1 //===-- InstSelectSimple.cpp - A simple instruction selector for x86 ------===//
3 // This file defines a simple peephole instruction selector for the x86 target
5 //===----------------------------------------------------------------------===//
8 #include "X86InstrInfo.h"
9 #include "X86InstrBuilder.h"
10 #include "llvm/Function.h"
11 #include "llvm/Instructions.h"
12 #include "llvm/DerivedTypes.h"
13 #include "llvm/Constants.h"
14 #include "llvm/Pass.h"
15 #include "llvm/Intrinsics.h"
16 #include "llvm/CodeGen/MachineFunction.h"
17 #include "llvm/CodeGen/MachineInstrBuilder.h"
18 #include "llvm/CodeGen/SSARegMap.h"
19 #include "llvm/CodeGen/MachineFrameInfo.h"
20 #include "llvm/CodeGen/MachineConstantPool.h"
21 #include "llvm/Target/TargetMachine.h"
22 #include "llvm/Target/MRegisterInfo.h"
23 #include "llvm/Support/InstVisitor.h"
25 /// BMI - A special BuildMI variant that takes an iterator to insert the
26 /// instruction at as well as a basic block. This is the version for when you
27 /// have a destination register in mind.
28 inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB,
29 MachineBasicBlock::iterator &I,
30 int Opcode, unsigned NumOperands,
32 assert(I >= MBB->begin() && I <= MBB->end() && "Bad iterator!");
33 MachineInstr *MI = new MachineInstr(Opcode, NumOperands+1, true, true);
34 I = MBB->insert(I, MI)+1;
35 return MachineInstrBuilder(MI).addReg(DestReg, MOTy::Def);
38 /// BMI - A special BuildMI variant that takes an iterator to insert the
39 /// instruction at as well as a basic block.
40 inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB,
41 MachineBasicBlock::iterator &I,
42 int Opcode, unsigned NumOperands) {
43 assert(I >= MBB->begin() && I <= MBB->end() && "Bad iterator!");
44 MachineInstr *MI = new MachineInstr(Opcode, NumOperands, true, true);
45 I = MBB->insert(I, MI)+1;
46 return MachineInstrBuilder(MI);
51 struct ISel : public FunctionPass, InstVisitor<ISel> {
53 MachineFunction *F; // The function we are compiling into
54 MachineBasicBlock *BB; // The current MBB we are compiling
55 int VarArgsFrameIndex; // FrameIndex for start of varargs area
57 std::map<Value*, unsigned> RegMap; // Mapping between Val's and SSA Regs
59 // MBBMap - Mapping between LLVM BB -> Machine BB
60 std::map<const BasicBlock*, MachineBasicBlock*> MBBMap;
62 ISel(TargetMachine &tm) : TM(tm), F(0), BB(0) {}
64 /// runOnFunction - Top level implementation of instruction selection for
65 /// the entire function.
67 bool runOnFunction(Function &Fn) {
68 F = &MachineFunction::construct(&Fn, TM);
70 // Create all of the machine basic blocks for the function...
71 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
72 F->getBasicBlockList().push_back(MBBMap[I] = new MachineBasicBlock(I));
76 // Copy incoming arguments off of the stack...
77 LoadArgumentsToVirtualRegs(Fn);
79 // Instruction select everything except PHI nodes
82 // Select the PHI nodes
88 return false; // We never modify the LLVM itself.
91 virtual const char *getPassName() const {
92 return "X86 Simple Instruction Selection";
95 /// visitBasicBlock - This method is called when we are visiting a new basic
96 /// block. This simply creates a new MachineBasicBlock to emit code into
97 /// and adds it to the current MachineFunction. Subsequent visit* for
98 /// instructions will be invoked for all instructions in the basic block.
100 void visitBasicBlock(BasicBlock &LLVM_BB) {
101 BB = MBBMap[&LLVM_BB];
104 /// LoadArgumentsToVirtualRegs - Load all of the arguments to this function
105 /// from the stack into virtual registers.
107 void LoadArgumentsToVirtualRegs(Function &F);
109 /// SelectPHINodes - Insert machine code to generate phis. This is tricky
110 /// because we have to generate our sources into the source basic blocks,
111 /// not the current one.
113 void SelectPHINodes();
115 // Visitation methods for various instructions. These methods simply emit
116 // fixed X86 code for each instruction.
119 // Control flow operators
120 void visitReturnInst(ReturnInst &RI);
121 void visitBranchInst(BranchInst &BI);
126 ValueRecord(unsigned R, const Type *T) : Reg(R), Ty(T) {}
128 void doCall(const ValueRecord &Ret, MachineInstr *CallMI,
129 const std::vector<ValueRecord> &Args);
130 void visitCallInst(CallInst &I);
131 void visitIntrinsicCall(LLVMIntrinsic::ID ID, CallInst &I);
133 // Arithmetic operators
134 void visitSimpleBinary(BinaryOperator &B, unsigned OpcodeClass);
135 void visitAdd(BinaryOperator &B) { visitSimpleBinary(B, 0); }
136 void visitSub(BinaryOperator &B) { visitSimpleBinary(B, 1); }
137 void doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI,
138 unsigned DestReg, const Type *DestTy,
139 unsigned Op0Reg, unsigned Op1Reg);
140 void visitMul(BinaryOperator &B);
142 void visitDiv(BinaryOperator &B) { visitDivRem(B); }
143 void visitRem(BinaryOperator &B) { visitDivRem(B); }
144 void visitDivRem(BinaryOperator &B);
147 void visitAnd(BinaryOperator &B) { visitSimpleBinary(B, 2); }
148 void visitOr (BinaryOperator &B) { visitSimpleBinary(B, 3); }
149 void visitXor(BinaryOperator &B) { visitSimpleBinary(B, 4); }
151 // Comparison operators...
152 void visitSetCondInst(SetCondInst &I);
153 bool EmitComparisonGetSignedness(unsigned OpNum, Value *Op0, Value *Op1);
155 // Memory Instructions
156 MachineInstr *doFPLoad(MachineBasicBlock *MBB,
157 MachineBasicBlock::iterator &MBBI,
158 const Type *Ty, unsigned DestReg);
159 void visitLoadInst(LoadInst &I);
160 void doFPStore(const Type *Ty, unsigned DestAddrReg, unsigned SrcReg);
161 void visitStoreInst(StoreInst &I);
162 void visitGetElementPtrInst(GetElementPtrInst &I);
163 void visitAllocaInst(AllocaInst &I);
164 void visitMallocInst(MallocInst &I);
165 void visitFreeInst(FreeInst &I);
168 void visitShiftInst(ShiftInst &I);
169 void visitPHINode(PHINode &I) {} // PHI nodes handled by second pass
170 void visitCastInst(CastInst &I);
171 void visitVarArgInst(VarArgInst &I);
173 void visitInstruction(Instruction &I) {
174 std::cerr << "Cannot instruction select: " << I;
178 /// promote32 - Make a value 32-bits wide, and put it somewhere.
180 void promote32(unsigned targetReg, const ValueRecord &VR);
182 /// EmitByteSwap - Byteswap SrcReg into DestReg.
184 void EmitByteSwap(unsigned DestReg, unsigned SrcReg, unsigned Class);
186 /// emitGEPOperation - Common code shared between visitGetElementPtrInst and
187 /// constant expression GEP support.
189 void emitGEPOperation(MachineBasicBlock *BB, MachineBasicBlock::iterator&IP,
190 Value *Src, User::op_iterator IdxBegin,
191 User::op_iterator IdxEnd, unsigned TargetReg);
193 /// emitCastOperation - Common code shared between visitCastInst and
194 /// constant expression cast support.
195 void emitCastOperation(MachineBasicBlock *BB,MachineBasicBlock::iterator&IP,
196 Value *Src, const Type *DestTy, unsigned TargetReg);
198 /// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary
199 /// and constant expression support.
200 void emitSimpleBinaryOperation(MachineBasicBlock *BB,
201 MachineBasicBlock::iterator &IP,
202 Value *Op0, Value *Op1,
203 unsigned OperatorClass, unsigned TargetReg);
205 /// copyConstantToRegister - Output the instructions required to put the
206 /// specified constant into the specified register.
208 void copyConstantToRegister(MachineBasicBlock *MBB,
209 MachineBasicBlock::iterator &MBBI,
210 Constant *C, unsigned Reg);
212 /// makeAnotherReg - This method returns the next register number we haven't
215 /// Long values are handled somewhat specially. They are always allocated
216 /// as pairs of 32 bit integer values. The register number returned is the
217 /// lower 32 bits of the long value, and the regNum+1 is the upper 32 bits
218 /// of the long value.
220 unsigned makeAnotherReg(const Type *Ty) {
221 if (Ty == Type::LongTy || Ty == Type::ULongTy) {
222 const TargetRegisterClass *RC =
223 TM.getRegisterInfo()->getRegClassForType(Type::IntTy);
224 // Create the lower part
225 F->getSSARegMap()->createVirtualRegister(RC);
226 // Create the upper part.
227 return F->getSSARegMap()->createVirtualRegister(RC)-1;
230 // Add the mapping of regnumber => reg class to MachineFunction
231 const TargetRegisterClass *RC =
232 TM.getRegisterInfo()->getRegClassForType(Ty);
233 return F->getSSARegMap()->createVirtualRegister(RC);
236 /// getReg - This method turns an LLVM value into a register number. This
237 /// is guaranteed to produce the same register number for a particular value
238 /// every time it is queried.
240 unsigned getReg(Value &V) { return getReg(&V); } // Allow references
241 unsigned getReg(Value *V) {
242 // Just append to the end of the current bb.
243 MachineBasicBlock::iterator It = BB->end();
244 return getReg(V, BB, It);
246 unsigned getReg(Value *V, MachineBasicBlock *MBB,
247 MachineBasicBlock::iterator &IPt) {
248 unsigned &Reg = RegMap[V];
250 Reg = makeAnotherReg(V->getType());
254 // If this operand is a constant, emit the code to copy the constant into
255 // the register here...
257 if (Constant *C = dyn_cast<Constant>(V)) {
258 copyConstantToRegister(MBB, IPt, C, Reg);
259 RegMap.erase(V); // Assign a new name to this constant if ref'd again
260 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
261 // Move the address of the global into the register
262 BMI(MBB, IPt, X86::MOVir32, 1, Reg).addGlobalAddress(GV);
263 RegMap.erase(V); // Assign a new name to this address if ref'd again
271 /// TypeClass - Used by the X86 backend to group LLVM types by their basic X86
275 cByte, cShort, cInt, cFP, cLong
278 /// getClass - Turn a primitive type into a "class" number which is based on the
279 /// size of the type, and whether or not it is floating point.
281 static inline TypeClass getClass(const Type *Ty) {
282 switch (Ty->getPrimitiveID()) {
283 case Type::SByteTyID:
284 case Type::UByteTyID: return cByte; // Byte operands are class #0
285 case Type::ShortTyID:
286 case Type::UShortTyID: return cShort; // Short operands are class #1
289 case Type::PointerTyID: return cInt; // Int's and pointers are class #2
291 case Type::FloatTyID:
292 case Type::DoubleTyID: return cFP; // Floating Point is #3
295 case Type::ULongTyID: return cLong; // Longs are class #4
297 assert(0 && "Invalid type to getClass!");
298 return cByte; // not reached
302 // getClassB - Just like getClass, but treat boolean values as bytes.
303 static inline TypeClass getClassB(const Type *Ty) {
304 if (Ty == Type::BoolTy) return cByte;
309 /// copyConstantToRegister - Output the instructions required to put the
310 /// specified constant into the specified register.
