#include "llvm/CallingConv.h"
#include "llvm/DerivedTypes.h"
#include "llvm/Instructions.h"
+#include "llvm/Intrinsics.h"
#include "llvm/CodeGen/FastISel.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
public:
explicit X86FastISel(MachineFunction &mf,
MachineModuleInfo *mmi,
+ DwarfWriter *dw,
DenseMap<const Value *, unsigned> &vm,
DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
DenseMap<const AllocaInst *, int> &am
, SmallSet<Instruction*, 8> &cil
#endif
)
- : FastISel(mf, mmi, vm, bm, am
+ : FastISel(mf, mmi, dw, vm, bm, am
#ifndef NDEBUG
, cil
#endif
bool X86FastEmitLoad(MVT VT, const X86AddressMode &AM, unsigned &RR);
+ bool X86FastEmitStore(MVT VT, Value *Val,
+ const X86AddressMode &AM);
bool X86FastEmitStore(MVT VT, unsigned Val,
const X86AddressMode &AM);
bool X86SelectFPExt(Instruction *I);
bool X86SelectFPTrunc(Instruction *I);
+ bool X86SelectExtractValue(Instruction *I);
+
+ bool X86VisitIntrinsicCall(CallInst &I, unsigned Intrinsic);
bool X86SelectCall(Instruction *I);
CCAssignFn *CCAssignFnForCall(unsigned CC, bool isTailCall = false);
const X86AddressMode &AM) {
// Get opcode and regclass of the output for the given store instruction.
unsigned Opc = 0;
- const TargetRegisterClass *RC = NULL;
switch (VT.getSimpleVT()) {
+ case MVT::f80: // No f80 support yet.
default: return false;
- case MVT::i8:
- Opc = X86::MOV8mr;
- RC = X86::GR8RegisterClass;
- break;
- case MVT::i16:
- Opc = X86::MOV16mr;
- RC = X86::GR16RegisterClass;
- break;
- case MVT::i32:
- Opc = X86::MOV32mr;
- RC = X86::GR32RegisterClass;
- break;
- case MVT::i64:
- // Must be in x86-64 mode.
- Opc = X86::MOV64mr;
- RC = X86::GR64RegisterClass;
- break;
+ case MVT::i8: Opc = X86::MOV8mr; break;
+ case MVT::i16: Opc = X86::MOV16mr; break;
+ case MVT::i32: Opc = X86::MOV32mr; break;
+ case MVT::i64: Opc = X86::MOV64mr; break; // Must be in x86-64 mode.
case MVT::f32:
- if (Subtarget->hasSSE1()) {
- Opc = X86::MOVSSmr;
- RC = X86::FR32RegisterClass;
- } else {
- Opc = X86::ST_Fp32m;
- RC = X86::RFP32RegisterClass;
- }
+ Opc = Subtarget->hasSSE1() ? X86::MOVSSmr : X86::ST_Fp32m;
break;
case MVT::f64:
- if (Subtarget->hasSSE2()) {
- Opc = X86::MOVSDmr;
- RC = X86::FR64RegisterClass;
- } else {
- Opc = X86::ST_Fp64m;
- RC = X86::RFP64RegisterClass;
- }
+ Opc = Subtarget->hasSSE2() ? X86::MOVSDmr : X86::ST_Fp64m;
break;
- case MVT::f80:
- // No f80 support yet.
- return false;
}
-
+
addFullAddress(BuildMI(MBB, TII.get(Opc)), AM).addReg(Val);
return true;
}
+bool X86FastISel::X86FastEmitStore(MVT VT, Value *Val,
+ const X86AddressMode &AM) {
+ // Handle 'null' like i32/i64 0.
+ if (isa<ConstantPointerNull>(Val))
+ Val = Constant::getNullValue(TD.getIntPtrType());
+
+ // If this is a store of a simple constant, fold the constant into the store.
