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 (Argument &I : F->args())
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()->front();
472 // Declare this here so we can reuse it
473 IndicesVector Indices;
474 for (Instruction &I : EntryBlock)
475 if (LoadInst *LI = dyn_cast<LoadInst>(&I)) {
476 Value *V = LI->getPointerOperand();
477 if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(V)) {
478 V = GEP->getPointerOperand();
480 // This load actually loads (part of) Arg? Check the indices then.
481 Indices.reserve(GEP->getNumIndices());
482 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
484 if (ConstantInt *CI = dyn_cast<ConstantInt>(*II))
485 Indices.push_back(CI->getSExtValue());
487 // We found a non-constant GEP index for this argument? Bail out
488 // right away, can't promote this argument at all.
491 // Indices checked out, mark them as safe
492 MarkIndicesSafe(Indices, SafeToUnconditionallyLoad);
495 } else if (V == Arg) {
496 // Direct loads are equivalent to a GEP with a single 0 index.
497 MarkIndicesSafe(IndicesVector(1, 0), SafeToUnconditionallyLoad);
501 // Now, iterate all uses of the argument to see if there are any uses that are
502 // not (GEP+)loads, or any (GEP+)loads that are not safe to promote.
503 SmallVector<LoadInst*, 16> Loads;
504 IndicesVector Operands;
505 for (Use &U : Arg->uses()) {
506 User *UR = U.getUser();
508 if (LoadInst *LI = dyn_cast<LoadInst>(UR)) {
509 // Don't hack volatile/atomic loads
510 if (!LI->isSimple()) return false;
512 // Direct loads are equivalent to a GEP with a zero index and then a load.
513 Operands.push_back(0);
514 } else if (GetElementPtrInst *GEP = dyn_cast<GetElementPtrInst>(UR)) {
515 if (GEP->use_empty()) {
516 // Dead GEP's cause trouble later. Just remove them if we run into
518 GEP->eraseFromParent();
519 // TODO: This runs the above loop over and over again for dead GEPs
520 // Couldn't we just do increment the UI iterator earlier and erase the
522 return isSafeToPromoteArgument(Arg, isByValOrInAlloca, AAR);
525 // Ensure that all of the indices are constants.
526 for (User::op_iterator i = GEP->idx_begin(), e = GEP->idx_end();
528 if (ConstantInt *C = dyn_cast<ConstantInt>(*i))
529 Operands.push_back(C->getSExtValue());
531 return false; // Not a constant operand GEP!
533 // Ensure that the only users of the GEP are load instructions.
534 for (User *GEPU : GEP->users())
535 if (LoadInst *LI = dyn_cast<LoadInst>(GEPU)) {
536 // Don't hack volatile/atomic loads
537 if (!LI->isSimple()) return false;
540 // Other uses than load?
544 return false; // Not a load or a GEP.
547 // Now, see if it is safe to promote this load / loads of this GEP. Loading
548 // is safe if Operands, or a prefix of Operands, is marked as safe.
549 if (!PrefixIn(Operands, SafeToUnconditionallyLoad))
552 // See if we are already promoting a load with these indices. If not, check
553 // to make sure that we aren't promoting too many elements. If so, nothing
555 if (ToPromote.find(Operands) == ToPromote.end()) {
556 if (maxElements > 0 && ToPromote.size() == maxElements) {
557 DEBUG(dbgs() << "argpromotion not promoting argument '"
558 << Arg->getName() << "' because it would require adding more "
559 << "than " << maxElements << " arguments to the function.\n");
560 // We limit aggregate promotion to only promoting up to a fixed number
561 // of elements of the aggregate.
564 ToPromote.insert(std::move(Operands));
568 if (Loads.empty()) return true; // No users, this is a dead argument.
570 // Okay, now we know that the argument is only used by load instructions and
571 // it is safe to unconditionally perform all of them. Use alias analysis to
572 // check to see if the pointer is guaranteed to not be modified from entry of
573 // the function to each of the load instructions.
575 // Because there could be several/many load instructions, remember which
576 // blocks we know to be transparent to the load.
577 SmallPtrSet<BasicBlock*, 16> TranspBlocks;
579 for (unsigned i = 0, e = Loads.size(); i != e; ++i) {
580 // Check to see if the load is invalidated from the start of the block to
582 LoadInst *Load = Loads[i];
583 BasicBlock *BB = Load->getParent();
585 MemoryLocation Loc = MemoryLocation::get(Load);
586 if (AAR.canInstructionRangeModRef(BB->front(), *Load, Loc, MRI_Mod))
587 return false; // Pointer is invalidated!