312 void ISel::copyConstantToRegister(MachineBasicBlock *MBB,
313 MachineBasicBlock::iterator &IP,
314 Constant *C, unsigned R) {
315 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
317 switch (CE->getOpcode()) {
318 case Instruction::GetElementPtr:
319 emitGEPOperation(MBB, IP, CE->getOperand(0),
320 CE->op_begin()+1, CE->op_end(), R);
322 case Instruction::Cast:
323 emitCastOperation(MBB, IP, CE->getOperand(0), CE->getType(), R);
326 case Instruction::Xor: ++Class; // FALL THROUGH
327 case Instruction::Or: ++Class; // FALL THROUGH
328 case Instruction::And: ++Class; // FALL THROUGH
329 case Instruction::Sub: ++Class; // FALL THROUGH
330 case Instruction::Add:
331 emitSimpleBinaryOperation(MBB, IP, CE->getOperand(0), CE->getOperand(1),
336 std::cerr << "Offending expr: " << C << "\n";
337 assert(0 && "Constant expressions not yet handled!\n");
341 if (C->getType()->isIntegral()) {
342 unsigned Class = getClassB(C->getType());
344 if (Class == cLong) {
345 // Copy the value into the register pair.
347 if (C->getType()->isSigned())
348 Val = cast<ConstantSInt>(C)->getValue();
350 Val = cast<ConstantUInt>(C)->getValue();
352 BMI(MBB, IP, X86::MOVir32, 1, R).addZImm(Val & 0xFFFFFFFF);
353 BMI(MBB, IP, X86::MOVir32, 1, R+1).addZImm(Val >> 32);
357 assert(Class <= cInt && "Type not handled yet!");
359 static const unsigned IntegralOpcodeTab[] = {
360 X86::MOVir8, X86::MOVir16, X86::MOVir32
363 if (C->getType() == Type::BoolTy) {
364 BMI(MBB, IP, X86::MOVir8, 1, R).addZImm(C == ConstantBool::True);
365 } else if (C->getType()->isSigned()) {
366 ConstantSInt *CSI = cast<ConstantSInt>(C);
367 BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addZImm(CSI->getValue());
369 ConstantUInt *CUI = cast<ConstantUInt>(C);
370 BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addZImm(CUI->getValue());
372 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
373 double Value = CFP->getValue();
375 BMI(MBB, IP, X86::FLD0, 0, R);
376 else if (Value == +1.0)
377 BMI(MBB, IP, X86::FLD1, 0, R);
379 // Otherwise we need to spill the constant to memory...
380 MachineConstantPool *CP = F->getConstantPool();
381 unsigned CPI = CP->getConstantPoolIndex(CFP);
382 addConstantPoolReference(doFPLoad(MBB, IP, CFP->getType(), R), CPI);
385 } else if (isa<ConstantPointerNull>(C)) {
386 // Copy zero (null pointer) to the register.
387 BMI(MBB, IP, X86::MOVir32, 1, R).addZImm(0);
388 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
389 unsigned SrcReg = getReg(CPR->getValue(), MBB, IP);
390 BMI(MBB, IP, X86::MOVrr32, 1, R).addReg(SrcReg);
392 std::cerr << "Offending constant: " << C << "\n";
393 assert(0 && "Type not handled yet!");
397 /// LoadArgumentsToVirtualRegs - Load all of the arguments to this function from
398 /// the stack into virtual registers.
400 void ISel::LoadArgumentsToVirtualRegs(Function &Fn) {
401 // Emit instructions to load the arguments... On entry to a function on the
402 // X86, the stack frame looks like this:
404 // [ESP] -- return address
405 // [ESP + 4] -- first argument (leftmost lexically)
406 // [ESP + 8] -- second argument, if first argument is four bytes in size
409 unsigned ArgOffset = 0; // Frame mechanisms handle retaddr slot
410 MachineFrameInfo *MFI = F->getFrameInfo();
412 for (Function::aiterator I = Fn.abegin(), E = Fn.aend(); I != E; ++I) {
413 unsigned Reg = getReg(*I);
415 int FI; // Frame object index
416 switch (getClassB(I->getType())) {
418 FI = MFI->CreateFixedObject(1, ArgOffset);
419 addFrameReference(BuildMI(BB, X86::MOVmr8, 4, Reg), FI);
422 FI = MFI->CreateFixedObject(2, ArgOffset);
423 addFrameReference(BuildMI(BB, X86::MOVmr16, 4, Reg), FI);
426 FI = MFI->CreateFixedObject(4, ArgOffset);
427 addFrameReference(BuildMI(BB, X86::MOVmr32, 4, Reg), FI);
430 FI = MFI->CreateFixedObject(8, ArgOffset);
431 addFrameReference(BuildMI(BB, X86::MOVmr32, 4, Reg), FI);
432 addFrameReference(BuildMI(BB, X86::MOVmr32, 4, Reg+1), FI, 4);
433 ArgOffset += 4; // longs require 4 additional bytes
437 if (I->getType() == Type::FloatTy) {
438 Opcode = X86::FLDr32;
439 FI = MFI->CreateFixedObject(4, ArgOffset);
441 Opcode = X86::FLDr64;
442 FI = MFI->CreateFixedObject(8, ArgOffset);
443 ArgOffset += 4; // doubles require 4 additional bytes
445 addFrameReference(BuildMI(BB, Opcode, 4, Reg), FI);
448 assert(0 && "Unhandled argument type!");
450 ArgOffset += 4; // Each argument takes at least 4 bytes on the stack...
453 // If the function takes variable number of arguments, add a frame offset for
454 // the start of the first vararg value... this is used to expand
456 if (Fn.getFunctionType()->isVarArg())
457 VarArgsFrameIndex = MFI->CreateFixedObject(1, ArgOffset);
461 /// SelectPHINodes - Insert machine code to generate phis. This is tricky
462 /// because we have to generate our sources into the source basic blocks, not
465 void ISel::SelectPHINodes() {
466 const TargetInstrInfo &TII = TM.getInstrInfo();
467 const Function &LF = *F->getFunction(); // The LLVM function...
468 for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) {
469 const BasicBlock *BB = I;
470 MachineBasicBlock *MBB = MBBMap[I];
472 // Loop over all of the PHI nodes in the LLVM basic block...
473 unsigned NumPHIs = 0;
474 for (BasicBlock::const_iterator I = BB->begin();
475 PHINode *PN = (PHINode*)dyn_cast<PHINode>(I); ++I) {
477 // Create a new machine instr PHI node, and insert it.
478 unsigned PHIReg = getReg(*PN);
479 MachineInstr *PhiMI = BuildMI(X86::PHI, PN->getNumOperands(), PHIReg);
480 MBB->insert(MBB->begin()+NumPHIs++, PhiMI);
482 MachineInstr *LongPhiMI = 0;
483 if (PN->getType() == Type::LongTy || PN->getType() == Type::ULongTy) {
484 LongPhiMI = BuildMI(X86::PHI, PN->getNumOperands(), PHIReg+1);
485 MBB->insert(MBB->begin()+NumPHIs++, LongPhiMI);
488 // PHIValues - Map of blocks to incoming virtual registers. We use this
489 // so that we only initialize one incoming value for a particular block,
490 // even if the block has multiple entries in the PHI node.
492 std::map<MachineBasicBlock*, unsigned> PHIValues;
494 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
495 MachineBasicBlock *PredMBB = MBBMap[PN->getIncomingBlock(i)];
497 std::map<MachineBasicBlock*, unsigned>::iterator EntryIt =
498 PHIValues.lower_bound(PredMBB);
500 if (EntryIt != PHIValues.end() && EntryIt->first == PredMBB) {
501 // We already inserted an initialization of the register for this
502 // predecessor. Recycle it.
503 ValReg = EntryIt->second;
506 // Get the incoming value into a virtual register. If it is not
507 // already available in a virtual register, insert the computation
510 MachineBasicBlock::iterator PI = PredMBB->end();
511 while (PI != PredMBB->begin() &&
512 TII.isTerminatorInstr((*(PI-1))->getOpcode()))
514 ValReg = getReg(PN->getIncomingValue(i), PredMBB, PI);
516 // Remember that we inserted a value for this PHI for this predecessor
517 PHIValues.insert(EntryIt, std::make_pair(PredMBB, ValReg));
520 PhiMI->addRegOperand(ValReg);
521 PhiMI->addMachineBasicBlockOperand(PredMBB);
523 LongPhiMI->addRegOperand(ValReg+1);
524 LongPhiMI->addMachineBasicBlockOperand(PredMBB);
531 // canFoldSetCCIntoBranch - Return the setcc instruction if we can fold it into
532 // the conditional branch instruction which is the only user of the cc
533 // instruction. This is the case if the conditional branch is the only user of
534 // the setcc, and if the setcc is in the same basic block as the conditional
535 // branch. We also don't handle long arguments below, so we reject them here as
538 static SetCondInst *canFoldSetCCIntoBranch(Value *V) {
539 if (SetCondInst *SCI = dyn_cast<SetCondInst>(V))
540 if (SCI->use_size() == 1 && isa<BranchInst>(SCI->use_back()) &&
541 SCI->getParent() == cast<BranchInst>(SCI->use_back())->getParent()) {
542 const Type *Ty = SCI->getOperand(0)->getType();
543 if (Ty != Type::LongTy && Ty != Type::ULongTy)
549 // Return a fixed numbering for setcc instructions which does not depend on the
550 // order of the opcodes.
552 static unsigned getSetCCNumber(unsigned Opcode) {
554 default: assert(0 && "Unknown setcc instruction!");
555 case Instruction::SetEQ: return 0;
556 case Instruction::SetNE: return 1;
557 case Instruction::SetLT: return 2;
558 case Instruction::SetGE: return 3;
559 case Instruction::SetGT: return 4;
560 case Instruction::SetLE: return 5;
564 // LLVM -> X86 signed X86 unsigned
565 // ----- ---------- ------------
566 // seteq -> sete sete
567 // setne -> setne setne
568 // setlt -> setl setb
569 // setge -> setge setae
570 // setgt -> setg seta
571 // setle -> setle setbe
572 static const unsigned SetCCOpcodeTab[2][6] = {
573 {X86::SETEr, X86::SETNEr, X86::SETBr, X86::SETAEr, X86::SETAr, X86::SETBEr},
574 {X86::SETEr, X86::SETNEr, X86::SETLr, X86::SETGEr, X86::SETGr, X86::SETLEr},
577 bool ISel::EmitComparisonGetSignedness(unsigned OpNum, Value *Op0, Value *Op1) {
579 // The arguments are already supposed to be of the same type.
580 const Type *CompTy = Op0->getType();
581 bool isSigned = CompTy->isSigned();
582 unsigned Class = getClassB(CompTy);
583 unsigned Op0r = getReg(Op0);
585 // Special case handling of: cmp R, i
586 if (Class == cByte || Class == cShort || Class == cInt)
587 if (ConstantInt *CI = dyn_cast<ConstantInt>(Op1)) {
589 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(CI))
590 Op1v = CSI->getValue();
592 Op1v = cast<ConstantUInt>(CI)->getValue();
593 // Mask off any upper bits of the constant, if there are any...
