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/MCAsmLayout.h"
13 #include "llvm/MC/MCExpr.h"
14 #include "llvm/MC/MCSectionMachO.h"
15 #include "llvm/MC/MCSymbol.h"
16 #include "llvm/MC/MCValue.h"
17 #include "llvm/ADT/DenseMap.h"
18 #include "llvm/ADT/SmallString.h"
19 #include "llvm/ADT/Statistic.h"
20 #include "llvm/ADT/StringExtras.h"
21 #include "llvm/ADT/StringMap.h"
22 #include "llvm/ADT/Twine.h"
23 #include "llvm/Support/ErrorHandling.h"
24 #include "llvm/Support/MachO.h"
25 #include "llvm/Support/raw_ostream.h"
26 #include "llvm/Support/Debug.h"
27 #include "llvm/Target/TargetRegistry.h"
28 #include "llvm/Target/TargetAsmBackend.h"
31 #include "../Target/X86/X86FixupKinds.h"
36 class MachObjectWriter;
38 STATISTIC(EmittedFragments, "Number of emitted assembler fragments");
40 // FIXME FIXME FIXME: There are number of places in this file where we convert
41 // what is a 64-bit assembler value used for computation into a value in the
42 // object file, which may truncate it. We should detect that truncation where
43 // invalid and report errors back.
45 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
46 MachObjectWriter &MOW);
48 static uint64_t WriteNopData(uint64_t Count, MachObjectWriter &MOW);
50 /// isVirtualSection - Check if this is a section which does not actually exist
51 /// in the object file.
52 static bool isVirtualSection(const MCSection &Section) {
54 const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section);
55 return (SMO.getType() == MCSectionMachO::S_ZEROFILL);
58 static unsigned getFixupKindLog2Size(unsigned Kind) {
60 default: llvm_unreachable("invalid fixup kind!");
61 case X86::reloc_pcrel_1byte:
62 case FK_Data_1: return 0;
63 case FK_Data_2: return 1;
64 case X86::reloc_pcrel_4byte:
65 case X86::reloc_riprel_4byte:
66 case FK_Data_4: return 2;
67 case FK_Data_8: return 3;
71 static bool isFixupKindPCRel(unsigned Kind) {
75 case X86::reloc_pcrel_1byte:
76 case X86::reloc_pcrel_4byte:
77 case X86::reloc_riprel_4byte:
82 class MachObjectWriter {
83 // See <mach-o/loader.h>.
85 Header_Magic32 = 0xFEEDFACE,
86 Header_Magic64 = 0xFEEDFACF
92 SegmentLoadCommand32Size = 56,
93 SegmentLoadCommand64Size = 72,
96 SymtabLoadCommandSize = 24,
97 DysymtabLoadCommandSize = 80,
100 RelocationInfoSize = 8
103 enum HeaderFileType {
108 HF_SubsectionsViaSymbols = 0x2000
111 enum LoadCommandType {
118 // See <mach-o/nlist.h>.
119 enum SymbolTypeType {
120 STT_Undefined = 0x00,
125 enum SymbolTypeFlags {
126 // If any of these bits are set, then the entry is a stab entry number (see
127 // <mach-o/stab.h>. Otherwise the other masks apply.
128 STF_StabsEntryMask = 0xe0,
132 STF_PrivateExtern = 0x10
135 /// IndirectSymbolFlags - Flags for encoding special values in the indirect
137 enum IndirectSymbolFlags {
138 ISF_Local = 0x80000000,
139 ISF_Absolute = 0x40000000
142 /// RelocationFlags - Special flags for addresses.
143 enum RelocationFlags {
144 RF_Scattered = 0x80000000
147 enum RelocationInfoType {
151 RIT_PreboundLazyPointer = 3,
152 RIT_LocalDifference = 4
155 /// MachSymbolData - Helper struct for containing some precomputed information
157 struct MachSymbolData {
158 MCSymbolData *SymbolData;
159 uint64_t StringIndex;
160 uint8_t SectionIndex;
162 // Support lexicographic sorting.
163 bool operator<(const MachSymbolData &RHS) const {
164 const std::string &Name = SymbolData->getSymbol().getName();
165 return Name < RHS.SymbolData->getSymbol().getName();
170 unsigned Is64Bit : 1;
173 /// @name Relocation Data
176 struct MachRelocationEntry {
181 llvm::DenseMap<const MCSectionData*,
182 std::vector<MachRelocationEntry> > Relocations;
185 /// @name Symbol Table Data
187 SmallString<256> StringTable;
188 std::vector<MachSymbolData> LocalSymbolData;
189 std::vector<MachSymbolData> ExternalSymbolData;
190 std::vector<MachSymbolData> UndefinedSymbolData;
195 MachObjectWriter(raw_ostream &_OS, bool _Is64Bit, bool _IsLSB = true)
196 : OS(_OS), Is64Bit(_Is64Bit), IsLSB(_IsLSB) {
199 /// @name Helper Methods
202 void Write8(uint8_t Value) {
206 void Write16(uint16_t Value) {
208 Write8(uint8_t(Value >> 0));
209 Write8(uint8_t(Value >> 8));
211 Write8(uint8_t(Value >> 8));
212 Write8(uint8_t(Value >> 0));
216 void Write32(uint32_t Value) {
218 Write16(uint16_t(Value >> 0));
219 Write16(uint16_t(Value >> 16));
221 Write16(uint16_t(Value >> 16));
222 Write16(uint16_t(Value >> 0));
226 void Write64(uint64_t Value) {
228 Write32(uint32_t(Value >> 0));
229 Write32(uint32_t(Value >> 32));
231 Write32(uint32_t(Value >> 32));
232 Write32(uint32_t(Value >> 0));
236 void WriteZeros(unsigned N) {
237 const char Zeros[16] = { 0 };
239 for (unsigned i = 0, e = N / 16; i != e; ++i)
240 OS << StringRef(Zeros, 16);
242 OS << StringRef(Zeros, N % 16);
245 void WriteString(StringRef Str, unsigned ZeroFillSize = 0) {
248 WriteZeros(ZeroFillSize - Str.size());
253 void WriteHeader(unsigned NumLoadCommands, unsigned LoadCommandsSize,
254 bool SubsectionsViaSymbols) {
257 if (SubsectionsViaSymbols)
258 Flags |= HF_SubsectionsViaSymbols;
260 // struct mach_header (28 bytes) or
261 // struct mach_header_64 (32 bytes)
263 uint64_t Start = OS.tell();
266 Write32(Is64Bit ? Header_Magic64 : Header_Magic32);
268 // FIXME: Support cputype.
269 Write32(Is64Bit ? MachO::CPUTypeX86_64 : MachO::CPUTypeI386);
270 // FIXME: Support cpusubtype.
271 Write32(MachO::CPUSubType_I386_ALL);
273 Write32(NumLoadCommands); // Object files have a single load command, the
275 Write32(LoadCommandsSize);
278 Write32(0); // reserved
280 assert(OS.tell() - Start == Is64Bit ? Header64Size : Header32Size);
283 /// WriteSegmentLoadCommand - Write a segment load command.
285 /// \arg NumSections - The number of sections in this segment.
286 /// \arg SectionDataSize - The total size of the sections.
287 void WriteSegmentLoadCommand(unsigned NumSections,
289 uint64_t SectionDataStartOffset,
290 uint64_t SectionDataSize) {
291 // struct segment_command (56 bytes) or
292 // struct segment_command_64 (72 bytes)
294 uint64_t Start = OS.tell();
297 unsigned SegmentLoadCommandSize = Is64Bit ? SegmentLoadCommand64Size :
298 SegmentLoadCommand32Size;
299 Write32(Is64Bit ? LCT_Segment64 : LCT_Segment);
300 Write32(SegmentLoadCommandSize +
301 NumSections * (Is64Bit ? Section64Size : Section32Size));
305 Write64(0); // vmaddr
306 Write64(VMSize); // vmsize
307 Write64(SectionDataStartOffset); // file offset
308 Write64(SectionDataSize); // file size
310 Write32(0); // vmaddr
311 Write32(VMSize); // vmsize
312 Write32(SectionDataStartOffset); // file offset
313 Write32(SectionDataSize); // file size
315 Write32(0x7); // maxprot
316 Write32(0x7); // initprot
317 Write32(NumSections);
320 assert(OS.tell() - Start == SegmentLoadCommandSize);
323 void WriteSection(const MCSectionData &SD, uint64_t FileOffset,
324 uint64_t RelocationsStart, unsigned NumRelocations) {
325 // The offset is unused for virtual sections.
326 if (isVirtualSection(SD.getSection())) {
327 assert(SD.getFileSize() == 0 && "Invalid file size!");
331 // struct section (68 bytes) or
332 // struct section_64 (80 bytes)
334 uint64_t Start = OS.tell();
337 // FIXME: cast<> support!
338 const MCSectionMachO &Section =
339 static_cast<const MCSectionMachO&>(SD.getSection());
340 WriteString(Section.getSectionName(), 16);
341 WriteString(Section.getSegmentName(), 16);
343 Write64(SD.getAddress()); // address
344 Write64(SD.getSize()); // size
346 Write32(SD.getAddress()); // address
347 Write32(SD.getSize()); // size
351 unsigned Flags = Section.getTypeAndAttributes();
352 if (SD.hasInstructions())
353 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
355 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
356 Write32(Log2_32(SD.getAlignment()));
357 Write32(NumRelocations ? RelocationsStart : 0);
358 Write32(NumRelocations);
360 Write32(0); // reserved1
361 Write32(Section.getStubSize()); // reserved2
363 Write32(0); // reserved3
365 assert(OS.tell() - Start == Is64Bit ? Section64Size : Section32Size);
368 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
369 uint32_t StringTableOffset,
370 uint32_t StringTableSize) {
371 // struct symtab_command (24 bytes)
373 uint64_t Start = OS.tell();
377 Write32(SymtabLoadCommandSize);
378 Write32(SymbolOffset);
380 Write32(StringTableOffset);
381 Write32(StringTableSize);
383 assert(OS.tell() - Start == SymtabLoadCommandSize);
386 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
387 uint32_t NumLocalSymbols,
388 uint32_t FirstExternalSymbol,
389 uint32_t NumExternalSymbols,
390 uint32_t FirstUndefinedSymbol,
391 uint32_t NumUndefinedSymbols,
392 uint32_t IndirectSymbolOffset,
393 uint32_t NumIndirectSymbols) {
394 // struct dysymtab_command (80 bytes)
396 uint64_t Start = OS.tell();
399 Write32(LCT_Dysymtab);
400 Write32(DysymtabLoadCommandSize);
401 Write32(FirstLocalSymbol);
402 Write32(NumLocalSymbols);
403 Write32(FirstExternalSymbol);
404 Write32(NumExternalSymbols);
405 Write32(FirstUndefinedSymbol);
406 Write32(NumUndefinedSymbols);
407 Write32(0); // tocoff
409 Write32(0); // modtaboff
410 Write32(0); // nmodtab
411 Write32(0); // extrefsymoff
412 Write32(0); // nextrefsyms
413 Write32(IndirectSymbolOffset);
414 Write32(NumIndirectSymbols);
415 Write32(0); // extreloff
416 Write32(0); // nextrel
417 Write32(0); // locreloff
418 Write32(0); // nlocrel
420 assert(OS.tell() - Start == DysymtabLoadCommandSize);
423 void WriteNlist(MachSymbolData &MSD) {
424 MCSymbolData &Data = *MSD.SymbolData;
425 const MCSymbol &Symbol = Data.getSymbol();
427 uint16_t Flags = Data.getFlags();
428 uint32_t Address = 0;
430 // Set the N_TYPE bits. See <mach-o/nlist.h>.
432 // FIXME: Are the prebound or indirect fields possible here?
433 if (Symbol.isUndefined())
434 Type = STT_Undefined;
435 else if (Symbol.isAbsolute())
440 // FIXME: Set STAB bits.
442 if (Data.isPrivateExtern())
443 Type |= STF_PrivateExtern;
446 if (Data.isExternal() || Symbol.isUndefined())
447 Type |= STF_External;
449 // Compute the symbol address.
450 if (Symbol.isDefined()) {
451 if (Symbol.isAbsolute()) {
452 llvm_unreachable("FIXME: Not yet implemented!");
454 Address = Data.getAddress();
456 } else if (Data.isCommon()) {
457 // Common symbols are encoded with the size in the address
458 // field, and their alignment in the flags.
459 Address = Data.getCommonSize();
461 // Common alignment is packed into the 'desc' bits.
462 if (unsigned Align = Data.getCommonAlignment()) {
463 unsigned Log2Size = Log2_32(Align);
464 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
466 llvm_report_error("invalid 'common' alignment '" +
468 // FIXME: Keep this mask with the SymbolFlags enumeration.
469 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
473 // struct nlist (12 bytes)
475 Write32(MSD.StringIndex);
477 Write8(MSD.SectionIndex);
479 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
488 void RecordScatteredRelocation(MCAssembler &Asm, MCFragment &Fragment,
489 const MCAsmFixup &Fixup, MCValue Target,
490 uint64_t &FixedValue) {
491 uint32_t Address = Fragment.getOffset() + Fixup.Offset;
492 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
493 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
494 unsigned Type = RIT_Vanilla;
497 const MCSymbol *A = &Target.getSymA()->getSymbol();
498 MCSymbolData *A_SD = &Asm.getSymbolData(*A);
500 if (!A_SD->getFragment())
501 llvm_report_error("symbol '" + A->getName() +
502 "' can not be undefined in a subtraction expression");
504 uint32_t Value = A_SD->getAddress();
507 if (const MCSymbolRefExpr *B = Target.getSymB()) {
508 MCSymbolData *B_SD = &Asm.getSymbolData(B->getSymbol());
510 if (!B_SD->getFragment())
511 llvm_report_error("symbol '" + B->getSymbol().getName() +
512 "' can not be undefined in a subtraction expression");
514 // Select the appropriate difference relocation type.
516 // Note that there is no longer any semantic difference between these two
517 // relocation types from the linkers point of view, this is done solely
518 // for pedantic compatibility with 'as'.
519 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
520 Value2 = B_SD->getAddress();
523 // Relocations are written out in reverse order, so the PAIR comes first.
524 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
525 MachRelocationEntry MRE;
526 MRE.Word0 = ((0 << 0) |
532 Relocations[Fragment.getParent()].push_back(MRE);
535 MachRelocationEntry MRE;
536 MRE.Word0 = ((Address << 0) |
542 Relocations[Fragment.getParent()].push_back(MRE);
545 void RecordRelocation(MCAssembler &Asm, MCDataFragment &Fragment,
546 const MCAsmFixup &Fixup, MCValue Target,
547 uint64_t &FixedValue) {
548 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
549 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
551 // If this is a difference or a defined symbol plus an offset, then we need
552 // a scattered relocation entry.
553 uint32_t Offset = Target.getConstant();
555 Offset += 1 << Log2Size;
556 if (Target.getSymB() ||
557 (Target.getSymA() && !Target.getSymA()->getSymbol().isUndefined() &&
559 RecordScatteredRelocation(Asm, Fragment, Fixup, Target, FixedValue);
564 uint32_t Address = Fragment.getOffset() + Fixup.Offset;
567 unsigned IsExtern = 0;
570 if (Target.isAbsolute()) { // constant
571 // SymbolNum of 0 indicates the absolute section.
573 // FIXME: Currently, these are never generated (see code below). I cannot
574 // find a case where they are actually emitted.
578 const MCSymbol *Symbol = &Target.getSymA()->getSymbol();
579 MCSymbolData *SD = &Asm.getSymbolData(*Symbol);
581 if (Symbol->isUndefined()) {
583 Index = SD->getIndex();
586 // The index is the section ordinal.
590 MCAssembler::iterator it = Asm.begin(), ie = Asm.end();
591 for (; it != ie; ++it, ++Index)
592 if (&*it == SD->getFragment()->getParent())
594 assert(it != ie && "Unable to find section index!");
595 Value = SD->getAddress();
601 // struct relocation_info (8 bytes)
602 MachRelocationEntry MRE;
604 MRE.Word1 = ((Index << 0) |
609 Relocations[Fragment.getParent()].push_back(MRE);
612 void ComputeRelocationInfo(MCAssembler &Asm, MCDataFragment &Fragment,
614 // FIXME: Share layout object.
615 MCAsmLayout Layout(Asm);
617 // Evaluate the fixup; if the value was resolved, no relocation is needed.
619 if (Asm.EvaluateFixup(Layout, Fixup, &Fragment, Target, Fixup.FixedValue))
622 RecordRelocation(Asm, Fragment, Fixup, Target, Fixup.FixedValue);
625 void BindIndirectSymbols(MCAssembler &Asm) {
626 // This is the point where 'as' creates actual symbols for indirect symbols
627 // (in the following two passes). It would be easier for us to do this
628 // sooner when we see the attribute, but that makes getting the order in the
629 // symbol table much more complicated than it is worth.
631 // FIXME: Revisit this when the dust settles.
633 // Bind non lazy symbol pointers first.
634 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
635 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
636 // FIXME: cast<> support!
637 const MCSectionMachO &Section =
638 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
640 if (Section.getType() != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
643 Asm.getOrCreateSymbolData(*it->Symbol);
646 // Then lazy symbol pointers and symbol stubs.
647 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
648 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
649 // FIXME: cast<> support!
650 const MCSectionMachO &Section =
651 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
653 if (Section.getType() != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
654 Section.getType() != MCSectionMachO::S_SYMBOL_STUBS)
657 // Set the symbol type to undefined lazy, but only on construction.
659 // FIXME: Do not hardcode.
661 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
663 Entry.setFlags(Entry.getFlags() | 0x0001);
667 /// ComputeSymbolTable - Compute the symbol table data
669 /// \param StringTable [out] - The string table data.
670 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
672 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
673 std::vector<MachSymbolData> &LocalSymbolData,
674 std::vector<MachSymbolData> &ExternalSymbolData,
675 std::vector<MachSymbolData> &UndefinedSymbolData) {
676 // Build section lookup table.
677 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
679 for (MCAssembler::iterator it = Asm.begin(),
680 ie = Asm.end(); it != ie; ++it, ++Index)
681 SectionIndexMap[&it->getSection()] = Index;
682 assert(Index <= 256 && "Too many sections!");
684 // Index 0 is always the empty string.
685 StringMap<uint64_t> StringIndexMap;
686 StringTable += '\x00';
688 // Build the symbol arrays and the string table, but only for non-local
691 // The particular order that we collect the symbols and create the string
692 // table, then sort the symbols is chosen to match 'as'. Even though it
693 // doesn't matter for correctness, this is important for letting us diff .o
695 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
696 ie = Asm.symbol_end(); it != ie; ++it) {
697 const MCSymbol &Symbol = it->getSymbol();
699 // Ignore non-linker visible symbols.
700 if (!Asm.isSymbolLinkerVisible(it))
703 if (!it->isExternal() && !Symbol.isUndefined())
706 uint64_t &Entry = StringIndexMap[Symbol.getName()];
708 Entry = StringTable.size();
709 StringTable += Symbol.getName();
710 StringTable += '\x00';
715 MSD.StringIndex = Entry;
717 if (Symbol.isUndefined()) {
718 MSD.SectionIndex = 0;
719 UndefinedSymbolData.push_back(MSD);
720 } else if (Symbol.isAbsolute()) {
721 MSD.SectionIndex = 0;
722 ExternalSymbolData.push_back(MSD);
724 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
725 assert(MSD.SectionIndex && "Invalid section index!");
726 ExternalSymbolData.push_back(MSD);
730 // Now add the data for local symbols.
