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/ADT/OwningPtr.h"
17 #include "llvm/Support/MemoryBuffer.h"
18 #include "llvm/System/Signals.h"
19 #include "llvm/System/Process.h"
20 #include "llvm/ModuleProvider.h"
26 // Write an integer using variable bit rate encoding. This saves a few bytes
27 // per entry in the symbol table.
28 static inline void writeInteger(unsigned num, std::ofstream& ARFile) {
30 if (num < 0x80) { // done?
31 ARFile << (unsigned char)num;
35 // Nope, we are bigger than a character, output the next 7 bits and set the
36 // high bit to say that there is more coming...
37 ARFile << (unsigned char)(0x80 | ((unsigned char)num & 0x7F));
38 num >>= 7; // Shift out 7 bits now...
42 // Compute how many bytes are taken by a given VBR encoded value. This is needed
43 // to pre-compute the size of the symbol table.
44 static inline unsigned numVbrBytes(unsigned num) {
46 // Note that the following nested ifs are somewhat equivalent to a binary
47 // search. We split it in half by comparing against 2^14 first. This allows
48 // most reasonable values to be done in 2 comparisons instead of 1 for
49 // small ones and four for large ones. We expect this to access file offsets
50 // in the 2^10 to 2^24 range and symbol lengths in the 2^0 to 2^8 range,
51 // so this approach is reasonable.
63 return 5; // anything >= 2^28 takes 5 bytes
66 // Create an empty archive.
68 Archive::CreateEmpty(const sys::Path& FilePath ) {
69 Archive* result = new Archive(FilePath);
73 // Fill the ArchiveMemberHeader with the information from a member. If
74 // TruncateNames is true, names are flattened to 15 chars or less. The sz field
75 // is provided here instead of coming from the mbr because the member might be
76 // stored compressed and the compressed size is not the ArchiveMember's size.
77 // Furthermore compressed files have negative size fields to identify them as
80 Archive::fillHeader(const ArchiveMember &mbr, ArchiveMemberHeader& hdr,
81 int sz, bool TruncateNames) const {
83 // Set the permissions mode, uid and gid
86 sprintf(buffer, "%-8o", mbr.getMode());
87 memcpy(hdr.mode,buffer,8);
88 sprintf(buffer, "%-6u", mbr.getUser());
89 memcpy(hdr.uid,buffer,6);
90 sprintf(buffer, "%-6u", mbr.getGroup());
91 memcpy(hdr.gid,buffer,6);
93 // Set the last modification date
94 uint64_t secondsSinceEpoch = mbr.getModTime().toEpochTime();
95 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
96 memcpy(hdr.date,buffer,12);
98 // Get rid of trailing blanks in the name
99 std::string mbrPath = mbr.getPath().toString();
100 size_t mbrLen = mbrPath.length();
101 while (mbrLen > 0 && mbrPath[mbrLen-1] == ' ') {
102 mbrPath.erase(mbrLen-1,1);
106 // Set the name field in one of its various flavors.
107 bool writeLongName = false;
108 if (mbr.isStringTable()) {
109 memcpy(hdr.name,ARFILE_STRTAB_NAME,16);
110 } else if (mbr.isSVR4SymbolTable()) {
111 memcpy(hdr.name,ARFILE_SVR4_SYMTAB_NAME,16);
112 } else if (mbr.isBSD4SymbolTable()) {
113 memcpy(hdr.name,ARFILE_BSD4_SYMTAB_NAME,16);
114 } else if (mbr.isLLVMSymbolTable()) {
115 memcpy(hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
116 } else if (TruncateNames) {
117 const char* nm = mbrPath.c_str();
118 unsigned len = mbrPath.length();
119 size_t slashpos = mbrPath.rfind('/');
120 if (slashpos != std::string::npos) {
126 memcpy(hdr.name,nm,len);
128 } else if (mbrPath.length() < 16 && mbrPath.find('/') == std::string::npos) {
129 memcpy(hdr.name,mbrPath.c_str(),mbrPath.length());
130 hdr.name[mbrPath.length()] = '/';
132 std::string nm = "#1/";
133 nm += utostr(mbrPath.length());
134 memcpy(hdr.name,nm.data(),nm.length());
136 sz -= mbrPath.length();
138 sz += mbrPath.length();
139 writeLongName = true;
142 // Set the size field
145 sprintf(&buffer[1],"%-9u",(unsigned)-sz);
147 sprintf(buffer, "%-10u", (unsigned)sz);
149 memcpy(hdr.size,buffer,10);
151 return writeLongName;
154 // Insert a file into the archive before some other member. This also takes care
155 // of extracting the necessary flags and information from the file.
157 Archive::addFileBefore(const sys::Path& filePath, iterator where,
158 std::string* ErrMsg) {
159 if (!filePath.exists()) {
161 *ErrMsg = "Can not add a non-existent file to archive";
165 ArchiveMember* mbr = new ArchiveMember(this);
168 mbr->path = filePath;
169 const sys::FileStatus *FSInfo = mbr->path.getFileStatus(false, ErrMsg);
177 bool hasSlash = filePath.toString().find('/') != std::string::npos;
179 flags |= ArchiveMember::HasPathFlag;
180 if (hasSlash || filePath.toString().length() > 15)
181 flags |= ArchiveMember::HasLongFilenameFlag;
183 mbr->path.getMagicNumber(magic,4);
184 switch (sys::IdentifyFileType(magic.c_str(),4)) {
185 case sys::Bitcode_FileType:
186 flags |= ArchiveMember::BitcodeFlag;
192 members.insert(where,mbr);
196 // Write one member out to the file.
