| 1 | //===- GlobalMerge.cpp - Internal globals merging -------------------------===// |
| 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 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 | // However, merging globals can have tradeoffs: |
| 54 | // - it confuses debuggers, tools, and users |
| 55 | // - it makes linker optimizations less useful (order files, LOHs, ...) |
| 56 | // - it forces usage of indexed addressing (which isn't necessarily "free") |
| 57 | // - it can increase register pressure when the uses are disparate enough. |
| 58 | // |
| 59 | // We use heuristics to discover the best global grouping we can (cf cl::opts). |
| 60 | // |
| 61 | // ===---------------------------------------------------------------------===// |
| 62 | |
| 63 | #include "llvm/CodeGen/GlobalMerge.h" |
| 64 | #include "llvm/ADT/BitVector.h" |
| 65 | #include "llvm/ADT/DenseMap.h" |
| 66 | #include "llvm/ADT/MapVector.h" |
| 67 | #include "llvm/ADT/SetVector.h" |
| 68 | #include "llvm/ADT/SmallVector.h" |
| 69 | #include "llvm/ADT/Statistic.h" |
| 70 | #include "llvm/ADT/StringRef.h" |
| 71 | #include "llvm/ADT/Twine.h" |
| 72 | #include "llvm/CodeGen/Passes.h" |
| 73 | #include "llvm/IR/BasicBlock.h" |
| 74 | #include "llvm/IR/Constants.h" |
| 75 | #include "llvm/IR/DataLayout.h" |
| 76 | #include "llvm/IR/DerivedTypes.h" |
| 77 | #include "llvm/IR/Function.h" |
| 78 | #include "llvm/IR/GlobalAlias.h" |
| 79 | #include "llvm/IR/GlobalValue.h" |
| 80 | #include "llvm/IR/GlobalVariable.h" |
| 81 | #include "llvm/IR/Instruction.h" |
| 82 | #include "llvm/IR/IntrinsicInst.h" |
| 83 | #include "llvm/IR/Module.h" |
| 84 | #include "llvm/IR/Type.h" |
| 85 | #include "llvm/IR/Use.h" |
| 86 | #include "llvm/IR/User.h" |
| 87 | #include "llvm/InitializePasses.h" |
| 88 | #include "llvm/MC/SectionKind.h" |
| 89 | #include "llvm/Pass.h" |
| 90 | #include "llvm/Support/Casting.h" |
| 91 | #include "llvm/Support/CommandLine.h" |
| 92 | #include "llvm/Support/Debug.h" |
| 93 | #include "llvm/Support/raw_ostream.h" |
| 94 | #include "llvm/Target/TargetLoweringObjectFile.h" |
| 95 | #include "llvm/Target/TargetMachine.h" |
| 96 | #include "llvm/TargetParser/Triple.h" |
| 97 | #include <algorithm> |
| 98 | #include <cassert> |
| 99 | #include <cstddef> |
| 100 | #include <cstdint> |
| 101 | #include <string> |
| 102 | #include <vector> |
| 103 | |
| 104 | using namespace llvm; |
| 105 | |
| 106 | #define DEBUG_TYPE "global-merge" |
| 107 | |
| 108 | // FIXME: This is only useful as a last-resort way to disable the pass. |
| 109 | static cl::opt<bool> |
| 110 | EnableGlobalMerge("enable-global-merge" , cl::Hidden, |
| 111 | cl::desc("Enable the global merge pass" ), |
| 112 | cl::init(Val: true)); |
| 113 | |
| 114 | static cl::opt<unsigned> |
| 115 | GlobalMergeMaxOffset("global-merge-max-offset" , cl::Hidden, |
| 116 | cl::desc("Set maximum offset for global merge pass" ), |
| 117 | cl::init(Val: 0)); |
| 118 | |
| 119 | static cl::opt<bool> GlobalMergeGroupByUse( |
| 120 | "global-merge-group-by-use" , cl::Hidden, |
| 121 | cl::desc("Improve global merge pass to look at uses" ), cl::init(Val: true)); |
| 122 | |
| 123 | static cl::opt<bool> GlobalMergeAllConst( |
| 124 | "global-merge-all-const" , cl::Hidden, |
| 125 | cl::desc("Merge all const globals without looking at uses" ), |
| 126 | cl::init(Val: false)); |
| 127 | |
| 128 | static cl::opt<bool> GlobalMergeIgnoreSingleUse( |
| 129 | "global-merge-ignore-single-use" , cl::Hidden, |
| 130 | cl::desc("Improve global merge pass to ignore globals only used alone" ), |
| 131 | cl::init(Val: true)); |
| 132 | |
| 133 | static cl::opt<bool> |
| 134 | EnableGlobalMergeOnConst("global-merge-on-const" , cl::Hidden, |
| 135 | cl::desc("Enable global merge pass on constants" ), |
| 136 | cl::init(Val: false)); |
| 137 | |
| 138 | // FIXME: this could be a transitional option, and we probably need to remove |
| 139 | // it if only we are sure this optimization could always benefit all targets. |
