//===----------------------------------------------------------------------===//
//
// TBD:
-// FastLowerArguments: Handle simple cases.
+// fastLowerArguments: Handle simple cases.
// PPCMaterializeGV: Handle TLS.
// SelectCall: Handle function pointers.
// SelectCall: Handle multi-register return values.
const TargetMachine &TM;
const TargetInstrInfo &TII;
const TargetLowering &TLI;
- const PPCSubtarget &PPCSubTarget;
+ const PPCSubtarget *PPCSubTarget;
LLVMContext *Context;
public:
explicit PPCFastISel(FunctionLoweringInfo &FuncInfo,
const TargetLibraryInfo *LibInfo)
- : FastISel(FuncInfo, LibInfo),
- TM(FuncInfo.MF->getTarget()),
- TII(*TM.getInstrInfo()),
- TLI(*TM.getTargetLowering()),
- PPCSubTarget(
- *((static_cast<const PPCTargetMachine *>(&TM))->getSubtargetImpl())
- ),
- Context(&FuncInfo.Fn->getContext()) { }
+ : FastISel(FuncInfo, LibInfo), TM(FuncInfo.MF->getTarget()),
+ TII(*TM.getSubtargetImpl()->getInstrInfo()),
+ TLI(*TM.getSubtargetImpl()->getTargetLowering()),
+ PPCSubTarget(&TM.getSubtarget<PPCSubtarget>()),
+ Context(&FuncInfo.Fn->getContext()) {}
// Backend specific FastISel code.
private:
- virtual bool TargetSelectInstruction(const Instruction *I);
- virtual unsigned TargetMaterializeConstant(const Constant *C);
- virtual unsigned TargetMaterializeAlloca(const AllocaInst *AI);
- virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
- const LoadInst *LI);
- virtual bool FastLowerArguments();
- virtual unsigned FastEmit_i(MVT Ty, MVT RetTy, unsigned Opc, uint64_t Imm);
- virtual unsigned FastEmitInst_ri(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- uint64_t Imm);
- virtual unsigned FastEmitInst_r(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill);
- virtual unsigned FastEmitInst_rr(unsigned MachineInstOpcode,
- const TargetRegisterClass *RC,
- unsigned Op0, bool Op0IsKill,
- unsigned Op1, bool Op1IsKill);
+ bool fastSelectInstruction(const Instruction *I) override;
+ unsigned fastMaterializeConstant(const Constant *C) override;
+ unsigned fastMaterializeAlloca(const AllocaInst *AI) override;
+ bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
+ const LoadInst *LI) override;
+ bool fastLowerArguments() override;
+ unsigned fastEmit_i(MVT Ty, MVT RetTy, unsigned Opc, uint64_t Imm) override;
+ unsigned fastEmitInst_ri(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ uint64_t Imm);
+ unsigned fastEmitInst_r(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill);
+ unsigned fastEmitInst_rr(unsigned MachineInstOpcode,
+ const TargetRegisterClass *RC,
+ unsigned Op0, bool Op0IsKill,
+ unsigned Op1, bool Op1IsKill);
// Instruction selection routines.
private:
unsigned DestReg, bool IsZExt);
unsigned PPCMaterializeFP(const ConstantFP *CFP, MVT VT);
unsigned PPCMaterializeGV(const GlobalValue *GV, MVT VT);
- unsigned PPCMaterializeInt(const Constant *C, MVT VT);
+ unsigned PPCMaterializeInt(const Constant *C, MVT VT, bool UseSExt = true);
unsigned PPCMaterialize32BitInt(int64_t Imm,
const TargetRegisterClass *RC);
unsigned PPCMaterialize64BitInt(int64_t Imm,
unsigned ResultReg = 0;
if (!PPCEmitLoad(VT, ResultReg, Addr, RC))
return false;
- UpdateValueMap(I, ResultReg);
+ updateValueMap(I, ResultReg);
return true;
}
BuildMI(*BrBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::BCC))
.addImm(PPCPred).addReg(CondReg).addMBB(TBB);
- FastEmitBranch(FBB, DbgLoc);
+ fastEmitBranch(FBB, DbgLoc);
FuncInfo.MBB->addSuccessor(TBB);
return true;
dyn_cast<ConstantInt>(BI->getCondition())) {
uint64_t Imm = CI->getZExtValue();
MachineBasicBlock *Target = (Imm == 0) ? FBB : TBB;
- FastEmitBranch(Target, DbgLoc);
+ fastEmitBranch(Target, DbgLoc);
return true;
}
return false;
MVT SrcVT = SrcEVT.getSimpleVT();
- if (SrcVT == MVT::i1 && PPCSubTarget.useCRBits())
+ if (SrcVT == MVT::i1 && PPCSubTarget->useCRBits())
return false;
// See if operand 2 is an immediate encodeable in the compare.
