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 "InputFiles.h"
78#include "LinkerScript.h"
79#include "OutputSections.h"
80#include "SymbolTable.h"
81#include "Symbols.h"
82#include "SyntheticSections.h"
83#include "llvm/BinaryFormat/ELF.h"
84#include "llvm/Object/ELF.h"
85#include "llvm/Support/Parallel.h"
86#include "llvm/Support/TimeProfiler.h"
87#include "llvm/Support/xxhash.h"
88#include <algorithm>
89#include <atomic>
90
91using namespace llvm;
92using namespace llvm::ELF;
93using namespace llvm::object;
94using namespace lld;
95using namespace lld::elf;
96
97namespace {
98template <class ELFT> class ICF {
99public:
100 void run();
101
102private:
103 void segregate(size_t begin, size_t end, uint32_t eqClassBase, bool constant);
104
105 template <class RelTy>
106 bool constantEq(const InputSection *a, Relocs<RelTy> relsA,
107 const InputSection *b, Relocs<RelTy> relsB);
108
109 template <class RelTy>
110 bool variableEq(const InputSection *a, Relocs<RelTy> relsA,
111 const InputSection *b, Relocs<RelTy> relsB);
112
113 bool equalsConstant(const InputSection *a, const InputSection *b);
114 bool equalsVariable(const InputSection *a, const InputSection *b);
115
116 size_t findBoundary(size_t begin, size_t end);
117
118 void forEachClassRange(size_t begin, size_t end,
119 llvm::function_ref<void(size_t, size_t)> fn);
120
121 void forEachClass(llvm::function_ref<void(size_t, size_t)> fn);
122
123 SmallVector<InputSection *, 0> sections;
124
125 // We repeat the main loop while `Repeat` is true.
126 std::atomic<bool> repeat;
127
128 // The main loop counter.
129 int cnt = 0;
130
131 // We have two locations for equivalence classes. On the first iteration
132 // of the main loop, Class[0] has a valid value, and Class[1] contains
133 // garbage. We read equivalence classes from slot 0 and write to slot 1.
134 // So, Class[0] represents the current class, and Class[1] represents
135 // the next class. On each iteration, we switch their roles and use them
136 // alternately.
137 //
138 // Why are we doing this? Recall that other threads may be working on
139 // other equivalence classes in parallel. They may read sections that we
140 // are updating. We cannot update equivalence classes in place because
141 // it breaks the invariance that all possibly-identical sections must be
142 // in the same equivalence class at any moment. In other words, the for
143 // loop to update equivalence classes is not atomic, and that is
144 // observable from other threads. By writing new classes to other
145 // places, we can keep the invariance.
146 //
147 // Below, `Current` has the index of the current class, and `Next` has
148 // the index of the next class. If threading is enabled, they are either
149 // (0, 1) or (1, 0).
150 //
151 // Note on single-thread: if that's the case, they are always (0, 0)
152 // because we can safely read the next class without worrying about race
153 // conditions. Using the same location makes this algorithm converge
154 // faster because it uses results of the same iteration earlier.
155 int current = 0;
156 int next = 0;
157};
158}
159
160// Returns true if section S is subject of ICF.
161static bool isEligible(InputSection *s) {
162 if (!s->isLive() || s->keepUnique || !(s->flags & SHF_ALLOC))
163 return false;
164
165 // Don't merge writable sections. .data.rel.ro sections are marked as writable
166 // but are semantically read-only.
167 if ((s->flags & SHF_WRITE) && s->name != ".data.rel.ro" &&
168 !s->name.starts_with(Prefix: ".data.rel.ro."))
169 return false;
170
171 // SHF_LINK_ORDER sections are ICF'd as a unit with their dependent sections,
172 // so we don't consider them for ICF individually.
173 if (s->flags & SHF_LINK_ORDER)
174 return false;
175
176 // Don't merge synthetic sections as their Data member is not valid and empty.
177 // The Data member needs to be valid for ICF as it is used by ICF to determine
178 // the equality of section contents.
179 if (isa<SyntheticSection>(Val: s))
180 return false;
181
182 // .init and .fini contains instructions that must be executed to initialize
183 // and finalize the process. They cannot and should not be merged.
184 if (s->name == ".init" || s->name == ".fini")
185 return false;
186
187 // A user program may enumerate sections named with a C identifier using
188 // __start_* and __stop_* symbols. We cannot ICF any such sections because
189 // that could change program semantics.
190 if (isValidCIdentifier(s: s->name))
191 return false;
192
193 return true;
194}
195
196// Split an equivalence class into smaller classes.
197template <class ELFT>
198void ICF<ELFT>::segregate(size_t begin, size_t end, uint32_t eqClassBase,
199 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(a: sections[begin], b: s);
216 return equalsVariable(a: sections[begin], b: 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 the basis for
222 // the equivalence class ID because every group ends with a unique index.
223 // Add this to eqClassBase to avoid equality with unique IDs.
224 for (size_t i = begin; i < mid; ++i)
225 sections[i]->eqClass[next] = eqClassBase + mid;
226
227 // If we created a group, we need to iterate the main loop again.
228 if (mid != end)
229 repeat = true;
230
231 begin = mid;
232 }
233}
234
235// Compare two lists of relocations.
236template <class ELFT>
237template <class RelTy>
238bool ICF<ELFT>::constantEq(const InputSection *secA, Relocs<RelTy> ra,
239 const InputSection *secB, Relocs<RelTy> rb) {
240 if (ra.size() != rb.size())
241 return false;
242 auto rai = ra.begin(), rae = ra.end(), rbi = rb.begin();
243 for (; rai != rae; ++rai, ++rbi) {
244 if (rai->r_offset != rbi->r_offset ||
245 rai->getType(config->isMips64EL) != rbi->getType(config->isMips64EL))
246 return false;
247
248 uint64_t addA = getAddend<ELFT>(*rai);
249 uint64_t addB = getAddend<ELFT>(*rbi);
250
251 Symbol &sa = secA->file->getRelocTargetSym(*rai);
252 Symbol &sb = secB->file->getRelocTargetSym(*rbi);
253 if (&sa == &sb) {
254 if (addA == addB)
255 continue;
256 return false;
257 }
258
259 auto *da = dyn_cast<Defined>(Val: &sa);
260 auto *db = dyn_cast<Defined>(Val: &sb);
261
262 // Placeholder symbols generated by linker scripts look the same now but
263 // may have different values later.
264 if (!da || !db || da->scriptDefined || db->scriptDefined)
265 return false;
266
267 // When comparing a pair of relocations, if they refer to different symbols,
268 // and either symbol is preemptible, the containing sections should be
269 // considered different. This is because even if the sections are identical
270 // in this DSO, they may not be after preemption.
271 if (da->isPreemptible || db->isPreemptible)
272 return false;
273
274 // Relocations referring to absolute symbols are constant-equal if their
275 // values are equal.
276 if (!da->section && !db->section && da->value + addA == db->value + addB)
277 continue;
278 if (!da->section || !db->section)
279 return false;
280
281 if (da->section->kind() != db->section->kind())
282 return false;
283
284 // Relocations referring to InputSections are constant-equal if their
285 // section offsets are equal.
286 if (isa<InputSection>(Val: da->section)) {
287 if (da->value + addA == db->value + addB)
288 continue;
289 return false;
290 }
291
292 // Relocations referring to MergeInputSections are constant-equal if their
293 // offsets in the output section are equal.
294 auto *x = dyn_cast<MergeInputSection>(Val: da->section);
295 if (!x)
296 return false;
297 auto *y = cast<MergeInputSection>(Val: db->section);
298 if (x->getParent() != y->getParent())
299 return false;
300
301 uint64_t offsetA =
302 sa.isSection() ? x->getOffset(offset: addA) : x->getOffset(offset: da->value) + addA;
303 uint64_t offsetB =
304 sb.isSection() ? y->getOffset(offset: addB) : y->getOffset(offset: db->value) + addB;
305 if (offsetA != offsetB)
306 return false;
307 }
308
309 return true;
310}
311
312// Compare "non-moving" part of two InputSections, namely everything
313// except relocation targets.
314template <class ELFT>
315bool ICF<ELFT>::equalsConstant(const InputSection *a, const InputSection *b) {
316 if (a->flags != b->flags || a->getSize() != b->getSize() ||
317 a->content() != b->content())
318 return false;
319
320 // If two sections have different output sections, we cannot merge them.
321 assert(a->getParent() && b->getParent());
322 if (a->getParent() != b->getParent())
323 return false;
324
325 const RelsOrRelas<ELFT> ra = a->template relsOrRelas<ELFT>();
326 const RelsOrRelas<ELFT> rb = b->template relsOrRelas<ELFT>();
327 if (ra.areRelocsCrel())
328 return constantEq(a, ra.crels, b, rb.crels);
329 return ra.areRelocsRel() || rb.areRelocsRel()
330 ? constantEq(a, ra.rels, b, rb.rels)
331 : constantEq(a, ra.relas, b, rb.relas);
332}
333
334// Compare two lists of relocations. Returns true if all pairs of
335// relocations point to the same section in terms of ICF.
336template <class ELFT>
337template <class RelTy>
338bool ICF<ELFT>::variableEq(const InputSection *secA, Relocs<RelTy> ra,
339 const InputSection *secB, Relocs<RelTy> rb) {
340 assert(ra.size() == rb.size());
341
342 auto rai = ra.begin(), rae = ra.end(), rbi = rb.begin();
343 for (; rai != rae; ++rai, ++rbi) {
344 // The two sections must be identical.
345 Symbol &sa = secA->file->getRelocTargetSym(*rai);
346 Symbol &sb = secB->file->getRelocTargetSym(*rbi);
347 if (&sa == &sb)
348 continue;
349
350 auto *da = cast<Defined>(Val: &sa);
351 auto *db = cast<Defined>(Val: &sb);
352
353 // We already dealt with absolute and non-InputSection symbols in
354 // constantEq, and for InputSections we have already checked everything
355 // except the equivalence class.
356 if (!da->section)
357 continue;
358 auto *x = dyn_cast<InputSection>(Val: da->section);
359 if (!x)
360 continue;
361 auto *y = cast<InputSection>(Val: db->section);
362
363 // Sections that are in the special equivalence class 0, can never be the
364 // same in terms of the equivalence class.
365 if (x->eqClass[current] == 0)
366 return false;
367 if (x->eqClass[current] != y->eqClass[current])
368 return false;
369 };
370
371 return true;
372}
373
374// Compare "moving" part of two InputSections, namely relocation targets.
375template <class ELFT>
376bool ICF<ELFT>::equalsVariable(const InputSection *a, const InputSection *b) {
377 const RelsOrRelas<ELFT> ra = a->template relsOrRelas<ELFT>();
378 const RelsOrRelas<ELFT> rb = b->template relsOrRelas<ELFT>();
379 if (ra.areRelocsCrel())
380 return variableEq(a, ra.crels, b, rb.crels);
381 return ra.areRelocsRel() || rb.areRelocsRel()
382 ? variableEq(a, ra.rels, b, rb.rels)
383 : variableEq(a, ra.relas, b, rb.relas);
384}
385
386template <class ELFT> size_t ICF<ELFT>::findBoundary(size_t begin, size_t end) {
387 uint32_t eqClass = sections[begin]->eqClass[current];
388 for (size_t i = begin + 1; i < end; ++i)
389 if (eqClass != sections[i]->eqClass[current])
390 return i;
391 return end;
392}
393
394// Sections in the same equivalence class are contiguous in Sections
395// vector. Therefore, Sections vector can be considered as contiguous
396// groups of sections, grouped by the class.
397//
398// This function calls Fn on every group within [Begin, End).
399template <class ELFT>
400void ICF<ELFT>::forEachClassRange(size_t begin, size_t end,
401 llvm::function_ref<void(size_t, size_t)> fn) {
402 while (begin < end) {
403 size_t mid = findBoundary(begin, end);
404 fn(begin, mid);
405 begin = mid;
406 }
407}
408
409// Call Fn on each equivalence class.
410template <class ELFT>
411void ICF<ELFT>::forEachClass(llvm::function_ref<void(size_t, size_t)> fn) {
412 // If threading is disabled or the number of sections are
413 // too small to use threading, call Fn sequentially.
414 if (parallel::strategy.ThreadsRequested == 1 || sections.size() < 1024) {
415 forEachClassRange(begin: 0, end: sections.size(), fn);
416 ++cnt;
417 return;
418 }
419
420 current = cnt % 2;
421 next = (cnt + 1) % 2;
422
423 // Shard into non-overlapping intervals, and call Fn in parallel.
424 // The sharding must be completed before any calls to Fn are made
425 // so that Fn can modify the Chunks in its shard without causing data
426 // races.
427 const size_t numShards = 256;
428 size_t step = sections.size() / numShards;
429 size_t boundaries[numShards + 1];
430 boundaries[0] = 0;
431 boundaries[numShards] = sections.size();
432
433 parallelFor(1, numShards, [&](size_t i) {
434 boundaries[i] = findBoundary(begin: (i - 1) * step, end: sections.size());
435 });
436
437 parallelFor(1, numShards + 1, [&](size_t i) {
438 if (boundaries[i - 1] < boundaries[i])
439 forEachClassRange(begin: boundaries[i - 1], end: boundaries[i], fn);
440 });
441 ++cnt;
442}
443
444// Combine the hashes of the sections referenced by the given section into its
445// hash.
446template <class RelTy>
447static void combineRelocHashes(unsigned cnt, InputSection *isec,
448 Relocs<RelTy> rels) {
449 uint32_t hash = isec->eqClass[cnt % 2];
450 for (RelTy rel : rels) {
451 Symbol &s = isec->file->getRelocTargetSym(rel);
452 if (auto *d = dyn_cast<Defined>(Val: &s))
453 if (auto *relSec = dyn_cast_or_null<InputSection>(Val: d->section))
454 hash += relSec->eqClass[cnt % 2];
455 }
456 // Set MSB to 1 to avoid collisions with unique IDs.
457 isec->eqClass[(cnt + 1) % 2] = hash | (1U << 31);
458}
459
460static void print(const Twine &s) {
461 if (config->printIcfSections)
462 message(msg: s);
463}
464
465// The main function of ICF.
466template <class ELFT> void ICF<ELFT>::run() {
467 // Compute isPreemptible early. We may add more symbols later, so this loop
468 // cannot be merged with the later computeIsPreemptible() pass which is used
469 // by scanRelocations().
470 if (config->hasDynSymTab)
471 for (Symbol *sym : symtab.getSymbols())
472 sym->isPreemptible = computeIsPreemptible(sym: *sym);
473
474 // Two text sections may have identical content and relocations but different
475 // LSDA, e.g. the two functions may have catch blocks of different types. If a
476 // text section is referenced by a .eh_frame FDE with LSDA, it is not
477 // eligible. This is implemented by iterating over CIE/FDE and setting
478 // eqClass[0] to the referenced text section from a live FDE.
479 //
480 // If two .gcc_except_table have identical semantics (usually identical
481 // content with PC-relative encoding), we will lose folding opportunity.
482 uint32_t uniqueId = 0;
483 for (Partition &part : partitions)
484 part.ehFrame->iterateFDEWithLSDA<ELFT>(
485 [&](InputSection &s) { s.eqClass[0] = s.eqClass[1] = ++uniqueId; });
486
487 // Collect sections to merge.
488 for (InputSectionBase *sec : ctx.inputSections) {
489 auto *s = dyn_cast<InputSection>(Val: sec);
490 if (s && s->eqClass[0] == 0) {
491 if (isEligible(s))
492 sections.push_back(Elt: s);
493 else
494 // Ineligible sections are assigned unique IDs, i.e. each section
495 // belongs to an equivalence class of its own.
496 s->eqClass[0] = s->eqClass[1] = ++uniqueId;
497 }
498 }
499
500 // Initially, we use hash values to partition sections.
501 parallelForEach(sections, [&](InputSection *s) {
502 // Set MSB to 1 to avoid collisions with unique IDs.
