[X86] combineSCALAR_TO_VECTOR - fold (v4i32 (scalar_to_vector (i32 (bitcast (float))))) to MOVD/SS

[RKSimon]

Partial fix for #140693

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Simon Pilgrim (RKSimon)

Changes

Partial fix for #140693

---

Full diff: https://github.com/llvm/llvm-project/pull/140704.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+7-3)
• (modified) llvm/test/CodeGen/X86/scalarize-fp.ll (-2)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 2d52e65aca1fb..b5d930ed4f7c3 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -59800,14 +59800,18 @@ static SDValue combineSCALAR_TO_VECTOR(SDNode *N, SelectionDAG &DAG,
                                   DAG.getZExtOrTrunc(ZeroExt, DL, MVT::i32))));
   }
 
-  if (VT == MVT::v2i64 && Src.getOpcode() == ISD::BITCAST) {
+  if (Src.getOpcode() == ISD::BITCAST) {
     SDValue SrcOp = Src.getOperand(0);
+    // Combine (v4i32 (scalar_to_vector (i32 (bitcast (float))))) to MOVD.
+    if (VT == MVT::v4i32 && SrcOp.getValueType() == MVT::f32)
+      return DAG.getBitcast(
+          VT, DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v4f32, SrcOp));
     // Combine (v2i64 (scalar_to_vector (i64 (bitcast (double))))) to MOVQ.
-    if (SrcOp.getValueType() == MVT::f64)
+    if (VT == MVT::v2i64 && SrcOp.getValueType() == MVT::f64)
       return DAG.getBitcast(
           VT, DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64, SrcOp));
     // Combine (v2i64 (scalar_to_vector (i64 (bitcast (mmx))))) to MOVQ2DQ.
-    if (SrcOp.getValueType() == MVT::x86mmx)
+    if (VT == MVT::v2i64 && SrcOp.getValueType() == MVT::x86mmx)
       return DAG.getNode(X86ISD::MOVQ2DQ, DL, VT, SrcOp);
   }
 
diff --git a/llvm/test/CodeGen/X86/scalarize-fp.ll b/llvm/test/CodeGen/X86/scalarize-fp.ll
index 410de083a01ab..ea9b69f8f5b80 100644
--- a/llvm/test/CodeGen/X86/scalarize-fp.ll
+++ b/llvm/test/CodeGen/X86/scalarize-fp.ll
@@ -911,8 +911,6 @@ define <4 x float> @merge_fcmp_cmpeqss_v4f32(<4 x float> %x, <4 x float> %y) {
 ; AVX1-LABEL: merge_fcmp_cmpeqss_v4f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vcmpeqss %xmm0, %xmm1, %xmm1
-; AVX1-NEXT:    vmovd %xmm1, %eax
-; AVX1-NEXT:    vmovd %eax, %xmm1
 ; AVX1-NEXT:    vblendps {{.*#+}} xmm0 = xmm1[0],xmm0[1,2,3]
 ; AVX1-NEXT:    retq
 ;