1 //===- SparcV9BurgISel.cpp - SparcV9 BURG-based Instruction Selector ------===//
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 // SparcV9 BURG-based instruction selector. It uses the SSA graph to
11 // construct a forest of BURG instruction trees (class InstrForest) and then
12 // uses the BURG-generated tree grammar (BURM) to find the optimal instruction
13 // sequences for the SparcV9.
15 //===----------------------------------------------------------------------===//
17 #include "MachineInstrAnnot.h"
18 #include "SparcV9BurgISel.h"
19 #include "SparcV9InstrForest.h"
20 #include "SparcV9Internals.h"
21 #include "SparcV9TmpInstr.h"
22 #include "llvm/CodeGen/IntrinsicLowering.h"
23 #include "llvm/CodeGen/MachineConstantPool.h"
24 #include "llvm/CodeGen/MachineFunction.h"
25 #include "llvm/CodeGen/MachineFunctionInfo.h"
26 #include "llvm/CodeGen/MachineInstr.h"
27 #include "llvm/CodeGen/MachineInstrBuilder.h"
28 #include "llvm/Constants.h"
29 #include "llvm/DerivedTypes.h"
30 #include "llvm/GlobalValue.h"
31 #include "llvm/Instructions.h"
32 #include "llvm/Intrinsics.h"
33 #include "llvm/Module.h"
34 #include "llvm/Pass.h"
35 #include "llvm/Support/CFG.h"
36 #include "llvm/Target/TargetInstrInfo.h"
37 #include "llvm/Target/TargetMachine.h"
38 #include "llvm/Type.h"
39 #include "Config/alloca.h"
40 #include "Support/CommandLine.h"
41 #include "Support/LeakDetector.h"
42 #include "Support/MathExtras.h"
43 #include "Support/STLExtras.h"
44 #include "Support/hash_map"
50 //==------------------------------------------------------------------------==//
51 // InstrForest (V9ISel BURG instruction trees) implementation
52 //==------------------------------------------------------------------------==//
56 class InstructionNode : public InstrTreeNode {
57 bool codeIsFoldedIntoParent;
60 InstructionNode(Instruction *_instr);
62 Instruction *getInstruction() const {
63 assert(treeNodeType == NTInstructionNode);
64 return cast<Instruction>(val);
67 void markFoldedIntoParent() { codeIsFoldedIntoParent = true; }
68 bool isFoldedIntoParent() { return codeIsFoldedIntoParent; }
70 // Methods to support type inquiry through isa, cast, and dyn_cast:
71 static inline bool classof(const InstructionNode *N) { return true; }
72 static inline bool classof(const InstrTreeNode *N) {
73 return N->getNodeType() == InstrTreeNode::NTInstructionNode;
77 virtual void dumpNode(int indent) const;
80 class VRegListNode : public InstrTreeNode {
82 VRegListNode() : InstrTreeNode(NTVRegListNode, 0) { opLabel = VRegListOp; }
83 // Methods to support type inquiry through isa, cast, and dyn_cast:
84 static inline bool classof(const VRegListNode *N) { return true; }
85 static inline bool classof(const InstrTreeNode *N) {
86 return N->getNodeType() == InstrTreeNode::NTVRegListNode;
89 virtual void dumpNode(int indent) const;
92 class VRegNode : public InstrTreeNode {
94 VRegNode(Value* _val) : InstrTreeNode(NTVRegNode, _val) {
97 // Methods to support type inquiry through isa, cast, and dyn_cast:
98 static inline bool classof(const VRegNode *N) { return true; }
99 static inline bool classof(const InstrTreeNode *N) {
100 return N->getNodeType() == InstrTreeNode::NTVRegNode;
103 virtual void dumpNode(int indent) const;
106 class ConstantNode : public InstrTreeNode {
108 ConstantNode(Constant *constVal)
109 : InstrTreeNode(NTConstNode, (Value*)constVal) {
110 opLabel = ConstantNodeOp;
112 Constant *getConstVal() const { return (Constant*) val;}
113 // Methods to support type inquiry through isa, cast, and dyn_cast:
114 static inline bool classof(const ConstantNode *N) { return true; }
115 static inline bool classof(const InstrTreeNode *N) {
116 return N->getNodeType() == InstrTreeNode::NTConstNode;
119 virtual void dumpNode(int indent) const;
122 class LabelNode : public InstrTreeNode {
124 LabelNode(BasicBlock* BB) : InstrTreeNode(NTLabelNode, (Value*)BB) {
125 opLabel = LabelNodeOp;
127 BasicBlock *getBasicBlock() const { return (BasicBlock*)val;}
128 // Methods to support type inquiry through isa, cast, and dyn_cast:
129 static inline bool classof(const LabelNode *N) { return true; }
130 static inline bool classof(const InstrTreeNode *N) {
131 return N->getNodeType() == InstrTreeNode::NTLabelNode;
134 virtual void dumpNode(int indent) const;
137 /// InstrForest - A forest of instruction trees for a single function.
138 /// The goal of InstrForest is to group instructions into a single
139 /// tree if one or more of them might be potentially combined into a
140 /// single complex instruction in the target machine. We group two
141 /// instructions O and I if: (1) Instruction O computes an operand used
142 /// by instruction I, and (2) O and I are part of the same basic block,
143 /// and (3) O has only a single use, viz., I.
145 class InstrForest : private hash_map<const Instruction *, InstructionNode*> {
147 // Use a vector for the root set to get a deterministic iterator
148 // for stable code generation. Even though we need to erase nodes
149 // during forest construction, a vector should still be efficient
150 // because the elements to erase are nearly always near the end.
151 typedef std::vector<InstructionNode*> RootSet;
152 typedef RootSet:: iterator root_iterator;
153 typedef RootSet::const_iterator const_root_iterator;
159 /*ctor*/ InstrForest (Function *F);
160 /*dtor*/ ~InstrForest ();
162 /// getTreeNodeForInstr - Returns the tree node for an Instruction.
164 inline InstructionNode *getTreeNodeForInstr(Instruction* instr) {
165 return (*this)[instr];
168 /// Iterators for the root nodes for all the trees.
170 const_root_iterator roots_begin() const { return treeRoots.begin(); }
171 root_iterator roots_begin() { return treeRoots.begin(); }
172 const_root_iterator roots_end () const { return treeRoots.end(); }
173 root_iterator roots_end () { return treeRoots.end(); }
178 // Methods used to build the instruction forest.
179 void eraseRoot (InstructionNode* node);
180 void setLeftChild (InstrTreeNode* parent, InstrTreeNode* child);
181 void setRightChild(InstrTreeNode* parent, InstrTreeNode* child);
182 void setParent (InstrTreeNode* child, InstrTreeNode* parent);
183 void noteTreeNodeForInstr(Instruction* instr, InstructionNode* treeNode);
184 InstructionNode* buildTreeForInstruction(Instruction* instr);
187 void InstrTreeNode::dump(int dumpChildren, int indent) const {
192 LeftChild->dump(dumpChildren, indent+1);
194 RightChild->dump(dumpChildren, indent+1);
198 InstructionNode::InstructionNode(Instruction* I)
199 : InstrTreeNode(NTInstructionNode, I), codeIsFoldedIntoParent(false) {
200 opLabel = I->getOpcode();
202 // Distinguish special cases of some instructions such as Ret and Br
204 if (opLabel == Instruction::Ret && cast<ReturnInst>(I)->getReturnValue()) {
205 opLabel = RetValueOp; // ret(value) operation
207 else if (opLabel ==Instruction::Br && !cast<BranchInst>(I)->isUnconditional())
209 opLabel = BrCondOp; // br(cond) operation
210 } else if (opLabel >= Instruction::SetEQ && opLabel <= Instruction::SetGT) {
211 opLabel = SetCCOp; // common label for all SetCC ops
212 } else if (opLabel == Instruction::Alloca && I->getNumOperands() > 0) {
213 opLabel = AllocaN; // Alloca(ptr, N) operation
214 } else if (opLabel == Instruction::GetElementPtr &&
215 cast<GetElementPtrInst>(I)->hasIndices()) {
216 opLabel = opLabel + 100; // getElem with index vector
217 } else if (opLabel == Instruction::Xor &&
218 BinaryOperator::isNot(I)) {
219 opLabel = (I->getType() == Type::BoolTy)? NotOp // boolean Not operator
220 : BNotOp; // bitwise Not operator
221 } else if (opLabel == Instruction::And || opLabel == Instruction::Or ||
222 opLabel == Instruction::Xor) {
223 // Distinguish bitwise operators from logical operators!
224 if (I->getType() != Type::BoolTy)
225 opLabel = opLabel + 100; // bitwise operator
226 } else if (opLabel == Instruction::Cast) {
227 const Type *ITy = I->getType();
228 switch(ITy->getTypeID())
230 case Type::BoolTyID: opLabel = ToBoolTy; break;
231 case Type::UByteTyID: opLabel = ToUByteTy; break;
232 case Type::SByteTyID: opLabel = ToSByteTy; break;
233 case Type::UShortTyID: opLabel = ToUShortTy; break;
234 case Type::ShortTyID: opLabel = ToShortTy; break;
235 case Type::UIntTyID: opLabel = ToUIntTy; break;
236 case Type::IntTyID: opLabel = ToIntTy; break;
237 case Type::ULongTyID: opLabel = ToULongTy; break;
238 case Type::LongTyID: opLabel = ToLongTy; break;
239 case Type::FloatTyID: opLabel = ToFloatTy; break;
240 case Type::DoubleTyID: opLabel = ToDoubleTy; break;
241 case Type::ArrayTyID: opLabel = ToArrayTy; break;
242 case Type::PointerTyID: opLabel = ToPointerTy; break;
244 // Just use `Cast' opcode otherwise. It's probably ignored.
250 void InstructionNode::dumpNode(int indent) const {
251 for (int i=0; i < indent; i++)
253 std::cerr << getInstruction()->getOpcodeName()
254 << " [label " << getOpLabel() << "]" << "\n";
257 void VRegListNode::dumpNode(int indent) const {
258 for (int i=0; i < indent; i++)
261 std::cerr << "List" << "\n";
264 void VRegNode::dumpNode(int indent) const {
265 for (int i=0; i < indent; i++)
267 std::cerr << "VReg " << *getValue() << "\n";
270 void ConstantNode::dumpNode(int indent) const {
271 for (int i=0; i < indent; i++)
273 std::cerr << "Constant " << *getValue() << "\n";
276 void LabelNode::dumpNode(int indent) const {
277 for (int i=0; i < indent; i++)
280 std::cerr << "Label " << *getValue() << "\n";
283 /// InstrForest ctor - Create a forest of instruction trees for a
286 InstrForest::InstrForest(Function *F) {
287 for (Function::iterator BB = F->begin(), FE = F->end(); BB != FE; ++BB) {
288 for(BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; ++I)
289 buildTreeForInstruction(I);
293 InstrForest::~InstrForest() {
294 for_each(treeRoots.begin(), treeRoots.end(), deleter<InstructionNode>);
297 void InstrForest::dump() const {
298 for (const_root_iterator I = roots_begin(); I != roots_end(); ++I)
299 (*I)->dump(/*dumpChildren*/ 1, /*indent*/ 0);
302 inline void InstrForest::eraseRoot(InstructionNode* node) {
303 for (RootSet::reverse_iterator RI=treeRoots.rbegin(), RE=treeRoots.rend();
306 treeRoots.erase(RI.base()-1);
309 inline void InstrForest::noteTreeNodeForInstr(Instruction *instr,
310 InstructionNode *treeNode) {
311 (*this)[instr] = treeNode;
312 treeRoots.push_back(treeNode); // mark node as root of a new tree
315 inline void InstrForest::setLeftChild(InstrTreeNode *parent,
316 InstrTreeNode *child) {
317 parent->LeftChild = child;
318 child->Parent = parent;
319 if (InstructionNode* instrNode = dyn_cast<InstructionNode>(child))
320 eraseRoot(instrNode); // no longer a tree root
323 inline void InstrForest::setRightChild(InstrTreeNode *parent,
324 InstrTreeNode *child) {
325 parent->RightChild = child;
326 child->Parent = parent;
327 if (InstructionNode* instrNode = dyn_cast<InstructionNode>(child))
328 eraseRoot(instrNode); // no longer a tree root
331 InstructionNode* InstrForest::buildTreeForInstruction(Instruction *instr) {
332 InstructionNode *treeNode = getTreeNodeForInstr(instr);
334 // treeNode has already been constructed for this instruction
335 assert(treeNode->getInstruction() == instr);
339 // Otherwise, create a new tree node for this instruction.
340 treeNode = new InstructionNode(instr);
341 noteTreeNodeForInstr(instr, treeNode);
343 if (instr->getOpcode() == Instruction::Call) {
344 // Operands of call instruction
348 // If the instruction has more than 2 instruction operands,
349 // then we need to create artificial list nodes to hold them.
350 // (Note that we only count operands that get tree nodes, and not
351 // others such as branch labels for a branch or switch instruction.)
352 // To do this efficiently, we'll walk all operands, build treeNodes
353 // for all appropriate operands and save them in an array. We then
354 // insert children at the end, creating list nodes where needed.
355 // As a performance optimization, allocate a child array only
356 // if a fixed array is too small.
358 InstrTreeNode** childArray = new InstrTreeNode*[instr->getNumOperands()];
360 // Walk the operands of the instruction
361 for (Instruction::op_iterator O = instr->op_begin(); O!=instr->op_end();
365 // Check if the operand is a data value, not an branch label, type,
366 // method or module. If the operand is an address type (i.e., label
367 // or method) that is used in an non-branching operation, e.g., `add'.
368 // that should be considered a data value.
369 // Check latter condition here just to simplify the next IF.
370 bool includeAddressOperand =
371 (isa<BasicBlock>(operand) || isa<Function>(operand))
372 && !instr->isTerminator();
374 if (includeAddressOperand || isa<Instruction>(operand) ||
375 isa<Constant>(operand) || isa<Argument>(operand)) {
376 // This operand is a data value.
377 // An instruction that computes the incoming value is added as a
378 // child of the current instruction if:
379 // the value has only a single use
380 // AND both instructions are in the same basic block.
381 // AND the current instruction is not a PHI (because the incoming
382 // value is conceptually in a predecessor block,
383 // even though it may be in the same static block)
384 // (Note that if the value has only a single use (viz., `instr'),
385 // the def of the value can be safely moved just before instr
386 // and therefore it is safe to combine these two instructions.)
387 // In all other cases, the virtual register holding the value
388 // is used directly, i.e., made a child of the instruction node.
389 InstrTreeNode* opTreeNode;
390 if (isa<Instruction>(operand) && operand->hasOneUse() &&
391 cast<Instruction>(operand)->getParent() == instr->getParent() &&
392 instr->getOpcode() != Instruction::PHI &&
393 instr->getOpcode() != Instruction::Call) {
394 // Recursively create a treeNode for it.
395 opTreeNode = buildTreeForInstruction((Instruction*)operand);
396 } else if (Constant *CPV = dyn_cast<Constant>(operand)) {
397 if (isa<GlobalValue>(CPV))
398 opTreeNode = new VRegNode(operand);
400 // Create a leaf node for a constant
401 opTreeNode = new ConstantNode(CPV);
403 // Create a leaf node for the virtual register
404 opTreeNode = new VRegNode(operand);
407 childArray[numChildren++] = opTreeNode;
411 // Add any selected operands as children in the tree.
412 // Certain instructions can have more than 2 in some instances (viz.,
413 // a CALL or a memory access -- LOAD, STORE, and GetElemPtr -- to an
414 // array or struct). Make the operands of every such instruction into
415 // a right-leaning binary tree with the operand nodes at the leaves
416 // and VRegList nodes as internal nodes.
417 InstrTreeNode *parent = treeNode;
419 if (numChildren > 2) {
420 unsigned instrOpcode = treeNode->getInstruction()->getOpcode();
421 assert(instrOpcode == Instruction::PHI ||
422 instrOpcode == Instruction::Call ||
423 instrOpcode == Instruction::Load ||
424 instrOpcode == Instruction::Store ||
425 instrOpcode == Instruction::GetElementPtr);
428 // Insert the first child as a direct child
429 if (numChildren >= 1)
430 setLeftChild(parent, childArray[0]);
434 // Create a list node for children 2 .. N-1, if any
435 for (n = numChildren-1; n >= 2; n--) {
436 // We have more than two children
437 InstrTreeNode *listNode = new VRegListNode();
438 setRightChild(parent, listNode);
439 setLeftChild(listNode, childArray[numChildren - n]);
443 // Now insert the last remaining child (if any).
444 if (numChildren >= 2) {
446 setRightChild(parent, childArray[numChildren - 1]);
449 delete [] childArray;
452 //==------------------------------------------------------------------------==//
453 // V9ISel Command-line options and declarations
454 //==------------------------------------------------------------------------==//
457 /// Allow the user to select the amount of debugging information printed
460 enum SelectDebugLevel_t {
462 Select_PrintMachineCode,
463 Select_DebugInstTrees,
464 Select_DebugBurgTrees,
466 cl::opt<SelectDebugLevel_t>
467 SelectDebugLevel("dselect", cl::Hidden,
468 cl::desc("enable instruction selection debug information"),
470 clEnumValN(Select_NoDebugInfo, "n", "disable debug output"),
471 clEnumValN(Select_PrintMachineCode, "y", "print generated machine code"),
472 clEnumValN(Select_DebugInstTrees, "i",
473 "print debugging info for instruction selection"),
474 clEnumValN(Select_DebugBurgTrees, "b", "print burg trees"),
478 /// V9ISel - This is the FunctionPass that drives the instruction selection
479 /// process on the SparcV9 target.
481 class V9ISel : public FunctionPass {
482 TargetMachine &Target;
483 void InsertCodeForPhis(Function &F);
484 void InsertPhiElimInstructions(BasicBlock *BB,
485 const std::vector<MachineInstr*>& CpVec);
486 void SelectInstructionsForTree(InstrTreeNode* treeRoot, int goalnt);
487 void PostprocessMachineCodeForTree(InstructionNode* instrNode,
488 int ruleForNode, short* nts);
490 V9ISel(TargetMachine &TM) : Target(TM) {}
492 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
493 AU.setPreservesCFG();
496 bool runOnFunction(Function &F);
497 virtual const char *getPassName() const {
498 return "SparcV9 BURG Instruction Selector";
504 //==------------------------------------------------------------------------==//
505 // Various V9ISel helper functions
506 //==------------------------------------------------------------------------==//
508 static const uint32_t MAXLO = (1 << 10) - 1; // set bits set by %lo(*)
509 static const uint32_t MAXSIMM = (1 << 12) - 1; // set bits in simm13 field of OR
511 /// ConvertConstantToIntType - Function to get the value of an integral
512 /// constant in the form that must be put into the machine register. The
513 /// specified constant is interpreted as (i.e., converted if necessary to) the
514 /// specified destination type. The result is always returned as an uint64_t,
515 /// since the representation of int64_t and uint64_t are identical. The
516 /// argument can be any known const. isValidConstant is set to true if a valid
517 /// constant was found.
519 uint64_t ConvertConstantToIntType(const TargetMachine &target, const Value *V,
520 const Type *destType, bool &isValidConstant) {
521 isValidConstant = false;
524 if (! destType->isIntegral() && ! isa<PointerType>(destType))
527 if (! isa<Constant>(V) || isa<GlobalValue>(V))
530 // GlobalValue: no conversions needed: get value and return it
531 if (const GlobalValue* GV = dyn_cast<GlobalValue>(V)) {
532 isValidConstant = true; // may be overwritten by recursive call
533 return ConvertConstantToIntType(target, GV, destType, isValidConstant);
536 // ConstantBool: no conversions needed: get value and return it
537 if (const ConstantBool *CB = dyn_cast<ConstantBool>(V)) {
538 isValidConstant = true;
539 return (uint64_t) CB->getValue();
542 // ConstantPointerNull: it's really just a big, shiny version of zero.
543 if (const ConstantPointerNull *CPN = dyn_cast<ConstantPointerNull>(V)) {
544 isValidConstant = true;
548 // For other types of constants, some conversion may be needed.
549 // First, extract the constant operand according to its own type
550 if (const ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
551 switch(CE->getOpcode()) {
552 case Instruction::Cast: // recursively get the value as cast
553 C = ConvertConstantToIntType(target, CE->getOperand(0), CE->getType(),
556 default: // not simplifying other ConstantExprs
559 else if (const ConstantInt *CI = dyn_cast<ConstantInt>(V)) {
560 isValidConstant = true;
561 C = CI->getRawValue();
562 } else if (const ConstantFP *CFP = dyn_cast<ConstantFP>(V)) {
563 isValidConstant = true;
564 double fC = CFP->getValue();
565 C = (destType->isSigned()? (uint64_t) (int64_t) fC
569 // Now if a valid value was found, convert it to destType.
570 if (isValidConstant) {
571 unsigned opSize = target.getTargetData().getTypeSize(V->getType());
572 unsigned destSize = target.getTargetData().getTypeSize(destType);
573 uint64_t maskHi = (destSize < 8)? (1U << 8*destSize) - 1 : ~0;
574 assert(opSize <= 8 && destSize <= 8 && ">8-byte int type unexpected");
576 if (destType->isSigned()) {
577 if (opSize > destSize) // operand is larger than dest:
578 C = C & maskHi; // mask high bits
580 if (opSize > destSize ||
581 (opSize == destSize && ! V->getType()->isSigned()))
582 if (C & (1U << (8*destSize - 1)))
583 C = C | ~maskHi; // sign-extend from destSize to 64 bits
586 if (opSize > destSize || (V->getType()->isSigned() && destSize < 8)) {
587 // operand is larger than dest,
588 // OR both are equal but smaller than the full register size
589 // AND operand is signed, so it may have extra sign bits:
599 /// CreateSETUWConst - Copy a 32-bit unsigned constant into the register
600 /// `dest', using SETHI, OR in the worst case. This function correctly emulates
601 /// the SETUW pseudo-op for SPARC v9 (if argument isSigned == false). The
602 /// isSigned=true case is used to implement SETSW without duplicating code. It
603 /// optimizes some common cases:
604 /// (1) Small value that fits in simm13 field of OR: don't need SETHI.
605 /// (2) isSigned = true and C is a small negative signed value, i.e.,
606 /// high bits are 1, and the remaining bits fit in simm13(OR).
608 CreateSETUWConst(const TargetMachine& target, uint32_t C,
609 Instruction* dest, std::vector<MachineInstr*>& mvec,
610 bool isSigned = false) {
611 MachineInstr *miSETHI = NULL, *miOR = NULL;
613 // In order to get efficient code, we should not generate the SETHI if
614 // all high bits are 1 (i.e., this is a small signed value that fits in
615 // the simm13 field of OR). So we check for and handle that case specially.
616 // NOTE: The value C = 0x80000000 is bad: sC < 0 *and* -sC < 0.
617 // In fact, sC == -sC, so we have to check for this explicitly.
618 int32_t sC = (int32_t) C;
619 bool smallNegValue =isSigned && sC < 0 && sC != -sC && -sC < (int32_t)MAXSIMM;
621 // Set the high 22 bits in dest if non-zero and simm13 field of OR not enough
622 if (!smallNegValue && (C & ~MAXLO) && C > MAXSIMM) {
623 miSETHI = BuildMI(V9::SETHI, 2).addZImm(C).addRegDef(dest);
624 miSETHI->getOperand(0).markHi32();
625 mvec.push_back(miSETHI);
628 // Set the low 10 or 12 bits in dest. This is necessary if no SETHI
629 // was generated, or if the low 10 bits are non-zero.
