xref: /freebsd/contrib/llvm-project/lld/COFF/ICF.cpp (revision a90b9d0159070121c221b966469c3e36d912bf82)
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 "COFFLinkerContext.h"
22 #include "Chunks.h"
23 #include "Symbols.h"
24 #include "lld/Common/ErrorHandler.h"
25 #include "lld/Common/Timer.h"
26 #include "llvm/ADT/Hashing.h"
27 #include "llvm/Support/Debug.h"
28 #include "llvm/Support/Parallel.h"
29 #include "llvm/Support/TimeProfiler.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include "llvm/Support/xxhash.h"
32 #include <algorithm>
33 #include <atomic>
34 #include <vector>
35 
36 using namespace llvm;
37 
38 namespace lld::coff {
39 
40 class ICF {
41 public:
42   ICF(COFFLinkerContext &c) : ctx(c){};
43   void run();
44 
45 private:
46   void segregate(size_t begin, size_t end, bool constant);
47 
48   bool assocEquals(const SectionChunk *a, const SectionChunk *b);
49 
50   bool equalsConstant(const SectionChunk *a, const SectionChunk *b);
51   bool equalsVariable(const SectionChunk *a, const SectionChunk *b);
52 
53   bool isEligible(SectionChunk *c);
54 
55   size_t findBoundary(size_t begin, size_t end);
56 
57   void forEachClassRange(size_t begin, size_t end,
58                          std::function<void(size_t, size_t)> fn);
59 
60   void forEachClass(std::function<void(size_t, size_t)> fn);
61 
62   std::vector<SectionChunk *> chunks;
63   int cnt = 0;
64   std::atomic<bool> repeat = {false};
65 
66   COFFLinkerContext &ctx;
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 ((ctx.config.doICF == 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('$').first;
93   if (outSecName == ".pdata" || outSecName == ".xdata")
94     return true;
95 
96   // So are vtables.
97   const char *itaniumVtablePrefix =
98       ctx.config.machine == I386 ? "__ZTV" : "_ZTV";
99   if (c->sym && (c->sym->getName().starts_with("??_7") ||
100                  c->sym->getName().starts_with(itaniumVtablePrefix)))
101     return true;
102 
103   // Anything else not in an address-significance table is eligible.
104   return !c->keepUnique;
105 }
106 
107 // Split an equivalence class into smaller classes.
108 void ICF::segregate(size_t begin, size_t end, bool constant) {
109   while (begin < end) {
110     // Divide [Begin, End) into two. Let Mid be the start index of the
111     // second group.
112     auto bound = std::stable_partition(
113         chunks.begin() + begin + 1, chunks.begin() + end, [&](SectionChunk *s) {
114           if (constant)
115             return equalsConstant(chunks[begin], s);
116           return equalsVariable(chunks[begin], s);
117         });
118     size_t mid = bound - chunks.begin();
119 
120     // Split [Begin, End) into [Begin, Mid) and [Mid, End). We use Mid as an
121     // equivalence class ID because every group ends with a unique index.
122     for (size_t i = begin; i < mid; ++i)
123       chunks[i]->eqClass[(cnt + 1) % 2] = mid;
124 
125     // If we created a group, we need to iterate the main loop again.
126     if (mid != end)
127       repeat = true;
128 
129     begin = mid;
130   }
131 }
132 
133 // Returns true if two sections' associative children are equal.
134 bool ICF::assocEquals(const SectionChunk *a, const SectionChunk *b) {
135   // Ignore associated metadata sections that don't participate in ICF, such as
136   // debug info and CFGuard metadata.
