1 //===-- InstSelectSimple.cpp - A simple instruction selector for x86 ------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a simple peephole instruction selector for the x86 target
12 //===----------------------------------------------------------------------===//
15 #include "X86InstrBuilder.h"
16 #include "X86InstrInfo.h"
17 #include "llvm/Constants.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/Function.h"
20 #include "llvm/Instructions.h"
21 #include "llvm/IntrinsicLowering.h"
22 #include "llvm/Pass.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFrameInfo.h"
25 #include "llvm/CodeGen/MachineFunction.h"
26 #include "llvm/CodeGen/SSARegMap.h"
27 #include "llvm/Target/MRegisterInfo.h"
28 #include "llvm/Target/TargetMachine.h"
29 #include "llvm/Support/GetElementPtrTypeIterator.h"
30 #include "llvm/Support/InstVisitor.h"
31 #include "llvm/Support/CFG.h"
32 #include "Support/Statistic.h"
37 NumFPKill("x86-codegen", "Number of FP_REG_KILL instructions added");
39 /// TypeClass - Used by the X86 backend to group LLVM types by their basic X86
43 cByte, cShort, cInt, cFP, cLong
47 /// getClass - Turn a primitive type into a "class" number which is based on the
48 /// size of the type, and whether or not it is floating point.
50 static inline TypeClass getClass(const Type *Ty) {
51 switch (Ty->getPrimitiveID()) {
53 case Type::UByteTyID: return cByte; // Byte operands are class #0
55 case Type::UShortTyID: return cShort; // Short operands are class #1
58 case Type::PointerTyID: return cInt; // Int's and pointers are class #2
61 case Type::DoubleTyID: return cFP; // Floating Point is #3
64 case Type::ULongTyID: return cLong; // Longs are class #4
66 assert(0 && "Invalid type to getClass!");
67 return cByte; // not reached
71 // getClassB - Just like getClass, but treat boolean values as bytes.
72 static inline TypeClass getClassB(const Type *Ty) {
73 if (Ty == Type::BoolTy) return cByte;
78 struct ISel : public FunctionPass, InstVisitor<ISel> {
80 MachineFunction *F; // The function we are compiling into
81 MachineBasicBlock *BB; // The current MBB we are compiling
82 int VarArgsFrameIndex; // FrameIndex for start of varargs area
83 int ReturnAddressIndex; // FrameIndex for the return address
85 std::map<Value*, unsigned> RegMap; // Mapping between Val's and SSA Regs
87 // MBBMap - Mapping between LLVM BB -> Machine BB
88 std::map<const BasicBlock*, MachineBasicBlock*> MBBMap;
90 ISel(TargetMachine &tm) : TM(tm), F(0), BB(0) {}
92 /// runOnFunction - Top level implementation of instruction selection for
93 /// the entire function.
95 bool runOnFunction(Function &Fn) {
96 // First pass over the function, lower any unknown intrinsic functions
97 // with the IntrinsicLowering class.
98 LowerUnknownIntrinsicFunctionCalls(Fn);
100 F = &MachineFunction::construct(&Fn, TM);
102 // Create all of the machine basic blocks for the function...
103 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
104 F->getBasicBlockList().push_back(MBBMap[I] = new MachineBasicBlock(I));
108 // Set up a frame object for the return address. This is used by the
109 // llvm.returnaddress & llvm.frameaddress intrinisics.
110 ReturnAddressIndex = F->getFrameInfo()->CreateFixedObject(4, -4);
112 // Copy incoming arguments off of the stack...
113 LoadArgumentsToVirtualRegs(Fn);
115 // Instruction select everything except PHI nodes
118 // Select the PHI nodes
121 // Insert the FP_REG_KILL instructions into blocks that need them.
127 // We always build a machine code representation for the function
131 virtual const char *getPassName() const {
132 return "X86 Simple Instruction Selection";
135 /// visitBasicBlock - This method is called when we are visiting a new basic
136 /// block. This simply creates a new MachineBasicBlock to emit code into
137 /// and adds it to the current MachineFunction. Subsequent visit* for
138 /// instructions will be invoked for all instructions in the basic block.
140 void visitBasicBlock(BasicBlock &LLVM_BB) {
141 BB = MBBMap[&LLVM_BB];
144 /// LowerUnknownIntrinsicFunctionCalls - This performs a prepass over the
145 /// function, lowering any calls to unknown intrinsic functions into the
146 /// equivalent LLVM code.
148 void LowerUnknownIntrinsicFunctionCalls(Function &F);
150 /// LoadArgumentsToVirtualRegs - Load all of the arguments to this function
151 /// from the stack into virtual registers.
153 void LoadArgumentsToVirtualRegs(Function &F);
155 /// SelectPHINodes - Insert machine code to generate phis. This is tricky
156 /// because we have to generate our sources into the source basic blocks,
157 /// not the current one.
159 void SelectPHINodes();
161 /// InsertFPRegKills - Insert FP_REG_KILL instructions into basic blocks
162 /// that need them. This only occurs due to the floating point stackifier
163 /// not being aggressive enough to handle arbitrary global stackification.
165 void InsertFPRegKills();
167 // Visitation methods for various instructions. These methods simply emit
168 // fixed X86 code for each instruction.
171 // Control flow operators
172 void visitReturnInst(ReturnInst &RI);
173 void visitBranchInst(BranchInst &BI);
179 ValueRecord(unsigned R, const Type *T) : Val(0), Reg(R), Ty(T) {}
180 ValueRecord(Value *V) : Val(V), Reg(0), Ty(V->getType()) {}
182 void doCall(const ValueRecord &Ret, MachineInstr *CallMI,
183 const std::vector<ValueRecord> &Args);
184 void visitCallInst(CallInst &I);
185 void visitIntrinsicCall(Intrinsic::ID ID, CallInst &I);
187 // Arithmetic operators
188 void visitSimpleBinary(BinaryOperator &B, unsigned OpcodeClass);
189 void visitAdd(BinaryOperator &B) { visitSimpleBinary(B, 0); }
190 void visitSub(BinaryOperator &B) { visitSimpleBinary(B, 1); }
191 void visitMul(BinaryOperator &B);
193 void visitDiv(BinaryOperator &B) { visitDivRem(B); }
194 void visitRem(BinaryOperator &B) { visitDivRem(B); }
195 void visitDivRem(BinaryOperator &B);
198 void visitAnd(BinaryOperator &B) { visitSimpleBinary(B, 2); }
199 void visitOr (BinaryOperator &B) { visitSimpleBinary(B, 3); }
200 void visitXor(BinaryOperator &B) { visitSimpleBinary(B, 4); }
202 // Comparison operators...
203 void visitSetCondInst(SetCondInst &I);
204 unsigned EmitComparison(unsigned OpNum, Value *Op0, Value *Op1,
205 MachineBasicBlock *MBB,
206 MachineBasicBlock::iterator MBBI);
207 void visitSelectInst(SelectInst &SI);
210 // Memory Instructions
211 void visitLoadInst(LoadInst &I);
212 void visitStoreInst(StoreInst &I);
213 void visitGetElementPtrInst(GetElementPtrInst &I);
214 void visitAllocaInst(AllocaInst &I);
215 void visitMallocInst(MallocInst &I);
216 void visitFreeInst(FreeInst &I);
219 void visitShiftInst(ShiftInst &I);
220 void visitPHINode(PHINode &I) {} // PHI nodes handled by second pass
221 void visitCastInst(CastInst &I);
222 void visitVANextInst(VANextInst &I);
223 void visitVAArgInst(VAArgInst &I);
225 void visitInstruction(Instruction &I) {
226 std::cerr << "Cannot instruction select: " << I;
230 /// promote32 - Make a value 32-bits wide, and put it somewhere.
232 void promote32(unsigned targetReg, const ValueRecord &VR);
234 /// getAddressingMode - Get the addressing mode to use to address the
235 /// specified value. The returned value should be used with addFullAddress.
236 void getAddressingMode(Value *Addr, unsigned &BaseReg, unsigned &Scale,
237 unsigned &IndexReg, unsigned &Disp);
240 /// getGEPIndex - This is used to fold GEP instructions into X86 addressing
242 void getGEPIndex(MachineBasicBlock *MBB, MachineBasicBlock::iterator IP,
243 std::vector<Value*> &GEPOps,
244 std::vector<const Type*> &GEPTypes, unsigned &BaseReg,
245 unsigned &Scale, unsigned &IndexReg, unsigned &Disp);
247 /// isGEPFoldable - Return true if the specified GEP can be completely
248 /// folded into the addressing mode of a load/store or lea instruction.
249 bool isGEPFoldable(MachineBasicBlock *MBB,
250 Value *Src, User::op_iterator IdxBegin,
251 User::op_iterator IdxEnd, unsigned &BaseReg,
252 unsigned &Scale, unsigned &IndexReg, unsigned &Disp);
254 /// emitGEPOperation - Common code shared between visitGetElementPtrInst and
255 /// constant expression GEP support.
257 void emitGEPOperation(MachineBasicBlock *BB, MachineBasicBlock::iterator IP,
258 Value *Src, User::op_iterator IdxBegin,
259 User::op_iterator IdxEnd, unsigned TargetReg);
261 /// emitCastOperation - Common code shared between visitCastInst and
262 /// constant expression cast support.
264 void emitCastOperation(MachineBasicBlock *BB,MachineBasicBlock::iterator IP,
265 Value *Src, const Type *DestTy, unsigned TargetReg);
267 /// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary
268 /// and constant expression support.
270 void emitSimpleBinaryOperation(MachineBasicBlock *BB,
271 MachineBasicBlock::iterator IP,
272 Value *Op0, Value *Op1,
273 unsigned OperatorClass, unsigned TargetReg);
275 /// emitBinaryFPOperation - This method handles emission of floating point
276 /// Add (0), Sub (1), Mul (2), and Div (3) operations.
277 void emitBinaryFPOperation(MachineBasicBlock *BB,
278 MachineBasicBlock::iterator IP,
279 Value *Op0, Value *Op1,
280 unsigned OperatorClass, unsigned TargetReg);
282 void emitMultiply(MachineBasicBlock *BB, MachineBasicBlock::iterator IP,
283 Value *Op0, Value *Op1, unsigned TargetReg);
285 void doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator MBBI,
286 unsigned DestReg, const Type *DestTy,
287 unsigned Op0Reg, unsigned Op1Reg);
288 void doMultiplyConst(MachineBasicBlock *MBB,
289 MachineBasicBlock::iterator MBBI,
290 unsigned DestReg, const Type *DestTy,
291 unsigned Op0Reg, unsigned Op1Val);
293 void emitDivRemOperation(MachineBasicBlock *BB,
294 MachineBasicBlock::iterator IP,
295 Value *Op0, Value *Op1, bool isDiv,
298 /// emitSetCCOperation - Common code shared between visitSetCondInst and
299 /// constant expression support.
301 void emitSetCCOperation(MachineBasicBlock *BB,
302 MachineBasicBlock::iterator IP,
303 Value *Op0, Value *Op1, unsigned Opcode,
306 /// emitShiftOperation - Common code shared between visitShiftInst and
307 /// constant expression support.
309 void emitShiftOperation(MachineBasicBlock *MBB,
310 MachineBasicBlock::iterator IP,
311 Value *Op, Value *ShiftAmount, bool isLeftShift,
312 const Type *ResultTy, unsigned DestReg);
314 /// emitSelectOperation - Common code shared between visitSelectInst and the
315 /// constant expression support.
316 void emitSelectOperation(MachineBasicBlock *MBB,
317 MachineBasicBlock::iterator IP,
318 Value *Cond, Value *TrueVal, Value *FalseVal,
321 /// copyConstantToRegister - Output the instructions required to put the
322 /// specified constant into the specified register.
324 void copyConstantToRegister(MachineBasicBlock *MBB,
325 MachineBasicBlock::iterator MBBI,
326 Constant *C, unsigned Reg);
328 /// makeAnotherReg - This method returns the next register number we haven't
331 /// Long values are handled somewhat specially. They are always allocated
332 /// as pairs of 32 bit integer values. The register number returned is the
333 /// lower 32 bits of the long value, and the regNum+1 is the upper 32 bits
334 /// of the long value.
336 unsigned makeAnotherReg(const Type *Ty) {
337 assert(dynamic_cast<const X86RegisterInfo*>(TM.getRegisterInfo()) &&
338 "Current target doesn't have X86 reg info??");
339 const X86RegisterInfo *MRI =
340 static_cast<const X86RegisterInfo*>(TM.getRegisterInfo());
341 if (Ty == Type::LongTy || Ty == Type::ULongTy) {
342 const TargetRegisterClass *RC = MRI->getRegClassForType(Type::IntTy);
343 // Create the lower part
344 F->getSSARegMap()->createVirtualRegister(RC);
345 // Create the upper part.
346 return F->getSSARegMap()->createVirtualRegister(RC)-1;
349 // Add the mapping of regnumber => reg class to MachineFunction
350 const TargetRegisterClass *RC = MRI->getRegClassForType(Ty);
351 return F->getSSARegMap()->createVirtualRegister(RC);
354 /// getReg - This method turns an LLVM value into a register number. This
355 /// is guaranteed to produce the same register number for a particular value
356 /// every time it is queried.
358 unsigned getReg(Value &V) { return getReg(&V); } // Allow references
359 unsigned getReg(Value *V) {
360 // Just append to the end of the current bb.
361 MachineBasicBlock::iterator It = BB->end();
362 return getReg(V, BB, It);
364 unsigned getReg(Value *V, MachineBasicBlock *MBB,
365 MachineBasicBlock::iterator IPt) {
366 // If this operand is a constant, emit the code to copy the constant into
367 // the register here...
369 if (Constant *C = dyn_cast<Constant>(V)) {
370 unsigned Reg = makeAnotherReg(V->getType());
371 copyConstantToRegister(MBB, IPt, C, Reg);
373 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
374 unsigned Reg = makeAnotherReg(V->getType());
375 // Move the address of the global into the register
376 BuildMI(*MBB, IPt, X86::MOV32ri, 1, Reg).addGlobalAddress(GV);
378 } else if (CastInst *CI = dyn_cast<CastInst>(V)) {
379 // Do not emit noop casts at all.
380 if (getClassB(CI->getType()) == getClassB(CI->getOperand(0)->getType()))
381 return getReg(CI->getOperand(0), MBB, IPt);
384 unsigned &Reg = RegMap[V];
386 Reg = makeAnotherReg(V->getType());
395 /// copyConstantToRegister - Output the instructions required to put the
396 /// specified constant into the specified register.
398 void ISel::copyConstantToRegister(MachineBasicBlock *MBB,
399 MachineBasicBlock::iterator IP,
400 Constant *C, unsigned R) {
401 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
403 switch (CE->getOpcode()) {
404 case Instruction::GetElementPtr:
405 emitGEPOperation(MBB, IP, CE->getOperand(0),
406 CE->op_begin()+1, CE->op_end(), R);
408 case Instruction::Cast:
409 emitCastOperation(MBB, IP, CE->getOperand(0), CE->getType(), R);
412 case Instruction::Xor: ++Class; // FALL THROUGH
413 case Instruction::Or: ++Class; // FALL THROUGH
414 case Instruction::And: ++Class; // FALL THROUGH
415 case Instruction::Sub: ++Class; // FALL THROUGH
416 case Instruction::Add:
417 emitSimpleBinaryOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1),
421 case Instruction::Mul:
422 emitMultiply(MBB, IP, CE->getOperand(0), CE->getOperand(1), R);
425 case Instruction::Div:
426 case Instruction::Rem:
427 emitDivRemOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1),
428 CE->getOpcode() == Instruction::Div, R);
431 case Instruction::SetNE:
432 case Instruction::SetEQ:
433 case Instruction::SetLT:
434 case Instruction::SetGT:
435 case Instruction::SetLE:
436 case Instruction::SetGE:
437 emitSetCCOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1),
441 case Instruction::Shl:
442 case Instruction::Shr:
443 emitShiftOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1),
444 CE->getOpcode() == Instruction::Shl, CE->getType(), R);
447 case Instruction::Select:
448 emitSelectOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1),
449 CE->getOperand(2), R);
453 std::cerr << "Offending expr: " << C << "\n";
454 assert(0 && "Constant expression not yet handled!\n");
458 if (C->getType()->isIntegral()) {
459 unsigned Class = getClassB(C->getType());
461 if (Class == cLong) {
462 // Copy the value into the register pair.
463 uint64_t Val = cast<ConstantInt>(C)->getRawValue();
464 BuildMI(*MBB, IP, X86::MOV32ri, 1, R).addImm(Val & 0xFFFFFFFF);
465 BuildMI(*MBB, IP, X86::MOV32ri, 1, R+1).addImm(Val >> 32);
469 assert(Class <= cInt && "Type not handled yet!");
471 static const unsigned IntegralOpcodeTab[] = {
472 X86::MOV8ri, X86::MOV16ri, X86::MOV32ri
475 if (C->getType() == Type::BoolTy) {
476 BuildMI(*MBB, IP, X86::MOV8ri, 1, R).addImm(C == ConstantBool::True);
478 ConstantInt *CI = cast<ConstantInt>(C);
479 BuildMI(*MBB, IP, IntegralOpcodeTab[Class],1,R).addImm(CI->getRawValue());
481 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
482 if (CFP->isExactlyValue(+0.0))
483 BuildMI(*MBB, IP, X86::FLD0, 0, R);
484 else if (CFP->isExactlyValue(+1.0))
485 BuildMI(*MBB, IP, X86::FLD1, 0, R);
487 // Otherwise we need to spill the constant to memory...
488 MachineConstantPool *CP = F->getConstantPool();
489 unsigned CPI = CP->getConstantPoolIndex(CFP);
490 const Type *Ty = CFP->getType();
492 assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
493 unsigned LoadOpcode = Ty == Type::FloatTy ? X86::FLD32m : X86::FLD64m;
494 addConstantPoolReference(BuildMI(*MBB, IP, LoadOpcode, 4, R), CPI);
497 } else if (isa<ConstantPointerNull>(C)) {
498 // Copy zero (null pointer) to the register.
499 BuildMI(*MBB, IP, X86::MOV32ri, 1, R).addImm(0);
500 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
501 BuildMI(*MBB, IP, X86::MOV32ri, 1, R).addGlobalAddress(CPR->getValue());
503 std::cerr << "Offending constant: " << C << "\n";
504 assert(0 && "Type not handled yet!");
508 /// LoadArgumentsToVirtualRegs - Load all of the arguments to this function from
509 /// the stack into virtual registers.
