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. This 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 // In ICF, two sections are considered identical if they have the same
15 // section flags, section data, and relocations. Relocations are tricky,
16 // because two relocations are considered the same if they have the same
17 // relocation types, values, and if they point to the same sections *in
18 // terms of ICF*.
19 //
20 // Here is an example. If foo and bar defined below are compiled to the
21 // same machine instructions, ICF can and should merge the two, although
22 // their relocations point to each other.
23 //
24 // void foo() { bar(); }
25 // void bar() { foo(); }
26 //
27 // If you merge the two, their relocations point to the same section and
28 // thus you know they are mergeable, but how do you know they are
29 // mergeable in the first place? This is not an easy problem to solve.
30 //
31 // What we are doing in LLD is to partition sections into equivalence
32 // classes. Sections in the same equivalence class when the algorithm
33 // terminates are considered identical. Here are details:
34 //
35 // 1. First, we partition sections using their hash values as keys. Hash
36 // values contain section types, section contents and numbers of
37 // relocations. During this step, relocation targets are not taken into
38 // account. We just put sections that apparently differ into different
39 // equivalence classes.
40 //
41 // 2. Next, for each equivalence class, we visit sections to compare
42 // relocation targets. Relocation targets are considered equivalent if
43 // their targets are in the same equivalence class. Sections with
44 // different relocation targets are put into different equivalence
45 // classes.
46 //
47 // 3. If we split an equivalence class in step 2, two relocations
48 // previously target the same equivalence class may now target
49 // different equivalence classes. Therefore, we repeat step 2 until a
50 // convergence is obtained.
51 //
52 // 4. For each equivalence class C, pick an arbitrary section in C, and
53 // merge all the other sections in C with it.
54 //
55 // For small programs, this algorithm needs 3-5 iterations. For large
56 // programs such as Chromium, it takes more than 20 iterations.
57 //
58 // This algorithm was mentioned as an "optimistic algorithm" in [1],
59 // though gold implements a different algorithm than this.
60 //
61 // We parallelize each step so that multiple threads can work on different
62 // equivalence classes concurrently. That gave us a large performance
63 // boost when applying ICF on large programs. For example, MSVC link.exe
64 // or GNU gold takes 10-20 seconds to apply ICF on Chromium, whose output
65 // size is about 1.5 GB, but LLD can finish it in less than 2 seconds on a
66 // 2.8 GHz 40 core machine. Even without threading, LLD's ICF is still
67 // faster than MSVC or gold though.
68 //
69 // [1] Safe ICF: Pointer Safe and Unwinding aware Identical Code Folding
70 // in the Gold Linker
71 // http://static.googleusercontent.com/media/research.google.com/en//pubs/archive/36912.pdf
72 //
73 //===----------------------------------------------------------------------===//
74
75 #include "ICF.h"
76 #include "Config.h"
77 #include "LinkerScript.h"
78 #include "OutputSections.h"
79 #include "SymbolTable.h"
80 #include "Symbols.h"
81 #include "SyntheticSections.h"
82 #include "Writer.h"
83 #include "llvm/ADT/StringExtras.h"
84 #include "llvm/BinaryFormat/ELF.h"
85 #include "llvm/Object/ELF.h"
86 #include "llvm/Support/Parallel.h"
87 #include "llvm/Support/TimeProfiler.h"
88 #include "llvm/Support/xxhash.h"
89 #include <algorithm>
90 #include <atomic>
91
92 using namespace llvm;
93 using namespace llvm::ELF;
94 using namespace llvm::object;
95 using namespace lld;
96 using namespace lld::elf;
97
98 namespace {
99 template <class ELFT> class ICF {
100 public:
101 void run();
102
103 private:
104 void segregate(size_t begin, size_t end, bool constant);
105
106 template <class RelTy>
107 bool constantEq(const InputSection *a, ArrayRef<RelTy> relsA,
108 const InputSection *b, ArrayRef<RelTy> relsB);
109
110 template <class RelTy>
111 bool variableEq(const InputSection *a, ArrayRef<RelTy> relsA,
112 const InputSection *b, ArrayRef<RelTy> relsB);
113
114 bool equalsConstant(const InputSection *a, const InputSection *b);
115 bool equalsVariable(const InputSection *a, const InputSection *b);
116
117 size_t findBoundary(size_t begin, size_t end);
118
119 void forEachClassRange(size_t begin, size_t end,
120 llvm::function_ref<void(size_t, size_t)> fn);
121
122 void forEachClass(llvm::function_ref<void(size_t, size_t)> fn);
123
124 std::vector<InputSection *> sections;
125
126 // We repeat the main loop while `Repeat` is true.
