[X86] Fold (v4i32 (scalar_to_vector (i32 (zextext (bitcast (f16)))))) -> (v4i32 bitcast (shuffle (v8f16 scalar_to_vector)))

Extension to #123338

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -58721,12 +58721,20 @@ static SDValue combineSCALAR_TO_VECTOR(SDNode *N, SelectionDAG &DAG,
 
   if (VT == MVT::v4i32) {
     SDValue HalfSrc;
-    // Combine (v4i32 (scalar_to_vector (i32 (anyext (bitcast (f16))))))
+    // Combine (v4i32 (scalar_to_vector (i32 (a/zext (bitcast (f16))))))
     // to remove XMM->GPR->XMM moves.
     if (sd_match(Src, m_AnyExt(m_BitCast(
                           m_AllOf(m_SpecificVT(MVT::f16), m_Value(HalfSrc))))))
       return DAG.getBitcast(
           VT, DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v8f16, HalfSrc));
+    if (sd_match(Src, m_ZExt(m_BitCast(m_AllOf(m_SpecificVT(MVT::f16),
+                                               m_Value(HalfSrc)))))) {
+      SDValue R = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v8f16, HalfSrc);
+      R = DAG.getVectorShuffle(MVT::v8f16, DL, R,
+                               getZeroVector(MVT::v8f16, Subtarget, DAG, DL),
+                               {0, 8, -1, -1, -1, -1, -1, -1});
+      return DAG.getBitcast(VT, R);
+    }
   }
 
   // See if we're broadcasting the scalar value, in which case just reuse that.

llvm/test/CodeGen/X86/canonicalize-vars-f16-type.ll

@@ -43,13 +43,11 @@ define void @v_test_canonicalize__half(half addrspace(1)* %out) nounwind {
 ;
 ; AVX512-LABEL: v_test_canonicalize__half:
 ; AVX512:       # %bb.0: # %entry
-; AVX512-NEXT:    movzwl (%rdi), %eax
-; AVX512-NEXT:    movzwl {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %ecx
-; AVX512-NEXT:    vmovd %ecx, %xmm0
+; AVX512-NEXT:    vpinsrw $0, (%rdi), %xmm0, %xmm0
+; AVX512-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
-; AVX512-NEXT:    vmovd %eax, %xmm1
-; AVX512-NEXT:    vcvtph2ps %xmm1, %xmm1
-; AVX512-NEXT:    vmulss %xmm0, %xmm1, %xmm0
+; AVX512-NEXT:    vcvtph2ps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
+; AVX512-NEXT:    vmulss %xmm1, %xmm0, %xmm0
 ; AVX512-NEXT:    vxorps %xmm1, %xmm1, %xmm1
 ; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm1[1,2,3]
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
@@ -144,9 +142,7 @@ define half @complex_canonicalize_fmul_half(half %a, half %b) nounwind {
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
-; AVX512-NEXT:    movzwl {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %eax
-; AVX512-NEXT:    vmovd %eax, %xmm2
-; AVX512-NEXT:    vcvtph2ps %xmm2, %xmm2
+; AVX512-NEXT:    vcvtph2ps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm2
 ; AVX512-NEXT:    vmulss %xmm2, %xmm0, %xmm0
 ; AVX512-NEXT:    vxorps %xmm2, %xmm2, %xmm2
 ; AVX512-NEXT:    vblendps {{.*#+}} xmm0 = xmm0[0],xmm2[1,2,3]
@@ -228,9 +224,7 @@ define void @v_test_canonicalize_v2half(<2 x half> addrspace(1)* %out) nounwind
 ; AVX512-LABEL: v_test_canonicalize_v2half:
 ; AVX512:       # %bb.0: # %entry
 ; AVX512-NEXT:    vmovd {{.*#+}} xmm0 = mem[0],zero,zero,zero
-; AVX512-NEXT:    movzwl {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %eax
-; AVX512-NEXT:    vmovd %eax, %xmm1
-; AVX512-NEXT:    vcvtph2ps %xmm1, %xmm1
+; AVX512-NEXT:    vcvtph2ps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm1
 ; AVX512-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[2,3],zero,zero,zero,zero,zero,zero,xmm0[u,u,u,u,u,u,u,u]
 ; AVX512-NEXT:    vcvtph2ps %xmm2, %xmm2
 ; AVX512-NEXT:    vmulss %xmm1, %xmm2, %xmm2

llvm/test/CodeGen/X86/fminimumnum-fmaximumnum.ll

@@ -1840,9 +1840,7 @@ define <4 x half> @test_fmaximumnum_v4f16(<4 x half> %x, <4 x half> %y) nounwind
 ; AVX512-NEXT:    vcvtps2ph $4, %xmm0, %xmm0
 ; AVX512-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; AVX512-NEXT:    vcvtph2ps %xmm0, %xmm0
-; AVX512-NEXT:    movzwl {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %eax
-; AVX512-NEXT:    vmovd %eax, %xmm1
-; AVX512-NEXT:    vcvtph2ps %xmm1, %xmm9
+; AVX512-NEXT:    vcvtph2ps {{\.?LCPI[0-9]+_[0-9]+}}(%rip), %xmm9
 ; AVX512-NEXT:    vmulss %xmm0, %xmm9, %xmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm4[3,3,3,3]
 ; AVX512-NEXT:    vcvtph2ps %xmm1, %xmm1

llvm/test/CodeGen/X86/fp-round.ll

@@ -50,9 +50,7 @@ define half @round_f16(half %h) {
 ;
 ; AVX512F-LABEL: round_f16:
 ; AVX512F:       # %bb.0: # %entry
-; AVX512F-NEXT:    vpextrw $0, %xmm0, %eax
-; AVX512F-NEXT:    movzwl %ax, %eax
-; AVX512F-NEXT:    vmovd %eax, %xmm0
+; AVX512F-NEXT:    vpmovzxwq {{.*#+}} xmm0 = xmm0[0],zero,zero,zero,xmm0[1],zero,zero,zero
 ; AVX512F-NEXT:    vcvtph2ps %xmm0, %xmm0
 ; AVX512F-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [4.9999997E-1,4.9999997E-1,4.9999997E-1,4.9999997E-1]
 ; AVX512F-NEXT:    vpternlogd {{.*#+}} xmm1 = xmm1 | (xmm0 & mem)