1//===-- GlobalDCE.cpp - DCE unreachable internal functions ----------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This transform is designed to eliminate unreachable internal globals from the
10// program. It uses an aggressive algorithm, searching out globals that are
11// known to be alive. After it finds all of the globals which are needed, it
12// deletes whatever is left over. This allows it to delete recursive chunks of
13// the program which are unreachable.
14//
15//===----------------------------------------------------------------------===//
16
17#include "llvm/Transforms/IPO/GlobalDCE.h"
18#include "llvm/ADT/SmallPtrSet.h"
19#include "llvm/ADT/Statistic.h"
20#include "llvm/Analysis/TypeMetadataUtils.h"
21#include "llvm/IR/Instructions.h"
22#include "llvm/IR/IntrinsicInst.h"
23#include "llvm/IR/Module.h"
24#include "llvm/InitializePasses.h"
25#include "llvm/Pass.h"
26#include "llvm/Support/CommandLine.h"
27#include "llvm/Transforms/IPO.h"
28#include "llvm/Transforms/Utils/CtorUtils.h"
29#include "llvm/Transforms/Utils/GlobalStatus.h"
30
31using namespace llvm;
32
33#define DEBUG_TYPE "globaldce"
34
35namespace {
36class GlobalDCELegacyPass : public ModulePass {
37public:
38 static char ID; // Pass identification, replacement for typeid
39 GlobalDCELegacyPass() : ModulePass(ID) {}
40 bool runOnModule(Module &M) override {
41 if (skipModule(M))
42 return false;
43 // Note: GlobalDCEPass does not use any analyses, so we're safe to call the
44 // new-pm style pass with a default-initialized analysis manager here
45 ModuleAnalysisManager MAM;
46 auto PA = Impl.run(M, MAM);
47 return !PA.areAllPreserved();
48 }
49
50private:
51 GlobalDCEPass Impl;
52};
53} // namespace
54
55char GlobalDCELegacyPass::ID = 0;
56INITIALIZE_PASS(GlobalDCELegacyPass, "globaldce", "Dead Global Elimination",
57 false, false)
58
59// Public interface to the GlobalDCEPass.
60ModulePass *llvm::createGlobalDCEPass() { return new GlobalDCELegacyPass(); }
61
62static cl::opt<bool>
63 ClEnableVFE("enable-vfe", cl::Hidden, cl::init(Val: true),
64 cl::desc("Enable virtual function elimination"));
65
66STATISTIC(NumAliases , "Number of global aliases removed");
67STATISTIC(NumFunctions, "Number of functions removed");
68STATISTIC(NumIFuncs, "Number of indirect functions removed");
69STATISTIC(NumVariables, "Number of global variables removed");
70STATISTIC(NumVFuncs, "Number of virtual functions removed");
71
72/// Returns true if F is effectively empty.
73static bool isEmptyFunction(Function *F) {
74 // Skip external functions.
75 if (F->isDeclaration())
76 return false;
77 BasicBlock &Entry = F->getEntryBlock();
78 for (auto &I : Entry) {
79 if (I.isDebugOrPseudoInst())
80 continue;
81 if (auto *RI = dyn_cast<ReturnInst>(Val: &I))
82 return !RI->getReturnValue();
83 break;
84 }
85 return false;
86}
87
88/// Compute the set of GlobalValue that depends from V.
89/// The recursion stops as soon as a GlobalValue is met.
90void GlobalDCEPass::ComputeDependencies(Value *V,
91 SmallPtrSetImpl<GlobalValue *> &Deps) {
92 if (auto *I = dyn_cast<Instruction>(Val: V)) {
93 Function *Parent = I->getParent()->getParent();
94 Deps.insert(Ptr: Parent);
95 } else if (auto *GV = dyn_cast<GlobalValue>(Val: V)) {
96 Deps.insert(Ptr: GV);
97 } else if (auto *CE = dyn_cast<Constant>(Val: V)) {
98 // Avoid walking the whole tree of a big ConstantExprs multiple times.
99 auto [Where, Inserted] = ConstantDependenciesCache.try_emplace(k: CE);
100 SmallPtrSetImpl<GlobalValue *> &LocalDeps = Where->second;
101 if (Inserted) {
102 for (User *CEUser : CE->users())
103 ComputeDependencies(V: CEUser, Deps&: LocalDeps);
104 }
105 Deps.insert_range(R&: LocalDeps);
106 }
107}
108
109void GlobalDCEPass::UpdateGVDependencies(GlobalValue &GV) {
110 SmallPtrSet<GlobalValue *, 8> Deps;
111 for (User *User : GV.users())
112 ComputeDependencies(V: User, Deps);
113 Deps.erase(Ptr: &GV); // Remove self-reference.
114 for (GlobalValue *GVU : Deps) {
115 // If this is a dep from a vtable to a virtual function, and we have
116 // complete information about all virtual call sites which could call
117 // though this vtable, then skip it, because the call site information will
118 // be more precise.
119 if (VFESafeVTables.count(Ptr: GVU) && isa<Function>(Val: &GV)) {
120 LLVM_DEBUG(dbgs() << "Ignoring dep " << GVU->getName() << " -> "
121 << GV.getName() << "\n");
122 continue;
123 }
124 GVDependencies[GVU].insert(Ptr: &GV);
125 }
126}
127
128/// Mark Global value as Live
129void GlobalDCEPass::MarkLive(GlobalValue &GV,
130 SmallVectorImpl<GlobalValue *> *Updates) {
131 auto const Ret = AliveGlobals.insert(Ptr: &GV);
132 if (!Ret.second)
133 return;
134
135 if (Updates)
136 Updates->push_back(Elt: &GV);
137 if (Comdat *C = GV.getComdat()) {
138 for (auto &&CM : make_range(p: ComdatMembers.equal_range(x: C))) {
139 MarkLive(GV&: *CM.second, Updates); // Recursion depth is only two because only
140 // globals in the same comdat are visited.
141 }
142 }
143}
144
145void GlobalDCEPass::ScanVTables(Module &M) {
146 SmallVector<MDNode *, 2> Types;
147 LLVM_DEBUG(dbgs() << "Building type info -> vtable map\n");
148
149 for (GlobalVariable &GV : M.globals()) {
150 Types.clear();
151 GV.getMetadata(KindID: LLVMContext::MD_type, MDs&: Types);
152 if (GV.isDeclaration() || Types.empty())
153 continue;
154
155 // Use the typeid metadata on the vtable to build a mapping from typeids to
156 // the list of (GV, offset) pairs which are the possible vtables for that
157 // typeid.
158 for (MDNode *Type : Types) {
159 Metadata *TypeID = Type->getOperand(I: 1).get();
160
161 uint64_t Offset =
162 cast<ConstantInt>(
163 Val: cast<ConstantAsMetadata>(Val: Type->getOperand(I: 0))->getValue())
164 ->getZExtValue();
165
166 TypeIdMap[TypeID].insert(V: std::make_pair(x: &GV, y&: Offset));
167 }
168
169 // If the type corresponding to the vtable is private to this translation
170 // unit, we know that we can see all virtual functions which might use it,
171 // so VFE is safe.
172 if (auto GO = dyn_cast<GlobalObject>(Val: &GV)) {
173 GlobalObject::VCallVisibility TypeVis = GO->getVCallVisibility();
174 if (TypeVis == GlobalObject::VCallVisibilityTranslationUnit ||
175 (InLTOPostLink &&
176 TypeVis == GlobalObject::VCallVisibilityLinkageUnit)) {
177 LLVM_DEBUG(dbgs() << GV.getName() << " is safe for VFE\n");
178 VFESafeVTables.insert(Ptr: &GV);
179 }
180 }
181 }
182}
183
184void GlobalDCEPass::ScanVTableLoad(Function *Caller, Metadata *TypeId,
185 uint64_t CallOffset) {
186 for (const auto &VTableInfo : TypeIdMap[TypeId]) {
187 GlobalVariable *VTable = VTableInfo.first;
188 uint64_t VTableOffset = VTableInfo.second;
189
190 Constant *Ptr =
191 getPointerAtOffset(I: VTable->getInitializer(), Offset: VTableOffset + CallOffset,
192 M&: *Caller->getParent(), TopLevelGlobal: VTable);
193 if (!Ptr) {
194 LLVM_DEBUG(dbgs() << "can't find pointer in vtable!\n");
195 VFESafeVTables.erase(Ptr: VTable);
196 continue;
197 }
198
199 auto Callee = dyn_cast<Function>(Val: Ptr->stripPointerCasts());
200 if (!Callee) {
201 LLVM_DEBUG(dbgs() << "vtable entry is not function pointer!\n");
202 VFESafeVTables.erase(Ptr: VTable);
203 continue;
204 }
205
206 LLVM_DEBUG(dbgs() << "vfunc dep " << Caller->getName() << " -> "
207 << Callee->getName() << "\n");
208 GVDependencies[Caller].insert(Ptr: Callee);
209 }
210}
211
212void GlobalDCEPass::ScanTypeCheckedLoadIntrinsics(Module &M) {
213 LLVM_DEBUG(dbgs() << "Scanning type.checked.load intrinsics\n");
214 Function *TypeCheckedLoadFunc =
215 Intrinsic::getDeclarationIfExists(M: &M, id: Intrinsic::type_checked_load);
216 Function *TypeCheckedLoadRelativeFunc = Intrinsic::getDeclarationIfExists(
217 M: &M, id: Intrinsic::type_checked_load_relative);
218
219 auto scan = [&](Function *CheckedLoadFunc) {
220 if (!CheckedLoadFunc)
221 return;
222
223 for (auto *U : CheckedLoadFunc->users()) {
224 auto CI = dyn_cast<CallInst>(Val: U);
225 if (!CI)
226 continue;
227
228 auto *Offset = dyn_cast<ConstantInt>(Val: CI->getArgOperand(i: 1));
229 Value *TypeIdValue = CI->getArgOperand(i: 2);
230 auto *TypeId = cast<MetadataAsValue>(Val: TypeIdValue)->getMetadata();
231
232 if (Offset) {
233 ScanVTableLoad(Caller: CI->getFunction(), TypeId, CallOffset: Offset->getZExtValue());
234 } else {
235 // type.checked.load with a non-constant offset, so assume every entry
236 // in every matching vtable is used.
237 for (const auto &VTableInfo : TypeIdMap[TypeId]) {
238 VFESafeVTables.erase(Ptr: VTableInfo.first);
239 }
240 }
241 }
242 };
243
244 scan(TypeCheckedLoadFunc);
245 scan(TypeCheckedLoadRelativeFunc);
246}
247
248void GlobalDCEPass::AddVirtualFunctionDependencies(Module &M) {
249 if (!ClEnableVFE)
250 return;
251
252 // If the Virtual Function Elim module flag is present and set to zero, then
253 // the vcall_visibility metadata was inserted for another optimization (WPD)
254 // and we may not have type checked loads on all accesses to the vtable.
255 // Don't attempt VFE in that case.
