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 // FIXME: Infer the maximum possible offset depending on the actual users
121 // (these max offsets are different for the users inside Thumb or ARM
122 // functions), see the code that passes in the offset in the ARM backend
123 // for more information.
126 /// Whether we should try to optimize for size only.
127 /// Currently, this applies a dead simple heuristic: only consider globals
128 /// used in minsize functions for merging.
129 /// FIXME: This could learn about optsize, and be used in the cost model.
130 bool OnlyOptimizeForSize;
132 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
133 Module &M, bool isConst, unsigned AddrSpace) const;
134 /// \brief Merge everything in \p Globals for which the corresponding bit
135 /// in \p GlobalSet is set.
136 bool doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
137 const BitVector &GlobalSet, Module &M, bool isConst,
138 unsigned AddrSpace) const;
140 /// \brief Check if the given variable has been identified as must keep
141 /// \pre setMustKeepGlobalVariables must have been called on the Module that
143 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
144 return MustKeepGlobalVariables.count(GV);
147 /// Collect every variables marked as "used" or used in a landing pad
148 /// instruction for this Module.
149 void setMustKeepGlobalVariables(Module &M);
151 /// Collect every variables marked as "used"
152 void collectUsedGlobalVariables(Module &M);
154 /// Keep track of the GlobalVariable that must not be merged away
155 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
158 static char ID; // Pass identification, replacement for typeid.
159 explicit GlobalMerge(const TargetMachine *TM = nullptr,
160 unsigned MaximalOffset = 0,
161 bool OnlyOptimizeForSize = false)
162 : FunctionPass(ID), TM(TM), MaxOffset(MaximalOffset),
163 OnlyOptimizeForSize(OnlyOptimizeForSize) {
164 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
167 bool doInitialization(Module &M) override;
168 bool runOnFunction(Function &F) override;
169 bool doFinalization(Module &M) override;
171 const char *getPassName() const override {
172 return "Merge internal globals";
175 void getAnalysisUsage(AnalysisUsage &AU) const override {
176 AU.setPreservesCFG();
177 FunctionPass::getAnalysisUsage(AU);
180 } // end anonymous namespace
182 char GlobalMerge::ID = 0;
183 INITIALIZE_PASS_BEGIN(GlobalMerge, "global-merge", "Merge global variables",
185 INITIALIZE_PASS_END(GlobalMerge, "global-merge", "Merge global variables",
188 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
189 Module &M, bool isConst, unsigned AddrSpace) const {
190 auto &DL = M.getDataLayout();
191 // FIXME: Find better heuristics
193 Globals.begin(), Globals.end(),
194 [&DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
195 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
196 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
198 return (DL.getTypeAllocSize(Ty1) < DL.getTypeAllocSize(Ty2));
201 // If we want to just blindly group all globals together, do so.
202 if (!GlobalMergeGroupByUse) {
203 BitVector AllGlobals(Globals.size());
205 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
208 // If we want to be smarter, look at all uses of each global, to try to
209 // discover all sets of globals used together, and how many times each of
210 // these sets occured.
212 // Keep this reasonably efficient, by having an append-only list of all sets
213 // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
214 // code (currently, a Function) to the set of globals seen so far that are
215 // used together in that unit (GlobalUsesByFunction).
217 // When we look at the Nth global, we now that any new set is either:
218 // - the singleton set {N}, containing this global only, or
219 // - the union of {N} and a previously-discovered set, containing some
220 // combination of the previous N-1 globals.
221 // Using that knowledge, when looking at the Nth global, we can keep:
222 // - a reference to the singleton set {N} (CurGVOnlySetIdx)
223 // - a list mapping each previous set to its union with {N} (EncounteredUGS),
224 // if it actually occurs.
226 // We keep track of the sets of globals used together "close enough".
227 struct UsedGlobalSet {
228 UsedGlobalSet(size_t Size) : Globals(Size), UsageCount(1) {}
233 // Each set is unique in UsedGlobalSets.
234 std::vector<UsedGlobalSet> UsedGlobalSets;
236 // Avoid repeating the create-global-set pattern.
237 auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
238 UsedGlobalSets.emplace_back(Globals.size());
239 return UsedGlobalSets.back();
242 // The first set is the empty set.
243 CreateGlobalSet().UsageCount = 0;
245 // We define "close enough" to be "in the same function".
246 // FIXME: Grouping uses by function is way too aggressive, so we should have
247 // a better metric for distance between uses.
248 // The obvious alternative would be to group by BasicBlock, but that's in
249 // turn too conservative..
250 // Anything in between wouldn't be trivial to compute, so just stick with
251 // per-function grouping.
253 // The value type is an index into UsedGlobalSets.
254 // The default (0) conveniently points to the empty set.
255 DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
257 // Now, look at each merge-eligible global in turn.
259 // Keep track of the sets we already encountered to which we added the
261 // Each element matches the same-index element in UsedGlobalSets.
262 // This lets us efficiently tell whether a set has already been expanded to
263 // include the current global.
264 std::vector<size_t> EncounteredUGS;
266 for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
267 GlobalVariable *GV = Globals[GI];
269 // Reset the encountered sets for this global...
270 std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
271 // ...and grow it in case we created new sets for the previous global.
272 EncounteredUGS.resize(UsedGlobalSets.size());
274 // We might need to create a set that only consists of the current global.
275 // Keep track of its index into UsedGlobalSets.
276 size_t CurGVOnlySetIdx = 0;
278 // For each global, look at all its Uses.
279 for (auto &U : GV->uses()) {
280 // This Use might be a ConstantExpr. We're interested in Instruction
281 // users, so look through ConstantExpr...
283 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
286 UI = &*CE->use_begin();
288 } else if (isa<Instruction>(U.getUser())) {
295 // ...to iterate on all the instruction users of the global.
296 // Note that we iterate on Uses and not on Users to be able to getNext().
297 for (; UI != UE; UI = UI->getNext()) {
298 Instruction *I = dyn_cast<Instruction>(UI->getUser());
302 Function *ParentFn = I->getParent()->getParent();
304 // If we're only optimizing for size, ignore non-minsize functions.
305 if (OnlyOptimizeForSize &&
306 !ParentFn->hasFnAttribute(Attribute::MinSize))
309 size_t UGSIdx = GlobalUsesByFunction[ParentFn];
311 // If this is the first global the basic block uses, map it to the set
312 // consisting of this global only.
314 // If that set doesn't exist yet, create it.
315 if (!CurGVOnlySetIdx) {
316 CurGVOnlySetIdx = UsedGlobalSets.size();
317 CreateGlobalSet().Globals.set(GI);
319 ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
322 GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
326 // If we already encountered this BB, just increment the counter.
327 if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
328 ++UsedGlobalSets[UGSIdx].UsageCount;
332 // If not, the previous set wasn't actually used in this function.
333 --UsedGlobalSets[UGSIdx].UsageCount;
335 // If we already expanded the previous set to include this global, just
336 // reuse that expanded set.
337 if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
338 ++UsedGlobalSets[ExpandedIdx].UsageCount;
339 GlobalUsesByFunction[ParentFn] = ExpandedIdx;
343 // If not, create a new set consisting of the union of the previous set
344 // and this global. Mark it as encountered, so we can reuse it later.
345 GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
346 UsedGlobalSets.size();
348 UsedGlobalSet &NewUGS = CreateGlobalSet();
349 NewUGS.Globals.set(GI);
350 NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
355 // Now we found a bunch of sets of globals used together. We accumulated
356 // the number of times we encountered the sets (i.e., the number of blocks
357 // that use that exact set of globals).
359 // Multiply that by the size of the set to give us a crude profitability
361 std::sort(UsedGlobalSets.begin(), UsedGlobalSets.end(),
362 [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
363 return UGS1.Globals.count() * UGS1.UsageCount <
364 UGS2.Globals.count() * UGS2.UsageCount;
367 // We can choose to merge all globals together, but ignore globals never used
368 // with another global. This catches the obviously non-profitable cases of
369 // having a single global, but is aggressive enough for any other case.
370 if (GlobalMergeIgnoreSingleUse) {
371 BitVector AllGlobals(Globals.size());
372 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
373 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
374 if (UGS.UsageCount == 0)
376 if (UGS.Globals.count() > 1)
377 AllGlobals |= UGS.Globals;
379 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
382 // Starting from the sets with the best (=biggest) profitability, find a
384 // The ideal (and expensive) solution can only be found by trying all
385 // combinations, looking for the one with the best profitability.
386 // Don't be smart about it, and just pick the first compatible combination,
387 // starting with the sets with the best profitability.
388 BitVector PickedGlobals(Globals.size());
389 bool Changed = false;
391 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
392 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
393 if (UGS.UsageCount == 0)
395 if (PickedGlobals.anyCommon(UGS.Globals))
397 PickedGlobals |= UGS.Globals;
398 // If the set only contains one global, there's no point in merging.
399 // Ignore the global for inclusion in other sets though, so keep it in
401 if (UGS.Globals.count() < 2)
403 Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
409 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
410 const BitVector &GlobalSet, Module &M, bool isConst,
411 unsigned AddrSpace) const {
413 Type *Int32Ty = Type::getInt32Ty(M.getContext());
414 auto &DL = M.getDataLayout();
416 assert(Globals.size() > 1);
418 DEBUG(dbgs() << " Trying to merge set, starts with #"
419 << GlobalSet.find_first() << "\n");
421 ssize_t i = GlobalSet.find_first();
424 uint64_t MergedSize = 0;
425 std::vector<Type*> Tys;
426 std::vector<Constant*> Inits;
428 bool HasExternal = false;
429 GlobalVariable *TheFirstExternal = 0;
430 for (j = i; j != -1; j = GlobalSet.find_next(j)) {
431 Type *Ty = Globals[j]->getType()->getElementType();
432 MergedSize += DL.getTypeAllocSize(Ty);
433 if (MergedSize > MaxOffset) {
437 Inits.push_back(Globals[j]->getInitializer());
439 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
441 TheFirstExternal = Globals[j];
445 // If merged variables doesn't have external linkage, we needn't to expose
446 // the symbol after merging.
447 GlobalValue::LinkageTypes Linkage = HasExternal
448 ? GlobalValue::ExternalLinkage
449 : GlobalValue::InternalLinkage;
451 StructType *MergedTy = StructType::get(M.getContext(), Tys);
452 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
454 // If merged variables have external linkage, we use symbol name of the
455 // first variable merged as the suffix of global symbol name. This would
456 // be able to avoid the link-time naming conflict for globalm symbols.
457 GlobalVariable *MergedGV = new GlobalVariable(
458 M, MergedTy, isConst, Linkage, MergedInit,
459 HasExternal ? "_MergedGlobals_" + TheFirstExternal->getName()
461 nullptr, GlobalVariable::NotThreadLocal, AddrSpace);
463 for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k)) {
464 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
465 std::string Name = Globals[k]->getName();
468 ConstantInt::get(Int32Ty, 0),
469 ConstantInt::get(Int32Ty, idx++)
472 ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
473 Globals[k]->replaceAllUsesWith(GEP);
474 Globals[k]->eraseFromParent();
476 if (Linkage != GlobalValue::InternalLinkage) {
477 // Generate a new alias...
478 auto *PTy = cast<PointerType>(GEP->getType());
479 GlobalAlias::create(PTy, Linkage, Name, GEP, &M);
490 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
491 // Extract global variables from llvm.used array
492 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
493 if (!GV || !GV->hasInitializer()) return;
495 // Should be an array of 'i8*'.
496 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
498 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
499 if (const GlobalVariable *G =
500 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
501 MustKeepGlobalVariables.insert(G);
504 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
505 collectUsedGlobalVariables(M);
507 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
509 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
510 IBB != IEndBB; ++IBB) {
511 // Follow the invoke link to find the landing pad instruction
512 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
515 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
516 // Look for globals in the clauses of the landing pad instruction
517 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
518 Idx != NumClauses; ++Idx)
519 if (const GlobalVariable *GV =
520 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
521 ->stripPointerCasts()))
522 MustKeepGlobalVariables.insert(GV);
527 bool GlobalMerge::doInitialization(Module &M) {
528 if (!EnableGlobalMerge)
531 auto &DL = M.getDataLayout();
532 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
534 bool Changed = false;
535 setMustKeepGlobalVariables(M);
537 // Grab all non-const globals.
538 for (Module::global_iterator I = M.global_begin(),
539 E = M.global_end(); I != E; ++I) {
540 // Merge is safe for "normal" internal or external globals only
541 if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
544 if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage()) &&
545 !I->hasInternalLinkage())
548 PointerType *PT = dyn_cast<PointerType>(I->getType());
549 assert(PT && "Global variable is not a pointer!");
551 unsigned AddressSpace = PT->getAddressSpace();
553 // Ignore fancy-aligned globals for now.
554 unsigned Alignment = DL.getPreferredAlignment(I);
555 Type *Ty = I->getType()->getElementType();
556 if (Alignment > DL.getABITypeAlignment(Ty))
559 // Ignore all 'special' globals.
560 if (I->getName().startswith("llvm.") ||
561 I->getName().startswith(".llvm."))
564 // Ignore all "required" globals:
565 if (isMustKeepGlobalVariable(I))
568 if (DL.getTypeAllocSize(Ty) < MaxOffset) {
569 if (TargetLoweringObjectFile::getKindForGlobal(I, *TM).isBSSLocal())
570 BSSGlobals[AddressSpace].push_back(I);
571 else if (I->isConstant())
572 ConstGlobals[AddressSpace].push_back(I);
574 Globals[AddressSpace].push_back(I);
578 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
579 I = Globals.begin(), E = Globals.end(); I != E; ++I)
580 if (I->second.size() > 1)
581 Changed |= doMerge(I->second, M, false, I->first);
583 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
584 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
585 if (I->second.size() > 1)
586 Changed |= doMerge(I->second, M, false, I->first);
588 if (EnableGlobalMergeOnConst)
589 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
590 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
591 if (I->second.size() > 1)
592 Changed |= doMerge(I->second, M, true, I->first);
597 bool GlobalMerge::runOnFunction(Function &F) {
601 bool GlobalMerge::doFinalization(Module &M) {
602 MustKeepGlobalVariables.clear();
606 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
607 bool OnlyOptimizeForSize) {
608 return new GlobalMerge(TM, Offset, OnlyOptimizeForSize);