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/AssumptionCache.h"
38 #include "llvm/Analysis/BasicAliasAnalysis.h"
39 #include "llvm/Analysis/CallGraph.h"
40 #include "llvm/Analysis/CallGraphSCCPass.h"
41 #include "llvm/Analysis/TargetLibraryInfo.h"
42 #include "llvm/Analysis/ValueTracking.h"
43 #include "llvm/IR/CFG.h"
44 #include "llvm/IR/CallSite.h"
45 #include "llvm/IR/Constants.h"
46 #include "llvm/IR/DataLayout.h"
47 #include "llvm/IR/DebugInfo.h"
48 #include "llvm/IR/DerivedTypes.h"
49 #include "llvm/IR/Instructions.h"
50 #include "llvm/IR/LLVMContext.h"
51 #include "llvm/IR/Module.h"
52 #include "llvm/Support/Debug.h"
53 #include "llvm/Support/raw_ostream.h"
57 #define DEBUG_TYPE "argpromotion"
59 STATISTIC(NumArgumentsPromoted , "Number of pointer arguments promoted");
60 STATISTIC(NumAggregatesPromoted, "Number of aggregate arguments promoted");
61 STATISTIC(NumByValArgsPromoted , "Number of byval arguments promoted");
62 STATISTIC(NumArgumentsDead , "Number of dead pointer args eliminated");
65 /// ArgPromotion - The 'by reference' to 'by value' argument promotion pass.
67 struct ArgPromotion : public CallGraphSCCPass {
68 void getAnalysisUsage(AnalysisUsage &AU) const override {
69 AU.addRequired<AssumptionCacheTracker>();
70 AU.addRequired<TargetLibraryInfoWrapperPass>();
71 CallGraphSCCPass::getAnalysisUsage(AU);
74 bool runOnSCC(CallGraphSCC &SCC) override;
75 static char ID; // Pass identification, replacement for typeid
76 explicit ArgPromotion(unsigned maxElements = 3)
77 : CallGraphSCCPass(ID), maxElements(maxElements) {
78 initializeArgPromotionPass(*PassRegistry::getPassRegistry());
81 /// A vector used to hold the indices of a single GEP instruction
82 typedef std::vector<uint64_t> IndicesVector;
85 bool isDenselyPacked(Type *type, const DataLayout &DL);
86 bool canPaddingBeAccessed(Argument *Arg);
87 CallGraphNode *PromoteArguments(CallGraphNode *CGN);
88 bool isSafeToPromoteArgument(Argument *Arg, bool isByVal,
89 AAResults &AAR) const;
90 CallGraphNode *DoPromotion(Function *F,
91 SmallPtrSetImpl<Argument*> &ArgsToPromote,
92 SmallPtrSetImpl<Argument*> &ByValArgsToTransform);
94 using llvm::Pass::doInitialization;
95 bool doInitialization(CallGraph &CG) override;
96 /// The maximum number of elements to expand, or 0 for unlimited.
98 DenseMap<const Function *, DISubprogram *> FunctionDIs;
102 char ArgPromotion::ID = 0;
103 INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion",
104 "Promote 'by reference' arguments to scalars", false, false)
105 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
106 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
107 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
108 INITIALIZE_PASS_END(ArgPromotion, "argpromotion",
109 "Promote 'by reference' arguments to scalars", false, false)
111 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
112 return new ArgPromotion(maxElements);
115 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
116 bool Changed = false, LocalChange;
118 do { // Iterate until we stop promoting from this SCC.
120 // Attempt to promote arguments from all functions in this SCC.
121 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
122 if (CallGraphNode *CGN = PromoteArguments(*I)) {
124 SCC.ReplaceNode(*I, CGN);
127 Changed |= LocalChange; // Remember that we changed something.
128 } while (LocalChange);
133 /// \brief Checks if a type could have padding bytes.
134 bool ArgPromotion::isDenselyPacked(Type *type, const DataLayout &DL) {
136 // There is no size information, so be conservative.
137 if (!type->isSized())
140 // If the alloc size is not equal to the storage size, then there are padding
141 // bytes. For x86_fp80 on x86-64, size: 80 alloc size: 128.
142 if (DL.getTypeSizeInBits(type) != DL.getTypeAllocSizeInBits(type))
145 if (!isa<CompositeType>(type))
148 // For homogenous sequential types, check for padding within members.
149 if (SequentialType *seqTy = dyn_cast<SequentialType>(type))
150 return isa<PointerType>(seqTy) ||
151 isDenselyPacked(seqTy->getElementType(), DL);
153 // Check for padding within and between elements of a struct.
154 StructType *StructTy = cast<StructType>(type);
155 const StructLayout *Layout = DL.getStructLayout(StructTy);
156 uint64_t StartPos = 0;
157 for (unsigned i = 0, E = StructTy->getNumElements(); i < E; ++i) {
158 Type *ElTy = StructTy->getElementType(i);
159 if (!isDenselyPacked(ElTy, DL))
161 if (StartPos != Layout->getElementOffsetInBits(i))
163 StartPos += DL.getTypeAllocSizeInBits(ElTy);
169 /// \brief Checks if the padding bytes of an argument could be accessed.
170 bool ArgPromotion::canPaddingBeAccessed(Argument *arg) {
172 assert(arg->hasByValAttr());
174 // Track all the pointers to the argument to make sure they are not captured.
175 SmallPtrSet<Value *, 16> PtrValues;
176 PtrValues.insert(arg);
178 // Track all of the stores.
179 SmallVector<StoreInst *, 16> Stores;
181 // Scan through the uses recursively to make sure the pointer is always used
183 SmallVector<Value *, 16> WorkList;
184 WorkList.insert(WorkList.end(), arg->user_begin(), arg->user_end());
185 while (!WorkList.empty()) {
186 Value *V = WorkList.back();
188 if (isa<GetElementPtrInst>(V) || isa<PHINode>(V)) {
189 if (PtrValues.insert(V).second)
190 WorkList.insert(WorkList.end(), V->user_begin(), V->user_end());
191 } else if (StoreInst *Store = dyn_cast<StoreInst>(V)) {
192 Stores.push_back(Store);
193 } else if (!isa<LoadInst>(V)) {
198 // Check to make sure the pointers aren't captured
199 for (StoreInst *Store : Stores)
200 if (PtrValues.count(Store->getValueOperand()))
206 /// PromoteArguments - This method checks the specified function to see if there
207 /// are any promotable arguments and if it is safe to promote the function (for
208 /// example, all callers are direct). If safe to promote some arguments, it
209 /// calls the DoPromotion method.
211 CallGraphNode *ArgPromotion::PromoteArguments(CallGraphNode *CGN) {
212 Function *F = CGN->getFunction();
214 // Make sure that it is local to this module.
215 if (!F || !F->hasLocalLinkage()) return nullptr;
217 // Don't promote arguments for variadic functions. Adding, removing, or
218 // changing non-pack parameters can change the classification of pack
219 // parameters. Frontends encode that classification at the call site in the
220 // IR, while in the callee the classification is determined dynamically based
221 // on the number of registers consumed so far.
222 if (F->isVarArg()) return nullptr;
224 // First check: see if there are any pointer arguments! If not, quick exit.
225 SmallVector<Argument*, 16> PointerArgs;
226 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E; ++I)
227 if (I->getType()->isPointerTy())
228 PointerArgs.push_back(I);
229 if (PointerArgs.empty()) return nullptr;
231 // Second check: make sure that all callers are direct callers. We can't
232 // transform functions that have indirect callers. Also see if the function
233 // is self-recursive.
234 bool isSelfRecursive = false;
235 for (Use &U : F->uses()) {
236 CallSite CS(U.getUser());
237 // Must be a direct call.
238 if (CS.getInstruction() == nullptr || !CS.isCallee(&U)) return nullptr;
240 if (CS.getInstruction()->getParent()->getParent() == F)
241 isSelfRecursive = true;
244 const DataLayout &DL = F->getParent()->getDataLayout();
246 // We need to manually construct BasicAA directly in order to disable its use
247 // of other function analyses.
248 BasicAAResult BAR(createLegacyPMBasicAAResult(*this, *F));
250 // Construct our own AA results for this function. We do this manually to
251 // work around the limitations of the legacy pass manager.
252 AAResults AAR(createLegacyPMAAResults(*this, *F, BAR));
254 // Check to see which arguments are promotable. If an argument is promotable,
255 // add it to ArgsToPromote.
