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 "Symbols.h" 10 #include "SyntheticSections.h" 11 #include "Target.h" 12 #include "Thunks.h" 13 #include "lld/Common/ErrorHandler.h" 14 #include "llvm/Object/ELF.h" 15 #include "llvm/Support/Endian.h" 16 17 using namespace llvm; 18 using namespace llvm::support::endian; 19 using namespace llvm::ELF; 20 using namespace lld; 21 using namespace lld::elf; 22 23 // Page(Expr) is the page address of the expression Expr, defined 24 // as (Expr & ~0xFFF). (This applies even if the machine page size 25 // supported by the platform has a different value.) 26 uint64_t elf::getAArch64Page(uint64_t expr) { 27 return expr & ~static_cast<uint64_t>(0xFFF); 28 } 29 30 namespace { 31 class AArch64 : public TargetInfo { 32 public: 33 AArch64(); 34 RelExpr getRelExpr(RelType type, const Symbol &s, 35 const uint8_t *loc) const override; 36 RelType getDynRel(RelType type) const override; 37 int64_t getImplicitAddend(const uint8_t *buf, RelType type) const override; 38 void writeGotPlt(uint8_t *buf, const Symbol &s) const override; 39 void writePltHeader(uint8_t *buf) const override; 40 void writePlt(uint8_t *buf, const Symbol &sym, 41 uint64_t pltEntryAddr) const override; 42 bool needsThunk(RelExpr expr, RelType type, const InputFile *file, 43 uint64_t branchAddr, const Symbol &s, 44 int64_t a) const override; 45 uint32_t getThunkSectionSpacing() const override; 46 bool inBranchRange(RelType type, uint64_t src, uint64_t dst) const override; 47 bool usesOnlyLowPageBits(RelType type) const override; 48 void relocate(uint8_t *loc, const Relocation &rel, 49 uint64_t val) const override; 50 RelExpr adjustTlsExpr(RelType type, RelExpr expr) const override; 51 void relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, 52 uint64_t val) const override; 53 void relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, 54 uint64_t val) const override; 55 void relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, 56 uint64_t val) const override; 57 }; 58 } // namespace 59 60 AArch64::AArch64() { 61 copyRel = R_AARCH64_COPY; 62 relativeRel = R_AARCH64_RELATIVE; 63 iRelativeRel = R_AARCH64_IRELATIVE; 64 gotRel = R_AARCH64_GLOB_DAT; 65 noneRel = R_AARCH64_NONE; 66 pltRel = R_AARCH64_JUMP_SLOT; 67 symbolicRel = R_AARCH64_ABS64; 68 tlsDescRel = R_AARCH64_TLSDESC; 69 tlsGotRel = R_AARCH64_TLS_TPREL64; 70 pltHeaderSize = 32; 71 pltEntrySize = 16; 72 ipltEntrySize = 16; 73 defaultMaxPageSize = 65536; 74 gotBaseSymInGotPlt = false; 75 76 // Align to the 2 MiB page size (known as a superpage or huge page). 77 // FreeBSD automatically promotes 2 MiB-aligned allocations. 78 defaultImageBase = 0x200000; 79 80 needsThunks = true; 81 } 82 83 RelExpr AArch64::getRelExpr(RelType type, const Symbol &s, 84 const uint8_t *loc) const { 85 switch (type) { 86 case R_AARCH64_ABS16: 87 case R_AARCH64_ABS32: 88 case R_AARCH64_ABS64: 89 case R_AARCH64_ADD_ABS_LO12_NC: 90 case R_AARCH64_LDST128_ABS_LO12_NC: 91 case R_AARCH64_LDST16_ABS_LO12_NC: 92 case R_AARCH64_LDST32_ABS_LO12_NC: 93 case R_AARCH64_LDST64_ABS_LO12_NC: 94 case R_AARCH64_LDST8_ABS_LO12_NC: 95 case R_AARCH64_MOVW_SABS_G0: 96 case R_AARCH64_MOVW_SABS_G1: 97 case R_AARCH64_MOVW_SABS_G2: 98 case R_AARCH64_MOVW_UABS_G0: 99 case R_AARCH64_MOVW_UABS_G0_NC: 100 case R_AARCH64_MOVW_UABS_G1: 101 case R_AARCH64_MOVW_UABS_G1_NC: 102 case R_AARCH64_MOVW_UABS_G2: 103 case R_AARCH64_MOVW_UABS_G2_NC: 104 case R_AARCH64_MOVW_UABS_G3: 105 return R_ABS; 106 case R_AARCH64_TLSDESC_ADR_PAGE21: 107 return R_AARCH64_TLSDESC_PAGE; 108 case R_AARCH64_TLSDESC_LD64_LO12: 109 case R_AARCH64_TLSDESC_ADD_LO12: 110 return R_TLSDESC; 111 case R_AARCH64_TLSDESC_CALL: 112 return R_TLSDESC_CALL; 113 case R_AARCH64_TLSLE_ADD_TPREL_HI12: 114 case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC: 115 case R_AARCH64_TLSLE_LDST8_TPREL_LO12_NC: 116 case R_AARCH64_TLSLE_LDST16_TPREL_LO12_NC: 117 case R_AARCH64_TLSLE_LDST32_TPREL_LO12_NC: 118 case R_AARCH64_TLSLE_LDST64_TPREL_LO12_NC: 119 case R_AARCH64_TLSLE_LDST128_TPREL_LO12_NC: 120 case R_AARCH64_TLSLE_MOVW_TPREL_G0: 121 case R_AARCH64_TLSLE_MOVW_TPREL_G0_NC: 122 case R_AARCH64_TLSLE_MOVW_TPREL_G1: 123 case R_AARCH64_TLSLE_MOVW_TPREL_G1_NC: 124 case R_AARCH64_TLSLE_MOVW_TPREL_G2: 125 return R_TPREL; 126 case R_AARCH64_CALL26: 127 case R_AARCH64_CONDBR19: 128 case R_AARCH64_JUMP26: 129 case R_AARCH64_TSTBR14: 130 case R_AARCH64_PLT32: 131 return R_PLT_PC; 132 case R_AARCH64_PREL16: 133 case R_AARCH64_PREL32: 134 case R_AARCH64_PREL64: 135 case R_AARCH64_ADR_PREL_LO21: 136 case R_AARCH64_LD_PREL_LO19: 137 case R_AARCH64_MOVW_PREL_G0: 138 case R_AARCH64_MOVW_PREL_G0_NC: 139 case R_AARCH64_MOVW_PREL_G1: 140 case R_AARCH64_MOVW_PREL_G1_NC: 141 case R_AARCH64_MOVW_PREL_G2: 142 case R_AARCH64_MOVW_PREL_G2_NC: 143 case R_AARCH64_MOVW_PREL_G3: 144 return R_PC; 145 case R_AARCH64_ADR_PREL_PG_HI21: 146 case R_AARCH64_ADR_PREL_PG_HI21_NC: 147 return R_AARCH64_PAGE_PC; 148 case R_AARCH64_LD64_GOT_LO12_NC: 149 case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: 150 return R_GOT; 151 case R_AARCH64_LD64_GOTPAGE_LO15: 152 return R_AARCH64_GOT_PAGE; 153 case R_AARCH64_ADR_GOT_PAGE: 154 case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21: 155 return R_AARCH64_GOT_PAGE_PC; 156 case R_AARCH64_NONE: 157 return R_NONE; 158 default: 159 error(getErrorLocation(loc) + "unknown relocation (" + Twine(type) + 160 ") against symbol " + toString(s)); 161 return R_NONE; 162 } 163 } 164 165 RelExpr AArch64::adjustTlsExpr(RelType type, RelExpr expr) const { 166 if (expr == R_RELAX_TLS_GD_TO_IE) { 167 if (type == R_AARCH64_TLSDESC_ADR_PAGE21) 168 return R_AARCH64_RELAX_TLS_GD_TO_IE_PAGE_PC; 169 return R_RELAX_TLS_GD_TO_IE_ABS; 170 } 171 return expr; 172 } 173 174 bool AArch64::usesOnlyLowPageBits(RelType type) const { 175 switch (type) { 176 default: 177 return false; 178 case R_AARCH64_ADD_ABS_LO12_NC: 179 case R_AARCH64_LD64_GOT_LO12_NC: 180 case R_AARCH64_LDST128_ABS_LO12_NC: 181 case R_AARCH64_LDST16_ABS_LO12_NC: 182 case R_AARCH64_LDST32_ABS_LO12_NC: 183 case R_AARCH64_LDST64_ABS_LO12_NC: 184 case R_AARCH64_LDST8_ABS_LO12_NC: 185 case R_AARCH64_TLSDESC_ADD_LO12: 186 case R_AARCH64_TLSDESC_LD64_LO12: 187 case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: 188 return true; 189 } 190 } 191 192 RelType AArch64::getDynRel(RelType type) const { 193 if (type == R_AARCH64_ABS64) 194 return type; 195 return R_AARCH64_NONE; 196 } 197 198 int64_t AArch64::getImplicitAddend(const uint8_t *buf, RelType type) const { 199 switch (type) { 200 case R_AARCH64_TLSDESC: 201 return read64(buf + 8); 202 default: 203 internalLinkerError(getErrorLocation(buf), 204 "cannot read addend for relocation " + toString(type)); 205 return 0; 206 } 207 } 208 209 void AArch64::writeGotPlt(uint8_t *buf, const Symbol &) const { 210 write64(buf, in.plt->getVA()); 211 } 212 213 void AArch64::writePltHeader(uint8_t *buf) const { 214 const uint8_t pltData[] = { 215 0xf0, 0x7b, 0xbf, 0xa9, // stp x16, x30, [sp,#-16]! 216 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.plt.got[2])) 217 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.plt.got[2]))] 218 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.plt.got[2])) 219 0x20, 0x02, 0x1f, 0xd6, // br x17 220 0x1f, 0x20, 0x03, 0xd5, // nop 221 0x1f, 0x20, 0x03, 0xd5, // nop 222 0x1f, 0x20, 0x03, 0xd5 // nop 223 }; 224 memcpy(buf, pltData, sizeof(pltData)); 225 226 uint64_t got = in.gotPlt->getVA(); 227 uint64_t plt = in.plt->getVA(); 228 relocateNoSym(buf + 4, R_AARCH64_ADR_PREL_PG_HI21, 229 getAArch64Page(got + 16) - getAArch64Page(plt + 4)); 230 relocateNoSym(buf + 8, R_AARCH64_LDST64_ABS_LO12_NC, got + 16); 231 relocateNoSym(buf + 12, R_AARCH64_ADD_ABS_LO12_NC, got + 16); 232 } 233 234 void AArch64::writePlt(uint8_t *buf, const Symbol &sym, 235 uint64_t pltEntryAddr) const { 236 const uint8_t inst[] = { 237 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.plt.got[n])) 238 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.plt.got[n]))] 239 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.plt.got[n])) 240 0x20, 0x02, 0x1f, 0xd6 // br x17 241 }; 242 memcpy(buf, inst, sizeof(inst)); 243 244 uint64_t gotPltEntryAddr = sym.getGotPltVA(); 245 relocateNoSym(buf, R_AARCH64_ADR_PREL_PG_HI21, 246 getAArch64Page(gotPltEntryAddr) - getAArch64Page(pltEntryAddr)); 247 relocateNoSym(buf + 4, R_AARCH64_LDST64_ABS_LO12_NC, gotPltEntryAddr); 248 relocateNoSym(buf + 8, R_AARCH64_ADD_ABS_LO12_NC, gotPltEntryAddr); 249 } 250 251 bool AArch64::needsThunk(RelExpr expr, RelType type, const InputFile *file, 252 uint64_t branchAddr, const Symbol &s, 253 int64_t a) const { 254 // If s is an undefined weak symbol and does not have a PLT entry then it 255 // will be resolved as a branch to the next instruction. 256 if (s.isUndefWeak() && !s.isInPlt()) 257 return false; 258 // ELF for the ARM 64-bit architecture, section Call and Jump relocations 259 // only permits range extension thunks for R_AARCH64_CALL26 and 260 // R_AARCH64_JUMP26 relocation types. 261 if (type != R_AARCH64_CALL26 && type != R_AARCH64_JUMP26 && 262 type != R_AARCH64_PLT32) 263 return false; 264 uint64_t dst = expr == R_PLT_PC ? s.getPltVA() : s.getVA(a); 265 return !inBranchRange(type, branchAddr, dst); 266 } 267 268 uint32_t AArch64::getThunkSectionSpacing() const { 269 // See comment in Arch/ARM.cpp for a more detailed explanation of 270 // getThunkSectionSpacing(). For AArch64 the only branches we are permitted to 271 // Thunk have a range of +/- 128 MiB 272 return (128 * 1024 * 1024) - 0x30000; 273 } 274 275 bool AArch64::inBranchRange(RelType type, uint64_t src, uint64_t dst) const { 276 if (type != R_AARCH64_CALL26 && type != R_AARCH64_JUMP26 && 277 type != R_AARCH64_PLT32) 278 return true; 279 // The AArch64 call and unconditional branch instructions have a range of 280 // +/- 128 MiB. The PLT32 relocation supports a range up to +/- 2 GiB. 281 uint64_t range = 282 type == R_AARCH64_PLT32 ? (UINT64_C(1) << 31) : (128 * 1024 * 1024); 283 if (dst > src) { 284 // Immediate of branch is signed. 285 range -= 4; 286 return dst - src <= range; 287 } 288 return src - dst <= range; 289 } 290 291 static void write32AArch64Addr(uint8_t *l, uint64_t imm) { 292 uint32_t immLo = (imm & 0x3) << 29; 293 uint32_t immHi = (imm & 0x1FFFFC) << 3; 294 uint64_t mask = (0x3 << 29) | (0x1FFFFC << 3); 295 write32le(l, (read32le(l) & ~mask) | immLo | immHi); 296 } 297 298 // Return the bits [Start, End] from Val shifted Start bits. 299 // For instance, getBits(0xF0, 4, 8) returns 0xF. 300 static uint64_t getBits(uint64_t val, int start, int end) { 301 uint64_t mask = ((uint64_t)1 << (end + 1 - start)) - 1; 302 return (val >> start) & mask; 303 } 304 305 static void or32le(uint8_t *p, int32_t v) { write32le(p, read32le(p) | v); } 306 307 // Update the immediate field in a AARCH64 ldr, str, and add instruction. 308 static void or32AArch64Imm(uint8_t *l, uint64_t imm) { 309 or32le(l, (imm & 0xFFF) << 10); 310 } 311 312 // Update the immediate field in an AArch64 movk, movn or movz instruction 313 // for a signed relocation, and update the opcode of a movn or movz instruction 314 // to match the sign of the operand. 315 static void writeSMovWImm(uint8_t *loc, uint32_t imm) { 316 uint32_t inst = read32le(loc); 317 // Opcode field is bits 30, 29, with 10 = movz, 00 = movn and 11 = movk. 318 if (!(inst & (1 << 29))) { 319 // movn or movz. 320 if (imm & 0x10000) { 321 // Change opcode to movn, which takes an inverted operand. 322 imm ^= 0xFFFF; 323 inst &= ~(1 << 30); 324 } else { 325 // Change opcode to movz. 326 inst |= 1 << 30; 327 } 328 } 329 write32le(loc, inst | ((imm & 0xFFFF) << 5)); 330 } 331 332 void AArch64::relocate(uint8_t *loc, const Relocation &rel, 333 uint64_t val) const { 334 switch (rel.type) { 335 case R_AARCH64_ABS16: 336 case R_AARCH64_PREL16: 337 checkIntUInt(loc, val, 16, rel); 338 write16(loc, val); 339 break; 340 case R_AARCH64_ABS32: 341 case R_AARCH64_PREL32: 342 checkIntUInt(loc, val, 32, rel); 343 write32(loc, val); 344 break; 345 case R_AARCH64_PLT32: 346 checkInt(loc, val, 32, rel); 347 write32(loc, val); 348 break; 349 case R_AARCH64_ABS64: 350 case R_AARCH64_PREL64: 351 write64(loc, val); 352 break; 353 case R_AARCH64_ADD_ABS_LO12_NC: 354 or32AArch64Imm(loc, val); 355 break; 356 case R_AARCH64_ADR_GOT_PAGE: 357 case R_AARCH64_ADR_PREL_PG_HI21: 358 case R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21: 359 case R_AARCH64_TLSDESC_ADR_PAGE21: 360 checkInt(loc, val, 33, rel); 361 LLVM_FALLTHROUGH; 362 case R_AARCH64_ADR_PREL_PG_HI21_NC: 363 write32AArch64Addr(loc, val >> 12); 364 break; 365 case R_AARCH64_ADR_PREL_LO21: 366 checkInt(loc, val, 21, rel); 367 write32AArch64Addr(loc, val); 368 break; 369 case R_AARCH64_JUMP26: 370 // Normally we would just write the bits of the immediate field, however 371 // when patching instructions for the cpu errata fix -fix-cortex-a53-843419 372 // we want to replace a non-branch instruction with a branch immediate 373 // instruction. By writing all the bits of the instruction including the 374 // opcode and the immediate (0 001 | 01 imm26) we can do this 375 // transformation by placing a R_AARCH64_JUMP26 relocation at the offset of 376 // the instruction we want to patch. 377 write32le(loc, 0x14000000); 378 LLVM_FALLTHROUGH; 379 case R_AARCH64_CALL26: 380 checkInt(loc, val, 28, rel); 381 or32le(loc, (val & 0x0FFFFFFC) >> 2); 382 break; 383 case R_AARCH64_CONDBR19: 384 case R_AARCH64_LD_PREL_LO19: 385 checkAlignment(loc, val, 4, rel); 386 checkInt(loc, val, 21, rel); 387 or32le(loc, (val & 0x1FFFFC) << 3); 388 break; 389 case R_AARCH64_LDST8_ABS_LO12_NC: 390 case R_AARCH64_TLSLE_LDST8_TPREL_LO12_NC: 391 or32AArch64Imm(loc, getBits(val, 0, 11)); 392 break; 393 case R_AARCH64_LDST16_ABS_LO12_NC: 394 case R_AARCH64_TLSLE_LDST16_TPREL_LO12_NC: 395 checkAlignment(loc, val, 2, rel); 396 or32AArch64Imm(loc, getBits(val, 1, 11)); 397 break; 398 case R_AARCH64_LDST32_ABS_LO12_NC: 399 case R_AARCH64_TLSLE_LDST32_TPREL_LO12_NC: 400 checkAlignment(loc, val, 4, rel); 401 or32AArch64Imm(loc, getBits(val, 2, 11)); 402 break; 403 case R_AARCH64_LDST64_ABS_LO12_NC: 404 case R_AARCH64_LD64_GOT_LO12_NC: 405 case R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC: 406 case R_AARCH64_TLSLE_LDST64_TPREL_LO12_NC: 407 case R_AARCH64_TLSDESC_LD64_LO12: 408 checkAlignment(loc, val, 8, rel); 409 or32AArch64Imm(loc, getBits(val, 3, 11)); 410 break; 411 case R_AARCH64_LDST128_ABS_LO12_NC: 412 case R_AARCH64_TLSLE_LDST128_TPREL_LO12_NC: 413 checkAlignment(loc, val, 16, rel); 414 or32AArch64Imm(loc, getBits(val, 4, 11)); 415 break; 416 case R_AARCH64_LD64_GOTPAGE_LO15: 417 checkAlignment(loc, val, 8, rel); 418 or32AArch64Imm(loc, getBits(val, 3, 14)); 419 break; 420 case R_AARCH64_MOVW_UABS_G0: 421 checkUInt(loc, val, 16, rel); 422 LLVM_FALLTHROUGH; 423 case R_AARCH64_MOVW_UABS_G0_NC: 424 or32le(loc, (val & 0xFFFF) << 5); 425 break; 426 case R_AARCH64_MOVW_UABS_G1: 427 checkUInt(loc, val, 32, rel); 428 LLVM_FALLTHROUGH; 429 case R_AARCH64_MOVW_UABS_G1_NC: 430 or32le(loc, (val & 0xFFFF0000) >> 11); 431 break; 432 case R_AARCH64_MOVW_UABS_G2: 433 checkUInt(loc, val, 48, rel); 434 LLVM_FALLTHROUGH; 435 case R_AARCH64_MOVW_UABS_G2_NC: 436 or32le(loc, (val & 0xFFFF00000000) >> 27); 437 break; 438 case R_AARCH64_MOVW_UABS_G3: 439 or32le(loc, (val & 0xFFFF000000000000) >> 43); 440 break; 441 case R_AARCH64_MOVW_PREL_G0: 442 case R_AARCH64_MOVW_SABS_G0: 443 case R_AARCH64_TLSLE_MOVW_TPREL_G0: 444 checkInt(loc, val, 17, rel); 445 LLVM_FALLTHROUGH; 446 case R_AARCH64_MOVW_PREL_G0_NC: 447 case R_AARCH64_TLSLE_MOVW_TPREL_G0_NC: 448 writeSMovWImm(loc, val); 449 break; 450 case R_AARCH64_MOVW_PREL_G1: 451 case R_AARCH64_MOVW_SABS_G1: 452 case R_AARCH64_TLSLE_MOVW_TPREL_G1: 453 checkInt(loc, val, 33, rel); 454 LLVM_FALLTHROUGH; 455 case R_AARCH64_MOVW_PREL_G1_NC: 456 case R_AARCH64_TLSLE_MOVW_TPREL_G1_NC: 457 writeSMovWImm(loc, val >> 16); 458 break; 459 case R_AARCH64_MOVW_PREL_G2: 460 case R_AARCH64_MOVW_SABS_G2: 461 case R_AARCH64_TLSLE_MOVW_TPREL_G2: 462 checkInt(loc, val, 49, rel); 463 LLVM_FALLTHROUGH; 464 case R_AARCH64_MOVW_PREL_G2_NC: 465 writeSMovWImm(loc, val >> 32); 466 break; 467 case R_AARCH64_MOVW_PREL_G3: 468 writeSMovWImm(loc, val >> 48); 469 break; 470 case R_AARCH64_TSTBR14: 471 checkInt(loc, val, 16, rel); 472 or32le(loc, (val & 0xFFFC) << 3); 473 break; 474 case R_AARCH64_TLSLE_ADD_TPREL_HI12: 475 checkUInt(loc, val, 24, rel); 476 or32AArch64Imm(loc, val >> 12); 477 break; 478 case R_AARCH64_TLSLE_ADD_TPREL_LO12_NC: 479 case R_AARCH64_TLSDESC_ADD_LO12: 480 or32AArch64Imm(loc, val); 481 break; 482 case R_AARCH64_TLSDESC: 483 // For R_AARCH64_TLSDESC the addend is stored in the second 64-bit word. 484 write64(loc + 8, val); 485 break; 486 default: 487 llvm_unreachable("unknown relocation"); 488 } 489 } 490 491 void AArch64::relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, 492 uint64_t val) const { 493 // TLSDESC Global-Dynamic relocation are in the form: 494 // adrp x0, :tlsdesc:v [R_AARCH64_TLSDESC_ADR_PAGE21] 495 // ldr x1, [x0, #:tlsdesc_lo12:v [R_AARCH64_TLSDESC_LD64_LO12] 496 // add x0, x0, :tlsdesc_los:v [R_AARCH64_TLSDESC_ADD_LO12] 497 // .tlsdesccall [R_AARCH64_TLSDESC_CALL] 498 // blr x1 499 // And it can optimized to: 500 // movz x0, #0x0, lsl #16 501 // movk x0, #0x10 502 // nop 503 // nop 504 checkUInt(loc, val, 32, rel); 505 506 switch (rel.type) { 507 case R_AARCH64_TLSDESC_ADD_LO12: 508 case R_AARCH64_TLSDESC_CALL: 509 write32le(loc, 0xd503201f); // nop 510 return; 511 case R_AARCH64_TLSDESC_ADR_PAGE21: 512 write32le(loc, 0xd2a00000 | (((val >> 16) & 0xffff) << 5)); // movz 513 return; 514 case R_AARCH64_TLSDESC_LD64_LO12: 515 write32le(loc, 0xf2800000 | ((val & 0xffff) << 5)); // movk 516 return; 517 default: 518 llvm_unreachable("unsupported relocation for TLS GD to LE relaxation"); 519 } 520 } 521 522 void AArch64::relaxTlsGdToIe(uint8_t *loc, const Relocation &rel, 523 uint64_t val) const { 524 // TLSDESC Global-Dynamic relocation are in the form: 525 // adrp x0, :tlsdesc:v [R_AARCH64_TLSDESC_ADR_PAGE21] 526 // ldr x1, [x0, #:tlsdesc_lo12:v [R_AARCH64_TLSDESC_LD64_LO12] 527 // add x0, x0, :tlsdesc_los:v [R_AARCH64_TLSDESC_ADD_LO12] 528 // .tlsdesccall [R_AARCH64_TLSDESC_CALL] 529 // blr x1 530 // And it can optimized to: 531 // adrp x0, :gottprel:v 532 // ldr x0, [x0, :gottprel_lo12:v] 533 // nop 534 // nop 535 536 switch (rel.type) { 537 case R_AARCH64_TLSDESC_ADD_LO12: 538 case R_AARCH64_TLSDESC_CALL: 539 write32le(loc, 0xd503201f); // nop 540 break; 541 case R_AARCH64_TLSDESC_ADR_PAGE21: 542 write32le(loc, 0x90000000); // adrp 543 relocateNoSym(loc, R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21, val); 544 break; 545 case R_AARCH64_TLSDESC_LD64_LO12: 546 write32le(loc, 0xf9400000); // ldr 547 relocateNoSym(loc, R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC, val); 548 break; 549 default: 550 llvm_unreachable("unsupported relocation for TLS GD to LE relaxation"); 551 } 552 } 553 554 void AArch64::relaxTlsIeToLe(uint8_t *loc, const Relocation &rel, 555 uint64_t val) const { 556 checkUInt(loc, val, 32, rel); 557 558 if (rel.type == R_AARCH64_TLSIE_ADR_GOTTPREL_PAGE21) { 559 // Generate MOVZ. 560 uint32_t regNo = read32le(loc) & 0x1f; 561 write32le(loc, (0xd2a00000 | regNo) | (((val >> 16) & 0xffff) << 5)); 562 return; 563 } 564 if (rel.type == R_AARCH64_TLSIE_LD64_GOTTPREL_LO12_NC) { 565 // Generate MOVK. 566 uint32_t regNo = read32le(loc) & 0x1f; 567 write32le(loc, (0xf2800000 | regNo) | ((val & 0xffff) << 5)); 568 return; 569 } 570 llvm_unreachable("invalid relocation for TLS IE to LE relaxation"); 571 } 572 573 // AArch64 may use security features in variant PLT sequences. These are: 574 // Pointer Authentication (PAC), introduced in armv8.3-a and Branch Target 575 // Indicator (BTI) introduced in armv8.5-a. The additional instructions used 576 // in the variant Plt sequences are encoded in the Hint space so they can be 577 // deployed on older architectures, which treat the instructions as a nop. 578 // PAC and BTI can be combined leading to the following combinations: 579 // writePltHeader 580 // writePltHeaderBti (no PAC Header needed) 581 // writePlt 582 // writePltBti (BTI only) 583 // writePltPac (PAC only) 584 // writePltBtiPac (BTI and PAC) 585 // 586 // When PAC is enabled the dynamic loader encrypts the address that it places 587 // in the .got.plt using the pacia1716 instruction which encrypts the value in 588 // x17 using the modifier in x16. The static linker places autia1716 before the 589 // indirect branch to x17 to authenticate the address in x17 with the modifier 590 // in x16. This makes it more difficult for an attacker to modify the value in 591 // the .got.plt. 592 // 593 // When BTI is enabled all indirect branches must land on a bti instruction. 594 // The static linker must place a bti instruction at the start of any PLT entry 595 // that may be the target of an indirect branch. As the PLT entries call the 596 // lazy resolver indirectly this must have a bti instruction at start. In 597 // general a bti instruction is not needed for a PLT entry as indirect calls 598 // are resolved to the function address and not the PLT entry for the function. 599 // There are a small number of cases where the PLT address can escape, such as 600 // taking the address of a function or ifunc via a non got-generating 601 // relocation, and a shared library refers to that symbol. 602 // 603 // We use the bti c variant of the instruction which permits indirect branches 604 // (br) via x16/x17 and indirect function calls (blr) via any register. The ABI 605 // guarantees that all indirect branches from code requiring BTI protection 606 // will go via x16/x17 607 608 namespace { 609 class AArch64BtiPac final : public AArch64 { 610 public: 611 AArch64BtiPac(); 612 void writePltHeader(uint8_t *buf) const override; 613 void writePlt(uint8_t *buf, const Symbol &sym, 614 uint64_t pltEntryAddr) const override; 615 616 private: 617 bool btiHeader; // bti instruction needed in PLT Header 618 bool btiEntry; // bti instruction needed in PLT Entry 619 bool pacEntry; // autia1716 instruction needed in PLT Entry 620 }; 621 } // namespace 622 623 AArch64BtiPac::AArch64BtiPac() { 624 btiHeader = (config->andFeatures & GNU_PROPERTY_AARCH64_FEATURE_1_BTI); 625 // A BTI (Branch Target Indicator) Plt Entry is only required if the 626 // address of the PLT entry can be taken by the program, which permits an 627 // indirect jump to the PLT entry. This can happen when the address 628 // of the PLT entry for a function is canonicalised due to the address of 629 // the function in an executable being taken by a shared library. 630 // FIXME: There is a potential optimization to omit the BTI if we detect 631 // that the address of the PLT entry isn't taken. 632 // The PAC PLT entries require dynamic loader support and this isn't known 633 // from properties in the objects, so we use the command line flag. 634 btiEntry = btiHeader && !config->shared; 635 pacEntry = config->zPacPlt; 636 637 if (btiEntry || pacEntry) { 638 pltEntrySize = 24; 639 ipltEntrySize = 24; 640 } 641 } 642 643 void AArch64BtiPac::writePltHeader(uint8_t *buf) const { 644 const uint8_t btiData[] = { 0x5f, 0x24, 0x03, 0xd5 }; // bti c 645 const uint8_t pltData[] = { 646 0xf0, 0x7b, 0xbf, 0xa9, // stp x16, x30, [sp,#-16]! 647 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.plt.got[2])) 648 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.plt.got[2]))] 649 0x10, 0x02, 0x00, 0x91, // add x16, x16, Offset(&(.plt.got[2])) 650 0x20, 0x02, 0x1f, 0xd6, // br x17 651 0x1f, 0x20, 0x03, 0xd5, // nop 652 0x1f, 0x20, 0x03, 0xd5 // nop 653 }; 654 const uint8_t nopData[] = { 0x1f, 0x20, 0x03, 0xd5 }; // nop 655 656 uint64_t got = in.gotPlt->getVA(); 657 uint64_t plt = in.plt->getVA(); 658 659 if (btiHeader) { 660 // PltHeader is called indirectly by plt[N]. Prefix pltData with a BTI C 661 // instruction. 662 memcpy(buf, btiData, sizeof(btiData)); 663 buf += sizeof(btiData); 664 plt += sizeof(btiData); 665 } 666 memcpy(buf, pltData, sizeof(pltData)); 667 668 relocateNoSym(buf + 4, R_AARCH64_ADR_PREL_PG_HI21, 669 getAArch64Page(got + 16) - getAArch64Page(plt + 8)); 670 relocateNoSym(buf + 8, R_AARCH64_LDST64_ABS_LO12_NC, got + 16); 671 relocateNoSym(buf + 12, R_AARCH64_ADD_ABS_LO12_NC, got + 16); 672 if (!btiHeader) 673 // We didn't add the BTI c instruction so round out size with NOP. 674 memcpy(buf + sizeof(pltData), nopData, sizeof(nopData)); 675 } 676 677 void AArch64BtiPac::writePlt(uint8_t *buf, const Symbol &sym, 678 uint64_t pltEntryAddr) const { 679 // The PLT entry is of the form: 680 // [btiData] addrInst (pacBr | stdBr) [nopData] 681 const uint8_t btiData[] = { 0x5f, 0x24, 0x03, 0xd5 }; // bti c 682 const uint8_t addrInst[] = { 683 0x10, 0x00, 0x00, 0x90, // adrp x16, Page(&(.plt.got[n])) 684 0x11, 0x02, 0x40, 0xf9, // ldr x17, [x16, Offset(&(.plt.got[n]))] 685 0x10, 0x02, 0x00, 0x91 // add x16, x16, Offset(&(.plt.got[n])) 686 }; 687 const uint8_t pacBr[] = { 688 0x9f, 0x21, 0x03, 0xd5, // autia1716 689 0x20, 0x02, 0x1f, 0xd6 // br x17 690 }; 691 const uint8_t stdBr[] = { 692 0x20, 0x02, 0x1f, 0xd6, // br x17 693 0x1f, 0x20, 0x03, 0xd5 // nop 694 }; 695 const uint8_t nopData[] = { 0x1f, 0x20, 0x03, 0xd5 }; // nop 696 697 if (btiEntry) { 698 memcpy(buf, btiData, sizeof(btiData)); 699 buf += sizeof(btiData); 700 pltEntryAddr += sizeof(btiData); 701 } 702 703 uint64_t gotPltEntryAddr = sym.getGotPltVA(); 704 memcpy(buf, addrInst, sizeof(addrInst)); 705 relocateNoSym(buf, R_AARCH64_ADR_PREL_PG_HI21, 706 getAArch64Page(gotPltEntryAddr) - getAArch64Page(pltEntryAddr)); 707 relocateNoSym(buf + 4, R_AARCH64_LDST64_ABS_LO12_NC, gotPltEntryAddr); 708 relocateNoSym(buf + 8, R_AARCH64_ADD_ABS_LO12_NC, gotPltEntryAddr); 709 710 if (pacEntry) 711 memcpy(buf + sizeof(addrInst), pacBr, sizeof(pacBr)); 712 else 713 memcpy(buf + sizeof(addrInst), stdBr, sizeof(stdBr)); 714 if (!btiEntry) 715 // We didn't add the BTI c instruction so round out size with NOP. 716 memcpy(buf + sizeof(addrInst) + sizeof(stdBr), nopData, sizeof(nopData)); 717 } 718 719 static TargetInfo *getTargetInfo() { 720 if (config->andFeatures & (GNU_PROPERTY_AARCH64_FEATURE_1_BTI | 721 GNU_PROPERTY_AARCH64_FEATURE_1_PAC)) { 722 static AArch64BtiPac t; 723 return &t; 724 } 725 static AArch64 t; 726 return &t; 727 } 728 729 TargetInfo *elf::getAArch64TargetInfo() { return getTargetInfo(); } 730