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 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.
60 return 5; // anything >= 2^28 takes 5 bytes
63 // Create an empty archive.
65 Archive::CreateEmpty(const sys::Path& FilePath ) {
66 Archive* result = new Archive(FilePath);
70 // Fill the ArchiveMemberHeader with the information from a member. If
71 // TruncateNames is true, names are flattened to 15 chars or less. The sz field
72 // is provided here instead of coming from the mbr because the member might be
73 // stored compressed and the compressed size is not the ArchiveMember's size.
74 // Furthermore compressed files have negative size fields to identify them as
77 Archive::fillHeader(const ArchiveMember &mbr, ArchiveMemberHeader& hdr,
78 int sz, bool TruncateNames) const {
80 // Set the permissions mode, uid and gid
83 sprintf(buffer, "%-8o", mbr.getMode());
84 memcpy(hdr.mode,buffer,8);
85 sprintf(buffer, "%-6u", mbr.getUser());
86 memcpy(hdr.uid,buffer,6);
87 sprintf(buffer, "%-6u", mbr.getGroup());
88 memcpy(hdr.gid,buffer,6);
90 // Set the last modification date
91 uint64_t secondsSinceEpoch = mbr.getModTime().toEpochTime();
92 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
93 memcpy(hdr.date,buffer,12);
95 // Get rid of trailing blanks in the name
96 std::string mbrPath = mbr.getPath().toString();
97 size_t mbrLen = mbrPath.length();
98 while (mbrLen > 0 && mbrPath[mbrLen-1] == ' ') {
99 mbrPath.erase(mbrLen-1,1);
103 // Set the name field in one of its various flavors.
104 bool writeLongName = false;
105 if (mbr.isStringTable()) {
106 memcpy(hdr.name,ARFILE_STRTAB_NAME,16);
107 } else if (mbr.isSVR4SymbolTable()) {
108 memcpy(hdr.name,ARFILE_SVR4_SYMTAB_NAME,16);
109 } else if (mbr.isBSD4SymbolTable()) {
110 memcpy(hdr.name,ARFILE_BSD4_SYMTAB_NAME,16);
111 } else if (mbr.isLLVMSymbolTable()) {
112 memcpy(hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
113 } else if (TruncateNames) {
114 const char* nm = mbrPath.c_str();
115 unsigned len = mbrPath.length();
116 size_t slashpos = mbrPath.rfind('/');
117 if (slashpos != std::string::npos) {
123 memcpy(hdr.name,nm,len);
125 } else if (mbrPath.length() < 16 && mbrPath.find('/') == std::string::npos) {
126 memcpy(hdr.name,mbrPath.c_str(),mbrPath.length());
127 hdr.name[mbrPath.length()] = '/';
129 std::string nm = "#1/";
130 nm += utostr(mbrPath.length());
131 memcpy(hdr.name,nm.data(),nm.length());
133 sz -= mbrPath.length();
135 sz += mbrPath.length();
136 writeLongName = true;
139 // Set the size field
142 sprintf(&buffer[1],"%-9u",(unsigned)-sz);
144 sprintf(buffer, "%-10u", (unsigned)sz);
146 memcpy(hdr.size,buffer,10);
148 return writeLongName;
151 // Insert a file into the archive before some other member. This also takes care
152 // of extracting the necessary flags and information from the file.
154 Archive::addFileBefore(const sys::Path& filePath, iterator where,
155 std::string* ErrMsg) {
156 if (!filePath.exists()) {
158 *ErrMsg = "Can not add a non-existent file to archive";
162 ArchiveMember* mbr = new ArchiveMember(this);
165 mbr->path = filePath;
166 const sys::FileStatus *FSInfo = mbr->path.getFileStatus(false, ErrMsg);
173 bool hasSlash = filePath.toString().find('/') != std::string::npos;
175 flags |= ArchiveMember::HasPathFlag;
176 if (hasSlash || filePath.toString().length() > 15)
177 flags |= ArchiveMember::HasLongFilenameFlag;
179 mbr->path.getMagicNumber(magic,4);
180 switch (sys::IdentifyFileType(magic.c_str(),4)) {
181 case sys::Bitcode_FileType:
182 case sys::Bytecode_FileType:
183 flags |= ArchiveMember::BytecodeFlag;
185 case sys::CompressedBytecode_FileType:
186 flags |= ArchiveMember::CompressedBytecodeFlag;
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 sys::MappedFile* mFile = 0;
216 mFile = new sys::MappedFile();
217 if (mFile->open(member.getPath(), sys::MappedFile::READ_ACCESS, ErrMsg))
219 if (!(data = (const char*) mFile->map(ErrMsg)))
221 fSize = mFile->size();
224 // Now that we have the data in memory, update the
225 // symbol table if its a bitcode file.
