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"
29 #include "../Target/X86/X86FixupKinds.h"
34 class MachObjectWriter;
36 STATISTIC(EmittedFragments, "Number of emitted assembler fragments");
38 // FIXME FIXME FIXME: There are number of places in this file where we convert
39 // what is a 64-bit assembler value used for computation into a value in the
40 // object file, which may truncate it. We should detect that truncation where
41 // invalid and report errors back.
43 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
44 MachObjectWriter &MOW);
46 static uint64_t WriteNopData(uint64_t Count, MachObjectWriter &MOW);
48 /// isVirtualSection - Check if this is a section which does not actually exist
49 /// in the object file.
50 static bool isVirtualSection(const MCSection &Section) {
52 const MCSectionMachO &SMO = static_cast<const MCSectionMachO&>(Section);
53 unsigned Type = SMO.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
54 return (Type == MCSectionMachO::S_ZEROFILL);
57 static unsigned getFixupKindLog2Size(unsigned Kind) {
59 default: llvm_unreachable("invalid fixup kind!");
60 case X86::reloc_pcrel_1byte:
61 case FK_Data_1: return 0;
62 case FK_Data_2: return 1;
63 case X86::reloc_pcrel_4byte:
64 case X86::reloc_riprel_4byte:
65 case FK_Data_4: return 2;
66 case FK_Data_8: return 3;
70 static bool isFixupKindPCRel(unsigned Kind) {
74 case X86::reloc_pcrel_1byte:
75 case X86::reloc_pcrel_4byte:
76 case X86::reloc_riprel_4byte:
81 class MachObjectWriter {
82 // See <mach-o/loader.h>.
84 Header_Magic32 = 0xFEEDFACE,
85 Header_Magic64 = 0xFEEDFACF
88 static const unsigned Header32Size = 28;
89 static const unsigned Header64Size = 32;
90 static const unsigned SegmentLoadCommand32Size = 56;
91 static const unsigned Section32Size = 68;
92 static const unsigned SymtabLoadCommandSize = 24;
93 static const unsigned DysymtabLoadCommandSize = 80;
94 static const unsigned Nlist32Size = 12;
95 static const unsigned RelocationInfoSize = 8;
102 HF_SubsectionsViaSymbols = 0x2000
105 enum LoadCommandType {
111 // See <mach-o/nlist.h>.
112 enum SymbolTypeType {
113 STT_Undefined = 0x00,
118 enum SymbolTypeFlags {
119 // If any of these bits are set, then the entry is a stab entry number (see
120 // <mach-o/stab.h>. Otherwise the other masks apply.
121 STF_StabsEntryMask = 0xe0,
125 STF_PrivateExtern = 0x10
128 /// IndirectSymbolFlags - Flags for encoding special values in the indirect
130 enum IndirectSymbolFlags {
131 ISF_Local = 0x80000000,
132 ISF_Absolute = 0x40000000
135 /// RelocationFlags - Special flags for addresses.
136 enum RelocationFlags {
137 RF_Scattered = 0x80000000
140 enum RelocationInfoType {
144 RIT_PreboundLazyPointer = 3,
145 RIT_LocalDifference = 4
148 /// MachSymbolData - Helper struct for containing some precomputed information
150 struct MachSymbolData {
151 MCSymbolData *SymbolData;
152 uint64_t StringIndex;
153 uint8_t SectionIndex;
155 // Support lexicographic sorting.
156 bool operator<(const MachSymbolData &RHS) const {
157 const std::string &Name = SymbolData->getSymbol().getName();
158 return Name < RHS.SymbolData->getSymbol().getName();
166 MachObjectWriter(raw_ostream &_OS, bool _IsLSB = true)
167 : OS(_OS), IsLSB(_IsLSB) {
170 /// @name Helper Methods
173 void Write8(uint8_t Value) {
177 void Write16(uint16_t Value) {
179 Write8(uint8_t(Value >> 0));
180 Write8(uint8_t(Value >> 8));
182 Write8(uint8_t(Value >> 8));
183 Write8(uint8_t(Value >> 0));
187 void Write32(uint32_t Value) {
189 Write16(uint16_t(Value >> 0));
190 Write16(uint16_t(Value >> 16));
192 Write16(uint16_t(Value >> 16));
193 Write16(uint16_t(Value >> 0));
197 void Write64(uint64_t Value) {
199 Write32(uint32_t(Value >> 0));
200 Write32(uint32_t(Value >> 32));
202 Write32(uint32_t(Value >> 32));
203 Write32(uint32_t(Value >> 0));
207 void WriteZeros(unsigned N) {
208 const char Zeros[16] = { 0 };
210 for (unsigned i = 0, e = N / 16; i != e; ++i)
211 OS << StringRef(Zeros, 16);
213 OS << StringRef(Zeros, N % 16);
216 void WriteString(StringRef Str, unsigned ZeroFillSize = 0) {
219 WriteZeros(ZeroFillSize - Str.size());
224 void WriteHeader32(unsigned NumLoadCommands, unsigned LoadCommandsSize,
225 bool SubsectionsViaSymbols) {
228 if (SubsectionsViaSymbols)
229 Flags |= HF_SubsectionsViaSymbols;
231 // struct mach_header (28 bytes)
233 uint64_t Start = OS.tell();
236 Write32(Header_Magic32);
238 // FIXME: Support cputype.
239 Write32(MachO::CPUTypeI386);
240 // FIXME: Support cpusubtype.
241 Write32(MachO::CPUSubType_I386_ALL);
243 Write32(NumLoadCommands); // Object files have a single load command, the
245 Write32(LoadCommandsSize);
248 assert(OS.tell() - Start == Header32Size);
251 /// WriteSegmentLoadCommand32 - Write a 32-bit segment load command.
253 /// \arg NumSections - The number of sections in this segment.
254 /// \arg SectionDataSize - The total size of the sections.
255 void WriteSegmentLoadCommand32(unsigned NumSections,
257 uint64_t SectionDataStartOffset,
258 uint64_t SectionDataSize) {
259 // struct segment_command (56 bytes)
261 uint64_t Start = OS.tell();
264 Write32(LCT_Segment);
265 Write32(SegmentLoadCommand32Size + NumSections * Section32Size);
268 Write32(0); // vmaddr
269 Write32(VMSize); // vmsize
270 Write32(SectionDataStartOffset); // file offset
271 Write32(SectionDataSize); // file size
272 Write32(0x7); // maxprot
273 Write32(0x7); // initprot
274 Write32(NumSections);
277 assert(OS.tell() - Start == SegmentLoadCommand32Size);
280 void WriteSection32(const MCSectionData &SD, uint64_t FileOffset,
281 uint64_t RelocationsStart, unsigned NumRelocations) {
282 // The offset is unused for virtual sections.
283 if (isVirtualSection(SD.getSection())) {
284 assert(SD.getFileSize() == 0 && "Invalid file size!");
288 // struct section (68 bytes)
290 uint64_t Start = OS.tell();
293 // FIXME: cast<> support!
294 const MCSectionMachO &Section =
295 static_cast<const MCSectionMachO&>(SD.getSection());
296 WriteString(Section.getSectionName(), 16);
297 WriteString(Section.getSegmentName(), 16);
298 Write32(SD.getAddress()); // address
299 Write32(SD.getSize()); // size
302 unsigned Flags = Section.getTypeAndAttributes();
303 if (SD.hasInstructions())
304 Flags |= MCSectionMachO::S_ATTR_SOME_INSTRUCTIONS;
306 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
307 Write32(Log2_32(SD.getAlignment()));
308 Write32(NumRelocations ? RelocationsStart : 0);
309 Write32(NumRelocations);
311 Write32(0); // reserved1
312 Write32(Section.getStubSize()); // reserved2
314 assert(OS.tell() - Start == Section32Size);
317 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
318 uint32_t StringTableOffset,
319 uint32_t StringTableSize) {
320 // struct symtab_command (24 bytes)
322 uint64_t Start = OS.tell();
326 Write32(SymtabLoadCommandSize);
327 Write32(SymbolOffset);
329 Write32(StringTableOffset);
330 Write32(StringTableSize);
332 assert(OS.tell() - Start == SymtabLoadCommandSize);
335 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
336 uint32_t NumLocalSymbols,
337 uint32_t FirstExternalSymbol,
338 uint32_t NumExternalSymbols,
339 uint32_t FirstUndefinedSymbol,
340 uint32_t NumUndefinedSymbols,
341 uint32_t IndirectSymbolOffset,
342 uint32_t NumIndirectSymbols) {
343 // struct dysymtab_command (80 bytes)
345 uint64_t Start = OS.tell();
348 Write32(LCT_Dysymtab);
349 Write32(DysymtabLoadCommandSize);
350 Write32(FirstLocalSymbol);
351 Write32(NumLocalSymbols);
352 Write32(FirstExternalSymbol);
353 Write32(NumExternalSymbols);
354 Write32(FirstUndefinedSymbol);
355 Write32(NumUndefinedSymbols);
356 Write32(0); // tocoff
358 Write32(0); // modtaboff
359 Write32(0); // nmodtab
360 Write32(0); // extrefsymoff
361 Write32(0); // nextrefsyms
362 Write32(IndirectSymbolOffset);
363 Write32(NumIndirectSymbols);
364 Write32(0); // extreloff
365 Write32(0); // nextrel
366 Write32(0); // locreloff
367 Write32(0); // nlocrel
369 assert(OS.tell() - Start == DysymtabLoadCommandSize);
372 void WriteNlist32(MachSymbolData &MSD) {
373 MCSymbolData &Data = *MSD.SymbolData;
374 const MCSymbol &Symbol = Data.getSymbol();
376 uint16_t Flags = Data.getFlags();
377 uint32_t Address = 0;
379 // Set the N_TYPE bits. See <mach-o/nlist.h>.
381 // FIXME: Are the prebound or indirect fields possible here?
382 if (Symbol.isUndefined())
383 Type = STT_Undefined;
384 else if (Symbol.isAbsolute())
389 // FIXME: Set STAB bits.
391 if (Data.isPrivateExtern())
392 Type |= STF_PrivateExtern;
395 if (Data.isExternal() || Symbol.isUndefined())
396 Type |= STF_External;
398 // Compute the symbol address.
399 if (Symbol.isDefined()) {
400 if (Symbol.isAbsolute()) {
401 llvm_unreachable("FIXME: Not yet implemented!");
403 Address = Data.getAddress();
405 } else if (Data.isCommon()) {
406 // Common symbols are encoded with the size in the address
407 // field, and their alignment in the flags.
408 Address = Data.getCommonSize();
410 // Common alignment is packed into the 'desc' bits.
411 if (unsigned Align = Data.getCommonAlignment()) {
412 unsigned Log2Size = Log2_32(Align);
413 assert((1U << Log2Size) == Align && "Invalid 'common' alignment!");
415 llvm_report_error("invalid 'common' alignment '" +
417 // FIXME: Keep this mask with the SymbolFlags enumeration.
418 Flags = (Flags & 0xF0FF) | (Log2Size << 8);
422 // struct nlist (12 bytes)
424 Write32(MSD.StringIndex);
426 Write8(MSD.SectionIndex);
428 // The Mach-O streamer uses the lowest 16-bits of the flags for the 'desc'
434 struct MachRelocationEntry {
438 void ComputeScatteredRelocationInfo(MCAssembler &Asm, MCFragment &Fragment,
440 const MCValue &Target,
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 = &Asm.getSymbolData(*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->getAddress();
458 if (const MCSymbol *B = Target.getSymB()) {
459 MCSymbolData *B_SD = &Asm.getSymbolData(*B);
461 if (!B_SD->getFragment())
462 llvm_report_error("symbol '" + B->getName() +
463 "' can not be undefined in a subtraction expression");
465 // Select the appropriate difference relocation type.
467 // Note that there is no longer any semantic difference between these two
468 // relocation types from the linkers point of view, this is done solely
469 // for pedantic compatibility with 'as'.
470 Type = A_SD->isExternal() ? RIT_Difference : RIT_LocalDifference;
471 Value2 = B_SD->getAddress();
474 // The value which goes in the fixup is current value of the expression.
475 Fixup.FixedValue = Value - Value2 + Target.getConstant();
477 Fixup.FixedValue -= Address;
479 // If this fixup is a vanilla PC relative relocation for a local label, we
480 // don't need a relocation.
482 // FIXME: Implement proper atom support.
483 if (IsPCRel && Target.getSymA() && Target.getSymA()->isTemporary() &&
487 MachRelocationEntry MRE;
488 MRE.Word0 = ((Address << 0) |
494 Relocs.push_back(MRE);
496 if (Type == RIT_Difference || Type == RIT_LocalDifference) {
497 MachRelocationEntry MRE;
498 MRE.Word0 = ((0 << 0) |
504 Relocs.push_back(MRE);
508 void ComputeRelocationInfo(MCAssembler &Asm, MCDataFragment &Fragment,
510 std::vector<MachRelocationEntry> &Relocs) {
511 unsigned IsPCRel = isFixupKindPCRel(Fixup.Kind);
512 unsigned Log2Size = getFixupKindLog2Size(Fixup.Kind);
514 // FIXME: Share layout object.
515 MCAsmLayout Layout(Asm);
518 if (!Fixup.Value->EvaluateAsRelocatable(Target, &Layout))
519 llvm_report_error("expected relocatable expression");
521 // If this is a difference or a defined symbol plus an offset, then we need
522 // a scattered relocation entry.
523 uint32_t Offset = Target.getConstant();
525 Offset += 1 << Log2Size;
526 if (Target.getSymB() ||
527 (Target.getSymA() && !Target.getSymA()->isUndefined() &&
529 return ComputeScatteredRelocationInfo(Asm, Fragment, Fixup, Target,
533 uint32_t Address = Fragment.getOffset() + Fixup.Offset;
536 unsigned IsExtern = 0;
539 if (Target.isAbsolute()) { // constant
540 // SymbolNum of 0 indicates the absolute section.
542 // FIXME: Currently, these are never generated (see code below). I cannot
543 // find a case where they are actually emitted.
547 const MCSymbol *Symbol = Target.getSymA();
548 MCSymbolData *SD = &Asm.getSymbolData(*Symbol);
550 if (Symbol->isUndefined()) {
552 Index = SD->getIndex();
555 // The index is the section ordinal.
559 MCAssembler::iterator it = Asm.begin(), ie = Asm.end();
560 for (; it != ie; ++it, ++Index)
561 if (&*it == SD->getFragment()->getParent())
563 assert(it != ie && "Unable to find section index!");
564 Value = SD->getAddress();
570 // The value which goes in the fixup is current value of the expression.
571 Fixup.FixedValue = Value + Target.getConstant();
573 Fixup.FixedValue -= Address;
575 // If the target evaluates to a constant, we don't need a relocation. This
576 // occurs with absolutized expressions which are not resolved to constants
577 // until after relaxation.
578 if (Target.isAbsolute())
581 // If this fixup is a vanilla PC relative relocation for a local label, we
582 // don't need a relocation.
584 // FIXME: Implement proper atom support.
585 if (IsPCRel && Target.getSymA() && Target.getSymA()->isTemporary())
588 // struct relocation_info (8 bytes)
589 MachRelocationEntry MRE;
591 MRE.Word1 = ((Index << 0) |
596 Relocs.push_back(MRE);
599 void BindIndirectSymbols(MCAssembler &Asm) {
600 // This is the point where 'as' creates actual symbols for indirect symbols
601 // (in the following two passes). It would be easier for us to do this
602 // sooner when we see the attribute, but that makes getting the order in the
603 // symbol table much more complicated than it is worth.
605 // FIXME: Revisit this when the dust settles.
607 // Bind non lazy symbol pointers first.
