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/ADT/DepthFirstIterator.h"
35 #include "llvm/ADT/Statistic.h"
36 #include "llvm/ADT/StringExtras.h"
37 #include "llvm/Analysis/AliasAnalysis.h"
38 #include "llvm/Analysis/CallGraph.h"
39 #include "llvm/Analysis/CallGraphSCCPass.h"
40 #include "llvm/IR/Constants.h"
41 #include "llvm/IR/DerivedTypes.h"
42 #include "llvm/IR/Instructions.h"
43 #include "llvm/IR/LLVMContext.h"
44 #include "llvm/IR/Module.h"
45 #include "llvm/Support/CFG.h"
46 #include "llvm/Support/CallSite.h"
47 #include "llvm/Support/Debug.h"
48 #include "llvm/Support/raw_ostream.h"
52 STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
53 STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
54 STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted");
55 STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated");
58 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
60 struct ArgPromotion : public CallGraphSCCPass {
61 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
62 AU.addRequired<AliasAnalysis>();
63 CallGraphSCCPass::getAnalysisUsage(AU);
66 virtual bool runOnSCC(CallGraphSCC &SCC);
67 static char ID; // Pass identification, replacement for typeid
68 explicit ArgPromotion(unsigned maxElements = 3)
69 : CallGraphSCCPass(ID), maxElements(maxElements) {
70 initializeArgPromotionPass(*PassRegistry::getPassRegistry());
73 /// A vector used to hold the indices of a single GEP instruction
74 typedef std::vector<uint64_t> IndicesVector;
77 CallGraphNode *PromoteArguments(CallGraphNode *CGN);
78 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const;
79 CallGraphNode *DoPromotion(Function *F,
80 SmallPtrSet<Argument*, 8> &ArgsToPromote,
81 SmallPtrSet<Argument*, 8> &ByValArgsToTransform);
82 /// The maximum number of elements to expand, or 0 for unlimited.
87 char ArgPromotion::ID = 0;
88 INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion",
89 "Promote 'by reference' arguments to scalars", false, false)
90 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
91 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
92 INITIALIZE_PASS_END(ArgPromotion, "argpromotion",
93 "Promote 'by reference' arguments to scalars", false, false)
95 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
96 return new ArgPromotion(maxElements);
99 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
100 bool Changed = false, LocalChange;
102 do { // Iterate until we stop promoting from this SCC.
104 // Attempt to promote arguments from all functions in this SCC.
105 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
106 if (CallGraphNode *CGN = PromoteArguments(*I)) {
108 SCC.ReplaceNode(*I, CGN);
111 Changed |= LocalChange; // Remember that we changed something.
112 } while (LocalChange);
117 /// PromoteArguments - This method checks the specified function to see if there
118 /// are any promotable arguments and if it is safe to promote the function (for
119 /// example, all callers are direct). If safe to promote some arguments, it
120 /// calls the DoPromotion method.
122 CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
123 Function *F = CGN->getFunction();
125 // Make sure that it is local to this module.
126 if (!F || !F->hasLocalLinkage()) return 0;
128 // First check: see if there are any pointer arguments! If not, quick exit.
129 SmallVector<Argument*, 16> PointerArgs;
130 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
131 if (I->getType()->isPointerTy())
132 PointerArgs.push_back(I);
133 if (PointerArgs.empty()) return 0;
135 // Second check: make sure that all callers are direct callers. We can't
136 // transform functions that have indirect callers. Also see if the function
137 // is self-recursive.
138 bool isSelfRecursive = false;
139 for (Value::use_iterator UI = F->use_begin(), E = F->use_end();
142 // Must be a direct call.
143 if (CS.getInstruction() == 0 || !CS.isCallee(UI)) return 0;
145 if (CS.getInstruction()->getParent()->getParent() == F)
146 isSelfRecursive = true;
149 // Check to see which arguments are promotable. If an argument is promotable,
150 // add it to ArgsToPromote.
151 SmallPtrSet<Argument*, 8> ArgsToPromote;
152 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
153 for (unsigned i = 0, e = PointerArgs.size(); i != e; ++i) {
154 Argument *PtrArg = PointerArgs[i];
155 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
157 // If this is a byval argument, and if the aggregate type is small, just
158 // pass the elements, which is always safe. This does not apply to
160 if (PtrArg->hasByValAttr()) {
161 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
162 if (maxElements > 0 && STy->getNumElements() > maxElements) {
163 DEBUG(dbgs() << "argpromotion disable promoting argument '"
164 << PtrArg->getName() << "' because it would require adding more"
165 << " than " << maxElements << " arguments to the function.\n");
169 // If all the elements are single-value types, we can promote it.
170 bool AllSimple = true;
171 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
172 if (!STy->getElementType(i)->isSingleValueType()) {
178 // Safe to transform, don't even bother trying to "promote" it.
179 // Passing the elements as a scalar will allow scalarrepl to hack on
180 // the new alloca we introduce.
182 ByValArgsToTransform.insert(PtrArg);
188 // If the argument is a recursive type and we're in a recursive
189 // function, we could end up infinitely peeling the function argument.
190 if (isSelfRecursive) {
191 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
192 bool RecursiveType = false;
193 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
194 if (STy->getElementType(i) == PtrArg->getType()) {
195 RecursiveType = true;
204 // Otherwise, see if we can promote the pointer to its value.
205 if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr()))
206 ArgsToPromote.insert(PtrArg);
209 // No promotable pointer arguments.
210 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
213 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
216 /// AllCallersPassInValidPointerForArgument - Return true if we can prove that
217 /// all callees pass in a valid pointer for the specified function argument.
218 static bool AllCallersPassInValidPointerForArgument(Argument *Arg) {
219 Function *Callee = Arg->getParent();
221 unsigned ArgNo = Arg->getArgNo();
223 // Look at all call sites of the function. At this pointer we know we only
224 // have direct callees.
225 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end();
228 assert(CS && "Should only have direct calls!");
230 if (!CS.getArgument(ArgNo)->isDereferenceablePointer())
236 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
237 /// that is greater than or equal to the size of prefix, and each of the
238 /// elements in Prefix is the same as the corresponding elements in Longer.
240 /// This means it also returns true when Prefix and Longer are equal!
241 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
242 const ArgPromotion::IndicesVector &Longer) {
243 if (Prefix.size() > Longer.size())
245 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin());
249 /// Checks if Indices, or a prefix of Indices, is in Set.
250 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
251 std::set<ArgPromotion::IndicesVector> &Set) {
252 std::set<ArgPromotion::IndicesVector>::iterator Low;
253 Low = Set.upper_bound(Indices);
254 if (Low != Set.begin())
256 // Low is now the last element smaller than or equal to Indices. This means
257 // it points to a prefix of Indices (possibly Indices itself), if such
260 // This load is safe if any prefix of its operands is safe to load.
261 return Low != Set.end() && IsPrefix(*Low, Indices);
264 /// Mark the given indices (ToMark) as safe in the given set of indices
265 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
266 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
267 /// already. Furthermore, any indices that Indices is itself a prefix of, are
268 /// removed from Safe (since they are implicitely safe because of Indices now).
269 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
270 std::set<ArgPromotion::IndicesVector> &Safe) {
271 std::set<ArgPromotion::IndicesVector>::iterator Low;
272 Low = Safe.upper_bound(ToMark);
273 // Guard against the case where Safe is empty
274 if (Low != Safe.begin())
276 // Low is now the last element smaller than or equal to Indices. This
277 // means it points to a prefix of Indices (possibly Indices itself), if
278 // such prefix exists.
279 if (Low != Safe.end()) {
280 if (IsPrefix(*Low, ToMark))
281 // If there is already a prefix of these indices (or exactly these
282 // indices) marked a safe, don't bother adding these indices
285 // Increment Low, so we can use it as a "insert before" hint
289 Low = Safe.insert(Low, ToMark);
291 // If there we're a prefix of longer index list(s), remove those
292 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
293 while (Low != End && IsPrefix(ToMark, *Low)) {
294 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
300 /// isSafeToPromoteArgument - As you might guess from the name of this method,
301 /// it checks to see if it is both safe and useful to promote the argument.
302 /// This method limits promotion of aggregates to only promote up to three
303 /// elements of the aggregate in order to avoid exploding the number of
304 /// arguments passed in.
305 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg,
306 bool isByValOrInAlloca) const {
307 typedef std::set<IndicesVector> GEPIndicesSet;
309 // Quick exit for unused arguments
310 if (Arg->use_empty())
313 // We can only promote this argument if all of the uses are loads, or are GEP
314 // instructions (with constant indices) that are subsequently loaded.
