[X86] Fold extract_subvector(fp_to_uint(x)) case to match existing fp_to_sint fold (necessary to fix llvm#83402 on AVX512 targets).

Prep work for llvm#83402

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -57167,8 +57167,10 @@ static SDValue combineEXTRACT_SUBVECTOR(SDNode *N, SelectionDAG &DAG,
         return DAG.getNode(X86ISD::VFPEXT, DL, VT, InVec.getOperand(0));
       }
     }
-    // v4i32 CVTPS2DQ(v4f32).
-    if (InOpcode == ISD::FP_TO_SINT && VT == MVT::v4i32) {
+    // v4i32 CVTPS2DQ(v4f32) / CVTPS2UDQ(v4f32).
+    if ((InOpcode == ISD::FP_TO_SINT ||
+         (InOpcode == ISD::FP_TO_UINT && Subtarget.hasVLX())) &&
+        VT == MVT::v4i32) {
       SDValue Src = InVec.getOperand(0);
       if (Src.getValueType().getScalarType() == MVT::f32)
         return DAG.getNode(InOpcode, DL, VT,

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

@@ -5233,8 +5233,7 @@ define <4 x i32> @fptoui_4f16_to_4i32(<4 x half> %a) nounwind {
 ; AVX512-FASTLANE-LABEL: fptoui_4f16_to_4i32:
 ; AVX512-FASTLANE:       # %bb.0:
 ; AVX512-FASTLANE-NEXT:    vcvtph2ps %xmm0, %ymm0
-; AVX512-FASTLANE-NEXT:    vcvttps2udq %ymm0, %ymm0
-; AVX512-FASTLANE-NEXT:    # kill: def $xmm0 killed $xmm0 killed $ymm0
+; AVX512-FASTLANE-NEXT:    vcvttps2udq %xmm0, %xmm0
 ; AVX512-FASTLANE-NEXT:    vzeroupper
 ; AVX512-FASTLANE-NEXT:    retq
   %cvt = fptoui <4 x half> %a to <4 x i32>