[X86][SSE] Added support for combining bit-shifts with shuffles.

Bit-shifts by a whole number of bytes can be represented as a shuffle mask suitable for combining.

Added a 'getFauxShuffleMask' function to allow us to create shuffle masks from other suitable operations.

llvm-svn: 288040

1 files changed, 57 insertions, 5 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 25ad59f919b..2863ec39421 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -5395,8 +5395,9 @@ static bool setTargetShuffleZeroElements(SDValue N,
   bool IsUnary;
   if (!isTargetShuffle(N.getOpcode()))
     return false;
-  if (!getTargetShuffleMask(N.getNode(), N.getSimpleValueType(), true, Ops,
-                            Mask, IsUnary))
+
+  MVT VT = N.getSimpleValueType();
+  if (!getTargetShuffleMask(N.getNode(), VT, true, Ops, Mask, IsUnary))
     return false;
 
   SDValue V1 = Ops[0];
@@ -5458,9 +5459,61 @@ static bool setTargetShuffleZeroElements(SDValue N,
     }
   }
 
+  assert(VT.getVectorNumElements() == Mask.size() &&
+         "Different mask size from vector size!");
   return true;
 }
 
+// Attempt to decode ops that could be represented as a shuffle mask.
+// The decoded shuffle mask may contain a different number of elements to the
+// destination value type.
+static bool getFauxShuffleMask(SDValue N, SmallVectorImpl<int> &Mask,
+                               SmallVectorImpl<SDValue> &Ops) {
+  Mask.clear();
+  Ops.clear();
+
+  MVT VT = N.getSimpleValueType();
+  unsigned NumElts = VT.getVectorNumElements();
+
+  unsigned Opcode = N.getOpcode();
+  switch (Opcode) {
+  case X86ISD::VSHLI:
+  case X86ISD::VSRLI: {
+    uint64_t ShiftVal = N.getConstantOperandVal(1);
+    // Out of range bit shifts are guaranteed to be zero.
+    if (VT.getScalarSizeInBits() <= ShiftVal) {
+      Mask.append(NumElts, SM_SentinelZero);
+      return true;
+    }
+
+    // We can only decode 'whole byte' bit shifts as shuffles.
+    if ((ShiftVal % 8) != 0)
+      break;
+
+    uint64_t ByteShift = ShiftVal / 8;
+    unsigned NumBytes = VT.getSizeInBits() / 8;
+    unsigned NumBytesPerElt = VT.getScalarSizeInBits() / 8;
+    Ops.push_back(N.getOperand(0));
+
+    // Clear mask to all zeros and insert the shifted byte indices.
+    Mask.append(NumBytes, SM_SentinelZero);
+
+    if (X86ISD::VSHLI == Opcode) {
+      for (unsigned i = 0; i != NumBytes; i += NumBytesPerElt)
+        for (unsigned j = ByteShift; j != NumBytesPerElt; ++j)
+          Mask[i + j] = i + j - ByteShift;
+    } else {
+      for (unsigned i = 0; i != NumBytes; i += NumBytesPerElt)
+        for (unsigned j = ByteShift; j != NumBytesPerElt; ++j)
+          Mask[i + j - ByteShift] = i + j;
+    }
+    return true;
+  }
+  }
+
+  return false;
+}
+
 /// Calls setTargetShuffleZeroElements to resolve a target shuffle mask's inputs
 /// and set the SM_SentinelUndef and SM_SentinelZero values. Then check the
 /// remaining input indices in case we now have a unary shuffle and adjust the
@@ -5470,7 +5523,8 @@ static bool resolveTargetShuffleInputs(SDValue Op, SDValue &Op0, SDValue &Op1,
                                        SmallVectorImpl<int> &Mask) {
   SmallVector<SDValue, 2> Ops;
   if (!setTargetShuffleZeroElements(Op, Mask, Ops))
-    return false;
+    if (!getFauxShuffleMask(Op, Mask, Ops))
+      return false;
 
   int NumElts = Mask.size();
   bool Op0InUse = any_of(Mask, [NumElts](int Idx) {
@@ -26299,8 +26353,6 @@ static bool combineX86ShufflesRecursively(ArrayRef<SDValue> SrcOps,
     Ops.push_back(Input1);
   }
 
-  assert(VT.getVectorNumElements() == OpMask.size() &&
-         "Different mask size from vector size!");
   assert(((RootMask.size() > OpMask.size() &&
            RootMask.size() % OpMask.size() == 0) ||
           (OpMask.size() > RootMask.size() &&