1 //===-- SparcInstrSelection.cpp -------------------------------------------===//
3 // BURS instruction selection for SPARC V9 architecture.
5 //===----------------------------------------------------------------------===//
7 #include "SparcInternals.h"
8 #include "SparcInstrSelectionSupport.h"
9 #include "SparcRegClassInfo.h"
10 #include "llvm/CodeGen/InstrSelectionSupport.h"
11 #include "llvm/CodeGen/MachineInstrBuilder.h"
12 #include "llvm/CodeGen/MachineInstrAnnot.h"
13 #include "llvm/CodeGen/InstrForest.h"
14 #include "llvm/CodeGen/InstrSelection.h"
15 #include "llvm/CodeGen/MachineFunction.h"
16 #include "llvm/CodeGen/MachineFunctionInfo.h"
17 #include "llvm/CodeGen/MachineCodeForInstruction.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/iTerminators.h"
20 #include "llvm/iMemory.h"
21 #include "llvm/iOther.h"
22 #include "llvm/Function.h"
23 #include "llvm/Constants.h"
24 #include "llvm/ConstantHandling.h"
25 #include "llvm/Intrinsics.h"
26 #include "Support/MathExtras.h"
29 static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
30 std::vector<MachineInstr*>& mvec) {
31 mvec.push_back(BuildMI(Opcode, 3).addReg(Node->leftChild()->getValue())
32 .addReg(Node->rightChild()->getValue())
33 .addRegDef(Node->getValue()));
38 //---------------------------------------------------------------------------
39 // Function: GetMemInstArgs
42 // Get the pointer value and the index vector for a memory operation
43 // (GetElementPtr, Load, or Store). If all indices of the given memory
44 // operation are constant, fold in constant indices in a chain of
45 // preceding GetElementPtr instructions (if any), and return the
46 // pointer value of the first instruction in the chain.
47 // All folded instructions are marked so no code is generated for them.
50 // Returns the pointer Value to use.
51 // Returns the resulting IndexVector in idxVec.
52 // Returns true/false in allConstantIndices if all indices are/aren't const.
53 //---------------------------------------------------------------------------
56 //---------------------------------------------------------------------------
57 // Function: FoldGetElemChain
60 // Fold a chain of GetElementPtr instructions containing only
61 // constant offsets into an equivalent (Pointer, IndexVector) pair.
62 // Returns the pointer Value, and stores the resulting IndexVector
63 // in argument chainIdxVec. This is a helper function for
64 // FoldConstantIndices that does the actual folding.
65 //---------------------------------------------------------------------------
68 // Check for a constant 0.
72 return (idx == ConstantSInt::getNullValue(idx->getType()));
76 FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
77 bool lastInstHasLeadingNonZero)
79 InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
80 GetElementPtrInst* gepInst =
81 dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
83 // ptr value is not computed in this tree or ptr value does not come from GEP
88 // Return NULL if we don't fold any instructions in.
91 // Now chase the chain of getElementInstr instructions, if any.
92 // Check for any non-constant indices and stop there.
93 // Also, stop if the first index of child is a non-zero array index
94 // and the last index of the current node is a non-array index:
95 // in that case, a non-array declared type is being accessed as an array
96 // which is not type-safe, but could be legal.
98 InstructionNode* ptrChild = gepNode;
99 while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
100 ptrChild->getOpLabel() == GetElemPtrIdx))
102 // Child is a GetElemPtr instruction
103 gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
104 User::op_iterator OI, firstIdx = gepInst->idx_begin();
105 User::op_iterator lastIdx = gepInst->idx_end();
106 bool allConstantOffsets = true;
108 // The first index of every GEP must be an array index.
109 assert((*firstIdx)->getType() == Type::LongTy &&
110 "INTERNAL ERROR: Structure index for a pointer type!");
112 // If the last instruction had a leading non-zero index, check if the
113 // current one references a sequential (i.e., indexable) type.
114 // If not, the code is not type-safe and we would create an illegal GEP
115 // by folding them, so don't fold any more instructions.
117 if (lastInstHasLeadingNonZero)
118 if (! isa<SequentialType>(gepInst->getType()->getElementType()))
119 break; // cannot fold in any preceding getElementPtr instrs.
121 // Check that all offsets are constant for this instruction
122 for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
123 allConstantOffsets = isa<ConstantInt>(*OI);
125 if (allConstantOffsets) {
126 // Get pointer value out of ptrChild.
127 ptrVal = gepInst->getPointerOperand();
129 // Remember if it has leading zero index: it will be discarded later.
130 lastInstHasLeadingNonZero = ! IsZero(*firstIdx);
132 // Insert its index vector at the start, skipping any leading [0]
133 chainIdxVec.insert(chainIdxVec.begin(),
134 firstIdx + !lastInstHasLeadingNonZero, lastIdx);
136 // Mark the folded node so no code is generated for it.
137 ((InstructionNode*) ptrChild)->markFoldedIntoParent();
139 // Get the previous GEP instruction and continue trying to fold
140 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)) {
194 // Append the index vector of the current instruction.
195 // Skip the leading [0] index if preceding GEPs were folded into this.
196 idxVec.insert(idxVec.end(),
197 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
203 //---------------------------------------------------------------------------
204 // Function: GetMemInstArgs
207 // Get the pointer value and the index vector for a memory operation
208 // (GetElementPtr, Load, or Store). If all indices of the given memory
209 // operation are constant, fold in constant indices in a chain of
210 // preceding GetElementPtr instructions (if any), and return the
211 // pointer value of the first instruction in the chain.
212 // All folded instructions are marked so no code is generated for them.
215 // Returns the pointer Value to use.
216 // Returns the resulting IndexVector in idxVec.
217 // Returns true/false in allConstantIndices if all indices are/aren't const.
218 //---------------------------------------------------------------------------
221 GetMemInstArgs(InstructionNode* memInstrNode,
222 std::vector<Value*>& idxVec,
223 bool& allConstantIndices)
225 allConstantIndices = false;
226 Instruction* memInst = memInstrNode->getInstruction();
227 assert(idxVec.size() == 0 && "Need empty vector to return indices");
229 // If there is a GetElemPtr instruction to fold in to this instr,
230 // it must be in the left child for Load and GetElemPtr, and in the
231 // right child for Store instructions.
232 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
233 ? memInstrNode->rightChild()
234 : memInstrNode->leftChild());
236 // Default pointer is the one from the current instruction.
237 Value* ptrVal = ptrChild->getValue();
239 // Find the "last" GetElemPtr instruction: this one or the immediate child.
240 // There will be none if this is a load or a store from a scalar pointer.
241 InstructionNode* gepNode = NULL;
242 if (isa<GetElementPtrInst>(memInst))
243 gepNode = memInstrNode;
244 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
245 // Child of load/store is a GEP and memInst is its only use.
246 // Use its indices and mark it as folded.
247 gepNode = cast<InstructionNode>(ptrChild);
248 gepNode->markFoldedIntoParent();
251 // If there are no indices, return the current pointer.
252 // Else extract the pointer from the GEP and fold the indices.
253 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
258 //************************ Internal Functions ******************************/
261 static inline MachineOpCode
262 ChooseBprInstruction(const InstructionNode* instrNode)
264 MachineOpCode opCode;
266 Instruction* setCCInstr =
267 ((InstructionNode*) instrNode->leftChild())->getInstruction();
269 switch(setCCInstr->getOpcode())
271 case Instruction::SetEQ: opCode = V9::BRZ; break;
272 case Instruction::SetNE: opCode = V9::BRNZ; break;
273 case Instruction::SetLE: opCode = V9::BRLEZ; break;
274 case Instruction::SetGE: opCode = V9::BRGEZ; break;
275 case Instruction::SetLT: opCode = V9::BRLZ; break;
276 case Instruction::SetGT: opCode = V9::BRGZ; break;
278 assert(0 && "Unrecognized VM instruction!");
279 opCode = V9::INVALID_OPCODE;
287 static inline MachineOpCode
288 ChooseBpccInstruction(const InstructionNode* instrNode,
289 const BinaryOperator* setCCInstr)
291 MachineOpCode opCode = V9::INVALID_OPCODE;
293 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
296 switch(setCCInstr->getOpcode())
298 case Instruction::SetEQ: opCode = V9::BE; break;
299 case Instruction::SetNE: opCode = V9::BNE; break;
300 case Instruction::SetLE: opCode = V9::BLE; break;
301 case Instruction::SetGE: opCode = V9::BGE; break;
302 case Instruction::SetLT: opCode = V9::BL; break;
303 case Instruction::SetGT: opCode = V9::BG; break;
305 assert(0 && "Unrecognized VM instruction!");
309 switch(setCCInstr->getOpcode())
311 case Instruction::SetEQ: opCode = V9::BE; break;
312 case Instruction::SetNE: opCode = V9::BNE; break;
313 case Instruction::SetLE: opCode = V9::BLEU; break;
314 case Instruction::SetGE: opCode = V9::BCC; break;
315 case Instruction::SetLT: opCode = V9::BCS; break;
316 case Instruction::SetGT: opCode = V9::BGU; break;
318 assert(0 && "Unrecognized VM instruction!");
326 static inline MachineOpCode
327 ChooseBFpccInstruction(const InstructionNode* instrNode,
328 const BinaryOperator* setCCInstr)
330 MachineOpCode opCode = V9::INVALID_OPCODE;
332 switch(setCCInstr->getOpcode())
334 case Instruction::SetEQ: opCode = V9::FBE; break;
335 case Instruction::SetNE: opCode = V9::FBNE; break;
336 case Instruction::SetLE: opCode = V9::FBLE; break;
337 case Instruction::SetGE: opCode = V9::FBGE; break;
338 case Instruction::SetLT: opCode = V9::FBL; break;
339 case Instruction::SetGT: opCode = V9::FBG; break;
341 assert(0 && "Unrecognized VM instruction!");
349 // Create a unique TmpInstruction for a boolean value,
350 // representing the CC register used by a branch on that value.
351 // For now, hack this using a little static cache of TmpInstructions.
352 // Eventually the entire BURG instruction selection should be put
353 // into a separate class that can hold such information.
354 // The static cache is not too bad because the memory for these
355 // TmpInstructions will be freed along with the rest of the Function anyway.
357 static TmpInstruction*
358 GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType)
360 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
361 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
362 static const Function *lastFunction = 0;// Use to flush cache between funcs
364 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
366 if (lastFunction != F) {
368 boolToTmpCache.clear();
371 // Look for tmpI and create a new one otherwise. The new value is
372 // directly written to map using the ref returned by operator[].
373 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
375 tmpI = new TmpInstruction(ccType, boolVal);
381 static inline MachineOpCode
382 ChooseBccInstruction(const InstructionNode* instrNode,
385 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
386 assert(setCCNode->getOpLabel() == SetCCOp);
387 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
388 const Type* setCCType = setCCInstr->getOperand(0)->getType();
390 isFPBranch = setCCType->isFloatingPoint(); // Return value: don't delete!
393 return ChooseBFpccInstruction(instrNode, setCCInstr);
395 return ChooseBpccInstruction(instrNode, setCCInstr);
399 static inline MachineOpCode
400 ChooseMovFpccInstruction(const InstructionNode* instrNode)
402 MachineOpCode opCode = V9::INVALID_OPCODE;
404 switch(instrNode->getInstruction()->getOpcode())
406 case Instruction::SetEQ: opCode = V9::MOVFE; break;
407 case Instruction::SetNE: opCode = V9::MOVFNE; break;
408 case Instruction::SetLE: opCode = V9::MOVFLE; break;
409 case Instruction::SetGE: opCode = V9::MOVFGE; break;
410 case Instruction::SetLT: opCode = V9::MOVFL; break;
411 case Instruction::SetGT: opCode = V9::MOVFG; break;
413 assert(0 && "Unrecognized VM instruction!");
421 // Assumes that SUBcc v1, v2 -> v3 has been executed.
422 // In most cases, we want to clear v3 and then follow it by instruction
424 // Set mustClearReg=false if v3 need not be cleared before conditional move.
425 // Set valueToMove=0 if we want to conditionally move 0 instead of 1
426 // (i.e., we want to test inverse of a condition)
427 // (The latter two cases do not seem to arise because SetNE needs nothing.)
430 ChooseMovpccAfterSub(const InstructionNode* instrNode,
434 MachineOpCode opCode = V9::INVALID_OPCODE;
438 switch(instrNode->getInstruction()->getOpcode())
440 case Instruction::SetEQ: opCode = V9::MOVE; break;
441 case Instruction::SetLE: opCode = V9::MOVLE; break;
442 case Instruction::SetGE: opCode = V9::MOVGE; break;
443 case Instruction::SetLT: opCode = V9::MOVL; break;
444 case Instruction::SetGT: opCode = V9::MOVG; break;
445 case Instruction::SetNE: assert(0 && "No move required!"); break;
446 default: assert(0 && "Unrecognized VM instr!"); break;
452 static inline MachineOpCode
453 ChooseConvertToFloatInstr(OpLabel vopCode, const Type* opType)
455 MachineOpCode opCode = V9::INVALID_OPCODE;
460 if (opType == Type::SByteTy || opType == Type::ShortTy ||
461 opType == Type::IntTy)
463 else if (opType == Type::LongTy)
465 else if (opType == Type::DoubleTy)
467 else if (opType == Type::FloatTy)
470 assert(0 && "Cannot convert this type to FLOAT on SPARC");
474 // This is usually used in conjunction with CreateCodeToCopyIntToFloat().
475 // Both functions should treat the integer as a 32-bit value for types
476 // of 4 bytes or less, and as a 64-bit value otherwise.
477 if (opType == Type::SByteTy || opType == Type::UByteTy ||
478 opType == Type::ShortTy || opType == Type::UShortTy ||
479 opType == Type::IntTy || opType == Type::UIntTy)
481 else if (opType == Type::LongTy || opType == Type::ULongTy)
483 else if (opType == Type::FloatTy)
485 else if (opType == Type::DoubleTy)
488 assert(0 && "Cannot convert this type to DOUBLE on SPARC");
498 static inline MachineOpCode
499 ChooseConvertFPToIntInstr(Type::PrimitiveID tid, const Type* opType)
501 MachineOpCode opCode = V9::INVALID_OPCODE;;
503 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
504 && "This function should only be called for FLOAT or DOUBLE");
506 if (tid == Type::UIntTyID) {
507 assert(tid != Type::UIntTyID && "FP-to-uint conversions must be expanded"
508 " into FP->long->uint for SPARC v9: SO RUN PRESELECTION PASS!");
509 } else if (tid == Type::SByteTyID || tid == Type::ShortTyID ||
510 tid == Type::IntTyID || tid == Type::UByteTyID ||
511 tid == Type::UShortTyID) {
512 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
513 } else if (tid == Type::LongTyID || tid == Type::ULongTyID) {
514 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
516 assert(0 && "Should not get here, Mo!");
522 CreateConvertFPToIntInstr(Type::PrimitiveID destTID,
523 Value* srcVal, Value* destVal)
525 MachineOpCode opCode = ChooseConvertFPToIntInstr(destTID, srcVal->getType());
526 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
527 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
530 // CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer
531 // The FP value must be converted to the dest type in an FP register,
532 // and the result is then copied from FP to int register via memory.
534 // Since fdtoi converts to signed integers, any FP value V between MAXINT+1
535 // and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly
536 // *only* when converting to an unsigned. (Unsigned byte, short or long
537 // don't have this problem.)
538 // For unsigned int, we therefore have to generate the code sequence:
540 // if (V > (float) MAXINT) {
541 // unsigned result = (unsigned) (V - (float) MAXINT);
542 // result = result + (unsigned) MAXINT;
545 // result = (unsigned) V;
548 CreateCodeToConvertFloatToInt(const TargetMachine& target,
551 std::vector<MachineInstr*>& mvec,
552 MachineCodeForInstruction& mcfi)
554 // Create a temporary to represent the FP register into which the
555 // int value will placed after conversion. The type of this temporary
556 // depends on the type of FP register to use: single-prec for a 32-bit
557 // int or smaller; double-prec for a 64-bit int.
559 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
560 const Type* destTypeToUse = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
561 TmpInstruction* destForCast = new TmpInstruction(destTypeToUse, opVal);
562 mcfi.addTemp(destForCast);
564 // Create the fp-to-int conversion code
565 MachineInstr* M =CreateConvertFPToIntInstr(destI->getType()->getPrimitiveID(),
569 // Create the fpreg-to-intreg copy code
570 target.getInstrInfo().
571 CreateCodeToCopyFloatToInt(target, destI->getParent()->getParent(),
572 destForCast, destI, mvec, mcfi);
576 static inline MachineOpCode
577 ChooseAddInstruction(const InstructionNode* instrNode)
579 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
583 static inline MachineInstr*
584 CreateMovFloatInstruction(const InstructionNode* instrNode,
585 const Type* resultType)
587 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
588 .addReg(instrNode->leftChild()->getValue())
589 .addRegDef(instrNode->getValue());
592 static inline MachineInstr*
593 CreateAddConstInstruction(const InstructionNode* instrNode)
595 MachineInstr* minstr = NULL;
597 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
598 assert(isa<Constant>(constOp));
600 // Cases worth optimizing are:
601 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
602 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
604 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
605 double dval = FPC->getValue();
607 minstr = CreateMovFloatInstruction(instrNode,
608 instrNode->getInstruction()->getType());
615 static inline MachineOpCode
616 ChooseSubInstructionByType(const Type* resultType)
618 MachineOpCode opCode = V9::INVALID_OPCODE;
620 if (resultType->isInteger() || isa<PointerType>(resultType)) {
623 switch(resultType->getPrimitiveID())
625 case Type::FloatTyID: opCode = V9::FSUBS; break;
626 case Type::DoubleTyID: opCode = V9::FSUBD; break;
627 default: assert(0 && "Invalid type for SUB instruction"); break;
635 static inline MachineInstr*
636 CreateSubConstInstruction(const InstructionNode* instrNode)
638 MachineInstr* minstr = NULL;
640 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
641 assert(isa<Constant>(constOp));
643 // Cases worth optimizing are:
644 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
645 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
647 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
648 double dval = FPC->getValue();
650 minstr = CreateMovFloatInstruction(instrNode,
651 instrNode->getInstruction()->getType());
658 static inline MachineOpCode
659 ChooseFcmpInstruction(const InstructionNode* instrNode)
661 MachineOpCode opCode = V9::INVALID_OPCODE;
663 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
664 switch(operand->getType()->getPrimitiveID()) {
665 case Type::FloatTyID: opCode = V9::FCMPS; break;
666 case Type::DoubleTyID: opCode = V9::FCMPD; break;
667 default: assert(0 && "Invalid type for FCMP instruction"); break;
674 // Assumes that leftArg and rightArg are both cast instructions.
677 BothFloatToDouble(const InstructionNode* instrNode)
679 InstrTreeNode* leftArg = instrNode->leftChild();
680 InstrTreeNode* rightArg = instrNode->rightChild();
681 InstrTreeNode* leftArgArg = leftArg->leftChild();
682 InstrTreeNode* rightArgArg = rightArg->leftChild();
683 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
685 // Check if both arguments are floats cast to double
686 return (leftArg->getValue()->getType() == Type::DoubleTy &&
687 leftArgArg->getValue()->getType() == Type::FloatTy &&
688 rightArgArg->getValue()->getType() == Type::FloatTy);
692 static inline MachineOpCode
693 ChooseMulInstructionByType(const Type* resultType)
695 MachineOpCode opCode = V9::INVALID_OPCODE;
697 if (resultType->isInteger())
700 switch(resultType->getPrimitiveID())
702 case Type::FloatTyID: opCode = V9::FMULS; break;
703 case Type::DoubleTyID: opCode = V9::FMULD; break;
704 default: assert(0 && "Invalid type for MUL instruction"); break;
712 static inline MachineInstr*
713 CreateIntNegInstruction(const TargetMachine& target,
716 return BuildMI(V9::SUBr, 3).addMReg(target.getRegInfo().getZeroRegNum())
717 .addReg(vreg).addRegDef(vreg);
721 // Create instruction sequence for any shift operation.
722 // SLL or SLLX on an operand smaller than the integer reg. size (64bits)
723 // requires a second instruction for explicit sign-extension.
724 // Note that we only have to worry about a sign-bit appearing in the
725 // most significant bit of the operand after shifting (e.g., bit 32 of
726 // Int or bit 16 of Short), so we do not have to worry about results
727 // that are as large as a normal integer register.
730 CreateShiftInstructions(const TargetMachine& target,
732 MachineOpCode shiftOpCode,
734 Value* optArgVal2, /* Use optArgVal2 if not NULL */
735 unsigned optShiftNum, /* else use optShiftNum */
736 Instruction* destVal,
737 std::vector<MachineInstr*>& mvec,
738 MachineCodeForInstruction& mcfi)
740 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
741 "Large shift sizes unexpected, but can be handled below: "
742 "You need to check whether or not it fits in immed field below");
744 // If this is a logical left shift of a type smaller than the standard
745 // integer reg. size, we have to extend the sign-bit into upper bits
746 // of dest, so we need to put the result of the SLL into a temporary.
748 Value* shiftDest = destVal;
749 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
750 if ((shiftOpCode == V9::SLLr6 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
751 // put SLL result into a temporary
752 shiftDest = new TmpInstruction(argVal1, optArgVal2, "sllTmp");
753 mcfi.addTemp(shiftDest);
756 MachineInstr* M = (optArgVal2 != NULL)
757 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
758 .addReg(shiftDest, MOTy::Def)
759 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
760 .addReg(shiftDest, MOTy::Def);
763 if (shiftDest != destVal) {
764 // extend the sign-bit of the result into all upper bits of dest
765 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
766 target.getInstrInfo().
767 CreateSignExtensionInstructions(target, F, shiftDest, destVal,
768 8*opSize, mvec, mcfi);
773 // Does not create any instructions if we cannot exploit constant to
774 // create a cheaper instruction.
775 // This returns the approximate cost of the instructions generated,
776 // which is used to pick the cheapest when both operands are constant.
778 CreateMulConstInstruction(const TargetMachine &target, Function* F,
779 Value* lval, Value* rval, Instruction* destVal,
780 std::vector<MachineInstr*>& mvec,
781 MachineCodeForInstruction& mcfi)
783 /* Use max. multiply cost, viz., cost of MULX */
784 unsigned cost = target.getInstrInfo().minLatency(V9::MULXr);
785 unsigned firstNewInstr = mvec.size();
787 Value* constOp = rval;
788 if (! isa<Constant>(constOp))
791 // Cases worth optimizing are:
792 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
793 // (2) Multiply by 2^x for integer types: replace with Shift
795 const Type* resultType = destVal->getType();
797 if (resultType->isInteger() || isa<PointerType>(resultType)) {
799 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
802 bool needNeg = false;
808 if (C == 0 || C == 1) {
809 cost = target.getInstrInfo().minLatency(V9::ADDr);
810 unsigned Zero = target.getRegInfo().getZeroRegNum();
813 M =BuildMI(V9::ADDr,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
815 M = BuildMI(V9::ADDr,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::SLLr6 : V9::SLLXr6;
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),
868 cast<Constant>(rval));
869 target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi);
871 else if (isa<Constant>(rval)) // rval is constant, but not lval
872 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
873 else if (isa<Constant>(lval)) // lval is constant, but not rval
874 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
876 // else neither is constant
880 // Return NULL if we cannot exploit constant to create a cheaper instruction
882 CreateMulInstruction(const TargetMachine &target, Function* F,
883 Value* lval, Value* rval, Instruction* destVal,
884 std::vector<MachineInstr*>& mvec,
885 MachineCodeForInstruction& mcfi,
886 MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE)
888 unsigned L = mvec.size();
889 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
890 if (mvec.size() == L) {
891 // no instructions were added so create MUL reg, reg, reg.
892 // Use FSMULD if both operands are actually floats cast to doubles.
893 // Otherwise, use the default opcode for the appropriate type.
894 MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE)
896 : ChooseMulInstructionByType(destVal->getType()));
897 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
898 .addRegDef(destVal));
903 // Generate a divide instruction for Div or Rem.
904 // For Rem, this assumes that the operand type will be signed if the result
905 // type is signed. This is correct because they must have the same sign.
907 static inline MachineOpCode
908 ChooseDivInstruction(TargetMachine &target,
909 const InstructionNode* instrNode)
911 MachineOpCode opCode = V9::INVALID_OPCODE;
913 const Type* resultType = instrNode->getInstruction()->getType();
915 if (resultType->isInteger())
916 opCode = resultType->isSigned()? V9::SDIVXr : V9::UDIVXr;
918 switch(resultType->getPrimitiveID())
920 case Type::FloatTyID: opCode = V9::FDIVS; break;
921 case Type::DoubleTyID: opCode = V9::FDIVD; break;
922 default: assert(0 && "Invalid type for DIV instruction"); break;
929 // Return if we cannot exploit constant to create a cheaper instruction
931 CreateDivConstInstruction(TargetMachine &target,
932 const InstructionNode* instrNode,
933 std::vector<MachineInstr*>& mvec)
935 Value* LHS = instrNode->leftChild()->getValue();
936 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
937 if (!isa<Constant>(constOp))
940 Instruction* destVal = instrNode->getInstruction();
941 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
943 // Cases worth optimizing are:
944 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
945 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
947 const Type* resultType = instrNode->getInstruction()->getType();
949 if (resultType->isInteger()) {
952 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
954 bool needNeg = false;
961 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
962 .addRegDef(destVal));
963 } else if (isPowerOf2(C, pow)) {
967 if (resultType->isSigned()) {
968 // The result may be negative and we need to add one before shifting
969 // a negative value. Use:
970 // srl i0, 31, x0; add x0, i0, i1 (if i0 is <= 32 bits)
972 // srlx i0, 63, x0; add x0, i0, i1 (if i0 is 64 bits)
973 // to compute i1=i0+1 if i0 < 0 and i1=i0 otherwise.
975 TmpInstruction *srlTmp, *addTmp;
976 MachineCodeForInstruction& mcfi
977 = MachineCodeForInstruction::get(destVal);
978 srlTmp = new TmpInstruction(resultType, LHS, 0, "getSign");
979 addTmp = new TmpInstruction(resultType, LHS, srlTmp, "incIfNeg");
980 mcfi.addTemp(srlTmp);
981 mcfi.addTemp(addTmp);
983 // Create the SRL or SRLX instruction to get the sign bit
984 mvec.push_back(BuildMI((resultType==Type::LongTy) ?
985 V9::SRLXi6 : V9::SRLi6, 3)
987 .addSImm((resultType==Type::LongTy)? 63 : 31)
990 // Create the ADD instruction to add 1 for negative values
991 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
994 // Get the shift operand and "right-shift" opcode to do the divide
995 shiftOperand = addTmp;
996 opCode = (resultType==Type::LongTy) ? V9::SRAXi6 : V9::SRAi6;
998 // Get the shift operand and "right-shift" opcode to do the divide
1000 opCode = (resultType==Type::LongTy) ? V9::SRLXi6 : V9::SRLi6;
1003 // Now do the actual shift!
1004 mvec.push_back(BuildMI(opCode, 3).addReg(shiftOperand).addZImm(pow)
1005 .addRegDef(destVal));
1008 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
1009 // insert <reg = SUB 0, reg> after the instr to flip the sign
1010 mvec.push_back(CreateIntNegInstruction(target, destVal));
1014 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
1015 double dval = FPC->getValue();
1016 if (fabs(dval) == 1) {
1018 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
1019 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
1021 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(destVal));
1029 CreateCodeForVariableSizeAlloca(const TargetMachine& target,
1030 Instruction* result,
1032 Value* numElementsVal,
1033 std::vector<MachineInstr*>& getMvec)
1035 Value* totalSizeVal;
1037 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
1038 Function *F = result->getParent()->getParent();
1040 // Enforce the alignment constraints on the stack pointer at
1041 // compile time if the total size is a known constant.
1042 if (isa<Constant>(numElementsVal)) {
1044 int64_t numElem = GetConstantValueAsSignedInt(numElementsVal, isValid);
1045 assert(isValid && "Unexpectedly large array dimension in alloca!");
1046 int64_t total = numElem * tsize;
1047 if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
1048 total += target.getFrameInfo().getStackFrameSizeAlignment() - extra;
1049 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
1051 // The size is not a constant. Generate code to compute it and
1052 // code to pad the size for stack alignment.
1053 // Create a Value to hold the (constant) element size
1054 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
1056 // Create temporary values to hold the result of MUL, SLL, SRL
1057 // THIS CASE IS INCOMPLETE AND WILL BE FIXED SHORTLY.
1058 TmpInstruction* tmpProd = new TmpInstruction(numElementsVal, tsizeVal);
1059 TmpInstruction* tmpSLL = new TmpInstruction(numElementsVal, tmpProd);
1060 TmpInstruction* tmpSRL = new TmpInstruction(numElementsVal, tmpSLL);
1061 mcfi.addTemp(tmpProd);
1062 mcfi.addTemp(tmpSLL);
1063 mcfi.addTemp(tmpSRL);
1065 // Instruction 1: mul numElements, typeSize -> tmpProd
1066 // This will optimize the MUL as far as possible.
1067 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd,getMvec,
1068 mcfi, INVALID_MACHINE_OPCODE);
1070 assert(0 && "Need to insert padding instructions here!");
1072 totalSizeVal = tmpProd;
1075 // Get the constant offset from SP for dynamically allocated storage
1076 // and create a temporary Value to hold it.
1077 MachineFunction& mcInfo = MachineFunction::get(F);
1079 ConstantSInt* dynamicAreaOffset =
1080 ConstantSInt::get(Type::IntTy,
1081 target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp));
1082 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
1084 unsigned SPReg = target.getRegInfo().getStackPointer();
1086 // Instruction 2: sub %sp, totalSizeVal -> %sp
1087 getMvec.push_back(BuildMI(V9::SUBr, 3).addMReg(SPReg).addReg(totalSizeVal)
1088 .addMReg(SPReg,MOTy::Def));
1090 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
1091 getMvec.push_back(BuildMI(V9::ADDr,3).addMReg(SPReg).addReg(dynamicAreaOffset)
1092 .addRegDef(result));
1097 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
1098 Instruction* result,
1100 unsigned numElements,
1101 std::vector<MachineInstr*>& getMvec)
1103 assert(tsize > 0 && "Illegal (zero) type size for alloca");
1104 assert(result && result->getParent() &&
1105 "Result value is not part of a function?");
1106 Function *F = result->getParent()->getParent();
1107 MachineFunction &mcInfo = MachineFunction::get(F);
1109 // Check if the offset would small enough to use as an immediate in
1110 // load/stores (check LDX because all load/stores have the same-size immediate
1111 // field). If not, put the variable in the dynamically sized area of the
1113 unsigned paddedSizeIgnored;
1114 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
1116 tsize * numElements);
1117 if (! target.getInstrInfo().constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
1118 CreateCodeForVariableSizeAlloca(target, result, tsize,
1119 ConstantSInt::get(Type::IntTy,numElements),
1124 // else offset fits in immediate field so go ahead and allocate it.
1125 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
1127 // Create a temporary Value to hold the constant offset.
1128 // This is needed because it may not fit in the immediate field.
1129 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
1131 // Instruction 1: add %fp, offsetFromFP -> result
1132 unsigned FPReg = target.getRegInfo().getFramePointer();
1133 getMvec.push_back(BuildMI(V9::ADDr, 3).addMReg(FPReg).addReg(offsetVal)
1134 .addRegDef(result));
1138 //------------------------------------------------------------------------
1139 // Function SetOperandsForMemInstr
1141 // Choose addressing mode for the given load or store instruction.
1142 // Use [reg+reg] if it is an indexed reference, and the index offset is
1143 // not a constant or if it cannot fit in the offset field.
1144 // Use [reg+offset] in all other cases.
1146 // This assumes that all array refs are "lowered" to one of these forms:
1147 // %x = load (subarray*) ptr, constant ; single constant offset
1148 // %x = load (subarray*) ptr, offsetVal ; single non-constant offset
1149 // Generally, this should happen via strength reduction + LICM.
1150 // Also, strength reduction should take care of using the same register for
1151 // the loop index variable and an array index, when that is profitable.
1152 //------------------------------------------------------------------------
1155 SetOperandsForMemInstr(unsigned Opcode,
1156 std::vector<MachineInstr*>& mvec,
1157 InstructionNode* vmInstrNode,
1158 const TargetMachine& target)
1160 Instruction* memInst = vmInstrNode->getInstruction();
1161 // Index vector, ptr value, and flag if all indices are const.
1162 std::vector<Value*> idxVec;
1163 bool allConstantIndices;
1164 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
1166 // Now create the appropriate operands for the machine instruction.
1167 // First, initialize so we default to storing the offset in a register.
1168 int64_t smallConstOffset = 0;
1169 Value* valueForRegOffset = NULL;
1170 MachineOperand::MachineOperandType offsetOpType =
1171 MachineOperand::MO_VirtualRegister;
1173 // Check if there is an index vector and if so, compute the
1174 // right offset for structures and for arrays
1176 if (!idxVec.empty()) {
1177 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
1179 // If all indices are constant, compute the combined offset directly.
1180 if (allConstantIndices) {
1181 // Compute the offset value using the index vector. Create a
1182 // virtual reg. for it since it may not fit in the immed field.
1183 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
1184 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
1186 // There is at least one non-constant offset. Therefore, this must
1187 // be an array ref, and must have been lowered to a single non-zero
1188 // offset. (An extra leading zero offset, if any, can be ignored.)
1189 // Generate code sequence to compute address from index.
1191 bool firstIdxIsZero = IsZero(idxVec[0]);
1192 assert(idxVec.size() == 1U + firstIdxIsZero
1193 && "Array refs must be lowered before Instruction Selection");
1195 Value* idxVal = idxVec[firstIdxIsZero];
1197 std::vector<MachineInstr*> mulVec;
1198 Instruction* addr = new TmpInstruction(Type::ULongTy, memInst);
1199 MachineCodeForInstruction::get(memInst).addTemp(addr);
1201 // Get the array type indexed by idxVal, and compute its element size.
1202 // The call to getTypeSize() will fail if size is not constant.
1203 const Type* vecType = (firstIdxIsZero
1204 ? GetElementPtrInst::getIndexedType(ptrType,
1205 std::vector<Value*>(1U, idxVec[0]),
1206 /*AllowCompositeLeaf*/ true)
1208 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
1209 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
1210 target.getTargetData().getTypeSize(eltType));
1212 // CreateMulInstruction() folds constants intelligently enough.
1213 CreateMulInstruction(target, memInst->getParent()->getParent(),
1214 idxVal, /* lval, not likely to be const*/
1215 eltSizeVal, /* rval, likely to be constant */
1217 mulVec, MachineCodeForInstruction::get(memInst),
1218 INVALID_MACHINE_OPCODE);
1220 assert(mulVec.size() > 0 && "No multiply code created?");
1221 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
1223 valueForRegOffset = addr;
1226 offsetOpType = MachineOperand::MO_SignExtendedImmed;
1227 smallConstOffset = 0;
1231 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
1232 // For LOAD or GET_ELEMENT_PTR,
1233 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
1235 unsigned offsetOpNum, ptrOpNum;
1237 if (memInst->getOpcode() == Instruction::Store) {
1238 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1239 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1240 .addReg(ptrVal).addReg(valueForRegOffset);
1242 Opcode = convertOpcodeFromRegToImm(Opcode);
1243 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1244 .addReg(ptrVal).addSImm(smallConstOffset);
1247 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1248 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
1249 .addRegDef(memInst);
1251 Opcode = convertOpcodeFromRegToImm(Opcode);
1252 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
1253 .addRegDef(memInst);
1261 // Substitute operand `operandNum' of the instruction in node `treeNode'
1262 // in place of the use(s) of that instruction in node `parent'.
1263 // Check both explicit and implicit operands!
1264 // Also make sure to skip over a parent who:
1265 // (1) is a list node in the Burg tree, or
1266 // (2) itself had its results forwarded to its parent
1269 ForwardOperand(InstructionNode* treeNode,
1270 InstrTreeNode* parent,
1273 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
1275 Instruction* unusedOp = treeNode->getInstruction();
1276 Value* fwdOp = unusedOp->getOperand(operandNum);
1278 // The parent itself may be a list node, so find the real parent instruction
1279 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode)
1281 parent = parent->parent();
1282 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
1284 InstructionNode* parentInstrNode = (InstructionNode*) parent;
1286 Instruction* userInstr = parentInstrNode->getInstruction();
1287 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
1289 // The parent's mvec would be empty if it was itself forwarded.
1290 // Recursively call ForwardOperand in that case...
1292 if (mvec.size() == 0) {
1293 assert(parent->parent() != NULL &&
1294 "Parent could not have been forwarded, yet has no instructions?");
1295 ForwardOperand(treeNode, parent->parent(), operandNum);
1297 for (unsigned i=0, N=mvec.size(); i < N; i++) {
1298 MachineInstr* minstr = mvec[i];
1299 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) {
1300 const MachineOperand& mop = minstr->getOperand(i);
1301 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
1302 mop.getVRegValue() == unusedOp)
1304 minstr->SetMachineOperandVal(i, MachineOperand::MO_VirtualRegister,
1309 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
1310 if (minstr->getImplicitRef(i) == unusedOp) {
1311 minstr->setImplicitRef(i, fwdOp,
1312 minstr->getImplicitOp(i).opIsDefOnly(),
1313 minstr->getImplicitOp(i).opIsDefAndUse());
1321 AllUsesAreBranches(const Instruction* setccI)
1323 for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end();
1325 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
1326 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
1331 // Generate code for any intrinsic that needs a special code sequence
1332 // instead of a regular call. If not that kind of intrinsic, do nothing.
1333 // Returns true if code was generated, otherwise false.
1335 bool CodeGenIntrinsic(LLVMIntrinsic::ID iid, CallInst &callInstr,
1336 TargetMachine &target,
1337 std::vector<MachineInstr*>& mvec)
1340 case LLVMIntrinsic::va_start: {
1341 // Get the address of the first vararg value on stack and copy it to
1342 // the argument of va_start(va_list* ap).
1344 Function* func = cast<Function>(callInstr.getParent()->getParent());
1345 int numFixedArgs = func->getFunctionType()->getNumParams();
1346 int fpReg = target.getFrameInfo().getIncomingArgBaseRegNum();
1347 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
1348 int firstVarArgOff = numFixedArgs * argSize + target.getFrameInfo().
1349 getFirstIncomingArgOffset(MachineFunction::get(func), ignore);
1350 mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
1351 addReg(callInstr.getOperand(1)));
1355 case LLVMIntrinsic::va_end:
1356 return true; // no-op on Sparc
1358 case LLVMIntrinsic::va_copy:
1359 // Simple copy of current va_list (arg2) to new va_list (arg1)
1360 mvec.push_back(BuildMI(V9::ORr, 3).
1361 addMReg(target.getRegInfo().getZeroRegNum()).
1362 addReg(callInstr.getOperand(2)).
1363 addReg(callInstr.getOperand(1)));
1371 //******************* Externally Visible Functions *************************/
1373 //------------------------------------------------------------------------
1374 // External Function: ThisIsAChainRule
1377 // Check if a given BURG rule is a chain rule.
1378 //------------------------------------------------------------------------
1381 ThisIsAChainRule(int eruleno)
1385 case 111: // stmt: reg
1409 return false; break;
1414 //------------------------------------------------------------------------
1415 // External Function: GetInstructionsByRule
1418 // Choose machine instructions for the SPARC according to the
1419 // patterns chosen by the BURG-generated parser.
1420 //------------------------------------------------------------------------
1423 GetInstructionsByRule(InstructionNode* subtreeRoot,
1426 TargetMachine &target,
1427 std::vector<MachineInstr*>& mvec)
1429 bool checkCast = false; // initialize here to use fall-through
1430 bool maskUnsignedResult = false;
1432 int forwardOperandNum = -1;
1433 unsigned allocaSize = 0;
1434 MachineInstr* M, *M2;
1439 // If the code for this instruction was folded into the parent (user),
1441 if (subtreeRoot->isFoldedIntoParent())
1445 // Let's check for chain rules outside the switch so that we don't have
1446 // to duplicate the list of chain rule production numbers here again
1448 if (ThisIsAChainRule(ruleForNode)) {
1449 // Chain rules have a single nonterminal on the RHS.
1450 // Get the rule that matches the RHS non-terminal and use that instead.
1452 assert(nts[0] && ! nts[1]
1453 && "A chain rule should have only one RHS non-terminal!");
1454 nextRule = burm_rule(subtreeRoot->state, nts[0]);
1455 nts = burm_nts[nextRule];
1456 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
1458 switch(ruleForNode) {
1459 case 1: // stmt: Ret
1460 case 2: // stmt: RetValue(reg)
1461 { // NOTE: Prepass of register allocation is responsible
1462 // for moving return value to appropriate register.
1463 // Mark the return-address register as a hidden virtual reg.
1464 // Mark the return value register as an implicit ref of
1465 // the machine instruction.
1466 // Finally put a NOP in the delay slot.
1467 ReturnInst *returnInstr =
1468 cast<ReturnInst>(subtreeRoot->getInstruction());
1469 assert(returnInstr->getOpcode() == Instruction::Ret);
1471 Instruction* returnReg = new TmpInstruction(returnInstr);
1472 MachineCodeForInstruction::get(returnInstr).addTemp(returnReg);
1474 M = BuildMI(V9::JMPLRETi, 3).addReg(returnReg).addSImm(8)
1475 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
1477 if (returnInstr->getReturnValue() != NULL)
1478 M->addImplicitRef(returnInstr->getReturnValue());
1481 mvec.push_back(BuildMI(V9::NOP, 0));
1486 case 3: // stmt: Store(reg,reg)
1487 case 4: // stmt: Store(reg,ptrreg)
1488 SetOperandsForMemInstr(ChooseStoreInstruction(
1489 subtreeRoot->leftChild()->getValue()->getType()),
1490 mvec, subtreeRoot, target);
1493 case 5: // stmt: BrUncond
1495 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1496 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
1499 mvec.push_back(BuildMI(V9::NOP, 0));
1503 case 206: // stmt: BrCond(setCCconst)
1504 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
1505 // If the constant is ZERO, we can use the branch-on-integer-register
1506 // instructions and avoid the SUBcc instruction entirely.
1507 // Otherwise this is just the same as case 5, so just fall through.
1509 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
1511 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
1512 Constant *constVal = cast<Constant>(constNode->getValue());
1515 if ((constVal->getType()->isInteger()
1516 || isa<PointerType>(constVal->getType()))
1517 && GetConstantValueAsSignedInt(constVal, isValidConst) == 0
1520 // That constant is a zero after all...
1521 // Use the left child of setCC as the first argument!
1522 // Mark the setCC node so that no code is generated for it.
1523 InstructionNode* setCCNode = (InstructionNode*)
1524 subtreeRoot->leftChild();
1525 assert(setCCNode->getOpLabel() == SetCCOp);
1526 setCCNode->markFoldedIntoParent();
1528 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
1530 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
1531 .addReg(setCCNode->leftChild()->getValue())
1532 .addPCDisp(brInst->getSuccessor(0));
1536 mvec.push_back(BuildMI(V9::NOP, 0));
1539 mvec.push_back(BuildMI(V9::BA, 1)
1540 .addPCDisp(brInst->getSuccessor(1)));
1543 mvec.push_back(BuildMI(V9::NOP, 0));
1546 // ELSE FALL THROUGH
1549 case 6: // stmt: BrCond(setCC)
1550 { // bool => boolean was computed with SetCC.
1551 // The branch to use depends on whether it is FP, signed, or unsigned.
1552 // If it is an integer CC, we also need to find the unique
1553 // TmpInstruction representing that CC.
1555 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
1557 unsigned Opcode = ChooseBccInstruction(subtreeRoot, isFPBranch);
1558 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
1559 brInst->getParent()->getParent(),
1560 isFPBranch? Type::FloatTy : Type::IntTy);
1561 M = BuildMI(Opcode, 2).addCCReg(ccValue)
1562 .addPCDisp(brInst->getSuccessor(0));
1566 mvec.push_back(BuildMI(V9::NOP, 0));
1569 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
1572 mvec.push_back(BuildMI(V9::NOP, 0));
1576 case 208: // stmt: BrCond(boolconst)
1578 // boolconst => boolean is a constant; use BA to first or second label
1579 Constant* constVal =
1580 cast<Constant>(subtreeRoot->leftChild()->getValue());
1581 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
1583 M = BuildMI(V9::BA, 1).addPCDisp(
1584 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
1588 mvec.push_back(BuildMI(V9::NOP, 0));
1592 case 8: // stmt: BrCond(boolreg)
1593 { // boolreg => boolean is stored in an existing register.
1594 // Just use the branch-on-integer-register instruction!
