1 //===- MergeFunctions.cpp - Merge identical functions ---------------------===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // This pass looks for equivalent functions that are mergable and folds them.
10 //
11 // Order relation is defined on set of functions. It was made through
12 // special function comparison procedure that returns
13 // 0 when functions are equal,
14 // -1 when Left function is less than right function, and
15 // 1 for opposite case. We need total-ordering, so we need to maintain
16 // four properties on the functions set:
17 // a <= a (reflexivity)
18 // if a <= b and b <= a then a = b (antisymmetry)
19 // if a <= b and b <= c then a <= c (transitivity).
20 // for all a and b: a <= b or b <= a (totality).
21 //
22 // Comparison iterates through each instruction in each basic block.
23 // Functions are kept on binary tree. For each new function F we perform
24 // lookup in binary tree.
25 // In practice it works the following way:
26 // -- We define Function* container class with custom "operator<" (FunctionPtr).
27 // -- "FunctionPtr" instances are stored in std::set collection, so every
28 // std::set::insert operation will give you result in log(N) time.
29 //
30 // As an optimization, a hash of the function structure is calculated first, and
31 // two functions are only compared if they have the same hash. This hash is
32 // cheap to compute, and has the property that if function F == G according to
33 // the comparison function, then hash(F) == hash(G). This consistency property
34 // is critical to ensuring all possible merging opportunities are exploited.
35 // Collisions in the hash affect the speed of the pass but not the correctness
36 // or determinism of the resulting transformation.
37 //
38 // When a match is found the functions are folded. If both functions are
39 // overridable, we move the functionality into a new internal function and
40 // leave two overridable thunks to it.
41 //
42 //===----------------------------------------------------------------------===//
43 //
44 // Future work:
45 //
46 // * virtual functions.
47 //
48 // Many functions have their address taken by the virtual function table for
49 // the object they belong to. However, as long as it's only used for a lookup
50 // and call, this is irrelevant, and we'd like to fold such functions.
51 //
52 // * be smarter about bitcasts.
53 //
54 // In order to fold functions, we will sometimes add either bitcast instructions
55 // or bitcast constant expressions. Unfortunately, this can confound further
56 // analysis since the two functions differ where one has a bitcast and the
57 // other doesn't. We should learn to look through bitcasts.
58 //
59 // * Compare complex types with pointer types inside.
60 // * Compare cross-reference cases.
61 // * Compare complex expressions.
62 //
63 // All the three issues above could be described as ability to prove that
64 // fA == fB == fC == fE == fF == fG in example below:
65 //
66 // void fA() {
67 // fB();
68 // }
69 // void fB() {
70 // fA();
71 // }
72 //
73 // void fE() {
74 // fF();
75 // }
76 // void fF() {
77 // fG();
78 // }
79 // void fG() {
80 // fE();
81 // }
82 //
83 // Simplest cross-reference case (fA <--> fB) was implemented in previous
84 // versions of MergeFunctions, though it presented only in two function pairs
85 // in test-suite (that counts >50k functions)
86 // Though possibility to detect complex cross-referencing (e.g.: A->B->C->D->A)
87 // could cover much more cases.
88 //
89 //===----------------------------------------------------------------------===//
90
91 #include "llvm/ADT/ArrayRef.h"
92 #include "llvm/ADT/SmallPtrSet.h"
93 #include "llvm/ADT/SmallVector.h"
94 #include "llvm/ADT/Statistic.h"
95 #include "llvm/IR/Argument.h"
96 #include "llvm/IR/Attributes.h"
97 #include "llvm/IR/BasicBlock.h"
98 #include "llvm/IR/Constant.h"
99 #include "llvm/IR/Constants.h"
100 #include "llvm/IR/DebugInfoMetadata.h"
101 #include "llvm/IR/DebugLoc.h"
102 #include "llvm/IR/DerivedTypes.h"
103 #include "llvm/IR/Function.h"
104 #include "llvm/IR/GlobalValue.h"
105 #include "llvm/IR/IRBuilder.h"
106 #include "llvm/IR/InstrTypes.h"
107 #include "llvm/IR/Instruction.h"
108 #include "llvm/IR/Instructions.h"
109 #include "llvm/IR/IntrinsicInst.h"
110 #include "llvm/IR/Module.h"
111 #include "llvm/IR/Type.h"
112 #include "llvm/IR/Use.h"
113 #include "llvm/IR/User.h"
114 #include "llvm/IR/Value.h"
115 #include "llvm/IR/ValueHandle.h"
116 #include "llvm/IR/ValueMap.h"
117 #include "llvm/InitializePasses.h"
118 #include "llvm/Pass.h"
119 #include "llvm/Support/Casting.h"
120 #include "llvm/Support/CommandLine.h"
121 #include "llvm/Support/Debug.h"
122 #include "llvm/Support/raw_ostream.h"
123 #include "llvm/Transforms/IPO.h"
124 #include "llvm/Transforms/IPO/MergeFunctions.h"
125 #include "llvm/Transforms/Utils/FunctionComparator.h"
126 #include <algorithm>
127 #include <cassert>
128 #include <iterator>
129 #include <set>
130 #include <utility>
131 #include <vector>
132
133 using namespace llvm;
134
135 #define DEBUG_TYPE "mergefunc"
136
137 STATISTIC(NumFunctionsMerged, "Number of functions merged");
138 STATISTIC(NumThunksWritten, "Number of thunks generated");
139 STATISTIC(NumAliasesWritten, "Number of aliases generated");
140 STATISTIC(NumDoubleWeak, "Number of new functions created");
141
142 static cl::opt<unsigned> NumFunctionsForSanityCheck(
143 "mergefunc-sanity",
144 cl::desc("How many functions in module could be used for "
145 "MergeFunctions pass sanity check. "
146 "'0' disables this check. Works only with '-debug' key."),
147 cl::init(0), cl::Hidden);
148
149 // Under option -mergefunc-preserve-debug-info we:
150 // - Do not create a new function for a thunk.