511 void ISel::LoadArgumentsToVirtualRegs(Function &Fn) {
512 // Emit instructions to load the arguments... On entry to a function on the
513 // X86, the stack frame looks like this:
515 // [ESP] -- return address
516 // [ESP + 4] -- first argument (leftmost lexically)
517 // [ESP + 8] -- second argument, if first argument is four bytes in size
520 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
521 MachineFrameInfo *MFI = F->getFrameInfo();
523 for (Function::aiterator I = Fn.abegin(), E = Fn.aend(); I != E; ++I) {
524 bool ArgLive = !I->use_empty();
525 unsigned Reg = ArgLive ? getReg(*I) : 0;
526 int FI; // Frame object index
528 switch (getClassB(I->getType())) {
531 FI = MFI->CreateFixedObject(1, ArgOffset);
532 addFrameReference(BuildMI(BB, X86::MOV8rm, 4, Reg), FI);
537 FI = MFI->CreateFixedObject(2, ArgOffset);
538 addFrameReference(BuildMI(BB, X86::MOV16rm, 4, Reg), FI);
543 FI = MFI->CreateFixedObject(4, ArgOffset);
544 addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg), FI);
549 FI = MFI->CreateFixedObject(8, ArgOffset);
550 addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg), FI);
551 addFrameReference(BuildMI(BB, X86::MOV32rm, 4, Reg+1), FI, 4);
553 ArgOffset += 4; // longs require 4 additional bytes
558 if (I->getType() == Type::FloatTy) {
559 Opcode = X86::FLD32m;
560 FI = MFI->CreateFixedObject(4, ArgOffset);
562 Opcode = X86::FLD64m;
563 FI = MFI->CreateFixedObject(8, ArgOffset);
565 addFrameReference(BuildMI(BB, Opcode, 4, Reg), FI);
567 if (I->getType() == Type::DoubleTy)
568 ArgOffset += 4; // doubles require 4 additional bytes
571 assert(0 && "Unhandled argument type!");
573 ArgOffset += 4; // Each argument takes at least 4 bytes on the stack...
576 // If the function takes variable number of arguments, add a frame offset for
577 // the start of the first vararg value... this is used to expand
579 if (Fn.getFunctionType()->isVarArg())
580 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
584 /// SelectPHINodes - Insert machine code to generate phis. This is tricky
585 /// because we have to generate our sources into the source basic blocks, not
588 void ISel::SelectPHINodes() {
589 const TargetInstrInfo &TII = TM.getInstrInfo();
590 const Function &LF = *F->getFunction(); // The LLVM function...
591 for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) {
592 const BasicBlock *BB = I;
593 MachineBasicBlock &MBB = *MBBMap[I];
595 // Loop over all of the PHI nodes in the LLVM basic block...
596 MachineBasicBlock::iterator PHIInsertPoint = MBB.begin();
597 for (BasicBlock::const_iterator I = BB->begin();
598 PHINode *PN = const_cast<PHINode*>(dyn_cast<PHINode>(I)); ++I) {
600 // Create a new machine instr PHI node, and insert it.
601 unsigned PHIReg = getReg(*PN);
602 MachineInstr *PhiMI = BuildMI(MBB, PHIInsertPoint,
603 X86::PHI, PN->getNumOperands(), PHIReg);
605 MachineInstr *LongPhiMI = 0;
606 if (PN->getType() == Type::LongTy || PN->getType() == Type::ULongTy)
607 LongPhiMI = BuildMI(MBB, PHIInsertPoint,
608 X86::PHI, PN->getNumOperands(), PHIReg+1);
610 // PHIValues - Map of blocks to incoming virtual registers. We use this
611 // so that we only initialize one incoming value for a particular block,
612 // even if the block has multiple entries in the PHI node.
614 std::map<MachineBasicBlock*, unsigned> PHIValues;
616 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
617 MachineBasicBlock *PredMBB = MBBMap[PN->getIncomingBlock(i)];
619 std::map<MachineBasicBlock*, unsigned>::iterator EntryIt =
620 PHIValues.lower_bound(PredMBB);
622 if (EntryIt != PHIValues.end() && EntryIt->first == PredMBB) {
623 // We already inserted an initialization of the register for this
624 // predecessor. Recycle it.
625 ValReg = EntryIt->second;
628 // Get the incoming value into a virtual register.
630 Value *Val = PN->getIncomingValue(i);
632 // If this is a constant or GlobalValue, we may have to insert code
633 // into the basic block to compute it into a virtual register.
634 if (isa<Constant>(Val) || isa<GlobalValue>(Val)) {
635 if (isa<ConstantExpr>(Val)) {
636 // Because we don't want to clobber any values which might be in
637 // physical registers with the computation of this constant (which
638 // might be arbitrarily complex if it is a constant expression),
639 // just insert the computation at the top of the basic block.
640 MachineBasicBlock::iterator PI = PredMBB->begin();
642 // Skip over any PHI nodes though!
643 while (PI != PredMBB->end() && PI->getOpcode() == X86::PHI)
646 ValReg = getReg(Val, PredMBB, PI);
648 // Simple constants get emitted at the end of the basic block,
649 // before any terminator instructions. We "know" that the code to
650 // move a constant into a register will never clobber any flags.
651 ValReg = getReg(Val, PredMBB, PredMBB->getFirstTerminator());
654 ValReg = getReg(Val);
657 // Remember that we inserted a value for this PHI for this predecessor
658 PHIValues.insert(EntryIt, std::make_pair(PredMBB, ValReg));
661 PhiMI->addRegOperand(ValReg);
662 PhiMI->addMachineBasicBlockOperand(PredMBB);
664 LongPhiMI->addRegOperand(ValReg+1);
665 LongPhiMI->addMachineBasicBlockOperand(PredMBB);
669 // Now that we emitted all of the incoming values for the PHI node, make
670 // sure to reposition the InsertPoint after the PHI that we just added.
671 // This is needed because we might have inserted a constant into this
672 // block, right after the PHI's which is before the old insert point!
673 PHIInsertPoint = LongPhiMI ? LongPhiMI : PhiMI;
679 /// RequiresFPRegKill - The floating point stackifier pass cannot insert
680 /// compensation code on critical edges. As such, it requires that we kill all
681 /// FP registers on the exit from any blocks that either ARE critical edges, or
682 /// branch to a block that has incoming critical edges.
684 /// Note that this kill instruction will eventually be eliminated when
685 /// restrictions in the stackifier are relaxed.
687 static bool RequiresFPRegKill(const BasicBlock *BB) {
689 for (succ_const_iterator SI = succ_begin(BB), E = succ_end(BB); SI!=E; ++SI) {
690 const BasicBlock *Succ = *SI;
691 pred_const_iterator PI = pred_begin(Succ), PE = pred_end(Succ);
692 ++PI; // Block have at least one predecessory
693 if (PI != PE) { // If it has exactly one, this isn't crit edge
694 // If this block has more than one predecessor, check all of the
695 // predecessors to see if they have multiple successors. If so, then the
696 // block we are analyzing needs an FPRegKill.
697 for (PI = pred_begin(Succ); PI != PE; ++PI) {
698 const BasicBlock *Pred = *PI;
699 succ_const_iterator SI2 = succ_begin(Pred);
700 ++SI2; // There must be at least one successor of this block.
701 if (SI2 != succ_end(Pred))
702 return true; // Yes, we must insert the kill on this edge.
706 // If we got this far, there is no need to insert the kill instruction.
713 // InsertFPRegKills - Insert FP_REG_KILL instructions into basic blocks that
714 // need them. This only occurs due to the floating point stackifier not being
715 // aggressive enough to handle arbitrary global stackification.
717 // Currently we insert an FP_REG_KILL instruction into each block that uses or
718 // defines a floating point virtual register.
720 // When the global register allocators (like linear scan) finally update live
721 // variable analysis, we can keep floating point values in registers across
722 // portions of the CFG that do not involve critical edges. This will be a big
723 // win, but we are waiting on the global allocators before we can do this.
725 // With a bit of work, the floating point stackifier pass can be enhanced to
726 // break critical edges as needed (to make a place to put compensation code),
727 // but this will require some infrastructure improvements as well.
729 void ISel::InsertFPRegKills() {
730 SSARegMap &RegMap = *F->getSSARegMap();
732 for (MachineFunction::iterator BB = F->begin(), E = F->end(); BB != E; ++BB) {
733 for (MachineBasicBlock::iterator I = BB->begin(), E = BB->end(); I!=E; ++I)
734 for (unsigned i = 0, e = I->getNumOperands(); i != e; ++i) {
735 MachineOperand& MO = I->getOperand(i);
736 if (MO.isRegister() && MO.getReg()) {
737 unsigned Reg = MO.getReg();
738 if (MRegisterInfo::isVirtualRegister(Reg))
739 if (RegMap.getRegClass(Reg)->getSize() == 10)
743 // If we haven't found an FP register use or def in this basic block, check
744 // to see if any of our successors has an FP PHI node, which will cause a
745 // copy to be inserted into this block.
746 for (succ_const_iterator SI = succ_begin(BB->getBasicBlock()),
747 E = succ_end(BB->getBasicBlock()); SI != E; ++SI) {
748 MachineBasicBlock *SBB = MBBMap[*SI];
749 for (MachineBasicBlock::iterator I = SBB->begin();
750 I != SBB->end() && I->getOpcode() == X86::PHI; ++I) {
751 if (RegMap.getRegClass(I->getOperand(0).getReg())->getSize() == 10)
757 // Okay, this block uses an FP register. If the block has successors (ie,
758 // it's not an unwind/return), insert the FP_REG_KILL instruction.
759 if (BB->getBasicBlock()->getTerminator()->getNumSuccessors() &&
760 RequiresFPRegKill(BB->getBasicBlock())) {
761 BuildMI(*BB, BB->getFirstTerminator(), X86::FP_REG_KILL, 0);
768 // canFoldSetCCIntoBranchOrSelect - Return the setcc instruction if we can fold
769 // it into the conditional branch or select instruction which is the only user
770 // of the cc instruction. This is the case if the conditional branch is the
771 // only user of the setcc, and if the setcc is in the same basic block as the
772 // conditional branch. We also don't handle long arguments below, so we reject
773 // them here as well.
775 static SetCondInst *canFoldSetCCIntoBranchOrSelect(Value *V) {
776 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
777 if (SCI->hasOneUse()) {
778 Instruction *User = cast<Instruction>(SCI->use_back());
779 if ((isa<BranchInst>(User) || isa<SelectInst>(User)) &&
780 SCI->getParent() == User->getParent() &&
781 (getClassB(SCI->getOperand(0)->getType()) != cLong ||
782 SCI->getOpcode() == Instruction::SetEQ ||
783 SCI->getOpcode() == Instruction::SetNE))
789 // Return a fixed numbering for setcc instructions which does not depend on the
790 // order of the opcodes.
792 static unsigned getSetCCNumber(unsigned Opcode) {
794 default: assert(0 && "Unknown setcc instruction!");
795 case Instruction::SetEQ: return 0;
796 case Instruction::SetNE: return 1;
797 case Instruction::SetLT: return 2;
798 case Instruction::SetGE: return 3;
799 case Instruction::SetGT: return 4;
800 case Instruction::SetLE: return 5;
804 // LLVM -> X86 signed X86 unsigned
805 // ----- ---------- ------------
806 // seteq -> sete sete
807 // setne -> setne setne
808 // setlt -> setl setb
809 // setge -> setge setae
810 // setgt -> setg seta
811 // setle -> setle setbe
813 // sets // Used by comparison with 0 optimization
815 static const unsigned SetCCOpcodeTab[2][8] = {
816 { X86::SETEr, X86::SETNEr, X86::SETBr, X86::SETAEr, X86::SETAr, X86::SETBEr,
818 { X86::SETEr, X86::SETNEr, X86::SETLr, X86::SETGEr, X86::SETGr, X86::SETLEr,
819 X86::SETSr, X86::SETNSr },
822 // EmitComparison - This function emits a comparison of the two operands,
823 // returning the extended setcc code to use.
824 unsigned ISel::EmitComparison(unsigned OpNum, Value *Op0, Value *Op1,
825 MachineBasicBlock *MBB,
826 MachineBasicBlock::iterator IP) {
827 // The arguments are already supposed to be of the same type.
828 const Type *CompTy = Op0->getType();
829 unsigned Class = getClassB(CompTy);
830 unsigned Op0r = getReg(Op0, MBB, IP);
832 // Special case handling of: cmp R, i
833 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
834 if (Class == cByte || Class == cShort || Class == cInt) {
835 unsigned Op1v = CI->getRawValue();
837 // Mask off any upper bits of the constant, if there are any...
838 Op1v &= (1ULL << (8 << Class)) - 1;
840 // If this is a comparison against zero, emit more efficient code. We
841 // can't handle unsigned comparisons against zero unless they are == or
842 // !=. These should have been strength reduced already anyway.
843 if (Op1v == 0 && (CompTy->isSigned() || OpNum < 2)) {
844 static const unsigned TESTTab[] = {
845 X86::TEST8rr, X86::TEST16rr, X86::TEST32rr
847 BuildMI(*MBB, IP, TESTTab[Class], 2).addReg(Op0r).addReg(Op0r);
849 if (OpNum == 2) return 6; // Map jl -> js
850 if (OpNum == 3) return 7; // Map jg -> jns
854 static const unsigned CMPTab[] = {
855 X86::CMP8ri, X86::CMP16ri, X86::CMP32ri
858 BuildMI(*MBB, IP, CMPTab[Class], 2).addReg(Op0r).addImm(Op1v);
861 assert(Class == cLong && "Unknown integer class!");
862 unsigned LowCst = CI->getRawValue();
863 unsigned HiCst = CI->getRawValue() >> 32;
864 if (OpNum < 2) { // seteq, setne
865 unsigned LoTmp = Op0r;
867 LoTmp = makeAnotherReg(Type::IntTy);
868 BuildMI(*MBB, IP, X86::XOR32ri, 2, LoTmp).addReg(Op0r).addImm(LowCst);
870 unsigned HiTmp = Op0r+1;
872 HiTmp = makeAnotherReg(Type::IntTy);
873 BuildMI(*MBB, IP, X86::XOR32ri, 2,HiTmp).addReg(Op0r+1).addImm(HiCst);
875 unsigned FinalTmp = makeAnotherReg(Type::IntTy);
876 BuildMI(*MBB, IP, X86::OR32rr, 2, FinalTmp).addReg(LoTmp).addReg(HiTmp);
879 // Emit a sequence of code which compares the high and low parts once
880 // each, then uses a conditional move to handle the overflow case. For
881 // example, a setlt for long would generate code like this:
883 // AL = lo(op1) < lo(op2) // Signedness depends on operands
884 // BL = hi(op1) < hi(op2) // Always unsigned comparison
885 // dest = hi(op1) == hi(op2) ? AL : BL;
888 // FIXME: This would be much better if we had hierarchical register
889 // classes! Until then, hardcode registers so that we can deal with
890 // their aliases (because we don't have conditional byte moves).
892 BuildMI(*MBB, IP, X86::CMP32ri, 2).addReg(Op0r).addImm(LowCst);
893 BuildMI(*MBB, IP, SetCCOpcodeTab[0][OpNum], 0, X86::AL);
894 BuildMI(*MBB, IP, X86::CMP32ri, 2).addReg(Op0r+1).addImm(HiCst);
895 BuildMI(*MBB, IP, SetCCOpcodeTab[CompTy->isSigned()][OpNum], 0,X86::BL);
896 BuildMI(*MBB, IP, X86::IMPLICIT_DEF, 0, X86::BH);
897 BuildMI(*MBB, IP, X86::IMPLICIT_DEF, 0, X86::AH);
898 BuildMI(*MBB, IP, X86::CMOVE16rr, 2, X86::BX).addReg(X86::BX)
900 // NOTE: visitSetCondInst knows that the value is dumped into the BL
901 // register at this point for long values...
907 // Special case handling of comparison against +/- 0.0
908 if (ConstantFP *CFP = dyn_cast<ConstantFP>(Op1))
909 if (CFP->isExactlyValue(+0.0) || CFP->isExactlyValue(-0.0)) {
910 BuildMI(*MBB, IP, X86::FTST, 1).addReg(Op0r);
911 BuildMI(*MBB, IP, X86::FNSTSW8r, 0);
912 BuildMI(*MBB, IP, X86::SAHF, 1);
916 unsigned Op1r = getReg(Op1, MBB, IP);
918 default: assert(0 && "Unknown type class!");
919 // Emit: cmp <var1>, <var2> (do the comparison). We can
920 // compare 8-bit with 8-bit, 16-bit with 16-bit, 32-bit with
923 BuildMI(*MBB, IP, X86::CMP8rr, 2).addReg(Op0r).addReg(Op1r);
926 BuildMI(*MBB, IP, X86::CMP16rr, 2).addReg(Op0r).addReg(Op1r);
929 BuildMI(*MBB, IP, X86::CMP32rr, 2).addReg(Op0r).addReg(Op1r);
932 BuildMI(*MBB, IP, X86::FpUCOM, 2).addReg(Op0r).addReg(Op1r);
933 BuildMI(*MBB, IP, X86::FNSTSW8r, 0);
934 BuildMI(*MBB, IP, X86::SAHF, 1);
938 if (OpNum < 2) { // seteq, setne
939 unsigned LoTmp = makeAnotherReg(Type::IntTy);
940 unsigned HiTmp = makeAnotherReg(Type::IntTy);
941 unsigned FinalTmp = makeAnotherReg(Type::IntTy);
942 BuildMI(*MBB, IP, X86::XOR32rr, 2, LoTmp).addReg(Op0r).addReg(Op1r);
943 BuildMI(*MBB, IP, X86::XOR32rr, 2, HiTmp).addReg(Op0r+1).addReg(Op1r+1);
944 BuildMI(*MBB, IP, X86::OR32rr, 2, FinalTmp).addReg(LoTmp).addReg(HiTmp);
945 break; // Allow the sete or setne to be generated from flags set by OR
947 // Emit a sequence of code which compares the high and low parts once
948 // each, then uses a conditional move to handle the overflow case. For
949 // example, a setlt for long would generate code like this:
951 // AL = lo(op1) < lo(op2) // Signedness depends on operands
952 // BL = hi(op1) < hi(op2) // Always unsigned comparison
953 // dest = hi(op1) == hi(op2) ? AL : BL;
956 // FIXME: This would be much better if we had hierarchical register
957 // classes! Until then, hardcode registers so that we can deal with their
958 // aliases (because we don't have conditional byte moves).
