STATISTIC(NumTailCalls, "Number of tail calls");
-static cl::opt<bool>
-DisableMMX("disable-mmx", cl::Hidden, cl::desc("Disable use of MMX"));
-
// Forward declarations.
static SDValue getMOVL(SelectionDAG &DAG, DebugLoc dl, EVT VT, SDValue V1,
SDValue V2);
static TargetLoweringObjectFile *createTLOF(X86TargetMachine &TM) {
-
+
bool is64Bit = TM.getSubtarget<X86Subtarget>().is64Bit();
-
+
if (TM.getSubtarget<X86Subtarget>().isTargetDarwin()) {
- if (is64Bit) return new X8664_MachoTargetObjectFile();
+ if (is64Bit)
+ return new X8664_MachoTargetObjectFile();
return new TargetLoweringObjectFileMachO();
- } else if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
- if (is64Bit) return new X8664_ELFTargetObjectFile(TM);
+ }
+
+ if (TM.getSubtarget<X86Subtarget>().isTargetELF() ){
+ if (is64Bit)
+ return new X8664_ELFTargetObjectFile(TM);
return new X8632_ELFTargetObjectFile(TM);
- } else if (TM.getSubtarget<X86Subtarget>().isTargetCOFF()) {
+ }
+ if (TM.getSubtarget<X86Subtarget>().isTargetCOFF())
return new TargetLoweringObjectFileCOFF();
- }
llvm_unreachable("unknown subtarget type");
}
X86TargetLowering::X86TargetLowering(X86TargetMachine &TM)
: TargetLowering(TM, createTLOF(TM)) {
Subtarget = &TM.getSubtarget<X86Subtarget>();
- X86ScalarSSEf64 = Subtarget->hasSSE2();
- X86ScalarSSEf32 = Subtarget->hasSSE1();
+ X86ScalarSSEf64 = Subtarget->hasXMMInt();
+ X86ScalarSSEf32 = Subtarget->hasXMM();
X86StackPtr = Subtarget->is64Bit() ? X86::RSP : X86::ESP;
RegInfo = TM.getRegisterInfo();
TD = getTargetData();
// Set up the TargetLowering object.
+ static MVT IntVTs[] = { MVT::i8, MVT::i16, MVT::i32, MVT::i64 };
// X86 is weird, it always uses i8 for shift amounts and setcc results.
setShiftAmountType(MVT::i8);
setSchedulingPreference(Sched::RegPressure);
setStackPointerRegisterToSaveRestore(X86StackPtr);
+ if (Subtarget->isTargetWindows() && !Subtarget->isTargetCygMing()) {
+ // Setup Windows compiler runtime calls.
+ setLibcallName(RTLIB::SDIV_I64, "_alldiv");
+ setLibcallName(RTLIB::UDIV_I64, "_aulldiv");
+ setLibcallName(RTLIB::FPTOUINT_F64_I64, "_ftol2");
+ setLibcallName(RTLIB::FPTOUINT_F32_I64, "_ftol2");
+ setLibcallCallingConv(RTLIB::SDIV_I64, CallingConv::X86_StdCall);
+ setLibcallCallingConv(RTLIB::UDIV_I64, CallingConv::X86_StdCall);
+ setLibcallCallingConv(RTLIB::FPTOUINT_F64_I64, CallingConv::C);
+ setLibcallCallingConv(RTLIB::FPTOUINT_F32_I64, CallingConv::C);
+ }
+
if (Subtarget->isTargetDarwin()) {
// Darwin should use _setjmp/_longjmp instead of setjmp/longjmp.
setUseUnderscoreSetJmp(false);
}
// TODO: when we have SSE, these could be more efficient, by using movd/movq.
- if (!X86ScalarSSEf64) {
- setOperationAction(ISD::BIT_CONVERT , MVT::f32 , Expand);
- setOperationAction(ISD::BIT_CONVERT , MVT::i32 , Expand);
+ if (!X86ScalarSSEf64) {
+ setOperationAction(ISD::BITCAST , MVT::f32 , Expand);
+ setOperationAction(ISD::BITCAST , MVT::i32 , Expand);
if (Subtarget->is64Bit()) {
- setOperationAction(ISD::BIT_CONVERT , MVT::f64 , Expand);
+ setOperationAction(ISD::BITCAST , MVT::f64 , Expand);
// Without SSE, i64->f64 goes through memory.
- setOperationAction(ISD::BIT_CONVERT , MVT::i64 , Expand);
+ setOperationAction(ISD::BITCAST , MVT::i64 , Expand);
}
}
// (low) operations are left as Legal, as there are single-result
// instructions for this in x86. Using the two-result multiply instructions
// when both high and low results are needed must be arranged by dagcombine.
- setOperationAction(ISD::MULHS , MVT::i8 , Expand);
- setOperationAction(ISD::MULHU , MVT::i8 , Expand);
- setOperationAction(ISD::SDIV , MVT::i8 , Expand);
- setOperationAction(ISD::UDIV , MVT::i8 , Expand);
- setOperationAction(ISD::SREM , MVT::i8 , Expand);
- setOperationAction(ISD::UREM , MVT::i8 , Expand);
- setOperationAction(ISD::MULHS , MVT::i16 , Expand);
- setOperationAction(ISD::MULHU , MVT::i16 , Expand);
- setOperationAction(ISD::SDIV , MVT::i16 , Expand);
- setOperationAction(ISD::UDIV , MVT::i16 , Expand);
- setOperationAction(ISD::SREM , MVT::i16 , Expand);
- setOperationAction(ISD::UREM , MVT::i16 , Expand);
- setOperationAction(ISD::MULHS , MVT::i32 , Expand);
- setOperationAction(ISD::MULHU , MVT::i32 , Expand);
- setOperationAction(ISD::SDIV , MVT::i32 , Expand);
- setOperationAction(ISD::UDIV , MVT::i32 , Expand);
- setOperationAction(ISD::SREM , MVT::i32 , Expand);
- setOperationAction(ISD::UREM , MVT::i32 , Expand);
- setOperationAction(ISD::MULHS , MVT::i64 , Expand);
- setOperationAction(ISD::MULHU , MVT::i64 , Expand);
- setOperationAction(ISD::SDIV , MVT::i64 , Expand);
- setOperationAction(ISD::UDIV , MVT::i64 , Expand);
- setOperationAction(ISD::SREM , MVT::i64 , Expand);
- setOperationAction(ISD::UREM , MVT::i64 , Expand);
+ for (unsigned i = 0, e = 4; i != e; ++i) {
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::MULHS, VT, Expand);
+ setOperationAction(ISD::MULHU, VT, Expand);
+ setOperationAction(ISD::SDIV, VT, Expand);
+ setOperationAction(ISD::UDIV, VT, Expand);
+ setOperationAction(ISD::SREM, VT, Expand);
+ setOperationAction(ISD::UREM, VT, Expand);
+
+ // Add/Sub overflow ops with MVT::Glues are lowered to EFLAGS dependences.
+ setOperationAction(ISD::ADDC, VT, Custom);
+ setOperationAction(ISD::ADDE, VT, Custom);
+ setOperationAction(ISD::SUBC, VT, Custom);
+ setOperationAction(ISD::SUBE, VT, Custom);
+ }
setOperationAction(ISD::BR_JT , MVT::Other, Expand);
setOperationAction(ISD::BRCOND , MVT::Other, Custom);
setOperationAction(ISD::FREM , MVT::f80 , Expand);
setOperationAction(ISD::FLT_ROUNDS_ , MVT::i32 , Custom);
- setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
setOperationAction(ISD::CTTZ , MVT::i8 , Custom);
setOperationAction(ISD::CTLZ , MVT::i8 , Custom);
- setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
setOperationAction(ISD::CTTZ , MVT::i16 , Custom);
setOperationAction(ISD::CTLZ , MVT::i16 , Custom);
- setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
setOperationAction(ISD::CTTZ , MVT::i32 , Custom);
setOperationAction(ISD::CTLZ , MVT::i32 , Custom);
if (Subtarget->is64Bit()) {
- setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
setOperationAction(ISD::CTTZ , MVT::i64 , Custom);
setOperationAction(ISD::CTLZ , MVT::i64 , Custom);
}
+ if (Subtarget->hasPOPCNT()) {
+ setOperationAction(ISD::CTPOP , MVT::i8 , Promote);
+ } else {
+ setOperationAction(ISD::CTPOP , MVT::i8 , Expand);
+ setOperationAction(ISD::CTPOP , MVT::i16 , Expand);
+ setOperationAction(ISD::CTPOP , MVT::i32 , Expand);
+ if (Subtarget->is64Bit())
+ setOperationAction(ISD::CTPOP , MVT::i64 , Expand);
+ }
+
setOperationAction(ISD::READCYCLECOUNTER , MVT::i64 , Custom);
setOperationAction(ISD::BSWAP , MVT::i16 , Expand);
setOperationAction(ISD::SELECT , MVT::i1 , Promote);
// X86 wants to expand cmov itself.
setOperationAction(ISD::SELECT , MVT::i8 , Custom);
- setOperationAction(ISD::SELECT , MVT::i16 , Custom);
+ setOperationAction(ISD::SELECT , MVT::i16 , Custom);
setOperationAction(ISD::SELECT , MVT::i32 , Custom);
setOperationAction(ISD::SELECT , MVT::f32 , Custom);
setOperationAction(ISD::SELECT , MVT::f64 , Custom);
setOperationAction(ISD::SRL_PARTS , MVT::i64 , Custom);
}
- if (Subtarget->hasSSE1())
+ if (Subtarget->hasXMM())
setOperationAction(ISD::PREFETCH , MVT::Other, Legal);
// We may not have a libcall for MEMBARRIER so we should lower this.
setOperationAction(ISD::MEMBARRIER , MVT::Other, Custom);
-
+
// On X86 and X86-64, atomic operations are lowered to locked instructions.
// Locked instructions, in turn, have implicit fence semantics (all memory
// operations are flushed before issuing the locked instruction, and they
setShouldFoldAtomicFences(true);
// Expand certain atomics
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i8, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i16, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i32, Custom);
- setOperationAction(ISD::ATOMIC_CMP_SWAP, MVT::i64, Custom);
-
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i8, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i16, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i32, Custom);
- setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
-
+ for (unsigned i = 0, e = 4; i != e; ++i) {
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::ATOMIC_CMP_SWAP, VT, Custom);
+ setOperationAction(ISD::ATOMIC_LOAD_SUB, VT, Custom);
+ }
+
if (!Subtarget->is64Bit()) {
setOperationAction(ISD::ATOMIC_LOAD_ADD, MVT::i64, Custom);
setOperationAction(ISD::ATOMIC_LOAD_SUB, MVT::i64, Custom);
setOperationAction(ISD::STACKRESTORE, MVT::Other, Expand);
if (Subtarget->is64Bit())
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i64, Expand);
- if (Subtarget->isTargetCygMing())
+ if (Subtarget->isTargetCygMing() || Subtarget->isTargetWindows())
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Custom);
else
setOperationAction(ISD::DYNAMIC_STACKALLOC, MVT::i32, Expand);
setOperationAction(ISD::UNDEF, MVT::f80, Expand);
setOperationAction(ISD::FCOPYSIGN, MVT::f80, Expand);
{
- bool ignored;
- APFloat TmpFlt(+0.0);
- TmpFlt.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
- &ignored);
+ APFloat TmpFlt = APFloat::getZero(APFloat::x87DoubleExtended);
addLegalFPImmediate(TmpFlt); // FLD0
TmpFlt.changeSign();
addLegalFPImmediate(TmpFlt); // FLD0/FCHS
+
+ bool ignored;
APFloat TmpFlt2(+1.0);
TmpFlt2.convert(APFloat::x87DoubleExtended, APFloat::rmNearestTiesToEven,
&ignored);
// FIXME: In order to prevent SSE instructions being expanded to MMX ones
// with -msoft-float, disable use of MMX as well.
- if (!UseSoftFloat && !DisableMMX && Subtarget->hasMMX()) {
- addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass, false);
+ if (!UseSoftFloat && Subtarget->hasMMX()) {
+ addRegisterClass(MVT::x86mmx, X86::VR64RegisterClass);
// No operations on x86mmx supported, everything uses intrinsics.
}
setOperationAction(ISD::SELECT, MVT::v4i16, Expand);
setOperationAction(ISD::SELECT, MVT::v2i32, Expand);
setOperationAction(ISD::SELECT, MVT::v1i64, Expand);
- setOperationAction(ISD::BIT_CONVERT, MVT::v8i8, Expand);
- setOperationAction(ISD::BIT_CONVERT, MVT::v4i16, Expand);
- setOperationAction(ISD::BIT_CONVERT, MVT::v2i32, Expand);
- setOperationAction(ISD::BIT_CONVERT, MVT::v1i64, Expand);
+ setOperationAction(ISD::BITCAST, MVT::v8i8, Expand);
+ setOperationAction(ISD::BITCAST, MVT::v4i16, Expand);
+ setOperationAction(ISD::BITCAST, MVT::v2i32, Expand);
+ setOperationAction(ISD::BITCAST, MVT::v1i64, Expand);
- if (!UseSoftFloat && Subtarget->hasSSE1()) {
+ if (!UseSoftFloat && Subtarget->hasXMM()) {
addRegisterClass(MVT::v4f32, X86::VR128RegisterClass);
setOperationAction(ISD::FADD, MVT::v4f32, Legal);
setOperationAction(ISD::VSETCC, MVT::v4f32, Custom);
}
- if (!UseSoftFloat && Subtarget->hasSSE2()) {
+ if (!UseSoftFloat && Subtarget->hasXMMInt()) {
addRegisterClass(MVT::v2f64, X86::VR128RegisterClass);
// FIXME: Unfortunately -soft-float and -no-implicit-float means XMM
// Do not attempt to promote non-128-bit vectors
if (!VT.is128BitVector())
continue;
-
+
setOperationAction(ISD::AND, SVT, Promote);
AddPromotedToType (ISD::AND, SVT, MVT::v2i64);
setOperationAction(ISD::OR, SVT, Promote);
}
}
- if (Subtarget->hasSSE42()) {
+ if (Subtarget->hasSSE42())
setOperationAction(ISD::VSETCC, MVT::v2i64, Custom);
- }
if (!UseSoftFloat && Subtarget->hasAVX()) {
addRegisterClass(MVT::v8f32, X86::VR256RegisterClass);
// We want to custom lower some of our intrinsics.
setOperationAction(ISD::INTRINSIC_WO_CHAIN, MVT::Other, Custom);
- // Add/Sub/Mul with overflow operations are custom lowered.
- setOperationAction(ISD::SADDO, MVT::i32, Custom);
- setOperationAction(ISD::UADDO, MVT::i32, Custom);
- setOperationAction(ISD::SSUBO, MVT::i32, Custom);
- setOperationAction(ISD::USUBO, MVT::i32, Custom);
- setOperationAction(ISD::SMULO, MVT::i32, Custom);
-
+
// Only custom-lower 64-bit SADDO and friends on 64-bit because we don't
// handle type legalization for these operations here.
//
// FIXME: We really should do custom legalization for addition and
// subtraction on x86-32 once PR3203 is fixed. We really can't do much better
// than generic legalization for 64-bit multiplication-with-overflow, though.
- if (Subtarget->is64Bit()) {
- setOperationAction(ISD::SADDO, MVT::i64, Custom);
- setOperationAction(ISD::UADDO, MVT::i64, Custom);
- setOperationAction(ISD::SSUBO, MVT::i64, Custom);
- setOperationAction(ISD::USUBO, MVT::i64, Custom);
- setOperationAction(ISD::SMULO, MVT::i64, Custom);
+ for (unsigned i = 0, e = 3+Subtarget->is64Bit(); i != e; ++i) {
+ // Add/Sub/Mul with overflow operations are custom lowered.
