1 //===-- InstSelectSimple.cpp - A simple instruction selector for x86 ------===//
3 // This file defines a simple peephole instruction selector for the x86 platform
5 //===----------------------------------------------------------------------===//
8 #include "X86InstrInfo.h"
9 #include "X86InstrBuilder.h"
10 #include "llvm/Function.h"
11 #include "llvm/iTerminators.h"
12 #include "llvm/iOperators.h"
13 #include "llvm/iOther.h"
14 #include "llvm/iPHINode.h"
15 #include "llvm/iMemory.h"
16 #include "llvm/Type.h"
17 #include "llvm/DerivedTypes.h"
18 #include "llvm/Constants.h"
19 #include "llvm/Pass.h"
20 #include "llvm/CodeGen/MachineFunction.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/Target/TargetMachine.h"
23 #include "llvm/Support/InstVisitor.h"
24 #include "llvm/Target/MRegisterInfo.h"
27 using namespace MOTy; // Get Use, Def, UseAndDef
30 /// BMI - A special BuildMI variant that takes an iterator to insert the
31 /// instruction at as well as a basic block.
32 /// this is the version for when you have a destination register in mind.
33 inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB,
34 MachineBasicBlock::iterator &I,
38 assert(I >= MBB->begin() && I <= MBB->end() && "Bad iterator!");
39 MachineInstr *MI = new MachineInstr(Opcode, NumOperands+1, true, true);
40 I = ++MBB->insert(I, MI);
41 return MachineInstrBuilder(MI).addReg(DestReg, MOTy::Def);
44 /// BMI - A special BuildMI variant that takes an iterator to insert the
45 /// instruction at as well as a basic block.
46 inline static MachineInstrBuilder BMI(MachineBasicBlock *MBB,
47 MachineBasicBlock::iterator &I,
49 unsigned NumOperands) {
50 assert(I > MBB->begin() && I <= MBB->end() && "Bad iterator!");
51 MachineInstr *MI = new MachineInstr(Opcode, NumOperands, true, true);
52 I = ++MBB->insert(I, MI);
53 return MachineInstrBuilder(MI);
58 struct ISel : public FunctionPass, InstVisitor<ISel> {
60 MachineFunction *F; // The function we are compiling into
61 MachineBasicBlock *BB; // The current MBB we are compiling
64 std::map<Value*, unsigned> RegMap; // Mapping between Val's and SSA Regs
66 // MBBMap - Mapping between LLVM BB -> Machine BB
67 std::map<const BasicBlock*, MachineBasicBlock*> MBBMap;
69 ISel(TargetMachine &tm)
70 : TM(tm), F(0), BB(0), CurReg(MRegisterInfo::FirstVirtualRegister) {}
72 /// runOnFunction - Top level implementation of instruction selection for
73 /// the entire function.
75 bool runOnFunction(Function &Fn) {
76 F = &MachineFunction::construct(&Fn, TM);
78 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
79 F->getBasicBlockList().push_back(MBBMap[I] = new MachineBasicBlock(I));
81 // Emit instructions to load the arguments... The function's arguments
84 // [EBP] -- copy of old EBP
85 // [EBP + 4] -- return address
86 // [EBP + 8] -- first argument (leftmost lexically)
88 // So we want to start with counter = 2.
91 unsigned ArgOffset = 8;
92 for (Function::aiterator I = Fn.abegin(), E = Fn.aend(); I != E;
93 ++I, ArgOffset += 4) {
94 unsigned Reg = getReg(*I);
96 // Load it out of the stack frame at EBP + 4*argPos.
98 // FIXME: This should load the argument of the appropriate size!!
99 addRegOffset(BuildMI(BB, X86::MOVmr32, 4, Reg), X86::EBP, ArgOffset);
102 // Instruction select everything except PHI nodes
105 // Select the PHI nodes
110 CurReg = MRegisterInfo::FirstVirtualRegister;
112 return false; // We never modify the LLVM itself.
115 virtual const char *getPassName() const {
116 return "X86 Simple Instruction Selection";
119 /// visitBasicBlock - This method is called when we are visiting a new basic
120 /// block. This simply creates a new MachineBasicBlock to emit code into
121 /// and adds it to the current MachineFunction. Subsequent visit* for
122 /// instructions will be invoked for all instructions in the basic block.
124 void visitBasicBlock(BasicBlock &LLVM_BB) {
125 BB = MBBMap[&LLVM_BB];
129 /// SelectPHINodes - Insert machine code to generate phis. This is tricky
130 /// because we have to generate our sources into the source basic blocks,
131 /// not the current one.
133 void SelectPHINodes();
135 // Visitation methods for various instructions. These methods simply emit
136 // fixed X86 code for each instruction.
139 // Control flow operators
140 void visitReturnInst(ReturnInst &RI);
141 void visitBranchInst(BranchInst &BI);
142 void visitCallInst(CallInst &I);
144 // Arithmetic operators
145 void visitSimpleBinary(BinaryOperator &B, unsigned OpcodeClass);
146 void visitAdd(BinaryOperator &B) { visitSimpleBinary(B, 0); }
147 void visitSub(BinaryOperator &B) { visitSimpleBinary(B, 1); }
148 void doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI,
149 unsigned destReg, const Type *resultType,
150 unsigned op0Reg, unsigned op1Reg);
151 void visitMul(BinaryOperator &B);
153 void visitDiv(BinaryOperator &B) { visitDivRem(B); }
154 void visitRem(BinaryOperator &B) { visitDivRem(B); }
155 void visitDivRem(BinaryOperator &B);
158 void visitAnd(BinaryOperator &B) { visitSimpleBinary(B, 2); }
159 void visitOr (BinaryOperator &B) { visitSimpleBinary(B, 3); }
160 void visitXor(BinaryOperator &B) { visitSimpleBinary(B, 4); }
162 // Binary comparison operators
163 void visitSetCCInst(SetCondInst &I, unsigned OpNum);
164 void visitSetEQ(SetCondInst &I) { visitSetCCInst(I, 0); }
165 void visitSetNE(SetCondInst &I) { visitSetCCInst(I, 1); }
166 void visitSetLT(SetCondInst &I) { visitSetCCInst(I, 2); }
167 void visitSetGT(SetCondInst &I) { visitSetCCInst(I, 3); }
168 void visitSetLE(SetCondInst &I) { visitSetCCInst(I, 4); }
169 void visitSetGE(SetCondInst &I) { visitSetCCInst(I, 5); }
171 // Memory Instructions
172 void visitLoadInst(LoadInst &I);
173 void visitStoreInst(StoreInst &I);
174 void visitGetElementPtrInst(GetElementPtrInst &I);
175 void visitMallocInst(MallocInst &I);
176 void visitFreeInst(FreeInst &I);
177 void visitAllocaInst(AllocaInst &I);
180 void visitShiftInst(ShiftInst &I);
181 void visitPHINode(PHINode &I) {} // PHI nodes handled by second pass
182 void visitCastInst(CastInst &I);
184 void visitInstruction(Instruction &I) {
185 std::cerr << "Cannot instruction select: " << I;
189 /// promote32 - Make a value 32-bits wide, and put it somewhere.
190 void promote32 (const unsigned targetReg, Value *v);
192 // emitGEPOperation - Common code shared between visitGetElementPtrInst and
193 // constant expression GEP support.
195 void emitGEPOperation(MachineBasicBlock *BB, MachineBasicBlock::iterator&IP,
196 Value *Src, User::op_iterator IdxBegin,
197 User::op_iterator IdxEnd, unsigned TargetReg);
199 /// copyConstantToRegister - Output the instructions required to put the
200 /// specified constant into the specified register.
202 void copyConstantToRegister(MachineBasicBlock *MBB,
203 MachineBasicBlock::iterator &MBBI,
204 Constant *C, unsigned Reg);
206 /// makeAnotherReg - This method returns the next register number
207 /// we haven't yet used.
208 unsigned makeAnotherReg(const Type *Ty) {
209 // Add the mapping of regnumber => reg class to MachineFunction
210 F->addRegMap(CurReg, TM.getRegisterInfo()->getRegClassForType(Ty));
214 /// getReg - This method turns an LLVM value into a register number. This
215 /// is guaranteed to produce the same register number for a particular value
216 /// every time it is queried.
218 unsigned getReg(Value &V) { return getReg(&V); } // Allow references
219 unsigned getReg(Value *V) {
220 // Just append to the end of the current bb.
221 MachineBasicBlock::iterator It = BB->end();
222 return getReg(V, BB, It);
224 unsigned getReg(Value *V, MachineBasicBlock *MBB,
225 MachineBasicBlock::iterator &IPt) {
226 unsigned &Reg = RegMap[V];
228 Reg = makeAnotherReg(V->getType());
232 // If this operand is a constant, emit the code to copy the constant into
233 // the register here...
235 if (Constant *C = dyn_cast<Constant>(V)) {
236 copyConstantToRegister(MBB, IPt, C, Reg);
237 RegMap.erase(V); // Assign a new name to this constant if ref'd again
238 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
239 // Move the address of the global into the register
240 BMI(MBB, IPt, X86::MOVir32, 1, Reg).addReg(GV);
241 RegMap.erase(V); // Assign a new name to this address if ref'd again
249 /// TypeClass - Used by the X86 backend to group LLVM types by their basic X86
253 cByte, cShort, cInt, cLong, cFloat, cDouble
256 /// getClass - Turn a primitive type into a "class" number which is based on the
257 /// size of the type, and whether or not it is floating point.
259 static inline TypeClass getClass(const Type *Ty) {
260 switch (Ty->getPrimitiveID()) {
261 case Type::SByteTyID:
262 case Type::UByteTyID: return cByte; // Byte operands are class #0
263 case Type::ShortTyID:
264 case Type::UShortTyID: return cShort; // Short operands are class #1
267 case Type::PointerTyID: return cInt; // Int's and pointers are class #2
270 case Type::ULongTyID: //return cLong; // Longs are class #3
271 return cInt; // FIXME: LONGS ARE TREATED AS INTS!
273 case Type::FloatTyID: return cFloat; // Float is class #4
274 case Type::DoubleTyID: return cDouble; // Doubles are class #5
276 assert(0 && "Invalid type to getClass!");
277 return cByte; // not reached
281 // getClassB - Just like getClass, but treat boolean values as bytes.
282 static inline TypeClass getClassB(const Type *Ty) {
283 if (Ty == Type::BoolTy) return cByte;
288 /// copyConstantToRegister - Output the instructions required to put the
289 /// specified constant into the specified register.
