[X86] Re-enable lowerUINT_TO_FP_vXi32 under fast-math by using an FSUB instead of an FADD.

[craig.topper]

Summary

We previously disabled this under fast math due to aggressive
reassociation by the machine combiner. But I think we can work
around this by using a FSUB instead of FADD for the first
operation.

This matches the similar algorithm we do for uint_to_fp i64->f64
in TargetLowering::expandUINT_TO_FP. If reassociation hasn't
been a problem for that, hopefully its not a problem here.

[spatel]

LGTM - we could do what the existing code comment suggests, but it's more work, and I don't think it's worth the effort.

llvm/lib/Target/X86/X86ISelLowering.cpp

19089–19091

We actually do have FMF on MIR now - see D52006 and related patches.

19137–19139

Better leave a code comment here to explain that this is fragile and requires subtracting the positive constant rather than adding the negative constant.

llvm/lib/Target/X86/X86ISelLowering.cpp

static SDValue lowerUINT_TO_FP_vXi32(SDValue Op, SelectionDAG &DAG,
   //                                 (uint4) 0x53000000, 0xaa);
   // #else
   //     uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
   //     uint4 hi = (v >> 16) | (uint4) 0x53000000;
   // #endif
   //     float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
   //     return (float4) lo + fhi;
 
-  // We shouldn't use it when unsafe-fp-math is enabled though: we might later
-  // reassociate the two FADDs, and if we do that, the algorithm fails
-  // spectacularly (PR24512).
-  // FIXME: If we ever have some kind of Machine FMF, this should be marked
-  // as non-fast and always be enabled. Why isn't SDAG FMF enough? Because
-  // there's also the MachineCombiner reassociations happening on Machine IR.
-  if (DAG.getTarget().Options.UnsafeFPMath)
-    return SDValue();
-
   bool Is128 = VecIntVT == MVT::v4i32;
   MVT VecFloatVT = Is128 ? MVT::v4f32 : MVT::v8f32;
   // If we convert to something else than the supported type, e.g., to v4f64,
   // abort early.
   if (VecFloatVT != Op->getSimpleValueType(0))
     return SDValue();
 
   // In the #idef/#else code, we have in common:
if (Subtarget.hasSSE41()) {
     //     uint4 lo = (v & (uint4) 0xffff) | (uint4) 0x4b000000;
     SDValue LowAnd = DAG.getNode(ISD::AND, DL, VecIntVT, V, VecCstMask);
     Low = DAG.getNode(ISD::OR, DL, VecIntVT, LowAnd, VecCstLow);
 
     //     uint4 hi = (v >> 16) | (uint4) 0x53000000;
     High = DAG.getNode(ISD::OR, DL, VecIntVT, HighShift, VecCstHigh);
   }
 
-  // Create the vector constant for -(0x1.0p39f + 0x1.0p23f).
-  SDValue VecCstFAdd = DAG.getConstantFP(
-      APFloat(APFloat::IEEEsingle(), APInt(32, 0xD3000080)), DL, VecFloatVT);
+  // Create the vector constant for (0x1.0p39f + 0x1.0p23f).
+  SDValue VecCstFSub = DAG.getConstantFP(
+      APFloat(APFloat::IEEEsingle(), APInt(32, 0x53000080)), DL, VecFloatVT);
 
   //     float4 fhi = (float4) hi - (0x1.0p39f + 0x1.0p23f);
+  // NOTE: By using fsub of a positive constant instead of fadd of a negative
+  // constant, we avoid reassociation in MachineCombiner when unsafe-fp-math is
+  // enabled. See PR24512.
   SDValue HighBitcast = DAG.getBitcast(VecFloatVT, High);
   // TODO: Are there any fast-math-flags to propagate here?
   //     (float4) lo;
   SDValue LowBitcast = DAG.getBitcast(VecFloatVT, Low);
   //     return (float4) lo + fhi;
   if (IsStrict) {
-    SDValue FHigh = DAG.getNode(ISD::STRICT_FADD, DL, {VecFloatVT, MVT::Other},
-                                {Op.getOperand(0), HighBitcast, VecCstFAdd});
+    SDValue FHigh = DAG.getNode(ISD::STRICT_FSUB, DL, {VecFloatVT, MVT::Other},
+                                {Op.getOperand(0), HighBitcast, VecCstFSub});
     return DAG.getNode(ISD::STRICT_FADD, DL, {VecFloatVT, MVT::Other},
                        {FHigh.getValue(1), LowBitcast, FHigh});
   }
 
   SDValue FHigh =
-      DAG.getNode(ISD::FADD, DL, VecFloatVT, HighBitcast, VecCstFAdd);
+      DAG.getNode(ISD::FSUB, DL, VecFloatVT, HighBitcast, VecCstFSub);
   return DAG.getNode(ISD::FADD, DL, VecFloatVT, LowBitcast, FHigh);
 }
 
 static SDValue lowerUINT_TO_FP_vec(SDValue Op, SelectionDAG &DAG,
                                    const X86Subtarget &Subtarget) {
   unsigned OpNo = Op.getNode()->isStrictFPOpcode() ? 1 : 0;
   SDValue N0 = Op.getOperand(OpNo);
   MVT SrcVT = N0.getSimpleValueType();

llvm/test/CodeGen/X86/ftrunc.ll

 ; SSE2-NEXT:    andps %xmm1, %xmm3
 ; SSE2-NEXT:    andnps %xmm0, %xmm1
 ; SSE2-NEXT:    orps %xmm3, %xmm1
 ; SSE2-NEXT:    movaps {{.*#+}} xmm0 = [65535,65535,65535,65535]
 ; SSE2-NEXT:    andps %xmm1, %xmm0
 ; SSE2-NEXT:    orps {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    psrld $16, %xmm1
 ; SSE2-NEXT:    por {{.*}}(%rip), %xmm1
-; SSE2-NEXT:    addps {{.*}}(%rip), %xmm1
+; SSE2-NEXT:    subps {{.*}}(%rip), %xmm1
 ; SSE2-NEXT:    addps %xmm0, %xmm1
 ; SSE2-NEXT:    movaps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: trunc_unsigned_v4f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    roundps $11, %xmm0, %xmm0
 ; SSE41-NEXT:    retq

llvm/test/CodeGen/X86/known-bits-vector.ll

 ; X32-LABEL: knownbits_smax_smin_shuffle_uitofp:
 ; X32:       # %bb.0:
 ; X32-NEXT:    vpminsd {{\.LCPI.*}}, %xmm0, %xmm0
 ; X32-NEXT:    vpmaxsd {{\.LCPI.*}}, %xmm0, %xmm0
 ; X32-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,3,3]
 ; X32-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; X32-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; X32-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X32-NEXT:    vaddps {{\.LCPI.*}}, %xmm0, %xmm0
+; X32-NEXT:    vsubps {{\.LCPI.*}}, %xmm0, %xmm0
 ; X32-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: knownbits_smax_smin_shuffle_uitofp:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vpminsd {{.*}}(%rip), %xmm0, %xmm0
 ; X64-NEXT:    vpmaxsd {{.*}}(%rip), %xmm0, %xmm0
 ; X64-NEXT:    vpshufd {{.*#+}} xmm0 = xmm0[0,0,3,3]
 ; X64-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; X64-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; X64-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X64-NEXT:    vaddps {{.*}}(%rip), %xmm0, %xmm0
+; X64-NEXT:    vsubps {{.*}}(%rip), %xmm0, %xmm0
 ; X64-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; X64-NEXT:    retq
   %1 = call <4 x i32> @llvm.x86.sse41.pminsd(<4 x i32> %a0, <4 x i32> <i32 0, i32 -65535, i32 -65535, i32 0>)
   %2 = call <4 x i32> @llvm.x86.sse41.pmaxsd(<4 x i32> %1, <4 x i32> <i32 65535, i32 -1, i32 -1, i32 131071>)
   %3 = shufflevector <4 x i32> %2, <4 x i32> undef, <4 x i32> <i32 0, i32 0, i32 3, i32 3>
   %4 = uitofp <4 x i32> %3 to <4 x float>
   ret <4 x float> %4
 }