+ MVT VT = IntVTs[i];
+ setOperationAction(ISD::SADDO, VT, Custom);
+ setOperationAction(ISD::UADDO, VT, Custom);
+ setOperationAction(ISD::SSUBO, VT, Custom);
+ setOperationAction(ISD::USUBO, VT, Custom);
+ setOperationAction(ISD::SMULO, VT, Custom);
+ setOperationAction(ISD::UMULO, VT, Custom);
}
+
+ // There are no 8-bit 3-address imul/mul instructions
+ setOperationAction(ISD::SMULO, MVT::i8, Expand);
+ setOperationAction(ISD::UMULO, MVT::i8, Expand);
if (!Subtarget->is64Bit()) {
// These libcalls are not available in 32-bit.
setTargetDAGCombine(ISD::SRA);
setTargetDAGCombine(ISD::SRL);
setTargetDAGCombine(ISD::OR);
+ setTargetDAGCombine(ISD::AND);
+ setTargetDAGCombine(ISD::ADD);
+ setTargetDAGCombine(ISD::SUB);
setTargetDAGCombine(ISD::STORE);
setTargetDAGCombine(ISD::ZERO_EXTEND);
if (Subtarget->is64Bit())
computeRegisterProperties();
- // FIXME: These should be based on subtarget info. Plus, the values should
- // be smaller when we are in optimizing for size mode.
+ // On Darwin, -Os means optimize for size without hurting performance,
+ // do not reduce the limit.
maxStoresPerMemset = 16; // For @llvm.memset -> sequence of stores
+ maxStoresPerMemsetOptSize = Subtarget->isTargetDarwin() ? 16 : 8;
maxStoresPerMemcpy = 8; // For @llvm.memcpy -> sequence of stores
- maxStoresPerMemmove = 3; // For @llvm.memmove -> sequence of stores
+ maxStoresPerMemcpyOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
+ maxStoresPerMemmove = 8; // For @llvm.memmove -> sequence of stores
+ maxStoresPerMemmoveOptSize = Subtarget->isTargetDarwin() ? 8 : 4;
setPrefLoopAlignment(16);
benefitFromCodePlacementOpt = true;
}
}
unsigned Align = 4;
- if (Subtarget->hasSSE1())
+ if (Subtarget->hasXMM())
getMaxByValAlign(Ty, Align);
return Align;
}
// linux. This is because the stack realignment code can't handle certain
// cases like PR2962. This should be removed when PR2962 is fixed.
const Function *F = MF.getFunction();
- if (NonScalarIntSafe &&
- !F->hasFnAttr(Attribute::NoImplicitFloat)) {
+ if (NonScalarIntSafe && !F->hasFnAttr(Attribute::NoImplicitFloat)) {
if (Size >= 16 &&
- (Subtarget->isUnalignedMemAccessFast() ||
- ((DstAlign == 0 || DstAlign >= 16) &&
- (SrcAlign == 0 || SrcAlign >= 16))) &&
Subtarget->getStackAlignment() >= 16) {
if (Subtarget->hasSSE2())
return MVT::v4i32;
} else if (!MemcpyStrSrc && Size >= 8 &&
!Subtarget->is64Bit() &&
Subtarget->getStackAlignment() >= 8 &&
- Subtarget->hasSSE2()) {
+ Subtarget->hasXMMInt()) {
// Do not use f64 to lower memcpy if source is string constant. It's
// better to use i32 to avoid the loads.
return MVT::f64;
if (getTargetMachine().getRelocationModel() == Reloc::PIC_ &&
Subtarget->isPICStyleGOT())
return MachineJumpTableInfo::EK_Custom32;
-
+
// Otherwise, use the normal jump table encoding heuristics.
return TargetLowering::getJumpTableEncoding();
}
-/// getPICBaseSymbol - Return the X86-32 PIC base.
-MCSymbol *
-X86TargetLowering::getPICBaseSymbol(const MachineFunction *MF,
- MCContext &Ctx) const {
- const MCAsmInfo &MAI = *getTargetMachine().getMCAsmInfo();
- return Ctx.GetOrCreateSymbol(Twine(MAI.getPrivateGlobalPrefix())+
- Twine(MF->getFunctionNumber())+"$pb");
-}
-
-
const MCExpr *
X86TargetLowering::LowerCustomJumpTableEntry(const MachineJumpTableInfo *MJTI,
const MachineBasicBlock *MBB,
return TargetLowering::getPICJumpTableRelocBaseExpr(MF, JTI, Ctx);
// Otherwise, the reference is relative to the PIC base.
- return MCSymbolRefExpr::Create(getPICBaseSymbol(MF, Ctx), Ctx);
+ return MCSymbolRefExpr::Create(MF->getPICBaseSymbol(), Ctx);
}
/// getFunctionAlignment - Return the Log2 alignment of this function.
unsigned
X86TargetLowering::getRegPressureLimit(const TargetRegisterClass *RC,
MachineFunction &MF) const {
- unsigned FPDiff = RegInfo->hasFP(MF) ? 1 : 0;
+ const TargetFrameInfo *TFI = MF.getTarget().getFrameInfo();
+
+ unsigned FPDiff = TFI->hasFP(MF) ? 1 : 0;
switch (RC->getID()) {
default:
return 0;
#include "X86GenCallingConv.inc"
-bool
+bool
X86TargetLowering::CanLowerReturn(CallingConv::ID CallConv, bool isVarArg,
const SmallVectorImpl<ISD::OutputArg> &Outs,
LLVMContext &Context) const {
// or SSE or MMX vectors.
if ((ValVT == MVT::f32 || ValVT == MVT::f64 ||
VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) &&
- (Subtarget->is64Bit() && !Subtarget->hasSSE1())) {
+ (Subtarget->is64Bit() && !Subtarget->hasXMM())) {
report_fatal_error("SSE register return with SSE disabled");
}
// Likewise we can't return F64 values with SSE1 only. gcc does so, but
// llvm-gcc has never done it right and no one has noticed, so this
// should be OK for now.
if (ValVT == MVT::f64 &&
- (Subtarget->is64Bit() && !Subtarget->hasSSE2()))
+ (Subtarget->is64Bit() && !Subtarget->hasXMMInt()))
report_fatal_error("SSE2 register return with SSE2 disabled");
// Returns in ST0/ST1 are handled specially: these are pushed as operands to
if (Subtarget->is64Bit()) {
if (ValVT == MVT::x86mmx) {
if (VA.getLocReg() == X86::XMM0 || VA.getLocReg() == X86::XMM1) {
- ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, ValToCopy);
+ ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::i64, ValToCopy);
ValToCopy = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64,
ValToCopy);
// If we don't have SSE2 available, convert to v4f32 so the generated
// register is legal.
if (!Subtarget->hasSSE2())
- ValToCopy = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32,ValToCopy);
+ ValToCopy = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32,ValToCopy);
}
}
}
-
+
Chain = DAG.getCopyToReg(Chain, dl, VA.getLocReg(), ValToCopy, Flag);
Flag = Chain.getValue(1);
}
MachineFunction &MF = DAG.getMachineFunction();
X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
unsigned Reg = FuncInfo->getSRetReturnReg();
- assert(Reg &&
+ assert(Reg &&
"SRetReturnReg should have been set in LowerFormalArguments().");
SDValue Val = DAG.getCopyFromReg(Chain, dl, Reg, getPointerTy());
MVT::Other, &RetOps[0], RetOps.size());
}
+bool X86TargetLowering::isUsedByReturnOnly(SDNode *N) const {
+ if (N->getNumValues() != 1)
+ return false;
+ if (!N->hasNUsesOfValue(1, 0))
+ return false;
+
+ SDNode *Copy = *N->use_begin();
+ if (Copy->getOpcode() != ISD::CopyToReg &&
+ Copy->getOpcode() != ISD::FP_EXTEND)
+ return false;
+
+ bool HasRet = false;
+ for (SDNode::use_iterator UI = Copy->use_begin(), UE = Copy->use_end();
+ UI != UE; ++UI) {
+ if (UI->getOpcode() != X86ISD::RET_FLAG)
+ return false;
+ HasRet = true;
+ }
+
+ return HasRet;
+}
+
/// LowerCallResult - Lower the result values of a call into the
/// appropriate copies out of appropriate physical registers.
///
// If this is x86-64, and we disabled SSE, we can't return FP values
if ((CopyVT == MVT::f32 || CopyVT == MVT::f64) &&
- ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasSSE1())) {
+ ((Is64Bit || Ins[i].Flags.isInReg()) && !Subtarget->hasXMM())) {
report_fatal_error("SSE register return with SSE disabled");
}
if (CopyVT == MVT::f64) Opc = isST0 ? X86::FpGET_ST0_64:X86::FpGET_ST1_64;
if (CopyVT == MVT::f80) Opc = isST0 ? X86::FpGET_ST0_80:X86::FpGET_ST1_80;
SDValue Ops[] = { Chain, InFlag };
- Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Flag,
+ Chain = SDValue(DAG.getMachineNode(Opc, dl, CopyVT, MVT::Other, MVT::Glue,
Ops, 2), 1);
Val = Chain.getValue(0);
MVT::i64, InFlag).getValue(1);
Val = Chain.getValue(0);
}
- Val = DAG.getNode(ISD::BIT_CONVERT, dl, CopyVT, Val);
+ Val = DAG.getNode(ISD::BITCAST, dl, CopyVT, Val);
} else {
Chain = DAG.getCopyFromReg(Chain, dl, VA.getLocReg(),
CopyVT, InFlag).getValue(1);
return Ins[0].Flags.isSRet();
}
-/// CCAssignFnForNode - Selects the correct CCAssignFn for a the
-/// given CallingConvention value.
-CCAssignFn *X86TargetLowering::CCAssignFnForNode(CallingConv::ID CC) const {
- if (Subtarget->is64Bit()) {
- if (CC == CallingConv::GHC)
- return CC_X86_64_GHC;
- else if (Subtarget->isTargetWin64())
- return CC_X86_Win64_C;
- else
- return CC_X86_64_C;
- }
-
- if (CC == CallingConv::X86_FastCall)
- return CC_X86_32_FastCall;
- else if (CC == CallingConv::X86_ThisCall)
- return CC_X86_32_ThisCall;
- else if (CC == CallingConv::Fast)
- return CC_X86_32_FastCC;
- else if (CC == CallingConv::GHC)
- return CC_X86_32_GHC;
- else
- return CC_X86_32_C;
-}
-
/// CreateCopyOfByValArgument - Make a copy of an aggregate at address specified
/// by "Src" to address "Dst" with size and alignment information specified by
/// the specific parameter attribute. The copy will be passed as a byval
ISD::ArgFlagsTy Flags, SelectionDAG &DAG,
DebugLoc dl) {
SDValue SizeNode = DAG.getConstant(Flags.getByValSize(), MVT::i32);
-
+
return DAG.getMemcpy(Chain, dl, Dst, Src, SizeNode, Flags.getByValAlign(),
/*isVolatile*/false, /*AlwaysInline=*/true,
MachinePointerInfo(), MachinePointerInfo());
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
- CCInfo.AnalyzeFormalArguments(Ins, CCAssignFnForNode(CallConv));
+ CCInfo.AnalyzeFormalArguments(Ins, CC_X86);
unsigned LastVal = ~0U;
SDValue ArgValue;
ArgValue = DAG.getNode(ISD::AssertZext, dl, RegVT, ArgValue,
DAG.getValueType(VA.getValVT()));
else if (VA.getLocInfo() == CCValAssign::BCvt)
- ArgValue = DAG.getNode(ISD::BIT_CONVERT, dl, VA.getValVT(), ArgValue);
+ ArgValue = DAG.getNode(ISD::BITCAST, dl, VA.getValVT(), ArgValue);
if (VA.isExtInLoc()) {
// Handle MMX values passed in XMM regs.
TotalNumIntRegs);
bool NoImplicitFloatOps = Fn->hasFnAttr(Attribute::NoImplicitFloat);
- assert(!(NumXMMRegs && !Subtarget->hasSSE1()) &&
+ assert(!(NumXMMRegs && !Subtarget->hasXMM()) &&
"SSE register cannot be used when SSE is disabled!");
assert(!(NumXMMRegs && UseSoftFloat && NoImplicitFloatOps) &&
"SSE register cannot be used when SSE is disabled!");
- if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasSSE1())
+ if (UseSoftFloat || NoImplicitFloatOps || !Subtarget->hasXMM())
// Kernel mode asks for SSE to be disabled, so don't push them
// on the stack.
TotalNumXMMRegs = 0;
if (IsWin64) {
+ const TargetFrameInfo &TFI = *getTargetMachine().getFrameInfo();
+ // Get to the caller-allocated home save location. Add 8 to account
+ // for the return address.
+ int HomeOffset = TFI.getOffsetOfLocalArea() + 8;
FuncInfo->setRegSaveFrameIndex(
- MFI->CreateFixedObject(1, NumIntRegs * 8, false));
+ MFI->CreateFixedObject(1, NumIntRegs * 8 + HomeOffset, false));
FuncInfo->setVarArgsFrameIndex(FuncInfo->getRegSaveFrameIndex());
} else {
// For X86-64, if there are vararg parameters that are passed via
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CallConv, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
- CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CallConv));
+ CCInfo.AnalyzeCallOperands(Outs, CC_X86);
// Get a count of how many bytes are to be pushed on the stack.
unsigned NumBytes = CCInfo.getNextStackOffset();
case CCValAssign::AExt:
if (RegVT.isVector() && RegVT.getSizeInBits() == 128) {
// Special case: passing MMX values in XMM registers.
- Arg = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::i64, Arg);
+ Arg = DAG.getNode(ISD::BITCAST, dl, MVT::i64, Arg);
Arg = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2i64, Arg);
Arg = getMOVL(DAG, dl, MVT::v2i64, DAG.getUNDEF(MVT::v2i64), Arg);
} else
Arg = DAG.getNode(ISD::ANY_EXTEND, dl, RegVT, Arg);
break;
case CCValAssign::BCvt:
- Arg = DAG.getNode(ISD::BIT_CONVERT, dl, RegVT, Arg);
+ Arg = DAG.getNode(ISD::BITCAST, dl, RegVT, Arg);
break;
case CCValAssign::Indirect: {
// Store the argument.
X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
};
unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
- assert((Subtarget->hasSSE1() || !NumXMMRegs)
+ assert((Subtarget->hasXMM() || !NumXMMRegs)
&& "SSE registers cannot be used when SSE is disabled");
Chain = DAG.getCopyToReg(Chain, dl, X86::AL,
} else if (ExternalSymbolSDNode *S = dyn_cast<ExternalSymbolSDNode>(Callee)) {
unsigned char OpFlags = 0;
- // On ELF targets, in either X86-64 or X86-32 mode, direct calls to external
- // symbols should go through the PLT.
+ // On ELF targets, in either X86-64 or X86-32 mode, direct calls to
+ // external symbols should go through the PLT.
if (Subtarget->isTargetELF() &&
getTargetMachine().getRelocationModel() == Reloc::PIC_) {
OpFlags = X86II::MO_PLT;
}
// Returns a chain & a flag for retval copy to use.
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
SmallVector<SDValue, 8> Ops;
if (!IsSibcall && isTailCall) {
SmallVector<CCValAssign, 16> ArgLocs;
CCState CCInfo(CalleeCC, isVarArg, getTargetMachine(),
ArgLocs, *DAG.getContext());
- CCInfo.AnalyzeCallOperands(Outs, CCAssignFnForNode(CalleeCC));
+ CCInfo.AnalyzeCallOperands(Outs, CC_X86);
if (CCInfo.getNextStackOffset()) {
MachineFunction &MF = DAG.getMachineFunction();
if (MF.getInfo<X86MachineFunctionInfo>()->getBytesToPopOnReturn())
}
}
+ // An stdcall caller is expected to clean up its arguments; the callee
+ // isn't going to do that.
