[llvm] [AMDGPU] Always emit SI_KILL_I1_PSEUDO for uniform floating point branches. (PR #124028)

[Jay Foad via llvm-commits]

jayfoad wrote:

Patch looks OK but can you come up with a test case? Maybe use `llvm-reduce` to reduce a pipeline that shows the problem?

As a follow up, we should probably rename SI_KILL_I1_PSEUDO and SI_KILL_F32_COND_IMM_PSEUDO, because the important different is not whether the comparison is integer or floating point. It is whether the comparison instruction is SALU or VALU (i.e. whether the result is in scc or vcc).

https://github.com/llvm/llvm-project/pull/124028