1596 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1597 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
1598 .addPCDisp(BI->getSuccessor(0));
1602 mvec.push_back(BuildMI(V9::NOP, 0));
1605 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
1608 mvec.push_back(BuildMI(V9::NOP, 0));
1612 case 9: // stmt: Switch(reg)
1613 assert(0 && "*** SWITCH instruction is not implemented yet.");
1616 case 10: // reg: VRegList(reg, reg)
1617 assert(0 && "VRegList should never be the topmost non-chain rule");
1620 case 21: // bool: Not(bool,reg): Both these are implemented as:
1621 case 421: // reg: BNot(reg,reg): reg = reg XOR-NOT 0
1622 { // First find the unary operand. It may be left or right, usually right.
1623 Value* notArg = BinaryOperator::getNotArgument(
1624 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1625 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1626 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(notArg).addMReg(ZeroReg)
1627 .addRegDef(subtreeRoot->getValue()));
1631 case 22: // reg: ToBoolTy(reg):
1633 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
1634 assert(opType->isIntegral() || isa<PointerType>(opType));
1635 forwardOperandNum = 0; // forward first operand to user
1639 case 23: // reg: ToUByteTy(reg)
1640 case 24: // reg: ToSByteTy(reg)
1641 case 25: // reg: ToUShortTy(reg)
1642 case 26: // reg: ToShortTy(reg)
1643 case 27: // reg: ToUIntTy(reg)
1644 case 28: // reg: ToIntTy(reg)
1646 //======================================================================
1647 // Rules for integer conversions:
1650 // From ISO 1998 C++ Standard, Sec. 4.7:
1652 // 2. If the destination type is unsigned, the resulting value is
1653 // the least unsigned integer congruent to the source integer
1654 // (modulo 2n where n is the number of bits used to represent the
1655 // unsigned type). [Note: In a two s complement representation,
1656 // this conversion is conceptual and there is no change in the
1657 // bit pattern (if there is no truncation). ]
1659 // 3. If the destination type is signed, the value is unchanged if
1660 // it can be represented in the destination type (and bitfield width);
1661 // otherwise, the value is implementation-defined.
1664 // Since we assume 2s complement representations, this implies:
1666 // -- if operand is smaller than destination, zero-extend or sign-extend
1667 // according to the signedness of the *operand*: source decides.
1668 // ==> we have to do nothing here!
1670 // -- if operand is same size as or larger than destination, and the
1671 // destination is *unsigned*, zero-extend the operand: dest. decides
1673 // -- if operand is same size as or larger than destination, and the
1674 // destination is *signed*, the choice is implementation defined:
1675 // we sign-extend the operand: i.e., again dest. decides.
1676 // Note: this matches both Sun's cc and gcc3.2.
1677 //======================================================================
1679 Instruction* destI = subtreeRoot->getInstruction();
1680 Value* opVal = subtreeRoot->leftChild()->getValue();
1681 const Type* opType = opVal->getType();
1682 if (opType->isIntegral() || isa<PointerType>(opType)) {
1683 unsigned opSize = target.getTargetData().getTypeSize(opType);
1685 target.getTargetData().getTypeSize(destI->getType());
1686 if (opSize >= destSize) {
1687 // Operand is same size as or larger than dest:
1688 // zero- or sign-extend, according to the signeddness of
1689 // the destination (see above).
1690 if (destI->getType()->isSigned())
1691 target.getInstrInfo().CreateSignExtensionInstructions(target,
1692 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1693 mvec, MachineCodeForInstruction::get(destI));
1695 target.getInstrInfo().CreateZeroExtensionInstructions(target,
1696 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1697 mvec, MachineCodeForInstruction::get(destI));
1699 forwardOperandNum = 0; // forward first operand to user
1700 } else if (opType->isFloatingPoint()) {
1701 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1702 MachineCodeForInstruction::get(destI));
1703 if (destI->getType()->isUnsigned())
1704 maskUnsignedResult = true; // not handled by fp->int code
1706 assert(0 && "Unrecognized operand type for convert-to-unsigned");
1711 case 29: // reg: ToULongTy(reg)
1712 case 30: // reg: ToLongTy(reg)
1714 Value* opVal = subtreeRoot->leftChild()->getValue();
1715 const Type* opType = opVal->getType();
1716 if (opType->isIntegral() || isa<PointerType>(opType))
1717 forwardOperandNum = 0; // forward first operand to user
1718 else if (opType->isFloatingPoint()) {
1719 Instruction* destI = subtreeRoot->getInstruction();
1720 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1721 MachineCodeForInstruction::get(destI));
1723 assert(0 && "Unrecognized operand type for convert-to-signed");
1727 case 31: // reg: ToFloatTy(reg):
1728 case 32: // reg: ToDoubleTy(reg):
1729 case 232: // reg: ToDoubleTy(Constant):
1731 // If this instruction has a parent (a user) in the tree
1732 // and the user is translated as an FsMULd instruction,
1733 // then the cast is unnecessary. So check that first.
1734 // In the future, we'll want to do the same for the FdMULq instruction,
1735 // so do the check here instead of only for ToFloatTy(reg).
1737 if (subtreeRoot->parent() != NULL) {
1738 const MachineCodeForInstruction& mcfi =
1739 MachineCodeForInstruction::get(
1740 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
1741 if (mcfi.size() == 0 || mcfi.front()->getOpCode() == V9::FSMULD)
1742 forwardOperandNum = 0; // forward first operand to user
1745 if (forwardOperandNum != 0) { // we do need the cast
1746 Value* leftVal = subtreeRoot->leftChild()->getValue();
1747 const Type* opType = leftVal->getType();
1748 MachineOpCode opCode=ChooseConvertToFloatInstr(
1749 subtreeRoot->getOpLabel(), opType);
1750 if (opCode == V9::INVALID_OPCODE) { // no conversion needed
1751 forwardOperandNum = 0; // forward first operand to user
1753 // If the source operand is a non-FP type it must be
1754 // first copied from int to float register via memory!
1755 Instruction *dest = subtreeRoot->getInstruction();
1758 if (! opType->isFloatingPoint()) {
1759 // Create a temporary to represent the FP register
1760 // into which the integer will be copied via memory.
1761 // The type of this temporary will determine the FP
1762 // register used: single-prec for a 32-bit int or smaller,
1763 // double-prec for a 64-bit int.
1766 target.getTargetData().getTypeSize(leftVal->getType());
1767 Type* tmpTypeToUse =
1768 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
1769 srcForCast = new TmpInstruction(tmpTypeToUse, dest);
1770 MachineCodeForInstruction &destMCFI =
1771 MachineCodeForInstruction::get(dest);
1772 destMCFI.addTemp(srcForCast);
1774 target.getInstrInfo().CreateCodeToCopyIntToFloat(target,
1775 dest->getParent()->getParent(),
1776 leftVal, cast<Instruction>(srcForCast),
1779 srcForCast = leftVal;
1781 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
1787 case 19: // reg: ToArrayTy(reg):
1788 case 20: // reg: ToPointerTy(reg):
1789 forwardOperandNum = 0; // forward first operand to user
1792 case 233: // reg: Add(reg, Constant)
1793 maskUnsignedResult = true;
1794 M = CreateAddConstInstruction(subtreeRoot);
1799 // ELSE FALL THROUGH
1801 case 33: // reg: Add(reg, reg)
1802 maskUnsignedResult = true;
1803 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
1806 case 234: // reg: Sub(reg, Constant)
1807 maskUnsignedResult = true;
1808 M = CreateSubConstInstruction(subtreeRoot);
1813 // ELSE FALL THROUGH
1815 case 34: // reg: Sub(reg, reg)
1816 maskUnsignedResult = true;
1817 Add3OperandInstr(ChooseSubInstructionByType(
1818 subtreeRoot->getInstruction()->getType()),
1822 case 135: // reg: Mul(todouble, todouble)
1826 case 35: // reg: Mul(reg, reg)
1828 maskUnsignedResult = true;
1829 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1831 : INVALID_MACHINE_OPCODE);
1832 Instruction* mulInstr = subtreeRoot->getInstruction();
1833 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1834 subtreeRoot->leftChild()->getValue(),
1835 subtreeRoot->rightChild()->getValue(),
1837 MachineCodeForInstruction::get(mulInstr),forceOp);
1840 case 335: // reg: Mul(todouble, todoubleConst)
1844 case 235: // reg: Mul(reg, Constant)
1846 maskUnsignedResult = true;
1847 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1849 : INVALID_MACHINE_OPCODE);
1850 Instruction* mulInstr = subtreeRoot->getInstruction();
1851 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1852 subtreeRoot->leftChild()->getValue(),
1853 subtreeRoot->rightChild()->getValue(),
1855 MachineCodeForInstruction::get(mulInstr),
1859 case 236: // reg: Div(reg, Constant)
1860 maskUnsignedResult = true;
1862 CreateDivConstInstruction(target, subtreeRoot, mvec);
1863 if (mvec.size() > L)
1865 // ELSE FALL THROUGH
1867 case 36: // reg: Div(reg, reg)
1868 maskUnsignedResult = true;
1869 Add3OperandInstr(ChooseDivInstruction(target, subtreeRoot),
1873 case 37: // reg: Rem(reg, reg)
1874 case 237: // reg: Rem(reg, Constant)
1876 maskUnsignedResult = true;
1877 Instruction* remInstr = subtreeRoot->getInstruction();
1879 TmpInstruction* quot = new TmpInstruction(
1880 subtreeRoot->leftChild()->getValue(),
1881 subtreeRoot->rightChild()->getValue());
1882 TmpInstruction* prod = new TmpInstruction(
1884 subtreeRoot->rightChild()->getValue());
1885 MachineCodeForInstruction::get(remInstr).addTemp(quot).addTemp(prod);
1887 M = BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
1888 .addReg(subtreeRoot->leftChild()->getValue())
1889 .addReg(subtreeRoot->rightChild()->getValue())
1893 unsigned MulOpcode =
1894 ChooseMulInstructionByType(subtreeRoot->getInstruction()->getType());
1895 Value *MulRHS = subtreeRoot->rightChild()->getValue();
1896 M = BuildMI(MulOpcode, 3).addReg(quot).addReg(MulRHS).addReg(prod,
1900 unsigned Opcode = ChooseSubInstructionByType(
1901 subtreeRoot->getInstruction()->getType());
1902 M = BuildMI(Opcode, 3).addReg(subtreeRoot->leftChild()->getValue())
1903 .addReg(prod).addRegDef(subtreeRoot->getValue());
1908 case 38: // bool: And(bool, bool)
1909 case 238: // bool: And(bool, boolconst)
1910 case 338: // reg : BAnd(reg, reg)
1911 case 538: // reg : BAnd(reg, Constant)
1912 Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
1915 case 138: // bool: And(bool, not)
1916 case 438: // bool: BAnd(bool, bnot)
1917 { // Use the argument of NOT as the second argument!
