[X86] canonicalizeShuffleWithOp - initial support for shuffle(cvt(x),cvt(y)) -> cvt(shuffle(x,y))

Initial support is just for UNPCKL(CVTPH2PS(X),CVTPH2PS(Y)) -> CVTPH2PS(UNPCKL(X,Y))

Making this more general for other shuffles/conversions will have to be done carefully as we have to handle changes in src/dst element width, so I just handled the CVTPH2PS regression case.

Fixes #83414

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -41500,6 +41500,21 @@ static SDValue canonicalizeShuffleWithOp(SDValue N, SelectionDAG &DAG,
             ShuffleVT,
             DAG.getNode(SrcOpcode, DL, OpVT, DAG.getBitcast(OpVT, Res)));
       }
+      // TODO: We can generalize this for other shuffles/conversions.
+      if (Opc == X86ISD::UNPCKL && SrcOpcode == X86ISD::CVTPH2PS &&
+          N1.getOpcode() == SrcOpcode &&
+          N0.getValueType() == N1.getValueType() &&
+          N0.getOperand(0).getValueType() == N1.getOperand(0).getValueType() &&
+          ShuffleVT.getScalarSizeInBits() == N0.getScalarValueSizeInBits() &&
+          IsSafeToMoveShuffle(N0, SrcOpcode) &&
+          IsSafeToMoveShuffle(N1, SrcOpcode)) {
+        EVT OpSrcVT = N0.getOperand(0).getValueType();
+        EVT OpDstVT = N0.getValueType();
+        SDValue Res =
+            DAG.getNode(Opc, DL, OpSrcVT, N0.getOperand(0), N1.getOperand(0));
+        return DAG.getBitcast(ShuffleVT,
+                              DAG.getNode(SrcOpcode, DL, OpDstVT, Res));
+      }
     }
     break;
   }

llvm/test/CodeGen/X86/vector-half-conversions.ll

@@ -4966,22 +4966,18 @@ define <4 x i32> @fptosi_2f16_to_4i32(<2 x half> %a) nounwind {
 ;
 ; F16C-LABEL: fptosi_2f16_to_4i32:
 ; F16C:       # %bb.0:
-; F16C-NEXT:    vpsrld $16, %xmm0, %xmm1
-; F16C-NEXT:    vcvtph2ps %xmm1, %xmm1
-; F16C-NEXT:    vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; F16C-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; F16C-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; F16C-NEXT:    vcvtph2ps %xmm0, %xmm0
-; F16C-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; F16C-NEXT:    vcvttps2dq %xmm0, %xmm0
 ; F16C-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
 ; F16C-NEXT:    retq
 ;
 ; AVX512-LABEL: fptosi_2f16_to_4i32:
 ; AVX512:       # %bb.0:
-; AVX512-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX512-NEXT:    vcvtph2ps %xmm1, %xmm1
-; AVX512-NEXT:    vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
-; AVX512-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; AVX512-NEXT:    vcvttps2dq %xmm0, %xmm0
 ; AVX512-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
 ; AVX512-NEXT:    retq
@@ -5084,11 +5080,9 @@ define <4 x i32> @fptoui_2f16_to_4i32(<2 x half> %a) nounwind {
 ;
 ; F16C-LABEL: fptoui_2f16_to_4i32:
 ; F16C:       # %bb.0:
-; F16C-NEXT:    vpsrld $16, %xmm0, %xmm1
-; F16C-NEXT:    vcvtph2ps %xmm1, %xmm1
-; F16C-NEXT:    vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; F16C-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; F16C-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; F16C-NEXT:    vcvtph2ps %xmm0, %xmm0
-; F16C-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; F16C-NEXT:    vcvttps2dq %xmm0, %xmm1
 ; F16C-NEXT:    vpsrad $31, %xmm1, %xmm2
 ; F16C-NEXT:    vsubps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm0, %xmm0
@@ -5100,23 +5094,19 @@ define <4 x i32> @fptoui_2f16_to_4i32(<2 x half> %a) nounwind {
 ;
 ; AVX512F-LABEL: fptoui_2f16_to_4i32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX512F-NEXT:    vcvtph2ps %xmm1, %xmm1
-; AVX512F-NEXT:    vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512F-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX512F-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX512F-NEXT:    vcvtph2ps %xmm0, %xmm0
-; AVX512F-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; AVX512F-NEXT:    vcvttps2udq %zmm0, %zmm0
 ; AVX512F-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512-FASTLANE-LABEL: fptoui_2f16_to_4i32:
 ; AVX512-FASTLANE:       # %bb.0:
-; AVX512-FASTLANE-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX512-FASTLANE-NEXT:    vcvtph2ps %xmm1, %xmm1
-; AVX512-FASTLANE-NEXT:    vpmovzxwd {{.*#+}} xmm0 = xmm0[0],zero,xmm0[1],zero,xmm0[2],zero,xmm0[3],zero
+; AVX512-FASTLANE-NEXT:    vxorps %xmm1, %xmm1, %xmm1
+; AVX512-FASTLANE-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX512-FASTLANE-NEXT:    vcvtph2ps %xmm0, %xmm0
-; AVX512-FASTLANE-NEXT:    vunpcklps {{.*#+}} xmm0 = xmm0[0],xmm1[0],xmm0[1],xmm1[1]
 ; AVX512-FASTLANE-NEXT:    vcvttps2udq %xmm0, %xmm0
 ; AVX512-FASTLANE-NEXT:    vmovq {{.*#+}} xmm0 = xmm0[0],zero
 ; AVX512-FASTLANE-NEXT:    retq