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 #include "llvm/MC/MCAssembler.h"
11 #include "llvm/MC/MCSectionMachO.h"
12 #include "llvm/Target/TargetMachOWriterInfo.h"
13 #include "llvm/ADT/DenseMap.h"
14 #include "llvm/ADT/SmallString.h"
15 #include "llvm/ADT/StringMap.h"
16 #include "llvm/ADT/Twine.h"
17 #include "llvm/Support/ErrorHandling.h"
18 #include "llvm/Support/raw_ostream.h"
22 class MachObjectWriter;
24 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
25 MachObjectWriter &MOW);
27 class MachObjectWriter {
28 // See <mach-o/loader.h>.
30 Header_Magic32 = 0xFEEDFACE,
31 Header_Magic64 = 0xFEEDFACF
34 static const unsigned Header32Size = 28;
35 static const unsigned Header64Size = 32;
36 static const unsigned SegmentLoadCommand32Size = 56;
37 static const unsigned Section32Size = 68;
38 static const unsigned SymtabLoadCommandSize = 24;
39 static const unsigned DysymtabLoadCommandSize = 80;
40 static const unsigned Nlist32Size = 12;
46 enum LoadCommandType {
52 // See <mach-o/nlist.h>.
59 enum SymbolTypeFlags {
60 // If any of these bits are set, then the entry is a stab entry number (see
61 // <mach-o/stab.h>. Otherwise the other masks apply.
62 STF_StabsEntryMask = 0xe0,
66 STF_PrivateExtern = 0x10
69 /// MachSymbolData - Helper struct for containing some precomputed information
71 struct MachSymbolData {
72 MCSymbolData *SymbolData;
76 // Support lexicographic sorting.
77 bool operator<(const MachSymbolData &RHS) const {
78 const std::string &Name = SymbolData->getSymbol().getName();
79 return Name < RHS.SymbolData->getSymbol().getName();
87 MachObjectWriter(raw_ostream &_OS, bool _IsLSB = true)
88 : OS(_OS), IsLSB(_IsLSB) {
91 /// @name Helper Methods
94 void Write8(uint8_t Value) {
98 void Write16(uint16_t Value) {
100 Write8(uint8_t(Value >> 0));
101 Write8(uint8_t(Value >> 8));
103 Write8(uint8_t(Value >> 8));
104 Write8(uint8_t(Value >> 0));
108 void Write32(uint32_t Value) {
110 Write16(uint16_t(Value >> 0));
111 Write16(uint16_t(Value >> 16));
113 Write16(uint16_t(Value >> 16));
114 Write16(uint16_t(Value >> 0));
118 void Write64(uint64_t Value) {
120 Write32(uint32_t(Value >> 0));
121 Write32(uint32_t(Value >> 32));
123 Write32(uint32_t(Value >> 32));
124 Write32(uint32_t(Value >> 0));
128 void WriteZeros(unsigned N) {
129 const char Zeros[16] = { 0 };
131 for (unsigned i = 0, e = N / 16; i != e; ++i)
132 OS << StringRef(Zeros, 16);
134 OS << StringRef(Zeros, N % 16);
137 void WriteString(const StringRef &Str, unsigned ZeroFillSize = 0) {
140 WriteZeros(ZeroFillSize - Str.size());
145 void WriteHeader32(unsigned NumLoadCommands, unsigned LoadCommandsSize) {
146 // struct mach_header (28 bytes)
148 uint64_t Start = OS.tell();
151 Write32(Header_Magic32);
153 // FIXME: Support cputype.
154 Write32(TargetMachOWriterInfo::HDR_CPU_TYPE_I386);
156 // FIXME: Support cpusubtype.
157 Write32(TargetMachOWriterInfo::HDR_CPU_SUBTYPE_I386_ALL);
161 // Object files have a single load command, the segment.
162 Write32(NumLoadCommands);
163 Write32(LoadCommandsSize);
166 assert(OS.tell() - Start == Header32Size);
169 /// WriteSegmentLoadCommand32 - Write a 32-bit segment load command.
171 /// \arg NumSections - The number of sections in this segment.
172 /// \arg SectionDataSize - The total size of the sections.
173 void WriteSegmentLoadCommand32(unsigned NumSections,
174 uint64_t SectionDataStartOffset,
175 uint64_t SectionDataSize) {
176 // struct segment_command (56 bytes)
178 uint64_t Start = OS.tell();
181 Write32(LCT_Segment);
182 Write32(SegmentLoadCommand32Size + NumSections * Section32Size);
185 Write32(0); // vmaddr
186 Write32(SectionDataSize); // vmsize
187 Write32(SectionDataStartOffset); // file offset
188 Write32(SectionDataSize); // file size
189 Write32(0x7); // maxprot
190 Write32(0x7); // initprot
191 Write32(NumSections);
194 assert(OS.tell() - Start == SegmentLoadCommand32Size);
197 void WriteSection32(const MCSectionData &SD, uint64_t FileOffset) {
198 // struct section (68 bytes)
200 uint64_t Start = OS.tell();
203 // FIXME: cast<> support!
204 const MCSectionMachO &Section =
205 static_cast<const MCSectionMachO&>(SD.getSection());
206 WriteString(Section.getSectionName(), 16);
207 WriteString(Section.getSegmentName(), 16);
208 Write32(0); // address
209 Write32(SD.getFileSize()); // size
212 assert(isPowerOf2_32(SD.getAlignment()) && "Invalid alignment!");
213 Write32(Log2_32(SD.getAlignment()));
214 Write32(0); // file offset of relocation entries
215 Write32(0); // number of relocation entrions
216 Write32(Section.getTypeAndAttributes());
217 Write32(0); // reserved1
218 Write32(Section.getStubSize()); // reserved2
220 assert(OS.tell() - Start == Section32Size);
223 void WriteSymtabLoadCommand(uint32_t SymbolOffset, uint32_t NumSymbols,
224 uint32_t StringTableOffset,
225 uint32_t StringTableSize) {
226 // struct symtab_command (24 bytes)
228 uint64_t Start = OS.tell();
232 Write32(SymtabLoadCommandSize);
233 Write32(SymbolOffset);
235 Write32(StringTableOffset);
236 Write32(StringTableSize);
238 assert(OS.tell() - Start == SymtabLoadCommandSize);
241 void WriteDysymtabLoadCommand(uint32_t FirstLocalSymbol,
242 uint32_t NumLocalSymbols,
243 uint32_t FirstExternalSymbol,
244 uint32_t NumExternalSymbols,
245 uint32_t FirstUndefinedSymbol,
246 uint32_t NumUndefinedSymbols,
247 uint32_t IndirectSymbolOffset,
248 uint32_t NumIndirectSymbols) {
249 // struct dysymtab_command (80 bytes)
251 uint64_t Start = OS.tell();
254 Write32(LCT_Dysymtab);
255 Write32(DysymtabLoadCommandSize);
256 Write32(FirstLocalSymbol);
257 Write32(NumLocalSymbols);
258 Write32(FirstExternalSymbol);
259 Write32(NumExternalSymbols);
260 Write32(FirstUndefinedSymbol);
261 Write32(NumUndefinedSymbols);
262 Write32(0); // tocoff
264 Write32(0); // modtaboff
265 Write32(0); // nmodtab
266 Write32(0); // extrefsymoff
267 Write32(0); // nextrefsyms
268 Write32(IndirectSymbolOffset);
269 Write32(NumIndirectSymbols);
270 Write32(0); // extreloff
271 Write32(0); // nextrel
272 Write32(0); // locreloff
273 Write32(0); // nlocrel
275 assert(OS.tell() - Start == DysymtabLoadCommandSize);
278 void WriteNlist32(MachSymbolData &MSD) {
279 MCSymbol &Symbol = MSD.SymbolData->getSymbol();
282 // Set the N_TYPE bits. See <mach-o/nlist.h>.
284 // FIXME: Are the prebound or indirect fields possible here?
285 if (Symbol.isUndefined())
286 Type = STT_Undefined;
287 else if (Symbol.isAbsolute())
292 // FIXME: Set STAB bits.
294 // FIXME: Set private external bit.
297 if (MSD.SymbolData->isExternal())
298 Type |= STF_External;
300 // struct nlist (12 bytes)
302 Write32(MSD.StringIndex);
304 Write8(MSD.SectionIndex);
305 Write16(0); // FIXME: Desc
306 Write32(0); // FIXME: Value
309 /// ComputeSymbolTable - Compute the symbol table data
311 /// \param StringTable [out] - The string table data.
312 /// \param StringIndexMap [out] - Map from symbol names to offsets in the
315 void ComputeSymbolTable(MCAssembler &Asm, SmallString<256> &StringTable,
316 std::vector<MachSymbolData> &LocalSymbolData,
317 std::vector<MachSymbolData> &ExternalSymbolData,
318 std::vector<MachSymbolData> &UndefinedSymbolData) {
319 // Build section lookup table.
