#include "X86.h"
#include "X86InstrBuilder.h"
+#include "X86ISelLowering.h"
#include "X86RegisterInfo.h"
#include "X86Subtarget.h"
#include "X86TargetMachine.h"
#include "llvm/CallingConv.h"
#include "llvm/DerivedTypes.h"
#include "llvm/GlobalVariable.h"
+#include "llvm/GlobalAlias.h"
#include "llvm/Instructions.h"
#include "llvm/IntrinsicInst.h"
#include "llvm/Operator.h"
bool X86FastEmitLoad(EVT VT, const X86AddressMode &AM, unsigned &RR);
- bool X86FastEmitStore(EVT VT, const Value *Val,
- const X86AddressMode &AM);
- bool X86FastEmitStore(EVT VT, unsigned Val,
- const X86AddressMode &AM);
+ bool X86FastEmitStore(EVT VT, const Value *Val, const X86AddressMode &AM);
+ bool X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM);
bool X86FastEmitExtend(ISD::NodeType Opc, EVT DstVT, unsigned Src, EVT SrcVT,
unsigned &ResultReg);
bool X86SelectFPExt(const Instruction *I);
bool X86SelectFPTrunc(const Instruction *I);
- bool X86SelectExtractValue(const Instruction *I);
-
bool X86VisitIntrinsicCall(const IntrinsicInst &I);
bool X86SelectCall(const Instruction *I);
+ bool DoSelectCall(const Instruction *I, const char *MemIntName);
+
const X86InstrInfo *getInstrInfo() const {
return getTargetMachine()->getInstrInfo();
}
unsigned TargetMaterializeAlloca(const AllocaInst *C);
+ unsigned TargetMaterializeFloatZero(const ConstantFP *CF);
+
/// isScalarFPTypeInSSEReg - Return true if the specified scalar FP type is
/// computed in an SSE register, not on the X87 floating point stack.
bool isScalarFPTypeInSSEReg(EVT VT) const {
(VT == MVT::f32 && X86ScalarSSEf32); // f32 is when SSE1
}
- bool isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1 = false);
+ bool isTypeLegal(Type *Ty, MVT &VT, bool AllowI1 = false);
+
+ bool IsMemcpySmall(uint64_t Len);
+
+ bool TryEmitSmallMemcpy(X86AddressMode DestAM,
+ X86AddressMode SrcAM, uint64_t Len);
};
} // end anonymous namespace.
-bool X86FastISel::isTypeLegal(const Type *Ty, MVT &VT, bool AllowI1) {
+bool X86FastISel::isTypeLegal(Type *Ty, MVT &VT, bool AllowI1) {
EVT evt = TLI.getValueType(Ty, /*HandleUnknown=*/true);
if (evt == MVT::Other || !evt.isSimple())
// Unhandled type. Halt "fast" selection and bail.
case MVT::i1:
case MVT::i8:
Opc = X86::MOV8rm;
- RC = X86::GR8RegisterClass;
+ RC = &X86::GR8RegClass;
break;
case MVT::i16:
Opc = X86::MOV16rm;
- RC = X86::GR16RegisterClass;
+ RC = &X86::GR16RegClass;
break;
case MVT::i32:
Opc = X86::MOV32rm;
- RC = X86::GR32RegisterClass;
+ RC = &X86::GR32RegClass;
break;
case MVT::i64:
// Must be in x86-64 mode.
Opc = X86::MOV64rm;
- RC = X86::GR64RegisterClass;
+ RC = &X86::GR64RegClass;
break;
case MVT::f32:
- if (Subtarget->hasSSE1()) {
- Opc = X86::MOVSSrm;
- RC = X86::FR32RegisterClass;
+ if (X86ScalarSSEf32) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
+ RC = &X86::FR32RegClass;
} else {
Opc = X86::LD_Fp32m;
- RC = X86::RFP32RegisterClass;
+ RC = &X86::RFP32RegClass;
}
break;
case MVT::f64:
- if (Subtarget->hasSSE2()) {
- Opc = X86::MOVSDrm;
- RC = X86::FR64RegisterClass;
+ if (X86ScalarSSEf64) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
+ RC = &X86::FR64RegClass;
} else {
Opc = X86::LD_Fp64m;
- RC = X86::RFP64RegisterClass;
+ RC = &X86::RFP64RegClass;
}
break;
case MVT::f80:
/// and a displacement offset, or a GlobalAddress,
/// i.e. V. Return true if it is possible.
bool
-X86FastISel::X86FastEmitStore(EVT VT, unsigned Val,
- const X86AddressMode &AM) {
+X86FastISel::X86FastEmitStore(EVT VT, unsigned Val, const X86AddressMode &AM) {
// Get opcode and regclass of the output for the given store instruction.
unsigned Opc = 0;
switch (VT.getSimpleVT().SimpleTy) {
default: return false;
case MVT::i1: {
// Mask out all but lowest bit.
- unsigned AndResult = createResultReg(X86::GR8RegisterClass);
+ unsigned AndResult = createResultReg(&X86::GR8RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(X86::AND8ri), AndResult).addReg(Val).addImm(1);
Val = AndResult;
case MVT::i32: Opc = X86::MOV32mr; break;
case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
case MVT::f32:
- Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
+ Opc = X86ScalarSSEf32 ?
+ (Subtarget->hasAVX() ? X86::VMOVSSmr : X86::MOVSSmr) : X86::ST_Fp32m;
break;
case MVT::f64:
- Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
+ Opc = X86ScalarSSEf64 ?
+ (Subtarget->hasAVX() ? X86::VMOVSDmr : X86::MOVSDmr) : X86::ST_Fp64m;
+ break;
+ case MVT::v4f32:
+ Opc = X86::MOVAPSmr;
+ break;
+ case MVT::v2f64:
+ Opc = X86::MOVAPDmr;
+ break;
+ case MVT::v4i32:
+ case MVT::v2i64:
+ case MVT::v8i16:
+ case MVT::v16i8:
+ Opc = X86::MOVDQAmr;
break;
}
U = C;
}
- if (
const PointerType *Ty = dyn_cast<PointerType>(V->getType()))
+ if (PointerType *Ty = dyn_cast<PointerType>(V->getType()))
if (Ty->getAddressSpace() > 255)
// Fast instruction selection doesn't support the special
// address spaces.
for (User::const_op_iterator i = U->op_begin() + 1, e = U->op_end();
i != e; ++i, ++GTI) {
const Value *Op = *i;
- if (const StructType *STy = dyn_cast<StructType>(*GTI)) {
+ if (StructType *STy = dyn_cast<StructType>(*GTI)) {
const StructLayout *SL = TD.getStructLayout(STy);
Disp += SL->getElementOffset(cast<ConstantInt>(Op)->getZExtValue());
continue;
}
-
+
// A array/variable index is always of the form i*S where S is the
// constant scale size. See if we can push the scale into immediates.
uint64_t S = TD.getTypeAllocSize(GTI.getIndexedType());
// Handle constant address.
if (const GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
- // Can't handle alternate code models or TLS yet.