960 BuildMI(*MBB, IP, X86::CMP32rr, 2).addReg(Op0r).addReg(Op1r);
961 BuildMI(*MBB, IP, SetCCOpcodeTab[0][OpNum], 0, X86::AL);
962 BuildMI(*MBB, IP, X86::CMP32rr, 2).addReg(Op0r+1).addReg(Op1r+1);
963 BuildMI(*MBB, IP, SetCCOpcodeTab[CompTy->isSigned()][OpNum], 0, X86::BL);
964 BuildMI(*MBB, IP, X86::IMPLICIT_DEF, 0, X86::BH);
965 BuildMI(*MBB, IP, X86::IMPLICIT_DEF, 0, X86::AH);
966 BuildMI(*MBB, IP, X86::CMOVE16rr, 2, X86::BX).addReg(X86::BX)
968 // NOTE: visitSetCondInst knows that the value is dumped into the BL
969 // register at this point for long values...
976 /// SetCC instructions - Here we just emit boilerplate code to set a byte-sized
977 /// register, then move it to wherever the result should be.
979 void ISel::visitSetCondInst(SetCondInst &I) {
980 if (canFoldSetCCIntoBranchOrSelect(&I))
981 return; // Fold this into a branch or select.
983 unsigned DestReg = getReg(I);
984 MachineBasicBlock::iterator MII = BB->end();
985 emitSetCCOperation(BB, MII, I.getOperand(0), I.getOperand(1), I.getOpcode(),
989 /// emitSetCCOperation - Common code shared between visitSetCondInst and
990 /// constant expression support.
992 void ISel::emitSetCCOperation(MachineBasicBlock *MBB,
993 MachineBasicBlock::iterator IP,
994 Value *Op0, Value *Op1, unsigned Opcode,
995 unsigned TargetReg) {
996 unsigned OpNum = getSetCCNumber(Opcode);
997 OpNum = EmitComparison(OpNum, Op0, Op1, MBB, IP);
999 const Type *CompTy = Op0->getType();
1000 unsigned CompClass = getClassB(CompTy);
1001 bool isSigned = CompTy->isSigned() && CompClass != cFP;
1003 if (CompClass != cLong || OpNum < 2) {
1004 // Handle normal comparisons with a setcc instruction...
1005 BuildMI(*MBB, IP, SetCCOpcodeTab[isSigned][OpNum], 0, TargetReg);
1007 // Handle long comparisons by copying the value which is already in BL into
1008 // the register we want...
1009 BuildMI(*MBB, IP, X86::MOV8rr, 1, TargetReg).addReg(X86::BL);
1013 void ISel::visitSelectInst(SelectInst &SI) {
1014 unsigned DestReg = getReg(SI);
1015 MachineBasicBlock::iterator MII = BB->end();
1016 emitSelectOperation(BB, MII, SI.getCondition(), SI.getTrueValue(),
1017 SI.getFalseValue(), DestReg);
1020 /// emitSelect - Common code shared between visitSelectInst and the constant
1021 /// expression support.
1022 void ISel::emitSelectOperation(MachineBasicBlock *MBB,
1023 MachineBasicBlock::iterator IP,
1024 Value *Cond, Value *TrueVal, Value *FalseVal,
1026 unsigned SelectClass = getClassB(TrueVal->getType());
1028 // We don't support 8-bit conditional moves. If we have incoming constants,
1029 // transform them into 16-bit constants to avoid having a run-time conversion.
1030 if (SelectClass == cByte) {
1031 if (Constant *T = dyn_cast<Constant>(TrueVal))
1032 TrueVal = ConstantExpr::getCast(T, Type::ShortTy);
1033 if (Constant *F = dyn_cast<Constant>(FalseVal))
1034 FalseVal = ConstantExpr::getCast(F, Type::ShortTy);
1039 if (SetCondInst *SCI = canFoldSetCCIntoBranchOrSelect(Cond)) {
1040 // We successfully folded the setcc into the select instruction.
1042 unsigned OpNum = getSetCCNumber(SCI->getOpcode());
1043 OpNum = EmitComparison(OpNum, SCI->getOperand(0), SCI->getOperand(1), MBB,
1046 const Type *CompTy = SCI->getOperand(0)->getType();
1047 bool isSigned = CompTy->isSigned() && getClassB(CompTy) != cFP;
1049 // LLVM -> X86 signed X86 unsigned
1050 // ----- ---------- ------------
1051 // seteq -> cmovNE cmovNE
1052 // setne -> cmovE cmovE
1053 // setlt -> cmovGE cmovAE
1054 // setge -> cmovL cmovB
1055 // setgt -> cmovLE cmovBE
1056 // setle -> cmovG cmovA
1058 // cmovNS // Used by comparison with 0 optimization
1061 switch (SelectClass) {
1062 default: assert(0 && "Unknown value class!");
1064 // Annoyingly, we don't have a full set of floating point conditional
1066 static const unsigned OpcodeTab[2][8] = {
1067 { X86::FCMOVNE, X86::FCMOVE, X86::FCMOVAE, X86::FCMOVB,
1068 X86::FCMOVBE, X86::FCMOVA, 0, 0 },
1069 { X86::FCMOVNE, X86::FCMOVE, 0, 0, 0, 0, 0, 0 },
1071 Opcode = OpcodeTab[isSigned][OpNum];
1073 // If opcode == 0, we hit a case that we don't support. Output a setcc
1074 // and compare the result against zero.
1076 unsigned CompClass = getClassB(CompTy);
1078 if (CompClass != cLong || OpNum < 2) {
1079 CondReg = makeAnotherReg(Type::BoolTy);
1080 // Handle normal comparisons with a setcc instruction...
1081 BuildMI(*MBB, IP, SetCCOpcodeTab[isSigned][OpNum], 0, CondReg);
1083 // Long comparisons end up in the BL register.
1087 BuildMI(*MBB, IP, X86::TEST8rr, 2).addReg(CondReg).addReg(CondReg);
1088 Opcode = X86::FCMOVE;
1094 static const unsigned OpcodeTab[2][8] = {
1095 { X86::CMOVNE16rr, X86::CMOVE16rr, X86::CMOVAE16rr, X86::CMOVB16rr,
1096 X86::CMOVBE16rr, X86::CMOVA16rr, 0, 0 },
1097 { X86::CMOVNE16rr, X86::CMOVE16rr, X86::CMOVGE16rr, X86::CMOVL16rr,
1098 X86::CMOVLE16rr, X86::CMOVG16rr, X86::CMOVNS16rr, X86::CMOVS16rr },
1100 Opcode = OpcodeTab[isSigned][OpNum];
1105 static const unsigned OpcodeTab[2][8] = {
1106 { X86::CMOVNE32rr, X86::CMOVE32rr, X86::CMOVAE32rr, X86::CMOVB32rr,
1107 X86::CMOVBE32rr, X86::CMOVA32rr, 0, 0 },
1108 { X86::CMOVNE32rr, X86::CMOVE32rr, X86::CMOVGE32rr, X86::CMOVL32rr,
1109 X86::CMOVLE32rr, X86::CMOVG32rr, X86::CMOVNS32rr, X86::CMOVS32rr },
1111 Opcode = OpcodeTab[isSigned][OpNum];
1116 // Get the value being branched on, and use it to set the condition codes.
1117 unsigned CondReg = getReg(Cond, MBB, IP);
1118 BuildMI(*MBB, IP, X86::TEST8rr, 2).addReg(CondReg).addReg(CondReg);
1119 switch (SelectClass) {
1120 default: assert(0 && "Unknown value class!");
1121 case cFP: Opcode = X86::FCMOVE; break;
1123 case cShort: Opcode = X86::CMOVE16rr; break;
1125 case cLong: Opcode = X86::CMOVE32rr; break;
1129 unsigned TrueReg = getReg(TrueVal, MBB, IP);
1130 unsigned FalseReg = getReg(FalseVal, MBB, IP);
1131 unsigned RealDestReg = DestReg;
1134 // Annoyingly enough, X86 doesn't HAVE 8-bit conditional moves. Because of
1135 // this, we have to promote the incoming values to 16 bits, perform a 16-bit
1136 // cmove, then truncate the result.
1137 if (SelectClass == cByte) {
1138 DestReg = makeAnotherReg(Type::ShortTy);
1139 if (getClassB(TrueVal->getType()) == cByte) {
1140 // Promote the true value, by storing it into AL, and reading from AX.
1141 BuildMI(*MBB, IP, X86::MOV8rr, 1, X86::AL).addReg(TrueReg);
1142 BuildMI(*MBB, IP, X86::MOV8ri, 1, X86::AH).addImm(0);
1143 TrueReg = makeAnotherReg(Type::ShortTy);
1144 BuildMI(*MBB, IP, X86::MOV16rr, 1, TrueReg).addReg(X86::AX);
1146 if (getClassB(FalseVal->getType()) == cByte) {
1147 // Promote the true value, by storing it into CL, and reading from CX.
1148 BuildMI(*MBB, IP, X86::MOV8rr, 1, X86::CL).addReg(FalseReg);
1149 BuildMI(*MBB, IP, X86::MOV8ri, 1, X86::CH).addImm(0);
1150 FalseReg = makeAnotherReg(Type::ShortTy);
1151 BuildMI(*MBB, IP, X86::MOV16rr, 1, FalseReg).addReg(X86::CX);
1155 BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(TrueReg).addReg(FalseReg);
1157 switch (SelectClass) {
1159 // We did the computation with 16-bit registers. Truncate back to our
1160 // result by copying into AX then copying out AL.
1161 BuildMI(*MBB, IP, X86::MOV16rr, 1, X86::AX).addReg(DestReg);
1162 BuildMI(*MBB, IP, X86::MOV8rr, 1, RealDestReg).addReg(X86::AL);
1165 // Move the upper half of the value as well.
1166 BuildMI(*MBB, IP, Opcode, 2,DestReg+1).addReg(TrueReg+1).addReg(FalseReg+1);
1173 /// promote32 - Emit instructions to turn a narrow operand into a 32-bit-wide
1174 /// operand, in the specified target register.
1176 void ISel::promote32(unsigned targetReg, const ValueRecord &VR) {
1177 bool isUnsigned = VR.Ty->isUnsigned();
1179 Value *Val = VR.Val;
1180 const Type *Ty = VR.Ty;
1182 if (Constant *C = dyn_cast<Constant>(Val)) {
1183 Val = ConstantExpr::getCast(C, Type::IntTy);
1187 // If this is a simple constant, just emit a MOVri directly to avoid the
1189 if (ConstantInt *CI = dyn_cast<ConstantInt>(Val)) {
1190 int TheVal = CI->getRawValue() & 0xFFFFFFFF;
1191 BuildMI(BB, X86::MOV32ri, 1, targetReg).addImm(TheVal);
1196 // Make sure we have the register number for this value...
1197 unsigned Reg = Val ? getReg(Val) : VR.Reg;
1199 switch (getClassB(Ty)) {
1201 // Extend value into target register (8->32)
1203 BuildMI(BB, X86::MOVZX32rr8, 1, targetReg).addReg(Reg);
1205 BuildMI(BB, X86::MOVSX32rr8, 1, targetReg).addReg(Reg);
1208 // Extend value into target register (16->32)
1210 BuildMI(BB, X86::MOVZX32rr16, 1, targetReg).addReg(Reg);
1212 BuildMI(BB, X86::MOVSX32rr16, 1, targetReg).addReg(Reg);
1215 // Move value into target register (32->32)
1216 BuildMI(BB, X86::MOV32rr, 1, targetReg).addReg(Reg);
1219 assert(0 && "Unpromotable operand class in promote32");
1223 /// 'ret' instruction - Here we are interested in meeting the x86 ABI. As such,
1224 /// we have the following possibilities:
1226 /// ret void: No return value, simply emit a 'ret' instruction
1227 /// ret sbyte, ubyte : Extend value into EAX and return
1228 /// ret short, ushort: Extend value into EAX and return
1229 /// ret int, uint : Move value into EAX and return
1230 /// ret pointer : Move value into EAX and return
1231 /// ret long, ulong : Move value into EAX/EDX and return
1232 /// ret float/double : Top of FP stack
1234 void ISel::visitReturnInst(ReturnInst &I) {
1235 if (I.getNumOperands() == 0) {
1236 BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
1240 Value *RetVal = I.getOperand(0);
1241 switch (getClassB(RetVal->getType())) {
1242 case cByte: // integral return values: extend or move into EAX and return
1245 promote32(X86::EAX, ValueRecord(RetVal));
1246 // Declare that EAX is live on exit
1247 BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::EAX).addReg(X86::ESP);
1249 case cFP: { // Floats & Doubles: Return in ST(0)
1250 unsigned RetReg = getReg(RetVal);
1251 BuildMI(BB, X86::FpSETRESULT, 1).addReg(RetReg);
1252 // Declare that top-of-stack is live on exit
1253 BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::ST0).addReg(X86::ESP);
1257 unsigned RetReg = getReg(RetVal);
1258 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(RetReg);
1259 BuildMI(BB, X86::MOV32rr, 1, X86::EDX).addReg(RetReg+1);
1260 // Declare that EAX & EDX are live on exit
1261 BuildMI(BB, X86::IMPLICIT_USE, 3).addReg(X86::EAX).addReg(X86::EDX)
1266 visitInstruction(I);
1268 // Emit a 'ret' instruction
1269 BuildMI(BB, X86::RET, 0);
1272 // getBlockAfter - Return the basic block which occurs lexically after the
1274 static inline BasicBlock *getBlockAfter(BasicBlock *BB) {
1275 Function::iterator I = BB; ++I; // Get iterator to next block
1276 return I != BB->getParent()->end() ? &*I : 0;
1279 /// visitBranchInst - Handle conditional and unconditional branches here. Note
1280 /// that since code layout is frozen at this point, that if we are trying to
1281 /// jump to a block that is the immediate successor of the current block, we can
1282 /// just make a fall-through (but we don't currently).
1284 void ISel::visitBranchInst(BranchInst &BI) {
1285 BasicBlock *NextBB = getBlockAfter(BI.getParent()); // BB after current one
1287 if (!BI.isConditional()) { // Unconditional branch?
1288 if (BI.getSuccessor(0) != NextBB)
1289 BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0));
1293 // See if we can fold the setcc into the branch itself...
1294 SetCondInst *SCI = canFoldSetCCIntoBranchOrSelect(BI.getCondition());
1296 // Nope, cannot fold setcc into this branch. Emit a branch on a condition
1297 // computed some other way...
1298 unsigned condReg = getReg(BI.getCondition());
1299 BuildMI(BB, X86::TEST8rr, 2).addReg(condReg).addReg(condReg);
1300 if (BI.getSuccessor(1) == NextBB) {
1301 if (BI.getSuccessor(0) != NextBB)
1302 BuildMI(BB, X86::JNE, 1).addPCDisp(BI.getSuccessor(0));
1304 BuildMI(BB, X86::JE, 1).addPCDisp(BI.getSuccessor(1));
1306 if (BI.getSuccessor(0) != NextBB)
1307 BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0));
1312 unsigned OpNum = getSetCCNumber(SCI->getOpcode());
1313 MachineBasicBlock::iterator MII = BB->end();
1314 OpNum = EmitComparison(OpNum, SCI->getOperand(0), SCI->getOperand(1), BB,MII);
1316 const Type *CompTy = SCI->getOperand(0)->getType();
1317 bool isSigned = CompTy->isSigned() && getClassB(CompTy) != cFP;
1320 // LLVM -> X86 signed X86 unsigned
1321 // ----- ---------- ------------
1329 // js // Used by comparison with 0 optimization
1332 static const unsigned OpcodeTab[2][8] = {
1333 { X86::JE, X86::JNE, X86::JB, X86::JAE, X86::JA, X86::JBE, 0, 0 },
1334 { X86::JE, X86::JNE, X86::JL, X86::JGE, X86::JG, X86::JLE,
1335 X86::JS, X86::JNS },
1338 if (BI.getSuccessor(0) != NextBB) {
1339 BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(0));
1340 if (BI.getSuccessor(1) != NextBB)
1341 BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(1));
1343 // Change to the inverse condition...
1344 if (BI.getSuccessor(1) != NextBB) {
1346 BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(1));
1352 /// doCall - This emits an abstract call instruction, setting up the arguments
1353 /// and the return value as appropriate. For the actual function call itself,
1354 /// it inserts the specified CallMI instruction into the stream.
1356 void ISel::doCall(const ValueRecord &Ret, MachineInstr *CallMI,
1357 const std::vector<ValueRecord> &Args) {
1359 // Count how many bytes are to be pushed on the stack...
1360 unsigned NumBytes = 0;
1362 if (!Args.empty()) {
1363 for (unsigned i = 0, e = Args.size(); i != e; ++i)
1364 switch (getClassB(Args[i].Ty)) {
1365 case cByte: case cShort: case cInt:
1366 NumBytes += 4; break;
1368 NumBytes += 8; break;
1370 NumBytes += Args[i].Ty == Type::FloatTy ? 4 : 8;
1372 default: assert(0 && "Unknown class!");
1375 // Adjust the stack pointer for the new arguments...
1376 BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addImm(NumBytes);
1378 // Arguments go on the stack in reverse order, as specified by the ABI.
1379 unsigned ArgOffset = 0;
1380 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
1382 switch (getClassB(Args[i].Ty)) {
1385 if (Args[i].Val && isa<ConstantInt>(Args[i].Val)) {
1386 // Zero/Sign extend constant, then stuff into memory.
1387 ConstantInt *Val = cast<ConstantInt>(Args[i].Val);
1388 Val = cast<ConstantInt>(ConstantExpr::getCast(Val, Type::IntTy));
1389 addRegOffset(BuildMI(BB, X86::MOV32mi, 5), X86::ESP, ArgOffset)
1390 .addImm(Val->getRawValue() & 0xFFFFFFFF);
1392 // Promote arg to 32 bits wide into a temporary register...
1393 ArgReg = makeAnotherReg(Type::UIntTy);
1394 promote32(ArgReg, Args[i]);
1395 addRegOffset(BuildMI(BB, X86::MOV32mr, 5),
1396 X86::ESP, ArgOffset).addReg(ArgReg);
1400 if (Args[i].Val && isa<ConstantInt>(Args[i].Val)) {
1401 unsigned Val = cast<ConstantInt>(Args[i].Val)->getRawValue();
1402 addRegOffset(BuildMI(BB, X86::MOV32mi, 5),
1403 X86::ESP, ArgOffset).addImm(Val);
1405 ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg;
1406 addRegOffset(BuildMI(BB, X86::MOV32mr, 5),
1407 X86::ESP, ArgOffset).addReg(ArgReg);
1411 if (Args[i].Val && isa<ConstantInt>(Args[i].Val)) {
1412 uint64_t Val = cast<ConstantInt>(Args[i].Val)->getRawValue();
1413 addRegOffset(BuildMI(BB, X86::MOV32mi, 5),
1414 X86::ESP, ArgOffset).addImm(Val & ~0U);
1415 addRegOffset(BuildMI(BB, X86::MOV32mi, 5),
1416 X86::ESP, ArgOffset+4).addImm(Val >> 32ULL);
1418 ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg;
1419 addRegOffset(BuildMI(BB, X86::MOV32mr, 5),
1420 X86::ESP, ArgOffset).addReg(ArgReg);
1421 addRegOffset(BuildMI(BB, X86::MOV32mr, 5),
1422 X86::ESP, ArgOffset+4).addReg(ArgReg+1);
1424 ArgOffset += 4; // 8 byte entry, not 4.
1428 ArgReg = Args[i].Val ? getReg(Args[i].Val) : Args[i].Reg;
1429 if (Args[i].Ty == Type::FloatTy) {
1430 addRegOffset(BuildMI(BB, X86::FST32m, 5),
1431 X86::ESP, ArgOffset).addReg(ArgReg);
1433 assert(Args[i].Ty == Type::DoubleTy && "Unknown FP type!");
1434 addRegOffset(BuildMI(BB, X86::FST64m, 5),
1435 X86::ESP, ArgOffset).addReg(ArgReg);
1436 ArgOffset += 4; // 8 byte entry, not 4.
1440 default: assert(0 && "Unknown class!");
1445 BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addImm(0);
1448 BB->push_back(CallMI);
1450 BuildMI(BB, X86::ADJCALLSTACKUP, 1).addImm(NumBytes);
1452 // If there is a return value, scavenge the result from the location the call
1455 if (Ret.Ty != Type::VoidTy) {
1456 unsigned DestClass = getClassB(Ret.Ty);
1457 switch (DestClass) {
1461 // Integral results are in %eax, or the appropriate portion
1463 static const unsigned regRegMove[] = {
1464 X86::MOV8rr, X86::MOV16rr, X86::MOV32rr
1466 static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX };
1467 BuildMI(BB, regRegMove[DestClass], 1, Ret.Reg).addReg(AReg[DestClass]);
1470 case cFP: // Floating-point return values live in %ST(0)
1471 BuildMI(BB, X86::FpGETRESULT, 1, Ret.Reg);
1473 case cLong: // Long values are left in EDX:EAX
1474 BuildMI(BB, X86::MOV32rr, 1, Ret.Reg).addReg(X86::EAX);
1475 BuildMI(BB, X86::MOV32rr, 1, Ret.Reg+1).addReg(X86::EDX);
1477 default: assert(0 && "Unknown class!");
1483 /// visitCallInst - Push args on stack and do a procedure call instruction.
1484 void ISel::visitCallInst(CallInst &CI) {
1485 MachineInstr *TheCall;
1486 if (Function *F = CI.getCalledFunction()) {
1487 // Is it an intrinsic function call?
1488 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
1489 visitIntrinsicCall(ID, CI); // Special intrinsics are not handled here
1493 // Emit a CALL instruction with PC-relative displacement.
1494 TheCall = BuildMI(X86::CALLpcrel32, 1).addGlobalAddress(F, true);
1495 } else { // Emit an indirect call...
1496 unsigned Reg = getReg(CI.getCalledValue());
1497 TheCall = BuildMI(X86::CALL32r, 1).addReg(Reg);
1500 std::vector<ValueRecord> Args;
1501 for (unsigned i = 1, e = CI.getNumOperands(); i != e; ++i)
1502 Args.push_back(ValueRecord(CI.getOperand(i)));
1504 unsigned DestReg = CI.getType() != Type::VoidTy ? getReg(CI) : 0;
1505 doCall(ValueRecord(DestReg, CI.getType()), TheCall, Args);
1509 /// LowerUnknownIntrinsicFunctionCalls - This performs a prepass over the
1510 /// function, lowering any calls to unknown intrinsic functions into the
1511 /// equivalent LLVM code.
1513 void ISel::LowerUnknownIntrinsicFunctionCalls(Function &F) {
1514 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1515 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1516 if (CallInst *CI = dyn_cast<CallInst>(I++))
1517 if (Function *F = CI->getCalledFunction())
1518 switch (F->getIntrinsicID()) {
1519 case Intrinsic::not_intrinsic:
1520 case Intrinsic::vastart:
1521 case Intrinsic::vacopy:
1522 case Intrinsic::vaend:
1523 case Intrinsic::returnaddress:
1524 case Intrinsic::frameaddress:
1525 case Intrinsic::memcpy:
1526 case Intrinsic::memset:
1527 case Intrinsic::readport:
1528 case Intrinsic::writeport:
1529 // We directly implement these intrinsics
1532 // All other intrinsic calls we must lower.
1533 Instruction *Before = CI->getPrev();
1534 TM.getIntrinsicLowering().LowerIntrinsicCall(CI);
1535 if (Before) { // Move iterator to instruction after call
1544 void ISel::visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI) {
1545 unsigned TmpReg1, TmpReg2;
1547 case Intrinsic::vastart:
1548 // Get the address of the first vararg value...
1549 TmpReg1 = getReg(CI);
1550 addFrameReference(BuildMI(BB, X86::LEA32r, 5, TmpReg1), VarArgsFrameIndex);
1553 case Intrinsic::vacopy:
1554 TmpReg1 = getReg(CI);
1555 TmpReg2 = getReg(CI.getOperand(1));
1556 BuildMI(BB, X86::MOV32rr, 1, TmpReg1).addReg(TmpReg2);
1558 case Intrinsic::vaend: return; // Noop on X86
1560 case Intrinsic::returnaddress:
1561 case Intrinsic::frameaddress:
1562 TmpReg1 = getReg(CI);
1563 if (cast<Constant>(CI.getOperand(1))->isNullValue()) {
1564 if (ID == Intrinsic::returnaddress) {
1565 // Just load the return address
1566 addFrameReference(BuildMI(BB, X86::MOV32rm, 4, TmpReg1),
1567 ReturnAddressIndex);
1569 addFrameReference(BuildMI(BB, X86::LEA32r, 4, TmpReg1),
1570 ReturnAddressIndex, -4);
1573 // Values other than zero are not implemented yet.
1574 BuildMI(BB, X86::MOV32ri, 1, TmpReg1).addImm(0);
1578 case Intrinsic::memcpy: {
1579 assert(CI.getNumOperands() == 5 && "Illegal llvm.memcpy call!");
1581 if (ConstantInt *AlignC = dyn_cast<ConstantInt>(CI.getOperand(4))) {
1582 Align = AlignC->getRawValue();
1583 if (Align == 0) Align = 1;
1586 // Turn the byte code into # iterations
1589 switch (Align & 3) {
1590 case 2: // WORD aligned
1591 if (ConstantInt *I = dyn_cast<ConstantInt>(CI.getOperand(3))) {
1592 CountReg = getReg(ConstantUInt::get(Type::UIntTy, I->getRawValue()/2));
1594 CountReg = makeAnotherReg(Type::IntTy);
1595 unsigned ByteReg = getReg(CI.getOperand(3));
1596 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
1598 Opcode = X86::REP_MOVSW;
1600 case 0: // DWORD aligned
1601 if (ConstantInt *I = dyn_cast<ConstantInt>(CI.getOperand(3))) {
1602 CountReg = getReg(ConstantUInt::get(Type::UIntTy, I->getRawValue()/4));
1604 CountReg = makeAnotherReg(Type::IntTy);
1605 unsigned ByteReg = getReg(CI.getOperand(3));
1606 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
1608 Opcode = X86::REP_MOVSD;
1610 default: // BYTE aligned
1611 CountReg = getReg(CI.getOperand(3));
1612 Opcode = X86::REP_MOVSB;
1616 // No matter what the alignment is, we put the source in ESI, the
1617 // destination in EDI, and the count in ECX.
1618 TmpReg1 = getReg(CI.getOperand(1));
1619 TmpReg2 = getReg(CI.getOperand(2));
1620 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
1621 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
1622 BuildMI(BB, X86::MOV32rr, 1, X86::ESI).addReg(TmpReg2);
1623 BuildMI(BB, Opcode, 0);
1626 case Intrinsic::memset: {
1627 assert(CI.getNumOperands() == 5 && "Illegal llvm.memset call!");
1629 if (ConstantInt *AlignC = dyn_cast<ConstantInt>(CI.getOperand(4))) {
1630 Align = AlignC->getRawValue();
1631 if (Align == 0) Align = 1;
1634 // Turn the byte code into # iterations
1637 if (ConstantInt *ValC = dyn_cast<ConstantInt>(CI.getOperand(2))) {
1638 unsigned Val = ValC->getRawValue() & 255;
1640 // If the value is a constant, then we can potentially use larger copies.
1641 switch (Align & 3) {
1642 case 2: // WORD aligned
1643 if (ConstantInt *I = dyn_cast<ConstantInt>(CI.getOperand(3))) {
1644 CountReg =getReg(ConstantUInt::get(Type::UIntTy, I->getRawValue()/2));
1646 CountReg = makeAnotherReg(Type::IntTy);
1647 unsigned ByteReg = getReg(CI.getOperand(3));
1648 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(1);
1650 BuildMI(BB, X86::MOV16ri, 1, X86::AX).addImm((Val << 8) | Val);
1651 Opcode = X86::REP_STOSW;
1653 case 0: // DWORD aligned
1654 if (ConstantInt *I = dyn_cast<ConstantInt>(CI.getOperand(3))) {
1655 CountReg =getReg(ConstantUInt::get(Type::UIntTy, I->getRawValue()/4));
1657 CountReg = makeAnotherReg(Type::IntTy);
1658 unsigned ByteReg = getReg(CI.getOperand(3));
1659 BuildMI(BB, X86::SHR32ri, 2, CountReg).addReg(ByteReg).addImm(2);
1661 Val = (Val << 8) | Val;
1662 BuildMI(BB, X86::MOV32ri, 1, X86::EAX).addImm((Val << 16) | Val);
1663 Opcode = X86::REP_STOSD;
1665 default: // BYTE aligned
1666 CountReg = getReg(CI.getOperand(3));
1667 BuildMI(BB, X86::MOV8ri, 1, X86::AL).addImm(Val);
1668 Opcode = X86::REP_STOSB;
1672 // If it's not a constant value we are storing, just fall back. We could
1673 // try to be clever to form 16 bit and 32 bit values, but we don't yet.
1674 unsigned ValReg = getReg(CI.getOperand(2));
1675 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(ValReg);
1676 CountReg = getReg(CI.getOperand(3));
1677 Opcode = X86::REP_STOSB;
1680 // No matter what the alignment is, we put the source in ESI, the
1681 // destination in EDI, and the count in ECX.
1682 TmpReg1 = getReg(CI.getOperand(1));
1683 //TmpReg2 = getReg(CI.getOperand(2));
1684 BuildMI(BB, X86::MOV32rr, 1, X86::ECX).addReg(CountReg);
1685 BuildMI(BB, X86::MOV32rr, 1, X86::EDI).addReg(TmpReg1);
1686 BuildMI(BB, Opcode, 0);
1690 case Intrinsic::readport:
1692 // First, determine that the size of the operand falls within the
1693 // acceptable range for this architecture.
1695 if ((CI.getOperand(1)->getType()->getPrimitiveSize()) != 2) {
1696 std::cerr << "llvm.readport: Address size is not 16 bits\n";
1701 // Now, move the I/O port address into the DX register and use the IN
1702 // instruction to get the input data.
1704 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(getReg(CI.getOperand(1)));
1705 switch (CI.getCalledFunction()->getReturnType()->getPrimitiveSize()) {
1707 BuildMI(BB, X86::IN8, 0);
1710 BuildMI(BB, X86::IN16, 0);
1713 BuildMI(BB, X86::IN32, 0);
1716 std::cerr << "Cannot do input on this data type";
1721 case Intrinsic::writeport:
1723 // First, determine that the size of the operand falls within the
1724 // acceptable range for this architecture.
1727 if ((CI.getOperand(2)->getType()->getPrimitiveSize()) != 2) {
1728 std::cerr << "llvm.writeport: Address size is not 16 bits\n";
1733 // Now, move the I/O port address into the DX register and the value to
1734 // write into the AL/AX/EAX register.
1736 BuildMI(BB, X86::MOV16rr, 1, X86::DX).addReg(getReg(CI.getOperand(2)));
1737 switch (CI.getOperand(1)->getType()->getPrimitiveSize()) {
1739 BuildMI(BB, X86::MOV8rr, 1, X86::AL).addReg(getReg(CI.getOperand(1)));
1740 BuildMI(BB, X86::OUT8, 0);
1743 BuildMI(BB, X86::MOV16rr, 1, X86::AX).addReg(getReg(CI.getOperand(1)));
1744 BuildMI(BB, X86::OUT16, 0);
1747 BuildMI(BB, X86::MOV32rr, 1, X86::EAX).addReg(getReg(CI.getOperand(1)));
1748 BuildMI(BB, X86::OUT32, 0);
1751 std::cerr << "Cannot do output on this data type";
1756 default: assert(0 && "Error: unknown intrinsics should have been lowered!");
1760 static bool isSafeToFoldLoadIntoInstruction(LoadInst &LI, Instruction &User) {
1761 if (LI.getParent() != User.getParent())
1763 BasicBlock::iterator It = &LI;
1764 // Check all of the instructions between the load and the user. We should
1765 // really use alias analysis here, but for now we just do something simple.
1766 for (++It; It != BasicBlock::iterator(&User); ++It) {
1767 switch (It->getOpcode()) {
1768 case Instruction::Free:
1769 case Instruction::Store:
1770 case Instruction::Call:
1771 case Instruction::Invoke:
1778 /// visitSimpleBinary - Implement simple binary operators for integral types...
1779 /// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or, 4 for
1782 void ISel::visitSimpleBinary(BinaryOperator &B, unsigned OperatorClass) {
1783 unsigned DestReg = getReg(B);
1784 MachineBasicBlock::iterator MI = BB->end();
1785 Value *Op0 = B.getOperand(0), *Op1 = B.getOperand(1);
1787 // Special case: op Reg, load [mem]
1788 if (isa<LoadInst>(Op0) && !isa<LoadInst>(Op1))
1789 if (!B.swapOperands())
1790 std::swap(Op0, Op1); // Make sure any loads are in the RHS.
1792 unsigned Class = getClassB(B.getType());
1793 if (isa<LoadInst>(Op1) && Class != cLong &&
1794 isSafeToFoldLoadIntoInstruction(*cast<LoadInst>(Op1), B)) {
1798 static const unsigned OpcodeTab[][3] = {
1799 // Arithmetic operators
1800 { X86::ADD8rm, X86::ADD16rm, X86::ADD32rm }, // ADD
1801 { X86::SUB8rm, X86::SUB16rm, X86::SUB32rm }, // SUB
1803 // Bitwise operators
1804 { X86::AND8rm, X86::AND16rm, X86::AND32rm }, // AND
1805 { X86:: OR8rm, X86:: OR16rm, X86:: OR32rm }, // OR
1806 { X86::XOR8rm, X86::XOR16rm, X86::XOR32rm }, // XOR
1808 Opcode = OpcodeTab[OperatorClass][Class];
1810 static const unsigned OpcodeTab[][2] = {
1811 { X86::FADD32m, X86::FADD64m }, // ADD
1812 { X86::FSUB32m, X86::FSUB64m }, // SUB
1814 const Type *Ty = Op0->getType();
1815 assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
1816 Opcode = OpcodeTab[OperatorClass][Ty == Type::DoubleTy];
1819 unsigned BaseReg, Scale, IndexReg, Disp;
1820 getAddressingMode(cast<LoadInst>(Op1)->getOperand(0), BaseReg,
1821 Scale, IndexReg, Disp);
1823 unsigned Op0r = getReg(Op0);
1824 addFullAddress(BuildMI(BB, Opcode, 2, DestReg).addReg(Op0r),
1825 BaseReg, Scale, IndexReg, Disp);
1829 // If this is a floating point subtract, check to see if we can fold the first
1831 if (Class == cFP && OperatorClass == 1 &&
1832 isa<LoadInst>(Op0) &&
1833 isSafeToFoldLoadIntoInstruction(*cast<LoadInst>(Op0), B)) {
1834 const Type *Ty = Op0->getType();
1835 assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
1836 unsigned Opcode = Ty == Type::FloatTy ? X86::FSUBR32m : X86::FSUBR64m;
1838 unsigned BaseReg, Scale, IndexReg, Disp;
1839 getAddressingMode(cast<LoadInst>(Op0)->getOperand(0), BaseReg,
1840 Scale, IndexReg, Disp);
1842 unsigned Op1r = getReg(Op1);
1843 addFullAddress(BuildMI(BB, Opcode, 2, DestReg).addReg(Op1r),
1844 BaseReg, Scale, IndexReg, Disp);
1848 emitSimpleBinaryOperation(BB, MI, Op0, Op1, OperatorClass, DestReg);
1852 /// emitBinaryFPOperation - This method handles emission of floating point
1853 /// Add (0), Sub (1), Mul (2), and Div (3) operations.
1854 void ISel::emitBinaryFPOperation(MachineBasicBlock *BB,
1855 MachineBasicBlock::iterator IP,
1856 Value *Op0, Value *Op1,
1857 unsigned OperatorClass, unsigned DestReg) {
1859 // Special case: op Reg, <const fp>
1860 if (ConstantFP *Op1C = dyn_cast<ConstantFP>(Op1))
1861 if (!Op1C->isExactlyValue(+0.0) && !Op1C->isExactlyValue(+1.0)) {
1862 // Create a constant pool entry for this constant.
