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 "Support/MathExtras.h"
28 static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
29 std::vector<MachineInstr*>& mvec) {
30 mvec.push_back(BuildMI(Opcode, 3).addReg(Node->leftChild()->getValue())
31 .addReg(Node->rightChild()->getValue())
32 .addRegDef(Node->getValue()));
37 //---------------------------------------------------------------------------
38 // Function: GetMemInstArgs
41 // Get the pointer value and the index vector for a memory operation
42 // (GetElementPtr, Load, or Store). If all indices of the given memory
43 // operation are constant, fold in constant indices in a chain of
44 // preceding GetElementPtr instructions (if any), and return the
45 // pointer value of the first instruction in the chain.
46 // All folded instructions are marked so no code is generated for them.
49 // Returns the pointer Value to use.
50 // Returns the resulting IndexVector in idxVec.
51 // Returns true/false in allConstantIndices if all indices are/aren't const.
52 //---------------------------------------------------------------------------
55 //---------------------------------------------------------------------------
56 // Function: FoldGetElemChain
59 // Fold a chain of GetElementPtr instructions containing only
60 // constant offsets into an equivalent (Pointer, IndexVector) pair.
61 // Returns the pointer Value, and stores the resulting IndexVector
62 // in argument chainIdxVec. This is a helper function for
63 // FoldConstantIndices that does the actual folding.
64 //---------------------------------------------------------------------------
67 // Check for a constant 0.
71 return (idx == ConstantSInt::getNullValue(idx->getType()));
75 FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
76 bool lastInstHasLeadingNonZero)
78 InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
79 GetElementPtrInst* gepInst =
80 dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
82 // ptr value is not computed in this tree or ptr value does not come from GEP
87 // Return NULL if we don't fold any instructions in.
90 // Now chase the chain of getElementInstr instructions, if any.
91 // Check for any non-constant indices and stop there.
92 // Also, stop if the first index of child is a non-zero array index
93 // and the last index of the current node is a non-array index:
94 // in that case, a non-array declared type is being accessed as an array
95 // which is not type-safe, but could be legal.
97 InstructionNode* ptrChild = gepNode;
98 while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
99 ptrChild->getOpLabel() == GetElemPtrIdx))
101 // Child is a GetElemPtr instruction
102 gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
103 User::op_iterator OI, firstIdx = gepInst->idx_begin();
104 User::op_iterator lastIdx = gepInst->idx_end();
105 bool allConstantOffsets = true;
107 // The first index of every GEP must be an array index.
108 assert((*firstIdx)->getType() == Type::LongTy &&
109 "INTERNAL ERROR: Structure index for a pointer type!");
111 // If the last instruction had a leading non-zero index, check if the
112 // current one references a sequential (i.e., indexable) type.
113 // If not, the code is not type-safe and we would create an illegal GEP
114 // by folding them, so don't fold any more instructions.
116 if (lastInstHasLeadingNonZero)
117 if (! isa<SequentialType>(gepInst->getType()->getElementType()))
118 break; // cannot fold in any preceding getElementPtr instrs.
120 // Check that all offsets are constant for this instruction
121 for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
122 allConstantOffsets = isa<ConstantInt>(*OI);
124 if (allConstantOffsets) {
125 // Get pointer value out of ptrChild.
126 ptrVal = gepInst->getPointerOperand();
128 // Remember if it has leading zero index: it will be discarded later.
129 lastInstHasLeadingNonZero = ! IsZero(*firstIdx);
131 // Insert its index vector at the start, skipping any leading [0]
132 chainIdxVec.insert(chainIdxVec.begin(),
133 firstIdx + !lastInstHasLeadingNonZero, lastIdx);
135 // Mark the folded node so no code is generated for it.
136 ((InstructionNode*) ptrChild)->markFoldedIntoParent();
138 // Get the previous GEP instruction and continue trying to fold
139 ptrChild = dyn_cast<InstructionNode>(ptrChild->leftChild());
140 } else // cannot fold this getElementPtr instr. or any preceding ones
144 // If the first getElementPtr instruction had a leading [0], add it back.
145 // Note that this instruction is the *last* one successfully folded above.
146 if (ptrVal && ! lastInstHasLeadingNonZero)
147 chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
153 //---------------------------------------------------------------------------
154 // Function: GetGEPInstArgs
157 // Helper function for GetMemInstArgs that handles the final getElementPtr
158 // instruction used by (or same as) the memory operation.
159 // Extracts the indices of the current instruction and tries to fold in
160 // preceding ones if all indices of the current one are constant.
161 //---------------------------------------------------------------------------
164 GetGEPInstArgs(InstructionNode* gepNode,
165 std::vector<Value*>& idxVec,
166 bool& allConstantIndices)
168 allConstantIndices = true;
169 GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
171 // Default pointer is the one from the current instruction.
172 Value* ptrVal = gepI->getPointerOperand();
173 InstrTreeNode* ptrChild = gepNode->leftChild();
175 // Extract the index vector of the GEP instructin.
176 // If all indices are constant and first index is zero, try to fold
177 // in preceding GEPs with all constant indices.
178 for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
179 allConstantIndices && OI != OE; ++OI)
180 if (! isa<Constant>(*OI))
181 allConstantIndices = false; // note: this also terminates loop!
183 // If we have only constant indices, fold chains of constant indices
184 // in this and any preceding GetElemPtr instructions.
185 bool foldedGEPs = false;
186 bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
187 if (allConstantIndices)
188 if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx)) {
193 // Append the index vector of the current instruction.
194 // Skip the leading [0] index if preceding GEPs were folded into this.
195 idxVec.insert(idxVec.end(),
196 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
202 //---------------------------------------------------------------------------
203 // Function: GetMemInstArgs
206 // Get the pointer value and the index vector for a memory operation
207 // (GetElementPtr, Load, or Store). If all indices of the given memory
208 // operation are constant, fold in constant indices in a chain of
209 // preceding GetElementPtr instructions (if any), and return the
210 // pointer value of the first instruction in the chain.
211 // All folded instructions are marked so no code is generated for them.
214 // Returns the pointer Value to use.
215 // Returns the resulting IndexVector in idxVec.
216 // Returns true/false in allConstantIndices if all indices are/aren't const.
217 //---------------------------------------------------------------------------
220 GetMemInstArgs(InstructionNode* memInstrNode,
221 std::vector<Value*>& idxVec,
222 bool& allConstantIndices)
224 allConstantIndices = false;
225 Instruction* memInst = memInstrNode->getInstruction();
226 assert(idxVec.size() == 0 && "Need empty vector to return indices");
228 // If there is a GetElemPtr instruction to fold in to this instr,
229 // it must be in the left child for Load and GetElemPtr, and in the
230 // right child for Store instructions.
231 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
232 ? memInstrNode->rightChild()
233 : memInstrNode->leftChild());
235 // Default pointer is the one from the current instruction.
236 Value* ptrVal = ptrChild->getValue();
238 // Find the "last" GetElemPtr instruction: this one or the immediate child.
239 // There will be none if this is a load or a store from a scalar pointer.
240 InstructionNode* gepNode = NULL;
241 if (isa<GetElementPtrInst>(memInst))
242 gepNode = memInstrNode;
243 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
244 // Child of load/store is a GEP and memInst is its only use.
245 // Use its indices and mark it as folded.
246 gepNode = cast<InstructionNode>(ptrChild);
247 gepNode->markFoldedIntoParent();
250 // If there are no indices, return the current pointer.
251 // Else extract the pointer from the GEP and fold the indices.
252 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
257 //************************ Internal Functions ******************************/
260 static inline MachineOpCode
261 ChooseBprInstruction(const InstructionNode* instrNode)
263 MachineOpCode opCode;
265 Instruction* setCCInstr =
266 ((InstructionNode*) instrNode->leftChild())->getInstruction();
268 switch(setCCInstr->getOpcode())
270 case Instruction::SetEQ: opCode = V9::BRZ; break;
271 case Instruction::SetNE: opCode = V9::BRNZ; break;
272 case Instruction::SetLE: opCode = V9::BRLEZ; break;
273 case Instruction::SetGE: opCode = V9::BRGEZ; break;
274 case Instruction::SetLT: opCode = V9::BRLZ; break;
275 case Instruction::SetGT: opCode = V9::BRGZ; break;
277 assert(0 && "Unrecognized VM instruction!");
278 opCode = V9::INVALID_OPCODE;
286 static inline MachineOpCode
287 ChooseBpccInstruction(const InstructionNode* instrNode,
288 const BinaryOperator* setCCInstr)
290 MachineOpCode opCode = V9::INVALID_OPCODE;
292 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
295 switch(setCCInstr->getOpcode())
297 case Instruction::SetEQ: opCode = V9::BE; break;
298 case Instruction::SetNE: opCode = V9::BNE; break;
299 case Instruction::SetLE: opCode = V9::BLE; break;
300 case Instruction::SetGE: opCode = V9::BGE; break;
301 case Instruction::SetLT: opCode = V9::BL; break;
302 case Instruction::SetGT: opCode = V9::BG; break;
304 assert(0 && "Unrecognized VM instruction!");
308 switch(setCCInstr->getOpcode())
310 case Instruction::SetEQ: opCode = V9::BE; break;
311 case Instruction::SetNE: opCode = V9::BNE; break;
312 case Instruction::SetLE: opCode = V9::BLEU; break;
313 case Instruction::SetGE: opCode = V9::BCC; break;
314 case Instruction::SetLT: opCode = V9::BCS; break;
315 case Instruction::SetGT: opCode = V9::BGU; break;
317 assert(0 && "Unrecognized VM instruction!");
325 static inline MachineOpCode
326 ChooseBFpccInstruction(const InstructionNode* instrNode,
327 const BinaryOperator* setCCInstr)
329 MachineOpCode opCode = V9::INVALID_OPCODE;
331 switch(setCCInstr->getOpcode())
333 case Instruction::SetEQ: opCode = V9::FBE; break;
334 case Instruction::SetNE: opCode = V9::FBNE; break;
335 case Instruction::SetLE: opCode = V9::FBLE; break;
336 case Instruction::SetGE: opCode = V9::FBGE; break;
337 case Instruction::SetLT: opCode = V9::FBL; break;
338 case Instruction::SetGT: opCode = V9::FBG; break;
340 assert(0 && "Unrecognized VM instruction!");
348 // Create a unique TmpInstruction for a boolean value,
349 // representing the CC register used by a branch on that value.
350 // For now, hack this using a little static cache of TmpInstructions.
351 // Eventually the entire BURG instruction selection should be put
352 // into a separate class that can hold such information.
353 // The static cache is not too bad because the memory for these
354 // TmpInstructions will be freed along with the rest of the Function anyway.
356 static TmpInstruction*
357 GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType)
359 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
360 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
361 static const Function *lastFunction = 0;// Use to flush cache between funcs
363 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
365 if (lastFunction != F) {
367 boolToTmpCache.clear();
370 // Look for tmpI and create a new one otherwise. The new value is
371 // directly written to map using the ref returned by operator[].
372 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
374 tmpI = new TmpInstruction(ccType, boolVal);
380 static inline MachineOpCode
381 ChooseBccInstruction(const InstructionNode* instrNode,
384 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
385 assert(setCCNode->getOpLabel() == SetCCOp);
386 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
387 const Type* setCCType = setCCInstr->getOperand(0)->getType();
389 isFPBranch = setCCType->isFloatingPoint(); // Return value: don't delete!
392 return ChooseBFpccInstruction(instrNode, setCCInstr);
394 return ChooseBpccInstruction(instrNode, setCCInstr);
398 static inline MachineOpCode
399 ChooseMovFpccInstruction(const InstructionNode* instrNode)
401 MachineOpCode opCode = V9::INVALID_OPCODE;
403 switch(instrNode->getInstruction()->getOpcode())
405 case Instruction::SetEQ: opCode = V9::MOVFE; break;
406 case Instruction::SetNE: opCode = V9::MOVFNE; break;
407 case Instruction::SetLE: opCode = V9::MOVFLE; break;
408 case Instruction::SetGE: opCode = V9::MOVFGE; break;
409 case Instruction::SetLT: opCode = V9::MOVFL; break;
410 case Instruction::SetGT: opCode = V9::MOVFG; break;
412 assert(0 && "Unrecognized VM instruction!");
420 // Assumes that SUBcc v1, v2 -> v3 has been executed.
421 // In most cases, we want to clear v3 and then follow it by instruction
423 // Set mustClearReg=false if v3 need not be cleared before conditional move.
424 // Set valueToMove=0 if we want to conditionally move 0 instead of 1
425 // (i.e., we want to test inverse of a condition)
426 // (The latter two cases do not seem to arise because SetNE needs nothing.)
429 ChooseMovpccAfterSub(const InstructionNode* instrNode,
433 MachineOpCode opCode = V9::INVALID_OPCODE;
437 switch(instrNode->getInstruction()->getOpcode())
439 case Instruction::SetEQ: opCode = V9::MOVE; break;
440 case Instruction::SetLE: opCode = V9::MOVLE; break;
441 case Instruction::SetGE: opCode = V9::MOVGE; break;
442 case Instruction::SetLT: opCode = V9::MOVL; break;
443 case Instruction::SetGT: opCode = V9::MOVG; break;
444 case Instruction::SetNE: assert(0 && "No move required!"); break;
445 default: assert(0 && "Unrecognized VM instr!"); break;
451 static inline MachineOpCode
452 ChooseConvertToFloatInstr(OpLabel vopCode, const Type* opType)
454 MachineOpCode opCode = V9::INVALID_OPCODE;
459 if (opType == Type::SByteTy || opType == Type::ShortTy ||
460 opType == Type::IntTy)
462 else if (opType == Type::LongTy)
464 else if (opType == Type::DoubleTy)
466 else if (opType == Type::FloatTy)
469 assert(0 && "Cannot convert this type to FLOAT on SPARC");
473 // This is usually used in conjunction with CreateCodeToCopyIntToFloat().
474 // Both functions should treat the integer as a 32-bit value for types
475 // of 4 bytes or less, and as a 64-bit value otherwise.
476 if (opType == Type::SByteTy || opType == Type::UByteTy ||
477 opType == Type::ShortTy || opType == Type::UShortTy ||
478 opType == Type::IntTy || opType == Type::UIntTy)
480 else if (opType == Type::LongTy || opType == Type::ULongTy)
482 else if (opType == Type::FloatTy)
484 else if (opType == Type::DoubleTy)
487 assert(0 && "Cannot convert this type to DOUBLE on SPARC");
497 static inline MachineOpCode
498 ChooseConvertFPToIntInstr(Type::PrimitiveID tid, const Type* opType)
500 MachineOpCode opCode = V9::INVALID_OPCODE;;
502 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
503 && "This function should only be called for FLOAT or DOUBLE");
505 if (tid == Type::UIntTyID) {
506 assert(tid != Type::UIntTyID && "FP-to-uint conversions must be expanded"
507 " into FP->long->uint for SPARC v9: SO RUN PRESELECTION PASS!");
508 } else if (tid == Type::SByteTyID || tid == Type::ShortTyID ||
509 tid == Type::IntTyID || tid == Type::UByteTyID ||
510 tid == Type::UShortTyID) {
511 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
512 } else if (tid == Type::LongTyID || tid == Type::ULongTyID) {
513 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
515 assert(0 && "Should not get here, Mo!");
521 CreateConvertFPToIntInstr(Type::PrimitiveID destTID,
522 Value* srcVal, Value* destVal)
524 MachineOpCode opCode = ChooseConvertFPToIntInstr(destTID, srcVal->getType());
525 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
526 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
529 // CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer
530 // The FP value must be converted to the dest type in an FP register,
531 // and the result is then copied from FP to int register via memory.
533 // Since fdtoi converts to signed integers, any FP value V between MAXINT+1
534 // and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly
535 // *only* when converting to an unsigned. (Unsigned byte, short or long
536 // don't have this problem.)
537 // For unsigned int, we therefore have to generate the code sequence:
539 // if (V > (float) MAXINT) {
540 // unsigned result = (unsigned) (V - (float) MAXINT);
541 // result = result + (unsigned) MAXINT;
544 // result = (unsigned) V;
547 CreateCodeToConvertFloatToInt(const TargetMachine& target,
550 std::vector<MachineInstr*>& mvec,
551 MachineCodeForInstruction& mcfi)
553 // Create a temporary to represent the FP register into which the
554 // int value will placed after conversion. The type of this temporary
555 // depends on the type of FP register to use: single-prec for a 32-bit
556 // int or smaller; double-prec for a 64-bit int.
558 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
559 const Type* destTypeToUse = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
560 TmpInstruction* destForCast = new TmpInstruction(destTypeToUse, opVal);
561 mcfi.addTemp(destForCast);
563 // Create the fp-to-int conversion code
564 MachineInstr* M =CreateConvertFPToIntInstr(destI->getType()->getPrimitiveID(),
568 // Create the fpreg-to-intreg copy code
569 target.getInstrInfo().
570 CreateCodeToCopyFloatToInt(target, destI->getParent()->getParent(),
571 destForCast, destI, mvec, mcfi);
575 static inline MachineOpCode
576 ChooseAddInstruction(const InstructionNode* instrNode)
578 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
582 static inline MachineInstr*
583 CreateMovFloatInstruction(const InstructionNode* instrNode,
584 const Type* resultType)
586 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
587 .addReg(instrNode->leftChild()->getValue())
588 .addRegDef(instrNode->getValue());
591 static inline MachineInstr*
592 CreateAddConstInstruction(const InstructionNode* instrNode)
594 MachineInstr* minstr = NULL;
596 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
597 assert(isa<Constant>(constOp));
599 // Cases worth optimizing are:
600 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
601 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
603 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
604 double dval = FPC->getValue();
606 minstr = CreateMovFloatInstruction(instrNode,
607 instrNode->getInstruction()->getType());
614 static inline MachineOpCode
615 ChooseSubInstructionByType(const Type* resultType)
617 MachineOpCode opCode = V9::INVALID_OPCODE;
619 if (resultType->isInteger() || isa<PointerType>(resultType)) {
622 switch(resultType->getPrimitiveID())
624 case Type::FloatTyID: opCode = V9::FSUBS; break;
625 case Type::DoubleTyID: opCode = V9::FSUBD; break;
626 default: assert(0 && "Invalid type for SUB instruction"); break;
634 static inline MachineInstr*
635 CreateSubConstInstruction(const InstructionNode* instrNode)
637 MachineInstr* minstr = NULL;
639 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
640 assert(isa<Constant>(constOp));
642 // Cases worth optimizing are:
643 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
644 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
646 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
647 double dval = FPC->getValue();
649 minstr = CreateMovFloatInstruction(instrNode,
650 instrNode->getInstruction()->getType());
657 static inline MachineOpCode
658 ChooseFcmpInstruction(const InstructionNode* instrNode)
660 MachineOpCode opCode = V9::INVALID_OPCODE;
662 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
663 switch(operand->getType()->getPrimitiveID()) {
664 case Type::FloatTyID: opCode = V9::FCMPS; break;
665 case Type::DoubleTyID: opCode = V9::FCMPD; break;
666 default: assert(0 && "Invalid type for FCMP instruction"); break;
673 // Assumes that leftArg and rightArg are both cast instructions.
676 BothFloatToDouble(const InstructionNode* instrNode)
678 InstrTreeNode* leftArg = instrNode->leftChild();
679 InstrTreeNode* rightArg = instrNode->rightChild();
680 InstrTreeNode* leftArgArg = leftArg->leftChild();
681 InstrTreeNode* rightArgArg = rightArg->leftChild();
682 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
684 // Check if both arguments are floats cast to double
685 return (leftArg->getValue()->getType() == Type::DoubleTy &&
686 leftArgArg->getValue()->getType() == Type::FloatTy &&
687 rightArgArg->getValue()->getType() == Type::FloatTy);
691 static inline MachineOpCode
692 ChooseMulInstructionByType(const Type* resultType)
694 MachineOpCode opCode = V9::INVALID_OPCODE;
696 if (resultType->isInteger())
699 switch(resultType->getPrimitiveID())
701 case Type::FloatTyID: opCode = V9::FMULS; break;
702 case Type::DoubleTyID: opCode = V9::FMULD; break;
703 default: assert(0 && "Invalid type for MUL instruction"); break;
711 static inline MachineInstr*
712 CreateIntNegInstruction(const TargetMachine& target,
715 return BuildMI(V9::SUB, 3).addMReg(target.getRegInfo().getZeroRegNum())
716 .addReg(vreg).addRegDef(vreg);
720 // Create instruction sequence for any shift operation.
721 // SLL or SLLX on an operand smaller than the integer reg. size (64bits)
722 // requires a second instruction for explicit sign-extension.
723 // Note that we only have to worry about a sign-bit appearing in the
724 // most significant bit of the operand after shifting (e.g., bit 32 of
725 // Int or bit 16 of Short), so we do not have to worry about results
726 // that are as large as a normal integer register.
729 CreateShiftInstructions(const TargetMachine& target,
731 MachineOpCode shiftOpCode,
733 Value* optArgVal2, /* Use optArgVal2 if not NULL */
734 unsigned optShiftNum, /* else use optShiftNum */
735 Instruction* destVal,
736 std::vector<MachineInstr*>& mvec,
737 MachineCodeForInstruction& mcfi)
739 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
740 "Large shift sizes unexpected, but can be handled below: "
741 "You need to check whether or not it fits in immed field below");
743 // If this is a logical left shift of a type smaller than the standard
744 // integer reg. size, we have to extend the sign-bit into upper bits
745 // of dest, so we need to put the result of the SLL into a temporary.
747 Value* shiftDest = destVal;
748 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
749 if ((shiftOpCode == V9::SLL || shiftOpCode == V9::SLLX) && opSize < 8)
750 { // put SLL result into a temporary
751 shiftDest = new TmpInstruction(argVal1, optArgVal2, "sllTmp");
752 mcfi.addTemp(shiftDest);
755 MachineInstr* M = (optArgVal2 != NULL)
756 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
757 .addReg(shiftDest, MOTy::Def)
758 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
759 .addReg(shiftDest, MOTy::Def);
762 if (shiftDest != destVal)
763 { // extend the sign-bit of the result into all upper bits of dest
764 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
765 target.getInstrInfo().
766 CreateSignExtensionInstructions(target, F, shiftDest, destVal,
767 8*opSize, mvec, mcfi);
772 // Does not create any instructions if we cannot exploit constant to
773 // create a cheaper instruction.
774 // This returns the approximate cost of the instructions generated,
775 // which is used to pick the cheapest when both operands are constant.
776 static inline unsigned
777 CreateMulConstInstruction(const TargetMachine &target, Function* F,
778 Value* lval, Value* rval, Instruction* destVal,
779 std::vector<MachineInstr*>& mvec,
780 MachineCodeForInstruction& mcfi)
782 /* Use max. multiply cost, viz., cost of MULX */
783 unsigned cost = target.getInstrInfo().minLatency(V9::MULX);
784 unsigned firstNewInstr = mvec.size();
786 Value* constOp = rval;
787 if (! isa<Constant>(constOp))
790 // Cases worth optimizing are:
791 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
792 // (2) Multiply by 2^x for integer types: replace with Shift
794 const Type* resultType = destVal->getType();
796 if (resultType->isInteger() || isa<PointerType>(resultType)) {
798 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
801 bool needNeg = false;
807 if (C == 0 || C == 1) {
808 cost = target.getInstrInfo().minLatency(V9::ADD);
809 unsigned Zero = target.getRegInfo().getZeroRegNum();
812 M = BuildMI(V9::ADD,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
814 M = BuildMI(V9::ADD,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
817 else if (isPowerOf2(C, pow)) {
818 unsigned opSize = target.getTargetData().getTypeSize(resultType);
819 MachineOpCode opCode = (opSize <= 32)? V9::SLL : V9::SLLX;
820 CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
821 destVal, mvec, mcfi);
824 if (mvec.size() > 0 && needNeg)
825 { // insert <reg = SUB 0, reg> after the instr to flip the sign
826 MachineInstr* M = CreateIntNegInstruction(target, destVal);
831 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
832 double dval = FPC->getValue();
833 if (fabs(dval) == 1) {
834 MachineOpCode opCode = (dval < 0)
835 ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
836 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
837 mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
842 if (firstNewInstr < mvec.size()) {
844 for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
845 cost += target.getInstrInfo().minLatency(mvec[i]->getOpCode());
852 // Does not create any instructions if we cannot exploit constant to
853 // create a cheaper instruction.
856 CreateCheapestMulConstInstruction(const TargetMachine &target,
858 Value* lval, Value* rval,
859 Instruction* destVal,
860 std::vector<MachineInstr*>& mvec,
861 MachineCodeForInstruction& mcfi)
864 if (isa<Constant>(lval) && isa<Constant>(rval))
865 { // both operands are constant: evaluate and "set" in dest
866 Constant* P = ConstantFoldBinaryInstruction(Instruction::Mul,
867 cast<Constant>(lval), cast<Constant>(rval));
868 target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi);
870 else if (isa<Constant>(rval)) // rval is constant, but not lval
871 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
872 else if (isa<Constant>(lval)) // lval is constant, but not rval
873 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
875 // else neither is constant
879 // Return NULL if we cannot exploit constant to create a cheaper instruction
881 CreateMulInstruction(const TargetMachine &target, Function* F,
882 Value* lval, Value* rval, Instruction* destVal,
883 std::vector<MachineInstr*>& mvec,
884 MachineCodeForInstruction& mcfi,
885 MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE)
887 unsigned L = mvec.size();
888 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
889 if (mvec.size() == L) {
890 // no instructions were added so create MUL reg, reg, reg.
891 // Use FSMULD if both operands are actually floats cast to doubles.
892 // Otherwise, use the default opcode for the appropriate type.
893 MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE)
895 : ChooseMulInstructionByType(destVal->getType()));
896 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
897 .addRegDef(destVal));
902 // Generate a divide instruction for Div or Rem.
903 // For Rem, this assumes that the operand type will be signed if the result
904 // type is signed. This is correct because they must have the same sign.
906 static inline MachineOpCode
907 ChooseDivInstruction(TargetMachine &target,
908 const InstructionNode* instrNode)
910 MachineOpCode opCode = V9::INVALID_OPCODE;
912 const Type* resultType = instrNode->getInstruction()->getType();
914 if (resultType->isInteger())
915 opCode = resultType->isSigned()? V9::SDIVX : V9::UDIVX;
917 switch(resultType->getPrimitiveID())
919 case Type::FloatTyID: opCode = V9::FDIVS; break;
920 case Type::DoubleTyID: opCode = V9::FDIVD; break;
921 default: assert(0 && "Invalid type for DIV instruction"); break;
928 // Return if we cannot exploit constant to create a cheaper instruction
930 CreateDivConstInstruction(TargetMachine &target,
931 const InstructionNode* instrNode,
932 std::vector<MachineInstr*>& mvec)
934 Value* LHS = instrNode->leftChild()->getValue();
935 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
936 if (!isa<Constant>(constOp))
939 Value* DestVal = instrNode->getValue();
940 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
942 // Cases worth optimizing are:
943 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
944 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
946 const Type* resultType = instrNode->getInstruction()->getType();
948 if (resultType->isInteger())
952 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
954 bool needNeg = false;
961 mvec.push_back(BuildMI(V9::ADD, 3).addReg(LHS).addMReg(ZeroReg)
962 .addRegDef(DestVal));
963 } else if (isPowerOf2(C, pow)) {
964 unsigned opCode= ((resultType->isSigned())
965 ? (resultType==Type::LongTy) ? V9::SRAX : V9::SRA
966 : (resultType==Type::LongTy) ? V9::SRLX : V9::SRL);
967 mvec.push_back(BuildMI(opCode, 3).addReg(LHS).addZImm(pow)
968 .addRegDef(DestVal));
971 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
972 // insert <reg = SUB 0, reg> after the instr to flip the sign
973 mvec.push_back(CreateIntNegInstruction(target, DestVal));
977 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
978 double dval = FPC->getValue();
979 if (fabs(dval) == 1) {
981 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
982 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
984 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(DestVal));
992 CreateCodeForVariableSizeAlloca(const TargetMachine& target,
995 Value* numElementsVal,
996 std::vector<MachineInstr*>& getMvec)
1000 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
1001 Function *F = result->getParent()->getParent();
1003 // Enforce the alignment constraints on the stack pointer at
1004 // compile time if the total size is a known constant.
1005 if (isa<Constant>(numElementsVal))
1008 int64_t numElem = GetConstantValueAsSignedInt(numElementsVal, isValid);
1009 assert(isValid && "Unexpectedly large array dimension in alloca!");
1010 int64_t total = numElem * tsize;
1011 if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
1012 total += target.getFrameInfo().getStackFrameSizeAlignment() - extra;
1013 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
1017 // The size is not a constant. Generate code to compute it and
1018 // code to pad the size for stack alignment.
1019 // Create a Value to hold the (constant) element size
1020 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
1022 // Create temporary values to hold the result of MUL, SLL, SRL
1023 // THIS CASE IS INCOMPLETE AND WILL BE FIXED SHORTLY.
1024 TmpInstruction* tmpProd = new TmpInstruction(numElementsVal, tsizeVal);
1025 TmpInstruction* tmpSLL = new TmpInstruction(numElementsVal, tmpProd);
1026 TmpInstruction* tmpSRL = new TmpInstruction(numElementsVal, tmpSLL);
1027 mcfi.addTemp(tmpProd);
1028 mcfi.addTemp(tmpSLL);
1029 mcfi.addTemp(tmpSRL);
1031 // Instruction 1: mul numElements, typeSize -> tmpProd
1032 // This will optimize the MUL as far as possible.
1033 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd,getMvec,
1034 mcfi, INVALID_MACHINE_OPCODE);
1036 assert(0 && "Need to insert padding instructions here!");
1038 totalSizeVal = tmpProd;
1041 // Get the constant offset from SP for dynamically allocated storage
1042 // and create a temporary Value to hold it.
1043 MachineFunction& mcInfo = MachineFunction::get(F);
1045 ConstantSInt* dynamicAreaOffset =
1046 ConstantSInt::get(Type::IntTy,
1047 target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp));
1048 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
1050 unsigned SPReg = target.getRegInfo().getStackPointer();
1052 // Instruction 2: sub %sp, totalSizeVal -> %sp
1053 getMvec.push_back(BuildMI(V9::SUB, 3).addMReg(SPReg).addReg(totalSizeVal)
1054 .addMReg(SPReg,MOTy::Def));
1056 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
1057 getMvec.push_back(BuildMI(V9::ADD, 3).addMReg(SPReg).addReg(dynamicAreaOffset)
1058 .addRegDef(result));
1063 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
1064 Instruction* result,
1066 unsigned numElements,
1067 std::vector<MachineInstr*>& getMvec)
1069 assert(tsize > 0 && "Illegal (zero) type size for alloca");
1070 assert(result && result->getParent() &&
1071 "Result value is not part of a function?");
1072 Function *F = result->getParent()->getParent();
1073 MachineFunction &mcInfo = MachineFunction::get(F);
1075 // Check if the offset would small enough to use as an immediate in
1076 // load/stores (check LDX because all load/stores have the same-size immediate
1077 // field). If not, put the variable in the dynamically sized area of the
1079 unsigned paddedSizeIgnored;
1080 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
1082 tsize * numElements);
1083 if (! target.getInstrInfo().constantFitsInImmedField(V9::LDX, offsetFromFP)) {
1084 CreateCodeForVariableSizeAlloca(target, result, tsize,
1085 ConstantSInt::get(Type::IntTy,numElements),
1090 // else offset fits in immediate field so go ahead and allocate it.
1091 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
1093 // Create a temporary Value to hold the constant offset.
1094 // This is needed because it may not fit in the immediate field.
1095 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
1097 // Instruction 1: add %fp, offsetFromFP -> result
1098 unsigned FPReg = target.getRegInfo().getFramePointer();
1099 getMvec.push_back(BuildMI(V9::ADD, 3).addMReg(FPReg).addReg(offsetVal)
1100 .addRegDef(result));
1104 //------------------------------------------------------------------------
1105 // Function SetOperandsForMemInstr
1107 // Choose addressing mode for the given load or store instruction.
1108 // Use [reg+reg] if it is an indexed reference, and the index offset is
1109 // not a constant or if it cannot fit in the offset field.
1110 // Use [reg+offset] in all other cases.
1112 // This assumes that all array refs are "lowered" to one of these forms:
1113 // %x = load (subarray*) ptr, constant ; single constant offset
1114 // %x = load (subarray*) ptr, offsetVal ; single non-constant offset
1115 // Generally, this should happen via strength reduction + LICM.
1116 // Also, strength reduction should take care of using the same register for
1117 // the loop index variable and an array index, when that is profitable.
1118 //------------------------------------------------------------------------
1121 SetOperandsForMemInstr(unsigned Opcode,
1122 std::vector<MachineInstr*>& mvec,
1123 InstructionNode* vmInstrNode,
1124 const TargetMachine& target)
1126 Instruction* memInst = vmInstrNode->getInstruction();
1127 // Index vector, ptr value, and flag if all indices are const.
1128 std::vector<Value*> idxVec;
1129 bool allConstantIndices;
1130 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
1132 // Now create the appropriate operands for the machine instruction.
1133 // First, initialize so we default to storing the offset in a register.
1134 int64_t smallConstOffset = 0;
1135 Value* valueForRegOffset = NULL;
1136 MachineOperand::MachineOperandType offsetOpType =
1137 MachineOperand::MO_VirtualRegister;
1139 // Check if there is an index vector and if so, compute the
1140 // right offset for structures and for arrays
1142 if (!idxVec.empty())
1144 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
1146 // If all indices are constant, compute the combined offset directly.
1147 if (allConstantIndices)
1149 // Compute the offset value using the index vector. Create a
1150 // virtual reg. for it since it may not fit in the immed field.
1151 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
1152 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
1156 // There is at least one non-constant offset. Therefore, this must
1157 // be an array ref, and must have been lowered to a single non-zero
1158 // offset. (An extra leading zero offset, if any, can be ignored.)
1159 // Generate code sequence to compute address from index.
1161 bool firstIdxIsZero = IsZero(idxVec[0]);
1162 assert(idxVec.size() == 1U + firstIdxIsZero
1163 && "Array refs must be lowered before Instruction Selection");
1165 Value* idxVal = idxVec[firstIdxIsZero];
1167 std::vector<MachineInstr*> mulVec;
1168 Instruction* addr = new TmpInstruction(Type::ULongTy, memInst);
1169 MachineCodeForInstruction::get(memInst).addTemp(addr);
1171 // Get the array type indexed by idxVal, and compute its element size.
1172 // The call to getTypeSize() will fail if size is not constant.
1173 const Type* vecType = (firstIdxIsZero
1174 ? GetElementPtrInst::getIndexedType(ptrType,
1175 std::vector<Value*>(1U, idxVec[0]),
1176 /*AllowCompositeLeaf*/ true)
1178 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
1179 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
1180 target.getTargetData().getTypeSize(eltType));
1182 // CreateMulInstruction() folds constants intelligently enough.
1183 CreateMulInstruction(target, memInst->getParent()->getParent(),
1184 idxVal, /* lval, not likely to be const*/
1185 eltSizeVal, /* rval, likely to be constant */
1187 mulVec, MachineCodeForInstruction::get(memInst),
1188 INVALID_MACHINE_OPCODE);
1190 assert(mulVec.size() > 0 && "No multiply code created?");
1191 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
1193 valueForRegOffset = addr;
1198 offsetOpType = MachineOperand::MO_SignExtendedImmed;
1199 smallConstOffset = 0;
1203 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
1204 // For LOAD or GET_ELEMENT_PTR,
1205 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
1207 unsigned offsetOpNum, ptrOpNum;
1209 if (memInst->getOpcode() == Instruction::Store) {
1210 if (offsetOpType == MachineOperand::MO_VirtualRegister)
1211 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1212 .addReg(ptrVal).addReg(valueForRegOffset);
1214 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1215 .addReg(ptrVal).addSImm(smallConstOffset);
1217 if (offsetOpType == MachineOperand::MO_VirtualRegister)
1218 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
1219 .addRegDef(memInst);
1221 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
1222 .addRegDef(memInst);
1229 // Substitute operand `operandNum' of the instruction in node `treeNode'
1230 // in place of the use(s) of that instruction in node `parent'.
1231 // Check both explicit and implicit operands!
1232 // Also make sure to skip over a parent who:
1233 // (1) is a list node in the Burg tree, or
1234 // (2) itself had its results forwarded to its parent
1237 ForwardOperand(InstructionNode* treeNode,
1238 InstrTreeNode* parent,
1241 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
1243 Instruction* unusedOp = treeNode->getInstruction();
1244 Value* fwdOp = unusedOp->getOperand(operandNum);
1246 // The parent itself may be a list node, so find the real parent instruction
1247 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode)
1249 parent = parent->parent();
1250 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
1252 InstructionNode* parentInstrNode = (InstructionNode*) parent;
1254 Instruction* userInstr = parentInstrNode->getInstruction();
1255 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
1257 // The parent's mvec would be empty if it was itself forwarded.
1258 // Recursively call ForwardOperand in that case...
1260 if (mvec.size() == 0)
1262 assert(parent->parent() != NULL &&
1263 "Parent could not have been forwarded, yet has no instructions?");
1264 ForwardOperand(treeNode, parent->parent(), operandNum);
1268 for (unsigned i=0, N=mvec.size(); i < N; i++)
1270 MachineInstr* minstr = mvec[i];
1271 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i)
1273 const MachineOperand& mop = minstr->getOperand(i);
1274 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
1275 mop.getVRegValue() == unusedOp)
1276 minstr->SetMachineOperandVal(i,
1277 MachineOperand::MO_VirtualRegister, fwdOp);
1280 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
1281 if (minstr->getImplicitRef(i) == unusedOp)
1282 minstr->setImplicitRef(i, fwdOp,
1283 minstr->implicitRefIsDefined(i),
1284 minstr->implicitRefIsDefinedAndUsed(i));
1291 AllUsesAreBranches(const Instruction* setccI)
1293 for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end();
1295 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
1296 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
1301 //******************* Externally Visible Functions *************************/
1303 //------------------------------------------------------------------------
1304 // External Function: ThisIsAChainRule
1307 // Check if a given BURG rule is a chain rule.
1308 //------------------------------------------------------------------------
1311 ThisIsAChainRule(int eruleno)
1315 case 111: // stmt: reg
1339 return false; break;
1344 //------------------------------------------------------------------------
1345 // External Function: GetInstructionsByRule
1348 // Choose machine instructions for the SPARC according to the
1349 // patterns chosen by the BURG-generated parser.
1350 //------------------------------------------------------------------------
1353 GetInstructionsByRule(InstructionNode* subtreeRoot,
1356 TargetMachine &target,
1357 std::vector<MachineInstr*>& mvec)
1359 bool checkCast = false; // initialize here to use fall-through
1360 bool maskUnsignedResult = false;
1362 int forwardOperandNum = -1;
1363 unsigned allocaSize = 0;
1364 MachineInstr* M, *M2;
1369 // If the code for this instruction was folded into the parent (user),
1371 if (subtreeRoot->isFoldedIntoParent())
1375 // Let's check for chain rules outside the switch so that we don't have
1376 // to duplicate the list of chain rule production numbers here again
1378 if (ThisIsAChainRule(ruleForNode))
1380 // Chain rules have a single nonterminal on the RHS.
1381 // Get the rule that matches the RHS non-terminal and use that instead.
1383 assert(nts[0] && ! nts[1]
1384 && "A chain rule should have only one RHS non-terminal!");
1385 nextRule = burm_rule(subtreeRoot->state, nts[0]);
1386 nts = burm_nts[nextRule];
1387 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
1391 switch(ruleForNode) {
1392 case 1: // stmt: Ret
1393 case 2: // stmt: RetValue(reg)
1394 { // NOTE: Prepass of register allocation is responsible
1395 // for moving return value to appropriate register.
1396 // Mark the return-address register as a hidden virtual reg.
1397 // Mark the return value register as an implicit ref of
1398 // the machine instruction.
1399 // Finally put a NOP in the delay slot.
1400 ReturnInst *returnInstr =
1401 cast<ReturnInst>(subtreeRoot->getInstruction());
1402 assert(returnInstr->getOpcode() == Instruction::Ret);
1404 Instruction* returnReg = new TmpInstruction(returnInstr);
1405 MachineCodeForInstruction::get(returnInstr).addTemp(returnReg);
1407 M = BuildMI(V9::JMPLRET, 3).addReg(returnReg).addSImm(8)
1408 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
1410 if (returnInstr->getReturnValue() != NULL)
1411 M->addImplicitRef(returnInstr->getReturnValue());
1414 mvec.push_back(BuildMI(V9::NOP, 0));
1419 case 3: // stmt: Store(reg,reg)
1420 case 4: // stmt: Store(reg,ptrreg)
1421 SetOperandsForMemInstr(ChooseStoreInstruction(
1422 subtreeRoot->leftChild()->getValue()->getType()),
1423 mvec, subtreeRoot, target);
1426 case 5: // stmt: BrUncond
1428 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1429 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
1432 mvec.push_back(BuildMI(V9::NOP, 0));
1436 case 206: // stmt: BrCond(setCCconst)
1437 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
1438 // If the constant is ZERO, we can use the branch-on-integer-register
1439 // instructions and avoid the SUBcc instruction entirely.
1440 // Otherwise this is just the same as case 5, so just fall through.
1442 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
1444 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
1445 Constant *constVal = cast<Constant>(constNode->getValue());
1448 if ((constVal->getType()->isInteger()
1449 || isa<PointerType>(constVal->getType()))
1450 && GetConstantValueAsSignedInt(constVal, isValidConst) == 0
1453 // That constant is a zero after all...
1454 // Use the left child of setCC as the first argument!
1455 // Mark the setCC node so that no code is generated for it.
1456 InstructionNode* setCCNode = (InstructionNode*)
1457 subtreeRoot->leftChild();
1458 assert(setCCNode->getOpLabel() == SetCCOp);
1459 setCCNode->markFoldedIntoParent();
1461 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
1463 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
1464 .addReg(setCCNode->leftChild()->getValue())
1465 .addPCDisp(brInst->getSuccessor(0));
1469 mvec.push_back(BuildMI(V9::NOP, 0));
1472 mvec.push_back(BuildMI(V9::BA, 1)
1473 .addPCDisp(brInst->getSuccessor(1)));
1476 mvec.push_back(BuildMI(V9::NOP, 0));
1479 // ELSE FALL THROUGH
1482 case 6: // stmt: BrCond(setCC)
1483 { // bool => boolean was computed with SetCC.
1484 // The branch to use depends on whether it is FP, signed, or unsigned.
1485 // If it is an integer CC, we also need to find the unique
1486 // TmpInstruction representing that CC.
1488 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
1490 unsigned Opcode = ChooseBccInstruction(subtreeRoot, isFPBranch);
1491 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
1492 brInst->getParent()->getParent(),
1493 isFPBranch? Type::FloatTy : Type::IntTy);
1494 M = BuildMI(Opcode, 2).addCCReg(ccValue)
1495 .addPCDisp(brInst->getSuccessor(0));
1499 mvec.push_back(BuildMI(V9::NOP, 0));
1502 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
1505 mvec.push_back(BuildMI(V9::NOP, 0));
1509 case 208: // stmt: BrCond(boolconst)
1511 // boolconst => boolean is a constant; use BA to first or second label
1512 Constant* constVal =
1513 cast<Constant>(subtreeRoot->leftChild()->getValue());
1514 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
1516 M = BuildMI(V9::BA, 1).addPCDisp(
1517 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
1521 mvec.push_back(BuildMI(V9::NOP, 0));
1525 case 8: // stmt: BrCond(boolreg)
1526 { // boolreg => boolean is stored in an existing register.
1527 // Just use the branch-on-integer-register instruction!
1529 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1530 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
1531 .addPCDisp(BI->getSuccessor(0));
1535 mvec.push_back(BuildMI(V9::NOP, 0));
1538 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
1541 mvec.push_back(BuildMI(V9::NOP, 0));
1545 case 9: // stmt: Switch(reg)
1546 assert(0 && "*** SWITCH instruction is not implemented yet.");
1549 case 10: // reg: VRegList(reg, reg)
1550 assert(0 && "VRegList should never be the topmost non-chain rule");
1553 case 21: // bool: Not(bool,reg): Both these are implemented as:
1554 case 421: // reg: BNot(reg,reg): reg = reg XOR-NOT 0
1555 { // First find the unary operand. It may be left or right, usually right.
1556 Value* notArg = BinaryOperator::getNotArgument(
1557 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1558 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1559 mvec.push_back(BuildMI(V9::XNOR, 3).addReg(notArg).addMReg(ZeroReg)
1560 .addRegDef(subtreeRoot->getValue()));
1564 case 22: // reg: ToBoolTy(reg):
1566 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
1567 assert(opType->isIntegral() || isa<PointerType>(opType));
1568 forwardOperandNum = 0; // forward first operand to user
1572 case 23: // reg: ToUByteTy(reg)
1573 case 24: // reg: ToSByteTy(reg)
1574 case 25: // reg: ToUShortTy(reg)
1575 case 26: // reg: ToShortTy(reg)
1576 case 27: // reg: ToUIntTy(reg)
1577 case 28: // reg: ToIntTy(reg)
1579 //======================================================================
1580 // Rules for integer conversions:
1583 // From ISO 1998 C++ Standard, Sec. 4.7:
1585 // 2. If the destination type is unsigned, the resulting value is
1586 // the least unsigned integer congruent to the source integer
1587 // (modulo 2n where n is the number of bits used to represent the
1588 // unsigned type). [Note: In a two s complement representation,
1589 // this conversion is conceptual and there is no change in the
1590 // bit pattern (if there is no truncation). ]
1592 // 3. If the destination type is signed, the value is unchanged if
1593 // it can be represented in the destination type (and bitfield width);
1594 // otherwise, the value is implementation-defined.
1597 // Since we assume 2s complement representations, this implies:
1599 // -- if operand is smaller than destination, zero-extend or sign-extend
1600 // according to the signedness of the *operand*: source decides.
1601 // ==> we have to do nothing here!
1603 // -- if operand is same size as or larger than destination, and the
1604 // destination is *unsigned*, zero-extend the operand: dest. decides
1606 // -- if operand is same size as or larger than destination, and the
1607 // destination is *signed*, the choice is implementation defined:
1608 // we sign-extend the operand: i.e., again dest. decides.
1609 // Note: this matches both Sun's cc and gcc3.2.
1610 //======================================================================
1612 Instruction* destI = subtreeRoot->getInstruction();
1613 Value* opVal = subtreeRoot->leftChild()->getValue();
1614 const Type* opType = opVal->getType();
1615 if (opType->isIntegral() || isa<PointerType>(opType))
1617 unsigned opSize = target.getTargetData().getTypeSize(opType);
1618 unsigned destSize = target.getTargetData().getTypeSize(destI->getType());
1619 if (opSize >= destSize)
1620 { // Operand is same size as or larger than dest:
1621 // zero- or sign-extend, according to the signeddness of
1622 // the destination (see above).
1623 if (destI->getType()->isSigned())
1624 target.getInstrInfo().CreateSignExtensionInstructions(target,
1625 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1626 mvec, MachineCodeForInstruction::get(destI));
1628 target.getInstrInfo().CreateZeroExtensionInstructions(target,
1629 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1630 mvec, MachineCodeForInstruction::get(destI));
1633 forwardOperandNum = 0; // forward first operand to user
1635 else if (opType->isFloatingPoint())
1637 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1638 MachineCodeForInstruction::get(destI));
1639 if (destI->getType()->isUnsigned())
1640 maskUnsignedResult = true; // not handled by fp->int code
1643 assert(0 && "Unrecognized operand type for convert-to-unsigned");
1648 case 29: // reg: ToULongTy(reg)
1649 case 30: // reg: ToLongTy(reg)
1651 Value* opVal = subtreeRoot->leftChild()->getValue();
1652 const Type* opType = opVal->getType();
1653 if (opType->isIntegral() || isa<PointerType>(opType))
1654 forwardOperandNum = 0; // forward first operand to user
1655 else if (opType->isFloatingPoint())
1657 Instruction* destI = subtreeRoot->getInstruction();
1658 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1659 MachineCodeForInstruction::get(destI));
1662 assert(0 && "Unrecognized operand type for convert-to-signed");
1666 case 31: // reg: ToFloatTy(reg):
1667 case 32: // reg: ToDoubleTy(reg):
1668 case 232: // reg: ToDoubleTy(Constant):
1670 // If this instruction has a parent (a user) in the tree
1671 // and the user is translated as an FsMULd instruction,
1672 // then the cast is unnecessary. So check that first.
1673 // In the future, we'll want to do the same for the FdMULq instruction,
1674 // so do the check here instead of only for ToFloatTy(reg).
1676 if (subtreeRoot->parent() != NULL)
1678 const MachineCodeForInstruction& mcfi =
1679 MachineCodeForInstruction::get(
1680 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
1681 if (mcfi.size() == 0 || mcfi.front()->getOpCode() == V9::FSMULD)
1682 forwardOperandNum = 0; // forward first operand to user
1685 if (forwardOperandNum != 0) // we do need the cast
1687 Value* leftVal = subtreeRoot->leftChild()->getValue();
1688 const Type* opType = leftVal->getType();
1689 MachineOpCode opCode=ChooseConvertToFloatInstr(
1690 subtreeRoot->getOpLabel(), opType);
1691 if (opCode == V9::INVALID_OPCODE) // no conversion needed
1693 forwardOperandNum = 0; // forward first operand to user
1697 // If the source operand is a non-FP type it must be
1698 // first copied from int to float register via memory!
1699 Instruction *dest = subtreeRoot->getInstruction();
1702 if (! opType->isFloatingPoint())
1704 // Create a temporary to represent the FP register
1705 // into which the integer will be copied via memory.
1706 // The type of this temporary will determine the FP
1707 // register used: single-prec for a 32-bit int or smaller,
1708 // double-prec for a 64-bit int.
1711 target.getTargetData().getTypeSize(leftVal->getType());
1712 Type* tmpTypeToUse =
1713 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
1714 srcForCast = new TmpInstruction(tmpTypeToUse, dest);
1715 MachineCodeForInstruction &destMCFI =
1716 MachineCodeForInstruction::get(dest);
1717 destMCFI.addTemp(srcForCast);
1719 target.getInstrInfo().CreateCodeToCopyIntToFloat(target,
1720 dest->getParent()->getParent(),
1721 leftVal, cast<Instruction>(srcForCast),
1725 srcForCast = leftVal;
1727 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
1733 case 19: // reg: ToArrayTy(reg):
1734 case 20: // reg: ToPointerTy(reg):
1735 forwardOperandNum = 0; // forward first operand to user
1738 case 233: // reg: Add(reg, Constant)
1739 maskUnsignedResult = true;
1740 M = CreateAddConstInstruction(subtreeRoot);
1746 // ELSE FALL THROUGH
1748 case 33: // reg: Add(reg, reg)
1749 maskUnsignedResult = true;
1750 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
1753 case 234: // reg: Sub(reg, Constant)
1754 maskUnsignedResult = true;
1755 M = CreateSubConstInstruction(subtreeRoot);
1761 // ELSE FALL THROUGH
1763 case 34: // reg: Sub(reg, reg)
1764 maskUnsignedResult = true;
1765 Add3OperandInstr(ChooseSubInstructionByType(
1766 subtreeRoot->getInstruction()->getType()),
1770 case 135: // reg: Mul(todouble, todouble)
1774 case 35: // reg: Mul(reg, reg)
1776 maskUnsignedResult = true;
1777 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1779 : INVALID_MACHINE_OPCODE);
1780 Instruction* mulInstr = subtreeRoot->getInstruction();
1781 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1782 subtreeRoot->leftChild()->getValue(),
1783 subtreeRoot->rightChild()->getValue(),
1785 MachineCodeForInstruction::get(mulInstr),forceOp);
1788 case 335: // reg: Mul(todouble, todoubleConst)
1792 case 235: // reg: Mul(reg, Constant)
1794 maskUnsignedResult = true;
1795 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1797 : INVALID_MACHINE_OPCODE);
1798 Instruction* mulInstr = subtreeRoot->getInstruction();
1799 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1800 subtreeRoot->leftChild()->getValue(),
1801 subtreeRoot->rightChild()->getValue(),
1803 MachineCodeForInstruction::get(mulInstr),
1807 case 236: // reg: Div(reg, Constant)
1808 maskUnsignedResult = true;
1810 CreateDivConstInstruction(target, subtreeRoot, mvec);
1811 if (mvec.size() > L)
1813 // ELSE FALL THROUGH
1815 case 36: // reg: Div(reg, reg)
1816 maskUnsignedResult = true;
1817 Add3OperandInstr(ChooseDivInstruction(target, subtreeRoot),
1821 case 37: // reg: Rem(reg, reg)
1822 case 237: // reg: Rem(reg, Constant)
1824 maskUnsignedResult = true;
1825 Instruction* remInstr = subtreeRoot->getInstruction();
1827 TmpInstruction* quot = new TmpInstruction(
1828 subtreeRoot->leftChild()->getValue(),
1829 subtreeRoot->rightChild()->getValue());
1830 TmpInstruction* prod = new TmpInstruction(
1832 subtreeRoot->rightChild()->getValue());
1833 MachineCodeForInstruction::get(remInstr).addTemp(quot).addTemp(prod);
1835 M = BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
1836 .addReg(subtreeRoot->leftChild()->getValue())
1837 .addReg(subtreeRoot->rightChild()->getValue())
1841 unsigned MulOpcode =
1842 ChooseMulInstructionByType(subtreeRoot->getInstruction()->getType());
1843 Value *MulRHS = subtreeRoot->rightChild()->getValue();
1844 M = BuildMI(MulOpcode, 3).addReg(quot).addReg(MulRHS).addReg(prod,
1848 unsigned Opcode = ChooseSubInstructionByType(
1849 subtreeRoot->getInstruction()->getType());
1850 M = BuildMI(Opcode, 3).addReg(subtreeRoot->leftChild()->getValue())
1851 .addReg(prod).addRegDef(subtreeRoot->getValue());
1856 case 38: // bool: And(bool, bool)
1857 case 238: // bool: And(bool, boolconst)
1858 case 338: // reg : BAnd(reg, reg)
1859 case 538: // reg : BAnd(reg, Constant)
1860 Add3OperandInstr(V9::AND, subtreeRoot, mvec);
1863 case 138: // bool: And(bool, not)
1864 case 438: // bool: BAnd(bool, bnot)
1865 { // Use the argument of NOT as the second argument!
1866 // Mark the NOT node so that no code is generated for it.
1867 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1868 Value* notArg = BinaryOperator::getNotArgument(
1869 cast<BinaryOperator>(notNode->getInstruction()));
1870 notNode->markFoldedIntoParent();
1871 Value *LHS = subtreeRoot->leftChild()->getValue();
1872 Value *Dest = subtreeRoot->getValue();
1873 mvec.push_back(BuildMI(V9::ANDN, 3).addReg(LHS).addReg(notArg)
1874 .addReg(Dest, MOTy::Def));
1878 case 39: // bool: Or(bool, bool)
1879 case 239: // bool: Or(bool, boolconst)
1880 case 339: // reg : BOr(reg, reg)
1881 case 539: // reg : BOr(reg, Constant)
1882 Add3OperandInstr(V9::OR, subtreeRoot, mvec);
1885 case 139: // bool: Or(bool, not)
1886 case 439: // bool: BOr(bool, bnot)
1887 { // Use the argument of NOT as the second argument!
1888 // Mark the NOT node so that no code is generated for it.
1889 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1890 Value* notArg = BinaryOperator::getNotArgument(
1891 cast<BinaryOperator>(notNode->getInstruction()));
1892 notNode->markFoldedIntoParent();
1893 Value *LHS = subtreeRoot->leftChild()->getValue();
1894 Value *Dest = subtreeRoot->getValue();
1895 mvec.push_back(BuildMI(V9::ORN, 3).addReg(LHS).addReg(notArg)
1896 .addReg(Dest, MOTy::Def));
1900 case 40: // bool: Xor(bool, bool)
1901 case 240: // bool: Xor(bool, boolconst)
1902 case 340: // reg : BXor(reg, reg)
1903 case 540: // reg : BXor(reg, Constant)
1904 Add3OperandInstr(V9::XOR, subtreeRoot, mvec);
1907 case 140: // bool: Xor(bool, not)
1908 case 440: // bool: BXor(bool, bnot)
1909 { // Use the argument of NOT as the second argument!
1910 // Mark the NOT node so that no code is generated for it.
1911 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1912 Value* notArg = BinaryOperator::getNotArgument(
1913 cast<BinaryOperator>(notNode->getInstruction()));
1914 notNode->markFoldedIntoParent();
1915 Value *LHS = subtreeRoot->leftChild()->getValue();
1916 Value *Dest = subtreeRoot->getValue();
1917 mvec.push_back(BuildMI(V9::XNOR, 3).addReg(LHS).addReg(notArg)
1918 .addReg(Dest, MOTy::Def));
1922 case 41: // boolconst: SetCC(reg, Constant)
1924 // If the SetCC was folded into the user (parent), it will be
1925 // caught above. All other cases are the same as case 42,
1926 // so just fall through.
1928 case 42: // bool: SetCC(reg, reg):
1930 // This generates a SUBCC instruction, putting the difference in
1931 // a result register, and setting a condition code.
1933 // If the boolean result of the SetCC is used by anything other
1934 // than a branch instruction, or if it is used outside the current
1935 // basic block, the boolean must be
1936 // computed and stored in the result register. Otherwise, discard
1937 // the difference (by using %g0) and keep only the condition code.
1939 // To compute the boolean result in a register we use a conditional
1940 // move, unless the result of the SUBCC instruction can be used as
1941 // the bool! This assumes that zero is FALSE and any non-zero
1944 InstructionNode* parentNode = (InstructionNode*) subtreeRoot->parent();
1945 Instruction* setCCInstr = subtreeRoot->getInstruction();
1947 bool keepBoolVal = parentNode == NULL ||
1948 ! AllUsesAreBranches(setCCInstr);
1949 bool subValIsBoolVal = setCCInstr->getOpcode() == Instruction::SetNE;
1950 bool keepSubVal = keepBoolVal && subValIsBoolVal;
1951 bool computeBoolVal = keepBoolVal && ! subValIsBoolVal;
1955 MachineOpCode movOpCode = 0;
1957 // Mark the 4th operand as being a CC register, and as a def
1958 // A TmpInstruction is created to represent the CC "result".
1959 // Unlike other instances of TmpInstruction, this one is used
1960 // by machine code of multiple LLVM instructions, viz.,
1961 // the SetCC and the branch. Make sure to get the same one!
1962 // Note that we do this even for FP CC registers even though they
1963 // are explicit operands, because the type of the operand
1964 // needs to be a floating point condition code, not an integer
1965 // condition code. Think of this as casting the bool result to
1966 // a FP condition code register.
1968 Value* leftVal = subtreeRoot->leftChild()->getValue();
1969 bool isFPCompare = leftVal->getType()->isFloatingPoint();
1971 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
1972 setCCInstr->getParent()->getParent(),
1973 isFPCompare ? Type::FloatTy : Type::IntTy);
1974 MachineCodeForInstruction::get(setCCInstr).addTemp(tmpForCC);
1978 // Integer condition: dest. should be %g0 or an integer register.
1979 // If result must be saved but condition is not SetEQ then we need
1980 // a separate instruction to compute the bool result, so discard
1981 // result of SUBcc instruction anyway.
1984 M = BuildMI(V9::SUBcc, 4)
1985 .addReg(subtreeRoot->leftChild()->getValue())
1986 .addReg(subtreeRoot->rightChild()->getValue())
1987 .addRegDef(subtreeRoot->getValue())
1988 .addCCReg(tmpForCC, MOTy::Def);
1990 M = BuildMI(V9::SUBcc, 4)
1991 .addReg(subtreeRoot->leftChild()->getValue())
1992 .addReg(subtreeRoot->rightChild()->getValue())
1993 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def)
1994 .addCCReg(tmpForCC, MOTy::Def);
1999 { // recompute bool using the integer condition codes
2001 ChooseMovpccAfterSub(subtreeRoot,mustClearReg,valueToMove);
2006 // FP condition: dest of FCMP should be some FCCn register
2007 M = BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
2008 .addCCReg(tmpForCC, MOTy::Def)
2009 .addReg(subtreeRoot->leftChild()->getValue())
2010 .addRegDef(subtreeRoot->rightChild()->getValue());
2014 {// recompute bool using the FP condition codes
2015 mustClearReg = true;
2017 movOpCode = ChooseMovFpccInstruction(subtreeRoot);
2024 {// Unconditionally set register to 0
2025 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
2029 // Now conditionally move `valueToMove' (0 or 1) into the register
2030 // Mark the register as a use (as well as a def) because the old
2031 // value should be retained if the condition is false.
2032 M = BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(valueToMove)
2033 .addReg(setCCInstr, MOTy::UseAndDef);
2039 case 51: // reg: Load(reg)
2040 case 52: // reg: Load(ptrreg)
2041 SetOperandsForMemInstr(ChooseLoadInstruction(
2042 subtreeRoot->getValue()->getType()),
2043 mvec, subtreeRoot, target);
2046 case 55: // reg: GetElemPtr(reg)
2047 case 56: // reg: GetElemPtrIdx(reg,reg)
2048 // If the GetElemPtr was folded into the user (parent), it will be
2049 // caught above. For other cases, we have to compute the address.
2050 SetOperandsForMemInstr(V9::ADD, mvec, subtreeRoot, target);
2053 case 57: // reg: Alloca: Implement as 1 instruction:
2054 { // add %fp, offsetFromFP -> result
2055 AllocationInst* instr =
2056 cast<AllocationInst>(subtreeRoot->getInstruction());
2058 target.getTargetData().getTypeSize(instr->getAllocatedType());
2060 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
2064 case 58: // reg: Alloca(reg): Implement as 3 instructions:
2065 // mul num, typeSz -> tmp
2066 // sub %sp, tmp -> %sp
2067 { // add %sp, frameSizeBelowDynamicArea -> result
2068 AllocationInst* instr =
2069 cast<AllocationInst>(subtreeRoot->getInstruction());
2070 const Type* eltType = instr->getAllocatedType();
2072 // If #elements is constant, use simpler code for fixed-size allocas
2073 int tsize = (int) target.getTargetData().getTypeSize(eltType);
2074 Value* numElementsVal = NULL;
2075 bool isArray = instr->isArrayAllocation();
2078 isa<Constant>(numElementsVal = instr->getArraySize()))
2079 { // total size is constant: generate code for fixed-size alloca
2080 unsigned numElements = isArray?
2081 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
2082 CreateCodeForFixedSizeAlloca(target, instr, tsize,
2085 else // total size is not constant.
2086 CreateCodeForVariableSizeAlloca(target, instr, tsize,
2087 numElementsVal, mvec);
2091 case 61: // reg: Call
2092 { // Generate a direct (CALL) or indirect (JMPL) call.
2093 // Mark the return-address register, the indirection
2094 // register (for indirect calls), the operands of the Call,
2095 // and the return value (if any) as implicit operands
2096 // of the machine instruction.
2098 // If this is a varargs function, floating point arguments
2099 // have to passed in integer registers so insert
2100 // copy-float-to-int instructions for each float operand.
2102 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
2103 Value *callee = callInstr->getCalledValue();
2105 // Create hidden virtual register for return address with type void*
2106 TmpInstruction* retAddrReg =
2107 new TmpInstruction(PointerType::get(Type::VoidTy), callInstr);
2108 MachineCodeForInstruction::get(callInstr).addTemp(retAddrReg);
2110 // Generate the machine instruction and its operands.
2111 // Use CALL for direct function calls; this optimistically assumes
2112 // the PC-relative address fits in the CALL address field (22 bits).
2113 // Use JMPL for indirect calls.
2115 if (isa<Function>(callee)) // direct function call
2116 M = BuildMI(V9::CALL, 1).addPCDisp(callee);
2117 else // indirect function call
2118 M = BuildMI(V9::JMPLCALL, 3).addReg(callee).addSImm((int64_t)0)
2119 .addRegDef(retAddrReg);
2122 const FunctionType* funcType =
2123 cast<FunctionType>(cast<PointerType>(callee->getType())
2124 ->getElementType());
2125 bool isVarArgs = funcType->isVarArg();
2126 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
2128 // Use a descriptor to pass information about call arguments
2129 // to the register allocator. This descriptor will be "owned"
2130 // and freed automatically when the MachineCodeForInstruction
2131 // object for the callInstr goes away.
2132 CallArgsDescriptor* argDesc = new CallArgsDescriptor(callInstr,
2133 retAddrReg, isVarArgs, noPrototype);
2135 assert(callInstr->getOperand(0) == callee
2136 && "This is assumed in the loop below!");
2138 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i)
2140 Value* argVal = callInstr->getOperand(i);
2141 Instruction* intArgReg = NULL;
2143 // Check for FP arguments to varargs functions.
2144 // Any such argument in the first $K$ args must be passed in an
2145 // integer register, where K = #integer argument registers.
2146 if (isVarArgs && argVal->getType()->isFloatingPoint())
2148 // If it is a function with no prototype, pass value
2149 // as an FP value as well as a varargs value
2151 argDesc->getArgInfo(i-1).setUseFPArgReg();
2153 // If this arg. is in the first $K$ regs, add a copy
2154 // float-to-int instruction to pass the value as an integer.
2155 if (i <= target.getRegInfo().GetNumOfIntArgRegs())
2157 MachineCodeForInstruction &destMCFI =
2158 MachineCodeForInstruction::get(callInstr);
2159 intArgReg = new TmpInstruction(Type::IntTy, argVal);
2160 destMCFI.addTemp(intArgReg);
2162 std::vector<MachineInstr*> copyMvec;
2163 target.getInstrInfo().CreateCodeToCopyFloatToInt(target,
2164 callInstr->getParent()->getParent(),
2165 argVal, (TmpInstruction*) intArgReg,
2166 copyMvec, destMCFI);
2167 mvec.insert(mvec.begin(),copyMvec.begin(),copyMvec.end());
2169 argDesc->getArgInfo(i-1).setUseIntArgReg();
2170 argDesc->getArgInfo(i-1).setArgCopy(intArgReg);
2173 // Cannot fit in first $K$ regs so pass the arg on the stack
2174 argDesc->getArgInfo(i-1).setUseStackSlot();
2178 mvec.back()->addImplicitRef(intArgReg);
2180 mvec.back()->addImplicitRef(argVal);
2183 // Add the return value as an implicit ref. The call operands
2184 // were added above.
2185 if (callInstr->getType() != Type::VoidTy)
2186 mvec.back()->addImplicitRef(callInstr, /*isDef*/ true);
2188 // For the CALL instruction, the ret. addr. reg. is also implicit
2189 if (isa<Function>(callee))
2190 mvec.back()->addImplicitRef(retAddrReg, /*isDef*/ true);
2193 mvec.push_back(BuildMI(V9::NOP, 0));
2197 case 62: // reg: Shl(reg, reg)
2199 Value* argVal1 = subtreeRoot->leftChild()->getValue();
2200 Value* argVal2 = subtreeRoot->rightChild()->getValue();
2201 Instruction* shlInstr = subtreeRoot->getInstruction();
2203 const Type* opType = argVal1->getType();
2204 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2205 "Shl unsupported for other types");
2207 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
2208 (opType == Type::LongTy)? V9::SLLX : V9::SLL,
2209 argVal1, argVal2, 0, shlInstr, mvec,
2210 MachineCodeForInstruction::get(shlInstr));
2214 case 63: // reg: Shr(reg, reg)
2215 { const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
2216 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2217 "Shr unsupported for other types");
2218 Add3OperandInstr(opType->isSigned()
2219 ? (opType == Type::LongTy ? V9::SRAX : V9::SRA)
2220 : (opType == Type::LongTy ? V9::SRLX : V9::SRL),
2225 case 64: // reg: Phi(reg,reg)
2226 break; // don't forward the value
2228 case 71: // reg: VReg
2229 case 72: // reg: Constant
2230 break; // don't forward the value
2233 assert(0 && "Unrecognized BURG rule");
2238 if (forwardOperandNum >= 0)
2239 { // We did not generate a machine instruction but need to use operand.
2240 // If user is in the same tree, replace Value in its machine operand.
2241 // If not, insert a copy instruction which should get coalesced away
2242 // by register allocation.
2243 if (subtreeRoot->parent() != NULL)
2244 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
2247 std::vector<MachineInstr*> minstrVec;
2248 Instruction* instr = subtreeRoot->getInstruction();
2249 target.getInstrInfo().
2250 CreateCopyInstructionsByType(target,
2251 instr->getParent()->getParent(),
2252 instr->getOperand(forwardOperandNum),
2254 MachineCodeForInstruction::get(instr));
2255 assert(minstrVec.size() > 0);
2256 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
2260 if (maskUnsignedResult)
2261 { // If result is unsigned and smaller than int reg size,
2262 // we need to clear high bits of result value.
2263 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
2264 Instruction* dest = subtreeRoot->getInstruction();
2265 if (dest->getType()->isUnsigned())
2267 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
2269 { // Mask high bits. Use a TmpInstruction to represent the
2270 // intermediate result before masking. Since those instructions
2271 // have already been generated, go back and substitute tmpI
2272 // for dest in the result position of each one of them.
2273 TmpInstruction *tmpI = new TmpInstruction(dest->getType(), dest,
2275 MachineCodeForInstruction::get(dest).addTemp(tmpI);
2277 for (unsigned i=0, N=mvec.size(); i < N; ++i)
2278 mvec[i]->substituteValue(dest, tmpI);
2280 M = BuildMI(V9::SRL, 3).addReg(tmpI).addZImm(8*(4-destSize))
2281 .addReg(dest, MOTy::Def);
2284 else if (destSize < 8)
2285 assert(0 && "Unsupported type size: 32 < size < 64 bits");