//
//===----------------------------------------------------------------------===//
-#define DEBUG_TYPE "x86-isel"
#include "X86ISelLowering.h"
#include "Utils/X86ShuffleDecode.h"
-#include "X86.h"
#include "X86CallingConv.h"
#include "X86InstrBuilder.h"
+#include "X86MachineFunctionInfo.h"
#include "X86TargetMachine.h"
#include "X86TargetObjectFile.h"
#include "llvm/ADT/SmallSet.h"
#include "llvm/CodeGen/MachineJumpTableInfo.h"
#include "llvm/CodeGen/MachineModuleInfo.h"
#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/CallSite.h"
#include "llvm/IR/CallingConv.h"
#include "llvm/IR/Constants.h"
#include "llvm/IR/DerivedTypes.h"
#include "llvm/MC/MCContext.h"
#include "llvm/MC/MCExpr.h"
#include "llvm/MC/MCSymbol.h"
-#include "llvm/Support/CallSite.h"
#include "llvm/Support/Debug.h"
#include "llvm/Support/ErrorHandling.h"
#include "llvm/Support/MathExtras.h"
#include <cctype>
using namespace llvm;
+#define DEBUG_TYPE "x86-isel"
+
STATISTIC(NumTailCalls, "Number of tail calls");
// Forward declarations.
return new X86LinuxTargetObjectFile();
if (Subtarget->isTargetELF())
return new TargetLoweringObjectFileELF();
+ if (Subtarget->isTargetKnownWindowsMSVC())
+ return new X86WindowsTargetObjectFile();
if (Subtarget->isTargetCOFF())
return new TargetLoweringObjectFileCOFF();
llvm_unreachable("unknown subtarget type");
addBypassSlowDiv(64, 16);
}
- if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
+ if (Subtarget->isTargetKnownWindowsMSVC()) {
// Setup Windows compiler runtime calls.
setLibcallName(RTLIB::SDIV_I64, "_alldiv");
setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
// The _ftol2 runtime function has an unusual calling conv, which
// is modeled by a special pseudo-instruction.
- setLibcallName(RTLIB::FPTOUINT_F64_I64, 0);
- setLibcallName(RTLIB::FPTOUINT_F32_I64, 0);
- setLibcallName(RTLIB::FPTOUINT_F64_I32, 0);
- setLibcallName(RTLIB::FPTOUINT_F32_I32, 0);
+ setLibcallName(RTLIB::FPTOUINT_F64_I64, nullptr);
+ setLibcallName(RTLIB::FPTOUINT_F32_I64, nullptr);
+ setLibcallName(RTLIB::FPTOUINT_F64_I32, nullptr);
+ setLibcallName(RTLIB::FPTOUINT_F32_I32, nullptr);
}
if (Subtarget->isTargetDarwin()) {
// Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
setUseUnderscoreSetJmp(false);
setUseUnderscoreLongJmp(false);
- } else if (Subtarget->isTargetMingw()) {
+ } else if (Subtarget->isTargetWindowsGNU()) {
// MS runtime is weird: it exports _setjmp, but longjmp!
setUseUnderscoreSetJmp(true);
setUseUnderscoreLongJmp(false);
}
setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
- setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
+
+ if (!Subtarget->hasMOVBE())
+ setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
// These should be promoted to a larger select which is supported.
setOperationAction(ISD::SELECT , MVT::i1 , Promote);
setOperationAction(ISD::STACKSAVE, MVT::Other, Expand);
setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
- if (Subtarget->isOSWindows() && !Subtarget->isTargetMacho())
- setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
- MVT::i64 : MVT::i32, Custom);
- else if (TM.Options.EnableSegmentedStacks)
- setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
- MVT::i64 : MVT::i32, Custom);
- else
- setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
- MVT::i64 : MVT::i32, Expand);
+ setOperationAction(ISD::DYNAMIC_STACKALLOC, Subtarget->is64Bit() ?
+ MVT::i64 : MVT::i32, Custom);
if (!TM.Options.UseSoftFloat && X86ScalarSSEf64) {
// f32 and f64 use SSE.
setOperationAction(ISD::FRINT, VT, Expand);
setOperationAction(ISD::FNEARBYINT, VT, Expand);
setOperationAction(ISD::SMUL_LOHI, VT, Expand);
+ setOperationAction(ISD::MULHS, VT, Expand);
setOperationAction(ISD::UMUL_LOHI, VT, Expand);
+ setOperationAction(ISD::MULHU, VT, Expand);
setOperationAction(ISD::SDIVREM, VT, Expand);
setOperationAction(ISD::UDIVREM, VT, Expand);
setOperationAction(ISD::FPOW, VT, Expand);
setOperationAction(ISD::ADD, MVT::v2i64, Legal);
setOperationAction(ISD::MUL, MVT::v4i32, Custom);
setOperationAction(ISD::MUL, MVT::v2i64, Custom);
+ setOperationAction(ISD::UMUL_LOHI, MVT::v4i32, Custom);
+ setOperationAction(ISD::MULHU, MVT::v8i16, Legal);
+ setOperationAction(ISD::MULHS, MVT::v8i16, Legal);
setOperationAction(ISD::SUB, MVT::v16i8, Legal);
setOperationAction(ISD::SUB, MVT::v8i16, Legal);
setOperationAction(ISD::SUB, MVT::v4i32, Legal);
setOperationAction(ISD::SHL, MVT::v4i32, Custom);
setOperationAction(ISD::SRA, MVT::v4i32, Custom);
-
- setOperationAction(ISD::SDIV, MVT::v8i16, Custom);
- setOperationAction(ISD::SDIV, MVT::v4i32, Custom);
}
if (!TM.Options.UseSoftFloat && Subtarget->hasFp256()) {
setOperationAction(ISD::FNEG, MVT::v4f64, Custom);
setOperationAction(ISD::FABS, MVT::v4f64, Custom);
- setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Custom);
-
+ // (fp_to_int:v8i16 (v8f32 ..)) requires the result type to be promoted
+ // even though v8i16 is a legal type.
+ setOperationAction(ISD::FP_TO_SINT, MVT::v8i16, Promote);
+ setOperationAction(ISD::FP_TO_UINT, MVT::v8i16, Promote);
setOperationAction(ISD::FP_TO_SINT, MVT::v8i32, Legal);
+
setOperationAction(ISD::SINT_TO_FP, MVT::v8i16, Promote);
setOperationAction(ISD::SINT_TO_FP, MVT::v8i32, Legal);
setOperationAction(ISD::FP_ROUND, MVT::v4f32, Legal);
setOperationAction(ISD::SRA, MVT::v16i16, Custom);
setOperationAction(ISD::SRA, MVT::v32i8, Custom);
- setOperationAction(ISD::SDIV, MVT::v16i16, Custom);
-
setOperationAction(ISD::SETCC, MVT::v32i8, Custom);
setOperationAction(ISD::SETCC, MVT::v16i16, Custom);
setOperationAction(ISD::SETCC, MVT::v8i32, Custom);
setOperationAction(ISD::MUL, MVT::v16i16, Legal);
// Don't lower v32i8 because there is no 128-bit byte mul
- setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
+ setOperationAction(ISD::UMUL_LOHI, MVT::v8i32, Custom);
+ setOperationAction(ISD::MULHU, MVT::v16i16, Legal);
+ setOperationAction(ISD::MULHS, MVT::v16i16, Legal);
- setOperationAction(ISD::SDIV, MVT::v8i32, Custom);
+ setOperationAction(ISD::VSELECT, MVT::v32i8, Legal);
} else {
setOperationAction(ISD::ADD, MVT::v4i64, Custom);
setOperationAction(ISD::ADD, MVT::v8i32, Custom);
setOperationAction(ISD::FNEG, MVT::v8f64, Custom);
setOperationAction(ISD::FMA, MVT::v8f64, Legal);
setOperationAction(ISD::FMA, MVT::v16f32, Legal);
- setOperationAction(ISD::SDIV, MVT::v16i32, Custom);
setOperationAction(ISD::FP_TO_SINT, MVT::i32, Legal);
setOperationAction(ISD::FP_TO_UINT, MVT::i32, Legal);
setOperationAction(ISD::FP_TO_SINT, MVT::v16i32, Legal);
setOperationAction(ISD::FP_TO_UINT, MVT::v16i32, Legal);
setOperationAction(ISD::FP_TO_UINT, MVT::v8i32, Legal);
+ setOperationAction(ISD::FP_TO_UINT, MVT::v4i32, Legal);
setOperationAction(ISD::SINT_TO_FP, MVT::v16i32, Legal);
setOperationAction(ISD::UINT_TO_FP, MVT::v16i32, Legal);
setOperationAction(ISD::UINT_TO_FP, MVT::v8i32, Legal);
+ setOperationAction(ISD::UINT_TO_FP, MVT::v4i32, Legal);
setOperationAction(ISD::FP_ROUND, MVT::v8f32, Legal);
setOperationAction(ISD::FP_EXTEND, MVT::v8f32, Legal);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v8i1, Custom);
setOperationAction(ISD::EXTRACT_VECTOR_ELT, MVT::v16i1, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v16i1, Custom);
+ setOperationAction(ISD::INSERT_VECTOR_ELT, MVT::v8i1, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v8i1, Custom);
setOperationAction(ISD::BUILD_VECTOR, MVT::v16i1, Custom);
setOperationAction(ISD::SELECT, MVT::v8f64, Custom);
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::Other, Custom);
setOperationAction(ISD::INTRINSIC_VOID, MVT::Other, Custom);
+ if (!Subtarget->is64Bit())
+ setOperationAction(ISD::INTRINSIC_W_CHAIN, MVT::i64, Custom);
// Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
// handle type legalization for these operations here.
if (!Subtarget->is64Bit()) {
// These libcalls are not available in 32-bit.
- setLibcallName(RTLIB::SHL_I128, 0);
- setLibcallName(RTLIB::SRL_I128, 0);
- setLibcallName(RTLIB::SRA_I128, 0);
+ setLibcallName(RTLIB::SHL_I128, nullptr);
+ setLibcallName(RTLIB::SRL_I128, nullptr);
+ setLibcallName(RTLIB::SRA_I128, nullptr);
}
// Combine sin / cos into one node or libcall if possible.
}
bool
-X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT, bool *Fast) const {
+X86TargetLowering::allowsUnalignedMemoryAccesses(EVT VT,
+ unsigned,
+ bool *Fast) const {
if (Fast)
*Fast = Subtarget->isUnalignedMemAccessFast();
return true;
// FIXME: Why this routine is here? Move to RegInfo!
std::pair<const TargetRegisterClass*, uint8_t>
X86TargetLowering::findRepresentativeClass(MVT VT) const{
- const TargetRegisterClass *RRC = 0;
+ const TargetRegisterClass *RRC = nullptr;
uint8_t Cost = 1;
switch (VT.SimpleTy) {
default:
return CCInfo.CheckReturn(Outs, RetCC_X86);
}
-const uint16_t *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
- static const uint16_t ScratchRegs[] = { X86::R11, 0 };
+const MCPhysReg *X86TargetLowering::getScratchRegisters(CallingConv::ID) const {
+ static const MCPhysReg ScratchRegs[] = { X86::R11, 0 };
return ScratchRegs;
}
else if (VA.getLocInfo() == CCValAssign::BCvt)
ValToCopy = DAG.getNode(ISD::BITCAST, dl, VA.getLocVT(), ValToCopy);
+ assert(VA.getLocInfo() != CCValAssign::FPExt &&
+ "Unexpected FP-extend for return value.");
+
// If this is x86-64, and we disabled SSE, we can't return FP values,
// or SSE or MMX vectors.
if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
// We saved the argument into a virtual register in the entry block,
// so now we copy the value out and into %rax/%eax.
if (DAG.getMachineFunction().getFunction()->hasStructRetAttr() &&
- (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
+ (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
MachineFunction &MF = DAG.getMachineFunction();
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
unsigned Reg = FuncInfo->getSRetReturnReg();
// Save the argument into a virtual register so that we can access it
// from the return points.
if (MF.getFunction()->hasStructRetAttr() &&
- (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
+ (Subtarget->is64Bit() || Subtarget->isTargetKnownWindowsMSVC())) {
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
unsigned Reg = FuncInfo->getSRetReturnReg();
if (!Reg) {
unsigned TotalNumIntRegs = 0, TotalNumXMMRegs = 0;
// FIXME: We should really autogenerate these arrays
- static const uint16_t GPR64ArgRegsWin64[] = {
+ static const MCPhysReg GPR64ArgRegsWin64[] = {
X86::RCX, X86::RDX, X86::R8, X86::R9
};
- static const uint16_t GPR64ArgRegs64Bit[] = {
+ static const MCPhysReg GPR64ArgRegs64Bit[] = {
X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8, X86::R9
};
- static const uint16_t XMMArgRegs64Bit[] = {
+ static const MCPhysReg XMMArgRegs64Bit[] = {
X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
};
- const uint16_t *GPR64ArgRegs;
+ const MCPhysReg *GPR64ArgRegs;
unsigned NumXMMRegs = 0;
if (IsWin64) {
MF.getFunction()->hasStructRetAttr(), CLI.RetTy,
Outs, OutVals, Ins, DAG);
+ if (!isTailCall && CLI.CS && CLI.CS->isMustTailCall())
+ report_fatal_error("failed to perform tail call elimination on a call "
+ "site marked musttail");
+
// Sibcalls are automatically detected tailcalls which do not require
// ABI changes.
if (!MF.getTarget().Options.GuaranteedTailCallOpt && isTailCall)
X86Info->setTCReturnAddrDelta(FPDiff);
}
+ unsigned NumBytesToPush = NumBytes;
+ unsigned NumBytesToPop = NumBytes;
+
+ // If we have an inalloca argument, all stack space has already been allocated
+ // for us and be right at the top of the stack. We don't support multiple
+ // arguments passed in memory when using inalloca.
+ if (!Outs.empty() && Outs.back().Flags.isInAlloca()) {
+ NumBytesToPush = 0;
+ assert(ArgLocs.back().getLocMemOffset() == 0 &&
+ "an inalloca argument must be the only memory argument");
+ }
+
if (!IsSibcall)
- Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(NumBytes, true),
- dl);
+ Chain = DAG.getCALLSEQ_START(
+ Chain, DAG.getIntPtrConstant(NumBytesToPush, true), dl);
SDValue RetAddrFrIdx;
// Load return address for tail calls.
const X86RegisterInfo *RegInfo =
static_cast<const X86RegisterInfo*>(getTargetMachine().getRegisterInfo());
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
+ // Skip inalloca arguments, they have already been written.
+ ISD::ArgFlagsTy Flags = Outs[i].Flags;
+ if (Flags.isInAlloca())
+ continue;
+
CCValAssign &VA = ArgLocs[i];
EVT RegVT = VA.getLocVT();
SDValue Arg = OutVals[i];
- ISD::ArgFlagsTy Flags = Outs[i].Flags;
bool isByVal = Flags.isByVal();
// Promote the value if needed.
}
} else if (!IsSibcall && (!isTailCall || isByVal)) {
assert(VA.isMemLoc());
- if (StackPtr.getNode() == 0)
+ if (!StackPtr.getNode())
StackPtr = DAG.getCopyFromReg(Chain, dl, RegInfo->getStackRegister(),
getPointerTy());
MemOpChains.push_back(LowerMemOpCallTo(Chain, StackPtr, Arg,
// registers used and is in the range 0 - 8 inclusive.
// Count the number of XMM registers allocated.
- static const uint16_t XMMArgRegs[] = {
+ static const MCPhysReg XMMArgRegs[] = {
X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
};
if (Flags.isByVal()) {
// Copy relative to framepointer.
SDValue Source = DAG.getIntPtrConstant(VA.getLocMemOffset());
- if (StackPtr.getNode() == 0)
+ if (!StackPtr.getNode())
StackPtr = DAG.getCopyFromReg(Chain, dl,
RegInfo->getStackRegister(),
getPointerTy());
// We should use extra load for direct calls to dllimported functions in
// non-JIT mode.
const GlobalValue *GV = G->getGlobal();
- if (!GV->hasDLLImportLinkage()) {
+ if (!GV->hasDLLImportStorageClass()) {
unsigned char OpFlags = 0;
bool ExtraLoad = false;
unsigned WrapperKind = ISD::DELETED_NODE;
SmallVector<SDValue, 8> Ops;
if (!IsSibcall && isTailCall) {
- Chain = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(NumBytes, true),
- DAG.getIntPtrConstant(0, true), InFlag, dl);
+ Chain = DAG.getCALLSEQ_END(Chain,
+ DAG.getIntPtrConstant(NumBytesToPop, true),
+ DAG.getIntPtrConstant(0, true), InFlag, dl);
InFlag = Chain.getValue(1);
}
InFlag = Chain.getValue(1);
// Create the CALLSEQ_END node.
