1//===- UnwindInfoSection.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 "UnwindInfoSection.h"
10#include "InputSection.h"
11#include "Layout.h"
12#include "OutputSection.h"
13#include "OutputSegment.h"
14#include "SymbolTable.h"
15#include "Symbols.h"
16#include "SyntheticSections.h"
17#include "Target.h"
18
19#include "lld/Common/ErrorHandler.h"
20#include "lld/Common/Memory.h"
21#include "llvm/ADT/DenseMap.h"
22#include "llvm/ADT/STLExtras.h"
23#include "llvm/BinaryFormat/MachO.h"
24#include "llvm/Support/Parallel.h"
25
26#include "mach-o/compact_unwind_encoding.h"
27
28#include <numeric>
29
30using namespace llvm;
31using namespace llvm::MachO;
32using namespace llvm::support::endian;
33using namespace lld;
34using namespace lld::macho;
35
36#define COMMON_ENCODINGS_MAX 127
37#define COMPACT_ENCODINGS_MAX 256
38
39#define SECOND_LEVEL_PAGE_BYTES 4096
40#define SECOND_LEVEL_PAGE_WORDS (SECOND_LEVEL_PAGE_BYTES / sizeof(uint32_t))
41#define REGULAR_SECOND_LEVEL_ENTRIES_MAX \
42 ((SECOND_LEVEL_PAGE_BYTES - \
43 sizeof(unwind_info_regular_second_level_page_header)) / \
44 sizeof(unwind_info_regular_second_level_entry))
45#define COMPRESSED_SECOND_LEVEL_ENTRIES_MAX \
46 ((SECOND_LEVEL_PAGE_BYTES - \
47 sizeof(unwind_info_compressed_second_level_page_header)) / \
48 sizeof(uint32_t))
49
50#define COMPRESSED_ENTRY_FUNC_OFFSET_BITS 24
51#define COMPRESSED_ENTRY_FUNC_OFFSET_MASK \
52 UNWIND_INFO_COMPRESSED_ENTRY_FUNC_OFFSET(~0)
53
54static_assert(static_cast<uint32_t>(UNWIND_X86_64_DWARF_SECTION_OFFSET) ==
55 static_cast<uint32_t>(UNWIND_ARM64_DWARF_SECTION_OFFSET) &&
56 static_cast<uint32_t>(UNWIND_X86_64_DWARF_SECTION_OFFSET) ==
57 static_cast<uint32_t>(UNWIND_X86_DWARF_SECTION_OFFSET));
58
59constexpr uint64_t DWARF_SECTION_OFFSET = UNWIND_X86_64_DWARF_SECTION_OFFSET;
60
61// Compact Unwind format is a Mach-O evolution of DWARF Unwind that
62// optimizes space and exception-time lookup. Most DWARF unwind
63// entries can be replaced with Compact Unwind entries, but the ones
64// that cannot are retained in DWARF form.
65//
66// This comment will address macro-level organization of the pre-link
67// and post-link compact unwind tables. For micro-level organization
68// pertaining to the bitfield layout of the 32-bit compact unwind
69// entries, see libunwind/include/mach-o/compact_unwind_encoding.h
70//
71// Important clarifying factoids:
72//
73// * __LD,__compact_unwind is the compact unwind format for compiler
74// output and linker input. It is never a final output. It could be
75// an intermediate output with the `-r` option which retains relocs.
76//
77// * __TEXT,__unwind_info is the compact unwind format for final
78// linker output. It is never an input.
79//
80// * __TEXT,__eh_frame is the DWARF format for both linker input and output.
81//
82// * __TEXT,__unwind_info entries are divided into 4 KiB pages (2nd
83// level) by ascending address, and the pages are referenced by an
84// index (1st level) in the section header.
85//
86// * Following the headers in __TEXT,__unwind_info, the bulk of the
87// section contains a vector of compact unwind entries
88// `{functionOffset, encoding}` sorted by ascending `functionOffset`.
89// Adjacent entries with the same encoding can be folded to great
90// advantage, achieving a 3-order-of-magnitude reduction in the
91// number of entries.
92//
93// Refer to the definition of unwind_info_section_header in
94// compact_unwind_encoding.h for an overview of the format we are encoding
95// here.
96
97// TODO(gkm): how do we align the 2nd-level pages?
98
99// The various fields in the on-disk representation of each compact unwind
100// entry.
101#define FOR_EACH_CU_FIELD(DO) \
102 DO(Ptr, functionAddress) \
103 DO(uint32_t, functionLength) \
104 DO(compact_unwind_encoding_t, encoding) \
105 DO(Ptr, personality) \
106 DO(Ptr, lsda)
107
108CREATE_LAYOUT_CLASS(CompactUnwind, FOR_EACH_CU_FIELD);
109
110#undef FOR_EACH_CU_FIELD
111
112// LLD's internal representation of a compact unwind entry.
113struct CompactUnwindEntry {
114 uint64_t functionAddress;
115 uint32_t functionLength;
116 compact_unwind_encoding_t encoding;
117 Symbol *personality;
118 InputSection *lsda;
119};
120
121using EncodingMap = DenseMap<compact_unwind_encoding_t, size_t>;
122
123struct SecondLevelPage {
124 uint32_t kind;
125 size_t entryIndex;
126 size_t entryCount;
127 size_t byteCount;
128 std::vector<compact_unwind_encoding_t> localEncodings;
129 EncodingMap localEncodingIndexes;
130};
131
132// UnwindInfoSectionImpl allows us to avoid cluttering our header file with a
133// lengthy definition of UnwindInfoSection.
134class UnwindInfoSectionImpl final : public UnwindInfoSection {
135public:
136 UnwindInfoSectionImpl() : cuLayout(target->wordSize) {}
137 uint64_t getSize() const override { return unwindInfoSize; }
138 void prepare() override;
139 void finalize() override;
140 void writeTo(uint8_t *buf) const override;
141
142private:
143 void prepareRelocations(ConcatInputSection *);
144 void relocateCompactUnwind(std::vector<CompactUnwindEntry> &);
145 void encodePersonalities();
146 Symbol *canonicalizePersonality(Symbol *);
147
148 uint64_t unwindInfoSize = 0;
149 SmallVector<decltype(symbols)::value_type, 0> symbolsVec;
150 CompactUnwindLayout cuLayout;
151 std::vector<std::pair<compact_unwind_encoding_t, size_t>> commonEncodings;
152 EncodingMap commonEncodingIndexes;
153 // The entries here will be in the same order as their originating symbols
154 // in symbolsVec.
155 std::vector<CompactUnwindEntry> cuEntries;
156 std::vector<Symbol *> personalities;
157 SmallDenseMap<std::pair<InputSection *, uint64_t /* addend */>, Symbol *>
158 personalityTable;
159 // Indices into cuEntries for CUEs with a non-null LSDA.
160 std::vector<size_t> entriesWithLsda;
161 // Map of cuEntries index to an index within the LSDA array.
162 DenseMap<size_t, uint32_t> lsdaIndex;
163 std::vector<SecondLevelPage> secondLevelPages;
164 uint64_t level2PagesOffset = 0;
165 // The highest-address function plus its size. The unwinder needs this to
166 // determine the address range that is covered by unwind info.
167 uint64_t cueEndBoundary = 0;
168};
169
170UnwindInfoSection::UnwindInfoSection()
171 : SyntheticSection(segment_names::text, section_names::unwindInfo) {
172 align = 4;
173}
174
175// Record function symbols that may need entries emitted in __unwind_info, which
176// stores unwind data for address ranges.
177//
178// Note that if several adjacent functions have the same unwind encoding and
179// personality function and no LSDA, they share one unwind entry. For this to
180// work, functions without unwind info need explicit "no unwind info" unwind
181// entries -- else the unwinder would think they have the unwind info of the
182// closest function with unwind info right before in the image. Thus, we add
183// function symbols for each unique address regardless of whether they have
184// associated unwind info.
185void UnwindInfoSection::addSymbol(const Defined *d) {
186 if (d->unwindEntry())
187 allEntriesAreOmitted = false;
188 // We don't yet know the final output address of this symbol, but we know that
189 // they are uniquely determined by a combination of the isec and value, so
190 // we use that as the key here.
191 auto p = symbols.insert(KV: {{d->isec(), d->value}, d});
192 // If we have multiple symbols at the same address, only one of them can have
193 // an associated unwind entry.
194 if (!p.second && d->unwindEntry()) {
195 assert(p.first->second == d || !p.first->second->unwindEntry());
196 p.first->second = d;
197 }
198}
199
200void UnwindInfoSectionImpl::prepare() {
201 // This iteration needs to be deterministic, since prepareRelocations may add
202 // entries to the GOT. Hence the use of a MapVector for
203 // UnwindInfoSection::symbols.
204 for (const Defined *d : make_second_range(c&: symbols))
205 if (d->unwindEntry()) {
206 if (d->unwindEntry()->getName() == section_names::compactUnwind) {
207 prepareRelocations(d->unwindEntry());
208 } else {
209 // We don't have to add entries to the GOT here because FDEs have
210 // explicit GOT relocations, so Writer::scanRelocations() will add those
211 // GOT entries. However, we still need to canonicalize the personality
212 // pointers (like prepareRelocations() does for CU entries) in order
213 // to avoid overflowing the 3-personality limit.
214 FDE &fde = cast<ObjFile>(Val: d->getFile())->fdes[d->unwindEntry()];
215 fde.personality = canonicalizePersonality(fde.personality);
216 }
217 }
218}
219
220// Compact unwind relocations have different semantics, so we handle them in a
221// separate code path from regular relocations. First, we do not wish to add
222// rebase opcodes for __LD,__compact_unwind, because that section doesn't
223// actually end up in the final binary. Second, personality pointers always
224// reside in the GOT and must be treated specially.
225void UnwindInfoSectionImpl::prepareRelocations(ConcatInputSection *isec) {
226 assert(!isec->shouldOmitFromOutput() &&
227 "__compact_unwind section should not be omitted");
228
229 // FIXME: Make this skip relocations for CompactUnwindEntries that
230 // point to dead-stripped functions. That might save some amount of
231 // work. But since there are usually just few personality functions
232 // that are referenced from many places, at least some of them likely
233 // live, it wouldn't reduce number of got entries.
234 for (size_t i = 0; i < isec->relocs.size(); ++i) {
235 Relocation &r = isec->relocs[i];
236 assert(target->hasAttr(r.type, RelocAttrBits::UNSIGNED));
237 // Since compact unwind sections aren't part of the inputSections vector,
238 // they don't get canonicalized by scanRelocations(), so we have to do the
239 // canonicalization here.
240 if (auto *referentIsec = r.referent.dyn_cast<InputSection *>())
241 r.referent = referentIsec->canonical();
242
243 // Functions and LSDA entries always reside in the same object file as the
244 // compact unwind entries that references them, and thus appear as section
245 // relocs. There is no need to prepare them. We only prepare relocs for
246 // personality functions.
247 if (r.offset != cuLayout.personalityOffset)
248 continue;
249
250 if (auto *s = r.referent.dyn_cast<Symbol *>()) {
251 // Personality functions are nearly always system-defined (e.g.,
252 // ___gxx_personality_v0 for C++) and relocated as dylib symbols. When an
253 // application provides its own personality function, it might be
254 // referenced by an extern Defined symbol reloc, or a local section reloc.
255 if (auto *defined = dyn_cast<Defined>(Val: s)) {
256 // XXX(vyng) This is a special case for handling duplicate personality
257 // symbols. Note that LD64's behavior is a bit different and it is
258 // inconsistent with how symbol resolution usually work
259 //
260 // So we've decided not to follow it. Instead, simply pick the symbol
261 // with the same name from the symbol table to replace the local one.
262 //
263 // (See discussions/alternatives already considered on D107533)
264 if (!defined->isExternal())
265 if (Symbol *sym = symtab->find(name: defined->getName()))
266 if (!sym->isLazy())
267 r.referent = s = sym;
268 }
269 if (auto *undefined = dyn_cast<Undefined>(Val: s)) {
270 treatUndefinedSymbol(*undefined, isec, offset: r.offset);
271 // treatUndefinedSymbol() can replace s with a DylibSymbol; re-check.
272 if (isa<Undefined>(Val: s))
273 continue;
274 }
275
276 // Similar to canonicalizePersonality(), but we also register a GOT entry.
277 if (auto *defined = dyn_cast<Defined>(Val: s)) {
278 // Check if we have created a synthetic symbol at the same address.
279 Symbol *&personality =
280 personalityTable[{defined->isec(), defined->value}];
281 if (personality == nullptr) {
282 personality = defined;
283 in.got->addEntry(sym: defined);
284 } else if (personality != defined) {
285 r.referent = personality;
286 }
287 continue;
288 }
289
290 assert(isa<DylibSymbol>(s));
291 in.got->addEntry(sym: s);
292 continue;
293 }
294
295 if (auto *referentIsec = r.referent.dyn_cast<InputSection *>()) {
296 assert(!isCoalescedWeak(referentIsec));
297 // Personality functions can be referenced via section relocations
298 // if they live in the same object file. Create placeholder synthetic
299 // symbols for them in the GOT. If the corresponding symbol is already
300 // in the GOT, use that to avoid creating a duplicate entry. All GOT
301 // entries needed by non-unwind sections will have already been added
302 // by this point.
303 Symbol *&s = personalityTable[{referentIsec, r.addend}];
304 if (s == nullptr) {
305 Defined *const *gotEntry =
306 llvm::find_if(Range&: referentIsec->symbols, P: [&](Defined const *d) {
307 return d->value == static_cast<uint64_t>(r.addend) &&
308 d->isInGot();
309 });
310 if (gotEntry != referentIsec->symbols.end()) {
311 s = *gotEntry;
312 } else {
313 // This runs after dead stripping, so the noDeadStrip argument does
314 // not matter.
315 s = make<Defined>(args: "<internal>", /*file=*/args: nullptr, args&: referentIsec,
316 args&: r.addend, /*size=*/args: 0, /*isWeakDef=*/args: false,
317 /*isExternal=*/args: false, /*isPrivateExtern=*/args: false,
318 /*includeInSymtab=*/args: true,
319 /*isReferencedDynamically=*/args: false,
320 /*noDeadStrip=*/args: false);
321 s->used = true;
322 in.got->addEntry(sym: s);
323 }
324 }
325 r.referent = s;
326 r.addend = 0;
327 }
328 }
329}
330
331Symbol *UnwindInfoSectionImpl::canonicalizePersonality(Symbol *personality) {
332 if (auto *defined = dyn_cast_or_null<Defined>(Val: personality)) {
333 // Check if we have created a synthetic symbol at the same address.
334 Symbol *&synth = personalityTable[{defined->isec(), defined->value}];
335 if (synth == nullptr)
336 synth = defined;
337 else if (synth != defined)
338 return synth;
339 }
340 return personality;
341}
342
343// We need to apply the relocations to the pre-link compact unwind section
344// before converting it to post-link form. There should only be absolute
345// relocations here: since we are not emitting the pre-link CU section, there
346// is no source address to make a relative location meaningful.
347void UnwindInfoSectionImpl::relocateCompactUnwind(
348 std::vector<CompactUnwindEntry> &cuEntries) {
349 parallelFor(Begin: 0, End: symbolsVec.size(), Fn: [&](size_t i) {
350 CompactUnwindEntry &cu = cuEntries[i];
351 const Defined *d = symbolsVec[i].second;
352 cu.functionAddress = d->getVA();
353 if (!d->unwindEntry())
354 return;
355
356 // If we have DWARF unwind info, create a slimmed-down CU entry that points
357 // to it.
358 if (d->unwindEntry()->getName() == section_names::ehFrame) {
359 // The unwinder will look for the DWARF entry starting at the hint,
360 // assuming the hint points to a valid CFI record start. If it
361 // fails to find the record, it proceeds in a linear search through the
362 // contiguous CFI records from the hint until the end of the section.
363 // Ideally, in the case where the offset is too large to be encoded, we
364 // would instead encode the largest possible offset to a valid CFI record,
365 // but since we don't keep track of that, just encode zero -- the start of
366 // the section is always the start of a CFI record.
367 uint64_t dwarfOffsetHint =
368 d->unwindEntry()->outSecOff <= DWARF_SECTION_OFFSET
369 ? d->unwindEntry()->outSecOff
370 : 0;
371 cu.encoding = target->modeDwarfEncoding | dwarfOffsetHint;
372 const FDE &fde = cast<ObjFile>(Val: d->getFile())->fdes[d->unwindEntry()];
373 cu.functionLength = fde.funcLength;
374 // Omit the DWARF personality from compact-unwind entry so that we
375 // don't need to encode it.
376 cu.personality = nullptr;
377 cu.lsda = fde.lsda;
378 return;
379 }
380
381 assert(d->unwindEntry()->getName() == section_names::compactUnwind);
382
383 auto buf =
384 reinterpret_cast<const uint8_t *>(d->unwindEntry()->data.data()) -
385 target->wordSize;
386 cu.functionLength =
387 support::endian::read32le(P: buf + cuLayout.functionLengthOffset);
388 cu.encoding = support::endian::read32le(P: buf + cuLayout.encodingOffset);
389 for (const Relocation &r : d->unwindEntry()->relocs) {
390 if (r.offset == cuLayout.personalityOffset)
391 cu.personality = cast<Symbol *>(Val: r.referent);
392 else if (r.offset == cuLayout.lsdaOffset)
393 cu.lsda = r.getReferentInputSection();
394 }
395 });
396}
397
398// There should only be a handful of unique personality pointers, so we can
399// encode them as 2-bit indices into a small array.
400void UnwindInfoSectionImpl::encodePersonalities() {
401 for (CompactUnwindEntry &cu : cuEntries) {
402 if (cu.personality == nullptr)
403 continue;
404 // Linear search is fast enough for a small array.
405 auto it = find(Range&: personalities, Val: cu.personality);
406 uint32_t personalityIndex; // 1-based index
407 if (it != personalities.end()) {
408 personalityIndex = std::distance(first: personalities.begin(), last: it) + 1;
409 } else {
410 personalities.push_back(x: cu.personality);
411 personalityIndex = personalities.size();
412 }
413 cu.encoding |=
414 personalityIndex << llvm::countr_zero(
415 Val: static_cast<compact_unwind_encoding_t>(UNWIND_PERSONALITY_MASK));
416 }
417 if (personalities.size() > 3)
418 error(msg: "too many personalities (" + Twine(personalities.size()) +
419 ") for compact unwind to encode");
420}
421
422static bool canFoldEncoding(compact_unwind_encoding_t encoding) {
423 // From compact_unwind_encoding.h:
424 // UNWIND_X86_64_MODE_STACK_IND:
425 // A "frameless" (RBP not used as frame pointer) function large constant
426 // stack size. This case is like the previous, except the stack size is too
427 // large to encode in the compact unwind encoding. Instead it requires that
428 // the function contains "subq $nnnnnnnn,RSP" in its prolog. The compact
429 // encoding contains the offset to the nnnnnnnn value in the function in
430 // UNWIND_X86_64_FRAMELESS_STACK_SIZE.
431 // Since this means the unwinder has to look at the `subq` in the function
432 // of the unwind info's unwind address, two functions that have identical
433 // unwind info can't be folded if it's using this encoding since both
434 // entries need unique addresses.
435 static_assert(static_cast<uint32_t>(UNWIND_X86_64_MODE_STACK_IND) ==
436 static_cast<uint32_t>(UNWIND_X86_MODE_STACK_IND));
437 if ((target->cpuType == CPU_TYPE_X86_64 || target->cpuType == CPU_TYPE_X86) &&
438 (encoding & UNWIND_MODE_MASK) == UNWIND_X86_64_MODE_STACK_IND) {
439 // FIXME: Consider passing in the two function addresses and getting
440 // their two stack sizes off the `subq` and only returning false if they're
441 // actually different.
442 return false;
443 }
444 return true;
445}
446
447// Scan the __LD,__compact_unwind entries and compute the space needs of
448// __TEXT,__unwind_info and __TEXT,__eh_frame.
449void UnwindInfoSectionImpl::finalize() {
450 if (symbols.empty())
451 return;
452
453 // At this point, the address space for __TEXT,__text has been
454 // assigned, so we can relocate the __LD,__compact_unwind entries
455 // into a temporary buffer. Relocation is necessary in order to sort
456 // the CU entries by function address. Sorting is necessary so that
457 // we can fold adjacent CU entries with identical encoding+personality
458 // and without any LSDA. Folding is necessary because it reduces the
459 // number of CU entries by as much as 3 orders of magnitude!
460 cuEntries.resize(new_size: symbols.size());
461 // The "map" part of the symbols MapVector was only needed for deduplication
462 // in addSymbol(). Now that we are done adding, move the contents to a plain
463 // std::vector for indexed access.
464 symbolsVec = symbols.takeVector();
465 relocateCompactUnwind(cuEntries);
466
467 // Sort the entries by address.
468 llvm::sort(C&: cuEntries, Comp: [&](auto &a, auto &b) {
469 return a.functionAddress < b.functionAddress;
470 });
471
472 // Record the ending boundary before we fold the entries.
473 cueEndBoundary =
474 cuEntries.back().functionAddress + cuEntries.back().functionLength;
475
476 // Fold adjacent entries with matching encoding+personality and without LSDA
477 // We use three iterators to fold in-situ:
478 // (1) `foldBegin` is the first of a potential sequence of matching entries
479 // (2) `foldEnd` is the first non-matching entry after `foldBegin`.
480 // The semi-open interval [ foldBegin .. foldEnd ) contains a range
481 // entries that can be folded into a single entry and written to ...
482 // (3) `foldWrite`
483 auto foldWrite = cuEntries.begin();
484 for (auto foldBegin = cuEntries.begin(); foldBegin != cuEntries.end();) {
485 auto foldEnd = foldBegin;
486 // Common LSDA encodings (e.g. for C++ and Objective-C) contain offsets from
487 // a base address. The base address is normally not contained directly in
488 // the LSDA, and in that case, the personality function treats the starting
489 // address of the function (which is computed by the unwinder) as the base
490 // address and interprets the LSDA accordingly. The unwinder computes the
491 // starting address of a function as the address associated with its CU
492 // entry. For this reason, we cannot fold adjacent entries if they have an
493 // LSDA, because folding would make the unwinder compute the wrong starting
494 // address for the functions with the folded entries, which in turn would
495 // cause the personality function to misinterpret the LSDA for those
496 // functions. In the very rare case where the base address is encoded
497 // directly in the LSDA, two functions at different addresses would
498 // necessarily have different LSDAs, so their CU entries would not have been
499 // folded anyway.
500 while (++foldEnd != cuEntries.end() &&
501 foldBegin->encoding == foldEnd->encoding && !foldBegin->lsda &&
502 !foldEnd->lsda &&
503 // If we've gotten to this point, we don't have an LSDA, which should
504 // also imply that we don't have a personality function, since in all
505 // likelihood a personality function needs the LSDA to do anything
506 // useful. It can be technically valid to have a personality function
507 // and no LSDA though (e.g. the C++ personality __gxx_personality_v0
508 // is just a no-op without LSDA), so we still check for personality
509 // function equivalence to handle that case.
510 foldBegin->personality == foldEnd->personality &&
511 canFoldEncoding(encoding: foldEnd->encoding))
512 ;
513 *foldWrite++ = *foldBegin;
514 foldBegin = foldEnd;
515 }
516 cuEntries.erase(first: foldWrite, last: cuEntries.end());
517
518 encodePersonalities();
519
520 // Count frequencies of the folded encodings
521 EncodingMap encodingFrequencies;
522 for (const CompactUnwindEntry &cu : cuEntries)
523 encodingFrequencies[cu.encoding]++;
524
525 // Make a vector of encodings, sorted by descending frequency
526 for (const auto &frequency : encodingFrequencies)
527 commonEncodings.emplace_back(args: frequency);
528 llvm::sort(C&: commonEncodings,
529 Comp: [](const std::pair<compact_unwind_encoding_t, size_t> &a,
530 const std::pair<compact_unwind_encoding_t, size_t> &b) {
531 // When frequencies match, secondarily sort on encoding
532 // to maintain parity with validate-unwind-info.py
533 return std::tie(args: a.second, args: a.first) > std::tie(args: b.second, args: b.first);
534 });
535
536 // Truncate the vector to 127 elements.
537 // Common encoding indexes are limited to 0..126, while encoding
538 // indexes 127..255 are local to each second-level page
539 if (commonEncodings.size() > COMMON_ENCODINGS_MAX)
540 commonEncodings.resize(COMMON_ENCODINGS_MAX);
541
542 // Create a map from encoding to common-encoding-table index
543 for (size_t i = 0; i < commonEncodings.size(); i++)
544 commonEncodingIndexes[commonEncodings[i].first] = i;
545
546 // Split folded encodings into pages, where each page is limited by ...
547 // (a) 4 KiB capacity
548 // (b) 24-bit difference between first & final function address
549 // (c) 8-bit compact-encoding-table index,
550 // for which 0..126 references the global common-encodings table,
551 // and 127..255 references a local per-second-level-page table.
552 // First we try the compact format and determine how many entries fit.
553 // If more entries fit in the regular format, we use that.
554 for (size_t i = 0; i < cuEntries.size();) {
555 secondLevelPages.emplace_back();
556 SecondLevelPage &page = secondLevelPages.back();
557 page.entryIndex = i;
558 uint64_t functionAddressMax =
559 cuEntries[i].functionAddress + COMPRESSED_ENTRY_FUNC_OFFSET_MASK;
560 size_t n = commonEncodings.size();
561 size_t wordsRemaining =
562 SECOND_LEVEL_PAGE_WORDS -
563 sizeof(unwind_info_compressed_second_level_page_header) /
564 sizeof(uint32_t);
565 while (wordsRemaining >= 1 && i < cuEntries.size()) {
566 const CompactUnwindEntry *cuPtr = &cuEntries[i];
567 if (cuPtr->functionAddress >= functionAddressMax)
568 break;
569 if (commonEncodingIndexes.count(Val: cuPtr->encoding) ||
570 page.localEncodingIndexes.count(Val: cuPtr->encoding)) {
571 i++;
572 wordsRemaining--;
573 } else if (wordsRemaining >= 2 && n < COMPACT_ENCODINGS_MAX) {
574 page.localEncodings.emplace_back(args: cuPtr->encoding);
575 page.localEncodingIndexes[cuPtr->encoding] = n++;
576 i++;
577 wordsRemaining -= 2;
578 } else {
579 break;
580 }
581 }
582 page.entryCount = i - page.entryIndex;
583
584 // If this is not the final page, see if it's possible to fit more entries
585 // by using the regular format. This can happen when there are many unique
586 // encodings, and we saturated the local encoding table early.
587 if (i < cuEntries.size() &&
588 page.entryCount < REGULAR_SECOND_LEVEL_ENTRIES_MAX) {
589 page.kind = UNWIND_SECOND_LEVEL_REGULAR;
590 page.entryCount = std::min(REGULAR_SECOND_LEVEL_ENTRIES_MAX,
591 b: cuEntries.size() - page.entryIndex);
592 i = page.entryIndex + page.entryCount;
593 } else {
594 page.kind = UNWIND_SECOND_LEVEL_COMPRESSED;
595 }
596 }
597
598 for (size_t i = 0; i < cuEntries.size(); ++i) {
599 lsdaIndex[i] = entriesWithLsda.size();
600 if (cuEntries[i].lsda)
601 entriesWithLsda.push_back(x: i);
602 }
603
604 // compute size of __TEXT,__unwind_info section
605 level2PagesOffset = sizeof(unwind_info_section_header) +
606 commonEncodings.size() * sizeof(uint32_t) +
607 personalities.size() * sizeof(uint32_t) +
608 // The extra second-level-page entry is for the sentinel
609 (secondLevelPages.size() + 1) *
610 sizeof(unwind_info_section_header_index_entry) +
611 entriesWithLsda.size() *
612 sizeof(unwind_info_section_header_lsda_index_entry);
613 unwindInfoSize =
614 level2PagesOffset + secondLevelPages.size() * SECOND_LEVEL_PAGE_BYTES;
615}
616
617// All inputs are relocated and output addresses are known, so write!
618
619void UnwindInfoSectionImpl::writeTo(uint8_t *buf) const {
620 assert(!cuEntries.empty() && "call only if there is unwind info");
621
622 // section header
623 auto *uip = reinterpret_cast<unwind_info_section_header *>(buf);
624 uip->version = 1;
625 uip->commonEncodingsArraySectionOffset = sizeof(unwind_info_section_header);
626 uip->commonEncodingsArrayCount = commonEncodings.size();
627 uip->personalityArraySectionOffset =
628 uip->commonEncodingsArraySectionOffset +
629 (uip->commonEncodingsArrayCount * sizeof(uint32_t));
630 uip->personalityArrayCount = personalities.size();
631 uip->indexSectionOffset = uip->personalityArraySectionOffset +
632 (uip->personalityArrayCount * sizeof(uint32_t));
633 uip->indexCount = secondLevelPages.size() + 1;
634
635 // Common encodings
636 auto *i32p = reinterpret_cast<uint32_t *>(&uip[1]);
637 for (const auto &encoding : commonEncodings)
638 *i32p++ = encoding.first;
639
640 // Personalities
641 for (const Symbol *personality : personalities)
642 *i32p++ = personality->getGotVA() - in.header->addr;
643
644 // FIXME: LD64 checks and warns aboutgaps or overlapse in cuEntries address
645 // ranges. We should do the same too
646
647 // Level-1 index
648 uint32_t lsdaOffset =
649 uip->indexSectionOffset +
650 uip->indexCount * sizeof(unwind_info_section_header_index_entry);
651 uint64_t l2PagesOffset = level2PagesOffset;
652 auto *iep = reinterpret_cast<unwind_info_section_header_index_entry *>(i32p);
653 for (const SecondLevelPage &page : secondLevelPages) {
654 size_t idx = page.entryIndex;
655 iep->functionOffset = cuEntries[idx].functionAddress - in.header->addr;
656 iep->secondLevelPagesSectionOffset = l2PagesOffset;
657 iep->lsdaIndexArraySectionOffset =
658 lsdaOffset + lsdaIndex.lookup(Val: idx) *
659 sizeof(unwind_info_section_header_lsda_index_entry);
660 iep++;
661 l2PagesOffset += SECOND_LEVEL_PAGE_BYTES;
662 }
663 // Level-1 sentinel
664 // XXX(vyng): Note that LD64 adds +1 here.
665 // Unsure whether it's a bug or it's their workaround for something else.
666 // See comments from https://reviews.llvm.org/D138320.
667 iep->functionOffset = cueEndBoundary - in.header->addr;
668 iep->secondLevelPagesSectionOffset = 0;
669 iep->lsdaIndexArraySectionOffset =
670 lsdaOffset + entriesWithLsda.size() *
671 sizeof(unwind_info_section_header_lsda_index_entry);
672 iep++;
673
674 // LSDAs
675 auto *lep =
676 reinterpret_cast<unwind_info_section_header_lsda_index_entry *>(iep);
677 for (size_t idx : entriesWithLsda) {
678 const CompactUnwindEntry &cu = cuEntries[idx];
679 lep->lsdaOffset = cu.lsda->getVA(/*off=*/0) - in.header->addr;
680 lep->functionOffset = cu.functionAddress - in.header->addr;
681 lep++;
682 }
683
684 // Level-2 pages
685 auto *pp = reinterpret_cast<uint32_t *>(lep);
686 for (const SecondLevelPage &page : secondLevelPages) {
687 if (page.kind == UNWIND_SECOND_LEVEL_COMPRESSED) {
688 uintptr_t functionAddressBase =
689 cuEntries[page.entryIndex].functionAddress;
690 auto *p2p =
691 reinterpret_cast<unwind_info_compressed_second_level_page_header *>(
692 pp);
693 p2p->kind = page.kind;
694 p2p->entryPageOffset =
695 sizeof(unwind_info_compressed_second_level_page_header);
696 p2p->entryCount = page.entryCount;
697 p2p->encodingsPageOffset =
698 p2p->entryPageOffset + p2p->entryCount * sizeof(uint32_t);
699 p2p->encodingsCount = page.localEncodings.size();
700 auto *ep = reinterpret_cast<uint32_t *>(&p2p[1]);
701 for (size_t i = 0; i < page.entryCount; i++) {
702 const CompactUnwindEntry &cue = cuEntries[page.entryIndex + i];
703 auto it = commonEncodingIndexes.find(Val: cue.encoding);
704 if (it == commonEncodingIndexes.end())
705 it = page.localEncodingIndexes.find(Val: cue.encoding);
706 *ep++ = (it->second << COMPRESSED_ENTRY_FUNC_OFFSET_BITS) |
707 (cue.functionAddress - functionAddressBase);
708 }
709 if (!page.localEncodings.empty())
710 memcpy(dest: ep, src: page.localEncodings.data(),
711 n: page.localEncodings.size() * sizeof(uint32_t));
712 } else {
713 auto *p2p =
714 reinterpret_cast<unwind_info_regular_second_level_page_header *>(pp);
715 p2p->kind = page.kind;
716 p2p->entryPageOffset =
717 sizeof(unwind_info_regular_second_level_page_header);
718 p2p->entryCount = page.entryCount;
719 auto *ep = reinterpret_cast<uint32_t *>(&p2p[1]);
720 for (size_t i = 0; i < page.entryCount; i++) {
721 const CompactUnwindEntry &cue = cuEntries[page.entryIndex + i];
722 *ep++ = cue.functionAddress;
723 *ep++ = cue.encoding;
724 }
725 }
726 pp += SECOND_LEVEL_PAGE_WORDS;
727 }
728}
729
730UnwindInfoSection *macho::makeUnwindInfoSection() {
731 return make<UnwindInfoSectionImpl>();
732}
733