226 if (CreateSymbolTable &&
227 (member.isBytecode() || member.isCompressedBytecode())) {
228 std::vector<std::string> symbols;
229 std::string FullMemberName = archPath.toString() + "(" +
230 member.getPath().toString()
233 GetBitcodeSymbols((const unsigned char*)data,fSize,
234 FullMemberName, symbols, ErrMsg);
236 // If the bitcode parsed successfully
238 for (std::vector<std::string>::iterator SI = symbols.begin(),
239 SE = symbols.end(); SI != SE; ++SI) {
241 std::pair<SymTabType::iterator,bool> Res =
242 symTab.insert(std::make_pair(*SI,filepos));
245 symTabSize += SI->length() +
246 numVbrBytes(SI->length()) +
247 numVbrBytes(filepos);
250 // We don't need this module any more.
258 *ErrMsg = "Can't parse bitcode member: " + member.getPath().toString()
266 // Compute the fields of the header
267 ArchiveMemberHeader Hdr;
268 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
270 // Write header to archive file
271 ARFile.write((char*)&Hdr, sizeof(Hdr));
273 // Write the long filename if its long
275 ARFile.write(member.getPath().toString().data(),
276 member.getPath().toString().length());
279 // Write the (possibly compressed) member's content to the file.
280 ARFile.write(data,fSize);
282 // Make sure the member is an even length
283 if ((ARFile.tellp() & 1) == 1)
284 ARFile << ARFILE_PAD;
286 // Close the mapped file if it was opened
294 // Write out the LLVM symbol table as an archive member to the file.
296 Archive::writeSymbolTable(std::ofstream& ARFile) {
298 // Construct the symbol table's header
299 ArchiveMemberHeader Hdr;
301 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
302 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
304 sprintf(buffer, "%-8o", 0644);
305 memcpy(Hdr.mode,buffer,8);
306 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
307 memcpy(Hdr.uid,buffer,6);
308 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
309 memcpy(Hdr.gid,buffer,6);
310 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
311 memcpy(Hdr.date,buffer,12);
312 sprintf(buffer,"%-10u",symTabSize);
313 memcpy(Hdr.size,buffer,10);
316 ARFile.write((char*)&Hdr, sizeof(Hdr));
318 // Save the starting position of the symbol tables data content.
319 unsigned startpos = ARFile.tellp();
321 // Write out the symbols sequentially
322 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
325 // Write out the file index
326 writeInteger(I->second, ARFile);
327 // Write out the length of the symbol
328 writeInteger(I->first.length(), ARFile);
329 // Write out the symbol
330 ARFile.write(I->first.data(), I->first.length());
333 // Now that we're done with the symbol table, get the ending file position
334 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->size() > 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 if (TmpArchive.exists())
377 TmpArchive.eraseFromDisk();
379 *ErrMsg = "Error opening archive file: " + archPath.toString();
383 // If we're creating a symbol table, reset it now
384 if (CreateSymbolTable) {
389 // Write magic string to archive.
390 ArchiveFile << ARFILE_MAGIC;
392 // Loop over all member files, and write them out. Note that this also
393 // builds the symbol table, symTab.
394 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
395 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
396 TruncateNames, Compress, ErrMsg)) {
397 if (TmpArchive.exists())
398 TmpArchive.eraseFromDisk();
404 // Close archive file.
407 // Write the symbol table
408 if (CreateSymbolTable) {
409 // At this point we have written a file that is a legal archive but it
410 // doesn't have a symbol table in it. To aid in faster reading and to
411 // ensure compatibility with other archivers we need to put the symbol
412 // table first in the file. Unfortunately, this means mapping the file
413 // we just wrote back in and copying it to the destination file.
415 // Map in the archive we just wrote.
416 sys::MappedFile arch;
417 if (arch.open(TmpArchive, sys::MappedFile::READ_ACCESS, ErrMsg))
420 if (!(base = (const char*) arch.map(ErrMsg)))
423 // Open another temporary file in order to avoid invalidating the
425 sys::Path FinalFilePath = archPath;
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 if (TmpArchive.exists())
433 TmpArchive.eraseFromDisk();
435 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
439 // Write the file magic number
440 FinalFile << ARFILE_MAGIC;
442 // If there is a foreign symbol table, put it into the file now. Most
443 // ar(1) implementations require the symbol table to be first but llvm-ar
444 // can deal with it being after a foreign symbol table. This ensures
445 // compatibility with other ar(1) implementations as well as allowing the
446 // archive to store both native .o and LLVM .bc files, both indexed.
448 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
450 if (TmpArchive.exists())
451 TmpArchive.eraseFromDisk();
456 // Put out the LLVM symbol table now.
457 writeSymbolTable(FinalFile);
459 // Copy the temporary file contents being sure to skip the file's magic
461 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
462 arch.size()-sizeof(ARFILE_MAGIC)+1);
468 // Move the final file over top of TmpArchive
469 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
473 // Before we replace the actual archive, we need to forget all the
474 // members, since they point to data in that old archive. We need to do
475 // this because we cannot replace an open file on Windows.
478 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))