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/Loads.h"
42 #include "llvm/Analysis/TargetLibraryInfo.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 getAAResultsAnalysisUsage(AU);
72 CallGraphSCCPass::getAnalysisUsage(AU);
75 bool runOnSCC(CallGraphSCC &SCC) override;
76 static char ID; // Pass identification, replacement for typeid
77 explicit ArgPromotion(unsigned maxElements = 3)
78 : CallGraphSCCPass(ID), maxElements(maxElements) {
79 initializeArgPromotionPass(*PassRegistry::getPassRegistry());
84 using llvm::Pass::doInitialization;
85 bool doInitialization(CallGraph &CG) override;
86 /// The maximum number of elements to expand, or 0 for unlimited.
91 /// A vector used to hold the indices of a single GEP instruction
92 typedef std::vector<uint64_t> IndicesVector;
94 static CallGraphNode *
95 PromoteArguments(CallGraphNode *CGN, CallGraph &CG,
96 function_ref<AAResults &(Function &F)> AARGetter,
97 unsigned MaxElements);
98 static bool isDenselyPacked(Type *type, const DataLayout &DL);
99 static bool canPaddingBeAccessed(Argument *Arg);
100 static bool isSafeToPromoteArgument(Argument *Arg, bool isByVal, AAResults &AAR,
101 unsigned MaxElements);
102 static CallGraphNode *
103 DoPromotion(Function *F, SmallPtrSetImpl<Argument *> &ArgsToPromote,
104 SmallPtrSetImpl<Argument *> &ByValArgsToTransform, CallGraph &CG);
106 char ArgPromotion::ID = 0;
107 INITIALIZE_PASS_BEGIN(ArgPromotion, "argpromotion",
108 "Promote 'by reference' arguments to scalars", false, false)
109 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
110 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
111 INITIALIZE_PASS_DEPENDENCY(TargetLibraryInfoWrapperPass)
112 INITIALIZE_PASS_END(ArgPromotion, "argpromotion",
113 "Promote 'by reference' arguments to scalars", false, false)
115 Pass *llvm::createArgumentPromotionPass(unsigned maxElements) {
116 return new ArgPromotion(maxElements);
119 static bool runImpl(CallGraphSCC &SCC, CallGraph &CG,
120 function_ref<AAResults &(Function &F)> AARGetter,
121 unsigned MaxElements) {
122 bool Changed = false, LocalChange;
124 do { // Iterate until we stop promoting from this SCC.
126 // Attempt to promote arguments from all functions in this SCC.
127 for (CallGraphNode *OldNode : SCC) {
128 if (CallGraphNode *NewNode =
129 PromoteArguments(OldNode, CG, AARGetter, MaxElements)) {
131 SCC.ReplaceNode(OldNode, NewNode);
134 Changed |= LocalChange; // Remember that we changed something.
135 } while (LocalChange);
140 bool ArgPromotion::runOnSCC(CallGraphSCC &SCC) {
144 // Get the callgraph information that we need to update to reflect our
146 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
148 // We compute dedicated AA results for each function in the SCC as needed. We
149 // use a lambda referencing external objects so that they live long enough to
150 // be queried, but we re-use them each time.
151 Optional<BasicAAResult> BAR;
152 Optional<AAResults> AAR;
153 auto AARGetter = [&](Function &F) -> AAResults & {
154 BAR.emplace(createLegacyPMBasicAAResult(*this, F));
155 AAR.emplace(createLegacyPMAAResults(*this, F, *BAR));
159 return runImpl(SCC, CG, AARGetter, maxElements);
162 /// \brief Checks if a type could have padding bytes.
163 static bool isDenselyPacked(Type *type, const DataLayout &DL) {
165 // There is no size information, so be conservative.
166 if (!type->isSized())
169 // If the alloc size is not equal to the storage size, then there are padding
170 // bytes. For x86_fp80 on x86-64, size: 80 alloc size: 128.
171 if (DL.getTypeSizeInBits(type) != DL.getTypeAllocSizeInBits(type))
174 if (!isa<CompositeType>(type))
177 // For homogenous sequential types, check for padding within members.
178 if (SequentialType *seqTy = dyn_cast<SequentialType>(type))
179 return isDenselyPacked(seqTy->getElementType(), DL);
181 // Check for padding within and between elements of a struct.
182 StructType *StructTy = cast<StructType>(type);
183 const StructLayout *Layout = DL.getStructLayout(StructTy);
184 uint64_t StartPos = 0;
185 for (unsigned i = 0, E = StructTy->getNumElements(); i < E; ++i) {
186 Type *ElTy = StructTy->getElementType(i);
187 if (!isDenselyPacked(ElTy, DL))
189 if (StartPos != Layout->getElementOffsetInBits(i))
191 StartPos += DL.getTypeAllocSizeInBits(ElTy);
197 /// \brief Checks if the padding bytes of an argument could be accessed.
198 static bool canPaddingBeAccessed(Argument *arg) {
200 assert(arg->hasByValAttr());
202 // Track all the pointers to the argument to make sure they are not captured.
203 SmallPtrSet<Value *, 16> PtrValues;
204 PtrValues.insert(arg);
206 // Track all of the stores.
207 SmallVector<StoreInst *, 16> Stores;
209 // Scan through the uses recursively to make sure the pointer is always used
211 SmallVector<Value *, 16> WorkList;
212 WorkList.insert(WorkList.end(), arg->user_begin(), arg->user_end());
213 while (!WorkList.empty()) {
214 Value *V = WorkList.back();
216 if (isa<GetElementPtrInst>(V) || isa<PHINode>(V)) {
217 if (PtrValues.insert(V).second)
218 WorkList.insert(WorkList.end(), V->user_begin(), V->user_end());
219 } else if (StoreInst *Store = dyn_cast<StoreInst>(V)) {
220 Stores.push_back(Store);
221 } else if (!isa<LoadInst>(V)) {
226 // Check to make sure the pointers aren't captured
227 for (StoreInst *Store : Stores)
228 if (PtrValues.count(Store->getValueOperand()))
234 /// PromoteArguments - This method checks the specified function to see if there
235 /// are any promotable arguments and if it is safe to promote the function (for
236 /// example, all callers are direct). If safe to promote some arguments, it
237 /// calls the DoPromotion method.
239 static CallGraphNode *
240 PromoteArguments(CallGraphNode *CGN, CallGraph &CG,
241 function_ref<AAResults &(Function &F)> AARGetter,
242 unsigned MaxElements) {
243 Function *F = CGN->getFunction();
245 // Make sure that it is local to this module.
246 if (!F || !F->hasLocalLinkage()) return nullptr;
248 // Don't promote arguments for variadic functions. Adding, removing, or
249 // changing non-pack parameters can change the classification of pack
250 // parameters. Frontends encode that classification at the call site in the
251 // IR, while in the callee the classification is determined dynamically based
252 // on the number of registers consumed so far.
253 if (F->isVarArg()) return nullptr;
255 // First check: see if there are any pointer arguments! If not, quick exit.
256 SmallVector<Argument*, 16> PointerArgs;
257 for (Argument &I : F->args())
258 if (I.getType()->isPointerTy())
259 PointerArgs.push_back(&I);
260 if (PointerArgs.empty()) return nullptr;
262 // Second check: make sure that all callers are direct callers. We can't
263 // transform functions that have indirect callers. Also see if the function
264 // is self-recursive.
265 bool isSelfRecursive = false;
266 for (Use &U : F->uses()) {
267 CallSite CS(U.getUser());
268 // Must be a direct call.
269 if (CS.getInstruction() == nullptr || !CS.isCallee(&U)) return nullptr;
271 if (CS.getInstruction()->getParent()->getParent() == F)
272 isSelfRecursive = true;
275 const DataLayout &DL = F->getParent()->getDataLayout();
277 AAResults &AAR = AARGetter(*F);
279 // Check to see which arguments are promotable. If an argument is promotable,
280 // add it to ArgsToPromote.
281 SmallPtrSet<Argument*, 8> ArgsToPromote;
282 SmallPtrSet<Argument*, 8> ByValArgsToTransform;
283 for (Argument *PtrArg : PointerArgs) {
284 Type *AgTy = cast<PointerType>(PtrArg->getType())->getElementType();
286 // Replace sret attribute with noalias. This reduces register pressure by
287 // avoiding a register copy.
288 if (PtrArg->hasStructRetAttr()) {
289 unsigned ArgNo = PtrArg->getArgNo();
292 .removeAttribute(F->getContext(), ArgNo + 1, Attribute::StructRet)
293 .addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
294 for (Use &U : F->uses()) {
295 CallSite CS(U.getUser());
298 .removeAttribute(F->getContext(), ArgNo + 1,
299 Attribute::StructRet)
300 .addAttribute(F->getContext(), ArgNo + 1, Attribute::NoAlias));
304 // If this is a byval argument, and if the aggregate type is small, just
305 // pass the elements, which is always safe, if the passed value is densely
306 // packed or if we can prove the padding bytes are never accessed. This does
307 // not apply to inalloca.
308 bool isSafeToPromote =
309 PtrArg->hasByValAttr() &&
310 (isDenselyPacked(AgTy, DL) || !canPaddingBeAccessed(PtrArg));
311 if (isSafeToPromote) {
312 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
313 if (MaxElements > 0 && STy->getNumElements() > MaxElements) {
314 DEBUG(dbgs() << "argpromotion disable promoting argument '"
315 << PtrArg->getName() << "' because it would require adding more"
316 << " than " << MaxElements << " arguments to the function.\n");
320 // If all the elements are single-value types, we can promote it.
321 bool AllSimple = true;
322 for (const auto *EltTy : STy->elements()) {
323 if (!EltTy->isSingleValueType()) {
329 // Safe to transform, don't even bother trying to "promote" it.
330 // Passing the elements as a scalar will allow sroa to hack on
331 // the new alloca we introduce.
333 ByValArgsToTransform.insert(PtrArg);
339 // If the argument is a recursive type and we're in a recursive
340 // function, we could end up infinitely peeling the function argument.
341 if (isSelfRecursive) {
342 if (StructType *STy = dyn_cast<StructType>(AgTy)) {
343 bool RecursiveType = false;
344 for (const auto *EltTy : STy->elements()) {
345 if (EltTy == PtrArg->getType()) {
346 RecursiveType = true;
355 // Otherwise, see if we can promote the pointer to its value.
356 if (isSafeToPromoteArgument(PtrArg, PtrArg->hasByValOrInAllocaAttr(), AAR,
358 ArgsToPromote.insert(PtrArg);
361 // No promotable pointer arguments.
362 if (ArgsToPromote.empty() && ByValArgsToTransform.empty())
365 return DoPromotion(F, ArgsToPromote, ByValArgsToTransform, CG);
368 /// AllCallersPassInValidPointerForArgument - Return true if we can prove that
369 /// all callees pass in a valid pointer for the specified function argument.
370 static bool AllCallersPassInValidPointerForArgument(Argument *Arg) {
371 Function *Callee = Arg->getParent();
372 const DataLayout &DL = Callee->getParent()->getDataLayout();
374 unsigned ArgNo = Arg->getArgNo();
376 // Look at all call sites of the function. At this point we know we only have
378 for (User *U : Callee->users()) {
380 assert(CS && "Should only have direct calls!");
382 if (!isDereferenceablePointer(CS.getArgument(ArgNo), DL))
388 /// Returns true if Prefix is a prefix of longer. That means, Longer has a size
389 /// that is greater than or equal to the size of prefix, and each of the
390 /// elements in Prefix is the same as the corresponding elements in Longer.
392 /// This means it also returns true when Prefix and Longer are equal!
393 static bool IsPrefix(const IndicesVector &Prefix, const IndicesVector &Longer) {
394 if (Prefix.size() > Longer.size())
396 return std::equal(Prefix.begin(), Prefix.end(), Longer.begin());
400 /// Checks if Indices, or a prefix of Indices, is in Set.
401 static bool PrefixIn(const IndicesVector &Indices,
402 std::set<IndicesVector> &Set) {
403 std::set<IndicesVector>::iterator Low;
404 Low = Set.upper_bound(Indices);
405 if (Low != Set.begin())
407 // Low is now the last element smaller than or equal to Indices. This means
408 // it points to a prefix of Indices (possibly Indices itself), if such
411 // This load is safe if any prefix of its operands is safe to load.
412 return Low != Set.end() && IsPrefix(*Low, Indices);
415 /// Mark the given indices (ToMark) as safe in the given set of indices
416 /// (Safe). Marking safe usually means adding ToMark to Safe. However, if there
417 /// is already a prefix of Indices in Safe, Indices are implicitely marked safe
418 /// already. Furthermore, any indices that Indices is itself a prefix of, are
419 /// removed from Safe (since they are implicitely safe because of Indices now).
420 static void MarkIndicesSafe(const IndicesVector &ToMark,
421 std::set<IndicesVector> &Safe) {
422 std::set<IndicesVector>::iterator Low;
423 Low = Safe.upper_bound(ToMark);
424 // Guard against the case where Safe is empty
425 if (Low != Safe.begin())
427 // Low is now the last element smaller than or equal to Indices. This
428 // means it points to a prefix of Indices (possibly Indices itself), if
429 // such prefix exists.
430 if (Low != Safe.end()) {
431 if (IsPrefix(*Low, ToMark))
432 // If there is already a prefix of these indices (or exactly these
433 // indices) marked a safe, don't bother adding these indices
436 // Increment Low, so we can use it as a "insert before" hint
440 Low = Safe.insert(Low, ToMark);
442 // If there we're a prefix of longer index list(s), remove those
443 std::set<IndicesVector>::iterator End = Safe.end();
444 while (Low != End && IsPrefix(ToMark, *Low)) {
445 std::set<IndicesVector>::iterator Remove = Low;
451 /// isSafeToPromoteArgument - As you might guess from the name of this method,
452 /// it checks to see if it is both safe and useful to promote the argument.
453 /// This method limits promotion of aggregates to only promote up to three
454 /// elements of the aggregate in order to avoid exploding the number of
455 /// arguments passed in.
456 static bool isSafeToPromoteArgument(Argument *Arg, bool isByValOrInAlloca,
457 AAResults &AAR, unsigned MaxElements) {
458 typedef std::set<IndicesVector> GEPIndicesSet;
460 // Quick exit for unused arguments
461 if (Arg->use_empty())
464 // We can only promote this argument if all of the uses are loads, or are GEP
465 // instructions (with constant indices) that are subsequently loaded.
467 // Promoting the argument causes it to be loaded in the caller
468 // unconditionally. This is only safe if we can prove that either the load
469 // would have happened in the callee anyway (ie, there is a load in the entry
470 // block) or the pointer passed in at every call site is guaranteed to be
472 // In the former case, invalid loads can happen, but would have happened
473 // anyway, in the latter case, invalid loads won't happen. This prevents us
474 // from introducing an invalid load that wouldn't have happened in the
477 // This set will contain all sets of indices that are loaded in the entry
478 // block, and thus are safe to unconditionally load in the caller.
480 // This optimization is also safe for InAlloca parameters, because it verifies
481 // that the address isn't captured.
482 GEPIndicesSet SafeToUnconditionallyLoad;
484 // This set contains all the sets of indices that we are planning to promote.
485 // This makes it possible to limit the number of arguments added.
486 GEPIndicesSet ToPromote;
488 // If the pointer is always valid, any load with first index 0 is valid.
489 if (isByValOrInAlloca || AllCallersPassInValidPointerForArgument(Arg))
490 SafeToUnconditionallyLoad.insert(IndicesVector(1, 0));
492 // First, iterate the entry block and mark loads of (geps of) arguments as
494 BasicBlock &EntryBlock = Arg->getParent()->front();
495 // Declare this here so we can reuse it
496 IndicesVector Indices;
497 for (Instruction &I : EntryBlock)
498 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
499 Value *V = LI->getPointerOperand();
500 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
501 V = GEP->getPointerOperand();
503 // This load actually loads (part of) Arg? Check the indices then.
504 Indices.reserve(GEP->getNumIndices());
505 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
507 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
508 Indices.push_back(CI->getSExtValue());
510 // We found a non-constant GEP index for this argument? Bail out
511 // right away, can't promote this argument at all.
514 // Indices checked out, mark them as safe
515 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
518 } else if (V == Arg) {
519 // Direct loads are equivalent to a GEP with a single 0 index.
520 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
524 // Now, iterate all uses of the argument to see if there are any uses that are
525 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
526 SmallVector<LoadInst*, 16> Loads;
527 IndicesVector Operands;
528 for (Use &U : Arg->uses()) {
529 User *UR = U.getUser();
531 if (LoadInst *LI = dyn_cast<LoadInst>(UR)) {
532 // Don't hack volatile/atomic loads
533 if (!LI->isSimple()) return false;
535 // Direct loads are equivalent to a GEP with a zero index and then a load.
536 Operands.push_back(0);
537 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UR)) {
538 if (GEP->use_empty()) {
539 // Dead GEP's cause trouble later. Just remove them if we run into
541 GEP->eraseFromParent();
542 // TODO: This runs the above loop over and over again for dead GEPs
543 // Couldn't we just do increment the UI iterator earlier and erase the
545 return isSafeToPromoteArgument(Arg, isByValOrInAlloca, AAR,
549 // Ensure that all of the indices are constants.
550 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
552 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
553 Operands.push_back(C->getSExtValue());
555 return false; // Not a constant operand GEP!
557 // Ensure that the only users of the GEP are load instructions.
558 for (User *GEPU : GEP->users())
559 if (LoadInst *LI = dyn_cast<LoadInst>(GEPU)) {
560 // Don't hack volatile/atomic loads
561 if (!LI->isSimple()) return false;
564 // Other uses than load?
568 return false; // Not a load or a GEP.
571 // Now, see if it is safe to promote this load / loads of this GEP. Loading
572 // is safe if Operands, or a prefix of Operands, is marked as safe.
573 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
576 // See if we are already promoting a load with these indices. If not, check
577 // to make sure that we aren't promoting too many elements. If so, nothing
579 if (ToPromote.find(Operands) == ToPromote.end()) {
580 if (MaxElements > 0 && ToPromote.size() == MaxElements) {
581 DEBUG(dbgs() << "argpromotion not promoting argument '"
582 << Arg->getName() << "' because it would require adding more "
583 << "than " << MaxElements << " arguments to the function.\n");
584 // We limit aggregate promotion to only promoting up to a fixed number
585 // of elements of the aggregate.
588 ToPromote.insert(std::move(Operands));
592 if (Loads.empty()) return true; // No users, this is a dead argument.
594 // Okay, now we know that the argument is only used by load instructions and
595 // it is safe to unconditionally perform all of them. Use alias analysis to
596 // check to see if the pointer is guaranteed to not be modified from entry of
597 // the function to each of the load instructions.
599 // Because there could be several/many load instructions, remember which
600 // blocks we know to be transparent to the load.
601 df_iterator_default_set<BasicBlock*, 16> TranspBlocks;
603 for (LoadInst *Load : Loads) {
604 // Check to see if the load is invalidated from the start of the block to
606 BasicBlock *BB = Load->getParent();
608 MemoryLocation Loc = MemoryLocation::get(Load);
609 if (AAR.canInstructionRangeModRef(BB->front(), *Load, Loc, MRI_Mod))
610 return false; // Pointer is invalidated!
612 // Now check every path from the entry block to the load for transparency.
613 // To do this, we perform a depth first search on the inverse CFG from the
615 for (BasicBlock *P : predecessors(BB)) {
616 for (BasicBlock *TranspBB : inverse_depth_first_ext(P, TranspBlocks))
617 if (AAR.canBasicBlockModify(*TranspBB, Loc))
622 // If the path from the entry of the function to each load is free of
623 // instructions that potentially invalidate the load, we can make the
628 /// DoPromotion - This method actually performs the promotion of the specified
629 /// arguments, and returns the new function. At this point, we know that it's
631 static CallGraphNode *
632 DoPromotion(Function *F, SmallPtrSetImpl<Argument *> &ArgsToPromote,
633 SmallPtrSetImpl<Argument *> &ByValArgsToTransform, CallGraph &CG) {
635 // Start by computing a new prototype for the function, which is the same as
636 // the old function, but has modified arguments.
637 FunctionType *FTy = F->getFunctionType();
638 std::vector<Type*> Params;
640 typedef std::set<std::pair<Type *, IndicesVector>> ScalarizeTable;
642 // ScalarizedElements - If we are promoting a pointer that has elements
643 // accessed out of it, keep track of which elements are accessed so that we
644 // can add one argument for each.
646 // Arguments that are directly loaded will have a zero element value here, to
647 // handle cases where there are both a direct load and GEP accesses.