return false;
// No code is generated for a FP extend.
- UpdateValueMap(I, SrcReg);
+ updateValueMap(I, SrcReg);
return true;
}
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::FRSP), DestReg)
.addReg(SrcReg);
- UpdateValueMap(I, DestReg);
+ updateValueMap(I, DestReg);
return true;
}
if (!IsSigned) {
LoadOpc = PPC::LFIWZX;
Addr.Offset = 4;
- } else if (PPCSubTarget.hasLFIWAX()) {
+ } else if (PPCSubTarget->hasLFIWAX()) {
LoadOpc = PPC::LFIWAX;
Addr.Offset = 4;
}
// We can only lower an unsigned convert if we have the newer
// floating-point conversion operations.
- if (!IsSigned && !PPCSubTarget.hasFPCVT())
+ if (!IsSigned && !PPCSubTarget->hasFPCVT())
return false;
// FIXME: For now we require the newer floating-point conversion operations
// to single-precision float. Otherwise we have to generate a lot of
// fiddly code to avoid double rounding. If necessary, the fiddly code
// can be found in PPCTargetLowering::LowerINT_TO_FP().
- if (DstVT == MVT::f32 && !PPCSubTarget.hasFPCVT())
+ if (DstVT == MVT::f32 && !PPCSubTarget->hasFPCVT())
return false;
// Extend the input if necessary.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), DestReg)
.addReg(FPReg);
- UpdateValueMap(I, DestReg);
+ updateValueMap(I, DestReg);
return true;
}
if (DstVT != MVT::i32 && DstVT != MVT::i64)
return false;
+ // If we don't have FCTIDUZ and we need it, punt to SelectionDAG.
+ if (DstVT == MVT::i64 && !IsSigned && !PPCSubTarget->hasFPCVT())
+ return false;
+
Value *Src = I->getOperand(0);
Type *SrcTy = Src->getType();
if (!isTypeLegal(SrcTy, SrcVT))
if (IsSigned)
Opc = PPC::FCTIWZ;
else
- Opc = PPCSubTarget.hasFPCVT() ? PPC::FCTIWUZ : PPC::FCTIDZ;
+ Opc = PPCSubTarget->hasFPCVT() ? PPC::FCTIWUZ : PPC::FCTIDZ;
else
Opc = IsSigned ? PPC::FCTIDZ : PPC::FCTIDUZ;
if (IntReg == 0)
return false;
- UpdateValueMap(I, IntReg);
+ updateValueMap(I, IntReg);
return true;
}
ResultReg)
.addReg(SrcReg1)
.addImm(Imm);
- UpdateValueMap(I, ResultReg);
+ updateValueMap(I, ResultReg);
return true;
}
}
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ResultReg)
.addReg(SrcReg1).addReg(SrcReg2);
- UpdateValueMap(I, ResultReg);
+ updateValueMap(I, ResultReg);
return true;
}
unsigned &NumBytes,
bool IsVarArg) {
SmallVector<CCValAssign, 16> ArgLocs;
- CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, TM, ArgLocs, *Context);
+ CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, ArgLocs, *Context);
+
+ // Reserve space for the linkage area on the stack.
+ bool isELFv2ABI = PPCSubTarget->isELFv2ABI();
+ unsigned LinkageSize = PPCFrameLowering::getLinkageSize(true, false,
+ isELFv2ABI);
+ CCInfo.AllocateStack(LinkageSize, 8);
+
CCInfo.AnalyzeCallOperands(ArgVTs, ArgFlags, CC_PPC64_ELF_FIS);
// Bail out if we can't handle any of the arguments.