- unsigned NumBytesForCalleeToPush;
+ unsigned NumBytesForCalleeToPop;
if (X86::isCalleePop(CallConv, Is64Bit, isVarArg,
getTargetMachine().Options.GuaranteedTailCallOpt))
- NumBytesForCalleeToPush = NumBytes; // Callee pops everything
+ NumBytesForCalleeToPop = NumBytes; // Callee pops everything
else if (!Is64Bit && !IsTailCallConvention(CallConv) &&
!Subtarget->getTargetTriple().isOSMSVCRT() &&
SR == StackStructReturn)
// pops the hidden struct pointer, so we have to push it back.
// This is common for Darwin/X86, Linux & Mingw32 targets.
// For MSVC Win32 targets, the caller pops the hidden struct pointer.
- NumBytesForCalleeToPush = 4;
+ NumBytesForCalleeToPop = 4;
else
- NumBytesForCalleeToPush = 0; // Callee pops nothing.
+ NumBytesForCalleeToPop = 0; // Callee pops nothing.
// Returns a flag for retval copy to use.
if (!IsSibcall) {
Chain = DAG.getCALLSEQ_END(Chain,
- DAG.getIntPtrConstant(NumBytes, true),
- DAG.getIntPtrConstant(NumBytesForCalleeToPush,
+ DAG.getIntPtrConstant(NumBytesToPop, true),
+ DAG.getIntPtrConstant(NumBytesForCalleeToPop,
true),
InFlag, dl);
InFlag = Chain.getValue(1);
return false;
}
+/// \brief Returns true if it is beneficial to convert a load of a constant
+/// to just the constant itself.
+bool X86TargetLowering::shouldConvertConstantLoadToIntImm(const APInt &Imm,
+ Type *Ty) const {
+ assert(Ty->isIntegerTy());
+
+ unsigned BitSize = Ty->getPrimitiveSizeInBits();
+ if (BitSize == 0 || BitSize > 64)
+ return false;
+ return true;
+}
+
/// isUndefOrInRange - Return true if Val is undef or if its value falls within
/// the specified range (L, H].
static bool isUndefOrInRange(int Val, int Low, int Hi) {
return true;
}
+/// isINSERTPSMask - Return true if the specified VECTOR_SHUFFLE operand
+/// specifies a shuffle of elements that is suitable for input to INSERTPS.
+/// i. e: If all but one element come from the same vector.
+static bool isINSERTPSMask(ArrayRef<int> Mask, MVT VT) {
+ // TODO: Deal with AVX's VINSERTPS
+ if (!VT.is128BitVector() || (VT != MVT::v4f32 && VT != MVT::v4i32))
+ return false;
+
+ unsigned CorrectPosV1 = 0;
+ unsigned CorrectPosV2 = 0;
+ for (int i = 0, e = (int)VT.getVectorNumElements(); i != e; ++i)
+ if (Mask[i] == i)
+ ++CorrectPosV1;
+ else if (Mask[i] == i + 4)
+ ++CorrectPosV2;
+
+ if (CorrectPosV1 == 3 || CorrectPosV2 == 3)
+ // We have 3 elements from one vector, and one from another.
+ return true;
+
+ return false;
+}
+
//
// Some special combinations that can be optimized.
//
/// isScalarLoadToVector - Returns true if the node is a scalar load that
/// is promoted to a vector. It also returns the LoadSDNode by reference if
/// required.
-static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = NULL) {
+static bool isScalarLoadToVector(SDNode *N, LoadSDNode **LD = nullptr) {
if (N->getOpcode() != ISD::SCALAR_TO_VECTOR)
return false;
N = N->getOperand(0).getNode();
SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v16i32, Ops, 16);
+ } else if (VT.getScalarType() == MVT::i1) {
+ assert(VT.getVectorNumElements() <= 16 && "Unexpected vector type");
+ SDValue Cst = DAG.getTargetConstant(0, MVT::i1);
+ SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst,
+ Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT,
+ Ops, VT.getVectorNumElements());
} else
llvm_unreachable("Unexpected vector type");
return SDValue();
SDLoc dl(Op);
- SDValue V(0, 0);
+ SDValue V;
bool First = true;
for (unsigned i = 0; i < 16; ++i) {
bool ThisIsNonZero = (NonZeros & (1 << i)) != 0;
}
if ((i & 1) != 0) {
- SDValue ThisElt(0, 0), LastElt(0, 0);
+ SDValue ThisElt, LastElt;
bool LastIsNonZero = (NonZeros & (1 << (i-1))) != 0;
if (LastIsNonZero) {
LastElt = DAG.getNode(ISD::ZERO_EXTEND, dl,
return SDValue();
SDLoc dl(Op);
- SDValue V(0, 0);
+ SDValue V;
bool First = true;
for (unsigned i = 0; i < 8; ++i) {
bool isNonZero = (NonZeros & (1 << i)) != 0;
EVT EltVT = VT.getVectorElementType();
unsigned NumElems = Elts.size();
- LoadSDNode *LDBase = NULL;
+ LoadSDNode *LDBase = nullptr;
unsigned LastLoadedElt = -1U;
// For each element in the initializer, see if we've found a load or an undef.
unsigned ScalarSize = CVT.getSizeInBits();
if (ScalarSize == 32 || (IsGE256 && ScalarSize == 64)) {
- const Constant *C = 0;
+ const Constant *C = nullptr;
if (ConstantSDNode *CI = dyn_cast<ConstantSDNode>(Ld))
C = CI->getConstantIntValue();
else if (ConstantFPSDNode *CF = dyn_cast<ConstantFPSDNode>(Ld))
return SDValue();
}
+/// \brief For an EXTRACT_VECTOR_ELT with a constant index return the real
+/// underlying vector and index.
+///
+/// Modifies \p ExtractedFromVec to the real vector and returns the real
+/// index.
+static int getUnderlyingExtractedFromVec(SDValue &ExtractedFromVec,
+ SDValue ExtIdx) {
+ int Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
+ if (!isa<ShuffleVectorSDNode>(ExtractedFromVec))
+ return Idx;
+
+ // For 256-bit vectors, LowerEXTRACT_VECTOR_ELT_SSE4 may have already
+ // lowered this:
+ // (extract_vector_elt (v8f32 %vreg1), Constant<6>)
+ // to:
+ // (extract_vector_elt (vector_shuffle<2,u,u,u>
+ // (extract_subvector (v8f32 %vreg0), Constant<4>),
+ // undef)
+ // Constant<0>)
+ // In this case the vector is the extract_subvector expression and the index
+ // is 2, as specified by the shuffle.
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(ExtractedFromVec);
+ SDValue ShuffleVec = SVOp->getOperand(0);
+ MVT ShuffleVecVT = ShuffleVec.getSimpleValueType();
+ assert(ShuffleVecVT.getVectorElementType() ==
+ ExtractedFromVec.getSimpleValueType().getVectorElementType());
+
+ int ShuffleIdx = SVOp->getMaskElt(Idx);
+ if (isUndefOrInRange(ShuffleIdx, 0, ShuffleVecVT.getVectorNumElements())) {
+ ExtractedFromVec = ShuffleVec;
+ return ShuffleIdx;
+ }
+ return Idx;
+}
+
static SDValue buildFromShuffleMostly(SDValue Op, SelectionDAG &DAG) {
MVT VT = Op.getSimpleValueType();
SDValue ExtractedFromVec = Op.getOperand(i).getOperand(0);
SDValue ExtIdx = Op.getOperand(i).getOperand(1);
+ // Quit if non-constant index.
+ if (!isa<ConstantSDNode>(ExtIdx))
+ return SDValue();
+ int Idx = getUnderlyingExtractedFromVec(ExtractedFromVec, ExtIdx);
// Quit if extracted from vector of different type.
if (ExtractedFromVec.getValueType() != VT)
return SDValue();
- // Quit if non-constant index.
- if (!isa<ConstantSDNode>(ExtIdx))
- return SDValue();
-
- if (VecIn1.getNode() == 0)
+ if (!VecIn1.getNode())
VecIn1 = ExtractedFromVec;
else if (VecIn1 != ExtractedFromVec) {
- if (VecIn2.getNode() == 0)
+ if (!VecIn2.getNode())
VecIn2 = ExtractedFromVec;
else if (VecIn2 != ExtractedFromVec)
// Quit if more than 2 vectors to shuffle
return SDValue();
}
- unsigned Idx = cast<ConstantSDNode>(ExtIdx)->getZExtValue();
-
if (ExtractedFromVec == VecIn1)
Mask[i] = Idx;
else if (ExtractedFromVec == VecIn2)
Mask[i] = Idx + NumElems;
}
- if (VecIn1.getNode() == 0)
+ if (!VecIn1.getNode())
return SDValue();
VecIn2 = VecIn2.getNode() ? VecIn2 : DAG.getUNDEF(VT);
bool AllContants = true;
uint64_t Immediate = 0;
+ int NonConstIdx = -1;
+ bool IsSplat = true;
+ unsigned NumNonConsts = 0;
+ unsigned NumConsts = 0;
for (unsigned idx = 0, e = Op.getNumOperands(); idx < e; ++idx) {
SDValue In = Op.getOperand(idx);
if (In.getOpcode() == ISD::UNDEF)
continue;
if (!isa<ConstantSDNode>(In)) {
AllContants = false;
- break;
+ NonConstIdx = idx;
+ NumNonConsts++;
}
- if (cast<ConstantSDNode>(In)->getZExtValue())
+ else {
+ NumConsts++;
+ if (cast<ConstantSDNode>(In)->getZExtValue())
Immediate |= (1ULL << idx);
+ }
+ if (In != Op.getOperand(0))
+ IsSplat = false;
}
if (AllContants) {
DAG.getIntPtrConstant(0));
}
- // Splat vector (with undefs)
- SDValue In = Op.getOperand(0);
- for (unsigned i = 1, e = Op.getNumOperands(); i != e; ++i) {
- if (Op.getOperand(i) != In && Op.getOperand(i).getOpcode() != ISD::UNDEF)
- llvm_unreachable("Unsupported predicate operation");
- }
-
- SDValue EFLAGS, X86CC;
- if (In.getOpcode() == ISD::SETCC) {
- SDValue Op0 = In.getOperand(0);
- SDValue Op1 = In.getOperand(1);
- ISD::CondCode CC = cast<CondCodeSDNode>(In.getOperand(2))->get();
- bool isFP = Op1.getValueType().isFloatingPoint();
- unsigned X86CCVal = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
-
- assert(X86CCVal != X86::COND_INVALID && "Unsupported predicate operation");
-
- X86CC = DAG.getConstant(X86CCVal, MVT::i8);
- EFLAGS = EmitCmp(Op0, Op1, X86CCVal, DAG);
- EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
- } else if (In.getOpcode() == X86ISD::SETCC) {
- X86CC = In.getOperand(0);
- EFLAGS = In.getOperand(1);
- } else {
- // The algorithm:
- // Bit1 = In & 0x1
- // if (Bit1 != 0)
- // ZF = 0
- // else
- // ZF = 1
- // if (ZF == 0)
- // res = allOnes ### CMOVNE -1, %res
- // else
- // res = allZero
- MVT InVT = In.getSimpleValueType();
- SDValue Bit1 = DAG.getNode(ISD::AND, dl, InVT, In, DAG.getConstant(1, InVT));
- EFLAGS = EmitTest(Bit1, X86::COND_NE, DAG);
- X86CC = DAG.getConstant(X86::COND_NE, MVT::i8);
- }
-
- if (VT == MVT::v16i1) {
- SDValue Cst1 = DAG.getConstant(-1, MVT::i16);
- SDValue Cst0 = DAG.getConstant(0, MVT::i16);
- SDValue CmovOp = DAG.getNode(X86ISD::CMOV, dl, MVT::i16,
- Cst0, Cst1, X86CC, EFLAGS);
- return DAG.getNode(ISD::BITCAST, dl, VT, CmovOp);
- }
-
- if (VT == MVT::v8i1) {
- SDValue Cst1 = DAG.getConstant(-1, MVT::i32);
- SDValue Cst0 = DAG.getConstant(0, MVT::i32);
- SDValue CmovOp = DAG.getNode(X86ISD::CMOV, dl, MVT::i32,
- Cst0, Cst1, X86CC, EFLAGS);
- CmovOp = DAG.getNode(ISD::TRUNCATE, dl, MVT::i8, CmovOp);
- return DAG.getNode(ISD::BITCAST, dl, VT, CmovOp);
- }
- llvm_unreachable("Unsupported predicate operation");
+ if (NumNonConsts == 1 && NonConstIdx != 0) {
+ SDValue DstVec;
+ if (NumConsts) {
+ SDValue VecAsImm = DAG.getConstant(Immediate,
+ MVT::getIntegerVT(VT.getSizeInBits()));
+ DstVec = DAG.getNode(ISD::BITCAST, dl, VT, VecAsImm);
+ }
+ else
+ DstVec = DAG.getUNDEF(VT);
+ return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, DstVec,
+ Op.getOperand(NonConstIdx),
+ DAG.getIntPtrConstant(NonConstIdx));
+ }
+ if (!IsSplat && (NonConstIdx != 0))
+ llvm_unreachable("Unsupported BUILD_VECTOR operation");
+ MVT SelectVT = (VT == MVT::v16i1)? MVT::i16 : MVT::i8;
+ SDValue Select;
+ if (IsSplat)
+ Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
+ DAG.getConstant(-1, SelectVT),
+ DAG.getConstant(0, SelectVT));
+ else
+ Select = DAG.getNode(ISD::SELECT, dl, SelectVT, Op.getOperand(0),
+ DAG.getConstant((Immediate | 1), SelectVT),
+ DAG.getConstant(Immediate, SelectVT));
+ return DAG.getNode(ISD::BITCAST, dl, VT, Select);
}
SDValue
// For AVX-length vectors, build the individual 128-bit pieces and use
// shuffles to put them in place.
- if (VT.is256BitVector()) {
- SmallVector<SDValue, 32> V;
+ if (VT.is256BitVector() || VT.is512BitVector()) {
+ SmallVector<SDValue, 64> V;
for (unsigned i = 0; i != NumElems; ++i)
V.push_back(Op.getOperand(i));
NumElems/2);
// Recreate the wider vector with the lower and upper part.
- return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
+ if (VT.is256BitVector())
+ return Concat128BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
+ return Concat256BitVectors(Lower, Upper, VT, NumElems, DAG, dl);
}
// Let legalizer expand 2-wide build_vectors.
return DAG.getNode(ISD::BITCAST, dl, VT, Ret);
}
+/// In vector type \p VT, return true if the element at index \p InputIdx
+/// falls on a different 128-bit lane than \p OutputIdx.
+static bool ShuffleCrosses128bitLane(MVT VT, unsigned InputIdx,
+ unsigned OutputIdx) {
+ unsigned EltSize = VT.getVectorElementType().getSizeInBits();
+ return InputIdx * EltSize / 128 != OutputIdx * EltSize / 128;
+}
+
+/// Generate a PSHUFB if possible. Selects elements from \p V1 according to
+/// \p MaskVals. MaskVals[OutputIdx] = InputIdx specifies that we want to
+/// shuffle the element at InputIdx in V1 to OutputIdx in the result. If \p
+/// MaskVals refers to elements outside of \p V1 or is undef (-1), insert a
+/// zero.
+static SDValue getPSHUFB(ArrayRef<int> MaskVals, SDValue V1, SDLoc &dl,
+ SelectionDAG &DAG) {
+ MVT VT = V1.getSimpleValueType();
+ assert(VT.is128BitVector() || VT.is256BitVector());
+
+ MVT EltVT = VT.getVectorElementType();
+ unsigned EltSizeInBytes = EltVT.getSizeInBits() / 8;
+ unsigned NumElts = VT.getVectorNumElements();
+
+ SmallVector<SDValue, 32> PshufbMask;
+ for (unsigned OutputIdx = 0; OutputIdx < NumElts; ++OutputIdx) {
+ int InputIdx = MaskVals[OutputIdx];
+ unsigned InputByteIdx;
+
+ if (InputIdx < 0 || NumElts <= (unsigned)InputIdx)
+ InputByteIdx = 0x80;
+ else {
+ // Cross lane is not allowed.
