i32 0, i32* [[A]], align 4
-; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 1, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*, i32*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32* nocapture nofree nonnull align 4 dereferenceable(4) [[A]])
+; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 1, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*, i32*)* @.omp_outlined..3 to void (i32*, i32*, ...)*), i32* nocapture nonnull align 4 dereferenceable(4) [[A]])
 ; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 1, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*, i32*)* @.omp_outlined..4 to void (i32*, i32*, ...)*), i32* nocapture nonnull align 4 dereferenceable(4) [[A]])
 ; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 1, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*, i32*)* @.omp_outlined..5 to void (i32*, i32*, ...)*), i32* nocapture nonnull align 4 dereferenceable(4) [[A]])
 ; CHECK-NEXT:    call void (%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...) @__kmpc_fork_call(%struct.ident_t* nonnull align 8 dereferenceable(24) @0, i32 1, void (i32*, i32*, ...)* nonnull bitcast (void (i32*, i32*, i32*)* @.omp_outlined..6 to void (i32*, i32*, ...)*), i32* nocapture nonnull align 4 dereferenceable(4) [[A]])
 ; CHECK-NEXT:    [[TMP1:%.*]] = bitcast i32* [[A]] to i8*
 ; CHECK-NEXT:    call void @llvm.lifetime.end.p0i8(i64 4, i8* nonnull [[TMP1]])
 ; CHECK-NEXT:    ret void
 ;
 entry:
omp_if.then:                                      ; preds = %entry
   call void @__kmpc_end_master(%struct.ident_t* nonnull @0, i32 %tmp)
   br label %omp_if.end
 
 omp_if.end:                                       ; preds = %entry, %omp_if.then
   ret void
 }
 
 declare void @llvm.lifetime.start.p0i8(i64 immarg, i8* nocapture)
 
-declare i32 @omp_get_thread_num() inaccessiblememonly nofree nosync nounwind readonly
+; CHECK: Function Attrs: inaccessiblememonly nofree nosync nounwind readonly
+; CHECK-NEXT: declare i32 @omp_get_thread_num
+declare i32 @omp_get_thread_num()
 
+; CHECK: Function Attrs: inaccessiblemem_or_argmemonly
+; CHECK-NEXT: declare i32 @__kmpc_master
 declare i32 @__kmpc_master(%struct.ident_t*, i32)
 
+; CHECK: Function Attrs: inaccessiblemem_or_argmemonly
+; CHECK-NEXT: declare void @__kmpc_end_master
 declare void @__kmpc_end_master(%struct.ident_t*, i32)
 
 define internal void @.omp_outlined..5(i32* noalias %.global_tid., i32* noalias %.bound_tid., i32* dereferenceable(4) %a) {
 entry:
   %omp_global_thread_num = call i32 @__kmpc_global_thread_num(%struct.ident_t* nonnull @0)
   %tmp = load i32, i32* %.global_tid., align 4
   %tmp1 = call i32 @__kmpc_single(%struct.ident_t* nonnull @0, i32 %tmp)
   %tmp2 = icmp eq i32 %tmp1, 0
entry:
   store i32 %add, i32* %tmp4, align 4
   ret void
 }
 
 declare i32 @__kmpc_single(%struct.ident_t*, i32)
 
 declare void @__kmpc_end_single(%struct.ident_t*, i32)
 
+; CHECK: Function Attrs: inaccessiblemem_or_argmemonly
+; CHECK-NEXT: declare void @__kmpc_barrier
 declare void @__kmpc_barrier(%struct.ident_t*, i32)
 
-declare i32 @__kmpc_global_thread_num(%struct.ident_t*) nofree nosync nounwind readonly
+; CHECK: Function Attrs: inaccessiblememonly nofree nosync nounwind readonly
+; CHECK-NEXT: declare i32 @__kmpc_global_thread_num
+declare i32 @__kmpc_global_thread_num(%struct.ident_t*)
 
 declare i32 @__kmpc_reduce_nowait(%struct.ident_t*, i32, i32, i64, i8*, void (i8*, i8*)*, [8 x i32]*)
 
 declare void @__kmpc_end_reduce_nowait(%struct.ident_t*, i32, [8 x i32]*)
 
 declare void @llvm.lifetime.end.p0i8(i64 immarg, i8* nocapture)
 
 declare !callback !2 void @__kmpc_fork_call(%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...)