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/MCValue.h"
20 #include "llvm/Support/ErrorHandling.h"
21 #include "llvm/Support/MachO.h"
22 #include "llvm/Target/TargetAsmBackend.h"
25 #include "../Target/X86/X86FixupKinds.h"
30 static unsigned getFixupKindLog2Size(unsigned Kind) {
32 default: llvm_unreachable("invalid fixup kind!");
33 case X86::reloc_pcrel_1byte:
34 case FK_Data_1: return 0;
35 case FK_Data_2: return 1;
36 case X86::reloc_pcrel_4byte:
37 case X86::reloc_riprel_4byte:
38 case X86::reloc_riprel_4byte_movq_load:
39 case FK_Data_4: return 2;
40 case FK_Data_8: return 3;
44 static bool isFixupKindPCRel(unsigned Kind) {
48 case X86::reloc_pcrel_1byte:
49 case X86::reloc_pcrel_4byte:
50 case X86::reloc_riprel_4byte:
51 case X86::reloc_riprel_4byte_movq_load:
56 static bool isFixupKindRIPRel(unsigned Kind) {
57 return Kind == X86::reloc_riprel_4byte ||
58 Kind == X86::reloc_riprel_4byte_movq_load;
63 class MachObjectWriterImpl {
64 // See <mach-o/loader.h>.
66 Header_Magic32 = 0xFEEDFACE,
67 Header_Magic64 = 0xFEEDFACF
73 SegmentLoadCommand32Size = 56,
74 SegmentLoadCommand64Size = 72,
77 SymtabLoadCommandSize = 24,
78 DysymtabLoadCommandSize = 80,
81 RelocationInfoSize = 8
89 HF_SubsectionsViaSymbols = 0x2000
92 enum LoadCommandType {
99 // See <mach-o/nlist.h>.
100 enum SymbolTypeType {
101 STT_Undefined = 0x00,
106 enum SymbolTypeFlags {
107 // If any of these bits are set, then the entry is a stab entry number (see
108 // <mach-o/stab.h>. Otherwise the other masks apply.
109 STF_StabsEntryMask = 0xe0,
113 STF_PrivateExtern = 0x10
116 /// IndirectSymbolFlags - Flags for encoding special values in the indirect
118 enum IndirectSymbolFlags {
119 ISF_Local = 0x80000000,
120 ISF_Absolute = 0x40000000
123 /// RelocationFlags - Special flags for addresses.
124 enum RelocationFlags {
125 RF_Scattered = 0x80000000
128 enum RelocationInfoType {
132 RIT_PreboundLazyPointer = 3,
133 RIT_LocalDifference = 4
136 /// X86_64 uses its own relocation types.
137 enum RelocationInfoTypeX86_64 {
138 RIT_X86_64_Unsigned = 0,
139 RIT_X86_64_Signed = 1,
140 RIT_X86_64_Branch = 2,
141 RIT_X86_64_GOTLoad = 3,
143 RIT_X86_64_Subtractor = 5,
144 RIT_X86_64_Signed1 = 6,
145 RIT_X86_64_Signed2 = 7,
146 RIT_X86_64_Signed4 = 8
149 /// MachSymbolData - Helper struct for containing some precomputed information
151 struct MachSymbolData {
152 MCSymbolData *SymbolData;
153 uint64_t StringIndex;
154 uint8_t SectionIndex;
156 // Support lexicographic sorting.
157 bool operator<(const MachSymbolData &RHS) const {
158 const std::string &Name = SymbolData->getSymbol().getName();
159 return Name < RHS.SymbolData->getSymbol().getName();
163 /// @name Relocation Data
166 struct MachRelocationEntry {
171 llvm::DenseMap<const MCSectionData*,
172 std::vector<MachRelocationEntry> > Relocations;
175 /// @name Symbol Table Data
178 SmallString<256> StringTable;
179 std::vector<MachSymbolData> LocalSymbolData;
180 std::vector<MachSymbolData> ExternalSymbolData;
181 std::vector<MachSymbolData> UndefinedSymbolData;
185 MachObjectWriter *Writer;
189 unsigned Is64Bit : 1;
192 MachObjectWriterImpl(MachObjectWriter *_Writer, bool _Is64Bit)
193 : Writer(_Writer), OS(Writer->getStream()), Is64Bit(_Is64Bit) {
196 void Write8(uint8_t Value) { Writer->Write8(Value); }
197 void Write16(uint16_t Value) { Writer->Write16(Value); }
198 void Write32(uint32_t Value) { Writer->Write32(Value); }
199 void Write64(uint64_t Value) { Writer->Write64(Value); }
200 void WriteZeros(unsigned N) { Writer->WriteZeros(N); }
201 void WriteBytes(StringRef Str, unsigned ZeroFillSize = 0) {
202 Writer->WriteBytes(Str, ZeroFillSize);
205 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
206 bool SubsectionsViaSymbols) {
209 if (SubsectionsViaSymbols)
210 Flags |= HF_SubsectionsViaSymbols;
212 // struct mach_header (28 bytes) or
213 // struct mach_header_64 (32 bytes)
215 uint64_t Start = OS.tell();
218 Write32(Is64Bit ? Header_Magic64 : Header_Magic32);
220 // FIXME: Support cputype.
221 Write32(Is64Bit ? MachO::CPUTypeX86_64 : MachO::CPUTypeI386);
222 // FIXME: Support cpusubtype.
223 Write32(MachO::CPUSubType_I386_ALL);
225 Write32(NumLoadCommands); // Object files have a single load command, the
227 Write32(LoadCommandsSize);
230 Write32(0); // reserved
232 assert(OS.tell() - Start == Is64Bit ? Header64Size : Header32Size);
235 /// WriteSegmentLoadCommand - Write a segment load command.
237 /// \arg NumSections - The number of sections in this segment.
238 /// \arg SectionDataSize - The total size of the sections.
239 void WriteSegmentLoadCommand(unsigned NumSections,
241 uint64_t SectionDataStartOffset,
242 uint64_t SectionDataSize) {
243 // struct segment_command (56 bytes) or
244 // struct segment_command_64 (72 bytes)
246 uint64_t Start = OS.tell();
249 unsigned SegmentLoadCommandSize = Is64Bit ? SegmentLoadCommand64Size :
250 SegmentLoadCommand32Size;
251 Write32(Is64Bit ? LCT_Segment64 : LCT_Segment);
252 Write32(SegmentLoadCommandSize +
253 NumSections * (Is64Bit ? Section64Size : Section32Size));
257 Write64(0); // vmaddr
258 Write64(VMSize); // vmsize
259 Write64(SectionDataStartOffset); // file offset
260 Write64(SectionDataSize); // file size
262 Write32(0); // vmaddr
263 Write32(VMSize); // vmsize
264 Write32(SectionDataStartOffset); // file offset
265 Write32(SectionDataSize); // file size
267 Write32(0x7); // maxprot
268 Write32(0x7); // initprot
269 Write32(NumSections);
272 assert(OS.tell() - Start == SegmentLoadCommandSize);
275 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
276 const MCSectionData &SD, uint64_t FileOffset,
277 uint64_t RelocationsStart, unsigned NumRelocations) {
278 uint64_t SectionSize = Layout.getSectionSize(&SD);
279 uint64_t SectionFileSize = Layout.getSectionFileSize(&SD);
281 // The offset is unused for virtual sections.
282 if (Asm.getBackend().isVirtualSection(SD.getSection())) {
283 assert(SectionFileSize == 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 llvm_unreachable("FIXME: Not yet implemented!");
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 llvm_report_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 += 1 << 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 llvm_report_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 llvm_report_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 llvm_report_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 llvm_report_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 // x86_64 almost always uses external relocations, except when there is no
553 // symbol to use as a base address (a local symbol with no preceeding
554 // non-local symbol).
556 Index = Base->getIndex();
559 // Add the local offset, if needed.
561 Value += Layout.getSymbolAddress(&SD) - Layout.getSymbolAddress(Base);
563 // The index is the section ordinal.
567 MCAssembler::const_iterator it = Asm.begin(), ie = Asm.end();
568 for (; it != ie; ++it, ++Index)
569 if (&*it == SD.getFragment()->getParent())
571 assert(it != ie && "Unable to find section index!");
573 Value += Layout.getSymbolAddress(&SD);
576 Value -= Address + (1 << Log2Size);
579 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
582 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
583 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
584 // rewrite the movq to an leaq at link time if the symbol ends up in
585 // the same linkage unit.
586 if (unsigned(Fixup.Kind) == X86::reloc_riprel_4byte_movq_load)
587 Type = RIT_X86_64_GOTLoad;
589 Type = RIT_X86_64_GOT;
590 } else if (Modifier != MCSymbolRefExpr::VK_None)
591 llvm_report_error("unsupported symbol modifier in relocation");
593 Type = RIT_X86_64_Signed;
595 if (Modifier != MCSymbolRefExpr::VK_None)
596 llvm_report_error("unsupported symbol modifier in branch "
599 Type = RIT_X86_64_Branch;
602 // The Darwin x86_64 relocation format has a problem where it cannot
603 // encode an address (L<foo> + <constant>) which is outside the atom
604 // containing L<foo>. Generally, this shouldn't occur but it does happen
605 // when we have a RIPrel instruction with data following the relocation
606 // entry (e.g., movb $012, L0(%rip)). Even with the PCrel adjustment
607 // Darwin x86_64 uses, the offset is still negative and the linker has
608 // no way to recognize this.
610 // To work around this, Darwin uses several special relocation types to
611 // indicate the offsets. However, the specification or implementation of
612 // these seems to also be incomplete; they should adjust the addend as
613 // well based on the actual encoded instruction (the additional bias),
614 // but instead appear to just look at the final offset.
616 switch (-(Target.getConstant() + (1 << Log2Size))) {
617 case 1: Type = RIT_X86_64_Signed1; break;
618 case 2: Type = RIT_X86_64_Signed2; break;
619 case 4: Type = RIT_X86_64_Signed4; break;
623 if (Modifier == MCSymbolRefExpr::VK_GOT)
624 Type = RIT_X86_64_GOT;
625 else if (Modifier != MCSymbolRefExpr::VK_None)
626 llvm_report_error("unsupported symbol modifier in relocation");
628 Type = RIT_X86_64_Unsigned;
632 // x86_64 always writes custom values into the fixups.
635 // struct relocation_info (8 bytes)
636 MachRelocationEntry MRE;
638 MRE.Word1 = ((Index << 0) |
643 Relocations[Fragment->getParent()].push_back(MRE);
646 void RecordScatteredRelocation(const MCAssembler &Asm,
647 const MCAsmLayout &Layout,
648 const MCFragment *Fragment,
649 const MCAsmFixup &Fixup, MCValue Target,
650 uint64_t &FixedValue) {
651 uint32_t Address = Layout.getFragmentOffset(Fragment) + Fixup.Offset;
652 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
653 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
654 unsigned Type = RIT_Vanilla;
657 const MCSymbol *A = &Target.getSymA()->getSymbol();
658 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
660 if (!A_SD->getFragment())
661 llvm_report_error("symbol '" + A->getName() +
662 "' can not be undefined in a subtraction expression");
664 uint32_t Value = Layout.getSymbolAddress(A_SD);
667 if (const MCSymbolRefExpr *B = Target.getSymB()) {
668 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
670 if (!B_SD->getFragment())
671 llvm_report_error("symbol '" + B->getSymbol().getName() +
672 "' can not be undefined in a subtraction expression");
674 // Select the appropriate difference relocation type.
676 // Note that there is no longer any semantic difference between these two
677 // relocation types from the linkers point of view, this is done solely
678 // for pedantic compatibility with 'as'.
679 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
680 Value2 = Layout.getSymbolAddress(B_SD);
683 // Relocations are written out in reverse order, so the PAIR comes first.
684 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
685 MachRelocationEntry MRE;
686 MRE.Word0 = ((0 << 0) |
692 Relocations[Fragment->getParent()].push_back(MRE);
695 MachRelocationEntry MRE;
696 MRE.Word0 = ((Address << 0) |
702 Relocations[Fragment->getParent()].push_back(MRE);
705 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
706 const MCFragment *Fragment, const MCAsmFixup &Fixup,
707 MCValue Target, uint64_t &FixedValue) {
709 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
713 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
714 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
716 // If this is a difference or a defined symbol plus an offset, then we need
717 // a scattered relocation entry.
718 uint32_t Offset = Target.getConstant();
720 Offset += 1 << Log2Size;
721 if (Target.getSymB() ||
722 (Target.getSymA() && !Target.getSymA()->getSymbol().isUndefined() &&
724 RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,Target,FixedValue);
729 uint32_t Address = Layout.getFragmentOffset(Fragment) + Fixup.Offset;
732 unsigned IsExtern = 0;
735 if (Target.isAbsolute()) { // constant
736 // SymbolNum of 0 indicates the absolute section.
738 // FIXME: Currently, these are never generated (see code below). I cannot
739 // find a case where they are actually emitted.
743 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
744 MCSymbolData *SD = &Asm.getSymbolData(*Symbol);
746 if (Symbol->isUndefined()) {
748 Index = SD->getIndex();
751 // The index is the section ordinal.
755 MCAssembler::const_iterator it = Asm.begin(), ie = Asm.end();
756 for (; it != ie; ++it, ++Index)
757 if (&*it == SD->getFragment()->getParent())
759 assert(it != ie && "Unable to find section index!");
760 Value = Layout.getSymbolAddress(SD);
766 // struct relocation_info (8 bytes)
767 MachRelocationEntry MRE;
769 MRE.Word1 = ((Index << 0) |
774 Relocations[Fragment->getParent()].push_back(MRE);
777 void BindIndirectSymbols(MCAssembler &Asm) {
778 // This is the point where 'as' creates actual symbols for indirect symbols
779 // (in the following two passes). It would be easier for us to do this
780 // sooner when we see the attribute, but that makes getting the order in the
781 // symbol table much more complicated than it is worth.
783 // FIXME: Revisit this when the dust settles.
785 // Bind non lazy symbol pointers first.
786 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
787 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
788 // FIXME: cast<> support!
789 const MCSectionMachO &Section =
790 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
792 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
795 Asm.getOrCreateSymbolData(*it->Symbol);
798 // Then lazy symbol pointers and symbol stubs.
799 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
800 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
801 // FIXME: cast<> support!
802 const MCSectionMachO &Section =
803 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
805 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
806 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
809 // Set the symbol type to undefined lazy, but only on construction.
811 // FIXME: Do not hardcode.
813 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
815 Entry.setFlags(Entry.getFlags() | 0x0001);
819 /// ComputeSymbolTable - Compute the symbol table data
821 /// \param StringTable [out] - The string table data.
822 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
824 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
825 std::vector<MachSymbolData> &LocalSymbolData,
826 std::vector<MachSymbolData> &ExternalSymbolData,
827 std::vector<MachSymbolData> &UndefinedSymbolData) {
828 // Build section lookup table.
829 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
831 for (MCAssembler::iterator it = Asm.begin(),
832 ie = Asm.end(); it != ie; ++it, ++Index)
833 SectionIndexMap[&it->getSection()] = Index;
834 assert(Index <= 256 && "Too many sections!");
836 // Index 0 is always the empty string.
837 StringMap<uint64_t> StringIndexMap;
838 StringTable += '\x00';
840 // Build the symbol arrays and the string table, but only for non-local
843 // The particular order that we collect the symbols and create the string
844 // table, then sort the symbols is chosen to match 'as'. Even though it
845 // doesn't matter for correctness, this is important for letting us diff .o
847 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
848 ie = Asm.symbol_end(); it != ie; ++it) {
849 const MCSymbol &Symbol = it->getSymbol();
851 // Ignore non-linker visible symbols.
852 if (!Asm.isSymbolLinkerVisible(it))
855 if (!it->isExternal() && !Symbol.isUndefined())
858 uint64_t &Entry = StringIndexMap[Symbol.getName()];
860 Entry = StringTable.size();
861 StringTable += Symbol.getName();
862 StringTable += '\x00';
867 MSD.StringIndex = Entry;
869 if (Symbol.isUndefined()) {
870 MSD.SectionIndex = 0;
871 UndefinedSymbolData.push_back(MSD);
872 } else if (Symbol.isAbsolute()) {
873 MSD.SectionIndex = 0;
874 ExternalSymbolData.push_back(MSD);
876 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
877 assert(MSD.SectionIndex && "Invalid section index!");
878 ExternalSymbolData.push_back(MSD);
882 // Now add the data for local symbols.
883 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
884 ie = Asm.symbol_end(); it != ie; ++it) {
885 const MCSymbol &Symbol = it->getSymbol();
887 // Ignore non-linker visible symbols.
888 if (!Asm.isSymbolLinkerVisible(it))
891 if (it->isExternal() || Symbol.isUndefined())
894 uint64_t &Entry = StringIndexMap[Symbol.getName()];
896 Entry = StringTable.size();
897 StringTable += Symbol.getName();
898 StringTable += '\x00';
903 MSD.StringIndex = Entry;
905 if (Symbol.isAbsolute()) {
906 MSD.SectionIndex = 0;
907 LocalSymbolData.push_back(MSD);
909 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
910 assert(MSD.SectionIndex && "Invalid section index!");
911 LocalSymbolData.push_back(MSD);
915 // External and undefined symbols are required to be in lexicographic order.
916 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
917 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
919 // Set the symbol indices.
921 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
922 LocalSymbolData[i].SymbolData->setIndex(Index++);
923 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
924 ExternalSymbolData[i].SymbolData->setIndex(Index++);
925 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
926 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
928 // The string table is padded to a multiple of 4.
929 while (StringTable.size() % 4)
930 StringTable += '\x00';
933 void ExecutePostLayoutBinding(MCAssembler &Asm) {
934 // Create symbol data for any indirect symbols.
935 BindIndirectSymbols(Asm);
937 // Compute symbol table information and bind symbol indices.
938 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
939 UndefinedSymbolData);
942 void WriteObject(const MCAssembler &Asm, const MCAsmLayout &Layout) {
943 unsigned NumSections = Asm.size();
945 // The section data starts after the header, the segment load command (and
946 // section headers) and the symbol table.
947 unsigned NumLoadCommands = 1;
948 uint64_t LoadCommandsSize = Is64Bit ?
949 SegmentLoadCommand64Size + NumSections * Section64Size :
950 SegmentLoadCommand32Size + NumSections * Section32Size;
952 // Add the symbol table load command sizes, if used.
953 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
954 UndefinedSymbolData.size();
956 NumLoadCommands += 2;
957 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
960 // Compute the total size of the section data, as well as its file size and
962 uint64_t SectionDataStart = (Is64Bit ? Header64Size : Header32Size)
964 uint64_t SectionDataSize = 0;
965 uint64_t SectionDataFileSize = 0;
967 for (MCAssembler::const_iterator it = Asm.begin(),
968 ie = Asm.end(); it != ie; ++it) {
969 const MCSectionData &SD = *it;
970 uint64_t Address = Layout.getSectionAddress(&SD);
971 uint64_t Size = Layout.getSectionSize(&SD);
972 uint64_t FileSize = Layout.getSectionFileSize(&SD);
974 VMSize = std::max(VMSize, Address + Size);
976 if (Asm.getBackend().isVirtualSection(SD.getSection()))
979 SectionDataSize = std::max(SectionDataSize, Address + Size);
980 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
983 // The section data is padded to 4 bytes.
985 // FIXME: Is this machine dependent?
986 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
987 SectionDataFileSize += SectionDataPadding;
989 // Write the prolog, starting with the header and load command...
990 WriteHeader(NumLoadCommands, LoadCommandsSize,
991 Asm.getSubsectionsViaSymbols());
992 WriteSegmentLoadCommand(NumSections, VMSize,
993 SectionDataStart, SectionDataSize);
995 // ... and then the section headers.
996 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
997 for (MCAssembler::const_iterator it = Asm.begin(),
998 ie = Asm.end(); it != ie; ++it) {
999 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1000 unsigned NumRelocs = Relocs.size();
1001 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
1002 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1003 RelocTableEnd += NumRelocs * RelocationInfoSize;
1006 // Write the symbol table load command, if used.
1008 unsigned FirstLocalSymbol = 0;
1009 unsigned NumLocalSymbols = LocalSymbolData.size();
1010 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1011 unsigned NumExternalSymbols = ExternalSymbolData.size();
1012 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1013 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1014 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1015 unsigned NumSymTabSymbols =
1016 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1017 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1018 uint64_t IndirectSymbolOffset = 0;
1020 // If used, the indirect symbols are written after the section data.
1021 if (NumIndirectSymbols)
1022 IndirectSymbolOffset = RelocTableEnd;
1024 // The symbol table is written after the indirect symbol data.
1025 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1027 // The string table is written after symbol table.
1028 uint64_t StringTableOffset =
1029 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? Nlist64Size :
1031 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1032 StringTableOffset, StringTable.size());
1034 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1035 FirstExternalSymbol, NumExternalSymbols,
1036 FirstUndefinedSymbol, NumUndefinedSymbols,
1037 IndirectSymbolOffset, NumIndirectSymbols);
1040 // Write the actual section data.
1041 for (MCAssembler::const_iterator it = Asm.begin(),
1042 ie = Asm.end(); it != ie; ++it)
1043 Asm.WriteSectionData(it, Layout, Writer);
1045 // Write the extra padding.
1046 WriteZeros(SectionDataPadding);
1048 // Write the relocation entries.
1049 for (MCAssembler::const_iterator it = Asm.begin(),
1050 ie = Asm.end(); it != ie; ++it) {
1051 // Write the section relocation entries, in reverse order to match 'as'
1052 // (approximately, the exact algorithm is more complicated than this).
1053 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1054 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1055 Write32(Relocs[e - i - 1].Word0);
1056 Write32(Relocs[e - i - 1].Word1);
1060 // Write the symbol table data, if used.
1062 // Write the indirect symbol entries.
1063 for (MCAssembler::const_indirect_symbol_iterator
1064 it = Asm.indirect_symbol_begin(),
1065 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1066 // Indirect symbols in the non lazy symbol pointer section have some
1067 // special handling.
1068 const MCSectionMachO &Section =
1069 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1070 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1071 // If this symbol is defined and internal, mark it as such.
1072 if (it->Symbol->isDefined() &&
1073 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1074 uint32_t Flags = ISF_Local;
1075 if (it->Symbol->isAbsolute())
1076 Flags |= ISF_Absolute;
1082 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1085 // FIXME: Check that offsets match computed ones.
1087 // Write the symbol table entries.
1088 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1089 WriteNlist(LocalSymbolData[i], Layout);
1090 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1091 WriteNlist(ExternalSymbolData[i], Layout);
1092 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1093 WriteNlist(UndefinedSymbolData[i], Layout);
1095 // Write the string table.
1096 OS << StringTable.str();
1103 MachObjectWriter::MachObjectWriter(raw_ostream &OS,
1105 bool IsLittleEndian)
1106 : MCObjectWriter(OS, IsLittleEndian)
1108 Impl = new MachObjectWriterImpl(this, Is64Bit);
1111 MachObjectWriter::~MachObjectWriter() {
1112 delete (MachObjectWriterImpl*) Impl;
1115 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm) {
1116 ((MachObjectWriterImpl*) Impl)->ExecutePostLayoutBinding(Asm);
1119 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1120 const MCAsmLayout &Layout,
1121 const MCFragment *Fragment,
1122 const MCAsmFixup &Fixup, MCValue Target,
1123 uint64_t &FixedValue) {
1124 ((MachObjectWriterImpl*) Impl)->RecordRelocation(Asm, Layout, Fragment, Fixup,
1125 Target, FixedValue);
1128 void MachObjectWriter::WriteObject(const MCAssembler &Asm,
1129 const MCAsmLayout &Layout) {
1130 ((MachObjectWriterImpl*) Impl)->WriteObject(Asm, Layout);