1 //===-- PoolAllocate.cpp - Pool Allocation Pass ---------------------------===//
3 // This transform changes programs so that disjoint data structures are
4 // allocated out of different pools of memory, increasing locality and shrinking
7 //===----------------------------------------------------------------------===//
10 #include "llvm/Pass.h"
12 #include "llvm/Transforms/IPO.h"
13 #include "llvm/Transforms/Utils/Cloning.h"
14 #include "llvm/Analysis/DataStructure.h"
15 #include "llvm/Module.h"
16 #include "llvm/iMemory.h"
17 #include "llvm/iTerminators.h"
18 #include "llvm/iPHINode.h"
19 #include "llvm/iOther.h"
20 #include "llvm/DerivedTypes.h"
21 #include "llvm/Constants.h"
22 #include "llvm/Target/TargetData.h"
23 #include "llvm/Support/InstVisitor.h"
24 #include "Support/DepthFirstIterator.h"
25 #include "Support/STLExtras.h"
33 // DEBUG_CREATE_POOLS - Enable this to turn on debug output for the pool
34 // creation phase in the top level function of a transformed data structure.
36 //#define DEBUG_CREATE_POOLS 1
38 // DEBUG_TRANSFORM_PROGRESS - Enable this to get lots of debug output on what
39 // the transformation is doing.
41 //#define DEBUG_TRANSFORM_PROGRESS 1
43 // DEBUG_POOLBASE_LOAD_ELIMINATOR - Turn this on to get statistics about how
44 // many static loads were eliminated from a function...
46 #define DEBUG_POOLBASE_LOAD_ELIMINATOR 1
48 #include "Support/CommandLine.h"
50 Ptr8bits, Ptr16bits, Ptr32bits
53 static cl::opt<PtrSize>
54 ReqPointerSize("poolalloc-ptr-size",
55 cl::desc("Set pointer size for -poolalloc pass"),
57 clEnumValN(Ptr32bits, "32", "Use 32 bit indices for pointers"),
58 clEnumValN(Ptr16bits, "16", "Use 16 bit indices for pointers"),
59 clEnumValN(Ptr8bits , "8", "Use 8 bit indices for pointers"),
63 DisableRLE("no-pool-load-elim", cl::Hidden,
64 cl::desc("Disable pool load elimination after poolalloc pass"));
66 const Type *POINTERTYPE;
68 // FIXME: This is dependant on the sparc backend layout conventions!!
69 static TargetData TargetData("test");
71 static const Type *getPointerTransformedType(const Type *Ty) {
72 if (const PointerType *PT = dyn_cast<PointerType>(Ty)) {
74 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
75 vector<const Type *> NewElTypes;
76 NewElTypes.reserve(STy->getElementTypes().size());
77 for (StructType::ElementTypes::const_iterator
78 I = STy->getElementTypes().begin(),
79 E = STy->getElementTypes().end(); I != E; ++I)
80 NewElTypes.push_back(getPointerTransformedType(*I));
81 return StructType::get(NewElTypes);
82 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
83 return ArrayType::get(getPointerTransformedType(ATy->getElementType()),
84 ATy->getNumElements());
86 assert(Ty->isPrimitiveType() && "Unknown derived type!");
93 DSNode *Node; // The node this pool allocation represents
94 Value *Handle; // LLVM value of the pool in the current context
95 const Type *NewType; // The transformed type of the memory objects
96 const Type *PoolType; // The type of the pool
98 const Type *getOldType() const { return Node->getType(); }
100 PoolInfo() { // Define a default ctor for map::operator[]
101 cerr << "Map subscript used to get element that doesn't exist!\n";
105 PoolInfo(DSNode *N, Value *H, const Type *NT, const Type *PT)
106 : Node(N), Handle(H), NewType(NT), PoolType(PT) {
107 // Handle can be null...
108 assert(N && NT && PT && "Pool info null!");
111 PoolInfo(DSNode *N) : Node(N), Handle(0), NewType(0), PoolType(0) {
112 assert(N && "Invalid pool info!");
114 // The new type of the memory object is the same as the old type, except
115 // that all of the pointer values are replaced with POINTERTYPE values.
116 NewType = getPointerTransformedType(getOldType());
120 // ScalarInfo - Information about an LLVM value that we know points to some
121 // datastructure we are processing.
124 Value *Val; // Scalar value in Current Function
125 PoolInfo Pool; // The pool the scalar points into
127 ScalarInfo(Value *V, const PoolInfo &PI) : Val(V), Pool(PI) {
128 assert(V && "Null value passed to ScalarInfo ctor!");
132 // CallArgInfo - Information on one operand for a call that got expanded.
134 int ArgNo; // Call argument number this corresponds to
135 DSNode *Node; // The graph node for the pool
136 Value *PoolHandle; // The LLVM value that is the pool pointer
138 CallArgInfo(int Arg, DSNode *N, Value *PH)
139 : ArgNo(Arg), Node(N), PoolHandle(PH) {
140 assert(Arg >= -1 && N && PH && "Illegal values to CallArgInfo ctor!");
143 // operator< when sorting, sort by argument number.
144 bool operator<(const CallArgInfo &CAI) const {
145 return ArgNo < CAI.ArgNo;
149 // TransformFunctionInfo - Information about how a function eeds to be
152 struct TransformFunctionInfo {
153 // ArgInfo - Maintain information about the arguments that need to be
154 // processed. Each CallArgInfo corresponds to an argument that needs to
155 // have a pool pointer passed into the transformed function with it.
157 // As a special case, "argument" number -1 corresponds to the return value.
159 vector<CallArgInfo> ArgInfo;
161 // Func - The function to be transformed...
164 // The call instruction that is used to map CallArgInfo PoolHandle values
165 // into the new function values.
169 TransformFunctionInfo() : Func(0), Call(0) {}
171 bool operator<(const TransformFunctionInfo &TFI) const {
172 if (Func < TFI.Func) return true;
173 if (Func > TFI.Func) return false;
174 if (ArgInfo.size() < TFI.ArgInfo.size()) return true;
175 if (ArgInfo.size() > TFI.ArgInfo.size()) return false;
176 return ArgInfo < TFI.ArgInfo;
179 void finalizeConstruction() {
180 // Sort the vector so that the return value is first, followed by the
181 // argument records, in order. Note that this must be a stable sort so
182 // that the entries with the same sorting criteria (ie they are multiple
183 // pool entries for the same argument) are kept in depth first order.
184 std::stable_sort(ArgInfo.begin(), ArgInfo.end());
187 // addCallInfo - For a specified function call CI, figure out which pool
188 // descriptors need to be passed in as arguments, and which arguments need
189 // to be transformed into indices. If Arg != -1, the specified call
190 // argument is passed in as a pointer to a data structure.
192 void addCallInfo(DataStructure *DS, CallInst *CI, int Arg,
193 DSNode *GraphNode, map<DSNode*, PoolInfo> &PoolDescs);
195 // Make sure that all dependant arguments are added to this transformation
196 // info. For example, if we call foo(null, P) and foo treats it's first and
197 // second arguments as belonging to the same data structure, the we MUST add
198 // entries to know that the null needs to be transformed into an index as
201 void ensureDependantArgumentsIncluded(DataStructure *DS,
202 map<DSNode*, PoolInfo> &PoolDescs);
206 // Define the pass class that we implement...
207 struct PoolAllocate : public Pass {
209 switch (ReqPointerSize) {
210 case Ptr32bits: POINTERTYPE = Type::UIntTy; break;
211 case Ptr16bits: POINTERTYPE = Type::UShortTy; break;
212 case Ptr8bits: POINTERTYPE = Type::UByteTy; break;
215 CurModule = 0; DS = 0;
216 PoolInit = PoolDestroy = PoolAlloc = PoolFree = 0;
219 // getPoolType - Get the type used by the backend for a pool of a particular
220 // type. This pool record is used to allocate nodes of type NodeType.
222 // Here, PoolTy = { NodeType*, sbyte*, uint }*
224 const StructType *getPoolType(const Type *NodeType) {
225 vector<const Type*> PoolElements;
226 PoolElements.push_back(PointerType::get(NodeType));
227 PoolElements.push_back(PointerType::get(Type::SByteTy));
228 PoolElements.push_back(Type::UIntTy);
229 StructType *Result = StructType::get(PoolElements);
231 // Add a name to the symbol table to correspond to the backend
232 // representation of this pool...
233 assert(CurModule && "No current module!?");
234 string Name = CurModule->getTypeName(NodeType);
235 if (Name.empty()) Name = CurModule->getTypeName(PoolElements[0]);
236 CurModule->addTypeName(Name+"oolbe", Result);
243 // getAnalysisUsage - This function requires data structure information
244 // to be able to see what is pool allocatable.
246 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
247 AU.addRequired<DataStructure>();
251 // CurModule - The module being processed.
254 // DS - The data structure graph for the module being processed.
257 // Prototypes that we add to support pool allocation...
258 Function *PoolInit, *PoolDestroy, *PoolAlloc, *PoolAllocArray, *PoolFree;
260 // The map of already transformed functions... note that the keys of this
261 // map do not have meaningful values for 'Call' or the 'PoolHandle' elements
262 // of the ArgInfo elements.
264 map<TransformFunctionInfo, Function*> TransformedFunctions;
266 // getTransformedFunction - Get a transformed function, or return null if
267 // the function specified hasn't been transformed yet.
269 Function *getTransformedFunction(TransformFunctionInfo &TFI) const {
270 map<TransformFunctionInfo, Function*>::const_iterator I =
271 TransformedFunctions.find(TFI);
272 if (I != TransformedFunctions.end()) return I->second;
277 // addPoolPrototypes - Add prototypes for the pool functions to the
278 // specified module and update the Pool* instance variables to point to
281 void addPoolPrototypes(Module &M);
284 // CreatePools - Insert instructions into the function we are processing to
285 // create all of the memory pool objects themselves. This also inserts
286 // destruction code. Add an alloca for each pool that is allocated to the
289 void CreatePools(Function *F, const vector<AllocDSNode*> &Allocs,
290 map<DSNode*, PoolInfo> &PoolDescs);
292 // processFunction - Convert a function to use pool allocation where
295 bool processFunction(Function *F);
297 // transformFunctionBody - This transforms the instruction in 'F' to use the
298 // pools specified in PoolDescs when modifying data structure nodes
299 // specified in the PoolDescs map. IPFGraph is the closed data structure
300 // graph for F, of which the PoolDescriptor nodes come from.
302 void transformFunctionBody(Function *F, FunctionDSGraph &IPFGraph,
303 map<DSNode*, PoolInfo> &PoolDescs);
305 // transformFunction - Transform the specified function the specified way.
306 // It we have already transformed that function that way, don't do anything.
307 // The nodes in the TransformFunctionInfo come out of callers data structure
308 // graph, and the PoolDescs passed in are the caller's.
310 void transformFunction(TransformFunctionInfo &TFI,
311 FunctionDSGraph &CallerIPGraph,
312 map<DSNode*, PoolInfo> &PoolDescs);
316 RegisterOpt<PoolAllocate> X("poolalloc",
317 "Pool allocate disjoint datastructures");
320 // isNotPoolableAlloc - This is a predicate that returns true if the specified
321 // allocation node in a data structure graph is eligable for pool allocation.
323 static bool isNotPoolableAlloc(const AllocDSNode *DS) {
324 if (DS->isAllocaNode()) return true; // Do not pool allocate alloca's.
328 // processFunction - Convert a function to use pool allocation where
331 bool PoolAllocate::processFunction(Function *F) {
332 // Get the closed datastructure graph for the current function... if there are
333 // any allocations in this graph that are not escaping, we need to pool
334 // allocate them here!
336 FunctionDSGraph &IPGraph = DS->getClosedDSGraph(F);
338 // Get all of the allocations that do not escape the current function. Since
339 // they are still live (they exist in the graph at all), this means we must
340 // have scalar references to these nodes, but the scalars are never returned.
342 vector<AllocDSNode*> Allocs;
343 IPGraph.getNonEscapingAllocations(Allocs);
345 // Filter out allocations that we cannot handle. Currently, this includes
346 // variable sized array allocations and alloca's (which we do not want to
349 Allocs.erase(std::remove_if(Allocs.begin(), Allocs.end(), isNotPoolableAlloc),
353 if (Allocs.empty()) return false; // Nothing to do.
355 #ifdef DEBUG_TRANSFORM_PROGRESS
356 cerr << "Transforming Function: " << F->getName() << "\n";
359 // Insert instructions into the function we are processing to create all of
360 // the memory pool objects themselves. This also inserts destruction code.
361 // This fills in the PoolDescs map to associate the alloc node with the
362 // allocation of the memory pool corresponding to it.
364 map<DSNode*, PoolInfo> PoolDescs;
365 CreatePools(F, Allocs, PoolDescs);
367 #ifdef DEBUG_TRANSFORM_PROGRESS
368 cerr << "Transformed Entry Function: \n" << F;
371 // Now we need to figure out what called functions we need to transform, and
372 // how. To do this, we look at all of the scalars, seeing which functions are
373 // either used as a scalar value (so they return a data structure), or are
374 // passed one of our scalar values.
376 transformFunctionBody(F, IPGraph, PoolDescs);
382 //===----------------------------------------------------------------------===//
384 // NewInstructionCreator - This class is used to traverse the function being
385 // modified, changing each instruction visit'ed to use and provide pointer
386 // indexes instead of real pointers. This is what changes the body of a
387 // function to use pool allocation.
389 class NewInstructionCreator : public InstVisitor<NewInstructionCreator> {
390 PoolAllocate &PoolAllocator;
391 vector<ScalarInfo> &Scalars;
392 map<CallInst*, TransformFunctionInfo> &CallMap;
393 map<Value*, Value*> &XFormMap; // Map old pointers to new indexes
396 Instruction *I; // Instruction to update
397 unsigned OpNum; // Operand number to update
398 Value *OldVal; // The old value it had
400 RefToUpdate(Instruction *i, unsigned o, Value *ov)
401 : I(i), OpNum(o), OldVal(ov) {}
403 vector<RefToUpdate> ReferencesToUpdate;
405 const ScalarInfo &getScalarRef(const Value *V) {
406 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
407 if (Scalars[i].Val == V) return Scalars[i];
409 cerr << "Could not find scalar " << V << " in scalar map!\n";
410 assert(0 && "Scalar not found in getScalar!");
415 const ScalarInfo *getScalar(const Value *V) {
416 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
417 if (Scalars[i].Val == V) return &Scalars[i];
421 BasicBlock::iterator ReplaceInstWith(Instruction &I, Instruction *New) {
422 BasicBlock *BB = I.getParent();
423 BasicBlock::iterator RI = &I;
424 BB->getInstList().remove(RI);
425 BB->getInstList().insert(RI, New);
430 Instruction *createPoolBaseInstruction(Value *PtrVal) {
431 const ScalarInfo &SC = getScalarRef(PtrVal);
432 vector<Value*> Args(3);
433 Args[0] = ConstantUInt::get(Type::UIntTy, 0); // No pointer offset
434 Args[1] = ConstantUInt::get(Type::UByteTy, 0); // Field #0 of pool descriptr
435 Args[2] = ConstantUInt::get(Type::UByteTy, 0); // Field #0 of poolalloc val
436 return new LoadInst(SC.Pool.Handle, Args, PtrVal->getName()+".poolbase");
441 NewInstructionCreator(PoolAllocate &PA, vector<ScalarInfo> &S,
442 map<CallInst*, TransformFunctionInfo> &C,
443 map<Value*, Value*> &X)
444 : PoolAllocator(PA), Scalars(S), CallMap(C), XFormMap(X) {}
447 // updateReferences - The NewInstructionCreator is responsible for creating
448 // new instructions to replace the old ones in the function, and then link up
449 // references to values to their new values. For it to do this, however, it
450 // keeps track of information about the value mapping of old values to new
451 // values that need to be patched up. Given this value map and a set of
452 // instruction operands to patch, updateReferences performs the updates.
454 void updateReferences() {
455 for (unsigned i = 0, e = ReferencesToUpdate.size(); i != e; ++i) {
456 RefToUpdate &Ref = ReferencesToUpdate[i];
457 Value *NewVal = XFormMap[Ref.OldVal];
460 if (isa<Constant>(Ref.OldVal) && // Refering to a null ptr?
461 cast<Constant>(Ref.OldVal)->isNullValue()) {
462 // Transform the null pointer into a null index... caching in XFormMap
463 XFormMap[Ref.OldVal] = NewVal = Constant::getNullValue(POINTERTYPE);
464 //} else if (isa<Argument>(Ref.OldVal)) {
466 cerr << "Unknown reference to: " << Ref.OldVal << "\n";
467 assert(XFormMap[Ref.OldVal] &&
468 "Reference to value that was not updated found!");
472 Ref.I->setOperand(Ref.OpNum, NewVal);
474 ReferencesToUpdate.clear();
477 //===--------------------------------------------------------------------===//
478 // Transformation methods:
479 // These methods specify how each type of instruction is transformed by the
480 // NewInstructionCreator instance...
481 //===--------------------------------------------------------------------===//
483 void visitGetElementPtrInst(GetElementPtrInst &I) {
484 assert(0 && "Cannot transform get element ptr instructions yet!");
487 // Replace the load instruction with a new one.
488 void visitLoadInst(LoadInst &I) {
489 vector<Instruction *> BeforeInsts;
491 // Cast our index to be a UIntTy so we can use it to index into the pool...
492 CastInst *Index = new CastInst(Constant::getNullValue(POINTERTYPE),
493 Type::UIntTy, I.getOperand(0)->getName());
494 BeforeInsts.push_back(Index);
495 ReferencesToUpdate.push_back(RefToUpdate(Index, 0, I.getOperand(0)));
497 // Include the pool base instruction...
498 Instruction *PoolBase = createPoolBaseInstruction(I.getOperand(0));
499 BeforeInsts.push_back(PoolBase);
501 Instruction *IdxInst =
502 BinaryOperator::create(Instruction::Add, *I.idx_begin(), Index,
504 BeforeInsts.push_back(IdxInst);
506 vector<Value*> Indices(I.idx_begin(), I.idx_end());
507 Indices[0] = IdxInst;
508 Instruction *Address = new GetElementPtrInst(PoolBase, Indices,
509 I.getName()+".addr");
510 BeforeInsts.push_back(Address);
512 Instruction *NewLoad = new LoadInst(Address, I.getName());
514 // Replace the load instruction with the new load instruction...
515 BasicBlock::iterator II = ReplaceInstWith(I, NewLoad);
517 // Add all of the instructions before the load...
518 NewLoad->getParent()->getInstList().insert(II, BeforeInsts.begin(),
521 // If not yielding a pool allocated pointer, use the new load value as the
522 // value in the program instead of the old load value...
525 I.replaceAllUsesWith(NewLoad);
528 // Replace the store instruction with a new one. In the store instruction,
529 // the value stored could be a pointer type, meaning that the new store may
530 // have to change one or both of it's operands.
532 void visitStoreInst(StoreInst &I) {
533 assert(getScalar(I.getOperand(1)) &&
534 "Store inst found only storing pool allocated pointer. "
537 Value *Val = I.getOperand(0); // The value to store...
539 // Check to see if the value we are storing is a data structure pointer...
540 //if (const ScalarInfo *ValScalar = getScalar(I.getOperand(0)))
541 if (isa<PointerType>(I.getOperand(0)->getType()))
542 Val = Constant::getNullValue(POINTERTYPE); // Yes, store a dummy
544 Instruction *PoolBase = createPoolBaseInstruction(I.getOperand(1));
546 // Cast our index to be a UIntTy so we can use it to index into the pool...
547 CastInst *Index = new CastInst(Constant::getNullValue(POINTERTYPE),
548 Type::UIntTy, I.getOperand(1)->getName());
549 ReferencesToUpdate.push_back(RefToUpdate(Index, 0, I.getOperand(1)));
551 // Instructions to add after the Index...
552 vector<Instruction*> AfterInsts;
554 Instruction *IdxInst =
555 BinaryOperator::create(Instruction::Add, *I.idx_begin(), Index, "idx");
556 AfterInsts.push_back(IdxInst);
558 vector<Value*> Indices(I.idx_begin(), I.idx_end());
559 Indices[0] = IdxInst;
560 Instruction *Address = new GetElementPtrInst(PoolBase, Indices,
561 I.getName()+"storeaddr");
562 AfterInsts.push_back(Address);
564 Instruction *NewStore = new StoreInst(Val, Address);
565 AfterInsts.push_back(NewStore);
566 if (Val != I.getOperand(0)) // Value stored was a pointer?
567 ReferencesToUpdate.push_back(RefToUpdate(NewStore, 0, I.getOperand(0)));
570 // Replace the store instruction with the cast instruction...
571 BasicBlock::iterator II = ReplaceInstWith(I, Index);
573 // Add the pool base calculator instruction before the index...
574 II = ++Index->getParent()->getInstList().insert(II, PoolBase);
577 // Add the instructions that go after the index...
578 Index->getParent()->getInstList().insert(II, AfterInsts.begin(),
583 // Create call to poolalloc for every malloc instruction
584 void visitMallocInst(MallocInst &I) {
585 const ScalarInfo &SCI = getScalarRef(&I);
589 if (!I.isArrayAllocation()) {
590 Args.push_back(SCI.Pool.Handle);
591 Call = new CallInst(PoolAllocator.PoolAlloc, Args, I.getName());
593 Args.push_back(I.getArraySize());
594 Args.push_back(SCI.Pool.Handle);
595 Call = new CallInst(PoolAllocator.PoolAllocArray, Args, I.getName());
598 ReplaceInstWith(I, Call);
601 // Convert a call to poolfree for every free instruction...
602 void visitFreeInst(FreeInst &I) {
603 // Create a new call to poolfree before the free instruction
605 Args.push_back(Constant::getNullValue(POINTERTYPE));
606 Args.push_back(getScalarRef(I.getOperand(0)).Pool.Handle);
607 Instruction *NewCall = new CallInst(PoolAllocator.PoolFree, Args);
608 ReplaceInstWith(I, NewCall);
609 ReferencesToUpdate.push_back(RefToUpdate(NewCall, 1, I.getOperand(0)));
612 // visitCallInst - Create a new call instruction with the extra arguments for
613 // all of the memory pools that the call needs.
615 void visitCallInst(CallInst &I) {
616 TransformFunctionInfo &TI = CallMap[&I];
618 // Start with all of the old arguments...
619 vector<Value*> Args(I.op_begin()+1, I.op_end());
621 for (unsigned i = 0, e = TI.ArgInfo.size(); i != e; ++i) {
622 // Replace all of the pointer arguments with our new pointer typed values.
623 if (TI.ArgInfo[i].ArgNo != -1)
624 Args[TI.ArgInfo[i].ArgNo] = Constant::getNullValue(POINTERTYPE);
626 // Add all of the pool arguments...
627 Args.push_back(TI.ArgInfo[i].PoolHandle);
630 Function *NF = PoolAllocator.getTransformedFunction(TI);
631 Instruction *NewCall = new CallInst(NF, Args, I.getName());
632 ReplaceInstWith(I, NewCall);
634 // Keep track of the mapping of operands so that we can resolve them to real
636 Value *RetVal = NewCall;
637 for (unsigned i = 0, e = TI.ArgInfo.size(); i != e; ++i)
638 if (TI.ArgInfo[i].ArgNo != -1)
639 ReferencesToUpdate.push_back(RefToUpdate(NewCall, TI.ArgInfo[i].ArgNo+1,
640 I.getOperand(TI.ArgInfo[i].ArgNo+1)));
642 RetVal = 0; // If returning a pointer, don't change retval...
644 // If not returning a pointer, use the new call as the value in the program
645 // instead of the old call...
648 I.replaceAllUsesWith(RetVal);
651 // visitPHINode - Create a new PHI node of POINTERTYPE for all of the old Phi
654 void visitPHINode(PHINode &PN) {
655 Value *DummyVal = Constant::getNullValue(POINTERTYPE);
656 PHINode *NewPhi = new PHINode(POINTERTYPE, PN.getName());
657 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
658 NewPhi->addIncoming(DummyVal, PN.getIncomingBlock(i));
659 ReferencesToUpdate.push_back(RefToUpdate(NewPhi, i*2,
660 PN.getIncomingValue(i)));
663 ReplaceInstWith(PN, NewPhi);
666 // visitReturnInst - Replace ret instruction with a new return...
667 void visitReturnInst(ReturnInst &I) {
668 Instruction *Ret = new ReturnInst(Constant::getNullValue(POINTERTYPE));
669 ReplaceInstWith(I, Ret);
670 ReferencesToUpdate.push_back(RefToUpdate(Ret, 0, I.getOperand(0)));
673 // visitSetCondInst - Replace a conditional test instruction with a new one
674 void visitSetCondInst(SetCondInst &SCI) {
675 BinaryOperator &I = (BinaryOperator&)SCI;
676 Value *DummyVal = Constant::getNullValue(POINTERTYPE);
677 BinaryOperator *New = BinaryOperator::create(I.getOpcode(), DummyVal,
678 DummyVal, I.getName());
679 ReplaceInstWith(I, New);
681 ReferencesToUpdate.push_back(RefToUpdate(New, 0, I.getOperand(0)));
682 ReferencesToUpdate.push_back(RefToUpdate(New, 1, I.getOperand(1)));
684 // Make sure branches refer to the new condition...
685 I.replaceAllUsesWith(New);
688 void visitInstruction(Instruction &I) {
689 cerr << "Unknown instruction to FunctionBodyTransformer:\n" << I;
694 // PoolBaseLoadEliminator - Every load and store through a pool allocated
695 // pointer causes a load of the real pool base out of the pool descriptor.
696 // Iterate through the function, doing a local elimination pass of duplicate
697 // loads. This attempts to turn the all too common:
699 // %reg109.poolbase22 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
700 // %reg207 = load %root.p* %reg109.poolbase22, uint %reg109, ubyte 0, ubyte 0
701 // %reg109.poolbase23 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
702 // store double %reg207, %root.p* %reg109.poolbase23, uint %reg109, ...
705 // %reg109.poolbase22 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
706 // %reg207 = load %root.p* %reg109.poolbase22, uint %reg109, ubyte 0, ubyte 0
707 // store double %reg207, %root.p* %reg109.poolbase22, uint %reg109, ...
710 class PoolBaseLoadEliminator : public InstVisitor<PoolBaseLoadEliminator> {
711 // PoolDescValues - Keep track of the values in the current function that are
712 // pool descriptors (loads from which we want to eliminate).
714 vector<Value*> PoolDescValues;
716 // PoolDescMap - As we are analyzing a BB, keep track of which load to use
717 // when referencing a pool descriptor.
719 map<Value*, LoadInst*> PoolDescMap;
721 // These two fields keep track of statistics of how effective we are, if
722 // debugging is enabled.
724 unsigned Eliminated, Remaining;
726 // Compact the pool descriptor map into a list of the pool descriptors in the
727 // current context that we should know about...
729 PoolBaseLoadEliminator(const map<DSNode*, PoolInfo> &PoolDescs) {
730 Eliminated = Remaining = 0;
731 for (map<DSNode*, PoolInfo>::const_iterator I = PoolDescs.begin(),
732 E = PoolDescs.end(); I != E; ++I)
733 PoolDescValues.push_back(I->second.Handle);
735 // Remove duplicates from the list of pool values
736 sort(PoolDescValues.begin(), PoolDescValues.end());
737 PoolDescValues.erase(unique(PoolDescValues.begin(), PoolDescValues.end()),
738 PoolDescValues.end());
741 #ifdef DEBUG_POOLBASE_LOAD_ELIMINATOR
742 void visitFunction(Function &F) {
743 cerr << "Pool Load Elim '" << F.getName() << "'\t";
745 ~PoolBaseLoadEliminator() {
746 unsigned Total = Eliminated+Remaining;
748 cerr << "removed " << Eliminated << "["
749 << Eliminated*100/Total << "%] loads, leaving "
750 << Remaining << ".\n";
754 // Loop over the function, looking for loads to eliminate. Because we are a
755 // local transformation, we reset all of our state when we enter a new basic
758 void visitBasicBlock(BasicBlock &) {
759 PoolDescMap.clear(); // Forget state.
762 // Starting with an empty basic block, we scan it looking for loads of the
763 // pool descriptor. When we find a load, we add it to the PoolDescMap,
764 // indicating that we have a value available to recycle next time we see the
765 // poolbase of this instruction being loaded.
767 void visitLoadInst(LoadInst &LI) {
768 Value *LoadAddr = LI.getPointerOperand();
769 map<Value*, LoadInst*>::iterator VIt = PoolDescMap.find(LoadAddr);
770 if (VIt != PoolDescMap.end()) { // We already have a value for this load?
771 LI.replaceAllUsesWith(VIt->second); // Make the current load dead
774 // This load might not be a load of a pool pointer, check to see if it is
775 if (LI.getNumOperands() == 4 && // load pool, uint 0, ubyte 0, ubyte 0
776 find(PoolDescValues.begin(), PoolDescValues.end(), LoadAddr) !=
777 PoolDescValues.end()) {
779 assert("Make sure it's a load of the pool base, not a chaining field" &&
780 LI.getOperand(1) == Constant::getNullValue(Type::UIntTy) &&
781 LI.getOperand(2) == Constant::getNullValue(Type::UByteTy) &&
782 LI.getOperand(3) == Constant::getNullValue(Type::UByteTy));
784 // If it is a load of a pool base, keep track of it for future reference
785 PoolDescMap.insert(std::make_pair(LoadAddr, &LI));
791 // If we run across a function call, forget all state... Calls to
792 // poolalloc/poolfree can invalidate the pool base pointer, so it should be
793 // reloaded the next time it is used. Furthermore, a call to a random
794 // function might call one of these functions, so be conservative. Through
795 // more analysis, this could be improved in the future.
797 void visitCallInst(CallInst &) {
802 static void addNodeMapping(DSNode *SrcNode, const PointerValSet &PVS,
803 map<DSNode*, PointerValSet> &NodeMapping) {
804 for (unsigned i = 0, e = PVS.size(); i != e; ++i)
805 if (NodeMapping[SrcNode].add(PVS[i])) { // Not in map yet?
806 assert(PVS[i].Index == 0 && "Node indexing not supported yet!");
807 DSNode *DestNode = PVS[i].Node;
809 // Loop over all of the outgoing links in the mapped graph
810 for (unsigned l = 0, le = DestNode->getNumOutgoingLinks(); l != le; ++l) {
811 PointerValSet &SrcSet = SrcNode->getOutgoingLink(l);
812 const PointerValSet &DestSet = DestNode->getOutgoingLink(l);
814 // Add all of the node mappings now!
815 for (unsigned si = 0, se = SrcSet.size(); si != se; ++si) {
816 assert(SrcSet[si].Index == 0 && "Can't handle node offset!");
817 addNodeMapping(SrcSet[si].Node, DestSet, NodeMapping);
823 // CalculateNodeMapping - There is a partial isomorphism between the graph
824 // passed in and the graph that is actually used by the function. We need to
825 // figure out what this mapping is so that we can transformFunctionBody the
826 // instructions in the function itself. Note that every node in the graph that
827 // we are interested in must be both in the local graph of the called function,
828 // and in the local graph of the calling function. Because of this, we only
829 // define the mapping for these nodes [conveniently these are the only nodes we
830 // CAN define a mapping for...]
832 // The roots of the graph that we are transforming is rooted in the arguments
833 // passed into the function from the caller. This is where we start our
834 // mapping calculation.
836 // The NodeMapping calculated maps from the callers graph to the called graph.
838 static void CalculateNodeMapping(Function *F, TransformFunctionInfo &TFI,
839 FunctionDSGraph &CallerGraph,
840 FunctionDSGraph &CalledGraph,
841 map<DSNode*, PointerValSet> &NodeMapping) {
843 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
844 // Figure out what nodes in the called graph the TFI.ArgInfo[i].Node node
847 // Only consider first node of sequence. Extra nodes may may be added
848 // to the TFI if the data structure requires more nodes than just the
849 // one the argument points to. We are only interested in the one the
850 // argument points to though.
852 if (TFI.ArgInfo[i].ArgNo != LastArgNo) {
853 if (TFI.ArgInfo[i].ArgNo == -1) {
854 addNodeMapping(TFI.ArgInfo[i].Node, CalledGraph.getRetNodes(),
857 // Figure out which node argument # ArgNo points to in the called graph.
858 Function::aiterator AI = F->abegin();
859 std::advance(AI, TFI.ArgInfo[i].ArgNo);
860 addNodeMapping(TFI.ArgInfo[i].Node, CalledGraph.getValueMap()[AI],
863 LastArgNo = TFI.ArgInfo[i].ArgNo;
871 // addCallInfo - For a specified function call CI, figure out which pool
872 // descriptors need to be passed in as arguments, and which arguments need to be
873 // transformed into indices. If Arg != -1, the specified call argument is
874 // passed in as a pointer to a data structure.
876 void TransformFunctionInfo::addCallInfo(DataStructure *DS, CallInst *CI,
877 int Arg, DSNode *GraphNode,
878 map<DSNode*, PoolInfo> &PoolDescs) {
879 assert(CI->getCalledFunction() && "Cannot handle indirect calls yet!");
880 assert(Func == 0 || Func == CI->getCalledFunction() &&
881 "Function call record should always call the same function!");
882 assert(Call == 0 || Call == CI &&
883 "Call element already filled in with different value!");
884 Func = CI->getCalledFunction();
886 //FunctionDSGraph &CalledGraph = DS->getClosedDSGraph(Func);
888 // For now, add the entire graph that is pointed to by the call argument.
889 // This graph can and should be pruned to only what the function itself will
890 // use, because often this will be a dramatically smaller subset of what we
893 // FIXME: This should use pool links instead of extra arguments!
895 for (df_iterator<DSNode*> I = df_begin(GraphNode), E = df_end(GraphNode);
897 ArgInfo.push_back(CallArgInfo(Arg, *I, PoolDescs[*I].Handle));
900 static void markReachableNodes(const PointerValSet &Vals,
901 set<DSNode*> &ReachableNodes) {
902 for (unsigned n = 0, ne = Vals.size(); n != ne; ++n) {
903 DSNode *N = Vals[n].Node;
904 if (ReachableNodes.count(N) == 0) // Haven't already processed node?
905 ReachableNodes.insert(df_begin(N), df_end(N)); // Insert all
909 // Make sure that all dependant arguments are added to this transformation info.
910 // For example, if we call foo(null, P) and foo treats it's first and second
911 // arguments as belonging to the same data structure, the we MUST add entries to
912 // know that the null needs to be transformed into an index as well.
914 void TransformFunctionInfo::ensureDependantArgumentsIncluded(DataStructure *DS,
915 map<DSNode*, PoolInfo> &PoolDescs) {
916 // FIXME: This does not work for indirect function calls!!!
917 if (Func == 0) return; // FIXME!
919 // Make sure argument entries are sorted.
920 finalizeConstruction();
922 // Loop over the function signature, checking to see if there are any pointer
923 // arguments that we do not convert... if there is something we haven't
924 // converted, set done to false.
928 if (isa<PointerType>(Func->getReturnType())) // Make sure we convert retval
929 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == -1) {
930 // We DO transform the ret val... skip all possible entries for retval
931 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == -1)
938 for (Function::aiterator I = Func->abegin(), E = Func->aend(); I!=E; ++I,++i){
939 if (isa<PointerType>(I->getType())) {
940 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == (int)i) {
941 // We DO transform this arg... skip all possible entries for argument
942 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == (int)i)
951 // If we already have entries for all pointer arguments and retvals, there
952 // certainly is no work to do. Bail out early to avoid building relatively
953 // expensive data structures.
957 #ifdef DEBUG_TRANSFORM_PROGRESS
958 cerr << "Must ensure dependant arguments for: " << Func->getName() << "\n";
961 // Otherwise, we MIGHT have to add the arguments/retval if they are part of
962 // the same datastructure graph as some other argument or retval that we ARE
965 // Get the data structure graph for the called function.
967 FunctionDSGraph &CalledDS = DS->getClosedDSGraph(Func);
969 // Build a mapping between the nodes in our current graph and the nodes in the
970 // called function's graph. We build it based on our _incomplete_
971 // transformation information, because it contains all of the info that we
974 map<DSNode*, PointerValSet> NodeMapping;
975 CalculateNodeMapping(Func, *this,
976 DS->getClosedDSGraph(Call->getParent()->getParent()),
977 CalledDS, NodeMapping);
979 // Build the inverted version of the node mapping, that maps from a node in
980 // the called functions graph to a single node in the caller graph.
982 map<DSNode*, DSNode*> InverseNodeMap;
983 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin(),
984 E = NodeMapping.end(); I != E; ++I) {
985 PointerValSet &CalledNodes = I->second;
986 for (unsigned i = 0, e = CalledNodes.size(); i != e; ++i)
987 InverseNodeMap[CalledNodes[i].Node] = I->first;
989 NodeMapping.clear(); // Done with information, free memory
991 // Build a set of reachable nodes from the arguments/retval that we ARE
993 set<DSNode*> ReachableNodes;
995 // Loop through all of the arguments, marking all of the reachable data
996 // structure nodes reachable if they are from this pointer...
998 for (unsigned i = 0, e = ArgInfo.size(); i != e; ++i) {
999 if (ArgInfo[i].ArgNo == -1) {
1000 if (i == 0) // Only process retvals once (performance opt)
1001 markReachableNodes(CalledDS.getRetNodes(), ReachableNodes);
1002 } else { // If it's an argument value...
1003 Function::aiterator AI = Func->abegin();
1004 std::advance(AI, ArgInfo[i].ArgNo);
1005 if (isa<PointerType>(AI->getType()))
1006 markReachableNodes(CalledDS.getValueMap()[AI], ReachableNodes);
1010 // Now that we know which nodes are already reachable, see if any of the
1011 // arguments that we are not passing values in for can reach one of the
1012 // existing nodes...
1015 // <FIXME> IN THEORY, we should allow arbitrary paths from the argument to
1016 // nodes we know about. The problem is that if we do this, then I don't know
1017 // how to get pool pointers for this head list. Since we are completely
1018 // deadline driven, I'll just allow direct accesses to the graph. </FIXME>
1022 if (isa<PointerType>(Func->getReturnType())) // Make sure we convert retval
1023 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == -1) {
1024 // We DO transform the ret val... skip all possible entries for retval
1025 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == -1)
1028 // See what the return value points to...
1030 // FIXME: This should generalize to any number of nodes, just see if any
1032 assert(CalledDS.getRetNodes().size() == 1 &&
1033 "Assumes only one node is returned");
1034 DSNode *N = CalledDS.getRetNodes()[0].Node;
1036 // If the return value is not marked as being passed in, but it NEEDS to
1037 // be transformed, then make it known now.
1039 if (ReachableNodes.count(N)) {
1040 #ifdef DEBUG_TRANSFORM_PROGRESS
1041 cerr << "ensure dependant arguments adds return value entry!\n";
1043 addCallInfo(DS, Call, -1, InverseNodeMap[N], PoolDescs);
1046 finalizeConstruction();
1051 for (Function::aiterator I = Func->abegin(), E = Func->aend(); I!=E; ++I, ++i)
1052 if (isa<PointerType>(I->getType())) {
1053 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == (int)i) {
1054 // We DO transform this arg... skip all possible entries for argument
1055 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == (int)i)
1058 // This should generalize to any number of nodes, just see if any are
1060 assert(CalledDS.getValueMap()[I].size() == 1 &&
1061 "Only handle case where pointing to one node so far!");
1063 // If the arg is not marked as being passed in, but it NEEDS to
1064 // be transformed, then make it known now.
1066 DSNode *N = CalledDS.getValueMap()[I][0].Node;
1067 if (ReachableNodes.count(N)) {
1068 #ifdef DEBUG_TRANSFORM_PROGRESS
1069 cerr << "ensure dependant arguments adds for arg #" << i << "\n";
1071 addCallInfo(DS, Call, i, InverseNodeMap[N], PoolDescs);
1074 finalizeConstruction();
1081 // transformFunctionBody - This transforms the instruction in 'F' to use the
1082 // pools specified in PoolDescs when modifying data structure nodes specified in
1083 // the PoolDescs map. Specifically, scalar values specified in the Scalars
1084 // vector must be remapped. IPFGraph is the closed data structure graph for F,
1085 // of which the PoolDescriptor nodes come from.
1087 void PoolAllocate::transformFunctionBody(Function *F, FunctionDSGraph &IPFGraph,
1088 map<DSNode*, PoolInfo> &PoolDescs) {
1090 // Loop through the value map looking for scalars that refer to nonescaping
1091 // allocations. Add them to the Scalars vector. Note that we may have
1092 // multiple entries in the Scalars vector for each value if it points to more
1095 map<Value*, PointerValSet> &ValMap = IPFGraph.getValueMap();
1096 vector<ScalarInfo> Scalars;
1098 #ifdef DEBUG_TRANSFORM_PROGRESS
1099 cerr << "Building scalar map for fn '" << F->getName() << "' body:\n";
1102 for (map<Value*, PointerValSet>::iterator I = ValMap.begin(),
1103 E = ValMap.end(); I != E; ++I) {
1104 const PointerValSet &PVS = I->second; // Set of things pointed to by scalar
1106 // Check to see if the scalar points to a data structure node...
1107 for (unsigned i = 0, e = PVS.size(); i != e; ++i) {
1108 if (PVS[i].Index) { cerr << "Problem in " << F->getName() << " for " << I->first << "\n"; }
1109 assert(PVS[i].Index == 0 && "Nonzero not handled yet!");
1111 // If the allocation is in the nonescaping set...
1112 map<DSNode*, PoolInfo>::iterator AI = PoolDescs.find(PVS[i].Node);
1113 if (AI != PoolDescs.end()) { // Add it to the list of scalars
1114 Scalars.push_back(ScalarInfo(I->first, AI->second));
1115 #ifdef DEBUG_TRANSFORM_PROGRESS
1116 cerr << "\nScalar Mapping from:" << I->first
1117 << "Scalar Mapping to: "; PVS.print(cerr);
1123 #ifdef DEBUG_TRANSFORM_PROGRESS
1124 cerr << "\nIn '" << F->getName()
1125 << "': Found the following values that point to poolable nodes:\n";
1127 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
1128 cerr << Scalars[i].Val;
1132 // CallMap - Contain an entry for every call instruction that needs to be
1133 // transformed. Each entry in the map contains information about what we need
1134 // to do to each call site to change it to work.
1136 map<CallInst*, TransformFunctionInfo> CallMap;
1138 // Now we need to figure out what called functions we need to transform, and
1139 // how. To do this, we look at all of the scalars, seeing which functions are
1140 // either used as a scalar value (so they return a data structure), or are
1141 // passed one of our scalar values.
1143 for (unsigned i = 0, e = Scalars.size(); i != e; ++i) {
1144 Value *ScalarVal = Scalars[i].Val;
1146 // Check to see if the scalar _IS_ a call...
1147 if (CallInst *CI = dyn_cast<CallInst>(ScalarVal))
1148 // If so, add information about the pool it will be returning...
1149 CallMap[CI].addCallInfo(DS, CI, -1, Scalars[i].Pool.Node, PoolDescs);
1151 // Check to see if the scalar is an operand to a call...
1152 for (Value::use_iterator UI = ScalarVal->use_begin(),
1153 UE = ScalarVal->use_end(); UI != UE; ++UI) {
1154 if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
1155 // Find out which operand this is to the call instruction...
1156 User::op_iterator OI = find(CI->op_begin(), CI->op_end(), ScalarVal);
1157 assert(OI != CI->op_end() && "Call on use list but not an operand!?");
1158 assert(OI != CI->op_begin() && "Pointer operand is call destination?");
1160 // FIXME: This is broken if the same pointer is passed to a call more
1161 // than once! It will get multiple entries for the first pointer.
1163 // Add the operand number and pool handle to the call table...
1164 CallMap[CI].addCallInfo(DS, CI, OI-CI->op_begin()-1,
1165 Scalars[i].Pool.Node, PoolDescs);
1170 // Make sure that all dependant arguments are added as well. For example, if
1171 // we call foo(null, P) and foo treats it's first and second arguments as
1172 // belonging to the same data structure, the we MUST set up the CallMap to
1173 // know that the null needs to be transformed into an index as well.
1175 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin();
1176 I != CallMap.end(); ++I)
1177 I->second.ensureDependantArgumentsIncluded(DS, PoolDescs);
1179 #ifdef DEBUG_TRANSFORM_PROGRESS
1180 // Print out call map...
1181 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin();
1182 I != CallMap.end(); ++I) {
1183 cerr << "For call: " << I->first;
1184 cerr << I->second.Func->getName() << " must pass pool pointer for args #";
1185 for (unsigned i = 0; i < I->second.ArgInfo.size(); ++i)
1186 cerr << I->second.ArgInfo[i].ArgNo << ", ";
1191 // Loop through all of the call nodes, recursively creating the new functions
1192 // that we want to call... This uses a map to prevent infinite recursion and
1193 // to avoid duplicating functions unneccesarily.
1195 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin(),
1196 E = CallMap.end(); I != E; ++I) {
1197 // Transform all of the functions we need, or at least ensure there is a
1198 // cached version available.
1199 transformFunction(I->second, IPFGraph, PoolDescs);
1202 // Now that all of the functions that we want to call are available, transform
1203 // the local function so that it uses the pools locally and passes them to the
1204 // functions that we just hacked up.
1207 // First step, find the instructions to be modified.
1208 vector<Instruction*> InstToFix;
1209 for (unsigned i = 0, e = Scalars.size(); i != e; ++i) {
1210 Value *ScalarVal = Scalars[i].Val;
1212 // Check to see if the scalar _IS_ an instruction. If so, it is involved.
1213 if (Instruction *Inst = dyn_cast<Instruction>(ScalarVal))
1214 InstToFix.push_back(Inst);
1216 // All all of the instructions that use the scalar as an operand...
1217 for (Value::use_iterator UI = ScalarVal->use_begin(),
1218 UE = ScalarVal->use_end(); UI != UE; ++UI)
1219 InstToFix.push_back(cast<Instruction>(*UI));
1222 // Make sure that we get return instructions that return a null value from the
1225 if (!IPFGraph.getRetNodes().empty()) {
1226 assert(IPFGraph.getRetNodes().size() == 1 && "Can only return one node?");
1227 PointerVal RetNode = IPFGraph.getRetNodes()[0];
1228 assert(RetNode.Index == 0 && "Subindexing not implemented yet!");
1230 // Only process return instructions if the return value of this function is
1231 // part of one of the data structures we are transforming...
1233 if (PoolDescs.count(RetNode.Node)) {
1234 // Loop over all of the basic blocks, adding return instructions...
1235 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
1236 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator()))
1237 InstToFix.push_back(RI);
1243 // Eliminate duplicates by sorting, then removing equal neighbors.
1244 sort(InstToFix.begin(), InstToFix.end());
1245 InstToFix.erase(unique(InstToFix.begin(), InstToFix.end()), InstToFix.end());
1247 // Loop over all of the instructions to transform, creating the new
1248 // replacement instructions for them. This also unlinks them from the
1249 // function so they can be safely deleted later.
1251 map<Value*, Value*> XFormMap;
1252 NewInstructionCreator NIC(*this, Scalars, CallMap, XFormMap);
1254 // Visit all instructions... creating the new instructions that we need and
1255 // unlinking the old instructions from the function...
1257 #ifdef DEBUG_TRANSFORM_PROGRESS
1258 for (unsigned i = 0, e = InstToFix.size(); i != e; ++i) {
1259 cerr << "Fixing: " << InstToFix[i];
1260 NIC.visit(*InstToFix[i]);
1263 NIC.visit(InstToFix.begin(), InstToFix.end());
1266 // Make all instructions we will delete "let go" of their operands... so that
1267 // we can safely delete Arguments whose types have changed...
1269 for_each(InstToFix.begin(), InstToFix.end(),
1270 std::mem_fun(&Instruction::dropAllReferences));
1272 // Loop through all of the pointer arguments coming into the function,
1273 // replacing them with arguments of POINTERTYPE to match the function type of
1276 FunctionType::ParamTypes::const_iterator TI =
1277 F->getFunctionType()->getParamTypes().begin();
1278 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I, ++TI) {
1279 if (I->getType() != *TI) {
1280 assert(isa<PointerType>(I->getType()) && *TI == POINTERTYPE);
1281 Argument *NewArg = new Argument(*TI, I->getName());
1282 XFormMap[I] = NewArg; // Map old arg into new arg...
1284 // Replace the old argument and then delete it...
1285 I = F->getArgumentList().erase(I);
1286 I = F->getArgumentList().insert(I, NewArg);
1290 // Now that all of the new instructions have been created, we can update all
1291 // of the references to dummy values to be references to the actual values
1292 // that are computed.
1294 NIC.updateReferences();
1296 #ifdef DEBUG_TRANSFORM_PROGRESS
1297 cerr << "TRANSFORMED FUNCTION:\n" << F;
1300 // Delete all of the "instructions to fix"
1301 for_each(InstToFix.begin(), InstToFix.end(), deleter<Instruction>);
1303 // Eliminate pool base loads that we can easily prove are redundant
1305 PoolBaseLoadEliminator(PoolDescs).visit(F);
1307 // Since we have liberally hacked the function to pieces, we want to inform
1308 // the datastructure pass that its internal representation is out of date.
1310 DS->invalidateFunction(F);
1315 // transformFunction - Transform the specified function the specified way. It
1316 // we have already transformed that function that way, don't do anything. The
1317 // nodes in the TransformFunctionInfo come out of callers data structure graph.
1319 void PoolAllocate::transformFunction(TransformFunctionInfo &TFI,
1320 FunctionDSGraph &CallerIPGraph,
1321 map<DSNode*, PoolInfo> &CallerPoolDesc) {
1322 if (getTransformedFunction(TFI)) return; // Function xformation already done?
1324 #ifdef DEBUG_TRANSFORM_PROGRESS
1325 cerr << "********** Entering transformFunction for "
1326 << TFI.Func->getName() << ":\n";
1327 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i)
1328 cerr << " ArgInfo[" << i << "] = " << TFI.ArgInfo[i].ArgNo << "\n";
1332 const FunctionType *OldFuncType = TFI.Func->getFunctionType();
1334 assert(!OldFuncType->isVarArg() && "Vararg functions not handled yet!");
1336 // Build the type for the new function that we are transforming
1337 vector<const Type*> ArgTys;
1338 ArgTys.reserve(OldFuncType->getNumParams()+TFI.ArgInfo.size());
1339 for (unsigned i = 0, e = OldFuncType->getNumParams(); i != e; ++i)
1340 ArgTys.push_back(OldFuncType->getParamType(i));
1342 const Type *RetType = OldFuncType->getReturnType();
1344 // Add one pool pointer for every argument that needs to be supplemented.
1345 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
1346 if (TFI.ArgInfo[i].ArgNo == -1)
1347 RetType = POINTERTYPE; // Return a pointer
1349 ArgTys[TFI.ArgInfo[i].ArgNo] = POINTERTYPE; // Pass a pointer
1350 ArgTys.push_back(PointerType::get(CallerPoolDesc.find(TFI.ArgInfo[i].Node)
1351 ->second.PoolType));
1354 // Build the new function type...
1355 const FunctionType *NewFuncType = FunctionType::get(RetType, ArgTys,
1356 OldFuncType->isVarArg());
1358 // The new function is internal, because we know that only we can call it.
1359 // This also helps subsequent IP transformations to eliminate duplicated pool
1360 // pointers (which look like the same value is always passed into a parameter,
1361 // allowing it to be easily eliminated).
1363 Function *NewFunc = new Function(NewFuncType, true,
1364 TFI.Func->getName()+".poolxform");
1365 CurModule->getFunctionList().push_back(NewFunc);
1368 #ifdef DEBUG_TRANSFORM_PROGRESS
1369 cerr << "Created function prototype: " << NewFunc << "\n";
1372 // Add the newly formed function to the TransformedFunctions table so that
1373 // infinite recursion does not occur!
1375 TransformedFunctions[TFI] = NewFunc;
1377 // Add arguments to the function... starting with all of the old arguments
1378 vector<Value*> ArgMap;
1379 for (Function::const_aiterator I = TFI.Func->abegin(), E = TFI.Func->aend();
1381 Argument *NFA = new Argument(I->getType(), I->getName());
1382 NewFunc->getArgumentList().push_back(NFA);
1383 ArgMap.push_back(NFA); // Keep track of the arguments
1386 // Now add all of the arguments corresponding to pools passed in...
1387 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
1388 CallArgInfo &AI = TFI.ArgInfo[i];
1393 Name = ArgMap[AI.ArgNo]->getName(); // Get the arg name
1394 const Type *Ty = PointerType::get(CallerPoolDesc[AI.Node].PoolType);
1395 Argument *NFA = new Argument(Ty, Name+".pool");
1396 NewFunc->getArgumentList().push_back(NFA);
1399 // Now clone the body of the old function into the new function...
1400 CloneFunctionInto(NewFunc, TFI.Func, ArgMap);
1402 // Okay, now we have a function that is identical to the old one, except that
1403 // it has extra arguments for the pools coming in. Now we have to get the
1404 // data structure graph for the function we are replacing, and figure out how
1405 // our graph nodes map to the graph nodes in the dest function.
1407 FunctionDSGraph &DSGraph = DS->getClosedDSGraph(NewFunc);
1409 // NodeMapping - Multimap from callers graph to called graph. We are
1410 // guaranteed that the called function graph has more nodes than the caller,
1411 // or exactly the same number of nodes. This is because the called function
1412 // might not know that two nodes are merged when considering the callers
1413 // context, but the caller obviously does. Because of this, a single node in
1414 // the calling function's data structure graph can map to multiple nodes in
1415 // the called functions graph.
1417 map<DSNode*, PointerValSet> NodeMapping;
1419 CalculateNodeMapping(NewFunc, TFI, CallerIPGraph, DSGraph,
1422 // Print out the node mapping...
1423 #ifdef DEBUG_TRANSFORM_PROGRESS
1424 cerr << "\nNode mapping for call of " << NewFunc->getName() << "\n";
1425 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin();
1426 I != NodeMapping.end(); ++I) {
1427 cerr << "Map: "; I->first->print(cerr);
1428 cerr << "To: "; I->second.print(cerr);
1433 // Fill in the PoolDescriptor information for the transformed function so that
1434 // it can determine which value holds the pool descriptor for each data
1435 // structure node that it accesses.
1437 map<DSNode*, PoolInfo> PoolDescs;
1439 #ifdef DEBUG_TRANSFORM_PROGRESS
1440 cerr << "\nCalculating the pool descriptor map:\n";
1443 // Calculate as much of the pool descriptor map as possible. Since we have
1444 // the node mapping between the caller and callee functions, and we have the
1445 // pool descriptor information of the caller, we can calculate a partical pool
1446 // descriptor map for the called function.
1448 // The nodes that we do not have complete information for are the ones that
1449 // are accessed by loading pointers derived from arguments passed in, but that
1450 // are not passed in directly. In this case, we have all of the information
1451 // except a pool value. If the called function refers to this pool, the pool
1452 // value will be loaded from the pool graph and added to the map as neccesary.
1454 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin();
1455 I != NodeMapping.end(); ++I) {
1456 DSNode *CallerNode = I->first;
1457 PoolInfo &CallerPI = CallerPoolDesc[CallerNode];
1459 // Check to see if we have a node pointer passed in for this value...
1460 Value *CalleeValue = 0;
1461 for (unsigned a = 0, ae = TFI.ArgInfo.size(); a != ae; ++a)
1462 if (TFI.ArgInfo[a].Node == CallerNode) {
1463 // Calculate the argument number that the pool is to the function
1464 // call... The call instruction should not have the pool operands added
1466 unsigned ArgNo = TFI.Call->getNumOperands()-1+a;
1467 #ifdef DEBUG_TRANSFORM_PROGRESS
1468 cerr << "Should be argument #: " << ArgNo << "[i = " << a << "]\n";
1470 assert(ArgNo < NewFunc->asize() &&
1471 "Call already has pool arguments added??");
1473 // Map the pool argument into the called function...
1474 Function::aiterator AI = NewFunc->abegin();
1475 std::advance(AI, ArgNo);
1477 break; // Found value, quit loop
1480 // Loop over all of the data structure nodes that this incoming node maps to
1481 // Creating a PoolInfo structure for them.
1482 for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
1483 assert(I->second[i].Index == 0 && "Doesn't handle subindexing yet!");
1484 DSNode *CalleeNode = I->second[i].Node;
1486 // Add the descriptor. We already know everything about it by now, much
1487 // of it is the same as the caller info.
1489 PoolDescs.insert(std::make_pair(CalleeNode,
1490 PoolInfo(CalleeNode, CalleeValue,
1492 CallerPI.PoolType)));
1496 // We must destroy the node mapping so that we don't have latent references
1497 // into the data structure graph for the new function. Otherwise we get
1498 // assertion failures when transformFunctionBody tries to invalidate the
1501 NodeMapping.clear();
1503 // Now that we know everything we need about the function, transform the body
1506 transformFunctionBody(NewFunc, DSGraph, PoolDescs);
1508 #ifdef DEBUG_TRANSFORM_PROGRESS
1509 cerr << "Function after transformation:\n" << NewFunc;
1513 static unsigned countPointerTypes(const Type *Ty) {
1514 if (isa<PointerType>(Ty)) {
1516 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1518 for (unsigned i = 0, e = STy->getElementTypes().size(); i != e; ++i)
1519 Num += countPointerTypes(STy->getElementTypes()[i]);
1521 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1522 return countPointerTypes(ATy->getElementType());
1524 assert(Ty->isPrimitiveType() && "Unknown derived type!");
1529 // CreatePools - Insert instructions into the function we are processing to
1530 // create all of the memory pool objects themselves. This also inserts
1531 // destruction code. Add an alloca for each pool that is allocated to the
1532 // PoolDescs vector.
1534 void PoolAllocate::CreatePools(Function *F, const vector<AllocDSNode*> &Allocs,
1535 map<DSNode*, PoolInfo> &PoolDescs) {
1536 // Find all of the return nodes in the function...
1537 vector<BasicBlock*> ReturnNodes;
1538 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
1539 if (isa<ReturnInst>(I->getTerminator()))
1540 ReturnNodes.push_back(I);
1542 #ifdef DEBUG_CREATE_POOLS
1543 cerr << "Allocs that we are pool allocating:\n";
1544 for (unsigned i = 0, e = Allocs.size(); i != e; ++i)
1548 map<DSNode*, PATypeHolder> AbsPoolTyMap;
1550 // First pass over the allocations to process...
1551 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1552 // Create the pooldescriptor mapping... with null entries for everything
1553 // except the node & NewType fields.
1555 map<DSNode*, PoolInfo>::iterator PI =
1556 PoolDescs.insert(std::make_pair(Allocs[i], PoolInfo(Allocs[i]))).first;
1558 // Add a symbol table entry for the new type if there was one for the old
1560 string OldName = CurModule->getTypeName(Allocs[i]->getType());
1561 if (OldName.empty()) OldName = "node";
1562 CurModule->addTypeName(OldName+".p", PI->second.NewType);
1564 // Create the abstract pool types that will need to be resolved in a second
1565 // pass once an abstract type is created for each pool.
1567 // Can only handle limited shapes for now...
1568 const Type *OldNodeTy = Allocs[i]->getType();
1569 vector<const Type*> PoolTypes;
1571 // Pool type is the first element of the pool descriptor type...
1572 PoolTypes.push_back(getPoolType(PoolDescs[Allocs[i]].NewType));
1574 unsigned NumPointers = countPointerTypes(OldNodeTy);
1575 while (NumPointers--) // Add a different opaque type for each pointer
1576 PoolTypes.push_back(OpaqueType::get());
1578 assert(Allocs[i]->getNumLinks() == PoolTypes.size()-1 &&
1579 "Node should have same number of pointers as pool!");
1581 StructType *PoolType = StructType::get(PoolTypes);
1583 // Add a symbol table entry for the pooltype if possible...
1584 CurModule->addTypeName(OldName+".pool", PoolType);
1586 // Create the pool type, with opaque values for pointers...
1587 AbsPoolTyMap.insert(std::make_pair(Allocs[i], PoolType));
1588 #ifdef DEBUG_CREATE_POOLS
1589 cerr << "POOL TY: " << AbsPoolTyMap.find(Allocs[i])->second.get() << "\n";
1593 // Now that we have types for all of the pool types, link them all together.
1594 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1595 PATypeHolder &PoolTyH = AbsPoolTyMap.find(Allocs[i])->second;
1597 // Resolve all of the outgoing pointer types of this pool node...
1598 for (unsigned p = 0, pe = Allocs[i]->getNumLinks(); p != pe; ++p) {
1599 PointerValSet &PVS = Allocs[i]->getLink(p);
1600 assert(!PVS.empty() && "Outgoing edge is empty, field unused, can"
1601 " probably just leave the type opaque or something dumb.");
1603 for (Out = 0; AbsPoolTyMap.count(PVS[Out].Node) == 0; ++Out)
1604 assert(Out != PVS.size() && "No edge to an outgoing allocation node!?");
1606 assert(PVS[Out].Index == 0 && "Subindexing not implemented yet!");
1608 // The actual struct type could change each time through the loop, so it's
1609 // NOT loop invariant.
1610 const StructType *PoolTy = cast<StructType>(PoolTyH.get());
1612 // Get the opaque type...
1613 DerivedType *ElTy = (DerivedType*)(PoolTy->getElementTypes()[p+1].get());
1615 #ifdef DEBUG_CREATE_POOLS
1616 cerr << "Refining " << ElTy << " of " << PoolTy << " to "
1617 << AbsPoolTyMap.find(PVS[Out].Node)->second.get() << "\n";
1620 const Type *RefPoolTy = AbsPoolTyMap.find(PVS[Out].Node)->second.get();
1621 ElTy->refineAbstractTypeTo(PointerType::get(RefPoolTy));
1623 #ifdef DEBUG_CREATE_POOLS
1624 cerr << "Result pool type is: " << PoolTyH.get() << "\n";
1629 // Create the code that goes in the entry and exit nodes for the function...
1630 vector<Instruction*> EntryNodeInsts;
1631 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1632 PoolInfo &PI = PoolDescs[Allocs[i]];
1634 // Fill in the pool type for this pool...
1635 PI.PoolType = AbsPoolTyMap.find(Allocs[i])->second.get();
1636 assert(!PI.PoolType->isAbstract() &&
1637 "Pool type should not be abstract anymore!");
1639 // Add an allocation and a free for each pool...
1640 AllocaInst *PoolAlloc = new AllocaInst(PI.PoolType, 0,
1641 CurModule->getTypeName(PI.PoolType));
1642 PI.Handle = PoolAlloc;
1643 EntryNodeInsts.push_back(PoolAlloc);
1644 AllocationInst *AI = Allocs[i]->getAllocation();
1646 // Initialize the pool. We need to know how big each allocation is. For
1647 // our purposes here, we assume we are allocating a scalar, or array of
1650 unsigned ElSize = TargetData.getTypeSize(PI.NewType);
1652 vector<Value*> Args;
1653 Args.push_back(ConstantUInt::get(Type::UIntTy, ElSize));
1654 Args.push_back(PoolAlloc); // Pool to initialize
1655 EntryNodeInsts.push_back(new CallInst(PoolInit, Args));
1657 // Add code to destroy the pool in all of the exit nodes of the function...
1659 Args.push_back(PoolAlloc); // Pool to initialize
1661 for (unsigned EN = 0, ENE = ReturnNodes.size(); EN != ENE; ++EN) {
1662 Instruction *Destroy = new CallInst(PoolDestroy, Args);
1664 // Insert it before the return instruction...
1665 BasicBlock *RetNode = ReturnNodes[EN];
1666 RetNode->getInstList().insert(RetNode->end()--, Destroy);
1670 // Now that all of the pool descriptors have been created, link them together
1671 // so that called functions can get links as neccesary...
1673 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1674 PoolInfo &PI = PoolDescs[Allocs[i]];
1676 // For every pointer in the data structure, initialize a link that
1677 // indicates which pool to access...
1679 vector<Value*> Indices(2);
1680 Indices[0] = ConstantUInt::get(Type::UIntTy, 0);
1681 for (unsigned l = 0, le = PI.Node->getNumLinks(); l != le; ++l)
1682 // Only store an entry for the field if the field is used!
1683 if (!PI.Node->getLink(l).empty()) {
1684 assert(PI.Node->getLink(l).size() == 1 && "Should have only one link!");
1685 PointerVal PV = PI.Node->getLink(l)[0];
1686 assert(PV.Index == 0 && "Subindexing not supported yet!");
1687 PoolInfo &LinkedPool = PoolDescs[PV.Node];
1688 Indices[1] = ConstantUInt::get(Type::UByteTy, 1+l);
1690 EntryNodeInsts.push_back(new StoreInst(LinkedPool.Handle, PI.Handle,
1695 // Insert the entry node code into the entry block...
1696 F->getEntryNode().getInstList().insert(++F->getEntryNode().begin(),
1697 EntryNodeInsts.begin(),
1698 EntryNodeInsts.end());
1702 // addPoolPrototypes - Add prototypes for the pool functions to the specified
1703 // module and update the Pool* instance variables to point to them.
1705 void PoolAllocate::addPoolPrototypes(Module &M) {
1706 // Get poolinit function...
1707 vector<const Type*> Args;
1708 Args.push_back(Type::UIntTy); // Num bytes per element
1709 FunctionType *PoolInitTy = FunctionType::get(Type::VoidTy, Args, true);
1710 PoolInit = M.getOrInsertFunction("poolinit", PoolInitTy);
1712 // Get pooldestroy function...
1713 Args.pop_back(); // Only takes a pool...
1714 FunctionType *PoolDestroyTy = FunctionType::get(Type::VoidTy, Args, true);
1715 PoolDestroy = M.getOrInsertFunction("pooldestroy", PoolDestroyTy);
1717 // Get the poolalloc function...
1718 FunctionType *PoolAllocTy = FunctionType::get(POINTERTYPE, Args, true);
1719 PoolAlloc = M.getOrInsertFunction("poolalloc", PoolAllocTy);
1721 // Get the poolfree function...
1722 Args.push_back(POINTERTYPE); // Pointer to free
1723 FunctionType *PoolFreeTy = FunctionType::get(Type::VoidTy, Args, true);
1724 PoolFree = M.getOrInsertFunction("poolfree", PoolFreeTy);
1726 Args[0] = Type::UIntTy; // Number of slots to allocate
1727 FunctionType *PoolAllocArrayTy = FunctionType::get(POINTERTYPE, Args, true);
1728 PoolAllocArray = M.getOrInsertFunction("poolallocarray", PoolAllocArrayTy);
1732 bool PoolAllocate::run(Module &M) {
1733 addPoolPrototypes(M);
1736 DS = &getAnalysis<DataStructure>();
1737 bool Changed = false;
1739 for (Module::iterator I = M.begin(); I != M.end(); ++I)
1740 if (!I->isExternal()) {
1741 Changed |= processFunction(I);
1743 cerr << "Only processing one function\n";
1754 // createPoolAllocatePass - Global function to access the functionality of this
1757 Pass *createPoolAllocatePass() {
1758 assert(0 && "Pool allocator disabled!");
1760 //return new PoolAllocate();