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;
51 GotRel = R_386_GLOB_DAT;
52 PltRel = R_386_JUMP_SLOT;
53 IRelativeRel = R_386_IRELATIVE;
54 RelativeRel = R_386_RELATIVE;
55 TlsGotRel = R_386_TLS_TPOFF;
56 TlsModuleIndexRel = R_386_TLS_DTPMOD32;
57 TlsOffsetRel = R_386_TLS_DTPOFF32;
63 TrapInstr = 0xcccccccc; // 0xcc = INT3
66 static bool hasBaseReg(uint8_t ModRM) { return (ModRM & 0xc7) != 0x5; }
68 RelExpr X86::getRelExpr(RelType Type, const Symbol &S,
69 const uint8_t *Loc) const {
74 case R_386_TLS_LDO_32:
87 return R_GOTONLY_PC_FROM_END;
92 // These relocations are arguably mis-designed because their calculations
93 // depend on the instructions they are applied to. This is bad because we
94 // usually don't care about whether the target section contains valid
95 // machine instructions or not. But this is part of the documented ABI, so
96 // we had to implement as the standard requires.
98 // x86 does not support PC-relative data access. Therefore, in order to
99 // access GOT contents, a GOT address needs to be known at link-time
100 // (which means non-PIC) or compilers have to emit code to get a GOT
101 // address at runtime (which means code is position-independent but
102 // compilers need to emit extra code for each GOT access.) This decision
103 // is made at compile-time. In the latter case, compilers emit code to
104 // load an GOT address to a register, which is usually %ebx.
106 // So, there are two ways to refer to symbol foo's GOT entry: foo@GOT or
109 // foo@GOT is not usable in PIC. If we are creating a PIC output and if we
110 // find such relocation, we should report an error. foo@GOT is resolved to
111 // an *absolute* address of foo's GOT entry, because both GOT address and
112 // foo's offset are known. In other words, it's G + A.
114 // foo@GOT(%reg) needs to be resolved to a *relative* offset from a GOT to
115 // foo's GOT entry in the table, because GOT address is not known but foo's
116 // offset in the table is known. It's G + A - GOT.
118 // It's unfortunate that compilers emit the same relocation for these
119 // different use cases. In order to distinguish them, we have to read a
120 // machine instruction.
122 // The following code implements it. We assume that Loc[0] is the first
123 // byte of a displacement or an immediate field of a valid machine
124 // instruction. That means a ModRM byte is at Loc[-1]. By taking a look at
125 // the byte, we can determine whether the instruction is register-relative
126 // (i.e. it was generated for foo@GOT(%reg)) or absolute (i.e. foo@GOT).
127 return hasBaseReg(Loc[-1]) ? R_GOT_FROM_END : R_GOT;
128 case R_386_TLS_GOTIE:
129 return R_GOT_FROM_END;
131 return R_GOTREL_FROM_END;
134 case R_386_TLS_LE_32:
143 RelExpr X86::adjustRelaxExpr(RelType Type, const uint8_t *Data,
144 RelExpr Expr) const {
148 case R_RELAX_TLS_GD_TO_IE:
149 return R_RELAX_TLS_GD_TO_IE_END;
150 case R_RELAX_TLS_GD_TO_LE:
151 return R_RELAX_TLS_GD_TO_LE_NEG;
155 void X86::writeGotPltHeader(uint8_t *Buf) const {
156 write32le(Buf, InX::Dynamic->getVA());
159 void X86::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
160 // Entries in .got.plt initially points back to the corresponding
161 // PLT entries with a fixed offset to skip the first instruction.
162 write32le(Buf, S.getPltVA() + 6);
165 void X86::writeIgotPlt(uint8_t *Buf, const Symbol &S) const {
166 // An x86 entry is the address of the ifunc resolver function.
167 write32le(Buf, S.getVA());
170 RelType X86::getDynRel(RelType Type) const {
171 if (Type == R_386_TLS_LE)
172 return R_386_TLS_TPOFF;
173 if (Type == R_386_TLS_LE_32)
174 return R_386_TLS_TPOFF32;
178 void X86::writePltHeader(uint8_t *Buf) const {
180 const uint8_t V[] = {
181 0xff, 0xb3, 0x04, 0x00, 0x00, 0x00, // pushl GOTPLT+4(%ebx)
182 0xff, 0xa3, 0x08, 0x00, 0x00, 0x00, // jmp *GOTPLT+8(%ebx)
183 0x90, 0x90, 0x90, 0x90 // nop
185 memcpy(Buf, V, sizeof(V));
187 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
188 uint32_t GotPlt = InX::GotPlt->getVA() - Ebx;
189 write32le(Buf + 2, GotPlt + 4);
190 write32le(Buf + 8, GotPlt + 8);
194 const uint8_t PltData[] = {
195 0xff, 0x35, 0, 0, 0, 0, // pushl (GOTPLT+4)
196 0xff, 0x25, 0, 0, 0, 0, // jmp *(GOTPLT+8)
197 0x90, 0x90, 0x90, 0x90, // nop
199 memcpy(Buf, PltData, sizeof(PltData));
200 uint32_t GotPlt = InX::GotPlt->getVA();
201 write32le(Buf + 2, GotPlt + 4);
202 write32le(Buf + 8, GotPlt + 8);
205 void X86::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
206 uint64_t PltEntryAddr, int32_t Index,
207 unsigned RelOff) const {
208 const uint8_t Inst[] = {
209 0xff, 0x00, 0, 0, 0, 0, // jmp *foo_in_GOT or jmp *foo@GOT(%ebx)
210 0x68, 0, 0, 0, 0, // pushl $reloc_offset
211 0xe9, 0, 0, 0, 0, // jmp .PLT0@PC
213 memcpy(Buf, Inst, sizeof(Inst));
216 // jmp *foo@GOT(%ebx)
217 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
219 write32le(Buf + 2, GotPltEntryAddr - Ebx);
223 write32le(Buf + 2, GotPltEntryAddr);
226 write32le(Buf + 7, RelOff);
227 write32le(Buf + 12, -Index * PltEntrySize - PltHeaderSize - 16);
230 int64_t X86::getImplicitAddend(const uint8_t *Buf, RelType Type) const {
234 return SignExtend64<8>(*Buf);
237 return SignExtend64<16>(read16le(Buf));
245 case R_386_TLS_LDO_32:
247 return SignExtend64<32>(read32le(Buf));
253 void X86::relocateOne(uint8_t *Loc, RelType Type, uint64_t Val) const {
256 // R_386_{PC,}{8,16} are not part of the i386 psABI, but they are
257 // being used for some 16-bit programs such as boot loaders, so
258 // we want to support them.
259 checkUInt<8>(Loc, Val, Type);
263 checkInt<8>(Loc, Val, Type);
267 checkUInt<16>(Loc, Val, Type);
271 // R_386_PC16 is normally used with 16 bit code. In that situation
272 // the PC is 16 bits, just like the addend. This means that it can
273 // point from any 16 bit address to any other if the possibility
274 // of wrapping is included.
275 // The only restriction we have to check then is that the destination
276 // address fits in 16 bits. That is impossible to do here. The problem is
277 // that we are passed the final value, which already had the
278 // current location subtracted from it.
279 // We just check that Val fits in 17 bits. This misses some cases, but
280 // should have no false positives.
281 checkInt<17>(Loc, Val, Type);
293 case R_386_TLS_DTPMOD32:
294 case R_386_TLS_DTPOFF32:
296 case R_386_TLS_GOTIE:
299 case R_386_TLS_LDO_32:
301 case R_386_TLS_LE_32:
302 case R_386_TLS_TPOFF:
303 case R_386_TLS_TPOFF32:
304 checkInt<32>(Loc, Val, Type);
308 error(getErrorLocation(Loc) + "unrecognized reloc " + Twine(Type));
312 void X86::relaxTlsGdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
314 // leal x@tlsgd(, %ebx, 1),
315 // call __tls_get_addr@plt
318 // subl $x@ntpoff,%eax
319 const uint8_t Inst[] = {
320 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0, %eax
321 0x81, 0xe8, 0, 0, 0, 0, // subl Val(%ebx), %eax
323 memcpy(Loc - 3, Inst, sizeof(Inst));
324 write32le(Loc + 5, Val);
327 void X86::relaxTlsGdToIe(uint8_t *Loc, RelType Type, uint64_t Val) const {
329 // leal x@tlsgd(, %ebx, 1),
330 // call __tls_get_addr@plt
333 // addl x@gotntpoff(%ebx), %eax
334 const uint8_t Inst[] = {
335 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0, %eax
336 0x03, 0x83, 0, 0, 0, 0, // addl Val(%ebx), %eax
338 memcpy(Loc - 3, Inst, sizeof(Inst));
339 write32le(Loc + 5, Val);
342 // In some conditions, relocations can be optimized to avoid using GOT.
343 // This function does that for Initial Exec to Local Exec case.
344 void X86::relaxTlsIeToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
345 // Ulrich's document section 6.2 says that @gotntpoff can
346 // be used with MOVL or ADDL instructions.
347 // @indntpoff is similar to @gotntpoff, but for use in
348 // position dependent code.
349 uint8_t Reg = (Loc[-1] >> 3) & 7;
351 if (Type == R_386_TLS_IE) {
352 if (Loc[-1] == 0xa1) {
353 // "movl foo@indntpoff,%eax" -> "movl $foo,%eax"
354 // This case is different from the generic case below because
355 // this is a 5 byte instruction while below is 6 bytes.
357 } else if (Loc[-2] == 0x8b) {
358 // "movl foo@indntpoff,%reg" -> "movl $foo,%reg"
360 Loc[-1] = 0xc0 | Reg;
362 // "addl foo@indntpoff,%reg" -> "addl $foo,%reg"
364 Loc[-1] = 0xc0 | Reg;
367 assert(Type == R_386_TLS_GOTIE);
368 if (Loc[-2] == 0x8b) {
369 // "movl foo@gottpoff(%rip),%reg" -> "movl $foo,%reg"
371 Loc[-1] = 0xc0 | Reg;
373 // "addl foo@gotntpoff(%rip),%reg" -> "leal foo(%reg),%reg"
375 Loc[-1] = 0x80 | (Reg << 3) | Reg;
381 void X86::relaxTlsLdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
382 if (Type == R_386_TLS_LDO_32) {
388 // leal foo(%reg),%eax
389 // call ___tls_get_addr
393 // leal 0(%esi,1),%esi
394 const uint8_t Inst[] = {
395 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0,%eax
397 0x8d, 0x74, 0x26, 0x00, // leal 0(%esi,1),%esi
399 memcpy(Loc - 2, Inst, sizeof(Inst));
403 class RetpolinePic : public X86 {
406 void writeGotPlt(uint8_t *Buf, const Symbol &S) const override;
407 void writePltHeader(uint8_t *Buf) const override;
408 void writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr, uint64_t PltEntryAddr,
409 int32_t Index, unsigned RelOff) const override;
412 class RetpolineNoPic : public X86 {
415 void writeGotPlt(uint8_t *Buf, const Symbol &S) const override;
416 void writePltHeader(uint8_t *Buf) const override;
417 void writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr, uint64_t PltEntryAddr,
418 int32_t Index, unsigned RelOff) const override;
