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 static unsigned getFixupKindLog2Size(unsigned Kind) {
33 default: llvm_unreachable("invalid fixup kind!");
34 case X86::reloc_pcrel_1byte:
35 case FK_Data_1: return 0;
36 case X86::reloc_pcrel_2byte:
37 case FK_Data_2: return 1;
38 case X86::reloc_pcrel_4byte:
39 case X86::reloc_riprel_4byte:
40 case X86::reloc_riprel_4byte_movq_load:
41 case FK_Data_4: return 2;
42 case FK_Data_8: return 3;
46 static bool isFixupKindPCRel(unsigned Kind) {
50 case X86::reloc_pcrel_1byte:
51 case X86::reloc_pcrel_2byte:
52 case X86::reloc_pcrel_4byte:
53 case X86::reloc_riprel_4byte:
54 case X86::reloc_riprel_4byte_movq_load:
59 static bool isFixupKindRIPRel(unsigned Kind) {
60 return Kind == X86::reloc_riprel_4byte ||
61 Kind == X86::reloc_riprel_4byte_movq_load;
64 static bool doesSymbolRequireExternRelocation(MCSymbolData *SD) {
65 // Undefined symbols are always extern.
66 if (SD->Symbol->isUndefined())
69 // References to weak definitions require external relocation entries; the
70 // definition may not always be the one in the same object file.
71 if (SD->getFlags() & SF_WeakDefinition)
74 // Otherwise, we can use an internal relocation.
80 class MachObjectWriterImpl {
81 // See <mach-o/loader.h>.
83 Header_Magic32 = 0xFEEDFACE,
84 Header_Magic64 = 0xFEEDFACF
90 SegmentLoadCommand32Size = 56,
91 SegmentLoadCommand64Size = 72,
94 SymtabLoadCommandSize = 24,
95 DysymtabLoadCommandSize = 80,
98 RelocationInfoSize = 8
101 enum HeaderFileType {
106 HF_SubsectionsViaSymbols = 0x2000
109 enum LoadCommandType {
116 // See <mach-o/nlist.h>.
117 enum SymbolTypeType {
118 STT_Undefined = 0x00,
123 enum SymbolTypeFlags {
124 // If any of these bits are set, then the entry is a stab entry number (see
125 // <mach-o/stab.h>. Otherwise the other masks apply.
126 STF_StabsEntryMask = 0xe0,
130 STF_PrivateExtern = 0x10
133 /// IndirectSymbolFlags - Flags for encoding special values in the indirect
135 enum IndirectSymbolFlags {
136 ISF_Local = 0x80000000,
137 ISF_Absolute = 0x40000000
140 /// RelocationFlags - Special flags for addresses.
141 enum RelocationFlags {
142 RF_Scattered = 0x80000000
145 enum RelocationInfoType {
149 RIT_PreboundLazyPointer = 3,
150 RIT_LocalDifference = 4,
154 /// X86_64 uses its own relocation types.
155 enum RelocationInfoTypeX86_64 {
156 RIT_X86_64_Unsigned = 0,
157 RIT_X86_64_Signed = 1,
158 RIT_X86_64_Branch = 2,
159 RIT_X86_64_GOTLoad = 3,
161 RIT_X86_64_Subtractor = 5,
162 RIT_X86_64_Signed1 = 6,
163 RIT_X86_64_Signed2 = 7,
164 RIT_X86_64_Signed4 = 8,
168 /// MachSymbolData - Helper struct for containing some precomputed information
170 struct MachSymbolData {
171 MCSymbolData *SymbolData;
172 uint64_t StringIndex;
173 uint8_t SectionIndex;
175 // Support lexicographic sorting.
176 bool operator<(const MachSymbolData &RHS) const {
177 return SymbolData->getSymbol().getName() <
178 RHS.SymbolData->getSymbol().getName();
182 /// @name Relocation Data
185 struct MachRelocationEntry {
190 llvm::DenseMap<const MCSectionData*,
191 std::vector<MachRelocationEntry> > Relocations;
192 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
195 /// @name Symbol Table Data
198 SmallString<256> StringTable;
199 std::vector<MachSymbolData> LocalSymbolData;
200 std::vector<MachSymbolData> ExternalSymbolData;
201 std::vector<MachSymbolData> UndefinedSymbolData;
205 MachObjectWriter *Writer;
209 unsigned Is64Bit : 1;
212 MachObjectWriterImpl(MachObjectWriter *_Writer, bool _Is64Bit)
213 : Writer(_Writer), OS(Writer->getStream()), Is64Bit(_Is64Bit) {
216 void Write8(uint8_t Value) { Writer->Write8(Value); }
217 void Write16(uint16_t Value) { Writer->Write16(Value); }
218 void Write32(uint32_t Value) { Writer->Write32(Value); }
219 void Write64(uint64_t Value) { Writer->Write64(Value); }
220 void WriteZeros(unsigned N) { Writer->WriteZeros(N); }
221 void WriteBytes(StringRef Str, unsigned ZeroFillSize = 0) {
222 Writer->WriteBytes(Str, ZeroFillSize);
225 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
226 bool SubsectionsViaSymbols) {
229 if (SubsectionsViaSymbols)
230 Flags |= HF_SubsectionsViaSymbols;
232 // struct mach_header (28 bytes) or
233 // struct mach_header_64 (32 bytes)
235 uint64_t Start = OS.tell();
238 Write32(Is64Bit ? Header_Magic64 : Header_Magic32);
240 // FIXME: Support cputype.
241 Write32(Is64Bit ? MachO::CPUTypeX86_64 : MachO::CPUTypeI386);
242 // FIXME: Support cpusubtype.
243 Write32(MachO::CPUSubType_I386_ALL);
245 Write32(NumLoadCommands); // Object files have a single load command, the
247 Write32(LoadCommandsSize);
250 Write32(0); // reserved
252 assert(OS.tell() - Start == Is64Bit ? Header64Size : Header32Size);
255 /// WriteSegmentLoadCommand - Write a segment load command.
257 /// \arg NumSections - The number of sections in this segment.
258 /// \arg SectionDataSize - The total size of the sections.
259 void WriteSegmentLoadCommand(unsigned NumSections,
261 uint64_t SectionDataStartOffset,
262 uint64_t SectionDataSize) {
263 // struct segment_command (56 bytes) or
264 // struct segment_command_64 (72 bytes)
266 uint64_t Start = OS.tell();
269 unsigned SegmentLoadCommandSize = Is64Bit ? SegmentLoadCommand64Size :
270 SegmentLoadCommand32Size;
271 Write32(Is64Bit ? LCT_Segment64 : LCT_Segment);
272 Write32(SegmentLoadCommandSize +
273 NumSections * (Is64Bit ? Section64Size : Section32Size));
277 Write64(0); // vmaddr
278 Write64(VMSize); // vmsize
279 Write64(SectionDataStartOffset); // file offset
280 Write64(SectionDataSize); // file size
282 Write32(0); // vmaddr
283 Write32(VMSize); // vmsize
284 Write32(SectionDataStartOffset); // file offset
285 Write32(SectionDataSize); // file size
287 Write32(0x7); // maxprot
288 Write32(0x7); // initprot
289 Write32(NumSections);
292 assert(OS.tell() - Start == SegmentLoadCommandSize);
295 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
296 const MCSectionData &SD, uint64_t FileOffset,
297 uint64_t RelocationsStart, unsigned NumRelocations) {
298 uint64_t SectionSize = Layout.getSectionSize(&SD);
300 // The offset is unused for virtual sections.
301 if (Asm.getBackend().isVirtualSection(SD.getSection())) {
302 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
306 // struct section (68 bytes) or
307 // struct section_64 (80 bytes)
309 uint64_t Start = OS.tell();
312 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
313 WriteBytes(Section.getSectionName(), 16);
314 WriteBytes(Section.getSegmentName(), 16);
316 Write64(Layout.getSectionAddress(&SD)); // address
317 Write64(SectionSize); // size
319 Write32(Layout.getSectionAddress(&SD)); // address
320 Write32(SectionSize); // size
324 unsigned Flags = Section.getTypeAndAttributes();
325 if (SD.hasInstructions())
326 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
328 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
329 Write32(Log2_32(SD.getAlignment()));
330 Write32(NumRelocations ? RelocationsStart : 0);
331 Write32(NumRelocations);
333 Write32(IndirectSymBase.lookup(&SD)); // reserved1
334 Write32(Section.getStubSize()); // reserved2
336 Write32(0); // reserved3
338 assert(OS.tell() - Start == Is64Bit ? Section64Size : Section32Size);
341 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
342 uint32_t StringTableOffset,
343 uint32_t StringTableSize) {
344 // struct symtab_command (24 bytes)
346 uint64_t Start = OS.tell();
350 Write32(SymtabLoadCommandSize);
351 Write32(SymbolOffset);
353 Write32(StringTableOffset);
354 Write32(StringTableSize);
356 assert(OS.tell() - Start == SymtabLoadCommandSize);
359 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
360 uint32_t NumLocalSymbols,
361 uint32_t FirstExternalSymbol,
362 uint32_t NumExternalSymbols,
363 uint32_t FirstUndefinedSymbol,
364 uint32_t NumUndefinedSymbols,
365 uint32_t IndirectSymbolOffset,
366 uint32_t NumIndirectSymbols) {
367 // struct dysymtab_command (80 bytes)
369 uint64_t Start = OS.tell();
372 Write32(LCT_Dysymtab);
373 Write32(DysymtabLoadCommandSize);
374 Write32(FirstLocalSymbol);
375 Write32(NumLocalSymbols);
376 Write32(FirstExternalSymbol);
377 Write32(NumExternalSymbols);
378 Write32(FirstUndefinedSymbol);
379 Write32(NumUndefinedSymbols);
380 Write32(0); // tocoff
382 Write32(0); // modtaboff
383 Write32(0); // nmodtab
384 Write32(0); // extrefsymoff
385 Write32(0); // nextrefsyms
386 Write32(IndirectSymbolOffset);
387 Write32(NumIndirectSymbols);
388 Write32(0); // extreloff
389 Write32(0); // nextrel
390 Write32(0); // locreloff
391 Write32(0); // nlocrel
393 assert(OS.tell() - Start == DysymtabLoadCommandSize);
396 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
397 MCSymbolData &Data = *MSD.SymbolData;
398 const MCSymbol &Symbol = Data.getSymbol();
400 uint16_t Flags = Data.getFlags();
401 uint32_t Address = 0;
403 // Set the N_TYPE bits. See <mach-o/nlist.h>.
405 // FIXME: Are the prebound or indirect fields possible here?
406 if (Symbol.isUndefined())
407 Type = STT_Undefined;
408 else if (Symbol.isAbsolute())
413 // FIXME: Set STAB bits.
415 if (Data.isPrivateExtern())
416 Type |= STF_PrivateExtern;
419 if (Data.isExternal() || Symbol.isUndefined())
420 Type |= STF_External;
422 // Compute the symbol address.
423 if (Symbol.isDefined()) {
424 if (Symbol.isAbsolute()) {
425 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
427 Address = Layout.getSymbolAddress(&Data);
429 } else if (Data.isCommon()) {
430 // Common symbols are encoded with the size in the address
431 // field, and their alignment in the flags.
432 Address = Data.getCommonSize();
434 // Common alignment is packed into the 'desc' bits.
435 if (unsigned Align = Data.getCommonAlignment()) {
436 unsigned Log2Size = Log2_32(Align);
437 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
439 report_fatal_error("invalid 'common' alignment '" +
441 // FIXME: Keep this mask with the SymbolFlags enumeration.
442 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
446 // struct nlist (12 bytes)
448 Write32(MSD.StringIndex);
450 Write8(MSD.SectionIndex);
452 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
461 // FIXME: We really need to improve the relocation validation. Basically, we
462 // want to implement a separate computation which evaluates the relocation
463 // entry as the linker would, and verifies that the resultant fixup value is
464 // exactly what the encoder wanted. This will catch several classes of
467 // - Relocation entry bugs, the two algorithms are unlikely to have the same
470 // - Relaxation issues, where we forget to relax something.
472 // - Input errors, where something cannot be correctly encoded. 'as' allows
473 // these through in many cases.
475 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
476 const MCFragment *Fragment,
477 const MCFixup &Fixup, MCValue Target,
478 uint64_t &FixedValue) {
479 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
480 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
481 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
484 uint32_t FixupOffset =
485 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
486 uint32_t FixupAddress =
487 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
490 unsigned IsExtern = 0;
493 Value = Target.getConstant();
496 // Compensate for the relocation offset, Darwin x86_64 relocations only
497 // have the addend and appear to have attempted to define it to be the
498 // actual expression addend without the PCrel bias. However, instructions
499 // with data following the relocation are not accomodated for (see comment
500 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
501 Value += 1LL << Log2Size;
504 if (Target.isAbsolute()) { // constant
505 // SymbolNum of 0 indicates the absolute section.
506 Type = RIT_X86_64_Unsigned;
509 // FIXME: I believe this is broken, I don't think the linker can
510 // understand it. I think it would require a local relocation, but I'm not
511 // sure if that would work either. The official way to get an absolute
512 // PCrel relocation is to use an absolute symbol (which we don't support
516 Type = RIT_X86_64_Branch;
518 } else if (Target.getSymB()) { // A - B + constant
519 const MCSymbol *A = &Target.getSymA()->getSymbol();
520 MCSymbolData &A_SD = Asm.getSymbolData(*A);
521 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
523 const MCSymbol *B = &Target.getSymB()->getSymbol();
524 MCSymbolData &B_SD = Asm.getSymbolData(*B);
525 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
527 // Neither symbol can be modified.
528 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
529 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
530 report_fatal_error("unsupported relocation of modified symbol");
532 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
533 // implement most of these correctly.
535 report_fatal_error("unsupported pc-relative relocation of difference");
537 // We don't currently support any situation where one or both of the
538 // symbols would require a local relocation. This is almost certainly
539 // unused and may not be possible to encode correctly.
540 if (!A_Base || !B_Base)
541 report_fatal_error("unsupported local relocations in difference");
543 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
544 // a single SIGNED relocation); reject it for now.
545 if (A_Base == B_Base)
546 report_fatal_error("unsupported relocation with identical base");
548 Value += Layout.getSymbolAddress(&A_SD) - Layout.getSymbolAddress(A_Base);
549 Value -= Layout.getSymbolAddress(&B_SD) - Layout.getSymbolAddress(B_Base);
551 Index = A_Base->getIndex();
553 Type = RIT_X86_64_Unsigned;
555 MachRelocationEntry MRE;
556 MRE.Word0 = FixupOffset;
557 MRE.Word1 = ((Index << 0) |
562 Relocations[Fragment->getParent()].push_back(MRE);
564 Index = B_Base->getIndex();
566 Type = RIT_X86_64_Subtractor;
568 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
569 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
570 const MCSymbolData *Base = Asm.getAtom(&SD);
572 // Relocations inside debug sections always use local relocations when
573 // possible. This seems to be done because the debugger doesn't fully
574 // understand x86_64 relocation entries, and expects to find values that
575 // have already been fixed up.
576 if (Symbol->isInSection()) {
577 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
578 Fragment->getParent()->getSection());
579 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
583 // x86_64 almost always uses external relocations, except when there is no
584 // symbol to use as a base address (a local symbol with no preceeding
585 // non-local symbol).
587 Index = Base->getIndex();
590 // Add the local offset, if needed.
592 Value += Layout.getSymbolAddress(&SD) - Layout.getSymbolAddress(Base);
593 } else if (Symbol->isInSection()) {
594 // The index is the section ordinal (1-based).
595 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
597 Value += Layout.getSymbolAddress(&SD);
600 Value -= FixupAddress + (1 << Log2Size);
602 report_fatal_error("unsupported relocation of undefined symbol '" +
603 Symbol->getName() + "'");
606 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
609 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
610 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
611 // rewrite the movq to an leaq at link time if the symbol ends up in
612 // the same linkage unit.
613 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
614 Type = RIT_X86_64_GOTLoad;
616 Type = RIT_X86_64_GOT;
617 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
618 Type = RIT_X86_64_TLV;
619 } else if (Modifier != MCSymbolRefExpr::VK_None) {
620 report_fatal_error("unsupported symbol modifier in relocation");
622 Type = RIT_X86_64_Signed;
624 // The Darwin x86_64 relocation format has a problem where it cannot
625 // encode an address (L<foo> + <constant>) which is outside the atom
626 // containing L<foo>. Generally, this shouldn't occur but it does
627 // happen when we have a RIPrel instruction with data following the
628 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
629 // adjustment Darwin x86_64 uses, the offset is still negative and
630 // the linker has no way to recognize this.
632 // To work around this, Darwin uses several special relocation types
633 // to indicate the offsets. However, the specification or
634 // implementation of these seems to also be incomplete; they should
635 // adjust the addend as well based on the actual encoded instruction
636 // (the additional bias), but instead appear to just look at the
638 switch (-(Target.getConstant() + (1LL << Log2Size))) {
639 case 1: Type = RIT_X86_64_Signed1; break;
640 case 2: Type = RIT_X86_64_Signed2; break;
641 case 4: Type = RIT_X86_64_Signed4; break;
645 if (Modifier != MCSymbolRefExpr::VK_None)
646 report_fatal_error("unsupported symbol modifier in branch "
649 Type = RIT_X86_64_Branch;
652 if (Modifier == MCSymbolRefExpr::VK_GOT) {
653 Type = RIT_X86_64_GOT;
654 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
655 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
656 // which case all we do is set the PCrel bit in the relocation entry;
657 // this is used with exception handling, for example. The source is
658 // required to include any necessary offset directly.
659 Type = RIT_X86_64_GOT;
661 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
662 report_fatal_error("TLVP symbol modifier should have been rip-rel");
663 } else if (Modifier != MCSymbolRefExpr::VK_None)
664 report_fatal_error("unsupported symbol modifier in relocation");
666 Type = RIT_X86_64_Unsigned;
670 // x86_64 always writes custom values into the fixups.
673 // struct relocation_info (8 bytes)
674 MachRelocationEntry MRE;
675 MRE.Word0 = FixupOffset;
676 MRE.Word1 = ((Index << 0) |
681 Relocations[Fragment->getParent()].push_back(MRE);
684 void RecordScatteredRelocation(const MCAssembler &Asm,
685 const MCAsmLayout &Layout,
686 const MCFragment *Fragment,
687 const MCFixup &Fixup, MCValue Target,
688 uint64_t &FixedValue) {
689 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
690 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
691 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
692 unsigned Type = RIT_Vanilla;
695 const MCSymbol *A = &Target.getSymA()->getSymbol();
696 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
698 if (!A_SD->getFragment())
699 report_fatal_error("symbol '" + A->getName() +
700 "' can not be undefined in a subtraction expression");
702 uint32_t Value = Layout.getSymbolAddress(A_SD);
705 if (const MCSymbolRefExpr *B = Target.getSymB()) {
706 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
708 if (!B_SD->getFragment())
709 report_fatal_error("symbol '" + B->getSymbol().getName() +
710 "' can not be undefined in a subtraction expression");
712 // Select the appropriate difference relocation type.
714 // Note that there is no longer any semantic difference between these two
715 // relocation types from the linkers point of view, this is done solely
716 // for pedantic compatibility with 'as'.
717 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
718 Value2 = Layout.getSymbolAddress(B_SD);
721 // Relocations are written out in reverse order, so the PAIR comes first.
722 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
723 MachRelocationEntry MRE;
724 MRE.Word0 = ((0 << 0) |
730 Relocations[Fragment->getParent()].push_back(MRE);
733 MachRelocationEntry MRE;
734 MRE.Word0 = ((FixupOffset << 0) |
740 Relocations[Fragment->getParent()].push_back(MRE);
743 void RecordTLVPRelocation(const MCAssembler &Asm,
744 const MCAsmLayout &Layout,
745 const MCFragment *Fragment,
746 const MCFixup &Fixup, MCValue Target,
747 uint64_t &FixedValue) {
748 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
750 "Should only be called with a 32-bit TLVP relocation!");
752 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
753 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
754 unsigned IsPCRel = 0;
756 // Get the symbol data.
757 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
758 unsigned Index = SD_A->getIndex();
760 // We're only going to have a second symbol in pic mode and it'll be a
761 // subtraction from the picbase. For 32-bit pic the addend is the difference
762 // between the picbase and the next address. For 32-bit static the addend
764 if (Target.getSymB()) {
765 // If this is a subtraction then we're pcrel.
766 uint32_t FixupAddress =
767 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
768 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
770 FixedValue = (FixupAddress - Layout.getSymbolAddress(SD_B) +
771 Target.getConstant());
772 FixedValue += 1ULL << Log2Size;
777 // struct relocation_info (8 bytes)
778 MachRelocationEntry MRE;
780 MRE.Word1 = ((Index << 0) |
783 (1 << 27) | // Extern
784 (RIT_TLV << 28)); // Type
785 Relocations[Fragment->getParent()].push_back(MRE);
788 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
789 const MCFragment *Fragment, const MCFixup &Fixup,
790 MCValue Target, uint64_t &FixedValue) {
792 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
796 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
797 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
799 // If this is a 32-bit TLVP reloc it's handled a bit differently.
