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/ADT/StringMap.h"
11 #include "llvm/ADT/Twine.h"
12 #include "llvm/MC/MCAssembler.h"
13 #include "llvm/MC/MCAsmLayout.h"
14 #include "llvm/MC/MCExpr.h"
15 #include "llvm/MC/MCObjectWriter.h"
16 #include "llvm/MC/MCSectionMachO.h"
17 #include "llvm/MC/MCSymbol.h"
18 #include "llvm/MC/MCMachOSymbolFlags.h"
19 #include "llvm/MC/MCValue.h"
20 #include "llvm/Object/MachOFormat.h"
21 #include "llvm/Support/ErrorHandling.h"
22 #include "llvm/Target/TargetAsmBackend.h"
25 #include "../Target/X86/X86FixupKinds.h"
29 using namespace llvm::object;
31 // FIXME: this has been copied from (or to) X86AsmBackend.cpp
32 static unsigned getFixupKindLog2Size(unsigned Kind) {
34 default: llvm_unreachable("invalid fixup kind!");
36 case FK_Data_1: return 0;
38 case FK_Data_2: return 1;
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) {
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 MachObjectWriter : public MCObjectWriter {
163 /// MachSymbolData - Helper struct for containing some precomputed information
165 struct MachSymbolData {
166 MCSymbolData *SymbolData;
167 uint64_t StringIndex;
168 uint8_t SectionIndex;
170 // Support lexicographic sorting.
171 bool operator<(const MachSymbolData &RHS) const {
172 return SymbolData->getSymbol().getName() <
173 RHS.SymbolData->getSymbol().getName();
177 /// @name Relocation Data
180 llvm::DenseMap<const MCSectionData*,
181 std::vector<macho::RelocationEntry> > Relocations;
182 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
185 /// @name Symbol Table Data
188 SmallString<256> StringTable;
189 std::vector<MachSymbolData> LocalSymbolData;
190 std::vector<MachSymbolData> ExternalSymbolData;
191 std::vector<MachSymbolData> UndefinedSymbolData;
195 unsigned Is64Bit : 1;
201 MachObjectWriter(raw_ostream &_OS,
202 bool _Is64Bit, uint32_t _CPUType, uint32_t _CPUSubtype,
203 bool _IsLittleEndian)
204 : MCObjectWriter(_OS, _IsLittleEndian),
205 Is64Bit(_Is64Bit), CPUType(_CPUType), CPUSubtype(_CPUSubtype) {
208 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
209 bool SubsectionsViaSymbols) {
212 if (SubsectionsViaSymbols)
213 Flags |= macho::HF_SubsectionsViaSymbols;
215 // struct mach_header (28 bytes) or
216 // struct mach_header_64 (32 bytes)
218 uint64_t Start = OS.tell();
221 Write32(Is64Bit ? macho::HM_Object64 : macho::HM_Object32);
226 Write32(macho::HFT_Object);
227 Write32(NumLoadCommands);
228 Write32(LoadCommandsSize);
231 Write32(0); // reserved
233 assert(OS.tell() - Start == Is64Bit ?
234 macho::Header64Size : macho::Header32Size);
237 /// WriteSegmentLoadCommand - Write a segment load command.
239 /// \arg NumSections - The number of sections in this segment.
240 /// \arg SectionDataSize - The total size of the sections.
241 void WriteSegmentLoadCommand(unsigned NumSections,
243 uint64_t SectionDataStartOffset,
244 uint64_t SectionDataSize) {
245 // struct segment_command (56 bytes) or
246 // struct segment_command_64 (72 bytes)
248 uint64_t Start = OS.tell();
251 unsigned SegmentLoadCommandSize = Is64Bit ? macho::SegmentLoadCommand64Size:
252 macho::SegmentLoadCommand32Size;
253 Write32(Is64Bit ? macho::LCT_Segment64 : macho::LCT_Segment);
254 Write32(SegmentLoadCommandSize +
255 NumSections * (Is64Bit ? macho::Section64Size :
256 macho::Section32Size));
260 Write64(0); // vmaddr
261 Write64(VMSize); // vmsize
262 Write64(SectionDataStartOffset); // file offset
263 Write64(SectionDataSize); // file size
265 Write32(0); // vmaddr
266 Write32(VMSize); // vmsize
267 Write32(SectionDataStartOffset); // file offset
268 Write32(SectionDataSize); // file size
270 Write32(0x7); // maxprot
271 Write32(0x7); // initprot
272 Write32(NumSections);
275 assert(OS.tell() - Start == SegmentLoadCommandSize);
278 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
279 const MCSectionData &SD, uint64_t FileOffset,
280 uint64_t RelocationsStart, unsigned NumRelocations) {
281 uint64_t SectionSize = Layout.getSectionSize(&SD);
283 // The offset is unused for virtual sections.
284 if (SD.getSection().isVirtualSection()) {
285 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
289 // struct section (68 bytes) or
290 // struct section_64 (80 bytes)
292 uint64_t Start = OS.tell();
295 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
296 WriteBytes(Section.getSectionName(), 16);
297 WriteBytes(Section.getSegmentName(), 16);
299 Write64(Layout.getSectionAddress(&SD)); // address
300 Write64(SectionSize); // size
302 Write32(Layout.getSectionAddress(&SD)); // address
303 Write32(SectionSize); // size
307 unsigned Flags = Section.getTypeAndAttributes();
308 if (SD.hasInstructions())
309 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
311 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
312 Write32(Log2_32(SD.getAlignment()));
313 Write32(NumRelocations ? RelocationsStart : 0);
314 Write32(NumRelocations);
316 Write32(IndirectSymBase.lookup(&SD)); // reserved1
317 Write32(Section.getStubSize()); // reserved2
319 Write32(0); // reserved3
321 assert(OS.tell() - Start == Is64Bit ? macho::Section64Size :
322 macho::Section32Size);
325 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
326 uint32_t StringTableOffset,
327 uint32_t StringTableSize) {
328 // struct symtab_command (24 bytes)
330 uint64_t Start = OS.tell();
333 Write32(macho::LCT_Symtab);
334 Write32(macho::SymtabLoadCommandSize);
335 Write32(SymbolOffset);
337 Write32(StringTableOffset);
338 Write32(StringTableSize);
340 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
343 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
344 uint32_t NumLocalSymbols,
345 uint32_t FirstExternalSymbol,
346 uint32_t NumExternalSymbols,
347 uint32_t FirstUndefinedSymbol,
348 uint32_t NumUndefinedSymbols,
349 uint32_t IndirectSymbolOffset,
350 uint32_t NumIndirectSymbols) {
351 // struct dysymtab_command (80 bytes)
353 uint64_t Start = OS.tell();
356 Write32(macho::LCT_Dysymtab);
357 Write32(macho::DysymtabLoadCommandSize);
358 Write32(FirstLocalSymbol);
359 Write32(NumLocalSymbols);
360 Write32(FirstExternalSymbol);
361 Write32(NumExternalSymbols);
362 Write32(FirstUndefinedSymbol);
363 Write32(NumUndefinedSymbols);
364 Write32(0); // tocoff
366 Write32(0); // modtaboff
367 Write32(0); // nmodtab
368 Write32(0); // extrefsymoff
369 Write32(0); // nextrefsyms
370 Write32(IndirectSymbolOffset);
371 Write32(NumIndirectSymbols);
372 Write32(0); // extreloff
373 Write32(0); // nextrel
374 Write32(0); // locreloff
375 Write32(0); // nlocrel
377 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
380 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
381 MCSymbolData &Data = *MSD.SymbolData;
382 const MCSymbol &Symbol = Data.getSymbol();
384 uint16_t Flags = Data.getFlags();
385 uint32_t Address = 0;
387 // Set the N_TYPE bits. See <mach-o/nlist.h>.
389 // FIXME: Are the prebound or indirect fields possible here?
390 if (Symbol.isUndefined())
391 Type = macho::STT_Undefined;
392 else if (Symbol.isAbsolute())
393 Type = macho::STT_Absolute;
395 Type = macho::STT_Section;
397 // FIXME: Set STAB bits.
399 if (Data.isPrivateExtern())
400 Type |= macho::STF_PrivateExtern;
403 if (Data.isExternal() || Symbol.isUndefined())
404 Type |= macho::STF_External;
406 // Compute the symbol address.
407 if (Symbol.isDefined()) {
408 if (Symbol.isAbsolute()) {
409 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
411 Address = Layout.getSymbolAddress(&Data);
413 } else if (Data.isCommon()) {
414 // Common symbols are encoded with the size in the address
415 // field, and their alignment in the flags.
416 Address = Data.getCommonSize();
418 // Common alignment is packed into the 'desc' bits.
419 if (unsigned Align = Data.getCommonAlignment()) {
420 unsigned Log2Size = Log2_32(Align);
421 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
423 report_fatal_error("invalid 'common' alignment '" +
425 // FIXME: Keep this mask with the SymbolFlags enumeration.
426 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
430 // struct nlist (12 bytes)
432 Write32(MSD.StringIndex);
434 Write8(MSD.SectionIndex);
436 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
445 // FIXME: We really need to improve the relocation validation. Basically, we
446 // want to implement a separate computation which evaluates the relocation
447 // entry as the linker would, and verifies that the resultant fixup value is
448 // exactly what the encoder wanted. This will catch several classes of
451 // - Relocation entry bugs, the two algorithms are unlikely to have the same
454 // - Relaxation issues, where we forget to relax something.
456 // - Input errors, where something cannot be correctly encoded. 'as' allows
457 // these through in many cases.
459 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
460 const MCFragment *Fragment,
461 const MCFixup &Fixup, MCValue Target,
462 uint64_t &FixedValue) {
463 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
464 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
465 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
468 uint32_t FixupOffset =
469 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
470 uint32_t FixupAddress =
471 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
474 unsigned IsExtern = 0;
477 Value = Target.getConstant();
480 // Compensate for the relocation offset, Darwin x86_64 relocations only
481 // have the addend and appear to have attempted to define it to be the
482 // actual expression addend without the PCrel bias. However, instructions
483 // with data following the relocation are not accomodated for (see comment
484 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
485 Value += 1LL << Log2Size;
488 if (Target.isAbsolute()) { // constant
489 // SymbolNum of 0 indicates the absolute section.
490 Type = macho::RIT_X86_64_Unsigned;
493 // FIXME: I believe this is broken, I don't think the linker can
494 // understand it. I think it would require a local relocation, but I'm not
495 // sure if that would work either. The official way to get an absolute
496 // PCrel relocation is to use an absolute symbol (which we don't support
500 Type = macho::RIT_X86_64_Branch;
502 } else if (Target.getSymB()) { // A - B + constant
503 const MCSymbol *A = &Target.getSymA()->getSymbol();
504 MCSymbolData &A_SD = Asm.getSymbolData(*A);
505 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
507 const MCSymbol *B = &Target.getSymB()->getSymbol();
508 MCSymbolData &B_SD = Asm.getSymbolData(*B);
509 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
511 // Neither symbol can be modified.
512 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
513 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
514 report_fatal_error("unsupported relocation of modified symbol");
516 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
517 // implement most of these correctly.
519 report_fatal_error("unsupported pc-relative relocation of difference");
521 // The support for the situation where one or both of the symbols would
522 // require a local relocation is handled just like if the symbols were
523 // external. This is certainly used in the case of debug sections where
524 // the section has only temporary symbols and thus the symbols don't have
525 // base symbols. This is encoded using the section ordinal and
526 // non-extern relocation entries.
528 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
529 // a single SIGNED relocation); reject it for now. Except the case where
530 // both symbols don't have a base, equal but both NULL.
531 if (A_Base == B_Base && A_Base)
532 report_fatal_error("unsupported relocation with identical base");
534 Value += Layout.getSymbolAddress(&A_SD) -
535 (A_Base == NULL ? 0 : Layout.getSymbolAddress(A_Base));
536 Value -= Layout.getSymbolAddress(&B_SD) -
537 (B_Base == NULL ? 0 : Layout.getSymbolAddress(B_Base));
540 Index = A_Base->getIndex();
544 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
547 Type = macho::RIT_X86_64_Unsigned;
549 macho::RelocationEntry MRE;
550 MRE.Word0 = FixupOffset;
551 MRE.Word1 = ((Index << 0) |
556 Relocations[Fragment->getParent()].push_back(MRE);
559 Index = B_Base->getIndex();
563 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
566 Type = macho::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 = macho::RIT_X86_64_GOTLoad;
616 Type = macho::RIT_X86_64_GOT;
617 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
618 Type = macho::RIT_X86_64_TLV;
619 } else if (Modifier != MCSymbolRefExpr::VK_None) {
620 report_fatal_error("unsupported symbol modifier in relocation");
622 Type = macho::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 = macho::RIT_X86_64_Signed1; break;
640 case 2: Type = macho::RIT_X86_64_Signed2; break;
641 case 4: Type = macho::RIT_X86_64_Signed4; break;
645 if (Modifier != MCSymbolRefExpr::VK_None)
646 report_fatal_error("unsupported symbol modifier in branch "
649 Type = macho::RIT_X86_64_Branch;
652 if (Modifier == MCSymbolRefExpr::VK_GOT) {
653 Type = macho::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 = macho::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 = macho::RIT_X86_64_Unsigned;
670 // x86_64 always writes custom values into the fixups.
673 // struct relocation_info (8 bytes)
674 macho::RelocationEntry 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 = macho::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() ? macho::RIT_Difference :
718 macho::RIT_LocalDifference;
719 Value2 = Layout.getSymbolAddress(B_SD);
722 // Relocations are written out in reverse order, so the PAIR comes first.
723 if (Type == macho::RIT_Difference || Type == macho::RIT_LocalDifference) {
724 macho::RelocationEntry MRE;
725 MRE.Word0 = ((0 << 0) |
726 (macho::RIT_Pair << 24) |
729 macho::RF_Scattered);
731 Relocations[Fragment->getParent()].push_back(MRE);
734 macho::RelocationEntry MRE;
735 MRE.Word0 = ((FixupOffset << 0) |
739 macho::RF_Scattered);
741 Relocations[Fragment->getParent()].push_back(MRE);
744 void RecordTLVPRelocation(const MCAssembler &Asm,
745 const MCAsmLayout &Layout,
746 const MCFragment *Fragment,
747 const MCFixup &Fixup, MCValue Target,
748 uint64_t &FixedValue) {
749 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
751 "Should only be called with a 32-bit TLVP relocation!");
753 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
754 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
755 unsigned IsPCRel = 0;
757 // Get the symbol data.
758 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
759 unsigned Index = SD_A->getIndex();
761 // We're only going to have a second symbol in pic mode and it'll be a
762 // subtraction from the picbase. For 32-bit pic the addend is the difference
763 // between the picbase and the next address. For 32-bit static the addend
765 if (Target.getSymB()) {
766 // If this is a subtraction then we're pcrel.
767 uint32_t FixupAddress =
768 Layout.getFragmentAddress(Fragment) + Fixup.getOffset();
769 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
771 FixedValue = (FixupAddress - Layout.getSymbolAddress(SD_B) +
772 Target.getConstant());
773 FixedValue += 1ULL << Log2Size;
778 // struct relocation_info (8 bytes)
779 macho::RelocationEntry MRE;
781 MRE.Word1 = ((Index << 0) |
784 (1 << 27) | // Extern
785 (macho::RIT_TLV << 28)); // Type
786 Relocations[Fragment->getParent()].push_back(MRE);
789 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
790 const MCFragment *Fragment, const MCFixup &Fixup,
791 MCValue Target, uint64_t &FixedValue) {
793 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
797 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
798 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
800 // If this is a 32-bit TLVP reloc it's handled a bit differently.
801 if (Target.getSymA() &&
802 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
803 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
807 // If this is a difference or a defined symbol plus an offset, then we need
808 // a scattered relocation entry.
809 // Differences always require scattered relocations.
810 if (Target.getSymB())
811 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
814 // Get the symbol data, if any.
815 MCSymbolData *SD = 0;
816 if (Target.getSymA())
817 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
819 // If this is an internal relocation with an offset, it also needs a
820 // scattered relocation entry.
821 uint32_t Offset = Target.getConstant();
823 Offset += 1 << Log2Size;
824 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
825 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
829 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
831 unsigned IsExtern = 0;
834 if (Target.isAbsolute()) { // constant
835 // SymbolNum of 0 indicates the absolute section.
837 // FIXME: Currently, these are never generated (see code below). I cannot
838 // find a case where they are actually emitted.
839 Type = macho::RIT_Vanilla;
841 // Check whether we need an external or internal relocation.
842 if (doesSymbolRequireExternRelocation(SD)) {
844 Index = SD->getIndex();
845 // For external relocations, make sure to offset the fixup value to
846 // compensate for the addend of the symbol address, if it was
847 // undefined. This occurs with weak definitions, for example.
848 if (!SD->Symbol->isUndefined())
849 FixedValue -= Layout.getSymbolAddress(SD);
851 // The index is the section ordinal (1-based).
852 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
855 Type = macho::RIT_Vanilla;
858 // struct relocation_info (8 bytes)
859 macho::RelocationEntry MRE;
860 MRE.Word0 = FixupOffset;
861 MRE.Word1 = ((Index << 0) |
866 Relocations[Fragment->getParent()].push_back(MRE);
869 void BindIndirectSymbols(MCAssembler &Asm) {
870 // This is the point where 'as' creates actual symbols for indirect symbols
871 // (in the following two passes). It would be easier for us to do this
872 // sooner when we see the attribute, but that makes getting the order in the
873 // symbol table much more complicated than it is worth.
875 // FIXME: Revisit this when the dust settles.
877 // Bind non lazy symbol pointers first.
878 unsigned IndirectIndex = 0;
879 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
880 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
881 const MCSectionMachO &Section =
882 cast<MCSectionMachO>(it->SectionData->getSection());
884 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
887 // Initialize the section indirect symbol base, if necessary.
888 if (!IndirectSymBase.count(it->SectionData))
889 IndirectSymBase[it->SectionData] = IndirectIndex;
891 Asm.getOrCreateSymbolData(*it->Symbol);
894 // Then lazy symbol pointers and symbol stubs.
896 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
897 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
898 const MCSectionMachO &Section =
899 cast<MCSectionMachO>(it->SectionData->getSection());
901 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
902 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
905 // Initialize the section indirect symbol base, if necessary.
906 if (!IndirectSymBase.count(it->SectionData))
907 IndirectSymBase[it->SectionData] = IndirectIndex;
909 // Set the symbol type to undefined lazy, but only on construction.
911 // FIXME: Do not hardcode.
913 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
915 Entry.setFlags(Entry.getFlags() | 0x0001);
919 /// ComputeSymbolTable - Compute the symbol table data
921 /// \param StringTable [out] - The string table data.
922 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
924 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
925 std::vector<MachSymbolData> &LocalSymbolData,
926 std::vector<MachSymbolData> &ExternalSymbolData,
927 std::vector<MachSymbolData> &UndefinedSymbolData) {
928 // Build section lookup table.
929 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
931 for (MCAssembler::iterator it = Asm.begin(),
932 ie = Asm.end(); it != ie; ++it, ++Index)
933 SectionIndexMap[&it->getSection()] = Index;
934 assert(Index <= 256 && "Too many sections!");
936 // Index 0 is always the empty string.
937 StringMap<uint64_t> StringIndexMap;
938 StringTable += '\x00';
940 // Build the symbol arrays and the string table, but only for non-local
943 // The particular order that we collect the symbols and create the string
944 // table, then sort the symbols is chosen to match 'as'. Even though it
945 // doesn't matter for correctness, this is important for letting us diff .o
947 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
948 ie = Asm.symbol_end(); it != ie; ++it) {
949 const MCSymbol &Symbol = it->getSymbol();
951 // Ignore non-linker visible symbols.
952 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
955 if (!it->isExternal() && !Symbol.isUndefined())
958 uint64_t &Entry = StringIndexMap[Symbol.getName()];
960 Entry = StringTable.size();
961 StringTable += Symbol.getName();
962 StringTable += '\x00';
967 MSD.StringIndex = Entry;
969 if (Symbol.isUndefined()) {
970 MSD.SectionIndex = 0;
971 UndefinedSymbolData.push_back(MSD);
972 } else if (Symbol.isAbsolute()) {
973 MSD.SectionIndex = 0;
974 ExternalSymbolData.push_back(MSD);
976 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
977 assert(MSD.SectionIndex && "Invalid section index!");
978 ExternalSymbolData.push_back(MSD);
982 // Now add the data for local symbols.
983 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
984 ie = Asm.symbol_end(); it != ie; ++it) {
985 const MCSymbol &Symbol = it->getSymbol();
987 // Ignore non-linker visible symbols.
988 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
991 if (it->isExternal() || Symbol.isUndefined())
994 uint64_t &Entry = StringIndexMap[Symbol.getName()];
996 Entry = StringTable.size();
997 StringTable += Symbol.getName();
998 StringTable += '\x00';
1002 MSD.SymbolData = it;
1003 MSD.StringIndex = Entry;
1005 if (Symbol.isAbsolute()) {
1006 MSD.SectionIndex = 0;
1007 LocalSymbolData.push_back(MSD);
1009 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1010 assert(MSD.SectionIndex && "Invalid section index!");
1011 LocalSymbolData.push_back(MSD);
1015 // External and undefined symbols are required to be in lexicographic order.
1016 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1017 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1019 // Set the symbol indices.
1021 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1022 LocalSymbolData[i].SymbolData->setIndex(Index++);
1023 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1024 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1025 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1026 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1028 // The string table is padded to a multiple of 4.
1029 while (StringTable.size() % 4)
1030 StringTable += '\x00';
1033 void ExecutePostLayoutBinding(MCAssembler &Asm) {
1034 // Create symbol data for any indirect symbols.
1035 BindIndirectSymbols(Asm);
1037 // Compute symbol table information and bind symbol indices.
1038 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1039 UndefinedSymbolData);
1043 bool IsFixupFullyResolved(const MCAssembler &Asm,
1044 const MCValue Target,
1046 const MCFragment *DF) const {
1047 // If we aren't using scattered symbols, the fixup is fully resolved.
1048 if (!Asm.getBackend().hasScatteredSymbols())
1051 // Otherwise, determine whether this value is actually resolved; scattering
1052 // may cause atoms to move.
1054 // Check if we are using the "simple" resolution algorithm (e.g.,
1056 if (!Asm.getBackend().hasReliableSymbolDifference()) {
1057 const MCSection *BaseSection = 0;
1059 BaseSection = &DF->getParent()->getSection();
1061 return isScatteredFixupFullyResolvedSimple(Asm, Target, BaseSection);
1064 // Otherwise, compute the proper answer as reliably as possible.
1066 // If this is a PCrel relocation, find the base atom (identified by its
1067 // symbol) that the fixup value is relative to.
1068 const MCSymbolData *BaseSymbol = 0;
1070 BaseSymbol = DF->getAtom();
1075 return isScatteredFixupFullyResolved(Asm, Target, BaseSymbol);
1078 void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout) {
1079 unsigned NumSections = Asm.size();
1081 // The section data starts after the header, the segment load command (and
1082 // section headers) and the symbol table.
1083 unsigned NumLoadCommands = 1;
1084 uint64_t LoadCommandsSize = Is64Bit ?
1085 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
1086 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
1088 // Add the symbol table load command sizes, if used.
1089 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1090 UndefinedSymbolData.size();
1092 NumLoadCommands += 2;
1093 LoadCommandsSize += (macho::SymtabLoadCommandSize +
1094 macho::DysymtabLoadCommandSize);
1097 // Compute the total size of the section data, as well as its file size and
1099 uint64_t SectionDataStart = (Is64Bit ? macho::Header64Size :
1100 macho::Header32Size) + LoadCommandsSize;
1101 uint64_t SectionDataSize = 0;
1102 uint64_t SectionDataFileSize = 0;
1103 uint64_t VMSize = 0;
1104 for (MCAssembler::const_iterator it = Asm.begin(),
1105 ie = Asm.end(); it != ie; ++it) {
1106 const MCSectionData &SD = *it;
1107 uint64_t Address = Layout.getSectionAddress(&SD);
1108 uint64_t Size = Layout.getSectionSize(&SD);
1109 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1111 VMSize = std::max(VMSize, Address + Size);
1113 if (SD.getSection().isVirtualSection())
1116 SectionDataSize = std::max(SectionDataSize, Address + Size);
1117 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1120 // The section data is padded to 4 bytes.
1122 // FIXME: Is this machine dependent?
1123 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1124 SectionDataFileSize += SectionDataPadding;
1126 // Write the prolog, starting with the header and load command...
1127 WriteHeader(NumLoadCommands, LoadCommandsSize,
1128 Asm.getSubsectionsViaSymbols());
1129 WriteSegmentLoadCommand(NumSections, VMSize,
1130 SectionDataStart, SectionDataSize);
1132 // ... and then the section headers.
1133 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1134 for (MCAssembler::const_iterator it = Asm.begin(),
1135 ie = Asm.end(); it != ie; ++it) {
1136 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1137 unsigned NumRelocs = Relocs.size();
1138 uint64_t SectionStart = SectionDataStart + Layout.getSectionAddress(it);
1139 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1140 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
1143 // Write the symbol table load command, if used.
1145 unsigned FirstLocalSymbol = 0;
1146 unsigned NumLocalSymbols = LocalSymbolData.size();
1147 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1148 unsigned NumExternalSymbols = ExternalSymbolData.size();
1149 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1150 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1151 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1152 unsigned NumSymTabSymbols =
1153 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1154 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1155 uint64_t IndirectSymbolOffset = 0;
1157 // If used, the indirect symbols are written after the section data.
1158 if (NumIndirectSymbols)
1159 IndirectSymbolOffset = RelocTableEnd;
1161 // The symbol table is written after the indirect symbol data.
1162 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1164 // The string table is written after symbol table.
1165 uint64_t StringTableOffset =
1166 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? macho::Nlist64Size :
1167 macho::Nlist32Size);
1168 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1169 StringTableOffset, StringTable.size());
1171 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1172 FirstExternalSymbol, NumExternalSymbols,
1173 FirstUndefinedSymbol, NumUndefinedSymbols,
1174 IndirectSymbolOffset, NumIndirectSymbols);
1177 // Write the actual section data.
1178 for (MCAssembler::const_iterator it = Asm.begin(),
1179 ie = Asm.end(); it != ie; ++it)
1180 Asm.WriteSectionData(it, Layout, this);
1182 // Write the extra padding.
1183 WriteZeros(SectionDataPadding);
1185 // Write the relocation entries.
1186 for (MCAssembler::const_iterator it = Asm.begin(),
1187 ie = Asm.end(); it != ie; ++it) {
1188 // Write the section relocation entries, in reverse order to match 'as'
1189 // (approximately, the exact algorithm is more complicated than this).
1190 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1191 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1192 Write32(Relocs[e - i - 1].Word0);
1193 Write32(Relocs[e - i - 1].Word1);
1197 // Write the symbol table data, if used.
1199 // Write the indirect symbol entries.
1200 for (MCAssembler::const_indirect_symbol_iterator
1201 it = Asm.indirect_symbol_begin(),
1202 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1203 // Indirect symbols in the non lazy symbol pointer section have some
1204 // special handling.
1205 const MCSectionMachO &Section =
1206 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1207 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1208 // If this symbol is defined and internal, mark it as such.
1209 if (it->Symbol->isDefined() &&
1210 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1211 uint32_t Flags = macho::ISF_Local;
1212 if (it->Symbol->isAbsolute())
1213 Flags |= macho::ISF_Absolute;
1219 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1222 // FIXME: Check that offsets match computed ones.
1224 // Write the symbol table entries.
1225 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1226 WriteNlist(LocalSymbolData[i], Layout);
1227 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1228 WriteNlist(ExternalSymbolData[i], Layout);
1229 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1230 WriteNlist(UndefinedSymbolData[i], Layout);
1232 // Write the string table.
1233 OS << StringTable.str();
1240 MCObjectWriter *llvm::createMachObjectWriter(raw_ostream &OS, bool is64Bit,
1242 uint32_t CPUSubtype,
1243 bool IsLittleEndian) {
1244 return new MachObjectWriter(OS, is64Bit, CPUType, CPUSubtype, IsLittleEndian);