1 //===- ArchiveWriter.cpp - ar File Format implementation --------*- C++ -*-===//
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 // This file defines the writeArchive function.
12 //===----------------------------------------------------------------------===//
14 #include "llvm/Object/ArchiveWriter.h"
15 #include "llvm/ADT/ArrayRef.h"
16 #include "llvm/ADT/StringRef.h"
17 #include "llvm/BinaryFormat/Magic.h"
18 #include "llvm/IR/LLVMContext.h"
19 #include "llvm/Object/Archive.h"
20 #include "llvm/Object/ObjectFile.h"
21 #include "llvm/Object/SymbolicFile.h"
22 #include "llvm/Support/EndianStream.h"
23 #include "llvm/Support/Errc.h"
24 #include "llvm/Support/ErrorHandling.h"
25 #include "llvm/Support/Format.h"
26 #include "llvm/Support/Path.h"
27 #include "llvm/Support/ToolOutputFile.h"
28 #include "llvm/Support/raw_ostream.h"
32 #if !defined(_MSC_VER) && !defined(__MINGW32__)
40 NewArchiveMember::NewArchiveMember(MemoryBufferRef BufRef)
41 : Buf(MemoryBuffer::getMemBuffer(BufRef, false)),
42 MemberName(BufRef.getBufferIdentifier()) {}
44 Expected<NewArchiveMember>
45 NewArchiveMember::getOldMember(const object::Archive::Child &OldMember,
47 Expected<llvm::MemoryBufferRef> BufOrErr = OldMember.getMemoryBufferRef();
49 return BufOrErr.takeError();
52 assert(M.IsNew == false);
53 M.Buf = MemoryBuffer::getMemBuffer(*BufOrErr, false);
54 M.MemberName = M.Buf->getBufferIdentifier();
56 auto ModTimeOrErr = OldMember.getLastModified();
58 return ModTimeOrErr.takeError();
59 M.ModTime = ModTimeOrErr.get();
60 Expected<unsigned> UIDOrErr = OldMember.getUID();
62 return UIDOrErr.takeError();
63 M.UID = UIDOrErr.get();
64 Expected<unsigned> GIDOrErr = OldMember.getGID();
66 return GIDOrErr.takeError();
67 M.GID = GIDOrErr.get();
68 Expected<sys::fs::perms> AccessModeOrErr = OldMember.getAccessMode();
70 return AccessModeOrErr.takeError();
71 M.Perms = AccessModeOrErr.get();
76 Expected<NewArchiveMember> NewArchiveMember::getFile(StringRef FileName,
78 sys::fs::file_status Status;
80 if (auto EC = sys::fs::openFileForRead(FileName, FD))
81 return errorCodeToError(EC);
84 if (auto EC = sys::fs::status(FD, Status))
85 return errorCodeToError(EC);
87 // Opening a directory doesn't make sense. Let it fail.
88 // Linux cannot open directories with open(2), although
89 // cygwin and *bsd can.
90 if (Status.type() == sys::fs::file_type::directory_file)
91 return errorCodeToError(make_error_code(errc::is_a_directory));
93 ErrorOr<std::unique_ptr<MemoryBuffer>> MemberBufferOrErr =
94 MemoryBuffer::getOpenFile(FD, FileName, Status.getSize(), false);
95 if (!MemberBufferOrErr)
96 return errorCodeToError(MemberBufferOrErr.getError());
99 return errorCodeToError(std::error_code(errno, std::generic_category()));
103 M.Buf = std::move(*MemberBufferOrErr);
104 M.MemberName = M.Buf->getBufferIdentifier();
105 if (!Deterministic) {
106 M.ModTime = std::chrono::time_point_cast<std::chrono::seconds>(
107 Status.getLastModificationTime());
108 M.UID = Status.getUser();
109 M.GID = Status.getGroup();
110 M.Perms = Status.permissions();
115 template <typename T>
116 static void printWithSpacePadding(raw_ostream &OS, T Data, unsigned Size) {
117 uint64_t OldPos = OS.tell();
119 unsigned SizeSoFar = OS.tell() - OldPos;
120 assert(SizeSoFar <= Size && "Data doesn't fit in Size");
121 OS.indent(Size - SizeSoFar);
124 static bool isDarwin(object::Archive::Kind Kind) {
125 return Kind == object::Archive::K_DARWIN ||
126 Kind == object::Archive::K_DARWIN64;
129 static bool isBSDLike(object::Archive::Kind Kind) {
131 case object::Archive::K_GNU:
132 case object::Archive::K_GNU64:
134 case object::Archive::K_BSD:
135 case object::Archive::K_DARWIN:
136 case object::Archive::K_DARWIN64:
138 case object::Archive::K_COFF:
141 llvm_unreachable("not supported for writting");
145 static void print(raw_ostream &Out, object::Archive::Kind Kind, T Val) {
146 support::endian::write(Out, Val,
147 isBSDLike(Kind) ? support::little : support::big);
150 static void printRestOfMemberHeader(
151 raw_ostream &Out, const sys::TimePoint<std::chrono::seconds> &ModTime,
152 unsigned UID, unsigned GID, unsigned Perms, unsigned Size) {
153 printWithSpacePadding(Out, sys::toTimeT(ModTime), 12);
155 // The format has only 6 chars for uid and gid. Truncate if the provided
157 printWithSpacePadding(Out, UID % 1000000, 6);
158 printWithSpacePadding(Out, GID % 1000000, 6);
160 printWithSpacePadding(Out, format("%o", Perms), 8);
161 printWithSpacePadding(Out, Size, 10);
166 printGNUSmallMemberHeader(raw_ostream &Out, StringRef Name,
167 const sys::TimePoint<std::chrono::seconds> &ModTime,
168 unsigned UID, unsigned GID, unsigned Perms,
170 printWithSpacePadding(Out, Twine(Name) + "/", 16);
171 printRestOfMemberHeader(Out, ModTime, UID, GID, Perms, Size);
175 printBSDMemberHeader(raw_ostream &Out, uint64_t Pos, StringRef Name,
176 const sys::TimePoint<std::chrono::seconds> &ModTime,
177 unsigned UID, unsigned GID, unsigned Perms,
179 uint64_t PosAfterHeader = Pos + 60 + Name.size();
180 // Pad so that even 64 bit object files are aligned.
181 unsigned Pad = OffsetToAlignment(PosAfterHeader, 8);
182 unsigned NameWithPadding = Name.size() + Pad;
183 printWithSpacePadding(Out, Twine("#1/") + Twine(NameWithPadding), 16);
184 printRestOfMemberHeader(Out, ModTime, UID, GID, Perms,
185 NameWithPadding + Size);
188 Out.write(uint8_t(0));
191 static bool useStringTable(bool Thin, StringRef Name) {
192 return Thin || Name.size() >= 16 || Name.contains('/');
195 // Compute the relative path from From to To.
196 static std::string computeRelativePath(StringRef From, StringRef To) {
197 if (sys::path::is_absolute(From) || sys::path::is_absolute(To))
200 StringRef DirFrom = sys::path::parent_path(From);
201 auto FromI = sys::path::begin(DirFrom);
202 auto ToI = sys::path::begin(To);
203 while (*FromI == *ToI) {
208 SmallString<128> Relative;
209 for (auto FromE = sys::path::end(DirFrom); FromI != FromE; ++FromI)
210 sys::path::append(Relative, "..");
212 for (auto ToE = sys::path::end(To); ToI != ToE; ++ToI)
213 sys::path::append(Relative, *ToI);
216 // Replace backslashes with slashes so that the path is portable between *nix
218 std::replace(Relative.begin(), Relative.end(), '\\', '/');
221 return Relative.str();
224 static bool is64BitKind(object::Archive::Kind Kind) {
226 case object::Archive::K_GNU:
227 case object::Archive::K_BSD:
228 case object::Archive::K_DARWIN:
229 case object::Archive::K_COFF:
231 case object::Archive::K_DARWIN64:
232 case object::Archive::K_GNU64:
235 llvm_unreachable("not supported for writting");
238 static void addToStringTable(raw_ostream &Out, StringRef ArcName,
239 const NewArchiveMember &M, bool Thin) {
240 StringRef ID = M.Buf->getBufferIdentifier();
243 Out << computeRelativePath(ArcName, ID);
251 static void printMemberHeader(raw_ostream &Out, uint64_t Pos,
252 raw_ostream &StringTable,
253 StringMap<uint64_t> &MemberNames,
254 object::Archive::Kind Kind, bool Thin,
255 StringRef ArcName, const NewArchiveMember &M,
256 sys::TimePoint<std::chrono::seconds> ModTime,
260 return printBSDMemberHeader(Out, Pos, M.MemberName, ModTime, M.UID, M.GID,
262 if (!useStringTable(Thin, M.MemberName))
263 return printGNUSmallMemberHeader(Out, M.MemberName, ModTime, M.UID, M.GID,
268 NamePos = StringTable.tell();
269 addToStringTable(StringTable, ArcName, M, Thin);
271 auto Insertion = MemberNames.insert({M.MemberName, uint64_t(0)});
272 if (Insertion.second) {
273 Insertion.first->second = StringTable.tell();
274 addToStringTable(StringTable, ArcName, M, Thin);
276 NamePos = Insertion.first->second;
278 printWithSpacePadding(Out, NamePos, 15);
279 printRestOfMemberHeader(Out, ModTime, M.UID, M.GID, M.Perms, Size);
284 std::vector<unsigned> Symbols;
291 static MemberData computeStringTable(StringRef Names) {
292 unsigned Size = Names.size();
293 unsigned Pad = OffsetToAlignment(Size, 2);
295 raw_string_ostream Out(Header);
296 printWithSpacePadding(Out, "//", 48);
297 printWithSpacePadding(Out, Size + Pad, 10);
300 return {{}, std::move(Header), Names, Pad ? "\n" : ""};
303 static sys::TimePoint<std::chrono::seconds> now(bool Deterministic) {
304 using namespace std::chrono;
307 return time_point_cast<seconds>(system_clock::now());
308 return sys::TimePoint<seconds>();
311 static bool isArchiveSymbol(const object::BasicSymbolRef &S) {
312 uint32_t Symflags = S.getFlags();
313 if (Symflags & object::SymbolRef::SF_FormatSpecific)
315 if (!(Symflags & object::SymbolRef::SF_Global))
317 if (Symflags & object::SymbolRef::SF_Undefined)
322 static void printNBits(raw_ostream &Out, object::Archive::Kind Kind,
324 if (is64BitKind(Kind))
325 print<uint64_t>(Out, Kind, Val);
327 print<uint32_t>(Out, Kind, Val);
330 static void writeSymbolTable(raw_ostream &Out, object::Archive::Kind Kind,
331 bool Deterministic, ArrayRef<MemberData> Members,
332 StringRef StringTable) {
333 // We don't write a symbol table on an archive with no members -- except on
334 // Darwin, where the linker will abort unless the archive has a symbol table.
335 if (StringTable.empty() && !isDarwin(Kind))
338 unsigned NumSyms = 0;
339 for (const MemberData &M : Members)
340 NumSyms += M.Symbols.size();
343 unsigned OffsetSize = is64BitKind(Kind) ? sizeof(uint64_t) : sizeof(uint32_t);
345 Size += OffsetSize; // Number of entries
347 Size += NumSyms * OffsetSize * 2; // Table
349 Size += NumSyms * OffsetSize; // Table
351 Size += OffsetSize; // byte count
352 Size += StringTable.size();
353 // ld64 expects the members to be 8-byte aligned for 64-bit content and at
354 // least 4-byte aligned for 32-bit content. Opt for the larger encoding
356 // We do this for all bsd formats because it simplifies aligning members.
357 unsigned Alignment = isBSDLike(Kind) ? 8 : 2;
358 unsigned Pad = OffsetToAlignment(Size, Alignment);
361 if (isBSDLike(Kind)) {
362 const char *Name = is64BitKind(Kind) ? "__.SYMDEF_64" : "__.SYMDEF";
363 printBSDMemberHeader(Out, Out.tell(), Name, now(Deterministic), 0, 0, 0,
366 const char *Name = is64BitKind(Kind) ? "/SYM64" : "";
367 printGNUSmallMemberHeader(Out, Name, now(Deterministic), 0, 0, 0, Size);
370 uint64_t Pos = Out.tell() + Size;
373 printNBits(Out, Kind, NumSyms * 2 * OffsetSize);
375 printNBits(Out, Kind, NumSyms);
377 for (const MemberData &M : Members) {
378 for (unsigned StringOffset : M.Symbols) {
380 printNBits(Out, Kind, StringOffset);
381 printNBits(Out, Kind, Pos); // member offset
383 Pos += M.Header.size() + M.Data.size() + M.Padding.size();
387 // byte count of the string table
388 printNBits(Out, Kind, StringTable.size());
392 Out.write(uint8_t(0));
395 static Expected<std::vector<unsigned>>
396 getSymbols(MemoryBufferRef Buf, raw_ostream &SymNames, bool &HasObject) {
397 std::vector<unsigned> Ret;
399 // In the scenario when LLVMContext is populated SymbolicFile will contain a
400 // reference to it, thus SymbolicFile should be destroyed first.
402 std::unique_ptr<object::SymbolicFile> Obj;
403 if (identify_magic(Buf.getBuffer()) == file_magic::bitcode) {
404 auto ObjOrErr = object::SymbolicFile::createSymbolicFile(
405 Buf, file_magic::bitcode, &Context);
407 // FIXME: check only for "not an object file" errors.
