1 //===-- SparcInstrSelection.cpp -------------------------------------------===//
3 // BURS instruction selection for SPARC V9 architecture.
5 //===----------------------------------------------------------------------===//
7 #include "SparcInternals.h"
8 #include "SparcInstrSelectionSupport.h"
9 #include "SparcRegClassInfo.h"
10 #include "llvm/CodeGen/InstrSelectionSupport.h"
11 #include "llvm/CodeGen/MachineInstrBuilder.h"
12 #include "llvm/CodeGen/MachineInstrAnnot.h"
13 #include "llvm/CodeGen/InstrForest.h"
14 #include "llvm/CodeGen/InstrSelection.h"
15 #include "llvm/CodeGen/MachineFunction.h"
16 #include "llvm/CodeGen/MachineFunctionInfo.h"
17 #include "llvm/CodeGen/MachineCodeForInstruction.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/iTerminators.h"
20 #include "llvm/iMemory.h"
21 #include "llvm/iOther.h"
22 #include "llvm/Function.h"
23 #include "llvm/Constants.h"
24 #include "llvm/ConstantHandling.h"
25 #include "llvm/Intrinsics.h"
26 #include "Support/MathExtras.h"
30 static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
31 std::vector<MachineInstr*>& mvec) {
32 mvec.push_back(BuildMI(Opcode, 3).addReg(Node->leftChild()->getValue())
33 .addReg(Node->rightChild()->getValue())
34 .addRegDef(Node->getValue()));
39 //---------------------------------------------------------------------------
40 // Function: GetMemInstArgs
43 // Get the pointer value and the index vector for a memory operation
44 // (GetElementPtr, Load, or Store). If all indices of the given memory
45 // operation are constant, fold in constant indices in a chain of
46 // preceding GetElementPtr instructions (if any), and return the
47 // pointer value of the first instruction in the chain.
48 // All folded instructions are marked so no code is generated for them.
51 // Returns the pointer Value to use.
52 // Returns the resulting IndexVector in idxVec.
53 // Returns true/false in allConstantIndices if all indices are/aren't const.
54 //---------------------------------------------------------------------------
57 //---------------------------------------------------------------------------
58 // Function: FoldGetElemChain
61 // Fold a chain of GetElementPtr instructions containing only
62 // constant offsets into an equivalent (Pointer, IndexVector) pair.
63 // Returns the pointer Value, and stores the resulting IndexVector
64 // in argument chainIdxVec. This is a helper function for
65 // FoldConstantIndices that does the actual folding.
66 //---------------------------------------------------------------------------
69 // Check for a constant 0.
73 return (idx == ConstantSInt::getNullValue(idx->getType()));
77 FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
78 bool lastInstHasLeadingNonZero)
80 InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
81 GetElementPtrInst* gepInst =
82 dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
84 // ptr value is not computed in this tree or ptr value does not come from GEP
89 // Return NULL if we don't fold any instructions in.
92 // Now chase the chain of getElementInstr instructions, if any.
93 // Check for any non-constant indices and stop there.
94 // Also, stop if the first index of child is a non-zero array index
95 // and the last index of the current node is a non-array index:
96 // in that case, a non-array declared type is being accessed as an array
97 // which is not type-safe, but could be legal.
99 InstructionNode* ptrChild = gepNode;
100 while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
101 ptrChild->getOpLabel() == GetElemPtrIdx))
103 // Child is a GetElemPtr instruction
104 gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
105 User::op_iterator OI, firstIdx = gepInst->idx_begin();
106 User::op_iterator lastIdx = gepInst->idx_end();
107 bool allConstantOffsets = true;
109 // The first index of every GEP must be an array index.
110 assert((*firstIdx)->getType() == Type::LongTy &&
111 "INTERNAL ERROR: Structure index for a pointer type!");
113 // If the last instruction had a leading non-zero index, check if the
114 // current one references a sequential (i.e., indexable) type.
115 // If not, the code is not type-safe and we would create an illegal GEP
116 // by folding them, so don't fold any more instructions.
118 if (lastInstHasLeadingNonZero)
119 if (! isa<SequentialType>(gepInst->getType()->getElementType()))
120 break; // cannot fold in any preceding getElementPtr instrs.
122 // Check that all offsets are constant for this instruction
123 for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
124 allConstantOffsets = isa<ConstantInt>(*OI);
126 if (allConstantOffsets) {
127 // Get pointer value out of ptrChild.
128 ptrVal = gepInst->getPointerOperand();
130 // Insert its index vector at the start, skipping any leading [0]
131 // Remember the old size to check if anything was inserted.
132 unsigned oldSize = chainIdxVec.size();
133 int firstIsZero = IsZero(*firstIdx);
134 chainIdxVec.insert(chainIdxVec.begin(), firstIdx + firstIsZero, lastIdx);
136 // Remember if it has leading zero index: it will be discarded later.
137 if (oldSize < chainIdxVec.size())
138 lastInstHasLeadingNonZero = !firstIsZero;
140 // Mark the folded node so no code is generated for it.
141 ((InstructionNode*) ptrChild)->markFoldedIntoParent();
143 // Get the previous GEP instruction and continue trying to fold
144 ptrChild = dyn_cast<InstructionNode>(ptrChild->leftChild());
145 } else // cannot fold this getElementPtr instr. or any preceding ones
149 // If the first getElementPtr instruction had a leading [0], add it back.
150 // Note that this instruction is the *last* one that was successfully
151 // folded *and* contributed any indices, in the loop above.
153 if (ptrVal && ! lastInstHasLeadingNonZero)
154 chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
160 //---------------------------------------------------------------------------
161 // Function: GetGEPInstArgs
164 // Helper function for GetMemInstArgs that handles the final getElementPtr
165 // instruction used by (or same as) the memory operation.
166 // Extracts the indices of the current instruction and tries to fold in
167 // preceding ones if all indices of the current one are constant.
168 //---------------------------------------------------------------------------
171 GetGEPInstArgs(InstructionNode* gepNode,
172 std::vector<Value*>& idxVec,
173 bool& allConstantIndices)
175 allConstantIndices = true;
176 GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
178 // Default pointer is the one from the current instruction.
179 Value* ptrVal = gepI->getPointerOperand();
180 InstrTreeNode* ptrChild = gepNode->leftChild();
182 // Extract the index vector of the GEP instructin.
183 // If all indices are constant and first index is zero, try to fold
184 // in preceding GEPs with all constant indices.
185 for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
186 allConstantIndices && OI != OE; ++OI)
187 if (! isa<Constant>(*OI))
188 allConstantIndices = false; // note: this also terminates loop!
190 // If we have only constant indices, fold chains of constant indices
191 // in this and any preceding GetElemPtr instructions.
192 bool foldedGEPs = false;
193 bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
194 if (allConstantIndices)
195 if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx)) {
200 // Append the index vector of the current instruction.
201 // Skip the leading [0] index if preceding GEPs were folded into this.
202 idxVec.insert(idxVec.end(),
203 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
209 //---------------------------------------------------------------------------
210 // Function: GetMemInstArgs
213 // Get the pointer value and the index vector for a memory operation
214 // (GetElementPtr, Load, or Store). If all indices of the given memory
215 // operation are constant, fold in constant indices in a chain of
216 // preceding GetElementPtr instructions (if any), and return the
217 // pointer value of the first instruction in the chain.
218 // All folded instructions are marked so no code is generated for them.
221 // Returns the pointer Value to use.
222 // Returns the resulting IndexVector in idxVec.
223 // Returns true/false in allConstantIndices if all indices are/aren't const.
224 //---------------------------------------------------------------------------
227 GetMemInstArgs(InstructionNode* memInstrNode,
228 std::vector<Value*>& idxVec,
229 bool& allConstantIndices)
231 allConstantIndices = false;
232 Instruction* memInst = memInstrNode->getInstruction();
233 assert(idxVec.size() == 0 && "Need empty vector to return indices");
235 // If there is a GetElemPtr instruction to fold in to this instr,
236 // it must be in the left child for Load and GetElemPtr, and in the
237 // right child for Store instructions.
238 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
239 ? memInstrNode->rightChild()
240 : memInstrNode->leftChild());
242 // Default pointer is the one from the current instruction.
243 Value* ptrVal = ptrChild->getValue();
245 // Find the "last" GetElemPtr instruction: this one or the immediate child.
246 // There will be none if this is a load or a store from a scalar pointer.
247 InstructionNode* gepNode = NULL;
248 if (isa<GetElementPtrInst>(memInst))
249 gepNode = memInstrNode;
250 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
251 // Child of load/store is a GEP and memInst is its only use.
252 // Use its indices and mark it as folded.
253 gepNode = cast<InstructionNode>(ptrChild);
254 gepNode->markFoldedIntoParent();
257 // If there are no indices, return the current pointer.
258 // Else extract the pointer from the GEP and fold the indices.
259 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
264 //************************ Internal Functions ******************************/
267 static inline MachineOpCode
268 ChooseBprInstruction(const InstructionNode* instrNode)
270 MachineOpCode opCode;
272 Instruction* setCCInstr =
273 ((InstructionNode*) instrNode->leftChild())->getInstruction();
275 switch(setCCInstr->getOpcode())
277 case Instruction::SetEQ: opCode = V9::BRZ; break;
278 case Instruction::SetNE: opCode = V9::BRNZ; break;
279 case Instruction::SetLE: opCode = V9::BRLEZ; break;
280 case Instruction::SetGE: opCode = V9::BRGEZ; break;
281 case Instruction::SetLT: opCode = V9::BRLZ; break;
282 case Instruction::SetGT: opCode = V9::BRGZ; break;
284 assert(0 && "Unrecognized VM instruction!");
285 opCode = V9::INVALID_OPCODE;
293 static inline MachineOpCode
294 ChooseBpccInstruction(const InstructionNode* instrNode,
295 const BinaryOperator* setCCInstr)
297 MachineOpCode opCode = V9::INVALID_OPCODE;
299 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
302 switch(setCCInstr->getOpcode())
304 case Instruction::SetEQ: opCode = V9::BE; break;
305 case Instruction::SetNE: opCode = V9::BNE; break;
306 case Instruction::SetLE: opCode = V9::BLE; break;
307 case Instruction::SetGE: opCode = V9::BGE; break;
308 case Instruction::SetLT: opCode = V9::BL; break;
309 case Instruction::SetGT: opCode = V9::BG; break;
311 assert(0 && "Unrecognized VM instruction!");
315 switch(setCCInstr->getOpcode())
317 case Instruction::SetEQ: opCode = V9::BE; break;
318 case Instruction::SetNE: opCode = V9::BNE; break;
319 case Instruction::SetLE: opCode = V9::BLEU; break;
320 case Instruction::SetGE: opCode = V9::BCC; break;
321 case Instruction::SetLT: opCode = V9::BCS; break;
322 case Instruction::SetGT: opCode = V9::BGU; break;
324 assert(0 && "Unrecognized VM instruction!");
332 static inline MachineOpCode
333 ChooseBFpccInstruction(const InstructionNode* instrNode,
334 const BinaryOperator* setCCInstr)
336 MachineOpCode opCode = V9::INVALID_OPCODE;
338 switch(setCCInstr->getOpcode())
340 case Instruction::SetEQ: opCode = V9::FBE; break;
341 case Instruction::SetNE: opCode = V9::FBNE; break;
342 case Instruction::SetLE: opCode = V9::FBLE; break;
343 case Instruction::SetGE: opCode = V9::FBGE; break;
344 case Instruction::SetLT: opCode = V9::FBL; break;
345 case Instruction::SetGT: opCode = V9::FBG; break;
347 assert(0 && "Unrecognized VM instruction!");
355 // Create a unique TmpInstruction for a boolean value,
356 // representing the CC register used by a branch on that value.
357 // For now, hack this using a little static cache of TmpInstructions.
358 // Eventually the entire BURG instruction selection should be put
359 // into a separate class that can hold such information.
360 // The static cache is not too bad because the memory for these
361 // TmpInstructions will be freed along with the rest of the Function anyway.
363 static TmpInstruction*
364 GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType,
365 MachineCodeForInstruction& mcfi)
367 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
368 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
369 static const Function *lastFunction = 0;// Use to flush cache between funcs
371 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
373 if (lastFunction != F) {
375 boolToTmpCache.clear();
378 // Look for tmpI and create a new one otherwise. The new value is
379 // directly written to map using the ref returned by operator[].
