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"
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,false);
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 assert(filePath.exists() && "Can't add a non-existent file");
158 ArchiveMember* mbr = new ArchiveMember(this);
161 mbr->path = filePath;
162 mbr->path.getStatusInfo(mbr->info);
165 bool hasSlash = filePath.toString().find('/') != std::string::npos;
167 flags |= ArchiveMember::HasPathFlag;
168 if (hasSlash || filePath.toString().length() > 15)
169 flags |= ArchiveMember::HasLongFilenameFlag;
171 mbr->path.getMagicNumber(magic,4);
172 switch (sys::IdentifyFileType(magic.c_str(),4)) {
173 case sys::BytecodeFileType:
174 flags |= ArchiveMember::BytecodeFlag;
176 case sys::CompressedBytecodeFileType:
177 flags |= ArchiveMember::CompressedBytecodeFlag;
183 members.insert(where,mbr);
186 // Write one member out to the file.
188 Archive::writeMember(
189 const ArchiveMember& member,
190 std::ofstream& ARFile,
191 bool CreateSymbolTable,
197 unsigned filepos = ARFile.tellp();
200 // Get the data and its size either from the
201 // member's in-memory data or directly from the file.
202 size_t fSize = member.getSize();
203 const char* data = (const char*)member.getData();
204 sys::MappedFile* mFile = 0;
206 mFile = new sys::MappedFile(member.getPath());
207 data = (const char*) mFile->map();
208 fSize = mFile->size();
211 // Now that we have the data in memory, update the
212 // symbol table if its a bytecode file.
213 if (CreateSymbolTable &&
214 (member.isBytecode() || member.isCompressedBytecode())) {
215 std::vector<std::string> symbols;
216 std::string FullMemberName = archPath.toString() + "(" +
217 member.getPath().toString()
219 ModuleProvider* MP = GetBytecodeSymbols(
220 (const unsigned char*)data,fSize,FullMemberName, symbols);
222 // If the bytecode parsed successfully
224 for (std::vector<std::string>::iterator SI = symbols.begin(),
225 SE = symbols.end(); SI != SE; ++SI) {
227 std::pair<SymTabType::iterator,bool> Res =
228 symTab.insert(std::make_pair(*SI,filepos));
231 symTabSize += SI->length() +
232 numVbrBytes(SI->length()) +
233 numVbrBytes(filepos);
236 // We don't need this module any more.
244 *error = "Can't parse bytecode member: " + member.getPath().toString();
248 // Determine if we actually should compress this member
251 !member.isCompressed() &&
252 !member.isCompressedBytecode() &&
253 !member.isLLVMSymbolTable() &&
254 !member.isSVR4SymbolTable() &&
255 !member.isBSD4SymbolTable());
257 // Perform the compression. Note that if the file is uncompressed bytecode
258 // then we turn the file into compressed bytecode rather than treating it as
259 // compressed data. This is necessary since it allows us to determine that the
260 // file contains bytecode instead of looking like a regular compressed data
261 // member. A compressed bytecode file has its content compressed but has a
262 // magic number of "llvc". This acounts for the +/-4 arithmetic in the code
267 if (member.isBytecode()) {
271 fSize = Compressor::compressToNewBuffer(data,fSize,output,error);
275 if (member.isBytecode())
283 // Compute the fields of the header
284 ArchiveMemberHeader Hdr;
285 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
287 // Write header to archive file
288 ARFile.write((char*)&Hdr, sizeof(Hdr));
290 // Write the long filename if its long
292 ARFile.write(member.getPath().toString().data(),
293 member.getPath().toString().length());
296 // Make sure we write the compressed bytecode magic number if we should.
297 if (willCompress && member.isBytecode())
298 ARFile.write("llvc",4);
300 // Write the (possibly compressed) member's content to the file.
301 ARFile.write(data,fSize);
303 // Make sure the member is an even length
304 if ((ARFile.tellp() & 1) == 1)
305 ARFile << ARFILE_PAD;
307 // Free the compressed data, if necessary
312 // Close the mapped file if it was opened
320 // Write out the LLVM symbol table as an archive member to the file.
322 Archive::writeSymbolTable(std::ofstream& ARFile) {
324 // Construct the symbol table's header
325 ArchiveMemberHeader Hdr;
327 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
328 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
330 sprintf(buffer, "%-8o", 0644);
331 memcpy(Hdr.mode,buffer,8);
332 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
333 memcpy(Hdr.uid,buffer,6);
334 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
335 memcpy(Hdr.gid,buffer,6);
336 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
337 memcpy(Hdr.date,buffer,12);
338 sprintf(buffer,"%-10u",symTabSize);
339 memcpy(Hdr.size,buffer,10);
342 ARFile.write((char*)&Hdr, sizeof(Hdr));
344 // Save the starting position of the symbol tables data content.
