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);
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 assert(filePath.exists() && "Can't add a non-existent file");
159 ArchiveMember* mbr = new ArchiveMember(this);
162 mbr->path = filePath;
163 if (mbr->path.getFileStatus(mbr->info, ErrMsg))
167 bool hasSlash = filePath.toString().find('/') != std::string::npos;
169 flags |= ArchiveMember::HasPathFlag;
170 if (hasSlash || filePath.toString().length() > 15)
171 flags |= ArchiveMember::HasLongFilenameFlag;
173 mbr->path.getMagicNumber(magic,4);
174 switch (sys::IdentifyFileType(magic.c_str(),4)) {
175 case sys::BytecodeFileType:
176 flags |= ArchiveMember::BytecodeFlag;
178 case sys::CompressedBytecodeFileType:
179 flags |= ArchiveMember::CompressedBytecodeFlag;
185 members.insert(where,mbr);
189 // Write one member out to the file.
191 Archive::writeMember(
192 const ArchiveMember& member,
193 std::ofstream& ARFile,
194 bool CreateSymbolTable,
200 unsigned filepos = ARFile.tellp();
203 // Get the data and its size either from the
204 // member's in-memory data or directly from the file.
205 size_t fSize = member.getSize();
206 const char* data = (const char*)member.getData();
207 sys::MappedFile* mFile = 0;
209 mFile = new sys::MappedFile();
210 if (mFile->open(member.getPath(), sys::MappedFile::READ_ACCESS, ErrMsg))
212 if (!(data = (const char*) mFile->map(ErrMsg)))
214 fSize = mFile->size();
217 // Now that we have the data in memory, update the
218 // symbol table if its a bytecode file.
219 if (CreateSymbolTable &&
220 (member.isBytecode() || member.isCompressedBytecode())) {
221 std::vector<std::string> symbols;
222 std::string FullMemberName = archPath.toString() + "(" +
223 member.getPath().toString()
225 ModuleProvider* MP = GetBytecodeSymbols(
226 (const unsigned char*)data,fSize,FullMemberName, symbols);
228 // If the bytecode parsed successfully
230 for (std::vector<std::string>::iterator SI = symbols.begin(),
231 SE = symbols.end(); SI != SE; ++SI) {
233 std::pair<SymTabType::iterator,bool> Res =
234 symTab.insert(std::make_pair(*SI,filepos));
237 symTabSize += SI->length() +
238 numVbrBytes(SI->length()) +
239 numVbrBytes(filepos);
242 // We don't need this module any more.
250 *ErrMsg = "Can't parse bytecode member: " + member.getPath().toString();
255 // Determine if we actually should compress this member
258 !member.isCompressed() &&
259 !member.isCompressedBytecode() &&
260 !member.isLLVMSymbolTable() &&
261 !member.isSVR4SymbolTable() &&
262 !member.isBSD4SymbolTable());
264 // Perform the compression. Note that if the file is uncompressed bytecode
265 // then we turn the file into compressed bytecode rather than treating it as
266 // compressed data. This is necessary since it allows us to determine that the
267 // file contains bytecode instead of looking like a regular compressed data
268 // member. A compressed bytecode file has its content compressed but has a
269 // magic number of "llvc". This acounts for the +/-4 arithmetic in the code
274 if (member.isBytecode()) {
278 fSize = Compressor::compressToNewBuffer(data,fSize,output,ErrMsg);
282 if (member.isBytecode())
290 // Compute the fields of the header
291 ArchiveMemberHeader Hdr;
292 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
294 // Write header to archive file
295 ARFile.write((char*)&Hdr, sizeof(Hdr));
297 // Write the long filename if its long
299 ARFile.write(member.getPath().toString().data(),
300 member.getPath().toString().length());
303 // Make sure we write the compressed bytecode magic number if we should.
304 if (willCompress && member.isBytecode())
305 ARFile.write("llvc",4);
307 // Write the (possibly compressed) member's content to the file.
308 ARFile.write(data,fSize);
310 // Make sure the member is an even length
311 if ((ARFile.tellp() & 1) == 1)
312 ARFile << ARFILE_PAD;
314 // Free the compressed data, if necessary
319 // Close the mapped file if it was opened
327 // Write out the LLVM symbol table as an archive member to the file.
329 Archive::writeSymbolTable(std::ofstream& ARFile) {
331 // Construct the symbol table's header
332 ArchiveMemberHeader Hdr;
334 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
335 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
337 sprintf(buffer, "%-8o", 0644);
338 memcpy(Hdr.mode,buffer,8);
339 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
340 memcpy(Hdr.uid,buffer,6);
341 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
342 memcpy(Hdr.gid,buffer,6);
343 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
344 memcpy(Hdr.date,buffer,12);
345 sprintf(buffer,"%-10u",symTabSize);
346 memcpy(Hdr.size,buffer,10);
349 ARFile.write((char*)&Hdr, sizeof(Hdr));
351 // Save the starting position of the symbol tables data content.
352 unsigned startpos = ARFile.tellp();
354 // Write out the symbols sequentially
355 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
358 // Write out the file index
359 writeInteger(I->second, ARFile);
360 // Write out the length of the symbol
361 writeInteger(I->first.length(), ARFile);
362 // Write out the symbol
363 ARFile.write(I->first.data(), I->first.length());
366 // Now that we're done with the symbol table, get the ending file position
367 unsigned endpos = ARFile.tellp();
369 // Make sure that the amount we wrote is what we pre-computed. This is
370 // critical for file integrity purposes.
371 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
373 // Make sure the symbol table is even sized
374 if (symTabSize % 2 != 0 )
375 ARFile << ARFILE_PAD;
378 // Write the entire archive to the file specified when the archive was created.
379 // This writes to a temporary file first. Options are for creating a symbol
380 // table, flattening the file names (no directories, 15 chars max) and
381 // compressing each archive member.
383 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
386 // Make sure they haven't opened up the file, not loaded it,
387 // but are now trying to write it which would wipe out the file.
388 assert(!(members.empty() && mapfile->size() > 8) &&
389 "Can't write an archive not opened for writing");
391 // Create a temporary file to store the archive in
392 sys::Path TmpArchive = archPath;
393 if (TmpArchive.createTemporaryFileOnDisk(ErrMsg))
396 // Make sure the temporary gets removed if we crash
397 sys::RemoveFileOnSignal(TmpArchive);
399 // Create archive file for output.
400 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
402 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
404 // Check for errors opening or creating archive file.
405 if (!ArchiveFile.is_open() || ArchiveFile.bad()) {
406 if (TmpArchive.exists())
407 TmpArchive.eraseFromDisk();
409 *ErrMsg = "Error opening archive file: " + archPath.toString();
413 // If we're creating a symbol table, reset it now
414 if (CreateSymbolTable) {
419 // Write magic string to archive.
420 ArchiveFile << ARFILE_MAGIC;
422 // Loop over all member files, and write them out. Note that this also
423 // builds the symbol table, symTab.
424 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
425 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
426 TruncateNames, Compress, ErrMsg)) {
427 if (TmpArchive.exists())
428 TmpArchive.eraseFromDisk();
434 // Close archive file.
437 // Write the symbol table
438 if (CreateSymbolTable) {
439 // At this point we have written a file that is a legal archive but it
440 // doesn't have a symbol table in it. To aid in faster reading and to
441 // ensure compatibility with other archivers we need to put the symbol
442 // table first in the file. Unfortunately, this means mapping the file
443 // we just wrote back in and copying it to the destination file.
445 // Map in the archive we just wrote.
446 sys::MappedFile arch;
447 if (arch.open(TmpArchive, sys::MappedFile::READ_ACCESS, ErrMsg))
450 if (!(base = (const char*) arch.map(ErrMsg)))
453 // Open another temporary file in order to avoid invalidating the
455 sys::Path FinalFilePath = archPath;
456 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
458 sys::RemoveFileOnSignal(FinalFilePath);
460 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
461 if (!FinalFile.is_open() || FinalFile.bad()) {
462 if (TmpArchive.exists())
463 TmpArchive.eraseFromDisk();
465 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
469 // Write the file magic number
470 FinalFile << ARFILE_MAGIC;
472 // If there is a foreign symbol table, put it into the file now. Most
473 // ar(1) implementations require the symbol table to be first but llvm-ar
474 // can deal with it being after a foreign symbol table. This ensures
475 // compatibility with other ar(1) implementations as well as allowing the
476 // archive to store both native .o and LLVM .bc files, both indexed.
478 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
480 if (TmpArchive.exists())
481 TmpArchive.eraseFromDisk();
486 // Put out the LLVM symbol table now.
487 writeSymbolTable(FinalFile);
489 // Copy the temporary file contents being sure to skip the file's magic
491 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
492 arch.size()-sizeof(ARFILE_MAGIC)+1);
498 // Move the final file over top of TmpArchive
499 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
503 // Before we replace the actual archive, we need to forget all the
504 // members, since they point to data in that old archive. We need to do
505 // this because we cannot replace an open file on Windows.
508 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))