+ if (ShuffleCrosses128bitLane(VT, InputIdx, OutputIdx))
+ return SDValue();
+ InputByteIdx = InputIdx * EltSizeInBytes;
+ // Index is an byte offset within the 128-bit lane.
+ InputByteIdx &= 0xf;
+ }
+
+ for (unsigned j = 0; j < EltSizeInBytes; ++j) {
+ PshufbMask.push_back(DAG.getConstant(InputByteIdx, MVT::i8));
+ if (InputByteIdx != 0x80)
+ ++InputByteIdx;
+ }
+ }
+
+ MVT ShufVT = MVT::getVectorVT(MVT::i8, PshufbMask.size());
+ if (ShufVT != VT)
+ V1 = DAG.getNode(ISD::BITCAST, dl, ShufVT, V1);
+ return DAG.getNode(X86ISD::PSHUFB, dl, ShufVT, V1,
+ DAG.getNode(ISD::BUILD_VECTOR, dl, ShufVT,
+ PshufbMask.data(), PshufbMask.size()));
+}
+
// v8i16 shuffles - Prefer shuffles in the following order:
// 1. [all] pshuflw, pshufhw, optional move
// 2. [ssse3] 1 x pshufb
// Determine if more than 1 of the words in each of the low and high quadwords
// of the result come from the same quadword of one of the two inputs. Undef
// mask values count as coming from any quadword, for better codegen.
+ //
+ // Lo/HiQuad[i] = j indicates how many words from the ith quad of the input
+ // feeds this quad. For i, 0 and 1 refer to V1, 2 and 3 refer to V2.
unsigned LoQuad[] = { 0, 0, 0, 0 };
unsigned HiQuad[] = { 0, 0, 0, 0 };
+ // Indices of quads used.
std::bitset<4> InputQuads;
for (unsigned i = 0; i < 8; ++i) {
unsigned *Quad = i < 4 ? LoQuad : HiQuad;
// For SSSE3, If all 8 words of the result come from only 1 quadword of each
// of the two input vectors, shuffle them into one input vector so only a
- // single pshufb instruction is necessary. If There are more than 2 input
+ // single pshufb instruction is necessary. If there are more than 2 input
// quads, disable the next transformation since it does not help SSSE3.
bool V1Used = InputQuads[0] || InputQuads[1];
bool V2Used = InputQuads[2] || InputQuads[3];
// mask, and elements that come from V1 in the V2 mask, so that the two
// results can be OR'd together.
bool TwoInputs = V1Used && V2Used;
- for (unsigned i = 0; i != 8; ++i) {
- int EltIdx = MaskVals[i] * 2;
- int Idx0 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx;
- int Idx1 = (TwoInputs && (EltIdx >= 16)) ? 0x80 : EltIdx+1;
- pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
- }
- V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V1);
- V1 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V1,
- DAG.getNode(ISD::BUILD_VECTOR, dl,
- MVT::v16i8, &pshufbMask[0], 16));
+ V1 = getPSHUFB(MaskVals, V1, dl, DAG);
if (!TwoInputs)
return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
// Calculate the shuffle mask for the second input, shuffle it, and
// OR it with the first shuffled input.
- pshufbMask.clear();
- for (unsigned i = 0; i != 8; ++i) {
- int EltIdx = MaskVals[i] * 2;
- int Idx0 = (EltIdx < 16) ? 0x80 : EltIdx - 16;
- int Idx1 = (EltIdx < 16) ? 0x80 : EltIdx - 15;
- pshufbMask.push_back(DAG.getConstant(Idx0, MVT::i8));
- pshufbMask.push_back(DAG.getConstant(Idx1, MVT::i8));
- }
- V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V2);
- V2 = DAG.getNode(X86ISD::PSHUFB, dl, MVT::v16i8, V2,
- DAG.getNode(ISD::BUILD_VECTOR, dl,
- MVT::v16i8, &pshufbMask[0], 16));
+ CommuteVectorShuffleMask(MaskVals, 8);
+ V2 = getPSHUFB(MaskVals, V2, dl, DAG);
V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
}
return NewV;
}
+/// \brief v16i16 shuffles
+///
+/// FIXME: We only support generation of a single pshufb currently. We can
+/// generalize the other applicable cases from LowerVECTOR_SHUFFLEv8i16 as
+/// well (e.g 2 x pshufb + 1 x por).
+static SDValue
+LowerVECTOR_SHUFFLEv16i16(SDValue Op, SelectionDAG &DAG) {
+ ShuffleVectorSDNode *SVOp = cast<ShuffleVectorSDNode>(Op);
+ SDValue V1 = SVOp->getOperand(0);
+ SDValue V2 = SVOp->getOperand(1);
+ SDLoc dl(SVOp);
+
+ if (V2.getOpcode() != ISD::UNDEF)
+ return SDValue();
+
+ SmallVector<int, 16> MaskVals(SVOp->getMask().begin(), SVOp->getMask().end());
+ return getPSHUFB(MaskVals, V1, dl, DAG);
+}
+
// v16i8 shuffles - Prefer shuffles in the following order:
// 1. [ssse3] 1 x pshufb
// 2. [ssse3] 2 x pshufb + 1 x por
CommuteVectorShuffleMask(MaskVals, 32);
V1 = V2;
}
- SmallVector<SDValue, 32> pshufbMask;
- for (unsigned i = 0; i != 32; i++) {
- int EltIdx = MaskVals[i];
- if (EltIdx < 0 || EltIdx >= 32)
- EltIdx = 0x80;
- else {
- if ((EltIdx >= 16 && i < 16) || (EltIdx < 16 && i >= 16))
- // Cross lane is not allowed.
- return SDValue();
- EltIdx &= 0xf;
- }
- pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
- }
- return DAG.getNode(X86ISD::PSHUFB, dl, MVT::v32i8, V1,
- DAG.getNode(ISD::BUILD_VECTOR, dl,
- MVT::v32i8, &pshufbMask[0], 32));
+ return getPSHUFB(MaskVals, V1, dl, DAG);
}
/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
SDValue SrcOp, SelectionDAG &DAG,
const X86Subtarget *Subtarget, SDLoc dl) {
if (VT == MVT::v2f64 || VT == MVT::v4f32) {
- LoadSDNode *LD = NULL;
+ LoadSDNode *LD = nullptr;
if (!isScalarLoadToVector(SrcOp.getNode(), &LD))
LD = dyn_cast<LoadSDNode>(SrcOp);
if (!LD) {
getShuffleSHUFImmediate(SVOp), DAG);
}
+// It is only safe to call this function if isINSERTPSMask is true for
+// this shufflevector mask.
+static SDValue getINSERTPS(ShuffleVectorSDNode *SVOp, SDLoc &dl,
+ SelectionDAG &DAG) {
+ // Generate an insertps instruction when inserting an f32 from memory onto a
+ // v4f32 or when copying a member from one v4f32 to another.
+ // We also use it for transferring i32 from one register to another,
+ // since it simply copies the same bits.
+ // If we're transfering an i32 from memory to a specific element in a
+ // register, we output a generic DAG that will match the PINSRD
+ // instruction.
+ // TODO: Optimize for AVX cases too (VINSERTPS)
+ MVT VT = SVOp->getSimpleValueType(0);
+ MVT EVT = VT.getVectorElementType();
+ SDValue V1 = SVOp->getOperand(0);
+ SDValue V2 = SVOp->getOperand(1);
+ auto Mask = SVOp->getMask();
+ assert((VT == MVT::v4f32 || VT == MVT::v4i32) &&
+ "unsupported vector type for insertps/pinsrd");
+
+ int FromV1 = std::count_if(Mask.begin(), Mask.end(),
+ [](const int &i) { return i < 4; });
+
+ SDValue From;
+ SDValue To;
+ unsigned DestIndex;
+ if (FromV1 == 1) {
+ From = V1;
+ To = V2;
+ DestIndex = std::find_if(Mask.begin(), Mask.end(),
+ [](const int &i) { return i < 4; }) -
+ Mask.begin();
+ } else {
+ From = V2;
+ To = V1;
+ DestIndex = std::find_if(Mask.begin(), Mask.end(),
+ [](const int &i) { return i >= 4; }) -
+ Mask.begin();
+ }
+
+ if (MayFoldLoad(From)) {
+ // Trivial case, when From comes from a load and is only used by the
+ // shuffle. Make it use insertps from the vector that we need from that
+ // load.
+ SDValue Addr = From.getOperand(1);
+ SDValue NewAddr =
+ DAG.getNode(ISD::ADD, dl, Addr.getSimpleValueType(), Addr,
+ DAG.getConstant(DestIndex * EVT.getStoreSize(),
+ Addr.getSimpleValueType()));
+
+ LoadSDNode *Load = cast<LoadSDNode>(From);
+ SDValue NewLoad =
+ DAG.getLoad(EVT, dl, Load->getChain(), NewAddr,
+ DAG.getMachineFunction().getMachineMemOperand(
+ Load->getMemOperand(), 0, EVT.getStoreSize()));
+
+ if (EVT == MVT::f32) {
+ // Create this as a scalar to vector to match the instruction pattern.
+ SDValue LoadScalarToVector =
+ DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, NewLoad);
+ SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4);
+ return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, LoadScalarToVector,
+ InsertpsMask);
+ } else { // EVT == MVT::i32
+ // If we're getting an i32 from memory, use an INSERT_VECTOR_ELT
+ // instruction, to match the PINSRD instruction, which loads an i32 to a
+ // certain vector element.
+ return DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, To, NewLoad,
+ DAG.getConstant(DestIndex, MVT::i32));
+ }
+ }
+
+ // Vector-element-to-vector
+ unsigned SrcIndex = Mask[DestIndex] % 4;
+ SDValue InsertpsMask = DAG.getIntPtrConstant(DestIndex << 4 | SrcIndex << 6);
+ return DAG.getNode(X86ISD::INSERTPS, dl, VT, To, From, InsertpsMask);
+}
+
// Reduce a vector shuffle to zext.
static SDValue LowerVectorIntExtend(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
if (V1IsUndef && V2IsUndef)
return DAG.getUNDEF(VT);
- assert(!V1IsUndef && "Op 1 of shuffle should not be undef");
+ // When we create a shuffle node we put the UNDEF node to second operand,
+ // but in some cases the first operand may be transformed to UNDEF.
+ // In this case we should just commute the node.
+ if (V1IsUndef)
+ return CommuteVectorShuffle(SVOp, DAG);
// Vector shuffle lowering takes 3 steps:
//
CommuteVectorShuffleMask(M, NumElems);
std::swap(V1, V2);
std::swap(V1IsSplat, V2IsSplat);
- Commuted = false;
if (isUNPCKLMask(M, VT, HasInt256))
return getTargetShuffleNode(X86ISD::UNPCKL, dl, VT, V1, V2, DAG);
if (BlendOp.getNode())
return BlendOp;
+ if (Subtarget->hasSSE41() && isINSERTPSMask(M, VT))
+ return getINSERTPS(SVOp, dl, DAG);
+
unsigned Imm8;
if (V2IsUndef && HasInt256 && isPermImmMask(M, VT, Imm8))
return getTargetShuffleNode(X86ISD::VPERMI, dl, VT, V1, Imm8, DAG);
// Bitcast is for VPERMPS since mask is v8i32 but node takes v8f32
return DAG.getNode(X86ISD::VPERMV, dl, VT,
DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1);
- return DAG.getNode(X86ISD::VPERMV3, dl, VT,
- DAG.getNode(ISD::BITCAST, dl, VT, Mask), V1, V2);
+ return DAG.getNode(X86ISD::VPERMV3, dl, VT, V1,
+ DAG.getNode(ISD::BITCAST, dl, VT, Mask), V2);
}
//===--------------------------------------------------------------------===//
return NewOp;
}
+ if (VT == MVT::v16i16 && Subtarget->hasInt256()) {
+ SDValue NewOp = LowerVECTOR_SHUFFLEv16i16(Op, DAG);
+ if (NewOp.getNode())
+ return NewOp;
+ }
+
if (VT == MVT::v16i8) {
SDValue NewOp = LowerVECTOR_SHUFFLEv16i8(SVOp, Subtarget, DAG);
if (NewOp.getNode())
/// Extract one bit from mask vector, like v16i1 or v8i1.
/// AVX-512 feature.
-static SDValue ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) {
+SDValue
+X86TargetLowering::ExtractBitFromMaskVector(SDValue Op, SelectionDAG &DAG) const {
SDValue Vec = Op.getOperand(0);
SDLoc dl(Vec);
MVT VecVT = Vec.getSimpleValueType();
}
unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
- if (IdxVal) {
- unsigned MaxSift = VecVT.getSizeInBits() - 1;
- Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
- DAG.getConstant(MaxSift - IdxVal, MVT::i8));
- Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
- DAG.getConstant(MaxSift, MVT::i8));
- }
- return DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i1, Vec,
+ const TargetRegisterClass* rc = getRegClassFor(VecVT);
+ unsigned MaxSift = rc->getSize()*8 - 1;
+ Vec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, Vec,
+ DAG.getConstant(MaxSift - IdxVal, MVT::i8));
+ Vec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, Vec,
+ DAG.getConstant(MaxSift, MVT::i8));
+ return DAG.getNode(X86ISD::VEXTRACT, dl, MVT::i1, Vec,
DAG.getIntPtrConstant(0));
}
return SDValue();
}
+/// Insert one bit to mask vector, like v16i1 or v8i1.
+/// AVX-512 feature.
+SDValue
+X86TargetLowering::InsertBitToMaskVector(SDValue Op, SelectionDAG &DAG) const {
+ SDLoc dl(Op);
+ SDValue Vec = Op.getOperand(0);
+ SDValue Elt = Op.getOperand(1);
+ SDValue Idx = Op.getOperand(2);
+ MVT VecVT = Vec.getSimpleValueType();
+
+ if (!isa<ConstantSDNode>(Idx)) {
+ // Non constant index. Extend source and destination,
+ // insert element and then truncate the result.
