1 //===-- InstSelectSimple.cpp - A simple instruction selector for SparcV8 --===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This file defines a simple peephole instruction selector for the V8 target
12 //===----------------------------------------------------------------------===//
15 #include "SparcV8InstrInfo.h"
16 #include "llvm/Support/Debug.h"
17 #include "llvm/Instructions.h"
18 #include "llvm/Pass.h"
19 #include "llvm/Constants.h"
20 #include "llvm/CodeGen/IntrinsicLowering.h"
21 #include "llvm/CodeGen/MachineInstrBuilder.h"
22 #include "llvm/CodeGen/MachineFrameInfo.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/SSARegMap.h"
26 #include "llvm/Target/TargetMachine.h"
27 #include "llvm/Support/GetElementPtrTypeIterator.h"
28 #include "llvm/Support/InstVisitor.h"
29 #include "llvm/Support/CFG.h"
33 struct V8ISel : public FunctionPass, public InstVisitor<V8ISel> {
35 MachineFunction *F; // The function we are compiling into
36 MachineBasicBlock *BB; // The current MBB we are compiling
37 int VarArgsOffset; // Offset from fp for start of varargs area
39 std::map<Value*, unsigned> RegMap; // Mapping between Val's and SSA Regs
41 // MBBMap - Mapping between LLVM BB -> Machine BB
42 std::map<const BasicBlock*, MachineBasicBlock*> MBBMap;
44 V8ISel(TargetMachine &tm) : TM(tm), F(0), BB(0) {}
46 /// runOnFunction - Top level implementation of instruction selection for
47 /// the entire function.
49 bool runOnFunction(Function &Fn);
51 virtual const char *getPassName() const {
52 return "SparcV8 Simple Instruction Selection";
55 /// emitGEPOperation - Common code shared between visitGetElementPtrInst and
56 /// constant expression GEP support.
58 void emitGEPOperation(MachineBasicBlock *BB, MachineBasicBlock::iterator IP,
59 Value *Src, User::op_iterator IdxBegin,
60 User::op_iterator IdxEnd, unsigned TargetReg);
62 /// emitCastOperation - Common code shared between visitCastInst and
63 /// constant expression cast support.
65 void emitCastOperation(MachineBasicBlock *BB,MachineBasicBlock::iterator IP,
66 Value *Src, const Type *DestTy, unsigned TargetReg);
68 /// emitIntegerCast, emitFPToIntegerCast - Helper methods for
69 /// emitCastOperation.
71 unsigned emitIntegerCast (MachineBasicBlock *BB,
72 MachineBasicBlock::iterator IP,
73 const Type *oldTy, unsigned SrcReg,
74 const Type *newTy, unsigned DestReg);
75 void emitFPToIntegerCast (MachineBasicBlock *BB,
76 MachineBasicBlock::iterator IP, const Type *oldTy,
77 unsigned SrcReg, const Type *newTy,
80 /// visitBasicBlock - This method is called when we are visiting a new basic
81 /// block. This simply creates a new MachineBasicBlock to emit code into
82 /// and adds it to the current MachineFunction. Subsequent visit* for
83 /// instructions will be invoked for all instructions in the basic block.
85 void visitBasicBlock(BasicBlock &LLVM_BB) {
86 BB = MBBMap[&LLVM_BB];
89 void emitOp64LibraryCall (MachineBasicBlock *MBB,
90 MachineBasicBlock::iterator IP,
91 unsigned DestReg, const char *FuncName,
92 unsigned Op0Reg, unsigned Op1Reg);
93 void emitShift64 (MachineBasicBlock *MBB, MachineBasicBlock::iterator IP,
94 Instruction &I, unsigned DestReg, unsigned Op0Reg,
96 void visitBinaryOperator(Instruction &I);
97 void visitShiftInst (ShiftInst &SI) { visitBinaryOperator (SI); }
98 void visitSetCondInst(SetCondInst &I);
99 void visitCallInst(CallInst &I);
100 void visitReturnInst(ReturnInst &I);
101 void visitBranchInst(BranchInst &I);
102 void visitUnreachableInst(UnreachableInst &I) {}
103 void visitCastInst(CastInst &I);
104 void visitVANextInst(VANextInst &I);
105 void visitVAArgInst(VAArgInst &I);
106 void visitLoadInst(LoadInst &I);
107 void visitStoreInst(StoreInst &I);
108 void visitPHINode(PHINode &I) {} // PHI nodes handled by second pass
109 void visitGetElementPtrInst(GetElementPtrInst &I);
110 void visitAllocaInst(AllocaInst &I);
112 void visitInstruction(Instruction &I) {
113 std::cerr << "Unhandled instruction: " << I;
117 /// LowerUnknownIntrinsicFunctionCalls - This performs a prepass over the
118 /// function, lowering any calls to unknown intrinsic functions into the
119 /// equivalent LLVM code.
120 void LowerUnknownIntrinsicFunctionCalls(Function &F);
121 void visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI);
123 void LoadArgumentsToVirtualRegs(Function *F);
125 /// SelectPHINodes - Insert machine code to generate phis. This is tricky
126 /// because we have to generate our sources into the source basic blocks,
127 /// not the current one.
129 void SelectPHINodes();
131 /// copyConstantToRegister - Output the instructions required to put the
132 /// specified constant into the specified register.
134 void copyConstantToRegister(MachineBasicBlock *MBB,
135 MachineBasicBlock::iterator IP,
136 Constant *C, unsigned R);
138 /// makeAnotherReg - This method returns the next register number we haven't
141 /// Long values are handled somewhat specially. They are always allocated
142 /// as pairs of 32 bit integer values. The register number returned is the
143 /// lower 32 bits of the long value, and the regNum+1 is the upper 32 bits
144 /// of the long value.
146 unsigned makeAnotherReg(const Type *Ty) {
147 assert(dynamic_cast<const SparcV8RegisterInfo*>(TM.getRegisterInfo()) &&
148 "Current target doesn't have SparcV8 reg info??");
149 const SparcV8RegisterInfo *MRI =
150 static_cast<const SparcV8RegisterInfo*>(TM.getRegisterInfo());
151 if (Ty == Type::LongTy || Ty == Type::ULongTy) {
152 const TargetRegisterClass *RC = MRI->getRegClassForType(Type::IntTy);
153 // Create the lower part
154 F->getSSARegMap()->createVirtualRegister(RC);
155 // Create the upper part.
156 return F->getSSARegMap()->createVirtualRegister(RC)-1;
159 // Add the mapping of regnumber => reg class to MachineFunction
160 const TargetRegisterClass *RC = MRI->getRegClassForType(Ty);
161 return F->getSSARegMap()->createVirtualRegister(RC);
164 unsigned getReg(Value &V) { return getReg (&V); } // allow refs.
165 unsigned getReg(Value *V) {
166 // Just append to the end of the current bb.
167 MachineBasicBlock::iterator It = BB->end();
168 return getReg(V, BB, It);
170 unsigned getReg(Value *V, MachineBasicBlock *MBB,
171 MachineBasicBlock::iterator IPt) {
172 unsigned &Reg = RegMap[V];
174 Reg = makeAnotherReg(V->getType());
177 // If this operand is a constant, emit the code to copy the constant into
178 // the register here...
180 if (Constant *C = dyn_cast<Constant>(V)) {
181 copyConstantToRegister(MBB, IPt, C, Reg);
182 RegMap.erase(V); // Assign a new name to this constant if ref'd again
183 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(V)) {
184 // Move the address of the global into the register
185 unsigned TmpReg = makeAnotherReg(V->getType());
186 BuildMI (*MBB, IPt, V8::SETHIi, 1, TmpReg).addGlobalAddress (GV);
187 BuildMI (*MBB, IPt, V8::ORri, 2, Reg).addReg (TmpReg)
188 .addGlobalAddress (GV);
189 RegMap.erase(V); // Assign a new name to this address if ref'd again
198 FunctionPass *llvm::createSparcV8SimpleInstructionSelector(TargetMachine &TM) {
199 return new V8ISel(TM);
203 cByte, cShort, cInt, cLong, cFloat, cDouble
206 static TypeClass getClass (const Type *T) {
207 switch (T->getTypeID()) {
208 case Type::UByteTyID: case Type::SByteTyID: return cByte;
209 case Type::UShortTyID: case Type::ShortTyID: return cShort;
210 case Type::PointerTyID:
211 case Type::UIntTyID: case Type::IntTyID: return cInt;
212 case Type::ULongTyID: case Type::LongTyID: return cLong;
213 case Type::FloatTyID: return cFloat;
214 case Type::DoubleTyID: return cDouble;
216 assert (0 && "Type of unknown class passed to getClass?");
221 static TypeClass getClassB(const Type *T) {
222 if (T == Type::BoolTy) return cByte;
226 /// copyConstantToRegister - Output the instructions required to put the
227 /// specified constant into the specified register.
