virtual bool TargetSelectInstruction(const Instruction *I);
- /// TryToFoldLoad - The specified machine instr operand is a vreg, and that
+ /// \brief 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
/// possible.
- virtual bool TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
- const LoadInst *LI);
+ virtual bool tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
+ const LoadInst *LI);
+
+ virtual bool FastLowerArguments();
#include "X86GenFastISel.inc"
bool X86SelectShift(const Instruction *I);
+ bool X86SelectDivRem(const Instruction *I);
+
bool X86SelectSelect(const Instruction *I);
bool X86SelectTrunc(const Instruction *I);
unsigned &ResultReg) {
unsigned RR = FastEmit_r(SrcVT.getSimpleVT(), DstVT.getSimpleVT(), Opc,
Src, /*TODO: Kill=*/false);
-
- if (RR != 0) {
- ResultReg = RR;
- return true;
- } else
+ if (RR == 0)
return false;
+
+ ResultReg = RR;
+ return true;
}
/// X86SelectAddress - Attempt to fill in an address from the given value.
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;
// Don't handle popping bytes on return for now.
if (X86MFInfo->getBytesToPopOnReturn() != 0)
- return 0;
+ return false;
// fastcc with -tailcallopt is intended to provide a guaranteed
// tail call optimization. Fastisel doesn't know how to do that.
// The x86-64 ABI for returning structs by value requires that we copy
// the sret argument into %rax for the return. We saved the argument into
// a virtual register in the entry block, so now we copy the value out
- // and into %rax.
- if (Subtarget->is64Bit() && F.hasStructRetAttr()) {
+ // and into %rax. We also do the same with %eax for Win32.
+ if (F.hasStructRetAttr() &&
+ (Subtarget->is64Bit() || Subtarget->isTargetWindows())) {
unsigned Reg = X86MFInfo->getSRetReturnReg();
assert(Reg &&
"SRetReturnReg should have been set in LowerFormalArguments()!");
+ unsigned RetReg = Subtarget->is64Bit() ? X86::RAX : X86::EAX;
BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
- X86::RAX).addReg(Reg);
- RetRegs.push_back(X86::RAX);
+ RetReg).addReg(Reg);
+ RetRegs.push_back(RetReg);
}
// Now emit the RET.
return true;
}
+bool X86FastISel::X86SelectDivRem(const Instruction *I) {
+ const static unsigned NumTypes = 4; // i8, i16, i32, i64
+ const static unsigned NumOps = 4; // SDiv, SRem, UDiv, URem
+ const static bool S = true; // IsSigned
+ const static bool U = false; // !IsSigned
+ const static unsigned Copy = TargetOpcode::COPY;
+ // For the X86 DIV/IDIV instruction, in most cases the dividend
+ // (numerator) must be in a specific register pair highreg:lowreg,
+ // producing the quotient in lowreg and the remainder in highreg.
+ // For most data types, to set up the instruction, the dividend is
+ // copied into lowreg, and lowreg is sign-extended or zero-extended
+ // into highreg. The exception is i8, where the dividend is defined
+ // as a single register rather than a register pair, and we
+ // therefore directly sign-extend or zero-extend the dividend into
+ // lowreg, instead of copying, and ignore the highreg.
+ const static struct DivRemEntry {
+ // The following portion depends only on the data type.
+ const TargetRegisterClass *RC;
+ unsigned LowInReg; // low part of the register pair
+ unsigned HighInReg; // high part of the register pair
+ // The following portion depends on both the data type and the operation.
+ struct DivRemResult {
+ unsigned OpDivRem; // The specific DIV/IDIV opcode to use.
+ unsigned OpSignExtend; // Opcode for sign-extending lowreg into
+ // highreg, or copying a zero into highreg.
+ unsigned OpCopy; // Opcode for copying dividend into lowreg, or
+ // zero/sign-extending into lowreg for i8.
+ unsigned DivRemResultReg; // Register containing the desired result.
+ bool IsOpSigned; // Whether to use signed or unsigned form.
