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/Compiler.h"
45 #include "llvm/Support/CFG.h"
46 #include "llvm/Support/Debug.h"
47 #include "llvm/Support/raw_ostream.h"
48 #include "llvm/ADT/DepthFirstIterator.h"
49 #include "llvm/ADT/Statistic.h"
50 #include "llvm/ADT/StringExtras.h"
54 STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
55 STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
56 STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted");
57 STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated");
60 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
62 struct VISIBILITY_HIDDEN ArgPromotion : public CallGraphSCCPass {
63 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
64 AU.addRequired<AliasAnalysis>();
65 AU.addRequired<TargetData>();
66 CallGraphSCCPass::getAnalysisUsage(AU);
69 virtual bool runOnSCC(const std::vector<CallGraphNode *> &SCC);
70 static char ID; // Pass identification, replacement for typeid
71 explicit ArgPromotion(unsigned maxElements = 3)
72 : CallGraphSCCPass(&ID), maxElements(maxElements) {}
74 /// A vector used to hold the indices of a single GEP instruction
75 typedef std::vector<uint64_t> IndicesVector;
78 bool PromoteArguments(CallGraphNode *CGN);
79 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const;
80 Function *DoPromotion(Function *F,
81 SmallPtrSet<Argument*, 8> &ArgsToPromote,
82 SmallPtrSet<Argument*, 8> &ByValArgsToTransform);
83 /// The maximum number of elements to expand, or 0 for unlimited.
88 char ArgPromotion::ID = 0;
89 static RegisterPass<ArgPromotion>
90 X("argpromotion", "Promote 'by reference' arguments to scalars");
92 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
93 return new ArgPromotion(maxElements);
96 bool ArgPromotion::runOnSCC(const std::vector<CallGraphNode *> &SCC) {
97 bool Changed = false, LocalChange;
99 do { // Iterate until we stop promoting from this SCC.
101 // Attempt to promote arguments from all functions in this SCC.
102 for (unsigned i = 0, e = SCC.size(); i != e; ++i)
103 LocalChange |= PromoteArguments(SCC[i]);
104 Changed |= LocalChange; // Remember that we changed something.
105 } while (LocalChange);
110 /// PromoteArguments - This method checks the specified function to see if there
111 /// are any promotable arguments and if it is safe to promote the function (for
112 /// example, all callers are direct). If safe to promote some arguments, it
113 /// calls the DoPromotion method.
115 bool ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
116 Function *F = CGN->getFunction();
118 // Make sure that it is local to this module.
119 if (!F || !F->hasLocalLinkage()) return false;
121 // First check: see if there are any pointer arguments! If not, quick exit.
122 SmallVector<std::pair<Argument*, unsigned>, 16> PointerArgs;
124 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
125 I != E; ++I, ++ArgNo)
126 if (isa<PointerType>(I->getType()))
127 PointerArgs.push_back(std::pair<Argument*, unsigned>(I, ArgNo));
128 if (PointerArgs.empty()) return false;
130 // Second check: make sure that all callers are direct callers. We can't
131 // transform functions that have indirect callers.
132 if (F->hasAddressTaken())
135 // Check to see which arguments are promotable. If an argument is promotable,
136 // add it to ArgsToPromote.
137 SmallPtrSet<Argument*, 8> ArgsToPromote;
138 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
139 for (unsigned i = 0; i != PointerArgs.size(); ++i) {
140 bool isByVal = F->paramHasAttr(PointerArgs[i].second+1, Attribute::ByVal);
142 // If this is a byval argument, and if the aggregate type is small, just
143 // pass the elements, which is always safe.
144 Argument *PtrArg = PointerArgs[i].first;
146 const Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
147 if (const StructType *STy = dyn_cast<StructType>(AgTy)) {
148 if (maxElements > 0 && STy->getNumElements() > maxElements) {
149 DEBUG(errs() << "argpromotion disable promoting argument '"
150 << PtrArg->getName() << "' because it would require adding more"
151 << " than " << maxElements << " arguments to the function.\n");
153 // If all the elements are single-value types, we can promote it.
154 bool AllSimple = true;
155 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
156 if (!STy->getElementType(i)->isSingleValueType()) {
161 // Safe to transform, don't even bother trying to "promote" it.
162 // Passing the elements as a scalar will allow scalarrepl to hack on
163 // the new alloca we introduce.
165 ByValArgsToTransform.insert(PtrArg);
172 // Otherwise, see if we can promote the pointer to its value.
173 if (isSafeToPromoteArgument(PtrArg, isByVal))
174 ArgsToPromote.insert(PtrArg);
177 // No promotable pointer arguments.
178 if (ArgsToPromote.empty() && ByValArgsToTransform.empty()) return false;
180 Function *NewF = DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
182 // Update the call graph to know that the function has been transformed.
183 getAnalysis<CallGraph>().changeFunction(F, NewF);
187 /// IsAlwaysValidPointer - Return true if the specified pointer is always legal
189 static bool IsAlwaysValidPointer(Value *V) {
190 if (isa<AllocaInst>(V) || isa<GlobalVariable>(V)) return true;
191 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V))
192 return IsAlwaysValidPointer(GEP->getOperand(0));
193 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(V))
194 if (CE->getOpcode() == Instruction::GetElementPtr)
195 return IsAlwaysValidPointer(CE->getOperand(0));
200 /// AllCalleesPassInValidPointerForArgument - Return true if we can prove that
201 /// all callees pass in a valid pointer for the specified function argument.
202 static bool AllCalleesPassInValidPointerForArgument(Argument *Arg) {
203 Function *Callee = Arg->getParent();
205 unsigned ArgNo = std::distance(Callee->arg_begin(),
206 Function::arg_iterator(Arg));
208 // Look at all call sites of the function. At this pointer we know we only
209 // have direct callees.
210 for (Value::use_iterator UI = Callee->use_begin(), E = Callee->use_end();
212 CallSite CS = CallSite::get(*UI);
213 assert(CS.getInstruction() && "Should only have direct calls!");
215 if (!IsAlwaysValidPointer(CS.getArgument(ArgNo)))
221 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
222 /// that is greater than or equal to the size of prefix, and each of the
223 /// elements in Prefix is the same as the corresponding elements in Longer.
225 /// This means it also returns true when Prefix and Longer are equal!
226 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
227 const ArgPromotion::IndicesVector &Longer) {
228 if (Prefix.size() > Longer.size())
230 for (unsigned i = 0, e = Prefix.size(); i != e; ++i)
231 if (Prefix[i] != Longer[i])
237 /// Checks if Indices, or a prefix of Indices, is in Set.
238 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
239 std::set<ArgPromotion::IndicesVector> &Set) {
240 std::set<ArgPromotion::IndicesVector>::iterator Low;
241 Low = Set.upper_bound(Indices);
242 if (Low != Set.begin())
244 // Low is now the last element smaller than or equal to Indices. This means
245 // it points to a prefix of Indices (possibly Indices itself), if such
248 // This load is safe if any prefix of its operands is safe to load.
249 return Low != Set.end() && IsPrefix(*Low, Indices);
252 /// Mark the given indices (ToMark) as safe in the the given set of indices
253 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
254 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
255 /// already. Furthermore, any indices that Indices is itself a prefix of, are
256 /// removed from Safe (since they are implicitely safe because of Indices now).
257 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
258 std::set<ArgPromotion::IndicesVector> &Safe) {
259 std::set<ArgPromotion::IndicesVector>::iterator Low;
260 Low = Safe.upper_bound(ToMark);
261 // Guard against the case where Safe is empty
262 if (Low != Safe.begin())
264 // Low is now the last element smaller than or equal to Indices. This
265 // means it points to a prefix of Indices (possibly Indices itself), if
266 // such prefix exists.