| 140 | static cl::opt<cl::boolOrDefault> |
| 141 | EnableGlobalMergeOnExternal("global-merge-on-external" , cl::Hidden, |
| 142 | cl::desc("Enable global merge pass on external linkage" )); |
| 143 | |
| 144 | static cl::opt<unsigned> |
| 145 | GlobalMergeMinDataSize("global-merge-min-data-size" , |
| 146 | cl::desc("The minimum size in bytes of each global " |
| 147 | "that should considered in merging." ), |
| 148 | cl::init(Val: 0), cl::Hidden); |
| 149 | |
| 150 | STATISTIC(NumMerged, "Number of globals merged" ); |
| 151 | |
| 152 | namespace { |
| 153 | |
| 154 | class GlobalMergeImpl { |
| 155 | const TargetMachine *TM = nullptr; |
| 156 | GlobalMergeOptions Opt; |
| 157 | bool IsMachO = false; |
| 158 | |
| 159 | private: |
| 160 | bool doMerge(SmallVectorImpl<GlobalVariable *> &Globals, Module &M, |
| 161 | bool isConst, unsigned AddrSpace) const; |
| 162 | |
| 163 | /// Merge everything in \p Globals for which the corresponding bit |
| 164 | /// in \p GlobalSet is set. |
| 165 | bool doMerge(const SmallVectorImpl<GlobalVariable *> &Globals, |
| 166 | const BitVector &GlobalSet, Module &M, bool isConst, |
| 167 | unsigned AddrSpace) const; |
| 168 | |
| 169 | /// Check if the given variable has been identified as must keep |
| 170 | /// \pre setMustKeepGlobalVariables must have been called on the Module that |
| 171 | /// contains GV |
| 172 | bool isMustKeepGlobalVariable(const GlobalVariable *GV) const { |
| 173 | return MustKeepGlobalVariables.count(key: GV); |
| 174 | } |
| 175 | |
| 176 | /// Collect every variables marked as "used" or used in a landing pad |
| 177 | /// instruction for this Module. |
| 178 | void setMustKeepGlobalVariables(Module &M); |
| 179 | |
| 180 | /// Collect every variables marked as "used" |
| 181 | void collectUsedGlobalVariables(Module &M, StringRef Name); |
| 182 | |
| 183 | /// Keep track of the GlobalVariable that must not be merged away |
| 184 | SmallSetVector<const GlobalVariable *, 16> MustKeepGlobalVariables; |
| 185 | |
| 186 | public: |
| 187 | GlobalMergeImpl(const TargetMachine *TM, GlobalMergeOptions Opt) |
| 188 | : TM(TM), Opt(Opt) {} |
| 189 | bool run(Module &M); |
| 190 | }; |
| 191 | |
| 192 | class GlobalMerge : public FunctionPass { |
| 193 | const TargetMachine *TM = nullptr; |
| 194 | GlobalMergeOptions Opt; |
| 195 | |
| 196 | public: |
| 197 | static char ID; // Pass identification, replacement for typeid. |
| 198 | |
| 199 | explicit GlobalMerge() : FunctionPass(ID) { |
| 200 | Opt.MaxOffset = GlobalMergeMaxOffset; |
| 201 | Opt.MergeConstantGlobals = EnableGlobalMergeOnConst; |
| 202 | Opt.MergeConstAggressive = GlobalMergeAllConst; |
| 203 | initializeGlobalMergePass(*PassRegistry::getPassRegistry()); |
| 204 | } |
| 205 | |
| 206 | explicit GlobalMerge(const TargetMachine *TM, unsigned MaximalOffset, |
| 207 | bool OnlyOptimizeForSize, bool MergeExternalGlobals, |
| 208 | bool MergeConstantGlobals, bool MergeConstAggressive) |
| 209 | : FunctionPass(ID), TM(TM) { |
| 210 | Opt.MaxOffset = MaximalOffset; |
| 211 | Opt.SizeOnly = OnlyOptimizeForSize; |
| 212 | Opt.MergeExternal = MergeExternalGlobals; |
| 213 | Opt.MergeConstantGlobals = MergeConstantGlobals; |
| 214 | Opt.MergeConstAggressive = MergeConstAggressive; |
| 215 | initializeGlobalMergePass(*PassRegistry::getPassRegistry()); |
| 216 | } |
| 217 | |
| 218 | bool doInitialization(Module &M) override { |
| 219 | auto GetSmallDataLimit = [](Module &M) -> std::optional<uint64_t> { |
| 220 | Metadata *SDL = M.getModuleFlag(Key: "SmallDataLimit" ); |
| 221 | if (!SDL) |
| 222 | return std::nullopt; |
| 223 | return mdconst::extract<ConstantInt>(MD&: SDL)->getZExtValue(); |
| 224 | }; |
| 225 | if (GlobalMergeMinDataSize.getNumOccurrences()) |
| 226 | Opt.MinSize = GlobalMergeMinDataSize; |
| 227 | else if (auto SDL = GetSmallDataLimit(M); SDL && *SDL > 0) |
| 228 | Opt.MinSize = *SDL + 1; |
| 229 | else |
| 230 | Opt.MinSize = 0; |
| 231 | |
| 232 | GlobalMergeImpl P(TM, Opt); |
| 233 | return P.run(M); |
| 234 | } |