256 auto *Val = mdconst::dyn_extract_or_null<ConstantInt>(
257 MD: M.getModuleFlag(Key: "Virtual Function Elim"));
258 if (!Val || Val->isZero())
259 return;
260
261 ScanVTables(M);
262
263 if (VFESafeVTables.empty())
264 return;
265
266 ScanTypeCheckedLoadIntrinsics(M);
267
268 LLVM_DEBUG(
269 dbgs() << "VFE safe vtables:\n";
270 for (auto *VTable : VFESafeVTables)
271 dbgs() << " " << VTable->getName() << "\n";
272 );
273}
274
275PreservedAnalyses GlobalDCEPass::run(Module &M, ModuleAnalysisManager &MAM) {
276 bool Changed = false;
277
278 // The algorithm first computes the set L of global variables that are
279 // trivially live. Then it walks the initialization of these variables to
280 // compute the globals used to initialize them, which effectively builds a
281 // directed graph where nodes are global variables, and an edge from A to B
282 // means B is used to initialize A. Finally, it propagates the liveness
283 // information through the graph starting from the nodes in L. Nodes note
284 // marked as alive are discarded.
285
286 // Remove empty functions from the global ctors list.
287 Changed |= optimizeGlobalCtorsList(
288 M, ShouldRemove: [](uint32_t, Function *F) { return isEmptyFunction(F); });
289
290 // Collect the set of members for each comdat.
291 for (Function &F : M)
292 if (Comdat *C = F.getComdat())
293 ComdatMembers.insert(x: std::make_pair(x&: C, y: &F));
294 for (GlobalVariable &GV : M.globals())
295 if (Comdat *C = GV.getComdat())
296 ComdatMembers.insert(x: std::make_pair(x&: C, y: &GV));
297 for (GlobalAlias &GA : M.aliases())
298 if (Comdat *C = GA.getComdat())
299 ComdatMembers.insert(x: std::make_pair(x&: C, y: &GA));
300
301 // Add dependencies between virtual call sites and the virtual functions they
302 // might call, if we have that information.
303 AddVirtualFunctionDependencies(M);
304
305 // Loop over the module, adding globals which are obviously necessary.
306 for (GlobalObject &GO : M.global_objects()) {
307 GO.removeDeadConstantUsers();
308 // Functions with external linkage are needed if they have a body.
309 // Externally visible & appending globals are needed, if they have an
310 // initializer.
311 if (!GO.isDeclaration())
312 if (!GO.isDiscardableIfUnused())
313 MarkLive(GV&: GO);
314
315 UpdateGVDependencies(GV&: GO);
316 }
317
318 // Compute direct dependencies of aliases.
319 for (GlobalAlias &GA : M.aliases()) {
320 GA.removeDeadConstantUsers();
321 // Externally visible aliases are needed.
322 if (!GA.isDiscardableIfUnused())
323 MarkLive(GV&: GA);
324
325 UpdateGVDependencies(GV&: GA);
326 }
327
328 // Compute direct dependencies of ifuncs.
329 for (GlobalIFunc &GIF : M.ifuncs()) {
330 GIF.removeDeadConstantUsers();
331 // Externally visible ifuncs are needed.
332 if (!GIF.isDiscardableIfUnused())
333 MarkLive(GV&: GIF);
334
335 UpdateGVDependencies(GV&: GIF);
336 }
337
338 // Propagate liveness from collected Global Values through the computed
339 // dependencies.
340 SmallVector<GlobalValue *, 8> NewLiveGVs{AliveGlobals.begin(),
341 AliveGlobals.end()};
342 while (!NewLiveGVs.empty()) {
343 GlobalValue *LGV = NewLiveGVs.pop_back_val();
344 for (auto *GVD : GVDependencies[LGV])
345 MarkLive(GV&: *GVD, Updates: &NewLiveGVs);
346 }
347
348 // Now that all globals which are needed are in the AliveGlobals set, we loop
349 // through the program, deleting those which are not alive.
350 //
351
352 // The first pass is to drop initializers of global variables which are dead.
353 std::vector<GlobalVariable *> DeadGlobalVars; // Keep track of dead globals
354 for (GlobalVariable &GV : M.globals())
355 if (!AliveGlobals.count(Ptr: &GV)) {
356 DeadGlobalVars.push_back(x: &GV); // Keep track of dead globals
357 if (GV.hasInitializer()) {
358 Constant *Init = GV.getInitializer();
359 GV.setInitializer(nullptr);
360 if (isSafeToDestroyConstant(C: Init))
361 Init->destroyConstant();
362 }
363 }
364
365 // The second pass drops the bodies of functions which are dead...