229 void V8ISel::copyConstantToRegister(MachineBasicBlock *MBB,
230 MachineBasicBlock::iterator IP,
231 Constant *C, unsigned R) {
232 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(C)) {
233 switch (CE->getOpcode()) {
234 case Instruction::GetElementPtr:
235 emitGEPOperation(MBB, IP, CE->getOperand(0),
236 CE->op_begin()+1, CE->op_end(), R);
238 case Instruction::Cast:
239 emitCastOperation(MBB, IP, CE->getOperand(0), CE->getType(), R);
242 std::cerr << "Copying this constant expr not yet handled: " << *CE;
245 } else if (isa<UndefValue>(C)) {
246 BuildMI(*MBB, IP, V8::IMPLICIT_DEF, 0, R);
247 if (getClassB (C->getType ()) == cLong)
248 BuildMI(*MBB, IP, V8::IMPLICIT_DEF, 0, R+1);
252 if (C->getType()->isIntegral ()) {
254 unsigned Class = getClassB (C->getType ());
255 if (Class == cLong) {
256 unsigned TmpReg = makeAnotherReg (Type::IntTy);
257 unsigned TmpReg2 = makeAnotherReg (Type::IntTy);
258 // Copy the value into the register pair.
259 // R = top(more-significant) half, R+1 = bottom(less-significant) half
260 uint64_t Val = cast<ConstantInt>(C)->getRawValue();
261 copyConstantToRegister(MBB, IP, ConstantUInt::get(Type::UIntTy,
263 copyConstantToRegister(MBB, IP, ConstantUInt::get(Type::UIntTy,
264 Val & 0xffffffffU), R+1);
268 assert(Class <= cInt && "Type not handled yet!");
270 if (C->getType() == Type::BoolTy) {
271 Val = (C == ConstantBool::True);
273 ConstantInt *CI = cast<ConstantInt> (C);
274 Val = CI->getRawValue ();
277 case cByte: Val = (int8_t) Val; break;
278 case cShort: Val = (int16_t) Val; break;
279 case cInt: Val = (int32_t) Val; break;
281 std::cerr << "Offending constant: " << *C << "\n";
282 assert (0 && "Can't copy this kind of constant into register yet");
286 BuildMI (*MBB, IP, V8::ORrr, 2, R).addReg (V8::G0).addReg(V8::G0);
287 } else if (((int64_t)Val >= -4096) && ((int64_t)Val <= 4095)) {
288 BuildMI (*MBB, IP, V8::ORri, 2, R).addReg (V8::G0).addSImm(Val);
290 unsigned TmpReg = makeAnotherReg (C->getType ());
291 BuildMI (*MBB, IP, V8::SETHIi, 1, TmpReg)
292 .addSImm (((uint32_t) Val) >> 10);
293 BuildMI (*MBB, IP, V8::ORri, 2, R).addReg (TmpReg)
294 .addSImm (((uint32_t) Val) & 0x03ff);
297 } else if (ConstantFP *CFP = dyn_cast<ConstantFP>(C)) {
298 // We need to spill the constant to memory...
299 MachineConstantPool *CP = F->getConstantPool();
300 unsigned CPI = CP->getConstantPoolIndex(CFP);
301 const Type *Ty = CFP->getType();
302 unsigned TmpReg = makeAnotherReg (Type::UIntTy);
303 unsigned AddrReg = makeAnotherReg (Type::UIntTy);
305 assert(Ty == Type::FloatTy || Ty == Type::DoubleTy && "Unknown FP type!");
306 unsigned LoadOpcode = Ty == Type::FloatTy ? V8::LDFri : V8::LDDFri;
307 BuildMI (*MBB, IP, V8::SETHIi, 1, TmpReg).addConstantPoolIndex (CPI);
308 BuildMI (*MBB, IP, V8::ORri, 2, AddrReg).addReg (TmpReg)
309 .addConstantPoolIndex (CPI);
310 BuildMI (*MBB, IP, LoadOpcode, 2, R).addReg (AddrReg).addSImm (0);
311 } else if (isa<ConstantPointerNull>(C)) {
312 // Copy zero (null pointer) to the register.
313 BuildMI (*MBB, IP, V8::ORri, 2, R).addReg (V8::G0).addSImm (0);
314 } else if (GlobalValue *GV = dyn_cast<GlobalValue>(C)) {
315 // Copy it with a SETHI/OR pair; the JIT + asmwriter should recognize
316 // that SETHI %reg,global == SETHI %reg,%hi(global) and
317 // OR %reg,global,%reg == OR %reg,%lo(global),%reg.
318 unsigned TmpReg = makeAnotherReg (C->getType ());
319 BuildMI (*MBB, IP, V8::SETHIi, 1, TmpReg).addGlobalAddress(GV);
320 BuildMI (*MBB, IP, V8::ORri, 2, R).addReg(TmpReg).addGlobalAddress(GV);
322 std::cerr << "Offending constant: " << *C << "\n";
323 assert (0 && "Can't copy this kind of constant into register yet");
327 void V8ISel::LoadArgumentsToVirtualRegs (Function *LF) {
328 static const unsigned IncomingArgRegs[] = { V8::I0, V8::I1, V8::I2,
329 V8::I3, V8::I4, V8::I5 };
331 // Add IMPLICIT_DEFs of input regs.
333 for (Function::aiterator I = LF->abegin(), E = LF->aend();
334 I != E && ArgNo < 6; ++I, ++ArgNo) {
335 switch (getClassB(I->getType())) {
340 BuildMI(BB, V8::IMPLICIT_DEF, 0, IncomingArgRegs[ArgNo]);
344 // Double and Long use register pairs.
345 BuildMI(BB, V8::IMPLICIT_DEF, 0, IncomingArgRegs[ArgNo]);
348 BuildMI(BB, V8::IMPLICIT_DEF, 0, IncomingArgRegs[ArgNo]);
351 assert (0 && "type not handled");
356 const unsigned *IAREnd = &IncomingArgRegs[6];
357 const unsigned *IAR = &IncomingArgRegs[0];
358 unsigned ArgOffset = 68;
360 // Store registers onto stack if this is a varargs function.
361 // FIXME: This doesn't really pertain to "loading arguments into
362 // virtual registers", so it's not clear that it really belongs here.
363 // FIXME: We could avoid storing any args onto the stack that don't
364 // need to be in memory, because they come before the ellipsis in the
365 // parameter list (and thus could never be accessed through va_arg).
366 if (LF->getFunctionType ()->isVarArg ()) {
367 for (unsigned i = 0; i < 6; ++i) {
368 int FI = F->getFrameInfo()->CreateFixedObject(4, ArgOffset);
369 assert (IAR != IAREnd
370 && "About to dereference past end of IncomingArgRegs");
371 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (0).addReg (*IAR++);
374 // Reset the pointers now that we're done.
376 IAR = &IncomingArgRegs[0];
379 // Copy args out of their incoming hard regs or stack slots into virtual regs.
380 for (Function::aiterator I = LF->abegin(), E = LF->aend(); I != E; ++I) {
382 unsigned ArgReg = getReg (A);
383 if (getClassB (A.getType ()) < cLong) {
384 // Get it out of the incoming arg register
385 if (ArgOffset < 92) {
386 assert (IAR != IAREnd
387 && "About to dereference past end of IncomingArgRegs");
388 BuildMI (BB, V8::ORrr, 2, ArgReg).addReg (V8::G0).addReg (*IAR++);
390 int FI = F->getFrameInfo()->CreateFixedObject(4, ArgOffset);
391 BuildMI (BB, V8::LD, 3, ArgReg).addFrameIndex (FI).addSImm (0);
394 } else if (getClassB (A.getType ()) == cFloat) {
395 if (ArgOffset < 92) {
396 // Single-fp args are passed in integer registers; go through
397 // memory to get them out of integer registers and back into fp. (Bleh!)
398 unsigned FltAlign = TM.getTargetData().getFloatAlignment();
399 int FI = F->getFrameInfo()->CreateStackObject(4, FltAlign);
400 assert (IAR != IAREnd
401 && "About to dereference past end of IncomingArgRegs");
402 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (0).addReg (*IAR++);
403 BuildMI (BB, V8::LDFri, 2, ArgReg).addFrameIndex (FI).addSImm (0);
405 int FI = F->getFrameInfo()->CreateFixedObject(4, ArgOffset);
406 BuildMI (BB, V8::LDFri, 3, ArgReg).addFrameIndex (FI).addSImm (0);
409 } else if (getClassB (A.getType ()) == cDouble) {
410 // Double-fp args are passed in pairs of integer registers; go through
411 // memory to get them out of integer registers and back into fp. (Bleh!)
412 // We'd like to 'ldd' these right out of the incoming-args area,
413 // but it might not be 8-byte aligned (e.g., call x(int x, double d)).
414 unsigned DblAlign = TM.getTargetData().getDoubleAlignment();
415 int FI = F->getFrameInfo()->CreateStackObject(8, DblAlign);
416 if (ArgOffset < 92 && IAR != IAREnd) {
417 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (0).addReg (*IAR++);
419 unsigned TempReg = makeAnotherReg (Type::IntTy);
420 BuildMI (BB, V8::LD, 2, TempReg).addFrameIndex (FI).addSImm (0);
421 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (0).addReg (TempReg);
424 if (ArgOffset < 92 && IAR != IAREnd) {
425 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (4).addReg (*IAR++);
427 unsigned TempReg = makeAnotherReg (Type::IntTy);
428 BuildMI (BB, V8::LD, 2, TempReg).addFrameIndex (FI).addSImm (4);
429 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (4).addReg (TempReg);
432 BuildMI (BB, V8::LDDFri, 2, ArgReg).addFrameIndex (FI).addSImm (0);
433 } else if (getClassB (A.getType ()) == cLong) {
434 // do the first half...
435 if (ArgOffset < 92) {
436 assert (IAR != IAREnd
437 && "About to dereference past end of IncomingArgRegs");
438 BuildMI (BB, V8::ORrr, 2, ArgReg).addReg (V8::G0).addReg (*IAR++);
440 int FI = F->getFrameInfo()->CreateFixedObject(4, ArgOffset);
441 BuildMI (BB, V8::LD, 2, ArgReg).addFrameIndex (FI).addSImm (0);
444 // ...then do the second half
445 if (ArgOffset < 92) {
446 assert (IAR != IAREnd
447 && "About to dereference past end of IncomingArgRegs");
448 BuildMI (BB, V8::ORrr, 2, ArgReg+1).addReg (V8::G0).addReg (*IAR++);
450 int FI = F->getFrameInfo()->CreateFixedObject(4, ArgOffset);
451 BuildMI (BB, V8::LD, 2, ArgReg+1).addFrameIndex (FI).addSImm (0);
455 assert (0 && "Unknown class?!");
459 // If the function takes variable number of arguments, remember the fp
460 // offset for the start of the first vararg value... this is used to expand
462 if (LF->getFunctionType ()->isVarArg ())
463 VarArgsOffset = ArgOffset;
466 void V8ISel::SelectPHINodes() {
467 const TargetInstrInfo &TII = *TM.getInstrInfo();
468 const Function &LF = *F->getFunction(); // The LLVM function...
469 for (Function::const_iterator I = LF.begin(), E = LF.end(); I != E; ++I) {
470 const BasicBlock *BB = I;
471 MachineBasicBlock &MBB = *MBBMap[I];
473 // Loop over all of the PHI nodes in the LLVM basic block...
474 MachineBasicBlock::iterator PHIInsertPoint = MBB.begin();
475 for (BasicBlock::const_iterator I = BB->begin();
476 PHINode *PN = const_cast<PHINode*>(dyn_cast<PHINode>(I)); ++I) {
478 // Create a new machine instr PHI node, and insert it.
479 unsigned PHIReg = getReg(*PN);
480 MachineInstr *PhiMI = BuildMI(MBB, PHIInsertPoint,
481 V8::PHI, PN->getNumOperands(), PHIReg);
483 MachineInstr *LongPhiMI = 0;
484 if (PN->getType() == Type::LongTy || PN->getType() == Type::ULongTy)
485 LongPhiMI = BuildMI(MBB, PHIInsertPoint,
486 V8::PHI, PN->getNumOperands(), PHIReg+1);
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 = 0;
496 for (MachineBasicBlock::pred_iterator PI = MBB.pred_begin (),
497 PE = MBB.pred_end (); PI != PE; ++PI)
498 if (PN->getIncomingBlock(i) == (*PI)->getBasicBlock()) {
502 assert (PredMBB && "Couldn't find incoming machine-cfg edge for phi");
505 std::map<MachineBasicBlock*, unsigned>::iterator EntryIt =
506 PHIValues.lower_bound(PredMBB);
508 if (EntryIt != PHIValues.end() && EntryIt->first == PredMBB) {
509 // We already inserted an initialization of the register for this
510 // predecessor. Recycle it.
511 ValReg = EntryIt->second;
514 // Get the incoming value into a virtual register.
516 Value *Val = PN->getIncomingValue(i);
518 // If this is a constant or GlobalValue, we may have to insert code
519 // into the basic block to compute it into a virtual register.
520 if ((isa<Constant>(Val) && !isa<ConstantExpr>(Val)) ||
521 isa<GlobalValue>(Val)) {
522 // Simple constants get emitted at the end of the basic block,
523 // before any terminator instructions. We "know" that the code to
524 // move a constant into a register will never clobber any flags.
525 ValReg = getReg(Val, PredMBB, PredMBB->getFirstTerminator());
527 // Because we don't want to clobber any values which might be in
528 // physical registers with the computation of this constant (which
529 // might be arbitrarily complex if it is a constant expression),
530 // just insert the computation at the top of the basic block.
531 MachineBasicBlock::iterator PI = PredMBB->begin();
533 // Skip over any PHI nodes though!
534 while (PI != PredMBB->end() && PI->getOpcode() == V8::PHI)
537 ValReg = getReg(Val, PredMBB, PI);
540 // Remember that we inserted a value for this PHI for this predecessor
541 PHIValues.insert(EntryIt, std::make_pair(PredMBB, ValReg));
544 PhiMI->addRegOperand(ValReg);
545 PhiMI->addMachineBasicBlockOperand(PredMBB);
547 LongPhiMI->addRegOperand(ValReg+1);
548 LongPhiMI->addMachineBasicBlockOperand(PredMBB);
552 // Now that we emitted all of the incoming values for the PHI node, make
553 // sure to reposition the InsertPoint after the PHI that we just added.
554 // This is needed because we might have inserted a constant into this
555 // block, right after the PHI's which is before the old insert point!
556 PHIInsertPoint = LongPhiMI ? LongPhiMI : PhiMI;
562 bool V8ISel::runOnFunction(Function &Fn) {
563 // First pass over the function, lower any unknown intrinsic functions
564 // with the IntrinsicLowering class.
565 LowerUnknownIntrinsicFunctionCalls(Fn);
567 F = &MachineFunction::construct(&Fn, TM);
569 // Create all of the machine basic blocks for the function...
570 for (Function::iterator I = Fn.begin(), E = Fn.end(); I != E; ++I)
571 F->getBasicBlockList().push_back(MBBMap[I] = new MachineBasicBlock(I));
575 // Set up a frame object for the return address. This is used by the
576 // llvm.returnaddress & llvm.frameaddress intrinisics.
577 //ReturnAddressIndex = F->getFrameInfo()->CreateFixedObject(4, -4);
579 // Copy incoming arguments off of the stack and out of fixed registers.
580 LoadArgumentsToVirtualRegs(&Fn);
582 // Instruction select everything except PHI nodes
585 // Select the PHI nodes
591 // We always build a machine code representation for the function
595 void V8ISel::visitCastInst(CastInst &I) {
596 Value *Op = I.getOperand(0);
597 unsigned DestReg = getReg(I);
598 MachineBasicBlock::iterator MI = BB->end();
599 emitCastOperation(BB, MI, Op, I.getType(), DestReg);
602 unsigned V8ISel::emitIntegerCast (MachineBasicBlock *BB,
603 MachineBasicBlock::iterator IP, const Type *oldTy,
604 unsigned SrcReg, const Type *newTy,
606 if (oldTy == newTy) {
607 // No-op cast - just emit a copy; assume the reg. allocator will zap it.
608 BuildMI (*BB, IP, V8::ORrr, 2, DestReg).addReg (V8::G0).addReg(SrcReg);
611 // Emit left-shift, then right-shift to sign- or zero-extend.
612 unsigned TmpReg = makeAnotherReg (newTy);
613 unsigned shiftWidth = 32 - (8 * TM.getTargetData ().getTypeSize (newTy));
614 BuildMI (*BB, IP, V8::SLLri, 2, TmpReg).addZImm (shiftWidth).addReg(SrcReg);
615 if (newTy->isSigned ()) { // sign-extend with SRA
616 BuildMI(*BB, IP, V8::SRAri, 2, DestReg).addZImm (shiftWidth).addReg(TmpReg);
617 } else { // zero-extend with SRL
618 BuildMI(*BB, IP, V8::SRLri, 2, DestReg).addZImm (shiftWidth).addReg(TmpReg);
620 // Return the temp reg. in case this is one half of a cast to long.
624 void V8ISel::emitFPToIntegerCast (MachineBasicBlock *BB,
625 MachineBasicBlock::iterator IP,
626 const Type *oldTy, unsigned SrcReg,
627 const Type *newTy, unsigned DestReg) {
628 unsigned FPCastOpcode, FPStoreOpcode, FPSize, FPAlign;
629 unsigned oldTyClass = getClassB(oldTy);
630 if (oldTyClass == cFloat) {
631 FPCastOpcode = V8::FSTOI; FPStoreOpcode = V8::STFri; FPSize = 4;
632 FPAlign = TM.getTargetData().getFloatAlignment();
633 } else { // it's a double
634 FPCastOpcode = V8::FDTOI; FPStoreOpcode = V8::STDFri; FPSize = 8;
635 FPAlign = TM.getTargetData().getDoubleAlignment();
637 unsigned TempReg = makeAnotherReg (oldTy);
638 BuildMI (*BB, IP, FPCastOpcode, 1, TempReg).addReg (SrcReg);
639 int FI = F->getFrameInfo()->CreateStackObject(FPSize, FPAlign);
640 BuildMI (*BB, IP, FPStoreOpcode, 3).addFrameIndex (FI).addSImm (0)
642 unsigned TempReg2 = makeAnotherReg (newTy);
643 BuildMI (*BB, IP, V8::LD, 3, TempReg2).addFrameIndex (FI).addSImm (0);
644 emitIntegerCast (BB, IP, Type::IntTy, TempReg2, newTy, DestReg);
647 /// emitCastOperation - Common code shared between visitCastInst and constant
648 /// expression cast support.
