[llvm] r297650 - AMDGPU: Fold icmp/fcmp into icmp intrinsic

[Matt Arsenault via llvm-commits]

> On Mar 31, 2017, at 03:03, Nicolai Hähnle via llvm-commits <llvm-commits at lists.llvm.org> wrote:
> 
> Hi Matt,
> 
> this commit causes the attached shader to miscompile. Specifically:
> 
>   opt -S -instcombine < $shader
> 
> causes the logical and in the endif8 block to disappear and one of the branch conditions to be come constant.
> 
> I haven't looked in more detail yet, just found this via a bisect, and reverting the commit fixes the issue.
> 
> Thanks,
> Nicolai

This testcase only has a use with constant inputs:   %25 = call i64 @llvm.amdgcn.icmp.i32(i32 -1, i32 0, i32 33) #2 which folds to true, so I would expect somewhere a branch condition would fold to a constant?

-Matt