[OpenMPOpt] initial tests for ICV tracking. Only nthreads is used.

[sstefan1]

Summary

Couple of tests to showcase what will be done and what to expect with ICV tracking.

[jdoerfert]

Thanks! One comment for the FIXME.

llvm/test/Transforms/OpenMP/icv_tracking.ll

62

I guess %5 should be replaced with min(%3, 10), right? We might want to use SCEV to represent the values and combine them.

[JonChesterfield]

Nice. Great to have tests committed before the optimisation as it means the diff attached to that optimisation makes it really obvious what the transform did.

It might be worth removing some more of the noise from the IR, e.g. the 'noalias nocapture' stuff on the arguments is unlikely to be relevant.

[sstefan1]

In D81114#2072328, @JonChesterfield wrote:

It might be worth removing some more of the noise from the IR, e.g. the 'noalias nocapture' stuff on the arguments is unlikely to be relevant.

Agreed. I already cleaned some stuff up. Will do more.

llvm/test/Transforms/OpenMP/icv_tracking.ll

62

Not sure I understand why min(%3, 10)? Without any transformations %5 is 10. Maybe I'm missing something?

Will look into using SCEV.

[jdoerfert]

llvm/test/Transforms/OpenMP/icv_tracking.ll

62

In reality, it is a value smaller or equal to 10, I think.

Let's say we have a quad core and the user sets the num_threads to 10, there are two reasonable outcomes. The runtime will cap it at 4, given that any more threads will most likely not help, or it will go with 10. Afaik, OpenMP doesn't guarantee 10 but not more than 10 if you set a limit. We might want to provide user hooks that simplify the logic, e.g., allow us to take 10, but otherwise min(10, %3) seems reasonable to me. WDYT?

[sstefan1]

llvm/test/Transforms/OpenMP/icv_tracking.ll

62

Oh, ok. Makes sense now. 10 was totally random. I'll update the comment to have min(10, %3).

[sstefan1]

changed fixme, IR cleanup

[jdoerfert]

LGTM

llvm/test/Transforms/OpenMP/icv_tracking.ll

2

Run it with old and new pass manager

llvm/test/Transforms/OpenMP/icv_tracking.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py
+; RUN: opt -S -openmpopt < %s | FileCheck %s
+
+%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
+
+define dso_local i32 @foo(i32 %0, i32 %1) {
+; CHECK-LABEL: @foo(
+; CHECK-NEXT:    tail call void @omp_set_num_threads(i32 [[TMP0:%.*]])
+; CHECK-NEXT:    [[TMP3:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @omp_set_num_threads(i32 [[TMP1:%.*]])
+; CHECK-NEXT:    [[TMP4:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    [[TMP5:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    [[TMP6:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP4]])
+; CHECK-NEXT:    tail call void @use(i32 [[TMP5]])
+; CHECK-NEXT:    tail 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*, ...)*))
+; CHECK-NEXT:    [[TMP7:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP7]])
+; CHECK-NEXT:    ret i32 0
+;
+  tail call void @omp_set_num_threads(i32 %0)
+  %3 = tail call i32 @omp_get_max_threads()
+; FIXME: this value should be tracked and the rest of the getters deduplicated and replaced with it.
+  tail call void @omp_set_num_threads(i32 %1)
+  %4 = tail call i32 @omp_get_max_threads()
+  %5 = tail call i32 @omp_get_max_threads()
+  %6 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %4)
+  tail call void @use(i32 %5)
+  tail 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*, ...)*))
+  %7 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %7)
+  ret i32 0
+}
+
+declare dso_local void @omp_set_num_threads(i32)
+
+declare dso_local i32 @omp_get_max_threads()
+
+declare dso_local void @use(i32)
+
+define internal void @.omp_outlined.(i32* %0, i32* %1) {
+; CHECK-LABEL: @.omp_outlined.(
+; CHECK-NEXT:    [[TMP3:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    [[TMP4:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP4]])
+; CHECK-NEXT:    tail call void @omp_set_num_threads(i32 10)
+; CHECK-NEXT:    [[TMP5:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP5]])
+; CHECK-NEXT:    ret void
+;
+; FIXME: this value should be tracked and the rest of the getters deduplicated and replaced with it.
+  %3 = tail call i32 @omp_get_max_threads()
+  %4 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %4)
+; FIXME: this value ( min(%3, 10) ) should be tracked and the rest of the getters deduplicated and replaced with it.
+  tail call void @omp_set_num_threads(i32 10)
+  %5 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %5)
+  ret void
+}
+
+declare !callback !0 void @__kmpc_fork_call(%struct.ident_t*, i32, void (i32*, i32*, ...)*, ...)
+
+define dso_local i32 @bar(i32 %0, i32 %1) {
+; CHECK-LABEL: @bar(
+; CHECK-NEXT:    [[TMP3:%.*]] = icmp sgt i32 [[TMP0:%.*]], [[TMP1:%.*]]
+; CHECK-NEXT:    [[TMP4:%.*]] = select i1 [[TMP3]], i32 [[TMP0]], i32 [[TMP1]]
+; CHECK-NEXT:    tail call void @omp_set_num_threads(i32 [[TMP4]])
+; CHECK-NEXT:    [[TMP5:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail 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..1 to void (i32*, i32*, ...)*))
+; CHECK-NEXT:    [[TMP6:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP6]])
+; CHECK-NEXT:    ret i32 0
+;
+  %3 = icmp sgt i32 %0, %1
+  %4 = select i1 %3, i32 %0, i32 %1
+; FIXME: getters can be replaced with %4
+  tail call void @omp_set_num_threads(i32 %4)
+  %5 = tail call i32 @omp_get_max_threads()
+  tail 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..1 to void (i32*, i32*, ...)*))
+  %6 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %6)
+  ret i32 0
+}
+
+define internal void @.omp_outlined..1(i32* %0, i32*  %1) {
+; CHECK-LABEL: @.omp_outlined..1(
+; CHECK-NEXT:    [[TMP3:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP3]])
+; CHECK-NEXT:    tail call void @omp_set_num_threads(i32 10)
+; CHECK-NEXT:    [[TMP4:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP4]])
+; CHECK-NEXT:    [[TMP5:%.*]] = tail call i32 @omp_get_max_threads()
+; CHECK-NEXT:    tail call void @use(i32 [[TMP5]])
+; CHECK-NEXT:    ret void
+;
+  %3 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %3)
+  tail call void @omp_set_num_threads(i32 10)
+  %4 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %4)
+  %5 = tail call i32 @omp_get_max_threads()
+  tail call void @use(i32 %5)
+  ret void
+}
+
+!0 = !{!1}
+!1 = !{i64 2, i64 -1, i64 -1, i1 true}