380 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
382 tmpI = new TmpInstruction(mcfi, ccType, boolVal);
388 static inline MachineOpCode
389 ChooseBccInstruction(const InstructionNode* instrNode,
390 const Type*& setCCType)
392 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
393 assert(setCCNode->getOpLabel() == SetCCOp);
394 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
395 setCCType = setCCInstr->getOperand(0)->getType();
397 if (setCCType->isFloatingPoint())
398 return ChooseBFpccInstruction(instrNode, setCCInstr);
400 return ChooseBpccInstruction(instrNode, setCCInstr);
404 // WARNING: since this function has only one caller, it always returns
405 // the opcode that expects an immediate and a register. If this function
406 // is ever used in cases where an opcode that takes two registers is required,
407 // then modify this function and use convertOpcodeFromRegToImm() where required.
409 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
410 // new cases of opcodes.
411 static inline MachineOpCode
412 ChooseMovFpcciInstruction(const InstructionNode* instrNode)
414 MachineOpCode opCode = V9::INVALID_OPCODE;
416 switch(instrNode->getInstruction()->getOpcode())
418 case Instruction::SetEQ: opCode = V9::MOVFEi; break;
419 case Instruction::SetNE: opCode = V9::MOVFNEi; break;
420 case Instruction::SetLE: opCode = V9::MOVFLEi; break;
421 case Instruction::SetGE: opCode = V9::MOVFGEi; break;
422 case Instruction::SetLT: opCode = V9::MOVFLi; break;
423 case Instruction::SetGT: opCode = V9::MOVFGi; break;
425 assert(0 && "Unrecognized VM instruction!");
433 // ChooseMovpcciForSetCC -- Choose a conditional-move instruction
434 // based on the type of SetCC operation.
436 // WARNING: since this function has only one caller, it always returns
437 // the opcode that expects an immediate and a register. If this function
438 // is ever used in cases where an opcode that takes two registers is required,
439 // then modify this function and use convertOpcodeFromRegToImm() where required.
441 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
442 // new cases of opcodes.
445 ChooseMovpcciForSetCC(const InstructionNode* instrNode)
447 MachineOpCode opCode = V9::INVALID_OPCODE;
449 const Type* opType = instrNode->leftChild()->getValue()->getType();
450 assert(opType->isIntegral() || isa<PointerType>(opType));
451 bool noSign = opType->isUnsigned() || isa<PointerType>(opType);
453 switch(instrNode->getInstruction()->getOpcode())
455 case Instruction::SetEQ: opCode = V9::MOVEi; break;
456 case Instruction::SetLE: opCode = noSign? V9::MOVLEUi : V9::MOVLEi; break;
457 case Instruction::SetGE: opCode = noSign? V9::MOVCCi : V9::MOVGEi; break;
458 case Instruction::SetLT: opCode = noSign? V9::MOVCSi : V9::MOVLi; break;
459 case Instruction::SetGT: opCode = noSign? V9::MOVGUi : V9::MOVGi; break;
460 case Instruction::SetNE: opCode = V9::MOVNEi; break;
461 default: assert(0 && "Unrecognized LLVM instr!"); break;
468 // ChooseMovpregiForSetCC -- Choose a conditional-move-on-register-value
469 // instruction based on the type of SetCC operation. These instructions
470 // compare a register with 0 and perform the move is the comparison is true.
472 // WARNING: like the previous function, this function it always returns
473 // the opcode that expects an immediate and a register. See above.
476 ChooseMovpregiForSetCC(const InstructionNode* instrNode)
478 MachineOpCode opCode = V9::INVALID_OPCODE;
480 switch(instrNode->getInstruction()->getOpcode())
482 case Instruction::SetEQ: opCode = V9::MOVRZi; break;
483 case Instruction::SetLE: opCode = V9::MOVRLEZi; break;
484 case Instruction::SetGE: opCode = V9::MOVRGEZi; break;
485 case Instruction::SetLT: opCode = V9::MOVRLZi; break;
486 case Instruction::SetGT: opCode = V9::MOVRGZi; break;
487 case Instruction::SetNE: opCode = V9::MOVRNZi; break;
488 default: assert(0 && "Unrecognized VM instr!"); break;
495 static inline MachineOpCode
496 ChooseConvertToFloatInstr(OpLabel vopCode, const Type* opType)
498 assert((vopCode == ToFloatTy || vopCode == ToDoubleTy) &&
499 "Unrecognized convert-to-float opcode!");
501 MachineOpCode opCode = V9::INVALID_OPCODE;
503 if (opType == Type::SByteTy || opType == Type::UByteTy ||
504 opType == Type::ShortTy || opType == Type::UShortTy ||
505 opType == Type::IntTy || opType == Type::UIntTy)
506 opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
507 else if (opType == Type::LongTy || opType == Type::ULongTy ||
508 isa<PointerType>(opType))
509 opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
510 else if (opType == Type::FloatTy)
511 opCode = (vopCode == ToFloatTy? V9::INVALID_OPCODE : V9::FSTOD);
512 else if (opType == Type::DoubleTy)
513 opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::INVALID_OPCODE);
515 assert(0 && "Trying to convert a non-scalar type to DOUBLE?");
520 static inline MachineOpCode
521 ChooseConvertFPToIntInstr(Type::PrimitiveID tid, const Type* opType)
523 MachineOpCode opCode = V9::INVALID_OPCODE;;
525 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
526 && "This function should only be called for FLOAT or DOUBLE");
528 // SPARC does not have a float-to-uint conversion, only a float-to-int.
529 // For converting an FP value to uint32_t, we first need to convert to
530 // uint64_t and then to uint32_t, or we may overflow the signed int
531 // representation even for legal uint32_t values. This expansion is
532 // done by the Preselection pass.
534 if (tid == Type::UIntTyID) {
535 assert(tid != Type::UIntTyID && "FP-to-uint conversions must be expanded"
536 " into FP->long->uint for SPARC v9: SO RUN PRESELECTION PASS!");
537 } else if (tid == Type::SByteTyID || tid == Type::ShortTyID ||
538 tid == Type::IntTyID || tid == Type::UByteTyID ||
539 tid == Type::UShortTyID) {
540 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
541 } else if (tid == Type::LongTyID || tid == Type::ULongTyID) {
542 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
544 assert(0 && "Should not get here, Mo!");
550 CreateConvertFPToIntInstr(Type::PrimitiveID destTID,
551 Value* srcVal, Value* destVal)
553 MachineOpCode opCode = ChooseConvertFPToIntInstr(destTID, srcVal->getType());
554 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
555 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
558 // CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer
559 // The FP value must be converted to the dest type in an FP register,
560 // and the result is then copied from FP to int register via memory.
562 // Since fdtoi converts to signed integers, any FP value V between MAXINT+1
563 // and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly
564 // *only* when converting to an unsigned. (Unsigned byte, short or long
565 // don't have this problem.)
566 // For unsigned int, we therefore have to generate the code sequence:
568 // if (V > (float) MAXINT) {
569 // unsigned result = (unsigned) (V - (float) MAXINT);
570 // result = result + (unsigned) MAXINT;
573 // result = (unsigned) V;
576 CreateCodeToConvertFloatToInt(const TargetMachine& target,
579 std::vector<MachineInstr*>& mvec,
580 MachineCodeForInstruction& mcfi)
582 // Create a temporary to represent the FP register into which the
583 // int value will placed after conversion. The type of this temporary
584 // depends on the type of FP register to use: single-prec for a 32-bit
585 // int or smaller; double-prec for a 64-bit int.
587 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
588 const Type* destTypeToUse = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
589 TmpInstruction* destForCast = new TmpInstruction(mcfi, destTypeToUse, opVal);
591 // Create the fp-to-int conversion code
592 MachineInstr* M =CreateConvertFPToIntInstr(destI->getType()->getPrimitiveID(),
596 // Create the fpreg-to-intreg copy code
597 target.getInstrInfo().
598 CreateCodeToCopyFloatToInt(target, destI->getParent()->getParent(),
599 destForCast, destI, mvec, mcfi);
603 static inline MachineOpCode
604 ChooseAddInstruction(const InstructionNode* instrNode)
606 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
610 static inline MachineInstr*
611 CreateMovFloatInstruction(const InstructionNode* instrNode,
612 const Type* resultType)
614 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
615 .addReg(instrNode->leftChild()->getValue())
616 .addRegDef(instrNode->getValue());
619 static inline MachineInstr*
620 CreateAddConstInstruction(const InstructionNode* instrNode)
622 MachineInstr* minstr = NULL;
624 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
625 assert(isa<Constant>(constOp));
627 // Cases worth optimizing are:
628 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
629 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
631 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
632 double dval = FPC->getValue();
634 minstr = CreateMovFloatInstruction(instrNode,
635 instrNode->getInstruction()->getType());
642 static inline MachineOpCode
643 ChooseSubInstructionByType(const Type* resultType)
645 MachineOpCode opCode = V9::INVALID_OPCODE;
647 if (resultType->isInteger() || isa<PointerType>(resultType)) {
650 switch(resultType->getPrimitiveID())
652 case Type::FloatTyID: opCode = V9::FSUBS; break;
653 case Type::DoubleTyID: opCode = V9::FSUBD; break;
654 default: assert(0 && "Invalid type for SUB instruction"); break;
662 static inline MachineInstr*
663 CreateSubConstInstruction(const InstructionNode* instrNode)
665 MachineInstr* minstr = NULL;
667 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
668 assert(isa<Constant>(constOp));
670 // Cases worth optimizing are:
671 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
672 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
674 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
675 double dval = FPC->getValue();
677 minstr = CreateMovFloatInstruction(instrNode,
678 instrNode->getInstruction()->getType());
685 static inline MachineOpCode
686 ChooseFcmpInstruction(const InstructionNode* instrNode)
688 MachineOpCode opCode = V9::INVALID_OPCODE;
690 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
691 switch(operand->getType()->getPrimitiveID()) {
692 case Type::FloatTyID: opCode = V9::FCMPS; break;
693 case Type::DoubleTyID: opCode = V9::FCMPD; break;
694 default: assert(0 && "Invalid type for FCMP instruction"); break;
701 // Assumes that leftArg and rightArg are both cast instructions.
704 BothFloatToDouble(const InstructionNode* instrNode)
706 InstrTreeNode* leftArg = instrNode->leftChild();
707 InstrTreeNode* rightArg = instrNode->rightChild();
708 InstrTreeNode* leftArgArg = leftArg->leftChild();
709 InstrTreeNode* rightArgArg = rightArg->leftChild();
710 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
712 // Check if both arguments are floats cast to double
713 return (leftArg->getValue()->getType() == Type::DoubleTy &&
714 leftArgArg->getValue()->getType() == Type::FloatTy &&
715 rightArgArg->getValue()->getType() == Type::FloatTy);
719 static inline MachineOpCode
720 ChooseMulInstructionByType(const Type* resultType)
722 MachineOpCode opCode = V9::INVALID_OPCODE;
724 if (resultType->isInteger())
727 switch(resultType->getPrimitiveID())
729 case Type::FloatTyID: opCode = V9::FMULS; break;
730 case Type::DoubleTyID: opCode = V9::FMULD; break;
731 default: assert(0 && "Invalid type for MUL instruction"); break;
739 static inline MachineInstr*
740 CreateIntNegInstruction(const TargetMachine& target,
743 return BuildMI(V9::SUBr, 3).addMReg(target.getRegInfo().getZeroRegNum())
744 .addReg(vreg).addRegDef(vreg);
748 // Create instruction sequence for any shift operation.
749 // SLL or SLLX on an operand smaller than the integer reg. size (64bits)
750 // requires a second instruction for explicit sign-extension.
751 // Note that we only have to worry about a sign-bit appearing in the
752 // most significant bit of the operand after shifting (e.g., bit 32 of
753 // Int or bit 16 of Short), so we do not have to worry about results
754 // that are as large as a normal integer register.