503 s->eqClass[0] = xxh3_64bits(data: s->content()) | (1U << 31);
504 });
505
506 // Perform 2 rounds of relocation hash propagation. 2 is an empirical value to
507 // reduce the average sizes of equivalence classes, i.e. segregate() which has
508 // a large time complexity will have less work to do.
509 for (unsigned cnt = 0; cnt != 2; ++cnt) {
510 parallelForEach(sections, [&](InputSection *s) {
511 const RelsOrRelas<ELFT> rels = s->template relsOrRelas<ELFT>();
512 if (rels.areRelocsCrel())
513 combineRelocHashes(cnt, s, rels.crels);
514 else if (rels.areRelocsRel())
515 combineRelocHashes(cnt, s, rels.rels);
516 else
517 combineRelocHashes(cnt, s, rels.relas);
518 });
519 }
520
521 // From now on, sections in Sections vector are ordered so that sections
522 // in the same equivalence class are consecutive in the vector.
523 llvm::stable_sort(sections, [](const InputSection *a, const InputSection *b) {
524 return a->eqClass[0] < b->eqClass[0];
525 });
526
527 // Compare static contents and assign unique equivalence class IDs for each
528 // static content. Use a base offset for these IDs to ensure no overlap with
529 // the unique IDs already assigned.
530 uint32_t eqClassBase = ++uniqueId;
531 forEachClass(fn: [&](size_t begin, size_t end) {
532 segregate(begin, end, eqClassBase, constant: true);
533 });
534
535 // Split groups by comparing relocations until convergence is obtained.
536 do {
537 repeat = false;
538 forEachClass(fn: [&](size_t begin, size_t end) {
539 segregate(begin, end, eqClassBase, constant: false);
540 });
541 } while (repeat);
542
543 log(msg: "ICF needed " + Twine(cnt) + " iterations");
544
545 // Merge sections by the equivalence class.
546 forEachClassRange(begin: 0, end: sections.size(), fn: [&](size_t begin, size_t end) {
547 if (end - begin == 1)
548 return;
549 print(s: "selected section " + toString(sections[begin]));
550 for (size_t i = begin + 1; i < end; ++i) {
551 print(s: " removing identical section " + toString(sections[i]));
552 sections[begin]->replace(other: sections[i]);
553
554 // At this point we know sections merged are fully identical and hence
555 // we want to remove duplicate implicit dependencies such as link order
556 // and relocation sections.
557 for (InputSection *isec : sections[i]->dependentSections)
558 isec->markDead();
559 }
560 });
561
562 // Change Defined symbol's section field to the canonical one.
563 auto fold = [](Symbol *sym) {
564 if (auto *d = dyn_cast<Defined>(Val: sym))
565 if (auto *sec = dyn_cast_or_null<InputSection>(Val: d->section))
566 if (sec->repl != d->section) {
567 d->section = sec->repl;
568 d->folded = true;
569 }
570 };
571 for (Symbol *sym : symtab.getSymbols())
572 fold(sym);
573 parallelForEach(ctx.objectFiles, [&](ELFFileBase *file) {
574 for (Symbol *sym : file->getLocalSymbols())
575 fold(sym);
576 });
577
578 // InputSectionDescription::sections is populated by processSectionCommands().
579 // ICF may fold some input sections assigned to output sections. Remove them.
580 for (SectionCommand *cmd : script->sectionCommands)
581 if (auto *osd = dyn_cast<OutputDesc>(Val: cmd))
582 for (SectionCommand *subCmd : osd->osec.commands)
583 if (auto *isd = dyn_cast<InputSectionDescription>(Val: subCmd))
584 llvm::erase_if(isd->sections,
585 [](InputSection *isec) { return !isec->isLive(); });
586}
587
588// ICF entry point function.
589template <class ELFT> void elf::doIcf() {
590 llvm::TimeTraceScope timeScope("ICF");
591 ICF<ELFT>().run();
592}
593
594template void elf::doIcf<ELF32LE>();
595template void elf::doIcf<ELF32BE>();
596template void elf::doIcf<ELF64LE>();
597template void elf::doIcf<ELF64BE>();
598