1 //===- lib/MC/MCAssembler.cpp - Assembler Backend Implementation ----------===//
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 #define DEBUG_TYPE "assembler"
11 #include "llvm/MC/MCAssembler.h"
12 #include "llvm/MC/MCExpr.h"
13 #include "llvm/MC/MCSectionMachO.h"
14 #include "llvm/MC/MCSymbol.h"
15 #include "llvm/MC/MCValue.h"
16 #include "llvm/ADT/DenseMap.h"
17 #include "llvm/ADT/SmallString.h"
18 #include "llvm/ADT/Statistic.h"
19 #include "llvm/ADT/StringExtras.h"
20 #include "llvm/ADT/StringMap.h"
21 #include "llvm/ADT/Twine.h"
22 #include "llvm/Support/ErrorHandling.h"
23 #include "llvm/Support/MachO.h"
24 #include "llvm/Support/raw_ostream.h"
25 #include "llvm/Support/Debug.h"
28 #include "../Target/X86/X86FixupKinds.h"
33 class MachObjectWriter;
35 STATISTIC(EmittedFragments, "Number of emitted assembler fragments");
37 // FIXME FIXME FIXME: There are number of places in this file where we convert
38 // what is a 64-bit assembler value used for computation into a value in the
39 // object file, which may truncate it. We should detect that truncation where
40 // invalid and report errors back.
42 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
43 MachObjectWriter &MOW);
45 static uint64_t WriteNopData(uint64_t Count, MachObjectWriter &MOW);
47 /// isVirtualSection - Check if this is a section which does not actually exist
48 /// in the object file.
49 static bool isVirtualSection(const MCSection &Section) {
51 const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section);
52 unsigned Type = SMO.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
53 return (Type == MCSectionMachO::S_ZEROFILL);
56 static unsigned getFixupKindLog2Size(unsigned Kind) {
58 default: llvm_unreachable("invalid fixup kind!");
59 case X86::reloc_pcrel_1byte:
60 case FK_Data_1: return 0;
61 case FK_Data_2: return 1;
62 case X86::reloc_pcrel_4byte:
63 case X86::reloc_riprel_4byte:
64 case FK_Data_4: return 2;
65 case FK_Data_8: return 3;
69 static bool isFixupKindPCRel(unsigned Kind) {
73 case X86::reloc_pcrel_1byte:
74 case X86::reloc_pcrel_4byte:
75 case X86::reloc_riprel_4byte:
80 class MachObjectWriter {
81 // See <mach-o/loader.h>.
83 Header_Magic32 = 0xFEEDFACE,
84 Header_Magic64 = 0xFEEDFACF
87 static const unsigned Header32Size = 28;
88 static const unsigned Header64Size = 32;
89 static const unsigned SegmentLoadCommand32Size = 56;
90 static const unsigned Section32Size = 68;
91 static const unsigned SymtabLoadCommandSize = 24;
92 static const unsigned DysymtabLoadCommandSize = 80;
93 static const unsigned Nlist32Size = 12;
94 static const unsigned RelocationInfoSize = 8;
101 HF_SubsectionsViaSymbols = 0x2000
104 enum LoadCommandType {
110 // See <mach-o/nlist.h>.
111 enum SymbolTypeType {
112 STT_Undefined = 0x00,
117 enum SymbolTypeFlags {
118 // If any of these bits are set, then the entry is a stab entry number (see
119 // <mach-o/stab.h>. Otherwise the other masks apply.
120 STF_StabsEntryMask = 0xe0,
124 STF_PrivateExtern = 0x10
127 /// IndirectSymbolFlags - Flags for encoding special values in the indirect
129 enum IndirectSymbolFlags {
130 ISF_Local = 0x80000000,
131 ISF_Absolute = 0x40000000
134 /// RelocationFlags - Special flags for addresses.
135 enum RelocationFlags {
136 RF_Scattered = 0x80000000
139 enum RelocationInfoType {
143 RIT_PreboundLazyPointer = 3,
144 RIT_LocalDifference = 4
147 /// MachSymbolData - Helper struct for containing some precomputed information
149 struct MachSymbolData {
150 MCSymbolData *SymbolData;
151 uint64_t StringIndex;
152 uint8_t SectionIndex;
154 // Support lexicographic sorting.