256 SmallPtrSet<Argument*, 8> ArgsToPromote;
257 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
258 for (unsigned i = 0, e = PointerArgs.size(); i != e; ++i) {
259 Argument *PtrArg = PointerArgs[i];
260 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
262 // Replace sret attribute with noalias. This reduces register pressure by
263 // avoiding a register copy.
264 if (PtrArg->hasStructRetAttr()) {
265 unsigned ArgNo = PtrArg->getArgNo();
268 .removeAttribute(F->getContext(), ArgNo + 1, Attribute::StructRet)
269 .addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
270 for (Use &U : F->uses()) {
271 CallSite CS(U.getUser());
274 .removeAttribute(F->getContext(), ArgNo + 1,
275 Attribute::StructRet)
276 .addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
280 // If this is a byval argument, and if the aggregate type is small, just
281 // pass the elements, which is always safe, if the passed value is densely
282 // packed or if we can prove the padding bytes are never accessed. This does
283 // not apply to inalloca.
284 bool isSafeToPromote =
285 PtrArg->hasByValAttr() &&
286 (isDenselyPacked(AgTy, DL) || !canPaddingBeAccessed(PtrArg));
287 if (isSafeToPromote) {
288 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
289 if (maxElements > 0 && STy->getNumElements() > maxElements) {
290 DEBUG(dbgs() << "argpromotion disable promoting argument '"
291 << PtrArg->getName() << "' because it would require adding more"
292 << " than " << maxElements << " arguments to the function.\n");
296 // If all the elements are single-value types, we can promote it.
297 bool AllSimple = true;
298 for (const auto *EltTy : STy->elements()) {
299 if (!EltTy->isSingleValueType()) {
305 // Safe to transform, don't even bother trying to "promote" it.
306 // Passing the elements as a scalar will allow scalarrepl to hack on
307 // the new alloca we introduce.
309 ByValArgsToTransform.insert(PtrArg);
315 // If the argument is a recursive type and we're in a recursive
316 // function, we could end up infinitely peeling the function argument.
317 if (isSelfRecursive) {
318 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
319 bool RecursiveType = false;
320 for (const auto *EltTy : STy->elements()) {
321 if (EltTy == PtrArg->getType()) {
322 RecursiveType = true;
331 // Otherwise, see if we can promote the pointer to its value.
332 if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr(), AAR))
333 ArgsToPromote.insert(PtrArg);
336 // No promotable pointer arguments.
337 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
340 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform);
343 /// AllCallersPassInValidPointerForArgument - Return true if we can prove that
344 /// all callees pass in a valid pointer for the specified function argument.
345 static bool AllCallersPassInValidPointerForArgument(Argument *Arg) {
346 Function *Callee = Arg->getParent();
347 const DataLayout &DL = Callee->getParent()->getDataLayout();
349 unsigned ArgNo = Arg->getArgNo();
351 // Look at all call sites of the function. At this pointer we know we only
352 // have direct callees.
353 for (User *U : Callee->users()) {
355 assert(CS && "Should only have direct calls!");
357 if (!isDereferenceablePointer(CS.getArgument(ArgNo), DL))
363 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
364 /// that is greater than or equal to the size of prefix, and each of the
365 /// elements in Prefix is the same as the corresponding elements in Longer.
367 /// This means it also returns true when Prefix and Longer are equal!
368 static bool IsPrefix(const ArgPromotion::IndicesVector &Prefix,
369 const ArgPromotion::IndicesVector &Longer) {
370 if (Prefix.size() > Longer.size())
372 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin());
376 /// Checks if Indices, or a prefix of Indices, is in Set.
377 static bool PrefixIn(const ArgPromotion::IndicesVector &Indices,
378 std::set<ArgPromotion::IndicesVector> &Set) {
379 std::set<ArgPromotion::IndicesVector>::iterator Low;
380 Low = Set.upper_bound(Indices);
381 if (Low != Set.begin())
383 // Low is now the last element smaller than or equal to Indices. This means
384 // it points to a prefix of Indices (possibly Indices itself), if such
387 // This load is safe if any prefix of its operands is safe to load.
