[LegalizeTypes] Add widening support for STRICT_FSETCC/FSETCCS

[craig.topper]

Summary

This patch adds widening which really just scalarizes because we don't have a strategy for the extra elements we would need to pad with.

[andrew.w.kaylor]

lgtm

llvm/lib/CodeGen/SelectionDAG/LegalizeTypes.h

private:
   SDValue WidenVecRes_LOAD(SDNode* N);
   SDValue WidenVecRes_MLOAD(MaskedLoadSDNode* N);
   SDValue WidenVecRes_MGATHER(MaskedGatherSDNode* N);
   SDValue WidenVecRes_SCALAR_TO_VECTOR(SDNode* N);
   SDValue WidenVecRes_SELECT(SDNode* N);
   SDValue WidenVSELECTAndMask(SDNode *N);
   SDValue WidenVecRes_SELECT_CC(SDNode* N);
   SDValue WidenVecRes_SETCC(SDNode* N);
+  SDValue WidenVecRes_STRICT_FSETCC(SDNode* N);
   SDValue WidenVecRes_UNDEF(SDNode *N);
   SDValue WidenVecRes_VECTOR_SHUFFLE(ShuffleVectorSDNode *N);
 
   SDValue WidenVecRes_Ternary(SDNode *N);
   SDValue WidenVecRes_Binary(SDNode *N);
   SDValue WidenVecRes_BinaryCanTrap(SDNode *N);
   SDValue WidenVecRes_BinaryWithExtraScalarOp(SDNode *N);
   SDValue WidenVecRes_StrictFP(SDNode *N);
private:
   SDValue WidenVecOp_EXTEND(SDNode *N);
   SDValue WidenVecOp_EXTRACT_VECTOR_ELT(SDNode *N);
   SDValue WidenVecOp_EXTRACT_SUBVECTOR(SDNode *N);
   SDValue WidenVecOp_STORE(SDNode* N);
   SDValue WidenVecOp_MSTORE(SDNode* N, unsigned OpNo);
   SDValue WidenVecOp_MGATHER(SDNode* N, unsigned OpNo);
   SDValue WidenVecOp_MSCATTER(SDNode* N, unsigned OpNo);
   SDValue WidenVecOp_SETCC(SDNode* N);
+  SDValue WidenVecOp_STRICT_FSETCC(SDNode* N);
   SDValue WidenVecOp_VSELECT(SDNode *N);
 
   SDValue WidenVecOp_Convert(SDNode *N);
   SDValue WidenVecOp_FCOPYSIGN(SDNode *N);
   SDValue WidenVecOp_VECREDUCE(SDNode *N);
 
   /// Helper function to generate a set of operations to perform
   /// a vector operation for a wider type.

llvm/lib/CodeGen/SelectionDAG/LegalizeVectorTypes.cpp

while (CurNumElts != 0) {
     }
   }
 
   return CollectOpsToWiden(DAG, TLI, ConcatOps, ConcatEnd, VT, MaxVT, WidenVT);
 }
 
 SDValue DAGTypeLegalizer::WidenVecRes_StrictFP(SDNode *N) {
   switch (N->getOpcode()) {
+  case ISD::STRICT_FSETCC:
+  case ISD::STRICT_FSETCCS:
+    return WidenVecRes_STRICT_FSETCC(N);
   case ISD::STRICT_FP_EXTEND:
   case ISD::STRICT_FP_ROUND:
   case ISD::STRICT_FP_TO_SINT:
   case ISD::STRICT_FP_TO_UINT:
   case ISD::STRICT_SINT_TO_FP:
   case ISD::STRICT_UINT_TO_FP:
    return WidenVecRes_Convert_StrictFP(N);
   default:
SDValue DAGTypeLegalizer::WidenVecRes_SETCC(SDNode *N) {
   assert(InOp1.getValueType() == WidenInVT &&
          InOp2.getValueType() == WidenInVT &&
          "Input not widened to expected type!");
   (void)WidenInVT;
   return DAG.getNode(ISD::SETCC, SDLoc(N),
                      WidenVT, InOp1, InOp2, N->getOperand(2));
 }
 
