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/ADT/DenseMap.h"
63 #include "llvm/ADT/SmallBitVector.h"
64 #include "llvm/ADT/SmallPtrSet.h"
65 #include "llvm/ADT/Statistic.h"
66 #include "llvm/CodeGen/Passes.h"
67 #include "llvm/IR/Attributes.h"
68 #include "llvm/IR/Constants.h"
69 #include "llvm/IR/DataLayout.h"
70 #include "llvm/IR/DerivedTypes.h"
71 #include "llvm/IR/Function.h"
72 #include "llvm/IR/GlobalVariable.h"
73 #include "llvm/IR/Instructions.h"
74 #include "llvm/IR/Intrinsics.h"
75 #include "llvm/IR/Module.h"
76 #include "llvm/Pass.h"
77 #include "llvm/Support/CommandLine.h"
78 #include "llvm/Support/Debug.h"
79 #include "llvm/Support/raw_ostream.h"
80 #include "llvm/Target/TargetLowering.h"
81 #include "llvm/Target/TargetLoweringObjectFile.h"
82 #include "llvm/Target/TargetSubtargetInfo.h"
86 #define DEBUG_TYPE "global-merge"
88 // FIXME: This is only useful as a last-resort way to disable the pass.
90 EnableGlobalMerge("enable-global-merge", cl::Hidden,
91 cl::desc("Enable the global merge pass"),
94 static cl::opt<unsigned>
95 GlobalMergeMaxOffset("global-merge-max-offset", cl::Hidden,
96 cl::desc("Set maximum offset for global merge pass"),
99 static cl::opt<bool> GlobalMergeGroupByUse(
100 "global-merge-group-by-use", cl::Hidden,
101 cl::desc("Improve global merge pass to look at uses"), cl::init(true));
103 static cl::opt<bool> GlobalMergeIgnoreSingleUse(
104 "global-merge-ignore-single-use", cl::Hidden,
105 cl::desc("Improve global merge pass to ignore globals only used alone"),
109 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
110 cl::desc("Enable global merge pass on constants"),
113 // FIXME: this could be a transitional option, and we probably need to remove
114 // it if only we are sure this optimization could always benefit all targets.
115 static cl::opt<cl::boolOrDefault>
116 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
117 cl::desc("Enable global merge pass on external linkage"));
119 STATISTIC(NumMerged, "Number of globals merged");
121 class GlobalMerge : public FunctionPass {
122 const TargetMachine *TM;
123 // FIXME: Infer the maximum possible offset depending on the actual users
124 // (these max offsets are different for the users inside Thumb or ARM
125 // functions), see the code that passes in the offset in the ARM backend
126 // for more information.
129 /// Whether we should try to optimize for size only.
130 /// Currently, this applies a dead simple heuristic: only consider globals
131 /// used in minsize functions for merging.
132 /// FIXME: This could learn about optsize, and be used in the cost model.
133 bool OnlyOptimizeForSize;
135 /// Whether we should merge global variables that have external linkage.
136 bool MergeExternalGlobals;
140 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
141 Module &M, bool isConst, unsigned AddrSpace) const;
142 /// \brief Merge everything in \p Globals for which the corresponding bit
143 /// in \p GlobalSet is set.
144 bool doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
145 const BitVector &GlobalSet, Module &M, bool isConst,
146 unsigned AddrSpace) const;
148 /// \brief Check if the given variable has been identified as must keep
149 /// \pre setMustKeepGlobalVariables must have been called on the Module that
151 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
152 return MustKeepGlobalVariables.count(GV);
155 /// Collect every variables marked as "used" or used in a landing pad
156 /// instruction for this Module.
157 void setMustKeepGlobalVariables(Module &M);
159 /// Collect every variables marked as "used"
160 void collectUsedGlobalVariables(Module &M);
162 /// Keep track of the GlobalVariable that must not be merged away
163 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
166 static char ID; // Pass identification, replacement for typeid.
167 explicit GlobalMerge()
168 : FunctionPass(ID), TM(nullptr), MaxOffset(GlobalMergeMaxOffset),
169 OnlyOptimizeForSize(false), MergeExternalGlobals(false) {
170 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
173 explicit GlobalMerge(const TargetMachine *TM, unsigned MaximalOffset,
174 bool OnlyOptimizeForSize, bool MergeExternalGlobals)
175 : FunctionPass(ID), TM(TM), MaxOffset(MaximalOffset),
176 OnlyOptimizeForSize(OnlyOptimizeForSize),
177 MergeExternalGlobals(MergeExternalGlobals) {
178 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
181 bool doInitialization(Module &M) override;
182 bool runOnFunction(Function &F) override;
183 bool doFinalization(Module &M) override;
185 StringRef getPassName() const override { return "Merge internal globals"; }
187 void getAnalysisUsage(AnalysisUsage &AU) const override {
188 AU.setPreservesCFG();
189 FunctionPass::getAnalysisUsage(AU);
192 } // end anonymous namespace
194 char GlobalMerge::ID = 0;
195 INITIALIZE_PASS(GlobalMerge, DEBUG_TYPE, "Merge global variables", false, false)
197 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
198 Module &M, bool isConst, unsigned AddrSpace) const {
199 auto &DL = M.getDataLayout();
200 // FIXME: Find better heuristics
201 std::stable_sort(Globals.begin(), Globals.end(),
202 [&DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
203 return DL.getTypeAllocSize(GV1->getValueType()) <
204 DL.getTypeAllocSize(GV2->getValueType());
207 // If we want to just blindly group all globals together, do so.
208 if (!GlobalMergeGroupByUse) {
209 BitVector AllGlobals(Globals.size());
211 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
214 // If we want to be smarter, look at all uses of each global, to try to
215 // discover all sets of globals used together, and how many times each of
216 // these sets occurred.
218 // Keep this reasonably efficient, by having an append-only list of all sets
219 // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of
220 // code (currently, a Function) to the set of globals seen so far that are
221 // used together in that unit (GlobalUsesByFunction).
223 // When we look at the Nth global, we now that any new set is either:
224 // - the singleton set {N}, containing this global only, or
225 // - the union of {N} and a previously-discovered set, containing some
226 // combination of the previous N-1 globals.
227 // Using that knowledge, when looking at the Nth global, we can keep:
228 // - a reference to the singleton set {N} (CurGVOnlySetIdx)
229 // - a list mapping each previous set to its union with {N} (EncounteredUGS),
230 // if it actually occurs.
232 // We keep track of the sets of globals used together "close enough".
233 struct UsedGlobalSet {
234 UsedGlobalSet(size_t Size) : Globals(Size), UsageCount(1) {}
239 // Each set is unique in UsedGlobalSets.
240 std::vector<UsedGlobalSet> UsedGlobalSets;
242 // Avoid repeating the create-global-set pattern.
243 auto CreateGlobalSet = [&]() -> UsedGlobalSet & {
244 UsedGlobalSets.emplace_back(Globals.size());
245 return UsedGlobalSets.back();
248 // The first set is the empty set.
249 CreateGlobalSet().UsageCount = 0;
251 // We define "close enough" to be "in the same function".
252 // FIXME: Grouping uses by function is way too aggressive, so we should have
253 // a better metric for distance between uses.
254 // The obvious alternative would be to group by BasicBlock, but that's in
255 // turn too conservative..
256 // Anything in between wouldn't be trivial to compute, so just stick with
257 // per-function grouping.
259 // The value type is an index into UsedGlobalSets.
260 // The default (0) conveniently points to the empty set.
261 DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction;
263 // Now, look at each merge-eligible global in turn.
265 // Keep track of the sets we already encountered to which we added the
267 // Each element matches the same-index element in UsedGlobalSets.
268 // This lets us efficiently tell whether a set has already been expanded to
269 // include the current global.
270 std::vector<size_t> EncounteredUGS;
272 for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) {
273 GlobalVariable *GV = Globals[GI];
275 // Reset the encountered sets for this global...
276 std::fill(EncounteredUGS.begin(), EncounteredUGS.end(), 0);
277 // ...and grow it in case we created new sets for the previous global.
278 EncounteredUGS.resize(UsedGlobalSets.size());
280 // We might need to create a set that only consists of the current global.
281 // Keep track of its index into UsedGlobalSets.
282 size_t CurGVOnlySetIdx = 0;
284 // For each global, look at all its Uses.
285 for (auto &U : GV->uses()) {
286 // This Use might be a ConstantExpr. We're interested in Instruction
287 // users, so look through ConstantExpr...
289 if (ConstantExpr *CE = dyn_cast<ConstantExpr>(U.getUser())) {
292 UI = &*CE->use_begin();
294 } else if (isa<Instruction>(U.getUser())) {
301 // ...to iterate on all the instruction users of the global.
302 // Note that we iterate on Uses and not on Users to be able to getNext().
303 for (; UI != UE; UI = UI->getNext()) {
304 Instruction *I = dyn_cast<Instruction>(UI->getUser());
308 Function *ParentFn = I->getParent()->getParent();
310 // If we're only optimizing for size, ignore non-minsize functions.
311 if (OnlyOptimizeForSize && !ParentFn->optForMinSize())
314 size_t UGSIdx = GlobalUsesByFunction[ParentFn];
316 // If this is the first global the basic block uses, map it to the set
317 // consisting of this global only.
319 // If that set doesn't exist yet, create it.
320 if (!CurGVOnlySetIdx) {
321 CurGVOnlySetIdx = UsedGlobalSets.size();
322 CreateGlobalSet().Globals.set(GI);
324 ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount;
327 GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx;
331 // If we already encountered this BB, just increment the counter.
332 if (UsedGlobalSets[UGSIdx].Globals.test(GI)) {
333 ++UsedGlobalSets[UGSIdx].UsageCount;
337 // If not, the previous set wasn't actually used in this function.
338 --UsedGlobalSets[UGSIdx].UsageCount;
340 // If we already expanded the previous set to include this global, just
341 // reuse that expanded set.
342 if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) {
343 ++UsedGlobalSets[ExpandedIdx].UsageCount;
344 GlobalUsesByFunction[ParentFn] = ExpandedIdx;
348 // If not, create a new set consisting of the union of the previous set
349 // and this global. Mark it as encountered, so we can reuse it later.
350 GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] =
351 UsedGlobalSets.size();
353 UsedGlobalSet &NewUGS = CreateGlobalSet();
354 NewUGS.Globals.set(GI);
355 NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals;
360 // Now we found a bunch of sets of globals used together. We accumulated
361 // the number of times we encountered the sets (i.e., the number of blocks
362 // that use that exact set of globals).
364 // Multiply that by the size of the set to give us a crude profitability
366 std::sort(UsedGlobalSets.begin(), UsedGlobalSets.end(),
367 [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) {
368 return UGS1.Globals.count() * UGS1.UsageCount <
369 UGS2.Globals.count() * UGS2.UsageCount;
372 // We can choose to merge all globals together, but ignore globals never used
373 // with another global. This catches the obviously non-profitable cases of
374 // having a single global, but is aggressive enough for any other case.
375 if (GlobalMergeIgnoreSingleUse) {
376 BitVector AllGlobals(Globals.size());
377 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
378 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
379 if (UGS.UsageCount == 0)
381 if (UGS.Globals.count() > 1)
382 AllGlobals |= UGS.Globals;
384 return doMerge(Globals, AllGlobals, M, isConst, AddrSpace);
387 // Starting from the sets with the best (=biggest) profitability, find a
389 // The ideal (and expensive) solution can only be found by trying all
390 // combinations, looking for the one with the best profitability.
391 // Don't be smart about it, and just pick the first compatible combination,
392 // starting with the sets with the best profitability.
393 BitVector PickedGlobals(Globals.size());
394 bool Changed = false;
396 for (size_t i = 0, e = UsedGlobalSets.size(); i != e; ++i) {
397 const UsedGlobalSet &UGS = UsedGlobalSets[e - i - 1];
398 if (UGS.UsageCount == 0)
400 if (PickedGlobals.anyCommon(UGS.Globals))
402 PickedGlobals |= UGS.Globals;
403 // If the set only contains one global, there's no point in merging.
404 // Ignore the global for inclusion in other sets though, so keep it in
406 if (UGS.Globals.count() < 2)
408 Changed |= doMerge(Globals, UGS.Globals, M, isConst, AddrSpace);
414 bool GlobalMerge::doMerge(const SmallVectorImpl<GlobalVariable *> &Globals,
415 const BitVector &GlobalSet, Module &M, bool isConst,
416 unsigned AddrSpace) const {
417 assert(Globals.size() > 1);
419 Type *Int32Ty = Type::getInt32Ty(M.getContext());
420 auto &DL = M.getDataLayout();
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 bool HasExternal = false;
433 StringRef FirstExternalName;
434 for (j = i; j != -1; j = GlobalSet.find_next(j)) {
435 Type *Ty = Globals[j]->getValueType();
436 MergedSize += DL.getTypeAllocSize(Ty);
437 if (MergedSize > MaxOffset) {
441 Inits.push_back(Globals[j]->getInitializer());
443 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
445 FirstExternalName = Globals[j]->getName();
449 // If merged variables doesn't have external linkage, we needn't to expose
450 // the symbol after merging.
451 GlobalValue::LinkageTypes Linkage = HasExternal
452 ? GlobalValue::ExternalLinkage
453 : GlobalValue::InternalLinkage;
454 StructType *MergedTy = StructType::get(M.getContext(), Tys);
455 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
457 // On Darwin external linkage needs to be preserved, otherwise
458 // dsymutil cannot preserve the debug info for the merged
459 // variables. If they have external linkage, use the symbol name
460 // of the first variable merged as the suffix of global symbol
461 // name. This avoids a link-time naming conflict for the
462 // _MergedGlobals symbols.
464 (IsMachO && HasExternal)
465 ? "_MergedGlobals_" + FirstExternalName
467 auto MergedLinkage = IsMachO ? Linkage : GlobalValue::PrivateLinkage;
468 auto *MergedGV = new GlobalVariable(
469 M, MergedTy, isConst, MergedLinkage, MergedInit, MergedName, nullptr,
470 GlobalVariable::NotThreadLocal, AddrSpace);
472 const StructLayout *MergedLayout = DL.getStructLayout(MergedTy);
474 for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(k), ++idx) {
475 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
476 std::string Name = Globals[k]->getName();
478 // Copy metadata while adjusting any debug info metadata by the original
479 // global's offset within the merged global.
480 MergedGV->copyMetadata(Globals[k], MergedLayout->getElementOffset(idx));
483 ConstantInt::get(Int32Ty, 0),
484 ConstantInt::get(Int32Ty, idx),
487 ConstantExpr::getInBoundsGetElementPtr(MergedTy, MergedGV, Idx);
488 Globals[k]->replaceAllUsesWith(GEP);
489 Globals[k]->eraseFromParent();
491 // When the linkage is not internal we must emit an alias for the original
492 // variable name as it may be accessed from another object. On non-Mach-O
493 // we can also emit an alias for internal linkage as it's safe to do so.
494 // It's not safe on Mach-O as the alias (and thus the portion of the
495 // MergedGlobals variable) may be dead stripped at link time.