+ MVT ExtVecVT = (VecVT == MVT::v8i1 ? MVT::v8i64 : MVT::v16i32);
+ MVT ExtEltVT = (VecVT == MVT::v8i1 ? MVT::i64 : MVT::i32);
+ SDValue ExtOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ExtVecVT,
+ DAG.getNode(ISD::ZERO_EXTEND, dl, ExtVecVT, Vec),
+ DAG.getNode(ISD::ZERO_EXTEND, dl, ExtEltVT, Elt), Idx);
+ return DAG.getNode(ISD::TRUNCATE, dl, VecVT, ExtOp);
+ }
+
+ unsigned IdxVal = cast<ConstantSDNode>(Idx)->getZExtValue();
+ SDValue EltInVec = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VecVT, Elt);
+ if (Vec.getOpcode() == ISD::UNDEF)
+ return DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
+ DAG.getConstant(IdxVal, MVT::i8));
+ const TargetRegisterClass* rc = getRegClassFor(VecVT);
+ unsigned MaxSift = rc->getSize()*8 - 1;
+ EltInVec = DAG.getNode(X86ISD::VSHLI, dl, VecVT, EltInVec,
+ DAG.getConstant(MaxSift, MVT::i8));
+ EltInVec = DAG.getNode(X86ISD::VSRLI, dl, VecVT, EltInVec,
+ DAG.getConstant(MaxSift - IdxVal, MVT::i8));
+ return DAG.getNode(ISD::OR, dl, VecVT, Vec, EltInVec);
+}
SDValue
X86TargetLowering::LowerINSERT_VECTOR_ELT(SDValue Op, SelectionDAG &DAG) const {
MVT VT = Op.getSimpleValueType();
MVT EltVT = VT.getVectorElementType();
+
+ if (EltVT == MVT::i1)
+ return InsertBitToMaskVector(Op, DAG);
SDLoc dl(Op);
SDValue N0 = Op.getOperand(0);
static SDValue
LowerToTLSGeneralDynamicModel64(GlobalAddressSDNode *GA, SelectionDAG &DAG,
const EVT PtrVT) {
- return GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT,
+ return GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT,
X86::RAX, X86II::MO_TLSGD);
}
SDValue Base;
if (is64Bit) {
- Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, NULL, PtrVT, X86::RAX,
+ Base = GetTLSADDR(DAG, DAG.getEntryNode(), GA, nullptr, PtrVT, X86::RAX,
X86II::MO_TLSLD, /*LocalDynamic=*/true);
} else {
SDValue InFlag;
Chain.getValue(1));
}
- if (Subtarget->isTargetWindows() || Subtarget->isTargetMingw()) {
+ if (Subtarget->isTargetKnownWindowsMSVC() ||
+ Subtarget->isTargetWindowsGNU()) {
// Just use the implicit TLS architecture
// Need to generate someting similar to:
// mov rdx, qword [gs:abs 58H]; Load pointer to ThreadLocalStorage
// If GV is an alias then use the aliasee for determining
// thread-localness.
if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
- GV = GA->resolveAliasedGlobal(false);
+ GV = GA->getAliasedGlobal();
SDLoc dl(GA);
SDValue Chain = DAG.getEntryNode();
: Type::getInt32PtrTy(*DAG.getContext(),
257));
- SDValue TlsArray = Subtarget->is64Bit() ? DAG.getIntPtrConstant(0x58) :
- (Subtarget->isTargetMingw() ? DAG.getIntPtrConstant(0x2C) :
- DAG.getExternalSymbol("_tls_array", getPointerTy()));
+ SDValue TlsArray =
+ Subtarget->is64Bit()
+ ? DAG.getIntPtrConstant(0x58)
+ : (Subtarget->isTargetWindowsGNU()
+ ? DAG.getIntPtrConstant(0x2C)
+ : DAG.getExternalSymbol("_tls_array", getPointerTy()));
- SDValue ThreadPointer = DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
- MachinePointerInfo(Ptr),
- false, false, false, 0);
+ SDValue ThreadPointer =
+ DAG.getLoad(getPointerTy(), dl, Chain, TlsArray,
+ MachinePointerInfo(Ptr), false, false, false, 0);
// Load the _tls_index variable
SDValue IDX = DAG.getExternalSymbol("_tls_index", getPointerTy());
return SDValue();
if (Subtarget->hasInt256())
- return DAG.getNode(X86ISD::VZEXT_MOVL, dl, VT, In);
+ return DAG.getNode(X86ISD::VZEXT, dl, VT, In);
SDValue ZeroVec = getZeroVector(InVT, Subtarget, DAG, dl);
SDValue Undef = DAG.getUNDEF(InVT);
In = DAG.getNode(ISD::SIGN_EXTEND, DL, ExtVT, In);
InVT = ExtVT;
}
+
SDValue Cst = DAG.getTargetConstant(1, InVT.getVectorElementType());
const Constant *C = (dyn_cast<ConstantSDNode>(Cst))->getConstantIntValue();
SDValue CP = DAG.getConstantPool(C, getPointerTy());
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);
+ static const int ShufMask[] = {0, 2, 4, 6};
+ return DAG.getVectorShuffle(VT, DL, OpLo, OpHi, ShufMask);
}
if ((VT == MVT::v8i16) && (InVT == MVT::v8i32)) {
SDValue X86TargetLowering::LowerFP_TO_SINT(SDValue Op,
SelectionDAG &DAG) const {
- MVT VT = Op.getSimpleValueType();
- if (VT.isVector()) {
- if (VT == MVT::v8i16)
- return DAG.getNode(ISD::TRUNCATE, SDLoc(Op), VT,
- DAG.getNode(ISD::FP_TO_SINT, SDLoc(Op),
- MVT::v8i32, Op.getOperand(0)));
- return SDValue();
- }
+ assert(!Op.getSimpleValueType().isVector());
std::pair<SDValue,SDValue> Vals = FP_TO_INTHelper(Op, DAG,
/*IsSigned=*/ true, /*IsReplace=*/ false);
SDValue FIST = Vals.first, StackSlot = Vals.second;
// If FP_TO_INTHelper failed, the node is actually supposed to be Legal.
- if (FIST.getNode() == 0) return Op;
+ if (!FIST.getNode()) return Op;
if (StackSlot.getNode())
// Load the result.
SmallVector<SDValue, 8> Opnds;
DenseMap<SDValue, unsigned> VecInMap;
+ SmallVector<SDValue, 8> VecIns;
EVT VT = MVT::Other;
// Recognize a special case where a vector is casted into wide integer to
VT != VecInMap.begin()->first.getValueType())
return SDValue();
M = VecInMap.insert(std::make_pair(ExtractedFromVec, 0)).first;
+ VecIns.push_back(ExtractedFromVec);
}
M->second |= 1U << cast<ConstantSDNode>(Idx)->getZExtValue();
}
"Not extracted from 128-/256-bit vector.");
unsigned FullMask = (1U << VT.getVectorNumElements()) - 1U;
- SmallVector<SDValue, 8> VecIns;
for (DenseMap<SDValue, unsigned>::const_iterator
I = VecInMap.begin(), E = VecInMap.end(); I != E; ++I) {
// Quit if not all elements are used.
if (I->second != FullMask)
return SDValue();
- VecIns.push_back(I->first);
}
EVT TestVT = VT.is128BitVector() ? MVT::v2i64 : MVT::v4i64;
VecIns.back(), VecIns.back());
}
+/// \brief return true if \c Op has a use that doesn't just read flags.
+static bool hasNonFlagsUse(SDValue Op) {
+ for (SDNode::use_iterator UI = Op->use_begin(), UE = Op->use_end(); UI != UE;
+ ++UI) {
+ SDNode *User = *UI;
+ unsigned UOpNo = UI.getOperandNo();
+ if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
+ // Look pass truncate.
+ UOpNo = User->use_begin().getOperandNo();
+ User = *User->use_begin();
+ }
+
+ if (User->getOpcode() != ISD::BRCOND && User->getOpcode() != ISD::SETCC &&
+ !(User->getOpcode() == ISD::SELECT && UOpNo == 0))
+ return true;
+ }
+ return false;
+}
+
/// Emit nodes that will be selected as "test Op0,Op0", or something
/// equivalent.
-SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC,
+SDValue X86TargetLowering::EmitTest(SDValue Op, unsigned X86CC, SDLoc dl,
SelectionDAG &DAG) const {
- SDLoc dl(Op);
+ if (Op.getValueType() == MVT::i1)
+ // KORTEST instruction should be selected
+ return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
+ DAG.getConstant(0, Op.getValueType()));
// CF and OF aren't always set the way we want. Determine which
// of these we need.
NeedOF = true;
break;
}
-
// See if we can use the EFLAGS value from the operand instead of
// doing a separate TEST. TEST always sets OF and CF to 0, so unless
// we prove that the arithmetic won't overflow, we can't use OF or CF.
if (Op.getResNo() != 0 || NeedOF || NeedCF) {
// Emit a CMP with 0, which is the TEST pattern.
- if (Op.getValueType() == MVT::i1)
- return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
- DAG.getConstant(0, MVT::i1));
+ //if (Op.getValueType() == MVT::i1)
+ // return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op,
+ // DAG.getConstant(0, MVT::i1));
return DAG.getNode(X86ISD::CMP, dl, MVT::i32, Op,
DAG.getConstant(0, Op.getValueType()));
}
Opcode = X86ISD::ADD;
NumOperands = 2;
break;
- case ISD::AND: {
- // If the primary and result isn't used, don't bother using X86ISD::AND,
- // because a TEST instruction will be better.
- bool NonFlagUse = false;
- for (SDNode::use_iterator UI = Op.getNode()->use_begin(),
- UE = Op.getNode()->use_end(); UI != UE; ++UI) {
- SDNode *User = *UI;
- unsigned UOpNo = UI.getOperandNo();
- if (User->getOpcode() == ISD::TRUNCATE && User->hasOneUse()) {
- // Look pass truncate.
- UOpNo = User->use_begin().getOperandNo();
- User = *User->use_begin();
- }
-
- if (User->getOpcode() != ISD::BRCOND &&
- User->getOpcode() != ISD::SETCC &&
- !(User->getOpcode() == ISD::SELECT && UOpNo == 0)) {
- NonFlagUse = true;
+ case ISD::SHL:
+ case ISD::SRL:
+ // If we have a constant logical shift that's only used in a comparison
+ // against zero turn it into an equivalent AND. This allows turning it into
+ // a TEST instruction later.
+ if ((X86CC == X86::COND_E || X86CC == X86::COND_NE) &&
+ isa<ConstantSDNode>(Op->getOperand(1)) && !hasNonFlagsUse(Op)) {
+ EVT VT = Op.getValueType();
+ unsigned BitWidth = VT.getSizeInBits();
+ unsigned ShAmt = Op->getConstantOperandVal(1);
+ if (ShAmt >= BitWidth) // Avoid undefined shifts.
break;
- }
+ APInt Mask = ArithOp.getOpcode() == ISD::SRL
+ ? APInt::getHighBitsSet(BitWidth, BitWidth - ShAmt)
+ : APInt::getLowBitsSet(BitWidth, BitWidth - ShAmt);
+ if (!Mask.isSignedIntN(32)) // Avoid large immediates.
+ break;
+ SDValue New = DAG.getNode(ISD::AND, dl, VT, Op->getOperand(0),
+ DAG.getConstant(Mask, VT));
+ DAG.ReplaceAllUsesWith(Op, New);
+ Op = New;
}
+ break;
- if (!NonFlagUse)
+ case ISD::AND:
+ // If the primary and result isn't used, don't bother using X86ISD::AND,
+ // because a TEST instruction will be better.
+ if (!hasNonFlagsUse(Op))
break;
- }
// FALL THROUGH
case ISD::SUB:
case ISD::OR:
/// Emit nodes that will be selected as "cmp Op0,Op1", or something
/// equivalent.
SDValue X86TargetLowering::EmitCmp(SDValue Op0, SDValue Op1, unsigned X86CC,
- SelectionDAG &DAG) const {
- SDLoc dl(Op0);
+ SDLoc dl, SelectionDAG &DAG) const {
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(Op1)) {
if (C->getAPIntValue() == 0)
- return EmitTest(Op0, X86CC, DAG);
+ return EmitTest(Op0, X86CC, dl, DAG);
- if (Op0.getValueType() == MVT::i1) {
- Op0 = DAG.getNode(ISD::XOR, dl, MVT::i1, Op0,
- DAG.getConstant(-1, MVT::i1));
- return DAG.getNode(X86ISD::CMP, dl, MVT::i1, Op0,
- DAG.getConstant(0, MVT::i1));
- }
+ if (Op0.getValueType() == MVT::i1)
+ llvm_unreachable("Unexpected comparison operation for MVT::i1 operands");
}
if ((Op0.getValueType() == MVT::i8 || Op0.getValueType() == MVT::i16 ||
Op0.getValueType() == MVT::i32 || Op0.getValueType() == MVT::i64)) {
- // Do the comparison at i32 if it's smaller. This avoids subregister
- // aliasing issues. Keep the smaller reference if we're optimizing for
- // size, however, as that'll allow better folding of memory operations.
+ // Do the comparison at i32 if it's smaller, besides the Atom case.
+ // This avoids subregister aliasing issues. Keep the smaller reference
+ // if we're optimizing for size, however, as that'll allow better folding
+ // of memory operations.
if (Op0.getValueType() != MVT::i32 && Op0.getValueType() != MVT::i64 &&
!DAG.getMachineFunction().getFunction()->getAttributes().hasAttribute(
- AttributeSet::FunctionIndex, Attribute::MinSize)) {
+ AttributeSet::FunctionIndex, Attribute::MinSize) &&
+ !Subtarget->isAtom()) {
unsigned ExtendOp =
isX86CCUnsigned(X86CC) ? ISD::ZERO_EXTEND : ISD::SIGN_EXTEND;
Op0 = DAG.getNode(ExtendOp, dl, MVT::i32, Op0);
DAG.getNode(Op.getOpcode(), dl, NewVT, LHS2, RHS2, CC));
}
-static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG) {
+static SDValue LowerIntVSETCC_AVX512(SDValue Op, SelectionDAG &DAG,
+ const X86Subtarget *Subtarget) {
SDValue Op0 = Op.getOperand(0);
SDValue Op1 = Op.getOperand(1);
SDValue CC = Op.getOperand(2);
MVT VT = Op.getSimpleValueType();
+ SDLoc dl(Op);
assert(Op0.getValueType().getVectorElementType().getSizeInBits() >= 32 &&
Op.getValueType().getScalarType() == MVT::i1 &&
"Cannot set masked compare for this operation");
ISD::CondCode SetCCOpcode = cast<CondCodeSDNode>(CC)->get();
- SDLoc dl(Op);
-
+ unsigned Opc = 0;
bool Unsigned = false;
+ bool Swap = false;
unsigned SSECC;
switch (SetCCOpcode) {
default: llvm_unreachable("Unexpected SETCC condition");
case ISD::SETNE: SSECC = 4; break;
- case ISD::SETEQ: SSECC = 0; break;
- case ISD::SETUGT: Unsigned = true;
- case ISD::SETGT: SSECC = 6; break; // NLE
- case ISD::SETULT: Unsigned = true;
- case ISD::SETLT: SSECC = 1; break;
- case ISD::SETUGE: Unsigned = true;
- case ISD::SETGE: SSECC = 5; break; // NLT
- case ISD::SETULE: Unsigned = true;
+ case ISD::SETEQ: Opc = X86ISD::PCMPEQM; break;
+ case ISD::SETUGT: SSECC = 6; Unsigned = true; break;
+ case ISD::SETLT: Swap = true; //fall-through
+ case ISD::SETGT: Opc = X86ISD::PCMPGTM; break;
+ case ISD::SETULT: SSECC = 1; Unsigned = true; break;
+ case ISD::SETUGE: SSECC = 5; Unsigned = true; break; //NLT
+ case ISD::SETGE: Swap = true; SSECC = 2; break; // LE + swap
+ case ISD::SETULE: Unsigned = true; //fall-through
case ISD::SETLE: SSECC = 2; break;
}
- unsigned Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
+
+ if (Swap)
+ std::swap(Op0, Op1);
+ if (Opc)
+ return DAG.getNode(Opc, dl, VT, Op0, Op1);
+ Opc = Unsigned ? X86ISD::CMPMU: X86ISD::CMPM;
return DAG.getNode(Opc, dl, VT, Op0, Op1,
DAG.getConstant(SSECC, MVT::i8));
+}
+
+/// \brief Try to turn a VSETULT into a VSETULE by modifying its second
+/// operand \p Op1. If non-trivial (for example because it's not constant)
+/// return an empty value.
+static SDValue ChangeVSETULTtoVSETULE(SDLoc dl, SDValue Op1, SelectionDAG &DAG)
+{
+ BuildVectorSDNode *BV = dyn_cast<BuildVectorSDNode>(Op1.getNode());
+ if (!BV)
+ return SDValue();
+ MVT VT = Op1.getSimpleValueType();
+ MVT EVT = VT.getVectorElementType();
+ unsigned n = VT.getVectorNumElements();
+ SmallVector<SDValue, 8> ULTOp1;
+
+ for (unsigned i = 0; i < n; ++i) {
+ ConstantSDNode *Elt = dyn_cast<ConstantSDNode>(BV->getOperand(i));
+ if (!Elt || Elt->isOpaque() || Elt->getValueType(0) != EVT)
+ return SDValue();
+
+ // Avoid underflow.
+ APInt Val = Elt->getAPIntValue();
+ if (Val == 0)
+ return SDValue();
+
+ ULTOp1.push_back(DAG.getConstant(Val - 1, EVT));
+ }
+
+ return DAG.getNode(ISD::BUILD_VECTOR, dl, VT, ULTOp1.data(), ULTOp1.size());
}
static SDValue LowerVSETCC(SDValue Op, const X86Subtarget *Subtarget,
if (Subtarget->hasAVX512()) {
if (Op1.getValueType().is512BitVector() ||
(MaskResult && OpVT.getVectorElementType().getSizeInBits() >= 32))
- return LowerIntVSETCC_AVX512(Op, DAG);
+ return LowerIntVSETCC_AVX512(Op, DAG, Subtarget);
// In AVX-512 architecture setcc returns mask with i1 elements,
// But there is no compare instruction for i8 and i16 elements.
