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 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
128 Module &M, bool isConst, unsigned AddrSpace) const;
129 /// \brief Merge everything in \p Globals for which the corresponding bit
130 /// in \p GlobalSet is set.
131 bool doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
132 const BitVector &GlobalSet, Module &M, bool isConst,
133 unsigned AddrSpace) const;
135 /// \brief Check if the given variable has been identified as must keep
136 /// \pre setMustKeepGlobalVariables must have been called on the Module that
138 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
139 return MustKeepGlobalVariables.count(GV);
142 /// Collect every variables marked as "used" or used in a landing pad
143 /// instruction for this Module.
144 void setMustKeepGlobalVariables(Module &M);
146 /// Collect every variables marked as "used"
147 void collectUsedGlobalVariables(Module &M);
149 /// Keep track of the GlobalVariable that must not be merged away
150 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
153 static char ID; // Pass identification, replacement for typeid.
154 explicit GlobalMerge(const TargetMachine *TM = nullptr,
155 unsigned MaximalOffset = 0)
156 : FunctionPass(ID), TM(TM), DL(TM->getDataLayout()),
157 MaxOffset(MaximalOffset) {
158 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
161 bool doInitialization(Module &M) override;
162 bool runOnFunction(Function &F) override;
163 bool doFinalization(Module &M) override;
165 const char *getPassName() const override {
166 return "Merge internal globals";
169 void getAnalysisUsage(AnalysisUsage &AU) const override {
170 AU.setPreservesCFG();
171 FunctionPass::getAnalysisUsage(AU);
174 } // end anonymous namespace
176 char GlobalMerge::ID = 0;
177 INITIALIZE_PASS_BEGIN(GlobalMerge, "global-merge", "Merge global variables",
179 INITIALIZE_PASS_END(GlobalMerge, "global-merge", "Merge global variables",
182 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
183 Module &M, bool isConst, unsigned AddrSpace) const {
184 // FIXME: Find better heuristics
185 std::stable_sort(Globals.begin(), Globals.end(),
186 [this](const GlobalVariable *GV1, const GlobalVariable *GV2) {
187 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
188 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
190 return (DL->getTypeAllocSize(Ty1) < DL->getTypeAllocSize(Ty2));
193 // If we want to just blindly group all globals together, do so.
194 if (!GlobalMergeGroupByUse) {
195 BitVector AllGlobals(Globals.size());
197 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
200 // If we want to be smarter, look at all uses of each global, to try to
201 // discover all sets of globals used together, and how many times each of
202 // these sets occured.
204 // Keep this reasonably efficient, by having an append-only list of all sets
205 // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
206 // code (currently, a Function) to the set of globals seen so far that are
207 // used together in that unit (GlobalUsesByFunction).
209 // When we look at the Nth global, we now that any new set is either:
210 // - the singleton set {N}, containing this global only, or
211 // - the union of {N} and a previously-discovered set, containing some
212 // combination of the previous N-1 globals.
213 // Using that knowledge, when looking at the Nth global, we can keep:
214 // - a reference to the singleton set {N} (CurGVOnlySetIdx)
215 // - a list mapping each previous set to its union with {N} (EncounteredUGS),
216 // if it actually occurs.
218 // We keep track of the sets of globals used together "close enough".
219 struct UsedGlobalSet {
220 UsedGlobalSet(size_t Size) : Globals(Size), UsageCount(1) {}
225 // Each set is unique in UsedGlobalSets.
226 std::vector<UsedGlobalSet> UsedGlobalSets;
228 // Avoid repeating the create-global-set pattern.
229 auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
230 UsedGlobalSets.emplace_back(Globals.size());
231 return UsedGlobalSets.back();
234 // The first set is the empty set.
235 CreateGlobalSet().UsageCount = 0;
237 // We define "close enough" to be "in the same function".
238 // FIXME: Grouping uses by function is way too aggressive, so we should have
239 // a better metric for distance between uses.
240 // The obvious alternative would be to group by BasicBlock, but that's in
241 // turn too conservative..
242 // Anything in between wouldn't be trivial to compute, so just stick with
243 // per-function grouping.
245 // The value type is an index into UsedGlobalSets.
246 // The default (0) conveniently points to the empty set.
247 DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
249 // Now, look at each merge-eligible global in turn.
251 // Keep track of the sets we already encountered to which we added the
253 // Each element matches the same-index element in UsedGlobalSets.
254 // This lets us efficiently tell whether a set has already been expanded to
255 // include the current global.
256 std::vector<size_t> EncounteredUGS;
258 for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
259 GlobalVariable *GV = Globals[GI];
261 // Reset the encountered sets for this global...
262 std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
263 // ...and grow it in case we created new sets for the previous global.
264 EncounteredUGS.resize(UsedGlobalSets.size());
266 // We might need to create a set that only consists of the current global.
267 // Keep track of its index into UsedGlobalSets.
268 size_t CurGVOnlySetIdx = 0;
270 // For each global, look at all its Uses.
271 for (auto &U : GV->uses()) {
272 // This Use might be a ConstantExpr. We're interested in Instruction
273 // users, so look through ConstantExpr...
275 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
276 UI = &*CE->use_begin();
278 } else if (isa<Instruction>(U.getUser())) {
285 // ...to iterate on all the instruction users of the global.
286 // Note that we iterate on Uses and not on Users to be able to getNext().
287 for (; UI != UE; UI = UI->getNext()) {
288 Instruction *I = dyn_cast<Instruction>(UI->getUser());
292 Function *ParentFn = I->getParent()->getParent();
293 size_t UGSIdx = GlobalUsesByFunction[ParentFn];
295 // If this is the first global the basic block uses, map it to the set
296 // consisting of this global only.
298 // If that set doesn't exist yet, create it.
299 if (!CurGVOnlySetIdx) {
300 CurGVOnlySetIdx = UsedGlobalSets.size();
301 CreateGlobalSet().Globals.set(GI);
303 ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
306 GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
310 // If we already encountered this BB, just increment the counter.
311 if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
312 ++UsedGlobalSets[UGSIdx].UsageCount;
316 // If not, the previous set wasn't actually used in this function.
317 --UsedGlobalSets[UGSIdx].UsageCount;
319 // If we already expanded the previous set to include this global, just
320 // reuse that expanded set.
321 if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
322 ++UsedGlobalSets[ExpandedIdx].UsageCount;
323 GlobalUsesByFunction[ParentFn] = ExpandedIdx;
327 // If not, create a new set consisting of the union of the previous set
328 // and this global. Mark it as encountered, so we can reuse it later.
329 GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
330 UsedGlobalSets.size();
332 UsedGlobalSet &NewUGS = CreateGlobalSet();
333 NewUGS.Globals.set(GI);
334 NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
339 // Now we found a bunch of sets of globals used together. We accumulated
340 // the number of times we encountered the sets (i.e., the number of blocks
341 // that use that exact set of globals).
343 // Multiply that by the size of the set to give us a crude profitability
345 std::sort(UsedGlobalSets.begin(), UsedGlobalSets.end(),
346 [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
347 return UGS1.Globals.count() * UGS1.UsageCount <
348 UGS2.Globals.count() * UGS2.UsageCount;
351 // We can choose to merge all globals together, but ignore globals never used
352 // with another global. This catches the obviously non-profitable cases of
353 // having a single global, but is aggressive enough for any other case.
354 if (GlobalMergeIgnoreSingleUse) {
355 BitVector AllGlobals(Globals.size());
356 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
357 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
358 if (UGS.UsageCount == 0)
360 if (UGS.Globals.count() > 1)
361 AllGlobals |= UGS.Globals;
363 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
366 // Starting from the sets with the best (=biggest) profitability, find a
368 // The ideal (and expensive) solution can only be found by trying all
369 // combinations, looking for the one with the best profitability.
370 // Don't be smart about it, and just pick the first compatible combination,
371 // starting with the sets with the best profitability.
372 BitVector PickedGlobals(Globals.size());
373 bool Changed = false;
375 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
376 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
377 if (UGS.UsageCount == 0)
379 if (PickedGlobals.anyCommon(UGS.Globals))
381 PickedGlobals |= UGS.Globals;
382 // If the set only contains one global, there's no point in merging.
383 // Ignore the global for inclusion in other sets though, so keep it in
385 if (UGS.Globals.count() < 2)
387 Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
393 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable *> &Globals,
394 const BitVector &GlobalSet, Module &M, bool isConst,
395 unsigned AddrSpace) const {
397 Type *Int32Ty = Type::getInt32Ty(M.getContext());
399 assert(Globals.size() > 1);
401 DEBUG(dbgs() << " Trying to merge set, starts with #"
402 << GlobalSet.find_first() << "\n");
404 ssize_t i = GlobalSet.find_first();
407 uint64_t MergedSize = 0;
408 std::vector<Type*> Tys;
409 std::vector<Constant*> Inits;
411 bool HasExternal = false;
412 GlobalVariable *TheFirstExternal = 0;
413 for (j = i; j != -1; j = GlobalSet.find_next(j)) {
414 Type *Ty = Globals[j]->getType()->getElementType();
415 MergedSize += DL->getTypeAllocSize(Ty);
416 if (MergedSize > MaxOffset) {
420 Inits.push_back(Globals[j]->getInitializer());
422 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
424 TheFirstExternal = Globals[j];
428 // If merged variables doesn't have external linkage, we needn't to expose
429 // the symbol after merging.
430 GlobalValue::LinkageTypes Linkage = HasExternal
431 ? GlobalValue::ExternalLinkage
432 : GlobalValue::InternalLinkage;
434 StructType *MergedTy = StructType::get(M.getContext(), Tys);
435 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
437 // If merged variables have external linkage, we use symbol name of the
438 // first variable merged as the suffix of global symbol name. This would
439 // be able to avoid the link-time naming conflict for globalm symbols.
440 GlobalVariable *MergedGV = new GlobalVariable(
441 M, MergedTy, isConst, Linkage, MergedInit,
442 HasExternal ? "_MergedGlobals_" + TheFirstExternal->getName()
444 nullptr, GlobalVariable::NotThreadLocal, AddrSpace);
446 for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k)) {
447 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
448 std::string Name = Globals[k]->getName();
451 ConstantInt::get(Int32Ty, 0),
452 ConstantInt::get(Int32Ty, idx++)
455 ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
456 Globals[k]->replaceAllUsesWith(GEP);
457 Globals[k]->eraseFromParent();
459 if (Linkage != GlobalValue::InternalLinkage) {
460 // Generate a new alias...
461 auto *PTy = cast<PointerType>(GEP->getType());
462 GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
463 Linkage, Name, GEP, &M);
474 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
475 // Extract global variables from llvm.used array
476 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
477 if (!GV || !GV->hasInitializer()) return;
479 // Should be an array of 'i8*'.
480 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
482 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
483 if (const GlobalVariable *G =
484 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
485 MustKeepGlobalVariables.insert(G);
488 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
489 collectUsedGlobalVariables(M);
491 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
493 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
494 IBB != IEndBB; ++IBB) {
495 // Follow the invoke link to find the landing pad instruction
496 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
499 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
500 // Look for globals in the clauses of the landing pad instruction
501 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
502 Idx != NumClauses; ++Idx)
503 if (const GlobalVariable *GV =
504 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
505 ->stripPointerCasts()))
506 MustKeepGlobalVariables.insert(GV);
511 bool GlobalMerge::doInitialization(Module &M) {
512 if (!EnableGlobalMerge)
515 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
517 bool Changed = false;
518 setMustKeepGlobalVariables(M);
520 // Grab all non-const globals.
521 for (Module::global_iterator I = M.global_begin(),
522 E = M.global_end(); I != E; ++I) {
523 // Merge is safe for "normal" internal or external globals only
524 if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
527 if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage()) &&
528 !I->hasInternalLinkage())
531 PointerType *PT = dyn_cast<PointerType>(I->getType());
532 assert(PT && "Global variable is not a pointer!");
534 unsigned AddressSpace = PT->getAddressSpace();
536 // Ignore fancy-aligned globals for now.
537 unsigned Alignment = DL->getPreferredAlignment(I);
538 Type *Ty = I->getType()->getElementType();
539 if (Alignment > DL->getABITypeAlignment(Ty))
542 // Ignore all 'special' globals.
543 if (I->getName().startswith("llvm.") ||
544 I->getName().startswith(".llvm."))
547 // Ignore all "required" globals:
548 if (isMustKeepGlobalVariable(I))
551 if (DL->getTypeAllocSize(Ty) < MaxOffset) {
552 if (TargetLoweringObjectFile::getKindForGlobal(I, *TM).isBSSLocal())
553 BSSGlobals[AddressSpace].push_back(I);
554 else if (I->isConstant())
555 ConstGlobals[AddressSpace].push_back(I);
557 Globals[AddressSpace].push_back(I);
561 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
562 I = Globals.begin(), E = Globals.end(); I != E; ++I)
563 if (I->second.size() > 1)
564 Changed |= doMerge(I->second, M, false, I->first);
566 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
567 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
568 if (I->second.size() > 1)
569 Changed |= doMerge(I->second, M, false, I->first);
571 if (EnableGlobalMergeOnConst)
572 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
573 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
574 if (I->second.size() > 1)
575 Changed |= doMerge(I->second, M, true, I->first);
580 bool GlobalMerge::runOnFunction(Function &F) {
584 bool GlobalMerge::doFinalization(Module &M) {
585 MustKeepGlobalVariables.clear();
589 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset) {
590 return new GlobalMerge(TM, Offset);