1 //===-- GlobalMerge.cpp - Internal globals merging -----------------------===//
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 //===----------------------------------------------------------------------===//
9 // This pass merges globals with internal linkage into one. This way all the
10 // globals which were merged into a biggest one can be addressed using offsets
11 // from the same base pointer (no need for separate base pointer for each of the
12 // global). Such a transformation can significantly reduce the register pressure
13 // when many globals are involved.
15 // For example, consider the code which touches several global variables at
18 // static int foo[N], bar[N], baz[N];
20 // for (i = 0; i < N; ++i) {
21 // foo[i] = bar[i] * baz[i];
24 // On ARM the addresses of 3 arrays should be kept in the registers, thus
25 // this code has quite large register pressure (loop body):
32 // Pass converts the code to something like:
40 // for (i = 0; i < N; ++i) {
41 // merged.foo[i] = merged.bar[i] * merged.baz[i];
44 // and in ARM code this becomes:
51 // note that we saved 2 registers here almostly "for free".
53 // However, merging globals can have tradeoffs:
54 // - it confuses debuggers, tools, and users
55 // - it makes linker optimizations less useful (order files, LOHs, ...)
56 // - it forces usage of indexed addressing (which isn't necessarily "free")
57 // - it can increase register pressure when the uses are disparate enough.
59 // We use heuristics to discover the best global grouping we can (cf cl::opts).
60 // ===---------------------------------------------------------------------===//
62 #include "llvm/Transforms/Scalar.h"
63 #include "llvm/ADT/DenseMap.h"
64 #include "llvm/ADT/SmallBitVector.h"
65 #include "llvm/ADT/SmallPtrSet.h"
66 #include "llvm/ADT/Statistic.h"
67 #include "llvm/CodeGen/Passes.h"
68 #include "llvm/IR/Attributes.h"
69 #include "llvm/IR/Constants.h"
70 #include "llvm/IR/DataLayout.h"
71 #include "llvm/IR/DerivedTypes.h"
72 #include "llvm/IR/Function.h"
73 #include "llvm/IR/GlobalVariable.h"
74 #include "llvm/IR/Instructions.h"
75 #include "llvm/IR/Intrinsics.h"
76 #include "llvm/IR/Module.h"
77 #include "llvm/Pass.h"
78 #include "llvm/Support/CommandLine.h"
79 #include "llvm/Support/Debug.h"
80 #include "llvm/Support/raw_ostream.h"
81 #include "llvm/Target/TargetLowering.h"
82 #include "llvm/Target/TargetLoweringObjectFile.h"
83 #include "llvm/Target/TargetSubtargetInfo.h"
87 #define DEBUG_TYPE "global-merge"
89 // FIXME: This is only useful as a last-resort way to disable the pass.
91 EnableGlobalMerge("enable-global-merge", cl::Hidden,
92 cl::desc("Enable the global merge pass"),
95 static cl::opt<bool> GlobalMergeGroupByUse(
96 "global-merge-group-by-use", cl::Hidden,
97 cl::desc("Improve global merge pass to look at uses"), cl::init(true));
99 static cl::opt<bool> GlobalMergeIgnoreSingleUse(
100 "global-merge-ignore-single-use", cl::Hidden,
101 cl::desc("Improve global merge pass to ignore globals only used alone"),
105 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
106 cl::desc("Enable global merge pass on constants"),
109 // FIXME: this could be a transitional option, and we probably need to remove
110 // it if only we are sure this optimization could always benefit all targets.
112 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
113 cl::desc("Enable global merge pass on external linkage"),
116 STATISTIC(NumMerged, "Number of globals merged");
118 class GlobalMerge : public FunctionPass {
119 const TargetMachine *TM;
120 const DataLayout *DL;
121 // FIXME: Infer the maximum possible offset depending on the actual users
122 // (these max offsets are different for the users inside Thumb or ARM
123 // functions), see the code that passes in the offset in the ARM backend
124 // for more information.
127 /// Whether we should try to optimize for size only.
128 /// Currently, this applies a dead simple heuristic: only consider globals
129 /// used in minsize functions for merging.
130 /// FIXME: This could learn about optsize, and be used in the cost model.
131 bool OnlyOptimizeForSize;
133 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
134 Module &M, bool isConst, unsigned AddrSpace) const;
135 /// \brief Merge everything in \p Globals for which the corresponding bit
136 /// in \p GlobalSet is set.
137 bool doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
138 const BitVector &GlobalSet, Module &M, bool isConst,
139 unsigned AddrSpace) const;
141 /// \brief Check if the given variable has been identified as must keep
142 /// \pre setMustKeepGlobalVariables must have been called on the Module that
144 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
145 return MustKeepGlobalVariables.count(GV);
148 /// Collect every variables marked as "used" or used in a landing pad
149 /// instruction for this Module.
150 void setMustKeepGlobalVariables(Module &M);
152 /// Collect every variables marked as "used"
153 void collectUsedGlobalVariables(Module &M);
155 /// Keep track of the GlobalVariable that must not be merged away
156 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
159 static char ID; // Pass identification, replacement for typeid.
160 explicit GlobalMerge(const TargetMachine *TM = nullptr,
161 unsigned MaximalOffset = 0,
162 bool OnlyOptimizeForSize = false)
163 : FunctionPass(ID), TM(TM), DL(TM->getDataLayout()),
164 MaxOffset(MaximalOffset), OnlyOptimizeForSize(OnlyOptimizeForSize) {
165 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
168 bool doInitialization(Module &M) override;
169 bool runOnFunction(Function &F) override;
170 bool doFinalization(Module &M) override;
172 const char *getPassName() const override {
173 return "Merge internal globals";
176 void getAnalysisUsage(AnalysisUsage &AU) const override {
177 AU.setPreservesCFG();
178 FunctionPass::getAnalysisUsage(AU);
181 } // end anonymous namespace
183 char GlobalMerge::ID = 0;
184 INITIALIZE_PASS_BEGIN(GlobalMerge, "global-merge", "Merge global variables",
186 INITIALIZE_PASS_END(GlobalMerge, "global-merge", "Merge global variables",
189 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
190 Module &M, bool isConst, unsigned AddrSpace) const {
191 // FIXME: Find better heuristics
192 std::stable_sort(Globals.begin(), Globals.end(),
193 [this](const GlobalVariable *GV1, const GlobalVariable *GV2) {
194 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
195 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
197 return (DL->getTypeAllocSize(Ty1) < DL->getTypeAllocSize(Ty2));
200 // If we want to just blindly group all globals together, do so.
201 if (!GlobalMergeGroupByUse) {
202 BitVector AllGlobals(Globals.size());
204 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
207 // If we want to be smarter, look at all uses of each global, to try to
208 // discover all sets of globals used together, and how many times each of
209 // these sets occured.
211 // Keep this reasonably efficient, by having an append-only list of all sets
212 // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
213 // code (currently, a Function) to the set of globals seen so far that are
214 // used together in that unit (GlobalUsesByFunction).
216 // When we look at the Nth global, we now that any new set is either:
217 // - the singleton set {N}, containing this global only, or
218 // - the union of {N} and a previously-discovered set, containing some
219 // combination of the previous N-1 globals.
220 // Using that knowledge, when looking at the Nth global, we can keep:
221 // - a reference to the singleton set {N} (CurGVOnlySetIdx)
222 // - a list mapping each previous set to its union with {N} (EncounteredUGS),
223 // if it actually occurs.
225 // We keep track of the sets of globals used together "close enough".
226 struct UsedGlobalSet {
227 UsedGlobalSet(size_t Size) : Globals(Size), UsageCount(1) {}
232 // Each set is unique in UsedGlobalSets.
233 std::vector<UsedGlobalSet> UsedGlobalSets;
235 // Avoid repeating the create-global-set pattern.
236 auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
237 UsedGlobalSets.emplace_back(Globals.size());
238 return UsedGlobalSets.back();
241 // The first set is the empty set.
242 CreateGlobalSet().UsageCount = 0;
244 // We define "close enough" to be "in the same function".
245 // FIXME: Grouping uses by function is way too aggressive, so we should have
246 // a better metric for distance between uses.
247 // The obvious alternative would be to group by BasicBlock, but that's in
248 // turn too conservative..
249 // Anything in between wouldn't be trivial to compute, so just stick with
250 // per-function grouping.
252 // The value type is an index into UsedGlobalSets.
253 // The default (0) conveniently points to the empty set.
254 DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
256 // Now, look at each merge-eligible global in turn.
258 // Keep track of the sets we already encountered to which we added the
260 // Each element matches the same-index element in UsedGlobalSets.
261 // This lets us efficiently tell whether a set has already been expanded to
262 // include the current global.
263 std::vector<size_t> EncounteredUGS;
265 for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
266 GlobalVariable *GV = Globals[GI];
268 // Reset the encountered sets for this global...
269 std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
270 // ...and grow it in case we created new sets for the previous global.
271 EncounteredUGS.resize(UsedGlobalSets.size());
273 // We might need to create a set that only consists of the current global.
274 // Keep track of its index into UsedGlobalSets.
275 size_t CurGVOnlySetIdx = 0;
277 // For each global, look at all its Uses.
278 for (auto &U : GV->uses()) {
279 // This Use might be a ConstantExpr. We're interested in Instruction
280 // users, so look through ConstantExpr...
282 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
285 UI = &*CE->use_begin();
287 } else if (isa<Instruction>(U.getUser())) {
294 // ...to iterate on all the instruction users of the global.
295 // Note that we iterate on Uses and not on Users to be able to getNext().
296 for (; UI != UE; UI = UI->getNext()) {
297 Instruction *I = dyn_cast<Instruction>(UI->getUser());
301 Function *ParentFn = I->getParent()->getParent();
303 // If we're only optimizing for size, ignore non-minsize functions.
304 if (OnlyOptimizeForSize &&
305 !ParentFn->hasFnAttribute(Attribute::MinSize))
308 size_t UGSIdx = GlobalUsesByFunction[ParentFn];
310 // If this is the first global the basic block uses, map it to the set
311 // consisting of this global only.
313 // If that set doesn't exist yet, create it.
314 if (!CurGVOnlySetIdx) {
315 CurGVOnlySetIdx = UsedGlobalSets.size();
316 CreateGlobalSet().Globals.set(GI);
318 ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
321 GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
325 // If we already encountered this BB, just increment the counter.
326 if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
327 ++UsedGlobalSets[UGSIdx].UsageCount;
331 // If not, the previous set wasn't actually used in this function.
332 --UsedGlobalSets[UGSIdx].UsageCount;
334 // If we already expanded the previous set to include this global, just
335 // reuse that expanded set.
336 if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
337 ++UsedGlobalSets[ExpandedIdx].UsageCount;
338 GlobalUsesByFunction[ParentFn] = ExpandedIdx;
342 // If not, create a new set consisting of the union of the previous set
343 // and this global. Mark it as encountered, so we can reuse it later.
344 GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
345 UsedGlobalSets.size();
347 UsedGlobalSet &NewUGS = CreateGlobalSet();
348 NewUGS.Globals.set(GI);
349 NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
354 // Now we found a bunch of sets of globals used together. We accumulated
355 // the number of times we encountered the sets (i.e., the number of blocks
356 // that use that exact set of globals).
358 // Multiply that by the size of the set to give us a crude profitability
360 std::sort(UsedGlobalSets.begin(), UsedGlobalSets.end(),
361 [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
362 return UGS1.Globals.count() * UGS1.UsageCount <
363 UGS2.Globals.count() * UGS2.UsageCount;
366 // We can choose to merge all globals together, but ignore globals never used
367 // with another global. This catches the obviously non-profitable cases of
368 // having a single global, but is aggressive enough for any other case.
369 if (GlobalMergeIgnoreSingleUse) {
370 BitVector AllGlobals(Globals.size());
371 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
372 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
373 if (UGS.UsageCount == 0)
375 if (UGS.Globals.count() > 1)
376 AllGlobals |= UGS.Globals;
378 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
381 // Starting from the sets with the best (=biggest) profitability, find a
383 // The ideal (and expensive) solution can only be found by trying all
384 // combinations, looking for the one with the best profitability.
385 // Don't be smart about it, and just pick the first compatible combination,
386 // starting with the sets with the best profitability.
387 BitVector PickedGlobals(Globals.size());
388 bool Changed = false;
390 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
391 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
392 if (UGS.UsageCount == 0)
394 if (PickedGlobals.anyCommon(UGS.Globals))
396 PickedGlobals |= UGS.Globals;
397 // If the set only contains one global, there's no point in merging.
398 // Ignore the global for inclusion in other sets though, so keep it in
400 if (UGS.Globals.count() < 2)
402 Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
408 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
409 const BitVector &GlobalSet, Module &M, bool isConst,
410 unsigned AddrSpace) const {
412 Type *Int32Ty = Type::getInt32Ty(M.getContext());
414 assert(Globals.size() > 1);
416 DEBUG(dbgs() << " Trying to merge set, starts with #"
417 << GlobalSet.find_first() << "\n");
419 ssize_t i = GlobalSet.find_first();
422 uint64_t MergedSize = 0;
423 std::vector<Type*> Tys;
424 std::vector<Constant*> Inits;
426 bool HasExternal = false;
427 GlobalVariable *TheFirstExternal = 0;
428 for (j = i; j != -1; j = GlobalSet.find_next(j)) {
429 Type *Ty = Globals[j]->getType()->getElementType();
430 MergedSize += DL->getTypeAllocSize(Ty);
431 if (MergedSize > MaxOffset) {
435 Inits.push_back(Globals[j]->getInitializer());
437 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
439 TheFirstExternal = Globals[j];
443 // If merged variables doesn't have external linkage, we needn't to expose
444 // the symbol after merging.
