[X86][SSE] Combine (VSRLI (VSRAI X, Y), (NumSignBits-1)) -> (VSRLI X, (NumSignBits-1))

Part 3 of 3.

Differential Revision: https://reviews.llvm.org/D31347

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@298782 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 547250375aa242bbace811e47a8eba3687bc2969..02e39dadea715922a62d296899e2f04e584d63b6 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -31075,13 +31075,14 @@ static SDValue combineVectorShift(SDNode *N, SelectionDAG &DAG,
   bool LogicalShift = X86ISD::VSHLI == Opcode || X86ISD::VSRLI == Opcode;
   EVT VT = N->getValueType(0);
   SDValue N0 = N->getOperand(0);
+  SDValue N1 = N->getOperand(1);
   unsigned NumBitsPerElt = VT.getScalarSizeInBits();
   assert(VT == N0.getValueType() && (NumBitsPerElt % 8) == 0 &&
          "Unexpected value type");
 
   // Out of range logical bit shifts are guaranteed to be zero.
   // Out of range arithmetic bit shifts splat the sign bit.
-  APInt ShiftVal = cast<ConstantSDNode>(N->getOperand(1))->getAPIntValue();
+  APInt ShiftVal = cast<ConstantSDNode>(N1)->getAPIntValue();
   if (ShiftVal.zextOrTrunc(8).uge(NumBitsPerElt)) {
     if (LogicalShift)
       return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, SDLoc(N));
@@ -31097,6 +31098,13 @@ static SDValue combineVectorShift(SDNode *N, SelectionDAG &DAG,
   if (ISD::isBuildVectorAllZeros(N0.getNode()))
     return getZeroVector(VT.getSimpleVT(), Subtarget, DAG, SDLoc(N));
 
+  // fold (VSRLI (VSRAI X, Y), 31) -> (VSRLI X, 31).
+  // This VSRLI only looks at the sign bit, which is unmodified by VSRAI.
+  // TODO - support other sra opcodes as needed.
+  if (Opcode == X86ISD::VSRLI && (ShiftVal + 1) == NumBitsPerElt &&
+      N0.getOpcode() == X86ISD::VSRAI)
+    return DAG.getNode(X86ISD::VSRLI, SDLoc(N), VT, N0.getOperand(0), N1);
+
   // We can decode 'whole byte' logical bit shifts as shuffles.
   if (LogicalShift && (ShiftVal.getZExtValue() % 8) == 0) {
     SDValue Op(N, 0);

diff --git a/test/CodeGen/X86/combine-and.ll b/test/CodeGen/X86/combine-and.ll
index 352705b48d9db8b1bd34ec199d71405a5f9b1dd9..f30fa61bbfbe65abd671e23eaac3543402365705 100644 (file)
--- a/test/CodeGen/X86/combine-and.ll
+++ b/test/CodeGen/X86/combine-and.ll
@@ -253,7 +253,6 @@ define <4 x i32> @and_or_zext_v4i16(<4 x i16> %a0) {
 define <8 x i16> @ashr_mask1_v8i16(<8 x i16> %a0) {
 ; CHECK-LABEL: ashr_mask1_v8i16:
 ; CHECK:       # BB#0:
-; CHECK-NEXT:    psraw $15, %xmm0
 ; CHECK-NEXT:    psrlw $15, %xmm0
 ; CHECK-NEXT:    retq
   %1 = ashr <8 x i16> %a0, <i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15, i16 15>