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/MCMachObjectWriter.h"
11 #include "llvm/ADT/OwningPtr.h"
12 #include "llvm/ADT/StringMap.h"
13 #include "llvm/ADT/Twine.h"
14 #include "llvm/MC/MCAssembler.h"
15 #include "llvm/MC/MCAsmLayout.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCObjectWriter.h"
18 #include "llvm/MC/MCSectionMachO.h"
19 #include "llvm/MC/MCSymbol.h"
20 #include "llvm/MC/MCMachOSymbolFlags.h"
21 #include "llvm/MC/MCValue.h"
22 #include "llvm/Object/MachOFormat.h"
23 #include "llvm/Support/ErrorHandling.h"
24 #include "llvm/Target/TargetAsmBackend.h"
27 #include "../Target/ARM/ARMFixupKinds.h"
28 #include "../Target/X86/X86FixupKinds.h"
32 using namespace llvm::object;
34 // FIXME: this has been copied from (or to) X86AsmBackend.cpp
35 static unsigned getFixupKindLog2Size(unsigned Kind) {
38 llvm_unreachable("invalid fixup kind!");
40 case FK_Data_1: return 0;
42 case FK_Data_2: return 1;
44 // FIXME: Remove these!!!
45 case X86::reloc_riprel_4byte:
46 case X86::reloc_riprel_4byte_movq_load:
47 case X86::reloc_signed_4byte:
48 case FK_Data_4: return 2;
49 case FK_Data_8: return 3;
53 static bool doesSymbolRequireExternRelocation(MCSymbolData *SD) {
54 // Undefined symbols are always extern.
55 if (SD->Symbol->isUndefined())
58 // References to weak definitions require external relocation entries; the
59 // definition may not always be the one in the same object file.
60 if (SD->getFlags() & SF_WeakDefinition)
63 // Otherwise, we can use an internal relocation.
67 static bool isScatteredFixupFullyResolved(const MCAssembler &Asm,
69 const MCSymbolData *BaseSymbol) {
70 // The effective fixup address is
71 // addr(atom(A)) + offset(A)
72 // - addr(atom(B)) - offset(B)
73 // - addr(BaseSymbol) + <fixup offset from base symbol>
74 // and the offsets are not relocatable, so the fixup is fully resolved when
75 // addr(atom(A)) - addr(atom(B)) - addr(BaseSymbol) == 0.
77 // Note that "false" is almost always conservatively correct (it means we emit
78 // a relocation which is unnecessary), except when it would force us to emit a
79 // relocation which the target cannot encode.
81 const MCSymbolData *A_Base = 0, *B_Base = 0;
82 if (const MCSymbolRefExpr *A = Target.getSymA()) {
83 // Modified symbol references cannot be resolved.
84 if (A->getKind() != MCSymbolRefExpr::VK_None)
87 A_Base = Asm.getAtom(&Asm.getSymbolData(A->getSymbol()));
92 if (const MCSymbolRefExpr *B = Target.getSymB()) {
93 // Modified symbol references cannot be resolved.
94 if (B->getKind() != MCSymbolRefExpr::VK_None)
97 B_Base = Asm.getAtom(&Asm.getSymbolData(B->getSymbol()));
102 // If there is no base, A and B have to be the same atom for this fixup to be
105 return A_Base == B_Base;
107 // Otherwise, B must be missing and A must be the base.
108 return !B_Base && BaseSymbol == A_Base;
111 static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm,
112 const MCValue Target,
113 const MCSection *BaseSection) {
114 // The effective fixup address is
115 // addr(atom(A)) + offset(A)
116 // - addr(atom(B)) - offset(B)
117 // - addr(<base symbol>) + <fixup offset from base symbol>
118 // and the offsets are not relocatable, so the fixup is fully resolved when
119 // addr(atom(A)) - addr(atom(B)) - addr(<base symbol>)) == 0.
121 // The simple (Darwin, except on x86_64) way of dealing with this was to
122 // assume that any reference to a temporary symbol *must* be a temporary
123 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
124 // relocation to a temporary symbol (in the same section) is fully
125 // resolved. This also works in conjunction with absolutized .set, which
126 // requires the compiler to use .set to absolutize the differences between
127 // symbols which the compiler knows to be assembly time constants, so we don't
128 // need to worry about considering symbol differences fully resolved.
130 // Non-relative fixups are only resolved if constant.
132 return Target.isAbsolute();
134 // Otherwise, relative fixups are only resolved if not a difference and the
135 // target is a temporary in the same section.
136 if (Target.isAbsolute() || Target.getSymB())
139 const MCSymbol *A = &Target.getSymA()->getSymbol();
140 if (!A->isTemporary() || !A->isInSection() ||
141 &A->getSection() != BaseSection)
149 class MachObjectWriter : public MCObjectWriter {
150 /// MachSymbolData - Helper struct for containing some precomputed information
152 struct MachSymbolData {
153 MCSymbolData *SymbolData;
154 uint64_t StringIndex;
155 uint8_t SectionIndex;
157 // Support lexicographic sorting.
158 bool operator<(const MachSymbolData &RHS) const {
159 return SymbolData->getSymbol().getName() <
160 RHS.SymbolData->getSymbol().getName();
164 /// The target specific Mach-O writer instance.
165 llvm::OwningPtr<MCMachObjectTargetWriter> TargetObjectWriter;
167 /// @name Relocation Data
170 llvm::DenseMap<const MCSectionData*,
171 std::vector<macho::RelocationEntry> > Relocations;
172 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
175 /// @name Symbol Table Data
178 SmallString<256> StringTable;
179 std::vector<MachSymbolData> LocalSymbolData;
180 std::vector<MachSymbolData> ExternalSymbolData;
181 std::vector<MachSymbolData> UndefinedSymbolData;
186 /// @name Utility Methods
189 bool isFixupKindPCRel(const MCAssembler &Asm, unsigned Kind) {
190 const MCFixupKindInfo &FKI = Asm.getBackend().getFixupKindInfo(
193 return FKI.Flags & MCFixupKindInfo::FKF_IsPCRel;
198 SectionAddrMap SectionAddress;
199 uint64_t getSectionAddress(const MCSectionData* SD) const {
200 return SectionAddress.lookup(SD);
202 uint64_t getSymbolAddress(const MCSymbolData* SD,
203 const MCAsmLayout &Layout) const {
204 return getSectionAddress(SD->getFragment()->getParent()) +
205 Layout.getSymbolOffset(SD);
207 uint64_t getFragmentAddress(const MCFragment *Fragment,
208 const MCAsmLayout &Layout) const {
209 return getSectionAddress(Fragment->getParent()) +
210 Layout.getFragmentOffset(Fragment);
213 uint64_t getPaddingSize(const MCSectionData *SD,
214 const MCAsmLayout &Layout) const {
215 uint64_t EndAddr = getSectionAddress(SD) + Layout.getSectionAddressSize(SD);
216 unsigned Next = SD->getLayoutOrder() + 1;
217 if (Next >= Layout.getSectionOrder().size())
220 const MCSectionData &NextSD = *Layout.getSectionOrder()[Next];
221 if (NextSD.getSection().isVirtualSection())
223 return OffsetToAlignment(EndAddr, NextSD.getAlignment());
227 MachObjectWriter(MCMachObjectTargetWriter *MOTW, raw_ostream &_OS,
228 bool _IsLittleEndian)
229 : MCObjectWriter(_OS, _IsLittleEndian), TargetObjectWriter(MOTW) {
232 /// @name Target Writer Proxy Accessors
235 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
237 uint32_t CPUType = TargetObjectWriter->getCPUType() & ~mach::CTFM_ArchMask;
238 return CPUType == mach::CTM_ARM;
243 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
244 bool SubsectionsViaSymbols) {
247 if (SubsectionsViaSymbols)
248 Flags |= macho::HF_SubsectionsViaSymbols;
250 // struct mach_header (28 bytes) or
251 // struct mach_header_64 (32 bytes)
253 uint64_t Start = OS.tell();
256 Write32(is64Bit() ? macho::HM_Object64 : macho::HM_Object32);
258 Write32(TargetObjectWriter->getCPUType());
259 Write32(TargetObjectWriter->getCPUSubtype());
261 Write32(macho::HFT_Object);
262 Write32(NumLoadCommands);
263 Write32(LoadCommandsSize);
266 Write32(0); // reserved
268 assert(OS.tell() - Start == is64Bit() ?
269 macho::Header64Size : macho::Header32Size);
272 /// WriteSegmentLoadCommand - Write a segment load command.
274 /// \arg NumSections - The number of sections in this segment.
275 /// \arg SectionDataSize - The total size of the sections.
276 void WriteSegmentLoadCommand(unsigned NumSections,
278 uint64_t SectionDataStartOffset,
279 uint64_t SectionDataSize) {
280 // struct segment_command (56 bytes) or
281 // struct segment_command_64 (72 bytes)
283 uint64_t Start = OS.tell();
286 unsigned SegmentLoadCommandSize =
287 is64Bit() ? macho::SegmentLoadCommand64Size:
288 macho::SegmentLoadCommand32Size;
289 Write32(is64Bit() ? macho::LCT_Segment64 : macho::LCT_Segment);
290 Write32(SegmentLoadCommandSize +
291 NumSections * (is64Bit() ? macho::Section64Size :
292 macho::Section32Size));
296 Write64(0); // vmaddr
297 Write64(VMSize); // vmsize
298 Write64(SectionDataStartOffset); // file offset
299 Write64(SectionDataSize); // file size
301 Write32(0); // vmaddr
302 Write32(VMSize); // vmsize
303 Write32(SectionDataStartOffset); // file offset
304 Write32(SectionDataSize); // file size
306 Write32(0x7); // maxprot
307 Write32(0x7); // initprot
308 Write32(NumSections);
311 assert(OS.tell() - Start == SegmentLoadCommandSize);
314 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
315 const MCSectionData &SD, uint64_t FileOffset,
316 uint64_t RelocationsStart, unsigned NumRelocations) {
317 uint64_t SectionSize = Layout.getSectionAddressSize(&SD);
319 // The offset is unused for virtual sections.
320 if (SD.getSection().isVirtualSection()) {
321 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
325 // struct section (68 bytes) or
326 // struct section_64 (80 bytes)
328 uint64_t Start = OS.tell();
331 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
332 WriteBytes(Section.getSectionName(), 16);
333 WriteBytes(Section.getSegmentName(), 16);
335 Write64(getSectionAddress(&SD)); // address
336 Write64(SectionSize); // size
338 Write32(getSectionAddress(&SD)); // address
339 Write32(SectionSize); // size
343 unsigned Flags = Section.getTypeAndAttributes();
344 if (SD.hasInstructions())
345 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
347 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
348 Write32(Log2_32(SD.getAlignment()));
349 Write32(NumRelocations ? RelocationsStart : 0);
350 Write32(NumRelocations);
352 Write32(IndirectSymBase.lookup(&SD)); // reserved1
353 Write32(Section.getStubSize()); // reserved2
355 Write32(0); // reserved3
357 assert(OS.tell() - Start == is64Bit() ? macho::Section64Size :
358 macho::Section32Size);
361 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
362 uint32_t StringTableOffset,
363 uint32_t StringTableSize) {
364 // struct symtab_command (24 bytes)
366 uint64_t Start = OS.tell();
369 Write32(macho::LCT_Symtab);
370 Write32(macho::SymtabLoadCommandSize);
371 Write32(SymbolOffset);
373 Write32(StringTableOffset);
374 Write32(StringTableSize);
376 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
379 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
380 uint32_t NumLocalSymbols,
381 uint32_t FirstExternalSymbol,
382 uint32_t NumExternalSymbols,
383 uint32_t FirstUndefinedSymbol,
384 uint32_t NumUndefinedSymbols,
385 uint32_t IndirectSymbolOffset,
386 uint32_t NumIndirectSymbols) {
387 // struct dysymtab_command (80 bytes)
389 uint64_t Start = OS.tell();
392 Write32(macho::LCT_Dysymtab);
393 Write32(macho::DysymtabLoadCommandSize);
394 Write32(FirstLocalSymbol);
395 Write32(NumLocalSymbols);
396 Write32(FirstExternalSymbol);
397 Write32(NumExternalSymbols);
398 Write32(FirstUndefinedSymbol);
399 Write32(NumUndefinedSymbols);
400 Write32(0); // tocoff
402 Write32(0); // modtaboff
403 Write32(0); // nmodtab
404 Write32(0); // extrefsymoff
405 Write32(0); // nextrefsyms
406 Write32(IndirectSymbolOffset);
407 Write32(NumIndirectSymbols);
408 Write32(0); // extreloff
409 Write32(0); // nextrel
410 Write32(0); // locreloff
411 Write32(0); // nlocrel
413 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
416 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
417 MCSymbolData &Data = *MSD.SymbolData;
418 const MCSymbol &Symbol = Data.getSymbol();
420 uint16_t Flags = Data.getFlags();
421 uint32_t Address = 0;
423 // Set the N_TYPE bits. See <mach-o/nlist.h>.
425 // FIXME: Are the prebound or indirect fields possible here?
426 if (Symbol.isUndefined())
427 Type = macho::STT_Undefined;
428 else if (Symbol.isAbsolute())
429 Type = macho::STT_Absolute;
431 Type = macho::STT_Section;
433 // FIXME: Set STAB bits.
435 if (Data.isPrivateExtern())
436 Type |= macho::STF_PrivateExtern;
439 if (Data.isExternal() || Symbol.isUndefined())
440 Type |= macho::STF_External;
442 // Compute the symbol address.
443 if (Symbol.isDefined()) {
444 if (Symbol.isAbsolute()) {
445 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
447 Address = getSymbolAddress(&Data, Layout);
449 } else if (Data.isCommon()) {
450 // Common symbols are encoded with the size in the address
451 // field, and their alignment in the flags.
452 Address = Data.getCommonSize();
454 // Common alignment is packed into the 'desc' bits.
455 if (unsigned Align = Data.getCommonAlignment()) {
456 unsigned Log2Size = Log2_32(Align);
457 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
459 report_fatal_error("invalid 'common' alignment '" +
461 // FIXME: Keep this mask with the SymbolFlags enumeration.
462 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
466 // struct nlist (12 bytes)
468 Write32(MSD.StringIndex);
470 Write8(MSD.SectionIndex);
472 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
481 // FIXME: We really need to improve the relocation validation. Basically, we
482 // want to implement a separate computation which evaluates the relocation
483 // entry as the linker would, and verifies that the resultant fixup value is
484 // exactly what the encoder wanted. This will catch several classes of
487 // - Relocation entry bugs, the two algorithms are unlikely to have the same
490 // - Relaxation issues, where we forget to relax something.
492 // - Input errors, where something cannot be correctly encoded. 'as' allows
493 // these through in many cases.
495 static bool isFixupKindRIPRel(unsigned Kind) {
496 return Kind == X86::reloc_riprel_4byte ||
497 Kind == X86::reloc_riprel_4byte_movq_load;
499 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
500 const MCFragment *Fragment,
501 const MCFixup &Fixup, MCValue Target,
502 uint64_t &FixedValue) {
503 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
504 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
505 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
508 uint32_t FixupOffset =
509 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
510 uint32_t FixupAddress =
511 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
514 unsigned IsExtern = 0;
517 Value = Target.getConstant();
520 // Compensate for the relocation offset, Darwin x86_64 relocations only
521 // have the addend and appear to have attempted to define it to be the
522 // actual expression addend without the PCrel bias. However, instructions
523 // with data following the relocation are not accomodated for (see comment
524 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
525 Value += 1LL << Log2Size;
528 if (Target.isAbsolute()) { // constant
529 // SymbolNum of 0 indicates the absolute section.
530 Type = macho::RIT_X86_64_Unsigned;
533 // FIXME: I believe this is broken, I don't think the linker can
534 // understand it. I think it would require a local relocation, but I'm not
535 // sure if that would work either. The official way to get an absolute
536 // PCrel relocation is to use an absolute symbol (which we don't support
540 Type = macho::RIT_X86_64_Branch;
542 } else if (Target.getSymB()) { // A - B + constant
543 const MCSymbol *A = &Target.getSymA()->getSymbol();
544 MCSymbolData &A_SD = Asm.getSymbolData(*A);
545 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
547 const MCSymbol *B = &Target.getSymB()->getSymbol();
548 MCSymbolData &B_SD = Asm.getSymbolData(*B);
549 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
551 // Neither symbol can be modified.
552 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
553 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
554 report_fatal_error("unsupported relocation of modified symbol");
556 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
557 // implement most of these correctly.
559 report_fatal_error("unsupported pc-relative relocation of difference");
561 // The support for the situation where one or both of the symbols would
562 // require a local relocation is handled just like if the symbols were
563 // external. This is certainly used in the case of debug sections where
564 // the section has only temporary symbols and thus the symbols don't have
565 // base symbols. This is encoded using the section ordinal and
566 // non-extern relocation entries.
568 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
569 // a single SIGNED relocation); reject it for now. Except the case where
570 // both symbols don't have a base, equal but both NULL.
571 if (A_Base == B_Base && A_Base)
572 report_fatal_error("unsupported relocation with identical base");
574 Value += getSymbolAddress(&A_SD, Layout) -
575 (A_Base == NULL ? 0 : getSymbolAddress(A_Base, Layout));
576 Value -= getSymbolAddress(&B_SD, Layout) -
577 (B_Base == NULL ? 0 : getSymbolAddress(B_Base, Layout));
580 Index = A_Base->getIndex();
584 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
587 Type = macho::RIT_X86_64_Unsigned;
589 macho::RelocationEntry MRE;
590 MRE.Word0 = FixupOffset;
591 MRE.Word1 = ((Index << 0) |
596 Relocations[Fragment->getParent()].push_back(MRE);
599 Index = B_Base->getIndex();
603 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
606 Type = macho::RIT_X86_64_Subtractor;
608 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
609 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
610 const MCSymbolData *Base = Asm.getAtom(&SD);
612 // Relocations inside debug sections always use local relocations when
613 // possible. This seems to be done because the debugger doesn't fully
614 // understand x86_64 relocation entries, and expects to find values that
615 // have already been fixed up.
616 if (Symbol->isInSection()) {
617 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
618 Fragment->getParent()->getSection());
619 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
623 // x86_64 almost always uses external relocations, except when there is no
624 // symbol to use as a base address (a local symbol with no preceeding
625 // non-local symbol).
627 Index = Base->getIndex();
630 // Add the local offset, if needed.
632 Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base);
633 } else if (Symbol->isInSection()) {
634 // The index is the section ordinal (1-based).
635 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
637 Value += getSymbolAddress(&SD, Layout);
640 Value -= FixupAddress + (1 << Log2Size);
641 } else if (Symbol->isVariable()) {
642 const MCExpr *Value = Symbol->getVariableValue();
644 bool isAbs = Value->EvaluateAsAbsolute(Res, Layout, SectionAddress);
649 report_fatal_error("unsupported relocation of variable '" +
650 Symbol->getName() + "'");
653 report_fatal_error("unsupported relocation of undefined symbol '" +
654 Symbol->getName() + "'");
657 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
660 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
661 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
662 // rewrite the movq to an leaq at link time if the symbol ends up in
663 // the same linkage unit.
664 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
665 Type = macho::RIT_X86_64_GOTLoad;
667 Type = macho::RIT_X86_64_GOT;
668 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
669 Type = macho::RIT_X86_64_TLV;
670 } else if (Modifier != MCSymbolRefExpr::VK_None) {
671 report_fatal_error("unsupported symbol modifier in relocation");
673 Type = macho::RIT_X86_64_Signed;
675 // The Darwin x86_64 relocation format has a problem where it cannot
676 // encode an address (L<foo> + <constant>) which is outside the atom
677 // containing L<foo>. Generally, this shouldn't occur but it does
678 // happen when we have a RIPrel instruction with data following the
679 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
680 // adjustment Darwin x86_64 uses, the offset is still negative and
681 // the linker has no way to recognize this.
683 // To work around this, Darwin uses several special relocation types
684 // to indicate the offsets. However, the specification or
685 // implementation of these seems to also be incomplete; they should
686 // adjust the addend as well based on the actual encoded instruction
687 // (the additional bias), but instead appear to just look at the
689 switch (-(Target.getConstant() + (1LL << Log2Size))) {
690 case 1: Type = macho::RIT_X86_64_Signed1; break;
691 case 2: Type = macho::RIT_X86_64_Signed2; break;
692 case 4: Type = macho::RIT_X86_64_Signed4; break;
696 if (Modifier != MCSymbolRefExpr::VK_None)
697 report_fatal_error("unsupported symbol modifier in branch "
700 Type = macho::RIT_X86_64_Branch;
703 if (Modifier == MCSymbolRefExpr::VK_GOT) {
704 Type = macho::RIT_X86_64_GOT;
705 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
706 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
707 // which case all we do is set the PCrel bit in the relocation entry;
708 // this is used with exception handling, for example. The source is
709 // required to include any necessary offset directly.
710 Type = macho::RIT_X86_64_GOT;
712 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
713 report_fatal_error("TLVP symbol modifier should have been rip-rel");
714 } else if (Modifier != MCSymbolRefExpr::VK_None)
715 report_fatal_error("unsupported symbol modifier in relocation");
717 Type = macho::RIT_X86_64_Unsigned;
721 // x86_64 always writes custom values into the fixups.
724 // struct relocation_info (8 bytes)
725 macho::RelocationEntry MRE;
726 MRE.Word0 = FixupOffset;
727 MRE.Word1 = ((Index << 0) |
732 Relocations[Fragment->getParent()].push_back(MRE);
735 void RecordScatteredRelocation(const MCAssembler &Asm,
736 const MCAsmLayout &Layout,
737 const MCFragment *Fragment,
738 const MCFixup &Fixup, MCValue Target,
739 uint64_t &FixedValue) {
740 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
741 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
742 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
743 unsigned Type = macho::RIT_Vanilla;
746 const MCSymbol *A = &Target.getSymA()->getSymbol();
747 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
749 if (!A_SD->getFragment())
750 report_fatal_error("symbol '" + A->getName() +
751 "' can not be undefined in a subtraction expression");
753 uint32_t Value = getSymbolAddress(A_SD, Layout);
754 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
755 FixedValue += SecAddr;
758 if (const MCSymbolRefExpr *B = Target.getSymB()) {
759 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
761 if (!B_SD->getFragment())
762 report_fatal_error("symbol '" + B->getSymbol().getName() +
763 "' can not be undefined in a subtraction expression");
765 // Select the appropriate difference relocation type.
767 // Note that there is no longer any semantic difference between these two
768 // relocation types from the linkers point of view, this is done solely
769 // for pedantic compatibility with 'as'.
770 Type = A_SD->isExternal() ? (unsigned)macho::RIT_Difference :
771 (unsigned)macho::RIT_Generic_LocalDifference;
772 Value2 = getSymbolAddress(B_SD, Layout);
773 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
776 // Relocations are written out in reverse order, so the PAIR comes first.
777 if (Type == macho::RIT_Difference ||
778 Type == macho::RIT_Generic_LocalDifference) {
779 macho::RelocationEntry MRE;
780 MRE.Word0 = ((0 << 0) |
781 (macho::RIT_Pair << 24) |
784 macho::RF_Scattered);
786 Relocations[Fragment->getParent()].push_back(MRE);
789 macho::RelocationEntry MRE;
790 MRE.Word0 = ((FixupOffset << 0) |
794 macho::RF_Scattered);
796 Relocations[Fragment->getParent()].push_back(MRE);
799 void RecordARMScatteredRelocation(const MCAssembler &Asm,
800 const MCAsmLayout &Layout,
801 const MCFragment *Fragment,
802 const MCFixup &Fixup, MCValue Target,
803 uint64_t &FixedValue) {
804 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
805 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
806 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
807 unsigned Type = macho::RIT_Vanilla;
810 const MCSymbol *A = &Target.getSymA()->getSymbol();
811 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
813 if (!A_SD->getFragment())
814 report_fatal_error("symbol '" + A->getName() +
815 "' can not be undefined in a subtraction expression");
817 uint32_t Value = getSymbolAddress(A_SD, Layout);
818 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
819 FixedValue += SecAddr;
822 if (const MCSymbolRefExpr *B = Target.getSymB()) {
823 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
825 if (!B_SD->getFragment())
826 report_fatal_error("symbol '" + B->getSymbol().getName() +
827 "' can not be undefined in a subtraction expression");
829 // Select the appropriate difference relocation type.
830 Type = macho::RIT_Difference;
831 Value2 = getSymbolAddress(B_SD, Layout);
832 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
835 // Relocations are written out in reverse order, so the PAIR comes first.
836 if (Type == macho::RIT_Difference ||
837 Type == macho::RIT_Generic_LocalDifference) {
838 macho::RelocationEntry MRE;
839 MRE.Word0 = ((0 << 0) |
840 (macho::RIT_Pair << 24) |
843 macho::RF_Scattered);
845 Relocations[Fragment->getParent()].push_back(MRE);
848 macho::RelocationEntry MRE;
849 MRE.Word0 = ((FixupOffset << 0) |
853 macho::RF_Scattered);
855 Relocations[Fragment->getParent()].push_back(MRE);
858 void RecordTLVPRelocation(const MCAssembler &Asm,
859 const MCAsmLayout &Layout,
860 const MCFragment *Fragment,
861 const MCFixup &Fixup, MCValue Target,
862 uint64_t &FixedValue) {
863 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
865 "Should only be called with a 32-bit TLVP relocation!");
867 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
868 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
869 unsigned IsPCRel = 0;
871 // Get the symbol data.
872 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
873 unsigned Index = SD_A->getIndex();
875 // We're only going to have a second symbol in pic mode and it'll be a
876 // subtraction from the picbase. For 32-bit pic the addend is the difference
877 // between the picbase and the next address. For 32-bit static the addend
879 if (Target.getSymB()) {
880 // If this is a subtraction then we're pcrel.
881 uint32_t FixupAddress =
882 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
883 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
885 FixedValue = (FixupAddress - getSymbolAddress(SD_B, Layout) +
886 Target.getConstant());
887 FixedValue += 1ULL << Log2Size;
892 // struct relocation_info (8 bytes)
893 macho::RelocationEntry MRE;
895 MRE.Word1 = ((Index << 0) |
898 (1 << 27) | // Extern
899 (macho::RIT_Generic_TLV << 28)); // Type
900 Relocations[Fragment->getParent()].push_back(MRE);
903 static bool getARMFixupKindMachOInfo(unsigned Kind, bool &Is24BitBranch,
904 unsigned &Log2Size) {
905 Is24BitBranch = false;
913 Log2Size = llvm::Log2_32(1);
916 Log2Size = llvm::Log2_32(2);
919 Log2Size = llvm::Log2_32(4);
922 Log2Size = llvm::Log2_32(8);
925 // Handle 24-bit branch kinds.
926 case ARM::fixup_arm_ldst_pcrel_12:
927 case ARM::fixup_arm_pcrel_10:
928 case ARM::fixup_arm_adr_pcrel_12:
929 case ARM::fixup_arm_branch:
930 Is24BitBranch = true;
931 // Report as 'long', even though that is not quite accurate.
932 Log2Size = llvm::Log2_32(4);
935 // Handle Thumb branches.
936 case ARM::fixup_arm_thumb_br:
937 Log2Size = llvm::Log2_32(2);
940 case ARM::fixup_arm_thumb_bl:
941 Log2Size = llvm::Log2_32(4);
944 case ARM::fixup_arm_thumb_blx:
945 // Report as 'long', even though that is not quite accurate.
946 Log2Size = llvm::Log2_32(4);
950 void RecordARMRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
951 const MCFragment *Fragment, const MCFixup &Fixup,
952 MCValue Target, uint64_t &FixedValue) {
953 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
956 if (!getARMFixupKindMachOInfo(Fixup.getKind(), Is24BitBranch, Log2Size)) {
957 report_fatal_error("unknown ARM fixup kind!");
961 // If this is a difference or a defined symbol plus an offset, then we need
962 // a scattered relocation entry. Differences always require scattered
964 if (Target.getSymB())
965 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup,
968 // Get the symbol data, if any.
969 MCSymbolData *SD = 0;
970 if (Target.getSymA())
971 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
973 // FIXME: For other platforms, we need to use scattered relocations for
974 // internal relocations with offsets. If this is an internal relocation
975 // with an offset, it also needs a scattered relocation entry.
977 // Is this right for ARM?
978 uint32_t Offset = Target.getConstant();
979 if (IsPCRel && !Is24BitBranch)
980 Offset += 1 << Log2Size;
981 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
982 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup,
986 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
988 unsigned IsExtern = 0;
991 if (Target.isAbsolute()) { // constant
993 report_fatal_error("FIXME: relocations to absolute targets "
994 "not yet implemented");
995 } else if (SD->getSymbol().isVariable()) {
997 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
998 Res, Layout, SectionAddress)) {
1003 report_fatal_error("unsupported relocation of variable '" +
1004 SD->getSymbol().getName() + "'");
1006 // Check whether we need an external or internal relocation.
1007 if (doesSymbolRequireExternRelocation(SD)) {
1009 Index = SD->getIndex();
1010 // For external relocations, make sure to offset the fixup value to
1011 // compensate for the addend of the symbol address, if it was
1012 // undefined. This occurs with weak definitions, for example.
1013 if (!SD->Symbol->isUndefined())
1014 FixedValue -= Layout.getSymbolOffset(SD);
1016 // The index is the section ordinal (1-based).
1017 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
1018 FixedValue += getSectionAddress(SD->getFragment()->getParent());
1021 FixedValue -= getSectionAddress(Fragment->getParent());
1023 // Determine the appropriate type based on the fixup kind.
1024 Type = Is24BitBranch ? (unsigned)macho::RIT_ARM_Branch24Bit :
1025 (unsigned)macho::RIT_Vanilla;
1028 // struct relocation_info (8 bytes)
1029 macho::RelocationEntry MRE;
1030 MRE.Word0 = FixupOffset;
1031 MRE.Word1 = ((Index << 0) |
1036 Relocations[Fragment->getParent()].push_back(MRE);
1039 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
1040 const MCFragment *Fragment, const MCFixup &Fixup,
1041 MCValue Target, uint64_t &FixedValue) {
1042 // FIXME: These needs to be factored into the target Mach-O writer.
1044 RecordARMRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1048 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1052 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
1053 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
1055 // If this is a 32-bit TLVP reloc it's handled a bit differently.
1056 if (Target.getSymA() &&
1057 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
1058 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1062 // If this is a difference or a defined symbol plus an offset, then we need
1063 // a scattered relocation entry.
1064 // Differences always require scattered relocations.
1065 if (Target.getSymB())
1066 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1067 Target, FixedValue);
1069 // Get the symbol data, if any.
1070 MCSymbolData *SD = 0;
1071 if (Target.getSymA())
1072 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1074 // If this is an internal relocation with an offset, it also needs a
1075 // scattered relocation entry.
1076 uint32_t Offset = Target.getConstant();
1078 Offset += 1 << Log2Size;
1079 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
1080 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1081 Target, FixedValue);
1084 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1086 unsigned IsExtern = 0;
1089 if (Target.isAbsolute()) { // constant
1090 // SymbolNum of 0 indicates the absolute section.
1092 // FIXME: Currently, these are never generated (see code below). I cannot
1093 // find a case where they are actually emitted.
1094 Type = macho::RIT_Vanilla;
1095 } else if (SD->getSymbol().isVariable()) {
1097 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
1098 Res, Layout, SectionAddress)) {
1103 report_fatal_error("unsupported relocation of variable '" +
1104 SD->getSymbol().getName() + "'");
1106 // Check whether we need an external or internal relocation.
1107 if (doesSymbolRequireExternRelocation(SD)) {
1109 Index = SD->getIndex();
1110 // For external relocations, make sure to offset the fixup value to
1111 // compensate for the addend of the symbol address, if it was
1112 // undefined. This occurs with weak definitions, for example.
1113 if (!SD->Symbol->isUndefined())
1114 FixedValue -= Layout.getSymbolOffset(SD);
1116 // The index is the section ordinal (1-based).
1117 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
1118 FixedValue += getSectionAddress(SD->getFragment()->getParent());
1121 FixedValue -= getSectionAddress(Fragment->getParent());
1123 Type = macho::RIT_Vanilla;
1126 // struct relocation_info (8 bytes)
1127 macho::RelocationEntry MRE;
1128 MRE.Word0 = FixupOffset;
1129 MRE.Word1 = ((Index << 0) |
1134 Relocations[Fragment->getParent()].push_back(MRE);
1137 void BindIndirectSymbols(MCAssembler &Asm) {
1138 // This is the point where 'as' creates actual symbols for indirect symbols
1139 // (in the following two passes). It would be easier for us to do this
1140 // sooner when we see the attribute, but that makes getting the order in the
1141 // symbol table much more complicated than it is worth.
1143 // FIXME: Revisit this when the dust settles.
1145 // Bind non lazy symbol pointers first.
1146 unsigned IndirectIndex = 0;
1147 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1148 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1149 const MCSectionMachO &Section =
1150 cast<MCSectionMachO>(it->SectionData->getSection());
1152 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
1155 // Initialize the section indirect symbol base, if necessary.
1156 if (!IndirectSymBase.count(it->SectionData))
1157 IndirectSymBase[it->SectionData] = IndirectIndex;
1159 Asm.getOrCreateSymbolData(*it->Symbol);
1162 // Then lazy symbol pointers and symbol stubs.
1164 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1165 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1166 const MCSectionMachO &Section =
1167 cast<MCSectionMachO>(it->SectionData->getSection());
1169 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
1170 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
1173 // Initialize the section indirect symbol base, if necessary.
1174 if (!IndirectSymBase.count(it->SectionData))
1175 IndirectSymBase[it->SectionData] = IndirectIndex;
1177 // Set the symbol type to undefined lazy, but only on construction.
1179 // FIXME: Do not hardcode.
1181 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
1183 Entry.setFlags(Entry.getFlags() | 0x0001);
1187 /// ComputeSymbolTable - Compute the symbol table data
1189 /// \param StringTable [out] - The string table data.
1190 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
1192 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
1193 std::vector<MachSymbolData> &LocalSymbolData,
1194 std::vector<MachSymbolData> &ExternalSymbolData,
1195 std::vector<MachSymbolData> &UndefinedSymbolData) {
1196 // Build section lookup table.
1197 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
1199 for (MCAssembler::iterator it = Asm.begin(),
1200 ie = Asm.end(); it != ie; ++it, ++Index)
1201 SectionIndexMap[&it->getSection()] = Index;
1202 assert(Index <= 256 && "Too many sections!");
1204 // Index 0 is always the empty string.
1205 StringMap<uint64_t> StringIndexMap;
1206 StringTable += '\x00';
1208 // Build the symbol arrays and the string table, but only for non-local
1211 // The particular order that we collect the symbols and create the string
1212 // table, then sort the symbols is chosen to match 'as'. Even though it
1213 // doesn't matter for correctness, this is important for letting us diff .o
1215 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1216 ie = Asm.symbol_end(); it != ie; ++it) {
1217 const MCSymbol &Symbol = it->getSymbol();
1219 // Ignore non-linker visible symbols.
1220 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1223 if (!it->isExternal() && !Symbol.isUndefined())
1226 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1228 Entry = StringTable.size();
1229 StringTable += Symbol.getName();
1230 StringTable += '\x00';
1234 MSD.SymbolData = it;
1235 MSD.StringIndex = Entry;
1237 if (Symbol.isUndefined()) {
1238 MSD.SectionIndex = 0;
1239 UndefinedSymbolData.push_back(MSD);
1240 } else if (Symbol.isAbsolute()) {
1241 MSD.SectionIndex = 0;
1242 ExternalSymbolData.push_back(MSD);
1244 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1245 assert(MSD.SectionIndex && "Invalid section index!");
1246 ExternalSymbolData.push_back(MSD);
1250 // Now add the data for local symbols.
1251 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1252 ie = Asm.symbol_end(); it != ie; ++it) {
1253 const MCSymbol &Symbol = it->getSymbol();
1255 // Ignore non-linker visible symbols.
1256 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1259 if (it->isExternal() || Symbol.isUndefined())
1262 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1264 Entry = StringTable.size();
1265 StringTable += Symbol.getName();
1266 StringTable += '\x00';
1270 MSD.SymbolData = it;
1271 MSD.StringIndex = Entry;
1273 if (Symbol.isAbsolute()) {
1274 MSD.SectionIndex = 0;
1275 LocalSymbolData.push_back(MSD);
1277 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1278 assert(MSD.SectionIndex && "Invalid section index!");
1279 LocalSymbolData.push_back(MSD);
1283 // External and undefined symbols are required to be in lexicographic order.
1284 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1285 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1287 // Set the symbol indices.
1289 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1290 LocalSymbolData[i].SymbolData->setIndex(Index++);
1291 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1292 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1293 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1294 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1296 // The string table is padded to a multiple of 4.
1297 while (StringTable.size() % 4)
1298 StringTable += '\x00';
1301 void computeSectionAddresses(const MCAssembler &Asm,
1302 const MCAsmLayout &Layout) {
1303 uint64_t StartAddress = 0;
1304 const SmallVectorImpl<MCSectionData*> &Order = Layout.getSectionOrder();
1305 for (int i = 0, n = Order.size(); i != n ; ++i) {
1306 const MCSectionData *SD = Order[i];
1307 StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment());
1308 SectionAddress[SD] = StartAddress;
1309 StartAddress += Layout.getSectionAddressSize(SD);
1310 // Explicitly pad the section to match the alignment requirements of the
1311 // following one. This is for 'gas' compatibility, it shouldn't
1312 /// strictly be necessary.
1313 StartAddress += getPaddingSize(SD, Layout);
1317 void ExecutePostLayoutBinding(MCAssembler &Asm, const MCAsmLayout &Layout) {
1318 computeSectionAddresses(Asm, Layout);
1320 // Create symbol data for any indirect symbols.
1321 BindIndirectSymbols(Asm);
1323 // Compute symbol table information and bind symbol indices.
1324 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1325 UndefinedSymbolData);
1328 bool IsSymbolRefDifferenceFullyResolved(const MCAssembler &Asm,
1329 const MCSymbolRefExpr *A,
1330 const MCSymbolRefExpr *B,
1335 if (!TargetObjectWriter->useAggressiveSymbolFolding())
1338 // The effective address is
1339 // addr(atom(A)) + offset(A)
1340 // - addr(atom(B)) - offset(B)
1341 // and the offsets are not relocatable, so the fixup is fully resolved when
1342 // addr(atom(A)) - addr(atom(B)) == 0.
1343 const MCSymbolData *A_Base = 0, *B_Base = 0;
1345 // Modified symbol references cannot be resolved.
1346 if (A->getKind() != MCSymbolRefExpr::VK_None ||
1347 B->getKind() != MCSymbolRefExpr::VK_None)
1350 A_Base = Asm.getAtom(&Asm.getSymbolData(A->getSymbol()));
1354 B_Base = Asm.getAtom(&Asm.getSymbolData(B->getSymbol()));
1358 // If the atoms are the same, they are guaranteed to have the same address.
1359 if (A_Base == B_Base)
1362 // Otherwise, we can't prove this is fully resolved.
1366 bool IsFixupFullyResolved(const MCAssembler &Asm,
1367 const MCValue Target,
1369 const MCFragment *DF) const {
1370 // Otherwise, determine whether this value is actually resolved; scattering
1371 // may cause atoms to move.
1373 // Check if we are using the "simple" resolution algorithm (e.g.,
1375 if (!Asm.getBackend().hasReliableSymbolDifference()) {
1376 const MCSection *BaseSection = 0;
1378 BaseSection = &DF->getParent()->getSection();
1380 return isScatteredFixupFullyResolvedSimple(Asm, Target, BaseSection);
1383 // Otherwise, compute the proper answer as reliably as possible.
1385 // If this is a PCrel relocation, find the base atom (identified by its
1386 // symbol) that the fixup value is relative to.
1387 const MCSymbolData *BaseSymbol = 0;
1389 BaseSymbol = DF->getAtom();
1394 return isScatteredFixupFullyResolved(Asm, Target, BaseSymbol);
1397 void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout) {
1398 unsigned NumSections = Asm.size();
1400 // The section data starts after the header, the segment load command (and
1401 // section headers) and the symbol table.
1402 unsigned NumLoadCommands = 1;
1403 uint64_t LoadCommandsSize = is64Bit() ?
1404 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
1405 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
1407 // Add the symbol table load command sizes, if used.
1408 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1409 UndefinedSymbolData.size();
1411 NumLoadCommands += 2;
1412 LoadCommandsSize += (macho::SymtabLoadCommandSize +
1413 macho::DysymtabLoadCommandSize);
1416 // Compute the total size of the section data, as well as its file size and
1418 uint64_t SectionDataStart = (is64Bit() ? macho::Header64Size :
1419 macho::Header32Size) + LoadCommandsSize;
1420 uint64_t SectionDataSize = 0;
1421 uint64_t SectionDataFileSize = 0;
1422 uint64_t VMSize = 0;
1423 for (MCAssembler::const_iterator it = Asm.begin(),
1424 ie = Asm.end(); it != ie; ++it) {
1425 const MCSectionData &SD = *it;
1426 uint64_t Address = getSectionAddress(&SD);
1427 uint64_t Size = Layout.getSectionAddressSize(&SD);
1428 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1429 FileSize += getPaddingSize(&SD, Layout);
1431 VMSize = std::max(VMSize, Address + Size);
1433 if (SD.getSection().isVirtualSection())
1436 SectionDataSize = std::max(SectionDataSize, Address + Size);
1437 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1440 // The section data is padded to 4 bytes.
1442 // FIXME: Is this machine dependent?
1443 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1444 SectionDataFileSize += SectionDataPadding;
1446 // Write the prolog, starting with the header and load command...
1447 WriteHeader(NumLoadCommands, LoadCommandsSize,
1448 Asm.getSubsectionsViaSymbols());
1449 WriteSegmentLoadCommand(NumSections, VMSize,
1450 SectionDataStart, SectionDataSize);
1452 // ... and then the section headers.
1453 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1454 for (MCAssembler::const_iterator it = Asm.begin(),
1455 ie = Asm.end(); it != ie; ++it) {
1456 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1457 unsigned NumRelocs = Relocs.size();
1458 uint64_t SectionStart = SectionDataStart + getSectionAddress(it);
1459 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1460 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
1463 // Write the symbol table load command, if used.
1465 unsigned FirstLocalSymbol = 0;
1466 unsigned NumLocalSymbols = LocalSymbolData.size();
1467 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1468 unsigned NumExternalSymbols = ExternalSymbolData.size();
1469 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1470 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1471 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1472 unsigned NumSymTabSymbols =
1473 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1474 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1475 uint64_t IndirectSymbolOffset = 0;
1477 // If used, the indirect symbols are written after the section data.
1478 if (NumIndirectSymbols)
1479 IndirectSymbolOffset = RelocTableEnd;
1481 // The symbol table is written after the indirect symbol data.
1482 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1484 // The string table is written after symbol table.
1485 uint64_t StringTableOffset =
1486 SymbolTableOffset + NumSymTabSymbols * (is64Bit() ? macho::Nlist64Size :
1487 macho::Nlist32Size);
1488 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1489 StringTableOffset, StringTable.size());
1491 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1492 FirstExternalSymbol, NumExternalSymbols,
1493 FirstUndefinedSymbol, NumUndefinedSymbols,
1494 IndirectSymbolOffset, NumIndirectSymbols);
1497 // Write the actual section data.
1498 for (MCAssembler::const_iterator it = Asm.begin(),
1499 ie = Asm.end(); it != ie; ++it) {
1500 Asm.WriteSectionData(it, Layout);
1502 uint64_t Pad = getPaddingSize(it, Layout);
1503 for (unsigned int i = 0; i < Pad; ++i)
1507 // Write the extra padding.
1508 WriteZeros(SectionDataPadding);
1510 // Write the relocation entries.
1511 for (MCAssembler::const_iterator it = Asm.begin(),
1512 ie = Asm.end(); it != ie; ++it) {
1513 // Write the section relocation entries, in reverse order to match 'as'
1514 // (approximately, the exact algorithm is more complicated than this).
1515 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1516 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1517 Write32(Relocs[e - i - 1].Word0);
1518 Write32(Relocs[e - i - 1].Word1);
1522 // Write the symbol table data, if used.
1524 // Write the indirect symbol entries.
1525 for (MCAssembler::const_indirect_symbol_iterator
1526 it = Asm.indirect_symbol_begin(),
1527 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1528 // Indirect symbols in the non lazy symbol pointer section have some
1529 // special handling.
1530 const MCSectionMachO &Section =
1531 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1532 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1533 // If this symbol is defined and internal, mark it as such.
1534 if (it->Symbol->isDefined() &&
1535 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1536 uint32_t Flags = macho::ISF_Local;
1537 if (it->Symbol->isAbsolute())
1538 Flags |= macho::ISF_Absolute;
1544 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1547 // FIXME: Check that offsets match computed ones.
1549 // Write the symbol table entries.
1550 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1551 WriteNlist(LocalSymbolData[i], Layout);
1552 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1553 WriteNlist(ExternalSymbolData[i], Layout);
1554 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1555 WriteNlist(UndefinedSymbolData[i], Layout);
1557 // Write the string table.
1558 OS << StringTable.str();
1565 MCObjectWriter *llvm::createMachObjectWriter(MCMachObjectTargetWriter *MOTW,
1567 bool IsLittleEndian) {
1568 return new MachObjectWriter(MOTW, OS, IsLittleEndian);