+ // Can't handle alternate code models yet.
if (TM.getCodeModel() != CodeModel::Small)
return false;
+ // Can't handle TLS yet.
if (const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV))
if (GVar->isThreadLocal())
return false;
-
+
+ // Can't handle TLS yet, part 2 (this is slightly crazy, but this is how
+ // it works...).
+ if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
+ if (const GlobalVariable *GVar =
+ dyn_cast_or_null<GlobalVariable>(GA->resolveAliasedGlobal(false)))
+ if (GVar->isThreadLocal())
+ return false;
+
// RIP-relative addresses can't have additional register operands, so if
// we've already folded stuff into the addressing mode, just force the
// global value into its own register, which we can use as the basereg.
if (TLI.getPointerTy() == MVT::i64) {
Opc = X86::MOV64rm;
- RC = X86::GR64RegisterClass;
+ RC = &X86::GR64RegClass;
if (Subtarget->isPICStyleRIPRel())
StubAM.Base.Reg = X86::RIP;
} else {
Opc = X86::MOV32rm;
- RC = X86::GR32RegisterClass;
+ RC = &X86::GR32RegClass;
}
LoadReg = createResultReg(RC);
/// X86SelectStore - Select and emit code to implement store instructions.
bool X86FastISel::X86SelectStore(const Instruction *I) {
+ // Atomic stores need special handling.
+ const StoreInst *S = cast<StoreInst>(I);
+
+ if (S->isAtomic())
+ return false;
+
+ unsigned SABIAlignment =
+ TD.getABITypeAlignment(S->getValueOperand()->getType());
+ if (S->getAlignment() != 0 && S->getAlignment() < SABIAlignment)
+ return false;
+
MVT VT;
if (!isTypeLegal(I->getOperand(0)->getType(), VT, /*AllowI1=*/true))
return false;
// fastcc with -tailcallopt is intended to provide a guaranteed
// tail call optimization. Fastisel doesn't know how to do that.
- if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
+ if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
return false;
// Let SDISel handle vararg functions.
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ValLocs;
- CCState CCInfo(CC, F.isVarArg(), TM, ValLocs, I->getContext());
+ CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,
+ I->getContext());
CCInfo.AnalyzeReturn(Outs, RetCC_X86);
const Value *RV = Ret->getOperand(0);
// Only handle register returns for now.
if (!VA.isRegLoc())
return false;
- // TODO: For now, don't try to handle cases where getLocInfo()
- // says Full but the types don't match.
- if (TLI.getValueType(RV->getType()) != VA.getValVT())
- return false;
// The calling-convention tables for x87 returns don't tell
// the whole story.
if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
return false;
- // Make the copy.
unsigned SrcReg = Reg + VA.getValNo();
+ EVT SrcVT = TLI.getValueType(RV->getType());
+ EVT DstVT = VA.getValVT();
+ // Special handling for extended integers.
+ if (SrcVT != DstVT) {
+ if (SrcVT != MVT::i1 && SrcVT != MVT::i8 && SrcVT != MVT::i16)
+ return false;
+
+ if (!Outs[0].Flags.isZExt() && !Outs[0].Flags.isSExt())
+ return false;
+
+ assert(DstVT == MVT::i32 && "X86 should always ext to i32");
+
+ if (SrcVT == MVT::i1) {
+ if (Outs[0].Flags.isSExt())
+ return false;
+ SrcReg = FastEmitZExtFromI1(MVT::i8, SrcReg, /*TODO: Kill=*/false);
+ SrcVT = MVT::i8;
+ }
+ unsigned Op = Outs[0].Flags.isZExt() ? ISD::ZERO_EXTEND :
+ ISD::SIGN_EXTEND;
+ SrcReg = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Op,
+ SrcReg, /*TODO: Kill=*/false);
+ }
+
+ // Make the copy.
unsigned DstReg = VA.getLocReg();
const TargetRegisterClass* SrcRC = MRI.getRegClass(SrcReg);
// Avoid a cross-class copy. This is very unlikely.
/// X86SelectLoad - Select and emit code to implement load instructions.
///
bool X86FastISel::X86SelectLoad(const Instruction *I) {
+ // Atomic loads need special handling.
+ if (cast<LoadInst>(I)->isAtomic())
+ return false;
+
MVT VT;
if (!isTypeLegal(I->getType(), VT, /*AllowI1=*/true))
return false;
}
static unsigned X86ChooseCmpOpcode(EVT VT, const X86Subtarget *Subtarget) {
+ bool HasAVX = Subtarget->hasAVX();
+ bool X86ScalarSSEf32 = Subtarget->hasSSE1();
+ bool X86ScalarSSEf64 = Subtarget->hasSSE2();
+
switch (VT.getSimpleVT().SimpleTy) {
default: return 0;
case MVT::i8: return X86::CMP8rr;
case MVT::i16: return X86::CMP16rr;
case MVT::i32: return X86::CMP32rr;
case MVT::i64: return X86::CMP64rr;
- case MVT::f32: return Subtarget->hasSSE1() ? X86::UCOMISSrr : 0;
- case MVT::f64: return Subtarget->hasSSE2() ? X86::UCOMISDrr : 0;
+ case MVT::f32:
+ return X86ScalarSSEf32 ? (HasAVX ? X86::VUCOMISSrr : X86::UCOMISSrr) : 0;
+ case MVT::f64:
+ return X86ScalarSSEf64 ? (HasAVX ? X86::VUCOMISDrr : X86::UCOMISDrr) : 0;
}
}
unsigned NEReg = createResultReg(&X86::GR8RegClass);
unsigned PReg = createResultReg(&X86::GR8RegClass);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(X86::SETNEr), NEReg);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(X86::SETPr), PReg);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(X86::OR8rr), ResultReg)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETNEr), NEReg);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::SETPr), PReg);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::OR8rr),ResultReg)
.addReg(PReg).addReg(NEReg);
UpdateValueMap(I, ResultReg);
return true;
bool X86FastISel::X86SelectZExt(const Instruction *I) {
// Handle zero-extension from i1 to i8, which is common.
- if (I->getType()->isIntegerTy(8) &&
- I->getOperand(0)->getType()->isIntegerTy(1)) {
- unsigned ResultReg = getRegForValue(I->getOperand(0));
- if (ResultReg == 0) return false;
- // Set the high bits to zero.
- ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
- if (ResultReg == 0) return false;
- UpdateValueMap(I, ResultReg);
- return true;
+ if (!I->getOperand(0)->getType()->isIntegerTy(1))
+ return false;
+
+ EVT DstVT = TLI.getValueType(I->getType());
+ if (!TLI.isTypeLegal(DstVT))
+ return false;
+
+ unsigned ResultReg = getRegForValue(I->getOperand(0));
+ if (ResultReg == 0)
+ return false;
+
+ // Set the high bits to zero.
+ ResultReg = FastEmitZExtFromI1(MVT::i8, ResultReg, /*TODO: Kill=*/false);
+ if (ResultReg == 0)
+ return false;
+
+ if (DstVT != MVT::i8) {
+ ResultReg = FastEmit_r(MVT::i8, DstVT.getSimpleVT(), ISD::ZERO_EXTEND,
+ ResultReg, /*Kill=*/true);
+ if (ResultReg == 0)
+ return false;
}
- return false;
+ UpdateValueMap(I, ResultReg);
+ return true;
}
FuncInfo.MBB->addSuccessor(TrueMBB);
return true;
}
- } else if (ExtractValueInst *EI =
- dyn_cast<ExtractValueInst>(BI->getCondition())) {
- // Check to see if the branch instruction is from an "arithmetic with
- // overflow" intrinsic. The main way these intrinsics are used is:
- //
- // %t = call { i32, i1 } @llvm.sadd.with.overflow.i32(i32 %v1, i32 %v2)
- // %sum = extractvalue { i32, i1 } %t, 0
- // %obit = extractvalue { i32, i1 } %t, 1
- // br i1 %obit, label %overflow, label %normal
- //
- // The %sum and %obit are converted in an ADD and a SETO/SETB before
- // reaching the branch. Therefore, we search backwards through the MBB
- // looking for the SETO/SETB instruction. If an instruction modifies the
- // EFLAGS register before we reach the SETO/SETB instruction, then we can't
- // convert the branch into a JO/JB instruction.
- if (const IntrinsicInst *CI =
- dyn_cast<IntrinsicInst>(EI->getAggregateOperand())){
- if (CI->getIntrinsicID() == Intrinsic::sadd_with_overflow ||
- CI->getIntrinsicID() == Intrinsic::uadd_with_overflow) {
- const MachineInstr *SetMI = 0;
- unsigned Reg = getRegForValue(EI);
-
- for (MachineBasicBlock::const_reverse_iterator
- RI = FuncInfo.MBB->rbegin(), RE = FuncInfo.MBB->rend();
- RI != RE; ++RI) {
- const MachineInstr &MI = *RI;
-
- if (MI.definesRegister(Reg)) {
- if (MI.isCopy()) {
- Reg = MI.getOperand(1).getReg();
- continue;
- }
-
- SetMI = &MI;
- break;
- }
-
- const TargetInstrDesc &TID = MI.getDesc();
- if (TID.hasImplicitDefOfPhysReg(X86::EFLAGS) ||
- MI.hasUnmodeledSideEffects())
- break;
- }
+ } else if (TruncInst *TI = dyn_cast<TruncInst>(BI->getCondition())) {
+ // Handle things like "%cond = trunc i32 %X to i1 / br i1 %cond", which
+ // typically happen for _Bool and C++ bools.
+ MVT SourceVT;
+ if (TI->hasOneUse() && TI->getParent() == I->getParent() &&
+ isTypeLegal(TI->getOperand(0)->getType(), SourceVT)) {
+ unsigned TestOpc = 0;
+ switch (SourceVT.SimpleTy) {
+ default: break;
+ case MVT::i8: TestOpc = X86::TEST8ri; break;
+ case MVT::i16: TestOpc = X86::TEST16ri; break;
+ case MVT::i32: TestOpc = X86::TEST32ri; break;
+ case MVT::i64: TestOpc = X86::TEST64ri32; break;
+ }
+ if (TestOpc) {
+ unsigned OpReg = getRegForValue(TI->getOperand(0));
+ if (OpReg == 0) return false;
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TestOpc))
+ .addReg(OpReg).addImm(1);
- if (SetMI) {
- unsigned OpCode = SetMI->getOpcode();
-
- if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(OpCode == X86::SETOr ? X86::JO_4 : X86::JB_4))
- .addMBB(TrueMBB);
- FastEmitBranch(FalseMBB, DL);
- FuncInfo.MBB->addSuccessor(TrueMBB);
- return true;
- }
+ unsigned JmpOpc = X86::JNE_4;
+ if (FuncInfo.MBB->isLayoutSuccessor(TrueMBB)) {
+ std::swap(TrueMBB, FalseMBB);
+ JmpOpc = X86::JE_4;
}
+
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(JmpOpc))
+ .addMBB(TrueMBB);
+ FastEmitBranch(FalseMBB, DL);
+ FuncInfo.MBB->addSuccessor(TrueMBB);
+ return true;
}
}
}
// Otherwise do a clumsy setcc and re-test it.
+ // Note that i1 essentially gets ANY_EXTEND'ed to i8 where it isn't used
+ // in an explicit cast, so make sure to handle that correctly.
unsigned OpReg = getRegForValue(BI->getCondition());
if (OpReg == 0) return false;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8rr))
- .addReg(OpReg).addReg(OpReg);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::TEST8ri))
+ .addReg(OpReg).addImm(1);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::JNE_4))
.addMBB(TrueMBB);
FastEmitBranch(FalseMBB, DL);
}
bool X86FastISel::X86SelectShift(const Instruction *I) {
- unsigned CReg = 0, OpReg = 0, OpImm = 0;
+ unsigned CReg = 0, OpReg = 0;
const TargetRegisterClass *RC = NULL;
if (I->getType()->isIntegerTy(8)) {
CReg = X86::CL;
RC = &X86::GR8RegClass;
switch (I->getOpcode()) {
- case Instruction::LShr: OpReg = X86::SHR8rCL; OpImm = X86::SHR8ri; break;
- case Instruction::AShr: OpReg = X86::SAR8rCL; OpImm = X86::SAR8ri; break;
- case Instruction::Shl: OpReg = X86::SHL8rCL; OpImm = X86::SHL8ri; break;
+ case Instruction::LShr: OpReg = X86::SHR8rCL; break;
+ case Instruction::AShr: OpReg = X86::SAR8rCL; break;
+ case Instruction::Shl: OpReg = X86::SHL8rCL; break;
default: return false;
}
} else if (I->getType()->isIntegerTy(16)) {
CReg = X86::CX;
RC = &X86::GR16RegClass;
switch (I->getOpcode()) {
- case Instruction::LShr: OpReg = X86::SHR16rCL; OpImm = X86::SHR16ri; break;
- case Instruction::AShr: OpReg = X86::SAR16rCL; OpImm = X86::SAR16ri; break;
- case Instruction::Shl: OpReg = X86::SHL16rCL; OpImm = X86::SHL16ri; break;
+ case Instruction::LShr: OpReg = X86::SHR16rCL; break;
+ case Instruction::AShr: OpReg = X86::SAR16rCL; break;
+ case Instruction::Shl: OpReg = X86::SHL16rCL; break;
default: return false;
}
} else if (I->getType()->isIntegerTy(32)) {
CReg = X86::ECX;
RC = &X86::GR32RegClass;
switch (I->getOpcode()) {
- case Instruction::LShr: OpReg = X86::SHR32rCL; OpImm = X86::SHR32ri; break;
- case Instruction::AShr: OpReg = X86::SAR32rCL; OpImm = X86::SAR32ri; break;
- case Instruction::Shl: OpReg = X86::SHL32rCL; OpImm = X86::SHL32ri; break;
+ case Instruction::LShr: OpReg = X86::SHR32rCL; break;
+ case Instruction::AShr: OpReg = X86::SAR32rCL; break;
+ case Instruction::Shl: OpReg = X86::SHL32rCL; break;
default: return false;
}
} else if (I->getType()->isIntegerTy(64)) {
CReg = X86::RCX;
RC = &X86::GR64RegClass;
switch (I->getOpcode()) {
- case Instruction::LShr: OpReg = X86::SHR64rCL; OpImm = X86::SHR64ri; break;
- case Instruction::AShr: OpReg = X86::SAR64rCL; OpImm = X86::SAR64ri; break;
- case Instruction::Shl: OpReg = X86::SHL64rCL; OpImm = X86::SHL64ri; break;
+ case Instruction::LShr: OpReg = X86::SHR64rCL; break;
+ case Instruction::AShr: OpReg = X86::SAR64rCL; break;
+ case Instruction::Shl: OpReg = X86::SHL64rCL; break;
default: return false;
}
} else {
unsigned Op0Reg = getRegForValue(I->getOperand(0));
if (Op0Reg == 0) return false;
- // Fold immediate in shl(x,3).
