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 // FIXME: this has been copied from (or to) X86AsmBackend.cpp
32 static unsigned getFixupKindLog2Size(unsigned Kind) {
34 default: llvm_unreachable("invalid fixup kind!");
35 case X86::reloc_pcrel_1byte:
36 case FK_Data_1: return 0;
37 case X86::reloc_pcrel_2byte:
38 case FK_Data_2: return 1;
39 case X86::reloc_pcrel_4byte:
40 case X86::reloc_riprel_4byte:
41 case X86::reloc_riprel_4byte_movq_load:
42 case X86::reloc_signed_4byte:
43 case FK_Data_4: return 2;
44 case FK_Data_8: return 3;
48 static bool isFixupKindPCRel(unsigned Kind) {
52 case X86::reloc_pcrel_1byte:
53 case X86::reloc_pcrel_2byte:
54 case X86::reloc_pcrel_4byte:
55 case X86::reloc_riprel_4byte:
56 case X86::reloc_riprel_4byte_movq_load:
61 static bool isFixupKindRIPRel(unsigned Kind) {
62 return Kind == X86::reloc_riprel_4byte ||
63 Kind == X86::reloc_riprel_4byte_movq_load;
66 static bool doesSymbolRequireExternRelocation(MCSymbolData *SD) {
67 // Undefined symbols are always extern.
68 if (SD->Symbol->isUndefined())
71 // References to weak definitions require external relocation entries; the
72 // definition may not always be the one in the same object file.
73 if (SD->getFlags() & SF_WeakDefinition)
76 // Otherwise, we can use an internal relocation.
80 static bool isScatteredFixupFullyResolved(const MCAssembler &Asm,
82 const MCSymbolData *BaseSymbol) {
83 // The effective fixup address is
84 // addr(atom(A)) + offset(A)
85 // - addr(atom(B)) - offset(B)
86 // - addr(BaseSymbol) + <fixup offset from base symbol>
87 // and the offsets are not relocatable, so the fixup is fully resolved when
88 // addr(atom(A)) - addr(atom(B)) - addr(BaseSymbol) == 0.
90 // Note that "false" is almost always conservatively correct (it means we emit
91 // a relocation which is unnecessary), except when it would force us to emit a
92 // relocation which the target cannot encode.
94 const MCSymbolData *A_Base = 0, *B_Base = 0;
95 if (const MCSymbolRefExpr *A = Target.getSymA()) {
96 // Modified symbol references cannot be resolved.
97 if (A->getKind() != MCSymbolRefExpr::VK_None)
100 A_Base = Asm.getAtom(&Asm.getSymbolData(A->getSymbol()));
105 if (const MCSymbolRefExpr *B = Target.getSymB()) {
106 // Modified symbol references cannot be resolved.
107 if (B->getKind() != MCSymbolRefExpr::VK_None)
110 B_Base = Asm.getAtom(&Asm.getSymbolData(B->getSymbol()));
115 // If there is no base, A and B have to be the same atom for this fixup to be
118 return A_Base == B_Base;
120 // Otherwise, B must be missing and A must be the base.
121 return !B_Base && BaseSymbol == A_Base;
124 static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm,
125 const MCValue Target,
126 const MCSection *BaseSection) {
127 // The effective fixup address is
128 // addr(atom(A)) + offset(A)
129 // - addr(atom(B)) - offset(B)
130 // - addr(<base symbol>) + <fixup offset from base symbol>
131 // and the offsets are not relocatable, so the fixup is fully resolved when
132 // addr(atom(A)) - addr(atom(B)) - addr(<base symbol>)) == 0.
134 // The simple (Darwin, except on x86_64) way of dealing with this was to
135 // assume that any reference to a temporary symbol *must* be a temporary
136 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
137 // relocation to a temporary symbol (in the same section) is fully
138 // resolved. This also works in conjunction with absolutized .set, which
139 // requires the compiler to use .set to absolutize the differences between
140 // symbols which the compiler knows to be assembly time constants, so we don't
141 // need to worry about considering symbol differences fully resolved.
143 // Non-relative fixups are only resolved if constant.
145 return Target.isAbsolute();
147 // Otherwise, relative fixups are only resolved if not a difference and the
148 // target is a temporary in the same section.
149 if (Target.isAbsolute() || Target.getSymB())
152 const MCSymbol *A = &Target.getSymA()->getSymbol();
153 if (!A->isTemporary() || !A->isInSection() ||
154 &A->getSection() != BaseSection)
162 class MachObjectWriterImpl {
163 // See <mach-o/loader.h>.
165 Header_Magic32 = 0xFEEDFACE,
166 Header_Magic64 = 0xFEEDFACF
172 SegmentLoadCommand32Size = 56,
173 SegmentLoadCommand64Size = 72,
176 SymtabLoadCommandSize = 24,
177 DysymtabLoadCommandSize = 80,
180 RelocationInfoSize = 8
183 enum HeaderFileType {
188 HF_SubsectionsViaSymbols = 0x2000
191 enum LoadCommandType {
198 // See <mach-o/nlist.h>.
199 enum SymbolTypeType {
200 STT_Undefined = 0x00,
205 enum SymbolTypeFlags {
206 // If any of these bits are set, then the entry is a stab entry number (see
207 // <mach-o/stab.h>. Otherwise the other masks apply.
208 STF_StabsEntryMask = 0xe0,
212 STF_PrivateExtern = 0x10
215 /// IndirectSymbolFlags - Flags for encoding special values in the indirect
217 enum IndirectSymbolFlags {
218 ISF_Local = 0x80000000,
219 ISF_Absolute = 0x40000000
222 /// RelocationFlags - Special flags for addresses.
223 enum RelocationFlags {
224 RF_Scattered = 0x80000000
227 enum RelocationInfoType {
231 RIT_PreboundLazyPointer = 3,
232 RIT_LocalDifference = 4,
236 /// X86_64 uses its own relocation types.
237 enum RelocationInfoTypeX86_64 {
238 RIT_X86_64_Unsigned = 0,
239 RIT_X86_64_Signed = 1,
240 RIT_X86_64_Branch = 2,
241 RIT_X86_64_GOTLoad = 3,
243 RIT_X86_64_Subtractor = 5,
244 RIT_X86_64_Signed1 = 6,
245 RIT_X86_64_Signed2 = 7,
246 RIT_X86_64_Signed4 = 8,
250 /// MachSymbolData - Helper struct for containing some precomputed information
252 struct MachSymbolData {
253 MCSymbolData *SymbolData;
254 uint64_t StringIndex;
255 uint8_t SectionIndex;
257 // Support lexicographic sorting.
258 bool operator<(const MachSymbolData &RHS) const {
259 return SymbolData->getSymbol().getName() <
260 RHS.SymbolData->getSymbol().getName();
264 /// @name Relocation Data
267 struct MachRelocationEntry {
272 llvm::DenseMap<const MCSectionData*,
273 std::vector<MachRelocationEntry> > Relocations;
274 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
277 /// @name Symbol Table Data
280 SmallString<256> StringTable;
281 std::vector<MachSymbolData> LocalSymbolData;
282 std::vector<MachSymbolData> ExternalSymbolData;
283 std::vector<MachSymbolData> UndefinedSymbolData;
287 MachObjectWriter *Writer;
291 unsigned Is64Bit : 1;
294 MachObjectWriterImpl(MachObjectWriter *_Writer, bool _Is64Bit)
295 : Writer(_Writer), OS(Writer->getStream()), Is64Bit(_Is64Bit) {
298 void Write8(uint8_t Value) { Writer->Write8(Value); }
299 void Write16(uint16_t Value) { Writer->Write16(Value); }
300 void Write32(uint32_t Value) { Writer->Write32(Value); }
301 void Write64(uint64_t Value) { Writer->Write64(Value); }
302 void WriteZeros(unsigned N) { Writer->WriteZeros(N); }
303 void WriteBytes(StringRef Str, unsigned ZeroFillSize = 0) {
304 Writer->WriteBytes(Str, ZeroFillSize);
307 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
308 bool SubsectionsViaSymbols) {
311 if (SubsectionsViaSymbols)
312 Flags |= HF_SubsectionsViaSymbols;
314 // struct mach_header (28 bytes) or
315 // struct mach_header_64 (32 bytes)
317 uint64_t Start = OS.tell();
320 Write32(Is64Bit ? Header_Magic64 : Header_Magic32);
322 // FIXME: Support cputype.
323 Write32(Is64Bit ? MachO::CPUTypeX86_64 : MachO::CPUTypeI386);
324 // FIXME: Support cpusubtype.
325 Write32(MachO::CPUSubType_I386_ALL);
327 Write32(NumLoadCommands); // Object files have a single load command, the
329 Write32(LoadCommandsSize);
332 Write32(0); // reserved
334 assert(OS.tell() - Start == Is64Bit ? Header64Size : Header32Size);
337 /// WriteSegmentLoadCommand - Write a segment load command.
339 /// \arg NumSections - The number of sections in this segment.
340 /// \arg SectionDataSize - The total size of the sections.
341 void WriteSegmentLoadCommand(unsigned NumSections,
343 uint64_t SectionDataStartOffset,
344 uint64_t SectionDataSize) {
345 // struct segment_command (56 bytes) or
346 // struct segment_command_64 (72 bytes)
348 uint64_t Start = OS.tell();
351 unsigned SegmentLoadCommandSize = Is64Bit ? SegmentLoadCommand64Size :
352 SegmentLoadCommand32Size;
353 Write32(Is64Bit ? LCT_Segment64 : LCT_Segment);
354 Write32(SegmentLoadCommandSize +
355 NumSections * (Is64Bit ? Section64Size : Section32Size));
359 Write64(0); // vmaddr
360 Write64(VMSize); // vmsize
361 Write64(SectionDataStartOffset); // file offset
362 Write64(SectionDataSize); // file size
364 Write32(0); // vmaddr
365 Write32(VMSize); // vmsize
366 Write32(SectionDataStartOffset); // file offset
367 Write32(SectionDataSize); // file size
369 Write32(0x7); // maxprot
370 Write32(0x7); // initprot
371 Write32(NumSections);
374 assert(OS.tell() - Start == SegmentLoadCommandSize);
377 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
378 const MCSectionData &SD, uint64_t FileOffset,
379 uint64_t RelocationsStart, unsigned NumRelocations) {
380 uint64_t SectionSize = Layout.getSectionSize(&SD);
382 // The offset is unused for virtual sections.
383 if (Asm.getBackend().isVirtualSection(SD.getSection())) {
384 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
388 // struct section (68 bytes) or
389 // struct section_64 (80 bytes)
391 uint64_t Start = OS.tell();
394 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
395 WriteBytes(Section.getSectionName(), 16);
396 WriteBytes(Section.getSegmentName(), 16);
398 Write64(Layout.getSectionAddress(&SD)); // address
399 Write64(SectionSize); // size
401 Write32(Layout.getSectionAddress(&SD)); // address
402 Write32(SectionSize); // size
406 unsigned Flags = Section.getTypeAndAttributes();
407 if (SD.hasInstructions())
408 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
410 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
411 Write32(Log2_32(SD.getAlignment()));
412 Write32(NumRelocations ? RelocationsStart : 0);
413 Write32(NumRelocations);
415 Write32(IndirectSymBase.lookup(&SD)); // reserved1
416 Write32(Section.getStubSize()); // reserved2
418 Write32(0); // reserved3
420 assert(OS.tell() - Start == Is64Bit ? Section64Size : Section32Size);
423 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
424 uint32_t StringTableOffset,
425 uint32_t StringTableSize) {
426 // struct symtab_command (24 bytes)
428 uint64_t Start = OS.tell();
432 Write32(SymtabLoadCommandSize);
433 Write32(SymbolOffset);
435 Write32(StringTableOffset);
436 Write32(StringTableSize);
438 assert(OS.tell() - Start == SymtabLoadCommandSize);
441 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
442 uint32_t NumLocalSymbols,
443 uint32_t FirstExternalSymbol,
444 uint32_t NumExternalSymbols,
445 uint32_t FirstUndefinedSymbol,
446 uint32_t NumUndefinedSymbols,
447 uint32_t IndirectSymbolOffset,
448 uint32_t NumIndirectSymbols) {
449 // struct dysymtab_command (80 bytes)
451 uint64_t Start = OS.tell();
454 Write32(LCT_Dysymtab);
455 Write32(DysymtabLoadCommandSize);
456 Write32(FirstLocalSymbol);
457 Write32(NumLocalSymbols);
458 Write32(FirstExternalSymbol);
459 Write32(NumExternalSymbols);
460 Write32(FirstUndefinedSymbol);
461 Write32(NumUndefinedSymbols);
462 Write32(0); // tocoff
464 Write32(0); // modtaboff
465 Write32(0); // nmodtab
466 Write32(0); // extrefsymoff
467 Write32(0); // nextrefsyms
468 Write32(IndirectSymbolOffset);
469 Write32(NumIndirectSymbols);
470 Write32(0); // extreloff
471 Write32(0); // nextrel
472 Write32(0); // locreloff
473 Write32(0); // nlocrel
475 assert(OS.tell() - Start == DysymtabLoadCommandSize);
478 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
479 MCSymbolData &Data = *MSD.SymbolData;
480 const MCSymbol &Symbol = Data.getSymbol();
482 uint16_t Flags = Data.getFlags();
483 uint32_t Address = 0;
485 // Set the N_TYPE bits. See <mach-o/nlist.h>.
487 // FIXME: Are the prebound or indirect fields possible here?
488 if (Symbol.isUndefined())
489 Type = STT_Undefined;
490 else if (Symbol.isAbsolute())
495 // FIXME: Set STAB bits.
497 if (Data.isPrivateExtern())
498 Type |= STF_PrivateExtern;
501 if (Data.isExternal() || Symbol.isUndefined())
502 Type |= STF_External;
504 // Compute the symbol address.
505 if (Symbol.isDefined()) {
506 if (Symbol.isAbsolute()) {
507 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
509 Address = Layout.getSymbolAddress(&Data);
511 } else if (Data.isCommon()) {
512 // Common symbols are encoded with the size in the address
513 // field, and their alignment in the flags.