366 std::vector<Function *> DeadFunctions;
367 for (Function &F : M)
368 if (!AliveGlobals.count(Ptr: &F)) {
369 DeadFunctions.push_back(x: &F); // Keep track of dead globals
370 if (!F.isDeclaration())
371 F.deleteBody();
372 }
373
374 // The third pass drops targets of aliases which are dead...
375 std::vector<GlobalAlias*> DeadAliases;
376 for (GlobalAlias &GA : M.aliases())
377 if (!AliveGlobals.count(Ptr: &GA)) {
378 DeadAliases.push_back(x: &GA);
379 GA.setAliasee(nullptr);
380 }
381
382 // The fourth pass drops targets of ifuncs which are dead...
383 std::vector<GlobalIFunc*> DeadIFuncs;
384 for (GlobalIFunc &GIF : M.ifuncs())
385 if (!AliveGlobals.count(Ptr: &GIF)) {
386 DeadIFuncs.push_back(x: &GIF);
387 GIF.setResolver(nullptr);
388 }
389
390 // Now that all interferences have been dropped, delete the actual objects
391 // themselves.
392 auto EraseUnusedGlobalValue = [&](GlobalValue *GV) {
393 GV->removeDeadConstantUsers();
394 GV->eraseFromParent();
395 Changed = true;
396 };
397
398 NumFunctions += DeadFunctions.size();
399 for (Function *F : DeadFunctions) {
400 if (!F->use_empty()) {
401 // Virtual functions might still be referenced by one or more vtables,
402 // but if we've proven them to be unused then it's safe to replace the
403 // virtual function pointers with null, allowing us to remove the
404 // function itself.
405 ++NumVFuncs;
406
407 // Detect vfuncs that are referenced as "relative pointers" which are used
408 // in Swift vtables, i.e. entries in the form of:
409 //
410 // i32 trunc (i64 sub (i64 ptrtoint @f, i64 ptrtoint ...)) to i32)
411 //
412 // In this case, replace the whole "sub" expression with constant 0 to
413 // avoid leaving a weird sub(0, symbol) expression behind.
414 replaceRelativePointerUsersWithZero(C: F);
415
416 F->replaceNonMetadataUsesWith(V: ConstantPointerNull::get(T: F->getType()));
417 }
418 EraseUnusedGlobalValue(F);
419 }
420
421 NumVariables += DeadGlobalVars.size();
422 for (GlobalVariable *GV : DeadGlobalVars)
423 EraseUnusedGlobalValue(GV);
424
425 NumAliases += DeadAliases.size();
426 for (GlobalAlias *GA : DeadAliases)
427 EraseUnusedGlobalValue(GA);
428
429 NumIFuncs += DeadIFuncs.size();
430 for (GlobalIFunc *GIF : DeadIFuncs)
431 EraseUnusedGlobalValue(GIF);
432
433 // Make sure that all memory is released
434 AliveGlobals.clear();
435 ConstantDependenciesCache.clear();
436 GVDependencies.clear();
437 ComdatMembers.clear();
438 TypeIdMap.clear();
439 VFESafeVTables.clear();
440
441 if (Changed)
442 return PreservedAnalyses::none();
443 return PreservedAnalyses::all();
444}
445
446void GlobalDCEPass::printPipeline(
447 raw_ostream &OS, function_ref<StringRef(StringRef)> MapClassName2PassName) {
448 static_cast<PassInfoMixin<GlobalDCEPass> *>(this)->printPipeline(
449 OS, MapClassName2PassName);
450 if (InLTOPostLink)
451 OS << "<vfe-linkage-unit-visibility>";
452}
453