[llvm] [AMDGPU] InstCombine: fold invalid calls to amdgcn intrinsics into poison values (PR #191904)

[Yoonseo Choi via llvm-commits]

yoonseoch wrote:

Hi @arsenm, @shiltian, 
This PR was reverted as it failed upstream HIP buildbot. What I found in one of the failed test is that  "amdgpu-no-workgroup-id-z", "amdgpu-no-workitem-id-x" are in the attribute but the kernel calls @llvm.amdgcn.workitem.id.x() and @llvm.amdgcn.workgroup.id.x().  (e.g.  saxpy.hip-hip-amdgcn-amd-amdhsa.hipfb.amdgcn.gfx90a.img.0.5.precodegen.bc) Isn't it against what is described in https://llvm.org/docs/AMDGPUUsage.html? 

I didn't look further into all failed tests listed at the bottom, but I guess there are similar issues. 
```
Failed Tests (7):
  test-suite :: External/HIP/InOneWeekend-hip-7.0.2.test
  test-suite :: External/HIP/TheNextWeek-hip-7.0.2.test
  test-suite :: External/HIP/builtin-logb-scalbn-hip-7.0.2.test
  test-suite :: External/HIP/memmove-hip-7.0.2.test
  test-suite :: External/HIP/memset-hip-7.0.2.test
  test-suite :: External/HIP/saxpy-hip-7.0.2.test
  test-suite :: External/HIP/split-kernel-args-hip-7.0.2.test
```

https://github.com/llvm/llvm-project/pull/191904