[llvm] [AMDGPU]Optimize SGPR spills (PR #93668)

[Yaxun Liu via llvm-commits]

yxsamliu wrote:

> > Hi, I believe this turned one of our buildbots red: https://lab.llvm.org/buildbot/#/builders/123/builds/253 If you require assistance in reproducing, let me know and I'll see if I can be of help.
> 
> Yeah, that would be great!

I can confirm that this PR caused regression in blender test of buildbot:

```
Fra:48 Mem:1531.14M (Peak 1531.14M) | Time:06:58.63 | Mem:2448.05M, Peak:2448.05M | Scene, View Layer | Sample 0/256
Warning: precise memory violation signal reporting is not enabled, reported
location may not be accurate.  See "show amdgpu precise-memory".

Thread 103 "blender" received signal SIGSEGV, Segmentation fault.
[Switching to thread 103, lane 9 (AMDGPU Lane 1:1:1:1/9 (0,0,0)[9,0,0])]
0x00007ffd7689a68c in integrator_intersect_shadow(KernelGlobalsGPU const*, int) () from memory://192866#offset=0x7ffd77f01240&size=4555968
```

Currently the buildbot has disabled all Blender test scenes so it is apparently green, but we'd get this fixed ASAP since this commit may get to amd-staging branch and cause other regressions.

debugger shows memfault happens at the highlighted line:

<img width="936" alt="Capture" src="https://github.com/llvm/llvm-project/assets/6145511/461c848d-2e8a-4aa2-ba88-8065f63ce77a">

the real instruction causing memfault is most likely the `buffer_load_dword` right above it, where `s[0:3]` contains unreasonably large values.



https://github.com/llvm/llvm-project/pull/93668