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 // FIXME: Until ARM has it's own relocation stuff spun off, it comes
35 // through here and we don't want it to puke all over. Any reasonable
36 // values will only come when ARM relocation support gets added, at which
37 // point this will be X86 only again and the llvm_unreachable can be
39 default: return 0;// llvm_unreachable("invalid fixup kind!");
41 case FK_Data_1: return 0;
43 case FK_Data_2: return 1;
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 isFixupKindPCRel(unsigned Kind) {
60 case X86::reloc_riprel_4byte:
61 case X86::reloc_riprel_4byte_movq_load:
66 static bool isFixupKindRIPRel(unsigned Kind) {
67 return Kind == X86::reloc_riprel_4byte ||
68 Kind == X86::reloc_riprel_4byte_movq_load;
71 static bool doesSymbolRequireExternRelocation(MCSymbolData *SD) {
72 // Undefined symbols are always extern.
73 if (SD->Symbol->isUndefined())
76 // References to weak definitions require external relocation entries; the
77 // definition may not always be the one in the same object file.
78 if (SD->getFlags() & SF_WeakDefinition)
81 // Otherwise, we can use an internal relocation.
85 static bool isScatteredFixupFullyResolved(const MCAssembler &Asm,
87 const MCSymbolData *BaseSymbol) {
88 // The effective fixup address is
89 // addr(atom(A)) + offset(A)
90 // - addr(atom(B)) - offset(B)
91 // - addr(BaseSymbol) + <fixup offset from base symbol>
92 // and the offsets are not relocatable, so the fixup is fully resolved when
93 // addr(atom(A)) - addr(atom(B)) - addr(BaseSymbol) == 0.
95 // Note that "false" is almost always conservatively correct (it means we emit
96 // a relocation which is unnecessary), except when it would force us to emit a
97 // relocation which the target cannot encode.
99 const MCSymbolData *A_Base = 0, *B_Base = 0;
100 if (const MCSymbolRefExpr *A = Target.getSymA()) {
101 // Modified symbol references cannot be resolved.
102 if (A->getKind() != MCSymbolRefExpr::VK_None)
105 A_Base = Asm.getAtom(&Asm.getSymbolData(A->getSymbol()));
110 if (const MCSymbolRefExpr *B = Target.getSymB()) {
111 // Modified symbol references cannot be resolved.
112 if (B->getKind() != MCSymbolRefExpr::VK_None)
115 B_Base = Asm.getAtom(&Asm.getSymbolData(B->getSymbol()));
120 // If there is no base, A and B have to be the same atom for this fixup to be
123 return A_Base == B_Base;
125 // Otherwise, B must be missing and A must be the base.
126 return !B_Base && BaseSymbol == A_Base;
129 static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm,
130 const MCValue Target,
131 const MCSection *BaseSection) {
132 // The effective fixup address is
133 // addr(atom(A)) + offset(A)
134 // - addr(atom(B)) - offset(B)
135 // - addr(<base symbol>) + <fixup offset from base symbol>
136 // and the offsets are not relocatable, so the fixup is fully resolved when
137 // addr(atom(A)) - addr(atom(B)) - addr(<base symbol>)) == 0.
139 // The simple (Darwin, except on x86_64) way of dealing with this was to
140 // assume that any reference to a temporary symbol *must* be a temporary
141 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
142 // relocation to a temporary symbol (in the same section) is fully
143 // resolved. This also works in conjunction with absolutized .set, which
144 // requires the compiler to use .set to absolutize the differences between
145 // symbols which the compiler knows to be assembly time constants, so we don't
146 // need to worry about considering symbol differences fully resolved.
148 // Non-relative fixups are only resolved if constant.
150 return Target.isAbsolute();
152 // Otherwise, relative fixups are only resolved if not a difference and the
153 // target is a temporary in the same section.
154 if (Target.isAbsolute() || Target.getSymB())
157 const MCSymbol *A = &Target.getSymA()->getSymbol();
158 if (!A->isTemporary() || !A->isInSection() ||
159 &A->getSection() != BaseSection)
167 class MachObjectWriter : public MCObjectWriter {
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 llvm::DenseMap<const MCSectionData*,
186 std::vector<macho::RelocationEntry> > Relocations;
187 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
190 /// @name Symbol Table Data
193 SmallString<256> StringTable;
194 std::vector<MachSymbolData> LocalSymbolData;
195 std::vector<MachSymbolData> ExternalSymbolData;
196 std::vector<MachSymbolData> UndefinedSymbolData;
200 SectionAddrMap SectionAddress;
201 uint64_t getSectionAddress(const MCSectionData* SD) const {
202 return SectionAddress.lookup(SD);
204 uint64_t getSymbolAddress(const MCSymbolData* SD,
205 const MCAsmLayout &Layout) const {
206 return getSectionAddress(SD->getFragment()->getParent()) +
207 Layout.getSymbolOffset(SD);
209 uint64_t getFragmentAddress(const MCFragment *Fragment,
210 const MCAsmLayout &Layout) const {
211 return getSectionAddress(Fragment->getParent()) +
212 Layout.getFragmentOffset(Fragment);
215 uint64_t getPaddingSize(const MCSectionData *SD,
216 const MCAsmLayout &Layout) const {
217 uint64_t EndAddr = getSectionAddress(SD) + Layout.getSectionAddressSize(SD);
218 unsigned Next = SD->getLayoutOrder() + 1;
219 if (Next >= Layout.getSectionOrder().size())
222 const MCSectionData &NextSD = *Layout.getSectionOrder()[Next];
223 if (NextSD.getSection().isVirtualSection())
225 return OffsetToAlignment(EndAddr, NextSD.getAlignment());
228 unsigned Is64Bit : 1;
234 MachObjectWriter(raw_ostream &_OS,
235 bool _Is64Bit, uint32_t _CPUType, uint32_t _CPUSubtype,
236 bool _IsLittleEndian)
237 : MCObjectWriter(_OS, _IsLittleEndian),
238 Is64Bit(_Is64Bit), CPUType(_CPUType), CPUSubtype(_CPUSubtype) {
241 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
242 bool SubsectionsViaSymbols) {
245 if (SubsectionsViaSymbols)
246 Flags |= macho::HF_SubsectionsViaSymbols;
248 // struct mach_header (28 bytes) or
249 // struct mach_header_64 (32 bytes)
251 uint64_t Start = OS.tell();
254 Write32(Is64Bit ? macho::HM_Object64 : macho::HM_Object32);
259 Write32(macho::HFT_Object);
260 Write32(NumLoadCommands);
261 Write32(LoadCommandsSize);
264 Write32(0); // reserved
266 assert(OS.tell() - Start == Is64Bit ?
267 macho::Header64Size : macho::Header32Size);
270 /// WriteSegmentLoadCommand - Write a segment load command.
272 /// \arg NumSections - The number of sections in this segment.
273 /// \arg SectionDataSize - The total size of the sections.
274 void WriteSegmentLoadCommand(unsigned NumSections,
276 uint64_t SectionDataStartOffset,
277 uint64_t SectionDataSize) {
278 // struct segment_command (56 bytes) or
279 // struct segment_command_64 (72 bytes)
281 uint64_t Start = OS.tell();
284 unsigned SegmentLoadCommandSize = Is64Bit ? macho::SegmentLoadCommand64Size:
285 macho::SegmentLoadCommand32Size;
286 Write32(Is64Bit ? macho::LCT_Segment64 : macho::LCT_Segment);
287 Write32(SegmentLoadCommandSize +
288 NumSections * (Is64Bit ? macho::Section64Size :
289 macho::Section32Size));
293 Write64(0); // vmaddr
294 Write64(VMSize); // vmsize
295 Write64(SectionDataStartOffset); // file offset
296 Write64(SectionDataSize); // file size
298 Write32(0); // vmaddr
299 Write32(VMSize); // vmsize
300 Write32(SectionDataStartOffset); // file offset
301 Write32(SectionDataSize); // file size
303 Write32(0x7); // maxprot
304 Write32(0x7); // initprot
305 Write32(NumSections);
308 assert(OS.tell() - Start == SegmentLoadCommandSize);
311 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
312 const MCSectionData &SD, uint64_t FileOffset,
313 uint64_t RelocationsStart, unsigned NumRelocations) {
314 uint64_t SectionSize = Layout.getSectionAddressSize(&SD);
316 // The offset is unused for virtual sections.
317 if (SD.getSection().isVirtualSection()) {
318 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
322 // struct section (68 bytes) or
323 // struct section_64 (80 bytes)
325 uint64_t Start = OS.tell();
328 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
329 WriteBytes(Section.getSectionName(), 16);
330 WriteBytes(Section.getSegmentName(), 16);
332 Write64(getSectionAddress(&SD)); // address
333 Write64(SectionSize); // size
335 Write32(getSectionAddress(&SD)); // address
336 Write32(SectionSize); // size
340 unsigned Flags = Section.getTypeAndAttributes();
341 if (SD.hasInstructions())
342 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
344 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
345 Write32(Log2_32(SD.getAlignment()));
346 Write32(NumRelocations ? RelocationsStart : 0);
347 Write32(NumRelocations);
349 Write32(IndirectSymBase.lookup(&SD)); // reserved1
350 Write32(Section.getStubSize()); // reserved2
352 Write32(0); // reserved3
354 assert(OS.tell() - Start == Is64Bit ? macho::Section64Size :
355 macho::Section32Size);
358 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
359 uint32_t StringTableOffset,
360 uint32_t StringTableSize) {
361 // struct symtab_command (24 bytes)
363 uint64_t Start = OS.tell();
366 Write32(macho::LCT_Symtab);
367 Write32(macho::SymtabLoadCommandSize);
368 Write32(SymbolOffset);
370 Write32(StringTableOffset);
371 Write32(StringTableSize);
373 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
376 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
377 uint32_t NumLocalSymbols,
378 uint32_t FirstExternalSymbol,
379 uint32_t NumExternalSymbols,
380 uint32_t FirstUndefinedSymbol,
381 uint32_t NumUndefinedSymbols,
382 uint32_t IndirectSymbolOffset,
383 uint32_t NumIndirectSymbols) {
384 // struct dysymtab_command (80 bytes)
386 uint64_t Start = OS.tell();
389 Write32(macho::LCT_Dysymtab);
390 Write32(macho::DysymtabLoadCommandSize);
391 Write32(FirstLocalSymbol);
392 Write32(NumLocalSymbols);
393 Write32(FirstExternalSymbol);
394 Write32(NumExternalSymbols);
395 Write32(FirstUndefinedSymbol);
396 Write32(NumUndefinedSymbols);
397 Write32(0); // tocoff
399 Write32(0); // modtaboff
400 Write32(0); // nmodtab
401 Write32(0); // extrefsymoff
402 Write32(0); // nextrefsyms
403 Write32(IndirectSymbolOffset);
404 Write32(NumIndirectSymbols);
405 Write32(0); // extreloff
406 Write32(0); // nextrel
407 Write32(0); // locreloff
408 Write32(0); // nlocrel
410 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
413 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
414 MCSymbolData &Data = *MSD.SymbolData;
415 const MCSymbol &Symbol = Data.getSymbol();
417 uint16_t Flags = Data.getFlags();
418 uint32_t Address = 0;
420 // Set the N_TYPE bits. See <mach-o/nlist.h>.
422 // FIXME: Are the prebound or indirect fields possible here?
423 if (Symbol.isUndefined())
424 Type = macho::STT_Undefined;
425 else if (Symbol.isAbsolute())
426 Type = macho::STT_Absolute;
428 Type = macho::STT_Section;
430 // FIXME: Set STAB bits.
432 if (Data.isPrivateExtern())
433 Type |= macho::STF_PrivateExtern;
436 if (Data.isExternal() || Symbol.isUndefined())
437 Type |= macho::STF_External;
439 // Compute the symbol address.
440 if (Symbol.isDefined()) {
441 if (Symbol.isAbsolute()) {
442 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
444 Address = getSymbolAddress(&Data, Layout);
446 } else if (Data.isCommon()) {
447 // Common symbols are encoded with the size in the address
448 // field, and their alignment in the flags.
449 Address = Data.getCommonSize();
451 // Common alignment is packed into the 'desc' bits.
452 if (unsigned Align = Data.getCommonAlignment()) {
453 unsigned Log2Size = Log2_32(Align);
454 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
456 report_fatal_error("invalid 'common' alignment '" +
458 // FIXME: Keep this mask with the SymbolFlags enumeration.
459 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
463 // struct nlist (12 bytes)
465 Write32(MSD.StringIndex);
467 Write8(MSD.SectionIndex);
469 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
478 // FIXME: We really need to improve the relocation validation. Basically, we
479 // want to implement a separate computation which evaluates the relocation
480 // entry as the linker would, and verifies that the resultant fixup value is
481 // exactly what the encoder wanted. This will catch several classes of
484 // - Relocation entry bugs, the two algorithms are unlikely to have the same
487 // - Relaxation issues, where we forget to relax something.
489 // - Input errors, where something cannot be correctly encoded. 'as' allows
490 // these through in many cases.
492 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
493 const MCFragment *Fragment,
494 const MCFixup &Fixup, MCValue Target,
495 uint64_t &FixedValue) {
496 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
497 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
498 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
501 uint32_t FixupOffset =
502 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
503 uint32_t FixupAddress =
504 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
507 unsigned IsExtern = 0;
510 Value = Target.getConstant();
513 // Compensate for the relocation offset, Darwin x86_64 relocations only
514 // have the addend and appear to have attempted to define it to be the
515 // actual expression addend without the PCrel bias. However, instructions
516 // with data following the relocation are not accomodated for (see comment
517 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
518 Value += 1LL << Log2Size;
521 if (Target.isAbsolute()) { // constant
522 // SymbolNum of 0 indicates the absolute section.
523 Type = macho::RIT_X86_64_Unsigned;
526 // FIXME: I believe this is broken, I don't think the linker can
527 // understand it. I think it would require a local relocation, but I'm not
528 // sure if that would work either. The official way to get an absolute
529 // PCrel relocation is to use an absolute symbol (which we don't support
533 Type = macho::RIT_X86_64_Branch;
535 } else if (Target.getSymB()) { // A - B + constant
536 const MCSymbol *A = &Target.getSymA()->getSymbol();
537 MCSymbolData &A_SD = Asm.getSymbolData(*A);
538 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
540 const MCSymbol *B = &Target.getSymB()->getSymbol();
541 MCSymbolData &B_SD = Asm.getSymbolData(*B);
542 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
544 // Neither symbol can be modified.
545 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
546 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
547 report_fatal_error("unsupported relocation of modified symbol");
549 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
550 // implement most of these correctly.
552 report_fatal_error("unsupported pc-relative relocation of difference");
554 // The support for the situation where one or both of the symbols would
555 // require a local relocation is handled just like if the symbols were
556 // external. This is certainly used in the case of debug sections where
557 // the section has only temporary symbols and thus the symbols don't have
558 // base symbols. This is encoded using the section ordinal and
559 // non-extern relocation entries.
561 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
562 // a single SIGNED relocation); reject it for now. Except the case where
563 // both symbols don't have a base, equal but both NULL.
564 if (A_Base == B_Base && A_Base)
565 report_fatal_error("unsupported relocation with identical base");
567 assert((A_Base == NULL) == (B_Base == NULL));
568 assert(A_SD.getFragment()->getParent() ==
569 B_SD.getFragment()->getParent());
571 Value += Layout.getSymbolOffset(&A_SD) -
572 (A_Base == NULL ? 0 : Layout.getSymbolOffset(A_Base));
573 Value -= Layout.getSymbolOffset(&B_SD) -
574 (B_Base == NULL ? 0 : Layout.getSymbolOffset(B_Base));
577 Index = A_Base->getIndex();
581 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
584 Type = macho::RIT_X86_64_Unsigned;
586 macho::RelocationEntry MRE;
587 MRE.Word0 = FixupOffset;
588 MRE.Word1 = ((Index << 0) |
593 Relocations[Fragment->getParent()].push_back(MRE);
596 Index = B_Base->getIndex();
600 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
603 Type = macho::RIT_X86_64_Subtractor;
605 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
606 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
607 const MCSymbolData *Base = Asm.getAtom(&SD);
609 // Relocations inside debug sections always use local relocations when
610 // possible. This seems to be done because the debugger doesn't fully
611 // understand x86_64 relocation entries, and expects to find values that
612 // have already been fixed up.
613 if (Symbol->isInSection()) {
614 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
615 Fragment->getParent()->getSection());
616 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
620 // x86_64 almost always uses external relocations, except when there is no
621 // symbol to use as a base address (a local symbol with no preceeding
622 // non-local symbol).
624 Index = Base->getIndex();
627 // Add the local offset, if needed.
629 Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base);
630 } else if (Symbol->isInSection()) {
631 // The index is the section ordinal (1-based).
632 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
634 Value += getSymbolAddress(&SD, Layout);
637 Value -= FixupAddress + (1 << Log2Size);
638 } else if (Symbol->isVariable()) {
639 const MCExpr *Value = Symbol->getVariableValue();
641 bool isAbs = Value->EvaluateAsAbsolute(Res, Layout, SectionAddress);
646 report_fatal_error("unsupported relocation of variable '" +
647 Symbol->getName() + "'");
650 report_fatal_error("unsupported relocation of undefined symbol '" +
651 Symbol->getName() + "'");
654 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
657 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
658 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
659 // rewrite the movq to an leaq at link time if the symbol ends up in
660 // the same linkage unit.
661 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
662 Type = macho::RIT_X86_64_GOTLoad;
664 Type = macho::RIT_X86_64_GOT;
665 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
666 Type = macho::RIT_X86_64_TLV;
667 } else if (Modifier != MCSymbolRefExpr::VK_None) {
668 report_fatal_error("unsupported symbol modifier in relocation");
670 Type = macho::RIT_X86_64_Signed;
672 // The Darwin x86_64 relocation format has a problem where it cannot
673 // encode an address (L<foo> + <constant>) which is outside the atom
674 // containing L<foo>. Generally, this shouldn't occur but it does
675 // happen when we have a RIPrel instruction with data following the
676 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
677 // adjustment Darwin x86_64 uses, the offset is still negative and
678 // the linker has no way to recognize this.
680 // To work around this, Darwin uses several special relocation types
681 // to indicate the offsets. However, the specification or
682 // implementation of these seems to also be incomplete; they should
683 // adjust the addend as well based on the actual encoded instruction
684 // (the additional bias), but instead appear to just look at the
686 switch (-(Target.getConstant() + (1LL << Log2Size))) {
687 case 1: Type = macho::RIT_X86_64_Signed1; break;
688 case 2: Type = macho::RIT_X86_64_Signed2; break;
689 case 4: Type = macho::RIT_X86_64_Signed4; break;
693 if (Modifier != MCSymbolRefExpr::VK_None)
694 report_fatal_error("unsupported symbol modifier in branch "
697 Type = macho::RIT_X86_64_Branch;
700 if (Modifier == MCSymbolRefExpr::VK_GOT) {
701 Type = macho::RIT_X86_64_GOT;
702 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
703 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
704 // which case all we do is set the PCrel bit in the relocation entry;
705 // this is used with exception handling, for example. The source is
706 // required to include any necessary offset directly.
707 Type = macho::RIT_X86_64_GOT;
709 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
710 report_fatal_error("TLVP symbol modifier should have been rip-rel");
711 } else if (Modifier != MCSymbolRefExpr::VK_None)
712 report_fatal_error("unsupported symbol modifier in relocation");
714 Type = macho::RIT_X86_64_Unsigned;
718 // x86_64 always writes custom values into the fixups.
721 // struct relocation_info (8 bytes)
722 macho::RelocationEntry MRE;
723 MRE.Word0 = FixupOffset;
724 MRE.Word1 = ((Index << 0) |
729 Relocations[Fragment->getParent()].push_back(MRE);
732 void RecordScatteredRelocation(const MCAssembler &Asm,
733 const MCAsmLayout &Layout,
734 const MCFragment *Fragment,
735 const MCFixup &Fixup, MCValue Target,
736 uint64_t &FixedValue) {
737 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
738 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
739 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
740 unsigned Type = macho::RIT_Vanilla;
743 const MCSymbol *A = &Target.getSymA()->getSymbol();
744 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
746 if (!A_SD->getFragment())
747 report_fatal_error("symbol '" + A->getName() +
748 "' can not be undefined in a subtraction expression");
750 uint32_t Value = getSymbolAddress(A_SD, Layout);
751 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
752 FixedValue += SecAddr;
755 if (const MCSymbolRefExpr *B = Target.getSymB()) {
756 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
758 if (!B_SD->getFragment())
759 report_fatal_error("symbol '" + B->getSymbol().getName() +
760 "' can not be undefined in a subtraction expression");
762 // Select the appropriate difference relocation type.
764 // Note that there is no longer any semantic difference between these two
765 // relocation types from the linkers point of view, this is done solely
766 // for pedantic compatibility with 'as'.
767 Type = A_SD->isExternal() ? macho::RIT_Difference :
768 macho::RIT_LocalDifference;
769 Value2 = getSymbolAddress(B_SD, Layout);
770 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
773 // Relocations are written out in reverse order, so the PAIR comes first.
774 if (Type == macho::RIT_Difference || Type == macho::RIT_LocalDifference) {
775 macho::RelocationEntry MRE;
776 MRE.Word0 = ((0 << 0) |
777 (macho::RIT_Pair << 24) |
780 macho::RF_Scattered);
782 Relocations[Fragment->getParent()].push_back(MRE);
785 macho::RelocationEntry MRE;
786 MRE.Word0 = ((FixupOffset << 0) |
790 macho::RF_Scattered);
792 Relocations[Fragment->getParent()].push_back(MRE);
795 void RecordTLVPRelocation(const MCAssembler &Asm,
796 const MCAsmLayout &Layout,
797 const MCFragment *Fragment,
798 const MCFixup &Fixup, MCValue Target,
799 uint64_t &FixedValue) {
800 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
802 "Should only be called with a 32-bit TLVP relocation!");
804 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
805 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
806 unsigned IsPCRel = 0;
808 // Get the symbol data.
809 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
810 unsigned Index = SD_A->getIndex();
812 // We're only going to have a second symbol in pic mode and it'll be a
813 // subtraction from the picbase. For 32-bit pic the addend is the difference
814 // between the picbase and the next address. For 32-bit static the addend
816 if (Target.getSymB()) {
817 // If this is a subtraction then we're pcrel.
818 uint32_t FixupAddress =
819 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
820 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
822 FixedValue = (FixupAddress - getSymbolAddress(SD_B, Layout) +
823 Target.getConstant());
824 FixedValue += 1ULL << Log2Size;
829 // struct relocation_info (8 bytes)
830 macho::RelocationEntry MRE;
832 MRE.Word1 = ((Index << 0) |
835 (1 << 27) | // Extern
836 (macho::RIT_TLV << 28)); // Type
837 Relocations[Fragment->getParent()].push_back(MRE);
840 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
841 const MCFragment *Fragment, const MCFixup &Fixup,
842 MCValue Target, uint64_t &FixedValue) {
844 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
848 unsigned IsPCRel = isFixupKindPCRel(Fixup.getKind());
849 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
851 // If this is a 32-bit TLVP reloc it's handled a bit differently.
852 if (Target.getSymA() &&
853 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
854 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
858 // If this is a difference or a defined symbol plus an offset, then we need
859 // a scattered relocation entry.
860 // Differences always require scattered relocations.
861 if (Target.getSymB())
862 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
865 // Get the symbol data, if any.
866 MCSymbolData *SD = 0;
867 if (Target.getSymA())
868 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
870 // If this is an internal relocation with an offset, it also needs a
871 // scattered relocation entry.
872 uint32_t Offset = Target.getConstant();
874 Offset += 1 << Log2Size;
875 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
876 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
880 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
882 unsigned IsExtern = 0;
885 if (Target.isAbsolute()) { // constant
886 // SymbolNum of 0 indicates the absolute section.
888 // FIXME: Currently, these are never generated (see code below). I cannot
889 // find a case where they are actually emitted.
890 Type = macho::RIT_Vanilla;
892 // Check whether we need an external or internal relocation.
893 if (doesSymbolRequireExternRelocation(SD)) {
895 Index = SD->getIndex();
896 // For external relocations, make sure to offset the fixup value to
897 // compensate for the addend of the symbol address, if it was
898 // undefined. This occurs with weak definitions, for example.
899 if (!SD->Symbol->isUndefined())
900 FixedValue -= getSymbolAddress(SD, Layout);
902 // The index is the section ordinal (1-based).
903 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
904 FixedValue += getSectionAddress(SD->getFragment()->getParent());
906 FixedValue -= getSectionAddress(Fragment->getParent());
909 Type = macho::RIT_Vanilla;
912 // struct relocation_info (8 bytes)
913 macho::RelocationEntry MRE;
914 MRE.Word0 = FixupOffset;
915 MRE.Word1 = ((Index << 0) |
920 Relocations[Fragment->getParent()].push_back(MRE);
923 void BindIndirectSymbols(MCAssembler &Asm) {
924 // This is the point where 'as' creates actual symbols for indirect symbols
925 // (in the following two passes). It would be easier for us to do this
926 // sooner when we see the attribute, but that makes getting the order in the
927 // symbol table much more complicated than it is worth.
929 // FIXME: Revisit this when the dust settles.
931 // Bind non lazy symbol pointers first.
932 unsigned IndirectIndex = 0;
933 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
934 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
935 const MCSectionMachO &Section =
936 cast<MCSectionMachO>(it->SectionData->getSection());
938 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
941 // Initialize the section indirect symbol base, if necessary.
942 if (!IndirectSymBase.count(it->SectionData))
943 IndirectSymBase[it->SectionData] = IndirectIndex;
945 Asm.getOrCreateSymbolData(*it->Symbol);
948 // Then lazy symbol pointers and symbol stubs.
950 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
951 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
952 const MCSectionMachO &Section =
953 cast<MCSectionMachO>(it->SectionData->getSection());
955 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
956 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
959 // Initialize the section indirect symbol base, if necessary.
960 if (!IndirectSymBase.count(it->SectionData))
961 IndirectSymBase[it->SectionData] = IndirectIndex;
963 // Set the symbol type to undefined lazy, but only on construction.
965 // FIXME: Do not hardcode.
967 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
969 Entry.setFlags(Entry.getFlags() | 0x0001);
973 /// ComputeSymbolTable - Compute the symbol table data
975 /// \param StringTable [out] - The string table data.
976 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
978 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
979 std::vector<MachSymbolData> &LocalSymbolData,
980 std::vector<MachSymbolData> &ExternalSymbolData,
981 std::vector<MachSymbolData> &UndefinedSymbolData) {
982 // Build section lookup table.
983 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
985 for (MCAssembler::iterator it = Asm.begin(),
986 ie = Asm.end(); it != ie; ++it, ++Index)
987 SectionIndexMap[&it->getSection()] = Index;
988 assert(Index <= 256 && "Too many sections!");
990 // Index 0 is always the empty string.
991 StringMap<uint64_t> StringIndexMap;
992 StringTable += '\x00';
994 // Build the symbol arrays and the string table, but only for non-local
997 // The particular order that we collect the symbols and create the string
998 // table, then sort the symbols is chosen to match 'as'. Even though it
999 // doesn't matter for correctness, this is important for letting us diff .o
1001 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1002 ie = Asm.symbol_end(); it != ie; ++it) {
1003 const MCSymbol &Symbol = it->getSymbol();
1005 // Ignore non-linker visible symbols.
1006 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1009 if (!it->isExternal() && !Symbol.isUndefined())
1012 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1014 Entry = StringTable.size();
1015 StringTable += Symbol.getName();
1016 StringTable += '\x00';
1020 MSD.SymbolData = it;
1021 MSD.StringIndex = Entry;
1023 if (Symbol.isUndefined()) {
1024 MSD.SectionIndex = 0;
1025 UndefinedSymbolData.push_back(MSD);
1026 } else if (Symbol.isAbsolute()) {
1027 MSD.SectionIndex = 0;
1028 ExternalSymbolData.push_back(MSD);
1030 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1031 assert(MSD.SectionIndex && "Invalid section index!");
1032 ExternalSymbolData.push_back(MSD);
1036 // Now add the data for local symbols.
1037 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1038 ie = Asm.symbol_end(); it != ie; ++it) {
1039 const MCSymbol &Symbol = it->getSymbol();
1041 // Ignore non-linker visible symbols.
1042 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1045 if (it->isExternal() || Symbol.isUndefined())
1048 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1050 Entry = StringTable.size();
1051 StringTable += Symbol.getName();
1052 StringTable += '\x00';
1056 MSD.SymbolData = it;
1057 MSD.StringIndex = Entry;
1059 if (Symbol.isAbsolute()) {
1060 MSD.SectionIndex = 0;
1061 LocalSymbolData.push_back(MSD);
1063 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1064 assert(MSD.SectionIndex && "Invalid section index!");
1065 LocalSymbolData.push_back(MSD);
1069 // External and undefined symbols are required to be in lexicographic order.
1070 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1071 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1073 // Set the symbol indices.
1075 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1076 LocalSymbolData[i].SymbolData->setIndex(Index++);
1077 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1078 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1079 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1080 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1082 // The string table is padded to a multiple of 4.
1083 while (StringTable.size() % 4)
1084 StringTable += '\x00';
1087 void computeSectionAddresses(const MCAssembler &Asm,
1088 const MCAsmLayout &Layout) {
1089 uint64_t StartAddress = 0;
1090 const SmallVectorImpl<MCSectionData*> &Order = Layout.getSectionOrder();
1091 for (int i = 0, n = Order.size(); i != n ; ++i) {
1092 const MCSectionData *SD = Order[i];
1093 StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment());
1094 SectionAddress[SD] = StartAddress;
1095 StartAddress += Layout.getSectionAddressSize(SD);
1096 // Explicitly pad the section to match the alignment requirements of the
1097 // following one. This is for 'gas' compatibility, it shouldn't
1098 /// strictly be necessary.
1099 StartAddress += getPaddingSize(SD, Layout);
1103 void ExecutePostLayoutBinding(MCAssembler &Asm, const MCAsmLayout &Layout) {
1104 computeSectionAddresses(Asm, Layout);
1106 // Create symbol data for any indirect symbols.
1107 BindIndirectSymbols(Asm);
1109 // Compute symbol table information and bind symbol indices.
1110 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1111 UndefinedSymbolData);
1115 bool IsFixupFullyResolved(const MCAssembler &Asm,
1116 const MCValue Target,
1118 const MCFragment *DF) const {
1119 // If we aren't using scattered symbols, the fixup is fully resolved.
1120 if (!Asm.getBackend().hasScatteredSymbols())
1123 // Otherwise, determine whether this value is actually resolved; scattering
1124 // may cause atoms to move.
1126 // Check if we are using the "simple" resolution algorithm (e.g.,
1128 if (!Asm.getBackend().hasReliableSymbolDifference()) {
1129 const MCSection *BaseSection = 0;
1131 BaseSection = &DF->getParent()->getSection();
1133 return isScatteredFixupFullyResolvedSimple(Asm, Target, BaseSection);
1136 // Otherwise, compute the proper answer as reliably as possible.
1138 // If this is a PCrel relocation, find the base atom (identified by its
1139 // symbol) that the fixup value is relative to.
1140 const MCSymbolData *BaseSymbol = 0;
1142 BaseSymbol = DF->getAtom();
1147 return isScatteredFixupFullyResolved(Asm, Target, BaseSymbol);
1150 void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout) {
1151 unsigned NumSections = Asm.size();
1153 // The section data starts after the header, the segment load command (and
1154 // section headers) and the symbol table.
1155 unsigned NumLoadCommands = 1;
1156 uint64_t LoadCommandsSize = Is64Bit ?
1157 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
1158 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
1160 // Add the symbol table load command sizes, if used.
1161 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1162 UndefinedSymbolData.size();
1164 NumLoadCommands += 2;
1165 LoadCommandsSize += (macho::SymtabLoadCommandSize +
1166 macho::DysymtabLoadCommandSize);
1169 // Compute the total size of the section data, as well as its file size and
1171 uint64_t SectionDataStart = (Is64Bit ? macho::Header64Size :
1172 macho::Header32Size) + LoadCommandsSize;
1173 uint64_t SectionDataSize = 0;
1174 uint64_t SectionDataFileSize = 0;
1175 uint64_t VMSize = 0;
1176 for (MCAssembler::const_iterator it = Asm.begin(),
1177 ie = Asm.end(); it != ie; ++it) {
1178 const MCSectionData &SD = *it;
1179 uint64_t Address = getSectionAddress(&SD);
1180 uint64_t Size = Layout.getSectionAddressSize(&SD);
1181 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1182 FileSize += getPaddingSize(&SD, Layout);
1184 VMSize = std::max(VMSize, Address + Size);
1186 if (SD.getSection().isVirtualSection())
1189 SectionDataSize = std::max(SectionDataSize, Address + Size);
1190 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1193 // The section data is padded to 4 bytes.
1195 // FIXME: Is this machine dependent?
1196 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1197 SectionDataFileSize += SectionDataPadding;
1199 // Write the prolog, starting with the header and load command...
1200 WriteHeader(NumLoadCommands, LoadCommandsSize,
1201 Asm.getSubsectionsViaSymbols());
1202 WriteSegmentLoadCommand(NumSections, VMSize,
1203 SectionDataStart, SectionDataSize);
1205 // ... and then the section headers.
1206 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1207 for (MCAssembler::const_iterator it = Asm.begin(),
1208 ie = Asm.end(); it != ie; ++it) {
1209 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1210 unsigned NumRelocs = Relocs.size();
1211 uint64_t SectionStart = SectionDataStart + getSectionAddress(it);
1212 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1213 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
1216 // Write the symbol table load command, if used.
1218 unsigned FirstLocalSymbol = 0;
1219 unsigned NumLocalSymbols = LocalSymbolData.size();
1220 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1221 unsigned NumExternalSymbols = ExternalSymbolData.size();
1222 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1223 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1224 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1225 unsigned NumSymTabSymbols =
1226 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1227 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1228 uint64_t IndirectSymbolOffset = 0;
1230 // If used, the indirect symbols are written after the section data.
1231 if (NumIndirectSymbols)
1232 IndirectSymbolOffset = RelocTableEnd;
1234 // The symbol table is written after the indirect symbol data.
1235 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1237 // The string table is written after symbol table.
1238 uint64_t StringTableOffset =
1239 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? macho::Nlist64Size :
1240 macho::Nlist32Size);
1241 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1242 StringTableOffset, StringTable.size());
1244 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1245 FirstExternalSymbol, NumExternalSymbols,
1246 FirstUndefinedSymbol, NumUndefinedSymbols,
1247 IndirectSymbolOffset, NumIndirectSymbols);
1250 // Write the actual section data.
1251 for (MCAssembler::const_iterator it = Asm.begin(),
1252 ie = Asm.end(); it != ie; ++it) {
1253 Asm.WriteSectionData(it, Layout, this);
1255 uint64_t Pad = getPaddingSize(it, Layout);
1256 for (unsigned int i = 0; i < Pad; ++i)
1260 // Write the extra padding.
1261 WriteZeros(SectionDataPadding);
1263 // Write the relocation entries.
1264 for (MCAssembler::const_iterator it = Asm.begin(),
1265 ie = Asm.end(); it != ie; ++it) {
1266 // Write the section relocation entries, in reverse order to match 'as'
1267 // (approximately, the exact algorithm is more complicated than this).
1268 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1269 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1270 Write32(Relocs[e - i - 1].Word0);
1271 Write32(Relocs[e - i - 1].Word1);
1275 // Write the symbol table data, if used.
1277 // Write the indirect symbol entries.
1278 for (MCAssembler::const_indirect_symbol_iterator
1279 it = Asm.indirect_symbol_begin(),
1280 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1281 // Indirect symbols in the non lazy symbol pointer section have some
1282 // special handling.
1283 const MCSectionMachO &Section =
1284 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1285 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1286 // If this symbol is defined and internal, mark it as such.
1287 if (it->Symbol->isDefined() &&
1288 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1289 uint32_t Flags = macho::ISF_Local;
1290 if (it->Symbol->isAbsolute())
1291 Flags |= macho::ISF_Absolute;
1297 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1300 // FIXME: Check that offsets match computed ones.
1302 // Write the symbol table entries.
1303 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1304 WriteNlist(LocalSymbolData[i], Layout);
1305 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1306 WriteNlist(ExternalSymbolData[i], Layout);
1307 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1308 WriteNlist(UndefinedSymbolData[i], Layout);
1310 // Write the string table.
1311 OS << StringTable.str();
1318 MCObjectWriter *llvm::createMachObjectWriter(raw_ostream &OS, bool is64Bit,
1320 uint32_t CPUSubtype,
1321 bool IsLittleEndian) {
1322 return new MachObjectWriter(OS, is64Bit, CPUType, CPUSubtype, IsLittleEndian);