+SDValue DAGTypeLegalizer::WidenVecRes_STRICT_FSETCC(SDNode *N) {
+  assert(N->getValueType(0).isVector() &&
+         N->getOperand(1).getValueType().isVector() &&
+         "Operands must be vectors");
+  EVT VT = N->getValueType(0);
+  EVT WidenVT = TLI.getTypeToTransformTo(*DAG.getContext(), VT);
+  unsigned WidenNumElts = WidenVT.getVectorNumElements();
+  unsigned NumElts = VT.getVectorNumElements();
+  EVT EltVT = VT.getVectorElementType();
+
+  SDLoc dl(N);
+  SDValue Chain = N->getOperand(0);
+  SDValue LHS = N->getOperand(1);
+  SDValue RHS = N->getOperand(2);
+  SDValue CC = N->getOperand(3);
+  EVT TmpEltVT = LHS.getValueType().getVectorElementType();
+
+  // Fully unroll and reassemble.
+  SmallVector<SDValue, 8> Scalars(WidenNumElts, DAG.getUNDEF(EltVT));
+  SmallVector<SDValue, 8> Chains(NumElts);
+  for (unsigned i = 0; i != NumElts; ++i) {
+    SDValue LHSElem = DAG.getNode(
+        ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
+        DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
+    SDValue RHSElem = DAG.getNode(
+        ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
+        DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
+
+    Scalars[i] = DAG.getNode(N->getOpcode(), dl, {MVT::i1, MVT::Other},
+                             {Chain, LHSElem, RHSElem, CC});
+    Chains[i] = Scalars[i].getValue(1);
+    Scalars[i] = DAG.getSelect(dl, EltVT, Scalars[i],
+                               DAG.getBoolConstant(true, dl, EltVT, VT),
+                               DAG.getBoolConstant(false, dl, EltVT, VT));
+  }
+
+  SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
+  ReplaceValueWith(SDValue(N, 1), NewChain);
+
+  return DAG.getBuildVector(WidenVT, dl, Scalars);
+}
 
 //===----------------------------------------------------------------------===//
 // Widen Vector Operand
 //===----------------------------------------------------------------------===//
 bool DAGTypeLegalizer::WidenVectorOperand(SDNode *N, unsigned OpNo) {
   LLVM_DEBUG(dbgs() << "Widen node operand " << OpNo << ": "; N->dump(&DAG);
              dbgs() << "\n");
   SDValue Res = SDValue();
#endif
   case ISD::CONCAT_VECTORS:     Res = WidenVecOp_CONCAT_VECTORS(N); break;
   case ISD::EXTRACT_SUBVECTOR:  Res = WidenVecOp_EXTRACT_SUBVECTOR(N); break;
   case ISD::EXTRACT_VECTOR_ELT: Res = WidenVecOp_EXTRACT_VECTOR_ELT(N); break;
   case ISD::STORE:              Res = WidenVecOp_STORE(N); break;
   case ISD::MSTORE:             Res = WidenVecOp_MSTORE(N, OpNo); break;
   case ISD::MGATHER:            Res = WidenVecOp_MGATHER(N, OpNo); break;
   case ISD::MSCATTER:           Res = WidenVecOp_MSCATTER(N, OpNo); break;
   case ISD::SETCC:              Res = WidenVecOp_SETCC(N); break;
+  case ISD::STRICT_FSETCC:
+  case ISD::STRICT_FSETCCS:     Res = WidenVecOp_STRICT_FSETCC(N); break;
   case ISD::VSELECT:            Res = WidenVecOp_VSELECT(N); break;
   case ISD::FCOPYSIGN:          Res = WidenVecOp_FCOPYSIGN(N); break;
 
   case ISD::ANY_EXTEND:
   case ISD::SIGN_EXTEND:
   case ISD::ZERO_EXTEND:
     Res = WidenVecOp_EXTEND(N);
     break;
SDValue CC = DAG.getNode(
       DAG.getConstant(0, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
 
   EVT OpVT = N->getOperand(0).getValueType();
   ISD::NodeType ExtendCode =
       TargetLowering::getExtendForContent(TLI.getBooleanContents(OpVT));
   return DAG.getNode(ExtendCode, dl, VT, CC);
 }
 
+SDValue DAGTypeLegalizer::WidenVecOp_STRICT_FSETCC(SDNode *N) {
+  SDValue Chain = N->getOperand(0);
+  SDValue LHS = GetWidenedVector(N->getOperand(1));
+  SDValue RHS = GetWidenedVector(N->getOperand(2));
+  SDValue CC = N->getOperand(3);
+  SDLoc dl(N);
+
+  EVT VT = N->getValueType(0);
+  EVT EltVT = VT.getVectorElementType();
+  EVT TmpEltVT = LHS.getValueType().getVectorElementType();
+  unsigned NumElts = VT.getVectorNumElements();
+
+  // Unroll into a build vector.
+  SmallVector<SDValue, 8> Scalars(NumElts);
+  SmallVector<SDValue, 8> Chains(NumElts);
+
+  for (unsigned i = 0; i != NumElts; ++i) {
+    SDValue LHSElem = DAG.getNode(
+        ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, LHS,
+        DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
+    SDValue RHSElem = DAG.getNode(
+        ISD::EXTRACT_VECTOR_ELT, dl, TmpEltVT, RHS,
+        DAG.getConstant(i, dl, TLI.getVectorIdxTy(DAG.getDataLayout())));
+
+    Scalars[i] = DAG.getNode(N->getOpcode(), dl, {MVT::i1, MVT::Other},
+                             {Chain, LHSElem, RHSElem, CC});
+    Chains[i] = Scalars[i].getValue(1);
+    Scalars[i] = DAG.getSelect(dl, EltVT, Scalars[i],
+                               DAG.getBoolConstant(true, dl, EltVT, VT),
+                               DAG.getBoolConstant(false, dl, EltVT, VT));
+  }
+
+  SDValue NewChain = DAG.getNode(ISD::TokenFactor, dl, MVT::Other, Chains);
+  ReplaceValueWith(SDValue(N, 1), NewChain);
+
+  return DAG.getBuildVector(VT, dl, Scalars);
+}
+
 SDValue DAGTypeLegalizer::WidenVecOp_VECREDUCE(SDNode *N) {
   SDLoc dl(N);
   SDValue Op = GetWidenedVector(N->getOperand(0));
   EVT OrigVT = N->getOperand(0).getValueType();
   EVT WideVT = Op.getValueType();
   EVT ElemVT = OrigVT.getVectorElementType();
 
   SDValue NeutralElem;

llvm/test/CodeGen/X86/vec-strict-cmp-sub128.ll

+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 -O3 | FileCheck %s --check-prefixes=CHECK,SSE,SSE-32
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 -O3 | FileCheck %s --check-prefixes=CHECK,SSE,SSE-64
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx -O3 | FileCheck %s --check-prefixes=CHECK,AVX,AVX-32
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx -O3 | FileCheck %s --check-prefixes=CHECK,AVX,AVX-64
+; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512f -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=CHECK,AVX512-32
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=CHECK,AVX512-64
+
+define <2 x i32> @test_v2f32_ogt_s(<2 x i32> %a, <2 x i32> %b, <2 x float> %f1, <2 x float> %f2) #0 {
+; SSE-32-LABEL: test_v2f32_ogt_s:
+; SSE-32:       # %bb.0:
+; SSE-32-NEXT:    pushl %ebp
+; SSE-32-NEXT:    movl %esp, %ebp
+; SSE-32-NEXT:    andl $-16, %esp
+; SSE-32-NEXT:    subl $16, %esp
+; SSE-32-NEXT:    movaps 8(%ebp), %xmm3
+; SSE-32-NEXT:    xorl %eax, %eax
+; SSE-32-NEXT:    comiss %xmm3, %xmm2
+; SSE-32-NEXT:    movl $-1, %ecx
+; SSE-32-NEXT:    movl $0, %edx
+; SSE-32-NEXT:    cmoval %ecx, %edx
+; SSE-32-NEXT:    movd %edx, %xmm4
+; SSE-32-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; SSE-32-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,2,3]
+; SSE-32-NEXT:    comiss %xmm3, %xmm2
+; SSE-32-NEXT:    cmoval %ecx, %eax
+; SSE-32-NEXT:    movd %eax, %xmm2
+; SSE-32-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
+; SSE-32-NEXT:    pand %xmm4, %xmm0
+; SSE-32-NEXT:    pandn %xmm1, %xmm4
+; SSE-32-NEXT:    por %xmm4, %xmm0
+; SSE-32-NEXT:    movl %ebp, %esp
+; SSE-32-NEXT:    popl %ebp
+; SSE-32-NEXT:    retl
+;
+; SSE-64-LABEL: test_v2f32_ogt_s:
+; SSE-64:       # %bb.0:
+; SSE-64-NEXT:    xorl %eax, %eax
+; SSE-64-NEXT:    comiss %xmm3, %xmm2
+; SSE-64-NEXT:    movl $-1, %ecx
+; SSE-64-NEXT:    movl $0, %edx
+; SSE-64-NEXT:    cmoval %ecx, %edx
+; SSE-64-NEXT:    movd %edx, %xmm4
+; SSE-64-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; SSE-64-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,2,3]
+; SSE-64-NEXT:    comiss %xmm3, %xmm2
+; SSE-64-NEXT:    cmoval %ecx, %eax
+; SSE-64-NEXT:    movd %eax, %xmm2
+; SSE-64-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
+; SSE-64-NEXT:    pand %xmm4, %xmm0
+; SSE-64-NEXT:    pandn %xmm1, %xmm4
+; SSE-64-NEXT:    por %xmm4, %xmm0
+; SSE-64-NEXT:    retq
+;
+; AVX-32-LABEL: test_v2f32_ogt_s:
+; AVX-32:       # %bb.0:
+; AVX-32-NEXT:    pushl %ebp
+; AVX-32-NEXT:    movl %esp, %ebp
+; AVX-32-NEXT:    andl $-16, %esp
+; AVX-32-NEXT:    subl $16, %esp
+; AVX-32-NEXT:    vmovaps 8(%ebp), %xmm3
+; AVX-32-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
+; AVX-32-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
+; AVX-32-NEXT:    xorl %eax, %eax
+; AVX-32-NEXT:    vcomiss %xmm4, %xmm5
+; AVX-32-NEXT:    movl $-1, %ecx
+; AVX-32-NEXT:    movl $0, %edx
+; AVX-32-NEXT:    cmoval %ecx, %edx
+; AVX-32-NEXT:    vcomiss %xmm3, %xmm2
+; AVX-32-NEXT:    cmoval %ecx, %eax
+; AVX-32-NEXT:    vmovd %eax, %xmm2
+; AVX-32-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
+; AVX-32-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX-32-NEXT:    movl %ebp, %esp
+; AVX-32-NEXT:    popl %ebp
+; AVX-32-NEXT:    retl
+;
+; AVX-64-LABEL: test_v2f32_ogt_s:
+; AVX-64:       # %bb.0:
+; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
+; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
+; AVX-64-NEXT:    xorl %eax, %eax
+; AVX-64-NEXT:    vcomiss %xmm4, %xmm5
+; AVX-64-NEXT:    movl $-1, %ecx
+; AVX-64-NEXT:    movl $0, %edx
+; AVX-64-NEXT:    cmoval %ecx, %edx
+; AVX-64-NEXT:    vcomiss %xmm3, %xmm2
+; AVX-64-NEXT:    cmoval %ecx, %eax
+; AVX-64-NEXT:    vmovd %eax, %xmm2
+; AVX-64-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
+; AVX-64-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX-64-NEXT:    retq
+;
+; AVX512-32-LABEL: test_v2f32_ogt_s:
+; AVX512-32:       # %bb.0:
+; AVX512-32-NEXT:    pushl %ebp
+; AVX512-32-NEXT:    movl %esp, %ebp
+; AVX512-32-NEXT:    andl $-16, %esp
+; AVX512-32-NEXT:    subl $16, %esp
+; AVX512-32-NEXT:    vmovaps 8(%ebp), %xmm3
+; AVX512-32-NEXT:    movw $-3, %ax
+; AVX512-32-NEXT:    kmovw %eax, %k0
+; AVX512-32-NEXT:    vcomiss %xmm3, %xmm2
+; AVX512-32-NEXT:    seta %al
+; AVX512-32-NEXT:    andl $1, %eax
+; AVX512-32-NEXT:    kmovw %eax, %k1
+; AVX512-32-NEXT:    kandw %k0, %k1, %k0
+; AVX512-32-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; AVX512-32-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; AVX512-32-NEXT:    vcomiss %xmm3, %xmm2
+; AVX512-32-NEXT:    seta %al
+; AVX512-32-NEXT:    kmovw %eax, %k1
+; AVX512-32-NEXT:    kshiftlw $15, %k1, %k1
+; AVX512-32-NEXT:    kshiftrw $14, %k1, %k1
+; AVX512-32-NEXT:    korw %k1, %k0, %k1
+; AVX512-32-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
+; AVX512-32-NEXT:    movl %ebp, %esp
+; AVX512-32-NEXT:    popl %ebp
+; AVX512-32-NEXT:    retl
+;
+; AVX512-64-LABEL: test_v2f32_ogt_s:
+; AVX512-64:       # %bb.0:
+; AVX512-64-NEXT:    movw $-3, %ax
+; AVX512-64-NEXT:    kmovw %eax, %k0
+; AVX512-64-NEXT:    vcomiss %xmm3, %xmm2
+; AVX512-64-NEXT:    seta %al
+; AVX512-64-NEXT:    andl $1, %eax
+; AVX512-64-NEXT:    kmovw %eax, %k1
+; AVX512-64-NEXT:    kandw %k0, %k1, %k0
+; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm3 = xmm3[1,1,3,3]
+; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm2 = xmm2[1,1,3,3]
+; AVX512-64-NEXT:    vcomiss %xmm3, %xmm2
+; AVX512-64-NEXT:    seta %al
+; AVX512-64-NEXT:    kmovw %eax, %k1
+; AVX512-64-NEXT:    kshiftlw $15, %k1, %k1
+; AVX512-64-NEXT:    kshiftrw $14, %k1, %k1
+; AVX512-64-NEXT:    korw %k1, %k0, %k1
+; AVX512-64-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
+; AVX512-64-NEXT:    retq
+  %cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f32(
+                                               <2 x float> %f1, <2 x float> %f2, metadata !"ogt",
+                                               metadata !"fpexcept.strict") #0
+  %res = select <2 x i1> %cond, <2 x i32> %a, <2 x i32> %b
+  ret <2 x i32> %res
+}
+
+define <2 x i32> @test_v2f32_oeq_q(<2 x i32> %a, <2 x i32> %b, <2 x float> %f1, <2 x float> %f2) #0 {
+; SSE-32-LABEL: test_v2f32_oeq_q:
+; SSE-32:       # %bb.0:
+; SSE-32-NEXT:    pushl %ebp
+; SSE-32-NEXT:    movl %esp, %ebp
+; SSE-32-NEXT:    andl $-16, %esp
+; SSE-32-NEXT:    subl $16, %esp
+; SSE-32-NEXT:    movaps 8(%ebp), %xmm3
+; SSE-32-NEXT:    xorl %eax, %eax
+; SSE-32-NEXT:    ucomiss %xmm3, %xmm2
+; SSE-32-NEXT:    movl $-1, %ecx
+; SSE-32-NEXT:    movl $-1, %edx
+; SSE-32-NEXT:    cmovnel %eax, %edx
+; SSE-32-NEXT:    cmovpl %eax, %edx
+; SSE-32-NEXT:    movd %edx, %xmm4
+; SSE-32-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; SSE-32-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,2,3]
+; SSE-32-NEXT:    ucomiss %xmm3, %xmm2
+; SSE-32-NEXT:    cmovnel %eax, %ecx
+; SSE-32-NEXT:    cmovpl %eax, %ecx
+; SSE-32-NEXT:    movd %ecx, %xmm2
+; SSE-32-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
+; SSE-32-NEXT:    pand %xmm4, %xmm0
+; SSE-32-NEXT:    pandn %xmm1, %xmm4
+; SSE-32-NEXT:    por %xmm4, %xmm0
+; SSE-32-NEXT:    movl %ebp, %esp
+; SSE-32-NEXT:    popl %ebp
+; SSE-32-NEXT:    retl
+;
+; SSE-64-LABEL: test_v2f32_oeq_q:
+; SSE-64:       # %bb.0:
+; SSE-64-NEXT:    xorl %eax, %eax
+; SSE-64-NEXT:    ucomiss %xmm3, %xmm2
+; SSE-64-NEXT:    movl $-1, %ecx
+; SSE-64-NEXT:    movl $-1, %edx
+; SSE-64-NEXT:    cmovnel %eax, %edx
+; SSE-64-NEXT:    cmovpl %eax, %edx
+; SSE-64-NEXT:    movd %edx, %xmm4
+; SSE-64-NEXT:    shufps {{.*#+}} xmm3 = xmm3[1,1,2,3]
+; SSE-64-NEXT:    shufps {{.*#+}} xmm2 = xmm2[1,1,2,3]
+; SSE-64-NEXT:    ucomiss %xmm3, %xmm2
+; SSE-64-NEXT:    cmovnel %eax, %ecx
+; SSE-64-NEXT:    cmovpl %eax, %ecx
+; SSE-64-NEXT:    movd %ecx, %xmm2
+; SSE-64-NEXT:    punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
+; SSE-64-NEXT:    pand %xmm4, %xmm0
+; SSE-64-NEXT:    pandn %xmm1, %xmm4
+; SSE-64-NEXT:    por %xmm4, %xmm0
+; SSE-64-NEXT:    retq
+;
+; AVX-32-LABEL: test_v2f32_oeq_q:
+; AVX-32:       # %bb.0:
+; AVX-32-NEXT:    pushl %ebp
+; AVX-32-NEXT:    movl %esp, %ebp
+; AVX-32-NEXT:    andl $-16, %esp
+; AVX-32-NEXT:    subl $16, %esp
+; AVX-32-NEXT:    vmovaps 8(%ebp), %xmm3
+; AVX-32-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
+; AVX-32-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
+; AVX-32-NEXT:    xorl %eax, %eax
+; AVX-32-NEXT:    vucomiss %xmm4, %xmm5
+; AVX-32-NEXT:    movl $-1, %ecx
+; AVX-32-NEXT:    movl $-1, %edx
+; AVX-32-NEXT:    cmovnel %eax, %edx
+; AVX-32-NEXT:    cmovpl %eax, %edx
+; AVX-32-NEXT:    vucomiss %xmm3, %xmm2
+; AVX-32-NEXT:    cmovnel %eax, %ecx
+; AVX-32-NEXT:    cmovpl %eax, %ecx
+; AVX-32-NEXT:    vmovd %ecx, %xmm2
+; AVX-32-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
+; AVX-32-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX-32-NEXT:    movl %ebp, %esp
+; AVX-32-NEXT:    popl %ebp
+; AVX-32-NEXT:    retl
+;
+; AVX-64-LABEL: test_v2f32_oeq_q:
+; AVX-64:       # %bb.0:
+; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
+; AVX-64-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
+; AVX-64-NEXT:    xorl %eax, %eax
+; AVX-64-NEXT:    vucomiss %xmm4, %xmm5
+; AVX-64-NEXT:    movl $-1, %ecx
+; AVX-64-NEXT:    movl $-1, %edx
+; AVX-64-NEXT:    cmovnel %eax, %edx
+; AVX-64-NEXT:    cmovpl %eax, %edx
+; AVX-64-NEXT:    vucomiss %xmm3, %xmm2
+; AVX-64-NEXT:    cmovnel %eax, %ecx
+; AVX-64-NEXT:    cmovpl %eax, %ecx
+; AVX-64-NEXT:    vmovd %ecx, %xmm2
+; AVX-64-NEXT:    vpinsrd $1, %edx, %xmm2, %xmm2
+; AVX-64-NEXT:    vblendvps %xmm2, %xmm0, %xmm1, %xmm0
+; AVX-64-NEXT:    retq
+;
+; AVX512-32-LABEL: test_v2f32_oeq_q:
+; AVX512-32:       # %bb.0:
+; AVX512-32-NEXT:    pushl %ebp
+; AVX512-32-NEXT:    movl %esp, %ebp
+; AVX512-32-NEXT:    andl $-16, %esp
+; AVX512-32-NEXT:    subl $16, %esp
+; AVX512-32-NEXT:    vmovaps 8(%ebp), %xmm3
+; AVX512-32-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
+; AVX512-32-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
+; AVX512-32-NEXT:    vucomiss %xmm4, %xmm5
+; AVX512-32-NEXT:    setnp %al
+; AVX512-32-NEXT:    sete %cl
+; AVX512-32-NEXT:    testb %al, %cl
+; AVX512-32-NEXT:    setne %al
+; AVX512-32-NEXT:    kmovw %eax, %k0
+; AVX512-32-NEXT:    kshiftlw $15, %k0, %k0
+; AVX512-32-NEXT:    kshiftrw $14, %k0, %k0
+; AVX512-32-NEXT:    vucomiss %xmm3, %xmm2
+; AVX512-32-NEXT:    setnp %al
+; AVX512-32-NEXT:    sete %cl
+; AVX512-32-NEXT:    testb %al, %cl
+; AVX512-32-NEXT:    setne %al
+; AVX512-32-NEXT:    andl $1, %eax
+; AVX512-32-NEXT:    kmovw %eax, %k1
+; AVX512-32-NEXT:    movw $-3, %ax
+; AVX512-32-NEXT:    kmovw %eax, %k2
+; AVX512-32-NEXT:    kandw %k2, %k1, %k1
+; AVX512-32-NEXT:    korw %k0, %k1, %k1
+; AVX512-32-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
+; AVX512-32-NEXT:    movl %ebp, %esp
+; AVX512-32-NEXT:    popl %ebp
+; AVX512-32-NEXT:    retl
+;
+; AVX512-64-LABEL: test_v2f32_oeq_q:
+; AVX512-64:       # %bb.0:
+; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm4 = xmm3[1,1,3,3]
+; AVX512-64-NEXT:    vmovshdup {{.*#+}} xmm5 = xmm2[1,1,3,3]
+; AVX512-64-NEXT:    vucomiss %xmm4, %xmm5
+; AVX512-64-NEXT:    setnp %al
+; AVX512-64-NEXT:    sete %cl
+; AVX512-64-NEXT:    testb %al, %cl
+; AVX512-64-NEXT:    setne %al
+; AVX512-64-NEXT:    kmovw %eax, %k0
+; AVX512-64-NEXT:    kshiftlw $15, %k0, %k0
+; AVX512-64-NEXT:    kshiftrw $14, %k0, %k0
+; AVX512-64-NEXT:    vucomiss %xmm3, %xmm2
+; AVX512-64-NEXT:    setnp %al
+; AVX512-64-NEXT:    sete %cl
+; AVX512-64-NEXT:    testb %al, %cl
+; AVX512-64-NEXT:    setne %al
+; AVX512-64-NEXT:    andl $1, %eax
+; AVX512-64-NEXT:    kmovw %eax, %k1
+; AVX512-64-NEXT:    movw $-3, %ax
+; AVX512-64-NEXT:    kmovw %eax, %k2
+; AVX512-64-NEXT:    kandw %k2, %k1, %k1
+; AVX512-64-NEXT:    korw %k0, %k1, %k1
+; AVX512-64-NEXT:    vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
+; AVX512-64-NEXT:    retq
+  %cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f32(
+                                               <2 x float> %f1, <2 x float> %f2, metadata !"oeq",
+                                               metadata !"fpexcept.strict") #0
+  %res = select <2 x i1> %cond, <2 x i32> %a, <2 x i32> %b
+  ret <2 x i32> %res
+}
+
+attributes #0 = { strictfp nounwind }
+
+declare <2 x i1> @llvm.experimental.constrained.fcmp.v2f32(<2 x float>, <2 x float>, metadata, metadata)
+declare <2 x i1> @llvm.experimental.constrained.fcmps.v2f32(<2 x float>, <2 x float>, metadata, metadata)