-//===-- InstSelectSimple.cpp - A simple instruction selector for x86 ------===//
+//===-- X86ISelSimple.cpp - A simple instruction selector for x86 ---------===//
//
// The LLVM Compiler Infrastructure
//
#include "llvm/DerivedTypes.h"
#include "llvm/Function.h"
#include "llvm/Instructions.h"
-#include "llvm/IntrinsicLowering.h"
#include "llvm/Pass.h"
+#include "llvm/CodeGen/IntrinsicLowering.h"
#include "llvm/CodeGen/MachineConstantPool.h"
#include "llvm/CodeGen/MachineFrameInfo.h"
#include "llvm/CodeGen/MachineFunction.h"
#include "llvm/Target/TargetMachine.h"
#include "llvm/Support/GetElementPtrTypeIterator.h"
#include "llvm/Support/InstVisitor.h"
-#include "llvm/Support/CFG.h"
-#include "Support/Statistic.h"
+#include "llvm/ADT/Statistic.h"
using namespace llvm;
namespace {
Statistic<>
NumFPKill("x86-codegen", "Number of FP_REG_KILL instructions added");
+
+ /// TypeClass - Used by the X86 backend to group LLVM types by their basic X86
+ /// Representation.
+ ///
+ enum TypeClass {
+ cByte, cShort, cInt, cFP, cLong
+ };
+}
+
+/// getClass - Turn a primitive type into a "class" number which is based on the
+/// size of the type, and whether or not it is floating point.
+///
+static inline TypeClass getClass(const Type *Ty) {
+ switch (Ty->getTypeID()) {
+ case Type::SByteTyID:
+ case Type::UByteTyID: return cByte; // Byte operands are class #0
+ case Type::ShortTyID:
+ case Type::UShortTyID: return cShort; // Short operands are class #1
+ case Type::IntTyID:
+ case Type::UIntTyID:
+ case Type::PointerTyID: return cInt; // Int's and pointers are class #2
+
+ case Type::FloatTyID:
+ case Type::DoubleTyID: return cFP; // Floating Point is #3
+
+ case Type::LongTyID:
+ case Type::ULongTyID: return cLong; // Longs are class #4
+ default:
+ assert(0 && "Invalid type to getClass!");
+ return cByte; // not reached
+ }
+}
+
+// getClassB - Just like getClass, but treat boolean values as bytes.
+static inline TypeClass getClassB(const Type *Ty) {
+ if (Ty == Type::BoolTy) return cByte;
+ return getClass(Ty);
}
namespace {
- struct ISel : public FunctionPass, InstVisitor<ISel> {
+ struct X86ISel : public FunctionPass, InstVisitor<X86ISel> {
TargetMachine &TM;
MachineFunction *F; // The function we are compiling into
MachineBasicBlock *BB; // The current MBB we are compiling
// MBBMap - Mapping between LLVM BB -> Machine BB
std::map<const BasicBlock*, MachineBasicBlock*> MBBMap;
- ISel(TargetMachine &tm) : TM(tm), F(0), BB(0) {}
+ // AllocaMap - Mapping from fixed sized alloca instructions to the
+ // FrameIndex for the alloca.
+ std::map<AllocaInst*, unsigned> AllocaMap;
+
+ X86ISel(TargetMachine &tm) : TM(tm), F(0), BB(0) {}
/// runOnFunction - Top level implementation of instruction selection for
/// the entire function.
RegMap.clear();
MBBMap.clear();
+ AllocaMap.clear();
F = 0;
// We always build a machine code representation for the function
return true;
// Control flow operators
void visitReturnInst(ReturnInst &RI);
void visitBranchInst(BranchInst &BI);
+ void visitUnreachableInst(UnreachableInst &UI) {}
struct ValueRecord {
Value *Val;
void visitSimpleBinary(BinaryOperator &B, unsigned OpcodeClass);
void visitAdd(BinaryOperator &B) { visitSimpleBinary(B, 0); }
void visitSub(BinaryOperator &B) { visitSimpleBinary(B, 1); }
- void doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator MBBI,
- unsigned DestReg, const Type *DestTy,
- unsigned Op0Reg, unsigned Op1Reg);
- void doMultiplyConst(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator MBBI,
- unsigned DestReg, const Type *DestTy,
- unsigned Op0Reg, unsigned Op1Val);
void visitMul(BinaryOperator &B);
void visitDiv(BinaryOperator &B) { visitDivRem(B); }
/// getAddressingMode - Get the addressing mode to use to address the
/// specified value. The returned value should be used with addFullAddress.
- void getAddressingMode(Value *Addr, unsigned &BaseReg, unsigned &Scale,
- unsigned &IndexReg, unsigned &Disp);
+ void getAddressingMode(Value *Addr, X86AddressMode &AM);
/// getGEPIndex - This is used to fold GEP instructions into X86 addressing
/// expressions.
void getGEPIndex(MachineBasicBlock *MBB, MachineBasicBlock::iterator IP,
std::vector<Value*> &GEPOps,
- std::vector<const Type*> &GEPTypes, unsigned &BaseReg,
- unsigned &Scale, unsigned &IndexReg, unsigned &Disp);
+ std::vector<const Type*> &GEPTypes,
+ X86AddressMode &AM);
/// isGEPFoldable - Return true if the specified GEP can be completely
/// folded into the addressing mode of a load/store or lea instruction.
bool isGEPFoldable(MachineBasicBlock *MBB,
Value *Src, User::op_iterator IdxBegin,
- User::op_iterator IdxEnd, unsigned &BaseReg,
- unsigned &Scale, unsigned &IndexReg, unsigned &Disp);
+ User::op_iterator IdxEnd, X86AddressMode &AM);
/// emitGEPOperation - Common code shared between visitGetElementPtrInst and
/// constant expression GEP support.
Value *Op0, Value *Op1,
unsigned OperatorClass, unsigned TargetReg);
+ /// emitBinaryFPOperation - This method handles emission of floating point
+ /// Add (0), Sub (1), Mul (2), and Div (3) operations.
+ void emitBinaryFPOperation(MachineBasicBlock *BB,
+ MachineBasicBlock::iterator IP,
+ Value *Op0, Value *Op1,
+ unsigned OperatorClass, unsigned TargetReg);
+
+ void emitMultiply(MachineBasicBlock *BB, MachineBasicBlock::iterator IP,
+ Value *Op0, Value *Op1, unsigned TargetReg);
+
+ void doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator MBBI,
+ unsigned DestReg, const Type *DestTy,
+ unsigned Op0Reg, unsigned Op1Reg);
+ void doMultiplyConst(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator MBBI,
+ unsigned DestReg, const Type *DestTy,
+ unsigned Op0Reg, unsigned Op1Val);
+
void emitDivRemOperation(MachineBasicBlock *BB,
MachineBasicBlock::iterator IP,
- unsigned Op0Reg, unsigned Op1Reg, bool isDiv,
- const Type *Ty, unsigned TargetReg);
+ Value *Op0, Value *Op1, bool isDiv,
+ unsigned TargetReg);
/// emitSetCCOperation - Common code shared between visitSetCondInst and
/// constant expression support.
MachineBasicBlock::iterator IP,
Value *Op, Value *ShiftAmount, bool isLeftShift,
const Type *ResultTy, unsigned DestReg);
+
+ // Emit code for a 'SHLD DestReg, Op0, Op1, Amt' operation, where Amt is a
+ // constant.
+ void doSHLDConst(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator MBBI,
+ unsigned DestReg, unsigned Op0Reg, unsigned Op1Reg,
+ unsigned Op1Val);
/// emitSelectOperation - Common code shared between visitSelectInst and the
/// constant expression support.
MachineBasicBlock::iterator MBBI,
Constant *C, unsigned Reg);
+ void emitUCOMr(MachineBasicBlock *MBB, MachineBasicBlock::iterator MBBI,
+ unsigned LHS, unsigned RHS);
+
/// makeAnotherReg - This method returns the next register number we haven't
/// yet used.
///
return F->getSSARegMap()->createVirtualRegister(RC);
}
- /// getReg - This method turns an LLVM value into a register number. This
- /// is guaranteed to produce the same register number for a particular value
- /// every time it is queried.
+ /// getReg - This method turns an LLVM value into a register number.
///
unsigned getReg(Value &V) { return getReg(&V); } // Allow references
unsigned getReg(Value *V) {
return getReg(V, BB, It);
}
unsigned getReg(Value *V, MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IPt) {
- unsigned &Reg = RegMap[V];
- if (Reg == 0) {
- Reg = makeAnotherReg(V->getType());
- RegMap[V] = Reg;
- }
-
- // If this operand is a constant, emit the code to copy the constant into
- // the register here...
- //
- if (Constant *C = dyn_cast<Constant>(V)) {
- copyConstantToRegister(MBB, IPt, C, Reg);
- RegMap.erase(V); // Assign a new name to this constant if ref'd again
- } else if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
- // Move the address of the global into the register
- BuildMI(*MBB, IPt, X86::MOV32ri, 1, Reg).addGlobalAddress(GV);
- RegMap.erase(V); // Assign a new name to this address if ref'd again
- }
+ MachineBasicBlock::iterator IPt);
- return Reg;
- }
+ /// getFixedSizedAllocaFI - Return the frame index for a fixed sized alloca
+ /// that is to be statically allocated with the initial stack frame
+ /// adjustment.
+ unsigned getFixedSizedAllocaFI(AllocaInst *AI);
};
}
-/// TypeClass - Used by the X86 backend to group LLVM types by their basic X86
-/// Representation.
-///
-enum TypeClass {
- cByte, cShort, cInt, cFP, cLong
-};
+/// dyn_castFixedAlloca - If the specified value is a fixed size alloca
+/// instruction in the entry block, return it. Otherwise, return a null
+/// pointer.
+static AllocaInst *dyn_castFixedAlloca(Value *V) {
+ if (AllocaInst *AI = dyn_cast<AllocaInst>(V)) {
+ BasicBlock *BB = AI->getParent();
+ if (isa<ConstantUInt>(AI->getArraySize()) && BB ==&BB->getParent()->front())
+ return AI;
+ }
+ return 0;
+}
-/// getClass - Turn a primitive type into a "class" number which is based on the
-/// size of the type, and whether or not it is floating point.
+/// getReg - This method turns an LLVM value into a register number.
///
-static inline TypeClass getClass(const Type *Ty) {
- switch (Ty->getPrimitiveID()) {
- case Type::SByteTyID:
- case Type::UByteTyID: return cByte; // Byte operands are class #0
- case Type::ShortTyID:
- case Type::UShortTyID: return cShort; // Short operands are class #1
- case Type::IntTyID:
- case Type::UIntTyID:
- case Type::PointerTyID: return cInt; // Int's and pointers are class #2
-
- case Type::FloatTyID:
- case Type::DoubleTyID: return cFP; // Floating Point is #3
-
- case Type::LongTyID:
- case Type::ULongTyID: return cLong; // Longs are class #4
- default:
- assert(0 && "Invalid type to getClass!");
- return cByte; // not reached
- }
+unsigned X86ISel::getReg(Value *V, MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IPt) {
+ // If this operand is a constant, emit the code to copy the constant into
+ // the register here...
+ if (Constant *C = dyn_cast<Constant>(V)) {
+ unsigned Reg = makeAnotherReg(V->getType());
+ copyConstantToRegister(MBB, IPt, C, Reg);
+ return Reg;
+ } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
+ // Do not emit noop casts at all, unless it's a double -> float cast.
+ if (getClassB(CI->getType()) == getClassB(CI->getOperand(0)->getType()) &&
+ (CI->getType() != Type::FloatTy ||
+ CI->getOperand(0)->getType() != Type::DoubleTy))
+ return getReg(CI->getOperand(0), MBB, IPt);
+ } else if (AllocaInst *AI = dyn_castFixedAlloca(V)) {
+ // If the alloca address couldn't be folded into the instruction addressing,
+ // emit an explicit LEA as appropriate.
+ unsigned Reg = makeAnotherReg(V->getType());
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ addFrameReference(BuildMI(*MBB, IPt, X86::LEA32r, 4, Reg), FI);
+ return Reg;
+ }
+
+ unsigned &Reg = RegMap[V];
+ if (Reg == 0) {
+ Reg = makeAnotherReg(V->getType());
+ RegMap[V] = Reg;
+ }
+
+ return Reg;
}
-// getClassB - Just like getClass, but treat boolean values as bytes.
-static inline TypeClass getClassB(const Type *Ty) {
- if (Ty == Type::BoolTy) return cByte;
- return getClass(Ty);
+/// getFixedSizedAllocaFI - Return the frame index for a fixed sized alloca
+/// that is to be statically allocated with the initial stack frame
+/// adjustment.
+unsigned X86ISel::getFixedSizedAllocaFI(AllocaInst *AI) {
+ // Already computed this?
+ std::map<AllocaInst*, unsigned>::iterator I = AllocaMap.lower_bound(AI);
+ if (I != AllocaMap.end() && I->first == AI) return I->second;
+
+ const Type *Ty = AI->getAllocatedType();
+ ConstantUInt *CUI = cast<ConstantUInt>(AI->getArraySize());
+ unsigned TySize = TM.getTargetData().getTypeSize(Ty);
+ TySize *= CUI->getValue(); // Get total allocated size...
+ unsigned Alignment = TM.getTargetData().getTypeAlignment(Ty);
+
+ // Create a new stack object using the frame manager...
+ int FrameIdx = F->getFrameInfo()->CreateStackObject(TySize, Alignment);
+ AllocaMap.insert(I, std::make_pair(AI, FrameIdx));
+ return FrameIdx;
}
/// copyConstantToRegister - Output the instructions required to put the
/// specified constant into the specified register.
///
-void ISel::copyConstantToRegister(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP,
- Constant *C, unsigned R) {
- if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
+void X86ISel::copyConstantToRegister(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ Constant *C, unsigned R) {
+ if (isa<UndefValue>(C)) {
+ switch (getClassB(C->getType())) {
+ case cFP:
+ // FIXME: SHOULD TEACH STACKIFIER ABOUT UNDEF VALUES!
+ BuildMI(*MBB, IP, X86::FLD0, 0, R);
+ return;
+ case cLong:
+ BuildMI(*MBB, IP, X86::IMPLICIT_DEF, 0, R+1);
+ // FALL THROUGH
+ default:
+ BuildMI(*MBB, IP, X86::IMPLICIT_DEF, 0, R);
+ return;
+ }
+ } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
unsigned Class = 0;
switch (CE->getOpcode()) {
case Instruction::GetElementPtr:
Class, R);
return;
- case Instruction::Mul: {
- unsigned Op0Reg = getReg(CE->getOperand(0), MBB, IP);
- unsigned Op1Reg = getReg(CE->getOperand(1), MBB, IP);
- doMultiply(MBB, IP, R, CE->getType(), Op0Reg, Op1Reg);
+ case Instruction::Mul:
+ emitMultiply(MBB, IP, CE->getOperand(0), CE->getOperand(1), R);
return;
- }
+
case Instruction::Div:
- case Instruction::Rem: {
- unsigned Op0Reg = getReg(CE->getOperand(0), MBB, IP);
- unsigned Op1Reg = getReg(CE->getOperand(1), MBB, IP);
- emitDivRemOperation(MBB, IP, Op0Reg, Op1Reg,
- CE->getOpcode() == Instruction::Div,
- CE->getType(), R);
+ case Instruction::Rem:
+ emitDivRemOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1),
+ CE->getOpcode() == Instruction::Div, R);
return;
- }
case Instruction::SetNE:
case Instruction::SetEQ:
return;
default:
- std::cerr << "Offending expr: " << C << "\n";
+ std::cerr << "Offending expr: " << *C << "\n";
assert(0 && "Constant expression not yet handled!\n");
}
}
} else if (isa<ConstantPointerNull>(C)) {
// Copy zero (null pointer) to the register.
BuildMI(*MBB, IP, X86::MOV32ri, 1, R).addImm(0);
- } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
- BuildMI(*MBB, IP, X86::MOV32ri, 1, R).addGlobalAddress(CPR->getValue());
+ } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) {
+ BuildMI(*MBB, IP, X86::MOV32ri, 1, R).addGlobalAddress(GV);
} else {
- std::cerr << "Offending constant: " << C << "\n";
+ std::cerr << "Offending constant: " << *C << "\n";
assert(0 && "Type not handled yet!");
}
}
/// LoadArgumentsToVirtualRegs - Load all of the arguments to this function from
/// the stack into virtual registers.
///
-void ISel::LoadArgumentsToVirtualRegs(Function &Fn) {
+void X86ISel::LoadArgumentsToVirtualRegs(Function &Fn) {
// Emit instructions to load the arguments... On entry to a function on the
// X86, the stack frame looks like this:
//
MachineFrameInfo *MFI = F->getFrameInfo();
for (Function::aiterator I = Fn.abegin(), E = Fn.aend(); I != E; ++I) {
- unsigned Reg = getReg(*I);
-
+ bool ArgLive = !I->use_empty();
+ unsigned Reg = ArgLive ? getReg(*I) : 0;
int FI; // Frame object index
+
switch (getClassB(I->getType())) {
case cByte:
- FI = MFI->CreateFixedObject(1, ArgOffset);
- addFrameReference(BuildMI(BB, X86::MOV8rm, 4, Reg), FI);
+ if (ArgLive) {
+ FI = MFI->CreateFixedObject(1, ArgOffset);
+ addFrameReference(BuildMI(BB, X86::MOV8rm, 4, Reg), FI);
+ }
break;
case cShort:
- FI = MFI->CreateFixedObject(2, ArgOffset);
- addFrameReference(BuildMI(BB, X86::MOV16rm, 4, Reg), FI);
+ if (ArgLive) {
+ FI = MFI->CreateFixedObject(2, ArgOffset);
+ addFrameReference(BuildMI(BB, X86::MOV16rm, 4, Reg), FI);
+ }
break;
case cInt:
- FI = MFI->CreateFixedObject(4, ArgOffset);
- addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg), FI);
+ if (ArgLive) {
+ FI = MFI->CreateFixedObject(4, ArgOffset);
+ addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg), FI);
+ }
break;
case cLong:
- FI = MFI->CreateFixedObject(8, ArgOffset);
- addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg), FI);
- addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg+1), FI, 4);
+ if (ArgLive) {
+ FI = MFI->CreateFixedObject(8, ArgOffset);
+ addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg), FI);
+ addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg+1), FI, 4);
+ }
ArgOffset += 4; // longs require 4 additional bytes
break;
case cFP:
- unsigned Opcode;
- if (I->getType() == Type::FloatTy) {
- Opcode = X86::FLD32m;
- FI = MFI->CreateFixedObject(4, ArgOffset);
- } else {
- Opcode = X86::FLD64m;
- FI = MFI->CreateFixedObject(8, ArgOffset);
- ArgOffset += 4; // doubles require 4 additional bytes
+ if (ArgLive) {
+ unsigned Opcode;
+ if (I->getType() == Type::FloatTy) {
+ Opcode = X86::FLD32m;
+ FI = MFI->CreateFixedObject(4, ArgOffset);
+ } else {
+ Opcode = X86::FLD64m;
+ FI = MFI->CreateFixedObject(8, ArgOffset);
+ }
+ addFrameReference(BuildMI(BB, Opcode, 4, Reg), FI);
}
- addFrameReference(BuildMI(BB, Opcode, 4, Reg), FI);
+ if (I->getType() == Type::DoubleTy)
+ ArgOffset += 4; // doubles require 4 additional bytes
break;
default:
assert(0 && "Unhandled argument type!");
/// because we have to generate our sources into the source basic blocks, not
/// the current one.
///
-void ISel::SelectPHINodes() {
- const TargetInstrInfo &TII = TM.getInstrInfo();
+void X86ISel::SelectPHINodes() {
+ const TargetInstrInfo &TII = *TM.getInstrInfo();
const Function &LF = *F->getFunction(); // The LLVM function...
for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) {
const BasicBlock *BB = I;
// Loop over all of the PHI nodes in the LLVM basic block...
MachineBasicBlock::iterator PHIInsertPoint = MBB.begin();
- for (BasicBlock::const_iterator I = BB->begin();
- PHINode *PN = const_cast<PHINode*>(dyn_cast<PHINode>(I)); ++I) {
+ for (BasicBlock::const_iterator I = BB->begin(); isa<PHINode>(I); ++I) {
+ PHINode *PN = const_cast<PHINode*>(dyn_cast<PHINode>(I));
// Create a new machine instr PHI node, and insert it.
unsigned PHIReg = getReg(*PN);
// If this is a constant or GlobalValue, we may have to insert code
// into the basic block to compute it into a virtual register.
- if (isa<Constant>(Val) || isa<GlobalValue>(Val)) {
- if (isa<ConstantExpr>(Val)) {
- // Because we don't want to clobber any values which might be in
- // physical registers with the computation of this constant (which
- // might be arbitrarily complex if it is a constant expression),
- // just insert the computation at the top of the basic block.
- MachineBasicBlock::iterator PI = PredMBB->begin();
-
- // Skip over any PHI nodes though!
- while (PI != PredMBB->end() && PI->getOpcode() == X86::PHI)
- ++PI;
-
- ValReg = getReg(Val, PredMBB, PI);
- } else {
- // Simple constants get emitted at the end of the basic block,
- // before any terminator instructions. We "know" that the code to
- // move a constant into a register will never clobber any flags.
- ValReg = getReg(Val, PredMBB, PredMBB->getFirstTerminator());
- }
+ if ((isa<Constant>(Val) && !isa<ConstantExpr>(Val))) {
+ // Simple constants get emitted at the end of the basic block,
+ // before any terminator instructions. We "know" that the code to
+ // move a constant into a register will never clobber any flags.
+ ValReg = getReg(Val, PredMBB, PredMBB->getFirstTerminator());
} else {
- ValReg = getReg(Val);
+ // Because we don't want to clobber any values which might be in
+ // physical registers with the computation of this constant (which
+ // might be arbitrarily complex if it is a constant expression),
+ // just insert the computation at the top of the basic block.
+ MachineBasicBlock::iterator PI = PredMBB->begin();
+
+ // Skip over any PHI nodes though!
+ while (PI != PredMBB->end() && PI->getOpcode() == X86::PHI)
+ ++PI;
+
+ ValReg = getReg(Val, PredMBB, PI);
}
// Remember that we inserted a value for this PHI for this predecessor
/// Note that this kill instruction will eventually be eliminated when
/// restrictions in the stackifier are relaxed.
///
-static bool RequiresFPRegKill(const BasicBlock *BB) {
+static bool RequiresFPRegKill(const MachineBasicBlock *MBB) {
#if 0
+ const BasicBlock *BB = MBB->getBasicBlock ();
for (succ_const_iterator SI = succ_begin(BB), E = succ_end(BB); SI!=E; ++SI) {
const BasicBlock *Succ = *SI;
pred_const_iterator PI = pred_begin(Succ), PE = pred_end(Succ);
// break critical edges as needed (to make a place to put compensation code),
// but this will require some infrastructure improvements as well.
//
-void ISel::InsertFPRegKills() {
+void X86ISel::InsertFPRegKills() {
SSARegMap &RegMap = *F->getSSARegMap();
for (MachineFunction::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
// If we haven't found an FP register use or def in this basic block, check
// to see if any of our successors has an FP PHI node, which will cause a
// copy to be inserted into this block.
- for (succ_const_iterator SI = succ_begin(BB->getBasicBlock()),
- E = succ_end(BB->getBasicBlock()); SI != E; ++SI) {
- MachineBasicBlock *SBB = MBBMap[*SI];
+ for (MachineBasicBlock::const_succ_iterator SI = BB->succ_begin(),
+ SE = BB->succ_end(); SI != SE; ++SI) {
+ MachineBasicBlock *SBB = *SI;
for (MachineBasicBlock::iterator I = SBB->begin();
I != SBB->end() && I->getOpcode() == X86::PHI; ++I) {
if (RegMap.getRegClass(I->getOperand(0).getReg())->getSize() == 10)
UsesFPReg:
// Okay, this block uses an FP register. If the block has successors (ie,
// it's not an unwind/return), insert the FP_REG_KILL instruction.
- if (BB->getBasicBlock()->getTerminator()->getNumSuccessors() &&
- RequiresFPRegKill(BB->getBasicBlock())) {
+ if (BB->succ_size () && RequiresFPRegKill(BB)) {
BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
++NumFPKill;
}
}
+void X86ISel::getAddressingMode(Value *Addr, X86AddressMode &AM) {
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = 0; AM.Scale = 1; AM.IndexReg = 0; AM.Disp = 0;
+ if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr)) {
+ if (isGEPFoldable(BB, GEP->getOperand(0), GEP->op_begin()+1, GEP->op_end(),
+ AM))
+ return;
+ } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
+ if (CE->getOpcode() == Instruction::GetElementPtr)
+ if (isGEPFoldable(BB, CE->getOperand(0), CE->op_begin()+1, CE->op_end(),
+ AM))
+ return;
+ } else if (AllocaInst *AI = dyn_castFixedAlloca(Addr)) {
+ AM.BaseType = X86AddressMode::FrameIndexBase;
+ AM.Base.FrameIndex = getFixedSizedAllocaFI(AI);
+ return;
+ } else if (GlobalValue *GV = dyn_cast<GlobalValue>(Addr)) {
+ AM.GV = GV;
+ return;
+ }
+
+ // If it's not foldable, reset addr mode.
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = getReg(Addr);
+ AM.Scale = 1; AM.IndexReg = 0; AM.Disp = 0;
+}
+
// canFoldSetCCIntoBranchOrSelect - Return the setcc instruction if we can fold
// it into the conditional branch or select instruction which is the only user
// of the cc instruction. This is the case if the conditional branch is the
-// only user of the setcc, and if the setcc is in the same basic block as the
-// conditional branch. We also don't handle long arguments below, so we reject
-// them here as well.
+// only user of the setcc. We also don't handle long arguments below, so we
+// reject them here as well.
//
static SetCondInst *canFoldSetCCIntoBranchOrSelect(Value *V) {
if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
if (SCI->hasOneUse()) {
Instruction *User = cast<Instruction>(SCI->use_back());
if ((isa<BranchInst>(User) || isa<SelectInst>(User)) &&
- SCI->getParent() == User->getParent() &&
(getClassB(SCI->getOperand(0)->getType()) != cLong ||
SCI->getOpcode() == Instruction::SetEQ ||
- SCI->getOpcode() == Instruction::SetNE))
+ SCI->getOpcode() == Instruction::SetNE) &&
+ (isa<BranchInst>(User) || User->getOperand(0) == V))
return SCI;
}
return 0;
X86::SETSr, X86::SETNSr },
};
+/// emitUCOMr - In the future when we support processors before the P6, this
+/// wraps the logic for emitting an FUCOMr vs FUCOMIr.
+void X86ISel::emitUCOMr(MachineBasicBlock *MBB, MachineBasicBlock::iterator IP,
+ unsigned LHS, unsigned RHS) {
+ if (0) { // for processors prior to the P6
+ BuildMI(*MBB, IP, X86::FUCOMr, 2).addReg(LHS).addReg(RHS);
+ BuildMI(*MBB, IP, X86::FNSTSW8r, 0);
+ BuildMI(*MBB, IP, X86::SAHF, 1);
+ } else {
+ BuildMI(*MBB, IP, X86::FUCOMIr, 2).addReg(LHS).addReg(RHS);
+ }
+}
+
// EmitComparison - This function emits a comparison of the two operands,
// returning the extended setcc code to use.
-unsigned ISel::EmitComparison(unsigned OpNum, Value *Op0, Value *Op1,
- MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP) {
+unsigned X86ISel::EmitComparison(unsigned OpNum, Value *Op0, Value *Op1,
+ MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP) {
// The arguments are already supposed to be of the same type.
const Type *CompTy = Op0->getType();
unsigned Class = getClassB(CompTy);
- unsigned Op0r = getReg(Op0, MBB, IP);
// Special case handling of: cmp R, i
- if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
+ if (isa<ConstantPointerNull>(Op1)) {
+ unsigned Op0r = getReg(Op0, MBB, IP);
+ if (OpNum < 2) // seteq/setne -> test
+ BuildMI(*MBB, IP, X86::TEST32rr, 2).addReg(Op0r).addReg(Op0r);
+ else
+ BuildMI(*MBB, IP, X86::CMP32ri, 2).addReg(Op0r).addImm(0);
+ return OpNum;
+
+ } else if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
if (Class == cByte || Class == cShort || Class == cInt) {
unsigned Op1v = CI->getRawValue();
// can't handle unsigned comparisons against zero unless they are == or
// !=. These should have been strength reduced already anyway.
if (Op1v == 0 && (CompTy->isSigned() || OpNum < 2)) {
+
+ // If this is a comparison against zero and the LHS is an and of a
+ // register with a constant, use the test to do the and.
+ if (Instruction *Op0I = dyn_cast<Instruction>(Op0))
+ if (Op0I->getOpcode() == Instruction::And && Op0->hasOneUse() &&
+ isa<ConstantInt>(Op0I->getOperand(1))) {
+ static const unsigned TESTTab[] = {
+ X86::TEST8ri, X86::TEST16ri, X86::TEST32ri
+ };
+
+ // Emit test X, i
+ unsigned LHS = getReg(Op0I->getOperand(0), MBB, IP);
+ unsigned Imm =
+ cast<ConstantInt>(Op0I->getOperand(1))->getRawValue();
+ BuildMI(*MBB, IP, TESTTab[Class], 2).addReg(LHS).addImm(Imm);
+
+ if (OpNum == 2) return 6; // Map jl -> js
+ if (OpNum == 3) return 7; // Map jg -> jns
+ return OpNum;
+ }
+
+ unsigned Op0r = getReg(Op0, MBB, IP);
static const unsigned TESTTab[] = {
X86::TEST8rr, X86::TEST16rr, X86::TEST32rr
};
X86::CMP8ri, X86::CMP16ri, X86::CMP32ri
};
+ unsigned Op0r = getReg(Op0, MBB, IP);
BuildMI(*MBB, IP, CMPTab[Class], 2).addReg(Op0r).addImm(Op1v);
return OpNum;
} else {
+ unsigned Op0r = getReg(Op0, MBB, IP);
assert(Class == cLong && "Unknown integer class!");
unsigned LowCst = CI->getRawValue();
unsigned HiCst = CI->getRawValue() >> 32;
// each, then uses a conditional move to handle the overflow case. For
// example, a setlt for long would generate code like this:
//
- // AL = lo(op1) < lo(op2) // Signedness depends on operands
- // BL = hi(op1) < hi(op2) // Always unsigned comparison
- // dest = hi(op1) == hi(op2) ? AL : BL;
+ // AL = lo(op1) < lo(op2) // Always unsigned comparison
+ // BL = hi(op1) < hi(op2) // Signedness depends on operands
+ // dest = hi(op1) == hi(op2) ? BL : AL;
//
// FIXME: This would be much better if we had hierarchical register
}
}
+ unsigned Op0r = getReg(Op0, MBB, IP);
+
// Special case handling of comparison against +/- 0.0
if (ConstantFP *CFP = dyn_cast<ConstantFP>(Op1))
if (CFP->isExactlyValue(+0.0) || CFP->isExactlyValue(-0.0)) {
BuildMI(*MBB, IP, X86::CMP32rr, 2).addReg(Op0r).addReg(Op1r);
break;
case cFP:
- BuildMI(*MBB, IP, X86::FpUCOM, 2).addReg(Op0r).addReg(Op1r);
- BuildMI(*MBB, IP, X86::FNSTSW8r, 0);
- BuildMI(*MBB, IP, X86::SAHF, 1);
+ emitUCOMr(MBB, IP, Op0r, Op1r);
break;
case cLong:
//
// AL = lo(op1) < lo(op2) // Signedness depends on operands
// BL = hi(op1) < hi(op2) // Always unsigned comparison
- // dest = hi(op1) == hi(op2) ? AL : BL;
+ // dest = hi(op1) == hi(op2) ? BL : AL;
//
// FIXME: This would be much better if we had hierarchical register
/// SetCC instructions - Here we just emit boilerplate code to set a byte-sized
/// register, then move it to wherever the result should be.
///
-void ISel::visitSetCondInst(SetCondInst &I) {
+void X86ISel::visitSetCondInst(SetCondInst &I) {
if (canFoldSetCCIntoBranchOrSelect(&I))
return; // Fold this into a branch or select.
/// emitSetCCOperation - Common code shared between visitSetCondInst and
/// constant expression support.
///
-void ISel::emitSetCCOperation(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP,
- Value *Op0, Value *Op1, unsigned Opcode,
- unsigned TargetReg) {
+void X86ISel::emitSetCCOperation(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ Value *Op0, Value *Op1, unsigned Opcode,
+ unsigned TargetReg) {
unsigned OpNum = getSetCCNumber(Opcode);
OpNum = EmitComparison(OpNum, Op0, Op1, MBB, IP);
}
}
-void ISel::visitSelectInst(SelectInst &SI) {
+void X86ISel::visitSelectInst(SelectInst &SI) {
unsigned DestReg = getReg(SI);
MachineBasicBlock::iterator MII = BB->end();
emitSelectOperation(BB, MII, SI.getCondition(), SI.getTrueValue(),
/// emitSelect - Common code shared between visitSelectInst and the constant
/// expression support.
