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; |
29 | using namespace llvm::support::endian; |
30 | using namespace llvm::COFF; |
31 | using llvm::support::ulittle32_t; |
32 | |
33 | namespace lld::coff { |
34 | |
35 | SectionChunk::SectionChunk(ObjFile *f, const coff_section *h, Kind k) |
36 | : Chunk(k), file(f), header(h), repl(this) { |
37 | // Initialize relocs. |
38 | if (file) |
39 | setRelocs(file->getCOFFObj()->getRelocations(Sec: header)); |
40 | |
41 | // Initialize sectionName. |
42 | StringRef sectionName; |
43 | if (file) { |
44 | if (Expected<StringRef> e = file->getCOFFObj()->getSectionName(Sec: header)) |
45 | sectionName = *e; |
46 | } |
47 | sectionNameData = sectionName.data(); |
48 | sectionNameSize = sectionName.size(); |
49 | |
50 | setAlignment(header->getAlignment()); |
51 | |
52 | hasData = !(header->Characteristics & IMAGE_SCN_CNT_UNINITIALIZED_DATA); |
53 | |
54 | // If linker GC is disabled, every chunk starts out alive. If linker GC is |
55 | // enabled, treat non-comdat sections as roots. Generally optimized object |
56 | // files will be built with -ffunction-sections or /Gy, so most things worth |
57 | // stripping will be in a comdat. |
58 | if (file) |
59 | live = !file->symtab.ctx.config.doGC || !isCOMDAT(); |
60 | else |
61 | live = true; |
62 | } |
63 | |
64 | MachineTypes SectionChunk::getMachine() const { |
65 | MachineTypes machine = file->getMachineType(); |
66 | // On ARM64EC, the IMAGE_SCN_GPREL flag is repurposed to indicate that section |
67 | // code is x86_64. This enables embedding x86_64 code within ARM64EC object |
68 | // files. MSVC uses this for export thunks in .exp files. |
69 | if (isArm64EC(Machine: machine) && (header->Characteristics & IMAGE_SCN_GPREL)) |
70 | machine = AMD64; |
71 | return machine; |
72 | } |
73 | |
74 | // SectionChunk is one of the most frequently allocated classes, so it is |
75 | // important to keep it as compact as possible. As of this writing, the number |
76 | // below is the size of this class on x64 platforms. |
77 | static_assert(sizeof(SectionChunk) <= 88, "SectionChunk grew unexpectedly" ); |
78 | |
79 | static void add16(uint8_t *p, int16_t v) { write16le(P: p, V: read16le(P: p) + v); } |
80 | static void add32(uint8_t *p, int32_t v) { write32le(P: p, V: read32le(P: p) + v); } |
81 | static void add64(uint8_t *p, int64_t v) { write64le(P: p, V: read64le(P: p) + v); } |
82 | static void or16(uint8_t *p, uint16_t v) { write16le(P: p, V: read16le(P: p) | v); } |
83 | static void or32(uint8_t *p, uint32_t v) { write32le(P: p, V: read32le(P: p) | v); } |
84 | |
85 | // Verify that given sections are appropriate targets for SECREL |
86 | // relocations. This check is relaxed because unfortunately debug |
87 | // sections have section-relative relocations against absolute symbols. |
88 | static bool checkSecRel(const SectionChunk *sec, OutputSection *os) { |
89 | if (os) |
90 | return true; |
91 | if (sec->isCodeView()) |
92 | return false; |
93 | error(msg: "SECREL relocation cannot be applied to absolute symbols" ); |
94 | return false; |
95 | } |
96 | |
97 | static void applySecRel(const SectionChunk *sec, uint8_t *off, |
98 | OutputSection *os, uint64_t s) { |
99 | if (!checkSecRel(sec, os)) |
100 | return; |
101 | uint64_t secRel = s - os->getRVA(); |
102 | if (secRel > UINT32_MAX) { |
103 | error(msg: "overflow in SECREL relocation in section: " + sec->getSectionName()); |
104 | return; |
105 | } |
106 | add32(p: off, v: secRel); |
107 | } |
108 | |
109 | static void applySecIdx(uint8_t *off, OutputSection *os, |
110 | unsigned numOutputSections) { |
111 | // numOutputSections is the largest valid section index. Make sure that |
112 | // it fits in 16 bits. |
113 | assert(numOutputSections <= 0xffff && "size of outputSections is too big" ); |
114 | |
115 | // Absolute symbol doesn't have section index, but section index relocation |
116 | // against absolute symbol should be resolved to one plus the last output |
117 | // section index. This is required for compatibility with MSVC. |
118 | if (os) |
119 | add16(p: off, v: os->sectionIndex); |
120 | else |
121 | add16(p: off, v: numOutputSections + 1); |
122 | } |
123 | |
124 | void SectionChunk::applyRelX64(uint8_t *off, uint16_t type, OutputSection *os, |
125 | uint64_t s, uint64_t p, |
126 | uint64_t imageBase) const { |
127 | switch (type) { |
128 | case IMAGE_REL_AMD64_ADDR32: |
129 | add32(p: off, v: s + imageBase); |
130 | break; |
131 | case IMAGE_REL_AMD64_ADDR64: |
132 | add64(p: off, v: s + imageBase); |
133 | break; |
134 | case IMAGE_REL_AMD64_ADDR32NB: add32(p: off, v: s); break; |
135 | case IMAGE_REL_AMD64_REL32: add32(p: off, v: s - p - 4); break; |
136 | case IMAGE_REL_AMD64_REL32_1: add32(p: off, v: s - p - 5); break; |
137 | case IMAGE_REL_AMD64_REL32_2: add32(p: off, v: s - p - 6); break; |
138 | case IMAGE_REL_AMD64_REL32_3: add32(p: off, v: s - p - 7); break; |
139 | case IMAGE_REL_AMD64_REL32_4: add32(p: off, v: s - p - 8); break; |
140 | case IMAGE_REL_AMD64_REL32_5: add32(p: off, v: s - p - 9); break; |
141 | case IMAGE_REL_AMD64_SECTION: |
142 | applySecIdx(off, os, numOutputSections: file->symtab.ctx.outputSections.size()); |
143 | break; |
144 | case IMAGE_REL_AMD64_SECREL: applySecRel(sec: this, off, os, s); break; |
145 | default: |
146 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
147 | toString(file)); |
148 | } |
149 | } |
150 | |
151 | void SectionChunk::applyRelX86(uint8_t *off, uint16_t type, OutputSection *os, |
152 | uint64_t s, uint64_t p, |
153 | uint64_t imageBase) const { |
154 | switch (type) { |
155 | case IMAGE_REL_I386_ABSOLUTE: break; |
156 | case IMAGE_REL_I386_DIR32: |
157 | add32(p: off, v: s + imageBase); |
158 | break; |
159 | case IMAGE_REL_I386_DIR32NB: add32(p: off, v: s); break; |
160 | case IMAGE_REL_I386_REL32: add32(p: off, v: s - p - 4); break; |
161 | case IMAGE_REL_I386_SECTION: |
162 | applySecIdx(off, os, numOutputSections: file->symtab.ctx.outputSections.size()); |
163 | break; |
164 | case IMAGE_REL_I386_SECREL: applySecRel(sec: this, off, os, s); break; |
165 | default: |
166 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
167 | toString(file)); |
168 | } |
169 | } |
170 | |
171 | static void applyMOV(uint8_t *off, uint16_t v) { |
172 | write16le(P: off, V: (read16le(P: off) & 0xfbf0) | ((v & 0x800) >> 1) | ((v >> 12) & 0xf)); |
173 | write16le(P: off + 2, V: (read16le(P: off + 2) & 0x8f00) | ((v & 0x700) << 4) | (v & 0xff)); |
174 | } |
175 | |
176 | static uint16_t readMOV(uint8_t *off, bool movt) { |
177 | uint16_t op1 = read16le(P: off); |
178 | if ((op1 & 0xfbf0) != (movt ? 0xf2c0 : 0xf240)) |
179 | error(msg: "unexpected instruction in " + Twine(movt ? "MOVT" : "MOVW" ) + |
180 | " instruction in MOV32T relocation" ); |
181 | uint16_t op2 = read16le(P: off + 2); |
182 | if ((op2 & 0x8000) != 0) |
183 | error(msg: "unexpected instruction in " + Twine(movt ? "MOVT" : "MOVW" ) + |
184 | " instruction in MOV32T relocation" ); |
185 | return (op2 & 0x00ff) | ((op2 >> 4) & 0x0700) | ((op1 << 1) & 0x0800) | |
186 | ((op1 & 0x000f) << 12); |
187 | } |
188 | |
189 | void applyMOV32T(uint8_t *off, uint32_t v) { |
190 | uint16_t immW = readMOV(off, movt: false); // read MOVW operand |
191 | uint16_t immT = readMOV(off: off + 4, movt: true); // read MOVT operand |
192 | uint32_t imm = immW | (immT << 16); |
193 | v += imm; // add the immediate offset |
194 | applyMOV(off, v); // set MOVW operand |
195 | applyMOV(off: off + 4, v: v >> 16); // set MOVT operand |
196 | } |
197 | |
198 | static void applyBranch20T(uint8_t *off, int32_t v) { |
199 | if (!isInt<21>(x: v)) |
200 | error(msg: "relocation out of range" ); |
201 | uint32_t s = v < 0 ? 1 : 0; |
202 | uint32_t j1 = (v >> 19) & 1; |
203 | uint32_t j2 = (v >> 18) & 1; |
204 | or16(p: off, v: (s << 10) | ((v >> 12) & 0x3f)); |
205 | or16(p: off + 2, v: (j1 << 13) | (j2 << 11) | ((v >> 1) & 0x7ff)); |
206 | } |
207 | |
208 | void applyBranch24T(uint8_t *off, int32_t v) { |
209 | if (!isInt<25>(x: v)) |
210 | error(msg: "relocation out of range" ); |
211 | uint32_t s = v < 0 ? 1 : 0; |
212 | uint32_t j1 = ((~v >> 23) & 1) ^ s; |
213 | uint32_t j2 = ((~v >> 22) & 1) ^ s; |
214 | or16(p: off, v: (s << 10) | ((v >> 12) & 0x3ff)); |
215 | // Clear out the J1 and J2 bits which may be set. |
216 | write16le(P: off + 2, V: (read16le(P: off + 2) & 0xd000) | (j1 << 13) | (j2 << 11) | ((v >> 1) & 0x7ff)); |
217 | } |
218 | |
219 | void SectionChunk::applyRelARM(uint8_t *off, uint16_t type, OutputSection *os, |
220 | uint64_t s, uint64_t p, |
221 | uint64_t imageBase) const { |
222 | // Pointer to thumb code must have the LSB set. |
223 | uint64_t sx = s; |
224 | if (os && (os->header.Characteristics & IMAGE_SCN_MEM_EXECUTE)) |
225 | sx |= 1; |
226 | switch (type) { |
227 | case IMAGE_REL_ARM_ADDR32: |
228 | add32(p: off, v: sx + imageBase); |
229 | break; |
230 | case IMAGE_REL_ARM_ADDR32NB: add32(p: off, v: sx); break; |
231 | case IMAGE_REL_ARM_MOV32T: |
232 | applyMOV32T(off, v: sx + imageBase); |
233 | break; |
234 | case IMAGE_REL_ARM_BRANCH20T: applyBranch20T(off, v: sx - p - 4); break; |
235 | case IMAGE_REL_ARM_BRANCH24T: applyBranch24T(off, v: sx - p - 4); break; |
236 | case IMAGE_REL_ARM_BLX23T: applyBranch24T(off, v: sx - p - 4); break; |
237 | case IMAGE_REL_ARM_SECTION: |
238 | applySecIdx(off, os, numOutputSections: file->symtab.ctx.outputSections.size()); |
239 | break; |
240 | case IMAGE_REL_ARM_SECREL: applySecRel(sec: this, off, os, s); break; |
241 | case IMAGE_REL_ARM_REL32: add32(p: off, v: sx - p - 4); break; |
242 | default: |
243 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
244 | toString(file)); |
245 | } |
246 | } |
247 | |
248 | // Interpret the existing immediate value as a byte offset to the |
249 | // target symbol, then update the instruction with the immediate as |
250 | // the page offset from the current instruction to the target. |
251 | void applyArm64Addr(uint8_t *off, uint64_t s, uint64_t p, int shift) { |
252 | uint32_t orig = read32le(P: off); |
253 | int64_t imm = |
254 | SignExtend64<21>(x: ((orig >> 29) & 0x3) | ((orig >> 3) & 0x1FFFFC)); |
255 | s += imm; |
256 | imm = (s >> shift) - (p >> shift); |
257 | uint32_t immLo = (imm & 0x3) << 29; |
258 | uint32_t immHi = (imm & 0x1FFFFC) << 3; |
259 | uint64_t mask = (0x3 << 29) | (0x1FFFFC << 3); |
260 | write32le(P: off, V: (orig & ~mask) | immLo | immHi); |
261 | } |
262 | |
263 | // Update the immediate field in a AARCH64 ldr, str, and add instruction. |
264 | // Optionally limit the range of the written immediate by one or more bits |
265 | // (rangeLimit). |
266 | void applyArm64Imm(uint8_t *off, uint64_t imm, uint32_t rangeLimit) { |
267 | uint32_t orig = read32le(P: off); |
268 | imm += (orig >> 10) & 0xFFF; |
269 | orig &= ~(0xFFF << 10); |
270 | write32le(P: off, V: orig | ((imm & (0xFFF >> rangeLimit)) << 10)); |
271 | } |
272 | |
273 | // Add the 12 bit page offset to the existing immediate. |
274 | // Ldr/str instructions store the opcode immediate scaled |
275 | // by the load/store size (giving a larger range for larger |
276 | // loads/stores). The immediate is always (both before and after |
277 | // fixing up the relocation) stored scaled similarly. |
278 | // Even if larger loads/stores have a larger range, limit the |
279 | // effective offset to 12 bit, since it is intended to be a |
280 | // page offset. |
281 | static void applyArm64Ldr(uint8_t *off, uint64_t imm) { |
282 | uint32_t orig = read32le(P: off); |
283 | uint32_t size = orig >> 30; |
284 | // 0x04000000 indicates SIMD/FP registers |
285 | // 0x00800000 indicates 128 bit |
286 | if ((orig & 0x4800000) == 0x4800000) |
287 | size += 4; |
288 | if ((imm & ((1 << size) - 1)) != 0) |
289 | error(msg: "misaligned ldr/str offset" ); |
290 | applyArm64Imm(off, imm: imm >> size, rangeLimit: size); |
291 | } |
292 | |
293 | static void applySecRelLow12A(const SectionChunk *sec, uint8_t *off, |
294 | OutputSection *os, uint64_t s) { |
295 | if (checkSecRel(sec, os)) |
296 | applyArm64Imm(off, imm: (s - os->getRVA()) & 0xfff, rangeLimit: 0); |
297 | } |
298 | |
299 | static void applySecRelHigh12A(const SectionChunk *sec, uint8_t *off, |
300 | OutputSection *os, uint64_t s) { |
301 | if (!checkSecRel(sec, os)) |
302 | return; |
303 | uint64_t secRel = (s - os->getRVA()) >> 12; |
304 | if (0xfff < secRel) { |
305 | error(msg: "overflow in SECREL_HIGH12A relocation in section: " + |
306 | sec->getSectionName()); |
307 | return; |
308 | } |
309 | applyArm64Imm(off, imm: secRel & 0xfff, rangeLimit: 0); |
310 | } |
311 | |
312 | static void applySecRelLdr(const SectionChunk *sec, uint8_t *off, |
313 | OutputSection *os, uint64_t s) { |
314 | if (checkSecRel(sec, os)) |
315 | applyArm64Ldr(off, imm: (s - os->getRVA()) & 0xfff); |
316 | } |
317 | |
318 | void applyArm64Branch26(uint8_t *off, int64_t v) { |
319 | if (!isInt<28>(x: v)) |
320 | error(msg: "relocation out of range" ); |
321 | or32(p: off, v: (v & 0x0FFFFFFC) >> 2); |
322 | } |
323 | |
324 | static void applyArm64Branch19(uint8_t *off, int64_t v) { |
325 | if (!isInt<21>(x: v)) |
326 | error(msg: "relocation out of range" ); |
327 | or32(p: off, v: (v & 0x001FFFFC) << 3); |
328 | } |
329 | |
330 | static void applyArm64Branch14(uint8_t *off, int64_t v) { |
331 | if (!isInt<16>(x: v)) |
332 | error(msg: "relocation out of range" ); |
333 | or32(p: off, v: (v & 0x0000FFFC) << 3); |
334 | } |
335 | |
336 | void SectionChunk::applyRelARM64(uint8_t *off, uint16_t type, OutputSection *os, |
337 | uint64_t s, uint64_t p, |
338 | uint64_t imageBase) const { |
339 | switch (type) { |
340 | case IMAGE_REL_ARM64_PAGEBASE_REL21: applyArm64Addr(off, s, p, shift: 12); break; |
341 | case IMAGE_REL_ARM64_REL21: applyArm64Addr(off, s, p, shift: 0); break; |
342 | case IMAGE_REL_ARM64_PAGEOFFSET_12A: applyArm64Imm(off, imm: s & 0xfff, rangeLimit: 0); break; |
343 | case IMAGE_REL_ARM64_PAGEOFFSET_12L: applyArm64Ldr(off, imm: s & 0xfff); break; |
344 | case IMAGE_REL_ARM64_BRANCH26: applyArm64Branch26(off, v: s - p); break; |
345 | case IMAGE_REL_ARM64_BRANCH19: applyArm64Branch19(off, v: s - p); break; |
346 | case IMAGE_REL_ARM64_BRANCH14: applyArm64Branch14(off, v: s - p); break; |
347 | case IMAGE_REL_ARM64_ADDR32: |
348 | add32(p: off, v: s + imageBase); |
349 | break; |
350 | case IMAGE_REL_ARM64_ADDR32NB: add32(p: off, v: s); break; |
351 | case IMAGE_REL_ARM64_ADDR64: |
352 | add64(p: off, v: s + imageBase); |
353 | break; |
354 | case IMAGE_REL_ARM64_SECREL: applySecRel(sec: this, off, os, s); break; |
355 | case IMAGE_REL_ARM64_SECREL_LOW12A: applySecRelLow12A(sec: this, off, os, s); break; |
356 | case IMAGE_REL_ARM64_SECREL_HIGH12A: applySecRelHigh12A(sec: this, off, os, s); break; |
357 | case IMAGE_REL_ARM64_SECREL_LOW12L: applySecRelLdr(sec: this, off, os, s); break; |
358 | case IMAGE_REL_ARM64_SECTION: |
359 | applySecIdx(off, os, numOutputSections: file->symtab.ctx.outputSections.size()); |
360 | break; |
361 | case IMAGE_REL_ARM64_REL32: add32(p: off, v: s - p - 4); break; |
362 | default: |
363 | error(msg: "unsupported relocation type 0x" + Twine::utohexstr(Val: type) + " in " + |
364 | toString(file)); |
365 | } |
366 | } |
367 | |
368 | static void maybeReportRelocationToDiscarded(const SectionChunk *fromChunk, |
369 | Defined *sym, |
370 | const coff_relocation &rel, |
371 | bool isMinGW) { |
372 | // Don't report these errors when the relocation comes from a debug info |
373 | // section or in mingw mode. MinGW mode object files (built by GCC) can |
374 | // have leftover sections with relocations against discarded comdat |
375 | // sections. Such sections are left as is, with relocations untouched. |
376 | if (fromChunk->isCodeView() || fromChunk->isDWARF() || isMinGW) |
377 | return; |
378 | |
379 | // Get the name of the symbol. If it's null, it was discarded early, so we |
380 | // have to go back to the object file. |
381 | ObjFile *file = fromChunk->file; |
382 | std::string name; |
383 | if (sym) { |
384 | name = toString(ctx: file->symtab.ctx, b&: *sym); |
385 | } else { |
386 | COFFSymbolRef coffSym = |
387 | check(e: file->getCOFFObj()->getSymbol(index: rel.SymbolTableIndex)); |
388 | name = maybeDemangleSymbol( |
389 | ctx: file->symtab.ctx, symName: check(e: file->getCOFFObj()->getSymbolName(Symbol: coffSym))); |
390 | } |
391 | |
392 | std::vector<std::string> symbolLocations = |
393 | getSymbolLocations(file, symIndex: rel.SymbolTableIndex); |
394 | |
395 | std::string out; |
396 | llvm::raw_string_ostream os(out); |
397 | os << "relocation against symbol in discarded section: " + name; |
398 | for (const std::string &s : symbolLocations) |
399 | os << s; |
400 | error(msg: out); |
401 | } |
402 | |
403 | void SectionChunk::writeTo(uint8_t *buf) const { |
404 | if (!hasData) |
405 | return; |
406 | // Copy section contents from source object file to output file. |
407 | ArrayRef<uint8_t> a = getContents(); |
408 | if (!a.empty()) |
409 | memcpy(dest: buf, src: a.data(), n: a.size()); |
410 | |
411 | // Apply relocations. |
412 | size_t inputSize = getSize(); |
413 | for (const coff_relocation &rel : getRelocs()) { |
414 | // Check for an invalid relocation offset. This check isn't perfect, because |
415 | // we don't have the relocation size, which is only known after checking the |
416 | // machine and relocation type. As a result, a relocation may overwrite the |
417 | // beginning of the following input section. |
418 | if (rel.VirtualAddress >= inputSize) { |
419 | error(msg: "relocation points beyond the end of its parent section" ); |
420 | continue; |
421 | } |
422 | |
423 | applyRelocation(off: buf + rel.VirtualAddress, rel); |
424 | } |
425 | |
426 | // Write the offset to EC entry thunk preceding section contents. The low bit |
427 | // is always set, so it's effectively an offset from the last byte of the |
428 | // offset. |
429 | if (Defined *entryThunk = getEntryThunk()) |
430 | write32le(P: buf - sizeof(uint32_t), V: entryThunk->getRVA() - rva + 1); |
431 | } |
432 | |
433 | void SectionChunk::applyRelocation(uint8_t *off, |
434 | const coff_relocation &rel) const { |
435 | auto *sym = dyn_cast_or_null<Defined>(Val: file->getSymbol(symbolIndex: rel.SymbolTableIndex)); |
436 | |
437 | // Get the output section of the symbol for this relocation. The output |
438 | // section is needed to compute SECREL and SECTION relocations used in debug |
439 | // info. |
440 | Chunk *c = sym ? sym->getChunk() : nullptr; |
441 | COFFLinkerContext &ctx = file->symtab.ctx; |
442 | OutputSection *os = c ? ctx.getOutputSection(c) : nullptr; |
443 | |
444 | // Skip the relocation if it refers to a discarded section, and diagnose it |
445 | // as an error if appropriate. If a symbol was discarded early, it may be |
446 | // null. If it was discarded late, the output section will be null, unless |
447 | // it was an absolute or synthetic symbol. |
448 | if (!sym || |
449 | (!os && !isa<DefinedAbsolute>(Val: sym) && !isa<DefinedSynthetic>(Val: sym))) { |
450 | maybeReportRelocationToDiscarded(fromChunk: this, sym, rel, isMinGW: ctx.config.mingw); |
451 | return; |
452 | } |
453 | |
454 | uint64_t s = sym->getRVA(); |
455 | |
456 | // Compute the RVA of the relocation for relative relocations. |
457 | uint64_t p = rva + rel.VirtualAddress; |
458 | uint64_t imageBase = ctx.config.imageBase; |
459 | switch (getArch()) { |
460 | case Triple::x86_64: |
461 | applyRelX64(off, type: rel.Type, os, s, p, imageBase); |
462 | break; |
463 | case Triple::x86: |
464 | applyRelX86(off, type: rel.Type, os, s, p, imageBase); |
465 | break; |
466 | case Triple::thumb: |
467 | applyRelARM(off, type: rel.Type, os, s, p, imageBase); |
468 | break; |
469 | case Triple::aarch64: |
470 | applyRelARM64(off, type: rel.Type, os, s, p, imageBase); |
471 | break; |
472 | default: |
473 | llvm_unreachable("unknown machine type" ); |
474 | } |
475 | } |
476 | |
477 | // Defend against unsorted relocations. This may be overly conservative. |
478 | void SectionChunk::sortRelocations() { |
479 | auto cmpByVa = [](const coff_relocation &l, const coff_relocation &r) { |
480 | return l.VirtualAddress < r.VirtualAddress; |
481 | }; |
482 | if (llvm::is_sorted(Range: getRelocs(), C: cmpByVa)) |
483 | return; |
484 | warn(msg: "some relocations in " + file->getName() + " are not sorted" ); |
485 | MutableArrayRef<coff_relocation> newRelocs( |
486 | bAlloc().Allocate<coff_relocation>(Num: relocsSize), relocsSize); |
487 | memcpy(dest: newRelocs.data(), src: relocsData, n: relocsSize * sizeof(coff_relocation)); |
488 | llvm::sort(C&: newRelocs, Comp: cmpByVa); |
489 | setRelocs(newRelocs); |
490 | } |
491 | |
492 | // Similar to writeTo, but suitable for relocating a subsection of the overall |
493 | // section. |
494 | void SectionChunk::writeAndRelocateSubsection(ArrayRef<uint8_t> sec, |
495 | ArrayRef<uint8_t> subsec, |
496 | uint32_t &nextRelocIndex, |
497 | uint8_t *buf) const { |
498 | assert(!subsec.empty() && !sec.empty()); |
499 | assert(sec.begin() <= subsec.begin() && subsec.end() <= sec.end() && |
500 | "subsection is not part of this section" ); |
501 | size_t vaBegin = std::distance(first: sec.begin(), last: subsec.begin()); |
502 | size_t vaEnd = std::distance(first: sec.begin(), last: subsec.end()); |
503 | memcpy(dest: buf, src: subsec.data(), n: subsec.size()); |
504 | for (; nextRelocIndex < relocsSize; ++nextRelocIndex) { |
505 | const coff_relocation &rel = relocsData[nextRelocIndex]; |
506 | // Only apply relocations that apply to this subsection. These checks |
507 | // assume that all subsections completely contain their relocations. |
508 | // Relocations must not straddle the beginning or end of a subsection. |
509 | if (rel.VirtualAddress < vaBegin) |
510 | continue; |
511 | if (rel.VirtualAddress + 1 >= vaEnd) |
512 | break; |
513 | applyRelocation(off: &buf[rel.VirtualAddress - vaBegin], rel); |
514 | } |
515 | } |
516 | |
517 | void SectionChunk::addAssociative(SectionChunk *child) { |
518 | // Insert the child section into the list of associated children. Keep the |
519 | // list ordered by section name so that ICF does not depend on section order. |
520 | assert(child->assocChildren == nullptr && |
521 | "associated sections cannot have their own associated children" ); |
522 | SectionChunk *prev = this; |
523 | SectionChunk *next = assocChildren; |
524 | for (; next != nullptr; prev = next, next = next->assocChildren) { |
525 | if (next->getSectionName() <= child->getSectionName()) |
526 | break; |
527 | } |
528 | |
529 | // Insert child between prev and next. |
530 | assert(prev->assocChildren == next); |
531 | prev->assocChildren = child; |
532 | child->assocChildren = next; |
533 | } |
534 | |
535 | static uint8_t getBaserelType(const coff_relocation &rel, |
536 | Triple::ArchType arch) { |
537 | switch (arch) { |
538 | case Triple::x86_64: |
539 | if (rel.Type == IMAGE_REL_AMD64_ADDR64) |
540 | return IMAGE_REL_BASED_DIR64; |
541 | if (rel.Type == IMAGE_REL_AMD64_ADDR32) |
542 | return IMAGE_REL_BASED_HIGHLOW; |
543 | return IMAGE_REL_BASED_ABSOLUTE; |
544 | case Triple::x86: |
545 | if (rel.Type == IMAGE_REL_I386_DIR32) |
546 | return IMAGE_REL_BASED_HIGHLOW; |
547 | return IMAGE_REL_BASED_ABSOLUTE; |
548 | case Triple::thumb: |
549 | if (rel.Type == IMAGE_REL_ARM_ADDR32) |
550 | return IMAGE_REL_BASED_HIGHLOW; |
551 | if (rel.Type == IMAGE_REL_ARM_MOV32T) |
552 | return IMAGE_REL_BASED_ARM_MOV32T; |
553 | return IMAGE_REL_BASED_ABSOLUTE; |
554 | case Triple::aarch64: |
555 | if (rel.Type == IMAGE_REL_ARM64_ADDR64) |
556 | return IMAGE_REL_BASED_DIR64; |
557 | return IMAGE_REL_BASED_ABSOLUTE; |
558 | default: |
559 | llvm_unreachable("unknown machine type" ); |
560 | } |
561 | } |
562 | |
563 | // Windows-specific. |
564 | // Collect all locations that contain absolute addresses, which need to be |
565 | // fixed by the loader if load-time relocation is needed. |
566 | // Only called when base relocation is enabled. |
567 | void SectionChunk::getBaserels(std::vector<Baserel> *res) { |
568 | for (const coff_relocation &rel : getRelocs()) { |
569 | uint8_t ty = getBaserelType(rel, arch: getArch()); |
570 | if (ty == IMAGE_REL_BASED_ABSOLUTE) |
571 | continue; |
572 | Symbol *target = file->getSymbol(symbolIndex: rel.SymbolTableIndex); |
573 | if (!target || isa<DefinedAbsolute>(Val: target)) |
574 | continue; |
575 | res->emplace_back(args: rva + rel.VirtualAddress, args&: ty); |
576 | } |
577 | |
578 | // Insert a 64-bit relocation for CHPEMetadataPointer in the native load |
579 | // config of a hybrid ARM64X image. Its value will be set in prepareLoadConfig |
580 | // to match the value in the EC load config, which is expected to be |
581 | // a relocatable pointer to the __chpe_metadata symbol. |
582 | COFFLinkerContext &ctx = file->symtab.ctx; |
583 | if (ctx.config.machine == ARM64X && ctx.hybridSymtab->loadConfigSym && |
584 | ctx.hybridSymtab->loadConfigSym->getChunk() == this && |
585 | ctx.symtab.loadConfigSym && |
586 | ctx.hybridSymtab->loadConfigSize >= |
587 | offsetof(coff_load_configuration64, CHPEMetadataPointer) + |
588 | sizeof(coff_load_configuration64::CHPEMetadataPointer)) |
589 | res->emplace_back( |
590 | args: ctx.hybridSymtab->loadConfigSym->getRVA() + |
591 | offsetof(coff_load_configuration64, CHPEMetadataPointer), |
592 | args: IMAGE_REL_BASED_DIR64); |
593 | } |
594 | |
595 | // MinGW specific. |
596 | // Check whether a static relocation of type Type can be deferred and |
597 | // handled at runtime as a pseudo relocation (for references to a module |
598 | // local variable, which turned out to actually need to be imported from |
599 | // another DLL) This returns the size the relocation is supposed to update, |
600 | // in bits, or 0 if the relocation cannot be handled as a runtime pseudo |
601 | // relocation. |
602 | static int getRuntimePseudoRelocSize(uint16_t type, Triple::ArchType arch) { |
603 | // Relocations that either contain an absolute address, or a plain |
604 | // relative offset, since the runtime pseudo reloc implementation |
605 | // adds 8/16/32/64 bit values to a memory address. |
606 | // |
607 | // Given a pseudo relocation entry, |
608 | // |
609 | // typedef struct { |
610 | // DWORD sym; |
611 | // DWORD target; |
612 | // DWORD flags; |
613 | // } runtime_pseudo_reloc_item_v2; |
614 | // |
615 | // the runtime relocation performs this adjustment: |
616 | // *(base + .target) += *(base + .sym) - (base + .sym) |
617 | // |
618 | // This works for both absolute addresses (IMAGE_REL_*_ADDR32/64, |
619 | // IMAGE_REL_I386_DIR32, where the memory location initially contains |
620 | // the address of the IAT slot, and for relative addresses (IMAGE_REL*_REL32), |
621 | // where the memory location originally contains the relative offset to the |
622 | // IAT slot. |
623 | // |
624 | // This requires the target address to be writable, either directly out of |
625 | // the image, or temporarily changed at runtime with VirtualProtect. |
626 | // Since this only operates on direct address values, it doesn't work for |
627 | // ARM/ARM64 relocations, other than the plain ADDR32/ADDR64 relocations. |
628 | switch (arch) { |
629 | case Triple::x86_64: |
630 | switch (type) { |
631 | case IMAGE_REL_AMD64_ADDR64: |
632 | return 64; |
633 | case IMAGE_REL_AMD64_ADDR32: |
634 | case IMAGE_REL_AMD64_REL32: |
635 | case IMAGE_REL_AMD64_REL32_1: |
636 | case IMAGE_REL_AMD64_REL32_2: |
637 | case IMAGE_REL_AMD64_REL32_3: |
638 | case IMAGE_REL_AMD64_REL32_4: |
639 | case IMAGE_REL_AMD64_REL32_5: |
640 | return 32; |
641 | default: |
642 | return 0; |
643 | } |
644 | case Triple::x86: |
645 | switch (type) { |
646 | case IMAGE_REL_I386_DIR32: |
647 | case IMAGE_REL_I386_REL32: |
648 | return 32; |
649 | default: |
650 | return 0; |
651 | } |
652 | case Triple::thumb: |
653 | switch (type) { |
654 | case IMAGE_REL_ARM_ADDR32: |
655 | return 32; |
656 | default: |
657 | return 0; |
658 | } |
659 | case Triple::aarch64: |
660 | switch (type) { |
661 | case IMAGE_REL_ARM64_ADDR64: |
662 | return 64; |
663 | case IMAGE_REL_ARM64_ADDR32: |
664 | return 32; |
665 | default: |
666 | return 0; |
667 | } |
668 | default: |
669 | llvm_unreachable("unknown machine type" ); |
670 | } |
671 | } |
672 | |
673 | // MinGW specific. |
674 | // Append information to the provided vector about all relocations that |
675 | // need to be handled at runtime as runtime pseudo relocations (references |
676 | // to a module local variable, which turned out to actually need to be |
677 | // imported from another DLL). |
678 | void SectionChunk::getRuntimePseudoRelocs( |
679 | std::vector<RuntimePseudoReloc> &res) { |
680 | for (const coff_relocation &rel : getRelocs()) { |
681 | auto *target = |
682 | dyn_cast_or_null<Defined>(Val: file->getSymbol(symbolIndex: rel.SymbolTableIndex)); |
683 | if (!target || !target->isRuntimePseudoReloc) |
684 | continue; |
685 | // If the target doesn't have a chunk allocated, it may be a |
686 | // DefinedImportData symbol which ended up unnecessary after GC. |
687 | // Normally we wouldn't eliminate section chunks that are referenced, but |
688 | // references within DWARF sections don't count for keeping section chunks |
689 | // alive. Thus such dangling references in DWARF sections are expected. |
690 | if (!target->getChunk()) |
691 | continue; |
692 | int sizeInBits = getRuntimePseudoRelocSize(type: rel.Type, arch: getArch()); |
693 | if (sizeInBits == 0) { |
694 | error(msg: "unable to automatically import from " + target->getName() + |
695 | " with relocation type " + |
696 | file->getCOFFObj()->getRelocationTypeName(Type: rel.Type) + " in " + |
697 | toString(file)); |
698 | continue; |
699 | } |
700 | int addressSizeInBits = file->symtab.ctx.config.is64() ? 64 : 32; |
701 | if (sizeInBits < addressSizeInBits) { |
702 | warn(msg: "runtime pseudo relocation in " + toString(file) + " against " + |
703 | "symbol " + target->getName() + " is too narrow (only " + |
704 | Twine(sizeInBits) + " bits wide); this can fail at runtime " + |
705 | "depending on memory layout" ); |
706 | } |
707 | // sizeInBits is used to initialize the Flags field; currently no |
708 | // other flags are defined. |
709 | res.emplace_back(args&: target, args: this, args: rel.VirtualAddress, args&: sizeInBits); |
710 | } |
711 | } |
712 | |
713 | bool SectionChunk::isCOMDAT() const { |
714 | return header->Characteristics & IMAGE_SCN_LNK_COMDAT; |
715 | } |
716 | |
717 | void SectionChunk::printDiscardedMessage() const { |
718 | // Removed by dead-stripping. If it's removed by ICF, ICF already |
719 | // printed out the name, so don't repeat that here. |
720 | if (sym && this == repl) |
721 | log(msg: "Discarded " + sym->getName()); |
722 | } |
723 | |
724 | StringRef SectionChunk::getDebugName() const { |
725 | if (sym) |
726 | return sym->getName(); |
727 | return "" ; |
728 | } |
729 | |
730 | ArrayRef<uint8_t> SectionChunk::getContents() const { |
731 | ArrayRef<uint8_t> a; |
732 | cantFail(Err: file->getCOFFObj()->getSectionContents(Sec: header, Res&: a)); |
733 | return a; |
734 | } |
735 | |
736 | ArrayRef<uint8_t> SectionChunk::consumeDebugMagic() { |
737 | assert(isCodeView()); |
738 | return consumeDebugMagic(data: getContents(), sectionName: getSectionName()); |
739 | } |
740 | |
741 | ArrayRef<uint8_t> SectionChunk::consumeDebugMagic(ArrayRef<uint8_t> data, |
742 | StringRef sectionName) { |
743 | if (data.empty()) |
744 | return {}; |
745 | |
746 | // First 4 bytes are section magic. |
747 | if (data.size() < 4) |
748 | fatal(msg: "the section is too short: " + sectionName); |
749 | |
750 | if (!sectionName.starts_with(Prefix: ".debug$" )) |
751 | fatal(msg: "invalid section: " + sectionName); |
752 | |
753 | uint32_t magic = support::endian::read32le(P: data.data()); |
754 | uint32_t expectedMagic = sectionName == ".debug$H" |
755 | ? DEBUG_HASHES_SECTION_MAGIC |
756 | : DEBUG_SECTION_MAGIC; |
757 | if (magic != expectedMagic) { |
758 | warn(msg: "ignoring section " + sectionName + " with unrecognized magic 0x" + |
759 | utohexstr(X: magic)); |
760 | return {}; |
761 | } |
762 | return data.slice(N: 4); |
763 | } |
764 | |
765 | SectionChunk *SectionChunk::findByName(ArrayRef<SectionChunk *> sections, |
766 | StringRef name) { |
767 | for (SectionChunk *c : sections) |
768 | if (c->getSectionName() == name) |
769 | return c; |
770 | return nullptr; |
771 | } |
772 | |
773 | void SectionChunk::replace(SectionChunk *other) { |
774 | p2Align = std::max(a: p2Align, b: other->p2Align); |
775 | other->repl = repl; |
776 | other->live = false; |
777 | } |
778 | |
779 | uint32_t SectionChunk::getSectionNumber() const { |
780 | DataRefImpl r; |
781 | r.p = reinterpret_cast<uintptr_t>(header); |
782 | SectionRef s(r, file->getCOFFObj()); |
783 | return s.getIndex() + 1; |
784 | } |
785 | |
786 | CommonChunk::CommonChunk(const COFFSymbolRef s) : sym(s) { |
787 | // The value of a common symbol is its size. Align all common symbols smaller |
788 | // than 32 bytes naturally, i.e. round the size up to the next power of two. |
789 | // This is what MSVC link.exe does. |
790 | setAlignment(std::min(a: 32U, b: uint32_t(PowerOf2Ceil(A: sym.getValue())))); |
791 | hasData = false; |
792 | } |
793 | |
794 | uint32_t CommonChunk::getOutputCharacteristics() const { |
795 | return IMAGE_SCN_CNT_UNINITIALIZED_DATA | IMAGE_SCN_MEM_READ | |
796 | IMAGE_SCN_MEM_WRITE; |
797 | } |
798 | |
799 | void StringChunk::writeTo(uint8_t *buf) const { |
800 | memcpy(dest: buf, src: str.data(), n: str.size()); |
801 | buf[str.size()] = '\0'; |
802 | } |
803 | |
804 | ImportThunkChunk::ImportThunkChunk(COFFLinkerContext &ctx, Defined *s) |
805 | : NonSectionCodeChunk(ImportThunkKind), live(!ctx.config.doGC), |
806 | impSymbol(s), ctx(ctx) {} |
807 | |
808 | ImportThunkChunkX64::ImportThunkChunkX64(COFFLinkerContext &ctx, Defined *s) |
809 | : ImportThunkChunk(ctx, s) { |
810 | // Intel Optimization Manual says that all branch targets |
811 | // should be 16-byte aligned. MSVC linker does this too. |
812 | setAlignment(16); |
813 | } |
814 | |
815 | void ImportThunkChunkX64::writeTo(uint8_t *buf) const { |
816 | memcpy(dest: buf, src: importThunkX86, n: sizeof(importThunkX86)); |
817 | // The first two bytes is a JMP instruction. Fill its operand. |
818 | write32le(P: buf + 2, V: impSymbol->getRVA() - rva - getSize()); |
819 | } |
820 | |
821 | void ImportThunkChunkX86::getBaserels(std::vector<Baserel> *res) { |
822 | res->emplace_back(args: getRVA() + 2, args&: ctx.config.machine); |
823 | } |
824 | |
825 | void ImportThunkChunkX86::writeTo(uint8_t *buf) const { |
826 | memcpy(dest: buf, src: importThunkX86, n: sizeof(importThunkX86)); |
827 | // The first two bytes is a JMP instruction. Fill its operand. |
828 | write32le(P: buf + 2, V: impSymbol->getRVA() + ctx.config.imageBase); |
829 | } |
830 | |
831 | void ImportThunkChunkARM::getBaserels(std::vector<Baserel> *res) { |
832 | res->emplace_back(args: getRVA(), args: IMAGE_REL_BASED_ARM_MOV32T); |
833 | } |
834 | |
835 | void ImportThunkChunkARM::writeTo(uint8_t *buf) const { |
836 | memcpy(dest: buf, src: importThunkARM, n: sizeof(importThunkARM)); |
837 | // Fix mov.w and mov.t operands. |
838 | applyMOV32T(off: buf, v: impSymbol->getRVA() + ctx.config.imageBase); |
839 | } |
840 | |
841 | void ImportThunkChunkARM64::writeTo(uint8_t *buf) const { |
842 | int64_t off = impSymbol->getRVA() & 0xfff; |
843 | memcpy(dest: buf, src: importThunkARM64, n: sizeof(importThunkARM64)); |
844 | applyArm64Addr(off: buf, s: impSymbol->getRVA(), p: rva, shift: 12); |
845 | applyArm64Ldr(off: buf + 4, imm: off); |
846 | } |
847 | |
848 | // A Thumb2, PIC, non-interworking range extension thunk. |
849 | const uint8_t armThunk[] = { |
850 | 0x40, 0xf2, 0x00, 0x0c, // P: movw ip,:lower16:S - (P + (L1-P) + 4) |
851 | 0xc0, 0xf2, 0x00, 0x0c, // movt ip,:upper16:S - (P + (L1-P) + 4) |
852 | 0xe7, 0x44, // L1: add pc, ip |
853 | }; |
854 | |
855 | size_t RangeExtensionThunkARM::getSize() const { |
856 | assert(ctx.config.machine == ARMNT); |
857 | (void)&ctx; |
858 | return sizeof(armThunk); |
859 | } |
860 | |
861 | void RangeExtensionThunkARM::writeTo(uint8_t *buf) const { |
862 | assert(ctx.config.machine == ARMNT); |
863 | uint64_t offset = target->getRVA() - rva - 12; |
864 | memcpy(dest: buf, src: armThunk, n: sizeof(armThunk)); |
865 | applyMOV32T(off: buf, v: uint32_t(offset)); |
866 | } |
867 | |
868 | // A position independent ARM64 adrp+add thunk, with a maximum range of |
869 | // +/- 4 GB, which is enough for any PE-COFF. |
870 | const uint8_t arm64Thunk[] = { |
871 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Dest |
872 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, :lo12:Dest |
873 | 0x00, 0x02, 0x1f, 0xd6, // br x16 |
874 | }; |
875 | |
876 | size_t RangeExtensionThunkARM64::getSize() const { return sizeof(arm64Thunk); } |
877 | |
878 | void RangeExtensionThunkARM64::writeTo(uint8_t *buf) const { |
879 | memcpy(dest: buf, src: arm64Thunk, n: sizeof(arm64Thunk)); |
880 | applyArm64Addr(off: buf + 0, s: target->getRVA(), p: rva, shift: 12); |
881 | applyArm64Imm(off: buf + 4, imm: target->getRVA() & 0xfff, rangeLimit: 0); |
882 | } |
883 | |
884 | LocalImportChunk::LocalImportChunk(COFFLinkerContext &c, Defined *s) |
885 | : sym(s), ctx(c) { |
886 | setAlignment(ctx.config.wordsize); |
887 | } |
888 | |
889 | void LocalImportChunk::getBaserels(std::vector<Baserel> *res) { |
890 | res->emplace_back(args: getRVA(), args&: ctx.config.machine); |
891 | } |
892 | |
893 | size_t LocalImportChunk::getSize() const { return ctx.config.wordsize; } |
894 | |
895 | void LocalImportChunk::writeTo(uint8_t *buf) const { |
896 | if (ctx.config.is64()) { |
897 | write64le(P: buf, V: sym->getRVA() + ctx.config.imageBase); |
898 | } else { |
899 | write32le(P: buf, V: sym->getRVA() + ctx.config.imageBase); |
900 | } |
901 | } |
902 | |
903 | void RVATableChunk::writeTo(uint8_t *buf) const { |
904 | ulittle32_t *begin = reinterpret_cast<ulittle32_t *>(buf); |
905 | size_t cnt = 0; |
906 | for (const ChunkAndOffset &co : syms) |
907 | begin[cnt++] = co.inputChunk->getRVA() + co.offset; |
908 | llvm::sort(Start: begin, End: begin + cnt); |
909 | assert(std::unique(begin, begin + cnt) == begin + cnt && |
910 | "RVA tables should be de-duplicated" ); |
911 | } |
912 | |
913 | void RVAFlagTableChunk::writeTo(uint8_t *buf) const { |
914 | struct RVAFlag { |
915 | ulittle32_t rva; |
916 | uint8_t flag; |
917 | }; |
918 | auto flags = |
919 | MutableArrayRef(reinterpret_cast<RVAFlag *>(buf), syms.size()); |
920 | for (auto t : zip(t: syms, u&: flags)) { |
921 | const auto &sym = std::get<0>(t&: t); |
922 | auto &flag = std::get<1>(t&: t); |
923 | flag.rva = sym.inputChunk->getRVA() + sym.offset; |
924 | flag.flag = 0; |
925 | } |
926 | llvm::sort(C&: flags, |
927 | Comp: [](const RVAFlag &a, const RVAFlag &b) { return a.rva < b.rva; }); |
928 | assert(llvm::unique(flags, [](const RVAFlag &a, |
929 | const RVAFlag &b) { return a.rva == b.rva; }) == |
930 | flags.end() && |
931 | "RVA tables should be de-duplicated" ); |
932 | } |
933 | |
934 | size_t ECCodeMapChunk::getSize() const { |
935 | return map.size() * sizeof(chpe_range_entry); |
936 | } |
937 | |
938 | void ECCodeMapChunk::writeTo(uint8_t *buf) const { |
939 | auto table = reinterpret_cast<chpe_range_entry *>(buf); |
940 | for (uint32_t i = 0; i < map.size(); i++) { |
941 | const ECCodeMapEntry &entry = map[i]; |
942 | uint32_t start = entry.first->getRVA(); |
943 | table[i].StartOffset = start | entry.type; |
944 | table[i].Length = entry.last->getRVA() + entry.last->getSize() - start; |
945 | } |
946 | } |
947 | |
948 | // MinGW specific, for the "automatic import of variables from DLLs" feature. |
949 | size_t PseudoRelocTableChunk::getSize() const { |
950 | if (relocs.empty()) |
951 | return 0; |
952 | return 12 + 12 * relocs.size(); |
953 | } |
954 | |
955 | // MinGW specific. |
956 | void PseudoRelocTableChunk::writeTo(uint8_t *buf) const { |
957 | if (relocs.empty()) |
958 | return; |
959 | |
960 | ulittle32_t *table = reinterpret_cast<ulittle32_t *>(buf); |
961 | // This is the list header, to signal the runtime pseudo relocation v2 |
962 | // format. |
963 | table[0] = 0; |
964 | table[1] = 0; |
965 | table[2] = 1; |
966 | |
967 | size_t idx = 3; |
968 | for (const RuntimePseudoReloc &rpr : relocs) { |
969 | table[idx + 0] = rpr.sym->getRVA(); |
970 | table[idx + 1] = rpr.target->getRVA() + rpr.targetOffset; |
971 | table[idx + 2] = rpr.flags; |
972 | idx += 3; |
973 | } |
974 | } |
975 | |
976 | // Windows-specific. This class represents a block in .reloc section. |
977 | // The format is described here. |
978 | // |
979 | // On Windows, each DLL is linked against a fixed base address and |
980 | // usually loaded to that address. However, if there's already another |
981 | // DLL that overlaps, the loader has to relocate it. To do that, DLLs |
982 | // contain .reloc sections which contain offsets that need to be fixed |
983 | // up at runtime. If the loader finds that a DLL cannot be loaded to its |
984 | // desired base address, it loads it to somewhere else, and add <actual |
985 | // base address> - <desired base address> to each offset that is |
986 | // specified by the .reloc section. In ELF terms, .reloc sections |
987 | // contain relative relocations in REL format (as opposed to RELA.) |
988 | // |
989 | // This already significantly reduces the size of relocations compared |
990 | // to ELF .rel.dyn, but Windows does more to reduce it (probably because |
991 | // it was invented for PCs in the late '80s or early '90s.) Offsets in |
992 | // .reloc are grouped by page where the page size is 12 bits, and |
993 | // offsets sharing the same page address are stored consecutively to |
994 | // represent them with less space. This is very similar to the page |
995 | // table which is grouped by (multiple stages of) pages. |
996 | // |
997 | // For example, let's say we have 0x00030, 0x00500, 0x00700, 0x00A00, |
998 | // 0x20004, and 0x20008 in a .reloc section for x64. The uppermost 4 |
999 | // bits have a type IMAGE_REL_BASED_DIR64 or 0xA. In the section, they |
1000 | // are represented like this: |
1001 | // |
1002 | // 0x00000 -- page address (4 bytes) |
1003 | // 16 -- size of this block (4 bytes) |
1004 | // 0xA030 -- entries (2 bytes each) |
1005 | // 0xA500 |
1006 | // 0xA700 |
1007 | // 0xAA00 |
1008 | // 0x20000 -- page address (4 bytes) |
1009 | // 12 -- size of this block (4 bytes) |
1010 | // 0xA004 -- entries (2 bytes each) |
1011 | // 0xA008 |
1012 | // |
1013 | // Usually we have a lot of relocations for each page, so the number of |
1014 | // bytes for one .reloc entry is close to 2 bytes on average. |
1015 | BaserelChunk::BaserelChunk(uint32_t page, Baserel *begin, Baserel *end) { |
1016 | // Block header consists of 4 byte page RVA and 4 byte block size. |
1017 | // Each entry is 2 byte. Last entry may be padding. |
1018 | data.resize(new_size: alignTo(Value: (end - begin) * 2 + 8, Align: 4)); |
1019 | uint8_t *p = data.data(); |
1020 | write32le(P: p, V: page); |
1021 | write32le(P: p + 4, V: data.size()); |
1022 | p += 8; |
1023 | for (Baserel *i = begin; i != end; ++i) { |
1024 | write16le(P: p, V: (i->type << 12) | (i->rva - page)); |
1025 | p += 2; |
1026 | } |
1027 | } |
1028 | |
1029 | void BaserelChunk::writeTo(uint8_t *buf) const { |
1030 | memcpy(dest: buf, src: data.data(), n: data.size()); |
1031 | } |
1032 | |
1033 | uint8_t Baserel::getDefaultType(llvm::COFF::MachineTypes machine) { |
1034 | return is64Bit(Machine: machine) ? IMAGE_REL_BASED_DIR64 : IMAGE_REL_BASED_HIGHLOW; |
1035 | } |
1036 | |
1037 | MergeChunk::MergeChunk(uint32_t alignment) |
1038 | : builder(StringTableBuilder::RAW, llvm::Align(alignment)) { |
1039 | setAlignment(alignment); |
1040 | } |
1041 | |
1042 | void MergeChunk::addSection(COFFLinkerContext &ctx, SectionChunk *c) { |
1043 | assert(isPowerOf2_32(c->getAlignment())); |
1044 | uint8_t p2Align = llvm::Log2_32(Value: c->getAlignment()); |
1045 | assert(p2Align < std::size(ctx.mergeChunkInstances)); |
1046 | auto *&mc = ctx.mergeChunkInstances[p2Align]; |
1047 | if (!mc) |
1048 | mc = make<MergeChunk>(args: c->getAlignment()); |
1049 | mc->sections.push_back(x: c); |
1050 | } |
1051 | |
1052 | void MergeChunk::finalizeContents() { |
1053 | assert(!finalized && "should only finalize once" ); |
1054 | for (SectionChunk *c : sections) |
1055 | if (c->live) |
1056 | builder.add(S: toStringRef(Input: c->getContents())); |
1057 | builder.finalize(); |
1058 | finalized = true; |
1059 | } |
1060 | |
1061 | void MergeChunk::assignSubsectionRVAs() { |
1062 | for (SectionChunk *c : sections) { |
1063 | if (!c->live) |
1064 | continue; |
1065 | size_t off = builder.getOffset(S: toStringRef(Input: c->getContents())); |
1066 | c->setRVA(rva + off); |
1067 | } |
1068 | } |
1069 | |
1070 | uint32_t MergeChunk::getOutputCharacteristics() const { |
1071 | return IMAGE_SCN_MEM_READ | IMAGE_SCN_CNT_INITIALIZED_DATA; |
1072 | } |
1073 | |
1074 | size_t MergeChunk::getSize() const { |
1075 | return builder.getSize(); |
1076 | } |
1077 | |
1078 | void MergeChunk::writeTo(uint8_t *buf) const { |
1079 | builder.write(Buf: buf); |
1080 | } |
1081 | |
1082 | // MinGW specific. |
1083 | size_t AbsolutePointerChunk::getSize() const { |
1084 | return symtab.ctx.config.wordsize; |
1085 | } |
1086 | |
1087 | void AbsolutePointerChunk::writeTo(uint8_t *buf) const { |
1088 | if (symtab.ctx.config.is64()) { |
1089 | write64le(P: buf, V: value); |
1090 | } else { |
1091 | write32le(P: buf, V: value); |
1092 | } |
1093 | } |
1094 | |
1095 | MachineTypes AbsolutePointerChunk::getMachine() const { return symtab.machine; } |
1096 | |
1097 | void ECExportThunkChunk::writeTo(uint8_t *buf) const { |
1098 | memcpy(dest: buf, src: ECExportThunkCode, n: sizeof(ECExportThunkCode)); |
1099 | write32le(P: buf + 10, V: target->getRVA() - rva - 14); |
1100 | } |
1101 | |
1102 | size_t CHPECodeRangesChunk::getSize() const { |
1103 | return exportThunks.size() * sizeof(chpe_code_range_entry); |
1104 | } |
1105 | |
1106 | void CHPECodeRangesChunk::writeTo(uint8_t *buf) const { |
1107 | auto ranges = reinterpret_cast<chpe_code_range_entry *>(buf); |
1108 | |
1109 | for (uint32_t i = 0; i < exportThunks.size(); i++) { |
1110 | Chunk *thunk = exportThunks[i].first; |
1111 | uint32_t start = thunk->getRVA(); |
1112 | ranges[i].StartRva = start; |
1113 | ranges[i].EndRva = start + thunk->getSize(); |
1114 | ranges[i].EntryPoint = start; |
1115 | } |
1116 | } |
1117 | |
1118 | size_t CHPERedirectionChunk::getSize() const { |
1119 | // Add an extra +1 for a terminator entry. |
1120 | return (exportThunks.size() + 1) * sizeof(chpe_redirection_entry); |
1121 | } |
1122 | |
1123 | void CHPERedirectionChunk::writeTo(uint8_t *buf) const { |
1124 | auto entries = reinterpret_cast<chpe_redirection_entry *>(buf); |
1125 | |
1126 | for (uint32_t i = 0; i < exportThunks.size(); i++) { |
1127 | entries[i].Source = exportThunks[i].first->getRVA(); |
1128 | entries[i].Destination = exportThunks[i].second->getRVA(); |
1129 | } |
1130 | } |
1131 | |
1132 | ImportThunkChunkARM64EC::ImportThunkChunkARM64EC(ImportFile *file) |
1133 | : ImportThunkChunk(file->symtab.ctx, file->impSym), file(file) {} |
1134 | |
1135 | size_t ImportThunkChunkARM64EC::getSize() const { |
1136 | if (!extended) |
1137 | return sizeof(importThunkARM64EC); |
1138 | // The last instruction is replaced with an inline range extension thunk. |
1139 | return sizeof(importThunkARM64EC) + sizeof(arm64Thunk) - sizeof(uint32_t); |
1140 | } |
1141 | |
1142 | void ImportThunkChunkARM64EC::writeTo(uint8_t *buf) const { |
1143 | memcpy(dest: buf, src: importThunkARM64EC, n: sizeof(importThunkARM64EC)); |
1144 | applyArm64Addr(off: buf, s: file->impSym->getRVA(), p: rva, shift: 12); |
1145 | applyArm64Ldr(off: buf + 4, imm: file->impSym->getRVA() & 0xfff); |
1146 | |
1147 | // The exit thunk may be missing. This can happen if the application only |
1148 | // references a function by its address (in which case the thunk is never |
1149 | // actually used, but is still required to fill the auxiliary IAT), or in |
1150 | // cases of hand-written assembly calling an imported ARM64EC function (where |
1151 | // the exit thunk is ignored by __icall_helper_arm64ec). In such cases, MSVC |
1152 | // link.exe uses 0 as the RVA. |
1153 | uint32_t exitThunkRVA = exitThunk ? exitThunk->getRVA() : 0; |
1154 | applyArm64Addr(off: buf + 8, s: exitThunkRVA, p: rva + 8, shift: 12); |
1155 | applyArm64Imm(off: buf + 12, imm: exitThunkRVA & 0xfff, rangeLimit: 0); |
1156 | |
1157 | Defined *helper = cast<Defined>(Val: file->symtab.ctx.config.arm64ECIcallHelper); |
1158 | if (extended) { |
1159 | // Replace last instruction with an inline range extension thunk. |
1160 | memcpy(dest: buf + 16, src: arm64Thunk, n: sizeof(arm64Thunk)); |
1161 | applyArm64Addr(off: buf + 16, s: helper->getRVA(), p: rva + 16, shift: 12); |
1162 | applyArm64Imm(off: buf + 20, imm: helper->getRVA() & 0xfff, rangeLimit: 0); |
1163 | } else { |
1164 | applyArm64Branch26(off: buf + 16, v: helper->getRVA() - rva - 16); |
1165 | } |
1166 | } |
1167 | |
1168 | bool ImportThunkChunkARM64EC::verifyRanges() { |
1169 | if (extended) |
1170 | return true; |
1171 | auto helper = cast<Defined>(Val: file->symtab.ctx.config.arm64ECIcallHelper); |
1172 | return isInt<28>(x: helper->getRVA() - rva - 16); |
1173 | } |
1174 | |
1175 | uint32_t ImportThunkChunkARM64EC::extendRanges() { |
1176 | if (extended || verifyRanges()) |
1177 | return 0; |
1178 | extended = true; |
1179 | // The last instruction is replaced with an inline range extension thunk. |
1180 | return sizeof(arm64Thunk) - sizeof(uint32_t); |
1181 | } |
1182 | |
1183 | uint64_t Arm64XRelocVal::get() const { |
1184 | return (sym ? sym->getRVA() : 0) + (chunk ? chunk->getRVA() : 0) + value; |
1185 | } |
1186 | |
1187 | size_t Arm64XDynamicRelocEntry::getSize() const { |
1188 | switch (type) { |
1189 | case IMAGE_DVRT_ARM64X_FIXUP_TYPE_ZEROFILL: |
1190 | return sizeof(uint16_t); // Just a header. |
1191 | case IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE: |
1192 | return sizeof(uint16_t) + size; // A header and a payload. |
1193 | case IMAGE_DVRT_ARM64X_FIXUP_TYPE_DELTA: |
1194 | return 2 * sizeof(uint16_t); // A header and a delta. |
1195 | } |
1196 | llvm_unreachable("invalid type" ); |
1197 | } |
1198 | |
1199 | void Arm64XDynamicRelocEntry::writeTo(uint8_t *buf) const { |
1200 | auto out = reinterpret_cast<ulittle16_t *>(buf); |
1201 | *out = (offset.get() & 0xfff) | (type << 12); |
1202 | |
1203 | switch (type) { |
1204 | case IMAGE_DVRT_ARM64X_FIXUP_TYPE_ZEROFILL: |
1205 | *out |= ((bit_width(Value: size) - 1) << 14); // Encode the size. |
1206 | break; |
1207 | case IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE: |
1208 | *out |= ((bit_width(Value: size) - 1) << 14); // Encode the size. |
1209 | switch (size) { |
1210 | case 2: |
1211 | out[1] = value.get(); |
1212 | break; |
1213 | case 4: |
1214 | *reinterpret_cast<ulittle32_t *>(out + 1) = value.get(); |
1215 | break; |
1216 | case 8: |
1217 | *reinterpret_cast<ulittle64_t *>(out + 1) = value.get(); |
1218 | break; |
1219 | default: |
1220 | llvm_unreachable("invalid size" ); |
1221 | } |
1222 | break; |
1223 | case IMAGE_DVRT_ARM64X_FIXUP_TYPE_DELTA: |
1224 | int delta = value.get(); |
1225 | // Negative offsets use a sign bit in the header. |
1226 | if (delta < 0) { |
1227 | *out |= 1 << 14; |
1228 | delta = -delta; |
1229 | } |
1230 | // Depending on the value, the delta is encoded with a shift of 2 or 3 bits. |
1231 | if (delta & 7) { |
1232 | assert(!(delta & 3)); |
1233 | delta >>= 2; |
1234 | } else { |
1235 | *out |= (1 << 15); |
1236 | delta >>= 3; |
1237 | } |
1238 | out[1] = delta; |
1239 | assert(!(delta & ~0xffff)); |
1240 | break; |
1241 | } |
1242 | } |
1243 | |
1244 | void DynamicRelocsChunk::finalize() { |
1245 | llvm::stable_sort(Range&: arm64xRelocs, C: [=](const Arm64XDynamicRelocEntry &a, |
1246 | const Arm64XDynamicRelocEntry &b) { |
1247 | return a.offset.get() < b.offset.get(); |
1248 | }); |
1249 | |
1250 | size = sizeof(coff_dynamic_reloc_table) + sizeof(coff_dynamic_relocation64); |
1251 | uint32_t prevPage = 0xfff; |
1252 | |
1253 | for (const Arm64XDynamicRelocEntry &entry : arm64xRelocs) { |
1254 | uint32_t page = entry.offset.get() & ~0xfff; |
1255 | if (page != prevPage) { |
1256 | size = alignTo(Value: size, Align: sizeof(uint32_t)) + |
1257 | sizeof(coff_base_reloc_block_header); |
1258 | prevPage = page; |
1259 | } |
1260 | size += entry.getSize(); |
1261 | } |
1262 | |
1263 | size = alignTo(Value: size, Align: sizeof(uint32_t)); |
1264 | } |
1265 | |
1266 | // Set the reloc value. The reloc entry must be allocated beforehand. |
1267 | void DynamicRelocsChunk::set(uint32_t rva, Arm64XRelocVal value) { |
1268 | auto entry = |
1269 | llvm::find_if(Range&: arm64xRelocs, P: [rva](const Arm64XDynamicRelocEntry &e) { |
1270 | return e.offset.get() == rva; |
1271 | }); |
1272 | assert(entry != arm64xRelocs.end()); |
1273 | assert(!entry->value.get()); |
1274 | entry->value = value; |
1275 | } |
1276 | |
1277 | void DynamicRelocsChunk::writeTo(uint8_t *buf) const { |
1278 | auto table = reinterpret_cast<coff_dynamic_reloc_table *>(buf); |
1279 | table->Version = 1; |
1280 | table->Size = sizeof(coff_dynamic_relocation64); |
1281 | buf += sizeof(*table); |
1282 | |
1283 | auto = reinterpret_cast<coff_dynamic_relocation64 *>(buf); |
1284 | header->Symbol = IMAGE_DYNAMIC_RELOCATION_ARM64X; |
1285 | buf += sizeof(*header); |
1286 | |
1287 | coff_base_reloc_block_header * = nullptr; |
1288 | size_t relocSize = 0; |
1289 | for (const Arm64XDynamicRelocEntry &entry : arm64xRelocs) { |
1290 | uint32_t page = entry.offset.get() & ~0xfff; |
1291 | if (!pageHeader || page != pageHeader->PageRVA) { |
1292 | relocSize = alignTo(Value: relocSize, Align: sizeof(uint32_t)); |
1293 | if (pageHeader) |
1294 | pageHeader->BlockSize = |
1295 | buf + relocSize - reinterpret_cast<uint8_t *>(pageHeader); |
1296 | pageHeader = |
1297 | reinterpret_cast<coff_base_reloc_block_header *>(buf + relocSize); |
1298 | pageHeader->PageRVA = page; |
1299 | relocSize += sizeof(*pageHeader); |
1300 | } |
1301 | |
1302 | entry.writeTo(buf: buf + relocSize); |
1303 | relocSize += entry.getSize(); |
1304 | } |
1305 | relocSize = alignTo(Value: relocSize, Align: sizeof(uint32_t)); |
1306 | pageHeader->BlockSize = |
1307 | buf + relocSize - reinterpret_cast<uint8_t *>(pageHeader); |
1308 | |
1309 | header->BaseRelocSize = relocSize; |
1310 | table->Size += relocSize; |
1311 | assert(size == sizeof(*table) + sizeof(*header) + relocSize); |
1312 | } |
1313 | |
1314 | } // namespace lld::coff |
1315 | |