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/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"
72 using namespace llvm;
73
74 #define DEBUG_TYPE "global-merge"
75
76 static cl::opt<bool>
77 EnableGlobalMerge("enable-global-merge", cl::Hidden,
78 cl::desc("Enable global merge pass"),
79 cl::init(true));
80
81 static cl::opt<bool>
82 EnableGlobalMergeOnConst("global-merge-on-const", cl::Hidden,
83 cl::desc("Enable global merge pass on constants"),
84 cl::init(false));
85
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.
88 static cl::opt<bool>
89 EnableGlobalMergeOnExternal("global-merge-on-external", cl::Hidden,
90 cl::desc("Enable global merge pass on external linkage"),
91 cl::init(false));
92
93 STATISTIC(NumMerged , "Number of globals merged");
94 namespace {
95 class GlobalMerge : public FunctionPass {
96 const TargetMachine *TM;
97
98 bool doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
99 Module &M, bool isConst, unsigned AddrSpace) const;
100
101 /// \brief Check if the given variable has been identified as must keep
102 /// \pre setMustKeepGlobalVariables must have been called on the Module that
103 /// contains GV
isMustKeepGlobalVariable(const GlobalVariable * GV) const104 bool isMustKeepGlobalVariable(const GlobalVariable *GV) const {
105 return MustKeepGlobalVariables.count(GV);
106 }
107
108 /// Collect every variables marked as "used" or used in a landing pad
109 /// instruction for this Module.
110 void setMustKeepGlobalVariables(Module &M);
111
112 /// Collect every variables marked as "used"
113 void collectUsedGlobalVariables(Module &M);
114
115 /// Keep track of the GlobalVariable that must not be merged away
116 SmallPtrSet<const GlobalVariable *, 16> MustKeepGlobalVariables;
117
118 public:
119 static char ID; // Pass identification, replacement for typeid.
GlobalMerge(const TargetMachine * TM=nullptr)120 explicit GlobalMerge(const TargetMachine *TM = nullptr)
121 : FunctionPass(ID), TM(TM) {
122 initializeGlobalMergePass(*PassRegistry::getPassRegistry());
123 }
124
125 bool doInitialization(Module &M) override;
126 bool runOnFunction(Function &F) override;
127 bool doFinalization(Module &M) override;
128
getPassName() const129 const char *getPassName() const override {
130 return "Merge internal globals";
131 }
132
getAnalysisUsage(AnalysisUsage & AU) const133 void getAnalysisUsage(AnalysisUsage &AU) const override {
134 AU.setPreservesCFG();
135 FunctionPass::getAnalysisUsage(AU);
136 }
137 };
138 } // end anonymous namespace
139
140 char GlobalMerge::ID = 0;
141 INITIALIZE_TM_PASS(GlobalMerge, "global-merge", "Merge global variables",
142 false, false)
143
doMerge(SmallVectorImpl<GlobalVariable * > & Globals,Module & M,bool isConst,unsigned AddrSpace) const144 bool GlobalMerge::doMerge(SmallVectorImpl<GlobalVariable*> &Globals,
145 Module &M, bool isConst, unsigned AddrSpace) const {
146 const TargetLowering *TLI = TM->getSubtargetImpl()->getTargetLowering();
147 const DataLayout *DL = TLI->getDataLayout();
148
149 // FIXME: Infer the maximum possible offset depending on the actual users
150 // (these max offsets are different for the users inside Thumb or ARM
151 // functions)
152 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
153
154 // FIXME: Find better heuristics
155 std::stable_sort(Globals.begin(), Globals.end(),
156 [DL](const GlobalVariable *GV1, const GlobalVariable *GV2) {
157 Type *Ty1 = cast<PointerType>(GV1->getType())->getElementType();
158 Type *Ty2 = cast<PointerType>(GV2->getType())->getElementType();
159
160 return (DL->getTypeAllocSize(Ty1) < DL->getTypeAllocSize(Ty2));
161 });
162
163 Type *Int32Ty = Type::getInt32Ty(M.getContext());
164
165 assert(Globals.size() > 1);
166
167 // FIXME: This simple solution merges globals all together as maximum as
168 // possible. However, with this solution it would be hard to remove dead
169 // global symbols at link-time. An alternative solution could be checking
170 // global symbols references function by function, and make the symbols
171 // being referred in the same function merged and we would probably need
172 // to introduce heuristic algorithm to solve the merge conflict from
173 // different functions.
174 for (size_t i = 0, e = Globals.size(); i != e; ) {
175 size_t j = 0;
176 uint64_t MergedSize = 0;
177 std::vector<Type*> Tys;
178 std::vector<Constant*> Inits;
179
180 bool HasExternal = false;
181 GlobalVariable *TheFirstExternal = 0;
182 for (j = i; j != e; ++j) {
183 Type *Ty = Globals[j]->getType()->getElementType();
184 MergedSize += DL->getTypeAllocSize(Ty);
185 if (MergedSize > MaxOffset) {
186 break;
187 }
188 Tys.push_back(Ty);
189 Inits.push_back(Globals[j]->getInitializer());
190
191 if (Globals[j]->hasExternalLinkage() && !HasExternal) {
192 HasExternal = true;
193 TheFirstExternal = Globals[j];
194 }
195 }
196
197 // If merged variables doesn't have external linkage, we needn't to expose
198 // the symbol after merging.
199 GlobalValue::LinkageTypes Linkage = HasExternal
200 ? GlobalValue::ExternalLinkage
201 : GlobalValue::InternalLinkage;
202
203 StructType *MergedTy = StructType::get(M.getContext(), Tys);
204 Constant *MergedInit = ConstantStruct::get(MergedTy, Inits);
205
206 // If merged variables have external linkage, we use symbol name of the
207 // first variable merged as the suffix of global symbol name. This would
208 // be able to avoid the link-time naming conflict for globalm symbols.
209 GlobalVariable *MergedGV = new GlobalVariable(
210 M, MergedTy, isConst, Linkage, MergedInit,
211 HasExternal ? "_MergedGlobals_" + TheFirstExternal->getName()
212 : "_MergedGlobals",
213 nullptr, GlobalVariable::NotThreadLocal, AddrSpace);
214
215 for (size_t k = i; k < j; ++k) {
216 GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage();
217 std::string Name = Globals[k]->getName();
218
219 Constant *Idx[2] = {
220 ConstantInt::get(Int32Ty, 0),
221 ConstantInt::get(Int32Ty, k-i)
222 };
223 Constant *GEP = ConstantExpr::getInBoundsGetElementPtr(MergedGV, Idx);
224 Globals[k]->replaceAllUsesWith(GEP);
225 Globals[k]->eraseFromParent();
226
227 if (Linkage != GlobalValue::InternalLinkage) {
228 // Generate a new alias...
229 auto *PTy = cast<PointerType>(GEP->getType());
230 GlobalAlias::create(PTy->getElementType(), PTy->getAddressSpace(),
231 Linkage, Name, GEP, &M);
232 }
233
234 NumMerged++;
235 }
236 i = j;
237 }
238
239 return true;
240 }
241
collectUsedGlobalVariables(Module & M)242 void GlobalMerge::collectUsedGlobalVariables(Module &M) {
243 // Extract global variables from llvm.used array
244 const GlobalVariable *GV = M.getGlobalVariable("llvm.used");
245 if (!GV || !GV->hasInitializer()) return;
246
247 // Should be an array of 'i8*'.
248 const ConstantArray *InitList = cast<ConstantArray>(GV->getInitializer());
249
250 for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i)
251 if (const GlobalVariable *G =
252 dyn_cast<GlobalVariable>(InitList->getOperand(i)->stripPointerCasts()))
253 MustKeepGlobalVariables.insert(G);
254 }
255
setMustKeepGlobalVariables(Module & M)256 void GlobalMerge::setMustKeepGlobalVariables(Module &M) {
257 collectUsedGlobalVariables(M);
258
259 for (Module::iterator IFn = M.begin(), IEndFn = M.end(); IFn != IEndFn;
260 ++IFn) {
261 for (Function::iterator IBB = IFn->begin(), IEndBB = IFn->end();
262 IBB != IEndBB; ++IBB) {
263 // Follow the invoke link to find the landing pad instruction
264 const InvokeInst *II = dyn_cast<InvokeInst>(IBB->getTerminator());
265 if (!II) continue;
266
267 const LandingPadInst *LPInst = II->getUnwindDest()->getLandingPadInst();
268 // Look for globals in the clauses of the landing pad instruction
269 for (unsigned Idx = 0, NumClauses = LPInst->getNumClauses();
270 Idx != NumClauses; ++Idx)
271 if (const GlobalVariable *GV =
272 dyn_cast<GlobalVariable>(LPInst->getClause(Idx)
273 ->stripPointerCasts()))
274 MustKeepGlobalVariables.insert(GV);
275 }
276 }
277 }
278
doInitialization(Module & M)279 bool GlobalMerge::doInitialization(Module &M) {
280 if (!EnableGlobalMerge)
281 return false;
282
283 DenseMap<unsigned, SmallVector<GlobalVariable*, 16> > Globals, ConstGlobals,
284 BSSGlobals;
285 const TargetLowering *TLI = TM->getSubtargetImpl()->getTargetLowering();
286 const DataLayout *DL = TLI->getDataLayout();
287 unsigned MaxOffset = TLI->getMaximalGlobalOffset();
288 bool Changed = false;
289 setMustKeepGlobalVariables(M);
290
291 // Grab all non-const globals.
292 for (Module::global_iterator I = M.global_begin(),
293 E = M.global_end(); I != E; ++I) {
294 // Merge is safe for "normal" internal or external globals only
295 if (I->isDeclaration() || I->isThreadLocal() || I->hasSection())
296 continue;
297
298 if (!(EnableGlobalMergeOnExternal && I->hasExternalLinkage()) &&
299 !I->hasInternalLinkage())
300 continue;
301
302 PointerType *PT = dyn_cast<PointerType>(I->getType());
303 assert(PT && "Global variable is not a pointer!");
304
305 unsigned AddressSpace = PT->getAddressSpace();
306
307 // Ignore fancy-aligned globals for now.
308 unsigned Alignment = DL->getPreferredAlignment(I);
309 Type *Ty = I->getType()->getElementType();
310 if (Alignment > DL->getABITypeAlignment(Ty))
311 continue;
312
313 // Ignore all 'special' globals.
314 if (I->getName().startswith("llvm.") ||
315 I->getName().startswith(".llvm."))
316 continue;
317
318 // Ignore all "required" globals:
319 if (isMustKeepGlobalVariable(I))
320 continue;
321
322 if (DL->getTypeAllocSize(Ty) < MaxOffset) {
323 if (TargetLoweringObjectFile::getKindForGlobal(I, *TM).isBSSLocal())
324 BSSGlobals[AddressSpace].push_back(I);
325 else if (I->isConstant())
326 ConstGlobals[AddressSpace].push_back(I);
327 else
328 Globals[AddressSpace].push_back(I);
329 }
330 }
331
332 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
333 I = Globals.begin(), E = Globals.end(); I != E; ++I)
334 if (I->second.size() > 1)
335 Changed |= doMerge(I->second, M, false, I->first);
336
337 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
338 I = BSSGlobals.begin(), E = BSSGlobals.end(); I != E; ++I)
339 if (I->second.size() > 1)
340 Changed |= doMerge(I->second, M, false, I->first);
341
342 if (EnableGlobalMergeOnConst)
343 for (DenseMap<unsigned, SmallVector<GlobalVariable*, 16> >::iterator
344 I = ConstGlobals.begin(), E = ConstGlobals.end(); I != E; ++I)
345 if (I->second.size() > 1)
346 Changed |= doMerge(I->second, M, true, I->first);
347
348 return Changed;
349 }
350
runOnFunction(Function & F)351 bool GlobalMerge::runOnFunction(Function &F) {
352 return false;
353 }
354
doFinalization(Module & M)355 bool GlobalMerge::doFinalization(Module &M) {
356 MustKeepGlobalVariables.clear();
357 return false;
358 }
359
createGlobalMergePass(const TargetMachine * TM)360 Pass *llvm::createGlobalMergePass(const TargetMachine *TM) {
361 return new GlobalMerge(TM);
362 }
363