1 | //===- OutputSections.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 "OutputSections.h" |
10 | #include "Config.h" |
11 | #include "InputFiles.h" |
12 | #include "LinkerScript.h" |
13 | #include "Symbols.h" |
14 | #include "SyntheticSections.h" |
15 | #include "Target.h" |
16 | #include "lld/Common/Arrays.h" |
17 | #include "lld/Common/Memory.h" |
18 | #include "llvm/BinaryFormat/Dwarf.h" |
19 | #include "llvm/Config/llvm-config.h" // LLVM_ENABLE_ZLIB |
20 | #include "llvm/Support/Compression.h" |
21 | #include "llvm/Support/LEB128.h" |
22 | #include "llvm/Support/Parallel.h" |
23 | #include "llvm/Support/Path.h" |
24 | #include "llvm/Support/TimeProfiler.h" |
25 | #if LLVM_ENABLE_ZLIB |
26 | // Avoid introducing max as a macro from Windows headers. |
27 | #define NOMINMAX |
28 | #include <zlib.h> |
29 | #endif |
30 | #if LLVM_ENABLE_ZSTD |
31 | #include <zstd.h> |
32 | #endif |
33 | |
34 | using namespace llvm; |
35 | using namespace llvm::dwarf; |
36 | using namespace llvm::object; |
37 | using namespace llvm::support::endian; |
38 | using namespace llvm::ELF; |
39 | using namespace lld; |
40 | using namespace lld::elf; |
41 | |
42 | uint8_t *Out::bufferStart; |
43 | PhdrEntry *Out::tlsPhdr; |
44 | OutputSection *Out::; |
45 | OutputSection *Out::; |
46 | OutputSection *Out::preinitArray; |
47 | OutputSection *Out::initArray; |
48 | OutputSection *Out::finiArray; |
49 | |
50 | SmallVector<OutputSection *, 0> elf::outputSections; |
51 | |
52 | uint32_t OutputSection::getPhdrFlags() const { |
53 | uint32_t ret = 0; |
54 | if (config->emachine != EM_ARM || !(flags & SHF_ARM_PURECODE)) |
55 | ret |= PF_R; |
56 | if (flags & SHF_WRITE) |
57 | ret |= PF_W; |
58 | if (flags & SHF_EXECINSTR) |
59 | ret |= PF_X; |
60 | return ret; |
61 | } |
62 | |
63 | template <class ELFT> |
64 | void OutputSection::(typename ELFT::Shdr *shdr) { |
65 | shdr->sh_entsize = entsize; |
66 | shdr->sh_addralign = addralign; |
67 | shdr->sh_type = type; |
68 | shdr->sh_offset = offset; |
69 | shdr->sh_flags = flags; |
70 | shdr->sh_info = info; |
71 | shdr->sh_link = link; |
72 | shdr->sh_addr = addr; |
73 | shdr->sh_size = size; |
74 | shdr->sh_name = shName; |
75 | } |
76 | |
77 | OutputSection::OutputSection(StringRef name, uint32_t type, uint64_t flags) |
78 | : SectionBase(Output, name, flags, /*Entsize*/ 0, /*Alignment*/ 1, type, |
79 | /*Info*/ 0, /*Link*/ 0) {} |
80 | |
81 | // We allow sections of types listed below to merged into a |
82 | // single progbits section. This is typically done by linker |
83 | // scripts. Merging nobits and progbits will force disk space |
84 | // to be allocated for nobits sections. Other ones don't require |
85 | // any special treatment on top of progbits, so there doesn't |
86 | // seem to be a harm in merging them. |
87 | // |
88 | // NOTE: clang since rL252300 emits SHT_X86_64_UNWIND .eh_frame sections. Allow |
89 | // them to be merged into SHT_PROGBITS .eh_frame (GNU as .cfi_*). |
90 | static bool canMergeToProgbits(unsigned type) { |
91 | return type == SHT_NOBITS || type == SHT_PROGBITS || type == SHT_INIT_ARRAY || |
92 | type == SHT_PREINIT_ARRAY || type == SHT_FINI_ARRAY || |
93 | type == SHT_NOTE || |
94 | (type == SHT_X86_64_UNWIND && config->emachine == EM_X86_64); |
95 | } |
96 | |
97 | // Record that isec will be placed in the OutputSection. isec does not become |
98 | // permanent until finalizeInputSections() is called. The function should not be |
99 | // used after finalizeInputSections() is called. If you need to add an |
100 | // InputSection post finalizeInputSections(), then you must do the following: |
101 | // |
102 | // 1. Find or create an InputSectionDescription to hold InputSection. |
103 | // 2. Add the InputSection to the InputSectionDescription::sections. |
104 | // 3. Call commitSection(isec). |
105 | void OutputSection::recordSection(InputSectionBase *isec) { |
106 | partition = isec->partition; |
107 | isec->parent = this; |
108 | if (commands.empty() || !isa<InputSectionDescription>(Val: commands.back())) |
109 | commands.push_back(Elt: make<InputSectionDescription>(args: "" )); |
110 | auto *isd = cast<InputSectionDescription>(Val: commands.back()); |
111 | isd->sectionBases.push_back(Elt: isec); |
112 | } |
113 | |
114 | // Update fields (type, flags, alignment, etc) according to the InputSection |
115 | // isec. Also check whether the InputSection flags and type are consistent with |
116 | // other InputSections. |
117 | void OutputSection::commitSection(InputSection *isec) { |
118 | if (LLVM_UNLIKELY(type != isec->type)) { |
119 | if (!hasInputSections && !typeIsSet) { |
120 | type = isec->type; |
121 | } else if (isStaticRelSecType(type) && isStaticRelSecType(type: isec->type) && |
122 | (type == SHT_CREL) != (isec->type == SHT_CREL)) { |
123 | // Combine mixed SHT_REL[A] and SHT_CREL to SHT_CREL. |
124 | type = SHT_CREL; |
125 | if (type == SHT_REL) { |
126 | if (name.consume_front(Prefix: ".rel" )) |
127 | name = saver().save(S: ".crel" + name); |
128 | } else if (name.consume_front(Prefix: ".rela" )) { |
129 | name = saver().save(S: ".crel" + name); |
130 | } |
131 | } else { |
132 | if (typeIsSet || !canMergeToProgbits(type) || |
133 | !canMergeToProgbits(type: isec->type)) { |
134 | // The (NOLOAD) changes the section type to SHT_NOBITS, the intention is |
135 | // that the contents at that address is provided by some other means. |
136 | // Some projects (e.g. |
137 | // https://github.com/ClangBuiltLinux/linux/issues/1597) rely on the |
138 | // behavior. Other types get an error. |
139 | if (type != SHT_NOBITS) { |
140 | errorOrWarn(msg: "section type mismatch for " + isec->name + "\n>>> " + |
141 | toString(isec) + ": " + |
142 | getELFSectionTypeName(Machine: config->emachine, Type: isec->type) + |
143 | "\n>>> output section " + name + ": " + |
144 | getELFSectionTypeName(Machine: config->emachine, Type: type)); |
145 | } |
146 | } |
147 | if (!typeIsSet) |
148 | type = SHT_PROGBITS; |
149 | } |
150 | } |
151 | if (!hasInputSections) { |
152 | // If IS is the first section to be added to this section, |
153 | // initialize type, entsize and flags from isec. |
154 | hasInputSections = true; |
155 | entsize = isec->entsize; |
156 | flags = isec->flags; |
157 | } else { |
158 | // Otherwise, check if new type or flags are compatible with existing ones. |
159 | if ((flags ^ isec->flags) & SHF_TLS) |
160 | error(msg: "incompatible section flags for " + name + "\n>>> " + |
161 | toString(isec) + ": 0x" + utohexstr(X: isec->flags) + |
162 | "\n>>> output section " + name + ": 0x" + utohexstr(X: flags)); |
163 | } |
164 | |
165 | isec->parent = this; |
166 | uint64_t andMask = |
167 | config->emachine == EM_ARM ? (uint64_t)SHF_ARM_PURECODE : 0; |
168 | uint64_t orMask = ~andMask; |
169 | uint64_t andFlags = (flags & isec->flags) & andMask; |
170 | uint64_t orFlags = (flags | isec->flags) & orMask; |
171 | flags = andFlags | orFlags; |
172 | if (nonAlloc) |
173 | flags &= ~(uint64_t)SHF_ALLOC; |
174 | |
175 | addralign = std::max(a: addralign, b: isec->addralign); |
176 | |
177 | // If this section contains a table of fixed-size entries, sh_entsize |
178 | // holds the element size. If it contains elements of different size we |
179 | // set sh_entsize to 0. |
180 | if (entsize != isec->entsize) |
181 | entsize = 0; |
182 | } |
183 | |
184 | static MergeSyntheticSection *createMergeSynthetic(StringRef name, |
185 | uint32_t type, |
186 | uint64_t flags, |
187 | uint32_t addralign) { |
188 | if ((flags & SHF_STRINGS) && config->optimize >= 2) |
189 | return make<MergeTailSection>(args&: name, args&: type, args&: flags, args&: addralign); |
190 | return make<MergeNoTailSection>(args&: name, args&: type, args&: flags, args&: addralign); |
191 | } |
192 | |
193 | // This function scans over the InputSectionBase list sectionBases to create |
194 | // InputSectionDescription::sections. |
195 | // |
196 | // It removes MergeInputSections from the input section array and adds |
197 | // new synthetic sections at the location of the first input section |
198 | // that it replaces. It then finalizes each synthetic section in order |
199 | // to compute an output offset for each piece of each input section. |
200 | void OutputSection::finalizeInputSections(LinkerScript *script) { |
201 | std::vector<MergeSyntheticSection *> mergeSections; |
202 | for (SectionCommand *cmd : commands) { |
203 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
204 | if (!isd) |
205 | continue; |
206 | isd->sections.reserve(N: isd->sectionBases.size()); |
207 | for (InputSectionBase *s : isd->sectionBases) { |
208 | MergeInputSection *ms = dyn_cast<MergeInputSection>(Val: s); |
209 | if (!ms) { |
210 | isd->sections.push_back(Elt: cast<InputSection>(Val: s)); |
211 | continue; |
212 | } |
213 | |
214 | // We do not want to handle sections that are not alive, so just remove |
215 | // them instead of trying to merge. |
216 | if (!ms->isLive()) |
217 | continue; |
218 | |
219 | auto i = llvm::find_if(Range&: mergeSections, P: [=](MergeSyntheticSection *sec) { |
220 | // While we could create a single synthetic section for two different |
221 | // values of Entsize, it is better to take Entsize into consideration. |
222 | // |
223 | // With a single synthetic section no two pieces with different Entsize |
224 | // could be equal, so we may as well have two sections. |
225 | // |
226 | // Using Entsize in here also allows us to propagate it to the synthetic |
227 | // section. |
228 | // |
229 | // SHF_STRINGS section with different alignments should not be merged. |
230 | return sec->flags == ms->flags && sec->entsize == ms->entsize && |
231 | (sec->addralign == ms->addralign || !(sec->flags & SHF_STRINGS)); |
232 | }); |
233 | if (i == mergeSections.end()) { |
234 | MergeSyntheticSection *syn = |
235 | createMergeSynthetic(name: s->name, type: ms->type, flags: ms->flags, addralign: ms->addralign); |
236 | mergeSections.push_back(x: syn); |
237 | i = std::prev(x: mergeSections.end()); |
238 | syn->entsize = ms->entsize; |
239 | isd->sections.push_back(Elt: syn); |
240 | // The merge synthetic section inherits the potential spill locations of |
241 | // its first contained section. |
242 | auto it = script->potentialSpillLists.find(Val: ms); |
243 | if (it != script->potentialSpillLists.end()) |
244 | script->potentialSpillLists.try_emplace(Key: syn, Args&: it->second); |
245 | } |
246 | (*i)->addSection(ms); |
247 | } |
248 | |
249 | // sectionBases should not be used from this point onwards. Clear it to |
250 | // catch misuses. |
251 | isd->sectionBases.clear(); |
252 | |
253 | // Some input sections may be removed from the list after ICF. |
254 | for (InputSection *s : isd->sections) |
255 | commitSection(isec: s); |
256 | } |
257 | for (auto *ms : mergeSections) |
258 | ms->finalizeContents(); |
259 | } |
260 | |
261 | static void sortByOrder(MutableArrayRef<InputSection *> in, |
262 | llvm::function_ref<int(InputSectionBase *s)> order) { |
263 | std::vector<std::pair<int, InputSection *>> v; |
264 | for (InputSection *s : in) |
265 | v.emplace_back(args: order(s), args&: s); |
266 | llvm::stable_sort(Range&: v, C: less_first()); |
267 | |
268 | for (size_t i = 0; i < v.size(); ++i) |
269 | in[i] = v[i].second; |
270 | } |
271 | |
272 | uint64_t elf::() { |
273 | if (config->oFormatBinary) |
274 | return 0; |
275 | return Out::elfHeader->size + Out::programHeaders->size; |
276 | } |
277 | |
278 | void OutputSection::sort(llvm::function_ref<int(InputSectionBase *s)> order) { |
279 | assert(isLive()); |
280 | for (SectionCommand *b : commands) |
281 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: b)) |
282 | sortByOrder(in: isd->sections, order); |
283 | } |
284 | |
285 | static void nopInstrFill(uint8_t *buf, size_t size) { |
286 | if (size == 0) |
287 | return; |
288 | unsigned i = 0; |
289 | if (size == 0) |
290 | return; |
291 | std::vector<std::vector<uint8_t>> nopFiller = *target->nopInstrs; |
292 | unsigned num = size / nopFiller.back().size(); |
293 | for (unsigned c = 0; c < num; ++c) { |
294 | memcpy(dest: buf + i, src: nopFiller.back().data(), n: nopFiller.back().size()); |
295 | i += nopFiller.back().size(); |
296 | } |
297 | unsigned remaining = size - i; |
298 | if (!remaining) |
299 | return; |
300 | assert(nopFiller[remaining - 1].size() == remaining); |
301 | memcpy(dest: buf + i, src: nopFiller[remaining - 1].data(), n: remaining); |
302 | } |
303 | |
304 | // Fill [Buf, Buf + Size) with Filler. |
305 | // This is used for linker script "=fillexp" command. |
306 | static void fill(uint8_t *buf, size_t size, |
307 | const std::array<uint8_t, 4> &filler) { |
308 | size_t i = 0; |
309 | for (; i + 4 < size; i += 4) |
310 | memcpy(dest: buf + i, src: filler.data(), n: 4); |
311 | memcpy(dest: buf + i, src: filler.data(), n: size - i); |
312 | } |
313 | |
314 | #if LLVM_ENABLE_ZLIB |
315 | static SmallVector<uint8_t, 0> deflateShard(ArrayRef<uint8_t> in, int level, |
316 | int flush) { |
317 | // 15 and 8 are default. windowBits=-15 is negative to generate raw deflate |
318 | // data with no zlib header or trailer. |
319 | z_stream s = {}; |
320 | auto res = deflateInit2(&s, level, Z_DEFLATED, -15, 8, Z_DEFAULT_STRATEGY); |
321 | if (res != 0) { |
322 | errorOrWarn(msg: "--compress-sections: deflateInit2 returned " + Twine(res)); |
323 | return {}; |
324 | } |
325 | s.next_in = const_cast<uint8_t *>(in.data()); |
326 | s.avail_in = in.size(); |
327 | |
328 | // Allocate a buffer of half of the input size, and grow it by 1.5x if |
329 | // insufficient. |
330 | SmallVector<uint8_t, 0> out; |
331 | size_t pos = 0; |
332 | out.resize_for_overwrite(N: std::max<size_t>(a: in.size() / 2, b: 64)); |
333 | do { |
334 | if (pos == out.size()) |
335 | out.resize_for_overwrite(N: out.size() * 3 / 2); |
336 | s.next_out = out.data() + pos; |
337 | s.avail_out = out.size() - pos; |
338 | (void)deflate(strm: &s, flush); |
339 | pos = s.next_out - out.data(); |
340 | } while (s.avail_out == 0); |
341 | assert(s.avail_in == 0); |
342 | |
343 | out.truncate(N: pos); |
344 | deflateEnd(strm: &s); |
345 | return out; |
346 | } |
347 | #endif |
348 | |
349 | // Compress certain non-SHF_ALLOC sections: |
350 | // |
351 | // * (if --compress-debug-sections is specified) non-empty .debug_* sections |
352 | // * (if --compress-sections is specified) matched sections |
353 | template <class ELFT> void OutputSection::maybeCompress() { |
354 | using Elf_Chdr = typename ELFT::Chdr; |
355 | (void)sizeof(Elf_Chdr); |
356 | |
357 | DebugCompressionType ctype = DebugCompressionType::None; |
358 | size_t compressedSize = sizeof(Elf_Chdr); |
359 | unsigned level = 0; // default compression level |
360 | if (!(flags & SHF_ALLOC) && config->compressDebugSections && |
361 | name.starts_with(Prefix: ".debug_" )) |
362 | ctype = *config->compressDebugSections; |
363 | for (auto &[glob, t, l] : config->compressSections) |
364 | if (glob.match(S: name)) |
