[AMDGPU] StackSlotColoring before LowerSGPRSpill causes memfault in Blender 4.1 on gfx908

[yxsamliu]

#93668 caused a regression in Blender 4.1 on gfx908 for all scenes.

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

[llvmbot]

@llvm/issue-subscribers-backend-amdgpu

Author: Yaxun (Sam) Liu (yxsamliu)

https://github.com//pull/93668 caused a regression in Blender 4.1 on gfx908 for all scenes.

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

[vg0204]

After spending time on ISA of good/bad cases, I found that s[0:3] represents the ScratchRsrcReg here, which is used as buffer descriptor(input) for buffer loads & stores, and thus most probably never get redefined or clobbered {since they are reserved & marked live throughout function}. Would it be better if I execute the given assembly & try to debug it?

[yxsamliu]

It seems DC's patches fix the issue, so we do not need to revert your patch.

93526 : [AMDGPU] Split VGPR regalloc pipeline

93525: [CodeGen] change prototype of regalloc filter function

[vg0204]

DC's patches

It seems DC's patches fix the issue, so we do not need to revert your patch.

93526 : [AMDGPU] Split VGPR regalloc pipeline

93525: [CodeGen] change prototype of regalloc filter function

How should I ensure that, as I already reverted the patch, before reverting the revert commit.

[yxsamliu]

I have verified that with those patches on top of amd-staging branch.

I will verify them with LLVM main branch.

Since you have reverted your PR. Let's re-enable blender test in buildbot and wait for these two PR's committed then reland your PR.

[vg0204]

Yeah, sounds good!