316 // Promoting the argument causes it to be loaded in the caller
317 // unconditionally. This is only safe if we can prove that either the load
318 // would have happened in the callee anyway (ie, there is a load in the entry
319 // block) or the pointer passed in at every call site is guaranteed to be
321 // In the former case, invalid loads can happen, but would have happened
322 // anyway, in the latter case, invalid loads won't happen. This prevents us
323 // from introducing an invalid load that wouldn't have happened in the
326 // This set will contain all sets of indices that are loaded in the entry
327 // block, and thus are safe to unconditionally load in the caller.
329 // This optimization is also safe for InAlloca parameters, because it verifies
330 // that the address isn't captured.
331 GEPIndicesSet SafeToUnconditionallyLoad;
333 // This set contains all the sets of indices that we are planning to promote.
334 // This makes it possible to limit the number of arguments added.
335 GEPIndicesSet ToPromote;
337 // If the pointer is always valid, any load with first index 0 is valid.
338 if (isByValOrInAlloca || AllCallersPassInValidPointerForArgument(Arg))
339 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
341 // First, iterate the entry block and mark loads of (geps of) arguments as
343 BasicBlock *EntryBlock = Arg->getParent()->begin();
344 // Declare this here so we can reuse it
345 IndicesVector Indices;
346 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
348 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
349 Value *V = LI->getPointerOperand();
350 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
351 V = GEP->getPointerOperand();
353 // This load actually loads (part of) Arg? Check the indices then.
354 Indices.reserve(GEP->getNumIndices());
355 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
357 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
358 Indices.push_back(CI->getSExtValue());
360 // We found a non-constant GEP index for this argument? Bail out
361 // right away, can't promote this argument at all.
364 // Indices checked out, mark them as safe
365 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
368 } else if (V == Arg) {
369 // Direct loads are equivalent to a GEP with a single 0 index.
370 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
374 // Now, iterate all uses of the argument to see if there are any uses that are
375 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
376 SmallVector<LoadInst*, 16> Loads;
377 IndicesVector Operands;
378 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
382 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
383 // Don't hack volatile/atomic loads
384 if (!LI->isSimple()) return false;
386 // Direct loads are equivalent to a GEP with a zero index and then a load.
387 Operands.push_back(0);
388 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
389 if (GEP->use_empty()) {
390 // Dead GEP's cause trouble later. Just remove them if we run into
392 getAnalysis<AliasAnalysis>().deleteValue(GEP);
393 GEP->eraseFromParent();
394 // TODO: This runs the above loop over and over again for dead GEPs
395 // Couldn't we just do increment the UI iterator earlier and erase the
397 return isSafeToPromoteArgument(Arg, isByValOrInAlloca);
400 // Ensure that all of the indices are constants.
401 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
403 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
404 Operands.push_back(C->getSExtValue());
406 return false; // Not a constant operand GEP!
408 // Ensure that the only users of the GEP are load instructions.
409 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
411 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
412 // Don't hack volatile/atomic loads
413 if (!LI->isSimple()) return false;
416 // Other uses than load?
420 return false; // Not a load or a GEP.
423 // Now, see if it is safe to promote this load / loads of this GEP. Loading
424 // is safe if Operands, or a prefix of Operands, is marked as safe.
425 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
428 // See if we are already promoting a load with these indices. If not, check
429 // to make sure that we aren't promoting too many elements. If so, nothing
431 if (ToPromote.find(Operands) == ToPromote.end()) {
432 if (maxElements > 0 && ToPromote.size() == maxElements) {
433 DEBUG(dbgs() << "argpromotion not promoting argument '"
434 << Arg->getName() << "' because it would require adding more "
435 << "than " << maxElements << " arguments to the function.\n");
436 // We limit aggregate promotion to only promoting up to a fixed number
437 // of elements of the aggregate.
440 ToPromote.insert(Operands);
444 if (Loads.empty()) return true; // No users, this is a dead argument.
446 // Okay, now we know that the argument is only used by load instructions and
447 // it is safe to unconditionally perform all of them. Use alias analysis to
448 // check to see if the pointer is guaranteed to not be modified from entry of
449 // the function to each of the load instructions.
451 // Because there could be several/many load instructions, remember which
452 // blocks we know to be transparent to the load.
453 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
455 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
457 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
458 // Check to see if the load is invalidated from the start of the block to
460 LoadInst *Load = Loads[i];
461 BasicBlock *BB = Load->getParent();
463 AliasAnalysis::Location Loc = AA.getLocation(Load);
464 if (AA.canInstructionRangeModify(BB->front(), *Load, Loc))
465 return false; // Pointer is invalidated!
467 // Now check every path from the entry block to the load for transparency.
468 // To do this, we perform a depth first search on the inverse CFG from the
470 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
472 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> >
473 I = idf_ext_begin(P, TranspBlocks),
474 E = idf_ext_end(P, TranspBlocks); I != E; ++I)
475 if (AA.canBasicBlockModify(**I, Loc))
480 // If the path from the entry of the function to each load is free of
481 // instructions that potentially invalidate the load, we can make the
486 /// DoPromotion - This method actually performs the promotion of the specified
487 /// arguments, and returns the new function. At this point, we know that it's
489 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
490 SmallPtrSet<Argument*, 8> &ArgsToPromote,
491 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {
493 // Start by computing a new prototype for the function, which is the same as
494 // the old function, but has modified arguments.
495 FunctionType *FTy = F->getFunctionType();
496 std::vector<Type*> Params;
498 typedef std::set<IndicesVector> ScalarizeTable;
500 // ScalarizedElements - If we are promoting a pointer that has elements
501 // accessed out of it, keep track of which elements are accessed so that we
502 // can add one argument for each.
504 // Arguments that are directly loaded will have a zero element value here, to
505 // handle cases where there are both a direct load and GEP accesses.
507 std::map<Argument*, ScalarizeTable> ScalarizedElements;
509 // OriginalLoads - Keep track of a representative load instruction from the
510 // original function so that we can tell the alias analysis implementation
511 // what the new GEP/Load instructions we are inserting look like.
512 // We need to keep the original loads for each argument and the elements
513 // of the argument that are accessed.
514 std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;
516 // Attribute - Keep track of the parameter attributes for the arguments
517 // that we are *not* promoting. For the ones that we do promote, the parameter
518 // attributes are lost
519 SmallVector<AttributeSet, 8> AttributesVec;
520 const AttributeSet &PAL = F->getAttributes();
522 // Add any return attributes.
523 if (PAL.hasAttributes(AttributeSet::ReturnIndex))
524 AttributesVec.push_back(AttributeSet::get(F->getContext(),
525 PAL.getRetAttributes()));
527 // First, determine the new argument list
528 unsigned ArgIndex = 1;
529 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
531 if (ByValArgsToTransform.count(I)) {
532 // Simple byval argument? Just add all the struct element types.
533 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
534 StructType *STy = cast<StructType>(AgTy);
535 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
536 Params.push_back(STy->getElementType(i));
537 ++NumByValArgsPromoted;
538 } else if (!ArgsToPromote.count(I)) {
539 // Unchanged argument
540 Params.push_back(I->getType());
541 AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
542 if (attrs.hasAttributes(ArgIndex)) {
543 AttrBuilder B(attrs, ArgIndex);
545 push_back(AttributeSet::get(F->getContext(), Params.size(), B));
547 } else if (I->use_empty()) {
548 // Dead argument (which are always marked as promotable)
551 // Okay, this is being promoted. This means that the only uses are loads
552 // or GEPs which are only used by loads
554 // In this table, we will track which indices are loaded from the argument
555 // (where direct loads are tracked as no indices).
556 ScalarizeTable &ArgIndices = ScalarizedElements[I];
557 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
559 Instruction *User = cast<Instruction>(*UI);
560 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
561 IndicesVector Indices;
562 Indices.reserve(User->getNumOperands() - 1);
563 // Since loads will only have a single operand, and GEPs only a single
564 // non-index operand, this will record direct loads without any indices,
565 // and gep+loads with the GEP indices.
566 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
568 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
569 // GEPs with a single 0 index can be merged with direct loads
570 if (Indices.size() == 1 && Indices.front() == 0)
572 ArgIndices.insert(Indices);
574 if (LoadInst *L = dyn_cast<LoadInst>(User))
577 // Take any load, we will use it only to update Alias Analysis
578 OrigLoad = cast<LoadInst>(User->use_back());
579 OriginalLoads[std::make_pair(I, Indices)] = OrigLoad;
582 // Add a parameter to the function for each element passed in.
583 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
584 E = ArgIndices.end(); SI != E; ++SI) {
585 // not allowed to dereference ->begin() if size() is 0
586 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(), *SI));
587 assert(Params.back());
590 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
591 ++NumArgumentsPromoted;
593 ++NumAggregatesPromoted;
597 // Add any function attributes.
598 if (PAL.hasAttributes(AttributeSet::FunctionIndex))
599 AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
600 PAL.getFnAttributes()));
602 Type *RetTy = FTy->getReturnType();
604 // Construct the new function type using the new arguments.
605 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
607 // Create the new function body and insert it into the module.
608 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
609 NF->copyAttributesFrom(F);
612 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
615 // Recompute the parameter attributes list based on the new arguments for
617 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
618 AttributesVec.clear();
620 F->getParent()->getFunctionList().insert(F, NF);
623 // Get the alias analysis information that we need to update to reflect our
625 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
627 // Get the callgraph information that we need to update to reflect our
629 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
631 // Get a new callgraph node for NF.
632 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
634 // Loop over all of the callers of the function, transforming the call sites
635 // to pass in the loaded pointers.
637 SmallVector<Value*, 16> Args;
638 while (!F->use_empty()) {
639 CallSite CS(F->use_back());
640 assert(CS.getCalledFunction() == F);
641 Instruction *Call = CS.getInstruction();
642 const AttributeSet &CallPAL = CS.getAttributes();
644 // Add any return attributes.
645 if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
646 AttributesVec.push_back(AttributeSet::get(F->getContext(),
647 CallPAL.getRetAttributes()));
649 // Loop over the operands, inserting GEP and loads in the caller as
651 CallSite::arg_iterator AI = CS.arg_begin();
653 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
654 I != E; ++I, ++AI, ++ArgIndex)
655 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
656 Args.push_back(*AI); // Unmodified argument
658 if (CallPAL.hasAttributes(ArgIndex)) {
659 AttrBuilder B(CallPAL, ArgIndex);
661 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
663 } else if (ByValArgsToTransform.count(I)) {
664 // Emit a GEP and load for each element of the struct.
665 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
666 StructType *STy = cast<StructType>(AgTy);
668 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
669 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
670 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
671 Value *Idx = GetElementPtrInst::Create(*AI, Idxs,
672 (*AI)->getName()+"."+utostr(i),
674 // TODO: Tell AA about the new values?
675 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
677 } else if (!I->use_empty()) {
678 // Non-dead argument: insert GEPs and loads as appropriate.
679 ScalarizeTable &ArgIndices = ScalarizedElements[I];
680 // Store the Value* version of the indices in here, but declare it now
682 std::vector<Value*> Ops;
683 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
684 E = ArgIndices.end(); SI != E; ++SI) {
686 LoadInst *OrigLoad = OriginalLoads[std::make_pair(I, *SI)];
688 Ops.reserve(SI->size());
689 Type *ElTy = V->getType();
690 for (IndicesVector::const_iterator II = SI->begin(),
691 IE = SI->end(); II != IE; ++II) {
692 // Use i32 to index structs, and i64 for others (pointers/arrays).
693 // This satisfies GEP constraints.
694 Type *IdxTy = (ElTy->isStructTy() ?
695 Type::getInt32Ty(F->getContext()) :
696 Type::getInt64Ty(F->getContext()));
697 Ops.push_back(ConstantInt::get(IdxTy, *II));
698 // Keep track of the type we're currently indexing.
699 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
701 // And create a GEP to extract those indices.
702 V = GetElementPtrInst::Create(V, Ops, V->getName()+".idx", Call);
704 AA.copyValue(OrigLoad->getOperand(0), V);
706 // Since we're replacing a load make sure we take the alignment
707 // of the previous load.
708 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
709 newLoad->setAlignment(OrigLoad->getAlignment());
710 // Transfer the TBAA info too.
711 newLoad->setMetadata(LLVMContext::MD_tbaa,
712 OrigLoad->getMetadata(LLVMContext::MD_tbaa));
713 Args.push_back(newLoad);
714 AA.copyValue(OrigLoad, Args.back());
718 // Push any varargs arguments on the list.
719 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
721 if (CallPAL.hasAttributes(ArgIndex)) {
722 AttrBuilder B(CallPAL, ArgIndex);
724 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
728 // Add any function attributes.
729 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
730 AttributesVec.push_back(AttributeSet::get(Call->getContext(),
731 CallPAL.getFnAttributes()));
734 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
735 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
737 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
738 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
741 New = CallInst::Create(NF, Args, "", Call);
742 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
743 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
745 if (cast<CallInst>(Call)->isTailCall())
746 cast<CallInst>(New)->setTailCall();
749 AttributesVec.clear();
751 // Update the alias analysis implementation to know that we are replacing
752 // the old call with a new one.
753 AA.replaceWithNewValue(Call, New);
755 // Update the callgraph to know that the callsite has been transformed.
756 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
757 CalleeNode->replaceCallEdge(Call, New, NF_CGN);
759 if (!Call->use_empty()) {
760 Call->replaceAllUsesWith(New);
764 // Finally, remove the old call from the program, reducing the use-count of
766 Call->eraseFromParent();
769 // Since we have now created the new function, splice the body of the old
770 // function right into the new function, leaving the old rotting hulk of the
772 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
774 // Loop over the argument list, transferring uses of the old arguments over to
775 // the new arguments, also transferring over the names as well.
777 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
778 I2 = NF->arg_begin(); I != E; ++I) {
779 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
780 // If this is an unmodified argument, move the name and users over to the
782 I->replaceAllUsesWith(I2);
784 AA.replaceWithNewValue(I, I2);
789 if (ByValArgsToTransform.count(I)) {
790 // In the callee, we create an alloca, and store each of the new incoming
791 // arguments into the alloca.
792 Instruction *InsertPt = NF->begin()->begin();
794 // Just add all the struct element types.
795 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
796 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
797 StructType *STy = cast<StructType>(AgTy);
799 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
801 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
802 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
804 GetElementPtrInst::Create(TheAlloca, Idxs,
805 TheAlloca->getName()+"."+Twine(i),
807 I2->setName(I->getName()+"."+Twine(i));
808 new StoreInst(I2++, Idx, InsertPt);
811 // Anything that used the arg should now use the alloca.
812 I->replaceAllUsesWith(TheAlloca);
813 TheAlloca->takeName(I);
814 AA.replaceWithNewValue(I, TheAlloca);
816 // If the alloca is used in a call, we must clear the tail flag since
817 // the callee now uses an alloca from the caller.
818 for (Value::use_iterator UI = TheAlloca->use_begin(),
819 E = TheAlloca->use_end(); UI != E; ++UI) {
820 CallInst *Call = dyn_cast<CallInst>(*UI);
823 Call->setTailCall(false);
828 if (I->use_empty()) {
833 // Otherwise, if we promoted this argument, then all users are load
834 // instructions (or GEPs with only load users), and all loads should be
835 // using the new argument that we added.
836 ScalarizeTable &ArgIndices = ScalarizedElements[I];
838 while (!I->use_empty()) {
839 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
840 assert(ArgIndices.begin()->empty() &&
841 "Load element should sort to front!");
842 I2->setName(I->getName()+".val");
843 LI->replaceAllUsesWith(I2);
844 AA.replaceWithNewValue(LI, I2);
845 LI->eraseFromParent();
846 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
847 << "' in function '" << F->getName() << "'\n");
849 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
850 IndicesVector Operands;
851 Operands.reserve(GEP->getNumIndices());
852 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
854 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
856 // GEPs with a single 0 index can be merged with direct loads
857 if (Operands.size() == 1 && Operands.front() == 0)
860 Function::arg_iterator TheArg = I2;
861 for (ScalarizeTable::iterator It = ArgIndices.begin();
862 *It != Operands; ++It, ++TheArg) {
863 assert(It != ArgIndices.end() && "GEP not handled??");
866 std::string NewName = I->getName();
867 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
868 NewName += "." + utostr(Operands[i]);
871 TheArg->setName(NewName);
873 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
874 << "' of function '" << NF->getName() << "'\n");
876 // All of the uses must be load instructions. Replace them all with
877 // the argument specified by ArgNo.
878 while (!GEP->use_empty()) {
879 LoadInst *L = cast<LoadInst>(GEP->use_back());
880 L->replaceAllUsesWith(TheArg);
881 AA.replaceWithNewValue(L, TheArg);
882 L->eraseFromParent();
885 GEP->eraseFromParent();
889 // Increment I2 past all of the arguments added for this promoted pointer.
890 std::advance(I2, ArgIndices.size());
893 // Tell the alias analysis that the old function is about to disappear.
894 AA.replaceWithNewValue(F, NF);
897 NF_CGN->stealCalledFunctionsFrom(CG[F]);
899 // Now that the old function is dead, delete it. If there is a dangling
900 // reference to the CallgraphNode, just leave the dead function around for
901 // someone else to nuke.
902 CallGraphNode *CGN = CG[F];
903 if (CGN->getNumReferences() == 0)
904 delete CG.removeFunctionFromModule(CGN);
906 F->setLinkage(Function::ExternalLinkage);