291 void ISel::copyConstantToRegister(MachineBasicBlock *MBB,
292 MachineBasicBlock::iterator &IP,
293 Constant *C, unsigned R) {
294 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
295 if (CE->getOpcode() == Instruction::GetElementPtr) {
296 emitGEPOperation(MBB, IP, CE->getOperand(0),
297 CE->op_begin()+1, CE->op_end(), R);
301 std::cerr << "Offending expr: " << C << "\n";
302 assert (0 && "Constant expressions not yet handled!\n");
305 if (C->getType()->isIntegral()) {
306 unsigned Class = getClassB(C->getType());
307 assert(Class != 3 && "Type not handled yet!");
309 static const unsigned IntegralOpcodeTab[] = {
310 X86::MOVir8, X86::MOVir16, X86::MOVir32
313 if (C->getType() == Type::BoolTy) {
314 BMI(MBB, IP, X86::MOVir8, 1, R).addZImm(C == ConstantBool::True);
315 } else if (C->getType()->isSigned()) {
316 ConstantSInt *CSI = cast<ConstantSInt>(C);
317 BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addSImm(CSI->getValue());
319 ConstantUInt *CUI = cast<ConstantUInt>(C);
320 BMI(MBB, IP, IntegralOpcodeTab[Class], 1, R).addZImm(CUI->getValue());
322 } else if (isa<ConstantPointerNull>(C)) {
323 // Copy zero (null pointer) to the register.
324 BMI(MBB, IP, X86::MOVir32, 1, R).addZImm(0);
325 } else if (ConstantPointerRef *CPR = dyn_cast<ConstantPointerRef>(C)) {
326 unsigned SrcReg = getReg(CPR->getValue(), MBB, IP);
327 BMI(MBB, IP, X86::MOVrr32, 1, R).addReg(SrcReg);
329 std::cerr << "Offending constant: " << C << "\n";
330 assert(0 && "Type not handled yet!");
334 /// SelectPHINodes - Insert machine code to generate phis. This is tricky
335 /// because we have to generate our sources into the source basic blocks, not
338 void ISel::SelectPHINodes() {
339 const Function &LF = *F->getFunction(); // The LLVM function...
340 for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) {
341 const BasicBlock *BB = I;
342 MachineBasicBlock *MBB = MBBMap[I];
344 // Loop over all of the PHI nodes in the LLVM basic block...
345 unsigned NumPHIs = 0;
346 for (BasicBlock::const_iterator I = BB->begin();
347 PHINode *PN = (PHINode*)dyn_cast<PHINode>(&*I); ++I) {
348 // Create a new machine instr PHI node, and insert it.
349 MachineInstr *MI = BuildMI(X86::PHI, PN->getNumOperands(), getReg(*PN));
350 MBB->insert(MBB->begin()+NumPHIs++, MI); // Insert it at the top of the BB
352 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
353 MachineBasicBlock *PredMBB = MBBMap[PN->getIncomingBlock(i)];
355 // Get the incoming value into a virtual register. If it is not already
356 // available in a virtual register, insert the computation code into
360 MachineBasicBlock::iterator PI = PredMBB->begin();
361 while ((*PI)->getOpcode() == X86::PHI) ++PI;
363 MI->addRegOperand(getReg(PN->getIncomingValue(i), PredMBB, PI));
364 MI->addMachineBasicBlockOperand(PredMBB);
372 /// SetCC instructions - Here we just emit boilerplate code to set a byte-sized
373 /// register, then move it to wherever the result should be.
374 /// We handle FP setcc instructions by pushing them, doing a
375 /// compare-and-pop-twice, and then copying the concodes to the main
376 /// processor's concodes (I didn't make this up, it's in the Intel manual)
378 void ISel::visitSetCCInst(SetCondInst &I, unsigned OpNum) {
379 // The arguments are already supposed to be of the same type.
380 const Type *CompTy = I.getOperand(0)->getType();
381 unsigned reg1 = getReg(I.getOperand(0));
382 unsigned reg2 = getReg(I.getOperand(1));
384 unsigned Class = getClass(CompTy);
386 // Emit: cmp <var1>, <var2> (do the comparison). We can
387 // compare 8-bit with 8-bit, 16-bit with 16-bit, 32-bit with
390 BuildMI (BB, X86::CMPrr8, 2).addReg (reg1).addReg (reg2);
393 BuildMI (BB, X86::CMPrr16, 2).addReg (reg1).addReg (reg2);
396 BuildMI (BB, X86::CMPrr32, 2).addReg (reg1).addReg (reg2);
399 // Push the variables on the stack with fldl opcodes.
400 // FIXME: assuming var1, var2 are in memory, if not, spill to
402 case cFloat: // Floats
403 BuildMI (BB, X86::FLDr32, 1).addReg (reg1);
404 BuildMI (BB, X86::FLDr32, 1).addReg (reg2);
406 case cDouble: // Doubles
407 BuildMI (BB, X86::FLDr64, 1).addReg (reg1);
408 BuildMI (BB, X86::FLDr64, 1).addReg (reg2);
415 if (CompTy->isFloatingPoint()) {
416 // (Non-trapping) compare and pop twice.
417 BuildMI (BB, X86::FUCOMPP, 0);
418 // Move fp status word (concodes) to ax.
419 BuildMI (BB, X86::FNSTSWr8, 1, X86::AX);
420 // Load real concodes from ax.
421 BuildMI (BB, X86::SAHF, 1).addReg(X86::AH);
424 // Emit setOp instruction (extract concode; clobbers ax),
425 // using the following mapping:
426 // LLVM -> X86 signed X86 unsigned
428 // seteq -> sete sete
429 // setne -> setne setne
430 // setlt -> setl setb
431 // setgt -> setg seta
432 // setle -> setle setbe
433 // setge -> setge setae
435 static const unsigned OpcodeTab[2][6] = {
436 {X86::SETEr, X86::SETNEr, X86::SETBr, X86::SETAr, X86::SETBEr, X86::SETAEr},
437 {X86::SETEr, X86::SETNEr, X86::SETLr, X86::SETGr, X86::SETLEr, X86::SETGEr},
440 BuildMI(BB, OpcodeTab[CompTy->isSigned()][OpNum], 0, getReg(I));
443 /// promote32 - Emit instructions to turn a narrow operand into a 32-bit-wide
444 /// operand, in the specified target register.
