1 //===-- SparcInstrSelection.cpp -------------------------------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by the LLVM research group and is distributed under
6 // the University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // BURS instruction selection for SPARC V9 architecture.
12 //===----------------------------------------------------------------------===//
14 #include "SparcInternals.h"
15 #include "SparcInstrSelectionSupport.h"
16 #include "SparcRegClassInfo.h"
17 #include "llvm/CodeGen/InstrSelectionSupport.h"
18 #include "llvm/CodeGen/MachineInstrBuilder.h"
19 #include "llvm/CodeGen/MachineInstrAnnot.h"
20 #include "llvm/CodeGen/InstrForest.h"
21 #include "llvm/CodeGen/InstrSelection.h"
22 #include "llvm/CodeGen/MachineFunction.h"
23 #include "llvm/CodeGen/MachineFunctionInfo.h"
24 #include "llvm/CodeGen/MachineCodeForInstruction.h"
25 #include "llvm/DerivedTypes.h"
26 #include "llvm/Instructions.h"
27 #include "llvm/Module.h"
28 #include "llvm/Constants.h"
29 #include "llvm/ConstantHandling.h"
30 #include "llvm/Intrinsics.h"
31 #include "Support/MathExtras.h"
35 static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
36 std::vector<MachineInstr*>& mvec) {
37 mvec.push_back(BuildMI(Opcode, 3).addReg(Node->leftChild()->getValue())
38 .addReg(Node->rightChild()->getValue())
39 .addRegDef(Node->getValue()));
43 //---------------------------------------------------------------------------
44 // Function: FoldGetElemChain
47 // Fold a chain of GetElementPtr instructions containing only
48 // constant offsets into an equivalent (Pointer, IndexVector) pair.
49 // Returns the pointer Value, and stores the resulting IndexVector
50 // in argument chainIdxVec. This is a helper function for
51 // FoldConstantIndices that does the actual folding.
52 //---------------------------------------------------------------------------
55 // Check for a constant 0.
59 return (idx == ConstantSInt::getNullValue(idx->getType()));
63 FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
64 bool lastInstHasLeadingNonZero)
66 InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
67 GetElementPtrInst* gepInst =
68 dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
70 // ptr value is not computed in this tree or ptr value does not come from GEP
75 // Return NULL if we don't fold any instructions in.
78 // Now chase the chain of getElementInstr instructions, if any.
79 // Check for any non-constant indices and stop there.
80 // Also, stop if the first index of child is a non-zero array index
81 // and the last index of the current node is a non-array index:
82 // in that case, a non-array declared type is being accessed as an array
83 // which is not type-safe, but could be legal.
85 InstructionNode* ptrChild = gepNode;
86 while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
87 ptrChild->getOpLabel() == GetElemPtrIdx))
89 // Child is a GetElemPtr instruction
90 gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
91 User::op_iterator OI, firstIdx = gepInst->idx_begin();
92 User::op_iterator lastIdx = gepInst->idx_end();
93 bool allConstantOffsets = true;
95 // The first index of every GEP must be an array index.
96 assert((*firstIdx)->getType() == Type::LongTy &&
97 "INTERNAL ERROR: Structure index for a pointer type!");
99 // If the last instruction had a leading non-zero index, check if the
100 // current one references a sequential (i.e., indexable) type.
101 // If not, the code is not type-safe and we would create an illegal GEP
102 // by folding them, so don't fold any more instructions.
104 if (lastInstHasLeadingNonZero)
105 if (! isa<SequentialType>(gepInst->getType()->getElementType()))
106 break; // cannot fold in any preceding getElementPtr instrs.
108 // Check that all offsets are constant for this instruction
109 for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
110 allConstantOffsets = isa<ConstantInt>(*OI);
112 if (allConstantOffsets) {
113 // Get pointer value out of ptrChild.
114 ptrVal = gepInst->getPointerOperand();
116 // Insert its index vector at the start, skipping any leading [0]
117 // Remember the old size to check if anything was inserted.
118 unsigned oldSize = chainIdxVec.size();
119 int firstIsZero = IsZero(*firstIdx);
120 chainIdxVec.insert(chainIdxVec.begin(), firstIdx + firstIsZero, lastIdx);
122 // Remember if it has leading zero index: it will be discarded later.
123 if (oldSize < chainIdxVec.size())
124 lastInstHasLeadingNonZero = !firstIsZero;
126 // Mark the folded node so no code is generated for it.
127 ((InstructionNode*) ptrChild)->markFoldedIntoParent();
129 // Get the previous GEP instruction and continue trying to fold
130 ptrChild = dyn_cast<InstructionNode>(ptrChild->leftChild());
131 } else // cannot fold this getElementPtr instr. or any preceding ones
135 // If the first getElementPtr instruction had a leading [0], add it back.
136 // Note that this instruction is the *last* one that was successfully
137 // folded *and* contributed any indices, in the loop above.
139 if (ptrVal && ! lastInstHasLeadingNonZero)
140 chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
146 //---------------------------------------------------------------------------
147 // Function: GetGEPInstArgs
150 // Helper function for GetMemInstArgs that handles the final getElementPtr
151 // instruction used by (or same as) the memory operation.
152 // Extracts the indices of the current instruction and tries to fold in
153 // preceding ones if all indices of the current one are constant.
154 //---------------------------------------------------------------------------
157 GetGEPInstArgs(InstructionNode* gepNode,
158 std::vector<Value*>& idxVec,
159 bool& allConstantIndices)
161 allConstantIndices = true;
162 GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
164 // Default pointer is the one from the current instruction.
165 Value* ptrVal = gepI->getPointerOperand();
166 InstrTreeNode* ptrChild = gepNode->leftChild();
168 // Extract the index vector of the GEP instruction.
169 // If all indices are constant and first index is zero, try to fold
170 // in preceding GEPs with all constant indices.
171 for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
172 allConstantIndices && OI != OE; ++OI)
173 if (! isa<Constant>(*OI))
174 allConstantIndices = false; // note: this also terminates loop!
176 // If we have only constant indices, fold chains of constant indices
177 // in this and any preceding GetElemPtr instructions.
178 bool foldedGEPs = false;
179 bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
180 if (allConstantIndices)
181 if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx)) {
186 // Append the index vector of the current instruction.
187 // Skip the leading [0] index if preceding GEPs were folded into this.
188 idxVec.insert(idxVec.end(),
189 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
195 //---------------------------------------------------------------------------
196 // Function: GetMemInstArgs
199 // Get the pointer value and the index vector for a memory operation
200 // (GetElementPtr, Load, or Store). If all indices of the given memory
201 // operation are constant, fold in constant indices in a chain of
202 // preceding GetElementPtr instructions (if any), and return the
203 // pointer value of the first instruction in the chain.
204 // All folded instructions are marked so no code is generated for them.
207 // Returns the pointer Value to use.
208 // Returns the resulting IndexVector in idxVec.
209 // Returns true/false in allConstantIndices if all indices are/aren't const.
210 //---------------------------------------------------------------------------
213 GetMemInstArgs(InstructionNode* memInstrNode,
214 std::vector<Value*>& idxVec,
215 bool& allConstantIndices)
217 allConstantIndices = false;
218 Instruction* memInst = memInstrNode->getInstruction();
219 assert(idxVec.size() == 0 && "Need empty vector to return indices");
221 // If there is a GetElemPtr instruction to fold in to this instr,
222 // it must be in the left child for Load and GetElemPtr, and in the
223 // right child for Store instructions.
224 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
225 ? memInstrNode->rightChild()
226 : memInstrNode->leftChild());
228 // Default pointer is the one from the current instruction.
229 Value* ptrVal = ptrChild->getValue();
231 // Find the "last" GetElemPtr instruction: this one or the immediate child.
232 // There will be none if this is a load or a store from a scalar pointer.
233 InstructionNode* gepNode = NULL;
234 if (isa<GetElementPtrInst>(memInst))
235 gepNode = memInstrNode;
236 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
237 // Child of load/store is a GEP and memInst is its only use.
238 // Use its indices and mark it as folded.
239 gepNode = cast<InstructionNode>(ptrChild);
240 gepNode->markFoldedIntoParent();
243 // If there are no indices, return the current pointer.
244 // Else extract the pointer from the GEP and fold the indices.
245 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
250 //************************ Internal Functions ******************************/
253 static inline MachineOpCode
254 ChooseBprInstruction(const InstructionNode* instrNode)
256 MachineOpCode opCode;
258 Instruction* setCCInstr =
259 ((InstructionNode*) instrNode->leftChild())->getInstruction();
261 switch(setCCInstr->getOpcode())
263 case Instruction::SetEQ: opCode = V9::BRZ; break;
264 case Instruction::SetNE: opCode = V9::BRNZ; break;
265 case Instruction::SetLE: opCode = V9::BRLEZ; break;
266 case Instruction::SetGE: opCode = V9::BRGEZ; break;
267 case Instruction::SetLT: opCode = V9::BRLZ; break;
268 case Instruction::SetGT: opCode = V9::BRGZ; break;
270 assert(0 && "Unrecognized VM instruction!");
271 opCode = V9::INVALID_OPCODE;
279 static inline MachineOpCode
280 ChooseBpccInstruction(const InstructionNode* instrNode,
281 const BinaryOperator* setCCInstr)
283 MachineOpCode opCode = V9::INVALID_OPCODE;
285 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
288 switch(setCCInstr->getOpcode())
290 case Instruction::SetEQ: opCode = V9::BE; break;
291 case Instruction::SetNE: opCode = V9::BNE; break;
292 case Instruction::SetLE: opCode = V9::BLE; break;
293 case Instruction::SetGE: opCode = V9::BGE; break;
294 case Instruction::SetLT: opCode = V9::BL; break;
295 case Instruction::SetGT: opCode = V9::BG; break;
297 assert(0 && "Unrecognized VM instruction!");
301 switch(setCCInstr->getOpcode())
303 case Instruction::SetEQ: opCode = V9::BE; break;
304 case Instruction::SetNE: opCode = V9::BNE; break;
305 case Instruction::SetLE: opCode = V9::BLEU; break;
306 case Instruction::SetGE: opCode = V9::BCC; break;
307 case Instruction::SetLT: opCode = V9::BCS; break;
308 case Instruction::SetGT: opCode = V9::BGU; break;
310 assert(0 && "Unrecognized VM instruction!");
318 static inline MachineOpCode
319 ChooseBFpccInstruction(const InstructionNode* instrNode,
320 const BinaryOperator* setCCInstr)
322 MachineOpCode opCode = V9::INVALID_OPCODE;
324 switch(setCCInstr->getOpcode())
326 case Instruction::SetEQ: opCode = V9::FBE; break;
327 case Instruction::SetNE: opCode = V9::FBNE; break;
328 case Instruction::SetLE: opCode = V9::FBLE; break;
329 case Instruction::SetGE: opCode = V9::FBGE; break;
330 case Instruction::SetLT: opCode = V9::FBL; break;
331 case Instruction::SetGT: opCode = V9::FBG; break;
333 assert(0 && "Unrecognized VM instruction!");
341 // Create a unique TmpInstruction for a boolean value,
342 // representing the CC register used by a branch on that value.
343 // For now, hack this using a little static cache of TmpInstructions.
344 // Eventually the entire BURG instruction selection should be put
345 // into a separate class that can hold such information.
346 // The static cache is not too bad because the memory for these
347 // TmpInstructions will be freed along with the rest of the Function anyway.
349 static TmpInstruction*
350 GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType,
351 MachineCodeForInstruction& mcfi)
353 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
354 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
355 static const Function *lastFunction = 0;// Use to flush cache between funcs
357 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
359 if (lastFunction != F) {
361 boolToTmpCache.clear();
364 // Look for tmpI and create a new one otherwise. The new value is
365 // directly written to map using the ref returned by operator[].
366 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
368 tmpI = new TmpInstruction(mcfi, ccType, boolVal);
374 static inline MachineOpCode
375 ChooseBccInstruction(const InstructionNode* instrNode,
376 const Type*& setCCType)
378 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
379 assert(setCCNode->getOpLabel() == SetCCOp);
380 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
381 setCCType = setCCInstr->getOperand(0)->getType();
383 if (setCCType->isFloatingPoint())
384 return ChooseBFpccInstruction(instrNode, setCCInstr);
386 return ChooseBpccInstruction(instrNode, setCCInstr);
390 // WARNING: since this function has only one caller, it always returns
391 // the opcode that expects an immediate and a register. If this function
392 // is ever used in cases where an opcode that takes two registers is required,
393 // then modify this function and use convertOpcodeFromRegToImm() where required.
395 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
396 // new cases of opcodes.
397 static inline MachineOpCode
398 ChooseMovFpcciInstruction(const InstructionNode* instrNode)
400 MachineOpCode opCode = V9::INVALID_OPCODE;
402 switch(instrNode->getInstruction()->getOpcode())
404 case Instruction::SetEQ: opCode = V9::MOVFEi; break;
405 case Instruction::SetNE: opCode = V9::MOVFNEi; break;
406 case Instruction::SetLE: opCode = V9::MOVFLEi; break;
407 case Instruction::SetGE: opCode = V9::MOVFGEi; break;
408 case Instruction::SetLT: opCode = V9::MOVFLi; break;
409 case Instruction::SetGT: opCode = V9::MOVFGi; break;
411 assert(0 && "Unrecognized VM instruction!");
419 // ChooseMovpcciForSetCC -- Choose a conditional-move instruction
420 // based on the type of SetCC operation.
422 // WARNING: since this function has only one caller, it always returns
423 // the opcode that expects an immediate and a register. If this function
424 // is ever used in cases where an opcode that takes two registers is required,
425 // then modify this function and use convertOpcodeFromRegToImm() where required.
427 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
428 // new cases of opcodes.
431 ChooseMovpcciForSetCC(const InstructionNode* instrNode)
433 MachineOpCode opCode = V9::INVALID_OPCODE;
435 const Type* opType = instrNode->leftChild()->getValue()->getType();
436 assert(opType->isIntegral() || isa<PointerType>(opType));
437 bool noSign = opType->isUnsigned() || isa<PointerType>(opType);
439 switch(instrNode->getInstruction()->getOpcode())
441 case Instruction::SetEQ: opCode = V9::MOVEi; break;
442 case Instruction::SetLE: opCode = noSign? V9::MOVLEUi : V9::MOVLEi; break;
443 case Instruction::SetGE: opCode = noSign? V9::MOVCCi : V9::MOVGEi; break;
444 case Instruction::SetLT: opCode = noSign? V9::MOVCSi : V9::MOVLi; break;
445 case Instruction::SetGT: opCode = noSign? V9::MOVGUi : V9::MOVGi; break;
446 case Instruction::SetNE: opCode = V9::MOVNEi; break;
447 default: assert(0 && "Unrecognized LLVM instr!"); break;
454 // ChooseMovpregiForSetCC -- Choose a conditional-move-on-register-value
455 // instruction based on the type of SetCC operation. These instructions
456 // compare a register with 0 and perform the move is the comparison is true.
458 // WARNING: like the previous function, this function it always returns
459 // the opcode that expects an immediate and a register. See above.
462 ChooseMovpregiForSetCC(const InstructionNode* instrNode)
464 MachineOpCode opCode = V9::INVALID_OPCODE;
466 switch(instrNode->getInstruction()->getOpcode())
468 case Instruction::SetEQ: opCode = V9::MOVRZi; break;
469 case Instruction::SetLE: opCode = V9::MOVRLEZi; break;
470 case Instruction::SetGE: opCode = V9::MOVRGEZi; break;
471 case Instruction::SetLT: opCode = V9::MOVRLZi; break;
472 case Instruction::SetGT: opCode = V9::MOVRGZi; break;
473 case Instruction::SetNE: opCode = V9::MOVRNZi; break;
474 default: assert(0 && "Unrecognized VM instr!"); break;
481 static inline MachineOpCode
482 ChooseConvertToFloatInstr(const TargetMachine& target,
483 OpLabel vopCode, const Type* opType)
485 assert((vopCode == ToFloatTy || vopCode == ToDoubleTy) &&
486 "Unrecognized convert-to-float opcode!");
487 assert((opType->isIntegral() || opType->isFloatingPoint() ||
488 isa<PointerType>(opType))
489 && "Trying to convert a non-scalar type to FLOAT/DOUBLE?");
491 MachineOpCode opCode = V9::INVALID_OPCODE;
493 unsigned opSize = target.getTargetData().getTypeSize(opType);
495 if (opType == Type::FloatTy)
496 opCode = (vopCode == ToFloatTy? V9::NOP : V9::FSTOD);
497 else if (opType == Type::DoubleTy)
498 opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::NOP);
499 else if (opSize <= 4)
500 opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
502 assert(opSize == 8 && "Unrecognized type size > 4 and < 8!");
503 opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
509 static inline MachineOpCode
510 ChooseConvertFPToIntInstr(const TargetMachine& target,
511 const Type* destType, const Type* opType)
513 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
514 && "This function should only be called for FLOAT or DOUBLE");
515 assert((destType->isIntegral() || isa<PointerType>(destType))
516 && "Trying to convert FLOAT/DOUBLE to a non-scalar type?");
518 MachineOpCode opCode = V9::INVALID_OPCODE;
520 unsigned destSize = target.getTargetData().getTypeSize(destType);
522 if (destType == Type::UIntTy)
523 assert(destType != Type::UIntTy && "Expand FP-to-uint beforehand.");
524 else if (destSize <= 4)
525 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
527 assert(destSize == 8 && "Unrecognized type size > 4 and < 8!");
528 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
535 CreateConvertFPToIntInstr(const TargetMachine& target,
538 const Type* destType)
540 MachineOpCode opCode = ChooseConvertFPToIntInstr(target, destType,
542 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
543 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
546 // CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer
547 // The FP value must be converted to the dest type in an FP register,
548 // and the result is then copied from FP to int register via memory.
549 // SPARC does not have a float-to-uint conversion, only a float-to-int (fdtoi).
550 // Since fdtoi converts to signed integers, any FP value V between MAXINT+1
551 // and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly.
552 // Therefore, for converting an FP value to uint32_t, we first need to convert
553 // to uint64_t and then to uint32_t.
556 CreateCodeToConvertFloatToInt(const TargetMachine& target,
559 std::vector<MachineInstr*>& mvec,
560 MachineCodeForInstruction& mcfi)
562 Function* F = destI->getParent()->getParent();
564 // Create a temporary to represent the FP register into which the
565 // int value will placed after conversion. The type of this temporary
566 // depends on the type of FP register to use: single-prec for a 32-bit
567 // int or smaller; double-prec for a 64-bit int.
569 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
571 const Type* castDestType = destI->getType(); // type for the cast instr result
572 const Type* castDestRegType; // type for cast instruction result reg
573 TmpInstruction* destForCast; // dest for cast instruction
574 Instruction* fpToIntCopyDest = destI; // dest for fp-reg-to-int-reg copy instr
576 // For converting an FP value to uint32_t, we first need to convert to
577 // uint64_t and then to uint32_t, as explained above.
578 if (destI->getType() == Type::UIntTy) {
579 castDestType = Type::ULongTy; // use this instead of type of destI
580 castDestRegType = Type::DoubleTy; // uint64_t needs 64-bit FP register.
581 destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
582 fpToIntCopyDest = new TmpInstruction(mcfi, castDestType, destForCast);
585 castDestRegType = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
586 destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
589 // Create the fp-to-int conversion instruction (src and dest regs are FP regs)
590 mvec.push_back(CreateConvertFPToIntInstr(target, opVal, destForCast,
593 // Create the fpreg-to-intreg copy code
594 target.getInstrInfo().CreateCodeToCopyFloatToInt(target, F, destForCast,
595 fpToIntCopyDest, mvec, mcfi);
597 // Create the uint64_t to uint32_t conversion, if needed
598 if (destI->getType() == Type::UIntTy)
599 target.getInstrInfo().
600 CreateZeroExtensionInstructions(target, F, fpToIntCopyDest, destI,
601 /*numLowBits*/ 32, mvec, mcfi);
605 static inline MachineOpCode
606 ChooseAddInstruction(const InstructionNode* instrNode)
608 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
612 static inline MachineInstr*
613 CreateMovFloatInstruction(const InstructionNode* instrNode,
614 const Type* resultType)
616 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
617 .addReg(instrNode->leftChild()->getValue())
618 .addRegDef(instrNode->getValue());
621 static inline MachineInstr*
622 CreateAddConstInstruction(const InstructionNode* instrNode)
624 MachineInstr* minstr = NULL;
626 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
627 assert(isa<Constant>(constOp));
629 // Cases worth optimizing are:
630 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
631 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
633 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
634 double dval = FPC->getValue();
636 minstr = CreateMovFloatInstruction(instrNode,
637 instrNode->getInstruction()->getType());
644 static inline MachineOpCode
645 ChooseSubInstructionByType(const Type* resultType)
647 MachineOpCode opCode = V9::INVALID_OPCODE;
649 if (resultType->isInteger() || isa<PointerType>(resultType)) {
652 switch(resultType->getPrimitiveID())
654 case Type::FloatTyID: opCode = V9::FSUBS; break;
655 case Type::DoubleTyID: opCode = V9::FSUBD; break;
656 default: assert(0 && "Invalid type for SUB instruction"); break;
664 static inline MachineInstr*
665 CreateSubConstInstruction(const InstructionNode* instrNode)
667 MachineInstr* minstr = NULL;
669 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
670 assert(isa<Constant>(constOp));
672 // Cases worth optimizing are:
673 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
674 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
676 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
677 double dval = FPC->getValue();
679 minstr = CreateMovFloatInstruction(instrNode,
680 instrNode->getInstruction()->getType());
687 static inline MachineOpCode
688 ChooseFcmpInstruction(const InstructionNode* instrNode)
690 MachineOpCode opCode = V9::INVALID_OPCODE;
692 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
693 switch(operand->getType()->getPrimitiveID()) {
694 case Type::FloatTyID: opCode = V9::FCMPS; break;
695 case Type::DoubleTyID: opCode = V9::FCMPD; break;
696 default: assert(0 && "Invalid type for FCMP instruction"); break;
703 // Assumes that leftArg and rightArg are both cast instructions.
706 BothFloatToDouble(const InstructionNode* instrNode)
708 InstrTreeNode* leftArg = instrNode->leftChild();
709 InstrTreeNode* rightArg = instrNode->rightChild();
710 InstrTreeNode* leftArgArg = leftArg->leftChild();
711 InstrTreeNode* rightArgArg = rightArg->leftChild();
712 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
714 // Check if both arguments are floats cast to double
715 return (leftArg->getValue()->getType() == Type::DoubleTy &&
716 leftArgArg->getValue()->getType() == Type::FloatTy &&
717 rightArgArg->getValue()->getType() == Type::FloatTy);
721 static inline MachineOpCode
722 ChooseMulInstructionByType(const Type* resultType)
724 MachineOpCode opCode = V9::INVALID_OPCODE;
726 if (resultType->isInteger())
729 switch(resultType->getPrimitiveID())
731 case Type::FloatTyID: opCode = V9::FMULS; break;
732 case Type::DoubleTyID: opCode = V9::FMULD; break;
733 default: assert(0 && "Invalid type for MUL instruction"); break;
741 static inline MachineInstr*
742 CreateIntNegInstruction(const TargetMachine& target,
745 return BuildMI(V9::SUBr, 3).addMReg(target.getRegInfo().getZeroRegNum())
746 .addReg(vreg).addRegDef(vreg);
750 // Create instruction sequence for any shift operation.
751 // SLL or SLLX on an operand smaller than the integer reg. size (64bits)
752 // requires a second instruction for explicit sign-extension.
753 // Note that we only have to worry about a sign-bit appearing in the
754 // most significant bit of the operand after shifting (e.g., bit 32 of
755 // Int or bit 16 of Short), so we do not have to worry about results
756 // that are as large as a normal integer register.
759 CreateShiftInstructions(const TargetMachine& target,
761 MachineOpCode shiftOpCode,
763 Value* optArgVal2, /* Use optArgVal2 if not NULL */
764 unsigned optShiftNum, /* else use optShiftNum */
765 Instruction* destVal,
766 std::vector<MachineInstr*>& mvec,
767 MachineCodeForInstruction& mcfi)
769 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
770 "Large shift sizes unexpected, but can be handled below: "
771 "You need to check whether or not it fits in immed field below");
773 // If this is a logical left shift of a type smaller than the standard
774 // integer reg. size, we have to extend the sign-bit into upper bits
775 // of dest, so we need to put the result of the SLL into a temporary.
777 Value* shiftDest = destVal;
778 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
780 if ((shiftOpCode == V9::SLLr5 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
781 // put SLL result into a temporary
782 shiftDest = new TmpInstruction(mcfi, argVal1, optArgVal2, "sllTmp");
785 MachineInstr* M = (optArgVal2 != NULL)
786 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
787 .addReg(shiftDest, MOTy::Def)
788 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
789 .addReg(shiftDest, MOTy::Def);
792 if (shiftDest != destVal) {
793 // extend the sign-bit of the result into all upper bits of dest
794 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
795 target.getInstrInfo().
796 CreateSignExtensionInstructions(target, F, shiftDest, destVal,
797 8*opSize, mvec, mcfi);
802 // Does not create any instructions if we cannot exploit constant to
803 // create a cheaper instruction.
804 // This returns the approximate cost of the instructions generated,
805 // which is used to pick the cheapest when both operands are constant.
807 CreateMulConstInstruction(const TargetMachine &target, Function* F,
808 Value* lval, Value* rval, Instruction* destVal,
809 std::vector<MachineInstr*>& mvec,
810 MachineCodeForInstruction& mcfi)
812 /* Use max. multiply cost, viz., cost of MULX */
813 unsigned cost = target.getInstrInfo().minLatency(V9::MULXr);
814 unsigned firstNewInstr = mvec.size();
816 Value* constOp = rval;
817 if (! isa<Constant>(constOp))
820 // Cases worth optimizing are:
821 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
822 // (2) Multiply by 2^x for integer types: replace with Shift
824 const Type* resultType = destVal->getType();
826 if (resultType->isInteger() || isa<PointerType>(resultType)) {
828 int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
829 constOp, constOp->getType(), isValidConst);
832 bool needNeg = false;
838 if (C == 0 || C == 1) {
839 cost = target.getInstrInfo().minLatency(V9::ADDr);
840 unsigned Zero = target.getRegInfo().getZeroRegNum();
843 M =BuildMI(V9::ADDr,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
845 M = BuildMI(V9::ADDr,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
847 } else if (isPowerOf2(C, pow)) {
848 unsigned opSize = target.getTargetData().getTypeSize(resultType);
849 MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
850 CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
851 destVal, mvec, mcfi);
854 if (mvec.size() > 0 && needNeg) {
855 // insert <reg = SUB 0, reg> after the instr to flip the sign
856 MachineInstr* M = CreateIntNegInstruction(target, destVal);
861 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
862 double dval = FPC->getValue();
863 if (fabs(dval) == 1) {
864 MachineOpCode opCode = (dval < 0)
865 ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
866 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
867 mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
872 if (firstNewInstr < mvec.size()) {
874 for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
875 cost += target.getInstrInfo().minLatency(mvec[i]->getOpCode());
882 // Does not create any instructions if we cannot exploit constant to
883 // create a cheaper instruction.
886 CreateCheapestMulConstInstruction(const TargetMachine &target,
888 Value* lval, Value* rval,
889 Instruction* destVal,
890 std::vector<MachineInstr*>& mvec,
891 MachineCodeForInstruction& mcfi)
894 if (isa<Constant>(lval) && isa<Constant>(rval)) {
895 // both operands are constant: evaluate and "set" in dest
896 Constant* P = ConstantFoldBinaryInstruction(Instruction::Mul,
897 cast<Constant>(lval),
898 cast<Constant>(rval));
899 target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi);
901 else if (isa<Constant>(rval)) // rval is constant, but not lval
902 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
903 else if (isa<Constant>(lval)) // lval is constant, but not rval
904 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
906 // else neither is constant
910 // Return NULL if we cannot exploit constant to create a cheaper instruction
912 CreateMulInstruction(const TargetMachine &target, Function* F,
913 Value* lval, Value* rval, Instruction* destVal,
914 std::vector<MachineInstr*>& mvec,
915 MachineCodeForInstruction& mcfi,
916 MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE)
918 unsigned L = mvec.size();
919 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
920 if (mvec.size() == L) {
921 // no instructions were added so create MUL reg, reg, reg.
922 // Use FSMULD if both operands are actually floats cast to doubles.
923 // Otherwise, use the default opcode for the appropriate type.
924 MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE)
926 : ChooseMulInstructionByType(destVal->getType()));
927 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
928 .addRegDef(destVal));
933 // Generate a divide instruction for Div or Rem.
934 // For Rem, this assumes that the operand type will be signed if the result
935 // type is signed. This is correct because they must have the same sign.
937 static inline MachineOpCode
938 ChooseDivInstruction(TargetMachine &target,
939 const InstructionNode* instrNode)
941 MachineOpCode opCode = V9::INVALID_OPCODE;
943 const Type* resultType = instrNode->getInstruction()->getType();
945 if (resultType->isInteger())
946 opCode = resultType->isSigned()? V9::SDIVXr : V9::UDIVXr;
948 switch(resultType->getPrimitiveID())
950 case Type::FloatTyID: opCode = V9::FDIVS; break;
951 case Type::DoubleTyID: opCode = V9::FDIVD; break;
952 default: assert(0 && "Invalid type for DIV instruction"); break;
959 // Return if we cannot exploit constant to create a cheaper instruction
961 CreateDivConstInstruction(TargetMachine &target,
962 const InstructionNode* instrNode,
963 std::vector<MachineInstr*>& mvec)
965 Value* LHS = instrNode->leftChild()->getValue();
966 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
967 if (!isa<Constant>(constOp))
970 Instruction* destVal = instrNode->getInstruction();
971 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
973 // Cases worth optimizing are:
974 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
975 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
977 const Type* resultType = instrNode->getInstruction()->getType();
979 if (resultType->isInteger()) {
982 int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
983 constOp, constOp->getType(), isValidConst);
985 bool needNeg = false;
992 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
993 .addRegDef(destVal));
994 } else if (isPowerOf2(C, pow)) {
997 unsigned opSize = target.getTargetData().getTypeSize(resultType);
999 if (resultType->isSigned()) {
1000 // For N / 2^k, if the operand N is negative,
1001 // we need to add (2^k - 1) before right-shifting by k, i.e.,
1003 // (N / 2^k) = N >> k, if N >= 0;
1004 // (N + 2^k - 1) >> k, if N < 0
1006 // If N is <= 32 bits, use:
1007 // sra N, 31, t1 // t1 = ~0, if N < 0, 0 else
1008 // srl t1, 32-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
1009 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
1010 // sra t3, k, result // result = N / 2^k
1012 // If N is 64 bits, use:
1013 // srax N, k-1, t1 // t1 = sign bit in high k positions
1014 // srlx t1, 64-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
1015 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
1016 // sra t3, k, result // result = N / 2^k
1018 TmpInstruction *sraTmp, *srlTmp, *addTmp;
1019 MachineCodeForInstruction& mcfi
1020 = MachineCodeForInstruction::get(destVal);
1021 sraTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
1022 srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getPlus2km1");
1023 addTmp = new TmpInstruction(mcfi, resultType, LHS, srlTmp,"incIfNeg");
1025 // Create the SRA or SRAX instruction to get the sign bit
1026 mvec.push_back(BuildMI((opSize > 4)? V9::SRAXi6 : V9::SRAi5, 3)
1028 .addSImm((resultType==Type::LongTy)? pow-1 : 31)
1029 .addRegDef(sraTmp));
1031 // Create the SRL or SRLX instruction to get the sign bit
1032 mvec.push_back(BuildMI((opSize > 4)? V9::SRLXi6 : V9::SRLi5, 3)
1034 .addSImm((resultType==Type::LongTy)? 64-pow : 32-pow)
1035 .addRegDef(srlTmp));
1037 // Create the ADD instruction to add 2^pow-1 for negative values
1038 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
1039 .addRegDef(addTmp));
1041 // Get the shift operand and "right-shift" opcode to do the divide
1042 shiftOperand = addTmp;
1043 opCode = (opSize > 4)? V9::SRAXi6 : V9::SRAi5;
1045 // Get the shift operand and "right-shift" opcode to do the divide
1047 opCode = (opSize > 4)? V9::SRLXi6 : V9::SRLi5;
1050 // Now do the actual shift!
1051 mvec.push_back(BuildMI(opCode, 3).addReg(shiftOperand).addZImm(pow)
1052 .addRegDef(destVal));
1055 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
1056 // insert <reg = SUB 0, reg> after the instr to flip the sign
1057 mvec.push_back(CreateIntNegInstruction(target, destVal));
1061 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
1062 double dval = FPC->getValue();
1063 if (fabs(dval) == 1) {
1065 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
1066 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
1068 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(destVal));
1076 CreateCodeForVariableSizeAlloca(const TargetMachine& target,
1077 Instruction* result,
1079 Value* numElementsVal,
1080 std::vector<MachineInstr*>& getMvec)
1082 Value* totalSizeVal;
1084 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
1085 Function *F = result->getParent()->getParent();
1087 // Enforce the alignment constraints on the stack pointer at
1088 // compile time if the total size is a known constant.
1089 if (isa<Constant>(numElementsVal)) {
1091 int64_t numElem = (int64_t) target.getInstrInfo().
1092 ConvertConstantToIntType(target, numElementsVal,
1093 numElementsVal->getType(), isValid);
1094 assert(isValid && "Unexpectedly large array dimension in alloca!");
1095 int64_t total = numElem * tsize;
1096 if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
1097 total += target.getFrameInfo().getStackFrameSizeAlignment() - extra;
1098 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
1100 // The size is not a constant. Generate code to compute it and
1101 // code to pad the size for stack alignment.
1102 // Create a Value to hold the (constant) element size
1103 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
1105 // Create temporary values to hold the result of MUL, SLL, SRL
1106 // To pad `size' to next smallest multiple of 16:
1107 // size = (size + 15) & (-16 = 0xfffffffffffffff0)
1109 TmpInstruction* tmpProd = new TmpInstruction(mcfi,numElementsVal, tsizeVal);
1110 TmpInstruction* tmpAdd15= new TmpInstruction(mcfi,numElementsVal, tmpProd);
1111 TmpInstruction* tmpAndf0= new TmpInstruction(mcfi,numElementsVal, tmpAdd15);
1113 // Instruction 1: mul numElements, typeSize -> tmpProd
1114 // This will optimize the MUL as far as possible.
1115 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec,
1116 mcfi, INVALID_MACHINE_OPCODE);
1118 // Instruction 2: andn tmpProd, 0x0f -> tmpAndn
1119 getMvec.push_back(BuildMI(V9::ADDi, 3).addReg(tmpProd).addSImm(15)
1120 .addReg(tmpAdd15, MOTy::Def));
1122 // Instruction 3: add tmpAndn, 0x10 -> tmpAdd16
1123 getMvec.push_back(BuildMI(V9::ANDi, 3).addReg(tmpAdd15).addSImm(-16)
1124 .addReg(tmpAndf0, MOTy::Def));
1126 totalSizeVal = tmpAndf0;
1129 // Get the constant offset from SP for dynamically allocated storage
1130 // and create a temporary Value to hold it.
1131 MachineFunction& mcInfo = MachineFunction::get(F);
1133 ConstantSInt* dynamicAreaOffset =
1134 ConstantSInt::get(Type::IntTy,
1135 target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp));
1136 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
1138 unsigned SPReg = target.getRegInfo().getStackPointer();
1140 // Instruction 2: sub %sp, totalSizeVal -> %sp
1141 getMvec.push_back(BuildMI(V9::SUBr, 3).addMReg(SPReg).addReg(totalSizeVal)
1142 .addMReg(SPReg,MOTy::Def));
1144 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
1145 getMvec.push_back(BuildMI(V9::ADDr,3).addMReg(SPReg).addReg(dynamicAreaOffset)
1146 .addRegDef(result));
1151 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
1152 Instruction* result,
1154 unsigned numElements,
1155 std::vector<MachineInstr*>& getMvec)
1157 assert(tsize > 0 && "Illegal (zero) type size for alloca");
1158 assert(result && result->getParent() &&
1159 "Result value is not part of a function?");
1160 Function *F = result->getParent()->getParent();
1161 MachineFunction &mcInfo = MachineFunction::get(F);
1163 // Put the variable in the dynamically sized area of the frame if either:
1164 // (a) The offset is too large to use as an immediate in load/stores
1165 // (check LDX because all load/stores have the same-size immed. field).
1166 // (b) The object is "large", so it could cause many other locals,
1167 // spills, and temporaries to have large offsets.
1168 // NOTE: We use LARGE = 8 * argSlotSize = 64 bytes.
1169 // You've gotta love having only 13 bits for constant offset values :-|.
1171 unsigned paddedSize;
1172 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
1174 tsize * numElements);
1176 if (((int)paddedSize) > 8 * target.getFrameInfo().getSizeOfEachArgOnStack() ||
1177 ! target.getInstrInfo().constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
1178 CreateCodeForVariableSizeAlloca(target, result, tsize,
1179 ConstantSInt::get(Type::IntTy,numElements),
1184 // else offset fits in immediate field so go ahead and allocate it.
1185 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
1187 // Create a temporary Value to hold the constant offset.
1188 // This is needed because it may not fit in the immediate field.
1189 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
1191 // Instruction 1: add %fp, offsetFromFP -> result
1192 unsigned FPReg = target.getRegInfo().getFramePointer();
1193 getMvec.push_back(BuildMI(V9::ADDr, 3).addMReg(FPReg).addReg(offsetVal)
1194 .addRegDef(result));
1198 //------------------------------------------------------------------------
1199 // Function SetOperandsForMemInstr
1201 // Choose addressing mode for the given load or store instruction.
1202 // Use [reg+reg] if it is an indexed reference, and the index offset is
1203 // not a constant or if it cannot fit in the offset field.
1204 // Use [reg+offset] in all other cases.
1206 // This assumes that all array refs are "lowered" to one of these forms:
1207 // %x = load (subarray*) ptr, constant ; single constant offset
1208 // %x = load (subarray*) ptr, offsetVal ; single non-constant offset
1209 // Generally, this should happen via strength reduction + LICM.
1210 // Also, strength reduction should take care of using the same register for
1211 // the loop index variable and an array index, when that is profitable.
1212 //------------------------------------------------------------------------
1215 SetOperandsForMemInstr(unsigned Opcode,
1216 std::vector<MachineInstr*>& mvec,
1217 InstructionNode* vmInstrNode,
1218 const TargetMachine& target)
1220 Instruction* memInst = vmInstrNode->getInstruction();
1221 // Index vector, ptr value, and flag if all indices are const.
1222 std::vector<Value*> idxVec;
1223 bool allConstantIndices;
1224 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
1226 // Now create the appropriate operands for the machine instruction.
1227 // First, initialize so we default to storing the offset in a register.
1228 int64_t smallConstOffset = 0;
1229 Value* valueForRegOffset = NULL;
1230 MachineOperand::MachineOperandType offsetOpType =
1231 MachineOperand::MO_VirtualRegister;
1233 // Check if there is an index vector and if so, compute the
1234 // right offset for structures and for arrays
1236 if (!idxVec.empty()) {
1237 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
1239 // If all indices are constant, compute the combined offset directly.
1240 if (allConstantIndices) {
1241 // Compute the offset value using the index vector. Create a
1242 // virtual reg. for it since it may not fit in the immed field.
1243 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
1244 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
1246 // There is at least one non-constant offset. Therefore, this must
1247 // be an array ref, and must have been lowered to a single non-zero
1248 // offset. (An extra leading zero offset, if any, can be ignored.)
1249 // Generate code sequence to compute address from index.
1251 bool firstIdxIsZero = IsZero(idxVec[0]);
1252 assert(idxVec.size() == 1U + firstIdxIsZero
1253 && "Array refs must be lowered before Instruction Selection");
1255 Value* idxVal = idxVec[firstIdxIsZero];
1257 std::vector<MachineInstr*> mulVec;
1259 new TmpInstruction(MachineCodeForInstruction::get(memInst),
1260 Type::ULongTy, memInst);
1262 // Get the array type indexed by idxVal, and compute its element size.
1263 // The call to getTypeSize() will fail if size is not constant.
1264 const Type* vecType = (firstIdxIsZero
1265 ? GetElementPtrInst::getIndexedType(ptrType,
1266 std::vector<Value*>(1U, idxVec[0]),
1267 /*AllowCompositeLeaf*/ true)
1269 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
1270 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
1271 target.getTargetData().getTypeSize(eltType));
1273 // CreateMulInstruction() folds constants intelligently enough.
1274 CreateMulInstruction(target, memInst->getParent()->getParent(),
1275 idxVal, /* lval, not likely to be const*/
1276 eltSizeVal, /* rval, likely to be constant */
1278 mulVec, MachineCodeForInstruction::get(memInst),
1279 INVALID_MACHINE_OPCODE);
1281 assert(mulVec.size() > 0 && "No multiply code created?");
1282 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
1284 valueForRegOffset = addr;
1287 offsetOpType = MachineOperand::MO_SignExtendedImmed;
1288 smallConstOffset = 0;
1292 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
1293 // For LOAD or GET_ELEMENT_PTR,
1294 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
1296 unsigned offsetOpNum, ptrOpNum;
1298 if (memInst->getOpcode() == Instruction::Store) {
1299 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1300 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1301 .addReg(ptrVal).addReg(valueForRegOffset);
1303 Opcode = convertOpcodeFromRegToImm(Opcode);
1304 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1305 .addReg(ptrVal).addSImm(smallConstOffset);
1308 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1309 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
1310 .addRegDef(memInst);
1312 Opcode = convertOpcodeFromRegToImm(Opcode);
1313 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
1314 .addRegDef(memInst);
1322 // Substitute operand `operandNum' of the instruction in node `treeNode'
1323 // in place of the use(s) of that instruction in node `parent'.
1324 // Check both explicit and implicit operands!
1325 // Also make sure to skip over a parent who:
1326 // (1) is a list node in the Burg tree, or
1327 // (2) itself had its results forwarded to its parent
1330 ForwardOperand(InstructionNode* treeNode,
1331 InstrTreeNode* parent,
1334 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
1336 Instruction* unusedOp = treeNode->getInstruction();
1337 Value* fwdOp = unusedOp->getOperand(operandNum);
1339 // The parent itself may be a list node, so find the real parent instruction
1340 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode)
1342 parent = parent->parent();
1343 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
1345 InstructionNode* parentInstrNode = (InstructionNode*) parent;
1347 Instruction* userInstr = parentInstrNode->getInstruction();
1348 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
1350 // The parent's mvec would be empty if it was itself forwarded.
1351 // Recursively call ForwardOperand in that case...
1353 if (mvec.size() == 0) {
1354 assert(parent->parent() != NULL &&
1355 "Parent could not have been forwarded, yet has no instructions?");
1356 ForwardOperand(treeNode, parent->parent(), operandNum);
1358 for (unsigned i=0, N=mvec.size(); i < N; i++) {
1359 MachineInstr* minstr = mvec[i];
1360 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) {
1361 const MachineOperand& mop = minstr->getOperand(i);
1362 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
1363 mop.getVRegValue() == unusedOp)
1365 minstr->SetMachineOperandVal(i, MachineOperand::MO_VirtualRegister,
1370 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
1371 if (minstr->getImplicitRef(i) == unusedOp)
1372 minstr->setImplicitRef(i, fwdOp);
1379 AllUsesAreBranches(const Instruction* setccI)
1381 for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end();
1383 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
1384 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
1389 // Generate code for any intrinsic that needs a special code sequence
1390 // instead of a regular call. If not that kind of intrinsic, do nothing.
1391 // Returns true if code was generated, otherwise false.
1393 bool CodeGenIntrinsic(LLVMIntrinsic::ID iid, CallInst &callInstr,
1394 TargetMachine &target,
1395 std::vector<MachineInstr*>& mvec)
1398 case LLVMIntrinsic::va_start: {
1399 // Get the address of the first incoming vararg argument on the stack
1401 Function* func = cast<Function>(callInstr.getParent()->getParent());
1402 int numFixedArgs = func->getFunctionType()->getNumParams();
1403 int fpReg = target.getFrameInfo().getIncomingArgBaseRegNum();
1404 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
1405 int firstVarArgOff = numFixedArgs * argSize + target.getFrameInfo().
1406 getFirstIncomingArgOffset(MachineFunction::get(func), ignore);
1407 mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
1408 addRegDef(&callInstr));
1412 case LLVMIntrinsic::va_end:
1413 return true; // no-op on Sparc
1415 case LLVMIntrinsic::va_copy:
1416 // Simple copy of current va_list (arg1) to new va_list (result)
1417 mvec.push_back(BuildMI(V9::ORr, 3).
1418 addMReg(target.getRegInfo().getZeroRegNum()).
1419 addReg(callInstr.getOperand(1)).
1420 addRegDef(&callInstr));
1423 case LLVMIntrinsic::sigsetjmp:
1424 case LLVMIntrinsic::setjmp: {
1425 // act as if we return 0
1426 unsigned g0 = target.getRegInfo().getZeroRegNum();
1427 mvec.push_back(BuildMI(V9::ORr,3).addMReg(g0).addMReg(g0)
1428 .addReg(&callInstr, MOTy::Def));
1432 case LLVMIntrinsic::siglongjmp:
1433 case LLVMIntrinsic::longjmp: {
1435 Module* M = callInstr.getParent()->getParent()->getParent();
1436 const FunctionType *voidvoidFuncTy =
1437 FunctionType::get(Type::VoidTy, std::vector<const Type*>(), false);
1438 Function *F = M->getOrInsertFunction("abort", voidvoidFuncTy);
1439 assert(F && "Unable to get or create `abort' function declaration");
1441 // Create hidden virtual register for return address with type void*
1442 TmpInstruction* retAddrReg =
1443 new TmpInstruction(MachineCodeForInstruction::get(&callInstr),
1444 PointerType::get(Type::VoidTy), &callInstr);
1446 // Use a descriptor to pass information about call arguments
1447 // to the register allocator. This descriptor will be "owned"
1448 // and freed automatically when the MachineCodeForInstruction
1449 // object for the callInstr goes away.
1450 CallArgsDescriptor* argDesc =
1451 new CallArgsDescriptor(&callInstr, retAddrReg, false, false);
1453 MachineInstr* callMI = BuildMI(V9::CALL, 1).addPCDisp(F);
1454 callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
1456 mvec.push_back(callMI);
1457 mvec.push_back(BuildMI(V9::NOP, 0));
1466 //******************* Externally Visible Functions *************************/
1468 //------------------------------------------------------------------------
1469 // External Function: ThisIsAChainRule
1472 // Check if a given BURG rule is a chain rule.
1473 //------------------------------------------------------------------------
1476 ThisIsAChainRule(int eruleno)
1480 case 111: // stmt: reg
1504 return false; break;
1509 //------------------------------------------------------------------------
1510 // External Function: GetInstructionsByRule
1513 // Choose machine instructions for the SPARC according to the
1514 // patterns chosen by the BURG-generated parser.
1515 //------------------------------------------------------------------------
1518 GetInstructionsByRule(InstructionNode* subtreeRoot,
1521 TargetMachine &target,
1522 std::vector<MachineInstr*>& mvec)
1524 bool checkCast = false; // initialize here to use fall-through
1525 bool maskUnsignedResult = false;
1527 int forwardOperandNum = -1;
1528 unsigned allocaSize = 0;
1529 MachineInstr* M, *M2;
1531 bool foldCase = false;
1535 // If the code for this instruction was folded into the parent (user),
1537 if (subtreeRoot->isFoldedIntoParent())
1541 // Let's check for chain rules outside the switch so that we don't have
1542 // to duplicate the list of chain rule production numbers here again
1544 if (ThisIsAChainRule(ruleForNode))
1546 // Chain rules have a single nonterminal on the RHS.
1547 // Get the rule that matches the RHS non-terminal and use that instead.
1549 assert(nts[0] && ! nts[1]
1550 && "A chain rule should have only one RHS non-terminal!");
1551 nextRule = burm_rule(subtreeRoot->state, nts[0]);
1552 nts = burm_nts[nextRule];
1553 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
1557 switch(ruleForNode) {
1558 case 1: // stmt: Ret
1559 case 2: // stmt: RetValue(reg)
1560 { // NOTE: Prepass of register allocation is responsible
1561 // for moving return value to appropriate register.
1562 // Copy the return value to the required return register.
1563 // Mark the return Value as an implicit ref of the RET instr..
1564 // Mark the return-address register as a hidden virtual reg.
1565 // Finally put a NOP in the delay slot.
1566 ReturnInst *returnInstr=cast<ReturnInst>(subtreeRoot->getInstruction());
1567 Value* retVal = returnInstr->getReturnValue();
1568 MachineCodeForInstruction& mcfi =
1569 MachineCodeForInstruction::get(returnInstr);
1571 // Create a hidden virtual reg to represent the return address register
1572 // used by the machine instruction but not represented in LLVM.
1574 Instruction* returnAddrTmp = new TmpInstruction(mcfi, returnInstr);
1576 MachineInstr* retMI =
1577 BuildMI(V9::JMPLRETi, 3).addReg(returnAddrTmp).addSImm(8)
1578 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
1580 // If there is a value to return, we need to:
1581 // (a) Sign-extend the value if it is smaller than 8 bytes (reg size)
1582 // (b) Insert a copy to copy the return value to the appropriate reg.
1583 // -- For FP values, create a FMOVS or FMOVD instruction
1584 // -- For non-FP values, create an add-with-0 instruction
1586 if (retVal != NULL) {
1587 const UltraSparcRegInfo& regInfo =
1588 (UltraSparcRegInfo&) target.getRegInfo();
1589 const Type* retType = retVal->getType();
1590 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
1591 unsigned retRegNum = (retType->isFloatingPoint()
1592 ? (unsigned) SparcFloatRegClass::f0
1593 : (unsigned) SparcIntRegClass::i0);
1594 retRegNum = regInfo.getUnifiedRegNum(regClassID, retRegNum);
1596 // () Insert sign-extension instructions for small signed values.
1598 Value* retValToUse = retVal;
1599 if (retType->isIntegral() && retType->isSigned()) {
1600 unsigned retSize = target.getTargetData().getTypeSize(retType);
1602 // create a temporary virtual reg. to hold the sign-extension
1603 retValToUse = new TmpInstruction(mcfi, retVal);
1605 // sign-extend retVal and put the result in the temporary reg.
1606 target.getInstrInfo().CreateSignExtensionInstructions
1607 (target, returnInstr->getParent()->getParent(),
1608 retVal, retValToUse, 8*retSize, mvec, mcfi);
1612 // (b) Now, insert a copy to to the appropriate register:
1613 // -- For FP values, create a FMOVS or FMOVD instruction
1614 // -- For non-FP values, create an add-with-0 instruction
1616 // First, create a virtual register to represent the register and
1617 // mark this vreg as being an implicit operand of the ret MI.
1618 TmpInstruction* retVReg =
1619 new TmpInstruction(mcfi, retValToUse, NULL, "argReg");
1621 retMI->addImplicitRef(retVReg);
1623 if (retType->isFloatingPoint())
1624 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
1625 .addReg(retValToUse).addReg(retVReg, MOTy::Def));
1627 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
1628 .addReg(retValToUse).addSImm((int64_t) 0)
1629 .addReg(retVReg, MOTy::Def));
1631 // Mark the operand with the register it should be assigned
1632 M->SetRegForOperand(M->getNumOperands()-1, retRegNum);
1633 retMI->SetRegForImplicitRef(retMI->getNumImplicitRefs()-1, retRegNum);
1638 // Now insert the RET instruction and a NOP for the delay slot
1639 mvec.push_back(retMI);
1640 mvec.push_back(BuildMI(V9::NOP, 0));
1645 case 3: // stmt: Store(reg,reg)
1646 case 4: // stmt: Store(reg,ptrreg)
1647 SetOperandsForMemInstr(ChooseStoreInstruction(
1648 subtreeRoot->leftChild()->getValue()->getType()),
1649 mvec, subtreeRoot, target);
1652 case 5: // stmt: BrUncond
1654 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1655 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
1658 mvec.push_back(BuildMI(V9::NOP, 0));
1662 case 206: // stmt: BrCond(setCCconst)
1663 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
1664 // If the constant is ZERO, we can use the branch-on-integer-register
1665 // instructions and avoid the SUBcc instruction entirely.
1666 // Otherwise this is just the same as case 5, so just fall through.
1668 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
1670 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
1671 Constant *constVal = cast<Constant>(constNode->getValue());
1674 if ((constVal->getType()->isInteger()
1675 || isa<PointerType>(constVal->getType()))
1676 && target.getInstrInfo().ConvertConstantToIntType(target,
1677 constVal, constVal->getType(), isValidConst) == 0
1680 // That constant is a zero after all...
1681 // Use the left child of setCC as the first argument!
1682 // Mark the setCC node so that no code is generated for it.
1683 InstructionNode* setCCNode = (InstructionNode*)
1684 subtreeRoot->leftChild();
1685 assert(setCCNode->getOpLabel() == SetCCOp);
1686 setCCNode->markFoldedIntoParent();
1688 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
1690 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
1691 .addReg(setCCNode->leftChild()->getValue())
1692 .addPCDisp(brInst->getSuccessor(0));
1696 mvec.push_back(BuildMI(V9::NOP, 0));
1699 mvec.push_back(BuildMI(V9::BA, 1)
1700 .addPCDisp(brInst->getSuccessor(1)));
1703 mvec.push_back(BuildMI(V9::NOP, 0));
1706 // ELSE FALL THROUGH
1709 case 6: // stmt: BrCond(setCC)
1710 { // bool => boolean was computed with SetCC.
1711 // The branch to use depends on whether it is FP, signed, or unsigned.
1712 // If it is an integer CC, we also need to find the unique
1713 // TmpInstruction representing that CC.
1715 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
1716 const Type* setCCType;
1717 unsigned Opcode = ChooseBccInstruction(subtreeRoot, setCCType);
1718 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
1719 brInst->getParent()->getParent(),
1721 MachineCodeForInstruction::get(brInst));
1722 M = BuildMI(Opcode, 2).addCCReg(ccValue)
1723 .addPCDisp(brInst->getSuccessor(0));
1727 mvec.push_back(BuildMI(V9::NOP, 0));
1730 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
1733 mvec.push_back(BuildMI(V9::NOP, 0));
1737 case 208: // stmt: BrCond(boolconst)
1739 // boolconst => boolean is a constant; use BA to first or second label
1740 Constant* constVal =
1741 cast<Constant>(subtreeRoot->leftChild()->getValue());
1742 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
1744 M = BuildMI(V9::BA, 1).addPCDisp(
1745 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
1749 mvec.push_back(BuildMI(V9::NOP, 0));
1753 case 8: // stmt: BrCond(boolreg)
1754 { // boolreg => boolean is recorded in an integer register.
1755 // Use branch-on-integer-register instruction.
1757 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1758 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
1759 .addPCDisp(BI->getSuccessor(0));
1763 mvec.push_back(BuildMI(V9::NOP, 0));
1766 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
1769 mvec.push_back(BuildMI(V9::NOP, 0));
1773 case 9: // stmt: Switch(reg)
1774 assert(0 && "*** SWITCH instruction is not implemented yet.");
1777 case 10: // reg: VRegList(reg, reg)
1778 assert(0 && "VRegList should never be the topmost non-chain rule");
1781 case 21: // bool: Not(bool,reg): Compute with a conditional-move-on-reg
1782 { // First find the unary operand. It may be left or right, usually right.
1783 Instruction* notI = subtreeRoot->getInstruction();
1784 Value* notArg = BinaryOperator::getNotArgument(
1785 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1786 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1788 // Unconditionally set register to 0
1789 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(notI));
1791 // Now conditionally move 1 into the register.
1792 // Mark the register as a use (as well as a def) because the old
1793 // value will be retained if the condition is false.
1794 mvec.push_back(BuildMI(V9::MOVRZi, 3).addReg(notArg).addZImm(1)
1795 .addReg(notI, MOTy::UseAndDef));
1800 case 421: // reg: BNot(reg,reg): Compute as reg = reg XOR-NOT 0
1801 { // First find the unary operand. It may be left or right, usually right.
1802 Value* notArg = BinaryOperator::getNotArgument(
1803 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1804 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1805 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(notArg).addMReg(ZeroReg)
1806 .addRegDef(subtreeRoot->getValue()));
1810 case 322: // reg: Not(tobool, reg):
1811 // Fold CAST-TO-BOOL with NOT by inverting the sense of cast-to-bool
1813 // Just fall through!
1815 case 22: // reg: ToBoolTy(reg):
1817 Instruction* castI = subtreeRoot->getInstruction();
1818 Value* opVal = subtreeRoot->leftChild()->getValue();
1819 assert(opVal->getType()->isIntegral() ||
1820 isa<PointerType>(opVal->getType()));
1822 // Unconditionally set register to 0
1823 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(castI));
1825 // Now conditionally move 1 into the register.
1826 // Mark the register as a use (as well as a def) because the old
1827 // value will be retained if the condition is false.
1828 MachineOpCode opCode = foldCase? V9::MOVRZi : V9::MOVRNZi;
1829 mvec.push_back(BuildMI(opCode, 3).addReg(opVal).addZImm(1)
1830 .addReg(castI, MOTy::UseAndDef));
1835 case 23: // reg: ToUByteTy(reg)
1836 case 24: // reg: ToSByteTy(reg)
1837 case 25: // reg: ToUShortTy(reg)
1838 case 26: // reg: ToShortTy(reg)
1839 case 27: // reg: ToUIntTy(reg)
1840 case 28: // reg: ToIntTy(reg)
1841 case 29: // reg: ToULongTy(reg)
1842 case 30: // reg: ToLongTy(reg)
1844 //======================================================================
1845 // Rules for integer conversions:
1848 // From ISO 1998 C++ Standard, Sec. 4.7:
1850 // 2. If the destination type is unsigned, the resulting value is
1851 // the least unsigned integer congruent to the source integer
1852 // (modulo 2n where n is the number of bits used to represent the
1853 // unsigned type). [Note: In a two s complement representation,
1854 // this conversion is conceptual and there is no change in the
1855 // bit pattern (if there is no truncation). ]
1857 // 3. If the destination type is signed, the value is unchanged if
1858 // it can be represented in the destination type (and bitfield width);
1859 // otherwise, the value is implementation-defined.
1862 // Since we assume 2s complement representations, this implies:
1864 // -- If operand is smaller than destination, zero-extend or sign-extend
1865 // according to the signedness of the *operand*: source decides:
1866 // (1) If operand is signed, sign-extend it.
1867 // If dest is unsigned, zero-ext the result!
1868 // (2) If operand is unsigned, our current invariant is that
1869 // it's high bits are correct, so zero-extension is not needed.
1871 // -- If operand is same size as or larger than destination,
1872 // zero-extend or sign-extend according to the signedness of
1873 // the *destination*: destination decides:
1874 // (1) If destination is signed, sign-extend (truncating if needed)
1875 // This choice is implementation defined. We sign-extend the
1876 // operand, which matches both Sun's cc and gcc3.2.
1877 // (2) If destination is unsigned, zero-extend (truncating if needed)
1878 //======================================================================
1880 Instruction* destI = subtreeRoot->getInstruction();
1881 Function* currentFunc = destI->getParent()->getParent();
1882 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(destI);
1884 Value* opVal = subtreeRoot->leftChild()->getValue();
1885 const Type* opType = opVal->getType();
1886 const Type* destType = destI->getType();
1887 unsigned opSize = target.getTargetData().getTypeSize(opType);
1888 unsigned destSize = target.getTargetData().getTypeSize(destType);
1890 bool isIntegral = opType->isIntegral() || isa<PointerType>(opType);
1892 if (opType == Type::BoolTy ||
1893 opType == destType ||
1894 isIntegral && opSize == destSize && opSize == 8) {
1895 // nothing to do in all these cases
1896 forwardOperandNum = 0; // forward first operand to user
1898 } else if (opType->isFloatingPoint()) {
1900 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, mcfi);
1901 if (destI->getType()->isUnsigned() && destI->getType() !=Type::UIntTy)
1902 maskUnsignedResult = true; // not handled by fp->int code
1904 } else if (isIntegral) {
1906 bool opSigned = opType->isSigned();
1907 bool destSigned = destType->isSigned();
1908 unsigned extSourceInBits = 8 * std::min<unsigned>(opSize, destSize);
1910 assert(! (opSize == destSize && opSigned == destSigned) &&
1911 "How can different int types have same size and signedness?");
1913 bool signExtend = (opSize < destSize && opSigned ||
1914 opSize >= destSize && destSigned);
1916 bool signAndZeroExtend = (opSize < destSize && destSize < 8u &&
1917 opSigned && !destSigned);
1918 assert(!signAndZeroExtend || signExtend);
1920 bool zeroExtendOnly = opSize >= destSize && !destSigned;
1921 assert(!zeroExtendOnly || !signExtend);
1924 Value* signExtDest = (signAndZeroExtend
1925 ? new TmpInstruction(mcfi, destType, opVal)
1928 target.getInstrInfo().CreateSignExtensionInstructions
1929 (target, currentFunc,opVal,signExtDest,extSourceInBits,mvec,mcfi);
1931 if (signAndZeroExtend)
1932 target.getInstrInfo().CreateZeroExtensionInstructions
1933 (target, currentFunc, signExtDest, destI, 8*destSize, mvec, mcfi);
1935 else if (zeroExtendOnly) {
1936 target.getInstrInfo().CreateZeroExtensionInstructions
1937 (target, currentFunc, opVal, destI, extSourceInBits, mvec, mcfi);
1940 forwardOperandNum = 0; // forward first operand to user
1943 assert(0 && "Unrecognized operand type for convert-to-integer");
1948 case 31: // reg: ToFloatTy(reg):
1949 case 32: // reg: ToDoubleTy(reg):
1950 case 232: // reg: ToDoubleTy(Constant):
1952 // If this instruction has a parent (a user) in the tree
1953 // and the user is translated as an FsMULd instruction,
1954 // then the cast is unnecessary. So check that first.
1955 // In the future, we'll want to do the same for the FdMULq instruction,
1956 // so do the check here instead of only for ToFloatTy(reg).
1958 if (subtreeRoot->parent() != NULL) {
1959 const MachineCodeForInstruction& mcfi =
1960 MachineCodeForInstruction::get(
1961 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
1962 if (mcfi.size() == 0 || mcfi.front()->getOpCode() == V9::FSMULD)
1963 forwardOperandNum = 0; // forward first operand to user
1966 if (forwardOperandNum != 0) { // we do need the cast
1967 Value* leftVal = subtreeRoot->leftChild()->getValue();
1968 const Type* opType = leftVal->getType();
1969 MachineOpCode opCode=ChooseConvertToFloatInstr(target,
1970 subtreeRoot->getOpLabel(), opType);
1971 if (opCode == V9::NOP) { // no conversion needed
1972 forwardOperandNum = 0; // forward first operand to user
1974 // If the source operand is a non-FP type it must be
1975 // first copied from int to float register via memory!
1976 Instruction *dest = subtreeRoot->getInstruction();
1979 if (! opType->isFloatingPoint()) {
1980 // Create a temporary to represent the FP register
1981 // into which the integer will be copied via memory.
1982 // The type of this temporary will determine the FP
1983 // register used: single-prec for a 32-bit int or smaller,
1984 // double-prec for a 64-bit int.
1987 target.getTargetData().getTypeSize(leftVal->getType());
1988 Type* tmpTypeToUse =
1989 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
1990 MachineCodeForInstruction &destMCFI =
1991 MachineCodeForInstruction::get(dest);
1992 srcForCast = new TmpInstruction(destMCFI, tmpTypeToUse, dest);
1994 target.getInstrInfo().CreateCodeToCopyIntToFloat(target,
1995 dest->getParent()->getParent(),
1996 leftVal, cast<Instruction>(srcForCast),
1999 srcForCast = leftVal;
2001 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
2007 case 19: // reg: ToArrayTy(reg):
2008 case 20: // reg: ToPointerTy(reg):
2009 forwardOperandNum = 0; // forward first operand to user
2012 case 233: // reg: Add(reg, Constant)
2013 maskUnsignedResult = true;
2014 M = CreateAddConstInstruction(subtreeRoot);
2019 // ELSE FALL THROUGH
2021 case 33: // reg: Add(reg, reg)
2022 maskUnsignedResult = true;
2023 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
2026 case 234: // reg: Sub(reg, Constant)
2027 maskUnsignedResult = true;
2028 M = CreateSubConstInstruction(subtreeRoot);
2033 // ELSE FALL THROUGH
2035 case 34: // reg: Sub(reg, reg)
2036 maskUnsignedResult = true;
2037 Add3OperandInstr(ChooseSubInstructionByType(
2038 subtreeRoot->getInstruction()->getType()),
2042 case 135: // reg: Mul(todouble, todouble)
2046 case 35: // reg: Mul(reg, reg)
2048 maskUnsignedResult = true;
2049 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
2051 : INVALID_MACHINE_OPCODE);
2052 Instruction* mulInstr = subtreeRoot->getInstruction();
2053 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
2054 subtreeRoot->leftChild()->getValue(),
2055 subtreeRoot->rightChild()->getValue(),
2057 MachineCodeForInstruction::get(mulInstr),forceOp);
2060 case 335: // reg: Mul(todouble, todoubleConst)
2064 case 235: // reg: Mul(reg, Constant)
2066 maskUnsignedResult = true;
2067 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
2069 : INVALID_MACHINE_OPCODE);
2070 Instruction* mulInstr = subtreeRoot->getInstruction();
2071 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
2072 subtreeRoot->leftChild()->getValue(),
2073 subtreeRoot->rightChild()->getValue(),
2075 MachineCodeForInstruction::get(mulInstr),
2079 case 236: // reg: Div(reg, Constant)
2080 maskUnsignedResult = true;
2082 CreateDivConstInstruction(target, subtreeRoot, mvec);
2083 if (mvec.size() > L)
2085 // ELSE FALL THROUGH
2087 case 36: // reg: Div(reg, reg)
2089 maskUnsignedResult = true;
2091 // If either operand of divide is smaller than 64 bits, we have
2092 // to make sure the unused top bits are correct because they affect
2093 // the result. These bits are already correct for unsigned values.
2094 // They may be incorrect for signed values, so sign extend to fill in.
2095 Instruction* divI = subtreeRoot->getInstruction();
2096 Value* divOp1 = subtreeRoot->leftChild()->getValue();
2097 Value* divOp2 = subtreeRoot->rightChild()->getValue();
2098 Value* divOp1ToUse = divOp1;
2099 Value* divOp2ToUse = divOp2;
2100 if (divI->getType()->isSigned()) {
2101 unsigned opSize=target.getTargetData().getTypeSize(divI->getType());
2103 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(divI);
2104 divOp1ToUse = new TmpInstruction(mcfi, divOp1);
2105 divOp2ToUse = new TmpInstruction(mcfi, divOp2);
2106 target.getInstrInfo().
2107 CreateSignExtensionInstructions(target,
2108 divI->getParent()->getParent(),
2109 divOp1, divOp1ToUse,
2110 8*opSize, mvec, mcfi);
2111 target.getInstrInfo().
2112 CreateSignExtensionInstructions(target,
2113 divI->getParent()->getParent(),
2114 divOp2, divOp2ToUse,
2115 8*opSize, mvec, mcfi);
2119 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
2120 .addReg(divOp1ToUse)
2121 .addReg(divOp2ToUse)
2127 case 37: // reg: Rem(reg, reg)
2128 case 237: // reg: Rem(reg, Constant)
2130 maskUnsignedResult = true;
2132 Instruction* remI = subtreeRoot->getInstruction();
2133 Value* divOp1 = subtreeRoot->leftChild()->getValue();
2134 Value* divOp2 = subtreeRoot->rightChild()->getValue();
2136 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(remI);
2138 // If second operand of divide is smaller than 64 bits, we have
2139 // to make sure the unused top bits are correct because they affect
2140 // the result. These bits are already correct for unsigned values.
2141 // They may be incorrect for signed values, so sign extend to fill in.
2143 Value* divOpToUse = divOp2;
2144 if (divOp2->getType()->isSigned()) {
2145 unsigned opSize=target.getTargetData().getTypeSize(divOp2->getType());
2147 divOpToUse = new TmpInstruction(mcfi, divOp2);
2148 target.getInstrInfo().
2149 CreateSignExtensionInstructions(target,
2150 remI->getParent()->getParent(),
2152 8*opSize, mvec, mcfi);
2156 // Now compute: result = rem V1, V2 as:
2157 // result = V1 - (V1 / signExtend(V2)) * signExtend(V2)
2159 TmpInstruction* quot = new TmpInstruction(mcfi, divOp1, divOpToUse);
2160 TmpInstruction* prod = new TmpInstruction(mcfi, quot, divOpToUse);
2162 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
2163 .addReg(divOp1).addReg(divOpToUse).addRegDef(quot));
2165 mvec.push_back(BuildMI(ChooseMulInstructionByType(remI->getType()), 3)
2166 .addReg(quot).addReg(divOpToUse).addRegDef(prod));
2168 mvec.push_back(BuildMI(ChooseSubInstructionByType(remI->getType()), 3)
2169 .addReg(divOp1).addReg(prod).addRegDef(remI));
2174 case 38: // bool: And(bool, bool)
2175 case 138: // bool: And(bool, not)
2176 case 238: // bool: And(bool, boolconst)
2177 case 338: // reg : BAnd(reg, reg)
2178 case 538: // reg : BAnd(reg, Constant)
2179 Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
2182 case 438: // bool: BAnd(bool, bnot)
2183 { // Use the argument of NOT as the second argument!
2184 // Mark the NOT node so that no code is generated for it.
2185 // If the type is boolean, set 1 or 0 in the result register.
2186 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2187 Value* notArg = BinaryOperator::getNotArgument(
2188 cast<BinaryOperator>(notNode->getInstruction()));
2189 notNode->markFoldedIntoParent();
2190 Value *lhs = subtreeRoot->leftChild()->getValue();
2191 Value *dest = subtreeRoot->getValue();
2192 mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(lhs).addReg(notArg)
2193 .addReg(dest, MOTy::Def));
2195 if (notArg->getType() == Type::BoolTy)
2196 { // set 1 in result register if result of above is non-zero
2197 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2198 .addReg(dest, MOTy::UseAndDef));
2204 case 39: // bool: Or(bool, bool)
2205 case 139: // bool: Or(bool, not)
2206 case 239: // bool: Or(bool, boolconst)
2207 case 339: // reg : BOr(reg, reg)
2208 case 539: // reg : BOr(reg, Constant)
2209 Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
2212 case 439: // bool: BOr(bool, bnot)
2213 { // Use the argument of NOT as the second argument!
2214 // Mark the NOT node so that no code is generated for it.
2215 // If the type is boolean, set 1 or 0 in the result register.
2216 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2217 Value* notArg = BinaryOperator::getNotArgument(
2218 cast<BinaryOperator>(notNode->getInstruction()));
2219 notNode->markFoldedIntoParent();
2220 Value *lhs = subtreeRoot->leftChild()->getValue();
2221 Value *dest = subtreeRoot->getValue();
2223 mvec.push_back(BuildMI(V9::ORNr, 3).addReg(lhs).addReg(notArg)
2224 .addReg(dest, MOTy::Def));
2226 if (notArg->getType() == Type::BoolTy)
2227 { // set 1 in result register if result of above is non-zero
2228 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2229 .addReg(dest, MOTy::UseAndDef));
2235 case 40: // bool: Xor(bool, bool)
2236 case 140: // bool: Xor(bool, not)
2237 case 240: // bool: Xor(bool, boolconst)
2238 case 340: // reg : BXor(reg, reg)
2239 case 540: // reg : BXor(reg, Constant)
2240 Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
2243 case 440: // bool: BXor(bool, bnot)
2244 { // Use the argument of NOT as the second argument!
2245 // Mark the NOT node so that no code is generated for it.
2246 // If the type is boolean, set 1 or 0 in the result register.
2247 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2248 Value* notArg = BinaryOperator::getNotArgument(
2249 cast<BinaryOperator>(notNode->getInstruction()));
2250 notNode->markFoldedIntoParent();
2251 Value *lhs = subtreeRoot->leftChild()->getValue();
2252 Value *dest = subtreeRoot->getValue();
2253 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(lhs).addReg(notArg)
2254 .addReg(dest, MOTy::Def));
2256 if (notArg->getType() == Type::BoolTy)
2257 { // set 1 in result register if result of above is non-zero
2258 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2259 .addReg(dest, MOTy::UseAndDef));
2264 case 41: // setCCconst: SetCC(reg, Constant)
2265 { // Comparison is with a constant:
2267 // If the bool result must be computed into a register (see below),
2268 // and the constant is int ZERO, we can use the MOVR[op] instructions
2269 // and avoid the SUBcc instruction entirely.
2270 // Otherwise this is just the same as case 42, so just fall through.
2272 // The result of the SetCC must be computed and stored in a register if
2273 // it is used outside the current basic block (so it must be computed
2274 // as a boolreg) or it is used by anything other than a branch.
2275 // We will use a conditional move to do this.
2277 Instruction* setCCInstr = subtreeRoot->getInstruction();
2278 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
2279 ! AllUsesAreBranches(setCCInstr));
2283 InstrTreeNode* constNode = subtreeRoot->rightChild();
2285 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
2286 Constant *constVal = cast<Constant>(constNode->getValue());
2289 if ((constVal->getType()->isInteger()
2290 || isa<PointerType>(constVal->getType()))
2291 && target.getInstrInfo().ConvertConstantToIntType(target,
2292 constVal, constVal->getType(), isValidConst) == 0
2295 // That constant is an integer zero after all...
2296 // Use a MOVR[op] to compute the boolean result
2297 // Unconditionally set register to 0
2298 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0)
2299 .addRegDef(setCCInstr));
2301 // Now conditionally move 1 into the register.
2302 // Mark the register as a use (as well as a def) because the old
2303 // value will be retained if the condition is false.
2304 MachineOpCode movOpCode = ChooseMovpregiForSetCC(subtreeRoot);
2305 mvec.push_back(BuildMI(movOpCode, 3)
2306 .addReg(subtreeRoot->leftChild()->getValue())
2307 .addZImm(1).addReg(setCCInstr, MOTy::UseAndDef));
2312 // ELSE FALL THROUGH
2315 case 42: // bool: SetCC(reg, reg):
2317 // This generates a SUBCC instruction, putting the difference in a
2318 // result reg. if needed, and/or setting a condition code if needed.
2320 Instruction* setCCInstr = subtreeRoot->getInstruction();
2321 Value* leftVal = subtreeRoot->leftChild()->getValue();
2322 Value* rightVal = subtreeRoot->rightChild()->getValue();
2323 const Type* opType = leftVal->getType();
2324 bool isFPCompare = opType->isFloatingPoint();
2326 // If the boolean result of the SetCC is used outside the current basic
2327 // block (so it must be computed as a boolreg) or is used by anything
2328 // other than a branch, the boolean must be computed and stored
2329 // in a result register. We will use a conditional move to do this.
2331 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
2332 ! AllUsesAreBranches(setCCInstr));
2334 // A TmpInstruction is created to represent the CC "result".
2335 // Unlike other instances of TmpInstruction, this one is used
2336 // by machine code of multiple LLVM instructions, viz.,
2337 // the SetCC and the branch. Make sure to get the same one!
2338 // Note that we do this even for FP CC registers even though they
2339 // are explicit operands, because the type of the operand
2340 // needs to be a floating point condition code, not an integer
2341 // condition code. Think of this as casting the bool result to
2342 // a FP condition code register.
2343 // Later, we mark the 4th operand as being a CC register, and as a def.
2345 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
2346 setCCInstr->getParent()->getParent(),
2348 MachineCodeForInstruction::get(setCCInstr));
2350 // If the operands are signed values smaller than 4 bytes, then they
2351 // must be sign-extended in order to do a valid 32-bit comparison
2352 // and get the right result in the 32-bit CC register (%icc).
2354 Value* leftOpToUse = leftVal;
2355 Value* rightOpToUse = rightVal;
2356 if (opType->isIntegral() && opType->isSigned()) {
2357 unsigned opSize = target.getTargetData().getTypeSize(opType);
2359 MachineCodeForInstruction& mcfi =
2360 MachineCodeForInstruction::get(setCCInstr);
2362 // create temporary virtual regs. to hold the sign-extensions
2363 leftOpToUse = new TmpInstruction(mcfi, leftVal);
2364 rightOpToUse = new TmpInstruction(mcfi, rightVal);
2366 // sign-extend each operand and put the result in the temporary reg.
2367 target.getInstrInfo().CreateSignExtensionInstructions
2368 (target, setCCInstr->getParent()->getParent(),
2369 leftVal, leftOpToUse, 8*opSize, mvec, mcfi);
2370 target.getInstrInfo().CreateSignExtensionInstructions
2371 (target, setCCInstr->getParent()->getParent(),
2372 rightVal, rightOpToUse, 8*opSize, mvec, mcfi);
2376 if (! isFPCompare) {
2377 // Integer condition: set CC and discard result.
2378 mvec.push_back(BuildMI(V9::SUBccr, 4)
2379 .addReg(leftOpToUse)
2380 .addReg(rightOpToUse)
2381 .addMReg(target.getRegInfo().getZeroRegNum(),MOTy::Def)
2382 .addCCReg(tmpForCC, MOTy::Def));
2384 // FP condition: dest of FCMP should be some FCCn register
2385 mvec.push_back(BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
2386 .addCCReg(tmpForCC, MOTy::Def)
2387 .addReg(leftOpToUse)
2388 .addReg(rightOpToUse));
2391 if (computeBoolVal) {
2392 MachineOpCode movOpCode = (isFPCompare
2393 ? ChooseMovFpcciInstruction(subtreeRoot)
2394 : ChooseMovpcciForSetCC(subtreeRoot));
2396 // Unconditionally set register to 0
2397 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
2400 // Now conditionally move 1 into the register.
2401 // Mark the register as a use (as well as a def) because the old
2402 // value will be retained if the condition is false.
2403 M = (BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(1)
2404 .addReg(setCCInstr, MOTy::UseAndDef));
2410 case 51: // reg: Load(reg)
2411 case 52: // reg: Load(ptrreg)
2412 SetOperandsForMemInstr(ChooseLoadInstruction(
2413 subtreeRoot->getValue()->getType()),
2414 mvec, subtreeRoot, target);
2417 case 55: // reg: GetElemPtr(reg)
2418 case 56: // reg: GetElemPtrIdx(reg,reg)
2419 // If the GetElemPtr was folded into the user (parent), it will be
2420 // caught above. For other cases, we have to compute the address.
2421 SetOperandsForMemInstr(V9::ADDr, mvec, subtreeRoot, target);
2424 case 57: // reg: Alloca: Implement as 1 instruction:
2425 { // add %fp, offsetFromFP -> result
2426 AllocationInst* instr =
2427 cast<AllocationInst>(subtreeRoot->getInstruction());
2429 target.getTargetData().getTypeSize(instr->getAllocatedType());
2431 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
2435 case 58: // reg: Alloca(reg): Implement as 3 instructions:
2436 // mul num, typeSz -> tmp
2437 // sub %sp, tmp -> %sp
2438 { // add %sp, frameSizeBelowDynamicArea -> result
2439 AllocationInst* instr =
2440 cast<AllocationInst>(subtreeRoot->getInstruction());
2441 const Type* eltType = instr->getAllocatedType();
2443 // If #elements is constant, use simpler code for fixed-size allocas
2444 int tsize = (int) target.getTargetData().getTypeSize(eltType);
2445 Value* numElementsVal = NULL;
2446 bool isArray = instr->isArrayAllocation();
2448 if (!isArray || isa<Constant>(numElementsVal = instr->getArraySize())) {
2449 // total size is constant: generate code for fixed-size alloca
2450 unsigned numElements = isArray?
2451 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
2452 CreateCodeForFixedSizeAlloca(target, instr, tsize,
2455 // total size is not constant.
2456 CreateCodeForVariableSizeAlloca(target, instr, tsize,
2457 numElementsVal, mvec);
2462 case 61: // reg: Call
2463 { // Generate a direct (CALL) or indirect (JMPL) call.
2464 // Mark the return-address register, the indirection
2465 // register (for indirect calls), the operands of the Call,
2466 // and the return value (if any) as implicit operands
2467 // of the machine instruction.
2469 // If this is a varargs function, floating point arguments
2470 // have to passed in integer registers so insert
2471 // copy-float-to-int instructions for each float operand.
2473 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
2474 Value *callee = callInstr->getCalledValue();
2475 Function* calledFunc = dyn_cast<Function>(callee);
2477 // Check if this is an intrinsic function that needs a special code
2478 // sequence (e.g., va_start). Indirect calls cannot be special.
2480 bool specialIntrinsic = false;
2481 LLVMIntrinsic::ID iid;
2482 if (calledFunc && (iid=(LLVMIntrinsic::ID)calledFunc->getIntrinsicID()))
2483 specialIntrinsic = CodeGenIntrinsic(iid, *callInstr, target, mvec);
2485 // If not, generate the normal call sequence for the function.
2486 // This can also handle any intrinsics that are just function calls.
2488 if (! specialIntrinsic) {
2489 Function* currentFunc = callInstr->getParent()->getParent();
2490 MachineFunction& MF = MachineFunction::get(currentFunc);
2491 MachineCodeForInstruction& mcfi =
2492 MachineCodeForInstruction::get(callInstr);
2493 const UltraSparcRegInfo& regInfo =
2494 (UltraSparcRegInfo&) target.getRegInfo();
2495 const TargetFrameInfo& frameInfo = target.getFrameInfo();
2497 // Create hidden virtual register for return address with type void*
2498 TmpInstruction* retAddrReg =
2499 new TmpInstruction(mcfi, PointerType::get(Type::VoidTy), callInstr);
2501 // Generate the machine instruction and its operands.
2502 // Use CALL for direct function calls; this optimistically assumes
2503 // the PC-relative address fits in the CALL address field (22 bits).
2504 // Use JMPL for indirect calls.
2505 // This will be added to mvec later, after operand copies.
2507 MachineInstr* callMI;
2508 if (calledFunc) // direct function call
2509 callMI = BuildMI(V9::CALL, 1).addPCDisp(callee);
2510 else // indirect function call
2511 callMI = (BuildMI(V9::JMPLCALLi,3).addReg(callee)
2512 .addSImm((int64_t)0).addRegDef(retAddrReg));
2514 const FunctionType* funcType =
2515 cast<FunctionType>(cast<PointerType>(callee->getType())
2516 ->getElementType());
2517 bool isVarArgs = funcType->isVarArg();
2518 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
2520 // Use a descriptor to pass information about call arguments
2521 // to the register allocator. This descriptor will be "owned"
2522 // and freed automatically when the MachineCodeForInstruction
2523 // object for the callInstr goes away.
2524 CallArgsDescriptor* argDesc =
2525 new CallArgsDescriptor(callInstr, retAddrReg,isVarArgs,noPrototype);
2526 assert(callInstr->getOperand(0) == callee
2527 && "This is assumed in the loop below!");
2529 // Insert sign-extension instructions for small signed values,
2530 // if this is an unknown function (i.e., called via a funcptr)
2531 // or an external one (i.e., which may not be compiled by llc).
2533 if (calledFunc == NULL || calledFunc->isExternal()) {
2534 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2535 Value* argVal = callInstr->getOperand(i);
2536 const Type* argType = argVal->getType();
2537 if (argType->isIntegral() && argType->isSigned()) {
2538 unsigned argSize = target.getTargetData().getTypeSize(argType);
2540 // create a temporary virtual reg. to hold the sign-extension
2541 TmpInstruction* argExtend = new TmpInstruction(mcfi, argVal);
2543 // sign-extend argVal and put the result in the temporary reg.
2544 target.getInstrInfo().CreateSignExtensionInstructions
2545 (target, currentFunc, argVal, argExtend,
2546 8*argSize, mvec, mcfi);
2548 // replace argVal with argExtend in CallArgsDescriptor
2549 argDesc->getArgInfo(i-1).replaceArgVal(argExtend);
2555 // Insert copy instructions to get all the arguments into
2556 // all the places that they need to be.
2558 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2560 CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
2561 Value* argVal = argInfo.getArgVal(); // don't use callInstr arg here
2562 const Type* argType = argVal->getType();
2563 unsigned regType = regInfo.getRegTypeForDataType(argType);
2564 unsigned argSize = target.getTargetData().getTypeSize(argType);
2565 int regNumForArg = TargetRegInfo::getInvalidRegNum();
2566 unsigned regClassIDOfArgReg;
2568 // Check for FP arguments to varargs functions.
2569 // Any such argument in the first $K$ args must be passed in an
2570 // integer register. If there is no prototype, it must also
2571 // be passed as an FP register.
2572 // K = #integer argument registers.
2573 bool isFPArg = argVal->getType()->isFloatingPoint();
2574 if (isVarArgs && isFPArg) {
2577 // It is a function with no prototype: pass value
2578 // as an FP value as well as a varargs value. The FP value
2579 // may go in a register or on the stack. The copy instruction
2580 // to the outgoing reg/stack is created by the normal argument
2581 // handling code since this is the "normal" passing mode.
2583 regNumForArg = regInfo.regNumForFPArg(regType,
2584 false, false, argNo,
2585 regClassIDOfArgReg);
2586 if (regNumForArg == regInfo.getInvalidRegNum())
2587 argInfo.setUseStackSlot();
2589 argInfo.setUseFPArgReg();
2592 // If this arg. is in the first $K$ regs, add special copy-
2593 // float-to-int instructions to pass the value as an int.
2594 // To check if it is in the first $K$, get the register
2595 // number for the arg #i. These copy instructions are
2596 // generated here because they are extra cases and not needed
2597 // for the normal argument handling (some code reuse is
2598 // possible though -- later).
2600 int copyRegNum = regInfo.regNumForIntArg(false, false, argNo,
2601 regClassIDOfArgReg);
2602 if (copyRegNum != regInfo.getInvalidRegNum()) {
2603 // Create a virtual register to represent copyReg. Mark
2604 // this vreg as being an implicit operand of the call MI
2605 const Type* loadTy = (argType == Type::FloatTy
2606 ? Type::IntTy : Type::LongTy);
2607 TmpInstruction* argVReg = new TmpInstruction(mcfi, loadTy,
2610 callMI->addImplicitRef(argVReg);
2612 // Get a temp stack location to use to copy
2613 // float-to-int via the stack.
2615 // FIXME: For now, we allocate permanent space because
2616 // the stack frame manager does not allow locals to be
2617 // allocated (e.g., for alloca) after a temp is
2620 // int tmpOffset = MF.getInfo()->pushTempValue(argSize);
2621 int tmpOffset = MF.getInfo()->allocateLocalVar(argVReg);
2623 // Generate the store from FP reg to stack
2624 unsigned StoreOpcode = ChooseStoreInstruction(argType);
2625 M = BuildMI(convertOpcodeFromRegToImm(StoreOpcode), 3)
2626 .addReg(argVal).addMReg(regInfo.getFramePointer())
2627 .addSImm(tmpOffset);
2630 // Generate the load from stack to int arg reg
2631 unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
2632 M = BuildMI(convertOpcodeFromRegToImm(LoadOpcode), 3)
2633 .addMReg(regInfo.getFramePointer()).addSImm(tmpOffset)
2634 .addReg(argVReg, MOTy::Def);
2636 // Mark operand with register it should be assigned
2637 // both for copy and for the callMI
2638 M->SetRegForOperand(M->getNumOperands()-1, copyRegNum);
2639 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2643 // Add info about the argument to the CallArgsDescriptor
2644 argInfo.setUseIntArgReg();
2645 argInfo.setArgCopy(copyRegNum);
2647 // Cannot fit in first $K$ regs so pass arg on stack
2648 argInfo.setUseStackSlot();
2650 } else if (isFPArg) {
2651 // Get the outgoing arg reg to see if there is one.
2652 regNumForArg = regInfo.regNumForFPArg(regType, false, false,
2653 argNo, regClassIDOfArgReg);
2654 if (regNumForArg == regInfo.getInvalidRegNum())
2655 argInfo.setUseStackSlot();
2657 argInfo.setUseFPArgReg();
2658 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2662 // Get the outgoing arg reg to see if there is one.
2663 regNumForArg = regInfo.regNumForIntArg(false,false,
2664 argNo, regClassIDOfArgReg);
2665 if (regNumForArg == regInfo.getInvalidRegNum())
2666 argInfo.setUseStackSlot();
2668 argInfo.setUseIntArgReg();
2669 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2675 // Now insert copy instructions to stack slot or arg. register
2677 if (argInfo.usesStackSlot()) {
2678 // Get the stack offset for this argument slot.
2679 // FP args on stack are right justified so adjust offset!
2680 // int arguments are also right justified but they are
2681 // always loaded as a full double-word so the offset does
2682 // not need to be adjusted.
2683 int argOffset = frameInfo.getOutgoingArgOffset(MF, argNo);
2684 if (argType->isFloatingPoint()) {
2685 unsigned slotSize = frameInfo.getSizeOfEachArgOnStack();
2686 assert(argSize <= slotSize && "Insufficient slot size!");
2687 argOffset += slotSize - argSize;
2690 // Now generate instruction to copy argument to stack
2691 MachineOpCode storeOpCode =
2692 (argType->isFloatingPoint()
2693 ? ((argSize == 4)? V9::STFi : V9::STDFi) : V9::STXi);
2695 M = BuildMI(storeOpCode, 3).addReg(argVal)
2696 .addMReg(regInfo.getStackPointer()).addSImm(argOffset);
2699 else if (regNumForArg != regInfo.getInvalidRegNum()) {
2701 // Create a virtual register to represent the arg reg. Mark
2702 // this vreg as being an implicit operand of the call MI.
2703 TmpInstruction* argVReg =
2704 new TmpInstruction(mcfi, argVal, NULL, "argReg");
2706 callMI->addImplicitRef(argVReg);
2708 // Generate the reg-to-reg copy into the outgoing arg reg.
2709 // -- For FP values, create a FMOVS or FMOVD instruction
2710 // -- For non-FP values, create an add-with-0 instruction
2711 if (argType->isFloatingPoint())
2712 M=(BuildMI(argType==Type::FloatTy? V9::FMOVS :V9::FMOVD,2)
2713 .addReg(argVal).addReg(argVReg, MOTy::Def));
2715 M = (BuildMI(ChooseAddInstructionByType(argType), 3)
2716 .addReg(argVal).addSImm((int64_t) 0)
2717 .addReg(argVReg, MOTy::Def));
2719 // Mark the operand with the register it should be assigned
2720 M->SetRegForOperand(M->getNumOperands()-1, regNumForArg);
2721 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2727 assert(argInfo.getArgCopy() != regInfo.getInvalidRegNum() &&
2728 "Arg. not in stack slot, primary or secondary register?");
2731 // add call instruction and delay slot before copying return value
2732 mvec.push_back(callMI);
2733 mvec.push_back(BuildMI(V9::NOP, 0));
2735 // Add the return value as an implicit ref. The call operands
2736 // were added above. Also, add code to copy out the return value.
2737 // This is always register-to-register for int or FP return values.
2739 if (callInstr->getType() != Type::VoidTy) {
2740 // Get the return value reg.
2741 const Type* retType = callInstr->getType();
2743 int regNum = (retType->isFloatingPoint()
2744 ? (unsigned) SparcFloatRegClass::f0
2745 : (unsigned) SparcIntRegClass::o0);
2746 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
2747 regNum = regInfo.getUnifiedRegNum(regClassID, regNum);
2749 // Create a virtual register to represent it and mark
2750 // this vreg as being an implicit operand of the call MI
2751 TmpInstruction* retVReg =
2752 new TmpInstruction(mcfi, callInstr, NULL, "argReg");
2754 callMI->addImplicitRef(retVReg, /*isDef*/ true);
2756 // Generate the reg-to-reg copy from the return value reg.
2757 // -- For FP values, create a FMOVS or FMOVD instruction
2758 // -- For non-FP values, create an add-with-0 instruction
2759 if (retType->isFloatingPoint())
2760 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
2761 .addReg(retVReg).addReg(callInstr, MOTy::Def));
2763 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
2764 .addReg(retVReg).addSImm((int64_t) 0)
2765 .addReg(callInstr, MOTy::Def));
2767 // Mark the operand with the register it should be assigned
2768 // Also mark the implicit ref of the call defining this operand
2769 M->SetRegForOperand(0, regNum);
2770 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,regNum);
2775 // For the CALL instruction, the ret. addr. reg. is also implicit
2776 if (isa<Function>(callee))
2777 callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
2779 MF.getInfo()->popAllTempValues(); // free temps used for this inst
2785 case 62: // reg: Shl(reg, reg)
2787 Value* argVal1 = subtreeRoot->leftChild()->getValue();
2788 Value* argVal2 = subtreeRoot->rightChild()->getValue();
2789 Instruction* shlInstr = subtreeRoot->getInstruction();
2791 const Type* opType = argVal1->getType();
2792 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2793 "Shl unsupported for other types");
2794 unsigned opSize = target.getTargetData().getTypeSize(opType);
2796 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
2797 (opSize > 4)? V9::SLLXr6:V9::SLLr5,
2798 argVal1, argVal2, 0, shlInstr, mvec,
2799 MachineCodeForInstruction::get(shlInstr));
2803 case 63: // reg: Shr(reg, reg)
2805 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
2806 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2807 "Shr unsupported for other types");
2808 unsigned opSize = target.getTargetData().getTypeSize(opType);
2809 Add3OperandInstr(opType->isSigned()
2810 ? (opSize > 4? V9::SRAXr6 : V9::SRAr5)
2811 : (opSize > 4? V9::SRLXr6 : V9::SRLr5),
2816 case 64: // reg: Phi(reg,reg)
2817 break; // don't forward the value
2819 case 65: // reg: VANext(reg): the va_next(va_list, type) instruction
2820 { // Increment the va_list pointer register according to the type.
2821 // All LLVM argument types are <= 64 bits, so use one doubleword.
2822 Instruction* vaNextI = subtreeRoot->getInstruction();
2823 assert(target.getTargetData().getTypeSize(vaNextI->getType()) <= 8 &&
2824 "We assumed that all LLVM parameter types <= 8 bytes!");
2825 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
2826 mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaNextI->getOperand(0)).
2827 addSImm(argSize).addRegDef(vaNextI));
2831 case 66: // reg: VAArg (reg): the va_arg instruction
2832 { // Load argument from stack using current va_list pointer value.
2833 // Use 64-bit load for all non-FP args, and LDDF or double for FP.
2834 Instruction* vaArgI = subtreeRoot->getInstruction();
2835 MachineOpCode loadOp = (vaArgI->getType()->isFloatingPoint()
2836 ? (vaArgI->getType() == Type::FloatTy
2837 ? V9::LDFi : V9::LDDFi)
2839 mvec.push_back(BuildMI(loadOp, 3).addReg(vaArgI->getOperand(0)).
2840 addSImm(0).addRegDef(vaArgI));
2844 case 71: // reg: VReg
2845 case 72: // reg: Constant
2846 break; // don't forward the value
2849 assert(0 && "Unrecognized BURG rule");
2854 if (forwardOperandNum >= 0) {
2855 // We did not generate a machine instruction but need to use operand.
2856 // If user is in the same tree, replace Value in its machine operand.
2857 // If not, insert a copy instruction which should get coalesced away
2858 // by register allocation.
2859 if (subtreeRoot->parent() != NULL)
2860 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
2862 std::vector<MachineInstr*> minstrVec;
2863 Instruction* instr = subtreeRoot->getInstruction();
2864 target.getInstrInfo().
2865 CreateCopyInstructionsByType(target,
2866 instr->getParent()->getParent(),
2867 instr->getOperand(forwardOperandNum),
2869 MachineCodeForInstruction::get(instr));
2870 assert(minstrVec.size() > 0);
2871 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
2875 if (maskUnsignedResult) {
2876 // If result is unsigned and smaller than int reg size,
2877 // we need to clear high bits of result value.
2878 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
2879 Instruction* dest = subtreeRoot->getInstruction();
2880 if (dest->getType()->isUnsigned()) {
2881 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
2882 if (destSize <= 4) {
2883 // Mask high 64 - N bits, where N = 4*destSize.
2885 // Use a TmpInstruction to represent the
2886 // intermediate result before masking. Since those instructions
2887 // have already been generated, go back and substitute tmpI
2888 // for dest in the result position of each one of them.
2890 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(dest);
2891 TmpInstruction *tmpI = new TmpInstruction(mcfi, dest->getType(),
2892 dest, NULL, "maskHi");
2893 Value* srlArgToUse = tmpI;
2895 unsigned numSubst = 0;
2896 for (unsigned i=0, N=mvec.size(); i < N; ++i) {
2898 // Make sure we substitute all occurrences of dest in these instrs.
2899 // Otherwise, we will have bogus code.
2900 bool someArgsWereIgnored = false;
2902 // Make sure not to substitute an upwards-exposed use -- that would
2903 // introduce a use of `tmpI' with no preceding def. Therefore,
2904 // substitute a use or def-and-use operand only if a previous def
2905 // operand has already been substituted (i.e., numSusbt > 0).
2907 numSubst += mvec[i]->substituteValue(dest, tmpI,
2908 /*defsOnly*/ numSubst == 0,
2909 /*notDefsAndUses*/ numSubst > 0,
2910 someArgsWereIgnored);
2911 assert(!someArgsWereIgnored &&
2912 "Operand `dest' exists but not replaced: probably bogus!");
2914 assert(numSubst > 0 && "Operand `dest' not replaced: probably bogus!");
2916 // Left shift 32-N if size (N) is less than 32 bits.
2917 // Use another tmp. virtual register to represent this result.
2919 srlArgToUse = new TmpInstruction(mcfi, dest->getType(),
2920 tmpI, NULL, "maskHi2");
2921 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpI)
2922 .addZImm(8*(4-destSize))
2923 .addReg(srlArgToUse, MOTy::Def));
2926 // Logical right shift 32-N to get zero extension in top 64-N bits.
2927 mvec.push_back(BuildMI(V9::SRLi5, 3).addReg(srlArgToUse)
2928 .addZImm(8*(4-destSize)).addReg(dest, MOTy::Def));
2930 } else if (destSize < 8) {
2931 assert(0 && "Unsupported type size: 32 < size < 64 bits");