267 if (Low != Safe.end()) {
268 if (IsPrefix(*Low, ToMark))
269 // If there is already a prefix of these indices (or exactly these
270 // indices) marked a safe, don't bother adding these indices
273 // Increment Low, so we can use it as a "insert before" hint
277 Low = Safe.insert(Low, ToMark);
279 // If there we're a prefix of longer index list(s), remove those
280 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
281 while (Low != End && IsPrefix(ToMark, *Low)) {
282 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
288 /// isSafeToPromoteArgument - As you might guess from the name of this method,
289 /// it checks to see if it is both safe and useful to promote the argument.
290 /// This method limits promotion of aggregates to only promote up to three
291 /// elements of the aggregate in order to avoid exploding the number of
292 /// arguments passed in.
293 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg, bool isByVal) const {
294 typedef std::set<IndicesVector> GEPIndicesSet;
296 // Quick exit for unused arguments
297 if (Arg->use_empty())
300 // We can only promote this argument if all of the uses are loads, or are GEP
301 // instructions (with constant indices) that are subsequently loaded.
303 // Promoting the argument causes it to be loaded in the caller
304 // unconditionally. This is only safe if we can prove that either the load
305 // would have happened in the callee anyway (ie, there is a load in the entry
306 // block) or the pointer passed in at every call site is guaranteed to be
308 // In the former case, invalid loads can happen, but would have happened
309 // anyway, in the latter case, invalid loads won't happen. This prevents us
310 // from introducing an invalid load that wouldn't have happened in the
313 // This set will contain all sets of indices that are loaded in the entry
314 // block, and thus are safe to unconditionally load in the caller.
315 GEPIndicesSet SafeToUnconditionallyLoad;
317 // This set contains all the sets of indices that we are planning to promote.
318 // This makes it possible to limit the number of arguments added.
319 GEPIndicesSet ToPromote;
321 // If the pointer is always valid, any load with first index 0 is valid.
322 if(isByVal || AllCalleesPassInValidPointerForArgument(Arg))
323 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
325 // First, iterate the entry block and mark loads of (geps of) arguments as
327 BasicBlock *EntryBlock = Arg->getParent()->begin();
328 // Declare this here so we can reuse it
329 IndicesVector Indices;
330 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
332 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
333 Value *V = LI->getPointerOperand();
334 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
335 V = GEP->getPointerOperand();
337 // This load actually loads (part of) Arg? Check the indices then.
338 Indices.reserve(GEP->getNumIndices());
339 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
341 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
342 Indices.push_back(CI->getSExtValue());
344 // We found a non-constant GEP index for this argument? Bail out
345 // right away, can't promote this argument at all.
348 // Indices checked out, mark them as safe
349 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
352 } else if (V == Arg) {
353 // Direct loads are equivalent to a GEP with a single 0 index.
354 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
358 // Now, iterate all uses of the argument to see if there are any uses that are
359 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
360 SmallVector<LoadInst*, 16> Loads;
361 IndicesVector Operands;
362 for (Value::use_iterator UI = Arg->use_begin(), E = Arg->use_end();
365 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
366 if (LI->isVolatile()) return false; // Don't hack volatile loads
368 // Direct loads are equivalent to a GEP with a zero index and then a load.
369 Operands.push_back(0);
370 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(*UI)) {
371 if (GEP->use_empty()) {
372 // Dead GEP's cause trouble later. Just remove them if we run into
374 getAnalysis<AliasAnalysis>().deleteValue(GEP);
375 GEP->eraseFromParent();
376 // TODO: This runs the above loop over and over again for dead GEPS
377 // Couldn't we just do increment the UI iterator earlier and erase the
379 return isSafeToPromoteArgument(Arg, isByVal);
382 // Ensure that all of the indices are constants.
383 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
385 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
386 Operands.push_back(C->getSExtValue());
388 return false; // Not a constant operand GEP!
390 // Ensure that the only users of the GEP are load instructions.
391 for (Value::use_iterator UI = GEP->use_begin(), E = GEP->use_end();
393 if (LoadInst *LI = dyn_cast<LoadInst>(*UI)) {
394 if (LI->isVolatile()) return false; // Don't hack volatile loads
397 // Other uses than load?
401 return false; // Not a load or a GEP.
404 // Now, see if it is safe to promote this load / loads of this GEP. Loading
405 // is safe if Operands, or a prefix of Operands, is marked as safe.
406 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
409 // See if we are already promoting a load with these indices. If not, check
410 // to make sure that we aren't promoting too many elements. If so, nothing
412 if (ToPromote.find(Operands) == ToPromote.end()) {
413 if (maxElements > 0 && ToPromote.size() == maxElements) {
414 DEBUG(errs() << "argpromotion not promoting argument '"
415 << Arg->getName() << "' because it would require adding more "
416 << "than " << maxElements << " arguments to the function.\n");
417 // We limit aggregate promotion to only promoting up to a fixed number
418 // of elements of the aggregate.
421 ToPromote.insert(Operands);
425 if (Loads.empty()) return true; // No users, this is a dead argument.
427 // Okay, now we know that the argument is only used by load instructions and
428 // it is safe to unconditionally perform all of them. Use alias analysis to
429 // check to see if the pointer is guaranteed to not be modified from entry of
430 // the function to each of the load instructions.
432 // Because there could be several/many load instructions, remember which
433 // blocks we know to be transparent to the load.
434 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
436 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
437 TargetData &TD = getAnalysis<TargetData>();
439 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
440 // Check to see if the load is invalidated from the start of the block to
442 LoadInst *Load = Loads[i];
443 BasicBlock *BB = Load->getParent();
445 const PointerType *LoadTy =
446 cast<PointerType>(Load->getPointerOperand()->getType());
447 unsigned LoadSize = (unsigned)TD.getTypeStoreSize(LoadTy->getElementType());
449 if (AA.canInstructionRangeModify(BB->front(), *Load, Arg, LoadSize))
450 return false; // Pointer is invalidated!
452 // Now check every path from the entry block to the load for transparency.
453 // To do this, we perform a depth first search on the inverse CFG from the
455 for (pred_iterator PI = pred_begin(BB), E = pred_end(BB); PI != E; ++PI)
456 for (idf_ext_iterator<BasicBlock*, SmallPtrSet<BasicBlock*, 16> >
457 I = idf_ext_begin(*PI, TranspBlocks),
458 E = idf_ext_end(*PI, TranspBlocks); I != E; ++I)
459 if (AA.canBasicBlockModify(**I, Arg, LoadSize))
463 // If the path from the entry of the function to each load is free of
464 // instructions that potentially invalidate the load, we can make the
469 /// DoPromotion - This method actually performs the promotion of the specified
470 /// arguments, and returns the new function. At this point, we know that it's
472 Function *ArgPromotion::DoPromotion(Function *F,
473 SmallPtrSet<Argument*, 8> &ArgsToPromote,
474 SmallPtrSet<Argument*, 8> &ByValArgsToTransform) {
476 // Start by computing a new prototype for the function, which is the same as
477 // the old function, but has modified arguments.
478 const FunctionType *FTy = F->getFunctionType();
479 std::vector<const Type*> Params;
481 typedef std::set<IndicesVector> ScalarizeTable;
483 // ScalarizedElements - If we are promoting a pointer that has elements
484 // accessed out of it, keep track of which elements are accessed so that we
485 // can add one argument for each.
487 // Arguments that are directly loaded will have a zero element value here, to
488 // handle cases where there are both a direct load and GEP accesses.
490 std::map<Argument*, ScalarizeTable> ScalarizedElements;
492 // OriginalLoads - Keep track of a representative load instruction from the
493 // original function so that we can tell the alias analysis implementation
494 // what the new GEP/Load instructions we are inserting look like.
495 std::map<IndicesVector, LoadInst*> OriginalLoads;
497 // Attributes - Keep track of the parameter attributes for the arguments
498 // that we are *not* promoting. For the ones that we do promote, the parameter
499 // attributes are lost
500 SmallVector<AttributeWithIndex, 8> AttributesVec;
501 const AttrListPtr &PAL = F->getAttributes();
503 // Add any return attributes.