-void ISel::emitSelectOperation(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP,
- Value *Cond, Value *TrueVal, Value *FalseVal,
- unsigned DestReg) {
+void X86ISel::emitSelectOperation(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ Value *Cond, Value *TrueVal, Value *FalseVal,
+ unsigned DestReg) {
unsigned SelectClass = getClassB(TrueVal->getType());
// We don't support 8-bit conditional moves. If we have incoming constants,
FalseVal = ConstantExpr::getCast(F, Type::ShortTy);
}
-
+ unsigned TrueReg = getReg(TrueVal, MBB, IP);
+ unsigned FalseReg = getReg(FalseVal, MBB, IP);
+ if (TrueReg == FalseReg) {
+ static const unsigned Opcode[] = {
+ X86::MOV8rr, X86::MOV16rr, X86::MOV32rr, X86::FpMOV, X86::MOV32rr
+ };
+ BuildMI(*MBB, IP, Opcode[SelectClass], 1, DestReg).addReg(TrueReg);
+ if (SelectClass == cLong)
+ BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg+1).addReg(TrueReg+1);
+ return;
+ }
+
unsigned Opcode;
if (SetCondInst *SCI = canFoldSetCCIntoBranchOrSelect(Cond)) {
// We successfully folded the setcc into the select instruction.
}
}
- unsigned TrueReg = getReg(TrueVal, MBB, IP);
- unsigned FalseReg = getReg(FalseVal, MBB, IP);
unsigned RealDestReg = DestReg;
/// promote32 - Emit instructions to turn a narrow operand into a 32-bit-wide
/// operand, in the specified target register.
///
-void ISel::promote32(unsigned targetReg, const ValueRecord &VR) {
- bool isUnsigned = VR.Ty->isUnsigned();
+void X86ISel::promote32(unsigned targetReg, const ValueRecord &VR) {
+ bool isUnsigned = VR.Ty->isUnsigned() || VR.Ty == Type::BoolTy;
Value *Val = VR.Val;
const Type *Ty = VR.Ty;
// copy.
if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
int TheVal = CI->getRawValue() & 0xFFFFFFFF;
- BuildMI(BB, X86::MOV32ri, 1, targetReg).addImm(TheVal);
+ BuildMI(BB, X86::MOV32ri, 1, targetReg).addImm(TheVal);
return;
}
}
/// ret long, ulong : Move value into EAX/EDX and return
/// ret float/double : Top of FP stack
///
-void ISel::visitReturnInst(ReturnInst &I) {
+void X86ISel::visitReturnInst(ReturnInst &I) {
if (I.getNumOperands() == 0) {
BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
return;
/// jump to a block that is the immediate successor of the current block, we can
/// just make a fall-through (but we don't currently).
///
-void ISel::visitBranchInst(BranchInst &BI) {
+void X86ISel::visitBranchInst(BranchInst &BI) {
+ // Update machine-CFG edges
+ BB->addSuccessor (MBBMap[BI.getSuccessor(0)]);
+ if (BI.isConditional())
+ BB->addSuccessor (MBBMap[BI.getSuccessor(1)]);
+
BasicBlock *NextBB = getBlockAfter(BI.getParent()); // BB after current one
if (!BI.isConditional()) { // Unconditional branch?
if (BI.getSuccessor(0) != NextBB)
- BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0));
+ BuildMI(BB, X86::JMP, 1).addMBB(MBBMap[BI.getSuccessor(0)]);
return;
}
BuildMI(BB, X86::TEST8rr, 2).addReg(condReg).addReg(condReg);
if (BI.getSuccessor(1) == NextBB) {
if (BI.getSuccessor(0) != NextBB)
- BuildMI(BB, X86::JNE, 1).addPCDisp(BI.getSuccessor(0));
+ BuildMI(BB, X86::JNE, 1).addMBB(MBBMap[BI.getSuccessor(0)]);
} else {
- BuildMI(BB, X86::JE, 1).addPCDisp(BI.getSuccessor(1));
+ BuildMI(BB, X86::JE, 1).addMBB(MBBMap[BI.getSuccessor(1)]);
if (BI.getSuccessor(0) != NextBB)
- BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0));
+ BuildMI(BB, X86::JMP, 1).addMBB(MBBMap[BI.getSuccessor(0)]);
}
return;
}
};
if (BI.getSuccessor(0) != NextBB) {
- BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(0));
+ BuildMI(BB, OpcodeTab[isSigned][OpNum], 1)
+ .addMBB(MBBMap[BI.getSuccessor(0)]);
if (BI.getSuccessor(1) != NextBB)
- BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(1));
+ BuildMI(BB, X86::JMP, 1).addMBB(MBBMap[BI.getSuccessor(1)]);
} else {
// Change to the inverse condition...
if (BI.getSuccessor(1) != NextBB) {
OpNum ^= 1;
- BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(1));
+ BuildMI(BB, OpcodeTab[isSigned][OpNum], 1)
+ .addMBB(MBBMap[BI.getSuccessor(1)]);
}
}
}
/// and the return value as appropriate. For the actual function call itself,
/// it inserts the specified CallMI instruction into the stream.
///
-void ISel::doCall(const ValueRecord &Ret, MachineInstr *CallMI,
- const std::vector<ValueRecord> &Args) {
-
+void X86ISel::doCall(const ValueRecord &Ret, MachineInstr *CallMI,
+ const std::vector<ValueRecord> &Args) {
// Count how many bytes are to be pushed on the stack...
unsigned NumBytes = 0;
unsigned ArgReg;
switch (getClassB(Args[i].Ty)) {
case cByte:
+ if (Args[i].Val && isa<ConstantBool>(Args[i].Val)) {
+ addRegOffset(BuildMI(BB, X86::MOV32mi, 5), X86::ESP, ArgOffset)
+ .addImm(Args[i].Val == ConstantBool::True);
+ break;
+ }
+ // FALL THROUGH
case cShort:
if (Args[i].Val && isa<ConstantInt>(Args[i].Val)) {
// Zero/Sign extend constant, then stuff into memory.
unsigned Val = cast<ConstantInt>(Args[i].Val)->getRawValue();
addRegOffset(BuildMI(BB, X86::MOV32mi, 5),
X86::ESP, ArgOffset).addImm(Val);
+ } else if (Args[i].Val && isa<ConstantPointerNull>(Args[i].Val)) {
+ addRegOffset(BuildMI(BB, X86::MOV32mi, 5),
+ X86::ESP, ArgOffset).addImm(0);
} else {
ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg;
addRegOffset(BuildMI(BB, X86::MOV32mr, 5),
/// visitCallInst - Push args on stack and do a procedure call instruction.
-void ISel::visitCallInst(CallInst &CI) {
+void X86ISel::visitCallInst(CallInst &CI) {
MachineInstr *TheCall;
if (Function *F = CI.getCalledFunction()) {
// Is it an intrinsic function call?
doCall(ValueRecord(DestReg, CI.getType()), TheCall, Args);
}
-
/// LowerUnknownIntrinsicFunctionCalls - This performs a prepass over the
/// function, lowering any calls to unknown intrinsic functions into the
/// equivalent LLVM code.
///
-void ISel::LowerUnknownIntrinsicFunctionCalls(Function &F) {
+void X86ISel::LowerUnknownIntrinsicFunctionCalls(Function &F) {
for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
if (CallInst *CI = dyn_cast<CallInst>(I++))
case Intrinsic::frameaddress:
case Intrinsic::memcpy:
case Intrinsic::memset:
+ case Intrinsic::isunordered:
case Intrinsic::readport:
case Intrinsic::writeport:
// We directly implement these intrinsics
break;
+ case Intrinsic::readio: {
+ // On X86, memory operations are in-order. Lower this intrinsic
+ // into a volatile load.
+ Instruction *Before = CI->getPrev();
+ LoadInst * LI = new LoadInst(CI->getOperand(1), "", true, CI);
+ CI->replaceAllUsesWith(LI);
+ BB->getInstList().erase(CI);
+ break;
+ }
+ case Intrinsic::writeio: {
+ // On X86, memory operations are in-order. Lower this intrinsic
+ // into a volatile store.
+ Instruction *Before = CI->getPrev();
+ StoreInst *LI = new StoreInst(CI->getOperand(1),
+ CI->getOperand(2), true, CI);
+ CI->replaceAllUsesWith(LI);
+ BB->getInstList().erase(CI);
+ break;
+ }
default:
// All other intrinsic calls we must lower.
Instruction *Before = CI->getPrev();
TM.getIntrinsicLowering().LowerIntrinsicCall(CI);
if (Before) { // Move iterator to instruction after call
- I = Before; ++I;
+ I = Before; ++I;
} else {
I = BB->begin();
}
}
-
}
-void ISel::visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI) {
+void X86ISel::visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI) {
unsigned TmpReg1, TmpReg2;
switch (ID) {
case Intrinsic::vastart:
}
return;
+ case Intrinsic::isunordered:
+ TmpReg1 = getReg(CI.getOperand(1));
+ TmpReg2 = getReg(CI.getOperand(2));
+ emitUCOMr(BB, BB->end(), TmpReg2, TmpReg1);
+ TmpReg2 = getReg(CI);
+ BuildMI(BB, X86::SETPr, 0, TmpReg2);
+ return;
+
case Intrinsic::memcpy: {
assert(CI.getNumOperands() == 5 && "Illegal llvm.memcpy call!");
unsigned Align = 1;
return;
}
- case Intrinsic::readport:
+ case Intrinsic::readport: {
+ // First, determine that the size of the operand falls within the acceptable
+ // range for this architecture.
//
- // First, determine that the size of the operand falls within the
- // acceptable range for this architecture.
- //
- if ((CI.getOperand(1)->getType()->getPrimitiveSize()) != 2) {
+ if (getClassB(CI.getOperand(1)->getType()) != cShort) {
std::cerr << "llvm.readport: Address size is not 16 bits\n";
- exit (1);
+ exit(1);
}
- //
// Now, move the I/O port address into the DX register and use the IN
// instruction to get the input data.
//
- BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(getReg(CI.getOperand(1)));
- switch (CI.getCalledFunction()->getReturnType()->getPrimitiveSize()) {
- case 1:
- BuildMI(BB, X86::IN8, 0);
- break;
- case 2:
- BuildMI(BB, X86::IN16, 0);
- break;
- case 4:
- BuildMI(BB, X86::IN32, 0);
- break;
- default:
- std::cerr << "Cannot do input on this data type";
- exit (1);
+ unsigned Class = getClass(CI.getCalledFunction()->getReturnType());
+ unsigned DestReg = getReg(CI);
+
+ // If the port is a single-byte constant, use the immediate form.
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CI.getOperand(1)))
+ if ((C->getRawValue() & 255) == C->getRawValue()) {
+ switch (Class) {
+ case cByte:
+ BuildMI(BB, X86::IN8ri, 1).addImm((unsigned char)C->getRawValue());
+ BuildMI(BB, X86::MOV8rr, 1, DestReg).addReg(X86::AL);
+ return;
+ case cShort:
+ BuildMI(BB, X86::IN16ri, 1).addImm((unsigned char)C->getRawValue());
+ BuildMI(BB, X86::MOV8rr, 1, DestReg).addReg(X86::AX);
+ return;
+ case cInt:
+ BuildMI(BB, X86::IN32ri, 1).addImm((unsigned char)C->getRawValue());
+ BuildMI(BB, X86::MOV8rr, 1, DestReg).addReg(X86::EAX);
+ return;
+ }
+ }
+
+ unsigned Reg = getReg(CI.getOperand(1));
+ BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(Reg);
+ switch (Class) {
+ case cByte:
+ BuildMI(BB, X86::IN8rr, 0);
+ BuildMI(BB, X86::MOV8rr, 1, DestReg).addReg(X86::AL);
+ break;
+ case cShort:
+ BuildMI(BB, X86::IN16rr, 0);
+ BuildMI(BB, X86::MOV8rr, 1, DestReg).addReg(X86::AX);
+ break;
+ case cInt:
+ BuildMI(BB, X86::IN32rr, 0);
+ BuildMI(BB, X86::MOV8rr, 1, DestReg).addReg(X86::EAX);
+ break;
+ default:
+ std::cerr << "Cannot do input on this data type";
+ exit (1);
}
return;
+ }
- case Intrinsic::writeport:
- //
+ case Intrinsic::writeport: {
// First, determine that the size of the operand falls within the
// acceptable range for this architecture.
- //
- //
- if ((CI.getOperand(2)->getType()->getPrimitiveSize()) != 2) {
+ if (getClass(CI.getOperand(2)->getType()) != cShort) {
std::cerr << "llvm.writeport: Address size is not 16 bits\n";
- exit (1);
+ exit(1);
}
- //
- // Now, move the I/O port address into the DX register and the value to
- // write into the AL/AX/EAX register.
- //
- BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(getReg(CI.getOperand(2)));
- switch (CI.getOperand(1)->getType()->getPrimitiveSize()) {
- case 1:
- BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(getReg(CI.getOperand(1)));
- BuildMI(BB, X86::OUT8, 0);
- break;
- case 2:
- BuildMI(BB, X86::MOV16rr, 1, X86::AX).addReg(getReg(CI.getOperand(1)));
- BuildMI(BB, X86::OUT16, 0);
- break;
- case 4:
- BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(getReg(CI.getOperand(1)));
- BuildMI(BB, X86::OUT32, 0);
- break;
- default:
- std::cerr << "Cannot do output on this data type";
- exit (1);
+ unsigned Class = getClassB(CI.getOperand(1)->getType());
+ unsigned ValReg = getReg(CI.getOperand(1));
+ switch (Class) {
+ case cByte:
+ BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(ValReg);
+ break;
+ case cShort:
+ BuildMI(BB, X86::MOV16rr, 1, X86::AX).addReg(ValReg);
+ break;
+ case cInt:
+ BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(ValReg);
+ break;
+ default:
+ std::cerr << "llvm.writeport: invalid data type for X86 target";
+ exit(1);
}
- return;
+
+ // If the port is a single-byte constant, use the immediate form.
+ if (ConstantInt *C = dyn_cast<ConstantInt>(CI.getOperand(2)))
+ if ((C->getRawValue() & 255) == C->getRawValue()) {
+ static const unsigned O[] = { X86::OUT8ir, X86::OUT16ir, X86::OUT32ir };
+ BuildMI(BB, O[Class], 1).addImm((unsigned char)C->getRawValue());
+ return;
+ }
+
+ // Otherwise, move the I/O port address into the DX register and the value
+ // to write into the AL/AX/EAX register.
+ static const unsigned Opc[] = { X86::OUT8rr, X86::OUT16rr, X86::OUT32rr };
+ unsigned Reg = getReg(CI.getOperand(2));
+ BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(Reg);
+ BuildMI(BB, Opc[Class], 0);
+ return;
+ }
+
default: assert(0 && "Error: unknown intrinsics should have been lowered!");
}
}
case Instruction::Call:
case Instruction::Invoke:
return false;
+ case Instruction::Load:
+ if (cast<LoadInst>(It)->isVolatile() && LI.isVolatile())
+ return false;
+ break;
}
}
return true;
}
-
/// visitSimpleBinary - Implement simple binary operators for integral types...
/// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or, 4 for
/// Xor.
///
-void ISel::visitSimpleBinary(BinaryOperator &B, unsigned OperatorClass) {
+void X86ISel::visitSimpleBinary(BinaryOperator &B, unsigned OperatorClass) {
unsigned DestReg = getReg(B);
MachineBasicBlock::iterator MI = BB->end();
Value *Op0 = B.getOperand(0), *Op1 = B.getOperand(1);
+ unsigned Class = getClassB(B.getType());
+
+ // If this is AND X, C, and it is only used by a setcc instruction, it will
+ // be folded. There is no need to emit this instruction.
+ if (B.hasOneUse() && OperatorClass == 2 && isa<ConstantInt>(Op1))
+ if (Class == cByte || Class == cShort || Class == cInt) {
+ Instruction *Use = cast<Instruction>(B.use_back());
+ if (isa<SetCondInst>(Use) &&
+ Use->getOperand(1) == Constant::getNullValue(B.getType())) {
+ switch (getSetCCNumber(Use->getOpcode())) {
+ case 0:
+ case 1:
+ return;
+ default:
+ if (B.getType()->isSigned()) return;
+ }
+ }
+ }
// Special case: op Reg, load [mem]
- if (isa<LoadInst>(Op0) && !isa<LoadInst>(Op1))
+ if (isa<LoadInst>(Op0) && !isa<LoadInst>(Op1) && Class != cLong &&
+ Op0->hasOneUse() &&
+ isSafeToFoldLoadIntoInstruction(*cast<LoadInst>(Op0), B))
if (!B.swapOperands())
std::swap(Op0, Op1); // Make sure any loads are in the RHS.
- unsigned Class = getClassB(B.getType());
- if (isa<LoadInst>(Op1) && Class < cFP &&
+ if (isa<LoadInst>(Op1) && Class != cLong && Op1->hasOneUse() &&
isSafeToFoldLoadIntoInstruction(*cast<LoadInst>(Op1), B)) {
- static const unsigned OpcodeTab[][3] = {
- // Arithmetic operators
- { X86::ADD8rm, X86::ADD16rm, X86::ADD32rm }, // ADD
- { X86::SUB8rm, X86::SUB16rm, X86::SUB32rm }, // SUB
-
- // Bitwise operators
- { X86::AND8rm, X86::AND16rm, X86::AND32rm }, // AND
- { X86:: OR8rm, X86:: OR16rm, X86:: OR32rm }, // OR
- { X86::XOR8rm, X86::XOR16rm, X86::XOR32rm }, // XOR
- };
-
- assert(Class < cFP && "General code handles 64-bit integer types!");
- unsigned Opcode = OpcodeTab[OperatorClass][Class];
-
- unsigned BaseReg, Scale, IndexReg, Disp;
- getAddressingMode(cast<LoadInst>(Op1)->getOperand(0), BaseReg,
- Scale, IndexReg, Disp);
+ unsigned Opcode;
+ if (Class != cFP) {
+ static const unsigned OpcodeTab[][3] = {
+ // Arithmetic operators
+ { X86::ADD8rm, X86::ADD16rm, X86::ADD32rm }, // ADD
+ { X86::SUB8rm, X86::SUB16rm, X86::SUB32rm }, // SUB
+
+ // Bitwise operators
+ { X86::AND8rm, X86::AND16rm, X86::AND32rm }, // AND
+ { X86:: OR8rm, X86:: OR16rm, X86:: OR32rm }, // OR
+ { X86::XOR8rm, X86::XOR16rm, X86::XOR32rm }, // XOR
+ };
+ Opcode = OpcodeTab[OperatorClass][Class];
+ } else {
+ static const unsigned OpcodeTab[][2] = {
+ { X86::FADD32m, X86::FADD64m }, // ADD
+ { X86::FSUB32m, X86::FSUB64m }, // SUB
+ };
+ const Type *Ty = Op0->getType();
+ assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
+ Opcode = OpcodeTab[OperatorClass][Ty == Type::DoubleTy];
+ }
unsigned Op0r = getReg(Op0);
- addFullAddress(BuildMI(BB, Opcode, 2, DestReg).addReg(Op0r),
- BaseReg, Scale, IndexReg, Disp);
- return;
+ if (AllocaInst *AI =
+ dyn_castFixedAlloca(cast<LoadInst>(Op1)->getOperand(0))) {
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ addFrameReference(BuildMI(BB, Opcode, 5, DestReg).addReg(Op0r), FI);
+
+ } else {
+ X86AddressMode AM;
+ getAddressingMode(cast<LoadInst>(Op1)->getOperand(0), AM);
+
+ addFullAddress(BuildMI(BB, Opcode, 5, DestReg).addReg(Op0r), AM);
+ }
+ return;
+ }
+
+ // If this is a floating point subtract, check to see if we can fold the first
+ // operand in.
+ if (Class == cFP && OperatorClass == 1 &&
+ isa<LoadInst>(Op0) &&
+ isSafeToFoldLoadIntoInstruction(*cast<LoadInst>(Op0), B)) {
+ const Type *Ty = Op0->getType();
+ assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
+ unsigned Opcode = Ty == Type::FloatTy ? X86::FSUBR32m : X86::FSUBR64m;
+
+ unsigned Op1r = getReg(Op1);
+ if (AllocaInst *AI =
+ dyn_castFixedAlloca(cast<LoadInst>(Op0)->getOperand(0))) {
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ addFrameReference(BuildMI(BB, Opcode, 5, DestReg).addReg(Op1r), FI);
+ } else {
+ X86AddressMode AM;
+ getAddressingMode(cast<LoadInst>(Op0)->getOperand(0), AM);
+
+ addFullAddress(BuildMI(BB, Opcode, 5, DestReg).addReg(Op1r), AM);
+ }
+ return;
}
emitSimpleBinaryOperation(BB, MI, Op0, Op1, OperatorClass, DestReg);
}
+
+/// emitBinaryFPOperation - This method handles emission of floating point
+/// Add (0), Sub (1), Mul (2), and Div (3) operations.
+void X86ISel::emitBinaryFPOperation(MachineBasicBlock *BB,
+ MachineBasicBlock::iterator IP,
+ Value *Op0, Value *Op1,
+ unsigned OperatorClass, unsigned DestReg) {
+ // Special case: op Reg, <const fp>
+ if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1))
+ if (!Op1C->isExactlyValue(+0.0) && !Op1C->isExactlyValue(+1.0)) {
+ // Create a constant pool entry for this constant.
+ MachineConstantPool *CP = F->getConstantPool();
+ unsigned CPI = CP->getConstantPoolIndex(Op1C);
+ const Type *Ty = Op1->getType();
+
+ static const unsigned OpcodeTab[][4] = {
+ { X86::FADD32m, X86::FSUB32m, X86::FMUL32m, X86::FDIV32m }, // Float
+ { X86::FADD64m, X86::FSUB64m, X86::FMUL64m, X86::FDIV64m }, // Double
+ };
+
+ assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
+ unsigned Opcode = OpcodeTab[Ty != Type::FloatTy][OperatorClass];
+ unsigned Op0r = getReg(Op0, BB, IP);
+ addConstantPoolReference(BuildMI(*BB, IP, Opcode, 5,
+ DestReg).addReg(Op0r), CPI);
+ return;
+ }
+
+ // Special case: R1 = op <const fp>, R2
+ if (ConstantFP *CFP = dyn_cast<ConstantFP>(Op0))
+ if (CFP->isExactlyValue(-0.0) && OperatorClass == 1) {
+ // -0.0 - X === -X
+ unsigned op1Reg = getReg(Op1, BB, IP);
+ BuildMI(*BB, IP, X86::FCHS, 1, DestReg).addReg(op1Reg);
+ return;
+ } else if (!CFP->isExactlyValue(+0.0) && !CFP->isExactlyValue(+1.0)) {
+ // R1 = op CST, R2 --> R1 = opr R2, CST
+
+ // Create a constant pool entry for this constant.
+ MachineConstantPool *CP = F->getConstantPool();
+ unsigned CPI = CP->getConstantPoolIndex(CFP);
+ const Type *Ty = CFP->getType();
+
+ static const unsigned OpcodeTab[][4] = {
+ { X86::FADD32m, X86::FSUBR32m, X86::FMUL32m, X86::FDIVR32m }, // Float
+ { X86::FADD64m, X86::FSUBR64m, X86::FMUL64m, X86::FDIVR64m }, // Double
+ };
+
+ assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
+ unsigned Opcode = OpcodeTab[Ty != Type::FloatTy][OperatorClass];
+ unsigned Op1r = getReg(Op1, BB, IP);
+ addConstantPoolReference(BuildMI(*BB, IP, Opcode, 5,
+ DestReg).addReg(Op1r), CPI);
+ return;
+ }
+
+ // General case.
+ static const unsigned OpcodeTab[4] = {
+ X86::FpADD, X86::FpSUB, X86::FpMUL, X86::FpDIV
+ };
+
+ unsigned Opcode = OpcodeTab[OperatorClass];
+ unsigned Op0r = getReg(Op0, BB, IP);
+ unsigned Op1r = getReg(Op1, BB, IP);
+ BuildMI(*BB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r);
+}
+
/// emitSimpleBinaryOperation - Implement simple binary operators for integral
/// types... OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for
/// Or, 4 for Xor.
/// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary
/// and constant expression support.
///
-void ISel::emitSimpleBinaryOperation(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP,
- Value *Op0, Value *Op1,
- unsigned OperatorClass, unsigned DestReg) {
+void X86ISel::emitSimpleBinaryOperation(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ Value *Op0, Value *Op1,
+ unsigned OperatorClass,
+ unsigned DestReg) {
unsigned Class = getClassB(Op0->getType());
- // sub 0, X -> neg X
- if (OperatorClass == 1)
- if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0)) {
+ if (Class == cFP) {
+ assert(OperatorClass < 2 && "No logical ops for FP!");
+ emitBinaryFPOperation(MBB, IP, Op0, Op1, OperatorClass, DestReg);
+ return;
+ }
+
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0))
+ if (OperatorClass == 1) {
+ static unsigned const NEGTab[] = {
+ X86::NEG8r, X86::NEG16r, X86::NEG32r, 0, X86::NEG32r
+ };
+
+ // sub 0, X -> neg X
if (CI->isNullValue()) {
unsigned op1Reg = getReg(Op1, MBB, IP);
- static unsigned const NEGTab[] = {
- X86::NEG8r, X86::NEG16r, X86::NEG32r, 0, X86::NEG32r
- };
BuildMI(*MBB, IP, NEGTab[Class], 1, DestReg).addReg(op1Reg);
-
+
if (Class == cLong) {
// We just emitted: Dl = neg Sl
// Now emit : T = addc Sh, 0
BuildMI(*MBB, IP, X86::NEG32r, 1, DestReg+1).addReg(T);
}
return;
- }
- } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(Op0))
- if (CFP->isExactlyValue(-0.0)) {
- // -0.0 - X === -X
+ } else if (Op1->hasOneUse() && Class != cLong) {
+ // sub C, X -> tmp = neg X; DestReg = add tmp, C. This is better
+ // than copying C into a temporary register, because of register
+ // pressure (tmp and destreg can share a register.
+ static unsigned const ADDRITab[] = {
+ X86::ADD8ri, X86::ADD16ri, X86::ADD32ri, 0, X86::ADD32ri
+ };
unsigned op1Reg = getReg(Op1, MBB, IP);
- BuildMI(*MBB, IP, X86::FCHS, 1, DestReg).addReg(op1Reg);
+ unsigned Tmp = makeAnotherReg(Op0->getType());
+ BuildMI(*MBB, IP, NEGTab[Class], 1, Tmp).addReg(op1Reg);
+ BuildMI(*MBB, IP, ADDRITab[Class], 2,
+ DestReg).addReg(Tmp).addImm(CI->getRawValue());
return;
}
+ }
- // Special case: op Reg, <const>
- if (isa<ConstantInt>(Op1)) {
- ConstantInt *Op1C = cast<ConstantInt>(Op1);
+ // Special case: op Reg, <const int>
+ if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
unsigned Op0r = getReg(Op0, MBB, IP);
// xor X, -1 -> not X
if (Class != cLong) {
BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addImm(Op1l);
return;
- } else {
- // If this is a long value and the high or low bits have a special
- // property, emit some special cases.
- unsigned Op1h = cast<ConstantInt>(Op1C)->getRawValue() >> 32LL;
-
- // If the constant is zero in the low 32-bits, just copy the low part
- // across and apply the normal 32-bit operation to the high parts. There
- // will be no carry or borrow into the top.
- if (Op1l == 0) {
- if (OperatorClass != 2) // All but and...
- BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg).addReg(Op0r);
- else
- BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
- BuildMI(*MBB, IP, OpcodeTab[OperatorClass][cLong], 2, DestReg+1)
- .addReg(Op0r+1).addImm(Op1h);
- return;
- }
-
- // If this is a logical operation and the top 32-bits are zero, just
- // operate on the lower 32.
- if (Op1h == 0 && OperatorClass > 1) {
- BuildMI(*MBB, IP, OpcodeTab[OperatorClass][cLong], 2, DestReg)
- .addReg(Op0r).addImm(Op1l);
- if (OperatorClass != 2) // All but and
- BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg+1).addReg(Op0r+1);
- else
- BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg+1).addImm(0);
- return;
- }
-
- // TODO: We could handle lots of other special cases here, such as AND'ing
- // with 0xFFFFFFFF00000000 -> noop, etc.
-
- // Otherwise, code generate the full operation with a constant.
- static const unsigned TopTab[] = {
- X86::ADC32ri, X86::SBB32ri, X86::AND32ri, X86::OR32ri, X86::XOR32ri
- };
-
- BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addImm(Op1l);
- BuildMI(*MBB, IP, TopTab[OperatorClass], 2, DestReg+1)
- .addReg(Op0r+1).addImm(Op1h);
+ }
+
+ // If this is a long value and the high or low bits have a special
+ // property, emit some special cases.
+ unsigned Op1h = cast<ConstantInt>(Op1C)->getRawValue() >> 32LL;
+
+ // If the constant is zero in the low 32-bits, just copy the low part
+ // across and apply the normal 32-bit operation to the high parts. There
+ // will be no carry or borrow into the top.
+ if (Op1l == 0) {
+ if (OperatorClass != 2) // All but and...
+ BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg).addReg(Op0r);
+ else
+ BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
+ BuildMI(*MBB, IP, OpcodeTab[OperatorClass][cLong], 2, DestReg+1)
+ .addReg(Op0r+1).addImm(Op1h);
+ return;
+ }
+
+ // If this is a logical operation and the top 32-bits are zero, just
+ // operate on the lower 32.
+ if (Op1h == 0 && OperatorClass > 1) {
+ BuildMI(*MBB, IP, OpcodeTab[OperatorClass][cLong], 2, DestReg)
+ .addReg(Op0r).addImm(Op1l);
+ if (OperatorClass != 2) // All but and
+ BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg+1).addReg(Op0r+1);
+ else
+ BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg+1).addImm(0);
return;
}
+
+ // TODO: We could handle lots of other special cases here, such as AND'ing
+ // with 0xFFFFFFFF00000000 -> noop, etc.
+
+ // Otherwise, code generate the full operation with a constant.
+ static const unsigned TopTab[] = {
+ X86::ADC32ri, X86::SBB32ri, X86::AND32ri, X86::OR32ri, X86::XOR32ri
+ };
+
+ BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addImm(Op1l);
+ BuildMI(*MBB, IP, TopTab[OperatorClass], 2, DestReg+1)
+ .addReg(Op0r+1).addImm(Op1h);
+ return;
}
// Finally, handle the general case now.
static const unsigned OpcodeTab[][5] = {
// Arithmetic operators
- { X86::ADD8rr, X86::ADD16rr, X86::ADD32rr, X86::FpADD, X86::ADD32rr },// ADD
- { X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, X86::FpSUB, X86::SUB32rr },// SUB
+ { X86::ADD8rr, X86::ADD16rr, X86::ADD32rr, 0, X86::ADD32rr }, // ADD
+ { X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, 0, X86::SUB32rr }, // SUB
// Bitwise operators
{ X86::AND8rr, X86::AND16rr, X86::AND32rr, 0, X86::AND32rr }, // AND
};
unsigned Opcode = OpcodeTab[OperatorClass][Class];
- assert(Opcode && "Floating point arguments to logical inst?");
unsigned Op0r = getReg(Op0, MBB, IP);
unsigned Op1r = getReg(Op1, MBB, IP);
BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r);
/// registers op0Reg and op1Reg, and put the result in DestReg. The type of the
/// result should be given as DestTy.
///
-void ISel::doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator MBBI,
- unsigned DestReg, const Type *DestTy,
- unsigned op0Reg, unsigned op1Reg) {
+void X86ISel::doMultiply(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator MBBI,
+ unsigned DestReg, const Type *DestTy,
+ unsigned op0Reg, unsigned op1Reg) {
unsigned Class = getClass(DestTy);
switch (Class) {
- case cFP: // Floating point multiply
- BuildMI(*MBB, MBBI, X86::FpMUL, 2, DestReg).addReg(op0Reg).addReg(op1Reg);
- return;
case cInt:
case cShort:
BuildMI(*MBB, MBBI, Class == cInt ? X86::IMUL32rr:X86::IMUL16rr, 2, DestReg)
// ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
// returns zero when the input is not exactly a power of two.
static unsigned ExactLog2(unsigned Val) {
- if (Val == 0) return 0;
+ if (Val == 0 || (Val & (Val-1))) return 0;
unsigned Count = 0;
while (Val != 1) {
- if (Val & 1) return 0;
Val >>= 1;
++Count;
}
return Count+1;
}
-void ISel::doMultiplyConst(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP,
- unsigned DestReg, const Type *DestTy,
- unsigned op0Reg, unsigned ConstRHS) {
+
+/// doMultiplyConst - This function is specialized to efficiently codegen an 8,
+/// 16, or 32-bit integer multiply by a constant.