757 CreateShiftInstructions(const TargetMachine& target,
759 MachineOpCode shiftOpCode,
761 Value* optArgVal2, /* Use optArgVal2 if not NULL */
762 unsigned optShiftNum, /* else use optShiftNum */
763 Instruction* destVal,
764 std::vector<MachineInstr*>& mvec,
765 MachineCodeForInstruction& mcfi)
767 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
768 "Large shift sizes unexpected, but can be handled below: "
769 "You need to check whether or not it fits in immed field below");
771 // If this is a logical left shift of a type smaller than the standard
772 // integer reg. size, we have to extend the sign-bit into upper bits
773 // of dest, so we need to put the result of the SLL into a temporary.
775 Value* shiftDest = destVal;
776 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
778 if ((shiftOpCode == V9::SLLr5 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
779 // put SLL result into a temporary
780 shiftDest = new TmpInstruction(mcfi, argVal1, optArgVal2, "sllTmp");
783 MachineInstr* M = (optArgVal2 != NULL)
784 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
785 .addReg(shiftDest, MOTy::Def)
786 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
787 .addReg(shiftDest, MOTy::Def);
790 if (shiftDest != destVal) {
791 // extend the sign-bit of the result into all upper bits of dest
792 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
793 target.getInstrInfo().
794 CreateSignExtensionInstructions(target, F, shiftDest, destVal,
795 8*opSize, mvec, mcfi);
800 // Does not create any instructions if we cannot exploit constant to
801 // create a cheaper instruction.
802 // This returns the approximate cost of the instructions generated,
803 // which is used to pick the cheapest when both operands are constant.
805 CreateMulConstInstruction(const TargetMachine &target, Function* F,
806 Value* lval, Value* rval, Instruction* destVal,
807 std::vector<MachineInstr*>& mvec,
808 MachineCodeForInstruction& mcfi)
810 /* Use max. multiply cost, viz., cost of MULX */
811 unsigned cost = target.getInstrInfo().minLatency(V9::MULXr);
812 unsigned firstNewInstr = mvec.size();
814 Value* constOp = rval;
815 if (! isa<Constant>(constOp))
818 // Cases worth optimizing are:
819 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
820 // (2) Multiply by 2^x for integer types: replace with Shift
822 const Type* resultType = destVal->getType();
824 if (resultType->isInteger() || isa<PointerType>(resultType)) {
826 int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
827 constOp, constOp->getType(), isValidConst);
830 bool needNeg = false;
836 if (C == 0 || C == 1) {
837 cost = target.getInstrInfo().minLatency(V9::ADDr);
838 unsigned Zero = target.getRegInfo().getZeroRegNum();
841 M =BuildMI(V9::ADDr,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
843 M = BuildMI(V9::ADDr,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
845 } else if (isPowerOf2(C, pow)) {
846 unsigned opSize = target.getTargetData().getTypeSize(resultType);
847 MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
848 CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
849 destVal, mvec, mcfi);
852 if (mvec.size() > 0 && needNeg) {
853 // insert <reg = SUB 0, reg> after the instr to flip the sign
854 MachineInstr* M = CreateIntNegInstruction(target, destVal);
859 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
860 double dval = FPC->getValue();
861 if (fabs(dval) == 1) {
862 MachineOpCode opCode = (dval < 0)
863 ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
864 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
865 mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
870 if (firstNewInstr < mvec.size()) {
872 for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
873 cost += target.getInstrInfo().minLatency(mvec[i]->getOpCode());
880 // Does not create any instructions if we cannot exploit constant to
881 // create a cheaper instruction.
884 CreateCheapestMulConstInstruction(const TargetMachine &target,
886 Value* lval, Value* rval,
887 Instruction* destVal,
888 std::vector<MachineInstr*>& mvec,
889 MachineCodeForInstruction& mcfi)
892 if (isa<Constant>(lval) && isa<Constant>(rval)) {
893 // both operands are constant: evaluate and "set" in dest
894 Constant* P = ConstantFoldBinaryInstruction(Instruction::Mul,
895 cast<Constant>(lval),
896 cast<Constant>(rval));
897 target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi);
899 else if (isa<Constant>(rval)) // rval is constant, but not lval
900 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
901 else if (isa<Constant>(lval)) // lval is constant, but not rval
902 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
904 // else neither is constant
908 // Return NULL if we cannot exploit constant to create a cheaper instruction
910 CreateMulInstruction(const TargetMachine &target, Function* F,
911 Value* lval, Value* rval, Instruction* destVal,
912 std::vector<MachineInstr*>& mvec,
913 MachineCodeForInstruction& mcfi,
914 MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE)
916 unsigned L = mvec.size();
917 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
918 if (mvec.size() == L) {
919 // no instructions were added so create MUL reg, reg, reg.
920 // Use FSMULD if both operands are actually floats cast to doubles.
921 // Otherwise, use the default opcode for the appropriate type.
922 MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE)
924 : ChooseMulInstructionByType(destVal->getType()));
925 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
926 .addRegDef(destVal));
931 // Generate a divide instruction for Div or Rem.
932 // For Rem, this assumes that the operand type will be signed if the result
933 // type is signed. This is correct because they must have the same sign.
935 static inline MachineOpCode
936 ChooseDivInstruction(TargetMachine &target,
937 const InstructionNode* instrNode)
939 MachineOpCode opCode = V9::INVALID_OPCODE;
941 const Type* resultType = instrNode->getInstruction()->getType();
943 if (resultType->isInteger())
944 opCode = resultType->isSigned()? V9::SDIVXr : V9::UDIVXr;
946 switch(resultType->getPrimitiveID())
948 case Type::FloatTyID: opCode = V9::FDIVS; break;
949 case Type::DoubleTyID: opCode = V9::FDIVD; break;
950 default: assert(0 && "Invalid type for DIV instruction"); break;
957 // Return if we cannot exploit constant to create a cheaper instruction
959 CreateDivConstInstruction(TargetMachine &target,
960 const InstructionNode* instrNode,
961 std::vector<MachineInstr*>& mvec)
963 Value* LHS = instrNode->leftChild()->getValue();
964 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
965 if (!isa<Constant>(constOp))
968 Instruction* destVal = instrNode->getInstruction();
969 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
971 // Cases worth optimizing are:
972 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
973 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
975 const Type* resultType = instrNode->getInstruction()->getType();
977 if (resultType->isInteger()) {
980 int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
981 constOp, constOp->getType(), isValidConst);
983 bool needNeg = false;
990 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
991 .addRegDef(destVal));
992 } else if (isPowerOf2(C, pow)) {
996 if (resultType->isSigned()) {
997 // For N / 2^k, if the operand N is negative,
998 // we need to add (2^k - 1) before right-shifting by k, i.e.,
1000 // (N / 2^k) = N >> k, if N >= 0;
1001 // (N + 2^k - 1) >> k, if N < 0
1003 // If N is <= 32 bits, use:
1004 // sra N, 31, t1 // t1 = ~0, if N < 0, 0 else
1005 // srl t1, 32-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
1006 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
1007 // sra t3, k, result // result = N / 2^k
1009 // If N is 64 bits, use:
1010 // srax N, k-1, t1 // t1 = sign bit in high k positions
1011 // srlx t1, 64-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
1012 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
1013 // sra t3, k, result // result = N / 2^k
1015 TmpInstruction *sraTmp, *srlTmp, *addTmp;
1016 MachineCodeForInstruction& mcfi
1017 = MachineCodeForInstruction::get(destVal);
1018 sraTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
1019 srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getPlus2km1");
1020 addTmp = new TmpInstruction(mcfi, resultType, LHS, srlTmp,"incIfNeg");
1022 // Create the SRA or SRAX instruction to get the sign bit
1023 mvec.push_back(BuildMI((resultType==Type::LongTy) ?
1024 V9::SRAXi6 : V9::SRAi5, 3)
1026 .addSImm((resultType==Type::LongTy)? pow-1 : 31)
1027 .addRegDef(sraTmp));
1029 // Create the SRL or SRLX instruction to get the sign bit
1030 mvec.push_back(BuildMI((resultType==Type::LongTy) ?
1031 V9::SRLXi6 : V9::SRLi5, 3)
1033 .addSImm((resultType==Type::LongTy)? 64-pow : 32-pow)
1034 .addRegDef(srlTmp));
1036 // Create the ADD instruction to add 2^pow-1 for negative values
1037 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
1038 .addRegDef(addTmp));
1040 // Get the shift operand and "right-shift" opcode to do the divide
1041 shiftOperand = addTmp;
1042 opCode = (resultType==Type::LongTy) ? V9::SRAXi6 : V9::SRAi5;
1044 // Get the shift operand and "right-shift" opcode to do the divide
1046 opCode = (resultType==Type::LongTy) ? V9::SRLXi6 : V9::SRLi5;
1049 // Now do the actual shift!
1050 mvec.push_back(BuildMI(opCode, 3).addReg(shiftOperand).addZImm(pow)
1051 .addRegDef(destVal));
1054 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
1055 // insert <reg = SUB 0, reg> after the instr to flip the sign
1056 mvec.push_back(CreateIntNegInstruction(target, destVal));
1060 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
1061 double dval = FPC->getValue();
1062 if (fabs(dval) == 1) {
1064 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
1065 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
1067 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(destVal));
1075 CreateCodeForVariableSizeAlloca(const TargetMachine& target,
1076 Instruction* result,
1078 Value* numElementsVal,
1079 std::vector<MachineInstr*>& getMvec)
1081 Value* totalSizeVal;
1083 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
1084 Function *F = result->getParent()->getParent();
1086 // Enforce the alignment constraints on the stack pointer at
1087 // compile time if the total size is a known constant.
1088 if (isa<Constant>(numElementsVal)) {
1090 int64_t numElem = (int64_t) target.getInstrInfo().
1091 ConvertConstantToIntType(target, numElementsVal,
1092 numElementsVal->getType(), isValid);
1093 assert(isValid && "Unexpectedly large array dimension in alloca!");
1094 int64_t total = numElem * tsize;
1095 if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
1096 total += target.getFrameInfo().getStackFrameSizeAlignment() - extra;
1097 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
1099 // The size is not a constant. Generate code to compute it and
1100 // code to pad the size for stack alignment.
1101 // Create a Value to hold the (constant) element size
1102 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
1104 // Create temporary values to hold the result of MUL, SLL, SRL
1105 // To pad `size' to next smallest multiple of 16:
1106 // size = (size + 15) & (-16 = 0xfffffffffffffff0)
1108 TmpInstruction* tmpProd = new TmpInstruction(mcfi,numElementsVal, tsizeVal);
1109 TmpInstruction* tmpAdd15= new TmpInstruction(mcfi,numElementsVal, tmpProd);
1110 TmpInstruction* tmpAndf0= new TmpInstruction(mcfi,numElementsVal, tmpAdd15);
1112 // Instruction 1: mul numElements, typeSize -> tmpProd
1113 // This will optimize the MUL as far as possible.
1114 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec,
1115 mcfi, INVALID_MACHINE_OPCODE);
1117 // Instruction 2: andn tmpProd, 0x0f -> tmpAndn
1118 getMvec.push_back(BuildMI(V9::ADDi, 3).addReg(tmpProd).addSImm(15)
1119 .addReg(tmpAdd15, MOTy::Def));
1121 // Instruction 3: add tmpAndn, 0x10 -> tmpAdd16
1122 getMvec.push_back(BuildMI(V9::ANDi, 3).addReg(tmpAdd15).addSImm(-16)
1123 .addReg(tmpAndf0, MOTy::Def));
1125 totalSizeVal = tmpAndf0;
1128 // Get the constant offset from SP for dynamically allocated storage
1129 // and create a temporary Value to hold it.
1130 MachineFunction& mcInfo = MachineFunction::get(F);
1132 ConstantSInt* dynamicAreaOffset =
1133 ConstantSInt::get(Type::IntTy,
1134 target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp));
1135 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
1137 unsigned SPReg = target.getRegInfo().getStackPointer();
1139 // Instruction 2: sub %sp, totalSizeVal -> %sp
1140 getMvec.push_back(BuildMI(V9::SUBr, 3).addMReg(SPReg).addReg(totalSizeVal)
1141 .addMReg(SPReg,MOTy::Def));
1143 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
1144 getMvec.push_back(BuildMI(V9::ADDr,3).addMReg(SPReg).addReg(dynamicAreaOffset)
1145 .addRegDef(result));
1150 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
1151 Instruction* result,
1153 unsigned numElements,
1154 std::vector<MachineInstr*>& getMvec)
1156 assert(tsize > 0 && "Illegal (zero) type size for alloca");
1157 assert(result && result->getParent() &&
1158 "Result value is not part of a function?");
1159 Function *F = result->getParent()->getParent();
1160 MachineFunction &mcInfo = MachineFunction::get(F);
1162 // Put the variable in the dynamically sized area of the frame if either:
1163 // (a) The offset is too large to use as an immediate in load/stores
1164 // (check LDX because all load/stores have the same-size immed. field).
