1 //===-- ArgumentPromotion.cpp - Promote by-reference arguments ------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // This pass promotes "by reference" arguments to be "by value" arguments. In
11 // practice, this means looking for internal functions that have pointer
12 // arguments. If it can prove, through the use of alias analysis, that an
13 // argument is *only* loaded, then it can pass the value into the function
14 // instead of the address of the value. This can cause recursive simplification
15 // of code and lead to the elimination of allocas (especially in C++ template
16 // code like the STL).
18 // This pass also handles aggregate arguments that are passed into a function,
19 // scalarizing them if the elements of the aggregate are only loaded. Note that
20 // by default it refuses to scalarize aggregates which would require passing in
21 // more than three operands to the function, because passing thousands of
22 // operands for a large array or structure is unprofitable! This limit can be
23 // configured or disabled, however.
25 // Note that this transformation could also be done for arguments that are only
26 // stored to (returning the value instead), but does not currently. This case
27 // would be best handled when and if LLVM begins supporting multiple return
28 // values from functions.
30 //===----------------------------------------------------------------------===//
32 #define DEBUG_TYPE "argpromotion"
33 #include "llvm/Transforms/IPO.h"
34 #include "llvm/Constants.h"
35 #include "llvm/DerivedTypes.h"
36 #include "llvm/Module.h"
37 #include "llvm/CallGraphSCCPass.h"
38 #include "llvm/Instructions.h"
39 #include "llvm/LLVMContext.h"
40 #include "llvm/Analysis/AliasAnalysis.h"
41 #include "llvm/Analysis/CallGraph.h"
42 #include "llvm/Target/TargetData.h"
43 #include "llvm/Support/CallSite.h"
44 #include "llvm/Support/CFG.h"
45 #include "llvm/Support/Debug.h"
46 #include "llvm/Support/raw_ostream.h"
47 #include "llvm/ADT/DepthFirstIterator.h"
48 #include "llvm/ADT/Statistic.h"
49 #include "llvm/ADT/StringExtras.h"
53 STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
54 STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
55 STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted");
56 STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated");
59 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
61 struct ArgPromotion : public CallGraphSCCPass {
62 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
63 AU.addRequired<AliasAnalysis>();
64 CallGraphSCCPass::getAnalysisUsage(AU);
67 virtual bool runOnSCC(CallGraphSCC &SCC);
68 static char ID; // Pass identification, replacement for typeid
69 explicit ArgPromotion(unsigned maxElements = 3)
70 : CallGraphSCCPass(ID), maxElements(maxElements) {}
72 /// A vector used to hold the indices of a single GEP instruction
73 typedef std::vector<uint64_t> IndicesVector;
76 CallGraphNode *PromoteArguments(CallGraphNode *CGN);
77 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const;
78 CallGraphNode *DoPromotion(Function *F,
79 SmallPtrSet<Argument*, 8> &ArgsToPromote,
80 SmallPtrSet<Argument*, 8> &ByValArgsToTransform);
81 /// The maximum number of elements to expand, or 0 for unlimited.
86 char ArgPromotion::ID = 0;
87 INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion",
88 "Promote 'by reference' arguments to scalars", false, false)
89 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
90 INITIALIZE_AG_DEPENDENCY(CallGraph)
91 INITIALIZE_PASS_END(ArgPromotion, "argpromotion",
92 "Promote 'by reference' arguments to scalars", false, false)
94 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
95 return new ArgPromotion(maxElements);
98 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
99 bool Changed = false, LocalChange;
101 do { // Iterate until we stop promoting from this SCC.
103 // Attempt to promote arguments from all functions in this SCC.
104 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
105 if (CallGraphNode *CGN = PromoteArguments(*I)) {
107 SCC.ReplaceNode(*I, CGN);
110 Changed |= LocalChange; // Remember that we changed something.
111 } while (LocalChange);
116 /// PromoteArguments - This method checks the specified function to see if there
117 /// are any promotable arguments and if it is safe to promote the function (for
118 /// example, all callers are direct). If safe to promote some arguments, it
119 /// calls the DoPromotion method.
121 CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
122 Function *F = CGN->getFunction();
124 // Make sure that it is local to this module.
125 if (!F || !F->hasLocalLinkage()) return 0;
127 // First check: see if there are any pointer arguments! If not, quick exit.
128 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs;
130 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
131 I != E; ++I, ++ArgNo)
132 if (I->getType()->isPointerTy())
133 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo));
134 if (PointerArgs.empty()) return 0;
136 // Second check: make sure that all callers are direct callers. We can't
137 // transform functions that have indirect callers.
138 if (F->hasAddressTaken())
141 // Check to see which arguments are promotable. If an argument is promotable,
142 // add it to ArgsToPromote.
143 SmallPtrSet<Argument*, 8> ArgsToPromote;
144 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
145 for (unsigned i = 0; i != PointerArgs.size(); ++i) {
146 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal);
148 // If this is a byval argument, and if the aggregate type is small, just
149 // pass the elements, which is always safe.
150 Argument *PtrArg = PointerArgs[i].first;
152 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
153 if (const StructType *STy = dyn_cast<StructType>(AgTy)) {
154 if (maxElements > 0 && STy->getNumElements() > maxElements) {
155 DEBUG(dbgs() << "argpromotion disable promoting argument '"
156 << PtrArg->getName() << "' because it would require adding more"
157 << " than " << maxElements << " arguments to the function.\n");
159 // If all the elements are single-value types, we can promote it.
160 bool AllSimple = true;
161 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
162 if (!STy->getElementType(i)->isSingleValueType()) {
167 // Safe to transform, don't even bother trying to "promote" it.
168 // Passing the elements as a scalar will allow scalarrepl to hack on
169 // the new alloca we introduce.
171 ByValArgsToTransform.insert(PtrArg);
178 // Otherwise, see if we can promote the pointer to its value.
179 if (isSafeToPromoteArgument(PtrArg, isByVal))
180 ArgsToPromote.insert(PtrArg);
183 // No promotable pointer arguments.
184 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
187 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
190 /// IsAlwaysValidPointer - Return true if the specified pointer is always legal
192 static bool IsAlwaysValidPointer(Value *V) {
193 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
194 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V))
195 return IsAlwaysValidPointer(GEP->getOperand(0));
196 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
197 if (CE->getOpcode() == Instruction::GetElementPtr)
198 return IsAlwaysValidPointer(CE->getOperand(0));
203 /// AllCalleesPassInValidPointerForArgument - Return true if we can prove that
204 /// all callees pass in a valid pointer for the specified function argument.
205 static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) {
206 Function *Callee = Arg->getParent();
208 unsigned ArgNo = std::distance(Callee->arg_begin(),
209 Function::arg_iterator(Arg));
211 // Look at all call sites of the function. At this pointer we know we only
212 // have direct callees.
213 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end();
216 assert(CS && "Should only have direct calls!");
218 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo)))
224 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
225 /// that is greater than or equal to the size of prefix, and each of the
226 /// elements in Prefix is the same as the corresponding elements in Longer.
228 /// This means it also returns true when Prefix and Longer are equal!
229 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
230 const ArgPromotion::IndicesVector &Longer) {
231 if (Prefix.size() > Longer.size())
233 for (unsigned i = 0, e = Prefix.size(); i != e; ++i)
234 if (Prefix[i] != Longer[i])
240 /// Checks if Indices, or a prefix of Indices, is in Set.
241 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
242 std::set<ArgPromotion::IndicesVector> &Set) {
243 std::set<ArgPromotion::IndicesVector>::iterator Low;
244 Low = Set.upper_bound(Indices);
245 if (Low != Set.begin())
247 // Low is now the last element smaller than or equal to Indices. This means
248 // it points to a prefix of Indices (possibly Indices itself), if such
251 // This load is safe if any prefix of its operands is safe to load.
252 return Low != Set.end() && IsPrefix(*Low, Indices);
255 /// Mark the given indices (ToMark) as safe in the given set of indices
256 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
257 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
258 /// already. Furthermore, any indices that Indices is itself a prefix of, are
259 /// removed from Safe (since they are implicitely safe because of Indices now).
260 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
261 std::set<ArgPromotion::IndicesVector> &Safe) {
262 std::set<ArgPromotion::IndicesVector>::iterator Low;
263 Low = Safe.upper_bound(ToMark);
264 // Guard against the case where Safe is empty
265 if (Low != Safe.begin())
267 // Low is now the last element smaller than or equal to Indices. This
268 // means it points to a prefix of Indices (possibly Indices itself), if
269 // such prefix exists.
270 if (Low != Safe.end()) {
271 if (IsPrefix(*Low, ToMark))
272 // If there is already a prefix of these indices (or exactly these
273 // indices) marked a safe, don't bother adding these indices
276 // Increment Low, so we can use it as a "insert before" hint
280 Low = Safe.insert(Low, ToMark);
282 // If there we're a prefix of longer index list(s), remove those
283 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
284 while (Low != End && IsPrefix(ToMark, *Low)) {
285 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
291 /// isSafeToPromoteArgument - As you might guess from the name of this method,
292 /// it checks to see if it is both safe and useful to promote the argument.
293 /// This method limits promotion of aggregates to only promote up to three
294 /// elements of the aggregate in order to avoid exploding the number of
295 /// arguments passed in.
296 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const {
297 typedef std::set<IndicesVector> GEPIndicesSet;
299 // Quick exit for unused arguments
300 if (Arg->use_empty())
303 // We can only promote this argument if all of the uses are loads, or are GEP
304 // instructions (with constant indices) that are subsequently loaded.
306 // Promoting the argument causes it to be loaded in the caller
307 // unconditionally. This is only safe if we can prove that either the load
308 // would have happened in the callee anyway (ie, there is a load in the entry
309 // block) or the pointer passed in at every call site is guaranteed to be
311 // In the former case, invalid loads can happen, but would have happened
312 // anyway, in the latter case, invalid loads won't happen. This prevents us
313 // from introducing an invalid load that wouldn't have happened in the
316 // This set will contain all sets of indices that are loaded in the entry
317 // block, and thus are safe to unconditionally load in the caller.
318 GEPIndicesSet SafeToUnconditionallyLoad;
320 // This set contains all the sets of indices that we are planning to promote.
321 // This makes it possible to limit the number of arguments added.
322 GEPIndicesSet ToPromote;
324 // If the pointer is always valid, any load with first index 0 is valid.
325 if (isByVal || AllCalleesPassInValidPointerForArgument(Arg))
326 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
328 // First, iterate the entry block and mark loads of (geps of) arguments as
330 BasicBlock *EntryBlock = Arg->getParent()->begin();
331 // Declare this here so we can reuse it
332 IndicesVector Indices;
333 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
335 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
336 Value *V = LI->getPointerOperand();
337 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
338 V = GEP->getPointerOperand();
340 // This load actually loads (part of) Arg? Check the indices then.
341 Indices.reserve(GEP->getNumIndices());
342 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
344 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
345 Indices.push_back(CI->getSExtValue());
347 // We found a non-constant GEP index for this argument? Bail out
348 // right away, can't promote this argument at all.
351 // Indices checked out, mark them as safe
352 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
355 } else if (V == Arg) {
356 // Direct loads are equivalent to a GEP with a single 0 index.
357 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
361 // Now, iterate all uses of the argument to see if there are any uses that are
362 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
363 SmallVector<LoadInst*, 16> Loads;
364 IndicesVector Operands;
365 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
369 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
370 if (LI->isVolatile()) return false; // Don't hack volatile loads
372 // Direct loads are equivalent to a GEP with a zero index and then a load.
373 Operands.push_back(0);
374 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
375 if (GEP->use_empty()) {
376 // Dead GEP's cause trouble later. Just remove them if we run into
378 getAnalysis<AliasAnalysis>().deleteValue(GEP);
379 GEP->eraseFromParent();
380 // TODO: This runs the above loop over and over again for dead GEPs
381 // Couldn't we just do increment the UI iterator earlier and erase the
383 return isSafeToPromoteArgument(Arg, isByVal);
386 // Ensure that all of the indices are constants.
387 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
389 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
390 Operands.push_back(C->getSExtValue());
392 return false; // Not a constant operand GEP!
394 // Ensure that the only users of the GEP are load instructions.
395 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
397 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
398 if (LI->isVolatile()) return false; // Don't hack volatile loads
401 // Other uses than load?
405 return false; // Not a load or a GEP.
408 // Now, see if it is safe to promote this load / loads of this GEP. Loading
409 // is safe if Operands, or a prefix of Operands, is marked as safe.
410 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
413 // See if we are already promoting a load with these indices. If not, check
414 // to make sure that we aren't promoting too many elements. If so, nothing
416 if (ToPromote.find(Operands) == ToPromote.end()) {
417 if (maxElements > 0 && ToPromote.size() == maxElements) {
418 DEBUG(dbgs() << "argpromotion not promoting argument '"
419 << Arg->getName() << "' because it would require adding more "
420 << "than " << maxElements << " arguments to the function.\n");
421 // We limit aggregate promotion to only promoting up to a fixed number
422 // of elements of the aggregate.
425 ToPromote.insert(Operands);
429 if (Loads.empty()) return true; // No users, this is a dead argument.
431 // Okay, now we know that the argument is only used by load instructions and
432 // it is safe to unconditionally perform all of them. Use alias analysis to
433 // check to see if the pointer is guaranteed to not be modified from entry of
434 // the function to each of the load instructions.
436 // Because there could be several/many load instructions, remember which
437 // blocks we know to be transparent to the load.
438 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
440 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
441 TargetData *TD = getAnalysisIfAvailable<TargetData>();
442 if (!TD) return false; // Without TargetData, assume the worst.
444 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
445 // Check to see if the load is invalidated from the start of the block to
447 LoadInst *Load = Loads[i];
448 BasicBlock *BB = Load->getParent();
450 const PointerType *LoadTy =
451 cast<PointerType>(Load->getPointerOperand()->getType());
452 unsigned LoadSize =(unsigned)TD->getTypeStoreSize(LoadTy->getElementType());
454 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize))
455 return false; // Pointer is invalidated!
457 // Now check every path from the entry block to the load for transparency.
458 // To do this, we perform a depth first search on the inverse CFG from the
460 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
462 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> >
463 I = idf_ext_begin(P, TranspBlocks),
464 E = idf_ext_end(P, TranspBlocks); I != E; ++I)
465 if (AA.canBasicBlockModify(**I, Arg, LoadSize))
470 // If the path from the entry of the function to each load is free of
471 // instructions that potentially invalidate the load, we can make the
476 /// DoPromotion - This method actually performs the promotion of the specified
477 /// arguments, and returns the new function. At this point, we know that it's
479 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
480 SmallPtrSet<Argument*, 8> &ArgsToPromote,
481 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {
483 // Start by computing a new prototype for the function, which is the same as
484 // the old function, but has modified arguments.
485 const FunctionType *FTy = F->getFunctionType();
486 std::vector<const Type*> Params;
488 typedef std::set<IndicesVector> ScalarizeTable;
490 // ScalarizedElements - If we are promoting a pointer that has elements
491 // accessed out of it, keep track of which elements are accessed so that we
492 // can add one argument for each.
494 // Arguments that are directly loaded will have a zero element value here, to
495 // handle cases where there are both a direct load and GEP accesses.
497 std::map<Argument*, ScalarizeTable> ScalarizedElements;
499 // OriginalLoads - Keep track of a representative load instruction from the
500 // original function so that we can tell the alias analysis implementation
501 // what the new GEP/Load instructions we are inserting look like.
502 std::map<IndicesVector, LoadInst*> OriginalLoads;
504 // Attributes - Keep track of the parameter attributes for the arguments
505 // that we are *not* promoting. For the ones that we do promote, the parameter
506 // attributes are lost
507 SmallVector<AttributeWithIndex, 8> AttributesVec;
508 const AttrListPtr &PAL = F->getAttributes();
510 // Add any return attributes.
511 if (Attributes attrs = PAL.getRetAttributes())
512 AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
514 // First, determine the new argument list
515 unsigned ArgIndex = 1;
516 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
518 if (ByValArgsToTransform.count(I)) {
519 // Simple byval argument? Just add all the struct element types.
520 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
521 const StructType *STy = cast<StructType>(AgTy);
522 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
523 Params.push_back(STy->getElementType(i));
524 ++NumByValArgsPromoted;
525 } else if (!ArgsToPromote.count(I)) {
526 // Unchanged argument
527 Params.push_back(I->getType());
528 if (Attributes attrs = PAL.getParamAttributes(ArgIndex))
529 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs));
530 } else if (I->use_empty()) {
531 // Dead argument (which are always marked as promotable)
534 // Okay, this is being promoted. This means that the only uses are loads
535 // or GEPs which are only used by loads
537 // In this table, we will track which indices are loaded from the argument
538 // (where direct loads are tracked as no indices).
539 ScalarizeTable &ArgIndices = ScalarizedElements[I];
540 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
542 Instruction *User = cast<Instruction>(*UI);
543 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
544 IndicesVector Indices;
545 Indices.reserve(User->getNumOperands() - 1);
546 // Since loads will only have a single operand, and GEPs only a single
547 // non-index operand, this will record direct loads without any indices,
548 // and gep+loads with the GEP indices.
549 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
551 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
552 // GEPs with a single 0 index can be merged with direct loads
553 if (Indices.size() == 1 && Indices.front() == 0)
555 ArgIndices.insert(Indices);
557 if (LoadInst *L = dyn_cast<LoadInst>(User))
560 // Take any load, we will use it only to update Alias Analysis
561 OrigLoad = cast<LoadInst>(User->use_back());
562 OriginalLoads[Indices] = OrigLoad;
565 // Add a parameter to the function for each element passed in.
566 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
567 E = ArgIndices.end(); SI != E; ++SI) {
568 // not allowed to dereference ->begin() if size() is 0
569 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(),
572 assert(Params.back());
575 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
576 ++NumArgumentsPromoted;
578 ++NumAggregatesPromoted;
582 // Add any function attributes.
583 if (Attributes attrs = PAL.getFnAttributes())
584 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
586 const Type *RetTy = FTy->getReturnType();
588 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
589 // have zero fixed arguments.
590 bool ExtraArgHack = false;
591 if (Params.empty() && FTy->isVarArg()) {
593 Params.push_back(Type::getInt32Ty(F->getContext()));
596 // Construct the new function type using the new arguments.
597 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
599 // Create the new function body and insert it into the module.
600 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
601 NF->copyAttributesFrom(F);
604 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
607 // Recompute the parameter attributes list based on the new arguments for
609 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(),
610 AttributesVec.end()));
611 AttributesVec.clear();
613 F->getParent()->getFunctionList().insert(F, NF);
616 // Get the alias analysis information that we need to update to reflect our
618 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
620 // Get the callgraph information that we need to update to reflect our
622 CallGraph &CG = getAnalysis<CallGraph>();
624 // Get a new callgraph node for NF.
625 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
627 // Loop over all of the callers of the function, transforming the call sites
628 // to pass in the loaded pointers.
630 SmallVector<Value*, 16> Args;
631 while (!F->use_empty()) {
632 CallSite CS(F->use_back());
633 assert(CS.getCalledFunction() == F);
634 Instruction *Call = CS.getInstruction();
635 const AttrListPtr &CallPAL = CS.getAttributes();
637 // Add any return attributes.
638 if (Attributes attrs = CallPAL.getRetAttributes())
639 AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
641 // Loop over the operands, inserting GEP and loads in the caller as
643 CallSite::arg_iterator AI = CS.arg_begin();
645 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
646 I != E; ++I, ++AI, ++ArgIndex)
647 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
648 Args.push_back(*AI); // Unmodified argument
650 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
651 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
653 } else if (ByValArgsToTransform.count(I)) {
654 // Emit a GEP and load for each element of the struct.
655 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
656 const StructType *STy = cast<StructType>(AgTy);
658 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
659 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
660 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
661 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2,
662 (*AI)->getName()+"."+utostr(i),
664 // TODO: Tell AA about the new values?
665 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
667 } else if (!I->use_empty()) {
668 // Non-dead argument: insert GEPs and loads as appropriate.
669 ScalarizeTable &ArgIndices = ScalarizedElements[I];
670 // Store the Value* version of the indices in here, but declare it now
672 std::vector<Value*> Ops;
673 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
674 E = ArgIndices.end(); SI != E; ++SI) {
676 LoadInst *OrigLoad = OriginalLoads[*SI];
678 Ops.reserve(SI->size());
679 const Type *ElTy = V->getType();
680 for (IndicesVector::const_iterator II = SI->begin(),
681 IE = SI->end(); II != IE; ++II) {
682 // Use i32 to index structs, and i64 for others (pointers/arrays).
683 // This satisfies GEP constraints.
684 const Type *IdxTy = (ElTy->isStructTy() ?
685 Type::getInt32Ty(F->getContext()) :
686 Type::getInt64Ty(F->getContext()));
687 Ops.push_back(ConstantInt::get(IdxTy, *II));
688 // Keep track of the type we're currently indexing.
689 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
691 // And create a GEP to extract those indices.
692 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(),
693 V->getName()+".idx", Call);
695 AA.copyValue(OrigLoad->getOperand(0), V);
697 // Since we're replacing a load make sure we take the alignment
698 // of the previous load.
699 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
700 newLoad->setAlignment(OrigLoad->getAlignment());
701 Args.push_back(newLoad);
702 AA.copyValue(OrigLoad, Args.back());
707 Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext())));
709 // Push any varargs arguments on the list.
710 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
712 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
713 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
716 // Add any function attributes.
717 if (Attributes attrs = CallPAL.getFnAttributes())
718 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
721 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
722 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
723 Args.begin(), Args.end(), "", Call);
724 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
725 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
726 AttributesVec.end()));
728 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
729 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
730 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
731 AttributesVec.end()));
732 if (cast<CallInst>(Call)->isTailCall())
733 cast<CallInst>(New)->setTailCall();
736 AttributesVec.clear();
738 // Update the alias analysis implementation to know that we are replacing
739 // the old call with a new one.
740 AA.replaceWithNewValue(Call, New);
742 // Update the callgraph to know that the callsite has been transformed.
743 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
744 CalleeNode->replaceCallEdge(Call, New, NF_CGN);
746 if (!Call->use_empty()) {
747 Call->replaceAllUsesWith(New);
751 // Finally, remove the old call from the program, reducing the use-count of
753 Call->eraseFromParent();
756 // Since we have now created the new function, splice the body of the old
757 // function right into the new function, leaving the old rotting hulk of the
759 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
761 // Loop over the argument list, transfering uses of the old arguments over to
762 // the new arguments, also transfering over the names as well.
764 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
765 I2 = NF->arg_begin(); I != E; ++I) {
766 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
767 // If this is an unmodified argument, move the name and users over to the
769 I->replaceAllUsesWith(I2);
771 AA.replaceWithNewValue(I, I2);
776 if (ByValArgsToTransform.count(I)) {
777 // In the callee, we create an alloca, and store each of the new incoming
778 // arguments into the alloca.
779 Instruction *InsertPt = NF->begin()->begin();
781 // Just add all the struct element types.
782 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
783 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
784 const StructType *STy = cast<StructType>(AgTy);
786 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
788 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
789 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
791 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2,
792 TheAlloca->getName()+"."+Twine(i),
794 I2->setName(I->getName()+"."+Twine(i));
795 new StoreInst(I2++, Idx, InsertPt);
798 // Anything that used the arg should now use the alloca.
799 I->replaceAllUsesWith(TheAlloca);
800 TheAlloca->takeName(I);
801 AA.replaceWithNewValue(I, TheAlloca);
805 if (I->use_empty()) {
810 // Otherwise, if we promoted this argument, then all users are load
811 // instructions (or GEPs with only load users), and all loads should be
812 // using the new argument that we added.
813 ScalarizeTable &ArgIndices = ScalarizedElements[I];
815 while (!I->use_empty()) {
816 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
817 assert(ArgIndices.begin()->empty() &&
818 "Load element should sort to front!");
819 I2->setName(I->getName()+".val");
820 LI->replaceAllUsesWith(I2);
821 AA.replaceWithNewValue(LI, I2);
822 LI->eraseFromParent();
823 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
824 << "' in function '" << F->getName() << "'\n");
826 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
827 IndicesVector Operands;
828 Operands.reserve(GEP->getNumIndices());
829 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
831 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
833 // GEPs with a single 0 index can be merged with direct loads
834 if (Operands.size() == 1 && Operands.front() == 0)
837 Function::arg_iterator TheArg = I2;
838 for (ScalarizeTable::iterator It = ArgIndices.begin();
839 *It != Operands; ++It, ++TheArg) {
840 assert(It != ArgIndices.end() && "GEP not handled??");
843 std::string NewName = I->getName();
844 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
845 NewName += "." + utostr(Operands[i]);
848 TheArg->setName(NewName);
850 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
851 << "' of function '" << NF->getName() << "'\n");
853 // All of the uses must be load instructions. Replace them all with
854 // the argument specified by ArgNo.
855 while (!GEP->use_empty()) {
856 LoadInst *L = cast<LoadInst>(GEP->use_back());
857 L->replaceAllUsesWith(TheArg);
858 AA.replaceWithNewValue(L, TheArg);
859 L->eraseFromParent();
862 GEP->eraseFromParent();
866 // Increment I2 past all of the arguments added for this promoted pointer.
867 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
871 // Notify the alias analysis implementation that we inserted a new argument.
873 AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())),
877 // Tell the alias analysis that the old function is about to disappear.
878 AA.replaceWithNewValue(F, NF);
881 NF_CGN->stealCalledFunctionsFrom(CG[F]);
883 // Now that the old function is dead, delete it. If there is a dangling
884 // reference to the CallgraphNode, just leave the dead function around for
885 // someone else to nuke.
886 CallGraphNode *CGN = CG[F];
887 if (CGN->getNumReferences() == 0)
888 delete CG.removeFunctionFromModule(CGN);
890 F->setLinkage(Function::ExternalLinkage);