Teach the x86 backend a couple tricks for dealing with v16i8 sra by a constant splat value. Fixes PR11289.

llvm-svn: 143498

2 files changed, 38 insertions, 0 deletions

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 2ddb1b7163e..c9b642242ae 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -966,6 +966,7 @@ X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
 
     setOperationAction(ISD::SRA,               MVT::v4i32, Custom);
     setOperationAction(ISD::SRA,               MVT::v8i16, Custom);
+    setOperationAction(ISD::SRA,               MVT::v16i8, Custom);
   }
 
   if (Subtarget->hasSSE42() || Subtarget->hasAVX())
@@ -9994,6 +9995,23 @@ SDValue X86TargetLowering::LowerShift(SDValue Op, SelectionDAG &DAG) const {
        return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
                      DAG.getConstant(Intrinsic::x86_sse2_psrai_w, MVT::i32),
                      R, DAG.getConstant(ShiftAmt, MVT::i32));
+
+      if (VT == MVT::v16i8 && Op.getOpcode() == ISD::SRA) {
+        if (ShiftAmt == 7) {
+          // R s>> 7  ===  R s< 0
+          SDValue Zeros = getZeroVector(VT, true /* HasXMMInt */, DAG, dl);
+          return DAG.getNode(X86ISD::PCMPGTB, dl, VT, Zeros, R);
+        }
+
+        // R s>> a === ((R u>> a) ^ m) - m
+        SDValue Res = DAG.getNode(ISD::SRL, dl, VT, R, Amt);
+        SmallVector<SDValue, 16> V(16, DAG.getConstant(128 >> ShiftAmt,
+                                                       MVT::i8));
+        SDValue Mask = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], 16);
+        Res = DAG.getNode(ISD::XOR, dl, VT, Res, Mask);
+        Res = DAG.getNode(ISD::SUB, dl, VT, Res, Mask);
+        return Res;
+      }
     }
   }
 
diff --git a/llvm/test/CodeGen/X86/x86-shifts.ll b/llvm/test/CodeGen/X86/x86-shifts.ll
index 3e44eafa240..20bccab8ff7 100644
--- a/llvm/test/CodeGen/X86/x86-shifts.ll
+++ b/llvm/test/CodeGen/X86/x86-shifts.ll
@@ -170,3 +170,23 @@ define <16 x i8> @shr9(<16 x i8> %A) nounwind {
 ; CHECK: pand
 ; CHECK: ret
 }
+
+define <16 x i8> @sra_v16i8_7(<16 x i8> %A) nounwind {
+  %B = ashr <16 x i8> %A, <i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7, i8 7>
+  ret <16 x i8> %B
+; CHECK: sra_v16i8_7:
+; CHECK: pxor
+; CHECK: pcmpgtb
+; CHECK: ret
+}
+
+define <16 x i8> @sra_v16i8(<16 x i8> %A) nounwind {
+  %B = ashr <16 x i8> %A, <i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3, i8 3>
+  ret <16 x i8> %B
+; CHECK: sra_v16i8:
+; CHECK: psrlw $3
+; CHECK: pand
+; CHECK: pxor
+; CHECK: psubb
+; CHECK: ret
+}