1 //===-- SparcInstrSelection.cpp -------------------------------------------===//
3 // BURS instruction selection for SPARC V9 architecture.
5 //===----------------------------------------------------------------------===//
7 #include "SparcInternals.h"
8 #include "SparcInstrSelectionSupport.h"
9 #include "SparcRegClassInfo.h"
10 #include "llvm/CodeGen/InstrSelectionSupport.h"
11 #include "llvm/CodeGen/MachineInstrBuilder.h"
12 #include "llvm/CodeGen/MachineInstrAnnot.h"
13 #include "llvm/CodeGen/InstrForest.h"
14 #include "llvm/CodeGen/InstrSelection.h"
15 #include "llvm/CodeGen/MachineFunction.h"
16 #include "llvm/CodeGen/MachineFunctionInfo.h"
17 #include "llvm/CodeGen/MachineCodeForInstruction.h"
18 #include "llvm/DerivedTypes.h"
19 #include "llvm/iTerminators.h"
20 #include "llvm/iMemory.h"
21 #include "llvm/iOther.h"
22 #include "llvm/Function.h"
23 #include "llvm/Constants.h"
24 #include "llvm/ConstantHandling.h"
25 #include "llvm/Intrinsics.h"
26 #include "Support/MathExtras.h"
29 static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
30 std::vector<MachineInstr*>& mvec) {
31 mvec.push_back(BuildMI(Opcode, 3).addReg(Node->leftChild()->getValue())
32 .addReg(Node->rightChild()->getValue())
33 .addRegDef(Node->getValue()));
38 //---------------------------------------------------------------------------
39 // Function: GetMemInstArgs
42 // Get the pointer value and the index vector for a memory operation
43 // (GetElementPtr, Load, or Store). If all indices of the given memory
44 // operation are constant, fold in constant indices in a chain of
45 // preceding GetElementPtr instructions (if any), and return the
46 // pointer value of the first instruction in the chain.
47 // All folded instructions are marked so no code is generated for them.
50 // Returns the pointer Value to use.
51 // Returns the resulting IndexVector in idxVec.
52 // Returns true/false in allConstantIndices if all indices are/aren't const.
53 //---------------------------------------------------------------------------
56 //---------------------------------------------------------------------------
57 // Function: FoldGetElemChain
60 // Fold a chain of GetElementPtr instructions containing only
61 // constant offsets into an equivalent (Pointer, IndexVector) pair.
62 // Returns the pointer Value, and stores the resulting IndexVector
63 // in argument chainIdxVec. This is a helper function for
64 // FoldConstantIndices that does the actual folding.
65 //---------------------------------------------------------------------------
68 // Check for a constant 0.
72 return (idx == ConstantSInt::getNullValue(idx->getType()));
76 FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
77 bool lastInstHasLeadingNonZero)
79 InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
80 GetElementPtrInst* gepInst =
81 dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
83 // ptr value is not computed in this tree or ptr value does not come from GEP
88 // Return NULL if we don't fold any instructions in.
91 // Now chase the chain of getElementInstr instructions, if any.
92 // Check for any non-constant indices and stop there.
93 // Also, stop if the first index of child is a non-zero array index
94 // and the last index of the current node is a non-array index:
95 // in that case, a non-array declared type is being accessed as an array
96 // which is not type-safe, but could be legal.
98 InstructionNode* ptrChild = gepNode;
99 while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
100 ptrChild->getOpLabel() == GetElemPtrIdx))
102 // Child is a GetElemPtr instruction
103 gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
104 User::op_iterator OI, firstIdx = gepInst->idx_begin();
105 User::op_iterator lastIdx = gepInst->idx_end();
106 bool allConstantOffsets = true;
108 // The first index of every GEP must be an array index.
109 assert((*firstIdx)->getType() == Type::LongTy &&
110 "INTERNAL ERROR: Structure index for a pointer type!");
112 // If the last instruction had a leading non-zero index, check if the
113 // current one references a sequential (i.e., indexable) type.
114 // If not, the code is not type-safe and we would create an illegal GEP
115 // by folding them, so don't fold any more instructions.
117 if (lastInstHasLeadingNonZero)
118 if (! isa<SequentialType>(gepInst->getType()->getElementType()))
119 break; // cannot fold in any preceding getElementPtr instrs.
121 // Check that all offsets are constant for this instruction
122 for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
123 allConstantOffsets = isa<ConstantInt>(*OI);
125 if (allConstantOffsets) {
126 // Get pointer value out of ptrChild.
127 ptrVal = gepInst->getPointerOperand();
129 // Insert its index vector at the start, skipping any leading [0]
130 // Remember the old size to check if anything was inserted.
131 unsigned oldSize = chainIdxVec.size();
132 int firstIsZero = IsZero(*firstIdx);
133 chainIdxVec.insert(chainIdxVec.begin(), firstIdx + firstIsZero, lastIdx);
135 // Remember if it has leading zero index: it will be discarded later.
136 if (oldSize < chainIdxVec.size())
137 lastInstHasLeadingNonZero = !firstIsZero;
139 // Mark the folded node so no code is generated for it.
140 ((InstructionNode*) ptrChild)->markFoldedIntoParent();
142 // Get the previous GEP instruction and continue trying to fold
143 ptrChild = dyn_cast<InstructionNode>(ptrChild->leftChild());
144 } else // cannot fold this getElementPtr instr. or any preceding ones
148 // If the first getElementPtr instruction had a leading [0], add it back.
149 // Note that this instruction is the *last* one that was successfully
150 // folded *and* contributed any indices, in the loop above.
152 if (ptrVal && ! lastInstHasLeadingNonZero)
153 chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
159 //---------------------------------------------------------------------------
160 // Function: GetGEPInstArgs
163 // Helper function for GetMemInstArgs that handles the final getElementPtr
164 // instruction used by (or same as) the memory operation.
165 // Extracts the indices of the current instruction and tries to fold in
166 // preceding ones if all indices of the current one are constant.
167 //---------------------------------------------------------------------------
170 GetGEPInstArgs(InstructionNode* gepNode,
171 std::vector<Value*>& idxVec,
172 bool& allConstantIndices)
174 allConstantIndices = true;
175 GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
177 // Default pointer is the one from the current instruction.
178 Value* ptrVal = gepI->getPointerOperand();
179 InstrTreeNode* ptrChild = gepNode->leftChild();
181 // Extract the index vector of the GEP instructin.
182 // If all indices are constant and first index is zero, try to fold
183 // in preceding GEPs with all constant indices.
184 for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
185 allConstantIndices && OI != OE; ++OI)
186 if (! isa<Constant>(*OI))
187 allConstantIndices = false; // note: this also terminates loop!
189 // If we have only constant indices, fold chains of constant indices
190 // in this and any preceding GetElemPtr instructions.
191 bool foldedGEPs = false;
192 bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
193 if (allConstantIndices)
194 if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx)) {
199 // Append the index vector of the current instruction.
200 // Skip the leading [0] index if preceding GEPs were folded into this.
201 idxVec.insert(idxVec.end(),
202 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
208 //---------------------------------------------------------------------------
209 // Function: GetMemInstArgs
212 // Get the pointer value and the index vector for a memory operation
213 // (GetElementPtr, Load, or Store). If all indices of the given memory
214 // operation are constant, fold in constant indices in a chain of
215 // preceding GetElementPtr instructions (if any), and return the
216 // pointer value of the first instruction in the chain.
217 // All folded instructions are marked so no code is generated for them.
220 // Returns the pointer Value to use.
221 // Returns the resulting IndexVector in idxVec.
222 // Returns true/false in allConstantIndices if all indices are/aren't const.
223 //---------------------------------------------------------------------------
226 GetMemInstArgs(InstructionNode* memInstrNode,
227 std::vector<Value*>& idxVec,
228 bool& allConstantIndices)
230 allConstantIndices = false;
231 Instruction* memInst = memInstrNode->getInstruction();
232 assert(idxVec.size() == 0 && "Need empty vector to return indices");
234 // If there is a GetElemPtr instruction to fold in to this instr,
235 // it must be in the left child for Load and GetElemPtr, and in the
236 // right child for Store instructions.
237 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
238 ? memInstrNode->rightChild()
239 : memInstrNode->leftChild());
241 // Default pointer is the one from the current instruction.
242 Value* ptrVal = ptrChild->getValue();
244 // Find the "last" GetElemPtr instruction: this one or the immediate child.
245 // There will be none if this is a load or a store from a scalar pointer.
246 InstructionNode* gepNode = NULL;
247 if (isa<GetElementPtrInst>(memInst))
248 gepNode = memInstrNode;
249 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
250 // Child of load/store is a GEP and memInst is its only use.
251 // Use its indices and mark it as folded.
252 gepNode = cast<InstructionNode>(ptrChild);
253 gepNode->markFoldedIntoParent();
256 // If there are no indices, return the current pointer.
257 // Else extract the pointer from the GEP and fold the indices.
258 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
263 //************************ Internal Functions ******************************/
266 static inline MachineOpCode
267 ChooseBprInstruction(const InstructionNode* instrNode)
269 MachineOpCode opCode;
271 Instruction* setCCInstr =
272 ((InstructionNode*) instrNode->leftChild())->getInstruction();
274 switch(setCCInstr->getOpcode())
276 case Instruction::SetEQ: opCode = V9::BRZ; break;
277 case Instruction::SetNE: opCode = V9::BRNZ; break;
278 case Instruction::SetLE: opCode = V9::BRLEZ; break;
279 case Instruction::SetGE: opCode = V9::BRGEZ; break;
280 case Instruction::SetLT: opCode = V9::BRLZ; break;
281 case Instruction::SetGT: opCode = V9::BRGZ; break;
283 assert(0 && "Unrecognized VM instruction!");
284 opCode = V9::INVALID_OPCODE;
292 static inline MachineOpCode
293 ChooseBpccInstruction(const InstructionNode* instrNode,
294 const BinaryOperator* setCCInstr)
296 MachineOpCode opCode = V9::INVALID_OPCODE;
298 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
301 switch(setCCInstr->getOpcode())
303 case Instruction::SetEQ: opCode = V9::BE; break;
304 case Instruction::SetNE: opCode = V9::BNE; break;
305 case Instruction::SetLE: opCode = V9::BLE; break;
306 case Instruction::SetGE: opCode = V9::BGE; break;
307 case Instruction::SetLT: opCode = V9::BL; break;
308 case Instruction::SetGT: opCode = V9::BG; break;
310 assert(0 && "Unrecognized VM instruction!");
314 switch(setCCInstr->getOpcode())
316 case Instruction::SetEQ: opCode = V9::BE; break;
317 case Instruction::SetNE: opCode = V9::BNE; break;
318 case Instruction::SetLE: opCode = V9::BLEU; break;
319 case Instruction::SetGE: opCode = V9::BCC; break;
320 case Instruction::SetLT: opCode = V9::BCS; break;
321 case Instruction::SetGT: opCode = V9::BGU; break;
323 assert(0 && "Unrecognized VM instruction!");
331 static inline MachineOpCode
332 ChooseBFpccInstruction(const InstructionNode* instrNode,
333 const BinaryOperator* setCCInstr)
335 MachineOpCode opCode = V9::INVALID_OPCODE;
337 switch(setCCInstr->getOpcode())
339 case Instruction::SetEQ: opCode = V9::FBE; break;
340 case Instruction::SetNE: opCode = V9::FBNE; break;
341 case Instruction::SetLE: opCode = V9::FBLE; break;
342 case Instruction::SetGE: opCode = V9::FBGE; break;
343 case Instruction::SetLT: opCode = V9::FBL; break;
344 case Instruction::SetGT: opCode = V9::FBG; break;
346 assert(0 && "Unrecognized VM instruction!");
354 // Create a unique TmpInstruction for a boolean value,
355 // representing the CC register used by a branch on that value.
356 // For now, hack this using a little static cache of TmpInstructions.
357 // Eventually the entire BURG instruction selection should be put
358 // into a separate class that can hold such information.
359 // The static cache is not too bad because the memory for these
360 // TmpInstructions will be freed along with the rest of the Function anyway.
362 static TmpInstruction*
363 GetTmpForCC(Value* boolVal, const Function *F, const Type* ccType,
364 MachineCodeForInstruction& mcfi)
366 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
367 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
368 static const Function *lastFunction = 0;// Use to flush cache between funcs
370 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
372 if (lastFunction != F) {
374 boolToTmpCache.clear();
377 // Look for tmpI and create a new one otherwise. The new value is
378 // directly written to map using the ref returned by operator[].
379 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
381 tmpI = new TmpInstruction(mcfi, ccType, boolVal);
387 static inline MachineOpCode
388 ChooseBccInstruction(const InstructionNode* instrNode,
391 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
392 assert(setCCNode->getOpLabel() == SetCCOp);
393 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
394 const Type* setCCType = setCCInstr->getOperand(0)->getType();
396 isFPBranch = setCCType->isFloatingPoint(); // Return value: don't delete!
399 return ChooseBFpccInstruction(instrNode, setCCInstr);
401 return ChooseBpccInstruction(instrNode, setCCInstr);
405 // WARNING: since this function has only one caller, it always returns
406 // the opcode that expects an immediate and a register. If this function
407 // is ever used in cases where an opcode that takes two registers is required,
408 // then modify this function and use convertOpcodeFromRegToImm() where required.
410 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
411 // new cases of opcodes.
412 static inline MachineOpCode
413 ChooseMovFpcciInstruction(const InstructionNode* instrNode)
415 MachineOpCode opCode = V9::INVALID_OPCODE;
417 switch(instrNode->getInstruction()->getOpcode())
419 case Instruction::SetEQ: opCode = V9::MOVFEi; break;
420 case Instruction::SetNE: opCode = V9::MOVFNEi; break;
421 case Instruction::SetLE: opCode = V9::MOVFLEi; break;
422 case Instruction::SetGE: opCode = V9::MOVFGEi; break;
423 case Instruction::SetLT: opCode = V9::MOVFLi; break;
424 case Instruction::SetGT: opCode = V9::MOVFGi; break;
426 assert(0 && "Unrecognized VM instruction!");
434 // Assumes that SUBcc v1, v2 -> v3 has been executed.
435 // In most cases, we want to clear v3 and then follow it by instruction
437 // Set mustClearReg=false if v3 need not be cleared before conditional move.
438 // Set valueToMove=0 if we want to conditionally move 0 instead of 1
439 // (i.e., we want to test inverse of a condition)
440 // (The latter two cases do not seem to arise because SetNE needs nothing.)
442 // WARNING: since this function has only one caller, it always returns
443 // the opcode that expects an immediate and a register. If this function
444 // is ever used in cases where an opcode that takes two registers is required,
445 // then modify this function and use convertOpcodeFromRegToImm() where required.
447 // It will be necessary to expand convertOpcodeFromRegToImm() to handle the
448 // new cases of opcodes.
450 ChooseMovpcciAfterSub(const InstructionNode* instrNode)
452 MachineOpCode opCode = V9::INVALID_OPCODE;
454 switch(instrNode->getInstruction()->getOpcode())
456 case Instruction::SetEQ: opCode = V9::MOVEi; break;
457 case Instruction::SetLE: opCode = V9::MOVLEi; break;
458 case Instruction::SetGE: opCode = V9::MOVGEi; break;
459 case Instruction::SetLT: opCode = V9::MOVLi; break;
460 case Instruction::SetGT: opCode = V9::MOVGi; break;
461 case Instruction::SetNE: opCode = V9::MOVNEi; break;
462 default: assert(0 && "Unrecognized VM instr!"); break;
468 static inline MachineOpCode
469 ChooseConvertToFloatInstr(OpLabel vopCode, const Type* opType)
471 assert((vopCode == ToFloatTy || vopCode == ToDoubleTy) &&
472 "Unrecognized convert-to-float opcode!");
474 MachineOpCode opCode = V9::INVALID_OPCODE;
476 if (opType == Type::SByteTy || opType == Type::UByteTy ||
477 opType == Type::ShortTy || opType == Type::UShortTy ||
478 opType == Type::IntTy || opType == Type::UIntTy)
479 opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
480 else if (opType == Type::LongTy || opType == Type::ULongTy ||
481 isa<PointerType>(opType))
482 opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
483 else if (opType == Type::FloatTy)
484 opCode = (vopCode == ToFloatTy? V9::INVALID_OPCODE : V9::FSTOD);
485 else if (opType == Type::DoubleTy)
486 opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::INVALID_OPCODE);
488 assert(0 && "Trying to convert a non-scalar type to DOUBLE?");
493 static inline MachineOpCode
494 ChooseConvertFPToIntInstr(Type::PrimitiveID tid, const Type* opType)
496 MachineOpCode opCode = V9::INVALID_OPCODE;;
498 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
499 && "This function should only be called for FLOAT or DOUBLE");
501 // SPARC does not have a float-to-uint conversion, only a float-to-int.
502 // For converting an FP value to uint32_t, we first need to convert to
503 // uint64_t and then to uint32_t, or we may overflow the signed int
504 // representation even for legal uint32_t values. This expansion is
505 // done by the Preselection pass.
507 if (tid == Type::UIntTyID) {
508 assert(tid != Type::UIntTyID && "FP-to-uint conversions must be expanded"
509 " into FP->long->uint for SPARC v9: SO RUN PRESELECTION PASS!");
510 } else if (tid == Type::SByteTyID || tid == Type::ShortTyID ||
511 tid == Type::IntTyID || tid == Type::UByteTyID ||
512 tid == Type::UShortTyID) {
513 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
514 } else if (tid == Type::LongTyID || tid == Type::ULongTyID) {
515 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
517 assert(0 && "Should not get here, Mo!");
523 CreateConvertFPToIntInstr(Type::PrimitiveID destTID,
524 Value* srcVal, Value* destVal)
526 MachineOpCode opCode = ChooseConvertFPToIntInstr(destTID, srcVal->getType());
527 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
528 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
531 // CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned integer
532 // The FP value must be converted to the dest type in an FP register,
533 // and the result is then copied from FP to int register via memory.
535 // Since fdtoi converts to signed integers, any FP value V between MAXINT+1
536 // and MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly
537 // *only* when converting to an unsigned. (Unsigned byte, short or long
538 // don't have this problem.)
539 // For unsigned int, we therefore have to generate the code sequence:
541 // if (V > (float) MAXINT) {
542 // unsigned result = (unsigned) (V - (float) MAXINT);
543 // result = result + (unsigned) MAXINT;
546 // result = (unsigned) V;
549 CreateCodeToConvertFloatToInt(const TargetMachine& target,
552 std::vector<MachineInstr*>& mvec,
553 MachineCodeForInstruction& mcfi)
555 // Create a temporary to represent the FP register into which the
556 // int value will placed after conversion. The type of this temporary
557 // depends on the type of FP register to use: single-prec for a 32-bit
558 // int or smaller; double-prec for a 64-bit int.
560 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
561 const Type* destTypeToUse = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
562 TmpInstruction* destForCast = new TmpInstruction(mcfi, destTypeToUse, opVal);
564 // Create the fp-to-int conversion code
565 MachineInstr* M =CreateConvertFPToIntInstr(destI->getType()->getPrimitiveID(),
569 // Create the fpreg-to-intreg copy code
570 target.getInstrInfo().
571 CreateCodeToCopyFloatToInt(target, destI->getParent()->getParent(),
572 destForCast, destI, mvec, mcfi);
576 static inline MachineOpCode
577 ChooseAddInstruction(const InstructionNode* instrNode)
579 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
583 static inline MachineInstr*
584 CreateMovFloatInstruction(const InstructionNode* instrNode,
585 const Type* resultType)
587 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
588 .addReg(instrNode->leftChild()->getValue())
589 .addRegDef(instrNode->getValue());
592 static inline MachineInstr*
593 CreateAddConstInstruction(const InstructionNode* instrNode)
595 MachineInstr* minstr = NULL;
597 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
598 assert(isa<Constant>(constOp));
600 // Cases worth optimizing are:
601 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
602 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
604 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
605 double dval = FPC->getValue();
607 minstr = CreateMovFloatInstruction(instrNode,
608 instrNode->getInstruction()->getType());
615 static inline MachineOpCode
616 ChooseSubInstructionByType(const Type* resultType)
618 MachineOpCode opCode = V9::INVALID_OPCODE;
620 if (resultType->isInteger() || isa<PointerType>(resultType)) {
623 switch(resultType->getPrimitiveID())
625 case Type::FloatTyID: opCode = V9::FSUBS; break;
626 case Type::DoubleTyID: opCode = V9::FSUBD; break;
627 default: assert(0 && "Invalid type for SUB instruction"); break;
635 static inline MachineInstr*
636 CreateSubConstInstruction(const InstructionNode* instrNode)
638 MachineInstr* minstr = NULL;
640 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
641 assert(isa<Constant>(constOp));
643 // Cases worth optimizing are:
644 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
645 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
647 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
648 double dval = FPC->getValue();
650 minstr = CreateMovFloatInstruction(instrNode,
651 instrNode->getInstruction()->getType());
658 static inline MachineOpCode
659 ChooseFcmpInstruction(const InstructionNode* instrNode)
661 MachineOpCode opCode = V9::INVALID_OPCODE;
663 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
664 switch(operand->getType()->getPrimitiveID()) {
665 case Type::FloatTyID: opCode = V9::FCMPS; break;
666 case Type::DoubleTyID: opCode = V9::FCMPD; break;
667 default: assert(0 && "Invalid type for FCMP instruction"); break;
674 // Assumes that leftArg and rightArg are both cast instructions.
