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/Module.h"
16 #include "llvm/ADT/OwningPtr.h"
17 #include "llvm/Bitcode/ReaderWriter.h"
18 #include "llvm/Support/FileSystem.h"
19 #include "llvm/Support/MemoryBuffer.h"
20 #include "llvm/Support/Process.h"
21 #include "llvm/Support/Signals.h"
22 #include "llvm/Support/system_error.h"
28 // Write an integer using variable bit rate encoding. This saves a few bytes
29 // per entry in the symbol table.
30 static inline void writeInteger(unsigned num, std::ofstream& ARFile) {
32 if (num < 0x80) { // done?
33 ARFile << (unsigned char)num;
37 // Nope, we are bigger than a character, output the next 7 bits and set the
38 // high bit to say that there is more coming...
39 ARFile << (unsigned char)(0x80 | ((unsigned char)num & 0x7F));
40 num >>= 7; // Shift out 7 bits now...
44 // Compute how many bytes are taken by a given VBR encoded value. This is needed
45 // to pre-compute the size of the symbol table.
46 static inline unsigned numVbrBytes(unsigned num) {
48 // Note that the following nested ifs are somewhat equivalent to a binary
49 // search. We split it in half by comparing against 2^14 first. This allows
50 // most reasonable values to be done in 2 comparisons instead of 1 for
51 // small ones and four for large ones. We expect this to access file offsets
52 // in the 2^10 to 2^24 range and symbol lengths in the 2^0 to 2^8 range,
53 // so this approach is reasonable.
65 return 5; // anything >= 2^28 takes 5 bytes
68 // Create an empty archive.
69 Archive* Archive::CreateEmpty(const sys::Path& FilePath, LLVMContext& C) {
70 Archive* result = new Archive(FilePath, C);
74 // Fill the ArchiveMemberHeader with the information from a member. If
75 // TruncateNames is true, names are flattened to 15 chars or less. The sz field
76 // is provided here instead of coming from the mbr because the member might be
77 // stored compressed and the compressed size is not the ArchiveMember's size.
78 // Furthermore compressed files have negative size fields to identify them as
81 Archive::fillHeader(const ArchiveMember &mbr, ArchiveMemberHeader& hdr,
82 int sz, bool TruncateNames) const {
84 // Set the permissions mode, uid and gid
87 sprintf(buffer, "%-8o", mbr.getMode());
88 memcpy(hdr.mode,buffer,8);
89 sprintf(buffer, "%-6u", mbr.getUser());
90 memcpy(hdr.uid,buffer,6);
91 sprintf(buffer, "%-6u", mbr.getGroup());
92 memcpy(hdr.gid,buffer,6);
94 // Set the last modification date
95 uint64_t secondsSinceEpoch = mbr.getModTime().toEpochTime();
96 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
97 memcpy(hdr.date,buffer,12);
99 // Get rid of trailing blanks in the name
100 std::string mbrPath = mbr.getPath().str();
101 size_t mbrLen = mbrPath.length();
102 while (mbrLen > 0 && mbrPath[mbrLen-1] == ' ') {
103 mbrPath.erase(mbrLen-1,1);
107 // Set the name field in one of its various flavors.
108 bool writeLongName = false;
109 if (mbr.isStringTable()) {
110 memcpy(hdr.name,ARFILE_STRTAB_NAME,16);
111 } else if (mbr.isSVR4SymbolTable()) {
112 memcpy(hdr.name,ARFILE_SVR4_SYMTAB_NAME,16);
113 } else if (mbr.isBSD4SymbolTable()) {
114 memcpy(hdr.name,ARFILE_BSD4_SYMTAB_NAME,16);
115 } else if (mbr.isLLVMSymbolTable()) {
116 memcpy(hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
117 } else if (TruncateNames) {
118 const char* nm = mbrPath.c_str();
119 unsigned len = mbrPath.length();
120 size_t slashpos = mbrPath.rfind('/');
121 if (slashpos != std::string::npos) {
127 memcpy(hdr.name,nm,len);
129 } else if (mbrPath.length() < 16 && mbrPath.find('/') == std::string::npos) {
130 memcpy(hdr.name,mbrPath.c_str(),mbrPath.length());
131 hdr.name[mbrPath.length()] = '/';
133 std::string nm = "#1/";
134 nm += utostr(mbrPath.length());
135 memcpy(hdr.name,nm.data(),nm.length());
137 sz -= mbrPath.length();
139 sz += mbrPath.length();
140 writeLongName = true;
143 // Set the size field
146 sprintf(&buffer[1],"%-9u",(unsigned)-sz);
148 sprintf(buffer, "%-10u", (unsigned)sz);
150 memcpy(hdr.size,buffer,10);
152 return writeLongName;
155 // Insert a file into the archive before some other member. This also takes care
156 // of extracting the necessary flags and information from the file.
