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_PASS_END(ArgPromotion, "argpromotion",
91 "Promote 'by reference' arguments to scalars", false, false)
93 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
94 return new ArgPromotion(maxElements);
97 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
98 bool Changed = false, LocalChange;
100 do { // Iterate until we stop promoting from this SCC.
102 // Attempt to promote arguments from all functions in this SCC.
103 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
104 if (CallGraphNode *CGN = PromoteArguments(*I)) {
106 SCC.ReplaceNode(*I, CGN);
109 Changed |= LocalChange; // Remember that we changed something.
110 } while (LocalChange);
115 /// PromoteArguments - This method checks the specified function to see if there
116 /// are any promotable arguments and if it is safe to promote the function (for
117 /// example, all callers are direct). If safe to promote some arguments, it
118 /// calls the DoPromotion method.
120 CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
121 Function *F = CGN->getFunction();
123 // Make sure that it is local to this module.
124 if (!F || !F->hasLocalLinkage()) return 0;
126 // First check: see if there are any pointer arguments! If not, quick exit.
127 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs;
129 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
130 I != E; ++I, ++ArgNo)
131 if (I->getType()->isPointerTy())
132 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo));
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.
137 if (F->hasAddressTaken())
140 // Check to see which arguments are promotable. If an argument is promotable,
141 // add it to ArgsToPromote.
142 SmallPtrSet<Argument*, 8> ArgsToPromote;
143 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
144 for (unsigned i = 0; i != PointerArgs.size(); ++i) {
145 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal);
147 // If this is a byval argument, and if the aggregate type is small, just
148 // pass the elements, which is always safe.
149 Argument *PtrArg = PointerArgs[i].first;
151 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
152 if (const StructType *STy = dyn_cast<StructType>(AgTy)) {
153 if (maxElements > 0 && STy->getNumElements() > maxElements) {
154 DEBUG(dbgs() << "argpromotion disable promoting argument '"
155 << PtrArg->getName() << "' because it would require adding more"
156 << " than " << maxElements << " arguments to the function.\n");
158 // If all the elements are single-value types, we can promote it.
159 bool AllSimple = true;
160 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
161 if (!STy->getElementType(i)->isSingleValueType()) {
166 // Safe to transform, don't even bother trying to "promote" it.
167 // Passing the elements as a scalar will allow scalarrepl to hack on
168 // the new alloca we introduce.
170 ByValArgsToTransform.insert(PtrArg);
177 // Otherwise, see if we can promote the pointer to its value.
178 if (isSafeToPromoteArgument(PtrArg, isByVal))
179 ArgsToPromote.insert(PtrArg);
182 // No promotable pointer arguments.
183 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
186 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
189 /// IsAlwaysValidPointer - Return true if the specified pointer is always legal
191 static bool IsAlwaysValidPointer(Value *V) {
192 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
193 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V))
194 return IsAlwaysValidPointer(GEP->getOperand(0));
195 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
196 if (CE->getOpcode() == Instruction::GetElementPtr)
197 return IsAlwaysValidPointer(CE->getOperand(0));
202 /// AllCalleesPassInValidPointerForArgument - Return true if we can prove that
203 /// all callees pass in a valid pointer for the specified function argument.
204 static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) {
205 Function *Callee = Arg->getParent();
207 unsigned ArgNo = std::distance(Callee->arg_begin(),
208 Function::arg_iterator(Arg));
210 // Look at all call sites of the function. At this pointer we know we only
211 // have direct callees.
212 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end();
215 assert(CS && "Should only have direct calls!");
217 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo)))
223 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
224 /// that is greater than or equal to the size of prefix, and each of the
225 /// elements in Prefix is the same as the corresponding elements in Longer.
227 /// This means it also returns true when Prefix and Longer are equal!
228 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
229 const ArgPromotion::IndicesVector &Longer) {
230 if (Prefix.size() > Longer.size())
232 for (unsigned i = 0, e = Prefix.size(); i != e; ++i)
233 if (Prefix[i] != Longer[i])
239 /// Checks if Indices, or a prefix of Indices, is in Set.
240 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
241 std::set<ArgPromotion::IndicesVector> &Set) {
242 std::set<ArgPromotion::IndicesVector>::iterator Low;
243 Low = Set.upper_bound(Indices);
244 if (Low != Set.begin())
246 // Low is now the last element smaller than or equal to Indices. This means
247 // it points to a prefix of Indices (possibly Indices itself), if such
250 // This load is safe if any prefix of its operands is safe to load.
251 return Low != Set.end() && IsPrefix(*Low, Indices);
254 /// Mark the given indices (ToMark) as safe in the given set of indices
255 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
256 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
257 /// already. Furthermore, any indices that Indices is itself a prefix of, are
258 /// removed from Safe (since they are implicitely safe because of Indices now).
259 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
260 std::set<ArgPromotion::IndicesVector> &Safe) {
261 std::set<ArgPromotion::IndicesVector>::iterator Low;
262 Low = Safe.upper_bound(ToMark);
263 // Guard against the case where Safe is empty
264 if (Low != Safe.begin())
266 // Low is now the last element smaller than or equal to Indices. This
267 // means it points to a prefix of Indices (possibly Indices itself), if
268 // such prefix exists.
269 if (Low != Safe.end()) {
270 if (IsPrefix(*Low, ToMark))
271 // If there is already a prefix of these indices (or exactly these
272 // indices) marked a safe, don't bother adding these indices
275 // Increment Low, so we can use it as a "insert before" hint
279 Low = Safe.insert(Low, ToMark);
281 // If there we're a prefix of longer index list(s), remove those
282 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
283 while (Low != End && IsPrefix(ToMark, *Low)) {
284 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
290 /// isSafeToPromoteArgument - As you might guess from the name of this method,
291 /// it checks to see if it is both safe and useful to promote the argument.
292 /// This method limits promotion of aggregates to only promote up to three
293 /// elements of the aggregate in order to avoid exploding the number of
294 /// arguments passed in.
295 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const {
296 typedef std::set<IndicesVector> GEPIndicesSet;
298 // Quick exit for unused arguments
299 if (Arg->use_empty())
302 // We can only promote this argument if all of the uses are loads, or are GEP
303 // instructions (with constant indices) that are subsequently loaded.
305 // Promoting the argument causes it to be loaded in the caller
306 // unconditionally. This is only safe if we can prove that either the load
307 // would have happened in the callee anyway (ie, there is a load in the entry
308 // block) or the pointer passed in at every call site is guaranteed to be
310 // In the former case, invalid loads can happen, but would have happened
311 // anyway, in the latter case, invalid loads won't happen. This prevents us
312 // from introducing an invalid load that wouldn't have happened in the
315 // This set will contain all sets of indices that are loaded in the entry
316 // block, and thus are safe to unconditionally load in the caller.
317 GEPIndicesSet SafeToUnconditionallyLoad;
319 // This set contains all the sets of indices that we are planning to promote.
320 // This makes it possible to limit the number of arguments added.
321 GEPIndicesSet ToPromote;
323 // If the pointer is always valid, any load with first index 0 is valid.
324 if (isByVal || AllCalleesPassInValidPointerForArgument(Arg))
325 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
327 // First, iterate the entry block and mark loads of (geps of) arguments as
329 BasicBlock *EntryBlock = Arg->getParent()->begin();
330 // Declare this here so we can reuse it
331 IndicesVector Indices;
332 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
334 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
335 Value *V = LI->getPointerOperand();
336 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
337 V = GEP->getPointerOperand();
339 // This load actually loads (part of) Arg? Check the indices then.
340 Indices.reserve(GEP->getNumIndices());
341 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
343 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
344 Indices.push_back(CI->getSExtValue());
346 // We found a non-constant GEP index for this argument? Bail out
347 // right away, can't promote this argument at all.
350 // Indices checked out, mark them as safe
351 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
354 } else if (V == Arg) {
355 // Direct loads are equivalent to a GEP with a single 0 index.
356 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
360 // Now, iterate all uses of the argument to see if there are any uses that are
361 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
362 SmallVector<LoadInst*, 16> Loads;
363 IndicesVector Operands;
364 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
368 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
369 if (LI->isVolatile()) return false; // Don't hack volatile loads
371 // Direct loads are equivalent to a GEP with a zero index and then a load.
372 Operands.push_back(0);
373 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
374 if (GEP->use_empty()) {
375 // Dead GEP's cause trouble later. Just remove them if we run into
377 getAnalysis<AliasAnalysis>().deleteValue(GEP);
378 GEP->eraseFromParent();
379 // TODO: This runs the above loop over and over again for dead GEPs
380 // Couldn't we just do increment the UI iterator earlier and erase the
382 return isSafeToPromoteArgument(Arg, isByVal);
385 // Ensure that all of the indices are constants.
386 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
388 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
389 Operands.push_back(C->getSExtValue());
391 return false; // Not a constant operand GEP!
393 // Ensure that the only users of the GEP are load instructions.
394 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
396 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
397 if (LI->isVolatile()) return false; // Don't hack volatile loads
400 // Other uses than load?
404 return false; // Not a load or a GEP.
407 // Now, see if it is safe to promote this load / loads of this GEP. Loading
408 // is safe if Operands, or a prefix of Operands, is marked as safe.
409 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
412 // See if we are already promoting a load with these indices. If not, check
413 // to make sure that we aren't promoting too many elements. If so, nothing
415 if (ToPromote.find(Operands) == ToPromote.end()) {
416 if (maxElements > 0 && ToPromote.size() == maxElements) {
417 DEBUG(dbgs() << "argpromotion not promoting argument '"
418 << Arg->getName() << "' because it would require adding more "
419 << "than " << maxElements << " arguments to the function.\n");
420 // We limit aggregate promotion to only promoting up to a fixed number
421 // of elements of the aggregate.
424 ToPromote.insert(Operands);
428 if (Loads.empty()) return true; // No users, this is a dead argument.
430 // Okay, now we know that the argument is only used by load instructions and
431 // it is safe to unconditionally perform all of them. Use alias analysis to
432 // check to see if the pointer is guaranteed to not be modified from entry of
433 // the function to each of the load instructions.
435 // Because there could be several/many load instructions, remember which
436 // blocks we know to be transparent to the load.
437 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
439 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
440 TargetData *TD = getAnalysisIfAvailable<TargetData>();
441 if (!TD) return false; // Without TargetData, assume the worst.
443 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
444 // Check to see if the load is invalidated from the start of the block to
446 LoadInst *Load = Loads[i];
447 BasicBlock *BB = Load->getParent();
449 const PointerType *LoadTy =
450 cast<PointerType>(Load->getPointerOperand()->getType());
451 unsigned LoadSize =(unsigned)TD->getTypeStoreSize(LoadTy->getElementType());
453 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize))
454 return false; // Pointer is invalidated!
456 // Now check every path from the entry block to the load for transparency.
457 // To do this, we perform a depth first search on the inverse CFG from the
459 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI) {
461 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> >
462 I = idf_ext_begin(P, TranspBlocks),
463 E = idf_ext_end(P, TranspBlocks); I != E; ++I)
464 if (AA.canBasicBlockModify(**I, Arg, LoadSize))
469 // If the path from the entry of the function to each load is free of
470 // instructions that potentially invalidate the load, we can make the
475 /// DoPromotion - This method actually performs the promotion of the specified
476 /// arguments, and returns the new function. At this point, we know that it's
478 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
479 SmallPtrSet<Argument*, 8> &ArgsToPromote,
480 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {
482 // Start by computing a new prototype for the function, which is the same as
483 // the old function, but has modified arguments.
