[X86][SSE] Shuffle combine blends to OR(X,Y) if the relevant elements are known zero.

This allows us to remove the (depth violating) code in getFauxShuffleMask where we were combining the OR(SHUFFLE,SHUFFLE) shuffle inputs as well, and not just the OR().

This is a minor step toward being able to shuffle combine from/to SELECT/BLENDV as a faux shuffle.

5 files changed, 61 insertions, 35 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index e9bb50aacec0..b2bfcc2698f4 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -7401,8 +7401,8 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
 
     // Handle OR(SHUFFLE,SHUFFLE) case where one source is zero and the other
     // is a valid shuffle index.
-    SDValue N0 = peekThroughOneUseBitcasts(N.getOperand(0));
-    SDValue N1 = peekThroughOneUseBitcasts(N.getOperand(1));
+    SDValue N0 = peekThroughBitcasts(N.getOperand(0));
+    SDValue N1 = peekThroughBitcasts(N.getOperand(1));
     if (!N0.getValueType().isVector() || !N1.getValueType().isVector())
       return false;
     SmallVector<int, 64> SrcMask0, SrcMask1;
@@ -7413,34 +7413,24 @@ static bool getFauxShuffleMask(SDValue N, const APInt &DemandedElts,
                                 true))
       return false;
 
-    // Shuffle inputs must be the same size as the result.
-    if (llvm::any_of(SrcInputs0, [VT](SDValue Op) {
-          return VT.getSizeInBits() != Op.getValueSizeInBits();
-        }))
-      return false;
-    if (llvm::any_of(SrcInputs1, [VT](SDValue Op) {
-          return VT.getSizeInBits() != Op.getValueSizeInBits();
-        }))
-      return false;
-
     size_t MaskSize = std::max(SrcMask0.size(), SrcMask1.size());
     SmallVector<int, 64> Mask0, Mask1;
     narrowShuffleMaskElts(MaskSize / SrcMask0.size(), SrcMask0, Mask0);
     narrowShuffleMaskElts(MaskSize / SrcMask1.size(), SrcMask1, Mask1);
-    for (size_t i = 0; i != MaskSize; ++i) {
+    for (int i = 0; i != (int)MaskSize; ++i) {
       if (Mask0[i] == SM_SentinelUndef && Mask1[i] == SM_SentinelUndef)
         Mask.push_back(SM_SentinelUndef);
       else if (Mask0[i] == SM_SentinelZero && Mask1[i] == SM_SentinelZero)
         Mask.push_back(SM_SentinelZero);
       else if (Mask1[i] == SM_SentinelZero)
-        Mask.push_back(Mask0[i]);
+        Mask.push_back(i);
       else if (Mask0[i] == SM_SentinelZero)
-        Mask.push_back(Mask1[i] + (int)(MaskSize * SrcInputs0.size()));
+        Mask.push_back(i + MaskSize);
       else
         return false;
     }
-    Ops.append(SrcInputs0.begin(), SrcInputs0.end());
-    Ops.append(SrcInputs1.begin(), SrcInputs1.end());
+    Ops.push_back(N0);
+    Ops.push_back(N1);
     return true;
   }
   case ISD::INSERT_SUBVECTOR: {
@@ -34219,6 +34209,7 @@ static bool matchBinaryShuffle(MVT MaskVT, ArrayRef<int> Mask,
                                SelectionDAG &DAG, const X86Subtarget &Subtarget,
                                unsigned &Shuffle, MVT &SrcVT, MVT &DstVT,
                                bool IsUnary) {
+  unsigned NumMaskElts = Mask.size();
   unsigned EltSizeInBits = MaskVT.getScalarSizeInBits();
 
   if (MaskVT.is128BitVector()) {
@@ -34276,6 +34267,46 @@ static bool matchBinaryShuffle(MVT MaskVT, ArrayRef<int> Mask,
     }
   }
 
+  // Attempt to match against a OR if we're performing a blend shuffle and the
+  // non-blended source element is zero in each case.
+  if ((EltSizeInBits % V1.getScalarValueSizeInBits()) == 0 &&
+      (EltSizeInBits % V2.getScalarValueSizeInBits()) == 0) {
+    bool IsBlend = true;
+    unsigned NumV1Elts = V1.getValueType().getVectorNumElements();
+    unsigned NumV2Elts = V2.getValueType().getVectorNumElements();
+    unsigned Scale1 = NumV1Elts / NumMaskElts;
+    unsigned Scale2 = NumV2Elts / NumMaskElts;
+    APInt DemandedZeroV1 = APInt::getNullValue(NumV1Elts);
+    APInt DemandedZeroV2 = APInt::getNullValue(NumV2Elts);
+    for (unsigned i = 0; i != NumMaskElts; ++i) {
+      int M = Mask[i];
+      if (M == SM_SentinelUndef)
+        continue;
+      if (M == SM_SentinelZero) {
+        DemandedZeroV1.setBits(i * Scale1, (i + 1) * Scale1);
+        DemandedZeroV2.setBits(i * Scale2, (i + 1) * Scale2);
+        continue;
+      }
+      if (M == i) {
+        DemandedZeroV2.setBits(i * Scale2, (i + 1) * Scale2);
+        continue;
+      }
+      if (M == (i + NumMaskElts)) {
+        DemandedZeroV1.setBits(i * Scale1, (i + 1) * Scale1);
+        continue;
+      }
+      IsBlend = false;
+      break;
+    }
+    if (IsBlend &&
+        DAG.computeKnownBits(V1, DemandedZeroV1).isZero() &&
+        DAG.computeKnownBits(V2, DemandedZeroV2).isZero()) {
+      Shuffle = ISD::OR;
+      SrcVT = DstVT = EVT(MaskVT).changeTypeToInteger().getSimpleVT();
+      return true;
+    }
+  }
+
   return false;
 }
 
diff --git a/llvm/test/CodeGen/X86/insertelement-ones.ll b/llvm/test/CodeGen/X86/insertelement-ones.ll
index 3d8e42b9c07d..6a9a401264c5 100644
--- a/llvm/test/CodeGen/X86/insertelement-ones.ll
+++ b/llvm/test/CodeGen/X86/insertelement-ones.ll
@@ -389,11 +389,9 @@ define <32 x i8> @insert_v32i8_x123456789ABCDEzGHIJKLMNOPQRSTxx(<32 x i8> %a) {
 ; SSE2-NEXT:    movdqa %xmm3, %xmm4
 ; SSE2-NEXT:    pslldq {{.*#+}} xmm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm4[0]
 ; SSE2-NEXT:    por %xmm4, %xmm0
-; SSE2-NEXT:    movdqa {{.*#+}} xmm5 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,255]
-; SSE2-NEXT:    pand %xmm5, %xmm1
+; SSE2-NEXT:    pand {{.*}}(%rip), %xmm1
 ; SSE2-NEXT:    pslldq {{.*#+}} xmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1]
-; SSE2-NEXT:    pandn %xmm3, %xmm5
-; SSE2-NEXT:    por %xmm5, %xmm1
+; SSE2-NEXT:    por %xmm3, %xmm1
 ; SSE2-NEXT:    pand %xmm2, %xmm1
 ; SSE2-NEXT:    por %xmm4, %xmm1
 ; SSE2-NEXT:    retq
@@ -411,11 +409,9 @@ define <32 x i8> @insert_v32i8_x123456789ABCDEzGHIJKLMNOPQRSTxx(<32 x i8> %a) {
 ; SSE3-NEXT:    movdqa %xmm3, %xmm4
 ; SSE3-NEXT:    pslldq {{.*#+}} xmm4 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm4[0]
 ; SSE3-NEXT:    por %xmm4, %xmm0
-; SSE3-NEXT:    movdqa {{.*#+}} xmm5 = [255,255,255,255,255,255,255,255,255,255,255,255,255,255,0,255]
-; SSE3-NEXT:    pand %xmm5, %xmm1
+; SSE3-NEXT:    pand {{.*}}(%rip), %xmm1
 ; SSE3-NEXT:    pslldq {{.*#+}} xmm3 = zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,zero,xmm3[0,1]
-; SSE3-NEXT:    pandn %xmm3, %xmm5
-; SSE3-NEXT:    por %xmm5, %xmm1
+; SSE3-NEXT:    por %xmm3, %xmm1
 ; SSE3-NEXT:    pand %xmm2, %xmm1
 ; SSE3-NEXT:    por %xmm4, %xmm1
 ; SSE3-NEXT:    retq
diff --git a/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll b/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll
index 6b49f22f21f1..9256a43f8e33 100644
--- a/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll
+++ b/llvm/test/CodeGen/X86/shuffle-vs-trunc-256.ll
@@ -1314,10 +1314,10 @@ define void @trunc_v4i64_to_v4i8(<32 x i8>* %L, <4 x i8>* %S) nounwind {
 define <16 x i8> @negative(<32 x i8> %v, <32 x i8> %w) nounwind {
 ; AVX1-LABEL: negative:
 ; AVX1:       # %bb.0:
-; AVX1-NEXT:    vpshufb {{.*#+}} xmm2 = xmm0[u,2,4,6,8,10,12,14],zero,zero,zero,zero,zero,zero,zero,zero
-; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm0
-; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u],zero,zero,zero,zero,zero,zero,zero,xmm0[0,2,4,6,8,10,12,14]
-; AVX1-NEXT:    vpor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vextractf128 $1, %ymm0, %xmm2
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm2 = xmm2[u,u,u,u,u,u,u,u,0,2,4,6,8,10,12,14]
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u,2,4,6,8,10,12,14,u,u,u,u,u,u,u,u]
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5,6,7]
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [0,255,255,255,255,255,255,255,255,255,255,255,255,255,255,255]
 ; AVX1-NEXT:    vpblendvb %xmm2, %xmm0, %xmm1, %xmm0
 ; AVX1-NEXT:    vzeroupper
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll b/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll
index f448f41cf522..86423ce76065 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-128-v8.ll
@@ -1713,9 +1713,8 @@ define <8 x i16> @shuffle_v8i16_XX4X8acX(<8 x i16> %a, <8 x i16> %b) {
 ;
 ; SSSE3-LABEL: shuffle_v8i16_XX4X8acX:
 ; SSSE3:       # %bb.0:
-; SSSE3-NEXT:    pshufb {{.*#+}} xmm0 = xmm0[u,u,u,u,8,9,u,u],zero,zero,zero,zero,zero,zero,xmm0[u,u]
-; SSSE3-NEXT:    pshufb {{.*#+}} xmm1 = xmm1[u,u,u,u],zero,zero,xmm1[u,u,0,1,4,5,8,9,u,u]
-; SSSE3-NEXT:    por %xmm1, %xmm0
+; SSSE3-NEXT:    pshufb {{.*#+}} xmm1 = xmm1[u,u,u,u,u,u,u,u,0,1,4,5,8,9,u,u]
+; SSSE3-NEXT:    shufps {{.*#+}} xmm0 = xmm0[0,2],xmm1[2,3]
 ; SSSE3-NEXT:    retq
 ;
 ; SSE41-LABEL: shuffle_v8i16_XX4X8acX:
diff --git a/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll b/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll
index 82df05e5ae06..e5285aebda69 100644
--- a/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll
+++ b/llvm/test/CodeGen/X86/vector-shuffle-256-v32.ll
@@ -3358,9 +3358,9 @@ define <32 x i8> @shuffle_v32i8_42_45_12_13_35_35_60_40_17_22_29_44_33_12_48_51_
 ; AVX1-NEXT:    vpshufb {{.*#+}} xmm2 = zero,zero,xmm2[u,u],zero,zero,xmm2[12],zero,xmm2[u,u,u],zero,zero,xmm2[u,0,3]
 ; AVX1-NEXT:    vpshufb {{.*#+}} xmm1 = xmm1[10,13,u,u,3,3],zero,xmm1[8,u,u,u,12,1,u],zero,zero
 ; AVX1-NEXT:    vpor %xmm2, %xmm1, %xmm1
-; AVX1-NEXT:    vpshufb {{.*#+}} xmm2 = xmm4[u,u],zero,zero,xmm4[u,u,u,u,1,6,13,u,u],zero,xmm4[u,u]
-; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u,u,12,13,u,u,u,u],zero,zero,zero,xmm0[u,u,12,u,u]
-; AVX1-NEXT:    vpor %xmm2, %xmm0, %xmm0
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm2 = xmm4[u,u,u,u,u,u,u,u,1,6,13,u,u,u,u,u]
+; AVX1-NEXT:    vpshufb {{.*#+}} xmm0 = xmm0[u,u,12,13,u,u,u,u,u,u,u,u,u,12,u,u]
+; AVX1-NEXT:    vpblendw {{.*#+}} xmm0 = xmm0[0,1,2,3],xmm2[4,5],xmm0[6,7]
 ; AVX1-NEXT:    vmovdqa {{.*#+}} xmm2 = [255,255,0,0,255,255,255,255,0,0,0,255,255,0,255,255]
 ; AVX1-NEXT:    vpblendvb %xmm2, %xmm1, %xmm0, %xmm0
 ; AVX1-NEXT:    vinsertf128 $1, %xmm3, %ymm0, %ymm0