xref: /freebsd/contrib/llvm-project/lld/COFF/ICF.cpp (revision 480093f4440d54b30b3025afeac24b48f2ba7a2e)
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/Threads.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/raw_ostream.h"
30 #include "llvm/Support/xxhash.h"
31 #include <algorithm>
32 #include <atomic>
33 #include <vector>
34 
35 using namespace llvm;
36 
37 namespace lld {
38 namespace coff {
39 
40 static Timer icfTimer("ICF", Timer::root());
41 
42 class ICF {
43 public:
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 };
67 
68 // Returns true if section S is subject of ICF.
69 //
70 // Microsoft's documentation
71 // (https://msdn.microsoft.com/en-us/library/bxwfs976.aspx; visited April
72 // 2017) says that /opt:icf folds both functions and read-only data.
73 // Despite that, the MSVC linker folds only functions. We found
74 // a few instances of programs that are not safe for data merging.
75 // Therefore, we merge only functions just like the MSVC tool. However, we also
76 // merge read-only sections in a couple of cases where the address of the
77 // section is insignificant to the user program and the behaviour matches that
78 // of the Visual C++ linker.
79 bool ICF::isEligible(SectionChunk *c) {
80   // Non-comdat chunks, dead chunks, and writable chunks are not eligible.
81   bool writable = c->getOutputCharacteristics() & llvm::COFF::IMAGE_SCN_MEM_WRITE;
82   if (!c->isCOMDAT() || !c->live || writable)
83     return false;
84 
85   // Code sections are eligible.
86   if (c->getOutputCharacteristics() & llvm::COFF::IMAGE_SCN_MEM_EXECUTE)
87     return true;
88 
89   // .pdata and .xdata unwind info sections are eligible.
90   StringRef outSecName = c->getSectionName().split('$').first;
91   if (outSecName == ".pdata" || outSecName == ".xdata")
92     return true;
93 
94   // So are vtables.
95   if (c->sym && c->sym->getName().startswith("??_7"))
96     return true;
97 
98   // Anything else not in an address-significance table is eligible.
99   return !c->keepUnique;
100 }
101 
102 // Split an equivalence class into smaller classes.
103 void ICF::segregate(size_t begin, size_t end, bool constant) {
104   while (begin < end) {
105     // Divide [Begin, End) into two. Let Mid be the start index of the
106     // second group.
107     auto bound = std::stable_partition(
108         chunks.begin() + begin + 1, chunks.begin() + end, [&](SectionChunk *s) {
109           if (constant)
110             return equalsConstant(chunks[begin], s);
111           return equalsVariable(chunks[begin], s);
112         });
113     size_t mid = bound - chunks.begin();
114 
115     // Split [Begin, End) into [Begin, Mid) and [Mid, End). We use Mid as an
116     // equivalence class ID because every group ends with a unique index.
117     for (size_t i = begin; i < mid; ++i)
118       chunks[i]->eqClass[(cnt + 1) % 2] = mid;
119 
120     // If we created a group, we need to iterate the main loop again.
121     if (mid != end)
122       repeat = true;
123 
124     begin = mid;
125   }
126 }
127 
128 // Returns true if two sections' associative children are equal.
129 bool ICF::assocEquals(const SectionChunk *a, const SectionChunk *b) {
130   auto childClasses = [&](const SectionChunk *sc) {
131     std::vector<uint32_t> classes;
132     for (const SectionChunk &c : sc->children())
133       if (!c.getSectionName().startswith(".debug") &&
134           c.getSectionName() != ".gfids$y" && c.getSectionName() != ".gljmp$y")
135         classes.push_back(c.eqClass[cnt % 2]);
136     return classes;
137   };
138   return childClasses(a) == childClasses(b);
139 }
140 
141 // Compare "non-moving" part of two sections, namely everything
142 // except relocation targets.
143 bool ICF::equalsConstant(const SectionChunk *a, const SectionChunk *b) {
144   if (a->relocsSize != b->relocsSize)
145     return false;
146 
147   // Compare relocations.
148   auto eq = [&](const coff_relocation &r1, const coff_relocation &r2) {
149     if (r1.Type != r2.Type ||
150         r1.VirtualAddress != r2.VirtualAddress) {
151       return false;
152     }
153     Symbol *b1 = a->file->getSymbol(r1.SymbolTableIndex);
154     Symbol *b2 = b->file->getSymbol(r2.SymbolTableIndex);
155     if (b1 == b2)
156       return true;
157     if (auto *d1 = dyn_cast<DefinedRegular>(b1))
158       if (auto *d2 = dyn_cast<DefinedRegular>(b2))
159         return d1->getValue() == d2->getValue() &&
160                d1->getChunk()->eqClass[cnt % 2] == d2->getChunk()->eqClass[cnt % 2];
161     return false;
162   };
163   if (!std::equal(a->getRelocs().begin(), a->getRelocs().end(),
164                   b->getRelocs().begin(), eq))
165     return false;
166 
167   // Compare section attributes and contents.
168   return a->getOutputCharacteristics() == b->getOutputCharacteristics() &&
169          a->getSectionName() == b->getSectionName() &&
170          a->header->SizeOfRawData == b->header->SizeOfRawData &&
171          a->checksum == b->checksum && a->getContents() == b->getContents() &&
172          assocEquals(a, b);
173 }
174 
175 // Compare "moving" part of two sections, namely relocation targets.
176 bool ICF::equalsVariable(const SectionChunk *a, const SectionChunk *b) {
177   // Compare relocations.
178   auto eq = [&](const coff_relocation &r1, const coff_relocation &r2) {
179     Symbol *b1 = a->file->getSymbol(r1.SymbolTableIndex);
180     Symbol *b2 = b->file->getSymbol(r2.SymbolTableIndex);
181     if (b1 == b2)
182       return true;
183     if (auto *d1 = dyn_cast<DefinedRegular>(b1))
184       if (auto *d2 = dyn_cast<DefinedRegular>(b2))
185         return d1->getChunk()->eqClass[cnt % 2] == d2->getChunk()->eqClass[cnt % 2];
186     return false;
187   };
188   return std::equal(a->getRelocs().begin(), a->getRelocs().end(),
189                     b->getRelocs().begin(), eq) &&
190          assocEquals(a, b);
191 }
192 
193 // Find the first Chunk after Begin that has a different class from Begin.
194 size_t ICF::findBoundary(size_t begin, size_t end) {
195   for (size_t i = begin + 1; i < end; ++i)
196     if (chunks[begin]->eqClass[cnt % 2] != chunks[i]->eqClass[cnt % 2])
197       return i;
198   return end;
199 }
200 
201 void ICF::forEachClassRange(size_t begin, size_t end,
202                             std::function<void(size_t, size_t)> fn) {
203   while (begin < end) {
204     size_t mid = findBoundary(begin, end);
205     fn(begin, mid);
206     begin = mid;
207   }
208 }
209 
210 // Call Fn on each class group.
211 void ICF::forEachClass(std::function<void(size_t, size_t)> fn) {
212   // If the number of sections are too small to use threading,
213   // call Fn sequentially.
214   if (chunks.size() < 1024) {
215     forEachClassRange(0, chunks.size(), fn);
216     ++cnt;
217     return;
218   }
219 
220   // Shard into non-overlapping intervals, and call Fn in parallel.
221   // The sharding must be completed before any calls to Fn are made
222   // so that Fn can modify the Chunks in its shard without causing data
223   // races.
224   const size_t numShards = 256;
225   size_t step = chunks.size() / numShards;
226   size_t boundaries[numShards + 1];
227   boundaries[0] = 0;
228   boundaries[numShards] = chunks.size();
229   parallelForEachN(1, numShards, [&](size_t i) {
230     boundaries[i] = findBoundary((i - 1) * step, chunks.size());
231   });
232   parallelForEachN(1, numShards + 1, [&](size_t i) {
233     if (boundaries[i - 1] < boundaries[i]) {
234       forEachClassRange(boundaries[i - 1], boundaries[i], fn);
235     }
236   });
237   ++cnt;
238 }
239 
240 // Merge identical COMDAT sections.
241 // Two sections are considered the same if their section headers,
242 // contents and relocations are all the same.
243 void ICF::run(ArrayRef<Chunk *> vec) {
244   ScopedTimer t(icfTimer);
245 
246   // Collect only mergeable sections and group by hash value.
247   uint32_t nextId = 1;
248   for (Chunk *c : vec) {
249     if (auto *sc = dyn_cast<SectionChunk>(c)) {
250       if (isEligible(sc))
251         chunks.push_back(sc);
252       else
253         sc->eqClass[0] = nextId++;
254     }
255   }
256 
257   // Make sure that ICF doesn't merge sections that are being handled by string
258   // tail merging.
259   for (MergeChunk *mc : MergeChunk::instances)
260     if (mc)
261       for (SectionChunk *sc : mc->sections)
262         sc->eqClass[0] = nextId++;
263 
264   // Initially, we use hash values to partition sections.
265   parallelForEach(chunks, [&](SectionChunk *sc) {
266     sc->eqClass[0] = xxHash64(sc->getContents());
267   });
268 
269   // Combine the hashes of the sections referenced by each section into its
270   // hash.
271   for (unsigned cnt = 0; cnt != 2; ++cnt) {
272     parallelForEach(chunks, [&](SectionChunk *sc) {
273       uint32_t hash = sc->eqClass[cnt % 2];
274       for (Symbol *b : sc->symbols())
275         if (auto *sym = dyn_cast_or_null<DefinedRegular>(b))
276           hash += sym->getChunk()->eqClass[cnt % 2];
277       // Set MSB to 1 to avoid collisions with non-hash classes.
278       sc->eqClass[(cnt + 1) % 2] = hash | (1U << 31);
279     });
280   }
281 
282   // From now on, sections in Chunks are ordered so that sections in
283   // the same group are consecutive in the vector.
284   llvm::stable_sort(chunks, [](const SectionChunk *a, const SectionChunk *b) {
285     return a->eqClass[0] < b->eqClass[0];
286   });
287 
288   // Compare static contents and assign unique IDs for each static content.
289   forEachClass([&](size_t begin, size_t end) { segregate(begin, end, true); });
290 
291   // Split groups by comparing relocations until convergence is obtained.
292   do {
293     repeat = false;
294     forEachClass(
295         [&](size_t begin, size_t end) { segregate(begin, end, false); });
296   } while (repeat);
297 
298   log("ICF needed " + Twine(cnt) + " iterations");
299 
300   // Merge sections in the same classes.
301   forEachClass([&](size_t begin, size_t end) {
302     if (end - begin == 1)
303       return;
304 
305     log("Selected " + chunks[begin]->getDebugName());
306     for (size_t i = begin + 1; i < end; ++i) {
307       log("  Removed " + chunks[i]->getDebugName());
308       chunks[begin]->replace(chunks[i]);
309     }
310   });
311 }
312 
313 // Entry point to ICF.
314 void doICF(ArrayRef<Chunk *> chunks) { ICF().run(chunks); }
315 
316 } // namespace coff
317 } // namespace lld
318