+void X86ISel::doMultiplyConst(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ unsigned DestReg, const Type *DestTy,
+ unsigned op0Reg, unsigned ConstRHS) {
static const unsigned MOVrrTab[] = {X86::MOV8rr, X86::MOV16rr, X86::MOV32rr};
static const unsigned MOVriTab[] = {X86::MOV8ri, X86::MOV16ri, X86::MOV32ri};
+ static const unsigned ADDrrTab[] = {X86::ADD8rr, X86::ADD16rr, X86::ADD32rr};
+ static const unsigned NEGrTab[] = {X86::NEG8r , X86::NEG16r , X86::NEG32r };
unsigned Class = getClass(DestTy);
-
- if (ConstRHS == 0) {
+ unsigned TmpReg;
+
+ // Handle special cases here.
+ switch (ConstRHS) {
+ case -2:
+ TmpReg = makeAnotherReg(DestTy);
+ BuildMI(*MBB, IP, NEGrTab[Class], 1, TmpReg).addReg(op0Reg);
+ BuildMI(*MBB, IP, ADDrrTab[Class], 1,DestReg).addReg(TmpReg).addReg(TmpReg);
+ return;
+ case -1:
+ BuildMI(*MBB, IP, NEGrTab[Class], 1, DestReg).addReg(op0Reg);
+ return;
+ case 0:
BuildMI(*MBB, IP, MOVriTab[Class], 1, DestReg).addImm(0);
return;
- } else if (ConstRHS == 1) {
+ case 1:
BuildMI(*MBB, IP, MOVrrTab[Class], 1, DestReg).addReg(op0Reg);
return;
+ case 2:
+ BuildMI(*MBB, IP, ADDrrTab[Class], 1,DestReg).addReg(op0Reg).addReg(op0Reg);
+ return;
+ case 3:
+ case 5:
+ case 9:
+ if (Class == cInt) {
+ X86AddressMode AM;
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = op0Reg;
+ AM.Scale = ConstRHS-1;
+ AM.IndexReg = op0Reg;
+ AM.Disp = 0;
+ addFullAddress(BuildMI(*MBB, IP, X86::LEA32r, 5, DestReg), AM);
+ return;
+ }
+ case -3:
+ case -5:
+ case -9:
+ if (Class == cInt) {
+ TmpReg = makeAnotherReg(DestTy);
+ X86AddressMode AM;
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = op0Reg;
+ AM.Scale = -ConstRHS-1;
+ AM.IndexReg = op0Reg;
+ AM.Disp = 0;
+ addFullAddress(BuildMI(*MBB, IP, X86::LEA32r, 5, TmpReg), AM);
+ BuildMI(*MBB, IP, NEGrTab[Class], 1, DestReg).addReg(TmpReg);
+ return;
+ }
}
// If the element size is exactly a power of 2, use a shift to get it.
switch (Class) {
default: assert(0 && "Unknown class for this function!");
case cByte:
- BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
+ BuildMI(*MBB, IP, X86::SHL8ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
return;
case cShort:
- BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
+ BuildMI(*MBB, IP, X86::SHL16ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
return;
case cInt:
BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
return;
}
}
+
+ // If the element size is a negative power of 2, use a shift/neg to get it.
+ if (unsigned Shift = ExactLog2(-ConstRHS)) {
+ TmpReg = makeAnotherReg(DestTy);
+ BuildMI(*MBB, IP, NEGrTab[Class], 1, TmpReg).addReg(op0Reg);
+ switch (Class) {
+ default: assert(0 && "Unknown class for this function!");
+ case cByte:
+ BuildMI(*MBB, IP, X86::SHL8ri,2, DestReg).addReg(TmpReg).addImm(Shift-1);
+ return;
+ case cShort:
+ BuildMI(*MBB, IP, X86::SHL16ri,2, DestReg).addReg(TmpReg).addImm(Shift-1);
+ return;
+ case cInt:
+ BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(TmpReg).addImm(Shift-1);
+ return;
+ }
+ }
if (Class == cShort) {
BuildMI(*MBB, IP, X86::IMUL16rri,2,DestReg).addReg(op0Reg).addImm(ConstRHS);
}
// Most general case, emit a normal multiply...
- unsigned TmpReg = makeAnotherReg(DestTy);
+ TmpReg = makeAnotherReg(DestTy);
BuildMI(*MBB, IP, MOVriTab[Class], 1, TmpReg).addImm(ConstRHS);
// Emit a MUL to multiply the register holding the index by
/// visitMul - Multiplies are not simple binary operators because they must deal
/// with the EAX register explicitly.
///
-void ISel::visitMul(BinaryOperator &I) {
- unsigned Op0Reg = getReg(I.getOperand(0));
- unsigned DestReg = getReg(I);
+void X86ISel::visitMul(BinaryOperator &I) {
+ unsigned ResultReg = getReg(I);
+
+ Value *Op0 = I.getOperand(0);
+ Value *Op1 = I.getOperand(1);
+
+ // Fold loads into floating point multiplies.
+ if (getClass(Op0->getType()) == cFP) {
+ if (isa<LoadInst>(Op0) && !isa<LoadInst>(Op1))
+ if (!I.swapOperands())
+ std::swap(Op0, Op1); // Make sure any loads are in the RHS.
+ if (LoadInst *LI = dyn_cast<LoadInst>(Op1))
+ if (isSafeToFoldLoadIntoInstruction(*LI, I)) {
+ const Type *Ty = Op0->getType();
+ assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
+ unsigned Opcode = Ty == Type::FloatTy ? X86::FMUL32m : X86::FMUL64m;
+
+ unsigned Op0r = getReg(Op0);
+ if (AllocaInst *AI = dyn_castFixedAlloca(LI->getOperand(0))) {
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ addFrameReference(BuildMI(BB, Opcode, 5, ResultReg).addReg(Op0r), FI);
+ } else {
+ X86AddressMode AM;
+ getAddressingMode(LI->getOperand(0), AM);
+
+ addFullAddress(BuildMI(BB, Opcode, 5, ResultReg).addReg(Op0r), AM);
+ }
+ return;
+ }
+ }
+
+ MachineBasicBlock::iterator IP = BB->end();
+ emitMultiply(BB, IP, Op0, Op1, ResultReg);
+}
+
+void X86ISel::emitMultiply(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ Value *Op0, Value *Op1, unsigned DestReg) {
+ MachineBasicBlock &BB = *MBB;
+ TypeClass Class = getClass(Op0->getType());
// Simple scalar multiply?
- if (getClass(I.getType()) != cLong) {
- if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1))) {
- unsigned Val = (unsigned)CI->getRawValue(); // Cannot be 64-bit constant
- MachineBasicBlock::iterator MBBI = BB->end();
- doMultiplyConst(BB, MBBI, DestReg, I.getType(), Op0Reg, Val);
+ unsigned Op0Reg = getReg(Op0, &BB, IP);
+ switch (Class) {
+ case cByte:
+ case cShort:
+ case cInt:
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
+ unsigned Val = (unsigned)CI->getRawValue(); // Isn't a 64-bit constant
+ doMultiplyConst(&BB, IP, DestReg, Op0->getType(), Op0Reg, Val);
} else {
- unsigned Op1Reg = getReg(I.getOperand(1));
- MachineBasicBlock::iterator MBBI = BB->end();
- doMultiply(BB, MBBI, DestReg, I.getType(), Op0Reg, Op1Reg);
+ unsigned Op1Reg = getReg(Op1, &BB, IP);
+ doMultiply(&BB, IP, DestReg, Op1->getType(), Op0Reg, Op1Reg);
}
- } else {
- // Long value. We have to do things the hard way...
- if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(1))) {
- unsigned CLow = CI->getRawValue();
- unsigned CHi = CI->getRawValue() >> 32;
-
- if (CLow == 0) {
- // If the low part of the constant is all zeros, things are simple.
- BuildMI(BB, X86::MOV32ri, 1, DestReg).addImm(0);
- doMultiplyConst(BB, BB->end(), DestReg+1, Type::UIntTy, Op0Reg, CHi);
- return;
- }
+ return;
+ case cFP:
+ emitBinaryFPOperation(MBB, IP, Op0, Op1, 2, DestReg);
+ return;
+ case cLong:
+ break;
+ }
- // Multiply the two low parts... capturing carry into EDX
- unsigned OverflowReg = 0;
- if (CLow == 1) {
- BuildMI(BB, X86::MOV32rr, 1, DestReg).addReg(Op0Reg);
- } else {
- unsigned Op1RegL = makeAnotherReg(Type::UIntTy);
- OverflowReg = makeAnotherReg(Type::UIntTy);
- BuildMI(BB, X86::MOV32ri, 1, Op1RegL).addImm(CLow);
- BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Op0Reg);
- BuildMI(BB, X86::MUL32r, 1).addReg(Op1RegL); // AL*BL
-
- BuildMI(BB, X86::MOV32rr, 1, DestReg).addReg(X86::EAX); // AL*BL
- BuildMI(BB, X86::MOV32rr, 1,OverflowReg).addReg(X86::EDX);// AL*BL >> 32
- }
-
- unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL
- doMultiplyConst(BB, BB->end(), AHBLReg, Type::UIntTy, Op0Reg+1, CLow);
-
- unsigned AHBLplusOverflowReg;
- if (OverflowReg) {
- AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy);
- BuildMI(BB, X86::ADD32rr, 2, // AH*BL+(AL*BL >> 32)
- AHBLplusOverflowReg).addReg(AHBLReg).addReg(OverflowReg);
- } else {
- AHBLplusOverflowReg = AHBLReg;
- }
-
- if (CHi == 0) {
- BuildMI(BB, X86::MOV32rr, 1, DestReg+1).addReg(AHBLplusOverflowReg);
- } else {
- unsigned ALBHReg = makeAnotherReg(Type::UIntTy); // AL*BH
- doMultiplyConst(BB, BB->end(), ALBHReg, Type::UIntTy, Op0Reg, CHi);
-
- BuildMI(BB, X86::ADD32rr, 2, // AL*BH + AH*BL + (AL*BL >> 32)
- DestReg+1).addReg(AHBLplusOverflowReg).addReg(ALBHReg);
- }
+ // Long value. We have to do things the hard way...
+ if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
+ unsigned CLow = CI->getRawValue();
+ unsigned CHi = CI->getRawValue() >> 32;
+
+ if (CLow == 0) {
+ // If the low part of the constant is all zeros, things are simple.
+ BuildMI(BB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
+ doMultiplyConst(&BB, IP, DestReg+1, Type::UIntTy, Op0Reg, CHi);
+ return;
+ }
+
+ // Multiply the two low parts... capturing carry into EDX
+ unsigned OverflowReg = 0;
+ if (CLow == 1) {
+ BuildMI(BB, IP, X86::MOV32rr, 1, DestReg).addReg(Op0Reg);
} else {
- unsigned Op1Reg = getReg(I.getOperand(1));
- // Multiply the two low parts... capturing carry into EDX
- BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(Op0Reg);
- BuildMI(BB, X86::MUL32r, 1).addReg(Op1Reg); // AL*BL
-
- unsigned OverflowReg = makeAnotherReg(Type::UIntTy);
- BuildMI(BB, X86::MOV32rr, 1, DestReg).addReg(X86::EAX); // AL*BL
- BuildMI(BB, X86::MOV32rr, 1, OverflowReg).addReg(X86::EDX); // AL*BL >> 32
+ unsigned Op1RegL = makeAnotherReg(Type::UIntTy);
+ OverflowReg = makeAnotherReg(Type::UIntTy);
+ BuildMI(BB, IP, X86::MOV32ri, 1, Op1RegL).addImm(CLow);
+ BuildMI(BB, IP, X86::MOV32rr, 1, X86::EAX).addReg(Op0Reg);
+ BuildMI(BB, IP, X86::MUL32r, 1).addReg(Op1RegL); // AL*BL
- MachineBasicBlock::iterator MBBI = BB->end();
- unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL
- BuildMI(*BB, MBBI, X86::IMUL32rr, 2,
- AHBLReg).addReg(Op0Reg+1).addReg(Op1Reg);
-
- unsigned AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy);
- BuildMI(*BB, MBBI, X86::ADD32rr, 2, // AH*BL+(AL*BL >> 32)
+ BuildMI(BB, IP, X86::MOV32rr, 1, DestReg).addReg(X86::EAX); // AL*BL
+ BuildMI(BB, IP, X86::MOV32rr, 1,
+ OverflowReg).addReg(X86::EDX); // AL*BL >> 32
+ }
+
+ unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL
+ doMultiplyConst(&BB, IP, AHBLReg, Type::UIntTy, Op0Reg+1, CLow);
+
+ unsigned AHBLplusOverflowReg;
+ if (OverflowReg) {
+ AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy);
+ BuildMI(BB, IP, X86::ADD32rr, 2, // AH*BL+(AL*BL >> 32)
AHBLplusOverflowReg).addReg(AHBLReg).addReg(OverflowReg);
-
- MBBI = BB->end();
+ } else {
+ AHBLplusOverflowReg = AHBLReg;
+ }
+
+ if (CHi == 0) {
+ BuildMI(BB, IP, X86::MOV32rr, 1, DestReg+1).addReg(AHBLplusOverflowReg);
+ } else {
unsigned ALBHReg = makeAnotherReg(Type::UIntTy); // AL*BH
- BuildMI(*BB, MBBI, X86::IMUL32rr, 2,
- ALBHReg).addReg(Op0Reg).addReg(Op1Reg+1);
+ doMultiplyConst(&BB, IP, ALBHReg, Type::UIntTy, Op0Reg, CHi);
- BuildMI(*BB, MBBI, X86::ADD32rr, 2, // AL*BH + AH*BL + (AL*BL >> 32)
+ BuildMI(BB, IP, X86::ADD32rr, 2, // AL*BH + AH*BL + (AL*BL >> 32)
DestReg+1).addReg(AHBLplusOverflowReg).addReg(ALBHReg);
}
+ return;
}
+
+ // General 64x64 multiply
+
+ unsigned Op1Reg = getReg(Op1, &BB, IP);
+ // Multiply the two low parts... capturing carry into EDX
+ BuildMI(BB, IP, X86::MOV32rr, 1, X86::EAX).addReg(Op0Reg);
+ BuildMI(BB, IP, X86::MUL32r, 1).addReg(Op1Reg); // AL*BL
+
+ unsigned OverflowReg = makeAnotherReg(Type::UIntTy);
+ BuildMI(BB, IP, X86::MOV32rr, 1, DestReg).addReg(X86::EAX); // AL*BL
+ BuildMI(BB, IP, X86::MOV32rr, 1,
+ OverflowReg).addReg(X86::EDX); // AL*BL >> 32
+
+ unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL
+ BuildMI(BB, IP, X86::IMUL32rr, 2,
+ AHBLReg).addReg(Op0Reg+1).addReg(Op1Reg);
+
+ unsigned AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy);
+ BuildMI(BB, IP, X86::ADD32rr, 2, // AH*BL+(AL*BL >> 32)
+ AHBLplusOverflowReg).addReg(AHBLReg).addReg(OverflowReg);
+
+ unsigned ALBHReg = makeAnotherReg(Type::UIntTy); // AL*BH
+ BuildMI(BB, IP, X86::IMUL32rr, 2,
+ ALBHReg).addReg(Op0Reg).addReg(Op1Reg+1);
+
+ BuildMI(BB, IP, X86::ADD32rr, 2, // AL*BH + AH*BL + (AL*BL >> 32)
+ DestReg+1).addReg(AHBLplusOverflowReg).addReg(ALBHReg);
}
/// select the result from a different register. Note that both of these
/// instructions work differently for signed and unsigned operands.
///
-void ISel::visitDivRem(BinaryOperator &I) {
- unsigned Op0Reg = getReg(I.getOperand(0));
- unsigned Op1Reg = getReg(I.getOperand(1));
+void X86ISel::visitDivRem(BinaryOperator &I) {
unsigned ResultReg = getReg(I);
+ Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
+
+ // Fold loads into floating point divides.