1918 // Mark the NOT node so that no code is generated for it.
1919 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1920 Value* notArg = BinaryOperator::getNotArgument(
1921 cast<BinaryOperator>(notNode->getInstruction()));
1922 notNode->markFoldedIntoParent();
1923 Value *LHS = subtreeRoot->leftChild()->getValue();
1924 Value *Dest = subtreeRoot->getValue();
1925 mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(LHS).addReg(notArg)
1926 .addReg(Dest, MOTy::Def));
1930 case 39: // bool: Or(bool, bool)
1931 case 239: // bool: Or(bool, boolconst)
1932 case 339: // reg : BOr(reg, reg)
1933 case 539: // reg : BOr(reg, Constant)
1934 Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
1937 case 139: // bool: Or(bool, not)
1938 case 439: // bool: BOr(bool, bnot)
1939 { // Use the argument of NOT as the second argument!
1940 // Mark the NOT node so that no code is generated for it.
1941 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1942 Value* notArg = BinaryOperator::getNotArgument(
1943 cast<BinaryOperator>(notNode->getInstruction()));
1944 notNode->markFoldedIntoParent();
1945 Value *LHS = subtreeRoot->leftChild()->getValue();
1946 Value *Dest = subtreeRoot->getValue();
1947 mvec.push_back(BuildMI(V9::ORNr, 3).addReg(LHS).addReg(notArg)
1948 .addReg(Dest, MOTy::Def));
1952 case 40: // bool: Xor(bool, bool)
1953 case 240: // bool: Xor(bool, boolconst)
1954 case 340: // reg : BXor(reg, reg)
1955 case 540: // reg : BXor(reg, Constant)
1956 Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
1959 case 140: // bool: Xor(bool, not)
1960 case 440: // bool: BXor(bool, bnot)
1961 { // Use the argument of NOT as the second argument!
1962 // Mark the NOT node so that no code is generated for it.
1963 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1964 Value* notArg = BinaryOperator::getNotArgument(
1965 cast<BinaryOperator>(notNode->getInstruction()));
1966 notNode->markFoldedIntoParent();
1967 Value *LHS = subtreeRoot->leftChild()->getValue();
1968 Value *Dest = subtreeRoot->getValue();
1969 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(LHS).addReg(notArg)
1970 .addReg(Dest, MOTy::Def));
1974 case 41: // boolconst: SetCC(reg, Constant)
1976 // If the SetCC was folded into the user (parent), it will be
1977 // caught above. All other cases are the same as case 42,
1978 // so just fall through.
1980 case 42: // bool: SetCC(reg, reg):
1982 // This generates a SUBCC instruction, putting the difference in
1983 // a result register, and setting a condition code.
1985 // If the boolean result of the SetCC is used by anything other
1986 // than a branch instruction, or if it is used outside the current
1987 // basic block, the boolean must be
1988 // computed and stored in the result register. Otherwise, discard
1989 // the difference (by using %g0) and keep only the condition code.
1991 // To compute the boolean result in a register we use a conditional
1992 // move, unless the result of the SUBCC instruction can be used as
1993 // the bool! This assumes that zero is FALSE and any non-zero
1996 InstructionNode* parentNode = (InstructionNode*) subtreeRoot->parent();
1997 Instruction* setCCInstr = subtreeRoot->getInstruction();
1999 bool keepBoolVal = parentNode == NULL ||
2000 ! AllUsesAreBranches(setCCInstr);
2001 bool subValIsBoolVal = setCCInstr->getOpcode() == Instruction::SetNE;
2002 bool keepSubVal = keepBoolVal && subValIsBoolVal;
2003 bool computeBoolVal = keepBoolVal && ! subValIsBoolVal;
2007 MachineOpCode movOpCode = 0;
2009 // Mark the 4th operand as being a CC register, and as a def
2010 // A TmpInstruction is created to represent the CC "result".
2011 // Unlike other instances of TmpInstruction, this one is used
2012 // by machine code of multiple LLVM instructions, viz.,
2013 // the SetCC and the branch. Make sure to get the same one!
2014 // Note that we do this even for FP CC registers even though they
2015 // are explicit operands, because the type of the operand
2016 // needs to be a floating point condition code, not an integer
2017 // condition code. Think of this as casting the bool result to
2018 // a FP condition code register.
2020 Value* leftVal = subtreeRoot->leftChild()->getValue();
2021 bool isFPCompare = leftVal->getType()->isFloatingPoint();
2023 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
2024 setCCInstr->getParent()->getParent(),
2025 isFPCompare ? Type::FloatTy : Type::IntTy);
2026 MachineCodeForInstruction::get(setCCInstr).addTemp(tmpForCC);
2028 if (! isFPCompare) {
2029 // Integer condition: dest. should be %g0 or an integer register.
2030 // If result must be saved but condition is not SetEQ then we need
2031 // a separate instruction to compute the bool result, so discard
2032 // result of SUBcc instruction anyway.
2035 M = BuildMI(V9::SUBccr, 4)
2036 .addReg(subtreeRoot->leftChild()->getValue())
2037 .addReg(subtreeRoot->rightChild()->getValue())
2038 .addRegDef(subtreeRoot->getValue())
2039 .addCCReg(tmpForCC, MOTy::Def);
2041 M = BuildMI(V9::SUBccr, 4)
2042 .addReg(subtreeRoot->leftChild()->getValue())
2043 .addReg(subtreeRoot->rightChild()->getValue())
2044 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def)
2045 .addCCReg(tmpForCC, MOTy::Def);
2049 if (computeBoolVal) {
2050 // recompute bool using the integer condition codes
2052 ChooseMovpccAfterSub(subtreeRoot,mustClearReg,valueToMove);
2055 // FP condition: dest of FCMP should be some FCCn register
2056 M = BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
2057 .addCCReg(tmpForCC, MOTy::Def)
2058 .addReg(subtreeRoot->leftChild()->getValue())
2059 .addRegDef(subtreeRoot->rightChild()->getValue());
2062 if (computeBoolVal) {
2063 // recompute bool using the FP condition codes
2064 mustClearReg = true;
2066 movOpCode = ChooseMovFpccInstruction(subtreeRoot);
2070 if (computeBoolVal) {
2072 // Unconditionally set register to 0
2073 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
2077 // Now conditionally move `valueToMove' (0 or 1) into the register
2078 // Mark the register as a use (as well as a def) because the old
2079 // value should be retained if the condition is false.
2080 M = BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(valueToMove)
2081 .addReg(setCCInstr, MOTy::UseAndDef);
2087 case 51: // reg: Load(reg)
2088 case 52: // reg: Load(ptrreg)
2089 SetOperandsForMemInstr(ChooseLoadInstruction(
2090 subtreeRoot->getValue()->getType()),
2091 mvec, subtreeRoot, target);
2094 case 55: // reg: GetElemPtr(reg)
2095 case 56: // reg: GetElemPtrIdx(reg,reg)
2096 // If the GetElemPtr was folded into the user (parent), it will be
2097 // caught above. For other cases, we have to compute the address.
2098 SetOperandsForMemInstr(V9::ADDr, mvec, subtreeRoot, target);
2101 case 57: // reg: Alloca: Implement as 1 instruction:
2102 { // add %fp, offsetFromFP -> result
2103 AllocationInst* instr =
2104 cast<AllocationInst>(subtreeRoot->getInstruction());
2106 target.getTargetData().getTypeSize(instr->getAllocatedType());
2108 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
2112 case 58: // reg: Alloca(reg): Implement as 3 instructions:
2113 // mul num, typeSz -> tmp
2114 // sub %sp, tmp -> %sp
2115 { // add %sp, frameSizeBelowDynamicArea -> result
2116 AllocationInst* instr =
2117 cast<AllocationInst>(subtreeRoot->getInstruction());
2118 const Type* eltType = instr->getAllocatedType();
2120 // If #elements is constant, use simpler code for fixed-size allocas
2121 int tsize = (int) target.getTargetData().getTypeSize(eltType);
2122 Value* numElementsVal = NULL;
2123 bool isArray = instr->isArrayAllocation();
2126 isa<Constant>(numElementsVal = instr->getArraySize()))
2128 // total size is constant: generate code for fixed-size alloca
2129 unsigned numElements = isArray?
2130 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
2131 CreateCodeForFixedSizeAlloca(target, instr, tsize,
2134 else // total size is not constant.
2135 CreateCodeForVariableSizeAlloca(target, instr, tsize,
2136 numElementsVal, mvec);
2140 case 61: // reg: Call
2141 { // Generate a direct (CALL) or indirect (JMPL) call.
2142 // Mark the return-address register, the indirection
2143 // register (for indirect calls), the operands of the Call,
2144 // and the return value (if any) as implicit operands
2145 // of the machine instruction.
2147 // If this is a varargs function, floating point arguments
2148 // have to passed in integer registers so insert
2149 // copy-float-to-int instructions for each float operand.
2151 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
2152 Value *callee = callInstr->getCalledValue();
2153 Function* calledFunc = dyn_cast<Function>(callee);
2155 // Check if this is an intrinsic function that needs a special code
2156 // sequence (e.g., va_start). Indirect calls cannot be special.
2158 bool specialIntrinsic = false;
2159 LLVMIntrinsic::ID iid;
2160 if (calledFunc && (iid=(LLVMIntrinsic::ID)calledFunc->getIntrinsicID()))
2161 specialIntrinsic = CodeGenIntrinsic(iid, *callInstr, target, mvec);
2163 // If not, generate the normal call sequence for the function.
2164 // This can also handle any intrinsics that are just function calls.
2166 if (! specialIntrinsic) {
2167 // Create hidden virtual register for return address with type void*
2168 TmpInstruction* retAddrReg =
2169 new TmpInstruction(PointerType::get(Type::VoidTy), callInstr);
2170 MachineCodeForInstruction::get(callInstr).addTemp(retAddrReg);
2172 // Generate the machine instruction and its operands.
2173 // Use CALL for direct function calls; this optimistically assumes
2174 // the PC-relative address fits in the CALL address field (22 bits).
2175 // Use JMPL for indirect calls.
2177 if (calledFunc) // direct function call
2178 M = BuildMI(V9::CALL, 1).addPCDisp(callee);
2179 else // indirect function call
2180 M = BuildMI(V9::JMPLCALLi, 3).addReg(callee).addSImm((int64_t)0)
2181 .addRegDef(retAddrReg);
2184 const FunctionType* funcType =
2185 cast<FunctionType>(cast<PointerType>(callee->getType())
2186 ->getElementType());
2187 bool isVarArgs = funcType->isVarArg();
2188 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
2190 // Use a descriptor to pass information about call arguments
2191 // to the register allocator. This descriptor will be "owned"
2192 // and freed automatically when the MachineCodeForInstruction
2193 // object for the callInstr goes away.
2194 CallArgsDescriptor* argDesc = new CallArgsDescriptor(callInstr,
2195 retAddrReg, isVarArgs,noPrototype);
2197 assert(callInstr->getOperand(0) == callee
2198 && "This is assumed in the loop below!");
2200 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2201 Value* argVal = callInstr->getOperand(i);
2202 Instruction* intArgReg = NULL;
2204 // Check for FP arguments to varargs functions.
2205 // Any such argument in the first $K$ args must be passed in an
2206 // integer register, where K = #integer argument registers.
2207 if (isVarArgs && argVal->getType()->isFloatingPoint()) {
2208 // If it is a function with no prototype, pass value
2209 // as an FP value as well as a varargs value
2211 argDesc->getArgInfo(i-1).setUseFPArgReg();
2213 // If this arg. is in the first $K$ regs, add a copy
2214 // float-to-int instruction to pass the value as an integer.
2215 if (i <= target.getRegInfo().getNumOfIntArgRegs()) {
2216 MachineCodeForInstruction &destMCFI =
2217 MachineCodeForInstruction::get(callInstr);
2218 intArgReg = new TmpInstruction(Type::IntTy, argVal);
2219 destMCFI.addTemp(intArgReg);
2221 std::vector<MachineInstr*> copyMvec;
2222 target.getInstrInfo().CreateCodeToCopyFloatToInt(target,
2223 callInstr->getParent()->getParent(),
2224 argVal, (TmpInstruction*) intArgReg,
2225 copyMvec, destMCFI);
2226 mvec.insert(mvec.begin(),copyMvec.begin(),copyMvec.end());
2228 argDesc->getArgInfo(i-1).setUseIntArgReg();
2229 argDesc->getArgInfo(i-1).setArgCopy(intArgReg);
2231 // Cannot fit in first $K$ regs so pass arg on stack
2232 argDesc->getArgInfo(i-1).setUseStackSlot();
2236 mvec.back()->addImplicitRef(intArgReg);
2238 mvec.back()->addImplicitRef(argVal);
2241 // Add the return value as an implicit ref. The call operands
2242 // were added above.
2243 if (callInstr->getType() != Type::VoidTy)
2244 mvec.back()->addImplicitRef(callInstr, /*isDef*/ true);
2246 // For the CALL instruction, the ret. addr. reg. is also implicit
2247 if (isa<Function>(callee))
2248 mvec.back()->addImplicitRef(retAddrReg, /*isDef*/ true);
2251 mvec.push_back(BuildMI(V9::NOP, 0));
2257 case 62: // reg: Shl(reg, reg)
2259 Value* argVal1 = subtreeRoot->leftChild()->getValue();
2260 Value* argVal2 = subtreeRoot->rightChild()->getValue();
2261 Instruction* shlInstr = subtreeRoot->getInstruction();
2263 const Type* opType = argVal1->getType();
2264 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2265 "Shl unsupported for other types");
2267 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
2268 (opType == Type::LongTy)? V9::SLLXr6:V9::SLLr6,
2269 argVal1, argVal2, 0, shlInstr, mvec,
2270 MachineCodeForInstruction::get(shlInstr));
2274 case 63: // reg: Shr(reg, reg)
2276 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
2277 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2278 "Shr unsupported for other types");
2279 Add3OperandInstr(opType->isSigned()
2280 ? (opType == Type::LongTy ? V9::SRAXr6 : V9::SRAr6)
2281 : (opType == Type::LongTy ? V9::SRLXr6 : V9::SRLr6),
2286 case 64: // reg: Phi(reg,reg)
2287 break; // don't forward the value
2289 case 65: // reg: VaArg(reg)
2291 // Use value initialized by va_start as pointer to args on the stack.
2292 // Load argument via current pointer value, then increment pointer.
2293 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
2294 Instruction* vaArgI = subtreeRoot->getInstruction();
2295 mvec.push_back(BuildMI(V9::LDXi, 3).addReg(vaArgI->getOperand(0)).
2296 addSImm(0).addRegDef(vaArgI));
2297 mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaArgI->getOperand(0)).
2298 addSImm(argSize).addRegDef(vaArgI->getOperand(0)));
2302 case 71: // reg: VReg
2303 case 72: // reg: Constant
2304 break; // don't forward the value
2307 assert(0 && "Unrecognized BURG rule");
2312 if (forwardOperandNum >= 0) {
2313 // We did not generate a machine instruction but need to use operand.
2314 // If user is in the same tree, replace Value in its machine operand.
2315 // If not, insert a copy instruction which should get coalesced away
2316 // by register allocation.
2317 if (subtreeRoot->parent() != NULL)
2318 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
2320 std::vector<MachineInstr*> minstrVec;
2321 Instruction* instr = subtreeRoot->getInstruction();
2322 target.getInstrInfo().
2323 CreateCopyInstructionsByType(target,
2324 instr->getParent()->getParent(),
2325 instr->getOperand(forwardOperandNum),
2327 MachineCodeForInstruction::get(instr));
2328 assert(minstrVec.size() > 0);
2329 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
2333 if (maskUnsignedResult) {
2334 // If result is unsigned and smaller than int reg size,
2335 // we need to clear high bits of result value.
2336 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
2337 Instruction* dest = subtreeRoot->getInstruction();
2338 if (dest->getType()->isUnsigned()) {
2339 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
2340 if (destSize <= 4) {
2341 // Mask high bits. Use a TmpInstruction to represent the
2342 // intermediate result before masking. Since those instructions
2343 // have already been generated, go back and substitute tmpI
2344 // for dest in the result position of each one of them.
2345 TmpInstruction *tmpI = new TmpInstruction(dest->getType(), dest,
2347 MachineCodeForInstruction::get(dest).addTemp(tmpI);
2349 for (unsigned i=0, N=mvec.size(); i < N; ++i)
2350 mvec[i]->substituteValue(dest, tmpI);
2352 M = BuildMI(V9::SRLi6, 3).addReg(tmpI).addZImm(8*(4-destSize))
2353 .addReg(dest, MOTy::Def);
2355 } else if (destSize < 8) {
2356 assert(0 && "Unsupported type size: 32 < size < 64 bits");