158 Archive::addFileBefore(const sys::Path& filePath, iterator where,
159 std::string* ErrMsg) {
161 if (sys::fs::exists(filePath.str(), Exists) || !Exists) {
163 *ErrMsg = "Can not add a non-existent file to archive";
167 ArchiveMember* mbr = new ArchiveMember(this);
170 mbr->path = filePath;
171 const sys::FileStatus *FSInfo = mbr->path.getFileStatus(false, ErrMsg);
179 bool hasSlash = filePath.str().find('/') != std::string::npos;
181 flags |= ArchiveMember::HasPathFlag;
182 if (hasSlash || filePath.str().length() > 15)
183 flags |= ArchiveMember::HasLongFilenameFlag;
185 mbr->path.getMagicNumber(magic,4);
186 switch (sys::IdentifyFileType(magic.c_str(),4)) {
187 case sys::Bitcode_FileType:
188 flags |= ArchiveMember::BitcodeFlag;
194 members.insert(where,mbr);
198 // Write one member out to the file.
200 Archive::writeMember(
201 const ArchiveMember& member,
202 std::ofstream& ARFile,
203 bool CreateSymbolTable,
209 unsigned filepos = ARFile.tellp();
212 // Get the data and its size either from the
213 // member's in-memory data or directly from the file.
214 size_t fSize = member.getSize();
215 const char *data = (const char*)member.getData();
216 MemoryBuffer *mFile = 0;
218 OwningPtr<MemoryBuffer> File;
219 if (error_code ec = MemoryBuffer::getFile(member.getPath().c_str(), File)) {
221 *ErrMsg = ec.message();
225 data = mFile->getBufferStart();
226 fSize = mFile->getBufferSize();
229 // Now that we have the data in memory, update the
230 // symbol table if it's a bitcode file.
231 if (CreateSymbolTable && member.isBitcode()) {
232 std::vector<std::string> symbols;
233 std::string FullMemberName = archPath.str() + "(" + member.getPath().str()
236 GetBitcodeSymbols(data, fSize, FullMemberName, Context, symbols, ErrMsg);
238 // If the bitcode parsed successfully
240 for (std::vector<std::string>::iterator SI = symbols.begin(),
241 SE = symbols.end(); SI != SE; ++SI) {
243 std::pair<SymTabType::iterator,bool> Res =
244 symTab.insert(std::make_pair(*SI,filepos));
247 symTabSize += SI->length() +
248 numVbrBytes(SI->length()) +
249 numVbrBytes(filepos);
252 // We don't need this module any more.
257 *ErrMsg = "Can't parse bitcode member: " + member.getPath().str()
265 // Compute the fields of the header
266 ArchiveMemberHeader Hdr;
267 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
269 // Write header to archive file
270 ARFile.write((char*)&Hdr, sizeof(Hdr));
272 // Write the long filename if its long
274 ARFile.write(member.getPath().str().data(),
275 member.getPath().str().length());
278 // Write the (possibly compressed) member's content to the file.
279 ARFile.write(data,fSize);
281 // Make sure the member is an even length
282 if ((ARFile.tellp() & 1) == 1)
283 ARFile << ARFILE_PAD;
285 // 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));
315 // Save the starting position of the symbol tables data content.
