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, LLVM_ENABLE_ZSTD |
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 | #undef in |
26 | #if LLVM_ENABLE_ZLIB |
27 | // Avoid introducing max as a macro from Windows headers. |
28 | #define NOMINMAX |
29 | #include <zlib.h> |
30 | #endif |
31 | #if LLVM_ENABLE_ZSTD |
32 | #include <zstd.h> |
33 | #endif |
34 | |
35 | using namespace llvm; |
36 | using namespace llvm::dwarf; |
37 | using namespace llvm::object; |
38 | using namespace llvm::support::endian; |
39 | using namespace llvm::ELF; |
40 | using namespace lld; |
41 | using namespace lld::elf; |
42 | |
43 | uint32_t OutputSection::getPhdrFlags() const { |
44 | uint32_t ret = 0; |
45 | bool purecode = |
46 | (ctx.arg.emachine == EM_ARM && (flags & SHF_ARM_PURECODE)) || |
47 | (ctx.arg.emachine == EM_AARCH64 && (flags & SHF_AARCH64_PURECODE)); |
48 | if (!purecode) |
49 | ret |= PF_R; |
50 | if (flags & SHF_WRITE) |
51 | ret |= PF_W; |
52 | if (flags & SHF_EXECINSTR) |
53 | ret |= PF_X; |
54 | return ret; |
55 | } |
56 | |
57 | template <class ELFT> |
58 | void OutputSection::writeHeaderTo(typename ELFT::Shdr *shdr) { |
59 | shdr->sh_entsize = entsize; |
60 | shdr->sh_addralign = addralign; |
61 | shdr->sh_type = type; |
62 | shdr->sh_offset = offset; |
63 | shdr->sh_flags = flags; |
64 | shdr->sh_info = info; |
65 | shdr->sh_link = link; |
66 | shdr->sh_addr = addr; |
67 | shdr->sh_size = size; |
68 | shdr->sh_name = shName; |
69 | } |
70 | |
71 | OutputSection::OutputSection(Ctx &ctx, StringRef name, uint32_t type, |
72 | uint64_t flags) |
73 | : SectionBase(Output, ctx.internalFile, name, type, flags, /*link=*/0, |
74 | /*info=*/0, /*addralign=*/1, /*entsize=*/0), |
75 | ctx(ctx) {} |
76 | |
77 | uint64_t OutputSection::getLMA() const { |
78 | return ptLoad ? addr + ptLoad->lmaOffset : addr; |
79 | } |
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(Ctx &ctx, 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 && ctx.arg.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 = ctx.saver.save(S: ".crel"+ name); |
128 | } else if (name.consume_front(Prefix: ".rela")) { |
129 | name = ctx.saver.save(S: ".crel"+ name); |
130 | } |
131 | } else { |
132 | if (typeIsSet || !canMergeToProgbits(ctx, type) || |
133 | !canMergeToProgbits(ctx, 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 | Err(ctx) << "section type mismatch for "<< isec->name << "\n>>> " |
141 | << isec << ": " |
142 | << getELFSectionTypeName(Machine: ctx.arg.emachine, Type: isec->type) |
143 | << "\n>>> output section "<< name << ": " |
144 | << getELFSectionTypeName(Machine: ctx.arg.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 | ErrAlways(ctx) << "incompatible section flags for "<< name << "\n>>> " |
161 | << isec << ": 0x"<< utohexstr(X: isec->flags, LowerCase: true) |
162 | << "\n>>> output section "<< name << ": 0x" |
163 | << utohexstr(X: flags, LowerCase: true); |
164 | } |
165 | |
166 | isec->parent = this; |
167 | uint64_t andMask = 0; |
168 | if (ctx.arg.emachine == EM_ARM) |
169 | andMask |= (uint64_t)SHF_ARM_PURECODE; |
170 | if (ctx.arg.emachine == EM_AARCH64) |
171 | andMask |= (uint64_t)SHF_AARCH64_PURECODE; |
172 | uint64_t orMask = ~andMask; |
173 | uint64_t andFlags = (flags & isec->flags) & andMask; |
174 | uint64_t orFlags = (flags | isec->flags) & orMask; |
175 | flags = andFlags | orFlags; |
176 | if (nonAlloc) |
177 | flags &= ~(uint64_t)SHF_ALLOC; |
178 | |
179 | addralign = std::max(a: addralign, b: isec->addralign); |
180 | |
181 | // If this section contains a table of fixed-size entries, sh_entsize |
182 | // holds the element size. If it contains elements of different size we |
183 | // set sh_entsize to 0. |
184 | if (entsize != isec->entsize) |
185 | entsize = 0; |
186 | } |
187 | |
188 | static MergeSyntheticSection *createMergeSynthetic(Ctx &ctx, StringRef name, |
189 | uint32_t type, |
190 | uint64_t flags, |
191 | uint32_t addralign) { |
192 | if ((flags & SHF_STRINGS) && ctx.arg.optimize >= 2) |
193 | return make<MergeTailSection>(args&: ctx, args&: name, args&: type, args&: flags, args&: addralign); |
194 | return make<MergeNoTailSection>(args&: ctx, args&: name, args&: type, args&: flags, args&: addralign); |
195 | } |
196 | |
197 | // This function scans over the InputSectionBase list sectionBases to create |
198 | // InputSectionDescription::sections. |
199 | // |
200 | // It removes MergeInputSections from the input section array and adds |
201 | // new synthetic sections at the location of the first input section |
202 | // that it replaces. It then finalizes each synthetic section in order |
203 | // to compute an output offset for each piece of each input section. |
204 | void OutputSection::finalizeInputSections() { |
205 | auto *script = ctx.script; |
206 | std::vector<MergeSyntheticSection *> mergeSections; |
207 | for (SectionCommand *cmd : commands) { |
208 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
209 | if (!isd) |
210 | continue; |
211 | isd->sections.reserve(N: isd->sectionBases.size()); |
212 | for (InputSectionBase *s : isd->sectionBases) { |
213 | MergeInputSection *ms = dyn_cast<MergeInputSection>(Val: s); |
214 | if (!ms) { |
215 | isd->sections.push_back(Elt: cast<InputSection>(Val: s)); |
216 | continue; |
217 | } |
218 | |
219 | // We do not want to handle sections that are not alive, so just remove |
220 | // them instead of trying to merge. |
221 | if (!ms->isLive()) |
222 | continue; |
223 | |
224 | auto i = llvm::find_if(Range&: mergeSections, P: [=](MergeSyntheticSection *sec) { |
225 | // While we could create a single synthetic section for two different |
226 | // values of Entsize, it is better to take Entsize into consideration. |
227 | // |
228 | // With a single synthetic section no two pieces with different Entsize |
229 | // could be equal, so we may as well have two sections. |
230 | // |
231 | // Using Entsize in here also allows us to propagate it to the synthetic |
232 | // section. |
233 | // |
234 | // SHF_STRINGS section with different alignments should not be merged. |
235 | return sec->flags == ms->flags && sec->entsize == ms->entsize && |
236 | (sec->addralign == ms->addralign || !(sec->flags & SHF_STRINGS)); |
237 | }); |
238 | if (i == mergeSections.end()) { |
239 | MergeSyntheticSection *syn = createMergeSynthetic( |
240 | ctx, name: s->name, type: ms->type, flags: ms->flags, addralign: ms->addralign); |
241 | mergeSections.push_back(x: syn); |
242 | i = std::prev(x: mergeSections.end()); |
243 | syn->entsize = ms->entsize; |
244 | isd->sections.push_back(Elt: syn); |
245 | // The merge synthetic section inherits the potential spill locations of |
246 | // its first contained section. |
247 | auto it = script->potentialSpillLists.find(Val: ms); |
248 | if (it != script->potentialSpillLists.end()) |
249 | script->potentialSpillLists.try_emplace(Key: syn, Args&: it->second); |
250 | } |
251 | (*i)->addSection(ms); |
252 | } |
253 | |
254 | // sectionBases should not be used from this point onwards. Clear it to |
255 | // catch misuses. |
256 | isd->sectionBases.clear(); |
257 | |
258 | // Some input sections may be removed from the list after ICF. |
259 | for (InputSection *s : isd->sections) |
260 | commitSection(isec: s); |
261 | } |
262 | for (auto *ms : mergeSections) { |
263 | // Merging may have increased the alignment of a spillable section. Update |
264 | // the alignment of potential spill sections and their containing output |
265 | // sections. |
266 | if (auto it = script->potentialSpillLists.find(Val: ms); |
267 | it != script->potentialSpillLists.end()) { |
268 | for (PotentialSpillSection *s = it->second.head; s; s = s->next) { |
269 | s->addralign = std::max(a: s->addralign, b: ms->addralign); |
270 | s->parent->addralign = std::max(a: s->parent->addralign, b: s->addralign); |
271 | } |
272 | } |
273 | |
274 | ms->finalizeContents(); |
275 | } |
276 | } |
277 | |
278 | static void sortByOrder(MutableArrayRef<InputSection *> in, |
279 | llvm::function_ref<int(InputSectionBase *s)> order) { |
280 | std::vector<std::pair<int, InputSection *>> v; |
281 | for (InputSection *s : in) |
282 | v.emplace_back(args: order(s), args&: s); |
283 | llvm::stable_sort(Range&: v, C: less_first()); |
284 | |
285 | for (size_t i = 0; i < v.size(); ++i) |
286 | in[i] = v[i].second; |
287 | } |
288 | |
289 | uint64_t elf::getHeaderSize(Ctx &ctx) { |
290 | if (ctx.arg.oFormatBinary) |
291 | return 0; |
292 | return ctx.out.elfHeader->size + ctx.out.programHeaders->size; |
293 | } |
294 | |
295 | void OutputSection::sort(llvm::function_ref<int(InputSectionBase *s)> order) { |
296 | assert(isLive()); |
297 | for (SectionCommand *b : commands) |
298 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: b)) |
299 | sortByOrder(in: isd->sections, order); |
300 | } |
301 | |
302 | static void nopInstrFill(Ctx &ctx, uint8_t *buf, size_t size) { |
303 | if (size == 0) |
304 | return; |
305 | unsigned i = 0; |
306 | if (size == 0) |
307 | return; |
308 | std::vector<std::vector<uint8_t>> nopFiller = *ctx.target->nopInstrs; |
309 | unsigned num = size / nopFiller.back().size(); |
310 | for (unsigned c = 0; c < num; ++c) { |
311 | memcpy(dest: buf + i, src: nopFiller.back().data(), n: nopFiller.back().size()); |
312 | i += nopFiller.back().size(); |
313 | } |
314 | unsigned remaining = size - i; |
315 | if (!remaining) |
316 | return; |
317 | assert(nopFiller[remaining - 1].size() == remaining); |
318 | memcpy(dest: buf + i, src: nopFiller[remaining - 1].data(), n: remaining); |
319 | } |
320 | |
321 | // Fill [Buf, Buf + Size) with Filler. |
322 | // This is used for linker script "=fillexp" command. |
323 | static void fill(uint8_t *buf, size_t size, |
324 | const std::array<uint8_t, 4> &filler) { |
325 | size_t i = 0; |
326 | for (; i + 4 < size; i += 4) |
327 | memcpy(dest: buf + i, src: filler.data(), n: 4); |
328 | memcpy(dest: buf + i, src: filler.data(), n: size - i); |
329 | } |
330 | |
331 | #if LLVM_ENABLE_ZLIB |
332 | static SmallVector<uint8_t, 0> deflateShard(Ctx &ctx, ArrayRef<uint8_t> in, |
333 | int level, int flush) { |
334 | // 15 and 8 are default. windowBits=-15 is negative to generate raw deflate |
335 | // data with no zlib header or trailer. |
336 | z_stream s = {}; |
337 | auto res = deflateInit2(&s, level, Z_DEFLATED, -15, 8, Z_DEFAULT_STRATEGY); |
338 | if (res != 0) { |
339 | Err(ctx) << "--compress-sections: deflateInit2 returned "<< res; |
340 | return {}; |
341 | } |
342 | s.next_in = const_cast<uint8_t *>(in.data()); |
343 | s.avail_in = in.size(); |
344 | |
345 | // Allocate a buffer of half of the input size, and grow it by 1.5x if |
346 | // insufficient. |
347 | SmallVector<uint8_t, 0> out; |
348 | size_t pos = 0; |
349 | out.resize_for_overwrite(N: std::max<size_t>(a: in.size() / 2, b: 64)); |
350 | do { |
351 | if (pos == out.size()) |
352 | out.resize_for_overwrite(N: out.size() * 3 / 2); |
353 | s.next_out = out.data() + pos; |
354 | s.avail_out = out.size() - pos; |
355 | (void)deflate(strm: &s, flush); |
356 | pos = s.next_out - out.data(); |
357 | } while (s.avail_out == 0); |
358 | assert(s.avail_in == 0); |
359 | |
360 | out.truncate(N: pos); |
361 | deflateEnd(strm: &s); |
362 | return out; |
363 | } |
364 | #endif |
365 | |
366 | // Compress certain non-SHF_ALLOC sections: |
367 | // |
368 | // * (if --compress-debug-sections is specified) non-empty .debug_* sections |
369 | // * (if --compress-sections is specified) matched sections |
370 | template <class ELFT> void OutputSection::maybeCompress(Ctx &ctx) { |
371 | using Elf_Chdr = typename ELFT::Chdr; |
372 | (void)sizeof(Elf_Chdr); |
373 | |
374 | DebugCompressionType ctype = DebugCompressionType::None; |
375 | size_t compressedSize = sizeof(Elf_Chdr); |
376 | unsigned level = 0; // default compression level |
377 | if (!(flags & SHF_ALLOC) && ctx.arg.compressDebugSections && |
378 | name.starts_with(Prefix: ".debug_")) |
379 | ctype = *ctx.arg.compressDebugSections; |
380 | for (auto &[glob, t, l] : ctx.arg.compressSections) |
381 | if (glob.match(S: name)) |
382 | std::tie(args&: ctype, args&: level) = {t, l}; |
383 | if (ctype == DebugCompressionType::None) |
384 | return; |
385 | if (flags & SHF_ALLOC) { |
386 | Err(ctx) << "--compress-sections: section '"<< name |
387 | << "' with the SHF_ALLOC flag cannot be compressed"; |
388 | return; |
389 | } |
390 | |
391 | llvm::TimeTraceScope timeScope("Compress sections"); |
392 | auto buf = std::make_unique<uint8_t[]>(num: size); |
393 | // Write uncompressed data to a temporary zero-initialized buffer. |
394 | { |
395 | parallel::TaskGroup tg; |
396 | writeTo<ELFT>(ctx, buf.get(), tg); |
397 | } |
398 | // The generic ABI specifies "The sh_size and sh_addralign fields of the |
399 | // section header for a compressed section reflect the requirements of the |
400 | // compressed section." However, 1-byte alignment has been wildly accepted |
401 | // and utilized for a long time. Removing alignment padding is particularly |
402 | // useful when there are many compressed output sections. |
403 | addralign = 1; |
404 | |
405 | // Split input into 1-MiB shards. |
406 | [[maybe_unused]] constexpr size_t shardSize = 1 << 20; |
407 | auto shardsIn = split(arr: ArrayRef<uint8_t>(buf.get(), size), chunkSize: shardSize); |
408 | const size_t numShards = shardsIn.size(); |
409 | auto shardsOut = std::make_unique<SmallVector<uint8_t, 0>[]>(num: numShards); |
410 | |
411 | #if LLVM_ENABLE_ZSTD |
412 | // Use ZSTD's streaming compression API. See |
413 | // http://facebook.github.io/zstd/zstd_manual.html "Streaming compression - |
414 | // HowTo". |
415 | if (ctype == DebugCompressionType::Zstd) { |
416 | parallelFor(0, numShards, [&](size_t i) { |
417 | SmallVector<uint8_t, 0> out; |
418 | ZSTD_CCtx *cctx = ZSTD_createCCtx(); |
419 | ZSTD_CCtx_setParameter(cctx, param: ZSTD_c_compressionLevel, value: level); |
420 | ZSTD_inBuffer zib = {.src: shardsIn[i].data(), .size: shardsIn[i].size(), .pos: 0}; |
421 | ZSTD_outBuffer zob = {.dst: nullptr, .size: 0, .pos: 0}; |
422 | size_t size; |
423 | do { |
424 | // Allocate a buffer of half of the input size, and grow it by 1.5x if |
425 | // insufficient. |
426 | if (zob.pos == zob.size) { |
427 | out.resize_for_overwrite( |
428 | N: zob.size ? zob.size * 3 / 2 : std::max<size_t>(a: zib.size / 4, b: 64)); |
429 | zob = {.dst: out.data(), .size: out.size(), .pos: zob.pos}; |
430 | } |
431 | size = ZSTD_compressStream2(cctx, output: &zob, input: &zib, endOp: ZSTD_e_end); |
432 | assert(!ZSTD_isError(size)); |
433 | } while (size != 0); |
434 | out.truncate(N: zob.pos); |
435 | ZSTD_freeCCtx(cctx); |
436 | shardsOut[i] = std::move(out); |
437 | }); |
438 | compressed.type = ELFCOMPRESS_ZSTD; |
439 | for (size_t i = 0; i != numShards; ++i) |
440 | compressedSize += shardsOut[i].size(); |
441 | } |
442 | #endif |
443 | |
444 | #if LLVM_ENABLE_ZLIB |
445 | // We chose 1 (Z_BEST_SPEED) as the default compression level because it is |
446 | // fast and provides decent compression ratios. |
447 | if (ctype == DebugCompressionType::Zlib) { |
448 | if (!level) |
449 | level = Z_BEST_SPEED; |
450 | |
451 | // Compress shards and compute Alder-32 checksums. Use Z_SYNC_FLUSH for all |
452 | // shards but the last to flush the output to a byte boundary to be |
453 | // concatenated with the next shard. |
454 | auto shardsAdler = std::make_unique<uint32_t[]>(num: numShards); |
455 | parallelFor(0, numShards, [&](size_t i) { |
456 | shardsOut[i] = deflateShard(ctx, in: shardsIn[i], level, |
457 | flush: i != numShards - 1 ? Z_SYNC_FLUSH : Z_FINISH); |
458 | shardsAdler[i] = adler32(adler: 1, buf: shardsIn[i].data(), len: shardsIn[i].size()); |
459 | }); |
460 | |
461 | // Update section size and combine Alder-32 checksums. |
462 | uint32_t checksum = 1; // Initial Adler-32 value |
463 | compressedSize += 2; // Elf_Chdir and zlib header |
464 | for (size_t i = 0; i != numShards; ++i) { |
465 | compressedSize += shardsOut[i].size(); |
466 | checksum = adler32_combine(checksum, shardsAdler[i], shardsIn[i].size()); |
467 | } |
468 | compressedSize += 4; // checksum |
469 | compressed.type = ELFCOMPRESS_ZLIB; |
470 | compressed.checksum = checksum; |
471 | } |
472 | #endif |
473 | |
474 | if (compressedSize >= size) |
475 | return; |
476 | compressed.uncompressedSize = size; |
477 | compressed.shards = std::move(shardsOut); |
478 | compressed.numShards = numShards; |
479 | size = compressedSize; |
480 | flags |= SHF_COMPRESSED; |
481 | } |
482 | |
483 | static void writeInt(Ctx &ctx, uint8_t *buf, uint64_t data, uint64_t size) { |
484 | if (size == 1) |
485 | *buf = data; |
486 | else if (size == 2) |
487 | write16(ctx, p: buf, v: data); |
488 | else if (size == 4) |
489 | write32(ctx, p: buf, v: data); |
490 | else if (size == 8) |
491 | write64(ctx, p: buf, v: data); |
492 | else |
493 | llvm_unreachable("unsupported Size argument"); |
494 | } |
495 | |
496 | template <class ELFT> |
497 | void OutputSection::writeTo(Ctx &ctx, uint8_t *buf, parallel::TaskGroup &tg) { |
498 | llvm::TimeTraceScope timeScope("Write sections", name); |
499 | if (type == SHT_NOBITS) |
500 | return; |
501 | if (type == SHT_CREL && !(flags & SHF_ALLOC)) { |
502 | buf += encodeULEB128(Value: crelHeader, p: buf); |
503 | memcpy(dest: buf, src: crelBody.data(), n: crelBody.size()); |
504 | return; |
505 | } |
506 | |
507 | // If the section is compressed due to |
508 | // --compress-debug-section/--compress-sections, the content is already known. |
509 | if (compressed.shards) { |
510 | auto *chdr = reinterpret_cast<typename ELFT::Chdr *>(buf); |
511 | chdr->ch_type = compressed.type; |
512 | chdr->ch_size = compressed.uncompressedSize; |
513 | chdr->ch_addralign = addralign; |
514 | buf += sizeof(*chdr); |
515 | |
516 | auto offsets = std::make_unique<size_t[]>(num: compressed.numShards); |
517 | if (compressed.type == ELFCOMPRESS_ZLIB) { |
518 | buf[0] = 0x78; // CMF |
519 | buf[1] = 0x01; // FLG: best speed |
520 | offsets[0] = 2; // zlib header |
521 | write32be(P: buf + (size - sizeof(*chdr) - 4), V: compressed.checksum); |
522 | } |
523 | |
524 | // Compute shard offsets. |
525 | for (size_t i = 1; i != compressed.numShards; ++i) |
526 | offsets[i] = offsets[i - 1] + compressed.shards[i - 1].size(); |
527 | parallelFor(0, compressed.numShards, [&](size_t i) { |
528 | memcpy(dest: buf + offsets[i], src: compressed.shards[i].data(), |
529 | n: compressed.shards[i].size()); |
530 | }); |
531 | return; |
532 | } |
533 | |
534 | // Write leading padding. |
535 | ArrayRef<InputSection *> sections = getInputSections(os: *this, storage); |
536 | std::array<uint8_t, 4> filler = getFiller(ctx); |
537 | bool nonZeroFiller = read32(ctx, p: filler.data()) != 0; |
538 | if (nonZeroFiller) |
539 | fill(buf, size: sections.empty() ? size : sections[0]->outSecOff, filler); |
540 | |
541 | if (type == SHT_CREL && !(flags & SHF_ALLOC)) { |
542 | buf += encodeULEB128(Value: crelHeader, p: buf); |
543 | memcpy(dest: buf, src: crelBody.data(), n: crelBody.size()); |
544 | return; |
545 | } |
546 | |
547 | auto fn = [=, &ctx](size_t begin, size_t end) { |
548 | size_t numSections = sections.size(); |
549 | for (size_t i = begin; i != end; ++i) { |
550 | InputSection *isec = sections[i]; |
551 | if (auto *s = dyn_cast<SyntheticSection>(Val: isec)) |
552 | s->writeTo(buf: buf + isec->outSecOff); |
553 | else |
554 | isec->writeTo<ELFT>(ctx, buf + isec->outSecOff); |
555 | |
556 | // When in Arm BE8 mode, the linker has to convert the big-endian |
557 | // instructions to little-endian, leaving the data big-endian. |
558 | if (ctx.arg.emachine == EM_ARM && !ctx.arg.isLE && ctx.arg.armBe8 && |
559 | (flags & SHF_EXECINSTR)) |
560 | convertArmInstructionstoBE8(ctx, sec: isec, buf: buf + isec->outSecOff); |
561 | |
562 | // Fill gaps between sections. |
563 | if (nonZeroFiller) { |
564 | uint8_t *start = buf + isec->outSecOff + isec->getSize(); |
565 | uint8_t *end; |
566 | if (i + 1 == numSections) |
567 | end = buf + size; |
568 | else |
569 | end = buf + sections[i + 1]->outSecOff; |
570 | if (isec->nopFiller) { |
571 | assert(ctx.target->nopInstrs); |
572 | nopInstrFill(ctx, buf: start, size: end - start); |
573 | } else |
574 | fill(buf: start, size: end - start, filler); |
575 | } |
576 | } |
577 | }; |
578 | |
579 | // If there is any BYTE()-family command (rare), write the section content |
580 | // first then process BYTE to overwrite the filler content. The write is |
581 | // serial due to the limitation of llvm/Support/Parallel.h. |
582 | bool written = false; |
583 | size_t numSections = sections.size(); |
584 | for (SectionCommand *cmd : commands) |
585 | if (auto *data = dyn_cast<ByteCommand>(Val: cmd)) { |
586 | if (!std::exchange(obj&: written, new_val: true)) |
587 | fn(0, numSections); |
588 | writeInt(ctx, buf: buf + data->offset, data: data->expression().getValue(), |
589 | size: data->size); |
590 | } |
591 | if (written || !numSections) |
592 | return; |
593 | |
594 | // There is no data command. Write content asynchronously to overlap the write |
595 | // time with other output sections. Note, if a linker script specifies |
596 | // overlapping output sections (needs --noinhibit-exec or --no-check-sections |
597 | // to supress the error), the output may be non-deterministic. |
598 | const size_t taskSizeLimit = 4 << 20; |
599 | for (size_t begin = 0, i = 0, taskSize = 0;;) { |
600 | taskSize += sections[i]->getSize(); |
601 | bool done = ++i == numSections; |
602 | if (done || taskSize >= taskSizeLimit) { |
603 | tg.spawn(f: [=] { fn(begin, i); }); |
604 | if (done) |
605 | break; |
606 | begin = i; |
607 | taskSize = 0; |
608 | } |
609 | } |
610 | } |
611 | |
612 | static void finalizeShtGroup(Ctx &ctx, OutputSection *os, |
613 | InputSection *section) { |
614 | // sh_link field for SHT_GROUP sections should contain the section index of |
615 | // the symbol table. |
616 | os->link = ctx.in.symTab->getParent()->sectionIndex; |
617 | |
618 | if (!section) |
619 | return; |
620 | |
621 | // sh_info then contain index of an entry in symbol table section which |
622 | // provides signature of the section group. |
623 | ArrayRef<Symbol *> symbols = section->file->getSymbols(); |
624 | os->info = ctx.in.symTab->getSymbolIndex(sym: *symbols[section->info]); |
625 | |
626 | // Some group members may be combined or discarded, so we need to compute the |
627 | // new size. The content will be rewritten in InputSection::copyShtGroup. |
628 | DenseSet<uint32_t> seen; |
629 | ArrayRef<InputSectionBase *> sections = section->file->getSections(); |
630 | for (const uint32_t &idx : section->getDataAs<uint32_t>().slice(N: 1)) |
631 | if (OutputSection *osec = sections[read32(ctx, p: &idx)]->getOutputSection()) |
632 | seen.insert(V: osec->sectionIndex); |
633 | os->size = (1 + seen.size()) * sizeof(uint32_t); |
634 | } |
635 | |
636 | template <class uint> |
637 | LLVM_ATTRIBUTE_ALWAYS_INLINE static void |
638 | encodeOneCrel(Ctx &ctx, raw_svector_ostream &os, |
639 | Elf_Crel<sizeof(uint) == 8> &out, uint offset, const Symbol &sym, |
640 | uint32_t type, uint addend) { |
641 | const auto deltaOffset = static_cast<uint64_t>(offset - out.r_offset); |
642 | out.r_offset = offset; |
643 | int64_t symidx = ctx.in.symTab->getSymbolIndex(sym); |
644 | if (sym.type == STT_SECTION) { |
645 | auto *d = dyn_cast<Defined>(Val: &sym); |
646 | if (d) { |
647 | SectionBase *section = d->section; |
648 | assert(section->isLive()); |
649 | addend = sym.getVA(ctx, addend) - section->getOutputSection()->addr; |
650 | } else { |
651 | // Encode R_*_NONE(symidx=0). |
652 | symidx = type = addend = 0; |
653 | } |
654 | } |
655 | |
656 | // Similar to llvm::ELF::encodeCrel. |
657 | uint8_t b = deltaOffset * 8 + (out.r_symidx != symidx) + |
658 | (out.r_type != type ? 2 : 0) + |
659 | (uint(out.r_addend) != addend ? 4 : 0); |
660 | if (deltaOffset < 0x10) { |
661 | os << char(b); |
662 | } else { |
663 | os << char(b | 0x80); |
664 | encodeULEB128(Value: deltaOffset >> 4, OS&: os); |
665 | } |
666 | if (b & 1) { |
667 | encodeSLEB128(Value: static_cast<int32_t>(symidx - out.r_symidx), OS&: os); |
668 | out.r_symidx = symidx; |
669 | } |
670 | if (b & 2) { |
671 | encodeSLEB128(Value: static_cast<int32_t>(type - out.r_type), OS&: os); |
672 | out.r_type = type; |
673 | } |
674 | if (b & 4) { |
675 | encodeSLEB128(std::make_signed_t<uint>(addend - out.r_addend), os); |
676 | out.r_addend = addend; |
677 | } |
678 | } |
679 | |
680 | template <class ELFT> |
681 | static size_t relToCrel(Ctx &ctx, raw_svector_ostream &os, |
682 | Elf_Crel<ELFT::Is64Bits> &out, InputSection *relSec, |
683 | InputSectionBase *sec) { |
684 | const auto &file = *cast<ELFFileBase>(Val: relSec->file); |
685 | if (relSec->type == SHT_REL) { |
686 | // REL conversion is complex and unsupported yet. |
687 | Err(ctx) << relSec << ": REL cannot be converted to CREL"; |
688 | return 0; |
689 | } |
690 | auto rels = relSec->getDataAs<typename ELFT::Rela>(); |
691 | for (auto rel : rels) { |
692 | encodeOneCrel<typename ELFT::uint>( |
693 | ctx, os, out, sec->getVA(offset: rel.r_offset), file.getRelocTargetSym(rel), |
694 | rel.getType(ctx.arg.isMips64EL), getAddend<ELFT>(rel)); |
695 | } |
696 | return rels.size(); |
697 | } |
698 | |
699 | // Compute the content of a non-alloc CREL section due to -r or --emit-relocs. |
700 | // Input CREL sections are decoded while REL[A] need to be converted. |
701 | template <bool is64> void OutputSection::finalizeNonAllocCrel(Ctx &ctx) { |
702 | using uint = typename Elf_Crel_Impl<is64>::uint; |
703 | raw_svector_ostream os(crelBody); |
704 | uint64_t totalCount = 0; |
705 | Elf_Crel<is64> out{}; |
706 | assert(commands.size() == 1); |
707 | auto *isd = cast<InputSectionDescription>(Val: commands[0]); |
708 | for (InputSection *relSec : isd->sections) { |
709 | const auto &file = *cast<ELFFileBase>(Val: relSec->file); |
710 | InputSectionBase *sec = relSec->getRelocatedSection(); |
711 | if (relSec->type == SHT_CREL) { |
712 | RelocsCrel<is64> entries(relSec->content_); |
713 | totalCount += entries.size(); |
714 | for (Elf_Crel_Impl<is64> r : entries) { |
715 | encodeOneCrel<uint>(ctx, os, out, uint(sec->getVA(offset: r.r_offset)), |
716 | file.getSymbol(symbolIndex: r.r_symidx), r.r_type, r.r_addend); |
717 | } |
718 | continue; |
719 | } |
720 | |
721 | // Convert REL[A] to CREL. |
722 | if constexpr (is64) { |
723 | totalCount += ctx.arg.isLE |
724 | ? relToCrel<ELF64LE>(ctx, os, out, relSec, sec) |
725 | : relToCrel<ELF64BE>(ctx, os, out, relSec, sec); |
726 | } else { |
727 | totalCount += ctx.arg.isLE |
728 | ? relToCrel<ELF32LE>(ctx, os, out, relSec, sec) |
729 | : relToCrel<ELF32BE>(ctx, os, out, relSec, sec); |
730 | } |
731 | } |
732 | |
733 | crelHeader = totalCount * 8 + 4; |
734 | size = getULEB128Size(Value: crelHeader) + crelBody.size(); |
735 | } |
736 | |
737 | void OutputSection::finalize(Ctx &ctx) { |
738 | InputSection *first = getFirstInputSection(os: this); |
739 | |
740 | if (flags & SHF_LINK_ORDER) { |
741 | // We must preserve the link order dependency of sections with the |
742 | // SHF_LINK_ORDER flag. The dependency is indicated by the sh_link field. We |
743 | // need to translate the InputSection sh_link to the OutputSection sh_link, |
744 | // all InputSections in the OutputSection have the same dependency. |
745 | if (auto *ex = dyn_cast<ARMExidxSyntheticSection>(Val: first)) |
746 | link = ex->getLinkOrderDep()->getParent()->sectionIndex; |
747 | else if (first->flags & SHF_LINK_ORDER) |
748 | if (auto *d = first->getLinkOrderDep()) |
749 | link = d->getParent()->sectionIndex; |
750 | } |
751 | |
752 | if (type == SHT_GROUP) { |
753 | finalizeShtGroup(ctx, os: this, section: first); |
754 | return; |
755 | } |
756 | |
757 | if (!ctx.arg.copyRelocs || !isStaticRelSecType(type)) |
758 | return; |
759 | |
760 | // Skip if 'first' is synthetic, i.e. not a section created by --emit-relocs. |
761 | // Normally 'type' was changed by 'first' so 'first' should be non-null. |
762 | // However, if the output section is .rela.dyn, 'type' can be set by the empty |
763 | // synthetic .rela.plt and first can be null. |
764 | if (!first || isa<SyntheticSection>(Val: first)) |
765 | return; |
766 | |
767 | link = ctx.in.symTab->getParent()->sectionIndex; |
768 | // sh_info for SHT_REL[A] sections should contain the section header index of |
769 | // the section to which the relocation applies. |
770 | InputSectionBase *s = first->getRelocatedSection(); |
771 | info = s->getOutputSection()->sectionIndex; |
772 | flags |= SHF_INFO_LINK; |
773 | // Finalize the content of non-alloc CREL. |
774 | if (type == SHT_CREL) { |
775 | if (ctx.arg.is64) |
776 | finalizeNonAllocCrel<true>(ctx); |
777 | else |
778 | finalizeNonAllocCrel<false>(ctx); |
779 | } |
780 | } |
781 | |
782 | // Returns true if S is in one of the many forms the compiler driver may pass |
783 | // crtbegin files. |
784 | // |
785 | // Gcc uses any of crtbegin[<empty>|S|T].o. |
786 | // Clang uses Gcc's plus clang_rt.crtbegin[-<arch>|<empty>].o. |
787 | |
788 | static bool isCrt(StringRef s, StringRef beginEnd) { |
789 | s = sys::path::filename(path: s); |
790 | if (!s.consume_back(Suffix: ".o")) |
791 | return false; |
792 | if (s.consume_front(Prefix: "clang_rt.")) |
793 | return s.consume_front(Prefix: beginEnd); |
794 | return s.consume_front(Prefix: beginEnd) && s.size() <= 1; |
795 | } |
796 | |
797 | // .ctors and .dtors are sorted by this order: |
798 | // |
799 | // 1. .ctors/.dtors in crtbegin (which contains a sentinel value -1). |
800 | // 2. The section is named ".ctors" or ".dtors" (priority: 65536). |
801 | // 3. The section has an optional priority value in the form of ".ctors.N" or |
802 | // ".dtors.N" where N is a number in the form of %05u (priority: 65535-N). |
803 | // 4. .ctors/.dtors in crtend (which contains a sentinel value 0). |
804 | // |
805 | // For 2 and 3, the sections are sorted by priority from high to low, e.g. |
806 | // .ctors (65536), .ctors.00100 (65436), .ctors.00200 (65336). In GNU ld's |
807 | // internal linker scripts, the sorting is by string comparison which can |
808 | // achieve the same goal given the optional priority values are of the same |
809 | // length. |
810 | // |
811 | // In an ideal world, we don't need this function because .init_array and |
812 | // .ctors are duplicate features (and .init_array is newer.) However, there |
813 | // are too many real-world use cases of .ctors, so we had no choice to |
814 | // support that with this rather ad-hoc semantics. |
815 | static bool compCtors(const InputSection *a, const InputSection *b) { |
816 | bool beginA = isCrt(s: a->file->getName(), beginEnd: "crtbegin"); |
817 | bool beginB = isCrt(s: b->file->getName(), beginEnd: "crtbegin"); |
818 | if (beginA != beginB) |
819 | return beginA; |
820 | bool endA = isCrt(s: a->file->getName(), beginEnd: "crtend"); |
821 | bool endB = isCrt(s: b->file->getName(), beginEnd: "crtend"); |
822 | if (endA != endB) |
823 | return endB; |
824 | return getPriority(s: a->name) > getPriority(s: b->name); |
825 | } |
826 | |
827 | // Sorts input sections by the special rules for .ctors and .dtors. |
828 | // Unfortunately, the rules are different from the one for .{init,fini}_array. |
829 | // Read the comment above. |
830 | void OutputSection::sortCtorsDtors() { |
831 | assert(commands.size() == 1); |
832 | auto *isd = cast<InputSectionDescription>(Val: commands[0]); |
833 | llvm::stable_sort(Range&: isd->sections, C: compCtors); |
834 | } |
835 | |
836 | // If an input string is in the form of "foo.N" where N is a number, return N |
837 | // (65535-N if .ctors.N or .dtors.N). Otherwise, returns 65536, which is one |
838 | // greater than the lowest priority. |
839 | int elf::getPriority(StringRef s) { |
840 | size_t pos = s.rfind(C: '.'); |
841 | if (pos == StringRef::npos) |
842 | return 65536; |
843 | int v = 65536; |
844 | if (to_integer(S: s.substr(Start: pos + 1), Num&: v, Base: 10) && |
845 | (pos == 6 && (s.starts_with(Prefix: ".ctors") || s.starts_with(Prefix: ".dtors")))) |
846 | v = 65535 - v; |
847 | return v; |
848 | } |
849 | |
850 | InputSection *elf::getFirstInputSection(const OutputSection *os) { |
851 | for (SectionCommand *cmd : os->commands) |
852 | if (auto *isd = dyn_cast<InputSectionDescription>(Val: cmd)) |
853 | if (!isd->sections.empty()) |
854 | return isd->sections[0]; |
855 | return nullptr; |
856 | } |
857 | |
858 | ArrayRef<InputSection *> |
859 | elf::getInputSections(const OutputSection &os, |
860 | SmallVector<InputSection *, 0> &storage) { |
861 | ArrayRef<InputSection *> ret; |
862 | storage.clear(); |
863 | for (SectionCommand *cmd : os.commands) { |
864 | auto *isd = dyn_cast<InputSectionDescription>(Val: cmd); |
865 | if (!isd) |
866 | continue; |
867 | if (ret.empty()) { |
868 | ret = isd->sections; |
869 | } else { |
870 | if (storage.empty()) |
871 | storage.assign(in_start: ret.begin(), in_end: ret.end()); |
872 | storage.insert(I: storage.end(), From: isd->sections.begin(), To: isd->sections.end()); |
873 | } |
874 | } |
875 | return storage.empty() ? ret : ArrayRef(storage); |
876 | } |
877 | |
878 | // Sorts input sections by section name suffixes, so that .foo.N comes |
879 | // before .foo.M if N < M. Used to sort .{init,fini}_array.N sections. |
880 | // We want to keep the original order if the priorities are the same |
881 | // because the compiler keeps the original initialization order in a |
882 | // translation unit and we need to respect that. |
883 | // For more detail, read the section of the GCC's manual about init_priority. |
884 | void OutputSection::sortInitFini() { |
885 | // Sort sections by priority. |
886 | sort(order: [](InputSectionBase *s) { return getPriority(s: s->name); }); |
887 | } |
888 | |
889 | std::array<uint8_t, 4> OutputSection::getFiller(Ctx &ctx) { |
890 | if (filler) |
891 | return *filler; |
892 | if (flags & SHF_EXECINSTR) |
893 | return ctx.target->trapInstr; |
894 | return {0, 0, 0, 0}; |
895 | } |
896 | |
897 | void OutputSection::checkDynRelAddends(Ctx &ctx) { |
898 | assert(ctx.arg.writeAddends && ctx.arg.checkDynamicRelocs); |
899 | assert(isStaticRelSecType(type)); |
900 | SmallVector<InputSection *, 0> storage; |
901 | ArrayRef<InputSection *> sections = getInputSections(os: *this, storage); |
902 | parallelFor(Begin: 0, End: sections.size(), Fn: [&](size_t i) { |
903 | // When linking with -r or --emit-relocs we might also call this function |
904 | // for input .rel[a].<sec> sections which we simply pass through to the |
905 | // output. We skip over those and only look at the synthetic relocation |
906 | // sections created during linking. |
907 | if (!SyntheticSection::classof(sec: sections[i]) || |
908 | !is_contained(Set: {ELF::SHT_REL, ELF::SHT_RELA, ELF::SHT_RELR}, |
909 | Element: sections[i]->type)) |
910 | return; |
911 | const auto *sec = cast<RelocationBaseSection>(Val: sections[i]); |
912 | if (!sec) |
913 | return; |
914 | for (const DynamicReloc &rel : sec->relocs) { |
915 | int64_t addend = rel.addend; |
916 | const OutputSection *relOsec = rel.inputSec->getOutputSection(); |
917 | assert(relOsec != nullptr && "missing output section for relocation"); |
918 | // Some targets have NOBITS synthetic sections with dynamic relocations |
919 | // with non-zero addends. Skip such sections. |
920 | if (is_contained(Set: {EM_PPC, EM_PPC64}, Element: ctx.arg.emachine) && |
921 | (rel.inputSec == ctx.in.ppc64LongBranchTarget.get() || |
922 | rel.inputSec == ctx.in.igotPlt.get())) |
923 | continue; |
924 | const uint8_t *relocTarget = ctx.bufferStart + relOsec->offset + |
925 | rel.inputSec->getOffset(offset: rel.offsetInSec); |
926 | // For SHT_NOBITS the written addend is always zero. |
927 | int64_t writtenAddend = |
928 | relOsec->type == SHT_NOBITS |
929 | ? 0 |
930 | : ctx.target->getImplicitAddend(buf: relocTarget, type: rel.type); |
931 | if (addend != writtenAddend) |
932 | InternalErr(ctx, buf: relocTarget) |
933 | << "wrote incorrect addend value 0x"<< utohexstr(X: writtenAddend) |
934 | << " instead of 0x"<< utohexstr(X: addend) |
935 | << " for dynamic relocation "<< rel.type << " at offset 0x" |
936 | << utohexstr(X: rel.getOffset()) |
937 | << (rel.sym ? " against symbol "+ rel.sym->getName() : ""); |
938 | } |
939 | }); |
940 | } |
941 | |
942 | template void OutputSection::writeHeaderTo<ELF32LE>(ELF32LE::Shdr *Shdr); |
943 | template void OutputSection::writeHeaderTo<ELF32BE>(ELF32BE::Shdr *Shdr); |
944 | template void OutputSection::writeHeaderTo<ELF64LE>(ELF64LE::Shdr *Shdr); |
945 | template void OutputSection::writeHeaderTo<ELF64BE>(ELF64BE::Shdr *Shdr); |
946 | |
947 | template void OutputSection::writeTo<ELF32LE>(Ctx &, uint8_t *, |
948 | llvm::parallel::TaskGroup &); |
949 | template void OutputSection::writeTo<ELF32BE>(Ctx &, uint8_t *, |
950 | llvm::parallel::TaskGroup &); |
951 | template void OutputSection::writeTo<ELF64LE>(Ctx &, uint8_t *, |
952 | llvm::parallel::TaskGroup &); |
953 | template void OutputSection::writeTo<ELF64BE>(Ctx &, uint8_t *, |
954 | llvm::parallel::TaskGroup &); |
955 | |
956 | template void OutputSection::maybeCompress<ELF32LE>(Ctx &); |
957 | template void OutputSection::maybeCompress<ELF32BE>(Ctx &); |
958 | template void OutputSection::maybeCompress<ELF64LE>(Ctx &); |
959 | template void OutputSection::maybeCompress<ELF64BE>(Ctx &); |
960 |
Definitions
- getPhdrFlags
- writeHeaderTo
- OutputSection
- getLMA
- canMergeToProgbits
- recordSection
- commitSection
- createMergeSynthetic
- finalizeInputSections
- sortByOrder
- getHeaderSize
- sort
- nopInstrFill
- fill
- deflateShard
- maybeCompress
- writeInt
- writeTo
- finalizeShtGroup
- encodeOneCrel
- relToCrel
- finalizeNonAllocCrel
- finalize
- isCrt
- compCtors
- sortCtorsDtors
- getPriority
- getFirstInputSection
- getInputSections
- sortInitFini
- getFiller
Update your C++ knowledge – Modern C++11/14/17 Training
Find out more