446 ISel::promote32 (unsigned targetReg, Value *v)
448 unsigned vReg = getReg (v);
449 unsigned Class = getClass (v->getType ());
450 bool isUnsigned = v->getType ()->isUnsigned ();
451 assert (((Class == cByte) || (Class == cShort) || (Class == cInt))
452 && "Unpromotable operand class in promote32");
456 // Extend value into target register (8->32)
458 BuildMI (BB, X86::MOVZXr32r8, 1, targetReg).addReg (vReg);
460 BuildMI (BB, X86::MOVSXr32r8, 1, targetReg).addReg (vReg);
463 // Extend value into target register (16->32)
465 BuildMI (BB, X86::MOVZXr32r16, 1, targetReg).addReg (vReg);
467 BuildMI (BB, X86::MOVSXr32r16, 1, targetReg).addReg (vReg);
470 // Move value into target register (32->32)
471 BuildMI (BB, X86::MOVrr32, 1, targetReg).addReg (vReg);
476 /// 'ret' instruction - Here we are interested in meeting the x86 ABI. As such,
477 /// we have the following possibilities:
479 /// ret void: No return value, simply emit a 'ret' instruction
480 /// ret sbyte, ubyte : Extend value into EAX and return
481 /// ret short, ushort: Extend value into EAX and return
482 /// ret int, uint : Move value into EAX and return
483 /// ret pointer : Move value into EAX and return
484 /// ret long, ulong : Move value into EAX/EDX and return
485 /// ret float/double : Top of FP stack
488 ISel::visitReturnInst (ReturnInst &I)
490 if (I.getNumOperands () == 0)
492 // Emit a 'ret' instruction
493 BuildMI (BB, X86::RET, 0);
496 Value *rv = I.getOperand (0);
497 unsigned Class = getClass (rv->getType ());
500 // integral return values: extend or move into EAX and return.
504 promote32 (X86::EAX, rv);
506 // ret float/double: top of FP stack
508 case cFloat: // Floats
509 BuildMI (BB, X86::FLDr32, 1).addReg (getReg (rv));
511 case cDouble: // Doubles
512 BuildMI (BB, X86::FLDr64, 1).addReg (getReg (rv));
515 // ret long: use EAX(least significant 32 bits)/EDX (most
516 // significant 32)...uh, I think so Brain, but how do i call
517 // up the two parts of the value from inside this mouse
520 visitInstruction (I);
522 // Emit a 'ret' instruction
523 BuildMI (BB, X86::RET, 0);
526 /// visitBranchInst - Handle conditional and unconditional branches here. Note
527 /// that since code layout is frozen at this point, that if we are trying to
528 /// jump to a block that is the immediate successor of the current block, we can
529 /// just make a fall-through. (but we don't currently).
532 ISel::visitBranchInst (BranchInst & BI)
534 if (BI.isConditional ())
536 BasicBlock *ifTrue = BI.getSuccessor (0);
537 BasicBlock *ifFalse = BI.getSuccessor (1); // this is really unobvious
539 // simplest thing I can think of: compare condition with zero,
540 // followed by jump-if-equal to ifFalse, and jump-if-nonequal to
542 unsigned int condReg = getReg (BI.getCondition ());
543 BuildMI (BB, X86::CMPri8, 2).addReg (condReg).addZImm (0);
544 BuildMI (BB, X86::JNE, 1).addPCDisp (BI.getSuccessor (0));
545 BuildMI (BB, X86::JE, 1).addPCDisp (BI.getSuccessor (1));
547 else // unconditional branch
549 BuildMI (BB, X86::JMP, 1).addPCDisp (BI.getSuccessor (0));
553 /// visitCallInst - Push args on stack and do a procedure call instruction.
555 ISel::visitCallInst (CallInst & CI)
557 // keep a counter of how many bytes we pushed on the stack
558 unsigned bytesPushed = 0;
560 // Push the arguments on the stack in reverse order, as specified by
562 for (unsigned i = CI.getNumOperands()-1; i >= 1; --i)
564 Value *v = CI.getOperand (i);
565 switch (getClass (v->getType ()))
569 // Promote V to 32 bits wide, and move the result into EAX,
571 promote32 (X86::EAX, v);
572 BuildMI (BB, X86::PUSHr32, 1).addReg (X86::EAX);
577 unsigned Reg = getReg(v);
578 BuildMI (BB, X86::PUSHr32, 1).addReg(Reg);
583 // FIXME: long/ulong/double args not handled.
584 visitInstruction (CI);
589 if (Function *F = CI.getCalledFunction()) {
590 // Emit a CALL instruction with PC-relative displacement.
