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 // The support for the situation where one or both of the symbols would
620 // require a local relocation is handled just like if the symbols were
621 // external. This is certainly used in the case of debug sections where
622 // the section has only temporary symbols and thus the symbols don't have
623 // base symbols. This is encoded using the section ordinal and
624 // non-extern relocation entries.
626 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
627 // a single SIGNED relocation); reject it for now. Except the case where
628 // both symbols don't have a base, equal but both NULL.
629 if (A_Base == B_Base && A_Base)
630 report_fatal_error("unsupported relocation with identical base");
632 Value += Layout.getSymbolAddress(&A_SD) -
633 (A_Base == NULL ? 0 : Layout.getSymbolAddress(A_Base));
634 Value -= Layout.getSymbolAddress(&B_SD) -
635 (B_Base == NULL ? 0 : Layout.getSymbolAddress(B_Base));
638 Index = A_Base->getIndex();
642 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
645 Type = RIT_X86_64_Unsigned;
647 MachRelocationEntry MRE;
648 MRE.Word0 = FixupOffset;
649 MRE.Word1 = ((Index << 0) |
654 Relocations[Fragment->getParent()].push_back(MRE);
657 Index = B_Base->getIndex();
661 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
664 Type = RIT_X86_64_Subtractor;
666 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
667 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
668 const MCSymbolData *Base = Asm.getAtom(&SD);
670 // Relocations inside debug sections always use local relocations when
671 // possible. This seems to be done because the debugger doesn't fully
672 // understand x86_64 relocation entries, and expects to find values that
673 // have already been fixed up.
674 if (Symbol->isInSection()) {
675 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
676 Fragment->getParent()->getSection());
677 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
681 // x86_64 almost always uses external relocations, except when there is no
682 // symbol to use as a base address (a local symbol with no preceeding
683 // non-local symbol).
685 Index = Base->getIndex();
688 // Add the local offset, if needed.
690 Value += Layout.getSymbolAddress(&SD) - Layout.getSymbolAddress(Base);
691 } else if (Symbol->isInSection()) {
692 // The index is the section ordinal (1-based).
693 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
695 Value += Layout.getSymbolAddress(&SD);
698 Value -= FixupAddress + (1 << Log2Size);
700 report_fatal_error("unsupported relocation of undefined symbol '" +
701 Symbol->getName() + "'");
704 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
707 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
708 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
709 // rewrite the movq to an leaq at link time if the symbol ends up in
710 // the same linkage unit.
711 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
712 Type = RIT_X86_64_GOTLoad;
714 Type = RIT_X86_64_GOT;
715 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
716 Type = RIT_X86_64_TLV;
717 } else if (Modifier != MCSymbolRefExpr::VK_None) {
718 report_fatal_error("unsupported symbol modifier in relocation");
720 Type = RIT_X86_64_Signed;
722 // The Darwin x86_64 relocation format has a problem where it cannot
723 // encode an address (L<foo> + <constant>) which is outside the atom
724 // containing L<foo>. Generally, this shouldn't occur but it does
725 // happen when we have a RIPrel instruction with data following the
726 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
727 // adjustment Darwin x86_64 uses, the offset is still negative and
728 // the linker has no way to recognize this.
730 // To work around this, Darwin uses several special relocation types
731 // to indicate the offsets. However, the specification or
732 // implementation of these seems to also be incomplete; they should
733 // adjust the addend as well based on the actual encoded instruction
734 // (the additional bias), but instead appear to just look at the
736 switch (-(Target.getConstant() + (1LL << Log2Size))) {
737 case 1: Type = RIT_X86_64_Signed1; break;
738 case 2: Type = RIT_X86_64_Signed2; break;
739 case 4: Type = RIT_X86_64_Signed4; break;
743 if (Modifier != MCSymbolRefExpr::VK_None)
744 report_fatal_error("unsupported symbol modifier in branch "
747 Type = RIT_X86_64_Branch;
750 if (Modifier == MCSymbolRefExpr::VK_GOT) {
751 Type = RIT_X86_64_GOT;
752 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
753 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
754 // which case all we do is set the PCrel bit in the relocation entry;
755 // this is used with exception handling, for example. The source is
756 // required to include any necessary offset directly.
757 Type = RIT_X86_64_GOT;
759 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
760 report_fatal_error("TLVP symbol modifier should have been rip-rel");
761 } else if (Modifier != MCSymbolRefExpr::VK_None)
762 report_fatal_error("unsupported symbol modifier in relocation");
764 Type = RIT_X86_64_Unsigned;
768 // x86_64 always writes custom values into the fixups.
771 // struct relocation_info (8 bytes)
772 MachRelocationEntry MRE;
773 MRE.Word0 = FixupOffset;
774 MRE.Word1 = ((Index << 0) |
779 Relocations[Fragment->getParent()].push_back(MRE);
782 void RecordScatteredRelocation(const MCAssembler &Asm,
783 const MCAsmLayout &Layout,
784 const MCFragment *Fragment,
785 const MCFixup &Fixup, MCValue Target,
786 uint64_t &FixedValue) {
787 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
788 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
789 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
790 unsigned Type = RIT_Vanilla;
793 const MCSymbol *A = &Target.getSymA()->getSymbol();
794 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
796 if (!A_SD->getFragment())
797 report_fatal_error("symbol '" + A->getName() +
798 "' can not be undefined in a subtraction expression");
800 uint32_t Value = Layout.getSymbolAddress(A_SD);
803 if (const MCSymbolRefExpr *B = Target.getSymB()) {
804 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
806 if (!B_SD->getFragment())
807 report_fatal_error("symbol '" + B->getSymbol().getName() +
808 "' can not be undefined in a subtraction expression");
810 // Select the appropriate difference relocation type.
812 // Note that there is no longer any semantic difference between these two
813 // relocation types from the linkers point of view, this is done solely
814 // for pedantic compatibility with 'as'.
815 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
816 Value2 = Layout.getSymbolAddress(B_SD);
819 // Relocations are written out in reverse order, so the PAIR comes first.
820 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
821 MachRelocationEntry MRE;
822 MRE.Word0 = ((0 << 0) |
828 Relocations[Fragment->getParent()].push_back(MRE);
831 MachRelocationEntry MRE;
832 MRE.Word0 = ((FixupOffset << 0) |
838 Relocations[Fragment->getParent()].push_back(MRE);
841 void RecordTLVPRelocation(const MCAssembler &Asm,
842 const MCAsmLayout &Layout,
843 const MCFragment *Fragment,
844 const MCFixup &Fixup, MCValue Target,
845 uint64_t &FixedValue) {
846 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
848 "Should only be called with a 32-bit TLVP relocation!");
850 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
851 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
852 unsigned IsPCRel = 0;
854 // Get the symbol data.
855 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
856 unsigned Index = SD_A->getIndex();
858 // We're only going to have a second symbol in pic mode and it'll be a
859 // subtraction from the picbase. For 32-bit pic the addend is the difference
860 // between the picbase and the next address. For 32-bit static the addend
862 if (Target.getSymB()) {
863 // If this is a subtraction then we're pcrel.
864 uint32_t FixupAddress =
865 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
866 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
868 FixedValue = (FixupAddress - Layout.getSymbolAddress(SD_B) +
869 Target.getConstant());
870 FixedValue += 1ULL << Log2Size;
875 // struct relocation_info (8 bytes)
876 MachRelocationEntry MRE;
878 MRE.Word1 = ((Index << 0) |
881 (1 << 27) | // Extern
882 (RIT_TLV << 28)); // Type
883 Relocations[Fragment->getParent()].push_back(MRE);
886 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
887 const MCFragment *Fragment, const MCFixup &Fixup,
888 MCValue Target, uint64_t &FixedValue) {
890 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
894 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
895 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
897 // If this is a 32-bit TLVP reloc it's handled a bit differently.
898 if (Target.getSymA() &&
899 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
900 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
904 // If this is a difference or a defined symbol plus an offset, then we need
905 // a scattered relocation entry.
906 // Differences always require scattered relocations.
907 if (Target.getSymB())
908 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
911 // Get the symbol data, if any.
912 MCSymbolData *SD = 0;
913 if (Target.getSymA())
914 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
916 // If this is an internal relocation with an offset, it also needs a
917 // scattered relocation entry.
918 uint32_t Offset = Target.getConstant();
920 Offset += 1 << Log2Size;
921 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
922 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
926 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
928 unsigned IsExtern = 0;
931 if (Target.isAbsolute()) { // constant
932 // SymbolNum of 0 indicates the absolute section.
934 // FIXME: Currently, these are never generated (see code below). I cannot
935 // find a case where they are actually emitted.
938 // Check whether we need an external or internal relocation.
939 if (doesSymbolRequireExternRelocation(SD)) {
941 Index = SD->getIndex();
942 // For external relocations, make sure to offset the fixup value to
943 // compensate for the addend of the symbol address, if it was
944 // undefined. This occurs with weak definitions, for example.
945 if (!SD->Symbol->isUndefined())
946 FixedValue -= Layout.getSymbolAddress(SD);
948 // The index is the section ordinal (1-based).
949 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
955 // struct relocation_info (8 bytes)
956 MachRelocationEntry MRE;
957 MRE.Word0 = FixupOffset;
958 MRE.Word1 = ((Index << 0) |
963 Relocations[Fragment->getParent()].push_back(MRE);
966 void BindIndirectSymbols(MCAssembler &Asm) {
967 // This is the point where 'as' creates actual symbols for indirect symbols
968 // (in the following two passes). It would be easier for us to do this
969 // sooner when we see the attribute, but that makes getting the order in the
970 // symbol table much more complicated than it is worth.
972 // FIXME: Revisit this when the dust settles.
974 // Bind non lazy symbol pointers first.
975 unsigned IndirectIndex = 0;
976 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
977 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
978 const MCSectionMachO &Section =
979 cast<MCSectionMachO>(it->SectionData->getSection());
981 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
984 // Initialize the section indirect symbol base, if necessary.
985 if (!IndirectSymBase.count(it->SectionData))
986 IndirectSymBase[it->SectionData] = IndirectIndex;
988 Asm.getOrCreateSymbolData(*it->Symbol);
991 // Then lazy symbol pointers and symbol stubs.
993 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
994 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
995 const MCSectionMachO &Section =
996 cast<MCSectionMachO>(it->SectionData->getSection());
998 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
999 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
1002 // Initialize the section indirect symbol base, if necessary.
1003 if (!IndirectSymBase.count(it->SectionData))
1004 IndirectSymBase[it->SectionData] = IndirectIndex;
1006 // Set the symbol type to undefined lazy, but only on construction.
1008 // FIXME: Do not hardcode.
1010 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
1012 Entry.setFlags(Entry.getFlags() | 0x0001);
1016 /// ComputeSymbolTable - Compute the symbol table data
1018 /// \param StringTable [out] - The string table data.
1019 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
1021 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
1022 std::vector<MachSymbolData> &LocalSymbolData,
1023 std::vector<MachSymbolData> &ExternalSymbolData,
1024 std::vector<MachSymbolData> &UndefinedSymbolData) {
1025 // Build section lookup table.
1026 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
1028 for (MCAssembler::iterator it = Asm.begin(),
1029 ie = Asm.end(); it != ie; ++it, ++Index)
1030 SectionIndexMap[&it->getSection()] = Index;
1031 assert(Index <= 256 && "Too many sections!");
1033 // Index 0 is always the empty string.
1034 StringMap<uint64_t> StringIndexMap;
1035 StringTable += '\x00';
1037 // Build the symbol arrays and the string table, but only for non-local
1040 // The particular order that we collect the symbols and create the string
1041 // table, then sort the symbols is chosen to match 'as'. Even though it
1042 // doesn't matter for correctness, this is important for letting us diff .o
1044 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1045 ie = Asm.symbol_end(); it != ie; ++it) {
1046 const MCSymbol &Symbol = it->getSymbol();
1048 // Ignore non-linker visible symbols.
1049 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1052 if (!it->isExternal() && !Symbol.isUndefined())
1055 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1057 Entry = StringTable.size();
1058 StringTable += Symbol.getName();
1059 StringTable += '\x00';
1063 MSD.SymbolData = it;
1064 MSD.StringIndex = Entry;
1066 if (Symbol.isUndefined()) {
1067 MSD.SectionIndex = 0;
1068 UndefinedSymbolData.push_back(MSD);
1069 } else if (Symbol.isAbsolute()) {
1070 MSD.SectionIndex = 0;
1071 ExternalSymbolData.push_back(MSD);
1073 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1074 assert(MSD.SectionIndex && "Invalid section index!");
1075 ExternalSymbolData.push_back(MSD);
1079 // Now add the data for local symbols.
1080 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1081 ie = Asm.symbol_end(); it != ie; ++it) {
1082 const MCSymbol &Symbol = it->getSymbol();
1084 // Ignore non-linker visible symbols.