320 DenseMap<const MCSection*, uint8_t> SectionIndexMap;
322 for (MCAssembler::iterator it = Asm.begin(),
323 ie = Asm.end(); it != ie; ++it, ++Index)
324 SectionIndexMap[&it->getSection()] = Index;
325 assert(Index <= 256 && "Too many sections!");
327 // Index 0 is always the empty string.
328 StringMap<uint64_t> StringIndexMap;
329 StringTable += '\x00';
331 // Build the symbol arrays and the string table, but only for non-local
334 // The particular order that we collect the symbols and create the string
335 // table, then sort the symbols is chosen to match 'as'. Even though it
336 // doesn't matter for correctness, this is important for letting us diff .o
338 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
339 ie = Asm.symbol_end(); it != ie; ++it) {
340 MCSymbol &Symbol = it->getSymbol();
342 if (!it->isExternal())
345 uint64_t &Entry = StringIndexMap[Symbol.getName()];
347 Entry = StringTable.size();
348 StringTable += Symbol.getName();
349 StringTable += '\x00';
354 MSD.StringIndex = Entry;
356 if (Symbol.isUndefined()) {
357 MSD.SectionIndex = 0;
358 UndefinedSymbolData.push_back(MSD);
359 } else if (Symbol.isAbsolute()) {
360 MSD.SectionIndex = 0;
361 ExternalSymbolData.push_back(MSD);
363 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
364 assert(MSD.SectionIndex && "Invalid section index!");
365 ExternalSymbolData.push_back(MSD);
369 // Now add the data for local symbols.
370 for (MCAssembler::symbol_iterator it = Asm.symbol_begin(),
371 ie = Asm.symbol_end(); it != ie; ++it) {
372 MCSymbol &Symbol = it->getSymbol();
374 if (it->isExternal())
377 uint64_t &Entry = StringIndexMap[Symbol.getName()];
379 Entry = StringTable.size();
380 StringTable += Symbol.getName();
381 StringTable += '\x00';
386 MSD.StringIndex = Entry;
388 assert(!Symbol.isUndefined() && "Local symbol can not be undefined!");
389 if (Symbol.isAbsolute()) {
390 MSD.SectionIndex = 0;
391 LocalSymbolData.push_back(MSD);
393 MSD.SectionIndex = SectionIndexMap.lookup(&Symbol.getSection());
394 assert(MSD.SectionIndex && "Invalid section index!");
395 LocalSymbolData.push_back(MSD);
399 // External and undefined symbols are required to be in lexicographic order.
400 std::sort(ExternalSymbolData.begin(), ExternalSymbolData.end());
401 std::sort(UndefinedSymbolData.begin(), UndefinedSymbolData.end());
403 // The string table is padded to a multiple of 4.
405 // FIXME: Check to see if this varies per arch.
406 while (StringTable.size() % 4)
407 StringTable += '\x00';
410 void WriteObject(MCAssembler &Asm) {
411 unsigned NumSections = Asm.size();
413 // Compute symbol table information.
414 SmallString<256> StringTable;
415 std::vector<MachSymbolData> LocalSymbolData;
416 std::vector<MachSymbolData> ExternalSymbolData;
417 std::vector<MachSymbolData> UndefinedSymbolData;
418 unsigned NumSymbols = Asm.symbol_size();
420 // No symbol table command is written if there are no symbols.
422 ComputeSymbolTable(Asm, StringTable, LocalSymbolData, ExternalSymbolData,
423 UndefinedSymbolData);
425 // Compute the file offsets for all the sections in advance, so that we can
426 // write things out in order.
427 SmallVector<uint64_t, 16> SectionFileOffsets;
428 SectionFileOffsets.resize(NumSections);
430 // The section data starts after the header, the segment load command (and
431 // section headers) and the symbol table.
432 unsigned NumLoadCommands = 1;
433 uint64_t LoadCommandsSize =
434 SegmentLoadCommand32Size + NumSections * Section32Size;
436 // Add the symbol table load command sizes, if used.
438 NumLoadCommands += 2;
439 LoadCommandsSize += SymtabLoadCommandSize + DysymtabLoadCommandSize;
442 uint64_t FileOffset = Header32Size + LoadCommandsSize;
443 uint64_t SectionDataStartOffset = FileOffset;
444 uint64_t SectionDataSize = 0;
446 for (MCAssembler::iterator it = Asm.begin(),
447 ie = Asm.end(); it != ie; ++it, ++Index) {
448 SectionFileOffsets[Index] = FileOffset;
449 FileOffset += it->getFileSize();
450 SectionDataSize += it->getFileSize();
453 // Write the prolog, starting with the header and load command...