608 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
609 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
610 // FIXME: cast<> support!
611 const MCSectionMachO &Section =
612 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
615 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
616 if (Type != MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS)
619 Asm.getOrCreateSymbolData(*it->Symbol);
622 // Then lazy symbol pointers and symbol stubs.
623 for (MCAssembler::indirect_symbol_iterator it = Asm.indirect_symbol_begin(),
624 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
625 // FIXME: cast<> support!
626 const MCSectionMachO &Section =
627 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
630 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
631 if (Type != MCSectionMachO::S_LAZY_SYMBOL_POINTERS &&
632 Type != MCSectionMachO::S_SYMBOL_STUBS)
635 // Set the symbol type to undefined lazy, but only on construction.
637 // FIXME: Do not hardcode.
639 MCSymbolData &Entry = Asm.getOrCreateSymbolData(*it->Symbol, &Created);
641 Entry.setFlags(Entry.getFlags() | 0x0001);
645 /// ComputeSymbolTable - Compute the symbol table data
647 /// \param StringTable [out] - The string table data.
648 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
650 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
651 std::vector<MachSymbolData> &LocalSymbolData,
652 std::vector<MachSymbolData> &ExternalSymbolData,
653 std::vector<MachSymbolData> &UndefinedSymbolData) {
654 // Build section lookup table.
655 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
657 for (MCAssembler::iterator it = Asm.begin(),
658 ie = Asm.end(); it != ie; ++it, ++Index)
659 SectionIndexMap[&it->getSection()] = Index;
660 assert(Index <= 256 && "Too many sections!");
662 // Index 0 is always the empty string.
663 StringMap<uint64_t> StringIndexMap;
664 StringTable += '\x00';
666 // Build the symbol arrays and the string table, but only for non-local
669 // The particular order that we collect the symbols and create the string
670 // table, then sort the symbols is chosen to match 'as'. Even though it
671 // doesn't matter for correctness, this is important for letting us diff .o
673 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
674 ie = Asm.symbol_end(); it != ie; ++it) {
675 const MCSymbol &Symbol = it->getSymbol();
677 // Ignore assembler temporaries.
678 if (it->getSymbol().isTemporary())
681 if (!it->isExternal() && !Symbol.isUndefined())
684 uint64_t &Entry = StringIndexMap[Symbol.getName()];
686 Entry = StringTable.size();
687 StringTable += Symbol.getName();
688 StringTable += '\x00';
693 MSD.StringIndex = Entry;
695 if (Symbol.isUndefined()) {
696 MSD.SectionIndex = 0;
697 UndefinedSymbolData.push_back(MSD);
698 } else if (Symbol.isAbsolute()) {
699 MSD.SectionIndex = 0;
700 ExternalSymbolData.push_back(MSD);
702 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
703 assert(MSD.SectionIndex && "Invalid section index!");
704 ExternalSymbolData.push_back(MSD);
708 // Now add the data for local symbols.
709 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
710 ie = Asm.symbol_end(); it != ie; ++it) {
711 const MCSymbol &Symbol = it->getSymbol();
713 // Ignore assembler temporaries.
714 if (it->getSymbol().isTemporary())
717 if (it->isExternal() || Symbol.isUndefined())
720 uint64_t &Entry = StringIndexMap[Symbol.getName()];
722 Entry = StringTable.size();
723 StringTable += Symbol.getName();
724 StringTable += '\x00';
729 MSD.StringIndex = Entry;
731 if (Symbol.isAbsolute()) {
732 MSD.SectionIndex = 0;
733 LocalSymbolData.push_back(MSD);
735 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
736 assert(MSD.SectionIndex && "Invalid section index!");
737 LocalSymbolData.push_back(MSD);
741 // External and undefined symbols are required to be in lexicographic order.
742 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
743 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
745 // Set the symbol indices.
747 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
748 LocalSymbolData[i].SymbolData->setIndex(Index++);
749 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
750 ExternalSymbolData[i].SymbolData->setIndex(Index++);
751 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
752 UndefinedSymbolData[i].SymbolData->setIndex(Index++);
754 // The string table is padded to a multiple of 4.
755 while (StringTable.size() % 4)
756 StringTable += '\x00';
759 void WriteObject(MCAssembler &Asm) {
760 unsigned NumSections = Asm.size();
762 // Create symbol data for any indirect symbols.
763 BindIndirectSymbols(Asm);
765 // Compute symbol table information.
766 SmallString<256> StringTable;
767 std::vector<MachSymbolData> LocalSymbolData;
768 std::vector<MachSymbolData> ExternalSymbolData;
769 std::vector<MachSymbolData> UndefinedSymbolData;
770 unsigned NumSymbols = Asm.symbol_size();
772 // No symbol table command is written if there are no symbols.
774 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
775 UndefinedSymbolData);
777 // The section data starts after the header, the segment load command (and
778 // section headers) and the symbol table.
779 unsigned NumLoadCommands = 1;
780 uint64_t LoadCommandsSize =
781 SegmentLoadCommand32Size + NumSections * Section32Size;
783 // Add the symbol table load command sizes, if used.
785 NumLoadCommands += 2;
786 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
789 // Compute the total size of the section data, as well as its file size and
791 uint64_t SectionDataStart = Header32Size + LoadCommandsSize;
792 uint64_t SectionDataSize = 0;
793 uint64_t SectionDataFileSize = 0;
795 for (MCAssembler::iterator it = Asm.begin(),
796 ie = Asm.end(); it != ie; ++it) {
797 MCSectionData &SD = *it;
799 VMSize = std::max(VMSize, SD.getAddress() + SD.getSize());
801 if (isVirtualSection(SD.getSection()))
804 SectionDataSize = std::max(SectionDataSize,
805 SD.getAddress() + SD.getSize());
806 SectionDataFileSize = std::max(SectionDataFileSize,
807 SD.getAddress() + SD.getFileSize());
810 // The section data is padded to 4 bytes.
812 // FIXME: Is this machine dependent?
813 unsigned SectionDataPadding = OffsetToAlignment(SectionDataFileSize, 4);
814 SectionDataFileSize += SectionDataPadding;
816 // Write the prolog, starting with the header and load command...
817 WriteHeader32(NumLoadCommands, LoadCommandsSize,
818 Asm.getSubsectionsViaSymbols());
819 WriteSegmentLoadCommand32(NumSections, VMSize,
820 SectionDataStart, SectionDataSize);
822 // ... and then the section headers.
824 // We also compute the section relocations while we do this. Note that
825 // computing relocation info will also update the fixup to have the correct
826 // value; this will overwrite the appropriate data in the fragment when it
828 std::vector<MachRelocationEntry> RelocInfos;
829 uint64_t RelocTableEnd = SectionDataStart + SectionDataFileSize;
830 for (MCAssembler::iterator it = Asm.begin(),
831 ie = Asm.end(); it != ie; ++it) {
832 MCSectionData &SD = *it;
834 // The assembler writes relocations in the reverse order they were seen.
836 // FIXME: It is probably more complicated than this.
837 unsigned NumRelocsStart = RelocInfos.size();
838 for (MCSectionData::reverse_iterator it2 = SD.rbegin(),
839 ie2 = SD.rend(); it2 != ie2; ++it2)
840 if (MCDataFragment *DF = dyn_cast<MCDataFragment>(&*it2))
841 for (unsigned i = 0, e = DF->fixup_size(); i != e; ++i)
842 ComputeRelocationInfo(Asm, *DF, DF->getFixups()[e - i - 1],
845 unsigned NumRelocs = RelocInfos.size() - NumRelocsStart;
846 uint64_t SectionStart = SectionDataStart + SD.getAddress();
847 WriteSection32(SD, SectionStart, RelocTableEnd, NumRelocs);
848 RelocTableEnd += NumRelocs * RelocationInfoSize;
851 // Write the symbol table load command, if used.
853 unsigned FirstLocalSymbol = 0;
854 unsigned NumLocalSymbols = LocalSymbolData.size();
855 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
856 unsigned NumExternalSymbols = ExternalSymbolData.size();
857 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
858 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
859 unsigned NumIndirectSymbols = Asm.indirect_symbol_size();
860 unsigned NumSymTabSymbols =
861 NumLocalSymbols + NumExternalSymbols + NumUndefinedSymbols;
862 uint64_t IndirectSymbolSize = NumIndirectSymbols * 4;
863 uint64_t IndirectSymbolOffset = 0;
865 // If used, the indirect symbols are written after the section data.
866 if (NumIndirectSymbols)
867 IndirectSymbolOffset = RelocTableEnd;
869 // The symbol table is written after the indirect symbol data.
870 uint64_t SymbolTableOffset = RelocTableEnd + IndirectSymbolSize;
872 // The string table is written after symbol table.
873 uint64_t StringTableOffset =
874 SymbolTableOffset + NumSymTabSymbols * Nlist32Size;
875 WriteSymtabLoadCommand(SymbolTableOffset, NumSymTabSymbols,
876 StringTableOffset, StringTable.size());
878 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
879 FirstExternalSymbol, NumExternalSymbols,
880 FirstUndefinedSymbol, NumUndefinedSymbols,
881 IndirectSymbolOffset, NumIndirectSymbols);
884 // Write the actual section data.
885 for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it)
886 WriteFileData(OS, *it, *this);
888 // Write the extra padding.
889 WriteZeros(SectionDataPadding);
891 // Write the relocation entries.
892 for (unsigned i = 0, e = RelocInfos.size(); i != e; ++i) {
893 Write32(RelocInfos[i].Word0);
894 Write32(RelocInfos[i].Word1);
897 // Write the symbol table data, if used.
899 // Write the indirect symbol entries.
900 for (MCAssembler::indirect_symbol_iterator
901 it = Asm.indirect_symbol_begin(),
902 ie = Asm.indirect_symbol_end(); it != ie; ++it) {
903 // Indirect symbols in the non lazy symbol pointer section have some
905 const MCSectionMachO &Section =
906 static_cast<const MCSectionMachO&>(it->SectionData->getSection());
908 Section.getTypeAndAttributes() & MCSectionMachO::SECTION_TYPE;
909 if (Type == MCSectionMachO::S_NON_LAZY_SYMBOL_POINTERS) {
910 // If this symbol is defined and internal, mark it as such.
911 if (it->Symbol->isDefined() &&
912 !Asm.getSymbolData(*it->Symbol).isExternal()) {
913 uint32_t Flags = ISF_Local;
914 if (it->Symbol->isAbsolute())
915 Flags |= ISF_Absolute;
921 Write32(Asm.getSymbolData(*it->Symbol).getIndex());
924 // FIXME: Check that offsets match computed ones.
926 // Write the symbol table entries.
927 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
928 WriteNlist32(LocalSymbolData[i]);
929 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
930 WriteNlist32(ExternalSymbolData[i]);
931 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
932 WriteNlist32(UndefinedSymbolData[i]);
934 // Write the string table.
935 OS << StringTable.str();
939 void ApplyFixup(const MCAsmFixup &Fixup, MCDataFragment &DF) {
940 unsigned Size = 1 << getFixupKindLog2Size(Fixup.Kind);
942 // FIXME: Endianness assumption.
943 assert(Fixup.Offset + Size <= DF.getContents().size() &&
944 "Invalid fixup offset!");
945 for (unsigned i = 0; i != Size; ++i)
946 DF.getContents()[Fixup.Offset + i] = uint8_t(Fixup.FixedValue >> (i * 8));
952 MCFragment::MCFragment() : Kind(FragmentType(~0)) {
955 MCFragment::MCFragment(FragmentType _Kind, MCSectionData *_Parent)
958 FileSize(~UINT64_C(0))
961 Parent->getFragmentList().push_back(this);
964 MCFragment::~MCFragment() {
967 uint64_t MCFragment::getAddress() const {
968 assert(getParent() && "Missing Section!");
969 return getParent()->getAddress() + Offset;
974 MCSectionData::MCSectionData() : Section(0) {}
976 MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A)
977 : Section(&_Section),
979 Address(~UINT64_C(0)),
981 FileSize(~UINT64_C(0)),
982 HasInstructions(false)
985 A->getSectionList().push_back(this);
990 MCSymbolData::MCSymbolData() : Symbol(0) {}
992 MCSymbolData::MCSymbolData(const MCSymbol &_Symbol, MCFragment *_Fragment,
993 uint64_t _Offset, MCAssembler *A)
994 : Symbol(&_Symbol), Fragment(_Fragment), Offset(_Offset),
995 IsExternal(false), IsPrivateExtern(false),
996 CommonSize(0), CommonAlign(0), Flags(0), Index(0)
999 A->getSymbolList().push_back(this);
1004 MCAssembler::MCAssembler(MCContext &_Context, TargetAsmBackend &_Backend,
1006 : Context(_Context), Backend(_Backend), OS(_OS), SubsectionsViaSymbols(false)
1010 MCAssembler::~MCAssembler() {
1013 void MCAssembler::LayoutSection(MCSectionData &SD) {
1014 MCAsmLayout Layout(*this);
1015 uint64_t Address = SD.getAddress();
1017 for (MCSectionData::iterator it = SD.begin(), ie = SD.end(); it != ie; ++it) {
1018 MCFragment &F = *it;
1020 F.setOffset(Address - SD.getAddress());
1022 // Evaluate fragment size.
1023 switch (F.getKind()) {
1024 case MCFragment::FT_Align: {
1025 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1027 uint64_t Size = OffsetToAlignment(Address, AF.getAlignment());
1028 if (Size > AF.getMaxBytesToEmit())
1031 AF.setFileSize(Size);
1035 case MCFragment::FT_Data:
1036 case MCFragment::FT_Fill:
1037 F.setFileSize(F.getMaxFileSize());
1040 case MCFragment::FT_Org: {
1041 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1043 int64_t TargetLocation;
1044 if (!OF.getOffset().EvaluateAsAbsolute(TargetLocation, &Layout))
1045 llvm_report_error("expected assembly-time absolute expression");
1047 // FIXME: We need a way to communicate this error.
1048 int64_t Offset = TargetLocation - F.getOffset();
1050 llvm_report_error("invalid .org offset '" + Twine(TargetLocation) +
1051 "' (at offset '" + Twine(F.getOffset()) + "'");
1053 F.setFileSize(Offset);
1057 case MCFragment::FT_ZeroFill: {
1058 MCZeroFillFragment &ZFF = cast<MCZeroFillFragment>(F);
1060 // Align the fragment offset; it is safe to adjust the offset freely since
1061 // this is only in virtual sections.
1062 Address = RoundUpToAlignment(Address, ZFF.getAlignment());
1063 F.setOffset(Address - SD.getAddress());
1065 // FIXME: This is misnamed.
1066 F.setFileSize(ZFF.getSize());
1071 Address += F.getFileSize();
1074 // Set the section sizes.
1075 SD.setSize(Address - SD.getAddress());
1076 if (isVirtualSection(SD.getSection()))
1079 SD.setFileSize(Address - SD.getAddress());
1082 /// WriteNopData - Write optimal nops to the output file for the \arg Count
1083 /// bytes. This returns the number of bytes written. It may return 0 if
1084 /// the \arg Count is more than the maximum optimal nops.
1086 /// FIXME this is X86 32-bit specific and should move to a better place.
1087 static uint64_t WriteNopData(uint64_t Count, MachObjectWriter &MOW) {
1088 static const uint8_t Nops[16][16] = {
1096 {0x0f, 0x1f, 0x40, 0x00},
1097 // nopl 0(%[re]ax,%[re]ax,1)
1098 {0x0f, 0x1f, 0x44, 0x00, 0x00},
1099 // nopw 0(%[re]ax,%[re]ax,1)
1100 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1102 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1103 // nopl 0L(%[re]ax,%[re]ax,1)
1104 {0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1105 // nopw 0L(%[re]ax,%[re]ax,1)
1106 {0x66, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1107 // nopw %cs:0L(%[re]ax,%[re]ax,1)
1108 {0x66, 0x2e, 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00},
1109 // nopl 0(%[re]ax,%[re]ax,1)
1110 // nopw 0(%[re]ax,%[re]ax,1)
1111 {0x0f, 0x1f, 0x44, 0x00, 0x00,
1112 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1113 // nopw 0(%[re]ax,%[re]ax,1)
1114 // nopw 0(%[re]ax,%[re]ax,1)
1115 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1116 0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00},
1117 // nopw 0(%[re]ax,%[re]ax,1)
1118 // nopl 0L(%[re]ax) */
1119 {0x66, 0x0f, 0x1f, 0x44, 0x00, 0x00,
1120 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1123 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1124 0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00},
1126 // nopl 0L(%[re]ax,%[re]ax,1)
1127 {0x0f, 0x1f, 0x80, 0x00, 0x00, 0x00, 0x00,
1128 0x0f, 0x1f, 0x84, 0x00, 0x00, 0x00, 0x00, 0x00}
1134 for (uint64_t i = 0; i < Count; i++)
1135 MOW.Write8 (uint8_t(Nops[Count - 1][i]));
1140 /// WriteFileData - Write the \arg F data to the output file.
1141 static void WriteFileData(raw_ostream &OS, const MCFragment &F,
1142 MachObjectWriter &MOW) {
1143 uint64_t Start = OS.tell();
1148 // FIXME: Embed in fragments instead?
1149 switch (F.getKind()) {
1150 case MCFragment::FT_Align: {
1151 MCAlignFragment &AF = cast<MCAlignFragment>(F);
1152 uint64_t Count = AF.getFileSize() / AF.getValueSize();
1154 // FIXME: This error shouldn't actually occur (the front end should emit
1155 // multiple .align directives to enforce the semantics it wants), but is
1156 // severe enough that we want to report it. How to handle this?
1157 if (Count * AF.getValueSize() != AF.getFileSize())
1158 llvm_report_error("undefined .align directive, value size '" +
1159 Twine(AF.getValueSize()) +
1160 "' is not a divisor of padding size '" +
1161 Twine(AF.getFileSize()) + "'");
1163 // See if we are aligning with nops, and if so do that first to try to fill
1164 // the Count bytes. Then if that did not fill any bytes or there are any
1165 // bytes left to fill use the the Value and ValueSize to fill the rest.
1166 if (AF.getEmitNops()) {
1167 uint64_t NopByteCount = WriteNopData(Count, MOW);
1168 Count -= NopByteCount;
1171 for (uint64_t i = 0; i != Count; ++i) {
1172 switch (AF.getValueSize()) {
1174 assert(0 && "Invalid size!");
1175 case 1: MOW.Write8 (uint8_t (AF.getValue())); break;
1176 case 2: MOW.Write16(uint16_t(AF.getValue())); break;
1177 case 4: MOW.Write32(uint32_t(AF.getValue())); break;
1178 case 8: MOW.Write64(uint64_t(AF.getValue())); break;
1184 case MCFragment::FT_Data: {
1185 MCDataFragment &DF = cast<MCDataFragment>(F);
1187 // Apply the fixups.
1189 // FIXME: Move elsewhere.
1190 for (MCDataFragment::const_fixup_iterator it = DF.fixup_begin(),
1191 ie = DF.fixup_end(); it != ie; ++it)
1192 MOW.ApplyFixup(*it, DF);
1194 OS << cast<MCDataFragment>(F).getContents().str();
1198 case MCFragment::FT_Fill: {
1199 MCFillFragment &FF = cast<MCFillFragment>(F);
1200 for (uint64_t i = 0, e = FF.getCount(); i != e; ++i) {
1201 switch (FF.getValueSize()) {
1203 assert(0 && "Invalid size!");
1204 case 1: MOW.Write8 (uint8_t (FF.getValue())); break;
1205 case 2: MOW.Write16(uint16_t(FF.getValue())); break;
1206 case 4: MOW.Write32(uint32_t(FF.getValue())); break;
1207 case 8: MOW.Write64(uint64_t(FF.getValue())); break;
1213 case MCFragment::FT_Org: {
1214 MCOrgFragment &OF = cast<MCOrgFragment>(F);
1216 for (uint64_t i = 0, e = OF.getFileSize(); i != e; ++i)
1217 MOW.Write8(uint8_t(OF.getValue()));
1222 case MCFragment::FT_ZeroFill: {
1223 assert(0 && "Invalid zero fill fragment in concrete section!");
1228 assert(OS.tell() - Start == F.getFileSize());
1231 /// WriteFileData - Write the \arg SD data to the output file.
1232 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
1233 MachObjectWriter &MOW) {
1234 // Ignore virtual sections.
1235 if (isVirtualSection(SD.getSection())) {
1236 assert(SD.getFileSize() == 0);
1240 uint64_t Start = OS.tell();
1243 for (MCSectionData::const_iterator it = SD.begin(),
1244 ie = SD.end(); it != ie; ++it)
1245 WriteFileData(OS, *it, MOW);
1247 // Add section padding.
1248 assert(SD.getFileSize() >= SD.getSize() && "Invalid section sizes!");
1249 MOW.WriteZeros(SD.getFileSize() - SD.getSize());
1251 assert(OS.tell() - Start == SD.getFileSize());
1254 void MCAssembler::Finish() {
1255 DEBUG_WITH_TYPE("mc-dump", {
1256 llvm::errs() << "assembler backend - pre-layout\n--\n";
1259 // Layout the concrete sections and fragments.
1260 uint64_t Address = 0;
1261 MCSectionData *Prev = 0;
1262 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1263 MCSectionData &SD = *it;
1265 // Skip virtual sections.
1266 if (isVirtualSection(SD.getSection()))
1269 // Align this section if necessary by adding padding bytes to the previous
1271 if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment())) {
1272 assert(Prev && "Missing prev section!");
1273 Prev->setFileSize(Prev->getFileSize() + Pad);
1277 // Layout the section fragments and its size.
1278 SD.setAddress(Address);
1280 Address += SD.getFileSize();
1285 // Layout the virtual sections.
1286 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1287 MCSectionData &SD = *it;
1289 if (!isVirtualSection(SD.getSection()))
1292 // Align this section if necessary by adding padding bytes to the previous
1294 if (uint64_t Pad = OffsetToAlignment(Address, it->getAlignment()))
1297 SD.setAddress(Address);
1299 Address += SD.getSize();
1303 DEBUG_WITH_TYPE("mc-dump", {
1304 llvm::errs() << "assembler backend - post-layout\n--\n";
1307 // Write the object file.
1308 MachObjectWriter MOW(OS);
1309 MOW.WriteObject(*this);
1315 // Debugging methods
1319 raw_ostream &operator<<(raw_ostream &OS, const MCAsmFixup &AF) {
1320 OS << "<MCAsmFixup" << " Offset:" << AF.Offset << " Value:" << *AF.Value
1321 << " Kind:" << AF.Kind << ">";
1327 void MCFragment::dump() {
1328 raw_ostream &OS = llvm::errs();
1330 OS << "<MCFragment " << (void*) this << " Offset:" << Offset
1331 << " FileSize:" << FileSize;
1336 void MCAlignFragment::dump() {
1337 raw_ostream &OS = llvm::errs();
1339 OS << "<MCAlignFragment ";
1340 this->MCFragment::dump();
1342 OS << " Alignment:" << getAlignment()
1343 << " Value:" << getValue() << " ValueSize:" << getValueSize()
1344 << " MaxBytesToEmit:" << getMaxBytesToEmit() << ">";
1347 void MCDataFragment::dump() {
1348 raw_ostream &OS = llvm::errs();
1350 OS << "<MCDataFragment ";
1351 this->MCFragment::dump();
1353 OS << " Contents:[";
1354 for (unsigned i = 0, e = getContents().size(); i != e; ++i) {
1356 OS << hexdigit((Contents[i] >> 4) & 0xF) << hexdigit(Contents[i] & 0xF);
1358 OS << "] (" << getContents().size() << " bytes)";
1360 if (!getFixups().empty()) {
1363 for (fixup_iterator it = fixup_begin(), ie = fixup_end(); it != ie; ++it) {
1364 if (it != fixup_begin()) OS << ",\n ";
1373 void MCFillFragment::dump() {
1374 raw_ostream &OS = llvm::errs();
1376 OS << "<MCFillFragment ";
1377 this->MCFragment::dump();
1379 OS << " Value:" << getValue() << " ValueSize:" << getValueSize()
1380 << " Count:" << getCount() << ">";
1383 void MCOrgFragment::dump() {
1384 raw_ostream &OS = llvm::errs();
1386 OS << "<MCOrgFragment ";
1387 this->MCFragment::dump();
1389 OS << " Offset:" << getOffset() << " Value:" << getValue() << ">";
1392 void MCZeroFillFragment::dump() {
1393 raw_ostream &OS = llvm::errs();
1395 OS << "<MCZeroFillFragment ";
1396 this->MCFragment::dump();
1398 OS << " Size:" << getSize() << " Alignment:" << getAlignment() << ">";
1401 void MCSectionData::dump() {
1402 raw_ostream &OS = llvm::errs();
1404 OS << "<MCSectionData";
1405 OS << " Alignment:" << getAlignment() << " Address:" << Address
1406 << " Size:" << Size << " FileSize:" << FileSize
1407 << " Fragments:[\n ";
1408 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1409 if (it != begin()) OS << ",\n ";
1415 void MCSymbolData::dump() {
1416 raw_ostream &OS = llvm::errs();
1418 OS << "<MCSymbolData Symbol:" << getSymbol()
1419 << " Fragment:" << getFragment() << " Offset:" << getOffset()
1420 << " Flags:" << getFlags() << " Index:" << getIndex();
1422 OS << " (common, size:" << getCommonSize()
1423 << " align: " << getCommonAlignment() << ")";
1425 OS << " (external)";
1426 if (isPrivateExtern())
1427 OS << " (private extern)";
1431 void MCAssembler::dump() {
1432 raw_ostream &OS = llvm::errs();
1434 OS << "<MCAssembler\n";
1435 OS << " Sections:[\n ";
1436 for (iterator it = begin(), ie = end(); it != ie; ++it) {
1437 if (it != begin()) OS << ",\n ";
1443 for (symbol_iterator it = symbol_begin(), ie = symbol_end(); it != ie; ++it) {
1444 if (it != symbol_begin()) OS << ",\n ";