+ if (!CCMatch && CallerCC==CallingConv::X86_StdCall)
+ return false;
+
return true;
}
static bool isPALIGNRMask(const SmallVectorImpl<int> &Mask, EVT VT,
bool hasSSSE3) {
int i, e = VT.getVectorNumElements();
-
+
// Do not handle v2i64 / v2f64 shuffles with palignr.
if (e < 4 || !hasSSSE3)
return false;
-
+
for (i = 0; i != e; ++i)
if (Mask[i] >= 0)
break;
-
+
// All undef, not a palignr.
if (i == e)
return false;
bool NeedsUnary = false;
int s = Mask[i] - i;
-
+
// Check the rest of the elements to see if they are consecutive.
for (++i; i != e; ++i) {
int m = Mask[i];
- if (m < 0)
+ if (m < 0)
continue;
-
+
Unary = Unary && (m < (int)e);
NeedsUnary = NeedsUnary || (m < s);
/// <2, 3, 2, 3>
bool X86::isMOVHLPS_v_undef_Mask(ShuffleVectorSDNode *N) {
unsigned NumElems = N->getValueType(0).getVectorNumElements();
-
+
if (NumElems != 4)
return false;
-
+
return isUndefOrEqual(N->getMaskElt(0), 2) &&
isUndefOrEqual(N->getMaskElt(1), 3) &&
isUndefOrEqual(N->getMaskElt(2), 2) &&
SDValue Ops[] = { Cst, Cst, Cst, Cst, Cst, Cst, Cst, Cst };
Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v8f32, Ops, 8);
}
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
+ return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
}
/// getOnesVector - Returns a vector of specified type with all bits set.
SDValue Cst = DAG.getTargetConstant(~0U, MVT::i32);
SDValue Vec;
Vec = DAG.getNode(ISD::BUILD_VECTOR, dl, MVT::v4i32, Cst, Cst, Cst, Cst);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Vec);
+ return DAG.getNode(ISD::BITCAST, dl, VT, Vec);
}
// Perform the splat.
int SplatMask[4] = { EltNo, EltNo, EltNo, EltNo };
- V1 = DAG.getNode(ISD::BIT_CONVERT, dl, PVT, V1);
+ V1 = DAG.getNode(ISD::BITCAST, dl, PVT, V1);
V1 = DAG.getVectorShuffle(PVT, dl, V1, DAG.getUNDEF(PVT), &SplatMask[0]);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, V1);
+ return DAG.getNode(ISD::BITCAST, dl, VT, V1);
}
/// getShuffleVectorZeroOrUndef - Return a vector_shuffle of the specified
}
// Actual nodes that may contain scalar elements
- if (Opcode == ISD::BIT_CONVERT) {
+ if (Opcode == ISD::BITCAST) {
V = V.getOperand(0);
EVT SrcVT = V.getValueType();
unsigned NumElems = VT.getVectorNumElements();
}
}
- return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V);
+ return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, V);
}
/// LowerBuildVectorv8i16 - Custom lower build_vector of v8i16.
const TargetLowering &TLI, DebugLoc dl) {
EVT ShVT = MVT::v2i64;
unsigned Opc = isLeft ? X86ISD::VSHL : X86ISD::VSRL;
- SrcOp = DAG.getNode(ISD::BIT_CONVERT, dl, ShVT, SrcOp);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ SrcOp = DAG.getNode(ISD::BITCAST, dl, ShVT, SrcOp);
+ return DAG.getNode(ISD::BITCAST, dl, VT,
DAG.getNode(Opc, dl, ShVT, SrcOp,
DAG.getConstant(NumBits, TLI.getShiftAmountTy())));
}
SDValue
X86TargetLowering::LowerAsSplatVectorLoad(SDValue SrcOp, EVT VT, DebugLoc dl,
SelectionDAG &DAG) const {
-
+
// Check if the scalar load can be widened into a vector load. And if
// the address is "base + cst" see if the cst can be "absorbed" into
// the shuffle mask.
LD->getPointerInfo().getWithOffset(StartOffset),
false, false, 0);
// Canonicalize it to a v4i32 shuffle.
- V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32, V1);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v4i32, V1);
+ return DAG.getNode(ISD::BITCAST, dl, VT,
DAG.getVectorShuffle(MVT::v4i32, dl, V1,
DAG.getUNDEF(MVT::v4i32),&Mask[0]));
}
return SDValue();
}
-/// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
-/// vector of type 'VT', see if the elements can be replaced by a single large
+/// EltsFromConsecutiveLoads - Given the initializing elements 'Elts' of a
+/// vector of type 'VT', see if the elements can be replaced by a single large
/// load which has the same value as a build_vector whose operands are 'elts'.
///
/// Example: <load i32 *a, load i32 *a+4, undef, undef> -> zextload a
-///
+///
/// FIXME: we'd also like to handle the case where the last elements are zero
/// rather than undef via VZEXT_LOAD, but we do not detect that case today.
/// There's even a handy isZeroNode for that purpose.
DebugLoc &DL, SelectionDAG &DAG) {
EVT EltVT = VT.getVectorElementType();
unsigned NumElems = Elts.size();
-
+
LoadSDNode *LDBase = NULL;
unsigned LastLoadedElt = -1U;
-
+
// For each element in the initializer, see if we've found a load or an undef.
- // If we don't find an initial load element, or later load elements are
+ // If we don't find an initial load element, or later load elements are
// non-consecutive, bail out.
for (unsigned i = 0; i < NumElems; ++i) {
SDValue Elt = Elts[i];
-
+
if (!Elt.getNode() ||
(Elt.getOpcode() != ISD::UNDEF && !ISD::isNON_EXTLoad(Elt.getNode())))
return SDValue();
SDValue ResNode = DAG.getMemIntrinsicNode(X86ISD::VZEXT_LOAD, DL, Tys,
Ops, 2, MVT::i32,
LDBase->getMemOperand());
- return DAG.getNode(ISD::BIT_CONVERT, DL, VT, ResNode);
+ return DAG.getNode(ISD::BITCAST, DL, VT, ResNode);
}
return SDValue();
}
DAG.getUNDEF(Item.getValueType()),
&Mask[0]);
}
- return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(), Item);
+ return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(), Item);
}
}
Item = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MiddleVT, Item);
Item = getShuffleVectorZeroOrUndef(Item, 0, true,
Subtarget->hasSSE2(), DAG);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, Item);
+ return DAG.getNode(ISD::BITCAST, dl, VT, Item);
}
}
// Check for a build vector of consecutive loads.
for (unsigned i = 0; i < NumElems; ++i)
V[i] = Op.getOperand(i);
-
+
// Check for elements which are consecutive loads.
SDValue LD = EltsFromConsecutiveLoads(VT, V, dl, DAG);
if (LD.getNode())
return LD;
-
- // For SSE 4.1, use insertps to put the high elements into the low element.
+
+ // For SSE 4.1, use insertps to put the high elements into the low element.
if (getSubtarget()->hasSSE41()) {
SDValue Result;
if (Op.getOperand(0).getOpcode() != ISD::UNDEF)
Result = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, VT, Op.getOperand(0));
else
Result = DAG.getUNDEF(VT);
-
+
for (unsigned i = 1; i < NumElems; ++i) {
if (Op.getOperand(i).getOpcode() == ISD::UNDEF) continue;
Result = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, VT, Result,
}
return Result;
}
-
+
// Otherwise, expand into a number of unpckl*, start by extending each of
// our (non-undef) elements to the full vector width with the element in the
// bottom slot of the vector (which generates no code for SSE).
if (V[i+EltStride].getOpcode() == ISD::UNDEF &&
EltStride == NumElems/2)
continue;
-
+
V[i] = getUnpackl(DAG, dl, VT, V[i], V[i + EltStride]);
}
EltStride >>= 1;
assert(ResVT == MVT::v2i64 || ResVT == MVT::v4i32 ||
ResVT == MVT::v8i16 || ResVT == MVT::v16i8);
int Mask[2];
- SDValue InVec = DAG.getNode(ISD::BIT_CONVERT,dl, MVT::v1i64, Op.getOperand(0));
+ SDValue InVec = DAG.getNode(ISD::BITCAST,dl, MVT::v1i64, Op.getOperand(0));
SDValue VecOp = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
InVec = Op.getOperand(1);
if (InVec.getOpcode() == ISD::SCALAR_TO_VECTOR) {
unsigned NumElts = ResVT.getVectorNumElements();
- VecOp = DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
+ VecOp = DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
VecOp = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, ResVT, VecOp,
InVec.getOperand(0), DAG.getIntPtrConstant(NumElts/2+1));
} else {
- InVec = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v1i64, InVec);
+ InVec = DAG.getNode(ISD::BITCAST, dl, MVT::v1i64, InVec);
SDValue VecOp2 = DAG.getNode(X86ISD::MOVQ2DQ, dl, MVT::v2i64, InVec);
Mask[0] = 0; Mask[1] = 2;
VecOp = DAG.getVectorShuffle(MVT::v2i64, dl, VecOp, VecOp2, Mask);
}
- return DAG.getNode(ISD::BIT_CONVERT, dl, ResVT, VecOp);
+ return DAG.getNode(ISD::BITCAST, dl, ResVT, VecOp);
}
// v8i16 shuffles - Prefer shuffles in the following order:
MaskV.push_back(BestLoQuad < 0 ? 0 : BestLoQuad);
MaskV.push_back(BestHiQuad < 0 ? 1 : BestHiQuad);
NewV = DAG.getVectorShuffle(MVT::v2i64, dl,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V1),
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, V2), &MaskV[0]);
- NewV = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, NewV);
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V1),
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, V2), &MaskV[0]);
+ NewV = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, NewV);
// Rewrite the MaskVals and assign NewV to V1 if NewV now contains all the
// source words for the shuffle, to aid later transformations.
pshufbMask.push_back(DAG.getConstant(EltIdx, MVT::i8));
pshufbMask.push_back(DAG.getConstant(EltIdx+1, MVT::i8));
}
- V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V1);
+ 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));
if (!TwoInputs)
- return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
+ 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.push_back(DAG.getConstant(EltIdx - 16, MVT::i8));
pshufbMask.push_back(DAG.getConstant(EltIdx - 15, MVT::i8));
}
- V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, V2);
+ 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));
V1 = DAG.getNode(ISD::OR, dl, MVT::v16i8, V1, V2);
- return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
+ return DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
}
// If BestLoQuad >= 0, generate a pshuflw to put the low elements in order,
// No SSSE3 - Calculate in place words and then fix all out of place words
// With 0-16 extracts & inserts. Worst case is 16 bytes out of order from
// the 16 different words that comprise the two doublequadword input vectors.
- V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V1);
- V2 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v8i16, V2);
+ V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V1);
+ V2 = DAG.getNode(ISD::BITCAST, dl, MVT::v8i16, V2);
SDValue NewV = V2Only ? V2 : V1;
for (int i = 0; i != 8; ++i) {
int Elt0 = MaskVals[i*2];
NewV = DAG.getNode(ISD::INSERT_VECTOR_ELT, dl, MVT::v8i16, NewV, InsElt,
DAG.getIntPtrConstant(i));
}
- return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v16i8, NewV);
+ return DAG.getNode(ISD::BITCAST, dl, MVT::v16i8, NewV);
}
/// RewriteAsNarrowerShuffle - Try rewriting v8i16 and v16i8 shuffles as 4 wide
MaskVec.push_back(StartIdx / Scale);
}
- V1 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V1);
- V2 = DAG.getNode(ISD::BIT_CONVERT, dl, NewVT, V2);
+ V1 = DAG.getNode(ISD::BITCAST, dl, NewVT, V1);
+ V2 = DAG.getNode(ISD::BITCAST, dl, NewVT, V2);
return DAG.getVectorShuffle(NewVT, dl, V1, V2, &MaskVec[0]);
}
// movssrr and movsdrr do not clear top bits. Try to use movd, movq
// instead.
MVT ExtVT = (OpVT == MVT::v2f64) ? MVT::i64 : MVT::i32;
- if ((ExtVT.SimpleTy != MVT::i64 || Subtarget->is64Bit()) &&
+ if ((ExtVT != MVT::i64 || Subtarget->is64Bit()) &&
SrcOp.getOpcode() == ISD::SCALAR_TO_VECTOR &&
- SrcOp.getOperand(0).getOpcode() == ISD::BIT_CONVERT &&
+ SrcOp.getOperand(0).getOpcode() == ISD::BITCAST &&
SrcOp.getOperand(0).getOperand(0).getValueType() == ExtVT) {
// PR2108
OpVT = (OpVT == MVT::v2f64) ? MVT::v2i64 : MVT::v4i32;
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ return DAG.getNode(ISD::BITCAST, dl, VT,
DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
OpVT,
}
}
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ return DAG.getNode(ISD::BITCAST, dl, VT,
DAG.getNode(X86ISD::VZEXT_MOVL, dl, OpVT,
- DAG.getNode(ISD::BIT_CONVERT, dl,
+ DAG.getNode(ISD::BITCAST, dl,
OpVT, SrcOp)));
}
}
static bool MayFoldVectorLoad(SDValue V) {
- if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
+ if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
V = V.getOperand(0);
if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
V = V.getOperand(0);
// one use. Remove this version after this bug get fixed.
// rdar://8434668, PR8156
static bool RelaxedMayFoldVectorLoad(SDValue V) {
- if (V.hasOneUse() && V.getOpcode() == ISD::BIT_CONVERT)
+ if (V.hasOneUse() && V.getOpcode() == ISD::BITCAST)
V = V.getOperand(0);
if (V.hasOneUse() && V.getOpcode() == ISD::SCALAR_TO_VECTOR)
V = V.getOperand(0);
// If the bit convert changed the number of elements, it is unsafe
// to examine the mask.
bool HasShuffleIntoBitcast = false;
- if (V.getOpcode() == ISD::BIT_CONVERT) {
+ if (V.getOpcode() == ISD::BITCAST) {
EVT SrcVT = V.getOperand(0).getValueType();
if (SrcVT.getVectorNumElements() != VT.getVectorNumElements())
return false;
V = (Idx < (int)NumElems) ? V.getOperand(0) : V.getOperand(1);
// Skip one more bit_convert if necessary
- if (V.getOpcode() == ISD::BIT_CONVERT)
+ if (V.getOpcode() == ISD::BITCAST)
V = V.getOperand(0);
if (ISD::isNormalLoad(V.getNode())) {
EVT VT = Op.getValueType();
// Canonizalize to v2f64.
- V1 = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, V1);
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ V1 = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, V1);
+ return DAG.getNode(ISD::BITCAST, dl, VT,
getTargetShuffleNode(X86ISD::MOVDDUP, dl, MVT::v2f64,
V1, DAG));
}
if (VT == MVT::v8i16 || VT == MVT::v16i8) {
SDValue NewOp = RewriteAsNarrowerShuffle(SVOp, DAG, dl);
if (NewOp.getNode())
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT, NewOp);
+ return DAG.getNode(ISD::BITCAST, dl, VT, NewOp);
} else if ((VT == MVT::v4i32 || (VT == MVT::v4f32 && Subtarget->hasSSE2()))) {
// FIXME: Figure out a cleaner way to do this.
// Try to make use of movq to zero out the top part.
if (Idx == 0)
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
- DAG.getNode(ISD::BIT_CONVERT, dl,
+ DAG.getNode(ISD::BITCAST, dl,
MVT::v4i32,
Op.getOperand(0)),
Op.getOperand(1)));
if ((User->getOpcode() != ISD::STORE ||
(isa<ConstantSDNode>(Op.getOperand(1)) &&
cast<ConstantSDNode>(Op.getOperand(1))->isNullValue())) &&
- (User->getOpcode() != ISD::BIT_CONVERT ||
+ (User->getOpcode() != ISD::BITCAST ||
User->getValueType(0) != MVT::i32))
return SDValue();
SDValue Extract = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4i32,
+ DAG.getNode(ISD::BITCAST, dl, MVT::v4i32,
Op.getOperand(0)),
Op.getOperand(1));
- return DAG.getNode(ISD::BIT_CONVERT, dl, MVT::f32, Extract);
+ return DAG.getNode(ISD::BITCAST, dl, MVT::f32, Extract);
} else if (VT == MVT::i32) {
// ExtractPS works with constant index.
if (isa<ConstantSDNode>(Op.getOperand(1)))
if (Idx == 0)
return DAG.getNode(ISD::TRUNCATE, dl, MVT::i16,
DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::i32,
- DAG.getNode(ISD::BIT_CONVERT, dl,
+ DAG.getNode(ISD::BITCAST, dl,
MVT::v4i32, Vec),
Op.getOperand(1)));
// Transform it so it match pextrw which produces a 32-bit result.
