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 if (mbr->path.getFileStatus(mbr->info, false, ErrMsg))
170 bool hasSlash = filePath.toString().find('/') != std::string::npos;
172 flags |= ArchiveMember::HasPathFlag;
173 if (hasSlash || filePath.toString().length() > 15)
174 flags |= ArchiveMember::HasLongFilenameFlag;
176 mbr->path.getMagicNumber(magic,4);
177 switch (sys::IdentifyFileType(magic.c_str(),4)) {
178 case sys::BytecodeFileType:
179 flags |= ArchiveMember::BytecodeFlag;
181 case sys::CompressedBytecodeFileType:
182 flags |= ArchiveMember::CompressedBytecodeFlag;
188 members.insert(where,mbr);
192 // Write one member out to the file.
194 Archive::writeMember(
195 const ArchiveMember& member,
196 std::ofstream& ARFile,
197 bool CreateSymbolTable,
203 unsigned filepos = ARFile.tellp();
206 // Get the data and its size either from the
207 // member's in-memory data or directly from the file.
208 size_t fSize = member.getSize();
209 const char* data = (const char*)member.getData();
210 sys::MappedFile* mFile = 0;
212 mFile = new sys::MappedFile();
213 if (mFile->open(member.getPath(), sys::MappedFile::READ_ACCESS, ErrMsg))
215 if (!(data = (const char*) mFile->map(ErrMsg)))
217 fSize = mFile->size();
220 // Now that we have the data in memory, update the
221 // symbol table if its a bytecode file.
222 if (CreateSymbolTable &&
223 (member.isBytecode() || member.isCompressedBytecode())) {
224 std::vector<std::string> symbols;
225 std::string FullMemberName = archPath.toString() + "(" +
226 member.getPath().toString()
229 GetBytecodeSymbols((const unsigned char*)data,fSize,
230 FullMemberName, symbols,
231 Compressor::decompressToNewBuffer, ErrMsg);
233 // If the bytecode parsed successfully
235 for (std::vector<std::string>::iterator SI = symbols.begin(),
236 SE = symbols.end(); SI != SE; ++SI) {
238 std::pair<SymTabType::iterator,bool> Res =
239 symTab.insert(std::make_pair(*SI,filepos));
242 symTabSize += SI->length() +
243 numVbrBytes(SI->length()) +
244 numVbrBytes(filepos);
247 // We don't need this module any more.
255 *ErrMsg = "Can't parse bytecode member: " + member.getPath().toString()
261 // Determine if we actually should compress this member
264 !member.isCompressed() &&
265 !member.isCompressedBytecode() &&
266 !member.isLLVMSymbolTable() &&
267 !member.isSVR4SymbolTable() &&
268 !member.isBSD4SymbolTable());
270 // Perform the compression. Note that if the file is uncompressed bytecode
271 // then we turn the file into compressed bytecode rather than treating it as
272 // compressed data. This is necessary since it allows us to determine that the
273 // file contains bytecode instead of looking like a regular compressed data
274 // member. A compressed bytecode file has its content compressed but has a
275 // magic number of "llvc". This acounts for the +/-4 arithmetic in the code
280 if (member.isBytecode()) {
284 fSize = Compressor::compressToNewBuffer(data,fSize,output,ErrMsg);
288 if (member.isBytecode())
296 // Compute the fields of the header
297 ArchiveMemberHeader Hdr;
298 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
300 // Write header to archive file
301 ARFile.write((char*)&Hdr, sizeof(Hdr));
303 // Write the long filename if its long
305 ARFile.write(member.getPath().toString().data(),
306 member.getPath().toString().length());
309 // Make sure we write the compressed bytecode magic number if we should.
310 if (willCompress && member.isBytecode())
311 ARFile.write("llvc",4);
313 // Write the (possibly compressed) member's content to the file.
314 ARFile.write(data,fSize);
316 // Make sure the member is an even length
317 if ((ARFile.tellp() & 1) == 1)
318 ARFile << ARFILE_PAD;
320 // Free the compressed data, if necessary
325 // Close the mapped file if it was opened
333 // Write out the LLVM symbol table as an archive member to the file.
335 Archive::writeSymbolTable(std::ofstream& ARFile) {
337 // Construct the symbol table's header
338 ArchiveMemberHeader Hdr;
340 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
341 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
343 sprintf(buffer, "%-8o", 0644);
344 memcpy(Hdr.mode,buffer,8);
345 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
346 memcpy(Hdr.uid,buffer,6);
347 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
348 memcpy(Hdr.gid,buffer,6);
349 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
350 memcpy(Hdr.date,buffer,12);
351 sprintf(buffer,"%-10u",symTabSize);
352 memcpy(Hdr.size,buffer,10);
355 ARFile.write((char*)&Hdr, sizeof(Hdr));
357 // Save the starting position of the symbol tables data content.
358 unsigned startpos = ARFile.tellp();
360 // Write out the symbols sequentially
361 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
364 // Write out the file index
365 writeInteger(I->second, ARFile);
366 // Write out the length of the symbol
367 writeInteger(I->first.length(), ARFile);
368 // Write out the symbol
369 ARFile.write(I->first.data(), I->first.length());
372 // Now that we're done with the symbol table, get the ending file position
373 unsigned endpos = ARFile.tellp();
375 // Make sure that the amount we wrote is what we pre-computed. This is
376 // critical for file integrity purposes.
377 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
379 // Make sure the symbol table is even sized
380 if (symTabSize % 2 != 0 )
381 ARFile << ARFILE_PAD;
384 // Write the entire archive to the file specified when the archive was created.
385 // This writes to a temporary file first. Options are for creating a symbol
386 // table, flattening the file names (no directories, 15 chars max) and
387 // compressing each archive member.
389 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
392 // Make sure they haven't opened up the file, not loaded it,
393 // but are now trying to write it which would wipe out the file.
394 if (members.empty() && mapfile->size() > 8) {
396 *ErrMsg = "Can't write an archive not opened for writing";
400 // Create a temporary file to store the archive in
401 sys::Path TmpArchive = archPath;
402 if (TmpArchive.createTemporaryFileOnDisk(ErrMsg))
405 // Make sure the temporary gets removed if we crash
406 sys::RemoveFileOnSignal(TmpArchive);
408 // Create archive file for output.
409 std::ios::openmode io_mode = std::ios::out | std::ios::trunc |
411 std::ofstream ArchiveFile(TmpArchive.c_str(), io_mode);
413 // Check for errors opening or creating archive file.
414 if (!ArchiveFile.is_open() || ArchiveFile.bad()) {
415 if (TmpArchive.exists())
416 TmpArchive.eraseFromDisk();
418 *ErrMsg = "Error opening archive file: " + archPath.toString();
422 // If we're creating a symbol table, reset it now
423 if (CreateSymbolTable) {
428 // Write magic string to archive.
429 ArchiveFile << ARFILE_MAGIC;
431 // Loop over all member files, and write them out. Note that this also
432 // builds the symbol table, symTab.
433 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
434 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
435 TruncateNames, Compress, ErrMsg)) {
436 if (TmpArchive.exists())
437 TmpArchive.eraseFromDisk();
443 // Close archive file.
446 // Write the symbol table
447 if (CreateSymbolTable) {
448 // At this point we have written a file that is a legal archive but it
449 // doesn't have a symbol table in it. To aid in faster reading and to
450 // ensure compatibility with other archivers we need to put the symbol
451 // table first in the file. Unfortunately, this means mapping the file
452 // we just wrote back in and copying it to the destination file.
454 // Map in the archive we just wrote.
455 sys::MappedFile arch;
456 if (arch.open(TmpArchive, sys::MappedFile::READ_ACCESS, ErrMsg))
459 if (!(base = (const char*) arch.map(ErrMsg)))
462 // Open another temporary file in order to avoid invalidating the
464 sys::Path FinalFilePath = archPath;
465 if (FinalFilePath.createTemporaryFileOnDisk(ErrMsg))
467 sys::RemoveFileOnSignal(FinalFilePath);
469 std::ofstream FinalFile(FinalFilePath.c_str(), io_mode);
470 if (!FinalFile.is_open() || FinalFile.bad()) {
471 if (TmpArchive.exists())
472 TmpArchive.eraseFromDisk();
474 *ErrMsg = "Error opening archive file: " + FinalFilePath.toString();
478 // Write the file magic number
479 FinalFile << ARFILE_MAGIC;
481 // If there is a foreign symbol table, put it into the file now. Most
482 // ar(1) implementations require the symbol table to be first but llvm-ar
483 // can deal with it being after a foreign symbol table. This ensures
484 // compatibility with other ar(1) implementations as well as allowing the
485 // archive to store both native .o and LLVM .bc files, both indexed.
487 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
489 if (TmpArchive.exists())
490 TmpArchive.eraseFromDisk();
495 // Put out the LLVM symbol table now.
496 writeSymbolTable(FinalFile);
498 // Copy the temporary file contents being sure to skip the file's magic
500 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
501 arch.size()-sizeof(ARFILE_MAGIC)+1);
507 // Move the final file over top of TmpArchive
508 if (FinalFilePath.renamePathOnDisk(TmpArchive, ErrMsg))
512 // Before we replace the actual archive, we need to forget all the
513 // members, since they point to data in that old archive. We need to do
514 // this because we cannot replace an open file on Windows.
517 if (TmpArchive.renamePathOnDisk(archPath, ErrMsg))