lld/COFF/ICF.cpp

   1 //===- ICF.cpp ------------------------------------------------------------===//
   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 // ICF is short for Identical Code Folding. That is a size optimization to
  10 // identify and merge two or more read-only sections (typically functions)
  11 // that happened to have the same contents. It usually reduces output size
  12 // by a few percent.
  13 //
  14 // On Windows, ICF is enabled by default.
  15 //
  16 // See ELF/ICF.cpp for the details about the algorithm.
  17 //
  18 //===----------------------------------------------------------------------===//
  19
  20 #include "ICF.h"
  21 #include "Chunks.h"
  22 #include "Symbols.h"
  23 #include "lld/Common/ErrorHandler.h"
  24 #include "lld/Common/Timer.h"
  25 #include "llvm/ADT/Hashing.h"
  26 #include "llvm/Support/Debug.h"
  27 #include "llvm/Support/Parallel.h"
  28 #include "llvm/Support/raw_ostream.h"
  29 #include "llvm/Support/xxhash.h"
  30 #include <algorithm>
  31 #include <atomic>
  32 #include <vector>
  33
  34 using namespace llvm;
  35
  36 namespace lld {
  37 namespace coff {
  38
  39 static Timer icfTimer("ICF", Timer::root());
  40
  41 class ICF {
  42 public:
  43   ICF(ICFLevel icfLevel) : icfLevel(icfLevel){};
  44   void run(ArrayRef<Chunk *> v);
  45
  46 private:
  47   void segregate(size_t begin, size_t end, bool constant);
  48
  49   bool assocEquals(const SectionChunk *a, const SectionChunk *b);
  50
  51   bool equalsConstant(const SectionChunk *a, const SectionChunk *b);
  52   bool equalsVariable(const SectionChunk *a, const SectionChunk *b);
  53
  54   bool isEligible(SectionChunk *c);
  55
  56   size_t findBoundary(size_t begin, size_t end);
  57
  58   void forEachClassRange(size_t begin, size_t end,
  59                          std::function<void(size_t, size_t)> fn);
  60
  61   void forEachClass(std::function<void(size_t, size_t)> fn);
  62
  63   std::vector<SectionChunk *> chunks;
  64   int cnt = 0;
  65   std::atomic<bool> repeat = {false};
  66   ICFLevel icfLevel = ICFLevel::All;
  67 };
  68
  69 // Returns true if section S is subject of ICF.
  70 //
  71 // Microsoft's documentation
  72 // (https://msdn.microsoft.com/en-us/library/bxwfs976.aspx; visited April
  73 // 2017) says that /opt:icf folds both functions and read-only data.
  74 // Despite that, the MSVC linker folds only functions. We found
  75 // a few instances of programs that are not safe for data merging.
  76 // Therefore, we merge only functions just like the MSVC tool. However, we also
  77 // merge read-only sections in a couple of cases where the address of the
  78 // section is insignificant to the user program and the behaviour matches that
  79 // of the Visual C++ linker.
  80 bool ICF::isEligible(SectionChunk *c) {
  81   // Non-comdat chunks, dead chunks, and writable chunks are not eligible.
  82   bool writable = c->getOutputCharacteristics() & llvm::COFF::IMAGE_SCN_MEM_WRITE;
  83   if (!c->isCOMDAT() || !c->live || writable)
  84     return false;
  85
  86   // Under regular (not safe) ICF, all code sections are eligible.
  87   if ((icfLevel == ICFLevel::All) &&
  88       c->getOutputCharacteristics() & llvm::COFF::IMAGE_SCN_MEM_EXECUTE)
  89     return true;
  90
  91   // .pdata and .xdata unwind info sections are eligible.
  92   StringRef outSecName = c->getSectionName().split('$').first;
  93   if (outSecName == ".pdata" || outSecName == ".xdata")
  94     return true;
  95
  96   // So are vtables.
  97   if (c->sym && c->sym->getName().startswith("??_7"))
  98     return true;
  99
 100   // Anything else not in an address-significance table is eligible.
 101   return !c->keepUnique;
 102 }
 103
 104 // Split an equivalence class into smaller classes.
 105 void ICF::segregate(size_t begin, size_t end, bool constant) {
 106   while (begin < end) {
 107     // Divide [Begin, End) into two. Let Mid be the start index of the
 108     // second group.
 109     auto bound = std::stable_partition(
 110         chunks.begin() + begin + 1, chunks.begin() + end, [&](SectionChunk *s) {
 111           if (constant)
 112             return equalsConstant(chunks[begin], s);
 113           return equalsVariable(chunks[begin], s);
 114         });
 115     size_t mid = bound - chunks.begin();
 116
 117     // Split [Begin, End) into [Begin, Mid) and [Mid, End). We use Mid as an
 118     // equivalence class ID because every group ends with a unique index.
 119     for (size_t i = begin; i < mid; ++i)
 120       chunks[i]->eqClass[(cnt + 1) % 2] = mid;
 121
 122     // If we created a group, we need to iterate the main loop again.
 123     if (mid != end)
 124       repeat = true;
 125
 126     begin = mid;
 127   }
 128 }
 129
 130 // Returns true if two sections' associative children are equal.
 131 bool ICF::assocEquals(const SectionChunk *a, const SectionChunk *b) {
 132   // Ignore associated metadata sections that don't participate in ICF, such as
 133   // debug info and CFGuard metadata.
 134   auto considerForICF = [](const SectionChunk &assoc) {
 135     StringRef Name = assoc.getSectionName();
 136     return !(Name.startswith(".debug") || Name == ".gfids$y" ||
 137              Name == ".giats$y" || Name == ".gljmp$y");
 138   };
 139   auto ra = make_filter_range(a->children(), considerForICF);
 140   auto rb = make_filter_range(b->children(), considerForICF);
 141   return std::equal(ra.begin(), ra.end(), rb.begin(), rb.end(),
 142                     [&](const SectionChunk &ia, const SectionChunk &ib) {
 143                       return ia.eqClass[cnt % 2] == ib.eqClass[cnt % 2];
 144                     });
 145 }
 146
 147 // Compare "non-moving" part of two sections, namely everything
 148 // except relocation targets.
 149 bool ICF::equalsConstant(const SectionChunk *a, const SectionChunk *b) {
 150   if (a->relocsSize != b->relocsSize)
 151     return false;
 152
 153   // Compare relocations.
 154   auto eq = [&](const coff_relocation &r1, const coff_relocation &r2) {
 155     if (r1.Type != r2.Type ||
 156         r1.VirtualAddress != r2.VirtualAddress) {
 157       return false;
 158     }
 159     Symbol *b1 = a->file->getSymbol(r1.SymbolTableIndex);
 160     Symbol *b2 = b->file->getSymbol(r2.SymbolTableIndex);
 161     if (b1 == b2)
 162       return true;
 163     if (auto *d1 = dyn_cast<DefinedRegular>(b1))
 164       if (auto *d2 = dyn_cast<DefinedRegular>(b2))
 165         return d1->getValue() == d2->getValue() &&
 166                d1->getChunk()->eqClass[cnt % 2] == d2->getChunk()->eqClass[cnt % 2];
 167     return false;
 168   };
 169   if (!std::equal(a->getRelocs().begin(), a->getRelocs().end(),
 170                   b->getRelocs().begin(), eq))
 171     return false;
 172
 173   // Compare section attributes and contents.
 174   return a->getOutputCharacteristics() == b->getOutputCharacteristics() &&
 175          a->getSectionName() == b->getSectionName() &&
 176          a->header->SizeOfRawData == b->header->SizeOfRawData &&
 177          a->checksum == b->checksum && a->getContents() == b->getContents() &&
 178          assocEquals(a, b);
 179 }
 180
 181 // Compare "moving" part of two sections, namely relocation targets.
 182 bool ICF::equalsVariable(const SectionChunk *a, const SectionChunk *b) {
 183   // Compare relocations.
 184   auto eq = [&](const coff_relocation &r1, const coff_relocation &r2) {
 185     Symbol *b1 = a->file->getSymbol(r1.SymbolTableIndex);
 186     Symbol *b2 = b->file->getSymbol(r2.SymbolTableIndex);
 187     if (b1 == b2)
 188       return true;
 189     if (auto *d1 = dyn_cast<DefinedRegular>(b1))
 190       if (auto *d2 = dyn_cast<DefinedRegular>(b2))
 191         return d1->getChunk()->eqClass[cnt % 2] == d2->getChunk()->eqClass[cnt % 2];
 192     return false;
 193   };
 194   return std::equal(a->getRelocs().begin(), a->getRelocs().end(),
 195                     b->getRelocs().begin(), eq) &&
 196          assocEquals(a, b);
 197 }
 198
 199 // Find the first Chunk after Begin that has a different class from Begin.
 200 size_t ICF::findBoundary(size_t begin, size_t end) {
 201   for (size_t i = begin + 1; i < end; ++i)
 202     if (chunks[begin]->eqClass[cnt % 2] != chunks[i]->eqClass[cnt % 2])
 203       return i;
 204   return end;
 205 }
 206
 207 void ICF::forEachClassRange(size_t begin, size_t end,
 208                             std::function<void(size_t, size_t)> fn) {
 209   while (begin < end) {
 210     size_t mid = findBoundary(begin, end);
 211     fn(begin, mid);
 212     begin = mid;
 213   }
 214 }
 215
 216 // Call Fn on each class group.
 217 void ICF::forEachClass(std::function<void(size_t, size_t)> fn) {
 218   // If the number of sections are too small to use threading,
 219   // call Fn sequentially.
 220   if (chunks.size() < 1024) {
 221     forEachClassRange(0, chunks.size(), fn);
 222     ++cnt;
 223     return;
 224   }
 225
 226   // Shard into non-overlapping intervals, and call Fn in parallel.
 227   // The sharding must be completed before any calls to Fn are made
 228   // so that Fn can modify the Chunks in its shard without causing data
 229   // races.
 230   const size_t numShards = 256;
 231   size_t step = chunks.size() / numShards;
 232   size_t boundaries[numShards + 1];
 233   boundaries[0] = 0;
 234   boundaries[numShards] = chunks.size();
 235   parallelForEachN(1, numShards, [&](size_t i) {
 236     boundaries[i] = findBoundary((i - 1) * step, chunks.size());
 237   });
 238   parallelForEachN(1, numShards + 1, [&](size_t i) {
 239     if (boundaries[i - 1] < boundaries[i]) {
 240       forEachClassRange(boundaries[i - 1], boundaries[i], fn);
 241     }
 242   });
 243   ++cnt;
 244 }
 245
 246 // Merge identical COMDAT sections.
 247 // Two sections are considered the same if their section headers,
 248 // contents and relocations are all the same.
 249 void ICF::run(ArrayRef<Chunk *> vec) {
 250   ScopedTimer t(icfTimer);
 251
 252   // Collect only mergeable sections and group by hash value.
 253   uint32_t nextId = 1;
 254   for (Chunk *c : vec) {
 255     if (auto *sc = dyn_cast<SectionChunk>(c)) {
 256       if (isEligible(sc))
 257         chunks.push_back(sc);
 258       else
 259         sc->eqClass[0] = nextId++;
 260     }
 261   }
 262
 263   // Make sure that ICF doesn't merge sections that are being handled by string
 264   // tail merging.
 265   for (MergeChunk *mc : MergeChunk::instances)
 266     if (mc)
 267       for (SectionChunk *sc : mc->sections)
 268         sc->eqClass[0] = nextId++;
 269
 270   // Initially, we use hash values to partition sections.
 271   parallelForEach(chunks, [&](SectionChunk *sc) {
 272     sc->eqClass[0] = xxHash64(sc->getContents());
 273   });
 274
 275   // Combine the hashes of the sections referenced by each section into its
 276   // hash.
 277   for (unsigned cnt = 0; cnt != 2; ++cnt) {
 278     parallelForEach(chunks, [&](SectionChunk *sc) {
 279       uint32_t hash = sc->eqClass[cnt % 2];
 280       for (Symbol *b : sc->symbols())
 281         if (auto *sym = dyn_cast_or_null<DefinedRegular>(b))
 282           hash += sym->getChunk()->eqClass[cnt % 2];
 283       // Set MSB to 1 to avoid collisions with non-hash classes.
 284       sc->eqClass[(cnt + 1) % 2] = hash | (1U << 31);
 285     });
 286   }
 287
 288   // From now on, sections in Chunks are ordered so that sections in
 289   // the same group are consecutive in the vector.
 290   llvm::stable_sort(chunks, [](const SectionChunk *a, const SectionChunk *b) {
 291     return a->eqClass[0] < b->eqClass[0];
 292   });
 293
 294   // Compare static contents and assign unique IDs for each static content.
 295   forEachClass([&](size_t begin, size_t end) { segregate(begin, end, true); });
 296
 297   // Split groups by comparing relocations until convergence is obtained.
 298   do {
 299     repeat = false;
 300     forEachClass(
 301         [&](size_t begin, size_t end) { segregate(begin, end, false); });
 302   } while (repeat);
 303
 304   log("ICF needed " + Twine(cnt) + " iterations");
 305
 306   // Merge sections in the same classes.
 307   forEachClass([&](size_t begin, size_t end) {
 308     if (end - begin == 1)
 309       return;
 310
 311     log("Selected " + chunks[begin]->getDebugName());
 312     for (size_t i = begin + 1; i < end; ++i) {
 313       log("  Removed " + chunks[i]->getDebugName());
 314       chunks[begin]->replace(chunks[i]);
 315     }
 316   });
 317 }
 318
 319 // Entry point to ICF.
 320 void doICF(ArrayRef<Chunk *> chunks, ICFLevel icfLevel) {
 321   ICF(icfLevel).run(chunks);
 322 }
 323
 324 } // namespace coff
 325 } // namespace lld