1 //===- FunctionAttrs.cpp - Pass which marks functions attributes ----------===//
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 //===----------------------------------------------------------------------===//
11 /// This file implements interprocedural passes which walk the
12 /// call-graph deducing and/or propagating function attributes.
14 //===----------------------------------------------------------------------===//
16 #include "llvm/Transforms/IPO/FunctionAttrs.h"
17 #include "llvm/Transforms/IPO.h"
18 #include "llvm/ADT/SCCIterator.h"
19 #include "llvm/ADT/SetVector.h"
20 #include "llvm/ADT/SmallSet.h"
21 #include "llvm/ADT/Statistic.h"
22 #include "llvm/ADT/StringSwitch.h"
23 #include "llvm/Analysis/AliasAnalysis.h"
24 #include "llvm/Analysis/AssumptionCache.h"
25 #include "llvm/Analysis/BasicAliasAnalysis.h"
26 #include "llvm/Analysis/CallGraph.h"
27 #include "llvm/Analysis/CallGraphSCCPass.h"
28 #include "llvm/Analysis/CaptureTracking.h"
29 #include "llvm/Analysis/TargetLibraryInfo.h"
30 #include "llvm/Analysis/ValueTracking.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/InstIterator.h"
33 #include "llvm/IR/IntrinsicInst.h"
34 #include "llvm/IR/LLVMContext.h"
35 #include "llvm/Support/Debug.h"
36 #include "llvm/Support/raw_ostream.h"
37 #include "llvm/Analysis/TargetLibraryInfo.h"
40 #define DEBUG_TYPE "functionattrs"
42 STATISTIC(NumReadNone, "Number of functions marked readnone");
43 STATISTIC(NumReadOnly, "Number of functions marked readonly");
44 STATISTIC(NumNoCapture, "Number of arguments marked nocapture");
45 STATISTIC(NumReadNoneArg, "Number of arguments marked readnone");
46 STATISTIC(NumReadOnlyArg, "Number of arguments marked readonly");
47 STATISTIC(NumNoAlias, "Number of function returns marked noalias");
48 STATISTIC(NumNonNullReturn, "Number of function returns marked nonnull");
49 STATISTIC(NumNoRecurse, "Number of functions marked as norecurse");
52 typedef SmallSetVector<Function *, 8> SCCNodeSet;
56 /// The three kinds of memory access relevant to 'readonly' and
57 /// 'readnone' attributes.
58 enum MemoryAccessKind {
65 static MemoryAccessKind checkFunctionMemoryAccess(Function &F, AAResults &AAR,
66 const SCCNodeSet &SCCNodes) {
67 FunctionModRefBehavior MRB = AAR.getModRefBehavior(&F);
68 if (MRB == FMRB_DoesNotAccessMemory)
72 // Non-exact function definitions may not be selected at link time, and an
73 // alternative version that writes to memory may be selected. See the comment
74 // on GlobalValue::isDefinitionExact for more details.
75 if (!F.hasExactDefinition()) {
76 if (AliasAnalysis::onlyReadsMemory(MRB))
79 // Conservatively assume it writes to memory.
83 // Scan the function body for instructions that may read or write memory.
84 bool ReadsMemory = false;
85 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
86 Instruction *I = &*II;
88 // Some instructions can be ignored even if they read or write memory.
89 // Detect these now, skipping to the next instruction if one is found.
90 CallSite CS(cast<Value>(I));
92 // Ignore calls to functions in the same SCC, as long as the call sites
93 // don't have operand bundles. Calls with operand bundles are allowed to
94 // have memory effects not described by the memory effects of the call
96 if (!CS.hasOperandBundles() && CS.getCalledFunction() &&
97 SCCNodes.count(CS.getCalledFunction()))
99 FunctionModRefBehavior MRB = AAR.getModRefBehavior(CS);
101 // If the call doesn't access memory, we're done.
102 if (!(MRB & MRI_ModRef))
105 if (!AliasAnalysis::onlyAccessesArgPointees(MRB)) {
106 // The call could access any memory. If that includes writes, give up.
109 // If it reads, note it.
115 // Check whether all pointer arguments point to local memory, and
116 // ignore calls that only access local memory.
117 for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
120 if (!Arg->getType()->isPtrOrPtrVectorTy())
124 I->getAAMetadata(AAInfo);
125 MemoryLocation Loc(Arg, MemoryLocation::UnknownSize, AAInfo);
127 // Skip accesses to local or constant memory as they don't impact the
128 // externally visible mod/ref behavior.
129 if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
133 // Writes non-local memory. Give up.
136 // Ok, it reads non-local memory.
140 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
141 // Ignore non-volatile loads from local memory. (Atomic is okay here.)
142 if (!LI->isVolatile()) {
143 MemoryLocation Loc = MemoryLocation::get(LI);
144 if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
147 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
148 // Ignore non-volatile stores to local memory. (Atomic is okay here.)
149 if (!SI->isVolatile()) {
150 MemoryLocation Loc = MemoryLocation::get(SI);
151 if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
154 } else if (VAArgInst *VI = dyn_cast<VAArgInst>(I)) {
155 // Ignore vaargs on local memory.
