1 | //===- Chunks.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 "Chunks.h" |
10 | #include "COFFLinkerContext.h" |
11 | #include "InputFiles.h" |
12 | #include "SymbolTable.h" |
13 | #include "Symbols.h" |
14 | #include "Writer.h" |
15 | #include "llvm/ADT/STLExtras.h" |
16 | #include "llvm/ADT/StringExtras.h" |
17 | #include "llvm/ADT/Twine.h" |
18 | #include "llvm/BinaryFormat/COFF.h" |
19 | #include "llvm/Object/COFF.h" |
20 | #include "llvm/Support/Debug.h" |
21 | #include "llvm/Support/Endian.h" |
22 | #include "llvm/Support/raw_ostream.h" |
23 | #include <algorithm> |
24 | #include <iterator> |
25 | |
26 | using namespace llvm; |
27 | using namespace llvm::object; |
28 | using namespace llvm::support::endian; |
29 | using namespace llvm::COFF; |
30 | using llvm::support::ulittle32_t; |
31 | |
32 | namespace lld::coff { |
33 | |
34 | SectionChunk::SectionChunk(ObjFile *f, const coff_section *h) |
35 | : Chunk(SectionKind), file(f), header(h), repl(this) { |
36 | // Initialize relocs. |
37 | if (file) |
38 | setRelocs(file->getCOFFObj()->getRelocations(Sec: header)); |
39 | |
40 | // Initialize sectionName. |
41 | StringRef sectionName; |
42 | if (file) { |
43 | if (Expected<StringRef> e = file->getCOFFObj()->getSectionName(Sec: header)) |
44 | sectionName = *e; |
45 | } |
46 | sectionNameData = sectionName.data(); |
47 | sectionNameSize = sectionName.size(); |
48 | |
49 | setAlignment(header->getAlignment()); |
50 | |
51 | hasData = !(header->Characteristics & IMAGE_SCN_CNT_UNINITIALIZED_DATA); |
52 | |
53 | // If linker GC is disabled, every chunk starts out alive. If linker GC is |
54 | // enabled, treat non-comdat sections as roots. Generally optimized object |
55 | // files will be built with -ffunction-sections or /Gy, so most things worth |
56 | // stripping will be in a comdat. |
57 | if (file) |
58 | live = !file->ctx.config.doGC || !isCOMDAT(); |
59 | else |
60 | live = true; |
61 | } |
62 | |
63 | // SectionChunk is one of the most frequently allocated classes, so it is |
64 | // important to keep it as compact as possible. As of this writing, the number |
65 | // below is the size of this class on x64 platforms. |
66 | static_assert(sizeof(SectionChunk) <= 88, "SectionChunk grew unexpectedly" ); |
67 | |
68 | static void add16(uint8_t *p, int16_t v) { write16le(P: p, V: read16le(P: p) + v); } |
69 | static void add32(uint8_t *p, int32_t v) { write32le(P: p, V: read32le(P: p) + v); } |
70 | static void add64(uint8_t *p, int64_t v) { write64le(P: p, V: read64le(P: p) + v); } |
71 | static void or16(uint8_t *p, uint16_t v) { write16le(P: p, V: read16le(P: p) | v); } |
72 | static void or32(uint8_t *p, uint32_t v) { write32le(P: p, V: read32le(P: p) | v); } |
73 | |
74 | // Verify that given sections are appropriate targets for SECREL |
75 | // relocations. This check is relaxed because unfortunately debug |
76 | // sections have section-relative relocations against absolute symbols. |
77 | static bool checkSecRel(const SectionChunk *sec, OutputSection *os) { |
78 | if (os) |
79 | return true; |
80 | if (sec->isCodeView()) |
81 | return false; |
82 | error(msg: "SECREL relocation cannot be applied to absolute symbols" ); |
83 | return false; |
84 | } |
85 | |
86 | static void applySecRel(const SectionChunk *sec, uint8_t *off, |
87 | OutputSection *os, uint64_t s) { |
88 | if (!checkSecRel(sec, os)) |
89 | return; |
90 | uint64_t secRel = s - os->getRVA(); |
91 | if (secRel > UINT32_MAX) { |
92 | error(msg: "overflow in SECREL relocation in section: " + sec->getSectionName()); |
93 | return; |
94 | } |
95 | add32(p: off, v: secRel); |
96 | } |
97 | |
98 | static void applySecIdx(uint8_t *off, OutputSection *os, |
99 | unsigned numOutputSections) { |
100 | // numOutputSections is the largest valid section index. Make sure that |
101 | // it fits in 16 bits. |
102 | assert(numOutputSections <= 0xffff && "size of outputSections is too big" ); |
103 | |
104 | // Absolute symbol doesn't have section index, but section index relocation |
105 | // against absolute symbol should be resolved to one plus the last output |
106 | // section index. This is required for compatibility with MSVC. |
107 | if (os) |
108 | add16(p: off, v: os->sectionIndex); |
109 | else |
110 | add16(p: off, v: numOutputSections + 1); |
111 | } |
112 | |
113 | void SectionChunk::applyRelX64(uint8_t *off, uint16_t type, OutputSection *os, |
114 | uint64_t s, uint64_t p, |
115 | uint64_t imageBase) const { |
116 | switch (type) { |
117 | case IMAGE_REL_AMD64_ADDR32: |
118 | add32(p: off, v: s + imageBase); |
119 | break; |
120 | case IMAGE_REL_AMD64_ADDR64: |
121 | add64(p: off, v: s + imageBase); |
122 | break; |
123 | case IMAGE_REL_AMD64_ADDR32NB: add32(p: off, v: s); break; |
124 | case IMAGE_REL_AMD64_REL32: add32(p: off, v: s - p - 4); break; |
125 | case IMAGE_REL_AMD64_REL32_1: add32(p: off, v: s - p - 5); break; |
126 | case IMAGE_REL_AMD64_REL32_2: add32(p: off, v: s - p - 6); break; |
127 | case IMAGE_REL_AMD64_REL32_3: add32(p: off, v: s - p - 7); break; |
128 | case IMAGE_REL_AMD64_REL32_4: add32(p: off, v: s - p - 8); break; |
129 | case IMAGE_REL_AMD64_REL32_5: add32(p: off, v: s - p - 9); break; |
130 | case IMAGE_REL_AMD64_SECTION: |
131 | applySecIdx(off, os, numOutputSections: file->ctx.outputSections.size()); |
132 | break; |
133 | case IMAGE_REL_AMD64_SECREL: applySecRel(sec: this, off, os, s); break; |
134 | default: |
135 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
136 | toString(file)); |
137 | } |
138 | } |
139 | |
140 | void SectionChunk::applyRelX86(uint8_t *off, uint16_t type, OutputSection *os, |
141 | uint64_t s, uint64_t p, |
142 | uint64_t imageBase) const { |
143 | switch (type) { |
144 | case IMAGE_REL_I386_ABSOLUTE: break; |
145 | case IMAGE_REL_I386_DIR32: |
146 | add32(p: off, v: s + imageBase); |
147 | break; |
148 | case IMAGE_REL_I386_DIR32NB: add32(p: off, v: s); break; |
149 | case IMAGE_REL_I386_REL32: add32(p: off, v: s - p - 4); break; |
150 | case IMAGE_REL_I386_SECTION: |
151 | applySecIdx(off, os, numOutputSections: file->ctx.outputSections.size()); |
152 | break; |
153 | case IMAGE_REL_I386_SECREL: applySecRel(sec: this, off, os, s); break; |
154 | default: |
155 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
156 | toString(file)); |
157 | } |
158 | } |
159 | |
160 | static void applyMOV(uint8_t *off, uint16_t v) { |
161 | write16le(P: off, V: (read16le(P: off) & 0xfbf0) | ((v & 0x800) >> 1) | ((v >> 12) & 0xf)); |
162 | write16le(P: off + 2, V: (read16le(P: off + 2) & 0x8f00) | ((v & 0x700) << 4) | (v & 0xff)); |
163 | } |
164 | |
165 | static uint16_t readMOV(uint8_t *off, bool movt) { |
166 | uint16_t op1 = read16le(P: off); |
167 | if ((op1 & 0xfbf0) != (movt ? 0xf2c0 : 0xf240)) |
168 | error(msg: "unexpected instruction in " + Twine(movt ? "MOVT" : "MOVW" ) + |
169 | " instruction in MOV32T relocation" ); |
170 | uint16_t op2 = read16le(P: off + 2); |
171 | if ((op2 & 0x8000) != 0) |
172 | error(msg: "unexpected instruction in " + Twine(movt ? "MOVT" : "MOVW" ) + |
173 | " instruction in MOV32T relocation" ); |
174 | return (op2 & 0x00ff) | ((op2 >> 4) & 0x0700) | ((op1 << 1) & 0x0800) | |
175 | ((op1 & 0x000f) << 12); |
176 | } |
177 | |
178 | void applyMOV32T(uint8_t *off, uint32_t v) { |
179 | uint16_t immW = readMOV(off, movt: false); // read MOVW operand |
180 | uint16_t immT = readMOV(off: off + 4, movt: true); // read MOVT operand |
181 | uint32_t imm = immW | (immT << 16); |
182 | v += imm; // add the immediate offset |
183 | applyMOV(off, v); // set MOVW operand |
184 | applyMOV(off: off + 4, v: v >> 16); // set MOVT operand |
185 | } |
186 | |
187 | static void applyBranch20T(uint8_t *off, int32_t v) { |
188 | if (!isInt<21>(x: v)) |
189 | error(msg: "relocation out of range" ); |
190 | uint32_t s = v < 0 ? 1 : 0; |
191 | uint32_t j1 = (v >> 19) & 1; |
192 | uint32_t j2 = (v >> 18) & 1; |
193 | or16(p: off, v: (s << 10) | ((v >> 12) & 0x3f)); |
194 | or16(p: off + 2, v: (j1 << 13) | (j2 << 11) | ((v >> 1) & 0x7ff)); |
195 | } |
196 | |
197 | void applyBranch24T(uint8_t *off, int32_t v) { |
198 | if (!isInt<25>(x: v)) |
199 | error(msg: "relocation out of range" ); |
200 | uint32_t s = v < 0 ? 1 : 0; |
201 | uint32_t j1 = ((~v >> 23) & 1) ^ s; |
202 | uint32_t j2 = ((~v >> 22) & 1) ^ s; |
203 | or16(p: off, v: (s << 10) | ((v >> 12) & 0x3ff)); |
204 | // Clear out the J1 and J2 bits which may be set. |
205 | write16le(P: off + 2, V: (read16le(P: off + 2) & 0xd000) | (j1 << 13) | (j2 << 11) | ((v >> 1) & 0x7ff)); |
206 | } |
207 | |
208 | void SectionChunk::applyRelARM(uint8_t *off, uint16_t type, OutputSection *os, |
209 | uint64_t s, uint64_t p, |
210 | uint64_t imageBase) const { |
211 | // Pointer to thumb code must have the LSB set. |
212 | uint64_t sx = s; |
213 | if (os && (os->header.Characteristics & IMAGE_SCN_MEM_EXECUTE)) |
214 | sx |= 1; |
215 | switch (type) { |
216 | case IMAGE_REL_ARM_ADDR32: |
217 | add32(p: off, v: sx + imageBase); |
218 | break; |
219 | case IMAGE_REL_ARM_ADDR32NB: add32(p: off, v: sx); break; |
220 | case IMAGE_REL_ARM_MOV32T: |
221 | applyMOV32T(off, v: sx + imageBase); |
222 | break; |
223 | case IMAGE_REL_ARM_BRANCH20T: applyBranch20T(off, v: sx - p - 4); break; |
224 | case IMAGE_REL_ARM_BRANCH24T: applyBranch24T(off, v: sx - p - 4); break; |
225 | case IMAGE_REL_ARM_BLX23T: applyBranch24T(off, v: sx - p - 4); break; |
226 | case IMAGE_REL_ARM_SECTION: |
227 | applySecIdx(off, os, numOutputSections: file->ctx.outputSections.size()); |
228 | break; |
229 | case IMAGE_REL_ARM_SECREL: applySecRel(sec: this, off, os, s); break; |
230 | case IMAGE_REL_ARM_REL32: add32(p: off, v: sx - p - 4); break; |
231 | default: |
232 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
233 | toString(file)); |
234 | } |
235 | } |
236 | |
237 | // Interpret the existing immediate value as a byte offset to the |
238 | // target symbol, then update the instruction with the immediate as |
239 | // the page offset from the current instruction to the target. |
240 | void applyArm64Addr(uint8_t *off, uint64_t s, uint64_t p, int shift) { |
241 | uint32_t orig = read32le(P: off); |
242 | int64_t imm = |
243 | SignExtend64<21>(x: ((orig >> 29) & 0x3) | ((orig >> 3) & 0x1FFFFC)); |
244 | s += imm; |
245 | imm = (s >> shift) - (p >> shift); |
246 | uint32_t immLo = (imm & 0x3) << 29; |
247 | uint32_t immHi = (imm & 0x1FFFFC) << 3; |
248 | uint64_t mask = (0x3 << 29) | (0x1FFFFC << 3); |
249 | write32le(P: off, V: (orig & ~mask) | immLo | immHi); |
250 | } |
251 | |
252 | // Update the immediate field in a AARCH64 ldr, str, and add instruction. |
253 | // Optionally limit the range of the written immediate by one or more bits |
254 | // (rangeLimit). |
255 | void applyArm64Imm(uint8_t *off, uint64_t imm, uint32_t rangeLimit) { |
256 | uint32_t orig = read32le(P: off); |
257 | imm += (orig >> 10) & 0xFFF; |
258 | orig &= ~(0xFFF << 10); |
259 | write32le(P: off, V: orig | ((imm & (0xFFF >> rangeLimit)) << 10)); |
260 | } |
261 | |
262 | // Add the 12 bit page offset to the existing immediate. |
263 | // Ldr/str instructions store the opcode immediate scaled |
264 | // by the load/store size (giving a larger range for larger |
265 | // loads/stores). The immediate is always (both before and after |
266 | // fixing up the relocation) stored scaled similarly. |
267 | // Even if larger loads/stores have a larger range, limit the |
268 | // effective offset to 12 bit, since it is intended to be a |
269 | // page offset. |
270 | static void applyArm64Ldr(uint8_t *off, uint64_t imm) { |
271 | uint32_t orig = read32le(P: off); |
272 | uint32_t size = orig >> 30; |
273 | // 0x04000000 indicates SIMD/FP registers |
274 | // 0x00800000 indicates 128 bit |
275 | if ((orig & 0x4800000) == 0x4800000) |
276 | size += 4; |
277 | if ((imm & ((1 << size) - 1)) != 0) |
278 | error(msg: "misaligned ldr/str offset" ); |
279 | applyArm64Imm(off, imm: imm >> size, rangeLimit: size); |
280 | } |
281 | |
282 | static void applySecRelLow12A(const SectionChunk *sec, uint8_t *off, |
283 | OutputSection *os, uint64_t s) { |
284 | if (checkSecRel(sec, os)) |
285 | applyArm64Imm(off, imm: (s - os->getRVA()) & 0xfff, rangeLimit: 0); |
286 | } |
287 | |
288 | static void applySecRelHigh12A(const SectionChunk *sec, uint8_t *off, |
289 | OutputSection *os, uint64_t s) { |
290 | if (!checkSecRel(sec, os)) |
291 | return; |
292 | uint64_t secRel = (s - os->getRVA()) >> 12; |
293 | if (0xfff < secRel) { |
294 | error(msg: "overflow in SECREL_HIGH12A relocation in section: " + |
295 | sec->getSectionName()); |
296 | return; |
297 | } |
298 | applyArm64Imm(off, imm: secRel & 0xfff, rangeLimit: 0); |
299 | } |
300 | |
301 | static void applySecRelLdr(const SectionChunk *sec, uint8_t *off, |
302 | OutputSection *os, uint64_t s) { |
303 | if (checkSecRel(sec, os)) |
304 | applyArm64Ldr(off, imm: (s - os->getRVA()) & 0xfff); |
305 | } |
306 | |
307 | void applyArm64Branch26(uint8_t *off, int64_t v) { |
308 | if (!isInt<28>(x: v)) |
309 | error(msg: "relocation out of range" ); |
310 | or32(p: off, v: (v & 0x0FFFFFFC) >> 2); |
311 | } |
312 | |
313 | static void applyArm64Branch19(uint8_t *off, int64_t v) { |
314 | if (!isInt<21>(x: v)) |
315 | error(msg: "relocation out of range" ); |
316 | or32(p: off, v: (v & 0x001FFFFC) << 3); |
317 | } |
318 | |
319 | static void applyArm64Branch14(uint8_t *off, int64_t v) { |
320 | if (!isInt<16>(x: v)) |
321 | error(msg: "relocation out of range" ); |
322 | or32(p: off, v: (v & 0x0000FFFC) << 3); |
323 | } |
324 | |
325 | void SectionChunk::applyRelARM64(uint8_t *off, uint16_t type, OutputSection *os, |
326 | uint64_t s, uint64_t p, |
327 | uint64_t imageBase) const { |
328 | switch (type) { |
329 | case IMAGE_REL_ARM64_PAGEBASE_REL21: applyArm64Addr(off, s, p, shift: 12); break; |
330 | case IMAGE_REL_ARM64_REL21: applyArm64Addr(off, s, p, shift: 0); break; |
331 | case IMAGE_REL_ARM64_PAGEOFFSET_12A: applyArm64Imm(off, imm: s & 0xfff, rangeLimit: 0); break; |
332 | case IMAGE_REL_ARM64_PAGEOFFSET_12L: applyArm64Ldr(off, imm: s & 0xfff); break; |
333 | case IMAGE_REL_ARM64_BRANCH26: applyArm64Branch26(off, v: s - p); break; |
334 | case IMAGE_REL_ARM64_BRANCH19: applyArm64Branch19(off, v: s - p); break; |
335 | case IMAGE_REL_ARM64_BRANCH14: applyArm64Branch14(off, v: s - p); break; |
336 | case IMAGE_REL_ARM64_ADDR32: |
337 | add32(p: off, v: s + imageBase); |
338 | break; |
339 | case IMAGE_REL_ARM64_ADDR32NB: add32(p: off, v: s); break; |
340 | case IMAGE_REL_ARM64_ADDR64: |
341 | add64(p: off, v: s + imageBase); |
342 | break; |
343 | case IMAGE_REL_ARM64_SECREL: applySecRel(sec: this, off, os, s); break; |
344 | case IMAGE_REL_ARM64_SECREL_LOW12A: applySecRelLow12A(sec: this, off, os, s); break; |
345 | case IMAGE_REL_ARM64_SECREL_HIGH12A: applySecRelHigh12A(sec: this, off, os, s); break; |
346 | case IMAGE_REL_ARM64_SECREL_LOW12L: applySecRelLdr(sec: this, off, os, s); break; |
347 | case IMAGE_REL_ARM64_SECTION: |
348 | applySecIdx(off, os, numOutputSections: file->ctx.outputSections.size()); |
349 | break; |
350 | case IMAGE_REL_ARM64_REL32: add32(p: off, v: s - p - 4); break; |
351 | default: |
352 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
353 | toString(file)); |
354 | } |
355 | } |
356 | |
357 | static void maybeReportRelocationToDiscarded(const SectionChunk *fromChunk, |
358 | Defined *sym, |
359 | const coff_relocation &rel, |
360 | bool isMinGW) { |
361 | // Don't report these errors when the relocation comes from a debug info |
362 | // section or in mingw mode. MinGW mode object files (built by GCC) can |
363 | // have leftover sections with relocations against discarded comdat |
364 | // sections. Such sections are left as is, with relocations untouched. |
365 | if (fromChunk->isCodeView() || fromChunk->isDWARF() || isMinGW) |
366 | return; |
367 | |
368 | // Get the name of the symbol. If it's null, it was discarded early, so we |
369 | // have to go back to the object file. |
370 | ObjFile *file = fromChunk->file; |
371 | StringRef name; |
372 | if (sym) { |
373 | name = sym->getName(); |
374 | } else { |
375 | COFFSymbolRef coffSym = |
376 | check(e: file->getCOFFObj()->getSymbol(index: rel.SymbolTableIndex)); |
377 | name = check(e: file->getCOFFObj()->getSymbolName(Symbol: coffSym)); |
378 | } |
379 | |
380 | std::vector<std::string> symbolLocations = |
381 | getSymbolLocations(file, symIndex: rel.SymbolTableIndex); |
382 | |
383 | std::string out; |
384 | llvm::raw_string_ostream os(out); |
385 | os << "relocation against symbol in discarded section: " + name; |
386 | for (const std::string &s : symbolLocations) |
387 | os << s; |
388 | error(msg: os.str()); |
389 | } |
390 | |
391 | void SectionChunk::writeTo(uint8_t *buf) const { |
392 | if (!hasData) |
393 | return; |
394 | // Copy section contents from source object file to output file. |
395 | ArrayRef<uint8_t> a = getContents(); |
396 | if (!a.empty()) |
397 | memcpy(dest: buf, src: a.data(), n: a.size()); |
398 | |
399 | // Apply relocations. |
400 | size_t inputSize = getSize(); |
401 | for (const coff_relocation &rel : getRelocs()) { |
402 | // Check for an invalid relocation offset. This check isn't perfect, because |
403 | // we don't have the relocation size, which is only known after checking the |
404 | // machine and relocation type. As a result, a relocation may overwrite the |
405 | // beginning of the following input section. |
406 | if (rel.VirtualAddress >= inputSize) { |
407 | error(msg: "relocation points beyond the end of its parent section" ); |
408 | continue; |
409 | } |
410 | |
411 | applyRelocation(off: buf + rel.VirtualAddress, rel); |
412 | } |
413 | } |
414 | |
415 | void SectionChunk::applyRelocation(uint8_t *off, |
416 | const coff_relocation &rel) const { |
417 | auto *sym = dyn_cast_or_null<Defined>(Val: file->getSymbol(symbolIndex: rel.SymbolTableIndex)); |
418 | |
419 | // Get the output section of the symbol for this relocation. The output |
420 | // section is needed to compute SECREL and SECTION relocations used in debug |
421 | // info. |
422 | Chunk *c = sym ? sym->getChunk() : nullptr; |
423 | OutputSection *os = c ? file->ctx.getOutputSection(c) : nullptr; |
424 | |
425 | // Skip the relocation if it refers to a discarded section, and diagnose it |
426 | // as an error if appropriate. If a symbol was discarded early, it may be |
427 | // null. If it was discarded late, the output section will be null, unless |
428 | // it was an absolute or synthetic symbol. |
429 | if (!sym || |
430 | (!os && !isa<DefinedAbsolute>(Val: sym) && !isa<DefinedSynthetic>(Val: sym))) { |
431 | maybeReportRelocationToDiscarded(fromChunk: this, sym, rel, isMinGW: file->ctx.config.mingw); |
432 | return; |
433 | } |
434 | |
435 | uint64_t s = sym->getRVA(); |
436 | |
437 | // Compute the RVA of the relocation for relative relocations. |
438 | uint64_t p = rva + rel.VirtualAddress; |
439 | uint64_t imageBase = file->ctx.config.imageBase; |
440 | switch (getArch()) { |
441 | case Triple::x86_64: |
442 | applyRelX64(off, type: rel.Type, os, s, p, imageBase); |
443 | break; |
444 | case Triple::x86: |
445 | applyRelX86(off, type: rel.Type, os, s, p, imageBase); |
446 | break; |
447 | case Triple::thumb: |
448 | applyRelARM(off, type: rel.Type, os, s, p, imageBase); |
449 | break; |
450 | case Triple::aarch64: |
451 | applyRelARM64(off, type: rel.Type, os, s, p, imageBase); |
452 | break; |
453 | default: |
454 | llvm_unreachable("unknown machine type" ); |
455 | } |
456 | } |
457 | |
458 | // Defend against unsorted relocations. This may be overly conservative. |
459 | void SectionChunk::sortRelocations() { |
460 | auto cmpByVa = [](const coff_relocation &l, const coff_relocation &r) { |
461 | return l.VirtualAddress < r.VirtualAddress; |
462 | }; |
463 | if (llvm::is_sorted(Range: getRelocs(), C: cmpByVa)) |
464 | return; |
465 | warn(msg: "some relocations in " + file->getName() + " are not sorted" ); |
466 | MutableArrayRef<coff_relocation> newRelocs( |
467 | bAlloc().Allocate<coff_relocation>(Num: relocsSize), relocsSize); |
468 | memcpy(dest: newRelocs.data(), src: relocsData, n: relocsSize * sizeof(coff_relocation)); |
469 | llvm::sort(C&: newRelocs, Comp: cmpByVa); |
470 | setRelocs(newRelocs); |
471 | } |
472 | |
473 | // Similar to writeTo, but suitable for relocating a subsection of the overall |
474 | // section. |
475 | void SectionChunk::writeAndRelocateSubsection(ArrayRef<uint8_t> sec, |
476 | ArrayRef<uint8_t> subsec, |
477 | uint32_t &nextRelocIndex, |
478 | uint8_t *buf) const { |
479 | assert(!subsec.empty() && !sec.empty()); |
480 | assert(sec.begin() <= subsec.begin() && subsec.end() <= sec.end() && |
481 | "subsection is not part of this section" ); |
482 | size_t vaBegin = std::distance(first: sec.begin(), last: subsec.begin()); |
483 | size_t vaEnd = std::distance(first: sec.begin(), last: subsec.end()); |
484 | memcpy(dest: buf, src: subsec.data(), n: subsec.size()); |
485 | for (; nextRelocIndex < relocsSize; ++nextRelocIndex) { |
486 | const coff_relocation &rel = relocsData[nextRelocIndex]; |
487 | // Only apply relocations that apply to this subsection. These checks |
488 | // assume that all subsections completely contain their relocations. |
489 | // Relocations must not straddle the beginning or end of a subsection. |
490 | if (rel.VirtualAddress < vaBegin) |
491 | continue; |
492 | if (rel.VirtualAddress + 1 >= vaEnd) |
493 | break; |
494 | applyRelocation(off: &buf[rel.VirtualAddress - vaBegin], rel); |
495 | } |
496 | } |
497 | |
498 | void SectionChunk::addAssociative(SectionChunk *child) { |
499 | // Insert the child section into the list of associated children. Keep the |
500 | // list ordered by section name so that ICF does not depend on section order. |
501 | assert(child->assocChildren == nullptr && |
502 | "associated sections cannot have their own associated children" ); |
503 | SectionChunk *prev = this; |
504 | SectionChunk *next = assocChildren; |
505 | for (; next != nullptr; prev = next, next = next->assocChildren) { |
506 | if (next->getSectionName() <= child->getSectionName()) |
507 | break; |
508 | } |
509 | |
510 | // Insert child between prev and next. |
511 | assert(prev->assocChildren == next); |
512 | prev->assocChildren = child; |
513 | child->assocChildren = next; |
514 | } |
515 | |
516 | static uint8_t getBaserelType(const coff_relocation &rel, |
517 | Triple::ArchType arch) { |
518 | switch (arch) { |
519 | case Triple::x86_64: |
520 | if (rel.Type == IMAGE_REL_AMD64_ADDR64) |
521 | return IMAGE_REL_BASED_DIR64; |
522 | if (rel.Type == IMAGE_REL_AMD64_ADDR32) |
523 | return IMAGE_REL_BASED_HIGHLOW; |
524 | return IMAGE_REL_BASED_ABSOLUTE; |
525 | case Triple::x86: |
526 | if (rel.Type == IMAGE_REL_I386_DIR32) |
527 | return IMAGE_REL_BASED_HIGHLOW; |
528 | return IMAGE_REL_BASED_ABSOLUTE; |
529 | case Triple::thumb: |
530 | if (rel.Type == IMAGE_REL_ARM_ADDR32) |
531 | return IMAGE_REL_BASED_HIGHLOW; |
532 | if (rel.Type == IMAGE_REL_ARM_MOV32T) |
533 | return IMAGE_REL_BASED_ARM_MOV32T; |
534 | return IMAGE_REL_BASED_ABSOLUTE; |
535 | case Triple::aarch64: |
536 | if (rel.Type == IMAGE_REL_ARM64_ADDR64) |
537 | return IMAGE_REL_BASED_DIR64; |
538 | return IMAGE_REL_BASED_ABSOLUTE; |
539 | default: |
540 | llvm_unreachable("unknown machine type" ); |
541 | } |
542 | } |
543 | |
544 | // Windows-specific. |
545 | // Collect all locations that contain absolute addresses, which need to be |
546 | // fixed by the loader if load-time relocation is needed. |
547 | // Only called when base relocation is enabled. |
548 | void SectionChunk::getBaserels(std::vector<Baserel> *res) { |
549 | for (const coff_relocation &rel : getRelocs()) { |
550 | uint8_t ty = getBaserelType(rel, arch: getArch()); |
551 | if (ty == IMAGE_REL_BASED_ABSOLUTE) |
552 | continue; |
553 | Symbol *target = file->getSymbol(symbolIndex: rel.SymbolTableIndex); |
554 | if (!target || isa<DefinedAbsolute>(Val: target)) |
555 | continue; |
556 | res->emplace_back(args: rva + rel.VirtualAddress, args&: ty); |
557 | } |
558 | } |
559 | |
560 | // MinGW specific. |
561 | // Check whether a static relocation of type Type can be deferred and |
562 | // handled at runtime as a pseudo relocation (for references to a module |
563 | // local variable, which turned out to actually need to be imported from |
564 | // another DLL) This returns the size the relocation is supposed to update, |
565 | // in bits, or 0 if the relocation cannot be handled as a runtime pseudo |
566 | // relocation. |
567 | static int getRuntimePseudoRelocSize(uint16_t type, |
568 | llvm::COFF::MachineTypes machine) { |
569 | // Relocations that either contain an absolute address, or a plain |
570 | // relative offset, since the runtime pseudo reloc implementation |
571 | // adds 8/16/32/64 bit values to a memory address. |
572 | // |
573 | // Given a pseudo relocation entry, |
574 | // |
575 | // typedef struct { |
576 | // DWORD sym; |
577 | // DWORD target; |
578 | // DWORD flags; |
579 | // } runtime_pseudo_reloc_item_v2; |
580 | // |
581 | // the runtime relocation performs this adjustment: |
582 | // *(base + .target) += *(base + .sym) - (base + .sym) |
583 | // |
584 | // This works for both absolute addresses (IMAGE_REL_*_ADDR32/64, |
585 | // IMAGE_REL_I386_DIR32, where the memory location initially contains |
586 | // the address of the IAT slot, and for relative addresses (IMAGE_REL*_REL32), |
587 | // where the memory location originally contains the relative offset to the |
588 | // IAT slot. |
589 | // |
590 | // This requires the target address to be writable, either directly out of |
591 | // the image, or temporarily changed at runtime with VirtualProtect. |
592 | // Since this only operates on direct address values, it doesn't work for |
593 | // ARM/ARM64 relocations, other than the plain ADDR32/ADDR64 relocations. |
594 | switch (machine) { |
595 | case AMD64: |
596 | switch (type) { |
597 | case IMAGE_REL_AMD64_ADDR64: |
598 | return 64; |
599 | case IMAGE_REL_AMD64_ADDR32: |
600 | case IMAGE_REL_AMD64_REL32: |
601 | case IMAGE_REL_AMD64_REL32_1: |
602 | case IMAGE_REL_AMD64_REL32_2: |
603 | case IMAGE_REL_AMD64_REL32_3: |
604 | case IMAGE_REL_AMD64_REL32_4: |
605 | case IMAGE_REL_AMD64_REL32_5: |
606 | return 32; |
607 | default: |
608 | return 0; |
609 | } |
610 | case I386: |
611 | switch (type) { |
612 | case IMAGE_REL_I386_DIR32: |
613 | case IMAGE_REL_I386_REL32: |
614 | return 32; |
615 | default: |
616 | return 0; |
617 | } |
618 | case ARMNT: |
619 | switch (type) { |
620 | case IMAGE_REL_ARM_ADDR32: |
621 | return 32; |
622 | default: |
623 | return 0; |
624 | } |
625 | case ARM64: |
626 | switch (type) { |
627 | case IMAGE_REL_ARM64_ADDR64: |
628 | return 64; |
629 | case IMAGE_REL_ARM64_ADDR32: |
630 | return 32; |
631 | default: |
632 | return 0; |
633 | } |
634 | default: |
635 | llvm_unreachable("unknown machine type" ); |
636 | } |
637 | } |
638 | |
639 | // MinGW specific. |
640 | // Append information to the provided vector about all relocations that |
641 | // need to be handled at runtime as runtime pseudo relocations (references |
642 | // to a module local variable, which turned out to actually need to be |
643 | // imported from another DLL). |
644 | void SectionChunk::getRuntimePseudoRelocs( |
645 | std::vector<RuntimePseudoReloc> &res) { |
646 | for (const coff_relocation &rel : getRelocs()) { |
647 | auto *target = |
648 | dyn_cast_or_null<Defined>(Val: file->getSymbol(symbolIndex: rel.SymbolTableIndex)); |
649 | if (!target || !target->isRuntimePseudoReloc) |
650 | continue; |
651 | // If the target doesn't have a chunk allocated, it may be a |
652 | // DefinedImportData symbol which ended up unnecessary after GC. |
653 | // Normally we wouldn't eliminate section chunks that are referenced, but |
654 | // references within DWARF sections don't count for keeping section chunks |
655 | // alive. Thus such dangling references in DWARF sections are expected. |
656 | if (!target->getChunk()) |
657 | continue; |
658 | int sizeInBits = |
659 | getRuntimePseudoRelocSize(type: rel.Type, machine: file->ctx.config.machine); |
660 | if (sizeInBits == 0) { |
661 | error(msg: "unable to automatically import from " + target->getName() + |
662 | " with relocation type " + |
663 | file->getCOFFObj()->getRelocationTypeName(Type: rel.Type) + " in " + |
664 | toString(file)); |
665 | continue; |
666 | } |
667 | int addressSizeInBits = file->ctx.config.is64() ? 64 : 32; |
668 | if (sizeInBits < addressSizeInBits) { |
669 | warn(msg: "runtime pseudo relocation in " + toString(file) + " against " + |
670 | "symbol " + target->getName() + " is too narrow (only " + |
671 | Twine(sizeInBits) + " bits wide); this can fail at runtime " + |
672 | "depending on memory layout" ); |
673 | } |
674 | // sizeInBits is used to initialize the Flags field; currently no |
675 | // other flags are defined. |
676 | res.emplace_back(args&: target, args: this, args: rel.VirtualAddress, args&: sizeInBits); |
677 | } |
678 | } |
679 | |
680 | bool SectionChunk::isCOMDAT() const { |
681 | return header->Characteristics & IMAGE_SCN_LNK_COMDAT; |
682 | } |
683 | |
684 | void SectionChunk::printDiscardedMessage() const { |
685 | // Removed by dead-stripping. If it's removed by ICF, ICF already |
686 | // printed out the name, so don't repeat that here. |
687 | if (sym && this == repl) |
688 | log(msg: "Discarded " + sym->getName()); |
689 | } |
690 | |
691 | StringRef SectionChunk::getDebugName() const { |
692 | if (sym) |
693 | return sym->getName(); |
694 | return "" ; |
695 | } |
696 | |
697 | ArrayRef<uint8_t> SectionChunk::getContents() const { |
698 | ArrayRef<uint8_t> a; |
699 | cantFail(Err: file->getCOFFObj()->getSectionContents(Sec: header, Res&: a)); |
700 | return a; |
701 | } |
702 | |
703 | ArrayRef<uint8_t> SectionChunk::consumeDebugMagic() { |
704 | assert(isCodeView()); |
705 | return consumeDebugMagic(data: getContents(), sectionName: getSectionName()); |
706 | } |
707 | |
708 | ArrayRef<uint8_t> SectionChunk::consumeDebugMagic(ArrayRef<uint8_t> data, |
709 | StringRef sectionName) { |
710 | if (data.empty()) |
711 | return {}; |
712 | |
713 | // First 4 bytes are section magic. |
714 | if (data.size() < 4) |
715 | fatal(msg: "the section is too short: " + sectionName); |
716 | |
717 | if (!sectionName.starts_with(Prefix: ".debug$" )) |
718 | fatal(msg: "invalid section: " + sectionName); |
719 | |
720 | uint32_t magic = support::endian::read32le(P: data.data()); |
721 | uint32_t expectedMagic = sectionName == ".debug$H" |
722 | ? DEBUG_HASHES_SECTION_MAGIC |
723 | : DEBUG_SECTION_MAGIC; |
724 | if (magic != expectedMagic) { |
725 | warn(msg: "ignoring section " + sectionName + " with unrecognized magic 0x" + |
726 | utohexstr(X: magic)); |
727 | return {}; |
728 | } |
729 | return data.slice(N: 4); |
730 | } |
731 | |
732 | SectionChunk *SectionChunk::findByName(ArrayRef<SectionChunk *> sections, |
733 | StringRef name) { |
734 | for (SectionChunk *c : sections) |
735 | if (c->getSectionName() == name) |
736 | return c; |
737 | return nullptr; |
738 | } |
739 | |
740 | void SectionChunk::replace(SectionChunk *other) { |
741 | p2Align = std::max(a: p2Align, b: other->p2Align); |
742 | other->repl = repl; |
743 | other->live = false; |
744 | } |
745 | |
746 | uint32_t SectionChunk::getSectionNumber() const { |
747 | DataRefImpl r; |
748 | r.p = reinterpret_cast<uintptr_t>(header); |
749 | SectionRef s(r, file->getCOFFObj()); |
750 | return s.getIndex() + 1; |
751 | } |
752 | |
753 | CommonChunk::CommonChunk(const COFFSymbolRef s) : sym(s) { |
754 | // The value of a common symbol is its size. Align all common symbols smaller |
755 | // than 32 bytes naturally, i.e. round the size up to the next power of two. |
756 | // This is what MSVC link.exe does. |
757 | setAlignment(std::min(a: 32U, b: uint32_t(PowerOf2Ceil(A: sym.getValue())))); |
758 | hasData = false; |
759 | } |
760 | |
761 | uint32_t CommonChunk::getOutputCharacteristics() const { |
762 | return IMAGE_SCN_CNT_UNINITIALIZED_DATA | IMAGE_SCN_MEM_READ | |
763 | IMAGE_SCN_MEM_WRITE; |
764 | } |
765 | |
766 | void StringChunk::writeTo(uint8_t *buf) const { |
767 | memcpy(dest: buf, src: str.data(), n: str.size()); |
768 | buf[str.size()] = '\0'; |
769 | } |
770 | |
771 | ImportThunkChunkX64::ImportThunkChunkX64(COFFLinkerContext &ctx, Defined *s) |
772 | : ImportThunkChunk(ctx, s) { |
773 | // Intel Optimization Manual says that all branch targets |
774 | // should be 16-byte aligned. MSVC linker does this too. |
775 | setAlignment(16); |
776 | } |
777 | |
778 | void ImportThunkChunkX64::writeTo(uint8_t *buf) const { |
779 | memcpy(dest: buf, src: importThunkX86, n: sizeof(importThunkX86)); |
780 | // The first two bytes is a JMP instruction. Fill its operand. |
781 | write32le(P: buf + 2, V: impSymbol->getRVA() - rva - getSize()); |
782 | } |
783 | |
784 | void ImportThunkChunkX86::getBaserels(std::vector<Baserel> *res) { |
785 | res->emplace_back(args: getRVA() + 2, args&: ctx.config.machine); |
786 | } |
787 | |
788 | void ImportThunkChunkX86::writeTo(uint8_t *buf) const { |
789 | memcpy(dest: buf, src: importThunkX86, n: sizeof(importThunkX86)); |
790 | // The first two bytes is a JMP instruction. Fill its operand. |
791 | write32le(P: buf + 2, V: impSymbol->getRVA() + ctx.config.imageBase); |
792 | } |
793 | |
794 | void ImportThunkChunkARM::getBaserels(std::vector<Baserel> *res) { |
795 | res->emplace_back(args: getRVA(), args: IMAGE_REL_BASED_ARM_MOV32T); |
796 | } |
797 | |
798 | void ImportThunkChunkARM::writeTo(uint8_t *buf) const { |
799 | memcpy(dest: buf, src: importThunkARM, n: sizeof(importThunkARM)); |
800 | // Fix mov.w and mov.t operands. |
801 | applyMOV32T(off: buf, v: impSymbol->getRVA() + ctx.config.imageBase); |
802 | } |
803 | |
804 | void ImportThunkChunkARM64::writeTo(uint8_t *buf) const { |
805 | int64_t off = impSymbol->getRVA() & 0xfff; |
806 | memcpy(dest: buf, src: importThunkARM64, n: sizeof(importThunkARM64)); |
807 | applyArm64Addr(off: buf, s: impSymbol->getRVA(), p: rva, shift: 12); |
808 | applyArm64Ldr(off: buf + 4, imm: off); |
809 | } |
810 | |
811 | // A Thumb2, PIC, non-interworking range extension thunk. |
812 | const uint8_t armThunk[] = { |
813 | 0x40, 0xf2, 0x00, 0x0c, // P: movw ip,:lower16:S - (P + (L1-P) + 4) |
814 | 0xc0, 0xf2, 0x00, 0x0c, // movt ip,:upper16:S - (P + (L1-P) + 4) |
815 | 0xe7, 0x44, // L1: add pc, ip |
816 | }; |
817 | |
818 | size_t RangeExtensionThunkARM::getSize() const { |
819 | assert(ctx.config.machine == ARMNT); |
820 | (void)&ctx; |
821 | return sizeof(armThunk); |
822 | } |
823 | |
824 | void RangeExtensionThunkARM::writeTo(uint8_t *buf) const { |
825 | assert(ctx.config.machine == ARMNT); |
826 | uint64_t offset = target->getRVA() - rva - 12; |
827 | memcpy(dest: buf, src: armThunk, n: sizeof(armThunk)); |
828 | applyMOV32T(off: buf, v: uint32_t(offset)); |
829 | } |
830 | |
831 | // A position independent ARM64 adrp+add thunk, with a maximum range of |
832 | // +/- 4 GB, which is enough for any PE-COFF. |
833 | const uint8_t arm64Thunk[] = { |
834 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Dest |
835 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, :lo12:Dest |
836 | 0x00, 0x02, 0x1f, 0xd6, // br x16 |
837 | }; |
838 | |
839 | size_t RangeExtensionThunkARM64::getSize() const { |
840 | assert(ctx.config.machine == ARM64); |
841 | (void)&ctx; |
842 | return sizeof(arm64Thunk); |
843 | } |
844 | |
845 | void RangeExtensionThunkARM64::writeTo(uint8_t *buf) const { |
846 | assert(ctx.config.machine == ARM64); |
847 | memcpy(dest: buf, src: arm64Thunk, n: sizeof(arm64Thunk)); |
848 | applyArm64Addr(off: buf + 0, s: target->getRVA(), p: rva, shift: 12); |
849 | applyArm64Imm(off: buf + 4, imm: target->getRVA() & 0xfff, rangeLimit: 0); |
850 | } |
851 | |
852 | LocalImportChunk::LocalImportChunk(COFFLinkerContext &c, Defined *s) |
853 | : sym(s), ctx(c) { |
854 | setAlignment(ctx.config.wordsize); |
855 | } |
856 | |
857 | void LocalImportChunk::getBaserels(std::vector<Baserel> *res) { |
858 | res->emplace_back(args: getRVA(), args&: ctx.config.machine); |
859 | } |
860 | |
861 | size_t LocalImportChunk::getSize() const { return ctx.config.wordsize; } |
862 | |
863 | void LocalImportChunk::writeTo(uint8_t *buf) const { |
864 | if (ctx.config.is64()) { |
865 | write64le(P: buf, V: sym->getRVA() + ctx.config.imageBase); |
866 | } else { |
867 | write32le(P: buf, V: sym->getRVA() + ctx.config.imageBase); |
868 | } |
869 | } |
870 | |
871 | void RVATableChunk::writeTo(uint8_t *buf) const { |
872 | ulittle32_t *begin = reinterpret_cast<ulittle32_t *>(buf); |
873 | size_t cnt = 0; |
874 | for (const ChunkAndOffset &co : syms) |
875 | begin[cnt++] = co.inputChunk->getRVA() + co.offset; |
876 | llvm::sort(Start: begin, End: begin + cnt); |
877 | assert(std::unique(begin, begin + cnt) == begin + cnt && |
878 | "RVA tables should be de-duplicated" ); |
879 | } |
880 | |
881 | void RVAFlagTableChunk::writeTo(uint8_t *buf) const { |
882 | struct RVAFlag { |
883 | ulittle32_t rva; |
884 | uint8_t flag; |
885 | }; |
886 | auto flags = |
887 | MutableArrayRef(reinterpret_cast<RVAFlag *>(buf), syms.size()); |
888 | for (auto t : zip(t: syms, u&: flags)) { |
889 | const auto &sym = std::get<0>(t&: t); |
890 | auto &flag = std::get<1>(t&: t); |
891 | flag.rva = sym.inputChunk->getRVA() + sym.offset; |
892 | flag.flag = 0; |
893 | } |
894 | llvm::sort(C&: flags, |
895 | Comp: [](const RVAFlag &a, const RVAFlag &b) { return a.rva < b.rva; }); |
896 | assert(llvm::unique(flags, [](const RVAFlag &a, |
897 | const RVAFlag &b) { return a.rva == b.rva; }) == |
898 | flags.end() && |
899 | "RVA tables should be de-duplicated" ); |
900 | } |
901 | |
902 | size_t ECCodeMapChunk::getSize() const { |
903 | return map.size() * sizeof(chpe_range_entry); |
904 | } |
905 | |
906 | void ECCodeMapChunk::writeTo(uint8_t *buf) const { |
907 | auto table = reinterpret_cast<chpe_range_entry *>(buf); |
908 | for (uint32_t i = 0; i < map.size(); i++) { |
909 | const ECCodeMapEntry &entry = map[i]; |
910 | uint32_t start = entry.first->getRVA(); |
911 | table[i].StartOffset = start | entry.type; |
912 | table[i].Length = entry.last->getRVA() + entry.last->getSize() - start; |
913 | } |
914 | } |
915 | |
916 | // MinGW specific, for the "automatic import of variables from DLLs" feature. |
917 | size_t PseudoRelocTableChunk::getSize() const { |
918 | if (relocs.empty()) |
919 | return 0; |
920 | return 12 + 12 * relocs.size(); |
921 | } |
922 | |
923 | // MinGW specific. |
924 | void PseudoRelocTableChunk::writeTo(uint8_t *buf) const { |
925 | if (relocs.empty()) |
926 | return; |
927 | |
928 | ulittle32_t *table = reinterpret_cast<ulittle32_t *>(buf); |
929 | // This is the list header, to signal the runtime pseudo relocation v2 |
930 | // format. |
931 | table[0] = 0; |
932 | table[1] = 0; |
933 | table[2] = 1; |
934 | |
935 | size_t idx = 3; |
936 | for (const RuntimePseudoReloc &rpr : relocs) { |
937 | table[idx + 0] = rpr.sym->getRVA(); |
938 | table[idx + 1] = rpr.target->getRVA() + rpr.targetOffset; |
939 | table[idx + 2] = rpr.flags; |
940 | idx += 3; |
941 | } |
942 | } |
943 | |
944 | // Windows-specific. This class represents a block in .reloc section. |
945 | // The format is described here. |
946 | // |
947 | // On Windows, each DLL is linked against a fixed base address and |
948 | // usually loaded to that address. However, if there's already another |
949 | // DLL that overlaps, the loader has to relocate it. To do that, DLLs |
950 | // contain .reloc sections which contain offsets that need to be fixed |
951 | // up at runtime. If the loader finds that a DLL cannot be loaded to its |
952 | // desired base address, it loads it to somewhere else, and add <actual |
953 | // base address> - <desired base address> to each offset that is |
954 | // specified by the .reloc section. In ELF terms, .reloc sections |
955 | // contain relative relocations in REL format (as opposed to RELA.) |
956 | // |
957 | // This already significantly reduces the size of relocations compared |
958 | // to ELF .rel.dyn, but Windows does more to reduce it (probably because |
959 | // it was invented for PCs in the late '80s or early '90s.) Offsets in |
960 | // .reloc are grouped by page where the page size is 12 bits, and |
961 | // offsets sharing the same page address are stored consecutively to |
962 | // represent them with less space. This is very similar to the page |
963 | // table which is grouped by (multiple stages of) pages. |
964 | // |
965 | // For example, let's say we have 0x00030, 0x00500, 0x00700, 0x00A00, |
966 | // 0x20004, and 0x20008 in a .reloc section for x64. The uppermost 4 |
967 | // bits have a type IMAGE_REL_BASED_DIR64 or 0xA. In the section, they |
968 | // are represented like this: |
969 | // |
970 | // 0x00000 -- page address (4 bytes) |
971 | // 16 -- size of this block (4 bytes) |
972 | // 0xA030 -- entries (2 bytes each) |
973 | // 0xA500 |
974 | // 0xA700 |
975 | // 0xAA00 |
976 | // 0x20000 -- page address (4 bytes) |
977 | // 12 -- size of this block (4 bytes) |
978 | // 0xA004 -- entries (2 bytes each) |
979 | // 0xA008 |
980 | // |
981 | // Usually we have a lot of relocations for each page, so the number of |
982 | // bytes for one .reloc entry is close to 2 bytes on average. |
983 | BaserelChunk::BaserelChunk(uint32_t page, Baserel *begin, Baserel *end) { |
984 | // Block header consists of 4 byte page RVA and 4 byte block size. |
985 | // Each entry is 2 byte. Last entry may be padding. |
986 | data.resize(new_size: alignTo(Value: (end - begin) * 2 + 8, Align: 4)); |
987 | uint8_t *p = data.data(); |
988 | write32le(P: p, V: page); |
989 | write32le(P: p + 4, V: data.size()); |
990 | p += 8; |
991 | for (Baserel *i = begin; i != end; ++i) { |
992 | write16le(P: p, V: (i->type << 12) | (i->rva - page)); |
993 | p += 2; |
994 | } |
995 | } |
996 | |
997 | void BaserelChunk::writeTo(uint8_t *buf) const { |
998 | memcpy(dest: buf, src: data.data(), n: data.size()); |
999 | } |
1000 | |
1001 | uint8_t Baserel::getDefaultType(llvm::COFF::MachineTypes machine) { |
1002 | switch (machine) { |
1003 | case AMD64: |
1004 | case ARM64: |
1005 | return IMAGE_REL_BASED_DIR64; |
1006 | case I386: |
1007 | case ARMNT: |
1008 | return IMAGE_REL_BASED_HIGHLOW; |
1009 | default: |
1010 | llvm_unreachable("unknown machine type" ); |
1011 | } |
1012 | } |
1013 | |
1014 | MergeChunk::MergeChunk(uint32_t alignment) |
1015 | : builder(StringTableBuilder::RAW, llvm::Align(alignment)) { |
1016 | setAlignment(alignment); |
1017 | } |
1018 | |
1019 | void MergeChunk::addSection(COFFLinkerContext &ctx, SectionChunk *c) { |
1020 | assert(isPowerOf2_32(c->getAlignment())); |
1021 | uint8_t p2Align = llvm::Log2_32(Value: c->getAlignment()); |
1022 | assert(p2Align < std::size(ctx.mergeChunkInstances)); |
1023 | auto *&mc = ctx.mergeChunkInstances[p2Align]; |
1024 | if (!mc) |
1025 | mc = make<MergeChunk>(args: c->getAlignment()); |
1026 | mc->sections.push_back(x: c); |
1027 | } |
1028 | |
1029 | void MergeChunk::finalizeContents() { |
1030 | assert(!finalized && "should only finalize once" ); |
1031 | for (SectionChunk *c : sections) |
1032 | if (c->live) |
1033 | builder.add(S: toStringRef(Input: c->getContents())); |
1034 | builder.finalize(); |
1035 | finalized = true; |
1036 | } |
1037 | |
1038 | void MergeChunk::assignSubsectionRVAs() { |
1039 | for (SectionChunk *c : sections) { |
1040 | if (!c->live) |
1041 | continue; |
1042 | size_t off = builder.getOffset(S: toStringRef(Input: c->getContents())); |
1043 | c->setRVA(rva + off); |
1044 | } |
1045 | } |
1046 | |
1047 | uint32_t MergeChunk::getOutputCharacteristics() const { |
1048 | return IMAGE_SCN_MEM_READ | IMAGE_SCN_CNT_INITIALIZED_DATA; |
1049 | } |
1050 | |
1051 | size_t MergeChunk::getSize() const { |
1052 | return builder.getSize(); |
1053 | } |
1054 | |
1055 | void MergeChunk::writeTo(uint8_t *buf) const { |
1056 | builder.write(Buf: buf); |
1057 | } |
1058 | |
1059 | // MinGW specific. |
1060 | size_t AbsolutePointerChunk::getSize() const { return ctx.config.wordsize; } |
1061 | |
1062 | void AbsolutePointerChunk::writeTo(uint8_t *buf) const { |
1063 | if (ctx.config.is64()) { |
1064 | write64le(P: buf, V: value); |
1065 | } else { |
1066 | write32le(P: buf, V: value); |
1067 | } |
1068 | } |
1069 | |
1070 | } // namespace lld::coff |
1071 | |