591 BuildMI(BB, X86::CALLpcrel32, 1).addPCDisp(F);
593 unsigned Reg = getReg(CI.getCalledValue());
594 BuildMI(BB, X86::CALLr32, 1).addReg(Reg);
597 // Adjust the stack by `bytesPushed' amount if non-zero
599 BuildMI (BB, X86::ADDri32,2,X86::ESP).addReg(X86::ESP).addZImm(bytesPushed);
601 // If there is a return value, scavenge the result from the location the call
604 if (CI.getType() != Type::VoidTy) {
605 unsigned resultTypeClass = getClass (CI.getType ());
606 switch (resultTypeClass) {
610 // Integral results are in %eax, or the appropriate portion
612 static const unsigned regRegMove[] = {
613 X86::MOVrr8, X86::MOVrr16, X86::MOVrr32
615 static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX };
616 BuildMI (BB, regRegMove[resultTypeClass], 1,
617 getReg (CI)).addReg (AReg[resultTypeClass]);
621 // Floating-point return values live in %st(0) (i.e., the top of
622 // the FP stack.) The general way to approach this is to do a
623 // FSTP to save the top of the FP stack on the real stack, then
624 // do a MOV to load the top of the real stack into the target
626 visitInstruction (CI); // FIXME: add the right args for the calls below
627 // BuildMI (BB, X86::FSTPm32, 0);
628 // BuildMI (BB, X86::MOVmr32, 0);
631 std::cerr << "Cannot get return value for call of type '"
632 << *CI.getType() << "'\n";
633 visitInstruction(CI);
638 /// visitSimpleBinary - Implement simple binary operators for integral types...
639 /// OperatorClass is one of: 0 for Add, 1 for Sub, 2 for And, 3 for Or,
642 void ISel::visitSimpleBinary(BinaryOperator &B, unsigned OperatorClass) {
643 if (B.getType() == Type::BoolTy) // FIXME: Handle bools for logicals
646 unsigned Class = getClass(B.getType());
647 if (Class > 2) // FIXME: Handle longs
650 static const unsigned OpcodeTab[][4] = {
651 // Arithmetic operators
652 { X86::ADDrr8, X86::ADDrr16, X86::ADDrr32, 0 }, // ADD
653 { X86::SUBrr8, X86::SUBrr16, X86::SUBrr32, 0 }, // SUB
656 { X86::ANDrr8, X86::ANDrr16, X86::ANDrr32, 0 }, // AND
657 { X86:: ORrr8, X86:: ORrr16, X86:: ORrr32, 0 }, // OR
658 { X86::XORrr8, X86::XORrr16, X86::XORrr32, 0 }, // XOR
661 unsigned Opcode = OpcodeTab[OperatorClass][Class];
662 unsigned Op0r = getReg(B.getOperand(0));
663 unsigned Op1r = getReg(B.getOperand(1));
664 BuildMI(BB, Opcode, 2, getReg(B)).addReg(Op0r).addReg(Op1r);
667 /// doMultiply - Emit appropriate instructions to multiply together
668 /// the registers op0Reg and op1Reg, and put the result in destReg.
669 /// The type of the result should be given as resultType.
670 void ISel::doMultiply(MachineBasicBlock *MBB, MachineBasicBlock::iterator &MBBI,
671 unsigned destReg, const Type *resultType,
672 unsigned op0Reg, unsigned op1Reg) {
673 unsigned Class = getClass (resultType);
676 assert (Class <= 2 && "Someday, we will learn how to multiply"
677 "longs and floating-point numbers. This is not that day.");
679 static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX };
680 static const unsigned MulOpcode[]={ X86::MULrr8, X86::MULrr16, X86::MULrr32 };
681 static const unsigned MovOpcode[]={ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 };
682 unsigned Reg = Regs[Class];
684 // Emit a MOV to put the first operand into the appropriately-sized
686 BMI(MBB, MBBI, MovOpcode[Class], 1, Reg).addReg (op0Reg);
688 // Emit the appropriate multiply instruction.
689 BMI(MBB, MBBI, MulOpcode[Class], 1).addReg (op1Reg);
691 // Emit another MOV to put the result into the destination register.
692 BMI(MBB, MBBI, MovOpcode[Class], 1, destReg).addReg (Reg);
695 /// visitMul - Multiplies are not simple binary operators because they must deal
696 /// with the EAX register explicitly.
698 void ISel::visitMul(BinaryOperator &I) {
699 unsigned DestReg = getReg(I);
700 unsigned Op0Reg = getReg(I.getOperand(0));
701 unsigned Op1Reg = getReg(I.getOperand(1));
702 MachineBasicBlock::iterator MBBI = BB->end();
703 doMultiply(BB, MBBI, DestReg, I.getType(), Op0Reg, Op1Reg);
707 /// visitDivRem - Handle division and remainder instructions... these
708 /// instruction both require the same instructions to be generated, they just
709 /// select the result from a different register. Note that both of these
710 /// instructions work differently for signed and unsigned operands.
712 void ISel::visitDivRem(BinaryOperator &I) {
713 unsigned Class = getClass(I.getType());
714 if (Class > 2) // FIXME: Handle longs
717 static const unsigned Regs[] ={ X86::AL , X86::AX , X86::EAX };
718 static const unsigned MovOpcode[]={ X86::MOVrr8, X86::MOVrr16, X86::MOVrr32 };
719 static const unsigned ExtOpcode[]={ X86::CBW , X86::CWD , X86::CDQ };
720 static const unsigned ClrOpcode[]={ X86::XORrr8, X86::XORrr16, X86::XORrr32 };
721 static const unsigned ExtRegs[] ={ X86::AH , X86::DX , X86::EDX };
723 static const unsigned DivOpcode[][4] = {
724 { X86::DIVrr8 , X86::DIVrr16 , X86::DIVrr32 , 0 }, // Unsigned division
725 { X86::IDIVrr8, X86::IDIVrr16, X86::IDIVrr32, 0 }, // Signed division
728 bool isSigned = I.getType()->isSigned();
729 unsigned Reg = Regs[Class];
730 unsigned ExtReg = ExtRegs[Class];
731 unsigned Op0Reg = getReg(I.getOperand(0));
732 unsigned Op1Reg = getReg(I.getOperand(1));
734 // Put the first operand into one of the A registers...
735 BuildMI(BB, MovOpcode[Class], 1, Reg).addReg(Op0Reg);
738 // Emit a sign extension instruction...
739 BuildMI(BB, ExtOpcode[Class], 0);
741 // If unsigned, emit a zeroing instruction... (reg = xor reg, reg)
742 BuildMI(BB, ClrOpcode[Class], 2, ExtReg).addReg(ExtReg).addReg(ExtReg);
745 // Emit the appropriate divide or remainder instruction...
746 BuildMI(BB, DivOpcode[isSigned][Class], 1).addReg(Op1Reg);
748 // Figure out which register we want to pick the result out of...
749 unsigned DestReg = (I.getOpcode() == Instruction::Div) ? Reg : ExtReg;
751 // Put the result into the destination register...
752 BuildMI(BB, MovOpcode[Class], 1, getReg(I)).addReg(DestReg);
756 /// Shift instructions: 'shl', 'sar', 'shr' - Some special cases here
757 /// for constant immediate shift values, and for constant immediate
758 /// shift values equal to 1. Even the general case is sort of special,
759 /// because the shift amount has to be in CL, not just any old register.
761 void ISel::visitShiftInst (ShiftInst &I) {
762 unsigned Op0r = getReg (I.getOperand(0));
763 unsigned DestReg = getReg(I);
764 bool isLeftShift = I.getOpcode() == Instruction::Shl;
765 bool isOperandSigned = I.getType()->isUnsigned();
766 unsigned OperandClass = getClass(I.getType());
768 if (OperandClass > 2)
769 visitInstruction(I); // Can't handle longs yet!
771 if (ConstantUInt *CUI = dyn_cast <ConstantUInt> (I.getOperand (1)))
773 // The shift amount is constant, guaranteed to be a ubyte. Get its value.
774 assert(CUI->getType() == Type::UByteTy && "Shift amount not a ubyte?");
775 unsigned char shAmt = CUI->getValue();
777 static const unsigned ConstantOperand[][4] = {
778 { X86::SHRir8, X86::SHRir16, X86::SHRir32, 0 }, // SHR
779 { X86::SARir8, X86::SARir16, X86::SARir32, 0 }, // SAR
780 { X86::SHLir8, X86::SHLir16, X86::SHLir32, 0 }, // SHL
781 { X86::SHLir8, X86::SHLir16, X86::SHLir32, 0 }, // SAL = SHL
784 const unsigned *OpTab = // Figure out the operand table to use
785 ConstantOperand[isLeftShift*2+isOperandSigned];
787 // Emit: <insn> reg, shamt (shift-by-immediate opcode "ir" form.)
788 BuildMI(BB, OpTab[OperandClass], 2, DestReg).addReg(Op0r).addZImm(shAmt);
792 // The shift amount is non-constant.
794 // In fact, you can only shift with a variable shift amount if
795 // that amount is already in the CL register, so we have to put it
799 // Emit: move cl, shiftAmount (put the shift amount in CL.)
800 BuildMI(BB, X86::MOVrr8, 1, X86::CL).addReg(getReg(I.getOperand(1)));
802 // This is a shift right (SHR).
803 static const unsigned NonConstantOperand[][4] = {
804 { X86::SHRrr8, X86::SHRrr16, X86::SHRrr32, 0 }, // SHR
805 { X86::SARrr8, X86::SARrr16, X86::SARrr32, 0 }, // SAR
806 { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32, 0 }, // SHL
807 { X86::SHLrr8, X86::SHLrr16, X86::SHLrr32, 0 }, // SAL = SHL
810 const unsigned *OpTab = // Figure out the operand table to use
811 NonConstantOperand[isLeftShift*2+isOperandSigned];
813 BuildMI(BB, OpTab[OperandClass], 1, DestReg).addReg(Op0r);
818 /// visitLoadInst - Implement LLVM load instructions in terms of the x86 'mov'
821 void ISel::visitLoadInst(LoadInst &I) {
822 unsigned Class = getClass(I.getType());
823 if (Class > 2) // FIXME: Handle longs and others...
826 static const unsigned Opcode[] = { X86::MOVmr8, X86::MOVmr16, X86::MOVmr32 };
828 unsigned AddressReg = getReg(I.getOperand(0));
829 addDirectMem(BuildMI(BB, Opcode[Class], 4, getReg(I)), AddressReg);
833 /// visitStoreInst - Implement LLVM store instructions in terms of the x86 'mov'
836 void ISel::visitStoreInst(StoreInst &I) {
837 unsigned Class = getClass(I.getOperand(0)->getType());
838 if (Class > 2) // FIXME: Handle longs and others...
841 static const unsigned Opcode[] = { X86::MOVrm8, X86::MOVrm16, X86::MOVrm32 };
843 unsigned ValReg = getReg(I.getOperand(0));
844 unsigned AddressReg = getReg(I.getOperand(1));
845 addDirectMem(BuildMI(BB, Opcode[Class], 1+4), AddressReg).addReg(ValReg);
849 /// visitCastInst - Here we have various kinds of copying with or without
850 /// sign extension going on.