- if (const ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
- unsigned ResultReg = createResultReg(RC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpImm),
- ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
- UpdateValueMap(I, ResultReg);
- return true;
- }
-
unsigned Op1Reg = getRegForValue(I->getOperand(1));
if (Op1Reg == 0) return false;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
bool X86FastISel::X86SelectFPExt(const Instruction *I) {
// fpext from float to double.
- if (Subtarget->hasSSE2() &&
+ if (X86ScalarSSEf64 &&
I->getType()->isDoubleTy()) {
const Value *V = I->getOperand(0);
if (V->getType()->isFloatTy()) {
unsigned OpReg = getRegForValue(V);
if (OpReg == 0) return false;
- unsigned ResultReg = createResultReg(X86::FR64RegisterClass);
+ unsigned ResultReg = createResultReg(&X86::FR64RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(X86::CVTSS2SDrr), ResultReg)
.addReg(OpReg);
}
bool X86FastISel::X86SelectFPTrunc(const Instruction *I) {
- if (Subtarget->hasSSE2()) {
+ if (X86ScalarSSEf64) {
if (I->getType()->isFloatTy()) {
const Value *V = I->getOperand(0);
if (V->getType()->isDoubleTy()) {
unsigned OpReg = getRegForValue(V);
if (OpReg == 0) return false;
- unsigned ResultReg = createResultReg(X86::FR32RegisterClass);
+ unsigned ResultReg = createResultReg(&X86::FR32RegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(X86::CVTSD2SSrr), ResultReg)
.addReg(OpReg);
}
bool X86FastISel::X86SelectTrunc(const Instruction *I) {
- if (Subtarget->is64Bit())
- // All other cases should be handled by the tblgen generated code.
- return false;
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
EVT DstVT = TLI.getValueType(I->getType());
- // This code only handles truncation to byte right now.
+ // This code only handles truncation to byte.
if (DstVT != MVT::i8 && DstVT != MVT::i1)
- // All other cases should be handled by the tblgen generated code.
return false;
- if (SrcVT != MVT::i16 && SrcVT != MVT::i32)
- // All other cases should be handled by the tblgen generated code.
+ if (!TLI.isTypeLegal(SrcVT))
return false;
unsigned InputReg = getRegForValue(I->getOperand(0));
// Unhandled operand. Halt "fast" selection and bail.
return false;
- // First issue a copy to GR16_ABCD or GR32_ABCD.
- const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16)
- ? X86::GR16_ABCDRegisterClass : X86::GR32_ABCDRegisterClass;
- unsigned CopyReg = createResultReg(CopyRC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- CopyReg).addReg(InputReg);
+ if (SrcVT == MVT::i8) {
+ // Truncate from i8 to i1; no code needed.
+ UpdateValueMap(I, InputReg);
+ return true;
+ }
+
+ if (!Subtarget->is64Bit()) {
+ // If we're on x86-32; we can't extract an i8 from a general register.
+ // First issue a copy to GR16_ABCD or GR32_ABCD.
+ const TargetRegisterClass *CopyRC = (SrcVT == MVT::i16) ?
+ (const TargetRegisterClass*)&X86::GR16_ABCDRegClass :
+ (const TargetRegisterClass*)&X86::GR32_ABCDRegClass;
+ unsigned CopyReg = createResultReg(CopyRC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+ CopyReg).addReg(InputReg);
+ InputReg = CopyReg;
+ }
- // Then issue an extract_subreg.
+ // Issue an extract_subreg.
unsigned ResultReg = FastEmitInst_extractsubreg(MVT::i8,
- CopyReg, /*Kill=*/true,
+ InputReg, /*Kill=*/true,
X86::sub_8bit);
if (!ResultReg)
return false;
return true;
}
-bool X86FastISel::X86SelectExtractValue(const Instruction *I) {
- const ExtractValueInst *EI = cast<ExtractValueInst>(I);
- const Value *Agg = EI->getAggregateOperand();
+bool X86FastISel::IsMemcpySmall(uint64_t Len) {
+ return Len <= (Subtarget->is64Bit() ? 32 : 16);
+}
- if (const IntrinsicInst *CI = dyn_cast<IntrinsicInst>(Agg)) {
- switch (CI->getIntrinsicID()) {
- default: break;
- case Intrinsic::sadd_with_overflow:
- case Intrinsic::uadd_with_overflow: {
- // Cheat a little. We know that the registers for "add" and "seto" are
- // allocated sequentially. However, we only keep track of the register
- // for "add" in the value map. Use extractvalue's index to get the
- // correct register for "seto".
- unsigned OpReg = getRegForValue(Agg);
- if (OpReg == 0)
- return false;
- UpdateValueMap(I, OpReg + *EI->idx_begin());
- return true;
- }
+bool X86FastISel::TryEmitSmallMemcpy(X86AddressMode DestAM,
+ X86AddressMode SrcAM, uint64_t Len) {
+
+ // Make sure we don't bloat code by inlining very large memcpy's.
+ if (!IsMemcpySmall(Len))
+ return false;
+
+ bool i64Legal = Subtarget->is64Bit();
+
+ // We don't care about alignment here since we just emit integer accesses.
+ while (Len) {
+ MVT VT;
+ if (Len >= 8 && i64Legal)
+ VT = MVT::i64;
+ else if (Len >= 4)
+ VT = MVT::i32;
+ else if (Len >= 2)
+ VT = MVT::i16;
+ else {
+ assert(Len == 1);
+ VT = MVT::i8;
}
+
+ unsigned Reg;
+ bool RV = X86FastEmitLoad(VT, SrcAM, Reg);
+ RV &= X86FastEmitStore(VT, Reg, DestAM);
+ assert(RV && "Failed to emit load or store??");
+
+ unsigned Size = VT.getSizeInBits()/8;
+ Len -= Size;
+ DestAM.Disp += Size;
+ SrcAM.Disp += Size;
}
- return false;
+ return true;
}
bool X86FastISel::X86VisitIntrinsicCall(const IntrinsicInst &I) {
// FIXME: Handle more intrinsics.
switch (I.getIntrinsicID()) {
default: return false;
- case Intrinsic::stackprotector: {
- // Emit code inline code to store the stack guard onto the stack.
- EVT PtrTy = TLI.getPointerTy();
+ case Intrinsic::memcpy: {
+ const MemCpyInst &MCI = cast<MemCpyInst>(I);
+ // Don't handle volatile or variable length memcpys.
