1 | //===- AArch64.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 "InputFiles.h" |
10 | #include "OutputSections.h" |
11 | #include "Symbols.h" |
12 | #include "SyntheticSections.h" |
13 | #include "Target.h" |
14 | #include "lld/Common/ErrorHandler.h" |
15 | #include "llvm/BinaryFormat/ELF.h" |
16 | #include "llvm/Support/Endian.h" |
17 | |
18 | using namespace llvm; |
19 | using namespace llvm::support::endian; |
20 | using namespace llvm::ELF; |
21 | using namespace lld; |
22 | using namespace lld::elf; |
23 | |
24 | // Page(Expr) is the page address of the expression Expr, defined |
25 | // as (Expr & ~0xFFF). (This applies even if the machine page size |
26 | // supported by the platform has a different value.) |
27 | uint64_t elf::getAArch64Page(uint64_t expr) { |
28 | return expr & ~static_cast<uint64_t>(0xFFF); |
29 | } |
30 | |
31 | namespace { |
32 | class AArch64 : public TargetInfo { |
33 | public: |
34 | AArch64(); |
35 | RelExpr getRelExpr(RelType type, const Symbol &s, |
36 | const uint8_t *loc) const override; |
37 | RelType getDynRel(RelType type) const override; |
38 | int64_t getImplicitAddend(const uint8_t *buf, RelType type) const override; |
39 | void writeGotPlt(uint8_t *buf, const Symbol &s) const override; |
40 | void writeIgotPlt(uint8_t *buf, const Symbol &s) const override; |
41 | void writePltHeader(uint8_t *buf) const override; |
42 | void writePlt(uint8_t *buf, const Symbol &sym, |
43 | uint64_t pltEntryAddr) const override; |
44 | bool needsThunk(RelExpr expr, RelType type, const InputFile *file, |
45 | uint64_t branchAddr, const Symbol &s, |
46 | int64_t a) const override; |
47 | uint32_t getThunkSectionSpacing() const override; |
48 | bool inBranchRange(RelType type, uint64_t src, uint64_t dst) const override; |
49 | bool usesOnlyLowPageBits(RelType type) const override; |
50 | void relocate(uint8_t *loc, const Relocation &rel, |
51 | uint64_t val) const override; |
52 | RelExpr adjustTlsExpr(RelType type, RelExpr expr) const override; |
53 | void relocateAlloc(InputSectionBase &sec, uint8_t *buf) const override; |
54 | |
55 | private: |
56 | void relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
57 | void relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
58 | void relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const; |
59 | }; |
60 | |
61 | struct AArch64Relaxer { |
62 | bool safeToRelaxAdrpLdr = false; |
63 | |
64 | AArch64Relaxer(ArrayRef<Relocation> relocs); |
65 | bool tryRelaxAdrpAdd(const Relocation &adrpRel, const Relocation &addRel, |
66 | uint64_t secAddr, uint8_t *buf) const; |
67 | bool tryRelaxAdrpLdr(const Relocation &adrpRel, const Relocation &ldrRel, |
68 | uint64_t secAddr, uint8_t *buf) const; |
69 | }; |
70 | } // namespace |
71 | |
72 | // Return the bits [Start, End] from Val shifted Start bits. |
73 | // For instance, getBits(0xF0, 4, 8) returns 0xF. |
74 | static uint64_t getBits(uint64_t val, int start, int end) { |
75 | uint64_t mask = ((uint64_t)1 << (end + 1 - start)) - 1; |
76 | return (val >> start) & mask; |
77 | } |
78 | |
79 | AArch64::AArch64() { |
80 | copyRel = R_AARCH64_COPY; |
81 | relativeRel = R_AARCH64_RELATIVE; |
82 | iRelativeRel = R_AARCH64_IRELATIVE; |
83 | gotRel = R_AARCH64_GLOB_DAT; |
84 | pltRel = R_AARCH64_JUMP_SLOT; |
85 | symbolicRel = R_AARCH64_ABS64; |
86 | tlsDescRel = R_AARCH64_TLSDESC; |
87 | tlsGotRel = R_AARCH64_TLS_TPREL64; |
88 | pltHeaderSize = 32; |
89 | pltEntrySize = 16; |
90 | ipltEntrySize = 16; |
91 | defaultMaxPageSize = 65536; |
92 | |
93 | // Align to the 2 MiB page size (known as a superpage or huge page). |
94 | // FreeBSD automatically promotes 2 MiB-aligned allocations. |
95 | defaultImageBase = 0x200000; |
96 | |
97 | needsThunks = true; |
98 | } |
99 | |
100 | RelExpr AArch64::getRelExpr(RelType type, const Symbol &s, |
101 | const uint8_t *loc) const { |
102 | switch (type) { |
103 | case R_AARCH64_ABS16: |
104 | case R_AARCH64_ABS32: |
105 | case R_AARCH64_ABS64: |
106 | case R_AARCH64_ADD_ABS_LO12_NC: |
107 | case R_AARCH64_LDST128_ABS_LO12_NC: |
108 | case R_AARCH64_LDST16_ABS_LO12_NC: |
109 | case R_AARCH64_LDST32_ABS_LO12_NC: |
110 | case R_AARCH64_LDST64_ABS_LO12_NC: |
111 | case R_AARCH64_LDST8_ABS_LO12_NC: |
112 | case R_AARCH64_MOVW_SABS_G0: |
113 | case R_AARCH64_MOVW_SABS_G1: |
114 | case R_AARCH64_MOVW_SABS_G2: |
115 | case R_AARCH64_MOVW_UABS_G0: |
116 | case R_AARCH64_MOVW_UABS_G0_NC: |
117 | case R_AARCH64_MOVW_UABS_G1: |
118 | case R_AARCH64_MOVW_UABS_G1_NC: |
119 | case R_AARCH64_MOVW_UABS_G2: |
120 | case R_AARCH64_MOVW_UABS_G2_NC: |
121 | case R_AARCH64_MOVW_UABS_G3: |
122 | return R_ABS; |
123 | case R_AARCH64_AUTH_ABS64: |
124 | return R_AARCH64_AUTH; |
125 | case R_AARCH64_TLSDESC_ADR_PAGE21: |
126 | return R_AARCH64_TLSDESC_PAGE; |
127 | case R_AARCH64_TLSDESC_LD64_LO12: |
128 | case R_AARCH64_TLSDESC_ADD_LO12: |
129 | return R_TLSDESC; |
130 | case R_AARCH64_TLSDESC_CALL: |
131 | return R_TLSDESC_CALL; |
132 | case R_AARCH64_TLSLE_ADD_TPREL_HI12: |
133 | case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC: |
134 | case R_AARCH64_TLSLE_LDST8_TPREL_LO12_NC: |
135 | case R_AARCH64_TLSLE_LDST16_TPREL_LO12_NC: |
136 | case R_AARCH64_TLSLE_LDST32_TPREL_LO12_NC: |
137 | case R_AARCH64_TLSLE_LDST64_TPREL_LO12_NC: |
138 | case R_AARCH64_TLSLE_LDST128_TPREL_LO12_NC: |
139 | case R_AARCH64_TLSLE_MOVW_TPREL_G0: |
140 | case R_AARCH64_TLSLE_MOVW_TPREL_G0_NC: |
141 | case R_AARCH64_TLSLE_MOVW_TPREL_G1: |
142 | case R_AARCH64_TLSLE_MOVW_TPREL_G1_NC: |
143 | case R_AARCH64_TLSLE_MOVW_TPREL_G2: |
144 | return R_TPREL; |
145 | case R_AARCH64_CALL26: |
146 | case R_AARCH64_CONDBR19: |
147 | case R_AARCH64_JUMP26: |
148 | case R_AARCH64_TSTBR14: |
149 | return R_PLT_PC; |
150 | case R_AARCH64_PLT32: |
151 | const_cast<Symbol &>(s).thunkAccessed = true; |
152 | return R_PLT_PC; |
153 | case R_AARCH64_PREL16: |
154 | case R_AARCH64_PREL32: |
