1 //===- X86.cpp ------------------------------------------------------------===//
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "InputFiles.h"
12 #include "SyntheticSections.h"
14 #include "lld/Common/ErrorHandler.h"
15 #include "llvm/Support/Endian.h"
18 using namespace llvm::support::endian;
19 using namespace llvm::ELF;
21 using namespace lld::elf;
24 class X86 : public TargetInfo {
27 RelExpr getRelExpr(RelType Type, const Symbol &S,
28 const uint8_t *Loc) const override;
29 int64_t getImplicitAddend(const uint8_t *Buf, RelType Type) const override;
30 void writeGotPltHeader(uint8_t *Buf) const override;
31 RelType getDynRel(RelType Type) const override;
32 void writeGotPlt(uint8_t *Buf, const Symbol &S) const override;
33 void writeIgotPlt(uint8_t *Buf, const Symbol &S) const override;
34 void writePltHeader(uint8_t *Buf) const override;
35 void writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr, uint64_t PltEntryAddr,
36 int32_t Index, unsigned RelOff) const override;
37 void relocateOne(uint8_t *Loc, RelType Type, uint64_t Val) const override;
39 RelExpr adjustRelaxExpr(RelType Type, const uint8_t *Data,
40 RelExpr Expr) const override;
41 void relaxTlsGdToIe(uint8_t *Loc, RelType Type, uint64_t Val) const override;
42 void relaxTlsGdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const override;
43 void relaxTlsIeToLe(uint8_t *Loc, RelType Type, uint64_t Val) const override;
44 void relaxTlsLdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const override;
50 GotRel = R_386_GLOB_DAT;
51 PltRel = R_386_JUMP_SLOT;
52 IRelativeRel = R_386_IRELATIVE;
53 RelativeRel = R_386_RELATIVE;
54 TlsGotRel = R_386_TLS_TPOFF;
55 TlsModuleIndexRel = R_386_TLS_DTPMOD32;
56 TlsOffsetRel = R_386_TLS_DTPOFF32;
62 TrapInstr = 0xcccccccc; // 0xcc = INT3
65 static bool hasBaseReg(uint8_t ModRM) { return (ModRM & 0xc7) != 0x5; }
67 RelExpr X86::getRelExpr(RelType Type, const Symbol &S,
68 const uint8_t *Loc) const {
73 case R_386_TLS_LDO_32:
76 return R_TLSGD_GOT_FROM_END;
78 return R_TLSLD_GOT_FROM_END;
86 return R_GOTONLY_PC_FROM_END;
91 // These relocations are arguably mis-designed because their calculations
92 // depend on the instructions they are applied to. This is bad because we
93 // usually don't care about whether the target section contains valid
94 // machine instructions or not. But this is part of the documented ABI, so
95 // we had to implement as the standard requires.
97 // x86 does not support PC-relative data access. Therefore, in order to
98 // access GOT contents, a GOT address needs to be known at link-time
99 // (which means non-PIC) or compilers have to emit code to get a GOT
100 // address at runtime (which means code is position-independent but
101 // compilers need to emit extra code for each GOT access.) This decision
102 // is made at compile-time. In the latter case, compilers emit code to
103 // load an GOT address to a register, which is usually %ebx.
105 // So, there are two ways to refer to symbol foo's GOT entry: foo@GOT or
108 // foo@GOT is not usable in PIC. If we are creating a PIC output and if we
109 // find such relocation, we should report an error. foo@GOT is resolved to
110 // an *absolute* address of foo's GOT entry, because both GOT address and
111 // foo's offset are known. In other words, it's G + A.
113 // foo@GOT(%reg) needs to be resolved to a *relative* offset from a GOT to
114 // foo's GOT entry in the table, because GOT address is not known but foo's
115 // offset in the table is known. It's G + A - GOT.
117 // It's unfortunate that compilers emit the same relocation for these
118 // different use cases. In order to distinguish them, we have to read a
119 // machine instruction.
121 // The following code implements it. We assume that Loc[0] is the first
122 // byte of a displacement or an immediate field of a valid machine
123 // instruction. That means a ModRM byte is at Loc[-1]. By taking a look at
124 // the byte, we can determine whether the instruction is register-relative
125 // (i.e. it was generated for foo@GOT(%reg)) or absolute (i.e. foo@GOT).
126 return hasBaseReg(Loc[-1]) ? R_GOT_FROM_END : R_GOT;
127 case R_386_TLS_GOTIE:
128 return R_GOT_FROM_END;
130 return R_GOTREL_FROM_END;
133 case R_386_TLS_LE_32:
142 RelExpr X86::adjustRelaxExpr(RelType Type, const uint8_t *Data,
143 RelExpr Expr) const {
147 case R_RELAX_TLS_GD_TO_IE:
148 return R_RELAX_TLS_GD_TO_IE_END;
149 case R_RELAX_TLS_GD_TO_LE:
150 return R_RELAX_TLS_GD_TO_LE_NEG;
154 void X86::writeGotPltHeader(uint8_t *Buf) const {
155 write32le(Buf, InX::Dynamic->getVA());
158 void X86::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
159 // Entries in .got.plt initially points back to the corresponding
160 // PLT entries with a fixed offset to skip the first instruction.
161 write32le(Buf, S.getPltVA() + 6);
164 void X86::writeIgotPlt(uint8_t *Buf, const Symbol &S) const {
165 // An x86 entry is the address of the ifunc resolver function.
166 write32le(Buf, S.getVA());
169 RelType X86::getDynRel(RelType Type) const {
170 if (Type == R_386_TLS_LE)
171 return R_386_TLS_TPOFF;
172 if (Type == R_386_TLS_LE_32)
173 return R_386_TLS_TPOFF32;
177 void X86::writePltHeader(uint8_t *Buf) const {
179 const uint8_t V[] = {
180 0xff, 0xb3, 0x04, 0x00, 0x00, 0x00, // pushl GOTPLT+4(%ebx)
181 0xff, 0xa3, 0x08, 0x00, 0x00, 0x00, // jmp *GOTPLT+8(%ebx)
182 0x90, 0x90, 0x90, 0x90 // nop
184 memcpy(Buf, V, sizeof(V));
186 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
187 uint32_t GotPlt = InX::GotPlt->getVA() - Ebx;
188 write32le(Buf + 2, GotPlt + 4);
189 write32le(Buf + 8, GotPlt + 8);
193 const uint8_t PltData[] = {
194 0xff, 0x35, 0, 0, 0, 0, // pushl (GOTPLT+4)
195 0xff, 0x25, 0, 0, 0, 0, // jmp *(GOTPLT+8)
196 0x90, 0x90, 0x90, 0x90, // nop
198 memcpy(Buf, PltData, sizeof(PltData));
199 uint32_t GotPlt = InX::GotPlt->getVA();
200 write32le(Buf + 2, GotPlt + 4);
201 write32le(Buf + 8, GotPlt + 8);
204 void X86::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
205 uint64_t PltEntryAddr, int32_t Index,
206 unsigned RelOff) const {
207 const uint8_t Inst[] = {
208 0xff, 0x00, 0, 0, 0, 0, // jmp *foo_in_GOT or jmp *foo@GOT(%ebx)
209 0x68, 0, 0, 0, 0, // pushl $reloc_offset
210 0xe9, 0, 0, 0, 0, // jmp .PLT0@PC
212 memcpy(Buf, Inst, sizeof(Inst));
215 // jmp *foo@GOT(%ebx)
216 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
218 write32le(Buf + 2, GotPltEntryAddr - Ebx);
222 write32le(Buf + 2, GotPltEntryAddr);
225 write32le(Buf + 7, RelOff);
226 write32le(Buf + 12, -getPltEntryOffset(Index) - 16);
229 int64_t X86::getImplicitAddend(const uint8_t *Buf, RelType Type) const {
233 return SignExtend64<8>(*Buf);
236 return SignExtend64<16>(read16le(Buf));
244 case R_386_TLS_LDO_32:
246 return SignExtend64<32>(read32le(Buf));
252 void X86::relocateOne(uint8_t *Loc, RelType Type, uint64_t Val) const {
255 // R_386_{PC,}{8,16} are not part of the i386 psABI, but they are
256 // being used for some 16-bit programs such as boot loaders, so
257 // we want to support them.
258 checkIntUInt(Loc, Val, 8, Type);
262 checkInt(Loc, Val, 8, Type);
266 checkIntUInt(Loc, Val, 16, Type);
270 // R_386_PC16 is normally used with 16 bit code. In that situation
271 // the PC is 16 bits, just like the addend. This means that it can
272 // point from any 16 bit address to any other if the possibility
273 // of wrapping is included.
274 // The only restriction we have to check then is that the destination
275 // address fits in 16 bits. That is impossible to do here. The problem is
276 // that we are passed the final value, which already had the
277 // current location subtracted from it.
278 // We just check that Val fits in 17 bits. This misses some cases, but
279 // should have no false positives.
280 checkInt(Loc, Val, 17, Type);
292 case R_386_TLS_DTPMOD32:
293 case R_386_TLS_DTPOFF32:
295 case R_386_TLS_GOTIE:
298 case R_386_TLS_LDO_32:
300 case R_386_TLS_LE_32:
301 case R_386_TLS_TPOFF:
302 case R_386_TLS_TPOFF32:
303 checkInt(Loc, Val, 32, Type);
307 error(getErrorLocation(Loc) + "unrecognized reloc " + Twine(Type));
311 void X86::relaxTlsGdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
313 // leal x@tlsgd(, %ebx, 1),
314 // call __tls_get_addr@plt
317 // subl $x@ntpoff,%eax
318 const uint8_t Inst[] = {
319 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0, %eax
320 0x81, 0xe8, 0, 0, 0, 0, // subl Val(%ebx), %eax
322 memcpy(Loc - 3, Inst, sizeof(Inst));
323 write32le(Loc + 5, Val);
326 void X86::relaxTlsGdToIe(uint8_t *Loc, RelType Type, uint64_t Val) const {
328 // leal x@tlsgd(, %ebx, 1),
329 // call __tls_get_addr@plt
332 // addl x@gotntpoff(%ebx), %eax
333 const uint8_t Inst[] = {
334 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0, %eax
335 0x03, 0x83, 0, 0, 0, 0, // addl Val(%ebx), %eax
337 memcpy(Loc - 3, Inst, sizeof(Inst));
338 write32le(Loc + 5, Val);
341 // In some conditions, relocations can be optimized to avoid using GOT.
342 // This function does that for Initial Exec to Local Exec case.
343 void X86::relaxTlsIeToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
344 // Ulrich's document section 6.2 says that @gotntpoff can
345 // be used with MOVL or ADDL instructions.
346 // @indntpoff is similar to @gotntpoff, but for use in
347 // position dependent code.
348 uint8_t Reg = (Loc[-1] >> 3) & 7;
350 if (Type == R_386_TLS_IE) {
351 if (Loc[-1] == 0xa1) {
352 // "movl foo@indntpoff,%eax" -> "movl $foo,%eax"
353 // This case is different from the generic case below because
354 // this is a 5 byte instruction while below is 6 bytes.
356 } else if (Loc[-2] == 0x8b) {
357 // "movl foo@indntpoff,%reg" -> "movl $foo,%reg"
359 Loc[-1] = 0xc0 | Reg;
361 // "addl foo@indntpoff,%reg" -> "addl $foo,%reg"
363 Loc[-1] = 0xc0 | Reg;
366 assert(Type == R_386_TLS_GOTIE);
367 if (Loc[-2] == 0x8b) {
368 // "movl foo@gottpoff(%rip),%reg" -> "movl $foo,%reg"
370 Loc[-1] = 0xc0 | Reg;
372 // "addl foo@gotntpoff(%rip),%reg" -> "leal foo(%reg),%reg"
374 Loc[-1] = 0x80 | (Reg << 3) | Reg;
380 void X86::relaxTlsLdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
381 if (Type == R_386_TLS_LDO_32) {
387 // leal foo(%reg),%eax
388 // call ___tls_get_addr
392 // leal 0(%esi,1),%esi
393 const uint8_t Inst[] = {
394 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0,%eax
396 0x8d, 0x74, 0x26, 0x00, // leal 0(%esi,1),%esi
398 memcpy(Loc - 2, Inst, sizeof(Inst));
402 class RetpolinePic : public X86 {
405 void writeGotPlt(uint8_t *Buf, const Symbol &S) const override;
406 void writePltHeader(uint8_t *Buf) const override;
407 void writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr, uint64_t PltEntryAddr,
408 int32_t Index, unsigned RelOff) const override;
411 class RetpolineNoPic : public X86 {
414 void writeGotPlt(uint8_t *Buf, const Symbol &S) const override;
415 void writePltHeader(uint8_t *Buf) const override;
416 void writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr, uint64_t PltEntryAddr,
417 int32_t Index, unsigned RelOff) const override;
421 RetpolinePic::RetpolinePic() {
426 void RetpolinePic::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
427 write32le(Buf, S.getPltVA() + 17);
430 void RetpolinePic::writePltHeader(uint8_t *Buf) const {
431 const uint8_t Insn[] = {
432 0xff, 0xb3, 0, 0, 0, 0, // 0: pushl GOTPLT+4(%ebx)
433 0x50, // 6: pushl %eax
434 0x8b, 0x83, 0, 0, 0, 0, // 7: mov GOTPLT+8(%ebx), %eax
435 0xe8, 0x0e, 0x00, 0x00, 0x00, // d: call next
436 0xf3, 0x90, // 12: loop: pause
437 0x0f, 0xae, 0xe8, // 14: lfence
438 0xeb, 0xf9, // 17: jmp loop
439 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // 19: int3; .align 16
440 0x89, 0x0c, 0x24, // 20: next: mov %ecx, (%esp)
441 0x8b, 0x4c, 0x24, 0x04, // 23: mov 0x4(%esp), %ecx
442 0x89, 0x44, 0x24, 0x04, // 27: mov %eax ,0x4(%esp)
443 0x89, 0xc8, // 2b: mov %ecx, %eax
444 0x59, // 2d: pop %ecx
446 0xcc, // 2f: int3; padding
448 memcpy(Buf, Insn, sizeof(Insn));
450 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
451 uint32_t GotPlt = InX::GotPlt->getVA() - Ebx;
452 write32le(Buf + 2, GotPlt + 4);
453 write32le(Buf + 9, GotPlt + 8);
456 void RetpolinePic::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
457 uint64_t PltEntryAddr, int32_t Index,
458 unsigned RelOff) const {
459 const uint8_t Insn[] = {
461 0x8b, 0x83, 0, 0, 0, 0, // mov foo@GOT(%ebx), %eax
462 0xe8, 0, 0, 0, 0, // call plt+0x20
463 0xe9, 0, 0, 0, 0, // jmp plt+0x12
464 0x68, 0, 0, 0, 0, // pushl $reloc_offset
465 0xe9, 0, 0, 0, 0, // jmp plt+0
466 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // int3; padding
468 memcpy(Buf, Insn, sizeof(Insn));
470 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
471 unsigned Off = getPltEntryOffset(Index);
472 write32le(Buf + 3, GotPltEntryAddr - Ebx);
473 write32le(Buf + 8, -Off - 12 + 32);
474 write32le(Buf + 13, -Off - 17 + 18);
475 write32le(Buf + 18, RelOff);
476 write32le(Buf + 23, -Off - 27);
479 RetpolineNoPic::RetpolineNoPic() {
484 void RetpolineNoPic::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
485 write32le(Buf, S.getPltVA() + 16);
488 void RetpolineNoPic::writePltHeader(uint8_t *Buf) const {
489 const uint8_t Insn[] = {
490 0xff, 0x35, 0, 0, 0, 0, // 0: pushl GOTPLT+4
491 0x50, // 6: pushl %eax
492 0xa1, 0, 0, 0, 0, // 7: mov GOTPLT+8, %eax
493 0xe8, 0x0f, 0x00, 0x00, 0x00, // c: call next
494 0xf3, 0x90, // 11: loop: pause
495 0x0f, 0xae, 0xe8, // 13: lfence
496 0xeb, 0xf9, // 16: jmp loop
497 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // 18: int3
498 0xcc, 0xcc, 0xcc, // 1f: int3; .align 16
499 0x89, 0x0c, 0x24, // 20: next: mov %ecx, (%esp)
500 0x8b, 0x4c, 0x24, 0x04, // 23: mov 0x4(%esp), %ecx
501 0x89, 0x44, 0x24, 0x04, // 27: mov %eax ,0x4(%esp)
502 0x89, 0xc8, // 2b: mov %ecx, %eax
503 0x59, // 2d: pop %ecx
505 0xcc, // 2f: int3; padding
507 memcpy(Buf, Insn, sizeof(Insn));
509 uint32_t GotPlt = InX::GotPlt->getVA();
510 write32le(Buf + 2, GotPlt + 4);
511 write32le(Buf + 8, GotPlt + 8);
514 void RetpolineNoPic::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
515 uint64_t PltEntryAddr, int32_t Index,
516 unsigned RelOff) const {
517 const uint8_t Insn[] = {
518 0x50, // 0: pushl %eax
519 0xa1, 0, 0, 0, 0, // 1: mov foo_in_GOT, %eax
520 0xe8, 0, 0, 0, 0, // 6: call plt+0x20
521 0xe9, 0, 0, 0, 0, // b: jmp plt+0x11
522 0x68, 0, 0, 0, 0, // 10: pushl $reloc_offset
523 0xe9, 0, 0, 0, 0, // 15: jmp plt+0
524 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // 1a: int3; padding
525 0xcc, // 1f: int3; padding
527 memcpy(Buf, Insn, sizeof(Insn));
529 unsigned Off = getPltEntryOffset(Index);
530 write32le(Buf + 2, GotPltEntryAddr);
531 write32le(Buf + 7, -Off - 11 + 32);
532 write32le(Buf + 12, -Off - 16 + 17);
533 write32le(Buf + 17, RelOff);
534 write32le(Buf + 22, -Off - 26);
537 TargetInfo *elf::getX86TargetInfo() {
538 if (Config->ZRetpolineplt) {
540 static RetpolinePic T;
543 static RetpolineNoPic T;