589 // Now check every path from the entry block to the load for transparency.
590 // To do this, we perform a depth first search on the inverse CFG from the
592 for (BasicBlock *P : predecessors(BB)) {
593 for (BasicBlock *TranspBB : inverse_depth_first_ext(P, TranspBlocks))
594 if (AAR.canBasicBlockModify(*TranspBB, Loc))
599 // If the path from the entry of the function to each load is free of
600 // instructions that potentially invalidate the load, we can make the
605 /// DoPromotion - This method actually performs the promotion of the specified
606 /// arguments, and returns the new function. At this point, we know that it's
608 CallGraphNode *ArgPromotion::DoPromotion(Function *F,
609 SmallPtrSetImpl<Argument*> &ArgsToPromote,
610 SmallPtrSetImpl<Argument*> &ByValArgsToTransform) {
612 // Start by computing a new prototype for the function, which is the same as
613 // the old function, but has modified arguments.
614 FunctionType *FTy = F->getFunctionType();
615 std::vector<Type*> Params;
617 typedef std::set<std::pair<Type *, IndicesVector>> ScalarizeTable;
619 // ScalarizedElements - If we are promoting a pointer that has elements
620 // accessed out of it, keep track of which elements are accessed so that we
621 // can add one argument for each.
623 // Arguments that are directly loaded will have a zero element value here, to
624 // handle cases where there are both a direct load and GEP accesses.
626 std::map<Argument*, ScalarizeTable> ScalarizedElements;
628 // OriginalLoads - Keep track of a representative load instruction from the
629 // original function so that we can tell the alias analysis implementation
630 // what the new GEP/Load instructions we are inserting look like.
631 // We need to keep the original loads for each argument and the elements
632 // of the argument that are accessed.
633 std::map<std::pair<Argument*, IndicesVector>, LoadInst*> OriginalLoads;
635 // Attribute - Keep track of the parameter attributes for the arguments
636 // that we are *not* promoting. For the ones that we do promote, the parameter
637 // attributes are lost
638 SmallVector<AttributeSet, 8> AttributesVec;
639 const AttributeSet &PAL = F->getAttributes();
641 // Add any return attributes.
642 if (PAL.hasAttributes(AttributeSet::ReturnIndex))
643 AttributesVec.push_back(AttributeSet::get(F->getContext(),
644 PAL.getRetAttributes()));
646 // First, determine the new argument list
647 unsigned ArgIndex = 1;
648 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(); I != E;
650 if (ByValArgsToTransform.count(&*I)) {
651 // Simple byval argument? Just add all the struct element types.
652 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
653 StructType *STy = cast<StructType>(AgTy);
654 Params.insert(Params.end(), STy->element_begin(), STy->element_end());
655 ++NumByValArgsPromoted;
656 } else if (!ArgsToPromote.count(&*I)) {
657 // Unchanged argument
658 Params.push_back(I->getType());
659 AttributeSet attrs = PAL.getParamAttributes(ArgIndex);
660 if (attrs.hasAttributes(ArgIndex)) {
661 AttrBuilder B(attrs, ArgIndex);
663 push_back(AttributeSet::get(F->getContext(), Params.size(), B));
665 } else if (I->use_empty()) {
666 // Dead argument (which are always marked as promotable)
669 // Okay, this is being promoted. This means that the only uses are loads
670 // or GEPs which are only used by loads
672 // In this table, we will track which indices are loaded from the argument
673 // (where direct loads are tracked as no indices).
674 ScalarizeTable &ArgIndices = ScalarizedElements[&*I];
675 for (User *U : I->users()) {
676 Instruction *UI = cast<Instruction>(U);
678 if (LoadInst *L = dyn_cast<LoadInst>(UI))
679 SrcTy = L->getType();
681 SrcTy = cast<GetElementPtrInst>(UI)->getSourceElementType();
682 IndicesVector Indices;
683 Indices.reserve(UI->getNumOperands() - 1);
684 // Since loads will only have a single operand, and GEPs only a single
685 // non-index operand, this will record direct loads without any indices,
686 // and gep+loads with the GEP indices.