+ if (getClass(Op0->getType()) == cFP) {
+ if (LoadInst *LI = dyn_cast<LoadInst>(Op1))
+ if (isSafeToFoldLoadIntoInstruction(*LI, I)) {
+ const Type *Ty = Op0->getType();
+ assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
+ unsigned Opcode = Ty == Type::FloatTy ? X86::FDIV32m : X86::FDIV64m;
+
+ unsigned Op0r = getReg(Op0);
+ if (AllocaInst *AI = dyn_castFixedAlloca(LI->getOperand(0))) {
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ addFrameReference(BuildMI(BB, Opcode, 5, ResultReg).addReg(Op0r), FI);
+ } else {
+ X86AddressMode AM;
+ getAddressingMode(LI->getOperand(0), AM);
+
+ addFullAddress(BuildMI(BB, Opcode, 5, ResultReg).addReg(Op0r), AM);
+ }
+ return;
+ }
+
+ if (LoadInst *LI = dyn_cast<LoadInst>(Op0))
+ if (isSafeToFoldLoadIntoInstruction(*LI, I)) {
+ const Type *Ty = Op0->getType();
+ assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
+ unsigned Opcode = Ty == Type::FloatTy ? X86::FDIVR32m : X86::FDIVR64m;
+
+ unsigned Op1r = getReg(Op1);
+ if (AllocaInst *AI = dyn_castFixedAlloca(LI->getOperand(0))) {
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ addFrameReference(BuildMI(BB, Opcode, 5, ResultReg).addReg(Op1r), FI);
+ } else {
+ X86AddressMode AM;
+ getAddressingMode(LI->getOperand(0), AM);
+ addFullAddress(BuildMI(BB, Opcode, 5, ResultReg).addReg(Op1r), AM);
+ }
+ return;
+ }
+ }
+
MachineBasicBlock::iterator IP = BB->end();
- emitDivRemOperation(BB, IP, Op0Reg, Op1Reg, I.getOpcode() == Instruction::Div,
- I.getType(), ResultReg);
+ emitDivRemOperation(BB, IP, Op0, Op1,
+ I.getOpcode() == Instruction::Div, ResultReg);
}
-void ISel::emitDivRemOperation(MachineBasicBlock *BB,
- MachineBasicBlock::iterator IP,
- unsigned Op0Reg, unsigned Op1Reg, bool isDiv,
- const Type *Ty, unsigned ResultReg) {
+void X86ISel::emitDivRemOperation(MachineBasicBlock *BB,
+ MachineBasicBlock::iterator IP,
+ Value *Op0, Value *Op1, bool isDiv,
+ unsigned ResultReg) {
+ const Type *Ty = Op0->getType();
unsigned Class = getClass(Ty);
switch (Class) {
case cFP: // Floating point divide
if (isDiv) {
- BuildMI(*BB, IP, X86::FpDIV, 2, ResultReg).addReg(Op0Reg).addReg(Op1Reg);
+ emitBinaryFPOperation(BB, IP, Op0, Op1, 3, ResultReg);
+ return;
} else { // Floating point remainder...
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ unsigned Op1Reg = getReg(Op1, BB, IP);
MachineInstr *TheCall =
BuildMI(X86::CALLpcrel32, 1).addExternalSymbol("fmod", true);
std::vector<ValueRecord> Args;
case cLong: {
static const char *FnName[] =
{ "__moddi3", "__divdi3", "__umoddi3", "__udivdi3" };
-
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ unsigned Op1Reg = getReg(Op1, BB, IP);
unsigned NameIdx = Ty->isUnsigned()*2 + isDiv;
MachineInstr *TheCall =
BuildMI(X86::CALLpcrel32, 1).addExternalSymbol(FnName[NameIdx], true);
default: assert(0 && "Unknown class!");
}
- static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX };
static const unsigned MovOpcode[]={ X86::MOV8rr, X86::MOV16rr, X86::MOV32rr };
- static const unsigned SarOpcode[]={ X86::SAR8ri, X86::SAR16ri, X86::SAR32ri };
+ static const unsigned NEGOpcode[]={ X86::NEG8r, X86::NEG16r, X86::NEG32r };
+ static const unsigned SAROpcode[]={ X86::SAR8ri, X86::SAR16ri, X86::SAR32ri };
+ static const unsigned SHROpcode[]={ X86::SHR8ri, X86::SHR16ri, X86::SHR32ri };
+ static const unsigned ADDOpcode[]={ X86::ADD8rr, X86::ADD16rr, X86::ADD32rr };
+
+ // Special case signed division by power of 2.
+ if (ConstantSInt *CI = dyn_cast<ConstantSInt>(Op1))
+ if (isDiv) {
+ assert(Class != cLong && "This doesn't handle 64-bit divides!");
+ int V = CI->getValue();
+
+ if (V == 1) { // X /s 1 => X
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ BuildMI(*BB, IP, MovOpcode[Class], 1, ResultReg).addReg(Op0Reg);
+ return;
+ }
+
+ if (V == -1) { // X /s -1 => -X
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ BuildMI(*BB, IP, NEGOpcode[Class], 1, ResultReg).addReg(Op0Reg);
+ return;
+ }
+
+ if (V == 2 || V == -2) { // X /s 2
+ static const unsigned CMPOpcode[] = {
+ X86::CMP8ri, X86::CMP16ri, X86::CMP32ri
+ };
+ static const unsigned SBBOpcode[] = {
+ X86::SBB8ri, X86::SBB16ri, X86::SBB32ri
+ };
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ unsigned SignBit = 1 << (CI->getType()->getPrimitiveSize()*8-1);
+ BuildMI(*BB, IP, CMPOpcode[Class], 2).addReg(Op0Reg).addImm(SignBit);
+
+ unsigned TmpReg = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, SBBOpcode[Class], 2, TmpReg).addReg(Op0Reg).addImm(-1);
+
+ unsigned TmpReg2 = V == 2 ? ResultReg : makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, SAROpcode[Class], 2, TmpReg2).addReg(TmpReg).addImm(1);
+ if (V == -2) {
+ BuildMI(*BB, IP, NEGOpcode[Class], 1, ResultReg).addReg(TmpReg2);
+ }
+ return;
+ }
+
+ bool isNeg = false;
+ if (V < 0) { // Not a positive power of 2?
+ V = -V;
+ isNeg = true; // Maybe it's a negative power of 2.
+ }
+ if (unsigned Log = ExactLog2(V)) {
+ --Log;
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ unsigned TmpReg = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, SAROpcode[Class], 2, TmpReg)
+ .addReg(Op0Reg).addImm(Log-1);
+ unsigned TmpReg2 = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, SHROpcode[Class], 2, TmpReg2)
+ .addReg(TmpReg).addImm(32-Log);
+ unsigned TmpReg3 = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, ADDOpcode[Class], 2, TmpReg3)
+ .addReg(Op0Reg).addReg(TmpReg2);
+
+ unsigned TmpReg4 = isNeg ? makeAnotherReg(Op0->getType()) : ResultReg;
+ BuildMI(*BB, IP, SAROpcode[Class], 2, TmpReg4)
+ .addReg(TmpReg3).addImm(Log);
+ if (isNeg)
+ BuildMI(*BB, IP, NEGOpcode[Class], 1, ResultReg).addReg(TmpReg4);
+ return;
+ }
+ } else { // X % C
+ assert(Class != cLong && "This doesn't handle 64-bit remainder!");
+ int V = CI->getValue();
+
+ if (V == 2 || V == -2) { // X % 2, X % -2
+ static const unsigned SExtOpcode[] = { X86::CBW, X86::CWD, X86::CDQ };
+ static const unsigned BaseReg[] = { X86::AL , X86::AX , X86::EAX };
+ static const unsigned SExtReg[] = { X86::AH , X86::DX , X86::EDX };
+ static const unsigned ANDOpcode[] = {
+ X86::AND8ri, X86::AND16ri, X86::AND32ri
+ };
+ static const unsigned XOROpcode[] = {
+ X86::XOR8rr, X86::XOR16rr, X86::XOR32rr
+ };
+ static const unsigned SUBOpcode[] = {
+ X86::SUB8rr, X86::SUB16rr, X86::SUB32rr
+ };
+
+ // Sign extend result into reg of -1 or 0.
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ BuildMI(*BB, IP, MovOpcode[Class], 1, BaseReg[Class]).addReg(Op0Reg);
+ BuildMI(*BB, IP, SExtOpcode[Class], 0);
+ unsigned TmpReg0 = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, MovOpcode[Class], 1, TmpReg0).addReg(SExtReg[Class]);
+
+ unsigned TmpReg1 = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, ANDOpcode[Class], 2, TmpReg1).addReg(Op0Reg).addImm(1);
+
+ unsigned TmpReg2 = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, XOROpcode[Class], 2,
+ TmpReg2).addReg(TmpReg1).addReg(TmpReg0);
+ BuildMI(*BB, IP, SUBOpcode[Class], 2,
+ ResultReg).addReg(TmpReg2).addReg(TmpReg0);
+ return;
+ }
+ }
+
+ static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX };
static const unsigned ClrOpcode[]={ X86::MOV8ri, X86::MOV16ri, X86::MOV32ri };
static const unsigned ExtRegs[] ={ X86::AH , X86::DX , X86::EDX };
{ X86::IDIV8r, X86::IDIV16r, X86::IDIV32r, 0 }, // Signed division
};
- bool isSigned = Ty->isSigned();
unsigned Reg = Regs[Class];
unsigned ExtReg = ExtRegs[Class];
// Put the first operand into one of the A registers...
+ unsigned Op0Reg = getReg(Op0, BB, IP);
+ unsigned Op1Reg = getReg(Op1, BB, IP);
BuildMI(*BB, IP, MovOpcode[Class], 1, Reg).addReg(Op0Reg);
- if (isSigned) {
+ if (Ty->isSigned()) {
// Emit a sign extension instruction...
- unsigned ShiftResult = makeAnotherReg(Ty);
- BuildMI(*BB, IP, SarOpcode[Class], 2,ShiftResult).addReg(Op0Reg).addImm(31);
+ unsigned ShiftResult = makeAnotherReg(Op0->getType());
+ BuildMI(*BB, IP, SAROpcode[Class], 2,ShiftResult).addReg(Op0Reg).addImm(31);
BuildMI(*BB, IP, MovOpcode[Class], 1, ExtReg).addReg(ShiftResult);
+
+ // Emit the appropriate divide or remainder instruction...
+ BuildMI(*BB, IP, DivOpcode[1][Class], 1).addReg(Op1Reg);
} else {
// If unsigned, emit a zeroing instruction... (reg = 0)
BuildMI(*BB, IP, ClrOpcode[Class], 2, ExtReg).addImm(0);
- }
- // Emit the appropriate divide or remainder instruction...
- BuildMI(*BB, IP, DivOpcode[isSigned][Class], 1).addReg(Op1Reg);
+ // Emit the appropriate divide or remainder instruction...
+ BuildMI(*BB, IP, DivOpcode[0][Class], 1).addReg(Op1Reg);
+ }
// Figure out which register we want to pick the result out of...
unsigned DestReg = isDiv ? Reg : ExtReg;
/// shift values equal to 1. Even the general case is sort of special,
/// because the shift amount has to be in CL, not just any old register.
///
-void ISel::visitShiftInst(ShiftInst &I) {
+void X86ISel::visitShiftInst(ShiftInst &I) {
MachineBasicBlock::iterator IP = BB->end ();
emitShiftOperation (BB, IP, I.getOperand (0), I.getOperand (1),
I.getOpcode () == Instruction::Shl, I.getType (),
getReg (I));
}
+/// Emit code for a 'SHLD DestReg, Op0, Op1, Amt' operation, where Amt is a
+/// constant.
+void X86ISel::doSHLDConst(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ unsigned DestReg, unsigned Op0Reg, unsigned Op1Reg,
+ unsigned Amt) {
+ // SHLD is a very inefficient operation on every processor, try to do
+ // somethign simpler for common values of 'Amt'.
+ if (Amt == 0) {
+ BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg).addReg(Op0Reg);
+ } else if (Amt == 1) {
+ unsigned Tmp = makeAnotherReg(Type::UIntTy);
+ BuildMI(*MBB, IP, X86::ADD32rr, 2, Tmp).addReg(Op1Reg).addReg(Op1Reg);
+ BuildMI(*MBB, IP, X86::ADC32rr, 2, DestReg).addReg(Op0Reg).addReg(Op0Reg);
+ } else if (Amt == 2 || Amt == 3) {
+ // On the P4 and Athlon it is cheaper to replace shld ..., 2|3 with a
+ // shift/lea pair. NOTE: This should not be done on the P6 family!
+ unsigned Tmp = makeAnotherReg(Type::UIntTy);
+ BuildMI(*MBB, IP, X86::SHR32ri, 2, Tmp).addReg(Op1Reg).addImm(32-Amt);
+ X86AddressMode AM;
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = Tmp;
+ AM.Scale = 1 << Amt;
+ AM.IndexReg = Op0Reg;
+ AM.Disp = 0;
+ addFullAddress(BuildMI(*MBB, IP, X86::LEA32r, 4, DestReg), AM);
+ } else {
+ // NOTE: It is always cheaper on the P4 to emit SHLD as two shifts and an OR
+ // than it is to emit a real SHLD.
+
+ BuildMI(*MBB, IP, X86::SHLD32rri8, 3,
+ DestReg).addReg(Op0Reg).addReg(Op1Reg).addImm(Amt);
+ }
+}
+
/// emitShiftOperation - Common code shared between visitShiftInst and
/// constant expression support.
-void ISel::emitShiftOperation(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP,
- Value *Op, Value *ShiftAmount, bool isLeftShift,
- const Type *ResultTy, unsigned DestReg) {
+void X86ISel::emitShiftOperation(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ Value *Op, Value *ShiftAmount,
+ bool isLeftShift, const Type *ResultTy,
+ unsigned DestReg) {
unsigned SrcReg = getReg (Op, MBB, IP);
bool isSigned = ResultTy->isSigned ();
unsigned Class = getClass (ResultTy);
-
- static const unsigned ConstantOperand[][4] = {
- { X86::SHR8ri, X86::SHR16ri, X86::SHR32ri, X86::SHRD32rri8 }, // SHR
- { X86::SAR8ri, X86::SAR16ri, X86::SAR32ri, X86::SHRD32rri8 }, // SAR
- { X86::SHL8ri, X86::SHL16ri, X86::SHL32ri, X86::SHLD32rri8 }, // SHL
- { X86::SHL8ri, X86::SHL16ri, X86::SHL32ri, X86::SHLD32rri8 }, // SAL = SHL
+
+ static const unsigned ConstantOperand[][3] = {
+ { X86::SHR8ri, X86::SHR16ri, X86::SHR32ri }, // SHR
+ { X86::SAR8ri, X86::SAR16ri, X86::SAR32ri }, // SAR
+ { X86::SHL8ri, X86::SHL16ri, X86::SHL32ri }, // SHL
+ { X86::SHL8ri, X86::SHL16ri, X86::SHL32ri }, // SAL = SHL
};
- static const unsigned NonConstantOperand[][4] = {
+ static const unsigned NonConstantOperand[][3] = {
{ X86::SHR8rCL, X86::SHR16rCL, X86::SHR32rCL }, // SHR
{ X86::SAR8rCL, X86::SAR16rCL, X86::SAR32rCL }, // SAR
{ X86::SHL8rCL, X86::SHL16rCL, X86::SHL32rCL }, // SHL
{ X86::SHL8rCL, X86::SHL16rCL, X86::SHL32rCL }, // SAL = SHL
};
- // Longs, as usual, are handled specially...
+ // Longs, as usual, are handled specially.
if (Class == cLong) {
- // If we have a constant shift, we can generate much more efficient code
- // than otherwise...
