#define DEBUG_TYPE "x86-isel"
#include "X86ISelLowering.h"
+#include "Utils/X86ShuffleDecode.h"
#include "X86.h"
#include "X86InstrBuilder.h"
#include "X86TargetMachine.h"
#include "X86TargetObjectFile.h"
-#include "Utils/X86ShuffleDecode.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/ADT/StringExtras.h"
+#include "llvm/ADT/VariadicFunction.h"
#include "llvm/CallingConv.h"
-#include "llvm/Constants.h"
-#include "llvm/DerivedTypes.h"
-#include "llvm/GlobalAlias.h"
-#include "llvm/GlobalVariable.h"
-#include "llvm/Function.h"
-#include "llvm/Instructions.h"
-#include "llvm/Intrinsics.h"
-#include "llvm/LLVMContext.h"
#include "llvm/CodeGen/IntrinsicLowering.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/Constants.h"
+#include "llvm/DerivedTypes.h"
+#include "llvm/Function.h"
+#include "llvm/GlobalAlias.h"
+#include "llvm/GlobalVariable.h"
+#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
+#include "llvm/LLVMContext.h"
#include "llvm/MC/MCAsmInfo.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCSymbol.h"
-#include "llvm/ADT/SmallSet.h"
-#include "llvm/ADT/Statistic.h"
-#include "llvm/ADT/StringExtras.h"
-#include "llvm/ADT/VariadicFunction.h"
#include "llvm/Support/CallSite.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
setOperationAction(ISD::ADD, MVT::v8i16, Legal);
setOperationAction(ISD::ADD, MVT::v4i32, Legal);
setOperationAction(ISD::ADD, MVT::v2i64, Legal);
+ setOperationAction(ISD::MUL, MVT::v4i32, Custom);
setOperationAction(ISD::MUL, MVT::v2i64, Custom);
setOperationAction(ISD::SUB, MVT::v16i8, Legal);
setOperationAction(ISD::SUB, MVT::v8i16, Legal);
setOperationAction(ISD::FABS, MVT::v4f64, Custom);
setOperationAction(ISD::TRUNCATE, MVT::v8i16, Custom);
+ setOperationAction(ISD::TRUNCATE, MVT::v4i32, Custom);
setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
setOperationAction(ISD::VSELECT, MVT::v8i32, Legal);
setOperationAction(ISD::VSELECT, MVT::v8f32, Legal);
+ setOperationAction(ISD::SIGN_EXTEND, MVT::v4i64, Custom);
+ setOperationAction(ISD::SIGN_EXTEND, MVT::v8i32, Custom);
+ setOperationAction(ISD::ZERO_EXTEND, MVT::v4i64, Custom);
+ setOperationAction(ISD::ZERO_EXTEND, MVT::v8i32, Custom);
+ setOperationAction(ISD::ANY_EXTEND, MVT::v4i64, Custom);
+ setOperationAction(ISD::ANY_EXTEND, MVT::v8i32, Custom);
+
if (Subtarget->hasFMA() || Subtarget->hasFMA4()) {
setOperationAction(ISD::FMA, MVT::v8f32, Legal);
setOperationAction(ISD::FMA, MVT::v4f64, Legal);
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
-
// Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
// handle type legalization for these operations here.
//
setPrefFunctionAlignment(4); // 2^4 bytes.
}
-
EVT X86TargetLowering::getSetCCResultType(EVT VT) const {
if (!VT.isVector()) return MVT::i8;
return VT.changeVectorElementTypeToInteger();
}
-
/// getMaxByValAlign - Helper for getByValTypeAlignment to determine
/// the desired ByVal argument alignment.
static void getMaxByValAlign(Type *Ty, unsigned &MaxAlign) {
/// lowering. If DstAlign is zero that means it's safe to destination
/// alignment can satisfy any constraint. Similarly if SrcAlign is zero it
/// means there isn't a need to check it against alignment requirement,
-/// probably because the source does not need to be loaded. If
-/// 'IsZeroVal' is true, that means it's safe to return a
-/// non-scalar-integer type, e.g. empty string source, constant, or loaded
-/// from memory. 'MemcpyStrSrc' indicates whether the memcpy source is
-/// constant so it does not need to be loaded.
+/// probably because the source does not need to be loaded. If 'IsMemset' is
+/// true, that means it's expanding a memset. If 'ZeroMemset' is true, that
+/// means it's a memset of zero. 'MemcpyStrSrc' indicates whether the memcpy
+/// source is constant so it does not need to be loaded.
/// It returns EVT::Other if the type should be determined using generic
/// target-independent logic.
EVT
X86TargetLowering::getOptimalMemOpType(uint64_t Size,
unsigned DstAlign, unsigned SrcAlign,
- bool IsZeroVal,
+ bool IsMemset, bool ZeroMemset,
bool MemcpyStrSrc,
MachineFunction &MF) const {
const Function *F = MF.getFunction();
- if (IsZeroVal &&
- !F->getFnAttributes().hasAttribute(Attributes::NoImplicitFloat)) {
+ if ((!IsMemset || ZeroMemset) &&
+ !F->getFnAttributes().hasAttribute(Attribute::NoImplicitFloat)) {
if (Size >= 16 &&
(Subtarget->isUnalignedMemAccessFast() ||
((DstAlign == 0 || DstAlign >= 16) &&
return MVT::i32;
}
+bool X86TargetLowering::isSafeMemOpType(MVT VT) const {
+ if (VT == MVT::f32)
+ return X86ScalarSSEf32;
+ else if (VT == MVT::f64)
+ return X86ScalarSSEf64;
+ return true;
+}
+
+bool
+X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
+ if (Fast)
+ *Fast = Subtarget->isUnalignedMemAccessFast();
+ return true;
+}
+
/// getJumpTableEncoding - Return the entry encoding for a jump table in the
/// current function. The returned value is a member of the
/// MachineJumpTableInfo::JTEntryKind enum.
// FIXME: Why this routine is here? Move to RegInfo!
std::pair<const TargetRegisterClass*, uint8_t>
-X86TargetLowering::findRepresentativeClass(EVT VT) const{
+X86TargetLowering::findRepresentativeClass(MVT VT) const{
const TargetRegisterClass *RRC = 0;
uint8_t Cost = 1;
- switch (VT.getSimpleVT().SimpleTy) {
+ switch (VT.SimpleTy) {
default:
return TargetLowering::findRepresentativeClass(VT);
case MVT::i8: case MVT::i16: case MVT::i32: case MVT::i64:
return true;
}
-
//===----------------------------------------------------------------------===//
// Return Value Calling Convention Implementation
//===----------------------------------------------------------------------===//
return true;
}
-EVT
-X86TargetLowering::getTypeForExtArgOrReturn(LLVMContext &Context, EVT VT,
+MVT
+X86TargetLowering::getTypeForExtArgOrReturn(MVT VT,
ISD::NodeType ExtendKind) const {
MVT ReturnMVT;
// TODO: Is this also valid on 32-bit?
else
ReturnMVT = MVT::i32;
- EVT MinVT = getRegisterType(Context, ReturnMVT);
+ MVT MinVT = getRegisterType(ReturnMVT);
return VT.bitsLT(MinVT) ? MinVT : VT;
}
CCInfo.AnalyzeCallResult(Ins, RetCC_X86);
// Copy all of the result registers out of their specified physreg.
- for (unsigned i = 0; i != RVLocs.size(); ++i) {
+ for (unsigned i = 0, e = RVLocs.size(); i != e; ++i) {
CCValAssign &VA = RVLocs[i];
EVT CopyVT = VA.getValVT();
return Chain;
}
-
//===----------------------------------------------------------------------===//
// C & StdCall & Fast Calling Convention implementation
//===----------------------------------------------------------------------===//
if (VA.isExtInLoc()) {
// Handle MMX values passed in XMM regs.
- if (RegVT.isVector()) {
- ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(),
- ArgValue);
- } else
+ if (RegVT.isVector())
+ ArgValue = DAG.getNode(X86ISD::MOVDQ2Q, dl, VA.getValVT(), ArgValue);
+ else
ArgValue = DAG.getNode(ISD::TRUNCATE, dl, VA.getValVT(), ArgValue);
}
} else {
TotalNumIntRegs);
bool NoImplicitFloatOps = Fn->getFnAttributes().
- hasAttribute(Attributes::NoImplicitFloat);
+ hasAttribute(Attribute::NoImplicitFloat);
assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
"SSE register cannot be used when SSE is disabled!");
assert(!(NumXMMRegs && MF.getTarget().Options.UseSoftFloat &&
} else if (Subtarget->isPICStyleRIPRel() &&
isa<Function>(GV) &&
cast<Function>(GV)->getFnAttributes().
- hasAttribute(Attributes::NonLazyBind)) {
+ hasAttribute(Attribute::NonLazyBind)) {
// If the function is marked as non-lazy, generate an indirect call
// which loads from the GOT directly. This avoids runtime overhead
// at the cost of eager binding (and one extra byte of encoding).
Ins, dl, DAG, InVals);
}
-
//===----------------------------------------------------------------------===//
// Fast Calling Convention (tail call) implementation
//===----------------------------------------------------------------------===//
return X86::createFastISel(funcInfo, libInfo);
}
-
//===----------------------------------------------------------------------===//
// Other Lowering Hooks
//===----------------------------------------------------------------------===//
return DAG.getFrameIndex(ReturnAddrIndex, getPointerTy());
}
-
bool X86::isOffsetSuitableForCodeModel(int64_t Offset, CodeModel::Model M,
bool hasSymbolicDisplacement) {
// Offset should fit into 32 bit immediate field.
