3 files changed, 64 insertions, 2 deletions

diff --git a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp
index f007349d184..768f688a09d 100644
--- a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp
+++ b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.cpp
@@ -342,7 +342,46 @@ void DecodeBLENDMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
   }
 }
 
-/// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
+void DecodeVPPERMMask(ArrayRef<uint64_t> RawMask,
+                      SmallVectorImpl<int> &ShuffleMask) {
+  assert(RawMask.size() == 16 && "Illegal VPPERM shuffle mask size");
+
+  // VPPERM Operation
+  // Bits[4:0] - Byte Index (0 - 31)
+  // Bits[7:5] - Permute Operation
+  //
+  // Permute Operation:
+  // 0 - Source byte (no logical operation).
+  // 1 - Invert source byte.
+  // 2 - Bit reverse of source byte.
+  // 3 - Bit reverse of inverted source byte.
+  // 4 - 00h (zero - fill).
+  // 5 - FFh (ones - fill).
+  // 6 - Most significant bit of source byte replicated in all bit positions.
+  // 7 - Invert most significant bit of source byte and replicate in all bit positions.
+  for (int i = 0, e = RawMask.size(); i < e; ++i) {
+    uint64_t M = RawMask[i];
+    if (M == (uint64_t)SM_SentinelUndef) {
+      ShuffleMask.push_back(M);
+      continue;
+    }
+
+    uint64_t PermuteOp = (M >> 5) & 0x3;
+    if (PermuteOp == 4) {
+      ShuffleMask.push_back(SM_SentinelZero);
+      continue;
+    }
+    if (PermuteOp != 0) {
+      ShuffleMask.clear();
+      return;
+    }
+
+    uint64_t Index = M & 0x1F;
+    ShuffleMask.push_back((int)Index);
+  }
+}
+
+  /// DecodeVPERMMask - this decodes the shuffle masks for VPERMQ/VPERMPD.
 /// No VT provided since it only works on 256-bit, 4 element vectors.
 void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask) {
   for (unsigned i = 0; i != 4; ++i) {
diff --git a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h
index 0c92b39e07f..4a4d33b3689 100644
--- a/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h
+++ b/llvm/lib/Target/X86/Utils/X86ShuffleDecode.h
@@ -73,7 +73,7 @@ void DecodePSHUFLWMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask);
 /// Decodes a PSWAPD 3DNow! instruction.
 void DecodePSWAPMask(MVT VT, SmallVectorImpl<int> &ShuffleMask);
 
-/// Decodes the shuffle masks for shufp*. 
+/// Decodes the shuffle masks for shufp*.
 /// VT indicates the type of the vector allowing it to handle different
 /// datatypes and vector widths.
 void DecodeSHUFPMask(MVT VT, unsigned Imm, SmallVectorImpl<int> &ShuffleMask);
@@ -108,6 +108,13 @@ void decodeVSHUF64x2FamilyMask(MVT VT, unsigned Imm,
 /// No VT provided since it only works on 256-bit, 4 element vectors.
 void DecodeVPERMMask(unsigned Imm, SmallVectorImpl<int> &ShuffleMask);
 
+/// Decode a VPPERM mask from a raw array of constants such as from
+/// BUILD_VECTOR.
+/// This can only basic masks (permutes + zeros), not any of the other
+/// operations that VPPERM can perform.
+void DecodeVPPERMMask(ArrayRef<uint64_t> RawMask,
+                      SmallVectorImpl<int> &ShuffleMask);
+
 /// Decode a zero extension instruction as a shuffle mask.
 void DecodeZeroExtendMask(MVT SrcScalarVT, MVT DstVT,
                           SmallVectorImpl<int> &ShuffleMask);
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 7f699762262..284925c9030 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -3871,6 +3871,7 @@ static bool isTargetShuffle(unsigned Opcode) {
   case X86ISD::VPERMILPV:
   case X86ISD::VPERM2X128:
   case X86ISD::VPERMI:
+  case X86ISD::VPPERM:
   case X86ISD::VPERMV:
   case X86ISD::VPERMV3:
   case X86ISD::VZEXT_MOVL:
@@ -5008,6 +5009,20 @@ static bool getTargetShuffleMask(SDNode *N, MVT VT, bool AllowSentinelZero,
   case X86ISD::MOVLPS:
     // Not yet implemented
     return false;
+  case X86ISD::VPPERM: {
+    IsUnary = IsFakeUnary = N->getOperand(0) == N->getOperand(1);
+    SDValue MaskNode = N->getOperand(2);
+    SmallVector<uint64_t, 32> RawMask;
+    if (getTargetShuffleMaskIndices(MaskNode, 8, RawMask)) {
+      DecodeVPPERMMask(RawMask, Mask);
+      break;
+    }
+    if (auto *C = getTargetShuffleMaskConstant(MaskNode)) {
+      DecodeVPPERMMask(C, Mask);
+      break;
+    }
+    return false;
+  }
   case X86ISD::VPERMV: {
     IsUnary = true;
     // Unlike most shuffle nodes, VPERMV's mask operand is operand 0.
@@ -29688,6 +29703,7 @@ SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
   case X86ISD::MOVDDUP:
   case X86ISD::MOVSS:
   case X86ISD::MOVSD:
+  case X86ISD::VPPERM:
   case X86ISD::VPERMV3:
   case X86ISD::VPERMILPI:
   case X86ISD::VPERMILPV: