Optimization for SIGN_EXTEND operation on AVX.
Special handling was added for v4i32 -> v4i64 and v8i16 -> v8i32
extensions.
git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@149600 91177308-0d34-0410-b5e6-96231b3b80d8
diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 76b21ac..7b8221a 100644
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -1221,6 +1221,7 @@
setTargetDAGCombine(ISD::LOAD);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::ZERO_EXTEND);
+ setTargetDAGCombine(ISD::SIGN_EXTEND);
setTargetDAGCombine(ISD::TRUNCATE);
setTargetDAGCombine(ISD::SINT_TO_FP);
if (Subtarget->is64Bit())
@@ -14641,6 +14642,55 @@
return SDValue();
}
+static SDValue PerformSExtCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ if (!DCI.isBeforeLegalizeOps())
+ return SDValue();
+
+ if (!Subtarget->hasAVX()) return SDValue();
+
+ // Optimize vectors in AVX mode
+ // Sign extend v8i16 to v8i32 and
+ // v4i32 to v4i64
+ //
+ // Divide input vector into two parts
+ // for v4i32 the shuffle mask will be { 0, 1, -1, -1} {2, 3, -1, -1}
+ // use vpmovsx instruction to extend v4i32 -> v2i64; v8i16 -> v4i32
+ // concat the vectors to original VT
+
+ EVT VT = N->getValueType(0);
+ SDValue Op = N->getOperand(0);
+ EVT OpVT = Op.getValueType();
+ DebugLoc dl = N->getDebugLoc();
+
+ if (((VT == MVT::v4i64) && (OpVT == MVT::v4i32)) ||
+ ((VT == MVT::v8i32) && (OpVT == MVT::v8i16))) {
+
+ unsigned NumElems = OpVT.getVectorNumElements();
+ SmallVector<int,8> ShufMask1(NumElems, -1);
+ for (unsigned i=0; i< NumElems/2; i++) ShufMask1[i] = i;
+
+ SDValue OpLo = DAG.getVectorShuffle(OpVT, dl, Op, DAG.getUNDEF(OpVT),
+ ShufMask1.data());
+
+ SmallVector<int,8> ShufMask2(NumElems, -1);
+ for (unsigned i=0; i< NumElems/2; i++) ShufMask2[i] = i+NumElems/2;
+
+ SDValue OpHi = DAG.getVectorShuffle(OpVT, dl, Op, DAG.getUNDEF(OpVT),
+ ShufMask2.data());
+
+ EVT HalfVT = EVT::getVectorVT(*DAG.getContext(), VT.getScalarType(),
+ VT.getVectorNumElements()/2);
+
+ OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
+ OpHi = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpHi);
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
+ }
+ return SDValue();
+}
+
static SDValue PerformZExtCombine(SDNode *N, SelectionDAG &DAG,
const X86Subtarget *Subtarget) {
// (i32 zext (and (i8 x86isd::setcc_carry), 1)) ->
@@ -14886,6 +14936,7 @@
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG, Subtarget);
+ case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG, DCI);
case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
case X86ISD::SHUFP: // Handle all target specific shuffles
diff --git a/lib/Target/X86/X86ISelLowering.h b/lib/Target/X86/X86ISelLowering.h
index 1a38921..9689bcd 100644
--- a/lib/Target/X86/X86ISelLowering.h
+++ b/lib/Target/X86/X86ISelLowering.h
@@ -219,6 +219,9 @@
// VZEXT_MOVL - Vector move low and zero extend.
VZEXT_MOVL,
+ // VZEXT_MOVL - Vector move low and sign extend.
+ VSEXT_MOVL,
+
// VSHL, VSRL - 128-bit vector logical left / right shift
VSHLDQ, VSRLDQ,
diff --git a/lib/Target/X86/X86InstrFragmentsSIMD.td b/lib/Target/X86/X86InstrFragmentsSIMD.td
index 98ed34a..f239509 100644
--- a/lib/Target/X86/X86InstrFragmentsSIMD.td
+++ b/lib/Target/X86/X86InstrFragmentsSIMD.td
@@ -71,6 +71,9 @@
SDTCisVT<2, v4f32>, SDTCisPtrTy<3>]>>;
def X86vzmovl : SDNode<"X86ISD::VZEXT_MOVL",
SDTypeProfile<1, 1, [SDTCisSameAs<0,1>]>>;
+def X86vsmovl : SDNode<"X86ISD::VSEXT_MOVL",
+ SDTypeProfile<1, 1, [SDTCisVec<0>, SDTCisInt<1>, SDTCisInt<0>]>>;
+
def X86vzload : SDNode<"X86ISD::VZEXT_LOAD", SDTLoad,
[SDNPHasChain, SDNPMayLoad, SDNPMemOperand]>;
def X86vshldq : SDNode<"X86ISD::VSHLDQ", SDTIntShiftOp>;
diff --git a/lib/Target/X86/X86InstrSSE.td b/lib/Target/X86/X86InstrSSE.td
index 8f6df30..34478a0 100644
--- a/lib/Target/X86/X86InstrSSE.td
+++ b/lib/Target/X86/X86InstrSSE.td
@@ -5478,6 +5478,16 @@
(PMOVZXDQrm addr:$src)>;
}
+let Predicates = [HasAVX] in {
+def : Pat<(v2i64 (X86vsmovl (v4i32 VR128:$src))), (VPMOVSXDQrr VR128:$src)>;
+def : Pat<(v4i32 (X86vsmovl (v8i16 VR128:$src))), (VPMOVSXWDrr VR128:$src)>;
+}
+
+let Predicates = [HasSSE41] in {
+def : Pat<(v2i64 (X86vsmovl (v4i32 VR128:$src))), (PMOVSXDQrr VR128:$src)>;
+def : Pat<(v4i32 (X86vsmovl (v8i16 VR128:$src))), (PMOVSXWDrr VR128:$src)>;
+}
+
multiclass SS41I_binop_rm_int4<bits<8> opc, string OpcodeStr, Intrinsic IntId> {
def rr : SS48I<opc, MRMSrcReg, (outs VR128:$dst), (ins VR128:$src),