514 Address = Data.getCommonSize();
516 // Common alignment is packed into the 'desc' bits.
517 if (unsigned Align = Data.getCommonAlignment()) {
518 unsigned Log2Size = Log2_32(Align);
519 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
521 report_fatal_error("invalid 'common' alignment '" +
523 // FIXME: Keep this mask with the SymbolFlags enumeration.
524 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
528 // struct nlist (12 bytes)
530 Write32(MSD.StringIndex);
532 Write8(MSD.SectionIndex);
534 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
543 // FIXME: We really need to improve the relocation validation. Basically, we
544 // want to implement a separate computation which evaluates the relocation
545 // entry as the linker would, and verifies that the resultant fixup value is
546 // exactly what the encoder wanted. This will catch several classes of
549 // - Relocation entry bugs, the two algorithms are unlikely to have the same
552 // - Relaxation issues, where we forget to relax something.
554 // - Input errors, where something cannot be correctly encoded. 'as' allows
555 // these through in many cases.
557 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
558 const MCFragment *Fragment,
559 const MCFixup &Fixup, MCValue Target,
560 uint64_t &FixedValue) {
561 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
562 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
563 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
566 uint32_t FixupOffset =
567 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
568 uint32_t FixupAddress =
569 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
572 unsigned IsExtern = 0;
575 Value = Target.getConstant();
578 // Compensate for the relocation offset, Darwin x86_64 relocations only
579 // have the addend and appear to have attempted to define it to be the
580 // actual expression addend without the PCrel bias. However, instructions
581 // with data following the relocation are not accomodated for (see comment
582 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
583 Value += 1LL << Log2Size;
586 if (Target.isAbsolute()) { // constant
587 // SymbolNum of 0 indicates the absolute section.
588 Type = RIT_X86_64_Unsigned;
591 // FIXME: I believe this is broken, I don't think the linker can
592 // understand it. I think it would require a local relocation, but I'm not
593 // sure if that would work either. The official way to get an absolute
594 // PCrel relocation is to use an absolute symbol (which we don't support
598 Type = RIT_X86_64_Branch;
600 } else if (Target.getSymB()) { // A - B + constant
601 const MCSymbol *A = &Target.getSymA()->getSymbol();
602 MCSymbolData &A_SD = Asm.getSymbolData(*A);
603 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
605 const MCSymbol *B = &Target.getSymB()->getSymbol();
606 MCSymbolData &B_SD = Asm.getSymbolData(*B);
607 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
609 // Neither symbol can be modified.
610 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
611 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
612 report_fatal_error("unsupported relocation of modified symbol");
614 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
615 // implement most of these correctly.
617 report_fatal_error("unsupported pc-relative relocation of difference");
619 // We don't currently support any situation where one or both of the
620 // symbols would require a local relocation. This is almost certainly
621 // unused and may not be possible to encode correctly.
622 if (!A_Base || !B_Base)
623 report_fatal_error("unsupported local relocations in difference");
625 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
626 // a single SIGNED relocation); reject it for now.
627 if (A_Base == B_Base)
628 report_fatal_error("unsupported relocation with identical base");
630 Value += Layout.getSymbolAddress(&A_SD) - Layout.getSymbolAddress(A_Base);
631 Value -= Layout.getSymbolAddress(&B_SD) - Layout.getSymbolAddress(B_Base);
633 Index = A_Base->getIndex();
635 Type = RIT_X86_64_Unsigned;
637 MachRelocationEntry MRE;
638 MRE.Word0 = FixupOffset;
639 MRE.Word1 = ((Index << 0) |
644 Relocations[Fragment->getParent()].push_back(MRE);
646 Index = B_Base->getIndex();
648 Type = RIT_X86_64_Subtractor;
650 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
651 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
652 const MCSymbolData *Base = Asm.getAtom(&SD);
654 // Relocations inside debug sections always use local relocations when
655 // possible. This seems to be done because the debugger doesn't fully
656 // understand x86_64 relocation entries, and expects to find values that
657 // have already been fixed up.
658 if (Symbol->isInSection()) {
659 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
660 Fragment->getParent()->getSection());
661 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
665 // x86_64 almost always uses external relocations, except when there is no
666 // symbol to use as a base address (a local symbol with no preceeding
667 // non-local symbol).
669 Index = Base->getIndex();
672 // Add the local offset, if needed.
674 Value += Layout.getSymbolAddress(&SD) - Layout.getSymbolAddress(Base);
675 } else if (Symbol->isInSection()) {
676 // The index is the section ordinal (1-based).
677 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
679 Value += Layout.getSymbolAddress(&SD);
682 Value -= FixupAddress + (1 << Log2Size);
684 report_fatal_error("unsupported relocation of undefined symbol '" +
685 Symbol->getName() + "'");
688 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
691 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
692 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
693 // rewrite the movq to an leaq at link time if the symbol ends up in
694 // the same linkage unit.
695 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
696 Type = RIT_X86_64_GOTLoad;
698 Type = RIT_X86_64_GOT;
699 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
700 Type = RIT_X86_64_TLV;
701 } else if (Modifier != MCSymbolRefExpr::VK_None) {
702 report_fatal_error("unsupported symbol modifier in relocation");
704 Type = RIT_X86_64_Signed;
706 // The Darwin x86_64 relocation format has a problem where it cannot
707 // encode an address (L<foo> + <constant>) which is outside the atom
708 // containing L<foo>. Generally, this shouldn't occur but it does
709 // happen when we have a RIPrel instruction with data following the
710 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
711 // adjustment Darwin x86_64 uses, the offset is still negative and
712 // the linker has no way to recognize this.
714 // To work around this, Darwin uses several special relocation types
715 // to indicate the offsets. However, the specification or
716 // implementation of these seems to also be incomplete; they should
717 // adjust the addend as well based on the actual encoded instruction
718 // (the additional bias), but instead appear to just look at the
720 switch (-(Target.getConstant() + (1LL << Log2Size))) {
721 case 1: Type = RIT_X86_64_Signed1; break;
722 case 2: Type = RIT_X86_64_Signed2; break;
723 case 4: Type = RIT_X86_64_Signed4; break;
727 if (Modifier != MCSymbolRefExpr::VK_None)
728 report_fatal_error("unsupported symbol modifier in branch "
731 Type = RIT_X86_64_Branch;
734 if (Modifier == MCSymbolRefExpr::VK_GOT) {
735 Type = RIT_X86_64_GOT;
736 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
737 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
738 // which case all we do is set the PCrel bit in the relocation entry;
739 // this is used with exception handling, for example. The source is
740 // required to include any necessary offset directly.
741 Type = RIT_X86_64_GOT;
743 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
744 report_fatal_error("TLVP symbol modifier should have been rip-rel");
745 } else if (Modifier != MCSymbolRefExpr::VK_None)
746 report_fatal_error("unsupported symbol modifier in relocation");
748 Type = RIT_X86_64_Unsigned;
752 // x86_64 always writes custom values into the fixups.
755 // struct relocation_info (8 bytes)
756 MachRelocationEntry MRE;
757 MRE.Word0 = FixupOffset;
758 MRE.Word1 = ((Index << 0) |
763 Relocations[Fragment->getParent()].push_back(MRE);
766 void RecordScatteredRelocation(const MCAssembler &Asm,
767 const MCAsmLayout &Layout,
768 const MCFragment *Fragment,
769 const MCFixup &Fixup, MCValue Target,
770 uint64_t &FixedValue) {
771 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
772 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
773 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
774 unsigned Type = RIT_Vanilla;
777 const MCSymbol *A = &Target.getSymA()->getSymbol();
778 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
780 if (!A_SD->getFragment())
781 report_fatal_error("symbol '" + A->getName() +
782 "' can not be undefined in a subtraction expression");
784 uint32_t Value = Layout.getSymbolAddress(A_SD);
787 if (const MCSymbolRefExpr *B = Target.getSymB()) {
788 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
790 if (!B_SD->getFragment())
791 report_fatal_error("symbol '" + B->getSymbol().getName() +
792 "' can not be undefined in a subtraction expression");
794 // Select the appropriate difference relocation type.
796 // Note that there is no longer any semantic difference between these two
797 // relocation types from the linkers point of view, this is done solely
798 // for pedantic compatibility with 'as'.
799 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
800 Value2 = Layout.getSymbolAddress(B_SD);
803 // Relocations are written out in reverse order, so the PAIR comes first.
804 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
805 MachRelocationEntry MRE;
806 MRE.Word0 = ((0 << 0) |
812 Relocations[Fragment->getParent()].push_back(MRE);
815 MachRelocationEntry MRE;
816 MRE.Word0 = ((FixupOffset << 0) |
822 Relocations[Fragment->getParent()].push_back(MRE);
825 void RecordTLVPRelocation(const MCAssembler &Asm,
826 const MCAsmLayout &Layout,
827 const MCFragment *Fragment,
828 const MCFixup &Fixup, MCValue Target,
829 uint64_t &FixedValue) {
830 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
832 "Should only be called with a 32-bit TLVP relocation!");
834 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
835 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
836 unsigned IsPCRel = 0;
838 // Get the symbol data.
839 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
840 unsigned Index = SD_A->getIndex();
842 // We're only going to have a second symbol in pic mode and it'll be a
843 // subtraction from the picbase. For 32-bit pic the addend is the difference
844 // between the picbase and the next address. For 32-bit static the addend
846 if (Target.getSymB()) {
847 // If this is a subtraction then we're pcrel.
848 uint32_t FixupAddress =
849 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
850 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
852 FixedValue = (FixupAddress - Layout.getSymbolAddress(SD_B) +
853 Target.getConstant());
854 FixedValue += 1ULL << Log2Size;
859 // struct relocation_info (8 bytes)
860 MachRelocationEntry MRE;
862 MRE.Word1 = ((Index << 0) |
865 (1 << 27) | // Extern
866 (RIT_TLV << 28)); // Type
867 Relocations[Fragment->getParent()].push_back(MRE);
870 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
871 const MCFragment *Fragment, const MCFixup &Fixup,
872 MCValue Target, uint64_t &FixedValue) {
874 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
878 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
879 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
881 // If this is a 32-bit TLVP reloc it's handled a bit differently.
882 if (Target.getSymA() &&
883 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
884 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
888 // If this is a difference or a defined symbol plus an offset, then we need
889 // a scattered relocation entry.
890 // Differences always require scattered relocations.
891 if (Target.getSymB())
892 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
895 // Get the symbol data, if any.
896 MCSymbolData *SD = 0;
897 if (Target.getSymA())
898 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
900 // If this is an internal relocation with an offset, it also needs a
901 // scattered relocation entry.
902 uint32_t Offset = Target.getConstant();
904 Offset += 1 << Log2Size;
905 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
906 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
910 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
912 unsigned IsExtern = 0;
915 if (Target.isAbsolute()) { // constant
916 // SymbolNum of 0 indicates the absolute section.
918 // FIXME: Currently, these are never generated (see code below). I cannot
919 // find a case where they are actually emitted.
922 // Check whether we need an external or internal relocation.
923 if (doesSymbolRequireExternRelocation(SD)) {
925 Index = SD->getIndex();
926 // For external relocations, make sure to offset the fixup value to
927 // compensate for the addend of the symbol address, if it was
928 // undefined. This occurs with weak definitions, for example.
929 if (!SD->Symbol->isUndefined())
930 FixedValue -= Layout.getSymbolAddress(SD);
932 // The index is the section ordinal (1-based).
933 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
939 // struct relocation_info (8 bytes)
940 MachRelocationEntry MRE;
941 MRE.Word0 = FixupOffset;
942 MRE.Word1 = ((Index << 0) |
947 Relocations[Fragment->getParent()].push_back(MRE);
950 void BindIndirectSymbols(MCAssembler &Asm) {
951 // This is the point where 'as' creates actual symbols for indirect symbols
952 // (in the following two passes). It would be easier for us to do this
953 // sooner when we see the attribute, but that makes getting the order in the
954 // symbol table much more complicated than it is worth.
956 // FIXME: Revisit this when the dust settles.
958 // Bind non lazy symbol pointers first.
959 unsigned IndirectIndex = 0;
960 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
961 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
962 const MCSectionMachO &Section =
963 cast<MCSectionMachO>(it->SectionData->getSection());
965 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
968 // Initialize the section indirect symbol base, if necessary.
969 if (!IndirectSymBase.count(it->SectionData))
970 IndirectSymBase[it->SectionData] = IndirectIndex;
972 Asm.getOrCreateSymbolData(*it->Symbol);
975 // Then lazy symbol pointers and symbol stubs.
977 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
978 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
979 const MCSectionMachO &Section =
980 cast<MCSectionMachO>(it->SectionData->getSection());
982 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
983 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
986 // Initialize the section indirect symbol base, if necessary.
987 if (!IndirectSymBase.count(it->SectionData))
988 IndirectSymBase[it->SectionData] = IndirectIndex;
990 // Set the symbol type to undefined lazy, but only on construction.
992 // FIXME: Do not hardcode.
994 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
996 Entry.setFlags(Entry.getFlags() | 0x0001);
1000 /// ComputeSymbolTable - Compute the symbol table data
1002 /// \param StringTable [out] - The string table data.
1003 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
1005 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
1006 std::vector<MachSymbolData> &LocalSymbolData,
1007 std::vector<MachSymbolData> &ExternalSymbolData,
1008 std::vector<MachSymbolData> &UndefinedSymbolData) {
1009 // Build section lookup table.
1010 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
1012 for (MCAssembler::iterator it = Asm.begin(),
1013 ie = Asm.end(); it != ie; ++it, ++Index)
1014 SectionIndexMap[&it->getSection()] = Index;
1015 assert(Index <= 256 && "Too many sections!");
1017 // Index 0 is always the empty string.
1018 StringMap<uint64_t> StringIndexMap;
1019 StringTable += '\x00';
1021 // Build the symbol arrays and the string table, but only for non-local
1024 // The particular order that we collect the symbols and create the string
1025 // table, then sort the symbols is chosen to match 'as'. Even though it
1026 // doesn't matter for correctness, this is important for letting us diff .o
1028 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1029 ie = Asm.symbol_end(); it != ie; ++it) {
1030 const MCSymbol &Symbol = it->getSymbol();
1032 // Ignore non-linker visible symbols.
