[llvm] [SelectionDAG][AArch64] Add dot product lowering in NEON for PARTIAL_REDUCE_*MLA ISD nodes (PR #140075)
Sander de Smalen via llvm-commits
llvm-commits at lists.llvm.org
Mon May 26 23:32:33 PDT 2025
================
@@ -1451,8 +1451,16 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
for (MVT VT : { MVT::v32i8, MVT::v16i16, MVT::v8i32, MVT::v4i64 })
setOperationAction(ISD::ADD, VT, Custom);
// FADDP custom lowering
- for (MVT VT : { MVT::v16f16, MVT::v8f32, MVT::v4f64 })
+ for (MVT VT : {MVT::v16f16, MVT::v8f32, MVT::v4f64})
setOperationAction(ISD::FADD, VT, Custom);
+
+ if (EnablePartialReduceNodes && Subtarget->hasDotProd()) {
+ setPartialReduceMLAAction(MVT::v2i64, MVT::v8i16, Legal);
----------------
sdesmalen-arm wrote:
One more thing; I don't think this one is legal (the neon udot/sdot instructions only support i8 -> i32).
I guess this case is missing test coverage, because the following would fail to compile:
```
define <2 x i64> @foo(<2 x i64> %acc, <8 x i16> %input){
entry:
%input.wide = zext <8 x i16> %input to <8 x i64>
%partial.reduce = tail call <2 x i64> @llvm.experimental.vector.partial.reduce.add(<2 x i64> %acc, <8 x i64> %input.wide)
ret <2 x i64> %partial.reduce
}
```
https://github.com/llvm/llvm-project/pull/140075
More information about the llvm-commits
mailing list