852 ISel::visitCastInst (CastInst &CI)
854 const Type *targetType = CI.getType ();
855 Value *operand = CI.getOperand (0);
856 unsigned int operandReg = getReg (operand);
857 const Type *sourceType = operand->getType ();
858 unsigned int destReg = getReg (CI);
860 // Currently we handle:
864 // 2) cast {sbyte, ubyte} to {sbyte, ubyte}
865 // cast {short, ushort} to {ushort, short}
866 // cast {int, uint, ptr} to {int, uint, ptr}
868 // 3) cast {sbyte, ubyte} to {ushort, short}
869 // cast {sbyte, ubyte} to {int, uint, ptr}
870 // cast {short, ushort} to {int, uint, ptr}
872 // 4) cast {int, uint, ptr} to {short, ushort}
873 // cast {int, uint, ptr} to {sbyte, ubyte}
874 // cast {short, ushort} to {sbyte, ubyte}
876 // 1) Implement casts to bool by using compare on the operand followed
877 // by set if not zero on the result.
878 if (targetType == Type::BoolTy)
880 BuildMI (BB, X86::CMPri8, 2).addReg (operandReg).addZImm (0);
881 BuildMI (BB, X86::SETNEr, 1, destReg);
885 // 2) Implement casts between values of the same type class (as determined
886 // by getClass) by using a register-to-register move.
887 unsigned srcClass = getClassB (sourceType);
888 unsigned targClass = getClass (targetType);
889 static const unsigned regRegMove[] = {
890 X86::MOVrr8, X86::MOVrr16, X86::MOVrr32
892 if ((srcClass < cLong) && (targClass < cLong) && (srcClass == targClass))
894 BuildMI (BB, regRegMove[srcClass], 1, destReg).addReg (operandReg);
897 // 3) Handle cast of SMALLER int to LARGER int using a move with sign
898 // extension or zero extension, depending on whether the source type
900 if ((srcClass < cLong) && (targClass < cLong) && (srcClass < targClass))
902 static const unsigned ops[] = {
903 X86::MOVSXr16r8, X86::MOVSXr32r8, X86::MOVSXr32r16,
904 X86::MOVZXr16r8, X86::MOVZXr32r8, X86::MOVZXr32r16
906 unsigned srcSigned = sourceType->isSigned ();
907 BuildMI (BB, ops[3 * srcSigned + srcClass + targClass - 1], 1,
908 destReg).addReg (operandReg);
911 // 4) Handle cast of LARGER int to SMALLER int using a move to EAX
912 // followed by a move out of AX or AL.
913 if ((srcClass < cLong) && (targClass < cLong) && (srcClass > targClass))
915 static const unsigned AReg[] = { X86::AL, X86::AX, X86::EAX };
916 BuildMI (BB, regRegMove[srcClass], 1,
917 AReg[srcClass]).addReg (operandReg);
918 BuildMI (BB, regRegMove[targClass], 1, destReg).addReg (AReg[srcClass]);
921 // Anything we haven't handled already, we can't (yet) handle at all.
923 // FP to integral casts can be handled with FISTP to store onto the
924 // stack while converting to integer, followed by a MOV to load from
925 // the stack into the result register. Integral to FP casts can be
926 // handled with MOV to store onto the stack, followed by a FILD to
927 // load from the stack while converting to FP. For the moment, I
928 // can't quite get straight in my head how to borrow myself some
929 // stack space and write on it. Otherwise, this would be trivial.
930 visitInstruction (CI);
933 // ExactLog2 - This function solves for (Val == 1 << (N-1)) and returns N. It
934 // returns zero when the input is not exactly a power of two.
935 static unsigned ExactLog2(unsigned Val) {
936 if (Val == 0) return 0;
939 if (Val & 1) return 0;
946 /// visitGetElementPtrInst - I don't know, most programs don't have
947 /// getelementptr instructions, right? That means we can put off
948 /// implementing this, right? Right. This method emits machine
949 /// instructions to perform type-safe pointer arithmetic. I am
950 /// guessing this could be cleaned up somewhat to use fewer temporary
953 ISel::visitGetElementPtrInst (GetElementPtrInst &I)
955 unsigned outputReg = getReg (I);
956 MachineBasicBlock::iterator MI = BB->end();
957 emitGEPOperation(BB, MI, I.getOperand(0),
958 I.op_begin()+1, I.op_end(), outputReg);
961 void ISel::emitGEPOperation(MachineBasicBlock *MBB,
962 MachineBasicBlock::iterator &IP,
963 Value *Src, User::op_iterator IdxBegin,
964 User::op_iterator IdxEnd, unsigned TargetReg) {
965 const TargetData &TD = TM.getTargetData();
966 const Type *Ty = Src->getType();
967 unsigned basePtrReg = getReg(Src, MBB, IP);
969 // GEPs have zero or more indices; we must perform a struct access
970 // or array access for each one.
971 for (GetElementPtrInst::op_iterator oi = IdxBegin,
972 oe = IdxEnd; oi != oe; ++oi) {
974 unsigned nextBasePtrReg = makeAnotherReg(Type::UIntTy);
975 if (const StructType *StTy = dyn_cast <StructType> (Ty)) {
976 // It's a struct access. idx is the index into the structure,
977 // which names the field. This index must have ubyte type.
978 const ConstantUInt *CUI = cast <ConstantUInt> (idx);
979 assert (CUI->getType () == Type::UByteTy
980 && "Funny-looking structure index in GEP");
981 // Use the TargetData structure to pick out what the layout of
982 // the structure is in memory. Since the structure index must
983 // be constant, we can get its value and use it to find the
984 // right byte offset from the StructLayout class's list of
985 // structure member offsets.
986 unsigned idxValue = CUI->getValue ();
987 unsigned memberOffset =
988 TD.getStructLayout (StTy)->MemberOffsets[idxValue];
989 // Emit an ADD to add memberOffset to the basePtr.