/// isUndefOrEqual - Val is either less than zero (undef) or equal to the
/// specified value.
static bool isUndefOrEqual(int Val, int CmpVal) {
- if (Val < 0 || Val == CmpVal)
- return true;
- return false;
+ return (Val < 0 || Val == CmpVal);
}
/// isSequentialOrUndefInRange - Return true if every element in Mask, beginning
SDValue V1 = SVOp->getOperand(0);
SDValue V2 = SVOp->getOperand(1);
DebugLoc dl = SVOp->getDebugLoc();
- MVT VT = SVOp->getValueType(0).getSimpleVT();
+ EVT VT = SVOp->getValueType(0);
+ EVT EltVT = VT.getVectorElementType();
unsigned NumElems = VT.getVectorNumElements();
- if (!Subtarget->hasSSE41())
+ if (!Subtarget->hasSSE41() || EltVT == MVT::i8)
+ return SDValue();
+ if (!Subtarget->hasInt256() && VT == MVT::v16i16)
return SDValue();
- unsigned ISDNo = 0;
- MVT OpTy;
-
- switch (VT.SimpleTy) {
- default: return SDValue();
- case MVT::v8i16:
- ISDNo = X86ISD::BLENDPW;
- OpTy = MVT::v8i16;
- break;
- case MVT::v4i32:
- case MVT::v4f32:
- ISDNo = X86ISD::BLENDPS;
- OpTy = MVT::v4f32;
- break;
- case MVT::v2i64:
- case MVT::v2f64:
- ISDNo = X86ISD::BLENDPD;
- OpTy = MVT::v2f64;
- break;
- case MVT::v8i32:
- case MVT::v8f32:
- if (!Subtarget->hasFp256())
- return SDValue();
- ISDNo = X86ISD::BLENDPS;
- OpTy = MVT::v8f32;
- break;
- case MVT::v4i64:
- case MVT::v4f64:
- if (!Subtarget->hasFp256())
- return SDValue();
- ISDNo = X86ISD::BLENDPD;
- OpTy = MVT::v4f64;
- break;
- }
- assert(ISDNo && "Invalid Op Number");
+ // Check the mask for BLEND and build the value.
+ unsigned MaskValue = 0;
+ // There are 2 lanes if (NumElems > 8), and 1 lane otherwise.
+ unsigned NumLanes = (NumElems-1)/8 + 1;
+ unsigned NumElemsInLane = NumElems / NumLanes;
- unsigned MaskVals = 0;
+ // Blend for v16i16 should be symetric for the both lanes.
+ for (unsigned i = 0; i < NumElemsInLane; ++i) {
- for (unsigned i = 0; i != NumElems; ++i) {
+ int SndLaneEltIdx = (NumLanes == 2) ?
+ SVOp->getMaskElt(i + NumElemsInLane) : -1;
int EltIdx = SVOp->getMaskElt(i);
- if (EltIdx == (int)i || EltIdx < 0)
- MaskVals |= (1<<i);
- else if (EltIdx == (int)(i + NumElems))
- continue; // Bit is set to zero;
- else
+
+ if ((EltIdx == -1 || EltIdx == (int)i) &&
+ (SndLaneEltIdx == -1 || SndLaneEltIdx == (int)(i + NumElemsInLane)))
+ continue;
+
+ if (((unsigned)EltIdx == (i + NumElems)) &&
+ (SndLaneEltIdx == -1 ||
+ (unsigned)SndLaneEltIdx == i + NumElems + NumElemsInLane))
+ MaskValue |= (1<<i);
+ else
return SDValue();
}
- V1 = DAG.getNode(ISD::BITCAST, dl, OpTy, V1);
- V2 = DAG.getNode(ISD::BITCAST, dl, OpTy, V2);
- SDValue Ret = DAG.getNode(ISDNo, dl, OpTy, V1, V2,
- DAG.getConstant(MaskVals, MVT::i32));
+ // Convert i32 vectors to floating point if it is not AVX2.
+ // AVX2 introduced VPBLENDD instruction for 128 and 256-bit vectors.
+ EVT BlendVT = VT;
+ if (EltVT == MVT::i64 || (EltVT == MVT::i32 && !Subtarget->hasInt256())) {
+ BlendVT = EVT::getVectorVT(*DAG.getContext(),
+ EVT::getFloatingPointVT(EltVT.getSizeInBits()),
+ NumElems);
+ V1 = DAG.getNode(ISD::BITCAST, dl, VT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, dl, VT, V2);
+ }
+
+ SDValue Ret = DAG.getNode(X86ISD::BLENDI, dl, BlendVT, V1, V2,
+ DAG.getConstant(MaskValue, MVT::i32));
return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
}
return MayFoldLoad(V);
}
-// FIXME: the version above should always be used. Since there's
-// a bug where several vector shuffles can't be folded because the
-// DAG is not updated during lowering and a node claims to have two
-// uses while it only has one, use this version, and let isel match
-// another instruction if the load really happens to have more than
-// one use. Remove this version after this bug get fixed.
-// rdar://8434668, PR8156
-static bool RelaxedMayFoldVectorLoad(SDValue V) {
- if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
- V = V.getOperand(0);
- if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
- V = V.getOperand(0);
- if (ISD::isNormalLoad(V.getNode()))
- return true;
- return false;
-}
-
static
SDValue getMOVDDup(SDValue &Op, DebugLoc &dl, SDValue V1, SelectionDAG &DAG) {
EVT VT = Op.getValueType();
bool HasInt256 = Subtarget->hasInt256();
MachineFunction &MF = DAG.getMachineFunction();
bool OptForSize = MF.getFunction()->getFnAttributes().
- hasAttribute(Attributes::OptimizeForSize);
+ hasAttribute(Attribute::OptimizeForSize);
assert(VT.getSizeInBits() != 64 && "Can't lower MMX shuffles");
return getTargetShuffleNode(X86ISD::UNPCKH, dl, VT, V1, V1, DAG);
if (isMOVDDUPMask(M, VT) && Subtarget->hasSSE3() &&
- V2IsUndef && RelaxedMayFoldVectorLoad(V1))
+ V2IsUndef && MayFoldVectorLoad(V1))
return getMOVDDup(Op, dl, V1, DAG);
if (isMOVHLPS_v_undef_Mask(M, VT))
unsigned TargetMask = getShuffleSHUFImmediate(SVOp);
- if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
- return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask, DAG);
-
if (HasSSE2 && (VT == MVT::v4f32 || VT == MVT::v4i32))
return getTargetShuffleNode(X86ISD::PSHUFD, dl, VT, V1, TargetMask, DAG);
+ if (HasFp256 && (VT == MVT::v4f32 || VT == MVT::v2f64))
+ return getTargetShuffleNode(X86ISD::VPERMILP, dl, VT, V1, TargetMask,
+ DAG);
+
return getTargetShuffleNode(X86ISD::SHUFP, dl, VT, V1, V1,
TargetMask, DAG);
}
return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1,
getShuffleCLImmediate(SVOp), DAG);
-
//===--------------------------------------------------------------------===//
// Since no target specific shuffle was selected for this generic one,
// lower it into other known shuffles. FIXME: this isn't true yet, but
return SDValue();
}
-
SDValue
X86TargetLowering::LowerEXTRACT_VECTOR_ELT(SDValue Op,
SelectionDAG &DAG) const {
DebugLoc DL = Op.getDebugLoc();
Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
-
// With PIC, the address is actually $g + Offset.
if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
!Subtarget->is64Bit()) {
llvm_unreachable("TLS not implemented for this target.");
}
-
/// LowerShiftParts - Lower SRA_PARTS and friends, which return two i32 values
/// and take a 2 x i32 value to shift plus a shift amount.
SDValue X86TargetLowering::LowerShiftParts(SDValue Op, SelectionDAG &DAG) const{
}
}
-SDValue X86TargetLowering::lowerZERO_EXTEND(SDValue Op, SelectionDAG &DAG) const {
+static SDValue LowerAVXExtend(SDValue Op, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ EVT VT = Op->getValueType(0);
+ SDValue In = Op->getOperand(0);
+ EVT InVT = In.getValueType();
+ DebugLoc dl = Op->getDebugLoc();
+
+ // Optimize vectors in AVX mode:
+ //
+ // v8i16 -> v8i32
+ // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
+ // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
+ // Concat upper and lower parts.
+ //
+ // v4i32 -> v4i64
+ // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
+ // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
+ // Concat upper and lower parts.
+ //
+
+ if (((VT != MVT::v8i32) || (InVT != MVT::v8i16)) &&
+ ((VT != MVT::v4i64) || (InVT != MVT::v4i32)))
+ return SDValue();
+
+ if (Subtarget->hasInt256())
+ return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, In);
+
+ SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
+ SDValue Undef = DAG.getUNDEF(InVT);
+ bool NeedZero = Op.getOpcode() == ISD::ZERO_EXTEND;
+ SDValue OpLo = getUnpackl(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
+ SDValue OpHi = getUnpackh(DAG, dl, InVT, In, NeedZero ? ZeroVec : Undef);
+
+ EVT HVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
+ VT.getVectorNumElements()/2);
+
+ OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
+ OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
+
+ return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
+}
+
+SDValue X86TargetLowering::LowerANY_EXTEND(SDValue Op,
+ SelectionDAG &DAG) const {
+ if (Subtarget->hasFp256()) {
+ SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
+ if (Res.getNode())
+ return Res;
+ }
+
+ return SDValue();
+}
+SDValue X86TargetLowering::LowerZERO_EXTEND(SDValue Op,
+ SelectionDAG &DAG) const {
DebugLoc DL = Op.getDebugLoc();
EVT VT = Op.getValueType();
SDValue In = Op.getOperand(0);
EVT SVT = In.getValueType();
+ if (Subtarget->hasFp256()) {
+ SDValue Res = LowerAVXExtend(Op, DAG, Subtarget);
+ if (Res.getNode())
+ return Res;
+ }
+
if (!VT.is256BitVector() || !SVT.is128BitVector() ||
VT.getVectorNumElements() != SVT.getVectorNumElements())
return SDValue();
SDValue Lo = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32, In);
static const int Mask[] = {4, 5, 6, 7, -1, -1, -1, -1};
SDValue Hi = DAG.getNode(X86ISD::VZEXT, DL, MVT::v4i32,
- DAG.getVectorShuffle(MVT::v8i16, DL, In, DAG.getUNDEF(MVT::v8i16), &Mask[0]));
+ DAG.getVectorShuffle(MVT::v8i16, DL, In,
+ DAG.getUNDEF(MVT::v8i16),
+ &Mask[0]));
return DAG.getNode(ISD::CONCAT_VECTORS, DL, MVT::v8i32, Lo, Hi);
}
SDValue X86TargetLowering::lowerTRUNCATE(SDValue Op, SelectionDAG &DAG) const {
DebugLoc DL = Op.getDebugLoc();
EVT VT = Op.getValueType();
- EVT SVT = Op.getOperand(0).getValueType();
+ SDValue In = Op.getOperand(0);
+ EVT SVT = In.getValueType();
- if (!VT.is128BitVector() || !SVT.is256BitVector() ||
- VT.getVectorNumElements() != SVT.getVectorNumElements())
+ if ((VT == MVT::v4i32) && (SVT == MVT::v4i64)) {
+ // On AVX2, v4i64 -> v4i32 becomes VPERMD.
+ if (Subtarget->hasInt256()) {
+ static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
+ In = DAG.getNode(ISD::BITCAST, DL, MVT::v8i32, In);
+ In = DAG.getVectorShuffle(MVT::v8i32, DL, In, DAG.getUNDEF(MVT::v8i32),
+ ShufMask);
+ return DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, VT, In,
+ DAG.getIntPtrConstant(0));
+ }
+
+ // On AVX, v4i64 -> v4i32 becomes a sequence that uses PSHUFD and MOVLHPS.
+ SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
+ DAG.getIntPtrConstant(0));
+ SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
+ DAG.getIntPtrConstant(2));
+
+ OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
+ OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
+
+ // The PSHUFD mask:
+ static const int ShufMask1[] = {0, 2, 0, 0};
+ SDValue Undef = DAG.getUNDEF(VT);
+ OpLo = DAG.getVectorShuffle(VT, DL, OpLo, Undef, ShufMask1);
+ OpHi = DAG.getVectorShuffle(VT, DL, OpHi, Undef, ShufMask1);
+
+ // The MOVLHPS mask:
+ static const int ShufMask2[] = {0, 1, 4, 5};
+ return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask2);
+ }
+
+ if ((VT == MVT::v8i16) && (SVT == MVT::v8i32)) {
+ // On AVX2, v8i32 -> v8i16 becomed PSHUFB.
+ if (Subtarget->hasInt256()) {
+ In = DAG.getNode(ISD::BITCAST, DL, MVT::v32i8, In);
+
+ SmallVector<SDValue,32> pshufbMask;
+ for (unsigned i = 0; i < 2; ++i) {
+ pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
+ pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
+ for (unsigned j = 0; j < 8; ++j)
+ pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
+ }
+ SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, DL, MVT::v32i8,
+ &pshufbMask[0], 32);
+ In = DAG.getNode(X86ISD::PSHUFB, DL, MVT::v32i8, In, BV);
+ In = DAG.getNode(ISD::BITCAST, DL, MVT::v4i64, In);
+
+ static const int ShufMask[] = {0, 2, -1, -1};
+ In = DAG.getVectorShuffle(MVT::v4i64, DL, In, DAG.getUNDEF(MVT::v4i64),
+ &ShufMask[0]);
+ In = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v2i64, In,
+ DAG.getIntPtrConstant(0));
+ return DAG.getNode(ISD::BITCAST, DL, VT, In);
+ }
+
+ SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
+ DAG.getIntPtrConstant(0));
+
+ SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, DL, MVT::v4i32, In,
+ DAG.getIntPtrConstant(4));
+
+ OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpLo);
+ OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, OpHi);
+
+ // The PSHUFB mask:
+ static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
+ -1, -1, -1, -1, -1, -1, -1, -1};
+
+ SDValue Undef = DAG.getUNDEF(MVT::v16i8);
+ OpLo = DAG.getVectorShuffle(MVT::v16i8, DL, OpLo, Undef, ShufMask1);
+ OpHi = DAG.getVectorShuffle(MVT::v16i8, DL, OpHi, Undef, ShufMask1);
+
+ OpLo = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpLo);
+ OpHi = DAG.getNode(ISD::BITCAST, DL, MVT::v4i32, OpHi);
+
+ // The MOVLHPS Mask:
+ static const int ShufMask2[] = {0, 1, 4, 5};
+ SDValue res = DAG.getVectorShuffle(MVT::v4i32, DL, OpLo, OpHi, ShufMask2);
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v8i16, res);
+ }
+
+ // Handle truncation of V256 to V128 using shuffles.
+ if (!VT.is128BitVector() || !SVT.is256BitVector())
return SDValue();
- assert(Subtarget->hasFp256() && "256-bit vector is observed without AVX!");
+ assert(VT.getVectorNumElements() != SVT.getVectorNumElements() &&
+ "Invalid op");
+ assert(Subtarget->hasFp256() && "256-bit vector without AVX!");
unsigned NumElems = VT.getVectorNumElements();
EVT NVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
NumElems * 2);
- SDValue In = Op.getOperand(0);
SmallVector<int, 16> MaskVec(NumElems * 2, -1);
// Prepare truncation shuffle mask
for (unsigned i = 0; i != NumElems; ++i)
return DAG.getNode(X86ISD::SAHF, dl, MVT::i32, TruncSrl);
}
+static bool isAllOnes(SDValue V) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
+ return C && C->isAllOnesValue();
+}
+
/// LowerToBT - Result of 'and' is compared against zero. Turn it into a BT node
/// if it's possible.
SDValue X86TargetLowering::LowerToBT(SDValue And, ISD::CondCode CC,
}
if (LHS.getNode()) {
+ // If the LHS is of the form (x ^ -1) then replace the LHS with x and flip
+ // the condition code later.
+ bool Invert = false;
+ if (LHS.getOpcode() == ISD::XOR && isAllOnes(LHS.getOperand(1))) {
+ Invert = true;
+ LHS = LHS.getOperand(0);
+ }
+
// If LHS is i8, promote it to i32 with any_extend. There is no i8 BT
// instruction. Since the shift amount is in-range-or-undefined, we know
// that doing a bittest on the i32 value is ok. We extend to i32 because
RHS = DAG.getNode(ISD::ANY_EXTEND, dl, LHS.getValueType(), RHS);
SDValue BT = DAG.getNode(X86ISD::BT, dl, MVT::i32, LHS, RHS);
- unsigned Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
+ X86::CondCode Cond = CC == ISD::SETEQ ? X86::COND_AE : X86::COND_B;
+ // Flip the condition if the LHS was a not instruction
+ if (Invert)
+ Cond = X86::GetOppositeBranchCondition(Cond);
return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
DAG.getConstant(Cond, MVT::i8), BT);
}
DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
}
-
SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
SDValue Cond;
SDValue Op0 = Op.getOperand(0);
if (VT == MVT::v2i64) {
if (Opc == X86ISD::PCMPGT && !Subtarget->hasSSE42())
return SDValue();
- if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41())
- return SDValue();
+ if (Opc == X86ISD::PCMPEQ && !Subtarget->hasSSE41()) {
+ // If pcmpeqq is missing but pcmpeqd is available synthesize pcmpeqq with
+ // pcmpeqd + pshufd + pand.
+ assert(Subtarget->hasSSE2() && !FlipSigns && "Don't know how to lower!");
+
+ // First cast everything to the right type,
+ Op0 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op0);
+ Op1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, Op1);
+
+ // Do the compare.
+ SDValue Result = DAG.getNode(Opc, dl, MVT::v4i32, Op0, Op1);
+
+ // Make sure the lower and upper halves are both all-ones.
+ const int Mask[] = { 1, 0, 3, 2 };
+ SDValue Shuf = DAG.getVectorShuffle(MVT::v4i32, dl, Result, Result, Mask);
+ Result = DAG.getNode(ISD::AND, dl, MVT::v4i32, Result, Shuf);
+
+ if (Invert)
+ Result = DAG.getNOT(dl, Result, MVT::v4i32);
+
+ return DAG.getNode(ISD::BITCAST, dl, VT, Result);
+ }
}
// Since SSE has no unsigned integer comparisons, we need to flip the sign
return C && C->isNullValue();
}
-static bool isAllOnes(SDValue V) {
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
- return C && C->isAllOnesValue();
-}
-
static bool isTruncWithZeroHighBitsInput(SDValue V, SelectionDAG &DAG) {
if (V.getOpcode() != ISD::TRUNCATE)
return false;
return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
}
+SDValue X86TargetLowering::LowerSIGN_EXTEND(SDValue Op,
+ SelectionDAG &DAG) const {
+ EVT VT = Op->getValueType(0);
+ SDValue In = Op->getOperand(0);
+ EVT InVT = In.getValueType();
+ DebugLoc dl = Op->getDebugLoc();
+
+ if ((VT != MVT::v4i64 || InVT != MVT::v4i32) &&
+ (VT != MVT::v8i32 || InVT != MVT::v8i16))
+ return SDValue();
+
+ if (Subtarget->hasInt256())
+ return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, In);
+
+ // 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
+
+ unsigned NumElems = InVT.getVectorNumElements();
+ SDValue Undef = DAG.getUNDEF(InVT);
+
+ SmallVector<int,8> ShufMask1(NumElems, -1);
+ for (unsigned i = 0; i != NumElems/2; ++i)
+ ShufMask1[i] = i;
+
+ SDValue OpLo = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask1[0]);
+
+ SmallVector<int,8> ShufMask2(NumElems, -1);
+ for (unsigned i = 0; i != NumElems/2; ++i)
+ ShufMask2[i] = i + NumElems/2;
+
+ SDValue OpHi = DAG.getVectorShuffle(InVT, dl, In, Undef, &ShufMask2[0]);
+
+ 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);
+}
+
// isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
// ISD::OR of two X86ISD::SETCC nodes each of which has no other use apart
// from the AND / OR.
