1 //===-- ArchiveWriter.cpp - Write LLVM archive files ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file was developed by Reid Spencer and is distributed under the
6 // University of Illinois Open Source License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Builds up an LLVM archive file (.a) containing LLVM bytecode.
12 //===----------------------------------------------------------------------===//
14 #include "ArchiveInternals.h"
15 #include "llvm/Bytecode/Reader.h"
16 #include "llvm/Bitcode/ReaderWriter.h"
17 #include "llvm/Support/Compressor.h"
18 #include "llvm/System/Signals.h"
19 #include "llvm/System/Process.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.
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 case sys::Bytecode_FileType:
184 flags |= ArchiveMember::BytecodeFlag;
186 case sys::CompressedBytecode_FileType:
187 flags |= ArchiveMember::CompressedBytecodeFlag;
193 members.insert(where,mbr);
197 // Write one member out to the file.
199 Archive::writeMember(
200 const ArchiveMember& member,
201 std::ofstream& ARFile,
202 bool CreateSymbolTable,
208 unsigned filepos = ARFile.tellp();
211 // Get the data and its size either from the
212 // member's in-memory data or directly from the file.
213 size_t fSize = member.getSize();
214 const char* data = (const char*)member.getData();
215 sys::MappedFile* mFile = 0;
217 mFile = new sys::MappedFile();
218 if (mFile->open(member.getPath(), sys::MappedFile::READ_ACCESS, ErrMsg))
220 if (!(data = (const char*) mFile->map(ErrMsg)))
222 fSize = mFile->size();
225 // Now that we have the data in memory, update the
226 // symbol table if its a bytecode file.
227 if (CreateSymbolTable &&
228 (member.isBytecode() || member.isCompressedBytecode())) {
229 std::vector<std::string> symbols;
230 std::string FullMemberName = archPath.toString() + "(" +
231 member.getPath().toString()
234 GetBytecodeSymbols((const unsigned char*)data,fSize,
235 FullMemberName, symbols,
236 Compressor::decompressToNewBuffer, ErrMsg);
238 // If the bytecode 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.
260 *ErrMsg = "Can't parse bytecode member: " + member.getPath().toString()
268 // Compute the fields of the header
269 ArchiveMemberHeader Hdr;
270 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
272 // Write header to archive file
273 ARFile.write((char*)&Hdr, sizeof(Hdr));
275 // Write the long filename if its long
277 ARFile.write(member.getPath().toString().data(),
278 member.getPath().toString().length());
281 // Write the (possibly compressed) member's content to the file.
282 ARFile.write(data,fSize);
284 // Make sure the member is an even length
285 if ((ARFile.tellp() & 1) == 1)
286 ARFile << ARFILE_PAD;
288 // Close the mapped file if it was opened
296 // Write out the LLVM symbol table as an archive member to the file.
298 Archive::writeSymbolTable(std::ofstream& ARFile) {
300 // Construct the symbol table's header
301 ArchiveMemberHeader Hdr;
303 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
304 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
306 sprintf(buffer, "%-8o", 0644);
307 memcpy(Hdr.mode,buffer,8);
308 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
309 memcpy(Hdr.uid,buffer,6);
310 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
311 memcpy(Hdr.gid,buffer,6);
312 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
313 memcpy(Hdr.date,buffer,12);
314 sprintf(buffer,"%-10u",symTabSize);
315 memcpy(Hdr.size,buffer,10);
318 ARFile.write((char*)&Hdr, sizeof(Hdr));
320 // Save the starting position of the symbol tables data content.
321 unsigned startpos = ARFile.tellp();
323 // Write out the symbols sequentially
324 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
327 // Write out the file index
328 writeInteger(I->second, ARFile);
329 // Write out the length of the symbol
330 writeInteger(I->first.length(), ARFile);
331 // Write out the symbol
332 ARFile.write(I->first.data(), I->first.length());
335 // Now that we're done with the symbol table, get the ending file position
336 unsigned endpos = ARFile.tellp();
338 // Make sure that the amount we wrote is what we pre-computed. This is
339 // critical for file integrity purposes.
340 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
342 // Make sure the symbol table is even sized
343 if (symTabSize % 2 != 0 )
344 ARFile << ARFILE_PAD;
347 // Write the entire archive to the file specified when the archive was created.
348 // This writes to a temporary file first. Options are for creating a symbol
349 // table, flattening the file names (no directories, 15 chars max) and
350 // compressing each archive member.
352 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
355 // Make sure they haven't opened up the file, not loaded it,
356 // but are now trying to write it which would wipe out the file.
357 if (members.empty() && mapfile->size() > 8) {
359 *ErrMsg = "Can't write an archive not opened for writing";
363 // Create a temporary file to store the archive in
364 sys::Path TmpArchive = archPath;
365 if (TmpArchive.createTemporaryFileOnDisk(ErrMsg))
368 // Make sure the temporary gets removed if we crash
369 sys::RemoveFileOnSignal(TmpArchive);
371 // Create archive file for output.
372 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
374 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
376 // Check for errors opening or creating archive file.
377 if (!ArchiveFile.is_open() || ArchiveFile.bad()) {
378 if (TmpArchive.exists())
379 TmpArchive.eraseFromDisk();
381 *ErrMsg = "Error opening archive file: " + archPath.toString();
385 // If we're creating a symbol table, reset it now
386 if (CreateSymbolTable) {
391 // Write magic string to archive.
392 ArchiveFile << ARFILE_MAGIC;
394 // Loop over all member files, and write them out. Note that this also
395 // builds the symbol table, symTab.
396 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
397 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
398 TruncateNames, Compress, ErrMsg)) {
399 if (TmpArchive.exists())
400 TmpArchive.eraseFromDisk();
406 // Close archive file.
409 // Write the symbol table
410 if (CreateSymbolTable) {
411 // At this point we have written a file that is a legal archive but it
412 // doesn't have a symbol table in it. To aid in faster reading and to
413 // ensure compatibility with other archivers we need to put the symbol
414 // table first in the file. Unfortunately, this means mapping the file
415 // we just wrote back in and copying it to the destination file.
417 // Map in the archive we just wrote.
418 sys::MappedFile arch;
419 if (arch.open(TmpArchive, sys::MappedFile::READ_ACCESS, ErrMsg))
422 if (!(base = (const char*) arch.map(ErrMsg)))
425 // Open another temporary file in order to avoid invalidating the
427 sys::Path FinalFilePath = archPath;
428 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
430 sys::RemoveFileOnSignal(FinalFilePath);
432 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
433 if (!FinalFile.is_open() || FinalFile.bad()) {
434 if (TmpArchive.exists())
435 TmpArchive.eraseFromDisk();
437 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
441 // Write the file magic number
442 FinalFile << ARFILE_MAGIC;
444 // If there is a foreign symbol table, put it into the file now. Most
445 // ar(1) implementations require the symbol table to be first but llvm-ar
446 // can deal with it being after a foreign symbol table. This ensures
447 // compatibility with other ar(1) implementations as well as allowing the
448 // archive to store both native .o and LLVM .bc files, both indexed.
450 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
452 if (TmpArchive.exists())
453 TmpArchive.eraseFromDisk();
458 // Put out the LLVM symbol table now.
459 writeSymbolTable(FinalFile);
461 // Copy the temporary file contents being sure to skip the file's magic
463 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
464 arch.size()-sizeof(ARFILE_MAGIC)+1);
470 // Move the final file over top of TmpArchive
471 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
475 // Before we replace the actual archive, we need to forget all the
476 // members, since they point to data in that old archive. We need to do
477 // this because we cannot replace an open file on Windows.
480 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))