687 for (User::op_iterator II = UI->op_begin() + 1, IE = UI->op_end();
689 Indices.push_back(cast<ConstantInt>(*II)->getSExtValue());
690 // GEPs with a single 0 index can be merged with direct loads
691 if (Indices.size() == 1 && Indices.front() == 0)
693 ArgIndices.insert(std::make_pair(SrcTy, Indices));
695 if (LoadInst *L = dyn_cast<LoadInst>(UI))
698 // Take any load, we will use it only to update Alias Analysis
699 OrigLoad = cast<LoadInst>(UI->user_back());
700 OriginalLoads[std::make_pair(&*I, Indices)] = OrigLoad;
703 // Add a parameter to the function for each element passed in.
704 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
705 E = ArgIndices.end(); SI != E; ++SI) {
706 // not allowed to dereference ->begin() if size() is 0
707 Params.push_back(GetElementPtrInst::getIndexedType(
708 cast<PointerType>(I->getType()->getScalarType())->getElementType(),
710 assert(Params.back());
713 if (ArgIndices.size() == 1 && ArgIndices.begin()->second.empty())
714 ++NumArgumentsPromoted;
716 ++NumAggregatesPromoted;
720 // Add any function attributes.
721 if (PAL.hasAttributes(AttributeSet::FunctionIndex))
722 AttributesVec.push_back(AttributeSet::get(FTy->getContext(),
723 PAL.getFnAttributes()));
725 Type *RetTy = FTy->getReturnType();
727 // Construct the new function type using the new arguments.
728 FunctionType *NFTy = FunctionType::get(RetTy, Params, FTy->isVarArg());
730 // Create the new function body and insert it into the module.
731 Function *NF = Function::Create(NFTy, F->getLinkage(), F->getName());
732 NF->copyAttributesFrom(F);
734 // Patch the pointer to LLVM function in debug info descriptor.
735 auto DI = FunctionDIs.find(F);
736 if (DI != FunctionDIs.end()) {
737 DISubprogram *SP = DI->second;
738 SP->replaceFunction(NF);
739 // Ensure the map is updated so it can be reused on subsequent argument
740 // promotions of the same function.
741 FunctionDIs.erase(DI);
742 FunctionDIs[NF] = SP;
745 DEBUG(dbgs() << "ARG PROMOTION: Promoting to:" << *NF << "\n"
748 // Recompute the parameter attributes list based on the new arguments for
750 NF->setAttributes(AttributeSet::get(F->getContext(), AttributesVec));
751 AttributesVec.clear();
753 F->getParent()->getFunctionList().insert(F->getIterator(), NF);
756 // Get the callgraph information that we need to update to reflect our
758 CallGraph &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
760 // Get a new callgraph node for NF.
761 CallGraphNode *NF_CGN = CG.getOrInsertFunction(NF);
763 // Loop over all of the callers of the function, transforming the call sites
764 // to pass in the loaded pointers.
766 SmallVector<Value*, 16> Args;
767 while (!F->use_empty()) {
768 CallSite CS(F->user_back());
769 assert(CS.getCalledFunction() == F);
770 Instruction *Call = CS.getInstruction();
771 const AttributeSet &CallPAL = CS.getAttributes();
773 // Add any return attributes.
774 if (CallPAL.hasAttributes(AttributeSet::ReturnIndex))
775 AttributesVec.push_back(AttributeSet::get(F->getContext(),
776 CallPAL.getRetAttributes()));
778 // Loop over the operands, inserting GEP and loads in the caller as
780 CallSite::arg_iterator AI = CS.arg_begin();
782 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end();
783 I != E; ++I, ++AI, ++ArgIndex)
784 if (!ArgsToPromote.count(&*I) && !ByValArgsToTransform.count(&*I)) {
785 Args.push_back(*AI); // Unmodified argument
787 if (CallPAL.hasAttributes(ArgIndex)) {
788 AttrBuilder B(CallPAL, ArgIndex);
790 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
792 } else if (ByValArgsToTransform.count(&*I)) {
793 // Emit a GEP and load for each element of the struct.
794 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
795 StructType *STy = cast<StructType>(AgTy);
797 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
798 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
799 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
800 Value *Idx = GetElementPtrInst::Create(
801 STy, *AI, Idxs, (*AI)->getName() + "." + Twine(i), Call);
802 // TODO: Tell AA about the new values?
803 Args.push_back(new LoadInst(Idx, Idx->getName()+".val", Call));
805 } else if (!I->use_empty()) {
806 // Non-dead argument: insert GEPs and loads as appropriate.
807 ScalarizeTable &ArgIndices = ScalarizedElements[&*I];
808 // Store the Value* version of the indices in here, but declare it now
810 std::vector<Value*> Ops;
811 for (ScalarizeTable::iterator SI = ArgIndices.begin(),
812 E = ArgIndices.end(); SI != E; ++SI) {
814 LoadInst *OrigLoad = OriginalLoads[std::make_pair(&*I, SI->second)];
815 if (!SI->second.empty()) {
816 Ops.reserve(SI->second.size());
817 Type *ElTy = V->getType();
818 for (IndicesVector::const_iterator II = SI->second.begin(),
819 IE = SI->second.end();
821 // Use i32 to index structs, and i64 for others (pointers/arrays).
822 // This satisfies GEP constraints.
823 Type *IdxTy = (ElTy->isStructTy() ?
824 Type::getInt32Ty(F->getContext()) :
825 Type::getInt64Ty(F->getContext()));
826 Ops.push_back(ConstantInt::get(IdxTy, *II));
827 // Keep track of the type we're currently indexing.
828 ElTy = cast<CompositeType>(ElTy)->getTypeAtIndex(*II);
830 // And create a GEP to extract those indices.
831 V = GetElementPtrInst::Create(SI->first, V, Ops,
832 V->getName() + ".idx", Call);
835 // Since we're replacing a load make sure we take the alignment
836 // of the previous load.
837 LoadInst *newLoad = new LoadInst(V, V->getName()+".val", Call);
838 newLoad->setAlignment(OrigLoad->getAlignment());
839 // Transfer the AA info too.
841 OrigLoad->getAAMetadata(AAInfo);
842 newLoad->setAAMetadata(AAInfo);
844 Args.push_back(newLoad);
848 // Push any varargs arguments on the list.
849 for (; AI != CS.arg_end(); ++AI, ++ArgIndex) {
851 if (CallPAL.hasAttributes(ArgIndex)) {
852 AttrBuilder B(CallPAL, ArgIndex);
854 push_back(AttributeSet::get(F->getContext(), Args.size(), B));
858 // Add any function attributes.
859 if (CallPAL.hasAttributes(AttributeSet::FunctionIndex))
860 AttributesVec.push_back(AttributeSet::get(Call->getContext(),
861 CallPAL.getFnAttributes()));
864 if (InvokeInst *II = dyn_cast<InvokeInst>(Call)) {
865 New = InvokeInst::Create(NF, II->getNormalDest(), II->getUnwindDest(),
867 cast<InvokeInst>(New)->setCallingConv(CS.getCallingConv());
868 cast<InvokeInst>(New)->setAttributes(AttributeSet::get(II->getContext(),
871 New = CallInst::Create(NF, Args, "", Call);
872 cast<CallInst>(New)->setCallingConv(CS.getCallingConv());
873 cast<CallInst>(New)->setAttributes(AttributeSet::get(New->getContext(),
875 if (cast<CallInst>(Call)->isTailCall())
876 cast<CallInst>(New)->setTailCall();
878 New->setDebugLoc(Call->getDebugLoc());
880 AttributesVec.clear();
882 // Update the callgraph to know that the callsite has been transformed.
883 CallGraphNode *CalleeNode = CG[Call->getParent()->getParent()];
884 CalleeNode->replaceCallEdge(CS, CallSite(New), NF_CGN);
886 if (!Call->use_empty()) {
887 Call->replaceAllUsesWith(New);
891 // Finally, remove the old call from the program, reducing the use-count of
893 Call->eraseFromParent();
896 // Since we have now created the new function, splice the body of the old
897 // function right into the new function, leaving the old rotting hulk of the
899 NF->getBasicBlockList().splice(NF->begin(), F->getBasicBlockList());
901 // Loop over the argument list, transferring uses of the old arguments over to
902 // the new arguments, also transferring over the names as well.
904 for (Function::arg_iterator I = F->arg_begin(), E = F->arg_end(),
905 I2 = NF->arg_begin(); I != E; ++I) {
906 if (!ArgsToPromote.count(&*I) && !ByValArgsToTransform.count(&*I)) {
907 // If this is an unmodified argument, move the name and users over to the
909 I->replaceAllUsesWith(&*I2);
915 if (ByValArgsToTransform.count(&*I)) {
916 // In the callee, we create an alloca, and store each of the new incoming
917 // arguments into the alloca.
918 Instruction *InsertPt = &NF->begin()->front();
920 // Just add all the struct element types.
921 Type *AgTy = cast<PointerType>(I->getType())->getElementType();
922 Value *TheAlloca = new AllocaInst(AgTy, nullptr, "", InsertPt);
923 StructType *STy = cast<StructType>(AgTy);
925 ConstantInt::get(Type::getInt32Ty(F->getContext()), 0), nullptr };
927 for (unsigned i = 0, e = STy->getNumElements(); i != e; ++i) {
928 Idxs[1] = ConstantInt::get(Type::getInt32Ty(F->getContext()), i);
929 Value *Idx = GetElementPtrInst::Create(
930 AgTy, TheAlloca, Idxs, TheAlloca->getName() + "." + Twine(i),
932 I2->setName(I->getName()+"."+Twine(i));
933 new StoreInst(&*I2++, Idx, InsertPt);
936 // Anything that used the arg should now use the alloca.
937 I->replaceAllUsesWith(TheAlloca);
938 TheAlloca->takeName(&*I);
940 // If the alloca is used in a call, we must clear the tail flag since
941 // the callee now uses an alloca from the caller.
942 for (User *U : TheAlloca->users()) {
943 CallInst *Call = dyn_cast<CallInst>(U);
946 Call->setTailCall(false);
954 // Otherwise, if we promoted this argument, then all users are load
955 // instructions (or GEPs with only load users), and all loads should be
956 // using the new argument that we added.
957 ScalarizeTable &ArgIndices = ScalarizedElements[&*I];
959 while (!I->use_empty()) {
960 if (LoadInst *LI = dyn_cast<LoadInst>(I->user_back())) {
961 assert(ArgIndices.begin()->second.empty() &&
962 "Load element should sort to front!");
963 I2->setName(I->getName()+".val");
964 LI->replaceAllUsesWith(&*I2);
965 LI->eraseFromParent();
966 DEBUG(dbgs() << "*** Promoted load of argument '" << I->getName()
967 << "' in function '" << F->getName() << "'\n");
969 GetElementPtrInst *GEP = cast<GetElementPtrInst>(I->user_back());
970 IndicesVector Operands;
971 Operands.reserve(GEP->getNumIndices());
972 for (User::op_iterator II = GEP->idx_begin(), IE = GEP->idx_end();
974 Operands.push_back(cast<ConstantInt>(*II)->getSExtValue());
976 // GEPs with a single 0 index can be merged with direct loads
977 if (Operands.size() == 1 && Operands.front() == 0)
980 Function::arg_iterator TheArg = I2;
981 for (ScalarizeTable::iterator It = ArgIndices.begin();
982 It->second != Operands; ++It, ++TheArg) {
983 assert(It != ArgIndices.end() && "GEP not handled??");
986 std::string NewName = I->getName();
987 for (unsigned i = 0, e = Operands.size(); i != e; ++i) {
988 NewName += "." + utostr(Operands[i]);
991 TheArg->setName(NewName);
993 DEBUG(dbgs() << "*** Promoted agg argument '" << TheArg->getName()
994 << "' of function '" << NF->getName() << "'\n");
996 // All of the uses must be load instructions. Replace them all with
997 // the argument specified by ArgNo.
998 while (!GEP->use_empty()) {
999 LoadInst *L = cast<LoadInst>(GEP->user_back());
1000 L->replaceAllUsesWith(&*TheArg);
1001 L->eraseFromParent();
1003 GEP->eraseFromParent();
1007 // Increment I2 past all of the arguments added for this promoted pointer.
1008 std::advance(I2, ArgIndices.size());
1011 NF_CGN->stealCalledFunctionsFrom(CG[F]);
1013 // Now that the old function is dead, delete it. If there is a dangling
1014 // reference to the CallgraphNode, just leave the dead function around for
1015 // someone else to nuke.
1016 CallGraphNode *CGN = CG[F];
1017 if (CGN->getNumReferences() == 0)
1018 delete CG.removeFunctionFromModule(CGN);
1020 F->setLinkage(Function::ExternalLinkage);
1025 bool ArgPromotion::doInitialization(CallGraph &CG) {
1026 FunctionDIs = makeSubprogramMap(CG.getModule());
1027 return CallGraphSCCPass::doInitialization(CG);