Chain, Dest, CC, Cond);
}
-
// Lower dynamic stack allocation to _alloca call for Cygwin/Mingw targets.
// Calls to _alloca is needed to probe the stack when allocating more than 4k
// bytes in one go. Touching the stack at 4K increments is necessary to ensure
assert(!getTargetMachine().Options.UseSoftFloat &&
!(DAG.getMachineFunction()
.getFunction()->getFnAttributes()
- .hasAttribute(Attributes::NoImplicitFloat)) &&
+ .hasAttribute(Attribute::NoImplicitFloat)) &&
Subtarget->hasSSE1());
}
return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
Op.getOperand(1), Op.getOperand(2));
+ // SSE2/AVX2 sub with unsigned saturation intrinsics
+ case Intrinsic::x86_sse2_psubus_b:
+ case Intrinsic::x86_sse2_psubus_w:
+ case Intrinsic::x86_avx2_psubus_b:
+ case Intrinsic::x86_avx2_psubus_w:
+ return DAG.getNode(X86ISD::SUBUS, dl, Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2));
+
// SSE3/AVX horizontal add/sub intrinsics
case Intrinsic::x86_sse3_hadd_ps:
case Intrinsic::x86_sse3_hadd_pd:
Op.getOperand(1), Op.getOperand(2));
}
+ // SSE2/SSE41/AVX2 integer max/min intrinsics.
+ case Intrinsic::x86_sse2_pmaxu_b:
+ case Intrinsic::x86_sse41_pmaxuw:
+ case Intrinsic::x86_sse41_pmaxud:
+ case Intrinsic::x86_avx2_pmaxu_b:
+ case Intrinsic::x86_avx2_pmaxu_w:
+ case Intrinsic::x86_avx2_pmaxu_d:
+ case Intrinsic::x86_sse2_pminu_b:
+ case Intrinsic::x86_sse41_pminuw:
+ case Intrinsic::x86_sse41_pminud:
+ case Intrinsic::x86_avx2_pminu_b:
+ case Intrinsic::x86_avx2_pminu_w:
+ case Intrinsic::x86_avx2_pminu_d:
+ case Intrinsic::x86_sse41_pmaxsb:
+ case Intrinsic::x86_sse2_pmaxs_w:
+ case Intrinsic::x86_sse41_pmaxsd:
+ case Intrinsic::x86_avx2_pmaxs_b:
+ case Intrinsic::x86_avx2_pmaxs_w:
+ case Intrinsic::x86_avx2_pmaxs_d:
+ case Intrinsic::x86_sse41_pminsb:
+ case Intrinsic::x86_sse2_pmins_w:
+ case Intrinsic::x86_sse41_pminsd:
+ case Intrinsic::x86_avx2_pmins_b:
+ case Intrinsic::x86_avx2_pmins_w:
+ case Intrinsic::x86_avx2_pmins_d: {
+ unsigned Opcode;
+ switch (IntNo) {
+ default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
+ case Intrinsic::x86_sse2_pmaxu_b:
+ case Intrinsic::x86_sse41_pmaxuw:
+ case Intrinsic::x86_sse41_pmaxud:
+ case Intrinsic::x86_avx2_pmaxu_b:
+ case Intrinsic::x86_avx2_pmaxu_w:
+ case Intrinsic::x86_avx2_pmaxu_d:
+ Opcode = X86ISD::UMAX;
+ break;
+ case Intrinsic::x86_sse2_pminu_b:
+ case Intrinsic::x86_sse41_pminuw:
+ case Intrinsic::x86_sse41_pminud:
+ case Intrinsic::x86_avx2_pminu_b:
+ case Intrinsic::x86_avx2_pminu_w:
+ case Intrinsic::x86_avx2_pminu_d:
+ Opcode = X86ISD::UMIN;
+ break;
+ case Intrinsic::x86_sse41_pmaxsb:
+ case Intrinsic::x86_sse2_pmaxs_w:
+ case Intrinsic::x86_sse41_pmaxsd:
+ case Intrinsic::x86_avx2_pmaxs_b:
+ case Intrinsic::x86_avx2_pmaxs_w:
+ case Intrinsic::x86_avx2_pmaxs_d:
+ Opcode = X86ISD::SMAX;
+ break;
+ case Intrinsic::x86_sse41_pminsb:
+ case Intrinsic::x86_sse2_pmins_w:
+ case Intrinsic::x86_sse41_pminsd:
+ case Intrinsic::x86_avx2_pmins_b:
+ case Intrinsic::x86_avx2_pmins_w:
+ case Intrinsic::x86_avx2_pmins_d:
+ Opcode = X86ISD::SMIN;
+ break;
+ }
+ return DAG.getNode(Opcode, dl, Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2));
+ }
+
+ // SSE/SSE2/AVX floating point max/min intrinsics.
+ case Intrinsic::x86_sse_max_ps:
+ case Intrinsic::x86_sse2_max_pd:
+ case Intrinsic::x86_avx_max_ps_256:
+ case Intrinsic::x86_avx_max_pd_256:
+ case Intrinsic::x86_sse_min_ps:
+ case Intrinsic::x86_sse2_min_pd:
+ case Intrinsic::x86_avx_min_ps_256:
+ case Intrinsic::x86_avx_min_pd_256: {
+ unsigned Opcode;
+ switch (IntNo) {
+ default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
+ case Intrinsic::x86_sse_max_ps:
+ case Intrinsic::x86_sse2_max_pd:
+ case Intrinsic::x86_avx_max_ps_256:
+ case Intrinsic::x86_avx_max_pd_256:
+ Opcode = X86ISD::FMAX;
+ break;
+ case Intrinsic::x86_sse_min_ps:
+ case Intrinsic::x86_sse2_min_pd:
+ case Intrinsic::x86_avx_min_ps_256:
+ case Intrinsic::x86_avx_min_pd_256:
+ Opcode = X86ISD::FMIN;
+ break;
+ }
+ return DAG.getNode(Opcode, dl, Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2));
+ }
+
// AVX2 variable shift intrinsics
case Intrinsic::x86_avx2_psllv_d:
case Intrinsic::x86_avx2_psllv_q:
return DAG.getNode(X86ISD::VPERMV, dl, Op.getValueType(),
Op.getOperand(2), Op.getOperand(1));
+ case Intrinsic::x86_sse_sqrt_ps:
+ case Intrinsic::x86_sse2_sqrt_pd:
+ case Intrinsic::x86_avx_sqrt_ps_256:
+ case Intrinsic::x86_avx_sqrt_pd_256:
+ return DAG.getNode(ISD::FSQRT, dl, Op.getValueType(), Op.getOperand(1));
+
// ptest and testp intrinsics. The intrinsic these come from are designed to
// return an integer value, not just an instruction so lower it to the ptest
// or testp pattern and a setcc for the result.
// Check that ECX wasn't needed by an 'inreg' parameter.
FunctionType *FTy = Func->getFunctionType();
- const AttrListPtr &Attrs = Func->getAttributes();
+ const AttributeSet &Attrs = Func->getAttributes();
if (!Attrs.isEmpty() && !Func->isVarArg()) {
unsigned InRegCount = 0;
for (FunctionType::param_iterator I = FTy->param_begin(),
E = FTy->param_end(); I != E; ++I, ++Idx)
- if (Attrs.getParamAttributes(Idx).hasAttribute(Attributes::InReg))
+ if (Attrs.getParamAttributes(Idx).hasAttribute(Attribute::InReg))
// FIXME: should only count parameters that are lowered to integers.
InRegCount += (TD->getTypeSizeInBits(*I) + 31) / 32;
int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
-
MachineMemOperand *MMO =
MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
MachineMemOperand::MOStore, 2, 2);
DAG.getConstant(1, MVT::i16)),
DAG.getConstant(3, MVT::i16));
-
return DAG.getNode((VT.getSizeInBits() < 16 ?
ISD::TRUNCATE : ISD::ZERO_EXTEND), DL, VT, RetVal);
}
static SDValue LowerMUL(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
+ DebugLoc dl = Op.getDebugLoc();
EVT VT = Op.getValueType();
// Decompose 256-bit ops into smaller 128-bit ops.
if (VT.is256BitVector() && !Subtarget->hasInt256())
return Lower256IntArith(Op, DAG);
- assert((VT == MVT::v2i64 || VT == MVT::v4i64) &&
- "Only know how to lower V2I64/V4I64 multiply");
+ SDValue A = Op.getOperand(0);
+ SDValue B = Op.getOperand(1);
- DebugLoc dl = Op.getDebugLoc();
+ // Lower v4i32 mul as 2x shuffle, 2x pmuludq, 2x shuffle.
+ if (VT == MVT::v4i32) {
+ assert(Subtarget->hasSSE2() && !Subtarget->hasSSE41() &&
+ "Should not custom lower when pmuldq is available!");
- // Ahi = psrlqi(a, 32);
+ // Extract the odd parts.
+ const int UnpackMask[] = { 1, -1, 3, -1 };
+ SDValue Aodds = DAG.getVectorShuffle(VT, dl, A, A, UnpackMask);
+ SDValue Bodds = DAG.getVectorShuffle(VT, dl, B, B, UnpackMask);
+
+ // Multiply the even parts.
+ SDValue Evens = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, A, B);
+ // Now multiply odd parts.
+ SDValue Odds = DAG.getNode(X86ISD::PMULUDQ, dl, MVT::v2i64, Aodds, Bodds);
+
+ Evens = DAG.getNode(ISD::BITCAST, dl, VT, Evens);
+ Odds = DAG.getNode(ISD::BITCAST, dl, VT, Odds);
+
+ // Merge the two vectors back together with a shuffle. This expands into 2
+ // shuffles.
+ const int ShufMask[] = { 0, 4, 2, 6 };
+ return DAG.getVectorShuffle(VT, dl, Evens, Odds, ShufMask);
+ }
+
+ assert((VT == MVT::v2i64 || VT == MVT::v4i64) &&
+ "Only know how to lower V2I64/V4I64 multiply");
+
+ // Ahi = psrlqi(a, 32);
// Bhi = psrlqi(b, 32);
//
// AloBlo = pmuludq(a, b);
// AhiBlo = psllqi(AhiBlo, 32);
// return AloBlo + AloBhi + AhiBlo;
- SDValue A = Op.getOperand(0);
- SDValue B = Op.getOperand(1);
-
SDValue ShAmt = DAG.getConstant(32, MVT::i32);
SDValue Ahi = DAG.getNode(X86ISD::VSRLI, dl, VT, A, ShAmt);
}
}
-
static SDValue LowerMEMBARRIER(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
DebugLoc dl = Op.getDebugLoc();
return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
}
-
static SDValue LowerCMP_SWAP(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
EVT T = Op.getValueType();
case ISD::SINT_TO_FP: return LowerSINT_TO_FP(Op, DAG);
case ISD::UINT_TO_FP: return LowerUINT_TO_FP(Op, DAG);
case ISD::TRUNCATE: return lowerTRUNCATE(Op, DAG);
- case ISD::ZERO_EXTEND: return lowerZERO_EXTEND(Op, DAG);
+ case ISD::ZERO_EXTEND: return LowerZERO_EXTEND(Op, DAG);
+ case ISD::SIGN_EXTEND: return LowerSIGN_EXTEND(Op, DAG);
+ case ISD::ANY_EXTEND: return LowerANY_EXTEND(Op, DAG);
case ISD::FP_TO_SINT: return LowerFP_TO_SINT(Op, DAG);
case ISD::FP_TO_UINT: return LowerFP_TO_UINT(Op, DAG);
case ISD::FP_EXTEND: return lowerFP_EXTEND(Op, DAG);
SmallVectorImpl<SDValue>&Results,
SelectionDAG &DAG) const {
DebugLoc dl = N->getDebugLoc();
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
switch (N->getOpcode()) {
default:
llvm_unreachable("Do not know how to custom type legalize this operation!");
return;
}
case ISD::FP_ROUND: {
+ if (!TLI.isTypeLegal(N->getOperand(0).getValueType()))
+ return;
SDValue V = DAG.getNode(X86ISD::VFPROUND, dl, MVT::v4f32, N->getOperand(0));
Results.push_back(V);
return;
case X86ISD::ANDNP: return "X86ISD::ANDNP";
case X86ISD::PSIGN: return "X86ISD::PSIGN";
case X86ISD::BLENDV: return "X86ISD::BLENDV";
- case X86ISD::BLENDPW: return "X86ISD::BLENDPW";
- case X86ISD::BLENDPS: return "X86ISD::BLENDPS";
- case X86ISD::BLENDPD: return "X86ISD::BLENDPD";
+ case X86ISD::BLENDI: return "X86ISD::BLENDI";
+ case X86ISD::SUBUS: return "X86ISD::SUBUS";
case X86ISD::HADD: return "X86ISD::HADD";
case X86ISD::HSUB: return "X86ISD::HSUB";
case X86ISD::FHADD: return "X86ISD::FHADD";
case X86ISD::FHSUB: return "X86ISD::FHSUB";
+ case X86ISD::UMAX: return "X86ISD::UMAX";
+ case X86ISD::UMIN: return "X86ISD::UMIN";
+ case X86ISD::SMAX: return "X86ISD::SMAX";
+ case X86ISD::SMIN: return "X86ISD::SMIN";
case X86ISD::FMAX: return "X86ISD::FMAX";
case X86ISD::FMIN: return "X86ISD::FMIN";
case X86ISD::FMAXC: return "X86ISD::FMAXC";
case X86ISD::OR: return "X86ISD::OR";
case X86ISD::XOR: return "X86ISD::XOR";
case X86ISD::AND: return "X86ISD::AND";
- case X86ISD::ANDN: return "X86ISD::ANDN";
case X86ISD::BLSI: return "X86ISD::BLSI";
case X86ISD::BLSMSK: return "X86ISD::BLSMSK";
case X86ISD::BLSR: return "X86ISD::BLSR";
return true;
}
-
bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
return false;
unsigned NumBits1 = Ty1->getPrimitiveSizeInBits();
unsigned NumBits2 = Ty2->getPrimitiveSizeInBits();
- if (NumBits1 <= NumBits2)
- return false;
- return true;
+ return NumBits1 > NumBits2;
}
bool X86TargetLowering::isLegalICmpImmediate(int64_t Imm) const {
- return Imm == (int32_t)Imm;
+ return isInt<32>(Imm);
}
bool X86TargetLowering::isLegalAddImmediate(int64_t Imm) const {
// Can also use sub to handle negated immediates.
- return Imm == (int32_t)Imm;
+ return isInt<32>(Imm);
}
bool X86TargetLowering::isTruncateFree(EVT VT1, EVT VT2) const {
return false;
unsigned NumBits1 = VT1.getSizeInBits();
unsigned NumBits2 = VT2.getSizeInBits();
- if (NumBits1 <= NumBits2)
- return false;
- return true;
+ return NumBits1 > NumBits2;
}
bool X86TargetLowering::isZExtFree(Type *Ty1, Type *Ty2) const {
return VT1 == MVT::i32 && VT2 == MVT::i64 && Subtarget->is64Bit();
}
+bool X86TargetLowering::isZExtFree(SDValue Val, EVT VT2) const {
+ EVT VT1 = Val.getValueType();
+ if (isZExtFree(VT1, VT2))
+ return true;
+
+ if (Val.getOpcode() != ISD::LOAD)
+ return false;
+
+ if (!VT1.isSimple() || !VT1.isInteger() ||
+ !VT2.isSimple() || !VT2.isInteger())
+ return false;
+
+ switch (VT1.getSimpleVT().SimpleTy) {
+ default: break;
+ case MVT::i8:
+ case MVT::i16:
+ case MVT::i32:
+ // X86 has 8, 16, and 32-bit zero-extending loads.
+ return true;
+ }
+
+ return false;
+}
+
bool X86TargetLowering::isNarrowingProfitable(EVT VT1, EVT VT2) const {
// i16 instructions are longer (0x66 prefix) and potentially slower.
return !(VT1 == MVT::i32 && VT2 == MVT::i16);
return EltsFromConsecutiveLoads(VT, Elts, dl, DAG);
}
-
/// PerformTruncateCombine - Converts truncate operation to
/// a sequence of vector shuffle operations.
/// It is possible when we truncate 256-bit vector to 128-bit vector
static SDValue PerformTruncateCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
- if (!DCI.isBeforeLegalizeOps())
- return SDValue();
-
- if (!Subtarget->hasFp256())
- return SDValue();
-
- EVT VT = N->getValueType(0);
- SDValue Op = N->getOperand(0);
- EVT OpVT = Op.getValueType();
- DebugLoc dl = N->getDebugLoc();
-
- if ((VT == MVT::v4i32) && (OpVT == MVT::v4i64)) {
-
- if (Subtarget->hasInt256()) {
- // AVX2: v4i64 -> v4i32
-
- // VPERMD
- static const int ShufMask[] = {0, 2, 4, 6, -1, -1, -1, -1};
-
- Op = DAG.getNode(ISD::BITCAST, dl, MVT::v8i32, Op);
- Op = DAG.getVectorShuffle(MVT::v8i32, dl, Op, DAG.getUNDEF(MVT::v8i32),
- ShufMask);
-
- return DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, VT, Op,
- DAG.getIntPtrConstant(0));
- }
-
- // AVX: v4i64 -> v4i32
- SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i64, Op,
- DAG.getIntPtrConstant(0));
-
- SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i64, Op,
- DAG.getIntPtrConstant(2));
-
- OpLo = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpLo);
- OpHi = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpHi);
-
- // PSHUFD
- static const int ShufMask1[] = {0, 2, 0, 0};
-
- SDValue Undef = DAG.getUNDEF(VT);
- OpLo = DAG.getVectorShuffle(VT, dl, OpLo, Undef, ShufMask1);
- OpHi = DAG.getVectorShuffle(VT, dl, OpHi, Undef, ShufMask1);
-
- // MOVLHPS
- static const int ShufMask2[] = {0, 1, 4, 5};
-
- return DAG.getVectorShuffle(VT, dl, OpLo, OpHi, ShufMask2);
- }
-
- if ((VT == MVT::v8i16) && (OpVT == MVT::v8i32)) {
-
- if (Subtarget->hasInt256()) {
- // AVX2: v8i32 -> v8i16
-
- Op = DAG.getNode(ISD::BITCAST, dl, MVT::v32i8, Op);
-
- // PSHUFB
- SmallVector<SDValue,32> pshufbMask;
- for (unsigned i = 0; i < 2; ++i) {
- pshufbMask.push_back(DAG.getConstant(0x0, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0x1, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0x4, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0x5, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0x8, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0x9, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0xc, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(0xd, MVT::i8));
- for (unsigned j = 0; j < 8; ++j)
- pshufbMask.push_back(DAG.getConstant(0x80, MVT::i8));
- }
- SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v32i8,
- &pshufbMask[0], 32);
- Op = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v32i8, Op, BV);
-
- Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4i64, Op);
-
- static const int ShufMask[] = {0, 2, -1, -1};
- Op = DAG.getVectorShuffle(MVT::v4i64, dl, Op, DAG.getUNDEF(MVT::v4i64),
- &ShufMask[0]);
-
- Op = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v2i64, Op,
- DAG.getIntPtrConstant(0));
-
- return DAG.getNode(ISD::BITCAST, dl, VT, Op);
- }
-
- SDValue OpLo = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i32, Op,
- DAG.getIntPtrConstant(0));
-
- SDValue OpHi = DAG.getNode(ISD::EXTRACT_SUBVECTOR, dl, MVT::v4i32, Op,
- DAG.getIntPtrConstant(4));
-
- OpLo = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpLo);
- OpHi = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, OpHi);
-
- // PSHUFB
- static const int ShufMask1[] = {0, 1, 4, 5, 8, 9, 12, 13,
- -1, -1, -1, -1, -1, -1, -1, -1};
-
- SDValue Undef = DAG.getUNDEF(MVT::v16i8);
- OpLo = DAG.getVectorShuffle(MVT::v16i8, dl, OpLo, Undef, ShufMask1);
- OpHi = DAG.getVectorShuffle(MVT::v16i8, dl, OpHi, Undef, ShufMask1);
-
- OpLo = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpLo);
- OpHi = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, OpHi);
-
- // MOVLHPS
- static const int ShufMask2[] = {0, 1, 4, 5};
-
- SDValue res = DAG.getVectorShuffle(MVT::v4i32, dl, OpLo, OpHi, ShufMask2);
- return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, res);
- }
-
return SDValue();
}
return SDValue();
}
+/// \brief Matches a VSELECT onto min/max or return 0 if the node doesn't match.
+static unsigned matchIntegerMINMAX(SDValue Cond, EVT VT, SDValue LHS,
+ SDValue RHS, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
+ if (!VT.isVector())
+ return 0;
+
+ switch (VT.getSimpleVT().SimpleTy) {
+ default: return 0;
+ case MVT::v32i8:
+ case MVT::v16i16:
+ case MVT::v8i32:
+ if (!Subtarget->hasAVX2())
+ return 0;
+ case MVT::v16i8:
+ case MVT::v8i16:
+ case MVT::v4i32:
+ if (!Subtarget->hasSSE2())
+ return 0;
+ }
+
+ // SSE2 has only a small subset of the operations.
+ bool hasUnsigned = Subtarget->hasSSE41() ||
+ (Subtarget->hasSSE2() && VT == MVT::v16i8);
+ bool hasSigned = Subtarget->hasSSE41() ||
+ (Subtarget->hasSSE2() && VT == MVT::v8i16);
+
+ ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+
+ // Check for x CC y ? x : y.
+ if (DAG.isEqualTo(LHS, Cond.getOperand(0)) &&
+ DAG.isEqualTo(RHS, Cond.getOperand(1))) {
+ switch (CC) {
+ default: break;
+ case ISD::SETULT:
+ case ISD::SETULE:
+ return hasUnsigned ? X86ISD::UMIN : 0;
+ case ISD::SETUGT:
+ case ISD::SETUGE:
+ return hasUnsigned ? X86ISD::UMAX : 0;
+ case ISD::SETLT:
+ case ISD::SETLE:
+ return hasSigned ? X86ISD::SMIN : 0;
+ case ISD::SETGT:
+ case ISD::SETGE:
+ return hasSigned ? X86ISD::SMAX : 0;
+ }
+ // Check for x CC y ? y : x -- a min/max with reversed arms.
+ } else if (DAG.isEqualTo(LHS, Cond.getOperand(1)) &&
+ DAG.isEqualTo(RHS, Cond.getOperand(0))) {
+ switch (CC) {
+ default: break;
+ case ISD::SETULT:
+ case ISD::SETULE:
+ return hasUnsigned ? X86ISD::UMAX : 0;
+ case ISD::SETUGT:
+ case ISD::SETUGE:
+ return hasUnsigned ? X86ISD::UMIN : 0;
+ case ISD::SETLT:
+ case ISD::SETLE:
+ return hasSigned ? X86ISD::SMAX : 0;
+ case ISD::SETGT:
+ case ISD::SETGE:
+ return hasSigned ? X86ISD::SMIN : 0;
+ }
+ }
+
+ return 0;
+}
+
/// PerformSELECTCombine - Do target-specific dag combines on SELECT and VSELECT
/// nodes.
static SDValue PerformSELECTCombine(SDNode *N, SelectionDAG &DAG,
}
}
+ // Match VSELECTs into subs with unsigned saturation.
+ if (!DCI.isBeforeLegalize() &&
+ N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC &&
+ // psubus is available in SSE2 and AVX2 for i8 and i16 vectors.
+ ((Subtarget->hasSSE2() && (VT == MVT::v16i8 || VT == MVT::v8i16)) ||
+ (Subtarget->hasAVX2() && (VT == MVT::v32i8 || VT == MVT::v16i16)))) {
+ ISD::CondCode CC = cast<CondCodeSDNode>(Cond.getOperand(2))->get();
+
+ // Check if one of the arms of the VSELECT is a zero vector. If it's on the
+ // left side invert the predicate to simplify logic below.
+ SDValue Other;
+ if (ISD::isBuildVectorAllZeros(LHS.getNode())) {
+ Other = RHS;
+ CC = ISD::getSetCCInverse(CC, true);
+ } else if (ISD::isBuildVectorAllZeros(RHS.getNode())) {
+ Other = LHS;
+ }
+
+ if (Other.getNode() && Other->getNumOperands() == 2 &&
+ DAG.isEqualTo(Other->getOperand(0), Cond.getOperand(0))) {
+ SDValue OpLHS = Other->getOperand(0), OpRHS = Other->getOperand(1);
+ SDValue CondRHS = Cond->getOperand(1);
+
+ // Look for a general sub with unsigned saturation first.
+ // x >= y ? x-y : 0 --> subus x, y
+ // x > y ? x-y : 0 --> subus x, y
+ if ((CC == ISD::SETUGE || CC == ISD::SETUGT) &&
+ Other->getOpcode() == ISD::SUB && DAG.isEqualTo(OpRHS, CondRHS))
+ return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
+
+ // If the RHS is a constant we have to reverse the const canonicalization.
+ // x > C-1 ? x+-C : 0 --> subus x, C
+ if (CC == ISD::SETUGT && Other->getOpcode() == ISD::ADD &&
+ isSplatVector(CondRHS.getNode()) && isSplatVector(OpRHS.getNode())) {
+ APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
+ if (CondRHS.getConstantOperandVal(0) == -A-1) {
+ SmallVector<SDValue, 32> V(VT.getVectorNumElements(),
+ DAG.getConstant(-A, VT.getScalarType()));
+ return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS,
+ DAG.getNode(ISD::BUILD_VECTOR, DL, VT,
+ V.data(), V.size()));
+ }
+ }
+
+ // Another special case: If C was a sign bit, the sub has been
+ // canonicalized into a xor.
+ // FIXME: Would it be better to use ComputeMaskedBits to determine whether
+ // it's safe to decanonicalize the xor?
+ // x s< 0 ? x^C : 0 --> subus x, C
+ if (CC == ISD::SETLT && Other->getOpcode() == ISD::XOR &&
+ ISD::isBuildVectorAllZeros(CondRHS.getNode()) &&
+ isSplatVector(OpRHS.getNode())) {
+ APInt A = cast<ConstantSDNode>(OpRHS.getOperand(0))->getAPIntValue();
+ if (A.isSignBit())
+ return DAG.getNode(X86ISD::SUBUS, DL, VT, OpLHS, OpRHS);
+ }
+ }
+ }
+
+ // Try to match a min/max vector operation.
+ if (!DCI.isBeforeLegalize() &&
+ N->getOpcode() == ISD::VSELECT && Cond.getOpcode() == ISD::SETCC)
+ if (unsigned Op = matchIntegerMINMAX(Cond, VT, LHS, RHS, DAG, Subtarget))
+ return DAG.getNode(Op, DL, N->getValueType(0), LHS, RHS);
+
// If we know that this node is legal then we know that it is going to be
// matched by one of the SSE/AVX BLEND instructions. These instructions only
// depend on the highest bit in each word. Try to use SimplifyDemandedBits
return SDValue();
}
-
/// PerformMulCombine - Optimize a single multiply with constant into two
/// in order to implement it with two cheaper instructions, e.g.
/// LEA + SHL, LEA + LEA.
}
}
-
// Hardware support for vector shifts is sparse which makes us scalarize the
// vector operations in many cases. Also, on sandybridge ADD is faster than
// shl.
}
}
-
// CMPEQCombine - Recognize the distinctive (AND (setcc ...) (setcc ..))
// where both setccs reference the same FP CMP, and rewrite for CMPEQSS
// and friends. Likewise for OR -> CMPNEQSS.
return false;
}
+// On AVX/AVX2 the type v8i1 is legalized to v8i16, which is an XMM sized
+// register. In most cases we actually compare or select YMM-sized registers
+// and mixing the two types creates horrible code. This method optimizes
+// some of the transition sequences.
+static SDValue WidenMaskArithmetic(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
+ if (VT.getSizeInBits() != 256)
+ return SDValue();
+
+ assert((N->getOpcode() == ISD::ANY_EXTEND ||
+ N->getOpcode() == ISD::ZERO_EXTEND ||
+ N->getOpcode() == ISD::SIGN_EXTEND) && "Invalid Node");
+
+ SDValue Narrow = N->getOperand(0);
+ EVT NarrowVT = Narrow->getValueType(0);
+ if (NarrowVT.getSizeInBits() != 128)
+ return SDValue();
+
+ if (Narrow->getOpcode() != ISD::XOR &&
+ Narrow->getOpcode() != ISD::AND &&
+ Narrow->getOpcode() != ISD::OR)
+ return SDValue();
+
+ SDValue N0 = Narrow->getOperand(0);
+ SDValue N1 = Narrow->getOperand(1);
+ DebugLoc DL = Narrow->getDebugLoc();
+
+ // The Left side has to be a trunc.
+ if (N0.getOpcode() != ISD::TRUNCATE)
+ return SDValue();
+
+ // The type of the truncated inputs.
+ EVT WideVT = N0->getOperand(0)->getValueType(0);
+ if (WideVT != VT)
+ return SDValue();
+
+ // The right side has to be a 'trunc' or a constant vector.
+ bool RHSTrunc = N1.getOpcode() == ISD::TRUNCATE;
+ bool RHSConst = (isSplatVector(N1.getNode()) &&
+ isa<ConstantSDNode>(N1->getOperand(0)));
+ if (!RHSTrunc && !RHSConst)
+ return SDValue();
+
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+
+ if (!TLI.isOperationLegalOrPromote(Narrow->getOpcode(), WideVT))
+ return SDValue();
+
+ // Set N0 and N1 to hold the inputs to the new wide operation.
+ N0 = N0->getOperand(0);
+ if (RHSConst) {
+ N1 = DAG.getNode(ISD::ZERO_EXTEND, DL, WideVT.getScalarType(),
+ N1->getOperand(0));
+ SmallVector<SDValue, 8> C(WideVT.getVectorNumElements(), N1);
+ N1 = DAG.getNode(ISD::BUILD_VECTOR, DL, WideVT, &C[0], C.size());
+ } else if (RHSTrunc) {
+ N1 = N1->getOperand(0);
+ }
+
+ // Generate the wide operation.
+ SDValue Op = DAG.getNode(N->getOpcode(), DL, WideVT, N0, N1);
+ unsigned Opcode = N->getOpcode();
+ switch (Opcode) {
+ case ISD::ANY_EXTEND:
+ return Op;
+ case ISD::ZERO_EXTEND: {
+ unsigned InBits = NarrowVT.getScalarType().getSizeInBits();
+ APInt Mask = APInt::getAllOnesValue(InBits);
+ Mask = Mask.zext(VT.getScalarType().getSizeInBits());
+ return DAG.getNode(ISD::AND, DL, VT,
+ Op, DAG.getConstant(Mask, VT));
+ }
+ case ISD::SIGN_EXTEND:
+ return DAG.getNode(ISD::SIGN_EXTEND_INREG, DL, VT,
+ Op, DAG.getValueType(NarrowVT));
+ default:
+ llvm_unreachable("Unexpected opcode");
+ }
+}
+
static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
if (R.getNode())
return R;
- EVT VT = N->getValueType(0);
-
- // Create ANDN, BLSI, and BLSR instructions
+ // Create BLSI, and BLSR instructions
// BLSI is X & (-X)
// BLSR is X & (X-1)
if (Subtarget->hasBMI() && (VT == MVT::i32 || VT == MVT::i64)) {
SDValue N1 = N->getOperand(1);
DebugLoc DL = N->getDebugLoc();
- // Check LHS for not
- if (N0.getOpcode() == ISD::XOR && isAllOnes(N0.getOperand(1)))
- return DAG.getNode(X86ISD::ANDN, DL, VT, N0.getOperand(0), N1);
- // Check RHS for not
- if (N1.getOpcode() == ISD::XOR && isAllOnes(N1.getOperand(1)))
- return DAG.getNode(X86ISD::ANDN, DL, VT, N1.getOperand(0), N0);
-
// Check LHS for neg
if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
isZero(N0.getOperand(0)))
static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
if (R.getNode())
return R;
- EVT VT = N->getValueType(0);
-
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
DebugLoc DL = N->getDebugLoc();
+ // We are going to replace the AND, OR, NAND with either BLEND
+ // or PSIGN, which only look at the MSB. The VSRAI instruction
+ // does not affect the highest bit, so we can get rid of it.
+ Mask = Mask.getOperand(0);
+
// Now we know we at least have a plendvb with the mask val. See if
// we can form a psignb/w/d.
// psign = x.type == y.type == mask.type && y = sub(0, x);
X.getValueType() == MaskVT && Y.getValueType() == MaskVT) {
assert((EltBits == 8 || EltBits == 16 || EltBits == 32) &&
"Unsupported VT for PSIGN");
- Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask.getOperand(0));
+ Mask = DAG.getNode(X86ISD::PSIGN, DL, MaskVT, X, Mask);
return DAG.getNode(ISD::BITCAST, DL, VT, Mask);
}
// PBLENDVB only available on SSE 4.1
static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
+ EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
if (!Subtarget->hasBMI())
return SDValue();
- EVT VT = N->getValueType(0);
-
if (VT != MVT::i32 && VT != MVT::i64)
return SDValue();
ISD::LoadExtType Ext = Ld->getExtensionType();
// If this is a vector EXT Load then attempt to optimize it using a
- // shuffle. We need SSSE3 shuffles.
+ // shuffle. If SSSE3 is not available we may emit an illegal shuffle but the
+ // expansion is still better than scalar code.
+ // We generate X86ISD::VSEXT for SEXTLOADs if it's available, otherwise we'll
+ // emit a shuffle and a arithmetic shift.
// TODO: It is possible to support ZExt by zeroing the undef values
// during the shuffle phase or after the shuffle.
- if (RegVT.isVector() && RegVT.isInteger() &&
- Ext == ISD::EXTLOAD && Subtarget->hasSSSE3()) {
+ if (RegVT.isVector() && RegVT.isInteger() && Subtarget->hasSSE2() &&
+ (Ext == ISD::EXTLOAD || Ext == ISD::SEXTLOAD)) {
assert(MemVT != RegVT && "Cannot extend to the same type");
assert(MemVT.isVector() && "Must load a vector from memory");
unsigned MemSz = MemVT.getSizeInBits();
assert(RegSz > MemSz && "Register size must be greater than the mem size");
+ if (Ext == ISD::SEXTLOAD && RegSz == 256 && !Subtarget->hasInt256())
+ return SDValue();
+
// All sizes must be a power of two.
if (!isPowerOf2_32(RegSz * MemSz * NumElems))
return SDValue();
// Calculate the number of scalar loads that we need to perform
// in order to load our vector from memory.
unsigned NumLoads = MemSz / SclrLoadTy.getSizeInBits();
+ if (Ext == ISD::SEXTLOAD && NumLoads > 1)
+ return SDValue();
+
+ unsigned loadRegZize = RegSz;
+ if (Ext == ISD::SEXTLOAD && RegSz == 256)
+ loadRegZize /= 2;
// Represent our vector as a sequence of elements which are the
// largest scalar that we can load.
EVT LoadUnitVecVT = EVT::getVectorVT(*DAG.getContext(), SclrLoadTy,
- RegSz/SclrLoadTy.getSizeInBits());
+ loadRegZize/SclrLoadTy.getSizeInBits());
// Represent the data using the same element type that is stored in
// memory. In practice, we ''widen'' MemVT.
- EVT WideVecVT = EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
- RegSz/MemVT.getScalarType().getSizeInBits());
+ EVT WideVecVT =
+ EVT::getVectorVT(*DAG.getContext(), MemVT.getScalarType(),
+ loadRegZize/MemVT.getScalarType().getSizeInBits());
assert(WideVecVT.getSizeInBits() == LoadUnitVecVT.getSizeInBits() &&
"Invalid vector type");
SDValue SlicedVec = DAG.getNode(ISD::BITCAST, dl, WideVecVT, Res);
unsigned SizeRatio = RegSz/MemSz;
+ if (Ext == ISD::SEXTLOAD) {
+ // If we have SSE4.1 we can directly emit a VSEXT node.
+ if (Subtarget->hasSSE41()) {
+ SDValue Sext = DAG.getNode(X86ISD::VSEXT, dl, RegVT, SlicedVec);
+ return DCI.CombineTo(N, Sext, TF, true);
+ }
+
+ // Otherwise we'll shuffle the small elements in the high bits of the
+ // larger type and perform an arithmetic shift. If the shift is not legal
+ // it's better to scalarize.
+ if (!TLI.isOperationLegalOrCustom(ISD::SRA, RegVT))
+ return SDValue();
+
+ // Redistribute the loaded elements into the different locations.
+ SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
+ for (unsigned i = 0; i != NumElems; ++i)
+ ShuffleVec[i*SizeRatio + SizeRatio-1] = i;
+
+ SDValue Shuff = DAG.getVectorShuffle(WideVecVT, dl, SlicedVec,
+ DAG.getUNDEF(WideVecVT),
+ &ShuffleVec[0]);
+
+ Shuff = DAG.getNode(ISD::BITCAST, dl, RegVT, Shuff);
+
+ // Build the arithmetic shift.