151 // - Retain the debug info for a thunk's parameters (and associated
152 // instructions for the debug info) from the entry block.
153 // Note: -debug will display the algorithm at work.
154 // - Create debug-info for the call (to the shared implementation) made by
155 // a thunk and its return value.
156 // - Erase the rest of the function, retaining the (minimally sized) entry
157 // block to create a thunk.
158 // - Preserve a thunk's call site to point to the thunk even when both occur
159 // within the same translation unit, to aid debugability. Note that this
160 // behaviour differs from the underlying -mergefunc implementation which
161 // modifies the thunk's call site to point to the shared implementation
162 // when both occur within the same translation unit.
163 static cl::opt<bool>
164 MergeFunctionsPDI("mergefunc-preserve-debug-info", cl::Hidden,
165 cl::init(false),
166 cl::desc("Preserve debug info in thunk when mergefunc "
167 "transformations are made."));
168
169 static cl::opt<bool>
170 MergeFunctionsAliases("mergefunc-use-aliases", cl::Hidden,
171 cl::init(false),
172 cl::desc("Allow mergefunc to create aliases"));
173
174 namespace {
175
176 class FunctionNode {
177 mutable AssertingVH<Function> F;
178 FunctionComparator::FunctionHash Hash;
179
180 public:
181 // Note the hash is recalculated potentially multiple times, but it is cheap.
FunctionNode(Function * F)182 FunctionNode(Function *F)
183 : F(F), Hash(FunctionComparator::functionHash(*F)) {}
184
getFunc() const185 Function *getFunc() const { return F; }
getHash() const186 FunctionComparator::FunctionHash getHash() const { return Hash; }
187
188 /// Replace the reference to the function F by the function G, assuming their
189 /// implementations are equal.
replaceBy(Function * G) const190 void replaceBy(Function *G) const {
191 F = G;
192 }
193 };
194
195 /// MergeFunctions finds functions which will generate identical machine code,
196 /// by considering all pointer types to be equivalent. Once identified,
197 /// MergeFunctions will fold them by replacing a call to one to a call to a
198 /// bitcast of the other.
199 class MergeFunctions {
200 public:
MergeFunctions()201 MergeFunctions() : FnTree(FunctionNodeCmp(&GlobalNumbers)) {
202 }
203
204 bool runOnModule(Module &M);
205
206 private:
207 // The function comparison operator is provided here so that FunctionNodes do
208 // not need to become larger with another pointer.
209 class FunctionNodeCmp {
210 GlobalNumberState* GlobalNumbers;
211
212 public:
FunctionNodeCmp(GlobalNumberState * GN)213 FunctionNodeCmp(GlobalNumberState* GN) : GlobalNumbers(GN) {}
214
operator ()(const FunctionNode & LHS,const FunctionNode & RHS) const215 bool operator()(const FunctionNode &LHS, const FunctionNode &RHS) const {
216 // Order first by hashes, then full function comparison.
217 if (LHS.getHash() != RHS.getHash())
218 return LHS.getHash() < RHS.getHash();
219 FunctionComparator FCmp(LHS.getFunc(), RHS.getFunc(), GlobalNumbers);
220 return FCmp.compare() == -1;
221 }
222 };
223 using FnTreeType = std::set<FunctionNode, FunctionNodeCmp>;
224
225 GlobalNumberState GlobalNumbers;
226
227 /// A work queue of functions that may have been modified and should be
228 /// analyzed again.
229 std::vector<WeakTrackingVH> Deferred;
230
231 #ifndef NDEBUG
232 /// Checks the rules of order relation introduced among functions set.
233 /// Returns true, if sanity check has been passed, and false if failed.
234 bool doSanityCheck(std::vector<WeakTrackingVH> &Worklist);
235 #endif
236
237 /// Insert a ComparableFunction into the FnTree, or merge it away if it's
238 /// equal to one that's already present.
239 bool insert(Function *NewFunction);
240
241 /// Remove a Function from the FnTree and queue it up for a second sweep of
242 /// analysis.
243 void remove(Function *F);
244
245 /// Find the functions that use this Value and remove them from FnTree and
246 /// queue the functions.
247 void removeUsers(Value *V);
248
249 /// Replace all direct calls of Old with calls of New. Will bitcast New if
250 /// necessary to make types match.
251 void replaceDirectCallers(Function *Old, Function *New);
252
253 /// Merge two equivalent functions. Upon completion, G may be deleted, or may
254 /// be converted into a thunk. In either case, it should never be visited
255 /// again.
256 void mergeTwoFunctions(Function *F, Function *G);
257
258 /// Fill PDIUnrelatedWL with instructions from the entry block that are
259 /// unrelated to parameter related debug info.
