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());
141 else // cannot fold this getElementPtr instr. or any preceding ones
145 // If the first getElementPtr instruction had a leading [0], add it back.
146 // Note that this instruction is the *last* one successfully folded above.
147 if (ptrVal && ! lastInstHasLeadingNonZero)
148 chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
154 //---------------------------------------------------------------------------
155 // Function: GetGEPInstArgs
158 // Helper function for GetMemInstArgs that handles the final getElementPtr
159 // instruction used by (or same as) the memory operation.
160 // Extracts the indices of the current instruction and tries to fold in
161 // preceding ones if all indices of the current one are constant.
162 //---------------------------------------------------------------------------
165 GetGEPInstArgs(InstructionNode* gepNode,
166 std::vector<Value*>& idxVec,
167 bool& allConstantIndices)
169 allConstantIndices = true;
170 GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
172 // Default pointer is the one from the current instruction.
173 Value* ptrVal = gepI->getPointerOperand();
174 InstrTreeNode* ptrChild = gepNode->leftChild();
176 // Extract the index vector of the GEP instructin.
177 // If all indices are constant and first index is zero, try to fold
178 // in preceding GEPs with all constant indices.
179 for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
180 allConstantIndices && OI != OE; ++OI)
181 if (! isa<Constant>(*OI))
182 allConstantIndices = false; // note: this also terminates loop!
184 // If we have only constant indices, fold chains of constant indices
185 // in this and any preceding GetElemPtr instructions.
186 bool foldedGEPs = false;
187 bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
188 if (allConstantIndices)
189 if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx))
195 // Append the index vector of the current instruction.
196 // Skip the leading [0] index if preceding GEPs were folded into this.
197 idxVec.insert(idxVec.end(),
198 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
204 //---------------------------------------------------------------------------
205 // Function: GetMemInstArgs
208 // Get the pointer value and the index vector for a memory operation
209 // (GetElementPtr, Load, or Store). If all indices of the given memory
210 // operation are constant, fold in constant indices in a chain of
211 // preceding GetElementPtr instructions (if any), and return the
212 // pointer value of the first instruction in the chain.
213 // All folded instructions are marked so no code is generated for them.
216 // Returns the pointer Value to use.
217 // Returns the resulting IndexVector in idxVec.
218 // Returns true/false in allConstantIndices if all indices are/aren't const.
219 //---------------------------------------------------------------------------
222 GetMemInstArgs(InstructionNode* memInstrNode,
223 std::vector<Value*>& idxVec,
224 bool& allConstantIndices)
226 allConstantIndices = false;
227 Instruction* memInst = memInstrNode->getInstruction();
228 assert(idxVec.size() == 0 && "Need empty vector to return indices");
230 // If there is a GetElemPtr instruction to fold in to this instr,
231 // it must be in the left child for Load and GetElemPtr, and in the
232 // right child for Store instructions.
233 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
234 ? memInstrNode->rightChild()
235 : memInstrNode->leftChild());
237 // Default pointer is the one from the current instruction.
238 Value* ptrVal = ptrChild->getValue();
240 // Find the "last" GetElemPtr instruction: this one or the immediate child.
241 // There will be none if this is a load or a store from a scalar pointer.
242 InstructionNode* gepNode = NULL;
243 if (isa<GetElementPtrInst>(memInst))
244 gepNode = memInstrNode;
245 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal))
246 { // Child of load/store is a GEP and memInst is its only use.
247 // Use its indices and mark it as folded.
248 gepNode = cast<InstructionNode>(ptrChild);
249 gepNode->markFoldedIntoParent();
252 // If there are no indices, return the current pointer.
253 // Else extract the pointer from the GEP and fold the indices.
254 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
259 //************************ Internal Functions ******************************/
262 static inline MachineOpCode
263 ChooseBprInstruction(const InstructionNode* instrNode)
265 MachineOpCode opCode;
267 Instruction* setCCInstr =
268 ((InstructionNode*) instrNode->leftChild())->getInstruction();
270 switch(setCCInstr->getOpcode())
272 case Instruction::SetEQ: opCode = V9::BRZ; break;
273 case Instruction::SetNE: opCode = V9::BRNZ; break;
274 case Instruction::SetLE: opCode = V9::BRLEZ; break;
275 case Instruction::SetGE: opCode = V9::BRGEZ; break;
276 case Instruction::SetLT: opCode = V9::BRLZ; break;
277 case Instruction::SetGT: opCode = V9::BRGZ; break;
279 assert(0 && "Unrecognized VM instruction!");
280 opCode = V9::INVALID_OPCODE;
288 static inline MachineOpCode
289 ChooseBpccInstruction(const InstructionNode* instrNode,
290 const BinaryOperator* setCCInstr)
292 MachineOpCode opCode = V9::INVALID_OPCODE;
294 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
298 switch(setCCInstr->getOpcode())
300 case Instruction::SetEQ: opCode = V9::BE; break;
301 case Instruction::SetNE: opCode = V9::BNE; break;
302 case Instruction::SetLE: opCode = V9::BLE; break;
303 case Instruction::SetGE: opCode = V9::BGE; break;
304 case Instruction::SetLT: opCode = V9::BL; break;
305 case Instruction::SetGT: opCode = V9::BG; break;
307 assert(0 && "Unrecognized VM instruction!");
313 switch(setCCInstr->getOpcode())
315 case Instruction::SetEQ: opCode = V9::BE; break;
316 case Instruction::SetNE: opCode = V9::BNE; break;
317 case Instruction::SetLE: opCode = V9::BLEU; break;
318 case Instruction::SetGE: opCode = V9::BCC; break;
319 case Instruction::SetLT: opCode = V9::BCS; break;
320 case Instruction::SetGT: opCode = V9::BGU; break;
322 assert(0 && "Unrecognized VM instruction!");
330 static inline MachineOpCode
331 ChooseBFpccInstruction(const InstructionNode* instrNode,
332 const BinaryOperator* setCCInstr)
334 MachineOpCode opCode = V9::INVALID_OPCODE;
336 switch(setCCInstr->getOpcode())
338 case Instruction::SetEQ: opCode = V9::FBE; break;
339 case Instruction::SetNE: opCode = V9::FBNE; break;
340 case Instruction::SetLE: opCode = V9::FBLE; break;
341 case Instruction::SetGE: opCode = V9::FBGE; break;
342 case Instruction::SetLT: opCode = V9::FBL; break;
343 case Instruction::SetGT: opCode = V9::FBG; break;
345 assert(0 && "Unrecognized VM instruction!");
353 // Create a unique TmpInstruction for a boolean value,
354 // representing the CC register used by a branch on that value.
355 // For now, hack this using a little static cache of TmpInstructions.
356 // Eventually the entire BURG instruction selection should be put
357 // into a separate class that can hold such information.
358 // The static cache is not too bad because the memory for these
359 // TmpInstructions will be freed along with the rest of the Function anyway.
361 static TmpInstruction*
362 GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType)
364 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
365 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
366 static const Function *lastFunction = 0;// Use to flush cache between funcs
368 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
370 if (lastFunction != F)
373 boolToTmpCache.clear();
376 // Look for tmpI and create a new one otherwise. The new value is
377 // directly written to map using the ref returned by operator[].
378 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
380 tmpI = new TmpInstruction(ccType, boolVal);
386 static inline MachineOpCode
387 ChooseBccInstruction(const InstructionNode* instrNode,
390 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
391 assert(setCCNode->getOpLabel() == SetCCOp);
392 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
393 const Type* setCCType = setCCInstr->getOperand(0)->getType();
395 isFPBranch = setCCType->isFloatingPoint(); // Return value: don't delete!
398 return ChooseBFpccInstruction(instrNode, setCCInstr);
400 return ChooseBpccInstruction(instrNode, setCCInstr);
404 static inline MachineOpCode
405 ChooseMovFpccInstruction(const InstructionNode* instrNode)
407 MachineOpCode opCode = V9::INVALID_OPCODE;
409 switch(instrNode->getInstruction()->getOpcode())
411 case Instruction::SetEQ: opCode = V9::MOVFE; break;
412 case Instruction::SetNE: opCode = V9::MOVFNE; break;
413 case Instruction::SetLE: opCode = V9::MOVFLE; break;
414 case Instruction::SetGE: opCode = V9::MOVFGE; break;
415 case Instruction::SetLT: opCode = V9::MOVFL; break;
416 case Instruction::SetGT: opCode = V9::MOVFG; break;
418 assert(0 && "Unrecognized VM instruction!");
426 // Assumes that SUBcc v1, v2 -> v3 has been executed.
427 // In most cases, we want to clear v3 and then follow it by instruction
429 // Set mustClearReg=false if v3 need not be cleared before conditional move.
430 // Set valueToMove=0 if we want to conditionally move 0 instead of 1
431 // (i.e., we want to test inverse of a condition)
432 // (The latter two cases do not seem to arise because SetNE needs nothing.)
435 ChooseMovpccAfterSub(const InstructionNode* instrNode,
439 MachineOpCode opCode = V9::INVALID_OPCODE;
443 switch(instrNode->getInstruction()->getOpcode())
445 case Instruction::SetEQ: opCode = V9::MOVE; break;
446 case Instruction::SetLE: opCode = V9::MOVLE; break;
447 case Instruction::SetGE: opCode = V9::MOVGE; break;
448 case Instruction::SetLT: opCode = V9::MOVL; break;
449 case Instruction::SetGT: opCode = V9::MOVG; break;
450 case Instruction::SetNE: assert(0 && "No move required!"); break;
451 default: assert(0 && "Unrecognized VM instr!"); break;
457 static inline MachineOpCode
458 ChooseConvertToFloatInstr(OpLabel vopCode, const Type* opType)
460 MachineOpCode opCode = V9::INVALID_OPCODE;
465 if (opType == Type::SByteTy || opType == Type::ShortTy || opType == Type::IntTy)
467 else if (opType == Type::LongTy)
469 else if (opType == Type::DoubleTy)
471 else if (opType == Type::FloatTy)
474 assert(0 && "Cannot convert this type to FLOAT on SPARC");
478 // This is usually used in conjunction with CreateCodeToCopyIntToFloat().
479 // Both functions should treat the integer as a 32-bit value for types
480 // of 4 bytes or less, and as a 64-bit value otherwise.
481 if (opType == Type::SByteTy || opType == Type::UByteTy ||
482 opType == Type::ShortTy || opType == Type::UShortTy ||
483 opType == Type::IntTy || opType == Type::UIntTy)
485 else if (opType == Type::LongTy || opType == Type::ULongTy)
487 else if (opType == Type::FloatTy)
489 else if (opType == Type::DoubleTy)
492 assert(0 && "Cannot convert this type to DOUBLE on SPARC");
502 static inline MachineOpCode
503 ChooseConvertFPToIntInstr(Type::PrimitiveID tid, const Type* opType)
505 MachineOpCode opCode = V9::INVALID_OPCODE;;
507 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
508 && "This function should only be called for FLOAT or DOUBLE");
510 if (tid==Type::UIntTyID)
512 assert(tid != Type::UIntTyID && "FP-to-uint conversions must be expanded"
513 " into FP->long->uint for SPARC v9: SO RUN PRESELECTION PASS!");
515 else if (tid==Type::SByteTyID || tid==Type::ShortTyID || tid==Type::IntTyID ||
516 tid==Type::UByteTyID || tid==Type::UShortTyID)
518 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
520 else if (tid==Type::LongTyID || tid==Type::ULongTyID)
522 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
525 assert(0 && "Should not get here, Mo!");
531 CreateConvertFPToIntInstr(Type::PrimitiveID destTID,
532 Value* srcVal, Value* destVal)
534 MachineOpCode opCode = ChooseConvertFPToIntInstr(destTID, srcVal->getType());
535 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
536 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
539 // CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer
540 // The FP value must be converted to the dest type in an FP register,
541 // and the result is then copied from FP to int register via memory.
543 // Since fdtoi converts to signed integers, any FP value V between MAXINT+1
544 // and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly
545 // *only* when converting to an unsigned. (Unsigned byte, short or long
546 // don't have this problem.)
547 // For unsigned int, we therefore have to generate the code sequence:
549 // if (V > (float) MAXINT) {
550 // unsigned result = (unsigned) (V - (float) MAXINT);
551 // result = result + (unsigned) MAXINT;
554 // result = (unsigned) V;
557 CreateCodeToConvertFloatToInt(const TargetMachine& target,
560 std::vector<MachineInstr*>& mvec,
561 MachineCodeForInstruction& mcfi)
563 // Create a temporary to represent the FP register into which the
564 // int value will placed after conversion. The type of this temporary
565 // depends on the type of FP register to use: single-prec for a 32-bit
566 // int or smaller; double-prec for a 64-bit int.
568 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
569 const Type* destTypeToUse = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
570 TmpInstruction* destForCast = new TmpInstruction(destTypeToUse, opVal);
571 mcfi.addTemp(destForCast);
573 // Create the fp-to-int conversion code
574 MachineInstr* M =CreateConvertFPToIntInstr(destI->getType()->getPrimitiveID(),
578 // Create the fpreg-to-intreg copy code
579 target.getInstrInfo().
580 CreateCodeToCopyFloatToInt(target, destI->getParent()->getParent(),
581 destForCast, destI, mvec, mcfi);
585 static inline MachineOpCode
586 ChooseAddInstruction(const InstructionNode* instrNode)
588 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
592 static inline MachineInstr*
593 CreateMovFloatInstruction(const InstructionNode* instrNode,
594 const Type* resultType)
596 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
597 .addReg(instrNode->leftChild()->getValue())
598 .addRegDef(instrNode->getValue());
601 static inline MachineInstr*
602 CreateAddConstInstruction(const InstructionNode* instrNode)
604 MachineInstr* minstr = NULL;
606 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
607 assert(isa<Constant>(constOp));
609 // Cases worth optimizing are:
610 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
611 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
613 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
614 double dval = FPC->getValue();
616 minstr = CreateMovFloatInstruction(instrNode,
617 instrNode->getInstruction()->getType());
624 static inline MachineOpCode
625 ChooseSubInstructionByType(const Type* resultType)
627 MachineOpCode opCode = V9::INVALID_OPCODE;
629 if (resultType->isInteger() || isa<PointerType>(resultType))
634 switch(resultType->getPrimitiveID())
636 case Type::FloatTyID: opCode = V9::FSUBS; break;
637 case Type::DoubleTyID: opCode = V9::FSUBD; break;
638 default: assert(0 && "Invalid type for SUB instruction"); break;
645 static inline MachineInstr*
646 CreateSubConstInstruction(const InstructionNode* instrNode)
648 MachineInstr* minstr = NULL;
650 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
651 assert(isa<Constant>(constOp));
653 // Cases worth optimizing are:
654 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
655 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
657 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
658 double dval = FPC->getValue();
660 minstr = CreateMovFloatInstruction(instrNode,
661 instrNode->getInstruction()->getType());
668 static inline MachineOpCode
669 ChooseFcmpInstruction(const InstructionNode* instrNode)
671 MachineOpCode opCode = V9::INVALID_OPCODE;
673 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
674 switch(operand->getType()->getPrimitiveID()) {
675 case Type::FloatTyID: opCode = V9::FCMPS; break;
676 case Type::DoubleTyID: opCode = V9::FCMPD; break;
677 default: assert(0 && "Invalid type for FCMP instruction"); break;
684 // Assumes that leftArg and rightArg are both cast instructions.
687 BothFloatToDouble(const InstructionNode* instrNode)
689 InstrTreeNode* leftArg = instrNode->leftChild();
690 InstrTreeNode* rightArg = instrNode->rightChild();
691 InstrTreeNode* leftArgArg = leftArg->leftChild();
692 InstrTreeNode* rightArgArg = rightArg->leftChild();
693 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
695 // Check if both arguments are floats cast to double
696 return (leftArg->getValue()->getType() == Type::DoubleTy &&
697 leftArgArg->getValue()->getType() == Type::FloatTy &&
698 rightArgArg->getValue()->getType() == Type::FloatTy);
702 static inline MachineOpCode
703 ChooseMulInstructionByType(const Type* resultType)
705 MachineOpCode opCode = V9::INVALID_OPCODE;
707 if (resultType->isInteger())
710 switch(resultType->getPrimitiveID())
712 case Type::FloatTyID: opCode = V9::FMULS; break;
713 case Type::DoubleTyID: opCode = V9::FMULD; break;
714 default: assert(0 && "Invalid type for MUL instruction"); break;
722 static inline MachineInstr*
723 CreateIntNegInstruction(const TargetMachine& target,
726 return BuildMI(V9::SUB, 3).addMReg(target.getRegInfo().getZeroRegNum())
727 .addReg(vreg).addRegDef(vreg);
731 // Create instruction sequence for any shift operation.
732 // SLL or SLLX on an operand smaller than the integer reg. size (64bits)
733 // requires a second instruction for explicit sign-extension.
734 // Note that we only have to worry about a sign-bit appearing in the
735 // most significant bit of the operand after shifting (e.g., bit 32 of
736 // Int or bit 16 of Short), so we do not have to worry about results
737 // that are as large as a normal integer register.
740 CreateShiftInstructions(const TargetMachine& target,
742 MachineOpCode shiftOpCode,
744 Value* optArgVal2, /* Use optArgVal2 if not NULL */
745 unsigned optShiftNum, /* else use optShiftNum */
746 Instruction* destVal,
747 std::vector<MachineInstr*>& mvec,
748 MachineCodeForInstruction& mcfi)
750 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
751 "Large shift sizes unexpected, but can be handled below: "
752 "You need to check whether or not it fits in immed field below");
754 // If this is a logical left shift of a type smaller than the standard
755 // integer reg. size, we have to extend the sign-bit into upper bits
756 // of dest, so we need to put the result of the SLL into a temporary.
758 Value* shiftDest = destVal;
759 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
760 if ((shiftOpCode == V9::SLL || shiftOpCode == V9::SLLX) && opSize < 8)
761 { // put SLL result into a temporary
762 shiftDest = new TmpInstruction(argVal1, optArgVal2, "sllTmp");
763 mcfi.addTemp(shiftDest);
766 MachineInstr* M = (optArgVal2 != NULL)
767 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
768 .addReg(shiftDest, MOTy::Def)
769 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
770 .addReg(shiftDest, MOTy::Def);
773 if (shiftDest != destVal)
774 { // extend the sign-bit of the result into all upper bits of dest
775 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
776 target.getInstrInfo().
777 CreateSignExtensionInstructions(target, F, shiftDest, destVal,
778 8*opSize, mvec, mcfi);
783 // Does not create any instructions if we cannot exploit constant to
784 // create a cheaper instruction.
785 // This returns the approximate cost of the instructions generated,
786 // which is used to pick the cheapest when both operands are constant.
787 static inline unsigned
788 CreateMulConstInstruction(const TargetMachine &target, Function* F,
789 Value* lval, Value* rval, Instruction* destVal,
790 std::vector<MachineInstr*>& mvec,
791 MachineCodeForInstruction& mcfi)
793 /* Use max. multiply cost, viz., cost of MULX */
794 unsigned cost = target.getInstrInfo().minLatency(V9::MULX);
795 unsigned firstNewInstr = mvec.size();
797 Value* constOp = rval;
798 if (! isa<Constant>(constOp))
801 // Cases worth optimizing are:
802 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
803 // (2) Multiply by 2^x for integer types: replace with Shift
805 const Type* resultType = destVal->getType();
807 if (resultType->isInteger() || isa<PointerType>(resultType)) {
809 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
812 bool needNeg = false;
818 if (C == 0 || C == 1) {
819 cost = target.getInstrInfo().minLatency(V9::ADD);
820 unsigned Zero = target.getRegInfo().getZeroRegNum();
823 M = BuildMI(V9::ADD,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
825 M = BuildMI(V9::ADD,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
828 else if (isPowerOf2(C, pow)) {
829 unsigned opSize = target.getTargetData().getTypeSize(resultType);
830 MachineOpCode opCode = (opSize <= 32)? V9::SLL : V9::SLLX;
831 CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
832 destVal, mvec, mcfi);
835 if (mvec.size() > 0 && needNeg)
836 { // insert <reg = SUB 0, reg> after the instr to flip the sign
837 MachineInstr* M = CreateIntNegInstruction(target, destVal);
842 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
843 double dval = FPC->getValue();
844 if (fabs(dval) == 1) {
845 MachineOpCode opCode = (dval < 0)
846 ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
847 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
848 mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
853 if (firstNewInstr < mvec.size()) {
855 for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
856 cost += target.getInstrInfo().minLatency(mvec[i]->getOpCode());
863 // Does not create any instructions if we cannot exploit constant to
864 // create a cheaper instruction.
867 CreateCheapestMulConstInstruction(const TargetMachine &target,
869 Value* lval, Value* rval,
870 Instruction* destVal,
871 std::vector<MachineInstr*>& mvec,
872 MachineCodeForInstruction& mcfi)
875 if (isa<Constant>(lval) && isa<Constant>(rval))
876 { // both operands are constant: evaluate and "set" in dest
877 Constant* P = ConstantFoldBinaryInstruction(Instruction::Mul,
878 cast<Constant>(lval), cast<Constant>(rval));
879 target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi);
881 else if (isa<Constant>(rval)) // rval is constant, but not lval
882 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
883 else if (isa<Constant>(lval)) // lval is constant, but not rval
884 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
886 // else neither is constant
890 // Return NULL if we cannot exploit constant to create a cheaper instruction
892 CreateMulInstruction(const TargetMachine &target, Function* F,
893 Value* lval, Value* rval, Instruction* destVal,
894 std::vector<MachineInstr*>& mvec,
895 MachineCodeForInstruction& mcfi,
896 MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE)
898 unsigned L = mvec.size();
899 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
900 if (mvec.size() == L) {
901 // no instructions were added so create MUL reg, reg, reg.
902 // Use FSMULD if both operands are actually floats cast to doubles.
903 // Otherwise, use the default opcode for the appropriate type.
904 MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE)
906 : ChooseMulInstructionByType(destVal->getType()));
907 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
908 .addRegDef(destVal));
913 // Generate a divide instruction for Div or Rem.
914 // For Rem, this assumes that the operand type will be signed if the result
915 // type is signed. This is correct because they must have the same sign.
917 static inline MachineOpCode
918 ChooseDivInstruction(TargetMachine &target,
919 const InstructionNode* instrNode)
921 MachineOpCode opCode = V9::INVALID_OPCODE;
923 const Type* resultType = instrNode->getInstruction()->getType();
925 if (resultType->isInteger())
926 opCode = resultType->isSigned()? V9::SDIVX : V9::UDIVX;
928 switch(resultType->getPrimitiveID())
930 case Type::FloatTyID: opCode = V9::FDIVS; break;
931 case Type::DoubleTyID: opCode = V9::FDIVD; break;
932 default: assert(0 && "Invalid type for DIV instruction"); break;
939 // Return if we cannot exploit constant to create a cheaper instruction
941 CreateDivConstInstruction(TargetMachine &target,
942 const InstructionNode* instrNode,
943 std::vector<MachineInstr*>& mvec)
945 Value* LHS = instrNode->leftChild()->getValue();
946 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
947 if (!isa<Constant>(constOp))
950 Value* DestVal = instrNode->getValue();
951 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
953 // Cases worth optimizing are:
954 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
955 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
957 const Type* resultType = instrNode->getInstruction()->getType();
959 if (resultType->isInteger())
963 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
965 bool needNeg = false;
972 mvec.push_back(BuildMI(V9::ADD, 3).addReg(LHS).addMReg(ZeroReg)
973 .addRegDef(DestVal));
974 } else if (isPowerOf2(C, pow)) {
975 unsigned opCode= ((resultType->isSigned())
976 ? (resultType==Type::LongTy) ? V9::SRAX : V9::SRA
977 : (resultType==Type::LongTy) ? V9::SRLX : V9::SRL);
978 mvec.push_back(BuildMI(opCode, 3).addReg(LHS).addZImm(pow)
979 .addRegDef(DestVal));
982 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
983 // insert <reg = SUB 0, reg> after the instr to flip the sign
984 mvec.push_back(CreateIntNegInstruction(target, DestVal));
988 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
989 double dval = FPC->getValue();
990 if (fabs(dval) == 1) {
992 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
993 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
995 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(DestVal));
1003 CreateCodeForVariableSizeAlloca(const TargetMachine& target,
1004 Instruction* result,
1006 Value* numElementsVal,
1007 std::vector<MachineInstr*>& getMvec)
1009 Value* totalSizeVal;
1011 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
1012 Function *F = result->getParent()->getParent();
1014 // Enforce the alignment constraints on the stack pointer at
1015 // compile time if the total size is a known constant.
1016 if (isa<Constant>(numElementsVal))
1019 int64_t numElem = GetConstantValueAsSignedInt(numElementsVal, isValid);
1020 assert(isValid && "Unexpectedly large array dimension in alloca!");
1021 int64_t total = numElem * tsize;
1022 if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
1023 total += target.getFrameInfo().getStackFrameSizeAlignment() - extra;
1024 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
1028 // The size is not a constant. Generate code to compute it and
1029 // code to pad the size for stack alignment.
1030 // Create a Value to hold the (constant) element size
1031 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
1033 // Create temporary values to hold the result of MUL, SLL, SRL
1034 // THIS CASE IS INCOMPLETE AND WILL BE FIXED SHORTLY.
1035 TmpInstruction* tmpProd = new TmpInstruction(numElementsVal, tsizeVal);
1036 TmpInstruction* tmpSLL = new TmpInstruction(numElementsVal, tmpProd);
1037 TmpInstruction* tmpSRL = new TmpInstruction(numElementsVal, tmpSLL);
1038 mcfi.addTemp(tmpProd);
1039 mcfi.addTemp(tmpSLL);
1040 mcfi.addTemp(tmpSRL);
1042 // Instruction 1: mul numElements, typeSize -> tmpProd
1043 // This will optimize the MUL as far as possible.
1044 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd,getMvec,
1045 mcfi, INVALID_MACHINE_OPCODE);
1047 assert(0 && "Need to insert padding instructions here!");
1049 totalSizeVal = tmpProd;
1052 // Get the constant offset from SP for dynamically allocated storage
1053 // and create a temporary Value to hold it.
1054 MachineFunction& mcInfo = MachineFunction::get(F);
1056 ConstantSInt* dynamicAreaOffset =
1057 ConstantSInt::get(Type::IntTy,
1058 target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp));
1059 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
1061 unsigned SPReg = target.getRegInfo().getStackPointer();
1063 // Instruction 2: sub %sp, totalSizeVal -> %sp
1064 getMvec.push_back(BuildMI(V9::SUB, 3).addMReg(SPReg).addReg(totalSizeVal)
1065 .addMReg(SPReg,MOTy::Def));
1067 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
1068 getMvec.push_back(BuildMI(V9::ADD, 3).addMReg(SPReg).addReg(dynamicAreaOffset)
1069 .addRegDef(result));
1074 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
1075 Instruction* result,
1077 unsigned numElements,
1078 std::vector<MachineInstr*>& getMvec)
1080 assert(tsize > 0 && "Illegal (zero) type size for alloca");
1081 assert(result && result->getParent() &&
1082 "Result value is not part of a function?");
1083 Function *F = result->getParent()->getParent();
1084 MachineFunction &mcInfo = MachineFunction::get(F);
1086 // Check if the offset would small enough to use as an immediate in
1087 // load/stores (check LDX because all load/stores have the same-size immediate
1088 // field). If not, put the variable in the dynamically sized area of the
1090 unsigned paddedSizeIgnored;
1091 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
1093 tsize * numElements);
1094 if (! target.getInstrInfo().constantFitsInImmedField(V9::LDX, offsetFromFP)) {
1095 CreateCodeForVariableSizeAlloca(target, result, tsize,
1096 ConstantSInt::get(Type::IntTy,numElements),
1101 // else offset fits in immediate field so go ahead and allocate it.
1102 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
1104 // Create a temporary Value to hold the constant offset.
1105 // This is needed because it may not fit in the immediate field.
1106 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
1108 // Instruction 1: add %fp, offsetFromFP -> result
1109 unsigned FPReg = target.getRegInfo().getFramePointer();
1110 getMvec.push_back(BuildMI(V9::ADD, 3).addMReg(FPReg).addReg(offsetVal)
1111 .addRegDef(result));
1115 //------------------------------------------------------------------------
1116 // Function SetOperandsForMemInstr
1118 // Choose addressing mode for the given load or store instruction.
1119 // Use [reg+reg] if it is an indexed reference, and the index offset is
1120 // not a constant or if it cannot fit in the offset field.
1121 // Use [reg+offset] in all other cases.
1123 // This assumes that all array refs are "lowered" to one of these forms:
1124 // %x = load (subarray*) ptr, constant ; single constant offset
1125 // %x = load (subarray*) ptr, offsetVal ; single non-constant offset
1126 // Generally, this should happen via strength reduction + LICM.
1127 // Also, strength reduction should take care of using the same register for
1128 // the loop index variable and an array index, when that is profitable.
1129 //------------------------------------------------------------------------
1132 SetOperandsForMemInstr(unsigned Opcode,
1133 std::vector<MachineInstr*>& mvec,
1134 InstructionNode* vmInstrNode,
1135 const TargetMachine& target)
1137 Instruction* memInst = vmInstrNode->getInstruction();
1138 // Index vector, ptr value, and flag if all indices are const.
1139 std::vector<Value*> idxVec;
1140 bool allConstantIndices;
1141 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
1143 // Now create the appropriate operands for the machine instruction.
1144 // First, initialize so we default to storing the offset in a register.
1145 int64_t smallConstOffset = 0;
1146 Value* valueForRegOffset = NULL;
1147 MachineOperand::MachineOperandType offsetOpType =
1148 MachineOperand::MO_VirtualRegister;
1150 // Check if there is an index vector and if so, compute the
1151 // right offset for structures and for arrays
1153 if (!idxVec.empty())
1155 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
1157 // If all indices are constant, compute the combined offset directly.
1158 if (allConstantIndices)
1160 // Compute the offset value using the index vector. Create a
1161 // virtual reg. for it since it may not fit in the immed field.
1162 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
1163 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
1167 // There is at least one non-constant offset. Therefore, this must
1168 // be an array ref, and must have been lowered to a single non-zero
1169 // offset. (An extra leading zero offset, if any, can be ignored.)
1170 // Generate code sequence to compute address from index.
1172 bool firstIdxIsZero = IsZero(idxVec[0]);
1173 assert(idxVec.size() == 1U + firstIdxIsZero
1174 && "Array refs must be lowered before Instruction Selection");
1176 Value* idxVal = idxVec[firstIdxIsZero];
1178 std::vector<MachineInstr*> mulVec;
1179 Instruction* addr = new TmpInstruction(Type::ULongTy, memInst);
1180 MachineCodeForInstruction::get(memInst).addTemp(addr);
1182 // Get the array type indexed by idxVal, and compute its element size.
1183 // The call to getTypeSize() will fail if size is not constant.
1184 const Type* vecType = (firstIdxIsZero
1185 ? GetElementPtrInst::getIndexedType(ptrType,
1186 std::vector<Value*>(1U, idxVec[0]),
1187 /*AllowCompositeLeaf*/ true)
1189 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
1190 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
1191 target.getTargetData().getTypeSize(eltType));
1193 // CreateMulInstruction() folds constants intelligently enough.
1194 CreateMulInstruction(target, memInst->getParent()->getParent(),
1195 idxVal, /* lval, not likely to be const*/
1196 eltSizeVal, /* rval, likely to be constant */
1198 mulVec, MachineCodeForInstruction::get(memInst),
1199 INVALID_MACHINE_OPCODE);
1201 assert(mulVec.size() > 0 && "No multiply code created?");
1202 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
1204 valueForRegOffset = addr;
1209 offsetOpType = MachineOperand::MO_SignExtendedImmed;
1210 smallConstOffset = 0;
1214 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
1215 // For LOAD or GET_ELEMENT_PTR,
1216 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
1218 unsigned offsetOpNum, ptrOpNum;
1220 if (memInst->getOpcode() == Instruction::Store) {
1221 if (offsetOpType == MachineOperand::MO_VirtualRegister)
1222 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1223 .addReg(ptrVal).addReg(valueForRegOffset);
1225 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1226 .addReg(ptrVal).addSImm(smallConstOffset);
1228 if (offsetOpType == MachineOperand::MO_VirtualRegister)
1229 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
1230 .addRegDef(memInst);
1232 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
1233 .addRegDef(memInst);
1240 // Substitute operand `operandNum' of the instruction in node `treeNode'
1241 // in place of the use(s) of that instruction in node `parent'.
1242 // Check both explicit and implicit operands!
1243 // Also make sure to skip over a parent who:
1244 // (1) is a list node in the Burg tree, or
1245 // (2) itself had its results forwarded to its parent
1248 ForwardOperand(InstructionNode* treeNode,
1249 InstrTreeNode* parent,
1252 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
1254 Instruction* unusedOp = treeNode->getInstruction();
1255 Value* fwdOp = unusedOp->getOperand(operandNum);
1257 // The parent itself may be a list node, so find the real parent instruction
1258 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode)
1260 parent = parent->parent();
1261 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
1263 InstructionNode* parentInstrNode = (InstructionNode*) parent;
1265 Instruction* userInstr = parentInstrNode->getInstruction();
1266 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
1268 // The parent's mvec would be empty if it was itself forwarded.
1269 // Recursively call ForwardOperand in that case...
1271 if (mvec.size() == 0)
1273 assert(parent->parent() != NULL &&
1274 "Parent could not have been forwarded, yet has no instructions?");
1275 ForwardOperand(treeNode, parent->parent(), operandNum);
1279 for (unsigned i=0, N=mvec.size(); i < N; i++)
1281 MachineInstr* minstr = mvec[i];
1282 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i)
1284 const MachineOperand& mop = minstr->getOperand(i);
1285 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
1286 mop.getVRegValue() == unusedOp)
1287 minstr->SetMachineOperandVal(i,
1288 MachineOperand::MO_VirtualRegister, fwdOp);
1291 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
1292 if (minstr->getImplicitRef(i) == unusedOp)
1293 minstr->setImplicitRef(i, fwdOp,
1294 minstr->implicitRefIsDefined(i),
1295 minstr->implicitRefIsDefinedAndUsed(i));
1302 AllUsesAreBranches(const Instruction* setccI)
1304 for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end();
1306 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
1307 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
1312 //******************* Externally Visible Functions *************************/
1314 //------------------------------------------------------------------------
1315 // External Function: ThisIsAChainRule
1318 // Check if a given BURG rule is a chain rule.
1319 //------------------------------------------------------------------------
1322 ThisIsAChainRule(int eruleno)
1326 case 111: // stmt: reg
1350 return false; break;
1355 //------------------------------------------------------------------------
1356 // External Function: GetInstructionsByRule
1359 // Choose machine instructions for the SPARC according to the
1360 // patterns chosen by the BURG-generated parser.
1361 //------------------------------------------------------------------------
1364 GetInstructionsByRule(InstructionNode* subtreeRoot,
1367 TargetMachine &target,
1368 std::vector<MachineInstr*>& mvec)
1370 bool checkCast = false; // initialize here to use fall-through
1371 bool maskUnsignedResult = false;
1373 int forwardOperandNum = -1;
1374 unsigned allocaSize = 0;
1375 MachineInstr* M, *M2;
1380 // If the code for this instruction was folded into the parent (user),
1382 if (subtreeRoot->isFoldedIntoParent())
1386 // Let's check for chain rules outside the switch so that we don't have
1387 // to duplicate the list of chain rule production numbers here again
1389 if (ThisIsAChainRule(ruleForNode))
1391 // Chain rules have a single nonterminal on the RHS.
1392 // Get the rule that matches the RHS non-terminal and use that instead.
1394 assert(nts[0] && ! nts[1]
1395 && "A chain rule should have only one RHS non-terminal!");
1396 nextRule = burm_rule(subtreeRoot->state, nts[0]);
1397 nts = burm_nts[nextRule];
1398 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
1402 switch(ruleForNode) {
1403 case 1: // stmt: Ret
1404 case 2: // stmt: RetValue(reg)
1405 { // NOTE: Prepass of register allocation is responsible
1406 // for moving return value to appropriate register.
1407 // Mark the return-address register as a hidden virtual reg.
1408 // Mark the return value register as an implicit ref of
1409 // the machine instruction.
1410 // Finally put a NOP in the delay slot.
1411 ReturnInst *returnInstr =
1412 cast<ReturnInst>(subtreeRoot->getInstruction());
1413 assert(returnInstr->getOpcode() == Instruction::Ret);
1415 Instruction* returnReg = new TmpInstruction(returnInstr);
1416 MachineCodeForInstruction::get(returnInstr).addTemp(returnReg);
1418 M = BuildMI(V9::JMPLRET, 3).addReg(returnReg).addSImm(8)
1419 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
1421 if (returnInstr->getReturnValue() != NULL)
1422 M->addImplicitRef(returnInstr->getReturnValue());
1425 mvec.push_back(BuildMI(V9::NOP, 0));
1430 case 3: // stmt: Store(reg,reg)
1431 case 4: // stmt: Store(reg,ptrreg)
1432 SetOperandsForMemInstr(ChooseStoreInstruction(
1433 subtreeRoot->leftChild()->getValue()->getType()),
1434 mvec, subtreeRoot, target);
1437 case 5: // stmt: BrUncond
1439 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1440 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
1443 mvec.push_back(BuildMI(V9::NOP, 0));
1447 case 206: // stmt: BrCond(setCCconst)
1448 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
1449 // If the constant is ZERO, we can use the branch-on-integer-register
1450 // instructions and avoid the SUBcc instruction entirely.
1451 // Otherwise this is just the same as case 5, so just fall through.
1453 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
1455 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
1456 Constant *constVal = cast<Constant>(constNode->getValue());
1459 if ((constVal->getType()->isInteger()
1460 || isa<PointerType>(constVal->getType()))
1461 && GetConstantValueAsSignedInt(constVal, isValidConst) == 0
1464 // That constant is a zero after all...
1465 // Use the left child of setCC as the first argument!
1466 // Mark the setCC node so that no code is generated for it.
1467 InstructionNode* setCCNode = (InstructionNode*)
1468 subtreeRoot->leftChild();
1469 assert(setCCNode->getOpLabel() == SetCCOp);
1470 setCCNode->markFoldedIntoParent();
1472 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
1474 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
1475 .addReg(setCCNode->leftChild()->getValue())
1476 .addPCDisp(brInst->getSuccessor(0));
1480 mvec.push_back(BuildMI(V9::NOP, 0));
1483 mvec.push_back(BuildMI(V9::BA, 1)
1484 .addPCDisp(brInst->getSuccessor(1)));
1487 mvec.push_back(BuildMI(V9::NOP, 0));
1490 // ELSE FALL THROUGH
1493 case 6: // stmt: BrCond(setCC)
1494 { // bool => boolean was computed with SetCC.
1495 // The branch to use depends on whether it is FP, signed, or unsigned.
1496 // If it is an integer CC, we also need to find the unique
1497 // TmpInstruction representing that CC.
1499 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
1501 unsigned Opcode = ChooseBccInstruction(subtreeRoot, isFPBranch);
1502 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
1503 brInst->getParent()->getParent(),
1504 isFPBranch? Type::FloatTy : Type::IntTy);
1505 M = BuildMI(Opcode, 2).addCCReg(ccValue)
1506 .addPCDisp(brInst->getSuccessor(0));
1510 mvec.push_back(BuildMI(V9::NOP, 0));
1513 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
1516 mvec.push_back(BuildMI(V9::NOP, 0));
1520 case 208: // stmt: BrCond(boolconst)
1522 // boolconst => boolean is a constant; use BA to first or second label
1523 Constant* constVal =
1524 cast<Constant>(subtreeRoot->leftChild()->getValue());
1525 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
1527 M = BuildMI(V9::BA, 1).addPCDisp(
1528 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
1532 mvec.push_back(BuildMI(V9::NOP, 0));
1536 case 8: // stmt: BrCond(boolreg)
1537 { // boolreg => boolean is stored in an existing register.
1538 // Just use the branch-on-integer-register instruction!
1540 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1541 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
1542 .addPCDisp(BI->getSuccessor(0));
1546 mvec.push_back(BuildMI(V9::NOP, 0));
1549 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
1552 mvec.push_back(BuildMI(V9::NOP, 0));
1556 case 9: // stmt: Switch(reg)
1557 assert(0 && "*** SWITCH instruction is not implemented yet.");
1560 case 10: // reg: VRegList(reg, reg)
1561 assert(0 && "VRegList should never be the topmost non-chain rule");
1564 case 21: // bool: Not(bool,reg): Both these are implemented as:
1565 case 421: // reg: BNot(reg,reg): reg = reg XOR-NOT 0
1566 { // First find the unary operand. It may be left or right, usually right.
1567 Value* notArg = BinaryOperator::getNotArgument(
1568 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1569 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1570 mvec.push_back(BuildMI(V9::XNOR, 3).addReg(notArg).addMReg(ZeroReg)
1571 .addRegDef(subtreeRoot->getValue()));
1575 case 22: // reg: ToBoolTy(reg):
1577 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
1578 assert(opType->isIntegral() || isa<PointerType>(opType));
1579 forwardOperandNum = 0; // forward first operand to user
1583 case 23: // reg: ToUByteTy(reg)
1584 case 24: // reg: ToSByteTy(reg)
1585 case 25: // reg: ToUShortTy(reg)
1586 case 26: // reg: ToShortTy(reg)
1587 case 27: // reg: ToUIntTy(reg)
1588 case 28: // reg: ToIntTy(reg)
1590 //======================================================================
1591 // Rules for integer conversions:
1594 // From ISO 1998 C++ Standard, Sec. 4.7:
1596 // 2. If the destination type is unsigned, the resulting value is
1597 // the least unsigned integer congruent to the source integer
1598 // (modulo 2n where n is the number of bits used to represent the
1599 // unsigned type). [Note: In a two s complement representation,
1600 // this conversion is conceptual and there is no change in the
1601 // bit pattern (if there is no truncation). ]
1603 // 3. If the destination type is signed, the value is unchanged if
1604 // it can be represented in the destination type (and bitfield width);
1605 // otherwise, the value is implementation-defined.
1608 // Since we assume 2s complement representations, this implies:
1610 // -- if operand is smaller than destination, zero-extend or sign-extend
1611 // according to the signedness of the *operand*: source decides.
1612 // ==> we have to do nothing here!
1614 // -- if operand is same size as or larger than destination, and the
1615 // destination is *unsigned*, zero-extend the operand: dest. decides
1617 // -- if operand is same size as or larger than destination, and the
1618 // destination is *signed*, the choice is implementation defined:
1619 // we sign-extend the operand: i.e., again dest. decides.
1620 // Note: this matches both Sun's cc and gcc3.2.
1621 //======================================================================
1623 Instruction* destI = subtreeRoot->getInstruction();
1624 Value* opVal = subtreeRoot->leftChild()->getValue();
1625 const Type* opType = opVal->getType();
1626 if (opType->isIntegral() || isa<PointerType>(opType))
1628 unsigned opSize = target.getTargetData().getTypeSize(opType);
1629 unsigned destSize = target.getTargetData().getTypeSize(destI->getType());
1630 if (opSize >= destSize)
1631 { // Operand is same size as or larger than dest:
1632 // zero- or sign-extend, according to the signeddness of
1633 // the destination (see above).
1634 if (destI->getType()->isSigned())
1635 target.getInstrInfo().CreateSignExtensionInstructions(target,
1636 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1637 mvec, MachineCodeForInstruction::get(destI));
1639 target.getInstrInfo().CreateZeroExtensionInstructions(target,
1640 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1641 mvec, MachineCodeForInstruction::get(destI));
1644 forwardOperandNum = 0; // forward first operand to user
1646 else if (opType->isFloatingPoint())
1648 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1649 MachineCodeForInstruction::get(destI));
1650 if (destI->getType()->isUnsigned())
1651 maskUnsignedResult = true; // not handled by fp->int code
1654 assert(0 && "Unrecognized operand type for convert-to-unsigned");
1659 case 29: // reg: ToULongTy(reg)
1660 case 30: // reg: ToLongTy(reg)
1662 Value* opVal = subtreeRoot->leftChild()->getValue();
1663 const Type* opType = opVal->getType();
1664 if (opType->isIntegral() || isa<PointerType>(opType))
1665 forwardOperandNum = 0; // forward first operand to user
1666 else if (opType->isFloatingPoint())
1668 Instruction* destI = subtreeRoot->getInstruction();
1669 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1670 MachineCodeForInstruction::get(destI));
1673 assert(0 && "Unrecognized operand type for convert-to-signed");
1677 case 31: // reg: ToFloatTy(reg):
1678 case 32: // reg: ToDoubleTy(reg):
1679 case 232: // reg: ToDoubleTy(Constant):
1681 // If this instruction has a parent (a user) in the tree
1682 // and the user is translated as an FsMULd instruction,
1683 // then the cast is unnecessary. So check that first.
1684 // In the future, we'll want to do the same for the FdMULq instruction,
1685 // so do the check here instead of only for ToFloatTy(reg).
1687 if (subtreeRoot->parent() != NULL)
1689 const MachineCodeForInstruction& mcfi =
1690 MachineCodeForInstruction::get(
1691 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
1692 if (mcfi.size() == 0 || mcfi.front()->getOpCode() == V9::FSMULD)
1693 forwardOperandNum = 0; // forward first operand to user
1696 if (forwardOperandNum != 0) // we do need the cast
1698 Value* leftVal = subtreeRoot->leftChild()->getValue();
1699 const Type* opType = leftVal->getType();
1700 MachineOpCode opCode=ChooseConvertToFloatInstr(
1701 subtreeRoot->getOpLabel(), opType);
1702 if (opCode == V9::INVALID_OPCODE) // no conversion needed
1704 forwardOperandNum = 0; // forward first operand to user
1708 // If the source operand is a non-FP type it must be
1709 // first copied from int to float register via memory!
1710 Instruction *dest = subtreeRoot->getInstruction();
1713 if (! opType->isFloatingPoint())
1715 // Create a temporary to represent the FP register
1716 // into which the integer will be copied via memory.
1717 // The type of this temporary will determine the FP
1718 // register used: single-prec for a 32-bit int or smaller,
1719 // double-prec for a 64-bit int.
1722 target.getTargetData().getTypeSize(leftVal->getType());
1723 Type* tmpTypeToUse =
1724 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
1725 srcForCast = new TmpInstruction(tmpTypeToUse, dest);
1726 MachineCodeForInstruction &destMCFI =
1727 MachineCodeForInstruction::get(dest);
1728 destMCFI.addTemp(srcForCast);
1730 target.getInstrInfo().CreateCodeToCopyIntToFloat(target,
1731 dest->getParent()->getParent(),
1732 leftVal, cast<Instruction>(srcForCast),
1736 srcForCast = leftVal;
1738 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
1744 case 19: // reg: ToArrayTy(reg):
1745 case 20: // reg: ToPointerTy(reg):
1746 forwardOperandNum = 0; // forward first operand to user
1749 case 233: // reg: Add(reg, Constant)
1750 maskUnsignedResult = true;
1751 M = CreateAddConstInstruction(subtreeRoot);
1757 // ELSE FALL THROUGH
1759 case 33: // reg: Add(reg, reg)
1760 maskUnsignedResult = true;
1761 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
1764 case 234: // reg: Sub(reg, Constant)
1765 maskUnsignedResult = true;
1766 M = CreateSubConstInstruction(subtreeRoot);
1772 // ELSE FALL THROUGH
1774 case 34: // reg: Sub(reg, reg)
1775 maskUnsignedResult = true;
1776 Add3OperandInstr(ChooseSubInstructionByType(
1777 subtreeRoot->getInstruction()->getType()),
1781 case 135: // reg: Mul(todouble, todouble)
1785 case 35: // reg: Mul(reg, reg)
1787 maskUnsignedResult = true;
1788 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1790 : INVALID_MACHINE_OPCODE);
1791 Instruction* mulInstr = subtreeRoot->getInstruction();
1792 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1793 subtreeRoot->leftChild()->getValue(),
1794 subtreeRoot->rightChild()->getValue(),
1796 MachineCodeForInstruction::get(mulInstr),forceOp);
1799 case 335: // reg: Mul(todouble, todoubleConst)
1803 case 235: // reg: Mul(reg, Constant)
1805 maskUnsignedResult = true;
1806 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1808 : INVALID_MACHINE_OPCODE);
1809 Instruction* mulInstr = subtreeRoot->getInstruction();
1810 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1811 subtreeRoot->leftChild()->getValue(),
1812 subtreeRoot->rightChild()->getValue(),
1814 MachineCodeForInstruction::get(mulInstr),
1818 case 236: // reg: Div(reg, Constant)
1819 maskUnsignedResult = true;
1821 CreateDivConstInstruction(target, subtreeRoot, mvec);
1822 if (mvec.size() > L)
1824 // ELSE FALL THROUGH
1826 case 36: // reg: Div(reg, reg)
1827 maskUnsignedResult = true;
1828 Add3OperandInstr(ChooseDivInstruction(target, subtreeRoot),
1832 case 37: // reg: Rem(reg, reg)
1833 case 237: // reg: Rem(reg, Constant)
1835 maskUnsignedResult = true;
1836 Instruction* remInstr = subtreeRoot->getInstruction();
1838 TmpInstruction* quot = new TmpInstruction(
1839 subtreeRoot->leftChild()->getValue(),
1840 subtreeRoot->rightChild()->getValue());
1841 TmpInstruction* prod = new TmpInstruction(
1843 subtreeRoot->rightChild()->getValue());
1844 MachineCodeForInstruction::get(remInstr).addTemp(quot).addTemp(prod);
1846 M = BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
1847 .addReg(subtreeRoot->leftChild()->getValue())
1848 .addReg(subtreeRoot->rightChild()->getValue())
1852 unsigned MulOpcode =
1853 ChooseMulInstructionByType(subtreeRoot->getInstruction()->getType());
1854 Value *MulRHS = subtreeRoot->rightChild()->getValue();
1855 M = BuildMI(MulOpcode, 3).addReg(quot).addReg(MulRHS).addReg(prod,
1859 unsigned Opcode = ChooseSubInstructionByType(
1860 subtreeRoot->getInstruction()->getType());
1861 M = BuildMI(Opcode, 3).addReg(subtreeRoot->leftChild()->getValue())
1862 .addReg(prod).addRegDef(subtreeRoot->getValue());
1867 case 38: // bool: And(bool, bool)
1868 case 238: // bool: And(bool, boolconst)
1869 case 338: // reg : BAnd(reg, reg)
1870 case 538: // reg : BAnd(reg, Constant)
1871 Add3OperandInstr(V9::AND, subtreeRoot, mvec);
1874 case 138: // bool: And(bool, not)
1875 case 438: // bool: BAnd(bool, bnot)
1876 { // Use the argument of NOT as the second argument!
1877 // Mark the NOT node so that no code is generated for it.
1878 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1879 Value* notArg = BinaryOperator::getNotArgument(
1880 cast<BinaryOperator>(notNode->getInstruction()));
1881 notNode->markFoldedIntoParent();
1882 Value *LHS = subtreeRoot->leftChild()->getValue();
1883 Value *Dest = subtreeRoot->getValue();
1884 mvec.push_back(BuildMI(V9::ANDN, 3).addReg(LHS).addReg(notArg)
1885 .addReg(Dest, MOTy::Def));
1889 case 39: // bool: Or(bool, bool)
1890 case 239: // bool: Or(bool, boolconst)
1891 case 339: // reg : BOr(reg, reg)
1892 case 539: // reg : BOr(reg, Constant)
1893 Add3OperandInstr(V9::OR, subtreeRoot, mvec);
1896 case 139: // bool: Or(bool, not)
1897 case 439: // bool: BOr(bool, bnot)
1898 { // Use the argument of NOT as the second argument!
1899 // Mark the NOT node so that no code is generated for it.
1900 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1901 Value* notArg = BinaryOperator::getNotArgument(
1902 cast<BinaryOperator>(notNode->getInstruction()));
1903 notNode->markFoldedIntoParent();
1904 Value *LHS = subtreeRoot->leftChild()->getValue();
1905 Value *Dest = subtreeRoot->getValue();
1906 mvec.push_back(BuildMI(V9::ORN, 3).addReg(LHS).addReg(notArg)
1907 .addReg(Dest, MOTy::Def));
1911 case 40: // bool: Xor(bool, bool)
1912 case 240: // bool: Xor(bool, boolconst)
1913 case 340: // reg : BXor(reg, reg)
1914 case 540: // reg : BXor(reg, Constant)
1915 Add3OperandInstr(V9::XOR, subtreeRoot, mvec);
1918 case 140: // bool: Xor(bool, not)
1919 case 440: // bool: BXor(bool, bnot)
1920 { // Use the argument of NOT as the second argument!
1921 // Mark the NOT node so that no code is generated for it.
1922 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1923 Value* notArg = BinaryOperator::getNotArgument(
1924 cast<BinaryOperator>(notNode->getInstruction()));
1925 notNode->markFoldedIntoParent();
1926 Value *LHS = subtreeRoot->leftChild()->getValue();
1927 Value *Dest = subtreeRoot->getValue();
1928 mvec.push_back(BuildMI(V9::XNOR, 3).addReg(LHS).addReg(notArg)
1929 .addReg(Dest, MOTy::Def));
1933 case 41: // boolconst: SetCC(reg, Constant)
1935 // If the SetCC was folded into the user (parent), it will be
1936 // caught above. All other cases are the same as case 42,
1937 // so just fall through.
1939 case 42: // bool: SetCC(reg, reg):
1941 // This generates a SUBCC instruction, putting the difference in
1942 // a result register, and setting a condition code.
1944 // If the boolean result of the SetCC is used by anything other
1945 // than a branch instruction, or if it is used outside the current
1946 // basic block, the boolean must be
1947 // computed and stored in the result register. Otherwise, discard
1948 // the difference (by using %g0) and keep only the condition code.
1950 // To compute the boolean result in a register we use a conditional
1951 // move, unless the result of the SUBCC instruction can be used as
1952 // the bool! This assumes that zero is FALSE and any non-zero
1955 InstructionNode* parentNode = (InstructionNode*) subtreeRoot->parent();
1956 Instruction* setCCInstr = subtreeRoot->getInstruction();
1958 bool keepBoolVal = parentNode == NULL ||
1959 ! AllUsesAreBranches(setCCInstr);
1960 bool subValIsBoolVal = setCCInstr->getOpcode() == Instruction::SetNE;
1961 bool keepSubVal = keepBoolVal && subValIsBoolVal;
1962 bool computeBoolVal = keepBoolVal && ! subValIsBoolVal;
1966 MachineOpCode movOpCode = 0;
1968 // Mark the 4th operand as being a CC register, and as a def
1969 // A TmpInstruction is created to represent the CC "result".
1970 // Unlike other instances of TmpInstruction, this one is used
1971 // by machine code of multiple LLVM instructions, viz.,
1972 // the SetCC and the branch. Make sure to get the same one!
1973 // Note that we do this even for FP CC registers even though they
1974 // are explicit operands, because the type of the operand
1975 // needs to be a floating point condition code, not an integer
1976 // condition code. Think of this as casting the bool result to
1977 // a FP condition code register.
1979 Value* leftVal = subtreeRoot->leftChild()->getValue();
1980 bool isFPCompare = leftVal->getType()->isFloatingPoint();
1982 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
1983 setCCInstr->getParent()->getParent(),
1984 isFPCompare ? Type::FloatTy : Type::IntTy);
1985 MachineCodeForInstruction::get(setCCInstr).addTemp(tmpForCC);
1989 // Integer condition: dest. should be %g0 or an integer register.
1990 // If result must be saved but condition is not SetEQ then we need
1991 // a separate instruction to compute the bool result, so discard
1992 // result of SUBcc instruction anyway.
1995 M = BuildMI(V9::SUBcc, 4)
1996 .addReg(subtreeRoot->leftChild()->getValue())
1997 .addReg(subtreeRoot->rightChild()->getValue())
1998 .addRegDef(subtreeRoot->getValue())
1999 .addCCReg(tmpForCC, MOTy::Def);
2001 M = BuildMI(V9::SUBcc, 4)
2002 .addReg(subtreeRoot->leftChild()->getValue())
2003 .addReg(subtreeRoot->rightChild()->getValue())
2004 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def)
2005 .addCCReg(tmpForCC, MOTy::Def);
2010 { // recompute bool using the integer condition codes
2012 ChooseMovpccAfterSub(subtreeRoot,mustClearReg,valueToMove);
2017 // FP condition: dest of FCMP should be some FCCn register
2018 M = BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
2019 .addCCReg(tmpForCC, MOTy::Def)
2020 .addReg(subtreeRoot->leftChild()->getValue())
2021 .addRegDef(subtreeRoot->rightChild()->getValue());
2025 {// recompute bool using the FP condition codes
2026 mustClearReg = true;
2028 movOpCode = ChooseMovFpccInstruction(subtreeRoot);
2035 {// Unconditionally set register to 0
2036 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
2040 // Now conditionally move `valueToMove' (0 or 1) into the register
2041 // Mark the register as a use (as well as a def) because the old
2042 // value should be retained if the condition is false.
2043 M = BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(valueToMove)
2044 .addReg(setCCInstr, MOTy::UseAndDef);
2050 case 51: // reg: Load(reg)
2051 case 52: // reg: Load(ptrreg)
2052 SetOperandsForMemInstr(ChooseLoadInstruction(
2053 subtreeRoot->getValue()->getType()),
2054 mvec, subtreeRoot, target);
2057 case 55: // reg: GetElemPtr(reg)
2058 case 56: // reg: GetElemPtrIdx(reg,reg)
2059 // If the GetElemPtr was folded into the user (parent), it will be
2060 // caught above. For other cases, we have to compute the address.
2061 SetOperandsForMemInstr(V9::ADD, mvec, subtreeRoot, target);
2064 case 57: // reg: Alloca: Implement as 1 instruction:
2065 { // add %fp, offsetFromFP -> result
2066 AllocationInst* instr =
2067 cast<AllocationInst>(subtreeRoot->getInstruction());
2069 target.getTargetData().getTypeSize(instr->getAllocatedType());
2071 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
2075 case 58: // reg: Alloca(reg): Implement as 3 instructions:
2076 // mul num, typeSz -> tmp
2077 // sub %sp, tmp -> %sp
2078 { // add %sp, frameSizeBelowDynamicArea -> result
2079 AllocationInst* instr =
2080 cast<AllocationInst>(subtreeRoot->getInstruction());
2081 const Type* eltType = instr->getAllocatedType();
2083 // If #elements is constant, use simpler code for fixed-size allocas
2084 int tsize = (int) target.getTargetData().getTypeSize(eltType);
2085 Value* numElementsVal = NULL;
2086 bool isArray = instr->isArrayAllocation();
2089 isa<Constant>(numElementsVal = instr->getArraySize()))
2090 { // total size is constant: generate code for fixed-size alloca
2091 unsigned numElements = isArray?
2092 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
2093 CreateCodeForFixedSizeAlloca(target, instr, tsize,
2096 else // total size is not constant.
2097 CreateCodeForVariableSizeAlloca(target, instr, tsize,
2098 numElementsVal, mvec);
2102 case 61: // reg: Call
2103 { // Generate a direct (CALL) or indirect (JMPL) call.
2104 // Mark the return-address register, the indirection
2105 // register (for indirect calls), the operands of the Call,
2106 // and the return value (if any) as implicit operands
2107 // of the machine instruction.
2109 // If this is a varargs function, floating point arguments
2110 // have to passed in integer registers so insert
2111 // copy-float-to-int instructions for each float operand.
2113 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
2114 Value *callee = callInstr->getCalledValue();
2116 // Create hidden virtual register for return address with type void*
2117 TmpInstruction* retAddrReg =
2118 new TmpInstruction(PointerType::get(Type::VoidTy), callInstr);
2119 MachineCodeForInstruction::get(callInstr).addTemp(retAddrReg);
2121 // Generate the machine instruction and its operands.
2122 // Use CALL for direct function calls; this optimistically assumes
2123 // the PC-relative address fits in the CALL address field (22 bits).
2124 // Use JMPL for indirect calls.
2126 if (isa<Function>(callee)) // direct function call
2127 M = BuildMI(V9::CALL, 1).addPCDisp(callee);
2128 else // indirect function call
2129 M = BuildMI(V9::JMPLCALL, 3).addReg(callee).addSImm((int64_t)0)
2130 .addRegDef(retAddrReg);
2133 const FunctionType* funcType =
2134 cast<FunctionType>(cast<PointerType>(callee->getType())
2135 ->getElementType());
2136 bool isVarArgs = funcType->isVarArg();
2137 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
2139 // Use a descriptor to pass information about call arguments
2140 // to the register allocator. This descriptor will be "owned"
2141 // and freed automatically when the MachineCodeForInstruction
2142 // object for the callInstr goes away.
2143 CallArgsDescriptor* argDesc = new CallArgsDescriptor(callInstr,
2144 retAddrReg, isVarArgs, noPrototype);
2146 assert(callInstr->getOperand(0) == callee
2147 && "This is assumed in the loop below!");
2149 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i)
2151 Value* argVal = callInstr->getOperand(i);
2152 Instruction* intArgReg = NULL;
2154 // Check for FP arguments to varargs functions.
2155 // Any such argument in the first $K$ args must be passed in an
2156 // integer register, where K = #integer argument registers.
2157 if (isVarArgs && argVal->getType()->isFloatingPoint())
2159 // If it is a function with no prototype, pass value
2160 // as an FP value as well as a varargs value
2162 argDesc->getArgInfo(i-1).setUseFPArgReg();
2164 // If this arg. is in the first $K$ regs, add a copy
2165 // float-to-int instruction to pass the value as an integer.
2166 if (i <= target.getRegInfo().GetNumOfIntArgRegs())
2168 MachineCodeForInstruction &destMCFI =
2169 MachineCodeForInstruction::get(callInstr);
2170 intArgReg = new TmpInstruction(Type::IntTy, argVal);
2171 destMCFI.addTemp(intArgReg);
2173 std::vector<MachineInstr*> copyMvec;
2174 target.getInstrInfo().CreateCodeToCopyFloatToInt(target,
2175 callInstr->getParent()->getParent(),
2176 argVal, (TmpInstruction*) intArgReg,
2177 copyMvec, destMCFI);
2178 mvec.insert(mvec.begin(),copyMvec.begin(),copyMvec.end());
2180 argDesc->getArgInfo(i-1).setUseIntArgReg();
2181 argDesc->getArgInfo(i-1).setArgCopy(intArgReg);
2184 // Cannot fit in first $K$ regs so pass the arg on the stack
2185 argDesc->getArgInfo(i-1).setUseStackSlot();
2189 mvec.back()->addImplicitRef(intArgReg);
2191 mvec.back()->addImplicitRef(argVal);
2194 // Add the return value as an implicit ref. The call operands
2195 // were added above.
2196 if (callInstr->getType() != Type::VoidTy)
2197 mvec.back()->addImplicitRef(callInstr, /*isDef*/ true);
2199 // For the CALL instruction, the ret. addr. reg. is also implicit
2200 if (isa<Function>(callee))
2201 mvec.back()->addImplicitRef(retAddrReg, /*isDef*/ true);
2204 mvec.push_back(BuildMI(V9::NOP, 0));
2208 case 62: // reg: Shl(reg, reg)
2210 Value* argVal1 = subtreeRoot->leftChild()->getValue();
2211 Value* argVal2 = subtreeRoot->rightChild()->getValue();
2212 Instruction* shlInstr = subtreeRoot->getInstruction();
2214 const Type* opType = argVal1->getType();
2215 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2216 "Shl unsupported for other types");
2218 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
2219 (opType == Type::LongTy)? V9::SLLX : V9::SLL,
2220 argVal1, argVal2, 0, shlInstr, mvec,
2221 MachineCodeForInstruction::get(shlInstr));
2225 case 63: // reg: Shr(reg, reg)
2226 { const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
2227 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2228 "Shr unsupported for other types");
2229 Add3OperandInstr(opType->isSigned()
2230 ? (opType == Type::LongTy ? V9::SRAX : V9::SRA)
2231 : (opType == Type::LongTy ? V9::SRLX : V9::SRL),
2236 case 64: // reg: Phi(reg,reg)
2237 break; // don't forward the value
2239 case 71: // reg: VReg
2240 case 72: // reg: Constant
2241 break; // don't forward the value
2244 assert(0 && "Unrecognized BURG rule");
2249 if (forwardOperandNum >= 0)
2250 { // We did not generate a machine instruction but need to use operand.
2251 // If user is in the same tree, replace Value in its machine operand.
2252 // If not, insert a copy instruction which should get coalesced away
2253 // by register allocation.
2254 if (subtreeRoot->parent() != NULL)
2255 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
2258 std::vector<MachineInstr*> minstrVec;
2259 Instruction* instr = subtreeRoot->getInstruction();
2260 target.getInstrInfo().
2261 CreateCopyInstructionsByType(target,
2262 instr->getParent()->getParent(),
2263 instr->getOperand(forwardOperandNum),
2265 MachineCodeForInstruction::get(instr));
2266 assert(minstrVec.size() > 0);
2267 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
2271 if (maskUnsignedResult)
2272 { // If result is unsigned and smaller than int reg size,
2273 // we need to clear high bits of result value.
2274 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
2275 Instruction* dest = subtreeRoot->getInstruction();
2276 if (dest->getType()->isUnsigned())
2278 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
2280 { // Mask high bits. Use a TmpInstruction to represent the
2281 // intermediate result before masking. Since those instructions
2282 // have already been generated, go back and substitute tmpI
2283 // for dest in the result position of each one of them.
2284 TmpInstruction *tmpI = new TmpInstruction(dest->getType(), dest,
2286 MachineCodeForInstruction::get(dest).addTemp(tmpI);
2288 for (unsigned i=0, N=mvec.size(); i < N; ++i)
2289 mvec[i]->substituteValue(dest, tmpI);
2291 M = BuildMI(V9::SRL, 3).addReg(tmpI).addZImm(8*(4-destSize))
2292 .addReg(dest, MOTy::Def);
2295 else if (destSize < 8)
2296 assert(0 && "Unsupported type size: 32 < size < 64 bits");