+ } ResultTable[NumOps];
+ } OpTable[NumTypes] = {
+ { &X86::GR8RegClass, X86::AX, 0, {
+ { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AL, S }, // SDiv
+ { X86::IDIV8r, 0, X86::MOVSX16rr8, X86::AH, S }, // SRem
+ { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AL, U }, // UDiv
+ { X86::DIV8r, 0, X86::MOVZX16rr8, X86::AH, U }, // URem
+ }
+ }, // i8
+ { &X86::GR16RegClass, X86::AX, X86::DX, {
+ { X86::IDIV16r, X86::CWD, Copy, X86::AX, S }, // SDiv
+ { X86::IDIV16r, X86::CWD, Copy, X86::DX, S }, // SRem
+ { X86::DIV16r, X86::MOV16r0, Copy, X86::AX, U }, // UDiv
+ { X86::DIV16r, X86::MOV16r0, Copy, X86::DX, U }, // URem
+ }
+ }, // i16
+ { &X86::GR32RegClass, X86::EAX, X86::EDX, {
+ { X86::IDIV32r, X86::CDQ, Copy, X86::EAX, S }, // SDiv
+ { X86::IDIV32r, X86::CDQ, Copy, X86::EDX, S }, // SRem
+ { X86::DIV32r, X86::MOV32r0, Copy, X86::EAX, U }, // UDiv
+ { X86::DIV32r, X86::MOV32r0, Copy, X86::EDX, U }, // URem
+ }
+ }, // i32
+ { &X86::GR64RegClass, X86::RAX, X86::RDX, {
+ { X86::IDIV64r, X86::CQO, Copy, X86::RAX, S }, // SDiv
+ { X86::IDIV64r, X86::CQO, Copy, X86::RDX, S }, // SRem
+ { X86::DIV64r, X86::MOV64r0, Copy, X86::RAX, U }, // UDiv
+ { X86::DIV64r, X86::MOV64r0, Copy, X86::RDX, U }, // URem
+ }
+ }, // i64
+ };
+
+ MVT VT;
+ if (!isTypeLegal(I->getType(), VT))
+ return false;
+
+ unsigned TypeIndex, OpIndex;
+ switch (VT.SimpleTy) {
+ default: return false;
+ case MVT::i8: TypeIndex = 0; break;
+ case MVT::i16: TypeIndex = 1; break;
+ case MVT::i32: TypeIndex = 2; break;
+ case MVT::i64: TypeIndex = 3;
+ if (!Subtarget->is64Bit())
+ return false;
+ break;
+ }
+
+ switch (I->getOpcode()) {
+ default: llvm_unreachable("Unexpected div/rem opcode");
+ case Instruction::SDiv: OpIndex = 0; break;
+ case Instruction::SRem: OpIndex = 1; break;
+ case Instruction::UDiv: OpIndex = 2; break;
+ case Instruction::URem: OpIndex = 3; break;
+ }
+
+ const DivRemEntry &TypeEntry = OpTable[TypeIndex];
+ const DivRemEntry::DivRemResult &OpEntry = TypeEntry.ResultTable[OpIndex];
+ unsigned Op0Reg = getRegForValue(I->getOperand(0));
+ if (Op0Reg == 0)
+ return false;
+ unsigned Op1Reg = getRegForValue(I->getOperand(1));
+ if (Op1Reg == 0)
+ return false;
+
+ // Move op0 into low-order input register.
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(OpEntry.OpCopy), TypeEntry.LowInReg).addReg(Op0Reg);
+ // Zero-extend or sign-extend into high-order input register.
+ if (OpEntry.OpSignExtend) {
+ if (OpEntry.IsOpSigned)
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(OpEntry.OpSignExtend));
+ else
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(OpEntry.OpSignExtend), TypeEntry.HighInReg);
+ }
+ // Generate the DIV/IDIV instruction.
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(OpEntry.OpDivRem)).addReg(Op1Reg);
+ // Copy output register into result register.