1165 // (b) The object is "large", so it could cause many other locals,
1166 // spills, and temporaries to have large offsets.
1167 // NOTE: We use LARGE = 8 * argSlotSize = 64 bytes.
1168 // You've gotta love having only 13 bits for constant offset values :-|.
1170 unsigned paddedSize;
1171 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
1173 tsize * numElements);
1175 if (((int)paddedSize) > 8 * target.getFrameInfo().getSizeOfEachArgOnStack() ||
1176 ! target.getInstrInfo().constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
1177 CreateCodeForVariableSizeAlloca(target, result, tsize,
1178 ConstantSInt::get(Type::IntTy,numElements),
1183 // else offset fits in immediate field so go ahead and allocate it.
1184 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
1186 // Create a temporary Value to hold the constant offset.
1187 // This is needed because it may not fit in the immediate field.
1188 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
1190 // Instruction 1: add %fp, offsetFromFP -> result
1191 unsigned FPReg = target.getRegInfo().getFramePointer();
1192 getMvec.push_back(BuildMI(V9::ADDr, 3).addMReg(FPReg).addReg(offsetVal)
1193 .addRegDef(result));
1197 //------------------------------------------------------------------------
1198 // Function SetOperandsForMemInstr
1200 // Choose addressing mode for the given load or store instruction.
1201 // Use [reg+reg] if it is an indexed reference, and the index offset is
1202 // not a constant or if it cannot fit in the offset field.
1203 // Use [reg+offset] in all other cases.
1205 // This assumes that all array refs are "lowered" to one of these forms:
1206 // %x = load (subarray*) ptr, constant ; single constant offset
1207 // %x = load (subarray*) ptr, offsetVal ; single non-constant offset
1208 // Generally, this should happen via strength reduction + LICM.
1209 // Also, strength reduction should take care of using the same register for
1210 // the loop index variable and an array index, when that is profitable.
1211 //------------------------------------------------------------------------
1214 SetOperandsForMemInstr(unsigned Opcode,
1215 std::vector<MachineInstr*>& mvec,
1216 InstructionNode* vmInstrNode,
1217 const TargetMachine& target)
1219 Instruction* memInst = vmInstrNode->getInstruction();
1220 // Index vector, ptr value, and flag if all indices are const.
1221 std::vector<Value*> idxVec;
1222 bool allConstantIndices;
1223 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
1225 // Now create the appropriate operands for the machine instruction.
1226 // First, initialize so we default to storing the offset in a register.
1227 int64_t smallConstOffset = 0;
1228 Value* valueForRegOffset = NULL;
1229 MachineOperand::MachineOperandType offsetOpType =
1230 MachineOperand::MO_VirtualRegister;
1232 // Check if there is an index vector and if so, compute the
1233 // right offset for structures and for arrays
1235 if (!idxVec.empty()) {
1236 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
1238 // If all indices are constant, compute the combined offset directly.
1239 if (allConstantIndices) {
1240 // Compute the offset value using the index vector. Create a
1241 // virtual reg. for it since it may not fit in the immed field.
1242 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
1243 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
1245 // There is at least one non-constant offset. Therefore, this must
1246 // be an array ref, and must have been lowered to a single non-zero
1247 // offset. (An extra leading zero offset, if any, can be ignored.)
1248 // Generate code sequence to compute address from index.
1250 bool firstIdxIsZero = IsZero(idxVec[0]);
1251 assert(idxVec.size() == 1U + firstIdxIsZero
1252 && "Array refs must be lowered before Instruction Selection");
1254 Value* idxVal = idxVec[firstIdxIsZero];
1256 std::vector<MachineInstr*> mulVec;
1258 new TmpInstruction(MachineCodeForInstruction::get(memInst),
1259 Type::ULongTy, memInst);
1261 // Get the array type indexed by idxVal, and compute its element size.
1262 // The call to getTypeSize() will fail if size is not constant.
1263 const Type* vecType = (firstIdxIsZero
1264 ? GetElementPtrInst::getIndexedType(ptrType,
1265 std::vector<Value*>(1U, idxVec[0]),
1266 /*AllowCompositeLeaf*/ true)
1268 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
1269 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
1270 target.getTargetData().getTypeSize(eltType));
1272 // CreateMulInstruction() folds constants intelligently enough.
1273 CreateMulInstruction(target, memInst->getParent()->getParent(),
1274 idxVal, /* lval, not likely to be const*/
1275 eltSizeVal, /* rval, likely to be constant */
1277 mulVec, MachineCodeForInstruction::get(memInst),
1278 INVALID_MACHINE_OPCODE);
1280 assert(mulVec.size() > 0 && "No multiply code created?");
1281 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
1283 valueForRegOffset = addr;
1286 offsetOpType = MachineOperand::MO_SignExtendedImmed;
1287 smallConstOffset = 0;
1291 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
1292 // For LOAD or GET_ELEMENT_PTR,
1293 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
1295 unsigned offsetOpNum, ptrOpNum;
1297 if (memInst->getOpcode() == Instruction::Store) {
1298 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1299 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1300 .addReg(ptrVal).addReg(valueForRegOffset);
1302 Opcode = convertOpcodeFromRegToImm(Opcode);
1303 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1304 .addReg(ptrVal).addSImm(smallConstOffset);
1307 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1308 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
1309 .addRegDef(memInst);
1311 Opcode = convertOpcodeFromRegToImm(Opcode);
1312 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
1313 .addRegDef(memInst);
1321 // Substitute operand `operandNum' of the instruction in node `treeNode'
1322 // in place of the use(s) of that instruction in node `parent'.
1323 // Check both explicit and implicit operands!
1324 // Also make sure to skip over a parent who:
1325 // (1) is a list node in the Burg tree, or
1326 // (2) itself had its results forwarded to its parent
1329 ForwardOperand(InstructionNode* treeNode,
1330 InstrTreeNode* parent,
1333 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
1335 Instruction* unusedOp = treeNode->getInstruction();
1336 Value* fwdOp = unusedOp->getOperand(operandNum);
1338 // The parent itself may be a list node, so find the real parent instruction
1339 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode)
1341 parent = parent->parent();
1342 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
1344 InstructionNode* parentInstrNode = (InstructionNode*) parent;
1346 Instruction* userInstr = parentInstrNode->getInstruction();
1347 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
1349 // The parent's mvec would be empty if it was itself forwarded.
1350 // Recursively call ForwardOperand in that case...
1352 if (mvec.size() == 0) {
1353 assert(parent->parent() != NULL &&
1354 "Parent could not have been forwarded, yet has no instructions?");
1355 ForwardOperand(treeNode, parent->parent(), operandNum);
1357 for (unsigned i=0, N=mvec.size(); i < N; i++) {
1358 MachineInstr* minstr = mvec[i];
1359 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) {
1360 const MachineOperand& mop = minstr->getOperand(i);
1361 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
1362 mop.getVRegValue() == unusedOp)
1364 minstr->SetMachineOperandVal(i, MachineOperand::MO_VirtualRegister,
1369 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
1370 if (minstr->getImplicitRef(i) == unusedOp) {
1371 minstr->setImplicitRef(i, fwdOp,
1372 minstr->getImplicitOp(i).opIsDefOnly(),
1373 minstr->getImplicitOp(i).opIsDefAndUse());
1381 AllUsesAreBranches(const Instruction* setccI)
1383 for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end();
1385 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
1386 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
1391 // Generate code for any intrinsic that needs a special code sequence
1392 // instead of a regular call. If not that kind of intrinsic, do nothing.
1393 // Returns true if code was generated, otherwise false.
1395 bool CodeGenIntrinsic(LLVMIntrinsic::ID iid, CallInst &callInstr,
1396 TargetMachine &target,
1397 std::vector<MachineInstr*>& mvec)
1400 case LLVMIntrinsic::va_start: {
1401 // Get the address of the first vararg value on stack and copy it to
1402 // the argument of va_start(va_list* ap).
1404 Function* func = cast<Function>(callInstr.getParent()->getParent());
1405 int numFixedArgs = func->getFunctionType()->getNumParams();
1406 int fpReg = target.getFrameInfo().getIncomingArgBaseRegNum();
1407 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
1408 int firstVarArgOff = numFixedArgs * argSize + target.getFrameInfo().
1409 getFirstIncomingArgOffset(MachineFunction::get(func), ignore);
1410 mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
1411 addReg(callInstr.getOperand(1)));
1415 case LLVMIntrinsic::va_end:
1416 return true; // no-op on Sparc
1418 case LLVMIntrinsic::va_copy:
1419 // Simple copy of current va_list (arg2) to new va_list (arg1)
1420 mvec.push_back(BuildMI(V9::ORr, 3).
1421 addMReg(target.getRegInfo().getZeroRegNum()).
1422 addReg(callInstr.getOperand(2)).
1423 addReg(callInstr.getOperand(1)));
1431 //******************* Externally Visible Functions *************************/
1433 //------------------------------------------------------------------------
1434 // External Function: ThisIsAChainRule
1437 // Check if a given BURG rule is a chain rule.
1438 //------------------------------------------------------------------------
1441 ThisIsAChainRule(int eruleno)
1445 case 111: // stmt: reg
1469 return false; break;
1474 //------------------------------------------------------------------------
1475 // External Function: GetInstructionsByRule
1478 // Choose machine instructions for the SPARC according to the
1479 // patterns chosen by the BURG-generated parser.
1480 //------------------------------------------------------------------------
1483 GetInstructionsByRule(InstructionNode* subtreeRoot,
1486 TargetMachine &target,
1487 std::vector<MachineInstr*>& mvec)
1489 bool checkCast = false; // initialize here to use fall-through
1490 bool maskUnsignedResult = false;
1492 int forwardOperandNum = -1;
1493 unsigned allocaSize = 0;
1494 MachineInstr* M, *M2;
1496 bool foldCase = false;
1500 // If the code for this instruction was folded into the parent (user),
1502 if (subtreeRoot->isFoldedIntoParent())
1506 // Let's check for chain rules outside the switch so that we don't have
1507 // to duplicate the list of chain rule production numbers here again
1509 if (ThisIsAChainRule(ruleForNode))
1511 // Chain rules have a single nonterminal on the RHS.
1512 // Get the rule that matches the RHS non-terminal and use that instead.
1514 assert(nts[0] && ! nts[1]
1515 && "A chain rule should have only one RHS non-terminal!");
1516 nextRule = burm_rule(subtreeRoot->state, nts[0]);
1517 nts = burm_nts[nextRule];
1518 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
1522 switch(ruleForNode) {
1523 case 1: // stmt: Ret
1524 case 2: // stmt: RetValue(reg)
1525 { // NOTE: Prepass of register allocation is responsible
1526 // for moving return value to appropriate register.
1527 // Copy the return value to the required return register.
1528 // Mark the return Value as an implicit ref of the RET instr..
1529 // Mark the return-address register as a hidden virtual reg.
1530 // Finally put a NOP in the delay slot.
1531 ReturnInst *returnInstr=cast<ReturnInst>(subtreeRoot->getInstruction());
1532 Value* retVal = returnInstr->getReturnValue();
1533 MachineCodeForInstruction& mcfi =
1534 MachineCodeForInstruction::get(returnInstr);
1536 // Create a hidden virtual reg to represent the return address register
1537 // used by the machine instruction but not represented in LLVM.
1539 Instruction* returnAddrTmp = new TmpInstruction(mcfi, returnInstr);
1541 MachineInstr* retMI =
1542 BuildMI(V9::JMPLRETi, 3).addReg(returnAddrTmp).addSImm(8)
1543 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
1545 // If there is a value to return, we need to:
1546 // (a) Sign-extend the value if it is smaller than 8 bytes (reg size)
1547 // (b) Insert a copy to copy the return value to the appropriate reg.
1548 // -- For FP values, create a FMOVS or FMOVD instruction
1549 // -- For non-FP values, create an add-with-0 instruction
1551 if (retVal != NULL) {
1552 const UltraSparcRegInfo& regInfo =
1553 (UltraSparcRegInfo&) target.getRegInfo();
1554 const Type* retType = retVal->getType();
1555 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
1556 unsigned retRegNum = (retType->isFloatingPoint()
1557 ? (unsigned) SparcFloatRegClass::f0
1558 : (unsigned) SparcIntRegClass::i0);
1559 retRegNum = regInfo.getUnifiedRegNum(regClassID, retRegNum);
1561 // () Insert sign-extension instructions for small signed values.
1563 Value* retValToUse = retVal;
1564 if (retType->isIntegral() && retType->isSigned()) {
1565 unsigned retSize = target.getTargetData().getTypeSize(retType);
1567 // create a temporary virtual reg. to hold the sign-extension
1568 retValToUse = new TmpInstruction(mcfi, retVal);
1570 // sign-extend retVal and put the result in the temporary reg.
1571 target.getInstrInfo().CreateSignExtensionInstructions
1572 (target, returnInstr->getParent()->getParent(),
1573 retVal, retValToUse, 8*retSize, mvec, mcfi);
1577 // (b) Now, insert a copy to to the appropriate register:
1578 // -- For FP values, create a FMOVS or FMOVD instruction
1579 // -- For non-FP values, create an add-with-0 instruction
1581 // First, create a virtual register to represent the register and
1582 // mark this vreg as being an implicit operand of the ret MI.
1583 TmpInstruction* retVReg =
1584 new TmpInstruction(mcfi, retValToUse, NULL, "argReg");
1586 retMI->addImplicitRef(retVReg);
1588 if (retType->isFloatingPoint())
1589 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
1590 .addReg(retValToUse).addReg(retVReg, MOTy::Def));
1592 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
1593 .addReg(retValToUse).addSImm((int64_t) 0)
1594 .addReg(retVReg, MOTy::Def));
1596 // Mark the operand with the register it should be assigned
1597 M->SetRegForOperand(M->getNumOperands()-1, retRegNum);
1598 retMI->SetRegForImplicitRef(retMI->getNumImplicitRefs()-1, retRegNum);
1603 // Now insert the RET instruction and a NOP for the delay slot
1604 mvec.push_back(retMI);
1605 mvec.push_back(BuildMI(V9::NOP, 0));
1610 case 3: // stmt: Store(reg,reg)
1611 case 4: // stmt: Store(reg,ptrreg)
1612 SetOperandsForMemInstr(ChooseStoreInstruction(
1613 subtreeRoot->leftChild()->getValue()->getType()),
1614 mvec, subtreeRoot, target);
1617 case 5: // stmt: BrUncond
1619 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1620 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
1623 mvec.push_back(BuildMI(V9::NOP, 0));
1627 case 206: // stmt: BrCond(setCCconst)
1628 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
1629 // If the constant is ZERO, we can use the branch-on-integer-register
1630 // instructions and avoid the SUBcc instruction entirely.
1631 // Otherwise this is just the same as case 5, so just fall through.
1633 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
1635 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
1636 Constant *constVal = cast<Constant>(constNode->getValue());
1639 if ((constVal->getType()->isInteger()
1640 || isa<PointerType>(constVal->getType()))
1641 && target.getInstrInfo().ConvertConstantToIntType(target,
1642 constVal, constVal->getType(), isValidConst) == 0
1645 // That constant is a zero after all...
1646 // Use the left child of setCC as the first argument!
1647 // Mark the setCC node so that no code is generated for it.
1648 InstructionNode* setCCNode = (InstructionNode*)
1649 subtreeRoot->leftChild();
1650 assert(setCCNode->getOpLabel() == SetCCOp);
1651 setCCNode->markFoldedIntoParent();
1653 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
1655 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
1656 .addReg(setCCNode->leftChild()->getValue())
1657 .addPCDisp(brInst->getSuccessor(0));
1661 mvec.push_back(BuildMI(V9::NOP, 0));
1664 mvec.push_back(BuildMI(V9::BA, 1)
1665 .addPCDisp(brInst->getSuccessor(1)));
1668 mvec.push_back(BuildMI(V9::NOP, 0));
1671 // ELSE FALL THROUGH
1674 case 6: // stmt: BrCond(setCC)
1675 { // bool => boolean was computed with SetCC.
1676 // The branch to use depends on whether it is FP, signed, or unsigned.
1677 // If it is an integer CC, we also need to find the unique
1678 // TmpInstruction representing that CC.
1680 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
1681 const Type* setCCType;
1682 unsigned Opcode = ChooseBccInstruction(subtreeRoot, setCCType);
1683 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
1684 brInst->getParent()->getParent(),
1686 MachineCodeForInstruction::get(brInst));
1687 M = BuildMI(Opcode, 2).addCCReg(ccValue)
1688 .addPCDisp(brInst->getSuccessor(0));
1692 mvec.push_back(BuildMI(V9::NOP, 0));
1695 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
1698 mvec.push_back(BuildMI(V9::NOP, 0));
1702 case 208: // stmt: BrCond(boolconst)
1704 // boolconst => boolean is a constant; use BA to first or second label
1705 Constant* constVal =
1706 cast<Constant>(subtreeRoot->leftChild()->getValue());
1707 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
1709 M = BuildMI(V9::BA, 1).addPCDisp(
1710 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
1714 mvec.push_back(BuildMI(V9::NOP, 0));
1718 case 8: // stmt: BrCond(boolreg)
1719 { // boolreg => boolean is recorded in an integer register.
1720 // Use branch-on-integer-register instruction.
1722 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1723 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
1724 .addPCDisp(BI->getSuccessor(0));
1728 mvec.push_back(BuildMI(V9::NOP, 0));
1731 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
1734 mvec.push_back(BuildMI(V9::NOP, 0));
1738 case 9: // stmt: Switch(reg)
1739 assert(0 && "*** SWITCH instruction is not implemented yet.");
1742 case 10: // reg: VRegList(reg, reg)
1743 assert(0 && "VRegList should never be the topmost non-chain rule");
1746 case 21: // bool: Not(bool,reg): Compute with a conditional-move-on-reg
1747 { // First find the unary operand. It may be left or right, usually right.
1748 Instruction* notI = subtreeRoot->getInstruction();
1749 Value* notArg = BinaryOperator::getNotArgument(
1750 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1751 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1753 // Unconditionally set register to 0
1754 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(notI));
1756 // Now conditionally move 1 into the register.
1757 // Mark the register as a use (as well as a def) because the old
1758 // value will be retained if the condition is false.
1759 mvec.push_back(BuildMI(V9::MOVRZi, 3).addReg(notArg).addZImm(1)
1760 .addReg(notI, MOTy::UseAndDef));
1765 case 421: // reg: BNot(reg,reg): Compute as reg = reg XOR-NOT 0
1766 { // First find the unary operand. It may be left or right, usually right.
1767 Value* notArg = BinaryOperator::getNotArgument(
1768 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1769 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1770 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(notArg).addMReg(ZeroReg)
1771 .addRegDef(subtreeRoot->getValue()));
1775 case 322: // reg: Not(tobool, reg):
1776 // Fold CAST-TO-BOOL with NOT by inverting the sense of cast-to-bool
1778 // Just fall through!
1780 case 22: // reg: ToBoolTy(reg):
1782 Instruction* castI = subtreeRoot->getInstruction();
1783 Value* opVal = subtreeRoot->leftChild()->getValue();
1784 assert(opVal->getType()->isIntegral() ||
1785 isa<PointerType>(opVal->getType()));
1787 // Unconditionally set register to 0
1788 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(castI));
1790 // Now conditionally move 1 into the register.
1791 // Mark the register as a use (as well as a def) because the old
1792 // value will be retained if the condition is false.
1793 MachineOpCode opCode = foldCase? V9::MOVRZi : V9::MOVRNZi;
1794 mvec.push_back(BuildMI(opCode, 3).addReg(opVal).addZImm(1)
1795 .addReg(castI, MOTy::UseAndDef));
1800 case 23: // reg: ToUByteTy(reg)
1801 case 24: // reg: ToSByteTy(reg)
1802 case 25: // reg: ToUShortTy(reg)
1803 case 26: // reg: ToShortTy(reg)
1804 case 27: // reg: ToUIntTy(reg)
1805 case 28: // reg: ToIntTy(reg)
1806 case 29: // reg: ToULongTy(reg)
1807 case 30: // reg: ToLongTy(reg)
1809 //======================================================================
1810 // Rules for integer conversions:
1813 // From ISO 1998 C++ Standard, Sec. 4.7:
1815 // 2. If the destination type is unsigned, the resulting value is
1816 // the least unsigned integer congruent to the source integer
1817 // (modulo 2n where n is the number of bits used to represent the
1818 // unsigned type). [Note: In a two s complement representation,
1819 // this conversion is conceptual and there is no change in the
1820 // bit pattern (if there is no truncation). ]
1822 // 3. If the destination type is signed, the value is unchanged if
1823 // it can be represented in the destination type (and bitfield width);
1824 // otherwise, the value is implementation-defined.
1827 // Since we assume 2s complement representations, this implies:
1829 // -- If operand is smaller than destination, zero-extend or sign-extend
1830 // according to the signedness of the *operand*: source decides:
1831 // (1) If operand is signed, sign-extend it.
1832 // If dest is unsigned, zero-ext the result!
1833 // (2) If operand is unsigned, our current invariant is that
1834 // it's high bits are correct, so zero-extension is not needed.
1836 // -- If operand is same size as or larger than destination,
1837 // zero-extend or sign-extend according to the signedness of
1838 // the *destination*: destination decides:
1839 // (1) If destination is signed, sign-extend (truncating if needed)
1840 // This choice is implementation defined. We sign-extend the
1841 // operand, which matches both Sun's cc and gcc3.2.
1842 // (2) If destination is unsigned, zero-extend (truncating if needed)
1843 //======================================================================
1845 Instruction* destI = subtreeRoot->getInstruction();
1846 Function* currentFunc = destI->getParent()->getParent();
1847 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(destI);
1849 Value* opVal = subtreeRoot->leftChild()->getValue();
1850 const Type* opType = opVal->getType();
1851 const Type* destType = destI->getType();
1852 unsigned opSize = target.getTargetData().getTypeSize(opType);
1853 unsigned destSize = target.getTargetData().getTypeSize(destType);
1855 bool isIntegral = opType->isIntegral() || isa<PointerType>(opType);
1857 if (opType == Type::BoolTy ||
1858 opType == destType ||
1859 isIntegral && opSize == destSize && opSize == 8) {
1860 // nothing to do in all these cases
1861 forwardOperandNum = 0; // forward first operand to user
1863 } else if (opType->isFloatingPoint()) {
1865 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, mcfi);
1866 if (destI->getType()->isUnsigned())
1867 maskUnsignedResult = true; // not handled by fp->int code
1869 } else if (isIntegral) {
1871 bool opSigned = opType->isSigned();
1872 bool destSigned = destType->isSigned();
1873 unsigned extSourceInBits = 8 * std::min<unsigned>(opSize, destSize);
1875 assert(! (opSize == destSize && opSigned == destSigned) &&
1876 "How can different int types have same size and signedness?");
1878 bool signExtend = (opSize < destSize && opSigned ||
1879 opSize >= destSize && destSigned);
1881 bool signAndZeroExtend = (opSize < destSize && destSize < 8u &&
1882 opSigned && !destSigned);
1883 assert(!signAndZeroExtend || signExtend);
1885 bool zeroExtendOnly = opSize >= destSize && !destSigned;
1886 assert(!zeroExtendOnly || !signExtend);
1889 Value* signExtDest = (signAndZeroExtend
1890 ? new TmpInstruction(mcfi, destType, opVal)
1893 target.getInstrInfo().CreateSignExtensionInstructions
1894 (target, currentFunc,opVal,signExtDest,extSourceInBits,mvec,mcfi);
1896 if (signAndZeroExtend)
1897 target.getInstrInfo().CreateZeroExtensionInstructions
1898 (target, currentFunc, signExtDest, destI, 8*destSize, mvec, mcfi);
1900 else if (zeroExtendOnly) {
1901 target.getInstrInfo().CreateZeroExtensionInstructions
1902 (target, currentFunc, opVal, destI, extSourceInBits, mvec, mcfi);
1905 forwardOperandNum = 0; // forward first operand to user
1908 assert(0 && "Unrecognized operand type for convert-to-integer");
1913 case 31: // reg: ToFloatTy(reg):
1914 case 32: // reg: ToDoubleTy(reg):
1915 case 232: // reg: ToDoubleTy(Constant):
1917 // If this instruction has a parent (a user) in the tree
1918 // and the user is translated as an FsMULd instruction,
1919 // then the cast is unnecessary. So check that first.
