bool X86SelectFPTrunc(const Instruction *I);
const X86InstrInfo *getInstrInfo() const {
- return getTargetMachine()->getInstrInfo();
+ return getTargetMachine()->getSubtargetImpl()->getInstrInfo();
}
const X86TargetMachine *getTargetMachine() const {
return static_cast<const X86TargetMachine *>(&TM);
bool handleConstantAddresses(const Value *V, X86AddressMode &AM);
+ unsigned X86MaterializeInt(const ConstantInt *CI, MVT VT);
+ unsigned X86MaterializeFP(const ConstantFP *CFP, MVT VT);
+ unsigned X86MaterializeGV(const GlobalValue *GV,MVT VT);
unsigned TargetMaterializeConstant(const Constant *C) override;
unsigned TargetMaterializeAlloca(const AllocaInst *C) override;
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ValLocs;
- CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs,
- I->getContext());
+ CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, I->getContext());
CCInfo.AnalyzeReturn(Outs, RetCC_X86);
const Value *RV = Ret->getOperand(0);
// The calling-convention tables for x87 returns don't tell
// the whole story.
- if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1)
+ if (VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1)
return false;
unsigned SrcReg = Reg + VA.getValNo();
MachineFrameInfo *MFI = FuncInfo.MF->getFrameInfo();
MFI->setFrameAddressIsTaken(true);
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ TM.getSubtargetImpl()->getRegisterInfo());
unsigned FrameReg = RegInfo->getFrameRegister(*(FuncInfo.MF));
assert(((FrameReg == X86::RBP && VT == MVT::i64) ||
(FrameReg == X86::EBP && VT == MVT::i32)) &&
isCommutativeIntrinsic(II))
std::swap(LHS, RHS);
+ bool UseIncDec = false;
+ if (isa<ConstantInt>(RHS) && cast<ConstantInt>(RHS)->isOne())
+ UseIncDec = true;
+
unsigned BaseOpc, CondOpc;
switch (II->getIntrinsicID()) {
default: llvm_unreachable("Unexpected intrinsic!");
case Intrinsic::sadd_with_overflow:
- BaseOpc = ISD::ADD; CondOpc = X86::SETOr; break;
+ BaseOpc = UseIncDec ? unsigned(X86ISD::INC) : unsigned(ISD::ADD);
+ CondOpc = X86::SETOr;
+ break;
case Intrinsic::uadd_with_overflow:
BaseOpc = ISD::ADD; CondOpc = X86::SETBr; break;
case Intrinsic::ssub_with_overflow:
- BaseOpc = ISD::SUB; CondOpc = X86::SETOr; break;
+ BaseOpc = UseIncDec ? unsigned(X86ISD::DEC) : unsigned(ISD::SUB);
+ CondOpc = X86::SETOr;
+ break;
case Intrinsic::usub_with_overflow:
BaseOpc = ISD::SUB; CondOpc = X86::SETBr; break;
case Intrinsic::smul_with_overflow:
unsigned ResultReg = 0;
// Check if we have an immediate version.
- if (auto const *C = dyn_cast<ConstantInt>(RHS)) {
- ResultReg = FastEmit_ri(VT, VT, BaseOpc, LHSReg, LHSIsKill,
- C->getZExtValue());
+ if (const auto *CI = dyn_cast<ConstantInt>(RHS)) {
+ static const unsigned Opc[2][2][4] = {
+ { { X86::INC8r, X86::INC16r, X86::INC32r, X86::INC64r },
+ { X86::DEC8r, X86::DEC16r, X86::DEC32r, X86::DEC64r } },
+ { { X86::INC8r, X86::INC64_16r, X86::INC64_32r, X86::INC64r },
+ { X86::DEC8r, X86::DEC64_16r, X86::DEC64_32r, X86::DEC64r } }
+ };
+
+ if (BaseOpc == X86ISD::INC || BaseOpc == X86ISD::DEC) {
+ ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ bool Is64Bit = Subtarget->is64Bit();
+ bool IsDec = BaseOpc == X86ISD::DEC;
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(Opc[Is64Bit][IsDec][VT.SimpleTy-MVT::i8]), ResultReg)
+ .addReg(LHSReg, getKillRegState(LHSIsKill));
+ } else
+ ResultReg = FastEmit_ri(VT, VT, BaseOpc, LHSReg, LHSIsKill,
+ CI->getZExtValue());
}
unsigned RHSReg;
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, TM, ArgLocs,
- CLI.RetTy->getContext());
+ CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, ArgLocs, CLI.RetTy->getContext());
// Allocate shadow area for Win64
if (IsWin64)
.addImm(NumBytes);
// Walk the register/memloc assignments, inserting copies/loads.
- const X86RegisterInfo *RegInfo =
- static_cast<const X86RegisterInfo *>(TM.getRegisterInfo());
+ const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
+ TM.getSubtargetImpl()->getRegisterInfo());
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
CCValAssign const &VA = ArgLocs[i];
const Value *ArgVal = OutVals[VA.getValNo()];
OutRegs.push_back(VA.getLocReg());
} else {
assert(VA.isMemLoc());
+
+ // Don't emit stores for undef values.
+ if (isa<UndefValue>(ArgVal))
+ continue;
+
unsigned LocMemOffset = VA.getLocMemOffset();
X86AddressMode AM;
AM.Base.Reg = RegInfo->getStackRegister();
// Now handle call return values.
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCRetInfo(CC, IsVarArg, *FuncInfo.MF, TM, RVLocs,
+ CCState CCRetInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs,
CLI.RetTy->getContext());
CCRetInfo.AnalyzeCallResult(Ins, RetCC_X86);
report_fatal_error("SSE register return with SSE disabled");
}
- // If this is a call to a function that returns an fp value on the floating
- // point stack, we must guarantee the value is popped from the stack, so
- // a COPY is not good enough - the copy instruction may be eliminated if the
- // return value is not used. We use the FpPOP_RETVAL instruction instead.
- if (VA.getLocReg() == X86::ST0 || VA.getLocReg() == X86::ST1) {
- // If 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 (isScalarFPTypeInSSEReg(VA.getValVT())) {
- CopyVT = MVT::f80;
- CopyReg = createResultReg(&X86::RFP80RegClass);
- }
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
- TII.get(X86::FpPOP_RETVAL), CopyReg);
-
- // Round the f80 to the right size, which also moves it to the appropriate
- // xmm register. This is accomplished by storing the f80 value in memory
- // and then loading it back.
- if (CopyVT != VA.getValVT()) {
- EVT ResVT = VA.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, DbgLoc,
- TII.get(Opc)), FI)
- .addReg(CopyReg);
- Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
- addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
- TII.get(Opc), ResultReg + i), FI);
- }
- } else {
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
- TII.get(TargetOpcode::COPY), CopyReg).addReg(VA.getLocReg());
- InRegs.push_back(VA.getLocReg());
+ // If 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 ((VA.getLocReg() == X86::FP0 || VA.getLocReg() == X86::FP1) &&
+ isScalarFPTypeInSSEReg(VA.getValVT())) {
+ CopyVT = MVT::f80;
+ CopyReg = createResultReg(&X86::RFP80RegClass);
+ }
+
+ // Copy out the result.
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(TargetOpcode::COPY), CopyReg).addReg(VA.getLocReg());
+ InRegs.push_back(VA.getLocReg());
+
+ // Round the f80 to the right size, which also moves it to the appropriate
+ // xmm register. This is accomplished by storing the f80 value in memory
+ // and then loading it back.
+ if (CopyVT != VA.getValVT()) {
+ EVT ResVT = VA.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, DbgLoc,
+ TII.get(Opc)), FI)
+ .addReg(CopyReg);
+ Opc = ResVT == MVT::f32 ? X86::MOVSSrm : X86::MOVSDrm;
+ addFrameReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(Opc), ResultReg + i), FI);
}
}
return false;
}
-unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
- MVT VT;
- if (!isTypeLegal(C->getType(), VT))
+unsigned X86FastISel::X86MaterializeInt(const ConstantInt *CI, MVT VT) {
+ if (VT > MVT::i64)
return 0;
+ return FastEmit_i(VT, VT, ISD::Constant, CI->getZExtValue());
+}
+
+unsigned X86FastISel::X86MaterializeFP(const ConstantFP *CFP, MVT VT) {
+ if (CFP->isNullValue())
+ return TargetMaterializeFloatZero(CFP);
// Can't handle alternate code models yet.
if (TM.getCodeModel() != CodeModel::Small)
const TargetRegisterClass *RC = nullptr;
switch (VT.SimpleTy) {
default: return 0;
- case MVT::i8:
- Opc = X86::MOV8rm;
- RC = &X86::GR8RegClass;
- break;
- case MVT::i16:
- Opc = X86::MOV16rm;
- RC = &X86::GR16RegClass;
- break;
- case MVT::i32:
- Opc = X86::MOV32rm;
- RC = &X86::GR32RegClass;
- break;
- case MVT::i64:
- // Must be in x86-64 mode.
- Opc = X86::MOV64rm;
- RC = &X86::GR64RegClass;
- break;
case MVT::f32:
if (X86ScalarSSEf32) {
Opc = Subtarget->hasAVX() ? X86::VMOVSSrm : X86::MOVSSrm;
return 0;
}
- // Materialize addresses with LEA/MOV instructions.
- if (isa<GlobalValue>(C)) {
- X86AddressMode AM;
- if (X86SelectAddress(C, AM)) {
- // If the expression is just a basereg, then we're done, otherwise we need
- // to emit an LEA.
- if (AM.BaseType == X86AddressMode::RegBase &&
- AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr)
- return AM.Base.Reg;
-
- unsigned ResultReg = createResultReg(RC);
- if (TM.getRelocationModel() == Reloc::Static &&
- TLI.getPointerTy() == MVT::i64) {
- // The displacement code be more than 32 bits away so we need to use
- // an instruction with a 64 bit immediate
- Opc = X86::MOV64ri;
- BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
- TII.get(Opc), ResultReg).addGlobalAddress(cast<GlobalValue>(C));
- } else {
- Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r;
- addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
- TII.get(Opc), ResultReg), AM);
- }
- return ResultReg;
- }
- return 0;
- }
-
// MachineConstantPool wants an explicit alignment.
- unsigned Align = DL.getPrefTypeAlignment(C->getType());
+ unsigned Align = DL.getPrefTypeAlignment(CFP->getType());
if (Align == 0) {
- // Alignment of vector types. FIXME!
- Align = DL.getTypeAllocSize(C->getType());
+ // Alignment of vector types. FIXME!
+ Align = DL.getTypeAllocSize(CFP->getType());
}
// x86-32 PIC requires a PIC base register for constant pools.
}
// Create the load from the constant pool.
- unsigned MCPOffset = MCP.getConstantPoolIndex(C, Align);
+ unsigned CPI = MCP.getConstantPoolIndex(CFP, Align);
unsigned ResultReg = createResultReg(RC);
+
addConstantPoolReference(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(Opc), ResultReg),
- MCPOffset, PICBase, OpFlag);
-
+ CPI, PICBase, OpFlag);
return ResultReg;
}
+unsigned X86FastISel::X86MaterializeGV(const GlobalValue *GV, MVT VT) {
+ // Can't handle alternate code models yet.
+ if (TM.getCodeModel() != CodeModel::Small)
+ return 0;
+
+ // Materialize addresses with LEA/MOV instructions.
+ X86AddressMode AM;
+ if (X86SelectAddress(GV, AM)) {
+ // If the expression is just a basereg, then we're done, otherwise we need
+ // to emit an LEA.
+ if (AM.BaseType == X86AddressMode::RegBase &&
+ AM.IndexReg == 0 && AM.Disp == 0 && AM.GV == nullptr)
+ return AM.Base.Reg;
+
+ unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ if (TM.getRelocationModel() == Reloc::Static &&
+ TLI.getPointerTy() == MVT::i64) {
+ // The displacement code could be more than 32 bits away so we need to use
+ // an instruction with a 64 bit immediate
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(X86::MOV64ri),
+ ResultReg)
+ .addGlobalAddress(GV);
+ } else {
+ unsigned Opc = TLI.getPointerTy() == MVT::i32 ? X86::LEA32r : X86::LEA64r;
+ addFullAddress(BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
+ TII.get(Opc), ResultReg), AM);
+ }
+ return ResultReg;
+ }
+ return 0;
+}
+
+unsigned X86FastISel::TargetMaterializeConstant(const Constant *C) {
+ EVT CEVT = TLI.getValueType(C->getType(), true);
+
+ // Only handle simple types.
+ if (!CEVT.isSimple())
+ return 0;
+ MVT VT = CEVT.getSimpleVT();
+
+ if (const auto *CI = dyn_cast<ConstantInt>(C))
+ return X86MaterializeInt(CI, VT);
+ else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(C))
+ return X86MaterializeFP(CFP, VT);
+ else if (const GlobalValue *GV = dyn_cast<GlobalValue>(C))
+ return X86MaterializeGV(GV, VT);
+
+ return 0;
+}
+
unsigned X86FastISel::TargetMaterializeAlloca(const AllocaInst *C) {
// Fail on dynamic allocas. At this point, getRegForValue has already
// checked its CSE maps, so if we're here trying to handle a dynamic
projects / oota-llvm.git / blobdiff