+ if (MCI.isVolatile())
+ return false;
- const Value *Op1 = I.getArgOperand(0); // The guard's value.
- const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
+ if (isa<ConstantInt>(MCI.getLength())) {
+ // Small memcpy's are common enough that we want to do them
+ // without a call if possible.
+ uint64_t Len = cast<ConstantInt>(MCI.getLength())->getZExtValue();
+ if (IsMemcpySmall(Len)) {
+ X86AddressMode DestAM, SrcAM;
+ if (!X86SelectAddress(MCI.getRawDest(), DestAM) ||
+ !X86SelectAddress(MCI.getRawSource(), SrcAM))
+ return false;
+ TryEmitSmallMemcpy(DestAM, SrcAM, Len);
+ return true;
+ }
+ }
- // Grab the frame index.
- X86AddressMode AM;
- if (!X86SelectAddress(Slot, AM)) return false;
+ unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
+ if (!MCI.getLength()->getType()->isIntegerTy(SizeWidth))
+ return false;
- if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
+ if (MCI.getSourceAddressSpace() > 255 || MCI.getDestAddressSpace() > 255)
+ return false;
- return true;
+ return DoSelectCall(&I, "memcpy");
}
- case Intrinsic::objectsize: {
- ConstantInt *CI = dyn_cast<ConstantInt>(I.getArgOperand(1));
- const Type *Ty = I.getCalledFunction()->getReturnType();
+ case Intrinsic::memset: {
+ const MemSetInst &MSI = cast<MemSetInst>(I);
- assert(CI && "Non-constant type in Intrinsic::objectsize?");
+ if (MSI.isVolatile())
+ return false;
- MVT VT;
- if (!isTypeLegal(Ty, VT))
+ unsigned SizeWidth = Subtarget->is64Bit() ? 64 : 32;
+ if (!MSI.getLength()->getType()->isIntegerTy(SizeWidth))
return false;
- unsigned OpC = 0;
- if (VT == MVT::i32)
- OpC = X86::MOV32ri;
- else if (VT == MVT::i64)
- OpC = X86::MOV64ri;
- else
+ if (MSI.getDestAddressSpace() > 255)
return false;
- unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg).
- addImm(CI->isZero() ? -1ULL : 0);
- UpdateValueMap(&I, ResultReg);
+ return DoSelectCall(&I, "memset");
+ }
+ case Intrinsic::stackprotector: {
+ // Emit code inline code to store the stack guard onto the stack.
+ EVT PtrTy = TLI.getPointerTy();
+
+ const Value *Op1 = I.getArgOperand(0); // The guard's value.
+ const AllocaInst *Slot = cast<AllocaInst>(I.getArgOperand(1));
+
+ // Grab the frame index.
+ X86AddressMode AM;
+ if (!X86SelectAddress(Slot, AM)) return false;
+ if (!X86FastEmitStore(PtrTy, Op1, AM)) return false;
return true;
}
case Intrinsic::dbg_declare: {
assert(DI->getAddress() && "Null address should be checked earlier!");
if (!X86SelectAddress(DI->getAddress(), AM))
return false;
- const TargetInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
+ const MCInstrDesc &II = TII.get(TargetOpcode::DBG_VALUE);
// FIXME may need to add RegState::Debug to any registers produced,
// although ESP/EBP should be the only ones at the moment.
addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, II), AM).
}
case Intrinsic::sadd_with_overflow:
case Intrinsic::uadd_with_overflow: {
+ // FIXME: Should fold immediates.
+
// Replace "add with overflow" intrinsics with an "add" instruction followed
- // by a seto/setc instruction. Later on, when the "extractvalue"
- // instructions are encountered, we use the fact that two registers were
- // created sequentially to get the correct registers for the "sum" and the
- // "overflow bit".
+ // by a seto/setc instruction.
const Function *Callee = I.getCalledFunction();
- const Type *RetTy =
+ Type *RetTy =
cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
MVT VT;
else
return false;
- unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ // The call to CreateRegs builds two sequential registers, to store the
+ // both the the returned values.
+ unsigned ResultReg = FuncInfo.CreateRegs(I.getType());
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(OpC), ResultReg)
.addReg(Reg1).addReg(Reg2);
- unsigned DestReg1 = UpdateValueMap(&I, ResultReg);
-
- // If the add with overflow is an intra-block value then we just want to
- // create temporaries for it like normal. If it is a cross-block value then
- // UpdateValueMap will return the cross-block register used. Since we
- // *really* want the value to be live in the register pair known by
- // UpdateValueMap, we have to use DestReg1+1 as the destination register in
- // the cross block case. In the non-cross-block case, we should just make
- // another register for the value.
- if (DestReg1 != ResultReg)
- ResultReg = DestReg1+1;
- else
- ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
unsigned Opc = X86::SETBr;
if (I.getIntrinsicID() == Intrinsic::sadd_with_overflow)
Opc = X86::SETOr;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg+1);
+
+ UpdateValueMap(&I, ResultReg, 2);
return true;
}
}
if (const IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI))
return X86VisitIntrinsicCall(*II);
+ return DoSelectCall(I, 0);
+}
+
+// Select either a call, or an llvm.memcpy/memmove/memset intrinsic
+bool X86FastISel::DoSelectCall(const Instruction *I, const char *MemIntName) {
+ const CallInst *CI = cast<CallInst>(I);
+ const Value *Callee = CI->getCalledValue();
+
// Handle only C and fastcc calling conventions for now.
ImmutableCallSite CS(CI);
CallingConv::ID CC = CS.getCallingConv();
- if (CC != CallingConv::C &&
- CC != CallingConv::Fast &&
+ if (CC != CallingConv::C && CC != CallingConv::Fast &&
CC != CallingConv::X86_FastCall)
return false;
// fastcc with -tailcallopt is intended to provide a guaranteed
// tail call optimization. Fastisel doesn't know how to do that.
- if (CC == CallingConv::Fast && GuaranteedTailCallOpt)
+ if (CC == CallingConv::Fast && TM.Options.GuaranteedTailCallOpt)
return false;
- // Let SDISel handle vararg functions.
- const PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
- const FunctionType *FTy = cast<FunctionType>(PT->getElementType());
- if (FTy->isVarArg())
+ PointerType *PT = cast<PointerType>(CS.getCalledValue()->getType());
+ FunctionType *FTy = cast<FunctionType>(PT->getElementType());
+ bool isVarArg = FTy->isVarArg();
+
+ // Don't know how to handle Win64 varargs yet. Nothing special needed for
+ // x86-32. Special handling for x86-64 is implemented.
+ if (isVarArg && Subtarget->isTargetWin64())
return false;
// Fast-isel doesn't know about callee-pop yet.
- if (Subtarget->IsCalleePop(FTy->isVarArg(), CC))
+ if (X86::isCalleePop(CC, Subtarget->is64Bit(), isVarArg,
+ TM.Options.GuaranteedTailCallOpt))
return false;
- // Handle *simple* calls for now.
- const Type *RetTy = CS.getType();
- MVT RetVT;
- if (RetTy->isVoidTy())
- RetVT = MVT::isVoid;
- else if (!isTypeLegal(RetTy, RetVT, true))
+ // Check whether the function can return without sret-demotion.
