[DAGCombiner] Don't reassociate operations that have the vector reduction flag set.

[craig.topper]

Summary

When nodes are reassociated the vector-reduction flag gets lost.

The test case is here is what would happen if you had a sum of absolute differences loop that started with a non-zero but contant sum and that loop was unrolled. The vectorizer will generate a constant vector for the initial value. And DAGCombiner reassociate tries to move it down the addition tree erasing the vector-reduction flag. Interestingly this moves constants the opposite direction of the reassociate IR pass.

I've chosen to just punt on the reassociate, but I suppose we could maybe preserve the flag if both nodes have it set.

[spatel]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

953–954

Not sure if this is possible: what if the root node and N1 do not have the reduction flag set? In that case, we want to allow the 'if' clause at line 975 to fire even though N0 has the reduction flag set rather than bailing out?

test/CodeGen/X86/sad.ll

1401

Please commit the test with baseline checks as a preliminary step.

[craig.topper]

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

953–954

Oops. I'll fix that.

[craig.topper]

Rebase and fix review comment

[spatel]

LGTM

lib/CodeGen/SelectionDAG/DAGCombiner.cpp

private:
     SDValue visitFP16_TO_FP(SDNode *N);
 
     SDValue visitFADDForFMACombine(SDNode *N);
     SDValue visitFSUBForFMACombine(SDNode *N);
     SDValue visitFMULForFMADistributiveCombine(SDNode *N);
 
     SDValue XformToShuffleWithZero(SDNode *N);
     SDValue ReassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0,
-                           SDValue N1);
+                           SDValue N1, SDNodeFlags Flags);
 
     SDValue visitShiftByConstant(SDNode *N, ConstantSDNode *Amt);
 
     SDValue foldSelectOfConstants(SDNode *N);
     SDValue foldVSelectOfConstants(SDNode *N);
     SDValue foldBinOpIntoSelect(SDNode *BO);
     bool SimplifySelectOps(SDNode *SELECT, SDValue LHS, SDValue RHS);
     SDValue SimplifyBinOpWithSameOpcodeHands(SDNode *N);

 // Determines if a BUILD_VECTOR is composed of all-constants possibly mixed with
 // undef's.
 static bool isAnyConstantBuildVector(const SDNode *N) {
   return ISD::isBuildVectorOfConstantSDNodes(N) ||
          ISD::isBuildVectorOfConstantFPSDNodes(N);
 }
 
 SDValue DAGCombiner::ReassociateOps(unsigned Opc, const SDLoc &DL, SDValue N0,
-                                    SDValue N1) {
+                                    SDValue N1, SDNodeFlags Flags) {
+  // Don't reassociate reductions.
+  if (Flags.hasVectorReduction())
+    return SDValue();
+
   EVT VT = N0.getValueType();
-  if (N0.getOpcode() == Opc) {
+  if (N0.getOpcode() == Opc && !N0->getFlags().hasVectorReduction()) {
     if (SDNode *L = DAG.isConstantIntBuildVectorOrConstantInt(N0.getOperand(1))) {
       if (SDNode *R = DAG.isConstantIntBuildVectorOrConstantInt(N1)) {
         // reassoc. (op (op x, c1), c2) -> (op x, (op c1, c2))
         if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, L, R))
           return DAG.getNode(Opc, DL, VT, N0.getOperand(0), OpNode);
         return SDValue();
       }
       if (N0.hasOneUse()) {
         // reassoc. (op (op x, c1), y) -> (op (op x, y), c1) iff x+c1 has one
         // use
         SDValue OpNode = DAG.getNode(Opc, SDLoc(N0), VT, N0.getOperand(0), N1);
         if (!OpNode.getNode())
           return SDValue();
         AddToWorklist(OpNode.getNode());
         return DAG.getNode(Opc, DL, VT, OpNode, N0.getOperand(1));
       }
     }
   }
 
