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 | case R_AARCH64_GOT_LD_PREL19: |
179 | return R_GOT_PC; |
180 | case R_AARCH64_NONE: |
181 | return R_NONE; |
182 | default: |
183 | error(msg: getErrorLocation(loc) + "unknown relocation (" + Twine(type) + |
184 | ") against symbol " + toString(s)); |
185 | return R_NONE; |
186 | } |
187 | } |
188 | |
189 | RelExpr AArch64::adjustTlsExpr(RelType type, RelExpr expr) const { |
190 | if (expr == R_RELAX_TLS_GD_TO_IE) { |
191 | if (type == R_AARCH64_TLSDESC_ADR_PAGE21) |
192 | return R_AARCH64_RELAX_TLS_GD_TO_IE_PAGE_PC; |
193 | return R_RELAX_TLS_GD_TO_IE_ABS; |
194 | } |
195 | return expr; |
196 | } |
197 | |
198 | bool AArch64::usesOnlyLowPageBits(RelType type) const { |
199 | switch (type) { |
200 | default: |
201 | return false; |
202 | case R_AARCH64_ADD_ABS_LO12_NC: |
203 | case R_AARCH64_LD64_GOT_LO12_NC: |
204 | case R_AARCH64_LDST128_ABS_LO12_NC: |
205 | case R_AARCH64_LDST16_ABS_LO12_NC: |
206 | case R_AARCH64_LDST32_ABS_LO12_NC: |
207 | case R_AARCH64_LDST64_ABS_LO12_NC: |
208 | case R_AARCH64_LDST8_ABS_LO12_NC: |
209 | case R_AARCH64_TLSDESC_ADD_LO12: |
210 | case R_AARCH64_TLSDESC_LD64_LO12: |
211 | case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: |
212 | return true; |
213 | } |
214 | } |
215 | |
216 | RelType AArch64::getDynRel(RelType type) const { |
217 | if (type == R_AARCH64_ABS64 || type == R_AARCH64_AUTH_ABS64) |
218 | return type; |
219 | return R_AARCH64_NONE; |
220 | } |
221 | |
222 | int64_t AArch64::getImplicitAddend(const uint8_t *buf, RelType type) const { |
223 | switch (type) { |
224 | case R_AARCH64_TLSDESC: |
225 | return read64(p: buf + 8); |
226 | case R_AARCH64_NONE: |
227 | case R_AARCH64_GLOB_DAT: |
228 | case R_AARCH64_JUMP_SLOT: |
229 | return 0; |
230 | case R_AARCH64_ABS16: |
231 | case R_AARCH64_PREL16: |
232 | return SignExtend64<16>(x: read16(p: buf)); |
233 | case R_AARCH64_ABS32: |
234 | case R_AARCH64_PREL32: |
235 | return SignExtend64<32>(x: read32(p: buf)); |
236 | case R_AARCH64_ABS64: |
237 | case R_AARCH64_PREL64: |
238 | case R_AARCH64_RELATIVE: |
239 | case R_AARCH64_IRELATIVE: |
240 | case R_AARCH64_TLS_TPREL64: |
241 | return read64(p: buf); |
242 | |
243 | // The following relocation types all point at instructions, and |
244 | // relocate an immediate field in the instruction. |
245 | // |
246 | // The general rule, from AAELF64 §5.7.2 "Addends and PC-bias", |
247 | // says: "If the relocation relocates an instruction the immediate |
248 | // field of the instruction is extracted, scaled as required by |
249 | // the instruction field encoding, and sign-extended to 64 bits". |
250 | |
251 | // The R_AARCH64_MOVW family operates on wide MOV/MOVK/MOVZ |
252 | // instructions, which have a 16-bit immediate field with its low |
253 | // bit in bit 5 of the instruction encoding. When the immediate |
254 | // field is used as an implicit addend for REL-type relocations, |
255 | // it is treated as added to the low bits of the output value, not |
256 | // shifted depending on the relocation type. |
257 | // |
258 | // This allows REL relocations to express the requirement 'please |
259 | // add 12345 to this symbol value and give me the four 16-bit |
260 | // chunks of the result', by putting the same addend 12345 in all |
261 | // four instructions. Carries between the 16-bit chunks are |
262 | // handled correctly, because the whole 64-bit addition is done |
263 | // once per relocation. |
264 | case R_AARCH64_MOVW_UABS_G0: |
265 | case R_AARCH64_MOVW_UABS_G0_NC: |
266 | case R_AARCH64_MOVW_UABS_G1: |
267 | case R_AARCH64_MOVW_UABS_G1_NC: |
268 | case R_AARCH64_MOVW_UABS_G2: |
269 | case R_AARCH64_MOVW_UABS_G2_NC: |
270 | case R_AARCH64_MOVW_UABS_G3: |
271 | return SignExtend64<16>(x: getBits(val: read32(p: buf), start: 5, end: 20)); |
272 | |
273 | // R_AARCH64_TSTBR14 points at a TBZ or TBNZ instruction, which |
274 | // has a 14-bit offset measured in instructions, i.e. shifted left |
275 | // by 2. |
276 | case R_AARCH64_TSTBR14: |
277 | return SignExtend64<16>(x: getBits(val: read32(p: buf), start: 5, end: 18) << 2); |
278 | |
279 | // R_AARCH64_CONDBR19 operates on the ordinary B.cond instruction, |
280 | // which has a 19-bit offset measured in instructions. |
281 | // |
282 | // R_AARCH64_LD_PREL_LO19 operates on the LDR (literal) |
283 | // instruction, which also has a 19-bit offset, measured in 4-byte |
284 | // chunks. So the calculation is the same as for |
285 | // R_AARCH64_CONDBR19. |
286 | case R_AARCH64_CONDBR19: |
287 | case R_AARCH64_LD_PREL_LO19: |
288 | return SignExtend64<21>(x: getBits(val: read32(p: buf), start: 5, end: 23) << 2); |
289 | |
290 | // R_AARCH64_ADD_ABS_LO12_NC operates on ADD (immediate). The |
291 | // immediate can optionally be shifted left by 12 bits, but this |
292 | // relocation is intended for the case where it is not. |
293 | case R_AARCH64_ADD_ABS_LO12_NC: |
294 | return SignExtend64<12>(x: getBits(val: read32(p: buf), start: 10, end: 21)); |
295 | |
296 | // R_AARCH64_ADR_PREL_LO21 operates on an ADR instruction, whose |
297 | // 21-bit immediate is split between two bits high up in the word |
298 | // (in fact the two _lowest_ order bits of the value) and 19 bits |
299 | // lower down. |
300 | // |
301 | // R_AARCH64_ADR_PREL_PG_HI21[_NC] operate on an ADRP instruction, |
302 | // which encodes the immediate in the same way, but will shift it |
303 | // left by 12 bits when the instruction executes. For the same |
304 | // reason as the MOVW family, we don't apply that left shift here. |
305 | case R_AARCH64_ADR_PREL_LO21: |
306 | case R_AARCH64_ADR_PREL_PG_HI21: |
307 | case R_AARCH64_ADR_PREL_PG_HI21_NC: |
308 | return SignExtend64<21>(x: (getBits(val: read32(p: buf), start: 5, end: 23) << 2) | |
309 | getBits(val: read32(p: buf), start: 29, end: 30)); |
310 | |
311 | // R_AARCH64_{JUMP,CALL}26 operate on B and BL, which have a |
312 | // 26-bit offset measured in instructions. |
313 | case R_AARCH64_JUMP26: |
314 | case R_AARCH64_CALL26: |
315 | return SignExtend64<28>(x: getBits(val: read32(p: buf), start: 0, end: 25) << 2); |
316 | |
317 | default: |
318 | internalLinkerError(loc: getErrorLocation(loc: buf), |
319 | msg: "cannot read addend for relocation " + toString(type)); |
320 | return 0; |
321 | } |
322 | } |
323 | |
324 | void AArch64::writeGotPlt(uint8_t *buf, const Symbol &) const { |
325 | write64(p: buf, v: in.plt->getVA()); |
326 | } |
327 | |
328 | void AArch64::writeIgotPlt(uint8_t *buf, const Symbol &s) const { |
329 | if (config->writeAddends) |
330 | write64(p: buf, v: s.getVA()); |
331 | } |
332 | |
333 | void AArch64::(uint8_t *buf) const { |
334 | const uint8_t pltData[] = { |
335 | 0xf0, 0x7b, 0xbf, 0xa9, // stp x16, x30, [sp,#-16]! |
336 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[2])) |
337 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[2]))] |
338 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.got.plt[2])) |
339 | 0x20, 0x02, 0x1f, 0xd6, // br x17 |
340 | 0x1f, 0x20, 0x03, 0xd5, // nop |
341 | 0x1f, 0x20, 0x03, 0xd5, // nop |
342 | 0x1f, 0x20, 0x03, 0xd5 // nop |
343 | }; |
344 | memcpy(dest: buf, src: pltData, n: sizeof(pltData)); |
345 | |
346 | uint64_t got = in.gotPlt->getVA(); |
347 | uint64_t plt = in.plt->getVA(); |
348 | relocateNoSym(loc: buf + 4, type: R_AARCH64_ADR_PREL_PG_HI21, |
349 | val: getAArch64Page(expr: got + 16) - getAArch64Page(expr: plt + 4)); |
350 | relocateNoSym(loc: buf + 8, type: R_AARCH64_LDST64_ABS_LO12_NC, val: got + 16); |
351 | relocateNoSym(loc: buf + 12, type: R_AARCH64_ADD_ABS_LO12_NC, val: got + 16); |
352 | } |
353 | |
354 | void AArch64::writePlt(uint8_t *buf, const Symbol &sym, |
355 | uint64_t pltEntryAddr) const { |
356 | const uint8_t inst[] = { |
357 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[n])) |
358 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[n]))] |
359 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.got.plt[n])) |
360 | 0x20, 0x02, 0x1f, 0xd6 // br x17 |
361 | }; |
362 | memcpy(dest: buf, src: inst, n: sizeof(inst)); |
363 | |
364 | uint64_t gotPltEntryAddr = sym.getGotPltVA(); |
365 | relocateNoSym(loc: buf, type: R_AARCH64_ADR_PREL_PG_HI21, |
366 | val: getAArch64Page(expr: gotPltEntryAddr) - getAArch64Page(expr: pltEntryAddr)); |
367 | relocateNoSym(loc: buf + 4, type: R_AARCH64_LDST64_ABS_LO12_NC, val: gotPltEntryAddr); |
368 | relocateNoSym(loc: buf + 8, type: R_AARCH64_ADD_ABS_LO12_NC, val: gotPltEntryAddr); |
369 | } |
370 | |
371 | bool AArch64::needsThunk(RelExpr expr, RelType type, const InputFile *file, |
372 | uint64_t branchAddr, const Symbol &s, |
373 | int64_t a) const { |
374 | // If s is an undefined weak symbol and does not have a PLT entry then it will |
375 | // be resolved as a branch to the next instruction. If it is hidden, its |
376 | // binding has been converted to local, so we just check isUndefined() here. A |
377 | // undefined non-weak symbol will have been errored. |
378 | if (s.isUndefined() && !s.isInPlt()) |
379 | return false; |
380 | // ELF for the ARM 64-bit architecture, section Call and Jump relocations |
381 | // only permits range extension thunks for R_AARCH64_CALL26 and |
382 | // R_AARCH64_JUMP26 relocation types. |
383 | if (type != R_AARCH64_CALL26 && type != R_AARCH64_JUMP26 && |
384 | type != R_AARCH64_PLT32) |
385 | return false; |
386 | uint64_t dst = expr == R_PLT_PC ? s.getPltVA() : s.getVA(addend: a); |
387 | return !inBranchRange(type, src: branchAddr, dst); |
388 | } |
389 | |
390 | uint32_t AArch64::getThunkSectionSpacing() const { |
391 | // See comment in Arch/ARM.cpp for a more detailed explanation of |
392 | // getThunkSectionSpacing(). For AArch64 the only branches we are permitted to |
393 | // Thunk have a range of +/- 128 MiB |
394 | return (128 * 1024 * 1024) - 0x30000; |
395 | } |
396 | |
397 | bool AArch64::inBranchRange(RelType type, uint64_t src, uint64_t dst) const { |
398 | if (type != R_AARCH64_CALL26 && type != R_AARCH64_JUMP26 && |
399 | type != R_AARCH64_PLT32) |
400 | return true; |
401 | // The AArch64 call and unconditional branch instructions have a range of |
402 | // +/- 128 MiB. The PLT32 relocation supports a range up to +/- 2 GiB. |
403 | uint64_t range = |
404 | type == R_AARCH64_PLT32 ? (UINT64_C(1) << 31) : (128 * 1024 * 1024); |
405 | if (dst > src) { |
406 | // Immediate of branch is signed. |
407 | range -= 4; |
408 | return dst - src <= range; |
409 | } |
410 | return src - dst <= range; |
411 | } |
412 | |
413 | static void write32AArch64Addr(uint8_t *l, uint64_t imm) { |
414 | uint32_t immLo = (imm & 0x3) << 29; |
415 | uint32_t immHi = (imm & 0x1FFFFC) << 3; |
416 | uint64_t mask = (0x3 << 29) | (0x1FFFFC << 3); |
417 | write32le(P: l, V: (read32le(P: l) & ~mask) | immLo | immHi); |
418 | } |
419 | |
420 | static void writeMaskedBits32le(uint8_t *p, int32_t v, uint32_t mask) { |
421 | write32le(P: p, V: (read32le(P: p) & ~mask) | v); |
422 | } |
423 | |
424 | // Update the immediate field in a AARCH64 ldr, str, and add instruction. |
425 | static void write32Imm12(uint8_t *l, uint64_t imm) { |
