[x86] Refactor the bit shift code the same as I just did the byte shift

code.

While this didn't have the miscompile (it used MatchLeft consistently)
it missed some cases where it could use right shifts. I've added a test
case Craig Topper came up with to exercise the right shift matching.

This code is really identical between the two. I'm going to merge them
next so that we don't keep two copies of all of this logic.

llvm-svn: 229655

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -7946,39 +7946,34 @@ static SDValue lowerVectorShuffleAsBitShift(SDLoc DL, MVT VT, SDValue V1,
   // PSHL : (little-endian) left bit shift.
   // [ zz, 0, zz,  2 ]
   // [ -1, 4, zz, -1 ]
-  auto MatchBitShift = [&](int Shift, int Scale) -> SDValue {
+
+  auto CheckZeros = [&](int Shift, int Scale, bool Left) {
+    for (int i = 0; i < Size; i += Scale)
+      for (int j = 0; j < Shift; ++j)
+        if (!Zeroable[i + j + (Left ? 0 : (Scale - Shift))])
+          return false;
+
+    return true;
+  };
+
+  auto MatchBitShift = [&](int Shift, int Scale, bool Left, SDValue V) {
     MVT ShiftSVT = MVT::getIntegerVT(VT.getScalarSizeInBits() * Scale);
     MVT ShiftVT = MVT::getVectorVT(ShiftSVT, Size / Scale);
     assert(DAG.getTargetLoweringInfo().isTypeLegal(ShiftVT) &&
            "Illegal integer vector type");
 
-    bool MatchLeft = true, MatchRight = true;
     for (int i = 0; i != Size; i += Scale) {
-      for (int j = 0; j != Shift; ++j) {
-        MatchLeft &= Zeroable[i + j];
-      }
-      for (int j = Scale - Shift; j != Scale; ++j) {
-        MatchRight &= Zeroable[i + j];
-      }
-    }
-    if (!(MatchLeft || MatchRight))
-      return SDValue();
-
-    bool MatchV1 = true, MatchV2 = true;
-    for (int i = 0; i != Size; i += Scale) {
-      unsigned Pos = MatchLeft ? i + Shift : i;
-      unsigned Low = MatchLeft ? i : i + Shift;
+      unsigned Pos = Left ? i + Shift : i;
+      unsigned Low = Left ? i : i + Shift;
       unsigned Len = Scale - Shift;
-      MatchV1 &= isSequentialOrUndefInRange(Mask, Pos, Len, Low);
-      MatchV2 &= isSequentialOrUndefInRange(Mask, Pos, Len, Low + Size);
+      if (!isSequentialOrUndefInRange(Mask, Pos, Len,
+                                      Low + (V == V1 ? 0 : Size)))
+        return SDValue();
     }
-    if (!(MatchV1 || MatchV2))
-      return SDValue();
 
     // Cast the inputs to ShiftVT to match VSRLI/VSHLI and back again.
-    unsigned OpCode = MatchLeft ? X86ISD::VSHLI : X86ISD::VSRLI;
+    unsigned OpCode = Left ? X86ISD::VSHLI : X86ISD::VSRLI;
     int ShiftAmt = Shift * VT.getScalarSizeInBits();
-    SDValue V = MatchV1 ? V1 : V2;
     V = DAG.getNode(ISD::BITCAST, DL, ShiftVT, V);
     V = DAG.getNode(OpCode, DL, ShiftVT, V, DAG.getConstant(ShiftAmt, MVT::i8));
     return DAG.getNode(ISD::BITCAST, DL, VT, V);
@@ -7992,8 +7987,11 @@ static SDValue lowerVectorShuffleAsBitShift(SDLoc DL, MVT VT, SDValue V1,
   // and that the shifted in elements are all zeroable.
   for (int Scale = 2; Scale * VT.getScalarSizeInBits() <= 64; Scale *= 2)
     for (int Shift = 1; Shift != Scale; ++Shift)
-      if (SDValue BitShift = MatchBitShift(Shift, Scale))
-        return BitShift;
+      for (bool Left : {true, false})
+        if (CheckZeros(Shift, Scale, Left))
+          for (SDValue V : {V1, V2})
+            if (SDValue BitShift = MatchBitShift(Shift, Scale, Left, V))
+              return BitShift;
 
   // no match
   return SDValue();

llvm/test/CodeGen/X86/vector-shuffle-128-v16.ll

@@ -1434,3 +1434,17 @@ entry:
   %0 = shufflevector <16 x i8> %inval1, <16 x i8> %inval2, <16 x i32> <i32 0, i32 2, i32 4, i32 6, i32 8, i32 10, i32 12, i32 14, i32 16, i32 18, i32 20, i32 22, i32 24, i32 26, i32 28, i32 30>
   ret <16 x i8> %0
 }
+
+define <16 x i8> @shuffle_v16i8_uu_02_03_zz_uu_06_07_zz_uu_10_11_zz_uu_14_15_zz(<16 x i8> %a) {
+; SSE-LABEL: shuffle_v16i8_uu_02_03_zz_uu_06_07_zz_uu_10_11_zz_uu_14_15_zz:
+; SSE:       # BB#0:
+; SSE-NEXT:    psrld $8, %xmm0
+; SSE-NEXT:    retq
+;
+; AVX-LABEL: shuffle_v16i8_uu_02_03_zz_uu_06_07_zz_uu_10_11_zz_uu_14_15_zz:
+; AVX:       # BB#0:
+; AVX-NEXT:    vpsrld $8, %xmm0, %xmm0
+; AVX-NEXT:    retq
+  %shuffle = shufflevector <16 x i8> %a, <16 x i8> zeroinitializer, <16 x i32> <i32 undef, i32 2, i32 3, i32 16, i32 undef, i32 6, i32 7, i32 16, i32 undef, i32 10, i32 11, i32 16, i32 undef, i32 14, i32 15, i32 16>
+  ret <16 x i8> %shuffle
+}