| 235 | bool runOnFunction(Function &F) override { return false; } |
| 236 | |
| 237 | StringRef getPassName() const override { return "Merge internal globals" ; } |
| 238 | |
| 239 | void getAnalysisUsage(AnalysisUsage &AU) const override { |
| 240 | AU.setPreservesCFG(); |
| 241 | FunctionPass::getAnalysisUsage(AU); |
| 242 | } |
| 243 | }; |
| 244 | |
| 245 | } // end anonymous namespace |
| 246 | |
| 247 | PreservedAnalyses GlobalMergePass::run(Module &M, ModuleAnalysisManager &) { |
| 248 | GlobalMergeImpl P(TM, Options); |
| 249 | bool Changed = P.run(M); |
| 250 | if (!Changed) |
| 251 | return PreservedAnalyses::all(); |
| 252 | |
| 253 | PreservedAnalyses PA; |
| 254 | PA.preserveSet<CFGAnalyses>(); |
| 255 | return PA; |
| 256 | } |
| 257 | |
| 258 | char GlobalMerge::ID = 0; |
| 259 | |
| 260 | INITIALIZE_PASS(GlobalMerge, DEBUG_TYPE, "Merge global variables" , false, false) |
| 261 | |
| 262 | bool GlobalMergeImpl::doMerge(SmallVectorImpl<GlobalVariable *> &Globals, |
| 263 | Module &M, bool isConst, |
| 264 | unsigned AddrSpace) const { |
| 265 | auto &DL = M.getDataLayout(); |
| 266 | // FIXME: Find better heuristics |
| 267 | llvm::stable_sort( |
| 268 | Range&: Globals, C: [&DL](const GlobalVariable *GV1, const GlobalVariable *GV2) { |
| 269 | // We don't support scalable global variables. |
| 270 | return DL.getTypeAllocSize(Ty: GV1->getValueType()).getFixedValue() < |
| 271 | DL.getTypeAllocSize(Ty: GV2->getValueType()).getFixedValue(); |
| 272 | }); |
| 273 | |
| 274 | // If we want to just blindly group all globals together, do so. |
| 275 | if (!GlobalMergeGroupByUse || (Opt.MergeConstAggressive && isConst)) { |
| 276 | BitVector AllGlobals(Globals.size(), true); |
| 277 | return doMerge(Globals, GlobalSet: AllGlobals, M, isConst, AddrSpace); |
| 278 | } |
| 279 | |
| 280 | // If we want to be smarter, look at all uses of each global, to try to |
| 281 | // discover all sets of globals used together, and how many times each of |
| 282 | // these sets occurred. |
| 283 | // |
| 284 | // Keep this reasonably efficient, by having an append-only list of all sets |
| 285 | // discovered so far (UsedGlobalSet), and mapping each "together-ness" unit of |
| 286 | // code (currently, a Function) to the set of globals seen so far that are |
| 287 | // used together in that unit (GlobalUsesByFunction). |
| 288 | // |
| 289 | // When we look at the Nth global, we know that any new set is either: |
| 290 | // - the singleton set {N}, containing this global only, or |
| 291 | // - the union of {N} and a previously-discovered set, containing some |
| 292 | // combination of the previous N-1 globals. |
| 293 | // Using that knowledge, when looking at the Nth global, we can keep: |
| 294 | // - a reference to the singleton set {N} (CurGVOnlySetIdx) |
| 295 | // - a list mapping each previous set to its union with {N} (EncounteredUGS), |
| 296 | // if it actually occurs. |
| 297 | |
| 298 | // We keep track of the sets of globals used together "close enough". |
| 299 | struct UsedGlobalSet { |
| 300 | BitVector Globals; |
| 301 | unsigned UsageCount = 1; |
| 302 | |
| 303 | UsedGlobalSet(size_t Size) : Globals(Size) {} |
| 304 | }; |
| 305 | |
| 306 | // Each set is unique in UsedGlobalSets. |
| 307 | std::vector<UsedGlobalSet> UsedGlobalSets; |
| 308 | |
| 309 | // Avoid repeating the create-global-set pattern. |
| 310 | auto CreateGlobalSet = [&]() -> UsedGlobalSet & { |
| 311 | UsedGlobalSets.emplace_back(args: Globals.size()); |
| 312 | return UsedGlobalSets.back(); |
| 313 | }; |
| 314 | |
| 315 | // The first set is the empty set. |
| 316 | CreateGlobalSet().UsageCount = 0; |
| 317 | |
| 318 | // We define "close enough" to be "in the same function". |
| 319 | // FIXME: Grouping uses by function is way too aggressive, so we should have |
| 320 | // a better metric for distance between uses. |
| 321 | // The obvious alternative would be to group by BasicBlock, but that's in |
| 322 | // turn too conservative.. |