- //
if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(ShiftAmount)) {
unsigned Amount = CUI->getValue();
- if (Amount < 32) {
+ if (Amount == 1 && isLeftShift) { // X << 1 == X+X
+ BuildMI(*MBB, IP, X86::ADD32rr, 2,
+ DestReg).addReg(SrcReg).addReg(SrcReg);
+ BuildMI(*MBB, IP, X86::ADC32rr, 2,
+ DestReg+1).addReg(SrcReg+1).addReg(SrcReg+1);
+ } else if (Amount < 32) {
const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned];
if (isLeftShift) {
- BuildMI(*MBB, IP, Opc[3], 3,
- DestReg+1).addReg(SrcReg+1).addReg(SrcReg).addImm(Amount);
+ doSHLDConst(MBB, IP, DestReg+1, SrcReg+1, SrcReg, Amount);
BuildMI(*MBB, IP, Opc[2], 2, DestReg).addReg(SrcReg).addImm(Amount);
} else {
- BuildMI(*MBB, IP, Opc[3], 3,
- DestReg).addReg(SrcReg ).addReg(SrcReg+1).addImm(Amount);
+ BuildMI(*MBB, IP, X86::SHRD32rri8, 3,
+ DestReg).addReg(SrcReg ).addReg(SrcReg+1).addImm(Amount);
BuildMI(*MBB, IP, Opc[2],2,DestReg+1).addReg(SrcReg+1).addImm(Amount);
}
- } else { // Shifting more than 32 bits
- Amount -= 32;
+ } else if (Amount == 32) {
if (isLeftShift) {
- if (Amount != 0) {
- BuildMI(*MBB, IP, X86::SHL32ri, 2,
- DestReg + 1).addReg(SrcReg).addImm(Amount);
- } else {
- BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg+1).addReg(SrcReg);
- }
+ BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg+1).addReg(SrcReg);
BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
} else {
- if (Amount != 0) {
- BuildMI(*MBB, IP, isSigned ? X86::SAR32ri : X86::SHR32ri, 2,
- DestReg).addReg(SrcReg+1).addImm(Amount);
+ BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg).addReg(SrcReg+1);
+ if (!isSigned) {
+ BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg+1).addImm(0);
} else {
- BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg).addReg(SrcReg+1);
+ BuildMI(*MBB, IP, X86::SAR32ri, 2,
+ DestReg+1).addReg(SrcReg).addImm(31);
}
+ }
+ } else { // Shifting more than 32 bits
+ Amount -= 32;
+ if (isLeftShift) {
+ BuildMI(*MBB, IP, X86::SHL32ri, 2,
+ DestReg + 1).addReg(SrcReg).addImm(Amount);
+ BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
+ } else {
+ BuildMI(*MBB, IP, isSigned ? X86::SAR32ri : X86::SHR32ri, 2,
+ DestReg).addReg(SrcReg+1).addImm(Amount);
BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg+1).addImm(0);
}
}
} else {
unsigned TmpReg = makeAnotherReg(Type::IntTy);
-
if (!isLeftShift && isSigned) {
// If this is a SHR of a Long, then we need to do funny sign extension
// stuff. TmpReg gets the value to use as the high-part if we are
// The shift amount is constant, guaranteed to be a ubyte. Get its value.
assert(CUI->getType() == Type::UByteTy && "Shift amount not a ubyte?");
- const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned];
- BuildMI(*MBB, IP, Opc[Class], 2,
- DestReg).addReg(SrcReg).addImm(CUI->getValue());
+ if (CUI->getValue() == 1 && isLeftShift) { // X << 1 -> X+X
+ static const int AddOpC[] = { X86::ADD8rr, X86::ADD16rr, X86::ADD32rr };
+ BuildMI(*MBB, IP, AddOpC[Class], 2,DestReg).addReg(SrcReg).addReg(SrcReg);
+ } else {
+ const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned];
+ BuildMI(*MBB, IP, Opc[Class], 2,
+ DestReg).addReg(SrcReg).addImm(CUI->getValue());
+ }
} else { // The shift amount is non-constant.
unsigned ShiftAmountReg = getReg (ShiftAmount, MBB, IP);
BuildMI(*MBB, IP, X86::MOV8rr, 1, X86::CL).addReg(ShiftAmountReg);
}
-void ISel::getAddressingMode(Value *Addr, unsigned &BaseReg, unsigned &Scale,
- unsigned &IndexReg, unsigned &Disp) {
- BaseReg = 0; Scale = 1; IndexReg = 0; Disp = 0;
- if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr)) {
- if (isGEPFoldable(BB, GEP->getOperand(0), GEP->op_begin()+1, GEP->op_end(),
- BaseReg, Scale, IndexReg, Disp))
- return;
- } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
- if (CE->getOpcode() == Instruction::GetElementPtr)
- if (isGEPFoldable(BB, CE->getOperand(0), CE->op_begin()+1, CE->op_end(),
- BaseReg, Scale, IndexReg, Disp))
- return;
- }
-
- // If it's not foldable, reset addr mode.
- BaseReg = getReg(Addr);
- Scale = 1; IndexReg = 0; Disp = 0;
-}
-
-
/// visitLoadInst - Implement LLVM load instructions in terms of the x86 'mov'
/// instruction. The load and store instructions are the only place where we
/// need to worry about the memory layout of the target machine.
///
-void ISel::visitLoadInst(LoadInst &I) {
+void X86ISel::visitLoadInst(LoadInst &I) {
// Check to see if this load instruction is going to be folded into a binary
// instruction, like add. If so, we don't want to emit it. Wouldn't a real
// pattern matching instruction selector be nice?
- if (I.hasOneUse() && getClassB(I.getType()) < cFP) {
+ unsigned Class = getClassB(I.getType());
+ if (I.hasOneUse()) {
Instruction *User = cast<Instruction>(I.use_back());
switch (User->getOpcode()) {
- default: User = 0; break;
+ case Instruction::Cast:
+ // If this is a cast from a signed-integer type to a floating point type,
+ // fold the cast here.
+ if (getClassB(User->getType()) == cFP &&
+ (I.getType() == Type::ShortTy || I.getType() == Type::IntTy ||
+ I.getType() == Type::LongTy)) {
+ unsigned DestReg = getReg(User);
+ static const unsigned Opcode[] = {
+ 0/*BYTE*/, X86::FILD16m, X86::FILD32m, 0/*FP*/, X86::FILD64m
+ };
+
+ if (AllocaInst *AI = dyn_castFixedAlloca(I.getOperand(0))) {
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ addFrameReference(BuildMI(BB, Opcode[Class], 4, DestReg), FI);
+ } else {
+ X86AddressMode AM;
+ getAddressingMode(I.getOperand(0), AM);
+ addFullAddress(BuildMI(BB, Opcode[Class], 4, DestReg), AM);
+ }
+ return;
+ } else {
+ User = 0;
+ }
+ break;
+
case Instruction::Add:
case Instruction::Sub:
case Instruction::And:
case Instruction::Or:
case Instruction::Xor:
+ if (Class == cLong) User = 0;
break;
+ case Instruction::Mul:
+ case Instruction::Div:
+ if (Class != cFP) User = 0;
+ break; // Folding only implemented for floating point.
+ default: User = 0; break;
}
if (User) {
// Okay, we found a user. If the load is the first operand and there is
// no second operand load, reverse the operand ordering. Note that this
// can fail for a subtract (ie, no change will be made).
+ bool Swapped = false;
if (!isa<LoadInst>(User->getOperand(1)))
- cast<BinaryOperator>(User)->swapOperands();
+ Swapped = !cast<BinaryOperator>(User)->swapOperands();
// Okay, now that everything is set up, if this load is used by the second
// operand, and if there are no instructions that invalidate the load
if (User->getOperand(1) == &I &&
isSafeToFoldLoadIntoInstruction(I, *User))
return; // Eliminate the load!
- }
- }
- unsigned DestReg = getReg(I);
- unsigned BaseReg = 0, Scale = 1, IndexReg = 0, Disp = 0;
- getAddressingMode(I.getOperand(0), BaseReg, Scale, IndexReg, Disp);
+ // If this is a floating point sub or div, we won't be able to swap the
+ // operands, but we will still be able to eliminate the load.
+ if (Class == cFP && User->getOperand(0) == &I &&
+ !isa<LoadInst>(User->getOperand(1)) &&
+ (User->getOpcode() == Instruction::Sub ||
+ User->getOpcode() == Instruction::Div) &&
+ isSafeToFoldLoadIntoInstruction(I, *User))
+ return; // Eliminate the load!
- unsigned Class = getClassB(I.getType());
- if (Class == cLong) {
- addFullAddress(BuildMI(BB, X86::MOV32rm, 4, DestReg),
- BaseReg, Scale, IndexReg, Disp);
- addFullAddress(BuildMI(BB, X86::MOV32rm, 4, DestReg+1),
- BaseReg, Scale, IndexReg, Disp+4);
- return;
+ // If we swapped the operands to the instruction, but couldn't fold the
+ // load anyway, swap them back. We don't want to break add X, int
+ // folding.
+ if (Swapped) cast<BinaryOperator>(User)->swapOperands();
+ }
}
static const unsigned Opcodes[] = {
- X86::MOV8rm, X86::MOV16rm, X86::MOV32rm, X86::FLD32m
+ X86::MOV8rm, X86::MOV16rm, X86::MOV32rm, X86::FLD32m, X86::MOV32rm
};
unsigned Opcode = Opcodes[Class];
if (I.getType() == Type::DoubleTy) Opcode = X86::FLD64m;
- addFullAddress(BuildMI(BB, Opcode, 4, DestReg),
- BaseReg, Scale, IndexReg, Disp);
+
+ unsigned DestReg = getReg(I);
+
+ if (AllocaInst *AI = dyn_castFixedAlloca(I.getOperand(0))) {
+ unsigned FI = getFixedSizedAllocaFI(AI);
+ if (Class == cLong) {
+ addFrameReference(BuildMI(BB, X86::MOV32rm, 4, DestReg), FI);
+ addFrameReference(BuildMI(BB, X86::MOV32rm, 4, DestReg+1), FI, 4);
+ } else {
+ addFrameReference(BuildMI(BB, Opcode, 4, DestReg), FI);
+ }
+ } else {
+ X86AddressMode AM;
+ getAddressingMode(I.getOperand(0), AM);
+
+ if (Class == cLong) {
+ addFullAddress(BuildMI(BB, X86::MOV32rm, 4, DestReg), AM);
+ AM.Disp += 4;
+ addFullAddress(BuildMI(BB, X86::MOV32rm, 4, DestReg+1), AM);
+ } else {
+ addFullAddress(BuildMI(BB, Opcode, 4, DestReg), AM);
+ }
+ }
}
/// visitStoreInst - Implement LLVM store instructions in terms of the x86 'mov'
/// instruction.
///
-void ISel::visitStoreInst(StoreInst &I) {
- unsigned BaseReg, Scale, IndexReg, Disp;
- getAddressingMode(I.getOperand(1), BaseReg, Scale, IndexReg, Disp);
+void X86ISel::visitStoreInst(StoreInst &I) {
+ X86AddressMode AM;
+ getAddressingMode(I.getOperand(1), AM);
const Type *ValTy = I.getOperand(0)->getType();
unsigned Class = getClassB(ValTy);
if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(0))) {
uint64_t Val = CI->getRawValue();
if (Class == cLong) {
- addFullAddress(BuildMI(BB, X86::MOV32mi, 5),
- BaseReg, Scale, IndexReg, Disp).addImm(Val & ~0U);
- addFullAddress(BuildMI(BB, X86::MOV32mi, 5),
- BaseReg, Scale, IndexReg, Disp+4).addImm(Val>>32);
+ addFullAddress(BuildMI(BB, X86::MOV32mi, 5), AM).addImm(Val & ~0U);
+ AM.Disp += 4;
+ addFullAddress(BuildMI(BB, X86::MOV32mi, 5), AM).addImm(Val>>32);
} else {
static const unsigned Opcodes[] = {
X86::MOV8mi, X86::MOV16mi, X86::MOV32mi
};
unsigned Opcode = Opcodes[Class];
- addFullAddress(BuildMI(BB, Opcode, 5),
- BaseReg, Scale, IndexReg, Disp).addImm(Val);
+ addFullAddress(BuildMI(BB, Opcode, 5), AM).addImm(Val);
}
+ } else if (isa<ConstantPointerNull>(I.getOperand(0))) {
+ addFullAddress(BuildMI(BB, X86::MOV32mi, 5), AM).addImm(0);
} else if (ConstantBool *CB = dyn_cast<ConstantBool>(I.getOperand(0))) {
- addFullAddress(BuildMI(BB, X86::MOV8mi, 5),
- BaseReg, Scale, IndexReg, Disp).addImm(CB->getValue());
- } else {
- if (Class == cLong) {
- unsigned ValReg = getReg(I.getOperand(0));
- addFullAddress(BuildMI(BB, X86::MOV32mr, 5),
- BaseReg, Scale, IndexReg, Disp).addReg(ValReg);
- addFullAddress(BuildMI(BB, X86::MOV32mr, 5),
- BaseReg, Scale, IndexReg, Disp+4).addReg(ValReg+1);
+ addFullAddress(BuildMI(BB, X86::MOV8mi, 5), AM).addImm(CB->getValue());
+ } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(I.getOperand(0))) {
+ // Store constant FP values with integer instructions to avoid having to
+ // load the constants from the constant pool then do a store.
+ if (CFP->getType() == Type::FloatTy) {
+ union {
+ unsigned I;
+ float F;
+ } V;
+ V.F = CFP->getValue();
+ addFullAddress(BuildMI(BB, X86::MOV32mi, 5), AM).addImm(V.I);
} else {
- unsigned ValReg = getReg(I.getOperand(0));
- static const unsigned Opcodes[] = {
- X86::MOV8mr, X86::MOV16mr, X86::MOV32mr, X86::FST32m
- };
- unsigned Opcode = Opcodes[Class];
- if (ValTy == Type::DoubleTy) Opcode = X86::FST64m;
- addFullAddress(BuildMI(BB, Opcode, 1+4),
- BaseReg, Scale, IndexReg, Disp).addReg(ValReg);
+ union {
+ uint64_t I;
+ double F;
+ } V;
+ V.F = CFP->getValue();
+ addFullAddress(BuildMI(BB, X86::MOV32mi, 5), AM).addImm((unsigned)V.I);
+ AM.Disp += 4;
+ addFullAddress(BuildMI(BB, X86::MOV32mi, 5), AM).addImm(
+ unsigned(V.I >> 32));
}
+
+ } else if (Class == cLong) {
+ unsigned ValReg = getReg(I.getOperand(0));
+ addFullAddress(BuildMI(BB, X86::MOV32mr, 5), AM).addReg(ValReg);
+ AM.Disp += 4;
+ addFullAddress(BuildMI(BB, X86::MOV32mr, 5), AM).addReg(ValReg+1);
+ } else {
+ // FIXME: stop emitting these two instructions:
+ // movl $global,%eax
+ // movl %eax,(%ebx)
+ // when one instruction will suffice. That includes when the global
+ // has an offset applied to it.
+ unsigned ValReg = getReg(I.getOperand(0));
+ static const unsigned Opcodes[] = {
+ X86::MOV8mr, X86::MOV16mr, X86::MOV32mr, X86::FST32m
+ };
+ unsigned Opcode = Opcodes[Class];
+ if (ValTy == Type::DoubleTy) Opcode = X86::FST64m;
+
+ addFullAddress(BuildMI(BB, Opcode, 1+4), AM).addReg(ValReg);
}
}
/// visitCastInst - Here we have various kinds of copying with or without sign
/// extension going on.
///
-void ISel::visitCastInst(CastInst &CI) {
+void X86ISel::visitCastInst(CastInst &CI) {
Value *Op = CI.getOperand(0);
+
+ unsigned SrcClass = getClassB(Op->getType());
+ unsigned DestClass = getClassB(CI.getType());
+ // Noop casts are not emitted: getReg will return the source operand as the
+ // register to use for any uses of the noop cast.
+ if (DestClass == SrcClass) {
+ // The only detail in this plan is that casts from double -> float are
+ // truncating operations that we have to codegen through memory (despite
+ // the fact that the source/dest registers are the same class).
+ if (CI.getType() != Type::FloatTy || Op->getType() != Type::DoubleTy)
+ return;
+ }
+
// If this is a cast from a 32-bit integer to a Long type, and the only uses
// of the case are GEP instructions, then the cast does not need to be
// generated explicitly, it will be folded into the GEP.
- if (CI.getType() == Type::LongTy &&
- (Op->getType() == Type::IntTy || Op->getType() == Type::UIntTy)) {
+ if (DestClass == cLong && SrcClass == cInt) {
bool AllUsesAreGEPs = true;
for (Value::use_iterator I = CI.use_begin(), E = CI.use_end(); I != E; ++I)
if (!isa<GetElementPtrInst>(*I)) {
if (AllUsesAreGEPs) return;
}
+ // If this cast converts a load from a short,int, or long integer to a FP
+ // value, we will have folded this cast away.
+ if (DestClass == cFP && isa<LoadInst>(Op) && Op->hasOneUse() &&
+ (Op->getType() == Type::ShortTy || Op->getType() == Type::IntTy ||
+ Op->getType() == Type::LongTy))
+ return;
+
+
unsigned DestReg = getReg(CI);
MachineBasicBlock::iterator MI = BB->end();
emitCastOperation(BB, MI, Op, CI.getType(), DestReg);
/// emitCastOperation - Common code shared between visitCastInst and constant
/// expression cast support.
///
-void ISel::emitCastOperation(MachineBasicBlock *BB,
- MachineBasicBlock::iterator IP,
- Value *Src, const Type *DestTy,
- unsigned DestReg) {
- unsigned SrcReg = getReg(Src, BB, IP);
+void X86ISel::emitCastOperation(MachineBasicBlock *BB,
+ MachineBasicBlock::iterator IP,
+ Value *Src, const Type *DestTy,
+ unsigned DestReg) {
const Type *SrcTy = Src->getType();
unsigned SrcClass = getClassB(SrcTy);
unsigned DestClass = getClassB(DestTy);
+ unsigned SrcReg = getReg(Src, BB, IP);
// Implement casts to bool by using compare on the operand followed by set if
// not zero on the result.
