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 #include "llvm/Transforms/IPO.h"
33 #include "llvm/ADT/DepthFirstIterator.h"
34 #include "llvm/ADT/Statistic.h"
35 #include "llvm/ADT/StringExtras.h"
36 #include "llvm/Analysis/AliasAnalysis.h"
37 #include "llvm/Analysis/CallGraph.h"
38 #include "llvm/Analysis/CallGraphSCCPass.h"
39 #include "llvm/Analysis/ValueTracking.h"
40 #include "llvm/IR/CFG.h"
41 #include "llvm/IR/CallSite.h"
42 #include "llvm/IR/Constants.h"
43 #include "llvm/IR/DataLayout.h"
44 #include "llvm/IR/DebugInfo.h"
45 #include "llvm/IR/DerivedTypes.h"
46 #include "llvm/IR/Instructions.h"
47 #include "llvm/IR/LLVMContext.h"
48 #include "llvm/IR/Module.h"
49 #include "llvm/Support/Debug.h"
50 #include "llvm/Support/raw_ostream.h"
54 #define DEBUG_TYPE "argpromotion"
56 STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
57 STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
58 STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted");
59 STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated");
62 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
64 struct ArgPromotion : public CallGraphSCCPass {
65 void getAnalysisUsage(AnalysisUsage &AU) const override {
66 AU.addRequired<AliasAnalysis>();
67 CallGraphSCCPass::getAnalysisUsage(AU);
70 bool runOnSCC(CallGraphSCC &SCC) override;
71 static char ID; // Pass identification, replacement for typeid
72 explicit ArgPromotion(unsigned maxElements = 3)
73 : CallGraphSCCPass(ID), maxElements(maxElements) {
74 initializeArgPromotionPass(*PassRegistry::getPassRegistry());
77 /// A vector used to hold the indices of a single GEP instruction
78 typedef std::vector<uint64_t> IndicesVector;
81 bool isDenselyPacked(Type *type, const DataLayout &DL);
82 bool canPaddingBeAccessed(Argument *Arg);
83 CallGraphNode *PromoteArguments(CallGraphNode *CGN);
84 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal) const;
85 CallGraphNode *DoPromotion(Function *F,
86 SmallPtrSetImpl<Argument*> &ArgsToPromote,
87 SmallPtrSetImpl<Argument*> &ByValArgsToTransform);
89 using llvm::Pass::doInitialization;
90 bool doInitialization(CallGraph &CG) override;
91 /// The maximum number of elements to expand, or 0 for unlimited.
93 DenseMap<const Function *, DISubprogram *> FunctionDIs;
97 char ArgPromotion::ID = 0;
98 INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion",
99 "Promote 'by reference' arguments to scalars", false, false)
100 INITIALIZE_AG_DEPENDENCY(AliasAnalysis)
101 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
102 INITIALIZE_PASS_END(ArgPromotion, "argpromotion",
103 "Promote 'by reference' arguments to scalars", false, false)
105 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
106 return new ArgPromotion(maxElements);
109 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
110 bool Changed = false, LocalChange;
112 do { // Iterate until we stop promoting from this SCC.
114 // Attempt to promote arguments from all functions in this SCC.
115 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
116 if (CallGraphNode *CGN = PromoteArguments(*I)) {
118 SCC.ReplaceNode(*I, CGN);
121 Changed |= LocalChange; // Remember that we changed something.
122 } while (LocalChange);
127 /// \brief Checks if a type could have padding bytes.
128 bool ArgPromotion::isDenselyPacked(Type *type, const DataLayout &DL) {
130 // There is no size information, so be conservative.
131 if (!type->isSized())
134 // If the alloc size is not equal to the storage size, then there are padding
135 // bytes. For x86_fp80 on x86-64, size: 80 alloc size: 128.
136 if (DL.getTypeSizeInBits(type) != DL.getTypeAllocSizeInBits(type))
139 if (!isa<CompositeType>(type))
142 // For homogenous sequential types, check for padding within members.
143 if (SequentialType *seqTy = dyn_cast<SequentialType>(type))
144 return isa<PointerType>(seqTy) ||
145 isDenselyPacked(seqTy->getElementType(), DL);
147 // Check for padding within and between elements of a struct.
148 StructType *StructTy = cast<StructType>(type);
149 const StructLayout *Layout = DL.getStructLayout(StructTy);
150 uint64_t StartPos = 0;
151 for (unsigned i = 0, E = StructTy->getNumElements(); i < E; ++i) {
152 Type *ElTy = StructTy->getElementType(i);
153 if (!isDenselyPacked(ElTy, DL))
155 if (StartPos != Layout->getElementOffsetInBits(i))
157 StartPos += DL.getTypeAllocSizeInBits(ElTy);
163 /// \brief Checks if the padding bytes of an argument could be accessed.
164 bool ArgPromotion::canPaddingBeAccessed(Argument *arg) {
166 assert(arg->hasByValAttr());
168 // Track all the pointers to the argument to make sure they are not captured.
169 SmallPtrSet<Value *, 16> PtrValues;
170 PtrValues.insert(arg);
172 // Track all of the stores.
173 SmallVector<StoreInst *, 16> Stores;
175 // Scan through the uses recursively to make sure the pointer is always used
177 SmallVector<Value *, 16> WorkList;
178 WorkList.insert(WorkList.end(), arg->user_begin(), arg->user_end());
179 while (!WorkList.empty()) {
180 Value *V = WorkList.back();
182 if (isa<GetElementPtrInst>(V) || isa<PHINode>(V)) {
183 if (PtrValues.insert(V).second)
184 WorkList.insert(WorkList.end(), V->user_begin(), V->user_end());
185 } else if (StoreInst *Store = dyn_cast<StoreInst>(V)) {
186 Stores.push_back(Store);
187 } else if (!isa<LoadInst>(V)) {
192 // Check to make sure the pointers aren't captured
193 for (StoreInst *Store : Stores)
194 if (PtrValues.count(Store->getValueOperand()))
200 /// PromoteArguments - This method checks the specified function to see if there
201 /// are any promotable arguments and if it is safe to promote the function (for
202 /// example, all callers are direct). If safe to promote some arguments, it
203 /// calls the DoPromotion method.
205 CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
206 Function *F = CGN->getFunction();
208 // Make sure that it is local to this module.
209 if (!F || !F->hasLocalLinkage()) return nullptr;
211 // Don't promote arguments for variadic functions. Adding, removing, or
212 // changing non-pack parameters can change the classification of pack
213 // parameters. Frontends encode that classification at the call site in the
214 // IR, while in the callee the classification is determined dynamically based
215 // on the number of registers consumed so far.
216 if (F->isVarArg()) return nullptr;
218 // First check: see if there are any pointer arguments! If not, quick exit.
219 SmallVector<Argument*, 16> PointerArgs;
220 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
221 if (I->getType()->isPointerTy())
222 PointerArgs.push_back(I);
223 if (PointerArgs.empty()) return nullptr;
225 // Second check: make sure that all callers are direct callers. We can't
226 // transform functions that have indirect callers. Also see if the function
227 // is self-recursive.
228 bool isSelfRecursive = false;
229 for (Use &U : F->uses()) {
230 CallSite CS(U.getUser());
231 // Must be a direct call.
232 if (CS.getInstruction() == nullptr || !CS.isCallee(&U)) return nullptr;
234 if (CS.getInstruction()->getParent()->getParent() == F)
235 isSelfRecursive = true;
238 const DataLayout &DL = F->getParent()->getDataLayout();
240 // Check to see which arguments are promotable. If an argument is promotable,
241 // add it to ArgsToPromote.
242 SmallPtrSet<Argument*, 8> ArgsToPromote;
243 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
244 for (unsigned i = 0, e = PointerArgs.size(); i != e; ++i) {
245 Argument *PtrArg = PointerArgs[i];
246 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
248 // Replace sret attribute with noalias. This reduces register pressure by
249 // avoiding a register copy.
250 if (PtrArg->hasStructRetAttr()) {
251 unsigned ArgNo = PtrArg->getArgNo();
254 .removeAttribute(F->getContext(), ArgNo + 1, Attribute::StructRet)
255 .addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
256 for (Use &U : F->uses()) {
257 CallSite CS(U.getUser());
260 .removeAttribute(F->getContext(), ArgNo + 1,
261 Attribute::StructRet)
262 .addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
266 // If this is a byval argument, and if the aggregate type is small, just
267 // pass the elements, which is always safe, if the passed value is densely
268 // packed or if we can prove the padding bytes are never accessed. This does
269 // not apply to inalloca.
270 bool isSafeToPromote =
271 PtrArg->hasByValAttr() &&
272 (isDenselyPacked(AgTy, DL) || !canPaddingBeAccessed(PtrArg));
273 if (isSafeToPromote) {
274 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
275 if (maxElements > 0 && STy->getNumElements() > maxElements) {
276 DEBUG(dbgs() << "argpromotion disable promoting argument '"
277 << PtrArg->getName() << "' because it would require adding more"
278 << " than " << maxElements << " arguments to the function.\n");
282 // If all the elements are single-value types, we can promote it.
283 bool AllSimple = true;
284 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
285 if (!STy->getElementType(i)->isSingleValueType()) {
291 // Safe to transform, don't even bother trying to "promote" it.
292 // Passing the elements as a scalar will allow scalarrepl to hack on
293 // the new alloca we introduce.
295 ByValArgsToTransform.insert(PtrArg);
301 // If the argument is a recursive type and we're in a recursive
302 // function, we could end up infinitely peeling the function argument.
303 if (isSelfRecursive) {
304 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
305 bool RecursiveType = false;
306 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
307 if (STy->getElementType(i) == PtrArg->getType()) {
308 RecursiveType = true;
317 // Otherwise, see if we can promote the pointer to its value.
318 if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr()))
319 ArgsToPromote.insert(PtrArg);
322 // No promotable pointer arguments.
323 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
326 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
329 /// AllCallersPassInValidPointerForArgument - Return true if we can prove that
330 /// all callees pass in a valid pointer for the specified function argument.
331 static bool AllCallersPassInValidPointerForArgument(Argument *Arg) {
332 Function *Callee = Arg->getParent();
333 const DataLayout &DL = Callee->getParent()->getDataLayout();
335 unsigned ArgNo = Arg->getArgNo();
337 // Look at all call sites of the function. At this pointer we know we only
338 // have direct callees.
339 for (User *U : Callee->users()) {
341 assert(CS && "Should only have direct calls!");
343 if (!isDereferenceablePointer(CS.getArgument(ArgNo), DL))
349 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
350 /// that is greater than or equal to the size of prefix, and each of the
351 /// elements in Prefix is the same as the corresponding elements in Longer.
353 /// This means it also returns true when Prefix and Longer are equal!
354 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
355 const ArgPromotion::IndicesVector &Longer) {
356 if (Prefix.size() > Longer.size())
358 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin());
362 /// Checks if Indices, or a prefix of Indices, is in Set.
363 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
364 std::set<ArgPromotion::IndicesVector> &Set) {
365 std::set<ArgPromotion::IndicesVector>::iterator Low;
366 Low = Set.upper_bound(Indices);
367 if (Low != Set.begin())
369 // Low is now the last element smaller than or equal to Indices. This means
370 // it points to a prefix of Indices (possibly Indices itself), if such
373 // This load is safe if any prefix of its operands is safe to load.
374 return Low != Set.end() && IsPrefix(*Low, Indices);
377 /// Mark the given indices (ToMark) as safe in the given set of indices
378 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
379 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
380 /// already. Furthermore, any indices that Indices is itself a prefix of, are
381 /// removed from Safe (since they are implicitely safe because of Indices now).
382 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
383 std::set<ArgPromotion::IndicesVector> &Safe) {
384 std::set<ArgPromotion::IndicesVector>::iterator Low;
385 Low = Safe.upper_bound(ToMark);
386 // Guard against the case where Safe is empty
387 if (Low != Safe.begin())
389 // Low is now the last element smaller than or equal to Indices. This
390 // means it points to a prefix of Indices (possibly Indices itself), if
391 // such prefix exists.
392 if (Low != Safe.end()) {
393 if (IsPrefix(*Low, ToMark))
394 // If there is already a prefix of these indices (or exactly these
395 // indices) marked a safe, don't bother adding these indices
398 // Increment Low, so we can use it as a "insert before" hint
402 Low = Safe.insert(Low, ToMark);
404 // If there we're a prefix of longer index list(s), remove those
405 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
406 while (Low != End && IsPrefix(ToMark, *Low)) {
407 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
413 /// isSafeToPromoteArgument - As you might guess from the name of this method,
414 /// it checks to see if it is both safe and useful to promote the argument.
415 /// This method limits promotion of aggregates to only promote up to three
416 /// elements of the aggregate in order to avoid exploding the number of
417 /// arguments passed in.
418 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg,
419 bool isByValOrInAlloca) const {
420 typedef std::set<IndicesVector> GEPIndicesSet;
422 // Quick exit for unused arguments
423 if (Arg->use_empty())
426 // We can only promote this argument if all of the uses are loads, or are GEP
427 // instructions (with constant indices) that are subsequently loaded.
429 // Promoting the argument causes it to be loaded in the caller
430 // unconditionally. This is only safe if we can prove that either the load
431 // would have happened in the callee anyway (ie, there is a load in the entry
432 // block) or the pointer passed in at every call site is guaranteed to be
434 // In the former case, invalid loads can happen, but would have happened
435 // anyway, in the latter case, invalid loads won't happen. This prevents us
436 // from introducing an invalid load that wouldn't have happened in the
439 // This set will contain all sets of indices that are loaded in the entry
440 // block, and thus are safe to unconditionally load in the caller.
442 // This optimization is also safe for InAlloca parameters, because it verifies
443 // that the address isn't captured.
444 GEPIndicesSet SafeToUnconditionallyLoad;
446 // This set contains all the sets of indices that we are planning to promote.
447 // This makes it possible to limit the number of arguments added.
448 GEPIndicesSet ToPromote;
450 // If the pointer is always valid, any load with first index 0 is valid.
451 if (isByValOrInAlloca || AllCallersPassInValidPointerForArgument(Arg))
452 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
454 // First, iterate the entry block and mark loads of (geps of) arguments as
456 BasicBlock *EntryBlock = Arg->getParent()->begin();
457 // Declare this here so we can reuse it
458 IndicesVector Indices;
459 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
461 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
462 Value *V = LI->getPointerOperand();
463 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
464 V = GEP->getPointerOperand();
466 // This load actually loads (part of) Arg? Check the indices then.
467 Indices.reserve(GEP->getNumIndices());
468 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
470 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
471 Indices.push_back(CI->getSExtValue());
473 // We found a non-constant GEP index for this argument? Bail out
474 // right away, can't promote this argument at all.
477 // Indices checked out, mark them as safe
478 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
481 } else if (V == Arg) {
482 // Direct loads are equivalent to a GEP with a single 0 index.
483 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
487 // Now, iterate all uses of the argument to see if there are any uses that are
488 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
489 SmallVector<LoadInst*, 16> Loads;
490 IndicesVector Operands;
491 for (Use &U : Arg->uses()) {
492 User *UR = U.getUser();
494 if (LoadInst *LI = dyn_cast<LoadInst>(UR)) {
495 // Don't hack volatile/atomic loads
496 if (!LI->isSimple()) return false;
498 // Direct loads are equivalent to a GEP with a zero index and then a load.
499 Operands.push_back(0);
500 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UR)) {
501 if (GEP->use_empty()) {
502 // Dead GEP's cause trouble later. Just remove them if we run into
504 GEP->eraseFromParent();
505 // TODO: This runs the above loop over and over again for dead GEPs
506 // Couldn't we just do increment the UI iterator earlier and erase the
508 return isSafeToPromoteArgument(Arg, isByValOrInAlloca);
511 // Ensure that all of the indices are constants.
512 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
514 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
515 Operands.push_back(C->getSExtValue());
517 return false; // Not a constant operand GEP!
519 // Ensure that the only users of the GEP are load instructions.
520 for (User *GEPU : GEP->users())
521 if (LoadInst *LI = dyn_cast<LoadInst>(GEPU)) {
522 // Don't hack volatile/atomic loads
523 if (!LI->isSimple()) return false;
526 // Other uses than load?
530 return false; // Not a load or a GEP.
533 // Now, see if it is safe to promote this load / loads of this GEP. Loading
534 // is safe if Operands, or a prefix of Operands, is marked as safe.
535 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
538 // See if we are already promoting a load with these indices. If not, check
539 // to make sure that we aren't promoting too many elements. If so, nothing
541 if (ToPromote.find(Operands) == ToPromote.end()) {
542 if (maxElements > 0 && ToPromote.size() == maxElements) {
543 DEBUG(dbgs() << "argpromotion not promoting argument '"
544 << Arg->getName() << "' because it would require adding more "
545 << "than " << maxElements << " arguments to the function.\n");
546 // We limit aggregate promotion to only promoting up to a fixed number
547 // of elements of the aggregate.
550 ToPromote.insert(std::move(Operands));
554 if (Loads.empty()) return true; // No users, this is a dead argument.
556 // Okay, now we know that the argument is only used by load instructions and
557 // it is safe to unconditionally perform all of them. Use alias analysis to
558 // check to see if the pointer is guaranteed to not be modified from entry of
559 // the function to each of the load instructions.
561 // Because there could be several/many load instructions, remember which
562 // blocks we know to be transparent to the load.
563 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
565 AliasAnalysis &AA = getAnalysis<AliasAnalysis>();
567 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
568 // Check to see if the load is invalidated from the start of the block to
570 LoadInst *Load = Loads[i];
571 BasicBlock *BB = Load->getParent();
573 MemoryLocation Loc = MemoryLocation::get(Load);
574 if (AA.canInstructionRangeModRef(BB->front(), *Load, Loc,
576 return false; // Pointer is invalidated!
578 // Now check every path from the entry block to the load for transparency.
579 // To do this, we perform a depth first search on the inverse CFG from the
581 for (BasicBlock *P : predecessors(BB)) {
582 for (BasicBlock *TranspBB : inverse_depth_first_ext(P, TranspBlocks))
583 if (AA.canBasicBlockModify(*TranspBB, Loc))
588 // If the path from the entry of the function to each load is free of
589 // instructions that potentially invalidate the load, we can make the
594 /// DoPromotion - This method actually performs the promotion of the specified
595 /// arguments, and returns the new function. At this point, we know that it's
597 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
598 SmallPtrSetImpl<Argument*> &ArgsToPromote,
599 SmallPtrSetImpl<Argument*> &ByValArgsToTransform) {
601 // Start by computing a new prototype for the function, which is the same as
602 // the old function, but has modified arguments.
603 FunctionType *FTy = F->getFunctionType();
604 std::vector<Type*> Params;
606 typedef std::set<std::pair<Type *, IndicesVector>> ScalarizeTable;
608 // ScalarizedElements - If we are promoting a pointer that has elements
609 // accessed out of it, keep track of which elements are accessed so that we
610 // can add one argument for each.
612 // Arguments that are directly loaded will have a zero element value here, to
613 // handle cases where there are both a direct load and GEP accesses.
615 std::map<Argument*, ScalarizeTable> ScalarizedElements;
617 // OriginalLoads - Keep track of a representative load instruction from the
618 // original function so that we can tell the alias analysis implementation
619 // what the new GEP/Load instructions we are inserting look like.
620 // We need to keep the original loads for each argument and the elements
621 // of the argument that are accessed.
622 std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;
624 // Attribute - Keep track of the parameter attributes for the arguments
625 // that we are *not* promoting. For the ones that we do promote, the parameter
626 // attributes are lost
627 SmallVector<AttributeSet, 8> AttributesVec;
628 const AttributeSet &PAL = F->getAttributes();
630 // Add any return attributes.
631 if (PAL.hasAttributes(AttributeSet::ReturnIndex))
632 AttributesVec.push_back(AttributeSet::get(F->getContext(),
633 PAL.getRetAttributes()));
635 // First, determine the new argument list
636 unsigned ArgIndex = 1;
637 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
639 if (ByValArgsToTransform.count(I)) {
640 // Simple byval argument? Just add all the struct element types.
641 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
642 StructType *STy = cast<StructType>(AgTy);
643 Params.insert(Params.end(), STy->element_begin(), STy->element_end());
644 ++NumByValArgsPromoted;
645 } else if (!ArgsToPromote.count(I)) {
646 // Unchanged argument
647 Params.push_back(I->getType());
648 AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
649 if (attrs.hasAttributes(ArgIndex)) {
650 AttrBuilder B(attrs, ArgIndex);
652 push_back(AttributeSet::get(F->getContext(), Params.size(), B));
654 } else if (I->use_empty()) {
655 // Dead argument (which are always marked as promotable)
658 // Okay, this is being promoted. This means that the only uses are loads
659 // or GEPs which are only used by loads
661 // In this table, we will track which indices are loaded from the argument
662 // (where direct loads are tracked as no indices).
663 ScalarizeTable &ArgIndices = ScalarizedElements[I];
664 for (User *U : I->users()) {
665 Instruction *UI = cast<Instruction>(U);
667 if (LoadInst *L = dyn_cast<LoadInst>(UI))
668 SrcTy = L->getType();
670 SrcTy = cast<GetElementPtrInst>(UI)->getSourceElementType();
671 IndicesVector Indices;
672 Indices.reserve(UI->getNumOperands() - 1);
673 // Since loads will only have a single operand, and GEPs only a single
674 // non-index operand, this will record direct loads without any indices,
675 // and gep+loads with the GEP indices.
676 for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end();
678 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
679 // GEPs with a single 0 index can be merged with direct loads
680 if (Indices.size() == 1 && Indices.front() == 0)
682 ArgIndices.insert(std::make_pair(SrcTy, Indices));
684 if (LoadInst *L = dyn_cast<LoadInst>(UI))
687 // Take any load, we will use it only to update Alias Analysis
688 OrigLoad = cast<LoadInst>(UI->user_back());
689 OriginalLoads[std::make_pair(I, Indices)] = OrigLoad;
692 // Add a parameter to the function for each element passed in.
693 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
694 E = ArgIndices.end(); SI != E; ++SI) {
695 // not allowed to dereference ->begin() if size() is 0
696 Params.push_back(GetElementPtrInst::getIndexedType(
697 cast<PointerType>(I->getType()->getScalarType())->getElementType(),
699 assert(Params.back());
702 if (ArgIndices.size() == 1 && ArgIndices.begin()->second.empty())
703 ++NumArgumentsPromoted;
705 ++NumAggregatesPromoted;
709 // Add any function attributes.
710 if (PAL.hasAttributes(AttributeSet::FunctionIndex))
711 AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
712 PAL.getFnAttributes()));
714 Type *RetTy = FTy->getReturnType();
716 // Construct the new function type using the new arguments.
717 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
719 // Create the new function body and insert it into the module.
720 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
721 NF->copyAttributesFrom(F);
723 // Patch the pointer to LLVM function in debug info descriptor.
724 auto DI = FunctionDIs.find(F);
725 if (DI != FunctionDIs.end()) {
726 DISubprogram *SP = DI->second;
727 SP->replaceFunction(NF);
728 // Ensure the map is updated so it can be reused on subsequent argument
729 // promotions of the same function.
730 FunctionDIs.erase(DI);
731 FunctionDIs[NF] = SP;
734 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
737 // Recompute the parameter attributes list based on the new arguments for
739 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
740 AttributesVec.clear();
742 F->getParent()->getFunctionList().insert(F, NF);
745 // Get the callgraph information that we need to update to reflect our
747 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
749 // Get a new callgraph node for NF.
750 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
752 // Loop over all of the callers of the function, transforming the call sites
753 // to pass in the loaded pointers.
755 SmallVector<Value*, 16> Args;
756 while (!F->use_empty()) {
757 CallSite CS(F->user_back());
758 assert(CS.getCalledFunction() == F);
759 Instruction *Call = CS.getInstruction();
760 const AttributeSet &CallPAL = CS.getAttributes();
762 // Add any return attributes.
763 if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
764 AttributesVec.push_back(AttributeSet::get(F->getContext(),
765 CallPAL.getRetAttributes()));
767 // Loop over the operands, inserting GEP and loads in the caller as
769 CallSite::arg_iterator AI = CS.arg_begin();
771 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
772 I != E; ++I, ++AI, ++ArgIndex)
773 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
774 Args.push_back(*AI); // Unmodified argument
776 if (CallPAL.hasAttributes(ArgIndex)) {
777 AttrBuilder B(CallPAL, ArgIndex);
779 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
781 } else if (ByValArgsToTransform.count(I)) {
782 // Emit a GEP and load for each element of the struct.
783 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
784 StructType *STy = cast<StructType>(AgTy);
786 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
787 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
788 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
789 Value *Idx = GetElementPtrInst::Create(
790 STy, *AI, Idxs, (*AI)->getName() + "." + Twine(i), Call);
791 // TODO: Tell AA about the new values?
792 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
794 } else if (!I->use_empty()) {
795 // Non-dead argument: insert GEPs and loads as appropriate.
796 ScalarizeTable &ArgIndices = ScalarizedElements[I];
797 // Store the Value* version of the indices in here, but declare it now
799 std::vector<Value*> Ops;
800 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
801 E = ArgIndices.end(); SI != E; ++SI) {
803 LoadInst *OrigLoad = OriginalLoads[std::make_pair(I, SI->second)];
804 if (!SI->second.empty()) {
805 Ops.reserve(SI->second.size());
806 Type *ElTy = V->getType();
807 for (IndicesVector::const_iterator II = SI->second.begin(),
808 IE = SI->second.end();
810 // Use i32 to index structs, and i64 for others (pointers/arrays).
811 // This satisfies GEP constraints.
812 Type *IdxTy = (ElTy->isStructTy() ?
813 Type::getInt32Ty(F->getContext()) :
814 Type::getInt64Ty(F->getContext()));
815 Ops.push_back(ConstantInt::get(IdxTy, *II));
816 // Keep track of the type we're currently indexing.
817 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
819 // And create a GEP to extract those indices.
820 V = GetElementPtrInst::Create(SI->first, V, Ops,
821 V->getName() + ".idx", Call);
824 // Since we're replacing a load make sure we take the alignment
825 // of the previous load.
826 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
827 newLoad->setAlignment(OrigLoad->getAlignment());
828 // Transfer the AA info too.
830 OrigLoad->getAAMetadata(AAInfo);
831 newLoad->setAAMetadata(AAInfo);
833 Args.push_back(newLoad);
837 // Push any varargs arguments on the list.
838 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
840 if (CallPAL.hasAttributes(ArgIndex)) {
841 AttrBuilder B(CallPAL, ArgIndex);
843 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
847 // Add any function attributes.
848 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
849 AttributesVec.push_back(AttributeSet::get(Call->getContext(),
850 CallPAL.getFnAttributes()));
853 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
854 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
856 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
857 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
860 New = CallInst::Create(NF, Args, "", Call);
861 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
862 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
864 if (cast<CallInst>(Call)->isTailCall())
865 cast<CallInst>(New)->setTailCall();
867 New->setDebugLoc(Call->getDebugLoc());
869 AttributesVec.clear();
871 // Update the callgraph to know that the callsite has been transformed.
872 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
873 CalleeNode->replaceCallEdge(CS, CallSite(New), NF_CGN);
875 if (!Call->use_empty()) {
876 Call->replaceAllUsesWith(New);
880 // Finally, remove the old call from the program, reducing the use-count of
882 Call->eraseFromParent();
885 // Since we have now created the new function, splice the body of the old
886 // function right into the new function, leaving the old rotting hulk of the
888 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
890 // Loop over the argument list, transferring uses of the old arguments over to
891 // the new arguments, also transferring over the names as well.
