1 //===- FunctionAttrs.cpp - Pass which marks functions readnone or readonly ===//
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 file implements a simple interprocedural pass which walks the
11 // call-graph, looking for functions which do not access or only read
12 // non-local memory, and marking them readnone/readonly. In addition,
13 // it marks function arguments (of pointer type) 'nocapture' if a call
14 // to the function does not create any copies of the pointer value that
15 // outlive the call. This more or less means that the pointer is only
16 // dereferenced, and not returned from the function or stored in a global.
17 // This pass is implemented as a bottom-up traversal of the call-graph.
19 //===----------------------------------------------------------------------===//
21 #define DEBUG_TYPE "functionattrs"
22 #include "llvm/Transforms/IPO.h"
23 #include "llvm/ADT/SCCIterator.h"
24 #include "llvm/ADT/SetVector.h"
25 #include "llvm/ADT/SmallSet.h"
26 #include "llvm/ADT/Statistic.h"
27 #include "llvm/Analysis/AliasAnalysis.h"
28 #include "llvm/Analysis/CallGraph.h"
29 #include "llvm/Analysis/CallGraphSCCPass.h"
30 #include "llvm/Analysis/CaptureTracking.h"
31 #include "llvm/IR/GlobalVariable.h"
32 #include "llvm/IR/IntrinsicInst.h"
33 #include "llvm/IR/LLVMContext.h"
34 #include "llvm/Support/InstIterator.h"
37 STATISTIC(NumReadNone, "Number of functions marked readnone");
38 STATISTIC(NumReadOnly, "Number of functions marked readonly");
39 STATISTIC(NumNoCapture, "Number of arguments marked nocapture");
40 STATISTIC(NumNoAlias, "Number of function returns marked noalias");
43 struct FunctionAttrs : public CallGraphSCCPass {
44 static char ID; // Pass identification, replacement for typeid
45 FunctionAttrs() : CallGraphSCCPass(ID), AA(0) {
46 initializeFunctionAttrsPass(*PassRegistry::getPassRegistry());
49 // runOnSCC - Analyze the SCC, performing the transformation if possible.
50 bool runOnSCC(CallGraphSCC &SCC);
52 // AddReadAttrs - Deduce readonly/readnone attributes for the SCC.
53 bool AddReadAttrs(const CallGraphSCC &SCC);
55 // AddNoCaptureAttrs - Deduce nocapture attributes for the SCC.
56 bool AddNoCaptureAttrs(const CallGraphSCC &SCC);
58 // IsFunctionMallocLike - Does this function allocate new memory?
59 bool IsFunctionMallocLike(Function *F,
60 SmallPtrSet<Function*, 8> &) const;
62 // AddNoAliasAttrs - Deduce noalias attributes for the SCC.
63 bool AddNoAliasAttrs(const CallGraphSCC &SCC);
65 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
67 AU.addRequired<AliasAnalysis>();
68 CallGraphSCCPass::getAnalysisUsage(AU);
76 char FunctionAttrs::ID = 0;
77 INITIALIZE_PASS_BEGIN(FunctionAttrs, "functionattrs",
78 "Deduce function attributes", false, false)
79 INITIALIZE_AG_DEPENDENCY(CallGraph)
80 INITIALIZE_PASS_END(FunctionAttrs, "functionattrs",
81 "Deduce function attributes", false, false)
83 Pass *llvm::createFunctionAttrsPass() { return new FunctionAttrs(); }
86 /// AddReadAttrs - Deduce readonly/readnone attributes for the SCC.
87 bool FunctionAttrs::AddReadAttrs(const CallGraphSCC &SCC) {
88 SmallPtrSet<Function*, 8> SCCNodes;
90 // Fill SCCNodes with the elements of the SCC. Used for quickly
91 // looking up whether a given CallGraphNode is in this SCC.
92 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I)
93 SCCNodes.insert((*I)->getFunction());
95 // Check if any of the functions in the SCC read or write memory. If they
96 // write memory then they can't be marked readnone or readonly.
97 bool ReadsMemory = false;
98 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
99 Function *F = (*I)->getFunction();
102 // External node - may write memory. Just give up.
105 AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(F);
106 if (MRB == AliasAnalysis::DoesNotAccessMemory)
110 // Definitions with weak linkage may be overridden at linktime with
111 // something that writes memory, so treat them like declarations.
