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/Parallel.h" |
22 | #include "llvm/Support/Path.h" |
23 | #include "llvm/Support/TimeProfiler.h" |
24 | #if LLVM_ENABLE_ZLIB |
25 | // Avoid introducing max as a macro from Windows headers. |
26 | #define NOMINMAX |
27 | #include <zlib.h> |
28 | #endif |
29 | #if LLVM_ENABLE_ZSTD |
30 | #include <zstd.h> |
31 | #endif |
32 | |
33 | using namespace llvm; |
34 | using namespace llvm::dwarf; |
35 | using namespace llvm::object; |
36 | using namespace llvm::support::endian; |
37 | using namespace llvm::ELF; |
38 | using namespace lld; |
39 | using namespace lld::elf; |
40 | |
41 | uint8_t *Out::bufferStart; |
42 | PhdrEntry *Out::tlsPhdr; |
43 | OutputSection *Out::; |
44 | OutputSection *Out::; |
45 | OutputSection *Out::preinitArray; |
46 | OutputSection *Out::initArray; |
47 | OutputSection *Out::finiArray; |
48 | |
49 | SmallVector<OutputSection *, 0> elf::outputSections; |
50 | |
51 | uint32_t OutputSection::getPhdrFlags() const { |
52 | uint32_t ret = 0; |
53 | if (config->emachine != EM_ARM || !(flags & SHF_ARM_PURECODE)) |
54 | ret |= PF_R; |
55 | if (flags & SHF_WRITE) |
56 | ret |= PF_W; |
57 | if (flags & SHF_EXECINSTR) |
58 | ret |= PF_X; |
59 | return ret; |
60 | } |
61 | |
62 | template <class ELFT> |
63 | void OutputSection::(typename ELFT::Shdr *shdr) { |
64 | shdr->sh_entsize = entsize; |
65 | shdr->sh_addralign = addralign; |
66 | shdr->sh_type = type; |
67 | shdr->sh_offset = offset; |
68 | shdr->sh_flags = flags; |
69 | shdr->sh_info = info; |
70 | shdr->sh_link = link; |
71 | shdr->sh_addr = addr; |
72 | shdr->sh_size = size; |
73 | shdr->sh_name = shName; |
74 | } |
75 | |
76 | OutputSection::OutputSection(StringRef name, uint32_t type, uint64_t flags) |
77 | : SectionBase(Output, name, flags, /*Entsize*/ 0, /*Alignment*/ 1, type, |
78 | /*Info*/ 0, /*Link*/ 0) {} |
79 | |
80 | // We allow sections of types listed below to merged into a |
81 | // single progbits section. This is typically done by linker |
82 | // scripts. Merging nobits and progbits will force disk space |
83 | // to be allocated for nobits sections. Other ones don't require |
84 | // any special treatment on top of progbits, so there doesn't |
85 | // seem to be a harm in merging them. |
86 | // |
87 | // NOTE: clang since rL252300 emits SHT_X86_64_UNWIND .eh_frame sections. Allow |
88 | // them to be merged into SHT_PROGBITS .eh_frame (GNU as .cfi_*). |
89 | static bool canMergeToProgbits(unsigned type) { |
90 | return type == SHT_NOBITS || type == SHT_PROGBITS || type == SHT_INIT_ARRAY || |
91 | type == SHT_PREINIT_ARRAY || type == SHT_FINI_ARRAY || |
92 | type == SHT_NOTE || |
93 | (type == SHT_X86_64_UNWIND && config->emachine == EM_X86_64); |
94 | } |
95 | |
96 | // Record that isec will be placed in the OutputSection. isec does not become |
97 | // permanent until finalizeInputSections() is called. The function should not be |
98 | // used after finalizeInputSections() is called. If you need to add an |
99 | // InputSection post finalizeInputSections(), then you must do the following: |
100 | // |
101 | // 1. Find or create an InputSectionDescription to hold InputSection. |
102 | // 2. Add the InputSection to the InputSectionDescription::sections. |
103 | // 3. Call commitSection(isec). |
104 | void OutputSection::recordSection(InputSectionBase *isec) { |
105 | partition = isec->partition; |
106 | isec->parent = this; |
107 | if (commands.empty() || !isa<InputSectionDescription>(Val: commands.back())) |
108 | commands.push_back(Elt: make<InputSectionDescription>(args: "" )); |
109 | auto *isd = cast<InputSectionDescription>(Val: commands.back()); |
110 | isd->sectionBases.push_back(Elt: isec); |
111 | } |
112 | |
113 | // Update fields (type, flags, alignment, etc) according to the InputSection |
114 | // isec. Also check whether the InputSection flags and type are consistent with |
115 | // other InputSections. |
116 | void OutputSection::commitSection(InputSection *isec) { |
117 | if (LLVM_UNLIKELY(type != isec->type)) { |
118 | if (hasInputSections || typeIsSet) { |
119 | if (typeIsSet || !canMergeToProgbits(type) || |
120 | !canMergeToProgbits(type: isec->type)) { |
121 | // The (NOLOAD) changes the section type to SHT_NOBITS, the intention is |
122 | // that the contents at that address is provided by some other means. |
123 | // Some projects (e.g. |
124 | // https://github.com/ClangBuiltLinux/linux/issues/1597) rely on the |
125 | // behavior. Other types get an error. |
126 | if (type != SHT_NOBITS) { |
127 | errorOrWarn(msg: "section type mismatch for " + isec->name + "\n>>> " + |
128 | toString(isec) + ": " + |
129 | getELFSectionTypeName(Machine: config->emachine, Type: isec->type) + |
130 | "\n>>> output section " + name + ": " + |
131 | getELFSectionTypeName(Machine: config->emachine, Type: type)); |
132 | } |
133 | } |
134 | if (!typeIsSet) |
135 | type = SHT_PROGBITS; |
136 | } else { |
137 | type = isec->type; |
138 | } |
139 | } |
140 | if (!hasInputSections) { |
141 | // If IS is the first section to be added to this section, |
142 | // initialize type, entsize and flags from isec. |
143 | hasInputSections = true; |
144 | entsize = isec->entsize; |
145 | flags = isec->flags; |
146 | } else { |
147 | // Otherwise, check if new type or flags are compatible with existing ones. |
148 | if ((flags ^ isec->flags) & SHF_TLS) |
149 | error(msg: "incompatible section flags for " + name + "\n>>> " + |
150 | toString(isec) + ": 0x" + utohexstr(X: isec->flags) + |
151 | "\n>>> output section " + name + ": 0x" + utohexstr(X: flags)); |
152 | } |
153 | |
154 | isec->parent = this; |
155 | uint64_t andMask = |
156 | config->emachine == EM_ARM ? (uint64_t)SHF_ARM_PURECODE : 0; |
157 | uint64_t orMask = ~andMask; |
158 | uint64_t andFlags = (flags & isec->flags) & andMask; |
159 | uint64_t orFlags = (flags | isec->flags) & orMask; |
160 | flags = andFlags | orFlags; |
161 | if (nonAlloc) |
162 | flags &= ~(uint64_t)SHF_ALLOC; |
163 | |
164 | addralign = std::max(a: addralign, b: isec->addralign); |
165 | |
166 | // If this section contains a table of fixed-size entries, sh_entsize |
167 | // holds the element size. If it contains elements of different size we |
168 | // set sh_entsize to 0. |
169 | if (entsize != isec->entsize) |
170 | entsize = 0; |
171 | } |
172 | |
173 | static MergeSyntheticSection *createMergeSynthetic(StringRef name, |
174 | uint32_t type, |
175 | uint64_t flags, |
176 | uint32_t addralign) { |
177 | if ((flags & SHF_STRINGS) && config->optimize >= 2) |
178 | return make<MergeTailSection>(args&: name, args&: type, args&: flags, args&: addralign); |
179 | return make<MergeNoTailSection>(args&: name, args&: type, args&: flags, args&: addralign); |
180 | } |
181 | |
182 | // This function scans over the InputSectionBase list sectionBases to create |
183 | // InputSectionDescription::sections. |
184 | // |
185 | // It removes MergeInputSections from the input section array and adds |
186 | // new synthetic sections at the location of the first input section |
187 | // that it replaces. It then finalizes each synthetic section in order |
188 | // to compute an output offset for each piece of each input section. |
189 | void OutputSection::finalizeInputSections() { |
190 | std::vector<MergeSyntheticSection *> mergeSections; |
191 | for (SectionCommand *cmd : commands) { |
192 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
193 | if (!isd) |
194 | continue; |
195 | isd->sections.reserve(N: isd->sectionBases.size()); |
196 | for (InputSectionBase *s : isd->sectionBases) { |
197 | MergeInputSection *ms = dyn_cast<MergeInputSection>(Val: s); |
198 | if (!ms) { |
199 | isd->sections.push_back(Elt: cast<InputSection>(Val: s)); |
200 | continue; |
201 | } |
202 | |
203 | // We do not want to handle sections that are not alive, so just remove |
204 | // them instead of trying to merge. |
205 | if (!ms->isLive()) |
206 | continue; |
207 | |
208 | auto i = llvm::find_if(Range&: mergeSections, P: [=](MergeSyntheticSection *sec) { |
209 | // While we could create a single synthetic section for two different |
210 | // values of Entsize, it is better to take Entsize into consideration. |
211 | // |
212 | // With a single synthetic section no two pieces with different Entsize |
213 | // could be equal, so we may as well have two sections. |
214 | // |
215 | // Using Entsize in here also allows us to propagate it to the synthetic |
216 | // section. |
217 | // |
218 | // SHF_STRINGS section with different alignments should not be merged. |
219 | return sec->flags == ms->flags && sec->entsize == ms->entsize && |
220 | (sec->addralign == ms->addralign || !(sec->flags & SHF_STRINGS)); |
221 | }); |
222 | if (i == mergeSections.end()) { |
223 | MergeSyntheticSection *syn = |
224 | createMergeSynthetic(name: s->name, type: ms->type, flags: ms->flags, addralign: ms->addralign); |
225 | mergeSections.push_back(x: syn); |
226 | i = std::prev(x: mergeSections.end()); |
227 | syn->entsize = ms->entsize; |
228 | isd->sections.push_back(Elt: syn); |
229 | } |
230 | (*i)->addSection(ms); |
231 | } |
232 | |
233 | // sectionBases should not be used from this point onwards. Clear it to |
234 | // catch misuses. |
235 | isd->sectionBases.clear(); |
236 | |
237 | // Some input sections may be removed from the list after ICF. |
238 | for (InputSection *s : isd->sections) |
239 | commitSection(isec: s); |
240 | } |
241 | for (auto *ms : mergeSections) |
242 | ms->finalizeContents(); |
243 | } |
244 | |
245 | static void sortByOrder(MutableArrayRef<InputSection *> in, |
246 | llvm::function_ref<int(InputSectionBase *s)> order) { |
247 | std::vector<std::pair<int, InputSection *>> v; |
248 | for (InputSection *s : in) |
249 | v.emplace_back(args: order(s), args&: s); |
250 | llvm::stable_sort(Range&: v, C: less_first()); |
251 | |
252 | for (size_t i = 0; i < v.size(); ++i) |
253 | in[i] = v[i].second; |
254 | } |
255 | |
256 | uint64_t elf::() { |
257 | if (config->oFormatBinary) |
258 | return 0; |
259 | return Out::elfHeader->size + Out::programHeaders->size; |
260 | } |
261 | |
262 | void OutputSection::sort(llvm::function_ref<int(InputSectionBase *s)> order) { |
263 | assert(isLive()); |
264 | for (SectionCommand *b : commands) |
265 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: b)) |
266 | sortByOrder(in: isd->sections, order); |
267 | } |
268 | |
269 | static void nopInstrFill(uint8_t *buf, size_t size) { |
270 | if (size == 0) |
271 | return; |
272 | unsigned i = 0; |
273 | if (size == 0) |
274 | return; |
275 | std::vector<std::vector<uint8_t>> nopFiller = *target->nopInstrs; |
276 | unsigned num = size / nopFiller.back().size(); |
277 | for (unsigned c = 0; c < num; ++c) { |
278 | memcpy(dest: buf + i, src: nopFiller.back().data(), n: nopFiller.back().size()); |
279 | i += nopFiller.back().size(); |
280 | } |
281 | unsigned remaining = size - i; |
282 | if (!remaining) |
283 | return; |
284 | assert(nopFiller[remaining - 1].size() == remaining); |
285 | memcpy(dest: buf + i, src: nopFiller[remaining - 1].data(), n: remaining); |
286 | } |
287 | |
288 | // Fill [Buf, Buf + Size) with Filler. |
289 | // This is used for linker script "=fillexp" command. |
290 | static void fill(uint8_t *buf, size_t size, |
291 | const std::array<uint8_t, 4> &filler) { |
292 | size_t i = 0; |
293 | for (; i + 4 < size; i += 4) |
294 | memcpy(dest: buf + i, src: filler.data(), n: 4); |
295 | memcpy(dest: buf + i, src: filler.data(), n: size - i); |
296 | } |
297 | |
298 | #if LLVM_ENABLE_ZLIB |
299 | static SmallVector<uint8_t, 0> deflateShard(ArrayRef<uint8_t> in, int level, |
300 | int flush) { |
301 | // 15 and 8 are default. windowBits=-15 is negative to generate raw deflate |
302 | // data with no zlib header or trailer. |
303 | z_stream s = {}; |
304 | deflateInit2(&s, level, Z_DEFLATED, -15, 8, Z_DEFAULT_STRATEGY); |
305 | s.next_in = const_cast<uint8_t *>(in.data()); |
306 | s.avail_in = in.size(); |
307 | |
308 | // Allocate a buffer of half of the input size, and grow it by 1.5x if |
309 | // insufficient. |
310 | SmallVector<uint8_t, 0> out; |
311 | size_t pos = 0; |
312 | out.resize_for_overwrite(N: std::max<size_t>(a: in.size() / 2, b: 64)); |
313 | do { |
314 | if (pos == out.size()) |
315 | out.resize_for_overwrite(N: out.size() * 3 / 2); |
316 | s.next_out = out.data() + pos; |
317 | s.avail_out = out.size() - pos; |
318 | (void)deflate(strm: &s, flush); |
319 | pos = s.next_out - out.data(); |
320 | } while (s.avail_out == 0); |
321 | assert(s.avail_in == 0); |
322 | |
323 | out.truncate(N: pos); |
324 | deflateEnd(strm: &s); |
325 | return out; |
326 | } |
327 | #endif |
328 | |
329 | // Compress certain non-SHF_ALLOC sections: |
330 | // |
331 | // * (if --compress-debug-sections is specified) non-empty .debug_* sections |
332 | // * (if --compress-sections is specified) matched sections |
333 | template <class ELFT> void OutputSection::maybeCompress() { |
334 | using Elf_Chdr = typename ELFT::Chdr; |
335 | (void)sizeof(Elf_Chdr); |
336 | |
337 | DebugCompressionType ctype = DebugCompressionType::None; |
338 | for (auto &[glob, t] : config->compressSections) |
339 | if (glob.match(S: name)) |
340 | ctype = t; |
341 | if (!(flags & SHF_ALLOC) && config->compressDebugSections && |
342 | name.starts_with(Prefix: ".debug_" ) && size) |
343 | ctype = *config->compressDebugSections; |
344 | if (ctype == DebugCompressionType::None) |
345 | return; |
346 | if (flags & SHF_ALLOC) { |
347 | errorOrWarn(msg: "--compress-sections: section '" + name + |
348 | "' with the SHF_ALLOC flag cannot be compressed" ); |
349 | return; |
350 | } |
351 | |
352 | llvm::TimeTraceScope timeScope("Compress sections" ); |
353 | compressed.uncompressedSize = size; |
354 | auto buf = std::make_unique<uint8_t[]>(num: size); |
355 | // Write uncompressed data to a temporary zero-initialized buffer. |
356 | { |
357 | parallel::TaskGroup tg; |
358 | writeTo<ELFT>(buf.get(), tg); |
359 | } |
360 | // The generic ABI specifies "The sh_size and sh_addralign fields of the |
361 | // section header for a compressed section reflect the requirements of the |
362 | // compressed section." However, 1-byte alignment has been wildly accepted |
363 | // and utilized for a long time. Removing alignment padding is particularly |
364 | // useful when there are many compressed output sections. |
365 | addralign = 1; |
366 | |
367 | #if LLVM_ENABLE_ZSTD |
368 | // Use ZSTD's streaming compression API which permits parallel workers working |
369 | // on the stream. See http://facebook.github.io/zstd/zstd_manual.html |
370 | // "Streaming compression - HowTo". |
371 | if (ctype == DebugCompressionType::Zstd) { |
372 | // Allocate a buffer of half of the input size, and grow it by 1.5x if |
373 | // insufficient. |
374 | compressed.type = ELFCOMPRESS_ZSTD; |
375 | compressed.shards = std::make_unique<SmallVector<uint8_t, 0>[]>(num: 1); |
376 | SmallVector<uint8_t, 0> &out = compressed.shards[0]; |
377 | out.resize_for_overwrite(N: std::max<size_t>(a: size / 2, b: 32)); |
378 | size_t pos = 0; |
379 | |
380 | ZSTD_CCtx *cctx = ZSTD_createCCtx(); |
381 | // Ignore error if zstd was not built with ZSTD_MULTITHREAD. |
382 | (void)ZSTD_CCtx_setParameter(cctx, param: ZSTD_c_nbWorkers, |
383 | value: parallel::strategy.compute_thread_count()); |
384 | ZSTD_outBuffer zob = {.dst: out.data(), .size: out.size(), .pos: 0}; |
385 | ZSTD_EndDirective directive = ZSTD_e_continue; |
386 | const size_t blockSize = ZSTD_CStreamInSize(); |
387 | do { |
388 | const size_t n = std::min(a: static_cast<size_t>(size - pos), b: blockSize); |
389 | if (n == size - pos) |
390 | directive = ZSTD_e_end; |
391 | ZSTD_inBuffer zib = {.src: buf.get() + pos, .size: n, .pos: 0}; |
392 | size_t bytesRemaining = 0; |
393 | while (zib.pos != zib.size || |
394 | (directive == ZSTD_e_end && bytesRemaining != 0)) { |
395 | if (zob.pos == zob.size) { |
396 | out.resize_for_overwrite(N: out.size() * 3 / 2); |
397 | zob.dst = out.data(); |
398 | zob.size = out.size(); |
399 | } |
400 | bytesRemaining = ZSTD_compressStream2(cctx, output: &zob, input: &zib, endOp: directive); |
401 | assert(!ZSTD_isError(bytesRemaining)); |
402 | } |
403 | pos += n; |
404 | } while (directive != ZSTD_e_end); |
405 | out.resize(N: zob.pos); |
406 | ZSTD_freeCCtx(cctx); |
407 | |
408 | size = sizeof(Elf_Chdr) + out.size(); |
409 | flags |= SHF_COMPRESSED; |
410 | return; |
411 | } |
412 | #endif |
413 | |
414 | #if LLVM_ENABLE_ZLIB |
415 | // We chose 1 (Z_BEST_SPEED) as the default compression level because it is |
416 | // the fastest. If -O2 is given, we use level 6 to compress debug info more by |
417 | // ~15%. We found that level 7 to 9 doesn't make much difference (~1% more |
418 | // compression) while they take significant amount of time (~2x), so level 6 |
419 | // seems enough. |
420 | const int level = config->optimize >= 2 ? 6 : Z_BEST_SPEED; |
421 | |
422 | // Split input into 1-MiB shards. |
423 | constexpr size_t shardSize = 1 << 20; |
424 | auto shardsIn = split(arr: ArrayRef<uint8_t>(buf.get(), size), chunkSize: shardSize); |
425 | const size_t numShards = shardsIn.size(); |
426 | |
427 | // Compress shards and compute Alder-32 checksums. Use Z_SYNC_FLUSH for all |
428 | // shards but the last to flush the output to a byte boundary to be |
429 | // concatenated with the next shard. |
430 | auto shardsOut = std::make_unique<SmallVector<uint8_t, 0>[]>(num: numShards); |
431 | auto shardsAdler = std::make_unique<uint32_t[]>(num: numShards); |
432 | parallelFor(0, numShards, [&](size_t i) { |
433 | shardsOut[i] = deflateShard(in: shardsIn[i], level, |
434 | flush: i != numShards - 1 ? Z_SYNC_FLUSH : Z_FINISH); |
435 | shardsAdler[i] = adler32(adler: 1, buf: shardsIn[i].data(), len: shardsIn[i].size()); |
436 | }); |
437 | |
438 | // Update section size and combine Alder-32 checksums. |
439 | uint32_t checksum = 1; // Initial Adler-32 value |
440 | size = sizeof(Elf_Chdr) + 2; // Elf_Chdir and zlib header |
441 | for (size_t i = 0; i != numShards; ++i) { |
442 | size += shardsOut[i].size(); |
443 | checksum = adler32_combine(checksum, shardsAdler[i], shardsIn[i].size()); |
444 | } |
445 | size += 4; // checksum |
446 | |
447 | compressed.type = ELFCOMPRESS_ZLIB; |
448 | compressed.shards = std::move(shardsOut); |
449 | compressed.numShards = numShards; |
450 | compressed.checksum = checksum; |
451 | flags |= SHF_COMPRESSED; |
452 | #endif |
453 | } |
454 | |
455 | static void writeInt(uint8_t *buf, uint64_t data, uint64_t size) { |
456 | if (size == 1) |
457 | *buf = data; |
458 | else if (size == 2) |
459 | write16(p: buf, v: data); |
460 | else if (size == 4) |
461 | write32(p: buf, v: data); |
462 | else if (size == 8) |
463 | write64(p: buf, v: data); |
464 | else |
465 | llvm_unreachable("unsupported Size argument" ); |
466 | } |
467 | |
468 | template <class ELFT> |
469 | void OutputSection::writeTo(uint8_t *buf, parallel::TaskGroup &tg) { |
470 | llvm::TimeTraceScope timeScope("Write sections" , name); |
471 | if (type == SHT_NOBITS) |
472 | return; |
473 | |
474 | // If the section is compressed due to |
475 | // --compress-debug-section/--compress-sections, the content is already known. |
476 | if (compressed.shards) { |
477 | auto *chdr = reinterpret_cast<typename ELFT::Chdr *>(buf); |
478 | chdr->ch_type = compressed.type; |
479 | chdr->ch_size = compressed.uncompressedSize; |
480 | chdr->ch_addralign = addralign; |
481 | buf += sizeof(*chdr); |
482 | if (compressed.type == ELFCOMPRESS_ZSTD) { |
483 | memcpy(dest: buf, src: compressed.shards[0].data(), n: compressed.shards[0].size()); |
484 | return; |
485 | } |
486 | |
487 | // Compute shard offsets. |
488 | auto offsets = std::make_unique<size_t[]>(num: compressed.numShards); |
489 | offsets[0] = 2; // zlib header |
490 | for (size_t i = 1; i != compressed.numShards; ++i) |
491 | offsets[i] = offsets[i - 1] + compressed.shards[i - 1].size(); |
492 | |
493 | buf[0] = 0x78; // CMF |
494 | buf[1] = 0x01; // FLG: best speed |
495 | parallelFor(0, compressed.numShards, [&](size_t i) { |
496 | memcpy(dest: buf + offsets[i], src: compressed.shards[i].data(), |
497 | n: compressed.shards[i].size()); |
498 | }); |
499 | |
500 | write32be(P: buf + (size - sizeof(*chdr) - 4), V: compressed.checksum); |
501 | return; |
502 | } |
503 | |
504 | // Write leading padding. |
505 | ArrayRef<InputSection *> sections = getInputSections(os: *this, storage); |
506 | std::array<uint8_t, 4> filler = getFiller(); |
507 | bool nonZeroFiller = read32(p: filler.data()) != 0; |
508 | if (nonZeroFiller) |
509 | fill(buf, size: sections.empty() ? size : sections[0]->outSecOff, filler); |
510 | |
511 | auto fn = [=](size_t begin, size_t end) { |
512 | size_t numSections = sections.size(); |
513 | for (size_t i = begin; i != end; ++i) { |
514 | InputSection *isec = sections[i]; |
515 | if (auto *s = dyn_cast<SyntheticSection>(Val: isec)) |
516 | s->writeTo(buf: buf + isec->outSecOff); |
517 | else |
518 | isec->writeTo<ELFT>(buf + isec->outSecOff); |
519 | |
520 | // When in Arm BE8 mode, the linker has to convert the big-endian |
521 | // instructions to little-endian, leaving the data big-endian. |
522 | if (config->emachine == EM_ARM && !config->isLE && config->armBe8 && |
523 | (flags & SHF_EXECINSTR)) |
524 | convertArmInstructionstoBE8(sec: isec, buf: buf + isec->outSecOff); |
525 | |
526 | // Fill gaps between sections. |
527 | if (nonZeroFiller) { |
528 | uint8_t *start = buf + isec->outSecOff + isec->getSize(); |
529 | uint8_t *end; |
530 | if (i + 1 == numSections) |
531 | end = buf + size; |
532 | else |
533 | end = buf + sections[i + 1]->outSecOff; |
534 | if (isec->nopFiller) { |
535 | assert(target->nopInstrs); |
536 | nopInstrFill(buf: start, size: end - start); |
537 | } else |
538 | fill(buf: start, size: end - start, filler); |
539 | } |
540 | } |
541 | }; |
542 | |
543 | // If there is any BYTE()-family command (rare), write the section content |
544 | // first then process BYTE to overwrite the filler content. The write is |
545 | // serial due to the limitation of llvm/Support/Parallel.h. |
546 | bool written = false; |
547 | size_t numSections = sections.size(); |
548 | for (SectionCommand *cmd : commands) |
549 | if (auto *data = dyn_cast<ByteCommand>(Val: cmd)) { |
550 | if (!std::exchange(obj&: written, new_val: true)) |
551 | fn(0, numSections); |
552 | writeInt(buf: buf + data->offset, data: data->expression().getValue(), size: data->size); |
553 | } |
554 | if (written || !numSections) |
555 | return; |
556 | |
557 | // There is no data command. Write content asynchronously to overlap the write |
558 | // time with other output sections. Note, if a linker script specifies |
559 | // overlapping output sections (needs --noinhibit-exec or --no-check-sections |
560 | // to supress the error), the output may be non-deterministic. |
561 | const size_t taskSizeLimit = 4 << 20; |
562 | for (size_t begin = 0, i = 0, taskSize = 0;;) { |
563 | taskSize += sections[i]->getSize(); |
564 | bool done = ++i == numSections; |
565 | if (done || taskSize >= taskSizeLimit) { |
566 | tg.spawn(f: [=] { fn(begin, i); }); |
567 | if (done) |
568 | break; |
569 | begin = i; |
570 | taskSize = 0; |
571 | } |
572 | } |
573 | } |
574 | |
575 | static void finalizeShtGroup(OutputSection *os, InputSection *section) { |
576 | // sh_link field for SHT_GROUP sections should contain the section index of |
577 | // the symbol table. |
578 | os->link = in.symTab->getParent()->sectionIndex; |
579 | |
580 | if (!section) |
581 | return; |
582 | |
583 | // sh_info then contain index of an entry in symbol table section which |
584 | // provides signature of the section group. |
585 | ArrayRef<Symbol *> symbols = section->file->getSymbols(); |
586 | os->info = in.symTab->getSymbolIndex(sym: *symbols[section->info]); |
587 | |
588 | // Some group members may be combined or discarded, so we need to compute the |
589 | // new size. The content will be rewritten in InputSection::copyShtGroup. |
590 | DenseSet<uint32_t> seen; |
591 | ArrayRef<InputSectionBase *> sections = section->file->getSections(); |
592 | for (const uint32_t &idx : section->getDataAs<uint32_t>().slice(N: 1)) |
593 | if (OutputSection *osec = sections[read32(p: &idx)]->getOutputSection()) |
594 | seen.insert(V: osec->sectionIndex); |
595 | os->size = (1 + seen.size()) * sizeof(uint32_t); |
596 | } |
597 | |
598 | void OutputSection::finalize() { |
599 | InputSection *first = getFirstInputSection(os: this); |
600 | |
601 | if (flags & SHF_LINK_ORDER) { |
602 | // We must preserve the link order dependency of sections with the |
603 | // SHF_LINK_ORDER flag. The dependency is indicated by the sh_link field. We |
604 | // need to translate the InputSection sh_link to the OutputSection sh_link, |
605 | // all InputSections in the OutputSection have the same dependency. |
606 | if (auto *ex = dyn_cast<ARMExidxSyntheticSection>(Val: first)) |
607 | link = ex->getLinkOrderDep()->getParent()->sectionIndex; |
608 | else if (first->flags & SHF_LINK_ORDER) |
609 | if (auto *d = first->getLinkOrderDep()) |
610 | link = d->getParent()->sectionIndex; |
611 | } |
612 | |
613 | if (type == SHT_GROUP) { |
614 | finalizeShtGroup(os: this, section: first); |
615 | return; |
616 | } |
617 | |
618 | if (!config->copyRelocs || !isStaticRelSecType(type)) |
619 | return; |
620 | |
621 | // Skip if 'first' is synthetic, i.e. not a section created by --emit-relocs. |
622 | // Normally 'type' was changed by 'first' so 'first' should be non-null. |
623 | // However, if the output section is .rela.dyn, 'type' can be set by the empty |
624 | // synthetic .rela.plt and first can be null. |
625 | if (!first || isa<SyntheticSection>(Val: first)) |
626 | return; |
627 | |
628 | link = in.symTab->getParent()->sectionIndex; |
629 | // sh_info for SHT_REL[A] sections should contain the section header index of |
630 | // the section to which the relocation applies. |
631 | InputSectionBase *s = first->getRelocatedSection(); |
632 | info = s->getOutputSection()->sectionIndex; |
633 | flags |= SHF_INFO_LINK; |
634 | } |
635 | |
636 | // Returns true if S is in one of the many forms the compiler driver may pass |
637 | // crtbegin files. |
638 | // |
639 | // Gcc uses any of crtbegin[<empty>|S|T].o. |
640 | // Clang uses Gcc's plus clang_rt.crtbegin[-<arch>|<empty>].o. |
641 | |
642 | static bool isCrt(StringRef s, StringRef beginEnd) { |
643 | s = sys::path::filename(path: s); |
644 | if (!s.consume_back(Suffix: ".o" )) |
645 | return false; |
646 | if (s.consume_front(Prefix: "clang_rt." )) |
647 | return s.consume_front(Prefix: beginEnd); |
648 | return s.consume_front(Prefix: beginEnd) && s.size() <= 1; |
649 | } |
650 | |
651 | // .ctors and .dtors are sorted by this order: |
652 | // |
653 | // 1. .ctors/.dtors in crtbegin (which contains a sentinel value -1). |
654 | // 2. The section is named ".ctors" or ".dtors" (priority: 65536). |
655 | // 3. The section has an optional priority value in the form of ".ctors.N" or |
656 | // ".dtors.N" where N is a number in the form of %05u (priority: 65535-N). |
657 | // 4. .ctors/.dtors in crtend (which contains a sentinel value 0). |
658 | // |
659 | // For 2 and 3, the sections are sorted by priority from high to low, e.g. |
660 | // .ctors (65536), .ctors.00100 (65436), .ctors.00200 (65336). In GNU ld's |
661 | // internal linker scripts, the sorting is by string comparison which can |
662 | // achieve the same goal given the optional priority values are of the same |
663 | // length. |
664 | // |
665 | // In an ideal world, we don't need this function because .init_array and |
666 | // .ctors are duplicate features (and .init_array is newer.) However, there |
667 | // are too many real-world use cases of .ctors, so we had no choice to |
668 | // support that with this rather ad-hoc semantics. |
669 | static bool compCtors(const InputSection *a, const InputSection *b) { |
670 | bool beginA = isCrt(s: a->file->getName(), beginEnd: "crtbegin" ); |
671 | bool beginB = isCrt(s: b->file->getName(), beginEnd: "crtbegin" ); |
672 | if (beginA != beginB) |
673 | return beginA; |
674 | bool endA = isCrt(s: a->file->getName(), beginEnd: "crtend" ); |
675 | bool endB = isCrt(s: b->file->getName(), beginEnd: "crtend" ); |
676 | if (endA != endB) |
677 | return endB; |
678 | return getPriority(s: a->name) > getPriority(s: b->name); |
679 | } |
680 | |
681 | // Sorts input sections by the special rules for .ctors and .dtors. |
682 | // Unfortunately, the rules are different from the one for .{init,fini}_array. |
683 | // Read the comment above. |
684 | void OutputSection::sortCtorsDtors() { |
685 | assert(commands.size() == 1); |
686 | auto *isd = cast<InputSectionDescription>(Val: commands[0]); |
687 | llvm::stable_sort(Range&: isd->sections, C: compCtors); |
688 | } |
689 | |
690 | // If an input string is in the form of "foo.N" where N is a number, return N |
691 | // (65535-N if .ctors.N or .dtors.N). Otherwise, returns 65536, which is one |
692 | // greater than the lowest priority. |
693 | int elf::getPriority(StringRef s) { |
694 | size_t pos = s.rfind(C: '.'); |
695 | if (pos == StringRef::npos) |
696 | return 65536; |
697 | int v = 65536; |
698 | if (to_integer(S: s.substr(Start: pos + 1), Num&: v, Base: 10) && |
699 | (pos == 6 && (s.starts_with(Prefix: ".ctors" ) || s.starts_with(Prefix: ".dtors" )))) |
700 | v = 65535 - v; |
701 | return v; |
702 | } |
703 | |
704 | InputSection *elf::getFirstInputSection(const OutputSection *os) { |
705 | for (SectionCommand *cmd : os->commands) |
706 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) |
707 | if (!isd->sections.empty()) |
708 | return isd->sections[0]; |
709 | return nullptr; |
710 | } |
711 | |
712 | ArrayRef<InputSection *> |
713 | elf::getInputSections(const OutputSection &os, |
714 | SmallVector<InputSection *, 0> &storage) { |
715 | ArrayRef<InputSection *> ret; |
716 | storage.clear(); |
717 | for (SectionCommand *cmd : os.commands) { |
718 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
719 | if (!isd) |
720 | continue; |
721 | if (ret.empty()) { |
722 | ret = isd->sections; |
723 | } else { |
724 | if (storage.empty()) |
725 | storage.assign(in_start: ret.begin(), in_end: ret.end()); |
726 | storage.insert(I: storage.end(), From: isd->sections.begin(), To: isd->sections.end()); |
727 | } |
728 | } |
729 | return storage.empty() ? ret : ArrayRef(storage); |
730 | } |
731 | |
732 | // Sorts input sections by section name suffixes, so that .foo.N comes |
733 | // before .foo.M if N < M. Used to sort .{init,fini}_array.N sections. |
734 | // We want to keep the original order if the priorities are the same |
735 | // because the compiler keeps the original initialization order in a |
736 | // translation unit and we need to respect that. |
737 | // For more detail, read the section of the GCC's manual about init_priority. |
738 | void OutputSection::sortInitFini() { |
739 | // Sort sections by priority. |
740 | sort(order: [](InputSectionBase *s) { return getPriority(s: s->name); }); |
741 | } |
742 | |
743 | std::array<uint8_t, 4> OutputSection::getFiller() { |
744 | if (filler) |
745 | return *filler; |
746 | if (flags & SHF_EXECINSTR) |
747 | return target->trapInstr; |
748 | return {0, 0, 0, 0}; |
749 | } |
750 | |
751 | void OutputSection::checkDynRelAddends(const uint8_t *bufStart) { |
752 | assert(config->writeAddends && config->checkDynamicRelocs); |
753 | assert(isStaticRelSecType(type)); |
754 | SmallVector<InputSection *, 0> storage; |
755 | ArrayRef<InputSection *> sections = getInputSections(os: *this, storage); |
756 | parallelFor(Begin: 0, End: sections.size(), Fn: [&](size_t i) { |
757 | // When linking with -r or --emit-relocs we might also call this function |
758 | // for input .rel[a].<sec> sections which we simply pass through to the |
759 | // output. We skip over those and only look at the synthetic relocation |
760 | // sections created during linking. |
761 | const auto *sec = dyn_cast<RelocationBaseSection>(Val: sections[i]); |
762 | if (!sec) |
763 | return; |
764 | for (const DynamicReloc &rel : sec->relocs) { |
765 | int64_t addend = rel.addend; |
766 | const OutputSection *relOsec = rel.inputSec->getOutputSection(); |
767 | assert(relOsec != nullptr && "missing output section for relocation" ); |
768 | // Some targets have NOBITS synthetic sections with dynamic relocations |
769 | // with non-zero addends. Skip such sections. |
770 | if (is_contained(Set: {EM_PPC, EM_PPC64}, Element: config->emachine) && |
771 | (rel.inputSec == in.ppc64LongBranchTarget.get() || |
772 | rel.inputSec == in.igotPlt.get())) |
773 | continue; |
774 | const uint8_t *relocTarget = |
775 | bufStart + relOsec->offset + rel.inputSec->getOffset(offset: rel.offsetInSec); |
776 | // For SHT_NOBITS the written addend is always zero. |
777 | int64_t writtenAddend = |
778 | relOsec->type == SHT_NOBITS |
779 | ? 0 |
780 | : target->getImplicitAddend(buf: relocTarget, type: rel.type); |
781 | if (addend != writtenAddend) |
782 | internalLinkerError( |
783 | loc: getErrorLocation(loc: relocTarget), |
784 | msg: "wrote incorrect addend value 0x" + utohexstr(X: writtenAddend) + |
785 | " instead of 0x" + utohexstr(X: addend) + |
786 | " for dynamic relocation " + toString(type: rel.type) + |
787 | " at offset 0x" + utohexstr(X: rel.getOffset()) + |
788 | (rel.sym ? " against symbol " + toString(*rel.sym) : "" )); |
789 | } |
790 | }); |
791 | } |
792 | |
793 | template void OutputSection::writeHeaderTo<ELF32LE>(ELF32LE::Shdr *Shdr); |
794 | template void OutputSection::writeHeaderTo<ELF32BE>(ELF32BE::Shdr *Shdr); |
795 | template void OutputSection::writeHeaderTo<ELF64LE>(ELF64LE::Shdr *Shdr); |
796 | template void OutputSection::writeHeaderTo<ELF64BE>(ELF64BE::Shdr *Shdr); |
797 | |
798 | template void OutputSection::writeTo<ELF32LE>(uint8_t *, |
799 | llvm::parallel::TaskGroup &); |
800 | template void OutputSection::writeTo<ELF32BE>(uint8_t *, |
801 | llvm::parallel::TaskGroup &); |
802 | template void OutputSection::writeTo<ELF64LE>(uint8_t *, |
803 | llvm::parallel::TaskGroup &); |
804 | template void OutputSection::writeTo<ELF64BE>(uint8_t *, |
805 | llvm::parallel::TaskGroup &); |
806 | |
807 | template void OutputSection::maybeCompress<ELF32LE>(); |
808 | template void OutputSection::maybeCompress<ELF32BE>(); |
809 | template void OutputSection::maybeCompress<ELF64LE>(); |
810 | template void OutputSection::maybeCompress<ELF64BE>(); |
811 | |