454 WriteHeader32(NumLoadCommands, LoadCommandsSize);
455 WriteSegmentLoadCommand32(NumSections, SectionDataStartOffset,
458 // ... and then the section headers.
460 for (MCAssembler::iterator it = Asm.begin(),
461 ie = Asm.end(); it != ie; ++it, ++Index)
462 WriteSection32(*it, SectionFileOffsets[Index]);
464 // Write the symbol table load command, if used.
466 // The string table is written after all the section data.
467 uint64_t SymbolTableOffset = SectionDataStartOffset + SectionDataSize;
468 uint64_t StringTableOffset =
469 SymbolTableOffset + NumSymbols * Nlist32Size;
470 WriteSymtabLoadCommand(SymbolTableOffset, NumSymbols,
471 StringTableOffset, StringTable.size());
473 unsigned FirstLocalSymbol = 0;
474 unsigned NumLocalSymbols = LocalSymbolData.size();
475 unsigned FirstExternalSymbol = FirstLocalSymbol + NumLocalSymbols;
476 unsigned NumExternalSymbols = ExternalSymbolData.size();
477 unsigned FirstUndefinedSymbol = FirstExternalSymbol + NumExternalSymbols;
478 unsigned NumUndefinedSymbols = UndefinedSymbolData.size();
479 // FIXME: Get correct symbol indices and counts for indirect symbols.
480 unsigned IndirectSymbolOffset = 0;
481 unsigned NumIndirectSymbols = 0;
482 WriteDysymtabLoadCommand(FirstLocalSymbol, NumLocalSymbols,
483 FirstExternalSymbol, NumExternalSymbols,
484 FirstUndefinedSymbol, NumUndefinedSymbols,
485 IndirectSymbolOffset, NumIndirectSymbols);
488 // Write the actual section data.
489 for (MCAssembler::iterator it = Asm.begin(), ie = Asm.end(); it != ie; ++it)
490 WriteFileData(OS, *it, *this);
492 // Write the symbol table data, if used.
494 // FIXME: Check that offsets match computed ones.
496 // FIXME: Some of these are ordered by name to help the linker.
498 // Write the symbol table entries.
499 for (unsigned i = 0, e = LocalSymbolData.size(); i != e; ++i)
500 WriteNlist32(LocalSymbolData[i]);
501 for (unsigned i = 0, e = ExternalSymbolData.size(); i != e; ++i)
502 WriteNlist32(ExternalSymbolData[i]);
503 for (unsigned i = 0, e = UndefinedSymbolData.size(); i != e; ++i)
504 WriteNlist32(UndefinedSymbolData[i]);
506 // Write the string table.
507 OS << StringTable.str();
514 MCFragment::MCFragment() : Kind(FragmentType(~0)) {
517 MCFragment::MCFragment(FragmentType _Kind, MCSectionData *SD)
519 FileSize(~UINT64_C(0))
522 SD->getFragmentList().push_back(this);
525 MCFragment::~MCFragment() {
530 MCSectionData::MCSectionData() : Section(*(MCSection*)0) {}
532 MCSectionData::MCSectionData(const MCSection &_Section, MCAssembler *A)
535 FileSize(~UINT64_C(0))
538 A->getSectionList().push_back(this);
543 MCSymbolData::MCSymbolData() : Symbol(*(MCSymbol*)0) {}
545 MCSymbolData::MCSymbolData(MCSymbol &_Symbol, MCFragment *_Fragment,
546 uint64_t _Offset, MCAssembler *A)
547 : Symbol(_Symbol), Fragment(_Fragment), Offset(_Offset),
551 A->getSymbolList().push_back(this);
556 MCAssembler::MCAssembler(raw_ostream &_OS) : OS(_OS) {}
558 MCAssembler::~MCAssembler() {
561 void MCAssembler::LayoutSection(MCSectionData &SD) {
564 for (MCSectionData::iterator it = SD.begin(), ie = SD.end(); it != ie; ++it) {
569 // Evaluate fragment size.
570 switch (F.getKind()) {
571 case MCFragment::FT_Align: {
572 MCAlignFragment &AF = cast<MCAlignFragment>(F);
574 uint64_t AlignedOffset = RoundUpToAlignment(Offset, AF.getAlignment());
575 uint64_t PaddingBytes = AlignedOffset - Offset;
577 if (PaddingBytes > AF.getMaxBytesToEmit())
580 AF.setFileSize(PaddingBytes);
584 case MCFragment::FT_Data:
585 case MCFragment::FT_Fill:
586 F.setFileSize(F.getMaxFileSize());
589 case MCFragment::FT_Org: {
590 MCOrgFragment &OF = cast<MCOrgFragment>(F);
592 if (!OF.getOffset().isAbsolute())
593 llvm_unreachable("FIXME: Not yet implemented!");
594 uint64_t OrgOffset = OF.getOffset().getConstant();
596 // FIXME: We need a way to communicate this error.
597 if (OrgOffset < Offset)
598 llvm_report_error("invalid .org offset '" + Twine(OrgOffset) +
599 "' (section offset '" + Twine(Offset) + "'");
601 F.setFileSize(OrgOffset - Offset);
606 Offset += F.getFileSize();
609 // FIXME: Pad section?
610 SD.setFileSize(Offset);
613 /// WriteFileData - Write the \arg F data to the output file.
614 static void WriteFileData(raw_ostream &OS, const MCFragment &F,
615 MachObjectWriter &MOW) {
616 uint64_t Start = OS.tell();
619 // FIXME: Embed in fragments instead?
620 switch (F.getKind()) {
621 case MCFragment::FT_Align: {
622 MCAlignFragment &AF = cast<MCAlignFragment>(F);
623 uint64_t Count = AF.getFileSize() / AF.getValueSize();
625 // FIXME: This error shouldn't actually occur (the front end should emit
626 // multiple .align directives to enforce the semantics it wants), but is
627 // severe enough that we want to report it. How to handle this?
628 if (Count * AF.getValueSize() != AF.getFileSize())
629 llvm_report_error("undefined .align directive, value size '" +
630 Twine(AF.getValueSize()) +
631 "' is not a divisor of padding size '" +
632 Twine(AF.getFileSize()) + "'");
634 for (uint64_t i = 0; i != Count; ++i) {
635 switch (AF.getValueSize()) {
637 assert(0 && "Invalid size!");
638 case 1: MOW.Write8 (uint8_t (AF.getValue())); break;
639 case 2: MOW.Write16(uint16_t(AF.getValue())); break;
640 case 4: MOW.Write32(uint32_t(AF.getValue())); break;
641 case 8: MOW.Write64(uint64_t(AF.getValue())); break;
647 case MCFragment::FT_Data:
648 OS << cast<MCDataFragment>(F).getContents().str();
651 case MCFragment::FT_Fill: {
652 MCFillFragment &FF = cast<MCFillFragment>(F);
654 if (!FF.getValue().isAbsolute())
655 llvm_unreachable("FIXME: Not yet implemented!");
656 int64_t Value = FF.getValue().getConstant();
658 for (uint64_t i = 0, e = FF.getCount(); i != e; ++i) {
659 switch (FF.getValueSize()) {
661 assert(0 && "Invalid size!");
662 case 1: MOW.Write8 (uint8_t (Value)); break;
663 case 2: MOW.Write16(uint16_t(Value)); break;
664 case 4: MOW.Write32(uint32_t(Value)); break;
665 case 8: MOW.Write64(uint64_t(Value)); break;
671 case MCFragment::FT_Org: {
672 MCOrgFragment &OF = cast<MCOrgFragment>(F);
674 for (uint64_t i = 0, e = OF.getFileSize(); i != e; ++i)
675 MOW.Write8(uint8_t(OF.getValue()));
681 assert(OS.tell() - Start == F.getFileSize());
684 /// WriteFileData - Write the \arg SD data to the output file.
685 static void WriteFileData(raw_ostream &OS, const MCSectionData &SD,
686 MachObjectWriter &MOW) {
687 uint64_t Start = OS.tell();
690 for (MCSectionData::const_iterator it = SD.begin(),
691 ie = SD.end(); it != ie; ++it)
692 WriteFileData(OS, *it, MOW);
694 assert(OS.tell() - Start == SD.getFileSize());
697 void MCAssembler::Finish() {
698 // Layout the sections and fragments.
699 for (iterator it = begin(), ie = end(); it != ie; ++it)
702 // Write the object file.
703 MachObjectWriter MOW(OS);
704 MOW.WriteObject(*this);