{ X86::MOVZX16rr8, X86::MOVZX32rr8, X86::MOVZX32rr16, X86::MOV32rr } // u
};
- bool isUnsigned = SrcTy->isUnsigned();
+ bool isUnsigned = SrcTy->isUnsigned() || SrcTy == Type::BoolTy;
BuildMI(*BB, IP, Opc[isUnsigned][SrcClass + DestClass - 1], 1,
DestReg).addReg(SrcReg);
const Type *PromoteType = 0;
unsigned PromoteOpcode = 0;
unsigned RealDestReg = DestReg;
- switch (SrcTy->getPrimitiveID()) {
+ switch (SrcTy->getTypeID()) {
case Type::BoolTyID:
case Type::SByteTyID:
// We don't have the facilities for directly loading byte sized data from
PromoteType = Type::IntTy;
PromoteOpcode = X86::MOVZX32rr16;
break;
- case Type::UIntTyID: {
- // Make a 64 bit temporary... and zero out the top of it...
- unsigned TmpReg = makeAnotherReg(Type::LongTy);
- BuildMI(*BB, IP, X86::MOV32rr, 1, TmpReg).addReg(SrcReg);
- BuildMI(*BB, IP, X86::MOV32ri, 1, TmpReg+1).addImm(0);
- SrcTy = Type::LongTy;
- SrcClass = cLong;
- SrcReg = TmpReg;
- break;
- }
case Type::ULongTyID:
+ case Type::UIntTyID:
// Don't fild into the read destination.
DestReg = makeAnotherReg(Type::DoubleTy);
break;
{ 0/*byte*/, X86::FILD16m, X86::FILD32m, 0/*FP*/, X86::FILD64m };
addFrameReference(BuildMI(*BB, IP, Op2[SrcClass], 5, DestReg), FrameIdx);
- // We need special handling for unsigned 64-bit integer sources. If the
- // input number has the "sign bit" set, then we loaded it incorrectly as a
- // negative 64-bit number. In this case, add an offset value.
- if (SrcTy == Type::ULongTy) {
+ if (SrcTy == Type::UIntTy) {
+ // If this is a cast from uint -> double, we need to be careful about if
+ // the "sign" bit is set. If so, we don't want to make a negative number,
+ // we want to make a positive number. Emit code to add an offset if the
+ // sign bit is set.
+
+ // Compute whether the sign bit is set by shifting the reg right 31 bits.
+ unsigned IsNeg = makeAnotherReg(Type::IntTy);
+ BuildMI(BB, X86::SHR32ri, 2, IsNeg).addReg(SrcReg).addImm(31);
+
+ // Create a CP value that has the offset in one word and 0 in the other.
+ static ConstantInt *TheOffset = ConstantUInt::get(Type::ULongTy,
+ 0x4f80000000000000ULL);
+ unsigned CPI = F->getConstantPool()->getConstantPoolIndex(TheOffset);
+ BuildMI(BB, X86::FADD32m, 5, RealDestReg).addReg(DestReg)
+ .addConstantPoolIndex(CPI).addZImm(4).addReg(IsNeg).addSImm(0);
+
+ } else if (SrcTy == Type::ULongTy) {
+ // We need special handling for unsigned 64-bit integer sources. If the
+ // input number has the "sign bit" set, then we loaded it incorrectly as a
+ // negative 64-bit number. In this case, add an offset value.
+
// Emit a test instruction to see if the dynamic input value was signed.
BuildMI(*BB, IP, X86::TEST32rr, 2).addReg(SrcReg+1).addReg(SrcReg+1);
/// visitVANextInst - Implement the va_next instruction...
///
-void ISel::visitVANextInst(VANextInst &I) {
+void X86ISel::visitVANextInst(VANextInst &I) {
unsigned VAList = getReg(I.getOperand(0));
unsigned DestReg = getReg(I);
unsigned Size;
- switch (I.getArgType()->getPrimitiveID()) {
+ switch (I.getArgType()->getTypeID()) {
default:
std::cerr << I;
assert(0 && "Error: bad type for va_next instruction!");
BuildMI(BB, X86::ADD32ri, 2, DestReg).addReg(VAList).addImm(Size);
}
-void ISel::visitVAArgInst(VAArgInst &I) {
+void X86ISel::visitVAArgInst(VAArgInst &I) {
unsigned VAList = getReg(I.getOperand(0));
unsigned DestReg = getReg(I);
- switch (I.getType()->getPrimitiveID()) {
+ switch (I.getType()->getTypeID()) {
default:
std::cerr << I;
assert(0 && "Error: bad type for va_next instruction!");
/// visitGetElementPtrInst - instruction-select GEP instructions
///
-void ISel::visitGetElementPtrInst(GetElementPtrInst &I) {
+void X86ISel::visitGetElementPtrInst(GetElementPtrInst &I) {
// If this GEP instruction will be folded into all of its users, we don't need
// to explicitly calculate it!
- unsigned A, B, C, D;
- if (isGEPFoldable(0, I.getOperand(0), I.op_begin()+1, I.op_end(), A,B,C,D)) {
+ X86AddressMode AM;
+ if (isGEPFoldable(0, I.getOperand(0), I.op_begin()+1, I.op_end(), AM)) {
// Check all of the users of the instruction to see if they are loads and
// stores.
bool AllWillFold = true;
///
/// Note that there is one fewer entry in GEPTypes than there is in GEPOps.
///
-void ISel::getGEPIndex(MachineBasicBlock *MBB, MachineBasicBlock::iterator IP,
- std::vector<Value*> &GEPOps,
- std::vector<const Type*> &GEPTypes, unsigned &BaseReg,
- unsigned &Scale, unsigned &IndexReg, unsigned &Disp) {
+void X86ISel::getGEPIndex(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ std::vector<Value*> &GEPOps,
+ std::vector<const Type*> &GEPTypes,
+ X86AddressMode &AM) {
const TargetData &TD = TM.getTargetData();
// Clear out the state we are working with...
- BaseReg = 0; // No base register
- Scale = 1; // Unit scale
- IndexReg = 0; // No index register
- Disp = 0; // No displacement
+ AM.BaseType = X86AddressMode::RegBase;
+ AM.Base.Reg = 0; // No base register
+ AM.Scale = 1; // Unit scale
+ AM.IndexReg = 0; // No index register
+ AM.Disp = 0; // No displacement
// While there are GEP indexes that can be folded into the current address,
// keep processing them.
// structure is in memory. Since the structure index must be constant, we
// can get its value and use it to find the right byte offset from the
// StructLayout class's list of structure member offsets.
- Disp += TD.getStructLayout(StTy)->MemberOffsets[CUI->getValue()];
+ AM.Disp += TD.getStructLayout(StTy)->MemberOffsets[CUI->getValue()];
GEPOps.pop_back(); // Consume a GEP operand
GEPTypes.pop_back();
} else {
// If idx is a constant, fold it into the offset.
unsigned TypeSize = TD.getTypeSize(SqTy->getElementType());
if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(idx)) {
- Disp += TypeSize*CSI->getValue();
+ AM.Disp += TypeSize*CSI->getValue();
} else if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(idx)) {
- Disp += TypeSize*CUI->getValue();
+ AM.Disp += TypeSize*CUI->getValue();
} else {
// If the index reg is already taken, we can't handle this index.
- if (IndexReg) return;
+ if (AM.IndexReg) return;
// If this is a size that we can handle, then add the index as
switch (TypeSize) {
case 1: case 2: case 4: case 8:
// These are all acceptable scales on X86.
- Scale = TypeSize;
+ AM.Scale = TypeSize;
break;
default:
// Otherwise, we can't handle this scale
CI->getOperand(0)->getType() == Type::UIntTy)
idx = CI->getOperand(0);
- IndexReg = MBB ? getReg(idx, MBB, IP) : 1;
+ AM.IndexReg = MBB ? getReg(idx, MBB, IP) : 1;
}
GEPOps.pop_back(); // Consume a GEP operand
}
}
- // GEPTypes is empty, which means we have a single operand left. See if we
- // can set it as the base register.
+ // GEPTypes is empty, which means we have a single operand left. Set it as
+ // the base register.
//
- // FIXME: When addressing modes are more powerful/correct, we could load
- // global addresses directly as 32-bit immediates.
- assert(BaseReg == 0);
- BaseReg = MBB ? getReg(GEPOps[0], MBB, IP) : 1;
+ assert(AM.Base.Reg == 0);
+
+ if (AllocaInst *AI = dyn_castFixedAlloca(GEPOps.back())) {
+ AM.BaseType = X86AddressMode::FrameIndexBase;
+ AM.Base.FrameIndex = getFixedSizedAllocaFI(AI);
+ GEPOps.pop_back();
+ return;
+ }
+
+ if (GlobalValue *GV = dyn_cast<GlobalValue>(GEPOps.back())) {
+ AM.GV = GV;
+ GEPOps.pop_back();
+ return;
+ }
+
+ AM.Base.Reg = MBB ? getReg(GEPOps[0], MBB, IP) : 1;
GEPOps.pop_back(); // Consume the last GEP operand
}
/// isGEPFoldable - Return true if the specified GEP can be completely
/// folded into the addressing mode of a load/store or lea instruction.
-bool ISel::isGEPFoldable(MachineBasicBlock *MBB,
- Value *Src, User::op_iterator IdxBegin,
- User::op_iterator IdxEnd, unsigned &BaseReg,
- unsigned &Scale, unsigned &IndexReg, unsigned &Disp) {
- if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Src))
- Src = CPR->getValue();
+bool X86ISel::isGEPFoldable(MachineBasicBlock *MBB,
+ Value *Src, User::op_iterator IdxBegin,
+ User::op_iterator IdxEnd, X86AddressMode &AM) {
std::vector<Value*> GEPOps;
GEPOps.resize(IdxEnd-IdxBegin+1);
GEPOps[0] = Src;
std::copy(IdxBegin, IdxEnd, GEPOps.begin()+1);
- std::vector<const Type*> GEPTypes;
- GEPTypes.assign(gep_type_begin(Src->getType(), IdxBegin, IdxEnd),
- gep_type_end(Src->getType(), IdxBegin, IdxEnd));
+ std::vector<const Type*>
+ GEPTypes(gep_type_begin(Src->getType(), IdxBegin, IdxEnd),
+ gep_type_end(Src->getType(), IdxBegin, IdxEnd));
MachineBasicBlock::iterator IP;
if (MBB) IP = MBB->end();
- getGEPIndex(MBB, IP, GEPOps, GEPTypes, BaseReg, Scale, IndexReg, Disp);
+ getGEPIndex(MBB, IP, GEPOps, GEPTypes, AM);
// We can fold it away iff the getGEPIndex call eliminated all operands.
return GEPOps.empty();
}
-void ISel::emitGEPOperation(MachineBasicBlock *MBB,
- MachineBasicBlock::iterator IP,
- Value *Src, User::op_iterator IdxBegin,
- User::op_iterator IdxEnd, unsigned TargetReg) {
+void X86ISel::emitGEPOperation(MachineBasicBlock *MBB,
+ MachineBasicBlock::iterator IP,
+ Value *Src, User::op_iterator IdxBegin,
+ User::op_iterator IdxEnd, unsigned TargetReg) {
const TargetData &TD = TM.getTargetData();
- if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Src))
- Src = CPR->getValue();
+
+ // If this is a getelementptr null, with all constant integer indices, just
+ // replace it with TargetReg = 42.
+ if (isa<ConstantPointerNull>(Src)) {
+ User::op_iterator I = IdxBegin;
+ for (; I != IdxEnd; ++I)
+ if (!isa<ConstantInt>(*I))
+ break;
+ if (I == IdxEnd) { // All constant indices
+ unsigned Offset = TD.getIndexedOffset(Src->getType(),
+ std::vector<Value*>(IdxBegin, IdxEnd));
+ BuildMI(*MBB, IP, X86::MOV32ri, 1, TargetReg).addImm(Offset);
+ return;
+ }
+ }
std::vector<Value*> GEPOps;
GEPOps.resize(IdxEnd-IdxBegin+1);
// Keep emitting instructions until we consume the entire GEP instruction.
while (!GEPOps.empty()) {
unsigned OldSize = GEPOps.size();
- unsigned BaseReg, Scale, IndexReg, Disp;
- getGEPIndex(MBB, IP, GEPOps, GEPTypes, BaseReg, Scale, IndexReg, Disp);
+ X86AddressMode AM;
+ getGEPIndex(MBB, IP, GEPOps, GEPTypes, AM);
if (GEPOps.size() != OldSize) {
// getGEPIndex consumed some of the input. Build an LEA instruction here.
unsigned NextTarget = 0;
if (!GEPOps.empty()) {
- assert(BaseReg == 0 &&
+ assert(AM.Base.Reg == 0 &&
"getGEPIndex should have left the base register open for chaining!");
- NextTarget = BaseReg = makeAnotherReg(Type::UIntTy);
+ NextTarget = AM.Base.Reg = makeAnotherReg(Type::UIntTy);
}
- if (IndexReg == 0 && Disp == 0)
- BuildMI(*MBB, IP, X86::MOV32rr, 1, TargetReg).addReg(BaseReg);
+ if (AM.BaseType == X86AddressMode::RegBase &&
+ AM.IndexReg == 0 && AM.Disp == 0 && !AM.GV)
+ BuildMI(*MBB, IP, X86::MOV32rr, 1, TargetReg).addReg(AM.Base.Reg);
+ else if (AM.BaseType == X86AddressMode::RegBase && AM.Base.Reg == 0 &&
+ AM.IndexReg == 0 && AM.Disp == 0)
+ BuildMI(*MBB, IP, X86::MOV32ri, 1, TargetReg).addGlobalAddress(AM.GV);
else
- addFullAddress(BuildMI(*MBB, IP, X86::LEA32r, 5, TargetReg),
- BaseReg, Scale, IndexReg, Disp);
+ addFullAddress(BuildMI(*MBB, IP, X86::LEA32r, 5, TargetReg), AM);
--IP;
TargetReg = NextTarget;
} else if (GEPTypes.empty()) {
}
}
-
/// visitAllocaInst - If this is a fixed size alloca, allocate space from the
/// frame manager, otherwise do it the hard way.
///
-void ISel::visitAllocaInst(AllocaInst &I) {
+void X86ISel::visitAllocaInst(AllocaInst &I) {
+ // If this is a fixed size alloca in the entry block for the function, we
+ // statically stack allocate the space, so we don't need to do anything here.
+ //
+ if (dyn_castFixedAlloca(&I)) return;
+
// Find the data size of the alloca inst's getAllocatedType.
const Type *Ty = I.getAllocatedType();
unsigned TySize = TM.getTargetData().getTypeSize(Ty);
- // If this is a fixed size alloca in the entry block for the function,
- // statically stack allocate the space.
- //
- if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I.getArraySize())) {
- if (I.getParent() == I.getParent()->getParent()->begin()) {
- TySize *= CUI->getValue(); // Get total allocated size...
- unsigned Alignment = TM.getTargetData().getTypeAlignment(Ty);
-
- // Create a new stack object using the frame manager...
- int FrameIdx = F->getFrameInfo()->CreateStackObject(TySize, Alignment);
- addFrameReference(BuildMI(BB, X86::LEA32r, 5, getReg(I)), FrameIdx);
- return;
- }
- }
-
// Create a register to hold the temporary result of multiplying the type size
// constant by the variable amount.
unsigned TotalSizeReg = makeAnotherReg(Type::UIntTy);
/// visitMallocInst - Malloc instructions are code generated into direct calls
/// to the library malloc.
///
-void ISel::visitMallocInst(MallocInst &I) {
+void X86ISel::visitMallocInst(MallocInst &I) {
unsigned AllocSize = TM.getTargetData().getTypeSize(I.getAllocatedType());
unsigned Arg;
/// visitFreeInst - Free instructions are code gen'd to call the free libc
/// function.
///
-void ISel::visitFreeInst(FreeInst &I) {
+void X86ISel::visitFreeInst(FreeInst &I) {
std::vector<ValueRecord> Args;
Args.push_back(ValueRecord(I.getOperand(0)));
MachineInstr *TheCall = BuildMI(X86::CALLpcrel32,
/// generated code sucks but the implementation is nice and simple.
///
FunctionPass *llvm::createX86SimpleInstructionSelector(TargetMachine &TM) {
- return new ISel(TM);
+ return new X86ISel(TM);
}