650 void V8ISel::emitCastOperation(MachineBasicBlock *BB,
651 MachineBasicBlock::iterator IP, Value *Src,
652 const Type *DestTy, unsigned DestReg) {
653 const Type *SrcTy = Src->getType();
654 unsigned SrcClass = getClassB(SrcTy);
655 unsigned DestClass = getClassB(DestTy);
656 unsigned SrcReg = getReg(Src, BB, IP);
658 const Type *oldTy = SrcTy;
659 const Type *newTy = DestTy;
660 unsigned oldTyClass = SrcClass;
661 unsigned newTyClass = DestClass;
663 if (oldTyClass < cLong && newTyClass < cLong) {
664 emitIntegerCast (BB, IP, oldTy, SrcReg, newTy, DestReg);
665 } else switch (newTyClass) {
669 switch (oldTyClass) {
671 // Treat it like a cast from the lower half of the value.
672 emitIntegerCast (BB, IP, Type::IntTy, SrcReg+1, newTy, DestReg);
676 emitFPToIntegerCast (BB, IP, oldTy, SrcReg, newTy, DestReg);
678 default: goto not_yet;
683 switch (oldTyClass) {
684 case cLong: goto not_yet;
686 BuildMI (*BB, IP, V8::FMOVS, 1, DestReg).addReg (SrcReg);
689 BuildMI (*BB, IP, V8::FDTOS, 1, DestReg).addReg (SrcReg);
692 unsigned FltAlign = TM.getTargetData().getFloatAlignment();
693 // cast integer type to float. Store it to a stack slot and then load
694 // it using ldf into a floating point register. then do fitos.
695 unsigned TmpReg = makeAnotherReg (newTy);
696 int FI = F->getFrameInfo()->CreateStackObject(4, FltAlign);
697 BuildMI (*BB, IP, V8::ST, 3).addFrameIndex (FI).addSImm (0)
699 BuildMI (*BB, IP, V8::LDFri, 2, TmpReg).addFrameIndex (FI).addSImm (0);
700 BuildMI (*BB, IP, V8::FITOS, 1, DestReg).addReg(TmpReg);
707 switch (oldTyClass) {
708 case cLong: goto not_yet;
710 BuildMI (*BB, IP, V8::FSTOD, 1, DestReg).addReg (SrcReg);
712 case cDouble: // use double move pseudo-instr
713 BuildMI (*BB, IP, V8::FpMOVD, 1, DestReg).addReg (SrcReg);
716 unsigned DoubleAlignment = TM.getTargetData().getDoubleAlignment();
717 unsigned TmpReg = makeAnotherReg (newTy);
718 int FI = F->getFrameInfo()->CreateStackObject(8, DoubleAlignment);
719 BuildMI (*BB, IP, V8::ST, 3).addFrameIndex (FI).addSImm (0)
721 BuildMI (*BB, IP, V8::LDDFri, 2, TmpReg).addFrameIndex (FI).addSImm (0);
722 BuildMI (*BB, IP, V8::FITOD, 1, DestReg).addReg(TmpReg);
729 switch (oldTyClass) {
733 // Cast to (u)int in the bottom half, and sign(zero) extend in the top
735 const Type *OldHalfTy = oldTy->isSigned() ? Type::IntTy : Type::UIntTy;
736 const Type *NewHalfTy = newTy->isSigned() ? Type::IntTy : Type::UIntTy;
737 unsigned TempReg = emitIntegerCast (BB, IP, OldHalfTy, SrcReg,
738 NewHalfTy, DestReg+1);
739 if (newTy->isSigned ()) {
740 BuildMI (*BB, IP, V8::SRAri, 2, DestReg).addReg (TempReg)
743 BuildMI (*BB, IP, V8::ORrr, 2, DestReg).addReg (V8::G0)
749 // Just copy both halves.
750 BuildMI (*BB, IP, V8::ORrr, 2, DestReg).addReg (V8::G0).addReg (SrcReg);
751 BuildMI (*BB, IP, V8::ORrr, 2, DestReg+1).addReg (V8::G0)
754 default: goto not_yet;
758 default: goto not_yet;
762 std::cerr << "Sorry, cast still unsupported: SrcTy = " << *SrcTy
763 << ", DestTy = " << *DestTy << "\n";
767 void V8ISel::visitLoadInst(LoadInst &I) {
768 unsigned DestReg = getReg (I);
769 unsigned PtrReg = getReg (I.getOperand (0));
770 switch (getClassB (I.getType ())) {
772 if (I.getType ()->isSigned ())
773 BuildMI (BB, V8::LDSB, 2, DestReg).addReg (PtrReg).addSImm(0);
775 BuildMI (BB, V8::LDUB, 2, DestReg).addReg (PtrReg).addSImm(0);
778 if (I.getType ()->isSigned ())
779 BuildMI (BB, V8::LDSH, 2, DestReg).addReg (PtrReg).addSImm(0);
781 BuildMI (BB, V8::LDUH, 2, DestReg).addReg (PtrReg).addSImm(0);
784 BuildMI (BB, V8::LD, 2, DestReg).addReg (PtrReg).addSImm(0);
787 BuildMI (BB, V8::LD, 2, DestReg).addReg (PtrReg).addSImm(0);
788 BuildMI (BB, V8::LD, 2, DestReg+1).addReg (PtrReg).addSImm(4);
791 BuildMI (BB, V8::LDFri, 2, DestReg).addReg (PtrReg).addSImm(0);
794 BuildMI (BB, V8::LDDFri, 2, DestReg).addReg (PtrReg).addSImm(0);
797 std::cerr << "Load instruction not handled: " << I;
803 void V8ISel::visitStoreInst(StoreInst &I) {
804 Value *SrcVal = I.getOperand (0);
805 unsigned SrcReg = getReg (SrcVal);
806 unsigned PtrReg = getReg (I.getOperand (1));
807 switch (getClassB (SrcVal->getType ())) {
809 BuildMI (BB, V8::STB, 3).addReg (PtrReg).addSImm (0).addReg (SrcReg);
812 BuildMI (BB, V8::STH, 3).addReg (PtrReg).addSImm (0).addReg (SrcReg);
815 BuildMI (BB, V8::ST, 3).addReg (PtrReg).addSImm (0).addReg (SrcReg);
818 BuildMI (BB, V8::ST, 3).addReg (PtrReg).addSImm (0).addReg (SrcReg);
819 BuildMI (BB, V8::ST, 3).addReg (PtrReg).addSImm (4).addReg (SrcReg+1);
822 BuildMI (BB, V8::STFri, 3).addReg (PtrReg).addSImm (0).addReg (SrcReg);
825 BuildMI (BB, V8::STDFri, 3).addReg (PtrReg).addSImm (0).addReg (SrcReg);
828 std::cerr << "Store instruction not handled: " << I;
834 void V8ISel::visitCallInst(CallInst &I) {
835 MachineInstr *TheCall;
836 // Is it an intrinsic function call?
837 if (Function *F = I.getCalledFunction()) {
838 if (Intrinsic::ID ID = (Intrinsic::ID)F->getIntrinsicID()) {
839 visitIntrinsicCall(ID, I); // Special intrinsics are not handled here
844 // How much extra call stack will we need?
846 for (unsigned i = 0; i < I.getNumOperands (); ++i) {
847 switch (getClassB (I.getOperand (i)->getType ())) {
848 case cLong: extraStack += 8; break;
849 case cFloat: extraStack += 4; break;
850 case cDouble: extraStack += 8; break;
851 default: extraStack += 4; break;
855 if (extraStack < 0) {
858 // Round up extra stack size to the nearest doubleword.
859 extraStack = (extraStack + 7) & ~7;
863 static const unsigned OutgoingArgRegs[] = { V8::O0, V8::O1, V8::O2, V8::O3,
865 const unsigned *OAREnd = &OutgoingArgRegs[6];
866 const unsigned *OAR = &OutgoingArgRegs[0];
867 unsigned ArgOffset = 68;
868 if (extraStack) BuildMI (BB, V8::ADJCALLSTACKDOWN, 1).addImm (extraStack);
869 for (unsigned i = 1; i < I.getNumOperands (); ++i) {
870 unsigned ArgReg = getReg (I.getOperand (i));
871 if (getClassB (I.getOperand (i)->getType ()) < cLong) {
872 // Schlep it over into the incoming arg register
873 if (ArgOffset < 92) {
874 assert (OAR != OAREnd && "About to dereference past end of OutgoingArgRegs");
875 BuildMI (BB, V8::ORrr, 2, *OAR++).addReg (V8::G0).addReg (ArgReg);
877 BuildMI (BB, V8::ST, 3).addReg (V8::SP).addSImm (ArgOffset).addReg (ArgReg);
880 } else if (getClassB (I.getOperand (i)->getType ()) == cFloat) {
881 if (ArgOffset < 92) {
882 // Single-fp args are passed in integer registers; go through
883 // memory to get them out of FP registers. (Bleh!)