 ; If we don't know that the input isn't INT_MIN we can't combine to sitofp
 define <4 x float> @knownbits_abs_uitofp(<4 x i32> %a0) {
 ; X32-LABEL: knownbits_abs_uitofp:
 ; X32:       # %bb.0:
 ; X32-NEXT:    vpabsd %xmm0, %xmm0
 ; X32-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; X32-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; X32-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X32-NEXT:    vaddps {{\.LCPI.*}}, %xmm0, %xmm0
+; X32-NEXT:    vsubps {{\.LCPI.*}}, %xmm0, %xmm0
 ; X32-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; X32-NEXT:    retl
 ;
 ; X64-LABEL: knownbits_abs_uitofp:
 ; X64:       # %bb.0:
 ; X64-NEXT:    vpabsd %xmm0, %xmm0
 ; X64-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; X64-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; X64-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; X64-NEXT:    vaddps {{.*}}(%rip), %xmm0, %xmm0
+; X64-NEXT:    vsubps {{.*}}(%rip), %xmm0, %xmm0
 ; X64-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; X64-NEXT:    retq
   %1 = sub <4 x i32> zeroinitializer, %a0
   %2 = icmp slt <4 x i32> %a0, zeroinitializer
   %3 = select <4 x i1> %2, <4 x i32> %1, <4 x i32> %a0
   %4 = uitofp <4 x i32> %3 to <4 x float>
   ret <4 x float> %4
 }

llvm/test/CodeGen/X86/vec-strict-inttofp-128.ll

 define <4 x float> @uitofp_v4i32_v4f32(<4 x i32> %x) #0 {
 ; SSE-32-LABEL: uitofp_v4i32_v4f32:
 ; SSE-32:       # %bb.0:
 ; SSE-32-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
 ; SSE-32-NEXT:    pand %xmm0, %xmm1
 ; SSE-32-NEXT:    por {{\.LCPI.*}}, %xmm1
 ; SSE-32-NEXT:    psrld $16, %xmm0
 ; SSE-32-NEXT:    por {{\.LCPI.*}}, %xmm0
-; SSE-32-NEXT:    addps {{\.LCPI.*}}, %xmm0
+; SSE-32-NEXT:    subps {{\.LCPI.*}}, %xmm0
 ; SSE-32-NEXT:    addps %xmm1, %xmm0
 ; SSE-32-NEXT:    retl
 ;
 ; SSE-64-LABEL: uitofp_v4i32_v4f32:
 ; SSE-64:       # %bb.0:
 ; SSE-64-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
 ; SSE-64-NEXT:    pand %xmm0, %xmm1
 ; SSE-64-NEXT:    por {{.*}}(%rip), %xmm1
 ; SSE-64-NEXT:    psrld $16, %xmm0
 ; SSE-64-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE-64-NEXT:    addps {{.*}}(%rip), %xmm0
+; SSE-64-NEXT:    subps {{.*}}(%rip), %xmm0
 ; SSE-64-NEXT:    addps %xmm1, %xmm0
 ; SSE-64-NEXT:    retq
 ;
 ; AVX1-32-LABEL: uitofp_v4i32_v4f32:
 ; AVX1-32:       # %bb.0:
 ; AVX1-32-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; AVX1-32-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-32-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; AVX1-32-NEXT:    vaddps {{\.LCPI.*}}, %xmm0, %xmm0
+; AVX1-32-NEXT:    vsubps {{\.LCPI.*}}, %xmm0, %xmm0
 ; AVX1-32-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX1-32-NEXT:    retl
 ;
 ; AVX1-64-LABEL: uitofp_v4i32_v4f32:
 ; AVX1-64:       # %bb.0:
 ; AVX1-64-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; AVX1-64-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-64-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; AVX1-64-NEXT:    vaddps {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-64-NEXT:    vsubps {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-64-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX1-64-NEXT:    retq
 ;
 ; AVX512F-LABEL: uitofp_v4i32_v4f32:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovaps %xmm0, %xmm0
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0

llvm/test/CodeGen/X86/vec-strict-inttofp-256.ll

 ;
 ; AVX2-LABEL: uitofp_v8i32_v8f32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7],ymm0[8],ymm1[9],ymm0[10],ymm1[11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
 ; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
-; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; AVX2-NEXT:    vaddps %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    ret{{[l|q]}}
 ;
 ; AVX512F-LABEL: uitofp_v8i32_v8f32:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovaps %ymm0, %ymm0
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512F-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0

llvm/test/CodeGen/X86/vec_int_to_fp.ll

 define <4 x float> @uitofp_4i32_to_4f32(<4 x i32> %a) {
 ; SSE2-LABEL: uitofp_4i32_to_4f32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
 ; SSE2-NEXT:    pand %xmm0, %xmm1
 ; SSE2-NEXT:    por {{.*}}(%rip), %xmm1
 ; SSE2-NEXT:    psrld $16, %xmm0
 ; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    addps {{.*}}(%rip), %xmm0
+; SSE2-NEXT:    subps {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    addps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: uitofp_4i32_to_4f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [1258291200,1258291200,1258291200,1258291200]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
 ; SSE41-NEXT:    psrld $16, %xmm0
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; SSE41-NEXT:    addps {{.*}}(%rip), %xmm0
+; SSE41-NEXT:    subps {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    addps %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: uitofp_4i32_to_4f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; AVX1-NEXT:    vaddps {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vsubps {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: uitofp_4i32_to_4f32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [1258291200,1258291200,1258291200,1258291200]
 ; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
 ; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [1392508928,1392508928,1392508928,1392508928]
 ; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
-; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; AVX2-NEXT:    vaddps %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: uitofp_4i32_to_4f32:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0

