1 //===-- GlobalMerge.cpp - Internal globals merging -----------------------===//
2 //
3 // The LLVM Compiler Infrastructure
4 //
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
7 //
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.
14 //
15 // For example, consider the code which touches several global variables at
16 // once:
17 //
18 // static int foo[N], bar[N], baz[N];
19 //
20 // for (i = 0; i < N; ++i) {
21 // foo[i] = bar[i] * baz[i];
22 // }
23 //
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):
26 //
27 // ldr r1, [r5], #4
28 // ldr r2, [r6], #4
29 // mul r1, r2, r1
30 // str r1, [r0], #4
31 //
32 // Pass converts the code to something like:
33 //
34 // static struct {
35 // int foo[N];
36 // int bar[N];
37 // int baz[N];
38 // } merged;
39 //
40 // for (i = 0; i < N; ++i) {
41 // merged.foo[i] = merged.bar[i] * merged.baz[i];
42 // }
43 //
44 // and in ARM code this becomes:
45 //
46 // ldr r0, [r5, #40]
47 // ldr r1, [r5, #80]
48 // mul r0, r1, r0
49 // str r0, [r5], #4
50 //
51 // note that we saved 2 registers here almostly "for free".
52 // ===---------------------------------------------------------------------===//
53
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/CodeGen/Passes.h"
68 #include "llvm/Support/CommandLine.h"
69 #include "llvm/Target/TargetLowering.h"
70 #include "llvm/Target/TargetLoweringObjectFile.h"
71 using namespace llvm;
72
73 #define DEBUG_TYPE "global-merge"
74
75 cl::opt<bool>
76 EnableGlobalMerge("enable-global-merge", cl::Hidden,
77 cl::desc("Enable global merge pass"),
78 cl::init(true));
79
80 static cl::opt<bool>
81 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
82 cl::desc("Enable global merge pass on constants"),
83 cl::init(false));
84
85 // FIXME: this could be a transitional option, and we probably need to remove
86 // it if only we are sure this optimization could always benefit all targets.
87 static cl::opt<bool>
88 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
89 cl::desc("Enable global merge pass on external linkage"),
90 cl::init(false));
91
92 STATISTIC(NumMerged , "Number of globals merged");
93 namespace {
94 class GlobalMerge : public FunctionPass {
95 const TargetMachine *TM;
96
97 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
98 Module &M, bool isConst, unsigned AddrSpace) const;
99
100 /// \brief Check if the given variable has been identified as must keep
101 /// \pre setMustKeepGlobalVariables must have been called on the Module that
102 /// contains GV
isMustKeepGlobalVariable(const GlobalVariable * GV) const103 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
104 return MustKeepGlobalVariables.count(GV);
105 }
106
107 /// Collect every variables marked as "used" or used in a landing pad
108 /// instruction for this Module.
109 void setMustKeepGlobalVariables(Module &M);
110
111 /// Collect every variables marked as "used"
112 void collectUsedGlobalVariables(Module &M);
113
114 /// Keep track of the GlobalVariable that must not be merged away
115 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
116
117 public:
118 static char ID; // Pass identification, replacement for typeid.
GlobalMerge(const TargetMachine * TM=nullptr)119 explicit GlobalMerge(const TargetMachine *TM = nullptr)
120 : FunctionPass(ID), TM(TM) {
121 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
122 }
123
124 bool doInitialization(Module &M) override;
125 bool runOnFunction(Function &F) override;
126 bool doFinalization(Module &M) override;
127
getPassName() const128 const char *getPassName() const override {
129 return "Merge internal globals";
130 }
131
getAnalysisUsage(AnalysisUsage & AU) const132 void getAnalysisUsage(AnalysisUsage &AU) const override {
133 AU.setPreservesCFG();
134 FunctionPass::getAnalysisUsage(AU);
135 }
136 };
137 } // end anonymous namespace
138
139 char GlobalMerge::ID = 0;
140 INITIALIZE_TM_PASS(GlobalMerge, "global-merge", "Merge global variables",
141 false, false)
142
doMerge(SmallVectorImpl<GlobalVariable * > & Globals,Module & M,bool isConst,unsigned AddrSpace) const143 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
144 Module &M, bool isConst, unsigned AddrSpace) const {
145 const TargetLowering *TLI = TM->getTargetLowering();
146 const DataLayout *DL = TLI->getDataLayout();
147
148 // FIXME: Infer the maximum possible offset depending on the actual users
149 // (these max offsets are different for the users inside Thumb or ARM
150 // functions)
151 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
152
153 // FIXME: Find better heuristics
154 std::stable_sort(Globals.begin(), Globals.end(),
155 [DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
156 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
157 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
158
159 return (DL->getTypeAllocSize(Ty1) < DL->getTypeAllocSize(Ty2));
160 });
161
162 Type *Int32Ty = Type::getInt32Ty(M.getContext());
163
164 assert(Globals.size() > 1);
165
166 // FIXME: This simple solution merges globals all together as maximum as
167 // possible. However, with this solution it would be hard to remove dead
168 // global symbols at link-time. An alternative solution could be checking
169 // global symbols references function by function, and make the symbols
170 // being referred in the same function merged and we would probably need
171 // to introduce heuristic algorithm to solve the merge conflict from
172 // different functions.
173 for (size_t i = 0, e = Globals.size(); i != e; ) {
174 size_t j = 0;
175 uint64_t MergedSize = 0;
176 std::vector<Type*> Tys;
177 std::vector<Constant*> Inits;
178
179 bool HasExternal = false;
180 GlobalVariable *TheFirstExternal = 0;
181 for (j = i; j != e; ++j) {
182 Type *Ty = Globals[j]->getType()->getElementType();
183 MergedSize += DL->getTypeAllocSize(Ty);
184 if (MergedSize > MaxOffset) {
185 break;
186 }
187 Tys.push_back(Ty);
188 Inits.push_back(Globals[j]->getInitializer());
189
190 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
191 HasExternal = true;
192 TheFirstExternal = Globals[j];
193 }
194 }
195
196 // If merged variables doesn't have external linkage, we needn't to expose
197 // the symbol after merging.
198 GlobalValue::LinkageTypes Linkage = HasExternal
199 ? GlobalValue::ExternalLinkage
200 : GlobalValue::InternalLinkage;
201
202 StructType *MergedTy = StructType::get(M.getContext(), Tys);
203 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
204
205 // If merged variables have external linkage, we use symbol name of the
206 // first variable merged as the suffix of global symbol name. This would
207 // be able to avoid the link-time naming conflict for globalm symbols.
208 GlobalVariable *MergedGV = new GlobalVariable(
209 M, MergedTy, isConst, Linkage, MergedInit,
210 HasExternal ? "_MergedGlobals_" + TheFirstExternal->getName()
211 : "_MergedGlobals",
212 nullptr, GlobalVariable::NotThreadLocal, AddrSpace);
213
214 for (size_t k = i; k < j; ++k) {
215 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
216 std::string Name = Globals[k]->getName();
217
218 Constant *Idx[2] = {
219 ConstantInt::get(Int32Ty, 0),
220 ConstantInt::get(Int32Ty, k-i)
221 };
222 Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx);
223 Globals[k]->replaceAllUsesWith(GEP);
224 Globals[k]->eraseFromParent();
225
226 if (Linkage != GlobalValue::InternalLinkage) {
227 // Generate a new alias...
228 auto *PTy = cast<PointerType>(GEP->getType());
229 GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
230 Linkage, Name, GEP, &M);
231 }
232
233 NumMerged++;
234 }
235 i = j;
236 }
237
238 return true;
239 }
240
collectUsedGlobalVariables(Module & M)241 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
242 // Extract global variables from llvm.used array
243 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
244 if (!GV || !GV->hasInitializer()) return;
245
246 // Should be an array of 'i8*'.
247 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
248
249 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
250 if (const GlobalVariable *G =
251 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
252 MustKeepGlobalVariables.insert(G);
253 }
254
setMustKeepGlobalVariables(Module & M)255 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
256 collectUsedGlobalVariables(M);
257
258 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
259 ++IFn) {
260 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
261 IBB != IEndBB; ++IBB) {
262 // Follow the invoke link to find the landing pad instruction
263 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
264 if (!II) continue;
265
266 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
267 // Look for globals in the clauses of the landing pad instruction
268 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
269 Idx != NumClauses; ++Idx)
270 if (const GlobalVariable *GV =
271 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
272 ->stripPointerCasts()))
273 MustKeepGlobalVariables.insert(GV);
274 }
275 }
276 }
277
doInitialization(Module & M)278 bool GlobalMerge::doInitialization(Module &M) {
279 if (!EnableGlobalMerge)
280 return false;
281
282 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
283 BSSGlobals;
284 const TargetLowering *TLI = TM->getTargetLowering();
285 const DataLayout *DL = TLI->getDataLayout();
286 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
287 bool Changed = false;
288 setMustKeepGlobalVariables(M);
289
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())
295 continue;
296
297 if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage()) &&
298 !I->hasInternalLinkage())
299 continue;
300
301 PointerType *PT = dyn_cast<PointerType>(I->getType());
302 assert(PT && "Global variable is not a pointer!");
303
304 unsigned AddressSpace = PT->getAddressSpace();
305
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))
310 continue;
311
312 // Ignore all 'special' globals.
313 if (I->getName().startswith("llvm.") ||
314 I->getName().startswith(".llvm."))
315 continue;
316
317 // Ignore all "required" globals:
318 if (isMustKeepGlobalVariable(I))
319 continue;
320
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);
326 else
327 Globals[AddressSpace].push_back(I);
328 }
329 }
330
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);
335
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);
340
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);
346
347 return Changed;
348 }
349
runOnFunction(Function & F)350 bool GlobalMerge::runOnFunction(Function &F) {
351 return false;
352 }
353
doFinalization(Module & M)354 bool GlobalMerge::doFinalization(Module &M) {
355 MustKeepGlobalVariables.clear();
356 return false;
357 }
358
createGlobalMergePass(const TargetMachine * TM)359 Pass *llvm::createGlobalMergePass(const TargetMachine *TM) {
360 return new GlobalMerge(TM);
361 }
362