[AMDGPU] Rework architected SGPRs implementation #79001
Conversation
Rework the architected SGPRs implementation such that workgroup id values (which live in TTMP registers) are available in all functions and do not rely on calling allocateSystemSGPRs to set them up.
| if (!AMDGPU::isGraphics(CC) || | ||
| (CC == CallingConv::AMDGPU_CS && ST.hasArchitectedSGPRs())) { | ||
| if (!AMDGPU::isGraphics(CC) || CC == CallingConv::AMDGPU_CS || | ||
| ST.hasArchitectedSGPRs()) { |
There was a problem hiding this comment.
This change looks redundant as this patch always allocate the TTMP* for subtargets with architectedSGPR enabled.
This was initially added with 2171f04.
You can revert this check to just have only !AMDGPU::isGraphics(CC)
There was a problem hiding this comment.
But I need WorkGroupIDZ to be set correctly when architected SGPRs are enabled. It is used below, line 177.
| ; GCN: s_mov_b64 s[4:5], s[0:1] | ||
| ; GCN: buffer_store_dword v{{[0-9]+}}, off, s[4:7], 0 offset:4 | ||
| ; GCN: s_mov_b64 s[8:9], s[0:1] | ||
| ; GCN: buffer_store_dword v{{[0-9]+}}, off, s[8:11], 0 offset:4 |
There was a problem hiding this comment.
No, I guess I've broken something.
There was a problem hiding this comment.
The following change? It now enables workgroup IDs for AMDGPU_CS always.
SIMachineFunctionInfo.cpp:110
- if (!AMDGPU::isGraphics(CC) ||
- (CC == CallingConv::AMDGPU_CS && ST.hasArchitectedSGPRs())) {
- if (!AMDGPU::isGraphics(CC) || CC == CallingConv::AMDGPU_CS ||
- ST.hasArchitectedSGPRs()) {
|
The WaveID support is missing. Hope that will be coming in a separate patch. |
| ArgDescriptor PrivateSegmentWaveByteOffset; | ||
|
|
||
| // System TTMPs. | ||
| ArgDescriptor ArchitectedWorkGroupIDX; |
There was a problem hiding this comment.
I don't think these really need to be tracked in ArgumentUsageInfo; they aren't arguments anymore
There was a problem hiding this comment.
OK. I started again from scratch: #79120
| } | ||
|
|
||
| if (STI->hasArchitectedSGPRs()) { | ||
| ArgInfo.ArchitectedWorkGroupIDX = |
There was a problem hiding this comment.
I think it's OK for the lowering to directly consume the hardcoded register numbers
There was a problem hiding this comment.
You mean not go via these ArgDescriptors? I'm only using them because they handle the shifting and masking.
3 participants
Rework the architected SGPRs implementation such that workgroup id
values (which live in TTMP registers) are available in all functions and
do not rely on calling allocateSystemSGPRs to set them up.