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#include "ICF.h"
10#include "ConcatOutputSection.h"
11#include "Config.h"
12#include "InputSection.h"
13#include "SymbolTable.h"
14#include "Symbols.h"
15
16#include "lld/Common/CommonLinkerContext.h"
17#include "llvm/Support/Parallel.h"
18#include "llvm/Support/TimeProfiler.h"
19#include "llvm/Support/xxhash.h"
20
21#include <atomic>
22
23using namespace llvm;
24using namespace lld;
25using namespace lld::macho;
26
27static constexpr bool verboseDiagnostics = false;
28// This counter is used to generate unique thunk names.
29static uint64_t icfThunkCounter = 0;
30
31class ICF {
32public:
33 ICF(std::vector<ConcatInputSection *> &inputs);
34 void run();
35
36 using EqualsFn = bool (ICF::*)(const ConcatInputSection *,
37 const ConcatInputSection *);
38 void segregate(size_t begin, size_t end, EqualsFn);
39 size_t findBoundary(size_t begin, size_t end);
40 void forEachClassRange(size_t begin, size_t end,
41 llvm::function_ref<void(size_t, size_t)> func);
42 void forEachClass(llvm::function_ref<void(size_t, size_t)> func);
43
44 bool equalsConstant(const ConcatInputSection *ia,
45 const ConcatInputSection *ib);
46 bool equalsVariable(const ConcatInputSection *ia,
47 const ConcatInputSection *ib);
48 void applySafeThunksToRange(size_t begin, size_t end);
49
50 // ICF needs a copy of the inputs vector because its equivalence-class
51 // segregation algorithm destroys the proper sequence.
52 std::vector<ConcatInputSection *> icfInputs;
53
54 unsigned icfPass = 0;
55 std::atomic<bool> icfRepeat{false};
56 std::atomic<uint64_t> equalsConstantCount{0};
57 std::atomic<uint64_t> equalsVariableCount{0};
58};
59
60ICF::ICF(std::vector<ConcatInputSection *> &inputs) {
61 icfInputs.assign(first: inputs.begin(), last: inputs.end());
62}
63
64// ICF = Identical Code Folding
65//
66// We only fold __TEXT,__text, so this is really "code" folding, and not
67// "COMDAT" folding. String and scalar constant literals are deduplicated
68// elsewhere.
69//
70// Summary of segments & sections:
71//
72// The __TEXT segment is readonly at the MMU. Some sections are already
73// deduplicated elsewhere (__TEXT,__cstring & __TEXT,__literal*) and some are
74// synthetic and inherently free of duplicates (__TEXT,__stubs &
75// __TEXT,__unwind_info). Note that we don't yet run ICF on __TEXT,__const,
76// because doing so induces many test failures.
77//
78// The __LINKEDIT segment is readonly at the MMU, yet entirely synthetic, and
79// thus ineligible for ICF.
80//
81// The __DATA_CONST segment is read/write at the MMU, but is logically const to
82// the application after dyld applies fixups to pointer data. We currently
83// fold only the __DATA_CONST,__cfstring section.
84//
85// The __DATA segment is read/write at the MMU, and as application-writeable
86// data, none of its sections are eligible for ICF.
87//
88// Please see the large block comment in lld/ELF/ICF.cpp for an explanation
89// of the segregation algorithm.
90//
91// FIXME(gkm): implement keep-unique attributes
92// FIXME(gkm): implement address-significance tables for MachO object files
93
94// Compare "non-moving" parts of two ConcatInputSections, namely everything
95// except references to other ConcatInputSections.
96bool ICF::equalsConstant(const ConcatInputSection *ia,
97 const ConcatInputSection *ib) {
98 if (verboseDiagnostics)
99 ++equalsConstantCount;
100 // We can only fold within the same OutputSection.
101 if (ia->parent != ib->parent)
102 return false;
103 if (ia->data.size() != ib->data.size())
104 return false;
105 if (ia->data != ib->data)
106 return false;
107 if (ia->relocs.size() != ib->relocs.size())
108 return false;
109 auto f = [](const Reloc &ra, const Reloc &rb) {
110 if (ra.type != rb.type)
111 return false;
112 if (ra.pcrel != rb.pcrel)
113 return false;
114 if (ra.length != rb.length)
115 return false;
116 if (ra.offset != rb.offset)
117 return false;
118 if (isa<Symbol *>(Val: ra.referent) != isa<Symbol *>(Val: rb.referent))
119 return false;
120
121 InputSection *isecA, *isecB;
122
123 uint64_t valueA = 0;
124 uint64_t valueB = 0;
125 if (isa<Symbol *>(Val: ra.referent)) {
126 const auto *sa = cast<Symbol *>(Val: ra.referent);
127 const auto *sb = cast<Symbol *>(Val: rb.referent);
128 if (sa->kind() != sb->kind())
129 return false;
130 // ICF runs before Undefineds are treated (and potentially converted into
131 // DylibSymbols).
132 if (isa<DylibSymbol>(Val: sa) || isa<Undefined>(Val: sa))
133 return sa == sb && ra.addend == rb.addend;
134 assert(isa<Defined>(sa));
135 const auto *da = cast<Defined>(Val: sa);
136 const auto *db = cast<Defined>(Val: sb);
137 if (!da->isec() || !db->isec()) {
138 assert(da->isAbsolute() && db->isAbsolute());
139 return da->value + ra.addend == db->value + rb.addend;
140 }
141 isecA = da->isec();
142 valueA = da->value;
143 isecB = db->isec();
144 valueB = db->value;
145 } else {
146 isecA = cast<InputSection *>(Val: ra.referent);
147 isecB = cast<InputSection *>(Val: rb.referent);
148 }
149
150 // Typically, we should not encounter sections marked with `keepUnique` at
151 // this point as they would have resulted in different hashes and therefore
152 // no need for a full comparison.
153 // However, in `safe_thunks` mode, it's possible for two different
154 // relocations to reference identical `keepUnique` functions that will be
155 // distinguished later via thunks - so we need to handle this case
156 // explicitly.
157 if ((isecA != isecB) && ((isecA->keepUnique && isCodeSection(isecA)) ||
158 (isecB->keepUnique && isCodeSection(isecB))))
159 return false;
160
161 if (isecA->parent != isecB->parent)
162 return false;
163 // Sections with identical parents should be of the same kind.
164 assert(isecA->kind() == isecB->kind());
165 // We will compare ConcatInputSection contents in equalsVariable.
166 if (isa<ConcatInputSection>(Val: isecA))
167 return ra.addend == rb.addend;
168 // Else we have two literal sections. References to them are equal iff their
169 // offsets in the output section are equal.
170 if (isa<Symbol *>(Val: ra.referent))
171 // For symbol relocs, we compare the contents at the symbol address. We
172 // don't do `getOffset(value + addend)` because value + addend may not be
173 // a valid offset in the literal section.
174 return isecA->getOffset(off: valueA) == isecB->getOffset(off: valueB) &&
175 ra.addend == rb.addend;
176 else {
177 assert(valueA == 0 && valueB == 0);
178 // For section relocs, we compare the content at the section offset.
179 return isecA->getOffset(off: ra.addend) == isecB->getOffset(off: rb.addend);
180 }
181 };
182 return std::equal(first1: ia->relocs.begin(), last1: ia->relocs.end(), first2: ib->relocs.begin(),
183 binary_pred: f);
184}
185
186// Compare the "moving" parts of two ConcatInputSections -- i.e. everything not
187// handled by equalsConstant().
188bool ICF::equalsVariable(const ConcatInputSection *ia,
189 const ConcatInputSection *ib) {
190 if (verboseDiagnostics)
191 ++equalsVariableCount;
192 assert(ia->relocs.size() == ib->relocs.size());
193 auto f = [this](const Reloc &ra, const Reloc &rb) {
194 // We already filtered out mismatching values/addends in equalsConstant.
195 if (ra.referent == rb.referent)
196 return true;
197 const ConcatInputSection *isecA, *isecB;
198 if (isa<Symbol *>(Val: ra.referent)) {
199 // Matching DylibSymbols are already filtered out by the
200 // identical-referent check above. Non-matching DylibSymbols were filtered
201 // out in equalsConstant(). So we can safely cast to Defined here.
202 const auto *da = cast<Defined>(Val: cast<Symbol *>(Val: ra.referent));
203 const auto *db = cast<Defined>(Val: cast<Symbol *>(Val: rb.referent));
204 if (da->isAbsolute())
205 return true;
206 isecA = dyn_cast<ConcatInputSection>(Val: da->isec());
207 if (!isecA)
208 return true; // literal sections were checked in equalsConstant.
209 isecB = cast<ConcatInputSection>(Val: db->isec());
210 } else {
211 const auto *sa = cast<InputSection *>(Val: ra.referent);
212 const auto *sb = cast<InputSection *>(Val: rb.referent);
213 isecA = dyn_cast<ConcatInputSection>(Val: sa);
214 if (!isecA)
215 return true;
216 isecB = cast<ConcatInputSection>(Val: sb);
217 }
218 return isecA->icfEqClass[icfPass % 2] == isecB->icfEqClass[icfPass % 2];
219 };
220 if (!std::equal(first1: ia->relocs.begin(), last1: ia->relocs.end(), first2: ib->relocs.begin(), binary_pred: f))
221 return false;
222
223 // If there are symbols with associated unwind info, check that the unwind
224 // info matches. For simplicity, we only handle the case where there are only
225 // symbols at offset zero within the section (which is typically the case with
226 // .subsections_via_symbols.)
227 auto hasUnwind = [](Defined *d) { return d->unwindEntry() != nullptr; };
228 const auto *itA = llvm::find_if(Range: ia->symbols, P: hasUnwind);
229 const auto *itB = llvm::find_if(Range: ib->symbols, P: hasUnwind);
230 if (itA == ia->symbols.end())
231 return itB == ib->symbols.end();
232 if (itB == ib->symbols.end())
233 return false;
234 const Defined *da = *itA;
235 const Defined *db = *itB;
236 if (da->unwindEntry()->icfEqClass[icfPass % 2] !=
237 db->unwindEntry()->icfEqClass[icfPass % 2] ||
238 da->value != 0 || db->value != 0)
239 return false;
240 auto isZero = [](Defined *d) { return d->value == 0; };
241 return std::find_if_not(first: std::next(x: itA), last: ia->symbols.end(), pred: isZero) ==
242 ia->symbols.end() &&
243 std::find_if_not(first: std::next(x: itB), last: ib->symbols.end(), pred: isZero) ==
244 ib->symbols.end();
245}
246
247// Find the first InputSection after BEGIN whose equivalence class differs
248size_t ICF::findBoundary(size_t begin, size_t end) {
249 uint64_t beginHash = icfInputs[begin]->icfEqClass[icfPass % 2];
250 for (size_t i = begin + 1; i < end; ++i)
251 if (beginHash != icfInputs[i]->icfEqClass[icfPass % 2])
252 return i;
253 return end;
254}
255
256// Invoke FUNC on subranges with matching equivalence class
257void ICF::forEachClassRange(size_t begin, size_t end,
258 llvm::function_ref<void(size_t, size_t)> func) {
259 while (begin < end) {
260 size_t mid = findBoundary(begin, end);
261 func(begin, mid);
262 begin = mid;
263 }
264}
265
266// Find or create a symbol at offset 0 in the given section
267static Symbol *getThunkTargetSymbol(ConcatInputSection *isec) {
268 for (Symbol *sym : isec->symbols)
269 if (auto *d = dyn_cast<Defined>(Val: sym))
270 if (d->value == 0)
271 return sym;
272
273 std::string thunkName;
274 if (isec->symbols.size() == 0)
275 thunkName = isec->getName().str() + ".icf.0";
276 else
277 thunkName = isec->getName().str() + "icf.thunk.target" +
278 std::to_string(val: icfThunkCounter++);
279
280 // If no symbol found at offset 0, create one
281 auto *sym = make<Defined>(args&: thunkName, /*file=*/args: nullptr, args&: isec,
282 /*value=*/args: 0, /*size=*/args: isec->getSize(),
283 /*isWeakDef=*/args: false, /*isExternal=*/args: false,
284 /*isPrivateExtern=*/args: false, /*isThumb=*/args: false,
285 /*isReferencedDynamically=*/args: false,
286 /*noDeadStrip=*/args: false);
287 isec->symbols.push_back(NewVal: sym);
288 return sym;
289}
290
291// Given a range of identical icfInputs, replace address significant functions
292// with a thunk that is just a direct branch to the first function in the
293// series. This way we keep only one main body of the function but we still
294// retain the address uniqueness of relevant functions by having them be a
295// direct branch thunk rather than containing a full copy of the actual function
296// body.
297void ICF::applySafeThunksToRange(size_t begin, size_t end) {
298 // When creating a unique ICF thunk, use the first section as the section that
299 // all thunks will branch to.
300 ConcatInputSection *masterIsec = icfInputs[begin];
301
302 // If the first section is not address significant, sorting guarantees that
303 // there are no address significant functions. So we can skip this range.
304 if (!masterIsec->keepUnique)
305 return;
306
307 // Skip anything that is not a code section.
308 if (!isCodeSection(masterIsec))
309 return;
310
311 // If the functions we're dealing with are smaller than the thunk size, then
312 // just leave them all as-is - creating thunks would be a net loss.
313 uint32_t thunkSize = target->getICFSafeThunkSize();
314 if (masterIsec->data.size() <= thunkSize)
315 return;
316
317 // Get the symbol that all thunks will branch to.
318 Symbol *masterSym = getThunkTargetSymbol(isec: masterIsec);
319
320 for (size_t i = begin + 1; i < end; ++i) {
321 ConcatInputSection *isec = icfInputs[i];
322 // When we're done processing keepUnique entries, we can stop. Sorting
323 // guaratees that all keepUnique will be at the front.
324 if (!isec->keepUnique)
325 break;
326
327 ConcatInputSection *thunk =
328 makeSyntheticInputSection(segName: isec->getSegName(), sectName: isec->getName());
329 addInputSection(inputSection: thunk);
330
331 target->initICFSafeThunkBody(thunk, targetSym: masterSym);
332 thunk->foldIdentical(redundant: isec, foldKind: Symbol::ICFFoldKind::Thunk);
333
334 // Since we're folding the target function into a thunk, we need to adjust
335 // the symbols that now got relocated from the target function to the thunk.
336 // Since the thunk is only one branch, we move all symbols to offset 0 and
337 // make sure that the size of all non-zero-size symbols is equal to the size
338 // of the branch.
339 for (auto *sym : thunk->symbols) {
340 sym->value = 0;
341 if (sym->size != 0)
342 sym->size = thunkSize;
343 }
344 }
345}
346
347// Split icfInputs into shards, then parallelize invocation of FUNC on subranges
348// with matching equivalence class
349void ICF::forEachClass(llvm::function_ref<void(size_t, size_t)> func) {
350 // Only use threads when the benefits outweigh the overhead.
351 const size_t threadingThreshold = 1024;
352 if (icfInputs.size() < threadingThreshold) {
353 forEachClassRange(begin: 0, end: icfInputs.size(), func);
354 ++icfPass;
355 return;
356 }
357
358 // Shard into non-overlapping intervals, and call FUNC in parallel. The
359 // sharding must be completed before any calls to FUNC are made so that FUNC
360 // can modify the InputSection in its shard without causing data races.
361 const size_t shards = 256;
362 size_t step = icfInputs.size() / shards;
363 size_t boundaries[shards + 1];
364 boundaries[0] = 0;
365 boundaries[shards] = icfInputs.size();
366 parallelFor(Begin: 1, End: shards, Fn: [&](size_t i) {
367 boundaries[i] = findBoundary(begin: (i - 1) * step, end: icfInputs.size());
368 });
369 parallelFor(Begin: 1, End: shards + 1, Fn: [&](size_t i) {
370 if (boundaries[i - 1] < boundaries[i]) {
371 forEachClassRange(begin: boundaries[i - 1], end: boundaries[i], func);
372 }
373 });
374 ++icfPass;
375}
376
377void ICF::run() {
378 // Into each origin-section hash, combine all reloc referent section hashes.
379 for (icfPass = 0; icfPass < 2; ++icfPass) {
380 parallelForEach(R&: icfInputs, Fn: [&](ConcatInputSection *isec) {
381 uint32_t hash = isec->icfEqClass[icfPass % 2];
382 for (const Reloc &r : isec->relocs) {
383 if (auto *sym = r.referent.dyn_cast<Symbol *>()) {
384 if (auto *defined = dyn_cast<Defined>(Val: sym)) {
385 if (defined->isec()) {
386 if (auto *referentIsec =
387 dyn_cast<ConcatInputSection>(Val: defined->isec()))
388 hash += defined->value + referentIsec->icfEqClass[icfPass % 2];
389 else
390 hash += defined->isec()->kind() +
391 defined->isec()->getOffset(off: defined->value);
392 } else {
393 hash += defined->value;
394 }
395 } else {
396 // ICF runs before Undefined diags
397 assert(isa<Undefined>(sym) || isa<DylibSymbol>(sym));
398 }
399 }
400 }
401 // Set MSB to 1 to avoid collisions with non-hashed classes.
402 isec->icfEqClass[(icfPass + 1) % 2] = hash | (1ull << 31);
403 });
404 }
405
406 llvm::stable_sort(
407 Range&: icfInputs, C: [](const ConcatInputSection *a, const ConcatInputSection *b) {
408 // When using safe_thunks, ensure that we first sort by icfEqClass and
409 // then by keepUnique (descending). This guarantees that within an
410 // equivalence class, the keepUnique inputs are always first.
411 if (config->icfLevel == ICFLevel::safe_thunks)
412 if (a->icfEqClass[0] == b->icfEqClass[0])
413 return a->keepUnique > b->keepUnique;
414 return a->icfEqClass[0] < b->icfEqClass[0];
415 });
416 forEachClass(func: [&](size_t begin, size_t end) {
417 segregate(begin, end, &ICF::equalsConstant);
418 });
419
420 // Split equivalence groups by comparing relocations until convergence
421 do {
422 icfRepeat = false;
423 forEachClass(func: [&](size_t begin, size_t end) {
424 segregate(begin, end, &ICF::equalsVariable);
425 });
426 } while (icfRepeat);
427 log(msg: "ICF needed " + Twine(icfPass) + " iterations");
428 if (verboseDiagnostics) {
429 log(msg: "equalsConstant() called " + Twine(equalsConstantCount) + " times");
430 log(msg: "equalsVariable() called " + Twine(equalsVariableCount) + " times");
431 }
432
433 // When using safe_thunks, we need to create thunks for all keepUnique
434 // functions that can be deduplicated. Since we're creating / adding new
435 // InputSections, we can't paralellize this.
436 if (config->icfLevel == ICFLevel::safe_thunks)
437 forEachClassRange(begin: 0, end: icfInputs.size(), func: [&](size_t begin, size_t end) {
438 applySafeThunksToRange(begin, end);
439 });
440
441 // Fold sections within equivalence classes
442 forEachClass(func: [&](size_t begin, size_t end) {
443 if (end - begin < 2)
444 return;
445 bool useSafeThunks = config->icfLevel == ICFLevel::safe_thunks;
446
447 // For ICF level safe_thunks, replace keepUnique function bodies with
448 // thunks. For all other ICF levles, directly merge the functions.
449
450 ConcatInputSection *beginIsec = icfInputs[begin];
451 for (size_t i = begin + 1; i < end; ++i) {
452 // Skip keepUnique inputs when using safe_thunks (already handeled above)
453 if (useSafeThunks && icfInputs[i]->keepUnique) {
454 // Assert keepUnique sections are either small or replaced with thunks.
455 assert(!icfInputs[i]->live ||
456 icfInputs[i]->data.size() <= target->getICFSafeThunkSize());
457 assert(!icfInputs[i]->replacement ||
458 icfInputs[i]->replacement->data.size() ==
459 target->getICFSafeThunkSize());
460 continue;
461 }
462 beginIsec->foldIdentical(redundant: icfInputs[i]);
463 }
464 });
465}
466
467// Split an equivalence class into smaller classes.
468void ICF::segregate(size_t begin, size_t end, EqualsFn equals) {
469 while (begin < end) {
470 // Divide [begin, end) into two. Let mid be the start index of the
471 // second group.
472 auto bound = std::stable_partition(
473 first: icfInputs.begin() + begin + 1, last: icfInputs.begin() + end,
474 pred: [&](ConcatInputSection *isec) {
475 return (this->*equals)(icfInputs[begin], isec);
476 });
477 size_t mid = bound - icfInputs.begin();
478
479 // Split [begin, end) into [begin, mid) and [mid, end). We use mid as an
480 // equivalence class ID because every group ends with a unique index.
481 for (size_t i = begin; i < mid; ++i)
482 icfInputs[i]->icfEqClass[(icfPass + 1) % 2] = mid;
483
484 // If we created a group, we need to iterate the main loop again.
485 if (mid != end)
486 icfRepeat = true;
487
488 begin = mid;
489 }
490}
491
492void macho::markSymAsAddrSig(Symbol *s) {
493 if (auto *d = dyn_cast_or_null<Defined>(Val: s))
494 if (d->isec())
495 d->isec()->keepUnique = true;
496}
497
498void macho::markAddrSigSymbols() {
499 TimeTraceScope timeScope("Mark addrsig symbols");
500 for (InputFile *file : inputFiles) {
501 ObjFile *obj = dyn_cast<ObjFile>(Val: file);
502 if (!obj)
503 continue;
504
505 Section *addrSigSection = obj->addrSigSection;
506 if (!addrSigSection)
507 continue;
508 assert(addrSigSection->subsections.size() == 1);
509
510 const InputSection *isec = addrSigSection->subsections[0].isec;
511
512 for (const Reloc &r : isec->relocs) {
513 if (auto *sym = r.referent.dyn_cast<Symbol *>())
514 markSymAsAddrSig(s: sym);
515 else
516 error(msg: toString(isec) + ": unexpected section relocation");
517 }
518 }
519}
520
521// Given a symbol that was folded into a thunk, return the symbol pointing to
522// the actual body of the function. We use this approach rather than storing the
523// needed info in the Defined itself in order to minimize memory usage.
524Defined *macho::getBodyForThunkFoldedSym(Defined *foldedSym) {
525 assert(isa<ConcatInputSection>(foldedSym->originalIsec) &&
526 "thunk-folded ICF symbol expected to be on a ConcatInputSection");
527 // foldedSec is the InputSection that was marked as deleted upon fold
528 ConcatInputSection *foldedSec =
529 cast<ConcatInputSection>(Val: foldedSym->originalIsec);
530
531 // thunkBody is the actual live thunk, containing the code that branches to
532 // the actual body of the function.
533 InputSection *thunkBody = foldedSec->replacement;
534
535 // The symbol of the merged body of the function that the thunk jumps to. This
536 // will end up in the final binary.
537 Symbol *targetSym = target->getThunkBranchTarget(thunk: thunkBody);
538
539 return cast<Defined>(Val: targetSym);
540}
541void macho::foldIdenticalSections(bool onlyCfStrings) {
542 TimeTraceScope timeScope("Fold Identical Code Sections");
543 // The ICF equivalence-class segregation algorithm relies on pre-computed
544 // hashes of InputSection::data for the ConcatOutputSection::inputs and all
545 // sections referenced by their relocs. We could recursively traverse the
546 // relocs to find every referenced InputSection, but that precludes easy
547 // parallelization. Therefore, we hash every InputSection here where we have
548 // them all accessible as simple vectors.
549
550 // If an InputSection is ineligible for ICF, we give it a unique ID to force
551 // it into an unfoldable singleton equivalence class. Begin the unique-ID
552 // space at inputSections.size(), so that it will never intersect with
553 // equivalence-class IDs which begin at 0. Since hashes & unique IDs never
554 // coexist with equivalence-class IDs, this is not necessary, but might help
555 // someone keep the numbers straight in case we ever need to debug the
556 // ICF::segregate()
557 std::vector<ConcatInputSection *> foldable;
558 uint64_t icfUniqueID = inputSections.size();
559 // Reset the thunk counter for each run of ICF.
560 icfThunkCounter = 0;
561 for (ConcatInputSection *isec : inputSections) {
562 bool isFoldableWithAddendsRemoved = isCfStringSection(isec) ||
563 isClassRefsSection(isec) ||
564 isSelRefsSection(isec);
565 // NOTE: __objc_selrefs is typically marked as no_dead_strip by MC, but we
566 // can still fold it.
567 bool hasFoldableFlags = (isSelRefsSection(isec) ||
568 sectionType(flags: isec->getFlags()) == MachO::S_REGULAR);
569
570 bool isCodeSec = isCodeSection(isec);
571
572 // When keepUnique is true, the section is not foldable. Unless we are at
573 // icf level safe_thunks, in which case we still want to fold code sections.
574 // When using safe_thunks we'll apply the safe_thunks logic at merge time
575 // based on the 'keepUnique' flag.
576 bool noUniqueRequirement =
577 !isec->keepUnique ||
578 ((config->icfLevel == ICFLevel::safe_thunks) && isCodeSec);
579
580 // FIXME: consider non-code __text sections as foldable?
581 bool isFoldable = (!onlyCfStrings || isCfStringSection(isec)) &&
582 (isCodeSec || isFoldableWithAddendsRemoved ||
583 isGccExceptTabSection(isec)) &&
584 noUniqueRequirement && !isec->hasAltEntry &&
585 !isec->shouldOmitFromOutput() && hasFoldableFlags;
586 if (isFoldable) {
587 foldable.push_back(x: isec);
588 for (Defined *d : isec->symbols)
589 if (d->unwindEntry())
590 foldable.push_back(x: d->unwindEntry());
591
592 // Some sections have embedded addends that foil ICF's hashing / equality
593 // checks. (We can ignore embedded addends when doing ICF because the same
594 // information gets recorded in our Reloc structs.) We therefore create a
595 // mutable copy of the section data and zero out the embedded addends
596 // before performing any hashing / equality checks.
597 if (isFoldableWithAddendsRemoved) {
598 // We have to do this copying serially as the BumpPtrAllocator is not
599 // thread-safe. FIXME: Make a thread-safe allocator.
600 MutableArrayRef<uint8_t> copy = isec->data.copy(A&: bAlloc());
601 for (const Reloc &r : isec->relocs)
602 target->relocateOne(loc: copy.data() + r.offset, r, /*va=*/0,
603 /*relocVA=*/0);
604 isec->data = copy;
605 }
606 } else if (!isEhFrameSection(isec)) {
607 // EH frames are gathered as foldables from unwindEntry above; give a
608 // unique ID to everything else.
609 isec->icfEqClass[0] = ++icfUniqueID;
610 }
611 }
612 parallelForEach(R&: foldable, Fn: [](ConcatInputSection *isec) {
613 assert(isec->icfEqClass[0] == 0); // don't overwrite a unique ID!
614 // Turn-on the top bit to guarantee that valid hashes have no collisions
615 // with the small-integer unique IDs for ICF-ineligible sections
616 isec->icfEqClass[0] = xxh3_64bits(data: isec->data) | (1ull << 31);
617 });
618 // Now that every input section is either hashed or marked as unique, run the
619 // segregation algorithm to detect foldable subsections.
620 ICF(foldable).run();
621}
622