shylie/llvm-project
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

AMDGPU: Fix -amdgpu-mfma-vgpr-form flag on gfx908 (#150599)

Browse Source 
This should be ignored since there are no VGPR forms. This
makes it possible to flip the default for the flag to true.
This commit is contained in:
Matt Arsenault 2025-07-25 19:49:56 +09:00 committed by GitHub
parent 129a35454c
commit 2b1ce25e21
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 12 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
View File

@ -81,11 +81,15 @@ SIMachineFunctionInfo::SIMachineFunctionInfo(const Function &F,
PSInputAddr = AMDGPU::getInitialPSInputAddr(F);
}
MayNeedAGPRs = ST.hasMAIInsts() && !MFMAVGPRForm;
if (!MFMAVGPRForm && ST.hasGFX90AInsts() &&
ST.getMaxNumVGPRs(F) <= AMDGPU::VGPR_32RegClass.getNumRegs() &&
!mayUseAGPRs(F))
MayNeedAGPRs = ST.hasMAIInsts();
if (ST.hasGFX90AInsts()) {
// FIXME: MayNeedAGPRs is a misnomer for how this is used. MFMA selection
// should be separated from availability of AGPRs
if (MFMAVGPRForm ||
(ST.getMaxNumVGPRs(F) <= AMDGPU::VGPR_32RegClass.getNumRegs() &&
!mayUseAGPRs(F)))
MayNeedAGPRs = false; // We will select all MAI with VGPR operands.
}
if (AMDGPU::isChainCC(CC)) {
// Chain functions don't receive an SP from their caller, but are free to

3
llvm/test/CodeGen/AMDGPU/gfx90a-enc.ll
View File

@ -1,4 +1,7 @@
; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX9,GFX908 %s
; Make sure flag is ignored
; RUN: llc -mtriple=amdgcn -mcpu=gfx908 -amdgpu-mfma-vgpr-form=1 -show-mc-encoding < %s | FileCheck -check-prefixes=GFX9,GFX908 %s
; RUN: llc -mtriple=amdgcn -mcpu=gfx90a -show-mc-encoding < %s | FileCheck -check-prefixes=GFX9,GFX90A %s
; GFX9-DAG: buffer_load_format_xyzw v[{{[0-9:]+}}], v{{[0-9]+}}, s[{{[0-9:]+}}], 0 idxen ; encoding: