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.
111 static cl::opt<cl::boolOrDefault>
112 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
113 cl::desc("Enable global merge pass on external linkage"));
115 STATISTIC(NumMerged, "Number of globals merged");
117 class GlobalMerge : public FunctionPass {
118 const TargetMachine *TM;
119 // FIXME: Infer the maximum possible offset depending on the actual users
120 // (these max offsets are different for the users inside Thumb or ARM
121 // functions), see the code that passes in the offset in the ARM backend
122 // for more information.
125 /// Whether we should try to optimize for size only.
126 /// Currently, this applies a dead simple heuristic: only consider globals
127 /// used in minsize functions for merging.
128 /// FIXME: This could learn about optsize, and be used in the cost model.
129 bool OnlyOptimizeForSize;
131 /// Whether we should merge global variables that have external linkage.
132 bool MergeExternalGlobals;
134 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
135 Module &M, bool isConst, unsigned AddrSpace) const;
136 /// \brief Merge everything in \p Globals for which the corresponding bit
137 /// in \p GlobalSet is set.
138 bool doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
139 const BitVector &GlobalSet, Module &M, bool isConst,
140 unsigned AddrSpace) const;
142 /// \brief Check if the given variable has been identified as must keep
143 /// \pre setMustKeepGlobalVariables must have been called on the Module that
145 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
146 return MustKeepGlobalVariables.count(GV);
149 /// Collect every variables marked as "used" or used in a landing pad
150 /// instruction for this Module.
151 void setMustKeepGlobalVariables(Module &M);
153 /// Collect every variables marked as "used"
154 void collectUsedGlobalVariables(Module &M);
156 /// Keep track of the GlobalVariable that must not be merged away
157 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
160 static char ID; // Pass identification, replacement for typeid.
161 explicit GlobalMerge(const TargetMachine *TM = nullptr,
162 unsigned MaximalOffset = 0,
163 bool OnlyOptimizeForSize = false,
164 bool MergeExternalGlobals = false)
165 : FunctionPass(ID), TM(TM), MaxOffset(MaximalOffset),
166 OnlyOptimizeForSize(OnlyOptimizeForSize),
167 MergeExternalGlobals(MergeExternalGlobals) {
168 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
171 bool doInitialization(Module &M) override;
172 bool runOnFunction(Function &F) override;
173 bool doFinalization(Module &M) override;
175 const char *getPassName() const override {
176 return "Merge internal globals";
179 void getAnalysisUsage(AnalysisUsage &AU) const override {
180 AU.setPreservesCFG();
181 FunctionPass::getAnalysisUsage(AU);
184 } // end anonymous namespace
186 char GlobalMerge::ID = 0;
187 INITIALIZE_PASS_BEGIN(GlobalMerge, "global-merge", "Merge global variables",
189 INITIALIZE_PASS_END(GlobalMerge, "global-merge", "Merge global variables",
192 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
193 Module &M, bool isConst, unsigned AddrSpace) const {
194 auto &DL = M.getDataLayout();
195 // FIXME: Find better heuristics
197 Globals.begin(), Globals.end(),
198 [&DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
199 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
200 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
202 return (DL.getTypeAllocSize(Ty1) < DL.getTypeAllocSize(Ty2));
205 // If we want to just blindly group all globals together, do so.
206 if (!GlobalMergeGroupByUse) {
207 BitVector AllGlobals(Globals.size());
209 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
212 // If we want to be smarter, look at all uses of each global, to try to
213 // discover all sets of globals used together, and how many times each of
214 // these sets occurred.
216 // Keep this reasonably efficient, by having an append-only list of all sets
217 // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
218 // code (currently, a Function) to the set of globals seen so far that are
219 // used together in that unit (GlobalUsesByFunction).
221 // When we look at the Nth global, we now that any new set is either:
222 // - the singleton set {N}, containing this global only, or
223 // - the union of {N} and a previously-discovered set, containing some
224 // combination of the previous N-1 globals.
225 // Using that knowledge, when looking at the Nth global, we can keep:
226 // - a reference to the singleton set {N} (CurGVOnlySetIdx)
227 // - a list mapping each previous set to its union with {N} (EncounteredUGS),
228 // if it actually occurs.
230 // We keep track of the sets of globals used together "close enough".
231 struct UsedGlobalSet {
232 UsedGlobalSet(size_t Size) : Globals(Size), UsageCount(1) {}
237 // Each set is unique in UsedGlobalSets.
238 std::vector<UsedGlobalSet> UsedGlobalSets;
240 // Avoid repeating the create-global-set pattern.
241 auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
242 UsedGlobalSets.emplace_back(Globals.size());
243 return UsedGlobalSets.back();
246 // The first set is the empty set.
247 CreateGlobalSet().UsageCount = 0;
249 // We define "close enough" to be "in the same function".
250 // FIXME: Grouping uses by function is way too aggressive, so we should have
251 // a better metric for distance between uses.
252 // The obvious alternative would be to group by BasicBlock, but that's in
253 // turn too conservative..
254 // Anything in between wouldn't be trivial to compute, so just stick with
255 // per-function grouping.
257 // The value type is an index into UsedGlobalSets.
258 // The default (0) conveniently points to the empty set.
259 DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
261 // Now, look at each merge-eligible global in turn.
263 // Keep track of the sets we already encountered to which we added the
265 // Each element matches the same-index element in UsedGlobalSets.
266 // This lets us efficiently tell whether a set has already been expanded to
267 // include the current global.
268 std::vector<size_t> EncounteredUGS;
270 for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
271 GlobalVariable *GV = Globals[GI];
273 // Reset the encountered sets for this global...
274 std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
275 // ...and grow it in case we created new sets for the previous global.
276 EncounteredUGS.resize(UsedGlobalSets.size());
278 // We might need to create a set that only consists of the current global.
279 // Keep track of its index into UsedGlobalSets.
280 size_t CurGVOnlySetIdx = 0;
282 // For each global, look at all its Uses.
283 for (auto &U : GV->uses()) {
284 // This Use might be a ConstantExpr. We're interested in Instruction
285 // users, so look through ConstantExpr...
287 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
290 UI = &*CE->use_begin();
292 } else if (isa<Instruction>(U.getUser())) {
299 // ...to iterate on all the instruction users of the global.
300 // Note that we iterate on Uses and not on Users to be able to getNext().
301 for (; UI != UE; UI = UI->getNext()) {
302 Instruction *I = dyn_cast<Instruction>(UI->getUser());
306 Function *ParentFn = I->getParent()->getParent();
308 // If we're only optimizing for size, ignore non-minsize functions.
309 if (OnlyOptimizeForSize &&
310 !ParentFn->hasFnAttribute(Attribute::MinSize))
313 size_t UGSIdx = GlobalUsesByFunction[ParentFn];
315 // If this is the first global the basic block uses, map it to the set
316 // consisting of this global only.
318 // If that set doesn't exist yet, create it.
319 if (!CurGVOnlySetIdx) {
320 CurGVOnlySetIdx = UsedGlobalSets.size();
321 CreateGlobalSet().Globals.set(GI);
323 ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
326 GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
330 // If we already encountered this BB, just increment the counter.
331 if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
332 ++UsedGlobalSets[UGSIdx].UsageCount;
336 // If not, the previous set wasn't actually used in this function.
337 --UsedGlobalSets[UGSIdx].UsageCount;
339 // If we already expanded the previous set to include this global, just
340 // reuse that expanded set.
341 if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
342 ++UsedGlobalSets[ExpandedIdx].UsageCount;
343 GlobalUsesByFunction[ParentFn] = ExpandedIdx;
347 // If not, create a new set consisting of the union of the previous set
348 // and this global. Mark it as encountered, so we can reuse it later.
349 GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
350 UsedGlobalSets.size();
352 UsedGlobalSet &NewUGS = CreateGlobalSet();
353 NewUGS.Globals.set(GI);
354 NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
359 // Now we found a bunch of sets of globals used together. We accumulated
360 // the number of times we encountered the sets (i.e., the number of blocks
361 // that use that exact set of globals).
363 // Multiply that by the size of the set to give us a crude profitability
365 std::sort(UsedGlobalSets.begin(), UsedGlobalSets.end(),
366 [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
367 return UGS1.Globals.count() * UGS1.UsageCount <
368 UGS2.Globals.count() * UGS2.UsageCount;
371 // We can choose to merge all globals together, but ignore globals never used
372 // with another global. This catches the obviously non-profitable cases of
373 // having a single global, but is aggressive enough for any other case.
374 if (GlobalMergeIgnoreSingleUse) {
375 BitVector AllGlobals(Globals.size());
376 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
377 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
378 if (UGS.UsageCount == 0)
380 if (UGS.Globals.count() > 1)
381 AllGlobals |= UGS.Globals;
383 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
386 // Starting from the sets with the best (=biggest) profitability, find a
388 // The ideal (and expensive) solution can only be found by trying all
389 // combinations, looking for the one with the best profitability.
390 // Don't be smart about it, and just pick the first compatible combination,
391 // starting with the sets with the best profitability.
392 BitVector PickedGlobals(Globals.size());
393 bool Changed = false;
395 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
396 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
397 if (UGS.UsageCount == 0)
399 if (PickedGlobals.anyCommon(UGS.Globals))
401 PickedGlobals |= UGS.Globals;
402 // If the set only contains one global, there's no point in merging.