884 unsigned FltAlign = TM.getTargetData().getFloatAlignment();
885 int FI = F->getFrameInfo()->CreateStackObject(4, FltAlign);
886 BuildMI (BB, V8::STFri, 3).addFrameIndex (FI).addSImm (0).addReg (ArgReg);
887 assert (OAR != OAREnd && "About to dereference past end of OutgoingArgRegs");
888 BuildMI (BB, V8::LD, 2, *OAR++).addFrameIndex (FI).addSImm (0);
890 BuildMI (BB, V8::STFri, 3).addReg (V8::SP).addSImm (ArgOffset).addReg (ArgReg);
893 } else if (getClassB (I.getOperand (i)->getType ()) == cDouble) {
894 // Double-fp args are passed in pairs of integer registers; go through
895 // memory to get them out of FP registers. (Bleh!)
896 // We'd like to 'std' these right onto the outgoing-args area, but it might
897 // not be 8-byte aligned (e.g., call x(int x, double d)). sigh.
898 unsigned DblAlign = TM.getTargetData().getDoubleAlignment();
899 int FI = F->getFrameInfo()->CreateStackObject(8, DblAlign);
900 BuildMI (BB, V8::STDFri, 3).addFrameIndex (FI).addSImm (0).addReg (ArgReg);
901 if (ArgOffset < 92 && OAR != OAREnd) {
902 assert (OAR != OAREnd && "About to dereference past end of OutgoingArgRegs");
903 BuildMI (BB, V8::LD, 2, *OAR++).addFrameIndex (FI).addSImm (0);
905 unsigned TempReg = makeAnotherReg (Type::IntTy);
906 BuildMI (BB, V8::LD, 2, TempReg).addFrameIndex (FI).addSImm (0);
907 BuildMI (BB, V8::ST, 3).addReg (V8::SP).addSImm (ArgOffset).addReg (TempReg);
910 if (ArgOffset < 92 && OAR != OAREnd) {
911 assert (OAR != OAREnd && "About to dereference past end of OutgoingArgRegs");
912 BuildMI (BB, V8::LD, 2, *OAR++).addFrameIndex (FI).addSImm (4);
914 unsigned TempReg = makeAnotherReg (Type::IntTy);
915 BuildMI (BB, V8::LD, 2, TempReg).addFrameIndex (FI).addSImm (4);
916 BuildMI (BB, V8::ST, 3).addReg (V8::SP).addSImm (ArgOffset).addReg (TempReg);
919 } else if (getClassB (I.getOperand (i)->getType ()) == cLong) {
920 // do the first half...
921 if (ArgOffset < 92) {
922 assert (OAR != OAREnd && "About to dereference past end of OutgoingArgRegs");
923 BuildMI (BB, V8::ORrr, 2, *OAR++).addReg (V8::G0).addReg (ArgReg);
925 BuildMI (BB, V8::ST, 3).addReg (V8::SP).addSImm (ArgOffset).addReg (ArgReg);
928 // ...then do the second half
929 if (ArgOffset < 92) {
930 assert (OAR != OAREnd && "About to dereference past end of OutgoingArgRegs");
931 BuildMI (BB, V8::ORrr, 2, *OAR++).addReg (V8::G0).addReg (ArgReg+1);
933 BuildMI (BB, V8::ST, 3).addReg (V8::SP).addSImm (ArgOffset).addReg (ArgReg+1);
937 assert (0 && "Unknown class?!");
941 // Emit call instruction
942 if (Function *F = I.getCalledFunction ()) {
943 BuildMI (BB, V8::CALL, 1).addGlobalAddress (F, true);
944 } else { // Emit an indirect call...
945 unsigned Reg = getReg (I.getCalledValue ());
946 BuildMI (BB, V8::JMPLrr, 3, V8::O7).addReg (Reg).addReg (V8::G0);
949 if (extraStack) BuildMI (BB, V8::ADJCALLSTACKUP, 1).addImm (extraStack);
951 // Deal w/ return value: schlep it over into the destination register
952 if (I.getType () == Type::VoidTy)
954 unsigned DestReg = getReg (I);
955 switch (getClassB (I.getType ())) {
959 BuildMI (BB, V8::ORrr, 2, DestReg).addReg(V8::G0).addReg(V8::O0);
962 BuildMI (BB, V8::FMOVS, 2, DestReg).addReg(V8::F0);
965 BuildMI (BB, V8::FpMOVD, 2, DestReg).addReg(V8::D0);
968 BuildMI (BB, V8::ORrr, 2, DestReg).addReg(V8::G0).addReg(V8::O0);
969 BuildMI (BB, V8::ORrr, 2, DestReg+1).addReg(V8::G0).addReg(V8::O1);
972 std::cerr << "Return type of call instruction not handled: " << I;
977 void V8ISel::visitReturnInst(ReturnInst &I) {
978 if (I.getNumOperands () == 1) {
979 unsigned RetValReg = getReg (I.getOperand (0));
980 switch (getClassB (I.getOperand (0)->getType ())) {
984 // Schlep it over into i0 (where it will become o0 after restore).
985 BuildMI (BB, V8::ORrr, 2, V8::I0).addReg(V8::G0).addReg(RetValReg);
988 BuildMI (BB, V8::FMOVS, 1, V8::F0).addReg(RetValReg);
991 BuildMI (BB, V8::FpMOVD, 1, V8::D0).addReg(RetValReg);
994 BuildMI (BB, V8::ORrr, 2, V8::I0).addReg(V8::G0).addReg(RetValReg);
995 BuildMI (BB, V8::ORrr, 2, V8::I1).addReg(V8::G0).addReg(RetValReg+1);
998 std::cerr << "Return instruction of this type not handled: " << I;
1003 // Just emit a 'retl' instruction to return.
1004 BuildMI(BB, V8::RETL, 0);
1008 static inline BasicBlock *getBlockAfter(BasicBlock *BB) {
1009 Function::iterator I = BB; ++I; // Get iterator to next block
1010 return I != BB->getParent()->end() ? &*I : 0;
1013 /// visitBranchInst - Handles conditional and unconditional branches.
1015 void V8ISel::visitBranchInst(BranchInst &I) {
1016 BasicBlock *takenSucc = I.getSuccessor (0);
1017 MachineBasicBlock *takenSuccMBB = MBBMap[takenSucc];
1018 BB->addSuccessor (takenSuccMBB);
1019 if (I.isConditional()) { // conditional branch
1020 BasicBlock *notTakenSucc = I.getSuccessor (1);
1021 MachineBasicBlock *notTakenSuccMBB = MBBMap[notTakenSucc];
1022 BB->addSuccessor (notTakenSuccMBB);
1024 // CondReg=(<condition>);
1025 // If (CondReg==0) goto notTakenSuccMBB;
1026 unsigned CondReg = getReg (I.getCondition ());
1027 BuildMI (BB, V8::CMPri, 2).addSImm (0).addReg (CondReg);
1028 BuildMI (BB, V8::BE, 1).addMBB (notTakenSuccMBB);
1030 // goto takenSuccMBB;
1031 BuildMI (BB, V8::BA, 1).addMBB (takenSuccMBB);
1034 /// emitGEPOperation - Common code shared between visitGetElementPtrInst and
1035 /// constant expression GEP support.
1037 void V8ISel::emitGEPOperation (MachineBasicBlock *MBB,
1038 MachineBasicBlock::iterator IP,
1039 Value *Src, User::op_iterator IdxBegin,
1040 User::op_iterator IdxEnd, unsigned TargetReg) {
1041 const TargetData &TD = TM.getTargetData ();
1042 const Type *Ty = Src->getType ();
1043 unsigned basePtrReg = getReg (Src, MBB, IP);
1045 // GEPs have zero or more indices; we must perform a struct access
1046 // or array access for each one.
1047 for (GetElementPtrInst::op_iterator oi = IdxBegin, oe = IdxEnd; oi != oe;
1050 unsigned nextBasePtrReg = makeAnotherReg (Type::UIntTy);
1051 if (const StructType *StTy = dyn_cast<StructType> (Ty)) {
1052 // It's a struct access. idx is the index into the structure,
1053 // which names the field. Use the TargetData structure to
1054 // pick out what the layout of the structure is in memory.
1055 // Use the (constant) structure index's value to find the
1056 // right byte offset from the StructLayout class's list of
1057 // structure member offsets.
1058 unsigned fieldIndex = cast<ConstantUInt> (idx)->getValue ();
1059 unsigned memberOffset =
1060 TD.getStructLayout (StTy)->MemberOffsets[fieldIndex];
1061 // Emit an ADD to add memberOffset to the basePtr.
