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/Signals.h"
17 #include "llvm/System/Process.h"
18 #include "llvm/ModuleProvider.h"
24 // Write an integer using variable bit rate encoding. This saves a few bytes
25 // per entry in the symbol table.
26 inline void writeInteger(unsigned num, std::ofstream& ARFile) {
28 if (num < 0x80) { // done?
29 ARFile << (unsigned char)num;
33 // Nope, we are bigger than a character, output the next 7 bits and set the
34 // high bit to say that there is more coming...
35 ARFile << (unsigned char)(0x80 | ((unsigned char)num & 0x7F));
36 num >>= 7; // Shift out 7 bits now...
40 // Compute how many bytes are taken by a given VBR encoded value. This is needed
41 // to pre-compute the size of the symbol table.
42 inline unsigned numVbrBytes(unsigned num) {
44 // Note that the following nested ifs are somewhat equivalent to a binary
45 // search. We split it in half by comparing against 2^14 first. This allows
46 // most reasonable values to be done in 2 comparisons instead of 1 for
47 // small ones and four for large ones. We expect this to access file offsets
48 // in the 2^10 to 2^24 range and symbol lengths in the 2^0 to 2^8 range,
49 // so this approach is reasonable.
61 return 5; // anything >= 2^28 takes 5 bytes
64 // Create an empty archive.
66 Archive::CreateEmpty(const sys::Path& FilePath ) {
67 Archive* result = new Archive(FilePath);
71 // Fill the ArchiveMemberHeader with the information from a member. If
72 // TruncateNames is true, names are flattened to 15 chars or less. The sz field
73 // is provided here instead of coming from the mbr because the member might be
74 // stored compressed and the compressed size is not the ArchiveMember's size.
75 // Furthermore compressed files have negative size fields to identify them as
78 Archive::fillHeader(const ArchiveMember &mbr, ArchiveMemberHeader& hdr,
79 int sz, bool TruncateNames) const {
81 // Set the permissions mode, uid and gid
84 sprintf(buffer, "%-8o", mbr.getMode());
85 memcpy(hdr.mode,buffer,8);
86 sprintf(buffer, "%-6u", mbr.getUser());
87 memcpy(hdr.uid,buffer,6);
88 sprintf(buffer, "%-6u", mbr.getGroup());
89 memcpy(hdr.gid,buffer,6);
91 // Set the last modification date
92 uint64_t secondsSinceEpoch = mbr.getModTime().toEpochTime();
93 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
94 memcpy(hdr.date,buffer,12);
96 // Get rid of trailing blanks in the name
97 std::string mbrPath = mbr.getPath().toString();
98 size_t mbrLen = mbrPath.length();
99 while (mbrLen > 0 && mbrPath[mbrLen-1] == ' ') {
100 mbrPath.erase(mbrLen-1,1);
104 // Set the name field in one of its various flavors.
105 bool writeLongName = false;
106 if (mbr.isStringTable()) {
107 memcpy(hdr.name,ARFILE_STRTAB_NAME,16);
108 } else if (mbr.isSVR4SymbolTable()) {
109 memcpy(hdr.name,ARFILE_SVR4_SYMTAB_NAME,16);
110 } else if (mbr.isBSD4SymbolTable()) {
111 memcpy(hdr.name,ARFILE_BSD4_SYMTAB_NAME,16);
112 } else if (mbr.isLLVMSymbolTable()) {
113 memcpy(hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
114 } else if (TruncateNames) {
115 const char* nm = mbrPath.c_str();
116 unsigned len = mbrPath.length();
117 size_t slashpos = mbrPath.rfind('/');
118 if (slashpos != std::string::npos) {
124 memcpy(hdr.name,nm,len);
126 } else if (mbrPath.length() < 16 && mbrPath.find('/') == std::string::npos) {
127 memcpy(hdr.name,mbrPath.c_str(),mbrPath.length());
128 hdr.name[mbrPath.length()] = '/';
130 std::string nm = "#1/";
131 nm += utostr(mbrPath.length());
132 memcpy(hdr.name,nm.data(),nm.length());
134 sz -= mbrPath.length();
136 sz += mbrPath.length();
137 writeLongName = true;
140 // Set the size field
143 sprintf(&buffer[1],"%-9u",(unsigned)-sz);
145 sprintf(buffer, "%-10u", (unsigned)sz);
147 memcpy(hdr.size,buffer,10);
149 return writeLongName;
152 // Insert a file into the archive before some other member. This also takes care
153 // of extracting the necessary flags and information from the file.