1085 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1088 if (it->isExternal() || Symbol.isUndefined())
1091 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1093 Entry = StringTable.size();
1094 StringTable += Symbol.getName();
1095 StringTable += '\x00';
1099 MSD.SymbolData = it;
1100 MSD.StringIndex = Entry;
1102 if (Symbol.isAbsolute()) {
1103 MSD.SectionIndex = 0;
1104 LocalSymbolData.push_back(MSD);
1106 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1107 assert(MSD.SectionIndex && "Invalid section index!");
1108 LocalSymbolData.push_back(MSD);
1112 // External and undefined symbols are required to be in lexicographic order.
1113 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1114 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1116 // Set the symbol indices.
1118 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1119 LocalSymbolData[i].SymbolData->setIndex(Index++);
1120 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1121 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1122 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1123 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1125 // The string table is padded to a multiple of 4.
1126 while (StringTable.size() % 4)
1127 StringTable += '\x00';
1130 void ExecutePostLayoutBinding(MCAssembler &Asm) {
1131 // Create symbol data for any indirect symbols.
1132 BindIndirectSymbols(Asm);
1134 // Compute symbol table information and bind symbol indices.
1135 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1136 UndefinedSymbolData);
1140 bool IsFixupFullyResolved(const MCAssembler &Asm,
1141 const MCValue Target,
1143 const MCFragment *DF) const {
1144 // If we are using scattered symbols, determine whether this value is
1145 // actually resolved; scattering may cause atoms to move.
1146 if (Asm.getBackend().hasScatteredSymbols()) {
1147 if (Asm.getBackend().hasReliableSymbolDifference()) {
1148 // If this is a PCrel relocation, find the base atom (identified by its
1149 // symbol) that the fixup value is relative to.
1150 const MCSymbolData *BaseSymbol = 0;
1152 BaseSymbol = DF->getAtom();
1157 return isScatteredFixupFullyResolved(Asm, Target, BaseSymbol);
1159 const MCSection *BaseSection = 0;
1161 BaseSection = &DF->getParent()->getSection();
1163 return isScatteredFixupFullyResolvedSimple(Asm, Target, BaseSection);
1169 void WriteObject(const MCAssembler &Asm, const MCAsmLayout &Layout) {
1170 unsigned NumSections = Asm.size();
1172 // The section data starts after the header, the segment load command (and
1173 // section headers) and the symbol table.
1174 unsigned NumLoadCommands = 1;
1175 uint64_t LoadCommandsSize = Is64Bit ?
1176 SegmentLoadCommand64Size + NumSections * Section64Size :
1177 SegmentLoadCommand32Size + NumSections * Section32Size;
1179 // Add the symbol table load command sizes, if used.
1180 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1181 UndefinedSymbolData.size();
1183 NumLoadCommands += 2;
1184 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
1187 // Compute the total size of the section data, as well as its file size and
1189 uint64_t SectionDataStart = (Is64Bit ? Header64Size : Header32Size)
1191 uint64_t SectionDataSize = 0;
1192 uint64_t SectionDataFileSize = 0;
1193 uint64_t VMSize = 0;
1194 for (MCAssembler::const_iterator it = Asm.begin(),
1195 ie = Asm.end(); it != ie; ++it) {
1196 const MCSectionData &SD = *it;
1197 uint64_t Address = Layout.getSectionAddress(&SD);
1198 uint64_t Size = Layout.getSectionSize(&SD);
1199 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1201 VMSize = std::max(VMSize, Address + Size);
1203 if (Asm.getBackend().isVirtualSection(SD.getSection()))
1206 SectionDataSize = std::max(SectionDataSize, Address + Size);
1207 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1210 // The section data is padded to 4 bytes.
1212 // FIXME: Is this machine dependent?
1213 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1214 SectionDataFileSize += SectionDataPadding;
1216 // Write the prolog, starting with the header and load command...
1217 WriteHeader(NumLoadCommands, LoadCommandsSize,
1218 Asm.getSubsectionsViaSymbols());
1219 WriteSegmentLoadCommand(NumSections, VMSize,
1220 SectionDataStart, SectionDataSize);
1222 // ... and then the section headers.
1223 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1224 for (MCAssembler::const_iterator it = Asm.begin(),
1225 ie = Asm.end(); it != ie; ++it) {
1226 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1227 unsigned NumRelocs = Relocs.size();
1228 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
1229 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1230 RelocTableEnd += NumRelocs * RelocationInfoSize;
1233 // Write the symbol table load command, if used.
1235 unsigned FirstLocalSymbol = 0;
1236 unsigned NumLocalSymbols = LocalSymbolData.size();
1237 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1238 unsigned NumExternalSymbols = ExternalSymbolData.size();
1239 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1240 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1241 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1242 unsigned NumSymTabSymbols =
1243 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1244 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1245 uint64_t IndirectSymbolOffset = 0;
1247 // If used, the indirect symbols are written after the section data.
1248 if (NumIndirectSymbols)
1249 IndirectSymbolOffset = RelocTableEnd;
1251 // The symbol table is written after the indirect symbol data.
1252 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1254 // The string table is written after symbol table.
1255 uint64_t StringTableOffset =
1256 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? Nlist64Size :
1258 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1259 StringTableOffset, StringTable.size());
1261 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1262 FirstExternalSymbol, NumExternalSymbols,
1263 FirstUndefinedSymbol, NumUndefinedSymbols,
1264 IndirectSymbolOffset, NumIndirectSymbols);
1267 // Write the actual section data.
1268 for (MCAssembler::const_iterator it = Asm.begin(),
1269 ie = Asm.end(); it != ie; ++it)
1270 Asm.WriteSectionData(it, Layout, Writer);
1272 // Write the extra padding.
1273 WriteZeros(SectionDataPadding);
1275 // Write the relocation entries.
1276 for (MCAssembler::const_iterator it = Asm.begin(),
1277 ie = Asm.end(); it != ie; ++it) {
1278 // Write the section relocation entries, in reverse order to match 'as'
1279 // (approximately, the exact algorithm is more complicated than this).
1280 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1281 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1282 Write32(Relocs[e - i - 1].Word0);
1283 Write32(Relocs[e - i - 1].Word1);
1287 // Write the symbol table data, if used.
1289 // Write the indirect symbol entries.
1290 for (MCAssembler::const_indirect_symbol_iterator
1291 it = Asm.indirect_symbol_begin(),
1292 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1293 // Indirect symbols in the non lazy symbol pointer section have some
1294 // special handling.
1295 const MCSectionMachO &Section =
1296 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1297 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1298 // If this symbol is defined and internal, mark it as such.
1299 if (it->Symbol->isDefined() &&
1300 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1301 uint32_t Flags = ISF_Local;
1302 if (it->Symbol->isAbsolute())
1303 Flags |= ISF_Absolute;
1309 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1312 // FIXME: Check that offsets match computed ones.
1314 // Write the symbol table entries.
1315 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1316 WriteNlist(LocalSymbolData[i], Layout);
1317 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1318 WriteNlist(ExternalSymbolData[i], Layout);
1319 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1320 WriteNlist(UndefinedSymbolData[i], Layout);
1322 // Write the string table.
1323 OS << StringTable.str();
1330 MachObjectWriter::MachObjectWriter(raw_ostream &OS,
1332 bool IsLittleEndian)
1333 : MCObjectWriter(OS, IsLittleEndian)
1335 Impl = new MachObjectWriterImpl(this, Is64Bit);
1338 MachObjectWriter::~MachObjectWriter() {
1339 delete (MachObjectWriterImpl*) Impl;
1342 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm) {
1343 ((MachObjectWriterImpl*) Impl)->ExecutePostLayoutBinding(Asm);
1346 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1347 const MCAsmLayout &Layout,
1348 const MCFragment *Fragment,
1349 const MCFixup &Fixup, MCValue Target,
1350 uint64_t &FixedValue) {
1351 ((MachObjectWriterImpl*) Impl)->RecordRelocation(Asm, Layout, Fragment, Fixup,
1352 Target, FixedValue);
1355 bool MachObjectWriter::IsFixupFullyResolved(const MCAssembler &Asm,
1356 const MCValue Target,
1358 const MCFragment *DF) const {
1359 return ((MachObjectWriterImpl*) Impl)->IsFixupFullyResolved(Asm, Target,
1363 void MachObjectWriter::WriteObject(MCAssembler &Asm,
1364 const MCAsmLayout &Layout) {
1365 ((MachObjectWriterImpl*) Impl)->WriteObject(Asm, Layout);