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".
52 // ===---------------------------------------------------------------------===//
54 #include "llvm/Transforms/Scalar.h"
55 #include "llvm/ADT/SmallPtrSet.h"
56 #include "llvm/ADT/Statistic.h"
57 #include "llvm/CodeGen/Passes.h"
58 #include "llvm/IR/Attributes.h"
59 #include "llvm/IR/Constants.h"
60 #include "llvm/IR/DataLayout.h"
61 #include "llvm/IR/DerivedTypes.h"
62 #include "llvm/IR/Function.h"
63 #include "llvm/IR/GlobalVariable.h"
64 #include "llvm/IR/Instructions.h"
65 #include "llvm/IR/Intrinsics.h"
66 #include "llvm/IR/Module.h"
67 #include "llvm/Pass.h"
68 #include "llvm/Support/CommandLine.h"
69 #include "llvm/Target/TargetLowering.h"
70 #include "llvm/Target/TargetLoweringObjectFile.h"
71 #include "llvm/Target/TargetSubtargetInfo.h"
74 #define DEBUG_TYPE "global-merge"
77 EnableGlobalMerge("enable-global-merge", cl::Hidden,
78 cl::desc("Enable global merge pass"),
82 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
83 cl::desc("Enable global merge pass on constants"),
86 // FIXME: this could be a transitional option, and we probably need to remove
87 // it if only we are sure this optimization could always benefit all targets.
89 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
90 cl::desc("Enable global merge pass on external linkage"),
93 STATISTIC(NumMerged, "Number of globals merged");
95 class GlobalMerge : public FunctionPass {
96 const TargetMachine *TM;
98 // FIXME: Infer the maximum possible offset depending on the actual users
99 // (these max offsets are different for the users inside Thumb or ARM
100 // functions), see the code that passes in the offset in the ARM backend
101 // for more information.
104 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
105 Module &M, bool isConst, unsigned AddrSpace) const;
107 /// \brief Check if the given variable has been identified as must keep
108 /// \pre setMustKeepGlobalVariables must have been called on the Module that
110 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
111 return MustKeepGlobalVariables.count(GV);
114 /// Collect every variables marked as "used" or used in a landing pad
115 /// instruction for this Module.
116 void setMustKeepGlobalVariables(Module &M);
118 /// Collect every variables marked as "used"
119 void collectUsedGlobalVariables(Module &M);
121 /// Keep track of the GlobalVariable that must not be merged away
122 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
125 static char ID; // Pass identification, replacement for typeid.
126 explicit GlobalMerge(const TargetMachine *TM = nullptr,
127 unsigned MaximalOffset = 0)
128 : FunctionPass(ID), TM(TM), DL(TM->getDataLayout()),
129 MaxOffset(MaximalOffset) {
130 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
133 bool doInitialization(Module &M) override;
134 bool runOnFunction(Function &F) override;
135 bool doFinalization(Module &M) override;
137 const char *getPassName() const override {
138 return "Merge internal globals";
141 void getAnalysisUsage(AnalysisUsage &AU) const override {
142 AU.setPreservesCFG();
143 FunctionPass::getAnalysisUsage(AU);
146 } // end anonymous namespace
148 char GlobalMerge::ID = 0;
149 INITIALIZE_PASS_BEGIN(GlobalMerge, "global-merge", "Merge global variables",
151 INITIALIZE_PASS_END(GlobalMerge, "global-merge", "Merge global variables",
154 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
155 Module &M, bool isConst, unsigned AddrSpace) const {
156 // FIXME: Find better heuristics
157 std::stable_sort(Globals.begin(), Globals.end(),
158 [this](const GlobalVariable *GV1, const GlobalVariable *GV2) {
159 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
160 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
162 return (DL->getTypeAllocSize(Ty1) < DL->getTypeAllocSize(Ty2));
165 Type *Int32Ty = Type::getInt32Ty(M.getContext());
167 assert(Globals.size() > 1);
169 // FIXME: This simple solution merges globals all together as maximum as
170 // possible. However, with this solution it would be hard to remove dead
171 // global symbols at link-time. An alternative solution could be checking
172 // global symbols references function by function, and make the symbols
173 // being referred in the same function merged and we would probably need
174 // to introduce heuristic algorithm to solve the merge conflict from
175 // different functions.
176 for (size_t i = 0, e = Globals.size(); i != e; ) {
178 uint64_t MergedSize = 0;
179 std::vector<Type*> Tys;
180 std::vector<Constant*> Inits;
182 bool HasExternal = false;
183 GlobalVariable *TheFirstExternal = 0;
184 for (j = i; j != e; ++j) {
185 Type *Ty = Globals[j]->getType()->getElementType();
186 MergedSize += DL->getTypeAllocSize(Ty);
187 if (MergedSize > MaxOffset) {
191 Inits.push_back(Globals[j]->getInitializer());
193 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
195 TheFirstExternal = Globals[j];
199 // If merged variables doesn't have external linkage, we needn't to expose
200 // the symbol after merging.
201 GlobalValue::LinkageTypes Linkage = HasExternal
202 ? GlobalValue::ExternalLinkage
203 : GlobalValue::InternalLinkage;
205 StructType *MergedTy = StructType::get(M.getContext(), Tys);
206 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
208 // If merged variables have external linkage, we use symbol name of the
209 // first variable merged as the suffix of global symbol name. This would
210 // be able to avoid the link-time naming conflict for globalm symbols.
211 GlobalVariable *MergedGV = new GlobalVariable(
212 M, MergedTy, isConst, Linkage, MergedInit,
213 HasExternal ? "_MergedGlobals_" + TheFirstExternal->getName()
215 nullptr, GlobalVariable::NotThreadLocal, AddrSpace);
217 for (size_t k = i; k < j; ++k) {
218 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
219 std::string Name = Globals[k]->getName();
222 ConstantInt::get(Int32Ty, 0),
223 ConstantInt::get(Int32Ty, k-i)
225 Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx);
226 Globals[k]->replaceAllUsesWith(GEP);
227 Globals[k]->eraseFromParent();
229 if (Linkage != GlobalValue::InternalLinkage) {
230 // Generate a new alias...
231 auto *PTy = cast<PointerType>(GEP->getType());
232 GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
233 Linkage, Name, GEP, &M);
244 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
245 // Extract global variables from llvm.used array
246 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
247 if (!GV || !GV->hasInitializer()) return;
249 // Should be an array of 'i8*'.
250 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
252 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
253 if (const GlobalVariable *G =
254 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
255 MustKeepGlobalVariables.insert(G);
258 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
259 collectUsedGlobalVariables(M);
261 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
263 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
264 IBB != IEndBB; ++IBB) {
265 // Follow the invoke link to find the landing pad instruction
266 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
269 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
270 // Look for globals in the clauses of the landing pad instruction
271 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
272 Idx != NumClauses; ++Idx)
273 if (const GlobalVariable *GV =
274 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
275 ->stripPointerCasts()))
276 MustKeepGlobalVariables.insert(GV);
281 bool GlobalMerge::doInitialization(Module &M) {
282 if (!EnableGlobalMerge)
285 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
287 bool Changed = false;
288 setMustKeepGlobalVariables(M);
290 // Grab all non-const globals.
291 for (Module::global_iterator I = M.global_begin(),
292 E = M.global_end(); I != E; ++I) {
293 // Merge is safe for "normal" internal or external globals only
294 if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
297 if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage()) &&
298 !I->hasInternalLinkage())
301 PointerType *PT = dyn_cast<PointerType>(I->getType());
302 assert(PT && "Global variable is not a pointer!");
304 unsigned AddressSpace = PT->getAddressSpace();
306 // Ignore fancy-aligned globals for now.
307 unsigned Alignment = DL->getPreferredAlignment(I);
308 Type *Ty = I->getType()->getElementType();
309 if (Alignment > DL->getABITypeAlignment(Ty))
312 // Ignore all 'special' globals.
313 if (I->getName().startswith("llvm.") ||
314 I->getName().startswith(".llvm."))
317 // Ignore all "required" globals:
318 if (isMustKeepGlobalVariable(I))
321 if (DL->getTypeAllocSize(Ty) < MaxOffset) {
322 if (TargetLoweringObjectFile::getKindForGlobal(I, *TM).isBSSLocal())
323 BSSGlobals[AddressSpace].push_back(I);
324 else if (I->isConstant())
325 ConstGlobals[AddressSpace].push_back(I);
327 Globals[AddressSpace].push_back(I);
331 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
332 I = Globals.begin(), E = Globals.end(); I != E; ++I)
333 if (I->second.size() > 1)
334 Changed |= doMerge(I->second, M, false, I->first);
336 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
337 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
338 if (I->second.size() > 1)
339 Changed |= doMerge(I->second, M, false, I->first);
341 if (EnableGlobalMergeOnConst)
342 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
343 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
344 if (I->second.size() > 1)
345 Changed |= doMerge(I->second, M, true, I->first);
350 bool GlobalMerge::runOnFunction(Function &F) {
354 bool GlobalMerge::doFinalization(Module &M) {
355 MustKeepGlobalVariables.clear();
359 Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset) {
360 return new GlobalMerge(TM, Offset);