345 unsigned startpos = ARFile.tellp();
347 // Write out the symbols sequentially
348 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
351 // Write out the file index
352 writeInteger(I->second, ARFile);
353 // Write out the length of the symbol
354 writeInteger(I->first.length(), ARFile);
355 // Write out the symbol
356 ARFile.write(I->first.data(), I->first.length());
359 // Now that we're done with the symbol table, get the ending file position
360 unsigned endpos = ARFile.tellp();
362 // Make sure that the amount we wrote is what we pre-computed. This is
363 // critical for file integrity purposes.
364 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
366 // Make sure the symbol table is even sized
367 if (symTabSize % 2 != 0 )
368 ARFile << ARFILE_PAD;
371 // Write the entire archive to the file specified when the archive was created.
372 // This writes to a temporary file first. Options are for creating a symbol
373 // table, flattening the file names (no directories, 15 chars max) and
374 // compressing each archive member.
376 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
379 // Make sure they haven't opened up the file, not loaded it,
380 // but are now trying to write it which would wipe out the file.
381 assert(!(members.empty() && mapfile->size() > 8) &&
382 "Can't write an archive not opened for writing");
384 // Create a temporary file to store the archive in
385 sys::Path TmpArchive = archPath;
386 TmpArchive.createTemporaryFileOnDisk();
388 // Make sure the temporary gets removed if we crash
389 sys::RemoveFileOnSignal(TmpArchive);
391 // Create archive file for output.
392 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
394 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
396 // Check for errors opening or creating archive file.
397 if ( !ArchiveFile.is_open() || ArchiveFile.bad() ) {
398 if (TmpArchive.exists())
399 TmpArchive.eraseFromDisk();
401 *error = "Error opening archive file: " + archPath.toString();
405 // If we're creating a symbol table, reset it now
406 if (CreateSymbolTable) {
411 // Write magic string to archive.
412 ArchiveFile << ARFILE_MAGIC;
414 // Loop over all member files, and write them out. Note that this also
415 // builds the symbol table, symTab.
416 for ( MembersList::iterator I = begin(), E = end(); I != E; ++I) {
417 if (!writeMember(*I,ArchiveFile,CreateSymbolTable,
418 TruncateNames,Compress,error))
420 if (TmpArchive.exists())
421 TmpArchive.eraseFromDisk();
427 // Close archive file.
430 // Write the symbol table
431 if (CreateSymbolTable) {
432 // At this point we have written a file that is a legal archive but it
433 // doesn't have a symbol table in it. To aid in faster reading and to
434 // ensure compatibility with other archivers we need to put the symbol
435 // table first in the file. Unfortunately, this means mapping the file
436 // we just wrote back in and copying it to the destination file.
438 // Map in the archive we just wrote.
439 sys::MappedFile arch(TmpArchive);
440 const char* base = (const char*) arch.map();
442 // Open another temporary file in order to avoid invalidating the
444 sys::Path FinalFilePath = archPath;
445 FinalFilePath.createTemporaryFileOnDisk();
446 sys::RemoveFileOnSignal(FinalFilePath);
448 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
449 if ( !FinalFile.is_open() || FinalFile.bad() ) {
450 if (TmpArchive.exists())
451 TmpArchive.eraseFromDisk();
453 *error = "Error opening archive file: " + FinalFilePath.toString();
457 // Write the file magic number
458 FinalFile << ARFILE_MAGIC;
460 // If there is a foreign symbol table, put it into the file now. Most
461 // ar(1) implementations require the symbol table to be first but llvm-ar
462 // can deal with it being after a foreign symbol table. This ensures
463 // compatibility with other ar(1) implementations as well as allowing the
464 // archive to store both native .o and LLVM .bc files, both indexed.
466 writeMember(*foreignST, FinalFile, false, false, false);
469 // Put out the LLVM symbol table now.
470 writeSymbolTable(FinalFile);
472 // Copy the temporary file contents being sure to skip the file's magic
474 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
475 arch.size()-sizeof(ARFILE_MAGIC)+1);
481 // Move the final file over top of TmpArchive
482 FinalFilePath.renamePathOnDisk(TmpArchive);
485 // Before we replace the actual archive, we need to forget all the
486 // members, since they point to data in that old archive. We need to do
487 // this because we cannot replace an open file on Windows.
490 TmpArchive.renamePathOnDisk(archPath);