630 if (miSETHI==NULL || C & MAXLO) {
632 // unsigned value with high-order bits set using SETHI
633 miOR = BuildMI(V9::ORi,3).addReg(dest).addZImm(C).addRegDef(dest);
634 miOR->getOperand(1).markLo32();
636 // unsigned or small signed value that fits in simm13 field of OR
637 assert(smallNegValue || (C & ~MAXSIMM) == 0);
638 miOR = BuildMI(V9::ORi, 3).addMReg(target.getRegInfo()->getZeroRegNum())
639 .addSImm(sC).addRegDef(dest);
641 mvec.push_back(miOR);
644 assert((miSETHI || miOR) && "Oops, no code was generated!");
647 /// CreateSETSWConst - Set a 32-bit signed constant in the register `dest',
648 /// with sign-extension to 64 bits. This uses SETHI, OR, SRA in the worst case.
649 /// This function correctly emulates the SETSW pseudo-op for SPARC v9. It
650 /// optimizes the same cases as SETUWConst, plus:
651 /// (1) SRA is not needed for positive or small negative values.
654 CreateSETSWConst(const TargetMachine& target, int32_t C,
655 Instruction* dest, std::vector<MachineInstr*>& mvec) {
656 // Set the low 32 bits of dest
657 CreateSETUWConst(target, (uint32_t) C, dest, mvec, /*isSigned*/true);
659 // Sign-extend to the high 32 bits if needed.
660 // NOTE: The value C = 0x80000000 is bad: -C == C and so -C is < MAXSIMM
661 if (C < 0 && (C == -C || -C > (int32_t) MAXSIMM))
662 mvec.push_back(BuildMI(V9::SRAi5,3).addReg(dest).addZImm(0).addRegDef(dest));
665 /// CreateSETXConst - Set a 64-bit signed or unsigned constant in the
666 /// register `dest'. Use SETUWConst for each 32 bit word, plus a
667 /// left-shift-by-32 in between. This function correctly emulates the SETX
668 /// pseudo-op for SPARC v9. It optimizes the same cases as SETUWConst for each
672 CreateSETXConst(const TargetMachine& target, uint64_t C,
673 Instruction* tmpReg, Instruction* dest,
674 std::vector<MachineInstr*>& mvec) {
675 assert(C > (unsigned int) ~0 && "Use SETUW/SETSW for 32-bit values!");
679 // Code to set the upper 32 bits of the value in register `tmpReg'
680 CreateSETUWConst(target, (C >> 32), tmpReg, mvec);
682 // Shift tmpReg left by 32 bits
683 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpReg).addZImm(32)
686 // Code to set the low 32 bits of the value in register `dest'
687 CreateSETUWConst(target, C, dest, mvec);
689 // dest = OR(tmpReg, dest)
690 mvec.push_back(BuildMI(V9::ORr,3).addReg(dest).addReg(tmpReg).addRegDef(dest));
693 /// CreateSETUWLabel - Set a 32-bit constant (given by a symbolic label) in
694 /// the register `dest'.
697 CreateSETUWLabel(const TargetMachine& target, Value* val,
698 Instruction* dest, std::vector<MachineInstr*>& mvec) {
701 // Set the high 22 bits in dest
702 MI = BuildMI(V9::SETHI, 2).addReg(val).addRegDef(dest);
703 MI->getOperand(0).markHi32();
706 // Set the low 10 bits in dest
707 MI = BuildMI(V9::ORr, 3).addReg(dest).addReg(val).addRegDef(dest);
708 MI->getOperand(1).markLo32();
712 /// CreateSETXLabel - Set a 64-bit constant (given by a symbolic label) in the
716 CreateSETXLabel(const TargetMachine& target, Value* val, Instruction* tmpReg,
717 Instruction* dest, std::vector<MachineInstr*>& mvec) {
718 assert(isa<Constant>(val) &&
719 "I only know about constant values and global addresses");
723 MI = BuildMI(V9::SETHI, 2).addPCDisp(val).addRegDef(tmpReg);
724 MI->getOperand(0).markHi64();
727 MI = BuildMI(V9::ORi, 3).addReg(tmpReg).addPCDisp(val).addRegDef(tmpReg);
728 MI->getOperand(1).markLo64();
731 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpReg).addZImm(32)
733 MI = BuildMI(V9::SETHI, 2).addPCDisp(val).addRegDef(dest);
734 MI->getOperand(0).markHi32();
737 MI = BuildMI(V9::ORr, 3).addReg(dest).addReg(tmpReg).addRegDef(dest);
740 MI = BuildMI(V9::ORi, 3).addReg(dest).addPCDisp(val).addRegDef(dest);
741 MI->getOperand(1).markLo32();
745 /// CreateUIntSetInstruction - Create code to Set an unsigned constant in the
746 /// register `dest'. Uses CreateSETUWConst, CreateSETSWConst or CreateSETXConst
747 /// as needed. CreateSETSWConst is an optimization for the case that the
748 /// unsigned value has all ones in the 33 high bits (so that sign-extension sets
752 CreateUIntSetInstruction(const TargetMachine& target,
753 uint64_t C, Instruction* dest,
754 std::vector<MachineInstr*>& mvec,
755 MachineCodeForInstruction& mcfi) {
756 static const uint64_t lo32 = (uint32_t) ~0;
757 if (C <= lo32) // High 32 bits are 0. Set low 32 bits.
758 CreateSETUWConst(target, (uint32_t) C, dest, mvec);
759 else if ((C & ~lo32) == ~lo32 && (C & (1U << 31))) {
760 // All high 33 (not 32) bits are 1s: sign-extension will take care
761 // of high 32 bits, so use the sequence for signed int
762 CreateSETSWConst(target, (int32_t) C, dest, mvec);
763 } else if (C > lo32) {
764 // C does not fit in 32 bits
765 TmpInstruction* tmpReg = new TmpInstruction(mcfi, Type::IntTy);
766 CreateSETXConst(target, C, tmpReg, dest, mvec);
770 /// CreateIntSetInstruction - Create code to Set a signed constant in the
771 /// register `dest'. Really the same as CreateUIntSetInstruction.
774 CreateIntSetInstruction(const TargetMachine& target,
775 int64_t C, Instruction* dest,
776 std::vector<MachineInstr*>& mvec,
777 MachineCodeForInstruction& mcfi) {
778 CreateUIntSetInstruction(target, (uint64_t) C, dest, mvec, mcfi);
781 /// MaxConstantsTableTy - Table mapping LLVM opcodes to the max. immediate
782 /// constant usable for that operation in the SparcV9 backend. Used by
783 /// ConstantMayNotFitInImmedField().
785 struct MaxConstantsTableTy {
786 // Entry == 0 ==> no immediate constant field exists at all.
787 // Entry > 0 ==> abs(immediate constant) <= Entry
788 std::vector<int> tbl;
790 int getMaxConstantForInstr (unsigned llvmOpCode);
791 MaxConstantsTableTy ();
792 unsigned size() const { return tbl.size (); }
793 int &operator[] (unsigned index) { return tbl[index]; }
796 int MaxConstantsTableTy::getMaxConstantForInstr(unsigned llvmOpCode) {
797 int modelOpCode = -1;
799 if (llvmOpCode >= Instruction::BinaryOpsBegin &&
800 llvmOpCode < Instruction::BinaryOpsEnd)
801 modelOpCode = V9::ADDi;
804 case Instruction::Ret: modelOpCode = V9::JMPLCALLi; break;
806 case Instruction::Malloc:
807 case Instruction::Alloca:
808 case Instruction::GetElementPtr:
809 case Instruction::PHI:
810 case Instruction::Cast:
811 case Instruction::Call: modelOpCode = V9::ADDi; break;
813 case Instruction::Shl:
814 case Instruction::Shr: modelOpCode = V9::SLLXi6; break;
819 return (modelOpCode < 0)? 0: SparcV9MachineInstrDesc[modelOpCode].maxImmedConst;
822 MaxConstantsTableTy::MaxConstantsTableTy () : tbl (Instruction::OtherOpsEnd) {
824 assert(tbl.size() == Instruction::OtherOpsEnd &&
825 "assignments below will be illegal!");
826 for (op = Instruction::TermOpsBegin; op < Instruction::TermOpsEnd; ++op)
827 tbl[op] = getMaxConstantForInstr(op);
828 for (op = Instruction::BinaryOpsBegin; op < Instruction::BinaryOpsEnd; ++op)
829 tbl[op] = getMaxConstantForInstr(op);
830 for (op = Instruction::MemoryOpsBegin; op < Instruction::MemoryOpsEnd; ++op)
831 tbl[op] = getMaxConstantForInstr(op);
832 for (op = Instruction::OtherOpsBegin; op < Instruction::OtherOpsEnd; ++op)
833 tbl[op] = getMaxConstantForInstr(op);
836 bool ConstantMayNotFitInImmedField(const Constant* CV, const Instruction* I) {
837 // The one and only MaxConstantsTable, used only by this function.
838 static MaxConstantsTableTy MaxConstantsTable;
840 if (I->getOpcode() >= MaxConstantsTable.size()) // user-defined op (or bug!)
843 if (isa<ConstantPointerNull>(CV)) // can always use %g0
846 if (isa<SwitchInst>(I)) // Switch instructions will be lowered!
849 if (const ConstantInt* CI = dyn_cast<ConstantInt>(CV))
850 return labs((int64_t)CI->getRawValue()) > MaxConstantsTable[I->getOpcode()];
852 if (isa<ConstantBool>(CV))
853 return 1 > MaxConstantsTable[I->getOpcode()];
858 /// ChooseLoadInstruction - Return the appropriate load instruction opcode
859 /// based on the given LLVM value type.
861 static inline MachineOpCode ChooseLoadInstruction(const Type *DestTy) {
862 switch (DestTy->getTypeID()) {
864 case Type::UByteTyID: return V9::LDUBr;
865 case Type::SByteTyID: return V9::LDSBr;
866 case Type::UShortTyID: return V9::LDUHr;
867 case Type::ShortTyID: return V9::LDSHr;
868 case Type::UIntTyID: return V9::LDUWr;
869 case Type::IntTyID: return V9::LDSWr;
870 case Type::PointerTyID:
871 case Type::ULongTyID:
872 case Type::LongTyID: return V9::LDXr;
873 case Type::FloatTyID: return V9::LDFr;
874 case Type::DoubleTyID: return V9::LDDFr;
875 default: assert(0 && "Invalid type for Load instruction");
880 /// ChooseStoreInstruction - Return the appropriate store instruction opcode
881 /// based on the given LLVM value type.
883 static inline MachineOpCode ChooseStoreInstruction(const Type *DestTy) {
884 switch (DestTy->getTypeID()) {
886 case Type::UByteTyID:
887 case Type::SByteTyID: return V9::STBr;
888 case Type::UShortTyID:
889 case Type::ShortTyID: return V9::STHr;
891 case Type::IntTyID: return V9::STWr;
892 case Type::PointerTyID:
893 case Type::ULongTyID:
894 case Type::LongTyID: return V9::STXr;
895 case Type::FloatTyID: return V9::STFr;
896 case Type::DoubleTyID: return V9::STDFr;
897 default: assert(0 && "Invalid type for Store instruction");
902 static inline MachineOpCode ChooseAddInstructionByType(const Type* resultType) {
903 MachineOpCode opCode = V9::INVALID_OPCODE;
904 if (resultType->isIntegral() || isa<PointerType>(resultType)
905 || isa<FunctionType>(resultType) || resultType == Type::LabelTy) {
908 switch(resultType->getTypeID()) {
909 case Type::FloatTyID: opCode = V9::FADDS; break;
910 case Type::DoubleTyID: opCode = V9::FADDD; break;
911 default: assert(0 && "Invalid type for ADD instruction"); break;
917 /// convertOpcodeFromRegToImm - Because the SparcV9 instruction selector likes
918 /// to re-write operands to instructions, making them change from a Value*
919 /// (virtual register) to a Constant* (making an immediate field), we need to
920 /// change the opcode from a register-based instruction to an immediate-based
921 /// instruction, hence this mapping.
923 static unsigned convertOpcodeFromRegToImm(unsigned Opcode) {
926 case V9::ADDr: return V9::ADDi;
927 case V9::ADDccr: return V9::ADDcci;
928 case V9::ADDCr: return V9::ADDCi;
929 case V9::ADDCccr: return V9::ADDCcci;
930 case V9::SUBr: return V9::SUBi;
931 case V9::SUBccr: return V9::SUBcci;
932 case V9::SUBCr: return V9::SUBCi;
933 case V9::SUBCccr: return V9::SUBCcci;
934 case V9::MULXr: return V9::MULXi;
935 case V9::SDIVXr: return V9::SDIVXi;
936 case V9::UDIVXr: return V9::UDIVXi;
939 case V9::ANDr: return V9::ANDi;
940 case V9::ANDccr: return V9::ANDcci;
941 case V9::ANDNr: return V9::ANDNi;
942 case V9::ANDNccr: return V9::ANDNcci;
943 case V9::ORr: return V9::ORi;
944 case V9::ORccr: return V9::ORcci;
945 case V9::ORNr: return V9::ORNi;
946 case V9::ORNccr: return V9::ORNcci;
947 case V9::XORr: return V9::XORi;
948 case V9::XORccr: return V9::XORcci;
949 case V9::XNORr: return V9::XNORi;
950 case V9::XNORccr: return V9::XNORcci;
953 case V9::SLLr5: return V9::SLLi5;
954 case V9::SRLr5: return V9::SRLi5;
955 case V9::SRAr5: return V9::SRAi5;
956 case V9::SLLXr6: return V9::SLLXi6;
957 case V9::SRLXr6: return V9::SRLXi6;
958 case V9::SRAXr6: return V9::SRAXi6;
960 /* Conditional move on int comparison with zero */
961 case V9::MOVRZr: return V9::MOVRZi;
962 case V9::MOVRLEZr: return V9::MOVRLEZi;
963 case V9::MOVRLZr: return V9::MOVRLZi;
964 case V9::MOVRNZr: return V9::MOVRNZi;
965 case V9::MOVRGZr: return V9::MOVRGZi;
966 case V9::MOVRGEZr: return V9::MOVRGEZi;
969 /* Conditional move on int condition code */
970 case V9::MOVAr: return V9::MOVAi;
971 case V9::MOVNr: return V9::MOVNi;
972 case V9::MOVNEr: return V9::MOVNEi;
973 case V9::MOVEr: return V9::MOVEi;
974 case V9::MOVGr: return V9::MOVGi;
975 case V9::MOVLEr: return V9::MOVLEi;
976 case V9::MOVGEr: return V9::MOVGEi;
977 case V9::MOVLr: return V9::MOVLi;
978 case V9::MOVGUr: return V9::MOVGUi;
979 case V9::MOVLEUr: return V9::MOVLEUi;
980 case V9::MOVCCr: return V9::MOVCCi;
981 case V9::MOVCSr: return V9::MOVCSi;
982 case V9::MOVPOSr: return V9::MOVPOSi;
983 case V9::MOVNEGr: return V9::MOVNEGi;
984 case V9::MOVVCr: return V9::MOVVCi;
985 case V9::MOVVSr: return V9::MOVVSi;
987 /* Conditional move of int reg on fp condition code */
988 case V9::MOVFAr: return V9::MOVFAi;
989 case V9::MOVFNr: return V9::MOVFNi;
990 case V9::MOVFUr: return V9::MOVFUi;
991 case V9::MOVFGr: return V9::MOVFGi;
992 case V9::MOVFUGr: return V9::MOVFUGi;
993 case V9::MOVFLr: return V9::MOVFLi;
994 case V9::MOVFULr: return V9::MOVFULi;
995 case V9::MOVFLGr: return V9::MOVFLGi;
996 case V9::MOVFNEr: return V9::MOVFNEi;
997 case V9::MOVFEr: return V9::MOVFEi;
998 case V9::MOVFUEr: return V9::MOVFUEi;
999 case V9::MOVFGEr: return V9::MOVFGEi;
1000 case V9::MOVFUGEr: return V9::MOVFUGEi;
1001 case V9::MOVFLEr: return V9::MOVFLEi;
1002 case V9::MOVFULEr: return V9::MOVFULEi;
1003 case V9::MOVFOr: return V9::MOVFOi;
1006 case V9::LDSBr: return V9::LDSBi;
1007 case V9::LDSHr: return V9::LDSHi;
1008 case V9::LDSWr: return V9::LDSWi;
1009 case V9::LDUBr: return V9::LDUBi;
1010 case V9::LDUHr: return V9::LDUHi;
1011 case V9::LDUWr: return V9::LDUWi;
1012 case V9::LDXr: return V9::LDXi;
1013 case V9::LDFr: return V9::LDFi;
1014 case V9::LDDFr: return V9::LDDFi;
1015 case V9::LDQFr: return V9::LDQFi;
1016 case V9::LDFSRr: return V9::LDFSRi;
1017 case V9::LDXFSRr: return V9::LDXFSRi;
1020 case V9::STBr: return V9::STBi;
1021 case V9::STHr: return V9::STHi;
1022 case V9::STWr: return V9::STWi;
1023 case V9::STXr: return V9::STXi;
1024 case V9::STFr: return V9::STFi;
1025 case V9::STDFr: return V9::STDFi;
1026 case V9::STFSRr: return V9::STFSRi;
1027 case V9::STXFSRr: return V9::STXFSRi;
1030 case V9::JMPLCALLr: return V9::JMPLCALLi;
1031 case V9::JMPLRETr: return V9::JMPLRETi;
1033 /* save and restore */
1034 case V9::SAVEr: return V9::SAVEi;
1035 case V9::RESTOREr: return V9::RESTOREi;
1038 // It's already in correct format
1039 // Or, it's just not handled yet, but an assert() would break LLC
1041 std::cerr << "Unhandled opcode in convertOpcodeFromRegToImm(): " << Opcode
1048 /// CreateCodeToLoadConst - Create an instruction sequence to put the
1049 /// constant `val' into the virtual register `dest'. `val' may be a Constant or
1050 /// a GlobalValue, viz., the constant address of a global variable or function.
1051 /// The generated instructions are returned in `mvec'. Any temp. registers
1052 /// (TmpInstruction) created are recorded in mcfi. Any stack space required is
1053 /// allocated via MachineFunction.
1055 void CreateCodeToLoadConst(const TargetMachine& target, Function* F,
1056 Value* val, Instruction* dest,
1057 std::vector<MachineInstr*>& mvec,
1058 MachineCodeForInstruction& mcfi) {
1059 assert(isa<Constant>(val) &&
1060 "I only know about constant values and global addresses");
1062 // Use a "set" instruction for known constants or symbolic constants (labels)
1063 // that can go in an integer reg.
1064 // We have to use a "load" instruction for all other constants,
1065 // in particular, floating point constants.
1066 const Type* valType = val->getType();
1068 if (isa<GlobalValue>(val)) {
1069 TmpInstruction* tmpReg =
1070 new TmpInstruction(mcfi, PointerType::get(val->getType()), val);
1071 CreateSETXLabel(target, val, tmpReg, dest, mvec);
1076 uint64_t C = ConvertConstantToIntType(target, val, dest->getType(), isValid);
1078 if (dest->getType()->isSigned())
1079 CreateUIntSetInstruction(target, C, dest, mvec, mcfi);
1081 CreateIntSetInstruction(target, (int64_t) C, dest, mvec, mcfi);
1084 // Make an instruction sequence to load the constant, viz:
1085 // SETX <addr-of-constant>, tmpReg, addrReg
1086 // LOAD /*addr*/ addrReg, /*offset*/ 0, dest
1087 // First, create a tmp register to be used by the SETX sequence.
1088 TmpInstruction* tmpReg =
1089 new TmpInstruction(mcfi, PointerType::get(val->getType()));
1091 // Create another TmpInstruction for the address register
1092 TmpInstruction* addrReg =
1093 new TmpInstruction(mcfi, PointerType::get(val->getType()));
1095 // Get the constant pool index for this constant
1096 MachineConstantPool *CP = MachineFunction::get(F).getConstantPool();
1097 Constant *C = cast<Constant>(val);
1098 unsigned CPI = CP->getConstantPoolIndex(C);
1100 // Put the address of the constant into a register
1103 MI = BuildMI(V9::SETHI, 2).addConstantPoolIndex(CPI).addRegDef(tmpReg);
1104 MI->getOperand(0).markHi64();
1107 MI = BuildMI(V9::ORi, 3).addReg(tmpReg).addConstantPoolIndex(CPI)
1109 MI->getOperand(1).markLo64();
1112 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpReg).addZImm(32)
1113 .addRegDef(tmpReg));
1114 MI = BuildMI(V9::SETHI, 2).addConstantPoolIndex(CPI).addRegDef(addrReg);
1115 MI->getOperand(0).markHi32();
1118 MI = BuildMI(V9::ORr, 3).addReg(addrReg).addReg(tmpReg).addRegDef(addrReg);
1121 MI = BuildMI(V9::ORi, 3).addReg(addrReg).addConstantPoolIndex(CPI)
1122 .addRegDef(addrReg);
1123 MI->getOperand(1).markLo32();
1126 // Now load the constant from out ConstantPool label
1127 unsigned Opcode = ChooseLoadInstruction(val->getType());
1128 Opcode = convertOpcodeFromRegToImm(Opcode);
1129 mvec.push_back(BuildMI(Opcode, 3)
1130 .addReg(addrReg).addSImm((int64_t)0).addRegDef(dest));
1134 /// CreateCodeToCopyFloatToInt - Similarly, create an instruction sequence
1135 /// to copy an FP register `val' to an integer register `dest' by copying to
1136 /// memory and back. The generated instructions are returned in `mvec'. Any
1137 /// temp. virtual registers (TmpInstruction) created are recorded in mcfi.
1138 /// Temporary stack space required is allocated via MachineFunction.
1140 void CreateCodeToCopyFloatToInt(const TargetMachine& target, Function* F,
1141 Value* val, Instruction* dest,
1142 std::vector<MachineInstr*>& mvec,
1143 MachineCodeForInstruction& mcfi) {
1144 const Type* opTy = val->getType();
1145 const Type* destTy = dest->getType();
1146 assert(opTy->isFloatingPoint() && "Source type must be float/double");
1147 assert((destTy->isIntegral() || isa<PointerType>(destTy))
1148 && "Dest type must be integer, bool or pointer");
1150 // FIXME: For now, we allocate permanent space because the stack frame
1151 // manager does not allow locals to be allocated (e.g., for alloca) after
1152 // a temp is allocated!
1153 int offset = MachineFunction::get(F).getInfo()->allocateLocalVar(val);
1155 unsigned FPReg = target.getRegInfo()->getFramePointer();
1157 // Store instruction stores `val' to [%fp+offset].
1158 // The store opCode is based only the source value being copied.
1159 unsigned StoreOpcode = ChooseStoreInstruction(opTy);
1160 StoreOpcode = convertOpcodeFromRegToImm(StoreOpcode);
1161 mvec.push_back(BuildMI(StoreOpcode, 3)
1162 .addReg(val).addMReg(FPReg).addSImm(offset));
1164 // Load instruction loads [%fp+offset] to `dest'.
1165 // The type of the load opCode is the integer type that matches the
1166 // source type in size:
1167 // On SparcV9: int for float, long for double.