365 | std::tie(args&: ctype, args&: level) = {t, l}; |
366 | if (ctype == DebugCompressionType::None) |
367 | return; |
368 | if (flags & SHF_ALLOC) { |
369 | errorOrWarn(msg: "--compress-sections: section '" + name + |
370 | "' with the SHF_ALLOC flag cannot be compressed" ); |
371 | return; |
372 | } |
373 | |
374 | llvm::TimeTraceScope timeScope("Compress sections" ); |
375 | auto buf = std::make_unique<uint8_t[]>(num: size); |
376 | // Write uncompressed data to a temporary zero-initialized buffer. |
377 | { |
378 | parallel::TaskGroup tg; |
379 | writeTo<ELFT>(buf.get(), tg); |
380 | } |
381 | // The generic ABI specifies "The sh_size and sh_addralign fields of the |
382 | // section header for a compressed section reflect the requirements of the |
383 | // compressed section." However, 1-byte alignment has been wildly accepted |
384 | // and utilized for a long time. Removing alignment padding is particularly |
385 | // useful when there are many compressed output sections. |
386 | addralign = 1; |
387 | |
388 | // Split input into 1-MiB shards. |
389 | [[maybe_unused]] constexpr size_t shardSize = 1 << 20; |
390 | auto shardsIn = split(arr: ArrayRef<uint8_t>(buf.get(), size), chunkSize: shardSize); |
391 | const size_t numShards = shardsIn.size(); |
392 | auto shardsOut = std::make_unique<SmallVector<uint8_t, 0>[]>(num: numShards); |
393 | |
394 | #if LLVM_ENABLE_ZSTD |
395 | // Use ZSTD's streaming compression API. See |
396 | // http://facebook.github.io/zstd/zstd_manual.html "Streaming compression - |
397 | // HowTo". |
398 | if (ctype == DebugCompressionType::Zstd) { |
399 | parallelFor(0, numShards, [&](size_t i) { |
400 | SmallVector<uint8_t, 0> out; |
401 | ZSTD_CCtx *cctx = ZSTD_createCCtx(); |
402 | ZSTD_CCtx_setParameter(cctx, param: ZSTD_c_compressionLevel, value: level); |
403 | ZSTD_inBuffer zib = {.src: shardsIn[i].data(), .size: shardsIn[i].size(), .pos: 0}; |
404 | ZSTD_outBuffer zob = {.dst: nullptr, .size: 0, .pos: 0}; |
405 | size_t size; |
406 | do { |
407 | // Allocate a buffer of half of the input size, and grow it by 1.5x if |
408 | // insufficient. |
409 | if (zob.pos == zob.size) { |
410 | out.resize_for_overwrite( |
411 | N: zob.size ? zob.size * 3 / 2 : std::max<size_t>(a: zib.size / 4, b: 64)); |
412 | zob = {.dst: out.data(), .size: out.size(), .pos: zob.pos}; |
413 | } |
414 | size = ZSTD_compressStream2(cctx, output: &zob, input: &zib, endOp: ZSTD_e_end); |
415 | assert(!ZSTD_isError(size)); |
416 | } while (size != 0); |
417 | out.truncate(N: zob.pos); |
418 | ZSTD_freeCCtx(cctx); |
419 | shardsOut[i] = std::move(out); |
420 | }); |
421 | compressed.type = ELFCOMPRESS_ZSTD; |
422 | for (size_t i = 0; i != numShards; ++i) |
423 | compressedSize += shardsOut[i].size(); |
424 | } |
425 | #endif |
426 | |
427 | #if LLVM_ENABLE_ZLIB |
428 | // We chose 1 (Z_BEST_SPEED) as the default compression level because it is |
429 | // fast and provides decent compression ratios. |
430 | if (ctype == DebugCompressionType::Zlib) { |
431 | if (!level) |
432 | level = Z_BEST_SPEED; |
433 | |
434 | // Compress shards and compute Alder-32 checksums. Use Z_SYNC_FLUSH for all |
435 | // shards but the last to flush the output to a byte boundary to be |
436 | // concatenated with the next shard. |
437 | auto shardsAdler = std::make_unique<uint32_t[]>(num: numShards); |
438 | parallelFor(0, numShards, [&](size_t i) { |
439 | shardsOut[i] = deflateShard(in: shardsIn[i], level, |
440 | flush: i != numShards - 1 ? Z_SYNC_FLUSH : Z_FINISH); |
441 | shardsAdler[i] = adler32(adler: 1, buf: shardsIn[i].data(), len: shardsIn[i].size()); |
442 | }); |
443 | |
444 | // Update section size and combine Alder-32 checksums. |
445 | uint32_t checksum = 1; // Initial Adler-32 value |
446 | compressedSize += 2; // Elf_Chdir and zlib header |
447 | for (size_t i = 0; i != numShards; ++i) { |
448 | compressedSize += shardsOut[i].size(); |
449 | checksum = adler32_combine(checksum, shardsAdler[i], shardsIn[i].size()); |
450 | } |
451 | compressedSize += 4; // checksum |
452 | compressed.type = ELFCOMPRESS_ZLIB; |
453 | compressed.checksum = checksum; |
454 | } |
455 | #endif |
456 | |
457 | if (compressedSize >= size) |
458 | return; |
459 | compressed.uncompressedSize = size; |
460 | compressed.shards = std::move(shardsOut); |
461 | compressed.numShards = numShards; |
462 | size = compressedSize; |
463 | flags |= SHF_COMPRESSED; |
464 | } |
465 | |
466 | static void writeInt(uint8_t *buf, uint64_t data, uint64_t size) { |
467 | if (size == 1) |
468 | *buf = data; |
469 | else if (size == 2) |
470 | write16(p: buf, v: data); |
471 | else if (size == 4) |
472 | write32(p: buf, v: data); |
473 | else if (size == 8) |
474 | write64(p: buf, v: data); |
475 | else |
476 | llvm_unreachable("unsupported Size argument" ); |
477 | } |
478 | |
479 | template <class ELFT> |
480 | void OutputSection::writeTo(uint8_t *buf, parallel::TaskGroup &tg) { |
481 | llvm::TimeTraceScope timeScope("Write sections" , name); |
482 | if (type == SHT_NOBITS) |
483 | return; |
484 | if (type == SHT_CREL && !(flags & SHF_ALLOC)) { |
485 | buf += encodeULEB128(Value: crelHeader, p: buf); |
486 | memcpy(dest: buf, src: crelBody.data(), n: crelBody.size()); |
487 | return; |
488 | } |
489 | |
490 | // If the section is compressed due to |
491 | // --compress-debug-section/--compress-sections, the content is already known. |
492 | if (compressed.shards) { |
493 | auto *chdr = reinterpret_cast<typename ELFT::Chdr *>(buf); |
494 | chdr->ch_type = compressed.type; |
495 | chdr->ch_size = compressed.uncompressedSize; |
496 | chdr->ch_addralign = addralign; |
497 | buf += sizeof(*chdr); |
498 | |
499 | auto offsets = std::make_unique<size_t[]>(num: compressed.numShards); |
500 | if (compressed.type == ELFCOMPRESS_ZLIB) { |
501 | buf[0] = 0x78; // CMF |
502 | buf[1] = 0x01; // FLG: best speed |
503 | offsets[0] = 2; // zlib header |
504 | write32be(P: buf + (size - sizeof(*chdr) - 4), V: compressed.checksum); |
505 | } |
506 | |
507 | // Compute shard offsets. |
508 | for (size_t i = 1; i != compressed.numShards; ++i) |
509 | offsets[i] = offsets[i - 1] + compressed.shards[i - 1].size(); |
510 | parallelFor(0, compressed.numShards, [&](size_t i) { |
511 | memcpy(dest: buf + offsets[i], src: compressed.shards[i].data(), |
512 | n: compressed.shards[i].size()); |
513 | }); |
514 | return; |
515 | } |
516 | |
517 | // Write leading padding. |
518 | ArrayRef<InputSection *> sections = getInputSections(os: *this, storage); |
519 | std::array<uint8_t, 4> filler = getFiller(); |
520 | bool nonZeroFiller = read32(p: filler.data()) != 0; |
521 | if (nonZeroFiller) |
522 | fill(buf, size: sections.empty() ? size : sections[0]->outSecOff, filler); |
523 | |
524 | if (type == SHT_CREL && !(flags & SHF_ALLOC)) { |
525 | buf += encodeULEB128(Value: crelHeader, p: buf); |
526 | memcpy(dest: buf, src: crelBody.data(), n: crelBody.size()); |
527 | return; |
528 | } |
529 | |
530 | auto fn = [=](size_t begin, size_t end) { |
531 | size_t numSections = sections.size(); |
532 | for (size_t i = begin; i != end; ++i) { |
533 | InputSection *isec = sections[i]; |
534 | if (auto *s = dyn_cast<SyntheticSection>(Val: isec)) |
535 | s->writeTo(buf: buf + isec->outSecOff); |
536 | else |
537 | isec->writeTo<ELFT>(buf + isec->outSecOff); |
538 | |
539 | // When in Arm BE8 mode, the linker has to convert the big-endian |
540 | // instructions to little-endian, leaving the data big-endian. |
541 | if (config->emachine == EM_ARM && !config->isLE && config->armBe8 && |
542 | (flags & SHF_EXECINSTR)) |
543 | convertArmInstructionstoBE8(sec: isec, buf: buf + isec->outSecOff); |
544 | |
545 | // Fill gaps between sections. |
546 | if (nonZeroFiller) { |
547 | uint8_t *start = buf + isec->outSecOff + isec->getSize(); |
548 | uint8_t *end; |
549 | if (i + 1 == numSections) |
550 | end = buf + size; |
551 | else |
552 | end = buf + sections[i + 1]->outSecOff; |
553 | if (isec->nopFiller) { |
554 | assert(target->nopInstrs); |
555 | nopInstrFill(buf: start, size: end - start); |
556 | } else |
557 | fill(buf: start, size: end - start, filler); |
558 | } |
559 | } |
560 | }; |
561 | |
562 | // If there is any BYTE()-family command (rare), write the section content |
563 | // first then process BYTE to overwrite the filler content. The write is |
564 | // serial due to the limitation of llvm/Support/Parallel.h. |
565 | bool written = false; |
566 | size_t numSections = sections.size(); |
567 | for (SectionCommand *cmd : commands) |
568 | if (auto *data = dyn_cast<ByteCommand>(Val: cmd)) { |
569 | if (!std::exchange(obj&: written, new_val: true)) |
570 | fn(0, numSections); |
571 | writeInt(buf: buf + data->offset, data: data->expression().getValue(), size: data->size); |
572 | } |
573 | if (written || !numSections) |
574 | return; |
575 | |
576 | // There is no data command. Write content asynchronously to overlap the write |
577 | // time with other output sections. Note, if a linker script specifies |
578 | // overlapping output sections (needs --noinhibit-exec or --no-check-sections |
579 | // to supress the error), the output may be non-deterministic. |
580 | const size_t taskSizeLimit = 4 << 20; |
581 | for (size_t begin = 0, i = 0, taskSize = 0;;) { |
582 | taskSize += sections[i]->getSize(); |
583 | bool done = ++i == numSections; |
584 | if (done || taskSize >= taskSizeLimit) { |
585 | tg.spawn(f: [=] { fn(begin, i); }); |
586 | if (done) |
587 | break; |
588 | begin = i; |
589 | taskSize = 0; |
590 | } |
591 | } |
592 | } |
593 | |
594 | static void finalizeShtGroup(OutputSection *os, InputSection *section) { |
595 | // sh_link field for SHT_GROUP sections should contain the section index of |
596 | // the symbol table. |
597 | os->link = in.symTab->getParent()->sectionIndex; |
598 | |
599 | if (!section) |
600 | return; |
601 | |
602 | // sh_info then contain index of an entry in symbol table section which |
603 | // provides signature of the section group. |
604 | ArrayRef<Symbol *> symbols = section->file->getSymbols(); |
605 | os->info = in.symTab->getSymbolIndex(sym: *symbols[section->info]); |
606 | |
607 | // Some group members may be combined or discarded, so we need to compute the |
608 | // new size. The content will be rewritten in InputSection::copyShtGroup. |
609 | DenseSet<uint32_t> seen; |
610 | ArrayRef<InputSectionBase *> sections = section->file->getSections(); |
611 | for (const uint32_t &idx : section->getDataAs<uint32_t>().slice(N: 1)) |
612 | if (OutputSection *osec = sections[read32(p: &idx)]->getOutputSection()) |
613 | seen.insert(V: osec->sectionIndex); |
614 | os->size = (1 + seen.size()) * sizeof(uint32_t); |
615 | } |
616 | |
617 | template <class uint> |
618 | LLVM_ATTRIBUTE_ALWAYS_INLINE static void |
619 | encodeOneCrel(raw_svector_ostream &os, Elf_Crel<sizeof(uint) == 8> &out, |
620 | uint offset, const Symbol &sym, uint32_t type, uint addend) { |
621 | const auto deltaOffset = static_cast<uint64_t>(offset - out.r_offset); |
622 | out.r_offset = offset; |
623 | int64_t symidx = in.symTab->getSymbolIndex(sym); |
624 | if (sym.type == STT_SECTION) { |
625 | auto *d = dyn_cast<Defined>(Val: &sym); |
626 | if (d) { |
627 | SectionBase *section = d->section; |
628 | assert(section->isLive()); |
629 | addend = sym.getVA(addend) - section->getOutputSection()->addr; |
630 | } else { |
631 | // Encode R_*_NONE(symidx=0). |
632 | symidx = type = addend = 0; |
633 | } |
634 | } |
635 | |
636 | // Similar to llvm::ELF::encodeCrel. |
637 | uint8_t b = deltaOffset * 8 + (out.r_symidx != symidx) + |
638 | (out.r_type != type ? 2 : 0) + |
639 | (uint(out.r_addend) != addend ? 4 : 0); |
640 | if (deltaOffset < 0x10) { |
641 | os << char(b); |
642 | } else { |
643 | os << char(b | 0x80); |
644 | encodeULEB128(Value: deltaOffset >> 4, OS&: os); |
645 | } |
646 | if (b & 1) { |
647 | encodeSLEB128(Value: static_cast<int32_t>(symidx - out.r_symidx), OS&: os); |
648 | out.r_symidx = symidx; |
649 | } |
650 | if (b & 2) { |
651 | encodeSLEB128(Value: static_cast<int32_t>(type - out.r_type), OS&: os); |
652 | out.r_type = type; |
653 | } |
654 | if (b & 4) { |
655 | encodeSLEB128(std::make_signed_t<uint>(addend - out.r_addend), os); |
656 | out.r_addend = addend; |
657 | } |
658 | } |
659 | |
660 | template <class ELFT> |
661 | static size_t relToCrel(raw_svector_ostream &os, Elf_Crel<ELFT::Is64Bits> &out, |
662 | InputSection *relSec, InputSectionBase *sec) { |
663 | const auto &file = *cast<ELFFileBase>(Val: relSec->file); |
664 | if (relSec->type == SHT_REL) { |
665 | // REL conversion is complex and unsupported yet. |
666 | errorOrWarn(msg: toString(relSec) + ": REL cannot be converted to CREL" ); |
667 | return 0; |
668 | } |
669 | auto rels = relSec->getDataAs<typename ELFT::Rela>(); |
670 | for (auto rel : rels) { |
671 | encodeOneCrel<typename ELFT::uint>( |
672 | os, out, sec->getVA(offset: rel.r_offset), file.getRelocTargetSym(rel), |
673 | rel.getType(config->isMips64EL), getAddend<ELFT>(rel)); |
674 | } |
675 | return rels.size(); |
676 | } |
677 | |
678 | // Compute the content of a non-alloc CREL section due to -r or --emit-relocs. |
679 | // Input CREL sections are decoded while REL[A] need to be converted. |
680 | template <bool is64> void OutputSection::finalizeNonAllocCrel() { |
681 | using uint = typename Elf_Crel_Impl<is64>::uint; |
682 | raw_svector_ostream os(crelBody); |
683 | uint64_t totalCount = 0; |
684 | Elf_Crel<is64> out{}; |
685 | assert(commands.size() == 1); |
686 | auto *isd = cast<InputSectionDescription>(Val: commands[0]); |
687 | for (InputSection *relSec : isd->sections) { |
688 | const auto &file = *cast<ELFFileBase>(Val: relSec->file); |
689 | InputSectionBase *sec = relSec->getRelocatedSection(); |
690 | if (relSec->type == SHT_CREL) { |
691 | RelocsCrel<is64> entries(relSec->content_); |
692 | totalCount += entries.size(); |
693 | for (Elf_Crel_Impl<is64> r : entries) { |
694 | encodeOneCrel<uint>(os, out, uint(sec->getVA(offset: r.r_offset)), |
695 | file.getSymbol(symbolIndex: r.r_symidx), r.r_type, r.r_addend); |
696 | } |
697 | continue; |
698 | } |
699 | |
700 | // Convert REL[A] to CREL. |
701 | if constexpr (is64) { |
702 | totalCount += config->isLE ? relToCrel<ELF64LE>(os, out, relSec, sec) |
703 | : relToCrel<ELF64BE>(os, out, relSec, sec); |
704 | } else { |
705 | totalCount += config->isLE ? relToCrel<ELF32LE>(os, out, relSec, sec) |
706 | : relToCrel<ELF32BE>(os, out, relSec, sec); |
707 | } |
708 | } |
709 | |
710 | crelHeader = totalCount * 8 + 4; |
711 | size = getULEB128Size(Value: crelHeader) + crelBody.size(); |
712 | } |
713 | |
714 | void OutputSection::finalize() { |
715 | InputSection *first = getFirstInputSection(os: this); |
716 | |
717 | if (flags & SHF_LINK_ORDER) { |
718 | // We must preserve the link order dependency of sections with the |
719 | // SHF_LINK_ORDER flag. The dependency is indicated by the sh_link field. We |
720 | // need to translate the InputSection sh_link to the OutputSection sh_link, |
721 | // all InputSections in the OutputSection have the same dependency. |
722 | if (auto *ex = dyn_cast<ARMExidxSyntheticSection>(Val: first)) |
723 | link = ex->getLinkOrderDep()->getParent()->sectionIndex; |
724 | else if (first->flags & SHF_LINK_ORDER) |
725 | if (auto *d = first->getLinkOrderDep()) |
726 | link = d->getParent()->sectionIndex; |
727 | } |
728 | |
729 | if (type == SHT_GROUP) { |
730 | finalizeShtGroup(os: this, section: first); |
731 | return; |
732 | } |
733 | |
734 | if (!config->copyRelocs || !isStaticRelSecType(type)) |
735 | return; |
736 | |
737 | // Skip if 'first' is synthetic, i.e. not a section created by --emit-relocs. |
738 | // Normally 'type' was changed by 'first' so 'first' should be non-null. |
739 | // However, if the output section is .rela.dyn, 'type' can be set by the empty |
740 | // synthetic .rela.plt and first can be null. |
741 | if (!first || isa<SyntheticSection>(Val: first)) |
742 | return; |
743 | |
744 | link = in.symTab->getParent()->sectionIndex; |
745 | // sh_info for SHT_REL[A] sections should contain the section header index of |
746 | // the section to which the relocation applies. |
747 | InputSectionBase *s = first->getRelocatedSection(); |
748 | info = s->getOutputSection()->sectionIndex; |
749 | flags |= SHF_INFO_LINK; |
750 | // Finalize the content of non-alloc CREL. |
751 | if (type == SHT_CREL) { |
752 | if (config->is64) |
753 | finalizeNonAllocCrel<true>(); |
754 | else |
755 | finalizeNonAllocCrel<false>(); |
756 | } |
757 | } |
758 | |
759 | // Returns true if S is in one of the many forms the compiler driver may pass |
760 | // crtbegin files. |
761 | // |
762 | // Gcc uses any of crtbegin[<empty>|S|T].o. |
763 | // Clang uses Gcc's plus clang_rt.crtbegin[-<arch>|<empty>].o. |
764 | |
765 | static bool isCrt(StringRef s, StringRef beginEnd) { |
766 | s = sys::path::filename(path: s); |
767 | if (!s.consume_back(Suffix: ".o" )) |
768 | return false; |
769 | if (s.consume_front(Prefix: "clang_rt." )) |
770 | return s.consume_front(Prefix: beginEnd); |
771 | return s.consume_front(Prefix: beginEnd) && s.size() <= 1; |
772 | } |
773 | |
774 | // .ctors and .dtors are sorted by this order: |
775 | // |
776 | // 1. .ctors/.dtors in crtbegin (which contains a sentinel value -1). |
777 | // 2. The section is named ".ctors" or ".dtors" (priority: 65536). |
778 | // 3. The section has an optional priority value in the form of ".ctors.N" or |
779 | // ".dtors.N" where N is a number in the form of %05u (priority: 65535-N). |
780 | // 4. .ctors/.dtors in crtend (which contains a sentinel value 0). |
781 | // |
782 | // For 2 and 3, the sections are sorted by priority from high to low, e.g. |
783 | // .ctors (65536), .ctors.00100 (65436), .ctors.00200 (65336). In GNU ld's |
784 | // internal linker scripts, the sorting is by string comparison which can |
785 | // achieve the same goal given the optional priority values are of the same |
786 | // length. |
787 | // |
788 | // In an ideal world, we don't need this function because .init_array and |
789 | // .ctors are duplicate features (and .init_array is newer.) However, there |
790 | // are too many real-world use cases of .ctors, so we had no choice to |
791 | // support that with this rather ad-hoc semantics. |
792 | static bool compCtors(const InputSection *a, const InputSection *b) { |
793 | bool beginA = isCrt(s: a->file->getName(), beginEnd: "crtbegin" ); |
794 | bool beginB = isCrt(s: b->file->getName(), beginEnd: "crtbegin" ); |
795 | if (beginA != beginB) |
796 | return beginA; |
797 | bool endA = isCrt(s: a->file->getName(), beginEnd: "crtend" ); |
798 | bool endB = isCrt(s: b->file->getName(), beginEnd: "crtend" ); |
799 | if (endA != endB) |
800 | return endB; |
801 | return getPriority(s: a->name) > getPriority(s: b->name); |
802 | } |
803 | |
804 | // Sorts input sections by the special rules for .ctors and .dtors. |
805 | // Unfortunately, the rules are different from the one for .{init,fini}_array. |
806 | // Read the comment above. |
807 | void OutputSection::sortCtorsDtors() { |
808 | assert(commands.size() == 1); |
809 | auto *isd = cast<InputSectionDescription>(Val: commands[0]); |
810 | llvm::stable_sort(Range&: isd->sections, C: compCtors); |
811 | } |
812 | |
813 | // If an input string is in the form of "foo.N" where N is a number, return N |
814 | // (65535-N if .ctors.N or .dtors.N). Otherwise, returns 65536, which is one |
815 | // greater than the lowest priority. |
816 | int elf::getPriority(StringRef s) { |
817 | size_t pos = s.rfind(C: '.'); |
818 | if (pos == StringRef::npos) |
819 | return 65536; |
820 | int v = 65536; |
821 | if (to_integer(S: s.substr(Start: pos + 1), Num&: v, Base: 10) && |
822 | (pos == 6 && (s.starts_with(Prefix: ".ctors" ) || s.starts_with(Prefix: ".dtors" )))) |
823 | v = 65535 - v; |
824 | return v; |
825 | } |
826 | |
827 | InputSection *elf::getFirstInputSection(const OutputSection *os) { |
828 | for (SectionCommand *cmd : os->commands) |
829 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) |
830 | if (!isd->sections.empty()) |
831 | return isd->sections[0]; |
832 | return nullptr; |
833 | } |
834 | |
835 | ArrayRef<InputSection *> |
836 | elf::getInputSections(const OutputSection &os, |
837 | SmallVector<InputSection *, 0> &storage) { |
838 | ArrayRef<InputSection *> ret; |
839 | storage.clear(); |
840 | for (SectionCommand *cmd : os.commands) { |
841 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
842 | if (!isd) |
843 | continue; |
844 | if (ret.empty()) { |
845 | ret = isd->sections; |
846 | } else { |
847 | if (storage.empty()) |
848 | storage.assign(in_start: ret.begin(), in_end: ret.end()); |
849 | storage.insert(I: storage.end(), From: isd->sections.begin(), To: isd->sections.end()); |
850 | } |
851 | } |
852 | return storage.empty() ? ret : ArrayRef(storage); |
853 | } |
854 | |
855 | // Sorts input sections by section name suffixes, so that .foo.N comes |
856 | // before .foo.M if N < M. Used to sort .{init,fini}_array.N sections. |
857 | // We want to keep the original order if the priorities are the same |
858 | // because the compiler keeps the original initialization order in a |
859 | // translation unit and we need to respect that. |
860 | // For more detail, read the section of the GCC's manual about init_priority. |
861 | void OutputSection::sortInitFini() { |
862 | // Sort sections by priority. |
863 | sort(order: [](InputSectionBase *s) { return getPriority(s: s->name); }); |
864 | } |
865 | |
866 | std::array<uint8_t, 4> OutputSection::getFiller() { |
867 | if (filler) |
868 | return *filler; |
869 | if (flags & SHF_EXECINSTR) |
870 | return target->trapInstr; |
871 | return {0, 0, 0, 0}; |
872 | } |
873 | |
874 | void OutputSection::checkDynRelAddends(const uint8_t *bufStart) { |
875 | assert(config->writeAddends && config->checkDynamicRelocs); |
876 | assert(isStaticRelSecType(type)); |
877 | SmallVector<InputSection *, 0> storage; |
878 | ArrayRef<InputSection *> sections = getInputSections(os: *this, storage); |
879 | parallelFor(Begin: 0, End: sections.size(), Fn: [&](size_t i) { |
880 | // When linking with -r or --emit-relocs we might also call this function |
881 | // for input .rel[a].<sec> sections which we simply pass through to the |
882 | // output. We skip over those and only look at the synthetic relocation |
883 | // sections created during linking. |
884 | const auto *sec = dyn_cast<RelocationBaseSection>(Val: sections[i]); |
885 | if (!sec) |
886 | return; |
887 | for (const DynamicReloc &rel : sec->relocs) { |
888 | int64_t addend = rel.addend; |
889 | const OutputSection *relOsec = rel.inputSec->getOutputSection(); |
890 | assert(relOsec != nullptr && "missing output section for relocation" ); |
891 | // Some targets have NOBITS synthetic sections with dynamic relocations |
892 | // with non-zero addends. Skip such sections. |
893 | if (is_contained(Set: {EM_PPC, EM_PPC64}, Element: config->emachine) && |
894 | (rel.inputSec == in.ppc64LongBranchTarget.get() || |
895 | rel.inputSec == in.igotPlt.get())) |
896 | continue; |
897 | const uint8_t *relocTarget = |
898 | bufStart + relOsec->offset + rel.inputSec->getOffset(offset: rel.offsetInSec); |
899 | // For SHT_NOBITS the written addend is always zero. |
900 | int64_t writtenAddend = |
901 | relOsec->type == SHT_NOBITS |
902 | ? 0 |
903 | : target->getImplicitAddend(buf: relocTarget, type: rel.type); |
904 | if (addend != writtenAddend) |
905 | internalLinkerError( |
906 | loc: getErrorLocation(loc: relocTarget), |
907 | msg: "wrote incorrect addend value 0x" + utohexstr(X: writtenAddend) + |
908 | " instead of 0x" + utohexstr(X: addend) + |
909 | " for dynamic relocation " + toString(type: rel.type) + |
910 | " at offset 0x" + utohexstr(X: rel.getOffset()) + |
911 | (rel.sym ? " against symbol " + toString(*rel.sym) : "" )); |
912 | } |
913 | }); |
914 | } |
915 | |
916 | template void OutputSection::writeHeaderTo<ELF32LE>(ELF32LE::Shdr *Shdr); |
917 | template void OutputSection::writeHeaderTo<ELF32BE>(ELF32BE::Shdr *Shdr); |
918 | template void OutputSection::writeHeaderTo<ELF64LE>(ELF64LE::Shdr *Shdr); |
919 | template void OutputSection::writeHeaderTo<ELF64BE>(ELF64BE::Shdr *Shdr); |
920 | |
921 | template void OutputSection::writeTo<ELF32LE>(uint8_t *, |
922 | llvm::parallel::TaskGroup &); |
923 | template void OutputSection::writeTo<ELF32BE>(uint8_t *, |
924 | llvm::parallel::TaskGroup &); |
925 | template void OutputSection::writeTo<ELF64LE>(uint8_t *, |
926 | llvm::parallel::TaskGroup &); |
927 | template void OutputSection::writeTo<ELF64BE>(uint8_t *, |
928 | llvm::parallel::TaskGroup &); |
929 | |
930 | template void OutputSection::maybeCompress<ELF32LE>(); |
931 | template void OutputSection::maybeCompress<ELF32BE>(); |
932 | template void OutputSection::maybeCompress<ELF64LE>(); |
933 | template void OutputSection::maybeCompress<ELF64BE>(); |
934 | |