[InstCombine] Fix test with experimental.guard + nounwind (NFC)

[nikic]

The function is marked nounwind, so experimental.guard becomes nounwind as well, and in conjunction with willreturn becomes guaranteed-to-transfer. Thus the optimization that this test was supposed to be guarding against is performed. This regressed in fbfb1c7 when experimental.guard was marked willreturn.

I believe this is an error in the test case, and it should not be using nounwind. However, I wanted to double check whether experimental.guard should be willreturn or not. (And possibly nounwind instead -- it depends on how one wants to model them.)

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Changes

The function is marked nounwind, so experimental.guard becomes nounwind as well, and in conjunction with willreturn becomes guaranteed-to-transfer. Thus the optimization that this test was supposed to be guarding against is performed. This regressed in fbfb1c7 when experimental.guard was marked willreturn.

I believe this is an error in the test case, and it should not be using nounwind. However, I wanted to double check whether experimental.guard should be willreturn or not. (And possibly nounwind instead -- it depends on how one wants to model them.)

---

Full diff: https://github.com/llvm/llvm-project/pull/67641.diff

1 Files Affected:

• (modified) llvm/test/Transforms/InstCombine/sdiv-guard.ll (+4-3)

diff --git a/llvm/test/Transforms/InstCombine/sdiv-guard.ll b/llvm/test/Transforms/InstCombine/sdiv-guard.ll
index ba9670924108b19..b0f96dafa18c0bd 100644
--- a/llvm/test/Transforms/InstCombine/sdiv-guard.ll
+++ b/llvm/test/Transforms/InstCombine/sdiv-guard.ll
@@ -4,10 +4,11 @@
 declare void @llvm.experimental.guard(i1, ...)
 
 ; Regression test. If %flag is false then %s == 0 and guard should be triggered.
-define i32 @a(i1 %flag, i32 %X) nounwind readnone {
+define i32 @a(i1 %flag, i32 %X) {
 ; CHECK-LABEL: @a(
-; CHECK-NEXT:    [[CMP:%.*]] = icmp ne i32 [[X:%.*]], 0
-; CHECK-NEXT:    call void (i1, ...) @llvm.experimental.guard(i1 [[CMP]]) #[[ATTR2:[0-9]+]] [ "deopt"() ]
+; CHECK-NEXT:    [[CMP1:%.*]] = icmp ne i32 [[X:%.*]], 0
+; CHECK-NEXT:    [[CMP:%.*]] = select i1 [[FLAG:%.*]], i1 [[CMP1]], i1 false
+; CHECK-NEXT:    call void (i1, ...) @llvm.experimental.guard(i1 [[CMP]]) [ "deopt"() ]
 ; CHECK-NEXT:    [[R:%.*]] = sdiv i32 100, [[X]]
 ; CHECK-NEXT:    ret i32 [[R]]
 ;

[nikic]

ping

[annamthomas]

This regressed in fbfb1c7 when experimental.guard was marked willreturn.

The problem is thatexperimental.guard should not be marked as willreturn. It has the same properties as experimental.deoptimize in that guards are nounwind, but are not willreturn. We missed this regression since we moved to the llvm.experimental.widenable.condition form instead of generating guards. We're waiting for couple of releases for this mode to stabilize and we should remove the guard form from upstream (as long as no one else is using this representation for some other purposes).

So, the testcase should still have nounwind and the guaranteed-to-transfer will not happen since the guard is not willreturn.

[nikic]

Okay, in that case I've put up a new patch that returns the willreturn attribute instead: #69433