// operations may be required for some comparisons.
unsigned Opc;
bool Swap = false, Invert = false, FlipSigns = false, MinMax = false;
+ bool Subus = false;
switch (SetCCOpcode) {
default: llvm_unreachable("Unexpected SETCC condition");
case ISD::SETNE: Invert = true;
- case ISD::SETEQ: Opc = MaskResult? X86ISD::PCMPEQM: X86ISD::PCMPEQ; break;
+ case ISD::SETEQ: Opc = X86ISD::PCMPEQ; break;
case ISD::SETLT: Swap = true;
- case ISD::SETGT: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT; break;
+ case ISD::SETGT: Opc = X86ISD::PCMPGT; break;
case ISD::SETGE: Swap = true;
- case ISD::SETLE: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
+ case ISD::SETLE: Opc = X86ISD::PCMPGT;
Invert = true; break;
case ISD::SETULT: Swap = true;
- case ISD::SETUGT: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
+ case ISD::SETUGT: Opc = X86ISD::PCMPGT;
FlipSigns = true; break;
case ISD::SETUGE: Swap = true;
- case ISD::SETULE: Opc = MaskResult? X86ISD::PCMPGTM: X86ISD::PCMPGT;
+ case ISD::SETULE: Opc = X86ISD::PCMPGT;
FlipSigns = true; Invert = true; break;
}
if (MinMax) { Swap = false; Invert = false; FlipSigns = false; }
}
+ bool hasSubus = Subtarget->hasSSE2() && (VET == MVT::i8 || VET == MVT::i16);
+ if (!MinMax && hasSubus) {
+ // As another special case, use PSUBUS[BW] when it's profitable. E.g. for
+ // Op0 u<= Op1:
+ // t = psubus Op0, Op1
+ // pcmpeq t, <0..0>
+ switch (SetCCOpcode) {
+ default: break;
+ case ISD::SETULT: {
+ // If the comparison is against a constant we can turn this into a
+ // setule. With psubus, setule does not require a swap. This is
+ // beneficial because the constant in the register is no longer
+ // destructed as the destination so it can be hoisted out of a loop.
+ // Only do this pre-AVX since vpcmp* is no longer destructive.
+ if (Subtarget->hasAVX())
+ break;
+ SDValue ULEOp1 = ChangeVSETULTtoVSETULE(dl, Op1, DAG);
+ if (ULEOp1.getNode()) {
+ Op1 = ULEOp1;
+ Subus = true; Invert = false; Swap = false;
+ }
+ break;
+ }
+ // Psubus is better than flip-sign because it requires no inversion.
+ case ISD::SETUGE: Subus = true; Invert = false; Swap = true; break;
+ case ISD::SETULE: Subus = true; Invert = false; Swap = false; break;
+ }
+
+ if (Subus) {
+ Opc = X86ISD::SUBUS;
+ FlipSigns = false;
+ }
+ }
+
if (Swap)
std::swap(Op0, Op1);
if (MinMax)
Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Op0, Result);
+ if (Subus)
+ Result = DAG.getNode(X86ISD::PCMPEQ, dl, VT, Result,
+ getZeroVector(VT, Subtarget, DAG, dl));
+
return Result;
}
X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
bool Invert = (CC == ISD::SETNE) ^
cast<ConstantSDNode>(Op1)->isNullValue();
- if (!Invert) return Op0;
+ if (!Invert)
+ return Op0;
CCode = X86::GetOppositeBranchCondition(CCode);
- return DAG.getNode(X86ISD::SETCC, dl, VT,
- DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(CCode, MVT::i8),
+ Op0.getOperand(1));
+ if (VT == MVT::i1)
+ return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
+ return SetCC;
}
}
+ if ((Op0.getValueType() == MVT::i1) && (Op1.getOpcode() == ISD::Constant) &&
+ (cast<ConstantSDNode>(Op1)->getZExtValue() == 1) &&
+ (CC == ISD::SETEQ || CC == ISD::SETNE)) {
+
+ ISD::CondCode NewCC = ISD::getSetCCInverse(CC, true);
+ return DAG.getSetCC(dl, VT, Op0, DAG.getConstant(0, MVT::i1), NewCC);
+ }
bool isFP = Op1.getSimpleValueType().isFloatingPoint();
unsigned X86CC = TranslateX86CC(CC, isFP, Op0, Op1, DAG);
if (X86CC == X86::COND_INVALID)
return SDValue();
- SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
+ SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, dl, DAG);
EFLAGS = ConvertCmpIfNecessary(EFLAGS, DAG);
- return DAG.getNode(X86ISD::SETCC, dl, VT,
- DAG.getConstant(X86CC, MVT::i8), EFLAGS);
+ SDValue SetCC = DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), EFLAGS);
+ if (VT == MVT::i1)
+ return DAG.getNode(ISD::TRUNCATE, dl, MVT::i1, SetCC);
+ return SetCC;
}
// isX86LogicalCmp - Return true if opcode is a X86 logical comparison.
Res = DAG.getNOT(DL, Res, Res.getValueType());
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
- if (N2C == 0 || !N2C->isNullValue())
+ if (!N2C || !N2C->isNullValue())
Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
return Res;
}
if (addTest) {
CC = DAG.getConstant(X86::COND_NE, MVT::i8);
- Cond = EmitTest(Cond, X86::COND_NE, DAG);
+ Cond = EmitTest(Cond, X86::COND_NE, DL, DAG);
}
// a < b ? -1 : 0 -> RES = ~setcc_carry
return SDValue();
if (Subtarget->hasInt256())
- return DAG.getNode(X86ISD::VSEXT_MOVL, dl, VT, In);
+ return DAG.getNode(X86ISD::VSEXT, dl, VT, In);
// Optimize vectors in AVX mode
// Sign extend v8i16 to v8i32 and
MVT HalfVT = MVT::getVectorVT(VT.getScalarType(),
VT.getVectorNumElements()/2);
- OpLo = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpLo);
- OpHi = DAG.getNode(X86ISD::VSEXT_MOVL, dl, HalfVT, OpHi);
+ OpLo = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpLo);
+ OpHi = DAG.getNode(X86ISD::VSEXT, dl, HalfVT, OpHi);
return DAG.getNode(ISD::CONCAT_VECTORS, dl, VT, OpLo, OpHi);
}
unsigned X86Opcode;
unsigned X86Cond;
SDVTList VTs;
+ // Keep this in sync with LowerXALUO, otherwise we might create redundant
+ // instructions that can't be removed afterwards (i.e. X86ISD::ADD and
+ // X86ISD::INC).
switch (CondOpcode) {
case ISD::UADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_B; break;
- case ISD::SADDO: X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
+ case ISD::SADDO:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
+ if (C->isOne()) {
+ X86Opcode = X86ISD::INC; X86Cond = X86::COND_O;
+ break;
+ }
+ X86Opcode = X86ISD::ADD; X86Cond = X86::COND_O; break;
case ISD::USUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_B; break;
- case ISD::SSUBO: X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
+ case ISD::SSUBO:
+ if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(RHS))
+ if (C->isOne()) {
+ X86Opcode = X86ISD::DEC; X86Cond = X86::COND_O;
+ break;
+ }
+ X86Opcode = X86ISD::SUB; X86Cond = X86::COND_O; break;
case ISD::UMULO: X86Opcode = X86ISD::UMUL; X86Cond = X86::COND_O; break;
case ISD::SMULO: X86Opcode = X86ISD::SMUL; X86Cond = X86::COND_O; break;
default: llvm_unreachable("unexpected overflowing operator");
if (addTest) {
CC = DAG.getConstant(X86::COND_NE, MVT::i8);
- Cond = EmitTest(Cond, X86::COND_NE, DAG);
+ Cond = EmitTest(Cond, X86::COND_NE, dl, DAG);
}
Cond = ConvertCmpIfNecessary(Cond, DAG);
return DAG.getNode(X86ISD::BRCOND, dl, Op.getValueType(),
SDValue
X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
SelectionDAG &DAG) const {
- assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows() ||
- getTargetMachine().Options.EnableSegmentedStacks) &&
- "This should be used only on Windows targets or when segmented stacks "
- "are being used");
- assert(!Subtarget->isTargetMacho() && "Not implemented");
+ MachineFunction &MF = DAG.getMachineFunction();
+ bool SplitStack = MF.shouldSplitStack();
+ bool Lower = (Subtarget->isOSWindows() && !Subtarget->isTargetMacho()) ||
+ SplitStack;
SDLoc dl(Op);
+ if (!Lower) {
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ SDNode* Node = Op.getNode();
+
+ unsigned SPReg = TLI.getStackPointerRegisterToSaveRestore();
+ assert(SPReg && "Target cannot require DYNAMIC_STACKALLOC expansion and"
+ " not tell us which reg is the stack pointer!");
+ EVT VT = Node->getValueType(0);
+ SDValue Tmp1 = SDValue(Node, 0);
+ SDValue Tmp2 = SDValue(Node, 1);
+ SDValue Tmp3 = Node->getOperand(2);
+ SDValue Chain = Tmp1.getOperand(0);
+
+ // Chain the dynamic stack allocation so that it doesn't modify the stack
+ // pointer when other instructions are using the stack.
+ Chain = DAG.getCALLSEQ_START(Chain, DAG.getIntPtrConstant(0, true),
+ SDLoc(Node));
+
+ SDValue Size = Tmp2.getOperand(1);
+ SDValue SP = DAG.getCopyFromReg(Chain, dl, SPReg, VT);
+ Chain = SP.getValue(1);
+ unsigned Align = cast<ConstantSDNode>(Tmp3)->getZExtValue();
+ const TargetFrameLowering &TFI = *getTargetMachine().getFrameLowering();
+ unsigned StackAlign = TFI.getStackAlignment();
+ Tmp1 = DAG.getNode(ISD::SUB, dl, VT, SP, Size); // Value
+ if (Align > StackAlign)
+ Tmp1 = DAG.getNode(ISD::AND, dl, VT, Tmp1,
+ DAG.getConstant(-(uint64_t)Align, VT));
+ Chain = DAG.getCopyToReg(Chain, dl, SPReg, Tmp1); // Output chain
+
+ Tmp2 = DAG.getCALLSEQ_END(Chain, DAG.getIntPtrConstant(0, true),
+ DAG.getIntPtrConstant(0, true), SDValue(),
+ SDLoc(Node));
+
+ SDValue Ops[2] = { Tmp1, Tmp2 };
+ return DAG.getMergeValues(Ops, 2, dl);
+ }
+
// Get the inputs.
SDValue Chain = Op.getOperand(0);
SDValue Size = Op.getOperand(1);
bool Is64Bit = Subtarget->is64Bit();
EVT SPTy = Is64Bit ? MVT::i64 : MVT::i32;
- if (getTargetMachine().Options.EnableSegmentedStacks) {
- MachineFunction &MF = DAG.getMachineFunction();
+ if (SplitStack) {
MachineRegisterInfo &MRI = MF.getRegInfo();
if (Is64Bit) {
&& "Unknown target vector shift-by-constant node");
// Fold this packed vector shift into a build vector if SrcOp is a
- // vector of ConstantSDNodes or UNDEFs.
- if (ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
+ // vector of Constants or UNDEFs, and SrcOp valuetype is the same as VT.
+ if (VT == SrcOp.getSimpleValueType() &&
+ ISD::isBuildVectorOfConstantSDNodes(SrcOp.getNode())) {
SmallVector<SDValue, 8> Elts;
unsigned NumElts = SrcOp->getNumOperands();
ConstantSDNode *ND;
return DAG.getNode(X86ISD::PMULUDQ, dl, Op.getValueType(),
Op.getOperand(1), Op.getOperand(2));
+ case Intrinsic::x86_sse2_pmulhu_w:
+ case Intrinsic::x86_avx2_pmulhu_w:
+ return DAG.getNode(ISD::MULHU, dl, Op.getValueType(),
+ Op.getOperand(1), Op.getOperand(2));
+
+ case Intrinsic::x86_sse2_pmulh_w:
+ case Intrinsic::x86_avx2_pmulh_w:
+ return DAG.getNode(ISD::MULHS, 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:
return SDValue(Res, 1);
}
+// getReadTimeStampCounter - Handles the lowering of builtin intrinsics that
+// read the time stamp counter (x86_rdtsc and x86_rdtscp). This function is
+// also used to custom lower READCYCLECOUNTER nodes.
+static void getReadTimeStampCounter(SDNode *N, SDLoc DL, unsigned Opcode,
+ SelectionDAG &DAG, const X86Subtarget *Subtarget,
+ SmallVectorImpl<SDValue> &Results) {
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
+ SDValue TheChain = N->getOperand(0);
+ SDValue rd = DAG.getNode(Opcode, DL, Tys, &TheChain, 1);
+ SDValue LO, HI;
+
+ // The processor's time-stamp counter (a 64-bit MSR) is stored into the
+ // EDX:EAX registers. EDX is loaded with the high-order 32 bits of the MSR
+ // and the EAX register is loaded with the low-order 32 bits.
+ if (Subtarget->is64Bit()) {
+ LO = DAG.getCopyFromReg(rd, DL, X86::RAX, MVT::i64, rd.getValue(1));
+ HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::RDX, MVT::i64,
+ LO.getValue(2));
+ } else {
+ LO = DAG.getCopyFromReg(rd, DL, X86::EAX, MVT::i32, rd.getValue(1));
+ HI = DAG.getCopyFromReg(LO.getValue(1), DL, X86::EDX, MVT::i32,
+ LO.getValue(2));
+ }
+ SDValue Chain = HI.getValue(1);
+
+ if (Opcode == X86ISD::RDTSCP_DAG) {
+ assert(N->getNumOperands() == 3 && "Unexpected number of operands!");
+
+ // Instruction RDTSCP loads the IA32:TSC_AUX_MSR (address C000_0103H) into
+ // the ECX register. Add 'ecx' explicitly to the chain.
+ SDValue ecx = DAG.getCopyFromReg(Chain, DL, X86::ECX, MVT::i32,
+ HI.getValue(2));
+ // Explicitly store the content of ECX at the location passed in input
+ // to the 'rdtscp' intrinsic.
+ Chain = DAG.getStore(ecx.getValue(1), DL, ecx, N->getOperand(2),
+ MachinePointerInfo(), false, false, 0);
+ }
+
+ if (Subtarget->is64Bit()) {
+ // The EDX register is loaded with the high-order 32 bits of the MSR, and
+ // the EAX register is loaded with the low-order 32 bits.
+ SDValue Tmp = DAG.getNode(ISD::SHL, DL, MVT::i64, HI,
+ DAG.getConstant(32, MVT::i8));
+ Results.push_back(DAG.getNode(ISD::OR, DL, MVT::i64, LO, Tmp));
+ Results.push_back(Chain);
+ return;
+ }
+
+ // Use a buildpair to merge the two 32-bit values into a 64-bit one.
+ SDValue Ops[] = { LO, HI };
+ SDValue Pair = DAG.getNode(ISD::BUILD_PAIR, DL, MVT::i64, Ops,
+ array_lengthof(Ops));
+ Results.push_back(Pair);
+ Results.push_back(Chain);
+}
+
+static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SmallVector<SDValue, 2> Results;
+ SDLoc DL(Op);
+ getReadTimeStampCounter(Op.getNode(), DL, X86ISD::RDTSC_DAG, DAG, Subtarget,
+ Results);
+ return DAG.getMergeValues(&Results[0], Results.size(), DL);
+}
+
static SDValue LowerINTRINSIC_W_CHAIN(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
SDLoc dl(Op);
SDValue Scale = Op.getOperand(6);
return getMScatterNode(Opc, Op, DAG, Src, Mask, Base, Index, Scale, Chain);
}
+ // Read Time Stamp Counter (RDTSC).
+ case Intrinsic::x86_rdtsc:
+ // Read Time Stamp Counter and Processor ID (RDTSCP).
+ case Intrinsic::x86_rdtscp: {
+ unsigned Opc;
+ switch (IntNo) {
+ default: llvm_unreachable("Impossible intrinsic"); // Can't reach here.