731 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
732 ie = Asm.symbol_end(); it != ie; ++it) {
733 const MCSymbol &Symbol = it->getSymbol();
735 // Ignore non-linker visible symbols.
736 if (!Asm.isSymbolLinkerVisible(it))
739 if (it->isExternal() || Symbol.isUndefined())
742 uint64_t &Entry = StringIndexMap[Symbol.getName()];
744 Entry = StringTable.size();
745 StringTable += Symbol.getName();
746 StringTable += '\x00';
751 MSD.StringIndex = Entry;
753 if (Symbol.isAbsolute()) {
754 MSD.SectionIndex = 0;
755 LocalSymbolData.push_back(MSD);
757 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
758 assert(MSD.SectionIndex && "Invalid section index!");
759 LocalSymbolData.push_back(MSD);
763 // External and undefined symbols are required to be in lexicographic order.
764 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
765 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
767 // Set the symbol indices.
769 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
770 LocalSymbolData[i].SymbolData->setIndex(Index++);
771 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
772 ExternalSymbolData[i].SymbolData->setIndex(Index++);
773 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
774 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
776 // The string table is padded to a multiple of 4.
777 while (StringTable.size() % 4)
778 StringTable += '\x00';
781 void WriteObject(MCAssembler &Asm) {
782 unsigned NumSections = Asm.size();
784 // Create symbol data for any indirect symbols.
785 BindIndirectSymbols(Asm);
787 // Compute symbol table information.
788 SmallString<256> StringTable;
789 std::vector<MachSymbolData> LocalSymbolData;
790 std::vector<MachSymbolData> ExternalSymbolData;
791 std::vector<MachSymbolData> UndefinedSymbolData;
792 unsigned NumSymbols = Asm.symbol_size();
794 // No symbol table command is written if there are no symbols.
796 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
797 UndefinedSymbolData);
799 // The section data starts after the header, the segment load command (and
800 // section headers) and the symbol table.
801 unsigned NumLoadCommands = 1;
802 uint64_t LoadCommandsSize = Is64Bit ?
803 SegmentLoadCommand64Size + NumSections * Section64Size :
804 SegmentLoadCommand32Size + NumSections * Section32Size;
806 // Add the symbol table load command sizes, if used.
808 NumLoadCommands += 2;
809 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
812 // Compute the total size of the section data, as well as its file size and
814 uint64_t SectionDataStart = (Is64Bit ? Header64Size : Header32Size)
816 uint64_t SectionDataSize = 0;
817 uint64_t SectionDataFileSize = 0;
819 for (MCAssembler::iterator it = Asm.begin(),
820 ie = Asm.end(); it != ie; ++it) {
821 MCSectionData &SD = *it;
823 VMSize = std::max(VMSize, SD.getAddress() + SD.getSize());
825 if (isVirtualSection(SD.getSection()))
828 SectionDataSize = std::max(SectionDataSize,
829 SD.getAddress() + SD.getSize());
830 SectionDataFileSize = std::max(SectionDataFileSize,
831 SD.getAddress() + SD.getFileSize());
834 // The section data is padded to 4 bytes.
836 // FIXME: Is this machine dependent?
837 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
838 SectionDataFileSize += SectionDataPadding;
840 // Write the prolog, starting with the header and load command...
841 WriteHeader(NumLoadCommands, LoadCommandsSize,
842 Asm.getSubsectionsViaSymbols());
843 WriteSegmentLoadCommand(NumSections, VMSize,
844 SectionDataStart, SectionDataSize);
846 for (MCAssembler::iterator it = Asm.begin(),
847 ie = Asm.end(); it != ie; ++it) {
848 MCSectionData &SD = *it;
849 for (MCSectionData::iterator it2 = SD.begin(),
850 ie2 = SD.end(); it2 != ie2; ++it2)
851 if (MCDataFragment *DF = dyn_cast<MCDataFragment>(&*it2))
852 for (unsigned i = 0, e = DF->fixup_size(); i != e; ++i)
853 ComputeRelocationInfo(Asm, *DF, DF->getFixups()[i]);
856 // ... and then the section headers.
857 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
858 for (MCAssembler::iterator it = Asm.begin(),
859 ie = Asm.end(); it != ie; ++it) {
860 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
861 unsigned NumRelocs = Relocs.size();
862 uint64_t SectionStart = SectionDataStart + it->getAddress();
863 WriteSection(*it, SectionStart, RelocTableEnd, NumRelocs);
864 RelocTableEnd += NumRelocs * RelocationInfoSize;
867 // Write the symbol table load command, if used.
869 unsigned FirstLocalSymbol = 0;
870 unsigned NumLocalSymbols = LocalSymbolData.size();
871 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
872 unsigned NumExternalSymbols = ExternalSymbolData.size();
873 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
874 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
875 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
876 unsigned NumSymTabSymbols =
877 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
878 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
879 uint64_t IndirectSymbolOffset = 0;
881 // If used, the indirect symbols are written after the section data.
882 if (NumIndirectSymbols)
883 IndirectSymbolOffset = RelocTableEnd;
885 // The symbol table is written after the indirect symbol data.
886 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
888 // The string table is written after symbol table.
889 uint64_t StringTableOffset =
890 SymbolTableOffset + NumSymTabSymbols * (Is64Bit ? Nlist64Size :
892 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
893 StringTableOffset, StringTable.size());
895 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
896 FirstExternalSymbol, NumExternalSymbols,
897 FirstUndefinedSymbol, NumUndefinedSymbols,
898 IndirectSymbolOffset, NumIndirectSymbols);
901 // Write the actual section data.
902 for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it)
903 WriteFileData(OS, *it, *this);
905 // Write the extra padding.
906 WriteZeros(SectionDataPadding);
908 // Write the relocation entries.
909 for (MCAssembler::iterator it = Asm.begin(),
910 ie = Asm.end(); it != ie; ++it) {
911 // Write the section relocation entries, in reverse order to match 'as'
912 // (approximately, the exact algorithm is more complicated than this).
913 std::vector<MachRelocationEntry> &Relocs = Relocations[it];
914 for (unsigned i = 0, e = Relocs.size(); i != e; ++i) {
915 Write32(Relocs[e - i - 1].Word0);
916 Write32(Relocs[e - i - 1].Word1);
920 // Write the symbol table data, if used.
922 // Write the indirect symbol entries.
923 for (MCAssembler::indirect_symbol_iterator
924 it = Asm.indirect_symbol_begin(),
925 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
926 // Indirect symbols in the non lazy symbol pointer section have some
928 const MCSectionMachO &Section =
929 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
930 if (Section.getType() == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
931 // If this symbol is defined and internal, mark it as such.
932 if (it->Symbol->isDefined() &&
933 !Asm.getSymbolData(*it->Symbol).isExternal()) {
934 uint32_t Flags = ISF_Local;
935 if (it->Symbol->isAbsolute())
936 Flags |= ISF_Absolute;
942 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
945 // FIXME: Check that offsets match computed ones.
947 // Write the symbol table entries.
948 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
949 WriteNlist(LocalSymbolData[i]);
950 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
951 WriteNlist(ExternalSymbolData[i]);
952 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
953 WriteNlist(UndefinedSymbolData[i]);
955 // Write the string table.
956 OS << StringTable.str();
960 void ApplyFixup(const MCAsmFixup &Fixup, MCDataFragment &DF) {
961 unsigned Size = 1 << getFixupKindLog2Size(Fixup.Kind);
963 // FIXME: Endianness assumption.
964 assert(Fixup.Offset + Size <= DF.getContents().size() &&
965 "Invalid fixup offset!");
966 for (unsigned i = 0; i != Size; ++i)
967 DF.getContents()[Fixup.Offset + i] = uint8_t(Fixup.FixedValue >> (i * 8));
973 MCFragment::MCFragment() : Kind(FragmentType(~0)) {
976 MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent)
979 FileSize(~UINT64_C(0))
982 Parent->getFragmentList().push_back(this);
985 MCFragment::~MCFragment() {
988 uint64_t MCFragment::getAddress() const {
989 assert(getParent() && "Missing Section!");
990 return getParent()->getAddress() + Offset;
995 MCSectionData::MCSectionData() : Section(0) {}
997 MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A)
998 : Section(&_Section),
1000 Address(~UINT64_C(0)),
1002 FileSize(~UINT64_C(0)),
1003 HasInstructions(false)
1006 A->getSectionList().push_back(this);
1011 MCSymbolData::MCSymbolData() : Symbol(0) {}
1013 MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment,
1014 uint64_t _Offset, MCAssembler *A)
1015 : Symbol(&_Symbol), Fragment(_Fragment), Offset(_Offset),
1016 IsExternal(false), IsPrivateExtern(false),
1017 CommonSize(0), CommonAlign(0), Flags(0), Index(0)
1020 A->getSymbolList().push_back(this);
1025 MCAssembler::MCAssembler(MCContext &_Context, TargetAsmBackend &_Backend,
1027 : Context(_Context), Backend(_Backend), OS(_OS), SubsectionsViaSymbols(false)
1031 MCAssembler::~MCAssembler() {
1034 static bool isScatteredFixupFullyResolvedSimple(const MCAssembler &Asm,
1035 const MCAsmFixup &Fixup,
1036 const MCDataFragment *DF,
1037 const MCValue Target,
1038 const MCSection *BaseSection) {
1039 // The effective fixup address is
1040 // addr(atom(A)) + offset(A)
1041 // - addr(atom(B)) - offset(B)
1042 // - addr(<base symbol>) + <fixup offset from base symbol>
1043 // and the offsets are not relocatable, so the fixup is fully resolved when
1044 // addr(atom(A)) - addr(atom(B)) - addr(<base symbol>)) == 0.
1046 // The simple (Darwin, except on x86_64) way of dealing with this was to
1047 // assume that any reference to a temporary symbol *must* be a temporary
1048 // symbol in the same atom, unless the sections differ. Therefore, any PCrel
1049 // relocation to a temporary symbol (in the same section) is fully
1050 // resolved. This also works in conjunction with absolutized .set, which
1051 // requires the compiler to use .set to absolutize the differences between
1052 // symbols which the compiler knows to be assembly time constants, so we don't
1053 // need to worry about consider symbol differences fully resolved.
1055 // Non-relative fixups are only resolved if constant.
1057 return Target.isAbsolute();
1059 // Otherwise, relative fixups are only resolved if not a difference and the
1060 // target is a temporary in the same section.
1061 if (Target.isAbsolute() || Target.getSymB())
1064 const MCSymbol *A = &Target.getSymA()->getSymbol();
1065 if (!A->isTemporary() || !A->isInSection() ||
1066 &A->getSection() != BaseSection)
1072 static bool isScatteredFixupFullyResolved(const MCAssembler &Asm,
1073 const MCAsmFixup &Fixup,
1074 const MCDataFragment *DF,
1075 const MCValue Target,
1076 const MCSymbolData *BaseSymbol) {
1077 // The effective fixup address is
1078 // addr(atom(A)) + offset(A)
1079 // - addr(atom(B)) - offset(B)
1080 // - addr(BaseSymbol) + <fixup offset from base symbol>
1081 // and the offsets are not relocatable, so the fixup is fully resolved when
1082 // addr(atom(A)) - addr(atom(B)) - addr(BaseSymbol) == 0.
1084 // Note that "false" is almost always conservatively correct (it means we emit
1085 // a relocation which is unnecessary), except when it would force us to emit a
1086 // relocation which the target cannot encode.
1088 const MCSymbolData *A_Base = 0, *B_Base = 0;
1089 if (const MCSymbolRefExpr *A = Target.getSymA()) {
1090 // Modified symbol references cannot be resolved.
1091 if (A->getKind() != MCSymbolRefExpr::VK_None)
1094 A_Base = Asm.getAtom(&Asm.getSymbolData(A->getSymbol()));
1099 if (const MCSymbolRefExpr *B = Target.getSymB()) {
1100 // Modified symbol references cannot be resolved.
1101 if (B->getKind() != MCSymbolRefExpr::VK_None)
1104 B_Base = Asm.getAtom(&Asm.getSymbolData(B->getSymbol()));
1109 // If there is no base, A and B have to be the same atom for this fixup to be
1112 return A_Base == B_Base;
1114 // Otherwise, B must be missing and A must be the base.
1115 return !B_Base && BaseSymbol == A_Base;
1118 bool MCAssembler::isSymbolLinkerVisible(const MCSymbolData *SD) const {
1119 // Non-temporary labels should always be visible to the linker.
1120 if (!SD->getSymbol().isTemporary())
1123 // Absolute temporary labels are never visible.
1124 if (!SD->getFragment())
1127 // Otherwise, check if the section requires symbols even for temporary labels.
1128 return getBackend().doesSectionRequireSymbols(
1129 SD->getFragment()->getParent()->getSection());
1132 const MCSymbolData *MCAssembler::getAtomForAddress(const MCSectionData *Section,
1133 uint64_t Address) const {
1134 const MCSymbolData *Best = 0;
1135 for (MCAssembler::const_symbol_iterator it = symbol_begin(),
1136 ie = symbol_end(); it != ie; ++it) {
1137 // Ignore non-linker visible symbols.
1138 if (!isSymbolLinkerVisible(it))
1141 // Ignore symbols not in the same section.
1142 if (!it->getFragment() || it->getFragment()->getParent() != Section)
1145 // Otherwise, find the closest symbol preceding this address (ties are
1146 // resolved in favor of the last defined symbol).
1147 if (it->getAddress() <= Address &&
1148 (!Best || it->getAddress() >= Best->getAddress()))
1155 const MCSymbolData *MCAssembler::getAtom(const MCSymbolData *SD) const {
1156 // Linker visible symbols define atoms.
1157 if (isSymbolLinkerVisible(SD))
1160 // Absolute and undefined symbols have no defining atom.
1161 if (!SD->getFragment())
1164 // Otherwise, search by address.
1165 return getAtomForAddress(SD->getFragment()->getParent(), SD->getAddress());
1168 bool MCAssembler::EvaluateFixup(const MCAsmLayout &Layout, MCAsmFixup &Fixup,
1170 MCValue &Target, uint64_t &Value) const {
1171 if (!Fixup.Value->EvaluateAsRelocatable(Target, &Layout))
1172 llvm_report_error("expected relocatable expression");
1174 // FIXME: How do non-scattered symbols work in ELF? I presume the linker
1175 // doesn't support small relocations, but then under what criteria does the
1176 // assembler allow symbol differences?
1178 Value = Target.getConstant();
1180 bool IsResolved = true, IsPCRel = isFixupKindPCRel(Fixup.Kind);
1181 if (const MCSymbolRefExpr *A = Target.getSymA()) {
1182 if (A->getSymbol().isDefined())
1183 Value += getSymbolData(A->getSymbol()).getAddress();
1187 if (const MCSymbolRefExpr *B = Target.getSymB()) {
1188 if (B->getSymbol().isDefined())
1189 Value -= getSymbolData(B->getSymbol()).getAddress();
1194 // If we are using scattered symbols, determine whether this value is actually
1195 // resolved; scattering may cause atoms to move.
1196 if (IsResolved && getBackend().hasScatteredSymbols()) {
1197 if (getBackend().hasReliableSymbolDifference()) {
1198 // If this is a PCrel relocation, find the base atom (identified by its
1199 // symbol) that the fixup value is relative to.
1200 const MCSymbolData *BaseSymbol = 0;
1202 BaseSymbol = getAtomForAddress(
1203 DF->getParent(), DF->getAddress() + Fixup.Offset);
1209 IsResolved = isScatteredFixupFullyResolved(*this, Fixup, DF, Target,
1212 const MCSection *BaseSection = 0;
1214 BaseSection = &DF->getParent()->getSection();
1216 IsResolved = isScatteredFixupFullyResolvedSimple(*this, Fixup, DF, Target,
1222 Value -= DF->getAddress() + Fixup.Offset;
1227 void MCAssembler::LayoutSection(MCSectionData &SD) {
1228 MCAsmLayout Layout(*this);
1229 uint64_t Address = SD.getAddress();
1231 for (MCSectionData::iterator it = SD.begin(), ie = SD.end(); it != ie; ++it) {
1232 MCFragment &F = *it;
1234 F.setOffset(Address - SD.getAddress());
1236 // Evaluate fragment size.
1237 switch (F.getKind()) {
1238 case MCFragment::FT_Align: {
1239 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1241 uint64_t Size = OffsetToAlignment(Address, AF.getAlignment());
1242 if (Size > AF.getMaxBytesToEmit())
1245 AF.setFileSize(Size);
1249 case MCFragment::FT_Data:
1250 case MCFragment::FT_Fill:
1251 F.setFileSize(F.getMaxFileSize());
1254 case MCFragment::FT_Org: {
1255 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1257 int64_t TargetLocation;
1258 if (!OF.getOffset().EvaluateAsAbsolute(TargetLocation, &Layout))
1259 llvm_report_error("expected assembly-time absolute expression");
1261 // FIXME: We need a way to communicate this error.
1262 int64_t Offset = TargetLocation - F.getOffset();
1264 llvm_report_error("invalid .org offset '" + Twine(TargetLocation) +
1265 "' (at offset '" + Twine(F.getOffset()) + "'");
1267 F.setFileSize(Offset);
1271 case MCFragment::FT_ZeroFill: {
1272 MCZeroFillFragment &ZFF = cast<MCZeroFillFragment>(F);
1274 // Align the fragment offset; it is safe to adjust the offset freely since
1275 // this is only in virtual sections.
1276 Address = RoundUpToAlignment(Address, ZFF.getAlignment());
1277 F.setOffset(Address - SD.getAddress());
1279 // FIXME: This is misnamed.
1280 F.setFileSize(ZFF.getSize());
1285 Address += F.getFileSize();
1288 // Set the section sizes.
1289 SD.setSize(Address - SD.getAddress());
1290 if (isVirtualSection(SD.getSection()))
1293 SD.setFileSize(Address - SD.getAddress());
1296 /// WriteNopData - Write optimal nops to the output file for the \arg Count
1297 /// bytes. This returns the number of bytes written. It may return 0 if
1298 /// the \arg Count is more than the maximum optimal nops.
1300 /// FIXME this is X86 32-bit specific and should move to a better place.
1301 static uint64_t WriteNopData(uint64_t Count, MachObjectWriter &MOW) {
1302 static const uint8_t Nops[16][16] = {
1310 {0x0f, 0x1f, 0x40, 0x00},
1311 // nopl 0(%[re]ax,%[re]ax,1)
1312 {0x0f, 0x1f, 0x44, 0x00, 0x00},
1313 // nopw 0(%[re]ax,%[re]ax,1)
1314 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1316 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1317 // nopl 0L(%[re]ax,%[re]ax,1)
1318 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1319 // nopw 0L(%[re]ax,%[re]ax,1)
1320 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1321 // nopw %cs:0L(%[re]ax,%[re]ax,1)
1322 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1323 // nopl 0(%[re]ax,%[re]ax,1)
1324 // nopw 0(%[re]ax,%[re]ax,1)
1325 {0x0f, 0x1f, 0x44, 0x00, 0x00,
1326 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1327 // nopw 0(%[re]ax,%[re]ax,1)
1328 // nopw 0(%[re]ax,%[re]ax,1)
1329 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1330 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1331 // nopw 0(%[re]ax,%[re]ax,1)
1332 // nopl 0L(%[re]ax) */
1333 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1334 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1337 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1338 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1340 // nopl 0L(%[re]ax,%[re]ax,1)
1341 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1342 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}
1348 for (uint64_t i = 0; i < Count; i++)
1349 MOW.Write8 (uint8_t(Nops[Count - 1][i]));
1354 /// WriteFileData - Write the \arg F data to the output file.
1355 static void WriteFileData(raw_ostream &OS, const MCFragment &F,
1356 MachObjectWriter &MOW) {
1357 uint64_t Start = OS.tell();
1362 // FIXME: Embed in fragments instead?
1363 switch (F.getKind()) {
1364 case MCFragment::FT_Align: {
1365 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1366 uint64_t Count = AF.getFileSize() / AF.getValueSize();
1368 // FIXME: This error shouldn't actually occur (the front end should emit
1369 // multiple .align directives to enforce the semantics it wants), but is
1370 // severe enough that we want to report it. How to handle this?
1371 if (Count * AF.getValueSize() != AF.getFileSize())
1372 llvm_report_error("undefined .align directive, value size '" +
1373 Twine(AF.getValueSize()) +
1374 "' is not a divisor of padding size '" +
1375 Twine(AF.getFileSize()) + "'");
1377 // See if we are aligning with nops, and if so do that first to try to fill
1378 // the Count bytes. Then if that did not fill any bytes or there are any
1379 // bytes left to fill use the the Value and ValueSize to fill the rest.
1380 if (AF.getEmitNops()) {
1381 uint64_t NopByteCount = WriteNopData(Count, MOW);
1382 Count -= NopByteCount;
1385 for (uint64_t i = 0; i != Count; ++i) {
1386 switch (AF.getValueSize()) {
1388 assert(0 && "Invalid size!");
1389 case 1: MOW.Write8 (uint8_t (AF.getValue())); break;
1390 case 2: MOW.Write16(uint16_t(AF.getValue())); break;
1391 case 4: MOW.Write32(uint32_t(AF.getValue())); break;
1392 case 8: MOW.Write64(uint64_t(AF.getValue())); break;
1398 case MCFragment::FT_Data: {
1399 MCDataFragment &DF = cast<MCDataFragment>(F);
1401 // Apply the fixups.
1403 // FIXME: Move elsewhere.
1404 for (MCDataFragment::const_fixup_iterator it = DF.fixup_begin(),
1405 ie = DF.fixup_end(); it != ie; ++it)
1406 MOW.ApplyFixup(*it, DF);
1408 OS << cast<MCDataFragment>(F).getContents().str();
1412 case MCFragment::FT_Fill: {
1413 MCFillFragment &FF = cast<MCFillFragment>(F);
1414 for (uint64_t i = 0, e = FF.getCount(); i != e; ++i) {
1415 switch (FF.getValueSize()) {
1417 assert(0 && "Invalid size!");
1418 case 1: MOW.Write8 (uint8_t (FF.getValue())); break;
1419 case 2: MOW.Write16(uint16_t(FF.getValue())); break;
1420 case 4: MOW.Write32(uint32_t(FF.getValue())); break;
1421 case 8: MOW.Write64(uint64_t(FF.getValue())); break;
1427 case MCFragment::FT_Org: {
1428 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1430 for (uint64_t i = 0, e = OF.getFileSize(); i != e; ++i)
1431 MOW.Write8(uint8_t(OF.getValue()));
1436 case MCFragment::FT_ZeroFill: {
1437 assert(0 && "Invalid zero fill fragment in concrete section!");
1442 assert(OS.tell() - Start == F.getFileSize());
1445 /// WriteFileData - Write the \arg SD data to the output file.
1446 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
1447 MachObjectWriter &MOW) {
1448 // Ignore virtual sections.
1449 if (isVirtualSection(SD.getSection())) {
1450 assert(SD.getFileSize() == 0);
1454 uint64_t Start = OS.tell();
1457 for (MCSectionData::const_iterator it = SD.begin(),
1458 ie = SD.end(); it != ie; ++it)
1459 WriteFileData(OS, *it, MOW);
1461 // Add section padding.
1462 assert(SD.getFileSize() >= SD.getSize() && "Invalid section sizes!");
1463 MOW.WriteZeros(SD.getFileSize() - SD.getSize());
1465 assert(OS.tell() - Start == SD.getFileSize());
1468 void MCAssembler::Finish() {
1469 DEBUG_WITH_TYPE("mc-dump", {
1470 llvm::errs() << "assembler backend - pre-layout\n--\n";
1473 // Layout until everything fits.
1474 while (LayoutOnce())
1477 DEBUG_WITH_TYPE("mc-dump", {
1478 llvm::errs() << "assembler backend - post-layout\n--\n";
1481 // Write the object file.
1483 // FIXME: Factor out MCObjectWriter.
1484 bool Is64Bit = StringRef(getBackend().getTarget().getName()) == "x86-64";
1485 MachObjectWriter MOW(OS, Is64Bit);
1486 MOW.WriteObject(*this);
1491 bool MCAssembler::FixupNeedsRelaxation(MCAsmFixup &Fixup, MCDataFragment *DF) {
1492 // FIXME: Share layout object.
1493 MCAsmLayout Layout(*this);
1495 // Currently we only need to relax X86::reloc_pcrel_1byte.
1496 if (unsigned(Fixup.Kind) != X86::reloc_pcrel_1byte)
1499 // If we cannot resolve the fixup value, it requires relaxation.
1502 if (!EvaluateFixup(Layout, Fixup, DF, Target, Value))
1505 // Otherwise, relax if the value is too big for a (signed) i8.
1506 return int64_t(Value) != int64_t(int8_t(Value));
1509 bool MCAssembler::LayoutOnce() {
1510 // Layout the concrete sections and fragments.
1511 uint64_t Address = 0;
1512 MCSectionData *Prev = 0;
1513 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1514 MCSectionData &SD = *it;
1516 // Skip virtual sections.
1517 if (isVirtualSection(SD.getSection()))
1520 // Align this section if necessary by adding padding bytes to the previous
1522 if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment())) {
1523 assert(Prev && "Missing prev section!");
1524 Prev->setFileSize(Prev->getFileSize() + Pad);
1528 // Layout the section fragments and its size.
1529 SD.setAddress(Address);
1531 Address += SD.getFileSize();
1536 // Layout the virtual sections.
1537 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1538 MCSectionData &SD = *it;
1540 if (!isVirtualSection(SD.getSection()))
1543 // Align this section if necessary by adding padding bytes to the previous
1545 if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment()))
1548 SD.setAddress(Address);
1550 Address += SD.getSize();
1553 // Scan the fixups in order and relax any that don't fit.
1554 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1555 MCSectionData &SD = *it;
1557 for (MCSectionData::iterator it2 = SD.begin(),
1558 ie2 = SD.end(); it2 != ie2; ++it2) {
1559 MCDataFragment *DF = dyn_cast<MCDataFragment>(it2);
1563 for (MCDataFragment::fixup_iterator it3 = DF->fixup_begin(),
1564 ie3 = DF->fixup_end(); it3 != ie3; ++it3) {
1565 MCAsmFixup &Fixup = *it3;
1567 // Check whether we need to relax this fixup.
1568 if (!FixupNeedsRelaxation(Fixup, DF))
1571 // Relax the instruction.
1573 // FIXME: This is a huge temporary hack which just looks for x86
1574 // branches; the only thing we need to relax on x86 is
1575 // 'X86::reloc_pcrel_1byte'. Once we have MCInst fragments, this will be
1576 // replaced by a TargetAsmBackend hook (most likely tblgen'd) to relax
1577 // an individual MCInst.
1578 SmallVectorImpl<char> &C = DF->getContents();
1579 uint64_t PrevOffset = Fixup.Offset;
1583 if (unsigned(C[Fixup.Offset-1]) >= 0x70 &&
1584 unsigned(C[Fixup.Offset-1]) <= 0x7f) {
1585 C[Fixup.Offset] = C[Fixup.Offset-1] + 0x10;
1586 C[Fixup.Offset-1] = char(0x0f);
1591 } else if (C[Fixup.Offset-1] == char(0xeb)) {
1592 C[Fixup.Offset-1] = char(0xe9);
1596 llvm_unreachable("unknown 1 byte pcrel instruction!");
1598 Fixup.Value = MCBinaryExpr::Create(
1599 MCBinaryExpr::Sub, Fixup.Value,
1600 MCConstantExpr::Create(3, getContext()),
1602 C.insert(C.begin() + Fixup.Offset, Amt, char(0));
1603 Fixup.Kind = MCFixupKind(X86::reloc_pcrel_4byte);
1605 // Update the remaining fixups, which have slid.
1607 // FIXME: This is bad for performance, but will be eliminated by the
1608 // move to MCInst specific fragments.
1610 for (; it3 != ie3; ++it3)
1613 // Update all the symbols for this fragment, which may have slid.
1615 // FIXME: This is really really bad for performance, but will be
1616 // eliminated by the move to MCInst specific fragments.
1617 for (MCAssembler::symbol_iterator it = symbol_begin(),
1618 ie = symbol_end(); it != ie; ++it) {
1619 MCSymbolData &SD = *it;
1621 if (it->getFragment() != DF)
1624 if (SD.getOffset() > PrevOffset)
1625 SD.setOffset(SD.getOffset() + Amt);
1630 // FIXME: This is O(N^2), but will be eliminated once we have a smart
1631 // MCAsmLayout object.
1640 // Debugging methods
1644 raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) {
1645 OS << "<MCAsmFixup" << " Offset:" << AF.Offset << " Value:" << *AF.Value
1646 << " Kind:" << AF.Kind << ">";
1652 void MCFragment::dump() {
1653 raw_ostream &OS = llvm::errs();
1655 OS << "<MCFragment " << (void*) this << " Offset:" << Offset
1656 << " FileSize:" << FileSize;
1661 void MCAlignFragment::dump() {
1662 raw_ostream &OS = llvm::errs();
1664 OS << "<MCAlignFragment ";
1665 this->MCFragment::dump();
1667 OS << " Alignment:" << getAlignment()
1668 << " Value:" << getValue() << " ValueSize:" << getValueSize()
1669 << " MaxBytesToEmit:" << getMaxBytesToEmit() << ">";
1672 void MCDataFragment::dump() {
1673 raw_ostream &OS = llvm::errs();
1675 OS << "<MCDataFragment ";
1676 this->MCFragment::dump();
1678 OS << " Contents:[";
1679 for (unsigned i = 0, e = getContents().size(); i != e; ++i) {
1681 OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF);
1683 OS << "] (" << getContents().size() << " bytes)";
1685 if (!getFixups().empty()) {
1688 for (fixup_iterator it = fixup_begin(), ie = fixup_end(); it != ie; ++it) {
1689 if (it != fixup_begin()) OS << ",\n ";
1698 void MCFillFragment::dump() {
1699 raw_ostream &OS = llvm::errs();
1701 OS << "<MCFillFragment ";
1702 this->MCFragment::dump();
1704 OS << " Value:" << getValue() << " ValueSize:" << getValueSize()
1705 << " Count:" << getCount() << ">";
1708 void MCOrgFragment::dump() {
1709 raw_ostream &OS = llvm::errs();
1711 OS << "<MCOrgFragment ";
1712 this->MCFragment::dump();
1714 OS << " Offset:" << getOffset() << " Value:" << getValue() << ">";
1717 void MCZeroFillFragment::dump() {
1718 raw_ostream &OS = llvm::errs();
1720 OS << "<MCZeroFillFragment ";
1721 this->MCFragment::dump();
1723 OS << " Size:" << getSize() << " Alignment:" << getAlignment() << ">";
1726 void MCSectionData::dump() {
1727 raw_ostream &OS = llvm::errs();
1729 OS << "<MCSectionData";
1730 OS << " Alignment:" << getAlignment() << " Address:" << Address
1731 << " Size:" << Size << " FileSize:" << FileSize
1732 << " Fragments:[\n ";
1733 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1734 if (it != begin()) OS << ",\n ";
1740 void MCSymbolData::dump() {
1741 raw_ostream &OS = llvm::errs();
1743 OS << "<MCSymbolData Symbol:" << getSymbol()
1744 << " Fragment:" << getFragment() << " Offset:" << getOffset()
1745 << " Flags:" << getFlags() << " Index:" << getIndex();
1747 OS << " (common, size:" << getCommonSize()
1748 << " align: " << getCommonAlignment() << ")";
1750 OS << " (external)";
1751 if (isPrivateExtern())
1752 OS << " (private extern)";
1756 void MCAssembler::dump() {
1757 raw_ostream &OS = llvm::errs();
1759 OS << "<MCAssembler\n";
1760 OS << " Sections:[\n ";
1761 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1762 if (it != begin()) OS << ",\n ";
1768 for (symbol_iterator it = symbol_begin(), ie = symbol_end(); it != ie; ++it) {
1769 if (it != symbol_begin()) OS << ",\n ";