 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; SSE2-NEXT:    pand %xmm2, %xmm3
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [1258291200,1258291200,1258291200,1258291200]
 ; SSE2-NEXT:    por %xmm4, %xmm3
 ; SSE2-NEXT:    psrld $16, %xmm0
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [1392508928,1392508928,1392508928,1392508928]
 ; SSE2-NEXT:    por %xmm5, %xmm0
-; SSE2-NEXT:    movaps {{.*#+}} xmm6 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; SSE2-NEXT:    addps %xmm6, %xmm0
+; SSE2-NEXT:    movaps {{.*#+}} xmm6 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE2-NEXT:    subps %xmm6, %xmm0
 ; SSE2-NEXT:    addps %xmm3, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm2
 ; SSE2-NEXT:    por %xmm4, %xmm2
 ; SSE2-NEXT:    psrld $16, %xmm1
 ; SSE2-NEXT:    por %xmm5, %xmm1
-; SSE2-NEXT:    addps %xmm6, %xmm1
+; SSE2-NEXT:    subps %xmm6, %xmm1
 ; SSE2-NEXT:    addps %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: uitofp_8i32_to_8f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [1258291200,1258291200,1258291200,1258291200]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm3
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1],xmm3[2],xmm2[3],xmm3[4],xmm2[5],xmm3[6],xmm2[7]
 ; SSE41-NEXT:    psrld $16, %xmm0
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [1392508928,1392508928,1392508928,1392508928]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2],xmm4[3],xmm0[4],xmm4[5],xmm0[6],xmm4[7]
-; SSE41-NEXT:    movaps {{.*#+}} xmm5 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; SSE41-NEXT:    addps %xmm5, %xmm0
+; SSE41-NEXT:    movaps {{.*#+}} xmm5 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE41-NEXT:    subps %xmm5, %xmm0
 ; SSE41-NEXT:    addps %xmm3, %xmm0
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4],xmm2[5],xmm1[6],xmm2[7]
 ; SSE41-NEXT:    psrld $16, %xmm1
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2],xmm4[3],xmm1[4],xmm4[5],xmm1[6],xmm4[7]
-; SSE41-NEXT:    addps %xmm5, %xmm1
+; SSE41-NEXT:    subps %xmm5, %xmm1
 ; SSE41-NEXT:    addps %xmm2, %xmm1
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: uitofp_8i32_to_8f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm1
 ; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
 ; AVX1-NEXT:    vpsrld $16, %xmm2, %xmm2
 ; AVX1-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
 ; AVX1-NEXT:    vcvtdq2ps %ymm1, %ymm1
 ; AVX1-NEXT:    vmulps {{.*}}(%rip), %ymm1, %ymm1
 ; AVX1-NEXT:    vandps {{.*}}(%rip), %ymm0, %ymm0
 ; AVX1-NEXT:    vcvtdq2ps %ymm0, %ymm0
 ; AVX1-NEXT:    vaddps %ymm0, %ymm1, %ymm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: uitofp_8i32_to_8f32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7],ymm0[8],ymm1[9],ymm0[10],ymm1[11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
 ; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
-; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; AVX2-NEXT:    vaddps %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: uitofp_8i32_to_8f32:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    # kill: def $ymm0 killed $ymm0 def $zmm0
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512F-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0

 ; SSE2-LABEL: uitofp_load_4i32_to_4f32:
 ; SSE2:       # %bb.0:
 ; SSE2-NEXT:    movdqa (%rdi), %xmm0
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
 ; SSE2-NEXT:    pand %xmm0, %xmm1
 ; SSE2-NEXT:    por {{.*}}(%rip), %xmm1
 ; SSE2-NEXT:    psrld $16, %xmm0
 ; SSE2-NEXT:    por {{.*}}(%rip), %xmm0
-; SSE2-NEXT:    addps {{.*}}(%rip), %xmm0
+; SSE2-NEXT:    subps {{.*}}(%rip), %xmm0
 ; SSE2-NEXT:    addps %xmm1, %xmm0
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: uitofp_load_4i32_to_4f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa (%rdi), %xmm0
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm1 = [1258291200,1258291200,1258291200,1258291200]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
 ; SSE41-NEXT:    psrld $16, %xmm0
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; SSE41-NEXT:    addps {{.*}}(%rip), %xmm0
+; SSE41-NEXT:    subps {{.*}}(%rip), %xmm0
 ; SSE41-NEXT:    addps %xmm1, %xmm0
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: uitofp_load_4i32_to_4f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; AVX1-NEXT:    vaddps {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vsubps {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: uitofp_load_4i32_to_4f32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa (%rdi), %xmm0
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm1 = [1258291200,1258291200,1258291200,1258291200]
 ; AVX2-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
 ; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} xmm2 = [1392508928,1392508928,1392508928,1392508928]
 ; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
-; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; AVX2-NEXT:    vaddps %xmm2, %xmm0, %xmm0
+; AVX2-NEXT:    vbroadcastss {{.*#+}} xmm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %xmm2, %xmm0, %xmm0
 ; AVX2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: uitofp_load_4i32_to_4f32:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovaps (%rdi), %xmm0
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0

 ; SSE2-NEXT:    movdqa {{.*#+}} xmm2 = [65535,65535,65535,65535]
 ; SSE2-NEXT:    movdqa %xmm0, %xmm3
 ; SSE2-NEXT:    pand %xmm2, %xmm3
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [1258291200,1258291200,1258291200,1258291200]
 ; SSE2-NEXT:    por %xmm4, %xmm3
 ; SSE2-NEXT:    psrld $16, %xmm0
 ; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [1392508928,1392508928,1392508928,1392508928]
 ; SSE2-NEXT:    por %xmm5, %xmm0
-; SSE2-NEXT:    movaps {{.*#+}} xmm6 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; SSE2-NEXT:    addps %xmm6, %xmm0
+; SSE2-NEXT:    movaps {{.*#+}} xmm6 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE2-NEXT:    subps %xmm6, %xmm0
 ; SSE2-NEXT:    addps %xmm3, %xmm0
 ; SSE2-NEXT:    pand %xmm1, %xmm2
 ; SSE2-NEXT:    por %xmm4, %xmm2
 ; SSE2-NEXT:    psrld $16, %xmm1
 ; SSE2-NEXT:    por %xmm5, %xmm1
-; SSE2-NEXT:    addps %xmm6, %xmm1
+; SSE2-NEXT:    subps %xmm6, %xmm1
 ; SSE2-NEXT:    addps %xmm2, %xmm1
 ; SSE2-NEXT:    retq
 ;
 ; SSE41-LABEL: uitofp_load_8i32_to_8f32:
 ; SSE41:       # %bb.0:
 ; SSE41-NEXT:    movdqa (%rdi), %xmm0
 ; SSE41-NEXT:    movdqa 16(%rdi), %xmm1
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm2 = [1258291200,1258291200,1258291200,1258291200]
 ; SSE41-NEXT:    movdqa %xmm0, %xmm3
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm3 = xmm3[0],xmm2[1],xmm3[2],xmm2[3],xmm3[4],xmm2[5],xmm3[6],xmm2[7]
 ; SSE41-NEXT:    psrld $16, %xmm0
 ; SSE41-NEXT:    movdqa {{.*#+}} xmm4 = [1392508928,1392508928,1392508928,1392508928]
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2],xmm4[3],xmm0[4],xmm4[5],xmm0[6],xmm4[7]
-; SSE41-NEXT:    movaps {{.*#+}} xmm5 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; SSE41-NEXT:    addps %xmm5, %xmm0
+; SSE41-NEXT:    movaps {{.*#+}} xmm5 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE41-NEXT:    subps %xmm5, %xmm0
 ; SSE41-NEXT:    addps %xmm3, %xmm0
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm2 = xmm1[0],xmm2[1],xmm1[2],xmm2[3],xmm1[4],xmm2[5],xmm1[6],xmm2[7]
 ; SSE41-NEXT:    psrld $16, %xmm1
 ; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2],xmm4[3],xmm1[4],xmm4[5],xmm1[6],xmm4[7]
-; SSE41-NEXT:    addps %xmm5, %xmm1
+; SSE41-NEXT:    subps %xmm5, %xmm1
 ; SSE41-NEXT:    addps %xmm2, %xmm1
 ; SSE41-NEXT:    retq
 ;
 ; AVX1-LABEL: uitofp_load_8i32_to_8f32:
 ; AVX1:       # %bb.0:
 ; AVX1-NEXT:    vmovaps (%rdi), %ymm0
 ; AVX1-NEXT:    vmovdqa (%rdi), %xmm1
 ; AVX1-NEXT:    vmovdqa 16(%rdi), %xmm2

 ; AVX2-LABEL: uitofp_load_8i32_to_8f32:
 ; AVX2:       # %bb.0:
 ; AVX2-NEXT:    vmovdqa (%rdi), %ymm0
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm1 = [1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200,1258291200]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm1 = ymm0[0],ymm1[1],ymm0[2],ymm1[3],ymm0[4],ymm1[5],ymm0[6],ymm1[7],ymm0[8],ymm1[9],ymm0[10],ymm1[11],ymm0[12],ymm1[13],ymm0[14],ymm1[15]
 ; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm0
 ; AVX2-NEXT:    vpbroadcastd {{.*#+}} ymm2 = [1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928,1392508928]
 ; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
-; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; AVX2-NEXT:    vaddps %ymm2, %ymm0, %ymm0
+; AVX2-NEXT:    vbroadcastss {{.*#+}} ymm2 = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; AVX2-NEXT:    vsubps %ymm2, %ymm0, %ymm0
 ; AVX2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    retq
 ;
 ; AVX512F-LABEL: uitofp_load_8i32_to_8f32:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    vmovaps (%rdi), %ymm0
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512F-NEXT:    # kill: def $ymm0 killed $ymm0 killed $zmm0

llvm/test/CodeGen/X86/vec_uint_to_fp-fastmath.ll

 ; RUN: llc < %s -mtriple=x86_64 -enable-unsafe-fp-math \
 ; RUN:   | FileCheck %s --check-prefix=CHECK --check-prefix=CST --check-prefix=SSE --check-prefix=SSE2
 ; RUN: llc < %s -mtriple=x86_64 -enable-unsafe-fp-math -mattr=+sse4.1 \
 ; RUN:   | FileCheck %s --check-prefix=CHECK --check-prefix=CST --check-prefix=SSE --check-prefix=SSE41
 ; RUN: llc < %s -mtriple=x86_64 -enable-unsafe-fp-math -mattr=+avx \
 ; RUN:   | FileCheck %s --check-prefix=CHECK --check-prefix=CST --check-prefix=AVX
 ; RUN: llc < %s -mtriple=x86_64 -enable-unsafe-fp-math -mattr=+avx2 \
 ; RUN:   | FileCheck %s --check-prefix=CHECK --check-prefix=AVX2
 ; RUN: llc < %s -mtriple=x86_64 -enable-unsafe-fp-math -mattr=+avx512f \
 ; RUN:   | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512F
 ; RUN: llc < %s -mtriple=x86_64 -enable-unsafe-fp-math -mattr=+avx512vl \
 ; RUN:   | FileCheck %s --check-prefix=CHECK --check-prefix=AVX512VL
 
+; Check that the constant used in the vectors are the right ones.
 ; SSE2: [[MASKCSTADDR:.LCPI[0-9_]+]]:
 ; SSE2-NEXT: .long 65535 # 0xffff
 ; SSE2-NEXT: .long 65535 # 0xffff
 ; SSE2-NEXT: .long 65535 # 0xffff
 ; SSE2-NEXT: .long 65535 # 0xffff
 
-; CST: [[FPMASKCSTADDR:.LCPI[0-9_]+]]:
-; CST-NEXT: .long 1199570944 # float 65536
-; CST-NEXT: .long 1199570944 # float 65536
-; CST-NEXT: .long 1199570944 # float 65536
-; CST-NEXT: .long 1199570944 # float 65536
+; CST: [[LOWCSTADDR:.LCPI[0-9_]+]]:
+; CST-NEXT: .long 1258291200 # 0x4b000000
+; CST-NEXT: .long 1258291200 # 0x4b000000
+; CST-NEXT: .long 1258291200 # 0x4b000000
+; CST-NEXT: .long 1258291200 # 0x4b000000
+
+; CST: [[HIGHCSTADDR:.LCPI[0-9_]+]]:
+; CST-NEXT: .long 1392508928 # 0x53000000
+; CST-NEXT: .long 1392508928 # 0x53000000
+; CST-NEXT: .long 1392508928 # 0x53000000
+; CST-NEXT: .long 1392508928 # 0x53000000
+
+; CST: [[MAGICCSTADDR:.LCPI[0-9_]+]]:
+; CST-NEXT: .long 1392509056 # float 5.49764202E+11
+; CST-NEXT: .long 1392509056 # float 5.49764202E+11
+; CST-NEXT: .long 1392509056 # float 5.49764202E+11
+; CST-NEXT: .long 1392509056 # float 5.49764202E+11
 
-; AVX2: [[FPMASKCSTADDR:.LCPI[0-9_]+]]:
-; AVX2-NEXT: .long 1199570944 # float 65536
+; AVX2: [[LOWCSTADDR:.LCPI[0-9_]+]]:
+; AVX2-NEXT: .long 1258291200 # 0x4b000000
+
+; AVX2: [[HIGHCSTADDR:.LCPI[0-9_]+]]:
+; AVX2-NEXT: .long 1392508928 # 0x53000000
+
+; AVX2: [[MAGICCSTADDR:.LCPI[0-9_]+]]:
+; AVX2-NEXT: .long 1392509056 # float 5.49764202E+11
 