155 bool operator<(const MachSymbolData &RHS) const {
156 const std::string &Name = SymbolData->getSymbol().getName();
157 return Name < RHS.SymbolData->getSymbol().getName();
165 MachObjectWriter(raw_ostream &_OS, bool _IsLSB = true)
166 : OS(_OS), IsLSB(_IsLSB) {
169 /// @name Helper Methods
172 void Write8(uint8_t Value) {
176 void Write16(uint16_t Value) {
178 Write8(uint8_t(Value >> 0));
179 Write8(uint8_t(Value >> 8));
181 Write8(uint8_t(Value >> 8));
182 Write8(uint8_t(Value >> 0));
186 void Write32(uint32_t Value) {
188 Write16(uint16_t(Value >> 0));
189 Write16(uint16_t(Value >> 16));
191 Write16(uint16_t(Value >> 16));
192 Write16(uint16_t(Value >> 0));
196 void Write64(uint64_t Value) {
198 Write32(uint32_t(Value >> 0));
199 Write32(uint32_t(Value >> 32));
201 Write32(uint32_t(Value >> 32));
202 Write32(uint32_t(Value >> 0));
206 void WriteZeros(unsigned N) {
207 const char Zeros[16] = { 0 };
209 for (unsigned i = 0, e = N / 16; i != e; ++i)
210 OS << StringRef(Zeros, 16);
212 OS << StringRef(Zeros, N % 16);
215 void WriteString(StringRef Str, unsigned ZeroFillSize = 0) {
218 WriteZeros(ZeroFillSize - Str.size());
223 void WriteHeader32(unsigned NumLoadCommands, unsigned LoadCommandsSize,
224 bool SubsectionsViaSymbols) {
227 if (SubsectionsViaSymbols)
228 Flags |= HF_SubsectionsViaSymbols;
230 // struct mach_header (28 bytes)
232 uint64_t Start = OS.tell();
235 Write32(Header_Magic32);
237 // FIXME: Support cputype.
238 Write32(MachO::CPUTypeI386);
239 // FIXME: Support cpusubtype.
240 Write32(MachO::CPUSubType_I386_ALL);
242 Write32(NumLoadCommands); // Object files have a single load command, the
244 Write32(LoadCommandsSize);
247 assert(OS.tell() - Start == Header32Size);
250 /// WriteSegmentLoadCommand32 - Write a 32-bit segment load command.
252 /// \arg NumSections - The number of sections in this segment.
253 /// \arg SectionDataSize - The total size of the sections.
254 void WriteSegmentLoadCommand32(unsigned NumSections,
256 uint64_t SectionDataStartOffset,
257 uint64_t SectionDataSize) {
258 // struct segment_command (56 bytes)
260 uint64_t Start = OS.tell();
263 Write32(LCT_Segment);
264 Write32(SegmentLoadCommand32Size + NumSections * Section32Size);
267 Write32(0); // vmaddr
268 Write32(VMSize); // vmsize
269 Write32(SectionDataStartOffset); // file offset
270 Write32(SectionDataSize); // file size
271 Write32(0x7); // maxprot
272 Write32(0x7); // initprot
273 Write32(NumSections);
276 assert(OS.tell() - Start == SegmentLoadCommand32Size);
279 void WriteSection32(const MCSectionData &SD, uint64_t FileOffset,
280 uint64_t RelocationsStart, unsigned NumRelocations) {
281 // The offset is unused for virtual sections.
282 if (isVirtualSection(SD.getSection())) {
283 assert(SD.getFileSize() == 0 && "Invalid file size!");
287 // struct section (68 bytes)
289 uint64_t Start = OS.tell();
292 // FIXME: cast<> support!
293 const MCSectionMachO &Section =
294 static_cast<const MCSectionMachO&>(SD.getSection());
295 WriteString(Section.getSectionName(), 16);
296 WriteString(Section.getSegmentName(), 16);
297 Write32(SD.getAddress()); // address
298 Write32(SD.getSize()); // size
301 unsigned Flags = Section.getTypeAndAttributes();
302 if (SD.hasInstructions())
303 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
305 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
306 Write32(Log2_32(SD.getAlignment()));
307 Write32(NumRelocations ? RelocationsStart : 0);
308 Write32(NumRelocations);
310 Write32(0); // reserved1
311 Write32(Section.getStubSize()); // reserved2
313 assert(OS.tell() - Start == Section32Size);
316 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
317 uint32_t StringTableOffset,
318 uint32_t StringTableSize) {
319 // struct symtab_command (24 bytes)
321 uint64_t Start = OS.tell();
325 Write32(SymtabLoadCommandSize);
326 Write32(SymbolOffset);
328 Write32(StringTableOffset);
329 Write32(StringTableSize);
331 assert(OS.tell() - Start == SymtabLoadCommandSize);
334 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
335 uint32_t NumLocalSymbols,
336 uint32_t FirstExternalSymbol,
337 uint32_t NumExternalSymbols,
338 uint32_t FirstUndefinedSymbol,
339 uint32_t NumUndefinedSymbols,
340 uint32_t IndirectSymbolOffset,
341 uint32_t NumIndirectSymbols) {
342 // struct dysymtab_command (80 bytes)
344 uint64_t Start = OS.tell();
347 Write32(LCT_Dysymtab);
348 Write32(DysymtabLoadCommandSize);
349 Write32(FirstLocalSymbol);
350 Write32(NumLocalSymbols);
351 Write32(FirstExternalSymbol);
352 Write32(NumExternalSymbols);
353 Write32(FirstUndefinedSymbol);
354 Write32(NumUndefinedSymbols);
355 Write32(0); // tocoff
357 Write32(0); // modtaboff
358 Write32(0); // nmodtab
359 Write32(0); // extrefsymoff
360 Write32(0); // nextrefsyms
361 Write32(IndirectSymbolOffset);
362 Write32(NumIndirectSymbols);
363 Write32(0); // extreloff
364 Write32(0); // nextrel
365 Write32(0); // locreloff
366 Write32(0); // nlocrel
368 assert(OS.tell() - Start == DysymtabLoadCommandSize);
371 void WriteNlist32(MachSymbolData &MSD) {
372 MCSymbolData &Data = *MSD.SymbolData;
373 const MCSymbol &Symbol = Data.getSymbol();
375 uint16_t Flags = Data.getFlags();
376 uint32_t Address = 0;
378 // Set the N_TYPE bits. See <mach-o/nlist.h>.
380 // FIXME: Are the prebound or indirect fields possible here?
381 if (Symbol.isUndefined())
382 Type = STT_Undefined;
383 else if (Symbol.isAbsolute())
388 // FIXME: Set STAB bits.
390 if (Data.isPrivateExtern())
391 Type |= STF_PrivateExtern;
394 if (Data.isExternal() || Symbol.isUndefined())
395 Type |= STF_External;
397 // Compute the symbol address.
398 if (Symbol.isDefined()) {
399 if (Symbol.isAbsolute()) {
400 llvm_unreachable("FIXME: Not yet implemented!");
402 Address = Data.getFragment()->getAddress() + Data.getOffset();
404 } else if (Data.isCommon()) {
405 // Common symbols are encoded with the size in the address
406 // field, and their alignment in the flags.
407 Address = Data.getCommonSize();
409 // Common alignment is packed into the 'desc' bits.
410 if (unsigned Align = Data.getCommonAlignment()) {
411 unsigned Log2Size = Log2_32(Align);
412 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
414 llvm_report_error("invalid 'common' alignment '" +
416 // FIXME: Keep this mask with the SymbolFlags enumeration.
417 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
421 // struct nlist (12 bytes)
423 Write32(MSD.StringIndex);
425 Write8(MSD.SectionIndex);
427 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
433 struct MachRelocationEntry {
437 void ComputeScatteredRelocationInfo(MCAssembler &Asm, MCFragment &Fragment,
439 const MCValue &Target,
440 DenseMap<const MCSymbol*,MCSymbolData*> &SymbolMap,
441 std::vector<MachRelocationEntry> &Relocs) {
442 uint32_t Address = Fragment.getOffset() + Fixup.Offset;
443 unsigned IsPCRel = 0;
444 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
445 unsigned Type = RIT_Vanilla;
448 const MCSymbol *A = Target.getSymA();
449 MCSymbolData *SD = SymbolMap.lookup(A);
450 uint32_t Value = SD->getFragment()->getAddress() + SD->getOffset();
453 if (const MCSymbol *B = Target.getSymB()) {
454 Type = RIT_LocalDifference;
456 MCSymbolData *SD = SymbolMap.lookup(B);
457 Value2 = SD->getFragment()->getAddress() + SD->getOffset();
460 // The value which goes in the fixup is current value of the expression.
461 Fixup.FixedValue = Value - Value2 + Target.getConstant();
462 if (isFixupKindPCRel(Fixup.Kind)) {
463 Fixup.FixedValue -= Address;
467 MachRelocationEntry MRE;
468 MRE.Word0 = ((Address << 0) |
474 Relocs.push_back(MRE);
476 if (Type == RIT_LocalDifference) {
479 MachRelocationEntry MRE;
480 MRE.Word0 = ((0 << 0) |
486 Relocs.push_back(MRE);
490 void ComputeRelocationInfo(MCAssembler &Asm, MCDataFragment &Fragment,
492 DenseMap<const MCSymbol*,MCSymbolData*> &SymbolMap,
493 std::vector<MachRelocationEntry> &Relocs) {
495 if (!Fixup.Value->EvaluateAsRelocatable(Target))
496 llvm_report_error("expected relocatable expression");
498 // If this is a difference or a local symbol plus an offset, then we need a
499 // scattered relocation entry.
500 if (Target.getSymB() ||
501 (Target.getSymA() && !Target.getSymA()->isUndefined() &&
502 Target.getConstant()))
503 return ComputeScatteredRelocationInfo(Asm, Fragment, Fixup, Target,
507 uint32_t Address = Fragment.getOffset() + Fixup.Offset;
510 unsigned IsPCRel = 0;
511 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
512 unsigned IsExtern = 0;
515 if (Target.isAbsolute()) { // constant
516 // SymbolNum of 0 indicates the absolute section.
518 // FIXME: When is this generated?
521 llvm_unreachable("FIXME: Not yet implemented!");
523 const MCSymbol *Symbol = Target.getSymA();
524 MCSymbolData *SD = SymbolMap.lookup(Symbol);
526 if (Symbol->isUndefined()) {
528 Index = SD->getIndex();
531 // The index is the section ordinal.
535 MCAssembler::iterator it = Asm.begin(), ie = Asm.end();
536 for (; it != ie; ++it, ++Index)
537 if (&*it == SD->getFragment()->getParent())
539 assert(it != ie && "Unable to find section index!");
540 Value = SD->getFragment()->getAddress() + SD->getOffset();
546 // The value which goes in the fixup is current value of the expression.
547 Fixup.FixedValue = Value + Target.getConstant();
549 if (isFixupKindPCRel(Fixup.Kind)) {
550 Fixup.FixedValue -= Address;
554 // struct relocation_info (8 bytes)
555 MachRelocationEntry MRE;
557 MRE.Word1 = ((Index << 0) |
562 Relocs.push_back(MRE);
565 void BindIndirectSymbols(MCAssembler &Asm,
566 DenseMap<const MCSymbol*,MCSymbolData*> &SymbolMap) {
567 // This is the point where 'as' creates actual symbols for indirect symbols
568 // (in the following two passes). It would be easier for us to do this
569 // sooner when we see the attribute, but that makes getting the order in the
570 // symbol table much more complicated than it is worth.
572 // FIXME: Revisit this when the dust settles.
574 // Bind non lazy symbol pointers first.
575 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
576 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
577 // FIXME: cast<> support!
578 const MCSectionMachO &Section =
579 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
582 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
583 if (Type != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
586 MCSymbolData *&Entry = SymbolMap[it->Symbol];
588 Entry = new MCSymbolData(*it->Symbol, 0, 0, &Asm);
591 // Then lazy symbol pointers and symbol stubs.
592 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
593 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
594 // FIXME: cast<> support!
595 const MCSectionMachO &Section =
596 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
599 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
600 if (Type != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
601 Type != MCSectionMachO::S_SYMBOL_STUBS)
604 MCSymbolData *&Entry = SymbolMap[it->Symbol];
606 Entry = new MCSymbolData(*it->Symbol, 0, 0, &Asm);
608 // Set the symbol type to undefined lazy, but only on construction.
610 // FIXME: Do not hardcode.
611 Entry->setFlags(Entry->getFlags() | 0x0001);
616 /// ComputeSymbolTable - Compute the symbol table data
618 /// \param StringTable [out] - The string table data.
619 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
621 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
622 std::vector<MachSymbolData> &LocalSymbolData,
623 std::vector<MachSymbolData> &ExternalSymbolData,
624 std::vector<MachSymbolData> &UndefinedSymbolData) {
625 // Build section lookup table.
626 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
628 for (MCAssembler::iterator it = Asm.begin(),
629 ie = Asm.end(); it != ie; ++it, ++Index)
630 SectionIndexMap[&it->getSection()] = Index;
631 assert(Index <= 256 && "Too many sections!");
633 // Index 0 is always the empty string.
634 StringMap<uint64_t> StringIndexMap;
635 StringTable += '\x00';
637 // Build the symbol arrays and the string table, but only for non-local
640 // The particular order that we collect the symbols and create the string
641 // table, then sort the symbols is chosen to match 'as'. Even though it
642 // doesn't matter for correctness, this is important for letting us diff .o
644 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
645 ie = Asm.symbol_end(); it != ie; ++it) {
646 const MCSymbol &Symbol = it->getSymbol();
648 // Ignore assembler temporaries.
649 if (it->getSymbol().isTemporary())
652 if (!it->isExternal() && !Symbol.isUndefined())
655 uint64_t &Entry = StringIndexMap[Symbol.getName()];
657 Entry = StringTable.size();
658 StringTable += Symbol.getName();
659 StringTable += '\x00';
664 MSD.StringIndex = Entry;
666 if (Symbol.isUndefined()) {
667 MSD.SectionIndex = 0;
668 UndefinedSymbolData.push_back(MSD);
669 } else if (Symbol.isAbsolute()) {
670 MSD.SectionIndex = 0;
671 ExternalSymbolData.push_back(MSD);
673 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
674 assert(MSD.SectionIndex && "Invalid section index!");
675 ExternalSymbolData.push_back(MSD);
679 // Now add the data for local symbols.
680 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
681 ie = Asm.symbol_end(); it != ie; ++it) {
682 const MCSymbol &Symbol = it->getSymbol();
684 // Ignore assembler temporaries.
685 if (it->getSymbol().isTemporary())
688 if (it->isExternal() || Symbol.isUndefined())
691 uint64_t &Entry = StringIndexMap[Symbol.getName()];
693 Entry = StringTable.size();
694 StringTable += Symbol.getName();
695 StringTable += '\x00';
700 MSD.StringIndex = Entry;
702 if (Symbol.isAbsolute()) {
703 MSD.SectionIndex = 0;
704 LocalSymbolData.push_back(MSD);
706 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
707 assert(MSD.SectionIndex && "Invalid section index!");
708 LocalSymbolData.push_back(MSD);
712 // External and undefined symbols are required to be in lexicographic order.
713 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
714 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
716 // Set the symbol indices.
718 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
719 LocalSymbolData[i].SymbolData->setIndex(Index++);
720 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
721 ExternalSymbolData[i].SymbolData->setIndex(Index++);
722 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
723 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
725 // The string table is padded to a multiple of 4.
726 while (StringTable.size() % 4)
727 StringTable += '\x00';
730 void WriteObject(MCAssembler &Asm) {
731 unsigned NumSections = Asm.size();
733 // Compute the symbol -> symbol data map.
735 // FIXME: This should not be here.
736 DenseMap<const MCSymbol*, MCSymbolData *> SymbolMap;
737 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
738 ie = Asm.symbol_end(); it != ie; ++it)
739 SymbolMap[&it->getSymbol()] = it;
741 // Create symbol data for any indirect symbols.
742 BindIndirectSymbols(Asm, SymbolMap);
744 // Compute symbol table information.
745 SmallString<256> StringTable;
746 std::vector<MachSymbolData> LocalSymbolData;
747 std::vector<MachSymbolData> ExternalSymbolData;
748 std::vector<MachSymbolData> UndefinedSymbolData;
749 unsigned NumSymbols = Asm.symbol_size();
751 // No symbol table command is written if there are no symbols.
753 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
754 UndefinedSymbolData);
756 // The section data starts after the header, the segment load command (and
757 // section headers) and the symbol table.
758 unsigned NumLoadCommands = 1;
759 uint64_t LoadCommandsSize =
760 SegmentLoadCommand32Size + NumSections * Section32Size;
762 // Add the symbol table load command sizes, if used.
764 NumLoadCommands += 2;
765 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
768 // Compute the total size of the section data, as well as its file size and
770 uint64_t SectionDataStart = Header32Size + LoadCommandsSize;
771 uint64_t SectionDataSize = 0;
772 uint64_t SectionDataFileSize = 0;
774 for (MCAssembler::iterator it = Asm.begin(),
775 ie = Asm.end(); it != ie; ++it) {
776 MCSectionData &SD = *it;
778 VMSize = std::max(VMSize, SD.getAddress() + SD.getSize());
780 if (isVirtualSection(SD.getSection()))
783 SectionDataSize = std::max(SectionDataSize,
784 SD.getAddress() + SD.getSize());
785 SectionDataFileSize = std::max(SectionDataFileSize,
786 SD.getAddress() + SD.getFileSize());
789 // The section data is padded to 4 bytes.
791 // FIXME: Is this machine dependent?
792 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
793 SectionDataFileSize += SectionDataPadding;
795 // Write the prolog, starting with the header and load command...
796 WriteHeader32(NumLoadCommands, LoadCommandsSize,
797 Asm.getSubsectionsViaSymbols());
798 WriteSegmentLoadCommand32(NumSections, VMSize,
799 SectionDataStart, SectionDataSize);
801 // ... and then the section headers.
803 // We also compute the section relocations while we do this. Note that
804 // computing relocation info will also update the fixup to have the correct
805 // value; this will overwrite the appropriate data in the fragment when it
807 std::vector<MachRelocationEntry> RelocInfos;
808 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
809 for (MCAssembler::iterator it = Asm.begin(),
810 ie = Asm.end(); it != ie; ++it) {
811 MCSectionData &SD = *it;
813 // The assembler writes relocations in the reverse order they were seen.
815 // FIXME: It is probably more complicated than this.
816 unsigned NumRelocsStart = RelocInfos.size();
817 for (MCSectionData::reverse_iterator it2 = SD.rbegin(),
818 ie2 = SD.rend(); it2 != ie2; ++it2)
819 if (MCDataFragment *DF = dyn_cast<MCDataFragment>(&*it2))
820 for (unsigned i = 0, e = DF->fixup_size(); i != e; ++i)
821 ComputeRelocationInfo(Asm, *DF, DF->getFixups()[e - i - 1],
822 SymbolMap, RelocInfos);
824 unsigned NumRelocs = RelocInfos.size() - NumRelocsStart;
825 uint64_t SectionStart = SectionDataStart + SD.getAddress();
826 WriteSection32(SD, SectionStart, RelocTableEnd, NumRelocs);
827 RelocTableEnd += NumRelocs * RelocationInfoSize;
830 // Write the symbol table load command, if used.
832 unsigned FirstLocalSymbol = 0;
833 unsigned NumLocalSymbols = LocalSymbolData.size();
834 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
835 unsigned NumExternalSymbols = ExternalSymbolData.size();
836 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
837 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
838 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
839 unsigned NumSymTabSymbols =
840 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
841 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
842 uint64_t IndirectSymbolOffset = 0;
844 // If used, the indirect symbols are written after the section data.
845 if (NumIndirectSymbols)
846 IndirectSymbolOffset = RelocTableEnd;
848 // The symbol table is written after the indirect symbol data.
849 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
851 // The string table is written after symbol table.
852 uint64_t StringTableOffset =
853 SymbolTableOffset + NumSymTabSymbols * Nlist32Size;
854 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
855 StringTableOffset, StringTable.size());
857 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
858 FirstExternalSymbol, NumExternalSymbols,
859 FirstUndefinedSymbol, NumUndefinedSymbols,
860 IndirectSymbolOffset, NumIndirectSymbols);
863 // Write the actual section data.
864 for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it)
865 WriteFileData(OS, *it, *this);
867 // Write the extra padding.
868 WriteZeros(SectionDataPadding);
870 // Write the relocation entries.
871 for (unsigned i = 0, e = RelocInfos.size(); i != e; ++i) {
872 Write32(RelocInfos[i].Word0);
873 Write32(RelocInfos[i].Word1);
876 // Write the symbol table data, if used.
878 // Write the indirect symbol entries.
879 for (MCAssembler::indirect_symbol_iterator
880 it = Asm.indirect_symbol_begin(),
881 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
882 // Indirect symbols in the non lazy symbol pointer section have some
884 const MCSectionMachO &Section =
885 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
887 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
888 if (Type == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
889 // If this symbol is defined and internal, mark it as such.
890 if (it->Symbol->isDefined() &&
891 !SymbolMap.lookup(it->Symbol)->isExternal()) {
892 uint32_t Flags = ISF_Local;
893 if (it->Symbol->isAbsolute())
894 Flags |= ISF_Absolute;
900 Write32(SymbolMap[it->Symbol]->getIndex());
903 // FIXME: Check that offsets match computed ones.
905 // Write the symbol table entries.
906 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
907 WriteNlist32(LocalSymbolData[i]);
908 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
909 WriteNlist32(ExternalSymbolData[i]);
910 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
911 WriteNlist32(UndefinedSymbolData[i]);
913 // Write the string table.
914 OS << StringTable.str();
918 void ApplyFixup(const MCAsmFixup &Fixup, MCDataFragment &DF) {
919 unsigned Size = 1 << getFixupKindLog2Size(Fixup.Kind);
921 // FIXME: Endianness assumption.
922 assert(Fixup.Offset + Size <= DF.getContents().size() &&
923 "Invalid fixup offset!");
924 for (unsigned i = 0; i != Size; ++i)
925 DF.getContents()[Fixup.Offset + i] = uint8_t(Fixup.FixedValue >> (i * 8));
931 MCFragment::MCFragment() : Kind(FragmentType(~0)) {
934 MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent)
937 FileSize(~UINT64_C(0))
940 Parent->getFragmentList().push_back(this);
943 MCFragment::~MCFragment() {
946 uint64_t MCFragment::getAddress() const {
947 assert(getParent() && "Missing Section!");
948 return getParent()->getAddress() + Offset;
953 MCSectionData::MCSectionData() : Section(0) {}
955 MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A)
956 : Section(&_Section),
958 Address(~UINT64_C(0)),
960 FileSize(~UINT64_C(0)),
961 HasInstructions(false)
964 A->getSectionList().push_back(this);
969 MCSymbolData::MCSymbolData() : Symbol(0) {}
971 MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment,
972 uint64_t _Offset, MCAssembler *A)
973 : Symbol(&_Symbol), Fragment(_Fragment), Offset(_Offset),
974 IsExternal(false), IsPrivateExtern(false),
975 CommonSize(0), CommonAlign(0), Flags(0), Index(0)
978 A->getSymbolList().push_back(this);
983 MCAssembler::MCAssembler(MCContext &_Context, raw_ostream &_OS)
984 : Context(_Context), OS(_OS), SubsectionsViaSymbols(false)
988 MCAssembler::~MCAssembler() {
991 void MCAssembler::LayoutSection(MCSectionData &SD) {
992 uint64_t Address = SD.getAddress();
994 for (MCSectionData::iterator it = SD.begin(), ie = SD.end(); it != ie; ++it) {
997 F.setOffset(Address - SD.getAddress());
999 // Evaluate fragment size.
1000 switch (F.getKind()) {
1001 case MCFragment::FT_Align: {
1002 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1004 uint64_t Size = OffsetToAlignment(Address, AF.getAlignment());
1005 if (Size > AF.getMaxBytesToEmit())
1008 AF.setFileSize(Size);
1012 case MCFragment::FT_Data:
1013 case MCFragment::FT_Fill:
1014 F.setFileSize(F.getMaxFileSize());
1017 case MCFragment::FT_Org: {
1018 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1021 if (!OF.getOffset().EvaluateAsRelocatable(Target))
1022 llvm_report_error("expected relocatable expression");
1024 if (!Target.isAbsolute())
1025 llvm_unreachable("FIXME: Not yet implemented!");
1026 uint64_t OrgOffset = Target.getConstant();
1027 uint64_t Offset = Address - SD.getAddress();
1029 // FIXME: We need a way to communicate this error.
1030 if (OrgOffset < Offset)
1031 llvm_report_error("invalid .org offset '" + Twine(OrgOffset) +
1032 "' (at offset '" + Twine(Offset) + "'");
1034 F.setFileSize(OrgOffset - Offset);
1038 case MCFragment::FT_ZeroFill: {
1039 MCZeroFillFragment &ZFF = cast<MCZeroFillFragment>(F);
1041 // Align the fragment offset; it is safe to adjust the offset freely since
1042 // this is only in virtual sections.
1043 uint64_t Aligned = RoundUpToAlignment(Address, ZFF.getAlignment());
1044 F.setOffset(Aligned - SD.getAddress());
1046 // FIXME: This is misnamed.
1047 F.setFileSize(ZFF.getSize());
1052 Address += F.getFileSize();
1055 // Set the section sizes.
1056 SD.setSize(Address - SD.getAddress());
1057 if (isVirtualSection(SD.getSection()))
1060 SD.setFileSize(Address - SD.getAddress());
1063 /// WriteNopData - Write optimal nops to the output file for the \arg Count
1064 /// bytes. This returns the number of bytes written. It may return 0 if
1065 /// the \arg Count is more than the maximum optimal nops.
1067 /// FIXME this is X86 32-bit specific and should move to a better place.
1068 static uint64_t WriteNopData(uint64_t Count, MachObjectWriter &MOW) {
1069 static const uint8_t Nops[16][16] = {
1077 {0x0f, 0x1f, 0x40, 0x00},
1078 // nopl 0(%[re]ax,%[re]ax,1)
1079 {0x0f, 0x1f, 0x44, 0x00, 0x00},
1080 // nopw 0(%[re]ax,%[re]ax,1)
1081 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1083 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1084 // nopl 0L(%[re]ax,%[re]ax,1)
1085 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1086 // nopw 0L(%[re]ax,%[re]ax,1)
1087 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1088 // nopw %cs:0L(%[re]ax,%[re]ax,1)
1089 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1090 // nopl 0(%[re]ax,%[re]ax,1)
1091 // nopw 0(%[re]ax,%[re]ax,1)
1092 {0x0f, 0x1f, 0x44, 0x00, 0x00,
1093 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1094 // nopw 0(%[re]ax,%[re]ax,1)
1095 // nopw 0(%[re]ax,%[re]ax,1)
1096 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1097 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1098 // nopw 0(%[re]ax,%[re]ax,1)
1099 // nopl 0L(%[re]ax) */
1100 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1101 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1104 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1105 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1107 // nopl 0L(%[re]ax,%[re]ax,1)
1108 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1109 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}
1115 for (uint64_t i = 0; i < Count; i++)
1116 MOW.Write8 (uint8_t(Nops[Count - 1][i]));
1121 /// WriteFileData - Write the \arg F data to the output file.
1122 static void WriteFileData(raw_ostream &OS, const MCFragment &F,
1123 MachObjectWriter &MOW) {
1124 uint64_t Start = OS.tell();
1129 // FIXME: Embed in fragments instead?
1130 switch (F.getKind()) {
1131 case MCFragment::FT_Align: {
1132 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1133 uint64_t Count = AF.getFileSize() / AF.getValueSize();
1135 // FIXME: This error shouldn't actually occur (the front end should emit
1136 // multiple .align directives to enforce the semantics it wants), but is
1137 // severe enough that we want to report it. How to handle this?
1138 if (Count * AF.getValueSize() != AF.getFileSize())
1139 llvm_report_error("undefined .align directive, value size '" +
1140 Twine(AF.getValueSize()) +
1141 "' is not a divisor of padding size '" +
1142 Twine(AF.getFileSize()) + "'");
1144 // See if we are aligning with nops, and if so do that first to try to fill
1145 // the Count bytes. Then if that did not fill any bytes or there are any
1146 // bytes left to fill use the the Value and ValueSize to fill the rest.
1147 if (AF.getEmitNops()) {
1148 uint64_t NopByteCount = WriteNopData(Count, MOW);
1149 Count -= NopByteCount;
1152 for (uint64_t i = 0; i != Count; ++i) {
1153 switch (AF.getValueSize()) {
1155 assert(0 && "Invalid size!");
1156 case 1: MOW.Write8 (uint8_t (AF.getValue())); break;
1157 case 2: MOW.Write16(uint16_t(AF.getValue())); break;
1158 case 4: MOW.Write32(uint32_t(AF.getValue())); break;
1159 case 8: MOW.Write64(uint64_t(AF.getValue())); break;
1165 case MCFragment::FT_Data: {
1166 MCDataFragment &DF = cast<MCDataFragment>(F);
1168 // Apply the fixups.
1170 // FIXME: Move elsewhere.
1171 for (MCDataFragment::const_fixup_iterator it = DF.fixup_begin(),
1172 ie = DF.fixup_end(); it != ie; ++it)
1173 MOW.ApplyFixup(*it, DF);
1175 OS << cast<MCDataFragment>(F).getContents().str();
1179 case MCFragment::FT_Fill: {
1180 MCFillFragment &FF = cast<MCFillFragment>(F);
1181 for (uint64_t i = 0, e = FF.getCount(); i != e; ++i) {
1182 switch (FF.getValueSize()) {
1184 assert(0 && "Invalid size!");
1185 case 1: MOW.Write8 (uint8_t (FF.getValue())); break;
1186 case 2: MOW.Write16(uint16_t(FF.getValue())); break;
1187 case 4: MOW.Write32(uint32_t(FF.getValue())); break;
1188 case 8: MOW.Write64(uint64_t(FF.getValue())); break;
1194 case MCFragment::FT_Org: {
1195 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1197 for (uint64_t i = 0, e = OF.getFileSize(); i != e; ++i)
1198 MOW.Write8(uint8_t(OF.getValue()));
1203 case MCFragment::FT_ZeroFill: {
1204 assert(0 && "Invalid zero fill fragment in concrete section!");
1209 assert(OS.tell() - Start == F.getFileSize());
1212 /// WriteFileData - Write the \arg SD data to the output file.
1213 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
1214 MachObjectWriter &MOW) {
1215 // Ignore virtual sections.
1216 if (isVirtualSection(SD.getSection())) {
1217 assert(SD.getFileSize() == 0);
1221 uint64_t Start = OS.tell();
1224 for (MCSectionData::const_iterator it = SD.begin(),
1225 ie = SD.end(); it != ie; ++it)
1226 WriteFileData(OS, *it, MOW);
1228 // Add section padding.
1229 assert(SD.getFileSize() >= SD.getSize() && "Invalid section sizes!");
1230 MOW.WriteZeros(SD.getFileSize() - SD.getSize());
1232 assert(OS.tell() - Start == SD.getFileSize());
1235 void MCAssembler::Finish() {
1236 DEBUG_WITH_TYPE("mc-dump", {
1237 llvm::errs() << "assembler backend - pre-layout\n--\n";
1240 // Layout the concrete sections and fragments.
1241 uint64_t Address = 0;
1242 MCSectionData *Prev = 0;
1243 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1244 MCSectionData &SD = *it;
1246 // Skip virtual sections.
1247 if (isVirtualSection(SD.getSection()))
1250 // Align this section if necessary by adding padding bytes to the previous
1252 if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment())) {
1253 assert(Prev && "Missing prev section!");
1254 Prev->setFileSize(Prev->getFileSize() + Pad);
1258 // Layout the section fragments and its size.
1259 SD.setAddress(Address);
1261 Address += SD.getFileSize();
1266 // Layout the virtual sections.
1267 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1268 MCSectionData &SD = *it;
1270 if (!isVirtualSection(SD.getSection()))
1273 SD.setAddress(Address);
1275 Address += SD.getSize();
1278 DEBUG_WITH_TYPE("mc-dump", {
1279 llvm::errs() << "assembler backend - post-layout\n--\n";
1282 // Write the object file.
1283 MachObjectWriter MOW(OS);
1284 MOW.WriteObject(*this);
1290 // Debugging methods
1294 raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) {
1295 OS << "<MCAsmFixup" << " Offset:" << AF.Offset << " Value:" << *AF.Value
1296 << " Kind:" << AF.Kind << ">";
1302 void MCFragment::dump() {
1303 raw_ostream &OS = llvm::errs();
1305 OS << "<MCFragment " << (void*) this << " Offset:" << Offset
1306 << " FileSize:" << FileSize;
1311 void MCAlignFragment::dump() {
1312 raw_ostream &OS = llvm::errs();
1314 OS << "<MCAlignFragment ";
1315 this->MCFragment::dump();
1317 OS << " Alignment:" << getAlignment()
1318 << " Value:" << getValue() << " ValueSize:" << getValueSize()
1319 << " MaxBytesToEmit:" << getMaxBytesToEmit() << ">";
1322 void MCDataFragment::dump() {
1323 raw_ostream &OS = llvm::errs();
1325 OS << "<MCDataFragment ";
1326 this->MCFragment::dump();
1328 OS << " Contents:[";
1329 for (unsigned i = 0, e = getContents().size(); i != e; ++i) {
1331 OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF);
1333 OS << "] (" << getContents().size() << " bytes)";
1335 if (!getFixups().empty()) {
1338 for (fixup_iterator it = fixup_begin(), ie = fixup_end(); it != ie; ++it) {
1339 if (it != fixup_begin()) OS << ",\n ";
1348 void MCFillFragment::dump() {
1349 raw_ostream &OS = llvm::errs();
1351 OS << "<MCFillFragment ";
1352 this->MCFragment::dump();
1354 OS << " Value:" << getValue() << " ValueSize:" << getValueSize()
1355 << " Count:" << getCount() << ">";
1358 void MCOrgFragment::dump() {
1359 raw_ostream &OS = llvm::errs();
1361 OS << "<MCOrgFragment ";
1362 this->MCFragment::dump();
1364 OS << " Offset:" << getOffset() << " Value:" << getValue() << ">";
1367 void MCZeroFillFragment::dump() {
1368 raw_ostream &OS = llvm::errs();
1370 OS << "<MCZeroFillFragment ";
1371 this->MCFragment::dump();
1373 OS << " Size:" << getSize() << " Alignment:" << getAlignment() << ">";
1376 void MCSectionData::dump() {
1377 raw_ostream &OS = llvm::errs();
1379 OS << "<MCSectionData";
1380 OS << " Alignment:" << getAlignment() << " Address:" << Address
1381 << " Size:" << Size << " FileSize:" << FileSize
1383 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1384 if (it != begin()) OS << ",\n ";
1390 void MCSymbolData::dump() {
1391 raw_ostream &OS = llvm::errs();
1393 OS << "<MCSymbolData Symbol:" << getSymbol()
1394 << " Fragment:" << getFragment() << " Offset:" << getOffset()
1395 << " Flags:" << getFlags() << " Index:" << getIndex();
1397 OS << " (common, size:" << getCommonSize()
1398 << " align: " << getCommonAlignment() << ")";
1400 OS << " (external)";
1401 if (isPrivateExtern())
1402 OS << " (private extern)";
1406 void MCAssembler::dump() {
1407 raw_ostream &OS = llvm::errs();
1409 OS << "<MCAssembler\n";
1410 OS << " Sections:[";
1411 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1412 if (it != begin()) OS << ",\n ";
1418 for (symbol_iterator it = symbol_begin(), ie = symbol_end(); it != ie; ++it) {
1419 if (it != symbol_begin()) OS << ",\n ";