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/FileUtilities.h"
17 #include "llvm/Support/Compressor.h"
18 #include "llvm/System/Signals.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().get();
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.isForeignSymbolTable()) {
109 memcpy(hdr.name,ARFILE_SYMTAB_NAME,16);
110 } else if (mbr.isLLVMSymbolTable()) {
111 memcpy(hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
112 } else if (TruncateNames) {
113 const char* nm = mbrPath.c_str();
114 unsigned len = mbrPath.length();
115 size_t slashpos = mbrPath.rfind('/');
116 if (slashpos != std::string::npos) {
122 memcpy(hdr.name,nm,len);
124 } else if (mbrPath.length() < 16 && mbrPath.find('/') == std::string::npos) {
125 memcpy(hdr.name,mbrPath.c_str(),mbrPath.length());
126 hdr.name[mbrPath.length()] = '/';
128 std::string nm = "#1/";
129 nm += utostr(mbrPath.length());
130 memcpy(hdr.name,nm.data(),nm.length());
132 sz -= mbrPath.length();
134 sz += mbrPath.length();
135 writeLongName = true;
138 // Set the size field
141 sprintf(&buffer[1],"%-9u",(unsigned)-sz);
143 sprintf(buffer, "%-10u", (unsigned)sz);
145 memcpy(hdr.size,buffer,10);
147 return writeLongName;
150 // Insert a file into the archive before some other member. This also takes care
151 // of extracting the necessary flags and information from the file.
153 Archive::addFileBefore(const sys::Path& filePath, iterator where) {
154 assert(filePath.exists() && "Can't add a non-existent file");
156 ArchiveMember* mbr = new ArchiveMember(this);
159 mbr->path = filePath;
160 mbr->path.getStatusInfo(mbr->info);
163 bool hasSlash = filePath.get().find('/') != std::string::npos;
165 flags |= ArchiveMember::HasPathFlag;
166 if (hasSlash || filePath.get().length() > 15)
167 flags |= ArchiveMember::HasLongFilenameFlag;
169 mbr->path.getMagicNumber(magic,4);
170 switch (sys::IdentifyFileType(magic.c_str(),4)) {
171 case sys::BytecodeFileType:
172 flags |= ArchiveMember::BytecodeFlag;
174 case sys::CompressedBytecodeFileType:
175 flags |= ArchiveMember::CompressedBytecodeFlag;
181 members.insert(where,mbr);
184 // Write one member out to the file.
186 Archive::writeMember(
187 const ArchiveMember& member,
188 std::ofstream& ARFile,
189 bool CreateSymbolTable,
194 unsigned filepos = ARFile.tellp();
197 // Get the data and its size either from the
198 // member's in-memory data or directly from the file.
199 size_t fSize = member.getSize();
200 const char* data = (const char*)member.getData();
201 sys::MappedFile* mFile = 0;
203 mFile = new sys::MappedFile(member.getPath());
204 data = (const char*) mFile->map();
205 fSize = mFile->size();
208 // Now that we have the data in memory, update the
209 // symbol table if its a bytecode file.
210 if (CreateSymbolTable &&
211 (member.isBytecode() || member.isCompressedBytecode())) {
212 std::vector<std::string> symbols;
213 std::string FullMemberName = archPath.get() + "(" + member.getPath().get()
215 ModuleProvider* MP = GetBytecodeSymbols(
216 (const unsigned char*)data,fSize,FullMemberName, symbols);
218 // If the bytecode parsed successfully
220 for (std::vector<std::string>::iterator SI = symbols.begin(),
221 SE = symbols.end(); SI != SE; ++SI) {
223 std::pair<SymTabType::iterator,bool> Res =
224 symTab.insert(std::make_pair(*SI,filepos));
227 symTabSize += SI->length() +
228 numVbrBytes(SI->length()) +
229 numVbrBytes(filepos);
232 // We don't need this module any more.
235 throw std::string("Can't parse bytecode member: ") +
236 member.getPath().get();
240 // Determine if we actually should compress this member
243 !member.isForeignSymbolTable() &&
244 !member.isLLVMSymbolTable() &&
245 !member.isCompressed() &&
246 !member.isCompressedBytecode());
248 // Perform the compression. Note that if the file is uncompressed bytecode
249 // then we turn the file into compressed bytecode rather than treating it as
250 // compressed data. This is necessary since it allows us to determine that the
251 // file contains bytecode instead of looking like a regular compressed data
252 // member. A compressed bytecode file has its content compressed but has a
253 // magic number of "llvc". This acounts for the +/-4 arithmetic in the code
258 if (member.isBytecode()) {
262 fSize = Compressor::compressToNewBuffer(
263 data,fSize,output,Compressor::COMP_TYPE_ZLIB);
265 if (member.isBytecode())
273 // Compute the fields of the header
274 ArchiveMemberHeader Hdr;
275 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
277 // Write header to archive file
278 ARFile.write((char*)&Hdr, sizeof(Hdr));
280 // Write the long filename if its long
282 ARFile.write(member.getPath().get().data(),member.getPath().get().length());
285 // Make sure we write the compressed bytecode magic number if we should.
286 if (willCompress && member.isBytecode())
287 ARFile.write("llvc",4);
289 // Write the (possibly compressed) member's content to the file.
290 ARFile.write(data,fSize);
292 // Make sure the member is an even length
293 if (ARFile.tellp() & 1 == 1)
294 ARFile << ARFILE_PAD;
296 // Free the compressed data, if necessary
301 // Close the mapped file if it was opened
308 // Write out the LLVM symbol table as an archive member to the file.
310 Archive::writeSymbolTable(std::ofstream& ARFile) {
312 // Construct the symbol table's header
313 ArchiveMemberHeader Hdr;
315 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
316 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
318 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
319 memcpy(Hdr.date,buffer,12);
320 sprintf(buffer,"%-10u",symTabSize);
321 memcpy(Hdr.size,buffer,10);
324 ARFile.write((char*)&Hdr, sizeof(Hdr));
326 // Save the starting position of the symbol tables data content.
327 unsigned startpos = ARFile.tellp();
329 // Write out the symbols sequentially
330 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
333 // Write out the file index
334 writeInteger(I->second, ARFile);
335 // Write out the length of the symbol
336 writeInteger(I->first.length(), ARFile);
337 // Write out the symbol
338 ARFile.write(I->first.data(), I->first.length());
341 // Now that we're done with the symbol table, get the ending file position
342 unsigned endpos = ARFile.tellp();
344 // Make sure that the amount we wrote is what we pre-computed. This is
345 // critical for file integrity purposes.
346 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
348 // Make sure the symbol table is even sized
349 if (symTabSize % 2 != 0 )
350 ARFile << ARFILE_PAD;
353 // Write the entire archive to the file specified when the archive was created.
354 // This writes to a temporary file first. Options are for creating a symbol
355 // table, flattening the file names (no directories, 15 chars max) and
356 // compressing each archive member.
358 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress){
360 // Make sure they haven't opened up the file, not loaded it,
361 // but are now trying to write it which would wipe out the file.
362 assert(!(members.empty() && mapfile->size() > 8) &&
363 "Can't write an archive not opened for writing");
365 // Create a temporary file to store the archive in
366 sys::Path TmpArchive = archPath;
367 TmpArchive.createTemporaryFile();
369 // Make sure the temporary gets removed if we crash
370 sys::RemoveFileOnSignal(TmpArchive);
372 // Ensure we can remove the temporary even in the face of an exception
374 // Create archive file for output.
375 std::ofstream ArchiveFile(TmpArchive.c_str());
377 // Check for errors opening or creating archive file.
378 if ( !ArchiveFile.is_open() || ArchiveFile.bad() ) {
379 throw std::string("Error opening archive file: ") + archPath.get();
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 writeMember(*I,ArchiveFile,CreateSymbolTable,TruncateNames,Compress);
397 // Close archive file.
400 // Write the symbol table
401 if (CreateSymbolTable) {
402 // At this point we have written a file that is a legal archive but it
403 // doesn't have a symbol table in it. To aid in faster reading and to
404 // ensure compatibility with other archivers we need to put the symbol
405 // table first in the file. Unfortunately, this means mapping the file
406 // we just wrote back in and copying it to the destination file.
408 // Map in the archive we just wrote.
409 sys::MappedFile arch(TmpArchive);
410 const char* base = (const char*) arch.map();
412 // Open the final file to write and check it.
413 std::ofstream FinalFile(archPath.c_str());
414 if ( !FinalFile.is_open() || FinalFile.bad() ) {
415 throw std::string("Error opening archive file: ") + archPath.get();
418 // Write the file magic number
419 FinalFile << ARFILE_MAGIC;
421 // If there is a foreign symbol table, put it into the file now.
423 writeMember(*foreignST, FinalFile, false, false, false);
426 // Put out the LLVM symbol table now.
427 writeSymbolTable(FinalFile);
429 // Copy the temporary file contents being sure to skip the file's magic
431 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
432 arch.size()-sizeof(ARFILE_MAGIC)+1);
437 TmpArchive.destroyFile();
440 // We don't have to insert the symbol table, so just renaming the temp
441 // file to the correct name will suffice.
442 TmpArchive.renameFile(archPath);
445 // Make sure we clean up.
446 if (TmpArchive.exists())
447 TmpArchive.destroyFile();