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
65 // Align to the non-PAE large page size (known as a superpage or huge page).
66 // FreeBSD automatically promotes large, superpage-aligned allocations.
67 DefaultImageBase = 0x400000;
70 static bool hasBaseReg(uint8_t ModRM) { return (ModRM & 0xc7) != 0x5; }
72 RelExpr X86::getRelExpr(RelType Type, const Symbol &S,
73 const uint8_t *Loc) const {
78 case R_386_TLS_LDO_32:
91 return R_GOTONLY_PC_FROM_END;
96 // These relocations are arguably mis-designed because their calculations
97 // depend on the instructions they are applied to. This is bad because we
98 // usually don't care about whether the target section contains valid
99 // machine instructions or not. But this is part of the documented ABI, so
100 // we had to implement as the standard requires.
102 // x86 does not support PC-relative data access. Therefore, in order to
103 // access GOT contents, a GOT address needs to be known at link-time
104 // (which means non-PIC) or compilers have to emit code to get a GOT
105 // address at runtime (which means code is position-independent but
106 // compilers need to emit extra code for each GOT access.) This decision
107 // is made at compile-time. In the latter case, compilers emit code to
108 // load an GOT address to a register, which is usually %ebx.
110 // So, there are two ways to refer to symbol foo's GOT entry: foo@GOT or
113 // foo@GOT is not usable in PIC. If we are creating a PIC output and if we
114 // find such relocation, we should report an error. foo@GOT is resolved to
115 // an *absolute* address of foo's GOT entry, because both GOT address and
116 // foo's offset are known. In other words, it's G + A.
118 // foo@GOT(%reg) needs to be resolved to a *relative* offset from a GOT to
119 // foo's GOT entry in the table, because GOT address is not known but foo's
120 // offset in the table is known. It's G + A - GOT.
122 // It's unfortunate that compilers emit the same relocation for these
123 // different use cases. In order to distinguish them, we have to read a
124 // machine instruction.
126 // The following code implements it. We assume that Loc[0] is the first
127 // byte of a displacement or an immediate field of a valid machine
128 // instruction. That means a ModRM byte is at Loc[-1]. By taking a look at
129 // the byte, we can determine whether the instruction is register-relative
130 // (i.e. it was generated for foo@GOT(%reg)) or absolute (i.e. foo@GOT).
131 return hasBaseReg(Loc[-1]) ? R_GOT_FROM_END : R_GOT;
132 case R_386_TLS_GOTIE:
133 return R_GOT_FROM_END;
135 return R_GOTREL_FROM_END;
138 case R_386_TLS_LE_32:
147 RelExpr X86::adjustRelaxExpr(RelType Type, const uint8_t *Data,
148 RelExpr Expr) const {
152 case R_RELAX_TLS_GD_TO_IE:
153 return R_RELAX_TLS_GD_TO_IE_END;
154 case R_RELAX_TLS_GD_TO_LE:
155 return R_RELAX_TLS_GD_TO_LE_NEG;
159 void X86::writeGotPltHeader(uint8_t *Buf) const {
160 write32le(Buf, InX::Dynamic->getVA());
163 void X86::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
164 // Entries in .got.plt initially points back to the corresponding
165 // PLT entries with a fixed offset to skip the first instruction.
166 write32le(Buf, S.getPltVA() + 6);
169 void X86::writeIgotPlt(uint8_t *Buf, const Symbol &S) const {
170 // An x86 entry is the address of the ifunc resolver function.
171 write32le(Buf, S.getVA());
174 RelType X86::getDynRel(RelType Type) const {
175 if (Type == R_386_TLS_LE)
176 return R_386_TLS_TPOFF;
177 if (Type == R_386_TLS_LE_32)
178 return R_386_TLS_TPOFF32;
182 void X86::writePltHeader(uint8_t *Buf) const {
184 const uint8_t V[] = {
185 0xff, 0xb3, 0x04, 0x00, 0x00, 0x00, // pushl GOTPLT+4(%ebx)
186 0xff, 0xa3, 0x08, 0x00, 0x00, 0x00, // jmp *GOTPLT+8(%ebx)
187 0x90, 0x90, 0x90, 0x90 // nop
189 memcpy(Buf, V, sizeof(V));
191 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
192 uint32_t GotPlt = InX::GotPlt->getVA() - Ebx;
193 write32le(Buf + 2, GotPlt + 4);
194 write32le(Buf + 8, GotPlt + 8);
198 const uint8_t PltData[] = {
199 0xff, 0x35, 0, 0, 0, 0, // pushl (GOTPLT+4)
200 0xff, 0x25, 0, 0, 0, 0, // jmp *(GOTPLT+8)
201 0x90, 0x90, 0x90, 0x90, // nop
203 memcpy(Buf, PltData, sizeof(PltData));
204 uint32_t GotPlt = InX::GotPlt->getVA();
205 write32le(Buf + 2, GotPlt + 4);
206 write32le(Buf + 8, GotPlt + 8);
209 void X86::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
210 uint64_t PltEntryAddr, int32_t Index,
211 unsigned RelOff) const {
212 const uint8_t Inst[] = {
213 0xff, 0x00, 0, 0, 0, 0, // jmp *foo_in_GOT or jmp *foo@GOT(%ebx)
214 0x68, 0, 0, 0, 0, // pushl $reloc_offset
215 0xe9, 0, 0, 0, 0, // jmp .PLT0@PC
217 memcpy(Buf, Inst, sizeof(Inst));
220 // jmp *foo@GOT(%ebx)
221 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
223 write32le(Buf + 2, GotPltEntryAddr - Ebx);
227 write32le(Buf + 2, GotPltEntryAddr);
230 write32le(Buf + 7, RelOff);
231 write32le(Buf + 12, -Index * PltEntrySize - PltHeaderSize - 16);
234 int64_t X86::getImplicitAddend(const uint8_t *Buf, RelType Type) const {
238 return SignExtend64<8>(*Buf);
241 return SignExtend64<16>(read16le(Buf));
249 case R_386_TLS_LDO_32:
251 return SignExtend64<32>(read32le(Buf));
257 void X86::relocateOne(uint8_t *Loc, RelType Type, uint64_t Val) const {
260 // R_386_{PC,}{8,16} are not part of the i386 psABI, but they are
261 // being used for some 16-bit programs such as boot loaders, so
262 // we want to support them.
263 checkUInt<8>(Loc, Val, Type);
267 checkInt<8>(Loc, Val, Type);
271 checkUInt<16>(Loc, Val, Type);
275 // R_386_PC16 is normally used with 16 bit code. In that situation
276 // the PC is 16 bits, just like the addend. This means that it can
277 // point from any 16 bit address to any other if the possibility
278 // of wrapping is included.
279 // The only restriction we have to check then is that the destination
280 // address fits in 16 bits. That is impossible to do here. The problem is
281 // that we are passed the final value, which already had the
282 // current location subtracted from it.
283 // We just check that Val fits in 17 bits. This misses some cases, but
284 // should have no false positives.
285 checkInt<17>(Loc, Val, Type);
297 case R_386_TLS_DTPMOD32:
298 case R_386_TLS_DTPOFF32:
300 case R_386_TLS_GOTIE:
303 case R_386_TLS_LDO_32:
305 case R_386_TLS_LE_32:
306 case R_386_TLS_TPOFF:
307 case R_386_TLS_TPOFF32:
308 checkInt<32>(Loc, Val, Type);
312 error(getErrorLocation(Loc) + "unrecognized reloc " + Twine(Type));
316 void X86::relaxTlsGdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
318 // leal x@tlsgd(, %ebx, 1),
319 // call __tls_get_addr@plt
322 // subl $x@ntpoff,%eax
323 const uint8_t Inst[] = {
324 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0, %eax
325 0x81, 0xe8, 0, 0, 0, 0, // subl Val(%ebx), %eax
327 memcpy(Loc - 3, Inst, sizeof(Inst));
328 write32le(Loc + 5, Val);
331 void X86::relaxTlsGdToIe(uint8_t *Loc, RelType Type, uint64_t Val) const {
333 // leal x@tlsgd(, %ebx, 1),
334 // call __tls_get_addr@plt
337 // addl x@gotntpoff(%ebx), %eax
338 const uint8_t Inst[] = {
339 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0, %eax
340 0x03, 0x83, 0, 0, 0, 0, // addl Val(%ebx), %eax
342 memcpy(Loc - 3, Inst, sizeof(Inst));
343 write32le(Loc + 5, Val);
346 // In some conditions, relocations can be optimized to avoid using GOT.
347 // This function does that for Initial Exec to Local Exec case.
348 void X86::relaxTlsIeToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
349 // Ulrich's document section 6.2 says that @gotntpoff can
350 // be used with MOVL or ADDL instructions.
351 // @indntpoff is similar to @gotntpoff, but for use in
352 // position dependent code.
353 uint8_t Reg = (Loc[-1] >> 3) & 7;
355 if (Type == R_386_TLS_IE) {
356 if (Loc[-1] == 0xa1) {
357 // "movl foo@indntpoff,%eax" -> "movl $foo,%eax"
358 // This case is different from the generic case below because
359 // this is a 5 byte instruction while below is 6 bytes.
361 } else if (Loc[-2] == 0x8b) {
362 // "movl foo@indntpoff,%reg" -> "movl $foo,%reg"
364 Loc[-1] = 0xc0 | Reg;
366 // "addl foo@indntpoff,%reg" -> "addl $foo,%reg"
368 Loc[-1] = 0xc0 | Reg;
371 assert(Type == R_386_TLS_GOTIE);
372 if (Loc[-2] == 0x8b) {
373 // "movl foo@gottpoff(%rip),%reg" -> "movl $foo,%reg"
375 Loc[-1] = 0xc0 | Reg;
377 // "addl foo@gotntpoff(%rip),%reg" -> "leal foo(%reg),%reg"
379 Loc[-1] = 0x80 | (Reg << 3) | Reg;
385 void X86::relaxTlsLdToLe(uint8_t *Loc, RelType Type, uint64_t Val) const {
386 if (Type == R_386_TLS_LDO_32) {
392 // leal foo(%reg),%eax
393 // call ___tls_get_addr
397 // leal 0(%esi,1),%esi
398 const uint8_t Inst[] = {
399 0x65, 0xa1, 0x00, 0x00, 0x00, 0x00, // movl %gs:0,%eax
401 0x8d, 0x74, 0x26, 0x00, // leal 0(%esi,1),%esi
403 memcpy(Loc - 2, Inst, sizeof(Inst));
407 class RetpolinePic : public X86 {
410 void writeGotPlt(uint8_t *Buf, const Symbol &S) const override;
411 void writePltHeader(uint8_t *Buf) const override;
412 void writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr, uint64_t PltEntryAddr,
413 int32_t Index, unsigned RelOff) const override;
416 class RetpolineNoPic : public X86 {
419 void writeGotPlt(uint8_t *Buf, const Symbol &S) const override;
420 void writePltHeader(uint8_t *Buf) const override;
421 void writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr, uint64_t PltEntryAddr,
422 int32_t Index, unsigned RelOff) const override;
426 RetpolinePic::RetpolinePic() {
431 void RetpolinePic::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
432 write32le(Buf, S.getPltVA() + 17);
435 void RetpolinePic::writePltHeader(uint8_t *Buf) const {
436 const uint8_t Insn[] = {
437 0xff, 0xb3, 0, 0, 0, 0, // 0: pushl GOTPLT+4(%ebx)
438 0x50, // 6: pushl %eax
439 0x8b, 0x83, 0, 0, 0, 0, // 7: mov GOTPLT+8(%ebx), %eax
440 0xe8, 0x0e, 0x00, 0x00, 0x00, // d: call next
441 0xf3, 0x90, // 12: loop: pause
442 0x0f, 0xae, 0xe8, // 14: lfence
443 0xeb, 0xf9, // 17: jmp loop
444 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // 19: int3; .align 16
445 0x89, 0x0c, 0x24, // 20: next: mov %ecx, (%esp)
446 0x8b, 0x4c, 0x24, 0x04, // 23: mov 0x4(%esp), %ecx
447 0x89, 0x44, 0x24, 0x04, // 27: mov %eax ,0x4(%esp)
448 0x89, 0xc8, // 2b: mov %ecx, %eax
449 0x59, // 2d: pop %ecx
452 memcpy(Buf, Insn, sizeof(Insn));
454 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
455 uint32_t GotPlt = InX::GotPlt->getVA() - Ebx;
456 write32le(Buf + 2, GotPlt + 4);
457 write32le(Buf + 9, GotPlt + 8);
460 void RetpolinePic::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
461 uint64_t PltEntryAddr, int32_t Index,
462 unsigned RelOff) const {
463 const uint8_t Insn[] = {
465 0x8b, 0x83, 0, 0, 0, 0, // mov foo@GOT(%ebx), %eax
466 0xe8, 0, 0, 0, 0, // call plt+0x20
467 0xe9, 0, 0, 0, 0, // jmp plt+0x12
468 0x68, 0, 0, 0, 0, // pushl $reloc_offset
469 0xe9, 0, 0, 0, 0, // jmp plt+0
471 memcpy(Buf, Insn, sizeof(Insn));
473 uint32_t Ebx = InX::Got->getVA() + InX::Got->getSize();
474 write32le(Buf + 3, GotPltEntryAddr - Ebx);
475 write32le(Buf + 8, -Index * PltEntrySize - PltHeaderSize - 12 + 32);
476 write32le(Buf + 13, -Index * PltEntrySize - PltHeaderSize - 17 + 18);
477 write32le(Buf + 18, RelOff);
478 write32le(Buf + 23, -Index * PltEntrySize - PltHeaderSize - 27);
481 RetpolineNoPic::RetpolineNoPic() {
486 void RetpolineNoPic::writeGotPlt(uint8_t *Buf, const Symbol &S) const {
487 write32le(Buf, S.getPltVA() + 16);
490 void RetpolineNoPic::writePltHeader(uint8_t *Buf) const {
491 const uint8_t PltData[] = {
492 0xff, 0x35, 0, 0, 0, 0, // 0: pushl GOTPLT+4
493 0x50, // 6: pushl %eax
494 0xa1, 0, 0, 0, 0, // 7: mov GOTPLT+8, %eax
495 0xe8, 0x0f, 0x00, 0x00, 0x00, // c: call next
496 0xf3, 0x90, // 11: loop: pause
497 0x0f, 0xae, 0xe8, // 13: lfence
498 0xeb, 0xf9, // 16: jmp loop
499 0xcc, 0xcc, 0xcc, 0xcc, 0xcc, // 18: int3
500 0xcc, 0xcc, 0xcc, // 1f: int3; .align 16
501 0x89, 0x0c, 0x24, // 20: next: mov %ecx, (%esp)
502 0x8b, 0x4c, 0x24, 0x04, // 23: mov 0x4(%esp), %ecx
503 0x89, 0x44, 0x24, 0x04, // 27: mov %eax ,0x4(%esp)
504 0x89, 0xc8, // 2b: mov %ecx, %eax
505 0x59, // 2d: pop %ecx
508 memcpy(Buf, PltData, sizeof(PltData));
510 uint32_t GotPlt = InX::GotPlt->getVA();
511 write32le(Buf + 2, GotPlt + 4);
512 write32le(Buf + 8, GotPlt + 8);
515 void RetpolineNoPic::writePlt(uint8_t *Buf, uint64_t GotPltEntryAddr,
516 uint64_t PltEntryAddr, int32_t Index,
517 unsigned RelOff) const {
518 const uint8_t Insn[] = {
519 0x50, // 0: pushl %eax
520 0xa1, 0, 0, 0, 0, // 1: mov foo_in_GOT, %eax
521 0xe8, 0, 0, 0, 0, // 6: call plt+0x20
522 0xe9, 0, 0, 0, 0, // b: jmp plt+0x11
523 0x68, 0, 0, 0, 0, // 10: pushl $reloc_offset
524 0xe9, 0, 0, 0, 0, // 15: jmp plt+0
526 memcpy(Buf, Insn, sizeof(Insn));
528 write32le(Buf + 2, GotPltEntryAddr);
529 write32le(Buf + 7, -Index * PltEntrySize - PltHeaderSize - 11 + 32);
530 write32le(Buf + 12, -Index * PltEntrySize - PltHeaderSize - 16 + 17);
531 write32le(Buf + 17, RelOff);
532 write32le(Buf + 22, -Index * PltEntrySize - PltHeaderSize - 26);
535 TargetInfo *elf::getX86TargetInfo() {
536 if (Config->ZRetpolineplt) {
538 static RetpolinePic T;
541 static RetpolineNoPic T;