677 BothFloatToDouble(const InstructionNode* instrNode)
679 InstrTreeNode* leftArg = instrNode->leftChild();
680 InstrTreeNode* rightArg = instrNode->rightChild();
681 InstrTreeNode* leftArgArg = leftArg->leftChild();
682 InstrTreeNode* rightArgArg = rightArg->leftChild();
683 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
685 // Check if both arguments are floats cast to double
686 return (leftArg->getValue()->getType() == Type::DoubleTy &&
687 leftArgArg->getValue()->getType() == Type::FloatTy &&
688 rightArgArg->getValue()->getType() == Type::FloatTy);
692 static inline MachineOpCode
693 ChooseMulInstructionByType(const Type* resultType)
695 MachineOpCode opCode = V9::INVALID_OPCODE;
697 if (resultType->isInteger())
700 switch(resultType->getPrimitiveID())
702 case Type::FloatTyID: opCode = V9::FMULS; break;
703 case Type::DoubleTyID: opCode = V9::FMULD; break;
704 default: assert(0 && "Invalid type for MUL instruction"); break;
712 static inline MachineInstr*
713 CreateIntNegInstruction(const TargetMachine& target,
716 return BuildMI(V9::SUBr, 3).addMReg(target.getRegInfo().getZeroRegNum())
717 .addReg(vreg).addRegDef(vreg);
721 // Create instruction sequence for any shift operation.
722 // SLL or SLLX on an operand smaller than the integer reg. size (64bits)
723 // requires a second instruction for explicit sign-extension.
724 // Note that we only have to worry about a sign-bit appearing in the
725 // most significant bit of the operand after shifting (e.g., bit 32 of
726 // Int or bit 16 of Short), so we do not have to worry about results
727 // that are as large as a normal integer register.
730 CreateShiftInstructions(const TargetMachine& target,
732 MachineOpCode shiftOpCode,
734 Value* optArgVal2, /* Use optArgVal2 if not NULL */
735 unsigned optShiftNum, /* else use optShiftNum */
736 Instruction* destVal,
737 std::vector<MachineInstr*>& mvec,
738 MachineCodeForInstruction& mcfi)
740 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
741 "Large shift sizes unexpected, but can be handled below: "
742 "You need to check whether or not it fits in immed field below");
744 // If this is a logical left shift of a type smaller than the standard
745 // integer reg. size, we have to extend the sign-bit into upper bits
746 // of dest, so we need to put the result of the SLL into a temporary.
748 Value* shiftDest = destVal;
749 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
751 if ((shiftOpCode == V9::SLLr5 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
752 // put SLL result into a temporary
753 shiftDest = new TmpInstruction(mcfi, argVal1, optArgVal2, "sllTmp");
756 MachineInstr* M = (optArgVal2 != NULL)
757 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
758 .addReg(shiftDest, MOTy::Def)
759 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
760 .addReg(shiftDest, MOTy::Def);
763 if (shiftDest != destVal) {
764 // extend the sign-bit of the result into all upper bits of dest
765 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
766 target.getInstrInfo().
767 CreateSignExtensionInstructions(target, F, shiftDest, destVal,
768 8*opSize, mvec, mcfi);
773 // Does not create any instructions if we cannot exploit constant to
774 // create a cheaper instruction.
775 // This returns the approximate cost of the instructions generated,
776 // which is used to pick the cheapest when both operands are constant.
778 CreateMulConstInstruction(const TargetMachine &target, Function* F,
779 Value* lval, Value* rval, Instruction* destVal,
780 std::vector<MachineInstr*>& mvec,
781 MachineCodeForInstruction& mcfi)
783 /* Use max. multiply cost, viz., cost of MULX */
784 unsigned cost = target.getInstrInfo().minLatency(V9::MULXr);
785 unsigned firstNewInstr = mvec.size();
787 Value* constOp = rval;
788 if (! isa<Constant>(constOp))
791 // Cases worth optimizing are:
792 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
793 // (2) Multiply by 2^x for integer types: replace with Shift
795 const Type* resultType = destVal->getType();
797 if (resultType->isInteger() || isa<PointerType>(resultType)) {
799 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
802 bool needNeg = false;
808 if (C == 0 || C == 1) {
809 cost = target.getInstrInfo().minLatency(V9::ADDr);
810 unsigned Zero = target.getRegInfo().getZeroRegNum();
813 M =BuildMI(V9::ADDr,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
815 M = BuildMI(V9::ADDr,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
817 } else if (isPowerOf2(C, pow)) {
818 unsigned opSize = target.getTargetData().getTypeSize(resultType);
819 MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
820 CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
821 destVal, mvec, mcfi);
824 if (mvec.size() > 0 && needNeg) {
825 // insert <reg = SUB 0, reg> after the instr to flip the sign
826 MachineInstr* M = CreateIntNegInstruction(target, destVal);
831 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
832 double dval = FPC->getValue();
833 if (fabs(dval) == 1) {
834 MachineOpCode opCode = (dval < 0)
835 ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
836 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
837 mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
842 if (firstNewInstr < mvec.size()) {
844 for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
845 cost += target.getInstrInfo().minLatency(mvec[i]->getOpCode());
852 // Does not create any instructions if we cannot exploit constant to
853 // create a cheaper instruction.
856 CreateCheapestMulConstInstruction(const TargetMachine &target,
858 Value* lval, Value* rval,
859 Instruction* destVal,
860 std::vector<MachineInstr*>& mvec,
861 MachineCodeForInstruction& mcfi)
864 if (isa<Constant>(lval) && isa<Constant>(rval)) {
865 // both operands are constant: evaluate and "set" in dest
866 Constant* P = ConstantFoldBinaryInstruction(Instruction::Mul,
867 cast<Constant>(lval),
868 cast<Constant>(rval));
869 target.getInstrInfo().CreateCodeToLoadConst(target,F,P,destVal,mvec,mcfi);
871 else if (isa<Constant>(rval)) // rval is constant, but not lval
872 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
873 else if (isa<Constant>(lval)) // lval is constant, but not rval
874 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
876 // else neither is constant
880 // Return NULL if we cannot exploit constant to create a cheaper instruction
882 CreateMulInstruction(const TargetMachine &target, Function* F,
883 Value* lval, Value* rval, Instruction* destVal,
884 std::vector<MachineInstr*>& mvec,
885 MachineCodeForInstruction& mcfi,
886 MachineOpCode forceMulOp = INVALID_MACHINE_OPCODE)
888 unsigned L = mvec.size();
889 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
890 if (mvec.size() == L) {
891 // no instructions were added so create MUL reg, reg, reg.
892 // Use FSMULD if both operands are actually floats cast to doubles.
893 // Otherwise, use the default opcode for the appropriate type.
894 MachineOpCode mulOp = ((forceMulOp != INVALID_MACHINE_OPCODE)
896 : ChooseMulInstructionByType(destVal->getType()));
897 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
898 .addRegDef(destVal));
903 // Generate a divide instruction for Div or Rem.
904 // For Rem, this assumes that the operand type will be signed if the result
905 // type is signed. This is correct because they must have the same sign.
907 static inline MachineOpCode
908 ChooseDivInstruction(TargetMachine &target,
909 const InstructionNode* instrNode)
911 MachineOpCode opCode = V9::INVALID_OPCODE;
913 const Type* resultType = instrNode->getInstruction()->getType();
915 if (resultType->isInteger())
916 opCode = resultType->isSigned()? V9::SDIVXr : V9::UDIVXr;
918 switch(resultType->getPrimitiveID())
920 case Type::FloatTyID: opCode = V9::FDIVS; break;
921 case Type::DoubleTyID: opCode = V9::FDIVD; break;
922 default: assert(0 && "Invalid type for DIV instruction"); break;
929 // Return if we cannot exploit constant to create a cheaper instruction
931 CreateDivConstInstruction(TargetMachine &target,
932 const InstructionNode* instrNode,
933 std::vector<MachineInstr*>& mvec)
935 Value* LHS = instrNode->leftChild()->getValue();
936 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
937 if (!isa<Constant>(constOp))
940 Instruction* destVal = instrNode->getInstruction();
941 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
943 // Cases worth optimizing are:
944 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
945 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
947 const Type* resultType = instrNode->getInstruction()->getType();
949 if (resultType->isInteger()) {
952 int64_t C = GetConstantValueAsSignedInt(constOp, isValidConst);
954 bool needNeg = false;
961 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
962 .addRegDef(destVal));
963 } else if (isPowerOf2(C, pow)) {
967 if (resultType->isSigned()) {
968 // The result may be negative and we need to add one before shifting
969 // a negative value. Use:
970 // srl i0, 31, x0; add x0, i0, i1 (if i0 is <= 32 bits)
972 // srlx i0, 63, x0; add x0, i0, i1 (if i0 is 64 bits)
973 // to compute i1=i0+1 if i0 < 0 and i1=i0 otherwise.
975 TmpInstruction *srlTmp, *addTmp;
976 MachineCodeForInstruction& mcfi
977 = MachineCodeForInstruction::get(destVal);
978 srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
979 addTmp = new TmpInstruction(mcfi, resultType, LHS, srlTmp,"incIfNeg");
981 // Create the SRL or SRLX instruction to get the sign bit
982 mvec.push_back(BuildMI((resultType==Type::LongTy) ?
983 V9::SRLXi6 : V9::SRLi5, 3)
985 .addSImm((resultType==Type::LongTy)? 63 : 31)
988 // Create the ADD instruction to add 1 for negative values
989 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
992 // Get the shift operand and "right-shift" opcode to do the divide
993 shiftOperand = addTmp;
994 opCode = (resultType==Type::LongTy) ? V9::SRAXi6 : V9::SRAi5;
996 // Get the shift operand and "right-shift" opcode to do the divide
998 opCode = (resultType==Type::LongTy) ? V9::SRLXi6 : V9::SRLi5;
1001 // Now do the actual shift!
1002 mvec.push_back(BuildMI(opCode, 3).addReg(shiftOperand).addZImm(pow)
1003 .addRegDef(destVal));
1006 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
1007 // insert <reg = SUB 0, reg> after the instr to flip the sign
1008 mvec.push_back(CreateIntNegInstruction(target, destVal));
1012 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
1013 double dval = FPC->getValue();
1014 if (fabs(dval) == 1) {
1016 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
1017 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
1019 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(destVal));
1027 CreateCodeForVariableSizeAlloca(const TargetMachine& target,
1028 Instruction* result,
1030 Value* numElementsVal,
1031 std::vector<MachineInstr*>& getMvec)
1033 Value* totalSizeVal;
1035 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
1036 Function *F = result->getParent()->getParent();
1038 // Enforce the alignment constraints on the stack pointer at
1039 // compile time if the total size is a known constant.
1040 if (isa<Constant>(numElementsVal)) {
1042 int64_t numElem = GetConstantValueAsSignedInt(numElementsVal, isValid);
1043 assert(isValid && "Unexpectedly large array dimension in alloca!");
1044 int64_t total = numElem * tsize;
1045 if (int extra= total % target.getFrameInfo().getStackFrameSizeAlignment())
1046 total += target.getFrameInfo().getStackFrameSizeAlignment() - extra;
1047 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
1049 // The size is not a constant. Generate code to compute it and
1050 // code to pad the size for stack alignment.
1051 // Create a Value to hold the (constant) element size
1052 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
1054 // Create temporary values to hold the result of MUL, SLL, SRL
1055 // To pad `size' to next smallest multiple of 16:
1056 // size = (size + 15) & (-16 = 0xfffffffffffffff0)
1058 TmpInstruction* tmpProd = new TmpInstruction(mcfi,numElementsVal, tsizeVal);
1059 TmpInstruction* tmpAdd15= new TmpInstruction(mcfi,numElementsVal, tmpProd);
1060 TmpInstruction* tmpAndf0= new TmpInstruction(mcfi,numElementsVal, tmpAdd15);
1062 // Instruction 1: mul numElements, typeSize -> tmpProd
1063 // This will optimize the MUL as far as possible.
1064 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec,
1065 mcfi, INVALID_MACHINE_OPCODE);
1067 // Instruction 2: andn tmpProd, 0x0f -> tmpAndn
1068 getMvec.push_back(BuildMI(V9::ADDi, 3).addReg(tmpProd).addSImm(15)
1069 .addReg(tmpAdd15, MOTy::Def));
1071 // Instruction 3: add tmpAndn, 0x10 -> tmpAdd16
1072 getMvec.push_back(BuildMI(V9::ANDi, 3).addReg(tmpAdd15).addSImm(-16)
1073 .addReg(tmpAndf0, MOTy::Def));
1075 totalSizeVal = tmpAndf0;
1078 // Get the constant offset from SP for dynamically allocated storage
1079 // and create a temporary Value to hold it.
1080 MachineFunction& mcInfo = MachineFunction::get(F);
1082 ConstantSInt* dynamicAreaOffset =
1083 ConstantSInt::get(Type::IntTy,
1084 target.getFrameInfo().getDynamicAreaOffset(mcInfo,growUp));
1085 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
1087 unsigned SPReg = target.getRegInfo().getStackPointer();
1089 // Instruction 2: sub %sp, totalSizeVal -> %sp
1090 getMvec.push_back(BuildMI(V9::SUBr, 3).addMReg(SPReg).addReg(totalSizeVal)
1091 .addMReg(SPReg,MOTy::Def));
1093 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
1094 getMvec.push_back(BuildMI(V9::ADDr,3).addMReg(SPReg).addReg(dynamicAreaOffset)
1095 .addRegDef(result));
1100 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
1101 Instruction* result,
1103 unsigned numElements,
1104 std::vector<MachineInstr*>& getMvec)
1106 assert(tsize > 0 && "Illegal (zero) type size for alloca");
1107 assert(result && result->getParent() &&
1108 "Result value is not part of a function?");
1109 Function *F = result->getParent()->getParent();
1110 MachineFunction &mcInfo = MachineFunction::get(F);
1112 // Check if the offset would small enough to use as an immediate in
1113 // load/stores (check LDX because all load/stores have the same-size immediate
1114 // field). If not, put the variable in the dynamically sized area of the
1116 unsigned paddedSizeIgnored;
1117 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
1119 tsize * numElements);
1120 if (! target.getInstrInfo().constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
1121 CreateCodeForVariableSizeAlloca(target, result, tsize,
1122 ConstantSInt::get(Type::IntTy,numElements),
1127 // else offset fits in immediate field so go ahead and allocate it.
1128 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
1130 // Create a temporary Value to hold the constant offset.
1131 // This is needed because it may not fit in the immediate field.
1132 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
1134 // Instruction 1: add %fp, offsetFromFP -> result
1135 unsigned FPReg = target.getRegInfo().getFramePointer();
1136 getMvec.push_back(BuildMI(V9::ADDr, 3).addMReg(FPReg).addReg(offsetVal)
1137 .addRegDef(result));
1141 //------------------------------------------------------------------------
1142 // Function SetOperandsForMemInstr
1144 // Choose addressing mode for the given load or store instruction.
1145 // Use [reg+reg] if it is an indexed reference, and the index offset is
1146 // not a constant or if it cannot fit in the offset field.
1147 // Use [reg+offset] in all other cases.
1149 // This assumes that all array refs are "lowered" to one of these forms:
1150 // %x = load (subarray*) ptr, constant ; single constant offset
1151 // %x = load (subarray*) ptr, offsetVal ; single non-constant offset
1152 // Generally, this should happen via strength reduction + LICM.
1153 // Also, strength reduction should take care of using the same register for
1154 // the loop index variable and an array index, when that is profitable.
1155 //------------------------------------------------------------------------
1158 SetOperandsForMemInstr(unsigned Opcode,
1159 std::vector<MachineInstr*>& mvec,
1160 InstructionNode* vmInstrNode,
1161 const TargetMachine& target)
1163 Instruction* memInst = vmInstrNode->getInstruction();
1164 // Index vector, ptr value, and flag if all indices are const.
1165 std::vector<Value*> idxVec;
1166 bool allConstantIndices;
1167 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
1169 // Now create the appropriate operands for the machine instruction.
1170 // First, initialize so we default to storing the offset in a register.
1171 int64_t smallConstOffset = 0;
1172 Value* valueForRegOffset = NULL;
1173 MachineOperand::MachineOperandType offsetOpType =
1174 MachineOperand::MO_VirtualRegister;
1176 // Check if there is an index vector and if so, compute the
1177 // right offset for structures and for arrays
1179 if (!idxVec.empty()) {
1180 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
1182 // If all indices are constant, compute the combined offset directly.
1183 if (allConstantIndices) {
1184 // Compute the offset value using the index vector. Create a
1185 // virtual reg. for it since it may not fit in the immed field.
1186 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
1187 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
1189 // There is at least one non-constant offset. Therefore, this must
1190 // be an array ref, and must have been lowered to a single non-zero
1191 // offset. (An extra leading zero offset, if any, can be ignored.)
1192 // Generate code sequence to compute address from index.
1194 bool firstIdxIsZero = IsZero(idxVec[0]);
1195 assert(idxVec.size() == 1U + firstIdxIsZero
1196 && "Array refs must be lowered before Instruction Selection");
1198 Value* idxVal = idxVec[firstIdxIsZero];
1200 std::vector<MachineInstr*> mulVec;
1202 new TmpInstruction(MachineCodeForInstruction::get(memInst),
1203 Type::ULongTy, memInst);
1205 // Get the array type indexed by idxVal, and compute its element size.
1206 // The call to getTypeSize() will fail if size is not constant.
1207 const Type* vecType = (firstIdxIsZero
1208 ? GetElementPtrInst::getIndexedType(ptrType,
1209 std::vector<Value*>(1U, idxVec[0]),
1210 /*AllowCompositeLeaf*/ true)
1212 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
1213 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
1214 target.getTargetData().getTypeSize(eltType));
1216 // CreateMulInstruction() folds constants intelligently enough.
1217 CreateMulInstruction(target, memInst->getParent()->getParent(),
1218 idxVal, /* lval, not likely to be const*/
1219 eltSizeVal, /* rval, likely to be constant */
1221 mulVec, MachineCodeForInstruction::get(memInst),
1222 INVALID_MACHINE_OPCODE);
1224 assert(mulVec.size() > 0 && "No multiply code created?");
1225 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
1227 valueForRegOffset = addr;
1230 offsetOpType = MachineOperand::MO_SignExtendedImmed;
1231 smallConstOffset = 0;
1235 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
1236 // For LOAD or GET_ELEMENT_PTR,
1237 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
1239 unsigned offsetOpNum, ptrOpNum;
1241 if (memInst->getOpcode() == Instruction::Store) {
1242 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1243 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1244 .addReg(ptrVal).addReg(valueForRegOffset);
1246 Opcode = convertOpcodeFromRegToImm(Opcode);
1247 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
1248 .addReg(ptrVal).addSImm(smallConstOffset);
1251 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
1252 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
1253 .addRegDef(memInst);
1255 Opcode = convertOpcodeFromRegToImm(Opcode);
1256 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
1257 .addRegDef(memInst);
1265 // Substitute operand `operandNum' of the instruction in node `treeNode'
1266 // in place of the use(s) of that instruction in node `parent'.
1267 // Check both explicit and implicit operands!
1268 // Also make sure to skip over a parent who:
1269 // (1) is a list node in the Burg tree, or
1270 // (2) itself had its results forwarded to its parent
1273 ForwardOperand(InstructionNode* treeNode,
1274 InstrTreeNode* parent,
1277 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
1279 Instruction* unusedOp = treeNode->getInstruction();
1280 Value* fwdOp = unusedOp->getOperand(operandNum);
1282 // The parent itself may be a list node, so find the real parent instruction
1283 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode)
1285 parent = parent->parent();
1286 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
1288 InstructionNode* parentInstrNode = (InstructionNode*) parent;
1290 Instruction* userInstr = parentInstrNode->getInstruction();
1291 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
1293 // The parent's mvec would be empty if it was itself forwarded.
1294 // Recursively call ForwardOperand in that case...
1296 if (mvec.size() == 0) {
1297 assert(parent->parent() != NULL &&
1298 "Parent could not have been forwarded, yet has no instructions?");
1299 ForwardOperand(treeNode, parent->parent(), operandNum);
1301 for (unsigned i=0, N=mvec.size(); i < N; i++) {
1302 MachineInstr* minstr = mvec[i];
1303 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) {
1304 const MachineOperand& mop = minstr->getOperand(i);
1305 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
1306 mop.getVRegValue() == unusedOp)
1308 minstr->SetMachineOperandVal(i, MachineOperand::MO_VirtualRegister,
1313 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
1314 if (minstr->getImplicitRef(i) == unusedOp) {
1315 minstr->setImplicitRef(i, fwdOp,
1316 minstr->getImplicitOp(i).opIsDefOnly(),
1317 minstr->getImplicitOp(i).opIsDefAndUse());
1325 AllUsesAreBranches(const Instruction* setccI)
1327 for (Value::use_const_iterator UI=setccI->use_begin(), UE=setccI->use_end();
1329 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
1330 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
1335 // Generate code for any intrinsic that needs a special code sequence
1336 // instead of a regular call. If not that kind of intrinsic, do nothing.
1337 // Returns true if code was generated, otherwise false.
1339 bool CodeGenIntrinsic(LLVMIntrinsic::ID iid, CallInst &callInstr,
1340 TargetMachine &target,
1341 std::vector<MachineInstr*>& mvec)
1344 case LLVMIntrinsic::va_start: {
1345 // Get the address of the first vararg value on stack and copy it to
1346 // the argument of va_start(va_list* ap).
1348 Function* func = cast<Function>(callInstr.getParent()->getParent());
1349 int numFixedArgs = func->getFunctionType()->getNumParams();
1350 int fpReg = target.getFrameInfo().getIncomingArgBaseRegNum();
1351 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
1352 int firstVarArgOff = numFixedArgs * argSize + target.getFrameInfo().
1353 getFirstIncomingArgOffset(MachineFunction::get(func), ignore);
1354 mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
1355 addReg(callInstr.getOperand(1)));
1359 case LLVMIntrinsic::va_end:
1360 return true; // no-op on Sparc
1362 case LLVMIntrinsic::va_copy:
1363 // Simple copy of current va_list (arg2) to new va_list (arg1)
1364 mvec.push_back(BuildMI(V9::ORr, 3).
1365 addMReg(target.getRegInfo().getZeroRegNum()).
1366 addReg(callInstr.getOperand(2)).
1367 addReg(callInstr.getOperand(1)));
1375 //******************* Externally Visible Functions *************************/
1377 //------------------------------------------------------------------------
1378 // External Function: ThisIsAChainRule
1381 // Check if a given BURG rule is a chain rule.
1382 //------------------------------------------------------------------------
1385 ThisIsAChainRule(int eruleno)
1389 case 111: // stmt: reg
1413 return false; break;
1418 //------------------------------------------------------------------------
1419 // External Function: GetInstructionsByRule
1422 // Choose machine instructions for the SPARC according to the
1423 // patterns chosen by the BURG-generated parser.
1424 //------------------------------------------------------------------------
1427 GetInstructionsByRule(InstructionNode* subtreeRoot,
1430 TargetMachine &target,
1431 std::vector<MachineInstr*>& mvec)
1433 bool checkCast = false; // initialize here to use fall-through
1434 bool maskUnsignedResult = false;
1436 int forwardOperandNum = -1;
1437 unsigned allocaSize = 0;
1438 MachineInstr* M, *M2;
1443 // If the code for this instruction was folded into the parent (user),
1445 if (subtreeRoot->isFoldedIntoParent())
1449 // Let's check for chain rules outside the switch so that we don't have
1450 // to duplicate the list of chain rule production numbers here again
1452 if (ThisIsAChainRule(ruleForNode))
1454 // Chain rules have a single nonterminal on the RHS.
1455 // Get the rule that matches the RHS non-terminal and use that instead.
1457 assert(nts[0] && ! nts[1]
1458 && "A chain rule should have only one RHS non-terminal!");
1459 nextRule = burm_rule(subtreeRoot->state, nts[0]);
1460 nts = burm_nts[nextRule];
1461 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
1465 switch(ruleForNode) {
1466 case 1: // stmt: Ret
1467 case 2: // stmt: RetValue(reg)
1468 { // NOTE: Prepass of register allocation is responsible
1469 // for moving return value to appropriate register.
1470 // Copy the return value to the required return register.
1471 // Mark the return Value as an implicit ref of the RET instr..
1472 // Mark the return-address register as a hidden virtual reg.
1473 // Finally put a NOP in the delay slot.
1474 ReturnInst *returnInstr=cast<ReturnInst>(subtreeRoot->getInstruction());
1475 Value* retVal = returnInstr->getReturnValue();
1476 MachineCodeForInstruction& mcfi =
1477 MachineCodeForInstruction::get(returnInstr);
1479 // Create a hidden virtual reg to represent the return address register
1480 // used by the machine instruction but not represented in LLVM.
1482 Instruction* returnAddrTmp = new TmpInstruction(mcfi, returnInstr);
1484 MachineInstr* retMI =
1485 BuildMI(V9::JMPLRETi, 3).addReg(returnAddrTmp).addSImm(8)
1486 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def);
1488 // Insert a copy to copy the return value to the appropriate register
1489 // -- For FP values, create a FMOVS or FMOVD instruction
1490 // -- For non-FP values, create an add-with-0 instruction
1492 if (retVal != NULL) {
1493 const UltraSparcRegInfo& regInfo =
1494 (UltraSparcRegInfo&) target.getRegInfo();
1495 const Type* retType = retVal->getType();
1496 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
1497 unsigned retRegNum = (retType->isFloatingPoint()
1498 ? (unsigned) SparcFloatRegClass::f0
1499 : (unsigned) SparcIntRegClass::i0);
1500 retRegNum = regInfo.getUnifiedRegNum(regClassID, retRegNum);
1502 // Create a virtual register to represent it and mark
1503 // this vreg as being an implicit operand of the ret MI
1504 TmpInstruction* retVReg =
1505 new TmpInstruction(mcfi, retVal, NULL, "argReg");
1507 retMI->addImplicitRef(retVReg);
1509 if (retType->isFloatingPoint())
1510 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
1511 .addReg(retVal).addReg(retVReg, MOTy::Def));
1513 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
1514 .addReg(retVal).addSImm((int64_t) 0)
1515 .addReg(retVReg, MOTy::Def));
1517 // Mark the operand with the register it should be assigned
1518 M->SetRegForOperand(M->getNumOperands()-1, retRegNum);
1519 retMI->SetRegForImplicitRef(retMI->getNumImplicitRefs()-1, retRegNum);
1524 // Now insert the RET instruction and a NOP for the delay slot
1525 mvec.push_back(retMI);
1526 mvec.push_back(BuildMI(V9::NOP, 0));
1531 case 3: // stmt: Store(reg,reg)
1532 case 4: // stmt: Store(reg,ptrreg)
1533 SetOperandsForMemInstr(ChooseStoreInstruction(
1534 subtreeRoot->leftChild()->getValue()->getType()),
1535 mvec, subtreeRoot, target);
1538 case 5: // stmt: BrUncond
1540 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1541 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
1544 mvec.push_back(BuildMI(V9::NOP, 0));
1548 case 206: // stmt: BrCond(setCCconst)
1549 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
1550 // If the constant is ZERO, we can use the branch-on-integer-register
1551 // instructions and avoid the SUBcc instruction entirely.
1552 // Otherwise this is just the same as case 5, so just fall through.
1554 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
1556 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
1557 Constant *constVal = cast<Constant>(constNode->getValue());
1560 if ((constVal->getType()->isInteger()
1561 || isa<PointerType>(constVal->getType()))
1562 && GetConstantValueAsSignedInt(constVal, isValidConst) == 0
1565 // That constant is a zero after all...
1566 // Use the left child of setCC as the first argument!
1567 // Mark the setCC node so that no code is generated for it.
1568 InstructionNode* setCCNode = (InstructionNode*)
1569 subtreeRoot->leftChild();
1570 assert(setCCNode->getOpLabel() == SetCCOp);
1571 setCCNode->markFoldedIntoParent();
1573 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
1575 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
1576 .addReg(setCCNode->leftChild()->getValue())
1577 .addPCDisp(brInst->getSuccessor(0));
1581 mvec.push_back(BuildMI(V9::NOP, 0));
1584 mvec.push_back(BuildMI(V9::BA, 1)
1585 .addPCDisp(brInst->getSuccessor(1)));
1588 mvec.push_back(BuildMI(V9::NOP, 0));
1591 // ELSE FALL THROUGH
1594 case 6: // stmt: BrCond(setCC)
1595 { // bool => boolean was computed with SetCC.
1596 // The branch to use depends on whether it is FP, signed, or unsigned.
1597 // If it is an integer CC, we also need to find the unique
1598 // TmpInstruction representing that CC.
1600 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
1602 unsigned Opcode = ChooseBccInstruction(subtreeRoot, isFPBranch);
1603 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
1604 brInst->getParent()->getParent(),
1605 isFPBranch? Type::FloatTy : Type::IntTy,
1606 MachineCodeForInstruction::get(brInst));
1607 M = BuildMI(Opcode, 2).addCCReg(ccValue)
1608 .addPCDisp(brInst->getSuccessor(0));
1612 mvec.push_back(BuildMI(V9::NOP, 0));
1615 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
1618 mvec.push_back(BuildMI(V9::NOP, 0));
1622 case 208: // stmt: BrCond(boolconst)
1624 // boolconst => boolean is a constant; use BA to first or second label
1625 Constant* constVal =
1626 cast<Constant>(subtreeRoot->leftChild()->getValue());
1627 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
1629 M = BuildMI(V9::BA, 1).addPCDisp(
1630 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
1634 mvec.push_back(BuildMI(V9::NOP, 0));
1638 case 8: // stmt: BrCond(boolreg)
1639 { // boolreg => boolean is recorded in an integer register.
1640 // Use branch-on-integer-register instruction.
1642 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
1643 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
1644 .addPCDisp(BI->getSuccessor(0));
1648 mvec.push_back(BuildMI(V9::NOP, 0));
1651 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
1654 mvec.push_back(BuildMI(V9::NOP, 0));
1658 case 9: // stmt: Switch(reg)
1659 assert(0 && "*** SWITCH instruction is not implemented yet.");
1662 case 10: // reg: VRegList(reg, reg)
1663 assert(0 && "VRegList should never be the topmost non-chain rule");
1666 case 21: // bool: Not(bool,reg): Both these are implemented as:
1667 case 421: // reg: BNot(reg,reg): reg = reg XOR-NOT 0
1668 { // First find the unary operand. It may be left or right, usually right.
1669 Value* notArg = BinaryOperator::getNotArgument(
1670 cast<BinaryOperator>(subtreeRoot->getInstruction()));
1671 unsigned ZeroReg = target.getRegInfo().getZeroRegNum();
1672 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(notArg).addMReg(ZeroReg)
1673 .addRegDef(subtreeRoot->getValue()));
1677 case 22: // reg: ToBoolTy(reg):
1679 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
1680 assert(opType->isIntegral() || isa<PointerType>(opType));
1681 forwardOperandNum = 0; // forward first operand to user
1685 case 23: // reg: ToUByteTy(reg)
1686 case 24: // reg: ToSByteTy(reg)
1687 case 25: // reg: ToUShortTy(reg)
1688 case 26: // reg: ToShortTy(reg)
1689 case 27: // reg: ToUIntTy(reg)
1690 case 28: // reg: ToIntTy(reg)
1692 //======================================================================
1693 // Rules for integer conversions:
1696 // From ISO 1998 C++ Standard, Sec. 4.7:
1698 // 2. If the destination type is unsigned, the resulting value is
1699 // the least unsigned integer congruent to the source integer
1700 // (modulo 2n where n is the number of bits used to represent the
1701 // unsigned type). [Note: In a two s complement representation,
1702 // this conversion is conceptual and there is no change in the
1703 // bit pattern (if there is no truncation). ]
1705 // 3. If the destination type is signed, the value is unchanged if
1706 // it can be represented in the destination type (and bitfield width);
1707 // otherwise, the value is implementation-defined.
1710 // Since we assume 2s complement representations, this implies:
1712 // -- if operand is smaller than destination, zero-extend or sign-extend
1713 // according to the signedness of the *operand*: source decides.
1714 // ==> we have to do nothing here!
1716 // -- if operand is same size as or larger than destination, and the
1717 // destination is *unsigned*, zero-extend the operand: dest. decides
1719 // -- if operand is same size as or larger than destination, and the
1720 // destination is *signed*, the choice is implementation defined:
1721 // we sign-extend the operand: i.e., again dest. decides.
1722 // Note: this matches both Sun's cc and gcc3.2.
1723 //======================================================================
1725 Instruction* destI = subtreeRoot->getInstruction();
1726 Value* opVal = subtreeRoot->leftChild()->getValue();
1727 const Type* opType = opVal->getType();
1728 if (opType->isIntegral() || isa<PointerType>(opType)) {
1729 unsigned opSize = target.getTargetData().getTypeSize(opType);
1731 target.getTargetData().getTypeSize(destI->getType());
1732 if (opSize >= destSize) {
1733 // Operand is same size as or larger than dest:
1734 // zero- or sign-extend, according to the signeddness of
1735 // the destination (see above).
1736 if (destI->getType()->isSigned())
1737 target.getInstrInfo().CreateSignExtensionInstructions(target,
1738 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1739 mvec, MachineCodeForInstruction::get(destI));
1741 target.getInstrInfo().CreateZeroExtensionInstructions(target,
1742 destI->getParent()->getParent(), opVal, destI, 8*destSize,
1743 mvec, MachineCodeForInstruction::get(destI));
1745 forwardOperandNum = 0; // forward first operand to user
1746 } else if (opType->isFloatingPoint()) {
1747 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1748 MachineCodeForInstruction::get(destI));
1749 if (destI->getType()->isUnsigned())
1750 maskUnsignedResult = true; // not handled by fp->int code
1752 assert(0 && "Unrecognized operand type for convert-to-unsigned");
1757 case 29: // reg: ToULongTy(reg)
1758 case 30: // reg: ToLongTy(reg)
1760 Value* opVal = subtreeRoot->leftChild()->getValue();
1761 const Type* opType = opVal->getType();
1762 if (opType->isIntegral() || isa<PointerType>(opType))
1763 forwardOperandNum = 0; // forward first operand to user
1764 else if (opType->isFloatingPoint()) {
1765 Instruction* destI = subtreeRoot->getInstruction();
1766 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec,
1767 MachineCodeForInstruction::get(destI));
1769 assert(0 && "Unrecognized operand type for convert-to-signed");
1773 case 31: // reg: ToFloatTy(reg):
1774 case 32: // reg: ToDoubleTy(reg):
1775 case 232: // reg: ToDoubleTy(Constant):
1777 // If this instruction has a parent (a user) in the tree
1778 // and the user is translated as an FsMULd instruction,
1779 // then the cast is unnecessary. So check that first.
1780 // In the future, we'll want to do the same for the FdMULq instruction,
1781 // so do the check here instead of only for ToFloatTy(reg).
1783 if (subtreeRoot->parent() != NULL) {
1784 const MachineCodeForInstruction& mcfi =
1785 MachineCodeForInstruction::get(
1786 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
1787 if (mcfi.size() == 0 || mcfi.front()->getOpCode() == V9::FSMULD)
1788 forwardOperandNum = 0; // forward first operand to user
1791 if (forwardOperandNum != 0) { // we do need the cast
1792 Value* leftVal = subtreeRoot->leftChild()->getValue();
1793 const Type* opType = leftVal->getType();
1794 MachineOpCode opCode=ChooseConvertToFloatInstr(
1795 subtreeRoot->getOpLabel(), opType);
1796 if (opCode == V9::INVALID_OPCODE) { // no conversion needed
1797 forwardOperandNum = 0; // forward first operand to user
1799 // If the source operand is a non-FP type it must be
1800 // first copied from int to float register via memory!
1801 Instruction *dest = subtreeRoot->getInstruction();
1804 if (! opType->isFloatingPoint()) {
1805 // Create a temporary to represent the FP register
1806 // into which the integer will be copied via memory.
1807 // The type of this temporary will determine the FP
1808 // register used: single-prec for a 32-bit int or smaller,
1809 // double-prec for a 64-bit int.
1812 target.getTargetData().getTypeSize(leftVal->getType());
1813 Type* tmpTypeToUse =
1814 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
1815 MachineCodeForInstruction &destMCFI =
1816 MachineCodeForInstruction::get(dest);
1817 srcForCast = new TmpInstruction(destMCFI, tmpTypeToUse, dest);
1819 target.getInstrInfo().CreateCodeToCopyIntToFloat(target,
1820 dest->getParent()->getParent(),
1821 leftVal, cast<Instruction>(srcForCast),
1824 srcForCast = leftVal;
1826 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
1832 case 19: // reg: ToArrayTy(reg):
1833 case 20: // reg: ToPointerTy(reg):
1834 forwardOperandNum = 0; // forward first operand to user
1837 case 233: // reg: Add(reg, Constant)
1838 maskUnsignedResult = true;
1839 M = CreateAddConstInstruction(subtreeRoot);
1844 // ELSE FALL THROUGH
1846 case 33: // reg: Add(reg, reg)
1847 maskUnsignedResult = true;
1848 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
1851 case 234: // reg: Sub(reg, Constant)
1852 maskUnsignedResult = true;
1853 M = CreateSubConstInstruction(subtreeRoot);
1858 // ELSE FALL THROUGH
1860 case 34: // reg: Sub(reg, reg)
1861 maskUnsignedResult = true;
1862 Add3OperandInstr(ChooseSubInstructionByType(
1863 subtreeRoot->getInstruction()->getType()),
1867 case 135: // reg: Mul(todouble, todouble)
1871 case 35: // reg: Mul(reg, reg)
1873 maskUnsignedResult = true;
1874 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1876 : INVALID_MACHINE_OPCODE);
1877 Instruction* mulInstr = subtreeRoot->getInstruction();
1878 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1879 subtreeRoot->leftChild()->getValue(),
1880 subtreeRoot->rightChild()->getValue(),
1882 MachineCodeForInstruction::get(mulInstr),forceOp);
1885 case 335: // reg: Mul(todouble, todoubleConst)
1889 case 235: // reg: Mul(reg, Constant)
1891 maskUnsignedResult = true;
1892 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
1894 : INVALID_MACHINE_OPCODE);
1895 Instruction* mulInstr = subtreeRoot->getInstruction();
1896 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
1897 subtreeRoot->leftChild()->getValue(),
1898 subtreeRoot->rightChild()->getValue(),
1900 MachineCodeForInstruction::get(mulInstr),
1904 case 236: // reg: Div(reg, Constant)
1905 maskUnsignedResult = true;
1907 CreateDivConstInstruction(target, subtreeRoot, mvec);
1908 if (mvec.size() > L)
1910 // ELSE FALL THROUGH
1912 case 36: // reg: Div(reg, reg)
1913 maskUnsignedResult = true;
1914 Add3OperandInstr(ChooseDivInstruction(target, subtreeRoot),
1918 case 37: // reg: Rem(reg, reg)
1919 case 237: // reg: Rem(reg, Constant)
1921 maskUnsignedResult = true;
1922 Instruction* remInstr = subtreeRoot->getInstruction();
1924 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(remInstr);
1925 TmpInstruction* quot = new TmpInstruction(mcfi,
1926 subtreeRoot->leftChild()->getValue(),
1927 subtreeRoot->rightChild()->getValue());
1928 TmpInstruction* prod = new TmpInstruction(mcfi,
1930 subtreeRoot->rightChild()->getValue());
1932 M = BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
1933 .addReg(subtreeRoot->leftChild()->getValue())
1934 .addReg(subtreeRoot->rightChild()->getValue())
1938 unsigned MulOpcode =
1939 ChooseMulInstructionByType(subtreeRoot->getInstruction()->getType());
1940 Value *MulRHS = subtreeRoot->rightChild()->getValue();
1941 M = BuildMI(MulOpcode, 3).addReg(quot).addReg(MulRHS).addReg(prod,
1945 unsigned Opcode = ChooseSubInstructionByType(
1946 subtreeRoot->getInstruction()->getType());
1947 M = BuildMI(Opcode, 3).addReg(subtreeRoot->leftChild()->getValue())
1948 .addReg(prod).addRegDef(subtreeRoot->getValue());
1953 case 38: // bool: And(bool, bool)
1954 case 238: // bool: And(bool, boolconst)
1955 case 338: // reg : BAnd(reg, reg)
1956 case 538: // reg : BAnd(reg, Constant)
1957 Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
1960 case 138: // bool: And(bool, not)
1961 case 438: // bool: BAnd(bool, bnot)
1962 { // Use the argument of NOT as the second argument!
1963 // Mark the NOT node so that no code is generated for it.
1964 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1965 Value* notArg = BinaryOperator::getNotArgument(
1966 cast<BinaryOperator>(notNode->getInstruction()));
1967 notNode->markFoldedIntoParent();
1968 Value *LHS = subtreeRoot->leftChild()->getValue();
1969 Value *Dest = subtreeRoot->getValue();
1970 mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(LHS).addReg(notArg)
1971 .addReg(Dest, MOTy::Def));
1975 case 39: // bool: Or(bool, bool)
1976 case 239: // bool: Or(bool, boolconst)
1977 case 339: // reg : BOr(reg, reg)
1978 case 539: // reg : BOr(reg, Constant)
1979 Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
1982 case 139: // bool: Or(bool, not)
1983 case 439: // bool: BOr(bool, bnot)
1984 { // Use the argument of NOT as the second argument!
1985 // Mark the NOT node so that no code is generated for it.
1986 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
1987 Value* notArg = BinaryOperator::getNotArgument(
1988 cast<BinaryOperator>(notNode->getInstruction()));
1989 notNode->markFoldedIntoParent();
1990 Value *LHS = subtreeRoot->leftChild()->getValue();
1991 Value *Dest = subtreeRoot->getValue();
1992 mvec.push_back(BuildMI(V9::ORNr, 3).addReg(LHS).addReg(notArg)
1993 .addReg(Dest, MOTy::Def));
1997 case 40: // bool: Xor(bool, bool)
1998 case 240: // bool: Xor(bool, boolconst)
1999 case 340: // reg : BXor(reg, reg)
2000 case 540: // reg : BXor(reg, Constant)
2001 Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
2004 case 140: // bool: Xor(bool, not)
2005 case 440: // bool: BXor(bool, bnot)
2006 { // Use the argument of NOT as the second argument!
2007 // Mark the NOT node so that no code is generated for it.
2008 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
2009 Value* notArg = BinaryOperator::getNotArgument(
2010 cast<BinaryOperator>(notNode->getInstruction()));
2011 notNode->markFoldedIntoParent();
2012 Value *LHS = subtreeRoot->leftChild()->getValue();
2013 Value *Dest = subtreeRoot->getValue();
2014 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(LHS).addReg(notArg)
2015 .addReg(Dest, MOTy::Def));
2019 case 41: // boolconst: SetCC(reg, Constant)
2021 // If the SetCC was folded into the user (parent), it will be
2022 // caught above. All other cases are the same as case 42,
2023 // so just fall through.
2025 case 42: // bool: SetCC(reg, reg):
2027 // This generates a SUBCC instruction, putting the difference in a
2028 // result reg. if needed, and/or setting a condition code if needed.
2030 Instruction* setCCInstr = subtreeRoot->getInstruction();
2031 Value* leftVal = subtreeRoot->leftChild()->getValue();
2032 bool isFPCompare = leftVal->getType()->isFloatingPoint();
2034 // If the boolean result of the SetCC is used outside the current basic
2035 // block (so it must be computed as a boolreg) or is used by anything
2036 // other than a branch, the boolean must be computed and stored
2037 // in a result register. We will use a conditional move to do this.
2039 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
2040 ! AllUsesAreBranches(setCCInstr));
2042 // A TmpInstruction is created to represent the CC "result".
2043 // Unlike other instances of TmpInstruction, this one is used
2044 // by machine code of multiple LLVM instructions, viz.,
2045 // the SetCC and the branch. Make sure to get the same one!
2046 // Note that we do this even for FP CC registers even though they
2047 // are explicit operands, because the type of the operand
2048 // needs to be a floating point condition code, not an integer
2049 // condition code. Think of this as casting the bool result to
2050 // a FP condition code register.
2051 // Later, we mark the 4th operand as being a CC register, and as a def.
2053 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
2054 setCCInstr->getParent()->getParent(),
2055 isFPCompare ? Type::FloatTy : Type::IntTy,
2056 MachineCodeForInstruction::get(setCCInstr));
2057 if (! isFPCompare) {
2058 // Integer condition: set CC and discard result.
2059 M = BuildMI(V9::SUBccr, 4)
2060 .addReg(subtreeRoot->leftChild()->getValue())
2061 .addReg(subtreeRoot->rightChild()->getValue())
2062 .addMReg(target.getRegInfo().getZeroRegNum(), MOTy::Def)
2063 .addCCReg(tmpForCC, MOTy::Def);
2065 // FP condition: dest of FCMP should be some FCCn register
2066 M = BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
2067 .addCCReg(tmpForCC, MOTy::Def)
2068 .addReg(subtreeRoot->leftChild()->getValue())
2069 .addReg(subtreeRoot->rightChild()->getValue());
2073 if (computeBoolVal) {
2074 MachineOpCode movOpCode = (isFPCompare
2075 ? ChooseMovFpcciInstruction(subtreeRoot)
2076 : ChooseMovpcciAfterSub(subtreeRoot));
2078 // Unconditionally set register to 0
2079 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
2082 // Now conditionally move 1 into the register.
2083 // Mark the register as a use (as well as a def) because the old
2084 // value will be retained if the condition is false.
2085 M = (BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(1)
2086 .addReg(setCCInstr, MOTy::UseAndDef));
2092 case 51: // reg: Load(reg)
2093 case 52: // reg: Load(ptrreg)
2094 SetOperandsForMemInstr(ChooseLoadInstruction(
2095 subtreeRoot->getValue()->getType()),
2096 mvec, subtreeRoot, target);
2099 case 55: // reg: GetElemPtr(reg)
2100 case 56: // reg: GetElemPtrIdx(reg,reg)
2101 // If the GetElemPtr was folded into the user (parent), it will be
2102 // caught above. For other cases, we have to compute the address.
2103 SetOperandsForMemInstr(V9::ADDr, mvec, subtreeRoot, target);
2106 case 57: // reg: Alloca: Implement as 1 instruction:
2107 { // add %fp, offsetFromFP -> result
2108 AllocationInst* instr =
2109 cast<AllocationInst>(subtreeRoot->getInstruction());
2111 target.getTargetData().getTypeSize(instr->getAllocatedType());
2113 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
2117 case 58: // reg: Alloca(reg): Implement as 3 instructions:
2118 // mul num, typeSz -> tmp
2119 // sub %sp, tmp -> %sp
2120 { // add %sp, frameSizeBelowDynamicArea -> result
2121 AllocationInst* instr =
2122 cast<AllocationInst>(subtreeRoot->getInstruction());
2123 const Type* eltType = instr->getAllocatedType();
2125 // If #elements is constant, use simpler code for fixed-size allocas
2126 int tsize = (int) target.getTargetData().getTypeSize(eltType);
2127 Value* numElementsVal = NULL;
2128 bool isArray = instr->isArrayAllocation();
2130 if (!isArray || isa<Constant>(numElementsVal = instr->getArraySize())) {
2131 // total size is constant: generate code for fixed-size alloca
2132 unsigned numElements = isArray?
2133 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
2134 CreateCodeForFixedSizeAlloca(target, instr, tsize,
2137 // total size is not constant.
2138 CreateCodeForVariableSizeAlloca(target, instr, tsize,
2139 numElementsVal, mvec);
2144 case 61: // reg: Call
2145 { // Generate a direct (CALL) or indirect (JMPL) call.
2146 // Mark the return-address register, the indirection
2147 // register (for indirect calls), the operands of the Call,
2148 // and the return value (if any) as implicit operands
2149 // of the machine instruction.
2151 // If this is a varargs function, floating point arguments
2152 // have to passed in integer registers so insert
2153 // copy-float-to-int instructions for each float operand.
2155 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
2156 Value *callee = callInstr->getCalledValue();
2157 Function* calledFunc = dyn_cast<Function>(callee);
2159 // Check if this is an intrinsic function that needs a special code
2160 // sequence (e.g., va_start). Indirect calls cannot be special.
2162 bool specialIntrinsic = false;
2163 LLVMIntrinsic::ID iid;
2164 if (calledFunc && (iid=(LLVMIntrinsic::ID)calledFunc->getIntrinsicID()))
2165 specialIntrinsic = CodeGenIntrinsic(iid, *callInstr, target, mvec);
2167 // If not, generate the normal call sequence for the function.
2168 // This can also handle any intrinsics that are just function calls.
2170 if (! specialIntrinsic) {
2171 MachineFunction& MF =
2172 MachineFunction::get(callInstr->getParent()->getParent());
2173 MachineCodeForInstruction& mcfi =
2174 MachineCodeForInstruction::get(callInstr);
2175 const UltraSparcRegInfo& regInfo =
2176 (UltraSparcRegInfo&) target.getRegInfo();
2177 const TargetFrameInfo& frameInfo = target.getFrameInfo();
2179 // Create hidden virtual register for return address with type void*
2180 TmpInstruction* retAddrReg =
2181 new TmpInstruction(mcfi, PointerType::get(Type::VoidTy), callInstr);
2183 // Generate the machine instruction and its operands.
2184 // Use CALL for direct function calls; this optimistically assumes
2185 // the PC-relative address fits in the CALL address field (22 bits).
2186 // Use JMPL for indirect calls.
2187 // This will be added to mvec later, after operand copies.
2189 MachineInstr* callMI;
2190 if (calledFunc) // direct function call
2191 callMI = BuildMI(V9::CALL, 1).addPCDisp(callee);
2192 else // indirect function call
2193 callMI = (BuildMI(V9::JMPLCALLi,3).addReg(callee)
2194 .addSImm((int64_t)0).addRegDef(retAddrReg));
2196 const FunctionType* funcType =
2197 cast<FunctionType>(cast<PointerType>(callee->getType())
2198 ->getElementType());
2199 bool isVarArgs = funcType->isVarArg();
2200 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
2202 // Use a descriptor to pass information about call arguments
2203 // to the register allocator. This descriptor will be "owned"
2204 // and freed automatically when the MachineCodeForInstruction
2205 // object for the callInstr goes away.
2206 CallArgsDescriptor* argDesc =
2207 new CallArgsDescriptor(callInstr, retAddrReg,isVarArgs,noPrototype);
2208 assert(callInstr->getOperand(0) == callee
2209 && "This is assumed in the loop below!");
2211 // Insert copy instructions to get all the arguments into
2212 // all the places that they need to be.
2214 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
2216 Value* argVal = callInstr->getOperand(i);
2217 const Type* argType = argVal->getType();
2218 unsigned regType = regInfo.getRegType(argType);
2219 unsigned argSize = target.getTargetData().getTypeSize(argType);
2220 int regNumForArg = TargetRegInfo::getInvalidRegNum();
2221 unsigned regClassIDOfArgReg;
2222 CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
2224 // Check for FP arguments to varargs functions.
2225 // Any such argument in the first $K$ args must be passed in an
2226 // integer register. If there is no prototype, it must also
2227 // be passed as an FP register.
2228 // K = #integer argument registers.
2229 bool isFPArg = argVal->getType()->isFloatingPoint();
2230 if (isVarArgs && isFPArg) {
2231 // If it is a function with no prototype, pass value
2232 // as an FP value as well as a varargs value
2234 argInfo.setUseFPArgReg();
2236 // If this arg. is in the first $K$ regs, add copy-
2237 // float-to-int instructions to pass the value as an int.
2238 // To check if it is in teh first $K$, get the register
2239 // number for the arg #i.
2240 int copyRegNum = regInfo.regNumForIntArg(false, false, argNo,
2241 regClassIDOfArgReg);
2242 if (copyRegNum != regInfo.getInvalidRegNum()) {
2243 // Create a virtual register to represent copyReg. Mark
2244 // this vreg as being an implicit operand of the call MI
2245 const Type* loadTy = (argType == Type::FloatTy
2246 ? Type::IntTy : Type::LongTy);
2247 TmpInstruction* argVReg = new TmpInstruction(mcfi, loadTy,
2250 callMI->addImplicitRef(argVReg);
2252 // Get a temp stack location to use to copy
2253 // float-to-int via the stack.
2255 // FIXME: For now, we allocate permanent space because
2256 // the stack frame manager does not allow locals to be
2257 // allocated (e.g., for alloca) after a temp is
2260 // int tmpOffset = MF.getInfo()->pushTempValue(argSize);
2261 int tmpOffset = MF.getInfo()->allocateLocalVar(argVReg);
2263 // Generate the store from FP reg to stack
2264 unsigned StoreOpcode = ChooseStoreInstruction(argType);
2265 M = BuildMI(convertOpcodeFromRegToImm(StoreOpcode), 3)
2266 .addReg(argVal).addMReg(regInfo.getFramePointer())
2267 .addSImm(tmpOffset);
2270 // Generate the load from stack to int arg reg
2271 unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
2272 M = BuildMI(convertOpcodeFromRegToImm(LoadOpcode), 3)
2273 .addMReg(regInfo.getFramePointer()).addSImm(tmpOffset)
2274 .addReg(argVReg, MOTy::Def);
2276 // Mark operand with register it should be assigned
2277 // both for copy and for the callMI
2278 M->SetRegForOperand(M->getNumOperands()-1, copyRegNum);
2279 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2283 // Add info about the argument to the CallArgsDescriptor
2284 argInfo.setUseIntArgReg();
2285 argInfo.setArgCopy(copyRegNum);
2287 // Cannot fit in first $K$ regs so pass arg on stack
2288 argInfo.setUseStackSlot();
2290 } else if (isFPArg) {
2291 // Get the outgoing arg reg to see if there is one.
2292 regNumForArg = regInfo.regNumForFPArg(regType, false, false,
2293 argNo, regClassIDOfArgReg);
2294 if (regNumForArg == regInfo.getInvalidRegNum())
2295 argInfo.setUseStackSlot();
2297 argInfo.setUseFPArgReg();
2298 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2302 // Get the outgoing arg reg to see if there is one.
2303 regNumForArg = regInfo.regNumForIntArg(false,false,
2304 argNo, regClassIDOfArgReg);
2305 if (regNumForArg == regInfo.getInvalidRegNum())
2306 argInfo.setUseStackSlot();
2308 argInfo.setUseIntArgReg();
2309 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
2315 // Now insert copy instructions to stack slot or arg. register
2317 if (argInfo.usesStackSlot()) {
2318 // Get the stack offset for this argument slot.
2319 // FP args on stack are right justified so adjust offset!
2320 // int arguments are also right justified but they are
2321 // always loaded as a full double-word so the offset does
2322 // not need to be adjusted.
2323 int argOffset = frameInfo.getOutgoingArgOffset(MF, argNo);
2324 if (argType->isFloatingPoint()) {
2325 unsigned slotSize = frameInfo.getSizeOfEachArgOnStack();
2326 assert(argSize <= slotSize && "Insufficient slot size!");
2327 argOffset += slotSize - argSize;
2330 // Now generate instruction to copy argument to stack
2331 MachineOpCode storeOpCode =
2332 (argType->isFloatingPoint()
2333 ? ((argSize == 4)? V9::STFi : V9::STDFi) : V9::STXi);
2335 M = BuildMI(storeOpCode, 3).addReg(argVal)
2336 .addMReg(regInfo.getStackPointer()).addSImm(argOffset);
2339 // Create a virtual register to represent the arg reg. Mark
2340 // this vreg as being an implicit operand of the call MI.
2341 TmpInstruction* argVReg =
2342 new TmpInstruction(mcfi, argVal, NULL, "argReg");
2344 callMI->addImplicitRef(argVReg);
2346 // Generate the reg-to-reg copy into the outgoing arg reg.
2347 // -- For FP values, create a FMOVS or FMOVD instruction
2348 // -- For non-FP values, create an add-with-0 instruction
2349 if (argType->isFloatingPoint())
2350 M=(BuildMI(argType==Type::FloatTy? V9::FMOVS :V9::FMOVD,2)
2351 .addReg(argVal).addReg(argVReg, MOTy::Def));
2353 M = (BuildMI(ChooseAddInstructionByType(argType), 3)
2354 .addReg(argVal).addSImm((int64_t) 0)
2355 .addReg(argVReg, MOTy::Def));
2357 // Mark the operand with the register it should be assigned
2358 M->SetRegForOperand(M->getNumOperands()-1, regNumForArg);
2359 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
2366 // add call instruction and delay slot before copying return value
2367 mvec.push_back(callMI);
2368 mvec.push_back(BuildMI(V9::NOP, 0));
2370 // Add the return value as an implicit ref. The call operands
2371 // were added above. Also, add code to copy out the return value.
2372 // This is always register-to-register for int or FP return values.
2374 if (callInstr->getType() != Type::VoidTy) {
2375 // Get the return value reg.
2376 const Type* retType = callInstr->getType();
2378 int regNum = (retType->isFloatingPoint()
2379 ? (unsigned) SparcFloatRegClass::f0
2380 : (unsigned) SparcIntRegClass::o0);
2381 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
2382 regNum = regInfo.getUnifiedRegNum(regClassID, regNum);
2384 // Create a virtual register to represent it and mark
2385 // this vreg as being an implicit operand of the call MI
2386 TmpInstruction* retVReg =
2387 new TmpInstruction(mcfi, callInstr, NULL, "argReg");
2389 callMI->addImplicitRef(retVReg, /*isDef*/ true);
2391 // Generate the reg-to-reg copy from the return value reg.
2392 // -- For FP values, create a FMOVS or FMOVD instruction
2393 // -- For non-FP values, create an add-with-0 instruction
2394 if (retType->isFloatingPoint())
2395 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
2396 .addReg(retVReg).addReg(callInstr, MOTy::Def));
2398 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
2399 .addReg(retVReg).addSImm((int64_t) 0)
2400 .addReg(callInstr, MOTy::Def));
2402 // Mark the operand with the register it should be assigned
2403 // Also mark the implicit ref of the call defining this operand
2404 M->SetRegForOperand(0, regNum);
2405 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,regNum);
2410 // For the CALL instruction, the ret. addr. reg. is also implicit
2411 if (isa<Function>(callee))
2412 callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
2414 MF.getInfo()->popAllTempValues(); // free temps used for this inst
2420 case 62: // reg: Shl(reg, reg)
2422 Value* argVal1 = subtreeRoot->leftChild()->getValue();
2423 Value* argVal2 = subtreeRoot->rightChild()->getValue();
2424 Instruction* shlInstr = subtreeRoot->getInstruction();
2426 const Type* opType = argVal1->getType();
2427 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2428 "Shl unsupported for other types");
2430 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
2431 (opType == Type::LongTy)? V9::SLLXr6:V9::SLLr5,
2432 argVal1, argVal2, 0, shlInstr, mvec,
2433 MachineCodeForInstruction::get(shlInstr));
2437 case 63: // reg: Shr(reg, reg)
2439 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
2440 assert((opType->isInteger() || isa<PointerType>(opType)) &&
2441 "Shr unsupported for other types");
2442 Add3OperandInstr(opType->isSigned()
2443 ? (opType == Type::LongTy ? V9::SRAXr6 : V9::SRAr5)
2444 : (opType == Type::LongTy ? V9::SRLXr6 : V9::SRLr5),
2449 case 64: // reg: Phi(reg,reg)
2450 break; // don't forward the value
2452 case 65: // reg: VaArg(reg)
2454 // Use value initialized by va_start as pointer to args on the stack.
2455 // Load argument via current pointer value, then increment pointer.
2456 int argSize = target.getFrameInfo().getSizeOfEachArgOnStack();
2457 Instruction* vaArgI = subtreeRoot->getInstruction();
2458 mvec.push_back(BuildMI(V9::LDXi, 3).addReg(vaArgI->getOperand(0)).
2459 addSImm(0).addRegDef(vaArgI));
2460 mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaArgI->getOperand(0)).
2461 addSImm(argSize).addRegDef(vaArgI->getOperand(0)));
2465 case 71: // reg: VReg
2466 case 72: // reg: Constant
2467 break; // don't forward the value
2470 assert(0 && "Unrecognized BURG rule");
2475 if (forwardOperandNum >= 0) {
2476 // We did not generate a machine instruction but need to use operand.
2477 // If user is in the same tree, replace Value in its machine operand.
2478 // If not, insert a copy instruction which should get coalesced away
2479 // by register allocation.
2480 if (subtreeRoot->parent() != NULL)
2481 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
2483 std::vector<MachineInstr*> minstrVec;
2484 Instruction* instr = subtreeRoot->getInstruction();
2485 target.getInstrInfo().
2486 CreateCopyInstructionsByType(target,
2487 instr->getParent()->getParent(),
2488 instr->getOperand(forwardOperandNum),
2490 MachineCodeForInstruction::get(instr));
2491 assert(minstrVec.size() > 0);
2492 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
2496 if (maskUnsignedResult) {
2497 // If result is unsigned and smaller than int reg size,
2498 // we need to clear high bits of result value.
2499 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
2500 Instruction* dest = subtreeRoot->getInstruction();
2501 if (dest->getType()->isUnsigned()) {
2502 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
2503 if (destSize <= 4) {
2504 // Mask high bits. Use a TmpInstruction to represent the
2505 // intermediate result before masking. Since those instructions
2506 // have already been generated, go back and substitute tmpI
2507 // for dest in the result position of each one of them.
2508 TmpInstruction *tmpI =
2509 new TmpInstruction(MachineCodeForInstruction::get(dest),
2510 dest->getType(), dest, NULL, "maskHi");
2512 for (unsigned i=0, N=mvec.size(); i < N; ++i)
2513 mvec[i]->substituteValue(dest, tmpI);
2515 M = BuildMI(V9::SRLi5, 3).addReg(tmpI).addZImm(8*(4-destSize))
2516 .addReg(dest, MOTy::Def);
2518 } else if (destSize < 8) {
2519 assert(0 && "Unsupported type size: 32 < size < 64 bits");