594 Op1v &= (1ULL << (8 << Class)) - 1;
597 case cByte: BuildMI(BB, X86::CMPri8, 2).addReg(Op0r).addZImm(Op1v);break;
598 case cShort: BuildMI(BB, X86::CMPri16,2).addReg(Op0r).addZImm(Op1v);break;
599 case cInt: BuildMI(BB, X86::CMPri32,2).addReg(Op0r).addZImm(Op1v);break;
601 assert(0 && "Invalid class!");
606 unsigned Op1r = getReg(Op1);
608 default: assert(0 && "Unknown type class!");
609 // Emit: cmp <var1>, <var2> (do the comparison). We can
610 // compare 8-bit with 8-bit, 16-bit with 16-bit, 32-bit with
613 BuildMI(BB, X86::CMPrr8, 2).addReg(Op0r).addReg(Op1r);
616 BuildMI(BB, X86::CMPrr16, 2).addReg(Op0r).addReg(Op1r);
619 BuildMI(BB, X86::CMPrr32, 2).addReg(Op0r).addReg(Op1r);
622 BuildMI(BB, X86::FpUCOM, 2).addReg(Op0r).addReg(Op1r);
623 BuildMI(BB, X86::FNSTSWr8, 0);
624 BuildMI(BB, X86::SAHF, 1);
625 isSigned = false; // Compare with unsigned operators
629 if (OpNum < 2) { // seteq, setne
630 unsigned LoTmp = makeAnotherReg(Type::IntTy);
631 unsigned HiTmp = makeAnotherReg(Type::IntTy);
632 unsigned FinalTmp = makeAnotherReg(Type::IntTy);
633 BuildMI(BB, X86::XORrr32, 2, LoTmp).addReg(Op0r).addReg(Op1r);
634 BuildMI(BB, X86::XORrr32, 2, HiTmp).addReg(Op0r+1).addReg(Op1r+1);
635 BuildMI(BB, X86::ORrr32, 2, FinalTmp).addReg(LoTmp).addReg(HiTmp);
636 break; // Allow the sete or setne to be generated from flags set by OR
638 // Emit a sequence of code which compares the high and low parts once
639 // each, then uses a conditional move to handle the overflow case. For
640 // example, a setlt for long would generate code like this:
642 // AL = lo(op1) < lo(op2) // Signedness depends on operands
643 // BL = hi(op1) < hi(op2) // Always unsigned comparison
644 // dest = hi(op1) == hi(op2) ? AL : BL;
647 // FIXME: This would be much better if we had hierarchical register
648 // classes! Until then, hardcode registers so that we can deal with their
649 // aliases (because we don't have conditional byte moves).
651 BuildMI(BB, X86::CMPrr32, 2).addReg(Op0r).addReg(Op1r);
652 BuildMI(BB, SetCCOpcodeTab[0][OpNum], 0, X86::AL);
653 BuildMI(BB, X86::CMPrr32, 2).addReg(Op0r+1).addReg(Op1r+1);
654 BuildMI(BB, SetCCOpcodeTab[isSigned][OpNum], 0, X86::BL);
655 BuildMI(BB, X86::CMOVErr16, 2, X86::BX).addReg(X86::BX).addReg(X86::AX);
656 // NOTE: visitSetCondInst knows that the value is dumped into the BL
657 // register at this point for long values...
665 /// SetCC instructions - Here we just emit boilerplate code to set a byte-sized
666 /// register, then move it to wherever the result should be.
668 void ISel::visitSetCondInst(SetCondInst &I) {
669 if (canFoldSetCCIntoBranch(&I)) return; // Fold this into a branch...
671 unsigned OpNum = getSetCCNumber(I.getOpcode());
672 unsigned DestReg = getReg(I);
673 bool isSigned = EmitComparisonGetSignedness(OpNum, I.getOperand(0),
676 if (getClassB(I.getOperand(0)->getType()) != cLong || OpNum < 2) {
677 // Handle normal comparisons with a setcc instruction...
678 BuildMI(BB, SetCCOpcodeTab[isSigned][OpNum], 0, DestReg);
680 // Handle long comparisons by copying the value which is already in BL into
681 // the register we want...
682 BuildMI(BB, X86::MOVrr8, 1, DestReg).addReg(X86::BL);
686 /// promote32 - Emit instructions to turn a narrow operand into a 32-bit-wide
687 /// operand, in the specified target register.
688 void ISel::promote32(unsigned targetReg, const ValueRecord &VR) {
689 bool isUnsigned = VR.Ty->isUnsigned();
690 switch (getClassB(VR.Ty)) {
692 // Extend value into target register (8->32)
694 BuildMI(BB, X86::MOVZXr32r8, 1, targetReg).addReg(VR.Reg);
696 BuildMI(BB, X86::MOVSXr32r8, 1, targetReg).addReg(VR.Reg);
699 // Extend value into target register (16->32)
701 BuildMI(BB, X86::MOVZXr32r16, 1, targetReg).addReg(VR.Reg);
703 BuildMI(BB, X86::MOVSXr32r16, 1, targetReg).addReg(VR.Reg);
706 // Move value into target register (32->32)
707 BuildMI(BB, X86::MOVrr32, 1, targetReg).addReg(VR.Reg);
710 assert(0 && "Unpromotable operand class in promote32");
714 /// 'ret' instruction - Here we are interested in meeting the x86 ABI. As such,
715 /// we have the following possibilities:
717 /// ret void: No return value, simply emit a 'ret' instruction
718 /// ret sbyte, ubyte : Extend value into EAX and return
719 /// ret short, ushort: Extend value into EAX and return
720 /// ret int, uint : Move value into EAX and return
721 /// ret pointer : Move value into EAX and return
722 /// ret long, ulong : Move value into EAX/EDX and return
723 /// ret float/double : Top of FP stack
725 void ISel::visitReturnInst(ReturnInst &I) {
726 if (I.getNumOperands() == 0) {
727 BuildMI(BB, X86::RET, 0); // Just emit a 'ret' instruction
731 Value *RetVal = I.getOperand(0);
732 unsigned RetReg = getReg(RetVal);
733 switch (getClassB(RetVal->getType())) {
734 case cByte: // integral return values: extend or move into EAX and return
737 promote32(X86::EAX, ValueRecord(RetReg, RetVal->getType()));
738 // Declare that EAX is live on exit
739 BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::EAX).addReg(X86::ESP);
741 case cFP: // Floats & Doubles: Return in ST(0)
742 BuildMI(BB, X86::FpSETRESULT, 1).addReg(RetReg);
743 // Declare that top-of-stack is live on exit
744 BuildMI(BB, X86::IMPLICIT_USE, 2).addReg(X86::ST0).addReg(X86::ESP);
747 BuildMI(BB, X86::MOVrr32, 1, X86::EAX).addReg(RetReg);
748 BuildMI(BB, X86::MOVrr32, 1, X86::EDX).addReg(RetReg+1);
749 // Declare that EAX & EDX are live on exit
750 BuildMI(BB, X86::IMPLICIT_USE, 3).addReg(X86::EAX).addReg(X86::EDX).addReg(X86::ESP);
755 // Emit a 'ret' instruction
756 BuildMI(BB, X86::RET, 0);
759 // getBlockAfter - Return the basic block which occurs lexically after the
761 static inline BasicBlock *getBlockAfter(BasicBlock *BB) {
762 Function::iterator I = BB; ++I; // Get iterator to next block
763 return I != BB->getParent()->end() ? &*I : 0;
766 /// visitBranchInst - Handle conditional and unconditional branches here. Note
767 /// that since code layout is frozen at this point, that if we are trying to
768 /// jump to a block that is the immediate successor of the current block, we can
769 /// just make a fall-through (but we don't currently).
771 void ISel::visitBranchInst(BranchInst &BI) {
772 BasicBlock *NextBB = getBlockAfter(BI.getParent()); // BB after current one
774 if (!BI.isConditional()) { // Unconditional branch?
775 if (BI.getSuccessor(0) != NextBB)
776 BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0));
780 // See if we can fold the setcc into the branch itself...
781 SetCondInst *SCI = canFoldSetCCIntoBranch(BI.getCondition());
783 // Nope, cannot fold setcc into this branch. Emit a branch on a condition
784 // computed some other way...
785 unsigned condReg = getReg(BI.getCondition());
786 BuildMI(BB, X86::CMPri8, 2).addReg(condReg).addZImm(0);
787 if (BI.getSuccessor(1) == NextBB) {
788 if (BI.getSuccessor(0) != NextBB)
789 BuildMI(BB, X86::JNE, 1).addPCDisp(BI.getSuccessor(0));
791 BuildMI(BB, X86::JE, 1).addPCDisp(BI.getSuccessor(1));
793 if (BI.getSuccessor(0) != NextBB)
794 BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(0));
799 unsigned OpNum = getSetCCNumber(SCI->getOpcode());
800 bool isSigned = EmitComparisonGetSignedness(OpNum, SCI->getOperand(0),
803 // LLVM -> X86 signed X86 unsigned
804 // ----- ---------- ------------
811 static const unsigned OpcodeTab[2][6] = {
812 { X86::JE, X86::JNE, X86::JB, X86::JAE, X86::JA, X86::JBE },
813 { X86::JE, X86::JNE, X86::JL, X86::JGE, X86::JG, X86::JLE },
816 if (BI.getSuccessor(0) != NextBB) {
817 BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(0));
818 if (BI.getSuccessor(1) != NextBB)
819 BuildMI(BB, X86::JMP, 1).addPCDisp(BI.getSuccessor(1));
821 // Change to the inverse condition...
822 if (BI.getSuccessor(1) != NextBB) {
824 BuildMI(BB, OpcodeTab[isSigned][OpNum], 1).addPCDisp(BI.getSuccessor(1));
830 /// doCall - This emits an abstract call instruction, setting up the arguments
831 /// and the return value as appropriate. For the actual function call itself,
832 /// it inserts the specified CallMI instruction into the stream.
834 void ISel::doCall(const ValueRecord &Ret, MachineInstr *CallMI,
835 const std::vector<ValueRecord> &Args) {
837 // Count how many bytes are to be pushed on the stack...
838 unsigned NumBytes = 0;
841 for (unsigned i = 0, e = Args.size(); i != e; ++i)
842 switch (getClassB(Args[i].Ty)) {
843 case cByte: case cShort: case cInt:
844 NumBytes += 4; break;
846 NumBytes += 8; break;
848 NumBytes += Args[i].Ty == Type::FloatTy ? 4 : 8;
850 default: assert(0 && "Unknown class!");
853 // Adjust the stack pointer for the new arguments...
854 BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addZImm(NumBytes);
856 // Arguments go on the stack in reverse order, as specified by the ABI.
857 unsigned ArgOffset = 0;
858 for (unsigned i = 0, e = Args.size(); i != e; ++i) {
859 unsigned ArgReg = Args[i].Reg;
860 switch (getClassB(Args[i].Ty)) {
863 // Promote arg to 32 bits wide into a temporary register...
864 unsigned R = makeAnotherReg(Type::UIntTy);
865 promote32(R, Args[i]);
866 addRegOffset(BuildMI(BB, X86::MOVrm32, 5),
867 X86::ESP, ArgOffset).addReg(R);
871 addRegOffset(BuildMI(BB, X86::MOVrm32, 5),
872 X86::ESP, ArgOffset).addReg(ArgReg);
875 addRegOffset(BuildMI(BB, X86::MOVrm32, 5),
876 X86::ESP, ArgOffset).addReg(ArgReg);
877 addRegOffset(BuildMI(BB, X86::MOVrm32, 5),
878 X86::ESP, ArgOffset+4).addReg(ArgReg+1);
879 ArgOffset += 4; // 8 byte entry, not 4.
883 if (Args[i].Ty == Type::FloatTy) {
884 addRegOffset(BuildMI(BB, X86::FSTr32, 5),
885 X86::ESP, ArgOffset).addReg(ArgReg);
887 assert(Args[i].Ty == Type::DoubleTy && "Unknown FP type!");
888 addRegOffset(BuildMI(BB, X86::FSTr64, 5),
889 X86::ESP, ArgOffset).addReg(ArgReg);
890 ArgOffset += 4; // 8 byte entry, not 4.
894 default: assert(0 && "Unknown class!");
899 BuildMI(BB, X86::ADJCALLSTACKDOWN, 1).addZImm(0);
902 BB->push_back(CallMI);
904 BuildMI(BB, X86::ADJCALLSTACKUP, 1).addZImm(NumBytes);
906 // If there is a return value, scavenge the result from the location the call
909 if (Ret.Ty != Type::VoidTy) {
910 unsigned DestClass = getClassB(Ret.Ty);
915 // Integral results are in %eax, or the appropriate portion
917 static const unsigned regRegMove[] = {
918 X86::MOVrr8, X86::MOVrr16, X86::MOVrr32
920 static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX };
921 BuildMI(BB, regRegMove[DestClass], 1, Ret.Reg).addReg(AReg[DestClass]);
924 case cFP: // Floating-point return values live in %ST(0)
925 BuildMI(BB, X86::FpGETRESULT, 1, Ret.Reg);
927 case cLong: // Long values are left in EDX:EAX
928 BuildMI(BB, X86::MOVrr32, 1, Ret.Reg).addReg(X86::EAX);
929 BuildMI(BB, X86::MOVrr32, 1, Ret.Reg+1).addReg(X86::EDX);
931 default: assert(0 && "Unknown class!");
937 /// visitCallInst - Push args on stack and do a procedure call instruction.
938 void ISel::visitCallInst(CallInst &CI) {
939 MachineInstr *TheCall;
940 if (Function *F = CI.getCalledFunction()) {
941 // Is it an intrinsic function call?
942 if (LLVMIntrinsic::ID ID = (LLVMIntrinsic::ID)F->getIntrinsicID()) {
943 visitIntrinsicCall(ID, CI); // Special intrinsics are not handled here
947 // Emit a CALL instruction with PC-relative displacement.
948 TheCall = BuildMI(X86::CALLpcrel32, 1).addGlobalAddress(F, true);
949 } else { // Emit an indirect call...
950 unsigned Reg = getReg(CI.getCalledValue());
951 TheCall = BuildMI(X86::CALLr32, 1).addReg(Reg);
954 std::vector<ValueRecord> Args;
955 for (unsigned i = 1, e = CI.getNumOperands(); i != e; ++i)
956 Args.push_back(ValueRecord(getReg(CI.getOperand(i)),
957 CI.getOperand(i)->getType()));
959 unsigned DestReg = CI.getType() != Type::VoidTy ? getReg(CI) : 0;
960 doCall(ValueRecord(DestReg, CI.getType()), TheCall, Args);
963 void ISel::visitIntrinsicCall(LLVMIntrinsic::ID ID, CallInst &CI) {
964 unsigned TmpReg1, TmpReg2;
966 case LLVMIntrinsic::va_start:
967 // Get the address of the first vararg value...
968 TmpReg1 = makeAnotherReg(Type::UIntTy);
969 addFrameReference(BuildMI(BB, X86::LEAr32, 5, TmpReg1), VarArgsFrameIndex);
970 TmpReg2 = getReg(CI.getOperand(1));
971 addDirectMem(BuildMI(BB, X86::MOVrm32, 5), TmpReg2).addReg(TmpReg1);
974 case LLVMIntrinsic::va_end: return; // Noop on X86
975 case LLVMIntrinsic::va_copy:
976 TmpReg1 = getReg(CI.getOperand(2)); // Get existing va_list
977 TmpReg2 = getReg(CI.getOperand(1)); // Get va_list* to store into
978 addDirectMem(BuildMI(BB, X86::MOVrm32, 5), TmpReg2).addReg(TmpReg1);
981 case LLVMIntrinsic::longjmp:
982 BuildMI(X86::CALLpcrel32, 1).addExternalSymbol("abort", true);
983 case LLVMIntrinsic::setjmp:
984 // Setjmp always returns zero...
985 BuildMI(BB, X86::MOVir32, 1, getReg(CI)).addZImm(0);
987 default: assert(0 && "Unknown intrinsic for X86!");
992 /// visitSimpleBinary - Implement simple binary operators for integral types...
993 /// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or, 4 for
995 void ISel::visitSimpleBinary(BinaryOperator &B, unsigned OperatorClass) {
996 unsigned DestReg = getReg(B);
997 MachineBasicBlock::iterator MI = BB->end();
998 emitSimpleBinaryOperation(BB, MI, B.getOperand(0), B.getOperand(1),
999 OperatorClass, DestReg);
1002 /// visitSimpleBinary - Implement simple binary operators for integral types...
1003 /// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or,
1006 /// emitSimpleBinaryOperation - Common code shared between visitSimpleBinary
1007 /// and constant expression support.
1008 void ISel::emitSimpleBinaryOperation(MachineBasicBlock *BB,
1009 MachineBasicBlock::iterator &IP,
1010 Value *Op0, Value *Op1,
1011 unsigned OperatorClass,unsigned TargetReg){
1012 unsigned Class = getClassB(Op0->getType());
1013 if (!isa<ConstantInt>(Op1) || Class == cLong) {
1014 static const unsigned OpcodeTab[][4] = {
1015 // Arithmetic operators
1016 { X86::ADDrr8, X86::ADDrr16, X86::ADDrr32, X86::FpADD }, // ADD
1017 { X86::SUBrr8, X86::SUBrr16, X86::SUBrr32, X86::FpSUB }, // SUB
1019 // Bitwise operators
1020 { X86::ANDrr8, X86::ANDrr16, X86::ANDrr32, 0 }, // AND
1021 { X86:: ORrr8, X86:: ORrr16, X86:: ORrr32, 0 }, // OR
1022 { X86::XORrr8, X86::XORrr16, X86::XORrr32, 0 }, // XOR
1025 bool isLong = false;
1026 if (Class == cLong) {
1028 Class = cInt; // Bottom 32 bits are handled just like ints
1031 unsigned Opcode = OpcodeTab[OperatorClass][Class];
1032 assert(Opcode && "Floating point arguments to logical inst?");
1033 unsigned Op0r = getReg(Op0, BB, IP);
1034 unsigned Op1r = getReg(Op1, BB, IP);
1035 BMI(BB, IP, Opcode, 2, TargetReg).addReg(Op0r).addReg(Op1r);
1037 if (isLong) { // Handle the upper 32 bits of long values...
1038 static const unsigned TopTab[] = {
1039 X86::ADCrr32, X86::SBBrr32, X86::ANDrr32, X86::ORrr32, X86::XORrr32
1041 BMI(BB, IP, TopTab[OperatorClass], 2,
1042 TargetReg+1).addReg(Op0r+1).addReg(Op1r+1);
1045 // Special case: op Reg, <const>
1046 ConstantInt *Op1C = cast<ConstantInt>(Op1);
1048 static const unsigned OpcodeTab[][3] = {
1049 // Arithmetic operators
1050 { X86::ADDri8, X86::ADDri16, X86::ADDri32 }, // ADD
1051 { X86::SUBri8, X86::SUBri16, X86::SUBri32 }, // SUB
1053 // Bitwise operators
1054 { X86::ANDri8, X86::ANDri16, X86::ANDri32 }, // AND
1055 { X86:: ORri8, X86:: ORri16, X86:: ORri32 }, // OR
1056 { X86::XORri8, X86::XORri16, X86::XORri32 }, // XOR
1059 assert(Class < 3 && "General code handles 64-bit integer types!");
1060 unsigned Opcode = OpcodeTab[OperatorClass][Class];
1061 unsigned Op0r = getReg(Op0, BB, IP);
1063 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(Op1C))
1064 Op1v = CSI->getValue();
1066 Op1v = cast<ConstantUInt>(Op1C)->getValue();
1068 // Mask off any upper bits of the constant, if there are any...
1069 Op1v &= (1ULL << (8 << Class)) - 1;
1070 BMI(BB, IP, Opcode, 2, TargetReg).addReg(Op0r).addZImm(Op1v);
1074 /// doMultiply - Emit appropriate instructions to multiply together the
1075 /// registers op0Reg and op1Reg, and put the result in DestReg. The type of the
1076 /// result should be given as DestTy.
1078 void ISel::doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI,
1079 unsigned DestReg, const Type *DestTy,
1080 unsigned op0Reg, unsigned op1Reg) {
1081 unsigned Class = getClass(DestTy);
1083 case cFP: // Floating point multiply
1084 BMI(BB, MBBI, X86::FpMUL, 2, DestReg).addReg(op0Reg).addReg(op1Reg);
1088 BMI(BB, MBBI, Class == cInt ? X86::IMULr32 : X86::IMULr16, 2, DestReg)
1089 .addReg(op0Reg).addReg(op1Reg);
1092 // Must use the MUL instruction, which forces use of AL...
1093 BMI(MBB, MBBI, X86::MOVrr8, 1, X86::AL).addReg(op0Reg);
1094 BMI(MBB, MBBI, X86::MULr8, 1).addReg(op1Reg);
1095 BMI(MBB, MBBI, X86::MOVrr8, 1, DestReg).addReg(X86::AL);
1098 case cLong: assert(0 && "doMultiply cannot operate on LONG values!");
1102 /// visitMul - Multiplies are not simple binary operators because they must deal
1103 /// with the EAX register explicitly.
1105 void ISel::visitMul(BinaryOperator &I) {
1106 unsigned Op0Reg = getReg(I.getOperand(0));
1107 unsigned Op1Reg = getReg(I.getOperand(1));
1108 unsigned DestReg = getReg(I);
1110 // Simple scalar multiply?
1111 if (I.getType() != Type::LongTy && I.getType() != Type::ULongTy) {
1112 MachineBasicBlock::iterator MBBI = BB->end();
1113 doMultiply(BB, MBBI, DestReg, I.getType(), Op0Reg, Op1Reg);
1115 // Long value. We have to do things the hard way...
1116 // Multiply the two low parts... capturing carry into EDX
1117 BuildMI(BB, X86::MOVrr32, 1, X86::EAX).addReg(Op0Reg);
1118 BuildMI(BB, X86::MULr32, 1).addReg(Op1Reg); // AL*BL
1120 unsigned OverflowReg = makeAnotherReg(Type::UIntTy);
1121 BuildMI(BB, X86::MOVrr32, 1, DestReg).addReg(X86::EAX); // AL*BL
1122 BuildMI(BB, X86::MOVrr32, 1, OverflowReg).addReg(X86::EDX); // AL*BL >> 32
1124 MachineBasicBlock::iterator MBBI = BB->end();
1125 unsigned AHBLReg = makeAnotherReg(Type::UIntTy); // AH*BL
1126 BMI(BB, MBBI, X86::IMULr32, 2, AHBLReg).addReg(Op0Reg+1).addReg(Op1Reg);
1128 unsigned AHBLplusOverflowReg = makeAnotherReg(Type::UIntTy);
1129 BuildMI(BB, X86::ADDrr32, 2, // AH*BL+(AL*BL >> 32)
1130 AHBLplusOverflowReg).addReg(AHBLReg).addReg(OverflowReg);
1133 unsigned ALBHReg = makeAnotherReg(Type::UIntTy); // AL*BH
1134 BMI(BB, MBBI, X86::IMULr32, 2, ALBHReg).addReg(Op0Reg).addReg(Op1Reg+1);
1136 BuildMI(BB, X86::ADDrr32, 2, // AL*BH + AH*BL + (AL*BL >> 32)
1137 DestReg+1).addReg(AHBLplusOverflowReg).addReg(ALBHReg);
1142 /// visitDivRem - Handle division and remainder instructions... these
1143 /// instruction both require the same instructions to be generated, they just
1144 /// select the result from a different register. Note that both of these
1145 /// instructions work differently for signed and unsigned operands.