| 323 | // Anything in between wouldn't be trivial to compute, so just stick with |
| 324 | // per-function grouping. |
| 325 | |
| 326 | // The value type is an index into UsedGlobalSets. |
| 327 | // The default (0) conveniently points to the empty set. |
| 328 | DenseMap<Function *, size_t /*UsedGlobalSetIdx*/> GlobalUsesByFunction; |
| 329 | |
| 330 | // Now, look at each merge-eligible global in turn. |
| 331 | |
| 332 | // Keep track of the sets we already encountered to which we added the |
| 333 | // current global. |
| 334 | // Each element matches the same-index element in UsedGlobalSets. |
| 335 | // This lets us efficiently tell whether a set has already been expanded to |
| 336 | // include the current global. |
| 337 | std::vector<size_t> EncounteredUGS; |
| 338 | |
| 339 | for (size_t GI = 0, GE = Globals.size(); GI != GE; ++GI) { |
| 340 | GlobalVariable *GV = Globals[GI]; |
| 341 | |
| 342 | // Reset the encountered sets for this global and grow it in case we created |
| 343 | // new sets for the previous global. |
| 344 | EncounteredUGS.assign(n: UsedGlobalSets.size(), val: 0); |
| 345 | |
| 346 | // We might need to create a set that only consists of the current global. |
| 347 | // Keep track of its index into UsedGlobalSets. |
| 348 | size_t CurGVOnlySetIdx = 0; |
| 349 | |
| 350 | // For each global, look at all its Uses. |
| 351 | for (auto &U : GV->uses()) { |
| 352 | // This Use might be a ConstantExpr. We're interested in Instruction |
| 353 | // users, so look through ConstantExpr... |
| 354 | Use *UI, *UE; |
| 355 | if (ConstantExpr *CE = dyn_cast<ConstantExpr>(Val: U.getUser())) { |
| 356 | if (CE->use_empty()) |
| 357 | continue; |
| 358 | UI = &*CE->use_begin(); |
| 359 | UE = nullptr; |
| 360 | } else if (isa<Instruction>(Val: U.getUser())) { |
| 361 | UI = &U; |
| 362 | UE = UI->getNext(); |
| 363 | } else { |
| 364 | continue; |
| 365 | } |
| 366 | |
| 367 | // ...to iterate on all the instruction users of the global. |
| 368 | // Note that we iterate on Uses and not on Users to be able to getNext(). |
| 369 | for (; UI != UE; UI = UI->getNext()) { |
| 370 | Instruction *I = dyn_cast<Instruction>(Val: UI->getUser()); |
| 371 | if (!I) |
| 372 | continue; |
| 373 | |
| 374 | Function *ParentFn = I->getParent()->getParent(); |
| 375 | |
| 376 | // If we're only optimizing for size, ignore non-minsize functions. |
| 377 | if (Opt.SizeOnly && !ParentFn->hasMinSize()) |
| 378 | continue; |
| 379 | |
| 380 | size_t UGSIdx = GlobalUsesByFunction[ParentFn]; |
| 381 | |
| 382 | // If this is the first global the function uses, map it to the set |
| 383 | // consisting of this global only. |
| 384 | if (!UGSIdx) { |
| 385 | // If that set doesn't exist yet, create it. |
| 386 | if (!CurGVOnlySetIdx) { |
| 387 | CurGVOnlySetIdx = UsedGlobalSets.size(); |
| 388 | CreateGlobalSet().Globals.set(GI); |
| 389 | } else { |
| 390 | ++UsedGlobalSets[CurGVOnlySetIdx].UsageCount; |
| 391 | } |
| 392 | |
| 393 | GlobalUsesByFunction[ParentFn] = CurGVOnlySetIdx; |
| 394 | continue; |
| 395 | } |
| 396 | |
| 397 | // If we already encountered a use of this global in this function, just |
| 398 | // increment the counter. |
| 399 | if (UsedGlobalSets[UGSIdx].Globals.test(Idx: GI)) { |
| 400 | ++UsedGlobalSets[UGSIdx].UsageCount; |
| 401 | continue; |
| 402 | } |
| 403 | |
| 404 | // If not, the previous set wasn't actually used in this function. |
| 405 | --UsedGlobalSets[UGSIdx].UsageCount; |
| 406 | |
| 407 | // If we already expanded the previous set to include this global, just |
| 408 | // reuse that expanded set. |
| 409 | if (size_t ExpandedIdx = EncounteredUGS[UGSIdx]) { |
| 410 | ++UsedGlobalSets[ExpandedIdx].UsageCount; |
| 411 | GlobalUsesByFunction[ParentFn] = ExpandedIdx; |
| 412 | continue; |
| 413 | } |
| 414 | |
| 415 | // If not, create a new set consisting of the union of the previous set |
| 416 | // and this global. Mark it as encountered, so we can reuse it later. |
| 417 | GlobalUsesByFunction[ParentFn] = EncounteredUGS[UGSIdx] = |