649 std::map<Argument*, ScalarizeTable> ScalarizedElements;
651 // OriginalLoads - Keep track of a representative load instruction from the
652 // original function so that we can tell the alias analysis implementation
653 // what the new GEP/Load instructions we are inserting look like.
654 // We need to keep the original loads for each argument and the elements
655 // of the argument that are accessed.
656 std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;
658 // Attribute - Keep track of the parameter attributes for the arguments
659 // that we are *not* promoting. For the ones that we do promote, the parameter
660 // attributes are lost
661 SmallVector<AttributeSet, 8> AttributesVec;
662 const AttributeSet &PAL = F->getAttributes();
664 // Add any return attributes.
665 if (PAL.hasAttributes(AttributeSet::ReturnIndex))
666 AttributesVec.push_back(AttributeSet::get(F->getContext(),
667 PAL.getRetAttributes()));
669 // First, determine the new argument list
670 unsigned ArgIndex = 1;
671 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
673 if (ByValArgsToTransform.count(&*I)) {
674 // Simple byval argument? Just add all the struct element types.
675 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
676 StructType *STy = cast<StructType>(AgTy);
677 Params.insert(Params.end(), STy->element_begin(), STy->element_end());
678 ++NumByValArgsPromoted;
679 } else if (!ArgsToPromote.count(&*I)) {
680 // Unchanged argument
681 Params.push_back(I->getType());
682 AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
683 if (attrs.hasAttributes(ArgIndex)) {
684 AttrBuilder B(attrs, ArgIndex);
686 push_back(AttributeSet::get(F->getContext(), Params.size(), B));
688 } else if (I->use_empty()) {
689 // Dead argument (which are always marked as promotable)
692 // Okay, this is being promoted. This means that the only uses are loads
693 // or GEPs which are only used by loads
695 // In this table, we will track which indices are loaded from the argument
696 // (where direct loads are tracked as no indices).
697 ScalarizeTable &ArgIndices = ScalarizedElements[&*I];
698 for (User *U : I->users()) {
699 Instruction *UI = cast<Instruction>(U);
701 if (LoadInst *L = dyn_cast<LoadInst>(UI))
702 SrcTy = L->getType();
704 SrcTy = cast<GetElementPtrInst>(UI)->getSourceElementType();
705 IndicesVector Indices;
706 Indices.reserve(UI->getNumOperands() - 1);
707 // Since loads will only have a single operand, and GEPs only a single
708 // non-index operand, this will record direct loads without any indices,
709 // and gep+loads with the GEP indices.
710 for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end();
712 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
713 // GEPs with a single 0 index can be merged with direct loads
714 if (Indices.size() == 1 && Indices.front() == 0)
716 ArgIndices.insert(std::make_pair(SrcTy, Indices));
718 if (LoadInst *L = dyn_cast<LoadInst>(UI))
721 // Take any load, we will use it only to update Alias Analysis
722 OrigLoad = cast<LoadInst>(UI->user_back());
723 OriginalLoads[std::make_pair(&*I, Indices)] = OrigLoad;
726 // Add a parameter to the function for each element passed in.
727 for (const auto &ArgIndex : ArgIndices) {
728 // not allowed to dereference ->begin() if size() is 0
729 Params.push_back(GetElementPtrInst::getIndexedType(
730 cast<PointerType>(I->getType()->getScalarType())->getElementType(),
732 assert(Params.back());
735 if (ArgIndices.size() == 1 && ArgIndices.begin()->second.empty())
736 ++NumArgumentsPromoted;
738 ++NumAggregatesPromoted;
742 // Add any function attributes.
743 if (PAL.hasAttributes(AttributeSet::FunctionIndex))
744 AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
745 PAL.getFnAttributes()));
747 Type *RetTy = FTy->getReturnType();
749 // Construct the new function type using the new arguments.
750 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
752 // Create the new function body and insert it into the module.
753 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
754 NF->copyAttributesFrom(F);
756 // Patch the pointer to LLVM function in debug info descriptor.
757 NF->setSubprogram(F->getSubprogram());
758 F->setSubprogram(nullptr);
760 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
763 // Recompute the parameter attributes list based on the new arguments for
765 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
766 AttributesVec.clear();
768 F->getParent()->getFunctionList().insert(F->getIterator(), NF);
771 // Get a new callgraph node for NF.
772 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
774 // Loop over all of the callers of the function, transforming the call sites
775 // to pass in the loaded pointers.
777 SmallVector<Value*, 16> Args;
778 while (!F->use_empty()) {
779 CallSite CS(F->user_back());
780 assert(CS.getCalledFunction() == F);
781 Instruction *Call = CS.getInstruction();
782 const AttributeSet &CallPAL = CS.getAttributes();
784 // Add any return attributes.
785 if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
786 AttributesVec.push_back(AttributeSet::get(F->getContext(),
787 CallPAL.getRetAttributes()));
789 // Loop over the operands, inserting GEP and loads in the caller as
791 CallSite::arg_iterator AI = CS.arg_begin();
793 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
794 I != E; ++I, ++AI, ++ArgIndex)
795 if (!ArgsToPromote.count(&*I) && !ByValArgsToTransform.count(&*I)) {
796 Args.push_back(*AI); // Unmodified argument
798 if (CallPAL.hasAttributes(ArgIndex)) {
799 AttrBuilder B(CallPAL, ArgIndex);
801 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
803 } else if (ByValArgsToTransform.count(&*I)) {
804 // Emit a GEP and load for each element of the struct.
805 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
806 StructType *STy = cast<StructType>(AgTy);
808 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
809 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
810 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
811 Value *Idx = GetElementPtrInst::Create(
812 STy, *AI, Idxs, (*AI)->getName() + "." + Twine(i), Call);
813 // TODO: Tell AA about the new values?
814 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
816 } else if (!I->use_empty()) {
817 // Non-dead argument: insert GEPs and loads as appropriate.
818 ScalarizeTable &ArgIndices = ScalarizedElements[&*I];
819 // Store the Value* version of the indices in here, but declare it now
821 std::vector<Value*> Ops;
822 for (const auto &ArgIndex : ArgIndices) {
825 OriginalLoads[std::make_pair(&*I, ArgIndex.second)];
826 if (!ArgIndex.second.empty()) {
827 Ops.reserve(ArgIndex.second.size());
828 Type *ElTy = V->getType();
829 for (unsigned long II : ArgIndex.second) {
830 // Use i32 to index structs, and i64 for others (pointers/arrays).
831 // This satisfies GEP constraints.
832 Type *IdxTy = (ElTy->isStructTy() ?
833 Type::getInt32Ty(F->getContext()) :
834 Type::getInt64Ty(F->getContext()));
835 Ops.push_back(ConstantInt::get(IdxTy, II));
836 // Keep track of the type we're currently indexing.
837 if (auto *ElPTy = dyn_cast<PointerType>(ElTy))
838 ElTy = ElPTy->getElementType();
840 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(II);
842 // And create a GEP to extract those indices.
843 V = GetElementPtrInst::Create(ArgIndex.first, V, Ops,
844 V->getName() + ".idx", Call);
847 // Since we're replacing a load make sure we take the alignment
848 // of the previous load.
849 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
850 newLoad->setAlignment(OrigLoad->getAlignment());
851 // Transfer the AA info too.
853 OrigLoad->getAAMetadata(AAInfo);
854 newLoad->setAAMetadata(AAInfo);
856 Args.push_back(newLoad);
860 // Push any varargs arguments on the list.
861 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
863 if (CallPAL.hasAttributes(ArgIndex)) {
864 AttrBuilder B(CallPAL, ArgIndex);
866 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
870 // Add any function attributes.
871 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
872 AttributesVec.push_back(AttributeSet::get(Call->getContext(),
873 CallPAL.getFnAttributes()));
875 SmallVector<OperandBundleDef, 1> OpBundles;
876 CS.getOperandBundlesAsDefs(OpBundles);
879 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
880 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
881 Args, OpBundles, "", Call);
882 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
883 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
886 New = CallInst::Create(NF, Args, OpBundles, "", Call);
887 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
888 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
890 cast<CallInst>(New)->setTailCallKind(
891 cast<CallInst>(Call)->getTailCallKind());
893 New->setDebugLoc(Call->getDebugLoc());
895 AttributesVec.clear();
897 // Update the callgraph to know that the callsite has been transformed.