1168 // Note that we *must* use signed loads even for unsigned dest types, to
1169 // ensure correct sign-extension for UByte, UShort or UInt:
1170 const Type* loadTy = (opTy == Type::FloatTy)? Type::IntTy : Type::LongTy;
1171 unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
1172 LoadOpcode = convertOpcodeFromRegToImm(LoadOpcode);
1173 mvec.push_back(BuildMI(LoadOpcode, 3).addMReg(FPReg)
1174 .addSImm(offset).addRegDef(dest));
1177 /// CreateBitExtensionInstructions - Helper function for sign-extension and
1178 /// zero-extension. For SPARC v9, we sign-extend the given operand using SLL;
1182 CreateBitExtensionInstructions(bool signExtend, const TargetMachine& target,
1183 Function* F, Value* srcVal, Value* destVal,
1184 unsigned int numLowBits,
1185 std::vector<MachineInstr*>& mvec,
1186 MachineCodeForInstruction& mcfi) {
1189 assert(numLowBits <= 32 && "Otherwise, nothing should be done here!");
1191 if (numLowBits < 32) {
1192 // SLL is needed since operand size is < 32 bits.
1193 TmpInstruction *tmpI = new TmpInstruction(mcfi, destVal->getType(),
1194 srcVal, destVal, "make32");
1195 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(srcVal)
1196 .addZImm(32-numLowBits).addRegDef(tmpI));
1200 mvec.push_back(BuildMI(signExtend? V9::SRAi5 : V9::SRLi5, 3)
1201 .addReg(srcVal).addZImm(32-numLowBits).addRegDef(destVal));
1204 /// CreateSignExtensionInstructions - Create instruction sequence to produce
1205 /// a sign-extended register value from an arbitrary-sized integer value (sized
1206 /// in bits, not bytes). The generated instructions are returned in `mvec'. Any
1207 /// temp. registers (TmpInstruction) created are recorded in mcfi. Any stack
1208 /// space required is allocated via MachineFunction.
1210 void CreateSignExtensionInstructions(const TargetMachine& target,
1211 Function* F, Value* srcVal, Value* destVal,
1212 unsigned int numLowBits,
1213 std::vector<MachineInstr*>& mvec,
1214 MachineCodeForInstruction& mcfi) {
1215 CreateBitExtensionInstructions(/*signExtend*/ true, target, F, srcVal,
1216 destVal, numLowBits, mvec, mcfi);
1219 /// CreateZeroExtensionInstructions - Create instruction sequence to produce
1220 /// a zero-extended register value from an arbitrary-sized integer value (sized
1221 /// in bits, not bytes). For SPARC v9, we sign-extend the given operand using
1222 /// SLL; SRL. The generated instructions are returned in `mvec'. Any temp.
1223 /// registers (TmpInstruction) created are recorded in mcfi. Any stack space
1224 /// required is allocated via MachineFunction.
1226 void CreateZeroExtensionInstructions(const TargetMachine& target,
1227 Function* F, Value* srcVal, Value* destVal,
1228 unsigned int numLowBits,
1229 std::vector<MachineInstr*>& mvec,
1230 MachineCodeForInstruction& mcfi) {
1231 CreateBitExtensionInstructions(/*signExtend*/ false, target, F, srcVal,
1232 destVal, numLowBits, mvec, mcfi);
1235 /// CreateCodeToCopyIntToFloat - Create an instruction sequence to copy an
1236 /// integer register `val' to a floating point register `dest' by copying to
1237 /// memory and back. val must be an integral type. dest must be a Float or
1238 /// Double. The generated instructions are returned in `mvec'. Any temp.
1239 /// registers (TmpInstruction) created are recorded in mcfi. Any stack space
1240 /// required is allocated via MachineFunction.
1242 void CreateCodeToCopyIntToFloat(const TargetMachine& target,
1243 Function* F, Value* val, Instruction* dest,
1244 std::vector<MachineInstr*>& mvec,
1245 MachineCodeForInstruction& mcfi) {
1246 assert((val->getType()->isIntegral() || isa<PointerType>(val->getType()))
1247 && "Source type must be integral (integer or bool) or pointer");
1248 assert(dest->getType()->isFloatingPoint()
1249 && "Dest type must be float/double");
1251 // Get a stack slot to use for the copy
1252 int offset = MachineFunction::get(F).getInfo()->allocateLocalVar(val);
1254 // Get the size of the source value being copied.
1255 size_t srcSize = target.getTargetData().getTypeSize(val->getType());
1257 // Store instruction stores `val' to [%fp+offset].
1258 // The store and load opCodes are based on the size of the source value.
1259 // If the value is smaller than 32 bits, we must sign- or zero-extend it
1260 // to 32 bits since the load-float will load 32 bits.
1261 // Note that the store instruction is the same for signed and unsigned ints.
1262 const Type* storeType = (srcSize <= 4)? Type::IntTy : Type::LongTy;
1263 Value* storeVal = val;
1264 if (srcSize < target.getTargetData().getTypeSize(Type::FloatTy)) {
1265 // sign- or zero-extend respectively
1266 storeVal = new TmpInstruction(mcfi, storeType, val);
1267 if (val->getType()->isSigned())
1268 CreateSignExtensionInstructions(target, F, val, storeVal, 8*srcSize,
1271 CreateZeroExtensionInstructions(target, F, val, storeVal, 8*srcSize,
1275 unsigned FPReg = target.getRegInfo()->getFramePointer();
1276 unsigned StoreOpcode = ChooseStoreInstruction(storeType);
1277 StoreOpcode = convertOpcodeFromRegToImm(StoreOpcode);
1278 mvec.push_back(BuildMI(StoreOpcode, 3)
1279 .addReg(storeVal).addMReg(FPReg).addSImm(offset));
1281 // Load instruction loads [%fp+offset] to `dest'.
1282 // The type of the load opCode is the floating point type that matches the
1283 // stored type in size:
1284 // On SparcV9: float for int or smaller, double for long.
1285 const Type* loadType = (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
1286 unsigned LoadOpcode = ChooseLoadInstruction(loadType);
1287 LoadOpcode = convertOpcodeFromRegToImm(LoadOpcode);
1288 mvec.push_back(BuildMI(LoadOpcode, 3)
1289 .addMReg(FPReg).addSImm(offset).addRegDef(dest));
1292 /// InsertCodeToLoadConstant - Generates code to load the constant
1293 /// into a TmpInstruction (virtual reg) and returns the virtual register.
1295 static TmpInstruction*
1296 InsertCodeToLoadConstant(Function *F, Value* opValue, Instruction* vmInstr,
1297 std::vector<MachineInstr*>& loadConstVec,
1298 TargetMachine& target) {
1299 // Create a tmp virtual register to hold the constant.
1300 MachineCodeForInstruction &mcfi = MachineCodeForInstruction::get(vmInstr);
1301 TmpInstruction* tmpReg = new TmpInstruction(mcfi, opValue);
1303 CreateCodeToLoadConst(target, F, opValue, tmpReg, loadConstVec, mcfi);
1305 // Record the mapping from the tmp VM instruction to machine instruction.
1306 // Do this for all machine instructions that were not mapped to any
1307 // other temp values created by
1308 // tmpReg->addMachineInstruction(loadConstVec.back());
1312 MachineOperand::MachineOperandType
1313 ChooseRegOrImmed(int64_t intValue, bool isSigned,
1314 MachineOpCode opCode, const TargetMachine& target,
1315 bool canUseImmed, unsigned int& getMachineRegNum,
1316 int64_t& getImmedValue) {
1317 MachineOperand::MachineOperandType opType=MachineOperand::MO_VirtualRegister;
1318 getMachineRegNum = 0;
1322 target.getInstrInfo()->constantFitsInImmedField(opCode, intValue)) {
1323 opType = isSigned? MachineOperand::MO_SignExtendedImmed
1324 : MachineOperand::MO_UnextendedImmed;
1325 getImmedValue = intValue;
1326 } else if (intValue == 0 &&
1327 target.getRegInfo()->getZeroRegNum() != (unsigned)-1) {
1328 opType = MachineOperand::MO_MachineRegister;
1329 getMachineRegNum = target.getRegInfo()->getZeroRegNum();
1335 MachineOperand::MachineOperandType
1336 ChooseRegOrImmed(Value* val,
1337 MachineOpCode opCode, const TargetMachine& target,
1338 bool canUseImmed, unsigned int& getMachineRegNum,
1339 int64_t& getImmedValue) {
1340 getMachineRegNum = 0;
1343 // To use reg or immed, constant needs to be integer, bool, or a NULL pointer.
1344 // ConvertConstantToIntType() does the right conversions.
1345 bool isValidConstant;
1346 uint64_t valueToUse =
1347 ConvertConstantToIntType(target, val, val->getType(), isValidConstant);
1348 if (! isValidConstant)
1349 return MachineOperand::MO_VirtualRegister;
1351 // Now check if the constant value fits in the IMMED field.
1352 return ChooseRegOrImmed((int64_t) valueToUse, val->getType()->isSigned(),
1353 opCode, target, canUseImmed,
1354 getMachineRegNum, getImmedValue);
1357 /// CreateCopyInstructionsByType - Create instruction(s) to copy src to dest,
1358 /// for arbitrary types. The generated instructions are returned in `mvec'. Any
1359 /// temp. registers (TmpInstruction) created are recorded in mcfi. Any stack
1360 /// space required is allocated via MachineFunction.
1362 void CreateCopyInstructionsByType(const TargetMachine& target,
1363 Function *F, Value* src, Instruction* dest,
1364 std::vector<MachineInstr*>& mvec,
1365 MachineCodeForInstruction& mcfi) {
1366 bool loadConstantToReg = false;
1367 const Type* resultType = dest->getType();
1368 MachineOpCode opCode = ChooseAddInstructionByType(resultType);
1369 assert (opCode != V9::INVALID_OPCODE
1370 && "Unsupported result type in CreateCopyInstructionsByType()");
1372 // If `src' is a constant that doesn't fit in the immed field or if it is
1373 // a global variable (i.e., a constant address), generate a load
1374 // instruction instead of an add.
1375 if (isa<GlobalValue>(src))
1376 loadConstantToReg = true;
1377 else if (isa<Constant>(src)) {
1378 unsigned int machineRegNum;
1380 MachineOperand::MachineOperandType opType =
1381 ChooseRegOrImmed(src, opCode, target, /*canUseImmed*/ true,
1382 machineRegNum, immedValue);
1384 if (opType == MachineOperand::MO_VirtualRegister)
1385 loadConstantToReg = true;
1388 if (loadConstantToReg) {
1389 // `src' is constant and cannot fit in immed field for the ADD.
1390 // Insert instructions to "load" the constant into a register.
1391 CreateCodeToLoadConst(target, F, src, dest, mvec, mcfi);
1393 // Create a reg-to-reg copy instruction for the given type:
1394 // -- For FP values, create a FMOVS or FMOVD instruction
1395 // -- For non-FP values, create an add-with-0 instruction (opCode as above)
1396 // Make `src' the second operand, in case it is a small constant!
1398 if (resultType->isFloatingPoint())
1399 MI = (BuildMI(resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
1400 .addReg(src).addRegDef(dest));
1402 const Type* Ty =isa<PointerType>(resultType)? Type::ULongTy :resultType;
1403 MI = (BuildMI(opCode, 3)
1404 .addSImm((int64_t) 0).addReg(src).addRegDef(dest));
1410 /// FixConstantOperandsForInstr - Make a machine instruction use its constant
1411 /// operands more efficiently. If the constant is 0, then use the hardwired 0
1412 /// register, if any. Else, if the constant fits in the IMMEDIATE field, then
1413 /// use that field. Otherwise, else create instructions to put the constant
1414 /// into a register, either directly or by loading explicitly from the constant
1415 /// pool. In the first 2 cases, the operand of `minstr' is modified in place.
1416 /// Returns a vector of machine instructions generated for operands that fall
1417 /// under case 3; these must be inserted before `minstr'.
1419 std::vector<MachineInstr*>
1420 FixConstantOperandsForInstr(Instruction* vmInstr, MachineInstr* minstr,
1421 TargetMachine& target) {
1422 std::vector<MachineInstr*> MVec;
1424 MachineOpCode opCode = minstr->getOpcode();
1425 const TargetInstrInfo& instrInfo = *target.getInstrInfo();
1426 int resultPos = instrInfo.getResultPos(opCode);
1427 int immedPos = instrInfo.getImmedConstantPos(opCode);
1429 Function *F = vmInstr->getParent()->getParent();
1431 for (unsigned op=0; op < minstr->getNumOperands(); op++) {
1432 const MachineOperand& mop = minstr->getOperand(op);
1434 // Skip the result position, preallocated machine registers, or operands
1435 // that cannot be constants (CC regs or PC-relative displacements)
1436 if (resultPos == (int)op ||
1437 mop.getType() == MachineOperand::MO_MachineRegister ||
1438 mop.getType() == MachineOperand::MO_CCRegister ||
1439 mop.getType() == MachineOperand::MO_PCRelativeDisp)
1442 bool constantThatMustBeLoaded = false;
1443 unsigned int machineRegNum = 0;
1444 int64_t immedValue = 0;
1445 Value* opValue = NULL;
1446 MachineOperand::MachineOperandType opType =
1447 MachineOperand::MO_VirtualRegister;
1449 // Operand may be a virtual register or a compile-time constant
1450 if (mop.getType() == MachineOperand::MO_VirtualRegister) {
1451 assert(mop.getVRegValue() != NULL);
1452 opValue = mop.getVRegValue();
1453 if (Constant *opConst = dyn_cast<Constant>(opValue))
1454 if (!isa<GlobalValue>(opConst)) {
1455 opType = ChooseRegOrImmed(opConst, opCode, target,
1456 (immedPos == (int)op), machineRegNum,
1458 if (opType == MachineOperand::MO_VirtualRegister)
1459 constantThatMustBeLoaded = true;
1462 // If the operand is from the constant pool, don't try to change it.
1463 if (mop.getType() == MachineOperand::MO_ConstantPoolIndex) {
1466 assert(mop.isImmediate());
1467 bool isSigned = mop.getType() == MachineOperand::MO_SignExtendedImmed;
1469 // Bit-selection flags indicate an instruction that is extracting
1470 // bits from its operand so ignore this even if it is a big constant.
1471 if (mop.isHiBits32() || mop.isLoBits32() ||
1472 mop.isHiBits64() || mop.isLoBits64())
1475 opType = ChooseRegOrImmed(mop.getImmedValue(), isSigned,
1476 opCode, target, (immedPos == (int)op),
1477 machineRegNum, immedValue);
1479 if (opType == MachineOperand::MO_SignExtendedImmed ||
1480 opType == MachineOperand::MO_UnextendedImmed) {
1481 // The optype is an immediate value
1482 // This means we need to change the opcode, e.g. ADDr -> ADDi
1483 unsigned newOpcode = convertOpcodeFromRegToImm(opCode);
1484 minstr->setOpcode(newOpcode);
1487 if (opType == mop.getType())
1488 continue; // no change: this is the most common case
1490 if (opType == MachineOperand::MO_VirtualRegister) {
1491 constantThatMustBeLoaded = true;
1493 ? (Value*)ConstantSInt::get(Type::LongTy, immedValue)
1494 : (Value*)ConstantUInt::get(Type::ULongTy,(uint64_t)immedValue);
1498 if (opType == MachineOperand::MO_MachineRegister)
1499 minstr->SetMachineOperandReg(op, machineRegNum);
1500 else if (opType == MachineOperand::MO_SignExtendedImmed ||
1501 opType == MachineOperand::MO_UnextendedImmed) {
1502 minstr->SetMachineOperandConst(op, opType, immedValue);
1503 // The optype is or has become an immediate
1504 // This means we need to change the opcode, e.g. ADDr -> ADDi
1505 unsigned newOpcode = convertOpcodeFromRegToImm(opCode);
1506 minstr->setOpcode(newOpcode);
1507 } else if (constantThatMustBeLoaded ||
1508 (opValue && isa<GlobalValue>(opValue)))
1509 { // opValue is a constant that must be explicitly loaded into a reg
1511 TmpInstruction* tmpReg = InsertCodeToLoadConstant(F, opValue, vmInstr,
1513 minstr->SetMachineOperandVal(op, MachineOperand::MO_VirtualRegister,
1518 // Also, check for implicit operands used by the machine instruction
1519 // (no need to check those defined since they cannot be constants).
1521 // -- arguments to a Call
1522 // -- return value of a Return
1523 // Any such operand that is a constant value needs to be fixed also.
1524 // The current instructions with implicit refs (viz., Call and Return)
1525 // have no immediate fields, so the constant always needs to be loaded
1527 bool isCall = instrInfo.isCall(opCode);
1528 unsigned lastCallArgNum = 0; // unused if not a call
1529 CallArgsDescriptor* argDesc = NULL; // unused if not a call
1531 argDesc = CallArgsDescriptor::get(minstr);
1533 for (unsigned i=0, N=minstr->getNumImplicitRefs(); i < N; ++i)
1534 if (isa<Constant>(minstr->getImplicitRef(i))) {
1535 Value* oldVal = minstr->getImplicitRef(i);
1536 TmpInstruction* tmpReg =
1537 InsertCodeToLoadConstant(F, oldVal, vmInstr, MVec, target);
1538 minstr->setImplicitRef(i, tmpReg);
1541 // find and replace the argument in the CallArgsDescriptor
1542 unsigned i=lastCallArgNum;
1543 while (argDesc->getArgInfo(i).getArgVal() != oldVal)
1545 assert(i < argDesc->getNumArgs() &&
1546 "Constant operands to a call *must* be in the arg list");
1548 argDesc->getArgInfo(i).replaceArgVal(tmpReg);
1555 static inline void Add3OperandInstr(unsigned Opcode, InstructionNode* Node,
1556 std::vector<MachineInstr*>& mvec) {
1557 mvec.push_back(BuildMI(Opcode, 3).addReg(Node->leftChild()->getValue())
1558 .addReg(Node->rightChild()->getValue())
1559 .addRegDef(Node->getValue()));
1562 /// IsZero - Check for a constant 0.
1564 static inline bool IsZero(Value* idx) {
1565 return (idx == ConstantSInt::getNullValue(idx->getType()));
1568 /// FoldGetElemChain - Fold a chain of GetElementPtr instructions containing
1569 /// only constant offsets into an equivalent (Pointer, IndexVector) pair.
1570 /// Returns the pointer Value, and stores the resulting IndexVector in argument
1571 /// chainIdxVec. This is a helper function for FoldConstantIndices that does the
1575 FoldGetElemChain(InstrTreeNode* ptrNode, std::vector<Value*>& chainIdxVec,
1576 bool lastInstHasLeadingNonZero) {
1577 InstructionNode* gepNode = dyn_cast<InstructionNode>(ptrNode);
1578 GetElementPtrInst* gepInst =
1579 dyn_cast_or_null<GetElementPtrInst>(gepNode ? gepNode->getInstruction() :0);
1581 // ptr value is not computed in this tree or ptr value does not come from GEP
1583 if (gepInst == NULL)
1586 // Return NULL if we don't fold any instructions in.
1587 Value* ptrVal = NULL;
1589 // Now chase the chain of getElementInstr instructions, if any.
1590 // Check for any non-constant indices and stop there.
1591 // Also, stop if the first index of child is a non-zero array index
1592 // and the last index of the current node is a non-array index:
1593 // in that case, a non-array declared type is being accessed as an array
1594 // which is not type-safe, but could be legal.
1595 InstructionNode* ptrChild = gepNode;
1596 while (ptrChild && (ptrChild->getOpLabel() == Instruction::GetElementPtr ||
1597 ptrChild->getOpLabel() == GetElemPtrIdx)) {
1598 // Child is a GetElemPtr instruction
1599 gepInst = cast<GetElementPtrInst>(ptrChild->getValue());
1600 User::op_iterator OI, firstIdx = gepInst->idx_begin();
1601 User::op_iterator lastIdx = gepInst->idx_end();
1602 bool allConstantOffsets = true;
1604 // The first index of every GEP must be an array index.
1605 assert((*firstIdx)->getType() == Type::LongTy &&
1606 "INTERNAL ERROR: Structure index for a pointer type!");
1608 // If the last instruction had a leading non-zero index, check if the
1609 // current one references a sequential (i.e., indexable) type.
1610 // If not, the code is not type-safe and we would create an illegal GEP
1611 // by folding them, so don't fold any more instructions.
1612 if (lastInstHasLeadingNonZero)
1613 if (! isa<SequentialType>(gepInst->getType()->getElementType()))
1614 break; // cannot fold in any preceding getElementPtr instrs.
1616 // Check that all offsets are constant for this instruction
1617 for (OI = firstIdx; allConstantOffsets && OI != lastIdx; ++OI)
1618 allConstantOffsets = isa<ConstantInt>(*OI);
1620 if (allConstantOffsets) {
1621 // Get pointer value out of ptrChild.
1622 ptrVal = gepInst->getPointerOperand();
1624 // Insert its index vector at the start, skipping any leading [0]
1625 // Remember the old size to check if anything was inserted.
1626 unsigned oldSize = chainIdxVec.size();
1627 int firstIsZero = IsZero(*firstIdx);
1628 chainIdxVec.insert(chainIdxVec.begin(), firstIdx + firstIsZero, lastIdx);
1630 // Remember if it has leading zero index: it will be discarded later.
1631 if (oldSize < chainIdxVec.size())
1632 lastInstHasLeadingNonZero = !firstIsZero;
1634 // Mark the folded node so no code is generated for it.
1635 ((InstructionNode*) ptrChild)->markFoldedIntoParent();
1637 // Get the previous GEP instruction and continue trying to fold
1638 ptrChild = dyn_cast<InstructionNode>(ptrChild->leftChild());
1639 } else // cannot fold this getElementPtr instr. or any preceding ones
1643 // If the first getElementPtr instruction had a leading [0], add it back.
1644 // Note that this instruction is the *last* one that was successfully
1645 // folded *and* contributed any indices, in the loop above.
1646 if (ptrVal && ! lastInstHasLeadingNonZero)
1647 chainIdxVec.insert(chainIdxVec.begin(), ConstantSInt::get(Type::LongTy,0));
1652 /// GetGEPInstArgs - Helper function for GetMemInstArgs that handles the
1653 /// final getElementPtr instruction used by (or same as) the memory operation.
1654 /// Extracts the indices of the current instruction and tries to fold in
1655 /// preceding ones if all indices of the current one are constant.
1657 static Value *GetGEPInstArgs(InstructionNode *gepNode,
1658 std::vector<Value *> &idxVec,
1659 bool &allConstantIndices) {
1660 allConstantIndices = true;
1661 GetElementPtrInst* gepI = cast<GetElementPtrInst>(gepNode->getInstruction());
1663 // Default pointer is the one from the current instruction.
1664 Value* ptrVal = gepI->getPointerOperand();
1665 InstrTreeNode* ptrChild = gepNode->leftChild();
1667 // Extract the index vector of the GEP instruction.
1668 // If all indices are constant and first index is zero, try to fold
1669 // in preceding GEPs with all constant indices.
1670 for (User::op_iterator OI=gepI->idx_begin(), OE=gepI->idx_end();
1671 allConstantIndices && OI != OE; ++OI)
1672 if (! isa<Constant>(*OI))
1673 allConstantIndices = false; // note: this also terminates loop!
1675 // If we have only constant indices, fold chains of constant indices
1676 // in this and any preceding GetElemPtr instructions.
1677 bool foldedGEPs = false;
1678 bool leadingNonZeroIdx = gepI && ! IsZero(*gepI->idx_begin());
1679 if (allConstantIndices)
1680 if (Value* newPtr = FoldGetElemChain(ptrChild, idxVec, leadingNonZeroIdx)) {
1685 // Append the index vector of the current instruction.