SDValue
X86TargetLowering::LowerSCALAR_TO_VECTOR(SDValue Op, SelectionDAG &DAG) const {
DebugLoc dl = Op.getDebugLoc();
-
+
if (Op.getValueType() == MVT::v1i64 &&
Op.getOperand(0).getValueType() == MVT::i64)
return DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v1i64, Op.getOperand(0));
SDValue AnyExt = DAG.getNode(ISD::ANY_EXTEND, dl, MVT::i32, Op.getOperand(0));
assert(Op.getValueType().getSimpleVT().getSizeInBits() == 128 &&
"Expected an SSE type!");
- return DAG.getNode(ISD::BIT_CONVERT, dl, Op.getValueType(),
+ return DAG.getNode(ISD::BITCAST, dl, Op.getValueType(),
DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v4i32,AnyExt));
}
Result = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
// With PIC, the address is actually $g + Offset.
- if (OpFlag) {
+ if (OpFlag)
Result = DAG.getNode(ISD::ADD, DL, getPointerTy(),
DAG.getNode(X86ISD::GlobalBaseReg,
DebugLoc(), getPointerTy()),
Result);
- }
return Result;
}
SDValue *InFlag, const EVT PtrVT, unsigned ReturnReg,
unsigned char OperandFlags) {
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
DebugLoc dl = GA->getDebugLoc();
SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
GA->getValueType(0),
const EVT PtrVT, TLSModel::Model model,
bool is64Bit) {
DebugLoc dl = GA->getDebugLoc();
-
+
// Get the Thread Pointer, which is %gs:0 (32-bit) or %fs:0 (64-bit).
Value *Ptr = Constant::getNullValue(Type::getInt8PtrTy(*DAG.getContext(),
is64Bit ? 257 : 256));
- SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
+ SDValue ThreadPointer = DAG.getLoad(PtrVT, dl, DAG.getEntryNode(),
DAG.getIntPtrConstant(0),
MachinePointerInfo(Ptr), false, false, 0);
// emit "addl x@ntpoff,%eax" (local exec) or "addl x@indntpoff,%eax" (initial
// exec)
- SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
+ SDValue TGA = DAG.getTargetGlobalAddress(GA->getGlobal(), dl,
GA->getValueType(0),
GA->getOffset(), OperandFlags);
SDValue Offset = DAG.getNode(WrapperKind, dl, PtrVT, TGA);
SDValue
X86TargetLowering::LowerGlobalTLSAddress(SDValue Op, SelectionDAG &DAG) const {
-
+
GlobalAddressSDNode *GA = cast<GlobalAddressSDNode>(Op);
const GlobalValue *GV = GA->getGlobal();
if (Subtarget->isTargetELF()) {
// TODO: implement the "local dynamic" model
// TODO: implement the "initial exec"model for pic executables
-
+
// 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);
-
- TLSModel::Model model
+
+ TLSModel::Model model
= getTLSModel(GV, getTargetMachine().getRelocationModel());
-
+
switch (model) {
case TLSModel::GeneralDynamic:
case TLSModel::LocalDynamic: // not implemented
if (Subtarget->is64Bit())
return LowerToTLSGeneralDynamicModel64(GA, DAG, getPointerTy());
return LowerToTLSGeneralDynamicModel32(GA, DAG, getPointerTy());
-
+
case TLSModel::InitialExec:
case TLSModel::LocalExec:
return LowerToTLSExecModel(GA, DAG, getPointerTy(), model,
unsigned char OpFlag = 0;
unsigned WrapperKind = Subtarget->isPICStyleRIPRel() ?
X86ISD::WrapperRIP : X86ISD::Wrapper;
-
+
// In PIC mode (unless we're in RIPRel PIC mode) we add an offset to the
// global base reg.
bool PIC32 = (getTargetMachine().getRelocationModel() == Reloc::PIC_) &&
OpFlag = X86II::MO_TLVP_PIC_BASE;
else
OpFlag = X86II::MO_TLVP;
- DebugLoc DL = Op.getDebugLoc();
+ DebugLoc DL = Op.getDebugLoc();
SDValue Result = DAG.getTargetGlobalAddress(GA->getGlobal(), DL,
- getPointerTy(),
+ GA->getValueType(0),
GA->getOffset(), OpFlag);
SDValue Offset = DAG.getNode(WrapperKind, DL, getPointerTy(), Result);
-
+
// With PIC32, the address is actually $g + Offset.
if (PIC32)
Offset = DAG.getNode(ISD::ADD, DL, getPointerTy(),
DAG.getNode(X86ISD::GlobalBaseReg,
DebugLoc(), getPointerTy()),
Offset);
-
+
// Lowering the machine isd will make sure everything is in the right
// location.
- SDValue Args[] = { Offset };
- SDValue Chain = DAG.getNode(X86ISD::TLSCALL, DL, MVT::Other, Args, 1);
-
+ SDValue Chain = DAG.getEntryNode();
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
+ SDValue Args[] = { Chain, Offset };
+ Chain = DAG.getNode(X86ISD::TLSCALL, DL, NodeTys, Args, 2);
+
// TLSCALL will be codegen'ed as call. Inform MFI that function has calls.
MachineFrameInfo *MFI = DAG.getMachineFunction().getFrameInfo();
MFI->setAdjustsStack(true);
-
+
// And our return value (tls address) is in the standard call return value
// location.
unsigned Reg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
return DAG.getCopyFromReg(Chain, DL, Reg, getPointerTy());
}
-
+
assert(false &&
"TLS not implemented for this target.");
}
SDValue X86TargetLowering::BuildFILD(SDValue Op, EVT SrcVT, SDValue Chain,
- SDValue StackSlot,
+ SDValue StackSlot,
SelectionDAG &DAG) const {
// Build the FILD
DebugLoc DL = Op.getDebugLoc();
SDVTList Tys;
bool useSSE = isScalarFPTypeInSSEReg(Op.getValueType());
if (useSSE)
- Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Flag);
+ Tys = DAG.getVTList(MVT::f64, MVT::Other, MVT::Glue);
else
Tys = DAG.getVTList(Op.getValueType(), MVT::Other);
-
+
unsigned ByteSize = SrcVT.getSizeInBits()/8;
-
+
int SSFI = cast<FrameIndexSDNode>(StackSlot)->getIndex();
MachineMemOperand *MMO =
DAG.getMachineFunction()
.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
MachineMemOperand::MOLoad, ByteSize, ByteSize);
-
+
SDValue Ops[] = { Chain, StackSlot, DAG.getValueType(SrcVT) };
SDValue Result = DAG.getMemIntrinsicNode(useSSE ? X86ISD::FILD_FLAG :
X86ISD::FILD, DL,
DAG.getMachineFunction()
.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
MachineMemOperand::MOStore, SSFISize, SSFISize);
-
+
Chain = DAG.getMemIntrinsicNode(X86ISD::FST, DL, Tys,
Ops, array_lengthof(Ops),
Op.getValueType(), MMO);
MachinePointerInfo::getConstantPool(),
false, false, 16);
SDValue Unpck2 = getUnpackl(DAG, dl, MVT::v4i32, Unpck1, CLod0);
- SDValue XR2F = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Unpck2);
+ SDValue XR2F = DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Unpck2);
SDValue CLod1 = DAG.getLoad(MVT::v2f64, dl, CLod0.getValue(1), CPIdx1,
MachinePointerInfo::getConstantPool(),
false, false, 16);
DAG.getIntPtrConstant(0)));
Load = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Load),
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Load),
DAG.getIntPtrConstant(0));
// Or the load with the bias.
SDValue Or = DAG.getNode(ISD::OR, dl, MVT::v2i64,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
MVT::v2f64, Load)),
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
DAG.getNode(ISD::SCALAR_TO_VECTOR, dl,
MVT::v2f64, Bias)));
Or = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f64,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2f64, Or),
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2f64, Or),
DAG.getIntPtrConstant(0));
// Subtract the bias.
DAG.getMachineFunction()
.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
MachineMemOperand::MOLoad, 8, 8);
-
+
SDVTList Tys = DAG.getVTList(MVT::f80, MVT::Other);
SDValue Ops[] = { Store, StackSlot, DAG.getValueType(MVT::i64) };
SDValue Fild = DAG.getMemIntrinsicNode(X86ISD::FILD, dl, Tys, Ops, 3,
int SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
-
-
+
+
unsigned Opc;
switch (DstTy.getSimpleVT().SimpleTy) {
default: llvm_unreachable("Invalid FP_TO_SINT to lower!");
SDValue Ops[] = {
Chain, StackSlot, DAG.getValueType(TheVT)
};
-
+
MachineMemOperand *MMO =
MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
MachineMemOperand::MOLoad, MemSize, MemSize);
SSFI = MF.getFrameInfo()->CreateStackObject(MemSize, MemSize, false);
StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
}
-
+
MachineMemOperand *MMO =
MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
MachineMemOperand::MOStore, MemSize, MemSize);
MachinePointerInfo::getConstantPool(),
false, false, 16);
if (VT.isVector()) {
- return DAG.getNode(ISD::BIT_CONVERT, dl, VT,
+ return DAG.getNode(ISD::BITCAST, dl, VT,
DAG.getNode(ISD::XOR, dl, MVT::v2i64,
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64,
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2i64,
Op.getOperand(0)),
- DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v2i64, Mask)));
+ DAG.getNode(ISD::BITCAST, dl, MVT::v2i64, Mask)));
} else {
return DAG.getNode(X86ISD::FXOR, dl, VT, Op.getOperand(0), Mask);
}
SignBit = DAG.getNode(ISD::SCALAR_TO_VECTOR, dl, MVT::v2f64, SignBit);
SignBit = DAG.getNode(X86ISD::FSRL, dl, MVT::v2f64, SignBit,
DAG.getConstant(32, MVT::i32));
- SignBit = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, SignBit);
+ SignBit = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, SignBit);
SignBit = DAG.getNode(ISD::EXTRACT_VECTOR_ELT, dl, MVT::f32, SignBit,
DAG.getIntPtrConstant(0));
}
// Lower (X & (1 << N)) == 0 to BT(X, N).
// Lower ((X >>u N) & 1) != 0 to BT(X, N).
// Lower ((X >>s N) & 1) != 0 to BT(X, N).
- if (Op0.getOpcode() == ISD::AND &&
- Op0.hasOneUse() &&
+ if (Op0.getOpcode() == ISD::AND && Op0.hasOneUse() &&
Op1.getOpcode() == ISD::Constant &&
cast<ConstantSDNode>(Op1)->isNullValue() &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
return NewSetCC;
}
- // Look for "(setcc) == / != 1" to avoid unncessary setcc.
- if (Op0.getOpcode() == X86ISD::SETCC &&
- Op1.getOpcode() == ISD::Constant &&
+ // Look for X == 0, X == 1, X != 0, or X != 1. We can simplify some forms of
+ // these.
+ if (Op1.getOpcode() == ISD::Constant &&
(cast<ConstantSDNode>(Op1)->getZExtValue() == 1 ||
cast<ConstantSDNode>(Op1)->isNullValue()) &&
(CC == ISD::SETEQ || CC == ISD::SETNE)) {
- X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
- bool Invert = (CC == ISD::SETNE) ^
- cast<ConstantSDNode>(Op1)->isNullValue();
- if (Invert)
+
+ // If the input is a setcc, then reuse the input setcc or use a new one with
+ // the inverted condition.
+ if (Op0.getOpcode() == X86ISD::SETCC) {
+ X86::CondCode CCode = (X86::CondCode)Op0.getConstantOperandVal(0);
+ bool Invert = (CC == ISD::SETNE) ^
+ cast<ConstantSDNode>(Op1)->isNullValue();
+ if (!Invert) return Op0;
+
CCode = X86::GetOppositeBranchCondition(CCode);
- return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
+ DAG.getConstant(CCode, MVT::i8), Op0.getOperand(1));
+ }
}
bool isFP = Op1.getValueType().isFloatingPoint();
if (X86CC == X86::COND_INVALID)
return SDValue();
- SDValue Cond = EmitCmp(Op0, Op1, X86CC, DAG);
-
- // Use sbb x, x to materialize carry bit into a GPR.
- if (X86CC == X86::COND_B)
- return DAG.getNode(ISD::AND, dl, MVT::i8,
- DAG.getNode(X86ISD::SETCC_CARRY, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond),
- DAG.getConstant(1, MVT::i8));
-
+ SDValue EFLAGS = EmitCmp(Op0, Op1, X86CC, DAG);
return DAG.getNode(X86ISD::SETCC, dl, MVT::i8,
- DAG.getConstant(X86CC, MVT::i8), Cond);
+ DAG.getConstant(X86CC, MVT::i8), EFLAGS);
}
SDValue X86TargetLowering::LowerVSETCC(SDValue Op, SelectionDAG &DAG) const {
if (Op.getResNo() == 1 &&
(Opc == X86ISD::ADD ||
Opc == X86ISD::SUB ||
+ Opc == X86ISD::ADC ||
+ Opc == X86ISD::SBB ||
Opc == X86ISD::SMUL ||
Opc == X86ISD::UMUL ||
Opc == X86ISD::INC ||
Opc == X86ISD::AND))
return true;
+ if (Op.getResNo() == 2 && Opc == X86ISD::UMUL)
+ return true;
+
return false;
}
+static bool isZero(SDValue V) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
+ return C && C->isNullValue();
+}
+
+static bool isAllOnes(SDValue V) {
+ ConstantSDNode *C = dyn_cast<ConstantSDNode>(V);
+ return C && C->isAllOnesValue();
+}
+
SDValue X86TargetLowering::LowerSELECT(SDValue Op, SelectionDAG &DAG) const {
bool addTest = true;
SDValue Cond = Op.getOperand(0);
- DebugLoc dl = Op.getDebugLoc();
+ SDValue Op1 = Op.getOperand(1);
+ SDValue Op2 = Op.getOperand(2);
+ DebugLoc DL = Op.getDebugLoc();
SDValue CC;
if (Cond.getOpcode() == ISD::SETCC) {
Cond = NewCond;
}
- // (select (x == 0), -1, 0) -> (sign_bit (x - 1))
- SDValue Op1 = Op.getOperand(1);
- SDValue Op2 = Op.getOperand(2);
+ // (select (x == 0), -1, y) -> (sign_bit (x - 1)) | y
+ // (select (x == 0), y, -1) -> ~(sign_bit (x - 1)) | y
+ // (select (x != 0), y, -1) -> (sign_bit (x - 1)) | y
+ // (select (x != 0), -1, y) -> ~(sign_bit (x - 1)) | y
if (Cond.getOpcode() == X86ISD::SETCC &&
- cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue() == X86::COND_E) {
+ Cond.getOperand(1).getOpcode() == X86ISD::CMP &&
+ isZero(Cond.getOperand(1).getOperand(1))) {
SDValue Cmp = Cond.getOperand(1);
- if (Cmp.getOpcode() == X86ISD::CMP) {
- ConstantSDNode *N1C = dyn_cast<ConstantSDNode>(Op1);
+
+ unsigned CondCode =cast<ConstantSDNode>(Cond.getOperand(0))->getZExtValue();
+
+ if ((isAllOnes(Op1) || isAllOnes(Op2)) &&
+ (CondCode == X86::COND_E || CondCode == X86::COND_NE)) {
+ SDValue Y = isAllOnes(Op2) ? Op1 : Op2;
+
+ SDValue CmpOp0 = Cmp.getOperand(0);
+ Cmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32,
+ CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
+
+ SDValue Res = // Res = 0 or -1.
