[X86] Don't assume that shuffle non-mask operands starts at #0.

That's not the case for VPERMV/VPERMV3, which cover all possible
combinations (the C intrinsics use a different order; the AVX vs
AVX512 intrinsics are different still).

Since:
  r246981 AVX-512: Lowering for 512-bit vector shuffles.
VPERMV is recognized in getTargetShuffleMask.

This breaks assumptions in most callers, as they expect
the non-mask operands to start at index 0.
VPERMV has the mask as operand #0; VPERMV3 has it in the middle.

Instead of the faulty assumption, have getTargetShuffleMask return
its operands as well.

One alternative we considered was to change the operand order of
VPERMV, but we agreed to stick to the instruction order, as there
are more AVX512 weirdness to cover (vpermt2/vpermi2 in particular).

Differential Revision: http://reviews.llvm.org/D17041

llvm-svn: 262627

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -4936,15 +4936,21 @@ static const Constant *getTargetShuffleMaskConstant(SDValue MaskNode) {
 }
 
 /// Calculates the shuffle mask corresponding to the target-specific opcode.
-/// Returns true if the Mask could be calculated. Sets IsUnary to true if only
-/// uses one source. Note that this will set IsUnary for shuffles which use a
-/// single input multiple times, and in those cases it will
-/// adjust the mask to only have indices within that single input.
+/// If the mask could be calculated, returns it in \p Mask, returns the shuffle
+/// operands in \p Ops, and returns true.
+/// Sets \p IsUnary to true if only one source is used. Note that this will set
+/// IsUnary for shuffles which use a single input multiple times, and in those
+/// cases it will adjust the mask to only have indices within that single input.
+/// It is an error to call this with non-empty Mask/Ops vectors.
 static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
+                                 SmallVectorImpl<SDValue> &Ops,
                                  SmallVectorImpl<int> &Mask, bool &IsUnary) {
   unsigned NumElems = VT.getVectorNumElements();
   SDValue ImmN;
 
+  assert(Mask.empty() && "getTargetShuffleMask expects an empty Mask vector");
+  assert(Ops.empty() && "getTargetShuffleMask expects an empty Ops vector");
+
   IsUnary = false;
   bool IsFakeUnary = false;
   switch(N->getOpcode()) {
@@ -5049,6 +5055,8 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
     return false;
   case X86ISD::VPERMV: {
     IsUnary = true;
+    // Unlike most shuffle nodes, VPERMV's mask operand is operand 0.
+    Ops.push_back(N->getOperand(1));
     SDValue MaskNode = N->getOperand(0);
     SmallVector<uint64_t, 32> RawMask;
     unsigned MaskLoBits = Log2_64(VT.getVectorNumElements());
@@ -5063,8 +5071,10 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
     return false;
   }
   case X86ISD::VPERMV3: {
+    // Unlike most shuffle nodes, VPERMV3's mask operand is the middle one.
+    Ops.push_back(N->getOperand(0));
+    Ops.push_back(N->getOperand(2));
     SDValue MaskNode = N->getOperand(1);
-
     SmallVector<uint64_t, 32> RawMask;
     unsigned MaskLoBits = Log2_64(VT.getVectorNumElements() * 2);
     if (getTargetShuffleMaskIndices(MaskNode, MaskLoBits, RawMask)) {
@@ -5098,6 +5108,14 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
       if (M >= (int)Mask.size())
         M -= Mask.size();
 
+  // If we didn't already add operands in the opcode-specific code, default to
+  // adding 1 or 2 operands starting at 0.
+  if (Ops.empty()) {
+    Ops.push_back(N->getOperand(0));
+    if (!IsUnary || IsFakeUnary)
+      Ops.push_back(N->getOperand(1));
+  }
+
   return true;
 }
 