1920 // In the future, we'll want to do the same for the FdMULq instruction,
1921 // so do the check here instead of only for ToFloatTy(reg).
1923 if (subtreeRoot->parent() != NULL) {
1924 const MachineCodeForInstruction& mcfi =
1925 MachineCodeForInstruction::get(
1926 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
1927 if (mcfi.size() == 0 || mcfi.front()->getOpCode() == V9::FSMULD)
1928 forwardOperandNum = 0; // forward first operand to user
1931 if (forwardOperandNum != 0) { // we do need the cast
1932 Value* leftVal = subtreeRoot->leftChild()->getValue();
1933 const Type* opType = leftVal->getType();
1934 MachineOpCode opCode=ChooseConvertToFloatInstr(
1935 subtreeRoot->getOpLabel(), opType);
1936 if (opCode == V9::INVALID_OPCODE) { // no conversion needed
1937 forwardOperandNum = 0; // forward first operand to user
1939 // If the source operand is a non-FP type it must be
1940 // first copied from int to float register via memory!
1941 Instruction *dest = subtreeRoot->getInstruction();
1944 if (! opType->isFloatingPoint()) {
1945 // Create a temporary to represent the FP register
1946 // into which the integer will be copied via memory.
1947 // The type of this temporary will determine the FP
1948 // register used: single-prec for a 32-bit int or smaller,
1949 // double-prec for a 64-bit int.
1952 target.getTargetData().getTypeSize(leftVal->getType());
1953 Type* tmpTypeToUse =
1954 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
1955 MachineCodeForInstruction &destMCFI =
1956 MachineCodeForInstruction::get(dest);
1957 srcForCast = new TmpInstruction(destMCFI, tmpTypeToUse, dest);
1959 target.getInstrInfo().CreateCodeToCopyIntToFloat(target,
1960 dest->getParent()->getParent(),
1961 leftVal, cast<Instruction>(srcForCast),
1964 srcForCast = leftVal;
1966 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
1972 case 19: // reg: ToArrayTy(reg):
1973 case 20: // reg: ToPointerTy(reg):
1974 forwardOperandNum = 0; // forward first operand to user
1977 case 233: // reg: Add(reg, Constant)
1978 maskUnsignedResult = true;
1979 M = CreateAddConstInstruction(subtreeRoot);
1984 // ELSE FALL THROUGH
1986 case 33: // reg: Add(reg, reg)
1987 maskUnsignedResult = true;
1988 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
1991 case 234: // reg: Sub(reg, Constant)
1992 maskUnsignedResult = true;
1993 M = CreateSubConstInstruction(subtreeRoot);
1998 // ELSE FALL THROUGH
2000 case 34: // reg: Sub(reg, reg)
2001 maskUnsignedResult = true;
2002 Add3OperandInstr(ChooseSubInstructionByType(
2003 subtreeRoot->getInstruction()->getType()),
2007 case 135: // reg: Mul(todouble, todouble)
2011 case 35: // reg: Mul(reg, reg)
2013 maskUnsignedResult = true;
2014 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
2016 : INVALID_MACHINE_OPCODE);
2017 Instruction* mulInstr = subtreeRoot->getInstruction();
2018 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
2019 subtreeRoot->leftChild()->getValue(),
2020 subtreeRoot->rightChild()->getValue(),
2022 MachineCodeForInstruction::get(mulInstr),forceOp);
2025 case 335: // reg: Mul(todouble, todoubleConst)
2029 case 235: // reg: Mul(reg, Constant)
2031 maskUnsignedResult = true;
2032 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
2034 : INVALID_MACHINE_OPCODE);
2035 Instruction* mulInstr = subtreeRoot->getInstruction();
2036 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
2037 subtreeRoot->leftChild()->getValue(),
2038 subtreeRoot->rightChild()->getValue(),
2040 MachineCodeForInstruction::get(mulInstr),
2044 case 236: // reg: Div(reg, Constant)
2045 maskUnsignedResult = true;
2047 CreateDivConstInstruction(target, subtreeRoot, mvec);
2048 if (mvec.size() > L)
2050 // ELSE FALL THROUGH
2052 case 36: // reg: Div(reg, reg)
2054 maskUnsignedResult = true;
2056 // If either operand of divide is smaller than 64 bits, we have
2057 // to make sure the unused top bits are correct because they affect
2058 // the result. These bits are already correct for unsigned values.
2059 // They may be incorrect for signed values, so sign extend to fill in.
2060 Instruction* divI = subtreeRoot->getInstruction();
2061 Value* divOp1 = subtreeRoot->leftChild()->getValue();
2062 Value* divOp2 = subtreeRoot->rightChild()->getValue();
2063 Value* divOp1ToUse = divOp1;
2064 Value* divOp2ToUse = divOp2;
2065 if (divI->getType()->isSigned()) {
2066 unsigned opSize=target.getTargetData().getTypeSize(divI->getType());
2068 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(divI);
2069 divOp1ToUse = new TmpInstruction(mcfi, divOp1);
2070 divOp2ToUse = new TmpInstruction(mcfi, divOp2);
2071 target.getInstrInfo().
2072 CreateSignExtensionInstructions(target,
2073 divI->getParent()->getParent(),
2074 divOp1, divOp1ToUse,
2075 8*opSize, mvec, mcfi);
2076 target.getInstrInfo().
2077 CreateSignExtensionInstructions(target,
2078 divI->getParent()->getParent(),
2079 divOp2, divOp2ToUse,
2080 8*opSize, mvec, mcfi);
2084 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
2085 .addReg(divOp1ToUse)
2086 .addReg(divOp2ToUse)
2092 case 37: // reg: Rem(reg, reg)
2093 case 237: // reg: Rem(reg, Constant)
2095 maskUnsignedResult = true;
2097 Instruction* remI = subtreeRoot->getInstruction();
2098 Value* divOp1 = subtreeRoot->leftChild()->getValue();
2099 Value* divOp2 = subtreeRoot->rightChild()->getValue();
2101 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(remI);
2103 // If second operand of divide is smaller than 64 bits, we have
2104 // to make sure the unused top bits are correct because they affect
2105 // the result. These bits are already correct for unsigned values.
2106 // They may be incorrect for signed values, so sign extend to fill in.
2108 Value* divOpToUse = divOp2;
2109 if (divOp2->getType()->isSigned()) {
2110 unsigned opSize=target.getTargetData().getTypeSize(divOp2->getType());
2112 divOpToUse = new TmpInstruction(mcfi, divOp2);
2113 target.getInstrInfo().
2114 CreateSignExtensionInstructions(target,
2115 remI->getParent()->getParent(),
2117 8*opSize, mvec, mcfi);
2121 // Now compute: result = rem V1, V2 as:
2122 // result = V1 - (V1 / signExtend(V2)) * signExtend(V2)
2124 TmpInstruction* quot = new TmpInstruction(mcfi, divOp1, divOpToUse);
2125 TmpInstruction* prod = new TmpInstruction(mcfi, quot, divOpToUse);
2127 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
2128 .addReg(divOp1).addReg(divOpToUse).addRegDef(quot));
2130 mvec.push_back(BuildMI(ChooseMulInstructionByType(remI->getType()), 3)
2131 .addReg(quot).addReg(divOpToUse).addRegDef(prod));
2133 mvec.push_back(BuildMI(ChooseSubInstructionByType(remI->getType()), 3)
2134 .addReg(divOp1).addReg(prod).addRegDef(remI));
2139 case 38: // bool: And(bool, bool)
2140 case 138: // bool: And(bool, not)
2141 case 238: // bool: And(bool, boolconst)
2142 case 338: // reg : BAnd(reg, reg)
2143 case 538: // reg : BAnd(reg, Constant)
2144 Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
2147 case 438: // bool: BAnd(bool, bnot)
2148 { // Use the argument of NOT as the second argument!
2149 // Mark the NOT node so that no code is generated for it.
2150 // If the type is boolean, set 1 or 0 in the result register.
2151 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2152 Value* notArg = BinaryOperator::getNotArgument(
2153 cast<BinaryOperator>(notNode->getInstruction()));
2154 notNode->markFoldedIntoParent();
2155 Value *lhs = subtreeRoot->leftChild()->getValue();
2156 Value *dest = subtreeRoot->getValue();
2157 mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(lhs).addReg(notArg)
2158 .addReg(dest, MOTy::Def));
2160 if (notArg->getType() == Type::BoolTy)
2161 { // set 1 in result register if result of above is non-zero
2162 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2163 .addReg(dest, MOTy::UseAndDef));
2169 case 39: // bool: Or(bool, bool)
2170 case 139: // bool: Or(bool, not)
2171 case 239: // bool: Or(bool, boolconst)
2172 case 339: // reg : BOr(reg, reg)
2173 case 539: // reg : BOr(reg, Constant)
2174 Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
2177 case 439: // bool: BOr(bool, bnot)
2178 { // Use the argument of NOT as the second argument!
2179 // Mark the NOT node so that no code is generated for it.
2180 // If the type is boolean, set 1 or 0 in the result register.
2181 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2182 Value* notArg = BinaryOperator::getNotArgument(
2183 cast<BinaryOperator>(notNode->getInstruction()));
2184 notNode->markFoldedIntoParent();
2185 Value *lhs = subtreeRoot->leftChild()->getValue();
2186 Value *dest = subtreeRoot->getValue();
2188 mvec.push_back(BuildMI(V9::ORNr, 3).addReg(lhs).addReg(notArg)
2189 .addReg(dest, MOTy::Def));
2191 if (notArg->getType() == Type::BoolTy)
2192 { // set 1 in result register if result of above is non-zero
2193 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2194 .addReg(dest, MOTy::UseAndDef));
2200 case 40: // bool: Xor(bool, bool)
2201 case 140: // bool: Xor(bool, not)
2202 case 240: // bool: Xor(bool, boolconst)
2203 case 340: // reg : BXor(reg, reg)
2204 case 540: // reg : BXor(reg, Constant)
2205 Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
2208 case 440: // bool: BXor(bool, bnot)
2209 { // Use the argument of NOT as the second argument!
2210 // Mark the NOT node so that no code is generated for it.
2211 // If the type is boolean, set 1 or 0 in the result register.
2212 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2213 Value* notArg = BinaryOperator::getNotArgument(
2214 cast<BinaryOperator>(notNode->getInstruction()));
2215 notNode->markFoldedIntoParent();
2216 Value *lhs = subtreeRoot->leftChild()->getValue();
2217 Value *dest = subtreeRoot->getValue();
2218 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(lhs).addReg(notArg)
2219 .addReg(dest, MOTy::Def));
2221 if (notArg->getType() == Type::BoolTy)
2222 { // set 1 in result register if result of above is non-zero
2223 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2224 .addReg(dest, MOTy::UseAndDef));
2229 case 41: // setCCconst: SetCC(reg, Constant)
2230 { // Comparison is with a constant:
2232 // If the bool result must be computed into a register (see below),
2233 // and the constant is int ZERO, we can use the MOVR[op] instructions
2234 // and avoid the SUBcc instruction entirely.
2235 // Otherwise this is just the same as case 42, so just fall through.
2237 // The result of the SetCC must be computed and stored in a register if
2238 // it is used outside the current basic block (so it must be computed
2239 // as a boolreg) or it is used by anything other than a branch.
2240 // We will use a conditional move to do this.
