[llvm] 2b1ce25 - AMDGPU: Fix -amdgpu-mfma-vgpr-form flag on gfx908 (#150599)

[via llvm-commits]

Author: Matt Arsenault
Date: 2025-07-25T19:49:56+09:00
New Revision: 2b1ce25e21765a07f69c48196bd15239d98cae92

URL: https://github.com/llvm/llvm-project/commit/2b1ce25e21765a07f69c48196bd15239d98cae92
DIFF: https://github.com/llvm/llvm-project/commit/2b1ce25e21765a07f69c48196bd15239d98cae92.diff

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

This should be ignored since there are no VGPR forms. This
makes it possible to flip the default for the flag to true.

Added: 
    

Modified: 
    llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
    llvm/test/CodeGen/AMDGPU/gfx90a-enc.ll

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
index f0be204cd9bdb..9a1448f1f95dc 100644
--- a/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIMachineFunctionInfo.cpp
@@ -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 = false; // We will select all MAI with VGPR operands.
+  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

diff  --git a/llvm/test/CodeGen/AMDGPU/gfx90a-enc.ll b/llvm/test/CodeGen/AMDGPU/gfx90a-enc.ll
index 99690e407d8da..fe8edd5d60912 100644
--- a/llvm/test/CodeGen/AMDGPU/gfx90a-enc.ll
+++ b/llvm/test/CodeGen/AMDGPU/gfx90a-enc.ll
@@ -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: