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 // If ICF folded the LSDA, point at the surviving copy.
217 if (fde.lsda)
218 fde.lsda = fde.lsda->canonical();
219 }
220 }
221}
222
223// Compact unwind relocations have different semantics, so we handle them in a
224// separate code path from regular relocations. First, we do not wish to add
225// rebase opcodes for __LD,__compact_unwind, because that section doesn't
226// actually end up in the final binary. Second, personality pointers always
227// reside in the GOT and must be treated specially.
228void UnwindInfoSectionImpl::prepareRelocations(ConcatInputSection *isec) {
229 assert(!isec->shouldOmitFromOutput() &&
230 "__compact_unwind section should not be omitted");
231
232 // FIXME: Make this skip relocations for CompactUnwindEntries that
233 // point to dead-stripped functions. That might save some amount of
234 // work. But since there are usually just few personality functions
235 // that are referenced from many places, at least some of them likely
236 // live, it wouldn't reduce number of got entries.
237 for (size_t i = 0; i < isec->relocs.size(); ++i) {
238 Relocation &r = isec->relocs[i];
239 assert(target->hasAttr(r.type, RelocAttrBits::UNSIGNED));
240 // Since compact unwind sections aren't part of the inputSections vector,
241 // they don't get canonicalized by scanRelocations(), so we have to do the
242 // canonicalization here.
243 if (auto *referentIsec = r.referent.dyn_cast<InputSection *>())
244 r.referent = referentIsec->canonical();
245
246 // Functions and LSDA entries always reside in the same object file as the
247 // compact unwind entries that references them, and thus appear as section
248 // relocs. There is no need to prepare them. We only prepare relocs for
249 // personality functions.
250 if (r.offset != cuLayout.personalityOffset)
251 continue;
252
253 if (auto *s = r.referent.dyn_cast<Symbol *>()) {
254 // Personality functions are nearly always system-defined (e.g.,
255 // ___gxx_personality_v0 for C++) and relocated as dylib symbols. When an
256 // application provides its own personality function, it might be
257 // referenced by an extern Defined symbol reloc, or a local section reloc.
258 if (auto *defined = dyn_cast<Defined>(Val: s)) {
259 // XXX(vyng) This is a special case for handling duplicate personality
260 // symbols. Note that LD64's behavior is a bit different and it is
261 // inconsistent with how symbol resolution usually work
262 //
263 // So we've decided not to follow it. Instead, simply pick the symbol
264 // with the same name from the symbol table to replace the local one.
265 //
266 // (See discussions/alternatives already considered on D107533)
267 if (!defined->isExternal())
268 if (Symbol *sym = symtab->find(name: defined->getName()))
269 if (!sym->isLazy())
270 r.referent = s = sym;
271 }
272 if (auto *undefined = dyn_cast<Undefined>(Val: s)) {
273 treatUndefinedSymbol(*undefined, isec, offset: r.offset);
274 // treatUndefinedSymbol() can replace s with a DylibSymbol; re-check.
275 if (isa<Undefined>(Val: s))
276 continue;
277 }
278
279 // Similar to canonicalizePersonality(), but we also register a GOT entry.
280 if (auto *defined = dyn_cast<Defined>(Val: s)) {
281 // Check if we have created a synthetic symbol at the same address.
282 Symbol *&personality =
283 personalityTable[{defined->isec(), defined->value}];
284 if (personality == nullptr) {
285 personality = defined;
286 in.got->addEntry(sym: defined);
287 } else if (personality != defined) {
288 r.referent = personality;
289 }
290 continue;
291 }
292
293 assert(isa<DylibSymbol>(s));
294 in.got->addEntry(sym: s);
295 continue;
296 }
297
298 if (auto *referentIsec = r.referent.dyn_cast<InputSection *>()) {
299 assert(!isCoalescedWeak(referentIsec));
300 // Personality functions can be referenced via section relocations
301 // if they live in the same object file. Create placeholder synthetic
302 // symbols for them in the GOT. If the corresponding symbol is already
303 // in the GOT, use that to avoid creating a duplicate entry. All GOT
304 // entries needed by non-unwind sections will have already been added
305 // by this point.
306 Symbol *&s = personalityTable[{referentIsec, r.addend}];
307 if (s == nullptr) {
308 Defined *const *gotEntry =
309 llvm::find_if(Range&: referentIsec->symbols, P: [&](Defined const *d) {
310 return d->value == static_cast<uint64_t>(r.addend) &&
311 d->isInGot();
312 });
313 if (gotEntry != referentIsec->symbols.end()) {
314 s = *gotEntry;
315 } else {
316 // This runs after dead stripping, so the noDeadStrip argument does
317 // not matter.
318 s = make<Defined>(args: "<internal>", /*file=*/args: nullptr, args&: referentIsec,
319 args&: r.addend, /*size=*/args: 0, /*isWeakDef=*/args: false,
320 /*isExternal=*/args: false, /*isPrivateExtern=*/args: false,
321 /*includeInSymtab=*/args: true,
322 /*isReferencedDynamically=*/args: false,
323 /*noDeadStrip=*/args: false);
324 s->used = true;
325 in.got->addEntry(sym: s);
326 }
327 }
328 r.referent = s;
329 r.addend = 0;
330 }
331 }
332}
333
334Symbol *UnwindInfoSectionImpl::canonicalizePersonality(Symbol *personality) {
335 if (auto *defined = dyn_cast_or_null<Defined>(Val: personality)) {
336 // Check if we have created a synthetic symbol at the same address.
337 Symbol *&synth = personalityTable[{defined->isec(), defined->value}];
338 if (synth == nullptr)
339 synth = defined;
340 else if (synth != defined)
341 return synth;
342 }
343 return personality;
344}
345
346// We need to apply the relocations to the pre-link compact unwind section
347// before converting it to post-link form. There should only be absolute
348// relocations here: since we are not emitting the pre-link CU section, there
349// is no source address to make a relative location meaningful.
350void UnwindInfoSectionImpl::relocateCompactUnwind(
351 std::vector<CompactUnwindEntry> &cuEntries) {
352 parallelFor(Begin: 0, End: symbolsVec.size(), Fn: [&](size_t i) {
353 CompactUnwindEntry &cu = cuEntries[i];
354 const Defined *d = symbolsVec[i].second;
355 cu.functionAddress = d->getVA();
356 if (!d->unwindEntry())
357 return;
358
359 // If we have DWARF unwind info, create a slimmed-down CU entry that points
360 // to it.
361 if (d->unwindEntry()->getName() == section_names::ehFrame) {
362 // The unwinder will look for the DWARF entry starting at the hint,
363 // assuming the hint points to a valid CFI record start. If it
364 // fails to find the record, it proceeds in a linear search through the
365 // contiguous CFI records from the hint until the end of the section.
366 // Ideally, in the case where the offset is too large to be encoded, we
367 // would instead encode the largest possible offset to a valid CFI record,
368 // but since we don't keep track of that, just encode zero -- the start of
369 // the section is always the start of a CFI record.
370 uint64_t dwarfOffsetHint =
371 d->unwindEntry()->outSecOff <= DWARF_SECTION_OFFSET
372 ? d->unwindEntry()->outSecOff
373 : 0;
374 cu.encoding = target->modeDwarfEncoding | dwarfOffsetHint;
375 const FDE &fde = cast<ObjFile>(Val: d->getFile())->fdes[d->unwindEntry()];
376 cu.functionLength = fde.funcLength;
377 // Omit the DWARF personality from compact-unwind entry so that we
378 // don't need to encode it.
379 cu.personality = nullptr;
380 cu.lsda = fde.lsda;
381 return;
382 }
383
384 assert(d->unwindEntry()->getName() == section_names::compactUnwind);
385
386 auto buf =
387 reinterpret_cast<const uint8_t *>(d->unwindEntry()->data.data()) -
388 target->wordSize;
389 cu.functionLength =
390 support::endian::read32le(P: buf + cuLayout.functionLengthOffset);
391 cu.encoding = support::endian::read32le(P: buf + cuLayout.encodingOffset);
392 for (const Relocation &r : d->unwindEntry()->relocs) {
393 if (r.offset == cuLayout.personalityOffset)
394 cu.personality = cast<Symbol *>(Val: r.referent);
395 else if (r.offset == cuLayout.lsdaOffset)
396 cu.lsda = r.getReferentInputSection();
397 }
398 });
399}
400
401// There should only be a handful of unique personality pointers, so we can
402// encode them as 2-bit indices into a small array.
403void UnwindInfoSectionImpl::encodePersonalities() {
404 for (CompactUnwindEntry &cu : cuEntries) {
405 if (cu.personality == nullptr)
406 continue;
407 // Linear search is fast enough for a small array.
408 auto it = find(Range&: personalities, Val: cu.personality);
409 uint32_t personalityIndex; // 1-based index
410 if (it != personalities.end()) {
411 personalityIndex = std::distance(first: personalities.begin(), last: it) + 1;
412 } else {
413 personalities.push_back(x: cu.personality);
414 personalityIndex = personalities.size();
415 }
416 cu.encoding |=
417 personalityIndex << llvm::countr_zero(
418 Val: static_cast<compact_unwind_encoding_t>(UNWIND_PERSONALITY_MASK));
419 }
420 if (personalities.size() > 3)
421 error(msg: "too many personalities (" + Twine(personalities.size()) +
422 ") for compact unwind to encode");
423}
424
425static bool canFoldEncoding(compact_unwind_encoding_t encoding) {
426 // From compact_unwind_encoding.h:
427 // UNWIND_X86_64_MODE_STACK_IND:
428 // A "frameless" (RBP not used as frame pointer) function large constant
429 // stack size. This case is like the previous, except the stack size is too
430 // large to encode in the compact unwind encoding. Instead it requires that
431 // the function contains "subq $nnnnnnnn,RSP" in its prolog. The compact
432 // encoding contains the offset to the nnnnnnnn value in the function in
433 // UNWIND_X86_64_FRAMELESS_STACK_SIZE.
434 // Since this means the unwinder has to look at the `subq` in the function
435 // of the unwind info's unwind address, two functions that have identical
436 // unwind info can't be folded if it's using this encoding since both
437 // entries need unique addresses.
438 static_assert(static_cast<uint32_t>(UNWIND_X86_64_MODE_STACK_IND) ==
439 static_cast<uint32_t>(UNWIND_X86_MODE_STACK_IND));
440 if ((target->cpuType == CPU_TYPE_X86_64 || target->cpuType == CPU_TYPE_X86) &&
441 (encoding & UNWIND_MODE_MASK) == UNWIND_X86_64_MODE_STACK_IND) {
442 // FIXME: Consider passing in the two function addresses and getting
443 // their two stack sizes off the `subq` and only returning false if they're
444 // actually different.
445 return false;
446 }
447 return true;
448}
449
450// Scan the __LD,__compact_unwind entries and compute the space needs of
451// __TEXT,__unwind_info and __TEXT,__eh_frame.
452void UnwindInfoSectionImpl::finalize() {
453 if (symbols.empty())
454 return;
455
456 // At this point, the address space for __TEXT,__text has been
457 // assigned, so we can relocate the __LD,__compact_unwind entries
458 // into a temporary buffer. Relocation is necessary in order to sort
459 // the CU entries by function address. Sorting is necessary so that
460 // we can fold adjacent CU entries with identical encoding+personality
461 // and without any LSDA. Folding is necessary because it reduces the
462 // number of CU entries by as much as 3 orders of magnitude!
463 cuEntries.resize(new_size: symbols.size());
464 // The "map" part of the symbols MapVector was only needed for deduplication
465 // in addSymbol(). Now that we are done adding, move the contents to a plain
466 // std::vector for indexed access.
467 symbolsVec = symbols.takeVector();
468 relocateCompactUnwind(cuEntries);
469
470 // Sort the entries by address.
471 llvm::sort(C&: cuEntries, Comp: [&](auto &a, auto &b) {
472 return a.functionAddress < b.functionAddress;
473 });
474
475 // Record the ending boundary before we fold the entries.
476 cueEndBoundary =
477 cuEntries.back().functionAddress + cuEntries.back().functionLength;
478
479 // Fold adjacent entries with matching encoding+personality and without LSDA
480 // We use three iterators to fold in-situ:
481 // (1) `foldBegin` is the first of a potential sequence of matching entries
482 // (2) `foldEnd` is the first non-matching entry after `foldBegin`.
483 // The semi-open interval [ foldBegin .. foldEnd ) contains a range
484 // entries that can be folded into a single entry and written to ...
485 // (3) `foldWrite`
486 auto foldWrite = cuEntries.begin();
487 for (auto foldBegin = cuEntries.begin(); foldBegin != cuEntries.end();) {
488 auto foldEnd = foldBegin;
489 // Common LSDA encodings (e.g. for C++ and Objective-C) contain offsets from
490 // a base address. The base address is normally not contained directly in
491 // the LSDA, and in that case, the personality function treats the starting
492 // address of the function (which is computed by the unwinder) as the base
493 // address and interprets the LSDA accordingly. The unwinder computes the
494 // starting address of a function as the address associated with its CU
495 // entry. For this reason, we cannot fold adjacent entries if they have an
496 // LSDA, because folding would make the unwinder compute the wrong starting
497 // address for the functions with the folded entries, which in turn would
498 // cause the personality function to misinterpret the LSDA for those
499 // functions. In the very rare case where the base address is encoded
500 // directly in the LSDA, two functions at different addresses would
501 // necessarily have different LSDAs, so their CU entries would not have been
502 // folded anyway.
503 while (++foldEnd != cuEntries.end() &&
504 foldBegin->encoding == foldEnd->encoding && !foldBegin->lsda &&
505 !foldEnd->lsda &&
506 // If we've gotten to this point, we don't have an LSDA, which should
507 // also imply that we don't have a personality function, since in all
508 // likelihood a personality function needs the LSDA to do anything
509 // useful. It can be technically valid to have a personality function
510 // and no LSDA though (e.g. the C++ personality __gxx_personality_v0
511 // is just a no-op without LSDA), so we still check for personality
512 // function equivalence to handle that case.
513 foldBegin->personality == foldEnd->personality &&
514 canFoldEncoding(encoding: foldEnd->encoding))
515 ;
516 *foldWrite++ = *foldBegin;
517 foldBegin = foldEnd;
518 }
519 cuEntries.erase(first: foldWrite, last: cuEntries.end());
520
521 encodePersonalities();
522
523 // Count frequencies of the folded encodings
524 EncodingMap encodingFrequencies;
525 for (const CompactUnwindEntry &cu : cuEntries)
526 encodingFrequencies[cu.encoding]++;
527
528 // Make a vector of encodings, sorted by descending frequency
529 for (const auto &frequency : encodingFrequencies)
530 commonEncodings.emplace_back(args: frequency);
531 llvm::sort(C&: commonEncodings,
532 Comp: [](const std::pair<compact_unwind_encoding_t, size_t> &a,
533 const std::pair<compact_unwind_encoding_t, size_t> &b) {
534 // When frequencies match, secondarily sort on encoding
535 // to maintain parity with validate-unwind-info.py
536 return std::tie(args: a.second, args: a.first) > std::tie(args: b.second, args: b.first);
537 });
538
539 // Truncate the vector to 127 elements.
540 // Common encoding indexes are limited to 0..126, while encoding
541 // indexes 127..255 are local to each second-level page
542 if (commonEncodings.size() > COMMON_ENCODINGS_MAX)
543 commonEncodings.resize(COMMON_ENCODINGS_MAX);
544
545 // Create a map from encoding to common-encoding-table index
546 for (size_t i = 0; i < commonEncodings.size(); i++)
547 commonEncodingIndexes[commonEncodings[i].first] = i;
548
549 // Split folded encodings into pages, where each page is limited by ...
550 // (a) 4 KiB capacity
551 // (b) 24-bit difference between first & final function address
552 // (c) 8-bit compact-encoding-table index,
553 // for which 0..126 references the global common-encodings table,
554 // and 127..255 references a local per-second-level-page table.
555 // First we try the compact format and determine how many entries fit.
556 // If more entries fit in the regular format, we use that.
557 for (size_t i = 0; i < cuEntries.size();) {
558 secondLevelPages.emplace_back();
559 SecondLevelPage &page = secondLevelPages.back();
560 page.entryIndex = i;
561 uint64_t functionAddressMax =
562 cuEntries[i].functionAddress + COMPRESSED_ENTRY_FUNC_OFFSET_MASK;
563 size_t n = commonEncodings.size();
564 size_t wordsRemaining =
565 SECOND_LEVEL_PAGE_WORDS -
566 sizeof(unwind_info_compressed_second_level_page_header) /
567 sizeof(uint32_t);
568 while (wordsRemaining >= 1 && i < cuEntries.size()) {
569 const CompactUnwindEntry *cuPtr = &cuEntries[i];
570 if (cuPtr->functionAddress >= functionAddressMax)
571 break;
572 if (commonEncodingIndexes.count(Val: cuPtr->encoding) ||
573 page.localEncodingIndexes.count(Val: cuPtr->encoding)) {
574 i++;
575 wordsRemaining--;
576 } else if (wordsRemaining >= 2 && n < COMPACT_ENCODINGS_MAX) {
577 page.localEncodings.emplace_back(args: cuPtr->encoding);
578 page.localEncodingIndexes[cuPtr->encoding] = n++;
579 i++;
580 wordsRemaining -= 2;
581 } else {
582 break;
583 }
584 }
585 page.entryCount = i - page.entryIndex;
586
587 // If this is not the final page, see if it's possible to fit more entries
588 // by using the regular format. This can happen when there are many unique
589 // encodings, and we saturated the local encoding table early.
590 if (i < cuEntries.size() &&
591 page.entryCount < REGULAR_SECOND_LEVEL_ENTRIES_MAX) {
592 page.kind = UNWIND_SECOND_LEVEL_REGULAR;
593 page.entryCount = std::min(REGULAR_SECOND_LEVEL_ENTRIES_MAX,
594 b: cuEntries.size() - page.entryIndex);
595 i = page.entryIndex + page.entryCount;
596 } else {
597 page.kind = UNWIND_SECOND_LEVEL_COMPRESSED;
598 }
599 }
600
601 for (size_t i = 0; i < cuEntries.size(); ++i) {
602 lsdaIndex[i] = entriesWithLsda.size();
603 if (cuEntries[i].lsda)
604 entriesWithLsda.push_back(x: i);
605 }
606
607 // compute size of __TEXT,__unwind_info section
608 level2PagesOffset = sizeof(unwind_info_section_header) +
609 commonEncodings.size() * sizeof(uint32_t) +
610 personalities.size() * sizeof(uint32_t) +
611 // The extra second-level-page entry is for the sentinel
612 (secondLevelPages.size() + 1) *
613 sizeof(unwind_info_section_header_index_entry) +
614 entriesWithLsda.size() *
615 sizeof(unwind_info_section_header_lsda_index_entry);
616 unwindInfoSize =
617 level2PagesOffset + secondLevelPages.size() * SECOND_LEVEL_PAGE_BYTES;
618}
619
620// All inputs are relocated and output addresses are known, so write!
621
622void UnwindInfoSectionImpl::writeTo(uint8_t *buf) const {
623 assert(!cuEntries.empty() && "call only if there is unwind info");
624
625 // section header
626 auto *uip = reinterpret_cast<unwind_info_section_header *>(buf);
627 uip->version = 1;
628 uip->commonEncodingsArraySectionOffset = sizeof(unwind_info_section_header);
629 uip->commonEncodingsArrayCount = commonEncodings.size();
630 uip->personalityArraySectionOffset =
631 uip->commonEncodingsArraySectionOffset +
632 (uip->commonEncodingsArrayCount * sizeof(uint32_t));
633 uip->personalityArrayCount = personalities.size();
634 uip->indexSectionOffset = uip->personalityArraySectionOffset +
635 (uip->personalityArrayCount * sizeof(uint32_t));
636 uip->indexCount = secondLevelPages.size() + 1;
637
638 // Common encodings
639 auto *i32p = reinterpret_cast<uint32_t *>(&uip[1]);
640 for (const auto &encoding : commonEncodings)
641 *i32p++ = encoding.first;
642
643 // Personalities
644 for (const Symbol *personality : personalities)
645 *i32p++ = personality->getGotVA() - in.header->addr;
646
647 // FIXME: LD64 checks and warns aboutgaps or overlapse in cuEntries address
648 // ranges. We should do the same too
649
650 // Level-1 index
651 uint32_t lsdaOffset =
652 uip->indexSectionOffset +
653 uip->indexCount * sizeof(unwind_info_section_header_index_entry);
654 uint64_t l2PagesOffset = level2PagesOffset;
655 auto *iep = reinterpret_cast<unwind_info_section_header_index_entry *>(i32p);
656 for (const SecondLevelPage &page : secondLevelPages) {
657 size_t idx = page.entryIndex;
658 iep->functionOffset = cuEntries[idx].functionAddress - in.header->addr;
659 iep->secondLevelPagesSectionOffset = l2PagesOffset;
660 iep->lsdaIndexArraySectionOffset =
661 lsdaOffset + lsdaIndex.lookup(Val: idx) *
662 sizeof(unwind_info_section_header_lsda_index_entry);
663 iep++;
664 l2PagesOffset += SECOND_LEVEL_PAGE_BYTES;
665 }
666 // Level-1 sentinel
667 // XXX(vyng): Note that LD64 adds +1 here.
668 // Unsure whether it's a bug or it's their workaround for something else.
669 // See comments from https://reviews.llvm.org/D138320.
670 iep->functionOffset = cueEndBoundary - in.header->addr;
671 iep->secondLevelPagesSectionOffset = 0;
672 iep->lsdaIndexArraySectionOffset =
673 lsdaOffset + entriesWithLsda.size() *
674 sizeof(unwind_info_section_header_lsda_index_entry);
675 iep++;
676
677 // LSDAs
678 auto *lep =
679 reinterpret_cast<unwind_info_section_header_lsda_index_entry *>(iep);
680 for (size_t idx : entriesWithLsda) {
681 const CompactUnwindEntry &cu = cuEntries[idx];
682 lep->lsdaOffset = cu.lsda->getVA(/*off=*/0) - in.header->addr;
683 lep->functionOffset = cu.functionAddress - in.header->addr;
684 lep++;
685 }
686
687 // Level-2 pages
688 auto *pp = reinterpret_cast<uint32_t *>(lep);
689 for (const SecondLevelPage &page : secondLevelPages) {
690 if (page.kind == UNWIND_SECOND_LEVEL_COMPRESSED) {
691 uintptr_t functionAddressBase =
692 cuEntries[page.entryIndex].functionAddress;
693 auto *p2p =
694 reinterpret_cast<unwind_info_compressed_second_level_page_header *>(
695 pp);
696 p2p->kind = page.kind;
697 p2p->entryPageOffset =
698 sizeof(unwind_info_compressed_second_level_page_header);
699 p2p->entryCount = page.entryCount;
700 p2p->encodingsPageOffset =
701 p2p->entryPageOffset + p2p->entryCount * sizeof(uint32_t);
702 p2p->encodingsCount = page.localEncodings.size();
703 auto *ep = reinterpret_cast<uint32_t *>(&p2p[1]);
704 for (size_t i = 0; i < page.entryCount; i++) {
705 const CompactUnwindEntry &cue = cuEntries[page.entryIndex + i];
706 auto it = commonEncodingIndexes.find(Val: cue.encoding);
707 if (it == commonEncodingIndexes.end())
708 it = page.localEncodingIndexes.find(Val: cue.encoding);
709 *ep++ = (it->second << COMPRESSED_ENTRY_FUNC_OFFSET_BITS) |
710 (cue.functionAddress - functionAddressBase);
711 }
712 if (!page.localEncodings.empty())
713 memcpy(dest: ep, src: page.localEncodings.data(),
714 n: page.localEncodings.size() * sizeof(uint32_t));
715 } else {
716 auto *p2p =
717 reinterpret_cast<unwind_info_regular_second_level_page_header *>(pp);
718 p2p->kind = page.kind;
719 p2p->entryPageOffset =
720 sizeof(unwind_info_regular_second_level_page_header);
721 p2p->entryCount = page.entryCount;
722 auto *ep = reinterpret_cast<uint32_t *>(&p2p[1]);
723 for (size_t i = 0; i < page.entryCount; i++) {
724 const CompactUnwindEntry &cue = cuEntries[page.entryIndex + i];
725 *ep++ = cue.functionAddress;
726 *ep++ = cue.encoding;
727 }
728 }
729 pp += SECOND_LEVEL_PAGE_WORDS;
730 }
731}
732
733UnwindInfoSection *macho::makeUnwindInfoSection() {
734 return make<UnwindInfoSectionImpl>();
735}
736