893 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
894 I2 = NF->arg_begin(); I != E; ++I) {
895 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
896 // If this is an unmodified argument, move the name and users over to the
898 I->replaceAllUsesWith(I2);
904 if (ByValArgsToTransform.count(I)) {
905 // In the callee, we create an alloca, and store each of the new incoming
906 // arguments into the alloca.
907 Instruction *InsertPt = NF->begin()->begin();
909 // Just add all the struct element types.
910 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
911 Value *TheAlloca = new AllocaInst(AgTy, nullptr, "", InsertPt);
912 StructType *STy = cast<StructType>(AgTy);
914 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
916 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
917 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
918 Value *Idx = GetElementPtrInst::Create(
919 AgTy, TheAlloca, Idxs, TheAlloca->getName() + "." + Twine(i),
921 I2->setName(I->getName()+"."+Twine(i));
922 new StoreInst(I2++, Idx, InsertPt);
925 // Anything that used the arg should now use the alloca.
926 I->replaceAllUsesWith(TheAlloca);
927 TheAlloca->takeName(I);
929 // If the alloca is used in a call, we must clear the tail flag since
930 // the callee now uses an alloca from the caller.
931 for (User *U : TheAlloca->users()) {
932 CallInst *Call = dyn_cast<CallInst>(U);
935 Call->setTailCall(false);
943 // Otherwise, if we promoted this argument, then all users are load
944 // instructions (or GEPs with only load users), and all loads should be
945 // using the new argument that we added.
946 ScalarizeTable &ArgIndices = ScalarizedElements[I];
948 while (!I->use_empty()) {
949 if (LoadInst *LI = dyn_cast<LoadInst>(I->user_back())) {
950 assert(ArgIndices.begin()->second.empty() &&
951 "Load element should sort to front!");
952 I2->setName(I->getName()+".val");
953 LI->replaceAllUsesWith(I2);
954 LI->eraseFromParent();
955 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
956 << "' in function '" << F->getName() << "'\n");
958 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_back());
959 IndicesVector Operands;
960 Operands.reserve(GEP->getNumIndices());
961 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
963 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
965 // GEPs with a single 0 index can be merged with direct loads
966 if (Operands.size() == 1 && Operands.front() == 0)
969 Function::arg_iterator TheArg = I2;
970 for (ScalarizeTable::iterator It = ArgIndices.begin();
971 It->second != Operands; ++It, ++TheArg) {
972 assert(It != ArgIndices.end() && "GEP not handled??");
975 std::string NewName = I->getName();
976 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
977 NewName += "." + utostr(Operands[i]);
980 TheArg->setName(NewName);
982 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
983 << "' of function '" << NF->getName() << "'\n");
985 // All of the uses must be load instructions. Replace them all with
986 // the argument specified by ArgNo.
987 while (!GEP->use_empty()) {
988 LoadInst *L = cast<LoadInst>(GEP->user_back());
989 L->replaceAllUsesWith(TheArg);
990 L->eraseFromParent();
992 GEP->eraseFromParent();
996 // Increment I2 past all of the arguments added for this promoted pointer.
997 std::advance(I2, ArgIndices.size());
1000 NF_CGN->stealCalledFunctionsFrom(CG[F]);
1002 // Now that the old function is dead, delete it. If there is a dangling
1003 // reference to the CallgraphNode, just leave the dead function around for
1004 // someone else to nuke.
1005 CallGraphNode *CGN = CG[F];
1006 if (CGN->getNumReferences() == 0)
1007 delete CG.removeFunctionFromModule(CGN);
1009 F->setLinkage(Function::ExternalLinkage);
1014 bool ArgPromotion::doInitialization(CallGraph &CG) {
1015 FunctionDIs = makeSubprogramMap(CG.getModule());
1016 return CallGraphSCCPass::doInitialization(CG);