 define <4 x float> @test_uitofp_v4i32_to_v4f32(<4 x i32> %arg) {
 ; SSE2-LABEL: test_uitofp_v4i32_to_v4f32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movaps {{.*#+}} xmm1 = [65535,65535,65535,65535]
-; SSE2-NEXT:    andps %xmm0, %xmm1
-; SSE2-NEXT:    cvtdq2ps %xmm1, %xmm1
-; SSE2-NEXT:    psrld $16, %xmm0
-; SSE2-NEXT:    cvtdq2ps %xmm0, %xmm0
-; SSE2-NEXT:    mulps [[FPMASKCSTADDR]](%rip), %xmm0
-; SSE2-NEXT:    addps %xmm1, %xmm0
-; SSE2-NEXT:    retq
+; SSE2: movdqa [[MASKCSTADDR]](%rip), [[MASK:%xmm[0-9]+]]
+; SSE2-NEXT: pand %xmm0, [[MASK]]
+; After this instruction, MASK will have the value of the low parts
+; of the vector.
+; SSE2-NEXT: por [[LOWCSTADDR]](%rip), [[MASK]]
+; SSE2-NEXT: psrld $16, %xmm0
+; SSE2-NEXT: por [[HIGHCSTADDR]](%rip), %xmm0
+; SSE2-NEXT: subps [[MAGICCSTADDR]](%rip), %xmm0
+; SSE2-NEXT: addps [[MASK]], %xmm0
+; SSE2-NEXT: retq
 ;
+; Currently we commute the arguments of the first blend, but this could be
+; improved to match the lowering of the second blend.
 ; SSE41-LABEL: test_uitofp_v4i32_to_v4f32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    pxor %xmm1, %xmm1
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
-; SSE41-NEXT:    cvtdq2ps %xmm1, %xmm1
-; SSE41-NEXT:    psrld $16, %xmm0
-; SSE41-NEXT:    cvtdq2ps %xmm0, %xmm0
-; SSE41-NEXT:    mulps [[FPMASKCSTADDR]](%rip), %xmm0
-; SSE41-NEXT:    addps %xmm1, %xmm0
-; SSE41-NEXT:    retq
+; SSE41: movdqa [[LOWCSTADDR]](%rip), [[LOWVEC:%xmm[0-9]+]]
+; SSE41-NEXT: pblendw $85, %xmm0, [[LOWVEC]]
+; SSE41-NEXT: psrld $16, %xmm0
+; SSE41-NEXT: pblendw $170, [[HIGHCSTADDR]](%rip), %xmm0
+; SSE41-NEXT: subps [[MAGICCSTADDR]](%rip), %xmm0
+; SSE41-NEXT: addps [[LOWVEC]], %xmm0
+; SSE41-NEXT: retq
 ;
 ; AVX-LABEL: test_uitofp_v4i32_to_v4f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpxor %xmm1, %xmm1, %xmm1
-; AVX-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],xmm1[1],xmm0[2],xmm1[3],xmm0[4],xmm1[5],xmm0[6],xmm1[7]
-; AVX-NEXT:    vcvtdq2ps %xmm1, %xmm1
-; AVX-NEXT:    vpsrld $16, %xmm0, %xmm0
-; AVX-NEXT:    vcvtdq2ps %xmm0, %xmm0
-; AVX-NEXT:    vmulps [[FPMASKCSTADDR]](%rip), %xmm0, %xmm0
-; AVX-NEXT:    vaddps %xmm1, %xmm0, %xmm0
-; AVX-NEXT:    retq
+; AVX: vpblendw $170, [[LOWCSTADDR]](%rip), %xmm0, [[LOWVEC:%xmm[0-9]+]]
+; AVX-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]]
+; AVX-NEXT: vpblendw $170, [[HIGHCSTADDR]](%rip), [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]]
+; AVX-NEXT: vsubps [[MAGICCSTADDR]](%rip), [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
+; AVX-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0
+; AVX-NEXT: retq
 ;
+; The lowering for AVX2 is a bit messy, because we select broadcast
+; instructions, instead of folding the constant loads.
 ; AVX2-LABEL: test_uitofp_v4i32_to_v4f32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX2-NEXT:    vcvtdq2ps %xmm1, %xmm1
-; AVX2-NEXT:    vbroadcastss [[FPMASKCSTADDR]](%rip), %xmm2
-; AVX2-NEXT:    vmulps %xmm2, %xmm1, %xmm1
-; AVX2-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; AVX2-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],xmm2[1],xmm0[2],xmm2[3],xmm0[4],xmm2[5],xmm0[6],xmm2[7]
-; AVX2-NEXT:    vcvtdq2ps %xmm0, %xmm0
-; AVX2-NEXT:    vaddps %xmm0, %xmm1, %xmm0
-; AVX2-NEXT:    retq
+; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%xmm[0-9]+]]
+; AVX2-NEXT: vpblendw $170, [[LOWCST]], %xmm0, [[LOWVEC:%xmm[0-9]+]]
+; AVX2-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]]
+; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%xmm[0-9]+]]
+; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]]
+; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%xmm[0-9]+]]
+; AVX2-NEXT: vsubps [[MAGICCST]], [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
+; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0
+; AVX2-NEXT: retq
 ;
 ; AVX512F-LABEL: test_uitofp_v4i32_to_v4f32:
 ; AVX512F:       # %bb.0:
-; AVX512F-NEXT:    # kill
+; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 def $zmm0
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
-; AVX512F-NEXT:    # kill
+; AVX512F-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0
 ; AVX512F-NEXT:    vzeroupper
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: test_uitofp_v4i32_to_v4f32:
 ; AVX512VL:       # %bb.0:
 ; AVX512VL-NEXT:    vcvtudq2ps %xmm0, %xmm0
 ; AVX512VL-NEXT:    retq
   %tmp = uitofp <4 x i32> %arg to <4 x float>
   ret <4 x float> %tmp
 }
 
-; AVX: [[FPMASKCSTADDR_v8:.LCPI[0-9_]+]]:
-; AVX-NEXT: .long 1199570944 # float 65536
-; AVX-NEXT: .long 1199570944 # float 65536
-; AVX-NEXT: .long 1199570944 # float 65536
-; AVX-NEXT: .long 1199570944 # float 65536
-
-; AVX: [[MASKCSTADDR_v8:.LCPI[0-9_]+]]:
-; AVX-NEXT: .long 65535 # 0xffff
-; AVX-NEXT: .long 65535 # 0xffff
-; AVX-NEXT: .long 65535 # 0xffff
-; AVX-NEXT: .long 65535 # 0xffff
+; Match the AVX2 constants used in the next function
+; AVX2: [[LOWCSTADDR:.LCPI[0-9_]+]]:
+; AVX2-NEXT: .long 1258291200 # 0x4b000000
+
+; AVX2: [[HIGHCSTADDR:.LCPI[0-9_]+]]:
+; AVX2-NEXT: .long 1392508928 # 0x53000000
 
-; AVX2: [[FPMASKCSTADDR_v8:.LCPI[0-9_]+]]:
-; AVX2-NEXT: .long 1199570944 # float 65536
+; AVX2: [[MAGICCSTADDR:.LCPI[0-9_]+]]:
+; AVX2-NEXT: .long 1392509056 # float 5.49764202E+11
 
 define <8 x float> @test_uitofp_v8i32_to_v8f32(<8 x i32> %arg) {
+; Legalization will break the thing is 2 x <4 x i32> on anthing prior AVX.
+; The constant used for in the vector instruction are shared between the
+; two sequences of instructions.
+;
 ; SSE2-LABEL: test_uitofp_v8i32_to_v8f32:
-; SSE2:       # %bb.0:
-; SSE2-NEXT:    movdqa %xmm0, %xmm2
-; SSE2-NEXT:    psrld $16, %xmm2
-; SSE2-NEXT:    cvtdq2ps %xmm2, %xmm2
-; SSE2-NEXT:    movaps {{.*#+}} xmm3 = [6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4]
-; SSE2-NEXT:    mulps %xmm3, %xmm2
-; SSE2-NEXT:    movdqa {{.*#+}} xmm4 = [65535,65535,65535,65535]
-; SSE2-NEXT:    pand %xmm4, %xmm0
-; SSE2-NEXT:    cvtdq2ps %xmm0, %xmm0
-; SSE2-NEXT:    addps %xmm2, %xmm0
-; SSE2-NEXT:    movdqa %xmm1, %xmm2
-; SSE2-NEXT:    psrld $16, %xmm2
-; SSE2-NEXT:    cvtdq2ps %xmm2, %xmm2
-; SSE2-NEXT:    mulps %xmm3, %xmm2
-; SSE2-NEXT:    pand %xmm4, %xmm1
-; SSE2-NEXT:    cvtdq2ps %xmm1, %xmm1
-; SSE2-NEXT:    addps %xmm2, %xmm1
-; SSE2-NEXT:    retq
+; SSE2: movdqa {{.*#+}} [[MASK:xmm[0-9]+]] = [65535,65535,65535,65535]
+; SSE2-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]]
+; SSE2-NEXT: pand %[[MASK]], [[VECLOW]]
+; SSE2-NEXT: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200]
+; SSE2-NEXT: por %[[LOWCST]], [[VECLOW]]
+; SSE2-NEXT: psrld $16, %xmm0
+; SSE2-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928]
+; SSE2-NEXT: por %[[HIGHCST]], %xmm0
+; SSE2-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE2-NEXT: subps %[[MAGICCST]], %xmm0
+; SSE2-NEXT: addps [[VECLOW]], %xmm0
+; MASK is the low vector of the second part after this point.
+; SSE2-NEXT: pand %xmm1, %[[MASK]]
+; SSE2-NEXT: por %[[LOWCST]], %[[MASK]]
+; SSE2-NEXT: psrld $16, %xmm1
+; SSE2-NEXT: por %[[HIGHCST]], %xmm1
+; SSE2-NEXT: subps %[[MAGICCST]], %xmm1
+; SSE2-NEXT: addps %[[MASK]], %xmm1
+; SSE2-NEXT: retq
 ;
 ; SSE41-LABEL: test_uitofp_v8i32_to_v8f32:
-; SSE41:       # %bb.0:
-; SSE41-NEXT:    movdqa %xmm0, %xmm2
-; SSE41-NEXT:    psrld $16, %xmm2
-; SSE41-NEXT:    cvtdq2ps %xmm2, %xmm2
-; SSE41-NEXT:    movaps {{.*#+}} xmm3 = [6.5536E+4,6.5536E+4,6.5536E+4,6.5536E+4]
-; SSE41-NEXT:    mulps %xmm3, %xmm2
-; SSE41-NEXT:    pxor %xmm4, %xmm4
-; SSE41-NEXT:    pblendw {{.*#+}} xmm0 = xmm0[0],xmm4[1],xmm0[2],xmm4[3],xmm0[4],xmm4[5],xmm0[6],xmm4[7]
-; SSE41-NEXT:    cvtdq2ps %xmm0, %xmm0
-; SSE41-NEXT:    addps %xmm2, %xmm0
-; SSE41-NEXT:    movdqa %xmm1, %xmm2
-; SSE41-NEXT:    psrld $16, %xmm2
-; SSE41-NEXT:    cvtdq2ps %xmm2, %xmm2
-; SSE41-NEXT:    mulps %xmm3, %xmm2
-; SSE41-NEXT:    pblendw {{.*#+}} xmm1 = xmm1[0],xmm4[1],xmm1[2],xmm4[3],xmm1[4],xmm4[5],xmm1[6],xmm4[7]
-; SSE41-NEXT:    cvtdq2ps %xmm1, %xmm1
-; SSE41-NEXT:    addps %xmm2, %xmm1
-; SSE41-NEXT:    retq
+; SSE41: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200]
+; SSE41-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]]
+; SSE41-NEXT: pblendw $170, %[[LOWCST]], [[VECLOW]]
+; SSE41-NEXT: psrld $16, %xmm0
+; SSE41-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928]
+; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm0
+; SSE41-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE41-NEXT: subps %[[MAGICCST]], %xmm0
+; SSE41-NEXT: addps [[VECLOW]], %xmm0
+; LOWCST is the low vector of the second part after this point.
+; The operands of the blend are inverted because we reuse xmm1
+; in the next shift.
+; SSE41-NEXT: pblendw $85, %xmm1, %[[LOWCST]]
+; SSE41-NEXT: psrld $16, %xmm1
+; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm1
+; SSE41-NEXT: subps %[[MAGICCST]], %xmm1
+; SSE41-NEXT: addps %[[LOWCST]], %xmm1
+; SSE41-NEXT: retq
 ;
-; AVX-LABEL: test_uitofp_v8i32_to_v8f32:
-; AVX:       # %bb.0:
-; AVX-NEXT:    vpsrld $16, %xmm0, %xmm1
-; AVX-NEXT:    vextractf128 $1, %ymm0, %xmm2
-; AVX-NEXT:    vpsrld $16, %xmm2, %xmm2
-; AVX-NEXT:    vinsertf128 $1, %xmm2, %ymm1, %ymm1
-; AVX-NEXT:    vcvtdq2ps %ymm1, %ymm1
-; AVX-NEXT:    vmulps [[FPMASKCSTADDR_v8]](%rip), %ymm1, %ymm1
-; AVX-NEXT:    vandps [[MASKCSTADDR_v8]](%rip), %ymm0, %ymm0
-; AVX-NEXT:    vcvtdq2ps %ymm0, %ymm0
-; AVX-NEXT:    vaddps %ymm0, %ymm1, %ymm0
-; AVX-NEXT:    retq
+; Test that we are not lowering uinttofp to scalars
+; AVX-NOT: cvtsd2ss
+; AVX: retq
 ;
 ; AVX2-LABEL: test_uitofp_v8i32_to_v8f32:
-; AVX2:       # %bb.0:
-; AVX2-NEXT:    vpsrld $16, %ymm0, %ymm1
-; AVX2-NEXT:    vcvtdq2ps %ymm1, %ymm1
-; AVX2-NEXT:    vbroadcastss [[FPMASKCSTADDR_v8]](%rip), %ymm2
-; AVX2-NEXT:    vmulps %ymm2, %ymm1, %ymm1
-; AVX2-NEXT:    vxorps %xmm2, %xmm2, %xmm2
-; AVX2-NEXT:    vpblendw {{.*#+}} ymm0 = ymm0[0],ymm2[1],ymm0[2],ymm2[3],ymm0[4],ymm2[5],ymm0[6],ymm2[7],ymm0[8],ymm2[9],ymm0[10],ymm2[11],ymm0[12],ymm2[13],ymm0[14],ymm2[15]
-; AVX2-NEXT:    vcvtdq2ps %ymm0, %ymm0
-; AVX2-NEXT:    vaddps %ymm0, %ymm1, %ymm0
-; AVX2-NEXT:    retq
+; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%ymm[0-9]+]]
+; AVX2-NEXT: vpblendw $170, [[LOWCST]], %ymm0, [[LOWVEC:%ymm[0-9]+]]
+; AVX2-NEXT: vpsrld $16, %ymm0, [[SHIFTVEC:%ymm[0-9]+]]
+; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%ymm[0-9]+]]
+; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%ymm[0-9]+]]
+; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%ymm[0-9]+]]
+; AVX2-NEXT: vsubps [[MAGICCST]], [[HIGHVEC]], [[TMP:%ymm[0-9]+]]
+; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %ymm0
+; AVX2-NEXT: retq
 ;
 ; AVX512F-LABEL: test_uitofp_v8i32_to_v8f32:
 ; AVX512F:       # %bb.0:
 ; AVX512F-NEXT:    # kill
 ; AVX512F-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512F-NEXT:    # kill
 ; AVX512F-NEXT:    retq
 ;
 ; AVX512VL-LABEL: test_uitofp_v8i32_to_v8f32:
 ; AVX512VL:       # %bb.0:
 ; AVX512VL-NEXT:    vcvtudq2ps %ymm0, %ymm0
 ; AVX512VL-NEXT:    retq
   %tmp = uitofp <8 x i32> %arg to <8 x float>
   ret <8 x float> %tmp
 }

llvm/test/CodeGen/X86/vec_uint_to_fp.ll

 
 ; CST: [[HIGHCSTADDR:LCPI0_[0-9]+]]:
 ; CST-NEXT: .long 1392508928              ## 0x53000000
 ; CST-NEXT: .long 1392508928              ## 0x53000000
 ; CST-NEXT: .long 1392508928              ## 0x53000000
 ; CST-NEXT: .long 1392508928              ## 0x53000000
 
 ; CST: [[MAGICCSTADDR:LCPI0_[0-9]+]]:
-; CST-NEXT: .long 3539992704              ## float -5.49764202E+11
-; CST-NEXT: .long 3539992704              ## float -5.49764202E+11
-; CST-NEXT: .long 3539992704              ## float -5.49764202E+11
-; CST-NEXT: .long 3539992704              ## float -5.49764202E+11
+; CST-NEXT: .long 1392509056              ## float 5.49764202E+11
+; CST-NEXT: .long 1392509056              ## float 5.49764202E+11
+; CST-NEXT: .long 1392509056              ## float 5.49764202E+11
+; CST-NEXT: .long 1392509056              ## float 5.49764202E+11
 
 ; AVX2: [[LOWCSTADDR:LCPI0_[0-9]+]]:
 ; AVX2-NEXT: .long  1258291200              ## 0x4b000000
 
 ; AVX2: [[HIGHCSTADDR:LCPI0_[0-9]+]]:
 ; AVX2-NEXT: .long  1392508928              ## 0x53000000
 
 ; AVX2: [[MAGICCSTADDR:LCPI0_[0-9]+]]:
-; AVX2-NEXT: .long  3539992704              ## float -5.49764202E+11
+; AVX2-NEXT: .long  1392509056              ## float 5.49764202E+11
 
 define <4 x float> @test1(<4 x i32> %A) nounwind {
 ; CHECK-LABEL: test1:
 ;
 ; SSE: movdqa [[MASKCSTADDR]](%rip), [[MASK:%xmm[0-9]+]]
 ; SSE-NEXT: pand %xmm0, [[MASK]]
 ; After this instruction, MASK will have the value of the low parts
 ; of the vector.
 ; SSE-NEXT: por [[LOWCSTADDR]](%rip), [[MASK]]
 ; SSE-NEXT: psrld $16, %xmm0
 ; SSE-NEXT: por [[HIGHCSTADDR]](%rip), %xmm0
-; SSE-NEXT: addps [[MAGICCSTADDR]](%rip), %xmm0
+; SSE-NEXT: subps [[MAGICCSTADDR]](%rip), %xmm0
 ; SSE-NEXT: addps [[MASK]], %xmm0
 ; SSE-NEXT: retq
 ;
 ; Currently we commute the arguments of the first blend, but this could be
 ; improved to match the lowering of the second blend.
 ; SSE41: movdqa [[LOWCSTADDR]](%rip), [[LOWVEC:%xmm[0-9]+]]
 ; SSE41-NEXT: pblendw $85, %xmm0, [[LOWVEC]]
 ; SSE41-NEXT: psrld $16, %xmm0
 ; SSE41-NEXT: pblendw $170, [[HIGHCSTADDR]](%rip), %xmm0
-; SSE41-NEXT: addps [[MAGICCSTADDR]](%rip), %xmm0
+; SSE41-NEXT: subps [[MAGICCSTADDR]](%rip), %xmm0
 ; SSE41-NEXT: addps [[LOWVEC]], %xmm0
 ; SSE41-NEXT: retq
 ;
 ; AVX: vpblendw $170, [[LOWCSTADDR]](%rip), %xmm0, [[LOWVEC:%xmm[0-9]+]]
 ; AVX-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]]
 ; AVX-NEXT: vpblendw $170, [[HIGHCSTADDR]](%rip), [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]]
-; AVX-NEXT: vaddps [[MAGICCSTADDR]](%rip), [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
+; AVX-NEXT: vsubps [[MAGICCSTADDR]](%rip), [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
 ; AVX-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0
 ; AVX-NEXT: retq
 ;
 ; The lowering for AVX2 is a bit messy, because we select broadcast
 ; instructions, instead of folding the constant loads.
 ; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%xmm[0-9]+]]
 ; AVX2-NEXT: vpblendw $170, [[LOWCST]], %xmm0, [[LOWVEC:%xmm[0-9]+]]
 ; AVX2-NEXT: vpsrld $16, %xmm0, [[SHIFTVEC:%xmm[0-9]+]]
 ; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%xmm[0-9]+]]
 ; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%xmm[0-9]+]]
 ; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%xmm[0-9]+]]
-; AVX2-NEXT: vaddps [[MAGICCST]], [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
+; AVX2-NEXT: vsubps [[MAGICCST]], [[HIGHVEC]], [[TMP:%xmm[0-9]+]]
 ; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %xmm0
 ; AVX2-NEXT: retq
   %C = uitofp <4 x i32> %A to <4 x float>
   ret <4 x float> %C
 }
 
 ; Match the AVX2 constants used in the next function
 ; AVX2: [[LOWCSTADDR:LCPI1_[0-9]+]]:
 ; AVX2-NEXT: .long  1258291200              ## 0x4b000000
 
 ; AVX2: [[HIGHCSTADDR:LCPI1_[0-9]+]]:
 ; AVX2-NEXT: .long  1392508928              ## 0x53000000
 
 ; AVX2: [[MAGICCSTADDR:LCPI1_[0-9]+]]:
-; AVX2-NEXT: .long  3539992704              ## float -5.49764202E+11
+; AVX2-NEXT: .long  1392509056              ## float 5.49764202E+11
 
 define <8 x float> @test2(<8 x i32> %A) nounwind {
 ; CHECK-LABEL: test2:
 ; Legalization will break the thing is 2 x <4 x i32> on anthing prior AVX.
 ; The constant used for in the vector instruction are shared between the
 ; two sequences of instructions.
 ;
 ; SSE: movdqa {{.*#+}} [[MASK:xmm[0-9]+]] = [65535,65535,65535,65535]
 ; SSE-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]]
 ; SSE-NEXT: pand %[[MASK]], [[VECLOW]]
 ; SSE-NEXT: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200]
 ; SSE-NEXT: por %[[LOWCST]], [[VECLOW]]
 ; SSE-NEXT: psrld $16, %xmm0
 ; SSE-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928]
 ; SSE-NEXT: por %[[HIGHCST]], %xmm0
-; SSE-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; SSE-NEXT: addps %[[MAGICCST]], %xmm0
+; SSE-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE-NEXT: subps %[[MAGICCST]], %xmm0
 ; SSE-NEXT: addps [[VECLOW]], %xmm0
 ; MASK is the low vector of the second part after this point.
 ; SSE-NEXT: pand %xmm1, %[[MASK]]
 ; SSE-NEXT: por %[[LOWCST]], %[[MASK]]
 ; SSE-NEXT: psrld $16, %xmm1
 ; SSE-NEXT: por %[[HIGHCST]], %xmm1
-; SSE-NEXT: addps %[[MAGICCST]], %xmm1
+; SSE-NEXT: subps %[[MAGICCST]], %xmm1
 ; SSE-NEXT: addps %[[MASK]], %xmm1
 ; SSE-NEXT: retq
 ;
 ; SSE41: movdqa {{.*#+}} [[LOWCST:xmm[0-9]+]] = [1258291200,1258291200,1258291200,1258291200]
 ; SSE41-NEXT: movdqa %xmm0, [[VECLOW:%xmm[0-9]+]]
 ; SSE41-NEXT: pblendw $170, %[[LOWCST]], [[VECLOW]]
 ; SSE41-NEXT: psrld $16, %xmm0
 ; SSE41-NEXT: movdqa {{.*#+}} [[HIGHCST:xmm[0-9]+]] = [1392508928,1392508928,1392508928,1392508928]
 ; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm0
-; SSE41-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [-5.49764202E+11,-5.49764202E+11,-5.49764202E+11,-5.49764202E+11]
-; SSE41-NEXT: addps %[[MAGICCST]], %xmm0
+; SSE41-NEXT: movaps {{.*#+}} [[MAGICCST:xmm[0-9]+]] = [5.49764202E+11,5.49764202E+11,5.49764202E+11,5.49764202E+11]
+; SSE41-NEXT: subps %[[MAGICCST]], %xmm0
 ; SSE41-NEXT: addps [[VECLOW]], %xmm0
 ; LOWCST is the low vector of the second part after this point.
 ; The operands of the blend are inverted because we reuse xmm1
 ; in the next shift.
 ; SSE41-NEXT: pblendw $85, %xmm1, %[[LOWCST]]
 ; SSE41-NEXT: psrld $16, %xmm1
 ; SSE41-NEXT: pblendw $170, %[[HIGHCST]], %xmm1
-; SSE41-NEXT: addps %[[MAGICCST]], %xmm1
+; SSE41-NEXT: subps %[[MAGICCST]], %xmm1
 ; SSE41-NEXT: addps %[[LOWCST]], %xmm1
 ; SSE41-NEXT: retq
 ;
 ; Test that we are not lowering uinttofp to scalars
 ; AVX-NOT: cvtsd2ss
 ; AVX: retq
 ;
 ; AVX2: vpbroadcastd [[LOWCSTADDR]](%rip), [[LOWCST:%ymm[0-9]+]]
 ; AVX2-NEXT: vpblendw $170, [[LOWCST]], %ymm0, [[LOWVEC:%ymm[0-9]+]]
 ; AVX2-NEXT: vpsrld $16, %ymm0, [[SHIFTVEC:%ymm[0-9]+]]
 ; AVX2-NEXT: vpbroadcastd [[HIGHCSTADDR]](%rip), [[HIGHCST:%ymm[0-9]+]]
 ; AVX2-NEXT: vpblendw $170, [[HIGHCST]], [[SHIFTVEC]], [[HIGHVEC:%ymm[0-9]+]]
 ; AVX2-NEXT: vbroadcastss [[MAGICCSTADDR]](%rip), [[MAGICCST:%ymm[0-9]+]]
-; AVX2-NEXT: vaddps [[MAGICCST]], [[HIGHVEC]], [[TMP:%ymm[0-9]+]]
+; AVX2-NEXT: vsubps [[MAGICCST]], [[HIGHVEC]], [[TMP:%ymm[0-9]+]]
 ; AVX2-NEXT: vaddps [[TMP]], [[LOWVEC]], %ymm0
 ; AVX2-NEXT: retq
   %C = uitofp <8 x i32> %A to <8 x float>
   ret <8 x float> %C
 }
 
 define <4 x double> @test3(<4 x i32> %arg) {
 ; CHECK-LABEL: test3:
 ; This test used to crash because we were custom lowering it as if it was
 ; a conversion between <4 x i32> and <4 x float>.
 ; AVX: vcvtdq2pd
 ; AVX2: vcvtdq2pd
 ; CHECK: retq
   %tmp = uitofp <4 x i32> %arg to <4 x double>
   ret <4 x double> %tmp
 }

llvm/test/CodeGen/X86/vector-constrained-fp-intrinsics.ll

 define <4 x float> @constrained_vector_uitofp_v4f32_v4i32(<4 x i32> %x) #0 {
 ; CHECK-LABEL: constrained_vector_uitofp_v4f32_v4i32:
 ; CHECK:       # %bb.0: # %entry
 ; CHECK-NEXT:    movdqa {{.*#+}} xmm1 = [65535,65535,65535,65535]
 ; CHECK-NEXT:    pand %xmm0, %xmm1
 ; CHECK-NEXT:    por {{.*}}(%rip), %xmm1
 ; CHECK-NEXT:    psrld $16, %xmm0
 ; CHECK-NEXT:    por {{.*}}(%rip), %xmm0
-; CHECK-NEXT:    addps {{.*}}(%rip), %xmm0
+; CHECK-NEXT:    subps {{.*}}(%rip), %xmm0
 ; CHECK-NEXT:    addps %xmm1, %xmm0
 ; CHECK-NEXT:    retq
 ;
 ; AVX1-LABEL: constrained_vector_uitofp_v4f32_v4i32:
 ; AVX1:       # %bb.0: # %entry
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm1 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
 ; AVX1-NEXT:    vpsrld $16, %xmm0, %xmm0
 ; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0],mem[1],xmm0[2],mem[3],xmm0[4],mem[5],xmm0[6],mem[7]
-; AVX1-NEXT:    vaddps {{.*}}(%rip), %xmm0, %xmm0
+; AVX1-NEXT:    vsubps {{.*}}(%rip), %xmm0, %xmm0
 ; AVX1-NEXT:    vaddps %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    retq
 ;
 ; AVX512-LABEL: constrained_vector_uitofp_v4f32_v4i32:
 ; AVX512:       # %bb.0: # %entry
 ; AVX512-NEXT:    vmovaps %xmm0, %xmm0
 ; AVX512-NEXT:    vcvtudq2ps %zmm0, %zmm0
 ; AVX512-NEXT:    # kill: def $xmm0 killed $xmm0 killed $zmm0