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 // The offset is unused for virtual sections.
279 if (Asm.getBackend().isVirtualSection(SD.getSection())) {
280 assert(SD.getFileSize() == 0 && "Invalid file size!");
284 // struct section (68 bytes) or
285 // struct section_64 (80 bytes)
287 uint64_t Start = OS.tell();
290 // FIXME: cast<> support!
291 const MCSectionMachO &Section =
292 static_cast<const MCSectionMachO&>(SD.getSection());
293 WriteBytes(Section.getSectionName(), 16);
294 WriteBytes(Section.getSegmentName(), 16);
296 Write64(Layout.getSectionAddress(&SD)); // address
297 Write64(SD.getSize()); // size
299 Write32(Layout.getSectionAddress(&SD)); // address
300 Write32(SD.getSize()); // size
304 unsigned Flags = Section.getTypeAndAttributes();
305 if (SD.hasInstructions())
306 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
308 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
309 Write32(Log2_32(SD.getAlignment()));
310 Write32(NumRelocations ? RelocationsStart : 0);
311 Write32(NumRelocations);
313 Write32(0); // reserved1
314 Write32(Section.getStubSize()); // reserved2
316 Write32(0); // reserved3
318 assert(OS.tell() - Start == Is64Bit ? Section64Size : Section32Size);
321 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
322 uint32_t StringTableOffset,
323 uint32_t StringTableSize) {
324 // struct symtab_command (24 bytes)
326 uint64_t Start = OS.tell();
330 Write32(SymtabLoadCommandSize);
331 Write32(SymbolOffset);
333 Write32(StringTableOffset);
334 Write32(StringTableSize);
336 assert(OS.tell() - Start == SymtabLoadCommandSize);
339 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
340 uint32_t NumLocalSymbols,
341 uint32_t FirstExternalSymbol,
342 uint32_t NumExternalSymbols,
343 uint32_t FirstUndefinedSymbol,
344 uint32_t NumUndefinedSymbols,
345 uint32_t IndirectSymbolOffset,
346 uint32_t NumIndirectSymbols) {
347 // struct dysymtab_command (80 bytes)
349 uint64_t Start = OS.tell();
352 Write32(LCT_Dysymtab);
353 Write32(DysymtabLoadCommandSize);
354 Write32(FirstLocalSymbol);
355 Write32(NumLocalSymbols);
356 Write32(FirstExternalSymbol);
357 Write32(NumExternalSymbols);
358 Write32(FirstUndefinedSymbol);
359 Write32(NumUndefinedSymbols);
360 Write32(0); // tocoff
362 Write32(0); // modtaboff
363 Write32(0); // nmodtab
364 Write32(0); // extrefsymoff
365 Write32(0); // nextrefsyms
366 Write32(IndirectSymbolOffset);
367 Write32(NumIndirectSymbols);
368 Write32(0); // extreloff
369 Write32(0); // nextrel
370 Write32(0); // locreloff
371 Write32(0); // nlocrel
373 assert(OS.tell() - Start == DysymtabLoadCommandSize);
376 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
377 MCSymbolData &Data = *MSD.SymbolData;
378 const MCSymbol &Symbol = Data.getSymbol();
380 uint16_t Flags = Data.getFlags();
381 uint32_t Address = 0;
383 // Set the N_TYPE bits. See <mach-o/nlist.h>.
385 // FIXME: Are the prebound or indirect fields possible here?
386 if (Symbol.isUndefined())
387 Type = STT_Undefined;
388 else if (Symbol.isAbsolute())
393 // FIXME: Set STAB bits.
395 if (Data.isPrivateExtern())
396 Type |= STF_PrivateExtern;
399 if (Data.isExternal() || Symbol.isUndefined())
400 Type |= STF_External;
402 // Compute the symbol address.
403 if (Symbol.isDefined()) {
404 if (Symbol.isAbsolute()) {
405 llvm_unreachable("FIXME: Not yet implemented!");
407 Address = Layout.getSymbolAddress(&Data);
409 } else if (Data.isCommon()) {
410 // Common symbols are encoded with the size in the address
411 // field, and their alignment in the flags.
412 Address = Data.getCommonSize();
414 // Common alignment is packed into the 'desc' bits.
415 if (unsigned Align = Data.getCommonAlignment()) {
416 unsigned Log2Size = Log2_32(Align);
417 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
419 llvm_report_error("invalid 'common' alignment '" +
421 // FIXME: Keep this mask with the SymbolFlags enumeration.
422 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
426 // struct nlist (12 bytes)
428 Write32(MSD.StringIndex);
430 Write8(MSD.SectionIndex);
432 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
441 // FIXME: We really need to improve the relocation validation. Basically, we
442 // want to implement a separate computation which evaluates the relocation
443 // entry as the linker would, and verifies that the resultant fixup value is
444 // exactly what the encoder wanted. This will catch several classes of
447 // - Relocation entry bugs, the two algorithms are unlikely to have the same
450 // - Relaxation issues, where we forget to relax something.
452 // - Input errors, where something cannot be correctly encoded. 'as' allows
453 // these through in many cases.
455 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
456 const MCFragment *Fragment,
457 const MCAsmFixup &Fixup, MCValue Target,
458 uint64_t &FixedValue) {
459 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
460 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.Kind);
461 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
464 uint32_t Address = Fragment->getOffset() + Fixup.Offset;
467 unsigned IsExtern = 0;
470 Value = Target.getConstant();
473 // Compensate for the relocation offset, Darwin x86_64 relocations only
474 // have the addend and appear to have attempted to define it to be the
475 // actual expression addend without the PCrel bias. However, instructions
476 // with data following the relocation are not accomodated for (see comment
477 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
478 Value += 1 << Log2Size;
481 if (Target.isAbsolute()) { // constant
482 // SymbolNum of 0 indicates the absolute section.
483 Type = RIT_X86_64_Unsigned;
486 // FIXME: I believe this is broken, I don't think the linker can
487 // understand it. I think it would require a local relocation, but I'm not
488 // sure if that would work either. The official way to get an absolute
489 // PCrel relocation is to use an absolute symbol (which we don't support
493 Type = RIT_X86_64_Branch;
495 } else if (Target.getSymB()) { // A - B + constant
496 const MCSymbol *A = &Target.getSymA()->getSymbol();
497 MCSymbolData &A_SD = Asm.getSymbolData(*A);
498 const MCSymbolData *A_Base = Asm.getAtom(Layout, &A_SD);
500 const MCSymbol *B = &Target.getSymB()->getSymbol();
501 MCSymbolData &B_SD = Asm.getSymbolData(*B);
502 const MCSymbolData *B_Base = Asm.getAtom(Layout, &B_SD);
504 // Neither symbol can be modified.
505 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
506 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
507 llvm_report_error("unsupported relocation of modified symbol");
509 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
510 // implement most of these correctly.
512 llvm_report_error("unsupported pc-relative relocation of difference");
514 // We don't currently support any situation where one or both of the
515 // symbols would require a local relocation. This is almost certainly
516 // unused and may not be possible to encode correctly.
517 if (!A_Base || !B_Base)
518 llvm_report_error("unsupported local relocations in difference");
520 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
521 // a single SIGNED relocation); reject it for now.
522 if (A_Base == B_Base)
523 llvm_report_error("unsupported relocation with identical base");
525 Value += Layout.getSymbolAddress(&A_SD) - Layout.getSymbolAddress(A_Base);
526 Value -= Layout.getSymbolAddress(&B_SD) - Layout.getSymbolAddress(B_Base);
528 Index = A_Base->getIndex();
530 Type = RIT_X86_64_Unsigned;
532 MachRelocationEntry MRE;
534 MRE.Word1 = ((Index << 0) |
539 Relocations[Fragment->getParent()].push_back(MRE);
541 Index = B_Base->getIndex();
543 Type = RIT_X86_64_Subtractor;
545 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
546 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
547 const MCSymbolData *Base = Asm.getAtom(Layout, &SD);
549 // x86_64 almost always uses external relocations, except when there is no
550 // symbol to use as a base address (a local symbol with no preceeding
551 // non-local symbol).
553 Index = Base->getIndex();
556 // Add the local offset, if needed.
558 Value += Layout.getSymbolAddress(&SD) - Layout.getSymbolAddress(Base);
560 // The index is the section ordinal.
564 MCAssembler::const_iterator it = Asm.begin(), ie = Asm.end();
565 for (; it != ie; ++it, ++Index)
566 if (&*it == SD.getFragment()->getParent())
568 assert(it != ie && "Unable to find section index!");
570 Value += Layout.getSymbolAddress(&SD);
573 Value -= Address + (1 << Log2Size);
576 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
579 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
580 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
581 // rewrite the movq to an leaq at link time if the symbol ends up in
582 // the same linkage unit.
583 if (unsigned(Fixup.Kind) == X86::reloc_riprel_4byte_movq_load)
584 Type = RIT_X86_64_GOTLoad;
586 Type = RIT_X86_64_GOT;
587 } else if (Modifier != MCSymbolRefExpr::VK_None)
588 llvm_report_error("unsupported symbol modifier in relocation");
590 Type = RIT_X86_64_Signed;
592 if (Modifier != MCSymbolRefExpr::VK_None)
593 llvm_report_error("unsupported symbol modifier in branch "
596 Type = RIT_X86_64_Branch;
599 // The Darwin x86_64 relocation format has a problem where it cannot
600 // encode an address (L<foo> + <constant>) which is outside the atom
601 // containing L<foo>. Generally, this shouldn't occur but it does happen
602 // when we have a RIPrel instruction with data following the relocation
603 // entry (e.g., movb $012, L0(%rip)). Even with the PCrel adjustment
604 // Darwin x86_64 uses, the offset is still negative and the linker has
605 // no way to recognize this.
607 // To work around this, Darwin uses several special relocation types to
608 // indicate the offsets. However, the specification or implementation of
609 // these seems to also be incomplete; they should adjust the addend as
610 // well based on the actual encoded instruction (the additional bias),
611 // but instead appear to just look at the final offset.
613 switch (-(Target.getConstant() + (1 << Log2Size))) {
614 case 1: Type = RIT_X86_64_Signed1; break;
615 case 2: Type = RIT_X86_64_Signed2; break;
616 case 4: Type = RIT_X86_64_Signed4; break;
620 if (Modifier == MCSymbolRefExpr::VK_GOT)
621 Type = RIT_X86_64_GOT;
622 else if (Modifier != MCSymbolRefExpr::VK_None)
623 llvm_report_error("unsupported symbol modifier in relocation");
625 Type = RIT_X86_64_Unsigned;
629 // x86_64 always writes custom values into the fixups.
632 // struct relocation_info (8 bytes)
633 MachRelocationEntry MRE;
635 MRE.Word1 = ((Index << 0) |
640 Relocations[Fragment->getParent()].push_back(MRE);
643 void RecordScatteredRelocation(const MCAssembler &Asm,
644 const MCAsmLayout &Layout,
645 const MCFragment *Fragment,
646 const MCAsmFixup &Fixup, MCValue Target,
647 uint64_t &FixedValue) {
648 uint32_t Address = Fragment->getOffset() + Fixup.Offset;
649 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
650 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
651 unsigned Type = RIT_Vanilla;
654 const MCSymbol *A = &Target.getSymA()->getSymbol();
655 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
657 if (!A_SD->getFragment())
658 llvm_report_error("symbol '" + A->getName() +
659 "' can not be undefined in a subtraction expression");
661 uint32_t Value = Layout.getSymbolAddress(A_SD);
664 if (const MCSymbolRefExpr *B = Target.getSymB()) {
665 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
667 if (!B_SD->getFragment())
668 llvm_report_error("symbol '" + B->getSymbol().getName() +
669 "' can not be undefined in a subtraction expression");
671 // Select the appropriate difference relocation type.
673 // Note that there is no longer any semantic difference between these two
674 // relocation types from the linkers point of view, this is done solely
675 // for pedantic compatibility with 'as'.
676 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
677 Value2 = Layout.getSymbolAddress(B_SD);
680 // Relocations are written out in reverse order, so the PAIR comes first.
681 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
682 MachRelocationEntry MRE;
683 MRE.Word0 = ((0 << 0) |
689 Relocations[Fragment->getParent()].push_back(MRE);
692 MachRelocationEntry MRE;
693 MRE.Word0 = ((Address << 0) |
699 Relocations[Fragment->getParent()].push_back(MRE);
702 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
703 const MCFragment *Fragment, const MCAsmFixup &Fixup,
704 MCValue Target, uint64_t &FixedValue) {
706 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
710 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
711 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
713 // If this is a difference or a defined symbol plus an offset, then we need
714 // a scattered relocation entry.
715 uint32_t Offset = Target.getConstant();
717 Offset += 1 << Log2Size;
718 if (Target.getSymB() ||
719 (Target.getSymA() && !Target.getSymA()->getSymbol().isUndefined() &&
721 RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,Target,FixedValue);
726 uint32_t Address = Fragment->getOffset() + Fixup.Offset;
729 unsigned IsExtern = 0;
732 if (Target.isAbsolute()) { // constant
733 // SymbolNum of 0 indicates the absolute section.
735 // FIXME: Currently, these are never generated (see code below). I cannot
736 // find a case where they are actually emitted.
740 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
741 MCSymbolData *SD = &Asm.getSymbolData(*Symbol);
743 if (Symbol->isUndefined()) {
745 Index = SD->getIndex();
748 // The index is the section ordinal.
752 MCAssembler::const_iterator it = Asm.begin(), ie = Asm.end();
753 for (; it != ie; ++it, ++Index)
754 if (&*it == SD->getFragment()->getParent())
756 assert(it != ie && "Unable to find section index!");
757 Value = Layout.getSymbolAddress(SD);
763 // struct relocation_info (8 bytes)
764 MachRelocationEntry MRE;
766 MRE.Word1 = ((Index << 0) |
771 Relocations[Fragment->getParent()].push_back(MRE);
774 void BindIndirectSymbols(MCAssembler &Asm) {
775 // This is the point where 'as' creates actual symbols for indirect symbols
776 // (in the following two passes). It would be easier for us to do this
777 // sooner when we see the attribute, but that makes getting the order in the
778 // symbol table much more complicated than it is worth.
780 // FIXME: Revisit this when the dust settles.
782 // Bind non lazy symbol pointers first.
783 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
784 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
785 // FIXME: cast<> support!
786 const MCSectionMachO &Section =
787 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
789 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
792 Asm.getOrCreateSymbolData(*it->Symbol);
795 // Then lazy symbol pointers and symbol stubs.
796 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
797 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
798 // FIXME: cast<> support!
799 const MCSectionMachO &Section =
800 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
802 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
803 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
806 // Set the symbol type to undefined lazy, but only on construction.
808 // FIXME: Do not hardcode.
810 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
812 Entry.setFlags(Entry.getFlags() | 0x0001);
816 /// ComputeSymbolTable - Compute the symbol table data
818 /// \param StringTable [out] - The string table data.
819 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
821 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
822 std::vector<MachSymbolData> &LocalSymbolData,
823 std::vector<MachSymbolData> &ExternalSymbolData,
824 std::vector<MachSymbolData> &UndefinedSymbolData) {
825 // Build section lookup table.
826 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
828 for (MCAssembler::iterator it = Asm.begin(),
829 ie = Asm.end(); it != ie; ++it, ++Index)
830 SectionIndexMap[&it->getSection()] = Index;
831 assert(Index <= 256 && "Too many sections!");
833 // Index 0 is always the empty string.
834 StringMap<uint64_t> StringIndexMap;
835 StringTable += '\x00';
837 // Build the symbol arrays and the string table, but only for non-local
840 // The particular order that we collect the symbols and create the string
841 // table, then sort the symbols is chosen to match 'as'. Even though it
842 // doesn't matter for correctness, this is important for letting us diff .o
844 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
845 ie = Asm.symbol_end(); it != ie; ++it) {
846 const MCSymbol &Symbol = it->getSymbol();
848 // Ignore non-linker visible symbols.
849 if (!Asm.isSymbolLinkerVisible(it))
852 if (!it->isExternal() && !Symbol.isUndefined())
855 uint64_t &Entry = StringIndexMap[Symbol.getName()];
857 Entry = StringTable.size();
858 StringTable += Symbol.getName();
859 StringTable += '\x00';
864 MSD.StringIndex = Entry;
866 if (Symbol.isUndefined()) {
867 MSD.SectionIndex = 0;
868 UndefinedSymbolData.push_back(MSD);
869 } else if (Symbol.isAbsolute()) {
870 MSD.SectionIndex = 0;
871 ExternalSymbolData.push_back(MSD);
873 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
874 assert(MSD.SectionIndex && "Invalid section index!");
875 ExternalSymbolData.push_back(MSD);
879 // Now add the data for local symbols.
880 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
881 ie = Asm.symbol_end(); it != ie; ++it) {
882 const MCSymbol &Symbol = it->getSymbol();
884 // Ignore non-linker visible symbols.
885 if (!Asm.isSymbolLinkerVisible(it))
888 if (it->isExternal() || Symbol.isUndefined())
891 uint64_t &Entry = StringIndexMap[Symbol.getName()];
893 Entry = StringTable.size();
894 StringTable += Symbol.getName();
895 StringTable += '\x00';
900 MSD.StringIndex = Entry;
902 if (Symbol.isAbsolute()) {
903 MSD.SectionIndex = 0;
904 LocalSymbolData.push_back(MSD);
906 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
907 assert(MSD.SectionIndex && "Invalid section index!");
908 LocalSymbolData.push_back(MSD);
912 // External and undefined symbols are required to be in lexicographic order.
913 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
914 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
916 // Set the symbol indices.
918 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
919 LocalSymbolData[i].SymbolData->setIndex(Index++);
920 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
921 ExternalSymbolData[i].SymbolData->setIndex(Index++);
922 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
923 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
925 // The string table is padded to a multiple of 4.
926 while (StringTable.size() % 4)
927 StringTable += '\x00';
930 void ExecutePostLayoutBinding(MCAssembler &Asm) {
931 // Create symbol data for any indirect symbols.
932 BindIndirectSymbols(Asm);
934 // Compute symbol table information and bind symbol indices.
935 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
936 UndefinedSymbolData);
939 void WriteObject(const MCAssembler &Asm, const MCAsmLayout &Layout) {
940 unsigned NumSections = Asm.size();
942 // The section data starts after the header, the segment load command (and
943 // section headers) and the symbol table.
944 unsigned NumLoadCommands = 1;
945 uint64_t LoadCommandsSize = Is64Bit ?
946 SegmentLoadCommand64Size + NumSections * Section64Size :
947 SegmentLoadCommand32Size + NumSections * Section32Size;
949 // Add the symbol table load command sizes, if used.
950 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
951 UndefinedSymbolData.size();
953 NumLoadCommands += 2;
954 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
957 // Compute the total size of the section data, as well as its file size and
959 uint64_t SectionDataStart = (Is64Bit ? Header64Size : Header32Size)
961 uint64_t SectionDataSize = 0;
962 uint64_t SectionDataFileSize = 0;
964 for (MCAssembler::const_iterator it = Asm.begin(),
965 ie = Asm.end(); it != ie; ++it) {
966 const MCSectionData &SD = *it;
967 uint64_t Address = Layout.getSectionAddress(&SD);
969 VMSize = std::max(VMSize, Address + SD.getSize());
971 if (Asm.getBackend().isVirtualSection(SD.getSection()))
974 SectionDataSize = std::max(SectionDataSize, Address + SD.getSize());
975 SectionDataFileSize = std::max(SectionDataFileSize,
976 Address + SD.getFileSize());
979 // The section data is padded to 4 bytes.
981 // FIXME: Is this machine dependent?
982 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
983 SectionDataFileSize += SectionDataPadding;
985 // Write the prolog, starting with the header and load command...
986 WriteHeader(NumLoadCommands, LoadCommandsSize,
987 Asm.getSubsectionsViaSymbols());
988 WriteSegmentLoadCommand(NumSections, VMSize,
989 SectionDataStart, SectionDataSize);
991 // ... and then the section headers.
992 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
993 for (MCAssembler::const_iterator it = Asm.begin(),
994 ie = Asm.end(); it != ie; ++it) {
995 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
996 unsigned NumRelocs = Relocs.size();
997 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
998 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
999 RelocTableEnd += NumRelocs * RelocationInfoSize;
1002 // Write the symbol table load command, if used.
1004 unsigned FirstLocalSymbol = 0;
1005 unsigned NumLocalSymbols = LocalSymbolData.size();
1006 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1007 unsigned NumExternalSymbols = ExternalSymbolData.size();
1008 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1009 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1010 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1011 unsigned NumSymTabSymbols =
1012 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1013 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1014 uint64_t IndirectSymbolOffset = 0;
1016 // If used, the indirect symbols are written after the section data.
1017 if (NumIndirectSymbols)
1018 IndirectSymbolOffset = RelocTableEnd;
1020 // The symbol table is written after the indirect symbol data.
1021 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1023 // The string table is written after symbol table.
1024 uint64_t StringTableOffset =
1025 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? Nlist64Size :
1027 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1028 StringTableOffset, StringTable.size());
1030 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1031 FirstExternalSymbol, NumExternalSymbols,
1032 FirstUndefinedSymbol, NumUndefinedSymbols,
1033 IndirectSymbolOffset, NumIndirectSymbols);
1036 // Write the actual section data.
1037 for (MCAssembler::const_iterator it = Asm.begin(),
1038 ie = Asm.end(); it != ie; ++it)
1039 Asm.WriteSectionData(it, Writer);
1041 // Write the extra padding.
1042 WriteZeros(SectionDataPadding);
1044 // Write the relocation entries.
1045 for (MCAssembler::const_iterator it = Asm.begin(),
1046 ie = Asm.end(); it != ie; ++it) {
1047 // Write the section relocation entries, in reverse order to match 'as'
1048 // (approximately, the exact algorithm is more complicated than this).
1049 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1050 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1051 Write32(Relocs[e - i - 1].Word0);
1052 Write32(Relocs[e - i - 1].Word1);
1056 // Write the symbol table data, if used.
1058 // Write the indirect symbol entries.
1059 for (MCAssembler::const_indirect_symbol_iterator
1060 it = Asm.indirect_symbol_begin(),
1061 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1062 // Indirect symbols in the non lazy symbol pointer section have some
1063 // special handling.
1064 const MCSectionMachO &Section =
1065 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1066 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1067 // If this symbol is defined and internal, mark it as such.
1068 if (it->Symbol->isDefined() &&
1069 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1070 uint32_t Flags = ISF_Local;
1071 if (it->Symbol->isAbsolute())
1072 Flags |= ISF_Absolute;
1078 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1081 // FIXME: Check that offsets match computed ones.
1083 // Write the symbol table entries.
1084 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1085 WriteNlist(LocalSymbolData[i], Layout);
1086 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1087 WriteNlist(ExternalSymbolData[i], Layout);
1088 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1089 WriteNlist(UndefinedSymbolData[i], Layout);
1091 // Write the string table.
1092 OS << StringTable.str();
1099 MachObjectWriter::MachObjectWriter(raw_ostream &OS,
1101 bool IsLittleEndian)
1102 : MCObjectWriter(OS, IsLittleEndian)
1104 Impl = new MachObjectWriterImpl(this, Is64Bit);
1107 MachObjectWriter::~MachObjectWriter() {
1108 delete (MachObjectWriterImpl*) Impl;
1111 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm) {
1112 ((MachObjectWriterImpl*) Impl)->ExecutePostLayoutBinding(Asm);
1115 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1116 const MCAsmLayout &Layout,
1117 const MCFragment *Fragment,
1118 const MCAsmFixup &Fixup, MCValue Target,
1119 uint64_t &FixedValue) {
1120 ((MachObjectWriterImpl*) Impl)->RecordRelocation(Asm, Layout, Fragment, Fixup,
1121 Target, FixedValue);
1124 void MachObjectWriter::WriteObject(const MCAssembler &Asm,
1125 const MCAsmLayout &Layout) {
1126 ((MachObjectWriterImpl*) Impl)->WriteObject(Asm, Layout);