260 void filterInstsUnrelatedToPDI(BasicBlock *GEntryBlock,
261 std::vector<Instruction *> &PDIUnrelatedWL);
262
263 /// Erase the rest of the CFG (i.e. barring the entry block).
264 void eraseTail(Function *G);
265
266 /// Erase the instructions in PDIUnrelatedWL as they are unrelated to the
267 /// parameter debug info, from the entry block.
268 void eraseInstsUnrelatedToPDI(std::vector<Instruction *> &PDIUnrelatedWL);
269
270 /// Replace G with a simple tail call to bitcast(F). Also (unless
271 /// MergeFunctionsPDI holds) replace direct uses of G with bitcast(F),
272 /// delete G.
273 void writeThunk(Function *F, Function *G);
274
275 // Replace G with an alias to F (deleting function G)
276 void writeAlias(Function *F, Function *G);
277
278 // Replace G with an alias to F if possible, or a thunk to F if possible.
279 // Returns false if neither is the case.
280 bool writeThunkOrAlias(Function *F, Function *G);
281
282 /// Replace function F with function G in the function tree.
283 void replaceFunctionInTree(const FunctionNode &FN, Function *G);
284
285 /// The set of all distinct functions. Use the insert() and remove() methods
286 /// to modify it. The map allows efficient lookup and deferring of Functions.
287 FnTreeType FnTree;
288
289 // Map functions to the iterators of the FunctionNode which contains them
290 // in the FnTree. This must be updated carefully whenever the FnTree is
291 // modified, i.e. in insert(), remove(), and replaceFunctionInTree(), to avoid
292 // dangling iterators into FnTree. The invariant that preserves this is that
293 // there is exactly one mapping F -> FN for each FunctionNode FN in FnTree.
294 DenseMap<AssertingVH<Function>, FnTreeType::iterator> FNodesInTree;
295 };
296
297 class MergeFunctionsLegacyPass : public ModulePass {
298 public:
299 static char ID;
300
MergeFunctionsLegacyPass()301 MergeFunctionsLegacyPass(): ModulePass(ID) {
302 initializeMergeFunctionsLegacyPassPass(*PassRegistry::getPassRegistry());
303 }
304
runOnModule(Module & M)305 bool runOnModule(Module &M) override {
306 if (skipModule(M))
307 return false;
308
309 MergeFunctions MF;
310 return MF.runOnModule(M);
311 }
312 };
313
314 } // end anonymous namespace
315
316 char MergeFunctionsLegacyPass::ID = 0;
317 INITIALIZE_PASS(MergeFunctionsLegacyPass, "mergefunc",
318 "Merge Functions", false, false)
319
createMergeFunctionsPass()320 ModulePass *llvm::createMergeFunctionsPass() {
321 return new MergeFunctionsLegacyPass();
322 }
323
run(Module & M,ModuleAnalysisManager & AM)324 PreservedAnalyses MergeFunctionsPass::run(Module &M,
325 ModuleAnalysisManager &AM) {
326 MergeFunctions MF;
327 if (!MF.runOnModule(M))
328 return PreservedAnalyses::all();
329 return PreservedAnalyses::none();
330 }
331
332 #ifndef NDEBUG
doSanityCheck(std::vector<WeakTrackingVH> & Worklist)333 bool MergeFunctions::doSanityCheck(std::vector<WeakTrackingVH> &Worklist) {
334 if (const unsigned Max = NumFunctionsForSanityCheck) {
335 unsigned TripleNumber = 0;
336 bool Valid = true;
337
338 dbgs() << "MERGEFUNC-SANITY: Started for first " << Max << " functions.\n";
339
340 unsigned i = 0;
341 for (std::vector<WeakTrackingVH>::iterator I = Worklist.begin(),
342 E = Worklist.end();
343 I != E && i < Max; ++I, ++i) {
344 unsigned j = i;
345 for (std::vector<WeakTrackingVH>::iterator J = I; J != E && j < Max;
346 ++J, ++j) {
347 Function *F1 = cast<Function>(*I);
348 Function *F2 = cast<Function>(*J);
349 int Res1 = FunctionComparator(F1, F2, &GlobalNumbers).compare();
350 int Res2 = FunctionComparator(F2, F1, &GlobalNumbers).compare();
351
352 // If F1 <= F2, then F2 >= F1, otherwise report failure.
353 if (Res1 != -Res2) {
354 dbgs() << "MERGEFUNC-SANITY: Non-symmetric; triple: " << TripleNumber
355 << "\n";
356 dbgs() << *F1 << '\n' << *F2 << '\n';
357 Valid = false;
358 }
359
360 if (Res1 == 0)
361 continue;
362
363 unsigned k = j;
364 for (std::vector<WeakTrackingVH>::iterator K = J; K != E && k < Max;
365 ++k, ++K, ++TripleNumber) {
366 if (K == J)
367 continue;
368
369 Function *F3 = cast<Function>(*K);
370 int Res3 = FunctionComparator(F1, F3, &GlobalNumbers).compare();
371 int Res4 = FunctionComparator(F2, F3, &GlobalNumbers).compare();
372
373 bool Transitive = true;
374
375 if (Res1 != 0 && Res1 == Res4) {
376 // F1 > F2, F2 > F3 => F1 > F3
377 Transitive = Res3 == Res1;
378 } else if (Res3 != 0 && Res3 == -Res4) {
379 // F1 > F3, F3 > F2 => F1 > F2
380 Transitive = Res3 == Res1;
381 } else if (Res4 != 0 && -Res3 == Res4) {
382 // F2 > F3, F3 > F1 => F2 > F1
383 Transitive = Res4 == -Res1;
384 }
385
386 if (!Transitive) {
387 dbgs() << "MERGEFUNC-SANITY: Non-transitive; triple: "
388 << TripleNumber << "\n";
389 dbgs() << "Res1, Res3, Res4: " << Res1 << ", " << Res3 << ", "
390 << Res4 << "\n";
391 dbgs() << *F1 << '\n' << *F2 << '\n' << *F3 << '\n';
392 Valid = false;
393 }
394 }
395 }
396 }
397
398 dbgs() << "MERGEFUNC-SANITY: " << (Valid ? "Passed." : "Failed.") << "\n";
399 return Valid;
400 }
401 return true;
402 }
403 #endif
404
405 /// Check whether \p F is eligible for function merging.
isEligibleForMerging(Function & F)406 static bool isEligibleForMerging(Function &F) {
407 return !F.isDeclaration() && !F.hasAvailableExternallyLinkage();
408 }
409
runOnModule(Module & M)410 bool MergeFunctions::runOnModule(Module &M) {
411 bool Changed = false;
412
413 // All functions in the module, ordered by hash. Functions with a unique
414 // hash value are easily eliminated.
415 std::vector<std::pair<FunctionComparator::FunctionHash, Function *>>
416 HashedFuncs;
417 for (Function &Func : M) {
418 if (isEligibleForMerging(Func)) {
419 HashedFuncs.push_back({FunctionComparator::functionHash(Func), &Func});
420 }
421 }
422
423 llvm::stable_sort(HashedFuncs, less_first());
424
425 auto S = HashedFuncs.begin();
426 for (auto I = HashedFuncs.begin(), IE = HashedFuncs.end(); I != IE; ++I) {
427 // If the hash value matches the previous value or the next one, we must
428 // consider merging it. Otherwise it is dropped and never considered again.
429 if ((I != S && std::prev(I)->first == I->first) ||
430 (std::next(I) != IE && std::next(I)->first == I->first) ) {
431 Deferred.push_back(WeakTrackingVH(I->second));
432 }
433 }
434
435 do {
436 std::vector<WeakTrackingVH> Worklist;
437 Deferred.swap(Worklist);
438
439 LLVM_DEBUG(doSanityCheck(Worklist));
440
441 LLVM_DEBUG(dbgs() << "size of module: " << M.size() << '\n');
442 LLVM_DEBUG(dbgs() << "size of worklist: " << Worklist.size() << '\n');
443
444 // Insert functions and merge them.
445 for (WeakTrackingVH &I : Worklist) {
446 if (!I)
447 continue;
448 Function *F = cast<Function>(I);
449 if (!F->isDeclaration() && !F->hasAvailableExternallyLinkage()) {
450 Changed |= insert(F);
451 }
452 }
453 LLVM_DEBUG(dbgs() << "size of FnTree: " << FnTree.size() << '\n');
454 } while (!Deferred.empty());
455
456 FnTree.clear();
457 FNodesInTree.clear();
458 GlobalNumbers.clear();
459
460 return Changed;
461 }
462
463 // Replace direct callers of Old with New.
replaceDirectCallers(Function * Old,Function * New)464 void MergeFunctions::replaceDirectCallers(Function *Old, Function *New) {
465 Constant *BitcastNew = ConstantExpr::getBitCast(New, Old->getType());
466 for (auto UI = Old->use_begin(), UE = Old->use_end(); UI != UE;) {
467 Use *U = &*UI;
468 ++UI;
469 CallBase *CB = dyn_cast<CallBase>(U->getUser());
470 if (CB && CB->isCallee(U)) {
471 // Do not copy attributes from the called function to the call-site.
472 // Function comparison ensures that the attributes are the same up to
473 // type congruences in byval(), in which case we need to keep the byval
474 // type of the call-site, not the callee function.
475 remove(CB->getFunction());
476 U->set(BitcastNew);
477 }
478 }
479 }
480
481 // Helper for writeThunk,
482 // Selects proper bitcast operation,
483 // but a bit simpler then CastInst::getCastOpcode.
createCast(IRBuilder<> & Builder,Value * V,Type * DestTy)484 static Value *createCast(IRBuilder<> &Builder, Value *V, Type *DestTy) {
485 Type *SrcTy = V->getType();
486 if (SrcTy->isStructTy()) {
487 assert(DestTy->isStructTy());
488 assert(SrcTy->getStructNumElements() == DestTy->getStructNumElements());
489 Value *Result = UndefValue::get(DestTy);
490 for (unsigned int I = 0, E = SrcTy->getStructNumElements(); I < E; ++I) {
491 Value *Element = createCast(
492 Builder, Builder.CreateExtractValue(V, makeArrayRef(I)),
493 DestTy->getStructElementType(I));
494
495 Result =
496 Builder.CreateInsertValue(Result, Element, makeArrayRef(I));
497 }
498 return Result;
499 }
500 assert(!DestTy->isStructTy());
501 if (SrcTy->isIntegerTy() && DestTy->isPointerTy())
502 return Builder.CreateIntToPtr(V, DestTy);
503 else if (SrcTy->isPointerTy() && DestTy->isIntegerTy())
504 return Builder.CreatePtrToInt(V, DestTy);
505 else
506 return Builder.CreateBitCast(V, DestTy);
507 }
508
509 // Erase the instructions in PDIUnrelatedWL as they are unrelated to the
510 // parameter debug info, from the entry block.
eraseInstsUnrelatedToPDI(std::vector<Instruction * > & PDIUnrelatedWL)511 void MergeFunctions::eraseInstsUnrelatedToPDI(
512 std::vector<Instruction *> &PDIUnrelatedWL) {
513 LLVM_DEBUG(
514 dbgs() << " Erasing instructions (in reverse order of appearance in "
515 "entry block) unrelated to parameter debug info from entry "
516 "block: {\n");
517 while (!PDIUnrelatedWL.empty()) {
518 Instruction *I = PDIUnrelatedWL.back();
519 LLVM_DEBUG(dbgs() << " Deleting Instruction: ");
520 LLVM_DEBUG(I->print(dbgs()));
521 LLVM_DEBUG(dbgs() << "\n");
522 I->eraseFromParent();
523 PDIUnrelatedWL.pop_back();
524 }
525 LLVM_DEBUG(dbgs() << " } // Done erasing instructions unrelated to parameter "
526 "debug info from entry block. \n");
527 }
528
529 // Reduce G to its entry block.
eraseTail(Function * G)530 void MergeFunctions::eraseTail(Function *G) {
531 std::vector<BasicBlock *> WorklistBB;
532 for (BasicBlock &BB : drop_begin(*G)) {
533 BB.dropAllReferences();
534 WorklistBB.push_back(&BB);
535 }
536 while (!WorklistBB.empty()) {
537 BasicBlock *BB = WorklistBB.back();
538 BB->eraseFromParent();
539 WorklistBB.pop_back();
540 }
541 }
542
543 // We are interested in the following instructions from the entry block as being
544 // related to parameter debug info:
545 // - @llvm.dbg.declare
546 // - stores from the incoming parameters to locations on the stack-frame
547 // - allocas that create these locations on the stack-frame
548 // - @llvm.dbg.value
549 // - the entry block's terminator
550 // The rest are unrelated to debug info for the parameters; fill up
551 // PDIUnrelatedWL with such instructions.
filterInstsUnrelatedToPDI(BasicBlock * GEntryBlock,std::vector<Instruction * > & PDIUnrelatedWL)552 void MergeFunctions::filterInstsUnrelatedToPDI(
553 BasicBlock *GEntryBlock, std::vector<Instruction *> &PDIUnrelatedWL) {
554 std::set<Instruction *> PDIRelated;
555 for (BasicBlock::iterator BI = GEntryBlock->begin(), BIE = GEntryBlock->end();
556 BI != BIE; ++BI) {
557 if (auto *DVI = dyn_cast<DbgValueInst>(&*BI)) {
558 LLVM_DEBUG(dbgs() << " Deciding: ");
559 LLVM_DEBUG(BI->print(dbgs()));
560 LLVM_DEBUG(dbgs() << "\n");
561 DILocalVariable *DILocVar = DVI->getVariable();
562 if (DILocVar->isParameter()) {
563 LLVM_DEBUG(dbgs() << " Include (parameter): ");
564 LLVM_DEBUG(BI->print(dbgs()));
565 LLVM_DEBUG(dbgs() << "\n");
566 PDIRelated.insert(&*BI);
567 } else {
568 LLVM_DEBUG(dbgs() << " Delete (!parameter): ");
569 LLVM_DEBUG(BI->print(dbgs()));
570 LLVM_DEBUG(dbgs() << "\n");
571 }
572 } else if (auto *DDI = dyn_cast<DbgDeclareInst>(&*BI)) {
573 LLVM_DEBUG(dbgs() << " Deciding: ");
574 LLVM_DEBUG(BI->print(dbgs()));
575 LLVM_DEBUG(dbgs() << "\n");
576 DILocalVariable *DILocVar = DDI->getVariable();
577 if (DILocVar->isParameter()) {
578 LLVM_DEBUG(dbgs() << " Parameter: ");
579 LLVM_DEBUG(DILocVar->print(dbgs()));
580 AllocaInst *AI = dyn_cast_or_null<AllocaInst>(DDI->getAddress());
581 if (AI) {
582 LLVM_DEBUG(dbgs() << " Processing alloca users: ");
583 LLVM_DEBUG(dbgs() << "\n");
584 for (User *U : AI->users()) {
585 if (StoreInst *SI = dyn_cast<StoreInst>(U)) {
586 if (Value *Arg = SI->getValueOperand()) {
587 if (isa<Argument>(Arg)) {
588 LLVM_DEBUG(dbgs() << " Include: ");
589 LLVM_DEBUG(AI->print(dbgs()));
590 LLVM_DEBUG(dbgs() << "\n");
591 PDIRelated.insert(AI);
592 LLVM_DEBUG(dbgs() << " Include (parameter): ");
593 LLVM_DEBUG(SI->print(dbgs()));
594 LLVM_DEBUG(dbgs() << "\n");
595 PDIRelated.insert(SI);
596 LLVM_DEBUG(dbgs() << " Include: ");
597 LLVM_DEBUG(BI->print(dbgs()));
598 LLVM_DEBUG(dbgs() << "\n");
599 PDIRelated.insert(&*BI);
600 } else {
601 LLVM_DEBUG(dbgs() << " Delete (!parameter): ");
602 LLVM_DEBUG(SI->print(dbgs()));
603 LLVM_DEBUG(dbgs() << "\n");
604 }
605 }
606 } else {
607 LLVM_DEBUG(dbgs() << " Defer: ");
608 LLVM_DEBUG(U->print(dbgs()));
609 LLVM_DEBUG(dbgs() << "\n");
610 }
611 }
612 } else {
613 LLVM_DEBUG(dbgs() << " Delete (alloca NULL): ");
614 LLVM_DEBUG(BI->print(dbgs()));
615 LLVM_DEBUG(dbgs() << "\n");
616 }
617 } else {
618 LLVM_DEBUG(dbgs() << " Delete (!parameter): ");
619 LLVM_DEBUG(BI->print(dbgs()));
620 LLVM_DEBUG(dbgs() << "\n");
621 }
622 } else if (BI->isTerminator() && &*BI == GEntryBlock->getTerminator()) {
623 LLVM_DEBUG(dbgs() << " Will Include Terminator: ");
624 LLVM_DEBUG(BI->print(dbgs()));
625 LLVM_DEBUG(dbgs() << "\n");
626 PDIRelated.insert(&*BI);
627 } else {
628 LLVM_DEBUG(dbgs() << " Defer: ");
629 LLVM_DEBUG(BI->print(dbgs()));
630 LLVM_DEBUG(dbgs() << "\n");
631 }
632 }
633 LLVM_DEBUG(
634 dbgs()
635 << " Report parameter debug info related/related instructions: {\n");
636 for (Instruction &I : *GEntryBlock) {
637 if (PDIRelated.find(&I) == PDIRelated.end()) {
638 LLVM_DEBUG(dbgs() << " !PDIRelated: ");
639 LLVM_DEBUG(I.print(dbgs()));
640 LLVM_DEBUG(dbgs() << "\n");
641 PDIUnrelatedWL.push_back(&I);
642 } else {
643 LLVM_DEBUG(dbgs() << " PDIRelated: ");
644 LLVM_DEBUG(I.print(dbgs()));
645 LLVM_DEBUG(dbgs() << "\n");
646 }
647 }
648 LLVM_DEBUG(dbgs() << " }\n");
649 }
650
651 /// Whether this function may be replaced by a forwarding thunk.
canCreateThunkFor(Function * F)652 static bool canCreateThunkFor(Function *F) {
653 if (F->isVarArg())
654 return false;
655
656 // Don't merge tiny functions using a thunk, since it can just end up
657 // making the function larger.
658 if (F->size() == 1) {
659 if (F->front().size() <= 2) {
660 LLVM_DEBUG(dbgs() << "canCreateThunkFor: " << F->getName()
661 << " is too small to bother creating a thunk for\n");
662 return false;
663 }
664 }
665 return true;
666 }
667
668 // Replace G with a simple tail call to bitcast(F). Also (unless
669 // MergeFunctionsPDI holds) replace direct uses of G with bitcast(F),
670 // delete G. Under MergeFunctionsPDI, we use G itself for creating
671 // the thunk as we preserve the debug info (and associated instructions)
672 // from G's entry block pertaining to G's incoming arguments which are
673 // passed on as corresponding arguments in the call that G makes to F.
674 // For better debugability, under MergeFunctionsPDI, we do not modify G's
675 // call sites to point to F even when within the same translation unit.
writeThunk(Function * F,Function * G)676 void MergeFunctions::writeThunk(Function *F, Function *G) {
677 BasicBlock *GEntryBlock = nullptr;
678 std::vector<Instruction *> PDIUnrelatedWL;
679 BasicBlock *BB = nullptr;
680 Function *NewG = nullptr;
681 if (MergeFunctionsPDI) {
682 LLVM_DEBUG(dbgs() << "writeThunk: (MergeFunctionsPDI) Do not create a new "
683 "function as thunk; retain original: "
684 << G->getName() << "()\n");
685 GEntryBlock = &G->getEntryBlock();
686 LLVM_DEBUG(
687 dbgs() << "writeThunk: (MergeFunctionsPDI) filter parameter related "
688 "debug info for "
689 << G->getName() << "() {\n");
690 filterInstsUnrelatedToPDI(GEntryBlock, PDIUnrelatedWL);
691 GEntryBlock->getTerminator()->eraseFromParent();
692 BB = GEntryBlock;
693 } else {
694 NewG = Function::Create(G->getFunctionType(), G->getLinkage(),
695 G->getAddressSpace(), "", G->getParent());
696 NewG->setComdat(G->getComdat());
697 BB = BasicBlock::Create(F->getContext(), "", NewG);
698 }
699
700 IRBuilder<> Builder(BB);
701 Function *H = MergeFunctionsPDI ? G : NewG;
702 SmallVector<Value *, 16> Args;
703 unsigned i = 0;
704 FunctionType *FFTy = F->getFunctionType();
705 for (Argument &AI : H->args()) {
706 Args.push_back(createCast(Builder, &AI, FFTy->getParamType(i)));
707 ++i;
708 }
709
710 CallInst *CI = Builder.CreateCall(F, Args);
711 ReturnInst *RI = nullptr;
712 CI->setTailCall();
713 CI->setCallingConv(F->getCallingConv());
714 CI->setAttributes(F->getAttributes());
715 if (H->getReturnType()->isVoidTy()) {
716 RI = Builder.CreateRetVoid();
717 } else {
718 RI = Builder.CreateRet(createCast(Builder, CI, H->getReturnType()));
719 }