2242 Instruction* setCCInstr = subtreeRoot->getInstruction();
2243 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
2244 ! AllUsesAreBranches(setCCInstr));
2248 InstrTreeNode* constNode = subtreeRoot->rightChild();
2250 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
2251 Constant *constVal = cast<Constant>(constNode->getValue());
2254 if ((constVal->getType()->isInteger()
2255 || isa<PointerType>(constVal->getType()))
2256 && target.getInstrInfo().ConvertConstantToIntType(target,
2257 constVal, constVal->getType(), isValidConst) == 0
2260 // That constant is an integer zero after all...
2261 // Use a MOVR[op] to compute the boolean result
2262 // Unconditionally set register to 0
2263 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0)
2264 .addRegDef(setCCInstr));
2266 // Now conditionally move 1 into the register.
2267 // Mark the register as a use (as well as a def) because the old
2268 // value will be retained if the condition is false.
2269 MachineOpCode movOpCode = ChooseMovpregiForSetCC(subtreeRoot);
2270 mvec.push_back(BuildMI(movOpCode, 3)
2271 .addReg(subtreeRoot->leftChild()->getValue())
2272 .addZImm(1).addReg(setCCInstr, MOTy::UseAndDef));
2277 // ELSE FALL THROUGH
2280 case 42: // bool: SetCC(reg, reg):
2282 // This generates a SUBCC instruction, putting the difference in a
2283 // result reg. if needed, and/or setting a condition code if needed.
2285 Instruction* setCCInstr = subtreeRoot->getInstruction();
2286 Value* leftVal = subtreeRoot->leftChild()->getValue();
2287 Value* rightVal = subtreeRoot->rightChild()->getValue();
2288 const Type* opType = leftVal->getType();
2289 bool isFPCompare = opType->isFloatingPoint();
2291 // If the boolean result of the SetCC is used outside the current basic
2292 // block (so it must be computed as a boolreg) or is used by anything
2293 // other than a branch, the boolean must be computed and stored
2294 // in a result register. We will use a conditional move to do this.
2296 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
2297 ! AllUsesAreBranches(setCCInstr));
2299 // A TmpInstruction is created to represent the CC "result".
2300 // Unlike other instances of TmpInstruction, this one is used
2301 // by machine code of multiple LLVM instructions, viz.,
2302 // the SetCC and the branch. Make sure to get the same one!
2303 // Note that we do this even for FP CC registers even though they
2304 // are explicit operands, because the type of the operand
2305 // needs to be a floating point condition code, not an integer
2306 // condition code. Think of this as casting the bool result to
2307 // a FP condition code register.
2308 // Later, we mark the 4th operand as being a CC register, and as a def.
2310 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
2311 setCCInstr->getParent()->getParent(),
2313 MachineCodeForInstruction::get(setCCInstr));
2315 // If the operands are signed values smaller than 4 bytes, then they
2316 // must be sign-extended in order to do a valid 32-bit comparison
2317 // and get the right result in the 32-bit CC register (%icc).
2319 Value* leftOpToUse = leftVal;
2320 Value* rightOpToUse = rightVal;
2321 if (opType->isIntegral() && opType->isSigned()) {
2322 unsigned opSize = target.getTargetData().getTypeSize(opType);
2324 MachineCodeForInstruction& mcfi =
2325 MachineCodeForInstruction::get(setCCInstr);
2327 // create temporary virtual regs. to hold the sign-extensions
2328 leftOpToUse = new TmpInstruction(mcfi, leftVal);
2329 rightOpToUse = new TmpInstruction(mcfi, rightVal);
2331 // sign-extend each operand and put the result in the temporary reg.
2332 target.getInstrInfo().CreateSignExtensionInstructions
2333 (target, setCCInstr->getParent()->getParent(),
2334 leftVal, leftOpToUse, 8*opSize, mvec, mcfi);
2335 target.getInstrInfo().CreateSignExtensionInstructions
2336 (target, setCCInstr->getParent()->getParent(),
2337 rightVal, rightOpToUse, 8*opSize, mvec, mcfi);
2341 if (! isFPCompare) {
2342 // Integer condition: set CC and discard result.
2343 mvec.push_back(BuildMI(V9::SUBccr, 4)
2344 .addReg(leftOpToUse)
2345 .addReg(rightOpToUse)
2346 .addMReg(target.getRegInfo().getZeroRegNum(),MOTy::Def)
2347 .addCCReg(tmpForCC, MOTy::Def));
2349 // FP condition: dest of FCMP should be some FCCn register
2350 mvec.push_back(BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
2351 .addCCReg(tmpForCC, MOTy::Def)
2352 .addReg(leftOpToUse)
2353 .addReg(rightOpToUse));
2356 if (computeBoolVal) {
2357 MachineOpCode movOpCode = (isFPCompare
2358 ? ChooseMovFpcciInstruction(subtreeRoot)
2359 : ChooseMovpcciForSetCC(subtreeRoot));
2361 // Unconditionally set register to 0
2362 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
2365 // Now conditionally move 1 into the register.
2366 // Mark the register as a use (as well as a def) because the old
2367 // value will be retained if the condition is false.
2368 M = (BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(1)
2369 .addReg(setCCInstr, MOTy::UseAndDef));
2375 case 51: // reg: Load(reg)
2376 case 52: // reg: Load(ptrreg)
2377 SetOperandsForMemInstr(ChooseLoadInstruction(
2378 subtreeRoot->getValue()->getType()),
2379 mvec, subtreeRoot, target);
2382 case 55: // reg: GetElemPtr(reg)
2383 case 56: // reg: GetElemPtrIdx(reg,reg)
2384 // If the GetElemPtr was folded into the user (parent), it will be
2385 // caught above. For other cases, we have to compute the address.
2386 SetOperandsForMemInstr(V9::ADDr, mvec, subtreeRoot, target);
2389 case 57: // reg: Alloca: Implement as 1 instruction:
2390 { // add %fp, offsetFromFP -> result
2391 AllocationInst* instr =
2392 cast<AllocationInst>(subtreeRoot->getInstruction());
2394 target.getTargetData().getTypeSize(instr->getAllocatedType());
2396 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
2400 case 58: // reg: Alloca(reg): Implement as 3 instructions:
2401 // mul num, typeSz -> tmp
2402 // sub %sp, tmp -> %sp
2403 { // add %sp, frameSizeBelowDynamicArea -> result
2404 AllocationInst* instr =
2405 cast<AllocationInst>(subtreeRoot->getInstruction());
2406 const Type* eltType = instr->getAllocatedType();
2408 // If #elements is constant, use simpler code for fixed-size allocas
2409 int tsize = (int) target.getTargetData().getTypeSize(eltType);
2410 Value* numElementsVal = NULL;
2411 bool isArray = instr->isArrayAllocation();
2413 if (!isArray || isa<Constant>(numElementsVal = instr->getArraySize())) {
2414 // total size is constant: generate code for fixed-size alloca
2415 unsigned numElements = isArray?
2416 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
2417 CreateCodeForFixedSizeAlloca(target, instr, tsize,
2420 // total size is not constant.
2421 CreateCodeForVariableSizeAlloca(target, instr, tsize,
2422 numElementsVal, mvec);
2427 case 61: // reg: Call
2428 { // Generate a direct (CALL) or indirect (JMPL) call.
2429 // Mark the return-address register, the indirection
2430 // register (for indirect calls), the operands of the Call,
2431 // and the return value (if any) as implicit operands
2432 // of the machine instruction.
2434 // If this is a varargs function, floating point arguments
2435 // have to passed in integer registers so insert
2436 // copy-float-to-int instructions for each float operand.
2438 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
2439 Value *callee = callInstr->getCalledValue();
2440 Function* calledFunc = dyn_cast<Function>(callee);
2442 // Check if this is an intrinsic function that needs a special code
2443 // sequence (e.g., va_start). Indirect calls cannot be special.
2445 bool specialIntrinsic = false;
2446 LLVMIntrinsic::ID iid;
2447 if (calledFunc && (iid=(LLVMIntrinsic::ID)calledFunc->getIntrinsicID()))
2448 specialIntrinsic = CodeGenIntrinsic(iid, *callInstr, target, mvec);
2450 // If not, generate the normal call sequence for the function.
2451 // This can also handle any intrinsics that are just function calls.
2453 if (! specialIntrinsic) {
2454 Function* currentFunc = callInstr->getParent()->getParent();
2455 MachineFunction& MF = MachineFunction::get(currentFunc);
2456 MachineCodeForInstruction& mcfi =
2457 MachineCodeForInstruction::get(callInstr);
2458 const UltraSparcRegInfo& regInfo =
2459 (UltraSparcRegInfo&) target.getRegInfo();
2460 const TargetFrameInfo& frameInfo = target.getFrameInfo();
2462 // Create hidden virtual register for return address with type void*
2463 TmpInstruction* retAddrReg =
2464 new TmpInstruction(mcfi, PointerType::get(Type::VoidTy), callInstr);
2466 // Generate the machine instruction and its operands.
2467 // Use CALL for direct function calls; this optimistically assumes
2468 // the PC-relative address fits in the CALL address field (22 bits).
2469 // Use JMPL for indirect calls.
2470 // This will be added to mvec later, after operand copies.
2472 MachineInstr* callMI;
2473 if (calledFunc) // direct function call
2474 callMI = BuildMI(V9::CALL, 1).addPCDisp(callee);
2475 else // indirect function call
2476 callMI = (BuildMI(V9::JMPLCALLi,3).addReg(callee)
2477 .addSImm((int64_t)0).addRegDef(retAddrReg));
2479 const FunctionType* funcType =
2480 cast<FunctionType>(cast<PointerType>(callee->getType())
2481 ->getElementType());
2482 bool isVarArgs = funcType->isVarArg();
2483 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
2485 // Use a descriptor to pass information about call arguments
2486 // to the register allocator. This descriptor will be "owned"
2487 // and freed automatically when the MachineCodeForInstruction
2488 // object for the callInstr goes away.
2489 CallArgsDescriptor* argDesc =
2490 new CallArgsDescriptor(callInstr, retAddrReg,isVarArgs,noPrototype);
2491 assert(callInstr->getOperand(0) == callee
2492 && "This is assumed in the loop below!");
2494 // Insert sign-extension instructions for small signed values,
2495 // if this is an unknown function (i.e., called via a funcptr)
2496 // or an external one (i.e., which may not be compiled by llc).
2498 if (calledFunc == NULL || calledFunc->isExternal()) {
2499 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2500 Value* argVal = callInstr->getOperand(i);
2501 const Type* argType = argVal->getType();
2502 if (argType->isIntegral() && argType->isSigned()) {
2503 unsigned argSize = target.getTargetData().getTypeSize(argType);
2505 // create a temporary virtual reg. to hold the sign-extension
2506 TmpInstruction* argExtend = new TmpInstruction(mcfi, argVal);
2508 // sign-extend argVal and put the result in the temporary reg.
2509 target.getInstrInfo().CreateSignExtensionInstructions
2510 (target, currentFunc, argVal, argExtend,
2511 8*argSize, mvec, mcfi);
2513 // replace argVal with argExtend in CallArgsDescriptor
2514 argDesc->getArgInfo(i-1).replaceArgVal(argExtend);
2520 // Insert copy instructions to get all the arguments into
2521 // all the places that they need to be.
2523 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2525 CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
2526 Value* argVal = argInfo.getArgVal(); // don't use callInstr arg here
2527 const Type* argType = argVal->getType();
2528 unsigned regType = regInfo.getRegTypeForDataType(argType);
2529 unsigned argSize = target.getTargetData().getTypeSize(argType);
2530 int regNumForArg = TargetRegInfo::getInvalidRegNum();
2531 unsigned regClassIDOfArgReg;
2533 // Check for FP arguments to varargs functions.
2534 // Any such argument in the first $K$ args must be passed in an
2535 // integer register. If there is no prototype, it must also
2536 // be passed as an FP register.
2537 // K = #integer argument registers.
2538 bool isFPArg = argVal->getType()->isFloatingPoint();
2539 if (isVarArgs && isFPArg) {
2542 // It is a function with no prototype: pass value
2543 // as an FP value as well as a varargs value. The FP value
2544 // may go in a register or on the stack. The copy instruction
2545 // to the outgoing reg/stack is created by the normal argument
2546 // handling code since this is the "normal" passing mode.
2548 regNumForArg = regInfo.regNumForFPArg(regType,
2549 false, false, argNo,
2550 regClassIDOfArgReg);
2551 if (regNumForArg == regInfo.getInvalidRegNum())
2552 argInfo.setUseStackSlot();
2554 argInfo.setUseFPArgReg();
2557 // If this arg. is in the first $K$ regs, add special copy-
2558 // float-to-int instructions to pass the value as an int.
2559 // To check if it is in the first $K$, get the register
2560 // number for the arg #i. These copy instructions are
2561 // generated here because they are extra cases and not needed
2562 // for the normal argument handling (some code reuse is
2563 // possible though -- later).
2565 int copyRegNum = regInfo.regNumForIntArg(false, false, argNo,
2566 regClassIDOfArgReg);
2567 if (copyRegNum != regInfo.getInvalidRegNum()) {
2568 // Create a virtual register to represent copyReg. Mark
2569 // this vreg as being an implicit operand of the call MI
2570 const Type* loadTy = (argType == Type::FloatTy
2571 ? Type::IntTy : Type::LongTy);
2572 TmpInstruction* argVReg = new TmpInstruction(mcfi, loadTy,
2575 callMI->addImplicitRef(argVReg);
2577 // Get a temp stack location to use to copy
2578 // float-to-int via the stack.
2580 // FIXME: For now, we allocate permanent space because
2581 // the stack frame manager does not allow locals to be
2582 // allocated (e.g., for alloca) after a temp is
2585 // int tmpOffset = MF.getInfo()->pushTempValue(argSize);
2586 int tmpOffset = MF.getInfo()->allocateLocalVar(argVReg);
2588 // Generate the store from FP reg to stack
2589 unsigned StoreOpcode = ChooseStoreInstruction(argType);
2590 M = BuildMI(convertOpcodeFromRegToImm(StoreOpcode), 3)
2591 .addReg(argVal).addMReg(regInfo.getFramePointer())
2592 .addSImm(tmpOffset);
2595 // Generate the load from stack to int arg reg
2596 unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
2597 M = BuildMI(convertOpcodeFromRegToImm(LoadOpcode), 3)
2598 .addMReg(regInfo.getFramePointer()).addSImm(tmpOffset)
2599 .addReg(argVReg, MOTy::Def);
2601 // Mark operand with register it should be assigned
2602 // both for copy and for the callMI
2603 M->SetRegForOperand(M->getNumOperands()-1, copyRegNum);
2604 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2608 // Add info about the argument to the CallArgsDescriptor
2609 argInfo.setUseIntArgReg();
2610 argInfo.setArgCopy(copyRegNum);
2612 // Cannot fit in first $K$ regs so pass arg on stack
2613 argInfo.setUseStackSlot();
2615 } else if (isFPArg) {
2616 // Get the outgoing arg reg to see if there is one.
2617 regNumForArg = regInfo.regNumForFPArg(regType, false, false,
2618 argNo, regClassIDOfArgReg);
2619 if (regNumForArg == regInfo.getInvalidRegNum())
2620 argInfo.setUseStackSlot();
2622 argInfo.setUseFPArgReg();
2623 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2627 // Get the outgoing arg reg to see if there is one.
2628 regNumForArg = regInfo.regNumForIntArg(false,false,
2629 argNo, regClassIDOfArgReg);
2630 if (regNumForArg == regInfo.getInvalidRegNum())
2631 argInfo.setUseStackSlot();
2633 argInfo.setUseIntArgReg();
2634 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2640 // Now insert copy instructions to stack slot or arg. register
2642 if (argInfo.usesStackSlot()) {
2643 // Get the stack offset for this argument slot.
2644 // FP args on stack are right justified so adjust offset!
2645 // int arguments are also right justified but they are
2646 // always loaded as a full double-word so the offset does
2647 // not need to be adjusted.
2648 int argOffset = frameInfo.getOutgoingArgOffset(MF, argNo);
2649 if (argType->isFloatingPoint()) {
2650 unsigned slotSize = frameInfo.getSizeOfEachArgOnStack();
2651 assert(argSize <= slotSize && "Insufficient slot size!");
2652 argOffset += slotSize - argSize;
2655 // Now generate instruction to copy argument to stack
2656 MachineOpCode storeOpCode =
2657 (argType->isFloatingPoint()
2658 ? ((argSize == 4)? V9::STFi : V9::STDFi) : V9::STXi);
2660 M = BuildMI(storeOpCode, 3).addReg(argVal)
2661 .addMReg(regInfo.getStackPointer()).addSImm(argOffset);
2664 else if (regNumForArg != regInfo.getInvalidRegNum()) {
2666 // Create a virtual register to represent the arg reg. Mark
2667 // this vreg as being an implicit operand of the call MI.
2668 TmpInstruction* argVReg =
2669 new TmpInstruction(mcfi, argVal, NULL, "argReg");
2671 callMI->addImplicitRef(argVReg);
2673 // Generate the reg-to-reg copy into the outgoing arg reg.
2674 // -- For FP values, create a FMOVS or FMOVD instruction
2675 // -- For non-FP values, create an add-with-0 instruction
2676 if (argType->isFloatingPoint())
2677 M=(BuildMI(argType==Type::FloatTy? V9::FMOVS :V9::FMOVD,2)
2678 .addReg(argVal).addReg(argVReg, MOTy::Def));
2680 M = (BuildMI(ChooseAddInstructionByType(argType), 3)
2681 .addReg(argVal).addSImm((int64_t) 0)
2682 .addReg(argVReg, MOTy::Def));
2684 // Mark the operand with the register it should be assigned
2685 M->SetRegForOperand(M->getNumOperands()-1, regNumForArg);
2686 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2692 assert(argInfo.getArgCopy() != regInfo.getInvalidRegNum() &&
2693 "Arg. not in stack slot, primary or secondary register?");
2696 // add call instruction and delay slot before copying return value
2697 mvec.push_back(callMI);
2698 mvec.push_back(BuildMI(V9::NOP, 0));
2700 // Add the return value as an implicit ref. The call operands
2701 // were added above. Also, add code to copy out the return value.
2702 // This is always register-to-register for int or FP return values.
2704 if (callInstr->getType() != Type::VoidTy) {
2705 // Get the return value reg.
2706 const Type* retType = callInstr->getType();
2708 int regNum = (retType->isFloatingPoint()
2709 ? (unsigned) SparcFloatRegClass::f0
2710 : (unsigned) SparcIntRegClass::o0);
2711 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
2712 regNum = regInfo.getUnifiedRegNum(regClassID, regNum);
2714 // Create a virtual register to represent it and mark
2715 // this vreg as being an implicit operand of the call MI
2716 TmpInstruction* retVReg =
2717 new TmpInstruction(mcfi, callInstr, NULL, "argReg");
2719 callMI->addImplicitRef(retVReg, /*isDef*/ true);
2721 // Generate the reg-to-reg copy from the return value reg.
2722 // -- For FP values, create a FMOVS or FMOVD instruction
2723 // -- For non-FP values, create an add-with-0 instruction
2724 if (retType->isFloatingPoint())
2725 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
2726 .addReg(retVReg).addReg(callInstr, MOTy::Def));
2728 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
2729 .addReg(retVReg).addSImm((int64_t) 0)
2730 .addReg(callInstr, MOTy::Def));
2732 // Mark the operand with the register it should be assigned
2733 // Also mark the implicit ref of the call defining this operand
2734 M->SetRegForOperand(0, regNum);
2735 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,regNum);
2740 // For the CALL instruction, the ret. addr. reg. is also implicit
2741 if (isa<Function>(callee))
2742 callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
2744 MF.getInfo()->popAllTempValues(); // free temps used for this inst
2750 case 62: // reg: Shl(reg, reg)
2752 Value* argVal1 = subtreeRoot->leftChild()->getValue();
2753 Value* argVal2 = subtreeRoot->rightChild()->getValue();
2754 Instruction* shlInstr = subtreeRoot->getInstruction();
2756 const Type* opType = argVal1->getType();
2757 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2758 "Shl unsupported for other types");
2760 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
2761 (opType == Type::LongTy)? V9::SLLXr6:V9::SLLr5,
2762 argVal1, argVal2, 0, shlInstr, mvec,
2763 MachineCodeForInstruction::get(shlInstr));
2767 case 63: // reg: Shr(reg, reg)
2769 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
2770 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2771 "Shr unsupported for other types");
2772 Add3OperandInstr(opType->isSigned()
2773 ? (opType == Type::LongTy ? V9::SRAXr6 : V9::SRAr5)
2774 : (opType == Type::ULongTy ? V9::SRLXr6 : V9::SRLr5),
2779 case 64: // reg: Phi(reg,reg)
2780 break; // don't forward the value
2782 case 65: // reg: VaArg(reg)
2784 // Use value initialized by va_start as pointer to args on the stack.
2785 // Load argument via current pointer value, then increment pointer.
2786 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
2787 Instruction* vaArgI = subtreeRoot->getInstruction();
2788 mvec.push_back(BuildMI(V9::LDXi, 3).addReg(vaArgI->getOperand(0)).
2789 addSImm(0).addRegDef(vaArgI));
2790 mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaArgI->getOperand(0)).
2791 addSImm(argSize).addRegDef(vaArgI->getOperand(0)));
2795 case 71: // reg: VReg
2796 case 72: // reg: Constant
2797 break; // don't forward the value
2800 assert(0 && "Unrecognized BURG rule");
2805 if (forwardOperandNum >= 0) {
2806 // We did not generate a machine instruction but need to use operand.
2807 // If user is in the same tree, replace Value in its machine operand.
2808 // If not, insert a copy instruction which should get coalesced away
2809 // by register allocation.
2810 if (subtreeRoot->parent() != NULL)
2811 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
2813 std::vector<MachineInstr*> minstrVec;
2814 Instruction* instr = subtreeRoot->getInstruction();
2815 target.getInstrInfo().
2816 CreateCopyInstructionsByType(target,
2817 instr->getParent()->getParent(),
2818 instr->getOperand(forwardOperandNum),
2820 MachineCodeForInstruction::get(instr));
2821 assert(minstrVec.size() > 0);
2822 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
2826 if (maskUnsignedResult) {
2827 // If result is unsigned and smaller than int reg size,
2828 // we need to clear high bits of result value.
2829 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
2830 Instruction* dest = subtreeRoot->getInstruction();
2831 if (dest->getType()->isUnsigned()) {
2832 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
2833 if (destSize <= 4) {
2834 // Mask high 64 - N bits, where N = 4*destSize.
2836 // Use a TmpInstruction to represent the
2837 // intermediate result before masking. Since those instructions
2838 // have already been generated, go back and substitute tmpI
2839 // for dest in the result position of each one of them.
2841 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(dest);
2842 TmpInstruction *tmpI = new TmpInstruction(mcfi, dest->getType(),
2843 dest, NULL, "maskHi");
2844 Value* srlArgToUse = tmpI;
2846 unsigned numSubst = 0;
2847 for (unsigned i=0, N=mvec.size(); i < N; ++i) {
2848 bool someArgsWereIgnored = false;
2849 numSubst += mvec[i]->substituteValue(dest, tmpI, /*defsOnly*/ true,
2850 /*defsAndUses*/ false,
2851 someArgsWereIgnored);
2852 assert(!someArgsWereIgnored &&
2853 "Operand `dest' exists but not replaced: probably bogus!");
2855 assert(numSubst > 0 && "Operand `dest' not replaced: probably bogus!");
2857 // Left shift 32-N if size (N) is less than 32 bits.
2858 // Use another tmp. virtual registe to represent this result.
2860 srlArgToUse = new TmpInstruction(mcfi, dest->getType(),
2861 tmpI, NULL, "maskHi2");
2862 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpI)
2863 .addZImm(8*(4-destSize))
2864 .addReg(srlArgToUse, MOTy::Def));
2867 // Logical right shift 32-N to get zero extension in top 64-N bits.
2868 mvec.push_back(BuildMI(V9::SRLi5, 3).addReg(srlArgToUse)
2869 .addZImm(8*(4-destSize)).addReg(dest, MOTy::Def));
2871 } else if (destSize < 8) {
2872 assert(0 && "Unsupported type size: 32 < size < 64 bits");