316 unsigned startpos = ARFile.tellp();
319 // Write out the symbols sequentially
320 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
323 // Write out the file index
324 writeInteger(I->second, ARFile);
325 // Write out the length of the symbol
326 writeInteger(I->first.length(), ARFile);
327 // Write out the symbol
328 ARFile.write(I->first.data(), I->first.length());
332 // Now that we're done with the symbol table, get the ending file position
333 unsigned endpos = ARFile.tellp();
336 // Make sure that the amount we wrote is what we pre-computed. This is
337 // critical for file integrity purposes.
338 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
340 // Make sure the symbol table is even sized
341 if (symTabSize % 2 != 0 )
342 ARFile << ARFILE_PAD;
345 // Write the entire archive to the file specified when the archive was created.
346 // This writes to a temporary file first. Options are for creating a symbol
347 // table, flattening the file names (no directories, 15 chars max) and
348 // compressing each archive member.
350 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
353 // Make sure they haven't opened up the file, not loaded it,
354 // but are now trying to write it which would wipe out the file.
355 if (members.empty() && mapfile && mapfile->getBufferSize() > 8) {
357 *ErrMsg = "Can't write an archive not opened for writing";
361 // Create a temporary file to store the archive in
362 sys::Path TmpArchive = archPath;
363 if (TmpArchive.createTemporaryFileOnDisk(ErrMsg))
366 // Make sure the temporary gets removed if we crash
367 sys::RemoveFileOnSignal(TmpArchive);
369 // Create archive file for output.
370 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
372 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
374 // Check for errors opening or creating archive file.
375 if (!ArchiveFile.is_open() || ArchiveFile.bad()) {
376 TmpArchive.eraseFromDisk();
378 *ErrMsg = "Error opening archive file: " + archPath.str();
382 // If we're creating a symbol table, reset it now
383 if (CreateSymbolTable) {
388 // Write magic string to archive.
389 ArchiveFile << ARFILE_MAGIC;
391 // Loop over all member files, and write them out. Note that this also
392 // builds the symbol table, symTab.
393 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
394 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
395 TruncateNames, Compress, ErrMsg)) {
396 TmpArchive.eraseFromDisk();
402 // Close archive file.
405 // Write the symbol table
406 if (CreateSymbolTable) {
407 // At this point we have written a file that is a legal archive but it
408 // doesn't have a symbol table in it. To aid in faster reading and to
409 // ensure compatibility with other archivers we need to put the symbol
410 // table first in the file. Unfortunately, this means mapping the file
411 // we just wrote back in and copying it to the destination file.
412 sys::Path FinalFilePath = archPath;
414 // Map in the archive we just wrote.
416 OwningPtr<MemoryBuffer> arch;
417 if (error_code ec = MemoryBuffer::getFile(TmpArchive.c_str(), arch)) {
419 *ErrMsg = ec.message();
422 const char* base = arch->getBufferStart();
424 // Open another temporary file in order to avoid invalidating the
426 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
428 sys::RemoveFileOnSignal(FinalFilePath);
430 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
431 if (!FinalFile.is_open() || FinalFile.bad()) {
432 TmpArchive.eraseFromDisk();
434 *ErrMsg = "Error opening archive file: " + FinalFilePath.str();
438 // Write the file magic number
439 FinalFile << ARFILE_MAGIC;
441 // If there is a foreign symbol table, put it into the file now. Most
442 // ar(1) implementations require the symbol table to be first but llvm-ar
443 // can deal with it being after a foreign symbol table. This ensures
444 // compatibility with other ar(1) implementations as well as allowing the
445 // archive to store both native .o and LLVM .bc files, both indexed.
447 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
449 TmpArchive.eraseFromDisk();
454 // Put out the LLVM symbol table now.
455 writeSymbolTable(FinalFile);
457 // Copy the temporary file contents being sure to skip the file's magic
459 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
460 arch->getBufferSize()-sizeof(ARFILE_MAGIC)+1);
466 // Move the final file over top of TmpArchive
467 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
471 // Before we replace the actual archive, we need to forget all the
472 // members, since they point to data in that old archive. We need to do
473 // this because we cannot replace an open file on Windows.
476 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))
479 // Set correct read and write permissions after temporary file is moved
480 // to final destination path.
481 if (archPath.makeReadableOnDisk(ErrMsg))
483 if (archPath.makeWriteableOnDisk(ErrMsg))