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
+ unsigned Opc = 0;
+ switch (VT.getSimpleVT()) {
+ default: break;
+ case MVT::i8: Opc = X86::MOV8mi; break;
+ case MVT::i16: Opc = X86::MOV16mi; break;
+ case MVT::i32: Opc = X86::MOV32mi; break;
+ case MVT::i64:
+ // Must be a 32-bit sign extended value.
+ if ((int)CI->getSExtValue() == CI->getSExtValue())
+ Opc = X86::MOV64mi32;
+ break;
+ }
+
+ if (Opc) {
+ addFullAddress(BuildMI(MBB, TII.get(Opc)), AM).addImm(CI->getSExtValue());
+ return true;
+ }
+ }
+
+ unsigned ValReg = getRegForValue(Val);
+ if (ValReg == 0)
+ return false;
+
+ return X86FastEmitStore(VT, ValReg, AM);
+}
+
/// X86FastEmitExtend - Emit a machine instruction to extend a value Src of
/// type SrcVT to type DstVT using the specified extension opcode Opc (e.g.
/// ISD::SIGN_EXTEND).
// Look past no-op inttoptrs.
if (TLI.getValueType(U->getOperand(0)->getType()) == TLI.getPointerTy())
return X86SelectAddress(U->getOperand(0), AM, isCall);
+ break;
case Instruction::PtrToInt:
// Look past no-op ptrtoints.
if (TLI.getValueType(U->getType()) == TLI.getPointerTy())
return X86SelectAddress(U->getOperand(0), AM, isCall);
+ break;
case Instruction::Alloca: {
if (isCall) break;
unsigned IndexReg = AM.IndexReg;
unsigned Scale = AM.Scale;
gep_type_iterator GTI = gep_type_begin(U);
- // Look at all but the last index. Constants can be folded,
- // and one dynamic index can be handled, if the scale is supported.
+ // Iterate through the indices, folding what we can. Constants can be
+ // folded, and one dynamic index can be handled, if the scale is supported.
for (User::op_iterator i = U->op_begin() + 1, e = U->op_end();
i != e; ++i, ++GTI) {
Value *Op = *i;
unsigned Idx = cast<ConstantInt>(Op)->getZExtValue();
Disp += SL->getElementOffset(Idx);
} else {
- uint64_t S = TD.getABITypeSize(GTI.getIndexedType());
+ uint64_t S = TD.getTypePaddedSize(GTI.getIndexedType());
if (ConstantInt *CI = dyn_cast<ConstantInt>(Op)) {
// Constant-offset addressing.
Disp += CI->getSExtValue() * S;
(S == 1 || S == 2 || S == 4 || S == 8)) {
// Scaled-index addressing.
Scale = S;
- IndexReg = getRegForValue(Op);
+ IndexReg = getRegForGEPIndex(Op);
if (IndexReg == 0)
return false;
} else
MVT VT;
if (!isTypeLegal(I->getOperand(0)->getType(), VT))
return false;
- unsigned Val = getRegForValue(I->getOperand(0));
- if (Val == 0)
- // Unhandled operand. Halt "fast" selection and bail.
- return false;
X86AddressMode AM;
if (!X86SelectAddress(I->getOperand(1), AM, false))
return false;
- return X86FastEmitStore(VT, Val, AM);
+ return X86FastEmitStore(VT, I->getOperand(0), AM);
}
/// X86SelectLoad - Select and emit code to implement load instructions.
case MVT::i64:
// 64-bit comparisons are only valid if the immediate fits in a 32-bit sext
// field.
- if (RHSC->getType() == Type::Int64Ty &&
- (int)RHSC->getSExtValue() == RHSC->getSExtValue())
+ if ((int)RHSC->getSExtValue() == RHSC->getSExtValue())
return X86::CMP64ri32;
return 0;
}
unsigned BranchOpc; // Opcode to jump on, e.g. "X86::JA"
switch (Predicate) {
+ case CmpInst::FCMP_OEQ:
+ std::swap(TrueMBB, FalseMBB);
+ Predicate = CmpInst::FCMP_UNE;
+ // FALL THROUGH
+ case CmpInst::FCMP_UNE: SwapArgs = false; BranchOpc = X86::JNE; break;
case CmpInst::FCMP_OGT: SwapArgs = false; BranchOpc = X86::JA; break;
case CmpInst::FCMP_OGE: SwapArgs = false; BranchOpc = X86::JAE; break;
case CmpInst::FCMP_OLT: SwapArgs = true; BranchOpc = X86::JA; break;
return false;
BuildMI(MBB, TII.get(BranchOpc)).addMBB(TrueMBB);
+
+ if (Predicate == CmpInst::FCMP_UNE) {
+ // X86 requires a second branch to handle UNE (and OEQ,
+ // which is mapped to UNE above).
+ BuildMI(MBB, TII.get(X86::JP)).addMBB(TrueMBB);
+ }
+
FastEmitBranch(FalseMBB);
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.
+
+ Value *Agg = EI->getAggregateOperand();
+
+ if (CallInst *CI = dyn_cast<CallInst>(Agg)) {
+ Function *F = CI->getCalledFunction();
+
+ if (F && F->isDeclaration()) {
+ switch (F->getIntrinsicID()) {
+ default: break;
+ case Intrinsic::sadd_with_overflow:
+ case Intrinsic::uadd_with_overflow: {
+ const MachineInstr *SetMI = 0;
+ unsigned Reg = lookUpRegForValue(EI);
+
+ for (MachineBasicBlock::const_reverse_iterator
+ RI = MBB->rbegin(), RE = MBB->rend(); RI != RE; ++RI) {
+ const MachineInstr &MI = *RI;
+
+ if (MI.modifiesRegister(Reg)) {
+ unsigned Src, Dst, SrcSR, DstSR;
+
+ if (getInstrInfo()->isMoveInstr(MI, Src, Dst, SrcSR, DstSR)) {
+ Reg = Src;
+ continue;
+ }
+
+ SetMI = &MI;
+ break;
+ }
+
+ const TargetInstrDesc &TID = MI.getDesc();
+ const unsigned *ImpDefs = TID.getImplicitDefs();
+
+ if (TID.hasUnmodeledSideEffects()) break;
+
+ bool ModifiesEFlags = false;
+
+ if (ImpDefs) {
+ for (unsigned u = 0; ImpDefs[u]; ++u)
+ if (ImpDefs[u] == X86::EFLAGS) {
+ ModifiesEFlags = true;
+ break;
+ }
+ }
+
+ if (ModifiesEFlags) break;
+ }
+
+ if (SetMI) {
+ unsigned OpCode = SetMI->getOpcode();
+
+ if (OpCode == X86::SETOr || OpCode == X86::SETBr) {
+ BuildMI(MBB, TII.get((OpCode == X86::SETOr) ?
+ X86::JO : X86::JB)).addMBB(TrueMBB);
+ FastEmitBranch(FalseMBB);
+ MBB->addSuccessor(TrueMBB);
+ return true;
+ }
+ }
+ }
+ }
+ }
+ }
}
// Otherwise do a clumsy setcc and re-test it.
if (ConstantInt *CI = dyn_cast<ConstantInt>(I->getOperand(1))) {
unsigned ResultReg = createResultReg(RC);
BuildMI(MBB, TII.get(OpImm),
- ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue());
+ ResultReg).addReg(Op0Reg).addImm(CI->getZExtValue() & 0xff);
UpdateValueMap(I, ResultReg);
return true;
}
BuildMI(MBB, TII.get(CopyOpc), CopyReg).addReg(InputReg);
// Then issue an extract_subreg.
- unsigned ResultReg = FastEmitInst_extractsubreg(CopyReg, X86::SUBREG_8BIT);
+ unsigned ResultReg = FastEmitInst_extractsubreg(DstVT.getSimpleVT(),
+ CopyReg, X86::SUBREG_8BIT);
if (!ResultReg)
return false;
return true;
}
+bool X86FastISel::X86SelectExtractValue(Instruction *I) {
+ ExtractValueInst *EI = cast<ExtractValueInst>(I);
+ Value *Agg = EI->getAggregateOperand();
+
+ if (CallInst *CI = dyn_cast<CallInst>(Agg)) {
+ Function *F = CI->getCalledFunction();
+
+ if (F && F->isDeclaration()) {
+ switch (F->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".
+ UpdateValueMap(I, lookUpRegForValue(Agg) + *EI->idx_begin());
+ return true;
+ }
+ }
+ }
+
+ return false;
+}
+
+bool X86FastISel::X86VisitIntrinsicCall(CallInst &I, unsigned Intrinsic) {
+ // FIXME: Handle more intrinsics.
+ switch (Intrinsic) {
+ default: return false;
+ case Intrinsic::sadd_with_overflow:
+ case Intrinsic::uadd_with_overflow: {
+ // 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".
+ MVT VT;
+ const Function *Callee = I.getCalledFunction();
+ const Type *RetTy =
+ cast<StructType>(Callee->getReturnType())->getTypeAtIndex(unsigned(0));
+
+ if (!isTypeLegal(RetTy, VT))
+ return false;
+
+ Value *Op1 = I.getOperand(1);
+ Value *Op2 = I.getOperand(2);
+ unsigned Reg1 = getRegForValue(Op1);
+ unsigned Reg2 = getRegForValue(Op2);
+
+ if (Reg1 == 0 || Reg2 == 0)
+ // FIXME: Handle values *not* in registers.
+ return false;
+
+ unsigned OpC = 0;
+
+ if (VT == MVT::i32)
+ OpC = X86::ADD32rr;
+ else if (VT == MVT::i64)
+ OpC = X86::ADD64rr;
+ else
+ return false;
+
+ unsigned ResultReg = createResultReg(TLI.getRegClassFor(VT));
+ BuildMI(MBB, TII.get(OpC), ResultReg).addReg(Reg1).addReg(Reg2);
+ UpdateValueMap(&I, ResultReg);
+
+ ResultReg = createResultReg(TLI.getRegClassFor(MVT::i8));
+ BuildMI(MBB, TII.get((Intrinsic == Intrinsic::sadd_with_overflow) ?
+ X86::SETOr : X86::SETBr), ResultReg);
+ return true;
+ }
+ }
+}
+
bool X86FastISel::X86SelectCall(Instruction *I) {
CallInst *CI = cast<CallInst>(I);
Value *Callee = I->getOperand(0);
if (isa<InlineAsm>(Callee))
return false;
- // FIXME: Handle some intrinsics.
- if (Function *F = CI->getCalledFunction()) {
- if (F->isDeclaration() &&F->getIntrinsicID())
- return false;
- }
+ // Handle intrinsic calls.
+ if (Function *F = CI->getCalledFunction())
+ if (F->isDeclaration())
+ if (unsigned IID = F->getIntrinsicID())
+ return X86VisitIntrinsicCall(*CI, IID);
// Handle only C and fastcc calling conventions for now.
CallSite CS(CI);
}
// Deal with call operands first.
- SmallVector<unsigned, 4> Args;
- SmallVector<MVT, 4> ArgVTs;
- SmallVector<ISD::ArgFlagsTy, 4> ArgFlags;
+ SmallVector<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());
for (CallSite::arg_iterator i = CS.arg_begin(), e = CS.arg_end();
Flags.setOrigAlign(OriginalAlignment);
Args.push_back(Arg);
+ ArgVals.push_back(*i);
ArgVTs.push_back(ArgVT);
ArgFlags.push_back(Flags);
}
unsigned AdjStackDown = TM.getRegisterInfo()->getCallFrameSetupOpcode();
BuildMI(MBB, TII.get(AdjStackDown)).addImm(NumBytes);
- // Process argumenet: walk the register/memloc assignments, inserting
+ // Process argument: walk the register/memloc assignments, inserting
// copies / loads.