898 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
899 CalleeNode->replaceCallEdge(CS, CallSite(New), NF_CGN);
901 if (!Call->use_empty()) {
902 Call->replaceAllUsesWith(New);
906 // Finally, remove the old call from the program, reducing the use-count of
908 Call->eraseFromParent();
911 // Since we have now created the new function, splice the body of the old
912 // function right into the new function, leaving the old rotting hulk of the
914 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
916 // Loop over the argument list, transferring uses of the old arguments over to
917 // the new arguments, also transferring over the names as well.
919 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
920 I2 = NF->arg_begin(); I != E; ++I) {
921 if (!ArgsToPromote.count(&*I) && !ByValArgsToTransform.count(&*I)) {
922 // If this is an unmodified argument, move the name and users over to the
924 I->replaceAllUsesWith(&*I2);
930 if (ByValArgsToTransform.count(&*I)) {
931 // In the callee, we create an alloca, and store each of the new incoming
932 // arguments into the alloca.
933 Instruction *InsertPt = &NF->begin()->front();
935 // Just add all the struct element types.
936 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
937 Value *TheAlloca = new AllocaInst(AgTy, nullptr, "", InsertPt);
938 StructType *STy = cast<StructType>(AgTy);
940 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
942 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
943 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
944 Value *Idx = GetElementPtrInst::Create(
945 AgTy, TheAlloca, Idxs, TheAlloca->getName() + "." + Twine(i),
947 I2->setName(I->getName()+"."+Twine(i));
948 new StoreInst(&*I2++, Idx, InsertPt);
951 // Anything that used the arg should now use the alloca.
952 I->replaceAllUsesWith(TheAlloca);
953 TheAlloca->takeName(&*I);
955 // If the alloca is used in a call, we must clear the tail flag since
956 // the callee now uses an alloca from the caller.
957 for (User *U : TheAlloca->users()) {
958 CallInst *Call = dyn_cast<CallInst>(U);
961 Call->setTailCall(false);
969 // Otherwise, if we promoted this argument, then all users are load
970 // instructions (or GEPs with only load users), and all loads should be
971 // using the new argument that we added.
972 ScalarizeTable &ArgIndices = ScalarizedElements[&*I];
974 while (!I->use_empty()) {
975 if (LoadInst *LI = dyn_cast<LoadInst>(I->user_back())) {
976 assert(ArgIndices.begin()->second.empty() &&
977 "Load element should sort to front!");
978 I2->setName(I->getName()+".val");
979 LI->replaceAllUsesWith(&*I2);
980 LI->eraseFromParent();
981 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
982 << "' in function '" << F->getName() << "'\n");
984 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_back());
985 IndicesVector Operands;
986 Operands.reserve(GEP->getNumIndices());
987 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
989 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
991 // GEPs with a single 0 index can be merged with direct loads
992 if (Operands.size() == 1 && Operands.front() == 0)
995 Function::arg_iterator TheArg = I2;
996 for (ScalarizeTable::iterator It = ArgIndices.begin();
997 It->second != Operands; ++It, ++TheArg) {
998 assert(It != ArgIndices.end() && "GEP not handled??");
1001 std::string NewName = I->getName();
1002 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
1003 NewName += "." + utostr(Operands[i]);
1006 TheArg->setName(NewName);
1008 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
1009 << "' of function '" << NF->getName() << "'\n");
1011 // All of the uses must be load instructions. Replace them all with
1012 // the argument specified by ArgNo.
1013 while (!GEP->use_empty()) {
1014 LoadInst *L = cast<LoadInst>(GEP->user_back());
1015 L->replaceAllUsesWith(&*TheArg);
1016 L->eraseFromParent();
1018 GEP->eraseFromParent();
1022 // Increment I2 past all of the arguments added for this promoted pointer.
1023 std::advance(I2, ArgIndices.size());
1026 NF_CGN->stealCalledFunctionsFrom(CG[F]);
1028 // Now that the old function is dead, delete it. If there is a dangling
1029 // reference to the CallgraphNode, just leave the dead function around for
1030 // someone else to nuke.
1031 CallGraphNode *CGN = CG[F];
1032 if (CGN->getNumReferences() == 0)
1033 delete CG.removeFunctionFromModule(CGN);
1035 F->setLinkage(Function::ExternalLinkage);
1040 bool ArgPromotion::doInitialization(CallGraph &CG) {
1041 return CallGraphSCCPass::doInitialization(CG);