[AArch64] Scalarize v2f16 vecreduce.fadd

[davemgreen]

This adds a custom lowering for v2f16 vecreduce.fadd to scalarize as opposed to padding with zeroes. This allows it to generate the more efficient faddp. Helps with #147583.

[llvmbot]

@llvm/pr-subscribers-backend-aarch64

Author: David Green (davemgreen)

Changes

This adds a custom lowering for v2f16 vecreduce.fadd to scalarize as opposed to padding with zeroes. This allows it to generate the more efficient faddp. Helps with #147583.

---

Full diff: https://github.com/llvm/llvm-project/pull/147783.diff

2 Files Affected:

• (modified) llvm/lib/Target/AArch64/AArch64ISelLowering.cpp (+14-1)
• (modified) llvm/test/CodeGen/AArch64/vecreduce-fadd.ll (-3)

diff --git a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
index 01be10be433fd..3cfe05d3559e0 100644
--- a/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
+++ b/llvm/lib/Target/AArch64/AArch64ISelLowering.cpp
@@ -1352,6 +1352,9 @@ AArch64TargetLowering::AArch64TargetLowering(const TargetMachine &TM,
         setOperationAction(ISD::VECREDUCE_FADD, VT, Legal);
       }
     }
+    if (Subtarget->hasFullFP16())
+      setOperationAction(ISD::VECREDUCE_FADD, MVT::v2f16, Custom);
+
     for (MVT VT : { MVT::v8i8, MVT::v4i16, MVT::v2i32,
                     MVT::v16i8, MVT::v8i16, MVT::v4i32 }) {
       setOperationAction(ISD::VECREDUCE_ADD, VT, Custom);
@@ -16046,9 +16049,19 @@ static SDValue getVectorBitwiseReduce(unsigned Opcode, SDValue Vec, EVT VT,
 SDValue AArch64TargetLowering::LowerVECREDUCE(SDValue Op,
                                               SelectionDAG &DAG) const {
   SDValue Src = Op.getOperand(0);
+  EVT SrcVT = Src.getValueType();
+
+  // Scalarize v2f32 to turn it into a faddp. This will be more efficient than
+  // widening by inserting zeroes.
+  if (Subtarget->hasFullFP16() && Op.getOpcode() == ISD::VECREDUCE_FADD &&
+      SrcVT == MVT::v2f16) {
+    SDLoc DL(Op);
+    return DAG.getNode(ISD::FADD, DL, MVT::f16,
+                       DAG.getExtractVectorElt(DL, MVT::f16, Src, 0),
+                       DAG.getExtractVectorElt(DL, MVT::f16, Src, 1));
+  }
 
   // Try to lower fixed length reductions to SVE.
-  EVT SrcVT = Src.getValueType();
   bool OverrideNEON = !Subtarget->isNeonAvailable() ||
                       Op.getOpcode() == ISD::VECREDUCE_AND ||
                       Op.getOpcode() == ISD::VECREDUCE_OR ||
diff --git a/llvm/test/CodeGen/AArch64/vecreduce-fadd.ll b/llvm/test/CodeGen/AArch64/vecreduce-fadd.ll
index 2e993a85760c6..8a84d3ca2328c 100644
--- a/llvm/test/CodeGen/AArch64/vecreduce-fadd.ll
+++ b/llvm/test/CodeGen/AArch64/vecreduce-fadd.ll
@@ -27,9 +27,6 @@ define half @add_v2HalfH(<2 x half> %bin.rdx)  {
 ; CHECK-SD-FP16-LABEL: add_v2HalfH:
 ; CHECK-SD-FP16:       // %bb.0:
 ; CHECK-SD-FP16-NEXT:    // kill: def $d0 killed $d0 def $q0
-; CHECK-SD-FP16-NEXT:    mov v0.h[2], wzr
-; CHECK-SD-FP16-NEXT:    mov v0.h[3], wzr
-; CHECK-SD-FP16-NEXT:    faddp v0.4h, v0.4h, v0.4h
 ; CHECK-SD-FP16-NEXT:    faddp h0, v0.2h
 ; CHECK-SD-FP16-NEXT:    ret
 ;

[SamTebbs33]

v2f32 -> v2f16 ?

[SamTebbs33]

Nice, LGTM.