[DAG] Add visitABD optimizations

[dmgreen]

Summary

This adds basic a visitABD to optimize ABDS and ABDU nodes, similar to the existing visitAVG method.

The fold I was initially interested in was folding shuffles though the binop. This also:

• Marks ABDS and ABDU as commutative binops (https://alive2.llvm.org/ce/z/oCDogb and https://alive2.llvm.org/ce/z/7zrs86).
• Add constant folding using max(x,y)-min(x,y)
• Canonicalizes constants to the RHS
• Folds abds x, 0 -> abs(x) (https://alive2.llvm.org/ce/z/4ZEibv)
• Folds abdu x, 0 -> x (https://alive2.llvm.org/ce/z/J_rKqx)
• Folds abd x, undef -> 0 (https://alive2.llvm.org/ce/z/NV6Nsv and https://alive2.llvm.org/ce/z/vs92hu).

[RKSimon]

Add reassociation folds as well?

[RKSimon]

LGTM - re-association combines might still be useful though

[dmgreen]

I was just updating it now :)

This adds reassociation, and also moves the lowering of NEON ABD intrinsics from lowering to a combine so that they can happen earlier which makes testing easier.

[RKSimon]

LGTM - cheers :)

llvm/include/llvm/CodeGen/TargetLowering.h

virtual bool isCommutativeBinOp(unsigned Opcode) const {
     case ISD::FMINNUM_IEEE:
     case ISD::FMAXNUM_IEEE:
     case ISD::FMINIMUM:
     case ISD::FMAXIMUM:
     case ISD::AVGFLOORS:
     case ISD::AVGFLOORU:
     case ISD::AVGCEILS:
     case ISD::AVGCEILU:
+    case ISD::ABDS:
+    case ISD::ABDU:
       return true;
     default: return false;
     }
   }
 
   /// Return true if the node is a math/logic binary operator.
   virtual bool isBinOp(unsigned Opcode) const {
     // A commutative binop must be a binop.

llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp

private:
     SDValue visitSDIV(SDNode *N);
     SDValue visitSDIVLike(SDValue N0, SDValue N1, SDNode *N);
     SDValue visitUDIV(SDNode *N);
     SDValue visitUDIVLike(SDValue N0, SDValue N1, SDNode *N);
     SDValue visitREM(SDNode *N);
     SDValue visitMULHU(SDNode *N);
     SDValue visitMULHS(SDNode *N);
     SDValue visitAVG(SDNode *N);
+    SDValue visitABD(SDNode *N);
     SDValue visitSMUL_LOHI(SDNode *N);
     SDValue visitUMUL_LOHI(SDNode *N);
     SDValue visitMULO(SDNode *N);
     SDValue visitIMINMAX(SDNode *N);
     SDValue visitAND(SDNode *N);
     SDValue visitANDLike(SDValue N0, SDValue N1, SDNode *N);
     SDValue visitOR(SDNode *N);
     SDValue visitORLike(SDValue N0, SDValue N1, SDNode *N);
SDValue DAGCombiner::visit(SDNode *N) {
   case ISD::SREM:
   case ISD::UREM:               return visitREM(N);
   case ISD::MULHU:              return visitMULHU(N);
   case ISD::MULHS:              return visitMULHS(N);
   case ISD::AVGFLOORS:
   case ISD::AVGFLOORU:
   case ISD::AVGCEILS:
   case ISD::AVGCEILU:           return visitAVG(N);
+  case ISD::ABDS:
+  case ISD::ABDU:               return visitABD(N);
   case ISD::SMUL_LOHI:          return visitSMUL_LOHI(N);
   case ISD::UMUL_LOHI:          return visitUMUL_LOHI(N);
   case ISD::SMULO:
   case ISD::UMULO:              return visitMULO(N);
   case ISD::SMIN:
   case ISD::SMAX:
   case ISD::UMIN:
   case ISD::UMAX:               return visitIMINMAX(N);
SDValue DAGCombiner::visitAVG(SDNode *N) {
   if (N1.isUndef())
     return N0;
 
   // TODO If we use avg for scalars anywhere, we can add (avgfl x, 0) -> x >> 1
 
   return SDValue();
 }
 
+SDValue DAGCombiner::visitABD(SDNode *N) {
+  unsigned Opcode = N->getOpcode();
+  SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
+  EVT VT = N->getValueType(0);
+  SDLoc DL(N);
+
+  // fold (abd c1, c2)
+  if (SDValue C = DAG.FoldConstantArithmetic(Opcode, DL, VT, {N0, N1}))
+    return C;
+
+  // canonicalize constant to RHS.
+  if (DAG.isConstantIntBuildVectorOrConstantInt(N0) &&
+      !DAG.isConstantIntBuildVectorOrConstantInt(N1))
+    return DAG.getNode(Opcode, DL, N->getVTList(), N1, N0);
+
+  if (VT.isVector()) {
+    if (SDValue FoldedVOp = SimplifyVBinOp(N, DL))
+      return FoldedVOp;
+
+    // fold (abds x, 0) -> abs x
+    // fold (abdu x, 0) -> x
+    if (ISD::isConstantSplatVectorAllZeros(N1.getNode())) {
+      if (Opcode == ISD::ABDS)
+        return DAG.getNode(ISD::ABS, DL, VT, N0);
+      if (Opcode == ISD::ABDU)
+        return N0;
+    }
+  }
+
+  // fold (abd x, undef) -> 0
+  if (N0.isUndef() || N1.isUndef())
+    return DAG.getConstant(0, DL, VT);
+
+  return SDValue();
+}
+
 /// Perform optimizations common to nodes that compute two values. LoOp and HiOp
 /// give the opcodes for the two computations that are being performed. Return
 /// true if a simplification was made.
 SDValue DAGCombiner::SimplifyNodeWithTwoResults(SDNode *N, unsigned LoOp,
                                                 unsigned HiOp) {
   // If the high half is not needed, just compute the low half.
   bool HiExists = N->hasAnyUseOfValue(1);
   if (!HiExists && (!LegalOperations ||

llvm/lib/CodeGen/SelectionDAG/SelectionDAG.cpp

case ISD::AVGCEILS: {
     return (C1Ext + C2Ext + 1).extractBits(C1.getBitWidth(), 1);
   }
   case ISD::AVGCEILU: {
     unsigned FullWidth = C1.getBitWidth() + 1;
     APInt C1Ext = C1.zext(FullWidth);
     APInt C2Ext = C2.zext(FullWidth);
     return (C1Ext + C2Ext + 1).extractBits(C1.getBitWidth(), 1);
   }
+  case ISD::ABDS:
+    return APIntOps::smax(C1, C2) - APIntOps::smin(C1, C2);
+  case ISD::ABDU:
+    return APIntOps::umax(C1, C2) - APIntOps::umin(C1, C2);
   }
   return std::nullopt;
 }
 
 // Handle constant folding with UNDEF.
 // TODO: Handle more cases.
 static std::optional<APInt> FoldValueWithUndef(unsigned Opcode, const APInt &C1,
                                                bool IsUndef1, const APInt &C2,

llvm/test/CodeGen/AArch64/abd-combine.ll

; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> %src1, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_const_lhs(<8 x i16> %src1) {
 ; CHECK-LABEL: abdu_i_const_lhs:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    movi v1.8h, #1
-; CHECK-NEXT:    uabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    uabd v0.8h, v0.8h, v1.8h
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_const_zero(<8 x i16> %src1) {
 ; CHECK-LABEL: abdu_i_const_zero:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    movi v1.2d, #0000000000000000
-; CHECK-NEXT:    uabd v0.8h, v1.8h, v0.8h
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_const_both() {
 ; CHECK-LABEL: abdu_i_const_both:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    movi v0.8h, #1
-; CHECK-NEXT:    movi v1.8h, #3
-; CHECK-NEXT:    uabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    movi v0.8h, #2
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_const_bothhigh() {
 ; CHECK-LABEL: abdu_i_const_bothhigh:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    movi v0.2d, #0xffffffffffffffff
-; CHECK-NEXT:    mvni v1.8h, #1
-; CHECK-NEXT:    uabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    movi v0.8h, #1
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> <i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534>, <8 x i16> <i16 65535, i16 65535, i16 65535, i16 65535, i16 65535, i16 65535, i16 65535, i16 65535>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_const_onehigh() {
 ; CHECK-LABEL: abdu_i_const_onehigh:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov w8, #32766
-; CHECK-NEXT:    movi v0.8h, #1
-; CHECK-NEXT:    dup v1.8h, w8
-; CHECK-NEXT:    uabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    mov w8, #32765
+; CHECK-NEXT:    dup v0.8h, w8
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> <i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766>, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_const_oneneg() {
 ; CHECK-LABEL: abdu_i_const_oneneg:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov w8, #32766
-; CHECK-NEXT:    mvni v1.8h, #1
-; CHECK-NEXT:    dup v0.8h, w8
-; CHECK-NEXT:    uabd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    movi v0.8h, #128, lsl #8
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> <i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766>, <8 x i16> <i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_zero(<8 x i16> %t, <8 x i16> %src1) {
 ; CHECK-LABEL: abdu_i_zero:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    movi v0.2d, #0000000000000000
-; CHECK-NEXT:    uabd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    mov v0.16b, v1.16b
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abdu_i_undef(<8 x i16> %t, <8 x i16> %src1) {
 ; CHECK-LABEL: abdu_i_undef:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    uabd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    movi v0.2d, #0000000000000000
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16> undef, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 
 
 
; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> %src1, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_const_lhs(<8 x i16> %src1) {
 ; CHECK-LABEL: abds_i_const_lhs:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    movi v1.8h, #1
-; CHECK-NEXT:    sabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    sabd v0.8h, v0.8h, v1.8h
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_const_zero(<8 x i16> %src1) {
 ; CHECK-LABEL: abds_i_const_zero:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    movi v1.2d, #0000000000000000
-; CHECK-NEXT:    sabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    abs v0.8h, v0.8h
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_const_both() {
 ; CHECK-LABEL: abds_i_const_both:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    movi v0.8h, #1
-; CHECK-NEXT:    movi v1.8h, #3
-; CHECK-NEXT:    sabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    movi v0.8h, #2
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> <i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3, i16 3>, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_const_bothhigh() {
 ; CHECK-LABEL: abds_i_const_bothhigh:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov w8, #32766
-; CHECK-NEXT:    mvni v1.8h, #128, lsl #8
-; CHECK-NEXT:    dup v0.8h, w8
-; CHECK-NEXT:    sabd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    movi v0.8h, #1
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> <i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766>, <8 x i16> <i16 32767, i16 32767, i16 32767, i16 32767, i16 32767, i16 32767, i16 32767, i16 32767>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_const_onehigh() {
 ; CHECK-LABEL: abds_i_const_onehigh:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov w8, #32766
-; CHECK-NEXT:    movi v0.8h, #1
-; CHECK-NEXT:    dup v1.8h, w8
-; CHECK-NEXT:    sabd v0.8h, v1.8h, v0.8h
+; CHECK-NEXT:    mov w8, #32765
+; CHECK-NEXT:    dup v0.8h, w8
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> <i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766>, <8 x i16> <i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1, i16 1>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_const_oneneg() {
 ; CHECK-LABEL: abds_i_const_oneneg:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    mov w8, #32766
-; CHECK-NEXT:    mvni v1.8h, #1
-; CHECK-NEXT:    dup v0.8h, w8
-; CHECK-NEXT:    sabd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    movi v0.8h, #128, lsl #8
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> <i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766, i16 32766>, <8 x i16> <i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534, i16 65534>)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_zero(<8 x i16> %t, <8 x i16> %src1) {
 ; CHECK-LABEL: abds_i_zero:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    movi v0.2d, #0000000000000000
-; CHECK-NEXT:    sabd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    abs v0.8h, v1.8h
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> <i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0, i16 0>, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 define <8 x i16> @abds_i_undef(<8 x i16> %t, <8 x i16> %src1) {
 ; CHECK-LABEL: abds_i_undef:
 ; CHECK:       // %bb.0:
-; CHECK-NEXT:    sabd v0.8h, v0.8h, v1.8h
+; CHECK-NEXT:    movi v0.2d, #0000000000000000
 ; CHECK-NEXT:    ret
   %result = call <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16> undef, <8 x i16> %src1)
   ret <8 x i16> %result
 }
 
 
 declare <8 x i16> @llvm.aarch64.neon.uabd.v8i16(<8 x i16>, <8 x i16>)
 declare <8 x i16> @llvm.aarch64.neon.sabd.v8i16(<8 x i16>, <8 x i16>)
 declare <8 x i32> @llvm.abs.v8i32(<8 x i32>, i1)

llvm/test/CodeGen/AArch64/arm64-neon-aba-abd.ll

; CHECK-NEXT:    ret
   %abd = call <2 x i32> @llvm.aarch64.neon.sabd.v2i32(<2 x i32> %lhs, <2 x i32> %rhs)
   ret <2 x i32> %abd
 }
 
 define <2 x i32> @test_sabd_v2i32_const() {
 ; CHECK-LABEL: test_sabd_v2i32_const:
 ; CHECK:       // %bb.0:
 ; CHECK-NEXT:    adrp x8, .LCPI19_0
-; CHECK-NEXT:    movi d0, #0x00ffffffff0000
-; CHECK-NEXT:    ldr d1, [x8, :lo12:.LCPI19_0]
-; CHECK-NEXT:    sabd v0.2s, v1.2s, v0.2s
+; CHECK-NEXT:    ldr d0, [x8, :lo12:.LCPI19_0]
 ; CHECK-NEXT:    ret
   %1 = tail call <2 x i32> @llvm.aarch64.neon.sabd.v2i32(
     <2 x i32> <i32 -2147483648, i32 2147450880>,
     <2 x i32> <i32 -65536, i32 65535>)
   ret <2 x i32> %1
 }
 
 define <2 x i32> @test_saba_v2i32(<2 x i32> %lhs, <2 x i32> %rhs) {

llvm/test/CodeGen/Thumb2/mve-vabdus.ll

 
 for.cond.cleanup:                                 ; preds = %vector.body
   ret void
 }
 
 define arm_aapcs_vfpcc <4 x i32> @vabd_v4u32_commutative(<4 x i32> %src1, <4 x i32> %src2) {
 ; CHECK-LABEL: vabd_v4u32_commutative:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    vabd.u32 q2, q1, q0
-; CHECK-NEXT:    vabd.u32 q0, q0, q1
-; CHECK-NEXT:    vadd.i32 q0, q0, q2
+; CHECK-NEXT:    vabd.u32 q0, q1, q0
+; CHECK-NEXT:    vadd.i32 q0, q0, q0
 ; CHECK-NEXT:    bx lr
   %azextsrc1 = zext <4 x i32> %src1 to <4 x i64>
   %azextsrc2 = zext <4 x i32> %src2 to <4 x i64>
   %aadd1 = sub <4 x i64> %azextsrc1, %azextsrc2
   %aadd2 = sub <4 x i64> zeroinitializer, %aadd1
   %ac = icmp sge <4 x i64> %aadd1, zeroinitializer
   %as = select <4 x i1> %ac, <4 x i64> %aadd1, <4 x i64> %aadd2
   %aresult = trunc <4 x i64> %as to <4 x i32>
   %bzextsrc1 = zext <4 x i32> %src2 to <4 x i64>
   %bzextsrc2 = zext <4 x i32> %src1 to <4 x i64>
   %badd1 = sub <4 x i64> %bzextsrc1, %bzextsrc2
   %badd2 = sub <4 x i64> zeroinitializer, %badd1
   %bc = icmp sge <4 x i64> %badd1, zeroinitializer
   %bs = select <4 x i1> %bc, <4 x i64> %badd1, <4 x i64> %badd2
   %bresult = trunc <4 x i64> %bs to <4 x i32>
   %r = add <4 x i32> %aresult, %bresult
   ret <4 x i32> %r
 }
 
 define arm_aapcs_vfpcc <4 x i32> @vabd_v4u32_shuffle(<4 x i32> %src1, <4 x i32> %src2) {
 ; CHECK-LABEL: vabd_v4u32_shuffle:
 ; CHECK:       @ %bb.0:
-; CHECK-NEXT:    vmov.f32 s8, s7
-; CHECK-NEXT:    vmov.f32 s9, s6
-; CHECK-NEXT:    vmov.f32 s10, s5
-; CHECK-NEXT:    vmov.f32 s11, s4
-; CHECK-NEXT:    vmov.f32 s4, s3
-; CHECK-NEXT:    vmov.f32 s5, s2
-; CHECK-NEXT:    vmov.f32 s6, s1
-; CHECK-NEXT:    vmov.f32 s7, s0
-; CHECK-NEXT:    vabd.u32 q0, q1, q2
+; CHECK-NEXT:    vabd.u32 q1, q0, q1
+; CHECK-NEXT:    vmov.f32 s0, s7
+; CHECK-NEXT:    vmov.f32 s1, s6
+; CHECK-NEXT:    vmov.f32 s2, s5
+; CHECK-NEXT:    vmov.f32 s3, s4
 ; CHECK-NEXT:    bx lr
   %s1 = shufflevector <4 x i32> %src1, <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   %s2 = shufflevector <4 x i32> %src2, <4 x i32> undef, <4 x i32> <i32 3, i32 2, i32 1, i32 0>
   %azextsrc1 = zext <4 x i32> %s1 to <4 x i64>
   %azextsrc2 = zext <4 x i32> %s2 to <4 x i64>
   %aadd1 = sub <4 x i64> %azextsrc1, %azextsrc2
   %aadd2 = sub <4 x i64> zeroinitializer, %aadd1
   %ac = icmp sge <4 x i64> %aadd1, zeroinitializer
   %as = select <4 x i1> %ac, <4 x i64> %aadd1, <4 x i64> %aadd2
   %aresult = trunc <4 x i64> %as to <4 x i32>
   ret <4 x i32> %aresult
 }