1062 BuildMI (*MBB, IP, V8::ADDri, 2,
1063 nextBasePtrReg).addReg (basePtrReg).addZImm (memberOffset);
1064 // The next type is the member of the structure selected by the
1066 Ty = StTy->getElementType (fieldIndex);
1067 } else if (const SequentialType *SqTy = dyn_cast<SequentialType> (Ty)) {
1068 // It's an array or pointer access: [ArraySize x ElementType].
1069 // We want to add basePtrReg to (idxReg * sizeof ElementType). First, we
1070 // must find the size of the pointed-to type (Not coincidentally, the next
1071 // type is the type of the elements in the array).
1072 Ty = SqTy->getElementType ();
1073 unsigned elementSize = TD.getTypeSize (Ty);
1074 unsigned idxReg = getReg (idx, MBB, IP);
1075 unsigned OffsetReg = makeAnotherReg (Type::IntTy);
1076 unsigned elementSizeReg = makeAnotherReg (Type::UIntTy);
1077 copyConstantToRegister (MBB, IP,
1078 ConstantUInt::get(Type::UIntTy, elementSize), elementSizeReg);
1079 // Emit a SMUL to multiply the register holding the index by
1080 // elementSize, putting the result in OffsetReg.
1081 BuildMI (*MBB, IP, V8::SMULrr, 2,
1082 OffsetReg).addReg (elementSizeReg).addReg (idxReg);
1083 // Emit an ADD to add OffsetReg to the basePtr.
1084 BuildMI (*MBB, IP, V8::ADDrr, 2,
1085 nextBasePtrReg).addReg (basePtrReg).addReg (OffsetReg);
1087 basePtrReg = nextBasePtrReg;
1089 // After we have processed all the indices, the result is left in
1090 // basePtrReg. Move it to the register where we were expected to
1092 BuildMI (BB, V8::ORrr, 1, TargetReg).addReg (V8::G0).addReg (basePtrReg);
1095 void V8ISel::visitGetElementPtrInst (GetElementPtrInst &I) {
1096 unsigned outputReg = getReg (I);
1097 emitGEPOperation (BB, BB->end (), I.getOperand (0),
1098 I.op_begin ()+1, I.op_end (), outputReg);
1101 void V8ISel::emitOp64LibraryCall (MachineBasicBlock *MBB,
1102 MachineBasicBlock::iterator IP,
1104 const char *FuncName,
1105 unsigned Op0Reg, unsigned Op1Reg) {
1106 BuildMI (*MBB, IP, V8::ORrr, 2, V8::O0).addReg (V8::G0).addReg (Op0Reg);
1107 BuildMI (*MBB, IP, V8::ORrr, 2, V8::O1).addReg (V8::G0).addReg (Op0Reg+1);
1108 BuildMI (*MBB, IP, V8::ORrr, 2, V8::O2).addReg (V8::G0).addReg (Op1Reg);
1109 BuildMI (*MBB, IP, V8::ORrr, 2, V8::O3).addReg (V8::G0).addReg (Op1Reg+1);
1110 BuildMI (*MBB, IP, V8::CALL, 1).addExternalSymbol (FuncName, true);
1111 BuildMI (*MBB, IP, V8::ORrr, 2, DestReg).addReg (V8::G0).addReg (V8::O0);
1112 BuildMI (*MBB, IP, V8::ORrr, 2, DestReg+1).addReg (V8::G0).addReg (V8::O1);
1115 void V8ISel::emitShift64 (MachineBasicBlock *MBB,
1116 MachineBasicBlock::iterator IP, Instruction &I,
1117 unsigned DestReg, unsigned SrcReg,
1118 unsigned ShiftAmtReg) {
1119 bool isSigned = I.getType()->isSigned();
1121 switch (I.getOpcode ()) {
1122 case Instruction::Shl:
1123 case Instruction::Shr:
1125 unsigned CarryReg = makeAnotherReg (Type::IntTy),
1126 ThirtyTwo = makeAnotherReg (Type::IntTy),
1127 HalfShiftReg = makeAnotherReg (Type::IntTy),
1128 NegHalfShiftReg = makeAnotherReg (Type::IntTy),
1129 TempReg = makeAnotherReg (Type::IntTy);
1130 unsigned OneShiftOutReg = makeAnotherReg (Type::ULongTy),
1131 TwoShiftsOutReg = makeAnotherReg (Type::ULongTy);
1136 // Check whether the shift amount is zero:
1137 V8::G0 = V8::SUBCCrr V8::G0, ShiftAmountReg
1138 V8::BE .lshr_continue
1141 .lshr_shift: // [preds: begin]
1142 // Calculate 32 - shamt:
1143 ThirtyTwo = V8::ORri V8::G0, 32
1144 HalfShiftReg = V8::SUBCCrr ThirtyTwo, ShiftAmountReg
1145 // See whether it was greater than 0:
1146 V8::BG .lshr_two_shifts
1147 V8::BA .lshr_one_shift
1149 .lshr_two_shifts: // [preds: shift]
1150 CarryReg = V8::SLLrr SrcReg, HalfShiftReg
1151 TwoShiftsOutReg = V8::SRLrr SrcReg, ShiftAmountReg
1152 TempReg = V8::SRLrr SrcReg+1, ShiftAmountReg
1153 TwoShiftsOutReg+1 = V8::ORrr TempReg, CarryReg
1154 V8::BA .lshr_continue
1156 .lshr_one_shift: // [preds: shift]
1157 OneShiftOutReg = V8::ORrr V8::G0, V8::G0
1158 NegHalfShiftReg = V8::SUBrr V8::G0, HalfShiftReg
1159 OneShiftOutReg+1 = V8::SRLrr SrcReg, NegHalfShiftReg
1160 V8::BA .lshr_continue
1162 .lshr_continue: // [preds: begin, do_one_shift, do_two_shifts]
1163 DestReg = V8::PHI (SrcReg, begin), (TwoShiftsOutReg, two_shifts),
1164 (OneShiftOutReg, one_shift)
1165 DestReg+1 = V8::PHI (SrcReg+1, begin), (TwoShiftsOutReg+1, two_shifts),
1166 (OneShiftOutReg+1, one_shift)
1170 std::cerr << "Sorry, 64-bit lshr is not yet supported:\n" << I;
1175 std::cerr << "Sorry, 64-bit shifts are not yet supported:\n" << I;
1180 void V8ISel::visitBinaryOperator (Instruction &I) {
1181 unsigned DestReg = getReg (I);
1182 unsigned Op0Reg = getReg (I.getOperand (0));
1183 unsigned Op1Reg = getReg (I.getOperand (1));
1185 unsigned Class = getClassB (I.getType());
1186 unsigned OpCase = ~0;
1188 if (Class > cLong) {
1189 switch (I.getOpcode ()) {
1190 case Instruction::Add: OpCase = 0; break;
1191 case Instruction::Sub: OpCase = 1; break;
1192 case Instruction::Mul: OpCase = 2; break;
1193 case Instruction::Div: OpCase = 3; break;
1194 default: visitInstruction (I); return;
1196 static unsigned Opcodes[] = { V8::FADDS, V8::FADDD,
1197 V8::FSUBS, V8::FSUBD,
1198 V8::FMULS, V8::FMULD,
1199 V8::FDIVS, V8::FDIVD };
1200 BuildMI (BB, Opcodes[2*OpCase + (Class - cFloat)], 2, DestReg)
1201 .addReg (Op0Reg).addReg (Op1Reg);
1205 unsigned ResultReg = DestReg;
1206 if (Class != cInt && Class != cLong)
1207 ResultReg = makeAnotherReg (I.getType ());
1209 if (Class == cLong) {
1210 const char *FuncName;
1211 DEBUG (std::cerr << "Class = cLong\n");
1212 DEBUG (std::cerr << "Op0Reg = " << Op0Reg << ", " << Op0Reg+1 << "\n");
1213 DEBUG (std::cerr << "Op1Reg = " << Op1Reg << ", " << Op1Reg+1 << "\n");
1214 DEBUG (std::cerr << "ResultReg = " << ResultReg << ", " << ResultReg+1 << "\n");
1215 DEBUG (std::cerr << "DestReg = " << DestReg << ", " << DestReg+1 << "\n");
1216 switch (I.getOpcode ()) {
1217 case Instruction::Add:
1218 BuildMI (BB, V8::ADDCCrr, 2, ResultReg+1).addReg (Op0Reg+1)
1220 BuildMI (BB, V8::ADDXrr, 2, ResultReg).addReg (Op0Reg).addReg (Op1Reg);
1222 case Instruction::Sub:
1223 BuildMI (BB, V8::SUBCCrr, 2, ResultReg+1).addReg (Op0Reg+1)
1225 BuildMI (BB, V8::SUBXrr, 2, ResultReg).addReg (Op0Reg).addReg (Op1Reg);
1227 case Instruction::Mul:
1228 FuncName = I.getType ()->isSigned () ? "__mul64" : "__umul64";
1229 emitOp64LibraryCall (BB, BB->end (), DestReg, FuncName, Op0Reg, Op1Reg);
1231 case Instruction::Div:
1232 FuncName = I.getType ()->isSigned () ? "__div64" : "__udiv64";
1233 emitOp64LibraryCall (BB, BB->end (), DestReg, FuncName, Op0Reg, Op1Reg);
1235 case Instruction::Rem:
1236 FuncName = I.getType ()->isSigned () ? "__rem64" : "__urem64";
1237 emitOp64LibraryCall (BB, BB->end (), DestReg, FuncName, Op0Reg, Op1Reg);
1239 case Instruction::Shl:
1240 case Instruction::Shr:
1241 emitShift64 (BB, BB->end (), I, DestReg, Op0Reg, Op1Reg);
1246 // FIXME: support long, ulong.
1247 switch (I.getOpcode ()) {
1248 case Instruction::Add: OpCase = 0; break;
1249 case Instruction::Sub: OpCase = 1; break;
1250 case Instruction::Mul: OpCase = 2; break;
1251 case Instruction::And: OpCase = 3; break;
1252 case Instruction::Or: OpCase = 4; break;
1253 case Instruction::Xor: OpCase = 5; break;
1254 case Instruction::Shl: OpCase = 6; break;
1255 case Instruction::Shr: OpCase = 7+I.getType()->isSigned(); break;
1257 case Instruction::Div:
1258 case Instruction::Rem: {
1259 unsigned Dest = ResultReg;
1260 if (I.getOpcode() == Instruction::Rem)
1261 Dest = makeAnotherReg(I.getType());
1263 // FIXME: this is probably only right for 32 bit operands.
1264 if (I.getType ()->isSigned()) {
1265 unsigned Tmp = makeAnotherReg (I.getType ());
1266 // Sign extend into the Y register
1267 BuildMI (BB, V8::SRAri, 2, Tmp).addReg (Op0Reg).addZImm (31);
1268 BuildMI (BB, V8::WRrr, 2, V8::Y).addReg (Tmp).addReg (V8::G0);
1269 BuildMI (BB, V8::SDIVrr, 2, Dest).addReg (Op0Reg).addReg (Op1Reg);
1271 // Zero extend into the Y register, ie, just set it to zero
1272 BuildMI (BB, V8::WRrr, 2, V8::Y).addReg (V8::G0).addReg (V8::G0);
1273 BuildMI (BB, V8::UDIVrr, 2, Dest).addReg (Op0Reg).addReg (Op1Reg);
1276 if (I.getOpcode() == Instruction::Rem) {
1277 unsigned Tmp = makeAnotherReg (I.getType ());
1278 BuildMI (BB, V8::SMULrr, 2, Tmp).addReg(Dest).addReg(Op1Reg);
1279 BuildMI (BB, V8::SUBrr, 2, ResultReg).addReg(Op0Reg).addReg(Tmp);
1284 visitInstruction (I);
1288 static const unsigned Opcodes[] = {
1289 V8::ADDrr, V8::SUBrr, V8::SMULrr, V8::ANDrr, V8::ORrr, V8::XORrr,
1290 V8::SLLrr, V8::SRLrr, V8::SRArr
1292 if (OpCase != ~0U) {
1293 BuildMI (BB, Opcodes[OpCase], 2, ResultReg).addReg (Op0Reg).addReg (Op1Reg);
1296 switch (getClassB (I.getType ())) {
1298 if (I.getType ()->isSigned ()) { // add byte
1299 BuildMI (BB, V8::ANDri, 2, DestReg).addReg (ResultReg).addZImm (0xff);
1300 } else { // add ubyte
1301 unsigned TmpReg = makeAnotherReg (I.getType ());
1302 BuildMI (BB, V8::SLLri, 2, TmpReg).addReg (ResultReg).addZImm (24);
1303 BuildMI (BB, V8::SRAri, 2, DestReg).addReg (TmpReg).addZImm (24);
1307 if (I.getType ()->isSigned ()) { // add short
1308 unsigned TmpReg = makeAnotherReg (I.getType ());
1309 BuildMI (BB, V8::SLLri, 2, TmpReg).addReg (ResultReg).addZImm (16);
1310 BuildMI (BB, V8::SRAri, 2, DestReg).addReg (TmpReg).addZImm (16);
1311 } else { // add ushort
1312 unsigned TmpReg = makeAnotherReg (I.getType ());
1313 BuildMI (BB, V8::SLLri, 2, TmpReg).addReg (ResultReg).addZImm (16);
1314 BuildMI (BB, V8::SRLri, 2, DestReg).addReg (TmpReg).addZImm (16);
1318 // Nothing to do here.
1321 // Only support and, or, xor here - others taken care of above.
1322 if (OpCase < 3 || OpCase > 5) {
1323 visitInstruction (I);
1326 // Do the other half of the value:
1327 BuildMI (BB, Opcodes[OpCase], 2, ResultReg+1).addReg (Op0Reg+1)
1331 visitInstruction (I);
1335 void V8ISel::visitSetCondInst(SetCondInst &I) {
1336 unsigned Op0Reg = getReg (I.getOperand (0));
1337 unsigned Op1Reg = getReg (I.getOperand (1));
1338 unsigned DestReg = getReg (I);
1339 const Type *Ty = I.getOperand (0)->getType ();
1341 // Compare the two values.
1342 if (getClass (Ty) < cLong) {
1343 BuildMI(BB, V8::SUBCCrr, 2, V8::G0).addReg(Op0Reg).addReg(Op1Reg);
1344 } else if (getClass (Ty) == cLong) {
1345 switch (I.getOpcode()) {
1346 default: assert(0 && "Unknown setcc instruction!");
1347 case Instruction::SetEQ:
1348 case Instruction::SetNE: {
1349 unsigned TempReg0 = makeAnotherReg (Type::IntTy),
1350 TempReg1 = makeAnotherReg (Type::IntTy),
1351 TempReg2 = makeAnotherReg (Type::IntTy),
1352 TempReg3 = makeAnotherReg (Type::IntTy);
1353 MachineOpCode Opcode;
1355 // These guys are special - no branches needed!
1356 BuildMI (BB, V8::XORrr, 2, TempReg0).addReg (Op0Reg+1).addReg (Op1Reg+1);
1357 BuildMI (BB, V8::XORrr, 2, TempReg1).addReg (Op0Reg).addReg (Op1Reg);
1358 BuildMI (BB, V8::SUBCCrr, 2, V8::G0).addReg (V8::G0).addReg (TempReg1);
1359 Opcode = I.getOpcode() == Instruction::SetEQ ? V8::SUBXri : V8::ADDXri;
1360 Immed = I.getOpcode() == Instruction::SetEQ ? -1 : 0;
1361 BuildMI (BB, Opcode, 2, TempReg2).addReg (V8::G0).addSImm (Immed);
1362 BuildMI (BB, V8::SUBCCrr, 2, V8::G0).addReg (V8::G0).addReg (TempReg0);
1363 BuildMI (BB, Opcode, 2, TempReg3).addReg (V8::G0).addSImm (Immed);
1364 Opcode = I.getOpcode() == Instruction::SetEQ ? V8::ANDrr : V8::ORrr;
1365 BuildMI (BB, Opcode, 2, DestReg).addReg (TempReg2).addReg (TempReg3);
1368 case Instruction::SetLT:
1369 case Instruction::SetGE:
1370 BuildMI (BB, V8::SUBCCrr, 2, V8::G0).addReg (Op0Reg+1).addReg (Op1Reg+1);
1371 BuildMI (BB, V8::SUBXCCrr, 2, V8::G0).addReg (Op0Reg).addReg (Op1Reg);
1373 case Instruction::SetGT:
1374 case Instruction::SetLE:
1375 BuildMI (BB, V8::SUBCCri, 2, V8::G0).addReg (V8::G0).addSImm (1);
1376 BuildMI (BB, V8::SUBXCCrr, 2, V8::G0).addReg (Op0Reg+1).addReg (Op1Reg+1);
1377 BuildMI (BB, V8::SUBXCCrr, 2, V8::G0).addReg (Op0Reg).addReg (Op1Reg);
1380 } else if (getClass (Ty) == cFloat) {
1381 BuildMI(BB, V8::FCMPS, 2).addReg(Op0Reg).addReg(Op1Reg);
1382 } else if (getClass (Ty) == cDouble) {
1383 BuildMI(BB, V8::FCMPD, 2).addReg(Op0Reg).addReg(Op1Reg);
1387 switch (I.getOpcode()) {
1388 default: assert(0 && "Unknown setcc instruction!");
1389 case Instruction::SetEQ: BranchIdx = 0; break;
1390 case Instruction::SetNE: BranchIdx = 1; break;
1391 case Instruction::SetLT: BranchIdx = 2; break;
1392 case Instruction::SetGT: BranchIdx = 3; break;
1393 case Instruction::SetLE: BranchIdx = 4; break;
1394 case Instruction::SetGE: BranchIdx = 5; break;
1397 unsigned Column = 0;
1398 if (Ty->isSigned() && !Ty->isFloatingPoint()) Column = 1;
1399 if (Ty->isFloatingPoint()) Column = 2;
1400 static unsigned OpcodeTab[3*6] = {
1402 // unsigned signed fp
1403 V8::BE, V8::BE, V8::FBE, // seteq = be be fbe
1404 V8::BNE, V8::BNE, V8::FBNE, // setne = bne bne fbne
1405 V8::BCS, V8::BL, V8::FBL, // setlt = bcs bl fbl
1406 V8::BGU, V8::BG, V8::FBG, // setgt = bgu bg fbg
1407 V8::BLEU, V8::BLE, V8::FBLE, // setle = bleu ble fble
1408 V8::BCC, V8::BGE, V8::FBGE // setge = bcc bge fbge
1410 unsigned Opcode = OpcodeTab[3*BranchIdx + Column];
1412 MachineBasicBlock *thisMBB = BB;
1413 const BasicBlock *LLVM_BB = BB->getBasicBlock ();
1416 // subcc %reg0, %reg1, %g0
1420 // FIXME: we wouldn't need copy0MBB (we could fold it into thisMBB)
1421 // if we could insert other, non-terminator instructions after the
1422 // bCC. But MBB->getFirstTerminator() can't understand this.
1423 MachineBasicBlock *copy1MBB = new MachineBasicBlock (LLVM_BB);
1424 F->getBasicBlockList ().push_back (copy1MBB);
1425 BuildMI (BB, Opcode, 1).addMBB (copy1MBB);
1426 MachineBasicBlock *copy0MBB = new MachineBasicBlock (LLVM_BB);
1427 F->getBasicBlockList ().push_back (copy0MBB);
1428 BuildMI (BB, V8::BA, 1).addMBB (copy0MBB);
1429 // Update machine-CFG edges
1430 BB->addSuccessor (copy1MBB);
1431 BB->addSuccessor (copy0MBB);
1434 // %FalseValue = or %G0, 0
1437 unsigned FalseValue = makeAnotherReg (I.getType ());
1438 BuildMI (BB, V8::ORri, 2, FalseValue).addReg (V8::G0).addZImm (0);
1439 MachineBasicBlock *sinkMBB = new MachineBasicBlock (LLVM_BB);
1440 F->getBasicBlockList ().push_back (sinkMBB);
1441 BuildMI (BB, V8::BA, 1).addMBB (sinkMBB);
1442 // Update machine-CFG edges
1443 BB->addSuccessor (sinkMBB);
1445 DEBUG (std::cerr << "thisMBB is at " << (void*)thisMBB << "\n");
1446 DEBUG (std::cerr << "copy1MBB is at " << (void*)copy1MBB << "\n");
1447 DEBUG (std::cerr << "copy0MBB is at " << (void*)copy0MBB << "\n");
1448 DEBUG (std::cerr << "sinkMBB is at " << (void*)sinkMBB << "\n");
1451 // %TrueValue = or %G0, 1
1454 unsigned TrueValue = makeAnotherReg (I.getType ());
1455 BuildMI (BB, V8::ORri, 2, TrueValue).addReg (V8::G0).addZImm (1);
1456 BuildMI (BB, V8::BA, 1).addMBB (sinkMBB);
1457 // Update machine-CFG edges
1458 BB->addSuccessor (sinkMBB);
1461 // %Result = phi [ %FalseValue, copy0MBB ], [ %TrueValue, copy1MBB ]
1464 BuildMI (BB, V8::PHI, 4, DestReg).addReg (FalseValue)
1465 .addMBB (copy0MBB).addReg (TrueValue).addMBB (copy1MBB);
1468 void V8ISel::visitAllocaInst(AllocaInst &I) {
1469 // Find the data size of the alloca inst's getAllocatedType.
1470 const Type *Ty = I.getAllocatedType();
1471 unsigned TySize = TM.getTargetData().getTypeSize(Ty);
1473 unsigned ArraySizeReg = getReg (I.getArraySize ());
1474 unsigned TySizeReg = getReg (ConstantUInt::get (Type::UIntTy, TySize));
1475 unsigned TmpReg1 = makeAnotherReg (Type::UIntTy);
1476 unsigned TmpReg2 = makeAnotherReg (Type::UIntTy);
1477 unsigned StackAdjReg = makeAnotherReg (Type::UIntTy);
1479 // StackAdjReg = (ArraySize * TySize) rounded up to nearest
1480 // doubleword boundary.
1481 BuildMI (BB, V8::UMULrr, 2, TmpReg1).addReg (ArraySizeReg).addReg (TySizeReg);
1483 // Round up TmpReg1 to nearest doubleword boundary:
1484 BuildMI (BB, V8::ADDri, 2, TmpReg2).addReg (TmpReg1).addSImm (7);
1485 BuildMI (BB, V8::ANDri, 2, StackAdjReg).addReg (TmpReg2).addSImm (-8);
1487 // Subtract size from stack pointer, thereby allocating some space.
1488 BuildMI (BB, V8::SUBrr, 2, V8::SP).addReg (V8::SP).addReg (StackAdjReg);
1490 // Put a pointer to the space into the result register, by copying
1491 // the stack pointer.
1492 BuildMI (BB, V8::ADDri, 2, getReg(I)).addReg (V8::SP).addSImm (96);
1494 // Inform the Frame Information that we have just allocated a variable-sized
1496 F->getFrameInfo()->CreateVariableSizedObject();
1499 /// LowerUnknownIntrinsicFunctionCalls - This performs a prepass over the
1500 /// function, lowering any calls to unknown intrinsic functions into the
1501 /// equivalent LLVM code.
1502 void V8ISel::LowerUnknownIntrinsicFunctionCalls(Function &F) {
1503 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
1504 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
1505 if (CallInst *CI = dyn_cast<CallInst>(I++))
1506 if (Function *F = CI->getCalledFunction())
1507 switch (F->getIntrinsicID()) {
1508 case Intrinsic::vastart:
1509 case Intrinsic::vacopy:
1510 case Intrinsic::vaend:
1511 // We directly implement these intrinsics
1512 case Intrinsic::not_intrinsic: break;
1514 // All other intrinsic calls we must lower.
1515 Instruction *Before = CI->getPrev();
1516 TM.getIntrinsicLowering().LowerIntrinsicCall(CI);
1517 if (Before) { // Move iterator to instruction after call
1526 void V8ISel::visitIntrinsicCall(Intrinsic::ID ID, CallInst &CI) {
1529 std::cerr << "Sorry, unknown intrinsic function call:\n" << CI; abort ();
1531 case Intrinsic::vastart: {
1532 // Add the VarArgsOffset to the frame pointer, and copy it to the result.
1533 unsigned DestReg = getReg (CI);
1534 BuildMI (BB, V8::ADDri, 2, DestReg).addReg (V8::FP).addSImm (VarArgsOffset);
1538 case Intrinsic::vaend:
1539 // va_end is a no-op on SparcV8.
1542 case Intrinsic::vacopy: {
1543 // Copy the va_list ptr (arg1) to the result.
1544 unsigned DestReg = getReg (CI), SrcReg = getReg (CI.getOperand (1));
1545 BuildMI (BB, V8::ORrr, 2, DestReg).addReg (V8::G0).addReg (SrcReg);
1551 void V8ISel::visitVANextInst (VANextInst &I) {
1552 // Add the type size to the vararg pointer (arg0).
1553 unsigned DestReg = getReg (I);
1554 unsigned SrcReg = getReg (I.getOperand (0));
1555 unsigned TySize = TM.getTargetData ().getTypeSize (I.getArgType ());
1556 BuildMI (BB, V8::ADDri, 2, DestReg).addReg (SrcReg).addSImm (TySize);
1559 void V8ISel::visitVAArgInst (VAArgInst &I) {
1560 unsigned VAList = getReg (I.getOperand (0));
1561 unsigned DestReg = getReg (I);
1563 switch (I.getType ()->getTypeID ()) {
1564 case Type::PointerTyID:
1565 case Type::UIntTyID:
1567 BuildMI (BB, V8::LD, 2, DestReg).addReg (VAList).addSImm (0);
1570 case Type::ULongTyID:
1571 case Type::LongTyID:
1572 BuildMI (BB, V8::LD, 2, DestReg).addReg (VAList).addSImm (0);
1573 BuildMI (BB, V8::LD, 2, DestReg+1).addReg (VAList).addSImm (4);
1576 case Type::DoubleTyID: {
1577 unsigned DblAlign = TM.getTargetData().getDoubleAlignment();
1578 unsigned TempReg = makeAnotherReg (Type::IntTy);
1579 unsigned TempReg2 = makeAnotherReg (Type::IntTy);
1580 int FI = F->getFrameInfo()->CreateStackObject(8, DblAlign);
1581 BuildMI (BB, V8::LD, 2, TempReg).addReg (VAList).addSImm (0);
1582 BuildMI (BB, V8::LD, 2, TempReg2).addReg (VAList).addSImm (4);
1583 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (0).addReg (TempReg);
1584 BuildMI (BB, V8::ST, 3).addFrameIndex (FI).addSImm (4).addReg (TempReg2);
1585 BuildMI (BB, V8::LDDFri, 2, DestReg).addFrameIndex (FI).addSImm (0);
1590 std::cerr << "Sorry, vaarg instruction of this type still unsupported:\n"