496 if (Linkage != GlobalValue::InternalLinkage || !IsMachO) {
497 GlobalAlias::create(Tys[idx], AddrSpace, Linkage, Name, GEP, &M);
508 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
509 // Extract global variables from llvm.used array
510 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
511 if (!GV || !GV->hasInitializer()) return;
513 // Should be an array of 'i8*'.
514 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
516 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
517 if (const GlobalVariable *G =
518 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
519 MustKeepGlobalVariables.insert(G);
522 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
523 collectUsedGlobalVariables(M);
525 for (Function &F : M) {
526 for (BasicBlock &BB : F) {
527 Instruction *Pad = BB.getFirstNonPHI();
531 // Keep globals used by landingpads and catchpads.
532 for (const Use &U : Pad->operands()) {
533 if (const GlobalVariable *GV =
534 dyn_cast<GlobalVariable>(U->stripPointerCasts()))
535 MustKeepGlobalVariables.insert(GV);
541 bool GlobalMerge::doInitialization(Module &M) {
542 if (!EnableGlobalMerge)
545 IsMachO = Triple(M.getTargetTriple()).isOSBinFormatMachO();
547 auto &DL = M.getDataLayout();
548 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
550 bool Changed = false;
551 setMustKeepGlobalVariables(M);
553 // Grab all non-const globals.
554 for (auto &GV : M.globals()) {
555 // Merge is safe for "normal" internal or external globals only
556 if (GV.isDeclaration() || GV.isThreadLocal() ||
557 GV.hasSection() || GV.hasImplicitSection())
560 // It's not safe to merge globals that may be preempted
561 if (TM && !TM->shouldAssumeDSOLocal(M, &GV))
564 if (!(MergeExternalGlobals && GV.hasExternalLinkage()) &&
565 !GV.hasInternalLinkage())
568 PointerType *PT = dyn_cast<PointerType>(GV.getType());
569 assert(PT && "Global variable is not a pointer!");
571 unsigned AddressSpace = PT->getAddressSpace();
573 // Ignore fancy-aligned globals for now.
574 unsigned Alignment = DL.getPreferredAlignment(&GV);
575 Type *Ty = GV.getValueType();
576 if (Alignment > DL.getABITypeAlignment(Ty))
579 // Ignore all 'special' globals.
580 if (GV.getName().startswith("llvm.") ||
581 GV.getName().startswith(".llvm."))
584 // Ignore all "required" globals:
585 if (isMustKeepGlobalVariable(&GV))
588 if (DL.getTypeAllocSize(Ty) < MaxOffset) {
590 TargetLoweringObjectFile::getKindForGlobal(&GV, *TM).isBSSLocal())
591 BSSGlobals[AddressSpace].push_back(&GV);
592 else if (GV.isConstant())
593 ConstGlobals[AddressSpace].push_back(&GV);
595 Globals[AddressSpace].push_back(&GV);
599 for (auto &P : Globals)
600 if (P.second.size() > 1)
601 Changed |= doMerge(P.second, M, false, P.first);
603 for (auto &P : BSSGlobals)
604 if (P.second.size() > 1)
605 Changed |= doMerge(P.second, M, false, P.first);
607 if (EnableGlobalMergeOnConst)
608 for (auto &P : ConstGlobals)
609 if (P.second.size() > 1)
610 Changed |= doMerge(P.second, M, true, P.first);
615 bool GlobalMerge::runOnFunction(Function &F) {
619 bool GlobalMerge::doFinalization(Module &M) {
620 MustKeepGlobalVariables.clear();
624 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset,
625 bool OnlyOptimizeForSize,
626 bool MergeExternalByDefault) {
627 bool MergeExternal = (EnableGlobalMergeOnExternal == cl::BOU_UNSET) ?
628 MergeExternalByDefault : (EnableGlobalMergeOnExternal == cl::BOU_TRUE);
629 return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal);