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/Support/Parallel.h"
85#include "llvm/Support/TimeProfiler.h"
86#include "llvm/Support/xxhash.h"
87#include <algorithm>
88#include <atomic>
89
90using namespace llvm;
91using namespace llvm::ELF;
92using namespace llvm::object;
93using namespace lld;
94using namespace lld::elf;
95
96namespace {
97template <class ELFT> class ICF {
98public:
99 ICF(Ctx &ctx) : ctx(ctx) {}
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 parallelForEachClass(llvm::function_ref<void(size_t, size_t)> fn);
122
123 Ctx &ctx;
124 SmallVector<InputSection *, 0> 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.
162static 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.starts_with(Prefix: ".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>(Val: 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: s->name))
192 return false;
193
194 return true;
195}
196
197// Split an equivalence class into smaller classes.
198template <class ELFT>
199void ICF<ELFT>::segregate(size_t begin, size_t end, uint32_t eqClassBase,
200 bool constant) {
201 // This loop rearranges sections in [Begin, End) so that all sections
202 // that are equal in terms of equals{Constant,Variable} are contiguous
203 // in [Begin, End).
204 //
205 // The algorithm is quadratic in the worst case, but that is not an
206 // issue in practice because the number of the distinct sections in
207 // each range is usually very small.
208
209 while (begin < end) {
210 // Divide [Begin, End) into two. Let Mid be the start index of the
211 // second group.
212 auto bound =
213 std::stable_partition(sections.begin() + begin + 1,
214 sections.begin() + end, [&](InputSection *s) {
215 if (constant)
216 return equalsConstant(a: sections[begin], b: s);
217 return equalsVariable(a: sections[begin], b: s);
218 });
219 size_t mid = bound - sections.begin();
220
221 // Now we split [Begin, End) into [Begin, Mid) and [Mid, End) by
222 // updating the sections in [Begin, Mid). We use Mid as the basis for
223 // the equivalence class ID because every group ends with a unique index.
224 // Add this to eqClassBase to avoid equality with unique IDs.
225 for (size_t i = begin; i < mid; ++i)
226 sections[i]->eqClass[next] = eqClassBase + mid;
227
228 // If we created a group, we need to iterate the main loop again.
229 if (mid != end)
230 repeat = true;
231
232 begin = mid;
233 }
234}
235
236// Compare two lists of relocations.
237template <class ELFT>
238template <class RelTy>
239bool ICF<ELFT>::constantEq(const InputSection *secA, Relocs<RelTy> ra,
240 const InputSection *secB, Relocs<RelTy> rb) {
241 if (ra.size() != rb.size())
242 return false;
243 auto rai = ra.begin(), rae = ra.end(), rbi = rb.begin();
244 for (; rai != rae; ++rai, ++rbi) {
245 if (rai->r_offset != rbi->r_offset ||
246 rai->getType(ctx.arg.isMips64EL) != rbi->getType(ctx.arg.isMips64EL))
247 return false;
248
249 uint64_t addA = getAddend<ELFT>(*rai);
250 uint64_t addB = getAddend<ELFT>(*rbi);
251
252 Symbol &sa = secA->file->getRelocTargetSym(*rai);
253 Symbol &sb = secB->file->getRelocTargetSym(*rbi);
254 if (&sa == &sb) {
255 if (addA == addB)
256 continue;
257 return false;
258 }
259
260 auto *da = dyn_cast<Defined>(Val: &sa);
261 auto *db = dyn_cast<Defined>(Val: &sb);
262
263 // Placeholder symbols generated by linker scripts look the same now but
264 // may have different values later.
265 if (!da || !db || da->scriptDefined || db->scriptDefined)
266 return false;
267
268 // When comparing a pair of relocations, if they refer to different symbols,
269 // and either symbol is preemptible, the containing sections should be
270 // considered different. This is because even if the sections are identical
271 // in this DSO, they may not be after preemption.
272 if (da->isPreemptible || db->isPreemptible)
273 return false;
274
275 // Relocations referring to absolute symbols are constant-equal if their
276 // values are equal.
277 if (!da->section && !db->section && da->value + addA == db->value + addB)
278 continue;
279 if (!da->section || !db->section)
280 return false;
281
282 if (da->section->kind() != db->section->kind())
283 return false;
284
285 // Relocations referring to InputSections are constant-equal if their
286 // section offsets are equal.
287 if (isa<InputSection>(Val: da->section)) {
288 if (da->value + addA == db->value + addB)
289 continue;
290 return false;
291 }
292
293 // Relocations referring to MergeInputSections are constant-equal if their
294 // offsets in the output section are equal.
295 auto *x = dyn_cast<MergeInputSection>(Val: da->section);
296 if (!x)
297 return false;
298 auto *y = cast<MergeInputSection>(Val: db->section);
299 if (x->getParent() != y->getParent())
300 return false;
301
302 uint64_t offsetA =
303 sa.isSection() ? x->getOffset(offset: addA) : x->getOffset(offset: da->value) + addA;
304 uint64_t offsetB =
305 sb.isSection() ? y->getOffset(offset: addB) : y->getOffset(offset: db->value) + addB;
306 if (offsetA != offsetB)
307 return false;
308 }
309
310 return true;
311}
312
313// Compare "non-moving" part of two InputSections, namely everything
314// except relocation targets.
315template <class ELFT>
316bool ICF<ELFT>::equalsConstant(const InputSection *a, const InputSection *b) {
317 if (a->flags != b->flags || a->getSize() != b->getSize() ||
318 a->content() != b->content())
319 return false;
320
321 // If two sections have different output sections, we cannot merge them.
322 assert(a->getParent() && b->getParent());
323 if (a->getParent() != b->getParent())
324 return false;
325
326 const RelsOrRelas<ELFT> ra = a->template relsOrRelas<ELFT>();
327 const RelsOrRelas<ELFT> rb = b->template relsOrRelas<ELFT>();
328 if (ra.areRelocsCrel() || rb.areRelocsCrel())
329 return constantEq(a, ra.crels, b, rb.crels);
330 return ra.areRelocsRel() || rb.areRelocsRel()
331 ? constantEq(a, ra.rels, b, rb.rels)
332 : constantEq(a, ra.relas, b, rb.relas);
333}
334
335// Compare two lists of relocations. Returns true if all pairs of
336// relocations point to the same section in terms of ICF.
337template <class ELFT>
338template <class RelTy>
339bool ICF<ELFT>::variableEq(const InputSection *secA, Relocs<RelTy> ra,
340 const InputSection *secB, Relocs<RelTy> rb) {
341 assert(ra.size() == rb.size());
342
343 auto rai = ra.begin(), rae = ra.end(), rbi = rb.begin();
344 for (; rai != rae; ++rai, ++rbi) {
345 // The two sections must be identical.
346 Symbol &sa = secA->file->getRelocTargetSym(*rai);
347 Symbol &sb = secB->file->getRelocTargetSym(*rbi);
348 if (&sa == &sb)
349 continue;
350
351 auto *da = cast<Defined>(Val: &sa);
352 auto *db = cast<Defined>(Val: &sb);
353
354 // We already dealt with absolute and non-InputSection symbols in
355 // constantEq, and for InputSections we have already checked everything
356 // except the equivalence class.
357 if (!da->section)
358 continue;
359 auto *x = dyn_cast<InputSection>(Val: da->section);
360 if (!x)
361 continue;
362 auto *y = cast<InputSection>(Val: db->section);
363
364 // Sections that are in the special equivalence class 0, can never be the
365 // same in terms of the equivalence class.
366 if (x->eqClass[current] == 0)
367 return false;
368 if (x->eqClass[current] != y->eqClass[current])
369 return false;
370 };
371
372 return true;
373}
374
375// Compare "moving" part of two InputSections, namely relocation targets.
376template <class ELFT>
377bool ICF<ELFT>::equalsVariable(const InputSection *a, const InputSection *b) {
378 const RelsOrRelas<ELFT> ra = a->template relsOrRelas<ELFT>();
379 const RelsOrRelas<ELFT> rb = b->template relsOrRelas<ELFT>();
380 if (ra.areRelocsCrel() || rb.areRelocsCrel())
381 return variableEq(a, ra.crels, b, rb.crels);
382 return ra.areRelocsRel() || rb.areRelocsRel()
383 ? variableEq(a, ra.rels, b, rb.rels)
384 : variableEq(a, ra.relas, b, rb.relas);
385}
386
387template <class ELFT> size_t ICF<ELFT>::findBoundary(size_t begin, size_t end) {
388 uint32_t eqClass = sections[begin]->eqClass[current];
389 for (size_t i = begin + 1; i < end; ++i)
390 if (eqClass != sections[i]->eqClass[current])
391 return i;
392 return end;
393}
394
395// Sections in the same equivalence class are contiguous in Sections
396// vector. Therefore, Sections vector can be considered as contiguous
397// groups of sections, grouped by the class.
398//
399// This function calls Fn on every group within [Begin, End).
400template <class ELFT>
401void ICF<ELFT>::forEachClassRange(size_t begin, size_t end,
402 llvm::function_ref<void(size_t, size_t)> fn) {
403 while (begin < end) {
404 size_t mid = findBoundary(begin, end);
405 fn(begin, mid);
406 begin = mid;
407 }
408}
409
410// Call Fn on each equivalence class.
411
412template <class ELFT>
413void ICF<ELFT>::parallelForEachClass(
414 llvm::function_ref<void(size_t, size_t)> fn) {
415 // If threading is disabled or the number of sections are
416 // too small to use threading, call Fn sequentially.
417 if (parallel::strategy.ThreadsRequested == 1 || sections.size() < 1024) {
418 forEachClassRange(begin: 0, end: sections.size(), fn);
419 ++cnt;
420 return;
421 }
422
423 current = cnt % 2;
424 next = (cnt + 1) % 2;
425
426 // Shard into non-overlapping intervals, and call Fn in parallel.
427 // The sharding must be completed before any calls to Fn are made
428 // so that Fn can modify the Chunks in its shard without causing data
429 // races.
430 const size_t numShards = 256;
431 size_t step = sections.size() / numShards;
432 size_t boundaries[numShards + 1];
433 boundaries[0] = 0;
434 boundaries[numShards] = sections.size();
435
436 parallelFor(1, numShards, [&](size_t i) {
437 boundaries[i] = findBoundary(begin: (i - 1) * step, end: sections.size());
438 });
439
440 parallelFor(1, numShards + 1, [&](size_t i) {
441 if (boundaries[i - 1] < boundaries[i])
442 forEachClassRange(begin: boundaries[i - 1], end: boundaries[i], fn);
443 });
444 ++cnt;
445}
446
447// Combine the hashes of the sections referenced by the given section into its
448// hash.
449template <class RelTy>
450static void combineRelocHashes(unsigned cnt, InputSection *isec,
451 Relocs<RelTy> rels) {
452 uint32_t hash = isec->eqClass[cnt % 2];
453 for (RelTy rel : rels) {
454 Symbol &s = isec->file->getRelocTargetSym(rel);
455 if (auto *d = dyn_cast<Defined>(Val: &s))
456 if (auto *relSec = dyn_cast_or_null<InputSection>(Val: d->section))
457 hash += relSec->eqClass[cnt % 2];
458 }
459 // Set MSB to 1 to avoid collisions with unique IDs.
460 isec->eqClass[(cnt + 1) % 2] = hash | (1U << 31);
461}
462
463// The main function of ICF.
464template <class ELFT> void ICF<ELFT>::run() {
465 // Two text sections may have identical content and relocations but different
466 // LSDA, e.g. the two functions may have catch blocks of different types. If a
467 // text section is referenced by a .eh_frame FDE with LSDA, it is not
468 // eligible. This is implemented by iterating over CIE/FDE and setting
469 // eqClass[0] to the referenced text section from a live FDE.
470 //
471 // If two .gcc_except_table have identical semantics (usually identical
472 // content with PC-relative encoding), we will lose folding opportunity.
473 uint32_t uniqueId = 0;
474 for (Partition &part : ctx.partitions)
475 part.ehFrame->iterateFDEWithLSDA<ELFT>(
476 [&](InputSection &s) { s.eqClass[0] = s.eqClass[1] = ++uniqueId; });
477
478 // Collect sections to merge.
479 for (InputSectionBase *sec : ctx.inputSections) {
480 auto *s = dyn_cast<InputSection>(Val: sec);
481 if (s && s->eqClass[0] == 0) {
482 if (isEligible(s))
483 sections.push_back(Elt: s);
484 else
485 // Ineligible sections are assigned unique IDs, i.e. each section
486 // belongs to an equivalence class of its own.
487 s->eqClass[0] = s->eqClass[1] = ++uniqueId;
488 }
489 }
490
491 // Initially, we use hash values to partition sections.
492 parallelForEach(sections, [&](InputSection *s) {
493 // Set MSB to 1 to avoid collisions with unique IDs.
494 s->eqClass[0] = xxh3_64bits(data: s->content()) | (1U << 31);
495 });
496
497 // Perform 2 rounds of relocation hash propagation. 2 is an empirical value to
498 // reduce the average sizes of equivalence classes, i.e. segregate() which has
499 // a large time complexity will have less work to do.
500 for (unsigned cnt = 0; cnt != 2; ++cnt) {
501 parallelForEach(sections, [&](InputSection *s) {
502 const RelsOrRelas<ELFT> rels = s->template relsOrRelas<ELFT>();
503 if (rels.areRelocsCrel())
504 combineRelocHashes(cnt, s, rels.crels);
505 else if (rels.areRelocsRel())
506 combineRelocHashes(cnt, s, rels.rels);
507 else
508 combineRelocHashes(cnt, s, rels.relas);
509 });
510 }
511
512 // From now on, sections in Sections vector are ordered so that sections
513 // in the same equivalence class are consecutive in the vector.
514 llvm::stable_sort(sections, [](const InputSection *a, const InputSection *b) {
515 return a->eqClass[0] < b->eqClass[0];
516 });
517
518 // Compare static contents and assign unique equivalence class IDs for each
519 // static content. Use a base offset for these IDs to ensure no overlap with
520 // the unique IDs already assigned.
521 uint32_t eqClassBase = ++uniqueId;
522 parallelForEachClass(fn: [&](size_t begin, size_t end) {
523 segregate(begin, end, eqClassBase, constant: true);
524 });
525
526 // Split groups by comparing relocations until convergence is obtained.
527 do {
528 repeat = false;
529 parallelForEachClass(fn: [&](size_t begin, size_t end) {
530 segregate(begin, end, eqClassBase, constant: false);
531 });
532 } while (repeat);
533
534 Log(ctx) << "ICF needed " << cnt << " iterations";
535
536 auto print = [&ctx = ctx]() -> ELFSyncStream {
537 return {ctx, ctx.arg.printIcfSections ? DiagLevel::Msg : DiagLevel::None};
538 };
539 // Merge sections by the equivalence class.
540 forEachClassRange(begin: 0, end: sections.size(), fn: [&](size_t begin, size_t end) {
541 if (end - begin == 1)
542 return;
543 print() << "selected section " << sections[begin];
544 for (size_t i = begin + 1; i < end; ++i) {
545 print() << " removing identical section " << sections[i];
546 sections[begin]->replace(other: sections[i]);
547
548 // At this point we know sections merged are fully identical and hence
549 // we want to remove duplicate implicit dependencies such as link order
550 // and relocation sections.
551 for (InputSection *isec : sections[i]->dependentSections)
552 isec->markDead();
553 }
554 });
555
556 // Change Defined symbol's section field to the canonical one.
557 auto fold = [](Symbol *sym) {
558 if (auto *d = dyn_cast<Defined>(Val: sym))
559 if (auto *sec = dyn_cast_or_null<InputSection>(Val: d->section))
560 if (sec->repl != d->section) {
561 d->section = sec->repl;
562 d->folded = true;
563 }
564 };
565 for (Symbol *sym : ctx.symtab->getSymbols())
566 fold(sym);
567 parallelForEach(ctx.objectFiles, [&](ELFFileBase *file) {
568 for (Symbol *sym : file->getLocalSymbols())
569 fold(sym);
570 });
571
572 // InputSectionDescription::sections is populated by processSectionCommands().
573 // ICF may fold some input sections assigned to output sections. Remove them.
574 for (SectionCommand *cmd : ctx.script->sectionCommands)
575 if (auto *osd = dyn_cast<OutputDesc>(Val: cmd))
576 for (SectionCommand *subCmd : osd->osec.commands)
577 if (auto *isd = dyn_cast<InputSectionDescription>(Val: subCmd))
578 llvm::erase_if(isd->sections,
579 [](InputSection *isec) { return !isec->isLive(); });
580}
581
582// ICF entry point function.
583template <class ELFT> void elf::doIcf(Ctx &ctx) {
584 llvm::TimeTraceScope timeScope("ICF");
585 ICF<ELFT>(ctx).run();
586}
587
588template void elf::doIcf<ELF32LE>(Ctx &);
589template void elf::doIcf<ELF32BE>(Ctx &);
590template void elf::doIcf<ELF64LE>(Ctx &);
591template void elf::doIcf<ELF64BE>(Ctx &);
592