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
+ DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
+
+ if (isAllOnes(Op1) != (CondCode == X86::COND_E))
+ Res = DAG.getNOT(DL, Res, Res.getValueType());
+
ConstantSDNode *N2C = dyn_cast<ConstantSDNode>(Op2);
- ConstantSDNode *RHSC =
- dyn_cast<ConstantSDNode>(Cmp.getOperand(1).getNode());
- if (N1C && N1C->isAllOnesValue() &&
- N2C && N2C->isNullValue() &&
- RHSC && RHSC->isNullValue()) {
- SDValue CmpOp0 = Cmp.getOperand(0);
- Cmp = DAG.getNode(X86ISD::CMP, dl, MVT::i32,
- CmpOp0, DAG.getConstant(1, CmpOp0.getValueType()));
- return DAG.getNode(X86ISD::SETCC_CARRY, dl, Op.getValueType(),
- DAG.getConstant(X86::COND_B, MVT::i8), Cmp);
- }
+ if (N2C == 0 || !N2C->isNullValue())
+ Res = DAG.getNode(ISD::OR, DL, Res.getValueType(), Res, Y);
+ return Res;
}
}
- // Look pass (and (setcc_carry (cmp ...)), 1).
+ // Look past (and (setcc_carry (cmp ...)), 1).
if (Cond.getOpcode() == ISD::AND &&
Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
- if (C && C->getAPIntValue() == 1)
+ if (C && C->getAPIntValue() == 1)
Cond = Cond.getOperand(0);
}
// We know the result of AND is compared against zero. Try to match
// it to BT.
- if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
- SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
+ if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
+ SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, DL, DAG);
if (NewSetCC.getNode()) {
CC = NewSetCC.getOperand(0);
Cond = NewSetCC.getOperand(1);
Cond = EmitTest(Cond, X86::COND_NE, DAG);
}
+ // a < b ? -1 : 0 -> RES = ~setcc_carry
+ // a < b ? 0 : -1 -> RES = setcc_carry
+ // a >= b ? -1 : 0 -> RES = setcc_carry
+ // a >= b ? 0 : -1 -> RES = ~setcc_carry
+ if (Cond.getOpcode() == X86ISD::CMP) {
+ unsigned CondCode = cast<ConstantSDNode>(CC)->getZExtValue();
+
+ if ((CondCode == X86::COND_AE || CondCode == X86::COND_B) &&
+ (isAllOnes(Op1) || isAllOnes(Op2)) && (isZero(Op1) || isZero(Op2))) {
+ SDValue Res = DAG.getNode(X86ISD::SETCC_CARRY, DL, Op.getValueType(),
+ DAG.getConstant(X86::COND_B, MVT::i8), Cond);
+ if (isAllOnes(Op1) != (CondCode == X86::COND_B))
+ return DAG.getNOT(DL, Res, Res.getValueType());
+ return Res;
+ }
+ }
+
// X86ISD::CMOV means set the result (which is operand 1) to the RHS if
// condition is true.
- SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Flag);
+ SDVTList VTs = DAG.getVTList(Op.getValueType(), MVT::Glue);
SDValue Ops[] = { Op2, Op1, CC, Cond };
- return DAG.getNode(X86ISD::CMOV, dl, VTs, Ops, array_lengthof(Ops));
+ return DAG.getNode(X86ISD::CMOV, DL, VTs, Ops, array_lengthof(Ops));
}
// isAndOrOfSingleUseSetCCs - Return true if node is an ISD::AND or
if (Cond.getOpcode() == ISD::AND &&
Cond.getOperand(0).getOpcode() == X86ISD::SETCC_CARRY) {
ConstantSDNode *C = dyn_cast<ConstantSDNode>(Cond.getOperand(1));
- if (C && C->getAPIntValue() == 1)
+ if (C && C->getAPIntValue() == 1)
Cond = Cond.getOperand(0);
}
// We know the result of AND is compared against zero. Try to match
// it to BT.
- if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
+ if (Cond.getOpcode() == ISD::AND && Cond.hasOneUse()) {
SDValue NewSetCC = LowerToBT(Cond, ISD::SETNE, dl, DAG);
if (NewSetCC.getNode()) {
CC = NewSetCC.getOperand(0);
SDValue
X86TargetLowering::LowerDYNAMIC_STACKALLOC(SDValue Op,
SelectionDAG &DAG) const {
- assert(Subtarget->isTargetCygMing() &&
- "This should be used only on Cygwin/Mingw targets");
+ assert((Subtarget->isTargetCygMing() || Subtarget->isTargetWindows()) &&
+ "This should be used only on Windows targets");
DebugLoc dl = Op.getDebugLoc();
// Get the inputs.
Chain = DAG.getCopyToReg(Chain, dl, X86::EAX, Size, Flag);
Flag = Chain.getValue(1);
- SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList NodeTys = DAG.getVTList(MVT::Other, MVT::Glue);
- Chain = DAG.getNode(X86ISD::MINGW_ALLOCA, dl, NodeTys, Chain, Flag);
+ Chain = DAG.getNode(X86ISD::WIN_ALLOCA, dl, NodeTys, Chain, Flag);
Flag = Chain.getValue(1);
Chain = DAG.getCopyFromReg(Chain, dl, X86StackPtr, SPTy).getValue(1);
}
SDValue X86TargetLowering::LowerVAARG(SDValue Op, SelectionDAG &DAG) const {
- // X86-64 va_list is a struct { i32, i32, i8*, i8* }.
- assert(Subtarget->is64Bit() && "This code only handles 64-bit va_arg!");
+ assert(Subtarget->is64Bit() &&
+ "LowerVAARG only handles 64-bit va_arg!");
+ assert((Subtarget->isTargetLinux() ||
+ Subtarget->isTargetDarwin()) &&
+ "Unhandled target in LowerVAARG");
+ assert(Op.getNode()->getNumOperands() == 4);
+ SDValue Chain = Op.getOperand(0);
+ SDValue SrcPtr = Op.getOperand(1);
+ const Value *SV = cast<SrcValueSDNode>(Op.getOperand(2))->getValue();
+ unsigned Align = Op.getConstantOperandVal(3);
+ DebugLoc dl = Op.getDebugLoc();
- report_fatal_error("VAArgInst is not yet implemented for x86-64!");
- return SDValue();
+ EVT ArgVT = Op.getNode()->getValueType(0);
+ const Type *ArgTy = ArgVT.getTypeForEVT(*DAG.getContext());
+ uint32_t ArgSize = getTargetData()->getTypeAllocSize(ArgTy);
+ uint8_t ArgMode;
+
+ // Decide which area this value should be read from.
+ // TODO: Implement the AMD64 ABI in its entirety. This simple
+ // selection mechanism works only for the basic types.
+ if (ArgVT == MVT::f80) {
+ llvm_unreachable("va_arg for f80 not yet implemented");
+ } else if (ArgVT.isFloatingPoint() && ArgSize <= 16 /*bytes*/) {
+ ArgMode = 2; // Argument passed in XMM register. Use fp_offset.
+ } else if (ArgVT.isInteger() && ArgSize <= 32 /*bytes*/) {
+ ArgMode = 1; // Argument passed in GPR64 register(s). Use gp_offset.
+ } else {
+ llvm_unreachable("Unhandled argument type in LowerVAARG");
+ }
+
+ if (ArgMode == 2) {
+ // Sanity Check: Make sure using fp_offset makes sense.
+ assert(!UseSoftFloat &&
+ !(DAG.getMachineFunction()
+ .getFunction()->hasFnAttr(Attribute::NoImplicitFloat)) &&
+ Subtarget->hasXMM());
+ }
+
+ // Insert VAARG_64 node into the DAG
+ // VAARG_64 returns two values: Variable Argument Address, Chain
+ SmallVector<SDValue, 11> InstOps;
+ InstOps.push_back(Chain);
+ InstOps.push_back(SrcPtr);
+ InstOps.push_back(DAG.getConstant(ArgSize, MVT::i32));
+ InstOps.push_back(DAG.getConstant(ArgMode, MVT::i8));
+ InstOps.push_back(DAG.getConstant(Align, MVT::i32));
+ SDVTList VTs = DAG.getVTList(getPointerTy(), MVT::Other);
+ SDValue VAARG = DAG.getMemIntrinsicNode(X86ISD::VAARG_64, dl,
+ VTs, &InstOps[0], InstOps.size(),
+ MVT::i64,
+ MachinePointerInfo(SV),
+ /*Align=*/0,
+ /*Volatile=*/false,
+ /*ReadMem=*/true,
+ /*WriteMem=*/true);
+ Chain = VAARG.getValue(1);
+
+ // Load the next argument and return it
+ return DAG.getLoad(ArgVT, dl,
+ Chain,
+ VAARG,
+ MachinePointerInfo(),
+ false, false, 0);
}
SDValue X86TargetLowering::LowerVACOPY(SDValue Op, SelectionDAG &DAG) const {
return DAG.getMemcpy(Chain, DL, DstPtr, SrcPtr,
DAG.getIntPtrConstant(24), 8, /*isVolatile*/false,
- false,
+ false,
MachinePointerInfo(DstSV), MachinePointerInfo(SrcSV));
}
}
EVT VT = Op.getValueType();
- ShAmt = DAG.getNode(ISD::BIT_CONVERT, dl, VT, ShAmt);
+ ShAmt = DAG.getNode(ISD::BITCAST, dl, VT, ShAmt);
return DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(NewIntNo, MVT::i32),
Op.getOperand(1), ShAmt);
int SSFI = MF.getFrameInfo()->CreateStackObject(2, StackAlignment, false);
SDValue StackSlot = DAG.getFrameIndex(SSFI, getPointerTy());
-
+
MachineMemOperand *MMO =
MF.getMachineMemOperand(MachinePointerInfo::getFixedStack(SSFI),
MachineMemOperand::MOStore, 2, 2);
-
+
SDValue Ops[] = { DAG.getEntryNode(), StackSlot };
SDValue Chain = DAG.getMemIntrinsicNode(X86ISD::FNSTCW16m, DL,
DAG.getVTList(MVT::Other),
Op.getOperand(1), DAG.getConstant(23, MVT::i32));
ConstantInt *CI = ConstantInt::get(*Context, APInt(32, 0x3f800000U));
-
+
std::vector<Constant*> CV(4, CI);
Constant *C = ConstantVector::get(CV);
SDValue CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
false, false, 16);
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Addend);
- Op = DAG.getNode(ISD::BIT_CONVERT, dl, MVT::v4f32, Op);
+ Op = DAG.getNode(ISD::BITCAST, dl, MVT::v4f32, Op);
Op = DAG.getNode(ISD::FP_TO_SINT, dl, VT, Op);
return DAG.getNode(ISD::MUL, dl, VT, Op, R);
}
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(4, MVT::i32));
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
- R, M, Op);
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
-
+
C = ConstantVector::get(CVM2);
CPIdx = DAG.getConstantPool(C, getPointerTy(), 16);
M = DAG.getLoad(VT, dl, DAG.getEntryNode(), CPIdx,
MachinePointerInfo::getConstantPool(),
false, false, 16);
-
+
// r = pblendv(r, psllw(r & (char16)63, 2), a);
M = DAG.getNode(ISD::AND, dl, VT, R, M);
M = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
DAG.getConstant(Intrinsic::x86_sse2_pslli_w, MVT::i32), M,
DAG.getConstant(2, MVT::i32));
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
- R, M, Op);
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT, R, M, Op);
// a += a
Op = DAG.getNode(ISD::ADD, dl, VT, Op, Op);
-
+
// return pblendv(r, r+r, a);
- R = DAG.getNode(ISD::INTRINSIC_WO_CHAIN, dl, VT,
- DAG.getConstant(Intrinsic::x86_sse41_pblendvb, MVT::i32),
+ R = DAG.getNode(X86ISD::PBLENDVB, dl, VT,
R, DAG.getNode(ISD::ADD, dl, VT, R, R), Op);
return R;
}
SDValue RHS = N->getOperand(1);
unsigned BaseOp = 0;
unsigned Cond = 0;
- DebugLoc dl = Op.getDebugLoc();
-
+ DebugLoc DL = Op.getDebugLoc();
switch (Op.getOpcode()) {
default: llvm_unreachable("Unknown ovf instruction!");
case ISD::SADDO:
BaseOp = X86ISD::SMUL;
Cond = X86::COND_O;
break;
- case ISD::UMULO:
- BaseOp = X86ISD::UMUL;
- Cond = X86::COND_B;
- break;
+ case ISD::UMULO: { // i64, i8 = umulo lhs, rhs --> i64, i64, i32 umul lhs,rhs
+ SDVTList VTs = DAG.getVTList(N->getValueType(0), N->getValueType(0),
+ MVT::i32);
+ SDValue Sum = DAG.getNode(X86ISD::UMUL, DL, VTs, LHS, RHS);
+
+ SDValue SetCC =
+ DAG.getNode(X86ISD::SETCC, DL, MVT::i8,
+ DAG.getConstant(X86::COND_O, MVT::i32),
+ SDValue(Sum.getNode(), 2));
+
+ DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
+ return Sum;
+ }
}
// Also sets EFLAGS.
SDVTList VTs = DAG.getVTList(N->getValueType(0), MVT::i32);
- SDValue Sum = DAG.getNode(BaseOp, dl, VTs, LHS, RHS);
+ SDValue Sum = DAG.getNode(BaseOp, DL, VTs, LHS, RHS);
SDValue SetCC =
- DAG.getNode(X86ISD::SETCC, dl, N->getValueType(1),
- DAG.getConstant(Cond, MVT::i32), SDValue(Sum.getNode(), 1));
+ DAG.getNode(X86ISD::SETCC, DL, N->getValueType(1),
+ DAG.getConstant(Cond, MVT::i32),
+ SDValue(Sum.getNode(), 1));
DAG.ReplaceAllUsesOfValueWith(SDValue(N, 1), SetCC);
return Sum;
SDValue X86TargetLowering::LowerMEMBARRIER(SDValue Op, SelectionDAG &DAG) const{
DebugLoc dl = Op.getDebugLoc();
-
+
if (!Subtarget->hasSSE2()) {
SDValue Chain = Op.getOperand(0);
- SDValue Zero = DAG.getConstant(0,
+ SDValue Zero = DAG.getConstant(0,
Subtarget->is64Bit() ? MVT::i64 : MVT::i32);
SDValue Ops[] = {
DAG.getRegister(X86::ESP, MVT::i32), // Base
Zero,
Chain
};
- SDNode *Res =
+ SDNode *Res =
DAG.getMachineNode(X86::OR32mrLocked, dl, MVT::Other, Ops,
array_lengthof(Ops));
return SDValue(Res, 0);
}
-
+
unsigned isDev = cast<ConstantSDNode>(Op.getOperand(5))->getZExtValue();
if (!isDev)
return DAG.getNode(X86ISD::MEMBARRIER, dl, MVT::Other, Op.getOperand(0));
-
+
unsigned Op1 = cast<ConstantSDNode>(Op.getOperand(1))->getZExtValue();
unsigned Op2 = cast<ConstantSDNode>(Op.getOperand(2))->getZExtValue();
unsigned Op3 = cast<ConstantSDNode>(Op.getOperand(3))->getZExtValue();
unsigned Op4 = cast<ConstantSDNode>(Op.getOperand(4))->getZExtValue();
-
+
// def : Pat<(membarrier (i8 0), (i8 0), (i8 0), (i8 1), (i8 1)), (SFENCE)>;
if (!Op1 && !Op2 && !Op3 && Op4)
return DAG.getNode(X86ISD::SFENCE, dl, MVT::Other, Op.getOperand(0));
-
+
// def : Pat<(membarrier (i8 1), (i8 0), (i8 0), (i8 0), (i8 1)), (LFENCE)>;
if (Op1 && !Op2 && !Op3 && !Op4)
return DAG.getNode(X86ISD::LFENCE, dl, MVT::Other, Op.getOperand(0));
-
- // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
+
+ // def : Pat<(membarrier (i8 imm), (i8 imm), (i8 imm), (i8 imm), (i8 1)),
// (MFENCE)>;
return DAG.getNode(X86ISD::MFENCE, dl, MVT::Other, Op.getOperand(0));
}
Op.getOperand(3),
DAG.getTargetConstant(size, MVT::i8),
cpIn.getValue(1) };
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
MachineMemOperand *MMO = cast<AtomicSDNode>(Op)->getMemOperand();
SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG_DAG, DL, Tys,
Ops, 5, T, MMO);
SDValue X86TargetLowering::LowerREADCYCLECOUNTER(SDValue Op,
SelectionDAG &DAG) const {
assert(Subtarget->is64Bit() && "Result not type legalized?");
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
SDValue TheChain = Op.getOperand(0);
DebugLoc dl = Op.getDebugLoc();
SDValue rd = DAG.getNode(X86ISD::RDTSC_DAG, dl, Tys, &TheChain, 1);
return DAG.getMergeValues(Ops, 2, dl);
}
-SDValue X86TargetLowering::LowerBIT_CONVERT(SDValue Op,
+SDValue X86TargetLowering::LowerBITCAST(SDValue Op,
SelectionDAG &DAG) const {
EVT SrcVT = Op.getOperand(0).getValueType();
EVT DstVT = Op.getValueType();
- assert((Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
- Subtarget->hasMMX() && !DisableMMX) &&
- "Unexpected custom BIT_CONVERT");
- assert((DstVT == MVT::i64 ||
+ assert(Subtarget->is64Bit() && !Subtarget->hasSSE2() &&
+ Subtarget->hasMMX() && "Unexpected custom BITCAST");
+ assert((DstVT == MVT::i64 ||
(DstVT.isVector() && DstVT.getSizeInBits()==64)) &&
- "Unexpected custom BIT_CONVERT");
+ "Unexpected custom BITCAST");
// i64 <=> MMX conversions are Legal.
if (SrcVT==MVT::i64 && DstVT.isVector())
return Op;
// All other conversions need to be expanded.
return SDValue();
}
+
SDValue X86TargetLowering::LowerLOAD_SUB(SDValue Op, SelectionDAG &DAG) const {
SDNode *Node = Op.getNode();
DebugLoc dl = Node->getDebugLoc();
cast<AtomicSDNode>(Node)->getAlignment());
}
+static SDValue LowerADDC_ADDE_SUBC_SUBE(SDValue Op, SelectionDAG &DAG) {
+ EVT VT = Op.getNode()->getValueType(0);
+
+ // Let legalize expand this if it isn't a legal type yet.
+ if (!DAG.getTargetLoweringInfo().isTypeLegal(VT))
+ return SDValue();
+
+ SDVTList VTs = DAG.getVTList(VT, MVT::i32);
+
+ unsigned Opc;
+ bool ExtraOp = false;
+ switch (Op.getOpcode()) {
+ default: assert(0 && "Invalid code");
+ case ISD::ADDC: Opc = X86ISD::ADD; break;
+ case ISD::ADDE: Opc = X86ISD::ADC; ExtraOp = true; break;
+ case ISD::SUBC: Opc = X86ISD::SUB; break;
+ case ISD::SUBE: Opc = X86ISD::SBB; ExtraOp = true; break;
+ }
+
+ if (!ExtraOp)
+ return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
+ Op.getOperand(1));
+ return DAG.getNode(Opc, Op->getDebugLoc(), VTs, Op.getOperand(0),
+ Op.getOperand(1), Op.getOperand(2));
+}
+
/// LowerOperation - Provide custom lowering hooks for some operations.
///
SDValue X86TargetLowering::LowerOperation(SDValue Op, SelectionDAG &DAG) const {
case ISD::SMULO:
case ISD::UMULO: return LowerXALUO(Op, DAG);
case ISD::READCYCLECOUNTER: return LowerREADCYCLECOUNTER(Op, DAG);
- case ISD::BIT_CONVERT: return LowerBIT_CONVERT(Op, DAG);
+ case ISD::BITCAST: return LowerBITCAST(Op, DAG);
+ case ISD::ADDC:
+ case ISD::ADDE:
+ case ISD::SUBC:
+ case ISD::SUBE: return LowerADDC_ADDE_SUBC_SUBE(Op, DAG);
}
}
default:
assert(false && "Do not know how to custom type legalize this operation!");
return;
+ case ISD::ADDC:
+ case ISD::ADDE:
+ case ISD::SUBC:
+ case ISD::SUBE:
+ // We don't want to expand or promote these.
+ return;
case ISD::FP_TO_SINT: {
std::pair<SDValue,SDValue> Vals =
FP_TO_INTHelper(SDValue(N, 0), DAG, true);
return;
}
case ISD::READCYCLECOUNTER: {
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
+ 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,
SDValue Ops[] = { swapInH.getValue(0),
N->getOperand(1),
swapInH.getValue(1) };
- SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Flag);
- SDValue Result = DAG.getNode(X86ISD::LCMPXCHG8_DAG, dl, Tys, Ops, 3);
+ SDVTList Tys = DAG.getVTList(MVT::Other, MVT::Glue);
+ MachineMemOperand *MMO = cast<AtomicSDNode>(N)->getMemOperand();
+ SDValue Result = DAG.getMemIntrinsicNode(X86ISD::LCMPXCHG8_DAG, dl, Tys,
+ Ops, 3, T, MMO);
SDValue cpOutL = DAG.getCopyFromReg(Result.getValue(0), dl, X86::EAX,
MVT::i32, Result.getValue(1));
SDValue cpOutH = DAG.getCopyFromReg(cpOutL.getValue(1), dl, X86::EDX,
case X86ISD::PINSRB: return "X86ISD::PINSRB";
case X86ISD::PINSRW: return "X86ISD::PINSRW";
case X86ISD::PSHUFB: return "X86ISD::PSHUFB";
+ case X86ISD::PANDN: return "X86ISD::PANDN";
+ case X86ISD::PSIGNB: return "X86ISD::PSIGNB";
+ case X86ISD::PSIGNW: return "X86ISD::PSIGNW";
+ case X86ISD::PSIGND: return "X86ISD::PSIGND";
+ case X86ISD::PBLENDVB: return "X86ISD::PBLENDVB";
case X86ISD::FMAX: return "X86ISD::FMAX";
case X86ISD::FMIN: return "X86ISD::FMIN";
case X86ISD::FRSQRT: return "X86ISD::FRSQRT";
case X86ISD::PCMPGTQ: return "X86ISD::PCMPGTQ";
case X86ISD::ADD: return "X86ISD::ADD";
case X86ISD::SUB: return "X86ISD::SUB";
+ case X86ISD::ADC: return "X86ISD::ADC";
+ case X86ISD::SBB: return "X86ISD::SBB";
case X86ISD::SMUL: return "X86ISD::SMUL";
case X86ISD::UMUL: return "X86ISD::UMUL";
case X86ISD::INC: return "X86ISD::INC";
case X86ISD::PUNPCKHDQ: return "X86ISD::PUNPCKHDQ";
case X86ISD::PUNPCKHQDQ: return "X86ISD::PUNPCKHQDQ";
case X86ISD::VASTART_SAVE_XMM_REGS: return "X86ISD::VASTART_SAVE_XMM_REGS";
- case X86ISD::MINGW_ALLOCA: return "X86ISD::MINGW_ALLOCA";
+ case X86ISD::VAARG_64: return "X86ISD::VAARG_64";
+ case X86ISD::WIN_ALLOCA: return "X86ISD::WIN_ALLOCA";
}
}
MachineBasicBlock *
X86TargetLowering::EmitPCMP(MachineInstr *MI, MachineBasicBlock *BB,
unsigned numArgs, bool memArg) const {
-
assert((Subtarget->hasSSE42() || Subtarget->hasAVX()) &&
"Target must have SSE4.2 or AVX features enabled");
DebugLoc dl = MI->getDebugLoc();
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
-
unsigned Opc;
-
if (!Subtarget->hasAVX()) {
if (memArg)
Opc = numArgs == 3 ? X86::PCMPISTRM128rm : X86::PCMPESTRM128rm;
Opc = numArgs == 3 ? X86::VPCMPISTRM128rr : X86::VPCMPESTRM128rr;
}
- MachineInstrBuilder MIB = BuildMI(BB, dl, TII->get(Opc));
-
+ MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(Opc));
for (unsigned i = 0; i < numArgs; ++i) {
MachineOperand &Op = MI->getOperand(i+1);
-
if (!(Op.isReg() && Op.isImplicit()))
MIB.addOperand(Op);
}
-
- BuildMI(BB, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
+ BuildMI(*BB, MI, dl, TII->get(X86::MOVAPSrr), MI->getOperand(0).getReg())
.addReg(X86::XMM0);
MI->eraseFromParent();
+ return BB;
+}
+
+MachineBasicBlock *
+X86TargetLowering::EmitMonitor(MachineInstr *MI, MachineBasicBlock *BB) const {
+ DebugLoc dl = MI->getDebugLoc();
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+
+ // Address into RAX/EAX, other two args into ECX, EDX.
+ unsigned MemOpc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
+ unsigned MemReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
+ MachineInstrBuilder MIB = BuildMI(*BB, MI, dl, TII->get(MemOpc), MemReg);
+ for (int i = 0; i < X86::AddrNumOperands; ++i)
+ MIB.addOperand(MI->getOperand(i));
+
+ unsigned ValOps = X86::AddrNumOperands;
+ BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
+ .addReg(MI->getOperand(ValOps).getReg());
+ BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EDX)
+ .addReg(MI->getOperand(ValOps+1).getReg());
+
+ // The instruction doesn't actually take any operands though.
+ BuildMI(*BB, MI, dl, TII->get(X86::MONITORrrr));
+
+ MI->eraseFromParent(); // The pseudo is gone now.
+ return BB;
+}
+MachineBasicBlock *
+X86TargetLowering::EmitMwait(MachineInstr *MI, MachineBasicBlock *BB) const {
+ DebugLoc dl = MI->getDebugLoc();
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+
+ // First arg in ECX, the second in EAX.
+ BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::ECX)
+ .addReg(MI->getOperand(0).getReg());
+ BuildMI(*BB, MI, dl, TII->get(TargetOpcode::COPY), X86::EAX)
+ .addReg(MI->getOperand(1).getReg());
+
+ // The instruction doesn't actually take any operands though.
+ BuildMI(*BB, MI, dl, TII->get(X86::MWAITrr));
+
+ MI->eraseFromParent(); // The pseudo is gone now.
return BB;
}
+MachineBasicBlock *
+X86TargetLowering::EmitVAARG64WithCustomInserter(
+ MachineInstr *MI,
+ MachineBasicBlock *MBB) const {
+ // Emit va_arg instruction on X86-64.
+
+ // Operands to this pseudo-instruction:
+ // 0 ) Output : destination address (reg)
+ // 1-5) Input : va_list address (addr, i64mem)
+ // 6 ) ArgSize : Size (in bytes) of vararg type
+ // 7 ) ArgMode : 0=overflow only, 1=use gp_offset, 2=use fp_offset
+ // 8 ) Align : Alignment of type
+ // 9 ) EFLAGS (implicit-def)
+
+ assert(MI->getNumOperands() == 10 && "VAARG_64 should have 10 operands!");
+ assert(X86::AddrNumOperands == 5 && "VAARG_64 assumes 5 address operands");
+
+ unsigned DestReg = MI->getOperand(0).getReg();
+ MachineOperand &Base = MI->getOperand(1);
+ MachineOperand &Scale = MI->getOperand(2);
+ MachineOperand &Index = MI->getOperand(3);
+ MachineOperand &Disp = MI->getOperand(4);
+ MachineOperand &Segment = MI->getOperand(5);
+ unsigned ArgSize = MI->getOperand(6).getImm();
+ unsigned ArgMode = MI->getOperand(7).getImm();
+ unsigned Align = MI->getOperand(8).getImm();
+
+ // Memory Reference
+ assert(MI->hasOneMemOperand() && "Expected VAARG_64 to have one memoperand");
+ MachineInstr::mmo_iterator MMOBegin = MI->memoperands_begin();
+ MachineInstr::mmo_iterator MMOEnd = MI->memoperands_end();
+
+ // Machine Information
+ const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
+ MachineRegisterInfo &MRI = MBB->getParent()->getRegInfo();
+ const TargetRegisterClass *AddrRegClass = getRegClassFor(MVT::i64);
+ const TargetRegisterClass *OffsetRegClass = getRegClassFor(MVT::i32);
+ DebugLoc DL = MI->getDebugLoc();
+
+ // struct va_list {
+ // i32 gp_offset
+ // i32 fp_offset
+ // i64 overflow_area (address)
+ // i64 reg_save_area (address)
+ // }
+ // sizeof(va_list) = 24
+ // alignment(va_list) = 8
+
+ unsigned TotalNumIntRegs = 6;
+ unsigned TotalNumXMMRegs = 8;
+ bool UseGPOffset = (ArgMode == 1);
+ bool UseFPOffset = (ArgMode == 2);
+ unsigned MaxOffset = TotalNumIntRegs * 8 +
+ (UseFPOffset ? TotalNumXMMRegs * 16 : 0);
+
+ /* Align ArgSize to a multiple of 8 */
+ unsigned ArgSizeA8 = (ArgSize + 7) & ~7;
+ bool NeedsAlign = (Align > 8);
+
+ MachineBasicBlock *thisMBB = MBB;
+ MachineBasicBlock *overflowMBB;
+ MachineBasicBlock *offsetMBB;
+ MachineBasicBlock *endMBB;
+
+ unsigned OffsetDestReg = 0; // Argument address computed by offsetMBB
+ unsigned OverflowDestReg = 0; // Argument address computed by overflowMBB
+ unsigned OffsetReg = 0;
+
+ if (!UseGPOffset && !UseFPOffset) {
+ // If we only pull from the overflow region, we don't create a branch.
+ // We don't need to alter control flow.
+ OffsetDestReg = 0; // unused
+ OverflowDestReg = DestReg;
+
+ offsetMBB = NULL;
+ overflowMBB = thisMBB;
+ endMBB = thisMBB;
+ } else {
+ // First emit code to check if gp_offset (or fp_offset) is below the bound.
+ // If so, pull the argument from reg_save_area. (branch to offsetMBB)
+ // If not, pull from overflow_area. (branch to overflowMBB)
+ //
+ // thisMBB
+ // | .
+ // | .
+ // offsetMBB overflowMBB
+ // | .
+ // | .
+ // endMBB
+
+ // Registers for the PHI in endMBB
+ OffsetDestReg = MRI.createVirtualRegister(AddrRegClass);
+ OverflowDestReg = MRI.createVirtualRegister(AddrRegClass);
+
+ const BasicBlock *LLVM_BB = MBB->getBasicBlock();
+ MachineFunction *MF = MBB->getParent();
+ overflowMBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ offsetMBB = MF->CreateMachineBasicBlock(LLVM_BB);
+ endMBB = MF->CreateMachineBasicBlock(LLVM_BB);
+
+ MachineFunction::iterator MBBIter = MBB;
+ ++MBBIter;
+
+ // Insert the new basic blocks
+ MF->insert(MBBIter, offsetMBB);
+ MF->insert(MBBIter, overflowMBB);
+ MF->insert(MBBIter, endMBB);
+
+ // Transfer the remainder of MBB and its successor edges to endMBB.
