1 //===-- X86MachObjectWriter.cpp - X86 Mach-O Writer -----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "MCTargetDesc/X86MCTargetDesc.h"
11 #include "MCTargetDesc/X86FixupKinds.h"
12 #include "llvm/ADT/Twine.h"
13 #include "llvm/MC/MCAsmInfo.h"
14 #include "llvm/MC/MCAsmLayout.h"
15 #include "llvm/MC/MCAssembler.h"
16 #include "llvm/MC/MCContext.h"
17 #include "llvm/MC/MCMachObjectWriter.h"
18 #include "llvm/MC/MCSectionMachO.h"
19 #include "llvm/MC/MCValue.h"
20 #include "llvm/Support/ErrorHandling.h"
21 #include "llvm/Support/Format.h"
22 #include "llvm/Support/MachO.h"
27 class X86MachObjectWriter : public MCMachObjectTargetWriter {
28 bool recordScatteredRelocation(MachObjectWriter *Writer,
29 const MCAssembler &Asm,
30 const MCAsmLayout &Layout,
31 const MCFragment *Fragment,
35 uint64_t &FixedValue);
36 void recordTLVPRelocation(MachObjectWriter *Writer,
37 const MCAssembler &Asm,
38 const MCAsmLayout &Layout,
39 const MCFragment *Fragment,
42 uint64_t &FixedValue);
44 void RecordX86Relocation(MachObjectWriter *Writer,
45 const MCAssembler &Asm,
46 const MCAsmLayout &Layout,
47 const MCFragment *Fragment,
50 uint64_t &FixedValue);
51 void RecordX86_64Relocation(MachObjectWriter *Writer, MCAssembler &Asm,
52 const MCAsmLayout &Layout,
53 const MCFragment *Fragment, const MCFixup &Fixup,
54 MCValue Target, uint64_t &FixedValue);
57 X86MachObjectWriter(bool Is64Bit, uint32_t CPUType, uint32_t CPUSubtype)
58 : MCMachObjectTargetWriter(Is64Bit, CPUType, CPUSubtype) {}
60 void recordRelocation(MachObjectWriter *Writer, MCAssembler &Asm,
61 const MCAsmLayout &Layout, const MCFragment *Fragment,
62 const MCFixup &Fixup, MCValue Target,
63 uint64_t &FixedValue) override {
64 if (Writer->is64Bit())
65 RecordX86_64Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
68 RecordX86Relocation(Writer, Asm, Layout, Fragment, Fixup, Target,
74 static bool isFixupKindRIPRel(unsigned Kind) {
75 return Kind == X86::reloc_riprel_4byte ||
76 Kind == X86::reloc_riprel_4byte_movq_load ||
77 Kind == X86::reloc_riprel_4byte_relax ||
78 Kind == X86::reloc_riprel_4byte_relax_rex;
81 static unsigned getFixupKindLog2Size(unsigned Kind) {
84 llvm_unreachable("invalid fixup kind!");
86 case FK_Data_1: return 0;
88 case FK_Data_2: return 1;
90 // FIXME: Remove these!!!
91 case X86::reloc_riprel_4byte:
92 case X86::reloc_riprel_4byte_relax:
93 case X86::reloc_riprel_4byte_relax_rex:
94 case X86::reloc_riprel_4byte_movq_load:
95 case X86::reloc_signed_4byte:
96 case X86::reloc_signed_4byte_relax:
97 case FK_Data_4: return 2;
98 case FK_Data_8: return 3;
102 void X86MachObjectWriter::RecordX86_64Relocation(
103 MachObjectWriter *Writer, MCAssembler &Asm, const MCAsmLayout &Layout,
104 const MCFragment *Fragment, const MCFixup &Fixup, MCValue Target,
105 uint64_t &FixedValue) {
106 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
107 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
108 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
111 uint32_t FixupOffset =
112 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
113 uint32_t FixupAddress =
114 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
117 unsigned IsExtern = 0;
119 const MCSymbol *RelSymbol = nullptr;
121 Value = Target.getConstant();
124 // Compensate for the relocation offset, Darwin x86_64 relocations only have
125 // the addend and appear to have attempted to define it to be the actual
126 // expression addend without the PCrel bias. However, instructions with data
127 // following the relocation are not accommodated for (see comment below
128 // regarding SIGNED{1,2,4}), so it isn't exactly that either.
129 Value += 1LL << Log2Size;
132 if (Target.isAbsolute()) { // constant
133 // SymbolNum of 0 indicates the absolute section.
134 Type = MachO::X86_64_RELOC_UNSIGNED;
136 // FIXME: I believe this is broken, I don't think the linker can understand
137 // it. I think it would require a local relocation, but I'm not sure if that
138 // would work either. The official way to get an absolute PCrel relocation
139 // is to use an absolute symbol (which we don't support yet).
142 Type = MachO::X86_64_RELOC_BRANCH;
144 } else if (Target.getSymB()) { // A - B + constant
145 const MCSymbol *A = &Target.getSymA()->getSymbol();
146 if (A->isTemporary())
147 A = &Writer->findAliasedSymbol(*A);
148 const MCSymbol *A_Base = Asm.getAtom(*A);
150 const MCSymbol *B = &Target.getSymB()->getSymbol();
151 if (B->isTemporary())
152 B = &Writer->findAliasedSymbol(*B);
153 const MCSymbol *B_Base = Asm.getAtom(*B);
155 // Neither symbol can be modified.
156 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
157 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None) {
158 Asm.getContext().reportError(Fixup.getLoc(),
159 "unsupported relocation of modified symbol");
163 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
164 // implement most of these correctly.
166 Asm.getContext().reportError(
167 Fixup.getLoc(), "unsupported pc-relative relocation of difference");
171 // The support for the situation where one or both of the symbols would
172 // require a local relocation is handled just like if the symbols were
173 // external. This is certainly used in the case of debug sections where the
174 // section has only temporary symbols and thus the symbols don't have base
175 // symbols. This is encoded using the section ordinal and non-extern
176 // relocation entries.
178 // Darwin 'as' doesn't emit correct relocations for this (it ends up with a
179 // single SIGNED relocation); reject it for now. Except the case where both
180 // symbols don't have a base, equal but both NULL.
181 if (A_Base == B_Base && A_Base) {
182 Asm.getContext().reportError(
183 Fixup.getLoc(), "unsupported relocation with identical base");
187 // A subtraction expression where either symbol is undefined is a
188 // non-relocatable expression.
189 if (A->isUndefined() || B->isUndefined()) {
190 StringRef Name = A->isUndefined() ? A->getName() : B->getName();
191 Asm.getContext().reportError(Fixup.getLoc(),
192 "unsupported relocation with subtraction expression, symbol '" +
193 Name + "' can not be undefined in a subtraction expression");
197 Value += Writer->getSymbolAddress(*A, Layout) -
198 (!A_Base ? 0 : Writer->getSymbolAddress(*A_Base, Layout));
199 Value -= Writer->getSymbolAddress(*B, Layout) -
200 (!B_Base ? 0 : Writer->getSymbolAddress(*B_Base, Layout));
203 Index = A->getFragment()->getParent()->getOrdinal() + 1;
204 Type = MachO::X86_64_RELOC_UNSIGNED;
206 MachO::any_relocation_info MRE;
207 MRE.r_word0 = FixupOffset;
209 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
210 Writer->addRelocation(A_Base, Fragment->getParent(), MRE);
215 Index = B->getFragment()->getParent()->getOrdinal() + 1;
216 Type = MachO::X86_64_RELOC_SUBTRACTOR;
218 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
219 if (Symbol->isTemporary() && Value) {
220 const MCSection &Sec = Symbol->getSection();
221 if (!Asm.getContext().getAsmInfo()->isSectionAtomizableBySymbols(Sec))
222 Symbol->setUsedInReloc();
224 RelSymbol = Asm.getAtom(*Symbol);
226 // Relocations inside debug sections always use local relocations when
227 // possible. This seems to be done because the debugger doesn't fully
228 // understand x86_64 relocation entries, and expects to find values that
229 // have already been fixed up.
230 if (Symbol->isInSection()) {
231 const MCSectionMachO &Section =
232 static_cast<const MCSectionMachO &>(*Fragment->getParent());
233 if (Section.hasAttribute(MachO::S_ATTR_DEBUG))
237 // x86_64 almost always uses external relocations, except when there is no
238 // symbol to use as a base address (a local symbol with no preceding
239 // non-local symbol).
241 // Add the local offset, if needed.
242 if (RelSymbol != Symbol)
243 Value += Layout.getSymbolOffset(*Symbol) -
244 Layout.getSymbolOffset(*RelSymbol);
245 } else if (Symbol->isInSection() && !Symbol->isVariable()) {
246 // The index is the section ordinal (1-based).
247 Index = Symbol->getFragment()->getParent()->getOrdinal() + 1;
248 Value += Writer->getSymbolAddress(*Symbol, Layout);
251 Value -= FixupAddress + (1 << Log2Size);
252 } else if (Symbol->isVariable()) {
253 const MCExpr *Value = Symbol->getVariableValue();
255 bool isAbs = Value->evaluateAsAbsolute(Res, Layout,
256 Writer->getSectionAddressMap());
261 Asm.getContext().reportError(Fixup.getLoc(),
262 "unsupported relocation of variable '" +
263 Symbol->getName() + "'");
267 Asm.getContext().reportError(
268 Fixup.getLoc(), "unsupported relocation of undefined symbol '" +
269 Symbol->getName() + "'");
273 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
276 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
277 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
278 // rewrite the movq to an leaq at link time if the symbol ends up in
279 // the same linkage unit.
280 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
281 Type = MachO::X86_64_RELOC_GOT_LOAD;
283 Type = MachO::X86_64_RELOC_GOT;
284 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
285 Type = MachO::X86_64_RELOC_TLV;
286 } else if (Modifier != MCSymbolRefExpr::VK_None) {
287 Asm.getContext().reportError(
288 Fixup.getLoc(), "unsupported symbol modifier in relocation");
291 Type = MachO::X86_64_RELOC_SIGNED;
293 // The Darwin x86_64 relocation format has a problem where it cannot
294 // encode an address (L<foo> + <constant>) which is outside the atom
295 // containing L<foo>. Generally, this shouldn't occur but it does
296 // happen when we have a RIPrel instruction with data following the
297 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
298 // adjustment Darwin x86_64 uses, the offset is still negative and the
299 // linker has no way to recognize this.
301 // To work around this, Darwin uses several special relocation types
302 // to indicate the offsets. However, the specification or
303 // implementation of these seems to also be incomplete; they should
304 // adjust the addend as well based on the actual encoded instruction
305 // (the additional bias), but instead appear to just look at the final
307 switch (-(Target.getConstant() + (1LL << Log2Size))) {
308 case 1: Type = MachO::X86_64_RELOC_SIGNED_1; break;
309 case 2: Type = MachO::X86_64_RELOC_SIGNED_2; break;
310 case 4: Type = MachO::X86_64_RELOC_SIGNED_4; break;
314 if (Modifier != MCSymbolRefExpr::VK_None) {
315 Asm.getContext().reportError(
317 "unsupported symbol modifier in branch relocation");
321 Type = MachO::X86_64_RELOC_BRANCH;
324 if (Modifier == MCSymbolRefExpr::VK_GOT) {
325 Type = MachO::X86_64_RELOC_GOT;
326 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
327 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which
328 // case all we do is set the PCrel bit in the relocation entry; this is
329 // used with exception handling, for example. The source is required to
330 // include any necessary offset directly.
331 Type = MachO::X86_64_RELOC_GOT;
333 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
334 Asm.getContext().reportError(
335 Fixup.getLoc(), "TLVP symbol modifier should have been rip-rel");
337 } else if (Modifier != MCSymbolRefExpr::VK_None) {
338 Asm.getContext().reportError(
339 Fixup.getLoc(), "unsupported symbol modifier in relocation");
342 Type = MachO::X86_64_RELOC_UNSIGNED;
343 unsigned Kind = Fixup.getKind();
344 if (Kind == X86::reloc_signed_4byte) {
345 Asm.getContext().reportError(
347 "32-bit absolute addressing is not supported in 64-bit mode");
354 // x86_64 always writes custom values into the fixups.
357 // struct relocation_info (8 bytes)
358 MachO::any_relocation_info MRE;
359 MRE.r_word0 = FixupOffset;
360 MRE.r_word1 = (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) |
361 (IsExtern << 27) | (Type << 28);
362 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
365 bool X86MachObjectWriter::recordScatteredRelocation(MachObjectWriter *Writer,
366 const MCAssembler &Asm,
367 const MCAsmLayout &Layout,
368 const MCFragment *Fragment,
369 const MCFixup &Fixup,
372 uint64_t &FixedValue) {
373 uint64_t OriginalFixedValue = FixedValue;
374 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
375 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
376 unsigned Type = MachO::GENERIC_RELOC_VANILLA;
379 const MCSymbol *A = &Target.getSymA()->getSymbol();
381 if (!A->getFragment()) {
382 Asm.getContext().reportError(
384 "symbol '" + A->getName() +
385 "' can not be undefined in a subtraction expression");
389 uint32_t Value = Writer->getSymbolAddress(*A, Layout);
390 uint64_t SecAddr = Writer->getSectionAddress(A->getFragment()->getParent());
391 FixedValue += SecAddr;
394 if (const MCSymbolRefExpr *B = Target.getSymB()) {
395 const MCSymbol *SB = &B->getSymbol();
397 if (!SB->getFragment()) {
398 Asm.getContext().reportError(
400 "symbol '" + B->getSymbol().getName() +
401 "' can not be undefined in a subtraction expression");
405 // Select the appropriate difference relocation type.
407 // Note that there is no longer any semantic difference between these two
408 // relocation types from the linkers point of view, this is done solely for
409 // pedantic compatibility with 'as'.
410 Type = A->isExternal() ? (unsigned)MachO::GENERIC_RELOC_SECTDIFF
411 : (unsigned)MachO::GENERIC_RELOC_LOCAL_SECTDIFF;
412 Value2 = Writer->getSymbolAddress(B->getSymbol(), Layout);
413 FixedValue -= Writer->getSectionAddress(SB->getFragment()->getParent());
416 // Relocations are written out in reverse order, so the PAIR comes first.
417 if (Type == MachO::GENERIC_RELOC_SECTDIFF ||
418 Type == MachO::GENERIC_RELOC_LOCAL_SECTDIFF) {
419 // If the offset is too large to fit in a scattered relocation,
420 // we're hosed. It's an unfortunate limitation of the MachO format.
421 if (FixupOffset > 0xffffff) {
423 format("0x%x", FixupOffset).print(Buffer, sizeof(Buffer));
424 Asm.getContext().reportError(Fixup.getLoc(),
425 Twine("Section too large, can't encode "
426 "r_address (") + Buffer +
427 ") into 24 bits of scattered "
428 "relocation entry.");
432 MachO::any_relocation_info MRE;
433 MRE.r_word0 = ((0 << 0) | // r_address
434 (MachO::GENERIC_RELOC_PAIR << 24) | // r_type
438 MRE.r_word1 = Value2;
439 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
441 // If the offset is more than 24-bits, it won't fit in a scattered
442 // relocation offset field, so we fall back to using a non-scattered
443 // relocation. This is a bit risky, as if the offset reaches out of
444 // the block and the linker is doing scattered loading on this
445 // symbol, things can go badly.
447 // Required for 'as' compatibility.
448 if (FixupOffset > 0xffffff) {
449 FixedValue = OriginalFixedValue;
454 MachO::any_relocation_info MRE;
455 MRE.r_word0 = ((FixupOffset << 0) |
461 Writer->addRelocation(nullptr, Fragment->getParent(), MRE);
465 void X86MachObjectWriter::recordTLVPRelocation(MachObjectWriter *Writer,
466 const MCAssembler &Asm,
467 const MCAsmLayout &Layout,
468 const MCFragment *Fragment,
469 const MCFixup &Fixup,
471 uint64_t &FixedValue) {
472 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
474 "Should only be called with a 32-bit TLVP relocation!");
476 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
477 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
478 unsigned IsPCRel = 0;
480 // We're only going to have a second symbol in pic mode and it'll be a
481 // subtraction from the picbase. For 32-bit pic the addend is the difference
482 // between the picbase and the next address. For 32-bit static the addend is
484 if (Target.getSymB()) {
485 // If this is a subtraction then we're pcrel.
486 uint32_t FixupAddress =
487 Writer->getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
491 Writer->getSymbolAddress(Target.getSymB()->getSymbol(), Layout) +
492 Target.getConstant();
493 FixedValue += 1ULL << Log2Size;
498 // struct relocation_info (8 bytes)
499 MachO::any_relocation_info MRE;
502 (IsPCRel << 24) | (Log2Size << 25) | (MachO::GENERIC_RELOC_TLV << 28);
503 Writer->addRelocation(&Target.getSymA()->getSymbol(), Fragment->getParent(),
507 void X86MachObjectWriter::RecordX86Relocation(MachObjectWriter *Writer,
508 const MCAssembler &Asm,
509 const MCAsmLayout &Layout,
510 const MCFragment *Fragment,
511 const MCFixup &Fixup,
513 uint64_t &FixedValue) {
514 unsigned IsPCRel = Writer->isFixupKindPCRel(Asm, Fixup.getKind());
515 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
517 // If this is a 32-bit TLVP reloc it's handled a bit differently.
518 if (Target.getSymA() &&
519 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
520 recordTLVPRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
525 // If this is a difference or a defined symbol plus an offset, then we need a
526 // scattered relocation entry. Differences always require scattered
528 if (Target.getSymB()) {
529 recordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup,
530 Target, Log2Size, FixedValue);
534 // Get the symbol data, if any.
535 const MCSymbol *A = nullptr;
536 if (Target.getSymA())
537 A = &Target.getSymA()->getSymbol();
539 // If this is an internal relocation with an offset, it also needs a scattered
541 uint32_t Offset = Target.getConstant();
543 Offset += 1 << Log2Size;
544 // Try to record the scattered relocation if needed. Fall back to non
545 // scattered if necessary (see comments in recordScatteredRelocation()
547 if (Offset && A && !Writer->doesSymbolRequireExternRelocation(*A) &&
548 recordScatteredRelocation(Writer, Asm, Layout, Fragment, Fixup, Target,
549 Log2Size, FixedValue))
553 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
556 const MCSymbol *RelSymbol = nullptr;
558 if (Target.isAbsolute()) { // constant
559 // SymbolNum of 0 indicates the absolute section.
561 // FIXME: Currently, these are never generated (see code below). I cannot
562 // find a case where they are actually emitted.
563 Type = MachO::GENERIC_RELOC_VANILLA;
565 // Resolve constant variables.
566 if (A->isVariable()) {
568 if (A->getVariableValue()->evaluateAsAbsolute(
569 Res, Layout, Writer->getSectionAddressMap())) {
575 // Check whether we need an external or internal relocation.
576 if (Writer->doesSymbolRequireExternRelocation(*A)) {
578 // For external relocations, make sure to offset the fixup value to
579 // compensate for the addend of the symbol address, if it was
580 // undefined. This occurs with weak definitions, for example.
581 if (!A->isUndefined())
582 FixedValue -= Layout.getSymbolOffset(*A);
584 // The index is the section ordinal (1-based).
585 const MCSection &Sec = A->getSection();
586 Index = Sec.getOrdinal() + 1;
587 FixedValue += Writer->getSectionAddress(&Sec);
590 FixedValue -= Writer->getSectionAddress(Fragment->getParent());
592 Type = MachO::GENERIC_RELOC_VANILLA;
595 // struct relocation_info (8 bytes)
596 MachO::any_relocation_info MRE;
597 MRE.r_word0 = FixupOffset;
599 (Index << 0) | (IsPCRel << 24) | (Log2Size << 25) | (Type << 28);
600 Writer->addRelocation(RelSymbol, Fragment->getParent(), MRE);
603 MCObjectWriter *llvm::createX86MachObjectWriter(raw_pwrite_stream &OS,
604 bool Is64Bit, uint32_t CPUType,
605 uint32_t CPUSubtype) {
606 return createMachObjectWriter(new X86MachObjectWriter(Is64Bit,
609 OS, /*IsLittleEndian=*/true);