1 //===-- ArchiveWriter.cpp - Write LLVM archive files ----------------------===//
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 // Builds up an LLVM archive file (.a) containing LLVM bitcode.
12 //===----------------------------------------------------------------------===//
14 #include "ArchiveInternals.h"
15 #include "llvm/Bitcode/ReaderWriter.h"
16 #include "llvm/System/MappedFile.h"
17 #include "llvm/System/Signals.h"
18 #include "llvm/System/Process.h"
19 #include "llvm/ModuleProvider.h"
25 // Write an integer using variable bit rate encoding. This saves a few bytes
26 // per entry in the symbol table.
27 inline void writeInteger(unsigned num, std::ofstream& ARFile) {
29 if (num < 0x80) { // done?
30 ARFile << (unsigned char)num;
34 // Nope, we are bigger than a character, output the next 7 bits and set the
35 // high bit to say that there is more coming...
36 ARFile << (unsigned char)(0x80 | ((unsigned char)num & 0x7F));
37 num >>= 7; // Shift out 7 bits now...
41 // Compute how many bytes are taken by a given VBR encoded value. This is needed
42 // to pre-compute the size of the symbol table.
43 inline unsigned numVbrBytes(unsigned num) {
45 // Note that the following nested ifs are somewhat equivalent to a binary
46 // search. We split it in half by comparing against 2^14 first. This allows
47 // most reasonable values to be done in 2 comparisons instead of 1 for
48 // small ones and four for large ones. We expect this to access file offsets
49 // in the 2^10 to 2^24 range and symbol lengths in the 2^0 to 2^8 range,
50 // so this approach is reasonable.
62 return 5; // anything >= 2^28 takes 5 bytes
65 // Create an empty archive.
67 Archive::CreateEmpty(const sys::Path& FilePath ) {
68 Archive* result = new Archive(FilePath);
72 // Fill the ArchiveMemberHeader with the information from a member. If
73 // TruncateNames is true, names are flattened to 15 chars or less. The sz field
74 // is provided here instead of coming from the mbr because the member might be
75 // stored compressed and the compressed size is not the ArchiveMember's size.
76 // Furthermore compressed files have negative size fields to identify them as
79 Archive::fillHeader(const ArchiveMember &mbr, ArchiveMemberHeader& hdr,
80 int sz, bool TruncateNames) const {
82 // Set the permissions mode, uid and gid
85 sprintf(buffer, "%-8o", mbr.getMode());
86 memcpy(hdr.mode,buffer,8);
87 sprintf(buffer, "%-6u", mbr.getUser());
88 memcpy(hdr.uid,buffer,6);
89 sprintf(buffer, "%-6u", mbr.getGroup());
90 memcpy(hdr.gid,buffer,6);
92 // Set the last modification date
93 uint64_t secondsSinceEpoch = mbr.getModTime().toEpochTime();
94 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
95 memcpy(hdr.date,buffer,12);
97 // Get rid of trailing blanks in the name
98 std::string mbrPath = mbr.getPath().toString();
99 size_t mbrLen = mbrPath.length();
100 while (mbrLen > 0 && mbrPath[mbrLen-1] == ' ') {
101 mbrPath.erase(mbrLen-1,1);
105 // Set the name field in one of its various flavors.
106 bool writeLongName = false;
107 if (mbr.isStringTable()) {
108 memcpy(hdr.name,ARFILE_STRTAB_NAME,16);
109 } else if (mbr.isSVR4SymbolTable()) {
110 memcpy(hdr.name,ARFILE_SVR4_SYMTAB_NAME,16);
111 } else if (mbr.isBSD4SymbolTable()) {
112 memcpy(hdr.name,ARFILE_BSD4_SYMTAB_NAME,16);
113 } else if (mbr.isLLVMSymbolTable()) {
114 memcpy(hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
115 } else if (TruncateNames) {
116 const char* nm = mbrPath.c_str();
117 unsigned len = mbrPath.length();
118 size_t slashpos = mbrPath.rfind('/');
119 if (slashpos != std::string::npos) {
125 memcpy(hdr.name,nm,len);
127 } else if (mbrPath.length() < 16 && mbrPath.find('/') == std::string::npos) {
128 memcpy(hdr.name,mbrPath.c_str(),mbrPath.length());
129 hdr.name[mbrPath.length()] = '/';
131 std::string nm = "#1/";
132 nm += utostr(mbrPath.length());
133 memcpy(hdr.name,nm.data(),nm.length());
135 sz -= mbrPath.length();
137 sz += mbrPath.length();
138 writeLongName = true;
141 // Set the size field
144 sprintf(&buffer[1],"%-9u",(unsigned)-sz);
146 sprintf(buffer, "%-10u", (unsigned)sz);
148 memcpy(hdr.size,buffer,10);
150 return writeLongName;
153 // Insert a file into the archive before some other member. This also takes care
154 // of extracting the necessary flags and information from the file.
156 Archive::addFileBefore(const sys::Path& filePath, iterator where,
157 std::string* ErrMsg) {
158 if (!filePath.exists()) {
160 *ErrMsg = "Can not add a non-existent file to archive";
164 ArchiveMember* mbr = new ArchiveMember(this);
167 mbr->path = filePath;
168 const sys::FileStatus *FSInfo = mbr->path.getFileStatus(false, ErrMsg);
175 bool hasSlash = filePath.toString().find('/') != std::string::npos;
177 flags |= ArchiveMember::HasPathFlag;
178 if (hasSlash || filePath.toString().length() > 15)
179 flags |= ArchiveMember::HasLongFilenameFlag;
181 mbr->path.getMagicNumber(magic,4);
182 switch (sys::IdentifyFileType(magic.c_str(),4)) {
183 case sys::Bitcode_FileType:
184 flags |= ArchiveMember::BitcodeFlag;
190 members.insert(where,mbr);
194 // Write one member out to the file.
196 Archive::writeMember(
197 const ArchiveMember& member,
198 std::ofstream& ARFile,
199 bool CreateSymbolTable,
205 unsigned filepos = ARFile.tellp();
208 // Get the data and its size either from the
209 // member's in-memory data or directly from the file.
210 size_t fSize = member.getSize();
211 const char* data = (const char*)member.getData();
212 sys::MappedFile* mFile = 0;
214 mFile = new sys::MappedFile();
215 if (mFile->open(member.getPath(), sys::MappedFile::READ_ACCESS, ErrMsg))
217 if (!(data = (const char*) mFile->map(ErrMsg)))
219 fSize = mFile->size();
222 // Now that we have the data in memory, update the
223 // symbol table if its a bitcode file.
224 if (CreateSymbolTable && member.isBitcode()) {
225 std::vector<std::string> symbols;
226 std::string FullMemberName = archPath.toString() + "(" +
227 member.getPath().toString()
230 GetBitcodeSymbols((const unsigned char*)data,fSize,
231 FullMemberName, symbols, ErrMsg);
233 // If the bitcode parsed successfully
235 for (std::vector<std::string>::iterator SI = symbols.begin(),
236 SE = symbols.end(); SI != SE; ++SI) {
238 std::pair<SymTabType::iterator,bool> Res =
239 symTab.insert(std::make_pair(*SI,filepos));
242 symTabSize += SI->length() +
243 numVbrBytes(SI->length()) +
244 numVbrBytes(filepos);
247 // We don't need this module any more.
255 *ErrMsg = "Can't parse bitcode member: " + member.getPath().toString()
263 // Compute the fields of the header
264 ArchiveMemberHeader Hdr;
265 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
267 // Write header to archive file
268 ARFile.write((char*)&Hdr, sizeof(Hdr));
270 // Write the long filename if its long
272 ARFile.write(member.getPath().toString().data(),
273 member.getPath().toString().length());
276 // Write the (possibly compressed) member's content to the file.
277 ARFile.write(data,fSize);
279 // Make sure the member is an even length
280 if ((ARFile.tellp() & 1) == 1)
281 ARFile << ARFILE_PAD;
283 // Close the mapped file if it was opened
291 // Write out the LLVM symbol table as an archive member to the file.
293 Archive::writeSymbolTable(std::ofstream& ARFile) {
295 // Construct the symbol table's header
296 ArchiveMemberHeader Hdr;
298 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
299 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
301 sprintf(buffer, "%-8o", 0644);
302 memcpy(Hdr.mode,buffer,8);
303 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
304 memcpy(Hdr.uid,buffer,6);
305 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
306 memcpy(Hdr.gid,buffer,6);
307 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
308 memcpy(Hdr.date,buffer,12);
309 sprintf(buffer,"%-10u",symTabSize);
310 memcpy(Hdr.size,buffer,10);
313 ARFile.write((char*)&Hdr, sizeof(Hdr));
315 // Save the starting position of the symbol tables data content.
316 unsigned startpos = ARFile.tellp();
318 // Write out the symbols sequentially
319 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
322 // Write out the file index
323 writeInteger(I->second, ARFile);
324 // Write out the length of the symbol
325 writeInteger(I->first.length(), ARFile);
326 // Write out the symbol
327 ARFile.write(I->first.data(), I->first.length());
330 // Now that we're done with the symbol table, get the ending file position
331 unsigned endpos = ARFile.tellp();
333 // Make sure that the amount we wrote is what we pre-computed. This is
334 // critical for file integrity purposes.
335 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
337 // Make sure the symbol table is even sized
338 if (symTabSize % 2 != 0 )
339 ARFile << ARFILE_PAD;
342 // Write the entire archive to the file specified when the archive was created.
343 // This writes to a temporary file first. Options are for creating a symbol
344 // table, flattening the file names (no directories, 15 chars max) and
345 // compressing each archive member.
347 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
350 // Make sure they haven't opened up the file, not loaded it,
351 // but are now trying to write it which would wipe out the file.
352 if (members.empty() && mapfile && mapfile->size() > 8) {
354 *ErrMsg = "Can't write an archive not opened for writing";
358 // Create a temporary file to store the archive in
359 sys::Path TmpArchive = archPath;
360 if (TmpArchive.createTemporaryFileOnDisk(ErrMsg))
363 // Make sure the temporary gets removed if we crash
364 sys::RemoveFileOnSignal(TmpArchive);
366 // Create archive file for output.
367 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
369 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
371 // Check for errors opening or creating archive file.
372 if (!ArchiveFile.is_open() || ArchiveFile.bad()) {
373 if (TmpArchive.exists())
374 TmpArchive.eraseFromDisk();
376 *ErrMsg = "Error opening archive file: " + archPath.toString();
380 // If we're creating a symbol table, reset it now
381 if (CreateSymbolTable) {
386 // Write magic string to archive.
387 ArchiveFile << ARFILE_MAGIC;
389 // Loop over all member files, and write them out. Note that this also
390 // builds the symbol table, symTab.
391 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
392 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
393 TruncateNames, Compress, ErrMsg)) {
394 if (TmpArchive.exists())
395 TmpArchive.eraseFromDisk();
401 // Close archive file.
404 // Write the symbol table
405 if (CreateSymbolTable) {
406 // At this point we have written a file that is a legal archive but it
407 // doesn't have a symbol table in it. To aid in faster reading and to
408 // ensure compatibility with other archivers we need to put the symbol
409 // table first in the file. Unfortunately, this means mapping the file
410 // we just wrote back in and copying it to the destination file.
412 // Map in the archive we just wrote.
413 sys::MappedFile arch;
414 if (arch.open(TmpArchive, sys::MappedFile::READ_ACCESS, ErrMsg))
417 if (!(base = (const char*) arch.map(ErrMsg)))
420 // Open another temporary file in order to avoid invalidating the
422 sys::Path FinalFilePath = archPath;
423 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
425 sys::RemoveFileOnSignal(FinalFilePath);
427 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
428 if (!FinalFile.is_open() || FinalFile.bad()) {
429 if (TmpArchive.exists())
430 TmpArchive.eraseFromDisk();
432 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
436 // Write the file magic number
437 FinalFile << ARFILE_MAGIC;
439 // If there is a foreign symbol table, put it into the file now. Most
440 // ar(1) implementations require the symbol table to be first but llvm-ar
441 // can deal with it being after a foreign symbol table. This ensures
442 // compatibility with other ar(1) implementations as well as allowing the
443 // archive to store both native .o and LLVM .bc files, both indexed.
445 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
447 if (TmpArchive.exists())
448 TmpArchive.eraseFromDisk();
453 // Put out the LLVM symbol table now.
454 writeSymbolTable(FinalFile);
456 // Copy the temporary file contents being sure to skip the file's magic
458 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
459 arch.size()-sizeof(ARFILE_MAGIC)+1);
465 // Move the final file over top of TmpArchive
466 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
470 // Before we replace the actual archive, we need to forget all the
471 // members, since they point to data in that old archive. We need to do
472 // this because we cannot replace an open file on Windows.
475 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))