1147 void ISel::visitDivRem(BinaryOperator &I) {
1148 unsigned Class = getClass(I.getType());
1149 unsigned Op0Reg = getReg(I.getOperand(0));
1150 unsigned Op1Reg = getReg(I.getOperand(1));
1151 unsigned ResultReg = getReg(I);
1154 case cFP: // Floating point divide
1155 if (I.getOpcode() == Instruction::Div)
1156 BuildMI(BB, X86::FpDIV, 2, ResultReg).addReg(Op0Reg).addReg(Op1Reg);
1157 else { // Floating point remainder...
1158 MachineInstr *TheCall =
1159 BuildMI(X86::CALLpcrel32, 1).addExternalSymbol("fmod", true);
1160 std::vector<ValueRecord> Args;
1161 Args.push_back(ValueRecord(Op0Reg, Type::DoubleTy));
1162 Args.push_back(ValueRecord(Op1Reg, Type::DoubleTy));
1163 doCall(ValueRecord(ResultReg, Type::DoubleTy), TheCall, Args);
1167 static const char *FnName[] =
1168 { "__moddi3", "__divdi3", "__umoddi3", "__udivdi3" };
1170 unsigned NameIdx = I.getType()->isUnsigned()*2;
1171 NameIdx += I.getOpcode() == Instruction::Div;
1172 MachineInstr *TheCall =
1173 BuildMI(X86::CALLpcrel32, 1).addExternalSymbol(FnName[NameIdx], true);
1175 std::vector<ValueRecord> Args;
1176 Args.push_back(ValueRecord(Op0Reg, Type::LongTy));
1177 Args.push_back(ValueRecord(Op1Reg, Type::LongTy));
1178 doCall(ValueRecord(ResultReg, Type::LongTy), TheCall, Args);
1181 case cByte: case cShort: case cInt:
1182 break; // Small integerals, handled below...
1183 default: assert(0 && "Unknown class!");
1186 static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX };
1187 static const unsigned MovOpcode[]={ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 };
1188 static const unsigned SarOpcode[]={ X86::SARir8, X86::SARir16, X86::SARir32 };
1189 static const unsigned ClrOpcode[]={ X86::XORrr8, X86::XORrr16, X86::XORrr32 };
1190 static const unsigned ExtRegs[] ={ X86::AH , X86::DX , X86::EDX };
1192 static const unsigned DivOpcode[][4] = {
1193 { X86::DIVr8 , X86::DIVr16 , X86::DIVr32 , 0 }, // Unsigned division
1194 { X86::IDIVr8, X86::IDIVr16, X86::IDIVr32, 0 }, // Signed division
1197 bool isSigned = I.getType()->isSigned();
1198 unsigned Reg = Regs[Class];
1199 unsigned ExtReg = ExtRegs[Class];
1201 // Put the first operand into one of the A registers...
1202 BuildMI(BB, MovOpcode[Class], 1, Reg).addReg(Op0Reg);
1205 // Emit a sign extension instruction...
1206 unsigned ShiftResult = makeAnotherReg(I.getType());
1207 BuildMI(BB, SarOpcode[Class], 2, ShiftResult).addReg(Op0Reg).addZImm(31);
1208 BuildMI(BB, MovOpcode[Class], 1, ExtReg).addReg(ShiftResult);
1210 // If unsigned, emit a zeroing instruction... (reg = xor reg, reg)
1211 BuildMI(BB, ClrOpcode[Class], 2, ExtReg).addReg(ExtReg).addReg(ExtReg);
1214 // Emit the appropriate divide or remainder instruction...
1215 BuildMI(BB, DivOpcode[isSigned][Class], 1).addReg(Op1Reg);
1217 // Figure out which register we want to pick the result out of...
1218 unsigned DestReg = (I.getOpcode() == Instruction::Div) ? Reg : ExtReg;
1220 // Put the result into the destination register...
1221 BuildMI(BB, MovOpcode[Class], 1, ResultReg).addReg(DestReg);
1225 /// Shift instructions: 'shl', 'sar', 'shr' - Some special cases here
1226 /// for constant immediate shift values, and for constant immediate
1227 /// shift values equal to 1. Even the general case is sort of special,
1228 /// because the shift amount has to be in CL, not just any old register.
1230 void ISel::visitShiftInst(ShiftInst &I) {
1231 unsigned SrcReg = getReg(I.getOperand(0));
1232 unsigned DestReg = getReg(I);
1233 bool isLeftShift = I.getOpcode() == Instruction::Shl;
1234 bool isSigned = I.getType()->isSigned();
1235 unsigned Class = getClass(I.getType());
1237 static const unsigned ConstantOperand[][4] = {
1238 { X86::SHRir8, X86::SHRir16, X86::SHRir32, X86::SHRDir32 }, // SHR
1239 { X86::SARir8, X86::SARir16, X86::SARir32, X86::SHRDir32 }, // SAR
1240 { X86::SHLir8, X86::SHLir16, X86::SHLir32, X86::SHLDir32 }, // SHL
1241 { X86::SHLir8, X86::SHLir16, X86::SHLir32, X86::SHLDir32 }, // SAL = SHL
1244 static const unsigned NonConstantOperand[][4] = {
1245 { X86::SHRrr8, X86::SHRrr16, X86::SHRrr32 }, // SHR
1246 { X86::SARrr8, X86::SARrr16, X86::SARrr32 }, // SAR
1247 { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32 }, // SHL
1248 { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32 }, // SAL = SHL
1251 // Longs, as usual, are handled specially...
1252 if (Class == cLong) {
1253 // If we have a constant shift, we can generate much more efficient code
1254 // than otherwise...
1256 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I.getOperand(1))) {
1257 unsigned Amount = CUI->getValue();
1259 const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned];
1261 BuildMI(BB, Opc[3], 3,
1262 DestReg+1).addReg(SrcReg+1).addReg(SrcReg).addZImm(Amount);
1263 BuildMI(BB, Opc[2], 2, DestReg).addReg(SrcReg).addZImm(Amount);
1265 BuildMI(BB, Opc[3], 3,
1266 DestReg).addReg(SrcReg ).addReg(SrcReg+1).addZImm(Amount);
1267 BuildMI(BB, Opc[2], 2, DestReg+1).addReg(SrcReg+1).addZImm(Amount);
1269 } else { // Shifting more than 32 bits
1272 BuildMI(BB, X86::SHLir32, 2,DestReg+1).addReg(SrcReg).addZImm(Amount);
1273 BuildMI(BB, X86::MOVir32, 1,DestReg ).addZImm(0);
1275 unsigned Opcode = isSigned ? X86::SARir32 : X86::SHRir32;
1276 BuildMI(BB, Opcode, 2, DestReg).addReg(SrcReg+1).addZImm(Amount);
1277 BuildMI(BB, X86::MOVir32, 1, DestReg+1).addZImm(0);
1281 unsigned TmpReg = makeAnotherReg(Type::IntTy);
1283 if (!isLeftShift && isSigned) {
1284 // If this is a SHR of a Long, then we need to do funny sign extension
1285 // stuff. TmpReg gets the value to use as the high-part if we are
1286 // shifting more than 32 bits.
1287 BuildMI(BB, X86::SARir32, 2, TmpReg).addReg(SrcReg).addZImm(31);
1289 // Other shifts use a fixed zero value if the shift is more than 32
1291 BuildMI(BB, X86::MOVir32, 1, TmpReg).addZImm(0);
1294 // Initialize CL with the shift amount...
1295 unsigned ShiftAmount = getReg(I.getOperand(1));
1296 BuildMI(BB, X86::MOVrr8, 1, X86::CL).addReg(ShiftAmount);
1298 unsigned TmpReg2 = makeAnotherReg(Type::IntTy);
1299 unsigned TmpReg3 = makeAnotherReg(Type::IntTy);
1301 // TmpReg2 = shld inHi, inLo
1302 BuildMI(BB, X86::SHLDrr32, 2, TmpReg2).addReg(SrcReg+1).addReg(SrcReg);
1303 // TmpReg3 = shl inLo, CL
1304 BuildMI(BB, X86::SHLrr32, 1, TmpReg3).addReg(SrcReg);
1306 // Set the flags to indicate whether the shift was by more than 32 bits.
1307 BuildMI(BB, X86::TESTri8, 2).addReg(X86::CL).addZImm(32);
1309 // DestHi = (>32) ? TmpReg3 : TmpReg2;
1310 BuildMI(BB, X86::CMOVNErr32, 2,
1311 DestReg+1).addReg(TmpReg2).addReg(TmpReg3);
1312 // DestLo = (>32) ? TmpReg : TmpReg3;
1313 BuildMI(BB, X86::CMOVNErr32, 2, DestReg).addReg(TmpReg3).addReg(TmpReg);
1315 // TmpReg2 = shrd inLo, inHi
1316 BuildMI(BB, X86::SHRDrr32, 2, TmpReg2).addReg(SrcReg).addReg(SrcReg+1);
1317 // TmpReg3 = s[ah]r inHi, CL
1318 BuildMI(BB, isSigned ? X86::SARrr32 : X86::SHRrr32, 1, TmpReg3)
1321 // Set the flags to indicate whether the shift was by more than 32 bits.
1322 BuildMI(BB, X86::TESTri8, 2).addReg(X86::CL).addZImm(32);
1324 // DestLo = (>32) ? TmpReg3 : TmpReg2;
1325 BuildMI(BB, X86::CMOVNErr32, 2,
1326 DestReg).addReg(TmpReg2).addReg(TmpReg3);
1328 // DestHi = (>32) ? TmpReg : TmpReg3;
1329 BuildMI(BB, X86::CMOVNErr32, 2,
1330 DestReg+1).addReg(TmpReg3).addReg(TmpReg);
1336 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I.getOperand(1))) {
1337 // The shift amount is constant, guaranteed to be a ubyte. Get its value.
1338 assert(CUI->getType() == Type::UByteTy && "Shift amount not a ubyte?");
1340 const unsigned *Opc = ConstantOperand[isLeftShift*2+isSigned];
1341 BuildMI(BB, Opc[Class], 2, DestReg).addReg(SrcReg).addZImm(CUI->getValue());
1342 } else { // The shift amount is non-constant.
1343 BuildMI(BB, X86::MOVrr8, 1, X86::CL).addReg(getReg(I.getOperand(1)));
1345 const unsigned *Opc = NonConstantOperand[isLeftShift*2+isSigned];
1346 BuildMI(BB, Opc[Class], 1, DestReg).addReg(SrcReg);
1351 /// doFPLoad - This method is used to load an FP value from memory using the
1352 /// current endianness. NOTE: This method returns a partially constructed load
1353 /// instruction which needs to have the memory source filled in still.
1355 MachineInstr *ISel::doFPLoad(MachineBasicBlock *MBB,
1356 MachineBasicBlock::iterator &MBBI,
1357 const Type *Ty, unsigned DestReg) {
1358 assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
1359 unsigned LoadOpcode = Ty == Type::FloatTy ? X86::FLDr32 : X86::FLDr64;
1361 if (TM.getTargetData().isLittleEndian()) // fast path...
1362 return BMI(MBB, MBBI, LoadOpcode, 4, DestReg);
1364 // If we are big-endian, start by creating an LEA instruction to represent the
1365 // address of the memory location to load from...
1367 unsigned SrcAddrReg = makeAnotherReg(Type::UIntTy);
1368 MachineInstr *Result = BMI(MBB, MBBI, X86::LEAr32, 5, SrcAddrReg);
1370 // Allocate a temporary stack slot to transform the value into...
1371 int FrameIdx = F->getFrameInfo()->CreateStackObject(Ty, TM.getTargetData());
1373 // Perform the bswaps 32 bits at a time...
1374 unsigned TmpReg1 = makeAnotherReg(Type::UIntTy);
1375 unsigned TmpReg2 = makeAnotherReg(Type::UIntTy);
1376 addDirectMem(BMI(MBB, MBBI, X86::MOVmr32, 4, TmpReg1), SrcAddrReg);
1377 BMI(MBB, MBBI, X86::BSWAPr32, 1, TmpReg2).addReg(TmpReg1);
1378 unsigned Offset = (Ty == Type::DoubleTy) << 2;
1379 addFrameReference(BMI(MBB, MBBI, X86::MOVrm32, 5),
1380 FrameIdx, Offset).addReg(TmpReg2);
1382 if (Ty == Type::DoubleTy) { // Swap the other 32 bits of a double value...
1383 TmpReg1 = makeAnotherReg(Type::UIntTy);
1384 TmpReg2 = makeAnotherReg(Type::UIntTy);
1386 addRegOffset(BMI(MBB, MBBI, X86::MOVmr32, 4, TmpReg1), SrcAddrReg, 4);
1387 BMI(MBB, MBBI, X86::BSWAPr32, 1, TmpReg2).addReg(TmpReg1);
1388 unsigned Offset = (Ty == Type::DoubleTy) << 2;
1389 addFrameReference(BMI(MBB, MBBI, X86::MOVrm32,5), FrameIdx).addReg(TmpReg2);
1392 // Now we can reload the final byteswapped result into the final destination.
1393 addFrameReference(BMI(MBB, MBBI, LoadOpcode, 4, DestReg), FrameIdx);
1397 /// EmitByteSwap - Byteswap SrcReg into DestReg.
1399 void ISel::EmitByteSwap(unsigned DestReg, unsigned SrcReg, unsigned Class) {
1400 // Emit the byte swap instruction...
1403 // No byteswap necessary for 8 bit value...
1404 BuildMI(BB, X86::MOVrr8, 1, DestReg).addReg(SrcReg);
1407 // Use the 32 bit bswap instruction to do a 32 bit swap...
1408 BuildMI(BB, X86::BSWAPr32, 1, DestReg).addReg(SrcReg);
1412 // For 16 bit we have to use an xchg instruction, because there is no
1413 // 16-bit bswap. XCHG is necessarily not in SSA form, so we force things
1414 // into AX to do the xchg.
1416 BuildMI(BB, X86::MOVrr16, 1, X86::AX).addReg(SrcReg);
1417 BuildMI(BB, X86::XCHGrr8, 2).addReg(X86::AL, MOTy::UseAndDef)
1418 .addReg(X86::AH, MOTy::UseAndDef);
1419 BuildMI(BB, X86::MOVrr16, 1, DestReg).addReg(X86::AX);
1421 default: assert(0 && "Cannot byteswap this class!");
1426 /// visitLoadInst - Implement LLVM load instructions in terms of the x86 'mov'
1427 /// instruction. The load and store instructions are the only place where we
1428 /// need to worry about the memory layout of the target machine.
1430 void ISel::visitLoadInst(LoadInst &I) {
1431 bool isLittleEndian = TM.getTargetData().isLittleEndian();
1432 bool hasLongPointers = TM.getTargetData().getPointerSize() == 8;
1433 unsigned SrcAddrReg = getReg(I.getOperand(0));
1434 unsigned DestReg = getReg(I);
1436 unsigned Class = getClassB(I.getType());
1439 MachineBasicBlock::iterator MBBI = BB->end();
1440 addDirectMem(doFPLoad(BB, MBBI, I.getType(), DestReg), SrcAddrReg);
1443 case cLong: case cInt: case cShort: case cByte:
1444 break; // Integers of various sizes handled below
1445 default: assert(0 && "Unknown memory class!");
1448 // We need to adjust the input pointer if we are emulating a big-endian
1449 // long-pointer target. On these systems, the pointer that we are interested
1450 // in is in the upper part of the eight byte memory image of the pointer. It
1451 // also happens to be byte-swapped, but this will be handled later.
1453 if (!isLittleEndian && hasLongPointers && isa<PointerType>(I.getType())) {
1454 unsigned R = makeAnotherReg(Type::UIntTy);
1455 BuildMI(BB, X86::ADDri32, 2, R).addReg(SrcAddrReg).addZImm(4);
1459 unsigned IReg = DestReg;
1460 if (!isLittleEndian) // If big endian we need an intermediate stage
1461 DestReg = makeAnotherReg(Class != cLong ? I.getType() : Type::UIntTy);
1463 static const unsigned Opcode[] = {
1464 X86::MOVmr8, X86::MOVmr16, X86::MOVmr32, 0, X86::MOVmr32
1466 addDirectMem(BuildMI(BB, Opcode[Class], 4, DestReg), SrcAddrReg);
1468 // Handle long values now...
1469 if (Class == cLong) {
1470 if (isLittleEndian) {
1471 addRegOffset(BuildMI(BB, X86::MOVmr32, 4, DestReg+1), SrcAddrReg, 4);
1473 EmitByteSwap(IReg+1, DestReg, cInt);
1474 unsigned TempReg = makeAnotherReg(Type::IntTy);
1475 addRegOffset(BuildMI(BB, X86::MOVmr32, 4, TempReg), SrcAddrReg, 4);
1476 EmitByteSwap(IReg, TempReg, cInt);
1481 if (!isLittleEndian)
1482 EmitByteSwap(IReg, DestReg, Class);
1486 /// doFPStore - This method is used to store an FP value to memory using the
1487 /// current endianness.
1489 void ISel::doFPStore(const Type *Ty, unsigned DestAddrReg, unsigned SrcReg) {
1490 assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
1491 unsigned StoreOpcode = Ty == Type::FloatTy ? X86::FSTr32 : X86::FSTr64;
1493 if (TM.getTargetData().isLittleEndian()) { // fast path...
1494 addDirectMem(BuildMI(BB, StoreOpcode,5), DestAddrReg).addReg(SrcReg);
1498 // Allocate a temporary stack slot to transform the value into...
1499 int FrameIdx = F->getFrameInfo()->CreateStackObject(Ty, TM.getTargetData());
1500 unsigned SrcAddrReg = makeAnotherReg(Type::UIntTy);
1501 addFrameReference(BuildMI(BB, X86::LEAr32, 5, SrcAddrReg), FrameIdx);
1503 // Store the value into a temporary stack slot...
1504 addDirectMem(BuildMI(BB, StoreOpcode, 5), SrcAddrReg).addReg(SrcReg);
1506 // Perform the bswaps 32 bits at a time...
1507 unsigned TmpReg1 = makeAnotherReg(Type::UIntTy);
1508 unsigned TmpReg2 = makeAnotherReg(Type::UIntTy);
1509 addDirectMem(BuildMI(BB, X86::MOVmr32, 4, TmpReg1), SrcAddrReg);
1510 BuildMI(BB, X86::BSWAPr32, 1, TmpReg2).addReg(TmpReg1);
1511 unsigned Offset = (Ty == Type::DoubleTy) << 2;
1512 addRegOffset(BuildMI(BB, X86::MOVrm32, 5),
1513 DestAddrReg, Offset).addReg(TmpReg2);
1515 if (Ty == Type::DoubleTy) { // Swap the other 32 bits of a double value...
1516 TmpReg1 = makeAnotherReg(Type::UIntTy);
1517 TmpReg2 = makeAnotherReg(Type::UIntTy);
1519 addRegOffset(BuildMI(BB, X86::MOVmr32, 4, TmpReg1), SrcAddrReg, 4);
1520 BuildMI(BB, X86::BSWAPr32, 1, TmpReg2).addReg(TmpReg1);
1521 unsigned Offset = (Ty == Type::DoubleTy) << 2;
1522 addDirectMem(BuildMI(BB, X86::MOVrm32, 5), DestAddrReg).addReg(TmpReg2);
1527 /// visitStoreInst - Implement LLVM store instructions in terms of the x86 'mov'
1530 void ISel::visitStoreInst(StoreInst &I) {
1531 bool isLittleEndian = TM.getTargetData().isLittleEndian();
1532 bool hasLongPointers = TM.getTargetData().getPointerSize() == 8;
1533 unsigned ValReg = getReg(I.getOperand(0));
1534 unsigned AddressReg = getReg(I.getOperand(1));
1536 unsigned Class = getClassB(I.getOperand(0)->getType());
1539 if (isLittleEndian) {
1540 addDirectMem(BuildMI(BB, X86::MOVrm32, 1+4), AddressReg).addReg(ValReg);
1541 addRegOffset(BuildMI(BB, X86::MOVrm32, 1+4),
1542 AddressReg, 4).addReg(ValReg+1);
1544 unsigned T1 = makeAnotherReg(Type::IntTy);
1545 unsigned T2 = makeAnotherReg(Type::IntTy);
1546 EmitByteSwap(T1, ValReg , cInt);
1547 EmitByteSwap(T2, ValReg+1, cInt);
1548 addDirectMem(BuildMI(BB, X86::MOVrm32, 1+4), AddressReg).addReg(T2);
1549 addRegOffset(BuildMI(BB, X86::MOVrm32, 1+4), AddressReg, 4).addReg(T1);
1553 doFPStore(I.getOperand(0)->getType(), AddressReg, ValReg);
1555 case cInt: case cShort: case cByte:
1556 break; // Integers of various sizes handled below
1557 default: assert(0 && "Unknown memory class!");
1560 if (!isLittleEndian && hasLongPointers &&
1561 isa<PointerType>(I.getOperand(0)->getType())) {
1562 unsigned R = makeAnotherReg(Type::UIntTy);
1563 BuildMI(BB, X86::ADDri32, 2, R).addReg(AddressReg).addZImm(4);
1567 if (!isLittleEndian && Class != cByte) {
1568 unsigned R = makeAnotherReg(I.getOperand(0)->getType());
1569 EmitByteSwap(R, ValReg, Class);
1573 static const unsigned Opcode[] = { X86::MOVrm8, X86::MOVrm16, X86::MOVrm32 };
1574 addDirectMem(BuildMI(BB, Opcode[Class], 1+4), AddressReg).addReg(ValReg);
1578 /// visitCastInst - Here we have various kinds of copying with or without
1579 /// sign extension going on.
1580 void ISel::visitCastInst(CastInst &CI) {
1581 Value *Op = CI.getOperand(0);
1582 // If this is a cast from a 32-bit integer to a Long type, and the only uses
1583 // of the case are GEP instructions, then the cast does not need to be
1584 // generated explicitly, it will be folded into the GEP.
1585 if (CI.getType() == Type::LongTy &&
1586 (Op->getType() == Type::IntTy || Op->getType() == Type::UIntTy)) {
1587 bool AllUsesAreGEPs = true;
1588 for (Value::use_iterator I = CI.use_begin(), E = CI.use_end(); I != E; ++I)
1589 if (!isa<GetElementPtrInst>(*I)) {
1590 AllUsesAreGEPs = false;
1594 // No need to codegen this cast if all users are getelementptr instrs...
1595 if (AllUsesAreGEPs) return;
1598 unsigned DestReg = getReg(CI);
1599 MachineBasicBlock::iterator MI = BB->end();
1600 emitCastOperation(BB, MI, Op, CI.getType(), DestReg);
1603 /// emitCastOperation - Common code shared between visitCastInst and
1604 /// constant expression cast support.
1605 void ISel::emitCastOperation(MachineBasicBlock *BB,
1606 MachineBasicBlock::iterator &IP,
1607 Value *Src, const Type *DestTy,
1609 unsigned SrcReg = getReg(Src, BB, IP);
1610 const Type *SrcTy = Src->getType();
1611 unsigned SrcClass = getClassB(SrcTy);
1612 unsigned DestClass = getClassB(DestTy);
1614 // Implement casts to bool by using compare on the operand followed by set if
1615 // not zero on the result.
1616 if (DestTy == Type::BoolTy) {
1619 BMI(BB, IP, X86::TESTrr8, 2).addReg(SrcReg).addReg(SrcReg);
1622 BMI(BB, IP, X86::TESTrr16, 2).addReg(SrcReg).addReg(SrcReg);
1625 BMI(BB, IP, X86::TESTrr32, 2).addReg(SrcReg).addReg(SrcReg);
1628 unsigned TmpReg = makeAnotherReg(Type::IntTy);
1629 BMI(BB, IP, X86::ORrr32, 2, TmpReg).addReg(SrcReg).addReg(SrcReg+1);
1633 assert(0 && "FIXME: implement cast FP to bool");
1637 // If the zero flag is not set, then the value is true, set the byte to
1639 BMI(BB, IP, X86::SETNEr, 1, DestReg);
1643 static const unsigned RegRegMove[] = {
1644 X86::MOVrr8, X86::MOVrr16, X86::MOVrr32, X86::FpMOV, X86::MOVrr32
1647 // Implement casts between values of the same type class (as determined by
1648 // getClass) by using a register-to-register move.
1649 if (SrcClass == DestClass) {
1650 if (SrcClass <= cInt || (SrcClass == cFP && SrcTy == DestTy)) {
1651 BMI(BB, IP, RegRegMove[SrcClass], 1, DestReg).addReg(SrcReg);
1652 } else if (SrcClass == cFP) {
1653 if (SrcTy == Type::FloatTy) { // double -> float
1654 assert(DestTy == Type::DoubleTy && "Unknown cFP member!");
1655 BMI(BB, IP, X86::FpMOV, 1, DestReg).addReg(SrcReg);
1656 } else { // float -> double
1657 assert(SrcTy == Type::DoubleTy && DestTy == Type::FloatTy &&
1658 "Unknown cFP member!");
1659 // Truncate from double to float by storing to memory as short, then
1661 unsigned FltAlign = TM.getTargetData().getFloatAlignment();
1662 int FrameIdx = F->getFrameInfo()->CreateStackObject(4, FltAlign);
1663 addFrameReference(BMI(BB, IP, X86::FSTr32, 5), FrameIdx).addReg(SrcReg);
1664 addFrameReference(BMI(BB, IP, X86::FLDr32, 5, DestReg), FrameIdx);
1666 } else if (SrcClass == cLong) {
1667 BMI(BB, IP, X86::MOVrr32, 1, DestReg).addReg(SrcReg);
1668 BMI(BB, IP, X86::MOVrr32, 1, DestReg+1).addReg(SrcReg+1);
1670 assert(0 && "Cannot handle this type of cast instruction!");
1676 // Handle cast of SMALLER int to LARGER int using a move with sign extension
1677 // or zero extension, depending on whether the source type was signed.
1678 if (SrcClass <= cInt && (DestClass <= cInt || DestClass == cLong) &&
1679 SrcClass < DestClass) {
1680 bool isLong = DestClass == cLong;
1681 if (isLong) DestClass = cInt;
1683 static const unsigned Opc[][4] = {
1684 { X86::MOVSXr16r8, X86::MOVSXr32r8, X86::MOVSXr32r16, X86::MOVrr32 }, // s
1685 { X86::MOVZXr16r8, X86::MOVZXr32r8, X86::MOVZXr32r16, X86::MOVrr32 } // u
1688 bool isUnsigned = SrcTy->isUnsigned();
1689 BMI(BB, IP, Opc[isUnsigned][SrcClass + DestClass - 1], 1,
1690 DestReg).addReg(SrcReg);
1692 if (isLong) { // Handle upper 32 bits as appropriate...
1693 if (isUnsigned) // Zero out top bits...
1694 BMI(BB, IP, X86::MOVir32, 1, DestReg+1).addZImm(0);
1695 else // Sign extend bottom half...
1696 BMI(BB, IP, X86::SARir32, 2, DestReg+1).addReg(DestReg).addZImm(31);
1701 // Special case long -> int ...
1702 if (SrcClass == cLong && DestClass == cInt) {
1703 BMI(BB, IP, X86::MOVrr32, 1, DestReg).addReg(SrcReg);
1707 // Handle cast of LARGER int to SMALLER int using a move to EAX followed by a
1708 // move out of AX or AL.
1709 if ((SrcClass <= cInt || SrcClass == cLong) && DestClass <= cInt
1710 && SrcClass > DestClass) {
1711 static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX, 0, X86::EAX };
1712 BMI(BB, IP, RegRegMove[SrcClass], 1, AReg[SrcClass]).addReg(SrcReg);
1713 BMI(BB, IP, RegRegMove[DestClass], 1, DestReg).addReg(AReg[DestClass]);
1717 // Handle casts from integer to floating point now...
1718 if (DestClass == cFP) {
1719 // Promote the integer to a type supported by FLD. We do this because there
1720 // are no unsigned FLD instructions, so we must promote an unsigned value to
1721 // a larger signed value, then use FLD on the larger value.
1723 const Type *PromoteType = 0;
1724 unsigned PromoteOpcode;
1725 switch (SrcTy->getPrimitiveID()) {
1726 case Type::BoolTyID:
1727 case Type::SByteTyID:
1728 // We don't have the facilities for directly loading byte sized data from
1729 // memory (even signed). Promote it to 16 bits.
1730 PromoteType = Type::ShortTy;
1731 PromoteOpcode = X86::MOVSXr16r8;
1733 case Type::UByteTyID:
1734 PromoteType = Type::ShortTy;
1735 PromoteOpcode = X86::MOVZXr16r8;
1737 case Type::UShortTyID:
1738 PromoteType = Type::IntTy;
1739 PromoteOpcode = X86::MOVZXr32r16;
1741 case Type::UIntTyID: {
1742 // Make a 64 bit temporary... and zero out the top of it...
1743 unsigned TmpReg = makeAnotherReg(Type::LongTy);
1744 BMI(BB, IP, X86::MOVrr32, 1, TmpReg).addReg(SrcReg);
1745 BMI(BB, IP, X86::MOVir32, 1, TmpReg+1).addZImm(0);
1746 SrcTy = Type::LongTy;
1751 case Type::ULongTyID:
1752 assert("FIXME: not implemented: cast ulong X to fp type!");
1753 default: // No promotion needed...
1758 unsigned TmpReg = makeAnotherReg(PromoteType);
1759 BMI(BB, IP, SrcTy->isSigned() ? X86::MOVSXr16r8 : X86::MOVZXr16r8,
1760 1, TmpReg).addReg(SrcReg);
1761 SrcTy = PromoteType;
1762 SrcClass = getClass(PromoteType);
1766 // Spill the integer to memory and reload it from there...
1768 F->getFrameInfo()->CreateStackObject(SrcTy, TM.getTargetData());
1770 if (SrcClass == cLong) {
1771 addFrameReference(BMI(BB, IP, X86::MOVrm32, 5), FrameIdx).addReg(SrcReg);
1772 addFrameReference(BMI(BB, IP, X86::MOVrm32, 5),
1773 FrameIdx, 4).addReg(SrcReg+1);
1775 static const unsigned Op1[] = { X86::MOVrm8, X86::MOVrm16, X86::MOVrm32 };
1776 addFrameReference(BMI(BB, IP, Op1[SrcClass], 5), FrameIdx).addReg(SrcReg);
1779 static const unsigned Op2[] =
1780 { 0/*byte*/, X86::FILDr16, X86::FILDr32, 0/*FP*/, X86::FILDr64 };
1781 addFrameReference(BMI(BB, IP, Op2[SrcClass], 5, DestReg), FrameIdx);
1785 // Handle casts from floating point to integer now...
1786 if (SrcClass == cFP) {
1787 // Change the floating point control register to use "round towards zero"
1788 // mode when truncating to an integer value.
1790 int CWFrameIdx = F->getFrameInfo()->CreateStackObject(2, 2);
1791 addFrameReference(BMI(BB, IP, X86::FNSTCWm16, 4), CWFrameIdx);
1793 // Load the old value of the high byte of the control word...
1794 unsigned HighPartOfCW = makeAnotherReg(Type::UByteTy);
1795 addFrameReference(BMI(BB, IP, X86::MOVmr8, 4, HighPartOfCW), CWFrameIdx, 1);
1797 // Set the high part to be round to zero...
1798 addFrameReference(BMI(BB, IP, X86::MOVim8, 5), CWFrameIdx, 1).addZImm(12);
1800 // Reload the modified control word now...
1801 addFrameReference(BMI(BB, IP, X86::FLDCWm16, 4), CWFrameIdx);
1803 // Restore the memory image of control word to original value
1804 addFrameReference(BMI(BB, IP, X86::MOVrm8, 5),
1805 CWFrameIdx, 1).addReg(HighPartOfCW);
1807 // We don't have the facilities for directly storing byte sized data to
1808 // memory. Promote it to 16 bits. We also must promote unsigned values to
1809 // larger classes because we only have signed FP stores.
1810 unsigned StoreClass = DestClass;
1811 const Type *StoreTy = DestTy;
1812 if (StoreClass == cByte || DestTy->isUnsigned())
1813 switch (StoreClass) {
1814 case cByte: StoreTy = Type::ShortTy; StoreClass = cShort; break;
1815 case cShort: StoreTy = Type::IntTy; StoreClass = cInt; break;
1816 case cInt: StoreTy = Type::LongTy; StoreClass = cLong; break;
1818 assert(0 &&"FIXME not implemented: cast FP to unsigned long long");
1820 default: assert(0 && "Unknown store class!");
1823 // Spill the integer to memory and reload it from there...
1825 F->getFrameInfo()->CreateStackObject(StoreTy, TM.getTargetData());
1827 static const unsigned Op1[] =
1828 { 0, X86::FISTr16, X86::FISTr32, 0, X86::FISTPr64 };
1829 addFrameReference(BMI(BB, IP, Op1[StoreClass], 5), FrameIdx).addReg(SrcReg);
1831 if (DestClass == cLong) {
1832 addFrameReference(BMI(BB, IP, X86::MOVmr32, 4, DestReg), FrameIdx);
1833 addFrameReference(BMI(BB, IP, X86::MOVmr32, 4, DestReg+1), FrameIdx, 4);
1835 static const unsigned Op2[] = { X86::MOVmr8, X86::MOVmr16, X86::MOVmr32 };
1836 addFrameReference(BMI(BB, IP, Op2[DestClass], 4, DestReg), FrameIdx);
1839 // Reload the original control word now...
1840 addFrameReference(BMI(BB, IP, X86::FLDCWm16, 4), CWFrameIdx);
1844 // Anything we haven't handled already, we can't (yet) handle at all.
1845 assert(0 && "Unhandled cast instruction!");
1849 /// visitVarArgInst - Implement the va_arg instruction...
1851 void ISel::visitVarArgInst(VarArgInst &I) {
1852 unsigned SrcReg = getReg(I.getOperand(0));
1853 unsigned DestReg = getReg(I);
1855 // Load the va_list into a register...
1856 unsigned VAList = makeAnotherReg(Type::UIntTy);
1857 addDirectMem(BuildMI(BB, X86::MOVmr32, 4, VAList), SrcReg);
1860 switch (I.getType()->getPrimitiveID()) {
1863 assert(0 && "Error: bad type for va_arg instruction!");
1865 case Type::PointerTyID:
1866 case Type::UIntTyID:
1869 addDirectMem(BuildMI(BB, X86::MOVmr32, 4, DestReg), VAList);
1871 case Type::ULongTyID:
1872 case Type::LongTyID:
1874 addDirectMem(BuildMI(BB, X86::MOVmr32, 4, DestReg), VAList);
1875 addRegOffset(BuildMI(BB, X86::MOVmr32, 4, DestReg+1), VAList, 4);
1877 case Type::DoubleTyID:
1879 addDirectMem(BuildMI(BB, X86::FLDr64, 4, DestReg), VAList);
1883 // Increment the VAList pointer...
1884 unsigned NextVAList = makeAnotherReg(Type::UIntTy);
1885 BuildMI(BB, X86::ADDri32, 2, NextVAList).addReg(VAList).addZImm(Size);
1887 // Update the VAList in memory...
1888 addDirectMem(BuildMI(BB, X86::MOVrm32, 5), SrcReg).addReg(NextVAList);
1892 // ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
1893 // returns zero when the input is not exactly a power of two.
1894 static unsigned ExactLog2(unsigned Val) {
1895 if (Val == 0) return 0;
1898 if (Val & 1) return 0;
1905 void ISel::visitGetElementPtrInst(GetElementPtrInst &I) {
1906 unsigned outputReg = getReg(I);
1907 MachineBasicBlock::iterator MI = BB->end();
1908 emitGEPOperation(BB, MI, I.getOperand(0),
1909 I.op_begin()+1, I.op_end(), outputReg);
1912 void ISel::emitGEPOperation(MachineBasicBlock *MBB,
1913 MachineBasicBlock::iterator &IP,
1914 Value *Src, User::op_iterator IdxBegin,
1915 User::op_iterator IdxEnd, unsigned TargetReg) {
1916 const TargetData &TD = TM.getTargetData();
1917 const Type *Ty = Src->getType();
1918 unsigned BaseReg = getReg(Src, MBB, IP);
1920 // GEPs have zero or more indices; we must perform a struct access
1921 // or array access for each one.
1922 for (GetElementPtrInst::op_iterator oi = IdxBegin,
1923 oe = IdxEnd; oi != oe; ++oi) {
1925 unsigned NextReg = BaseReg;
1926 if (const StructType *StTy = dyn_cast<StructType>(Ty)) {
1927 // It's a struct access. idx is the index into the structure,
1928 // which names the field. This index must have ubyte type.
1929 const ConstantUInt *CUI = cast<ConstantUInt>(idx);
1930 assert(CUI->getType() == Type::UByteTy
1931 && "Funny-looking structure index in GEP");
1932 // Use the TargetData structure to pick out what the layout of
1933 // the structure is in memory. Since the structure index must
1934 // be constant, we can get its value and use it to find the
1935 // right byte offset from the StructLayout class's list of
1936 // structure member offsets.
1937 unsigned idxValue = CUI->getValue();
1938 unsigned FieldOff = TD.getStructLayout(StTy)->MemberOffsets[idxValue];
1940 NextReg = makeAnotherReg(Type::UIntTy);
1941 // Emit an ADD to add FieldOff to the basePtr.
1942 BMI(MBB, IP, X86::ADDri32, 2,NextReg).addReg(BaseReg).addZImm(FieldOff);
1944 // The next type is the member of the structure selected by the
1946 Ty = StTy->getElementTypes()[idxValue];
1947 } else if (const SequentialType *SqTy = cast<SequentialType>(Ty)) {
1948 // It's an array or pointer access: [ArraySize x ElementType].
1950 // idx is the index into the array. Unlike with structure
1951 // indices, we may not know its actual value at code-generation
1953 assert(idx->getType() == Type::LongTy && "Bad GEP array index!");
1955 // Most GEP instructions use a [cast (int/uint) to LongTy] as their
1956 // operand on X86. Handle this case directly now...
1957 if (CastInst *CI = dyn_cast<CastInst>(idx))
1958 if (CI->getOperand(0)->getType() == Type::IntTy ||
1959 CI->getOperand(0)->getType() == Type::UIntTy)
1960 idx = CI->getOperand(0);
1962 // We want to add BaseReg to(idxReg * sizeof ElementType). First, we
1963 // must find the size of the pointed-to type (Not coincidentally, the next
1964 // type is the type of the elements in the array).
1965 Ty = SqTy->getElementType();
1966 unsigned elementSize = TD.getTypeSize(Ty);
1968 // If idxReg is a constant, we don't need to perform the multiply!
1969 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(idx)) {
1970 if (!CSI->isNullValue()) {
1971 unsigned Offset = elementSize*CSI->getValue();
1972 NextReg = makeAnotherReg(Type::UIntTy);
1973 BMI(MBB, IP, X86::ADDri32, 2,NextReg).addReg(BaseReg).addZImm(Offset);
1975 } else if (elementSize == 1) {
1976 // If the element size is 1, we don't have to multiply, just add
1977 unsigned idxReg = getReg(idx, MBB, IP);
1978 NextReg = makeAnotherReg(Type::UIntTy);
1979 BMI(MBB, IP, X86::ADDrr32, 2, NextReg).addReg(BaseReg).addReg(idxReg);
1981 unsigned idxReg = getReg(idx, MBB, IP);
1982 unsigned OffsetReg = makeAnotherReg(Type::UIntTy);
1983 if (unsigned Shift = ExactLog2(elementSize)) {
1984 // If the element size is exactly a power of 2, use a shift to get it.
1985 BMI(MBB, IP, X86::SHLir32, 2,
1986 OffsetReg).addReg(idxReg).addZImm(Shift-1);
1988 // Most general case, emit a multiply...
1989 unsigned elementSizeReg = makeAnotherReg(Type::LongTy);
1990 BMI(MBB, IP, X86::MOVir32, 1, elementSizeReg).addZImm(elementSize);
1992 // Emit a MUL to multiply the register holding the index by
1993 // elementSize, putting the result in OffsetReg.
1994 doMultiply(MBB, IP, OffsetReg, Type::IntTy, idxReg, elementSizeReg);
1996 // Emit an ADD to add OffsetReg to the basePtr.
1997 NextReg = makeAnotherReg(Type::UIntTy);
1998 BMI(MBB, IP, X86::ADDrr32, 2,NextReg).addReg(BaseReg).addReg(OffsetReg);
2001 // Now that we are here, further indices refer to subtypes of this
2002 // one, so we don't need to worry about BaseReg itself, anymore.
2005 // After we have processed all the indices, the result is left in
2006 // BaseReg. Move it to the register where we were expected to
2007 // put the answer. A 32-bit move should do it, because we are in
2009 BMI(MBB, IP, X86::MOVrr32, 1, TargetReg).addReg(BaseReg);
2013 /// visitAllocaInst - If this is a fixed size alloca, allocate space from the
2014 /// frame manager, otherwise do it the hard way.
2016 void ISel::visitAllocaInst(AllocaInst &I) {
2017 // Find the data size of the alloca inst's getAllocatedType.
2018 const Type *Ty = I.getAllocatedType();
2019 unsigned TySize = TM.getTargetData().getTypeSize(Ty);
2021 // If this is a fixed size alloca in the entry block for the function,
2022 // statically stack allocate the space.
2024 if (ConstantUInt *CUI = dyn_cast<ConstantUInt>(I.getArraySize())) {
2025 if (I.getParent() == I.getParent()->getParent()->begin()) {
2026 TySize *= CUI->getValue(); // Get total allocated size...
2027 unsigned Alignment = TM.getTargetData().getTypeAlignment(Ty);
2029 // Create a new stack object using the frame manager...
2030 int FrameIdx = F->getFrameInfo()->CreateStackObject(TySize, Alignment);
2031 addFrameReference(BuildMI(BB, X86::LEAr32, 5, getReg(I)), FrameIdx);
2036 // Create a register to hold the temporary result of multiplying the type size
2037 // constant by the variable amount.
2038 unsigned TotalSizeReg = makeAnotherReg(Type::UIntTy);
2039 unsigned SrcReg1 = getReg(I.getArraySize());
2040 unsigned SizeReg = makeAnotherReg(Type::UIntTy);
2041 BuildMI(BB, X86::MOVir32, 1, SizeReg).addZImm(TySize);
2043 // TotalSizeReg = mul <numelements>, <TypeSize>
2044 MachineBasicBlock::iterator MBBI = BB->end();
2045 doMultiply(BB, MBBI, TotalSizeReg, Type::UIntTy, SrcReg1, SizeReg);
2047 // AddedSize = add <TotalSizeReg>, 15
2048 unsigned AddedSizeReg = makeAnotherReg(Type::UIntTy);
2049 BuildMI(BB, X86::ADDri32, 2, AddedSizeReg).addReg(TotalSizeReg).addZImm(15);
2051 // AlignedSize = and <AddedSize>, ~15
2052 unsigned AlignedSize = makeAnotherReg(Type::UIntTy);
2053 BuildMI(BB, X86::ANDri32, 2, AlignedSize).addReg(AddedSizeReg).addZImm(~15);
2055 // Subtract size from stack pointer, thereby allocating some space.
2056 BuildMI(BB, X86::SUBrr32, 2, X86::ESP).addReg(X86::ESP).addReg(AlignedSize);
2058 // Put a pointer to the space into the result register, by copying
2059 // the stack pointer.
2060 BuildMI(BB, X86::MOVrr32, 1, getReg(I)).addReg(X86::ESP);
2062 // Inform the Frame Information that we have just allocated a variable-sized
2064 F->getFrameInfo()->CreateVariableSizedObject();
2067 /// visitMallocInst - Malloc instructions are code generated into direct calls
2068 /// to the library malloc.
2070 void ISel::visitMallocInst(MallocInst &I) {
2071 unsigned AllocSize = TM.getTargetData().getTypeSize(I.getAllocatedType());
2074 if (ConstantUInt *C = dyn_cast<ConstantUInt>(I.getOperand(0))) {
2075 Arg = getReg(ConstantUInt::get(Type::UIntTy, C->getValue() * AllocSize));
2077 Arg = makeAnotherReg(Type::UIntTy);
2078 unsigned Op0Reg = getReg(ConstantUInt::get(Type::UIntTy, AllocSize));
2079 unsigned Op1Reg = getReg(I.getOperand(0));
2080 MachineBasicBlock::iterator MBBI = BB->end();
2081 doMultiply(BB, MBBI, Arg, Type::UIntTy, Op0Reg, Op1Reg);
2086 std::vector<ValueRecord> Args;
2087 Args.push_back(ValueRecord(Arg, Type::UIntTy));
2088 MachineInstr *TheCall = BuildMI(X86::CALLpcrel32,
2089 1).addExternalSymbol("malloc", true);
2090 doCall(ValueRecord(getReg(I), I.getType()), TheCall, Args);
2094 /// visitFreeInst - Free instructions are code gen'd to call the free libc
2097 void ISel::visitFreeInst(FreeInst &I) {
2098 std::vector<ValueRecord> Args;
2099 Args.push_back(ValueRecord(getReg(I.getOperand(0)),
2100 I.getOperand(0)->getType()));
2101 MachineInstr *TheCall = BuildMI(X86::CALLpcrel32,
2102 1).addExternalSymbol("free", true);
2103 doCall(ValueRecord(0, Type::VoidTy), TheCall, Args);
2107 /// createSimpleX86InstructionSelector - This pass converts an LLVM function
2108 /// into a machine code representation is a very simple peep-hole fashion. The
2109 /// generated code sucks but the implementation is nice and simple.
2111 Pass *createSimpleX86InstructionSelector(TargetMachine &TM) {
2112 return new ISel(TM);