388 return Low != Set.end() && IsPrefix(*Low, Indices);
391 /// Mark the given indices (ToMark) as safe in the given set of indices
392 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
393 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
394 /// already. Furthermore, any indices that Indices is itself a prefix of, are
395 /// removed from Safe (since they are implicitely safe because of Indices now).
396 static void MarkIndicesSafe(const ArgPromotion::IndicesVector &ToMark,
397 std::set<ArgPromotion::IndicesVector> &Safe) {
398 std::set<ArgPromotion::IndicesVector>::iterator Low;
399 Low = Safe.upper_bound(ToMark);
400 // Guard against the case where Safe is empty
401 if (Low != Safe.begin())
403 // Low is now the last element smaller than or equal to Indices. This
404 // means it points to a prefix of Indices (possibly Indices itself), if
405 // such prefix exists.
406 if (Low != Safe.end()) {
407 if (IsPrefix(*Low, ToMark))
408 // If there is already a prefix of these indices (or exactly these
409 // indices) marked a safe, don't bother adding these indices
412 // Increment Low, so we can use it as a "insert before" hint
416 Low = Safe.insert(Low, ToMark);
418 // If there we're a prefix of longer index list(s), remove those
419 std::set<ArgPromotion::IndicesVector>::iterator End = Safe.end();
420 while (Low != End && IsPrefix(ToMark, *Low)) {
421 std::set<ArgPromotion::IndicesVector>::iterator Remove = Low;
427 /// isSafeToPromoteArgument - As you might guess from the name of this method,
428 /// it checks to see if it is both safe and useful to promote the argument.
429 /// This method limits promotion of aggregates to only promote up to three
430 /// elements of the aggregate in order to avoid exploding the number of
431 /// arguments passed in.
432 bool ArgPromotion::isSafeToPromoteArgument(Argument *Arg,
433 bool isByValOrInAlloca,
434 AAResults &AAR) const {
435 typedef std::set<IndicesVector> GEPIndicesSet;
437 // Quick exit for unused arguments
438 if (Arg->use_empty())
441 // We can only promote this argument if all of the uses are loads, or are GEP
442 // instructions (with constant indices) that are subsequently loaded.
444 // Promoting the argument causes it to be loaded in the caller
445 // unconditionally. This is only safe if we can prove that either the load
446 // would have happened in the callee anyway (ie, there is a load in the entry
447 // block) or the pointer passed in at every call site is guaranteed to be
449 // In the former case, invalid loads can happen, but would have happened
450 // anyway, in the latter case, invalid loads won't happen. This prevents us
451 // from introducing an invalid load that wouldn't have happened in the
454 // This set will contain all sets of indices that are loaded in the entry
455 // block, and thus are safe to unconditionally load in the caller.
457 // This optimization is also safe for InAlloca parameters, because it verifies
458 // that the address isn't captured.
459 GEPIndicesSet SafeToUnconditionallyLoad;
461 // This set contains all the sets of indices that we are planning to promote.
462 // This makes it possible to limit the number of arguments added.
463 GEPIndicesSet ToPromote;
465 // If the pointer is always valid, any load with first index 0 is valid.
466 if (isByValOrInAlloca || AllCallersPassInValidPointerForArgument(Arg))
467 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
469 // First, iterate the entry block and mark loads of (geps of) arguments as
471 BasicBlock *EntryBlock = Arg->getParent()->begin();
472 // Declare this here so we can reuse it
473 IndicesVector Indices;
474 for (BasicBlock::iterator I = EntryBlock->begin(), E = EntryBlock->end();
476 if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
477 Value *V = LI->getPointerOperand();
478 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
479 V = GEP->getPointerOperand();
481 // This load actually loads (part of) Arg? Check the indices then.
482 Indices.reserve(GEP->getNumIndices());
483 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
485 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
486 Indices.push_back(CI->getSExtValue());
488 // We found a non-constant GEP index for this argument? Bail out
489 // right away, can't promote this argument at all.
492 // Indices checked out, mark them as safe
493 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
496 } else if (V == Arg) {
497 // Direct loads are equivalent to a GEP with a single 0 index.
498 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
502 // Now, iterate all uses of the argument to see if there are any uses that are
503 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
504 SmallVector<LoadInst*, 16> Loads;
505 IndicesVector Operands;
506 for (Use &U : Arg->uses()) {
507 User *UR = U.getUser();
509 if (LoadInst *LI = dyn_cast<LoadInst>(UR)) {
510 // Don't hack volatile/atomic loads
511 if (!LI->isSimple()) return false;
513 // Direct loads are equivalent to a GEP with a zero index and then a load.
514 Operands.push_back(0);
515 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UR)) {
516 if (GEP->use_empty()) {
517 // Dead GEP's cause trouble later. Just remove them if we run into
519 GEP->eraseFromParent();
520 // TODO: This runs the above loop over and over again for dead GEPs
521 // Couldn't we just do increment the UI iterator earlier and erase the
523 return isSafeToPromoteArgument(Arg, isByValOrInAlloca, AAR);
526 // Ensure that all of the indices are constants.
527 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
529 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
530 Operands.push_back(C->getSExtValue());
532 return false; // Not a constant operand GEP!
534 // Ensure that the only users of the GEP are load instructions.
535 for (User *GEPU : GEP->users())
536 if (LoadInst *LI = dyn_cast<LoadInst>(GEPU)) {
537 // Don't hack volatile/atomic loads
538 if (!LI->isSimple()) return false;
541 // Other uses than load?
545 return false; // Not a load or a GEP.
548 // Now, see if it is safe to promote this load / loads of this GEP. Loading
549 // is safe if Operands, or a prefix of Operands, is marked as safe.
550 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
553 // See if we are already promoting a load with these indices. If not, check
554 // to make sure that we aren't promoting too many elements. If so, nothing
556 if (ToPromote.find(Operands) == ToPromote.end()) {
557 if (maxElements > 0 && ToPromote.size() == maxElements) {
558 DEBUG(dbgs() << "argpromotion not promoting argument '"
559 << Arg->getName() << "' because it would require adding more "
560 << "than " << maxElements << " arguments to the function.\n");
561 // We limit aggregate promotion to only promoting up to a fixed number
562 // of elements of the aggregate.
565 ToPromote.insert(std::move(Operands));
569 if (Loads.empty()) return true; // No users, this is a dead argument.
571 // Okay, now we know that the argument is only used by load instructions and
572 // it is safe to unconditionally perform all of them. Use alias analysis to
573 // check to see if the pointer is guaranteed to not be modified from entry of
574 // the function to each of the load instructions.
576 // Because there could be several/many load instructions, remember which
577 // blocks we know to be transparent to the load.
578 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
580 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
581 // Check to see if the load is invalidated from the start of the block to
583 LoadInst *Load = Loads[i];
584 BasicBlock *BB = Load->getParent();
586 MemoryLocation Loc = MemoryLocation::get(Load);
587 if (AAR.canInstructionRangeModRef(BB->front(), *Load, Loc, MRI_Mod))
588 return false; // Pointer is invalidated!
590 // Now check every path from the entry block to the load for transparency.
591 // To do this, we perform a depth first search on the inverse CFG from the
593 for (BasicBlock *P : predecessors(BB)) {
594 for (BasicBlock *TranspBB : inverse_depth_first_ext(P, TranspBlocks))
595 if (AAR.canBasicBlockModify(*TranspBB, Loc))
600 // If the path from the entry of the function to each load is free of
601 // instructions that potentially invalidate the load, we can make the
606 /// DoPromotion - This method actually performs the promotion of the specified
607 /// arguments, and returns the new function. At this point, we know that it's
609 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
610 SmallPtrSetImpl<Argument*> &ArgsToPromote,
611 SmallPtrSetImpl<Argument*> &ByValArgsToTransform) {
613 // Start by computing a new prototype for the function, which is the same as
614 // the old function, but has modified arguments.
615 FunctionType *FTy = F->getFunctionType();
616 std::vector<Type*> Params;
618 typedef std::set<std::pair<Type *, IndicesVector>> ScalarizeTable;
620 // ScalarizedElements - If we are promoting a pointer that has elements
621 // accessed out of it, keep track of which elements are accessed so that we
622 // can add one argument for each.
624 // Arguments that are directly loaded will have a zero element value here, to
625 // handle cases where there are both a direct load and GEP accesses.