@@ -5106,16 +5124,17 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
 /// (not just zeroable) from their inputs.
 /// Returns true if the target shuffle mask was decoded.
 static bool setTargetShuffleZeroElements(SDValue N,
-                                         SmallVectorImpl<int> &Mask) {
+                                         SmallVectorImpl<int> &Mask,
+                                         SmallVectorImpl<SDValue> &Ops) {
   bool IsUnary;
   if (!isTargetShuffle(N.getOpcode()))
     return false;
-  if (!getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), true, Mask,
-                            IsUnary))
+  if (!getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), true, Ops,
+                            Mask, IsUnary))
     return false;
 
-  SDValue V1 = N.getOperand(0);
-  SDValue V2 = IsUnary ? V1 : N.getOperand(1);
+  SDValue V1 = Ops[0];
+  SDValue V2 = IsUnary ? V1 : Ops[1];
 
   while (V1.getOpcode() == ISD::BITCAST)
     V1 = V1->getOperand(0);
@@ -5186,7 +5205,8 @@ static bool setTargetShuffleZeroElements(SDValue N,
 static bool resolveTargetShuffleInputs(SDValue Op, bool &IsUnary, SDValue &Op0,
                                        SDValue &Op1,
                                        SmallVectorImpl<int> &Mask) {
-  if (!setTargetShuffleZeroElements(Op, Mask))
+  SmallVector<SDValue, 2> Ops;
+  if (!setTargetShuffleZeroElements(Op, Mask, Ops))
     return false;
 
   int NumElts = Mask.size();
@@ -5196,8 +5216,8 @@ static bool resolveTargetShuffleInputs(SDValue Op, bool &IsUnary, SDValue &Op0,
   bool Op1InUse = std::any_of(Mask.begin(), Mask.end(),
                               [NumElts](int Idx) { return NumElts <= Idx; });
 
-  Op0 = Op0InUse ? Op.getOperand(0) : SDValue();
-  Op1 = Op1InUse ? Op.getOperand(1) : SDValue();
+  Op0 = Op0InUse ? Ops[0] : SDValue();
+  Op1 = Op1InUse ? Ops[1] : SDValue();
   IsUnary = !(Op0InUse && Op1InUse);
 
   if (!IsUnary)
@@ -5245,9 +5265,10 @@ static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
     MVT ShufSVT = ShufVT.getVectorElementType();
     int NumElems = (int)ShufVT.getVectorNumElements();
     SmallVector<int, 16> ShuffleMask;
+    SmallVector<SDValue, 16> ShuffleOps;
     bool IsUnary;
 
-    if (!getTargetShuffleMask(N, ShufVT, true, ShuffleMask, IsUnary))
+    if (!getTargetShuffleMask(N, ShufVT, true, ShuffleOps, ShuffleMask, IsUnary))
       return SDValue();
 
     int Elt = ShuffleMask[Index];
@@ -5258,7 +5279,7 @@ static SDValue getShuffleScalarElt(SDNode *N, unsigned Index, SelectionDAG &DAG,
       return DAG.getUNDEF(ShufSVT);
 
     assert(0 <= Elt && Elt < (2*NumElems) && "Shuffle index out of range");
-    SDValue NewV = (Elt < NumElems) ? N->getOperand(0) : N->getOperand(1);
+    SDValue NewV = (Elt < NumElems) ? ShuffleOps[0] : ShuffleOps[1];
     return getShuffleScalarElt(NewV.getNode(), Elt % NumElems, DAG,
                                Depth+1);
   }
@@ -24044,8 +24065,10 @@ static bool combineX86ShufflesRecursively(SDValue Op, SDValue Root,
 static SmallVector<int, 4> getPSHUFShuffleMask(SDValue N) {
   MVT VT = N.getSimpleValueType();
   SmallVector<int, 4> Mask;
+  SmallVector<SDValue, 2> Ops;
   bool IsUnary;
-  bool HaveMask = getTargetShuffleMask(N.getNode(), VT, false, Mask, IsUnary);
+  bool HaveMask =
+      getTargetShuffleMask(N.getNode(), VT, false, Ops, Mask, IsUnary);
   (void)HaveMask;
   assert(HaveMask);
 