403 // Ignore the global for inclusion in other sets though, so keep it in
405 if (UGS.Globals.count() < 2)
407 Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
413 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
414 const BitVector &GlobalSet, Module &M, bool isConst,
415 unsigned AddrSpace) const {
417 Type *Int32Ty = Type::getInt32Ty(M.getContext());
418 auto &DL = M.getDataLayout();
420 assert(Globals.size() > 1);
422 DEBUG(dbgs() << " Trying to merge set, starts with #"
423 << GlobalSet.find_first() << "\n");
425 ssize_t i = GlobalSet.find_first();
428 uint64_t MergedSize = 0;
429 std::vector<Type*> Tys;
430 std::vector<Constant*> Inits;
432 for (j = i; j != -1; j = GlobalSet.find_next(j)) {
433 Type *Ty = Globals[j]->getType()->getElementType();
434 MergedSize += DL.getTypeAllocSize(Ty);
435 if (MergedSize > MaxOffset) {
439 Inits.push_back(Globals[j]->getInitializer());
442 StructType *MergedTy = StructType::get(M.getContext(), Tys);
443 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
445 GlobalVariable *MergedGV = new GlobalVariable(
446 M, MergedTy, isConst, GlobalValue::PrivateLinkage, MergedInit,
447 "_MergedGlobals", nullptr, GlobalVariable::NotThreadLocal, AddrSpace);
449 for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k)) {
450 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
451 std::string Name = Globals[k]->getName();
454 ConstantInt::get(Int32Ty, 0),
455 ConstantInt::get(Int32Ty, idx++)
458 ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
459 Globals[k]->replaceAllUsesWith(GEP);
460 Globals[k]->eraseFromParent();
462 // When the linkage is not internal we must emit an alias for the original
463 // variable name as it may be accessed from another object. On non-Mach-O
464 // we can also emit an alias for internal linkage as it's safe to do so.
465 // It's not safe on Mach-O as the alias (and thus the portion of the
466 // MergedGlobals variable) may be dead stripped at link time.
467 if (Linkage != GlobalValue::InternalLinkage ||
468 !TM->getTargetTriple().isOSBinFormatMachO()) {
469 auto *PTy = cast<PointerType>(GEP->getType());
470 GlobalAlias::create(PTy, Linkage, Name, GEP, &M);
481 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
482 // Extract global variables from llvm.used array
483 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
484 if (!GV || !GV->hasInitializer()) return;
486 // Should be an array of 'i8*'.
487 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
489 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
490 if (const GlobalVariable *G =
491 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
492 MustKeepGlobalVariables.insert(G);
495 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
496 collectUsedGlobalVariables(M);
498 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
500 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
501 IBB != IEndBB; ++IBB) {
502 // Follow the invoke link to find the landing pad instruction
503 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
506 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
507 // Look for globals in the clauses of the landing pad instruction
508 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
509 Idx != NumClauses; ++Idx)
510 if (const GlobalVariable *GV =
511 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
512 ->stripPointerCasts()))
513 MustKeepGlobalVariables.insert(GV);
518 bool GlobalMerge::doInitialization(Module &M) {
519 if (!EnableGlobalMerge)
522 auto &DL = M.getDataLayout();
523 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
525 bool Changed = false;
526 setMustKeepGlobalVariables(M);
528 // Grab all non-const globals.
529 for (Module::global_iterator I = M.global_begin(),
530 E = M.global_end(); I != E; ++I) {
531 // Merge is safe for "normal" internal or external globals only
532 if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
535 if (!(MergeExternalGlobals && I->hasExternalLinkage()) &&
536 !I->hasInternalLinkage())
539 PointerType *PT = dyn_cast<PointerType>(I->getType());
540 assert(PT && "Global variable is not a pointer!");
542 unsigned AddressSpace = PT->getAddressSpace();
544 // Ignore fancy-aligned globals for now.
545 unsigned Alignment = DL.getPreferredAlignment(I);
546 Type *Ty = I->getType()->getElementType();
547 if (Alignment > DL.getABITypeAlignment(Ty))
550 // Ignore all 'special' globals.
551 if (I->getName().startswith("llvm.") ||
552 I->getName().startswith(".llvm."))
555 // Ignore all "required" globals:
556 if (isMustKeepGlobalVariable(I))
559 if (DL.getTypeAllocSize(Ty) < MaxOffset) {
560 if (TargetLoweringObjectFile::getKindForGlobal(I, *TM).isBSSLocal())
561 BSSGlobals[AddressSpace].push_back(I);
562 else if (I->isConstant())
563 ConstGlobals[AddressSpace].push_back(I);
565 Globals[AddressSpace].push_back(I);
569 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
570 I = Globals.begin(), E = Globals.end(); I != E; ++I)
571 if (I->second.size() > 1)
572 Changed |= doMerge(I->second, M, false, I->first);
574 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
575 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
576 if (I->second.size() > 1)
577 Changed |= doMerge(I->second, M, false, I->first);
579 if (EnableGlobalMergeOnConst)
580 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
581 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
582 if (I->second.size() > 1)
583 Changed |= doMerge(I->second, M, true, I->first);
588 bool GlobalMerge::runOnFunction(Function &F) {
592 bool GlobalMerge::doFinalization(Module &M) {
593 MustKeepGlobalVariables.clear();
597 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
598 bool OnlyOptimizeForSize,
599 bool MergeExternalByDefault) {
600 bool MergeExternal = (EnableGlobalMergeOnExternal == cl::BOU_UNSET) ?
601 MergeExternalByDefault : (EnableGlobalMergeOnExternal == cl::BOU_TRUE);
602 return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal);