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/MC/MachObjectWriter.h"
11 #include "llvm/ADT/StringMap.h"
12 #include "llvm/ADT/Twine.h"
13 #include "llvm/MC/MCAssembler.h"
14 #include "llvm/MC/MCAsmLayout.h"
15 #include "llvm/MC/MCExpr.h"
16 #include "llvm/MC/MCObjectWriter.h"
17 #include "llvm/MC/MCSectionMachO.h"
18 #include "llvm/MC/MCSymbol.h"
19 #include "llvm/MC/MCMachOSymbolFlags.h"
20 #include "llvm/MC/MCValue.h"
21 #include "llvm/Support/ErrorHandling.h"
22 #include "llvm/Support/MachO.h"
23 #include "llvm/Target/TargetAsmBackend.h"
26 #include "../Target/X86/X86FixupKinds.h"
31 static unsigned getFixupKindLog2Size(unsigned Kind) {
33 default: llvm_unreachable("invalid fixup kind!");
34 case X86::reloc_pcrel_1byte:
35 case FK_Data_1: return 0;
36 case FK_Data_2: return 1;
37 case X86::reloc_pcrel_4byte:
38 case X86::reloc_riprel_4byte:
39 case X86::reloc_riprel_4byte_movq_load:
40 case FK_Data_4: return 2;
41 case FK_Data_8: return 3;
45 static bool isFixupKindPCRel(unsigned Kind) {
49 case X86::reloc_pcrel_1byte:
50 case X86::reloc_pcrel_4byte:
51 case X86::reloc_riprel_4byte:
52 case X86::reloc_riprel_4byte_movq_load:
57 static bool isFixupKindRIPRel(unsigned Kind) {
58 return Kind == X86::reloc_riprel_4byte ||
59 Kind == X86::reloc_riprel_4byte_movq_load;
64 class MachObjectWriterImpl {
65 // See <mach-o/loader.h>.
67 Header_Magic32 = 0xFEEDFACE,
68 Header_Magic64 = 0xFEEDFACF
74 SegmentLoadCommand32Size = 56,
75 SegmentLoadCommand64Size = 72,
78 SymtabLoadCommandSize = 24,
79 DysymtabLoadCommandSize = 80,
82 RelocationInfoSize = 8
90 HF_SubsectionsViaSymbols = 0x2000
93 enum LoadCommandType {
100 // See <mach-o/nlist.h>.
101 enum SymbolTypeType {
102 STT_Undefined = 0x00,
107 enum SymbolTypeFlags {
108 // If any of these bits are set, then the entry is a stab entry number (see
109 // <mach-o/stab.h>. Otherwise the other masks apply.
110 STF_StabsEntryMask = 0xe0,
114 STF_PrivateExtern = 0x10
117 /// IndirectSymbolFlags - Flags for encoding special values in the indirect
119 enum IndirectSymbolFlags {
120 ISF_Local = 0x80000000,
121 ISF_Absolute = 0x40000000
124 /// RelocationFlags - Special flags for addresses.
125 enum RelocationFlags {
126 RF_Scattered = 0x80000000
129 enum RelocationInfoType {
133 RIT_PreboundLazyPointer = 3,
134 RIT_LocalDifference = 4
137 /// X86_64 uses its own relocation types.
138 enum RelocationInfoTypeX86_64 {
139 RIT_X86_64_Unsigned = 0,
140 RIT_X86_64_Signed = 1,
141 RIT_X86_64_Branch = 2,
142 RIT_X86_64_GOTLoad = 3,
144 RIT_X86_64_Subtractor = 5,
145 RIT_X86_64_Signed1 = 6,
146 RIT_X86_64_Signed2 = 7,
147 RIT_X86_64_Signed4 = 8
150 /// MachSymbolData - Helper struct for containing some precomputed information
152 struct MachSymbolData {
153 MCSymbolData *SymbolData;
154 uint64_t StringIndex;
155 uint8_t SectionIndex;
157 // Support lexicographic sorting.
158 bool operator<(const MachSymbolData &RHS) const {
159 const std::string &Name = SymbolData->getSymbol().getName();
160 return Name < RHS.SymbolData->getSymbol().getName();
164 /// @name Relocation Data
167 struct MachRelocationEntry {
172 llvm::DenseMap<const MCSectionData*,
173 std::vector<MachRelocationEntry> > Relocations;
176 /// @name Symbol Table Data
179 SmallString<256> StringTable;
180 std::vector<MachSymbolData> LocalSymbolData;
181 std::vector<MachSymbolData> ExternalSymbolData;
182 std::vector<MachSymbolData> UndefinedSymbolData;
186 MachObjectWriter *Writer;
190 unsigned Is64Bit : 1;
193 MachObjectWriterImpl(MachObjectWriter *_Writer, bool _Is64Bit)
194 : Writer(_Writer), OS(Writer->getStream()), Is64Bit(_Is64Bit) {
197 void Write8(uint8_t Value) { Writer->Write8(Value); }
198 void Write16(uint16_t Value) { Writer->Write16(Value); }
199 void Write32(uint32_t Value) { Writer->Write32(Value); }
200 void Write64(uint64_t Value) { Writer->Write64(Value); }
201 void WriteZeros(unsigned N) { Writer->WriteZeros(N); }
202 void WriteBytes(StringRef Str, unsigned ZeroFillSize = 0) {
203 Writer->WriteBytes(Str, ZeroFillSize);
206 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
207 bool SubsectionsViaSymbols) {
210 if (SubsectionsViaSymbols)
211 Flags |= HF_SubsectionsViaSymbols;
213 // struct mach_header (28 bytes) or
214 // struct mach_header_64 (32 bytes)
216 uint64_t Start = OS.tell();
219 Write32(Is64Bit ? Header_Magic64 : Header_Magic32);
221 // FIXME: Support cputype.
222 Write32(Is64Bit ? MachO::CPUTypeX86_64 : MachO::CPUTypeI386);
223 // FIXME: Support cpusubtype.
224 Write32(MachO::CPUSubType_I386_ALL);
226 Write32(NumLoadCommands); // Object files have a single load command, the
228 Write32(LoadCommandsSize);
231 Write32(0); // reserved
233 assert(OS.tell() - Start == Is64Bit ? Header64Size : Header32Size);
236 /// WriteSegmentLoadCommand - Write a segment load command.
238 /// \arg NumSections - The number of sections in this segment.
239 /// \arg SectionDataSize - The total size of the sections.
240 void WriteSegmentLoadCommand(unsigned NumSections,
242 uint64_t SectionDataStartOffset,
243 uint64_t SectionDataSize) {
244 // struct segment_command (56 bytes) or
245 // struct segment_command_64 (72 bytes)
247 uint64_t Start = OS.tell();
250 unsigned SegmentLoadCommandSize = Is64Bit ? SegmentLoadCommand64Size :
251 SegmentLoadCommand32Size;
252 Write32(Is64Bit ? LCT_Segment64 : LCT_Segment);
253 Write32(SegmentLoadCommandSize +
254 NumSections * (Is64Bit ? Section64Size : Section32Size));
258 Write64(0); // vmaddr
259 Write64(VMSize); // vmsize
260 Write64(SectionDataStartOffset); // file offset
261 Write64(SectionDataSize); // file size
263 Write32(0); // vmaddr
264 Write32(VMSize); // vmsize
265 Write32(SectionDataStartOffset); // file offset
266 Write32(SectionDataSize); // file size
268 Write32(0x7); // maxprot
269 Write32(0x7); // initprot
270 Write32(NumSections);
273 assert(OS.tell() - Start == SegmentLoadCommandSize);
276 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
277 const MCSectionData &SD, uint64_t FileOffset,
278 uint64_t RelocationsStart, unsigned NumRelocations) {
279 uint64_t SectionSize = Layout.getSectionSize(&SD);
281 // The offset is unused for virtual sections.
282 if (Asm.getBackend().isVirtualSection(SD.getSection())) {
283 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
287 // struct section (68 bytes) or
288 // struct section_64 (80 bytes)
290 uint64_t Start = OS.tell();
293 // FIXME: cast<> support!
294 const MCSectionMachO &Section =
295 static_cast<const MCSectionMachO&>(SD.getSection());
296 WriteBytes(Section.getSectionName(), 16);
297 WriteBytes(Section.getSegmentName(), 16);
299 Write64(Layout.getSectionAddress(&SD)); // address
300 Write64(SectionSize); // size
302 Write32(Layout.getSectionAddress(&SD)); // address
303 Write32(SectionSize); // size
307 unsigned Flags = Section.getTypeAndAttributes();
308 if (SD.hasInstructions())
309 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
311 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
312 Write32(Log2_32(SD.getAlignment()));
313 Write32(NumRelocations ? RelocationsStart : 0);
314 Write32(NumRelocations);
316 Write32(0); // reserved1
317 Write32(Section.getStubSize()); // reserved2
319 Write32(0); // reserved3
321 assert(OS.tell() - Start == Is64Bit ? Section64Size : Section32Size);
324 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
325 uint32_t StringTableOffset,
326 uint32_t StringTableSize) {
327 // struct symtab_command (24 bytes)
329 uint64_t Start = OS.tell();
333 Write32(SymtabLoadCommandSize);
334 Write32(SymbolOffset);
336 Write32(StringTableOffset);
337 Write32(StringTableSize);
339 assert(OS.tell() - Start == SymtabLoadCommandSize);
342 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
343 uint32_t NumLocalSymbols,
344 uint32_t FirstExternalSymbol,
345 uint32_t NumExternalSymbols,
346 uint32_t FirstUndefinedSymbol,
347 uint32_t NumUndefinedSymbols,
348 uint32_t IndirectSymbolOffset,
349 uint32_t NumIndirectSymbols) {
350 // struct dysymtab_command (80 bytes)
352 uint64_t Start = OS.tell();
355 Write32(LCT_Dysymtab);
356 Write32(DysymtabLoadCommandSize);
357 Write32(FirstLocalSymbol);
358 Write32(NumLocalSymbols);
359 Write32(FirstExternalSymbol);
360 Write32(NumExternalSymbols);
361 Write32(FirstUndefinedSymbol);
362 Write32(NumUndefinedSymbols);
363 Write32(0); // tocoff
365 Write32(0); // modtaboff
366 Write32(0); // nmodtab
367 Write32(0); // extrefsymoff
368 Write32(0); // nextrefsyms
369 Write32(IndirectSymbolOffset);
370 Write32(NumIndirectSymbols);
371 Write32(0); // extreloff
372 Write32(0); // nextrel
373 Write32(0); // locreloff
374 Write32(0); // nlocrel
376 assert(OS.tell() - Start == DysymtabLoadCommandSize);
379 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
380 MCSymbolData &Data = *MSD.SymbolData;
381 const MCSymbol &Symbol = Data.getSymbol();
383 uint16_t Flags = Data.getFlags();
384 uint32_t Address = 0;
386 // Set the N_TYPE bits. See <mach-o/nlist.h>.
388 // FIXME: Are the prebound or indirect fields possible here?
389 if (Symbol.isUndefined())
390 Type = STT_Undefined;
391 else if (Symbol.isAbsolute())
396 // FIXME: Set STAB bits.
398 if (Data.isPrivateExtern())
399 Type |= STF_PrivateExtern;
402 if (Data.isExternal() || Symbol.isUndefined())
403 Type |= STF_External;
405 // Compute the symbol address.
406 if (Symbol.isDefined()) {
407 if (Symbol.isAbsolute()) {
408 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
410 Address = Layout.getSymbolAddress(&Data);
412 } else if (Data.isCommon()) {
413 // Common symbols are encoded with the size in the address
414 // field, and their alignment in the flags.
415 Address = Data.getCommonSize();
417 // Common alignment is packed into the 'desc' bits.
418 if (unsigned Align = Data.getCommonAlignment()) {
419 unsigned Log2Size = Log2_32(Align);
420 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
422 report_fatal_error("invalid 'common' alignment '" +
424 // FIXME: Keep this mask with the SymbolFlags enumeration.
425 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
429 // struct nlist (12 bytes)
431 Write32(MSD.StringIndex);
433 Write8(MSD.SectionIndex);
435 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
444 // FIXME: We really need to improve the relocation validation. Basically, we
445 // want to implement a separate computation which evaluates the relocation
446 // entry as the linker would, and verifies that the resultant fixup value is
447 // exactly what the encoder wanted. This will catch several classes of
450 // - Relocation entry bugs, the two algorithms are unlikely to have the same
453 // - Relaxation issues, where we forget to relax something.
455 // - Input errors, where something cannot be correctly encoded. 'as' allows
456 // these through in many cases.
458 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
459 const MCFragment *Fragment,
460 const MCAsmFixup &Fixup, MCValue Target,
461 uint64_t &FixedValue) {
462 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
463 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.Kind);
464 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
467 uint32_t Address = Layout.getFragmentOffset(Fragment) + Fixup.Offset;
470 unsigned IsExtern = 0;
473 Value = Target.getConstant();
476 // Compensate for the relocation offset, Darwin x86_64 relocations only
477 // have the addend and appear to have attempted to define it to be the
478 // actual expression addend without the PCrel bias. However, instructions
479 // with data following the relocation are not accomodated for (see comment
480 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
481 Value += 1LL << Log2Size;
484 if (Target.isAbsolute()) { // constant
485 // SymbolNum of 0 indicates the absolute section.
486 Type = RIT_X86_64_Unsigned;
489 // FIXME: I believe this is broken, I don't think the linker can
490 // understand it. I think it would require a local relocation, but I'm not
491 // sure if that would work either. The official way to get an absolute
492 // PCrel relocation is to use an absolute symbol (which we don't support
496 Type = RIT_X86_64_Branch;
498 } else if (Target.getSymB()) { // A - B + constant
499 const MCSymbol *A = &Target.getSymA()->getSymbol();
500 MCSymbolData &A_SD = Asm.getSymbolData(*A);
501 const MCSymbolData *A_Base = Asm.getAtom(Layout, &A_SD);
503 const MCSymbol *B = &Target.getSymB()->getSymbol();
504 MCSymbolData &B_SD = Asm.getSymbolData(*B);
505 const MCSymbolData *B_Base = Asm.getAtom(Layout, &B_SD);
507 // Neither symbol can be modified.
508 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
509 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
510 report_fatal_error("unsupported relocation of modified symbol");
512 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
513 // implement most of these correctly.
515 report_fatal_error("unsupported pc-relative relocation of difference");
517 // We don't currently support any situation where one or both of the
518 // symbols would require a local relocation. This is almost certainly
519 // unused and may not be possible to encode correctly.
520 if (!A_Base || !B_Base)
521 report_fatal_error("unsupported local relocations in difference");
523 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
524 // a single SIGNED relocation); reject it for now.
525 if (A_Base == B_Base)
526 report_fatal_error("unsupported relocation with identical base");
528 Value += Layout.getSymbolAddress(&A_SD) - Layout.getSymbolAddress(A_Base);
529 Value -= Layout.getSymbolAddress(&B_SD) - Layout.getSymbolAddress(B_Base);
531 Index = A_Base->getIndex();
533 Type = RIT_X86_64_Unsigned;
535 MachRelocationEntry MRE;
537 MRE.Word1 = ((Index << 0) |
542 Relocations[Fragment->getParent()].push_back(MRE);
544 Index = B_Base->getIndex();
546 Type = RIT_X86_64_Subtractor;
548 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
549 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
550 const MCSymbolData *Base = Asm.getAtom(Layout, &SD);
552 // Relocations inside debug sections always use local relocations when
553 // possible. This seems to be done because the debugger doesn't fully
554 // understand x86_64 relocation entries, and expects to find values that
555 // have already been fixed up.
556 if (Symbol->isInSection()) {
557 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
558 Fragment->getParent()->getSection());
559 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
563 // x86_64 almost always uses external relocations, except when there is no
564 // symbol to use as a base address (a local symbol with no preceeding
565 // non-local symbol).
567 Index = Base->getIndex();
570 // Add the local offset, if needed.
572 Value += Layout.getSymbolAddress(&SD) - Layout.getSymbolAddress(Base);
574 // The index is the section ordinal (1-based).
575 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
577 Value += Layout.getSymbolAddress(&SD);
580 Value -= Address + (1 << Log2Size);
583 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
586 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
587 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
588 // rewrite the movq to an leaq at link time if the symbol ends up in
589 // the same linkage unit.
590 if (unsigned(Fixup.Kind) == X86::reloc_riprel_4byte_movq_load)
591 Type = RIT_X86_64_GOTLoad;
593 Type = RIT_X86_64_GOT;
594 } else if (Modifier != MCSymbolRefExpr::VK_None)
595 report_fatal_error("unsupported symbol modifier in relocation");
597 Type = RIT_X86_64_Signed;
599 if (Modifier != MCSymbolRefExpr::VK_None)
600 report_fatal_error("unsupported symbol modifier in branch "
603 Type = RIT_X86_64_Branch;
606 // The Darwin x86_64 relocation format has a problem where it cannot
607 // encode an address (L<foo> + <constant>) which is outside the atom
608 // containing L<foo>. Generally, this shouldn't occur but it does happen
609 // when we have a RIPrel instruction with data following the relocation
610 // entry (e.g., movb $012, L0(%rip)). Even with the PCrel adjustment
611 // Darwin x86_64 uses, the offset is still negative and the linker has
612 // no way to recognize this.
614 // To work around this, Darwin uses several special relocation types to
615 // indicate the offsets. However, the specification or implementation of
616 // these seems to also be incomplete; they should adjust the addend as
617 // well based on the actual encoded instruction (the additional bias),
618 // but instead appear to just look at the final offset.
620 switch (-(Target.getConstant() + (1LL << Log2Size))) {
621 case 1: Type = RIT_X86_64_Signed1; break;
622 case 2: Type = RIT_X86_64_Signed2; break;
623 case 4: Type = RIT_X86_64_Signed4; break;
627 if (Modifier == MCSymbolRefExpr::VK_GOT) {
628 Type = RIT_X86_64_GOT;
629 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
630 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
631 // which case all we do is set the PCrel bit in the relocation entry;
632 // this is used with exception handling, for example. The source is
633 // required to include any necessary offset directly.
634 Type = RIT_X86_64_GOT;
636 } else if (Modifier != MCSymbolRefExpr::VK_None)
637 report_fatal_error("unsupported symbol modifier in relocation");
639 Type = RIT_X86_64_Unsigned;
643 // x86_64 always writes custom values into the fixups.
646 // struct relocation_info (8 bytes)
647 MachRelocationEntry MRE;
649 MRE.Word1 = ((Index << 0) |
654 Relocations[Fragment->getParent()].push_back(MRE);
657 void RecordScatteredRelocation(const MCAssembler &Asm,
658 const MCAsmLayout &Layout,
659 const MCFragment *Fragment,
660 const MCAsmFixup &Fixup, MCValue Target,
661 uint64_t &FixedValue) {
662 uint32_t Address = Layout.getFragmentOffset(Fragment) + Fixup.Offset;
663 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
664 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
665 unsigned Type = RIT_Vanilla;
668 const MCSymbol *A = &Target.getSymA()->getSymbol();
669 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
671 if (!A_SD->getFragment())
672 report_fatal_error("symbol '" + A->getName() +
673 "' can not be undefined in a subtraction expression");
675 uint32_t Value = Layout.getSymbolAddress(A_SD);
678 if (const MCSymbolRefExpr *B = Target.getSymB()) {
679 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
681 if (!B_SD->getFragment())
682 report_fatal_error("symbol '" + B->getSymbol().getName() +
683 "' can not be undefined in a subtraction expression");
685 // Select the appropriate difference relocation type.
687 // Note that there is no longer any semantic difference between these two
688 // relocation types from the linkers point of view, this is done solely
689 // for pedantic compatibility with 'as'.
690 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
691 Value2 = Layout.getSymbolAddress(B_SD);
694 // Relocations are written out in reverse order, so the PAIR comes first.
695 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
696 MachRelocationEntry MRE;
697 MRE.Word0 = ((0 << 0) |
703 Relocations[Fragment->getParent()].push_back(MRE);
706 MachRelocationEntry MRE;
707 MRE.Word0 = ((Address << 0) |
713 Relocations[Fragment->getParent()].push_back(MRE);
716 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
717 const MCFragment *Fragment, const MCAsmFixup &Fixup,
718 MCValue Target, uint64_t &FixedValue) {
720 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
724 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
725 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
727 // If this is a difference or a defined symbol plus an offset, then we need
728 // a scattered relocation entry.
729 uint32_t Offset = Target.getConstant();
731 Offset += 1 << Log2Size;
732 if (Target.getSymB() ||
733 (Target.getSymA() && !Target.getSymA()->getSymbol().isUndefined() &&
735 RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,Target,FixedValue);
740 uint32_t Address = Layout.getFragmentOffset(Fragment) + Fixup.Offset;
743 unsigned IsExtern = 0;
746 if (Target.isAbsolute()) { // constant
747 // SymbolNum of 0 indicates the absolute section.
749 // FIXME: Currently, these are never generated (see code below). I cannot
750 // find a case where they are actually emitted.
754 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
755 MCSymbolData *SD = &Asm.getSymbolData(*Symbol);
757 // Both references to undefined symbols and references to Weak Definitions
758 // get external relocation entries. This is so the static and then the
759 // the dynamic linker can resolve them to the actual definition that will
760 // be used. And in the case of Weak Definitions a reference to one will
761 // not always be to the definition in the same object file.
762 if (Symbol->isUndefined() || (SD->getFlags() & SF_WeakDefinition)) {
764 Index = SD->getIndex();
765 // In the case of a Weak Definition the FixedValue needs to be set to
766 // to not have the address of the symbol. In the case of an undefined
767 // symbol you can't call getSymbolAddress().
768 if (SD->getFlags() & SF_WeakDefinition)
769 FixedValue -= Layout.getSymbolAddress(SD);
772 // The index is the section ordinal (1-based).
773 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
774 Value = Layout.getSymbolAddress(SD);
780 // struct relocation_info (8 bytes)
781 MachRelocationEntry MRE;
783 MRE.Word1 = ((Index << 0) |
788 Relocations[Fragment->getParent()].push_back(MRE);
791 void BindIndirectSymbols(MCAssembler &Asm) {
792 // This is the point where 'as' creates actual symbols for indirect symbols
793 // (in the following two passes). It would be easier for us to do this
794 // sooner when we see the attribute, but that makes getting the order in the
795 // symbol table much more complicated than it is worth.
797 // FIXME: Revisit this when the dust settles.
799 // Bind non lazy symbol pointers first.
800 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
801 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
802 // FIXME: cast<> support!
803 const MCSectionMachO &Section =
804 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
806 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
809 Asm.getOrCreateSymbolData(*it->Symbol);
812 // Then lazy symbol pointers and symbol stubs.
813 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
814 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
815 // FIXME: cast<> support!
816 const MCSectionMachO &Section =
817 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
819 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
820 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
823 // Set the symbol type to undefined lazy, but only on construction.
825 // FIXME: Do not hardcode.
827 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
829 Entry.setFlags(Entry.getFlags() | 0x0001);
833 /// ComputeSymbolTable - Compute the symbol table data
835 /// \param StringTable [out] - The string table data.
836 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
838 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
839 std::vector<MachSymbolData> &LocalSymbolData,
840 std::vector<MachSymbolData> &ExternalSymbolData,
841 std::vector<MachSymbolData> &UndefinedSymbolData) {
842 // Build section lookup table.
843 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
845 for (MCAssembler::iterator it = Asm.begin(),
846 ie = Asm.end(); it != ie; ++it, ++Index)
847 SectionIndexMap[&it->getSection()] = Index;
848 assert(Index <= 256 && "Too many sections!");
850 // Index 0 is always the empty string.
851 StringMap<uint64_t> StringIndexMap;
852 StringTable += '\x00';
854 // Build the symbol arrays and the string table, but only for non-local
857 // The particular order that we collect the symbols and create the string
858 // table, then sort the symbols is chosen to match 'as'. Even though it
859 // doesn't matter for correctness, this is important for letting us diff .o
861 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
862 ie = Asm.symbol_end(); it != ie; ++it) {
863 const MCSymbol &Symbol = it->getSymbol();
865 // Ignore non-linker visible symbols.
866 if (!Asm.isSymbolLinkerVisible(it))
869 if (!it->isExternal() && !Symbol.isUndefined())
872 uint64_t &Entry = StringIndexMap[Symbol.getName()];
874 Entry = StringTable.size();
875 StringTable += Symbol.getName();
876 StringTable += '\x00';
881 MSD.StringIndex = Entry;
883 if (Symbol.isUndefined()) {
884 MSD.SectionIndex = 0;
885 UndefinedSymbolData.push_back(MSD);
886 } else if (Symbol.isAbsolute()) {
887 MSD.SectionIndex = 0;
888 ExternalSymbolData.push_back(MSD);
890 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
891 assert(MSD.SectionIndex && "Invalid section index!");
892 ExternalSymbolData.push_back(MSD);
896 // Now add the data for local symbols.
897 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
898 ie = Asm.symbol_end(); it != ie; ++it) {
899 const MCSymbol &Symbol = it->getSymbol();
901 // Ignore non-linker visible symbols.
902 if (!Asm.isSymbolLinkerVisible(it))
905 if (it->isExternal() || Symbol.isUndefined())
908 uint64_t &Entry = StringIndexMap[Symbol.getName()];
910 Entry = StringTable.size();
911 StringTable += Symbol.getName();
912 StringTable += '\x00';
917 MSD.StringIndex = Entry;
919 if (Symbol.isAbsolute()) {
920 MSD.SectionIndex = 0;
921 LocalSymbolData.push_back(MSD);
923 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
924 assert(MSD.SectionIndex && "Invalid section index!");
925 LocalSymbolData.push_back(MSD);
929 // External and undefined symbols are required to be in lexicographic order.
930 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
931 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
933 // Set the symbol indices.
935 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
936 LocalSymbolData[i].SymbolData->setIndex(Index++);
937 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
938 ExternalSymbolData[i].SymbolData->setIndex(Index++);
939 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
940 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
942 // The string table is padded to a multiple of 4.
943 while (StringTable.size() % 4)
944 StringTable += '\x00';
947 void ExecutePostLayoutBinding(MCAssembler &Asm) {
948 // Create symbol data for any indirect symbols.
949 BindIndirectSymbols(Asm);
951 // Compute symbol table information and bind symbol indices.
952 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
953 UndefinedSymbolData);
956 void WriteObject(const MCAssembler &Asm, const MCAsmLayout &Layout) {
957 unsigned NumSections = Asm.size();
959 // The section data starts after the header, the segment load command (and
960 // section headers) and the symbol table.
961 unsigned NumLoadCommands = 1;
962 uint64_t LoadCommandsSize = Is64Bit ?
963 SegmentLoadCommand64Size + NumSections * Section64Size :
964 SegmentLoadCommand32Size + NumSections * Section32Size;
966 // Add the symbol table load command sizes, if used.
967 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
968 UndefinedSymbolData.size();
970 NumLoadCommands += 2;
971 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
974 // Compute the total size of the section data, as well as its file size and
976 uint64_t SectionDataStart = (Is64Bit ? Header64Size : Header32Size)
978 uint64_t SectionDataSize = 0;
979 uint64_t SectionDataFileSize = 0;
981 for (MCAssembler::const_iterator it = Asm.begin(),
982 ie = Asm.end(); it != ie; ++it) {
983 const MCSectionData &SD = *it;
984 uint64_t Address = Layout.getSectionAddress(&SD);
985 uint64_t Size = Layout.getSectionSize(&SD);
986 uint64_t FileSize = Layout.getSectionFileSize(&SD);
988 VMSize = std::max(VMSize, Address + Size);
990 if (Asm.getBackend().isVirtualSection(SD.getSection()))
993 SectionDataSize = std::max(SectionDataSize, Address + Size);
994 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
997 // The section data is padded to 4 bytes.
999 // FIXME: Is this machine dependent?
1000 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1001 SectionDataFileSize += SectionDataPadding;
1003 // Write the prolog, starting with the header and load command...
1004 WriteHeader(NumLoadCommands, LoadCommandsSize,
1005 Asm.getSubsectionsViaSymbols());
1006 WriteSegmentLoadCommand(NumSections, VMSize,
1007 SectionDataStart, SectionDataSize);
1009 // ... and then the section headers.
1010 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1011 for (MCAssembler::const_iterator it = Asm.begin(),
1012 ie = Asm.end(); it != ie; ++it) {
1013 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1014 unsigned NumRelocs = Relocs.size();
1015 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
1016 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1017 RelocTableEnd += NumRelocs * RelocationInfoSize;
1020 // Write the symbol table load command, if used.
1022 unsigned FirstLocalSymbol = 0;
1023 unsigned NumLocalSymbols = LocalSymbolData.size();
1024 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1025 unsigned NumExternalSymbols = ExternalSymbolData.size();
1026 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1027 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1028 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1029 unsigned NumSymTabSymbols =
1030 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1031 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1032 uint64_t IndirectSymbolOffset = 0;
1034 // If used, the indirect symbols are written after the section data.
1035 if (NumIndirectSymbols)
1036 IndirectSymbolOffset = RelocTableEnd;
1038 // The symbol table is written after the indirect symbol data.
1039 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1041 // The string table is written after symbol table.
1042 uint64_t StringTableOffset =
1043 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? Nlist64Size :
1045 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1046 StringTableOffset, StringTable.size());
1048 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1049 FirstExternalSymbol, NumExternalSymbols,
1050 FirstUndefinedSymbol, NumUndefinedSymbols,
1051 IndirectSymbolOffset, NumIndirectSymbols);
1054 // Write the actual section data.
1055 for (MCAssembler::const_iterator it = Asm.begin(),
1056 ie = Asm.end(); it != ie; ++it)
1057 Asm.WriteSectionData(it, Layout, Writer);
1059 // Write the extra padding.
1060 WriteZeros(SectionDataPadding);
1062 // Write the relocation entries.
1063 for (MCAssembler::const_iterator it = Asm.begin(),
1064 ie = Asm.end(); it != ie; ++it) {
1065 // Write the section relocation entries, in reverse order to match 'as'
1066 // (approximately, the exact algorithm is more complicated than this).
1067 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1068 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1069 Write32(Relocs[e - i - 1].Word0);
1070 Write32(Relocs[e - i - 1].Word1);
1074 // Write the symbol table data, if used.
1076 // Write the indirect symbol entries.
1077 for (MCAssembler::const_indirect_symbol_iterator
1078 it = Asm.indirect_symbol_begin(),
1079 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1080 // Indirect symbols in the non lazy symbol pointer section have some
1081 // special handling.
1082 const MCSectionMachO &Section =
1083 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1084 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1085 // If this symbol is defined and internal, mark it as such.
1086 if (it->Symbol->isDefined() &&
1087 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1088 uint32_t Flags = ISF_Local;
1089 if (it->Symbol->isAbsolute())
1090 Flags |= ISF_Absolute;
1096 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1099 // FIXME: Check that offsets match computed ones.
1101 // Write the symbol table entries.
1102 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1103 WriteNlist(LocalSymbolData[i], Layout);
1104 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1105 WriteNlist(ExternalSymbolData[i], Layout);
1106 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1107 WriteNlist(UndefinedSymbolData[i], Layout);
1109 // Write the string table.
1110 OS << StringTable.str();
1117 MachObjectWriter::MachObjectWriter(raw_ostream &OS,
1119 bool IsLittleEndian)
1120 : MCObjectWriter(OS, IsLittleEndian)
1122 Impl = new MachObjectWriterImpl(this, Is64Bit);
1125 MachObjectWriter::~MachObjectWriter() {
1126 delete (MachObjectWriterImpl*) Impl;
1129 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm) {
1130 ((MachObjectWriterImpl*) Impl)->ExecutePostLayoutBinding(Asm);
1133 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1134 const MCAsmLayout &Layout,
1135 const MCFragment *Fragment,
1136 const MCAsmFixup &Fixup, MCValue Target,
1137 uint64_t &FixedValue) {
1138 ((MachObjectWriterImpl*) Impl)->RecordRelocation(Asm, Layout, Fragment, Fixup,
1139 Target, FixedValue);
1142 void MachObjectWriter::WriteObject(const MCAssembler &Asm,
1143 const MCAsmLayout &Layout) {
1144 ((MachObjectWriterImpl*) Impl)->WriteObject(Asm, Layout);