422 RetpolinePic::RetpolinePic() {
427 void RetpolinePic::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
428 write32le(Buf, S.getPltVA() + 17);
431 void RetpolinePic::writePltHeader(uint8_t *Buf) const {
432 const uint8_t Insn[] = {
433 0xff, 0xb3, 0, 0, 0, 0, // 0: pushl GOTPLT+4(%ebx)
434 0x50, // 6: pushl %eax
435 0x8b, 0x83, 0, 0, 0, 0, // 7: mov GOTPLT+8(%ebx), %eax
436 0xe8, 0x0e, 0x00, 0x00, 0x00, // d: call next
437 0xf3, 0x90, // 12: loop: pause
438 0x0f, 0xae, 0xe8, // 14: lfence
439 0xeb, 0xf9, // 17: jmp loop
440 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // 19: int3; .align 16
441 0x89, 0x0c, 0x24, // 20: next: mov %ecx, (%esp)
442 0x8b, 0x4c, 0x24, 0x04, // 23: mov 0x4(%esp), %ecx
443 0x89, 0x44, 0x24, 0x04, // 27: mov %eax ,0x4(%esp)
444 0x89, 0xc8, // 2b: mov %ecx, %eax
445 0x59, // 2d: pop %ecx
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
467 memcpy(Buf, Insn, sizeof(Insn));
469 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
470 write32le(Buf + 3, GotPltEntryAddr - Ebx);
471 write32le(Buf + 8, -Index * PltEntrySize - PltHeaderSize - 12 + 32);
472 write32le(Buf + 13, -Index * PltEntrySize - PltHeaderSize - 17 + 18);
473 write32le(Buf + 18, RelOff);
474 write32le(Buf + 23, -Index * PltEntrySize - PltHeaderSize - 27);
477 RetpolineNoPic::RetpolineNoPic() {
482 void RetpolineNoPic::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
483 write32le(Buf, S.getPltVA() + 16);
486 void RetpolineNoPic::writePltHeader(uint8_t *Buf) const {
487 const uint8_t PltData[] = {
488 0xff, 0x35, 0, 0, 0, 0, // 0: pushl GOTPLT+4
489 0x50, // 6: pushl %eax
490 0xa1, 0, 0, 0, 0, // 7: mov GOTPLT+8, %eax
491 0xe8, 0x0f, 0x00, 0x00, 0x00, // c: call next
492 0xf3, 0x90, // 11: loop: pause
493 0x0f, 0xae, 0xe8, // 13: lfence
494 0xeb, 0xf9, // 16: jmp loop
495 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // 18: int3
496 0xcc, 0xcc, 0xcc, // 1f: int3; .align 16
497 0x89, 0x0c, 0x24, // 20: next: mov %ecx, (%esp)
498 0x8b, 0x4c, 0x24, 0x04, // 23: mov 0x4(%esp), %ecx
499 0x89, 0x44, 0x24, 0x04, // 27: mov %eax ,0x4(%esp)
500 0x89, 0xc8, // 2b: mov %ecx, %eax
501 0x59, // 2d: pop %ecx
504 memcpy(Buf, PltData, sizeof(PltData));
506 uint32_t GotPlt = InX::GotPlt->getVA();
507 write32le(Buf + 2, GotPlt + 4);
508 write32le(Buf + 8, GotPlt + 8);
511 void RetpolineNoPic::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
512 uint64_t PltEntryAddr, int32_t Index,
513 unsigned RelOff) const {
514 const uint8_t Insn[] = {
515 0x50, // 0: pushl %eax
516 0xa1, 0, 0, 0, 0, // 1: mov foo_in_GOT, %eax
517 0xe8, 0, 0, 0, 0, // 6: call plt+0x20
518 0xe9, 0, 0, 0, 0, // b: jmp plt+0x11
519 0x68, 0, 0, 0, 0, // 10: pushl $reloc_offset
520 0xe9, 0, 0, 0, 0, // 15: jmp plt+0
522 memcpy(Buf, Insn, sizeof(Insn));
524 write32le(Buf + 2, GotPltEntryAddr);
525 write32le(Buf + 7, -Index * PltEntrySize - PltHeaderSize - 11 + 32);
526 write32le(Buf + 12, -Index * PltEntrySize - PltHeaderSize - 16 + 17);
527 write32le(Buf + 17, RelOff);
528 write32le(Buf + 22, -Index * PltEntrySize - PltHeaderSize - 26);
531 TargetInfo *elf::getX86TargetInfo() {
532 if (Config->ZRetpolineplt) {
534 static RetpolinePic T;
537 static RetpolineNoPic T;