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 {
124 return getSectionAddress(SD->getFragment()->getParent()) +
125 Layout.getSymbolOffset(SD);
127 uint64_t getFragmentAddress(const MCFragment *Fragment,
128 const MCAsmLayout &Layout) const {
129 return getSectionAddress(Fragment->getParent()) +
130 Layout.getFragmentOffset(Fragment);
133 uint64_t getPaddingSize(const MCSectionData *SD,
134 const MCAsmLayout &Layout) const {
135 uint64_t EndAddr = getSectionAddress(SD) + Layout.getSectionAddressSize(SD);
136 unsigned Next = SD->getLayoutOrder() + 1;
137 if (Next >= Layout.getSectionOrder().size())
140 const MCSectionData &NextSD = *Layout.getSectionOrder()[Next];
141 if (NextSD.getSection().isVirtualSection())
143 return OffsetToAlignment(EndAddr, NextSD.getAlignment());
147 MachObjectWriter(MCMachObjectTargetWriter *MOTW, raw_ostream &_OS,
148 bool _IsLittleEndian)
149 : MCObjectWriter(_OS, _IsLittleEndian), TargetObjectWriter(MOTW) {
152 /// @name Target Writer Proxy Accessors
155 bool is64Bit() const { return TargetObjectWriter->is64Bit(); }
157 uint32_t CPUType = TargetObjectWriter->getCPUType() & ~mach::CTFM_ArchMask;
158 return CPUType == mach::CTM_ARM;
163 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
164 bool SubsectionsViaSymbols) {
167 if (SubsectionsViaSymbols)
168 Flags |= macho::HF_SubsectionsViaSymbols;
170 // struct mach_header (28 bytes) or
171 // struct mach_header_64 (32 bytes)
173 uint64_t Start = OS.tell();
176 Write32(is64Bit() ? macho::HM_Object64 : macho::HM_Object32);
178 Write32(TargetObjectWriter->getCPUType());
179 Write32(TargetObjectWriter->getCPUSubtype());
181 Write32(macho::HFT_Object);
182 Write32(NumLoadCommands);
183 Write32(LoadCommandsSize);
186 Write32(0); // reserved
188 assert(OS.tell() - Start == is64Bit() ?
189 macho::Header64Size : macho::Header32Size);
192 /// WriteSegmentLoadCommand - Write a segment load command.
194 /// \arg NumSections - The number of sections in this segment.
195 /// \arg SectionDataSize - The total size of the sections.
196 void WriteSegmentLoadCommand(unsigned NumSections,
198 uint64_t SectionDataStartOffset,
199 uint64_t SectionDataSize) {
200 // struct segment_command (56 bytes) or
201 // struct segment_command_64 (72 bytes)
203 uint64_t Start = OS.tell();
206 unsigned SegmentLoadCommandSize =
207 is64Bit() ? macho::SegmentLoadCommand64Size:
208 macho::SegmentLoadCommand32Size;
209 Write32(is64Bit() ? macho::LCT_Segment64 : macho::LCT_Segment);
210 Write32(SegmentLoadCommandSize +
211 NumSections * (is64Bit() ? macho::Section64Size :
212 macho::Section32Size));
216 Write64(0); // vmaddr
217 Write64(VMSize); // vmsize
218 Write64(SectionDataStartOffset); // file offset
219 Write64(SectionDataSize); // file size
221 Write32(0); // vmaddr
222 Write32(VMSize); // vmsize
223 Write32(SectionDataStartOffset); // file offset
224 Write32(SectionDataSize); // file size
226 Write32(0x7); // maxprot
227 Write32(0x7); // initprot
228 Write32(NumSections);
231 assert(OS.tell() - Start == SegmentLoadCommandSize);
234 void WriteSection(const MCAssembler &Asm, const MCAsmLayout &Layout,
235 const MCSectionData &SD, uint64_t FileOffset,
236 uint64_t RelocationsStart, unsigned NumRelocations) {
237 uint64_t SectionSize = Layout.getSectionAddressSize(&SD);
239 // The offset is unused for virtual sections.
240 if (SD.getSection().isVirtualSection()) {
241 assert(Layout.getSectionFileSize(&SD) == 0 && "Invalid file size!");
245 // struct section (68 bytes) or
246 // struct section_64 (80 bytes)
248 uint64_t Start = OS.tell();
251 const MCSectionMachO &Section = cast<MCSectionMachO>(SD.getSection());
252 WriteBytes(Section.getSectionName(), 16);
253 WriteBytes(Section.getSegmentName(), 16);
255 Write64(getSectionAddress(&SD)); // address
256 Write64(SectionSize); // size
258 Write32(getSectionAddress(&SD)); // address
259 Write32(SectionSize); // size
263 unsigned Flags = Section.getTypeAndAttributes();
264 if (SD.hasInstructions())
265 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
267 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
268 Write32(Log2_32(SD.getAlignment()));
269 Write32(NumRelocations ? RelocationsStart : 0);
270 Write32(NumRelocations);
272 Write32(IndirectSymBase.lookup(&SD)); // reserved1
273 Write32(Section.getStubSize()); // reserved2
275 Write32(0); // reserved3
277 assert(OS.tell() - Start == is64Bit() ? macho::Section64Size :
278 macho::Section32Size);
281 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
282 uint32_t StringTableOffset,
283 uint32_t StringTableSize) {
284 // struct symtab_command (24 bytes)
286 uint64_t Start = OS.tell();
289 Write32(macho::LCT_Symtab);
290 Write32(macho::SymtabLoadCommandSize);
291 Write32(SymbolOffset);
293 Write32(StringTableOffset);
294 Write32(StringTableSize);
296 assert(OS.tell() - Start == macho::SymtabLoadCommandSize);
299 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
300 uint32_t NumLocalSymbols,
301 uint32_t FirstExternalSymbol,
302 uint32_t NumExternalSymbols,
303 uint32_t FirstUndefinedSymbol,
304 uint32_t NumUndefinedSymbols,
305 uint32_t IndirectSymbolOffset,
306 uint32_t NumIndirectSymbols) {
307 // struct dysymtab_command (80 bytes)
309 uint64_t Start = OS.tell();
312 Write32(macho::LCT_Dysymtab);
313 Write32(macho::DysymtabLoadCommandSize);
314 Write32(FirstLocalSymbol);
315 Write32(NumLocalSymbols);
316 Write32(FirstExternalSymbol);
317 Write32(NumExternalSymbols);
318 Write32(FirstUndefinedSymbol);
319 Write32(NumUndefinedSymbols);
320 Write32(0); // tocoff
322 Write32(0); // modtaboff
323 Write32(0); // nmodtab
324 Write32(0); // extrefsymoff
325 Write32(0); // nextrefsyms
326 Write32(IndirectSymbolOffset);
327 Write32(NumIndirectSymbols);
328 Write32(0); // extreloff
329 Write32(0); // nextrel
330 Write32(0); // locreloff
331 Write32(0); // nlocrel
333 assert(OS.tell() - Start == macho::DysymtabLoadCommandSize);
336 void WriteNlist(MachSymbolData &MSD, const MCAsmLayout &Layout) {
337 MCSymbolData &Data = *MSD.SymbolData;
338 const MCSymbol &Symbol = Data.getSymbol();
340 uint16_t Flags = Data.getFlags();
341 uint32_t Address = 0;
343 // Set the N_TYPE bits. See <mach-o/nlist.h>.
345 // FIXME: Are the prebound or indirect fields possible here?
346 if (Symbol.isUndefined())
347 Type = macho::STT_Undefined;
348 else if (Symbol.isAbsolute())
349 Type = macho::STT_Absolute;
351 Type = macho::STT_Section;
353 // FIXME: Set STAB bits.
355 if (Data.isPrivateExtern())
356 Type |= macho::STF_PrivateExtern;
359 if (Data.isExternal() || Symbol.isUndefined())
360 Type |= macho::STF_External;
362 // Compute the symbol address.
363 if (Symbol.isDefined()) {
364 if (Symbol.isAbsolute()) {
365 Address = cast<MCConstantExpr>(Symbol.getVariableValue())->getValue();
367 Address = getSymbolAddress(&Data, Layout);
369 } else if (Data.isCommon()) {
370 // Common symbols are encoded with the size in the address
371 // field, and their alignment in the flags.
372 Address = Data.getCommonSize();
374 // Common alignment is packed into the 'desc' bits.
375 if (unsigned Align = Data.getCommonAlignment()) {
376 unsigned Log2Size = Log2_32(Align);
377 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
379 report_fatal_error("invalid 'common' alignment '" +
381 // FIXME: Keep this mask with the SymbolFlags enumeration.
382 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
386 // struct nlist (12 bytes)
388 Write32(MSD.StringIndex);
390 Write8(MSD.SectionIndex);
392 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
401 // FIXME: We really need to improve the relocation validation. Basically, we
402 // want to implement a separate computation which evaluates the relocation
403 // entry as the linker would, and verifies that the resultant fixup value is
404 // exactly what the encoder wanted. This will catch several classes of
407 // - Relocation entry bugs, the two algorithms are unlikely to have the same
410 // - Relaxation issues, where we forget to relax something.
412 // - Input errors, where something cannot be correctly encoded. 'as' allows
413 // these through in many cases.
415 static bool isFixupKindRIPRel(unsigned Kind) {
416 return Kind == X86::reloc_riprel_4byte ||
417 Kind == X86::reloc_riprel_4byte_movq_load;
419 void RecordX86_64Relocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
420 const MCFragment *Fragment,
421 const MCFixup &Fixup, MCValue Target,
422 uint64_t &FixedValue) {
423 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
424 unsigned IsRIPRel = isFixupKindRIPRel(Fixup.getKind());
425 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
428 uint32_t FixupOffset =
429 Layout.getFragmentOffset(Fragment) + Fixup.getOffset();
430 uint32_t FixupAddress =
431 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
434 unsigned IsExtern = 0;
437 Value = Target.getConstant();
440 // Compensate for the relocation offset, Darwin x86_64 relocations only
441 // have the addend and appear to have attempted to define it to be the
442 // actual expression addend without the PCrel bias. However, instructions
443 // with data following the relocation are not accomodated for (see comment
444 // below regarding SIGNED{1,2,4}), so it isn't exactly that either.
445 Value += 1LL << Log2Size;
448 if (Target.isAbsolute()) { // constant
449 // SymbolNum of 0 indicates the absolute section.
450 Type = macho::RIT_X86_64_Unsigned;
453 // FIXME: I believe this is broken, I don't think the linker can
454 // understand it. I think it would require a local relocation, but I'm not
455 // sure if that would work either. The official way to get an absolute
456 // PCrel relocation is to use an absolute symbol (which we don't support
460 Type = macho::RIT_X86_64_Branch;
462 } else if (Target.getSymB()) { // A - B + constant
463 const MCSymbol *A = &Target.getSymA()->getSymbol();
464 MCSymbolData &A_SD = Asm.getSymbolData(*A);
465 const MCSymbolData *A_Base = Asm.getAtom(&A_SD);
467 const MCSymbol *B = &Target.getSymB()->getSymbol();
468 MCSymbolData &B_SD = Asm.getSymbolData(*B);
469 const MCSymbolData *B_Base = Asm.getAtom(&B_SD);
471 // Neither symbol can be modified.
472 if (Target.getSymA()->getKind() != MCSymbolRefExpr::VK_None ||
473 Target.getSymB()->getKind() != MCSymbolRefExpr::VK_None)
474 report_fatal_error("unsupported relocation of modified symbol");
476 // We don't support PCrel relocations of differences. Darwin 'as' doesn't
477 // implement most of these correctly.
479 report_fatal_error("unsupported pc-relative relocation of difference");
481 // The support for the situation where one or both of the symbols would
482 // require a local relocation is handled just like if the symbols were
483 // external. This is certainly used in the case of debug sections where
484 // the section has only temporary symbols and thus the symbols don't have
485 // base symbols. This is encoded using the section ordinal and
486 // non-extern relocation entries.
488 // Darwin 'as' doesn't emit correct relocations for this (it ends up with
489 // a single SIGNED relocation); reject it for now. Except the case where
490 // both symbols don't have a base, equal but both NULL.
491 if (A_Base == B_Base && A_Base)
492 report_fatal_error("unsupported relocation with identical base");
494 Value += getSymbolAddress(&A_SD, Layout) -
495 (A_Base == NULL ? 0 : getSymbolAddress(A_Base, Layout));
496 Value -= getSymbolAddress(&B_SD, Layout) -
497 (B_Base == NULL ? 0 : getSymbolAddress(B_Base, Layout));
500 Index = A_Base->getIndex();
504 Index = A_SD.getFragment()->getParent()->getOrdinal() + 1;
507 Type = macho::RIT_X86_64_Unsigned;
509 macho::RelocationEntry MRE;
510 MRE.Word0 = FixupOffset;
511 MRE.Word1 = ((Index << 0) |
516 Relocations[Fragment->getParent()].push_back(MRE);
519 Index = B_Base->getIndex();
523 Index = B_SD.getFragment()->getParent()->getOrdinal() + 1;
526 Type = macho::RIT_X86_64_Subtractor;
528 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
529 MCSymbolData &SD = Asm.getSymbolData(*Symbol);
530 const MCSymbolData *Base = Asm.getAtom(&SD);
532 // Relocations inside debug sections always use local relocations when
533 // possible. This seems to be done because the debugger doesn't fully
534 // understand x86_64 relocation entries, and expects to find values that
535 // have already been fixed up.
536 if (Symbol->isInSection()) {
537 const MCSectionMachO &Section = static_cast<const MCSectionMachO&>(
538 Fragment->getParent()->getSection());
539 if (Section.hasAttribute(MCSectionMachO::S_ATTR_DEBUG))
543 // x86_64 almost always uses external relocations, except when there is no
544 // symbol to use as a base address (a local symbol with no preceeding
545 // non-local symbol).
547 Index = Base->getIndex();
550 // Add the local offset, if needed.
552 Value += Layout.getSymbolOffset(&SD) - Layout.getSymbolOffset(Base);
553 } else if (Symbol->isInSection()) {
554 // The index is the section ordinal (1-based).
555 Index = SD.getFragment()->getParent()->getOrdinal() + 1;
557 Value += getSymbolAddress(&SD, Layout);
560 Value -= FixupAddress + (1 << Log2Size);
561 } else if (Symbol->isVariable()) {
562 const MCExpr *Value = Symbol->getVariableValue();
564 bool isAbs = Value->EvaluateAsAbsolute(Res, Layout, SectionAddress);
569 report_fatal_error("unsupported relocation of variable '" +
570 Symbol->getName() + "'");
573 report_fatal_error("unsupported relocation of undefined symbol '" +
574 Symbol->getName() + "'");
577 MCSymbolRefExpr::VariantKind Modifier = Target.getSymA()->getKind();
580 if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
581 // x86_64 distinguishes movq foo@GOTPCREL so that the linker can
582 // rewrite the movq to an leaq at link time if the symbol ends up in
583 // the same linkage unit.
584 if (unsigned(Fixup.getKind()) == X86::reloc_riprel_4byte_movq_load)
585 Type = macho::RIT_X86_64_GOTLoad;
587 Type = macho::RIT_X86_64_GOT;
588 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
589 Type = macho::RIT_X86_64_TLV;
590 } else if (Modifier != MCSymbolRefExpr::VK_None) {
591 report_fatal_error("unsupported symbol modifier in relocation");
593 Type = macho::RIT_X86_64_Signed;
595 // The Darwin x86_64 relocation format has a problem where it cannot
596 // encode an address (L<foo> + <constant>) which is outside the atom
597 // containing L<foo>. Generally, this shouldn't occur but it does
598 // happen when we have a RIPrel instruction with data following the
599 // relocation entry (e.g., movb $012, L0(%rip)). Even with the PCrel
600 // adjustment Darwin x86_64 uses, the offset is still negative and
601 // the linker has no way to recognize this.
603 // To work around this, Darwin uses several special relocation types
604 // to indicate the offsets. However, the specification or
605 // implementation of these seems to also be incomplete; they should
606 // adjust the addend as well based on the actual encoded instruction
607 // (the additional bias), but instead appear to just look at the
609 switch (-(Target.getConstant() + (1LL << Log2Size))) {
610 case 1: Type = macho::RIT_X86_64_Signed1; break;
611 case 2: Type = macho::RIT_X86_64_Signed2; break;
612 case 4: Type = macho::RIT_X86_64_Signed4; break;
616 if (Modifier != MCSymbolRefExpr::VK_None)
617 report_fatal_error("unsupported symbol modifier in branch "
620 Type = macho::RIT_X86_64_Branch;
623 if (Modifier == MCSymbolRefExpr::VK_GOT) {
624 Type = macho::RIT_X86_64_GOT;
625 } else if (Modifier == MCSymbolRefExpr::VK_GOTPCREL) {
626 // GOTPCREL is allowed as a modifier on non-PCrel instructions, in
627 // which case all we do is set the PCrel bit in the relocation entry;
628 // this is used with exception handling, for example. The source is
629 // required to include any necessary offset directly.
630 Type = macho::RIT_X86_64_GOT;
632 } else if (Modifier == MCSymbolRefExpr::VK_TLVP) {
633 report_fatal_error("TLVP symbol modifier should have been rip-rel");
634 } else if (Modifier != MCSymbolRefExpr::VK_None)
635 report_fatal_error("unsupported symbol modifier in relocation");
637 Type = macho::RIT_X86_64_Unsigned;
641 // x86_64 always writes custom values into the fixups.
644 // struct relocation_info (8 bytes)
645 macho::RelocationEntry MRE;
646 MRE.Word0 = FixupOffset;
647 MRE.Word1 = ((Index << 0) |
652 Relocations[Fragment->getParent()].push_back(MRE);
655 void RecordScatteredRelocation(const MCAssembler &Asm,
656 const MCAsmLayout &Layout,
657 const MCFragment *Fragment,
658 const MCFixup &Fixup, MCValue Target,
660 uint64_t &FixedValue) {
661 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
662 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
663 unsigned Type = macho::RIT_Vanilla;
666 const MCSymbol *A = &Target.getSymA()->getSymbol();
667 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
669 if (!A_SD->getFragment())
670 report_fatal_error("symbol '" + A->getName() +
671 "' can not be undefined in a subtraction expression");
673 uint32_t Value = getSymbolAddress(A_SD, Layout);
674 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
675 FixedValue += SecAddr;
678 if (const MCSymbolRefExpr *B = Target.getSymB()) {
679 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
681 if (!B_SD->getFragment())
682 report_fatal_error("symbol '" + B->getSymbol().getName() +
683 "' can not be undefined in a subtraction expression");
685 // Select the appropriate difference relocation type.
687 // Note that there is no longer any semantic difference between these two
688 // relocation types from the linkers point of view, this is done solely
689 // for pedantic compatibility with 'as'.
690 Type = A_SD->isExternal() ? (unsigned)macho::RIT_Difference :
691 (unsigned)macho::RIT_Generic_LocalDifference;
692 Value2 = getSymbolAddress(B_SD, Layout);
693 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
696 // Relocations are written out in reverse order, so the PAIR comes first.
697 if (Type == macho::RIT_Difference ||
698 Type == macho::RIT_Generic_LocalDifference) {
699 macho::RelocationEntry MRE;
700 MRE.Word0 = ((0 << 0) |
701 (macho::RIT_Pair << 24) |
704 macho::RF_Scattered);
706 Relocations[Fragment->getParent()].push_back(MRE);
709 macho::RelocationEntry MRE;
710 MRE.Word0 = ((FixupOffset << 0) |
714 macho::RF_Scattered);
716 Relocations[Fragment->getParent()].push_back(MRE);
719 void RecordARMScatteredRelocation(const MCAssembler &Asm,
720 const MCAsmLayout &Layout,
721 const MCFragment *Fragment,
722 const MCFixup &Fixup, MCValue Target,
724 uint64_t &FixedValue) {
725 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
726 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
727 unsigned Type = macho::RIT_Vanilla;
730 const MCSymbol *A = &Target.getSymA()->getSymbol();
731 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
733 if (!A_SD->getFragment())
734 report_fatal_error("symbol '" + A->getName() +
735 "' can not be undefined in a subtraction expression");
737 uint32_t Value = getSymbolAddress(A_SD, Layout);
738 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
739 FixedValue += SecAddr;
742 if (const MCSymbolRefExpr *B = Target.getSymB()) {
743 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
745 if (!B_SD->getFragment())
746 report_fatal_error("symbol '" + B->getSymbol().getName() +
747 "' can not be undefined in a subtraction expression");
749 // Select the appropriate difference relocation type.
750 Type = macho::RIT_Difference;
751 Value2 = getSymbolAddress(B_SD, Layout);
752 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
755 // Relocations are written out in reverse order, so the PAIR comes first.
756 if (Type == macho::RIT_Difference ||
757 Type == macho::RIT_Generic_LocalDifference) {
758 macho::RelocationEntry MRE;
759 MRE.Word0 = ((0 << 0) |
760 (macho::RIT_Pair << 24) |
763 macho::RF_Scattered);
765 Relocations[Fragment->getParent()].push_back(MRE);
768 macho::RelocationEntry MRE;
769 MRE.Word0 = ((FixupOffset << 0) |
773 macho::RF_Scattered);
775 Relocations[Fragment->getParent()].push_back(MRE);
778 void RecordARMMovwMovtRelocation(const MCAssembler &Asm,
779 const MCAsmLayout &Layout,
780 const MCFragment *Fragment,
781 const MCFixup &Fixup, MCValue Target,
782 uint64_t &FixedValue) {
783 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
784 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
785 unsigned Type = macho::RIT_ARM_Half;
788 const MCSymbol *A = &Target.getSymA()->getSymbol();
789 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
791 if (!A_SD->getFragment())
792 report_fatal_error("symbol '" + A->getName() +
793 "' can not be undefined in a subtraction expression");
795 uint32_t Value = getSymbolAddress(A_SD, Layout);
797 uint64_t SecAddr = getSectionAddress(A_SD->getFragment()->getParent());
798 FixedValue += SecAddr;
800 if (const MCSymbolRefExpr *B = Target.getSymB()) {
801 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
803 if (!B_SD->getFragment())
804 report_fatal_error("symbol '" + B->getSymbol().getName() +
805 "' can not be undefined in a subtraction expression");
807 // Select the appropriate difference relocation type.
808 Type = macho::RIT_ARM_HalfDifference;
809 Value2 = getSymbolAddress(B_SD, Layout);
810 FixedValue -= getSectionAddress(B_SD->getFragment()->getParent());
813 // Relocations are written out in reverse order, so the PAIR comes first.
814 // ARM_RELOC_HALF and ARM_RELOC_HALF_SECTDIFF abuse the r_length field:
816 // For these two r_type relocations they always have a pair following them
817 // and the r_length bits are used differently. The encoding of the
818 // r_length is as follows:
819 // low bit of r_length:
820 // 0 - :lower16: for movw instructions
821 // 1 - :upper16: for movt instructions
822 // high bit of r_length:
823 // 0 - arm instructions
824 // 1 - thumb instructions
825 // the other half of the relocated expression is in the following pair
826 // relocation entry in the the low 16 bits of r_address field.
827 unsigned ThumbBit = 0;
828 unsigned MovtBit = 0;
829 switch (Fixup.getKind()) {
831 case ARM::fixup_arm_movt_hi16:
832 case ARM::fixup_arm_movt_hi16_pcrel:
835 case ARM::fixup_t2_movt_hi16:
836 case ARM::fixup_t2_movt_hi16_pcrel:
839 case ARM::fixup_t2_movw_lo16:
840 case ARM::fixup_t2_movw_lo16_pcrel:
846 if (Type == macho::RIT_ARM_HalfDifference) {
847 uint32_t OtherHalf = MovtBit
848 ? (FixedValue & 0xffff) : ((FixedValue & 0xffff0000) >> 16);
850 macho::RelocationEntry MRE;
851 MRE.Word0 = ((OtherHalf << 0) |
852 (macho::RIT_Pair << 24) |
856 macho::RF_Scattered);
858 Relocations[Fragment->getParent()].push_back(MRE);
861 macho::RelocationEntry MRE;
862 MRE.Word0 = ((FixupOffset << 0) |
867 macho::RF_Scattered);
869 Relocations[Fragment->getParent()].push_back(MRE);
872 void RecordTLVPRelocation(const MCAssembler &Asm,
873 const MCAsmLayout &Layout,
874 const MCFragment *Fragment,
875 const MCFixup &Fixup, MCValue Target,
876 uint64_t &FixedValue) {
877 assert(Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP &&
879 "Should only be called with a 32-bit TLVP relocation!");
881 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
882 uint32_t Value = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
883 unsigned IsPCRel = 0;
885 // Get the symbol data.
886 MCSymbolData *SD_A = &Asm.getSymbolData(Target.getSymA()->getSymbol());
887 unsigned Index = SD_A->getIndex();
889 // We're only going to have a second symbol in pic mode and it'll be a
890 // subtraction from the picbase. For 32-bit pic the addend is the difference
891 // between the picbase and the next address. For 32-bit static the addend
893 if (Target.getSymB()) {
894 // If this is a subtraction then we're pcrel.
895 uint32_t FixupAddress =
896 getFragmentAddress(Fragment, Layout) + Fixup.getOffset();
897 MCSymbolData *SD_B = &Asm.getSymbolData(Target.getSymB()->getSymbol());
899 FixedValue = (FixupAddress - getSymbolAddress(SD_B, Layout) +
900 Target.getConstant());
901 FixedValue += 1ULL << Log2Size;
906 // struct relocation_info (8 bytes)
907 macho::RelocationEntry MRE;
909 MRE.Word1 = ((Index << 0) |
912 (1 << 27) | // Extern
913 (macho::RIT_Generic_TLV << 28)); // Type
914 Relocations[Fragment->getParent()].push_back(MRE);
917 static bool getARMFixupKindMachOInfo(unsigned Kind, unsigned &RelocType,
918 unsigned &Log2Size) {
919 RelocType = unsigned(macho::RIT_Vanilla);
927 Log2Size = llvm::Log2_32(1);
930 Log2Size = llvm::Log2_32(2);
933 Log2Size = llvm::Log2_32(4);
936 Log2Size = llvm::Log2_32(8);
939 // Handle 24-bit branch kinds.
940 case ARM::fixup_arm_ldst_pcrel_12:
941 case ARM::fixup_arm_pcrel_10:
942 case ARM::fixup_arm_adr_pcrel_12:
943 case ARM::fixup_arm_branch:
944 RelocType = unsigned(macho::RIT_ARM_Branch24Bit);
945 // Report as 'long', even though that is not quite accurate.
946 Log2Size = llvm::Log2_32(4);
949 // Handle Thumb branches.
950 case ARM::fixup_arm_thumb_br:
951 RelocType = unsigned(macho::RIT_ARM_ThumbBranch22Bit);
952 Log2Size = llvm::Log2_32(2);
955 case ARM::fixup_arm_thumb_bl:
956 RelocType = unsigned(macho::RIT_ARM_ThumbBranch32Bit);
957 Log2Size = llvm::Log2_32(4);
960 case ARM::fixup_arm_thumb_blx:
961 RelocType = unsigned(macho::RIT_ARM_ThumbBranch22Bit);
962 // Report as 'long', even though that is not quite accurate.
963 Log2Size = llvm::Log2_32(4);
966 case ARM::fixup_arm_movt_hi16:
967 case ARM::fixup_arm_movt_hi16_pcrel:
968 case ARM::fixup_t2_movt_hi16:
969 case ARM::fixup_t2_movt_hi16_pcrel:
970 RelocType = unsigned(macho::RIT_ARM_HalfDifference);
971 // Report as 'long', even though that is not quite accurate.
972 Log2Size = llvm::Log2_32(4);
975 case ARM::fixup_arm_movw_lo16:
976 case ARM::fixup_arm_movw_lo16_pcrel:
977 case ARM::fixup_t2_movw_lo16:
978 case ARM::fixup_t2_movw_lo16_pcrel:
979 RelocType = unsigned(macho::RIT_ARM_Half);
980 // Report as 'long', even though that is not quite accurate.
981 Log2Size = llvm::Log2_32(4);
985 void RecordARMRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
986 const MCFragment *Fragment, const MCFixup &Fixup,
987 MCValue Target, uint64_t &FixedValue) {
988 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
990 unsigned RelocType = macho::RIT_Vanilla;
991 if (!getARMFixupKindMachOInfo(Fixup.getKind(), RelocType, Log2Size)) {
992 report_fatal_error("unknown ARM fixup kind!");
996 // If this is a difference or a defined symbol plus an offset, then we need
997 // a scattered relocation entry. Differences always require scattered
999 if (Target.getSymB()) {
1000 if (RelocType == macho::RIT_ARM_Half ||
1001 RelocType == macho::RIT_ARM_HalfDifference)
1002 return RecordARMMovwMovtRelocation(Asm, Layout, Fragment, Fixup,
1003 Target, FixedValue);
1004 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup,
1005 Target, Log2Size, FixedValue);
1008 // Get the symbol data, if any.
1009 MCSymbolData *SD = 0;
1010 if (Target.getSymA())
1011 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1013 // FIXME: For other platforms, we need to use scattered relocations for
1014 // internal relocations with offsets. If this is an internal relocation
1015 // with an offset, it also needs a scattered relocation entry.
1017 // Is this right for ARM?
1018 uint32_t Offset = Target.getConstant();
1019 if (IsPCRel && RelocType == macho::RIT_Vanilla)
1020 Offset += 1 << Log2Size;
1021 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
1022 return RecordARMScatteredRelocation(Asm, Layout, Fragment, Fixup, Target,
1023 Log2Size, FixedValue);
1026 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1028 unsigned IsExtern = 0;
1031 if (Target.isAbsolute()) { // constant
1033 report_fatal_error("FIXME: relocations to absolute targets "
1034 "not yet implemented");
1035 } else if (SD->getSymbol().isVariable()) {
1037 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
1038 Res, Layout, SectionAddress)) {
1043 report_fatal_error("unsupported relocation of variable '" +
1044 SD->getSymbol().getName() + "'");
1046 // Check whether we need an external or internal relocation.
1047 if (doesSymbolRequireExternRelocation(SD)) {
1049 Index = SD->getIndex();
1050 // For external relocations, make sure to offset the fixup value to
1051 // compensate for the addend of the symbol address, if it was
1052 // undefined. This occurs with weak definitions, for example.
1053 if (!SD->Symbol->isUndefined())
1054 FixedValue -= Layout.getSymbolOffset(SD);
1056 // The index is the section ordinal (1-based).
1057 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
1058 FixedValue += getSectionAddress(SD->getFragment()->getParent());
1061 FixedValue -= getSectionAddress(Fragment->getParent());
1063 // The type is determined by the fixup kind.
1067 // struct relocation_info (8 bytes)
1068 macho::RelocationEntry MRE;
1069 MRE.Word0 = FixupOffset;
1070 MRE.Word1 = ((Index << 0) |
1075 Relocations[Fragment->getParent()].push_back(MRE);
1078 void RecordRelocation(const MCAssembler &Asm, const MCAsmLayout &Layout,
1079 const MCFragment *Fragment, const MCFixup &Fixup,
1080 MCValue Target, uint64_t &FixedValue) {
1081 // FIXME: These needs to be factored into the target Mach-O writer.
1083 RecordARMRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1087 RecordX86_64Relocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1091 unsigned IsPCRel = isFixupKindPCRel(Asm, Fixup.getKind());
1092 unsigned Log2Size = getFixupKindLog2Size(Fixup.getKind());
1094 // If this is a 32-bit TLVP reloc it's handled a bit differently.
1095 if (Target.getSymA() &&
1096 Target.getSymA()->getKind() == MCSymbolRefExpr::VK_TLVP) {
1097 RecordTLVPRelocation(Asm, Layout, Fragment, Fixup, Target, FixedValue);
1101 // If this is a difference or a defined symbol plus an offset, then we need
1102 // a scattered relocation entry.
1103 // Differences always require scattered relocations.
1104 if (Target.getSymB())
1105 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1106 Target, Log2Size, FixedValue);
1108 // Get the symbol data, if any.
1109 MCSymbolData *SD = 0;
1110 if (Target.getSymA())
1111 SD = &Asm.getSymbolData(Target.getSymA()->getSymbol());
1113 // If this is an internal relocation with an offset, it also needs a
1114 // scattered relocation entry.
1115 uint32_t Offset = Target.getConstant();
1117 Offset += 1 << Log2Size;
1118 if (Offset && SD && !doesSymbolRequireExternRelocation(SD))
1119 return RecordScatteredRelocation(Asm, Layout, Fragment, Fixup,
1120 Target, Log2Size, FixedValue);
1123 uint32_t FixupOffset = Layout.getFragmentOffset(Fragment)+Fixup.getOffset();
1125 unsigned IsExtern = 0;
1128 if (Target.isAbsolute()) { // constant
1129 // SymbolNum of 0 indicates the absolute section.
1131 // FIXME: Currently, these are never generated (see code below). I cannot
1132 // find a case where they are actually emitted.
1133 Type = macho::RIT_Vanilla;
1134 } else if (SD->getSymbol().isVariable()) {
1136 if (SD->getSymbol().getVariableValue()->EvaluateAsAbsolute(
1137 Res, Layout, SectionAddress)) {
1142 report_fatal_error("unsupported relocation of variable '" +
1143 SD->getSymbol().getName() + "'");
1145 // Check whether we need an external or internal relocation.
1146 if (doesSymbolRequireExternRelocation(SD)) {
1148 Index = SD->getIndex();
1149 // For external relocations, make sure to offset the fixup value to
1150 // compensate for the addend of the symbol address, if it was
1151 // undefined. This occurs with weak definitions, for example.
1152 if (!SD->Symbol->isUndefined())
1153 FixedValue -= Layout.getSymbolOffset(SD);
1155 // The index is the section ordinal (1-based).
1156 Index = SD->getFragment()->getParent()->getOrdinal() + 1;
1157 FixedValue += getSectionAddress(SD->getFragment()->getParent());
1160 FixedValue -= getSectionAddress(Fragment->getParent());
1162 Type = macho::RIT_Vanilla;
1165 // struct relocation_info (8 bytes)
1166 macho::RelocationEntry MRE;
1167 MRE.Word0 = FixupOffset;
1168 MRE.Word1 = ((Index << 0) |
1173 Relocations[Fragment->getParent()].push_back(MRE);
1176 void BindIndirectSymbols(MCAssembler &Asm) {
1177 // This is the point where 'as' creates actual symbols for indirect symbols
1178 // (in the following two passes). It would be easier for us to do this
1179 // sooner when we see the attribute, but that makes getting the order in the
1180 // symbol table much more complicated than it is worth.
1182 // FIXME: Revisit this when the dust settles.
1184 // Bind non lazy symbol pointers first.
1185 unsigned IndirectIndex = 0;
1186 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1187 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1188 const MCSectionMachO &Section =
1189 cast<MCSectionMachO>(it->SectionData->getSection());
1191 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
1194 // Initialize the section indirect symbol base, if necessary.
1195 if (!IndirectSymBase.count(it->SectionData))
1196 IndirectSymBase[it->SectionData] = IndirectIndex;
1198 Asm.getOrCreateSymbolData(*it->Symbol);
1201 // Then lazy symbol pointers and symbol stubs.
1203 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
1204 ie = Asm.indirect_symbol_end(); it != ie; ++it, ++IndirectIndex) {
1205 const MCSectionMachO &Section =
1206 cast<MCSectionMachO>(it->SectionData->getSection());
1208 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
1209 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
1212 // Initialize the section indirect symbol base, if necessary.
1213 if (!IndirectSymBase.count(it->SectionData))
1214 IndirectSymBase[it->SectionData] = IndirectIndex;
1216 // Set the symbol type to undefined lazy, but only on construction.
1218 // FIXME: Do not hardcode.
1220 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
1222 Entry.setFlags(Entry.getFlags() | 0x0001);
1226 /// ComputeSymbolTable - Compute the symbol table data
1228 /// \param StringTable [out] - The string table data.
1229 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
1231 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
1232 std::vector<MachSymbolData> &LocalSymbolData,
1233 std::vector<MachSymbolData> &ExternalSymbolData,
1234 std::vector<MachSymbolData> &UndefinedSymbolData) {
1235 // Build section lookup table.
1236 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
1238 for (MCAssembler::iterator it = Asm.begin(),
1239 ie = Asm.end(); it != ie; ++it, ++Index)
1240 SectionIndexMap[&it->getSection()] = Index;
1241 assert(Index <= 256 && "Too many sections!");
1243 // Index 0 is always the empty string.
1244 StringMap<uint64_t> StringIndexMap;
1245 StringTable += '\x00';
1247 // Build the symbol arrays and the string table, but only for non-local
1250 // The particular order that we collect the symbols and create the string
1251 // table, then sort the symbols is chosen to match 'as'. Even though it
1252 // doesn't matter for correctness, this is important for letting us diff .o
1254 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1255 ie = Asm.symbol_end(); it != ie; ++it) {
1256 const MCSymbol &Symbol = it->getSymbol();
1258 // Ignore non-linker visible symbols.
1259 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1262 if (!it->isExternal() && !Symbol.isUndefined())
1265 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1267 Entry = StringTable.size();
1268 StringTable += Symbol.getName();
1269 StringTable += '\x00';
1273 MSD.SymbolData = it;
1274 MSD.StringIndex = Entry;
1276 if (Symbol.isUndefined()) {
1277 MSD.SectionIndex = 0;
1278 UndefinedSymbolData.push_back(MSD);
1279 } else if (Symbol.isAbsolute()) {
1280 MSD.SectionIndex = 0;
1281 ExternalSymbolData.push_back(MSD);
1283 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1284 assert(MSD.SectionIndex && "Invalid section index!");
1285 ExternalSymbolData.push_back(MSD);
1289 // Now add the data for local symbols.
1290 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
1291 ie = Asm.symbol_end(); it != ie; ++it) {
1292 const MCSymbol &Symbol = it->getSymbol();
1294 // Ignore non-linker visible symbols.
1295 if (!Asm.isSymbolLinkerVisible(it->getSymbol()))
1298 if (it->isExternal() || Symbol.isUndefined())
1301 uint64_t &Entry = StringIndexMap[Symbol.getName()];
1303 Entry = StringTable.size();
1304 StringTable += Symbol.getName();
1305 StringTable += '\x00';
1309 MSD.SymbolData = it;
1310 MSD.StringIndex = Entry;
1312 if (Symbol.isAbsolute()) {
1313 MSD.SectionIndex = 0;
1314 LocalSymbolData.push_back(MSD);
1316 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
1317 assert(MSD.SectionIndex && "Invalid section index!");
1318 LocalSymbolData.push_back(MSD);
1322 // External and undefined symbols are required to be in lexicographic order.
1323 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
1324 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
1326 // Set the symbol indices.
1328 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1329 LocalSymbolData[i].SymbolData->setIndex(Index++);
1330 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1331 ExternalSymbolData[i].SymbolData->setIndex(Index++);
1332 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1333 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
1335 // The string table is padded to a multiple of 4.
1336 while (StringTable.size() % 4)
1337 StringTable += '\x00';
1340 void computeSectionAddresses(const MCAssembler &Asm,
1341 const MCAsmLayout &Layout) {
1342 uint64_t StartAddress = 0;
1343 const SmallVectorImpl<MCSectionData*> &Order = Layout.getSectionOrder();
1344 for (int i = 0, n = Order.size(); i != n ; ++i) {
1345 const MCSectionData *SD = Order[i];
1346 StartAddress = RoundUpToAlignment(StartAddress, SD->getAlignment());
1347 SectionAddress[SD] = StartAddress;
1348 StartAddress += Layout.getSectionAddressSize(SD);
1349 // Explicitly pad the section to match the alignment requirements of the
1350 // following one. This is for 'gas' compatibility, it shouldn't
1351 /// strictly be necessary.
1352 StartAddress += getPaddingSize(SD, Layout);
1356 void ExecutePostLayoutBinding(MCAssembler &Asm, const MCAsmLayout &Layout) {
1357 computeSectionAddresses(Asm, Layout);
1359 // Create symbol data for any indirect symbols.
1360 BindIndirectSymbols(Asm);
1362 // Compute symbol table information and bind symbol indices.
1363 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
1364 UndefinedSymbolData);
1367 virtual bool IsSymbolRefDifferenceFullyResolvedImpl(const MCAssembler &Asm,
1368 const MCSymbolData &DataA,
1369 const MCFragment &FB,
1371 bool IsPCRel) const {
1375 // The effective address is
1376 // addr(atom(A)) + offset(A)
1377 // - addr(atom(B)) - offset(B)
1378 // and the offsets are not relocatable, so the fixup is fully resolved when
1379 // addr(atom(A)) - addr(atom(B)) == 0.
1380 const MCSymbolData *A_Base = 0, *B_Base = 0;
1382 const MCSymbol &SA = DataA.getSymbol().AliasedSymbol();
1383 const MCSection &SecA = SA.getSection();
1384 const MCSection &SecB = FB.getParent()->getSection();
1387 // The simple (Darwin, except on x86_64) way of dealing with this was to
1388 // assume that any reference to a temporary symbol *must* be a temporary
1389 // symbol in the same atom, unless the sections differ. Therefore, any
1390 // PCrel relocation to a temporary symbol (in the same section) is fully
1391 // resolved. This also works in conjunction with absolutized .set, which
1392 // requires the compiler to use .set to absolutize the differences between
1393 // symbols which the compiler knows to be assembly time constants, so we
1394 // don't need to worry about considering symbol differences fully
1397 if (!Asm.getBackend().hasReliableSymbolDifference()) {
1398 if (!SA.isTemporary() || !SA.isInSection() || &SecA != &SecB)
1403 if (!TargetObjectWriter->useAggressiveSymbolFolding())
1407 const MCFragment &FA = *Asm.getSymbolData(SA).getFragment();
1409 A_Base = FA.getAtom();
1413 B_Base = FB.getAtom();
1417 // If the atoms are the same, they are guaranteed to have the same address.
1418 if (A_Base == B_Base)
1421 // Otherwise, we can't prove this is fully resolved.
1425 void WriteObject(MCAssembler &Asm, const MCAsmLayout &Layout) {
1426 unsigned NumSections = Asm.size();
1428 // The section data starts after the header, the segment load command (and
1429 // section headers) and the symbol table.
1430 unsigned NumLoadCommands = 1;
1431 uint64_t LoadCommandsSize = is64Bit() ?
1432 macho::SegmentLoadCommand64Size + NumSections * macho::Section64Size :
1433 macho::SegmentLoadCommand32Size + NumSections * macho::Section32Size;
1435 // Add the symbol table load command sizes, if used.
1436 unsigned NumSymbols = LocalSymbolData.size() + ExternalSymbolData.size() +
1437 UndefinedSymbolData.size();
1439 NumLoadCommands += 2;
1440 LoadCommandsSize += (macho::SymtabLoadCommandSize +
1441 macho::DysymtabLoadCommandSize);
1444 // Compute the total size of the section data, as well as its file size and
1446 uint64_t SectionDataStart = (is64Bit() ? macho::Header64Size :
1447 macho::Header32Size) + LoadCommandsSize;
1448 uint64_t SectionDataSize = 0;
1449 uint64_t SectionDataFileSize = 0;
1450 uint64_t VMSize = 0;
1451 for (MCAssembler::const_iterator it = Asm.begin(),
1452 ie = Asm.end(); it != ie; ++it) {
1453 const MCSectionData &SD = *it;
1454 uint64_t Address = getSectionAddress(&SD);
1455 uint64_t Size = Layout.getSectionAddressSize(&SD);
1456 uint64_t FileSize = Layout.getSectionFileSize(&SD);
1457 FileSize += getPaddingSize(&SD, Layout);
1459 VMSize = std::max(VMSize, Address + Size);
1461 if (SD.getSection().isVirtualSection())
1464 SectionDataSize = std::max(SectionDataSize, Address + Size);
1465 SectionDataFileSize = std::max(SectionDataFileSize, Address + FileSize);
1468 // The section data is padded to 4 bytes.
1470 // FIXME: Is this machine dependent?
1471 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
1472 SectionDataFileSize += SectionDataPadding;
1474 // Write the prolog, starting with the header and load command...
1475 WriteHeader(NumLoadCommands, LoadCommandsSize,
1476 Asm.getSubsectionsViaSymbols());
1477 WriteSegmentLoadCommand(NumSections, VMSize,
1478 SectionDataStart, SectionDataSize);
1480 // ... and then the section headers.
1481 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
1482 for (MCAssembler::const_iterator it = Asm.begin(),
1483 ie = Asm.end(); it != ie; ++it) {
1484 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1485 unsigned NumRelocs = Relocs.size();
1486 uint64_t SectionStart = SectionDataStart + getSectionAddress(it);
1487 WriteSection(Asm, Layout, *it, SectionStart, RelocTableEnd, NumRelocs);
1488 RelocTableEnd += NumRelocs * macho::RelocationInfoSize;
1491 // Write the symbol table load command, if used.
1493 unsigned FirstLocalSymbol = 0;
1494 unsigned NumLocalSymbols = LocalSymbolData.size();
1495 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
1496 unsigned NumExternalSymbols = ExternalSymbolData.size();
1497 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
1498 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
1499 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
1500 unsigned NumSymTabSymbols =
1501 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
1502 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
1503 uint64_t IndirectSymbolOffset = 0;
1505 // If used, the indirect symbols are written after the section data.
1506 if (NumIndirectSymbols)
1507 IndirectSymbolOffset = RelocTableEnd;
1509 // The symbol table is written after the indirect symbol data.
1510 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
1512 // The string table is written after symbol table.
1513 uint64_t StringTableOffset =
1514 SymbolTableOffset + NumSymTabSymbols * (is64Bit() ? macho::Nlist64Size :
1515 macho::Nlist32Size);
1516 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
1517 StringTableOffset, StringTable.size());
1519 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
1520 FirstExternalSymbol, NumExternalSymbols,
1521 FirstUndefinedSymbol, NumUndefinedSymbols,
1522 IndirectSymbolOffset, NumIndirectSymbols);
1525 // Write the actual section data.
1526 for (MCAssembler::const_iterator it = Asm.begin(),
1527 ie = Asm.end(); it != ie; ++it) {
1528 Asm.WriteSectionData(it, Layout);
1530 uint64_t Pad = getPaddingSize(it, Layout);
1531 for (unsigned int i = 0; i < Pad; ++i)
1535 // Write the extra padding.
1536 WriteZeros(SectionDataPadding);
1538 // Write the relocation entries.
1539 for (MCAssembler::const_iterator it = Asm.begin(),
1540 ie = Asm.end(); it != ie; ++it) {
1541 // Write the section relocation entries, in reverse order to match 'as'
1542 // (approximately, the exact algorithm is more complicated than this).
1543 std::vector<macho::RelocationEntry> &Relocs = Relocations[it];
1544 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
1545 Write32(Relocs[e - i - 1].Word0);
1546 Write32(Relocs[e - i - 1].Word1);
1550 // Write the symbol table data, if used.
1552 // Write the indirect symbol entries.
1553 for (MCAssembler::const_indirect_symbol_iterator
1554 it = Asm.indirect_symbol_begin(),
1555 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
1556 // Indirect symbols in the non lazy symbol pointer section have some
1557 // special handling.
1558 const MCSectionMachO &Section =
1559 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
1560 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
1561 // If this symbol is defined and internal, mark it as such.
1562 if (it->Symbol->isDefined() &&
1563 !Asm.getSymbolData(*it->Symbol).isExternal()) {
1564 uint32_t Flags = macho::ISF_Local;
1565 if (it->Symbol->isAbsolute())
1566 Flags |= macho::ISF_Absolute;
1572 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
1575 // FIXME: Check that offsets match computed ones.
1577 // Write the symbol table entries.
1578 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
1579 WriteNlist(LocalSymbolData[i], Layout);
1580 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
1581 WriteNlist(ExternalSymbolData[i], Layout);
1582 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
1583 WriteNlist(UndefinedSymbolData[i], Layout);
1585 // Write the string table.
1586 OS << StringTable.str();
1593 MCObjectWriter *llvm::createMachObjectWriter(MCMachObjectTargetWriter *MOTW,
1595 bool IsLittleEndian) {
1596 return new MachObjectWriter(MOTW, OS, IsLittleEndian);