445 GlobalValue::LinkageTypes Linkage = HasExternal
446 ? GlobalValue::ExternalLinkage
447 : GlobalValue::InternalLinkage;
449 StructType *MergedTy = StructType::get(M.getContext(), Tys);
450 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
452 // If merged variables have external linkage, we use symbol name of the
453 // first variable merged as the suffix of global symbol name. This would
454 // be able to avoid the link-time naming conflict for globalm symbols.
455 GlobalVariable *MergedGV = new GlobalVariable(
456 M, MergedTy, isConst, Linkage, MergedInit,
457 HasExternal ? "_MergedGlobals_" + TheFirstExternal->getName()
459 nullptr, GlobalVariable::NotThreadLocal, AddrSpace);
461 for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k)) {
462 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
463 std::string Name = Globals[k]->getName();
466 ConstantInt::get(Int32Ty, 0),
467 ConstantInt::get(Int32Ty, idx++)
470 ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
471 Globals[k]->replaceAllUsesWith(GEP);
472 Globals[k]->eraseFromParent();
474 if (Linkage != GlobalValue::InternalLinkage) {
475 // Generate a new alias...
476 auto *PTy = cast<PointerType>(GEP->getType());
477 GlobalAlias::create(PTy, Linkage, Name, GEP, &M);
488 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
489 // Extract global variables from llvm.used array
490 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
491 if (!GV || !GV->hasInitializer()) return;
493 // Should be an array of 'i8*'.
494 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
496 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
497 if (const GlobalVariable *G =
498 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
499 MustKeepGlobalVariables.insert(G);
502 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
503 collectUsedGlobalVariables(M);
505 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
507 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
508 IBB != IEndBB; ++IBB) {
509 // Follow the invoke link to find the landing pad instruction
510 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
513 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
514 // Look for globals in the clauses of the landing pad instruction
515 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
516 Idx != NumClauses; ++Idx)
517 if (const GlobalVariable *GV =
518 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
519 ->stripPointerCasts()))
520 MustKeepGlobalVariables.insert(GV);
525 bool GlobalMerge::doInitialization(Module &M) {
526 if (!EnableGlobalMerge)
529 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
531 bool Changed = false;
532 setMustKeepGlobalVariables(M);
534 // Grab all non-const globals.
535 for (Module::global_iterator I = M.global_begin(),
536 E = M.global_end(); I != E; ++I) {
537 // Merge is safe for "normal" internal or external globals only
538 if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
541 if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage()) &&
542 !I->hasInternalLinkage())
545 PointerType *PT = dyn_cast<PointerType>(I->getType());
546 assert(PT && "Global variable is not a pointer!");
548 unsigned AddressSpace = PT->getAddressSpace();
550 // Ignore fancy-aligned globals for now.
551 unsigned Alignment = DL->getPreferredAlignment(I);
552 Type *Ty = I->getType()->getElementType();
553 if (Alignment > DL->getABITypeAlignment(Ty))
556 // Ignore all 'special' globals.
557 if (I->getName().startswith("llvm.") ||
558 I->getName().startswith(".llvm."))
561 // Ignore all "required" globals:
562 if (isMustKeepGlobalVariable(I))
565 if (DL->getTypeAllocSize(Ty) < MaxOffset) {
566 if (TargetLoweringObjectFile::getKindForGlobal(I, *TM).isBSSLocal())
567 BSSGlobals[AddressSpace].push_back(I);
568 else if (I->isConstant())
569 ConstGlobals[AddressSpace].push_back(I);
571 Globals[AddressSpace].push_back(I);
575 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
576 I = Globals.begin(), E = Globals.end(); I != E; ++I)
577 if (I->second.size() > 1)
578 Changed |= doMerge(I->second, M, false, I->first);
580 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
581 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
582 if (I->second.size() > 1)
583 Changed |= doMerge(I->second, M, false, I->first);
585 if (EnableGlobalMergeOnConst)
586 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
587 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
588 if (I->second.size() > 1)
589 Changed |= doMerge(I->second, M, true, I->first);
594 bool GlobalMerge::runOnFunction(Function &F) {
598 bool GlobalMerge::doFinalization(Module &M) {
599 MustKeepGlobalVariables.clear();
603 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
604 bool OnlyOptimizeForSize) {
605 return new GlobalMerge(TM, Offset, OnlyOptimizeForSize);