| 418 | UsedGlobalSets.size(); |
| 419 | |
| 420 | UsedGlobalSet &NewUGS = CreateGlobalSet(); |
| 421 | NewUGS.Globals.set(GI); |
| 422 | NewUGS.Globals |= UsedGlobalSets[UGSIdx].Globals; |
| 423 | } |
| 424 | } |
| 425 | } |
| 426 | |
| 427 | // We can choose to merge all globals together, but ignore globals never used |
| 428 | // with another global. This catches the obviously non-profitable cases of |
| 429 | // having a single global, but is aggressive enough for any other case. |
| 430 | if (GlobalMergeIgnoreSingleUse) { |
| 431 | BitVector AllGlobals(Globals.size()); |
| 432 | for (const UsedGlobalSet &UGS : UsedGlobalSets) { |
| 433 | if (UGS.UsageCount == 0) |
| 434 | continue; |
| 435 | if (UGS.Globals.count() > 1) |
| 436 | AllGlobals |= UGS.Globals; |
| 437 | } |
| 438 | return doMerge(Globals, GlobalSet: AllGlobals, M, isConst, AddrSpace); |
| 439 | } |
| 440 | |
| 441 | // Now we found a bunch of sets of globals used together. We accumulated |
| 442 | // the number of times we encountered the sets (i.e., the number of functions |
| 443 | // that use that exact set of globals). |
| 444 | // |
| 445 | // Multiply that by the size of the set to give us a crude profitability |
| 446 | // metric. |
| 447 | llvm::stable_sort(Range&: UsedGlobalSets, |
| 448 | C: [](const UsedGlobalSet &UGS1, const UsedGlobalSet &UGS2) { |
| 449 | return UGS1.Globals.count() * UGS1.UsageCount < |
| 450 | UGS2.Globals.count() * UGS2.UsageCount; |
| 451 | }); |
| 452 | |
| 453 | // Starting from the sets with the best (=biggest) profitability, find a |
| 454 | // good combination. |
| 455 | // The ideal (and expensive) solution can only be found by trying all |
| 456 | // combinations, looking for the one with the best profitability. |
| 457 | // Don't be smart about it, and just pick the first compatible combination, |
| 458 | // starting with the sets with the best profitability. |
| 459 | BitVector PickedGlobals(Globals.size()); |
| 460 | bool Changed = false; |
| 461 | |
| 462 | for (const UsedGlobalSet &UGS : llvm::reverse(C&: UsedGlobalSets)) { |
| 463 | if (UGS.UsageCount == 0) |
| 464 | continue; |
| 465 | if (PickedGlobals.anyCommon(RHS: UGS.Globals)) |
| 466 | continue; |
| 467 | PickedGlobals |= UGS.Globals; |
| 468 | // If the set only contains one global, there's no point in merging. |
| 469 | // Ignore the global for inclusion in other sets though, so keep it in |
| 470 | // PickedGlobals. |
| 471 | if (UGS.Globals.count() < 2) |
| 472 | continue; |
| 473 | Changed |= doMerge(Globals, GlobalSet: UGS.Globals, M, isConst, AddrSpace); |
| 474 | } |
| 475 | |
| 476 | return Changed; |
| 477 | } |
| 478 | |
| 479 | bool GlobalMergeImpl::doMerge(const SmallVectorImpl<GlobalVariable *> &Globals, |
| 480 | const BitVector &GlobalSet, Module &M, |
| 481 | bool isConst, unsigned AddrSpace) const { |
| 482 | assert(Globals.size() > 1); |
| 483 | |
| 484 | Type *Int32Ty = Type::getInt32Ty(C&: M.getContext()); |
| 485 | Type *Int8Ty = Type::getInt8Ty(C&: M.getContext()); |
| 486 | auto &DL = M.getDataLayout(); |
| 487 | |
| 488 | LLVM_DEBUG(dbgs() << " Trying to merge set, starts with #" |
| 489 | << GlobalSet.find_first() << ", total of " << Globals.size() |
| 490 | << "\n" ); |
| 491 | |
| 492 | bool Changed = false; |
| 493 | ssize_t i = GlobalSet.find_first(); |
| 494 | while (i != -1) { |
| 495 | ssize_t j = 0; |
| 496 | uint64_t MergedSize = 0; |
| 497 | std::vector<Type*> Tys; |
| 498 | std::vector<Constant*> Inits; |
| 499 | std::vector<unsigned> StructIdxs; |
| 500 | |
| 501 | bool HasExternal = false; |
| 502 | StringRef FirstExternalName; |
| 503 | Align MaxAlign; |
| 504 | unsigned CurIdx = 0; |
| 505 | for (j = i; j != -1; j = GlobalSet.find_next(Prev: j)) { |
| 506 | Type *Ty = Globals[j]->getValueType(); |
| 507 | |
| 508 | // Make sure we use the same alignment AsmPrinter would use. |
| 509 | Align Alignment = DL.getPreferredAlign(GV: Globals[j]); |
| 510 | unsigned Padding = alignTo(Size: MergedSize, A: Alignment) - MergedSize; |