1686 // Skip the leading [0] index if preceding GEPs were folded into this.
1687 idxVec.insert(idxVec.end(),
1688 gepI->idx_begin() + (foldedGEPs && !leadingNonZeroIdx),
1694 /// GetMemInstArgs - Get the pointer value and the index vector for a memory
1695 /// operation (GetElementPtr, Load, or Store). If all indices of the given
1696 /// memory operation are constant, fold in constant indices in a chain of
1697 /// preceding GetElementPtr instructions (if any), and return the pointer value
1698 /// of the first instruction in the chain. All folded instructions are marked so
1699 /// no code is generated for them. Returns the pointer Value to use, and
1700 /// returns the resulting IndexVector in idxVec. Sets allConstantIndices
1701 /// to true/false if all indices are/aren't const.
1703 static Value *GetMemInstArgs(InstructionNode *memInstrNode,
1704 std::vector<Value*> &idxVec,
1705 bool& allConstantIndices) {
1706 allConstantIndices = false;
1707 Instruction* memInst = memInstrNode->getInstruction();
1708 assert(idxVec.size() == 0 && "Need empty vector to return indices");
1710 // If there is a GetElemPtr instruction to fold in to this instr,
1711 // it must be in the left child for Load and GetElemPtr, and in the
1712 // right child for Store instructions.
1713 InstrTreeNode* ptrChild = (memInst->getOpcode() == Instruction::Store
1714 ? memInstrNode->rightChild()
1715 : memInstrNode->leftChild());
1717 // Default pointer is the one from the current instruction.
1718 Value* ptrVal = ptrChild->getValue();
1720 // Find the "last" GetElemPtr instruction: this one or the immediate child.
1721 // There will be none if this is a load or a store from a scalar pointer.
1722 InstructionNode* gepNode = NULL;
1723 if (isa<GetElementPtrInst>(memInst))
1724 gepNode = memInstrNode;
1725 else if (isa<InstructionNode>(ptrChild) && isa<GetElementPtrInst>(ptrVal)) {
1726 // Child of load/store is a GEP and memInst is its only use.
1727 // Use its indices and mark it as folded.
1728 gepNode = cast<InstructionNode>(ptrChild);
1729 gepNode->markFoldedIntoParent();
1732 // If there are no indices, return the current pointer.
1733 // Else extract the pointer from the GEP and fold the indices.
1734 return gepNode ? GetGEPInstArgs(gepNode, idxVec, allConstantIndices)
1738 static inline MachineOpCode
1739 ChooseBprInstruction(const InstructionNode* instrNode) {
1740 MachineOpCode opCode;
1742 Instruction* setCCInstr =
1743 ((InstructionNode*) instrNode->leftChild())->getInstruction();
1745 switch(setCCInstr->getOpcode()) {
1746 case Instruction::SetEQ: opCode = V9::BRZ; break;
1747 case Instruction::SetNE: opCode = V9::BRNZ; break;
1748 case Instruction::SetLE: opCode = V9::BRLEZ; break;
1749 case Instruction::SetGE: opCode = V9::BRGEZ; break;
1750 case Instruction::SetLT: opCode = V9::BRLZ; break;
1751 case Instruction::SetGT: opCode = V9::BRGZ; break;
1753 assert(0 && "Unrecognized VM instruction!");
1754 opCode = V9::INVALID_OPCODE;
1761 static inline MachineOpCode
1762 ChooseBpccInstruction(const InstructionNode* instrNode,
1763 const BinaryOperator* setCCInstr) {
1764 MachineOpCode opCode = V9::INVALID_OPCODE;
1766 bool isSigned = setCCInstr->getOperand(0)->getType()->isSigned();
1769 switch(setCCInstr->getOpcode()) {
1770 case Instruction::SetEQ: opCode = V9::BE; break;
1771 case Instruction::SetNE: opCode = V9::BNE; break;
1772 case Instruction::SetLE: opCode = V9::BLE; break;
1773 case Instruction::SetGE: opCode = V9::BGE; break;
1774 case Instruction::SetLT: opCode = V9::BL; break;
1775 case Instruction::SetGT: opCode = V9::BG; break;
1777 assert(0 && "Unrecognized VM instruction!");
1781 switch(setCCInstr->getOpcode()) {
1782 case Instruction::SetEQ: opCode = V9::BE; break;
1783 case Instruction::SetNE: opCode = V9::BNE; break;
1784 case Instruction::SetLE: opCode = V9::BLEU; break;
1785 case Instruction::SetGE: opCode = V9::BCC; break;
1786 case Instruction::SetLT: opCode = V9::BCS; break;
1787 case Instruction::SetGT: opCode = V9::BGU; break;
1789 assert(0 && "Unrecognized VM instruction!");
1797 static inline MachineOpCode
1798 ChooseBFpccInstruction(const InstructionNode* instrNode,
1799 const BinaryOperator* setCCInstr) {
1800 MachineOpCode opCode = V9::INVALID_OPCODE;
1802 switch(setCCInstr->getOpcode()) {
1803 case Instruction::SetEQ: opCode = V9::FBE; break;
1804 case Instruction::SetNE: opCode = V9::FBNE; break;
1805 case Instruction::SetLE: opCode = V9::FBLE; break;
1806 case Instruction::SetGE: opCode = V9::FBGE; break;
1807 case Instruction::SetLT: opCode = V9::FBL; break;
1808 case Instruction::SetGT: opCode = V9::FBG; break;
1810 assert(0 && "Unrecognized VM instruction!");
1817 // GetTmpForCC - Create a unique TmpInstruction for a boolean value,
1818 // representing the CC register used by a branch on that value.
1819 // For now, hack this using a little static cache of TmpInstructions.
1820 // Eventually the entire BURG instruction selection should be put
1821 // into a separate class that can hold such information.
1822 // The static cache is not too bad because the memory for these
1823 // TmpInstructions will be freed along with the rest of the Function anyway.
1825 static TmpInstruction *GetTmpForCC (Value* boolVal, const Function *F,
1827 MachineCodeForInstruction& mcfi) {
1828 typedef hash_map<const Value*, TmpInstruction*> BoolTmpCache;
1829 static BoolTmpCache boolToTmpCache; // Map boolVal -> TmpInstruction*
1830 static const Function *lastFunction = 0;// Use to flush cache between funcs
1832 assert(boolVal->getType() == Type::BoolTy && "Weird but ok! Delete assert");
1834 if (lastFunction != F) {
1836 boolToTmpCache.clear();
1839 // Look for tmpI and create a new one otherwise. The new value is
1840 // directly written to map using the ref returned by operator[].
1841 TmpInstruction*& tmpI = boolToTmpCache[boolVal];
1843 tmpI = new TmpInstruction(mcfi, ccType, boolVal);
1848 static inline MachineOpCode
1849 ChooseBccInstruction(const InstructionNode* instrNode, const Type*& setCCType) {
1850 InstructionNode* setCCNode = (InstructionNode*) instrNode->leftChild();
1851 assert(setCCNode->getOpLabel() == SetCCOp);
1852 BinaryOperator* setCCInstr =cast<BinaryOperator>(setCCNode->getInstruction());
1853 setCCType = setCCInstr->getOperand(0)->getType();
1855 if (setCCType->isFloatingPoint())
1856 return ChooseBFpccInstruction(instrNode, setCCInstr);
1858 return ChooseBpccInstruction(instrNode, setCCInstr);
1861 /// ChooseMovFpcciInstruction - WARNING: since this function has only one
1862 /// caller, it always returns the opcode that expects an immediate and a
1863 /// register. If this function is ever used in cases where an opcode that takes
1864 /// two registers is required, then modify this function and use
1865 /// convertOpcodeFromRegToImm() where required. It will be necessary to expand
1866 /// convertOpcodeFromRegToImm() to handle the new cases of opcodes.
1868 static inline MachineOpCode
1869 ChooseMovFpcciInstruction(const InstructionNode* instrNode) {
1870 MachineOpCode opCode = V9::INVALID_OPCODE;
1872 switch(instrNode->getInstruction()->getOpcode()) {
1873 case Instruction::SetEQ: opCode = V9::MOVFEi; break;
1874 case Instruction::SetNE: opCode = V9::MOVFNEi; break;
1875 case Instruction::SetLE: opCode = V9::MOVFLEi; break;
1876 case Instruction::SetGE: opCode = V9::MOVFGEi; break;
1877 case Instruction::SetLT: opCode = V9::MOVFLi; break;
1878 case Instruction::SetGT: opCode = V9::MOVFGi; break;
1880 assert(0 && "Unrecognized VM instruction!");
1887 /// ChooseMovpcciForSetCC -- Choose a conditional-move instruction
1888 /// based on the type of SetCC operation.
1890 /// WARNING: like the previous function, this function always returns
1891 /// the opcode that expects an immediate and a register. See above.
1893 static MachineOpCode ChooseMovpcciForSetCC(const InstructionNode* instrNode) {
1894 MachineOpCode opCode = V9::INVALID_OPCODE;
1896 const Type* opType = instrNode->leftChild()->getValue()->getType();
1897 assert(opType->isIntegral() || isa<PointerType>(opType));
1898 bool noSign = opType->isUnsigned() || isa<PointerType>(opType);
1900 switch(instrNode->getInstruction()->getOpcode()) {
1901 case Instruction::SetEQ: opCode = V9::MOVEi; break;
1902 case Instruction::SetLE: opCode = noSign? V9::MOVLEUi : V9::MOVLEi; break;
1903 case Instruction::SetGE: opCode = noSign? V9::MOVCCi : V9::MOVGEi; break;
1904 case Instruction::SetLT: opCode = noSign? V9::MOVCSi : V9::MOVLi; break;
1905 case Instruction::SetGT: opCode = noSign? V9::MOVGUi : V9::MOVGi; break;
1906 case Instruction::SetNE: opCode = V9::MOVNEi; break;
1907 default: assert(0 && "Unrecognized LLVM instr!"); break;
1913 /// ChooseMovpregiForSetCC -- Choose a conditional-move-on-register-value
1914 /// instruction based on the type of SetCC operation. These instructions
1915 /// compare a register with 0 and perform the move is the comparison is true.
1917 /// WARNING: like the previous function, this function it always returns
1918 /// the opcode that expects an immediate and a register. See above.
1920 static MachineOpCode ChooseMovpregiForSetCC(const InstructionNode* instrNode) {
1921 MachineOpCode opCode = V9::INVALID_OPCODE;
1923 switch(instrNode->getInstruction()->getOpcode()) {
1924 case Instruction::SetEQ: opCode = V9::MOVRZi; break;
1925 case Instruction::SetLE: opCode = V9::MOVRLEZi; break;
1926 case Instruction::SetGE: opCode = V9::MOVRGEZi; break;
1927 case Instruction::SetLT: opCode = V9::MOVRLZi; break;
1928 case Instruction::SetGT: opCode = V9::MOVRGZi; break;
1929 case Instruction::SetNE: opCode = V9::MOVRNZi; break;
1930 default: assert(0 && "Unrecognized VM instr!"); break;
1936 static inline MachineOpCode
1937 ChooseConvertToFloatInstr(const TargetMachine& target,
1938 OpLabel vopCode, const Type* opType) {
1939 assert((vopCode == ToFloatTy || vopCode == ToDoubleTy) &&
1940 "Unrecognized convert-to-float opcode!");
1941 assert((opType->isIntegral() || opType->isFloatingPoint() ||
1942 isa<PointerType>(opType))
1943 && "Trying to convert a non-scalar type to FLOAT/DOUBLE?");
1945 MachineOpCode opCode = V9::INVALID_OPCODE;
1947 unsigned opSize = target.getTargetData().getTypeSize(opType);
1949 if (opType == Type::FloatTy)
1950 opCode = (vopCode == ToFloatTy? V9::NOP : V9::FSTOD);
1951 else if (opType == Type::DoubleTy)
1952 opCode = (vopCode == ToFloatTy? V9::FDTOS : V9::NOP);
1953 else if (opSize <= 4)
1954 opCode = (vopCode == ToFloatTy? V9::FITOS : V9::FITOD);
1956 assert(opSize == 8 && "Unrecognized type size > 4 and < 8!");
1957 opCode = (vopCode == ToFloatTy? V9::FXTOS : V9::FXTOD);
1963 static inline MachineOpCode
1964 ChooseConvertFPToIntInstr(const TargetMachine& target,
1965 const Type* destType, const Type* opType) {
1966 assert((opType == Type::FloatTy || opType == Type::DoubleTy)
1967 && "This function should only be called for FLOAT or DOUBLE");
1968 assert((destType->isIntegral() || isa<PointerType>(destType))
1969 && "Trying to convert FLOAT/DOUBLE to a non-scalar type?");
1971 MachineOpCode opCode = V9::INVALID_OPCODE;
1973 unsigned destSize = target.getTargetData().getTypeSize(destType);
1975 if (destType == Type::UIntTy)
1976 assert(destType != Type::UIntTy && "Expand FP-to-uint beforehand.");
1977 else if (destSize <= 4)
1978 opCode = (opType == Type::FloatTy)? V9::FSTOI : V9::FDTOI;
1980 assert(destSize == 8 && "Unrecognized type size > 4 and < 8!");
1981 opCode = (opType == Type::FloatTy)? V9::FSTOX : V9::FDTOX;
1987 static MachineInstr*
1988 CreateConvertFPToIntInstr(const TargetMachine& target, Value* srcVal,
1989 Value* destVal, const Type* destType) {
1990 MachineOpCode opCode = ChooseConvertFPToIntInstr(target, destType,
1992 assert(opCode != V9::INVALID_OPCODE && "Expected to need conversion!");
1993 return BuildMI(opCode, 2).addReg(srcVal).addRegDef(destVal);
1996 /// CreateCodeToConvertFloatToInt: Convert FP value to signed or unsigned
1997 /// integer. The FP value must be converted to the dest type in an FP register,
1998 /// and the result is then copied from FP to int register via memory. SPARC
1999 /// does not have a float-to-uint conversion, only a float-to-int (fdtoi).
2000 /// Since fdtoi converts to signed integers, any FP value V between MAXINT+1 and
2001 /// MAXUNSIGNED (i.e., 2^31 <= V <= 2^32-1) would be converted incorrectly.
2002 /// Therefore, for converting an FP value to uint32_t, we first need to convert
2003 /// to uint64_t and then to uint32_t.
2006 CreateCodeToConvertFloatToInt(const TargetMachine& target,
2007 Value* opVal, Instruction* destI,
2008 std::vector<MachineInstr*>& mvec,
2009 MachineCodeForInstruction& mcfi) {
2010 Function* F = destI->getParent()->getParent();
2012 // Create a temporary to represent the FP register into which the
2013 // int value will placed after conversion. The type of this temporary
2014 // depends on the type of FP register to use: single-prec for a 32-bit
2015 // int or smaller; double-prec for a 64-bit int.
2016 size_t destSize = target.getTargetData().getTypeSize(destI->getType());
2018 const Type* castDestType = destI->getType(); // type for the cast instr result
2019 const Type* castDestRegType; // type for cast instruction result reg
2020 TmpInstruction* destForCast; // dest for cast instruction
2021 Instruction* fpToIntCopyDest = destI; // dest for fp-reg-to-int-reg copy instr
2023 // For converting an FP value to uint32_t, we first need to convert to
2024 // uint64_t and then to uint32_t, as explained above.
2025 if (destI->getType() == Type::UIntTy) {
2026 castDestType = Type::ULongTy; // use this instead of type of destI
2027 castDestRegType = Type::DoubleTy; // uint64_t needs 64-bit FP register.
2028 destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
2029 fpToIntCopyDest = new TmpInstruction(mcfi, castDestType, destForCast);
2031 castDestRegType = (destSize > 4)? Type::DoubleTy : Type::FloatTy;
2032 destForCast = new TmpInstruction(mcfi, castDestRegType, opVal);
2035 // Create the fp-to-int conversion instruction (src and dest regs are FP regs)
2036 mvec.push_back(CreateConvertFPToIntInstr(target, opVal, destForCast,
2039 // Create the fpreg-to-intreg copy code
2040 CreateCodeToCopyFloatToInt(target, F, destForCast, fpToIntCopyDest, mvec,
2043 // Create the uint64_t to uint32_t conversion, if needed
2044 if (destI->getType() == Type::UIntTy)
2045 CreateZeroExtensionInstructions(target, F, fpToIntCopyDest, destI,
2046 /*numLowBits*/ 32, mvec, mcfi);
2049 static inline MachineOpCode
2050 ChooseAddInstruction(const InstructionNode* instrNode) {
2051 return ChooseAddInstructionByType(instrNode->getInstruction()->getType());
2054 static inline MachineInstr*
2055 CreateMovFloatInstruction(const InstructionNode* instrNode,
2056 const Type* resultType) {
2057 return BuildMI((resultType == Type::FloatTy) ? V9::FMOVS : V9::FMOVD, 2)
2058 .addReg(instrNode->leftChild()->getValue())
2059 .addRegDef(instrNode->getValue());
2062 static inline MachineInstr*
2063 CreateAddConstInstruction(const InstructionNode* instrNode) {
2064 MachineInstr* minstr = NULL;
2066 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
2067 assert(isa<Constant>(constOp));
2069 // Cases worth optimizing are:
2070 // (1) Add with 0 for float or double: use an FMOV of appropriate type,
2071 // instead of an FADD (1 vs 3 cycles). There is no integer MOV.
2072 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
2073 double dval = FPC->getValue();
2075 minstr = CreateMovFloatInstruction(instrNode,
2076 instrNode->getInstruction()->getType());
2082 static inline MachineOpCode ChooseSubInstructionByType(const Type* resultType) {
2083 MachineOpCode opCode = V9::INVALID_OPCODE;
2085 if (resultType->isInteger() || isa<PointerType>(resultType)) {
2088 switch(resultType->getTypeID()) {
2089 case Type::FloatTyID: opCode = V9::FSUBS; break;
2090 case Type::DoubleTyID: opCode = V9::FSUBD; break;
2091 default: assert(0 && "Invalid type for SUB instruction"); break;
2098 static inline MachineInstr*
2099 CreateSubConstInstruction(const InstructionNode* instrNode) {
2100 MachineInstr* minstr = NULL;
2102 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
2103 assert(isa<Constant>(constOp));
2105 // Cases worth optimizing are:
2106 // (1) Sub with 0 for float or double: use an FMOV of appropriate type,
2107 // instead of an FSUB (1 vs 3 cycles). There is no integer MOV.
2108 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
2109 double dval = FPC->getValue();
2111 minstr = CreateMovFloatInstruction(instrNode,
2112 instrNode->getInstruction()->getType());
2118 static inline MachineOpCode
2119 ChooseFcmpInstruction(const InstructionNode* instrNode) {
2120 MachineOpCode opCode = V9::INVALID_OPCODE;
2122 Value* operand = ((InstrTreeNode*) instrNode->leftChild())->getValue();
2123 switch(operand->getType()->getTypeID()) {
2124 case Type::FloatTyID: opCode = V9::FCMPS; break;
2125 case Type::DoubleTyID: opCode = V9::FCMPD; break;
2126 default: assert(0 && "Invalid type for FCMP instruction"); break;
2132 /// BothFloatToDouble - Assumes that leftArg and rightArg of instrNode are both
2133 /// cast instructions. Returns true if both are floats cast to double.
2135 static inline bool BothFloatToDouble(const InstructionNode* instrNode) {
2136 InstrTreeNode* leftArg = instrNode->leftChild();
2137 InstrTreeNode* rightArg = instrNode->rightChild();
2138 InstrTreeNode* leftArgArg = leftArg->leftChild();
2139 InstrTreeNode* rightArgArg = rightArg->leftChild();
2140 assert(leftArg->getValue()->getType() == rightArg->getValue()->getType());
2141 return (leftArg->getValue()->getType() == Type::DoubleTy &&
2142 leftArgArg->getValue()->getType() == Type::FloatTy &&
2143 rightArgArg->getValue()->getType() == Type::FloatTy);
2146 static inline MachineOpCode ChooseMulInstructionByType(const Type* resultType) {
2147 MachineOpCode opCode = V9::INVALID_OPCODE;
2149 if (resultType->isInteger())
2152 switch(resultType->getTypeID()) {
2153 case Type::FloatTyID: opCode = V9::FMULS; break;
2154 case Type::DoubleTyID: opCode = V9::FMULD; break;
2155 default: assert(0 && "Invalid type for MUL instruction"); break;
2161 static inline MachineInstr*
2162 CreateIntNegInstruction(const TargetMachine& target, Value* vreg) {
2163 return BuildMI(V9::SUBr, 3).addMReg(target.getRegInfo()->getZeroRegNum())
2164 .addReg(vreg).addRegDef(vreg);
2167 /// CreateShiftInstructions - Create instruction sequence for any shift
2168 /// operation. SLL or SLLX on an operand smaller than the integer reg. size
2169 /// (64bits) requires a second instruction for explicit sign-extension. Note
2170 /// that we only have to worry about a sign-bit appearing in the most
2171 /// significant bit of the operand after shifting (e.g., bit 32 of Int or bit 16
2172 /// of Short), so we do not have to worry about results that are as large as a
2173 /// normal integer register.
2176 CreateShiftInstructions(const TargetMachine& target, Function* F,
2177 MachineOpCode shiftOpCode, Value* argVal1,
2178 Value* optArgVal2, /* Use optArgVal2 if not NULL */
2179 unsigned optShiftNum, /* else use optShiftNum */
2180 Instruction* destVal, std::vector<MachineInstr*>& mvec,
2181 MachineCodeForInstruction& mcfi) {
2182 assert((optArgVal2 != NULL || optShiftNum <= 64) &&
2183 "Large shift sizes unexpected, but can be handled below: "
2184 "You need to check whether or not it fits in immed field below");
2186 // If this is a logical left shift of a type smaller than the standard
2187 // integer reg. size, we have to extend the sign-bit into upper bits
2188 // of dest, so we need to put the result of the SLL into a temporary.
2189 Value* shiftDest = destVal;
2190 unsigned opSize = target.getTargetData().getTypeSize(argVal1->getType());
2192 if ((shiftOpCode == V9::SLLr5 || shiftOpCode == V9::SLLXr6) && opSize < 8) {
2193 // put SLL result into a temporary
2194 shiftDest = new TmpInstruction(mcfi, argVal1, optArgVal2, "sllTmp");
2197 MachineInstr* M = (optArgVal2 != NULL)
2198 ? BuildMI(shiftOpCode, 3).addReg(argVal1).addReg(optArgVal2)
2199 .addReg(shiftDest, MachineOperand::Def)
2200 : BuildMI(shiftOpCode, 3).addReg(argVal1).addZImm(optShiftNum)
2201 .addReg(shiftDest, MachineOperand::Def);
2204 if (shiftDest != destVal) {
2205 // extend the sign-bit of the result into all upper bits of dest
2206 assert(8*opSize <= 32 && "Unexpected type size > 4 and < IntRegSize?");
2207 CreateSignExtensionInstructions(target, F, shiftDest, destVal, 8*opSize,
2212 /// CreateMulConstInstruction - Does not create any instructions if we
2213 /// cannot exploit constant to create a cheaper instruction. This returns the
2214 /// approximate cost of the instructions generated, which is used to pick the
2215 /// cheapest when both operands are constant.
2218 CreateMulConstInstruction(const TargetMachine &target, Function* F,
2219 Value* lval, Value* rval, Instruction* destVal,
2220 std::vector<MachineInstr*>& mvec,
2221 MachineCodeForInstruction& mcfi) {
2222 // Use max. multiply cost, viz., cost of MULX
2223 unsigned cost = target.getInstrInfo()->minLatency(V9::MULXr);
2224 unsigned firstNewInstr = mvec.size();
2226 Value* constOp = rval;
2227 if (! isa<Constant>(constOp))
2230 // Cases worth optimizing are:
2231 // (1) Multiply by 0 or 1 for any type: replace with copy (ADD or FMOV)
2232 // (2) Multiply by 2^x for integer types: replace with Shift
2233 const Type* resultType = destVal->getType();
2235 if (resultType->isInteger() || isa<PointerType>(resultType)) {
2237 int64_t C = (int64_t) ConvertConstantToIntType(target, constOp,
2242 bool needNeg = false;
2248 if (C == 0 || C == 1) {
2249 cost = target.getInstrInfo()->minLatency(V9::ADDr);
2250 unsigned Zero = target.getRegInfo()->getZeroRegNum();
2253 M =BuildMI(V9::ADDr,3).addMReg(Zero).addMReg(Zero).addRegDef(destVal);
2255 M = BuildMI(V9::ADDr,3).addReg(lval).addMReg(Zero).addRegDef(destVal);
2257 } else if (isPowerOf2(C, pow)) {
2258 unsigned opSize = target.getTargetData().getTypeSize(resultType);
2259 MachineOpCode opCode = (opSize <= 32)? V9::SLLr5 : V9::SLLXr6;
2260 CreateShiftInstructions(target, F, opCode, lval, NULL, pow,
2261 destVal, mvec, mcfi);
2264 if (mvec.size() > 0 && needNeg) {
2265 // insert <reg = SUB 0, reg> after the instr to flip the sign
2266 MachineInstr* M = CreateIntNegInstruction(target, destVal);
2271 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
2272 double dval = FPC->getValue();
2273 if (fabs(dval) == 1) {
2274 MachineOpCode opCode = (dval < 0)
2275 ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
2276 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
2277 mvec.push_back(BuildMI(opCode,2).addReg(lval).addRegDef(destVal));
2282 if (firstNewInstr < mvec.size()) {
2284 for (unsigned i=firstNewInstr; i < mvec.size(); ++i)
2285 cost += target.getInstrInfo()->minLatency(mvec[i]->getOpcode());
2291 /// CreateCheapestMulConstInstruction - Does not create any instructions
2292 /// if we cannot exploit constant to create a cheaper instruction.
2295 CreateCheapestMulConstInstruction(const TargetMachine &target, Function* F,
2296 Value* lval, Value* rval,
2297 Instruction* destVal,
2298 std::vector<MachineInstr*>& mvec,
2299 MachineCodeForInstruction& mcfi) {
2301 if (isa<Constant>(lval) && isa<Constant>(rval)) {
2302 // both operands are constant: evaluate and "set" in dest
2303 Constant* P = ConstantExpr::get(Instruction::Mul,
2304 cast<Constant>(lval),
2305 cast<Constant>(rval));
2306 CreateCodeToLoadConst (target, F, P, destVal, mvec, mcfi);
2308 else if (isa<Constant>(rval)) // rval is constant, but not lval
2309 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
2310 else if (isa<Constant>(lval)) // lval is constant, but not rval
2311 CreateMulConstInstruction(target, F, lval, rval, destVal, mvec, mcfi);
2313 // else neither is constant
2317 /// CreateMulInstruction - Returns NULL if we cannot exploit constant
2318 /// to create a cheaper instruction.
2321 CreateMulInstruction(const TargetMachine &target, Function* F,
2322 Value* lval, Value* rval, Instruction* destVal,
2323 std::vector<MachineInstr*>& mvec,
2324 MachineCodeForInstruction& mcfi,
2325 MachineOpCode forceMulOp = -1) {
2326 unsigned L = mvec.size();
2327 CreateCheapestMulConstInstruction(target,F, lval, rval, destVal, mvec, mcfi);
2328 if (mvec.size() == L) {
2329 // no instructions were added so create MUL reg, reg, reg.
2330 // Use FSMULD if both operands are actually floats cast to doubles.
2331 // Otherwise, use the default opcode for the appropriate type.
2332 MachineOpCode mulOp = ((forceMulOp != -1)
2334 : ChooseMulInstructionByType(destVal->getType()));
2335 mvec.push_back(BuildMI(mulOp, 3).addReg(lval).addReg(rval)
2336 .addRegDef(destVal));
2340 /// ChooseDivInstruction - Generate a divide instruction for Div or Rem.
2341 /// For Rem, this assumes that the operand type will be signed if the result
2342 /// type is signed. This is correct because they must have the same sign.
2344 static inline MachineOpCode
2345 ChooseDivInstruction(TargetMachine &target, const InstructionNode* instrNode) {
2346 MachineOpCode opCode = V9::INVALID_OPCODE;
2348 const Type* resultType = instrNode->getInstruction()->getType();
2350 if (resultType->isInteger())
2351 opCode = resultType->isSigned()? V9::SDIVXr : V9::UDIVXr;
2353 switch(resultType->getTypeID()) {
2354 case Type::FloatTyID: opCode = V9::FDIVS; break;
2355 case Type::DoubleTyID: opCode = V9::FDIVD; break;
2356 default: assert(0 && "Invalid type for DIV instruction"); break;
2362 /// CreateDivConstInstruction - Return if we cannot exploit constant to create
2363 /// a cheaper instruction.
2365 static void CreateDivConstInstruction(TargetMachine &target,
2366 const InstructionNode* instrNode,
2367 std::vector<MachineInstr*>& mvec) {
2368 Value* LHS = instrNode->leftChild()->getValue();
2369 Value* constOp = ((InstrTreeNode*) instrNode->rightChild())->getValue();
2370 if (!isa<Constant>(constOp))
2373 Instruction* destVal = instrNode->getInstruction();
2374 unsigned ZeroReg = target.getRegInfo()->getZeroRegNum();
2376 // Cases worth optimizing are:
2377 // (1) Divide by 1 for any type: replace with copy (ADD or FMOV)
2378 // (2) Divide by 2^x for integer types: replace with SR[L or A]{X}
2379 const Type* resultType = instrNode->getInstruction()->getType();
2381 if (resultType->isInteger()) {
2384 int64_t C = (int64_t) ConvertConstantToIntType(target, constOp,
2388 bool needNeg = false;
2395 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addMReg(ZeroReg)
2396 .addRegDef(destVal));
2397 } else if (isPowerOf2(C, pow)) {
2399 Value* shiftOperand;
2400 unsigned opSize = target.getTargetData().getTypeSize(resultType);
2402 if (resultType->isSigned()) {
2403 // For N / 2^k, if the operand N is negative,
2404 // we need to add (2^k - 1) before right-shifting by k, i.e.,
2406 // (N / 2^k) = N >> k, if N >= 0;
2407 // (N + 2^k - 1) >> k, if N < 0
2409 // If N is <= 32 bits, use:
2410 // sra N, 31, t1 // t1 = ~0, if N < 0, 0 else
2411 // srl t1, 32-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
2412 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
2413 // sra t3, k, result // result = N / 2^k
2415 // If N is 64 bits, use:
2416 // srax N, k-1, t1 // t1 = sign bit in high k positions
2417 // srlx t1, 64-k, t2 // t2 = 2^k - 1, if N < 0, 0 else
2418 // add t2, N, t3 // t3 = N + 2^k -1, if N < 0, N else
2419 // sra t3, k, result // result = N / 2^k
2420 TmpInstruction *sraTmp, *srlTmp, *addTmp;
2421 MachineCodeForInstruction& mcfi
2422 = MachineCodeForInstruction::get(destVal);
2423 sraTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getSign");
2424 srlTmp = new TmpInstruction(mcfi, resultType, LHS, 0, "getPlus2km1");
2425 addTmp = new TmpInstruction(mcfi, resultType, LHS, srlTmp,"incIfNeg");
2427 // Create the SRA or SRAX instruction to get the sign bit
2428 mvec.push_back(BuildMI((opSize > 4)? V9::SRAXi6 : V9::SRAi5, 3)
2430 .addSImm((resultType==Type::LongTy)? pow-1 : 31)
2431 .addRegDef(sraTmp));
2433 // Create the SRL or SRLX instruction to get the sign bit
2434 mvec.push_back(BuildMI((opSize > 4)? V9::SRLXi6 : V9::SRLi5, 3)
2436 .addSImm((resultType==Type::LongTy)? 64-pow : 32-pow)
2437 .addRegDef(srlTmp));
2439 // Create the ADD instruction to add 2^pow-1 for negative values
2440 mvec.push_back(BuildMI(V9::ADDr, 3).addReg(LHS).addReg(srlTmp)
2441 .addRegDef(addTmp));
2443 // Get the shift operand and "right-shift" opcode to do the divide
2444 shiftOperand = addTmp;
2445 opCode = (opSize > 4)? V9::SRAXi6 : V9::SRAi5;
2447 // Get the shift operand and "right-shift" opcode to do the divide
2449 opCode = (opSize > 4)? V9::SRLXi6 : V9::SRLi5;
2452 // Now do the actual shift!
2453 mvec.push_back(BuildMI(opCode, 3).addReg(shiftOperand).addZImm(pow)
2454 .addRegDef(destVal));
2457 if (needNeg && (C == 1 || isPowerOf2(C, pow))) {
2458 // insert <reg = SUB 0, reg> after the instr to flip the sign
2459 mvec.push_back(CreateIntNegInstruction(target, destVal));
2463 if (ConstantFP *FPC = dyn_cast<ConstantFP>(constOp)) {
2464 double dval = FPC->getValue();
2465 if (fabs(dval) == 1) {
2467 (dval < 0) ? (resultType == Type::FloatTy? V9::FNEGS : V9::FNEGD)
2468 : (resultType == Type::FloatTy? V9::FMOVS : V9::FMOVD);
2470 mvec.push_back(BuildMI(opCode, 2).addReg(LHS).addRegDef(destVal));
2476 static void CreateCodeForVariableSizeAlloca(const TargetMachine& target,
2477 Instruction* result, unsigned tsize,
2478 Value* numElementsVal,
2479 std::vector<MachineInstr*>& getMvec)
2481 Value* totalSizeVal;
2483 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(result);
2484 Function *F = result->getParent()->getParent();
2486 // Enforce the alignment constraints on the stack pointer at
2487 // compile time if the total size is a known constant.
2488 if (isa<Constant>(numElementsVal)) {
2490 int64_t numElem = (int64_t)
2491 ConvertConstantToIntType(target, numElementsVal,
2492 numElementsVal->getType(), isValid);
2493 assert(isValid && "Unexpectedly large array dimension in alloca!");
2494 int64_t total = numElem * tsize;
2495 if (int extra= total % target.getFrameInfo()->getStackFrameSizeAlignment())
2496 total += target.getFrameInfo()->getStackFrameSizeAlignment() - extra;
2497 totalSizeVal = ConstantSInt::get(Type::IntTy, total);
2499 // The size is not a constant. Generate code to compute it and
2500 // code to pad the size for stack alignment.
2501 // Create a Value to hold the (constant) element size
2502 Value* tsizeVal = ConstantSInt::get(Type::IntTy, tsize);
2504 // Create temporary values to hold the result of MUL, SLL, SRL
2505 // To pad `size' to next smallest multiple of 16:
2506 // size = (size + 15) & (-16 = 0xfffffffffffffff0)
2507 TmpInstruction* tmpProd = new TmpInstruction(mcfi,numElementsVal, tsizeVal);
2508 TmpInstruction* tmpAdd15= new TmpInstruction(mcfi,numElementsVal, tmpProd);
2509 TmpInstruction* tmpAndf0= new TmpInstruction(mcfi,numElementsVal, tmpAdd15);
2511 // Instruction 1: mul numElements, typeSize -> tmpProd
2512 // This will optimize the MUL as far as possible.
2513 CreateMulInstruction(target, F, numElementsVal, tsizeVal, tmpProd, getMvec,
2516 // Instruction 2: andn tmpProd, 0x0f -> tmpAndn
2517 getMvec.push_back(BuildMI(V9::ADDi, 3).addReg(tmpProd).addSImm(15)
2518 .addReg(tmpAdd15, MachineOperand::Def));
2520 // Instruction 3: add tmpAndn, 0x10 -> tmpAdd16
2521 getMvec.push_back(BuildMI(V9::ANDi, 3).addReg(tmpAdd15).addSImm(-16)
2522 .addReg(tmpAndf0, MachineOperand::Def));
2524 totalSizeVal = tmpAndf0;
2527 // Get the constant offset from SP for dynamically allocated storage
2528 // and create a temporary Value to hold it.
2529 MachineFunction& mcInfo = MachineFunction::get(F);
2531 ConstantSInt* dynamicAreaOffset =
2532 ConstantSInt::get(Type::IntTy,
2533 target.getFrameInfo()->getDynamicAreaOffset(mcInfo,growUp));
2534 assert(! growUp && "Has SPARC v9 stack frame convention changed?");
2536 unsigned SPReg = target.getRegInfo()->getStackPointer();
2538 // Instruction 2: sub %sp, totalSizeVal -> %sp
2539 getMvec.push_back(BuildMI(V9::SUBr, 3).addMReg(SPReg).addReg(totalSizeVal)
2540 .addMReg(SPReg,MachineOperand::Def));
2542 // Instruction 3: add %sp, frameSizeBelowDynamicArea -> result
2543 getMvec.push_back(BuildMI(V9::ADDr,3).addMReg(SPReg).addReg(dynamicAreaOffset)
2544 .addRegDef(result));
2548 CreateCodeForFixedSizeAlloca(const TargetMachine& target,
2549 Instruction* result, unsigned tsize,
2550 unsigned numElements,
2551 std::vector<MachineInstr*>& getMvec) {
2552 assert(result && result->getParent() &&
2553 "Result value is not part of a function?");
2554 Function *F = result->getParent()->getParent();
2555 MachineFunction &mcInfo = MachineFunction::get(F);
2557 // If the alloca is of zero bytes (which is perfectly legal) we bump it up to
2558 // one byte. This is unnecessary, but I really don't want to break any
2559 // fragile logic in this code. FIXME.
2563 // Put the variable in the dynamically sized area of the frame if either:
2564 // (a) The offset is too large to use as an immediate in load/stores
2565 // (check LDX because all load/stores have the same-size immed. field).
2566 // (b) The object is "large", so it could cause many other locals,
2567 // spills, and temporaries to have large offsets.
2568 // NOTE: We use LARGE = 8 * argSlotSize = 64 bytes.
2569 // You've gotta love having only 13 bits for constant offset values :-|.
2571 unsigned paddedSize;
2572 int offsetFromFP = mcInfo.getInfo()->computeOffsetforLocalVar(result,
2574 tsize * numElements);
2576 if (((int)paddedSize) > 8 * target.getFrameInfo()->getSizeOfEachArgOnStack()||
2577 !target.getInstrInfo()->constantFitsInImmedField(V9::LDXi,offsetFromFP)) {
2578 CreateCodeForVariableSizeAlloca(target, result, tsize,
2579 ConstantSInt::get(Type::IntTy,numElements),
2584 // else offset fits in immediate field so go ahead and allocate it.
2585 offsetFromFP = mcInfo.getInfo()->allocateLocalVar(result, tsize *numElements);
2587 // Create a temporary Value to hold the constant offset.
2588 // This is needed because it may not fit in the immediate field.
2589 ConstantSInt* offsetVal = ConstantSInt::get(Type::IntTy, offsetFromFP);
2591 // Instruction 1: add %fp, offsetFromFP -> result
2592 unsigned FPReg = target.getRegInfo()->getFramePointer();
2593 getMvec.push_back(BuildMI(V9::ADDr, 3).addMReg(FPReg).addReg(offsetVal)
2594 .addRegDef(result));
2597 /// SetOperandsForMemInstr - Choose addressing mode for the given load or store
2598 /// instruction. Use [reg+reg] if it is an indexed reference, and the index
2599 /// offset is not a constant or if it cannot fit in the offset field. Use
2600 /// [reg+offset] in all other cases. This assumes that all array refs are
2601 /// "lowered" to one of these forms:
2602 /// %x = load (subarray*) ptr, constant ; single constant offset
2603 /// %x = load (subarray*) ptr, offsetVal ; single non-constant offset
2604 /// Generally, this should happen via strength reduction + LICM. Also, strength
2605 /// reduction should take care of using the same register for the loop index
2606 /// variable and an array index, when that is profitable.
2608 static void SetOperandsForMemInstr(unsigned Opcode,
2609 std::vector<MachineInstr*>& mvec,
2610 InstructionNode* vmInstrNode,
2611 const TargetMachine& target) {
2612 Instruction* memInst = vmInstrNode->getInstruction();
2613 // Index vector, ptr value, and flag if all indices are const.
2614 std::vector<Value*> idxVec;
2615 bool allConstantIndices;
2616 Value* ptrVal = GetMemInstArgs(vmInstrNode, idxVec, allConstantIndices);
2618 // Now create the appropriate operands for the machine instruction.
2619 // First, initialize so we default to storing the offset in a register.
2620 int64_t smallConstOffset = 0;
2621 Value* valueForRegOffset = NULL;
2622 MachineOperand::MachineOperandType offsetOpType =
2623 MachineOperand::MO_VirtualRegister;
2625 // Check if there is an index vector and if so, compute the
2626 // right offset for structures and for arrays
2627 if (!idxVec.empty()) {
2628 const PointerType* ptrType = cast<PointerType>(ptrVal->getType());
2630 // If all indices are constant, compute the combined offset directly.
2631 if (allConstantIndices) {
2632 // Compute the offset value using the index vector. Create a
2633 // virtual reg. for it since it may not fit in the immed field.
2634 uint64_t offset = target.getTargetData().getIndexedOffset(ptrType,idxVec);
2635 valueForRegOffset = ConstantSInt::get(Type::LongTy, offset);
2637 // There is at least one non-constant offset. Therefore, this must
2638 // be an array ref, and must have been lowered to a single non-zero
2639 // offset. (An extra leading zero offset, if any, can be ignored.)
2640 // Generate code sequence to compute address from index.
2641 bool firstIdxIsZero = IsZero(idxVec[0]);
2642 assert(idxVec.size() == 1U + firstIdxIsZero
2643 && "Array refs must be lowered before Instruction Selection");
2645 Value* idxVal = idxVec[firstIdxIsZero];
2647 std::vector<MachineInstr*> mulVec;
2649 new TmpInstruction(MachineCodeForInstruction::get(memInst),
2650 Type::ULongTy, memInst);
2652 // Get the array type indexed by idxVal, and compute its element size.
2653 // The call to getTypeSize() will fail if size is not constant.
2654 const Type* vecType = (firstIdxIsZero
2655 ? GetElementPtrInst::getIndexedType(ptrType,
2656 std::vector<Value*>(1U, idxVec[0]),
2657 /*AllowCompositeLeaf*/ true)
2659 const Type* eltType = cast<SequentialType>(vecType)->getElementType();
2660 ConstantUInt* eltSizeVal = ConstantUInt::get(Type::ULongTy,
2661 target.getTargetData().getTypeSize(eltType));
2663 // CreateMulInstruction() folds constants intelligently enough.
2664 CreateMulInstruction(target, memInst->getParent()->getParent(),
2665 idxVal, /* lval, not likely to be const*/
2666 eltSizeVal, /* rval, likely to be constant */
2668 mulVec, MachineCodeForInstruction::get(memInst),
2671 assert(mulVec.size() > 0 && "No multiply code created?");
2672 mvec.insert(mvec.end(), mulVec.begin(), mulVec.end());
2674 valueForRegOffset = addr;
2677 offsetOpType = MachineOperand::MO_SignExtendedImmed;
2678 smallConstOffset = 0;
2682 // Operand 0 is value, operand 1 is ptr, operand 2 is offset
2683 // For LOAD or GET_ELEMENT_PTR,
2684 // Operand 0 is ptr, operand 1 is offset, operand 2 is result.
2685 unsigned offsetOpNum, ptrOpNum;
2687 if (memInst->getOpcode() == Instruction::Store) {
2688 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
2689 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
2690 .addReg(ptrVal).addReg(valueForRegOffset);
2692 Opcode = convertOpcodeFromRegToImm(Opcode);
2693 MI = BuildMI(Opcode, 3).addReg(vmInstrNode->leftChild()->getValue())
2694 .addReg(ptrVal).addSImm(smallConstOffset);
2697 if (offsetOpType == MachineOperand::MO_VirtualRegister) {
2698 MI = BuildMI(Opcode, 3).addReg(ptrVal).addReg(valueForRegOffset)
2699 .addRegDef(memInst);
2701 Opcode = convertOpcodeFromRegToImm(Opcode);
2702 MI = BuildMI(Opcode, 3).addReg(ptrVal).addSImm(smallConstOffset)
2703 .addRegDef(memInst);
2709 /// ForwardOperand - Substitute operand `operandNum' of the instruction in
2710 /// node `treeNode' in place of the use(s) of that instruction in node `parent'.
2711 /// Check both explicit and implicit operands! Also make sure to skip over a
2712 /// parent who: (1) is a list node in the Burg tree, or (2) itself had its
2713 /// results forwarded to its parent.
2715 static void ForwardOperand (InstructionNode *treeNode, InstrTreeNode *parent,
2717 assert(treeNode && parent && "Invalid invocation of ForwardOperand");
2719 Instruction* unusedOp = treeNode->getInstruction();
2720 Value* fwdOp = unusedOp->getOperand(operandNum);
2722 // The parent itself may be a list node, so find the real parent instruction
2723 while (parent->getNodeType() != InstrTreeNode::NTInstructionNode) {
2724 parent = parent->parent();
2725 assert(parent && "ERROR: Non-instruction node has no parent in tree.");
2727 InstructionNode* parentInstrNode = (InstructionNode*) parent;
2729 Instruction* userInstr = parentInstrNode->getInstruction();
2730 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(userInstr);
2732 // The parent's mvec would be empty if it was itself forwarded.
2733 // Recursively call ForwardOperand in that case...
2735 if (mvec.size() == 0) {
2736 assert(parent->parent() != NULL &&
2737 "Parent could not have been forwarded, yet has no instructions?");
2738 ForwardOperand(treeNode, parent->parent(), operandNum);
2740 for (unsigned i=0, N=mvec.size(); i < N; i++) {
2741 MachineInstr* minstr = mvec[i];
2742 for (unsigned i=0, numOps=minstr->getNumOperands(); i < numOps; ++i) {
2743 const MachineOperand& mop = minstr->getOperand(i);
2744 if (mop.getType() == MachineOperand::MO_VirtualRegister &&
2745 mop.getVRegValue() == unusedOp) {
2746 minstr->SetMachineOperandVal(i, MachineOperand::MO_VirtualRegister,
2751 for (unsigned i=0,numOps=minstr->getNumImplicitRefs(); i<numOps; ++i)
2752 if (minstr->getImplicitRef(i) == unusedOp)
2753 minstr->setImplicitRef(i, fwdOp);
2758 /// AllUsesAreBranches - Returns true if all the uses of I are
2759 /// Branch instructions, false otherwise.
2761 inline bool AllUsesAreBranches(const Instruction* I) {
2762 for (Value::use_const_iterator UI=I->use_begin(), UE=I->use_end();
2764 if (! isa<TmpInstruction>(*UI) // ignore tmp instructions here
2765 && cast<Instruction>(*UI)->getOpcode() != Instruction::Br)
2770 /// CodeGenIntrinsic - Generate code for any intrinsic that needs a special
2771 /// code sequence instead of a regular call. If not that kind of intrinsic, do
2772 /// nothing. Returns true if code was generated, otherwise false.