426 | writeMaskedBits32le(p: l, v: (imm & 0xFFF) << 10, mask: 0xFFF << 10); |
427 | } |
428 | |
429 | // Update the immediate field in an AArch64 movk, movn or movz instruction |
430 | // for a signed relocation, and update the opcode of a movn or movz instruction |
431 | // to match the sign of the operand. |
432 | static void writeSMovWImm(uint8_t *loc, uint32_t imm) { |
433 | uint32_t inst = read32le(P: loc); |
434 | // Opcode field is bits 30, 29, with 10 = movz, 00 = movn and 11 = movk. |
435 | if (!(inst & (1 << 29))) { |
436 | // movn or movz. |
437 | if (imm & 0x10000) { |
438 | // Change opcode to movn, which takes an inverted operand. |
439 | imm ^= 0xFFFF; |
440 | inst &= ~(1 << 30); |
441 | } else { |
442 | // Change opcode to movz. |
443 | inst |= 1 << 30; |
444 | } |
445 | } |
446 | write32le(P: loc, V: inst | ((imm & 0xFFFF) << 5)); |
447 | } |
448 | |
449 | void AArch64::relocate(uint8_t *loc, const Relocation &rel, |
450 | uint64_t val) const { |
451 | switch (rel.type) { |
452 | case R_AARCH64_ABS16: |
453 | case R_AARCH64_PREL16: |
454 | checkIntUInt(loc, v: val, n: 16, rel); |
455 | write16(p: loc, v: val); |
456 | break; |
457 | case R_AARCH64_ABS32: |
458 | case R_AARCH64_PREL32: |
459 | checkIntUInt(loc, v: val, n: 32, rel); |
460 | write32(p: loc, v: val); |
461 | break; |
462 | case R_AARCH64_PLT32: |
463 | case R_AARCH64_GOTPCREL32: |
464 | checkInt(loc, v: val, n: 32, rel); |
465 | write32(p: loc, v: val); |
466 | break; |
467 | case R_AARCH64_ABS64: |
468 | // AArch64 relocations to tagged symbols have extended semantics, as |
469 | // described here: |
470 | // https://github.com/ARM-software/abi-aa/blob/main/memtagabielf64/memtagabielf64.rst#841extended-semantics-of-r_aarch64_relative. |
471 | // tl;dr: encode the symbol's special addend in the place, which is an |
472 | // offset to the point where the logical tag is derived from. Quick hack, if |
473 | // the addend is within the symbol's bounds, no need to encode the tag |
474 | // derivation offset. |
475 | if (rel.sym && rel.sym->isTagged() && |
476 | (rel.addend < 0 || |
477 | rel.addend >= static_cast<int64_t>(rel.sym->getSize()))) |
478 | write64(p: loc, v: -rel.addend); |
479 | else |
480 | write64(p: loc, v: val); |
481 | break; |
482 | case R_AARCH64_PREL64: |
483 | write64(p: loc, v: val); |
484 | break; |
485 | case R_AARCH64_AUTH_ABS64: |
486 | // If val is wider than 32 bits, the relocation must have been moved from |
487 | // .relr.auth.dyn to .rela.dyn, and the addend write is not needed. |
488 | // |
489 | // If val fits in 32 bits, we have two potential scenarios: |
490 | // * True RELR: Write the 32-bit `val`. |
491 | // * RELA: Even if the value now fits in 32 bits, it might have been |
492 | // converted from RELR during an iteration in |
493 | // finalizeAddressDependentContent(). Writing the value is harmless |
494 | // because dynamic linking ignores it. |
495 | if (isInt<32>(x: val)) |
496 | write32(p: loc, v: val); |
497 | break; |
498 | case R_AARCH64_ADD_ABS_LO12_NC: |
499 | write32Imm12(l: loc, imm: val); |
500 | break; |
501 | case R_AARCH64_ADR_GOT_PAGE: |
502 | case R_AARCH64_ADR_PREL_PG_HI21: |
503 | case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21: |
504 | case R_AARCH64_TLSDESC_ADR_PAGE21: |
505 | checkInt(loc, v: val, n: 33, rel); |
506 | [[fallthrough]]; |
507 | case R_AARCH64_ADR_PREL_PG_HI21_NC: |
508 | write32AArch64Addr(l: loc, imm: val >> 12); |
509 | break; |
510 | case R_AARCH64_ADR_PREL_LO21: |
511 | checkInt(loc, v: val, n: 21, rel); |
512 | write32AArch64Addr(l: loc, imm: val); |
513 | break; |
514 | case R_AARCH64_JUMP26: |
515 | // Normally we would just write the bits of the immediate field, however |
516 | // when patching instructions for the cpu errata fix -fix-cortex-a53-843419 |
517 | // we want to replace a non-branch instruction with a branch immediate |
518 | // instruction. By writing all the bits of the instruction including the |
519 | // opcode and the immediate (0 001 | 01 imm26) we can do this |
520 | // transformation by placing a R_AARCH64_JUMP26 relocation at the offset of |
521 | // the instruction we want to patch. |
522 | write32le(P: loc, V: 0x14000000); |
523 | [[fallthrough]]; |
524 | case R_AARCH64_CALL26: |
525 | checkInt(loc, v: val, n: 28, rel); |
526 | writeMaskedBits32le(p: loc, v: (val & 0x0FFFFFFC) >> 2, mask: 0x0FFFFFFC >> 2); |
527 | break; |
528 | case R_AARCH64_CONDBR19: |
529 | case R_AARCH64_LD_PREL_LO19: |
530 | case R_AARCH64_GOT_LD_PREL19: |
531 | checkAlignment(loc, v: val, n: 4, rel); |
532 | checkInt(loc, v: val, n: 21, rel); |
533 | writeMaskedBits32le(p: loc, v: (val & 0x1FFFFC) << 3, mask: 0x1FFFFC << 3); |
534 | break; |
535 | case R_AARCH64_LDST8_ABS_LO12_NC: |
536 | case R_AARCH64_TLSLE_LDST8_TPREL_LO12_NC: |
537 | write32Imm12(l: loc, imm: getBits(val, start: 0, end: 11)); |
538 | break; |
539 | case R_AARCH64_LDST16_ABS_LO12_NC: |
540 | case R_AARCH64_TLSLE_LDST16_TPREL_LO12_NC: |
541 | checkAlignment(loc, v: val, n: 2, rel); |
542 | write32Imm12(l: loc, imm: getBits(val, start: 1, end: 11)); |
543 | break; |
544 | case R_AARCH64_LDST32_ABS_LO12_NC: |
545 | case R_AARCH64_TLSLE_LDST32_TPREL_LO12_NC: |
546 | checkAlignment(loc, v: val, n: 4, rel); |
547 | write32Imm12(l: loc, imm: getBits(val, start: 2, end: 11)); |
548 | break; |
549 | case R_AARCH64_LDST64_ABS_LO12_NC: |
550 | case R_AARCH64_LD64_GOT_LO12_NC: |
551 | case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: |
552 | case R_AARCH64_TLSLE_LDST64_TPREL_LO12_NC: |
553 | case R_AARCH64_TLSDESC_LD64_LO12: |
554 | checkAlignment(loc, v: val, n: 8, rel); |
555 | write32Imm12(l: loc, imm: getBits(val, start: 3, end: 11)); |
556 | break; |
557 | case R_AARCH64_LDST128_ABS_LO12_NC: |
558 | case R_AARCH64_TLSLE_LDST128_TPREL_LO12_NC: |
559 | checkAlignment(loc, v: val, n: 16, rel); |
560 | write32Imm12(l: loc, imm: getBits(val, start: 4, end: 11)); |
561 | break; |
562 | case R_AARCH64_LD64_GOTPAGE_LO15: |
563 | checkAlignment(loc, v: val, n: 8, rel); |
564 | write32Imm12(l: loc, imm: getBits(val, start: 3, end: 14)); |
565 | break; |
566 | case R_AARCH64_MOVW_UABS_G0: |
567 | checkUInt(loc, v: val, n: 16, rel); |
568 | [[fallthrough]]; |
569 | case R_AARCH64_MOVW_UABS_G0_NC: |
570 | writeMaskedBits32le(p: loc, v: (val & 0xFFFF) << 5, mask: 0xFFFF << 5); |
571 | break; |
572 | case R_AARCH64_MOVW_UABS_G1: |
573 | checkUInt(loc, v: val, n: 32, rel); |
574 | [[fallthrough]]; |
575 | case R_AARCH64_MOVW_UABS_G1_NC: |
576 | writeMaskedBits32le(p: loc, v: (val & 0xFFFF0000) >> 11, mask: 0xFFFF0000 >> 11); |
577 | break; |
578 | case R_AARCH64_MOVW_UABS_G2: |
579 | checkUInt(loc, v: val, n: 48, rel); |
580 | [[fallthrough]]; |
581 | case R_AARCH64_MOVW_UABS_G2_NC: |
582 | writeMaskedBits32le(p: loc, v: (val & 0xFFFF00000000) >> 27, |
583 | mask: 0xFFFF00000000 >> 27); |
584 | break; |
585 | case R_AARCH64_MOVW_UABS_G3: |
586 | writeMaskedBits32le(p: loc, v: (val & 0xFFFF000000000000) >> 43, |
587 | mask: 0xFFFF000000000000 >> 43); |
588 | break; |
589 | case R_AARCH64_MOVW_PREL_G0: |
590 | case R_AARCH64_MOVW_SABS_G0: |
591 | case R_AARCH64_TLSLE_MOVW_TPREL_G0: |
592 | checkInt(loc, v: val, n: 17, rel); |
593 | [[fallthrough]]; |
594 | case R_AARCH64_MOVW_PREL_G0_NC: |
595 | case R_AARCH64_TLSLE_MOVW_TPREL_G0_NC: |
596 | writeSMovWImm(loc, imm: val); |
597 | break; |
598 | case R_AARCH64_MOVW_PREL_G1: |
599 | case R_AARCH64_MOVW_SABS_G1: |
600 | case R_AARCH64_TLSLE_MOVW_TPREL_G1: |
601 | checkInt(loc, v: val, n: 33, rel); |
602 | [[fallthrough]]; |
603 | case R_AARCH64_MOVW_PREL_G1_NC: |
604 | case R_AARCH64_TLSLE_MOVW_TPREL_G1_NC: |
605 | writeSMovWImm(loc, imm: val >> 16); |
606 | break; |
607 | case R_AARCH64_MOVW_PREL_G2: |
608 | case R_AARCH64_MOVW_SABS_G2: |
609 | case R_AARCH64_TLSLE_MOVW_TPREL_G2: |
610 | checkInt(loc, v: val, n: 49, rel); |
611 | [[fallthrough]]; |
612 | case R_AARCH64_MOVW_PREL_G2_NC: |
613 | writeSMovWImm(loc, imm: val >> 32); |
614 | break; |
615 | case R_AARCH64_MOVW_PREL_G3: |
616 | writeSMovWImm(loc, imm: val >> 48); |
617 | break; |
618 | case R_AARCH64_TSTBR14: |
619 | checkInt(loc, v: val, n: 16, rel); |
620 | writeMaskedBits32le(p: loc, v: (val & 0xFFFC) << 3, mask: 0xFFFC << 3); |
621 | break; |
622 | case R_AARCH64_TLSLE_ADD_TPREL_HI12: |
623 | checkUInt(loc, v: val, n: 24, rel); |
624 | write32Imm12(l: loc, imm: val >> 12); |
625 | break; |
626 | case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC: |
627 | case R_AARCH64_TLSDESC_ADD_LO12: |
628 | write32Imm12(l: loc, imm: val); |
629 | break; |
630 | case R_AARCH64_TLSDESC: |
631 | // For R_AARCH64_TLSDESC the addend is stored in the second 64-bit word. |
632 | write64(p: loc + 8, v: val); |
633 | break; |
634 | default: |
635 | llvm_unreachable("unknown relocation" ); |
636 | } |
637 | } |
638 | |
639 | void AArch64::relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, |
640 | uint64_t val) const { |
641 | // TLSDESC Global-Dynamic relocation are in the form: |
642 | // adrp x0, :tlsdesc:v [R_AARCH64_TLSDESC_ADR_PAGE21] |
643 | // ldr x1, [x0, #:tlsdesc_lo12:v [R_AARCH64_TLSDESC_LD64_LO12] |
644 | // add x0, x0, :tlsdesc_los:v [R_AARCH64_TLSDESC_ADD_LO12] |
645 | // .tlsdesccall [R_AARCH64_TLSDESC_CALL] |
646 | // blr x1 |
647 | // And it can optimized to: |
648 | // movz x0, #0x0, lsl #16 |
649 | // movk x0, #0x10 |
650 | // nop |
651 | // nop |
652 | checkUInt(loc, v: val, n: 32, rel); |
653 | |
654 | switch (rel.type) { |
655 | case R_AARCH64_TLSDESC_ADD_LO12: |
656 | case R_AARCH64_TLSDESC_CALL: |
657 | write32le(P: loc, V: 0xd503201f); // nop |
658 | return; |
659 | case R_AARCH64_TLSDESC_ADR_PAGE21: |
660 | write32le(P: loc, V: 0xd2a00000 | (((val >> 16) & 0xffff) << 5)); // movz |
661 | return; |
662 | case R_AARCH64_TLSDESC_LD64_LO12: |
663 | write32le(P: loc, V: 0xf2800000 | ((val & 0xffff) << 5)); // movk |
664 | return; |
665 | default: |
666 | llvm_unreachable("unsupported relocation for TLS GD to LE relaxation" ); |
667 | } |
668 | } |
669 | |
670 | void AArch64::relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, |
671 | uint64_t val) const { |
672 | // TLSDESC Global-Dynamic relocation are in the form: |
673 | // adrp x0, :tlsdesc:v [R_AARCH64_TLSDESC_ADR_PAGE21] |
674 | // ldr x1, [x0, #:tlsdesc_lo12:v [R_AARCH64_TLSDESC_LD64_LO12] |
675 | // add x0, x0, :tlsdesc_los:v [R_AARCH64_TLSDESC_ADD_LO12] |
676 | // .tlsdesccall [R_AARCH64_TLSDESC_CALL] |
677 | // blr x1 |
678 | // And it can optimized to: |
679 | // adrp x0, :gottprel:v |
680 | // ldr x0, [x0, :gottprel_lo12:v] |
681 | // nop |
682 | // nop |
683 | |
684 | switch (rel.type) { |
685 | case R_AARCH64_TLSDESC_ADD_LO12: |
686 | case R_AARCH64_TLSDESC_CALL: |
687 | write32le(P: loc, V: 0xd503201f); // nop |
688 | break; |
689 | case R_AARCH64_TLSDESC_ADR_PAGE21: |
690 | write32le(P: loc, V: 0x90000000); // adrp |
691 | relocateNoSym(loc, type: R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21, val); |
692 | break; |
693 | case R_AARCH64_TLSDESC_LD64_LO12: |
694 | write32le(P: loc, V: 0xf9400000); // ldr |
695 | relocateNoSym(loc, type: R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC, val); |
696 | break; |
697 | default: |
698 | llvm_unreachable("unsupported relocation for TLS GD to LE relaxation" ); |
699 | } |
700 | } |
701 | |
702 | void AArch64::relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, |
703 | uint64_t val) const { |
704 | checkUInt(loc, v: val, n: 32, rel); |
705 | |
706 | if (rel.type == R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21) { |
707 | // Generate MOVZ. |
708 | uint32_t regNo = read32le(P: loc) & 0x1f; |
709 | write32le(P: loc, V: (0xd2a00000 | regNo) | (((val >> 16) & 0xffff) << 5)); |
710 | return; |
711 | } |
712 | if (rel.type == R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC) { |
713 | // Generate MOVK. |
714 | uint32_t regNo = read32le(P: loc) & 0x1f; |
715 | write32le(P: loc, V: (0xf2800000 | regNo) | ((val & 0xffff) << 5)); |
716 | return; |
717 | } |
718 | llvm_unreachable("invalid relocation for TLS IE to LE relaxation" ); |
719 | } |
720 | |
721 | AArch64Relaxer::AArch64Relaxer(ArrayRef<Relocation> relocs) { |
722 | if (!config->relax) |
723 | return; |
724 | // Check if R_AARCH64_ADR_GOT_PAGE and R_AARCH64_LD64_GOT_LO12_NC |
725 | // always appear in pairs. |
726 | size_t i = 0; |
727 | const size_t size = relocs.size(); |
728 | for (; i != size; ++i) { |
729 | if (relocs[i].type == R_AARCH64_ADR_GOT_PAGE) { |
730 | if (i + 1 < size && relocs[i + 1].type == R_AARCH64_LD64_GOT_LO12_NC) { |
731 | ++i; |
732 | continue; |
733 | } |
734 | break; |
735 | } else if (relocs[i].type == R_AARCH64_LD64_GOT_LO12_NC) { |
736 | break; |
737 | } |
738 | } |
739 | safeToRelaxAdrpLdr = i == size; |
740 | } |
741 | |
742 | bool AArch64Relaxer::tryRelaxAdrpAdd(const Relocation &adrpRel, |
743 | const Relocation &addRel, uint64_t secAddr, |
744 | uint8_t *buf) const { |
745 | // When the address of sym is within the range of ADR then |
746 | // we may relax |
747 | // ADRP xn, sym |
748 | // ADD xn, xn, :lo12: sym |
749 | // to |
750 | // NOP |
751 | // ADR xn, sym |
752 | if (!config->relax || adrpRel.type != R_AARCH64_ADR_PREL_PG_HI21 || |
753 | addRel.type != R_AARCH64_ADD_ABS_LO12_NC) |
754 | return false; |
755 | // Check if the relocations apply to consecutive instructions. |
756 | if (adrpRel.offset + 4 != addRel.offset) |
757 | return false; |
758 | if (adrpRel.sym != addRel.sym) |
759 | return false; |
760 | if (adrpRel.addend != 0 || addRel.addend != 0) |
761 | return false; |
762 | |
763 | uint32_t adrpInstr = read32le(P: buf + adrpRel.offset); |
764 | uint32_t addInstr = read32le(P: buf + addRel.offset); |
765 | // Check if the first instruction is ADRP and the second instruction is ADD. |
766 | if ((adrpInstr & 0x9f000000) != 0x90000000 || |
767 | (addInstr & 0xffc00000) != 0x91000000) |
768 | return false; |
769 | uint32_t adrpDestReg = adrpInstr & 0x1f; |
770 | uint32_t addDestReg = addInstr & 0x1f; |
771 | uint32_t addSrcReg = (addInstr >> 5) & 0x1f; |
772 | if (adrpDestReg != addDestReg || adrpDestReg != addSrcReg) |
773 | return false; |
774 | |
775 | Symbol &sym = *adrpRel.sym; |
776 | // Check if the address difference is within 1MiB range. |
777 | int64_t val = sym.getVA() - (secAddr + addRel.offset); |
778 | if (val < -1024 * 1024 || val >= 1024 * 1024) |
779 | return false; |
780 | |
781 | Relocation adrRel = {.expr: R_ABS, .type: R_AARCH64_ADR_PREL_LO21, .offset: addRel.offset, |
782 | /*addend=*/0, .sym: &sym}; |
783 | // nop |
784 | write32le(P: buf + adrpRel.offset, V: 0xd503201f); |
785 | // adr x_<dest_reg> |
786 | write32le(P: buf + adrRel.offset, V: 0x10000000 | adrpDestReg); |
787 | target->relocate(loc: buf + adrRel.offset, rel: adrRel, val); |
788 | return true; |
789 | } |
790 | |
791 | bool AArch64Relaxer::tryRelaxAdrpLdr(const Relocation &adrpRel, |
792 | const Relocation &ldrRel, uint64_t secAddr, |
793 | uint8_t *buf) const { |
794 | if (!safeToRelaxAdrpLdr) |
795 | return false; |
796 | |
797 | // When the definition of sym is not preemptible then we may |
798 | // be able to relax |
799 | // ADRP xn, :got: sym |
800 | // LDR xn, [ xn :got_lo12: sym] |
801 | // to |
802 | // ADRP xn, sym |
803 | // ADD xn, xn, :lo_12: sym |
804 | |
805 | if (adrpRel.type != R_AARCH64_ADR_GOT_PAGE || |
806 | ldrRel.type != R_AARCH64_LD64_GOT_LO12_NC) |
807 | return false; |
808 | // Check if the relocations apply to consecutive instructions. |
809 | if (adrpRel.offset + 4 != ldrRel.offset) |
810 | return false; |
811 | // Check if the relocations reference the same symbol and |
812 | // skip undefined, preemptible and STT_GNU_IFUNC symbols. |
813 | if (!adrpRel.sym || adrpRel.sym != ldrRel.sym || !adrpRel.sym->isDefined() || |
814 | adrpRel.sym->isPreemptible || adrpRel.sym->isGnuIFunc()) |
815 | return false; |
816 | // Check if the addends of the both relocations are zero. |
817 | if (adrpRel.addend != 0 || ldrRel.addend != 0) |
818 | return false; |
819 | uint32_t adrpInstr = read32le(P: buf + adrpRel.offset); |
820 | uint32_t ldrInstr = read32le(P: buf + ldrRel.offset); |
821 | // Check if the first instruction is ADRP and the second instruction is LDR. |
822 | if ((adrpInstr & 0x9f000000) != 0x90000000 || |
823 | (ldrInstr & 0x3b000000) != 0x39000000) |
824 | return false; |
825 | // Check the value of the sf bit. |
826 | if (!(ldrInstr >> 31)) |
827 | return false; |
828 | uint32_t adrpDestReg = adrpInstr & 0x1f; |
829 | uint32_t ldrDestReg = ldrInstr & 0x1f; |
830 | uint32_t ldrSrcReg = (ldrInstr >> 5) & 0x1f; |
831 | // Check if ADPR and LDR use the same register. |
832 | if (adrpDestReg != ldrDestReg || adrpDestReg != ldrSrcReg) |
833 | return false; |
834 | |
835 | Symbol &sym = *adrpRel.sym; |
836 | // GOT references to absolute symbols can't be relaxed to use ADRP/ADD in |
837 | // position-independent code because these instructions produce a relative |
838 | // address. |
839 | if (config->isPic && !cast<Defined>(Val&: sym).section) |
840 | return false; |
841 | // Check if the address difference is within 4GB range. |
842 | int64_t val = |
843 | getAArch64Page(expr: sym.getVA()) - getAArch64Page(expr: secAddr + adrpRel.offset); |
844 | if (val != llvm::SignExtend64(X: val, B: 33)) |
845 | return false; |
846 | |
847 | Relocation adrpSymRel = {.expr: R_AARCH64_PAGE_PC, .type: R_AARCH64_ADR_PREL_PG_HI21, |
848 | .offset: adrpRel.offset, /*addend=*/0, .sym: &sym}; |
849 | Relocation addRel = {.expr: R_ABS, .type: R_AARCH64_ADD_ABS_LO12_NC, .offset: ldrRel.offset, |
850 | /*addend=*/0, .sym: &sym}; |
851 | |
852 | // adrp x_<dest_reg> |
853 | write32le(P: buf + adrpSymRel.offset, V: 0x90000000 | adrpDestReg); |
854 | // add x_<dest reg>, x_<dest reg> |
855 | write32le(P: buf + addRel.offset, V: 0x91000000 | adrpDestReg | (adrpDestReg << 5)); |
856 | |
857 | target->relocate(loc: buf + adrpSymRel.offset, rel: adrpSymRel, |
858 | val: SignExtend64(X: getAArch64Page(expr: sym.getVA()) - |
859 | getAArch64Page(expr: secAddr + adrpSymRel.offset), |
860 | B: 64)); |
861 | target->relocate(loc: buf + addRel.offset, rel: addRel, val: SignExtend64(X: sym.getVA(), B: 64)); |
862 | tryRelaxAdrpAdd(adrpRel: adrpSymRel, addRel, secAddr, buf); |
863 | return true; |
864 | } |
865 | |
866 | // Tagged symbols have upper address bits that are added by the dynamic loader, |
867 | // and thus need the full 64-bit GOT entry. Do not relax such symbols. |
868 | static bool needsGotForMemtag(const Relocation &rel) { |
869 | return rel.sym->isTagged() && needsGot(expr: rel.expr); |
870 | } |
871 | |
872 | void AArch64::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const { |
873 | uint64_t secAddr = sec.getOutputSection()->addr; |
874 | if (auto *s = dyn_cast<InputSection>(Val: &sec)) |
875 | secAddr += s->outSecOff; |
876 | else if (auto *ehIn = dyn_cast<EhInputSection>(Val: &sec)) |
877 | secAddr += ehIn->getParent()->outSecOff; |
878 | AArch64Relaxer relaxer(sec.relocs()); |
879 | for (size_t i = 0, size = sec.relocs().size(); i != size; ++i) { |
880 | const Relocation &rel = sec.relocs()[i]; |
881 | uint8_t *loc = buf + rel.offset; |
882 | const uint64_t val = |
883 | sec.getRelocTargetVA(File: sec.file, Type: rel.type, A: rel.addend, |
884 | P: secAddr + rel.offset, Sym: *rel.sym, Expr: rel.expr); |
885 | |
886 | if (needsGotForMemtag(rel)) { |
887 | relocate(loc, rel, val); |
888 | continue; |
889 | } |
890 | |
891 | switch (rel.expr) { |
892 | case R_AARCH64_GOT_PAGE_PC: |
893 | if (i + 1 < size && |
894 | relaxer.tryRelaxAdrpLdr(adrpRel: rel, ldrRel: sec.relocs()[i + 1], secAddr, buf)) { |
895 | ++i; |
896 | continue; |
897 | } |
898 | break; |
899 | case R_AARCH64_PAGE_PC: |
900 | if (i + 1 < size && |
901 | relaxer.tryRelaxAdrpAdd(adrpRel: rel, addRel: sec.relocs()[i + 1], secAddr, buf)) { |
902 | ++i; |
903 | continue; |
904 | } |
905 | break; |
906 | case R_AARCH64_RELAX_TLS_GD_TO_IE_PAGE_PC: |
907 | case R_RELAX_TLS_GD_TO_IE_ABS: |
908 | relaxTlsGdToIe(loc, rel, val); |
909 | continue; |
910 | case R_RELAX_TLS_GD_TO_LE: |
911 | relaxTlsGdToLe(loc, rel, val); |
912 | continue; |
913 | case R_RELAX_TLS_IE_TO_LE: |
914 | relaxTlsIeToLe(loc, rel, val); |
915 | continue; |
916 | default: |
917 | break; |
918 | } |
919 | relocate(loc, rel, val); |
920 | } |
921 | } |
922 | |
923 | // AArch64 may use security features in variant PLT sequences. These are: |
924 | // Pointer Authentication (PAC), introduced in armv8.3-a and Branch Target |
925 | // Indicator (BTI) introduced in armv8.5-a. The additional instructions used |
926 | // in the variant Plt sequences are encoded in the Hint space so they can be |
927 | // deployed on older architectures, which treat the instructions as a nop. |
928 | // PAC and BTI can be combined leading to the following combinations: |
929 | // writePltHeader |
930 | // writePltHeaderBti (no PAC Header needed) |
931 | // writePlt |
932 | // writePltBti (BTI only) |
933 | // writePltPac (PAC only) |
934 | // writePltBtiPac (BTI and PAC) |
935 | // |
936 | // When PAC is enabled the dynamic loader encrypts the address that it places |
937 | // in the .got.plt using the pacia1716 instruction which encrypts the value in |
938 | // x17 using the modifier in x16. The static linker places autia1716 before the |
939 | // indirect branch to x17 to authenticate the address in x17 with the modifier |
940 | // in x16. This makes it more difficult for an attacker to modify the value in |
941 | // the .got.plt. |
942 | // |
943 | // When BTI is enabled all indirect branches must land on a bti instruction. |
944 | // The static linker must place a bti instruction at the start of any PLT entry |
945 | // that may be the target of an indirect branch. As the PLT entries call the |
946 | // lazy resolver indirectly this must have a bti instruction at start. In |
947 | // general a bti instruction is not needed for a PLT entry as indirect calls |
948 | // are resolved to the function address and not the PLT entry for the function. |
949 | // There are a small number of cases where the PLT address can escape, such as |
950 | // taking the address of a function or ifunc via a non got-generating |
951 | // relocation, and a shared library refers to that symbol. |
952 | // |
953 | // We use the bti c variant of the instruction which permits indirect branches |
954 | // (br) via x16/x17 and indirect function calls (blr) via any register. The ABI |
955 | // guarantees that all indirect branches from code requiring BTI protection |
956 | // will go via x16/x17 |
957 | |
958 | namespace { |
959 | class AArch64BtiPac final : public AArch64 { |
960 | public: |
961 | AArch64BtiPac(); |
962 | void writePltHeader(uint8_t *buf) const override; |
963 | void writePlt(uint8_t *buf, const Symbol &sym, |
964 | uint64_t pltEntryAddr) const override; |
965 | |
966 | private: |
967 | bool ; // bti instruction needed in PLT Header and Entry |
968 | bool pacEntry; // autia1716 instruction needed in PLT Entry |
969 | }; |
970 | } // namespace |
971 | |
972 | AArch64BtiPac::AArch64BtiPac() { |
973 | btiHeader = (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI); |
974 | // A BTI (Branch Target Indicator) Plt Entry is only required if the |
975 | // address of the PLT entry can be taken by the program, which permits an |
976 | // indirect jump to the PLT entry. This can happen when the address |
977 | // of the PLT entry for a function is canonicalised due to the address of |
978 | // the function in an executable being taken by a shared library, or |
979 | // non-preemptible ifunc referenced by non-GOT-generating, non-PLT-generating |
980 | // relocations. |
981 | // The PAC PLT entries require dynamic loader support and this isn't known |
982 | // from properties in the objects, so we use the command line flag. |
983 | pacEntry = config->zPacPlt; |
984 | |
985 | if (btiHeader || pacEntry) { |
986 | pltEntrySize = 24; |
987 | ipltEntrySize = 24; |
988 | } |
989 | } |
990 | |
991 | void AArch64BtiPac::(uint8_t *buf) const { |
992 | const uint8_t btiData[] = { 0x5f, 0x24, 0x03, 0xd5 }; // bti c |
993 | const uint8_t pltData[] = { |
994 | 0xf0, 0x7b, 0xbf, 0xa9, // stp x16, x30, [sp,#-16]! |
995 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[2])) |
996 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[2]))] |
997 | 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.got.plt[2])) |
998 | 0x20, 0x02, 0x1f, 0xd6, // br x17 |
999 | 0x1f, 0x20, 0x03, 0xd5, // nop |
1000 | 0x1f, 0x20, 0x03, 0xd5 // nop |
1001 | }; |
1002 | const uint8_t nopData[] = { 0x1f, 0x20, 0x03, 0xd5 }; // nop |
1003 | |
1004 | uint64_t got = in.gotPlt->getVA(); |
1005 | uint64_t plt = in.plt->getVA(); |
1006 | |
1007 | if (btiHeader) { |
1008 | // PltHeader is called indirectly by plt[N]. Prefix pltData with a BTI C |
1009 | // instruction. |
1010 | memcpy(dest: buf, src: btiData, n: sizeof(btiData)); |
1011 | buf += sizeof(btiData); |
1012 | plt += sizeof(btiData); |
1013 | } |
1014 | memcpy(dest: buf, src: pltData, n: sizeof(pltData)); |
1015 | |
1016 | relocateNoSym(loc: buf + 4, type: R_AARCH64_ADR_PREL_PG_HI21, |
1017 | val: getAArch64Page(expr: got + 16) - getAArch64Page(expr: plt + 8)); |
1018 | relocateNoSym(loc: buf + 8, type: R_AARCH64_LDST64_ABS_LO12_NC, val: got + 16); |
1019 | relocateNoSym(loc: buf + 12, type: R_AARCH64_ADD_ABS_LO12_NC, val: got + 16); |
1020 | if (!btiHeader) |
1021 | // We didn't add the BTI c instruction so round out size with NOP. |
1022 | memcpy(dest: buf + sizeof(pltData), src: nopData, n: sizeof(nopData)); |
1023 | } |
1024 | |
1025 | void AArch64BtiPac::writePlt(uint8_t *buf, const Symbol &sym, |
1026 | uint64_t pltEntryAddr) const { |
1027 | // The PLT entry is of the form: |
1028 | // [btiData] addrInst (pacBr | stdBr) [nopData] |
1029 | const uint8_t btiData[] = { 0x5f, 0x24, 0x03, 0xd5 }; // bti c |
1030 | const uint8_t addrInst[] = { |
1031 | 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.got.plt[n])) |
1032 | 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.got.plt[n]))] |
1033 | 0x10, 0x02, 0x00, 0x91 // add x16, x16, Offset(&(.got.plt[n])) |
1034 | }; |
1035 | const uint8_t pacBr[] = { |
1036 | 0x9f, 0x21, 0x03, 0xd5, // autia1716 |
1037 | 0x20, 0x02, 0x1f, 0xd6 // br x17 |
1038 | }; |
1039 | const uint8_t stdBr[] = { |
1040 | 0x20, 0x02, 0x1f, 0xd6, // br x17 |
1041 | 0x1f, 0x20, 0x03, 0xd5 // nop |
1042 | }; |
1043 | const uint8_t nopData[] = { 0x1f, 0x20, 0x03, 0xd5 }; // nop |
1044 | |
1045 | // NEEDS_COPY indicates a non-ifunc canonical PLT entry whose address may |
1046 | // escape to shared objects. isInIplt indicates a non-preemptible ifunc. Its |
1047 | // address may escape if referenced by a direct relocation. If relative |
1048 | // vtables are used then if the vtable is in a shared object the offsets will |
1049 | // be to the PLT entry. The condition is conservative. |
1050 | bool hasBti = btiHeader && |
1051 | (sym.hasFlag(bit: NEEDS_COPY) || sym.isInIplt || sym.thunkAccessed); |
1052 | if (hasBti) { |
1053 | memcpy(dest: buf, src: btiData, n: sizeof(btiData)); |
1054 | buf += sizeof(btiData); |
1055 | pltEntryAddr += sizeof(btiData); |
1056 | } |
1057 | |
1058 | uint64_t gotPltEntryAddr = sym.getGotPltVA(); |
1059 | memcpy(dest: buf, src: addrInst, n: sizeof(addrInst)); |
1060 | relocateNoSym(loc: buf, type: R_AARCH64_ADR_PREL_PG_HI21, |
1061 | val: getAArch64Page(expr: gotPltEntryAddr) - getAArch64Page(expr: pltEntryAddr)); |
1062 | relocateNoSym(loc: buf + 4, type: R_AARCH64_LDST64_ABS_LO12_NC, val: gotPltEntryAddr); |
1063 | relocateNoSym(loc: buf + 8, type: R_AARCH64_ADD_ABS_LO12_NC, val: gotPltEntryAddr); |
1064 | |
1065 | if (pacEntry) |
1066 | memcpy(dest: buf + sizeof(addrInst), src: pacBr, n: sizeof(pacBr)); |
1067 | else |
1068 | memcpy(dest: buf + sizeof(addrInst), src: stdBr, n: sizeof(stdBr)); |
1069 | if (!hasBti) |
1070 | // We didn't add the BTI c instruction so round out size with NOP. |
1071 | memcpy(dest: buf + sizeof(addrInst) + sizeof(stdBr), src: nopData, n: sizeof(nopData)); |
1072 | } |
1073 | |
1074 | static TargetInfo *getTargetInfo() { |
1075 | if ((config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI) || |
1076 | config->zPacPlt) { |
1077 | static AArch64BtiPac t; |
1078 | return &t; |
1079 | } |
1080 | static AArch64 t; |
1081 | return &t; |
1082 | } |
1083 | |
1084 | TargetInfo *elf::getAArch64TargetInfo() { return getTargetInfo(); } |
1085 | |
1086 | template <class ELFT> |
1087 | static void |
1088 | addTaggedSymbolReferences(InputSectionBase &sec, |
1089 | DenseMap<Symbol *, unsigned> &referenceCount) { |
1090 | assert(sec.type == SHT_AARCH64_MEMTAG_GLOBALS_STATIC); |
1091 | |
1092 | const RelsOrRelas<ELFT> rels = sec.relsOrRelas<ELFT>(); |
1093 | if (rels.areRelocsRel()) |
1094 | error(msg: "non-RELA relocations are not allowed with memtag globals" ); |
1095 | |
1096 | for (const typename ELFT::Rela &rel : rels.relas) { |
1097 | Symbol &sym = sec.file->getRelocTargetSym(rel); |
1098 | // Linker-synthesized symbols such as __executable_start may be referenced |
1099 | // as tagged in input objfiles, and we don't want them to be tagged. A |
1100 | // cheap way to exclude them is the type check, but their type is |
1101 | // STT_NOTYPE. In addition, this save us from checking untaggable symbols, |
1102 | // like functions or TLS symbols. |
1103 | if (sym.type != STT_OBJECT) |
1104 | continue; |
1105 | // STB_LOCAL symbols can't be referenced from outside the object file, and |
1106 | // thus don't need to be checked for references from other object files. |
1107 | if (sym.binding == STB_LOCAL) { |
1108 | sym.setIsTagged(true); |
1109 | continue; |
1110 | } |
1111 | ++referenceCount[&sym]; |
1112 | } |
1113 | sec.markDead(); |
1114 | } |
1115 | |
1116 | // A tagged symbol must be denoted as being tagged by all references and the |
1117 | // chosen definition. For simplicity, here, it must also be denoted as tagged |
1118 | // for all definitions. Otherwise: |
1119 | // |
1120 | // 1. A tagged definition can be used by an untagged declaration, in which case |
1121 | // the untagged access may be PC-relative, causing a tag mismatch at |
1122 | // runtime. |
1123 | // 2. An untagged definition can be used by a tagged declaration, where the |
1124 | // compiler has taken advantage of the increased alignment of the tagged |
1125 | // declaration, but the alignment at runtime is wrong, causing a fault. |
1126 | // |
1127 | // Ideally, this isn't a problem, as any TU that imports or exports tagged |
1128 | // symbols should also be built with tagging. But, to handle these cases, we |
1129 | // demote the symbol to be untagged. |
1130 | void lld::elf::createTaggedSymbols(const SmallVector<ELFFileBase *, 0> &files) { |
1131 | assert(hasMemtag()); |
1132 | |
1133 | // First, collect all symbols that are marked as tagged, and count how many |
1134 | // times they're marked as tagged. |
1135 | DenseMap<Symbol *, unsigned> taggedSymbolReferenceCount; |
1136 | for (InputFile* file : files) { |
1137 | if (file->kind() != InputFile::ObjKind) |
1138 | continue; |
1139 | for (InputSectionBase *section : file->getSections()) { |
1140 | if (!section || section->type != SHT_AARCH64_MEMTAG_GLOBALS_STATIC || |
1141 | section == &InputSection::discarded) |
1142 | continue; |
1143 | invokeELFT(addTaggedSymbolReferences, *section, |
1144 | taggedSymbolReferenceCount); |
1145 | } |
1146 | } |
1147 | |
1148 | // Now, go through all the symbols. If the number of declarations + |
1149 | // definitions to a symbol exceeds the amount of times they're marked as |
1150 | // tagged, it means we have an objfile that uses the untagged variant of the |
1151 | // symbol. |
1152 | for (InputFile *file : files) { |
1153 | if (file->kind() != InputFile::BinaryKind && |
1154 | file->kind() != InputFile::ObjKind) |
1155 | continue; |
1156 | |
1157 | for (Symbol *symbol : file->getSymbols()) { |
1158 | // See `addTaggedSymbolReferences` for more details. |
1159 | if (symbol->type != STT_OBJECT || |
1160 | symbol->binding == STB_LOCAL) |
1161 | continue; |
1162 | auto it = taggedSymbolReferenceCount.find(Val: symbol); |
1163 | if (it == taggedSymbolReferenceCount.end()) continue; |
1164 | unsigned &remainingAllowedTaggedRefs = it->second; |
1165 | if (remainingAllowedTaggedRefs == 0) { |
1166 | taggedSymbolReferenceCount.erase(I: it); |
1167 | continue; |
1168 | } |
1169 | --remainingAllowedTaggedRefs; |
1170 | } |
1171 | } |
1172 | |
1173 | // `addTaggedSymbolReferences` has already checked that we have RELA |
1174 | // relocations, the only other way to get written addends is with |
1175 | // --apply-dynamic-relocs. |
1176 | if (!taggedSymbolReferenceCount.empty() && config->writeAddends) |
1177 | error(msg: "--apply-dynamic-relocs cannot be used with MTE globals" ); |
1178 | |
1179 | // Now, `taggedSymbolReferenceCount` should only contain symbols that are |
1180 | // defined as tagged exactly the same amount as it's referenced, meaning all |
1181 | // uses are tagged. |
1182 | for (auto &[symbol, remainingTaggedRefs] : taggedSymbolReferenceCount) { |
1183 | assert(remainingTaggedRefs == 0 && |
1184 | "Symbol is defined as tagged more times than it's used" ); |
1185 | symbol->setIsTagged(true); |
1186 | } |
1187 | } |
1188 | |