| 511 | MergedSize += Padding; |
| 512 | MergedSize += DL.getTypeAllocSize(Ty); |
| 513 | if (MergedSize > Opt.MaxOffset) { |
| 514 | break; |
| 515 | } |
| 516 | if (Padding) { |
| 517 | Tys.push_back(x: ArrayType::get(ElementType: Int8Ty, NumElements: Padding)); |
| 518 | Inits.push_back(x: ConstantAggregateZero::get(Ty: Tys.back())); |
| 519 | ++CurIdx; |
| 520 | } |
| 521 | Tys.push_back(x: Ty); |
| 522 | Inits.push_back(x: Globals[j]->getInitializer()); |
| 523 | StructIdxs.push_back(x: CurIdx++); |
| 524 | |
| 525 | MaxAlign = std::max(a: MaxAlign, b: Alignment); |
| 526 | |
| 527 | if (Globals[j]->hasExternalLinkage() && !HasExternal) { |
| 528 | HasExternal = true; |
| 529 | FirstExternalName = Globals[j]->getName(); |
| 530 | } |
| 531 | } |
| 532 | |
| 533 | // Exit early if there is only one global to merge. |
| 534 | if (Tys.size() < 2) { |
| 535 | i = j; |
| 536 | continue; |
| 537 | } |
| 538 | |
| 539 | // If merged variables doesn't have external linkage, we needn't to expose |
| 540 | // the symbol after merging. |
| 541 | GlobalValue::LinkageTypes Linkage = HasExternal |
| 542 | ? GlobalValue::ExternalLinkage |
| 543 | : GlobalValue::InternalLinkage; |
| 544 | // Use a packed struct so we can control alignment. |
| 545 | StructType *MergedTy = StructType::get(Context&: M.getContext(), Elements: Tys, isPacked: true); |
| 546 | Constant *MergedInit = ConstantStruct::get(T: MergedTy, V: Inits); |
| 547 | |
| 548 | // On Darwin external linkage needs to be preserved, otherwise |
| 549 | // dsymutil cannot preserve the debug info for the merged |
| 550 | // variables. If they have external linkage, use the symbol name |
| 551 | // of the first variable merged as the suffix of global symbol |
| 552 | // name. This avoids a link-time naming conflict for the |
| 553 | // _MergedGlobals symbols. |
| 554 | Twine MergedName = |
| 555 | (IsMachO && HasExternal) |
| 556 | ? "_MergedGlobals_" + FirstExternalName |
| 557 | : "_MergedGlobals" ; |
| 558 | auto MergedLinkage = IsMachO ? Linkage : GlobalValue::PrivateLinkage; |
| 559 | auto *MergedGV = new GlobalVariable( |
| 560 | M, MergedTy, isConst, MergedLinkage, MergedInit, MergedName, nullptr, |
| 561 | GlobalVariable::NotThreadLocal, AddrSpace); |
| 562 | |
| 563 | MergedGV->setAlignment(MaxAlign); |
| 564 | MergedGV->setSection(Globals[i]->getSection()); |
| 565 | |
| 566 | LLVM_DEBUG(dbgs() << "MergedGV: " << *MergedGV << "\n" ); |
| 567 | |
| 568 | const StructLayout *MergedLayout = DL.getStructLayout(Ty: MergedTy); |
| 569 | for (ssize_t k = i, idx = 0; k != j; k = GlobalSet.find_next(Prev: k), ++idx) { |
| 570 | GlobalValue::LinkageTypes Linkage = Globals[k]->getLinkage(); |
| 571 | std::string Name(Globals[k]->getName()); |
| 572 | GlobalValue::VisibilityTypes Visibility = Globals[k]->getVisibility(); |
| 573 | GlobalValue::DLLStorageClassTypes DLLStorage = |
| 574 | Globals[k]->getDLLStorageClass(); |
| 575 | |
| 576 | // Copy metadata while adjusting any debug info metadata by the original |
| 577 | // global's offset within the merged global. |
| 578 | MergedGV->copyMetadata(Src: Globals[k], |
| 579 | Offset: MergedLayout->getElementOffset(Idx: StructIdxs[idx])); |
| 580 | |
| 581 | Constant *Idx[2] = { |
| 582 | ConstantInt::get(Ty: Int32Ty, V: 0), |
| 583 | ConstantInt::get(Ty: Int32Ty, V: StructIdxs[idx]), |
| 584 | }; |
| 585 | Constant *GEP = |
| 586 | ConstantExpr::getInBoundsGetElementPtr(Ty: MergedTy, C: MergedGV, IdxList: Idx); |
| 587 | Globals[k]->replaceAllUsesWith(V: GEP); |
| 588 | Globals[k]->eraseFromParent(); |
| 589 | |
| 590 | // Emit an alias for the original variable name. This is necessary for an |
| 591 | // external symbol, as it may be accessed from another object. For |
| 592 | // internal symbols, it's not strictly required, but it's useful. |
| 593 | // |
| 594 | // This _should_ also work on Mach-O ever since '.alt_entry' support was |
| 595 | // added in 2016. Unfortunately, there's a bug in ld-prime (present at |