2774 static bool CodeGenIntrinsic(Intrinsic::ID iid, CallInst &callInstr,
2775 TargetMachine &target,
2776 std::vector<MachineInstr*>& mvec) {
2779 assert(0 && "Unknown intrinsic function call should have been lowered!");
2780 case Intrinsic::vastart: {
2781 // Get the address of the first incoming vararg argument on the stack
2783 Function* func = cast<Function>(callInstr.getParent()->getParent());
2784 int numFixedArgs = func->getFunctionType()->getNumParams();
2785 int fpReg = target.getFrameInfo()->getIncomingArgBaseRegNum();
2786 int argSize = target.getFrameInfo()->getSizeOfEachArgOnStack();
2787 int firstVarArgOff = numFixedArgs * argSize + target.getFrameInfo()->
2788 getFirstIncomingArgOffset(MachineFunction::get(func), ignore);
2789 mvec.push_back(BuildMI(V9::ADDi, 3).addMReg(fpReg).addSImm(firstVarArgOff).
2790 addRegDef(&callInstr));
2794 case Intrinsic::vaend:
2795 return true; // no-op on SparcV9
2797 case Intrinsic::vacopy:
2798 // Simple copy of current va_list (arg1) to new va_list (result)
2799 mvec.push_back(BuildMI(V9::ORr, 3).
2800 addMReg(target.getRegInfo()->getZeroRegNum()).
2801 addReg(callInstr.getOperand(1)).
2802 addRegDef(&callInstr));
2807 /// ThisIsAChainRule - returns true if the given BURG rule is a chain rule.
2809 extern bool ThisIsAChainRule(int eruleno) {
2811 case 111: // stmt: reg
2835 return false; break;
2839 /// GetInstructionsByRule - Choose machine instructions for the
2840 /// SPARC V9 according to the patterns chosen by the BURG-generated parser.
2841 /// This is where most of the work in the V9 instruction selector gets done.
2843 void GetInstructionsByRule(InstructionNode* subtreeRoot, int ruleForNode,
2844 short* nts, TargetMachine &target,
2845 std::vector<MachineInstr*>& mvec) {
2846 bool checkCast = false; // initialize here to use fall-through
2847 bool maskUnsignedResult = false;
2849 int forwardOperandNum = -1;
2850 unsigned allocaSize = 0;
2851 MachineInstr* M, *M2;
2853 bool foldCase = false;
2857 // If the code for this instruction was folded into the parent (user),
2859 if (subtreeRoot->isFoldedIntoParent())
2862 // Let's check for chain rules outside the switch so that we don't have
2863 // to duplicate the list of chain rule production numbers here again
2864 if (ThisIsAChainRule(ruleForNode)) {
2865 // Chain rules have a single nonterminal on the RHS.
2866 // Get the rule that matches the RHS non-terminal and use that instead.
2867 assert(nts[0] && ! nts[1]
2868 && "A chain rule should have only one RHS non-terminal!");
2869 nextRule = burm_rule(subtreeRoot->state, nts[0]);
2870 nts = burm_nts[nextRule];
2871 GetInstructionsByRule(subtreeRoot, nextRule, nts, target, mvec);
2873 switch(ruleForNode) {
2874 case 1: // stmt: Ret
2875 case 2: // stmt: RetValue(reg)
2876 { // NOTE: Prepass of register allocation is responsible
2877 // for moving return value to appropriate register.
2878 // Copy the return value to the required return register.
2879 // Mark the return Value as an implicit ref of the RET instr..
2880 // Mark the return-address register as a hidden virtual reg.
2881 // Finally put a NOP in the delay slot.
2882 ReturnInst *returnInstr=cast<ReturnInst>(subtreeRoot->getInstruction());
2883 Value* retVal = returnInstr->getReturnValue();
2884 MachineCodeForInstruction& mcfi =
2885 MachineCodeForInstruction::get(returnInstr);
2887 // Create a hidden virtual reg to represent the return address register
2888 // used by the machine instruction but not represented in LLVM.
2889 Instruction* returnAddrTmp = new TmpInstruction(mcfi, returnInstr);
2891 MachineInstr* retMI =
2892 BuildMI(V9::JMPLRETi, 3).addReg(returnAddrTmp).addSImm(8)
2893 .addMReg(target.getRegInfo()->getZeroRegNum(), MachineOperand::Def);
2895 // If there is a value to return, we need to:
2896 // (a) Sign-extend the value if it is smaller than 8 bytes (reg size)
2897 // (b) Insert a copy to copy the return value to the appropriate reg.
2898 // -- For FP values, create a FMOVS or FMOVD instruction
2899 // -- For non-FP values, create an add-with-0 instruction
2900 if (retVal != NULL) {
2901 const SparcV9RegInfo& regInfo =
2902 (SparcV9RegInfo&) *target.getRegInfo();
2903 const Type* retType = retVal->getType();
2904 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
2905 unsigned retRegNum = (retType->isFloatingPoint()
2906 ? (unsigned) SparcV9FloatRegClass::f0
2907 : (unsigned) SparcV9IntRegClass::i0);
2908 retRegNum = regInfo.getUnifiedRegNum(regClassID, retRegNum);
2910 // Insert sign-extension instructions for small signed values.
2911 Value* retValToUse = retVal;
2912 if (retType->isIntegral() && retType->isSigned()) {
2913 unsigned retSize = target.getTargetData().getTypeSize(retType);
2915 // Create a temporary virtual reg. to hold the sign-extension.
2916 retValToUse = new TmpInstruction(mcfi, retVal);
2918 // Sign-extend retVal and put the result in the temporary reg.
2919 CreateSignExtensionInstructions
2920 (target, returnInstr->getParent()->getParent(),
2921 retVal, retValToUse, 8*retSize, mvec, mcfi);
2925 // (b) Now, insert a copy to to the appropriate register:
2926 // -- For FP values, create a FMOVS or FMOVD instruction
2927 // -- For non-FP values, create an add-with-0 instruction
2928 // First, create a virtual register to represent the register and
2929 // mark this vreg as being an implicit operand of the ret MI.
2930 TmpInstruction* retVReg =
2931 new TmpInstruction(mcfi, retValToUse, NULL, "argReg");
2933 retMI->addImplicitRef(retVReg);
2935 if (retType->isFloatingPoint())
2936 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
2937 .addReg(retValToUse).addReg(retVReg, MachineOperand::Def));
2939 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
2940 .addReg(retValToUse).addSImm((int64_t) 0)
2941 .addReg(retVReg, MachineOperand::Def));
2943 // Mark the operand with the register it should be assigned
2944 M->SetRegForOperand(M->getNumOperands()-1, retRegNum);
2945 retMI->SetRegForImplicitRef(retMI->getNumImplicitRefs()-1, retRegNum);
2950 // Now insert the RET instruction and a NOP for the delay slot
2951 mvec.push_back(retMI);
2952 mvec.push_back(BuildMI(V9::NOP, 0));
2957 case 3: // stmt: Store(reg,reg)
2958 case 4: // stmt: Store(reg,ptrreg)
2959 SetOperandsForMemInstr(ChooseStoreInstruction(
2960 subtreeRoot->leftChild()->getValue()->getType()),
2961 mvec, subtreeRoot, target);
2964 case 5: // stmt: BrUncond
2966 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
2967 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(0)));
2970 mvec.push_back(BuildMI(V9::NOP, 0));
2974 case 206: // stmt: BrCond(setCCconst)
2975 { // setCCconst => boolean was computed with `%b = setCC type reg1 const'
2976 // If the constant is ZERO, we can use the branch-on-integer-register
2977 // instructions and avoid the SUBcc instruction entirely.
2978 // Otherwise this is just the same as case 5, so just fall through.
2980 InstrTreeNode* constNode = subtreeRoot->leftChild()->rightChild();
2982 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
2983 Constant *constVal = cast<Constant>(constNode->getValue());
2986 if ((constVal->getType()->isInteger()
2987 || isa<PointerType>(constVal->getType()))
2988 && ConvertConstantToIntType(target,
2989 constVal, constVal->getType(), isValidConst) == 0
2992 // That constant is a zero after all...
2993 // Use the left child of setCC as the first argument!
2994 // Mark the setCC node so that no code is generated for it.
2995 InstructionNode* setCCNode = (InstructionNode*)
2996 subtreeRoot->leftChild();
2997 assert(setCCNode->getOpLabel() == SetCCOp);
2998 setCCNode->markFoldedIntoParent();
3000 BranchInst* brInst=cast<BranchInst>(subtreeRoot->getInstruction());
3002 M = BuildMI(ChooseBprInstruction(subtreeRoot), 2)
3003 .addReg(setCCNode->leftChild()->getValue())
3004 .addPCDisp(brInst->getSuccessor(0));
3008 mvec.push_back(BuildMI(V9::NOP, 0));
3011 mvec.push_back(BuildMI(V9::BA, 1)
3012 .addPCDisp(brInst->getSuccessor(1)));
3015 mvec.push_back(BuildMI(V9::NOP, 0));
3018 // ELSE FALL THROUGH
3021 case 6: // stmt: BrCond(setCC)
3022 { // bool => boolean was computed with SetCC.
3023 // The branch to use depends on whether it is FP, signed, or unsigned.
3024 // If it is an integer CC, we also need to find the unique
3025 // TmpInstruction representing that CC.
3027 BranchInst* brInst = cast<BranchInst>(subtreeRoot->getInstruction());
3028 const Type* setCCType;
3029 unsigned Opcode = ChooseBccInstruction(subtreeRoot, setCCType);
3030 Value* ccValue = GetTmpForCC(subtreeRoot->leftChild()->getValue(),
3031 brInst->getParent()->getParent(),
3033 MachineCodeForInstruction::get(brInst));
3034 M = BuildMI(Opcode, 2).addCCReg(ccValue)
3035 .addPCDisp(brInst->getSuccessor(0));
3039 mvec.push_back(BuildMI(V9::NOP, 0));
3042 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(brInst->getSuccessor(1)));
3045 mvec.push_back(BuildMI(V9::NOP, 0));
3049 case 208: // stmt: BrCond(boolconst)
3051 // boolconst => boolean is a constant; use BA to first or second label
3052 Constant* constVal =
3053 cast<Constant>(subtreeRoot->leftChild()->getValue());
3054 unsigned dest = cast<ConstantBool>(constVal)->getValue()? 0 : 1;
3056 M = BuildMI(V9::BA, 1).addPCDisp(
3057 cast<BranchInst>(subtreeRoot->getInstruction())->getSuccessor(dest));
3061 mvec.push_back(BuildMI(V9::NOP, 0));
3065 case 8: // stmt: BrCond(boolreg)
3066 { // boolreg => boolean is recorded in an integer register.
3067 // Use branch-on-integer-register instruction.
3069 BranchInst *BI = cast<BranchInst>(subtreeRoot->getInstruction());
3070 M = BuildMI(V9::BRNZ, 2).addReg(subtreeRoot->leftChild()->getValue())
3071 .addPCDisp(BI->getSuccessor(0));
3075 mvec.push_back(BuildMI(V9::NOP, 0));
3078 mvec.push_back(BuildMI(V9::BA, 1).addPCDisp(BI->getSuccessor(1)));
3081 mvec.push_back(BuildMI(V9::NOP, 0));
3085 case 9: // stmt: Switch(reg)
3086 assert(0 && "*** SWITCH instruction is not implemented yet.");
3089 case 10: // reg: VRegList(reg, reg)
3090 assert(0 && "VRegList should never be the topmost non-chain rule");
3093 case 21: // bool: Not(bool,reg): Compute with a conditional-move-on-reg
3094 { // First find the unary operand. It may be left or right, usually right.
3095 Instruction* notI = subtreeRoot->getInstruction();
3096 Value* notArg = BinaryOperator::getNotArgument(
3097 cast<BinaryOperator>(subtreeRoot->getInstruction()));
3098 unsigned ZeroReg = target.getRegInfo()->getZeroRegNum();
3100 // Unconditionally set register to 0
3101 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(notI));
3103 // Now conditionally move 1 into the register.
3104 // Mark the register as a use (as well as a def) because the old
3105 // value will be retained if the condition is false.
3106 mvec.push_back(BuildMI(V9::MOVRZi, 3).addReg(notArg).addZImm(1)
3107 .addReg(notI, MachineOperand::UseAndDef));
3112 case 421: // reg: BNot(reg,reg): Compute as reg = reg XOR-NOT 0
3113 { // First find the unary operand. It may be left or right, usually right.
3114 Value* notArg = BinaryOperator::getNotArgument(
3115 cast<BinaryOperator>(subtreeRoot->getInstruction()));
3116 unsigned ZeroReg = target.getRegInfo()->getZeroRegNum();
3117 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(notArg).addMReg(ZeroReg)
3118 .addRegDef(subtreeRoot->getValue()));
3122 case 322: // reg: Not(tobool, reg):
3123 // Fold CAST-TO-BOOL with NOT by inverting the sense of cast-to-bool
3125 // Just fall through!
3127 case 22: // reg: ToBoolTy(reg):
3129 Instruction* castI = subtreeRoot->getInstruction();
3130 Value* opVal = subtreeRoot->leftChild()->getValue();
3131 assert(opVal->getType()->isIntegral() ||
3132 isa<PointerType>(opVal->getType()));
3134 // Unconditionally set register to 0
3135 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(castI));
3137 // Now conditionally move 1 into the register.
3138 // Mark the register as a use (as well as a def) because the old
3139 // value will be retained if the condition is false.
3140 MachineOpCode opCode = foldCase? V9::MOVRZi : V9::MOVRNZi;
3141 mvec.push_back(BuildMI(opCode, 3).addReg(opVal).addZImm(1)
3142 .addReg(castI, MachineOperand::UseAndDef));
3147 case 23: // reg: ToUByteTy(reg)
3148 case 24: // reg: ToSByteTy(reg)
3149 case 25: // reg: ToUShortTy(reg)
3150 case 26: // reg: ToShortTy(reg)
3151 case 27: // reg: ToUIntTy(reg)
3152 case 28: // reg: ToIntTy(reg)
3153 case 29: // reg: ToULongTy(reg)
3154 case 30: // reg: ToLongTy(reg)
3156 //======================================================================
3157 // Rules for integer conversions:
3160 // From ISO 1998 C++ Standard, Sec. 4.7:
3162 // 2. If the destination type is unsigned, the resulting value is
3163 // the least unsigned integer congruent to the source integer
3164 // (modulo 2n where n is the number of bits used to represent the
3165 // unsigned type). [Note: In a two s complement representation,
3166 // this conversion is conceptual and there is no change in the
3167 // bit pattern (if there is no truncation). ]
3169 // 3. If the destination type is signed, the value is unchanged if
3170 // it can be represented in the destination type (and bitfield width);
3171 // otherwise, the value is implementation-defined.
3174 // Since we assume 2s complement representations, this implies:
3176 // -- If operand is smaller than destination, zero-extend or sign-extend
3177 // according to the signedness of the *operand*: source decides:
3178 // (1) If operand is signed, sign-extend it.
3179 // If dest is unsigned, zero-ext the result!
3180 // (2) If operand is unsigned, our current invariant is that
3181 // it's high bits are correct, so zero-extension is not needed.
3183 // -- If operand is same size as or larger than destination,
3184 // zero-extend or sign-extend according to the signedness of
3185 // the *destination*: destination decides:
3186 // (1) If destination is signed, sign-extend (truncating if needed)
3187 // This choice is implementation defined. We sign-extend the
3188 // operand, which matches both Sun's cc and gcc3.2.
3189 // (2) If destination is unsigned, zero-extend (truncating if needed)
3190 //======================================================================
3192 Instruction* destI = subtreeRoot->getInstruction();
3193 Function* currentFunc = destI->getParent()->getParent();
3194 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(destI);
3196 Value* opVal = subtreeRoot->leftChild()->getValue();
3197 const Type* opType = opVal->getType();
3198 const Type* destType = destI->getType();
3199 unsigned opSize = target.getTargetData().getTypeSize(opType);
3200 unsigned destSize = target.getTargetData().getTypeSize(destType);
3202 bool isIntegral = opType->isIntegral() || isa<PointerType>(opType);
3204 if (opType == Type::BoolTy ||
3205 opType == destType ||
3206 isIntegral && opSize == destSize && opSize == 8) {
3207 // nothing to do in all these cases
3208 forwardOperandNum = 0; // forward first operand to user
3210 } else if (opType->isFloatingPoint()) {
3212 CreateCodeToConvertFloatToInt(target, opVal, destI, mvec, mcfi);
3213 if (destI->getType()->isUnsigned() && destI->getType() !=Type::UIntTy)
3214 maskUnsignedResult = true; // not handled by fp->int code
3216 } else if (isIntegral) {
3218 bool opSigned = opType->isSigned();
3219 bool destSigned = destType->isSigned();
3220 unsigned extSourceInBits = 8 * std::min<unsigned>(opSize, destSize);
3222 assert(! (opSize == destSize && opSigned == destSigned) &&
3223 "How can different int types have same size and signedness?");
3225 bool signExtend = (opSize < destSize && opSigned ||
3226 opSize >= destSize && destSigned);
3228 bool signAndZeroExtend = (opSize < destSize && destSize < 8u &&
3229 opSigned && !destSigned);
3230 assert(!signAndZeroExtend || signExtend);
3232 bool zeroExtendOnly = opSize >= destSize && !destSigned;
3233 assert(!zeroExtendOnly || !signExtend);
3236 Value* signExtDest = (signAndZeroExtend
3237 ? new TmpInstruction(mcfi, destType, opVal)
3240 CreateSignExtensionInstructions
3241 (target, currentFunc,opVal,signExtDest,extSourceInBits,mvec,mcfi);
3243 if (signAndZeroExtend)
3244 CreateZeroExtensionInstructions
3245 (target, currentFunc, signExtDest, destI, 8*destSize, mvec, mcfi);
3247 else if (zeroExtendOnly) {
3248 CreateZeroExtensionInstructions
3249 (target, currentFunc, opVal, destI, extSourceInBits, mvec, mcfi);
3252 forwardOperandNum = 0; // forward first operand to user
3255 assert(0 && "Unrecognized operand type for convert-to-integer");
3260 case 31: // reg: ToFloatTy(reg):
3261 case 32: // reg: ToDoubleTy(reg):
3262 case 232: // reg: ToDoubleTy(Constant):
3264 // If this instruction has a parent (a user) in the tree
3265 // and the user is translated as an FsMULd instruction,
3266 // then the cast is unnecessary. So check that first.
3267 // In the future, we'll want to do the same for the FdMULq instruction,
3268 // so do the check here instead of only for ToFloatTy(reg).
3270 if (subtreeRoot->parent() != NULL) {
3271 const MachineCodeForInstruction& mcfi =
3272 MachineCodeForInstruction::get(
3273 cast<InstructionNode>(subtreeRoot->parent())->getInstruction());
3274 if (mcfi.size() == 0 || mcfi.front()->getOpcode() == V9::FSMULD)
3275 forwardOperandNum = 0; // forward first operand to user
3278 if (forwardOperandNum != 0) { // we do need the cast
3279 Value* leftVal = subtreeRoot->leftChild()->getValue();
3280 const Type* opType = leftVal->getType();
3281 MachineOpCode opCode=ChooseConvertToFloatInstr(target,
3282 subtreeRoot->getOpLabel(), opType);
3283 if (opCode == V9::NOP) { // no conversion needed
3284 forwardOperandNum = 0; // forward first operand to user
3286 // If the source operand is a non-FP type it must be
3287 // first copied from int to float register via memory!
3288 Instruction *dest = subtreeRoot->getInstruction();
3291 if (! opType->isFloatingPoint()) {
3292 // Create a temporary to represent the FP register
3293 // into which the integer will be copied via memory.
3294 // The type of this temporary will determine the FP
3295 // register used: single-prec for a 32-bit int or smaller,
3296 // double-prec for a 64-bit int.
3299 target.getTargetData().getTypeSize(leftVal->getType());
3300 Type* tmpTypeToUse =
3301 (srcSize <= 4)? Type::FloatTy : Type::DoubleTy;
3302 MachineCodeForInstruction &destMCFI =
3303 MachineCodeForInstruction::get(dest);
3304 srcForCast = new TmpInstruction(destMCFI, tmpTypeToUse, dest);
3306 CreateCodeToCopyIntToFloat(target,
3307 dest->getParent()->getParent(),
3308 leftVal, cast<Instruction>(srcForCast),
3311 srcForCast = leftVal;
3313 M = BuildMI(opCode, 2).addReg(srcForCast).addRegDef(dest);
3319 case 19: // reg: ToArrayTy(reg):
3320 case 20: // reg: ToPointerTy(reg):
3321 forwardOperandNum = 0; // forward first operand to user
3324 case 233: // reg: Add(reg, Constant)
3325 maskUnsignedResult = true;
3326 M = CreateAddConstInstruction(subtreeRoot);
3331 // ELSE FALL THROUGH
3333 case 33: // reg: Add(reg, reg)
3334 maskUnsignedResult = true;
3335 Add3OperandInstr(ChooseAddInstruction(subtreeRoot), subtreeRoot, mvec);
3338 case 234: // reg: Sub(reg, Constant)
3339 maskUnsignedResult = true;
3340 M = CreateSubConstInstruction(subtreeRoot);
3345 // ELSE FALL THROUGH
3347 case 34: // reg: Sub(reg, reg)
3348 maskUnsignedResult = true;
3349 Add3OperandInstr(ChooseSubInstructionByType(
3350 subtreeRoot->getInstruction()->getType()),
3354 case 135: // reg: Mul(todouble, todouble)
3358 case 35: // reg: Mul(reg, reg)
3360 maskUnsignedResult = true;
3361 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
3362 ? (MachineOpCode)V9::FSMULD
3364 Instruction* mulInstr = subtreeRoot->getInstruction();
3365 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
3366 subtreeRoot->leftChild()->getValue(),
3367 subtreeRoot->rightChild()->getValue(),
3369 MachineCodeForInstruction::get(mulInstr),forceOp);
3372 case 335: // reg: Mul(todouble, todoubleConst)
3376 case 235: // reg: Mul(reg, Constant)
3378 maskUnsignedResult = true;
3379 MachineOpCode forceOp = ((checkCast && BothFloatToDouble(subtreeRoot))
3380 ? (MachineOpCode)V9::FSMULD
3382 Instruction* mulInstr = subtreeRoot->getInstruction();
3383 CreateMulInstruction(target, mulInstr->getParent()->getParent(),
3384 subtreeRoot->leftChild()->getValue(),
3385 subtreeRoot->rightChild()->getValue(),
3387 MachineCodeForInstruction::get(mulInstr),
3391 case 236: // reg: Div(reg, Constant)
3392 maskUnsignedResult = true;
3394 CreateDivConstInstruction(target, subtreeRoot, mvec);
3395 if (mvec.size() > L)
3397 // ELSE FALL THROUGH
3399 case 36: // reg: Div(reg, reg)
3401 maskUnsignedResult = true;
3403 // If either operand of divide is smaller than 64 bits, we have
3404 // to make sure the unused top bits are correct because they affect
3405 // the result. These bits are already correct for unsigned values.
3406 // They may be incorrect for signed values, so sign extend to fill in.
3407 Instruction* divI = subtreeRoot->getInstruction();
3408 Value* divOp1 = subtreeRoot->leftChild()->getValue();
3409 Value* divOp2 = subtreeRoot->rightChild()->getValue();
3410 Value* divOp1ToUse = divOp1;
3411 Value* divOp2ToUse = divOp2;
3412 if (divI->getType()->isSigned()) {
3413 unsigned opSize=target.getTargetData().getTypeSize(divI->getType());
3415 MachineCodeForInstruction& mcfi=MachineCodeForInstruction::get(divI);
3416 divOp1ToUse = new TmpInstruction(mcfi, divOp1);
3417 divOp2ToUse = new TmpInstruction(mcfi, divOp2);
3418 CreateSignExtensionInstructions(target,
3419 divI->getParent()->getParent(),
3420 divOp1, divOp1ToUse,
3421 8*opSize, mvec, mcfi);
3422 CreateSignExtensionInstructions(target,
3423 divI->getParent()->getParent(),
3424 divOp2, divOp2ToUse,
3425 8*opSize, mvec, mcfi);
3429 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
3430 .addReg(divOp1ToUse)
3431 .addReg(divOp2ToUse)
3437 case 37: // reg: Rem(reg, reg)
3438 case 237: // reg: Rem(reg, Constant)
3440 maskUnsignedResult = true;
3442 Instruction* remI = subtreeRoot->getInstruction();
3443 Value* divOp1 = subtreeRoot->leftChild()->getValue();
3444 Value* divOp2 = subtreeRoot->rightChild()->getValue();
3446 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(remI);
3448 // If second operand of divide is smaller than 64 bits, we have
3449 // to make sure the unused top bits are correct because they affect
3450 // the result. These bits are already correct for unsigned values.
3451 // They may be incorrect for signed values, so sign extend to fill in.
3453 Value* divOpToUse = divOp2;
3454 if (divOp2->getType()->isSigned()) {
3455 unsigned opSize=target.getTargetData().getTypeSize(divOp2->getType());
3457 divOpToUse = new TmpInstruction(mcfi, divOp2);
3458 CreateSignExtensionInstructions(target,
3459 remI->getParent()->getParent(),
3461 8*opSize, mvec, mcfi);
3465 // Now compute: result = rem V1, V2 as:
3466 // result = V1 - (V1 / signExtend(V2)) * signExtend(V2)
3468 TmpInstruction* quot = new TmpInstruction(mcfi, divOp1, divOpToUse);
3469 TmpInstruction* prod = new TmpInstruction(mcfi, quot, divOpToUse);
3471 mvec.push_back(BuildMI(ChooseDivInstruction(target, subtreeRoot), 3)
3472 .addReg(divOp1).addReg(divOpToUse).addRegDef(quot));
3474 mvec.push_back(BuildMI(ChooseMulInstructionByType(remI->getType()), 3)
3475 .addReg(quot).addReg(divOpToUse).addRegDef(prod));
3477 mvec.push_back(BuildMI(ChooseSubInstructionByType(remI->getType()), 3)
3478 .addReg(divOp1).addReg(prod).addRegDef(remI));
3483 case 38: // bool: And(bool, bool)
3484 case 138: // bool: And(bool, not)
3485 case 238: // bool: And(bool, boolconst)
3486 case 338: // reg : BAnd(reg, reg)
3487 case 538: // reg : BAnd(reg, Constant)
3488 Add3OperandInstr(V9::ANDr, subtreeRoot, mvec);
3491 case 438: // bool: BAnd(bool, bnot)
3492 { // Use the argument of NOT as the second argument!
3493 // Mark the NOT node so that no code is generated for it.
3494 // If the type is boolean, set 1 or 0 in the result register.
3495 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
3496 Value* notArg = BinaryOperator::getNotArgument(
3497 cast<BinaryOperator>(notNode->getInstruction()));
3498 notNode->markFoldedIntoParent();
3499 Value *lhs = subtreeRoot->leftChild()->getValue();
3500 Value *dest = subtreeRoot->getValue();
3501 mvec.push_back(BuildMI(V9::ANDNr, 3).addReg(lhs).addReg(notArg)
3502 .addReg(dest, MachineOperand::Def));
3504 if (notArg->getType() == Type::BoolTy) {
3505 // set 1 in result register if result of above is non-zero
3506 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
3507 .addReg(dest, MachineOperand::UseAndDef));
3513 case 39: // bool: Or(bool, bool)
3514 case 139: // bool: Or(bool, not)
3515 case 239: // bool: Or(bool, boolconst)
3516 case 339: // reg : BOr(reg, reg)
3517 case 539: // reg : BOr(reg, Constant)
3518 Add3OperandInstr(V9::ORr, subtreeRoot, mvec);
3521 case 439: // bool: BOr(bool, bnot)
3522 { // Use the argument of NOT as the second argument!
3523 // Mark the NOT node so that no code is generated for it.
3524 // If the type is boolean, set 1 or 0 in the result register.
3525 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
3526 Value* notArg = BinaryOperator::getNotArgument(
3527 cast<BinaryOperator>(notNode->getInstruction()));
3528 notNode->markFoldedIntoParent();
3529 Value *lhs = subtreeRoot->leftChild()->getValue();
3530 Value *dest = subtreeRoot->getValue();
3532 mvec.push_back(BuildMI(V9::ORNr, 3).addReg(lhs).addReg(notArg)
3533 .addReg(dest, MachineOperand::Def));
3535 if (notArg->getType() == Type::BoolTy) {
3536 // set 1 in result register if result of above is non-zero
3537 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
3538 .addReg(dest, MachineOperand::UseAndDef));
3544 case 40: // bool: Xor(bool, bool)
3545 case 140: // bool: Xor(bool, not)
3546 case 240: // bool: Xor(bool, boolconst)
3547 case 340: // reg : BXor(reg, reg)
3548 case 540: // reg : BXor(reg, Constant)
3549 Add3OperandInstr(V9::XORr, subtreeRoot, mvec);
3552 case 440: // bool: BXor(bool, bnot)
3553 { // Use the argument of NOT as the second argument!
3554 // Mark the NOT node so that no code is generated for it.
3555 // If the type is boolean, set 1 or 0 in the result register.
3556 InstructionNode* notNode = (InstructionNode*) subtreeRoot->rightChild();
3557 Value* notArg = BinaryOperator::getNotArgument(
3558 cast<BinaryOperator>(notNode->getInstruction()));
3559 notNode->markFoldedIntoParent();
3560 Value *lhs = subtreeRoot->leftChild()->getValue();
3561 Value *dest = subtreeRoot->getValue();
3562 mvec.push_back(BuildMI(V9::XNORr, 3).addReg(lhs).addReg(notArg)
3563 .addReg(dest, MachineOperand::Def));
3565 if (notArg->getType() == Type::BoolTy) {
3566 // set 1 in result register if result of above is non-zero
3567 mvec.push_back(BuildMI(V9::MOVRNZi, 3).addReg(dest).addZImm(1)
3568 .addReg(dest, MachineOperand::UseAndDef));
3573 case 41: // setCCconst: SetCC(reg, Constant)
3574 { // Comparison is with a constant:
3576 // If the bool result must be computed into a register (see below),
3577 // and the constant is int ZERO, we can use the MOVR[op] instructions
3578 // and avoid the SUBcc instruction entirely.
3579 // Otherwise this is just the same as case 42, so just fall through.
3581 // The result of the SetCC must be computed and stored in a register if
3582 // it is used outside the current basic block (so it must be computed
3583 // as a boolreg) or it is used by anything other than a branch.
3584 // We will use a conditional move to do this.
3586 Instruction* setCCInstr = subtreeRoot->getInstruction();
3587 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
3588 ! AllUsesAreBranches(setCCInstr));
3590 if (computeBoolVal) {
3591 InstrTreeNode* constNode = subtreeRoot->rightChild();
3593 constNode->getNodeType() ==InstrTreeNode::NTConstNode);
3594 Constant *constVal = cast<Constant>(constNode->getValue());
3597 if ((constVal->getType()->isInteger()
3598 || isa<PointerType>(constVal->getType()))
3599 && ConvertConstantToIntType(target,
3600 constVal, constVal->getType(), isValidConst) == 0
3603 // That constant is an integer zero after all...
3604 // Use a MOVR[op] to compute the boolean result
3605 // Unconditionally set register to 0
3606 mvec.push_back(BuildMI(V9::SETHI, 2).addZImm(0)
3607 .addRegDef(setCCInstr));
3609 // Now conditionally move 1 into the register.
3610 // Mark the register as a use (as well as a def) because the old
3611 // value will be retained if the condition is false.
3612 MachineOpCode movOpCode = ChooseMovpregiForSetCC(subtreeRoot);
3613 mvec.push_back(BuildMI(movOpCode, 3)
3614 .addReg(subtreeRoot->leftChild()->getValue())
3616 .addReg(setCCInstr, MachineOperand::UseAndDef));
3621 // ELSE FALL THROUGH
3624 case 42: // bool: SetCC(reg, reg):
3626 // This generates a SUBCC instruction, putting the difference in a
3627 // result reg. if needed, and/or setting a condition code if needed.
3629 Instruction* setCCInstr = subtreeRoot->getInstruction();
3630 Value* leftVal = subtreeRoot->leftChild()->getValue();
3631 Value* rightVal = subtreeRoot->rightChild()->getValue();
3632 const Type* opType = leftVal->getType();
3633 bool isFPCompare = opType->isFloatingPoint();
3635 // If the boolean result of the SetCC is used outside the current basic
3636 // block (so it must be computed as a boolreg) or is used by anything
3637 // other than a branch, the boolean must be computed and stored
3638 // in a result register. We will use a conditional move to do this.
3640 bool computeBoolVal = (subtreeRoot->parent() == NULL ||
3641 ! AllUsesAreBranches(setCCInstr));
3643 // A TmpInstruction is created to represent the CC "result".
3644 // Unlike other instances of TmpInstruction, this one is used
3645 // by machine code of multiple LLVM instructions, viz.,
3646 // the SetCC and the branch. Make sure to get the same one!
3647 // Note that we do this even for FP CC registers even though they
3648 // are explicit operands, because the type of the operand
3649 // needs to be a floating point condition code, not an integer
3650 // condition code. Think of this as casting the bool result to
3651 // a FP condition code register.
3652 // Later, we mark the 4th operand as being a CC register, and as a def.
3654 TmpInstruction* tmpForCC = GetTmpForCC(setCCInstr,
3655 setCCInstr->getParent()->getParent(),
3657 MachineCodeForInstruction::get(setCCInstr));
3659 // If the operands are signed values smaller than 4 bytes, then they
3660 // must be sign-extended in order to do a valid 32-bit comparison
3661 // and get the right result in the 32-bit CC register (%icc).
3663 Value* leftOpToUse = leftVal;
3664 Value* rightOpToUse = rightVal;
3665 if (opType->isIntegral() && opType->isSigned()) {
3666 unsigned opSize = target.getTargetData().getTypeSize(opType);
3668 MachineCodeForInstruction& mcfi =
3669 MachineCodeForInstruction::get(setCCInstr);
3671 // create temporary virtual regs. to hold the sign-extensions
3672 leftOpToUse = new TmpInstruction(mcfi, leftVal);
3673 rightOpToUse = new TmpInstruction(mcfi, rightVal);
3675 // sign-extend each operand and put the result in the temporary reg.
3676 CreateSignExtensionInstructions
3677 (target, setCCInstr->getParent()->getParent(),
3678 leftVal, leftOpToUse, 8*opSize, mvec, mcfi);
3679 CreateSignExtensionInstructions
3680 (target, setCCInstr->getParent()->getParent(),
3681 rightVal, rightOpToUse, 8*opSize, mvec, mcfi);
3685 if (! isFPCompare) {
3686 // Integer condition: set CC and discard result.
3687 mvec.push_back(BuildMI(V9::SUBccr, 4)
3688 .addReg(leftOpToUse)
3689 .addReg(rightOpToUse)
3690 .addMReg(target.getRegInfo()->
3691 getZeroRegNum(), MachineOperand::Def)
3692 .addCCReg(tmpForCC, MachineOperand::Def));
3694 // FP condition: dest of FCMP should be some FCCn register
3695 mvec.push_back(BuildMI(ChooseFcmpInstruction(subtreeRoot), 3)
3696 .addCCReg(tmpForCC, MachineOperand::Def)
3697 .addReg(leftOpToUse)
3698 .addReg(rightOpToUse));
3701 if (computeBoolVal) {
3702 MachineOpCode movOpCode = (isFPCompare
3703 ? ChooseMovFpcciInstruction(subtreeRoot)
3704 : ChooseMovpcciForSetCC(subtreeRoot));
3706 // Unconditionally set register to 0
3707 M = BuildMI(V9::SETHI, 2).addZImm(0).addRegDef(setCCInstr);
3710 // Now conditionally move 1 into the register.
3711 // Mark the register as a use (as well as a def) because the old
3712 // value will be retained if the condition is false.
3713 M = (BuildMI(movOpCode, 3).addCCReg(tmpForCC).addZImm(1)
3714 .addReg(setCCInstr, MachineOperand::UseAndDef));
3720 case 51: // reg: Load(reg)
3721 case 52: // reg: Load(ptrreg)
3722 SetOperandsForMemInstr(ChooseLoadInstruction(
3723 subtreeRoot->getValue()->getType()),
3724 mvec, subtreeRoot, target);
3727 case 55: // reg: GetElemPtr(reg)
3728 case 56: // reg: GetElemPtrIdx(reg,reg)
3729 // If the GetElemPtr was folded into the user (parent), it will be
3730 // caught above. For other cases, we have to compute the address.
3731 SetOperandsForMemInstr(V9::ADDr, mvec, subtreeRoot, target);
3734 case 57: // reg: Alloca: Implement as 1 instruction:
3735 { // add %fp, offsetFromFP -> result
3736 AllocationInst* instr =
3737 cast<AllocationInst>(subtreeRoot->getInstruction());
3739 target.getTargetData().getTypeSize(instr->getAllocatedType());
3741 CreateCodeForFixedSizeAlloca(target, instr, tsize, 1, mvec);
3745 case 58: // reg: Alloca(reg): Implement as 3 instructions:
3746 // mul num, typeSz -> tmp
3747 // sub %sp, tmp -> %sp
3748 { // add %sp, frameSizeBelowDynamicArea -> result
3749 AllocationInst* instr =
3750 cast<AllocationInst>(subtreeRoot->getInstruction());
3751 const Type* eltType = instr->getAllocatedType();
3753 // If #elements is constant, use simpler code for fixed-size allocas
3754 int tsize = (int) target.getTargetData().getTypeSize(eltType);
3755 Value* numElementsVal = NULL;
3756 bool isArray = instr->isArrayAllocation();
3758 if (!isArray || isa<Constant>(numElementsVal = instr->getArraySize())) {
3759 // total size is constant: generate code for fixed-size alloca
3760 unsigned numElements = isArray?
3761 cast<ConstantUInt>(numElementsVal)->getValue() : 1;
3762 CreateCodeForFixedSizeAlloca(target, instr, tsize,
3765 // total size is not constant.
3766 CreateCodeForVariableSizeAlloca(target, instr, tsize,
3767 numElementsVal, mvec);
3772 case 61: // reg: Call
3773 { // Generate a direct (CALL) or indirect (JMPL) call.
3774 // Mark the return-address register, the indirection
3775 // register (for indirect calls), the operands of the Call,
3776 // and the return value (if any) as implicit operands
3777 // of the machine instruction.
3779 // If this is a varargs function, floating point arguments
3780 // have to passed in integer registers so insert
3781 // copy-float-to-int instructions for each float operand.
3783 CallInst *callInstr = cast<CallInst>(subtreeRoot->getInstruction());
3784 Value *callee = callInstr->getCalledValue();
3785 Function* calledFunc = dyn_cast<Function>(callee);
3787 // Check if this is an intrinsic function that needs a special code
3788 // sequence (e.g., va_start). Indirect calls cannot be special.
3790 bool specialIntrinsic = false;
3792 if (calledFunc && (iid=(Intrinsic::ID)calledFunc->getIntrinsicID()))
3793 specialIntrinsic = CodeGenIntrinsic(iid, *callInstr, target, mvec);
3795 // If not, generate the normal call sequence for the function.
3796 // This can also handle any intrinsics that are just function calls.
3798 if (! specialIntrinsic) {
3799 Function* currentFunc = callInstr->getParent()->getParent();
3800 MachineFunction& MF = MachineFunction::get(currentFunc);
3801 MachineCodeForInstruction& mcfi =
3802 MachineCodeForInstruction::get(callInstr);
3803 const SparcV9RegInfo& regInfo =
3804 (SparcV9RegInfo&) *target.getRegInfo();
3805 const TargetFrameInfo& frameInfo = *target.getFrameInfo();
3807 // Create hidden virtual register for return address with type void*
3808 TmpInstruction* retAddrReg =
3809 new TmpInstruction(mcfi, PointerType::get(Type::VoidTy), callInstr);
3811 // Generate the machine instruction and its operands.
3812 // Use CALL for direct function calls; this optimistically assumes
3813 // the PC-relative address fits in the CALL address field (22 bits).
3814 // Use JMPL for indirect calls.
3815 // This will be added to mvec later, after operand copies.
3817 MachineInstr* callMI;
3818 if (calledFunc) // direct function call
3819 callMI = BuildMI(V9::CALL, 1).addPCDisp(callee);
3820 else // indirect function call
3821 callMI = (BuildMI(V9::JMPLCALLi,3).addReg(callee)
3822 .addSImm((int64_t)0).addRegDef(retAddrReg));
3824 const FunctionType* funcType =
3825 cast<FunctionType>(cast<PointerType>(callee->getType())
3826 ->getElementType());
3827 bool isVarArgs = funcType->isVarArg();
3828 bool noPrototype = isVarArgs && funcType->getNumParams() == 0;
3830 // Use a descriptor to pass information about call arguments
3831 // to the register allocator. This descriptor will be "owned"
3832 // and freed automatically when the MachineCodeForInstruction
3833 // object for the callInstr goes away.
3834 CallArgsDescriptor* argDesc =
3835 new CallArgsDescriptor(callInstr, retAddrReg,isVarArgs,noPrototype);
3836 assert(callInstr->getOperand(0) == callee
3837 && "This is assumed in the loop below!");
3839 // Insert sign-extension instructions for small signed values,
3840 // if this is an unknown function (i.e., called via a funcptr)
3841 // or an external one (i.e., which may not be compiled by llc).
3843 if (calledFunc == NULL || calledFunc->isExternal()) {
3844 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
3845 Value* argVal = callInstr->getOperand(i);
3846 const Type* argType = argVal->getType();
3847 if (argType->isIntegral() && argType->isSigned()) {
3848 unsigned argSize = target.getTargetData().getTypeSize(argType);
3850 // create a temporary virtual reg. to hold the sign-extension
3851 TmpInstruction* argExtend = new TmpInstruction(mcfi, argVal);
3853 // sign-extend argVal and put the result in the temporary reg.
3854 CreateSignExtensionInstructions
3855 (target, currentFunc, argVal, argExtend,
3856 8*argSize, mvec, mcfi);
3858 // replace argVal with argExtend in CallArgsDescriptor
3859 argDesc->getArgInfo(i-1).replaceArgVal(argExtend);
3865 // Insert copy instructions to get all the arguments into
3866 // all the places that they need to be.
3868 for (unsigned i=1, N=callInstr->getNumOperands(); i < N; ++i) {
3870 CallArgInfo& argInfo = argDesc->getArgInfo(argNo);
3871 Value* argVal = argInfo.getArgVal(); // don't use callInstr arg here
3872 const Type* argType = argVal->getType();
3873 unsigned regType = regInfo.getRegTypeForDataType(argType);
3874 unsigned argSize = target.getTargetData().getTypeSize(argType);
3875 int regNumForArg = SparcV9RegInfo::getInvalidRegNum();
3876 unsigned regClassIDOfArgReg;
3878 // Check for FP arguments to varargs functions.
3879 // Any such argument in the first $K$ args must be passed in an
3880 // integer register. If there is no prototype, it must also
3881 // be passed as an FP register.
3882 // K = #integer argument registers.
3883 bool isFPArg = argVal->getType()->isFloatingPoint();
3884 if (isVarArgs && isFPArg) {
3887 // It is a function with no prototype: pass value
3888 // as an FP value as well as a varargs value. The FP value
3889 // may go in a register or on the stack. The copy instruction
3890 // to the outgoing reg/stack is created by the normal argument
3891 // handling code since this is the "normal" passing mode.
3893 regNumForArg = regInfo.regNumForFPArg(regType,
3894 false, false, argNo,
3895 regClassIDOfArgReg);
3896 if (regNumForArg == regInfo.getInvalidRegNum())
3897 argInfo.setUseStackSlot();
3899 argInfo.setUseFPArgReg();
3902 // If this arg. is in the first $K$ regs, add special copy-
3903 // float-to-int instructions to pass the value as an int.
3904 // To check if it is in the first $K$, get the register
3905 // number for the arg #i. These copy instructions are
3906 // generated here because they are extra cases and not needed
3907 // for the normal argument handling (some code reuse is
3908 // possible though -- later).
3910 int copyRegNum = regInfo.regNumForIntArg(false, false, argNo,
3911 regClassIDOfArgReg);
3912 if (copyRegNum != regInfo.getInvalidRegNum()) {
3913 // Create a virtual register to represent copyReg. Mark
3914 // this vreg as being an implicit operand of the call MI
3915 const Type* loadTy = (argType == Type::FloatTy
3916 ? Type::IntTy : Type::LongTy);
3917 TmpInstruction* argVReg = new TmpInstruction(mcfi, loadTy,
3920 callMI->addImplicitRef(argVReg);
3922 // Get a temp stack location to use to copy
3923 // float-to-int via the stack.
3925 // FIXME: For now, we allocate permanent space because
3926 // the stack frame manager does not allow locals to be
3927 // allocated (e.g., for alloca) after a temp is
3930 // int tmpOffset = MF.getInfo()->pushTempValue(argSize);
3931 int tmpOffset = MF.getInfo()->allocateLocalVar(argVReg);
3933 // Generate the store from FP reg to stack
3934 unsigned StoreOpcode = ChooseStoreInstruction(argType);
3935 M = BuildMI(convertOpcodeFromRegToImm(StoreOpcode), 3)
3936 .addReg(argVal).addMReg(regInfo.getFramePointer())
3937 .addSImm(tmpOffset);
3940 // Generate the load from stack to int arg reg
3941 unsigned LoadOpcode = ChooseLoadInstruction(loadTy);
3942 M = BuildMI(convertOpcodeFromRegToImm(LoadOpcode), 3)
3943 .addMReg(regInfo.getFramePointer()).addSImm(tmpOffset)
3944 .addReg(argVReg, MachineOperand::Def);
3946 // Mark operand with register it should be assigned
3947 // both for copy and for the callMI
3948 M->SetRegForOperand(M->getNumOperands()-1, copyRegNum);
3949 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
3953 // Add info about the argument to the CallArgsDescriptor
3954 argInfo.setUseIntArgReg();
3955 argInfo.setArgCopy(copyRegNum);
3957 // Cannot fit in first $K$ regs so pass arg on stack
3958 argInfo.setUseStackSlot();
3960 } else if (isFPArg) {
3961 // Get the outgoing arg reg to see if there is one.
3962 regNumForArg = regInfo.regNumForFPArg(regType, false, false,
3963 argNo, regClassIDOfArgReg);
3964 if (regNumForArg == regInfo.getInvalidRegNum())
3965 argInfo.setUseStackSlot();
3967 argInfo.setUseFPArgReg();
3968 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
3972 // Get the outgoing arg reg to see if there is one.
3973 regNumForArg = regInfo.regNumForIntArg(false,false,
3974 argNo, regClassIDOfArgReg);
3975 if (regNumForArg == regInfo.getInvalidRegNum())
3976 argInfo.setUseStackSlot();
3978 argInfo.setUseIntArgReg();
3979 regNumForArg =regInfo.getUnifiedRegNum(regClassIDOfArgReg,
3985 // Now insert copy instructions to stack slot or arg. register
3987 if (argInfo.usesStackSlot()) {
3988 // Get the stack offset for this argument slot.
3989 // FP args on stack are right justified so adjust offset!
3990 // int arguments are also right justified but they are
3991 // always loaded as a full double-word so the offset does
3992 // not need to be adjusted.
3993 int argOffset = frameInfo.getOutgoingArgOffset(MF, argNo);
3994 if (argType->isFloatingPoint()) {
3995 unsigned slotSize = frameInfo.getSizeOfEachArgOnStack();
3996 assert(argSize <= slotSize && "Insufficient slot size!");
3997 argOffset += slotSize - argSize;
4000 // Now generate instruction to copy argument to stack
4001 MachineOpCode storeOpCode =
4002 (argType->isFloatingPoint()
4003 ? ((argSize == 4)? V9::STFi : V9::STDFi) : V9::STXi);
4005 M = BuildMI(storeOpCode, 3).addReg(argVal)
4006 .addMReg(regInfo.getStackPointer()).addSImm(argOffset);
4009 else if (regNumForArg != regInfo.getInvalidRegNum()) {
4011 // Create a virtual register to represent the arg reg. Mark
4012 // this vreg as being an implicit operand of the call MI.
4013 TmpInstruction* argVReg =
4014 new TmpInstruction(mcfi, argVal, NULL, "argReg");
4016 callMI->addImplicitRef(argVReg);
4018 // Generate the reg-to-reg copy into the outgoing arg reg.
4019 // -- For FP values, create a FMOVS or FMOVD instruction
4020 // -- For non-FP values, create an add-with-0 instruction
4021 if (argType->isFloatingPoint())
4022 M=(BuildMI(argType==Type::FloatTy? V9::FMOVS :V9::FMOVD,2)
4023 .addReg(argVal).addReg(argVReg, MachineOperand::Def));
4025 M = (BuildMI(ChooseAddInstructionByType(argType), 3)
4026 .addReg(argVal).addSImm((int64_t) 0)
4027 .addReg(argVReg, MachineOperand::Def));
4029 // Mark the operand with the register it should be assigned
4030 M->SetRegForOperand(M->getNumOperands()-1, regNumForArg);
4031 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,
4037 assert(argInfo.getArgCopy() != regInfo.getInvalidRegNum() &&
4038 "Arg. not in stack slot, primary or secondary register?");
4041 // add call instruction and delay slot before copying return value
4042 mvec.push_back(callMI);
4043 mvec.push_back(BuildMI(V9::NOP, 0));
4045 // Add the return value as an implicit ref. The call operands
4046 // were added above. Also, add code to copy out the return value.
4047 // This is always register-to-register for int or FP return values.
4049 if (callInstr->getType() != Type::VoidTy) {
4050 // Get the return value reg.
4051 const Type* retType = callInstr->getType();
4053 int regNum = (retType->isFloatingPoint()
4054 ? (unsigned) SparcV9FloatRegClass::f0
4055 : (unsigned) SparcV9IntRegClass::o0);
4056 unsigned regClassID = regInfo.getRegClassIDOfType(retType);
4057 regNum = regInfo.getUnifiedRegNum(regClassID, regNum);
4059 // Create a virtual register to represent it and mark
4060 // this vreg as being an implicit operand of the call MI
4061 TmpInstruction* retVReg =
4062 new TmpInstruction(mcfi, callInstr, NULL, "argReg");
4064 callMI->addImplicitRef(retVReg, /*isDef*/ true);
4066 // Generate the reg-to-reg copy from the return value reg.
4067 // -- For FP values, create a FMOVS or FMOVD instruction
4068 // -- For non-FP values, create an add-with-0 instruction
4069 if (retType->isFloatingPoint())
4070 M = (BuildMI(retType==Type::FloatTy? V9::FMOVS : V9::FMOVD, 2)
4071 .addReg(retVReg).addReg(callInstr, MachineOperand::Def));
4073 M = (BuildMI(ChooseAddInstructionByType(retType), 3)
4074 .addReg(retVReg).addSImm((int64_t) 0)
4075 .addReg(callInstr, MachineOperand::Def));
4077 // Mark the operand with the register it should be assigned
4078 // Also mark the implicit ref of the call defining this operand
4079 M->SetRegForOperand(0, regNum);
4080 callMI->SetRegForImplicitRef(callMI->getNumImplicitRefs()-1,regNum);
4085 // For the CALL instruction, the ret. addr. reg. is also implicit
4086 if (isa<Function>(callee))
4087 callMI->addImplicitRef(retAddrReg, /*isDef*/ true);
4089 MF.getInfo()->popAllTempValues(); // free temps used for this inst
4095 case 62: // reg: Shl(reg, reg)
4097 Value* argVal1 = subtreeRoot->leftChild()->getValue();
4098 Value* argVal2 = subtreeRoot->rightChild()->getValue();
4099 Instruction* shlInstr = subtreeRoot->getInstruction();
4101 const Type* opType = argVal1->getType();
4102 assert((opType->isInteger() || isa<PointerType>(opType)) &&
4103 "Shl unsupported for other types");
4104 unsigned opSize = target.getTargetData().getTypeSize(opType);
4106 CreateShiftInstructions(target, shlInstr->getParent()->getParent(),
4107 (opSize > 4)? V9::SLLXr6:V9::SLLr5,
4108 argVal1, argVal2, 0, shlInstr, mvec,
4109 MachineCodeForInstruction::get(shlInstr));
4113 case 63: // reg: Shr(reg, reg)
4115 const Type* opType = subtreeRoot->leftChild()->getValue()->getType();
4116 assert((opType->isInteger() || isa<PointerType>(opType)) &&
4117 "Shr unsupported for other types");
4118 unsigned opSize = target.getTargetData().getTypeSize(opType);
4119 Add3OperandInstr(opType->isSigned()
4120 ? (opSize > 4? V9::SRAXr6 : V9::SRAr5)
4121 : (opSize > 4? V9::SRLXr6 : V9::SRLr5),
4126 case 64: // reg: Phi(reg,reg)
4127 break; // don't forward the value
4129 case 65: // reg: VANext(reg): the va_next(va_list, type) instruction
4130 { // Increment the va_list pointer register according to the type.
4131 // All LLVM argument types are <= 64 bits, so use one doubleword.
4132 Instruction* vaNextI = subtreeRoot->getInstruction();
4133 assert(target.getTargetData().getTypeSize(vaNextI->getType()) <= 8 &&
4134 "We assumed that all LLVM parameter types <= 8 bytes!");
4135 int argSize = target.getFrameInfo()->getSizeOfEachArgOnStack();
4136 mvec.push_back(BuildMI(V9::ADDi, 3).addReg(vaNextI->getOperand(0)).
4137 addSImm(argSize).addRegDef(vaNextI));
4141 case 66: // reg: VAArg (reg): the va_arg instruction
4142 { // Load argument from stack using current va_list pointer value.
4143 // Use 64-bit load for all non-FP args, and LDDF or double for FP.
4144 Instruction* vaArgI = subtreeRoot->getInstruction();
4145 MachineOpCode loadOp = (vaArgI->getType()->isFloatingPoint()
4146 ? (vaArgI->getType() == Type::FloatTy
4147 ? V9::LDFi : V9::LDDFi)
4149 mvec.push_back(BuildMI(loadOp, 3).addReg(vaArgI->getOperand(0)).
4150 addSImm(0).addRegDef(vaArgI));
4154 case 71: // reg: VReg
4155 case 72: // reg: Constant
4156 break; // don't forward the value
4159 assert(0 && "Unrecognized BURG rule");
4164 if (forwardOperandNum >= 0) {
4165 // We did not generate a machine instruction but need to use operand.
4166 // If user is in the same tree, replace Value in its machine operand.
4167 // If not, insert a copy instruction which should get coalesced away
4168 // by register allocation.
4169 if (subtreeRoot->parent() != NULL)
4170 ForwardOperand(subtreeRoot, subtreeRoot->parent(), forwardOperandNum);
4172 std::vector<MachineInstr*> minstrVec;
4173 Instruction* instr = subtreeRoot->getInstruction();
4174 CreateCopyInstructionsByType(target,
4175 instr->getParent()->getParent(),
4176 instr->getOperand(forwardOperandNum),
4178 MachineCodeForInstruction::get(instr));
4179 assert(minstrVec.size() > 0);
4180 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
4184 if (maskUnsignedResult) {
4185 // If result is unsigned and smaller than int reg size,
4186 // we need to clear high bits of result value.
4187 assert(forwardOperandNum < 0 && "Need mask but no instruction generated");
4188 Instruction* dest = subtreeRoot->getInstruction();
4189 if (dest->getType()->isUnsigned()) {
4190 unsigned destSize=target.getTargetData().getTypeSize(dest->getType());
4191 if (destSize <= 4) {
4192 // Mask high 64 - N bits, where N = 4*destSize.
4194 // Use a TmpInstruction to represent the
4195 // intermediate result before masking. Since those instructions
4196 // have already been generated, go back and substitute tmpI
4197 // for dest in the result position of each one of them.
4199 MachineCodeForInstruction& mcfi = MachineCodeForInstruction::get(dest);
4200 TmpInstruction *tmpI = new TmpInstruction(mcfi, dest->getType(),
4201 dest, NULL, "maskHi");
4202 Value* srlArgToUse = tmpI;
4204 unsigned numSubst = 0;
4205 for (unsigned i=0, N=mvec.size(); i < N; ++i) {
4207 // Make sure we substitute all occurrences of dest in these instrs.
4208 // Otherwise, we will have bogus code.
4209 bool someArgsWereIgnored = false;
4211 // Make sure not to substitute an upwards-exposed use -- that would
4212 // introduce a use of `tmpI' with no preceding def. Therefore,
4213 // substitute a use or def-and-use operand only if a previous def
4214 // operand has already been substituted (i.e., numSubst > 0).
4216 numSubst += mvec[i]->substituteValue(dest, tmpI,
4217 /*defsOnly*/ numSubst == 0,
4218 /*notDefsAndUses*/ numSubst > 0,
4219 someArgsWereIgnored);
4220 assert(!someArgsWereIgnored &&
4221 "Operand `dest' exists but not replaced: probably bogus!");
4223 assert(numSubst > 0 && "Operand `dest' not replaced: probably bogus!");
4225 // Left shift 32-N if size (N) is less than 32 bits.
4226 // Use another tmp. virtual register to represent this result.
4228 srlArgToUse = new TmpInstruction(mcfi, dest->getType(),
4229 tmpI, NULL, "maskHi2");
4230 mvec.push_back(BuildMI(V9::SLLXi6, 3).addReg(tmpI)
4231 .addZImm(8*(4-destSize))
4232 .addReg(srlArgToUse, MachineOperand::Def));
4235 // Logical right shift 32-N to get zero extension in top 64-N bits.
4236 mvec.push_back(BuildMI(V9::SRLi5, 3).addReg(srlArgToUse)
4237 .addZImm(8*(4-destSize))
4238 .addReg(dest, MachineOperand::Def));
4240 } else if (destSize < 8) {
4241 assert(0 && "Unsupported type size: 32 < size < 64 bits");
4247 } // End llvm namespace
4249 //==------------------------------------------------------------------------==//
4250 // Class V9ISel Implementation
4251 //==------------------------------------------------------------------------==//
4253 bool V9ISel::runOnFunction(Function &F) {
4254 // First pass - Walk the function, lowering any calls to intrinsic functions
4255 // which the instruction selector cannot handle.
4256 for (Function::iterator BB = F.begin(), E = F.end(); BB != E; ++BB)
4257 for (BasicBlock::iterator I = BB->begin(), E = BB->end(); I != E; )
4258 if (CallInst *CI = dyn_cast<CallInst>(I++))
4259 if (Function *F = CI->getCalledFunction())
4260 switch (F->getIntrinsicID()) {
4261 case Intrinsic::not_intrinsic:
4262 case Intrinsic::vastart:
4263 case Intrinsic::vacopy:
4264 case Intrinsic::vaend:
4265 // We directly implement these intrinsics. Note that this knowledge
4266 // is incestuously entangled with the code in
4267 // SparcInstrSelection.cpp and must be updated when it is updated.
4268 // Since ALL of the code in this library is incestuously intertwined
4269 // with it already and sparc specific, we will live with this.
4272 // All other intrinsic calls we must lower.
4273 Instruction *Before = CI->getPrev();
4274 Target.getIntrinsicLowering().LowerIntrinsicCall(CI);
4275 if (Before) { // Move iterator to instruction after call
4282 // Build the instruction trees to be given as inputs to BURG.
4283 InstrForest instrForest(&F);
4284 if (SelectDebugLevel >= Select_DebugInstTrees) {
4285 std::cerr << "\n\n*** Input to instruction selection for function "
4286 << F.getName() << "\n\n" << F
4287 << "\n\n*** Instruction trees for function "
4288 << F.getName() << "\n\n";
4292 // Invoke BURG instruction selection for each tree
4293 for (InstrForest::const_root_iterator RI = instrForest.roots_begin();
4294 RI != instrForest.roots_end(); ++RI) {
4295 InstructionNode* basicNode = *RI;
4296 assert(basicNode->parent() == NULL && "A `root' node has a parent?");
4298 // Invoke BURM to label each tree node with a state
4299 burm_label(basicNode);
4300 if (SelectDebugLevel >= Select_DebugBurgTrees) {
4301 printcover(basicNode, 1, 0);
4302 std::cerr << "\nCover cost == " << treecost(basicNode, 1, 0) <<"\n\n";
4303 printMatches(basicNode);
4306 // Then recursively walk the tree to select instructions
4307 SelectInstructionsForTree(basicNode, /*goalnt*/1);
4310 // Create the MachineBasicBlocks and add all of the MachineInstrs
4311 // defined in the MachineCodeForInstruction objects to the MachineBasicBlocks.
4312 MachineFunction &MF = MachineFunction::get(&F);
4313 std::map<const BasicBlock *, MachineBasicBlock *> MBBMap;
4314 for (Function::iterator BI = F.begin(), BE = F.end(); BI != BE; ++BI) {
4315 MachineBasicBlock *MBB = new MachineBasicBlock(BI);
4316 MF.getBasicBlockList().push_back(MBB);
4319 for (BasicBlock::iterator II = BI->begin(); II != BI->end(); ++II) {
4320 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(II);
4321 MBB->insert(MBB->end(), mvec.begin(), mvec.end());
4325 // Initialize Machine-CFG for the function.
4326 for (MachineFunction::iterator i = MF.begin (), e = MF.end (); i != e; ++i) {
4327 MachineBasicBlock &MBB = *i;
4328 const BasicBlock *BB = MBB.getBasicBlock ();
4329 // for each successor S of BB, add MBBMap[S] as a successor of MBB.
4330 for (succ_const_iterator si = succ_begin(BB), se = succ_end(BB); si != se;
4332 MachineBasicBlock *succMBB = MBBMap[*si];
4333 assert (succMBB && "Can't find MachineBasicBlock for this successor");
4334 MBB.addSuccessor (succMBB);
4338 // Insert phi elimination code
4339 InsertCodeForPhis(F);
4341 if (SelectDebugLevel >= Select_PrintMachineCode) {
4342 std::cerr << "\n*** Machine instructions after INSTRUCTION SELECTION\n";
4343 MachineFunction::get(&F).dump();
4349 /// InsertCodeForPhis - This method inserts Phi elimination code for
4350 /// all Phi nodes in the given function. After this method is called,
4351 /// the Phi nodes still exist in the LLVM code, but copies are added to the
4354 void V9ISel::InsertCodeForPhis(Function &F) {
4355 // Iterate over every Phi node PN in F:
4356 MachineFunction &MF = MachineFunction::get(&F);
4357 for (MachineFunction::iterator BB = MF.begin(); BB != MF.end(); ++BB) {
4358 for (BasicBlock::const_iterator IIt = BB->getBasicBlock()->begin();
4359 const PHINode *PN = dyn_cast<PHINode>(IIt); ++IIt) {
4360 // Create a new temporary register to hold the result of the Phi copy.
4361 // The leak detector shouldn't track these nodes. They are not garbage,
4362 // even though their parent field is never filled in.
4363 Value *PhiCpRes = new PHINode(PN->getType(), PN->getName() + ":PhiCp");
4364 LeakDetector::removeGarbageObject(PhiCpRes);
4366 // For each of PN's incoming values, insert a copy in the corresponding
4367 // predecessor block.
4368 MachineCodeForInstruction &MCforPN = MachineCodeForInstruction::get (PN);
4369 for (unsigned i = 0; i < PN->getNumIncomingValues(); ++i) {
4370 std::vector<MachineInstr*> mvec, CpVec;
4371 Target.getRegInfo()->cpValue2Value(PN->getIncomingValue(i),
4373 for (std::vector<MachineInstr*>::iterator MI=mvec.begin();
4374 MI != mvec.end(); ++MI) {
4375 std::vector<MachineInstr*> CpVec2 =
4376 FixConstantOperandsForInstr(const_cast<PHINode*>(PN), *MI, Target);
4377 CpVec2.push_back(*MI);
4378 CpVec.insert(CpVec.end(), CpVec2.begin(), CpVec2.end());
4380 // Insert the copy instructions into the predecessor BB.
4381 InsertPhiElimInstructions(PN->getIncomingBlock(i), CpVec);
4382 MCforPN.insert (MCforPN.end (), CpVec.begin (), CpVec.end ());
4384 // Insert a copy instruction from PhiCpRes to PN.
4385 std::vector<MachineInstr*> mvec;
4386 Target.getRegInfo()->cpValue2Value(PhiCpRes, const_cast<PHINode*>(PN),
4388 BB->insert(BB->begin(), mvec.begin(), mvec.end());
4389 MCforPN.insert (MCforPN.end (), mvec.begin (), mvec.end ());
4390 } // for each Phi Instr in BB
4391 } // for all BBs in function
4394 /// InsertPhiElimInstructions - Inserts the instructions in CpVec into the
4395 /// MachineBasicBlock corresponding to BB, just before its terminator
4396 /// instruction. This is used by InsertCodeForPhis() to insert copies, above.
4398 void V9ISel::InsertPhiElimInstructions(BasicBlock *BB,
4399 const std::vector<MachineInstr*>& CpVec)
4401 Instruction *TermInst = (Instruction*)BB->getTerminator();
4402 MachineCodeForInstruction &MC4Term = MachineCodeForInstruction::get(TermInst);
4403 MachineInstr *FirstMIOfTerm = MC4Term.front();
4404 assert (FirstMIOfTerm && "No Machine Instrs for terminator");
4406 MachineBasicBlock *MBB = FirstMIOfTerm->getParent();
4407 assert(MBB && "Machine BB for predecessor's terminator not found");
4408 MachineBasicBlock::iterator MCIt = FirstMIOfTerm;
4409 assert(MCIt != MBB->end() && "Start inst of terminator not found");
4411 // Insert the copy instructions just before the first machine instruction
4412 // generated for the terminator.
4413 MBB->insert(MCIt, CpVec.begin(), CpVec.end());
4416 /// SelectInstructionsForTree - Recursively walk the tree to select
4417 /// instructions. Do this top-down so that child instructions can exploit
4418 /// decisions made at the child instructions.
4420 /// E.g., if br(setle(reg,const)) decides the constant is 0 and uses
4421 /// a branch-on-integer-register instruction, then the setle node
4422 /// can use that information to avoid generating the SUBcc instruction.
4424 /// Note that this cannot be done bottom-up because setle must do this
4425 /// only if it is a child of the branch (otherwise, the result of setle
4426 /// may be used by multiple instructions).
4428 void V9ISel::SelectInstructionsForTree(InstrTreeNode* treeRoot, int goalnt) {
4429 // Get the rule that matches this node.
4430 int ruleForNode = burm_rule(treeRoot->state, goalnt);
4432 if (ruleForNode == 0) {
4433 std::cerr << "Could not match instruction tree for instr selection\n";
4437 // Get this rule's non-terminals and the corresponding child nodes (if any)
4438 short *nts = burm_nts[ruleForNode];
4440 // First, select instructions for the current node and rule.
4441 // (If this is a list node, not an instruction, then skip this step).
4442 // This function is specific to the target architecture.
4443 if (treeRoot->opLabel != VRegListOp) {
4444 std::vector<MachineInstr*> minstrVec;
4445 InstructionNode* instrNode = (InstructionNode*)treeRoot;
4446 assert(instrNode->getNodeType() == InstrTreeNode::NTInstructionNode);
4447 GetInstructionsByRule(instrNode, ruleForNode, nts, Target, minstrVec);
4448 MachineCodeForInstruction &mvec =
4449 MachineCodeForInstruction::get(instrNode->getInstruction());
4450 mvec.insert(mvec.end(), minstrVec.begin(), minstrVec.end());
4453 // Then, recursively compile the child nodes, if any.
4456 // i.e., there is at least one kid
4457 InstrTreeNode* kids[2];
4458 int currentRule = ruleForNode;
4459 burm_kids(treeRoot, currentRule, kids);
4461 // First skip over any chain rules so that we don't visit
4462 // the current node again.
4463 while (ThisIsAChainRule(currentRule)) {
4464 currentRule = burm_rule(treeRoot->state, nts[0]);
4465 nts = burm_nts[currentRule];
4466 burm_kids(treeRoot, currentRule, kids);
4469 // Now we have the first non-chain rule so we have found
4470 // the actual child nodes. Recursively compile them.
4471 for (unsigned i = 0; nts[i]; i++) {
4473 InstrTreeNode::InstrTreeNodeType nodeType = kids[i]->getNodeType();
4474 if (nodeType == InstrTreeNode::NTVRegListNode ||
4475 nodeType == InstrTreeNode::NTInstructionNode)
4476 SelectInstructionsForTree(kids[i], nts[i]);
4480 // Finally, do any post-processing on this node after its children
4481 // have been translated.
4482 if (treeRoot->opLabel != VRegListOp)
4483 PostprocessMachineCodeForTree((InstructionNode*)treeRoot, ruleForNode, nts);
4486 /// PostprocessMachineCodeForTree - Apply any final cleanups to
4487 /// machine code for the root of a subtree after selection for all its
4488 /// children has been completed.
4490 void V9ISel::PostprocessMachineCodeForTree(InstructionNode *instrNode,
4491 int ruleForNode, short *nts) {
4492 // Fix up any constant operands in the machine instructions to either
4493 // use an immediate field or to load the constant into a register.
4494 // Walk backwards and use direct indexes to allow insertion before current.
4495 Instruction* vmInstr = instrNode->getInstruction();
4496 MachineCodeForInstruction &mvec = MachineCodeForInstruction::get(vmInstr);
4497 for (unsigned i = mvec.size(); i != 0; --i) {
4498 std::vector<MachineInstr*> loadConstVec =
4499 FixConstantOperandsForInstr(vmInstr, mvec[i-1], Target);
4500 mvec.insert(mvec.begin()+i-1, loadConstVec.begin(), loadConstVec.end());
4504 /// createSparcV9BurgInstSelector - Creates and returns a new SparcV9
4505 /// BURG-based instruction selection pass.
4507 FunctionPass *llvm::createSparcV9BurgInstSelector(TargetMachine &TM) {
4508 return new V9ISel(TM);