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 #define DEBUG_TYPE "global-merge"
55 #include "llvm/Transforms/Scalar.h"
56 #include "llvm/ADT/SmallPtrSet.h"
57 #include "llvm/ADT/Statistic.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"
74 EnableGlobalMerge("global-merge", cl::Hidden,
75 cl::desc("Enable global merge pass"),
79 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
80 cl::desc("Enable global merge pass on constants"),
83 STATISTIC(NumMerged , "Number of globals merged");
85 class GlobalMerge : public FunctionPass {
86 const TargetMachine *TM;
88 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
89 Module &M, bool isConst, unsigned AddrSpace) const;
91 /// \brief Check if the given variable has been identified as must keep
92 /// \pre setMustKeepGlobalVariables must have been called on the Module that
94 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
95 return MustKeepGlobalVariables.count(GV);
98 /// Collect every variables marked as "used" or used in a landing pad
99 /// instruction for this Module.
100 void setMustKeepGlobalVariables(Module &M);
102 /// Collect every variables marked as "used"
103 void collectUsedGlobalVariables(Module &M);
105 /// Keep track of the GlobalVariable that must not be merged away
106 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
109 static char ID; // Pass identification, replacement for typeid.
110 explicit GlobalMerge(const TargetMachine *TM = 0)
111 : FunctionPass(ID), TM(TM) {
112 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
115 virtual bool doInitialization(Module &M);
116 virtual bool runOnFunction(Function &F);
117 virtual bool doFinalization(Module &M);
119 const char *getPassName() const {
120 return "Merge internal globals";
123 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
124 AU.setPreservesCFG();
125 FunctionPass::getAnalysisUsage(AU);
128 } // end anonymous namespace
130 char GlobalMerge::ID = 0;
131 INITIALIZE_PASS(GlobalMerge, "global-merge",
132 "Global Merge", false, false)
135 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
136 Module &M, bool isConst, unsigned AddrSpace) const {
137 const TargetLowering *TLI = TM->getTargetLowering();
138 const DataLayout *DL = TLI->getDataLayout();
140 // FIXME: Infer the maximum possible offset depending on the actual users
141 // (these max offsets are different for the users inside Thumb or ARM
143 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
145 // FIXME: Find better heuristics
146 std::stable_sort(Globals.begin(), Globals.end(),
147 [DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
148 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
149 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
151 return (DL->getTypeAllocSize(Ty1) < DL->getTypeAllocSize(Ty2));
154 Type *Int32Ty = Type::getInt32Ty(M.getContext());
156 for (size_t i = 0, e = Globals.size(); i != e; ) {
158 uint64_t MergedSize = 0;
159 std::vector<Type*> Tys;
160 std::vector<Constant*> Inits;
161 for (j = i; j != e; ++j) {
162 Type *Ty = Globals[j]->getType()->getElementType();
163 MergedSize += DL->getTypeAllocSize(Ty);
164 if (MergedSize > MaxOffset) {
168 Inits.push_back(Globals[j]->getInitializer());
171 StructType *MergedTy = StructType::get(M.getContext(), Tys);
172 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
173 GlobalVariable *MergedGV = new GlobalVariable(M, MergedTy, isConst,
174 GlobalValue::InternalLinkage,
175 MergedInit, "_MergedGlobals",
176 0, GlobalVariable::NotThreadLocal,
178 for (size_t k = i; k < j; ++k) {
180 ConstantInt::get(Int32Ty, 0),
181 ConstantInt::get(Int32Ty, k-i)
183 Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx);
184 Globals[k]->replaceAllUsesWith(GEP);
185 Globals[k]->eraseFromParent();
194 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
195 // Extract global variables from llvm.used array
196 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
197 if (!GV || !GV->hasInitializer()) return;
199 // Should be an array of 'i8*'.
200 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
202 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
203 if (const GlobalVariable *G =
204 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
205 MustKeepGlobalVariables.insert(G);
208 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
209 collectUsedGlobalVariables(M);
211 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
213 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
214 IBB != IEndBB; ++IBB) {
215 // Follow the inwoke link to find the landing pad instruction
216 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
219 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
220 // Look for globals in the clauses of the landing pad instruction
221 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
222 Idx != NumClauses; ++Idx)
223 if (const GlobalVariable *GV =
224 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
225 ->stripPointerCasts()))
226 MustKeepGlobalVariables.insert(GV);
231 bool GlobalMerge::doInitialization(Module &M) {
232 if (!EnableGlobalMerge)
235 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
237 const TargetLowering *TLI = TM->getTargetLowering();
238 const DataLayout *DL = TLI->getDataLayout();
239 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
240 bool Changed = false;
241 setMustKeepGlobalVariables(M);
243 // Grab all non-const globals.
244 for (Module::global_iterator I = M.global_begin(),
245 E = M.global_end(); I != E; ++I) {
246 // Merge is safe for "normal" internal globals only
247 if (!I->hasLocalLinkage() || I->isThreadLocal() || I->hasSection())
250 PointerType *PT = dyn_cast<PointerType>(I->getType());
251 assert(PT && "Global variable is not a pointer!");
253 unsigned AddressSpace = PT->getAddressSpace();
255 // Ignore fancy-aligned globals for now.
256 unsigned Alignment = DL->getPreferredAlignment(I);
257 Type *Ty = I->getType()->getElementType();
258 if (Alignment > DL->getABITypeAlignment(Ty))
261 // Ignore all 'special' globals.
262 if (I->getName().startswith("llvm.") ||
263 I->getName().startswith(".llvm."))
266 // Ignore all "required" globals:
267 if (isMustKeepGlobalVariable(I))
270 if (DL->getTypeAllocSize(Ty) < MaxOffset) {
271 if (TargetLoweringObjectFile::getKindForGlobal(I, TLI->getTargetMachine())
273 BSSGlobals[AddressSpace].push_back(I);
274 else if (I->isConstant())
275 ConstGlobals[AddressSpace].push_back(I);
277 Globals[AddressSpace].push_back(I);
281 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
282 I = Globals.begin(), E = Globals.end(); I != E; ++I)
283 if (I->second.size() > 1)
284 Changed |= doMerge(I->second, M, false, I->first);
286 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
287 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
288 if (I->second.size() > 1)
289 Changed |= doMerge(I->second, M, false, I->first);
291 if (EnableGlobalMergeOnConst)
292 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
293 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
294 if (I->second.size() > 1)
295 Changed |= doMerge(I->second, M, true, I->first);
300 bool GlobalMerge::runOnFunction(Function &F) {
304 bool GlobalMerge::doFinalization(Module &M) {
305 MustKeepGlobalVariables.clear();
309 Pass *llvm::createGlobalMergePass(const TargetMachine *TM) {
310 return new GlobalMerge(TM);