800 if (Target.getSymA() &&
801 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
802 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
806 // If this is a difference or a defined symbol plus an offset, then we need
807 // a scattered relocation entry.
808 // Differences always require scattered relocations.
809 if (Target.getSymB())
810 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
813 // Get the symbol data, if any.
814 MCSymbolData *SD = 0;
815 if (Target.getSymA())
816 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
818 // If this is an internal relocation with an offset, it also needs a
819 // scattered relocation entry.
820 uint32_t Offset = Target.getConstant();
822 Offset += 1 << Log2Size;
823 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
824 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
828 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
830 unsigned IsExtern = 0;
833 if (Target.isAbsolute()) { // constant
834 // SymbolNum of 0 indicates the absolute section.
836 // FIXME: Currently, these are never generated (see code below). I cannot
837 // find a case where they are actually emitted.
840 // Check whether we need an external or internal relocation.
841 if (doesSymbolRequireExternRelocation(SD)) {
843 Index = SD->getIndex();
844 // For external relocations, make sure to offset the fixup value to
845 // compensate for the addend of the symbol address, if it was
846 // undefined. This occurs with weak definitions, for example.
847 if (!SD->Symbol->isUndefined())
848 FixedValue -= Layout.getSymbolAddress(SD);
850 // The index is the section ordinal (1-based).
851 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
857 // struct relocation_info (8 bytes)
858 MachRelocationEntry MRE;
859 MRE.Word0 = FixupOffset;
860 MRE.Word1 = ((Index << 0) |
865 Relocations[Fragment->getParent()].push_back(MRE);
868 void BindIndirectSymbols(MCAssembler &Asm) {
869 // This is the point where 'as' creates actual symbols for indirect symbols
870 // (in the following two passes). It would be easier for us to do this
871 // sooner when we see the attribute, but that makes getting the order in the
872 // symbol table much more complicated than it is worth.
874 // FIXME: Revisit this when the dust settles.
876 // Bind non lazy symbol pointers first.
877 unsigned IndirectIndex = 0;
878 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
879 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
880 const MCSectionMachO &Section =
881 cast<MCSectionMachO>(it->SectionData->getSection());
883 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
886 // Initialize the section indirect symbol base, if necessary.
887 if (!IndirectSymBase.count(it->SectionData))
888 IndirectSymBase[it->SectionData] = IndirectIndex;
890 Asm.getOrCreateSymbolData(*it->Symbol);
893 // Then lazy symbol pointers and symbol stubs.
895 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
896 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
897 const MCSectionMachO &Section =
898 cast<MCSectionMachO>(it->SectionData->getSection());
900 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
901 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
904 // Initialize the section indirect symbol base, if necessary.
905 if (!IndirectSymBase.count(it->SectionData))
906 IndirectSymBase[it->SectionData] = IndirectIndex;
908 // Set the symbol type to undefined lazy, but only on construction.
910 // FIXME: Do not hardcode.
912 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
914 Entry.setFlags(Entry.getFlags() | 0x0001);
918 /// ComputeSymbolTable - Compute the symbol table data
920 /// \param StringTable [out] - The string table data.
921 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
923 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
924 std::vector<MachSymbolData> &LocalSymbolData,
925 std::vector<MachSymbolData> &ExternalSymbolData,
926 std::vector<MachSymbolData> &UndefinedSymbolData) {
927 // Build section lookup table.
928 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
930 for (MCAssembler::iterator it = Asm.begin(),
931 ie = Asm.end(); it != ie; ++it, ++Index)
932 SectionIndexMap[&it->getSection()] = Index;
933 assert(Index <= 256 && "Too many sections!");
935 // Index 0 is always the empty string.
936 StringMap<uint64_t> StringIndexMap;
937 StringTable += '\x00';
939 // Build the symbol arrays and the string table, but only for non-local
942 // The particular order that we collect the symbols and create the string
943 // table, then sort the symbols is chosen to match 'as'. Even though it
944 // doesn't matter for correctness, this is important for letting us diff .o
946 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
947 ie = Asm.symbol_end(); it != ie; ++it) {
948 const MCSymbol &Symbol = it->getSymbol();
950 // Ignore non-linker visible symbols.
951 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
954 if (!it->isExternal() && !Symbol.isUndefined())
957 uint64_t &Entry = StringIndexMap[Symbol.getName()];
959 Entry = StringTable.size();
960 StringTable += Symbol.getName();
961 StringTable += '\x00';
966 MSD.StringIndex = Entry;
968 if (Symbol.isUndefined()) {
969 MSD.SectionIndex = 0;
970 UndefinedSymbolData.push_back(MSD);
971 } else if (Symbol.isAbsolute()) {
972 MSD.SectionIndex = 0;
973 ExternalSymbolData.push_back(MSD);
975 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
976 assert(MSD.SectionIndex && "Invalid section index!");
977 ExternalSymbolData.push_back(MSD);
981 // Now add the data for local symbols.
982 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
983 ie = Asm.symbol_end(); it != ie; ++it) {
984 const MCSymbol &Symbol = it->getSymbol();
986 // Ignore non-linker visible symbols.
987 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
990 if (it->isExternal() || Symbol.isUndefined())
993 uint64_t &Entry = StringIndexMap[Symbol.getName()];
995 Entry = StringTable.size();
996 StringTable += Symbol.getName();
997 StringTable += '\x00';
1001 MSD.SymbolData = it;
1002 MSD.StringIndex = Entry;
1004 if (Symbol.isAbsolute()) {
1005 MSD.SectionIndex = 0;
1006 LocalSymbolData.push_back(MSD);
1008 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1009 assert(MSD.SectionIndex && "Invalid section index!");
1010 LocalSymbolData.push_back(MSD);
1014 // External and undefined symbols are required to be in lexicographic order.
1015 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1016 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1018 // Set the symbol indices.
1020 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1021 LocalSymbolData[i].SymbolData->setIndex(Index++);
1022 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1023 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1024 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1025 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1027 // The string table is padded to a multiple of 4.
1028 while (StringTable.size() % 4)
1029 StringTable += '\x00';
1032 void ExecutePostLayoutBinding(MCAssembler &Asm) {
1033 // Create symbol data for any indirect symbols.
1034 BindIndirectSymbols(Asm);
1036 // Compute symbol table information and bind symbol indices.
1037 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1038 UndefinedSymbolData);
1041 void WriteObject(const MCAssembler &Asm, const MCAsmLayout &Layout) {
1042 unsigned NumSections = Asm.size();
1044 // The section data starts after the header, the segment load command (and
1045 // section headers) and the symbol table.
1046 unsigned NumLoadCommands = 1;
1047 uint64_t LoadCommandsSize = Is64Bit ?
1048 SegmentLoadCommand64Size + NumSections * Section64Size :
1049 SegmentLoadCommand32Size + NumSections * Section32Size;
1051 // Add the symbol table load command sizes, if used.
1052 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1053 UndefinedSymbolData.size();
1055 NumLoadCommands += 2;
1056 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
1059 // Compute the total size of the section data, as well as its file size and
1061 uint64_t SectionDataStart = (Is64Bit ? Header64Size : Header32Size)
1063 uint64_t SectionDataSize = 0;
1064 uint64_t SectionDataFileSize = 0;
1065 uint64_t VMSize = 0;
1066 for (MCAssembler::const_iterator it = Asm.begin(),
1067 ie = Asm.end(); it != ie; ++it) {
1068 const MCSectionData &SD = *it;
1069 uint64_t Address = Layout.getSectionAddress(&SD);
1070 uint64_t Size = Layout.getSectionSize(&SD);
1071 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1073 VMSize = std::max(VMSize, Address + Size);
1075 if (Asm.getBackend().isVirtualSection(SD.getSection()))
1078 SectionDataSize = std::max(SectionDataSize, Address + Size);
1079 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1082 // The section data is padded to 4 bytes.
1084 // FIXME: Is this machine dependent?
1085 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1086 SectionDataFileSize += SectionDataPadding;
1088 // Write the prolog, starting with the header and load command...
1089 WriteHeader(NumLoadCommands, LoadCommandsSize,
1090 Asm.getSubsectionsViaSymbols());
1091 WriteSegmentLoadCommand(NumSections, VMSize,
1092 SectionDataStart, SectionDataSize);
1094 // ... and then the section headers.
1095 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1096 for (MCAssembler::const_iterator it = Asm.begin(),
1097 ie = Asm.end(); it != ie; ++it) {
1098 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1099 unsigned NumRelocs = Relocs.size();
1100 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
1101 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1102 RelocTableEnd += NumRelocs * RelocationInfoSize;
1105 // Write the symbol table load command, if used.
1107 unsigned FirstLocalSymbol = 0;
1108 unsigned NumLocalSymbols = LocalSymbolData.size();
1109 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1110 unsigned NumExternalSymbols = ExternalSymbolData.size();
1111 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1112 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1113 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1114 unsigned NumSymTabSymbols =
1115 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1116 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1117 uint64_t IndirectSymbolOffset = 0;
1119 // If used, the indirect symbols are written after the section data.
1120 if (NumIndirectSymbols)
1121 IndirectSymbolOffset = RelocTableEnd;
1123 // The symbol table is written after the indirect symbol data.
1124 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1126 // The string table is written after symbol table.
1127 uint64_t StringTableOffset =
1128 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? Nlist64Size :
1130 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1131 StringTableOffset, StringTable.size());
1133 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1134 FirstExternalSymbol, NumExternalSymbols,
1135 FirstUndefinedSymbol, NumUndefinedSymbols,
1136 IndirectSymbolOffset, NumIndirectSymbols);
1139 // Write the actual section data.
1140 for (MCAssembler::const_iterator it = Asm.begin(),
1141 ie = Asm.end(); it != ie; ++it)
1142 Asm.WriteSectionData(it, Layout, Writer);
1144 // Write the extra padding.
1145 WriteZeros(SectionDataPadding);
1147 // Write the relocation entries.
1148 for (MCAssembler::const_iterator it = Asm.begin(),
1149 ie = Asm.end(); it != ie; ++it) {
1150 // Write the section relocation entries, in reverse order to match 'as'
1151 // (approximately, the exact algorithm is more complicated than this).
1152 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
1153 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1154 Write32(Relocs[e - i - 1].Word0);
1155 Write32(Relocs[e - i - 1].Word1);
1159 // Write the symbol table data, if used.
1161 // Write the indirect symbol entries.
1162 for (MCAssembler::const_indirect_symbol_iterator
1163 it = Asm.indirect_symbol_begin(),
1164 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1165 // Indirect symbols in the non lazy symbol pointer section have some
1166 // special handling.
1167 const MCSectionMachO &Section =
1168 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1169 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1170 // If this symbol is defined and internal, mark it as such.
1171 if (it->Symbol->isDefined() &&
1172 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1173 uint32_t Flags = ISF_Local;
1174 if (it->Symbol->isAbsolute())
1175 Flags |= ISF_Absolute;
1181 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1184 // FIXME: Check that offsets match computed ones.
1186 // Write the symbol table entries.
1187 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1188 WriteNlist(LocalSymbolData[i], Layout);
1189 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1190 WriteNlist(ExternalSymbolData[i], Layout);
1191 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1192 WriteNlist(UndefinedSymbolData[i], Layout);
1194 // Write the string table.
1195 OS << StringTable.str();
1202 MachObjectWriter::MachObjectWriter(raw_ostream &OS,
1204 bool IsLittleEndian)
1205 : MCObjectWriter(OS, IsLittleEndian)
1207 Impl = new MachObjectWriterImpl(this, Is64Bit);
1210 MachObjectWriter::~MachObjectWriter() {
1211 delete (MachObjectWriterImpl*) Impl;
1214 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm) {
1215 ((MachObjectWriterImpl*) Impl)->ExecutePostLayoutBinding(Asm);
1218 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1219 const MCAsmLayout &Layout,
1220 const MCFragment *Fragment,
1221 const MCFixup &Fixup, MCValue Target,
1222 uint64_t &FixedValue) {
1223 ((MachObjectWriterImpl*) Impl)->RecordRelocation(Asm, Layout, Fragment, Fixup,
1224 Target, FixedValue);
1227 void MachObjectWriter::WriteObject(const MCAssembler &Asm,
1228 const MCAsmLayout &Layout) {
1229 ((MachObjectWriterImpl*) Impl)->WriteObject(Asm, Layout);