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/Support/Compressor.h"
17 #include "llvm/System/Signals.h"
18 #include "llvm/System/Process.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::Bytecode_FileType:
182 flags |= ArchiveMember::BytecodeFlag;
184 case sys::CompressedBytecode_FileType:
185 flags |= ArchiveMember::CompressedBytecodeFlag;
191 members.insert(where,mbr);
195 // Write one member out to the file.
197 Archive::writeMember(
198 const ArchiveMember& member,
199 std::ofstream& ARFile,
200 bool CreateSymbolTable,
206 unsigned filepos = ARFile.tellp();
209 // Get the data and its size either from the
210 // member's in-memory data or directly from the file.
211 size_t fSize = member.getSize();
212 const char* data = (const char*)member.getData();
213 sys::MappedFile* mFile = 0;
215 mFile = new sys::MappedFile();
216 if (mFile->open(member.getPath(), sys::MappedFile::READ_ACCESS, ErrMsg))
218 if (!(data = (const char*) mFile->map(ErrMsg)))
220 fSize = mFile->size();
223 // Now that we have the data in memory, update the
224 // symbol table if its a bytecode file.
225 if (CreateSymbolTable &&
226 (member.isBytecode() || member.isCompressedBytecode())) {
227 std::vector<std::string> symbols;
228 std::string FullMemberName = archPath.toString() + "(" +
229 member.getPath().toString()
232 GetBytecodeSymbols((const unsigned char*)data,fSize,
233 FullMemberName, symbols,
234 Compressor::decompressToNewBuffer, ErrMsg);
236 // If the bytecode 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 bytecode member: " + member.getPath().toString()
264 // Determine if we actually should compress this member
267 !member.isCompressed() &&
268 !member.isCompressedBytecode() &&
269 !member.isLLVMSymbolTable() &&
270 !member.isSVR4SymbolTable() &&
271 !member.isBSD4SymbolTable());
273 // Perform the compression. Note that if the file is uncompressed bytecode
274 // then we turn the file into compressed bytecode rather than treating it as
275 // compressed data. This is necessary since it allows us to determine that the
276 // file contains bytecode instead of looking like a regular compressed data
277 // member. A compressed bytecode file has its content compressed but has a
278 // magic number of "llvc". This acounts for the +/-4 arithmetic in the code
283 if (member.isBytecode()) {
287 fSize = Compressor::compressToNewBuffer(data,fSize,output,ErrMsg);
291 if (member.isBytecode())
299 // Compute the fields of the header
300 ArchiveMemberHeader Hdr;
301 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
303 // Write header to archive file
304 ARFile.write((char*)&Hdr, sizeof(Hdr));
306 // Write the long filename if its long
308 ARFile.write(member.getPath().toString().data(),
309 member.getPath().toString().length());
312 // Make sure we write the compressed bytecode magic number if we should.
313 if (willCompress && member.isBytecode())
314 ARFile.write("llvc",4);
316 // Write the (possibly compressed) member's content to the file.
317 ARFile.write(data,fSize);
319 // Make sure the member is an even length
320 if ((ARFile.tellp() & 1) == 1)
321 ARFile << ARFILE_PAD;
323 // Free the compressed data, if necessary
328 // Close the mapped file if it was opened
336 // Write out the LLVM symbol table as an archive member to the file.
338 Archive::writeSymbolTable(std::ofstream& ARFile) {
340 // Construct the symbol table's header
341 ArchiveMemberHeader Hdr;
343 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
344 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
346 sprintf(buffer, "%-8o", 0644);
347 memcpy(Hdr.mode,buffer,8);
348 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
349 memcpy(Hdr.uid,buffer,6);
350 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
351 memcpy(Hdr.gid,buffer,6);
352 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
353 memcpy(Hdr.date,buffer,12);
354 sprintf(buffer,"%-10u",symTabSize);
355 memcpy(Hdr.size,buffer,10);
358 ARFile.write((char*)&Hdr, sizeof(Hdr));
360 // Save the starting position of the symbol tables data content.
361 unsigned startpos = ARFile.tellp();
363 // Write out the symbols sequentially
364 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
367 // Write out the file index
368 writeInteger(I->second, ARFile);
369 // Write out the length of the symbol
370 writeInteger(I->first.length(), ARFile);
371 // Write out the symbol
372 ARFile.write(I->first.data(), I->first.length());
375 // Now that we're done with the symbol table, get the ending file position
376 unsigned endpos = ARFile.tellp();
378 // Make sure that the amount we wrote is what we pre-computed. This is
379 // critical for file integrity purposes.
380 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
382 // Make sure the symbol table is even sized
383 if (symTabSize % 2 != 0 )
384 ARFile << ARFILE_PAD;
387 // Write the entire archive to the file specified when the archive was created.
388 // This writes to a temporary file first. Options are for creating a symbol
389 // table, flattening the file names (no directories, 15 chars max) and
390 // compressing each archive member.
392 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
395 // Make sure they haven't opened up the file, not loaded it,
396 // but are now trying to write it which would wipe out the file.
397 if (members.empty() && mapfile->size() > 8) {
399 *ErrMsg = "Can't write an archive not opened for writing";
403 // Create a temporary file to store the archive in
404 sys::Path TmpArchive = archPath;
405 if (TmpArchive.createTemporaryFileOnDisk(ErrMsg))
408 // Make sure the temporary gets removed if we crash
409 sys::RemoveFileOnSignal(TmpArchive);
411 // Create archive file for output.
412 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
414 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
416 // Check for errors opening or creating archive file.
417 if (!ArchiveFile.is_open() || ArchiveFile.bad()) {
418 if (TmpArchive.exists())
419 TmpArchive.eraseFromDisk();
421 *ErrMsg = "Error opening archive file: " + archPath.toString();
425 // If we're creating a symbol table, reset it now
426 if (CreateSymbolTable) {
431 // Write magic string to archive.
432 ArchiveFile << ARFILE_MAGIC;
434 // Loop over all member files, and write them out. Note that this also
435 // builds the symbol table, symTab.
436 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
437 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
438 TruncateNames, Compress, ErrMsg)) {
439 if (TmpArchive.exists())
440 TmpArchive.eraseFromDisk();
446 // Close archive file.
449 // Write the symbol table
450 if (CreateSymbolTable) {
451 // At this point we have written a file that is a legal archive but it
452 // doesn't have a symbol table in it. To aid in faster reading and to
453 // ensure compatibility with other archivers we need to put the symbol
454 // table first in the file. Unfortunately, this means mapping the file
455 // we just wrote back in and copying it to the destination file.
457 // Map in the archive we just wrote.
458 sys::MappedFile arch;
459 if (arch.open(TmpArchive, sys::MappedFile::READ_ACCESS, ErrMsg))
462 if (!(base = (const char*) arch.map(ErrMsg)))
465 // Open another temporary file in order to avoid invalidating the
467 sys::Path FinalFilePath = archPath;
468 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
470 sys::RemoveFileOnSignal(FinalFilePath);
472 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
473 if (!FinalFile.is_open() || FinalFile.bad()) {
474 if (TmpArchive.exists())
475 TmpArchive.eraseFromDisk();
477 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
481 // Write the file magic number
482 FinalFile << ARFILE_MAGIC;
484 // If there is a foreign symbol table, put it into the file now. Most
485 // ar(1) implementations require the symbol table to be first but llvm-ar
486 // can deal with it being after a foreign symbol table. This ensures
487 // compatibility with other ar(1) implementations as well as allowing the
488 // archive to store both native .o and LLVM .bc files, both indexed.
490 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
492 if (TmpArchive.exists())
493 TmpArchive.eraseFromDisk();
498 // Put out the LLVM symbol table now.
499 writeSymbolTable(FinalFile);
501 // Copy the temporary file contents being sure to skip the file's magic
503 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
504 arch.size()-sizeof(ARFILE_MAGIC)+1);
510 // Move the final file over top of TmpArchive
511 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
515 // Before we replace the actual archive, we need to forget all the
516 // members, since they point to data in that old archive. We need to do
517 // this because we cannot replace an open file on Windows.
520 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))