720
721 if (MergeFunctionsPDI) {
722 DISubprogram *DIS = G->getSubprogram();
723 if (DIS) {
724 DebugLoc CIDbgLoc =
725 DILocation::get(DIS->getContext(), DIS->getScopeLine(), 0, DIS);
726 DebugLoc RIDbgLoc =
727 DILocation::get(DIS->getContext(), DIS->getScopeLine(), 0, DIS);
728 CI->setDebugLoc(CIDbgLoc);
729 RI->setDebugLoc(RIDbgLoc);
730 } else {
731 LLVM_DEBUG(
732 dbgs() << "writeThunk: (MergeFunctionsPDI) No DISubprogram for "
733 << G->getName() << "()\n");
734 }
735 eraseTail(G);
736 eraseInstsUnrelatedToPDI(PDIUnrelatedWL);
737 LLVM_DEBUG(
738 dbgs() << "} // End of parameter related debug info filtering for: "
739 << G->getName() << "()\n");
740 } else {
741 NewG->copyAttributesFrom(G);
742 NewG->takeName(G);
743 removeUsers(G);
744 G->replaceAllUsesWith(NewG);
745 G->eraseFromParent();
746 }
747
748 LLVM_DEBUG(dbgs() << "writeThunk: " << H->getName() << '\n');
749 ++NumThunksWritten;
750 }
751
752 // Whether this function may be replaced by an alias
canCreateAliasFor(Function * F)753 static bool canCreateAliasFor(Function *F) {
754 if (!MergeFunctionsAliases || !F->hasGlobalUnnamedAddr())
755 return false;
756
757 // We should only see linkages supported by aliases here
758 assert(F->hasLocalLinkage() || F->hasExternalLinkage()
759 || F->hasWeakLinkage() || F->hasLinkOnceLinkage());
760 return true;
761 }
762
763 // Replace G with an alias to F (deleting function G)
writeAlias(Function * F,Function * G)764 void MergeFunctions::writeAlias(Function *F, Function *G) {
765 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
766 PointerType *PtrType = G->getType();
767 auto *GA = GlobalAlias::create(
768 PtrType->getElementType(), PtrType->getAddressSpace(),
769 G->getLinkage(), "", BitcastF, G->getParent());
770
771 F->setAlignment(MaybeAlign(std::max(F->getAlignment(), G->getAlignment())));
772 GA->takeName(G);
773 GA->setVisibility(G->getVisibility());
774 GA->setUnnamedAddr(GlobalValue::UnnamedAddr::Global);
775
776 removeUsers(G);
777 G->replaceAllUsesWith(GA);
778 G->eraseFromParent();
779
780 LLVM_DEBUG(dbgs() << "writeAlias: " << GA->getName() << '\n');
781 ++NumAliasesWritten;
782 }
783
784 // Replace G with an alias to F if possible, or a thunk to F if
785 // profitable. Returns false if neither is the case.
writeThunkOrAlias(Function * F,Function * G)786 bool MergeFunctions::writeThunkOrAlias(Function *F, Function *G) {
787 if (canCreateAliasFor(G)) {
788 writeAlias(F, G);
789 return true;
790 }
791 if (canCreateThunkFor(F)) {
792 writeThunk(F, G);
793 return true;
794 }
795 return false;
796 }
797
798 // Merge two equivalent functions. Upon completion, Function G is deleted.
mergeTwoFunctions(Function * F,Function * G)799 void MergeFunctions::mergeTwoFunctions(Function *F, Function *G) {
800 if (F->isInterposable()) {
801 assert(G->isInterposable());
802
803 // Both writeThunkOrAlias() calls below must succeed, either because we can
804 // create aliases for G and NewF, or because a thunk for F is profitable.
805 // F here has the same signature as NewF below, so that's what we check.
806 if (!canCreateThunkFor(F) &&
807 (!canCreateAliasFor(F) || !canCreateAliasFor(G)))
808 return;
809
810 // Make them both thunks to the same internal function.
811 Function *NewF = Function::Create(F->getFunctionType(), F->getLinkage(),
812 F->getAddressSpace(), "", F->getParent());
813 NewF->copyAttributesFrom(F);
814 NewF->takeName(F);
815 removeUsers(F);
816 F->replaceAllUsesWith(NewF);
817
818 MaybeAlign MaxAlignment(std::max(G->getAlignment(), NewF->getAlignment()));
819
820 writeThunkOrAlias(F, G);
821 writeThunkOrAlias(F, NewF);
822
823 F->setAlignment(MaxAlignment);
824 F->setLinkage(GlobalValue::PrivateLinkage);
825 ++NumDoubleWeak;
826 ++NumFunctionsMerged;
827 } else {
828 // For better debugability, under MergeFunctionsPDI, we do not modify G's
829 // call sites to point to F even when within the same translation unit.