SmallVector<unsigned, 4> RegArgs;
for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
case CCValAssign::SExt: {
bool Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
Arg, ArgVT, Arg);
- assert(Emitted && "Failed to emit a sext!");
+ assert(Emitted && "Failed to emit a sext!"); Emitted=Emitted;
+ Emitted = true;
ArgVT = VA.getLocVT();
break;
}
case CCValAssign::ZExt: {
bool Emitted = X86FastEmitExtend(ISD::ZERO_EXTEND, VA.getLocVT(),
Arg, ArgVT, Arg);
- assert(Emitted && "Failed to emit a zext!");
+ assert(Emitted && "Failed to emit a zext!"); Emitted=Emitted;
+ Emitted = true;
ArgVT = VA.getLocVT();
break;
}
Emitted = X86FastEmitExtend(ISD::SIGN_EXTEND, VA.getLocVT(),
Arg, ArgVT, Arg);
- assert(Emitted && "Failed to emit a aext!");
+ assert(Emitted && "Failed to emit a aext!"); Emitted=Emitted;
ArgVT = VA.getLocVT();
break;
}
TargetRegisterClass* RC = TLI.getRegClassFor(ArgVT);
bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), VA.getLocReg(),
Arg, RC, RC);
- assert(Emitted && "Failed to emit a copy instruction!");
+ assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
+ Emitted = true;
RegArgs.push_back(VA.getLocReg());
} else {
unsigned LocMemOffset = VA.getLocMemOffset();
X86AddressMode AM;
AM.Base.Reg = StackPtr;
AM.Disp = LocMemOffset;
- X86FastEmitStore(ArgVT, Arg, AM);
+ 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);
}
}
TargetRegisterClass *RC = X86::GR32RegisterClass;
unsigned Base = getInstrInfo()->getGlobalBaseReg(&MF);
bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), X86::EBX, Base, RC, RC);
- assert(Emitted && "Failed to emit a copy instruction!");
+ assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
+ Emitted = true;
}
// Issue the call.
unsigned ResultReg = createResultReg(DstRC);
bool Emitted = TII.copyRegToReg(*MBB, MBB->end(), ResultReg,
RVLocs[0].getLocReg(), DstRC, SrcRC);
- assert(Emitted && "Failed to emit a copy instruction!");
+ assert(Emitted && "Failed to emit a copy instruction!"); Emitted=Emitted;
+ Emitted = true;
if (CopyVT != RVLocs[0].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
return X86SelectFPExt(I);
case Instruction::FPTrunc:
return X86SelectFPTrunc(I);
+ case Instruction::ExtractValue:
+ return X86SelectExtractValue(I);
}
return false;
unsigned Align = TD.getPreferredTypeAlignmentShift(C->getType());
if (Align == 0) {
// Alignment of vector types. FIXME!
- Align = TD.getABITypeSize(C->getType());
+ Align = TD.getTypePaddedSize(C->getType());
Align = Log2_64(Align);
}
namespace llvm {
llvm::FastISel *X86::createFastISel(MachineFunction &mf,
MachineModuleInfo *mmi,
+ DwarfWriter *dw,
DenseMap<const Value *, unsigned> &vm,
DenseMap<const BasicBlock *, MachineBasicBlock *> &bm,
DenseMap<const AllocaInst *, int> &am
, SmallSet<Instruction*, 8> &cil
#endif
) {
- return new X86FastISel(mf, mmi, vm, bm, am
+ return new X86FastISel(mf, mmi, dw, vm, bm, am
#ifndef NDEBUG
, cil
#endif
projects / oota-llvm.git / blobdiff