627 std::map<Argument*, ScalarizeTable> ScalarizedElements;
629 // OriginalLoads - Keep track of a representative load instruction from the
630 // original function so that we can tell the alias analysis implementation
631 // what the new GEP/Load instructions we are inserting look like.
632 // We need to keep the original loads for each argument and the elements
633 // of the argument that are accessed.
634 std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;
636 // Attribute - Keep track of the parameter attributes for the arguments
637 // that we are *not* promoting. For the ones that we do promote, the parameter
638 // attributes are lost
639 SmallVector<AttributeSet, 8> AttributesVec;
640 const AttributeSet &PAL = F->getAttributes();
642 // Add any return attributes.
643 if (PAL.hasAttributes(AttributeSet::ReturnIndex))
644 AttributesVec.push_back(AttributeSet::get(F->getContext(),
645 PAL.getRetAttributes()));
647 // First, determine the new argument list
648 unsigned ArgIndex = 1;
649 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
651 if (ByValArgsToTransform.count(I)) {
652 // Simple byval argument? Just add all the struct element types.
653 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
654 StructType *STy = cast<StructType>(AgTy);
655 Params.insert(Params.end(), STy->element_begin(), STy->element_end());
656 ++NumByValArgsPromoted;
657 } else if (!ArgsToPromote.count(I)) {
658 // Unchanged argument
659 Params.push_back(I->getType());
660 AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
661 if (attrs.hasAttributes(ArgIndex)) {
662 AttrBuilder B(attrs, ArgIndex);
664 push_back(AttributeSet::get(F->getContext(), Params.size(), B));
666 } else if (I->use_empty()) {
667 // Dead argument (which are always marked as promotable)
670 // Okay, this is being promoted. This means that the only uses are loads
671 // or GEPs which are only used by loads
673 // In this table, we will track which indices are loaded from the argument
674 // (where direct loads are tracked as no indices).
675 ScalarizeTable &ArgIndices = ScalarizedElements[I];
676 for (User *U : I->users()) {
677 Instruction *UI = cast<Instruction>(U);
679 if (LoadInst *L = dyn_cast<LoadInst>(UI))
680 SrcTy = L->getType();
682 SrcTy = cast<GetElementPtrInst>(UI)->getSourceElementType();
683 IndicesVector Indices;
684 Indices.reserve(UI->getNumOperands() - 1);
685 // Since loads will only have a single operand, and GEPs only a single
686 // non-index operand, this will record direct loads without any indices,
687 // and gep+loads with the GEP indices.
688 for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end();
690 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
691 // GEPs with a single 0 index can be merged with direct loads
692 if (Indices.size() == 1 && Indices.front() == 0)
694 ArgIndices.insert(std::make_pair(SrcTy, Indices));
696 if (LoadInst *L = dyn_cast<LoadInst>(UI))
699 // Take any load, we will use it only to update Alias Analysis
700 OrigLoad = cast<LoadInst>(UI->user_back());
701 OriginalLoads[std::make_pair(I, Indices)] = OrigLoad;
704 // Add a parameter to the function for each element passed in.
705 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
706 E = ArgIndices.end(); SI != E; ++SI) {
707 // not allowed to dereference ->begin() if size() is 0
708 Params.push_back(GetElementPtrInst::getIndexedType(
709 cast<PointerType>(I->getType()->getScalarType())->getElementType(),
711 assert(Params.back());
714 if (ArgIndices.size() == 1 && ArgIndices.begin()->second.empty())
715 ++NumArgumentsPromoted;
717 ++NumAggregatesPromoted;
721 // Add any function attributes.
722 if (PAL.hasAttributes(AttributeSet::FunctionIndex))
723 AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
724 PAL.getFnAttributes()));
726 Type *RetTy = FTy->getReturnType();
728 // Construct the new function type using the new arguments.
729 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
731 // Create the new function body and insert it into the module.
732 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
733 NF->copyAttributesFrom(F);
735 // Patch the pointer to LLVM function in debug info descriptor.
736 auto DI = FunctionDIs.find(F);
737 if (DI != FunctionDIs.end()) {
738 DISubprogram *SP = DI->second;
739 SP->replaceFunction(NF);
740 // Ensure the map is updated so it can be reused on subsequent argument
741 // promotions of the same function.
742 FunctionDIs.erase(DI);
743 FunctionDIs[NF] = SP;
746 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
749 // Recompute the parameter attributes list based on the new arguments for
751 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
752 AttributesVec.clear();
754 F->getParent()->getFunctionList().insert(F, NF);
757 // Get the callgraph information that we need to update to reflect our
759 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
761 // Get a new callgraph node for NF.
762 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
764 // Loop over all of the callers of the function, transforming the call sites
765 // to pass in the loaded pointers.
767 SmallVector<Value*, 16> Args;
768 while (!F->use_empty()) {
769 CallSite CS(F->user_back());
770 assert(CS.getCalledFunction() == F);
771 Instruction *Call = CS.getInstruction();
772 const AttributeSet &CallPAL = CS.getAttributes();
774 // Add any return attributes.
775 if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
776 AttributesVec.push_back(AttributeSet::get(F->getContext(),
777 CallPAL.getRetAttributes()));
779 // Loop over the operands, inserting GEP and loads in the caller as
781 CallSite::arg_iterator AI = CS.arg_begin();
783 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
784 I != E; ++I, ++AI, ++ArgIndex)
785 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
786 Args.push_back(*AI); // Unmodified argument
788 if (CallPAL.hasAttributes(ArgIndex)) {
789 AttrBuilder B(CallPAL, ArgIndex);
791 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
793 } else if (ByValArgsToTransform.count(I)) {
794 // Emit a GEP and load for each element of the struct.
795 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
796 StructType *STy = cast<StructType>(AgTy);
798 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
799 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
800 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
801 Value *Idx = GetElementPtrInst::Create(
802 STy, *AI, Idxs, (*AI)->getName() + "." + Twine(i), Call);
803 // TODO: Tell AA about the new values?
804 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
806 } else if (!I->use_empty()) {
807 // Non-dead argument: insert GEPs and loads as appropriate.
808 ScalarizeTable &ArgIndices = ScalarizedElements[I];
809 // Store the Value* version of the indices in here, but declare it now
811 std::vector<Value*> Ops;
812 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
813 E = ArgIndices.end(); SI != E; ++SI) {
815 LoadInst *OrigLoad = OriginalLoads[std::make_pair(I, SI->second)];
816 if (!SI->second.empty()) {
817 Ops.reserve(SI->second.size());
818 Type *ElTy = V->getType();
819 for (IndicesVector::const_iterator II = SI->second.begin(),
820 IE = SI->second.end();
822 // Use i32 to index structs, and i64 for others (pointers/arrays).
823 // This satisfies GEP constraints.
824 Type *IdxTy = (ElTy->isStructTy() ?
825 Type::getInt32Ty(F->getContext()) :
826 Type::getInt64Ty(F->getContext()));
827 Ops.push_back(ConstantInt::get(IdxTy, *II));
828 // Keep track of the type we're currently indexing.
829 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
831 // And create a GEP to extract those indices.
832 V = GetElementPtrInst::Create(SI->first, V, Ops,
833 V->getName() + ".idx", Call);
836 // Since we're replacing a load make sure we take the alignment
837 // of the previous load.
838 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
839 newLoad->setAlignment(OrigLoad->getAlignment());
840 // Transfer the AA info too.
842 OrigLoad->getAAMetadata(AAInfo);
843 newLoad->setAAMetadata(AAInfo);
845 Args.push_back(newLoad);
849 // Push any varargs arguments on the list.
850 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
852 if (CallPAL.hasAttributes(ArgIndex)) {
853 AttrBuilder B(CallPAL, ArgIndex);
855 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
859 // Add any function attributes.
860 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
861 AttributesVec.push_back(AttributeSet::get(Call->getContext(),
862 CallPAL.getFnAttributes()));
865 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
866 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
868 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
869 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
872 New = CallInst::Create(NF, Args, "", Call);
873 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
874 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
876 if (cast<CallInst>(Call)->isTailCall())
877 cast<CallInst>(New)->setTailCall();
879 New->setDebugLoc(Call->getDebugLoc());
881 AttributesVec.clear();
883 // Update the callgraph to know that the callsite has been transformed.