127 std::atomic<bool> repeat;
128
129 // The main loop counter.
130 int cnt = 0;
131
132 // We have two locations for equivalence classes. On the first iteration
133 // of the main loop, Class[0] has a valid value, and Class[1] contains
134 // garbage. We read equivalence classes from slot 0 and write to slot 1.
135 // So, Class[0] represents the current class, and Class[1] represents
136 // the next class. On each iteration, we switch their roles and use them
137 // alternately.
138 //
139 // Why are we doing this? Recall that other threads may be working on
140 // other equivalence classes in parallel. They may read sections that we
141 // are updating. We cannot update equivalence classes in place because
142 // it breaks the invariance that all possibly-identical sections must be
143 // in the same equivalence class at any moment. In other words, the for
144 // loop to update equivalence classes is not atomic, and that is
145 // observable from other threads. By writing new classes to other
146 // places, we can keep the invariance.
147 //
148 // Below, `Current` has the index of the current class, and `Next` has
149 // the index of the next class. If threading is enabled, they are either
150 // (0, 1) or (1, 0).
151 //
152 // Note on single-thread: if that's the case, they are always (0, 0)
153 // because we can safely read the next class without worrying about race
154 // conditions. Using the same location makes this algorithm converge
155 // faster because it uses results of the same iteration earlier.
156 int current = 0;
157 int next = 0;
158 };
159 }
160
161 // Returns true if section S is subject of ICF.
isEligible(InputSection * s)162 static bool isEligible(InputSection *s) {
163 if (!s->isLive() || s->keepUnique || !(s->flags & SHF_ALLOC))
164 return false;
165
166 // Don't merge writable sections. .data.rel.ro sections are marked as writable
167 // but are semantically read-only.
168 if ((s->flags & SHF_WRITE) && s->name != ".data.rel.ro" &&
169 !s->name.startswith(".data.rel.ro."))
170 return false;
171
172 // SHF_LINK_ORDER sections are ICF'd as a unit with their dependent sections,
173 // so we don't consider them for ICF individually.
174 if (s->flags & SHF_LINK_ORDER)
175 return false;
176
177 // Don't merge synthetic sections as their Data member is not valid and empty.
178 // The Data member needs to be valid for ICF as it is used by ICF to determine
179 // the equality of section contents.
180 if (isa<SyntheticSection>(s))
181 return false;
182
183 // .init and .fini contains instructions that must be executed to initialize
184 // and finalize the process. They cannot and should not be merged.
185 if (s->name == ".init" || s->name == ".fini")
186 return false;
187
188 // A user program may enumerate sections named with a C identifier using
189 // __start_* and __stop_* symbols. We cannot ICF any such sections because
190 // that could change program semantics.
191 if (isValidCIdentifier(s->name))
192 return false;
193
194 return true;
195 }
196
197 // Split an equivalence class into smaller classes.
198 template <class ELFT>
segregate(size_t begin,size_t end,bool constant)199 void ICF<ELFT>::segregate(size_t begin, size_t end, bool constant) {
200 // This loop rearranges sections in [Begin, End) so that all sections
201 // that are equal in terms of equals{Constant,Variable} are contiguous
202 // in [Begin, End).
203 //
204 // The algorithm is quadratic in the worst case, but that is not an
205 // issue in practice because the number of the distinct sections in
206 // each range is usually very small.
207
208 while (begin < end) {
209 // Divide [Begin, End) into two. Let Mid be the start index of the
210 // second group.
211 auto bound =
212 std::stable_partition(sections.begin() + begin + 1,
213 sections.begin() + end, [&](InputSection *s) {
214 if (constant)
215 return equalsConstant(sections[begin], s);
216 return equalsVariable(sections[begin], s);
217 });
218 size_t mid = bound - sections.begin();
219
220 // Now we split [Begin, End) into [Begin, Mid) and [Mid, End) by
221 // updating the sections in [Begin, Mid). We use Mid as an equivalence
222 // class ID because every group ends with a unique index.
223 for (size_t i = begin; i < mid; ++i)
224 sections[i]->eqClass[next] = mid;
225
226 // If we created a group, we need to iterate the main loop again.
227 if (mid != end)
228 repeat = true;
229
230 begin = mid;
231 }
232 }
233
234 // Compare two lists of relocations.
235 template <class ELFT>
236 template <class RelTy>
constantEq(const InputSection * secA,ArrayRef<RelTy> ra,const InputSection * secB,ArrayRef<RelTy> rb)237 bool ICF<ELFT>::constantEq(const InputSection *secA, ArrayRef<RelTy> ra,
238 const InputSection *secB, ArrayRef<RelTy> rb) {
239 for (size_t i = 0; i < ra.size(); ++i) {
240 if (ra[i].r_offset != rb[i].r_offset ||
241 ra[i].getType(config->isMips64EL) != rb[i].getType(config->isMips64EL))
242 return false;
243
244 uint64_t addA = getAddend<ELFT>(ra[i]);
245 uint64_t addB = getAddend<ELFT>(rb[i]);
246
247 Symbol &sa = secA->template getFile<ELFT>()->getRelocTargetSym(ra[i]);
248 Symbol &sb = secB->template getFile<ELFT>()->getRelocTargetSym(rb[i]);
249 if (&sa == &sb) {
250 if (addA == addB)
251 continue;
252 return false;
253 }
254
255 auto *da = dyn_cast<Defined>(&sa);
256 auto *db = dyn_cast<Defined>(&sb);
257
258 // Placeholder symbols generated by linker scripts look the same now but
259 // may have different values later.
260 if (!da || !db || da->scriptDefined || db->scriptDefined)
261 return false;
262
263 // When comparing a pair of relocations, if they refer to different symbols,
264 // and either symbol is preemptible, the containing sections should be
265 // considered different. This is because even if the sections are identical
266 // in this DSO, they may not be after preemption.
267 if (da->isPreemptible || db->isPreemptible)
268 return false;
269
270 // Relocations referring to absolute symbols are constant-equal if their
271 // values are equal.
272 if (!da->section && !db->section && da->value + addA == db->value + addB)
273 continue;
274 if (!da->section || !db->section)
275 return false;
276
277 if (da->section->kind() != db->section->kind())
278 return false;
279
280 // Relocations referring to InputSections are constant-equal if their
281 // section offsets are equal.
282 if (isa<InputSection>(da->section)) {
283 if (da->value + addA == db->value + addB)
284 continue;
285 return false;
286 }
287
288 // Relocations referring to MergeInputSections are constant-equal if their
289 // offsets in the output section are equal.
290 auto *x = dyn_cast<MergeInputSection>(da->section);
291 if (!x)
292 return false;
293 auto *y = cast<MergeInputSection>(db->section);
294 if (x->getParent() != y->getParent())
295 return false;
296
297 uint64_t offsetA =
298 sa.isSection() ? x->getOffset(addA) : x->getOffset(da->value) + addA;
299 uint64_t offsetB =
300 sb.isSection() ? y->getOffset(addB) : y->getOffset(db->value) + addB;
301 if (offsetA != offsetB)
302 return false;
303 }
304
305 return true;
306 }
307
308 // Compare "non-moving" part of two InputSections, namely everything
309 // except relocation targets.
310 template <class ELFT>
equalsConstant(const InputSection * a,const InputSection * b)311 bool ICF<ELFT>::equalsConstant(const InputSection *a, const InputSection *b) {
312 if (a->numRelocations != b->numRelocations || a->flags != b->flags ||
313 a->getSize() != b->getSize() || a->data() != b->data())
314 return false;
315
316 // If two sections have different output sections, we cannot merge them.
317 assert(a->getParent() && b->getParent());
318 if (a->getParent() != b->getParent())
319 return false;
320
321 if (a->areRelocsRela)
322 return constantEq(a, a->template relas<ELFT>(), b,
323 b->template relas<ELFT>());
324 return constantEq(a, a->template rels<ELFT>(), b, b->template rels<ELFT>());
325 }
326
327 // Compare two lists of relocations. Returns true if all pairs of
328 // relocations point to the same section in terms of ICF.
329 template <class ELFT>
330 template <class RelTy>
variableEq(const InputSection * secA,ArrayRef<RelTy> ra,const InputSection * secB,ArrayRef<RelTy> rb)331 bool ICF<ELFT>::variableEq(const InputSection *secA, ArrayRef<RelTy> ra,
332 const InputSection *secB, ArrayRef<RelTy> rb) {
333 assert(ra.size() == rb.size());
334
335 for (size_t i = 0; i < ra.size(); ++i) {
336 // The two sections must be identical.
337 Symbol &sa = secA->template getFile<ELFT>()->getRelocTargetSym(ra[i]);
338 Symbol &sb = secB->template getFile<ELFT>()->getRelocTargetSym(rb[i]);
339 if (&sa == &sb)
340 continue;
341
342 auto *da = cast<Defined>(&sa);
343 auto *db = cast<Defined>(&sb);
344
345 // We already dealt with absolute and non-InputSection symbols in
346 // constantEq, and for InputSections we have already checked everything
347 // except the equivalence class.
348 if (!da->section)
349 continue;
350 auto *x = dyn_cast<InputSection>(da->section);
351 if (!x)
352 continue;
353 auto *y = cast<InputSection>(db->section);
354
355 // Ineligible sections are in the special equivalence class 0.
356 // They can never be the same in terms of the equivalence class.
357 if (x->eqClass[current] == 0)
358 return false;
359 if (x->eqClass[current] != y->eqClass[current])
360 return false;
361 };
362
363 return true;
364 }
365
366 // Compare "moving" part of two InputSections, namely relocation targets.
367 template <class ELFT>
equalsVariable(const InputSection * a,const InputSection * b)368 bool ICF<ELFT>::equalsVariable(const InputSection *a, const InputSection *b) {
369 if (a->areRelocsRela)
370 return variableEq(a, a->template relas<ELFT>(), b,
371 b->template relas<ELFT>());
372 return variableEq(a, a->template rels<ELFT>(), b, b->template rels<ELFT>());
373 }
374
findBoundary(size_t begin,size_t end)375 template <class ELFT> size_t ICF<ELFT>::findBoundary(size_t begin, size_t end) {
376 uint32_t eqClass = sections[begin]->eqClass[current];
377 for (size_t i = begin + 1; i < end; ++i)
378 if (eqClass != sections[i]->eqClass[current])
379 return i;
380 return end;
381 }
382
383 // Sections in the same equivalence class are contiguous in Sections
384 // vector. Therefore, Sections vector can be considered as contiguous
385 // groups of sections, grouped by the class.
386 //
387 // This function calls Fn on every group within [Begin, End).
388 template <class ELFT>
forEachClassRange(size_t begin,size_t end,llvm::function_ref<void (size_t,size_t)> fn)389 void ICF<ELFT>::forEachClassRange(size_t begin, size_t end,
390 llvm::function_ref<void(size_t, size_t)> fn) {
391 while (begin < end) {
392 size_t mid = findBoundary(begin, end);
393 fn(begin, mid);
394 begin = mid;
395 }
396 }
397
398 // Call Fn on each equivalence class.
399 template <class ELFT>
forEachClass(llvm::function_ref<void (size_t,size_t)> fn)400 void ICF<ELFT>::forEachClass(llvm::function_ref<void(size_t, size_t)> fn) {
401 // If threading is disabled or the number of sections are
402 // too small to use threading, call Fn sequentially.
403 if (parallel::strategy.ThreadsRequested == 1 || sections.size() < 1024) {
404 forEachClassRange(0, sections.size(), fn);
405 ++cnt;
406 return;
407 }
408
409 current = cnt % 2;
410 next = (cnt + 1) % 2;
411
412 // Shard into non-overlapping intervals, and call Fn in parallel.
413 // The sharding must be completed before any calls to Fn are made
414 // so that Fn can modify the Chunks in its shard without causing data
415 // races.
416 const size_t numShards = 256;
417 size_t step = sections.size() / numShards;
418 size_t boundaries[numShards + 1];
419 boundaries[0] = 0;
420 boundaries[numShards] = sections.size();
421
422 parallelForEachN(1, numShards, [&](size_t i) {
423 boundaries[i] = findBoundary((i - 1) * step, sections.size());
424 });
425
426 parallelForEachN(1, numShards + 1, [&](size_t i) {
427 if (boundaries[i - 1] < boundaries[i])
428 forEachClassRange(boundaries[i - 1], boundaries[i], fn);
429 });
430 ++cnt;
431 }
432
433 // Combine the hashes of the sections referenced by the given section into its
434 // hash.
435 template <class ELFT, class RelTy>
combineRelocHashes(unsigned cnt,InputSection * isec,ArrayRef<RelTy> rels)436 static void combineRelocHashes(unsigned cnt, InputSection *isec,
437 ArrayRef<RelTy> rels) {
438 uint32_t hash = isec->eqClass[cnt % 2];
439 for (RelTy rel : rels) {
440 Symbol &s = isec->template getFile<ELFT>()->getRelocTargetSym(rel);
441 if (auto *d = dyn_cast<Defined>(&s))
442 if (auto *relSec = dyn_cast_or_null<InputSection>(d->section))
443 hash += relSec->eqClass[cnt % 2];
444 }
445 // Set MSB to 1 to avoid collisions with non-hash IDs.
446 isec->eqClass[(cnt + 1) % 2] = hash | (1U << 31);
447 }
448
print(const Twine & s)449 static void print(const Twine &s) {
450 if (config->printIcfSections)
451 message(s);
452 }
453
454 // The main function of ICF.
run()455 template <class ELFT> void ICF<ELFT>::run() {
456 // Compute isPreemptible early. We may add more symbols later, so this loop
457 // cannot be merged with the later computeIsPreemptible() pass which is used
458 // by scanRelocations().
459 for (Symbol *sym : symtab->symbols())
460 sym->isPreemptible = computeIsPreemptible(*sym);
461
462 // Collect sections to merge.
463 for (InputSectionBase *sec : inputSections) {
464 auto *s = cast<InputSection>(sec);
465 if (isEligible(s))
466 sections.push_back(s);
467 }
468
469 // Initially, we use hash values to partition sections.
470 parallelForEach(
471 sections, [&](InputSection *s) { s->eqClass[0] = xxHash64(s->data()); });
472
473 for (unsigned cnt = 0; cnt != 2; ++cnt) {
474 parallelForEach(sections, [&](InputSection *s) {
475 if (s->areRelocsRela)
476 combineRelocHashes<ELFT>(cnt, s, s->template relas<ELFT>());
477 else
478 combineRelocHashes<ELFT>(cnt, s, s->template rels<ELFT>());
479 });
480 }
481
482 // From now on, sections in Sections vector are ordered so that sections
483 // in the same equivalence class are consecutive in the vector.
484 llvm::stable_sort(sections, [](const InputSection *a, const InputSection *b) {
485 return a->eqClass[0] < b->eqClass[0];
486 });
487
488 // Compare static contents and assign unique IDs for each static content.
489 forEachClass([&](size_t begin, size_t end) { segregate(begin, end, true); });
490
491 // Split groups by comparing relocations until convergence is obtained.
492 do {
493 repeat = false;
494 forEachClass(
495 [&](size_t begin, size_t end) { segregate(begin, end, false); });
496 } while (repeat);
497
498 log("ICF needed " + Twine(cnt) + " iterations");
499
500 // Merge sections by the equivalence class.
501 forEachClassRange(0, sections.size(), [&](size_t begin, size_t end) {
502 if (end - begin == 1)
503 return;
504 print("selected section " + toString(sections[begin]));
505 for (size_t i = begin + 1; i < end; ++i) {
506 print(" removing identical section " + toString(sections[i]));
507 sections[begin]->replace(sections[i]);
508
509 // At this point we know sections merged are fully identical and hence
510 // we want to remove duplicate implicit dependencies such as link order
511 // and relocation sections.
512 for (InputSection *isec : sections[i]->dependentSections)
513 isec->markDead();
514 }
515 });
516
517 // InputSectionDescription::sections is populated by processSectionCommands().
518 // ICF may fold some input sections assigned to output sections. Remove them.
519 for (BaseCommand *base : script->sectionCommands)
520 if (auto *sec = dyn_cast<OutputSection>(base))
521 for (BaseCommand *sub_base : sec->sectionCommands)
522 if (auto *isd = dyn_cast<InputSectionDescription>(sub_base))
523 llvm::erase_if(isd->sections,
524 [](InputSection *isec) { return !isec->isLive(); });
525 }
526
527 // ICF entry point function.
doIcf()528 template <class ELFT> void elf::doIcf() {
529 llvm::TimeTraceScope timeScope("ICF");
530 ICF<ELFT>().run();
531 }
532
533 template void elf::doIcf<ELF32LE>();
534 template void elf::doIcf<ELF32BE>();
535 template void elf::doIcf<ELF64LE>();
536 template void elf::doIcf<ELF64BE>();
537