137   auto considerForICF = [](const SectionChunk &assoc) {
138     StringRef Name = assoc.getSectionName();
139     return !(Name.starts_with(".debug") || Name == ".gfids$y" ||
140              Name == ".giats$y" || Name == ".gljmp$y");
141   };
142   auto ra = make_filter_range(a->children(), considerForICF);
143   auto rb = make_filter_range(b->children(), considerForICF);
144   return std::equal(ra.begin(), ra.end(), rb.begin(), rb.end(),
145                     [&](const SectionChunk &ia, const SectionChunk &ib) {
146                       return ia.eqClass[cnt % 2] == ib.eqClass[cnt % 2];
147                     });
148 }
149 
150 // Compare "non-moving" part of two sections, namely everything
151 // except relocation targets.
152 bool ICF::equalsConstant(const SectionChunk *a, const SectionChunk *b) {
153   if (a->relocsSize != b->relocsSize)
154     return false;
155 
156   // Compare relocations.
157   auto eq = [&](const coff_relocation &r1, const coff_relocation &r2) {
158     if (r1.Type != r2.Type ||
159         r1.VirtualAddress != r2.VirtualAddress) {
160       return false;
161     }
162     Symbol *b1 = a->file->getSymbol(r1.SymbolTableIndex);
163     Symbol *b2 = b->file->getSymbol(r2.SymbolTableIndex);
164     if (b1 == b2)
165       return true;
166     if (auto *d1 = dyn_cast<DefinedRegular>(b1))
167       if (auto *d2 = dyn_cast<DefinedRegular>(b2))
168         return d1->getValue() == d2->getValue() &&
169                d1->getChunk()->eqClass[cnt % 2] == d2->getChunk()->eqClass[cnt % 2];
170     return false;
171   };
172   if (!std::equal(a->getRelocs().begin(), a->getRelocs().end(),
173                   b->getRelocs().begin(), eq))
174     return false;
175 
176   // Compare section attributes and contents.
177   return a->getOutputCharacteristics() == b->getOutputCharacteristics() &&
178          a->getSectionName() == b->getSectionName() &&
179          a->header->SizeOfRawData == b->header->SizeOfRawData &&
180          a->checksum == b->checksum && a->getContents() == b->getContents() &&
181          assocEquals(a, b);
182 }
183 
184 // Compare "moving" part of two sections, namely relocation targets.
185 bool ICF::equalsVariable(const SectionChunk *a, const SectionChunk *b) {
186   // Compare relocations.
187   auto eq = [&](const coff_relocation &r1, const coff_relocation &r2) {
188     Symbol *b1 = a->file->getSymbol(r1.SymbolTableIndex);
189     Symbol *b2 = b->file->getSymbol(r2.SymbolTableIndex);
190     if (b1 == b2)
191       return true;
192     if (auto *d1 = dyn_cast<DefinedRegular>(b1))
193       if (auto *d2 = dyn_cast<DefinedRegular>(b2))
194         return d1->getChunk()->eqClass[cnt % 2] == d2->getChunk()->eqClass[cnt % 2];
195     return false;
196   };
197   return std::equal(a->getRelocs().begin(), a->getRelocs().end(),
198                     b->getRelocs().begin(), eq) &&
199          assocEquals(a, b);
200 }
201 
202 // Find the first Chunk after Begin that has a different class from Begin.
203 size_t ICF::findBoundary(size_t begin, size_t end) {
204   for (size_t i = begin + 1; i < end; ++i)
205     if (chunks[begin]->eqClass[cnt % 2] != chunks[i]->eqClass[cnt % 2])
206       return i;
207   return end;
208 }
209 
210 void ICF::forEachClassRange(size_t begin, size_t end,
211                             std::function<void(size_t, size_t)> fn) {
212   while (begin < end) {
213     size_t mid = findBoundary(begin, end);
214     fn(begin, mid);
215     begin = mid;
216   }
217 }
218 
219 // Call Fn on each class group.
220 void ICF::forEachClass(std::function<void(size_t, size_t)> fn) {
221   // If the number of sections are too small to use threading,
222   // call Fn sequentially.
223   if (chunks.size() < 1024) {
224     forEachClassRange(0, chunks.size(), fn);
225     ++cnt;
226     return;
227   }
228 
229   // Shard into non-overlapping intervals, and call Fn in parallel.
230   // The sharding must be completed before any calls to Fn are made
231   // so that Fn can modify the Chunks in its shard without causing data
232   // races.
233   const size_t numShards = 256;
234   size_t step = chunks.size() / numShards;
235   size_t boundaries[numShards + 1];
236   boundaries[0] = 0;
237   boundaries[numShards] = chunks.size();
238   parallelFor(1, numShards, [&](size_t i) {
239     boundaries[i] = findBoundary((i - 1) * step, chunks.size());
240   });
241   parallelFor(1, numShards + 1, [&](size_t i) {
242     if (boundaries[i - 1] < boundaries[i]) {
243       forEachClassRange(boundaries[i - 1], boundaries[i], fn);
244     }
245   });
246   ++cnt;
247 }
248 
249 // Merge identical COMDAT sections.
250 // Two sections are considered the same if their section headers,
251 // contents and relocations are all the same.
252 void ICF::run() {
253   llvm::TimeTraceScope timeScope("ICF");
254   ScopedTimer t(ctx.icfTimer);
255 
256   // Collect only mergeable sections and group by hash value.
257   uint32_t nextId = 1;
258   for (Chunk *c : ctx.symtab.getChunks()) {
259     if (auto *sc = dyn_cast<SectionChunk>(c)) {
260       if (isEligible(sc))
261         chunks.push_back(sc);
262       else
263         sc->eqClass[0] = nextId++;
264     }
265   }
266 
267   // Make sure that ICF doesn't merge sections that are being handled by string
268   // tail merging.
269   for (MergeChunk *mc : ctx.mergeChunkInstances)
270     if (mc)
271       for (SectionChunk *sc : mc->sections)
272         sc->eqClass[0] = nextId++;
273 
274   // Initially, we use hash values to partition sections.
275   parallelForEach(chunks, [&](SectionChunk *sc) {
276     sc->eqClass[0] = xxh3_64bits(sc->getContents());
277   });
278 
279   // Combine the hashes of the sections referenced by each section into its
280   // hash.
281   for (unsigned cnt = 0; cnt != 2; ++cnt) {
282     parallelForEach(chunks, [&](SectionChunk *sc) {
283       uint32_t hash = sc->eqClass[cnt % 2];
284       for (Symbol *b : sc->symbols())
285         if (auto *sym = dyn_cast_or_null<DefinedRegular>(b))
286           hash += sym->getChunk()->eqClass[cnt % 2];
287       // Set MSB to 1 to avoid collisions with non-hash classes.
288       sc->eqClass[(cnt + 1) % 2] = hash | (1U << 31);
289     });
290   }
291 
292   // From now on, sections in Chunks are ordered so that sections in
293   // the same group are consecutive in the vector.
294   llvm::stable_sort(chunks, [](const SectionChunk *a, const SectionChunk *b) {
295     return a->eqClass[0] < b->eqClass[0];
296   });
297 
298   // Compare static contents and assign unique IDs for each static content.
299   forEachClass([&](size_t begin, size_t end) { segregate(begin, end, true); });
300 
301   // Split groups by comparing relocations until convergence is obtained.
302   do {
303     repeat = false;
304     forEachClass(
305         [&](size_t begin, size_t end) { segregate(begin, end, false); });
306   } while (repeat);
307 
308   log("ICF needed " + Twine(cnt) + " iterations");
309 
310   // Merge sections in the same classes.
311   forEachClass([&](size_t begin, size_t end) {
312     if (end - begin == 1)
313       return;
314 
315     log("Selected " + chunks[begin]->getDebugName());
316     for (size_t i = begin + 1; i < end; ++i) {
317       log("  Removed " + chunks[i]->getDebugName());
318       chunks[begin]->replace(chunks[i]);
319     }
320   });
321 }
322 
323 // Entry point to ICF.
324 void doICF(COFFLinkerContext &ctx) { ICF(ctx).run(); }
325 
326 } // namespace lld::coff
327