+ case Intrinsic::x86_rdtsc:
+ Opc = X86ISD::RDTSC_DAG; break;
+ case Intrinsic::x86_rdtscp:
+ Opc = X86ISD::RDTSCP_DAG; break;
+ }
+ SmallVector<SDValue, 2> Results;
+ getReadTimeStampCounter(Op.getNode(), dl, Opc, DAG, Subtarget, Results);
+ return DAG.getMergeValues(&Results[0], Results.size(), dl);
+ }
// XTEST intrinsics.
case Intrinsic::x86_xtest: {
SDVTList VTs = DAG.getVTList(Op->getValueType(0), MVT::Other);
return DAG.getNode(ISD::ADD, dl, VT, Res, AhiBlo);
}
-static SDValue LowerSDIV(SDValue Op, SelectionDAG &DAG) {
- MVT VT = Op.getSimpleValueType();
- MVT EltTy = VT.getVectorElementType();
- unsigned NumElts = VT.getVectorNumElements();
- SDValue N0 = Op.getOperand(0);
+static SDValue LowerUMUL_LOHI(SDValue Op, const X86Subtarget *Subtarget,
+ SelectionDAG &DAG) {
+ SDValue Op0 = Op.getOperand(0), Op1 = Op.getOperand(1);
+ EVT VT = Op0.getValueType();
SDLoc dl(Op);
- // Lower sdiv X, pow2-const.
- BuildVectorSDNode *C = dyn_cast<BuildVectorSDNode>(Op.getOperand(1));
- if (!C)
- return SDValue();
+ assert((VT == MVT::v4i32 && Subtarget->hasSSE2()) ||
+ (VT == MVT::v8i32 && Subtarget->hasInt256()));
- APInt SplatValue, SplatUndef;
- unsigned SplatBitSize;
- bool HasAnyUndefs;
- if (!C->isConstantSplat(SplatValue, SplatUndef, SplatBitSize,
- HasAnyUndefs) ||
- EltTy.getSizeInBits() < SplatBitSize)
- return SDValue();
+ // Get the high parts.
+ const int Mask[] = {1, 2, 3, 4, 5, 6, 7, 8};
+ SDValue Hi0 = DAG.getVectorShuffle(VT, dl, Op0, Op0, Mask);
+ SDValue Hi1 = DAG.getVectorShuffle(VT, dl, Op1, Op1, Mask);
- if ((SplatValue != 0) &&
- (SplatValue.isPowerOf2() || (-SplatValue).isPowerOf2())) {
- unsigned Lg2 = SplatValue.countTrailingZeros();
- // Splat the sign bit.
- SmallVector<SDValue, 16> Sz(NumElts,
- DAG.getConstant(EltTy.getSizeInBits() - 1,
- EltTy));
- SDValue SGN = DAG.getNode(ISD::SRA, dl, VT, N0,
- DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Sz[0],
- NumElts));
- // Add (N0 < 0) ? abs2 - 1 : 0;
- SmallVector<SDValue, 16> Amt(NumElts,
- DAG.getConstant(EltTy.getSizeInBits() - Lg2,
- EltTy));
- SDValue SRL = DAG.getNode(ISD::SRL, dl, VT, SGN,
- DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Amt[0],
- NumElts));
- SDValue ADD = DAG.getNode(ISD::ADD, dl, VT, N0, SRL);
- SmallVector<SDValue, 16> Lg2Amt(NumElts, DAG.getConstant(Lg2, EltTy));
- SDValue SRA = DAG.getNode(ISD::SRA, dl, VT, ADD,
- DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Lg2Amt[0],
- NumElts));
-
- // If we're dividing by a positive value, we're done. Otherwise, we must
- // negate the result.
- if (SplatValue.isNonNegative())
- return SRA;
-
- SmallVector<SDValue, 16> V(NumElts, DAG.getConstant(0, EltTy));
- SDValue Zero = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &V[0], NumElts);
- return DAG.getNode(ISD::SUB, dl, VT, Zero, SRA);
- }
- return SDValue();
+ // Emit two multiplies, one for the lower 2 ints and one for the higher 2
+ // ints.
+ MVT MulVT = VT == MVT::v4i32 ? MVT::v2i64 : MVT::v4i64;
+ SDValue Mul1 = DAG.getNode(ISD::BITCAST, dl, VT,
+ DAG.getNode(X86ISD::PMULUDQ, dl, MulVT, Op0, Op1));
+ SDValue Mul2 = DAG.getNode(ISD::BITCAST, dl, VT,
+ DAG.getNode(X86ISD::PMULUDQ, dl, MulVT, Hi0, Hi1));
+
+ // Shuffle it back into the right order.
+ const int HighMask[] = {1, 3, 5, 7, 9, 11, 13, 15};
+ SDValue Highs = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, HighMask);
+ const int LowMask[] = {0, 2, 4, 6, 8, 10, 12, 14};
+ SDValue Lows = DAG.getVectorShuffle(VT, dl, Mul1, Mul2, LowMask);
+
+ return DAG.getNode(ISD::MERGE_VALUES, dl, Op.getValueType(), Highs, Lows);
}
static SDValue LowerScalarImmediateShift(SDValue Op, SelectionDAG &DAG,
uint64_t ShiftAmt = 0;
for (unsigned i = 0; i != Ratio; ++i) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Amt.getOperand(i));
- if (C == 0)
+ if (!C)
return SDValue();
// 6 == Log2(64)
ShiftAmt |= C->getZExtValue() << (i * (1 << (6 - RatioInLog2)));
for (unsigned j = 0; j != Ratio; ++j) {
ConstantSDNode *C =
dyn_cast<ConstantSDNode>(Amt.getOperand(i + j));
- if (C == 0)
+ if (!C)
return SDValue();
// 6 == Log2(64)
ShAmt |= C->getZExtValue() << (j * (1 << (6 - RatioInLog2)));
BaseShAmt = InVec.getOperand(1);
}
}
- if (BaseShAmt.getNode() == 0)
+ if (!BaseShAmt.getNode())
BaseShAmt = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, EltVT, Amt,
DAG.getIntPtrConstant(0));
}
return Op;
}
+ // If possible, lower this packed shift into a vector multiply instead of
+ // expanding it into a sequence of scalar shifts.
+ // Do this only if the vector shift count is a constant build_vector.
+ if (Op.getOpcode() == ISD::SHL &&
+ (VT == MVT::v8i16 || VT == MVT::v4i32 ||
+ (Subtarget->hasInt256() && VT == MVT::v16i16)) &&
+ ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
+ SmallVector<SDValue, 8> Elts;
+ EVT SVT = VT.getScalarType();
+ unsigned SVTBits = SVT.getSizeInBits();
+ const APInt &One = APInt(SVTBits, 1);
+ unsigned NumElems = VT.getVectorNumElements();
+
+ for (unsigned i=0; i !=NumElems; ++i) {
+ SDValue Op = Amt->getOperand(i);
+ if (Op->getOpcode() == ISD::UNDEF) {
+ Elts.push_back(Op);
+ continue;
+ }
+
+ ConstantSDNode *ND = cast<ConstantSDNode>(Op);
+ const APInt &C = APInt(SVTBits, ND->getAPIntValue().getZExtValue());
+ uint64_t ShAmt = C.getZExtValue();
+ if (ShAmt >= SVTBits) {
+ Elts.push_back(DAG.getUNDEF(SVT));
+ continue;
+ }
+ Elts.push_back(DAG.getConstant(One.shl(ShAmt), SVT));
+ }
+ SDValue BV = DAG.getNode(ISD::BUILD_VECTOR, dl, VT, &Elts[0], NumElems);
+ return DAG.getNode(ISD::MUL, dl, VT, R, BV);
+ }
+
// Lower SHL with variable shift amount.
if (VT == MVT::v4i32 && Op->getOpcode() == ISD::SHL) {
Op = DAG.getNode(ISD::SHL, dl, VT, Amt, DAG.getConstant(23, VT));
Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
return DAG.getNode(ISD::MUL, dl, VT, Op, R);
}
+
+ // If possible, lower this shift as a sequence of two shifts by
+ // constant plus a MOVSS/MOVSD instead of scalarizing it.
+ // Example:
+ // (v4i32 (srl A, (build_vector < X, Y, Y, Y>)))
+ //
+ // Could be rewritten as:
+ // (v4i32 (MOVSS (srl A, <Y,Y,Y,Y>), (srl A, <X,X,X,X>)))
+ //
+ // The advantage is that the two shifts from the example would be
+ // lowered as X86ISD::VSRLI nodes. This would be cheaper than scalarizing
+ // the vector shift into four scalar shifts plus four pairs of vector
+ // insert/extract.
+ if ((VT == MVT::v8i16 || VT == MVT::v4i32) &&
+ ISD::isBuildVectorOfConstantSDNodes(Amt.getNode())) {
+ unsigned TargetOpcode = X86ISD::MOVSS;
+ bool CanBeSimplified;
+ // The splat value for the first packed shift (the 'X' from the example).
+ SDValue Amt1 = Amt->getOperand(0);
+ // The splat value for the second packed shift (the 'Y' from the example).
+ SDValue Amt2 = (VT == MVT::v4i32) ? Amt->getOperand(1) :
+ Amt->getOperand(2);
+
+ // See if it is possible to replace this node with a sequence of
+ // two shifts followed by a MOVSS/MOVSD
+ if (VT == MVT::v4i32) {
+ // Check if it is legal to use a MOVSS.
+ CanBeSimplified = Amt2 == Amt->getOperand(2) &&
+ Amt2 == Amt->getOperand(3);
+ if (!CanBeSimplified) {
+ // Otherwise, check if we can still simplify this node using a MOVSD.
+ CanBeSimplified = Amt1 == Amt->getOperand(1) &&
+ Amt->getOperand(2) == Amt->getOperand(3);
+ TargetOpcode = X86ISD::MOVSD;
+ Amt2 = Amt->getOperand(2);
+ }
+ } else {
+ // Do similar checks for the case where the machine value type
+ // is MVT::v8i16.
+ CanBeSimplified = Amt1 == Amt->getOperand(1);
+ for (unsigned i=3; i != 8 && CanBeSimplified; ++i)
+ CanBeSimplified = Amt2 == Amt->getOperand(i);
+
+ if (!CanBeSimplified) {
+ TargetOpcode = X86ISD::MOVSD;
+ CanBeSimplified = true;
+ Amt2 = Amt->getOperand(4);
+ for (unsigned i=0; i != 4 && CanBeSimplified; ++i)
+ CanBeSimplified = Amt1 == Amt->getOperand(i);
+ for (unsigned j=4; j != 8 && CanBeSimplified; ++j)
+ CanBeSimplified = Amt2 == Amt->getOperand(j);
+ }
+ }
+
+ if (CanBeSimplified && isa<ConstantSDNode>(Amt1) &&
+ isa<ConstantSDNode>(Amt2)) {
+ // Replace this node with two shifts followed by a MOVSS/MOVSD.
+ EVT CastVT = MVT::v4i32;
+ SDValue Splat1 =
+ DAG.getConstant(cast<ConstantSDNode>(Amt1)->getAPIntValue(), VT);
+ SDValue Shift1 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat1);
+ SDValue Splat2 =
+ DAG.getConstant(cast<ConstantSDNode>(Amt2)->getAPIntValue(), VT);
+ SDValue Shift2 = DAG.getNode(Op->getOpcode(), dl, VT, R, Splat2);
+ if (TargetOpcode == X86ISD::MOVSD)
+ CastVT = MVT::v2i64;
+ SDValue BitCast1 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift1);
+ SDValue BitCast2 = DAG.getNode(ISD::BITCAST, dl, CastVT, Shift2);
+ SDValue Result = getTargetShuffleNode(TargetOpcode, dl, CastVT, BitCast2,
+ BitCast1, DAG);
+ return DAG.getNode(ISD::BITCAST, dl, VT, Result);
+ }
+ }
+
if (VT == MVT::v16i8 && Op->getOpcode() == ISD::SHL) {
assert(Subtarget->hasSSE2() && "Need SSE2 for pslli/pcmpeq.");
return R;
}
+ // It's worth extending once and using the v8i32 shifts for 16-bit types, but
+ // the extra overheads to get from v16i8 to v8i32 make the existing SSE
+ // solution better.
+ if (Subtarget->hasInt256() && VT == MVT::v8i16) {
+ MVT NewVT = VT == MVT::v8i16 ? MVT::v8i32 : MVT::v16i16;
+ unsigned ExtOpc =
+ Op.getOpcode() == ISD::SRA ? ISD::SIGN_EXTEND : ISD::ZERO_EXTEND;
+ R = DAG.getNode(ExtOpc, dl, NewVT, R);
+ Amt = DAG.getNode(ISD::ANY_EXTEND, dl, NewVT, Amt);
+ return DAG.getNode(ISD::TRUNCATE, dl, VT,
+ DAG.getNode(Op.getOpcode(), dl, NewVT, R, Amt));
+ }
+
// Decompose 256-bit shifts into smaller 128-bit shifts.
if (VT.is256BitVector()) {
unsigned NumElems = VT.getVectorNumElements();
return cpOut;
}
-static SDValue LowerREADCYCLECOUNTER(SDValue Op, const X86Subtarget *Subtarget,
- SelectionDAG &DAG) {
- assert(Subtarget->is64Bit() && "Result not type legalized?");
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
- SDValue TheChain = Op.getOperand(0);
- SDLoc dl(Op);
- SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
- SDValue rax = DAG.getCopyFromReg(rd, dl, X86::RAX, MVT::i64, rd.getValue(1));
- SDValue rdx = DAG.getCopyFromReg(rax.getValue(1), dl, X86::RDX, MVT::i64,
- rax.getValue(2));
- SDValue Tmp = DAG.getNode(ISD::SHL, dl, MVT::i64, rdx,
- DAG.getConstant(32, MVT::i8));
- SDValue Ops[] = {
- DAG.getNode(ISD::OR, dl, MVT::i64, rax, Tmp),
- rdx.getValue(1)
- };
- return DAG.getMergeValues(Ops, array_lengthof(Ops), dl);
-}
-
static SDValue LowerBITCAST(SDValue Op, const X86Subtarget *Subtarget,
SelectionDAG &DAG) {
MVT SrcVT = Op.getOperand(0).getSimpleValueType();
cast<AtomicSDNode>(Node)->getMemoryVT(),
Node->getOperand(0),
Node->getOperand(1), negOp,
- cast<AtomicSDNode>(Node)->getSrcValue(),
- cast<AtomicSDNode>(Node)->getAlignment(),
+ cast<AtomicSDNode>(Node)->getMemOperand(),
cast<AtomicSDNode>(Node)->getOrdering(),
cast<AtomicSDNode>(Node)->getSynchScope());
}
case ISD::CTLZ_ZERO_UNDEF: return LowerCTLZ_ZERO_UNDEF(Op, DAG);
case ISD::CTTZ: return LowerCTTZ(Op, DAG);
case ISD::MUL: return LowerMUL(Op, Subtarget, DAG);
+ case ISD::UMUL_LOHI: return LowerUMUL_LOHI(Op, Subtarget, DAG);
case ISD::SRA:
case ISD::SRL:
case ISD::SHL: return LowerShift(Op, Subtarget, DAG);
case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
case ISD::ADD: return LowerADD(Op, DAG);
case ISD::SUB: return LowerSUB(Op, DAG);
- case ISD::SDIV: return LowerSDIV(Op, DAG);
case ISD::FSINCOS: return LowerFSINCOS(Op, Subtarget, DAG);
}
}
Node->getOperand(1), Zero, Zero,
cast<AtomicSDNode>(Node)->getMemOperand(),
cast<AtomicSDNode>(Node)->getOrdering(),
+ cast<AtomicSDNode>(Node)->getOrdering(),
cast<AtomicSDNode>(Node)->getSynchScope());
Results.push_back(Swap.getValue(0));
Results.push_back(Swap.getValue(1));
std::pair<SDValue,SDValue> Vals =
FP_TO_INTHelper(SDValue(N, 0), DAG, IsSigned, /*IsReplace=*/ true);
SDValue FIST = Vals.first, StackSlot = Vals.second;
- if (FIST.getNode() != 0) {
+ if (FIST.getNode()) {
EVT VT = N->getValueType(0);
// Return a load from the stack slot.
- if (StackSlot.getNode() != 0)
+ if (StackSlot.getNode())
Results.push_back(DAG.getLoad(VT, dl, FIST, StackSlot,
MachinePointerInfo(),
false, false, false, 0));
Results.push_back(V);
return;
}
+ case ISD::INTRINSIC_W_CHAIN: {
+ unsigned IntNo = cast<ConstantSDNode>(N->getOperand(1))->getZExtValue();
+ switch (IntNo) {
+ default : llvm_unreachable("Do not know how to custom type "
+ "legalize this intrinsic operation!");
+ case Intrinsic::x86_rdtsc:
+ return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
+ Results);
+ case Intrinsic::x86_rdtscp:
+ return getReadTimeStampCounter(N, dl, X86ISD::RDTSCP_DAG, DAG, Subtarget,
+ Results);
+ }
+ }
case ISD::READCYCLECOUNTER: {
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
- SDValue TheChain = N->getOperand(0);
- SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
- SDValue eax = DAG.getCopyFromReg(rd, dl, X86::EAX, MVT::i32,
- rd.getValue(1));
- SDValue edx = DAG.getCopyFromReg(eax.getValue(1), dl, X86::EDX, MVT::i32,
- eax.getValue(2));
- // Use a buildpair to merge the two 32-bit values into a 64-bit one.
- SDValue Ops[] = { eax, edx };
- Results.push_back(DAG.getNode(ISD::BUILD_PAIR, dl, MVT::i64, Ops,
- array_lengthof(Ops)));
- Results.push_back(edx.getValue(1));
- return;
+ return getReadTimeStampCounter(N, dl, X86ISD::RDTSC_DAG, DAG, Subtarget,
+ Results);
}
case ISD::ATOMIC_CMP_SWAP: {
EVT T = N->getValueType(0);
const char *X86TargetLowering::getTargetNodeName(unsigned Opcode) const {
switch (Opcode) {
- default: return NULL;
+ default: return nullptr;
case X86ISD::BSF: return "X86ISD::BSF";
case X86ISD::BSR: return "X86ISD::BSR";
case X86ISD::SHLD: return "X86ISD::SHLD";
case X86ISD::ATOMAND64_DAG: return "X86ISD::ATOMAND64_DAG";
case X86ISD::ATOMNAND64_DAG: return "X86ISD::ATOMNAND64_DAG";
case X86ISD::VZEXT_MOVL: return "X86ISD::VZEXT_MOVL";
- case X86ISD::VSEXT_MOVL: return "X86ISD::VSEXT_MOVL";
case X86ISD::VZEXT_LOAD: return "X86ISD::VZEXT_LOAD";
case X86ISD::VZEXT: return "X86ISD::VZEXT";
case X86ISD::VSEXT: return "X86ISD::VSEXT";
case X86ISD::OR: return "X86ISD::OR";
case X86ISD::XOR: return "X86ISD::XOR";
case X86ISD::AND: return "X86ISD::AND";
- case X86ISD::BLSI: return "X86ISD::BLSI";
- case X86ISD::BLSMSK: return "X86ISD::BLSMSK";
- case X86ISD::BLSR: return "X86ISD::BLSR";
- case X86ISD::BZHI: return "X86ISD::BZHI";
case X86ISD::BEXTR: return "X86ISD::BEXTR";
case X86ISD::MUL_IMM: return "X86ISD::MUL_IMM";
case X86ISD::PTEST: return "X86ISD::PTEST";
case X86ISD::TESTP: return "X86ISD::TESTP";
case X86ISD::TESTM: return "X86ISD::TESTM";
+ case X86ISD::TESTNM: return "X86ISD::TESTNM";
case X86ISD::KORTEST: return "X86ISD::KORTEST";
case X86ISD::PALIGNR: return "X86ISD::PALIGNR";
case X86ISD::PSHUFD: return "X86ISD::PSHUFD";
case X86ISD::UNPCKH: return "X86ISD::UNPCKH";
case X86ISD::VBROADCAST: return "X86ISD::VBROADCAST";
case X86ISD::VBROADCASTM: return "X86ISD::VBROADCASTM";
+ case X86ISD::VEXTRACT: return "X86ISD::VEXTRACT";
case X86ISD::VPERMILP: return "X86ISD::VPERMILP";
case X86ISD::VPERM2X128: return "X86ISD::VPERM2X128";
case X86ISD::VPERMV: return "X86ISD::VPERMV";
case X86ISD::VPERMV3: return "X86ISD::VPERMV3";
+ case X86ISD::VPERMIV3: return "X86ISD::VPERMIV3";
case X86ISD::VPERMI: return "X86ISD::VPERMI";
case X86ISD::PMULUDQ: return "X86ISD::PMULUDQ";
case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
Reloc::Model R = getTargetMachine().getRelocationModel();
// X86 allows a sign-extended 32-bit immediate field as a displacement.
- if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != NULL))
+ if (!X86::isOffsetSuitableForCodeModel(AM.BaseOffs, M, AM.BaseGV != nullptr))
return false;
if (AM.BaseGV) {
return true;
}
+bool X86TargetLowering::isVectorShiftByScalarCheap(Type *Ty) const {
+ unsigned Bits = Ty->getScalarSizeInBits();
+
+ // 8-bit shifts are always expensive, but versions with a scalar amount aren't
+ // particularly cheaper than those without.
+ if (Bits == 8)
+ return false;
+
+ // On AVX2 there are new vpsllv[dq] instructions (and other shifts), that make
+ // variable shifts just as cheap as scalar ones.
+ if (Subtarget->hasInt256() && (Bits == 32 || Bits == 64))
+ return false;
+
+ // Otherwise, it's significantly cheaper to shift by a scalar amount than by a
+ // fully general vector.
+ return true;
+}
+
bool X86TargetLowering::isTruncateFree(Type *Ty1, Type *Ty2) const {
if (!Ty1->isIntegerTy() || !Ty2->isIntegerTy())
return false;
// Transfer the remainder of BB and its successor edges to sinkMBB.
sinkMBB->splice(sinkMBB->begin(), MBB,
- llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
+ std::next(MachineBasicBlock::iterator(MI)), MBB->end());
sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
// thisMBB:
// Transfer the remainder of BB and its successor edges to sinkMBB.
sinkMBB->splice(sinkMBB->begin(), MBB,
- llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
+ std::next(MachineBasicBlock::iterator(MI)), MBB->end());
sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
// thisMBB:
// Transfer the remainder of BB and its successor edges to sinkMBB.
sinkMBB->splice(sinkMBB->begin(), MBB,
- llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
+ std::next(MachineBasicBlock::iterator(MI)), MBB->end());
sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
// thisMBB:
OffsetDestReg = 0; // unused
OverflowDestReg = DestReg;
- offsetMBB = NULL;
+ offsetMBB = nullptr;
overflowMBB = thisMBB;
endMBB = thisMBB;
} else {
// Transfer the remainder of MBB and its successor edges to endMBB.
endMBB->splice(endMBB->begin(), thisMBB,
- llvm::next(MachineBasicBlock::iterator(MI)),
- thisMBB->end());
+ std::next(MachineBasicBlock::iterator(MI)), thisMBB->end());
endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
// Make offsetMBB and overflowMBB successors of thisMBB
// Transfer the remainder of MBB and its successor edges to EndMBB.
EndMBB->splice(EndMBB->begin(), MBB,
- llvm::next(MachineBasicBlock::iterator(MI)),
- MBB->end());
+ std::next(MachineBasicBlock::iterator(MI)), MBB->end());
EndMBB->transferSuccessorsAndUpdatePHIs(MBB);
// The original block will now fall through to the XMM save block.
MachineBasicBlock* BB,
const TargetRegisterInfo* TRI) {
// Scan forward through BB for a use/def of EFLAGS.
- MachineBasicBlock::iterator miI(llvm::next(SelectItr));
+ MachineBasicBlock::iterator miI(std::next(SelectItr));
for (MachineBasicBlock::iterator miE = BB->end(); miI != miE; ++miI) {
const MachineInstr& mi = *miI;
if (mi.readsRegister(X86::EFLAGS))
// Transfer the remainder of BB and its successor edges to sinkMBB.
sinkMBB->splice(sinkMBB->begin(), BB,
- llvm::next(MachineBasicBlock::iterator(MI)),
- BB->end());
+ std::next(MachineBasicBlock::iterator(MI)), BB->end());
sinkMBB->transferSuccessorsAndUpdatePHIs(BB);
// Add the true and fallthrough blocks as its successors.
MachineFunction *MF = BB->getParent();
const BasicBlock *LLVM_BB = BB->getBasicBlock();
- assert(getTargetMachine().Options.EnableSegmentedStacks);
+ assert(MF->shouldSplitStack());
unsigned TlsReg = Is64Bit ? X86::FS : X86::GS;
unsigned TlsOffset = Is64Bit ? 0x70 : 0x30;
MF->insert(MBBIter, mallocMBB);
MF->insert(MBBIter, continueMBB);
- continueMBB->splice(continueMBB->begin(), BB, llvm::next
- (MachineBasicBlock::iterator(MI)), BB->end());
+ continueMBB->splice(continueMBB->begin(), BB,
+ std::next(MachineBasicBlock::iterator(MI)), BB->end());
continueMBB->transferSuccessorsAndUpdatePHIs(BB);
// Add code to the main basic block to check if the stack limit has been hit,
}
} else {
const char *StackProbeSymbol =
- Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
+ Subtarget->isTargetKnownWindowsMSVC() ? "_chkstk" : "_alloca";
BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
.addExternalSymbol(StackProbeSymbol)
// Transfer the remainder of BB and its successor edges to sinkMBB.
sinkMBB->splice(sinkMBB->begin(), MBB,
- llvm::next(MachineBasicBlock::iterator(MI)), MBB->end());
+ std::next(MachineBasicBlock::iterator(MI)), MBB->end());
sinkMBB->transferSuccessorsAndUpdatePHIs(MBB);
// thisMBB:
return MBB;
}
+// Replace 213-type (isel default) FMA3 instructions with 231-type for
+// accumulator loops. Writing back to the accumulator allows the coalescer
+// to remove extra copies in the loop.
+MachineBasicBlock *
+X86TargetLowering::emitFMA3Instr(MachineInstr *MI,
+ MachineBasicBlock *MBB) const {
+ MachineOperand &AddendOp = MI->getOperand(3);
+
+ // Bail out early if the addend isn't a register - we can't switch these.
+ if (!AddendOp.isReg())
+ return MBB;
+
+ MachineFunction &MF = *MBB->getParent();
+ MachineRegisterInfo &MRI = MF.getRegInfo();
+
+ // Check whether the addend is defined by a PHI:
+ assert(MRI.hasOneDef(AddendOp.getReg()) && "Multiple defs in SSA?");
+ MachineInstr &AddendDef = *MRI.def_instr_begin(AddendOp.getReg());
+ if (!AddendDef.isPHI())
+ return MBB;
+
+ // Look for the following pattern:
+ // loop:
+ // %addend = phi [%entry, 0], [%loop, %result]
+ // ...
+ // %result<tied1> = FMA213 %m2<tied0>, %m1, %addend
+
+ // Replace with:
+ // loop:
+ // %addend = phi [%entry, 0], [%loop, %result]
+ // ...
+ // %result<tied1> = FMA231 %addend<tied0>, %m1, %m2
+
+ for (unsigned i = 1, e = AddendDef.getNumOperands(); i < e; i += 2) {
+ assert(AddendDef.getOperand(i).isReg());
+ MachineOperand PHISrcOp = AddendDef.getOperand(i);
+ MachineInstr &PHISrcInst = *MRI.def_instr_begin(PHISrcOp.getReg());
+ if (&PHISrcInst == MI) {
+ // Found a matching instruction.
+ unsigned NewFMAOpc = 0;
+ switch (MI->getOpcode()) {
+ case X86::VFMADDPDr213r: NewFMAOpc = X86::VFMADDPDr231r; break;
+ case X86::VFMADDPSr213r: NewFMAOpc = X86::VFMADDPSr231r; break;
+ case X86::VFMADDSDr213r: NewFMAOpc = X86::VFMADDSDr231r; break;
+ case X86::VFMADDSSr213r: NewFMAOpc = X86::VFMADDSSr231r; break;
+ case X86::VFMSUBPDr213r: NewFMAOpc = X86::VFMSUBPDr231r; break;
+ case X86::VFMSUBPSr213r: NewFMAOpc = X86::VFMSUBPSr231r; break;
+ case X86::VFMSUBSDr213r: NewFMAOpc = X86::VFMSUBSDr231r; break;
+ case X86::VFMSUBSSr213r: NewFMAOpc = X86::VFMSUBSSr231r; break;
+ case X86::VFNMADDPDr213r: NewFMAOpc = X86::VFNMADDPDr231r; break;
+ case X86::VFNMADDPSr213r: NewFMAOpc = X86::VFNMADDPSr231r; break;
+ case X86::VFNMADDSDr213r: NewFMAOpc = X86::VFNMADDSDr231r; break;
+ case X86::VFNMADDSSr213r: NewFMAOpc = X86::VFNMADDSSr231r; break;
+ case X86::VFNMSUBPDr213r: NewFMAOpc = X86::VFNMSUBPDr231r; break;
+ case X86::VFNMSUBPSr213r: NewFMAOpc = X86::VFNMSUBPSr231r; break;
+ case X86::VFNMSUBSDr213r: NewFMAOpc = X86::VFNMSUBSDr231r; break;
+ case X86::VFNMSUBSSr213r: NewFMAOpc = X86::VFNMSUBSSr231r; break;
+ case X86::VFMADDPDr213rY: NewFMAOpc = X86::VFMADDPDr231rY; break;
+ case X86::VFMADDPSr213rY: NewFMAOpc = X86::VFMADDPSr231rY; break;
+ case X86::VFMSUBPDr213rY: NewFMAOpc = X86::VFMSUBPDr231rY; break;
+ case X86::VFMSUBPSr213rY: NewFMAOpc = X86::VFMSUBPSr231rY; break;
+ case X86::VFNMADDPDr213rY: NewFMAOpc = X86::VFNMADDPDr231rY; break;
+ case X86::VFNMADDPSr213rY: NewFMAOpc = X86::VFNMADDPSr231rY; break;
+ case X86::VFNMSUBPDr213rY: NewFMAOpc = X86::VFNMSUBPDr231rY; break;
+ case X86::VFNMSUBPSr213rY: NewFMAOpc = X86::VFNMSUBPSr231rY; break;
+ default: llvm_unreachable("Unrecognized FMA variant.");
+ }
+
+ const TargetInstrInfo &TII = *MF.getTarget().getInstrInfo();
+ MachineInstrBuilder MIB =
+ BuildMI(MF, MI->getDebugLoc(), TII.get(NewFMAOpc))
+ .addOperand(MI->getOperand(0))
+ .addOperand(MI->getOperand(3))
+ .addOperand(MI->getOperand(2))
+ .addOperand(MI->getOperand(1));
+ MBB->insert(MachineBasicBlock::iterator(MI), MIB);
+ MI->eraseFromParent();
+ }
+ }
+
+ return MBB;
+}
+
MachineBasicBlock *
X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
MachineBasicBlock *BB) const {
case TargetOpcode::STACKMAP:
case TargetOpcode::PATCHPOINT:
return emitPatchPoint(MI, BB);
+
+ case X86::VFMADDPDr213r:
+ case X86::VFMADDPSr213r:
+ case X86::VFMADDSDr213r:
+ case X86::VFMADDSSr213r:
+ case X86::VFMSUBPDr213r:
+ case X86::VFMSUBPSr213r:
+ case X86::VFMSUBSDr213r:
+ case X86::VFMSUBSSr213r:
+ case X86::VFNMADDPDr213r:
+ case X86::VFNMADDPSr213r:
+ case X86::VFNMADDSDr213r:
+ case X86::VFNMADDSSr213r:
+ case X86::VFNMSUBPDr213r:
+ case X86::VFNMSUBPSr213r:
+ case X86::VFNMSUBSDr213r:
+ case X86::VFNMSUBSSr213r:
+ case X86::VFMADDPDr213rY:
+ case X86::VFMADDPSr213rY:
+ case X86::VFMSUBPDr213rY:
+ case X86::VFMSUBPSr213rY:
+ case X86::VFNMADDPDr213rY:
+ case X86::VFNMADDPSr213rY:
+ case X86::VFNMSUBPDr213rY:
+ case X86::VFNMSUBPSr213rY:
+ return emitFMA3Instr(MI, BB);
}
}
}
}
-unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(SDValue Op,
- unsigned Depth) const {
+unsigned X86TargetLowering::ComputeNumSignBitsForTargetNode(
+ SDValue Op,
+ const SelectionDAG &,
+ unsigned Depth) const {
// SETCC_CARRY sets the dest to ~0 for true or 0 for false.
if (Op.getOpcode() == X86ISD::SETCC_CARRY)
return Op.getValueType().getScalarType().getSizeInBits();
}
}
+ // Try to fold this VSELECT into a MOVSS/MOVSD
+ if (N->getOpcode() == ISD::VSELECT &&
+ Cond.getOpcode() == ISD::BUILD_VECTOR && !DCI.isBeforeLegalize()) {
+ if (VT == MVT::v4i32 || VT == MVT::v4f32 ||
+ (Subtarget->hasSSE2() && (VT == MVT::v2i64 || VT == MVT::v2f64))) {
+ bool CanFold = false;
+ unsigned NumElems = Cond.getNumOperands();
+ SDValue A = LHS;
+ SDValue B = RHS;
+
+ if (isZero(Cond.getOperand(0))) {
+ CanFold = true;
+
+ // fold (vselect <0,-1,-1,-1>, A, B) -> (movss A, B)
+ // fold (vselect <0,-1> -> (movsd A, B)
+ for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
+ CanFold = isAllOnes(Cond.getOperand(i));
+ } else if (isAllOnes(Cond.getOperand(0))) {
+ CanFold = true;
+ std::swap(A, B);
+
+ // fold (vselect <-1,0,0,0>, A, B) -> (movss B, A)
+ // fold (vselect <-1,0> -> (movsd B, A)
+ for (unsigned i = 1, e = NumElems; i != e && CanFold; ++i)
+ CanFold = isZero(Cond.getOperand(i));
+ }
+
+ if (CanFold) {
+ if (VT == MVT::v4i32 || VT == MVT::v4f32)
+ return getTargetShuffleNode(X86ISD::MOVSS, DL, VT, A, B, DAG);
+ return getTargetShuffleNode(X86ISD::MOVSD, DL, VT, A, B, DAG);
+ }
+
+ if (Subtarget->hasSSE2() && (VT == MVT::v4i32 || VT == MVT::v4f32)) {
+ // fold (v4i32: vselect <0,0,-1,-1>, A, B) ->
+ // (v4i32 (bitcast (movsd (v2i64 (bitcast A)),
+ // (v2i64 (bitcast B)))))
+ //
+ // fold (v4f32: vselect <0,0,-1,-1>, A, B) ->
+ // (v4f32 (bitcast (movsd (v2f64 (bitcast A)),
+ // (v2f64 (bitcast B)))))
+ //
+ // fold (v4i32: vselect <-1,-1,0,0>, A, B) ->
+ // (v4i32 (bitcast (movsd (v2i64 (bitcast B)),
+ // (v2i64 (bitcast A)))))
+ //
+ // fold (v4f32: vselect <-1,-1,0,0>, A, B) ->
+ // (v4f32 (bitcast (movsd (v2f64 (bitcast B)),
+ // (v2f64 (bitcast A)))))
+
+ CanFold = (isZero(Cond.getOperand(0)) &&
+ isZero(Cond.getOperand(1)) &&
+ isAllOnes(Cond.getOperand(2)) &&
+ isAllOnes(Cond.getOperand(3)));
+
+ if (!CanFold && isAllOnes(Cond.getOperand(0)) &&
+ isAllOnes(Cond.getOperand(1)) &&
+ isZero(Cond.getOperand(2)) &&
+ isZero(Cond.getOperand(3))) {
+ CanFold = true;
+ std::swap(LHS, RHS);
+ }
+
+ if (CanFold) {
+ EVT NVT = (VT == MVT::v4i32) ? MVT::v2i64 : MVT::v2f64;
+ SDValue NewA = DAG.getNode(ISD::BITCAST, DL, NVT, LHS);
+ SDValue NewB = DAG.getNode(ISD::BITCAST, DL, NVT, RHS);
+ SDValue Select = getTargetShuffleNode(X86ISD::MOVSD, DL, NVT, NewA,
+ NewB, DAG);
+ return DAG.getNode(ISD::BITCAST, DL, VT, Select);
+ }
+ }
+ }
+ }
+
// 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
SDValue Op2 = Cmp.getOperand(1);
SDValue SetCC;
- const ConstantSDNode* C = 0;
+ const ConstantSDNode* C = nullptr;
bool needOppositeCond = (CC == X86::COND_E);
bool checkAgainstTrue = false; // Is it a comparison against 1?
// the DCI.xxxx conditions are provided to postpone the optimization as
// late as possible.
- ConstantSDNode *CmpAgainst = 0;
+ ConstantSDNode *CmpAgainst = nullptr;
if ((Cond.getOpcode() == X86ISD::CMP || Cond.getOpcode() == X86ISD::SUB) &&
(CmpAgainst = dyn_cast<ConstantSDNode>(Cond.getOperand(1))) &&
!isa<ConstantSDNode>(Cond.getOperand(0))) {
if ((cc0 == X86::COND_E && cc1 == X86::COND_NP) ||
(cc0 == X86::COND_NE && cc1 == X86::COND_P)) {
- bool is64BitFP = (CMP00.getValueType() == MVT::f64);
// FIXME: need symbolic constants for these magic numbers.
// See X86ATTInstPrinter.cpp:printSSECC().
unsigned x86cc = (cc0 == X86::COND_E) ? 0 : 4;
if (Subtarget->hasAVX512()) {
- // SETCC type in AVX-512 is MVT::i1
- assert(N->getValueType(0) == MVT::i1 && "Unexpected AND node type");
- return DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00, CMP01,
- DAG.getConstant(x86cc, MVT::i8));
+ SDValue FSetCC = DAG.getNode(X86ISD::FSETCC, DL, MVT::i1, CMP00,
+ CMP01, DAG.getConstant(x86cc, MVT::i8));
+ if (N->getValueType(0) != MVT::i1)
+ return DAG.getNode(ISD::ZERO_EXTEND, DL, N->getValueType(0),
+ FSetCC);
+ return FSetCC;
}
- SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL, CMP00.getValueType(), CMP00, CMP01,
+ SDValue OnesOrZeroesF = DAG.getNode(X86ISD::FSETCC, DL,
+ CMP00.getValueType(), CMP00, CMP01,
DAG.getConstant(x86cc, MVT::i8));
- MVT IntVT = (is64BitFP ? MVT::i64 : MVT::i32);
- SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT,
- OnesOrZeroesF);
+
+ bool is64BitFP = (CMP00.getValueType() == MVT::f64);
+ MVT IntVT = is64BitFP ? MVT::i64 : MVT::i32;
+
+ if (is64BitFP && !Subtarget->is64Bit()) {
+ // On a 32-bit target, we cannot bitcast the 64-bit float to a
+ // 64-bit integer, since that's not a legal type. Since
+ // OnesOrZeroesF is all ones of all zeroes, we don't need all the
+ // bits, but can do this little dance to extract the lowest 32 bits
+ // and work with those going forward.
+ SDValue Vector64 = DAG.getNode(ISD::SCALAR_TO_VECTOR, DL, MVT::v2f64,
+ OnesOrZeroesF);
+ SDValue Vector32 = DAG.getNode(ISD::BITCAST, DL, MVT::v4f32,
+ Vector64);
+ OnesOrZeroesF = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, DL, MVT::f32,
+ Vector32, DAG.getIntPtrConstant(0));
+ IntVT = MVT::i32;
+ }
+
+ SDValue OnesOrZeroesI = DAG.getNode(ISD::BITCAST, DL, IntVT, OnesOrZeroesF);
SDValue ANDed = DAG.getNode(ISD::AND, DL, IntVT, OnesOrZeroesI,
DAG.getConstant(1, IntVT));
SDValue OneBitOfTruth = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, ANDed);
if (R.getNode())
return R;
- // Create BLSI, BLSR, and BZHI instructions
- // BLSI is X & (-X)
- // BLSR is X & (X-1)
- // BZHI is X & ((1 << Y) - 1)
+ // Create BEXTR instructions
// BEXTR is ((X >> imm) & (2**size-1))
if (VT == MVT::i32 || VT == MVT::i64) {
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
SDLoc DL(N);
- if (Subtarget->hasBMI()) {
- // Check LHS for neg
- if (N0.getOpcode() == ISD::SUB && N0.getOperand(1) == N1 &&
- isZero(N0.getOperand(0)))
- return DAG.getNode(X86ISD::BLSI, DL, VT, N1);
-
- // Check RHS for neg
- if (N1.getOpcode() == ISD::SUB && N1.getOperand(1) == N0 &&
- isZero(N1.getOperand(0)))
- return DAG.getNode(X86ISD::BLSI, DL, VT, N0);
-
- // Check LHS for X-1
- if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
- isAllOnes(N0.getOperand(1)))
- return DAG.getNode(X86ISD::BLSR, DL, VT, N1);
-
- // Check RHS for X-1
- if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
- isAllOnes(N1.getOperand(1)))
- return DAG.getNode(X86ISD::BLSR, DL, VT, N0);
- }
-
- if (Subtarget->hasBMI2()) {
- // Check for (and (add (shl 1, Y), -1), X)
- if (N0.getOpcode() == ISD::ADD && isAllOnes(N0.getOperand(1))) {
- SDValue N00 = N0.getOperand(0);
- if (N00.getOpcode() == ISD::SHL) {
- SDValue N001 = N00.getOperand(1);
- assert(N001.getValueType() == MVT::i8 && "unexpected type");
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(N00.getOperand(0));
- if (C && C->getZExtValue() == 1)
- return DAG.getNode(X86ISD::BZHI, DL, VT, N1, N001);
- }
- }
-
- // Check for (and X, (add (shl 1, Y), -1))
- if (N1.getOpcode() == ISD::ADD && isAllOnes(N1.getOperand(1))) {
- SDValue N10 = N1.getOperand(0);
- if (N10.getOpcode() == ISD::SHL) {
- SDValue N101 = N10.getOperand(1);
- assert(N101.getValueType() == MVT::i8 && "unexpected type");
- ConstantSDNode *C = dyn_cast<ConstantSDNode>(N10.getOperand(0));
- if (C && C->getZExtValue() == 1)
- return DAG.getNode(X86ISD::BZHI, DL, VT, N0, N101);
- }
- }
- }
-
// Check for BEXTR.
if ((Subtarget->hasBMI() || Subtarget->hasTBM()) &&
(N0.getOpcode() == ISD::SRA || N0.getOpcode() == ISD::SRL)) {
static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
- EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
+ EVT VT = N->getValueType(0);
// look for psign/blend
if (VT == MVT::v2i64 || VT == MVT::v4i64) {
static SDValue PerformXorCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
- EVT VT = N->getValueType(0);
if (DCI.isBeforeLegalizeOps())
return SDValue();
return RV;
}
- // Try forming BMI if it is available.
- if (!Subtarget->hasBMI())
- return SDValue();
-
- if (VT != MVT::i32 && VT != MVT::i64)
- return SDValue();
-
- assert(Subtarget->hasBMI() && "Creating BLSMSK requires BMI instructions");
-
- // Create BLSMSK instructions by finding X ^ (X-1)
- SDValue N0 = N->getOperand(0);
- SDValue N1 = N->getOperand(1);
- SDLoc DL(N);
-
- if (N0.getOpcode() == ISD::ADD && N0.getOperand(0) == N1 &&
- isAllOnes(N0.getOperand(1)))
- return DAG.getNode(X86ISD::BLSMSK, DL, VT, N1);
-
- if (N1.getOpcode() == ISD::ADD && N1.getOperand(0) == N0 &&
- isAllOnes(N1.getOperand(1)))
- return DAG.getNode(X86ISD::BLSMSK, DL, VT, N0);
-
return SDValue();
}
!cast<LoadSDNode>(St->getValue())->isVolatile() &&
St->getChain().hasOneUse() && !St->isVolatile()) {
SDNode* LdVal = St->getValue().getNode();
- LoadSDNode *Ld = 0;
+ LoadSDNode *Ld = nullptr;
int TokenFactorIndex = -1;
SmallVector<SDValue, 8> Ops;
SDNode* ChainVal = St->getChain().getNode();
// Optimize x == -y --> x+y == 0
// x != -y --> x+y != 0
-static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG) {
+static SDValue PerformISDSETCCCombine(SDNode *N, SelectionDAG &DAG,
+ const X86Subtarget* Subtarget) {
ISD::CondCode CC = cast<CondCodeSDNode>(N->getOperand(2))->get();
SDValue LHS = N->getOperand(0);
SDValue RHS = N->getOperand(1);
+ EVT VT = N->getValueType(0);
+ SDLoc DL(N);
if ((CC == ISD::SETNE || CC == ISD::SETEQ) && LHS.getOpcode() == ISD::SUB)
if (ConstantSDNode *C = dyn_cast<ConstantSDNode>(LHS.getOperand(0)))
return DAG.getSetCC(SDLoc(N), N->getValueType(0),
addV, DAG.getConstant(0, addV.getValueType()), CC);
}
+
+ if (VT.getScalarType() == MVT::i1) {
+ bool IsSEXT0 = (LHS.getOpcode() == ISD::SIGN_EXTEND) &&
+ (LHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
+ bool IsVZero0 = ISD::isBuildVectorAllZeros(LHS.getNode());
+ if (!IsSEXT0 && !IsVZero0)
+ return SDValue();
+ bool IsSEXT1 = (RHS.getOpcode() == ISD::SIGN_EXTEND) &&
+ (RHS.getOperand(0).getValueType().getScalarType() == MVT::i1);
+ bool IsVZero1 = ISD::isBuildVectorAllZeros(RHS.getNode());
+
+ if (!IsSEXT1 && !IsVZero1)
+ return SDValue();
+
+ if (IsSEXT0 && IsVZero1) {
+ assert(VT == LHS.getOperand(0).getValueType() && "Uexpected operand type");
+ if (CC == ISD::SETEQ)
+ return DAG.getNOT(DL, LHS.getOperand(0), VT);
+ return LHS.getOperand(0);
+ }
+ if (IsSEXT1 && IsVZero0) {
+ assert(VT == RHS.getOperand(0).getValueType() && "Uexpected operand type");
+ if (CC == ISD::SETEQ)
+ return DAG.getNOT(DL, RHS.getOperand(0), VT);
+ return RHS.getOperand(0);
+ }
+ }
+
return SDValue();
}
case ISD::SIGN_EXTEND: return PerformSExtCombine(N, DAG, DCI, Subtarget);
case ISD::SIGN_EXTEND_INREG: return PerformSIGN_EXTEND_INREGCombine(N, DAG, Subtarget);
case ISD::TRUNCATE: return PerformTruncateCombine(N, DAG,DCI,Subtarget);
- case ISD::SETCC: return PerformISDSETCCCombine(N, DAG);
+ case ISD::SETCC: return PerformISDSETCCCombine(N, DAG, Subtarget);
case X86ISD::SETCC: return PerformSETCCCombine(N, DAG, DCI, Subtarget);
case X86ISD::BRCOND: return PerformBrCondCombine(N, DAG, DCI, Subtarget);
case X86ISD::VZEXT: return performVZEXTCombine(N, DAG, DCI, Subtarget);
Value *CallOperandVal = info.CallOperandVal;
// If we don't have a value, we can't do a match,
// but allow it at the lowest weight.
- if (CallOperandVal == NULL)
+ if (!CallOperandVal)
return CW_Default;
Type *type = CallOperandVal->getType();
// Look at the constraint type.
std::string &Constraint,
std::vector<SDValue>&Ops,
SelectionDAG &DAG) const {
- SDValue Result(0, 0);
+ SDValue Result;
// Only support length 1 constraints for now.
if (Constraint.length() > 1) return;
// If we are in non-pic codegen mode, we allow the address of a global (with
// an optional displacement) to be used with 'i'.
- GlobalAddressSDNode *GA = 0;
+ GlobalAddressSDNode *GA = nullptr;
int64_t Offset = 0;
// Match either (GA), (GA+C), (GA+C1+C2), etc.
Res = TargetLowering::getRegForInlineAsmConstraint(Constraint, VT);
// Not found as a standard register?
- if (Res.second == 0) {
+ if (!Res.second) {
// Map st(0) -> st(7) -> ST0
if (Constraint.size() == 7 && Constraint[0] == '{' &&
tolower(Constraint[1]) == 's' &&
return Res;
}
+
+int X86TargetLowering::getScalingFactorCost(const AddrMode &AM,
+ Type *Ty) const {
+ // Scaling factors are not free at all.
+ // An indexed folded instruction, i.e., inst (reg1, reg2, scale),
+ // will take 2 allocations instead of 1 for plain addressing mode,
+ // i.e. inst (reg1).
+ if (isLegalAddressingMode(AM, Ty))
+ // Scale represents reg2 * scale, thus account for 1
+ // as soon as we use a second register.
+ return AM.Scale != 0;
+ return -1;
+}
projects / oota-llvm.git / blobdiff