408 consumeError(ObjOrErr.takeError());
411 Obj = std::move(*ObjOrErr);
413 auto ObjOrErr = object::SymbolicFile::createSymbolicFile(Buf);
415 // FIXME: check only for "not an object file" errors.
416 consumeError(ObjOrErr.takeError());
419 Obj = std::move(*ObjOrErr);
423 for (const object::BasicSymbolRef &S : Obj->symbols()) {
424 if (!isArchiveSymbol(S))
426 Ret.push_back(SymNames.tell());
427 if (auto EC = S.printName(SymNames))
428 return errorCodeToError(EC);
434 static Expected<std::vector<MemberData>>
435 computeMemberData(raw_ostream &StringTable, raw_ostream &SymNames,
436 object::Archive::Kind Kind, bool Thin, StringRef ArcName,
437 bool Deterministic, ArrayRef<NewArchiveMember> NewMembers) {
438 static char PaddingData[8] = {'\n', '\n', '\n', '\n', '\n', '\n', '\n', '\n'};
440 // This ignores the symbol table, but we only need the value mod 8 and the
441 // symbol table is aligned to be a multiple of 8 bytes
444 std::vector<MemberData> Ret;
445 bool HasObject = false;
447 // Deduplicate long member names in the string table and reuse earlier name
448 // offsets. This especially saves space for COFF Import libraries where all
449 // members have the same name.
450 StringMap<uint64_t> MemberNames;
452 // UniqueTimestamps is a special case to improve debugging on Darwin:
454 // The Darwin linker does not link debug info into the final
455 // binary. Instead, it emits entries of type N_OSO in in the output
456 // binary's symbol table, containing references to the linked-in
457 // object files. Using that reference, the debugger can read the
458 // debug data directly from the object files. Alternatively, an
459 // invocation of 'dsymutil' will link the debug data from the object
460 // files into a dSYM bundle, which can be loaded by the debugger,
461 // instead of the object files.
463 // For an object file, the N_OSO entries contain the absolute path
464 // path to the file, and the file's timestamp. For an object
465 // included in an archive, the path is formatted like
466 // "/absolute/path/to/archive.a(member.o)", and the timestamp is the
467 // archive member's timestamp, rather than the archive's timestamp.
469 // However, this doesn't always uniquely identify an object within
470 // an archive -- an archive file can have multiple entries with the
471 // same filename. (This will happen commonly if the original object
472 // files started in different directories.) The only way they get
473 // distinguished, then, is via the timestamp. But this process is
474 // unable to find the correct object file in the archive when there
475 // are two files of the same name and timestamp.
477 // Additionally, timestamp==0 is treated specially, and causes the
478 // timestamp to be ignored as a match criteria.
480 // That will "usually" work out okay when creating an archive not in
481 // deterministic timestamp mode, because the objects will probably
482 // have been created at different timestamps.
484 // To ameliorate this problem, in deterministic archive mode (which
485 // is the default), on Darwin we will emit a unique non-zero
486 // timestamp for each entry with a duplicated name. This is still
487 // deterministic: the only thing affecting that timestamp is the
488 // order of the files in the resultant archive.
490 // See also the functions that handle the lookup:
491 // in lldb: ObjectContainerBSDArchive::Archive::FindObject()
492 // in llvm/tools/dsymutil: BinaryHolder::GetArchiveMemberBuffers().
493 bool UniqueTimestamps = Deterministic && isDarwin(Kind);
494 std::map<StringRef, unsigned> FilenameCount;
495 if (UniqueTimestamps) {
496 for (const NewArchiveMember &M : NewMembers)
497 FilenameCount[M.MemberName]++;
498 for (auto &Entry : FilenameCount)
499 Entry.second = Entry.second > 1 ? 1 : 0;
502 for (const NewArchiveMember &M : NewMembers) {
504 raw_string_ostream Out(Header);
506 MemoryBufferRef Buf = M.Buf->getMemBufferRef();
507 StringRef Data = Thin ? "" : Buf.getBuffer();
509 // ld64 expects the members to be 8-byte aligned for 64-bit content and at
510 // least 4-byte aligned for 32-bit content. Opt for the larger encoding
511 // uniformly. This matches the behaviour with cctools and ensures that ld64
512 // is happy with archives that we generate.
513 unsigned MemberPadding =
514 isDarwin(Kind) ? OffsetToAlignment(Data.size(), 8) : 0;
515 unsigned TailPadding = OffsetToAlignment(Data.size() + MemberPadding, 2);
516 StringRef Padding = StringRef(PaddingData, MemberPadding + TailPadding);
518 sys::TimePoint<std::chrono::seconds> ModTime;
519 if (UniqueTimestamps)
520 // Increment timestamp for each file of a given name.