484 const FunctionType *FTy = F->getFunctionType();
485 std::vector<const Type*> Params;
487 typedef std::set<IndicesVector> ScalarizeTable;
489 // ScalarizedElements - If we are promoting a pointer that has elements
490 // accessed out of it, keep track of which elements are accessed so that we
491 // can add one argument for each.
493 // Arguments that are directly loaded will have a zero element value here, to
494 // handle cases where there are both a direct load and GEP accesses.
496 std::map<Argument*, ScalarizeTable> ScalarizedElements;
498 // OriginalLoads - Keep track of a representative load instruction from the
499 // original function so that we can tell the alias analysis implementation
500 // what the new GEP/Load instructions we are inserting look like.
501 std::map<IndicesVector, LoadInst*> OriginalLoads;
503 // Attributes - Keep track of the parameter attributes for the arguments
504 // that we are *not* promoting. For the ones that we do promote, the parameter
505 // attributes are lost
506 SmallVector<AttributeWithIndex, 8> AttributesVec;
507 const AttrListPtr &PAL = F->getAttributes();
509 // Add any return attributes.
510 if (Attributes attrs = PAL.getRetAttributes())
511 AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
513 // First, determine the new argument list
514 unsigned ArgIndex = 1;
515 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
517 if (ByValArgsToTransform.count(I)) {
518 // Simple byval argument? Just add all the struct element types.
519 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
520 const StructType *STy = cast<StructType>(AgTy);
521 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
522 Params.push_back(STy->getElementType(i));
523 ++NumByValArgsPromoted;
524 } else if (!ArgsToPromote.count(I)) {
525 // Unchanged argument
526 Params.push_back(I->getType());
527 if (Attributes attrs = PAL.getParamAttributes(ArgIndex))
528 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs));
529 } else if (I->use_empty()) {
530 // Dead argument (which are always marked as promotable)
533 // Okay, this is being promoted. This means that the only uses are loads
534 // or GEPs which are only used by loads
536 // In this table, we will track which indices are loaded from the argument
537 // (where direct loads are tracked as no indices).
538 ScalarizeTable &ArgIndices = ScalarizedElements[I];
539 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
541 Instruction *User = cast<Instruction>(*UI);
542 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
543 IndicesVector Indices;
544 Indices.reserve(User->getNumOperands() - 1);
545 // Since loads will only have a single operand, and GEPs only a single
546 // non-index operand, this will record direct loads without any indices,
547 // and gep+loads with the GEP indices.
548 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
550 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
551 // GEPs with a single 0 index can be merged with direct loads
552 if (Indices.size() == 1 && Indices.front() == 0)
554 ArgIndices.insert(Indices);
556 if (LoadInst *L = dyn_cast<LoadInst>(User))
559 // Take any load, we will use it only to update Alias Analysis
560 OrigLoad = cast<LoadInst>(User->use_back());
561 OriginalLoads[Indices] = OrigLoad;
564 // Add a parameter to the function for each element passed in.
565 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
566 E = ArgIndices.end(); SI != E; ++SI) {
567 // not allowed to dereference ->begin() if size() is 0
568 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(),
571 assert(Params.back());
574 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
575 ++NumArgumentsPromoted;
577 ++NumAggregatesPromoted;
581 // Add any function attributes.
582 if (Attributes attrs = PAL.getFnAttributes())
583 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
585 const Type *RetTy = FTy->getReturnType();
587 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
588 // have zero fixed arguments.
589 bool ExtraArgHack = false;
590 if (Params.empty() && FTy->isVarArg()) {
592 Params.push_back(Type::getInt32Ty(F->getContext()));
595 // Construct the new function type using the new arguments.
596 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
598 // Create the new function body and insert it into the module.
599 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
600 NF->copyAttributesFrom(F);
603 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
606 // Recompute the parameter attributes list based on the new arguments for
608 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(),
609 AttributesVec.end()));
610 AttributesVec.clear();
612 F->getParent()->getFunctionList().insert(F, NF);
615 // Get the alias analysis information that we need to update to reflect our
617 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
619 // Get the callgraph information that we need to update to reflect our
621 CallGraph &CG = getAnalysis<CallGraph>();
623 // Get a new callgraph node for NF.
624 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
626 // Loop over all of the callers of the function, transforming the call sites
627 // to pass in the loaded pointers.
629 SmallVector<Value*, 16> Args;
630 while (!F->use_empty()) {
631 CallSite CS(F->use_back());
632 assert(CS.getCalledFunction() == F);
633 Instruction *Call = CS.getInstruction();
634 const AttrListPtr &CallPAL = CS.getAttributes();
636 // Add any return attributes.
637 if (Attributes attrs = CallPAL.getRetAttributes())
638 AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
640 // Loop over the operands, inserting GEP and loads in the caller as
642 CallSite::arg_iterator AI = CS.arg_begin();
644 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
645 I != E; ++I, ++AI, ++ArgIndex)
646 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
647 Args.push_back(*AI); // Unmodified argument
649 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
650 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
652 } else if (ByValArgsToTransform.count(I)) {
653 // Emit a GEP and load for each element of the struct.
654 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
655 const StructType *STy = cast<StructType>(AgTy);
657 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
658 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
659 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
660 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2,
661 (*AI)->getName()+"."+utostr(i),
663 // TODO: Tell AA about the new values?
664 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
666 } else if (!I->use_empty()) {
667 // Non-dead argument: insert GEPs and loads as appropriate.
668 ScalarizeTable &ArgIndices = ScalarizedElements[I];
669 // Store the Value* version of the indices in here, but declare it now
671 std::vector<Value*> Ops;
672 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
673 E = ArgIndices.end(); SI != E; ++SI) {
675 LoadInst *OrigLoad = OriginalLoads[*SI];
677 Ops.reserve(SI->size());
678 const Type *ElTy = V->getType();
679 for (IndicesVector::const_iterator II = SI->begin(),
680 IE = SI->end(); II != IE; ++II) {
681 // Use i32 to index structs, and i64 for others (pointers/arrays).
682 // This satisfies GEP constraints.
683 const Type *IdxTy = (ElTy->isStructTy() ?
684 Type::getInt32Ty(F->getContext()) :
685 Type::getInt64Ty(F->getContext()));
686 Ops.push_back(ConstantInt::get(IdxTy, *II));
687 // Keep track of the type we're currently indexing.
688 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
690 // And create a GEP to extract those indices.
691 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(),
692 V->getName()+".idx", Call);
694 AA.copyValue(OrigLoad->getOperand(0), V);
696 // Since we're replacing a load make sure we take the alignment
697 // of the previous load.
698 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
699 newLoad->setAlignment(OrigLoad->getAlignment());
700 Args.push_back(newLoad);
701 AA.copyValue(OrigLoad, Args.back());
706 Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext())));
708 // Push any varargs arguments on the list.
709 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
711 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
712 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
715 // Add any function attributes.
716 if (Attributes attrs = CallPAL.getFnAttributes())
717 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
720 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
721 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
722 Args.begin(), Args.end(), "", Call);
723 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
724 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
725 AttributesVec.end()));
727 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
728 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
729 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
730 AttributesVec.end()));
731 if (cast<CallInst>(Call)->isTailCall())
732 cast<CallInst>(New)->setTailCall();
735 AttributesVec.clear();
737 // Update the alias analysis implementation to know that we are replacing
738 // the old call with a new one.
739 AA.replaceWithNewValue(Call, New);
741 // Update the callgraph to know that the callsite has been transformed.
742 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
743 CalleeNode->replaceCallEdge(Call, New, NF_CGN);
745 if (!Call->use_empty()) {
746 Call->replaceAllUsesWith(New);
750 // Finally, remove the old call from the program, reducing the use-count of
752 Call->eraseFromParent();
755 // Since we have now created the new function, splice the body of the old
756 // function right into the new function, leaving the old rotting hulk of the
758 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
760 // Loop over the argument list, transfering uses of the old arguments over to
761 // the new arguments, also transfering over the names as well.
763 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
764 I2 = NF->arg_begin(); I != E; ++I) {
765 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
766 // If this is an unmodified argument, move the name and users over to the
768 I->replaceAllUsesWith(I2);
770 AA.replaceWithNewValue(I, I2);
775 if (ByValArgsToTransform.count(I)) {
776 // In the callee, we create an alloca, and store each of the new incoming
777 // arguments into the alloca.
778 Instruction *InsertPt = NF->begin()->begin();
780 // Just add all the struct element types.
781 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
782 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
783 const StructType *STy = cast<StructType>(AgTy);
785 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
787 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
788 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
790 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2,
791 TheAlloca->getName()+"."+Twine(i),
793 I2->setName(I->getName()+"."+Twine(i));
794 new StoreInst(I2++, Idx, InsertPt);
797 // Anything that used the arg should now use the alloca.
798 I->replaceAllUsesWith(TheAlloca);
799 TheAlloca->takeName(I);
800 AA.replaceWithNewValue(I, TheAlloca);
804 if (I->use_empty()) {
809 // Otherwise, if we promoted this argument, then all users are load
810 // instructions (or GEPs with only load users), and all loads should be
811 // using the new argument that we added.
812 ScalarizeTable &ArgIndices = ScalarizedElements[I];
814 while (!I->use_empty()) {
815 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
816 assert(ArgIndices.begin()->empty() &&
817 "Load element should sort to front!");
818 I2->setName(I->getName()+".val");
819 LI->replaceAllUsesWith(I2);
820 AA.replaceWithNewValue(LI, I2);
821 LI->eraseFromParent();
822 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
823 << "' in function '" << F->getName() << "'\n");
825 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
826 IndicesVector Operands;
827 Operands.reserve(GEP->getNumIndices());
828 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
830 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
832 // GEPs with a single 0 index can be merged with direct loads
833 if (Operands.size() == 1 && Operands.front() == 0)
836 Function::arg_iterator TheArg = I2;
837 for (ScalarizeTable::iterator It = ArgIndices.begin();
838 *It != Operands; ++It, ++TheArg) {
839 assert(It != ArgIndices.end() && "GEP not handled??");
842 std::string NewName = I->getName();
843 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
844 NewName += "." + utostr(Operands[i]);
847 TheArg->setName(NewName);
849 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
850 << "' of function '" << NF->getName() << "'\n");
852 // All of the uses must be load instructions. Replace them all with
853 // the argument specified by ArgNo.
854 while (!GEP->use_empty()) {
855 LoadInst *L = cast<LoadInst>(GEP->use_back());
856 L->replaceAllUsesWith(TheArg);
857 AA.replaceWithNewValue(L, TheArg);
858 L->eraseFromParent();
861 GEP->eraseFromParent();
865 // Increment I2 past all of the arguments added for this promoted pointer.
866 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
870 // Notify the alias analysis implementation that we inserted a new argument.
872 AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())),
876 // Tell the alias analysis that the old function is about to disappear.
877 AA.replaceWithNewValue(F, NF);
880 NF_CGN->stealCalledFunctionsFrom(CG[F]);
882 // Now that the old function is dead, delete it. If there is a dangling
883 // reference to the CallgraphNode, just leave the dead function around for
884 // someone else to nuke.
885 CallGraphNode *CGN = CG[F];
886 if (CGN->getNumReferences() == 0)
887 delete CG.removeFunctionFromModule(CGN);
889 F->setLinkage(Function::ExternalLinkage);