830 if (!G->isInterposable() && !MergeFunctionsPDI) {
831 if (G->hasGlobalUnnamedAddr()) {
832 // G might have been a key in our GlobalNumberState, and it's illegal
833 // to replace a key in ValueMap<GlobalValue *> with a non-global.
834 GlobalNumbers.erase(G);
835 // If G's address is not significant, replace it entirely.
836 Constant *BitcastF = ConstantExpr::getBitCast(F, G->getType());
837 removeUsers(G);
838 G->replaceAllUsesWith(BitcastF);
839 } else {
840 // Redirect direct callers of G to F. (See note on MergeFunctionsPDI
841 // above).
842 replaceDirectCallers(G, F);
843 }
844 }
845
846 // If G was internal then we may have replaced all uses of G with F. If so,
847 // stop here and delete G. There's no need for a thunk. (See note on
848 // MergeFunctionsPDI above).
849 if (G->isDiscardableIfUnused() && G->use_empty() && !MergeFunctionsPDI) {
850 G->eraseFromParent();
851 ++NumFunctionsMerged;
852 return;
853 }
854
855 if (writeThunkOrAlias(F, G)) {
856 ++NumFunctionsMerged;
857 }
858 }
859 }
860
861 /// Replace function F by function G.
replaceFunctionInTree(const FunctionNode & FN,Function * G)862 void MergeFunctions::replaceFunctionInTree(const FunctionNode &FN,
863 Function *G) {
864 Function *F = FN.getFunc();
865 assert(FunctionComparator(F, G, &GlobalNumbers).compare() == 0 &&
866 "The two functions must be equal");
867
868 auto I = FNodesInTree.find(F);
869 assert(I != FNodesInTree.end() && "F should be in FNodesInTree");
870 assert(FNodesInTree.count(G) == 0 && "FNodesInTree should not contain G");
871
872 FnTreeType::iterator IterToFNInFnTree = I->second;
873 assert(&(*IterToFNInFnTree) == &FN && "F should map to FN in FNodesInTree.");
874 // Remove F -> FN and insert G -> FN
875 FNodesInTree.erase(I);
876 FNodesInTree.insert({G, IterToFNInFnTree});
877 // Replace F with G in FN, which is stored inside the FnTree.
878 FN.replaceBy(G);
879 }
880
881 // Ordering for functions that are equal under FunctionComparator
isFuncOrderCorrect(const Function * F,const Function * G)882 static bool isFuncOrderCorrect(const Function *F, const Function *G) {
883 if (F->isInterposable() != G->isInterposable()) {
884 // Strong before weak, because the weak function may call the strong
885 // one, but not the other way around.
886 return !F->isInterposable();
887 }
888 if (F->hasLocalLinkage() != G->hasLocalLinkage()) {
889 // External before local, because we definitely have to keep the external
890 // function, but may be able to drop the local one.
891 return !F->hasLocalLinkage();
892 }
893 // Impose a total order (by name) on the replacement of functions. This is
894 // important when operating on more than one module independently to prevent
895 // cycles of thunks calling each other when the modules are linked together.
896 return F->getName() <= G->getName();
897 }
898
899 // Insert a ComparableFunction into the FnTree, or merge it away if equal to one
900 // that was already inserted.
insert(Function * NewFunction)901 bool MergeFunctions::insert(Function *NewFunction) {
902 std::pair<FnTreeType::iterator, bool> Result =
903 FnTree.insert(FunctionNode(NewFunction));
904
905 if (Result.second) {
906 assert(FNodesInTree.count(NewFunction) == 0);
907 FNodesInTree.insert({NewFunction, Result.first});
908 LLVM_DEBUG(dbgs() << "Inserting as unique: " << NewFunction->getName()
909 << '\n');
910 return false;
911 }
912
913 const FunctionNode &OldF = *Result.first;
914
915 if (!isFuncOrderCorrect(OldF.getFunc(), NewFunction)) {
916 // Swap the two functions.
917 Function *F = OldF.getFunc();
918 replaceFunctionInTree(*Result.first, NewFunction);
919 NewFunction = F;
920 assert(OldF.getFunc() != F && "Must have swapped the functions.");
921 }
922
923 LLVM_DEBUG(dbgs() << " " << OldF.getFunc()->getName()
924 << " == " << NewFunction->getName() << '\n');
925
926 Function *DeleteF = NewFunction;
927 mergeTwoFunctions(OldF.getFunc(), DeleteF);
928 return true;
929 }
930
931 // Remove a function from FnTree. If it was already in FnTree, add
932 // it to Deferred so that we'll look at it in the next round.
remove(Function * F)933 void MergeFunctions::remove(Function *F) {
934 auto I = FNodesInTree.find(F);
935 if (I != FNodesInTree.end()) {
936 LLVM_DEBUG(dbgs() << "Deferred " << F->getName() << ".\n");
937 FnTree.erase(I->second);
938 // I->second has been invalidated, remove it from the FNodesInTree map to
939 // preserve the invariant.
940 FNodesInTree.erase(I);
941 Deferred.emplace_back(F);
942 }
943 }
944
945 // For each instruction used by the value, remove() the function that contains
946 // the instruction. This should happen right before a call to RAUW.
removeUsers(Value * V)947 void MergeFunctions::removeUsers(Value *V) {
948 for (User *U : V->users())
949 if (auto *I = dyn_cast<Instruction>(U))
950 remove(I->getFunction());
951 }
952