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/Support/CallSite.h"
43 #include "llvm/Support/CFG.h"
44 #include "llvm/Support/Debug.h"
45 #include "llvm/Support/raw_ostream.h"
46 #include "llvm/ADT/DepthFirstIterator.h"
47 #include "llvm/ADT/Statistic.h"
48 #include "llvm/ADT/StringExtras.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_AG_DEPENDENCY(CallGraph)
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<std::pair<Argument*, unsigned>, 16> PointerArgs;
131 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
132 I != E; ++I, ++ArgNo)
133 if (I->getType()->isPointerTy())
134 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo));
135 if (PointerArgs.empty()) return 0;
137 // Second check: make sure that all callers are direct callers. We can't
138 // transform functions that have indirect callers.
139 if (F->hasAddressTaken())
142 // Check to see which arguments are promotable. If an argument is promotable,
143 // add it to ArgsToPromote.
144 SmallPtrSet<Argument*, 8> ArgsToPromote;
145 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
146 for (unsigned i = 0; i != PointerArgs.size(); ++i) {
147 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal);
149 // If this is a byval argument, and if the aggregate type is small, just
150 // pass the elements, which is always safe.
151 Argument *PtrArg = PointerArgs[i].first;
153 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
154 if (const StructType *STy = dyn_cast<StructType>(AgTy)) {
155 if (maxElements > 0 && STy->getNumElements() > maxElements) {
156 DEBUG(dbgs() << "argpromotion disable promoting argument '"
157 << PtrArg->getName() << "' because it would require adding more"
158 << " than " << maxElements << " arguments to the function.\n");
160 // If all the elements are single-value types, we can promote it.
161 bool AllSimple = true;
162 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
163 if (!STy->getElementType(i)->isSingleValueType()) {
168 // Safe to transform, don't even bother trying to "promote" it.
169 // Passing the elements as a scalar will allow scalarrepl to hack on
170 // the new alloca we introduce.
172 ByValArgsToTransform.insert(PtrArg);
179 // Otherwise, see if we can promote the pointer to its value.
180 if (isSafeToPromoteArgument(PtrArg, isByVal))
181 ArgsToPromote.insert(PtrArg);
184 // No promotable pointer arguments.
185 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
188 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
191 /// IsAlwaysValidPointer - Return true if the specified pointer is always legal
193 static bool IsAlwaysValidPointer(Value *V) {
194 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
195 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V))
196 return IsAlwaysValidPointer(GEP->getOperand(0));
197 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
198 if (CE->getOpcode() == Instruction::GetElementPtr)
199 return IsAlwaysValidPointer(CE->getOperand(0));
204 /// AllCalleesPassInValidPointerForArgument - Return true if we can prove that
205 /// all callees pass in a valid pointer for the specified function argument.
206 static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) {
207 Function *Callee = Arg->getParent();
209 unsigned ArgNo = std::distance(Callee->arg_begin(),
210 Function::arg_iterator(Arg));
212 // Look at all call sites of the function. At this pointer we know we only
213 // have direct callees.
214 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end();
217 assert(CS && "Should only have direct calls!");
219 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo)))
225 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
226 /// that is greater than or equal to the size of prefix, and each of the
227 /// elements in Prefix is the same as the corresponding elements in Longer.
229 /// This means it also returns true when Prefix and Longer are equal!
230 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
231 const ArgPromotion::IndicesVector &Longer) {
232 if (Prefix.size() > Longer.size())
234 for (unsigned i = 0, e = Prefix.size(); i != e; ++i)
235 if (Prefix[i] != Longer[i])
241 /// Checks if Indices, or a prefix of Indices, is in Set.
242 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
243 std::set<ArgPromotion::IndicesVector> &Set) {
244 std::set<ArgPromotion::IndicesVector>::iterator Low;
245 Low = Set.upper_bound(Indices);
246 if (Low != Set.begin())
248 // Low is now the last element smaller than or equal to Indices. This means
249 // it points to a prefix of Indices (possibly Indices itself), if such
252 // This load is safe if any prefix of its operands is safe to load.
253 return Low != Set.end() && IsPrefix(*Low, Indices);
256 /// Mark the given indices (ToMark) as safe in the given set of indices
257 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
258 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
259 /// already. Furthermore, any indices that Indices is itself a prefix of, are
260 /// removed from Safe (since they are implicitely safe because of Indices now).
261 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
262 std::set<ArgPromotion::IndicesVector> &Safe) {
263 std::set<ArgPromotion::IndicesVector>::iterator Low;
264 Low = Safe.upper_bound(ToMark);
265 // Guard against the case where Safe is empty
266 if (Low != Safe.begin())
268 // Low is now the last element smaller than or equal to Indices. This
269 // means it points to a prefix of Indices (possibly Indices itself), if
270 // such prefix exists.
271 if (Low != Safe.end()) {
272 if (IsPrefix(*Low, ToMark))
273 // If there is already a prefix of these indices (or exactly these
274 // indices) marked a safe, don't bother adding these indices
277 // Increment Low, so we can use it as a "insert before" hint
281 Low = Safe.insert(Low, ToMark);
283 // If there we're a prefix of longer index list(s), remove those
284 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
285 while (Low != End && IsPrefix(ToMark, *Low)) {
286 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
292 /// isSafeToPromoteArgument - As you might guess from the name of this method,
293 /// it checks to see if it is both safe and useful to promote the argument.
294 /// This method limits promotion of aggregates to only promote up to three
295 /// elements of the aggregate in order to avoid exploding the number of
296 /// arguments passed in.
297 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const {
298 typedef std::set<IndicesVector> GEPIndicesSet;
300 // Quick exit for unused arguments
301 if (Arg->use_empty())
304 // We can only promote this argument if all of the uses are loads, or are GEP
305 // instructions (with constant indices) that are subsequently loaded.
307 // Promoting the argument causes it to be loaded in the caller
308 // unconditionally. This is only safe if we can prove that either the load
309 // would have happened in the callee anyway (ie, there is a load in the entry
310 // block) or the pointer passed in at every call site is guaranteed to be
312 // In the former case, invalid loads can happen, but would have happened
313 // anyway, in the latter case, invalid loads won't happen. This prevents us
314 // from introducing an invalid load that wouldn't have happened in the
317 // This set will contain all sets of indices that are loaded in the entry
318 // block, and thus are safe to unconditionally load in the caller.
319 GEPIndicesSet SafeToUnconditionallyLoad;
321 // This set contains all the sets of indices that we are planning to promote.
322 // This makes it possible to limit the number of arguments added.
323 GEPIndicesSet ToPromote;
325 // If the pointer is always valid, any load with first index 0 is valid.
326 if (isByVal || AllCalleesPassInValidPointerForArgument(Arg))
327 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
329 // First, iterate the entry block and mark loads of (geps of) arguments as
331 BasicBlock *EntryBlock = Arg->getParent()->begin();
332 // Declare this here so we can reuse it
333 IndicesVector Indices;
334 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
336 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
337 Value *V = LI->getPointerOperand();
338 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
339 V = GEP->getPointerOperand();
341 // This load actually loads (part of) Arg? Check the indices then.
342 Indices.reserve(GEP->getNumIndices());
343 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
345 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
346 Indices.push_back(CI->getSExtValue());
348 // We found a non-constant GEP index for this argument? Bail out
349 // right away, can't promote this argument at all.
352 // Indices checked out, mark them as safe
353 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
356 } else if (V == Arg) {
357 // Direct loads are equivalent to a GEP with a single 0 index.
358 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
362 // Now, iterate all uses of the argument to see if there are any uses that are
363 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
364 SmallVector<LoadInst*, 16> Loads;
365 IndicesVector Operands;
366 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
370 if (LoadInst *LI = dyn_cast<LoadInst>(U)) {
371 if (LI->isVolatile()) return false; // Don't hack volatile loads
373 // Direct loads are equivalent to a GEP with a zero index and then a load.
374 Operands.push_back(0);
375 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(U)) {
376 if (GEP->use_empty()) {
377 // Dead GEP's cause trouble later. Just remove them if we run into
379 getAnalysis<AliasAnalysis>().deleteValue(GEP);
380 GEP->eraseFromParent();
381 // TODO: This runs the above loop over and over again for dead GEPs
382 // Couldn't we just do increment the UI iterator earlier and erase the
384 return isSafeToPromoteArgument(Arg, isByVal);
387 // Ensure that all of the indices are constants.
388 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
390 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
391 Operands.push_back(C->getSExtValue());
393 return false; // Not a constant operand GEP!
395 // Ensure that the only users of the GEP are load instructions.
396 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
398 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
399 if (LI->isVolatile()) return false; // Don't hack volatile loads
402 // Other uses than load?
406 return false; // Not a load or a GEP.
409 // Now, see if it is safe to promote this load / loads of this GEP. Loading
410 // is safe if Operands, or a prefix of Operands, is marked as safe.
411 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
414 // See if we are already promoting a load with these indices. If not, check
415 // to make sure that we aren't promoting too many elements. If so, nothing
417 if (ToPromote.find(Operands) == ToPromote.end()) {
418 if (maxElements > 0 && ToPromote.size() == maxElements) {
419 DEBUG(dbgs() << "argpromotion not promoting argument '"
420 << Arg->getName() << "' because it would require adding more "
421 << "than " << maxElements << " arguments to the function.\n");
422 // We limit aggregate promotion to only promoting up to a fixed number
423 // of elements of the aggregate.
426 ToPromote.insert(Operands);
430 if (Loads.empty()) return true; // No users, this is a dead argument.
432 // Okay, now we know that the argument is only used by load instructions and
433 // it is safe to unconditionally perform all of them. Use alias analysis to
434 // check to see if the pointer is guaranteed to not be modified from entry of
435 // the function to each of the load instructions.
437 // Because there could be several/many load instructions, remember which
438 // blocks we know to be transparent to the load.
439 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
441 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
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 AliasAnalysis::Location Loc(Load->getPointerOperand(),
450 AA.getTypeStoreSize(Load->getType()),
451 Load->getMetadata(LLVMContext::MD_tbaa));
453 if (AA.canInstructionRangeModify(BB->front(), *Load, Loc))
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, Loc))
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 // Transfer the TBAA info too.
701 newLoad->setMetadata(LLVMContext::MD_tbaa,
702 OrigLoad->getMetadata(LLVMContext::MD_tbaa));
703 Args.push_back(newLoad);
704 AA.copyValue(OrigLoad, Args.back());
709 Args.push_back(Constant::getNullValue(Type::getInt32Ty(F->getContext())));
711 // Push any varargs arguments on the list.
712 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
714 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
715 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
718 // Add any function attributes.
719 if (Attributes attrs = CallPAL.getFnAttributes())
720 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
723 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
724 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
725 Args.begin(), Args.end(), "", Call);
726 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
727 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
728 AttributesVec.end()));
730 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
731 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
732 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
733 AttributesVec.end()));
734 if (cast<CallInst>(Call)->isTailCall())
735 cast<CallInst>(New)->setTailCall();
738 AttributesVec.clear();
740 // Update the alias analysis implementation to know that we are replacing
741 // the old call with a new one.
742 AA.replaceWithNewValue(Call, New);
744 // Update the callgraph to know that the callsite has been transformed.
745 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
746 CalleeNode->replaceCallEdge(Call, New, NF_CGN);
748 if (!Call->use_empty()) {
749 Call->replaceAllUsesWith(New);
753 // Finally, remove the old call from the program, reducing the use-count of
755 Call->eraseFromParent();
758 // Since we have now created the new function, splice the body of the old
759 // function right into the new function, leaving the old rotting hulk of the
761 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
763 // Loop over the argument list, transfering uses of the old arguments over to
764 // the new arguments, also transfering over the names as well.
766 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
767 I2 = NF->arg_begin(); I != E; ++I) {
768 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
769 // If this is an unmodified argument, move the name and users over to the
771 I->replaceAllUsesWith(I2);
773 AA.replaceWithNewValue(I, I2);
778 if (ByValArgsToTransform.count(I)) {
779 // In the callee, we create an alloca, and store each of the new incoming
780 // arguments into the alloca.
781 Instruction *InsertPt = NF->begin()->begin();
783 // Just add all the struct element types.
784 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
785 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
786 const StructType *STy = cast<StructType>(AgTy);
788 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), 0 };
790 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
791 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
793 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2,
794 TheAlloca->getName()+"."+Twine(i),
796 I2->setName(I->getName()+"."+Twine(i));
797 new StoreInst(I2++, Idx, InsertPt);
800 // Anything that used the arg should now use the alloca.
801 I->replaceAllUsesWith(TheAlloca);
802 TheAlloca->takeName(I);
803 AA.replaceWithNewValue(I, TheAlloca);
807 if (I->use_empty()) {
812 // Otherwise, if we promoted this argument, then all users are load
813 // instructions (or GEPs with only load users), and all loads should be
814 // using the new argument that we added.
815 ScalarizeTable &ArgIndices = ScalarizedElements[I];
817 while (!I->use_empty()) {
818 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
819 assert(ArgIndices.begin()->empty() &&
820 "Load element should sort to front!");
821 I2->setName(I->getName()+".val");
822 LI->replaceAllUsesWith(I2);
823 AA.replaceWithNewValue(LI, I2);
824 LI->eraseFromParent();
825 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
826 << "' in function '" << F->getName() << "'\n");
828 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
829 IndicesVector Operands;
830 Operands.reserve(GEP->getNumIndices());
831 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
833 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
835 // GEPs with a single 0 index can be merged with direct loads
836 if (Operands.size() == 1 && Operands.front() == 0)
839 Function::arg_iterator TheArg = I2;
840 for (ScalarizeTable::iterator It = ArgIndices.begin();
841 *It != Operands; ++It, ++TheArg) {
842 assert(It != ArgIndices.end() && "GEP not handled??");
845 std::string NewName = I->getName();
846 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
847 NewName += "." + utostr(Operands[i]);
850 TheArg->setName(NewName);
852 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
853 << "' of function '" << NF->getName() << "'\n");
855 // All of the uses must be load instructions. Replace them all with
856 // the argument specified by ArgNo.
857 while (!GEP->use_empty()) {
858 LoadInst *L = cast<LoadInst>(GEP->use_back());
859 L->replaceAllUsesWith(TheArg);
860 AA.replaceWithNewValue(L, TheArg);
861 L->eraseFromParent();
864 GEP->eraseFromParent();
868 // Increment I2 past all of the arguments added for this promoted pointer.
869 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
873 // Notify the alias analysis implementation that we inserted a new argument.
875 AA.copyValue(Constant::getNullValue(Type::getInt32Ty(F->getContext())),
879 // Tell the alias analysis that the old function is about to disappear.
880 AA.replaceWithNewValue(F, NF);
883 NF_CGN->stealCalledFunctionsFrom(CG[F]);
885 // Now that the old function is dead, delete it. If there is a dangling
886 // reference to the CallgraphNode, just leave the dead function around for
887 // someone else to nuke.
888 CallGraphNode *CGN = CG[F];
889 if (CGN->getNumReferences() == 0)
890 delete CG.removeFunctionFromModule(CGN);
892 F->setLinkage(Function::ExternalLinkage);