PPC: Implement efficient lowering of sign_extend_inreg.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@172269 91177308-0d34-0410-b5e6-96231b3b80d8

lib/CodeGen/SelectionDAG/LegalizeVectorOps.cpp

@@ -61,6 +61,8 @@ class VectorLegalizer {
   // Implements expansion for UINT_TO_FLOAT; falls back to UnrollVectorOp if
   // SINT_TO_FLOAT and SHR on vectors isn't legal.
   SDValue ExpandUINT_TO_FLOAT(SDValue Op);
+  // Implement expansion for SIGN_EXTEND_INREG using SRL and SRA.
+  SDValue ExpandSEXTINREG(SDValue Op);
   // Implement vselect in terms of XOR, AND, OR when blend is not supported
   // by the target.
   SDValue ExpandVSELECT(SDValue Op);
@@ -262,7 +264,9 @@ SDValue VectorLegalizer::LegalizeOp(SDValue Op) {
     // FALL THROUGH
   }
   case TargetLowering::Expand:
-    if (Node->getOpcode() == ISD::VSELECT)
+    if (Node->getOpcode() == ISD::SIGN_EXTEND_INREG)
+      Result = ExpandSEXTINREG(Op);
+    else if (Node->getOpcode() == ISD::VSELECT)
       Result = ExpandVSELECT(Op);
     else if (Node->getOpcode() == ISD::SELECT)
       Result = ExpandSELECT(Op);
@@ -501,6 +505,26 @@ SDValue VectorLegalizer::ExpandSELECT(SDValue Op) {
   return DAG.getNode(ISD::BITCAST, DL, Op.getValueType(), Val);
 }
 
+SDValue VectorLegalizer::ExpandSEXTINREG(SDValue Op) {
+  EVT VT = Op.getValueType();
+
+  // Make sure that the SRA and SRL instructions are available.
+  if (TLI.getOperationAction(ISD::SRA, VT) == TargetLowering::Expand ||
+      TLI.getOperationAction(ISD::SRL, VT) == TargetLowering::Expand)
+    return DAG.UnrollVectorOp(Op.getNode());
+
+  DebugLoc DL = Op.getDebugLoc();
+  EVT OrigTy = cast<VTSDNode>(Op->getOperand(1))->getVT();
+
+  unsigned BW = VT.getScalarType().getSizeInBits();
+  unsigned OrigBW = OrigTy.getScalarType().getSizeInBits();
+  SDValue ShiftSz = DAG.getConstant(BW - OrigBW, VT);
+
+  Op = Op.getOperand(0);
+  Op =   DAG.getNode(ISD::SRL, DL, VT, Op, ShiftSz);
+  return DAG.getNode(ISD::SRA, DL, VT, Op, ShiftSz);
+}
+
 SDValue VectorLegalizer::ExpandVSELECT(SDValue Op) {
   // Implement VSELECT in terms of XOR, AND, OR
   // on platforms which do not support blend natively.

test/CodeGen/PowerPC/vec_extload.ll

@@ -15,55 +15,9 @@ define <16 x i8> @v16si8_sext_in_reg(<16 x i8> %a) {
   ret <16 x i8> %c
 }
 ; CHECK: v16si8_sext_in_reg:
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
+; CHECK: vsrb
-; CHECK: lbz
+; CHECK: vsrab
-; CHECK: stb
+; CHECK: blr 
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lbz
-; CHECK: stb
-; CHECK: lvx 2, {{[0-9]+}}, {{[0-9]+}}
 
 ; The zero extend uses a more clever logic: a vector splat
 ; and a logic and to set higher bits to 0.
@@ -83,31 +37,9 @@ define <8 x i16> @v8si16_sext_in_reg(<8 x i16> %a) {
   ret <8 x i16> %c
 }
 ; CHECK: v8si16_sext_in_reg:
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
+; CHECK: vsrh
-; CHECK: lhz
+; CHECK: vsrah
-; CHECK: sth
+; CHECK: blr 
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lhz
-; CHECK: sth
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lhz
-; CHECK: sth
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lhz
-; CHECK: sth
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lhz
-; CHECK: sth
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lhz
-; CHECK: sth
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lhz
-; CHECK: sth
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lhz
-; CHECK: sth
-; CHECK: lvx 2, {{[0-9]+}}, {{[0-9]+}}
 
 ; Same as v8si16_sext_in_reg, but instead of creating the mask
 ; with a splat, loads it from memory.
@@ -129,19 +61,9 @@ define <4 x i32> @v4si32_sext_in_reg(<4 x i32> %a) {
   ret <4 x i32> %c
 }
 ; CHECK: v4si32_sext_in_reg:
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
+; CHECK: vsrw
-; CHECK: lha
+; CHECK: vsraw
-; CHECK: stw
+; CHECK: blr 
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lha
-; CHECK: stw
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lha
-; CHECK: stw
-; CHECK: stvx 2, {{[0-9]+}}, {{[0-9]+}}
-; CHECK: lha
-; CHECK: stw
-; CHECK: lvx 2, {{[0-9]+}}, {{[0-9]+}}
 
 ; Same as v8si16_sext_in_reg.
 define <4 x i32> @v4si32_zext_in_reg(<4 x i32> %a) {