// Get a count of how many bytes are to be pushed onto the stack.
NumBytes = CCInfo.getNextStackOffset();
+ // The prolog code of the callee may store up to 8 GPR argument registers to
+ // the stack, allowing va_start to index over them in memory if its varargs.
+ // Because we cannot tell if this is needed on the caller side, we have to
+ // conservatively assume that it is needed. As such, make sure we have at
+ // least enough stack space for the caller to store the 8 GPRs.
+ // FIXME: On ELFv2, it may be unnecessary to allocate the parameter area.
+ NumBytes = std::max(NumBytes, LinkageSize + 64);
+
// Issue CALLSEQ_START.
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(TII.getCallFrameSetupOpcode()))
// any real difficulties there.
if (RetVT != MVT::isVoid) {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs, *Context);
CCInfo.AnalyzeCallResult(RetVT, RetCC_PPC64_ELF_FIS);
CCValAssign &VA = RVLocs[0];
assert(RVLocs.size() == 1 && "No support for multi-reg return values!");
assert(ResultReg && "ResultReg unset!");
UsedRegs.push_back(SourcePhysReg);
- UpdateValueMap(I, ResultReg);
+ updateValueMap(I, ResultReg);
}
}
RetVT != MVT::i32 && RetVT != MVT::i64 && RetVT != MVT::f32 &&
RetVT != MVT::f64) {
SmallVector<CCValAssign, 16> RVLocs;
- CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, TM, RVLocs, *Context);
+ CCState CCInfo(CC, IsVarArg, *FuncInfo.MF, RVLocs, *Context);
CCInfo.AnalyzeCallResult(RetVT, RetCC_PPC64_ELF_FIS);
if (RVLocs.size() > 1)
return false;
for (unsigned II = 0, IE = RegArgs.size(); II != IE; ++II)
MIB.addReg(RegArgs[II], RegState::Implicit);
+ // Direct calls in the ELFv2 ABI need the TOC register live into the call.
+ if (PPCSubTarget->isELFv2ABI())
+ MIB.addReg(PPC::X2, RegState::Implicit);
+
// Add a register mask with the call-preserved registers. Proper
// defs for return values will be added by setPhysRegsDeadExcept().
MIB.addRegMask(TRI.getCallPreservedMask(CC));
// Analyze operands of the call, assigning locations to each operand.
SmallVector<CCValAssign, 16> ValLocs;
- CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, TM, ValLocs, *Context);
+ CCState CCInfo(CC, F.isVarArg(), *FuncInfo.MF, ValLocs, *Context);
CCInfo.AnalyzeReturn(Outs, RetCC_PPC64_ELF_FIS);
const Value *RV = Ret->getOperand(0);
// Special case for returning a constant integer of any size.
// Materialize the constant as an i64 and copy it to the return
- // register. This avoids an unnecessary extend or truncate.
+ // register. We still need to worry about properly extending the sign. E.g:
+ // If the constant has only one bit, it means it is a boolean. Therefore
+ // we can't use PPCMaterializeInt because it extends the sign which will
+ // cause negations of the returned value to be incorrect as they are
+ // implemented as the flip of the least significant bit.
if (isa<ConstantInt>(*RV)) {
const Constant *C = cast<Constant>(RV);
- unsigned SrcReg = PPCMaterializeInt(C, MVT::i64);
- unsigned RetReg = ValLocs[0].getLocReg();
+
+ CCValAssign &VA = ValLocs[0];
+
+ unsigned RetReg = VA.getLocReg();
+ unsigned SrcReg = PPCMaterializeInt(C, MVT::i64,
+ VA.getLocInfo() == CCValAssign::SExt);
+
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
- TII.get(TargetOpcode::COPY), RetReg).addReg(SrcReg);
+ TII.get(TargetOpcode::COPY), RetReg).addReg(SrcReg);
+
RetRegs.push_back(RetReg);
} else {
SrcReg = ResultReg;
}
- UpdateValueMap(I, SrcReg);
+ updateValueMap(I, SrcReg);
return true;
}
if (!PPCEmitIntExt(SrcVT, SrcReg, DestVT, ResultReg, IsZExt))
return false;
- UpdateValueMap(I, ResultReg);
+ updateValueMap(I, ResultReg);
return true;
}
// Attempt to fast-select an instruction that wasn't handled by
// the table-generated machinery.
-bool PPCFastISel::TargetSelectInstruction(const Instruction *I) {
+bool PPCFastISel::fastSelectInstruction(const Instruction *I) {
switch (I->getOpcode()) {
case Instruction::Load:
// FIXME: Jump tables are not yet required because fast-isel doesn't
// handle switches; if that changes, we need them as well. For now,
// what follows assumes everything's a generic (or TLS) global address.
- const GlobalVariable *GVar = dyn_cast<GlobalVariable>(GV);
- if (!GVar) {
- // If GV is an alias, use the aliasee for determining thread-locality.
- if (const GlobalAlias *GA = dyn_cast<GlobalAlias>(GV))
- GVar = dyn_cast_or_null<GlobalVariable>(GA->getAliasedGlobal());
- }
// FIXME: We don't yet handle the complexity of TLS.
- bool IsTLS = GVar && GVar->isThreadLocal();
- if (IsTLS)
+ if (GV->isThreadLocal())
return 0;
// For small code model, generate a simple TOC load.
.addGlobalAddress(GV)
.addReg(PPC::X2);
else {
- // If the address is an externally defined symbol, a symbol with
- // common or externally available linkage, a function address, or a
+ // If the address is an externally defined symbol, a symbol with common
+ // or externally available linkage, a non-local function address, or a
// jump table address (not yet needed), or if we are generating code
// for large code model, we generate:
// LDtocL(GV, ADDIStocHA(%X2, GV))
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::ADDIStocHA),
HighPartReg).addReg(PPC::X2).addGlobalAddress(GV);
- // !GVar implies a function address. An external variable is one
- // without an initializer.
// If/when switches are implemented, jump tables should be handled
// on the "if" path here.
- if (CModel == CodeModel::Large || !GVar || !GVar->hasInitializer() ||
- GVar->hasCommonLinkage() || GVar->hasAvailableExternallyLinkage())
+ if (CModel == CodeModel::Large ||
+ (GV->getType()->getElementType()->isFunctionTy() &&
+ (GV->isDeclaration() || GV->isWeakForLinker())) ||
+ GV->isDeclaration() || GV->hasCommonLinkage() ||
+ GV->hasAvailableExternallyLinkage())
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(PPC::LDtocL),
DestReg).addGlobalAddress(GV).addReg(HighPartReg);
else
// Materialize an integer constant into a register, and return
// the register number (or zero if we failed to handle it).
-unsigned PPCFastISel::PPCMaterializeInt(const Constant *C, MVT VT) {
+unsigned PPCFastISel::PPCMaterializeInt(const Constant *C, MVT VT,
+ bool UseSExt) {
// If we're using CR bit registers for i1 values, handle that as a special
// case first.
- if (VT == MVT::i1 && PPCSubTarget.useCRBits()) {
+ if (VT == MVT::i1 && PPCSubTarget->useCRBits()) {
const ConstantInt *CI = cast<ConstantInt>(C);
unsigned ImmReg = createResultReg(&PPC::CRBITRCRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
unsigned Opc = (VT == MVT::i64) ? PPC::LI8 : PPC::LI;
unsigned ImmReg = createResultReg(RC);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(Opc), ImmReg)
- .addImm(CI->getSExtValue());
+ .addImm( (UseSExt) ? CI->getSExtValue() : CI->getZExtValue() );
return ImmReg;
}
// Materialize a constant into a register, and return the register
// number (or zero if we failed to handle it).
-unsigned PPCFastISel::TargetMaterializeConstant(const Constant *C) {
+unsigned PPCFastISel::fastMaterializeConstant(const Constant *C) {
EVT CEVT = TLI.getValueType(C->getType(), true);
// Only handle simple types.
// Materialize the address created by an alloca into a register, and
// return the register number (or zero if we failed to handle it).
-unsigned PPCFastISel::TargetMaterializeAlloca(const AllocaInst *AI) {
+unsigned PPCFastISel::fastMaterializeAlloca(const AllocaInst *AI) {
// Don't handle dynamic allocas.
if (!FuncInfo.StaticAllocaMap.count(AI)) return 0;
// Attempt to lower call arguments in a faster way than done by
// the selection DAG code.
-bool PPCFastISel::FastLowerArguments() {
+bool PPCFastISel::fastLowerArguments() {
// Defer to normal argument lowering for now. It's reasonably
// efficient. Consider doing something like ARM to handle the
// case where all args fit in registers, no varargs, no float
// Handle materializing integer constants into a register. This is not
// automatically generated for PowerPC, so must be explicitly created here.
-unsigned PPCFastISel::FastEmit_i(MVT Ty, MVT VT, unsigned Opc, uint64_t Imm) {
+unsigned PPCFastISel::fastEmit_i(MVT Ty, MVT VT, unsigned Opc, uint64_t Imm) {
if (Opc != ISD::Constant)
return 0;
// If we're using CR bit registers for i1 values, handle that as a special
// case first.
- if (VT == MVT::i1 && PPCSubTarget.useCRBits()) {
+ if (VT == MVT::i1 && PPCSubTarget->useCRBits()) {
unsigned ImmReg = createResultReg(&PPC::CRBITRCRegClass);
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
TII.get(Imm == 0 ? PPC::CRUNSET : PPC::CRSET), ImmReg);
// assigning R0 or X0 to the output register for GPRC and G8RC
// register classes, as any such result could be used in ADDI, etc.,
// where those regs have another meaning.
-unsigned PPCFastISel::FastEmitInst_ri(unsigned MachineInstOpcode,
+unsigned PPCFastISel::fastEmitInst_ri(unsigned MachineInstOpcode,
const TargetRegisterClass *RC,
unsigned Op0, bool Op0IsKill,
uint64_t Imm) {
(RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass :
(RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC));
- return FastISel::FastEmitInst_ri(MachineInstOpcode, UseRC,
+ return FastISel::fastEmitInst_ri(MachineInstOpcode, UseRC,
Op0, Op0IsKill, Imm);
}
// Override for instructions with one register operand to avoid use of
// R0/X0. The automatic infrastructure isn't aware of the context so
// we must be conservative.
-unsigned PPCFastISel::FastEmitInst_r(unsigned MachineInstOpcode,
+unsigned PPCFastISel::fastEmitInst_r(unsigned MachineInstOpcode,
const TargetRegisterClass* RC,
unsigned Op0, bool Op0IsKill) {
const TargetRegisterClass *UseRC =
(RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass :
(RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC));
- return FastISel::FastEmitInst_r(MachineInstOpcode, UseRC, Op0, Op0IsKill);
+ return FastISel::fastEmitInst_r(MachineInstOpcode, UseRC, Op0, Op0IsKill);
}
// Override for instructions with two register operands to avoid use
// of R0/X0. The automatic infrastructure isn't aware of the context
// so we must be conservative.
-unsigned PPCFastISel::FastEmitInst_rr(unsigned MachineInstOpcode,
+unsigned PPCFastISel::fastEmitInst_rr(unsigned MachineInstOpcode,
const TargetRegisterClass* RC,
unsigned Op0, bool Op0IsKill,
unsigned Op1, bool Op1IsKill) {
(RC == &PPC::GPRCRegClass ? &PPC::GPRC_and_GPRC_NOR0RegClass :
(RC == &PPC::G8RCRegClass ? &PPC::G8RC_and_G8RC_NOX0RegClass : RC));
- return FastISel::FastEmitInst_rr(MachineInstOpcode, UseRC, Op0, Op0IsKill,
+ return FastISel::fastEmitInst_rr(MachineInstOpcode, UseRC, Op0, Op0IsKill,
Op1, Op1IsKill);
}
projects / oota-llvm.git / blobdiff