156 MemoryLocation Loc = MemoryLocation::get(VI);
157 if (AAR.pointsToConstantMemory(Loc, /*OrLocal=*/true))
161 // Any remaining instructions need to be taken seriously! Check if they
162 // read or write memory.
163 if (I->mayWriteToMemory())
164 // Writes memory. Just give up.
167 // If this instruction may read memory, remember that.
168 ReadsMemory |= I->mayReadFromMemory();
171 return ReadsMemory ? MAK_ReadOnly : MAK_ReadNone;
174 /// Deduce readonly/readnone attributes for the SCC.
175 template <typename AARGetterT>
176 static bool addReadAttrs(const SCCNodeSet &SCCNodes, AARGetterT AARGetter) {
177 // Check if any of the functions in the SCC read or write memory. If they
178 // write memory then they can't be marked readnone or readonly.
179 bool ReadsMemory = false;
180 for (Function *F : SCCNodes) {
181 // Call the callable parameter to look up AA results for this function.
182 AAResults &AAR = AARGetter(*F);
184 switch (checkFunctionMemoryAccess(*F, AAR, SCCNodes)) {
196 // Success! Functions in this SCC do not access memory, or only read memory.
197 // Give them the appropriate attribute.
198 bool MadeChange = false;
199 for (Function *F : SCCNodes) {
200 if (F->doesNotAccessMemory())
204 if (F->onlyReadsMemory() && ReadsMemory)
210 // Clear out any existing attributes.
212 B.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone);
214 AttributeSet::FunctionIndex,
215 AttributeSet::get(F->getContext(), AttributeSet::FunctionIndex, B));
217 // Add in the new attribute.
218 F->addAttribute(AttributeSet::FunctionIndex,
219 ReadsMemory ? Attribute::ReadOnly : Attribute::ReadNone);
231 /// For a given pointer Argument, this retains a list of Arguments of functions
232 /// in the same SCC that the pointer data flows into. We use this to build an
233 /// SCC of the arguments.
234 struct ArgumentGraphNode {
235 Argument *Definition;
236 SmallVector<ArgumentGraphNode *, 4> Uses;
239 class ArgumentGraph {
240 // We store pointers to ArgumentGraphNode objects, so it's important that
241 // that they not move around upon insert.
242 typedef std::map<Argument *, ArgumentGraphNode> ArgumentMapTy;
244 ArgumentMapTy ArgumentMap;
246 // There is no root node for the argument graph, in fact:
247 // void f(int *x, int *y) { if (...) f(x, y); }
248 // is an example where the graph is disconnected. The SCCIterator requires a
249 // single entry point, so we maintain a fake ("synthetic") root node that
250 // uses every node. Because the graph is directed and nothing points into
251 // the root, it will not participate in any SCCs (except for its own).
252 ArgumentGraphNode SyntheticRoot;
255 ArgumentGraph() { SyntheticRoot.Definition = nullptr; }
257 typedef SmallVectorImpl<ArgumentGraphNode *>::iterator iterator;
259 iterator begin() { return SyntheticRoot.Uses.begin(); }
260 iterator end() { return SyntheticRoot.Uses.end(); }
261 ArgumentGraphNode *getEntryNode() { return &SyntheticRoot; }
263 ArgumentGraphNode *operator[](Argument *A) {
264 ArgumentGraphNode &Node = ArgumentMap[A];
266 SyntheticRoot.Uses.push_back(&Node);
271 /// This tracker checks whether callees are in the SCC, and if so it does not
272 /// consider that a capture, instead adding it to the "Uses" list and
273 /// continuing with the analysis.
274 struct ArgumentUsesTracker : public CaptureTracker {
275 ArgumentUsesTracker(const SCCNodeSet &SCCNodes)
276 : Captured(false), SCCNodes(SCCNodes) {}
278 void tooManyUses() override { Captured = true; }
280 bool captured(const Use *U) override {
281 CallSite CS(U->getUser());
282 if (!CS.getInstruction()) {
287 Function *F = CS.getCalledFunction();
288 if (!F || !F->hasExactDefinition() || !SCCNodes.count(F)) {
293 // Note: the callee and the two successor blocks *follow* the argument
294 // operands. This means there is no need to adjust UseIndex to account for
298 std::distance(const_cast<const Use *>(CS.arg_begin()), U);
300 assert(UseIndex < CS.data_operands_size() &&
301 "Indirect function calls should have been filtered above!");
303 if (UseIndex >= CS.getNumArgOperands()) {
304 // Data operand, but not a argument operand -- must be a bundle operand
305 assert(CS.hasOperandBundles() && "Must be!");
307 // CaptureTracking told us that we're being captured by an operand bundle
308 // use. In this case it does not matter if the callee is within our SCC
309 // or not -- we've been captured in some unknown way, and we have to be
315 if (UseIndex >= F->arg_size()) {
316 assert(F->isVarArg() && "More params than args in non-varargs call");
321 Uses.push_back(&*std::next(F->arg_begin(), UseIndex));
325 bool Captured; // True only if certainly captured (used outside our SCC).
326 SmallVector<Argument *, 4> Uses; // Uses within our SCC.
328 const SCCNodeSet &SCCNodes;
333 template <> struct GraphTraits<ArgumentGraphNode *> {
334 typedef ArgumentGraphNode NodeType;
335 typedef ArgumentGraphNode *NodeRef;
336 typedef SmallVectorImpl<ArgumentGraphNode *>::iterator ChildIteratorType;
338 static inline NodeType *getEntryNode(NodeType *A) { return A; }
339 static inline ChildIteratorType child_begin(NodeType *N) {
340 return N->Uses.begin();
342 static inline ChildIteratorType child_end(NodeType *N) {
343 return N->Uses.end();
347 struct GraphTraits<ArgumentGraph *> : public GraphTraits<ArgumentGraphNode *> {
348 static NodeType *getEntryNode(ArgumentGraph *AG) {
349 return AG->getEntryNode();
351 static ChildIteratorType nodes_begin(ArgumentGraph *AG) {
354 static ChildIteratorType nodes_end(ArgumentGraph *AG) { return AG->end(); }
358 /// Returns Attribute::None, Attribute::ReadOnly or Attribute::ReadNone.
359 static Attribute::AttrKind
360 determinePointerReadAttrs(Argument *A,
361 const SmallPtrSet<Argument *, 8> &SCCNodes) {
363 SmallVector<Use *, 32> Worklist;
364 SmallSet<Use *, 32> Visited;
366 // inalloca arguments are always clobbered by the call.
367 if (A->hasInAllocaAttr())
368 return Attribute::None;
371 // We don't need to track IsWritten. If A is written to, return immediately.
373 for (Use &U : A->uses()) {
375 Worklist.push_back(&U);
378 while (!Worklist.empty()) {
379 Use *U = Worklist.pop_back_val();
380 Instruction *I = cast<Instruction>(U->getUser());
382 switch (I->getOpcode()) {
383 case Instruction::BitCast:
384 case Instruction::GetElementPtr:
385 case Instruction::PHI:
386 case Instruction::Select:
387 case Instruction::AddrSpaceCast:
388 // The original value is not read/written via this if the new value isn't.
389 for (Use &UU : I->uses())
390 if (Visited.insert(&UU).second)
391 Worklist.push_back(&UU);
394 case Instruction::Call:
395 case Instruction::Invoke: {
396 bool Captures = true;
398 if (I->getType()->isVoidTy())
401 auto AddUsersToWorklistIfCapturing = [&] {
403 for (Use &UU : I->uses())
404 if (Visited.insert(&UU).second)
405 Worklist.push_back(&UU);
409 if (CS.doesNotAccessMemory()) {
410 AddUsersToWorklistIfCapturing();
414 Function *F = CS.getCalledFunction();
416 if (CS.onlyReadsMemory()) {
418 AddUsersToWorklistIfCapturing();
421 return Attribute::None;
424 // Note: the callee and the two successor blocks *follow* the argument
425 // operands. This means there is no need to adjust UseIndex to account
428 unsigned UseIndex = std::distance(CS.arg_begin(), U);
430 // U cannot be the callee operand use: since we're exploring the
431 // transitive uses of an Argument, having such a use be a callee would
432 // imply the CallSite is an indirect call or invoke; and we'd take the
434 assert(UseIndex < CS.data_operands_size() &&
435 "Data operand use expected!");
437 bool IsOperandBundleUse = UseIndex >= CS.getNumArgOperands();
439 if (UseIndex >= F->arg_size() && !IsOperandBundleUse) {
440 assert(F->isVarArg() && "More params than args in non-varargs call");
441 return Attribute::None;
444 Captures &= !CS.doesNotCapture(UseIndex);
446 // Since the optimizer (by design) cannot see the data flow corresponding
447 // to a operand bundle use, these cannot participate in the optimistic SCC
448 // analysis. Instead, we model the operand bundle uses as arguments in
449 // call to a function external to the SCC.
450 if (!SCCNodes.count(&*std::next(F->arg_begin(), UseIndex)) ||
451 IsOperandBundleUse) {
453 // The accessors used on CallSite here do the right thing for calls and
454 // invokes with operand bundles.
456 if (!CS.onlyReadsMemory() && !CS.onlyReadsMemory(UseIndex))
457 return Attribute::None;
458 if (!CS.doesNotAccessMemory(UseIndex))
462 AddUsersToWorklistIfCapturing();
466 case Instruction::Load:
467 // A volatile load has side effects beyond what readonly can be relied
469 if (cast<LoadInst>(I)->isVolatile())
470 return Attribute::None;
475 case Instruction::ICmp:
476 case Instruction::Ret:
480 return Attribute::None;
484 return IsRead ? Attribute::ReadOnly : Attribute::ReadNone;
487 /// Deduce nocapture attributes for the SCC.
488 static bool addArgumentAttrs(const SCCNodeSet &SCCNodes) {
489 bool Changed = false;
494 B.addAttribute(Attribute::NoCapture);
496 // Check each function in turn, determining which pointer arguments are not
498 for (Function *F : SCCNodes) {
499 // We can infer and propagate function attributes only when we know that the
500 // definition we'll get at link time is *exactly* the definition we see now.
501 // For more details, see GlobalValue::mayBeDerefined.
502 if (!F->hasExactDefinition())
505 // Functions that are readonly (or readnone) and nounwind and don't return
506 // a value can't capture arguments. Don't analyze them.
507 if (F->onlyReadsMemory() && F->doesNotThrow() &&
508 F->getReturnType()->isVoidTy()) {
509 for (Function::arg_iterator A = F->arg_begin(), E = F->arg_end(); A != E;
511 if (A->getType()->isPointerTy() && !A->hasNoCaptureAttr()) {
512 A->addAttr(AttributeSet::get(F->getContext(), A->getArgNo() + 1, B));
520 for (Function::arg_iterator A = F->arg_begin(), E = F->arg_end(); A != E;
522 if (!A->getType()->isPointerTy())
524 bool HasNonLocalUses = false;
525 if (!A->hasNoCaptureAttr()) {
526 ArgumentUsesTracker Tracker(SCCNodes);
527 PointerMayBeCaptured(&*A, &Tracker);
528 if (!Tracker.Captured) {
529 if (Tracker.Uses.empty()) {
530 // If it's trivially not captured, mark it nocapture now.
532 AttributeSet::get(F->getContext(), A->getArgNo() + 1, B));
536 // If it's not trivially captured and not trivially not captured,
537 // then it must be calling into another function in our SCC. Save
538 // its particulars for Argument-SCC analysis later.
539 ArgumentGraphNode *Node = AG[&*A];
540 for (Argument *Use : Tracker.Uses) {
541 Node->Uses.push_back(AG[Use]);
543 HasNonLocalUses = true;
547 // Otherwise, it's captured. Don't bother doing SCC analysis on it.
549 if (!HasNonLocalUses && !A->onlyReadsMemory()) {
550 // Can we determine that it's readonly/readnone without doing an SCC?
551 // Note that we don't allow any calls at all here, or else our result
552 // will be dependent on the iteration order through the functions in the
554 SmallPtrSet<Argument *, 8> Self;
556 Attribute::AttrKind R = determinePointerReadAttrs(&*A, Self);
557 if (R != Attribute::None) {
560 A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
562 R == Attribute::ReadOnly ? ++NumReadOnlyArg : ++NumReadNoneArg;
568 // The graph we've collected is partial because we stopped scanning for
569 // argument uses once we solved the argument trivially. These partial nodes
570 // show up as ArgumentGraphNode objects with an empty Uses list, and for
571 // these nodes the final decision about whether they capture has already been
572 // made. If the definition doesn't have a 'nocapture' attribute by now, it
575 for (scc_iterator<ArgumentGraph *> I = scc_begin(&AG); !I.isAtEnd(); ++I) {
576 const std::vector<ArgumentGraphNode *> &ArgumentSCC = *I;
577 if (ArgumentSCC.size() == 1) {
578 if (!ArgumentSCC[0]->Definition)
579 continue; // synthetic root node
581 // eg. "void f(int* x) { if (...) f(x); }"
582 if (ArgumentSCC[0]->Uses.size() == 1 &&
583 ArgumentSCC[0]->Uses[0] == ArgumentSCC[0]) {
584 Argument *A = ArgumentSCC[0]->Definition;
585 A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
592 bool SCCCaptured = false;
593 for (auto I = ArgumentSCC.begin(), E = ArgumentSCC.end();
594 I != E && !SCCCaptured; ++I) {
595 ArgumentGraphNode *Node = *I;
596 if (Node->Uses.empty()) {
597 if (!Node->Definition->hasNoCaptureAttr())
604 SmallPtrSet<Argument *, 8> ArgumentSCCNodes;
605 // Fill ArgumentSCCNodes with the elements of the ArgumentSCC. Used for
606 // quickly looking up whether a given Argument is in this ArgumentSCC.
607 for (ArgumentGraphNode *I : ArgumentSCC) {
608 ArgumentSCCNodes.insert(I->Definition);
611 for (auto I = ArgumentSCC.begin(), E = ArgumentSCC.end();
612 I != E && !SCCCaptured; ++I) {
613 ArgumentGraphNode *N = *I;
614 for (ArgumentGraphNode *Use : N->Uses) {
615 Argument *A = Use->Definition;
616 if (A->hasNoCaptureAttr() || ArgumentSCCNodes.count(A))
625 for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) {
626 Argument *A = ArgumentSCC[i]->Definition;
627 A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
632 // We also want to compute readonly/readnone. With a small number of false
633 // negatives, we can assume that any pointer which is captured isn't going
634 // to be provably readonly or readnone, since by definition we can't
635 // analyze all uses of a captured pointer.
637 // The false negatives happen when the pointer is captured by a function
638 // that promises readonly/readnone behaviour on the pointer, then the
639 // pointer's lifetime ends before anything that writes to arbitrary memory.
640 // Also, a readonly/readnone pointer may be returned, but returning a
641 // pointer is capturing it.
643 Attribute::AttrKind ReadAttr = Attribute::ReadNone;
644 for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) {
645 Argument *A = ArgumentSCC[i]->Definition;
646 Attribute::AttrKind K = determinePointerReadAttrs(A, ArgumentSCCNodes);
647 if (K == Attribute::ReadNone)
649 if (K == Attribute::ReadOnly) {
650 ReadAttr = Attribute::ReadOnly;
657 if (ReadAttr != Attribute::None) {
659 B.addAttribute(ReadAttr);
660 R.addAttribute(Attribute::ReadOnly).addAttribute(Attribute::ReadNone);
661 for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) {
662 Argument *A = ArgumentSCC[i]->Definition;
663 // Clear out existing readonly/readnone attributes
664 A->removeAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, R));
665 A->addAttr(AttributeSet::get(A->getContext(), A->getArgNo() + 1, B));
666 ReadAttr == Attribute::ReadOnly ? ++NumReadOnlyArg : ++NumReadNoneArg;
675 /// Tests whether a function is "malloc-like".
677 /// A function is "malloc-like" if it returns either null or a pointer that
678 /// doesn't alias any other pointer visible to the caller.
679 static bool isFunctionMallocLike(Function *F, const SCCNodeSet &SCCNodes) {
680 SmallSetVector<Value *, 8> FlowsToReturn;
681 for (BasicBlock &BB : *F)
682 if (ReturnInst *Ret = dyn_cast<ReturnInst>(BB.getTerminator()))
683 FlowsToReturn.insert(Ret->getReturnValue());
685 for (unsigned i = 0; i != FlowsToReturn.size(); ++i) {
686 Value *RetVal = FlowsToReturn[i];
688 if (Constant *C = dyn_cast<Constant>(RetVal)) {
689 if (!C->isNullValue() && !isa<UndefValue>(C))
695 if (isa<Argument>(RetVal))
698 if (Instruction *RVI = dyn_cast<Instruction>(RetVal))
699 switch (RVI->getOpcode()) {
700 // Extend the analysis by looking upwards.
701 case Instruction::BitCast:
702 case Instruction::GetElementPtr:
703 case Instruction::AddrSpaceCast:
704 FlowsToReturn.insert(RVI->getOperand(0));
706 case Instruction::Select: {
707 SelectInst *SI = cast<SelectInst>(RVI);
708 FlowsToReturn.insert(SI->getTrueValue());
709 FlowsToReturn.insert(SI->getFalseValue());
712 case Instruction::PHI: {
713 PHINode *PN = cast<PHINode>(RVI);
714 for (Value *IncValue : PN->incoming_values())
715 FlowsToReturn.insert(IncValue);
719 // Check whether the pointer came from an allocation.
720 case Instruction::Alloca:
722 case Instruction::Call:
723 case Instruction::Invoke: {
725 if (CS.paramHasAttr(0, Attribute::NoAlias))
727 if (CS.getCalledFunction() && SCCNodes.count(CS.getCalledFunction()))
731 return false; // Did not come from an allocation.
734 if (PointerMayBeCaptured(RetVal, false, /*StoreCaptures=*/false))
741 /// Deduce noalias attributes for the SCC.
742 static bool addNoAliasAttrs(const SCCNodeSet &SCCNodes) {
743 // Check each function in turn, determining which functions return noalias
745 for (Function *F : SCCNodes) {
747 if (F->doesNotAlias(0))
750 // We can infer and propagate function attributes only when we know that the
751 // definition we'll get at link time is *exactly* the definition we see now.
752 // For more details, see GlobalValue::mayBeDerefined.
753 if (!F->hasExactDefinition())
756 // We annotate noalias return values, which are only applicable to
758 if (!F->getReturnType()->isPointerTy())
761 if (!isFunctionMallocLike(F, SCCNodes))
765 bool MadeChange = false;
766 for (Function *F : SCCNodes) {
767 if (F->doesNotAlias(0) || !F->getReturnType()->isPointerTy())
770 F->setDoesNotAlias(0);
778 /// Tests whether this function is known to not return null.
780 /// Requires that the function returns a pointer.
782 /// Returns true if it believes the function will not return a null, and sets
783 /// \p Speculative based on whether the returned conclusion is a speculative
784 /// conclusion due to SCC calls.
785 static bool isReturnNonNull(Function *F, const SCCNodeSet &SCCNodes,
787 assert(F->getReturnType()->isPointerTy() &&
788 "nonnull only meaningful on pointer types");
791 SmallSetVector<Value *, 8> FlowsToReturn;
792 for (BasicBlock &BB : *F)
793 if (auto *Ret = dyn_cast<ReturnInst>(BB.getTerminator()))
794 FlowsToReturn.insert(Ret->getReturnValue());
796 for (unsigned i = 0; i != FlowsToReturn.size(); ++i) {
797 Value *RetVal = FlowsToReturn[i];
799 // If this value is locally known to be non-null, we're good
800 if (isKnownNonNull(RetVal))
803 // Otherwise, we need to look upwards since we can't make any local
805 Instruction *RVI = dyn_cast<Instruction>(RetVal);
808 switch (RVI->getOpcode()) {
809 // Extend the analysis by looking upwards.
810 case Instruction::BitCast:
811 case Instruction::GetElementPtr:
812 case Instruction::AddrSpaceCast:
813 FlowsToReturn.insert(RVI->getOperand(0));
815 case Instruction::Select: {
816 SelectInst *SI = cast<SelectInst>(RVI);
817 FlowsToReturn.insert(SI->getTrueValue());
818 FlowsToReturn.insert(SI->getFalseValue());
821 case Instruction::PHI: {
822 PHINode *PN = cast<PHINode>(RVI);
823 for (int i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
824 FlowsToReturn.insert(PN->getIncomingValue(i));
827 case Instruction::Call:
828 case Instruction::Invoke: {
830 Function *Callee = CS.getCalledFunction();
831 // A call to a node within the SCC is assumed to return null until
833 if (Callee && SCCNodes.count(Callee)) {
840 return false; // Unknown source, may be null
842 llvm_unreachable("should have either continued or returned");
848 /// Deduce nonnull attributes for the SCC.
849 static bool addNonNullAttrs(const SCCNodeSet &SCCNodes) {
850 // Speculative that all functions in the SCC return only nonnull
851 // pointers. We may refute this as we analyze functions.
852 bool SCCReturnsNonNull = true;
854 bool MadeChange = false;
856 // Check each function in turn, determining which functions return nonnull
858 for (Function *F : SCCNodes) {
860 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
864 // We can infer and propagate function attributes only when we know that the
865 // definition we'll get at link time is *exactly* the definition we see now.
866 // For more details, see GlobalValue::mayBeDerefined.
867 if (!F->hasExactDefinition())
870 // We annotate nonnull return values, which are only applicable to
872 if (!F->getReturnType()->isPointerTy())
875 bool Speculative = false;
876 if (isReturnNonNull(F, SCCNodes, Speculative)) {
878 // Mark the function eagerly since we may discover a function
879 // which prevents us from speculating about the entire SCC
880 DEBUG(dbgs() << "Eagerly marking " << F->getName() << " as nonnull\n");
881 F->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
887 // At least one function returns something which could be null, can't
888 // speculate any more.
889 SCCReturnsNonNull = false;
892 if (SCCReturnsNonNull) {
893 for (Function *F : SCCNodes) {
894 if (F->getAttributes().hasAttribute(AttributeSet::ReturnIndex,
895 Attribute::NonNull) ||
896 !F->getReturnType()->isPointerTy())
899 DEBUG(dbgs() << "SCC marking " << F->getName() << " as nonnull\n");
900 F->addAttribute(AttributeSet::ReturnIndex, Attribute::NonNull);
909 /// Remove the convergent attribute from all functions in the SCC if every
910 /// callsite within the SCC is not convergent (except for calls to functions
911 /// within the SCC). Returns true if changes were made.
912 static bool removeConvergentAttrs(const SCCNodeSet &SCCNodes) {
913 // For every function in SCC, ensure that either
914 // * it is not convergent, or
915 // * we can remove its convergent attribute.
916 bool HasConvergentFn = false;
917 for (Function *F : SCCNodes) {
918 if (!F->isConvergent()) continue;
919 HasConvergentFn = true;
921 // Can't remove convergent from function declarations.
922 if (F->isDeclaration()) return false;
924 // Can't remove convergent if any of our functions has a convergent call to a
925 // function not in the SCC.
926 for (Instruction &I : instructions(*F)) {
928 // Bail if CS is a convergent call to a function not in the SCC.
929 if (CS && CS.isConvergent() &&
930 SCCNodes.count(CS.getCalledFunction()) == 0)
935 // If the SCC doesn't have any convergent functions, we have nothing to do.
936 if (!HasConvergentFn) return false;
938 // If we got here, all of the calls the SCC makes to functions not in the SCC
939 // are non-convergent. Therefore all of the SCC's functions can also be made
940 // non-convergent. We'll remove the attr from the callsites in
942 for (Function *F : SCCNodes) {
943 if (!F->isConvergent()) continue;
945 DEBUG(dbgs() << "Removing convergent attr from fn " << F->getName()
947 F->setNotConvergent();
952 static bool setDoesNotRecurse(Function &F) {
953 if (F.doesNotRecurse())
955 F.setDoesNotRecurse();
960 static bool addNoRecurseAttrs(const SCCNodeSet &SCCNodes) {
961 // Try and identify functions that do not recurse.
963 // If the SCC contains multiple nodes we know for sure there is recursion.
964 if (SCCNodes.size() != 1)
967 Function *F = *SCCNodes.begin();
968 if (!F || F->isDeclaration() || F->doesNotRecurse())
971 // If all of the calls in F are identifiable and are to norecurse functions, F
972 // is norecurse. This check also detects self-recursion as F is not currently
973 // marked norecurse, so any called from F to F will not be marked norecurse.
974 for (Instruction &I : instructions(*F))
975 if (auto CS = CallSite(&I)) {
976 Function *Callee = CS.getCalledFunction();
977 if (!Callee || Callee == F || !Callee->doesNotRecurse())
978 // Function calls a potentially recursive function.
982 // Every call was to a non-recursive function other than this function, and
983 // we have no indirect recursion as the SCC size is one. This function cannot
985 return setDoesNotRecurse(*F);
988 PreservedAnalyses PostOrderFunctionAttrsPass::run(LazyCallGraph::SCC &C,
989 CGSCCAnalysisManager &AM) {
990 FunctionAnalysisManager &FAM =
991 AM.getResult<FunctionAnalysisManagerCGSCCProxy>(C).getManager();
993 // We pass a lambda into functions to wire them up to the analysis manager
994 // for getting function analyses.
995 auto AARGetter = [&](Function &F) -> AAResults & {
996 return FAM.getResult<AAManager>(F);
999 // Fill SCCNodes with the elements of the SCC. Also track whether there are
1000 // any external or opt-none nodes that will prevent us from optimizing any
1002 SCCNodeSet SCCNodes;
1003 bool HasUnknownCall = false;
1004 for (LazyCallGraph::Node &N : C) {
1005 Function &F = N.getFunction();
1006 if (F.hasFnAttribute(Attribute::OptimizeNone)) {
1007 // Treat any function we're trying not to optimize as if it were an
1008 // indirect call and omit it from the node set used below.
1009 HasUnknownCall = true;
1012 // Track whether any functions in this SCC have an unknown call edge.
1013 // Note: if this is ever a performance hit, we can common it with
1014 // subsequent routines which also do scans over the instructions of the
1016 if (!HasUnknownCall)
1017 for (Instruction &I : instructions(F))
1018 if (auto CS = CallSite(&I))
1019 if (!CS.getCalledFunction()) {
1020 HasUnknownCall = true;
1024 SCCNodes.insert(&F);
1027 bool Changed = false;
1028 Changed |= addReadAttrs(SCCNodes, AARGetter);
1029 Changed |= addArgumentAttrs(SCCNodes);
1031 // If we have no external nodes participating in the SCC, we can deduce some
1032 // more precise attributes as well.
1033 if (!HasUnknownCall) {
1034 Changed |= addNoAliasAttrs(SCCNodes);
1035 Changed |= addNonNullAttrs(SCCNodes);
1036 Changed |= removeConvergentAttrs(SCCNodes);
1037 Changed |= addNoRecurseAttrs(SCCNodes);
1040 return Changed ? PreservedAnalyses::none() : PreservedAnalyses::all();
1044 struct PostOrderFunctionAttrsLegacyPass : public CallGraphSCCPass {
1045 static char ID; // Pass identification, replacement for typeid
1046 PostOrderFunctionAttrsLegacyPass() : CallGraphSCCPass(ID) {
1047 initializePostOrderFunctionAttrsLegacyPassPass(*PassRegistry::getPassRegistry());
1050 bool runOnSCC(CallGraphSCC &SCC) override;
1052 void getAnalysisUsage(AnalysisUsage &AU) const override {
1053 AU.setPreservesCFG();
1054 AU.addRequired<AssumptionCacheTracker>();
1055 getAAResultsAnalysisUsage(AU);
1056 CallGraphSCCPass::getAnalysisUsage(AU);
1061 char PostOrderFunctionAttrsLegacyPass::ID = 0;
1062 INITIALIZE_PASS_BEGIN(PostOrderFunctionAttrsLegacyPass, "functionattrs",
1063 "Deduce function attributes", false, false)
1064 INITIALIZE_PASS_DEPENDENCY(AssumptionCacheTracker)
1065 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
1066 INITIALIZE_PASS_END(PostOrderFunctionAttrsLegacyPass, "functionattrs",
1067 "Deduce function attributes", false, false)
1069 Pass *llvm::createPostOrderFunctionAttrsLegacyPass() { return new PostOrderFunctionAttrsLegacyPass(); }
1071 template <typename AARGetterT>
1072 static bool runImpl(CallGraphSCC &SCC, AARGetterT AARGetter) {
1073 bool Changed = false;
1075 // Fill SCCNodes with the elements of the SCC. Used for quickly looking up
1076 // whether a given CallGraphNode is in this SCC. Also track whether there are
1077 // any external or opt-none nodes that will prevent us from optimizing any
1079 SCCNodeSet SCCNodes;
1080 bool ExternalNode = false;
1081 for (CallGraphNode *I : SCC) {
1082 Function *F = I->getFunction();
1083 if (!F || F->hasFnAttribute(Attribute::OptimizeNone)) {
1084 // External node or function we're trying not to optimize - we both avoid
1085 // transform them and avoid leveraging information they provide.
1086 ExternalNode = true;
1093 Changed |= addReadAttrs(SCCNodes, AARGetter);
1094 Changed |= addArgumentAttrs(SCCNodes);
1096 // If we have no external nodes participating in the SCC, we can deduce some
1097 // more precise attributes as well.
1098 if (!ExternalNode) {
1099 Changed |= addNoAliasAttrs(SCCNodes);
1100 Changed |= addNonNullAttrs(SCCNodes);
1101 Changed |= removeConvergentAttrs(SCCNodes);
1102 Changed |= addNoRecurseAttrs(SCCNodes);
1108 bool PostOrderFunctionAttrsLegacyPass::runOnSCC(CallGraphSCC &SCC) {
1112 // We compute dedicated AA results for each function in the SCC as needed. We
1113 // use a lambda referencing external objects so that they live long enough to
1114 // be queried, but we re-use them each time.
1115 Optional<BasicAAResult> BAR;
1116 Optional<AAResults> AAR;
1117 auto AARGetter = [&](Function &F) -> AAResults & {
1118 BAR.emplace(createLegacyPMBasicAAResult(*this, F));
1119 AAR.emplace(createLegacyPMAAResults(*this, F, *BAR));
1123 return runImpl(SCC, AARGetter);
1127 struct ReversePostOrderFunctionAttrsLegacyPass : public ModulePass {
1128 static char ID; // Pass identification, replacement for typeid
1129 ReversePostOrderFunctionAttrsLegacyPass() : ModulePass(ID) {
1130 initializeReversePostOrderFunctionAttrsLegacyPassPass(*PassRegistry::getPassRegistry());
1133 bool runOnModule(Module &M) override;
1135 void getAnalysisUsage(AnalysisUsage &AU) const override {
1136 AU.setPreservesCFG();
1137 AU.addRequired<CallGraphWrapperPass>();
1138 AU.addPreserved<CallGraphWrapperPass>();
1143 char ReversePostOrderFunctionAttrsLegacyPass::ID = 0;
1144 INITIALIZE_PASS_BEGIN(ReversePostOrderFunctionAttrsLegacyPass, "rpo-functionattrs",
1145 "Deduce function attributes in RPO", false, false)
1146 INITIALIZE_PASS_DEPENDENCY(CallGraphWrapperPass)
1147 INITIALIZE_PASS_END(ReversePostOrderFunctionAttrsLegacyPass, "rpo-functionattrs",
1148 "Deduce function attributes in RPO", false, false)
1150 Pass *llvm::createReversePostOrderFunctionAttrsPass() {
1151 return new ReversePostOrderFunctionAttrsLegacyPass();
1154 static bool addNoRecurseAttrsTopDown(Function &F) {
1155 // We check the preconditions for the function prior to calling this to avoid
1156 // the cost of building up a reversible post-order list. We assert them here
1157 // to make sure none of the invariants this relies on were violated.
1158 assert(!F.isDeclaration() && "Cannot deduce norecurse without a definition!");
1159 assert(!F.doesNotRecurse() &&
1160 "This function has already been deduced as norecurs!");
1161 assert(F.hasInternalLinkage() &&
1162 "Can only do top-down deduction for internal linkage functions!");
1164 // If F is internal and all of its uses are calls from a non-recursive
1165 // functions, then none of its calls could in fact recurse without going
1166 // through a function marked norecurse, and so we can mark this function too
1167 // as norecurse. Note that the uses must actually be calls -- otherwise
1168 // a pointer to this function could be returned from a norecurse function but
1169 // this function could be recursively (indirectly) called. Note that this
1170 // also detects if F is directly recursive as F is not yet marked as
1171 // a norecurse function.
1172 for (auto *U : F.users()) {
1173 auto *I = dyn_cast<Instruction>(U);
1177 if (!CS || !CS.getParent()->getParent()->doesNotRecurse())
1180 return setDoesNotRecurse(F);
1183 static bool deduceFunctionAttributeInRPO(Module &M, CallGraph &CG) {
1184 // We only have a post-order SCC traversal (because SCCs are inherently
1185 // discovered in post-order), so we accumulate them in a vector and then walk
1186 // it in reverse. This is simpler than using the RPO iterator infrastructure
1187 // because we need to combine SCC detection and the PO walk of the call
1188 // graph. We can also cheat egregiously because we're primarily interested in
1189 // synthesizing norecurse and so we can only save the singular SCCs as SCCs
1190 // with multiple functions in them will clearly be recursive.
1191 SmallVector<Function *, 16> Worklist;
1192 for (scc_iterator<CallGraph *> I = scc_begin(&CG); !I.isAtEnd(); ++I) {
1196 Function *F = I->front()->getFunction();
1197 if (F && !F->isDeclaration() && !F->doesNotRecurse() &&
1198 F->hasInternalLinkage())
1199 Worklist.push_back(F);
1202 bool Changed = false;
1203 for (auto *F : reverse(Worklist))
1204 Changed |= addNoRecurseAttrsTopDown(*F);
1209 bool ReversePostOrderFunctionAttrsLegacyPass::runOnModule(Module &M) {
1213 auto &CG = getAnalysis<CallGraphWrapperPass>().getCallGraph();
1215 return deduceFunctionAttributeInRPO(M, CG);
1219 ReversePostOrderFunctionAttrsPass::run(Module &M, AnalysisManager<Module> &AM) {
1220 auto &CG = AM.getResult<CallGraphAnalysis>(M);
1222 bool Changed = deduceFunctionAttributeInRPO(M, CG);
1224 return PreservedAnalyses::all();
1225 PreservedAnalyses PA;
1226 PA.preserve<CallGraphAnalysis>();