[llvm] [AMDGPU] Add a feature for VOP3PX2 instructions incrementing VA_VDST twice (PR #202613)

[via llvm-commits]

llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

<details>
<summary>Changes</summary>



---
Full diff: https://github.com/llvm/llvm-project/pull/202613.diff


2 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/AMDGPU.td (+6) 
- (modified) llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp (+2-1) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPU.td b/llvm/lib/Target/AMDGPU/AMDGPU.td
index 2d014be12cad7..db375bc8803a1 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPU.td
+++ b/llvm/lib/Target/AMDGPU/AMDGPU.td
@@ -1403,6 +1403,10 @@ def FeatureUseAddPC64Inst : SubtargetFeature<"use-add-pc64-inst",
   "Use s_add_pc_i64 instruction."
 >;
 
+defm VOP3PX2IncrementsVaVdstTwice : AMDGPUSubtargetFeature<"vop3px2-increments-va-vdst-twice",
+  "VOP3PX2 encoded instructions increment the VaVdst counter twice", GenPredicate=false
+>;
+
 //===----------------------------------------------------------------------===//
 
 class GCNSubtargetFeatureGeneration <string Value,
@@ -2190,6 +2194,7 @@ def FeatureISAVersion12_50_Common : FeatureSet<
 def FeatureISAVersion12_50 : FeatureSet<
   !listconcat(FeatureISAVersion12_50_Common.Features,
   [FeatureAddressableLocalMemorySize327680,
+   FeatureVOP3PX2IncrementsVaVdstTwice,
    FeatureSetregVGPRMSBFixup,
    FeatureCubeInsts,
    FeatureLerpInst,
@@ -2207,6 +2212,7 @@ def FeatureISAVersion12_50 : FeatureSet<
 def FeatureISAVersion12_51 : FeatureSet<
   !listconcat(FeatureISAVersion12_50_Common.Features,
   [FeatureAddressableLocalMemorySize327680,
+   FeatureVOP3PX2IncrementsVaVdstTwice,
    FeatureDPALU_DPP,
    FeatureCubeInsts,
    FeatureLerpInst,
diff --git a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
index 7241c0db726ce..0839c6936c15e 100644
--- a/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInsertWaitcnts.cpp
@@ -1127,7 +1127,8 @@ void WaitcntBrackets::updateByEvent(WaitEventType E, MachineInstr &Inst) {
 
   unsigned UB = getScoreUB(T);
   unsigned Increment = 1;
-  if (T == AMDGPU::VA_VDST && AMDGPU::getHasMatrixScale(Inst.getOpcode())) {
+  if (T == AMDGPU::VA_VDST && AMDGPU::getHasMatrixScale(Inst.getOpcode()) &&
+      Context->ST.hasVOP3PX2IncrementsVaVdstTwice()) {
     // V_WMMA_SCALE instructions use VOP3PX2 encoding. Hardware treats this as
     // two VOP3P instructions and increments VA_VDST twice.
     Increment = 2;

``````````

</details>


https://github.com/llvm/llvm-project/pull/202613