1 //===- lib/MC/MachObjectWriter.cpp - Mach-O File 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 "llvm/ADT/StringMap.h"
11 #include "llvm/ADT/Twine.h"
12 #include "llvm/MC/MCAssembler.h"
13 #include "llvm/MC/MCAsmLayout.h"
14 #include "llvm/MC/MCExpr.h"
15 #include "llvm/MC/MCObjectWriter.h"
16 #include "llvm/MC/MCSectionMachO.h"
17 #include "llvm/MC/MCSymbol.h"
18 #include "llvm/MC/MCMachOSymbolFlags.h"
19 #include "llvm/MC/MCValue.h"
20 #include "llvm/Object/MachOFormat.h"
21 #include "llvm/Support/ErrorHandling.h"
22 #include "llvm/Target/TargetAsmBackend.h"
25 #include "../Target/X86/X86FixupKinds.h"
29 using namespace llvm::object;
31 // FIXME: this has been copied from (or to) X86AsmBackend.cpp
32 static unsigned getFixupKindLog2Size(unsigned Kind) {
34 // FIXME: Until ARM has it's own relocation stuff spun off, it comes
35 // through here and we don't want it to puke all over. Any reasonable
36 // values will only come when ARM relocation support gets added, at which
37 // point this will be X86 only again and the llvm_unreachable can be
39 default: return 0;// llvm_unreachable("invalid fixup kind!");
41 case FK_Data_1: return 0;
43 case FK_Data_2: return 1;
45 case X86::reloc_riprel_4byte:
46 case X86::reloc_riprel_4byte_movq_load:
47 case X86::reloc_signed_4byte:
48 case FK_Data_4: return 2;
49 case FK_Data_8: return 3;
53 static bool isFixupKindPCRel(unsigned Kind) {
60 case X86::reloc_riprel_4byte:
61 case X86::reloc_riprel_4byte_movq_load:
66 static bool isFixupKindRIPRel(unsigned Kind) {
67 return Kind == X86::reloc_riprel_4byte ||
68 Kind == X86::reloc_riprel_4byte_movq_load;
71 static bool doesSymbolRequireExternRelocation(MCSymbolData *SD) {
72 // Undefined symbols are always extern.
73 if (SD->Symbol->isUndefined())
76 // References to weak definitions require external relocation entries; the
77 // definition may not always be the one in the same object file.
78 if (SD->getFlags() & SF_WeakDefinition)
81 // Otherwise, we can use an internal relocation.
85 static bool isScatteredFixupFullyResolved(const MCAssembler &Asm,
87 const MCSymbolData *BaseSymbol) {
88 // The effective fixup address is
89 // addr(atom(A)) + offset(A)
90 // - addr(atom(B)) - offset(B)
91 // - addr(BaseSymbol) + <fixup offset from base symbol>
92 // and the offsets are not relocatable, so the fixup is fully resolved when
93 // addr(atom(A)) - addr(atom(B)) - addr(BaseSymbol) == 0.
95 // Note that "false" is almost always conservatively correct (it means we emit
96 // a relocation which is unnecessary), except when it would force us to emit a
97 // relocation which the target cannot encode.
99 const MCSymbolData *A_Base = 0, *B_Base = 0;
100 if (const MCSymbolRefExpr *A = Target.getSymA()) {
101 // Modified symbol references cannot be resolved.
102 if (A->getKind() != MCSymbolRefExpr::VK_None)
105 A_Base = Asm.getAtom(&Asm.getSymbolData(A->getSymbol()));
110 if (const MCSymbolRefExpr *B = Target.getSymB()) {
111 // Modified symbol references cannot be resolved.
112 if (B->getKind() != MCSymbolRefExpr::VK_None)
115 B_Base = Asm.getAtom(&Asm.getSymbolData(B->getSymbol()));
120 // If there is no base, A and B have to be the same atom for this fixup to be
123 return A_Base == B_Base;
125 // Otherwise, B must be missing and A must be the base.
126 return !B_Base && BaseSymbol == A_Base;
129 static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm,
130 const MCValue Target,
131 const MCSection *BaseSection) {
132 // The effective fixup address is
133 // addr(atom(A)) + offset(A)
134 // - addr(atom(B)) - offset(B)
135 // - addr(<base symbol>) + <fixup offset from base symbol>
136 // and the offsets are not relocatable, so the fixup is fully resolved when
137 // addr(atom(A)) - addr(atom(B)) - addr(<base symbol>)) == 0.
139 // The simple (Darwin, except on x86_64) way of dealing with this was to
140 // assume that any reference to a temporary symbol *must* be a temporary
141 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
142 // relocation to a temporary symbol (in the same section) is fully
143 // resolved. This also works in conjunction with absolutized .set, which
144 // requires the compiler to use .set to absolutize the differences between
145 // symbols which the compiler knows to be assembly time constants, so we don't
146 // need to worry about considering symbol differences fully resolved.
148 // Non-relative fixups are only resolved if constant.
150 return Target.isAbsolute();
152 // Otherwise, relative fixups are only resolved if not a difference and the
153 // target is a temporary in the same section.
154 if (Target.isAbsolute() || Target.getSymB())
157 const MCSymbol *A = &Target.getSymA()->getSymbol();
158 if (!A->isTemporary() || !A->isInSection() ||
159 &A->getSection() != BaseSection)
167 class MachObjectWriter : public MCObjectWriter {
168 /// MachSymbolData - Helper struct for containing some precomputed information
170 struct MachSymbolData {
171 MCSymbolData *SymbolData;
172 uint64_t StringIndex;
173 uint8_t SectionIndex;
175 // Support lexicographic sorting.
176 bool operator<(const MachSymbolData &RHS) const {
177 return SymbolData->getSymbol().getName() <
178 RHS.SymbolData->getSymbol().getName();
182 /// @name Relocation Data
185 llvm::DenseMap<const MCSectionData*,
186 std::vector<macho::RelocationEntry> > Relocations;
187 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
190 /// @name Symbol Table Data
193 SmallString<256> StringTable;
194 std::vector<MachSymbolData> LocalSymbolData;
195 std::vector<MachSymbolData> ExternalSymbolData;
196 std::vector<MachSymbolData> UndefinedSymbolData;
200 unsigned Is64Bit : 1;
206 MachObjectWriter(raw_ostream &_OS,
207 bool _Is64Bit, uint32_t _CPUType, uint32_t _CPUSubtype,
208 bool _IsLittleEndian)
209 : MCObjectWriter(_OS, _IsLittleEndian),
210 Is64Bit(_Is64Bit), CPUType(_CPUType), CPUSubtype(_CPUSubtype) {
213 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
214 bool SubsectionsViaSymbols) {
217 if (SubsectionsViaSymbols)
218 Flags |= macho::HF_SubsectionsViaSymbols;
220 // struct mach_header (28 bytes) or
221 // struct mach_header_64 (32 bytes)
223 uint64_t Start = OS.tell();
226 Write32(Is64Bit ? macho::HM_Object64 : macho::HM_Object32);
231 Write32(macho::HFT_Object);
232 Write32(NumLoadCommands);
233 Write32(LoadCommandsSize);
236 Write32(0); // reserved
238 assert(OS.tell() - Start == Is64Bit ?
239 macho::Header64Size : macho::Header32Size);
242 /// WriteSegmentLoadCommand - Write a segment load command.
244 /// \arg NumSections - The number of sections in this segment.
245 /// \arg SectionDataSize - The total size of the sections.
246 void WriteSegmentLoadCommand(unsigned NumSections,
248 uint64_t SectionDataStartOffset,
249 uint64_t SectionDataSize) {
250 // struct segment_command (56 bytes) or
251 // struct segment_command_64 (72 bytes)
253 uint64_t Start = OS.tell();
256 unsigned SegmentLoadCommandSize = Is64Bit ? macho::SegmentLoadCommand64Size:
257 macho::SegmentLoadCommand32Size;
258 Write32(Is64Bit ? macho::LCT_Segment64 : macho::LCT_Segment);
259 Write32(SegmentLoadCommandSize +
260 NumSections * (Is64Bit ? macho::Section64Size :
261 macho::Section32Size));
265 Write64(0); // vmaddr
266 Write64(VMSize); // vmsize
267 Write64(SectionDataStartOffset); // file offset
268 Write64(SectionDataSize); // file size
270 Write32(0); // vmaddr
271 Write32(VMSize); // vmsize
272 Write32(SectionDataStartOffset); // file offset
273 Write32(SectionDataSize); // file size
275 Write32(0x7); // maxprot
276 Write32(0x7); // initprot
277 Write32(NumSections);
280 assert(OS.tell() - Start == SegmentLoadCommandSize);
283 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
284 const MCSectionData &SD, uint64_t FileOffset,
285 uint64_t RelocationsStart, unsigned NumRelocations) {
286 uint64_t SectionSize = Layout.getSectionSize(&SD);
288 // The offset is unused for virtual sections.
289 if (SD.getSection().isVirtualSection()) {
290 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
294 // struct section (68 bytes) or
295 // struct section_64 (80 bytes)
297 uint64_t Start = OS.tell();
300 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
301 WriteBytes(Section.getSectionName(), 16);
302 WriteBytes(Section.getSegmentName(), 16);
304 Write64(Layout.getSectionAddress(&SD)); // address
305 Write64(SectionSize); // size
307 Write32(Layout.getSectionAddress(&SD)); // address
308 Write32(SectionSize); // size
312 unsigned Flags = Section.getTypeAndAttributes();
313 if (SD.hasInstructions())
314 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
316 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
317 Write32(Log2_32(SD.getAlignment()));
318 Write32(NumRelocations ? RelocationsStart : 0);
319 Write32(NumRelocations);
321 Write32(IndirectSymBase.lookup(&SD)); // reserved1
322 Write32(Section.getStubSize()); // reserved2
324 Write32(0); // reserved3
326 assert(OS.tell() - Start == Is64Bit ? macho::Section64Size :
327 macho::Section32Size);
330 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
331 uint32_t StringTableOffset,
332 uint32_t StringTableSize) {
333 // struct symtab_command (24 bytes)
335 uint64_t Start = OS.tell();
338 Write32(macho::LCT_Symtab);
339 Write32(macho::SymtabLoadCommandSize);
340 Write32(SymbolOffset);
342 Write32(StringTableOffset);
343 Write32(StringTableSize);
345 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
348 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
349 uint32_t NumLocalSymbols,
350 uint32_t FirstExternalSymbol,
351 uint32_t NumExternalSymbols,
352 uint32_t FirstUndefinedSymbol,
353 uint32_t NumUndefinedSymbols,
354 uint32_t IndirectSymbolOffset,
355 uint32_t NumIndirectSymbols) {
356 // struct dysymtab_command (80 bytes)
358 uint64_t Start = OS.tell();
361 Write32(macho::LCT_Dysymtab);
362 Write32(macho::DysymtabLoadCommandSize);
363 Write32(FirstLocalSymbol);
364 Write32(NumLocalSymbols);
365 Write32(FirstExternalSymbol);
366 Write32(NumExternalSymbols);
367 Write32(FirstUndefinedSymbol);
368 Write32(NumUndefinedSymbols);
369 Write32(0); // tocoff
371 Write32(0); // modtaboff
372 Write32(0); // nmodtab
373 Write32(0); // extrefsymoff
374 Write32(0); // nextrefsyms
375 Write32(IndirectSymbolOffset);
376 Write32(NumIndirectSymbols);
377 Write32(0); // extreloff
378 Write32(0); // nextrel
379 Write32(0); // locreloff
380 Write32(0); // nlocrel
382 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
385 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
386 MCSymbolData &Data = *MSD.SymbolData;
387 const MCSymbol &Symbol = Data.getSymbol();
389 uint16_t Flags = Data.getFlags();
390 uint32_t Address = 0;
392 // Set the N_TYPE bits. See <mach-o/nlist.h>.
394 // FIXME: Are the prebound or indirect fields possible here?
395 if (Symbol.isUndefined())
396 Type = macho::STT_Undefined;
397 else if (Symbol.isAbsolute())
398 Type = macho::STT_Absolute;
400 Type = macho::STT_Section;
402 // FIXME: Set STAB bits.
404 if (Data.isPrivateExtern())
405 Type |= macho::STF_PrivateExtern;
408 if (Data.isExternal() || Symbol.isUndefined())
409 Type |= macho::STF_External;
411 // Compute the symbol address.
412 if (Symbol.isDefined()) {
413 if (Symbol.isAbsolute()) {
414 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
416 Address = Layout.getSymbolAddress(&Data);
418 } else if (Data.isCommon()) {
419 // Common symbols are encoded with the size in the address
420 // field, and their alignment in the flags.
421 Address = Data.getCommonSize();
423 // Common alignment is packed into the 'desc' bits.
424 if (unsigned Align = Data.getCommonAlignment()) {
425 unsigned Log2Size = Log2_32(Align);
426 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
428 report_fatal_error("invalid 'common' alignment '" +
430 // FIXME: Keep this mask with the SymbolFlags enumeration.
431 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
435 // struct nlist (12 bytes)
437 Write32(MSD.StringIndex);
439 Write8(MSD.SectionIndex);
441 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
450 // FIXME: We really need to improve the relocation validation. Basically, we
451 // want to implement a separate computation which evaluates the relocation
452 // entry as the linker would, and verifies that the resultant fixup value is
453 // exactly what the encoder wanted. This will catch several classes of
456 // - Relocation entry bugs, the two algorithms are unlikely to have the same
459 // - Relaxation issues, where we forget to relax something.
461 // - Input errors, where something cannot be correctly encoded. 'as' allows
462 // these through in many cases.
464 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
465 const MCFragment *Fragment,
466 const MCFixup &Fixup, MCValue Target,
467 uint64_t &FixedValue) {
468 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
469 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
470 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
473 uint32_t FixupOffset =
474 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
475 uint32_t FixupAddress =
476 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
479 unsigned IsExtern = 0;
482 Value = Target.getConstant();
485 // Compensate for the relocation offset, Darwin x86_64 relocations only
486 // have the addend and appear to have attempted to define it to be the
487 // actual expression addend without the PCrel bias. However, instructions
488 // with data following the relocation are not accomodated for (see comment
489 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
490 Value += 1LL << Log2Size;
493 if (Target.isAbsolute()) { // constant
494 // SymbolNum of 0 indicates the absolute section.
495 Type = macho::RIT_X86_64_Unsigned;
498 // FIXME: I believe this is broken, I don't think the linker can
499 // understand it. I think it would require a local relocation, but I'm not
500 // sure if that would work either. The official way to get an absolute
501 // PCrel relocation is to use an absolute symbol (which we don't support
505 Type = macho::RIT_X86_64_Branch;
507 } else if (Target.getSymB()) { // A - B + constant
508 const MCSymbol *A = &Target.getSymA()->getSymbol();
509 MCSymbolData &A_SD = Asm.getSymbolData(*A);
510 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
512 const MCSymbol *B = &Target.getSymB()->getSymbol();
513 MCSymbolData &B_SD = Asm.getSymbolData(*B);
514 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
516 // Neither symbol can be modified.
517 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
518 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
519 report_fatal_error("unsupported relocation of modified symbol");
521 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
522 // implement most of these correctly.
524 report_fatal_error("unsupported pc-relative relocation of difference");
526 // The support for the situation where one or both of the symbols would
527 // require a local relocation is handled just like if the symbols were
528 // external. This is certainly used in the case of debug sections where
529 // the section has only temporary symbols and thus the symbols don't have
530 // base symbols. This is encoded using the section ordinal and
531 // non-extern relocation entries.
533 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
534 // a single SIGNED relocation); reject it for now. Except the case where
535 // both symbols don't have a base, equal but both NULL.
536 if (A_Base == B_Base && A_Base)
537 report_fatal_error("unsupported relocation with identical base");
539 assert((A_Base == NULL) == (B_Base == NULL));
540 assert(A_SD.getFragment()->getParent() ==
541 B_SD.getFragment()->getParent());
543 Value += Layout.getSymbolOffset(&A_SD) -
544 (A_Base == NULL ? 0 : Layout.getSymbolOffset(A_Base));
545 Value -= Layout.getSymbolOffset(&B_SD) -
546 (B_Base == NULL ? 0 : Layout.getSymbolOffset(B_Base));
549 Index = A_Base->getIndex();
553 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
556 Type = macho::RIT_X86_64_Unsigned;
558 macho::RelocationEntry MRE;
559 MRE.Word0 = FixupOffset;
560 MRE.Word1 = ((Index << 0) |
565 Relocations[Fragment->getParent()].push_back(MRE);
568 Index = B_Base->getIndex();
572 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
575 Type = macho::RIT_X86_64_Subtractor;
577 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
578 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
579 const MCSymbolData *Base = Asm.getAtom(&SD);
581 // Relocations inside debug sections always use local relocations when
582 // possible. This seems to be done because the debugger doesn't fully
583 // understand x86_64 relocation entries, and expects to find values that
584 // have already been fixed up.
585 if (Symbol->isInSection()) {
586 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
587 Fragment->getParent()->getSection());
588 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
592 // x86_64 almost always uses external relocations, except when there is no
593 // symbol to use as a base address (a local symbol with no preceeding
594 // non-local symbol).
596 Index = Base->getIndex();
599 // Add the local offset, if needed.
601 Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base);
602 } else if (Symbol->isInSection()) {
603 // The index is the section ordinal (1-based).
604 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
606 Value += Layout.getSymbolAddress(&SD);
609 Value -= FixupAddress + (1 << Log2Size);
611 report_fatal_error("unsupported relocation of undefined symbol '" +
612 Symbol->getName() + "'");
615 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
618 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
619 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
620 // rewrite the movq to an leaq at link time if the symbol ends up in
621 // the same linkage unit.
622 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
623 Type = macho::RIT_X86_64_GOTLoad;
625 Type = macho::RIT_X86_64_GOT;
626 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
627 Type = macho::RIT_X86_64_TLV;
628 } else if (Modifier != MCSymbolRefExpr::VK_None) {
629 report_fatal_error("unsupported symbol modifier in relocation");
631 Type = macho::RIT_X86_64_Signed;
633 // The Darwin x86_64 relocation format has a problem where it cannot
634 // encode an address (L<foo> + <constant>) which is outside the atom
635 // containing L<foo>. Generally, this shouldn't occur but it does
636 // happen when we have a RIPrel instruction with data following the
637 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
638 // adjustment Darwin x86_64 uses, the offset is still negative and
639 // the linker has no way to recognize this.
641 // To work around this, Darwin uses several special relocation types
642 // to indicate the offsets. However, the specification or
643 // implementation of these seems to also be incomplete; they should
644 // adjust the addend as well based on the actual encoded instruction
645 // (the additional bias), but instead appear to just look at the
647 switch (-(Target.getConstant() + (1LL << Log2Size))) {
648 case 1: Type = macho::RIT_X86_64_Signed1; break;
649 case 2: Type = macho::RIT_X86_64_Signed2; break;
650 case 4: Type = macho::RIT_X86_64_Signed4; break;
654 if (Modifier != MCSymbolRefExpr::VK_None)
655 report_fatal_error("unsupported symbol modifier in branch "
658 Type = macho::RIT_X86_64_Branch;
661 if (Modifier == MCSymbolRefExpr::VK_GOT) {
662 Type = macho::RIT_X86_64_GOT;
663 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
664 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
665 // which case all we do is set the PCrel bit in the relocation entry;
666 // this is used with exception handling, for example. The source is
667 // required to include any necessary offset directly.
668 Type = macho::RIT_X86_64_GOT;
670 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
671 report_fatal_error("TLVP symbol modifier should have been rip-rel");
672 } else if (Modifier != MCSymbolRefExpr::VK_None)
673 report_fatal_error("unsupported symbol modifier in relocation");
675 Type = macho::RIT_X86_64_Unsigned;
679 // x86_64 always writes custom values into the fixups.
682 // struct relocation_info (8 bytes)
683 macho::RelocationEntry MRE;
684 MRE.Word0 = FixupOffset;
685 MRE.Word1 = ((Index << 0) |
690 Relocations[Fragment->getParent()].push_back(MRE);
693 void RecordScatteredRelocation(const MCAssembler &Asm,
694 const MCAsmLayout &Layout,
695 const MCFragment *Fragment,
696 const MCFixup &Fixup, MCValue Target,
697 uint64_t &FixedValue) {
698 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
699 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
700 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
701 unsigned Type = macho::RIT_Vanilla;
704 const MCSymbol *A = &Target.getSymA()->getSymbol();
705 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
707 if (!A_SD->getFragment())
708 report_fatal_error("symbol '" + A->getName() +
709 "' can not be undefined in a subtraction expression");
711 uint32_t Value = Layout.getSymbolAddress(A_SD);
714 if (const MCSymbolRefExpr *B = Target.getSymB()) {
715 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
717 if (!B_SD->getFragment())
718 report_fatal_error("symbol '" + B->getSymbol().getName() +
719 "' can not be undefined in a subtraction expression");
721 // Select the appropriate difference relocation type.
723 // Note that there is no longer any semantic difference between these two
724 // relocation types from the linkers point of view, this is done solely
725 // for pedantic compatibility with 'as'.
726 Type = A_SD->isExternal() ? macho::RIT_Difference :
727 macho::RIT_LocalDifference;
728 Value2 = Layout.getSymbolAddress(B_SD);
731 // Relocations are written out in reverse order, so the PAIR comes first.
732 if (Type == macho::RIT_Difference || Type == macho::RIT_LocalDifference) {
733 macho::RelocationEntry MRE;
734 MRE.Word0 = ((0 << 0) |
735 (macho::RIT_Pair << 24) |
738 macho::RF_Scattered);
740 Relocations[Fragment->getParent()].push_back(MRE);
743 macho::RelocationEntry MRE;
744 MRE.Word0 = ((FixupOffset << 0) |
748 macho::RF_Scattered);
750 Relocations[Fragment->getParent()].push_back(MRE);
753 void RecordTLVPRelocation(const MCAssembler &Asm,
754 const MCAsmLayout &Layout,
755 const MCFragment *Fragment,
756 const MCFixup &Fixup, MCValue Target,
757 uint64_t &FixedValue) {
758 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
760 "Should only be called with a 32-bit TLVP relocation!");
762 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
763 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
764 unsigned IsPCRel = 0;
766 // Get the symbol data.
767 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
768 unsigned Index = SD_A->getIndex();
770 // We're only going to have a second symbol in pic mode and it'll be a
771 // subtraction from the picbase. For 32-bit pic the addend is the difference
772 // between the picbase and the next address. For 32-bit static the addend
774 if (Target.getSymB()) {
775 // If this is a subtraction then we're pcrel.
776 uint32_t FixupAddress =
777 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
778 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
780 FixedValue = (FixupAddress - Layout.getSymbolAddress(SD_B) +
781 Target.getConstant());
782 FixedValue += 1ULL << Log2Size;
787 // struct relocation_info (8 bytes)
788 macho::RelocationEntry MRE;
790 MRE.Word1 = ((Index << 0) |
793 (1 << 27) | // Extern
794 (macho::RIT_TLV << 28)); // Type
795 Relocations[Fragment->getParent()].push_back(MRE);
798 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
799 const MCFragment *Fragment, const MCFixup &Fixup,
800 MCValue Target, uint64_t &FixedValue) {
802 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
806 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
807 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
809 // If this is a 32-bit TLVP reloc it's handled a bit differently.
810 if (Target.getSymA() &&
811 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
812 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
816 // If this is a difference or a defined symbol plus an offset, then we need
817 // a scattered relocation entry.
818 // Differences always require scattered relocations.
819 if (Target.getSymB())
820 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
823 // Get the symbol data, if any.
824 MCSymbolData *SD = 0;
825 if (Target.getSymA())
826 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
828 // If this is an internal relocation with an offset, it also needs a
829 // scattered relocation entry.
830 uint32_t Offset = Target.getConstant();
832 Offset += 1 << Log2Size;
833 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
834 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
838 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
840 unsigned IsExtern = 0;
843 if (Target.isAbsolute()) { // constant
844 // SymbolNum of 0 indicates the absolute section.
846 // FIXME: Currently, these are never generated (see code below). I cannot
847 // find a case where they are actually emitted.
848 Type = macho::RIT_Vanilla;
850 // Check whether we need an external or internal relocation.
851 if (doesSymbolRequireExternRelocation(SD)) {
853 Index = SD->getIndex();
854 // For external relocations, make sure to offset the fixup value to
855 // compensate for the addend of the symbol address, if it was
856 // undefined. This occurs with weak definitions, for example.
857 if (!SD->Symbol->isUndefined())
858 FixedValue -= Layout.getSymbolAddress(SD);
860 // The index is the section ordinal (1-based).
861 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
864 Type = macho::RIT_Vanilla;
867 // struct relocation_info (8 bytes)
868 macho::RelocationEntry MRE;
869 MRE.Word0 = FixupOffset;
870 MRE.Word1 = ((Index << 0) |
875 Relocations[Fragment->getParent()].push_back(MRE);
878 void BindIndirectSymbols(MCAssembler &Asm) {
879 // This is the point where 'as' creates actual symbols for indirect symbols
880 // (in the following two passes). It would be easier for us to do this
881 // sooner when we see the attribute, but that makes getting the order in the
882 // symbol table much more complicated than it is worth.
884 // FIXME: Revisit this when the dust settles.
886 // Bind non lazy symbol pointers first.
887 unsigned IndirectIndex = 0;
888 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
889 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
890 const MCSectionMachO &Section =
891 cast<MCSectionMachO>(it->SectionData->getSection());
893 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
896 // Initialize the section indirect symbol base, if necessary.
897 if (!IndirectSymBase.count(it->SectionData))
898 IndirectSymBase[it->SectionData] = IndirectIndex;
900 Asm.getOrCreateSymbolData(*it->Symbol);
903 // Then lazy symbol pointers and symbol stubs.
905 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
906 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
907 const MCSectionMachO &Section =
908 cast<MCSectionMachO>(it->SectionData->getSection());
910 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
911 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
914 // Initialize the section indirect symbol base, if necessary.
915 if (!IndirectSymBase.count(it->SectionData))
916 IndirectSymBase[it->SectionData] = IndirectIndex;
918 // Set the symbol type to undefined lazy, but only on construction.
920 // FIXME: Do not hardcode.
922 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
924 Entry.setFlags(Entry.getFlags() | 0x0001);
928 /// ComputeSymbolTable - Compute the symbol table data
930 /// \param StringTable [out] - The string table data.
931 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
933 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
934 std::vector<MachSymbolData> &LocalSymbolData,
935 std::vector<MachSymbolData> &ExternalSymbolData,
936 std::vector<MachSymbolData> &UndefinedSymbolData) {
937 // Build section lookup table.
938 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
940 for (MCAssembler::iterator it = Asm.begin(),
941 ie = Asm.end(); it != ie; ++it, ++Index)
942 SectionIndexMap[&it->getSection()] = Index;
943 assert(Index <= 256 && "Too many sections!");
945 // Index 0 is always the empty string.
946 StringMap<uint64_t> StringIndexMap;
947 StringTable += '\x00';
949 // Build the symbol arrays and the string table, but only for non-local
952 // The particular order that we collect the symbols and create the string
953 // table, then sort the symbols is chosen to match 'as'. Even though it
954 // doesn't matter for correctness, this is important for letting us diff .o
956 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
957 ie = Asm.symbol_end(); it != ie; ++it) {
958 const MCSymbol &Symbol = it->getSymbol();
960 // Ignore non-linker visible symbols.
961 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
964 if (!it->isExternal() && !Symbol.isUndefined())
967 uint64_t &Entry = StringIndexMap[Symbol.getName()];
969 Entry = StringTable.size();
970 StringTable += Symbol.getName();
971 StringTable += '\x00';
976 MSD.StringIndex = Entry;
978 if (Symbol.isUndefined()) {
979 MSD.SectionIndex = 0;
980 UndefinedSymbolData.push_back(MSD);
981 } else if (Symbol.isAbsolute()) {
982 MSD.SectionIndex = 0;
983 ExternalSymbolData.push_back(MSD);
985 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
986 assert(MSD.SectionIndex && "Invalid section index!");
987 ExternalSymbolData.push_back(MSD);
991 // Now add the data for local symbols.
992 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
993 ie = Asm.symbol_end(); it != ie; ++it) {
994 const MCSymbol &Symbol = it->getSymbol();
996 // Ignore non-linker visible symbols.
997 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1000 if (it->isExternal() || Symbol.isUndefined())
1003 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1005 Entry = StringTable.size();
1006 StringTable += Symbol.getName();
1007 StringTable += '\x00';
1011 MSD.SymbolData = it;
1012 MSD.StringIndex = Entry;
1014 if (Symbol.isAbsolute()) {
1015 MSD.SectionIndex = 0;
1016 LocalSymbolData.push_back(MSD);
1018 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1019 assert(MSD.SectionIndex && "Invalid section index!");
1020 LocalSymbolData.push_back(MSD);
1024 // External and undefined symbols are required to be in lexicographic order.
1025 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1026 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1028 // Set the symbol indices.
1030 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1031 LocalSymbolData[i].SymbolData->setIndex(Index++);
1032 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1033 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1034 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1035 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1037 // The string table is padded to a multiple of 4.
1038 while (StringTable.size() % 4)
1039 StringTable += '\x00';
1042 void ExecutePostLayoutBinding(MCAssembler &Asm) {
1043 // Create symbol data for any indirect symbols.
1044 BindIndirectSymbols(Asm);
1046 // Compute symbol table information and bind symbol indices.
1047 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1048 UndefinedSymbolData);
1052 bool IsFixupFullyResolved(const MCAssembler &Asm,
1053 const MCValue Target,
1055 const MCFragment *DF) const {
1056 // If we aren't using scattered symbols, the fixup is fully resolved.
1057 if (!Asm.getBackend().hasScatteredSymbols())
1060 // Otherwise, determine whether this value is actually resolved; scattering
1061 // may cause atoms to move.
1063 // Check if we are using the "simple" resolution algorithm (e.g.,
1065 if (!Asm.getBackend().hasReliableSymbolDifference()) {
1066 const MCSection *BaseSection = 0;
1068 BaseSection = &DF->getParent()->getSection();
1070 return isScatteredFixupFullyResolvedSimple(Asm, Target, BaseSection);
1073 // Otherwise, compute the proper answer as reliably as possible.
1075 // If this is a PCrel relocation, find the base atom (identified by its
1076 // symbol) that the fixup value is relative to.
1077 const MCSymbolData *BaseSymbol = 0;
1079 BaseSymbol = DF->getAtom();
1084 return isScatteredFixupFullyResolved(Asm, Target, BaseSymbol);
1087 void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout) {
1088 unsigned NumSections = Asm.size();
1090 // The section data starts after the header, the segment load command (and
1091 // section headers) and the symbol table.
1092 unsigned NumLoadCommands = 1;
1093 uint64_t LoadCommandsSize = Is64Bit ?
1094 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
1095 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
1097 // Add the symbol table load command sizes, if used.
1098 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1099 UndefinedSymbolData.size();
1101 NumLoadCommands += 2;
1102 LoadCommandsSize += (macho::SymtabLoadCommandSize +
1103 macho::DysymtabLoadCommandSize);
1106 // Compute the total size of the section data, as well as its file size and
1108 uint64_t SectionDataStart = (Is64Bit ? macho::Header64Size :
1109 macho::Header32Size) + LoadCommandsSize;
1110 uint64_t SectionDataSize = 0;
1111 uint64_t SectionDataFileSize = 0;
1112 uint64_t VMSize = 0;
1113 for (MCAssembler::const_iterator it = Asm.begin(),
1114 ie = Asm.end(); it != ie; ++it) {
1115 const MCSectionData &SD = *it;
1116 uint64_t Address = Layout.getSectionAddress(&SD);
1117 uint64_t Size = Layout.getSectionSize(&SD);
1118 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1120 VMSize = std::max(VMSize, Address + Size);
1122 if (SD.getSection().isVirtualSection())
1125 SectionDataSize = std::max(SectionDataSize, Address + Size);
1126 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1129 // The section data is padded to 4 bytes.
1131 // FIXME: Is this machine dependent?
1132 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1133 SectionDataFileSize += SectionDataPadding;
1135 // Write the prolog, starting with the header and load command...
1136 WriteHeader(NumLoadCommands, LoadCommandsSize,
1137 Asm.getSubsectionsViaSymbols());
1138 WriteSegmentLoadCommand(NumSections, VMSize,
1139 SectionDataStart, SectionDataSize);
1141 // ... and then the section headers.
1142 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1143 for (MCAssembler::const_iterator it = Asm.begin(),
1144 ie = Asm.end(); it != ie; ++it) {
1145 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1146 unsigned NumRelocs = Relocs.size();
1147 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
1148 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1149 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
1152 // Write the symbol table load command, if used.
1154 unsigned FirstLocalSymbol = 0;
1155 unsigned NumLocalSymbols = LocalSymbolData.size();
1156 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1157 unsigned NumExternalSymbols = ExternalSymbolData.size();
1158 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1159 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1160 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1161 unsigned NumSymTabSymbols =
1162 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1163 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1164 uint64_t IndirectSymbolOffset = 0;
1166 // If used, the indirect symbols are written after the section data.
1167 if (NumIndirectSymbols)
1168 IndirectSymbolOffset = RelocTableEnd;
1170 // The symbol table is written after the indirect symbol data.
1171 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1173 // The string table is written after symbol table.
1174 uint64_t StringTableOffset =
1175 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? macho::Nlist64Size :
1176 macho::Nlist32Size);
1177 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1178 StringTableOffset, StringTable.size());
1180 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1181 FirstExternalSymbol, NumExternalSymbols,
1182 FirstUndefinedSymbol, NumUndefinedSymbols,
1183 IndirectSymbolOffset, NumIndirectSymbols);
1186 // Write the actual section data.
1187 for (MCAssembler::const_iterator it = Asm.begin(),
1188 ie = Asm.end(); it != ie; ++it)
1189 Asm.WriteSectionData(it, Layout, this);
1191 // Write the extra padding.
1192 WriteZeros(SectionDataPadding);
1194 // Write the relocation entries.
1195 for (MCAssembler::const_iterator it = Asm.begin(),
1196 ie = Asm.end(); it != ie; ++it) {
1197 // Write the section relocation entries, in reverse order to match 'as'
1198 // (approximately, the exact algorithm is more complicated than this).
1199 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1200 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1201 Write32(Relocs[e - i - 1].Word0);
1202 Write32(Relocs[e - i - 1].Word1);
1206 // Write the symbol table data, if used.
1208 // Write the indirect symbol entries.
1209 for (MCAssembler::const_indirect_symbol_iterator
1210 it = Asm.indirect_symbol_begin(),
1211 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1212 // Indirect symbols in the non lazy symbol pointer section have some
1213 // special handling.
1214 const MCSectionMachO &Section =
1215 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1216 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1217 // If this symbol is defined and internal, mark it as such.
1218 if (it->Symbol->isDefined() &&
1219 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1220 uint32_t Flags = macho::ISF_Local;
1221 if (it->Symbol->isAbsolute())
1222 Flags |= macho::ISF_Absolute;
1228 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1231 // FIXME: Check that offsets match computed ones.
1233 // Write the symbol table entries.
1234 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1235 WriteNlist(LocalSymbolData[i], Layout);
1236 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1237 WriteNlist(ExternalSymbolData[i], Layout);
1238 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1239 WriteNlist(UndefinedSymbolData[i], Layout);
1241 // Write the string table.
1242 OS << StringTable.str();
1249 MCObjectWriter *llvm::createMachObjectWriter(raw_ostream &OS, bool is64Bit,
1251 uint32_t CPUSubtype,
1252 bool IsLittleEndian) {
1253 return new MachObjectWriter(OS, is64Bit, CPUType, CPUSubtype, IsLittleEndian);