155 | case R_AARCH64_PREL64: |
156 | case R_AARCH64_ADR_PREL_LO21: |
157 | case R_AARCH64_LD_PREL_LO19: |
158 | case R_AARCH64_MOVW_PREL_G0: |
159 | case R_AARCH64_MOVW_PREL_G0_NC: |
160 | case R_AARCH64_MOVW_PREL_G1: |
161 | case R_AARCH64_MOVW_PREL_G1_NC: |
162 | case R_AARCH64_MOVW_PREL_G2: |
163 | case R_AARCH64_MOVW_PREL_G2_NC: |
164 | case R_AARCH64_MOVW_PREL_G3: |
165 | return R_PC; |
166 | case R_AARCH64_ADR_PREL_PG_HI21: |
167 | case R_AARCH64_ADR_PREL_PG_HI21_NC: |
168 | return R_AARCH64_PAGE_PC; |
169 | case R_AARCH64_LD64_GOT_LO12_NC: |
170 | case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: |
171 | return R_GOT; |
172 | case R_AARCH64_LD64_GOTPAGE_LO15: |
173 | return R_AARCH64_GOT_PAGE; |
174 | case R_AARCH64_ADR_GOT_PAGE: |
175 | case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21: |
176 | return R_AARCH64_GOT_PAGE_PC; |
177 | case R_AARCH64_GOTPCREL32: |
178 | return R_GOT_PC; |
179 | case R_AARCH64_NONE: |
180 | return R_NONE; |
181 | default: |
182 | error(msg: getErrorLocation(loc) + "unknown relocation (" + Twine(type) + |
183 | ") against symbol " + toString(s)); |
184 | return R_NONE; |
185 | } |
186 | } |
187 | |
188 | RelExpr AArch64::adjustTlsExpr(RelType type, RelExpr expr) const { |
189 | if (expr == R_RELAX_TLS_GD_TO_IE) { |
190 | if (type == R_AARCH64_TLSDESC_ADR_PAGE21) |
191 | return R_AARCH64_RELAX_TLS_GD_TO_IE_PAGE_PC; |
192 | return R_RELAX_TLS_GD_TO_IE_ABS; |
193 | } |
194 | return expr; |
195 | } |
196 | |
197 | bool AArch64::usesOnlyLowPageBits(RelType type) const { |
198 | switch (type) { |
199 | default: |
200 | return false; |
201 | case R_AARCH64_ADD_ABS_LO12_NC: |
202 | case R_AARCH64_LD64_GOT_LO12_NC: |
203 | case R_AARCH64_LDST128_ABS_LO12_NC: |
204 | case R_AARCH64_LDST16_ABS_LO12_NC: |
205 | case R_AARCH64_LDST32_ABS_LO12_NC: |
206 | case R_AARCH64_LDST64_ABS_LO12_NC: |
207 | case R_AARCH64_LDST8_ABS_LO12_NC: |
208 | case R_AARCH64_TLSDESC_ADD_LO12: |
209 | case R_AARCH64_TLSDESC_LD64_LO12: |
210 | case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: |
211 | return true; |
212 | } |
213 | } |
214 | |
215 | RelType AArch64::getDynRel(RelType type) const { |
216 | if (type == R_AARCH64_ABS64 || type == R_AARCH64_AUTH_ABS64) |
217 | return type; |
218 | return R_AARCH64_NONE; |
219 | } |
220 | |
221 | int64_t AArch64::getImplicitAddend(const uint8_t *buf, RelType type) const { |
222 | switch (type) { |
223 | case R_AARCH64_TLSDESC: |
224 | return read64(p: buf + 8); |
225 | case R_AARCH64_NONE: |
226 | case R_AARCH64_GLOB_DAT: |
227 | case R_AARCH64_JUMP_SLOT: |
228 | return 0; |
229 | case R_AARCH64_ABS16: |
230 | case R_AARCH64_PREL16: |
231 | return SignExtend64<16>(x: read16(p: buf)); |
232 | case R_AARCH64_ABS32: |
233 | case R_AARCH64_PREL32: |
234 | return SignExtend64<32>(x: read32(p: buf)); |
235 | case R_AARCH64_ABS64: |
236 | case R_AARCH64_PREL64: |
237 | case R_AARCH64_RELATIVE: |
238 | case R_AARCH64_IRELATIVE: |
239 | case R_AARCH64_TLS_TPREL64: |
240 | return read64(p: buf); |
241 | case R_AARCH64_MOVW_UABS_G0: |
242 | case R_AARCH64_MOVW_UABS_G0_NC: |
243 | return getBits(val: SignExtend64<16>(x: read16(p: buf)), start: 0, end: 15); |
244 | case R_AARCH64_MOVW_UABS_G1: |
245 | case R_AARCH64_MOVW_UABS_G1_NC: |
246 | return getBits(val: SignExtend64<32>(x: read32(p: buf)), start: 16, end: 31); |
247 | case R_AARCH64_MOVW_UABS_G2: |
248 | case R_AARCH64_MOVW_UABS_G2_NC: |
249 | return getBits(val: read64(p: buf), start: 32, end: 47); |
250 | case R_AARCH64_MOVW_UABS_G3: |
251 | return getBits(val: read64(p: buf), start: 48, end: 63); |
252 | case R_AARCH64_TSTBR14: |
253 | return getBits(val: SignExtend64<32>(x: read32(p: buf)), start: 2, end: 15); |
254 | case R_AARCH64_CONDBR19: |
255 | case R_AARCH64_LD_PREL_LO19: |
256 | return getBits(val: SignExtend64<32>(x: read32(p: buf)), start: 2, end: 20); |
257 | case R_AARCH64_ADD_ABS_LO12_NC: |
258 | return getBits(val: SignExtend64<16>(x: read16(p: buf)), start: 0, end: 11); |
259 | case R_AARCH64_ADR_PREL_PG_HI21: |
260 | case R_AARCH64_ADR_PREL_PG_HI21_NC: |
261 | return getBits(val: SignExtend64<32>(x: read32(p: buf)), start: 12, end: 32); |
262 | case R_AARCH64_JUMP26: |
263 | case R_AARCH64_CALL26: |
264 | return getBits(val: SignExtend64<32>(x: read32(p: buf)), start: 2, end: 27); |
265 | default: |
266 | internalLinkerError(loc: getErrorLocation(loc: buf), |
267 | msg: "cannot read addend for relocation " + toString(type)); |
268 | return 0; |
269 | } |
270 | } |
271 | |
272 | void AArch64::writeGotPlt(uint8_t *buf, const Symbol &) const { |
273 | write64(p: buf, v: in.plt->getVA()); |
274 | } |
275 | |
276 | void AArch64::writeIgotPlt(uint8_t *buf, const Symbol &s) const { |
277 | if (config->writeAddends) |
278 | write64(p: buf, v: s.getVA()); |
279 | } |
280 | |
281 | void AArch64::(uint8_t *buf) const { |
282 | const uint8_t pltData[] = { |
283 | 0xf0, 0x7b, 0xbf, 0xa9, // stp x16, x30, [sp,#-16]! |
284 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[2])) |
285 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[2]))] |
286 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.got.plt[2])) |
287 | 0x20, 0x02, 0x1f, 0xd6, // br x17 |
288 | 0x1f, 0x20, 0x03, 0xd5, // nop |
289 | 0x1f, 0x20, 0x03, 0xd5, // nop |
290 | 0x1f, 0x20, 0x03, 0xd5 // nop |
291 | }; |
292 | memcpy(dest: buf, src: pltData, n: sizeof(pltData)); |
293 | |
294 | uint64_t got = in.gotPlt->getVA(); |
295 | uint64_t plt = in.plt->getVA(); |
296 | relocateNoSym(loc: buf + 4, type: R_AARCH64_ADR_PREL_PG_HI21, |
297 | val: getAArch64Page(expr: got + 16) - getAArch64Page(expr: plt + 4)); |
298 | relocateNoSym(loc: buf + 8, type: R_AARCH64_LDST64_ABS_LO12_NC, val: got + 16); |
299 | relocateNoSym(loc: buf + 12, type: R_AARCH64_ADD_ABS_LO12_NC, val: got + 16); |
300 | } |
301 | |
302 | void AArch64::writePlt(uint8_t *buf, const Symbol &sym, |
303 | uint64_t pltEntryAddr) const { |
304 | const uint8_t inst[] = { |
305 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[n])) |
306 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[n]))] |
307 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.got.plt[n])) |
308 | 0x20, 0x02, 0x1f, 0xd6 // br x17 |
309 | }; |
310 | memcpy(dest: buf, src: inst, n: sizeof(inst)); |
311 | |
312 | uint64_t gotPltEntryAddr = sym.getGotPltVA(); |
313 | relocateNoSym(loc: buf, type: R_AARCH64_ADR_PREL_PG_HI21, |
314 | val: getAArch64Page(expr: gotPltEntryAddr) - getAArch64Page(expr: pltEntryAddr)); |
315 | relocateNoSym(loc: buf + 4, type: R_AARCH64_LDST64_ABS_LO12_NC, val: gotPltEntryAddr); |
316 | relocateNoSym(loc: buf + 8, type: R_AARCH64_ADD_ABS_LO12_NC, val: gotPltEntryAddr); |
317 | } |
318 | |
319 | bool AArch64::needsThunk(RelExpr expr, RelType type, const InputFile *file, |
320 | uint64_t branchAddr, const Symbol &s, |
321 | int64_t a) const { |
322 | // If s is an undefined weak symbol and does not have a PLT entry then it will |
323 | // be resolved as a branch to the next instruction. If it is hidden, its |
324 | // binding has been converted to local, so we just check isUndefined() here. A |
325 | // undefined non-weak symbol will have been errored. |
326 | if (s.isUndefined() && !s.isInPlt()) |
327 | return false; |
328 | // ELF for the ARM 64-bit architecture, section Call and Jump relocations |
329 | // only permits range extension thunks for R_AARCH64_CALL26 and |
330 | // R_AARCH64_JUMP26 relocation types. |
331 | if (type != R_AARCH64_CALL26 && type != R_AARCH64_JUMP26 && |
332 | type != R_AARCH64_PLT32) |
333 | return false; |
334 | uint64_t dst = expr == R_PLT_PC ? s.getPltVA() : s.getVA(addend: a); |
335 | return !inBranchRange(type, src: branchAddr, dst); |
336 | } |
337 | |
338 | uint32_t AArch64::getThunkSectionSpacing() const { |
339 | // See comment in Arch/ARM.cpp for a more detailed explanation of |
340 | // getThunkSectionSpacing(). For AArch64 the only branches we are permitted to |
341 | // Thunk have a range of +/- 128 MiB |
342 | return (128 * 1024 * 1024) - 0x30000; |
343 | } |
344 | |
345 | bool AArch64::inBranchRange(RelType type, uint64_t src, uint64_t dst) const { |
346 | if (type != R_AARCH64_CALL26 && type != R_AARCH64_JUMP26 && |
347 | type != R_AARCH64_PLT32) |
348 | return true; |
349 | // The AArch64 call and unconditional branch instructions have a range of |
350 | // +/- 128 MiB. The PLT32 relocation supports a range up to +/- 2 GiB. |
351 | uint64_t range = |
352 | type == R_AARCH64_PLT32 ? (UINT64_C(1) << 31) : (128 * 1024 * 1024); |
353 | if (dst > src) { |
354 | // Immediate of branch is signed. |
355 | range -= 4; |
356 | return dst - src <= range; |
357 | } |
358 | return src - dst <= range; |
359 | } |
360 | |
361 | static void write32AArch64Addr(uint8_t *l, uint64_t imm) { |
362 | uint32_t immLo = (imm & 0x3) << 29; |
363 | uint32_t immHi = (imm & 0x1FFFFC) << 3; |
364 | uint64_t mask = (0x3 << 29) | (0x1FFFFC << 3); |
365 | write32le(P: l, V: (read32le(P: l) & ~mask) | immLo | immHi); |
366 | } |
367 | |
368 | static void or32le(uint8_t *p, int32_t v) { write32le(P: p, V: read32le(P: p) | v); } |
369 | |
370 | // Update the immediate field in a AARCH64 ldr, str, and add instruction. |
371 | static void or32AArch64Imm(uint8_t *l, uint64_t imm) { |
372 | or32le(p: l, v: (imm & 0xFFF) << 10); |
373 | } |
374 | |
375 | // Update the immediate field in an AArch64 movk, movn or movz instruction |
376 | // for a signed relocation, and update the opcode of a movn or movz instruction |
377 | // to match the sign of the operand. |
378 | static void writeSMovWImm(uint8_t *loc, uint32_t imm) { |
379 | uint32_t inst = read32le(P: loc); |
380 | // Opcode field is bits 30, 29, with 10 = movz, 00 = movn and 11 = movk. |
381 | if (!(inst & (1 << 29))) { |
382 | // movn or movz. |
383 | if (imm & 0x10000) { |
384 | // Change opcode to movn, which takes an inverted operand. |
385 | imm ^= 0xFFFF; |
386 | inst &= ~(1 << 30); |
387 | } else { |
388 | // Change opcode to movz. |
389 | inst |= 1 << 30; |
390 | } |
391 | } |
392 | write32le(P: loc, V: inst | ((imm & 0xFFFF) << 5)); |
393 | } |
394 | |
395 | void AArch64::relocate(uint8_t *loc, const Relocation &rel, |
396 | uint64_t val) const { |
397 | switch (rel.type) { |
398 | case R_AARCH64_ABS16: |
399 | case R_AARCH64_PREL16: |
400 | checkIntUInt(loc, v: val, n: 16, rel); |
401 | write16(p: loc, v: val); |
402 | break; |
403 | case R_AARCH64_ABS32: |
404 | case R_AARCH64_PREL32: |
405 | checkIntUInt(loc, v: val, n: 32, rel); |
406 | write32(p: loc, v: val); |
407 | break; |
408 | case R_AARCH64_PLT32: |
409 | case R_AARCH64_GOTPCREL32: |
410 | checkInt(loc, v: val, n: 32, rel); |
411 | write32(p: loc, v: val); |
412 | break; |
413 | case R_AARCH64_ABS64: |
414 | // AArch64 relocations to tagged symbols have extended semantics, as |
415 | // described here: |
416 | // https://github.com/ARM-software/abi-aa/blob/main/memtagabielf64/memtagabielf64.rst#841extended-semantics-of-r_aarch64_relative. |
417 | // tl;dr: encode the symbol's special addend in the place, which is an |
418 | // offset to the point where the logical tag is derived from. Quick hack, if |
419 | // the addend is within the symbol's bounds, no need to encode the tag |
420 | // derivation offset. |
421 | if (rel.sym && rel.sym->isTagged() && |
422 | (rel.addend < 0 || |
423 | rel.addend >= static_cast<int64_t>(rel.sym->getSize()))) |
424 | write64(p: loc, v: -rel.addend); |
425 | else |
426 | write64(p: loc, v: val); |
427 | break; |
428 | case R_AARCH64_PREL64: |
429 | write64(p: loc, v: val); |
430 | break; |
431 | case R_AARCH64_ADD_ABS_LO12_NC: |
432 | or32AArch64Imm(l: loc, imm: val); |
433 | break; |
434 | case R_AARCH64_ADR_GOT_PAGE: |
435 | case R_AARCH64_ADR_PREL_PG_HI21: |
436 | case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21: |
437 | case R_AARCH64_TLSDESC_ADR_PAGE21: |
438 | checkInt(loc, v: val, n: 33, rel); |
439 | [[fallthrough]]; |
440 | case R_AARCH64_ADR_PREL_PG_HI21_NC: |
441 | write32AArch64Addr(l: loc, imm: val >> 12); |
442 | break; |
443 | case R_AARCH64_ADR_PREL_LO21: |
444 | checkInt(loc, v: val, n: 21, rel); |
445 | write32AArch64Addr(l: loc, imm: val); |
446 | break; |
447 | case R_AARCH64_JUMP26: |
448 | // Normally we would just write the bits of the immediate field, however |
449 | // when patching instructions for the cpu errata fix -fix-cortex-a53-843419 |
450 | // we want to replace a non-branch instruction with a branch immediate |
451 | // instruction. By writing all the bits of the instruction including the |
452 | // opcode and the immediate (0 001 | 01 imm26) we can do this |
453 | // transformation by placing a R_AARCH64_JUMP26 relocation at the offset of |
454 | // the instruction we want to patch. |
455 | write32le(P: loc, V: 0x14000000); |
456 | [[fallthrough]]; |
457 | case R_AARCH64_CALL26: |
458 | checkInt(loc, v: val, n: 28, rel); |
459 | or32le(p: loc, v: (val & 0x0FFFFFFC) >> 2); |
460 | break; |
461 | case R_AARCH64_CONDBR19: |
462 | case R_AARCH64_LD_PREL_LO19: |
463 | checkAlignment(loc, v: val, n: 4, rel); |
464 | checkInt(loc, v: val, n: 21, rel); |
465 | or32le(p: loc, v: (val & 0x1FFFFC) << 3); |
466 | break; |
467 | case R_AARCH64_LDST8_ABS_LO12_NC: |
468 | case R_AARCH64_TLSLE_LDST8_TPREL_LO12_NC: |
469 | or32AArch64Imm(l: loc, imm: getBits(val, start: 0, end: 11)); |
470 | break; |
471 | case R_AARCH64_LDST16_ABS_LO12_NC: |
472 | case R_AARCH64_TLSLE_LDST16_TPREL_LO12_NC: |
473 | checkAlignment(loc, v: val, n: 2, rel); |
474 | or32AArch64Imm(l: loc, imm: getBits(val, start: 1, end: 11)); |
475 | break; |
476 | case R_AARCH64_LDST32_ABS_LO12_NC: |
477 | case R_AARCH64_TLSLE_LDST32_TPREL_LO12_NC: |
478 | checkAlignment(loc, v: val, n: 4, rel); |
479 | or32AArch64Imm(l: loc, imm: getBits(val, start: 2, end: 11)); |
480 | break; |
481 | case R_AARCH64_LDST64_ABS_LO12_NC: |
482 | case R_AARCH64_LD64_GOT_LO12_NC: |
483 | case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: |
484 | case R_AARCH64_TLSLE_LDST64_TPREL_LO12_NC: |
485 | case R_AARCH64_TLSDESC_LD64_LO12: |
486 | checkAlignment(loc, v: val, n: 8, rel); |
487 | or32AArch64Imm(l: loc, imm: getBits(val, start: 3, end: 11)); |
488 | break; |
489 | case R_AARCH64_LDST128_ABS_LO12_NC: |
490 | case R_AARCH64_TLSLE_LDST128_TPREL_LO12_NC: |
491 | checkAlignment(loc, v: val, n: 16, rel); |
492 | or32AArch64Imm(l: loc, imm: getBits(val, start: 4, end: 11)); |
493 | break; |
494 | case R_AARCH64_LD64_GOTPAGE_LO15: |
495 | checkAlignment(loc, v: val, n: 8, rel); |
496 | or32AArch64Imm(l: loc, imm: getBits(val, start: 3, end: 14)); |
497 | break; |
498 | case R_AARCH64_MOVW_UABS_G0: |
499 | checkUInt(loc, v: val, n: 16, rel); |
500 | [[fallthrough]]; |
501 | case R_AARCH64_MOVW_UABS_G0_NC: |
502 | or32le(p: loc, v: (val & 0xFFFF) << 5); |
503 | break; |
504 | case R_AARCH64_MOVW_UABS_G1: |
505 | checkUInt(loc, v: val, n: 32, rel); |
506 | [[fallthrough]]; |
507 | case R_AARCH64_MOVW_UABS_G1_NC: |
508 | or32le(p: loc, v: (val & 0xFFFF0000) >> 11); |
509 | break; |
510 | case R_AARCH64_MOVW_UABS_G2: |
511 | checkUInt(loc, v: val, n: 48, rel); |
512 | [[fallthrough]]; |
513 | case R_AARCH64_MOVW_UABS_G2_NC: |
514 | or32le(p: loc, v: (val & 0xFFFF00000000) >> 27); |
515 | break; |
516 | case R_AARCH64_MOVW_UABS_G3: |
517 | or32le(p: loc, v: (val & 0xFFFF000000000000) >> 43); |
518 | break; |
519 | case R_AARCH64_MOVW_PREL_G0: |
520 | case R_AARCH64_MOVW_SABS_G0: |
521 | case R_AARCH64_TLSLE_MOVW_TPREL_G0: |
522 | checkInt(loc, v: val, n: 17, rel); |
523 | [[fallthrough]]; |
524 | case R_AARCH64_MOVW_PREL_G0_NC: |
525 | case R_AARCH64_TLSLE_MOVW_TPREL_G0_NC: |
526 | writeSMovWImm(loc, imm: val); |
527 | break; |
528 | case R_AARCH64_MOVW_PREL_G1: |
529 | case R_AARCH64_MOVW_SABS_G1: |
530 | case R_AARCH64_TLSLE_MOVW_TPREL_G1: |
531 | checkInt(loc, v: val, n: 33, rel); |
532 | [[fallthrough]]; |
533 | case R_AARCH64_MOVW_PREL_G1_NC: |
534 | case R_AARCH64_TLSLE_MOVW_TPREL_G1_NC: |
535 | writeSMovWImm(loc, imm: val >> 16); |
536 | break; |
537 | case R_AARCH64_MOVW_PREL_G2: |
538 | case R_AARCH64_MOVW_SABS_G2: |
539 | case R_AARCH64_TLSLE_MOVW_TPREL_G2: |
540 | checkInt(loc, v: val, n: 49, rel); |
541 | [[fallthrough]]; |
542 | case R_AARCH64_MOVW_PREL_G2_NC: |
543 | writeSMovWImm(loc, imm: val >> 32); |
544 | break; |
545 | case R_AARCH64_MOVW_PREL_G3: |
546 | writeSMovWImm(loc, imm: val >> 48); |
547 | break; |
548 | case R_AARCH64_TSTBR14: |
549 | checkInt(loc, v: val, n: 16, rel); |
550 | or32le(p: loc, v: (val & 0xFFFC) << 3); |
551 | break; |
552 | case R_AARCH64_TLSLE_ADD_TPREL_HI12: |
553 | checkUInt(loc, v: val, n: 24, rel); |
554 | or32AArch64Imm(l: loc, imm: val >> 12); |
555 | break; |
556 | case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC: |
557 | case R_AARCH64_TLSDESC_ADD_LO12: |
558 | or32AArch64Imm(l: loc, imm: val); |
559 | break; |
560 | case R_AARCH64_TLSDESC: |
561 | // For R_AARCH64_TLSDESC the addend is stored in the second 64-bit word. |
562 | write64(p: loc + 8, v: val); |
563 | break; |
564 | default: |
565 | llvm_unreachable("unknown relocation" ); |
566 | } |
567 | } |
568 | |
569 | void AArch64::relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, |
570 | uint64_t val) const { |
571 | // TLSDESC Global-Dynamic relocation are in the form: |
572 | // adrp x0, :tlsdesc:v [R_AARCH64_TLSDESC_ADR_PAGE21] |
573 | // ldr x1, [x0, #:tlsdesc_lo12:v [R_AARCH64_TLSDESC_LD64_LO12] |
574 | // add x0, x0, :tlsdesc_los:v [R_AARCH64_TLSDESC_ADD_LO12] |
575 | // .tlsdesccall [R_AARCH64_TLSDESC_CALL] |
576 | // blr x1 |
577 | // And it can optimized to: |
578 | // movz x0, #0x0, lsl #16 |
579 | // movk x0, #0x10 |
580 | // nop |
581 | // nop |
582 | checkUInt(loc, v: val, n: 32, rel); |
583 | |
584 | switch (rel.type) { |
585 | case R_AARCH64_TLSDESC_ADD_LO12: |
586 | case R_AARCH64_TLSDESC_CALL: |
587 | write32le(P: loc, V: 0xd503201f); // nop |
588 | return; |
589 | case R_AARCH64_TLSDESC_ADR_PAGE21: |
590 | write32le(P: loc, V: 0xd2a00000 | (((val >> 16) & 0xffff) << 5)); // movz |
591 | return; |
592 | case R_AARCH64_TLSDESC_LD64_LO12: |
593 | write32le(P: loc, V: 0xf2800000 | ((val & 0xffff) << 5)); // movk |
594 | return; |
595 | default: |
596 | llvm_unreachable("unsupported relocation for TLS GD to LE relaxation" ); |
597 | } |
598 | } |
599 | |
600 | void AArch64::relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, |
601 | uint64_t val) const { |
602 | // TLSDESC Global-Dynamic relocation are in the form: |
603 | // adrp x0, :tlsdesc:v [R_AARCH64_TLSDESC_ADR_PAGE21] |
604 | // ldr x1, [x0, #:tlsdesc_lo12:v [R_AARCH64_TLSDESC_LD64_LO12] |
605 | // add x0, x0, :tlsdesc_los:v [R_AARCH64_TLSDESC_ADD_LO12] |
606 | // .tlsdesccall [R_AARCH64_TLSDESC_CALL] |
607 | // blr x1 |
608 | // And it can optimized to: |
609 | // adrp x0, :gottprel:v |
610 | // ldr x0, [x0, :gottprel_lo12:v] |
611 | // nop |
612 | // nop |
613 | |
614 | switch (rel.type) { |
615 | case R_AARCH64_TLSDESC_ADD_LO12: |
616 | case R_AARCH64_TLSDESC_CALL: |
617 | write32le(P: loc, V: 0xd503201f); // nop |
618 | break; |
619 | case R_AARCH64_TLSDESC_ADR_PAGE21: |
620 | write32le(P: loc, V: 0x90000000); // adrp |
621 | relocateNoSym(loc, type: R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21, val); |
622 | break; |
623 | case R_AARCH64_TLSDESC_LD64_LO12: |
624 | write32le(P: loc, V: 0xf9400000); // ldr |
625 | relocateNoSym(loc, type: R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC, val); |
626 | break; |
627 | default: |
628 | llvm_unreachable("unsupported relocation for TLS GD to LE relaxation" ); |
629 | } |
630 | } |
631 | |
632 | void AArch64::relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, |
633 | uint64_t val) const { |
634 | checkUInt(loc, v: val, n: 32, rel); |
635 | |
636 | if (rel.type == R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21) { |
637 | // Generate MOVZ. |
638 | uint32_t regNo = read32le(P: loc) & 0x1f; |
639 | write32le(P: loc, V: (0xd2a00000 | regNo) | (((val >> 16) & 0xffff) << 5)); |
640 | return; |
641 | } |
642 | if (rel.type == R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC) { |
643 | // Generate MOVK. |
644 | uint32_t regNo = read32le(P: loc) & 0x1f; |
645 | write32le(P: loc, V: (0xf2800000 | regNo) | ((val & 0xffff) << 5)); |
646 | return; |
647 | } |
648 | llvm_unreachable("invalid relocation for TLS IE to LE relaxation" ); |
649 | } |
650 | |
651 | AArch64Relaxer::AArch64Relaxer(ArrayRef<Relocation> relocs) { |
652 | if (!config->relax) |
653 | return; |
654 | // Check if R_AARCH64_ADR_GOT_PAGE and R_AARCH64_LD64_GOT_LO12_NC |
655 | // always appear in pairs. |
656 | size_t i = 0; |
657 | const size_t size = relocs.size(); |
658 | for (; i != size; ++i) { |
659 | if (relocs[i].type == R_AARCH64_ADR_GOT_PAGE) { |
660 | if (i + 1 < size && relocs[i + 1].type == R_AARCH64_LD64_GOT_LO12_NC) { |
661 | ++i; |
662 | continue; |
663 | } |
664 | break; |
665 | } else if (relocs[i].type == R_AARCH64_LD64_GOT_LO12_NC) { |
666 | break; |
667 | } |
668 | } |
669 | safeToRelaxAdrpLdr = i == size; |
670 | } |
671 | |
672 | bool AArch64Relaxer::tryRelaxAdrpAdd(const Relocation &adrpRel, |
673 | const Relocation &addRel, uint64_t secAddr, |
674 | uint8_t *buf) const { |
675 | // When the address of sym is within the range of ADR then |
676 | // we may relax |
677 | // ADRP xn, sym |
678 | // ADD xn, xn, :lo12: sym |
679 | // to |
680 | // NOP |
681 | // ADR xn, sym |
682 | if (!config->relax || adrpRel.type != R_AARCH64_ADR_PREL_PG_HI21 || |
683 | addRel.type != R_AARCH64_ADD_ABS_LO12_NC) |
684 | return false; |
685 | // Check if the relocations apply to consecutive instructions. |
686 | if (adrpRel.offset + 4 != addRel.offset) |
687 | return false; |
688 | if (adrpRel.sym != addRel.sym) |
689 | return false; |
690 | if (adrpRel.addend != 0 || addRel.addend != 0) |
691 | return false; |
692 | |
693 | uint32_t adrpInstr = read32le(P: buf + adrpRel.offset); |
694 | uint32_t addInstr = read32le(P: buf + addRel.offset); |
695 | // Check if the first instruction is ADRP and the second instruction is ADD. |
696 | if ((adrpInstr & 0x9f000000) != 0x90000000 || |
697 | (addInstr & 0xffc00000) != 0x91000000) |
698 | return false; |
699 | uint32_t adrpDestReg = adrpInstr & 0x1f; |
700 | uint32_t addDestReg = addInstr & 0x1f; |
701 | uint32_t addSrcReg = (addInstr >> 5) & 0x1f; |
702 | if (adrpDestReg != addDestReg || adrpDestReg != addSrcReg) |
703 | return false; |
704 | |
705 | Symbol &sym = *adrpRel.sym; |
706 | // Check if the address difference is within 1MiB range. |
707 | int64_t val = sym.getVA() - (secAddr + addRel.offset); |
708 | if (val < -1024 * 1024 || val >= 1024 * 1024) |
709 | return false; |
710 | |
711 | Relocation adrRel = {.expr: R_ABS, .type: R_AARCH64_ADR_PREL_LO21, .offset: addRel.offset, |
712 | /*addend=*/0, .sym: &sym}; |
713 | // nop |
714 | write32le(P: buf + adrpRel.offset, V: 0xd503201f); |
715 | // adr x_<dest_reg> |
716 | write32le(P: buf + adrRel.offset, V: 0x10000000 | adrpDestReg); |
717 | target->relocate(loc: buf + adrRel.offset, rel: adrRel, val); |
718 | return true; |
719 | } |
720 | |
721 | bool AArch64Relaxer::tryRelaxAdrpLdr(const Relocation &adrpRel, |
722 | const Relocation &ldrRel, uint64_t secAddr, |
723 | uint8_t *buf) const { |
724 | if (!safeToRelaxAdrpLdr) |
725 | return false; |
726 | |
727 | // When the definition of sym is not preemptible then we may |
728 | // be able to relax |
729 | // ADRP xn, :got: sym |
730 | // LDR xn, [ xn :got_lo12: sym] |
731 | // to |
732 | // ADRP xn, sym |
733 | // ADD xn, xn, :lo_12: sym |
734 | |
735 | if (adrpRel.type != R_AARCH64_ADR_GOT_PAGE || |
736 | ldrRel.type != R_AARCH64_LD64_GOT_LO12_NC) |
737 | return false; |
738 | // Check if the relocations apply to consecutive instructions. |
739 | if (adrpRel.offset + 4 != ldrRel.offset) |
740 | return false; |
741 | // Check if the relocations reference the same symbol and |
742 | // skip undefined, preemptible and STT_GNU_IFUNC symbols. |
743 | if (!adrpRel.sym || adrpRel.sym != ldrRel.sym || !adrpRel.sym->isDefined() || |
744 | adrpRel.sym->isPreemptible || adrpRel.sym->isGnuIFunc()) |
745 | return false; |
746 | // Check if the addends of the both relocations are zero. |
747 | if (adrpRel.addend != 0 || ldrRel.addend != 0) |
748 | return false; |
749 | uint32_t adrpInstr = read32le(P: buf + adrpRel.offset); |
750 | uint32_t ldrInstr = read32le(P: buf + ldrRel.offset); |
751 | // Check if the first instruction is ADRP and the second instruction is LDR. |
752 | if ((adrpInstr & 0x9f000000) != 0x90000000 || |
753 | (ldrInstr & 0x3b000000) != 0x39000000) |
754 | return false; |
755 | // Check the value of the sf bit. |
756 | if (!(ldrInstr >> 31)) |
757 | return false; |
758 | uint32_t adrpDestReg = adrpInstr & 0x1f; |
759 | uint32_t ldrDestReg = ldrInstr & 0x1f; |
760 | uint32_t ldrSrcReg = (ldrInstr >> 5) & 0x1f; |
761 | // Check if ADPR and LDR use the same register. |
762 | if (adrpDestReg != ldrDestReg || adrpDestReg != ldrSrcReg) |
763 | return false; |
764 | |
765 | Symbol &sym = *adrpRel.sym; |
766 | // GOT references to absolute symbols can't be relaxed to use ADRP/ADD in |
767 | // position-independent code because these instructions produce a relative |
768 | // address. |
769 | if (config->isPic && !cast<Defined>(Val&: sym).section) |
770 | return false; |
771 | // Check if the address difference is within 4GB range. |
772 | int64_t val = |
773 | getAArch64Page(expr: sym.getVA()) - getAArch64Page(expr: secAddr + adrpRel.offset); |
774 | if (val != llvm::SignExtend64(X: val, B: 33)) |
775 | return false; |
776 | |
777 | Relocation adrpSymRel = {.expr: R_AARCH64_PAGE_PC, .type: R_AARCH64_ADR_PREL_PG_HI21, |
778 | .offset: adrpRel.offset, /*addend=*/0, .sym: &sym}; |
779 | Relocation addRel = {.expr: R_ABS, .type: R_AARCH64_ADD_ABS_LO12_NC, .offset: ldrRel.offset, |
780 | /*addend=*/0, .sym: &sym}; |
781 | |
782 | // adrp x_<dest_reg> |
783 | write32le(P: buf + adrpSymRel.offset, V: 0x90000000 | adrpDestReg); |
784 | // add x_<dest reg>, x_<dest reg> |
785 | write32le(P: buf + addRel.offset, V: 0x91000000 | adrpDestReg | (adrpDestReg << 5)); |
786 | |
787 | target->relocate(loc: buf + adrpSymRel.offset, rel: adrpSymRel, |
788 | val: SignExtend64(X: getAArch64Page(expr: sym.getVA()) - |
789 | getAArch64Page(expr: secAddr + adrpSymRel.offset), |
790 | B: 64)); |
791 | target->relocate(loc: buf + addRel.offset, rel: addRel, val: SignExtend64(X: sym.getVA(), B: 64)); |
792 | tryRelaxAdrpAdd(adrpRel: adrpSymRel, addRel, secAddr, buf); |
793 | return true; |
794 | } |
795 | |
796 | // Tagged symbols have upper address bits that are added by the dynamic loader, |
797 | // and thus need the full 64-bit GOT entry. Do not relax such symbols. |
798 | static bool needsGotForMemtag(const Relocation &rel) { |
799 | return rel.sym->isTagged() && needsGot(expr: rel.expr); |
800 | } |
801 | |
802 | void AArch64::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const { |
803 | uint64_t secAddr = sec.getOutputSection()->addr; |
804 | if (auto *s = dyn_cast<InputSection>(Val: &sec)) |
805 | secAddr += s->outSecOff; |
806 | else if (auto *ehIn = dyn_cast<EhInputSection>(Val: &sec)) |
807 | secAddr += ehIn->getParent()->outSecOff; |
808 | AArch64Relaxer relaxer(sec.relocs()); |
809 | for (size_t i = 0, size = sec.relocs().size(); i != size; ++i) { |
810 | const Relocation &rel = sec.relocs()[i]; |
811 | uint8_t *loc = buf + rel.offset; |
812 | const uint64_t val = |
813 | sec.getRelocTargetVA(File: sec.file, Type: rel.type, A: rel.addend, |
814 | P: secAddr + rel.offset, Sym: *rel.sym, Expr: rel.expr); |
815 | |
816 | if (needsGotForMemtag(rel)) { |
817 | relocate(loc, rel, val); |
818 | continue; |
819 | } |
820 | |
821 | switch (rel.expr) { |
822 | case R_AARCH64_GOT_PAGE_PC: |
823 | if (i + 1 < size && |
824 | relaxer.tryRelaxAdrpLdr(adrpRel: rel, ldrRel: sec.relocs()[i + 1], secAddr, buf)) { |
825 | ++i; |
826 | continue; |
827 | } |
828 | break; |
829 | case R_AARCH64_PAGE_PC: |
830 | if (i + 1 < size && |
831 | relaxer.tryRelaxAdrpAdd(adrpRel: rel, addRel: sec.relocs()[i + 1], secAddr, buf)) { |
832 | ++i; |
833 | continue; |
834 | } |
835 | break; |
836 | case R_AARCH64_RELAX_TLS_GD_TO_IE_PAGE_PC: |
837 | case R_RELAX_TLS_GD_TO_IE_ABS: |
838 | relaxTlsGdToIe(loc, rel, val); |
839 | continue; |
840 | case R_RELAX_TLS_GD_TO_LE: |
841 | relaxTlsGdToLe(loc, rel, val); |
842 | continue; |
843 | case R_RELAX_TLS_IE_TO_LE: |
844 | relaxTlsIeToLe(loc, rel, val); |
845 | continue; |
846 | default: |
847 | break; |
848 | } |
849 | relocate(loc, rel, val); |
850 | } |
851 | } |
852 | |
853 | // AArch64 may use security features in variant PLT sequences. These are: |
854 | // Pointer Authentication (PAC), introduced in armv8.3-a and Branch Target |
855 | // Indicator (BTI) introduced in armv8.5-a. The additional instructions used |
856 | // in the variant Plt sequences are encoded in the Hint space so they can be |
857 | // deployed on older architectures, which treat the instructions as a nop. |
858 | // PAC and BTI can be combined leading to the following combinations: |
859 | // writePltHeader |
860 | // writePltHeaderBti (no PAC Header needed) |
861 | // writePlt |
862 | // writePltBti (BTI only) |
863 | // writePltPac (PAC only) |
864 | // writePltBtiPac (BTI and PAC) |
865 | // |
866 | // When PAC is enabled the dynamic loader encrypts the address that it places |
867 | // in the .got.plt using the pacia1716 instruction which encrypts the value in |
868 | // x17 using the modifier in x16. The static linker places autia1716 before the |
869 | // indirect branch to x17 to authenticate the address in x17 with the modifier |
870 | // in x16. This makes it more difficult for an attacker to modify the value in |
871 | // the .got.plt. |
872 | // |
873 | // When BTI is enabled all indirect branches must land on a bti instruction. |
874 | // The static linker must place a bti instruction at the start of any PLT entry |
875 | // that may be the target of an indirect branch. As the PLT entries call the |
876 | // lazy resolver indirectly this must have a bti instruction at start. In |
877 | // general a bti instruction is not needed for a PLT entry as indirect calls |
878 | // are resolved to the function address and not the PLT entry for the function. |
879 | // There are a small number of cases where the PLT address can escape, such as |
880 | // taking the address of a function or ifunc via a non got-generating |
881 | // relocation, and a shared library refers to that symbol. |
882 | // |
883 | // We use the bti c variant of the instruction which permits indirect branches |
884 | // (br) via x16/x17 and indirect function calls (blr) via any register. The ABI |
885 | // guarantees that all indirect branches from code requiring BTI protection |
886 | // will go via x16/x17 |
887 | |
888 | namespace { |
889 | class AArch64BtiPac final : public AArch64 { |
890 | public: |
891 | AArch64BtiPac(); |
892 | void writePltHeader(uint8_t *buf) const override; |
893 | void writePlt(uint8_t *buf, const Symbol &sym, |
894 | uint64_t pltEntryAddr) const override; |
895 | |
896 | private: |
897 | bool ; // bti instruction needed in PLT Header and Entry |
898 | bool pacEntry; // autia1716 instruction needed in PLT Entry |
899 | }; |
900 | } // namespace |
901 | |
902 | AArch64BtiPac::AArch64BtiPac() { |
903 | btiHeader = (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI); |
904 | // A BTI (Branch Target Indicator) Plt Entry is only required if the |
905 | // address of the PLT entry can be taken by the program, which permits an |
906 | // indirect jump to the PLT entry. This can happen when the address |
907 | // of the PLT entry for a function is canonicalised due to the address of |
908 | // the function in an executable being taken by a shared library, or |
909 | // non-preemptible ifunc referenced by non-GOT-generating, non-PLT-generating |
910 | // relocations. |
911 | // The PAC PLT entries require dynamic loader support and this isn't known |
912 | // from properties in the objects, so we use the command line flag. |
913 | pacEntry = config->zPacPlt; |
914 | |
915 | if (btiHeader || pacEntry) { |
916 | pltEntrySize = 24; |
917 | ipltEntrySize = 24; |
918 | } |
919 | } |
920 | |
921 | void AArch64BtiPac::(uint8_t *buf) const { |
922 | const uint8_t btiData[] = { 0x5f, 0x24, 0x03, 0xd5 }; // bti c |
923 | const uint8_t pltData[] = { |
924 | 0xf0, 0x7b, 0xbf, 0xa9, // stp x16, x30, [sp,#-16]! |
925 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[2])) |
926 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[2]))] |
927 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.got.plt[2])) |
928 | 0x20, 0x02, 0x1f, 0xd6, // br x17 |
929 | 0x1f, 0x20, 0x03, 0xd5, // nop |
930 | 0x1f, 0x20, 0x03, 0xd5 // nop |
931 | }; |
932 | const uint8_t nopData[] = { 0x1f, 0x20, 0x03, 0xd5 }; // nop |
933 | |
934 | uint64_t got = in.gotPlt->getVA(); |
935 | uint64_t plt = in.plt->getVA(); |
936 | |
937 | if (btiHeader) { |
938 | // PltHeader is called indirectly by plt[N]. Prefix pltData with a BTI C |
939 | // instruction. |
940 | memcpy(dest: buf, src: btiData, n: sizeof(btiData)); |
941 | buf += sizeof(btiData); |
942 | plt += sizeof(btiData); |
943 | } |
944 | memcpy(dest: buf, src: pltData, n: sizeof(pltData)); |
945 | |
946 | relocateNoSym(loc: buf + 4, type: R_AARCH64_ADR_PREL_PG_HI21, |
947 | val: getAArch64Page(expr: got + 16) - getAArch64Page(expr: plt + 8)); |
948 | relocateNoSym(loc: buf + 8, type: R_AARCH64_LDST64_ABS_LO12_NC, val: got + 16); |
949 | relocateNoSym(loc: buf + 12, type: R_AARCH64_ADD_ABS_LO12_NC, val: got + 16); |
950 | if (!btiHeader) |
951 | // We didn't add the BTI c instruction so round out size with NOP. |
952 | memcpy(dest: buf + sizeof(pltData), src: nopData, n: sizeof(nopData)); |
953 | } |
954 | |
955 | void AArch64BtiPac::writePlt(uint8_t *buf, const Symbol &sym, |
956 | uint64_t pltEntryAddr) const { |
957 | // The PLT entry is of the form: |
958 | // [btiData] addrInst (pacBr | stdBr) [nopData] |
959 | const uint8_t btiData[] = { 0x5f, 0x24, 0x03, 0xd5 }; // bti c |
960 | const uint8_t addrInst[] = { |
961 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[n])) |
962 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[n]))] |
963 | 0x10, 0x02, 0x00, 0x91 // add x16, x16, Offset(&(.got.plt[n])) |
964 | }; |
965 | const uint8_t pacBr[] = { |
966 | 0x9f, 0x21, 0x03, 0xd5, // autia1716 |
967 | 0x20, 0x02, 0x1f, 0xd6 // br x17 |
968 | }; |
969 | const uint8_t stdBr[] = { |
970 | 0x20, 0x02, 0x1f, 0xd6, // br x17 |
971 | 0x1f, 0x20, 0x03, 0xd5 // nop |
972 | }; |
973 | const uint8_t nopData[] = { 0x1f, 0x20, 0x03, 0xd5 }; // nop |
974 | |
975 | // NEEDS_COPY indicates a non-ifunc canonical PLT entry whose address may |
976 | // escape to shared objects. isInIplt indicates a non-preemptible ifunc. Its |
977 | // address may escape if referenced by a direct relocation. If relative |
978 | // vtables are used then if the vtable is in a shared object the offsets will |
979 | // be to the PLT entry. The condition is conservative. |
980 | bool hasBti = btiHeader && |
981 | (sym.hasFlag(bit: NEEDS_COPY) || sym.isInIplt || sym.thunkAccessed); |
982 | if (hasBti) { |
983 | memcpy(dest: buf, src: btiData, n: sizeof(btiData)); |
984 | buf += sizeof(btiData); |
985 | pltEntryAddr += sizeof(btiData); |
986 | } |
987 | |
988 | uint64_t gotPltEntryAddr = sym.getGotPltVA(); |
989 | memcpy(dest: buf, src: addrInst, n: sizeof(addrInst)); |
990 | relocateNoSym(loc: buf, type: R_AARCH64_ADR_PREL_PG_HI21, |
991 | val: getAArch64Page(expr: gotPltEntryAddr) - getAArch64Page(expr: pltEntryAddr)); |
992 | relocateNoSym(loc: buf + 4, type: R_AARCH64_LDST64_ABS_LO12_NC, val: gotPltEntryAddr); |
993 | relocateNoSym(loc: buf + 8, type: R_AARCH64_ADD_ABS_LO12_NC, val: gotPltEntryAddr); |
994 | |
995 | if (pacEntry) |
996 | memcpy(dest: buf + sizeof(addrInst), src: pacBr, n: sizeof(pacBr)); |
997 | else |
998 | memcpy(dest: buf + sizeof(addrInst), src: stdBr, n: sizeof(stdBr)); |
999 | if (!hasBti) |
1000 | // We didn't add the BTI c instruction so round out size with NOP. |
1001 | memcpy(dest: buf + sizeof(addrInst) + sizeof(stdBr), src: nopData, n: sizeof(nopData)); |
1002 | } |
1003 | |
1004 | static TargetInfo *getTargetInfo() { |
1005 | if ((config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI) || |
1006 | config->zPacPlt) { |
1007 | static AArch64BtiPac t; |
1008 | return &t; |
1009 | } |
1010 | static AArch64 t; |
1011 | return &t; |
1012 | } |
1013 | |
1014 | TargetInfo *elf::getAArch64TargetInfo() { return getTargetInfo(); } |
1015 | |
1016 | template <class ELFT> |
1017 | static void |
1018 | addTaggedSymbolReferences(InputSectionBase &sec, |
1019 | DenseMap<Symbol *, unsigned> &referenceCount) { |
1020 | assert(sec.type == SHT_AARCH64_MEMTAG_GLOBALS_STATIC); |
1021 | |
1022 | const RelsOrRelas<ELFT> rels = sec.relsOrRelas<ELFT>(); |
1023 | if (rels.areRelocsRel()) |
1024 | error(msg: "non-RELA relocations are not allowed with memtag globals" ); |
1025 | |
1026 | for (const typename ELFT::Rela &rel : rels.relas) { |
1027 | Symbol &sym = sec.file->getRelocTargetSym(rel); |
1028 | // Linker-synthesized symbols such as __executable_start may be referenced |
1029 | // as tagged in input objfiles, and we don't want them to be tagged. A |
1030 | // cheap way to exclude them is the type check, but their type is |
1031 | // STT_NOTYPE. In addition, this save us from checking untaggable symbols, |
1032 | // like functions or TLS symbols. |
1033 | if (sym.type != STT_OBJECT) |
1034 | continue; |
1035 | // STB_LOCAL symbols can't be referenced from outside the object file, and |
1036 | // thus don't need to be checked for references from other object files. |
1037 | if (sym.binding == STB_LOCAL) { |
1038 | sym.setIsTagged(true); |
1039 | continue; |
1040 | } |
1041 | ++referenceCount[&sym]; |
1042 | } |
1043 | sec.markDead(); |
1044 | } |
1045 | |
1046 | // A tagged symbol must be denoted as being tagged by all references and the |
1047 | // chosen definition. For simplicity, here, it must also be denoted as tagged |
1048 | // for all definitions. Otherwise: |
1049 | // |
1050 | // 1. A tagged definition can be used by an untagged declaration, in which case |
1051 | // the untagged access may be PC-relative, causing a tag mismatch at |
1052 | // runtime. |
1053 | // 2. An untagged definition can be used by a tagged declaration, where the |
1054 | // compiler has taken advantage of the increased alignment of the tagged |
1055 | // declaration, but the alignment at runtime is wrong, causing a fault. |
1056 | // |
1057 | // Ideally, this isn't a problem, as any TU that imports or exports tagged |
1058 | // symbols should also be built with tagging. But, to handle these cases, we |
1059 | // demote the symbol to be untagged. |
1060 | void lld::elf::createTaggedSymbols(const SmallVector<ELFFileBase *, 0> &files) { |
1061 | assert(hasMemtag()); |
1062 | |
1063 | // First, collect all symbols that are marked as tagged, and count how many |
1064 | // times they're marked as tagged. |
1065 | DenseMap<Symbol *, unsigned> taggedSymbolReferenceCount; |
1066 | for (InputFile* file : files) { |
1067 | if (file->kind() != InputFile::ObjKind) |
1068 | continue; |
1069 | for (InputSectionBase *section : file->getSections()) { |
1070 | if (!section || section->type != SHT_AARCH64_MEMTAG_GLOBALS_STATIC || |
1071 | section == &InputSection::discarded) |
1072 | continue; |
1073 | invokeELFT(addTaggedSymbolReferences, *section, |
1074 | taggedSymbolReferenceCount); |
1075 | } |
1076 | } |
1077 | |
1078 | // Now, go through all the symbols. If the number of declarations + |
1079 | // definitions to a symbol exceeds the amount of times they're marked as |
1080 | // tagged, it means we have an objfile that uses the untagged variant of the |
1081 | // symbol. |
1082 | for (InputFile *file : files) { |
1083 | if (file->kind() != InputFile::BinaryKind && |
1084 | file->kind() != InputFile::ObjKind) |
1085 | continue; |
1086 | |
1087 | for (Symbol *symbol : file->getSymbols()) { |
1088 | // See `addTaggedSymbolReferences` for more details. |
1089 | if (symbol->type != STT_OBJECT || |
1090 | symbol->binding == STB_LOCAL) |
1091 | continue; |
1092 | auto it = taggedSymbolReferenceCount.find(Val: symbol); |
1093 | if (it == taggedSymbolReferenceCount.end()) continue; |
1094 | unsigned &remainingAllowedTaggedRefs = it->second; |
1095 | if (remainingAllowedTaggedRefs == 0) { |
1096 | taggedSymbolReferenceCount.erase(I: it); |
1097 | continue; |
1098 | } |
1099 | --remainingAllowedTaggedRefs; |
1100 | } |
1101 | } |
1102 | |
1103 | // `addTaggedSymbolReferences` has already checked that we have RELA |
1104 | // relocations, the only other way to get written addends is with |
1105 | // --apply-dynamic-relocs. |
1106 | if (!taggedSymbolReferenceCount.empty() && config->writeAddends) |
1107 | error(msg: "--apply-dynamic-relocs cannot be used with MTE globals" ); |
1108 | |
1109 | // Now, `taggedSymbolReferenceCount` should only contain symbols that are |
1110 | // defined as tagged exactly the same amount as it's referenced, meaning all |
1111 | // uses are tagged. |
1112 | for (auto &[symbol, remainingTaggedRefs] : taggedSymbolReferenceCount) { |
1113 | assert(remainingTaggedRefs == 0 && |
1114 | "Symbol is defined as tagged more times than it's used" ); |
1115 | symbol->setIsTagged(true); |
1116 | } |
1117 | } |
1118 | |