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 // This pass requires a DCE & instcombine pass to be run after it for best
10 //===----------------------------------------------------------------------===//
12 #include "llvm/Transforms/IPO/PoolAllocate.h"
13 #include "llvm/Transforms/Utils/CloneFunction.h"
14 #include "llvm/Analysis/DataStructureGraph.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"
28 // DEBUG_CREATE_POOLS - Enable this to turn on debug output for the pool
29 // creation phase in the top level function of a transformed data structure.
31 //#define DEBUG_CREATE_POOLS 1
33 // DEBUG_TRANSFORM_PROGRESS - Enable this to get lots of debug output on what
34 // the transformation is doing.
36 //#define DEBUG_TRANSFORM_PROGRESS 1
38 // DEBUG_POOLBASE_LOAD_ELIMINATOR - Turn this on to get statistics about how
39 // many static loads were eliminated from a function...
41 #define DEBUG_POOLBASE_LOAD_ELIMINATOR 1
43 #include "Support/CommandLine.h"
45 Ptr8bits, Ptr16bits, Ptr32bits
48 static cl::Enum<enum PtrSize> ReqPointerSize("ptrsize", 0,
49 "Set pointer size for -poolalloc pass",
50 clEnumValN(Ptr32bits, "32", "Use 32 bit indices for pointers"),
51 clEnumValN(Ptr16bits, "16", "Use 16 bit indices for pointers"),
52 clEnumValN(Ptr8bits , "8", "Use 8 bit indices for pointers"), 0);
54 static cl::Flag DisableRLE("no-pool-load-elim", "Disable pool load elimination after poolalloc pass", cl::Hidden);
56 const Type *POINTERTYPE;
58 // FIXME: This is dependant on the sparc backend layout conventions!!
59 static TargetData TargetData("test");
61 static const Type *getPointerTransformedType(const Type *Ty) {
62 if (const PointerType *PT = dyn_cast<PointerType>(Ty)) {
64 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
65 vector<const Type *> NewElTypes;
66 NewElTypes.reserve(STy->getElementTypes().size());
67 for (StructType::ElementTypes::const_iterator
68 I = STy->getElementTypes().begin(),
69 E = STy->getElementTypes().end(); I != E; ++I)
70 NewElTypes.push_back(getPointerTransformedType(*I));
71 return StructType::get(NewElTypes);
72 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
73 return ArrayType::get(getPointerTransformedType(ATy->getElementType()),
74 ATy->getNumElements());
76 assert(Ty->isPrimitiveType() && "Unknown derived type!");
83 DSNode *Node; // The node this pool allocation represents
84 Value *Handle; // LLVM value of the pool in the current context
85 const Type *NewType; // The transformed type of the memory objects
86 const Type *PoolType; // The type of the pool
88 const Type *getOldType() const { return Node->getType(); }
90 PoolInfo() { // Define a default ctor for map::operator[]
91 cerr << "Map subscript used to get element that doesn't exist!\n";
95 PoolInfo(DSNode *N, Value *H, const Type *NT, const Type *PT)
96 : Node(N), Handle(H), NewType(NT), PoolType(PT) {
97 // Handle can be null...
98 assert(N && NT && PT && "Pool info null!");
101 PoolInfo(DSNode *N) : Node(N), Handle(0), NewType(0), PoolType(0) {
102 assert(N && "Invalid pool info!");
104 // The new type of the memory object is the same as the old type, except
105 // that all of the pointer values are replaced with POINTERTYPE values.
106 NewType = getPointerTransformedType(getOldType());
110 // ScalarInfo - Information about an LLVM value that we know points to some
111 // datastructure we are processing.
114 Value *Val; // Scalar value in Current Function
115 PoolInfo Pool; // The pool the scalar points into
117 ScalarInfo(Value *V, const PoolInfo &PI) : Val(V), Pool(PI) {
118 assert(V && "Null value passed to ScalarInfo ctor!");
122 // CallArgInfo - Information on one operand for a call that got expanded.
124 int ArgNo; // Call argument number this corresponds to
125 DSNode *Node; // The graph node for the pool
126 Value *PoolHandle; // The LLVM value that is the pool pointer
128 CallArgInfo(int Arg, DSNode *N, Value *PH)
129 : ArgNo(Arg), Node(N), PoolHandle(PH) {
130 assert(Arg >= -1 && N && PH && "Illegal values to CallArgInfo ctor!");
133 // operator< when sorting, sort by argument number.
134 bool operator<(const CallArgInfo &CAI) const {
135 return ArgNo < CAI.ArgNo;
139 // TransformFunctionInfo - Information about how a function eeds to be
142 struct TransformFunctionInfo {
143 // ArgInfo - Maintain information about the arguments that need to be
144 // processed. Each CallArgInfo corresponds to an argument that needs to
145 // have a pool pointer passed into the transformed function with it.
147 // As a special case, "argument" number -1 corresponds to the return value.
149 vector<CallArgInfo> ArgInfo;
151 // Func - The function to be transformed...
154 // The call instruction that is used to map CallArgInfo PoolHandle values
155 // into the new function values.
159 TransformFunctionInfo() : Func(0), Call(0) {}
161 bool operator<(const TransformFunctionInfo &TFI) const {
162 if (Func < TFI.Func) return true;
163 if (Func > TFI.Func) return false;
164 if (ArgInfo.size() < TFI.ArgInfo.size()) return true;
165 if (ArgInfo.size() > TFI.ArgInfo.size()) return false;
166 return ArgInfo < TFI.ArgInfo;
169 void finalizeConstruction() {
170 // Sort the vector so that the return value is first, followed by the
171 // argument records, in order. Note that this must be a stable sort so
172 // that the entries with the same sorting criteria (ie they are multiple
173 // pool entries for the same argument) are kept in depth first order.
174 stable_sort(ArgInfo.begin(), ArgInfo.end());
177 // addCallInfo - For a specified function call CI, figure out which pool
178 // descriptors need to be passed in as arguments, and which arguments need
179 // to be transformed into indices. If Arg != -1, the specified call
180 // argument is passed in as a pointer to a data structure.
182 void addCallInfo(DataStructure *DS, CallInst *CI, int Arg,
183 DSNode *GraphNode, map<DSNode*, PoolInfo> &PoolDescs);
185 // Make sure that all dependant arguments are added to this transformation
186 // info. For example, if we call foo(null, P) and foo treats it's first and
187 // second arguments as belonging to the same data structure, the we MUST add
188 // entries to know that the null needs to be transformed into an index as
191 void ensureDependantArgumentsIncluded(DataStructure *DS,
192 map<DSNode*, PoolInfo> &PoolDescs);
196 // Define the pass class that we implement...
197 struct PoolAllocate : public Pass {
198 const char *getPassName() const { return "Pool Allocate"; }
201 switch (ReqPointerSize) {
202 case Ptr32bits: POINTERTYPE = Type::UIntTy; break;
203 case Ptr16bits: POINTERTYPE = Type::UShortTy; break;
204 case Ptr8bits: POINTERTYPE = Type::UByteTy; break;
207 CurModule = 0; DS = 0;
208 PoolInit = PoolDestroy = PoolAlloc = PoolFree = 0;
211 // getPoolType - Get the type used by the backend for a pool of a particular
212 // type. This pool record is used to allocate nodes of type NodeType.
214 // Here, PoolTy = { NodeType*, sbyte*, uint }*
216 const StructType *getPoolType(const Type *NodeType) {
217 vector<const Type*> PoolElements;
218 PoolElements.push_back(PointerType::get(NodeType));
219 PoolElements.push_back(PointerType::get(Type::SByteTy));
220 PoolElements.push_back(Type::UIntTy);
221 StructType *Result = StructType::get(PoolElements);
223 // Add a name to the symbol table to correspond to the backend
224 // representation of this pool...
225 assert(CurModule && "No current module!?");
226 string Name = CurModule->getTypeName(NodeType);
227 if (Name.empty()) Name = CurModule->getTypeName(PoolElements[0]);
228 CurModule->addTypeName(Name+"oolbe", Result);
235 // getAnalysisUsage - This function requires data structure information
236 // to be able to see what is pool allocatable.
238 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
239 AU.addRequired(DataStructure::ID);
243 // CurModule - The module being processed.
246 // DS - The data structure graph for the module being processed.
249 // Prototypes that we add to support pool allocation...
250 Function *PoolInit, *PoolDestroy, *PoolAlloc, *PoolAllocArray, *PoolFree;
252 // The map of already transformed functions... note that the keys of this
253 // map do not have meaningful values for 'Call' or the 'PoolHandle' elements
254 // of the ArgInfo elements.
256 map<TransformFunctionInfo, Function*> TransformedFunctions;
258 // getTransformedFunction - Get a transformed function, or return null if
259 // the function specified hasn't been transformed yet.
261 Function *getTransformedFunction(TransformFunctionInfo &TFI) const {
262 map<TransformFunctionInfo, Function*>::const_iterator I =
263 TransformedFunctions.find(TFI);
264 if (I != TransformedFunctions.end()) return I->second;
269 // addPoolPrototypes - Add prototypes for the pool functions to the
270 // specified module and update the Pool* instance variables to point to
273 void addPoolPrototypes(Module &M);
276 // CreatePools - Insert instructions into the function we are processing to
277 // create all of the memory pool objects themselves. This also inserts
278 // destruction code. Add an alloca for each pool that is allocated to the
281 void CreatePools(Function *F, const vector<AllocDSNode*> &Allocs,
282 map<DSNode*, PoolInfo> &PoolDescs);
284 // processFunction - Convert a function to use pool allocation where
287 bool processFunction(Function *F);
289 // transformFunctionBody - This transforms the instruction in 'F' to use the
290 // pools specified in PoolDescs when modifying data structure nodes
291 // specified in the PoolDescs map. IPFGraph is the closed data structure
292 // graph for F, of which the PoolDescriptor nodes come from.
294 void transformFunctionBody(Function *F, FunctionDSGraph &IPFGraph,
295 map<DSNode*, PoolInfo> &PoolDescs);
297 // transformFunction - Transform the specified function the specified way.
298 // It we have already transformed that function that way, don't do anything.
299 // The nodes in the TransformFunctionInfo come out of callers data structure
300 // graph, and the PoolDescs passed in are the caller's.
302 void transformFunction(TransformFunctionInfo &TFI,
303 FunctionDSGraph &CallerIPGraph,
304 map<DSNode*, PoolInfo> &PoolDescs);
309 // isNotPoolableAlloc - This is a predicate that returns true if the specified
310 // allocation node in a data structure graph is eligable for pool allocation.
312 static bool isNotPoolableAlloc(const AllocDSNode *DS) {
313 if (DS->isAllocaNode()) return true; // Do not pool allocate alloca's.
317 // processFunction - Convert a function to use pool allocation where
320 bool PoolAllocate::processFunction(Function *F) {
321 // Get the closed datastructure graph for the current function... if there are
322 // any allocations in this graph that are not escaping, we need to pool
323 // allocate them here!
325 FunctionDSGraph &IPGraph = DS->getClosedDSGraph(F);
327 // Get all of the allocations that do not escape the current function. Since
328 // they are still live (they exist in the graph at all), this means we must
329 // have scalar references to these nodes, but the scalars are never returned.
331 vector<AllocDSNode*> Allocs;
332 IPGraph.getNonEscapingAllocations(Allocs);
334 // Filter out allocations that we cannot handle. Currently, this includes
335 // variable sized array allocations and alloca's (which we do not want to
338 Allocs.erase(remove_if(Allocs.begin(), Allocs.end(), isNotPoolableAlloc),
342 if (Allocs.empty()) return false; // Nothing to do.
344 #ifdef DEBUG_TRANSFORM_PROGRESS
345 cerr << "Transforming Function: " << F->getName() << "\n";
348 // Insert instructions into the function we are processing to create all of
349 // the memory pool objects themselves. This also inserts destruction code.
350 // This fills in the PoolDescs map to associate the alloc node with the
351 // allocation of the memory pool corresponding to it.
353 map<DSNode*, PoolInfo> PoolDescs;
354 CreatePools(F, Allocs, PoolDescs);
356 #ifdef DEBUG_TRANSFORM_PROGRESS
357 cerr << "Transformed Entry Function: \n" << F;
360 // Now we need to figure out what called functions we need to transform, and
361 // how. To do this, we look at all of the scalars, seeing which functions are
362 // either used as a scalar value (so they return a data structure), or are
363 // passed one of our scalar values.
365 transformFunctionBody(F, IPGraph, PoolDescs);
371 //===----------------------------------------------------------------------===//
373 // NewInstructionCreator - This class is used to traverse the function being
374 // modified, changing each instruction visit'ed to use and provide pointer
375 // indexes instead of real pointers. This is what changes the body of a
376 // function to use pool allocation.
378 class NewInstructionCreator : public InstVisitor<NewInstructionCreator> {
379 PoolAllocate &PoolAllocator;
380 vector<ScalarInfo> &Scalars;
381 map<CallInst*, TransformFunctionInfo> &CallMap;
382 map<Value*, Value*> &XFormMap; // Map old pointers to new indexes
385 Instruction *I; // Instruction to update
386 unsigned OpNum; // Operand number to update
387 Value *OldVal; // The old value it had
389 RefToUpdate(Instruction *i, unsigned o, Value *ov)
390 : I(i), OpNum(o), OldVal(ov) {}
392 vector<RefToUpdate> ReferencesToUpdate;
394 const ScalarInfo &getScalarRef(const Value *V) {
395 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
396 if (Scalars[i].Val == V) return Scalars[i];
398 cerr << "Could not find scalar " << V << " in scalar map!\n";
399 assert(0 && "Scalar not found in getScalar!");
404 const ScalarInfo *getScalar(const Value *V) {
405 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
406 if (Scalars[i].Val == V) return &Scalars[i];
410 BasicBlock::iterator ReplaceInstWith(Instruction &I, Instruction *New) {
411 BasicBlock *BB = I.getParent();
412 BasicBlock::iterator RI = &I;
413 BB->getInstList().remove(RI);
414 BB->getInstList().insert(RI, New);
419 Instruction *createPoolBaseInstruction(Value *PtrVal) {
420 const ScalarInfo &SC = getScalarRef(PtrVal);
421 vector<Value*> Args(3);
422 Args[0] = ConstantUInt::get(Type::UIntTy, 0); // No pointer offset
423 Args[1] = ConstantUInt::get(Type::UByteTy, 0); // Field #0 of pool descriptr
424 Args[2] = ConstantUInt::get(Type::UByteTy, 0); // Field #0 of poolalloc val
425 return new LoadInst(SC.Pool.Handle, Args, PtrVal->getName()+".poolbase");
430 NewInstructionCreator(PoolAllocate &PA, vector<ScalarInfo> &S,
431 map<CallInst*, TransformFunctionInfo> &C,
432 map<Value*, Value*> &X)
433 : PoolAllocator(PA), Scalars(S), CallMap(C), XFormMap(X) {}
436 // updateReferences - The NewInstructionCreator is responsible for creating
437 // new instructions to replace the old ones in the function, and then link up
438 // references to values to their new values. For it to do this, however, it
439 // keeps track of information about the value mapping of old values to new
440 // values that need to be patched up. Given this value map and a set of
441 // instruction operands to patch, updateReferences performs the updates.
443 void updateReferences() {
444 for (unsigned i = 0, e = ReferencesToUpdate.size(); i != e; ++i) {
445 RefToUpdate &Ref = ReferencesToUpdate[i];
446 Value *NewVal = XFormMap[Ref.OldVal];
449 if (isa<Constant>(Ref.OldVal) && // Refering to a null ptr?
450 cast<Constant>(Ref.OldVal)->isNullValue()) {
451 // Transform the null pointer into a null index... caching in XFormMap
452 XFormMap[Ref.OldVal] = NewVal = Constant::getNullValue(POINTERTYPE);
453 //} else if (isa<Argument>(Ref.OldVal)) {
455 cerr << "Unknown reference to: " << Ref.OldVal << "\n";
456 assert(XFormMap[Ref.OldVal] &&
457 "Reference to value that was not updated found!");
461 Ref.I->setOperand(Ref.OpNum, NewVal);
463 ReferencesToUpdate.clear();
466 //===--------------------------------------------------------------------===//
467 // Transformation methods:
468 // These methods specify how each type of instruction is transformed by the
469 // NewInstructionCreator instance...
470 //===--------------------------------------------------------------------===//
472 void visitGetElementPtrInst(GetElementPtrInst &I) {
473 assert(0 && "Cannot transform get element ptr instructions yet!");
476 // Replace the load instruction with a new one.
477 void visitLoadInst(LoadInst &I) {
478 vector<Instruction *> BeforeInsts;
480 // Cast our index to be a UIntTy so we can use it to index into the pool...
481 CastInst *Index = new CastInst(Constant::getNullValue(POINTERTYPE),
482 Type::UIntTy, I.getOperand(0)->getName());
483 BeforeInsts.push_back(Index);
484 ReferencesToUpdate.push_back(RefToUpdate(Index, 0, I.getOperand(0)));
486 // Include the pool base instruction...
487 Instruction *PoolBase = createPoolBaseInstruction(I.getOperand(0));
488 BeforeInsts.push_back(PoolBase);
490 Instruction *IdxInst =
491 BinaryOperator::create(Instruction::Add, *I.idx_begin(), Index,
493 BeforeInsts.push_back(IdxInst);
495 vector<Value*> Indices(I.idx_begin(), I.idx_end());
496 Indices[0] = IdxInst;
497 Instruction *Address = new GetElementPtrInst(PoolBase, Indices,
498 I.getName()+".addr");
499 BeforeInsts.push_back(Address);
501 Instruction *NewLoad = new LoadInst(Address, I.getName());
503 // Replace the load instruction with the new load instruction...
504 BasicBlock::iterator II = ReplaceInstWith(I, NewLoad);
506 // Add all of the instructions before the load...
507 NewLoad->getParent()->getInstList().insert(II, BeforeInsts.begin(),
510 // If not yielding a pool allocated pointer, use the new load value as the
511 // value in the program instead of the old load value...
514 I.replaceAllUsesWith(NewLoad);
517 // Replace the store instruction with a new one. In the store instruction,
518 // the value stored could be a pointer type, meaning that the new store may
519 // have to change one or both of it's operands.
521 void visitStoreInst(StoreInst &I) {
522 assert(getScalar(I.getOperand(1)) &&
523 "Store inst found only storing pool allocated pointer. "
526 Value *Val = I.getOperand(0); // The value to store...
528 // Check to see if the value we are storing is a data structure pointer...
529 //if (const ScalarInfo *ValScalar = getScalar(I.getOperand(0)))
530 if (isa<PointerType>(I.getOperand(0)->getType()))
531 Val = Constant::getNullValue(POINTERTYPE); // Yes, store a dummy
533 Instruction *PoolBase = createPoolBaseInstruction(I.getOperand(1));
535 // Cast our index to be a UIntTy so we can use it to index into the pool...
536 CastInst *Index = new CastInst(Constant::getNullValue(POINTERTYPE),
537 Type::UIntTy, I.getOperand(1)->getName());
538 ReferencesToUpdate.push_back(RefToUpdate(Index, 0, I.getOperand(1)));
540 // Instructions to add after the Index...
541 vector<Instruction*> AfterInsts;
543 Instruction *IdxInst =
544 BinaryOperator::create(Instruction::Add, *I.idx_begin(), Index, "idx");
545 AfterInsts.push_back(IdxInst);
547 vector<Value*> Indices(I.idx_begin(), I.idx_end());
548 Indices[0] = IdxInst;
549 Instruction *Address = new GetElementPtrInst(PoolBase, Indices,
550 I.getName()+"storeaddr");
551 AfterInsts.push_back(Address);
553 Instruction *NewStore = new StoreInst(Val, Address);
554 AfterInsts.push_back(NewStore);
555 if (Val != I.getOperand(0)) // Value stored was a pointer?
556 ReferencesToUpdate.push_back(RefToUpdate(NewStore, 0, I.getOperand(0)));
559 // Replace the store instruction with the cast instruction...
560 BasicBlock::iterator II = ReplaceInstWith(I, Index);
562 // Add the pool base calculator instruction before the index...
563 II = ++Index->getParent()->getInstList().insert(II, PoolBase);
566 // Add the instructions that go after the index...
567 Index->getParent()->getInstList().insert(II, AfterInsts.begin(),
572 // Create call to poolalloc for every malloc instruction
573 void visitMallocInst(MallocInst &I) {
574 const ScalarInfo &SCI = getScalarRef(&I);
578 if (!I.isArrayAllocation()) {
579 Args.push_back(SCI.Pool.Handle);
580 Call = new CallInst(PoolAllocator.PoolAlloc, Args, I.getName());
582 Args.push_back(I.getArraySize());
583 Args.push_back(SCI.Pool.Handle);
584 Call = new CallInst(PoolAllocator.PoolAllocArray, Args, I.getName());
587 ReplaceInstWith(I, Call);
590 // Convert a call to poolfree for every free instruction...
591 void visitFreeInst(FreeInst &I) {
592 // Create a new call to poolfree before the free instruction
594 Args.push_back(Constant::getNullValue(POINTERTYPE));
595 Args.push_back(getScalarRef(I.getOperand(0)).Pool.Handle);
596 Instruction *NewCall = new CallInst(PoolAllocator.PoolFree, Args);
597 ReplaceInstWith(I, NewCall);
598 ReferencesToUpdate.push_back(RefToUpdate(NewCall, 1, I.getOperand(0)));
601 // visitCallInst - Create a new call instruction with the extra arguments for
602 // all of the memory pools that the call needs.
604 void visitCallInst(CallInst &I) {
605 TransformFunctionInfo &TI = CallMap[&I];
607 // Start with all of the old arguments...
608 vector<Value*> Args(I.op_begin()+1, I.op_end());
610 for (unsigned i = 0, e = TI.ArgInfo.size(); i != e; ++i) {
611 // Replace all of the pointer arguments with our new pointer typed values.
612 if (TI.ArgInfo[i].ArgNo != -1)
613 Args[TI.ArgInfo[i].ArgNo] = Constant::getNullValue(POINTERTYPE);
615 // Add all of the pool arguments...
616 Args.push_back(TI.ArgInfo[i].PoolHandle);
619 Function *NF = PoolAllocator.getTransformedFunction(TI);
620 Instruction *NewCall = new CallInst(NF, Args, I.getName());
621 ReplaceInstWith(I, NewCall);
623 // Keep track of the mapping of operands so that we can resolve them to real
625 Value *RetVal = NewCall;
626 for (unsigned i = 0, e = TI.ArgInfo.size(); i != e; ++i)
627 if (TI.ArgInfo[i].ArgNo != -1)
628 ReferencesToUpdate.push_back(RefToUpdate(NewCall, TI.ArgInfo[i].ArgNo+1,
629 I.getOperand(TI.ArgInfo[i].ArgNo+1)));
631 RetVal = 0; // If returning a pointer, don't change retval...
633 // If not returning a pointer, use the new call as the value in the program
634 // instead of the old call...
637 I.replaceAllUsesWith(RetVal);
640 // visitPHINode - Create a new PHI node of POINTERTYPE for all of the old Phi
643 void visitPHINode(PHINode &PN) {
644 Value *DummyVal = Constant::getNullValue(POINTERTYPE);
645 PHINode *NewPhi = new PHINode(POINTERTYPE, PN.getName());
646 for (unsigned i = 0, e = PN.getNumIncomingValues(); i != e; ++i) {
647 NewPhi->addIncoming(DummyVal, PN.getIncomingBlock(i));
648 ReferencesToUpdate.push_back(RefToUpdate(NewPhi, i*2,
649 PN.getIncomingValue(i)));
652 ReplaceInstWith(PN, NewPhi);
655 // visitReturnInst - Replace ret instruction with a new return...
656 void visitReturnInst(ReturnInst &I) {
657 Instruction *Ret = new ReturnInst(Constant::getNullValue(POINTERTYPE));
658 ReplaceInstWith(I, Ret);
659 ReferencesToUpdate.push_back(RefToUpdate(Ret, 0, I.getOperand(0)));
662 // visitSetCondInst - Replace a conditional test instruction with a new one
663 void visitSetCondInst(SetCondInst &SCI) {
664 BinaryOperator &I = (BinaryOperator&)SCI;
665 Value *DummyVal = Constant::getNullValue(POINTERTYPE);
666 BinaryOperator *New = BinaryOperator::create(I.getOpcode(), DummyVal,
667 DummyVal, I.getName());
668 ReplaceInstWith(I, New);
670 ReferencesToUpdate.push_back(RefToUpdate(New, 0, I.getOperand(0)));
671 ReferencesToUpdate.push_back(RefToUpdate(New, 1, I.getOperand(1)));
673 // Make sure branches refer to the new condition...
674 I.replaceAllUsesWith(New);
677 void visitInstruction(Instruction &I) {
678 cerr << "Unknown instruction to FunctionBodyTransformer:\n" << I;
683 // PoolBaseLoadEliminator - Every load and store through a pool allocated
684 // pointer causes a load of the real pool base out of the pool descriptor.
685 // Iterate through the function, doing a local elimination pass of duplicate
686 // loads. This attempts to turn the all too common:
688 // %reg109.poolbase22 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
689 // %reg207 = load %root.p* %reg109.poolbase22, uint %reg109, ubyte 0, ubyte 0
690 // %reg109.poolbase23 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
691 // store double %reg207, %root.p* %reg109.poolbase23, uint %reg109, ...
694 // %reg109.poolbase22 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
695 // %reg207 = load %root.p* %reg109.poolbase22, uint %reg109, ubyte 0, ubyte 0
696 // store double %reg207, %root.p* %reg109.poolbase22, uint %reg109, ...
699 class PoolBaseLoadEliminator : public InstVisitor<PoolBaseLoadEliminator> {
700 // PoolDescValues - Keep track of the values in the current function that are
701 // pool descriptors (loads from which we want to eliminate).
703 vector<Value*> PoolDescValues;
705 // PoolDescMap - As we are analyzing a BB, keep track of which load to use
706 // when referencing a pool descriptor.
708 map<Value*, LoadInst*> PoolDescMap;
710 // These two fields keep track of statistics of how effective we are, if
711 // debugging is enabled.
713 unsigned Eliminated, Remaining;
715 // Compact the pool descriptor map into a list of the pool descriptors in the
716 // current context that we should know about...
718 PoolBaseLoadEliminator(const map<DSNode*, PoolInfo> &PoolDescs) {
719 Eliminated = Remaining = 0;
720 for (map<DSNode*, PoolInfo>::const_iterator I = PoolDescs.begin(),
721 E = PoolDescs.end(); I != E; ++I)
722 PoolDescValues.push_back(I->second.Handle);
724 // Remove duplicates from the list of pool values
725 sort(PoolDescValues.begin(), PoolDescValues.end());
726 PoolDescValues.erase(unique(PoolDescValues.begin(), PoolDescValues.end()),
727 PoolDescValues.end());
730 #ifdef DEBUG_POOLBASE_LOAD_ELIMINATOR
731 void visitFunction(Function &F) {
732 cerr << "Pool Load Elim '" << F.getName() << "'\t";
734 ~PoolBaseLoadEliminator() {
735 unsigned Total = Eliminated+Remaining;
737 cerr << "removed " << Eliminated << "["
738 << Eliminated*100/Total << "%] loads, leaving "
739 << Remaining << ".\n";
743 // Loop over the function, looking for loads to eliminate. Because we are a
744 // local transformation, we reset all of our state when we enter a new basic
747 void visitBasicBlock(BasicBlock &) {
748 PoolDescMap.clear(); // Forget state.
751 // Starting with an empty basic block, we scan it looking for loads of the
752 // pool descriptor. When we find a load, we add it to the PoolDescMap,
753 // indicating that we have a value available to recycle next time we see the
754 // poolbase of this instruction being loaded.
756 void visitLoadInst(LoadInst &LI) {
757 Value *LoadAddr = LI.getPointerOperand();
758 map<Value*, LoadInst*>::iterator VIt = PoolDescMap.find(LoadAddr);
759 if (VIt != PoolDescMap.end()) { // We already have a value for this load?
760 LI.replaceAllUsesWith(VIt->second); // Make the current load dead
763 // This load might not be a load of a pool pointer, check to see if it is
764 if (LI.getNumOperands() == 4 && // load pool, uint 0, ubyte 0, ubyte 0
765 find(PoolDescValues.begin(), PoolDescValues.end(), LoadAddr) !=
766 PoolDescValues.end()) {
768 assert("Make sure it's a load of the pool base, not a chaining field" &&
769 LI.getOperand(1) == Constant::getNullValue(Type::UIntTy) &&
770 LI.getOperand(2) == Constant::getNullValue(Type::UByteTy) &&
771 LI.getOperand(3) == Constant::getNullValue(Type::UByteTy));
773 // If it is a load of a pool base, keep track of it for future reference
774 PoolDescMap.insert(make_pair(LoadAddr, &LI));
780 // If we run across a function call, forget all state... Calls to
781 // poolalloc/poolfree can invalidate the pool base pointer, so it should be
782 // reloaded the next time it is used. Furthermore, a call to a random
783 // function might call one of these functions, so be conservative. Through
784 // more analysis, this could be improved in the future.
786 void visitCallInst(CallInst &) {
791 static void addNodeMapping(DSNode *SrcNode, const PointerValSet &PVS,
792 map<DSNode*, PointerValSet> &NodeMapping) {
793 for (unsigned i = 0, e = PVS.size(); i != e; ++i)
794 if (NodeMapping[SrcNode].add(PVS[i])) { // Not in map yet?
795 assert(PVS[i].Index == 0 && "Node indexing not supported yet!");
796 DSNode *DestNode = PVS[i].Node;
798 // Loop over all of the outgoing links in the mapped graph
799 for (unsigned l = 0, le = DestNode->getNumOutgoingLinks(); l != le; ++l) {
800 PointerValSet &SrcSet = SrcNode->getOutgoingLink(l);
801 const PointerValSet &DestSet = DestNode->getOutgoingLink(l);
803 // Add all of the node mappings now!
804 for (unsigned si = 0, se = SrcSet.size(); si != se; ++si) {
805 assert(SrcSet[si].Index == 0 && "Can't handle node offset!");
806 addNodeMapping(SrcSet[si].Node, DestSet, NodeMapping);
812 // CalculateNodeMapping - There is a partial isomorphism between the graph
813 // passed in and the graph that is actually used by the function. We need to
814 // figure out what this mapping is so that we can transformFunctionBody the
815 // instructions in the function itself. Note that every node in the graph that
816 // we are interested in must be both in the local graph of the called function,
817 // and in the local graph of the calling function. Because of this, we only
818 // define the mapping for these nodes [conveniently these are the only nodes we
819 // CAN define a mapping for...]
821 // The roots of the graph that we are transforming is rooted in the arguments
822 // passed into the function from the caller. This is where we start our
823 // mapping calculation.
825 // The NodeMapping calculated maps from the callers graph to the called graph.
827 static void CalculateNodeMapping(Function *F, TransformFunctionInfo &TFI,
828 FunctionDSGraph &CallerGraph,
829 FunctionDSGraph &CalledGraph,
830 map<DSNode*, PointerValSet> &NodeMapping) {
832 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
833 // Figure out what nodes in the called graph the TFI.ArgInfo[i].Node node
836 // Only consider first node of sequence. Extra nodes may may be added
837 // to the TFI if the data structure requires more nodes than just the
838 // one the argument points to. We are only interested in the one the
839 // argument points to though.
841 if (TFI.ArgInfo[i].ArgNo != LastArgNo) {
842 if (TFI.ArgInfo[i].ArgNo == -1) {
843 addNodeMapping(TFI.ArgInfo[i].Node, CalledGraph.getRetNodes(),
846 // Figure out which node argument # ArgNo points to in the called graph.
847 Function::aiterator AI = F->abegin();
848 std::advance(AI, TFI.ArgInfo[i].ArgNo);
849 addNodeMapping(TFI.ArgInfo[i].Node, CalledGraph.getValueMap()[AI],
852 LastArgNo = TFI.ArgInfo[i].ArgNo;
860 // addCallInfo - For a specified function call CI, figure out which pool
861 // descriptors need to be passed in as arguments, and which arguments need to be
862 // transformed into indices. If Arg != -1, the specified call argument is
863 // passed in as a pointer to a data structure.
865 void TransformFunctionInfo::addCallInfo(DataStructure *DS, CallInst *CI,
866 int Arg, DSNode *GraphNode,
867 map<DSNode*, PoolInfo> &PoolDescs) {
868 assert(CI->getCalledFunction() && "Cannot handle indirect calls yet!");
869 assert(Func == 0 || Func == CI->getCalledFunction() &&
870 "Function call record should always call the same function!");
871 assert(Call == 0 || Call == CI &&
872 "Call element already filled in with different value!");
873 Func = CI->getCalledFunction();
875 //FunctionDSGraph &CalledGraph = DS->getClosedDSGraph(Func);
877 // For now, add the entire graph that is pointed to by the call argument.
878 // This graph can and should be pruned to only what the function itself will
879 // use, because often this will be a dramatically smaller subset of what we
882 // FIXME: This should use pool links instead of extra arguments!
884 for (df_iterator<DSNode*> I = df_begin(GraphNode), E = df_end(GraphNode);
886 ArgInfo.push_back(CallArgInfo(Arg, *I, PoolDescs[*I].Handle));
889 static void markReachableNodes(const PointerValSet &Vals,
890 set<DSNode*> &ReachableNodes) {
891 for (unsigned n = 0, ne = Vals.size(); n != ne; ++n) {
892 DSNode *N = Vals[n].Node;
893 if (ReachableNodes.count(N) == 0) // Haven't already processed node?
