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()));
44 //---------------------------------------------------------------------------
45 // Function: GetMemInstArgs
48 // Get the pointer value and the index vector for a memory operation
49 // (GetElementPtr, Load, or Store). If all indices of the given memory
50 // operation are constant, fold in constant indices in a chain of
51 // preceding GetElementPtr instructions (if any), and return the
52 // pointer value of the first instruction in the chain.
53 // All folded instructions are marked so no code is generated for them.
56 // Returns the pointer Value to use.
57 // Returns the resulting IndexVector in idxVec.
58 // Returns true/false in allConstantIndices if all indices are/aren't const.
59 //---------------------------------------------------------------------------
62 //---------------------------------------------------------------------------
63 // Function: FoldGetElemChain
66 // Fold a chain of GetElementPtr instructions containing only
67 // constant offsets into an equivalent (Pointer, IndexVector) pair.
68 // Returns the pointer Value, and stores the resulting IndexVector
69 // in argument chainIdxVec. This is a helper function for
70 // FoldConstantIndices that does the actual folding.
71 //---------------------------------------------------------------------------
74 // Check for a constant 0.
78 return (idx == ConstantSInt::getNullValue(idx->getType()));
82 FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
83 bool lastInstHasLeadingNonZero)
85 InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
86 GetElementPtrInst* gepInst =
87 dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
89 // ptr value is not computed in this tree or ptr value does not come from GEP
94 // Return NULL if we don't fold any instructions in.
97 // Now chase the chain of getElementInstr instructions, if any.
98 // Check for any non-constant indices and stop there.
99 // Also, stop if the first index of child is a non-zero array index
100 // and the last index of the current node is a non-array index:
101 // in that case, a non-array declared type is being accessed as an array
102 // which is not type-safe, but could be legal.
104 InstructionNode* ptrChild = gepNode;
105 while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
106 ptrChild->getOpLabel() == GetElemPtrIdx))
108 // Child is a GetElemPtr instruction
109 gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
110 User::op_iterator OI, firstIdx = gepInst->idx_begin();
111 User::op_iterator lastIdx = gepInst->idx_end();
112 bool allConstantOffsets = true;
114 // The first index of every GEP must be an array index.
115 assert((*firstIdx)->getType() == Type::LongTy &&
116 "INTERNAL ERROR: Structure index for a pointer type!");
118 // If the last instruction had a leading non-zero index, check if the
119 // current one references a sequential (i.e., indexable) type.
120 // If not, the code is not type-safe and we would create an illegal GEP
121 // by folding them, so don't fold any more instructions.
123 if (lastInstHasLeadingNonZero)
124 if (! isa<SequentialType>(gepInst->getType()->getElementType()))
125 break; // cannot fold in any preceding getElementPtr instrs.
127 // Check that all offsets are constant for this instruction
128 for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
129 allConstantOffsets = isa<ConstantInt>(*OI);
131 if (allConstantOffsets) {
132 // Get pointer value out of ptrChild.
133 ptrVal = gepInst->getPointerOperand();
135 // Insert its index vector at the start, skipping any leading [0]
136 // Remember the old size to check if anything was inserted.
137 unsigned oldSize = chainIdxVec.size();
138 int firstIsZero = IsZero(*firstIdx);
139 chainIdxVec.insert(chainIdxVec.begin(), firstIdx + firstIsZero, lastIdx);
141 // Remember if it has leading zero index: it will be discarded later.
142 if (oldSize < chainIdxVec.size())
143 lastInstHasLeadingNonZero = !firstIsZero;
145 // Mark the folded node so no code is generated for it.
146 ((InstructionNode*) ptrChild)->markFoldedIntoParent();
148 // Get the previous GEP instruction and continue trying to fold
149 ptrChild = dyn_cast<InstructionNode>(ptrChild->leftChild());
150 } else // cannot fold this getElementPtr instr. or any preceding ones
154 // If the first getElementPtr instruction had a leading [0], add it back.
155 // Note that this instruction is the *last* one that was successfully
156 // folded *and* contributed any indices, in the loop above.
158 if (ptrVal && ! lastInstHasLeadingNonZero)
159 chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
165 //---------------------------------------------------------------------------
166 // Function: GetGEPInstArgs
169 // Helper function for GetMemInstArgs that handles the final getElementPtr
170 // instruction used by (or same as) the memory operation.
171 // Extracts the indices of the current instruction and tries to fold in
172 // preceding ones if all indices of the current one are constant.
173 //---------------------------------------------------------------------------
176 GetGEPInstArgs(InstructionNode* gepNode,
177 std::vector<Value*>& idxVec,
178 bool& allConstantIndices)
180 allConstantIndices = true;
181 GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
183 // Default pointer is the one from the current instruction.
184 Value* ptrVal = gepI->getPointerOperand();
185 InstrTreeNode* ptrChild = gepNode->leftChild();
187 // Extract the index vector of the GEP instruction.
188 // If all indices are constant and first index is zero, try to fold
189 // in preceding GEPs with all constant indices.
190 for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
191 allConstantIndices && OI != OE; ++OI)
192 if (! isa<Constant>(*OI))
193 allConstantIndices = false; // note: this also terminates loop!
195 // If we have only constant indices, fold chains of constant indices
196 // in this and any preceding GetElemPtr instructions.
197 bool foldedGEPs = false;
198 bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
199 if (allConstantIndices)
200 if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx)) {
205 // Append the index vector of the current instruction.
206 // Skip the leading [0] index if preceding GEPs were folded into this.
207 idxVec.insert(idxVec.end(),
208 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
214 //---------------------------------------------------------------------------
215 // Function: GetMemInstArgs
218 // Get the pointer value and the index vector for a memory operation
219 // (GetElementPtr, Load, or Store). If all indices of the given memory
220 // operation are constant, fold in constant indices in a chain of
221 // preceding GetElementPtr instructions (if any), and return the
222 // pointer value of the first instruction in the chain.
223 // All folded instructions are marked so no code is generated for them.
226 // Returns the pointer Value to use.
227 // Returns the resulting IndexVector in idxVec.
228 // Returns true/false in allConstantIndices if all indices are/aren't const.
229 //---------------------------------------------------------------------------
232 GetMemInstArgs(InstructionNode* memInstrNode,
233 std::vector<Value*>& idxVec,
234 bool& allConstantIndices)
236 allConstantIndices = false;
237 Instruction* memInst = memInstrNode->getInstruction();
238 assert(idxVec.size() == 0 && "Need empty vector to return indices");
240 // If there is a GetElemPtr instruction to fold in to this instr,
241 // it must be in the left child for Load and GetElemPtr, and in the
242 // right child for Store instructions.
243 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
244 ? memInstrNode->rightChild()
245 : memInstrNode->leftChild());
247 // Default pointer is the one from the current instruction.
248 Value* ptrVal = ptrChild->getValue();
250 // Find the "last" GetElemPtr instruction: this one or the immediate child.
251 // There will be none if this is a load or a store from a scalar pointer.
252 InstructionNode* gepNode = NULL;
253 if (isa<GetElementPtrInst>(memInst))
254 gepNode = memInstrNode;
255 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
256 // Child of load/store is a GEP and memInst is its only use.
257 // Use its indices and mark it as folded.
258 gepNode = cast<InstructionNode>(ptrChild);
259 gepNode->markFoldedIntoParent();
262 // If there are no indices, return the current pointer.
263 // Else extract the pointer from the GEP and fold the indices.
264 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
269 //************************ Internal Functions ******************************/
272 static inline MachineOpCode
273 ChooseBprInstruction(const InstructionNode* instrNode)
275 MachineOpCode opCode;
277 Instruction* setCCInstr =
278 ((InstructionNode*) instrNode->leftChild())->getInstruction();
280 switch(setCCInstr->getOpcode())
282 case Instruction::SetEQ: opCode = V9::BRZ; break;
283 case Instruction::SetNE: opCode = V9::BRNZ; break;
284 case Instruction::SetLE: opCode = V9::BRLEZ; break;
285 case Instruction::SetGE: opCode = V9::BRGEZ; break;
286 case Instruction::SetLT: opCode = V9::BRLZ; break;
287 case Instruction::SetGT: opCode = V9::BRGZ; break;
289 assert(0 && "Unrecognized VM instruction!");
290 opCode = V9::INVALID_OPCODE;
298 static inline MachineOpCode
299 ChooseBpccInstruction(const InstructionNode* instrNode,
300 const BinaryOperator* setCCInstr)
302 MachineOpCode opCode = V9::INVALID_OPCODE;
304 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
307 switch(setCCInstr->getOpcode())
309 case Instruction::SetEQ: opCode = V9::BE; break;
310 case Instruction::SetNE: opCode = V9::BNE; break;
311 case Instruction::SetLE: opCode = V9::BLE; break;
312 case Instruction::SetGE: opCode = V9::BGE; break;
313 case Instruction::SetLT: opCode = V9::BL; break;
314 case Instruction::SetGT: opCode = V9::BG; break;
316 assert(0 && "Unrecognized VM instruction!");
320 switch(setCCInstr->getOpcode())
322 case Instruction::SetEQ: opCode = V9::BE; break;
323 case Instruction::SetNE: opCode = V9::BNE; break;
324 case Instruction::SetLE: opCode = V9::BLEU; break;
325 case Instruction::SetGE: opCode = V9::BCC; break;
326 case Instruction::SetLT: opCode = V9::BCS; break;
327 case Instruction::SetGT: opCode = V9::BGU; break;
329 assert(0 && "Unrecognized VM instruction!");
337 static inline MachineOpCode
338 ChooseBFpccInstruction(const InstructionNode* instrNode,
339 const BinaryOperator* setCCInstr)
341 MachineOpCode opCode = V9::INVALID_OPCODE;
343 switch(setCCInstr->getOpcode())
345 case Instruction::SetEQ: opCode = V9::FBE; break;
346 case Instruction::SetNE: opCode = V9::FBNE; break;
347 case Instruction::SetLE: opCode = V9::FBLE; break;
348 case Instruction::SetGE: opCode = V9::FBGE; break;
349 case Instruction::SetLT: opCode = V9::FBL; break;
350 case Instruction::SetGT: opCode = V9::FBG; break;
352 assert(0 && "Unrecognized VM instruction!");
360 // Create a unique TmpInstruction for a boolean value,
361 // representing the CC register used by a branch on that value.
362 // For now, hack this using a little static cache of TmpInstructions.
363 // Eventually the entire BURG instruction selection should be put
364 // into a separate class that can hold such information.
365 // The static cache is not too bad because the memory for these
366 // TmpInstructions will be freed along with the rest of the Function anyway.
368 static TmpInstruction*
369 GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType,
370 MachineCodeForInstruction& mcfi)
372 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
373 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
374 static const Function *lastFunction = 0;// Use to flush cache between funcs
376 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
378 if (lastFunction != F) {
380 boolToTmpCache.clear();
383 // Look for tmpI and create a new one otherwise. The new value is
384 // directly written to map using the ref returned by operator[].
385 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
387 tmpI = new TmpInstruction(mcfi, ccType, boolVal);
393 static inline MachineOpCode
394 ChooseBccInstruction(const InstructionNode* instrNode,
395 const Type*& setCCType)
397 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
398 assert(setCCNode->getOpLabel() == SetCCOp);
399 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
400 setCCType = setCCInstr->getOperand(0)->getType();
402 if (setCCType->isFloatingPoint())
403 return ChooseBFpccInstruction(instrNode, setCCInstr);
405 return ChooseBpccInstruction(instrNode, setCCInstr);
409 // WARNING: since this function has only one caller, it always returns
410 // the opcode that expects an immediate and a register. If this function
411 // is ever used in cases where an opcode that takes two registers is required,
412 // then modify this function and use convertOpcodeFromRegToImm() where required.
414 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
415 // new cases of opcodes.
416 static inline MachineOpCode
417 ChooseMovFpcciInstruction(const InstructionNode* instrNode)
419 MachineOpCode opCode = V9::INVALID_OPCODE;
421 switch(instrNode->getInstruction()->getOpcode())
423 case Instruction::SetEQ: opCode = V9::MOVFEi; break;
424 case Instruction::SetNE: opCode = V9::MOVFNEi; break;
425 case Instruction::SetLE: opCode = V9::MOVFLEi; break;
426 case Instruction::SetGE: opCode = V9::MOVFGEi; break;
427 case Instruction::SetLT: opCode = V9::MOVFLi; break;
428 case Instruction::SetGT: opCode = V9::MOVFGi; break;
430 assert(0 && "Unrecognized VM instruction!");
438 // ChooseMovpcciForSetCC -- Choose a conditional-move instruction
439 // based on the type of SetCC operation.
441 // WARNING: since this function has only one caller, it always returns
442 // the opcode that expects an immediate and a register. If this function
443 // is ever used in cases where an opcode that takes two registers is required,
444 // then modify this function and use convertOpcodeFromRegToImm() where required.
446 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
447 // new cases of opcodes.
450 ChooseMovpcciForSetCC(const InstructionNode* instrNode)
452 MachineOpCode opCode = V9::INVALID_OPCODE;
454 const Type* opType = instrNode->leftChild()->getValue()->getType();
455 assert(opType->isIntegral() || isa<PointerType>(opType));
456 bool noSign = opType->isUnsigned() || isa<PointerType>(opType);
458 switch(instrNode->getInstruction()->getOpcode())
460 case Instruction::SetEQ: opCode = V9::MOVEi; break;
461 case Instruction::SetLE: opCode = noSign? V9::MOVLEUi : V9::MOVLEi; break;
462 case Instruction::SetGE: opCode = noSign? V9::MOVCCi : V9::MOVGEi; break;
463 case Instruction::SetLT: opCode = noSign? V9::MOVCSi : V9::MOVLi; break;
464 case Instruction::SetGT: opCode = noSign? V9::MOVGUi : V9::MOVGi; break;
465 case Instruction::SetNE: opCode = V9::MOVNEi; break;
466 default: assert(0 && "Unrecognized LLVM instr!"); break;
473 // ChooseMovpregiForSetCC -- Choose a conditional-move-on-register-value
474 // instruction based on the type of SetCC operation. These instructions
475 // compare a register with 0 and perform the move is the comparison is true.
477 // WARNING: like the previous function, this function it always returns
478 // the opcode that expects an immediate and a register. See above.
481 ChooseMovpregiForSetCC(const InstructionNode* instrNode)
483 MachineOpCode opCode = V9::INVALID_OPCODE;
485 switch(instrNode->getInstruction()->getOpcode())
487 case Instruction::SetEQ: opCode = V9::MOVRZi; break;
488 case Instruction::SetLE: opCode = V9::MOVRLEZi; break;
489 case Instruction::SetGE: opCode = V9::MOVRGEZi; break;
490 case Instruction::SetLT: opCode = V9::MOVRLZi; break;
491 case Instruction::SetGT: opCode = V9::MOVRGZi; break;
492 case Instruction::SetNE: opCode = V9::MOVRNZi; break;
493 default: assert(0 && "Unrecognized VM instr!"); break;
500 static inline MachineOpCode
501 ChooseConvertToFloatInstr(const TargetMachine& target,
502 OpLabel vopCode, const Type* opType)
504 assert((vopCode == ToFloatTy || vopCode == ToDoubleTy) &&
505 "Unrecognized convert-to-float opcode!");
506 assert((opType->isIntegral() || opType->isFloatingPoint() ||
507 isa<PointerType>(opType))
508 && "Trying to convert a non-scalar type to FLOAT/DOUBLE?");
510 MachineOpCode opCode = V9::INVALID_OPCODE;
512 unsigned opSize = target.getTargetData().getTypeSize(opType);
514 if (opType == Type::FloatTy)
515 opCode = (vopCode == ToFloatTy? V9::NOP : V9::FSTOD);
516 else if (opType == Type::DoubleTy)
517 opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::NOP);
518 else if (opSize <= 4)
519 opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
521 assert(opSize == 8 && "Unrecognized type size > 4 and < 8!");
522 opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
528 static inline MachineOpCode
529 ChooseConvertFPToIntInstr(const TargetMachine& target,
530 const Type* destType, const Type* opType)
532 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
533 && "This function should only be called for FLOAT or DOUBLE");
534 assert((destType->isIntegral() || isa<PointerType>(destType))
535 && "Trying to convert FLOAT/DOUBLE to a non-scalar type?");
537 MachineOpCode opCode = V9::INVALID_OPCODE;
539 unsigned destSize = target.getTargetData().getTypeSize(destType);
541 if (destType == Type::UIntTy)
542 assert(destType != Type::UIntTy && "Expand FP-to-uint beforehand.");
543 else if (destSize <= 4)
544 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
546 assert(destSize == 8 && "Unrecognized type size > 4 and < 8!");
547 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
554 CreateConvertFPToIntInstr(const TargetMachine& target,
557 const Type* destType)
559 MachineOpCode opCode = ChooseConvertFPToIntInstr(target, destType,
561 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
562 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
565 // CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer
566 // The FP value must be converted to the dest type in an FP register,
567 // and the result is then copied from FP to int register via memory.
568 // SPARC does not have a float-to-uint conversion, only a float-to-int (fdtoi).
569 // Since fdtoi converts to signed integers, any FP value V between MAXINT+1
570 // and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly.
571 // Therefore, for converting an FP value to uint32_t, we first need to convert
572 // to uint64_t and then to uint32_t.
575 CreateCodeToConvertFloatToInt(const TargetMachine& target,
578 std::vector<MachineInstr*>& mvec,
579 MachineCodeForInstruction& mcfi)
581 Function* F = destI->getParent()->getParent();
583 // Create a temporary to represent the FP register into which the
584 // int value will placed after conversion. The type of this temporary
585 // depends on the type of FP register to use: single-prec for a 32-bit
586 // int or smaller; double-prec for a 64-bit int.
588 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
590 const Type* castDestType = destI->getType(); // type for the cast instr result
591 const Type* castDestRegType; // type for cast instruction result reg
592 TmpInstruction* destForCast; // dest for cast instruction
593 Instruction* fpToIntCopyDest = destI; // dest for fp-reg-to-int-reg copy instr
595 // For converting an FP value to uint32_t, we first need to convert to
596 // uint64_t and then to uint32_t, as explained above.
597 if (destI->getType() == Type::UIntTy) {
598 castDestType = Type::ULongTy; // use this instead of type of destI
599 castDestRegType = Type::DoubleTy; // uint64_t needs 64-bit FP register.
600 destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
601 fpToIntCopyDest = new TmpInstruction(mcfi, castDestType, destForCast);
604 castDestRegType = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
605 destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
608 // Create the fp-to-int conversion instruction (src and dest regs are FP regs)
609 mvec.push_back(CreateConvertFPToIntInstr(target, opVal, destForCast,
612 // Create the fpreg-to-intreg copy code
613 target.getInstrInfo().CreateCodeToCopyFloatToInt(target, F, destForCast,
614 fpToIntCopyDest, mvec, mcfi);
616 // Create the uint64_t to uint32_t conversion, if needed
617 if (destI->getType() == Type::UIntTy)
618 target.getInstrInfo().
619 CreateZeroExtensionInstructions(target, F, fpToIntCopyDest, destI,
620 /*numLowBits*/ 32, mvec, mcfi);
624 static inline MachineOpCode
625 ChooseAddInstruction(const InstructionNode* instrNode)
627 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
631 static inline MachineInstr*
632 CreateMovFloatInstruction(const InstructionNode* instrNode,
633 const Type* resultType)
635 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
636 .addReg(instrNode->leftChild()->getValue())
637 .addRegDef(instrNode->getValue());
640 static inline MachineInstr*
641 CreateAddConstInstruction(const InstructionNode* instrNode)
643 MachineInstr* minstr = NULL;
645 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
646 assert(isa<Constant>(constOp));
648 // Cases worth optimizing are:
649 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
650 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
652 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
653 double dval = FPC->getValue();
655 minstr = CreateMovFloatInstruction(instrNode,
656 instrNode->getInstruction()->getType());
663 static inline MachineOpCode
664 ChooseSubInstructionByType(const Type* resultType)
666 MachineOpCode opCode = V9::INVALID_OPCODE;
668 if (resultType->isInteger() || isa<PointerType>(resultType)) {
671 switch(resultType->getPrimitiveID())
673 case Type::FloatTyID: opCode = V9::FSUBS; break;
674 case Type::DoubleTyID: opCode = V9::FSUBD; break;
675 default: assert(0 && "Invalid type for SUB instruction"); break;
683 static inline MachineInstr*
684 CreateSubConstInstruction(const InstructionNode* instrNode)
686 MachineInstr* minstr = NULL;
688 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
689 assert(isa<Constant>(constOp));
691 // Cases worth optimizing are:
692 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
693 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
695 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
696 double dval = FPC->getValue();
698 minstr = CreateMovFloatInstruction(instrNode,
699 instrNode->getInstruction()->getType());
706 static inline MachineOpCode
707 ChooseFcmpInstruction(const InstructionNode* instrNode)
709 MachineOpCode opCode = V9::INVALID_OPCODE;
711 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
712 switch(operand->getType()->getPrimitiveID()) {
713 case Type::FloatTyID: opCode = V9::FCMPS; break;
714 case Type::DoubleTyID: opCode = V9::FCMPD; break;
715 default: assert(0 && "Invalid type for FCMP instruction"); break;
722 // Assumes that leftArg and rightArg are both cast instructions.
725 BothFloatToDouble(const InstructionNode* instrNode)
727 InstrTreeNode* leftArg = instrNode->leftChild();
728 InstrTreeNode* rightArg = instrNode->rightChild();
729 InstrTreeNode* leftArgArg = leftArg->leftChild();
730 InstrTreeNode* rightArgArg = rightArg->leftChild();
731 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
733 // Check if both arguments are floats cast to double
734 return (leftArg->getValue()->getType() == Type::DoubleTy &&
735 leftArgArg->getValue()->getType() == Type::FloatTy &&
736 rightArgArg->getValue()->getType() == Type::FloatTy);
740 static inline MachineOpCode
741 ChooseMulInstructionByType(const Type* resultType)
743 MachineOpCode opCode = V9::INVALID_OPCODE;
745 if (resultType->isInteger())
748 switch(resultType->getPrimitiveID())
750 case Type::FloatTyID: opCode = V9::FMULS; break;
751 case Type::DoubleTyID: opCode = V9::FMULD; break;
752 default: assert(0 && "Invalid type for MUL instruction"); break;
760 static inline MachineInstr*
761 CreateIntNegInstruction(const TargetMachine& target,
764 return BuildMI(V9::SUBr, 3).addMReg(target.getRegInfo().getZeroRegNum())
765 .addReg(vreg).addRegDef(vreg);
769 // Create instruction sequence for any shift operation.
770 // SLL or SLLX on an operand smaller than the integer reg. size (64bits)
771 // requires a second instruction for explicit sign-extension.
772 // Note that we only have to worry about a sign-bit appearing in the
773 // most significant bit of the operand after shifting (e.g., bit 32 of
774 // Int or bit 16 of Short), so we do not have to worry about results
775 // that are as large as a normal integer register.
778 CreateShiftInstructions(const TargetMachine& target,
780 MachineOpCode shiftOpCode,
782 Value* optArgVal2, /* Use optArgVal2 if not NULL */
783 unsigned optShiftNum, /* else use optShiftNum */
784 Instruction* destVal,
785 std::vector<MachineInstr*>& mvec,
786 MachineCodeForInstruction& mcfi)
788 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
789 "Large shift sizes unexpected, but can be handled below: "
790 "You need to check whether or not it fits in immed field below");
792 // If this is a logical left shift of a type smaller than the standard
793 // integer reg. size, we have to extend the sign-bit into upper bits
794 // of dest, so we need to put the result of the SLL into a temporary.
796 Value* shiftDest = destVal;
797 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
799 if ((shiftOpCode == V9::SLLr5 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
800 // put SLL result into a temporary
801 shiftDest = new TmpInstruction(mcfi, argVal1, optArgVal2, "sllTmp");
804 MachineInstr* M = (optArgVal2 != NULL)
805 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
806 .addReg(shiftDest, MOTy::Def)
807 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
808 .addReg(shiftDest, MOTy::Def);
811 if (shiftDest != destVal) {
812 // extend the sign-bit of the result into all upper bits of dest
813 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
814 target.getInstrInfo().
815 CreateSignExtensionInstructions(target, F, shiftDest, destVal,
816 8*opSize, mvec, mcfi);
821 // Does not create any instructions if we cannot exploit constant to
822 // create a cheaper instruction.
823 // This returns the approximate cost of the instructions generated,
824 // which is used to pick the cheapest when both operands are constant.
826 CreateMulConstInstruction(const TargetMachine &target, Function* F,
827 Value* lval, Value* rval, Instruction* destVal,
828 std::vector<MachineInstr*>& mvec,
829 MachineCodeForInstruction& mcfi)
831 /* Use max. multiply cost, viz., cost of MULX */
832 unsigned cost = target.getInstrInfo().minLatency(V9::MULXr);
833 unsigned firstNewInstr = mvec.size();
835 Value* constOp = rval;
836 if (! isa<Constant>(constOp))
839 // Cases worth optimizing are:
840 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
841 // (2) Multiply by 2^x for integer types: replace with Shift
843 const Type* resultType = destVal->getType();
845 if (resultType->isInteger() || isa<PointerType>(resultType)) {
847 int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
848 constOp, constOp->getType(), isValidConst);
851 bool needNeg = false;
857 if (C == 0 || C == 1) {
858 cost = target.getInstrInfo().minLatency(V9::ADDr);
859 unsigned Zero = target.getRegInfo().getZeroRegNum();
862 M =BuildMI(V9::ADDr,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
864 M = BuildMI(V9::ADDr,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
866 } else if (isPowerOf2(C, pow)) {
867 unsigned opSize = target.getTargetData().getTypeSize(resultType);
868 MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
869 CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
870 destVal, mvec, mcfi);
873 if (mvec.size() > 0 && needNeg) {
874 // insert <reg = SUB 0, reg> after the instr to flip the sign
875 MachineInstr* M = CreateIntNegInstruction(target, destVal);
880 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
881 double dval = FPC->getValue();
882 if (fabs(dval) == 1) {
883 MachineOpCode opCode = (dval < 0)
884 ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
885 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
886 mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
891 if (firstNewInstr < mvec.size()) {
893 for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
894 cost += target.getInstrInfo().minLatency(mvec[i]->getOpCode());
901 // Does not create any instructions if we cannot exploit constant to
902 // create a cheaper instruction.
905 CreateCheapestMulConstInstruction(const TargetMachine &target,
907 Value* lval, Value* rval,
908 Instruction* destVal,
909 std::vector<MachineInstr*>& mvec,
910 MachineCodeForInstruction& mcfi)
913 if (isa<Constant>(lval) && isa<Constant>(rval)) {
914 // both operands are constant: evaluate and "set" in dest
915 Constant* P = ConstantFoldBinaryInstruction(Instruction::Mul,
916 cast<Constant>(lval),
917 cast<Constant>(rval));
918 target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi);
920 else if (isa<Constant>(rval)) // rval is constant, but not lval
921 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
922 else if (isa<Constant>(lval)) // lval is constant, but not rval
923 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
925 // else neither is constant
929 // Return NULL if we cannot exploit constant to create a cheaper instruction
931 CreateMulInstruction(const TargetMachine &target, Function* F,
932 Value* lval, Value* rval, Instruction* destVal,
933 std::vector<MachineInstr*>& mvec,
934 MachineCodeForInstruction& mcfi,
935 MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE)
937 unsigned L = mvec.size();
938 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
939 if (mvec.size() == L) {
940 // no instructions were added so create MUL reg, reg, reg.
941 // Use FSMULD if both operands are actually floats cast to doubles.
942 // Otherwise, use the default opcode for the appropriate type.
943 MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE)
945 : ChooseMulInstructionByType(destVal->getType()));
946 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
947 .addRegDef(destVal));
952 // Generate a divide instruction for Div or Rem.
953 // For Rem, this assumes that the operand type will be signed if the result
954 // type is signed. This is correct because they must have the same sign.
956 static inline MachineOpCode
957 ChooseDivInstruction(TargetMachine &target,
958 const InstructionNode* instrNode)
960 MachineOpCode opCode = V9::INVALID_OPCODE;
962 const Type* resultType = instrNode->getInstruction()->getType();
964 if (resultType->isInteger())
965 opCode = resultType->isSigned()? V9::SDIVXr : V9::UDIVXr;
967 switch(resultType->getPrimitiveID())
969 case Type::FloatTyID: opCode = V9::FDIVS; break;
970 case Type::DoubleTyID: opCode = V9::FDIVD; break;
971 default: assert(0 && "Invalid type for DIV instruction"); break;
978 // Return if we cannot exploit constant to create a cheaper instruction
980 CreateDivConstInstruction(TargetMachine &target,
981 const InstructionNode* instrNode,
982 std::vector<MachineInstr*>& mvec)
984 Value* LHS = instrNode->leftChild()->getValue();
985 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
986 if (!isa<Constant>(constOp))
989 Instruction* destVal = instrNode->getInstruction();
990 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
992 // Cases worth optimizing are:
993 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
994 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
996 const Type* resultType = instrNode->getInstruction()->getType();
998 if (resultType->isInteger()) {
1001 int64_t C = (int64_t) target.getInstrInfo().ConvertConstantToIntType(target,
1002 constOp, constOp->getType(), isValidConst);
1004 bool needNeg = false;
1011 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
1012 .addRegDef(destVal));
1013 } else if (isPowerOf2(C, pow)) {
1015 Value* shiftOperand;
1016 unsigned opSize = target.getTargetData().getTypeSize(resultType);
1018 if (resultType->isSigned()) {
1019 // For N / 2^k, if the operand N is negative,
1020 // we need to add (2^k - 1) before right-shifting by k, i.e.,
1022 // (N / 2^k) = N >> k, if N >= 0;
1023 // (N + 2^k - 1) >> k, if N < 0
1025 // If N is <= 32 bits, use:
1026 // sra N, 31, t1 // t1 = ~0, if N < 0, 0 else
1027 // srl t1, 32-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
1028 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
1029 // sra t3, k, result // result = N / 2^k
1031 // If N is 64 bits, use:
1032 // srax N, k-1, t1 // t1 = sign bit in high k positions
1033 // srlx t1, 64-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
1034 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
1035 // sra t3, k, result // result = N / 2^k
1037 TmpInstruction *sraTmp, *srlTmp, *addTmp;
1038 MachineCodeForInstruction& mcfi
1039 = MachineCodeForInstruction::get(destVal);
1040 sraTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
1041 srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getPlus2km1");
1042 addTmp = new TmpInstruction(mcfi, resultType, LHS, srlTmp,"incIfNeg");
1044 // Create the SRA or SRAX instruction to get the sign bit
1045 mvec.push_back(BuildMI((opSize > 4)? V9::SRAXi6 : V9::SRAi5, 3)
1047 .addSImm((resultType==Type::LongTy)? pow-1 : 31)
1048 .addRegDef(sraTmp));
1050 // Create the SRL or SRLX instruction to get the sign bit
1051 mvec.push_back(BuildMI((opSize > 4)? V9::SRLXi6 : V9::SRLi5, 3)
1053 .addSImm((resultType==Type::LongTy)? 64-pow : 32-pow)
1054 .addRegDef(srlTmp));
1056 // Create the ADD instruction to add 2^pow-1 for negative values
1057 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
1058 .addRegDef(addTmp));
1060 // Get the shift operand and "right-shift" opcode to do the divide
1061 shiftOperand = addTmp;
1062 opCode = (opSize > 4)? V9::SRAXi6 : V9::SRAi5;
1064 // Get the shift operand and "right-shift" opcode to do the divide
1066 opCode = (opSize > 4)? V9::SRLXi6 : V9::SRLi5;
1069 // Now do the actual shift!
1070 mvec.push_back(BuildMI(opCode, 3).addReg(shiftOperand).addZImm(pow)
1071 .addRegDef(destVal));
1074 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
1075 // insert <reg = SUB 0, reg> after the instr to flip the sign
1076 mvec.push_back(CreateIntNegInstruction(target, destVal));
1080 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
1081 double dval = FPC->getValue();
1082 if (fabs(dval) == 1) {
1084 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
1085 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
1087 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(destVal));
1095 CreateCodeForVariableSizeAlloca(const TargetMachine& target,
1096 Instruction* result,
1098 Value* numElementsVal,
1099 std::vector<MachineInstr*>& getMvec)
1101 Value* totalSizeVal;
1103 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
1104 Function *F = result->getParent()->getParent();
1106 // Enforce the alignment constraints on the stack pointer at
1107 // compile time if the total size is a known constant.
1108 if (isa<Constant>(numElementsVal)) {
1110 int64_t numElem = (int64_t) target.getInstrInfo().
1111 ConvertConstantToIntType(target, numElementsVal,
1112 numElementsVal->getType(), isValid);
1113 assert(isValid && "Unexpectedly large array dimension in alloca!");
1114 int64_t total = numElem * tsize;
1115 if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
1116 total += target.getFrameInfo().getStackFrameSizeAlignment() - extra;
1117 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
1119 // The size is not a constant. Generate code to compute it and
1120 // code to pad the size for stack alignment.
1121 // Create a Value to hold the (constant) element size
1122 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
1124 // Create temporary values to hold the result of MUL, SLL, SRL
1125 // To pad `size' to next smallest multiple of 16:
1126 // size = (size + 15) & (-16 = 0xfffffffffffffff0)
1128 TmpInstruction* tmpProd = new TmpInstruction(mcfi,numElementsVal, tsizeVal);
1129 TmpInstruction* tmpAdd15= new TmpInstruction(mcfi,numElementsVal, tmpProd);
1130 TmpInstruction* tmpAndf0= new TmpInstruction(mcfi,numElementsVal, tmpAdd15);
1132 // Instruction 1: mul numElements, typeSize -> tmpProd
1133 // This will optimize the MUL as far as possible.
1134 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec,
1135 mcfi, INVALID_MACHINE_OPCODE);
1137 // Instruction 2: andn tmpProd, 0x0f -> tmpAndn
1138 getMvec.push_back(BuildMI(V9::ADDi, 3).addReg(tmpProd).addSImm(15)
1139 .addReg(tmpAdd15, MOTy::Def));
1141 // Instruction 3: add tmpAndn, 0x10 -> tmpAdd16
1142 getMvec.push_back(BuildMI(V9::ANDi, 3).addReg(tmpAdd15).addSImm(-16)
1143 .addReg(tmpAndf0, MOTy::Def));
1145 totalSizeVal = tmpAndf0;
1148 // Get the constant offset from SP for dynamically allocated storage
1149 // and create a temporary Value to hold it.
1150 MachineFunction& mcInfo = MachineFunction::get(F);
1152 ConstantSInt* dynamicAreaOffset =
1153 ConstantSInt::get(Type::IntTy,
1154 target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp));
1155 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
1157 unsigned SPReg = target.getRegInfo().getStackPointer();
1159 // Instruction 2: sub %sp, totalSizeVal -> %sp
1160 getMvec.push_back(BuildMI(V9::SUBr, 3).addMReg(SPReg).addReg(totalSizeVal)
1161 .addMReg(SPReg,MOTy::Def));
1163 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
1164 getMvec.push_back(BuildMI(V9::ADDr,3).addMReg(SPReg).addReg(dynamicAreaOffset)
1165 .addRegDef(result));
1170 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
1171 Instruction* result,
1173 unsigned numElements,
1174 std::vector<MachineInstr*>& getMvec)
1176 assert(tsize > 0 && "Illegal (zero) type size for alloca");
1177 assert(result && result->getParent() &&
1178 "Result value is not part of a function?");
1179 Function *F = result->getParent()->getParent();
1180 MachineFunction &mcInfo = MachineFunction::get(F);
1182 // Put the variable in the dynamically sized area of the frame if either:
1183 // (a) The offset is too large to use as an immediate in load/stores
1184 // (check LDX because all load/stores have the same-size immed. field).
1185 // (b) The object is "large", so it could cause many other locals,
1186 // spills, and temporaries to have large offsets.
1187 // NOTE: We use LARGE = 8 * argSlotSize = 64 bytes.
1188 // You've gotta love having only 13 bits for constant offset values :-|.
1190 unsigned paddedSize;
1191 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
1193 tsize * numElements);
1195 if (((int)paddedSize) > 8 * target.getFrameInfo().getSizeOfEachArgOnStack() ||
1196 ! target.getInstrInfo().constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
1197 CreateCodeForVariableSizeAlloca(target, result, tsize,
1198 ConstantSInt::get(Type::IntTy,numElements),
1203 // else offset fits in immediate field so go ahead and allocate it.
1204 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
1206 // Create a temporary Value to hold the constant offset.
1207 // This is needed because it may not fit in the immediate field.
1208 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
1210 // Instruction 1: add %fp, offsetFromFP -> result
1211 unsigned FPReg = target.getRegInfo().getFramePointer();
1212 getMvec.push_back(BuildMI(V9::ADDr, 3).addMReg(FPReg).addReg(offsetVal)
1213 .addRegDef(result));
1217 //------------------------------------------------------------------------
1218 // Function SetOperandsForMemInstr
1220 // Choose addressing mode for the given load or store instruction.
1221 // Use [reg+reg] if it is an indexed reference, and the index offset is
1222 // not a constant or if it cannot fit in the offset field.
1223 // Use [reg+offset] in all other cases.
1225 // This assumes that all array refs are "lowered" to one of these forms:
1226 // %x = load (subarray*) ptr, constant ; single constant offset
1227 // %x = load (subarray*) ptr, offsetVal ; single non-constant offset
1228 // Generally, this should happen via strength reduction + LICM.
1229 // Also, strength reduction should take care of using the same register for
1230 // the loop index variable and an array index, when that is profitable.
1231 //------------------------------------------------------------------------
1234 SetOperandsForMemInstr(unsigned Opcode,
1235 std::vector<MachineInstr*>& mvec,
1236 InstructionNode* vmInstrNode,
1237 const TargetMachine& target)
1239 Instruction* memInst = vmInstrNode->getInstruction();
1240 // Index vector, ptr value, and flag if all indices are const.
1241 std::vector<Value*> idxVec;
1242 bool allConstantIndices;
1243 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
1245 // Now create the appropriate operands for the machine instruction.
1246 // First, initialize so we default to storing the offset in a register.
1247 int64_t smallConstOffset = 0;
1248 Value* valueForRegOffset = NULL;
1249 MachineOperand::MachineOperandType offsetOpType =
1250 MachineOperand::MO_VirtualRegister;
1252 // Check if there is an index vector and if so, compute the
1253 // right offset for structures and for arrays
1255 if (!idxVec.empty()) {
1256 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
1258 // If all indices are constant, compute the combined offset directly.
1259 if (allConstantIndices) {
1260 // Compute the offset value using the index vector. Create a
1261 // virtual reg. for it since it may not fit in the immed field.
1262 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
1263 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
1265 // There is at least one non-constant offset. Therefore, this must
1266 // be an array ref, and must have been lowered to a single non-zero
1267 // offset. (An extra leading zero offset, if any, can be ignored.)
1268 // Generate code sequence to compute address from index.
1270 bool firstIdxIsZero = IsZero(idxVec[0]);
1271 assert(idxVec.size() == 1U + firstIdxIsZero
1272 && "Array refs must be lowered before Instruction Selection");
1274 Value* idxVal = idxVec[firstIdxIsZero];
1276 std::vector<MachineInstr*> mulVec;
1278 new TmpInstruction(MachineCodeForInstruction::get(memInst),
1279 Type::ULongTy, memInst);
1281 // Get the array type indexed by idxVal, and compute its element size.
1282 // The call to getTypeSize() will fail if size is not constant.
1283 const Type* vecType = (firstIdxIsZero
1284 ? GetElementPtrInst::getIndexedType(ptrType,
1285 std::vector<Value*>(1U, idxVec[0]),
1286 /*AllowCompositeLeaf*/ true)
1288 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
1289 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
1290 target.getTargetData().getTypeSize(eltType));
1292 // CreateMulInstruction() folds constants intelligently enough.
1293 CreateMulInstruction(target, memInst->getParent()->getParent(),
1294 idxVal, /* lval, not likely to be const*/
1295 eltSizeVal, /* rval, likely to be constant */
1297 mulVec, MachineCodeForInstruction::get(memInst),
1298 INVALID_MACHINE_OPCODE);
1300 assert(mulVec.size() > 0 && "No multiply code created?");
1301 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
1303 valueForRegOffset = addr;
1306 offsetOpType = MachineOperand::MO_SignExtendedImmed;
1307 smallConstOffset = 0;
1311 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
1312 // For LOAD or GET_ELEMENT_PTR,
1313 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
1315 unsigned offsetOpNum, ptrOpNum;
1317 if (memInst->getOpcode() == Instruction::Store) {
1318 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1319 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1320 .addReg(ptrVal).addReg(valueForRegOffset);
1322 Opcode = convertOpcodeFromRegToImm(Opcode);
1323 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1324 .addReg(ptrVal).addSImm(smallConstOffset);
1327 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1328 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
1329 .addRegDef(memInst);
1331 Opcode = convertOpcodeFromRegToImm(Opcode);
1332 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
1333 .addRegDef(memInst);
1341 // Substitute operand `operandNum' of the instruction in node `treeNode'
1342 // in place of the use(s) of that instruction in node `parent'.
1343 // Check both explicit and implicit operands!
1344 // Also make sure to skip over a parent who:
1345 // (1) is a list node in the Burg tree, or
1346 // (2) itself had its results forwarded to its parent
1349 ForwardOperand(InstructionNode* treeNode,
1350 InstrTreeNode* parent,
1353 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
1355 Instruction* unusedOp = treeNode->getInstruction();
1356 Value* fwdOp = unusedOp->getOperand(operandNum);
1358 // The parent itself may be a list node, so find the real parent instruction
1359 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode)
1361 parent = parent->parent();
1362 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
1364 InstructionNode* parentInstrNode = (InstructionNode*) parent;
1366 Instruction* userInstr = parentInstrNode->getInstruction();
1367 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
1369 // The parent's mvec would be empty if it was itself forwarded.
1370 // Recursively call ForwardOperand in that case...
1372 if (mvec.size() == 0) {
1373 assert(parent->parent() != NULL &&
1374 "Parent could not have been forwarded, yet has no instructions?");
1375 ForwardOperand(treeNode, parent->parent(), operandNum);
1377 for (unsigned i=0, N=mvec.size(); i < N; i++) {
1378 MachineInstr* minstr = mvec[i];
1379 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) {
1380 const MachineOperand& mop = minstr->getOperand(i);
1381 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
1382 mop.getVRegValue() == unusedOp)
1384 minstr->SetMachineOperandVal(i, MachineOperand::MO_VirtualRegister,
1389 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
1390 if (minstr->getImplicitRef(i) == unusedOp)
1391 minstr->setImplicitRef(i, fwdOp);
1398 AllUsesAreBranches(const Instruction* setccI)
1400 for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end();
1402 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
1403 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
1408 // Generate code for any intrinsic that needs a special code sequence
1409 // instead of a regular call. If not that kind of intrinsic, do nothing.
1410 // Returns true if code was generated, otherwise false.
1412 bool CodeGenIntrinsic(LLVMIntrinsic::ID iid, CallInst &callInstr,
1413 TargetMachine &target,
1414 std::vector<MachineInstr*>& mvec)
1417 case LLVMIntrinsic::va_start: {
1418 // FIXME: this needs to be updated!
1421 // Get the address of the first vararg value on stack and copy it to
1422 // the argument of va_start(va_list* ap).
1424 Function* func = cast<Function>(callInstr.getParent()->getParent());
1425 int numFixedArgs = func->getFunctionType()->getNumParams();
1426 int fpReg = target.getFrameInfo().getIncomingArgBaseRegNum();
1427 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
1428 int firstVarArgOff = numFixedArgs * argSize + target.getFrameInfo().
1429 getFirstIncomingArgOffset(MachineFunction::get(func), ignore);
1430 mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
1431 addRegDef(callInstr.getOperand(1)));
1435 case LLVMIntrinsic::va_end:
1436 return true; // no-op on Sparc
1438 case LLVMIntrinsic::va_copy:
1439 // FIXME: this needs to be updated!
1442 // Simple copy of current va_list (arg2) to new va_list (arg1)
1443 mvec.push_back(BuildMI(V9::ORr, 3).
1444 addMReg(target.getRegInfo().getZeroRegNum()).
1445 addReg(callInstr.getOperand(2)).
1446 addReg(callInstr.getOperand(1)));
1449 case LLVMIntrinsic::sigsetjmp:
1450 case LLVMIntrinsic::setjmp: {
1451 // act as if we return 0
1452 unsigned g0 = target.getRegInfo().getZeroRegNum();
1453 mvec.push_back(BuildMI(V9::ORr,3).addMReg(g0).addMReg(g0)
1454 .addReg(&callInstr, MOTy::Def));
1458 case LLVMIntrinsic::siglongjmp:
1459 case LLVMIntrinsic::longjmp: {
1461 Module* M = callInstr.getParent()->getParent()->getParent();
1462 const FunctionType *voidvoidFuncTy =
1463 FunctionType::get(Type::VoidTy, std::vector<const Type*>(), false);
1464 Function *F = M->getOrInsertFunction("abort", voidvoidFuncTy);
1465 assert(F && "Unable to get or create `abort' function declaration");
1467 // Create hidden virtual register for return address with type void*
1468 TmpInstruction* retAddrReg =
1469 new TmpInstruction(MachineCodeForInstruction::get(&callInstr),
1470 PointerType::get(Type::VoidTy), &callInstr);
1472 // Use a descriptor to pass information about call arguments
1473 // to the register allocator. This descriptor will be "owned"
1474 // and freed automatically when the MachineCodeForInstruction
1475 // object for the callInstr goes away.
1476 CallArgsDescriptor* argDesc =
1477 new CallArgsDescriptor(&callInstr, retAddrReg, false, false);
1479 MachineInstr* callMI = BuildMI(V9::CALL, 1).addPCDisp(F);
1480 callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
1482 mvec.push_back(callMI);
1483 mvec.push_back(BuildMI(V9::NOP, 0));
1492 //******************* Externally Visible Functions *************************/
1494 //------------------------------------------------------------------------
1495 // External Function: ThisIsAChainRule
1498 // Check if a given BURG rule is a chain rule.
1499 //------------------------------------------------------------------------
1502 ThisIsAChainRule(int eruleno)
1506 case 111: // stmt: reg
1530 return false; break;
1535 //------------------------------------------------------------------------
1536 // External Function: GetInstructionsByRule
1539 // Choose machine instructions for the SPARC according to the
1540 // patterns chosen by the BURG-generated parser.
1541 //------------------------------------------------------------------------
1544 GetInstructionsByRule(InstructionNode* subtreeRoot,
1547 TargetMachine &target,
1548 std::vector<MachineInstr*>& mvec)
1550 bool checkCast = false; // initialize here to use fall-through
1551 bool maskUnsignedResult = false;
1553 int forwardOperandNum = -1;
1554 unsigned allocaSize = 0;
1555 MachineInstr* M, *M2;
1557 bool foldCase = false;
1561 // If the code for this instruction was folded into the parent (user),
1563 if (subtreeRoot->isFoldedIntoParent())
1567 // Let's check for chain rules outside the switch so that we don't have
1568 // to duplicate the list of chain rule production numbers here again
1570 if (ThisIsAChainRule(ruleForNode))
1572 // Chain rules have a single nonterminal on the RHS.
1573 // Get the rule that matches the RHS non-terminal and use that instead.
1575 assert(nts[0] && ! nts[1]
1576 && "A chain rule should have only one RHS non-terminal!");
1577 nextRule = burm_rule(subtreeRoot->state, nts[0]);
1578 nts = burm_nts[nextRule];
1579 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
1583 switch(ruleForNode) {
1584 case 1: // stmt: Ret
1585 case 2: // stmt: RetValue(reg)
1586 { // NOTE: Prepass of register allocation is responsible
1587 // for moving return value to appropriate register.
1588 // Copy the return value to the required return register.
1589 // Mark the return Value as an implicit ref of the RET instr..
1590 // Mark the return-address register as a hidden virtual reg.
1591 // Finally put a NOP in the delay slot.
1592 ReturnInst *returnInstr=cast<ReturnInst>(subtreeRoot->getInstruction());
1593 Value* retVal = returnInstr->getReturnValue();
1594 MachineCodeForInstruction& mcfi =
1595 MachineCodeForInstruction::get(returnInstr);
1597 // Create a hidden virtual reg to represent the return address register
1598 // used by the machine instruction but not represented in LLVM.
1600 Instruction* returnAddrTmp = new TmpInstruction(mcfi, returnInstr);
1602 MachineInstr* retMI =
1603 BuildMI(V9::JMPLRETi, 3).addReg(returnAddrTmp).addSImm(8)
1604 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
1606 // If there is a value to return, we need to:
1607 // (a) Sign-extend the value if it is smaller than 8 bytes (reg size)
1608 // (b) Insert a copy to copy the return value to the appropriate reg.
1609 // -- For FP values, create a FMOVS or FMOVD instruction
1610 // -- For non-FP values, create an add-with-0 instruction
1612 if (retVal != NULL) {
1613 const UltraSparcRegInfo& regInfo =
1614 (UltraSparcRegInfo&) target.getRegInfo();
1615 const Type* retType = retVal->getType();
1616 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
1617 unsigned retRegNum = (retType->isFloatingPoint()
1618 ? (unsigned) SparcFloatRegClass::f0
1619 : (unsigned) SparcIntRegClass::i0);
1620 retRegNum = regInfo.getUnifiedRegNum(regClassID, retRegNum);
1622 // () Insert sign-extension instructions for small signed values.
1624 Value* retValToUse = retVal;
1625 if (retType->isIntegral() && retType->isSigned()) {
1626 unsigned retSize = target.getTargetData().getTypeSize(retType);
1628 // create a temporary virtual reg. to hold the sign-extension
1629 retValToUse = new TmpInstruction(mcfi, retVal);
1631 // sign-extend retVal and put the result in the temporary reg.
1632 target.getInstrInfo().CreateSignExtensionInstructions
1633 (target, returnInstr->getParent()->getParent(),
1634 retVal, retValToUse, 8*retSize, mvec, mcfi);
1638 // (b) Now, insert a copy to to the appropriate register:
1639 // -- For FP values, create a FMOVS or FMOVD instruction
1640 // -- For non-FP values, create an add-with-0 instruction
1642 // First, create a virtual register to represent the register and
1643 // mark this vreg as being an implicit operand of the ret MI.
1644 TmpInstruction* retVReg =
1645 new TmpInstruction(mcfi, retValToUse, NULL, "argReg");
1647 retMI->addImplicitRef(retVReg);
1649 if (retType->isFloatingPoint())
1650 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
1651 .addReg(retValToUse).addReg(retVReg, MOTy::Def));
1653 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
1654 .addReg(retValToUse).addSImm((int64_t) 0)
1655 .addReg(retVReg, MOTy::Def));
1657 // Mark the operand with the register it should be assigned
1658 M->SetRegForOperand(M->getNumOperands()-1, retRegNum);
1659 retMI->SetRegForImplicitRef(retMI->getNumImplicitRefs()-1, retRegNum);
1664 // Now insert the RET instruction and a NOP for the delay slot
1665 mvec.push_back(retMI);
1666 mvec.push_back(BuildMI(V9::NOP, 0));
1671 case 3: // stmt: Store(reg,reg)
1672 case 4: // stmt: Store(reg,ptrreg)
1673 SetOperandsForMemInstr(ChooseStoreInstruction(
1674 subtreeRoot->leftChild()->getValue()->getType()),
1675 mvec, subtreeRoot, target);
1678 case 5: // stmt: BrUncond
1680 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1681 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
1684 mvec.push_back(BuildMI(V9::NOP, 0));
1688 case 206: // stmt: BrCond(setCCconst)
1689 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
1690 // If the constant is ZERO, we can use the branch-on-integer-register
1691 // instructions and avoid the SUBcc instruction entirely.
1692 // Otherwise this is just the same as case 5, so just fall through.
1694 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
1696 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
1697 Constant *constVal = cast<Constant>(constNode->getValue());
1700 if ((constVal->getType()->isInteger()
1701 || isa<PointerType>(constVal->getType()))
1702 && target.getInstrInfo().ConvertConstantToIntType(target,
1703 constVal, constVal->getType(), isValidConst) == 0
1706 // That constant is a zero after all...
1707 // Use the left child of setCC as the first argument!
1708 // Mark the setCC node so that no code is generated for it.
1709 InstructionNode* setCCNode = (InstructionNode*)
1710 subtreeRoot->leftChild();
1711 assert(setCCNode->getOpLabel() == SetCCOp);
1712 setCCNode->markFoldedIntoParent();
1714 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
1716 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
1717 .addReg(setCCNode->leftChild()->getValue())
1718 .addPCDisp(brInst->getSuccessor(0));
1722 mvec.push_back(BuildMI(V9::NOP, 0));
1725 mvec.push_back(BuildMI(V9::BA, 1)
1726 .addPCDisp(brInst->getSuccessor(1)));
1729 mvec.push_back(BuildMI(V9::NOP, 0));
1732 // ELSE FALL THROUGH
1735 case 6: // stmt: BrCond(setCC)
1736 { // bool => boolean was computed with SetCC.
1737 // The branch to use depends on whether it is FP, signed, or unsigned.
1738 // If it is an integer CC, we also need to find the unique
1739 // TmpInstruction representing that CC.
1741 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
1742 const Type* setCCType;
1743 unsigned Opcode = ChooseBccInstruction(subtreeRoot, setCCType);
1744 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
1745 brInst->getParent()->getParent(),
1747 MachineCodeForInstruction::get(brInst));
1748 M = BuildMI(Opcode, 2).addCCReg(ccValue)
1749 .addPCDisp(brInst->getSuccessor(0));
1753 mvec.push_back(BuildMI(V9::NOP, 0));
1756 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
1759 mvec.push_back(BuildMI(V9::NOP, 0));
1763 case 208: // stmt: BrCond(boolconst)
1765 // boolconst => boolean is a constant; use BA to first or second label
1766 Constant* constVal =
1767 cast<Constant>(subtreeRoot->leftChild()->getValue());
1768 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
1770 M = BuildMI(V9::BA, 1).addPCDisp(
1771 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
1775 mvec.push_back(BuildMI(V9::NOP, 0));
1779 case 8: // stmt: BrCond(boolreg)
1780 { // boolreg => boolean is recorded in an integer register.
1781 // Use branch-on-integer-register instruction.
1783 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1784 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
1785 .addPCDisp(BI->getSuccessor(0));
1789 mvec.push_back(BuildMI(V9::NOP, 0));
1792 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
1795 mvec.push_back(BuildMI(V9::NOP, 0));
1799 case 9: // stmt: Switch(reg)
1800 assert(0 && "*** SWITCH instruction is not implemented yet.");
1803 case 10: // reg: VRegList(reg, reg)
1804 assert(0 && "VRegList should never be the topmost non-chain rule");
1807 case 21: // bool: Not(bool,reg): Compute with a conditional-move-on-reg
1808 { // First find the unary operand. It may be left or right, usually right.
1809 Instruction* notI = subtreeRoot->getInstruction();
1810 Value* notArg = BinaryOperator::getNotArgument(
1811 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1812 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1814 // Unconditionally set register to 0
1815 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(notI));
1817 // Now conditionally move 1 into the register.
1818 // Mark the register as a use (as well as a def) because the old
1819 // value will be retained if the condition is false.
1820 mvec.push_back(BuildMI(V9::MOVRZi, 3).addReg(notArg).addZImm(1)
1821 .addReg(notI, MOTy::UseAndDef));
1826 case 421: // reg: BNot(reg,reg): Compute as reg = reg XOR-NOT 0
1827 { // First find the unary operand. It may be left or right, usually right.
1828 Value* notArg = BinaryOperator::getNotArgument(
1829 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1830 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1831 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(notArg).addMReg(ZeroReg)
1832 .addRegDef(subtreeRoot->getValue()));
1836 case 322: // reg: Not(tobool, reg):
1837 // Fold CAST-TO-BOOL with NOT by inverting the sense of cast-to-bool
1839 // Just fall through!
1841 case 22: // reg: ToBoolTy(reg):
1843 Instruction* castI = subtreeRoot->getInstruction();
1844 Value* opVal = subtreeRoot->leftChild()->getValue();
1845 assert(opVal->getType()->isIntegral() ||
1846 isa<PointerType>(opVal->getType()));
1848 // Unconditionally set register to 0
1849 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(castI));
1851 // Now conditionally move 1 into the register.
1852 // Mark the register as a use (as well as a def) because the old
1853 // value will be retained if the condition is false.
1854 MachineOpCode opCode = foldCase? V9::MOVRZi : V9::MOVRNZi;
1855 mvec.push_back(BuildMI(opCode, 3).addReg(opVal).addZImm(1)
1856 .addReg(castI, MOTy::UseAndDef));
1861 case 23: // reg: ToUByteTy(reg)
1862 case 24: // reg: ToSByteTy(reg)
1863 case 25: // reg: ToUShortTy(reg)
1864 case 26: // reg: ToShortTy(reg)
1865 case 27: // reg: ToUIntTy(reg)
1866 case 28: // reg: ToIntTy(reg)
1867 case 29: // reg: ToULongTy(reg)
1868 case 30: // reg: ToLongTy(reg)
1870 //======================================================================
1871 // Rules for integer conversions:
1874 // From ISO 1998 C++ Standard, Sec. 4.7:
1876 // 2. If the destination type is unsigned, the resulting value is
1877 // the least unsigned integer congruent to the source integer
1878 // (modulo 2n where n is the number of bits used to represent the
1879 // unsigned type). [Note: In a two s complement representation,
1880 // this conversion is conceptual and there is no change in the
1881 // bit pattern (if there is no truncation). ]
1883 // 3. If the destination type is signed, the value is unchanged if
1884 // it can be represented in the destination type (and bitfield width);
1885 // otherwise, the value is implementation-defined.
1888 // Since we assume 2s complement representations, this implies:
1890 // -- If operand is smaller than destination, zero-extend or sign-extend
1891 // according to the signedness of the *operand*: source decides:
1892 // (1) If operand is signed, sign-extend it.
1893 // If dest is unsigned, zero-ext the result!
1894 // (2) If operand is unsigned, our current invariant is that
1895 // it's high bits are correct, so zero-extension is not needed.
1897 // -- If operand is same size as or larger than destination,
1898 // zero-extend or sign-extend according to the signedness of
1899 // the *destination*: destination decides:
1900 // (1) If destination is signed, sign-extend (truncating if needed)
1901 // This choice is implementation defined. We sign-extend the
1902 // operand, which matches both Sun's cc and gcc3.2.
1903 // (2) If destination is unsigned, zero-extend (truncating if needed)
1904 //======================================================================
1906 Instruction* destI = subtreeRoot->getInstruction();
1907 Function* currentFunc = destI->getParent()->getParent();
1908 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(destI);
1910 Value* opVal = subtreeRoot->leftChild()->getValue();
1911 const Type* opType = opVal->getType();
1912 const Type* destType = destI->getType();
1913 unsigned opSize = target.getTargetData().getTypeSize(opType);
1914 unsigned destSize = target.getTargetData().getTypeSize(destType);
1916 bool isIntegral = opType->isIntegral() || isa<PointerType>(opType);
1918 if (opType == Type::BoolTy ||
1919 opType == destType ||
1920 isIntegral && opSize == destSize && opSize == 8) {
1921 // nothing to do in all these cases
1922 forwardOperandNum = 0; // forward first operand to user
1924 } else if (opType->isFloatingPoint()) {
1926 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, mcfi);
1927 if (destI->getType()->isUnsigned() && destI->getType() !=Type::UIntTy)
1928 maskUnsignedResult = true; // not handled by fp->int code
1930 } else if (isIntegral) {
1932 bool opSigned = opType->isSigned();
1933 bool destSigned = destType->isSigned();
1934 unsigned extSourceInBits = 8 * std::min<unsigned>(opSize, destSize);
1936 assert(! (opSize == destSize && opSigned == destSigned) &&
1937 "How can different int types have same size and signedness?");
1939 bool signExtend = (opSize < destSize && opSigned ||
1940 opSize >= destSize && destSigned);
1942 bool signAndZeroExtend = (opSize < destSize && destSize < 8u &&
1943 opSigned && !destSigned);
1944 assert(!signAndZeroExtend || signExtend);
1946 bool zeroExtendOnly = opSize >= destSize && !destSigned;
1947 assert(!zeroExtendOnly || !signExtend);
1950 Value* signExtDest = (signAndZeroExtend
1951 ? new TmpInstruction(mcfi, destType, opVal)
1954 target.getInstrInfo().CreateSignExtensionInstructions
1955 (target, currentFunc,opVal,signExtDest,extSourceInBits,mvec,mcfi);
1957 if (signAndZeroExtend)
1958 target.getInstrInfo().CreateZeroExtensionInstructions
1959 (target, currentFunc, signExtDest, destI, 8*destSize, mvec, mcfi);
1961 else if (zeroExtendOnly) {
1962 target.getInstrInfo().CreateZeroExtensionInstructions
1963 (target, currentFunc, opVal, destI, extSourceInBits, mvec, mcfi);
1966 forwardOperandNum = 0; // forward first operand to user
1969 assert(0 && "Unrecognized operand type for convert-to-integer");
1974 case 31: // reg: ToFloatTy(reg):
1975 case 32: // reg: ToDoubleTy(reg):
1976 case 232: // reg: ToDoubleTy(Constant):
1978 // If this instruction has a parent (a user) in the tree
1979 // and the user is translated as an FsMULd instruction,
1980 // then the cast is unnecessary. So check that first.
1981 // In the future, we'll want to do the same for the FdMULq instruction,
1982 // so do the check here instead of only for ToFloatTy(reg).
1984 if (subtreeRoot->parent() != NULL) {
1985 const MachineCodeForInstruction& mcfi =
1986 MachineCodeForInstruction::get(
1987 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
1988 if (mcfi.size() == 0 || mcfi.front()->getOpCode() == V9::FSMULD)
1989 forwardOperandNum = 0; // forward first operand to user
1992 if (forwardOperandNum != 0) { // we do need the cast
1993 Value* leftVal = subtreeRoot->leftChild()->getValue();
1994 const Type* opType = leftVal->getType();
1995 MachineOpCode opCode=ChooseConvertToFloatInstr(target,
1996 subtreeRoot->getOpLabel(), opType);
1997 if (opCode == V9::NOP) { // no conversion needed
1998 forwardOperandNum = 0; // forward first operand to user
2000 // If the source operand is a non-FP type it must be
2001 // first copied from int to float register via memory!
2002 Instruction *dest = subtreeRoot->getInstruction();
2005 if (! opType->isFloatingPoint()) {
2006 // Create a temporary to represent the FP register
2007 // into which the integer will be copied via memory.
2008 // The type of this temporary will determine the FP
2009 // register used: single-prec for a 32-bit int or smaller,
2010 // double-prec for a 64-bit int.
2013 target.getTargetData().getTypeSize(leftVal->getType());
2014 Type* tmpTypeToUse =
2015 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
2016 MachineCodeForInstruction &destMCFI =
2017 MachineCodeForInstruction::get(dest);
2018 srcForCast = new TmpInstruction(destMCFI, tmpTypeToUse, dest);
2020 target.getInstrInfo().CreateCodeToCopyIntToFloat(target,
2021 dest->getParent()->getParent(),
2022 leftVal, cast<Instruction>(srcForCast),
2025 srcForCast = leftVal;
2027 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
2033 case 19: // reg: ToArrayTy(reg):
2034 case 20: // reg: ToPointerTy(reg):
2035 forwardOperandNum = 0; // forward first operand to user
2038 case 233: // reg: Add(reg, Constant)
2039 maskUnsignedResult = true;
2040 M = CreateAddConstInstruction(subtreeRoot);
2045 // ELSE FALL THROUGH
2047 case 33: // reg: Add(reg, reg)
2048 maskUnsignedResult = true;
2049 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
2052 case 234: // reg: Sub(reg, Constant)
2053 maskUnsignedResult = true;
2054 M = CreateSubConstInstruction(subtreeRoot);
2059 // ELSE FALL THROUGH
2061 case 34: // reg: Sub(reg, reg)
2062 maskUnsignedResult = true;
2063 Add3OperandInstr(ChooseSubInstructionByType(
2064 subtreeRoot->getInstruction()->getType()),
2068 case 135: // reg: Mul(todouble, todouble)
2072 case 35: // reg: Mul(reg, reg)
2074 maskUnsignedResult = true;
2075 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
2077 : INVALID_MACHINE_OPCODE);
2078 Instruction* mulInstr = subtreeRoot->getInstruction();
2079 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
2080 subtreeRoot->leftChild()->getValue(),
2081 subtreeRoot->rightChild()->getValue(),
2083 MachineCodeForInstruction::get(mulInstr),forceOp);
2086 case 335: // reg: Mul(todouble, todoubleConst)
2090 case 235: // reg: Mul(reg, Constant)
2092 maskUnsignedResult = true;
2093 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
2095 : INVALID_MACHINE_OPCODE);
2096 Instruction* mulInstr = subtreeRoot->getInstruction();
2097 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
2098 subtreeRoot->leftChild()->getValue(),
2099 subtreeRoot->rightChild()->getValue(),
2101 MachineCodeForInstruction::get(mulInstr),
2105 case 236: // reg: Div(reg, Constant)
2106 maskUnsignedResult = true;
2108 CreateDivConstInstruction(target, subtreeRoot, mvec);
2109 if (mvec.size() > L)
2111 // ELSE FALL THROUGH
2113 case 36: // reg: Div(reg, reg)
2115 maskUnsignedResult = true;
2117 // If either operand of divide is smaller than 64 bits, we have
2118 // to make sure the unused top bits are correct because they affect
2119 // the result. These bits are already correct for unsigned values.
2120 // They may be incorrect for signed values, so sign extend to fill in.
2121 Instruction* divI = subtreeRoot->getInstruction();
2122 Value* divOp1 = subtreeRoot->leftChild()->getValue();
2123 Value* divOp2 = subtreeRoot->rightChild()->getValue();
2124 Value* divOp1ToUse = divOp1;
2125 Value* divOp2ToUse = divOp2;
2126 if (divI->getType()->isSigned()) {
2127 unsigned opSize=target.getTargetData().getTypeSize(divI->getType());
2129 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(divI);
2130 divOp1ToUse = new TmpInstruction(mcfi, divOp1);
2131 divOp2ToUse = new TmpInstruction(mcfi, divOp2);
2132 target.getInstrInfo().
2133 CreateSignExtensionInstructions(target,
2134 divI->getParent()->getParent(),
2135 divOp1, divOp1ToUse,
2136 8*opSize, mvec, mcfi);
2137 target.getInstrInfo().
2138 CreateSignExtensionInstructions(target,
2139 divI->getParent()->getParent(),
2140 divOp2, divOp2ToUse,
2141 8*opSize, mvec, mcfi);
2145 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
2146 .addReg(divOp1ToUse)
2147 .addReg(divOp2ToUse)
2153 case 37: // reg: Rem(reg, reg)
2154 case 237: // reg: Rem(reg, Constant)
2156 maskUnsignedResult = true;
2158 Instruction* remI = subtreeRoot->getInstruction();
2159 Value* divOp1 = subtreeRoot->leftChild()->getValue();
2160 Value* divOp2 = subtreeRoot->rightChild()->getValue();
2162 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(remI);
2164 // If second operand of divide is smaller than 64 bits, we have
2165 // to make sure the unused top bits are correct because they affect
2166 // the result. These bits are already correct for unsigned values.
2167 // They may be incorrect for signed values, so sign extend to fill in.
2169 Value* divOpToUse = divOp2;
2170 if (divOp2->getType()->isSigned()) {
2171 unsigned opSize=target.getTargetData().getTypeSize(divOp2->getType());
2173 divOpToUse = new TmpInstruction(mcfi, divOp2);
2174 target.getInstrInfo().
2175 CreateSignExtensionInstructions(target,
2176 remI->getParent()->getParent(),
2178 8*opSize, mvec, mcfi);
2182 // Now compute: result = rem V1, V2 as:
2183 // result = V1 - (V1 / signExtend(V2)) * signExtend(V2)
2185 TmpInstruction* quot = new TmpInstruction(mcfi, divOp1, divOpToUse);
2186 TmpInstruction* prod = new TmpInstruction(mcfi, quot, divOpToUse);
2188 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
2189 .addReg(divOp1).addReg(divOpToUse).addRegDef(quot));
2191 mvec.push_back(BuildMI(ChooseMulInstructionByType(remI->getType()), 3)
2192 .addReg(quot).addReg(divOpToUse).addRegDef(prod));
2194 mvec.push_back(BuildMI(ChooseSubInstructionByType(remI->getType()), 3)
2195 .addReg(divOp1).addReg(prod).addRegDef(remI));
2200 case 38: // bool: And(bool, bool)
2201 case 138: // bool: And(bool, not)
2202 case 238: // bool: And(bool, boolconst)
2203 case 338: // reg : BAnd(reg, reg)
2204 case 538: // reg : BAnd(reg, Constant)
2205 Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
2208 case 438: // bool: BAnd(bool, bnot)
2209 { // Use the argument of NOT as the second argument!
2210 // Mark the NOT node so that no code is generated for it.
2211 // If the type is boolean, set 1 or 0 in the result register.
2212 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2213 Value* notArg = BinaryOperator::getNotArgument(
2214 cast<BinaryOperator>(notNode->getInstruction()));
2215 notNode->markFoldedIntoParent();
2216 Value *lhs = subtreeRoot->leftChild()->getValue();
2217 Value *dest = subtreeRoot->getValue();
2218 mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(lhs).addReg(notArg)
2219 .addReg(dest, MOTy::Def));
2221 if (notArg->getType() == Type::BoolTy)
2222 { // set 1 in result register if result of above is non-zero
2223 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2224 .addReg(dest, MOTy::UseAndDef));
2230 case 39: // bool: Or(bool, bool)
2231 case 139: // bool: Or(bool, not)
2232 case 239: // bool: Or(bool, boolconst)
2233 case 339: // reg : BOr(reg, reg)
2234 case 539: // reg : BOr(reg, Constant)
2235 Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
2238 case 439: // bool: BOr(bool, bnot)
2239 { // Use the argument of NOT as the second argument!
2240 // Mark the NOT node so that no code is generated for it.
2241 // If the type is boolean, set 1 or 0 in the result register.
2242 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2243 Value* notArg = BinaryOperator::getNotArgument(
2244 cast<BinaryOperator>(notNode->getInstruction()));
2245 notNode->markFoldedIntoParent();
2246 Value *lhs = subtreeRoot->leftChild()->getValue();
2247 Value *dest = subtreeRoot->getValue();
2249 mvec.push_back(BuildMI(V9::ORNr, 3).addReg(lhs).addReg(notArg)
2250 .addReg(dest, MOTy::Def));
2252 if (notArg->getType() == Type::BoolTy)
2253 { // set 1 in result register if result of above is non-zero
2254 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2255 .addReg(dest, MOTy::UseAndDef));
2261 case 40: // bool: Xor(bool, bool)
2262 case 140: // bool: Xor(bool, not)
2263 case 240: // bool: Xor(bool, boolconst)
2264 case 340: // reg : BXor(reg, reg)
2265 case 540: // reg : BXor(reg, Constant)
2266 Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
2269 case 440: // bool: BXor(bool, bnot)
2270 { // Use the argument of NOT as the second argument!
2271 // Mark the NOT node so that no code is generated for it.
2272 // If the type is boolean, set 1 or 0 in the result register.
2273 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2274 Value* notArg = BinaryOperator::getNotArgument(
2275 cast<BinaryOperator>(notNode->getInstruction()));
2276 notNode->markFoldedIntoParent();
2277 Value *lhs = subtreeRoot->leftChild()->getValue();
2278 Value *dest = subtreeRoot->getValue();
2279 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(lhs).addReg(notArg)
2280 .addReg(dest, MOTy::Def));
2282 if (notArg->getType() == Type::BoolTy)
2283 { // set 1 in result register if result of above is non-zero
2284 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
2285 .addReg(dest, MOTy::UseAndDef));
2290 case 41: // setCCconst: SetCC(reg, Constant)
2291 { // Comparison is with a constant:
2293 // If the bool result must be computed into a register (see below),
2294 // and the constant is int ZERO, we can use the MOVR[op] instructions
2295 // and avoid the SUBcc instruction entirely.
2296 // Otherwise this is just the same as case 42, so just fall through.
2298 // The result of the SetCC must be computed and stored in a register if
2299 // it is used outside the current basic block (so it must be computed
2300 // as a boolreg) or it is used by anything other than a branch.
2301 // We will use a conditional move to do this.
2303 Instruction* setCCInstr = subtreeRoot->getInstruction();
2304 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
2305 ! AllUsesAreBranches(setCCInstr));
2309 InstrTreeNode* constNode = subtreeRoot->rightChild();
2311 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
2312 Constant *constVal = cast<Constant>(constNode->getValue());
2315 if ((constVal->getType()->isInteger()
2316 || isa<PointerType>(constVal->getType()))
2317 && target.getInstrInfo().ConvertConstantToIntType(target,
2318 constVal, constVal->getType(), isValidConst) == 0
2321 // That constant is an integer zero after all...
2322 // Use a MOVR[op] to compute the boolean result
2323 // Unconditionally set register to 0
2324 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0)
2325 .addRegDef(setCCInstr));
2327 // Now conditionally move 1 into the register.
2328 // Mark the register as a use (as well as a def) because the old
2329 // value will be retained if the condition is false.
2330 MachineOpCode movOpCode = ChooseMovpregiForSetCC(subtreeRoot);
2331 mvec.push_back(BuildMI(movOpCode, 3)
2332 .addReg(subtreeRoot->leftChild()->getValue())
2333 .addZImm(1).addReg(setCCInstr, MOTy::UseAndDef));
2338 // ELSE FALL THROUGH
2341 case 42: // bool: SetCC(reg, reg):
2343 // This generates a SUBCC instruction, putting the difference in a
2344 // result reg. if needed, and/or setting a condition code if needed.
2346 Instruction* setCCInstr = subtreeRoot->getInstruction();
2347 Value* leftVal = subtreeRoot->leftChild()->getValue();
2348 Value* rightVal = subtreeRoot->rightChild()->getValue();
2349 const Type* opType = leftVal->getType();
2350 bool isFPCompare = opType->isFloatingPoint();
2352 // If the boolean result of the SetCC is used outside the current basic
2353 // block (so it must be computed as a boolreg) or is used by anything
2354 // other than a branch, the boolean must be computed and stored
2355 // in a result register. We will use a conditional move to do this.
2357 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
2358 ! AllUsesAreBranches(setCCInstr));
2360 // A TmpInstruction is created to represent the CC "result".
2361 // Unlike other instances of TmpInstruction, this one is used
2362 // by machine code of multiple LLVM instructions, viz.,
2363 // the SetCC and the branch. Make sure to get the same one!
2364 // Note that we do this even for FP CC registers even though they
2365 // are explicit operands, because the type of the operand
2366 // needs to be a floating point condition code, not an integer
2367 // condition code. Think of this as casting the bool result to
2368 // a FP condition code register.
2369 // Later, we mark the 4th operand as being a CC register, and as a def.
2371 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
2372 setCCInstr->getParent()->getParent(),
2374 MachineCodeForInstruction::get(setCCInstr));
2376 // If the operands are signed values smaller than 4 bytes, then they
2377 // must be sign-extended in order to do a valid 32-bit comparison
2378 // and get the right result in the 32-bit CC register (%icc).
2380 Value* leftOpToUse = leftVal;
2381 Value* rightOpToUse = rightVal;
2382 if (opType->isIntegral() && opType->isSigned()) {
2383 unsigned opSize = target.getTargetData().getTypeSize(opType);
2385 MachineCodeForInstruction& mcfi =
2386 MachineCodeForInstruction::get(setCCInstr);
2388 // create temporary virtual regs. to hold the sign-extensions
2389 leftOpToUse = new TmpInstruction(mcfi, leftVal);
2390 rightOpToUse = new TmpInstruction(mcfi, rightVal);
2392 // sign-extend each operand and put the result in the temporary reg.
2393 target.getInstrInfo().CreateSignExtensionInstructions
2394 (target, setCCInstr->getParent()->getParent(),
2395 leftVal, leftOpToUse, 8*opSize, mvec, mcfi);
2396 target.getInstrInfo().CreateSignExtensionInstructions
2397 (target, setCCInstr->getParent()->getParent(),
2398 rightVal, rightOpToUse, 8*opSize, mvec, mcfi);
2402 if (! isFPCompare) {
2403 // Integer condition: set CC and discard result.
2404 mvec.push_back(BuildMI(V9::SUBccr, 4)
2405 .addReg(leftOpToUse)
2406 .addReg(rightOpToUse)
2407 .addMReg(target.getRegInfo().getZeroRegNum(),MOTy::Def)
2408 .addCCReg(tmpForCC, MOTy::Def));
2410 // FP condition: dest of FCMP should be some FCCn register
2411 mvec.push_back(BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
2412 .addCCReg(tmpForCC, MOTy::Def)
2413 .addReg(leftOpToUse)
2414 .addReg(rightOpToUse));
2417 if (computeBoolVal) {
2418 MachineOpCode movOpCode = (isFPCompare
2419 ? ChooseMovFpcciInstruction(subtreeRoot)
2420 : ChooseMovpcciForSetCC(subtreeRoot));
2422 // Unconditionally set register to 0
2423 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
2426 // Now conditionally move 1 into the register.
2427 // Mark the register as a use (as well as a def) because the old
2428 // value will be retained if the condition is false.
2429 M = (BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(1)
2430 .addReg(setCCInstr, MOTy::UseAndDef));
2436 case 51: // reg: Load(reg)
2437 case 52: // reg: Load(ptrreg)
2438 SetOperandsForMemInstr(ChooseLoadInstruction(
2439 subtreeRoot->getValue()->getType()),
2440 mvec, subtreeRoot, target);
2443 case 55: // reg: GetElemPtr(reg)
2444 case 56: // reg: GetElemPtrIdx(reg,reg)
2445 // If the GetElemPtr was folded into the user (parent), it will be
2446 // caught above. For other cases, we have to compute the address.
2447 SetOperandsForMemInstr(V9::ADDr, mvec, subtreeRoot, target);
2450 case 57: // reg: Alloca: Implement as 1 instruction:
2451 { // add %fp, offsetFromFP -> result
2452 AllocationInst* instr =
2453 cast<AllocationInst>(subtreeRoot->getInstruction());
2455 target.getTargetData().getTypeSize(instr->getAllocatedType());
2457 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
2461 case 58: // reg: Alloca(reg): Implement as 3 instructions:
2462 // mul num, typeSz -> tmp
2463 // sub %sp, tmp -> %sp
2464 { // add %sp, frameSizeBelowDynamicArea -> result
2465 AllocationInst* instr =
2466 cast<AllocationInst>(subtreeRoot->getInstruction());
2467 const Type* eltType = instr->getAllocatedType();
2469 // If #elements is constant, use simpler code for fixed-size allocas
2470 int tsize = (int) target.getTargetData().getTypeSize(eltType);
2471 Value* numElementsVal = NULL;
2472 bool isArray = instr->isArrayAllocation();
2474 if (!isArray || isa<Constant>(numElementsVal = instr->getArraySize())) {
2475 // total size is constant: generate code for fixed-size alloca
2476 unsigned numElements = isArray?
2477 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
2478 CreateCodeForFixedSizeAlloca(target, instr, tsize,
2481 // total size is not constant.
2482 CreateCodeForVariableSizeAlloca(target, instr, tsize,
2483 numElementsVal, mvec);
2488 case 61: // reg: Call
2489 { // Generate a direct (CALL) or indirect (JMPL) call.
2490 // Mark the return-address register, the indirection
2491 // register (for indirect calls), the operands of the Call,
2492 // and the return value (if any) as implicit operands
2493 // of the machine instruction.
2495 // If this is a varargs function, floating point arguments
2496 // have to passed in integer registers so insert
2497 // copy-float-to-int instructions for each float operand.
2499 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
2500 Value *callee = callInstr->getCalledValue();
2501 Function* calledFunc = dyn_cast<Function>(callee);
2503 // Check if this is an intrinsic function that needs a special code
2504 // sequence (e.g., va_start). Indirect calls cannot be special.
2506 bool specialIntrinsic = false;
2507 LLVMIntrinsic::ID iid;
2508 if (calledFunc && (iid=(LLVMIntrinsic::ID)calledFunc->getIntrinsicID()))
2509 specialIntrinsic = CodeGenIntrinsic(iid, *callInstr, target, mvec);
2511 // If not, generate the normal call sequence for the function.
2512 // This can also handle any intrinsics that are just function calls.
2514 if (! specialIntrinsic) {
2515 Function* currentFunc = callInstr->getParent()->getParent();
2516 MachineFunction& MF = MachineFunction::get(currentFunc);
2517 MachineCodeForInstruction& mcfi =
2518 MachineCodeForInstruction::get(callInstr);
2519 const UltraSparcRegInfo& regInfo =
2520 (UltraSparcRegInfo&) target.getRegInfo();
2521 const TargetFrameInfo& frameInfo = target.getFrameInfo();
2523 // Create hidden virtual register for return address with type void*
2524 TmpInstruction* retAddrReg =
2525 new TmpInstruction(mcfi, PointerType::get(Type::VoidTy), callInstr);
2527 // Generate the machine instruction and its operands.
2528 // Use CALL for direct function calls; this optimistically assumes
2529 // the PC-relative address fits in the CALL address field (22 bits).
2530 // Use JMPL for indirect calls.
2531 // This will be added to mvec later, after operand copies.
2533 MachineInstr* callMI;
2534 if (calledFunc) // direct function call
2535 callMI = BuildMI(V9::CALL, 1).addPCDisp(callee);
2536 else // indirect function call
2537 callMI = (BuildMI(V9::JMPLCALLi,3).addReg(callee)
2538 .addSImm((int64_t)0).addRegDef(retAddrReg));
2540 const FunctionType* funcType =
2541 cast<FunctionType>(cast<PointerType>(callee->getType())
2542 ->getElementType());
2543 bool isVarArgs = funcType->isVarArg();
2544 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
2546 // Use a descriptor to pass information about call arguments
2547 // to the register allocator. This descriptor will be "owned"
2548 // and freed automatically when the MachineCodeForInstruction
2549 // object for the callInstr goes away.
2550 CallArgsDescriptor* argDesc =
2551 new CallArgsDescriptor(callInstr, retAddrReg,isVarArgs,noPrototype);
2552 assert(callInstr->getOperand(0) == callee
2553 && "This is assumed in the loop below!");
2555 // Insert sign-extension instructions for small signed values,
2556 // if this is an unknown function (i.e., called via a funcptr)
2557 // or an external one (i.e., which may not be compiled by llc).
2559 if (calledFunc == NULL || calledFunc->isExternal()) {
2560 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2561 Value* argVal = callInstr->getOperand(i);
2562 const Type* argType = argVal->getType();
2563 if (argType->isIntegral() && argType->isSigned()) {
2564 unsigned argSize = target.getTargetData().getTypeSize(argType);
2566 // create a temporary virtual reg. to hold the sign-extension
2567 TmpInstruction* argExtend = new TmpInstruction(mcfi, argVal);
2569 // sign-extend argVal and put the result in the temporary reg.
2570 target.getInstrInfo().CreateSignExtensionInstructions
2571 (target, currentFunc, argVal, argExtend,
2572 8*argSize, mvec, mcfi);
2574 // replace argVal with argExtend in CallArgsDescriptor
2575 argDesc->getArgInfo(i-1).replaceArgVal(argExtend);
2581 // Insert copy instructions to get all the arguments into
2582 // all the places that they need to be.
2584 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2586 CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
2587 Value* argVal = argInfo.getArgVal(); // don't use callInstr arg here
2588 const Type* argType = argVal->getType();
2589 unsigned regType = regInfo.getRegTypeForDataType(argType);
2590 unsigned argSize = target.getTargetData().getTypeSize(argType);
2591 int regNumForArg = TargetRegInfo::getInvalidRegNum();
2592 unsigned regClassIDOfArgReg;
2594 // Check for FP arguments to varargs functions.
2595 // Any such argument in the first $K$ args must be passed in an
2596 // integer register. If there is no prototype, it must also
2597 // be passed as an FP register.
2598 // K = #integer argument registers.
2599 bool isFPArg = argVal->getType()->isFloatingPoint();
2600 if (isVarArgs && isFPArg) {
2603 // It is a function with no prototype: pass value
2604 // as an FP value as well as a varargs value. The FP value
2605 // may go in a register or on the stack. The copy instruction
2606 // to the outgoing reg/stack is created by the normal argument
2607 // handling code since this is the "normal" passing mode.
2609 regNumForArg = regInfo.regNumForFPArg(regType,
2610 false, false, argNo,
2611 regClassIDOfArgReg);
2612 if (regNumForArg == regInfo.getInvalidRegNum())
2613 argInfo.setUseStackSlot();
2615 argInfo.setUseFPArgReg();
2618 // If this arg. is in the first $K$ regs, add special copy-
2619 // float-to-int instructions to pass the value as an int.
2620 // To check if it is in the first $K$, get the register
2621 // number for the arg #i. These copy instructions are
2622 // generated here because they are extra cases and not needed
2623 // for the normal argument handling (some code reuse is
2624 // possible though -- later).
2626 int copyRegNum = regInfo.regNumForIntArg(false, false, argNo,
2627 regClassIDOfArgReg);
2628 if (copyRegNum != regInfo.getInvalidRegNum()) {
2629 // Create a virtual register to represent copyReg. Mark
2630 // this vreg as being an implicit operand of the call MI
2631 const Type* loadTy = (argType == Type::FloatTy
2632 ? Type::IntTy : Type::LongTy);
2633 TmpInstruction* argVReg = new TmpInstruction(mcfi, loadTy,
2636 callMI->addImplicitRef(argVReg);
2638 // Get a temp stack location to use to copy
2639 // float-to-int via the stack.
2641 // FIXME: For now, we allocate permanent space because
2642 // the stack frame manager does not allow locals to be
2643 // allocated (e.g., for alloca) after a temp is
2646 // int tmpOffset = MF.getInfo()->pushTempValue(argSize);
2647 int tmpOffset = MF.getInfo()->allocateLocalVar(argVReg);
2649 // Generate the store from FP reg to stack
2650 unsigned StoreOpcode = ChooseStoreInstruction(argType);
2651 M = BuildMI(convertOpcodeFromRegToImm(StoreOpcode), 3)
2652 .addReg(argVal).addMReg(regInfo.getFramePointer())
2653 .addSImm(tmpOffset);
2656 // Generate the load from stack to int arg reg
2657 unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
2658 M = BuildMI(convertOpcodeFromRegToImm(LoadOpcode), 3)
2659 .addMReg(regInfo.getFramePointer()).addSImm(tmpOffset)
2660 .addReg(argVReg, MOTy::Def);
2662 // Mark operand with register it should be assigned
2663 // both for copy and for the callMI
2664 M->SetRegForOperand(M->getNumOperands()-1, copyRegNum);
2665 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2669 // Add info about the argument to the CallArgsDescriptor
2670 argInfo.setUseIntArgReg();
2671 argInfo.setArgCopy(copyRegNum);
2673 // Cannot fit in first $K$ regs so pass arg on stack
2674 argInfo.setUseStackSlot();
2676 } else if (isFPArg) {
2677 // Get the outgoing arg reg to see if there is one.
2678 regNumForArg = regInfo.regNumForFPArg(regType, false, false,
2679 argNo, regClassIDOfArgReg);
2680 if (regNumForArg == regInfo.getInvalidRegNum())
2681 argInfo.setUseStackSlot();
2683 argInfo.setUseFPArgReg();
2684 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2688 // Get the outgoing arg reg to see if there is one.
2689 regNumForArg = regInfo.regNumForIntArg(false,false,
2690 argNo, regClassIDOfArgReg);
2691 if (regNumForArg == regInfo.getInvalidRegNum())
2692 argInfo.setUseStackSlot();
2694 argInfo.setUseIntArgReg();
2695 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2701 // Now insert copy instructions to stack slot or arg. register
2703 if (argInfo.usesStackSlot()) {
2704 // Get the stack offset for this argument slot.
2705 // FP args on stack are right justified so adjust offset!
2706 // int arguments are also right justified but they are
2707 // always loaded as a full double-word so the offset does
2708 // not need to be adjusted.
2709 int argOffset = frameInfo.getOutgoingArgOffset(MF, argNo);
2710 if (argType->isFloatingPoint()) {
2711 unsigned slotSize = frameInfo.getSizeOfEachArgOnStack();
2712 assert(argSize <= slotSize && "Insufficient slot size!");
2713 argOffset += slotSize - argSize;
2716 // Now generate instruction to copy argument to stack
2717 MachineOpCode storeOpCode =
2718 (argType->isFloatingPoint()
2719 ? ((argSize == 4)? V9::STFi : V9::STDFi) : V9::STXi);
2721 M = BuildMI(storeOpCode, 3).addReg(argVal)
2722 .addMReg(regInfo.getStackPointer()).addSImm(argOffset);
2725 else if (regNumForArg != regInfo.getInvalidRegNum()) {
2727 // Create a virtual register to represent the arg reg. Mark
2728 // this vreg as being an implicit operand of the call MI.
2729 TmpInstruction* argVReg =
2730 new TmpInstruction(mcfi, argVal, NULL, "argReg");
2732 callMI->addImplicitRef(argVReg);
2734 // Generate the reg-to-reg copy into the outgoing arg reg.
2735 // -- For FP values, create a FMOVS or FMOVD instruction
2736 // -- For non-FP values, create an add-with-0 instruction
2737 if (argType->isFloatingPoint())
2738 M=(BuildMI(argType==Type::FloatTy? V9::FMOVS :V9::FMOVD,2)
2739 .addReg(argVal).addReg(argVReg, MOTy::Def));
2741 M = (BuildMI(ChooseAddInstructionByType(argType), 3)
2742 .addReg(argVal).addSImm((int64_t) 0)
2743 .addReg(argVReg, MOTy::Def));
2745 // Mark the operand with the register it should be assigned
2746 M->SetRegForOperand(M->getNumOperands()-1, regNumForArg);
2747 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2753 assert(argInfo.getArgCopy() != regInfo.getInvalidRegNum() &&
2754 "Arg. not in stack slot, primary or secondary register?");
2757 // add call instruction and delay slot before copying return value
2758 mvec.push_back(callMI);
2759 mvec.push_back(BuildMI(V9::NOP, 0));
2761 // Add the return value as an implicit ref. The call operands
2762 // were added above. Also, add code to copy out the return value.
2763 // This is always register-to-register for int or FP return values.
2765 if (callInstr->getType() != Type::VoidTy) {
2766 // Get the return value reg.
2767 const Type* retType = callInstr->getType();
2769 int regNum = (retType->isFloatingPoint()
2770 ? (unsigned) SparcFloatRegClass::f0
2771 : (unsigned) SparcIntRegClass::o0);
2772 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
2773 regNum = regInfo.getUnifiedRegNum(regClassID, regNum);
2775 // Create a virtual register to represent it and mark
2776 // this vreg as being an implicit operand of the call MI
2777 TmpInstruction* retVReg =
2778 new TmpInstruction(mcfi, callInstr, NULL, "argReg");
2780 callMI->addImplicitRef(retVReg, /*isDef*/ true);
2782 // Generate the reg-to-reg copy from the return value reg.
2783 // -- For FP values, create a FMOVS or FMOVD instruction
2784 // -- For non-FP values, create an add-with-0 instruction
2785 if (retType->isFloatingPoint())
2786 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
2787 .addReg(retVReg).addReg(callInstr, MOTy::Def));
2789 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
2790 .addReg(retVReg).addSImm((int64_t) 0)
2791 .addReg(callInstr, MOTy::Def));
2793 // Mark the operand with the register it should be assigned
2794 // Also mark the implicit ref of the call defining this operand
2795 M->SetRegForOperand(0, regNum);
2796 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,regNum);
2801 // For the CALL instruction, the ret. addr. reg. is also implicit
2802 if (isa<Function>(callee))
2803 callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
2805 MF.getInfo()->popAllTempValues(); // free temps used for this inst
2811 case 62: // reg: Shl(reg, reg)
2813 Value* argVal1 = subtreeRoot->leftChild()->getValue();
2814 Value* argVal2 = subtreeRoot->rightChild()->getValue();
2815 Instruction* shlInstr = subtreeRoot->getInstruction();
2817 const Type* opType = argVal1->getType();
2818 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2819 "Shl unsupported for other types");
2820 unsigned opSize = target.getTargetData().getTypeSize(opType);
2822 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
2823 (opSize > 4)? V9::SLLXr6:V9::SLLr5,
2824 argVal1, argVal2, 0, shlInstr, mvec,
2825 MachineCodeForInstruction::get(shlInstr));
2829 case 63: // reg: Shr(reg, reg)
2831 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
2832 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2833 "Shr unsupported for other types");
2834 unsigned opSize = target.getTargetData().getTypeSize(opType);
2835 Add3OperandInstr(opType->isSigned()
2836 ? (opSize > 4? V9::SRAXr6 : V9::SRAr5)
2837 : (opSize > 4? V9::SRLXr6 : V9::SRLr5),
2842 case 64: // reg: Phi(reg,reg)
2843 break; // don't forward the value
2845 case 65: // reg: VANext(reg): the va_next instruction
2846 case 66: // reg: VAArg (reg): the va_arg instruction
2848 abort(); // FIXME: This is incorrect!
2850 // Use value initialized by va_start as pointer to args on the stack.
2851 // Load argument via current pointer value, then increment pointer.
2852 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
2853 Instruction* vaArgI = subtreeRoot->getInstruction();
2854 MachineOpCode loadOp = vaArgI->getType()->isFloatingPoint()? V9::LDDFi
2856 mvec.push_back(BuildMI(loadOp, 3).addReg(vaArgI->getOperand(0)).
2857 addSImm(0).addRegDef(vaArgI));
2858 mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaArgI->getOperand(0)).
2859 addSImm(argSize).addRegDef(vaArgI->getOperand(0)));
2864 case 71: // reg: VReg
2865 case 72: // reg: Constant
2866 break; // don't forward the value
2869 assert(0 && "Unrecognized BURG rule");
2874 if (forwardOperandNum >= 0) {
2875 // We did not generate a machine instruction but need to use operand.
2876 // If user is in the same tree, replace Value in its machine operand.
2877 // If not, insert a copy instruction which should get coalesced away
2878 // by register allocation.
2879 if (subtreeRoot->parent() != NULL)
2880 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
2882 std::vector<MachineInstr*> minstrVec;
2883 Instruction* instr = subtreeRoot->getInstruction();
2884 target.getInstrInfo().
2885 CreateCopyInstructionsByType(target,
2886 instr->getParent()->getParent(),
2887 instr->getOperand(forwardOperandNum),
2889 MachineCodeForInstruction::get(instr));
2890 assert(minstrVec.size() > 0);
2891 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
2895 if (maskUnsignedResult) {
2896 // If result is unsigned and smaller than int reg size,
2897 // we need to clear high bits of result value.
2898 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
2899 Instruction* dest = subtreeRoot->getInstruction();
2900 if (dest->getType()->isUnsigned()) {
2901 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
2902 if (destSize <= 4) {
2903 // Mask high 64 - N bits, where N = 4*destSize.
2905 // Use a TmpInstruction to represent the
2906 // intermediate result before masking. Since those instructions
2907 // have already been generated, go back and substitute tmpI
2908 // for dest in the result position of each one of them.
2910 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(dest);
2911 TmpInstruction *tmpI = new TmpInstruction(mcfi, dest->getType(),
2912 dest, NULL, "maskHi");
2913 Value* srlArgToUse = tmpI;
2915 unsigned numSubst = 0;
2916 for (unsigned i=0, N=mvec.size(); i < N; ++i) {
2918 // Make sure we substitute all occurrences of dest in these instrs.
2919 // Otherwise, we will have bogus code.
2920 bool someArgsWereIgnored = false;
2922 // Make sure not to substitute an upwards-exposed use -- that would
2923 // introduce a use of `tmpI' with no preceding def. Therefore,
2924 // substitute a use or def-and-use operand only if a previous def
2925 // operand has already been substituted (i.e., numSusbt > 0).
2927 numSubst += mvec[i]->substituteValue(dest, tmpI,
2928 /*defsOnly*/ numSubst == 0,
2929 /*notDefsAndUses*/ numSubst > 0,
2930 someArgsWereIgnored);
2931 assert(!someArgsWereIgnored &&
2932 "Operand `dest' exists but not replaced: probably bogus!");
2934 assert(numSubst > 0 && "Operand `dest' not replaced: probably bogus!");
2936 // Left shift 32-N if size (N) is less than 32 bits.
2937 // Use another tmp. virtual register to represent this result.
2939 srlArgToUse = new TmpInstruction(mcfi, dest->getType(),
2940 tmpI, NULL, "maskHi2");
2941 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpI)
2942 .addZImm(8*(4-destSize))
2943 .addReg(srlArgToUse, MOTy::Def));
2946 // Logical right shift 32-N to get zero extension in top 64-N bits.
2947 mvec.push_back(BuildMI(V9::SRLi5, 3).addReg(srlArgToUse)
2948 .addZImm(8*(4-destSize)).addReg(dest, MOTy::Def));
2950 } else if (destSize < 8) {
2951 assert(0 && "Unsupported type size: 32 < size < 64 bits");