+ unsigned Amt = RegVT.getVectorElementType().getSizeInBits() -
+ MemVT.getVectorElementType().getSizeInBits();
+ SmallVector<SDValue, 8> C(NumElems,
+ DAG.getConstant(Amt, RegVT.getScalarType()));
+ SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, RegVT, &C[0], C.size());
+ Shuff = DAG.getNode(ISD::SRA, dl, RegVT, Shuff, BV);
+
+ return DCI.CombineTo(N, Shuff, TF, true);
+ }
+
// Redistribute the loaded elements into the different locations.
SmallVector<int, 8> ShuffleVec(NumElems * SizeRatio, -1);
for (unsigned i = 0; i != NumElems; ++i)
Chains.size());
}
-
// Turn load->store of MMX types into GPR load/stores. This avoids clobbering
// the FP state in cases where an emms may be missing.
// A preferable solution to the general problem is to figure out the right
const Function *F = DAG.getMachineFunction().getFunction();
bool NoImplicitFloatOps = F->getFnAttributes().
- hasAttribute(Attributes::NoImplicitFloat);
+ hasAttribute(Attribute::NoImplicitFloat);
bool F64IsLegal = !DAG.getTarget().Options.UseSoftFloat && !NoImplicitFloatOps
&& Subtarget->hasSSE2();
if ((VT.isVector() ||
N->getOperand(0), N->getOperand(1));
}
-
/// PerformFANDCombine - Do target-specific dag combines on X86ISD::FAND nodes.
static SDValue PerformFANDCombine(SDNode *N, SelectionDAG &DAG) {
// FAND(0.0, x) -> 0.0
return SDValue();
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)) {
-
- if (Subtarget->hasInt256())
- return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, Op);
-
- // 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
-
- unsigned NumElems = OpVT.getVectorNumElements();
- SDValue Undef = DAG.getUNDEF(OpVT);
-
- SmallVector<int,8> ShufMask1(NumElems, -1);
- for (unsigned i = 0; i != NumElems/2; ++i)
- ShufMask1[i] = i;
-
- SDValue OpLo = DAG.getVectorShuffle(OpVT, dl, Op, Undef, &ShufMask1[0]);
-
- 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, Undef, &ShufMask2[0]);
-
- 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);
+ if (VT.isVector() && VT.getSizeInBits() == 256) {
+ SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
+ if (R.getNode())
+ return R;
}
+
return SDValue();
}
DebugLoc dl = N->getDebugLoc();
SDValue N0 = N->getOperand(0);
EVT VT = N->getValueType(0);
- EVT OpVT = N0.getValueType();
if (N0.getOpcode() == ISD::AND &&
N0.hasOneUse() &&
N0.getOperand(0).hasOneUse()) {
SDValue N00 = N0.getOperand(0);
- if (N00.getOpcode() != X86ISD::SETCC_CARRY)
- return SDValue();
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
- if (!C || C->getZExtValue() != 1)
- return SDValue();
- return DAG.getNode(ISD::AND, dl, VT,
- DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
- N00.getOperand(0), N00.getOperand(1)),
- DAG.getConstant(1, VT));
+ if (N00.getOpcode() == X86ISD::SETCC_CARRY) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(N0.getOperand(1));
+ if (!C || C->getZExtValue() != 1)
+ return SDValue();
+ return DAG.getNode(ISD::AND, dl, VT,
+ DAG.getNode(X86ISD::SETCC_CARRY, dl, VT,
+ N00.getOperand(0), N00.getOperand(1)),
+ DAG.getConstant(1, VT));
+ }
}
- // Optimize vectors in AVX mode:
- //
- // v8i16 -> v8i32
- // Use vpunpcklwd for 4 lower elements v8i16 -> v4i32.
- // Use vpunpckhwd for 4 upper elements v8i16 -> v4i32.
- // Concat upper and lower parts.
- //
- // v4i32 -> v4i64
- // Use vpunpckldq for 4 lower elements v4i32 -> v2i64.
- // Use vpunpckhdq for 4 upper elements v4i32 -> v2i64.
- // Concat upper and lower parts.
- //
- if (!DCI.isBeforeLegalizeOps())
- return SDValue();
-
- if (!Subtarget->hasFp256())
- return SDValue();
-
- if (((VT == MVT::v8i32) && (OpVT == MVT::v8i16)) ||
- ((VT == MVT::v4i64) && (OpVT == MVT::v4i32))) {
-
- if (Subtarget->hasInt256())
- return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, N0);
-
- SDValue ZeroVec = getZeroVector(OpVT, Subtarget, DAG, dl);
- SDValue OpLo = getUnpackl(DAG, dl, OpVT, N0, ZeroVec);
- SDValue OpHi = getUnpackh(DAG, dl, OpVT, N0, ZeroVec);
-
- EVT HVT = EVT::getVectorVT(*DAG.getContext(), VT.getVectorElementType(),
- VT.getVectorNumElements()/2);
-
- OpLo = DAG.getNode(ISD::BITCAST, dl, HVT, OpLo);
- OpHi = DAG.getNode(ISD::BITCAST, dl, HVT, OpHi);
-
- return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
+ if (VT.isVector() && VT.getSizeInBits() == 256) {
+ SDValue R = WidenMaskArithmetic(N, DAG, DCI, Subtarget);
+ if (R.getNode())
+ return R;
}
return SDValue();
return false;
}
-
-
/// getConstraintType - Given a constraint letter, return the type of
/// constraint it is for this target.
X86TargetLowering::ConstraintType
case 'f':
case 't':
case 'u':
- if (type->isFloatingPointTy())
- weight = CW_SpecificReg;
- break;
+ if (type->isFloatingPointTy())
+ weight = CW_SpecificReg;
+ break;
case 'y':
- if (type->isX86_MMXTy() && Subtarget->hasMMX())
- weight = CW_SpecificReg;
- break;
+ if (type->isX86_MMXTy() && Subtarget->hasMMX())
+ weight = CW_SpecificReg;
+ break;
case 'x':
case 'Y':
if (((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasSSE1()) ||
return -1;
}
+ScalarTargetTransformInfo::PopcntHwSupport
+X86ScalarTargetTransformImpl::getPopcntHwSupport(unsigned TyWidth) const {
+ assert(isPowerOf2_32(TyWidth) && "Ty width must be power of 2");
+ const X86Subtarget &ST = TLI->getTargetMachine().getSubtarget<X86Subtarget>();
+
+ // TODO: Currently the __builtin_popcount() implementation using SSE3
+ // instructions is inefficient. Once the problem is fixed, we should
+ // call ST.hasSSE3() instead of ST.hasSSE4().
+ return ST.hasSSE41() ? Fast : None;
+}
+
unsigned
X86VectorTargetTransformInfo::getArithmeticInstrCost(unsigned Opcode,
Type *Ty) const {
return VectorTargetTransformImpl::getArithmeticInstrCost(Opcode, Ty);
}
+unsigned
+X86VectorTargetTransformInfo::getMemoryOpCost(unsigned Opcode, Type *Src,
+ unsigned Alignment,
+ unsigned AddressSpace) const {
+ // Legalize the type.
+ std::pair<unsigned, MVT> LT = getTypeLegalizationCost(Src);
+ assert((Opcode == Instruction::Load || Opcode == Instruction::Store) &&
+ "Invalid Opcode");
+
+ const X86Subtarget &ST =
+ TLI->getTargetMachine().getSubtarget<X86Subtarget>();
+
+ // Each load/store unit costs 1.
+ unsigned Cost = LT.first * 1;
+
+ // On Sandybridge 256bit load/stores are double pumped
+ // (but not on Haswell).
+ if (LT.second.getSizeInBits() > 128 && !ST.hasAVX2())
+ Cost*=2;
+
+ return Cost;
+}
+
unsigned
X86VectorTargetTransformInfo::getVectorInstrCost(unsigned Opcode, Type *Val,
unsigned Index) const {
{ ISD::SETCC, MVT::v32i8, 1 },
};
- if (ST.hasSSE42()) {
- int Idx = FindInTable(SSE42CostTbl, array_lengthof(SSE42CostTbl), ISD, MTy);
+ if (ST.hasAVX2()) {
+ int Idx = FindInTable(AVX2CostTbl, array_lengthof(AVX2CostTbl), ISD, MTy);
if (Idx != -1)
- return LT.first * SSE42CostTbl[Idx].Cost;
+ return LT.first * AVX2CostTbl[Idx].Cost;
}
if (ST.hasAVX()) {
return LT.first * AVX1CostTbl[Idx].Cost;
}
- if (ST.hasAVX2()) {
- int Idx = FindInTable(AVX2CostTbl, array_lengthof(AVX2CostTbl), ISD, MTy);
+ if (ST.hasSSE42()) {
+ int Idx = FindInTable(SSE42CostTbl, array_lengthof(SSE42CostTbl), ISD, MTy);
if (Idx != -1)
- return LT.first * AVX2CostTbl[Idx].Cost;
+ return LT.first * SSE42CostTbl[Idx].Cost;
}
return VectorTargetTransformImpl::getCmpSelInstrCost(Opcode, ValTy, CondTy);
return VectorTargetTransformImpl::getCastInstrCost(Opcode, Dst, Src);
}
+
+unsigned X86VectorTargetTransformInfo::getShuffleCost(ShuffleKind Kind, Type *Tp,
+ int Index) const {
+ // We only estimate the cost of reverse shuffles.
+ if (Kind != Reverse)
+ return VectorTargetTransformImpl::getShuffleCost(Kind, Tp, Index);
+
+ std::pair<unsigned, MVT> LT = getTypeLegalizationCost(Tp);
+ unsigned Cost = 1;
+ if (LT.second.getSizeInBits() > 128)
+ Cost = 3; // Extract + insert + copy.
+
+ // Multiple by the number of parts.
+ return Cost * LT.first;
+}
+