1863 MachineConstantPool *CP = F->getConstantPool();
1864 unsigned CPI = CP->getConstantPoolIndex(Op1C);
1865 const Type *Ty = Op1->getType();
1867 static const unsigned OpcodeTab[][4] = {
1868 { X86::FADD32m, X86::FSUB32m, X86::FMUL32m, X86::FDIV32m }, // Float
1869 { X86::FADD64m, X86::FSUB64m, X86::FMUL64m, X86::FDIV64m }, // Double
1872 assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
1873 unsigned Opcode = OpcodeTab[Ty != Type::FloatTy][OperatorClass];
1874 unsigned Op0r = getReg(Op0, BB, IP);
1875 addConstantPoolReference(BuildMI(*BB, IP, Opcode, 5,
1876 DestReg).addReg(Op0r), CPI);
1880 // Special case: R1 = op <const fp>, R2
1881 if (ConstantFP *CFP = dyn_cast<ConstantFP>(Op0))
1882 if (CFP->isExactlyValue(-0.0) && OperatorClass == 1) {
1884 unsigned op1Reg = getReg(Op1, BB, IP);
1885 BuildMI(*BB, IP, X86::FCHS, 1, DestReg).addReg(op1Reg);
1887 } else if (!CFP->isExactlyValue(+0.0) && !CFP->isExactlyValue(+1.0)) {
1888 // R1 = op CST, R2 --> R1 = opr R2, CST
1890 // Create a constant pool entry for this constant.
1891 MachineConstantPool *CP = F->getConstantPool();
1892 unsigned CPI = CP->getConstantPoolIndex(CFP);
1893 const Type *Ty = CFP->getType();
1895 static const unsigned OpcodeTab[][4] = {
1896 { X86::FADD32m, X86::FSUBR32m, X86::FMUL32m, X86::FDIVR32m }, // Float
1897 { X86::FADD64m, X86::FSUBR64m, X86::FMUL64m, X86::FDIVR64m }, // Double
1900 assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
1901 unsigned Opcode = OpcodeTab[Ty != Type::FloatTy][OperatorClass];
1902 unsigned Op1r = getReg(Op1, BB, IP);
1903 addConstantPoolReference(BuildMI(*BB, IP, Opcode, 5,
1904 DestReg).addReg(Op1r), CPI);
1909 static const unsigned OpcodeTab[4] = {
1910 X86::FpADD, X86::FpSUB, X86::FpMUL, X86::FpDIV
1913 unsigned Opcode = OpcodeTab[OperatorClass];
1914 unsigned Op0r = getReg(Op0, BB, IP);
1915 unsigned Op1r = getReg(Op1, BB, IP);
1916 BuildMI(*BB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r);
1919 /// emitSimpleBinaryOperation - Implement simple binary operators for integral
1920 /// types... OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for
1923 /// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary
1924 /// and constant expression support.
1926 void ISel::emitSimpleBinaryOperation(MachineBasicBlock *MBB,
1927 MachineBasicBlock::iterator IP,
1928 Value *Op0, Value *Op1,
1929 unsigned OperatorClass, unsigned DestReg) {
1930 unsigned Class = getClassB(Op0->getType());
1933 assert(OperatorClass < 2 && "No logical ops for FP!");
1934 emitBinaryFPOperation(MBB, IP, Op0, Op1, OperatorClass, DestReg);
1938 // sub 0, X -> neg X
1939 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op0))
1940 if (OperatorClass == 1 && CI->isNullValue()) {
1941 unsigned op1Reg = getReg(Op1, MBB, IP);
1942 static unsigned const NEGTab[] = {
1943 X86::NEG8r, X86::NEG16r, X86::NEG32r, 0, X86::NEG32r
1945 BuildMI(*MBB, IP, NEGTab[Class], 1, DestReg).addReg(op1Reg);
1947 if (Class == cLong) {
1948 // We just emitted: Dl = neg Sl
1949 // Now emit : T = addc Sh, 0
1951 unsigned T = makeAnotherReg(Type::IntTy);
1952 BuildMI(*MBB, IP, X86::ADC32ri, 2, T).addReg(op1Reg+1).addImm(0);
1953 BuildMI(*MBB, IP, X86::NEG32r, 1, DestReg+1).addReg(T);
1958 // Special case: op Reg, <const int>
1959 if (ConstantInt *Op1C = dyn_cast<ConstantInt>(Op1)) {
1960 unsigned Op0r = getReg(Op0, MBB, IP);
1962 // xor X, -1 -> not X
1963 if (OperatorClass == 4 && Op1C->isAllOnesValue()) {
1964 static unsigned const NOTTab[] = {
1965 X86::NOT8r, X86::NOT16r, X86::NOT32r, 0, X86::NOT32r
1967 BuildMI(*MBB, IP, NOTTab[Class], 1, DestReg).addReg(Op0r);
1968 if (Class == cLong) // Invert the top part too
1969 BuildMI(*MBB, IP, X86::NOT32r, 1, DestReg+1).addReg(Op0r+1);
1973 // add X, -1 -> dec X
1974 if (OperatorClass == 0 && Op1C->isAllOnesValue() && Class != cLong) {
1975 // Note that we can't use dec for 64-bit decrements, because it does not
1976 // set the carry flag!
1977 static unsigned const DECTab[] = { X86::DEC8r, X86::DEC16r, X86::DEC32r };
1978 BuildMI(*MBB, IP, DECTab[Class], 1, DestReg).addReg(Op0r);
1982 // add X, 1 -> inc X
1983 if (OperatorClass == 0 && Op1C->equalsInt(1) && Class != cLong) {
1984 // Note that we can't use inc for 64-bit increments, because it does not
1985 // set the carry flag!
1986 static unsigned const INCTab[] = { X86::INC8r, X86::INC16r, X86::INC32r };
1987 BuildMI(*MBB, IP, INCTab[Class], 1, DestReg).addReg(Op0r);
1991 static const unsigned OpcodeTab[][5] = {
1992 // Arithmetic operators
1993 { X86::ADD8ri, X86::ADD16ri, X86::ADD32ri, 0, X86::ADD32ri }, // ADD
1994 { X86::SUB8ri, X86::SUB16ri, X86::SUB32ri, 0, X86::SUB32ri }, // SUB
1996 // Bitwise operators
1997 { X86::AND8ri, X86::AND16ri, X86::AND32ri, 0, X86::AND32ri }, // AND
1998 { X86:: OR8ri, X86:: OR16ri, X86:: OR32ri, 0, X86::OR32ri }, // OR
1999 { X86::XOR8ri, X86::XOR16ri, X86::XOR32ri, 0, X86::XOR32ri }, // XOR
2002 unsigned Opcode = OpcodeTab[OperatorClass][Class];
2003 unsigned Op1l = cast<ConstantInt>(Op1C)->getRawValue();
2005 if (Class != cLong) {
2006 BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addImm(Op1l);
2010 // If this is a long value and the high or low bits have a special
2011 // property, emit some special cases.
2012 unsigned Op1h = cast<ConstantInt>(Op1C)->getRawValue() >> 32LL;
2014 // If the constant is zero in the low 32-bits, just copy the low part
2015 // across and apply the normal 32-bit operation to the high parts. There
2016 // will be no carry or borrow into the top.
2018 if (OperatorClass != 2) // All but and...
2019 BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg).addReg(Op0r);
2021 BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
2022 BuildMI(*MBB, IP, OpcodeTab[OperatorClass][cLong], 2, DestReg+1)
2023 .addReg(Op0r+1).addImm(Op1h);
2027 // If this is a logical operation and the top 32-bits are zero, just
2028 // operate on the lower 32.
2029 if (Op1h == 0 && OperatorClass > 1) {
2030 BuildMI(*MBB, IP, OpcodeTab[OperatorClass][cLong], 2, DestReg)
2031 .addReg(Op0r).addImm(Op1l);
2032 if (OperatorClass != 2) // All but and
2033 BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg+1).addReg(Op0r+1);
2035 BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg+1).addImm(0);
2039 // TODO: We could handle lots of other special cases here, such as AND'ing
2040 // with 0xFFFFFFFF00000000 -> noop, etc.
2042 // Otherwise, code generate the full operation with a constant.
2043 static const unsigned TopTab[] = {
2044 X86::ADC32ri, X86::SBB32ri, X86::AND32ri, X86::OR32ri, X86::XOR32ri
2047 BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addImm(Op1l);
2048 BuildMI(*MBB, IP, TopTab[OperatorClass], 2, DestReg+1)
2049 .addReg(Op0r+1).addImm(Op1h);
2053 // Finally, handle the general case now.
2054 static const unsigned OpcodeTab[][5] = {
2055 // Arithmetic operators
2056 { X86::ADD8rr, X86::ADD16rr, X86::ADD32rr, 0, X86::ADD32rr }, // ADD
2057 { X86::SUB8rr, X86::SUB16rr, X86::SUB32rr, 0, X86::SUB32rr }, // SUB
2059 // Bitwise operators
2060 { X86::AND8rr, X86::AND16rr, X86::AND32rr, 0, X86::AND32rr }, // AND
2061 { X86:: OR8rr, X86:: OR16rr, X86:: OR32rr, 0, X86:: OR32rr }, // OR
2062 { X86::XOR8rr, X86::XOR16rr, X86::XOR32rr, 0, X86::XOR32rr }, // XOR
2065 unsigned Opcode = OpcodeTab[OperatorClass][Class];
2066 unsigned Op0r = getReg(Op0, MBB, IP);
2067 unsigned Op1r = getReg(Op1, MBB, IP);
2068 BuildMI(*MBB, IP, Opcode, 2, DestReg).addReg(Op0r).addReg(Op1r);
2070 if (Class == cLong) { // Handle the upper 32 bits of long values...
2071 static const unsigned TopTab[] = {
2072 X86::ADC32rr, X86::SBB32rr, X86::AND32rr, X86::OR32rr, X86::XOR32rr
2074 BuildMI(*MBB, IP, TopTab[OperatorClass], 2,
2075 DestReg+1).addReg(Op0r+1).addReg(Op1r+1);
2079 /// doMultiply - Emit appropriate instructions to multiply together the
2080 /// registers op0Reg and op1Reg, and put the result in DestReg. The type of the
2081 /// result should be given as DestTy.
2083 void ISel::doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator MBBI,
2084 unsigned DestReg, const Type *DestTy,
2085 unsigned op0Reg, unsigned op1Reg) {
2086 unsigned Class = getClass(DestTy);
2090 BuildMI(*MBB, MBBI, Class == cInt ? X86::IMUL32rr:X86::IMUL16rr, 2, DestReg)
2091 .addReg(op0Reg).addReg(op1Reg);
2094 // Must use the MUL instruction, which forces use of AL...
2095 BuildMI(*MBB, MBBI, X86::MOV8rr, 1, X86::AL).addReg(op0Reg);
2096 BuildMI(*MBB, MBBI, X86::MUL8r, 1).addReg(op1Reg);
2097 BuildMI(*MBB, MBBI, X86::MOV8rr, 1, DestReg).addReg(X86::AL);
2100 case cLong: assert(0 && "doMultiply cannot operate on LONG values!");
2104 // ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
2105 // returns zero when the input is not exactly a power of two.
2106 static unsigned ExactLog2(unsigned Val) {
2107 if (Val == 0) return 0;
2110 if (Val & 1) return 0;
2118 /// doMultiplyConst - This function is specialized to efficiently codegen an 8,
2119 /// 16, or 32-bit integer multiply by a constant.
2120 void ISel::doMultiplyConst(MachineBasicBlock *MBB,
2121 MachineBasicBlock::iterator IP,
2122 unsigned DestReg, const Type *DestTy,
2123 unsigned op0Reg, unsigned ConstRHS) {
2124 static const unsigned MOVrrTab[] = {X86::MOV8rr, X86::MOV16rr, X86::MOV32rr};
2125 static const unsigned MOVriTab[] = {X86::MOV8ri, X86::MOV16ri, X86::MOV32ri};
2127 unsigned Class = getClass(DestTy);
2129 if (ConstRHS == 0) {
2130 BuildMI(*MBB, IP, MOVriTab[Class], 1, DestReg).addImm(0);
2132 } else if (ConstRHS == 1) {
2133 BuildMI(*MBB, IP, MOVrrTab[Class], 1, DestReg).addReg(op0Reg);
2137 // If the element size is exactly a power of 2, use a shift to get it.
2138 if (unsigned Shift = ExactLog2(ConstRHS)) {
2140 default: assert(0 && "Unknown class for this function!");
2142 BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
2145 BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
2148 BuildMI(*MBB, IP, X86::SHL32ri,2, DestReg).addReg(op0Reg).addImm(Shift-1);
2153 if (Class == cShort) {
2154 BuildMI(*MBB, IP, X86::IMUL16rri,2,DestReg).addReg(op0Reg).addImm(ConstRHS);
2156 } else if (Class == cInt) {
2157 BuildMI(*MBB, IP, X86::IMUL32rri,2,DestReg).addReg(op0Reg).addImm(ConstRHS);
2161 // Most general case, emit a normal multiply...
2162 unsigned TmpReg = makeAnotherReg(DestTy);
2163 BuildMI(*MBB, IP, MOVriTab[Class], 1, TmpReg).addImm(ConstRHS);
2165 // Emit a MUL to multiply the register holding the index by
2166 // elementSize, putting the result in OffsetReg.
2167 doMultiply(MBB, IP, DestReg, DestTy, op0Reg, TmpReg);
2170 /// visitMul - Multiplies are not simple binary operators because they must deal
2171 /// with the EAX register explicitly.
2173 void ISel::visitMul(BinaryOperator &I) {
2174 unsigned ResultReg = getReg(I);
2176 Value *Op0 = I.getOperand(0);
2177 Value *Op1 = I.getOperand(1);
2179 // Fold loads into floating point multiplies.
2180 if (getClass(Op0->getType()) == cFP) {
2181 if (isa<LoadInst>(Op0) && !isa<LoadInst>(Op1))
2182 if (!I.swapOperands())
2183 std::swap(Op0, Op1); // Make sure any loads are in the RHS.
2184 if (LoadInst *LI = dyn_cast<LoadInst>(Op1))
2185 if (isSafeToFoldLoadIntoInstruction(*LI, I)) {
2186 const Type *Ty = Op0->getType();
2187 assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
2188 unsigned Opcode = Ty == Type::FloatTy ? X86::FMUL32m : X86::FMUL64m;
2190 unsigned BaseReg, Scale, IndexReg, Disp;
2191 getAddressingMode(LI->getOperand(0), BaseReg,
2192 Scale, IndexReg, Disp);
2194 unsigned Op0r = getReg(Op0);
2195 addFullAddress(BuildMI(BB, Opcode, 2, ResultReg).addReg(Op0r),
2196 BaseReg, Scale, IndexReg, Disp);
2201 MachineBasicBlock::iterator IP = BB->end();
2202 emitMultiply(BB, IP, Op0, Op1, ResultReg);
2205 void ISel::emitMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator IP,
2206 Value *Op0, Value *Op1, unsigned DestReg) {
2207 MachineBasicBlock &BB = *MBB;
2208 TypeClass Class = getClass(Op0->getType());
2210 // Simple scalar multiply?
2211 unsigned Op0Reg = getReg(Op0, &BB, IP);
2216 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2217 unsigned Val = (unsigned)CI->getRawValue(); // Isn't a 64-bit constant
2218 doMultiplyConst(&BB, IP, DestReg, Op0->getType(), Op0Reg, Val);
2220 unsigned Op1Reg = getReg(Op1, &BB, IP);
2221 doMultiply(&BB, IP, DestReg, Op1->getType(), Op0Reg, Op1Reg);
2225 emitBinaryFPOperation(MBB, IP, Op0, Op1, 2, DestReg);
2231 // Long value. We have to do things the hard way...
2232 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
2233 unsigned CLow = CI->getRawValue();
2234 unsigned CHi = CI->getRawValue() >> 32;
2237 // If the low part of the constant is all zeros, things are simple.
2238 BuildMI(BB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
2239 doMultiplyConst(&BB, IP, DestReg+1, Type::UIntTy, Op0Reg, CHi);
2243 // Multiply the two low parts... capturing carry into EDX
2244 unsigned OverflowReg = 0;
2246 BuildMI(BB, IP, X86::MOV32rr, 1, DestReg).addReg(Op0Reg);
2248 unsigned Op1RegL = makeAnotherReg(Type::UIntTy);
2249 OverflowReg = makeAnotherReg(Type::UIntTy);
2250 BuildMI(BB, IP, X86::MOV32ri, 1, Op1RegL).addImm(CLow);
2251 BuildMI(BB, IP, X86::MOV32rr, 1, X86::EAX).addReg(Op0Reg);
2252 BuildMI(BB, IP, X86::MUL32r, 1).addReg(Op1RegL); // AL*BL
2254 BuildMI(BB, IP, X86::MOV32rr, 1, DestReg).addReg(X86::EAX); // AL*BL
2255 BuildMI(BB, IP, X86::MOV32rr, 1,
2256 OverflowReg).addReg(X86::EDX); // AL*BL >> 32
2259 unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL
2260 doMultiplyConst(&BB, IP, AHBLReg, Type::UIntTy, Op0Reg+1, CLow);
2262 unsigned AHBLplusOverflowReg;
2264 AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy);
2265 BuildMI(BB, IP, X86::ADD32rr, 2, // AH*BL+(AL*BL >> 32)
2266 AHBLplusOverflowReg).addReg(AHBLReg).addReg(OverflowReg);
2268 AHBLplusOverflowReg = AHBLReg;
2272 BuildMI(BB, IP, X86::MOV32rr, 1, DestReg+1).addReg(AHBLplusOverflowReg);
2274 unsigned ALBHReg = makeAnotherReg(Type::UIntTy); // AL*BH
2275 doMultiplyConst(&BB, IP, ALBHReg, Type::UIntTy, Op0Reg, CHi);
2277 BuildMI(BB, IP, X86::ADD32rr, 2, // AL*BH + AH*BL + (AL*BL >> 32)
2278 DestReg+1).addReg(AHBLplusOverflowReg).addReg(ALBHReg);
2283 // General 64x64 multiply
2285 unsigned Op1Reg = getReg(Op1, &BB, IP);
2286 // Multiply the two low parts... capturing carry into EDX
2287 BuildMI(BB, IP, X86::MOV32rr, 1, X86::EAX).addReg(Op0Reg);
2288 BuildMI(BB, IP, X86::MUL32r, 1).addReg(Op1Reg); // AL*BL
2290 unsigned OverflowReg = makeAnotherReg(Type::UIntTy);
2291 BuildMI(BB, IP, X86::MOV32rr, 1, DestReg).addReg(X86::EAX); // AL*BL
2292 BuildMI(BB, IP, X86::MOV32rr, 1,
2293 OverflowReg).addReg(X86::EDX); // AL*BL >> 32
2295 unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL
2296 BuildMI(BB, IP, X86::IMUL32rr, 2,
2297 AHBLReg).addReg(Op0Reg+1).addReg(Op1Reg);
2299 unsigned AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy);
2300 BuildMI(BB, IP, X86::ADD32rr, 2, // AH*BL+(AL*BL >> 32)
2301 AHBLplusOverflowReg).addReg(AHBLReg).addReg(OverflowReg);
2303 unsigned ALBHReg = makeAnotherReg(Type::UIntTy); // AL*BH
2304 BuildMI(BB, IP, X86::IMUL32rr, 2,
2305 ALBHReg).addReg(Op0Reg).addReg(Op1Reg+1);
2307 BuildMI(BB, IP, X86::ADD32rr, 2, // AL*BH + AH*BL + (AL*BL >> 32)
2308 DestReg+1).addReg(AHBLplusOverflowReg).addReg(ALBHReg);
2312 /// visitDivRem - Handle division and remainder instructions... these
2313 /// instruction both require the same instructions to be generated, they just
2314 /// select the result from a different register. Note that both of these
2315 /// instructions work differently for signed and unsigned operands.