+ SmallVector<ISD::OutputArg, 4> Outs;
+ SmallVector<uint64_t, 4> Offsets;
+ GetReturnInfo(I->getType(), CS.getAttributes().getRetAttributes(),
+ Outs, TLI, &Offsets);
+ bool CanLowerReturn = TLI.CanLowerReturn(CS.getCallingConv(),
+ *FuncInfo.MF, FTy->isVarArg(),
+ Outs, FTy->getContext());
+ if (!CanLowerReturn)
return false;
// Materialize callee address in a register. FIXME: GV address can be
} else
return false;
- // Allow calls which produce i1 results.
- bool AndToI1 = false;
- if (RetVT == MVT::i1) {
- RetVT = MVT::i8;
- AndToI1 = true;
- }
-
// Deal with call operands first.
SmallVector<const Value *, 8> ArgVals;
SmallVector<unsigned, 8> Args;
SmallVector<MVT, 8> ArgVTs;
SmallVector<ISD::ArgFlagsTy, 8> ArgFlags;
- Args.reserve(CS.arg_size());
- ArgVals.reserve(CS.arg_size());
- ArgVTs.reserve(CS.arg_size());
- ArgFlags.reserve(CS.arg_size());
+ unsigned arg_size = CS.arg_size();
+ Args.reserve(arg_size);
+ ArgVals.reserve(arg_size);
+ ArgVTs.reserve(arg_size);
+ ArgFlags.reserve(arg_size);
for (ImmutableCallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
i != e; ++i) {
- unsigned Arg = getRegForValue(*i);
- if (Arg == 0)
- return false;
+ // If we're lowering a mem intrinsic instead of a regular call, skip the
+ // last two arguments, which should not passed to the underlying functions.
+ if (MemIntName && e-i <= 2)
+ break;
+ Value *ArgVal = *i;
ISD::ArgFlagsTy Flags;
unsigned AttrInd = i - CS.arg_begin() + 1;
if (CS.paramHasAttr(AttrInd, Attribute::SExt))
if (CS.paramHasAttr(AttrInd, Attribute::ZExt))
Flags.setZExt();
- // FIXME: Only handle *easy* calls for now.
- if (CS.paramHasAttr(AttrInd, Attribute::InReg) ||
- CS.paramHasAttr(AttrInd, Attribute::StructRet) ||
- CS.paramHasAttr(AttrInd, Attribute::Nest) ||
- CS.paramHasAttr(AttrInd, Attribute::ByVal))
- return false;
+ if (CS.paramHasAttr(AttrInd, Attribute::ByVal)) {
+ PointerType *Ty = cast<PointerType>(ArgVal->getType());
+ Type *ElementTy = Ty->getElementType();
+ unsigned FrameSize = TD.getTypeAllocSize(ElementTy);
+ unsigned FrameAlign = CS.getParamAlignment(AttrInd);
+ if (!FrameAlign)
+ FrameAlign = TLI.getByValTypeAlignment(ElementTy);
+ Flags.setByVal();
+ Flags.setByValSize(FrameSize);
+ Flags.setByValAlign(FrameAlign);
+ if (!IsMemcpySmall(FrameSize))
+ return false;
+ }
+
+ if (CS.paramHasAttr(AttrInd, Attribute::InReg))
+ Flags.setInReg();
+ if (CS.paramHasAttr(AttrInd, Attribute::Nest))
+ Flags.setNest();
+
+ // If this is an i1/i8/i16 argument, promote to i32 to avoid an extra
+ // instruction. This is safe because it is common to all fastisel supported
+ // calling conventions on x86.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(ArgVal)) {
+ if (CI->getBitWidth() == 1 || CI->getBitWidth() == 8 ||
+ CI->getBitWidth() == 16) {
+ if (Flags.isSExt())
+ ArgVal = ConstantExpr::getSExt(CI,Type::getInt32Ty(CI->getContext()));
+ else
+ ArgVal = ConstantExpr::getZExt(CI,Type::getInt32Ty(CI->getContext()));
+ }
+ }
+
+ unsigned ArgReg;
+
+ // Passing bools around ends up doing a trunc to i1 and passing it.
+ // Codegen this as an argument + "and 1".
+ if (ArgVal->getType()->isIntegerTy(1) && isa<TruncInst>(ArgVal) &&
+ cast<TruncInst>(ArgVal)->getParent() == I->getParent() &&
+ ArgVal->hasOneUse()) {
+ ArgVal = cast<TruncInst>(ArgVal)->getOperand(0);
+ ArgReg = getRegForValue(ArgVal);
+ if (ArgReg == 0) return false;
- const Type *ArgTy = (*i)->getType();
+ MVT ArgVT;
+ if (!isTypeLegal(ArgVal->getType(), ArgVT)) return false;
+
+ ArgReg = FastEmit_ri(ArgVT, ArgVT, ISD::AND, ArgReg,
+ ArgVal->hasOneUse(), 1);
+ } else {
+ ArgReg = getRegForValue(ArgVal);
+ }
+
+ if (ArgReg == 0) return false;
+
+ Type *ArgTy = ArgVal->getType();
MVT ArgVT;
if (!isTypeLegal(ArgTy, ArgVT))
return false;
+ if (ArgVT == MVT::x86mmx)
+ return false;
unsigned OriginalAlignment = TD.getABITypeAlignment(ArgTy);
Flags.setOrigAlign(OriginalAlignment);
- Args.push_back(Arg);
- ArgVals.push_back(*i);
+ Args.push_back(ArgReg);
+ ArgVals.push_back(ArgVal);
ArgVTs.push_back(ArgVT);
ArgFlags.push_back(Flags);
}
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, false, TM, ArgLocs, I->getParent()->getContext());
+ CCState CCInfo(CC, isVarArg, *FuncInfo.MF, TM, ArgLocs,
+ I->getParent()->getContext());
// Allocate shadow area for Win64
- if (Subtarget->isTargetWin64()) {
+ if (Subtarget->isTargetWin64())
CCInfo.AllocateStack(32, 8);
- }
CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_X86);
unsigned NumBytes = CCInfo.getNextStackOffset();
// Issue CALLSEQ_START
- unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
+ unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackDown))
.addImm(NumBytes);
default: llvm_unreachable("Unknown loc info!");
case CCValAssign::Full: break;
case CCValAssign::SExt: {
+ assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
+ "Unexpected extend");
bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
Arg, ArgVT, Arg);
assert(Emitted && "Failed to emit a sext!"); (void)Emitted;
break;
}
case CCValAssign::ZExt: {
+ assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
+ "Unexpected extend");
bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
Arg, ArgVT, Arg);
assert(Emitted && "Failed to emit a zext!"); (void)Emitted;
break;
}
case CCValAssign::AExt: {
- // We don't handle MMX parameters yet.
- if (VA.getLocVT().isVector() && VA.getLocVT().getSizeInBits() == 128)
- return false;
+ assert(VA.getLocVT().isInteger() && !VA.getLocVT().isVector() &&
+ "Unexpected extend");
bool Emitted = X86FastEmitExtend(ISD::ANY_EXTEND, VA.getLocVT(),
Arg, ArgVT, Arg);
if (!Emitted)
AM.Base.Reg = StackPtr;
AM.Disp = LocMemOffset;
const Value *ArgVal = ArgVals[VA.getValNo()];
-
- // If this is a really simple value, emit this with the Value* version of
- // X86FastEmitStore. If it isn't simple, we don't want to do this, as it
- // can cause us to reevaluate the argument.
- if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal))
- X86FastEmitStore(ArgVT, ArgVal, AM);
- else
- X86FastEmitStore(ArgVT, Arg, AM);
+ ISD::ArgFlagsTy Flags = ArgFlags[VA.getValNo()];
+
+ if (Flags.isByVal()) {
+ X86AddressMode SrcAM;
+ SrcAM.Base.Reg = Arg;
+ bool Res = TryEmitSmallMemcpy(AM, SrcAM, Flags.getByValSize());
+ assert(Res && "memcpy length already checked!"); (void)Res;
+ } else if (isa<ConstantInt>(ArgVal) || isa<ConstantPointerNull>(ArgVal)) {
+ // If this is a really simple value, emit this with the Value* version
+ // of X86FastEmitStore. If it isn't simple, we don't want to do this,
+ // as it can cause us to reevaluate the argument.
+ if (!X86FastEmitStore(ArgVT, ArgVal, AM))
+ return false;
+ } else {
+ if (!X86FastEmitStore(ArgVT, Arg, AM))
+ return false;
+ }
}
}
X86::EBX).addReg(Base);
}
+ if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64()) {
+ // Count the number of XMM registers allocated.
+ static const uint16_t XMMArgRegs[] = {
+ X86::XMM0, X86::XMM1, X86::XMM2, X86::XMM3,
+ X86::XMM4, X86::XMM5, X86::XMM6, X86::XMM7
+ };
+ unsigned NumXMMRegs = CCInfo.getFirstUnallocated(XMMArgRegs, 8);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::MOV8ri),
+ X86::AL).addImm(NumXMMRegs);
+ }
+
// Issue the call.
MachineInstrBuilder MIB;
if (CalleeOp) {
// Register-indirect call.
unsigned CallOpc;
- if (Subtarget->isTargetWin64())
- CallOpc = X86::WINCALL64r;
- else if (Subtarget->is64Bit())
+ if (Subtarget->is64Bit())
CallOpc = X86::CALL64r;
else
CallOpc = X86::CALL32r;
// Direct call.
assert(GV && "Not a direct call");
unsigned CallOpc;
- if (Subtarget->isTargetWin64())
- CallOpc = X86::WINCALL64pcrel32;
- else if (Subtarget->is64Bit())
+ if (Subtarget->is64Bit())
CallOpc = X86::CALL64pcrel32;
else
CallOpc = X86::CALLpcrel32;
OpFlags = X86II::MO_PLT;
} else if (Subtarget->isPICStyleStubAny() &&
(GV->isDeclaration() || GV->isWeakForLinker()) &&
- Subtarget->getDarwinVers() < 9) {
+ (!Subtarget->getTargetTriple().isMacOSX() ||
+ Subtarget->getTargetTriple().isMacOSXVersionLT(10, 5))) {
// PC-relative references to external symbols should go through $stub,
// unless we're building with the leopard linker or later, which
// automatically synthesizes these stubs.
}
- MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc))
- .addGlobalAddress(GV, 0, OpFlags);
+ MIB = BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(CallOpc));
+ if (MemIntName)
+ MIB.addExternalSymbol(MemIntName, OpFlags);
+ else
+ MIB.addGlobalAddress(GV, 0, OpFlags);
}
// Add an implicit use GOT pointer in EBX.
if (Subtarget->isPICStyleGOT())
MIB.addReg(X86::EBX);
+ if (Subtarget->is64Bit() && isVarArg && !Subtarget->isTargetWin64())
+ MIB.addReg(X86::AL);
+
// Add implicit physical register uses to the call.
for (unsigned i = 0, e = RegArgs.size(); i != e; ++i)
MIB.addReg(RegArgs[i]);
+ // Add a register mask with the call-preserved registers.
+ // Proper defs for return values will be added by setPhysRegsDeadExcept().
+ MIB.addRegMask(TRI.getCallPreservedMask(CS.getCallingConv()));
+
// Issue CALLSEQ_END
- unsigned AdjStackUp = TM.getRegisterInfo()->getCallFrameDestroyOpcode();
+ unsigned AdjStackUp = TII.getCallFrameDestroyOpcode();
+ unsigned NumBytesCallee = 0;
+ if (!Subtarget->is64Bit() && !Subtarget->isTargetWindows() &&
+ CS.paramHasAttr(1, Attribute::StructRet))
+ NumBytesCallee = 4;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(AdjStackUp))
- .addImm(NumBytes).addImm(0);
+ .addImm(NumBytes).addImm(NumBytesCallee);
+
+ // Build info for return calling conv lowering code.
+ // FIXME: This is practically a copy-paste from TargetLowering::LowerCallTo.
+ SmallVector<ISD::InputArg, 32> Ins;
+ SmallVector<EVT, 4> RetTys;
+ ComputeValueVTs(TLI, I->getType(), RetTys);
+ for (unsigned i = 0, e = RetTys.size(); i != e; ++i) {
+ EVT VT = RetTys[i];
+ EVT RegisterVT = TLI.getRegisterType(I->getParent()->getContext(), VT);
+ unsigned NumRegs = TLI.getNumRegisters(I->getParent()->getContext(), VT);
+ for (unsigned j = 0; j != NumRegs; ++j) {
+ ISD::InputArg MyFlags;
+ MyFlags.VT = RegisterVT.getSimpleVT();
+ MyFlags.Used = !CS.getInstruction()->use_empty();
+ if (CS.paramHasAttr(0, Attribute::SExt))
+ MyFlags.Flags.setSExt();
+ if (CS.paramHasAttr(0, Attribute::ZExt))
+ MyFlags.Flags.setZExt();
+ if (CS.paramHasAttr(0, Attribute::InReg))
+ MyFlags.Flags.setInReg();
+ Ins.push_back(MyFlags);
+ }
+ }
- // Now handle call return value (if any).
+ // Now handle call return values.
SmallVector<unsigned, 4> UsedRegs;
- if (RetVT != MVT::isVoid) {
- SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CC, false, TM, RVLocs, I->getParent()->getContext());
- CCInfo.AnalyzeCallResult(RetVT, RetCC_X86);
-
- // Copy all of the result registers out of their specified physreg.
- assert(RVLocs.size() == 1 && "Can't handle multi-value calls!");
- EVT CopyVT = RVLocs[0].getValVT();
- TargetRegisterClass* DstRC = TLI.getRegClassFor(CopyVT);
+ SmallVector<CCValAssign, 16> RVLocs;
+ CCState CCRetInfo(CC, false, *FuncInfo.MF, TM, RVLocs,
+ I->getParent()->getContext());
+ unsigned ResultReg = FuncInfo.CreateRegs(I->getType());
+ CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86);
+ for (unsigned i = 0; i != RVLocs.size(); ++i) {
+ EVT CopyVT = RVLocs[i].getValVT();
+ unsigned CopyReg = ResultReg + i;
// If this is a call to a function that returns an fp value on the x87 fp
// stack, but where we prefer to use the value in xmm registers, copy it
// out as F80 and use a truncate to move it from fp stack reg to xmm reg.
- if ((RVLocs[0].getLocReg() == X86::ST0 ||
- RVLocs[0].getLocReg() == X86::ST1) &&
- isScalarFPTypeInSSEReg(RVLocs[0].getValVT())) {
- CopyVT = MVT::f80;
- DstRC = X86::RFP80RegisterClass;
+ if ((RVLocs[i].getLocReg() == X86::ST0 ||
+ RVLocs[i].getLocReg() == X86::ST1)) {
+ if (isScalarFPTypeInSSEReg(RVLocs[i].getValVT())) {
+ CopyVT = MVT::f80;
+ CopyReg = createResultReg(&X86::RFP80RegClass);
+ }
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(X86::FpPOP_RETVAL),
+ CopyReg);
+ } else {
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+ CopyReg).addReg(RVLocs[i].getLocReg());
+ UsedRegs.push_back(RVLocs[i].getLocReg());
}
- unsigned ResultReg = createResultReg(DstRC);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- ResultReg).addReg(RVLocs[0].getLocReg());
- UsedRegs.push_back(RVLocs[0].getLocReg());
-
- if (CopyVT != RVLocs[0].getValVT()) {
+ if (CopyVT != RVLocs[i].getValVT()) {
// Round the F80 the right size, which also moves to the appropriate xmm
// register. This is accomplished by storing the F80 value in memory and
// then loading it back. Ewww...
- EVT ResVT = RVLocs[0].getValVT();
+ EVT ResVT = RVLocs[i].getValVT();
unsigned Opc = ResVT == MVT::f32 ? X86::ST_Fp80m32 : X86::ST_Fp80m64;
unsigned MemSize = ResVT.getSizeInBits()/8;
int FI = MFI.CreateStackObject(MemSize, MemSize, false);
addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc)), FI)
- .addReg(ResultReg);
- DstRC = ResVT == MVT::f32
- ? X86::FR32RegisterClass : X86::FR64RegisterClass;
+ .addReg(CopyReg);
Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
- ResultReg = createResultReg(DstRC);
addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(Opc), ResultReg), FI);
- }
-
- if (AndToI1) {
- // Mask out all but lowest bit for some call which produces an i1.
- unsigned AndResult = createResultReg(X86::GR8RegisterClass);
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
- TII.get(X86::AND8ri), AndResult).addReg(ResultReg).addImm(1);
- ResultReg = AndResult;
+ TII.get(Opc), ResultReg + i), FI);
}
-
- UpdateValueMap(I, ResultReg);
}
+ if (RVLocs.size())
+ UpdateValueMap(I, ResultReg, RVLocs.size());
+
// Set all unused physreg defs as dead.
static_cast<MachineInstr *>(MIB)->setPhysRegsDeadExcept(UsedRegs, TRI);
return X86SelectFPExt(I);
case Instruction::FPTrunc:
return X86SelectFPTrunc(I);
- case Instruction::ExtractValue:
- return X86SelectExtractValue(I);
case Instruction::IntToPtr: // Deliberate fall-through.
case Instruction::PtrToInt: {
EVT SrcVT = TLI.getValueType(I->getOperand(0)->getType());
default: return false;
case MVT::i8:
Opc = X86::MOV8rm;
- RC = X86::GR8RegisterClass;
+ RC = &X86::GR8RegClass;
break;
case MVT::i16:
Opc = X86::MOV16rm;
- RC = X86::GR16RegisterClass;
+ RC = &X86::GR16RegClass;
break;
case MVT::i32:
Opc = X86::MOV32rm;
- RC = X86::GR32RegisterClass;
+ RC = &X86::GR32RegClass;
break;
case MVT::i64:
// Must be in x86-64 mode.
Opc = X86::MOV64rm;
- RC = X86::GR64RegisterClass;
+ RC = &X86::GR64RegClass;
break;
case MVT::f32:
- if (Subtarget->hasSSE1()) {
- Opc = X86::MOVSSrm;
- RC = X86::FR32RegisterClass;
+ if (X86ScalarSSEf32) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
+ RC = &X86::FR32RegClass;
} else {
Opc = X86::LD_Fp32m;
- RC = X86::RFP32RegisterClass;
+ RC = &X86::RFP32RegClass;
}
break;
case MVT::f64:
- if (Subtarget->hasSSE2()) {
- Opc = X86::MOVSDrm;
- RC = X86::FR64RegisterClass;
+ if (X86ScalarSSEf64) {
+ Opc = Subtarget->hasAVX() ? X86::VMOVSDrm : X86::MOVSDrm;
+ RC = &X86::FR64RegClass;
} else {
Opc = X86::LD_Fp64m;
- RC = X86::RFP64RegisterClass;
+ RC = &X86::RFP64RegClass;
}
break;
case MVT::f80:
if (AM.BaseType == X86AddressMode::RegBase &&
AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == 0)
return AM.Base.Reg;
-
+
Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r;
unsigned ResultReg = createResultReg(RC);
addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
if (!X86SelectAddress(C, AM))
return 0;
unsigned Opc = Subtarget->is64Bit() ? X86::LEA64r : X86::LEA32r;
- TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
+ const TargetRegisterClass* RC = TLI.getRegClassFor(TLI.getPointerTy());
unsigned ResultReg = createResultReg(RC);
addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
TII.get(Opc), ResultReg), AM);
return ResultReg;
}
+unsigned X86FastISel::TargetMaterializeFloatZero(const ConstantFP *CF) {
+ MVT VT;
+ if (!isTypeLegal(CF->getType(), VT))
+ return false;
+
+ // Get opcode and regclass for the given zero.
+ unsigned Opc = 0;
+ const TargetRegisterClass *RC = NULL;
+ switch (VT.SimpleTy) {
+ default: return false;
+ case MVT::f32:
+ if (X86ScalarSSEf32) {
+ Opc = X86::FsFLD0SS;
+ RC = &X86::FR32RegClass;
+ } else {
+ Opc = X86::LD_Fp032;
+ RC = &X86::RFP32RegClass;
+ }
+ break;
+ case MVT::f64:
+ if (X86ScalarSSEf64) {
+ Opc = X86::FsFLD0SD;
+ RC = &X86::FR64RegClass;
+ } else {
+ Opc = X86::LD_Fp064;
+ RC = &X86::RFP64RegClass;
+ }
+ break;
+ case MVT::f80:
+ // No f80 support yet.
+ return false;
+ }
+
+ unsigned ResultReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(Opc), ResultReg);
+ return ResultReg;
+}
+
+
/// TryToFoldLoad - The specified machine instr operand is a vreg, and that
/// vreg is being provided by the specified load instruction. If possible,
/// try to fold the load as an operand to the instruction, returning true if
namespace llvm {
- llvm::FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
+ FastISel *X86::createFastISel(FunctionLoweringInfo &funcInfo) {
return new X86FastISel(funcInfo);
}
}
projects / oota-llvm.git / blobdiff