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 = isFixupKindPCRel(Fixup.Kind);
444 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
445 unsigned Type = RIT_Vanilla;
448 const MCSymbol *A = Target.getSymA();
449 MCSymbolData *A_SD = SymbolMap.lookup(A);
451 if (!A_SD->getFragment())
452 llvm_report_error("symbol '" + A->getName() +
453 "' can not be undefined in a subtraction expression");
455 uint32_t Value = A_SD->getFragment()->getAddress() + A_SD->getOffset();
458 if (const MCSymbol *B = Target.getSymB()) {
459 MCSymbolData *B_SD = SymbolMap.lookup(B);
461 if (!B_SD->getFragment())
462 llvm_report_error("symbol '" + B->getName() +
463 "' can not be undefined in a subtraction expression");
465 // FIXME: This change of type based on the external bit doesn't make much
466 // sense, it seems to be redundant with the other information in the
468 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
469 Value2 = B_SD->getFragment()->getAddress() + B_SD->getOffset();
472 // The value which goes in the fixup is current value of the expression.
473 Fixup.FixedValue = Value - Value2 + Target.getConstant();
475 Fixup.FixedValue -= Address;
477 // If this fixup is a vanilla PC relative relocation for a local label, we
478 // don't need a relocation.
480 // FIXME: Implement proper atom support.
481 if (IsPCRel && Target.getSymA() && Target.getSymA()->isTemporary() &&
485 MachRelocationEntry MRE;
486 MRE.Word0 = ((Address << 0) |
492 Relocs.push_back(MRE);
494 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
495 MachRelocationEntry MRE;
496 MRE.Word0 = ((0 << 0) |
502 Relocs.push_back(MRE);
506 void ComputeRelocationInfo(MCAssembler &Asm, MCDataFragment &Fragment,
508 DenseMap<const MCSymbol*,MCSymbolData*> &SymbolMap,
509 std::vector<MachRelocationEntry> &Relocs) {
510 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
511 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
514 if (!Fixup.Value->EvaluateAsRelocatable(Target))
515 llvm_report_error("expected relocatable expression");
517 // If this is a difference or a defined symbol plus an offset, then we need
518 // a scattered relocation entry.
519 uint32_t Offset = Target.getConstant();
521 Offset += 1 << Log2Size;
522 if (Target.getSymB() ||
523 (Target.getSymA() && !Target.getSymA()->isUndefined() &&
525 return ComputeScatteredRelocationInfo(Asm, Fragment, Fixup, Target,
529 uint32_t Address = Fragment.getOffset() + Fixup.Offset;
532 unsigned IsExtern = 0;
535 if (Target.isAbsolute()) { // constant
536 // SymbolNum of 0 indicates the absolute section.
538 // FIXME: When is this generated?
541 llvm_unreachable("FIXME: Not yet implemented!");
543 const MCSymbol *Symbol = Target.getSymA();
544 MCSymbolData *SD = SymbolMap.lookup(Symbol);
546 if (Symbol->isUndefined()) {
548 Index = SD->getIndex();
551 // The index is the section ordinal.
555 MCAssembler::iterator it = Asm.begin(), ie = Asm.end();
556 for (; it != ie; ++it, ++Index)
557 if (&*it == SD->getFragment()->getParent())
559 assert(it != ie && "Unable to find section index!");
560 Value = SD->getFragment()->getAddress() + SD->getOffset();
566 // The value which goes in the fixup is current value of the expression.
567 Fixup.FixedValue = Value + Target.getConstant();
569 Fixup.FixedValue -= Address;
571 // If this fixup is a vanilla PC relative relocation for a local label, we
572 // don't need a relocation.
574 // FIXME: Implement proper atom support.
575 if (IsPCRel && Target.getSymA() && Target.getSymA()->isTemporary())
578 // struct relocation_info (8 bytes)
579 MachRelocationEntry MRE;
581 MRE.Word1 = ((Index << 0) |
586 Relocs.push_back(MRE);
589 void BindIndirectSymbols(MCAssembler &Asm,
590 DenseMap<const MCSymbol*,MCSymbolData*> &SymbolMap) {
591 // This is the point where 'as' creates actual symbols for indirect symbols
592 // (in the following two passes). It would be easier for us to do this
593 // sooner when we see the attribute, but that makes getting the order in the
594 // symbol table much more complicated than it is worth.
596 // FIXME: Revisit this when the dust settles.
598 // Bind non lazy symbol pointers first.
599 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
600 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
601 // FIXME: cast<> support!
602 const MCSectionMachO &Section =
603 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
606 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
607 if (Type != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
610 MCSymbolData *&Entry = SymbolMap[it->Symbol];
612 Entry = new MCSymbolData(*it->Symbol, 0, 0, &Asm);
615 // Then lazy symbol pointers and symbol stubs.
616 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
617 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
618 // FIXME: cast<> support!
619 const MCSectionMachO &Section =
620 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
623 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
624 if (Type != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
625 Type != MCSectionMachO::S_SYMBOL_STUBS)
628 MCSymbolData *&Entry = SymbolMap[it->Symbol];
630 Entry = new MCSymbolData(*it->Symbol, 0, 0, &Asm);
632 // Set the symbol type to undefined lazy, but only on construction.
634 // FIXME: Do not hardcode.
635 Entry->setFlags(Entry->getFlags() | 0x0001);
640 /// ComputeSymbolTable - Compute the symbol table data
642 /// \param StringTable [out] - The string table data.
643 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
645 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
646 std::vector<MachSymbolData> &LocalSymbolData,
647 std::vector<MachSymbolData> &ExternalSymbolData,
648 std::vector<MachSymbolData> &UndefinedSymbolData) {
649 // Build section lookup table.
650 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
652 for (MCAssembler::iterator it = Asm.begin(),
653 ie = Asm.end(); it != ie; ++it, ++Index)
654 SectionIndexMap[&it->getSection()] = Index;
655 assert(Index <= 256 && "Too many sections!");
657 // Index 0 is always the empty string.
658 StringMap<uint64_t> StringIndexMap;
659 StringTable += '\x00';
661 // Build the symbol arrays and the string table, but only for non-local
664 // The particular order that we collect the symbols and create the string
665 // table, then sort the symbols is chosen to match 'as'. Even though it
666 // doesn't matter for correctness, this is important for letting us diff .o
668 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
669 ie = Asm.symbol_end(); it != ie; ++it) {
670 const MCSymbol &Symbol = it->getSymbol();
672 // Ignore assembler temporaries.
673 if (it->getSymbol().isTemporary())
676 if (!it->isExternal() && !Symbol.isUndefined())
679 uint64_t &Entry = StringIndexMap[Symbol.getName()];
681 Entry = StringTable.size();
682 StringTable += Symbol.getName();
683 StringTable += '\x00';
688 MSD.StringIndex = Entry;
690 if (Symbol.isUndefined()) {
691 MSD.SectionIndex = 0;
692 UndefinedSymbolData.push_back(MSD);
693 } else if (Symbol.isAbsolute()) {
694 MSD.SectionIndex = 0;
695 ExternalSymbolData.push_back(MSD);
697 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
698 assert(MSD.SectionIndex && "Invalid section index!");
699 ExternalSymbolData.push_back(MSD);
703 // Now add the data for local symbols.
704 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
705 ie = Asm.symbol_end(); it != ie; ++it) {
706 const MCSymbol &Symbol = it->getSymbol();
708 // Ignore assembler temporaries.
709 if (it->getSymbol().isTemporary())
712 if (it->isExternal() || Symbol.isUndefined())
715 uint64_t &Entry = StringIndexMap[Symbol.getName()];
717 Entry = StringTable.size();
718 StringTable += Symbol.getName();
719 StringTable += '\x00';
724 MSD.StringIndex = Entry;
726 if (Symbol.isAbsolute()) {
727 MSD.SectionIndex = 0;
728 LocalSymbolData.push_back(MSD);
730 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
731 assert(MSD.SectionIndex && "Invalid section index!");
732 LocalSymbolData.push_back(MSD);
736 // External and undefined symbols are required to be in lexicographic order.
737 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
738 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
740 // Set the symbol indices.
742 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
743 LocalSymbolData[i].SymbolData->setIndex(Index++);
744 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
745 ExternalSymbolData[i].SymbolData->setIndex(Index++);
746 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
747 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
749 // The string table is padded to a multiple of 4.
750 while (StringTable.size() % 4)
751 StringTable += '\x00';
754 void WriteObject(MCAssembler &Asm) {
755 unsigned NumSections = Asm.size();
757 // Compute the symbol -> symbol data map.
759 // FIXME: This should not be here.
760 DenseMap<const MCSymbol*, MCSymbolData *> SymbolMap;
761 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
762 ie = Asm.symbol_end(); it != ie; ++it)
763 SymbolMap[&it->getSymbol()] = it;
765 // Create symbol data for any indirect symbols.
766 BindIndirectSymbols(Asm, SymbolMap);
768 // Compute symbol table information.
769 SmallString<256> StringTable;
770 std::vector<MachSymbolData> LocalSymbolData;
771 std::vector<MachSymbolData> ExternalSymbolData;
772 std::vector<MachSymbolData> UndefinedSymbolData;
773 unsigned NumSymbols = Asm.symbol_size();
775 // No symbol table command is written if there are no symbols.
777 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
778 UndefinedSymbolData);
780 // The section data starts after the header, the segment load command (and
781 // section headers) and the symbol table.
782 unsigned NumLoadCommands = 1;
783 uint64_t LoadCommandsSize =
784 SegmentLoadCommand32Size + NumSections * Section32Size;
786 // Add the symbol table load command sizes, if used.
788 NumLoadCommands += 2;
789 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
792 // Compute the total size of the section data, as well as its file size and
794 uint64_t SectionDataStart = Header32Size + LoadCommandsSize;
795 uint64_t SectionDataSize = 0;
796 uint64_t SectionDataFileSize = 0;
798 for (MCAssembler::iterator it = Asm.begin(),
799 ie = Asm.end(); it != ie; ++it) {
800 MCSectionData &SD = *it;
802 VMSize = std::max(VMSize, SD.getAddress() + SD.getSize());
804 if (isVirtualSection(SD.getSection()))
807 SectionDataSize = std::max(SectionDataSize,
808 SD.getAddress() + SD.getSize());
809 SectionDataFileSize = std::max(SectionDataFileSize,
810 SD.getAddress() + SD.getFileSize());
813 // The section data is padded to 4 bytes.
815 // FIXME: Is this machine dependent?
816 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
817 SectionDataFileSize += SectionDataPadding;
819 // Write the prolog, starting with the header and load command...
820 WriteHeader32(NumLoadCommands, LoadCommandsSize,
821 Asm.getSubsectionsViaSymbols());
822 WriteSegmentLoadCommand32(NumSections, VMSize,
823 SectionDataStart, SectionDataSize);
825 // ... and then the section headers.
827 // We also compute the section relocations while we do this. Note that
828 // computing relocation info will also update the fixup to have the correct
829 // value; this will overwrite the appropriate data in the fragment when it
831 std::vector<MachRelocationEntry> RelocInfos;
832 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
833 for (MCAssembler::iterator it = Asm.begin(),
834 ie = Asm.end(); it != ie; ++it) {
835 MCSectionData &SD = *it;
837 // The assembler writes relocations in the reverse order they were seen.
839 // FIXME: It is probably more complicated than this.
840 unsigned NumRelocsStart = RelocInfos.size();
841 for (MCSectionData::reverse_iterator it2 = SD.rbegin(),
842 ie2 = SD.rend(); it2 != ie2; ++it2)
843 if (MCDataFragment *DF = dyn_cast<MCDataFragment>(&*it2))
844 for (unsigned i = 0, e = DF->fixup_size(); i != e; ++i)
845 ComputeRelocationInfo(Asm, *DF, DF->getFixups()[e - i - 1],
846 SymbolMap, RelocInfos);
848 unsigned NumRelocs = RelocInfos.size() - NumRelocsStart;
849 uint64_t SectionStart = SectionDataStart + SD.getAddress();
850 WriteSection32(SD, SectionStart, RelocTableEnd, NumRelocs);
851 RelocTableEnd += NumRelocs * RelocationInfoSize;
854 // Write the symbol table load command, if used.
856 unsigned FirstLocalSymbol = 0;
857 unsigned NumLocalSymbols = LocalSymbolData.size();
858 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
859 unsigned NumExternalSymbols = ExternalSymbolData.size();
860 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
861 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
862 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
863 unsigned NumSymTabSymbols =
864 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
865 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
866 uint64_t IndirectSymbolOffset = 0;
868 // If used, the indirect symbols are written after the section data.
869 if (NumIndirectSymbols)
870 IndirectSymbolOffset = RelocTableEnd;
872 // The symbol table is written after the indirect symbol data.
873 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
875 // The string table is written after symbol table.
876 uint64_t StringTableOffset =
877 SymbolTableOffset + NumSymTabSymbols * Nlist32Size;
878 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
879 StringTableOffset, StringTable.size());
881 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
882 FirstExternalSymbol, NumExternalSymbols,
883 FirstUndefinedSymbol, NumUndefinedSymbols,
884 IndirectSymbolOffset, NumIndirectSymbols);
887 // Write the actual section data.
888 for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it)
889 WriteFileData(OS, *it, *this);
891 // Write the extra padding.
892 WriteZeros(SectionDataPadding);
894 // Write the relocation entries.
895 for (unsigned i = 0, e = RelocInfos.size(); i != e; ++i) {
896 Write32(RelocInfos[i].Word0);
897 Write32(RelocInfos[i].Word1);
900 // Write the symbol table data, if used.
902 // Write the indirect symbol entries.
903 for (MCAssembler::indirect_symbol_iterator
904 it = Asm.indirect_symbol_begin(),
905 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
906 // Indirect symbols in the non lazy symbol pointer section have some
908 const MCSectionMachO &Section =
909 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
911 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
912 if (Type == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
913 // If this symbol is defined and internal, mark it as such.
914 if (it->Symbol->isDefined() &&
915 !SymbolMap.lookup(it->Symbol)->isExternal()) {
916 uint32_t Flags = ISF_Local;
917 if (it->Symbol->isAbsolute())
918 Flags |= ISF_Absolute;
924 Write32(SymbolMap[it->Symbol]->getIndex());
927 // FIXME: Check that offsets match computed ones.
929 // Write the symbol table entries.
930 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
931 WriteNlist32(LocalSymbolData[i]);
932 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
933 WriteNlist32(ExternalSymbolData[i]);
934 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
935 WriteNlist32(UndefinedSymbolData[i]);
937 // Write the string table.
938 OS << StringTable.str();
942 void ApplyFixup(const MCAsmFixup &Fixup, MCDataFragment &DF) {
943 unsigned Size = 1 << getFixupKindLog2Size(Fixup.Kind);
945 // FIXME: Endianness assumption.
946 assert(Fixup.Offset + Size <= DF.getContents().size() &&
947 "Invalid fixup offset!");
948 for (unsigned i = 0; i != Size; ++i)
949 DF.getContents()[Fixup.Offset + i] = uint8_t(Fixup.FixedValue >> (i * 8));
955 MCFragment::MCFragment() : Kind(FragmentType(~0)) {
958 MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent)
961 FileSize(~UINT64_C(0))
964 Parent->getFragmentList().push_back(this);
967 MCFragment::~MCFragment() {
970 uint64_t MCFragment::getAddress() const {
971 assert(getParent() && "Missing Section!");
972 return getParent()->getAddress() + Offset;
977 MCSectionData::MCSectionData() : Section(0) {}
979 MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A)
980 : Section(&_Section),
982 Address(~UINT64_C(0)),
984 FileSize(~UINT64_C(0)),
985 HasInstructions(false)
988 A->getSectionList().push_back(this);
993 MCSymbolData::MCSymbolData() : Symbol(0) {}
995 MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment,
996 uint64_t _Offset, MCAssembler *A)
997 : Symbol(&_Symbol), Fragment(_Fragment), Offset(_Offset),
998 IsExternal(false), IsPrivateExtern(false),
999 CommonSize(0), CommonAlign(0), Flags(0), Index(0)
1002 A->getSymbolList().push_back(this);
1007 MCAssembler::MCAssembler(MCContext &_Context, raw_ostream &_OS)
1008 : Context(_Context), OS(_OS), SubsectionsViaSymbols(false)
1012 MCAssembler::~MCAssembler() {
1015 void MCAssembler::LayoutSection(MCSectionData &SD) {
1016 uint64_t Address = SD.getAddress();
1018 for (MCSectionData::iterator it = SD.begin(), ie = SD.end(); it != ie; ++it) {
1019 MCFragment &F = *it;
1021 F.setOffset(Address - SD.getAddress());
1023 // Evaluate fragment size.
1024 switch (F.getKind()) {
1025 case MCFragment::FT_Align: {
1026 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1028 uint64_t Size = OffsetToAlignment(Address, AF.getAlignment());
1029 if (Size > AF.getMaxBytesToEmit())
1032 AF.setFileSize(Size);
1036 case MCFragment::FT_Data:
1037 case MCFragment::FT_Fill:
1038 F.setFileSize(F.getMaxFileSize());
1041 case MCFragment::FT_Org: {
1042 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1045 if (!OF.getOffset().EvaluateAsRelocatable(Target))
1046 llvm_report_error("expected relocatable expression");
1048 if (!Target.isAbsolute())
1049 llvm_unreachable("FIXME: Not yet implemented!");
1050 uint64_t OrgOffset = Target.getConstant();
1051 uint64_t Offset = Address - SD.getAddress();
1053 // FIXME: We need a way to communicate this error.
1054 if (OrgOffset < Offset)
1055 llvm_report_error("invalid .org offset '" + Twine(OrgOffset) +
1056 "' (at offset '" + Twine(Offset) + "'");
1058 F.setFileSize(OrgOffset - Offset);
1062 case MCFragment::FT_ZeroFill: {
1063 MCZeroFillFragment &ZFF = cast<MCZeroFillFragment>(F);
1065 // Align the fragment offset; it is safe to adjust the offset freely since
1066 // this is only in virtual sections.
1067 Address = RoundUpToAlignment(Address, ZFF.getAlignment());
1068 F.setOffset(Address - SD.getAddress());
1070 // FIXME: This is misnamed.
1071 F.setFileSize(ZFF.getSize());
1076 Address += F.getFileSize();
1079 // Set the section sizes.
1080 SD.setSize(Address - SD.getAddress());
1081 if (isVirtualSection(SD.getSection()))
1084 SD.setFileSize(Address - SD.getAddress());
1087 /// WriteNopData - Write optimal nops to the output file for the \arg Count
1088 /// bytes. This returns the number of bytes written. It may return 0 if
1089 /// the \arg Count is more than the maximum optimal nops.
1091 /// FIXME this is X86 32-bit specific and should move to a better place.
1092 static uint64_t WriteNopData(uint64_t Count, MachObjectWriter &MOW) {
1093 static const uint8_t Nops[16][16] = {
1101 {0x0f, 0x1f, 0x40, 0x00},
1102 // nopl 0(%[re]ax,%[re]ax,1)
1103 {0x0f, 0x1f, 0x44, 0x00, 0x00},
1104 // nopw 0(%[re]ax,%[re]ax,1)
1105 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1107 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1108 // nopl 0L(%[re]ax,%[re]ax,1)
1109 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1110 // nopw 0L(%[re]ax,%[re]ax,1)
1111 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1112 // nopw %cs:0L(%[re]ax,%[re]ax,1)
1113 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1114 // nopl 0(%[re]ax,%[re]ax,1)
1115 // nopw 0(%[re]ax,%[re]ax,1)
1116 {0x0f, 0x1f, 0x44, 0x00, 0x00,
1117 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1118 // nopw 0(%[re]ax,%[re]ax,1)
1119 // nopw 0(%[re]ax,%[re]ax,1)
1120 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1121 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1122 // nopw 0(%[re]ax,%[re]ax,1)
1123 // nopl 0L(%[re]ax) */
1124 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1125 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1128 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1129 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1131 // nopl 0L(%[re]ax,%[re]ax,1)
1132 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1133 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}
1139 for (uint64_t i = 0; i < Count; i++)
1140 MOW.Write8 (uint8_t(Nops[Count - 1][i]));
1145 /// WriteFileData - Write the \arg F data to the output file.
1146 static void WriteFileData(raw_ostream &OS, const MCFragment &F,
1147 MachObjectWriter &MOW) {
1148 uint64_t Start = OS.tell();
1153 // FIXME: Embed in fragments instead?
1154 switch (F.getKind()) {
1155 case MCFragment::FT_Align: {
1156 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1157 uint64_t Count = AF.getFileSize() / AF.getValueSize();
1159 // FIXME: This error shouldn't actually occur (the front end should emit
1160 // multiple .align directives to enforce the semantics it wants), but is
1161 // severe enough that we want to report it. How to handle this?
1162 if (Count * AF.getValueSize() != AF.getFileSize())
1163 llvm_report_error("undefined .align directive, value size '" +
1164 Twine(AF.getValueSize()) +
1165 "' is not a divisor of padding size '" +
1166 Twine(AF.getFileSize()) + "'");
1168 // See if we are aligning with nops, and if so do that first to try to fill
1169 // the Count bytes. Then if that did not fill any bytes or there are any
1170 // bytes left to fill use the the Value and ValueSize to fill the rest.
1171 if (AF.getEmitNops()) {
1172 uint64_t NopByteCount = WriteNopData(Count, MOW);
1173 Count -= NopByteCount;
1176 for (uint64_t i = 0; i != Count; ++i) {
1177 switch (AF.getValueSize()) {
1179 assert(0 && "Invalid size!");
1180 case 1: MOW.Write8 (uint8_t (AF.getValue())); break;
1181 case 2: MOW.Write16(uint16_t(AF.getValue())); break;
1182 case 4: MOW.Write32(uint32_t(AF.getValue())); break;
1183 case 8: MOW.Write64(uint64_t(AF.getValue())); break;
1189 case MCFragment::FT_Data: {
1190 MCDataFragment &DF = cast<MCDataFragment>(F);
1192 // Apply the fixups.
1194 // FIXME: Move elsewhere.
1195 for (MCDataFragment::const_fixup_iterator it = DF.fixup_begin(),
1196 ie = DF.fixup_end(); it != ie; ++it)
1197 MOW.ApplyFixup(*it, DF);
1199 OS << cast<MCDataFragment>(F).getContents().str();
1203 case MCFragment::FT_Fill: {
1204 MCFillFragment &FF = cast<MCFillFragment>(F);
1205 for (uint64_t i = 0, e = FF.getCount(); i != e; ++i) {
1206 switch (FF.getValueSize()) {
1208 assert(0 && "Invalid size!");
1209 case 1: MOW.Write8 (uint8_t (FF.getValue())); break;
1210 case 2: MOW.Write16(uint16_t(FF.getValue())); break;
1211 case 4: MOW.Write32(uint32_t(FF.getValue())); break;
1212 case 8: MOW.Write64(uint64_t(FF.getValue())); break;
1218 case MCFragment::FT_Org: {
1219 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1221 for (uint64_t i = 0, e = OF.getFileSize(); i != e; ++i)
1222 MOW.Write8(uint8_t(OF.getValue()));
1227 case MCFragment::FT_ZeroFill: {
1228 assert(0 && "Invalid zero fill fragment in concrete section!");
1233 assert(OS.tell() - Start == F.getFileSize());
1236 /// WriteFileData - Write the \arg SD data to the output file.
1237 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
1238 MachObjectWriter &MOW) {
1239 // Ignore virtual sections.
1240 if (isVirtualSection(SD.getSection())) {
1241 assert(SD.getFileSize() == 0);
1245 uint64_t Start = OS.tell();
1248 for (MCSectionData::const_iterator it = SD.begin(),
1249 ie = SD.end(); it != ie; ++it)
1250 WriteFileData(OS, *it, MOW);
1252 // Add section padding.
1253 assert(SD.getFileSize() >= SD.getSize() && "Invalid section sizes!");
1254 MOW.WriteZeros(SD.getFileSize() - SD.getSize());
1256 assert(OS.tell() - Start == SD.getFileSize());
1259 void MCAssembler::Finish() {
1260 DEBUG_WITH_TYPE("mc-dump", {
1261 llvm::errs() << "assembler backend - pre-layout\n--\n";
1264 // Layout the concrete sections and fragments.
1265 uint64_t Address = 0;
1266 MCSectionData *Prev = 0;
1267 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1268 MCSectionData &SD = *it;
1270 // Skip virtual sections.
1271 if (isVirtualSection(SD.getSection()))
1274 // Align this section if necessary by adding padding bytes to the previous
1276 if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment())) {
1277 assert(Prev && "Missing prev section!");
1278 Prev->setFileSize(Prev->getFileSize() + Pad);
1282 // Layout the section fragments and its size.
1283 SD.setAddress(Address);
1285 Address += SD.getFileSize();
1290 // Layout the virtual sections.
1291 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1292 MCSectionData &SD = *it;
1294 if (!isVirtualSection(SD.getSection()))
1297 // Align this section if necessary by adding padding bytes to the previous
1299 if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment()))
1302 SD.setAddress(Address);
1304 Address += SD.getSize();
1308 DEBUG_WITH_TYPE("mc-dump", {
1309 llvm::errs() << "assembler backend - post-layout\n--\n";
1312 // Write the object file.
1313 MachObjectWriter MOW(OS);
1314 MOW.WriteObject(*this);
1320 // Debugging methods
1324 raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) {
1325 OS << "<MCAsmFixup" << " Offset:" << AF.Offset << " Value:" << *AF.Value
1326 << " Kind:" << AF.Kind << ">";
1332 void MCFragment::dump() {
1333 raw_ostream &OS = llvm::errs();
1335 OS << "<MCFragment " << (void*) this << " Offset:" << Offset
1336 << " FileSize:" << FileSize;
1341 void MCAlignFragment::dump() {
1342 raw_ostream &OS = llvm::errs();
1344 OS << "<MCAlignFragment ";
1345 this->MCFragment::dump();
1347 OS << " Alignment:" << getAlignment()
1348 << " Value:" << getValue() << " ValueSize:" << getValueSize()
1349 << " MaxBytesToEmit:" << getMaxBytesToEmit() << ">";
1352 void MCDataFragment::dump() {
1353 raw_ostream &OS = llvm::errs();
1355 OS << "<MCDataFragment ";
1356 this->MCFragment::dump();
1358 OS << " Contents:[";
1359 for (unsigned i = 0, e = getContents().size(); i != e; ++i) {
1361 OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF);
1363 OS << "] (" << getContents().size() << " bytes)";
1365 if (!getFixups().empty()) {
1368 for (fixup_iterator it = fixup_begin(), ie = fixup_end(); it != ie; ++it) {
1369 if (it != fixup_begin()) OS << ",\n ";
1378 void MCFillFragment::dump() {
1379 raw_ostream &OS = llvm::errs();
1381 OS << "<MCFillFragment ";
1382 this->MCFragment::dump();
1384 OS << " Value:" << getValue() << " ValueSize:" << getValueSize()
1385 << " Count:" << getCount() << ">";
1388 void MCOrgFragment::dump() {
1389 raw_ostream &OS = llvm::errs();
1391 OS << "<MCOrgFragment ";
1392 this->MCFragment::dump();
1394 OS << " Offset:" << getOffset() << " Value:" << getValue() << ">";
1397 void MCZeroFillFragment::dump() {
1398 raw_ostream &OS = llvm::errs();
1400 OS << "<MCZeroFillFragment ";
1401 this->MCFragment::dump();
1403 OS << " Size:" << getSize() << " Alignment:" << getAlignment() << ">";
1406 void MCSectionData::dump() {
1407 raw_ostream &OS = llvm::errs();
1409 OS << "<MCSectionData";
1410 OS << " Alignment:" << getAlignment() << " Address:" << Address
1411 << " Size:" << Size << " FileSize:" << FileSize
1412 << " Fragments:[\n ";
1413 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1414 if (it != begin()) OS << ",\n ";
1420 void MCSymbolData::dump() {
1421 raw_ostream &OS = llvm::errs();
1423 OS << "<MCSymbolData Symbol:" << getSymbol()
1424 << " Fragment:" << getFragment() << " Offset:" << getOffset()
1425 << " Flags:" << getFlags() << " Index:" << getIndex();
1427 OS << " (common, size:" << getCommonSize()
1428 << " align: " << getCommonAlignment() << ")";
1430 OS << " (external)";
1431 if (isPrivateExtern())
1432 OS << " (private extern)";
1436 void MCAssembler::dump() {
1437 raw_ostream &OS = llvm::errs();
1439 OS << "<MCAssembler\n";
1440 OS << " Sections:[\n ";
1441 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1442 if (it != begin()) OS << ",\n ";
1448 for (symbol_iterator it = symbol_begin(), ie = symbol_end(); it != ie; ++it) {
1449 if (it != symbol_begin()) OS << ",\n ";