1033 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1036 if (!it->isExternal() && !Symbol.isUndefined())
1039 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1041 Entry = StringTable.size();
1042 StringTable += Symbol.getName();
1043 StringTable += '\x00';
1047 MSD.SymbolData = it;
1048 MSD.StringIndex = Entry;
1050 if (Symbol.isUndefined()) {
1051 MSD.SectionIndex = 0;
1052 UndefinedSymbolData.push_back(MSD);
1053 } else if (Symbol.isAbsolute()) {
1054 MSD.SectionIndex = 0;
1055 ExternalSymbolData.push_back(MSD);
1057 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1058 assert(MSD.SectionIndex && "Invalid section index!");
1059 ExternalSymbolData.push_back(MSD);
1063 // Now add the data for local symbols.
1064 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1065 ie = Asm.symbol_end(); it != ie; ++it) {
1066 const MCSymbol &Symbol = it->getSymbol();
1068 // Ignore non-linker visible symbols.
1069 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1072 if (it->isExternal() || Symbol.isUndefined())
1075 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1077 Entry = StringTable.size();
1078 StringTable += Symbol.getName();
1079 StringTable += '\x00';
1083 MSD.SymbolData = it;
1084 MSD.StringIndex = Entry;
1086 if (Symbol.isAbsolute()) {
1087 MSD.SectionIndex = 0;
1088 LocalSymbolData.push_back(MSD);
1090 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1091 assert(MSD.SectionIndex && "Invalid section index!");
1092 LocalSymbolData.push_back(MSD);
1096 // External and undefined symbols are required to be in lexicographic order.
1097 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1098 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1100 // Set the symbol indices.
1102 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1103 LocalSymbolData[i].SymbolData->setIndex(Index++);
1104 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1105 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1106 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1107 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1109 // The string table is padded to a multiple of 4.
1110 while (StringTable.size() % 4)
1111 StringTable += '\x00';
1114 void ExecutePostLayoutBinding(MCAssembler &Asm) {
1115 // Create symbol data for any indirect symbols.
1116 BindIndirectSymbols(Asm);
1118 // Compute symbol table information and bind symbol indices.
1119 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1120 UndefinedSymbolData);
1124 bool IsFixupFullyResolved(const MCAssembler &Asm,
1125 const MCValue Target,
1127 const MCFragment *DF) const {
1128 // If we are using scattered symbols, determine whether this value is
1129 // actually resolved; scattering may cause atoms to move.
1130 if (Asm.getBackend().hasScatteredSymbols()) {
1131 if (Asm.getBackend().hasReliableSymbolDifference()) {
1132 // If this is a PCrel relocation, find the base atom (identified by its
1133 // symbol) that the fixup value is relative to.
1134 const MCSymbolData *BaseSymbol = 0;
1136 BaseSymbol = DF->getAtom();
1141 return isScatteredFixupFullyResolved(Asm, Target, BaseSymbol);
1143 const MCSection *BaseSection = 0;
1145 BaseSection = &DF->getParent()->getSection();
1147 return isScatteredFixupFullyResolvedSimple(Asm, Target, BaseSection);
1153 void WriteObject(const MCAssembler &Asm, const MCAsmLayout &Layout) {
1154 unsigned NumSections = Asm.size();
1156 // The section data starts after the header, the segment load command (and
1157 // section headers) and the symbol table.
1158 unsigned NumLoadCommands = 1;
1159 uint64_t LoadCommandsSize = Is64Bit ?
1160 SegmentLoadCommand64Size + NumSections * Section64Size :
1161 SegmentLoadCommand32Size + NumSections * Section32Size;
1163 // Add the symbol table load command sizes, if used.
1164 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1165 UndefinedSymbolData.size();
1167 NumLoadCommands += 2;
1168 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
1171 // Compute the total size of the section data, as well as its file size and
1173 uint64_t SectionDataStart = (Is64Bit ? Header64Size : Header32Size)
1175 uint64_t SectionDataSize = 0;
1176 uint64_t SectionDataFileSize = 0;
1177 uint64_t VMSize = 0;
1178 for (MCAssembler::const_iterator it = Asm.begin(),
1179 ie = Asm.end(); it != ie; ++it) {
1180 const MCSectionData &SD = *it;
1181 uint64_t Address = Layout.getSectionAddress(&SD);
1182 uint64_t Size = Layout.getSectionSize(&SD);
1183 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1185 VMSize = std::max(VMSize, Address + Size);
1187 if (Asm.getBackend().isVirtualSection(SD.getSection()))
1190 SectionDataSize = std::max(SectionDataSize, Address + Size);
1191 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1194 // The section data is padded to 4 bytes.
1196 // FIXME: Is this machine dependent?
1197 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1198 SectionDataFileSize += SectionDataPadding;
1200 // Write the prolog, starting with the header and load command...
1201 WriteHeader(NumLoadCommands, LoadCommandsSize,
1202 Asm.getSubsectionsViaSymbols());
1203 WriteSegmentLoadCommand(NumSections, VMSize,
1204 SectionDataStart, SectionDataSize);
1206 // ... and then the section headers.
1207 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1208 for (MCAssembler::const_iterator it = Asm.begin(),
1209 ie = Asm.end(); it != ie; ++it) {
1210 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1211 unsigned NumRelocs = Relocs.size();
1212 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
1213 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1214 RelocTableEnd += NumRelocs * RelocationInfoSize;
1217 // Write the symbol table load command, if used.
1219 unsigned FirstLocalSymbol = 0;
1220 unsigned NumLocalSymbols = LocalSymbolData.size();
1221 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1222 unsigned NumExternalSymbols = ExternalSymbolData.size();
1223 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1224 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1225 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1226 unsigned NumSymTabSymbols =
1227 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1228 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1229 uint64_t IndirectSymbolOffset = 0;
1231 // If used, the indirect symbols are written after the section data.
1232 if (NumIndirectSymbols)
1233 IndirectSymbolOffset = RelocTableEnd;
1235 // The symbol table is written after the indirect symbol data.
1236 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1238 // The string table is written after symbol table.
1239 uint64_t StringTableOffset =
1240 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? Nlist64Size :
1242 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1243 StringTableOffset, StringTable.size());
1245 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1246 FirstExternalSymbol, NumExternalSymbols,
1247 FirstUndefinedSymbol, NumUndefinedSymbols,
1248 IndirectSymbolOffset, NumIndirectSymbols);
1251 // Write the actual section data.
1252 for (MCAssembler::const_iterator it = Asm.begin(),
1253 ie = Asm.end(); it != ie; ++it)
1254 Asm.WriteSectionData(it, Layout, Writer);
1256 // Write the extra padding.
1257 WriteZeros(SectionDataPadding);
1259 // Write the relocation entries.
1260 for (MCAssembler::const_iterator it = Asm.begin(),
1261 ie = Asm.end(); it != ie; ++it) {
1262 // Write the section relocation entries, in reverse order to match 'as'
1263 // (approximately, the exact algorithm is more complicated than this).
1264 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1265 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1266 Write32(Relocs[e - i - 1].Word0);
1267 Write32(Relocs[e - i - 1].Word1);
1271 // Write the symbol table data, if used.
1273 // Write the indirect symbol entries.
1274 for (MCAssembler::const_indirect_symbol_iterator
1275 it = Asm.indirect_symbol_begin(),
1276 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1277 // Indirect symbols in the non lazy symbol pointer section have some
1278 // special handling.
1279 const MCSectionMachO &Section =
1280 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1281 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1282 // If this symbol is defined and internal, mark it as such.
1283 if (it->Symbol->isDefined() &&
1284 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1285 uint32_t Flags = ISF_Local;
1286 if (it->Symbol->isAbsolute())
1287 Flags |= ISF_Absolute;
1293 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1296 // FIXME: Check that offsets match computed ones.
1298 // Write the symbol table entries.
1299 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1300 WriteNlist(LocalSymbolData[i], Layout);
1301 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1302 WriteNlist(ExternalSymbolData[i], Layout);
1303 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1304 WriteNlist(UndefinedSymbolData[i], Layout);
1306 // Write the string table.
1307 OS << StringTable.str();
1314 MachObjectWriter::MachObjectWriter(raw_ostream &OS,
1316 bool IsLittleEndian)
1317 : MCObjectWriter(OS, IsLittleEndian)
1319 Impl = new MachObjectWriterImpl(this, Is64Bit);
1322 MachObjectWriter::~MachObjectWriter() {
1323 delete (MachObjectWriterImpl*) Impl;
1326 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm) {
1327 ((MachObjectWriterImpl*) Impl)->ExecutePostLayoutBinding(Asm);
1330 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1331 const MCAsmLayout &Layout,
1332 const MCFragment *Fragment,
1333 const MCFixup &Fixup, MCValue Target,
1334 uint64_t &FixedValue) {
1335 ((MachObjectWriterImpl*) Impl)->RecordRelocation(Asm, Layout, Fragment, Fixup,
1336 Target, FixedValue);
1339 bool MachObjectWriter::IsFixupFullyResolved(const MCAssembler &Asm,
1340 const MCValue Target,
1342 const MCFragment *DF) const {
1343 return ((MachObjectWriterImpl*) Impl)->IsFixupFullyResolved(Asm, Target,
1347 void MachObjectWriter::WriteObject(const MCAssembler &Asm,
1348 const MCAsmLayout &Layout) {
1349 ((MachObjectWriterImpl*) Impl)->WriteObject(Asm, Layout);