990 BMI(MBB, IP, X86::ADDri32, 2,
991 nextBasePtrReg).addReg (basePtrReg).addZImm (memberOffset);
992 // The next type is the member of the structure selected by the
994 Ty = StTy->getElementTypes ()[idxValue];
995 } else if (const SequentialType *SqTy = cast <SequentialType>(Ty)) {
996 // It's an array or pointer access: [ArraySize x ElementType].
998 // idx is the index into the array. Unlike with structure
999 // indices, we may not know its actual value at code-generation
1001 assert(idx->getType() == Type::LongTy && "Bad GEP array index!");
1003 // We want to add basePtrReg to (idxReg * sizeof ElementType). First, we
1004 // must find the size of the pointed-to type (Not coincidentally, the next
1005 // type is the type of the elements in the array).
1006 Ty = SqTy->getElementType();
1007 unsigned elementSize = TD.getTypeSize(Ty);
1009 // If idxReg is a constant, we don't need to perform the multiply!
1010 if (ConstantSInt *CSI = dyn_cast<ConstantSInt>(idx)) {
1011 if (CSI->isNullValue()) {
1012 BMI(MBB, IP, X86::MOVrr32, 1, nextBasePtrReg).addReg(basePtrReg);
1014 unsigned Offset = elementSize*CSI->getValue();
1016 BMI(MBB, IP, X86::ADDri32, 2,
1017 nextBasePtrReg).addReg(basePtrReg).addZImm(Offset);
1019 } else if (elementSize == 1) {
1020 // If the element size is 1, we don't have to multiply, just add
1021 unsigned idxReg = getReg(idx, MBB, IP);
1022 BMI(MBB, IP, X86::ADDrr32, 2,
1023 nextBasePtrReg).addReg(basePtrReg).addReg(idxReg);
1025 unsigned idxReg = getReg(idx, MBB, IP);
1026 unsigned OffsetReg = makeAnotherReg(Type::UIntTy);
1027 if (unsigned Shift = ExactLog2(elementSize)) {
1028 // If the element size is exactly a power of 2, use a shift to get it.
1030 BMI(MBB, IP, X86::SHLir32, 2,
1031 OffsetReg).addReg(idxReg).addZImm(Shift-1);
1033 // Most general case, emit a multiply...
1034 unsigned elementSizeReg = makeAnotherReg(Type::LongTy);
1035 BMI(MBB, IP, X86::MOVir32, 1, elementSizeReg).addZImm(elementSize);
1037 // Emit a MUL to multiply the register holding the index by
1038 // elementSize, putting the result in OffsetReg.
1039 doMultiply(MBB, IP, OffsetReg, Type::LongTy, idxReg, elementSizeReg);
1041 // Emit an ADD to add OffsetReg to the basePtr.
1042 BMI(MBB, IP, X86::ADDrr32, 2,
1043 nextBasePtrReg).addReg (basePtrReg).addReg (OffsetReg);
1046 // Now that we are here, further indices refer to subtypes of this
1047 // one, so we don't need to worry about basePtrReg itself, anymore.
1048 basePtrReg = nextBasePtrReg;
1050 // After we have processed all the indices, the result is left in
1051 // basePtrReg. Move it to the register where we were expected to
1052 // put the answer. A 32-bit move should do it, because we are in
1054 BMI(MBB, IP, X86::MOVrr32, 1, TargetReg).addReg (basePtrReg);
1058 /// visitMallocInst - I know that personally, whenever I want to remember
1059 /// something, I have to clear off some space in my brain.
1061 ISel::visitMallocInst (MallocInst &I)
1063 // We assume that by this point, malloc instructions have been
1064 // lowered to calls, and dlsym will magically find malloc for us.
1065 // So we do not want to see malloc instructions here.
1066 visitInstruction (I);
1070 /// visitFreeInst - same story as MallocInst
1072 ISel::visitFreeInst (FreeInst &I)
1074 // We assume that by this point, free instructions have been
1075 // lowered to calls, and dlsym will magically find free for us.
1076 // So we do not want to see free instructions here.
1077 visitInstruction (I);
1081 /// visitAllocaInst - I want some stack space. Come on, man, I said I
1082 /// want some freakin' stack space.
1084 ISel::visitAllocaInst (AllocaInst &I)
1086 // Find the data size of the alloca inst's getAllocatedType.
1087 const Type *allocatedType = I.getAllocatedType ();
1088 const TargetData &TD = TM.DataLayout;
1089 unsigned allocatedTypeSize = TD.getTypeSize (allocatedType);
1090 // Keep stack 32-bit aligned.
1091 unsigned int allocatedTypeWords = allocatedTypeSize / 4;
1092 if (allocatedTypeSize % 4 != 0) { allocatedTypeWords++; }
1093 // Subtract size from stack pointer, thereby allocating some space.
1094 BuildMI(BB, X86::SUBri32, 2,
1095 X86::ESP).addReg(X86::ESP).addZImm(allocatedTypeWords * 4);
1096 // Put a pointer to the space into the result register, by copying
1097 // the stack pointer.
1098 BuildMI (BB, X86::MOVrr32, 1, getReg (I)).addReg (X86::ESP);
1102 /// createSimpleX86InstructionSelector - This pass converts an LLVM function
1103 /// into a machine code representation is a very simple peep-hole fashion. The
1104 /// generated code sucks but the implementation is nice and simple.
1106 Pass *createSimpleX86InstructionSelector(TargetMachine &TM) {
1107 return new ISel(TM);