[llvm] [SelectionDAG] Scalarize binary ops of splats before legal types (PR #100749)
Luke Lau via llvm-commits
llvm-commits at lists.llvm.org
Wed Aug 7 01:54:04 PDT 2024
================
@@ -26993,11 +26993,20 @@ static SDValue scalarizeBinOpOfSplats(SDNode *N, SelectionDAG &DAG,
// TODO: use DAG.isSplatValue instead?
bool IsBothSplatVector = N0.getOpcode() == ISD::SPLAT_VECTOR &&
N1.getOpcode() == ISD::SPLAT_VECTOR;
+
+ // If binop is legal or custom on EltVT, scalarize should be profitable. The
+ // check is the same as isOperationLegalOrCustom without isTypeLegal. We
+ // can do this only before LegalTypes, because it may generate illegal `op
+ // EltVT` from legal `op VT (splat EltVT)`, where EltVT is not legal type but
+ // the result type of splat is legal.
+ auto EltAction = TLI.getOperationAction(Opcode, EltVT);
if (!Src0 || !Src1 || Index0 != Index1 ||
Src0.getValueType().getVectorElementType() != EltVT ||
Src1.getValueType().getVectorElementType() != EltVT ||
!(IsBothSplatVector || TLI.isExtractVecEltCheap(VT, Index0)) ||
- !TLI.isOperationLegalOrCustom(Opcode, EltVT))
+ (LegalTypes && !TLI.isTypeLegal(EltVT)) ||
+ !(EltAction == TargetLoweringBase::Legal ||
+ EltAction == TargetLoweringBase::Custom))
----------------
lukel97 wrote:
Sorry for the delay here: I took another look at this and I think this might actually be working by coincidence.
I'm not sure it makes much sense to check if an operation is legal for an illegal type. I think this just happens to work because by default all op actions are initiated to legal.
I think what we need to do is if we're before legal types, we check if the operation is legal on the legalized type:
```suggestion
// If before type legalization, allow scalar types that will eventually be
// made legal.
!TLI.isOperationLegalOrCustom(
Opcode, LegalTypes
? EltVT
: TLI.getTypeToTransformTo(*DAG.getContext(), EltVT)))
```
I tried this out locally and this actually fixed that `vmul_xx_nxv8i64` test case.
https://github.com/llvm/llvm-project/pull/100749
More information about the llvm-commits
mailing list