-  if (N1.getOpcode() == Opc) {
+  if (N1.getOpcode() == Opc && !N1->getFlags().hasVectorReduction()) {
     if (SDNode *R = DAG.isConstantIntBuildVectorOrConstantInt(N1.getOperand(1))) {
       if (SDNode *L = DAG.isConstantIntBuildVectorOrConstantInt(N0)) {
         // reassoc. (op c2, (op x, c1)) -> (op x, (op c1, c2))
         if (SDValue OpNode = DAG.FoldConstantArithmetic(Opc, DL, VT, R, L))
           return DAG.getNode(Opc, DL, VT, N1.getOperand(0), OpNode);
         return SDValue();
       }
       if (N1.hasOneUse()) {
if (N0.getOpcode() == ISD::OR &&
       return DAG.getNode(ISD::ADD, DL, VT, N0.getOperand(0), Add0);
     }
   }
 
   if (SDValue NewSel = foldBinOpIntoSelect(N))
     return NewSel;
 
   // reassociate add
-  if (SDValue RADD = ReassociateOps(ISD::ADD, DL, N0, N1))
+  if (SDValue RADD = ReassociateOps(ISD::ADD, DL, N0, N1, N->getFlags()))
     return RADD;
 
   // fold ((0-A) + B) -> B-A
   if (N0.getOpcode() == ISD::SUB &&
       isNullConstantOrNullSplatConstant(N0.getOperand(0)))
     return DAG.getNode(ISD::SUB, DL, VT, N1, N0.getOperand(1));
 
   // fold (A + (0-B)) -> A-B
if (DAG.isConstantIntBuildVectorOrConstantInt(N1) &&
       isMulAddWithConstProfitable(N, N0, N1))
       return DAG.getNode(ISD::ADD, SDLoc(N), VT,
                          DAG.getNode(ISD::MUL, SDLoc(N0), VT,
                                      N0.getOperand(0), N1),
                          DAG.getNode(ISD::MUL, SDLoc(N1), VT,
                                      N0.getOperand(1), N1));
 
   // reassociate mul
-  if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1))
+  if (SDValue RMUL = ReassociateOps(ISD::MUL, SDLoc(N), N0, N1, N->getFlags()))
     return RMUL;
 
   return SDValue();
 }
 
 /// Return true if divmod libcall is available.
 static bool isDivRemLibcallAvailable(SDNode *Node, bool isSigned,
                                      const TargetLowering &TLI) {
SDValue DAGCombiner::visitAND(SDNode *N) {
   if (N1C && DAG.MaskedValueIsZero(SDValue(N, 0),
                                    APInt::getAllOnesValue(BitWidth)))
     return DAG.getConstant(0, SDLoc(N), VT);
 
   if (SDValue NewSel = foldBinOpIntoSelect(N))
     return NewSel;
 
   // reassociate and
-  if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1))
+  if (SDValue RAND = ReassociateOps(ISD::AND, SDLoc(N), N0, N1, N->getFlags()))
     return RAND;
 
   // Try to convert a constant mask AND into a shuffle clear mask.
   if (VT.isVector())
     if (SDValue Shuffle = XformToShuffleWithZero(N))
       return Shuffle;
 
   // fold (and (or x, C), D) -> D if (C & D) == D
SDValue DAGCombiner::visitOR(SDNode *N) {
 
   // Recognize halfword bswaps as (bswap + rotl 16) or (bswap + shl 16)
   if (SDValue BSwap = MatchBSwapHWord(N, N0, N1))
     return BSwap;
   if (SDValue BSwap = MatchBSwapHWordLow(N, N0, N1))
     return BSwap;
 
   // reassociate or
-  if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1))
+  if (SDValue ROR = ReassociateOps(ISD::OR, SDLoc(N), N0, N1, N->getFlags()))
     return ROR;
 
   // Canonicalize (or (and X, c1), c2) -> (and (or X, c2), c1|c2)
   // iff (c1 & c2) != 0.
   auto MatchIntersect = [](ConstantSDNode *LHS, ConstantSDNode *RHS) {
     return LHS->getAPIntValue().intersects(RHS->getAPIntValue());
   };
   if (N0.getOpcode() == ISD::AND && N0.getNode()->hasOneUse() &&
SDValue DAGCombiner::visitXOR(SDNode *N) {
   // fold (xor x, 0) -> x
   if (isNullConstant(N1))
     return N0;
 
   if (SDValue NewSel = foldBinOpIntoSelect(N))
     return NewSel;
 
   // reassociate xor
-  if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1))
+  if (SDValue RXOR = ReassociateOps(ISD::XOR, SDLoc(N), N0, N1, N->getFlags()))
     return RXOR;
 
   // fold !(x cc y) -> (x !cc y)
   SDValue LHS, RHS, CC;
   if (TLI.isConstTrueVal(N1.getNode()) && isSetCCEquivalent(N0, LHS, RHS, CC)) {
     bool isInt = LHS.getValueType().isInteger();
     ISD::CondCode NotCC = ISD::getSetCCInverse(cast<CondCodeSDNode>(CC)->get(),
                                                isInt);

test/CodeGen/X86/sad.ll

; AVX512-NEXT:    retq
   %h3 = shufflevector <32 x i32> %sum2, <32 x i32> undef, <32 x i32> <i32 1, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef, i32 undef>
   %sum3 = add <32 x i32> %sum2, %h3
   %sum = extractelement <32 x i32> %sum3, i32 0
   ret i32 %sum
 }
 