112 if (F->isDeclaration() || F->mayBeOverridden()) {
113 if (!AliasAnalysis::onlyReadsMemory(MRB))
114 // May write memory. Just give up.
121 // Scan the function body for instructions that may read or write memory.
122 for (inst_iterator II = inst_begin(F), E = inst_end(F); II != E; ++II) {
123 Instruction *I = &*II;
125 // Some instructions can be ignored even if they read or write memory.
126 // Detect these now, skipping to the next instruction if one is found.
127 CallSite CS(cast<Value>(I));
129 // Ignore calls to functions in the same SCC.
130 if (CS.getCalledFunction() && SCCNodes.count(CS.getCalledFunction()))
132 AliasAnalysis::ModRefBehavior MRB = AA->getModRefBehavior(CS);
133 // If the call doesn't access arbitrary memory, we may be able to
134 // figure out something.
135 if (AliasAnalysis::onlyAccessesArgPointees(MRB)) {
136 // If the call does access argument pointees, check each argument.
137 if (AliasAnalysis::doesAccessArgPointees(MRB))
138 // Check whether all pointer arguments point to local memory, and
139 // ignore calls that only access local memory.
140 for (CallSite::arg_iterator CI = CS.arg_begin(), CE = CS.arg_end();
143 if (Arg->getType()->isPointerTy()) {
144 AliasAnalysis::Location Loc(Arg,
145 AliasAnalysis::UnknownSize,
146 I->getMetadata(LLVMContext::MD_tbaa));
147 if (!AA->pointsToConstantMemory(Loc, /*OrLocal=*/true)) {
148 if (MRB & AliasAnalysis::Mod)
149 // Writes non-local memory. Give up.
151 if (MRB & AliasAnalysis::Ref)
152 // Ok, it reads non-local memory.
159 // The call could access any memory. If that includes writes, give up.
160 if (MRB & AliasAnalysis::Mod)
162 // If it reads, note it.
163 if (MRB & AliasAnalysis::Ref)
166 } else if (LoadInst *LI = dyn_cast<LoadInst>(I)) {
167 // Ignore non-volatile loads from local memory. (Atomic is okay here.)
168 if (!LI->isVolatile()) {
169 AliasAnalysis::Location Loc = AA->getLocation(LI);
170 if (AA->pointsToConstantMemory(Loc, /*OrLocal=*/true))
173 } else if (StoreInst *SI = dyn_cast<StoreInst>(I)) {
174 // Ignore non-volatile stores to local memory. (Atomic is okay here.)
175 if (!SI->isVolatile()) {
176 AliasAnalysis::Location Loc = AA->getLocation(SI);
177 if (AA->pointsToConstantMemory(Loc, /*OrLocal=*/true))
180 } else if (VAArgInst *VI = dyn_cast<VAArgInst>(I)) {
181 // Ignore vaargs on local memory.
182 AliasAnalysis::Location Loc = AA->getLocation(VI);
183 if (AA->pointsToConstantMemory(Loc, /*OrLocal=*/true))
187 // Any remaining instructions need to be taken seriously! Check if they
188 // read or write memory.
189 if (I->mayWriteToMemory())
190 // Writes memory. Just give up.
193 // If this instruction may read memory, remember that.
194 ReadsMemory |= I->mayReadFromMemory();
198 // Success! Functions in this SCC do not access memory, or only read memory.
199 // Give them the appropriate attribute.
200 bool MadeChange = false;
201 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
202 Function *F = (*I)->getFunction();
204 if (F->doesNotAccessMemory())
208 if (F->onlyReadsMemory() && ReadsMemory)
214 // Clear out any existing attributes.
216 B.addAttribute(Attribute::ReadOnly)
217 .addAttribute(Attribute::ReadNone);
218 F->removeAttributes(AttributeSet::FunctionIndex,
219 AttributeSet::get(F->getContext(),
220 AttributeSet::FunctionIndex, B));
222 // Add in the new attribute.
223 F->addAttribute(AttributeSet::FunctionIndex,
224 ReadsMemory ? Attribute::ReadOnly : Attribute::ReadNone);
236 // For a given pointer Argument, this retains a list of Arguments of functions
237 // in the same SCC that the pointer data flows into. We use this to build an
238 // SCC of the arguments.
239 struct ArgumentGraphNode {
240 Argument *Definition;
241 SmallVector<ArgumentGraphNode*, 4> Uses;
244 class ArgumentGraph {
245 // We store pointers to ArgumentGraphNode objects, so it's important that
246 // that they not move around upon insert.
247 typedef std::map<Argument*, ArgumentGraphNode> ArgumentMapTy;
249 ArgumentMapTy ArgumentMap;
251 // There is no root node for the argument graph, in fact:
252 // void f(int *x, int *y) { if (...) f(x, y); }
253 // is an example where the graph is disconnected. The SCCIterator requires a
254 // single entry point, so we maintain a fake ("synthetic") root node that
255 // uses every node. Because the graph is directed and nothing points into
256 // the root, it will not participate in any SCCs (except for its own).
257 ArgumentGraphNode SyntheticRoot;
260 ArgumentGraph() { SyntheticRoot.Definition = 0; }
262 typedef SmallVectorImpl<ArgumentGraphNode*>::iterator iterator;
264 iterator begin() { return SyntheticRoot.Uses.begin(); }
265 iterator end() { return SyntheticRoot.Uses.end(); }
266 ArgumentGraphNode *getEntryNode() { return &SyntheticRoot; }
268 ArgumentGraphNode *operator[](Argument *A) {
269 ArgumentGraphNode &Node = ArgumentMap[A];
271 SyntheticRoot.Uses.push_back(&Node);
276 // This tracker checks whether callees are in the SCC, and if so it does not
277 // consider that a capture, instead adding it to the "Uses" list and
278 // continuing with the analysis.
279 struct ArgumentUsesTracker : public CaptureTracker {
280 ArgumentUsesTracker(const SmallPtrSet<Function*, 8> &SCCNodes)
281 : Captured(false), SCCNodes(SCCNodes) {}
283 void tooManyUses() { Captured = true; }
285 bool captured(Use *U) {
286 CallSite CS(U->getUser());
287 if (!CS.getInstruction()) { Captured = true; return true; }
289 Function *F = CS.getCalledFunction();
290 if (!F || !SCCNodes.count(F)) { Captured = true; return true; }
292 Function::arg_iterator AI = F->arg_begin(), AE = F->arg_end();
293 for (CallSite::arg_iterator PI = CS.arg_begin(), PE = CS.arg_end();
294 PI != PE; ++PI, ++AI) {
296 assert(F->isVarArg() && "More params than args in non-varargs call");
305 assert(!Uses.empty() && "Capturing call-site captured nothing?");
309 bool Captured; // True only if certainly captured (used outside our SCC).
310 SmallVector<Argument*, 4> Uses; // Uses within our SCC.
312 const SmallPtrSet<Function*, 8> &SCCNodes;
317 template<> struct GraphTraits<ArgumentGraphNode*> {
318 typedef ArgumentGraphNode NodeType;
319 typedef SmallVectorImpl<ArgumentGraphNode*>::iterator ChildIteratorType;
321 static inline NodeType *getEntryNode(NodeType *A) { return A; }
322 static inline ChildIteratorType child_begin(NodeType *N) {
323 return N->Uses.begin();
325 static inline ChildIteratorType child_end(NodeType *N) {
326 return N->Uses.end();
329 template<> struct GraphTraits<ArgumentGraph*>
330 : public GraphTraits<ArgumentGraphNode*> {
331 static NodeType *getEntryNode(ArgumentGraph *AG) {
332 return AG->getEntryNode();
334 static ChildIteratorType nodes_begin(ArgumentGraph *AG) {
337 static ChildIteratorType nodes_end(ArgumentGraph *AG) {
343 /// AddNoCaptureAttrs - Deduce nocapture attributes for the SCC.
344 bool FunctionAttrs::AddNoCaptureAttrs(const CallGraphSCC &SCC) {
345 bool Changed = false;
347 SmallPtrSet<Function*, 8> SCCNodes;
349 // Fill SCCNodes with the elements of the SCC. Used for quickly
350 // looking up whether a given CallGraphNode is in this SCC.
351 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
352 Function *F = (*I)->getFunction();
353 if (F && !F->isDeclaration() && !F->mayBeOverridden())
360 B.addAttribute(Attribute::NoCapture);
362 // Check each function in turn, determining which pointer arguments are not
364 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
365 Function *F = (*I)->getFunction();
368 // External node - only a problem for arguments that we pass to it.
371 // Definitions with weak linkage may be overridden at linktime with
372 // something that captures pointers, so treat them like declarations.
373 if (F->isDeclaration() || F->mayBeOverridden())
376 // Functions that are readonly (or readnone) and nounwind and don't return
377 // a value can't capture arguments. Don't analyze them.
378 if (F->onlyReadsMemory() && F->doesNotThrow() &&
379 F->getReturnType()->isVoidTy()) {
380 for (Function::arg_iterator A = F->arg_begin(), E = F->arg_end();
382 if (A->getType()->isPointerTy() && !A->hasNoCaptureAttr()) {
383 A->addAttr(Attribute::get(F->getContext(), B));
391 for (Function::arg_iterator A = F->arg_begin(), E = F->arg_end(); A!=E; ++A)
392 if (A->getType()->isPointerTy() && !A->hasNoCaptureAttr()) {
393 ArgumentUsesTracker Tracker(SCCNodes);
394 PointerMayBeCaptured(A, &Tracker);
395 if (!Tracker.Captured) {
396 if (Tracker.Uses.empty()) {
397 // If it's trivially not captured, mark it nocapture now.
398 A->addAttr(Attribute::get(F->getContext(), B));
402 // If it's not trivially captured and not trivially not captured,
403 // then it must be calling into another function in our SCC. Save
404 // its particulars for Argument-SCC analysis later.
405 ArgumentGraphNode *Node = AG[A];
406 for (SmallVectorImpl<Argument*>::iterator UI = Tracker.Uses.begin(),
407 UE = Tracker.Uses.end(); UI != UE; ++UI)
408 Node->Uses.push_back(AG[*UI]);
411 // Otherwise, it's captured. Don't bother doing SCC analysis on it.
415 // The graph we've collected is partial because we stopped scanning for
416 // argument uses once we solved the argument trivially. These partial nodes
417 // show up as ArgumentGraphNode objects with an empty Uses list, and for
418 // these nodes the final decision about whether they capture has already been
419 // made. If the definition doesn't have a 'nocapture' attribute by now, it
422 for (scc_iterator<ArgumentGraph*> I = scc_begin(&AG), E = scc_end(&AG);
424 std::vector<ArgumentGraphNode*> &ArgumentSCC = *I;
425 if (ArgumentSCC.size() == 1) {
426 if (!ArgumentSCC[0]->Definition) continue; // synthetic root node
428 // eg. "void f(int* x) { if (...) f(x); }"
429 if (ArgumentSCC[0]->Uses.size() == 1 &&
430 ArgumentSCC[0]->Uses[0] == ArgumentSCC[0]) {
433 addAttr(Attribute::get(ArgumentSCC[0]->Definition->getContext(), B));
440 bool SCCCaptured = false;
441 for (std::vector<ArgumentGraphNode*>::iterator I = ArgumentSCC.begin(),
442 E = ArgumentSCC.end(); I != E && !SCCCaptured; ++I) {
443 ArgumentGraphNode *Node = *I;
444 if (Node->Uses.empty()) {
445 if (!Node->Definition->hasNoCaptureAttr())
449 if (SCCCaptured) continue;
451 SmallPtrSet<Argument*, 8> ArgumentSCCNodes;
452 // Fill ArgumentSCCNodes with the elements of the ArgumentSCC. Used for
453 // quickly looking up whether a given Argument is in this ArgumentSCC.
454 for (std::vector<ArgumentGraphNode*>::iterator I = ArgumentSCC.begin(),
455 E = ArgumentSCC.end(); I != E; ++I) {
456 ArgumentSCCNodes.insert((*I)->Definition);
459 for (std::vector<ArgumentGraphNode*>::iterator I = ArgumentSCC.begin(),
460 E = ArgumentSCC.end(); I != E && !SCCCaptured; ++I) {
461 ArgumentGraphNode *N = *I;
462 for (SmallVectorImpl<ArgumentGraphNode*>::iterator UI = N->Uses.begin(),
463 UE = N->Uses.end(); UI != UE; ++UI) {
464 Argument *A = (*UI)->Definition;
465 if (A->hasNoCaptureAttr() || ArgumentSCCNodes.count(A))
471 if (SCCCaptured) continue;
473 for (unsigned i = 0, e = ArgumentSCC.size(); i != e; ++i) {
474 Argument *A = ArgumentSCC[i]->Definition;
475 A->addAttr(Attribute::get(A->getContext(), B));
484 /// IsFunctionMallocLike - A function is malloc-like if it returns either null
485 /// or a pointer that doesn't alias any other pointer visible to the caller.
486 bool FunctionAttrs::IsFunctionMallocLike(Function *F,
487 SmallPtrSet<Function*, 8> &SCCNodes) const {
488 SmallSetVector<Value *, 8> FlowsToReturn;
489 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
490 if (ReturnInst *Ret = dyn_cast<ReturnInst>(I->getTerminator()))
491 FlowsToReturn.insert(Ret->getReturnValue());
493 for (unsigned i = 0; i != FlowsToReturn.size(); ++i) {
494 Value *RetVal = FlowsToReturn[i];
496 if (Constant *C = dyn_cast<Constant>(RetVal)) {
497 if (!C->isNullValue() && !isa<UndefValue>(C))
503 if (isa<Argument>(RetVal))
506 if (Instruction *RVI = dyn_cast<Instruction>(RetVal))
507 switch (RVI->getOpcode()) {
508 // Extend the analysis by looking upwards.
509 case Instruction::BitCast:
510 case Instruction::GetElementPtr:
511 FlowsToReturn.insert(RVI->getOperand(0));
513 case Instruction::Select: {
514 SelectInst *SI = cast<SelectInst>(RVI);
515 FlowsToReturn.insert(SI->getTrueValue());
516 FlowsToReturn.insert(SI->getFalseValue());
519 case Instruction::PHI: {
520 PHINode *PN = cast<PHINode>(RVI);
521 for (int i = 0, e = PN->getNumIncomingValues(); i != e; ++i)
522 FlowsToReturn.insert(PN->getIncomingValue(i));
526 // Check whether the pointer came from an allocation.
527 case Instruction::Alloca:
529 case Instruction::Call:
530 case Instruction::Invoke: {
532 if (CS.paramHasAttr(0, Attribute::NoAlias))
534 if (CS.getCalledFunction() &&
535 SCCNodes.count(CS.getCalledFunction()))
539 return false; // Did not come from an allocation.
542 if (PointerMayBeCaptured(RetVal, false, /*StoreCaptures=*/false))
549 /// AddNoAliasAttrs - Deduce noalias attributes for the SCC.
550 bool FunctionAttrs::AddNoAliasAttrs(const CallGraphSCC &SCC) {
551 SmallPtrSet<Function*, 8> SCCNodes;
553 // Fill SCCNodes with the elements of the SCC. Used for quickly
554 // looking up whether a given CallGraphNode is in this SCC.
555 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I)
556 SCCNodes.insert((*I)->getFunction());
558 // Check each function in turn, determining which functions return noalias
560 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
561 Function *F = (*I)->getFunction();
564 // External node - skip it;
568 if (F->doesNotAlias(0))
571 // Definitions with weak linkage may be overridden at linktime, so
572 // treat them like declarations.
573 if (F->isDeclaration() || F->mayBeOverridden())
576 // We annotate noalias return values, which are only applicable to
578 if (!F->getReturnType()->isPointerTy())
581 if (!IsFunctionMallocLike(F, SCCNodes))
585 bool MadeChange = false;
586 for (CallGraphSCC::iterator I = SCC.begin(), E = SCC.end(); I != E; ++I) {
587 Function *F = (*I)->getFunction();
588 if (F->doesNotAlias(0) || !F->getReturnType()->isPointerTy())
591 F->setDoesNotAlias(0);
599 bool FunctionAttrs::runOnSCC(CallGraphSCC &SCC) {
600 AA = &getAnalysis<AliasAnalysis>();
602 bool Changed = AddReadAttrs(SCC);
603 Changed |= AddNoCaptureAttrs(SCC);
604 Changed |= AddNoAliasAttrs(SCC);