884 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
885 CalleeNode->replaceCallEdge(CS, CallSite(New), NF_CGN);
887 if (!Call->use_empty()) {
888 Call->replaceAllUsesWith(New);
892 // Finally, remove the old call from the program, reducing the use-count of
894 Call->eraseFromParent();
897 // Since we have now created the new function, splice the body of the old
898 // function right into the new function, leaving the old rotting hulk of the
900 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
902 // Loop over the argument list, transferring uses of the old arguments over to
903 // the new arguments, also transferring over the names as well.
905 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
906 I2 = NF->arg_begin(); I != E; ++I) {
907 if (!ArgsToPromote.count(I) && !ByValArgsToTransform.count(I)) {
908 // If this is an unmodified argument, move the name and users over to the
910 I->replaceAllUsesWith(I2);
916 if (ByValArgsToTransform.count(I)) {
917 // In the callee, we create an alloca, and store each of the new incoming
918 // arguments into the alloca.
919 Instruction *InsertPt = NF->begin()->begin();
921 // Just add all the struct element types.
922 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
923 Value *TheAlloca = new AllocaInst(AgTy, nullptr, "", InsertPt);
924 StructType *STy = cast<StructType>(AgTy);
926 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
928 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
929 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
930 Value *Idx = GetElementPtrInst::Create(
931 AgTy, TheAlloca, Idxs, TheAlloca->getName() + "." + Twine(i),
933 I2->setName(I->getName()+"."+Twine(i));
934 new StoreInst(I2++, Idx, InsertPt);
937 // Anything that used the arg should now use the alloca.
938 I->replaceAllUsesWith(TheAlloca);
939 TheAlloca->takeName(I);
941 // If the alloca is used in a call, we must clear the tail flag since
942 // the callee now uses an alloca from the caller.
943 for (User *U : TheAlloca->users()) {
944 CallInst *Call = dyn_cast<CallInst>(U);
947 Call->setTailCall(false);
955 // Otherwise, if we promoted this argument, then all users are load
956 // instructions (or GEPs with only load users), and all loads should be
957 // using the new argument that we added.
958 ScalarizeTable &ArgIndices = ScalarizedElements[I];
960 while (!I->use_empty()) {
961 if (LoadInst *LI = dyn_cast<LoadInst>(I->user_back())) {
962 assert(ArgIndices.begin()->second.empty() &&
963 "Load element should sort to front!");
964 I2->setName(I->getName()+".val");
965 LI->replaceAllUsesWith(I2);
966 LI->eraseFromParent();
967 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
968 << "' in function '" << F->getName() << "'\n");
970 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_back());
971 IndicesVector Operands;
972 Operands.reserve(GEP->getNumIndices());
973 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
975 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
977 // GEPs with a single 0 index can be merged with direct loads
978 if (Operands.size() == 1 && Operands.front() == 0)
981 Function::arg_iterator TheArg = I2;
982 for (ScalarizeTable::iterator It = ArgIndices.begin();
983 It->second != Operands; ++It, ++TheArg) {
984 assert(It != ArgIndices.end() && "GEP not handled??");
987 std::string NewName = I->getName();
988 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
989 NewName += "." + utostr(Operands[i]);
992 TheArg->setName(NewName);
994 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
995 << "' of function '" << NF->getName() << "'\n");
997 // All of the uses must be load instructions. Replace them all with
998 // the argument specified by ArgNo.
999 while (!GEP->use_empty()) {
1000 LoadInst *L = cast<LoadInst>(GEP->user_back());
1001 L->replaceAllUsesWith(TheArg);
1002 L->eraseFromParent();
1004 GEP->eraseFromParent();
1008 // Increment I2 past all of the arguments added for this promoted pointer.
1009 std::advance(I2, ArgIndices.size());
1012 NF_CGN->stealCalledFunctionsFrom(CG[F]);
1014 // Now that the old function is dead, delete it. If there is a dangling
1015 // reference to the CallgraphNode, just leave the dead function around for
1016 // someone else to nuke.
1017 CallGraphNode *CGN = CG[F];
1018 if (CGN->getNumReferences() == 0)
1019 delete CG.removeFunctionFromModule(CGN);
1021 F->setLinkage(Function::ExternalLinkage);
1026 bool ArgPromotion::doInitialization(CallGraph &CG) {
1027 FunctionDIs = makeSubprogramMap(CG.getModule());
1028 return CallGraphSCCPass::doInitialization(CG);