1 //===- lib/MC/MachObjectWriter.cpp - Mach-O File Writer -------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 #include "llvm/MC/MCMachObjectWriter.h"
11 #include "llvm/ADT/OwningPtr.h"
12 #include "llvm/ADT/StringMap.h"
13 #include "llvm/ADT/Twine.h"
14 #include "llvm/MC/MCAssembler.h"
15 #include "llvm/MC/MCAsmLayout.h"
16 #include "llvm/MC/MCExpr.h"
17 #include "llvm/MC/MCObjectWriter.h"
18 #include "llvm/MC/MCSectionMachO.h"
19 #include "llvm/MC/MCSymbol.h"
20 #include "llvm/MC/MCMachOSymbolFlags.h"
21 #include "llvm/MC/MCValue.h"
22 #include "llvm/Object/MachOFormat.h"
23 #include "llvm/Support/ErrorHandling.h"
24 #include "llvm/Target/TargetAsmBackend.h"
27 #include "../Target/ARM/ARMFixupKinds.h"
28 #include "../Target/X86/X86FixupKinds.h"
32 using namespace llvm::object;
34 // FIXME: this has been copied from (or to) X86AsmBackend.cpp
35 static unsigned getFixupKindLog2Size(unsigned Kind) {
38 llvm_unreachable("invalid fixup kind!");
40 case FK_Data_1: return 0;
42 case FK_Data_2: return 1;
44 // FIXME: Remove these!!!
45 case X86::reloc_riprel_4byte:
46 case X86::reloc_riprel_4byte_movq_load:
47 case X86::reloc_signed_4byte:
48 case FK_Data_4: return 2;
49 case FK_Data_8: return 3;
53 static bool doesSymbolRequireExternRelocation(MCSymbolData *SD) {
54 // Undefined symbols are always extern.
55 if (SD->Symbol->isUndefined())
58 // References to weak definitions require external relocation entries; the
59 // definition may not always be the one in the same object file.
60 if (SD->getFlags() & SF_WeakDefinition)
63 // Otherwise, we can use an internal relocation.
69 class MachObjectWriter : public MCObjectWriter {
70 /// MachSymbolData - Helper struct for containing some precomputed information
72 struct MachSymbolData {
73 MCSymbolData *SymbolData;
77 // Support lexicographic sorting.
78 bool operator<(const MachSymbolData &RHS) const {
79 return SymbolData->getSymbol().getName() <
80 RHS.SymbolData->getSymbol().getName();
84 /// The target specific Mach-O writer instance.
85 llvm::OwningPtr<MCMachObjectTargetWriter> TargetObjectWriter;
87 /// @name Relocation Data
90 llvm::DenseMap<const MCSectionData*,
91 std::vector<macho::RelocationEntry> > Relocations;
92 llvm::DenseMap<const MCSectionData*, unsigned> IndirectSymBase;
95 /// @name Symbol Table Data
98 SmallString<256> StringTable;
99 std::vector<MachSymbolData> LocalSymbolData;
100 std::vector<MachSymbolData> ExternalSymbolData;
101 std::vector<MachSymbolData> UndefinedSymbolData;
106 /// @name Utility Methods
109 bool isFixupKindPCRel(const MCAssembler &Asm, unsigned Kind) {
110 const MCFixupKindInfo &FKI = Asm.getBackend().getFixupKindInfo(
113 return FKI.Flags & MCFixupKindInfo::FKF_IsPCRel;
118 SectionAddrMap SectionAddress;
119 uint64_t getSectionAddress(const MCSectionData* SD) const {
120 return SectionAddress.lookup(SD);
122 uint64_t getSymbolAddress(const MCSymbolData* SD,
123 const MCAsmLayout &Layout) const;
125 uint64_t getFragmentAddress(const MCFragment *Fragment,
126 const MCAsmLayout &Layout) const {
127 return getSectionAddress(Fragment->getParent()) +
128 Layout.getFragmentOffset(Fragment);
131 uint64_t getPaddingSize(const MCSectionData *SD,
132 const MCAsmLayout &Layout) const;
134 MachObjectWriter(MCMachObjectTargetWriter *MOTW, raw_ostream &_OS,
135 bool _IsLittleEndian)
136 : MCObjectWriter(_OS, _IsLittleEndian), TargetObjectWriter(MOTW) {
139 /// @name Target Writer Proxy Accessors
142 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
144 uint32_t CPUType = TargetObjectWriter->getCPUType() & ~mach::CTFM_ArchMask;
145 return CPUType == mach::CTM_ARM;
150 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
151 bool SubsectionsViaSymbols);
153 /// WriteSegmentLoadCommand - Write a segment load command.
155 /// \arg NumSections - The number of sections in this segment.
156 /// \arg SectionDataSize - The total size of the sections.
157 void WriteSegmentLoadCommand(unsigned NumSections,
159 uint64_t SectionDataStartOffset,
160 uint64_t SectionDataSize);
162 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
163 const MCSectionData &SD, uint64_t FileOffset,
164 uint64_t RelocationsStart, unsigned NumRelocations);
166 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
167 uint32_t StringTableOffset,
168 uint32_t StringTableSize);
170 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
171 uint32_t NumLocalSymbols,
172 uint32_t FirstExternalSymbol,
173 uint32_t NumExternalSymbols,
174 uint32_t FirstUndefinedSymbol,
175 uint32_t NumUndefinedSymbols,
176 uint32_t IndirectSymbolOffset,
177 uint32_t NumIndirectSymbols);
179 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout);
181 // FIXME: We really need to improve the relocation validation. Basically, we
182 // want to implement a separate computation which evaluates the relocation
183 // entry as the linker would, and verifies that the resultant fixup value is
184 // exactly what the encoder wanted. This will catch several classes of
187 // - Relocation entry bugs, the two algorithms are unlikely to have the same
190 // - Relaxation issues, where we forget to relax something.
192 // - Input errors, where something cannot be correctly encoded. 'as' allows
193 // these through in many cases.
195 static bool isFixupKindRIPRel(unsigned Kind) {
196 return Kind == X86::reloc_riprel_4byte ||
197 Kind == X86::reloc_riprel_4byte_movq_load;
199 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
200 const MCFragment *Fragment,
201 const MCFixup &Fixup, MCValue Target,
202 uint64_t &FixedValue);
204 void RecordScatteredRelocation(const MCAssembler &Asm,
205 const MCAsmLayout &Layout,
206 const MCFragment *Fragment,
207 const MCFixup &Fixup, MCValue Target,
209 uint64_t &FixedValue);
211 void RecordARMScatteredRelocation(const MCAssembler &Asm,
212 const MCAsmLayout &Layout,
213 const MCFragment *Fragment,
214 const MCFixup &Fixup, MCValue Target,
216 uint64_t &FixedValue);
218 void RecordARMMovwMovtRelocation(const MCAssembler &Asm,
219 const MCAsmLayout &Layout,
220 const MCFragment *Fragment,
221 const MCFixup &Fixup, MCValue Target,
222 uint64_t &FixedValue);
224 void RecordTLVPRelocation(const MCAssembler &Asm,
225 const MCAsmLayout &Layout,
226 const MCFragment *Fragment,
227 const MCFixup &Fixup, MCValue Target,
228 uint64_t &FixedValue);
230 static bool getARMFixupKindMachOInfo(unsigned Kind, unsigned &RelocType,
233 void RecordARMRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
234 const MCFragment *Fragment, const MCFixup &Fixup,
235 MCValue Target, uint64_t &FixedValue);
237 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
238 const MCFragment *Fragment, const MCFixup &Fixup,
239 MCValue Target, uint64_t &FixedValue);
241 void BindIndirectSymbols(MCAssembler &Asm);
243 /// ComputeSymbolTable - Compute the symbol table data
245 /// \param StringTable [out] - The string table data.
246 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
248 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
249 std::vector<MachSymbolData> &LocalSymbolData,
250 std::vector<MachSymbolData> &ExternalSymbolData,
251 std::vector<MachSymbolData> &UndefinedSymbolData);
253 void computeSectionAddresses(const MCAssembler &Asm,
254 const MCAsmLayout &Layout);
256 void ExecutePostLayoutBinding(MCAssembler &Asm, const MCAsmLayout &Layout);
258 virtual bool IsSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
259 const MCSymbolData &DataA,
260 const MCFragment &FB,
264 void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout);
267 } // end anonymous namespace
269 uint64_t MachObjectWriter::getSymbolAddress(const MCSymbolData* SD,
270 const MCAsmLayout &Layout) const {
271 const MCSymbol &S = SD->getSymbol();
273 // If this is a variable, then recursively evaluate now.
274 if (S.isVariable()) {
276 if (!S.getVariableValue()->EvaluateAsRelocatable(Target, Layout))
277 report_fatal_error("unable to evaluate offset for variable '" +
280 // Verify that any used symbols are defined.
281 if (Target.getSymA() && Target.getSymA()->getSymbol().isUndefined())
282 report_fatal_error("unable to evaluate offset to undefined symbol '" +
283 Target.getSymA()->getSymbol().getName() + "'");
284 if (Target.getSymB() && Target.getSymB()->getSymbol().isUndefined())
285 report_fatal_error("unable to evaluate offset to undefined symbol '" +
286 Target.getSymB()->getSymbol().getName() + "'");
288 uint64_t Address = Target.getConstant();
289 if (Target.getSymA())
290 Address += getSymbolAddress(&Layout.getAssembler().getSymbolData(
291 Target.getSymA()->getSymbol()), Layout);
292 if (Target.getSymB())
293 Address += getSymbolAddress(&Layout.getAssembler().getSymbolData(
294 Target.getSymB()->getSymbol()), Layout);
298 return getSectionAddress(SD->getFragment()->getParent()) +
299 Layout.getSymbolOffset(SD);
302 uint64_t MachObjectWriter::getPaddingSize(const MCSectionData *SD,
303 const MCAsmLayout &Layout) const {
304 uint64_t EndAddr = getSectionAddress(SD) + Layout.getSectionAddressSize(SD);
305 unsigned Next = SD->getLayoutOrder() + 1;
306 if (Next >= Layout.getSectionOrder().size())
309 const MCSectionData &NextSD = *Layout.getSectionOrder()[Next];
310 if (NextSD.getSection().isVirtualSection())
312 return OffsetToAlignment(EndAddr, NextSD.getAlignment());
315 void MachObjectWriter::WriteHeader(unsigned NumLoadCommands,
316 unsigned LoadCommandsSize,
317 bool SubsectionsViaSymbols) {
320 if (SubsectionsViaSymbols)
321 Flags |= macho::HF_SubsectionsViaSymbols;
323 // struct mach_header (28 bytes) or
324 // struct mach_header_64 (32 bytes)
326 uint64_t Start = OS.tell();
329 Write32(is64Bit() ? macho::HM_Object64 : macho::HM_Object32);
331 Write32(TargetObjectWriter->getCPUType());
332 Write32(TargetObjectWriter->getCPUSubtype());
334 Write32(macho::HFT_Object);
335 Write32(NumLoadCommands);
336 Write32(LoadCommandsSize);
339 Write32(0); // reserved
341 assert(OS.tell() - Start ==
342 (is64Bit() ? macho::Header64Size : macho::Header32Size));
345 /// WriteSegmentLoadCommand - Write a segment load command.
347 /// \arg NumSections - The number of sections in this segment.
348 /// \arg SectionDataSize - The total size of the sections.
349 void MachObjectWriter::WriteSegmentLoadCommand(unsigned NumSections,
351 uint64_t SectionDataStartOffset,
352 uint64_t SectionDataSize) {
353 // struct segment_command (56 bytes) or
354 // struct segment_command_64 (72 bytes)
356 uint64_t Start = OS.tell();
359 unsigned SegmentLoadCommandSize =
360 is64Bit() ? macho::SegmentLoadCommand64Size:
361 macho::SegmentLoadCommand32Size;
362 Write32(is64Bit() ? macho::LCT_Segment64 : macho::LCT_Segment);
363 Write32(SegmentLoadCommandSize +
364 NumSections * (is64Bit() ? macho::Section64Size :
365 macho::Section32Size));
369 Write64(0); // vmaddr
370 Write64(VMSize); // vmsize
371 Write64(SectionDataStartOffset); // file offset
372 Write64(SectionDataSize); // file size
374 Write32(0); // vmaddr
375 Write32(VMSize); // vmsize
376 Write32(SectionDataStartOffset); // file offset
377 Write32(SectionDataSize); // file size
379 Write32(0x7); // maxprot
380 Write32(0x7); // initprot
381 Write32(NumSections);
384 assert(OS.tell() - Start == SegmentLoadCommandSize);
387 void MachObjectWriter::WriteSection(const MCAssembler &Asm,
388 const MCAsmLayout &Layout,
389 const MCSectionData &SD,
391 uint64_t RelocationsStart,
392 unsigned NumRelocations) {
393 uint64_t SectionSize = Layout.getSectionAddressSize(&SD);
395 // The offset is unused for virtual sections.
396 if (SD.getSection().isVirtualSection()) {
397 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
401 // struct section (68 bytes) or
402 // struct section_64 (80 bytes)
404 uint64_t Start = OS.tell();
407 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
408 WriteBytes(Section.getSectionName(), 16);
409 WriteBytes(Section.getSegmentName(), 16);
411 Write64(getSectionAddress(&SD)); // address
412 Write64(SectionSize); // size
414 Write32(getSectionAddress(&SD)); // address
415 Write32(SectionSize); // size
419 unsigned Flags = Section.getTypeAndAttributes();
420 if (SD.hasInstructions())
421 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
423 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
424 Write32(Log2_32(SD.getAlignment()));
425 Write32(NumRelocations ? RelocationsStart : 0);
426 Write32(NumRelocations);
428 Write32(IndirectSymBase.lookup(&SD)); // reserved1
429 Write32(Section.getStubSize()); // reserved2
431 Write32(0); // reserved3
433 assert(OS.tell() - Start == (is64Bit() ? macho::Section64Size :
434 macho::Section32Size));
437 void MachObjectWriter::WriteSymtabLoadCommand(uint32_t SymbolOffset,
439 uint32_t StringTableOffset,
440 uint32_t StringTableSize) {
441 // struct symtab_command (24 bytes)
443 uint64_t Start = OS.tell();
446 Write32(macho::LCT_Symtab);
447 Write32(macho::SymtabLoadCommandSize);
448 Write32(SymbolOffset);
450 Write32(StringTableOffset);
451 Write32(StringTableSize);
453 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
456 void MachObjectWriter::WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
457 uint32_t NumLocalSymbols,
458 uint32_t FirstExternalSymbol,
459 uint32_t NumExternalSymbols,
460 uint32_t FirstUndefinedSymbol,
461 uint32_t NumUndefinedSymbols,
462 uint32_t IndirectSymbolOffset,
463 uint32_t NumIndirectSymbols) {
464 // struct dysymtab_command (80 bytes)
466 uint64_t Start = OS.tell();
469 Write32(macho::LCT_Dysymtab);
470 Write32(macho::DysymtabLoadCommandSize);
471 Write32(FirstLocalSymbol);
472 Write32(NumLocalSymbols);
473 Write32(FirstExternalSymbol);
474 Write32(NumExternalSymbols);
475 Write32(FirstUndefinedSymbol);
476 Write32(NumUndefinedSymbols);
477 Write32(0); // tocoff
479 Write32(0); // modtaboff
480 Write32(0); // nmodtab
481 Write32(0); // extrefsymoff
482 Write32(0); // nextrefsyms
483 Write32(IndirectSymbolOffset);
484 Write32(NumIndirectSymbols);
485 Write32(0); // extreloff
486 Write32(0); // nextrel
487 Write32(0); // locreloff
488 Write32(0); // nlocrel
490 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
493 void MachObjectWriter::WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
494 MCSymbolData &Data = *MSD.SymbolData;
495 const MCSymbol &Symbol = Data.getSymbol();
497 uint16_t Flags = Data.getFlags();
498 uint32_t Address = 0;
500 // Set the N_TYPE bits. See <mach-o/nlist.h>.
502 // FIXME: Are the prebound or indirect fields possible here?
503 if (Symbol.isUndefined())
504 Type = macho::STT_Undefined;
505 else if (Symbol.isAbsolute())
506 Type = macho::STT_Absolute;
508 Type = macho::STT_Section;
510 // FIXME: Set STAB bits.
512 if (Data.isPrivateExtern())
513 Type |= macho::STF_PrivateExtern;
516 if (Data.isExternal() || Symbol.isUndefined())
517 Type |= macho::STF_External;
519 // Compute the symbol address.
520 if (Symbol.isDefined()) {
521 if (Symbol.isAbsolute()) {
522 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
524 Address = getSymbolAddress(&Data, Layout);
526 } else if (Data.isCommon()) {
527 // Common symbols are encoded with the size in the address
528 // field, and their alignment in the flags.
529 Address = Data.getCommonSize();
531 // Common alignment is packed into the 'desc' bits.
532 if (unsigned Align = Data.getCommonAlignment()) {
533 unsigned Log2Size = Log2_32(Align);
534 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
536 report_fatal_error("invalid 'common' alignment '" +
538 // FIXME: Keep this mask with the SymbolFlags enumeration.
539 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
543 // struct nlist (12 bytes)
545 Write32(MSD.StringIndex);
547 Write8(MSD.SectionIndex);
549 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
558 void MachObjectWriter::RecordX86_64Relocation(const MCAssembler &Asm,
559 const MCAsmLayout &Layout,
560 const MCFragment *Fragment,
561 const MCFixup &Fixup,
563 uint64_t &FixedValue) {
564 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
565 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
566 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
569 uint32_t FixupOffset =
570 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
571 uint32_t FixupAddress =
572 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
575 unsigned IsExtern = 0;
578 Value = Target.getConstant();
581 // Compensate for the relocation offset, Darwin x86_64 relocations only have
582 // the addend and appear to have attempted to define it to be the actual
583 // expression addend without the PCrel bias. However, instructions with data
584 // following the relocation are not accommodated for (see comment below
585 // regarding SIGNED{1,2,4}), so it isn't exactly that either.
586 Value += 1LL << Log2Size;
589 if (Target.isAbsolute()) { // constant
590 // SymbolNum of 0 indicates the absolute section.
591 Type = macho::RIT_X86_64_Unsigned;
594 // FIXME: I believe this is broken, I don't think the linker can understand
595 // it. I think it would require a local relocation, but I'm not sure if that
596 // would work either. The official way to get an absolute PCrel relocation
597 // is to use an absolute symbol (which we don't support yet).
600 Type = macho::RIT_X86_64_Branch;
602 } else if (Target.getSymB()) { // A - B + constant
603 const MCSymbol *A = &Target.getSymA()->getSymbol();
604 MCSymbolData &A_SD = Asm.getSymbolData(*A);
605 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
607 const MCSymbol *B = &Target.getSymB()->getSymbol();
608 MCSymbolData &B_SD = Asm.getSymbolData(*B);
609 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
611 // Neither symbol can be modified.
612 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
613 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
614 report_fatal_error("unsupported relocation of modified symbol");
616 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
617 // implement most of these correctly.
619 report_fatal_error("unsupported pc-relative relocation of difference");
621 // The support for the situation where one or both of the symbols would
622 // require a local relocation is handled just like if the symbols were
623 // external. This is certainly used in the case of debug sections where the
624 // section has only temporary symbols and thus the symbols don't have base
625 // symbols. This is encoded using the section ordinal and non-extern
626 // relocation entries.
628 // Darwin 'as' doesn't emit correct relocations for this (it ends up with a
629 // single SIGNED relocation); reject it for now. Except the case where both
630 // symbols don't have a base, equal but both NULL.
631 if (A_Base == B_Base && A_Base)
632 report_fatal_error("unsupported relocation with identical base");
634 Value += getSymbolAddress(&A_SD, Layout) -
635 (A_Base == NULL ? 0 : getSymbolAddress(A_Base, Layout));
636 Value -= getSymbolAddress(&B_SD, Layout) -
637 (B_Base == NULL ? 0 : getSymbolAddress(B_Base, Layout));
640 Index = A_Base->getIndex();
644 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
647 Type = macho::RIT_X86_64_Unsigned;
649 macho::RelocationEntry MRE;
650 MRE.Word0 = FixupOffset;
651 MRE.Word1 = ((Index << 0) |
656 Relocations[Fragment->getParent()].push_back(MRE);
659 Index = B_Base->getIndex();
663 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
666 Type = macho::RIT_X86_64_Subtractor;
668 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
669 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
670 const MCSymbolData *Base = Asm.getAtom(&SD);
672 // Relocations inside debug sections always use local relocations when
673 // possible. This seems to be done because the debugger doesn't fully
674 // understand x86_64 relocation entries, and expects to find values that
675 // have already been fixed up.
676 if (Symbol->isInSection()) {
677 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
678 Fragment->getParent()->getSection());
679 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
683 // x86_64 almost always uses external relocations, except when there is no
684 // symbol to use as a base address (a local symbol with no preceding
685 // non-local symbol).
687 Index = Base->getIndex();
690 // Add the local offset, if needed.
692 Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base);
693 } else if (Symbol->isInSection() && !Symbol->isVariable()) {
694 // The index is the section ordinal (1-based).
695 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
697 Value += getSymbolAddress(&SD, Layout);
700 Value -= FixupAddress + (1 << Log2Size);
701 } else if (Symbol->isVariable()) {
702 const MCExpr *Value = Symbol->getVariableValue();
704 bool isAbs = Value->EvaluateAsAbsolute(Res, Layout, SectionAddress);
709 report_fatal_error("unsupported relocation of variable '" +
710 Symbol->getName() + "'");
713 report_fatal_error("unsupported relocation of undefined symbol '" +
714 Symbol->getName() + "'");
717 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
720 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
721 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
722 // rewrite the movq to an leaq at link time if the symbol ends up in
723 // the same linkage unit.
724 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
725 Type = macho::RIT_X86_64_GOTLoad;
727 Type = macho::RIT_X86_64_GOT;
728 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
729 Type = macho::RIT_X86_64_TLV;
730 } else if (Modifier != MCSymbolRefExpr::VK_None) {
731 report_fatal_error("unsupported symbol modifier in relocation");
733 Type = macho::RIT_X86_64_Signed;
735 // The Darwin x86_64 relocation format has a problem where it cannot
736 // encode an address (L<foo> + <constant>) which is outside the atom
737 // containing L<foo>. Generally, this shouldn't occur but it does
738 // happen when we have a RIPrel instruction with data following the
739 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
740 // adjustment Darwin x86_64 uses, the offset is still negative and the
741 // linker has no way to recognize this.
743 // To work around this, Darwin uses several special relocation types
744 // to indicate the offsets. However, the specification or
745 // implementation of these seems to also be incomplete; they should
746 // adjust the addend as well based on the actual encoded instruction
747 // (the additional bias), but instead appear to just look at the final
749 switch (-(Target.getConstant() + (1LL << Log2Size))) {
750 case 1: Type = macho::RIT_X86_64_Signed1; break;
751 case 2: Type = macho::RIT_X86_64_Signed2; break;
752 case 4: Type = macho::RIT_X86_64_Signed4; break;
756 if (Modifier != MCSymbolRefExpr::VK_None)
757 report_fatal_error("unsupported symbol modifier in branch "
760 Type = macho::RIT_X86_64_Branch;
763 if (Modifier == MCSymbolRefExpr::VK_GOT) {
764 Type = macho::RIT_X86_64_GOT;
765 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
766 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in which
767 // case all we do is set the PCrel bit in the relocation entry; this is
768 // used with exception handling, for example. The source is required to
769 // include any necessary offset directly.
770 Type = macho::RIT_X86_64_GOT;
772 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
773 report_fatal_error("TLVP symbol modifier should have been rip-rel");
774 } else if (Modifier != MCSymbolRefExpr::VK_None)
775 report_fatal_error("unsupported symbol modifier in relocation");
777 Type = macho::RIT_X86_64_Unsigned;
781 // x86_64 always writes custom values into the fixups.
784 // struct relocation_info (8 bytes)
785 macho::RelocationEntry MRE;
786 MRE.Word0 = FixupOffset;
787 MRE.Word1 = ((Index << 0) |
792 Relocations[Fragment->getParent()].push_back(MRE);
795 void MachObjectWriter::RecordScatteredRelocation(const MCAssembler &Asm,
796 const MCAsmLayout &Layout,
797 const MCFragment *Fragment,
798 const MCFixup &Fixup,
801 uint64_t &FixedValue) {
802 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
803 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
804 unsigned Type = macho::RIT_Vanilla;
807 const MCSymbol *A = &Target.getSymA()->getSymbol();
808 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
810 if (!A_SD->getFragment())
811 report_fatal_error("symbol '" + A->getName() +
812 "' can not be undefined in a subtraction expression");
814 uint32_t Value = getSymbolAddress(A_SD, Layout);
815 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
816 FixedValue += SecAddr;
819 if (const MCSymbolRefExpr *B = Target.getSymB()) {
820 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
822 if (!B_SD->getFragment())
823 report_fatal_error("symbol '" + B->getSymbol().getName() +
824 "' can not be undefined in a subtraction expression");
826 // Select the appropriate difference relocation type.
828 // Note that there is no longer any semantic difference between these two
829 // relocation types from the linkers point of view, this is done solely for
830 // pedantic compatibility with 'as'.
831 Type = A_SD->isExternal() ? (unsigned)macho::RIT_Difference :
832 (unsigned)macho::RIT_Generic_LocalDifference;
833 Value2 = getSymbolAddress(B_SD, Layout);
834 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
837 // Relocations are written out in reverse order, so the PAIR comes first.
838 if (Type == macho::RIT_Difference ||
839 Type == macho::RIT_Generic_LocalDifference) {
840 macho::RelocationEntry MRE;
841 MRE.Word0 = ((0 << 0) |
842 (macho::RIT_Pair << 24) |
845 macho::RF_Scattered);
847 Relocations[Fragment->getParent()].push_back(MRE);
850 macho::RelocationEntry MRE;
851 MRE.Word0 = ((FixupOffset << 0) |
855 macho::RF_Scattered);
857 Relocations[Fragment->getParent()].push_back(MRE);
860 void MachObjectWriter::RecordARMScatteredRelocation(const MCAssembler &Asm,
861 const MCAsmLayout &Layout,
862 const MCFragment *Fragment,
863 const MCFixup &Fixup,
866 uint64_t &FixedValue) {
867 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
868 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
869 unsigned Type = macho::RIT_Vanilla;
872 const MCSymbol *A = &Target.getSymA()->getSymbol();
873 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
875 if (!A_SD->getFragment())
876 report_fatal_error("symbol '" + A->getName() +
877 "' can not be undefined in a subtraction expression");
879 uint32_t Value = getSymbolAddress(A_SD, Layout);
880 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
881 FixedValue += SecAddr;
884 if (const MCSymbolRefExpr *B = Target.getSymB()) {
885 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
887 if (!B_SD->getFragment())
888 report_fatal_error("symbol '" + B->getSymbol().getName() +
889 "' can not be undefined in a subtraction expression");
891 // Select the appropriate difference relocation type.
892 Type = macho::RIT_Difference;
893 Value2 = getSymbolAddress(B_SD, Layout);
894 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
897 // Relocations are written out in reverse order, so the PAIR comes first.
898 if (Type == macho::RIT_Difference ||
899 Type == macho::RIT_Generic_LocalDifference) {
900 macho::RelocationEntry MRE;
901 MRE.Word0 = ((0 << 0) |
902 (macho::RIT_Pair << 24) |
905 macho::RF_Scattered);
907 Relocations[Fragment->getParent()].push_back(MRE);
910 macho::RelocationEntry MRE;
911 MRE.Word0 = ((FixupOffset << 0) |
915 macho::RF_Scattered);
917 Relocations[Fragment->getParent()].push_back(MRE);
920 void MachObjectWriter::RecordARMMovwMovtRelocation(const MCAssembler &Asm,
921 const MCAsmLayout &Layout,
922 const MCFragment *Fragment,
923 const MCFixup &Fixup,
925 uint64_t &FixedValue) {
926 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
927 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
928 unsigned Type = macho::RIT_ARM_Half;
931 const MCSymbol *A = &Target.getSymA()->getSymbol();
932 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
934 if (!A_SD->getFragment())
935 report_fatal_error("symbol '" + A->getName() +
936 "' can not be undefined in a subtraction expression");
938 uint32_t Value = getSymbolAddress(A_SD, Layout);
940 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
941 FixedValue += SecAddr;
943 if (const MCSymbolRefExpr *B = Target.getSymB()) {
944 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
946 if (!B_SD->getFragment())
947 report_fatal_error("symbol '" + B->getSymbol().getName() +
948 "' can not be undefined in a subtraction expression");
950 // Select the appropriate difference relocation type.
951 Type = macho::RIT_ARM_HalfDifference;
952 Value2 = getSymbolAddress(B_SD, Layout);
953 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
956 // Relocations are written out in reverse order, so the PAIR comes first.
957 // ARM_RELOC_HALF and ARM_RELOC_HALF_SECTDIFF abuse the r_length field:
959 // For these two r_type relocations they always have a pair following them and
960 // the r_length bits are used differently. The encoding of the r_length is as
962 // low bit of r_length:
963 // 0 - :lower16: for movw instructions
964 // 1 - :upper16: for movt instructions
965 // high bit of r_length:
966 // 0 - arm instructions
967 // 1 - thumb instructions
968 // the other half of the relocated expression is in the following pair
969 // relocation entry in the the low 16 bits of r_address field.
970 unsigned ThumbBit = 0;
971 unsigned MovtBit = 0;
972 switch ((unsigned)Fixup.getKind()) {
974 case ARM::fixup_arm_movt_hi16:
975 case ARM::fixup_arm_movt_hi16_pcrel:
978 case ARM::fixup_t2_movt_hi16:
979 case ARM::fixup_t2_movt_hi16_pcrel:
982 case ARM::fixup_t2_movw_lo16:
983 case ARM::fixup_t2_movw_lo16_pcrel:
989 if (Type == macho::RIT_ARM_HalfDifference) {
990 uint32_t OtherHalf = MovtBit
991 ? (FixedValue & 0xffff) : ((FixedValue & 0xffff0000) >> 16);
993 macho::RelocationEntry MRE;
994 MRE.Word0 = ((OtherHalf << 0) |
995 (macho::RIT_Pair << 24) |
999 macho::RF_Scattered);
1001 Relocations[Fragment->getParent()].push_back(MRE);
1004 macho::RelocationEntry MRE;
1005 MRE.Word0 = ((FixupOffset << 0) |
1010 macho::RF_Scattered);
1012 Relocations[Fragment->getParent()].push_back(MRE);
1015 void MachObjectWriter::RecordTLVPRelocation(const MCAssembler &Asm,
1016 const MCAsmLayout &Layout,
1017 const MCFragment *Fragment,
1018 const MCFixup &Fixup,
1020 uint64_t &FixedValue) {
1021 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
1023 "Should only be called with a 32-bit TLVP relocation!");
1025 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
1026 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1027 unsigned IsPCRel = 0;
1029 // Get the symbol data.
1030 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1031 unsigned Index = SD_A->getIndex();
1033 // We're only going to have a second symbol in pic mode and it'll be a
1034 // subtraction from the picbase. For 32-bit pic the addend is the difference
1035 // between the picbase and the next address. For 32-bit static the addend is
1037 if (Target.getSymB()) {
1038 // If this is a subtraction then we're pcrel.
1039 uint32_t FixupAddress =
1040 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
1041 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
1043 FixedValue = (FixupAddress - getSymbolAddress(SD_B, Layout) +
1044 Target.getConstant());
1045 FixedValue += 1ULL << Log2Size;
1050 // struct relocation_info (8 bytes)
1051 macho::RelocationEntry MRE;
1053 MRE.Word1 = ((Index << 0) |
1056 (1 << 27) | // Extern
1057 (macho::RIT_Generic_TLV << 28)); // Type
1058 Relocations[Fragment->getParent()].push_back(MRE);
1061 bool MachObjectWriter::getARMFixupKindMachOInfo(unsigned Kind, unsigned &RelocType,
1062 unsigned &Log2Size) {
1063 RelocType = unsigned(macho::RIT_Vanilla);
1071 Log2Size = llvm::Log2_32(1);
1074 Log2Size = llvm::Log2_32(2);
1077 Log2Size = llvm::Log2_32(4);
1080 Log2Size = llvm::Log2_32(8);
1083 // Handle 24-bit branch kinds.
1084 case ARM::fixup_arm_ldst_pcrel_12:
1085 case ARM::fixup_arm_pcrel_10:
1086 case ARM::fixup_arm_adr_pcrel_12:
1087 case ARM::fixup_arm_condbranch:
1088 case ARM::fixup_arm_uncondbranch:
1089 RelocType = unsigned(macho::RIT_ARM_Branch24Bit);
1090 // Report as 'long', even though that is not quite accurate.
1091 Log2Size = llvm::Log2_32(4);
1094 // Handle Thumb branches.
1095 case ARM::fixup_arm_thumb_br:
1096 RelocType = unsigned(macho::RIT_ARM_ThumbBranch22Bit);
1097 Log2Size = llvm::Log2_32(2);
1100 case ARM::fixup_arm_thumb_bl:
1101 case ARM::fixup_arm_thumb_blx:
1102 RelocType = unsigned(macho::RIT_ARM_ThumbBranch22Bit);
1103 Log2Size = llvm::Log2_32(4);
1106 case ARM::fixup_arm_movt_hi16:
1107 case ARM::fixup_arm_movt_hi16_pcrel:
1108 case ARM::fixup_t2_movt_hi16:
1109 case ARM::fixup_t2_movt_hi16_pcrel:
1110 RelocType = unsigned(macho::RIT_ARM_HalfDifference);
1111 // Report as 'long', even though that is not quite accurate.
1112 Log2Size = llvm::Log2_32(4);
1115 case ARM::fixup_arm_movw_lo16:
1116 case ARM::fixup_arm_movw_lo16_pcrel:
1117 case ARM::fixup_t2_movw_lo16:
1118 case ARM::fixup_t2_movw_lo16_pcrel:
1119 RelocType = unsigned(macho::RIT_ARM_Half);
1120 // Report as 'long', even though that is not quite accurate.
1121 Log2Size = llvm::Log2_32(4);
1125 void MachObjectWriter::RecordARMRelocation(const MCAssembler &Asm,
1126 const MCAsmLayout &Layout,
1127 const MCFragment *Fragment,
1128 const MCFixup &Fixup,
1130 uint64_t &FixedValue) {
1131 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
1133 unsigned RelocType = macho::RIT_Vanilla;
1134 if (!getARMFixupKindMachOInfo(Fixup.getKind(), RelocType, Log2Size)) {
1135 report_fatal_error("unknown ARM fixup kind!");
1139 // If this is a difference or a defined symbol plus an offset, then we need a
1140 // scattered relocation entry. Differences always require scattered
1142 if (Target.getSymB()) {
1143 if (RelocType == macho::RIT_ARM_Half ||
1144 RelocType == macho::RIT_ARM_HalfDifference)
1145 return RecordARMMovwMovtRelocation(Asm, Layout, Fragment, Fixup,
1146 Target, FixedValue);
1147 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup,
1148 Target, Log2Size, FixedValue);
1151 // Get the symbol data, if any.
1152 MCSymbolData *SD = 0;
1153 if (Target.getSymA())
1154 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1156 // FIXME: For other platforms, we need to use scattered relocations for
1157 // internal relocations with offsets. If this is an internal relocation with
1158 // an offset, it also needs a scattered relocation entry.
1160 // Is this right for ARM?
1161 uint32_t Offset = Target.getConstant();
1162 if (IsPCRel && RelocType == macho::RIT_Vanilla)
1163 Offset += 1 << Log2Size;
1164 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
1165 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup, Target,
1166 Log2Size, FixedValue);
1169 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1171 unsigned IsExtern = 0;
1174 if (Target.isAbsolute()) { // constant
1176 report_fatal_error("FIXME: relocations to absolute targets "
1177 "not yet implemented");
1179 // Resolve constant variables.
1180 if (SD->getSymbol().isVariable()) {
1182 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
1183 Res, Layout, SectionAddress)) {
1189 // Check whether we need an external or internal relocation.
1190 if (doesSymbolRequireExternRelocation(SD)) {
1192 Index = SD->getIndex();
1194 // For external relocations, make sure to offset the fixup value to
1195 // compensate for the addend of the symbol address, if it was
1196 // undefined. This occurs with weak definitions, for example.
1197 if (!SD->Symbol->isUndefined())
1198 FixedValue -= Layout.getSymbolOffset(SD);
1200 // The index is the section ordinal (1-based).
1201 const MCSectionData &SymSD = Asm.getSectionData(
1202 SD->getSymbol().getSection());
1203 Index = SymSD.getOrdinal() + 1;
1204 FixedValue += getSectionAddress(&SymSD);
1207 FixedValue -= getSectionAddress(Fragment->getParent());
1209 // The type is determined by the fixup kind.
1213 // struct relocation_info (8 bytes)
1214 macho::RelocationEntry MRE;
1215 MRE.Word0 = FixupOffset;
1216 MRE.Word1 = ((Index << 0) |
1221 Relocations[Fragment->getParent()].push_back(MRE);
1224 void MachObjectWriter::RecordRelocation(const MCAssembler &Asm,
1225 const MCAsmLayout &Layout,
1226 const MCFragment *Fragment,
1227 const MCFixup &Fixup,
1229 uint64_t &FixedValue) {
1230 // FIXME: These needs to be factored into the target Mach-O writer.
1232 RecordARMRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1236 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1240 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
1241 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
1243 // If this is a 32-bit TLVP reloc it's handled a bit differently.
1244 if (Target.getSymA() &&
1245 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
1246 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1250 // If this is a difference or a defined symbol plus an offset, then we need a
1251 // scattered relocation entry. Differences always require scattered
1253 if (Target.getSymB())
1254 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1255 Target, Log2Size, FixedValue);
1257 // Get the symbol data, if any.
1258 MCSymbolData *SD = 0;
1259 if (Target.getSymA())
1260 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1262 // If this is an internal relocation with an offset, it also needs a scattered
1263 // relocation entry.
1264 uint32_t Offset = Target.getConstant();
1266 Offset += 1 << Log2Size;
1267 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
1268 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1269 Target, Log2Size, FixedValue);
1272 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1274 unsigned IsExtern = 0;
1277 if (Target.isAbsolute()) { // constant
1278 // SymbolNum of 0 indicates the absolute section.
1280 // FIXME: Currently, these are never generated (see code below). I cannot
1281 // find a case where they are actually emitted.
1282 Type = macho::RIT_Vanilla;
1284 // Resolve constant variables.
1285 if (SD->getSymbol().isVariable()) {
1287 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
1288 Res, Layout, SectionAddress)) {
1294 // Check whether we need an external or internal relocation.
1295 if (doesSymbolRequireExternRelocation(SD)) {
1297 Index = SD->getIndex();
1298 // For external relocations, make sure to offset the fixup value to
1299 // compensate for the addend of the symbol address, if it was
1300 // undefined. This occurs with weak definitions, for example.
1301 if (!SD->Symbol->isUndefined())
1302 FixedValue -= Layout.getSymbolOffset(SD);
1304 // The index is the section ordinal (1-based).
1305 const MCSectionData &SymSD = Asm.getSectionData(
1306 SD->getSymbol().getSection());
1307 Index = SymSD.getOrdinal() + 1;
1308 FixedValue += getSectionAddress(&SymSD);
1311 FixedValue -= getSectionAddress(Fragment->getParent());
1313 Type = macho::RIT_Vanilla;
1316 // struct relocation_info (8 bytes)
1317 macho::RelocationEntry MRE;
1318 MRE.Word0 = FixupOffset;
1319 MRE.Word1 = ((Index << 0) |
1324 Relocations[Fragment->getParent()].push_back(MRE);
1327 void MachObjectWriter::BindIndirectSymbols(MCAssembler &Asm) {
1328 // This is the point where 'as' creates actual symbols for indirect symbols
1329 // (in the following two passes). It would be easier for us to do this sooner
1330 // when we see the attribute, but that makes getting the order in the symbol
1331 // table much more complicated than it is worth.
1333 // FIXME: Revisit this when the dust settles.
1335 // Bind non lazy symbol pointers first.
1336 unsigned IndirectIndex = 0;
1337 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1338 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1339 const MCSectionMachO &Section =
1340 cast<MCSectionMachO>(it->SectionData->getSection());
1342 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
1345 // Initialize the section indirect symbol base, if necessary.
1346 if (!IndirectSymBase.count(it->SectionData))
1347 IndirectSymBase[it->SectionData] = IndirectIndex;
1349 Asm.getOrCreateSymbolData(*it->Symbol);
1352 // Then lazy symbol pointers and symbol stubs.
1354 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1355 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1356 const MCSectionMachO &Section =
1357 cast<MCSectionMachO>(it->SectionData->getSection());
1359 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
1360 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
1363 // Initialize the section indirect symbol base, if necessary.
1364 if (!IndirectSymBase.count(it->SectionData))
1365 IndirectSymBase[it->SectionData] = IndirectIndex;
1367 // Set the symbol type to undefined lazy, but only on construction.
1369 // FIXME: Do not hardcode.
1371 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
1373 Entry.setFlags(Entry.getFlags() | 0x0001);
1377 /// ComputeSymbolTable - Compute the symbol table data
1379 /// \param StringTable [out] - The string table data.
1380 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
1382 void MachObjectWriter::ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
1383 std::vector<MachSymbolData> &LocalSymbolData,
1384 std::vector<MachSymbolData> &ExternalSymbolData,
1385 std::vector<MachSymbolData> &UndefinedSymbolData) {
1386 // Build section lookup table.
1387 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
1389 for (MCAssembler::iterator it = Asm.begin(),
1390 ie = Asm.end(); it != ie; ++it, ++Index)
1391 SectionIndexMap[&it->getSection()] = Index;
1392 assert(Index <= 256 && "Too many sections!");
1394 // Index 0 is always the empty string.
1395 StringMap<uint64_t> StringIndexMap;
1396 StringTable += '\x00';
1398 // Build the symbol arrays and the string table, but only for non-local
1401 // The particular order that we collect the symbols and create the string
1402 // table, then sort the symbols is chosen to match 'as'. Even though it
1403 // doesn't matter for correctness, this is important for letting us diff .o
1405 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1406 ie = Asm.symbol_end(); it != ie; ++it) {
1407 const MCSymbol &Symbol = it->getSymbol();
1409 // Ignore non-linker visible symbols.
1410 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1413 if (!it->isExternal() && !Symbol.isUndefined())
1416 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1418 Entry = StringTable.size();
1419 StringTable += Symbol.getName();
1420 StringTable += '\x00';
1424 MSD.SymbolData = it;
1425 MSD.StringIndex = Entry;
1427 if (Symbol.isUndefined()) {
1428 MSD.SectionIndex = 0;
1429 UndefinedSymbolData.push_back(MSD);
1430 } else if (Symbol.isAbsolute()) {
1431 MSD.SectionIndex = 0;
1432 ExternalSymbolData.push_back(MSD);
1434 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1435 assert(MSD.SectionIndex && "Invalid section index!");
1436 ExternalSymbolData.push_back(MSD);
1440 // Now add the data for local symbols.
1441 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1442 ie = Asm.symbol_end(); it != ie; ++it) {
1443 const MCSymbol &Symbol = it->getSymbol();
1445 // Ignore non-linker visible symbols.
1446 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1449 if (it->isExternal() || Symbol.isUndefined())
1452 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1454 Entry = StringTable.size();
1455 StringTable += Symbol.getName();
1456 StringTable += '\x00';
1460 MSD.SymbolData = it;
1461 MSD.StringIndex = Entry;
1463 if (Symbol.isAbsolute()) {
1464 MSD.SectionIndex = 0;
1465 LocalSymbolData.push_back(MSD);
1467 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1468 assert(MSD.SectionIndex && "Invalid section index!");
1469 LocalSymbolData.push_back(MSD);
1473 // External and undefined symbols are required to be in lexicographic order.
1474 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1475 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1477 // Set the symbol indices.
1479 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1480 LocalSymbolData[i].SymbolData->setIndex(Index++);
1481 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1482 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1483 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1484 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1486 // The string table is padded to a multiple of 4.
1487 while (StringTable.size() % 4)
1488 StringTable += '\x00';
1491 void MachObjectWriter::computeSectionAddresses(const MCAssembler &Asm,
1492 const MCAsmLayout &Layout) {
1493 uint64_t StartAddress = 0;
1494 const SmallVectorImpl<MCSectionData*> &Order = Layout.getSectionOrder();
1495 for (int i = 0, n = Order.size(); i != n ; ++i) {
1496 const MCSectionData *SD = Order[i];
1497 StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment());
1498 SectionAddress[SD] = StartAddress;
1499 StartAddress += Layout.getSectionAddressSize(SD);
1501 // Explicitly pad the section to match the alignment requirements of the
1502 // following one. This is for 'gas' compatibility, it shouldn't
1503 /// strictly be necessary.
1504 StartAddress += getPaddingSize(SD, Layout);
1508 void MachObjectWriter::ExecutePostLayoutBinding(MCAssembler &Asm, const MCAsmLayout &Layout) {
1509 computeSectionAddresses(Asm, Layout);
1511 // Create symbol data for any indirect symbols.
1512 BindIndirectSymbols(Asm);
1514 // Compute symbol table information and bind symbol indices.
1515 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1516 UndefinedSymbolData);
1519 bool MachObjectWriter::IsSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
1520 const MCSymbolData &DataA,
1521 const MCFragment &FB,
1523 bool IsPCRel) const {
1527 // The effective address is
1528 // addr(atom(A)) + offset(A)
1529 // - addr(atom(B)) - offset(B)
1530 // and the offsets are not relocatable, so the fixup is fully resolved when
1531 // addr(atom(A)) - addr(atom(B)) == 0.
1532 const MCSymbolData *A_Base = 0, *B_Base = 0;
1534 const MCSymbol &SA = DataA.getSymbol().AliasedSymbol();
1535 const MCSection &SecA = SA.getSection();
1536 const MCSection &SecB = FB.getParent()->getSection();
1539 // The simple (Darwin, except on x86_64) way of dealing with this was to
1540 // assume that any reference to a temporary symbol *must* be a temporary
1541 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
1542 // relocation to a temporary symbol (in the same section) is fully
1543 // resolved. This also works in conjunction with absolutized .set, which
1544 // requires the compiler to use .set to absolutize the differences between
1545 // symbols which the compiler knows to be assembly time constants, so we
1546 // don't need to worry about considering symbol differences fully resolved.
1548 if (!Asm.getBackend().hasReliableSymbolDifference()) {
1549 if (!SA.isTemporary() || !SA.isInSection() || &SecA != &SecB)
1554 if (!TargetObjectWriter->useAggressiveSymbolFolding())
1558 const MCFragment &FA = *Asm.getSymbolData(SA).getFragment();
1560 A_Base = FA.getAtom();
1564 B_Base = FB.getAtom();
1568 // If the atoms are the same, they are guaranteed to have the same address.
1569 if (A_Base == B_Base)
1572 // Otherwise, we can't prove this is fully resolved.
1576 void MachObjectWriter::WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout) {
1577 unsigned NumSections = Asm.size();
1579 // The section data starts after the header, the segment load command (and
1580 // section headers) and the symbol table.
1581 unsigned NumLoadCommands = 1;
1582 uint64_t LoadCommandsSize = is64Bit() ?
1583 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
1584 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
1586 // Add the symbol table load command sizes, if used.
1587 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1588 UndefinedSymbolData.size();
1590 NumLoadCommands += 2;
1591 LoadCommandsSize += (macho::SymtabLoadCommandSize +
1592 macho::DysymtabLoadCommandSize);
1595 // Compute the total size of the section data, as well as its file size and vm
1597 uint64_t SectionDataStart = (is64Bit() ? macho::Header64Size :
1598 macho::Header32Size) + LoadCommandsSize;
1599 uint64_t SectionDataSize = 0;
1600 uint64_t SectionDataFileSize = 0;
1601 uint64_t VMSize = 0;
1602 for (MCAssembler::const_iterator it = Asm.begin(),
1603 ie = Asm.end(); it != ie; ++it) {
1604 const MCSectionData &SD = *it;
1605 uint64_t Address = getSectionAddress(&SD);
1606 uint64_t Size = Layout.getSectionAddressSize(&SD);
1607 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1608 FileSize += getPaddingSize(&SD, Layout);
1610 VMSize = std::max(VMSize, Address + Size);
1612 if (SD.getSection().isVirtualSection())
1615 SectionDataSize = std::max(SectionDataSize, Address + Size);
1616 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1619 // The section data is padded to 4 bytes.
1621 // FIXME: Is this machine dependent?
1622 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1623 SectionDataFileSize += SectionDataPadding;
1625 // Write the prolog, starting with the header and load command...
1626 WriteHeader(NumLoadCommands, LoadCommandsSize,
1627 Asm.getSubsectionsViaSymbols());
1628 WriteSegmentLoadCommand(NumSections, VMSize,
1629 SectionDataStart, SectionDataSize);
1631 // ... and then the section headers.
1632 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1633 for (MCAssembler::const_iterator it = Asm.begin(),
1634 ie = Asm.end(); it != ie; ++it) {
1635 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1636 unsigned NumRelocs = Relocs.size();
1637 uint64_t SectionStart = SectionDataStart + getSectionAddress(it);
1638 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1639 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
1642 // Write the symbol table load command, if used.
1644 unsigned FirstLocalSymbol = 0;
1645 unsigned NumLocalSymbols = LocalSymbolData.size();
1646 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1647 unsigned NumExternalSymbols = ExternalSymbolData.size();
1648 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1649 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1650 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1651 unsigned NumSymTabSymbols =
1652 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1653 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1654 uint64_t IndirectSymbolOffset = 0;
1656 // If used, the indirect symbols are written after the section data.
1657 if (NumIndirectSymbols)
1658 IndirectSymbolOffset = RelocTableEnd;
1660 // The symbol table is written after the indirect symbol data.
1661 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1663 // The string table is written after symbol table.
1664 uint64_t StringTableOffset =
1665 SymbolTableOffset + NumSymTabSymbols * (is64Bit() ? macho::Nlist64Size :
1666 macho::Nlist32Size);
1667 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1668 StringTableOffset, StringTable.size());
1670 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1671 FirstExternalSymbol, NumExternalSymbols,
1672 FirstUndefinedSymbol, NumUndefinedSymbols,
1673 IndirectSymbolOffset, NumIndirectSymbols);
1676 // Write the actual section data.
1677 for (MCAssembler::const_iterator it = Asm.begin(),
1678 ie = Asm.end(); it != ie; ++it) {
1679 Asm.WriteSectionData(it, Layout);
1681 uint64_t Pad = getPaddingSize(it, Layout);
1682 for (unsigned int i = 0; i < Pad; ++i)
1686 // Write the extra padding.
1687 WriteZeros(SectionDataPadding);
1689 // Write the relocation entries.
1690 for (MCAssembler::const_iterator it = Asm.begin(),
1691 ie = Asm.end(); it != ie; ++it) {
1692 // Write the section relocation entries, in reverse order to match 'as'
1693 // (approximately, the exact algorithm is more complicated than this).
1694 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1695 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1696 Write32(Relocs[e - i - 1].Word0);
1697 Write32(Relocs[e - i - 1].Word1);
1701 // Write the symbol table data, if used.
1703 // Write the indirect symbol entries.
1704 for (MCAssembler::const_indirect_symbol_iterator
1705 it = Asm.indirect_symbol_begin(),
1706 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1707 // Indirect symbols in the non lazy symbol pointer section have some
1708 // special handling.
1709 const MCSectionMachO &Section =
1710 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1711 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1712 // If this symbol is defined and internal, mark it as such.
1713 if (it->Symbol->isDefined() &&
1714 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1715 uint32_t Flags = macho::ISF_Local;
1716 if (it->Symbol->isAbsolute())
1717 Flags |= macho::ISF_Absolute;
1723 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1726 // FIXME: Check that offsets match computed ones.
1728 // Write the symbol table entries.
1729 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1730 WriteNlist(LocalSymbolData[i], Layout);
1731 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1732 WriteNlist(ExternalSymbolData[i], Layout);
1733 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1734 WriteNlist(UndefinedSymbolData[i], Layout);
1736 // Write the string table.
1737 OS << StringTable.str();
1741 MCObjectWriter *llvm::createMachObjectWriter(MCMachObjectTargetWriter *MOTW,
1743 bool IsLittleEndian) {
1744 return new MachObjectWriter(MOTW, OS, IsLittleEndian);