2317 void ISel::visitDivRem(BinaryOperator &I) {
2318 unsigned ResultReg = getReg(I);
2319 Value *Op0 = I.getOperand(0), *Op1 = I.getOperand(1);
2321 // Fold loads into floating point divides.
2322 if (getClass(Op0->getType()) == cFP) {
2323 if (LoadInst *LI = dyn_cast<LoadInst>(Op1))
2324 if (isSafeToFoldLoadIntoInstruction(*LI, I)) {
2325 const Type *Ty = Op0->getType();
2326 assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
2327 unsigned Opcode = Ty == Type::FloatTy ? X86::FDIV32m : X86::FDIV64m;
2329 unsigned BaseReg, Scale, IndexReg, Disp;
2330 getAddressingMode(LI->getOperand(0), BaseReg,
2331 Scale, IndexReg, Disp);
2333 unsigned Op0r = getReg(Op0);
2334 addFullAddress(BuildMI(BB, Opcode, 2, ResultReg).addReg(Op0r),
2335 BaseReg, Scale, IndexReg, Disp);
2339 if (LoadInst *LI = dyn_cast<LoadInst>(Op0))
2340 if (isSafeToFoldLoadIntoInstruction(*LI, I)) {
2341 const Type *Ty = Op0->getType();
2342 assert(Ty == Type::FloatTy||Ty == Type::DoubleTy && "Unknown FP type!");
2343 unsigned Opcode = Ty == Type::FloatTy ? X86::FDIVR32m : X86::FDIVR64m;
2345 unsigned BaseReg, Scale, IndexReg, Disp;
2346 getAddressingMode(LI->getOperand(0), BaseReg,
2347 Scale, IndexReg, Disp);
2349 unsigned Op1r = getReg(Op1);
2350 addFullAddress(BuildMI(BB, Opcode, 2, ResultReg).addReg(Op1r),
2351 BaseReg, Scale, IndexReg, Disp);
2357 MachineBasicBlock::iterator IP = BB->end();
2358 emitDivRemOperation(BB, IP, Op0, Op1,
2359 I.getOpcode() == Instruction::Div, ResultReg);
2362 void ISel::emitDivRemOperation(MachineBasicBlock *BB,
2363 MachineBasicBlock::iterator IP,
2364 Value *Op0, Value *Op1, bool isDiv,
2365 unsigned ResultReg) {
2366 const Type *Ty = Op0->getType();
2367 unsigned Class = getClass(Ty);
2369 case cFP: // Floating point divide
2371 emitBinaryFPOperation(BB, IP, Op0, Op1, 3, ResultReg);
2373 } else { // Floating point remainder...
2374 unsigned Op0Reg = getReg(Op0, BB, IP);
2375 unsigned Op1Reg = getReg(Op1, BB, IP);
2376 MachineInstr *TheCall =
2377 BuildMI(X86::CALLpcrel32, 1).addExternalSymbol("fmod", true);
2378 std::vector<ValueRecord> Args;
2379 Args.push_back(ValueRecord(Op0Reg, Type::DoubleTy));
2380 Args.push_back(ValueRecord(Op1Reg, Type::DoubleTy));
2381 doCall(ValueRecord(ResultReg, Type::DoubleTy), TheCall, Args);
2385 static const char *FnName[] =
2386 { "__moddi3", "__divdi3", "__umoddi3", "__udivdi3" };
2387 unsigned Op0Reg = getReg(Op0, BB, IP);
2388 unsigned Op1Reg = getReg(Op1, BB, IP);
2389 unsigned NameIdx = Ty->isUnsigned()*2 + isDiv;
2390 MachineInstr *TheCall =
2391 BuildMI(X86::CALLpcrel32, 1).addExternalSymbol(FnName[NameIdx], true);
2393 std::vector<ValueRecord> Args;
2394 Args.push_back(ValueRecord(Op0Reg, Type::LongTy));
2395 Args.push_back(ValueRecord(Op1Reg, Type::LongTy));
2396 doCall(ValueRecord(ResultReg, Type::LongTy), TheCall, Args);
2399 case cByte: case cShort: case cInt:
2400 break; // Small integrals, handled below...
2401 default: assert(0 && "Unknown class!");
2404 static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX };
2405 static const unsigned MovOpcode[]={ X86::MOV8rr, X86::MOV16rr, X86::MOV32rr };
2406 static const unsigned SarOpcode[]={ X86::SAR8ri, X86::SAR16ri, X86::SAR32ri };
2407 static const unsigned ClrOpcode[]={ X86::MOV8ri, X86::MOV16ri, X86::MOV32ri };
2408 static const unsigned ExtRegs[] ={ X86::AH , X86::DX , X86::EDX };
2410 static const unsigned DivOpcode[][4] = {
2411 { X86::DIV8r , X86::DIV16r , X86::DIV32r , 0 }, // Unsigned division
2412 { X86::IDIV8r, X86::IDIV16r, X86::IDIV32r, 0 }, // Signed division
2415 bool isSigned = Ty->isSigned();
2416 unsigned Reg = Regs[Class];
2417 unsigned ExtReg = ExtRegs[Class];
2419 // Put the first operand into one of the A registers...
2420 unsigned Op0Reg = getReg(Op0, BB, IP);
2421 unsigned Op1Reg = getReg(Op1, BB, IP);
2422 BuildMI(*BB, IP, MovOpcode[Class], 1, Reg).addReg(Op0Reg);
2425 // Emit a sign extension instruction...
2426 unsigned ShiftResult = makeAnotherReg(Op0->getType());
2427 BuildMI(*BB, IP, SarOpcode[Class], 2,ShiftResult).addReg(Op0Reg).addImm(31);
2428 BuildMI(*BB, IP, MovOpcode[Class], 1, ExtReg).addReg(ShiftResult);
2430 // If unsigned, emit a zeroing instruction... (reg = 0)
2431 BuildMI(*BB, IP, ClrOpcode[Class], 2, ExtReg).addImm(0);
2434 // Emit the appropriate divide or remainder instruction...
2435 BuildMI(*BB, IP, DivOpcode[isSigned][Class], 1).addReg(Op1Reg);
2437 // Figure out which register we want to pick the result out of...
2438 unsigned DestReg = isDiv ? Reg : ExtReg;
2440 // Put the result into the destination register...
2441 BuildMI(*BB, IP, MovOpcode[Class], 1, ResultReg).addReg(DestReg);
2445 /// Shift instructions: 'shl', 'sar', 'shr' - Some special cases here
2446 /// for constant immediate shift values, and for constant immediate
2447 /// shift values equal to 1. Even the general case is sort of special,
2448 /// because the shift amount has to be in CL, not just any old register.
2450 void ISel::visitShiftInst(ShiftInst &I) {
2451 MachineBasicBlock::iterator IP = BB->end ();
2452 emitShiftOperation (BB, IP, I.getOperand (0), I.getOperand (1),
2453 I.getOpcode () == Instruction::Shl, I.getType (),
2457 /// emitShiftOperation - Common code shared between visitShiftInst and
2458 /// constant expression support.
2459 void ISel::emitShiftOperation(MachineBasicBlock *MBB,
2460 MachineBasicBlock::iterator IP,
2461 Value *Op, Value *ShiftAmount, bool isLeftShift,
2462 const Type *ResultTy, unsigned DestReg) {
2463 unsigned SrcReg = getReg (Op, MBB, IP);
2464 bool isSigned = ResultTy->isSigned ();
2465 unsigned Class = getClass (ResultTy);
2467 static const unsigned ConstantOperand[][4] = {
2468 { X86::SHR8ri, X86::SHR16ri, X86::SHR32ri, X86::SHRD32rri8 }, // SHR
2469 { X86::SAR8ri, X86::SAR16ri, X86::SAR32ri, X86::SHRD32rri8 }, // SAR
2470 { X86::SHL8ri, X86::SHL16ri, X86::SHL32ri, X86::SHLD32rri8 }, // SHL
2471 { X86::SHL8ri, X86::SHL16ri, X86::SHL32ri, X86::SHLD32rri8 }, // SAL = SHL
2474 static const unsigned NonConstantOperand[][4] = {
2475 { X86::SHR8rCL, X86::SHR16rCL, X86::SHR32rCL }, // SHR
2476 { X86::SAR8rCL, X86::SAR16rCL, X86::SAR32rCL }, // SAR
2477 { X86::SHL8rCL, X86::SHL16rCL, X86::SHL32rCL }, // SHL
2478 { X86::SHL8rCL, X86::SHL16rCL, X86::SHL32rCL }, // SAL = SHL
2481 // Longs, as usual, are handled specially...
2482 if (Class == cLong) {
2483 // If we have a constant shift, we can generate much more efficient code
2484 // than otherwise...
2486 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(ShiftAmount)) {
2487 unsigned Amount = CUI->getValue();
2489 const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned];
2491 BuildMI(*MBB, IP, Opc[3], 3,
2492 DestReg+1).addReg(SrcReg+1).addReg(SrcReg).addImm(Amount);
2493 BuildMI(*MBB, IP, Opc[2], 2, DestReg).addReg(SrcReg).addImm(Amount);
2495 BuildMI(*MBB, IP, Opc[3], 3,
2496 DestReg).addReg(SrcReg ).addReg(SrcReg+1).addImm(Amount);
2497 BuildMI(*MBB, IP, Opc[2],2,DestReg+1).addReg(SrcReg+1).addImm(Amount);
2499 } else { // Shifting more than 32 bits
2503 BuildMI(*MBB, IP, X86::SHL32ri, 2,
2504 DestReg + 1).addReg(SrcReg).addImm(Amount);
2506 BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg+1).addReg(SrcReg);
2508 BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg).addImm(0);
2511 BuildMI(*MBB, IP, isSigned ? X86::SAR32ri : X86::SHR32ri, 2,
2512 DestReg).addReg(SrcReg+1).addImm(Amount);
2514 BuildMI(*MBB, IP, X86::MOV32rr, 1, DestReg).addReg(SrcReg+1);
2516 BuildMI(*MBB, IP, X86::MOV32ri, 1, DestReg+1).addImm(0);
2520 unsigned TmpReg = makeAnotherReg(Type::IntTy);
2522 if (!isLeftShift && isSigned) {
2523 // If this is a SHR of a Long, then we need to do funny sign extension
2524 // stuff. TmpReg gets the value to use as the high-part if we are
2525 // shifting more than 32 bits.
2526 BuildMI(*MBB, IP, X86::SAR32ri, 2, TmpReg).addReg(SrcReg).addImm(31);
2528 // Other shifts use a fixed zero value if the shift is more than 32
2530 BuildMI(*MBB, IP, X86::MOV32ri, 1, TmpReg).addImm(0);
2533 // Initialize CL with the shift amount...
2534 unsigned ShiftAmountReg = getReg(ShiftAmount, MBB, IP);
2535 BuildMI(*MBB, IP, X86::MOV8rr, 1, X86::CL).addReg(ShiftAmountReg);
2537 unsigned TmpReg2 = makeAnotherReg(Type::IntTy);
2538 unsigned TmpReg3 = makeAnotherReg(Type::IntTy);
2540 // TmpReg2 = shld inHi, inLo
2541 BuildMI(*MBB, IP, X86::SHLD32rrCL,2,TmpReg2).addReg(SrcReg+1)
2543 // TmpReg3 = shl inLo, CL
2544 BuildMI(*MBB, IP, X86::SHL32rCL, 1, TmpReg3).addReg(SrcReg);
2546 // Set the flags to indicate whether the shift was by more than 32 bits.
2547 BuildMI(*MBB, IP, X86::TEST8ri, 2).addReg(X86::CL).addImm(32);
2549 // DestHi = (>32) ? TmpReg3 : TmpReg2;
2550 BuildMI(*MBB, IP, X86::CMOVNE32rr, 2,
2551 DestReg+1).addReg(TmpReg2).addReg(TmpReg3);
2552 // DestLo = (>32) ? TmpReg : TmpReg3;
2553 BuildMI(*MBB, IP, X86::CMOVNE32rr, 2,
2554 DestReg).addReg(TmpReg3).addReg(TmpReg);
2556 // TmpReg2 = shrd inLo, inHi
2557 BuildMI(*MBB, IP, X86::SHRD32rrCL,2,TmpReg2).addReg(SrcReg)
2559 // TmpReg3 = s[ah]r inHi, CL
2560 BuildMI(*MBB, IP, isSigned ? X86::SAR32rCL : X86::SHR32rCL, 1, TmpReg3)
2563 // Set the flags to indicate whether the shift was by more than 32 bits.
2564 BuildMI(*MBB, IP, X86::TEST8ri, 2).addReg(X86::CL).addImm(32);
2566 // DestLo = (>32) ? TmpReg3 : TmpReg2;
2567 BuildMI(*MBB, IP, X86::CMOVNE32rr, 2,
2568 DestReg).addReg(TmpReg2).addReg(TmpReg3);
2570 // DestHi = (>32) ? TmpReg : TmpReg3;
2571 BuildMI(*MBB, IP, X86::CMOVNE32rr, 2,
2572 DestReg+1).addReg(TmpReg3).addReg(TmpReg);
2578 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(ShiftAmount)) {
2579 // The shift amount is constant, guaranteed to be a ubyte. Get its value.
2580 assert(CUI->getType() == Type::UByteTy && "Shift amount not a ubyte?");
2582 const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned];
2583 BuildMI(*MBB, IP, Opc[Class], 2,
2584 DestReg).addReg(SrcReg).addImm(CUI->getValue());
2585 } else { // The shift amount is non-constant.
2586 unsigned ShiftAmountReg = getReg (ShiftAmount, MBB, IP);
2587 BuildMI(*MBB, IP, X86::MOV8rr, 1, X86::CL).addReg(ShiftAmountReg);
2589 const unsigned *Opc = NonConstantOperand[isLeftShift*2+isSigned];
2590 BuildMI(*MBB, IP, Opc[Class], 1, DestReg).addReg(SrcReg);
2595 void ISel::getAddressingMode(Value *Addr, unsigned &BaseReg, unsigned &Scale,
2596 unsigned &IndexReg, unsigned &Disp) {
2597 BaseReg = 0; Scale = 1; IndexReg = 0; Disp = 0;
2598 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(Addr)) {
2599 if (isGEPFoldable(BB, GEP->getOperand(0), GEP->op_begin()+1, GEP->op_end(),
2600 BaseReg, Scale, IndexReg, Disp))
2602 } else if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Addr)) {
2603 if (CE->getOpcode() == Instruction::GetElementPtr)
2604 if (isGEPFoldable(BB, CE->getOperand(0), CE->op_begin()+1, CE->op_end(),
2605 BaseReg, Scale, IndexReg, Disp))
2609 // If it's not foldable, reset addr mode.
2610 BaseReg = getReg(Addr);
2611 Scale = 1; IndexReg = 0; Disp = 0;
2615 /// visitLoadInst - Implement LLVM load instructions in terms of the x86 'mov'
2616 /// instruction. The load and store instructions are the only place where we
2617 /// need to worry about the memory layout of the target machine.
2619 void ISel::visitLoadInst(LoadInst &I) {
2620 // Check to see if this load instruction is going to be folded into a binary
2621 // instruction, like add. If so, we don't want to emit it. Wouldn't a real
2622 // pattern matching instruction selector be nice?
2623 unsigned Class = getClassB(I.getType());
2624 if (I.hasOneUse()) {
2625 Instruction *User = cast<Instruction>(I.use_back());
2626 switch (User->getOpcode()) {
2627 case Instruction::Cast:
2628 // If this is a cast from a signed-integer type to a floating point type,
2629 // fold the cast here.
2630 if (getClass(User->getType()) == cFP &&
2631 (I.getType() == Type::ShortTy || I.getType() == Type::IntTy ||
2632 I.getType() == Type::LongTy)) {
2633 unsigned DestReg = getReg(User);
2634 static const unsigned Opcode[] = {
2635 0/*BYTE*/, X86::FILD16m, X86::FILD32m, 0/*FP*/, X86::FILD64m
2637 unsigned BaseReg = 0, Scale = 1, IndexReg = 0, Disp = 0;
2638 getAddressingMode(I.getOperand(0), BaseReg, Scale, IndexReg, Disp);
2639 addFullAddress(BuildMI(BB, Opcode[Class], 5, DestReg),
2640 BaseReg, Scale, IndexReg, Disp);
2647 case Instruction::Add:
2648 case Instruction::Sub:
2649 case Instruction::And:
2650 case Instruction::Or:
2651 case Instruction::Xor:
2652 if (Class == cLong) User = 0;
2654 case Instruction::Mul:
2655 case Instruction::Div:
2656 if (Class != cFP) User = 0;
2657 break; // Folding only implemented for floating point.
2658 default: User = 0; break;
2662 // Okay, we found a user. If the load is the first operand and there is
2663 // no second operand load, reverse the operand ordering. Note that this
2664 // can fail for a subtract (ie, no change will be made).
2665 if (!isa<LoadInst>(User->getOperand(1)))
2666 cast<BinaryOperator>(User)->swapOperands();
2668 // Okay, now that everything is set up, if this load is used by the second
2669 // operand, and if there are no instructions that invalidate the load
2670 // before the binary operator, eliminate the load.
2671 if (User->getOperand(1) == &I &&
2672 isSafeToFoldLoadIntoInstruction(I, *User))
2673 return; // Eliminate the load!
2675 // If this is a floating point sub or div, we won't be able to swap the
2676 // operands, but we will still be able to eliminate the load.