@@ -24360,7 +24383,8 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
 
       // Determine which elements are known to be zero.
       SmallVector<int, 8> TargetMask;
-      if (!setTargetShuffleZeroElements(N, TargetMask))
+      SmallVector<SDValue, 2> BlendOps;
+      if (!setTargetShuffleZeroElements(N, TargetMask, BlendOps))
         return SDValue();
 
       // Helper function to take inner insertps node and attempt to
@@ -24414,7 +24438,8 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
 
     // Attempt to merge insertps Op1 with an inner target shuffle node.
     SmallVector<int, 8> TargetMask1;
-    if (setTargetShuffleZeroElements(Op1, TargetMask1)) {
+    SmallVector<SDValue, 2> Ops1;
+    if (setTargetShuffleZeroElements(Op1, TargetMask1, Ops1)) {
       int M = TargetMask1[SrcIdx];
       if (isUndefOrZero(M)) {
         // Zero/UNDEF insertion - zero out element and remove dependency.
@@ -24425,14 +24450,15 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
       // Update insertps mask srcidx and reference the source input directly.
       assert(0 <= M && M < 8 && "Shuffle index out of range");
       InsertPSMask = (InsertPSMask & 0x3f) | ((M & 0x3) << 6);
-      Op1 = Op1.getOperand(M < 4 ? 0 : 1);
+      Op1 = Ops1[M < 4 ? 0 : 1];
       return DAG.getNode(X86ISD::INSERTPS, DL, VT, Op0, Op1,
                          DAG.getConstant(InsertPSMask, DL, MVT::i8));
     }
 
     // Attempt to merge insertps Op0 with an inner target shuffle node.
     SmallVector<int, 8> TargetMask0;
-    if (!setTargetShuffleZeroElements(Op0, TargetMask0))
+    SmallVector<SDValue, 2> Ops0;
+    if (!setTargetShuffleZeroElements(Op0, TargetMask0, Ops0))
       return SDValue();
 
     bool Updated = false;
@@ -24463,10 +24489,10 @@ static SDValue combineTargetShuffle(SDValue N, SelectionDAG &DAG,
     // referenced input directly.
     if (UseInput00 && !UseInput01) {
       Updated = true;
-      Op0 = Op0.getOperand(0);
+      Op0 = Ops0[0];
     } else if (!UseInput00 && UseInput01) {
       Updated = true;
-      Op0 = Op0.getOperand(1);
+      Op0 = Ops0[1];
     }
 
     if (Updated)
@@ -24767,9 +24793,10 @@ static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
     return SDValue();
 
   SmallVector<int, 16> ShuffleMask;
+  SmallVector<SDValue, 2> ShuffleOps;
   bool UnaryShuffle;
   if (!getTargetShuffleMask(InVec.getNode(), CurrentVT.getSimpleVT(), true,
-                            ShuffleMask, UnaryShuffle))
+                            ShuffleOps, ShuffleMask, UnaryShuffle))
     return SDValue();
 
   // Select the input vector, guarding against out of range extract vector.
@@ -24784,12 +24811,12 @@ static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
     return DAG.getUNDEF(EltVT);
 
   assert(0 <= Idx && Idx < (int)(2 * NumElems) && "Shuffle index out of range");
-  SDValue LdNode = (Idx < (int)NumElems) ? InVec.getOperand(0)
-                                         : InVec.getOperand(1);
+  SDValue LdNode = (Idx < (int)NumElems) ? ShuffleOps[0]
+                                         : ShuffleOps[1];
 
   // If inputs to shuffle are the same for both ops, then allow 2 uses
-  unsigned AllowedUses = InVec.getNumOperands() > 1 &&
-                         InVec.getOperand(0) == InVec.getOperand(1) ? 2 : 1;
+  unsigned AllowedUses =
+      (ShuffleOps.size() > 1 && ShuffleOps[0] == ShuffleOps[1]) ? 2 : 1;
 
   if (LdNode.getOpcode() == ISD::BITCAST) {
     // Don't duplicate a load with other uses.
@@ -24823,10 +24850,9 @@ static SDValue XFormVExtractWithShuffleIntoLoad(SDNode *N, SelectionDAG &DAG,
   SDLoc dl(N);
 
   // Create shuffle node taking into account the case that its a unary shuffle
-  SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT)
-                                   : InVec.getOperand(1);
+  SDValue Shuffle = (UnaryShuffle) ? DAG.getUNDEF(CurrentVT) : ShuffleOps[1];
   Shuffle = DAG.getVectorShuffle(CurrentVT, dl,
-                                 InVec.getOperand(0), Shuffle,
+                                 ShuffleOps[0], Shuffle,
                                  &ShuffleMask[0]);
   Shuffle = DAG.getBitcast(OriginalVT, Shuffle);
   return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, N->getValueType(0), Shuffle,

llvm/lib/Target/X86/X86ISelLowering.h

@@ -392,11 +392,21 @@ namespace llvm {
       UNPCKH,
       VPERMILPV,
       VPERMILPI,
+      VPERMI,
+      VPERM2X128,
+
+      // Variable Permute (VPERM)
+      // Res = VPERMV MaskV, V0
       VPERMV,
+
+      // 3-op Variable Permute (VPERMT2)
+      // Res = VPERMV3 V0, MaskV, V1
       VPERMV3,
+
+      // 3-op Variable Permute overwriting the index (VPERMI2)
+      // Res = VPERMIV3 V0, MaskV, V1
       VPERMIV3,
-      VPERMI,
-      VPERM2X128,
+
       // Bitwise ternary logic
       VPTERNLOG,
       // Fix Up Special Packed Float32/64 values

llvm/test/CodeGen/X86/vector-shuffle-combining-avx2.ll

@@ -0,0 +1,27 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc < %s -mtriple=x86_64-- -mattr=+avx2 | FileCheck %s
+
+declare <8 x i32> @llvm.x86.avx2.permd(<8 x i32>, <8 x i32>)
+declare <8 x float> @llvm.x86.avx2.permps(<8 x float>, <8 x i32>)
+
+define <32 x i8> @combine_pshufb_vpermd(<8 x i32> %a) {
+; CHECK-LABEL: combine_pshufb_vpermd:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,16,17,18,18]
+; CHECK-NEXT:    retq
+  %tmp0 = call <8 x i32> @llvm.x86.avx2.permd(<8 x i32> %a, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 4>)
+  %tmp1 = bitcast <8 x i32> %tmp0 to <32 x i8>
+  %tmp2 = shufflevector <32 x i8> %tmp1, <32 x i8> undef, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 30>
+  ret <32 x i8> %tmp2
+}
+
+define <32 x i8> @combine_pshufb_vpermps(<8 x float> %a) {
+; CHECK-LABEL: combine_pshufb_vpermps:
+; CHECK:       # BB#0:
+; CHECK-NEXT:    vpshufb {{.*#+}} ymm0 = ymm0[0,1,2,3,4,5,6,7,8,9,10,11,12,13,14,15,16,17,18,19,20,21,22,23,24,25,26,27,16,17,18,18]
+; CHECK-NEXT:    retq
+  %tmp0 = call <8 x float> @llvm.x86.avx2.permps(<8 x float> %a, <8 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 4>)
+  %tmp1 = bitcast <8 x float> %tmp0 to <32 x i8>
+  %tmp2 = shufflevector <32 x i8> %tmp1, <32 x i8> undef, <32 x i32> <i32 0, i32 1, i32 2, i32 3, i32 4, i32 5, i32 6, i32 7, i32 8, i32 9, i32 10, i32 11, i32 12, i32 13, i32 14, i32 15, i32 16, i32 17, i32 18, i32 19, i32 20, i32 21, i32 22, i32 23, i32 24, i32 25, i32 26, i32 27, i32 28, i32 29, i32 30, i32 30>
+  ret <32 x i8> %tmp2
+}