155 Archive::addFileBefore(const sys::Path& filePath, iterator where,
156 std::string* ErrMsg) {
157 if (!filePath.exists()) {
159 *ErrMsg = "Can not add a non-existent file to archive";
163 ArchiveMember* mbr = new ArchiveMember(this);
166 mbr->path = filePath;
167 const sys::FileStatus *FSInfo = mbr->path.getFileStatus(false, ErrMsg);
174 bool hasSlash = filePath.toString().find('/') != std::string::npos;
176 flags |= ArchiveMember::HasPathFlag;
177 if (hasSlash || filePath.toString().length() > 15)
178 flags |= ArchiveMember::HasLongFilenameFlag;
180 mbr->path.getMagicNumber(magic,4);
181 switch (sys::IdentifyFileType(magic.c_str(),4)) {
182 case sys::Bitcode_FileType:
183 flags |= ArchiveMember::BitcodeFlag;
189 members.insert(where,mbr);
193 // Write one member out to the file.
195 Archive::writeMember(
196 const ArchiveMember& member,
197 std::ofstream& ARFile,
198 bool CreateSymbolTable,
204 unsigned filepos = ARFile.tellp();
207 // Get the data and its size either from the
208 // member's in-memory data or directly from the file.
209 size_t fSize = member.getSize();
210 const char* data = (const char*)member.getData();
211 sys::MappedFile* mFile = 0;
213 mFile = new sys::MappedFile();
214 if (mFile->open(member.getPath(), sys::MappedFile::READ_ACCESS, ErrMsg))
216 if (!(data = (const char*) mFile->map(ErrMsg)))
218 fSize = mFile->size();
221 // Now that we have the data in memory, update the
222 // symbol table if its a bitcode file.
223 if (CreateSymbolTable && member.isBitcode()) {
224 std::vector<std::string> symbols;
225 std::string FullMemberName = archPath.toString() + "(" +
226 member.getPath().toString()
229 GetBitcodeSymbols((const unsigned char*)data,fSize,
230 FullMemberName, symbols, ErrMsg);
232 // If the bitcode parsed successfully
234 for (std::vector<std::string>::iterator SI = symbols.begin(),
235 SE = symbols.end(); SI != SE; ++SI) {
237 std::pair<SymTabType::iterator,bool> Res =
238 symTab.insert(std::make_pair(*SI,filepos));
241 symTabSize += SI->length() +
242 numVbrBytes(SI->length()) +
243 numVbrBytes(filepos);
246 // We don't need this module any more.
254 *ErrMsg = "Can't parse bitcode member: " + member.getPath().toString()
262 // Compute the fields of the header
263 ArchiveMemberHeader Hdr;
264 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
266 // Write header to archive file
267 ARFile.write((char*)&Hdr, sizeof(Hdr));
269 // Write the long filename if its long
271 ARFile.write(member.getPath().toString().data(),
272 member.getPath().toString().length());
275 // Write the (possibly compressed) member's content to the file.
276 ARFile.write(data,fSize);
278 // Make sure the member is an even length
279 if ((ARFile.tellp() & 1) == 1)
280 ARFile << ARFILE_PAD;
282 // Close the mapped file if it was opened
290 // Write out the LLVM symbol table as an archive member to the file.
292 Archive::writeSymbolTable(std::ofstream& ARFile) {
294 // Construct the symbol table's header
295 ArchiveMemberHeader Hdr;
297 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
298 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
300 sprintf(buffer, "%-8o", 0644);
301 memcpy(Hdr.mode,buffer,8);
302 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
303 memcpy(Hdr.uid,buffer,6);
304 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
305 memcpy(Hdr.gid,buffer,6);
306 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
307 memcpy(Hdr.date,buffer,12);
308 sprintf(buffer,"%-10u",symTabSize);
309 memcpy(Hdr.size,buffer,10);
312 ARFile.write((char*)&Hdr, sizeof(Hdr));
314 // Save the starting position of the symbol tables data content.
315 unsigned startpos = ARFile.tellp();
317 // Write out the symbols sequentially
318 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
321 // Write out the file index
322 writeInteger(I->second, ARFile);
323 // Write out the length of the symbol
324 writeInteger(I->first.length(), ARFile);
325 // Write out the symbol
326 ARFile.write(I->first.data(), I->first.length());
329 // Now that we're done with the symbol table, get the ending file position
330 unsigned endpos = ARFile.tellp();
332 // Make sure that the amount we wrote is what we pre-computed. This is
333 // critical for file integrity purposes.
334 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
336 // Make sure the symbol table is even sized
337 if (symTabSize % 2 != 0 )
338 ARFile << ARFILE_PAD;
341 // Write the entire archive to the file specified when the archive was created.
342 // This writes to a temporary file first. Options are for creating a symbol
343 // table, flattening the file names (no directories, 15 chars max) and
344 // compressing each archive member.
346 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
349 // Make sure they haven't opened up the file, not loaded it,
350 // but are now trying to write it which would wipe out the file.
351 if (members.empty() && mapfile && mapfile->size() > 8) {
353 *ErrMsg = "Can't write an archive not opened for writing";
357 // Create a temporary file to store the archive in
358 sys::Path TmpArchive = archPath;
359 if (TmpArchive.createTemporaryFileOnDisk(ErrMsg))
362 // Make sure the temporary gets removed if we crash
363 sys::RemoveFileOnSignal(TmpArchive);
365 // Create archive file for output.
366 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
368 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
370 // Check for errors opening or creating archive file.
371 if (!ArchiveFile.is_open() || ArchiveFile.bad()) {
372 if (TmpArchive.exists())
373 TmpArchive.eraseFromDisk();
375 *ErrMsg = "Error opening archive file: " + archPath.toString();
379 // If we're creating a symbol table, reset it now
380 if (CreateSymbolTable) {
385 // Write magic string to archive.
386 ArchiveFile << ARFILE_MAGIC;
388 // Loop over all member files, and write them out. Note that this also
389 // builds the symbol table, symTab.
390 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
391 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
392 TruncateNames, Compress, ErrMsg)) {
393 if (TmpArchive.exists())
394 TmpArchive.eraseFromDisk();
400 // Close archive file.
403 // Write the symbol table
404 if (CreateSymbolTable) {
405 // At this point we have written a file that is a legal archive but it
406 // doesn't have a symbol table in it. To aid in faster reading and to
407 // ensure compatibility with other archivers we need to put the symbol
408 // table first in the file. Unfortunately, this means mapping the file
409 // we just wrote back in and copying it to the destination file.
411 // Map in the archive we just wrote.
412 sys::MappedFile arch;
413 if (arch.open(TmpArchive, sys::MappedFile::READ_ACCESS, ErrMsg))
416 if (!(base = (const char*) arch.map(ErrMsg)))
419 // Open another temporary file in order to avoid invalidating the
421 sys::Path FinalFilePath = archPath;
422 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
424 sys::RemoveFileOnSignal(FinalFilePath);
426 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
427 if (!FinalFile.is_open() || FinalFile.bad()) {
428 if (TmpArchive.exists())
429 TmpArchive.eraseFromDisk();
431 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
435 // Write the file magic number
436 FinalFile << ARFILE_MAGIC;
438 // If there is a foreign symbol table, put it into the file now. Most
439 // ar(1) implementations require the symbol table to be first but llvm-ar
440 // can deal with it being after a foreign symbol table. This ensures
441 // compatibility with other ar(1) implementations as well as allowing the
442 // archive to store both native .o and LLVM .bc files, both indexed.
444 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
446 if (TmpArchive.exists())
447 TmpArchive.eraseFromDisk();
452 // Put out the LLVM symbol table now.
453 writeSymbolTable(FinalFile);
455 // Copy the temporary file contents being sure to skip the file's magic
457 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
458 arch.size()-sizeof(ARFILE_MAGIC)+1);
464 // Move the final file over top of TmpArchive
465 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
469 // Before we replace the actual archive, we need to forget all the
470 // members, since they point to data in that old archive. We need to do
471 // this because we cannot replace an open file on Windows.
474 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))