198 Archive::writeMember(
199 const ArchiveMember& member,
200 std::ofstream& ARFile,
201 bool CreateSymbolTable,
207 unsigned filepos = ARFile.tellp();
210 // Get the data and its size either from the
211 // member's in-memory data or directly from the file.
212 size_t fSize = member.getSize();
213 const char *data = (const char*)member.getData();
214 MemoryBuffer *mFile = 0;
216 mFile = MemoryBuffer::getFile(member.getPath().c_str(), ErrMsg);
219 data = mFile->getBufferStart();
220 fSize = mFile->getBufferSize();
223 // Now that we have the data in memory, update the
224 // symbol table if its a bitcode file.
225 if (CreateSymbolTable && member.isBitcode()) {
226 std::vector<std::string> symbols;
227 std::string FullMemberName = archPath.toString() + "(" +
228 member.getPath().toString()
231 GetBitcodeSymbols((const unsigned char*)data,fSize,
232 FullMemberName, symbols, ErrMsg);
234 // If the bitcode parsed successfully
236 for (std::vector<std::string>::iterator SI = symbols.begin(),
237 SE = symbols.end(); SI != SE; ++SI) {
239 std::pair<SymTabType::iterator,bool> Res =
240 symTab.insert(std::make_pair(*SI,filepos));
243 symTabSize += SI->length() +
244 numVbrBytes(SI->length()) +
245 numVbrBytes(filepos);
248 // We don't need this module any more.
253 *ErrMsg = "Can't parse bitcode member: " + member.getPath().toString()
261 // Compute the fields of the header
262 ArchiveMemberHeader Hdr;
263 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
265 // Write header to archive file
266 ARFile.write((char*)&Hdr, sizeof(Hdr));
268 // Write the long filename if its long
270 ARFile.write(member.getPath().toString().data(),
271 member.getPath().toString().length());
274 // Write the (possibly compressed) member's content to the file.
275 ARFile.write(data,fSize);
277 // Make sure the member is an even length
278 if ((ARFile.tellp() & 1) == 1)
279 ARFile << ARFILE_PAD;
281 // Close the mapped file if it was opened
286 // Write out the LLVM symbol table as an archive member to the file.
288 Archive::writeSymbolTable(std::ofstream& ARFile) {
290 // Construct the symbol table's header
291 ArchiveMemberHeader Hdr;
293 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
294 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
296 sprintf(buffer, "%-8o", 0644);
297 memcpy(Hdr.mode,buffer,8);
298 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
299 memcpy(Hdr.uid,buffer,6);
300 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
301 memcpy(Hdr.gid,buffer,6);
302 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
303 memcpy(Hdr.date,buffer,12);
304 sprintf(buffer,"%-10u",symTabSize);
305 memcpy(Hdr.size,buffer,10);
308 ARFile.write((char*)&Hdr, sizeof(Hdr));
311 // Save the starting position of the symbol tables data content.
312 unsigned startpos = ARFile.tellp();
315 // Write out the symbols sequentially
316 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
319 // Write out the file index
320 writeInteger(I->second, ARFile);
321 // Write out the length of the symbol
322 writeInteger(I->first.length(), ARFile);
323 // Write out the symbol
324 ARFile.write(I->first.data(), I->first.length());
328 // Now that we're done with the symbol table, get the ending file position
329 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->getBufferSize() > 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.
410 sys::Path FinalFilePath = archPath;
412 // Map in the archive we just wrote.
414 OwningPtr<MemoryBuffer> arch(MemoryBuffer::getFile(TmpArchive.c_str()));
415 if (arch == 0) return true;
416 const char* base = arch->getBufferStart();
418 // Open another temporary file in order to avoid invalidating the
420 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
422 sys::RemoveFileOnSignal(FinalFilePath);
424 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
425 if (!FinalFile.is_open() || FinalFile.bad()) {
426 if (TmpArchive.exists())
427 TmpArchive.eraseFromDisk();
429 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
433 // Write the file magic number
434 FinalFile << ARFILE_MAGIC;
436 // If there is a foreign symbol table, put it into the file now. Most
437 // ar(1) implementations require the symbol table to be first but llvm-ar
438 // can deal with it being after a foreign symbol table. This ensures
439 // compatibility with other ar(1) implementations as well as allowing the
440 // archive to store both native .o and LLVM .bc files, both indexed.
442 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
444 if (TmpArchive.exists())
445 TmpArchive.eraseFromDisk();
450 // Put out the LLVM symbol table now.
451 writeSymbolTable(FinalFile);
453 // Copy the temporary file contents being sure to skip the file's magic
455 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
456 arch->getBufferSize()-sizeof(ARFILE_MAGIC)+1);
462 // Move the final file over top of TmpArchive
463 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
467 // Before we replace the actual archive, we need to forget all the
468 // members, since they point to data in that old archive. We need to do
469 // this because we cannot replace an open file on Windows.
472 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))
475 // Set correct read and write permissions after temporary file is moved
476 // to final destination path.
477 if (archPath.makeReadableOnDisk(ErrMsg))
479 if (archPath.makeWriteableOnDisk(ErrMsg))