+ endMBB->splice(endMBB->begin(), thisMBB,
+ llvm::next(MachineBasicBlock::iterator(MI)),
+ thisMBB->end());
+ endMBB->transferSuccessorsAndUpdatePHIs(thisMBB);
+
+ // Make offsetMBB and overflowMBB successors of thisMBB
+ thisMBB->addSuccessor(offsetMBB);
+ thisMBB->addSuccessor(overflowMBB);
+
+ // endMBB is a successor of both offsetMBB and overflowMBB
+ offsetMBB->addSuccessor(endMBB);
+ overflowMBB->addSuccessor(endMBB);
+
+ // Load the offset value into a register
+ OffsetReg = MRI.createVirtualRegister(OffsetRegClass);
+ BuildMI(thisMBB, DL, TII->get(X86::MOV32rm), OffsetReg)
+ .addOperand(Base)
+ .addOperand(Scale)
+ .addOperand(Index)
+ .addDisp(Disp, UseFPOffset ? 4 : 0)
+ .addOperand(Segment)
+ .setMemRefs(MMOBegin, MMOEnd);
+
+ // Check if there is enough room left to pull this argument.
+ BuildMI(thisMBB, DL, TII->get(X86::CMP32ri))
+ .addReg(OffsetReg)
+ .addImm(MaxOffset + 8 - ArgSizeA8);
+
+ // Branch to "overflowMBB" if offset >= max
+ // Fall through to "offsetMBB" otherwise
+ BuildMI(thisMBB, DL, TII->get(X86::GetCondBranchFromCond(X86::COND_AE)))
+ .addMBB(overflowMBB);
+ }
+
+ // In offsetMBB, emit code to use the reg_save_area.
+ if (offsetMBB) {
+ assert(OffsetReg != 0);
+
+ // Read the reg_save_area address.
+ unsigned RegSaveReg = MRI.createVirtualRegister(AddrRegClass);
+ BuildMI(offsetMBB, DL, TII->get(X86::MOV64rm), RegSaveReg)
+ .addOperand(Base)
+ .addOperand(Scale)
+ .addOperand(Index)
+ .addDisp(Disp, 16)
+ .addOperand(Segment)
+ .setMemRefs(MMOBegin, MMOEnd);
+
+ // Zero-extend the offset
+ unsigned OffsetReg64 = MRI.createVirtualRegister(AddrRegClass);
+ BuildMI(offsetMBB, DL, TII->get(X86::SUBREG_TO_REG), OffsetReg64)
+ .addImm(0)
+ .addReg(OffsetReg)
+ .addImm(X86::sub_32bit);
+
+ // Add the offset to the reg_save_area to get the final address.
+ BuildMI(offsetMBB, DL, TII->get(X86::ADD64rr), OffsetDestReg)
+ .addReg(OffsetReg64)
+ .addReg(RegSaveReg);
+
+ // Compute the offset for the next argument
+ unsigned NextOffsetReg = MRI.createVirtualRegister(OffsetRegClass);
+ BuildMI(offsetMBB, DL, TII->get(X86::ADD32ri), NextOffsetReg)
+ .addReg(OffsetReg)
+ .addImm(UseFPOffset ? 16 : 8);
+
+ // Store it back into the va_list.
+ BuildMI(offsetMBB, DL, TII->get(X86::MOV32mr))
+ .addOperand(Base)
+ .addOperand(Scale)
+ .addOperand(Index)
+ .addDisp(Disp, UseFPOffset ? 4 : 0)
+ .addOperand(Segment)
+ .addReg(NextOffsetReg)
+ .setMemRefs(MMOBegin, MMOEnd);
+
+ // Jump to endMBB
+ BuildMI(offsetMBB, DL, TII->get(X86::JMP_4))
+ .addMBB(endMBB);
+ }
+
+ //
+ // Emit code to use overflow area
+ //
+
+ // Load the overflow_area address into a register.
+ unsigned OverflowAddrReg = MRI.createVirtualRegister(AddrRegClass);
+ BuildMI(overflowMBB, DL, TII->get(X86::MOV64rm), OverflowAddrReg)
+ .addOperand(Base)
+ .addOperand(Scale)
+ .addOperand(Index)
+ .addDisp(Disp, 8)
+ .addOperand(Segment)
+ .setMemRefs(MMOBegin, MMOEnd);
+
+ // If we need to align it, do so. Otherwise, just copy the address
+ // to OverflowDestReg.
+ if (NeedsAlign) {
+ // Align the overflow address
+ assert((Align & (Align-1)) == 0 && "Alignment must be a power of 2");
+ unsigned TmpReg = MRI.createVirtualRegister(AddrRegClass);
+
+ // aligned_addr = (addr + (align-1)) & ~(align-1)
+ BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), TmpReg)
+ .addReg(OverflowAddrReg)
+ .addImm(Align-1);
+
+ BuildMI(overflowMBB, DL, TII->get(X86::AND64ri32), OverflowDestReg)
+ .addReg(TmpReg)
+ .addImm(~(uint64_t)(Align-1));
+ } else {
+ BuildMI(overflowMBB, DL, TII->get(TargetOpcode::COPY), OverflowDestReg)
+ .addReg(OverflowAddrReg);
+ }
+
+ // Compute the next overflow address after this argument.
+ // (the overflow address should be kept 8-byte aligned)
+ unsigned NextAddrReg = MRI.createVirtualRegister(AddrRegClass);
+ BuildMI(overflowMBB, DL, TII->get(X86::ADD64ri32), NextAddrReg)
+ .addReg(OverflowDestReg)
+ .addImm(ArgSizeA8);
+
+ // Store the new overflow address.
+ BuildMI(overflowMBB, DL, TII->get(X86::MOV64mr))
+ .addOperand(Base)
+ .addOperand(Scale)
+ .addOperand(Index)
+ .addDisp(Disp, 8)
+ .addOperand(Segment)
+ .addReg(NextAddrReg)
+ .setMemRefs(MMOBegin, MMOEnd);
+
+ // If we branched, emit the PHI to the front of endMBB.
+ if (offsetMBB) {
+ BuildMI(*endMBB, endMBB->begin(), DL,
+ TII->get(X86::PHI), DestReg)
+ .addReg(OffsetDestReg).addMBB(offsetMBB)
+ .addReg(OverflowDestReg).addMBB(overflowMBB);
+ }
+
+ // Erase the pseudo instruction
+ MI->eraseFromParent();
+
+ return endMBB;
+}
+
MachineBasicBlock *
X86TargetLowering::EmitVAStartSaveXMMRegsWithCustomInserter(
MachineInstr *MI,
}
MachineBasicBlock *
-X86TargetLowering::EmitLoweredMingwAlloca(MachineInstr *MI,
+X86TargetLowering::EmitLoweredWinAlloca(MachineInstr *MI,
MachineBasicBlock *BB) const {
const TargetInstrInfo *TII = getTargetMachine().getInstrInfo();
DebugLoc DL = MI->getDebugLoc();
// FIXME: The code should be tweaked as soon as we'll try to do codegen for
// mingw-w64.
+ const char *StackProbeSymbol =
+ Subtarget->isTargetWindows() ? "_chkstk" : "_alloca";
+
BuildMI(*BB, MI, DL, TII->get(X86::CALLpcrel32))
- .addExternalSymbol("_alloca")
+ .addExternalSymbol(StackProbeSymbol)
.addReg(X86::EAX, RegState::Implicit)
.addReg(X86::ESP, RegState::Implicit)
.addReg(X86::EAX, RegState::Define | RegState::Implicit)
// our load from the relocation, sticking it in either RDI (x86-64)
// or EAX and doing an indirect call. The return value will then
// be in the normal return register.
- const X86InstrInfo *TII
+ const X86InstrInfo *TII
= static_cast<const X86InstrInfo*>(getTargetMachine().getInstrInfo());
DebugLoc DL = MI->getDebugLoc();
MachineFunction *F = BB->getParent();
assert(Subtarget->isTargetDarwin() && "Darwin only instr emitted?");
assert(MI->getOperand(3).isGlobal() && "This should be a global");
-
+
if (Subtarget->is64Bit()) {
MachineInstrBuilder MIB = BuildMI(*BB, MI, DL,
TII->get(X86::MOV64rm), X86::RDI)
.addReg(X86::RIP)
.addImm(0).addReg(0)
- .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
+ .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
MI->getOperand(3).getTargetFlags())
.addReg(0);
MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL64m));
TII->get(X86::MOV32rm), X86::EAX)
.addReg(0)
.addImm(0).addReg(0)
- .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
+ .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
MI->getOperand(3).getTargetFlags())
.addReg(0);
MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
TII->get(X86::MOV32rm), X86::EAX)
.addReg(TII->getGlobalBaseReg(F))
.addImm(0).addReg(0)
- .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
+ .addGlobalAddress(MI->getOperand(3).getGlobal(), 0,
MI->getOperand(3).getTargetFlags())
.addReg(0);
MIB = BuildMI(*BB, MI, DL, TII->get(X86::CALL32m));
addDirectMem(MIB, X86::EAX);
}
-
+
MI->eraseFromParent(); // The pseudo instruction is gone now.
return BB;
}
MachineBasicBlock *BB) const {
switch (MI->getOpcode()) {
default: assert(false && "Unexpected instr type to insert");
- case X86::MINGW_ALLOCA:
- return EmitLoweredMingwAlloca(MI, BB);
+ case X86::WIN_ALLOCA:
+ return EmitLoweredWinAlloca(MI, BB);
case X86::TLSCall_32:
case X86::TLSCall_64:
return EmitLoweredTLSCall(MI, BB);
case X86::VPCMPESTRM128MEM:
return EmitPCMP(MI, BB, 5, true /* in mem */);
+ // Thread synchronization.
+ case X86::MONITOR:
+ return EmitMonitor(MI, BB);
+ case X86::MWAIT:
+ return EmitMwait(MI, BB);
+
// Atomic Lowering.
case X86::ATOMAND32:
return EmitAtomicBitwiseWithCustomInserter(MI, BB, X86::AND32rr,
false);
case X86::VASTART_SAVE_XMM_REGS:
return EmitVAStartSaveXMMRegsWithCustomInserter(MI, BB);
+
+ case X86::VAARG_64:
+ return EmitVAARG64WithCustomInserter(MI, BB);
}
}
default: break;
case X86ISD::ADD:
case X86ISD::SUB:
+ case X86ISD::ADC:
+ case X86ISD::SBB:
case X86ISD::SMUL:
case X86ISD::UMUL:
case X86ISD::INC:
// SETCC_CARRY sets the dest to ~0 for true or 0 for false.
if (Op.getOpcode() == X86ISD::SETCC_CARRY)
return Op.getValueType().getScalarType().getSizeInBits();
-
+
// Fallback case.
return 1;
}
/// if the load addresses are consecutive, non-overlapping, and in the right
/// order.
static SDValue PerformShuffleCombine(SDNode *N, SelectionDAG &DAG,
- const TargetLowering &TLI) {
+ TargetLowering::DAGCombinerInfo &DCI) {
DebugLoc dl = N->getDebugLoc();
EVT VT = N->getValueType(0);
if (VT.getSizeInBits() != 128)
return SDValue();
+ // Don't create instructions with illegal types after legalize types has run.
+ const TargetLowering &TLI = DAG.getTargetLoweringInfo();
+ if (!DCI.isBeforeLegalize() && !TLI.isTypeLegal(VT.getVectorElementType()))
+ return SDValue();
+
SmallVector<SDValue, 16> Elts;
for (unsigned i = 0, e = VT.getVectorNumElements(); i != e; ++i)
Elts.push_back(getShuffleScalarElt(N, i, DAG, 0));
return SDValue();
}
+
+static SDValue PerformAndCombine(SDNode *N, SelectionDAG &DAG,
+ TargetLowering::DAGCombinerInfo &DCI,
+ const X86Subtarget *Subtarget) {
+ if (DCI.isBeforeLegalizeOps())
+ return SDValue();
+
+ // Want to form PANDN nodes, in the hopes of then easily combining them with
+ // OR and AND nodes to form PBLEND/PSIGN.
+ EVT VT = N->getValueType(0);
+ if (VT != MVT::v2i64)
+ return SDValue();
+
+ SDValue N0 = N->getOperand(0);
+ SDValue N1 = N->getOperand(1);
+ DebugLoc DL = N->getDebugLoc();
+
+ // Check LHS for vnot
+ if (N0.getOpcode() == ISD::XOR &&
+ ISD::isBuildVectorAllOnes(N0.getOperand(1).getNode()))
+ return DAG.getNode(X86ISD::PANDN, DL, VT, N0.getOperand(0), N1);
+
+ // Check RHS for vnot
+ if (N1.getOpcode() == ISD::XOR &&
+ ISD::isBuildVectorAllOnes(N1.getOperand(1).getNode()))
+ return DAG.getNode(X86ISD::PANDN, DL, VT, N1.getOperand(0), N0);
+
+ return SDValue();
+}
+
static SDValue PerformOrCombine(SDNode *N, SelectionDAG &DAG,
TargetLowering::DAGCombinerInfo &DCI,
const X86Subtarget *Subtarget) {
return SDValue();
EVT VT = N->getValueType(0);
- if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64)
+ if (VT != MVT::i16 && VT != MVT::i32 && VT != MVT::i64 && VT != MVT::v2i64)
return SDValue();
- // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
SDValue N0 = N->getOperand(0);
SDValue N1 = N->getOperand(1);
+
+ // look for psign/blend
+ if (Subtarget->hasSSSE3()) {
+ if (VT == MVT::v2i64) {
+ // Canonicalize pandn to RHS
+ if (N0.getOpcode() == X86ISD::PANDN)
+ std::swap(N0, N1);
+ // or (and (m, x), (pandn m, y))
+ if (N0.getOpcode() == ISD::AND && N1.getOpcode() == X86ISD::PANDN) {
+ SDValue Mask = N1.getOperand(0);
+ SDValue X = N1.getOperand(1);
+ SDValue Y;
+ if (N0.getOperand(0) == Mask)
+ Y = N0.getOperand(1);
+ if (N0.getOperand(1) == Mask)
+ Y = N0.getOperand(0);
+
+ // Check to see if the mask appeared in both the AND and PANDN and
+ if (!Y.getNode())
+ return SDValue();
+
+ // Validate that X, Y, and Mask are BIT_CONVERTS, and see through them.
+ if (Mask.getOpcode() != ISD::BITCAST ||
+ X.getOpcode() != ISD::BITCAST ||
+ Y.getOpcode() != ISD::BITCAST)
+ return SDValue();
+
+ // Look through mask bitcast.
+ Mask = Mask.getOperand(0);
+ EVT MaskVT = Mask.getValueType();
+
+ // Validate that the Mask operand is a vector sra node. The sra node
+ // will be an intrinsic.
+ if (Mask.getOpcode() != ISD::INTRINSIC_WO_CHAIN)
+ return SDValue();
+
+ // FIXME: what to do for bytes, since there is a psignb/pblendvb, but
+ // there is no psrai.b
+ switch (cast<ConstantSDNode>(Mask.getOperand(0))->getZExtValue()) {
+ case Intrinsic::x86_sse2_psrai_w:
+ case Intrinsic::x86_sse2_psrai_d:
+ break;
+ default: return SDValue();
+ }
+
+ // Check that the SRA is all signbits.
+ SDValue SraC = Mask.getOperand(2);
+ unsigned SraAmt = cast<ConstantSDNode>(SraC)->getZExtValue();
+ unsigned EltBits = MaskVT.getVectorElementType().getSizeInBits();
+ if ((SraAmt + 1) != EltBits)
+ return SDValue();
+
+ DebugLoc DL = N->getDebugLoc();
+
+ // Now we know we at least have a plendvb with the mask val. See if
+ // we can form a psignb/w/d.
+ // psign = x.type == y.type == mask.type && y = sub(0, x);
+ X = X.getOperand(0);
+ Y = Y.getOperand(0);
+ if (Y.getOpcode() == ISD::SUB && Y.getOperand(1) == X &&
+ ISD::isBuildVectorAllZeros(Y.getOperand(0).getNode()) &&
+ X.getValueType() == MaskVT && X.getValueType() == Y.getValueType()){
+ unsigned Opc = 0;
+ switch (EltBits) {
+ case 8: Opc = X86ISD::PSIGNB; break;
+ case 16: Opc = X86ISD::PSIGNW; break;
+ case 32: Opc = X86ISD::PSIGND; break;
+ default: break;
+ }
+ if (Opc) {
+ SDValue Sign = DAG.getNode(Opc, DL, MaskVT, X, Mask.getOperand(1));
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Sign);
+ }
+ }
+ // PBLENDVB only available on SSE 4.1
+ if (!Subtarget->hasSSE41())
+ return SDValue();
+
+ X = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, X);
+ Y = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Y);
+ Mask = DAG.getNode(ISD::BITCAST, DL, MVT::v16i8, Mask);
+ Mask = DAG.getNode(X86ISD::PBLENDVB, DL, MVT::v16i8, X, Y, Mask);
+ return DAG.getNode(ISD::BITCAST, DL, MVT::v2i64, Mask);
+ }
+ }
+ }
+
+ // fold (or (x << c) | (y >> (64 - c))) ==> (shld64 x, y, c)
if (N0.getOpcode() == ISD::SRL && N1.getOpcode() == ISD::SHL)
std::swap(N0, N1);
if (N0.getOpcode() != ISD::SHL || N1.getOpcode() != ISD::SRL)
DAG.getNode(ISD::TRUNCATE, DL,
MVT::i8, ShAmt0));
}
-
+
return SDValue();
}
static SDValue PerformVZEXT_MOVLCombine(SDNode *N, SelectionDAG &DAG) {
SDValue Op = N->getOperand(0);
- if (Op.getOpcode() == ISD::BIT_CONVERT)
+ if (Op.getOpcode() == ISD::BITCAST)
Op = Op.getOperand(0);
EVT VT = N->getValueType(0), OpVT = Op.getValueType();
if (Op.getOpcode() == X86ISD::VZEXT_LOAD &&
VT.getVectorElementType().getSizeInBits() ==
OpVT.getVectorElementType().getSizeInBits()) {
- return DAG.getNode(ISD::BIT_CONVERT, N->getDebugLoc(), VT, Op);
+ return DAG.getNode(ISD::BITCAST, N->getDebugLoc(), VT, Op);
}
return SDValue();
}
return SDValue();
}
+// Optimize RES = X86ISD::SETCC CONDCODE, EFLAG_INPUT
+static SDValue PerformSETCCCombine(SDNode *N, SelectionDAG &DAG) {
+ unsigned X86CC = N->getConstantOperandVal(0);
+ SDValue EFLAG = N->getOperand(1);
+ DebugLoc DL = N->getDebugLoc();
+
+ // Materialize "setb reg" as "sbb reg,reg", since it can be extended without
+ // a zext and produces an all-ones bit which is more useful than 0/1 in some
+ // cases.
+ if (X86CC == X86::COND_B)
+ return DAG.getNode(ISD::AND, DL, MVT::i8,
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, MVT::i8,
+ DAG.getConstant(X86CC, MVT::i8), EFLAG),
+ DAG.getConstant(1, MVT::i8));
+
+ return SDValue();
+}
+
+// Optimize RES, EFLAGS = X86ISD::ADC LHS, RHS, EFLAGS
+static SDValue PerformADCCombine(SDNode *N, SelectionDAG &DAG,
+ X86TargetLowering::DAGCombinerInfo &DCI) {
+ // If the LHS and RHS of the ADC node are zero, then it can't overflow and
+ // the result is either zero or one (depending on the input carry bit).
+ // Strength reduce this down to a "set on carry" aka SETCC_CARRY&1.
+ if (X86::isZeroNode(N->getOperand(0)) &&
+ X86::isZeroNode(N->getOperand(1)) &&
+ // We don't have a good way to replace an EFLAGS use, so only do this when
+ // dead right now.
+ SDValue(N, 1).use_empty()) {
+ DebugLoc DL = N->getDebugLoc();
+ EVT VT = N->getValueType(0);
+ SDValue CarryOut = DAG.getConstant(0, N->getValueType(1));
+ SDValue Res1 = DAG.getNode(ISD::AND, DL, VT,
+ DAG.getNode(X86ISD::SETCC_CARRY, DL, VT,
+ DAG.getConstant(X86::COND_B,MVT::i8),
+ N->getOperand(2)),
+ DAG.getConstant(1, VT));
+ return DCI.CombineTo(N, Res1, CarryOut);
+ }
+
+ return SDValue();
+}
+
+// fold (add Y, (sete X, 0)) -> adc 0, Y
+// (add Y, (setne X, 0)) -> sbb -1, Y
+// (sub (sete X, 0), Y) -> sbb 0, Y
+// (sub (setne X, 0), Y) -> adc -1, Y
+static SDValue OptimizeConditonalInDecrement(SDNode *N, SelectionDAG &DAG) {
+ DebugLoc DL = N->getDebugLoc();
+
+ // Look through ZExts.
+ SDValue Ext = N->getOperand(N->getOpcode() == ISD::SUB ? 1 : 0);
+ if (Ext.getOpcode() != ISD::ZERO_EXTEND || !Ext.hasOneUse())
+ return SDValue();
+
+ SDValue SetCC = Ext.getOperand(0);
+ if (SetCC.getOpcode() != X86ISD::SETCC || !SetCC.hasOneUse())
+ return SDValue();
+
+ X86::CondCode CC = (X86::CondCode)SetCC.getConstantOperandVal(0);
+ if (CC != X86::COND_E && CC != X86::COND_NE)
+ return SDValue();
+
+ SDValue Cmp = SetCC.getOperand(1);
+ if (Cmp.getOpcode() != X86ISD::CMP || !Cmp.hasOneUse() ||
+ !X86::isZeroNode(Cmp.getOperand(1)))
+ return SDValue();
+
+ SDValue CmpOp0 = Cmp.getOperand(0);
+ SDValue NewCmp = DAG.getNode(X86ISD::CMP, DL, MVT::i32, CmpOp0,
+ DAG.getConstant(1, CmpOp0.getValueType()));
+
+ SDValue OtherVal = N->getOperand(N->getOpcode() == ISD::SUB ? 0 : 1);
+ if (CC == X86::COND_NE)
+ return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::ADC : X86ISD::SBB,
+ DL, OtherVal.getValueType(), OtherVal,
+ DAG.getConstant(-1ULL, OtherVal.getValueType()), NewCmp);
+ return DAG.getNode(N->getOpcode() == ISD::SUB ? X86ISD::SBB : X86ISD::ADC,
+ DL, OtherVal.getValueType(), OtherVal,
+ DAG.getConstant(0, OtherVal.getValueType()), NewCmp);
+}
+
SDValue X86TargetLowering::PerformDAGCombine(SDNode *N,
DAGCombinerInfo &DCI) const {
SelectionDAG &DAG = DCI.DAG;
switch (N->getOpcode()) {
default: break;
case ISD::EXTRACT_VECTOR_ELT:
- return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
+ return PerformEXTRACT_VECTOR_ELTCombine(N, DAG, *this);
case ISD::SELECT: return PerformSELECTCombine(N, DAG, Subtarget);
case X86ISD::CMOV: return PerformCMOVCombine(N, DAG, DCI);
+ case ISD::ADD:
+ case ISD::SUB: return OptimizeConditonalInDecrement(N, DAG);
+ case X86ISD::ADC: return PerformADCCombine(N, DAG, DCI);
case ISD::MUL: return PerformMulCombine(N, DAG, DCI);
case ISD::SHL:
case ISD::SRA:
case ISD::SRL: return PerformShiftCombine(N, DAG, Subtarget);
+ case ISD::AND: return PerformAndCombine(N, DAG, DCI, Subtarget);
case ISD::OR: return PerformOrCombine(N, DAG, DCI, Subtarget);
case ISD::STORE: return PerformSTORECombine(N, DAG, Subtarget);
case X86ISD::FXOR:
case X86ISD::BT: return PerformBTCombine(N, DAG, DCI);
case X86ISD::VZEXT_MOVL: return PerformVZEXT_MOVLCombine(N, DAG);
case ISD::ZERO_EXTEND: return PerformZExtCombine(N, DAG);
+ case X86ISD::SETCC: return PerformSETCCCombine(N, DAG);
case X86ISD::SHUFPS: // Handle all target specific shuffles
case X86ISD::SHUFPD:
case X86ISD::PALIGN:
case X86ISD::PSHUFLW:
case X86ISD::MOVSS:
case X86ISD::MOVSD:
- case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, *this);
+ case ISD::VECTOR_SHUFFLE: return PerformShuffleCombine(N, DAG, DCI);
}
return SDValue();
bool X86TargetLowering::ExpandInlineAsm(CallInst *CI) const {
InlineAsm *IA = cast<InlineAsm>(CI->getCalledValue());
- std::vector<InlineAsm::ConstraintInfo> Constraints = IA->ParseConstraints();
+ InlineAsm::ConstraintInfoVector Constraints = IA->ParseConstraints();
std::string AsmStr = IA->getAsmString();
// TODO: should remove alternatives from the asmstring: "foo {a|b}" -> "foo a"
SmallVector<StringRef, 4> AsmPieces;
- SplitString(AsmStr, AsmPieces, "\n"); // ; as separator?
+ SplitString(AsmStr, AsmPieces, ";\n");
switch (AsmPieces.size()) {
default: return false;
}
break;
case 3:
+ if (CI->getType()->isIntegerTy(32) &&
+ IA->getConstraintString().compare(0, 5, "=r,0,") == 0) {
+ SmallVector<StringRef, 4> Words;
+ SplitString(AsmPieces[0], Words, " \t,");
+ if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
+ Words[2] == "${0:w}") {
+ Words.clear();
+ SplitString(AsmPieces[1], Words, " \t,");
+ if (Words.size() == 3 && Words[0] == "rorl" && Words[1] == "$$16" &&
+ Words[2] == "$0") {
+ Words.clear();
+ SplitString(AsmPieces[2], Words, " \t,");
+ if (Words.size() == 3 && Words[0] == "rorw" && Words[1] == "$$8" &&
+ Words[2] == "${0:w}") {
+ AsmPieces.clear();
+ const std::string &Constraints = IA->getConstraintString();
+ SplitString(StringRef(Constraints).substr(5), AsmPieces, ",");
+ std::sort(AsmPieces.begin(), AsmPieces.end());
+ if (AsmPieces.size() == 4 &&
+ AsmPieces[0] == "~{cc}" &&
+ AsmPieces[1] == "~{dirflag}" &&
+ AsmPieces[2] == "~{flags}" &&
+ AsmPieces[3] == "~{fpsr}") {
+ return LowerToBSwap(CI);
+ }
+ }
+ }
+ }
+ }
if (CI->getType()->isIntegerTy(64) &&
Constraints.size() >= 2 &&
Constraints[0].Codes.size() == 1 && Constraints[0].Codes[0] == "A" &&
X86TargetLowering::getConstraintType(const std::string &Constraint) const {
if (Constraint.size() == 1) {
switch (Constraint[0]) {
- case 'A':
- return C_Register;
- case 'f':
- case 'r':
case 'R':
- case 'l':
case 'q':
case 'Q':
- case 'x':
+ case 'f':
+ case 't':
+ case 'u':
case 'y':
+ case 'x':
case 'Y':
return C_RegisterClass;
+ case 'a':
+ case 'b':
+ case 'c':
+ case 'd':
+ case 'S':
+ case 'D':
+ case 'A':
+ return C_Register;
+ case 'I':
+ case 'J':
+ case 'K':
+ case 'L':
+ case 'M':
+ case 'N':
+ case 'G':
+ case 'C':
case 'e':
case 'Z':
return C_Other;
return TargetLowering::getConstraintType(Constraint);
}
-/// Examine constraint type and operand type and determine a weight value,
-/// where: -1 = invalid match, and 0 = so-so match to 3 = good match.
+/// Examine constraint type and operand type and determine a weight value.
/// This object must already have been set up with the operand type
/// and the current alternative constraint selected.
-int X86TargetLowering::getSingleConstraintMatchWeight(
+TargetLowering::ConstraintWeight
+ X86TargetLowering::getSingleConstraintMatchWeight(
AsmOperandInfo &info, const char *constraint) const {
- int weight = -1;
+ ConstraintWeight weight = CW_Invalid;
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)
- return 0;
+ return CW_Default;
+ const Type *type = CallOperandVal->getType();
// Look at the constraint type.
switch (*constraint) {
default:
- return TargetLowering::getSingleConstraintMatchWeight(info, constraint);
+ weight = TargetLowering::getSingleConstraintMatchWeight(info, constraint);
+ case 'R':
+ case 'q':
+ case 'Q':
+ case 'a':
+ case 'b':
+ case 'c':
+ case 'd':
+ case 'S':
+ case 'D':
+ case 'A':
+ if (CallOperandVal->getType()->isIntegerTy())
+ weight = CW_SpecificReg;
+ break;
+ case 'f':
+ case 't':
+ case 'u':
+ if (type->isFloatingPointTy())
+ weight = CW_SpecificReg;
+ break;
+ case 'y':
+ if (type->isX86_MMXTy() && Subtarget->hasMMX())
+ weight = CW_SpecificReg;
+ break;
+ case 'x':
+ case 'Y':
+ if ((type->getPrimitiveSizeInBits() == 128) && Subtarget->hasXMM())
+ weight = CW_Register;
break;
case 'I':
if (ConstantInt *C = dyn_cast<ConstantInt>(info.CallOperandVal)) {
if (C->getZExtValue() <= 31)
- weight = 3;
+ weight = CW_Constant;
+ }
+ break;
+ case 'J':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
+ if (C->getZExtValue() <= 63)
+ weight = CW_Constant;
+ }
+ break;
+ case 'K':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
+ if ((C->getSExtValue() >= -0x80) && (C->getSExtValue() <= 0x7f))
+ weight = CW_Constant;
+ }
+ break;
+ case 'L':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
+ if ((C->getZExtValue() == 0xff) || (C->getZExtValue() == 0xffff))
+ weight = CW_Constant;
+ }
+ break;
+ case 'M':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
+ if (C->getZExtValue() <= 3)
+ weight = CW_Constant;
+ }
+ break;
+ case 'N':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
+ if (C->getZExtValue() <= 0xff)
+ weight = CW_Constant;
+ }
+ break;
+ case 'G':
+ case 'C':
+ if (dyn_cast<ConstantFP>(CallOperandVal)) {
+ weight = CW_Constant;
+ }
+ break;
+ case 'e':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
+ if ((C->getSExtValue() >= -0x80000000LL) &&
+ (C->getSExtValue() <= 0x7fffffffLL))
+ weight = CW_Constant;
+ }
+ break;
+ case 'Z':
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CallOperandVal)) {
+ if (C->getZExtValue() <= 0xffffffff)
+ weight = CW_Constant;
}
break;
- // etc.
}
return weight;
}
// FP X constraints get lowered to SSE1/2 registers if available, otherwise
// 'f' like normal targets.
if (ConstraintVT.isFloatingPoint()) {
- if (Subtarget->hasSSE2())
+ if (Subtarget->hasXMMInt())
return "Y";
- if (Subtarget->hasSSE1())
+ if (Subtarget->hasXMM())
return "x";
}
if (!Subtarget->hasMMX()) break;
return std::make_pair(0U, X86::VR64RegisterClass);
case 'Y': // SSE_REGS if SSE2 allowed
- if (!Subtarget->hasSSE2()) break;
+ if (!Subtarget->hasXMMInt()) break;
// FALL THROUGH.
case 'x': // SSE_REGS if SSE1 allowed
- if (!Subtarget->hasSSE1()) break;
+ if (!Subtarget->hasXMM()) break;
switch (VT.getSimpleVT().SimpleTy) {
default: break;