521 ModTime = sys::toTimePoint(FilenameCount[M.MemberName]++);
524 printMemberHeader(Out, Pos, StringTable, MemberNames, Kind, Thin, ArcName,
525 M, ModTime, Buf.getBufferSize() + MemberPadding);
528 Expected<std::vector<unsigned>> Symbols =
529 getSymbols(Buf, SymNames, HasObject);
530 if (auto E = Symbols.takeError())
533 Pos += Header.size() + Data.size() + Padding.size();
534 Ret.push_back({std::move(*Symbols), std::move(Header), Data, Padding});
536 // If there are no symbols, emit an empty symbol table, to satisfy Solaris
537 // tools, older versions of which expect a symbol table in a non-empty
538 // archive, regardless of whether there are any symbols in it.
539 if (HasObject && SymNames.tell() == 0)
540 SymNames << '\0' << '\0' << '\0';
544 Error llvm::writeArchive(StringRef ArcName,
545 ArrayRef<NewArchiveMember> NewMembers,
546 bool WriteSymtab, object::Archive::Kind Kind,
547 bool Deterministic, bool Thin,
548 std::unique_ptr<MemoryBuffer> OldArchiveBuf) {
549 assert((!Thin || !isBSDLike(Kind)) && "Only the gnu format has a thin mode");
551 SmallString<0> SymNamesBuf;
552 raw_svector_ostream SymNames(SymNamesBuf);
553 SmallString<0> StringTableBuf;
554 raw_svector_ostream StringTable(StringTableBuf);
556 Expected<std::vector<MemberData>> DataOrErr = computeMemberData(
557 StringTable, SymNames, Kind, Thin, ArcName, Deterministic, NewMembers);
558 if (Error E = DataOrErr.takeError())
560 std::vector<MemberData> &Data = *DataOrErr;
562 if (!StringTableBuf.empty())
563 Data.insert(Data.begin(), computeStringTable(StringTableBuf));
565 // We would like to detect if we need to switch to a 64-bit symbol table.
567 uint64_t MaxOffset = 0;
568 uint64_t LastOffset = MaxOffset;
569 for (const auto &M : Data) {
570 // Record the start of the member's offset
571 LastOffset = MaxOffset;
572 // Account for the size of each part associated with the member.
573 MaxOffset += M.Header.size() + M.Data.size() + M.Padding.size();
574 // We assume 32-bit symbols to see if 32-bit symbols are possible or not.
575 MaxOffset += M.Symbols.size() * 4;
578 // The SYM64 format is used when an archive's member offsets are larger than
579 // 32-bits can hold. The need for this shift in format is detected by
580 // writeArchive. To test this we need to generate a file with a member that
581 // has an offset larger than 32-bits but this demands a very slow test. To
582 // speed the test up we use this environment variable to pretend like the
583 // cutoff happens before 32-bits and instead happens at some much smaller
585 const char *Sym64Env = std::getenv("SYM64_THRESHOLD");
586 int Sym64Threshold = 32;
588 StringRef(Sym64Env).getAsInteger(10, Sym64Threshold);
590 // If LastOffset isn't going to fit in a 32-bit varible we need to switch
591 // to 64-bit. Note that the file can be larger than 4GB as long as the last
592 // member starts before the 4GB offset.
593 if (LastOffset >= (1ULL << Sym64Threshold)) {
594 if (Kind == object::Archive::K_DARWIN)
595 Kind = object::Archive::K_DARWIN64;
597 Kind = object::Archive::K_GNU64;
601 Expected<sys::fs::TempFile> Temp =
602 sys::fs::TempFile::create(ArcName + ".temp-archive-%%%%%%%.a");
604 return Temp.takeError();
606 raw_fd_ostream Out(Temp->FD, false);
613 writeSymbolTable(Out, Kind, Deterministic, Data, SymNamesBuf);
615 for (const MemberData &M : Data)
616 Out << M.Header << M.Data << M.Padding;
620 // At this point, we no longer need whatever backing memory
621 // was used to generate the NewMembers. On Windows, this buffer
622 // could be a mapped view of the file we want to replace (if
623 // we're updating an existing archive, say). In that case, the
624 // rename would still succeed, but it would leave behind a
625 // temporary file (actually the original file renamed) because
626 // a file cannot be deleted while there's a handle open on it,
627 // only renamed. So by freeing this buffer, this ensures that
628 // the last open handle on the destination file, if any, is
629 // closed before we attempt to rename.
630 OldArchiveBuf.reset();
632 return Temp->keep(ArcName);