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/IR/Attributes.h"
58 #include "llvm/IR/Constants.h"
59 #include "llvm/IR/DataLayout.h"
60 #include "llvm/IR/DerivedTypes.h"
61 #include "llvm/IR/Function.h"
62 #include "llvm/IR/GlobalVariable.h"
63 #include "llvm/IR/Instructions.h"
64 #include "llvm/IR/Intrinsics.h"
65 #include "llvm/IR/Module.h"
66 #include "llvm/Pass.h"
67 #include "llvm/Support/CommandLine.h"
68 #include "llvm/Target/TargetLowering.h"
69 #include "llvm/Target/TargetLoweringObjectFile.h"
72 #define DEBUG_TYPE "global-merge"
75 EnableGlobalMerge("enable-global-merge", cl::NotHidden,
76 cl::desc("Enable global merge pass"),
80 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
81 cl::desc("Enable global merge pass on constants"),
85 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
86 cl::desc("Enable global merge pass on external linkage"),
90 EnableGlobalMergeAligned("global-merge-aligned", cl::Hidden,
91 cl::desc("Set target specific alignment for global merge pass"),
94 STATISTIC(NumMerged , "Number of globals merged");
96 class GlobalMerge : public FunctionPass {
97 const TargetMachine *TM;
99 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
100 Module &M, bool isConst, unsigned AddrSpace) const;
102 /// \brief Check if the given variable has been identified as must keep
103 /// \pre setMustKeepGlobalVariables must have been called on the Module that
105 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
106 return MustKeepGlobalVariables.count(GV);
109 /// Collect every variables marked as "used" or used in a landing pad
110 /// instruction for this Module.
111 void setMustKeepGlobalVariables(Module &M);
113 /// Collect every variables marked as "used"
114 void collectUsedGlobalVariables(Module &M);
116 /// Keep track of the GlobalVariable that must not be merged away
117 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
120 static char ID; // Pass identification, replacement for typeid.
121 explicit GlobalMerge(const TargetMachine *TM = nullptr)
122 : FunctionPass(ID), TM(TM) {
123 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
126 bool doInitialization(Module &M) override;
127 bool runOnFunction(Function &F) override;
128 bool doFinalization(Module &M) override;
130 const char *getPassName() const override {
131 return "Merge internal globals";
134 void getAnalysisUsage(AnalysisUsage &AU) const override {
135 AU.setPreservesCFG();
136 FunctionPass::getAnalysisUsage(AU);
139 } // end anonymous namespace
141 char GlobalMerge::ID = 0;
143 static void *initializeGlobalMergePassOnce(PassRegistry &Registry) {
144 PassInfo *PI = new PassInfo(
145 "Merge global variables",
146 "global-merge", &GlobalMerge::ID,
147 PassInfo::NormalCtor_t(callDefaultCtor<GlobalMerge>), false,
148 false, PassInfo::TargetMachineCtor_t(
149 callTargetMachineCtor<GlobalMerge>));
150 Registry.registerPass(*PI, true);
154 void llvm::initializeGlobalMergePass(PassRegistry &Registry) {
155 CALL_ONCE_INITIALIZATION(initializeGlobalMergePassOnce)
158 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
159 Module &M, bool isConst, unsigned AddrSpace) const {
160 const TargetLowering *TLI = TM->getTargetLowering();
161 const DataLayout *DL = TLI->getDataLayout();
163 // FIXME: Infer the maximum possible offset depending on the actual users
164 // (these max offsets are different for the users inside Thumb or ARM
166 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
168 // FIXME: Find better heuristics
169 std::stable_sort(Globals.begin(), Globals.end(),
170 [DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
171 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
172 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
174 return (DL->getTypeAllocSize(Ty1) < DL->getTypeAllocSize(Ty2));
177 Type *Int32Ty = Type::getInt32Ty(M.getContext());
179 assert (Globals.size() > 1);
181 for (size_t i = 0, e = Globals.size(); i != e; ) {
183 uint64_t MergedSize = 0;
184 std::vector<Type*> Tys;
185 std::vector<Constant*> Inits;
187 bool HasExternal = false;
188 GlobalVariable *TheFirstExternal = 0;
189 for (j = i; j != e; ++j) {
190 Type *Ty = Globals[j]->getType()->getElementType();
191 MergedSize += DL->getTypeAllocSize(Ty);
192 if (MergedSize > MaxOffset) {
196 Inits.push_back(Globals[j]->getInitializer());
198 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
200 TheFirstExternal = Globals[j];
204 // If merged variables doesn't have external linkage, we needn't to expose
205 // the symbol after merging.
206 GlobalValue::LinkageTypes Linkage = HasExternal ?
207 GlobalValue::ExternalLinkage :
208 GlobalValue::InternalLinkage ;
210 // If merged variables have external linkage, we use symbol name of the
211 // first variable merged as the suffix of global symbol name. This would
212 // be able to avoid the link-time naming conflict for globalm symbols.
213 Twine MergedGVName = HasExternal ?
214 "_MergedGlobals_" + TheFirstExternal->getName() :
217 StructType *MergedTy = StructType::get(M.getContext(), Tys);
218 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
220 GlobalVariable *MergedGV = new GlobalVariable(M, MergedTy, isConst,
221 Linkage, MergedInit, MergedGVName,
222 nullptr, GlobalVariable::NotThreadLocal,
225 if (EnableGlobalMergeAligned) {
226 unsigned Align = TLI->getGlobalMergeAlignment(MergedTy);
227 assert(((Align % DL->getABITypeAlignment(MergedTy)) == 0) &&
228 "Specified alignment doesn't meet natural alignment requirement.");
229 MergedGV->setAlignment(Align);
232 for (size_t k = i; k < j; ++k) {
233 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
234 std::string Name = Globals[k]->getName();
237 ConstantInt::get(Int32Ty, 0),
238 ConstantInt::get(Int32Ty, k-i)
240 Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx);
241 Globals[k]->replaceAllUsesWith(GEP);
242 Globals[k]->eraseFromParent();
244 if (Linkage != GlobalValue::InternalLinkage) {
245 // Generate a new alias...
246 new GlobalAlias(GEP->getType(), Linkage, Name, GEP, &M);
257 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
258 // Extract global variables from llvm.used array
259 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
260 if (!GV || !GV->hasInitializer()) return;
262 // Should be an array of 'i8*'.
263 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
265 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
266 if (const GlobalVariable *G =
267 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
268 MustKeepGlobalVariables.insert(G);
271 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
272 collectUsedGlobalVariables(M);
274 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
276 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
277 IBB != IEndBB; ++IBB) {
278 // Follow the invoke link to find the landing pad instruction
279 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
282 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
283 // Look for globals in the clauses of the landing pad instruction
284 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
285 Idx != NumClauses; ++Idx)
286 if (const GlobalVariable *GV =
287 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
288 ->stripPointerCasts()))
289 MustKeepGlobalVariables.insert(GV);
294 bool GlobalMerge::doInitialization(Module &M) {
295 if (!EnableGlobalMerge)
298 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
300 const TargetLowering *TLI = TM->getTargetLowering();
301 const DataLayout *DL = TLI->getDataLayout();
302 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
303 bool Changed = false;
304 setMustKeepGlobalVariables(M);
306 // Grab all non-const globals.
307 for (Module::global_iterator I = M.global_begin(),
308 E = M.global_end(); I != E; ++I) {
309 // Merge is safe for "normal" internal or external globals only
310 if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
313 if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage())
314 && !I->hasInternalLinkage())
317 PointerType *PT = dyn_cast<PointerType>(I->getType());
318 assert(PT && "Global variable is not a pointer!");
320 unsigned AddressSpace = PT->getAddressSpace();
322 // Ignore fancy-aligned globals for now.
323 unsigned Alignment = DL->getPreferredAlignment(I);
324 Type *Ty = I->getType()->getElementType();
325 if (Alignment > DL->getABITypeAlignment(Ty))
328 // Ignore all 'special' globals.
329 if (I->getName().startswith("llvm.") ||
330 I->getName().startswith(".llvm."))
333 // Ignore all "required" globals:
334 if (isMustKeepGlobalVariable(I))
337 if (DL->getTypeAllocSize(Ty) < MaxOffset) {
338 if (TargetLoweringObjectFile::getKindForGlobal(I, TLI->getTargetMachine())
340 BSSGlobals[AddressSpace].push_back(I);
341 else if (I->isConstant())
342 ConstGlobals[AddressSpace].push_back(I);
344 Globals[AddressSpace].push_back(I);
348 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
349 I = Globals.begin(), E = Globals.end(); I != E; ++I)
350 if (I->second.size() > 1)
351 Changed |= doMerge(I->second, M, false, I->first);
353 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
354 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
355 if (I->second.size() > 1)
356 Changed |= doMerge(I->second, M, false, I->first);
358 if (EnableGlobalMergeOnConst)
359 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
360 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
361 if (I->second.size() > 1)
362 Changed |= doMerge(I->second, M, true, I->first);
367 bool GlobalMerge::runOnFunction(Function &F) {
371 bool GlobalMerge::doFinalization(Module &M) {
372 MustKeepGlobalVariables.clear();
376 Pass *llvm::createGlobalMergePass(const TargetMachine *TM) {
377 return new GlobalMerge(TM);