 ; This contains an unrolled sad loop with a non-zero initial value.
 ; DAGCombiner reassociation previously rewrote the adds to move the constant vector further down the tree. This resulted in the vector-reduction flag being lost.
 define i32 @sad_unroll_nonzero_initial(<16 x i8>* %arg, <16 x i8>* %arg1, <16 x i8>* %arg2, <16 x i8>* %arg3) {
 ; SSE2-LABEL: sad_unroll_nonzero_initial:
 ; SSE2:       # %bb.0: # %bb
-; SSE2-NEXT:    movdqu (%rdi), %xmm2
-; SSE2-NEXT:    movdqu (%rsi), %xmm0
-; SSE2-NEXT:    pxor %xmm11, %xmm11
-; SSE2-NEXT:    movdqa %xmm2, %xmm12
-; SSE2-NEXT:    punpcklbw {{.*#+}} xmm12 = xmm12[0],xmm11[0],xmm12[1],xmm11[1],xmm12[2],xmm11[2],xmm12[3],xmm11[3],xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; SSE2-NEXT:    movdqa %xmm12, %xmm9
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm9 = xmm9[0],xmm11[0],xmm9[1],xmm11[1],xmm9[2],xmm11[2],xmm9[3],xmm11[3]
-; SSE2-NEXT:    punpckhbw {{.*#+}} xmm2 = xmm2[8],xmm11[8],xmm2[9],xmm11[9],xmm2[10],xmm11[10],xmm2[11],xmm11[11],xmm2[12],xmm11[12],xmm2[13],xmm11[13],xmm2[14],xmm11[14],xmm2[15],xmm11[15]
-; SSE2-NEXT:    movdqa %xmm2, %xmm10
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm10 = xmm10[0],xmm11[0],xmm10[1],xmm11[1],xmm10[2],xmm11[2],xmm10[3],xmm11[3]
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm12 = xmm12[4],xmm11[4],xmm12[5],xmm11[5],xmm12[6],xmm11[6],xmm12[7],xmm11[7]
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm2 = xmm2[4],xmm11[4],xmm2[5],xmm11[5],xmm2[6],xmm11[6],xmm2[7],xmm11[7]
-; SSE2-NEXT:    movdqa %xmm0, %xmm5
-; SSE2-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3],xmm5[4],xmm11[4],xmm5[5],xmm11[5],xmm5[6],xmm11[6],xmm5[7],xmm11[7]
-; SSE2-NEXT:    movdqa %xmm5, %xmm6
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3]
-; SSE2-NEXT:    psubd %xmm6, %xmm9
-; SSE2-NEXT:    punpckhbw {{.*#+}} xmm0 = xmm0[8],xmm11[8],xmm0[9],xmm11[9],xmm0[10],xmm11[10],xmm0[11],xmm11[11],xmm0[12],xmm11[12],xmm0[13],xmm11[13],xmm0[14],xmm11[14],xmm0[15],xmm11[15]
-; SSE2-NEXT:    movdqa %xmm0, %xmm6
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3]
-; SSE2-NEXT:    psubd %xmm6, %xmm10
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm5 = xmm5[4],xmm11[4],xmm5[5],xmm11[5],xmm5[6],xmm11[6],xmm5[7],xmm11[7]
-; SSE2-NEXT:    psubd %xmm5, %xmm12
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm11[4],xmm0[5],xmm11[5],xmm0[6],xmm11[6],xmm0[7],xmm11[7]
-; SSE2-NEXT:    psubd %xmm0, %xmm2
-; SSE2-NEXT:    movdqa %xmm9, %xmm0
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    paddd %xmm0, %xmm9
-; SSE2-NEXT:    pxor %xmm0, %xmm9
-; SSE2-NEXT:    movdqa %xmm10, %xmm0
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    paddd %xmm0, %xmm10
-; SSE2-NEXT:    pxor %xmm0, %xmm10
-; SSE2-NEXT:    movdqa %xmm12, %xmm0
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    paddd %xmm0, %xmm12
-; SSE2-NEXT:    pxor %xmm0, %xmm12
-; SSE2-NEXT:    movdqa %xmm2, %xmm0
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    paddd %xmm0, %xmm2
-; SSE2-NEXT:    pxor %xmm0, %xmm2
+; SSE2-NEXT:    movdqu (%rdi), %xmm0
+; SSE2-NEXT:    movdqu (%rsi), %xmm1
+; SSE2-NEXT:    psadbw %xmm0, %xmm1
 ; SSE2-NEXT:    movl $1, %eax
-; SSE2-NEXT:    movd %eax, %xmm8
-; SSE2-NEXT:    movdqu (%rdx), %xmm5
-; SSE2-NEXT:    movdqu (%rcx), %xmm3
-; SSE2-NEXT:    movdqa %xmm5, %xmm6
-; SSE2-NEXT:    punpckhbw {{.*#+}} xmm6 = xmm6[8],xmm11[8],xmm6[9],xmm11[9],xmm6[10],xmm11[10],xmm6[11],xmm11[11],xmm6[12],xmm11[12],xmm6[13],xmm11[13],xmm6[14],xmm11[14],xmm6[15],xmm11[15]
-; SSE2-NEXT:    movdqa %xmm6, %xmm0
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm0 = xmm0[4],xmm11[4],xmm0[5],xmm11[5],xmm0[6],xmm11[6],xmm0[7],xmm11[7]
-; SSE2-NEXT:    punpcklbw {{.*#+}} xmm5 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3],xmm5[4],xmm11[4],xmm5[5],xmm11[5],xmm5[6],xmm11[6],xmm5[7],xmm11[7]
-; SSE2-NEXT:    movdqa %xmm5, %xmm7
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm7 = xmm7[4],xmm11[4],xmm7[5],xmm11[5],xmm7[6],xmm11[6],xmm7[7],xmm11[7]
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm6 = xmm6[0],xmm11[0],xmm6[1],xmm11[1],xmm6[2],xmm11[2],xmm6[3],xmm11[3]
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm5 = xmm5[0],xmm11[0],xmm5[1],xmm11[1],xmm5[2],xmm11[2],xmm5[3],xmm11[3]
-; SSE2-NEXT:    movdqa %xmm3, %xmm4
-; SSE2-NEXT:    punpckhbw {{.*#+}} xmm4 = xmm4[8],xmm11[8],xmm4[9],xmm11[9],xmm4[10],xmm11[10],xmm4[11],xmm11[11],xmm4[12],xmm11[12],xmm4[13],xmm11[13],xmm4[14],xmm11[14],xmm4[15],xmm11[15]
-; SSE2-NEXT:    movdqa %xmm4, %xmm1
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm11[4],xmm1[5],xmm11[5],xmm1[6],xmm11[6],xmm1[7],xmm11[7]
-; SSE2-NEXT:    psubd %xmm1, %xmm0
-; SSE2-NEXT:    punpcklbw {{.*#+}} xmm3 = xmm3[0],xmm11[0],xmm3[1],xmm11[1],xmm3[2],xmm11[2],xmm3[3],xmm11[3],xmm3[4],xmm11[4],xmm3[5],xmm11[5],xmm3[6],xmm11[6],xmm3[7],xmm11[7]
-; SSE2-NEXT:    movdqa %xmm3, %xmm1
-; SSE2-NEXT:    punpckhwd {{.*#+}} xmm1 = xmm1[4],xmm11[4],xmm1[5],xmm11[5],xmm1[6],xmm11[6],xmm1[7],xmm11[7]
-; SSE2-NEXT:    psubd %xmm1, %xmm7
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm4 = xmm4[0],xmm11[0],xmm4[1],xmm11[1],xmm4[2],xmm11[2],xmm4[3],xmm11[3]
-; SSE2-NEXT:    psubd %xmm4, %xmm6
-; SSE2-NEXT:    punpcklwd {{.*#+}} xmm3 = xmm3[0],xmm11[0],xmm3[1],xmm11[1],xmm3[2],xmm11[2],xmm3[3],xmm11[3]
-; SSE2-NEXT:    psubd %xmm3, %xmm5
-; SSE2-NEXT:    movdqa %xmm0, %xmm1
-; SSE2-NEXT:    psrad $31, %xmm1
+; SSE2-NEXT:    movd %eax, %xmm0
 ; SSE2-NEXT:    paddd %xmm1, %xmm0
-; SSE2-NEXT:    pxor %xmm1, %xmm0
+; SSE2-NEXT:    movdqu (%rdx), %xmm1
+; SSE2-NEXT:    movdqu (%rcx), %xmm2
+; SSE2-NEXT:    psadbw %xmm1, %xmm2
+; SSE2-NEXT:    paddd %xmm0, %xmm2
+; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm2[2,3,0,1]
 ; SSE2-NEXT:    paddd %xmm2, %xmm0
-; SSE2-NEXT:    paddd %xmm12, %xmm0
-; SSE2-NEXT:    movdqa %xmm7, %xmm1
-; SSE2-NEXT:    psrad $31, %xmm1
-; SSE2-NEXT:    paddd %xmm1, %xmm7
-; SSE2-NEXT:    pxor %xmm1, %xmm7
-; SSE2-NEXT:    paddd %xmm0, %xmm7
-; SSE2-NEXT:    movdqa %xmm6, %xmm0
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    paddd %xmm0, %xmm6
-; SSE2-NEXT:    pxor %xmm0, %xmm6
-; SSE2-NEXT:    paddd %xmm10, %xmm6
-; SSE2-NEXT:    paddd %xmm9, %xmm6
-; SSE2-NEXT:    paddd %xmm7, %xmm6
-; SSE2-NEXT:    paddd %xmm8, %xmm6
-; SSE2-NEXT:    movdqa %xmm5, %xmm0
-; SSE2-NEXT:    psrad $31, %xmm0
-; SSE2-NEXT:    paddd %xmm0, %xmm5
-; SSE2-NEXT:    pxor %xmm0, %xmm5
-; SSE2-NEXT:    paddd %xmm6, %xmm5
-; SSE2-NEXT:    pshufd {{.*#+}} xmm0 = xmm5[2,3,0,1]
-; SSE2-NEXT:    paddd %xmm5, %xmm0
 ; SSE2-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]
 ; SSE2-NEXT:    paddd %xmm0, %xmm1
 ; SSE2-NEXT:    movd %xmm1, %eax
 ; SSE2-NEXT:    retq
 ;
 ; AVX1-LABEL: sad_unroll_nonzero_initial:
 ; AVX1:       # %bb.0: # %bb
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm3 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm4, %xmm0, %xmm0
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm4, %xmm1, %xmm1
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm4, %xmm2, %xmm2
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm4, %xmm3, %xmm3
-; AVX1-NEXT:    vpabsd %xmm0, %xmm8
-; AVX1-NEXT:    vpabsd %xmm1, %xmm1
-; AVX1-NEXT:    vpabsd %xmm2, %xmm2
-; AVX1-NEXT:    vpabsd %xmm3, %xmm3
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm5 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm6 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm7 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm0, %xmm4, %xmm0
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm4, %xmm5, %xmm4
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm5 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm5, %xmm6, %xmm5
-; AVX1-NEXT:    vpmovzxbd {{.*#+}} xmm6 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero
-; AVX1-NEXT:    vpsubd %xmm6, %xmm7, %xmm6
-; AVX1-NEXT:    vpabsd %xmm0, %xmm0
-; AVX1-NEXT:    vpaddd %xmm3, %xmm0, %xmm0
-; AVX1-NEXT:    vpaddd %xmm0, %xmm2, %xmm0
-; AVX1-NEXT:    vpabsd %xmm4, %xmm2
-; AVX1-NEXT:    vpaddd %xmm0, %xmm2, %xmm0
-; AVX1-NEXT:    vpabsd %xmm5, %xmm2
-; AVX1-NEXT:    vpaddd %xmm1, %xmm2, %xmm1
-; AVX1-NEXT:    vpaddd %xmm1, %xmm8, %xmm1
-; AVX1-NEXT:    vpabsd %xmm6, %xmm2
+; AVX1-NEXT:    vmovdqu (%rdi), %xmm0
+; AVX1-NEXT:    vpsadbw (%rsi), %xmm0, %xmm0
+; AVX1-NEXT:    vmovdqu (%rdx), %xmm1
+; AVX1-NEXT:    vpsadbw (%rcx), %xmm1, %xmm1
 ; AVX1-NEXT:    movl $1, %eax
-; AVX1-NEXT:    vmovd %eax, %xmm3
-; AVX1-NEXT:    vpaddd %xmm3, %xmm1, %xmm1
+; AVX1-NEXT:    vmovd %eax, %xmm2
+; AVX1-NEXT:    vpaddd %xmm2, %xmm0, %xmm0
 ; AVX1-NEXT:    vpaddd %xmm0, %xmm1, %xmm0
-; AVX1-NEXT:    vpaddd %xmm0, %xmm2, %xmm0
 ; AVX1-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX1-NEXT:    vpaddd %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vphaddd %xmm0, %xmm0, %xmm0
 ; AVX1-NEXT:    vmovd %xmm0, %eax
 ; AVX1-NEXT:    retq
 ;
 ; AVX2-LABEL: sad_unroll_nonzero_initial:
 ; AVX2:       # %bb.0: # %bb
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpsubd %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpsubd %ymm2, %ymm1, %ymm1
-; AVX2-NEXT:    vpabsd %ymm0, %ymm0
-; AVX2-NEXT:    vpabsd %ymm1, %ymm1
+; AVX2-NEXT:    vmovdqu (%rdi), %xmm0
+; AVX2-NEXT:    vpsadbw (%rsi), %xmm0, %xmm0
 ; AVX2-NEXT:    movl $1, %eax
-; AVX2-NEXT:    vmovd %eax, %xmm2
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm3 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm4 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm5 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpsubd %ymm5, %ymm3, %ymm3
-; AVX2-NEXT:    vpmovzxbd {{.*#+}} ymm5 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero
-; AVX2-NEXT:    vpsubd %ymm5, %ymm4, %ymm4
-; AVX2-NEXT:    vpabsd %ymm3, %ymm3
-; AVX2-NEXT:    vpaddd %ymm1, %ymm3, %ymm1
+; AVX2-NEXT:    vmovd %eax, %xmm1
 ; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm0
-; AVX2-NEXT:    vpaddd %ymm2, %ymm0, %ymm0
-; AVX2-NEXT:    vpabsd %ymm4, %ymm1
+; AVX2-NEXT:    vmovdqu (%rdx), %xmm1
+; AVX2-NEXT:    vpsadbw (%rcx), %xmm1, %xmm1
 ; AVX2-NEXT:    vpaddd %ymm0, %ymm1, %ymm0
 ; AVX2-NEXT:    vextracti128 $1, %ymm0, %xmm1
 ; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX2-NEXT:    vpaddd %ymm1, %ymm0, %ymm0
 ; AVX2-NEXT:    vphaddd %ymm0, %ymm0, %ymm0
 ; AVX2-NEXT:    vmovd %xmm0, %eax
 ; AVX2-NEXT:    vzeroupper
 ; AVX2-NEXT:    retq
 ;
 ; AVX512-LABEL: sad_unroll_nonzero_initial:
 ; AVX512:       # %bb.0: # %bb
-; AVX512-NEXT:    vpmovzxbd {{.*#+}} zmm0 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
-; AVX512-NEXT:    vpmovzxbd {{.*#+}} zmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
-; AVX512-NEXT:    vpsubd %zmm1, %zmm0, %zmm0
-; AVX512-NEXT:    vpabsd %zmm0, %zmm0
+; AVX512-NEXT:    vmovdqu (%rdi), %xmm0
+; AVX512-NEXT:    vpsadbw (%rsi), %xmm0, %xmm0
 ; AVX512-NEXT:    movl $1, %eax
 ; AVX512-NEXT:    vmovd %eax, %xmm1
 ; AVX512-NEXT:    vpaddd %zmm1, %zmm0, %zmm0
-; AVX512-NEXT:    vpmovzxbd {{.*#+}} zmm1 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
-; AVX512-NEXT:    vpmovzxbd {{.*#+}} zmm2 = mem[0],zero,zero,zero,mem[1],zero,zero,zero,mem[2],zero,zero,zero,mem[3],zero,zero,zero,mem[4],zero,zero,zero,mem[5],zero,zero,zero,mem[6],zero,zero,zero,mem[7],zero,zero,zero,mem[8],zero,zero,zero,mem[9],zero,zero,zero,mem[10],zero,zero,zero,mem[11],zero,zero,zero,mem[12],zero,zero,zero,mem[13],zero,zero,zero,mem[14],zero,zero,zero,mem[15],zero,zero,zero
-; AVX512-NEXT:    vpsubd %zmm2, %zmm1, %zmm1
-; AVX512-NEXT:    vpabsd %zmm1, %zmm1
+; AVX512-NEXT:    vmovdqu (%rdx), %xmm1
+; AVX512-NEXT:    vpsadbw (%rcx), %xmm1, %xmm1
 ; AVX512-NEXT:    vpaddd %zmm0, %zmm1, %zmm0
 ; AVX512-NEXT:    vextracti64x4 $1, %zmm0, %ymm1
 ; AVX512-NEXT:    vpaddd %zmm1, %zmm0, %zmm0
 ; AVX512-NEXT:    vextracti128 $1, %ymm0, %xmm1
 ; AVX512-NEXT:    vpaddd %zmm1, %zmm0, %zmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[2,3,0,1]
 ; AVX512-NEXT:    vpaddd %zmm1, %zmm0, %zmm0
 ; AVX512-NEXT:    vpshufd {{.*#+}} xmm1 = xmm0[1,1,2,3]