2677 if (Class == cFP && User->getOperand(0) == &I &&
2678 !isa<LoadInst>(User->getOperand(1)) &&
2679 (User->getOpcode() == Instruction::Sub ||
2680 User->getOpcode() == Instruction::Div) &&
2681 isSafeToFoldLoadIntoInstruction(I, *User))
2682 return; // Eliminate the load!
2686 unsigned DestReg = getReg(I);
2687 unsigned BaseReg = 0, Scale = 1, IndexReg = 0, Disp = 0;
2688 getAddressingMode(I.getOperand(0), BaseReg, Scale, IndexReg, Disp);
2690 if (Class == cLong) {
2691 addFullAddress(BuildMI(BB, X86::MOV32rm, 4, DestReg),
2692 BaseReg, Scale, IndexReg, Disp);
2693 addFullAddress(BuildMI(BB, X86::MOV32rm, 4, DestReg+1),
2694 BaseReg, Scale, IndexReg, Disp+4);
2698 static const unsigned Opcodes[] = {
2699 X86::MOV8rm, X86::MOV16rm, X86::MOV32rm, X86::FLD32m
2701 unsigned Opcode = Opcodes[Class];
2702 if (I.getType() == Type::DoubleTy) Opcode = X86::FLD64m;
2703 addFullAddress(BuildMI(BB, Opcode, 4, DestReg),
2704 BaseReg, Scale, IndexReg, Disp);
2707 /// visitStoreInst - Implement LLVM store instructions in terms of the x86 'mov'
2710 void ISel::visitStoreInst(StoreInst &I) {
2711 unsigned BaseReg, Scale, IndexReg, Disp;
2712 getAddressingMode(I.getOperand(1), BaseReg, Scale, IndexReg, Disp);
2714 const Type *ValTy = I.getOperand(0)->getType();
2715 unsigned Class = getClassB(ValTy);
2717 if (ConstantInt *CI = dyn_cast<ConstantInt>(I.getOperand(0))) {
2718 uint64_t Val = CI->getRawValue();
2719 if (Class == cLong) {
2720 addFullAddress(BuildMI(BB, X86::MOV32mi, 5),
2721 BaseReg, Scale, IndexReg, Disp).addImm(Val & ~0U);
2722 addFullAddress(BuildMI(BB, X86::MOV32mi, 5),
2723 BaseReg, Scale, IndexReg, Disp+4).addImm(Val>>32);
2725 static const unsigned Opcodes[] = {
2726 X86::MOV8mi, X86::MOV16mi, X86::MOV32mi
2728 unsigned Opcode = Opcodes[Class];
2729 addFullAddress(BuildMI(BB, Opcode, 5),
2730 BaseReg, Scale, IndexReg, Disp).addImm(Val);
2732 } else if (ConstantBool *CB = dyn_cast<ConstantBool>(I.getOperand(0))) {
2733 addFullAddress(BuildMI(BB, X86::MOV8mi, 5),
2734 BaseReg, Scale, IndexReg, Disp).addImm(CB->getValue());
2736 if (Class == cLong) {
2737 unsigned ValReg = getReg(I.getOperand(0));
2738 addFullAddress(BuildMI(BB, X86::MOV32mr, 5),
2739 BaseReg, Scale, IndexReg, Disp).addReg(ValReg);
2740 addFullAddress(BuildMI(BB, X86::MOV32mr, 5),
2741 BaseReg, Scale, IndexReg, Disp+4).addReg(ValReg+1);
2743 unsigned ValReg = getReg(I.getOperand(0));
2744 static const unsigned Opcodes[] = {
2745 X86::MOV8mr, X86::MOV16mr, X86::MOV32mr, X86::FST32m
2747 unsigned Opcode = Opcodes[Class];
2748 if (ValTy == Type::DoubleTy) Opcode = X86::FST64m;
2749 addFullAddress(BuildMI(BB, Opcode, 1+4),
2750 BaseReg, Scale, IndexReg, Disp).addReg(ValReg);
2756 /// visitCastInst - Here we have various kinds of copying with or without sign
2757 /// extension going on.
2759 void ISel::visitCastInst(CastInst &CI) {
2760 Value *Op = CI.getOperand(0);
2762 unsigned SrcClass = getClassB(Op->getType());
2763 unsigned DestClass = getClassB(CI.getType());
2764 // Noop casts are not emitted: getReg will return the source operand as the
2765 // register to use for any uses of the noop cast.
2766 if (DestClass == SrcClass)
2769 // If this is a cast from a 32-bit integer to a Long type, and the only uses
2770 // of the case are GEP instructions, then the cast does not need to be
2771 // generated explicitly, it will be folded into the GEP.
2772 if (DestClass == cLong && SrcClass == cInt) {
2773 bool AllUsesAreGEPs = true;
2774 for (Value::use_iterator I = CI.use_begin(), E = CI.use_end(); I != E; ++I)
2775 if (!isa<GetElementPtrInst>(*I)) {
2776 AllUsesAreGEPs = false;
2780 // No need to codegen this cast if all users are getelementptr instrs...
2781 if (AllUsesAreGEPs) return;
2784 // If this cast converts a load from a short,int, or long integer to a FP
2785 // value, we will have folded this cast away.
2786 if (DestClass == cFP && isa<LoadInst>(Op) && Op->hasOneUse() &&
2787 (Op->getType() == Type::ShortTy || Op->getType() == Type::IntTy ||
2788 Op->getType() == Type::LongTy))
2792 unsigned DestReg = getReg(CI);
2793 MachineBasicBlock::iterator MI = BB->end();
2794 emitCastOperation(BB, MI, Op, CI.getType(), DestReg);
2797 /// emitCastOperation - Common code shared between visitCastInst and constant
2798 /// expression cast support.
2800 void ISel::emitCastOperation(MachineBasicBlock *BB,
2801 MachineBasicBlock::iterator IP,
2802 Value *Src, const Type *DestTy,
2804 const Type *SrcTy = Src->getType();
2805 unsigned SrcClass = getClassB(SrcTy);
2806 unsigned DestClass = getClassB(DestTy);
2807 unsigned SrcReg = getReg(Src, BB, IP);
2809 // Implement casts to bool by using compare on the operand followed by set if
2810 // not zero on the result.
2811 if (DestTy == Type::BoolTy) {
2814 BuildMI(*BB, IP, X86::TEST8rr, 2).addReg(SrcReg).addReg(SrcReg);
2817 BuildMI(*BB, IP, X86::TEST16rr, 2).addReg(SrcReg).addReg(SrcReg);
2820 BuildMI(*BB, IP, X86::TEST32rr, 2).addReg(SrcReg).addReg(SrcReg);
2823 unsigned TmpReg = makeAnotherReg(Type::IntTy);
2824 BuildMI(*BB, IP, X86::OR32rr, 2, TmpReg).addReg(SrcReg).addReg(SrcReg+1);
2828 BuildMI(*BB, IP, X86::FTST, 1).addReg(SrcReg);
2829 BuildMI(*BB, IP, X86::FNSTSW8r, 0);
2830 BuildMI(*BB, IP, X86::SAHF, 1);
2834 // If the zero flag is not set, then the value is true, set the byte to
2836 BuildMI(*BB, IP, X86::SETNEr, 1, DestReg);
2840 static const unsigned RegRegMove[] = {
2841 X86::MOV8rr, X86::MOV16rr, X86::MOV32rr, X86::FpMOV, X86::MOV32rr
2844 // Implement casts between values of the same type class (as determined by
2845 // getClass) by using a register-to-register move.
2846 if (SrcClass == DestClass) {
2847 if (SrcClass <= cInt || (SrcClass == cFP && SrcTy == DestTy)) {
2848 BuildMI(*BB, IP, RegRegMove[SrcClass], 1, DestReg).addReg(SrcReg);
2849 } else if (SrcClass == cFP) {
2850 if (SrcTy == Type::FloatTy) { // double -> float
2851 assert(DestTy == Type::DoubleTy && "Unknown cFP member!");
2852 BuildMI(*BB, IP, X86::FpMOV, 1, DestReg).addReg(SrcReg);
2853 } else { // float -> double
2854 assert(SrcTy == Type::DoubleTy && DestTy == Type::FloatTy &&
2855 "Unknown cFP member!");
2856 // Truncate from double to float by storing to memory as short, then
2858 unsigned FltAlign = TM.getTargetData().getFloatAlignment();
2859 int FrameIdx = F->getFrameInfo()->CreateStackObject(4, FltAlign);
2860 addFrameReference(BuildMI(*BB, IP, X86::FST32m, 5), FrameIdx).addReg(SrcReg);
2861 addFrameReference(BuildMI(*BB, IP, X86::FLD32m, 5, DestReg), FrameIdx);
2863 } else if (SrcClass == cLong) {
2864 BuildMI(*BB, IP, X86::MOV32rr, 1, DestReg).addReg(SrcReg);
2865 BuildMI(*BB, IP, X86::MOV32rr, 1, DestReg+1).addReg(SrcReg+1);
2867 assert(0 && "Cannot handle this type of cast instruction!");
2873 // Handle cast of SMALLER int to LARGER int using a move with sign extension
2874 // or zero extension, depending on whether the source type was signed.
2875 if (SrcClass <= cInt && (DestClass <= cInt || DestClass == cLong) &&
2876 SrcClass < DestClass) {
2877 bool isLong = DestClass == cLong;
2878 if (isLong) DestClass = cInt;
2880 static const unsigned Opc[][4] = {
2881 { X86::MOVSX16rr8, X86::MOVSX32rr8, X86::MOVSX32rr16, X86::MOV32rr }, // s
2882 { X86::MOVZX16rr8, X86::MOVZX32rr8, X86::MOVZX32rr16, X86::MOV32rr } // u
2885 bool isUnsigned = SrcTy->isUnsigned();
2886 BuildMI(*BB, IP, Opc[isUnsigned][SrcClass + DestClass - 1], 1,
2887 DestReg).addReg(SrcReg);
2889 if (isLong) { // Handle upper 32 bits as appropriate...
2890 if (isUnsigned) // Zero out top bits...
2891 BuildMI(*BB, IP, X86::MOV32ri, 1, DestReg+1).addImm(0);
2892 else // Sign extend bottom half...
2893 BuildMI(*BB, IP, X86::SAR32ri, 2, DestReg+1).addReg(DestReg).addImm(31);
2898 // Special case long -> int ...
2899 if (SrcClass == cLong && DestClass == cInt) {
2900 BuildMI(*BB, IP, X86::MOV32rr, 1, DestReg).addReg(SrcReg);
2904 // Handle cast of LARGER int to SMALLER int using a move to EAX followed by a
2905 // move out of AX or AL.
2906 if ((SrcClass <= cInt || SrcClass == cLong) && DestClass <= cInt
2907 && SrcClass > DestClass) {
2908 static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX, 0, X86::EAX };
2909 BuildMI(*BB, IP, RegRegMove[SrcClass], 1, AReg[SrcClass]).addReg(SrcReg);
2910 BuildMI(*BB, IP, RegRegMove[DestClass], 1, DestReg).addReg(AReg[DestClass]);
2914 // Handle casts from integer to floating point now...
2915 if (DestClass == cFP) {
2916 // Promote the integer to a type supported by FLD. We do this because there
2917 // are no unsigned FLD instructions, so we must promote an unsigned value to
2918 // a larger signed value, then use FLD on the larger value.
2920 const Type *PromoteType = 0;
2921 unsigned PromoteOpcode = 0;
2922 unsigned RealDestReg = DestReg;
2923 switch (SrcTy->getPrimitiveID()) {
2924 case Type::BoolTyID:
2925 case Type::SByteTyID:
2926 // We don't have the facilities for directly loading byte sized data from
2927 // memory (even signed). Promote it to 16 bits.
2928 PromoteType = Type::ShortTy;
2929 PromoteOpcode = X86::MOVSX16rr8;
2931 case Type::UByteTyID:
2932 PromoteType = Type::ShortTy;
2933 PromoteOpcode = X86::MOVZX16rr8;
2935 case Type::UShortTyID:
2936 PromoteType = Type::IntTy;
2937 PromoteOpcode = X86::MOVZX32rr16;
2939 case Type::UIntTyID: {
2940 // Make a 64 bit temporary... and zero out the top of it...
2941 unsigned TmpReg = makeAnotherReg(Type::LongTy);
2942 BuildMI(*BB, IP, X86::MOV32rr, 1, TmpReg).addReg(SrcReg);
2943 BuildMI(*BB, IP, X86::MOV32ri, 1, TmpReg+1).addImm(0);
2944 SrcTy = Type::LongTy;
2949 case Type::ULongTyID:
2950 // Don't fild into the read destination.
2951 DestReg = makeAnotherReg(Type::DoubleTy);
2953 default: // No promotion needed...
2958 unsigned TmpReg = makeAnotherReg(PromoteType);
2959 BuildMI(*BB, IP, PromoteOpcode, 1, TmpReg).addReg(SrcReg);
2960 SrcTy = PromoteType;
2961 SrcClass = getClass(PromoteType);
2965 // Spill the integer to memory and reload it from there...
2967 F->getFrameInfo()->CreateStackObject(SrcTy, TM.getTargetData());
2969 if (SrcClass == cLong) {
2970 addFrameReference(BuildMI(*BB, IP, X86::MOV32mr, 5),
2971 FrameIdx).addReg(SrcReg);
2972 addFrameReference(BuildMI(*BB, IP, X86::MOV32mr, 5),
2973 FrameIdx, 4).addReg(SrcReg+1);
2975 static const unsigned Op1[] = { X86::MOV8mr, X86::MOV16mr, X86::MOV32mr };
2976 addFrameReference(BuildMI(*BB, IP, Op1[SrcClass], 5),
2977 FrameIdx).addReg(SrcReg);
2980 static const unsigned Op2[] =
2981 { 0/*byte*/, X86::FILD16m, X86::FILD32m, 0/*FP*/, X86::FILD64m };
2982 addFrameReference(BuildMI(*BB, IP, Op2[SrcClass], 5, DestReg), FrameIdx);
2984 // We need special handling for unsigned 64-bit integer sources. If the
2985 // input number has the "sign bit" set, then we loaded it incorrectly as a
2986 // negative 64-bit number. In this case, add an offset value.
2987 if (SrcTy == Type::ULongTy) {
2988 // Emit a test instruction to see if the dynamic input value was signed.
2989 BuildMI(*BB, IP, X86::TEST32rr, 2).addReg(SrcReg+1).addReg(SrcReg+1);
2991 // If the sign bit is set, get a pointer to an offset, otherwise get a
2992 // pointer to a zero.
2993 MachineConstantPool *CP = F->getConstantPool();
2994 unsigned Zero = makeAnotherReg(Type::IntTy);
2995 Constant *Null = Constant::getNullValue(Type::UIntTy);
2996 addConstantPoolReference(BuildMI(*BB, IP, X86::LEA32r, 5, Zero),
2997 CP->getConstantPoolIndex(Null));
2998 unsigned Offset = makeAnotherReg(Type::IntTy);
2999 Constant *OffsetCst = ConstantUInt::get(Type::UIntTy, 0x5f800000);
3001 addConstantPoolReference(BuildMI(*BB, IP, X86::LEA32r, 5, Offset),
3002 CP->getConstantPoolIndex(OffsetCst));
3003 unsigned Addr = makeAnotherReg(Type::IntTy);
3004 BuildMI(*BB, IP, X86::CMOVS32rr, 2, Addr).addReg(Zero).addReg(Offset);
3006 // Load the constant for an add. FIXME: this could make an 'fadd' that
3007 // reads directly from memory, but we don't support these yet.
3008 unsigned ConstReg = makeAnotherReg(Type::DoubleTy);
3009 addDirectMem(BuildMI(*BB, IP, X86::FLD32m, 4, ConstReg), Addr);
3011 BuildMI(*BB, IP, X86::FpADD, 2, RealDestReg)
3012 .addReg(ConstReg).addReg(DestReg);
3018 // Handle casts from floating point to integer now...
3019 if (SrcClass == cFP) {
3020 // Change the floating point control register to use "round towards zero"
3021 // mode when truncating to an integer value.
3023 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
3024 addFrameReference(BuildMI(*BB, IP, X86::FNSTCW16m, 4), CWFrameIdx);
3026 // Load the old value of the high byte of the control word...
3027 unsigned HighPartOfCW = makeAnotherReg(Type::UByteTy);
3028 addFrameReference(BuildMI(*BB, IP, X86::MOV8rm, 4, HighPartOfCW),
3031 // Set the high part to be round to zero...
3032 addFrameReference(BuildMI(*BB, IP, X86::MOV8mi, 5),
3033 CWFrameIdx, 1).addImm(12);
3035 // Reload the modified control word now...
3036 addFrameReference(BuildMI(*BB, IP, X86::FLDCW16m, 4), CWFrameIdx);
3038 // Restore the memory image of control word to original value
3039 addFrameReference(BuildMI(*BB, IP, X86::MOV8mr, 5),
3040 CWFrameIdx, 1).addReg(HighPartOfCW);
3042 // We don't have the facilities for directly storing byte sized data to
3043 // memory. Promote it to 16 bits. We also must promote unsigned values to
3044 // larger classes because we only have signed FP stores.
3045 unsigned StoreClass = DestClass;
3046 const Type *StoreTy = DestTy;
3047 if (StoreClass == cByte || DestTy->isUnsigned())
3048 switch (StoreClass) {
3049 case cByte: StoreTy = Type::ShortTy; StoreClass = cShort; break;
3050 case cShort: StoreTy = Type::IntTy; StoreClass = cInt; break;
3051 case cInt: StoreTy = Type::LongTy; StoreClass = cLong; break;
3052 // The following treatment of cLong may not be perfectly right,
3053 // but it survives chains of casts of the form
3054 // double->ulong->double.
3055 case cLong: StoreTy = Type::LongTy; StoreClass = cLong; break;
3056 default: assert(0 && "Unknown store class!");
3059 // Spill the integer to memory and reload it from there...
3061 F->getFrameInfo()->CreateStackObject(StoreTy, TM.getTargetData());
3063 static const unsigned Op1[] =
3064 { 0, X86::FIST16m, X86::FIST32m, 0, X86::FISTP64m };
3065 addFrameReference(BuildMI(*BB, IP, Op1[StoreClass], 5),
3066 FrameIdx).addReg(SrcReg);
3068 if (DestClass == cLong) {
3069 addFrameReference(BuildMI(*BB, IP, X86::MOV32rm, 4, DestReg), FrameIdx);
3070 addFrameReference(BuildMI(*BB, IP, X86::MOV32rm, 4, DestReg+1),
3073 static const unsigned Op2[] = { X86::MOV8rm, X86::MOV16rm, X86::MOV32rm };
3074 addFrameReference(BuildMI(*BB, IP, Op2[DestClass], 4, DestReg), FrameIdx);
3077 // Reload the original control word now...