504 if (Attributes attrs = PAL.getRetAttributes())
505 AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
507 // First, determine the new argument list
508 unsigned ArgIndex = 1;
509 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
511 if (ByValArgsToTransform.count(I)) {
512 // Simple byval argument? Just add all the struct element types.
513 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
514 const StructType *STy = cast<StructType>(AgTy);
515 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i)
516 Params.push_back(STy->getElementType(i));
517 ++NumByValArgsPromoted;
518 } else if (!ArgsToPromote.count(I)) {
519 // Unchanged argument
520 Params.push_back(I->getType());
521 if (Attributes attrs = PAL.getParamAttributes(ArgIndex))
522 AttributesVec.push_back(AttributeWithIndex::get(Params.size(), attrs));
523 } else if (I->use_empty()) {
524 // Dead argument (which are always marked as promotable)
527 // Okay, this is being promoted. This means that the only uses are loads
528 // or GEPs which are only used by loads
530 // In this table, we will track which indices are loaded from the argument
531 // (where direct loads are tracked as no indices).
532 ScalarizeTable &ArgIndices = ScalarizedElements[I];
533 for (Value::use_iterator UI = I->use_begin(), E = I->use_end(); UI != E;
535 Instruction *User = cast<Instruction>(*UI);
536 assert(isa<LoadInst>(User) || isa<GetElementPtrInst>(User));
537 IndicesVector Indices;
538 Indices.reserve(User->getNumOperands() - 1);
539 // Since loads will only have a single operand, and GEPs only a single
540 // non-index operand, this will record direct loads without any indices,
541 // and gep+loads with the GEP indices.
542 for (User::op_iterator II = User->op_begin() + 1, IE = User->op_end();
544 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
545 // GEPs with a single 0 index can be merged with direct loads
546 if (Indices.size() == 1 && Indices.front() == 0)
548 ArgIndices.insert(Indices);
550 if (LoadInst *L = dyn_cast<LoadInst>(User))
553 // Take any load, we will use it only to update Alias Analysis
554 OrigLoad = cast<LoadInst>(User->use_back());
555 OriginalLoads[Indices] = OrigLoad;
558 // Add a parameter to the function for each element passed in.
559 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
560 E = ArgIndices.end(); SI != E; ++SI) {
561 // not allowed to dereference ->begin() if size() is 0
562 Params.push_back(GetElementPtrInst::getIndexedType(I->getType(),
565 assert(Params.back());
568 if (ArgIndices.size() == 1 && ArgIndices.begin()->empty())
569 ++NumArgumentsPromoted;
571 ++NumAggregatesPromoted;
575 // Add any function attributes.
576 if (Attributes attrs = PAL.getFnAttributes())
577 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
579 const Type *RetTy = FTy->getReturnType();
580 LLVMContext &Context = RetTy->getContext();
582 // Work around LLVM bug PR56: the CWriter cannot emit varargs functions which
583 // have zero fixed arguments.
584 bool ExtraArgHack = false;
585 if (Params.empty() && FTy->isVarArg()) {
587 Params.push_back(Type::Int32Ty);
590 // Construct the new function type using the new arguments.
591 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
593 // Create the new function body and insert it into the module...
594 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
595 NF->copyAttributesFrom(F);
597 // Recompute the parameter attributes list based on the new arguments for
599 NF->setAttributes(AttrListPtr::get(AttributesVec.begin(), AttributesVec.end()));
600 AttributesVec.clear();
602 F->getParent()->getFunctionList().insert(F, NF);
605 // Get the alias analysis information that we need to update to reflect our
607 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
609 // Get the callgraph information that we need to update to reflect our
611 CallGraph &CG = getAnalysis<CallGraph>();
613 // Loop over all of the callers of the function, transforming the call sites
614 // to pass in the loaded pointers.
616 SmallVector<Value*, 16> Args;
617 while (!F->use_empty()) {
618 CallSite CS = CallSite::get(F->use_back());
619 Instruction *Call = CS.getInstruction();
620 const AttrListPtr &CallPAL = CS.getAttributes();
622 // Add any return attributes.
623 if (Attributes attrs = CallPAL.getRetAttributes())
624 AttributesVec.push_back(AttributeWithIndex::get(0, attrs));
626 // Loop over the operands, inserting GEP and loads in the caller as
628 CallSite::arg_iterator AI = CS.arg_begin();
630 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
631 I != E; ++I, ++AI, ++ArgIndex)
632 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
633 Args.push_back(*AI); // Unmodified argument
635 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
636 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
638 } else if (ByValArgsToTransform.count(I)) {
639 // Emit a GEP and load for each element of the struct.
640 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
641 const StructType *STy = cast<StructType>(AgTy);
642 Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 };
643 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
644 Idxs[1] = ConstantInt::get(Type::Int32Ty, i);
645 Value *Idx = GetElementPtrInst::Create(*AI, Idxs, Idxs+2,
646 (*AI)->getName()+"."+utostr(i),
648 // TODO: Tell AA about the new values?
649 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
651 } else if (!I->use_empty()) {
652 // Non-dead argument: insert GEPs and loads as appropriate.
653 ScalarizeTable &ArgIndices = ScalarizedElements[I];
654 // Store the Value* version of the indices in here, but declare it now
656 std::vector<Value*> Ops;
657 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
658 E = ArgIndices.end(); SI != E; ++SI) {
660 LoadInst *OrigLoad = OriginalLoads[*SI];
662 Ops.reserve(SI->size());
663 const Type *ElTy = V->getType();
664 for (IndicesVector::const_iterator II = SI->begin(),
665 IE = SI->end(); II != IE; ++II) {
666 // Use i32 to index structs, and i64 for others (pointers/arrays).
667 // This satisfies GEP constraints.
668 const Type *IdxTy = (isa<StructType>(ElTy) ? Type::Int32Ty : Type::Int64Ty);
669 Ops.push_back(ConstantInt::get(IdxTy, *II));
670 // Keep track of the type we're currently indexing
671 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
673 // And create a GEP to extract those indices
674 V = GetElementPtrInst::Create(V, Ops.begin(), Ops.end(),
675 V->getName()+".idx", Call);
677 AA.copyValue(OrigLoad->getOperand(0), V);
679 Args.push_back(new LoadInst(V, V->getName()+".val", Call));
680 AA.copyValue(OrigLoad, Args.back());
685 Args.push_back(Context.getNullValue(Type::Int32Ty));
687 // Push any varargs arguments on the list
688 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
690 if (Attributes Attrs = CallPAL.getParamAttributes(ArgIndex))
691 AttributesVec.push_back(AttributeWithIndex::get(Args.size(), Attrs));
694 // Add any function attributes.
695 if (Attributes attrs = CallPAL.getFnAttributes())
696 AttributesVec.push_back(AttributeWithIndex::get(~0, attrs));
699 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
700 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
701 Args.begin(), Args.end(), "", Call);
702 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
703 cast<InvokeInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
704 AttributesVec.end()));
706 New = CallInst::Create(NF, Args.begin(), Args.end(), "", Call);
707 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
708 cast<CallInst>(New)->setAttributes(AttrListPtr::get(AttributesVec.begin(),
709 AttributesVec.end()));
710 if (cast<CallInst>(Call)->isTailCall())
711 cast<CallInst>(New)->setTailCall();
714 AttributesVec.clear();
716 // Update the alias analysis implementation to know that we are replacing
717 // the old call with a new one.
718 AA.replaceWithNewValue(Call, New);
720 // Update the callgraph to know that the callsite has been transformed.
721 CG[Call->getParent()->getParent()]->replaceCallSite(Call, New);
723 if (!Call->use_empty()) {
724 Call->replaceAllUsesWith(New);
728 // Finally, remove the old call from the program, reducing the use-count of
730 Call->eraseFromParent();
733 // Since we have now created the new function, splice the body of the old
734 // function right into the new function, leaving the old rotting hulk of the
736 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
738 // Loop over the argument list, transfering uses of the old arguments over to
739 // the new arguments, also transfering over the names as well.
741 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
742 I2 = NF->arg_begin(); I != E; ++I) {
743 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
744 // If this is an unmodified argument, move the name and users over to the
746 I->replaceAllUsesWith(I2);
748 AA.replaceWithNewValue(I, I2);
753 if (ByValArgsToTransform.count(I)) {
754 // In the callee, we create an alloca, and store each of the new incoming
755 // arguments into the alloca.
756 Instruction *InsertPt = NF->begin()->begin();
758 // Just add all the struct element types.
759 const Type *AgTy = cast<PointerType>(I->getType())->getElementType();
760 Value *TheAlloca = new AllocaInst(AgTy, 0, "", InsertPt);
761 const StructType *STy = cast<StructType>(AgTy);
762 Value *Idxs[2] = { ConstantInt::get(Type::Int32Ty, 0), 0 };
764 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
765 Idxs[1] = ConstantInt::get(Type::Int32Ty, i);
767 GetElementPtrInst::Create(TheAlloca, Idxs, Idxs+2,
768 TheAlloca->getName()+"."+utostr(i),
770 I2->setName(I->getName()+"."+utostr(i));
771 new StoreInst(I2++, Idx, InsertPt);
774 // Anything that used the arg should now use the alloca.
775 I->replaceAllUsesWith(TheAlloca);
776 TheAlloca->takeName(I);
777 AA.replaceWithNewValue(I, TheAlloca);
781 if (I->use_empty()) {
786 // Otherwise, if we promoted this argument, then all users are load
787 // instructions (or GEPs with only load users), and all loads should be
788 // using the new argument that we added.
789 ScalarizeTable &ArgIndices = ScalarizedElements[I];
791 while (!I->use_empty()) {
792 if (LoadInst *LI = dyn_cast<LoadInst>(I->use_back())) {
793 assert(ArgIndices.begin()->empty() &&
794 "Load element should sort to front!");
795 I2->setName(I->getName()+".val");
796 LI->replaceAllUsesWith(I2);
797 AA.replaceWithNewValue(LI, I2);
798 LI->eraseFromParent();
799 DEBUG(errs() << "*** Promoted load of argument '" << I->getName()
800 << "' in function '" << F->getName() << "'\n");
802 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->use_back());
803 IndicesVector Operands;
804 Operands.reserve(GEP->getNumIndices());
805 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
807 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
809 // GEPs with a single 0 index can be merged with direct loads
810 if (Operands.size() == 1 && Operands.front() == 0)
813 Function::arg_iterator TheArg = I2;
814 for (ScalarizeTable::iterator It = ArgIndices.begin();
815 *It != Operands; ++It, ++TheArg) {
816 assert(It != ArgIndices.end() && "GEP not handled??");
819 std::string NewName = I->getName();
820 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
821 NewName += "." + utostr(Operands[i]);
824 TheArg->setName(NewName);
826 DEBUG(errs() << "*** Promoted agg argument '" << TheArg->getName()
827 << "' of function '" << NF->getName() << "'\n");
829 // All of the uses must be load instructions. Replace them all with
830 // the argument specified by ArgNo.
831 while (!GEP->use_empty()) {
832 LoadInst *L = cast<LoadInst>(GEP->use_back());
833 L->replaceAllUsesWith(TheArg);
834 AA.replaceWithNewValue(L, TheArg);
835 L->eraseFromParent();
838 GEP->eraseFromParent();
842 // Increment I2 past all of the arguments added for this promoted pointer.
843 for (unsigned i = 0, e = ArgIndices.size(); i != e; ++i)
847 // Notify the alias analysis implementation that we inserted a new argument.
849 AA.copyValue(Context.getNullValue(Type::Int32Ty), NF->arg_begin());
852 // Tell the alias analysis that the old function is about to disappear.
853 AA.replaceWithNewValue(F, NF);
855 // Now that the old function is dead, delete it.
856 F->eraseFromParent();