+ unsigned ResultReg = createResultReg(TypeEntry.RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL,
+ TII.get(Copy), ResultReg).addReg(OpEntry.DivRemResultReg);
+ UpdateValueMap(I, ResultReg);
+
+ return true;
+}
+
bool X86FastISel::X86SelectSelect(const Instruction *I) {
MVT VT;
if (!isTypeLegal(I->getType(), VT))
else if (Len >= 2)
VT = MVT::i16;
else {
- assert(Len == 1);
VT = MVT::i8;
}
}
}
+bool X86FastISel::FastLowerArguments() {
+ if (!FuncInfo.CanLowerReturn)
+ return false;
+
+ if (Subtarget->isTargetWin64())
+ return false;
+
+ const Function *F = FuncInfo.Fn;
+ if (F->isVarArg())
+ return false;
+
+ CallingConv::ID CC = F->getCallingConv();
+ if (CC != CallingConv::C)
+ return false;
+
+ if (!Subtarget->is64Bit())
+ return false;
+
+ // Only handle simple cases. i.e. Up to 6 i32/i64 scalar arguments.
+ unsigned Idx = 1;
+ for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I, ++Idx) {
+ if (Idx > 6)
+ return false;
+
+ if (F->getAttributes().hasAttribute(Idx, Attribute::ByVal) ||
+ F->getAttributes().hasAttribute(Idx, Attribute::InReg) ||
+ F->getAttributes().hasAttribute(Idx, Attribute::StructRet) ||
+ F->getAttributes().hasAttribute(Idx, Attribute::Nest))
+ return false;
+
+ Type *ArgTy = I->getType();
+ if (ArgTy->isStructTy() || ArgTy->isArrayTy() || ArgTy->isVectorTy())
+ return false;
+
+ EVT ArgVT = TLI.getValueType(ArgTy);
+ if (!ArgVT.isSimple()) return false;
+ switch (ArgVT.getSimpleVT().SimpleTy) {
+ case MVT::i32:
+ case MVT::i64:
+ break;
+ default:
+ return false;
+ }
+ }
+
+ static const uint16_t GPR32ArgRegs[] = {
+ X86::EDI, X86::ESI, X86::EDX, X86::ECX, X86::R8D, X86::R9D
+ };
+ static const uint16_t GPR64ArgRegs[] = {
+ X86::RDI, X86::RSI, X86::RDX, X86::RCX, X86::R8 , X86::R9
+ };
+
+ Idx = 0;
+ const TargetRegisterClass *RC32 = TLI.getRegClassFor(MVT::i32);
+ const TargetRegisterClass *RC64 = TLI.getRegClassFor(MVT::i64);
+ for (Function::const_arg_iterator I = F->arg_begin(), E = F->arg_end();
+ I != E; ++I, ++Idx) {
+ if (I->use_empty())
+ continue;
+ bool is32Bit = TLI.getValueType(I->getType()) == MVT::i32;
+ const TargetRegisterClass *RC = is32Bit ? RC32 : RC64;
+ unsigned SrcReg = is32Bit ? GPR32ArgRegs[Idx] : GPR64ArgRegs[Idx];
+ unsigned DstReg = FuncInfo.MF->addLiveIn(SrcReg, RC);
+ // FIXME: Unfortunately it's necessary to emit a copy from the livein copy.
+ // Without this, EmitLiveInCopies may eliminate the livein if its only
+ // use is a bitcast (which isn't turned into an instruction).
+ unsigned ResultReg = createResultReg(RC);
+ BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DL, TII.get(TargetOpcode::COPY),
+ ResultReg).addReg(DstReg, getKillRegState(true));
+ UpdateValueMap(I, ResultReg);
+ }
+ return true;
+}
+
bool X86FastISel::X86SelectCall(const Instruction *I) {
const CallInst *CI = cast<CallInst>(I);
const Value *Callee = CI->getCalledValue();
case Instruction::AShr:
case Instruction::Shl:
return X86SelectShift(I);
+ case Instruction::SDiv:
+ case Instruction::UDiv:
+ case Instruction::SRem:
+ case Instruction::URem:
+ return X86SelectDivRem(I);
case Instruction::Select:
return X86SelectSelect(I);
case Instruction::Trunc:
}
-/// 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
-/// possible.
-bool X86FastISel::TryToFoldLoad(MachineInstr *MI, unsigned OpNo,
- const LoadInst *LI) {
+bool X86FastISel::tryToFoldLoadIntoMI(MachineInstr *MI, unsigned OpNo,
+ const LoadInst *LI) {
X86AddressMode AM;
if (!X86SelectAddress(LI->getOperand(0), AM))
return false;