894 ReachableNodes.insert(df_begin(N), df_end(N)); // Insert all
898 // Make sure that all dependant arguments are added to this transformation info.
899 // For example, if we call foo(null, P) and foo treats it's first and second
900 // arguments as belonging to the same data structure, the we MUST add entries to
901 // know that the null needs to be transformed into an index as well.
903 void TransformFunctionInfo::ensureDependantArgumentsIncluded(DataStructure *DS,
904 map<DSNode*, PoolInfo> &PoolDescs) {
905 // FIXME: This does not work for indirect function calls!!!
906 if (Func == 0) return; // FIXME!
908 // Make sure argument entries are sorted.
909 finalizeConstruction();
911 // Loop over the function signature, checking to see if there are any pointer
912 // arguments that we do not convert... if there is something we haven't
913 // converted, set done to false.
917 if (isa<PointerType>(Func->getReturnType())) // Make sure we convert retval
918 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == -1) {
919 // We DO transform the ret val... skip all possible entries for retval
920 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == -1)
927 for (Function::aiterator I = Func->abegin(), E = Func->aend(); I!=E; ++I,++i){
928 if (isa<PointerType>(I->getType())) {
929 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == (int)i) {
930 // We DO transform this arg... skip all possible entries for argument
931 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == (int)i)
940 // If we already have entries for all pointer arguments and retvals, there
941 // certainly is no work to do. Bail out early to avoid building relatively
942 // expensive data structures.
946 #ifdef DEBUG_TRANSFORM_PROGRESS
947 cerr << "Must ensure dependant arguments for: " << Func->getName() << "\n";
950 // Otherwise, we MIGHT have to add the arguments/retval if they are part of
951 // the same datastructure graph as some other argument or retval that we ARE
954 // Get the data structure graph for the called function.
956 FunctionDSGraph &CalledDS = DS->getClosedDSGraph(Func);
958 // Build a mapping between the nodes in our current graph and the nodes in the
959 // called function's graph. We build it based on our _incomplete_
960 // transformation information, because it contains all of the info that we
963 map<DSNode*, PointerValSet> NodeMapping;
964 CalculateNodeMapping(Func, *this,
965 DS->getClosedDSGraph(Call->getParent()->getParent()),
966 CalledDS, NodeMapping);
968 // Build the inverted version of the node mapping, that maps from a node in
969 // the called functions graph to a single node in the caller graph.
971 map<DSNode*, DSNode*> InverseNodeMap;
972 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin(),
973 E = NodeMapping.end(); I != E; ++I) {
974 PointerValSet &CalledNodes = I->second;
975 for (unsigned i = 0, e = CalledNodes.size(); i != e; ++i)
976 InverseNodeMap[CalledNodes[i].Node] = I->first;
978 NodeMapping.clear(); // Done with information, free memory
980 // Build a set of reachable nodes from the arguments/retval that we ARE
982 set<DSNode*> ReachableNodes;
984 // Loop through all of the arguments, marking all of the reachable data
985 // structure nodes reachable if they are from this pointer...
987 for (unsigned i = 0, e = ArgInfo.size(); i != e; ++i) {
988 if (ArgInfo[i].ArgNo == -1) {
989 if (i == 0) // Only process retvals once (performance opt)
990 markReachableNodes(CalledDS.getRetNodes(), ReachableNodes);
991 } else { // If it's an argument value...
992 Function::aiterator AI = Func->abegin();
993 std::advance(AI, ArgInfo[i].ArgNo);
994 if (isa<PointerType>(AI->getType()))
995 markReachableNodes(CalledDS.getValueMap()[AI], ReachableNodes);
999 // Now that we know which nodes are already reachable, see if any of the
1000 // arguments that we are not passing values in for can reach one of the
1001 // existing nodes...
1004 // <FIXME> IN THEORY, we should allow arbitrary paths from the argument to
1005 // nodes we know about. The problem is that if we do this, then I don't know
1006 // how to get pool pointers for this head list. Since we are completely
1007 // deadline driven, I'll just allow direct accesses to the graph. </FIXME>
1011 if (isa<PointerType>(Func->getReturnType())) // Make sure we convert retval
1012 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == -1) {
1013 // We DO transform the ret val... skip all possible entries for retval
1014 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == -1)
1017 // See what the return value points to...
1019 // FIXME: This should generalize to any number of nodes, just see if any
1021 assert(CalledDS.getRetNodes().size() == 1 &&
1022 "Assumes only one node is returned");
1023 DSNode *N = CalledDS.getRetNodes()[0].Node;
1025 // If the return value is not marked as being passed in, but it NEEDS to
1026 // be transformed, then make it known now.
1028 if (ReachableNodes.count(N)) {
1029 #ifdef DEBUG_TRANSFORM_PROGRESS
1030 cerr << "ensure dependant arguments adds return value entry!\n";
1032 addCallInfo(DS, Call, -1, InverseNodeMap[N], PoolDescs);
1035 finalizeConstruction();
1040 for (Function::aiterator I = Func->abegin(), E = Func->aend(); I!=E; ++I, ++i)
1041 if (isa<PointerType>(I->getType())) {
1042 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == (int)i) {
1043 // We DO transform this arg... skip all possible entries for argument
1044 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == (int)i)
1047 // This should generalize to any number of nodes, just see if any are
1049 assert(CalledDS.getValueMap()[I].size() == 1 &&
1050 "Only handle case where pointing to one node so far!");
1052 // If the arg is not marked as being passed in, but it NEEDS to
1053 // be transformed, then make it known now.
1055 DSNode *N = CalledDS.getValueMap()[I][0].Node;
1056 if (ReachableNodes.count(N)) {
1057 #ifdef DEBUG_TRANSFORM_PROGRESS
1058 cerr << "ensure dependant arguments adds for arg #" << i << "\n";
1060 addCallInfo(DS, Call, i, InverseNodeMap[N], PoolDescs);
1063 finalizeConstruction();
1070 // transformFunctionBody - This transforms the instruction in 'F' to use the
1071 // pools specified in PoolDescs when modifying data structure nodes specified in
1072 // the PoolDescs map. Specifically, scalar values specified in the Scalars
1073 // vector must be remapped. IPFGraph is the closed data structure graph for F,
1074 // of which the PoolDescriptor nodes come from.
1076 void PoolAllocate::transformFunctionBody(Function *F, FunctionDSGraph &IPFGraph,
1077 map<DSNode*, PoolInfo> &PoolDescs) {
1079 // Loop through the value map looking for scalars that refer to nonescaping
1080 // allocations. Add them to the Scalars vector. Note that we may have
1081 // multiple entries in the Scalars vector for each value if it points to more
1084 map<Value*, PointerValSet> &ValMap = IPFGraph.getValueMap();
1085 vector<ScalarInfo> Scalars;
1087 #ifdef DEBUG_TRANSFORM_PROGRESS
1088 cerr << "Building scalar map for fn '" << F->getName() << "' body:\n";
1091 for (map<Value*, PointerValSet>::iterator I = ValMap.begin(),
1092 E = ValMap.end(); I != E; ++I) {
1093 const PointerValSet &PVS = I->second; // Set of things pointed to by scalar
1095 // Check to see if the scalar points to a data structure node...
1096 for (unsigned i = 0, e = PVS.size(); i != e; ++i) {
1097 if (PVS[i].Index) { cerr << "Problem in " << F->getName() << " for " << I->first << "\n"; }
1098 assert(PVS[i].Index == 0 && "Nonzero not handled yet!");
1100 // If the allocation is in the nonescaping set...
1101 map<DSNode*, PoolInfo>::iterator AI = PoolDescs.find(PVS[i].Node);
1102 if (AI != PoolDescs.end()) { // Add it to the list of scalars
1103 Scalars.push_back(ScalarInfo(I->first, AI->second));
1104 #ifdef DEBUG_TRANSFORM_PROGRESS
1105 cerr << "\nScalar Mapping from:" << I->first
1106 << "Scalar Mapping to: "; PVS.print(cerr);
1112 #ifdef DEBUG_TRANSFORM_PROGRESS
1113 cerr << "\nIn '" << F->getName()
1114 << "': Found the following values that point to poolable nodes:\n";
1116 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
1117 cerr << Scalars[i].Val;
1121 // CallMap - Contain an entry for every call instruction that needs to be
1122 // transformed. Each entry in the map contains information about what we need
1123 // to do to each call site to change it to work.
1125 map<CallInst*, TransformFunctionInfo> CallMap;
1127 // Now we need to figure out what called functions we need to transform, and
1128 // how. To do this, we look at all of the scalars, seeing which functions are
1129 // either used as a scalar value (so they return a data structure), or are
1130 // passed one of our scalar values.
1132 for (unsigned i = 0, e = Scalars.size(); i != e; ++i) {
1133 Value *ScalarVal = Scalars[i].Val;
1135 // Check to see if the scalar _IS_ a call...
1136 if (CallInst *CI = dyn_cast<CallInst>(ScalarVal))
1137 // If so, add information about the pool it will be returning...
1138 CallMap[CI].addCallInfo(DS, CI, -1, Scalars[i].Pool.Node, PoolDescs);
1140 // Check to see if the scalar is an operand to a call...
1141 for (Value::use_iterator UI = ScalarVal->use_begin(),
1142 UE = ScalarVal->use_end(); UI != UE; ++UI) {
1143 if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
1144 // Find out which operand this is to the call instruction...
1145 User::op_iterator OI = find(CI->op_begin(), CI->op_end(), ScalarVal);
1146 assert(OI != CI->op_end() && "Call on use list but not an operand!?");
1147 assert(OI != CI->op_begin() && "Pointer operand is call destination?");
1149 // FIXME: This is broken if the same pointer is passed to a call more
1150 // than once! It will get multiple entries for the first pointer.
1152 // Add the operand number and pool handle to the call table...
1153 CallMap[CI].addCallInfo(DS, CI, OI-CI->op_begin()-1,
1154 Scalars[i].Pool.Node, PoolDescs);
1159 // Make sure that all dependant arguments are added as well. For example, if
1160 // we call foo(null, P) and foo treats it's first and second arguments as
1161 // belonging to the same data structure, the we MUST set up the CallMap to
1162 // know that the null needs to be transformed into an index as well.
1164 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin();
1165 I != CallMap.end(); ++I)
1166 I->second.ensureDependantArgumentsIncluded(DS, PoolDescs);
1168 #ifdef DEBUG_TRANSFORM_PROGRESS
1169 // Print out call map...
1170 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin();
1171 I != CallMap.end(); ++I) {
1172 cerr << "For call: " << I->first;
1173 cerr << I->second.Func->getName() << " must pass pool pointer for args #";
1174 for (unsigned i = 0; i < I->second.ArgInfo.size(); ++i)
1175 cerr << I->second.ArgInfo[i].ArgNo << ", ";
1180 // Loop through all of the call nodes, recursively creating the new functions
1181 // that we want to call... This uses a map to prevent infinite recursion and
1182 // to avoid duplicating functions unneccesarily.
1184 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin(),
1185 E = CallMap.end(); I != E; ++I) {
1186 // Transform all of the functions we need, or at least ensure there is a
1187 // cached version available.
1188 transformFunction(I->second, IPFGraph, PoolDescs);
1191 // Now that all of the functions that we want to call are available, transform
1192 // the local function so that it uses the pools locally and passes them to the
1193 // functions that we just hacked up.
1196 // First step, find the instructions to be modified.
1197 vector<Instruction*> InstToFix;
1198 for (unsigned i = 0, e = Scalars.size(); i != e; ++i) {
1199 Value *ScalarVal = Scalars[i].Val;
1201 // Check to see if the scalar _IS_ an instruction. If so, it is involved.
1202 if (Instruction *Inst = dyn_cast<Instruction>(ScalarVal))
1203 InstToFix.push_back(Inst);
1205 // All all of the instructions that use the scalar as an operand...
1206 for (Value::use_iterator UI = ScalarVal->use_begin(),
1207 UE = ScalarVal->use_end(); UI != UE; ++UI)
1208 InstToFix.push_back(cast<Instruction>(*UI));
1211 // Make sure that we get return instructions that return a null value from the
1214 if (!IPFGraph.getRetNodes().empty()) {
1215 assert(IPFGraph.getRetNodes().size() == 1 && "Can only return one node?");
1216 PointerVal RetNode = IPFGraph.getRetNodes()[0];
1217 assert(RetNode.Index == 0 && "Subindexing not implemented yet!");
1219 // Only process return instructions if the return value of this function is
1220 // part of one of the data structures we are transforming...
1222 if (PoolDescs.count(RetNode.Node)) {
1223 // Loop over all of the basic blocks, adding return instructions...
1224 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
1225 if (ReturnInst *RI = dyn_cast<ReturnInst>(I->getTerminator()))
1226 InstToFix.push_back(RI);
1232 // Eliminate duplicates by sorting, then removing equal neighbors.
1233 sort(InstToFix.begin(), InstToFix.end());
1234 InstToFix.erase(unique(InstToFix.begin(), InstToFix.end()), InstToFix.end());
1236 // Loop over all of the instructions to transform, creating the new
1237 // replacement instructions for them. This also unlinks them from the
1238 // function so they can be safely deleted later.
1240 map<Value*, Value*> XFormMap;
1241 NewInstructionCreator NIC(*this, Scalars, CallMap, XFormMap);
1243 // Visit all instructions... creating the new instructions that we need and
1244 // unlinking the old instructions from the function...
1246 #ifdef DEBUG_TRANSFORM_PROGRESS
1247 for (unsigned i = 0, e = InstToFix.size(); i != e; ++i) {
1248 cerr << "Fixing: " << InstToFix[i];
1249 NIC.visit(*InstToFix[i]);
1252 NIC.visit(InstToFix.begin(), InstToFix.end());
1255 // Make all instructions we will delete "let go" of their operands... so that
1256 // we can safely delete Arguments whose types have changed...
1258 for_each(InstToFix.begin(), InstToFix.end(),
1259 mem_fun(&Instruction::dropAllReferences));
1261 // Loop through all of the pointer arguments coming into the function,
1262 // replacing them with arguments of POINTERTYPE to match the function type of
1265 FunctionType::ParamTypes::const_iterator TI =
1266 F->getFunctionType()->getParamTypes().begin();
1267 for (Function::aiterator I = F->abegin(), E = F->aend(); I != E; ++I, ++TI) {
1268 if (I->getType() != *TI) {
1269 assert(isa<PointerType>(I->getType()) && *TI == POINTERTYPE);
1270 Argument *NewArg = new Argument(*TI, I->getName());
1271 XFormMap[I] = NewArg; // Map old arg into new arg...
1273 // Replace the old argument and then delete it...
1274 I = F->getArgumentList().erase(I);
1275 I = F->getArgumentList().insert(I, NewArg);
1279 // Now that all of the new instructions have been created, we can update all
1280 // of the references to dummy values to be references to the actual values
1281 // that are computed.
1283 NIC.updateReferences();
1285 #ifdef DEBUG_TRANSFORM_PROGRESS
1286 cerr << "TRANSFORMED FUNCTION:\n" << F;
1289 // Delete all of the "instructions to fix"
1290 for_each(InstToFix.begin(), InstToFix.end(), deleter<Instruction>);
1292 // Eliminate pool base loads that we can easily prove are redundant
1294 PoolBaseLoadEliminator(PoolDescs).visit(F);
1296 // Since we have liberally hacked the function to pieces, we want to inform
1297 // the datastructure pass that its internal representation is out of date.
1299 DS->invalidateFunction(F);
1304 // transformFunction - Transform the specified function the specified way. It
1305 // we have already transformed that function that way, don't do anything. The
1306 // nodes in the TransformFunctionInfo come out of callers data structure graph.
1308 void PoolAllocate::transformFunction(TransformFunctionInfo &TFI,
1309 FunctionDSGraph &CallerIPGraph,
1310 map<DSNode*, PoolInfo> &CallerPoolDesc) {
1311 if (getTransformedFunction(TFI)) return; // Function xformation already done?
1313 #ifdef DEBUG_TRANSFORM_PROGRESS
1314 cerr << "********** Entering transformFunction for "
1315 << TFI.Func->getName() << ":\n";
1316 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i)
1317 cerr << " ArgInfo[" << i << "] = " << TFI.ArgInfo[i].ArgNo << "\n";
1321 const FunctionType *OldFuncType = TFI.Func->getFunctionType();
1323 assert(!OldFuncType->isVarArg() && "Vararg functions not handled yet!");
1325 // Build the type for the new function that we are transforming
1326 vector<const Type*> ArgTys;
1327 ArgTys.reserve(OldFuncType->getNumParams()+TFI.ArgInfo.size());
1328 for (unsigned i = 0, e = OldFuncType->getNumParams(); i != e; ++i)
1329 ArgTys.push_back(OldFuncType->getParamType(i));
1331 const Type *RetType = OldFuncType->getReturnType();
1333 // Add one pool pointer for every argument that needs to be supplemented.
1334 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
1335 if (TFI.ArgInfo[i].ArgNo == -1)
1336 RetType = POINTERTYPE; // Return a pointer
1338 ArgTys[TFI.ArgInfo[i].ArgNo] = POINTERTYPE; // Pass a pointer
1339 ArgTys.push_back(PointerType::get(CallerPoolDesc.find(TFI.ArgInfo[i].Node)
1340 ->second.PoolType));
1343 // Build the new function type...
1344 const FunctionType *NewFuncType = FunctionType::get(RetType, ArgTys,
1345 OldFuncType->isVarArg());
1347 // The new function is internal, because we know that only we can call it.
1348 // This also helps subsequent IP transformations to eliminate duplicated pool
1349 // pointers (which look like the same value is always passed into a parameter,
1350 // allowing it to be easily eliminated).
1352 Function *NewFunc = new Function(NewFuncType, true,
1353 TFI.Func->getName()+".poolxform");
1354 CurModule->getFunctionList().push_back(NewFunc);
1357 #ifdef DEBUG_TRANSFORM_PROGRESS
1358 cerr << "Created function prototype: " << NewFunc << "\n";
1361 // Add the newly formed function to the TransformedFunctions table so that
1362 // infinite recursion does not occur!
1364 TransformedFunctions[TFI] = NewFunc;
1366 // Add arguments to the function... starting with all of the old arguments
1367 vector<Value*> ArgMap;
1368 for (Function::const_aiterator I = TFI.Func->abegin(), E = TFI.Func->aend();
1370 Argument *NFA = new Argument(I->getType(), I->getName());
1371 NewFunc->getArgumentList().push_back(NFA);
1372 ArgMap.push_back(NFA); // Keep track of the arguments
1375 // Now add all of the arguments corresponding to pools passed in...
1376 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
1377 CallArgInfo &AI = TFI.ArgInfo[i];
1382 Name = ArgMap[AI.ArgNo]->getName(); // Get the arg name
1383 const Type *Ty = PointerType::get(CallerPoolDesc[AI.Node].PoolType);
1384 Argument *NFA = new Argument(Ty, Name+".pool");
1385 NewFunc->getArgumentList().push_back(NFA);
1388 // Now clone the body of the old function into the new function...
1389 CloneFunctionInto(NewFunc, TFI.Func, ArgMap);
1391 // Okay, now we have a function that is identical to the old one, except that
1392 // it has extra arguments for the pools coming in. Now we have to get the
1393 // data structure graph for the function we are replacing, and figure out how
1394 // our graph nodes map to the graph nodes in the dest function.
1396 FunctionDSGraph &DSGraph = DS->getClosedDSGraph(NewFunc);
1398 // NodeMapping - Multimap from callers graph to called graph. We are
1399 // guaranteed that the called function graph has more nodes than the caller,
1400 // or exactly the same number of nodes. This is because the called function
1401 // might not know that two nodes are merged when considering the callers
1402 // context, but the caller obviously does. Because of this, a single node in
1403 // the calling function's data structure graph can map to multiple nodes in
1404 // the called functions graph.
1406 map<DSNode*, PointerValSet> NodeMapping;
1408 CalculateNodeMapping(NewFunc, TFI, CallerIPGraph, DSGraph,
1411 // Print out the node mapping...
1412 #ifdef DEBUG_TRANSFORM_PROGRESS
1413 cerr << "\nNode mapping for call of " << NewFunc->getName() << "\n";
1414 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin();
1415 I != NodeMapping.end(); ++I) {
1416 cerr << "Map: "; I->first->print(cerr);
1417 cerr << "To: "; I->second.print(cerr);
1422 // Fill in the PoolDescriptor information for the transformed function so that
1423 // it can determine which value holds the pool descriptor for each data
1424 // structure node that it accesses.
1426 map<DSNode*, PoolInfo> PoolDescs;
1428 #ifdef DEBUG_TRANSFORM_PROGRESS
1429 cerr << "\nCalculating the pool descriptor map:\n";
1432 // Calculate as much of the pool descriptor map as possible. Since we have
1433 // the node mapping between the caller and callee functions, and we have the
1434 // pool descriptor information of the caller, we can calculate a partical pool
1435 // descriptor map for the called function.
1437 // The nodes that we do not have complete information for are the ones that
1438 // are accessed by loading pointers derived from arguments passed in, but that
1439 // are not passed in directly. In this case, we have all of the information
1440 // except a pool value. If the called function refers to this pool, the pool
1441 // value will be loaded from the pool graph and added to the map as neccesary.
1443 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin();
1444 I != NodeMapping.end(); ++I) {
1445 DSNode *CallerNode = I->first;
1446 PoolInfo &CallerPI = CallerPoolDesc[CallerNode];
1448 // Check to see if we have a node pointer passed in for this value...
1449 Value *CalleeValue = 0;
1450 for (unsigned a = 0, ae = TFI.ArgInfo.size(); a != ae; ++a)
1451 if (TFI.ArgInfo[a].Node == CallerNode) {
1452 // Calculate the argument number that the pool is to the function
1453 // call... The call instruction should not have the pool operands added
1455 unsigned ArgNo = TFI.Call->getNumOperands()-1+a;
1456 #ifdef DEBUG_TRANSFORM_PROGRESS
1457 cerr << "Should be argument #: " << ArgNo << "[i = " << a << "]\n";
1459 assert(ArgNo < NewFunc->asize() &&
1460 "Call already has pool arguments added??");
1462 // Map the pool argument into the called function...
1463 Function::aiterator AI = NewFunc->abegin();
1464 std::advance(AI, ArgNo);
1466 break; // Found value, quit loop
1469 // Loop over all of the data structure nodes that this incoming node maps to
1470 // Creating a PoolInfo structure for them.
1471 for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
1472 assert(I->second[i].Index == 0 && "Doesn't handle subindexing yet!");
1473 DSNode *CalleeNode = I->second[i].Node;
1475 // Add the descriptor. We already know everything about it by now, much
1476 // of it is the same as the caller info.
1478 PoolDescs.insert(make_pair(CalleeNode,
1479 PoolInfo(CalleeNode, CalleeValue,
1481 CallerPI.PoolType)));
1485 // We must destroy the node mapping so that we don't have latent references
1486 // into the data structure graph for the new function. Otherwise we get
1487 // assertion failures when transformFunctionBody tries to invalidate the
1490 NodeMapping.clear();
1492 // Now that we know everything we need about the function, transform the body
1495 transformFunctionBody(NewFunc, DSGraph, PoolDescs);
1497 #ifdef DEBUG_TRANSFORM_PROGRESS
1498 cerr << "Function after transformation:\n" << NewFunc;
1502 static unsigned countPointerTypes(const Type *Ty) {
1503 if (isa<PointerType>(Ty)) {
1505 } else if (const StructType *STy = dyn_cast<StructType>(Ty)) {
1507 for (unsigned i = 0, e = STy->getElementTypes().size(); i != e; ++i)
1508 Num += countPointerTypes(STy->getElementTypes()[i]);
1510 } else if (const ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1511 return countPointerTypes(ATy->getElementType());
1513 assert(Ty->isPrimitiveType() && "Unknown derived type!");
1518 // CreatePools - Insert instructions into the function we are processing to
1519 // create all of the memory pool objects themselves. This also inserts
1520 // destruction code. Add an alloca for each pool that is allocated to the
1521 // PoolDescs vector.
1523 void PoolAllocate::CreatePools(Function *F, const vector<AllocDSNode*> &Allocs,
1524 map<DSNode*, PoolInfo> &PoolDescs) {
1525 // Find all of the return nodes in the function...
1526 vector<BasicBlock*> ReturnNodes;
1527 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
1528 if (isa<ReturnInst>(I->getTerminator()))
1529 ReturnNodes.push_back(I);
1531 #ifdef DEBUG_CREATE_POOLS
1532 cerr << "Allocs that we are pool allocating:\n";
1533 for (unsigned i = 0, e = Allocs.size(); i != e; ++i)
1537 map<DSNode*, PATypeHolder> AbsPoolTyMap;
1539 // First pass over the allocations to process...
1540 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1541 // Create the pooldescriptor mapping... with null entries for everything
1542 // except the node & NewType fields.
1544 map<DSNode*, PoolInfo>::iterator PI =
1545 PoolDescs.insert(make_pair(Allocs[i], PoolInfo(Allocs[i]))).first;
1547 // Add a symbol table entry for the new type if there was one for the old
1549 string OldName = CurModule->getTypeName(Allocs[i]->getType());
1550 if (OldName.empty()) OldName = "node";
1551 CurModule->addTypeName(OldName+".p", PI->second.NewType);
1553 // Create the abstract pool types that will need to be resolved in a second
1554 // pass once an abstract type is created for each pool.
1556 // Can only handle limited shapes for now...
1557 const Type *OldNodeTy = Allocs[i]->getType();
1558 vector<const Type*> PoolTypes;
1560 // Pool type is the first element of the pool descriptor type...
1561 PoolTypes.push_back(getPoolType(PoolDescs[Allocs[i]].NewType));
1563 unsigned NumPointers = countPointerTypes(OldNodeTy);
1564 while (NumPointers--) // Add a different opaque type for each pointer
1565 PoolTypes.push_back(OpaqueType::get());
1567 assert(Allocs[i]->getNumLinks() == PoolTypes.size()-1 &&
1568 "Node should have same number of pointers as pool!");
1570 StructType *PoolType = StructType::get(PoolTypes);
1572 // Add a symbol table entry for the pooltype if possible...
1573 CurModule->addTypeName(OldName+".pool", PoolType);
1575 // Create the pool type, with opaque values for pointers...
1576 AbsPoolTyMap.insert(make_pair(Allocs[i], PoolType));
1577 #ifdef DEBUG_CREATE_POOLS
1578 cerr << "POOL TY: " << AbsPoolTyMap.find(Allocs[i])->second.get() << "\n";
1582 // Now that we have types for all of the pool types, link them all together.
1583 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1584 PATypeHolder &PoolTyH = AbsPoolTyMap.find(Allocs[i])->second;
1586 // Resolve all of the outgoing pointer types of this pool node...
1587 for (unsigned p = 0, pe = Allocs[i]->getNumLinks(); p != pe; ++p) {
1588 PointerValSet &PVS = Allocs[i]->getLink(p);
1589 assert(!PVS.empty() && "Outgoing edge is empty, field unused, can"
1590 " probably just leave the type opaque or something dumb.");
1592 for (Out = 0; AbsPoolTyMap.count(PVS[Out].Node) == 0; ++Out)
1593 assert(Out != PVS.size() && "No edge to an outgoing allocation node!?");
1595 assert(PVS[Out].Index == 0 && "Subindexing not implemented yet!");
1597 // The actual struct type could change each time through the loop, so it's
1598 // NOT loop invariant.
1599 const StructType *PoolTy = cast<StructType>(PoolTyH.get());
1601 // Get the opaque type...
1602 DerivedType *ElTy = (DerivedType*)(PoolTy->getElementTypes()[p+1].get());
1604 #ifdef DEBUG_CREATE_POOLS
1605 cerr << "Refining " << ElTy << " of " << PoolTy << " to "
1606 << AbsPoolTyMap.find(PVS[Out].Node)->second.get() << "\n";
1609 const Type *RefPoolTy = AbsPoolTyMap.find(PVS[Out].Node)->second.get();
1610 ElTy->refineAbstractTypeTo(PointerType::get(RefPoolTy));
1612 #ifdef DEBUG_CREATE_POOLS
1613 cerr << "Result pool type is: " << PoolTyH.get() << "\n";
1618 // Create the code that goes in the entry and exit nodes for the function...
1619 vector<Instruction*> EntryNodeInsts;
1620 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1621 PoolInfo &PI = PoolDescs[Allocs[i]];
1623 // Fill in the pool type for this pool...
1624 PI.PoolType = AbsPoolTyMap.find(Allocs[i])->second.get();
1625 assert(!PI.PoolType->isAbstract() &&
1626 "Pool type should not be abstract anymore!");
1628 // Add an allocation and a free for each pool...
1629 AllocaInst *PoolAlloc
1630 = new AllocaInst(PointerType::get(PI.PoolType), 0,
1631 CurModule->getTypeName(PI.PoolType));
1632 PI.Handle = PoolAlloc;
1633 EntryNodeInsts.push_back(PoolAlloc);
1634 AllocationInst *AI = Allocs[i]->getAllocation();
1636 // Initialize the pool. We need to know how big each allocation is. For
1637 // our purposes here, we assume we are allocating a scalar, or array of
1640 unsigned ElSize = TargetData.getTypeSize(PI.NewType);
1642 vector<Value*> Args;
1643 Args.push_back(ConstantUInt::get(Type::UIntTy, ElSize));
1644 Args.push_back(PoolAlloc); // Pool to initialize
1645 EntryNodeInsts.push_back(new CallInst(PoolInit, Args));
1647 // Add code to destroy the pool in all of the exit nodes of the function...
1649 Args.push_back(PoolAlloc); // Pool to initialize
1651 for (unsigned EN = 0, ENE = ReturnNodes.size(); EN != ENE; ++EN) {
1652 Instruction *Destroy = new CallInst(PoolDestroy, Args);
1654 // Insert it before the return instruction...
1655 BasicBlock *RetNode = ReturnNodes[EN];
1656 RetNode->getInstList().insert(RetNode->end()--, Destroy);
1660 // Now that all of the pool descriptors have been created, link them together
1661 // so that called functions can get links as neccesary...
1663 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1664 PoolInfo &PI = PoolDescs[Allocs[i]];
1666 // For every pointer in the data structure, initialize a link that
1667 // indicates which pool to access...
1669 vector<Value*> Indices(2);
1670 Indices[0] = ConstantUInt::get(Type::UIntTy, 0);
1671 for (unsigned l = 0, le = PI.Node->getNumLinks(); l != le; ++l)
1672 // Only store an entry for the field if the field is used!
1673 if (!PI.Node->getLink(l).empty()) {
1674 assert(PI.Node->getLink(l).size() == 1 && "Should have only one link!");
1675 PointerVal PV = PI.Node->getLink(l)[0];
1676 assert(PV.Index == 0 && "Subindexing not supported yet!");
1677 PoolInfo &LinkedPool = PoolDescs[PV.Node];
1678 Indices[1] = ConstantUInt::get(Type::UByteTy, 1+l);
1680 EntryNodeInsts.push_back(new StoreInst(LinkedPool.Handle, PI.Handle,
1685 // Insert the entry node code into the entry block...
1686 F->getEntryNode().getInstList().insert(++F->getEntryNode().begin(),
1687 EntryNodeInsts.begin(),
1688 EntryNodeInsts.end());
1692 // addPoolPrototypes - Add prototypes for the pool functions to the specified
1693 // module and update the Pool* instance variables to point to them.
1695 void PoolAllocate::addPoolPrototypes(Module &M) {
1696 // Get poolinit function...
1697 vector<const Type*> Args;
1698 Args.push_back(Type::UIntTy); // Num bytes per element
1699 FunctionType *PoolInitTy = FunctionType::get(Type::VoidTy, Args, true);
1700 PoolInit = M.getOrInsertFunction("poolinit", PoolInitTy);
1702 // Get pooldestroy function...
1703 Args.pop_back(); // Only takes a pool...
1704 FunctionType *PoolDestroyTy = FunctionType::get(Type::VoidTy, Args, true);
1705 PoolDestroy = M.getOrInsertFunction("pooldestroy", PoolDestroyTy);
1707 // Get the poolalloc function...
1708 FunctionType *PoolAllocTy = FunctionType::get(POINTERTYPE, Args, true);
1709 PoolAlloc = M.getOrInsertFunction("poolalloc", PoolAllocTy);
1711 // Get the poolfree function...
1712 Args.push_back(POINTERTYPE); // Pointer to free
1713 FunctionType *PoolFreeTy = FunctionType::get(Type::VoidTy, Args, true);
1714 PoolFree = M.getOrInsertFunction("poolfree", PoolFreeTy);
1716 Args[0] = Type::UIntTy; // Number of slots to allocate
1717 FunctionType *PoolAllocArrayTy = FunctionType::get(POINTERTYPE, Args, true);
1718 PoolAllocArray = M.getOrInsertFunction("poolallocarray", PoolAllocArrayTy);
1722 bool PoolAllocate::run(Module &M) {
1723 addPoolPrototypes(M);
1726 DS = &getAnalysis<DataStructure>();
1727 bool Changed = false;
1729 for (Module::iterator I = M.begin(); I != M.end(); ++I)
1730 if (!I->isExternal()) {
1731 Changed |= processFunction(I);
1733 cerr << "Only processing one function\n";
1744 // createPoolAllocatePass - Global function to access the functionality of this
1747 Pass *createPoolAllocatePass() { return new PoolAllocate(); }