| 596 | // least from Xcode 15.0 through Xcode 16.0), in which -dead_strip doesn't |
| 597 | // always honor alt_entry. To workaround this issue, we don't emit aliases |
| 598 | // on Mach-O. Except, we _must_ do so for external symbols. That means |
| 599 | // MergeExternal is broken with that linker. (That option is currently off |
| 600 | // by default on MachO). |
| 601 | if (!IsMachO || Linkage == GlobalValue::ExternalLinkage) { |
| 602 | GlobalAlias *GA = GlobalAlias::create(Ty: Tys[StructIdxs[idx]], AddressSpace: AddrSpace, |
| 603 | Linkage, Name, Aliasee: GEP, Parent: &M); |
| 604 | GA->setVisibility(Visibility); |
| 605 | GA->setDLLStorageClass(DLLStorage); |
| 606 | } |
| 607 | |
| 608 | NumMerged++; |
| 609 | } |
| 610 | Changed = true; |
| 611 | i = j; |
| 612 | } |
| 613 | |
| 614 | return Changed; |
| 615 | } |
| 616 | |
| 617 | void GlobalMergeImpl::collectUsedGlobalVariables(Module &M, StringRef Name) { |
| 618 | // Extract global variables from llvm.used array |
| 619 | const GlobalVariable *GV = M.getGlobalVariable(Name); |
| 620 | if (!GV || !GV->hasInitializer()) return; |
| 621 | |
| 622 | // Should be an array of 'i8*'. |
| 623 | const ConstantArray *InitList = cast<ConstantArray>(Val: GV->getInitializer()); |
| 624 | |
| 625 | for (unsigned i = 0, e = InitList->getNumOperands(); i != e; ++i) |
| 626 | if (const GlobalVariable *G = |
| 627 | dyn_cast<GlobalVariable>(Val: InitList->getOperand(i_nocapture: i)->stripPointerCasts())) |
| 628 | MustKeepGlobalVariables.insert(X: G); |
| 629 | } |
| 630 | |
| 631 | void GlobalMergeImpl::setMustKeepGlobalVariables(Module &M) { |
| 632 | collectUsedGlobalVariables(M, Name: "llvm.used" ); |
| 633 | collectUsedGlobalVariables(M, Name: "llvm.compiler.used" ); |
| 634 | |
| 635 | for (Function &F : M) { |
| 636 | for (BasicBlock &BB : F) { |
| 637 | BasicBlock::iterator Pad = BB.getFirstNonPHIIt(); |
| 638 | auto *II = dyn_cast<IntrinsicInst>(Val&: Pad); |
| 639 | if (!Pad->isEHPad() && |
| 640 | !(II && II->getIntrinsicID() == Intrinsic::eh_typeid_for)) |
| 641 | continue; |
| 642 | |
| 643 | // Keep globals used by landingpads, catchpads, |
| 644 | // or intrinsics that require a plain global. |
| 645 | for (const Use &U : Pad->operands()) { |
| 646 | if (const GlobalVariable *GV = |
| 647 | dyn_cast<GlobalVariable>(Val: U->stripPointerCasts())) |
| 648 | MustKeepGlobalVariables.insert(X: GV); |
| 649 | else if (const ConstantArray *CA = dyn_cast<ConstantArray>(Val: U->stripPointerCasts())) { |
| 650 | for (const Use &Elt : CA->operands()) { |
| 651 | if (const GlobalVariable *GV = |
| 652 | dyn_cast<GlobalVariable>(Val: Elt->stripPointerCasts())) |
| 653 | MustKeepGlobalVariables.insert(X: GV); |
| 654 | } |
| 655 | } |
| 656 | } |
| 657 | } |
| 658 | } |
| 659 | } |
| 660 | |
| 661 | // This function returns true if the given data Section name has custom |
| 662 | // subsection-splitting semantics in Mach-O (such as splitting by a fixed size) |
| 663 | // |
| 664 | // See also ObjFile::parseSections and getRecordSize in lld/MachO/InputFiles.cpp |
| 665 | static bool isSpecialMachOSection(StringRef Section) { |
| 666 | // Uses starts_with, since section attributes can appear at the end of the |
| 667 | // name. |
| 668 | return Section.starts_with(Prefix: "__DATA,__cfstring" ) || |
| 669 | Section.starts_with(Prefix: "__DATA,__objc_classrefs" ) || |
| 670 | Section.starts_with(Prefix: "__DATA,__objc_selrefs" ); |
| 671 | } |
| 672 | |
| 673 | bool GlobalMergeImpl::run(Module &M) { |
| 674 | if (!EnableGlobalMerge) |
| 675 | return false; |
| 676 | |
| 677 | IsMachO = M.getTargetTriple().isOSBinFormatMachO(); |
| 678 | |
| 679 | auto &DL = M.getDataLayout(); |
| 680 | MapVector<std::pair<unsigned, StringRef>, SmallVector<GlobalVariable *, 0>> |
| 681 | Globals, ConstGlobals, BSSGlobals; |
| 682 | bool Changed = false; |
| 683 | setMustKeepGlobalVariables(M); |
| 684 | |
| 685 | LLVM_DEBUG({ |
| 686 | dbgs() << "Number of GV that must be kept: " << |
| 687 | MustKeepGlobalVariables.size() << "\n" ; |
| 688 | for (const GlobalVariable *KeptGV : MustKeepGlobalVariables) |
| 689 | dbgs() << "Kept: " << *KeptGV << "\n" ; |
| 690 | }); |
| 691 | // Grab all non-const globals. |
| 692 | for (auto &GV : M.globals()) { |
| 693 | // Merge is safe for "normal" internal or external globals only |
| 694 | if (GV.isDeclaration() || GV.isThreadLocal() || GV.hasImplicitSection()) |
| 695 | continue; |
| 696 | |
| 697 | // It's not safe to merge globals that may be preempted |
| 698 | if (TM && !TM->shouldAssumeDSOLocal(GV: &GV)) |
| 699 | continue; |
| 700 | |
| 701 | if (!(Opt.MergeExternal && GV.hasExternalLinkage()) && |
| 702 | !GV.hasLocalLinkage()) |
| 703 | continue; |
| 704 | |
| 705 | PointerType *PT = dyn_cast<PointerType>(Val: GV.getType()); |
| 706 | assert(PT && "Global variable is not a pointer!" ); |
| 707 | |
| 708 | unsigned AddressSpace = PT->getAddressSpace(); |
| 709 | StringRef Section = GV.getSection(); |
| 710 | |
| 711 | // On Mach-O, some section names have special semantics. Don't merge these. |
| 712 | if (IsMachO && isSpecialMachOSection(Section)) |
| 713 | continue; |
| 714 | |
| 715 | // Ignore all 'special' globals. |
| 716 | if (GV.getName().starts_with(Prefix: "llvm." ) || |
| 717 | GV.getName().starts_with(Prefix: ".llvm." ) || Section == "llvm.metadata" ) |
| 718 | continue; |
| 719 | |
| 720 | // Ignore all "required" globals: |
| 721 | if (isMustKeepGlobalVariable(GV: &GV)) |
| 722 | continue; |
| 723 | |
| 724 | // Don't merge tagged globals, as each global should have its own unique |
| 725 | // memory tag at runtime. TODO(hctim): This can be relaxed: constant globals |
| 726 | // with compatible alignment and the same contents may be merged as long as |
| 727 | // the globals occupy the same number of tag granules (i.e. `size_a / 16 == |
| 728 | // size_b / 16`). |
| 729 | if (GV.isTagged()) |
| 730 | continue; |
| 731 | |
| 732 | Type *Ty = GV.getValueType(); |
| 733 | TypeSize AllocSize = DL.getTypeAllocSize(Ty); |
| 734 | bool CanMerge = AllocSize < Opt.MaxOffset && AllocSize >= Opt.MinSize; |
| 735 | if (CanMerge) { |
| 736 | if (TM && |
| 737 | TargetLoweringObjectFile::getKindForGlobal(GO: &GV, TM: *TM).isBSS()) |
| 738 | BSSGlobals[{AddressSpace, Section}].push_back(Elt: &GV); |
| 739 | else if (GV.isConstant()) |
| 740 | ConstGlobals[{AddressSpace, Section}].push_back(Elt: &GV); |
| 741 | else |
| 742 | Globals[{AddressSpace, Section}].push_back(Elt: &GV); |
| 743 | } |
| 744 | LLVM_DEBUG(dbgs() << "GV " << (CanMerge ? "" : "not " ) << "to merge: " << GV |
| 745 | << "\n" ); |
| 746 | } |
| 747 | |
| 748 | for (auto &P : Globals) |
| 749 | if (P.second.size() > 1) |
| 750 | Changed |= doMerge(Globals&: P.second, M, isConst: false, AddrSpace: P.first.first); |
| 751 | |
| 752 | for (auto &P : BSSGlobals) |
| 753 | if (P.second.size() > 1) |
| 754 | Changed |= doMerge(Globals&: P.second, M, isConst: false, AddrSpace: P.first.first); |
| 755 | |
| 756 | if (Opt.MergeConstantGlobals) |
| 757 | for (auto &P : ConstGlobals) |
| 758 | if (P.second.size() > 1) |
| 759 | Changed |= doMerge(Globals&: P.second, M, isConst: true, AddrSpace: P.first.first); |
| 760 | |
| 761 | return Changed; |
| 762 | } |
| 763 | |
| 764 | Pass *llvm::createGlobalMergePass(const TargetMachine *TM, unsigned Offset, |
| 765 | bool OnlyOptimizeForSize, |
| 766 | bool MergeExternalByDefault, |
| 767 | bool MergeConstantByDefault, |
| 768 | bool MergeConstAggressiveByDefault) { |
| 769 | bool MergeExternal = (EnableGlobalMergeOnExternal == cl::BOU_UNSET) ? |
| 770 | MergeExternalByDefault : (EnableGlobalMergeOnExternal == cl::BOU_TRUE); |
| 771 | bool MergeConstant = EnableGlobalMergeOnConst || MergeConstantByDefault; |
| 772 | bool MergeConstAggressive = GlobalMergeAllConst.getNumOccurrences() > 0 |
| 773 | ? GlobalMergeAllConst |
| 774 | : MergeConstAggressiveByDefault; |
| 775 | return new GlobalMerge(TM, Offset, OnlyOptimizeForSize, MergeExternal, |
| 776 | MergeConstant, MergeConstAggressive); |
| 777 | } |
| 778 | |