[X86] Try to zero elts when lowering 256-bit shuffle with PSHUFB.

Otherwise we fallback to a blend of PSHUFBs later on.

Differential Revision: http://reviews.llvm.org/D19661

llvm-svn: 271113

1 files changed, 66 insertions, 35 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 1d92d7b3392..f3b4c70b712 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -7170,6 +7170,59 @@ static SmallBitVector computeZeroableShuffleElements(ArrayRef<int> Mask,
   return Zeroable;
 }
 
+/// Mutate a shuffle mask, replacing zeroable elements with SM_SentinelZero.
+static void computeZeroableShuffleMask(MutableArrayRef<int> Mask,
+                                       SDValue V1, SDValue V2) {
+  SmallBitVector Zeroable = computeZeroableShuffleElements(Mask, V1, V2);
+  for (int i = 0, Size = Mask.size(); i < Size; ++i) {
+    if (Mask[i] != SM_SentinelUndef && Zeroable[i])
+      Mask[i] = SM_SentinelZero;
+  }
+}
+
+/// Try to lower a shuffle with a single PSHUFB of V1.
+/// This is only possible if V2 is unused (at all, or only for zero elements).
+static SDValue lowerVectorShuffleWithPSHUFB(SDLoc DL, MVT VT,
+                                            ArrayRef<int> Mask, SDValue V1,
+                                            SDValue V2,
+                                            const X86Subtarget &Subtarget,
+                                            SelectionDAG &DAG) {
+  const int NumBytes = VT.is128BitVector() ? 16 : 32;
+  const int NumEltBytes = VT.getScalarSizeInBits() / 8;
+
+  assert((Subtarget.hasSSSE3() && VT.is128BitVector()) ||
+         (Subtarget.hasAVX2() && VT.is256BitVector()));
+
+  SmallVector<int, 32> ZeroableMask(Mask.begin(), Mask.end());
+  computeZeroableShuffleMask(ZeroableMask, V1, V2);
+
+  if (!isSingleInputShuffleMask(ZeroableMask) ||
+      is128BitLaneCrossingShuffleMask(VT, Mask))
+    return SDValue();
+
+  SmallVector<SDValue, 32> PSHUFBMask(NumBytes);
+  // Sign bit set in i8 mask means zero element.
+  SDValue ZeroMask = DAG.getConstant(0x80, DL, MVT::i8);
+
+  for (int i = 0; i < NumBytes; ++i) {
+    int M = ZeroableMask[i / NumEltBytes];
+    if (M == SM_SentinelUndef) {
+      PSHUFBMask[i] = DAG.getUNDEF(MVT::i8);
+    } else if (M == SM_SentinelZero) {
+      PSHUFBMask[i] = ZeroMask;
+    } else {
+      M = M * NumEltBytes + (i % NumEltBytes);
+      M = i < 16 ? M : M - 16;
+      PSHUFBMask[i] = DAG.getConstant(M, DL, MVT::i8);
+    }
+  }
+
+  MVT I8VT = MVT::getVectorVT(MVT::i8, NumBytes);
+  return DAG.getBitcast(
+      VT, DAG.getNode(X86ISD::PSHUFB, DL, I8VT, DAG.getBitcast(I8VT, V1),
+                      DAG.getBuildVector(I8VT, DL, PSHUFBMask)));
+}
+
 // X86 has dedicated unpack instructions that can handle specific blend
 // operations: UNPCKH and UNPCKL.
 static SDValue lowerVectorShuffleWithUNPCK(SDLoc DL, MVT VT, ArrayRef<int> Mask,
@@ -11389,26 +11442,12 @@ static SDValue lowerV16I16VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
       return lowerV8I16GeneralSingleInputVectorShuffle(
           DL, MVT::v16i16, V1, RepeatedMask, Subtarget, DAG);
     }
-
-    SDValue PSHUFBMask[32];
-    for (int i = 0; i < 16; ++i) {
-      if (Mask[i] == -1) {
-        PSHUFBMask[2 * i] = PSHUFBMask[2 * i + 1] = DAG.getUNDEF(MVT::i8);
-        continue;
-      }
-
-      int M = i < 8 ? Mask[i] : Mask[i] - 8;
-      assert(M >= 0 && M < 8 && "Invalid single-input mask!");
-      PSHUFBMask[2 * i] = DAG.getConstant(2 * M, DL, MVT::i8);
-      PSHUFBMask[2 * i + 1] = DAG.getConstant(2 * M + 1, DL, MVT::i8);
-    }
-    return DAG.getBitcast(
-        MVT::v16i16,
-        DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8,
-                    DAG.getBitcast(MVT::v32i8, V1),
-                    DAG.getBuildVector(MVT::v32i8, DL, PSHUFBMask)));
   }
 
+  if (SDValue PSHUFB = lowerVectorShuffleWithPSHUFB(DL, MVT::v16i16, Mask, V1,
+                                                    V2, Subtarget, DAG))
+    return PSHUFB;
+
   // Try to simplify this by merging 128-bit lanes to enable a lane-based
   // shuffle.
   if (SDValue Result = lowerVectorShuffleByMerging128BitLanes(
@@ -11471,24 +11510,16 @@ static SDValue lowerV32I8VectorShuffle(SDValue Op, SDValue V1, SDValue V2,
           DL, MVT::v32i8, V1, V2, Mask, Subtarget, DAG))
     return V;
 
-  if (isSingleInputShuffleMask(Mask)) {
-    // There are no generalized cross-lane shuffle operations available on i8
-    // element types.
-    if (is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
-      return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2,
-                                                     Mask, DAG);
-
-    SDValue PSHUFBMask[32];
-    for (int i = 0; i < 32; ++i)
-      PSHUFBMask[i] =
-          Mask[i] < 0
-              ? DAG.getUNDEF(MVT::i8)
-              : DAG.getConstant(Mask[i] < 16 ? Mask[i] : Mask[i] - 16, DL,
-                                MVT::i8);
+  // There are no generalized cross-lane shuffle operations available on i8
+  // element types.
+  if (isSingleInputShuffleMask(Mask) &&
+      is128BitLaneCrossingShuffleMask(MVT::v32i8, Mask))
+    return lowerVectorShuffleAsLanePermuteAndBlend(DL, MVT::v32i8, V1, V2, Mask,
+                                                   DAG);
 
-    return DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, V1,
-                       DAG.getBuildVector(MVT::v32i8, DL, PSHUFBMask));
-  }
+  if (SDValue PSHUFB = lowerVectorShuffleWithPSHUFB(DL, MVT::v32i8, Mask, V1,
+                                                    V2, Subtarget, DAG))
+    return PSHUFB;
 
   // Try to simplify this by merging 128-bit lanes to enable a lane-based
   // shuffle.