3078 addFrameReference(BuildMI(*BB, IP, X86::FLDCW16m, 4), CWFrameIdx);
3082 // Anything we haven't handled already, we can't (yet) handle at all.
3083 assert(0 && "Unhandled cast instruction!");
3087 /// visitVANextInst - Implement the va_next instruction...
3089 void ISel::visitVANextInst(VANextInst &I) {
3090 unsigned VAList = getReg(I.getOperand(0));
3091 unsigned DestReg = getReg(I);
3094 switch (I.getArgType()->getPrimitiveID()) {
3097 assert(0 && "Error: bad type for va_next instruction!");
3099 case Type::PointerTyID:
3100 case Type::UIntTyID:
3104 case Type::ULongTyID:
3105 case Type::LongTyID:
3106 case Type::DoubleTyID:
3111 // Increment the VAList pointer...
3112 BuildMI(BB, X86::ADD32ri, 2, DestReg).addReg(VAList).addImm(Size);
3115 void ISel::visitVAArgInst(VAArgInst &I) {
3116 unsigned VAList = getReg(I.getOperand(0));
3117 unsigned DestReg = getReg(I);
3119 switch (I.getType()->getPrimitiveID()) {
3122 assert(0 && "Error: bad type for va_next instruction!");
3124 case Type::PointerTyID:
3125 case Type::UIntTyID:
3127 addDirectMem(BuildMI(BB, X86::MOV32rm, 4, DestReg), VAList);
3129 case Type::ULongTyID:
3130 case Type::LongTyID:
3131 addDirectMem(BuildMI(BB, X86::MOV32rm, 4, DestReg), VAList);
3132 addRegOffset(BuildMI(BB, X86::MOV32rm, 4, DestReg+1), VAList, 4);
3134 case Type::DoubleTyID:
3135 addDirectMem(BuildMI(BB, X86::FLD64m, 4, DestReg), VAList);
3140 /// visitGetElementPtrInst - instruction-select GEP instructions
3142 void ISel::visitGetElementPtrInst(GetElementPtrInst &I) {
3143 // If this GEP instruction will be folded into all of its users, we don't need
3144 // to explicitly calculate it!
3145 unsigned A, B, C, D;
3146 if (isGEPFoldable(0, I.getOperand(0), I.op_begin()+1, I.op_end(), A,B,C,D)) {
3147 // Check all of the users of the instruction to see if they are loads and
3149 bool AllWillFold = true;
3150 for (Value::use_iterator UI = I.use_begin(), E = I.use_end(); UI != E; ++UI)
3151 if (cast<Instruction>(*UI)->getOpcode() != Instruction::Load)
3152 if (cast<Instruction>(*UI)->getOpcode() != Instruction::Store ||
3153 cast<Instruction>(*UI)->getOperand(0) == &I) {
3154 AllWillFold = false;
3158 // If the instruction is foldable, and will be folded into all users, don't
3160 if (AllWillFold) return;
3163 unsigned outputReg = getReg(I);
3164 emitGEPOperation(BB, BB->end(), I.getOperand(0),
3165 I.op_begin()+1, I.op_end(), outputReg);
3168 /// getGEPIndex - Inspect the getelementptr operands specified with GEPOps and
3169 /// GEPTypes (the derived types being stepped through at each level). On return
3170 /// from this function, if some indexes of the instruction are representable as
3171 /// an X86 lea instruction, the machine operands are put into the Ops
3172 /// instruction and the consumed indexes are poped from the GEPOps/GEPTypes
3173 /// lists. Otherwise, GEPOps.size() is returned. If this returns a an
3174 /// addressing mode that only partially consumes the input, the BaseReg input of
3175 /// the addressing mode must be left free.
3177 /// Note that there is one fewer entry in GEPTypes than there is in GEPOps.
3179 void ISel::getGEPIndex(MachineBasicBlock *MBB, MachineBasicBlock::iterator IP,
3180 std::vector<Value*> &GEPOps,
3181 std::vector<const Type*> &GEPTypes, unsigned &BaseReg,
3182 unsigned &Scale, unsigned &IndexReg, unsigned &Disp) {
3183 const TargetData &TD = TM.getTargetData();
3185 // Clear out the state we are working with...
3186 BaseReg = 0; // No base register
3187 Scale = 1; // Unit scale
3188 IndexReg = 0; // No index register
3189 Disp = 0; // No displacement
3191 // While there are GEP indexes that can be folded into the current address,
3192 // keep processing them.
3193 while (!GEPTypes.empty()) {
3194 if (const StructType *StTy = dyn_cast<StructType>(GEPTypes.back())) {
3195 // It's a struct access. CUI is the index into the structure,
3196 // which names the field. This index must have unsigned type.
3197 const ConstantUInt *CUI = cast<ConstantUInt>(GEPOps.back());
3199 // Use the TargetData structure to pick out what the layout of the
3200 // structure is in memory. Since the structure index must be constant, we
3201 // can get its value and use it to find the right byte offset from the
3202 // StructLayout class's list of structure member offsets.
3203 Disp += TD.getStructLayout(StTy)->MemberOffsets[CUI->getValue()];
3204 GEPOps.pop_back(); // Consume a GEP operand
3205 GEPTypes.pop_back();
3207 // It's an array or pointer access: [ArraySize x ElementType].
3208 const SequentialType *SqTy = cast<SequentialType>(GEPTypes.back());
3209 Value *idx = GEPOps.back();
3211 // idx is the index into the array. Unlike with structure
3212 // indices, we may not know its actual value at code-generation
3215 // If idx is a constant, fold it into the offset.
3216 unsigned TypeSize = TD.getTypeSize(SqTy->getElementType());
3217 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(idx)) {
3218 Disp += TypeSize*CSI->getValue();
3219 } else if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(idx)) {
3220 Disp += TypeSize*CUI->getValue();
3222 // If the index reg is already taken, we can't handle this index.
3223 if (IndexReg) return;
3225 // If this is a size that we can handle, then add the index as
3227 case 1: case 2: case 4: case 8:
3228 // These are all acceptable scales on X86.
3232 // Otherwise, we can't handle this scale
3236 if (CastInst *CI = dyn_cast<CastInst>(idx))
3237 if (CI->getOperand(0)->getType() == Type::IntTy ||
3238 CI->getOperand(0)->getType() == Type::UIntTy)
3239 idx = CI->getOperand(0);
3241 IndexReg = MBB ? getReg(idx, MBB, IP) : 1;
3244 GEPOps.pop_back(); // Consume a GEP operand
3245 GEPTypes.pop_back();
3249 // GEPTypes is empty, which means we have a single operand left. See if we
3250 // can set it as the base register.
3252 // FIXME: When addressing modes are more powerful/correct, we could load
3253 // global addresses directly as 32-bit immediates.
3254 assert(BaseReg == 0);
3255 BaseReg = MBB ? getReg(GEPOps[0], MBB, IP) : 1;
3256 GEPOps.pop_back(); // Consume the last GEP operand
3260 /// isGEPFoldable - Return true if the specified GEP can be completely
3261 /// folded into the addressing mode of a load/store or lea instruction.
3262 bool ISel::isGEPFoldable(MachineBasicBlock *MBB,
3263 Value *Src, User::op_iterator IdxBegin,
3264 User::op_iterator IdxEnd, unsigned &BaseReg,
3265 unsigned &Scale, unsigned &IndexReg, unsigned &Disp) {
3266 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Src))
3267 Src = CPR->getValue();
3269 std::vector<Value*> GEPOps;
3270 GEPOps.resize(IdxEnd-IdxBegin+1);
3272 std::copy(IdxBegin, IdxEnd, GEPOps.begin()+1);
3274 std::vector<const Type*> GEPTypes;
3275 GEPTypes.assign(gep_type_begin(Src->getType(), IdxBegin, IdxEnd),
3276 gep_type_end(Src->getType(), IdxBegin, IdxEnd));
3278 MachineBasicBlock::iterator IP;
3279 if (MBB) IP = MBB->end();
3280 getGEPIndex(MBB, IP, GEPOps, GEPTypes, BaseReg, Scale, IndexReg, Disp);
3282 // We can fold it away iff the getGEPIndex call eliminated all operands.
3283 return GEPOps.empty();
3286 void ISel::emitGEPOperation(MachineBasicBlock *MBB,
3287 MachineBasicBlock::iterator IP,
3288 Value *Src, User::op_iterator IdxBegin,
3289 User::op_iterator IdxEnd, unsigned TargetReg) {
3290 const TargetData &TD = TM.getTargetData();
3291 if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(Src))
3292 Src = CPR->getValue();
3294 std::vector<Value*> GEPOps;
3295 GEPOps.resize(IdxEnd-IdxBegin+1);
3297 std::copy(IdxBegin, IdxEnd, GEPOps.begin()+1);
3299 std::vector<const Type*> GEPTypes;
3300 GEPTypes.assign(gep_type_begin(Src->getType(), IdxBegin, IdxEnd),
3301 gep_type_end(Src->getType(), IdxBegin, IdxEnd));
3303 // Keep emitting instructions until we consume the entire GEP instruction.
3304 while (!GEPOps.empty()) {
3305 unsigned OldSize = GEPOps.size();
3306 unsigned BaseReg, Scale, IndexReg, Disp;
3307 getGEPIndex(MBB, IP, GEPOps, GEPTypes, BaseReg, Scale, IndexReg, Disp);
3309 if (GEPOps.size() != OldSize) {
3310 // getGEPIndex consumed some of the input. Build an LEA instruction here.
3311 unsigned NextTarget = 0;
3312 if (!GEPOps.empty()) {
3313 assert(BaseReg == 0 &&
3314 "getGEPIndex should have left the base register open for chaining!");
3315 NextTarget = BaseReg = makeAnotherReg(Type::UIntTy);
3318 if (IndexReg == 0 && Disp == 0)
3319 BuildMI(*MBB, IP, X86::MOV32rr, 1, TargetReg).addReg(BaseReg);
3321 addFullAddress(BuildMI(*MBB, IP, X86::LEA32r, 5, TargetReg),
3322 BaseReg, Scale, IndexReg, Disp);
3324 TargetReg = NextTarget;
3325 } else if (GEPTypes.empty()) {
3326 // The getGEPIndex operation didn't want to build an LEA. Check to see if
3327 // all operands are consumed but the base pointer. If so, just load it
3328 // into the register.
3329 if (GlobalValue *GV = dyn_cast<GlobalValue>(GEPOps[0])) {
3330 BuildMI(*MBB, IP, X86::MOV32ri, 1, TargetReg).addGlobalAddress(GV);
3332 unsigned BaseReg = getReg(GEPOps[0], MBB, IP);
3333 BuildMI(*MBB, IP, X86::MOV32rr, 1, TargetReg).addReg(BaseReg);
3335 break; // we are now done
3338 // It's an array or pointer access: [ArraySize x ElementType].
3339 const SequentialType *SqTy = cast<SequentialType>(GEPTypes.back());
3340 Value *idx = GEPOps.back();
3341 GEPOps.pop_back(); // Consume a GEP operand
3342 GEPTypes.pop_back();
3344 // Many GEP instructions use a [cast (int/uint) to LongTy] as their
3345 // operand on X86. Handle this case directly now...
3346 if (CastInst *CI = dyn_cast<CastInst>(idx))
3347 if (CI->getOperand(0)->getType() == Type::IntTy ||
3348 CI->getOperand(0)->getType() == Type::UIntTy)
3349 idx = CI->getOperand(0);
3351 // We want to add BaseReg to(idxReg * sizeof ElementType). First, we
3352 // must find the size of the pointed-to type (Not coincidentally, the next
3353 // type is the type of the elements in the array).
3354 const Type *ElTy = SqTy->getElementType();
3355 unsigned elementSize = TD.getTypeSize(ElTy);
3357 // If idxReg is a constant, we don't need to perform the multiply!
3358 if (ConstantInt *CSI = dyn_cast<ConstantInt>(idx)) {
3359 if (!CSI->isNullValue()) {
3360 unsigned Offset = elementSize*CSI->getRawValue();
3361 unsigned Reg = makeAnotherReg(Type::UIntTy);
3362 BuildMI(*MBB, IP, X86::ADD32ri, 2, TargetReg)
3363 .addReg(Reg).addImm(Offset);
3364 --IP; // Insert the next instruction before this one.
3365 TargetReg = Reg; // Codegen the rest of the GEP into this
3367 } else if (elementSize == 1) {
3368 // If the element size is 1, we don't have to multiply, just add
3369 unsigned idxReg = getReg(idx, MBB, IP);
3370 unsigned Reg = makeAnotherReg(Type::UIntTy);
3371 BuildMI(*MBB, IP, X86::ADD32rr, 2,TargetReg).addReg(Reg).addReg(idxReg);
3372 --IP; // Insert the next instruction before this one.
3373 TargetReg = Reg; // Codegen the rest of the GEP into this
3375 unsigned idxReg = getReg(idx, MBB, IP);
3376 unsigned OffsetReg = makeAnotherReg(Type::UIntTy);
3378 // Make sure we can back the iterator up to point to the first
3379 // instruction emitted.
3380 MachineBasicBlock::iterator BeforeIt = IP;
3381 if (IP == MBB->begin())
3382 BeforeIt = MBB->end();
3385 doMultiplyConst(MBB, IP, OffsetReg, Type::IntTy, idxReg, elementSize);
3387 // Emit an ADD to add OffsetReg to the basePtr.
3388 unsigned Reg = makeAnotherReg(Type::UIntTy);
3389 BuildMI(*MBB, IP, X86::ADD32rr, 2, TargetReg)
3390 .addReg(Reg).addReg(OffsetReg);
3392 // Step to the first instruction of the multiply.
3393 if (BeforeIt == MBB->end())
3398 TargetReg = Reg; // Codegen the rest of the GEP into this
3405 /// visitAllocaInst - If this is a fixed size alloca, allocate space from the
3406 /// frame manager, otherwise do it the hard way.
3408 void ISel::visitAllocaInst(AllocaInst &I) {
3409 // Find the data size of the alloca inst's getAllocatedType.
3410 const Type *Ty = I.getAllocatedType();
3411 unsigned TySize = TM.getTargetData().getTypeSize(Ty);
3413 // If this is a fixed size alloca in the entry block for the function,
3414 // statically stack allocate the space.
3416 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I.getArraySize())) {
3417 if (I.getParent() == I.getParent()->getParent()->begin()) {
3418 TySize *= CUI->getValue(); // Get total allocated size...
3419 unsigned Alignment = TM.getTargetData().getTypeAlignment(Ty);
3421 // Create a new stack object using the frame manager...
3422 int FrameIdx = F->getFrameInfo()->CreateStackObject(TySize, Alignment);
3423 addFrameReference(BuildMI(BB, X86::LEA32r, 5, getReg(I)), FrameIdx);
3428 // Create a register to hold the temporary result of multiplying the type size
3429 // constant by the variable amount.
3430 unsigned TotalSizeReg = makeAnotherReg(Type::UIntTy);
3431 unsigned SrcReg1 = getReg(I.getArraySize());
3433 // TotalSizeReg = mul <numelements>, <TypeSize>
3434 MachineBasicBlock::iterator MBBI = BB->end();
3435 doMultiplyConst(BB, MBBI, TotalSizeReg, Type::UIntTy, SrcReg1, TySize);
3437 // AddedSize = add <TotalSizeReg>, 15
3438 unsigned AddedSizeReg = makeAnotherReg(Type::UIntTy);
3439 BuildMI(BB, X86::ADD32ri, 2, AddedSizeReg).addReg(TotalSizeReg).addImm(15);
3441 // AlignedSize = and <AddedSize>, ~15
3442 unsigned AlignedSize = makeAnotherReg(Type::UIntTy);
3443 BuildMI(BB, X86::AND32ri, 2, AlignedSize).addReg(AddedSizeReg).addImm(~15);
3445 // Subtract size from stack pointer, thereby allocating some space.
3446 BuildMI(BB, X86::SUB32rr, 2, X86::ESP).addReg(X86::ESP).addReg(AlignedSize);
3448 // Put a pointer to the space into the result register, by copying
3449 // the stack pointer.
3450 BuildMI(BB, X86::MOV32rr, 1, getReg(I)).addReg(X86::ESP);
3452 // Inform the Frame Information that we have just allocated a variable-sized
3454 F->getFrameInfo()->CreateVariableSizedObject();
3457 /// visitMallocInst - Malloc instructions are code generated into direct calls
3458 /// to the library malloc.
3460 void ISel::visitMallocInst(MallocInst &I) {
3461 unsigned AllocSize = TM.getTargetData().getTypeSize(I.getAllocatedType());
3464 if (ConstantUInt *C = dyn_cast<ConstantUInt>(I.getOperand(0))) {
3465 Arg = getReg(ConstantUInt::get(Type::UIntTy, C->getValue() * AllocSize));
3467 Arg = makeAnotherReg(Type::UIntTy);
3468 unsigned Op0Reg = getReg(I.getOperand(0));
3469 MachineBasicBlock::iterator MBBI = BB->end();
3470 doMultiplyConst(BB, MBBI, Arg, Type::UIntTy, Op0Reg, AllocSize);
3473 std::vector<ValueRecord> Args;
3474 Args.push_back(ValueRecord(Arg, Type::UIntTy));
3475 MachineInstr *TheCall = BuildMI(X86::CALLpcrel32,
3476 1).addExternalSymbol("malloc", true);
3477 doCall(ValueRecord(getReg(I), I.getType()), TheCall, Args);
3481 /// visitFreeInst - Free instructions are code gen'd to call the free libc
3484 void ISel::visitFreeInst(FreeInst &I) {
3485 std::vector<ValueRecord> Args;
3486 Args.push_back(ValueRecord(I.getOperand(0)));
3487 MachineInstr *TheCall = BuildMI(X86::CALLpcrel32,
3488 1).addExternalSymbol("free", true);
3489 doCall(ValueRecord(0, Type::VoidTy), TheCall, Args);
3492 /// createX86SimpleInstructionSelector - This pass converts an LLVM function
3493 /// into a machine code representation is a very simple peep-hole fashion. The
3494 /// generated code sucks but the implementation is nice and simple.
3496 FunctionPass *llvm::createX86SimpleInstructionSelector(TargetMachine &TM) {
3497 return new ISel(TM);