1 //===-- ArchiveWriter.cpp - Write LLVM archive files ----------------------===//
3 // The LLVM Compiler Infrastructure
5 // This file is distributed under the University of Illinois Open Source
6 // License. See LICENSE.TXT for details.
8 //===----------------------------------------------------------------------===//
10 // Builds up an LLVM archive file (.a) containing LLVM bitcode.
12 //===----------------------------------------------------------------------===//
14 #include "ArchiveInternals.h"
15 #include "llvm/Module.h"
16 #include "llvm/ADT/OwningPtr.h"
17 #include "llvm/Bitcode/ReaderWriter.h"
18 #include "llvm/Support/FileSystem.h"
19 #include "llvm/Support/MemoryBuffer.h"
20 #include "llvm/Support/Process.h"
21 #include "llvm/Support/raw_ostream.h"
22 #include "llvm/Support/Signals.h"
23 #include "llvm/Support/system_error.h"
29 // Write an integer using variable bit rate encoding. This saves a few bytes
30 // per entry in the symbol table.
31 static inline void writeInteger(unsigned num, raw_ostream& ARFile) {
33 if (num < 0x80) { // done?
34 ARFile << (unsigned char)num;
38 // Nope, we are bigger than a character, output the next 7 bits and set the
39 // high bit to say that there is more coming...
40 ARFile << (unsigned char)(0x80 | ((unsigned char)num & 0x7F));
41 num >>= 7; // Shift out 7 bits now...
45 // Compute how many bytes are taken by a given VBR encoded value. This is needed
46 // to pre-compute the size of the symbol table.
47 static inline unsigned numVbrBytes(unsigned num) {
49 // Note that the following nested ifs are somewhat equivalent to a binary
50 // search. We split it in half by comparing against 2^14 first. This allows
51 // most reasonable values to be done in 2 comparisons instead of 1 for
52 // small ones and four for large ones. We expect this to access file offsets
53 // in the 2^10 to 2^24 range and symbol lengths in the 2^0 to 2^8 range,
54 // so this approach is reasonable.
66 return 5; // anything >= 2^28 takes 5 bytes
69 // Create an empty archive.
70 Archive* Archive::CreateEmpty(const sys::Path& FilePath, LLVMContext& C) {
71 Archive* result = new Archive(FilePath, C);
75 // Fill the ArchiveMemberHeader with the information from a member. If
76 // TruncateNames is true, names are flattened to 15 chars or less. The sz field
77 // is provided here instead of coming from the mbr because the member might be
78 // stored compressed and the compressed size is not the ArchiveMember's size.
79 // Furthermore compressed files have negative size fields to identify them as
82 Archive::fillHeader(const ArchiveMember &mbr, ArchiveMemberHeader& hdr,
83 int sz, bool TruncateNames) const {
85 // Set the permissions mode, uid and gid
88 sprintf(buffer, "%-8o", mbr.getMode());
89 memcpy(hdr.mode,buffer,8);
90 sprintf(buffer, "%-6u", mbr.getUser());
91 memcpy(hdr.uid,buffer,6);
92 sprintf(buffer, "%-6u", mbr.getGroup());
93 memcpy(hdr.gid,buffer,6);
95 // Set the last modification date
96 uint64_t secondsSinceEpoch = mbr.getModTime().toEpochTime();
97 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
98 memcpy(hdr.date,buffer,12);
100 // Get rid of trailing blanks in the name
101 std::string mbrPath = mbr.getPath().str();
102 size_t mbrLen = mbrPath.length();
103 while (mbrLen > 0 && mbrPath[mbrLen-1] == ' ') {
104 mbrPath.erase(mbrLen-1,1);
108 // Set the name field in one of its various flavors.
109 bool writeLongName = false;
110 if (mbr.isStringTable()) {
111 memcpy(hdr.name,ARFILE_STRTAB_NAME,16);
112 } else if (mbr.isSVR4SymbolTable()) {
113 memcpy(hdr.name,ARFILE_SVR4_SYMTAB_NAME,16);
114 } else if (mbr.isBSD4SymbolTable()) {
115 memcpy(hdr.name,ARFILE_BSD4_SYMTAB_NAME,16);
116 } else if (mbr.isLLVMSymbolTable()) {
117 memcpy(hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
118 } else if (TruncateNames) {
119 const char* nm = mbrPath.c_str();
120 unsigned len = mbrPath.length();
121 size_t slashpos = mbrPath.rfind('/');
122 if (slashpos != std::string::npos) {
128 memcpy(hdr.name,nm,len);
130 } else if (mbrPath.length() < 16 && mbrPath.find('/') == std::string::npos) {
131 memcpy(hdr.name,mbrPath.c_str(),mbrPath.length());
132 hdr.name[mbrPath.length()] = '/';
134 std::string nm = "#1/";
135 nm += utostr(mbrPath.length());
136 memcpy(hdr.name,nm.data(),nm.length());
138 sz -= mbrPath.length();
140 sz += mbrPath.length();
141 writeLongName = true;
144 // Set the size field
147 sprintf(&buffer[1],"%-9u",(unsigned)-sz);
149 sprintf(buffer, "%-10u", (unsigned)sz);
151 memcpy(hdr.size,buffer,10);
153 return writeLongName;
156 // Insert a file into the archive before some other member. This also takes care
157 // of extracting the necessary flags and information from the file.
159 Archive::addFileBefore(const sys::Path& filePath, iterator where,
160 std::string* ErrMsg) {
162 if (sys::fs::exists(filePath.str(), Exists) || !Exists) {
164 *ErrMsg = "Can not add a non-existent file to archive";
168 ArchiveMember* mbr = new ArchiveMember(this);
171 mbr->path = filePath;
172 const sys::FileStatus *FSInfo = mbr->path.getFileStatus(false, ErrMsg);
180 bool hasSlash = filePath.str().find('/') != std::string::npos;
182 flags |= ArchiveMember::HasPathFlag;
183 if (hasSlash || filePath.str().length() > 15)
184 flags |= ArchiveMember::HasLongFilenameFlag;
186 sys::LLVMFileType type;
187 if (sys::fs::identify_magic(mbr->path.str(), type))
188 type = sys::Unknown_FileType;
190 case sys::Bitcode_FileType:
191 flags |= ArchiveMember::BitcodeFlag;
197 members.insert(where,mbr);
201 // Write one member out to the file.
203 Archive::writeMember(
204 const ArchiveMember& member,
206 bool CreateSymbolTable,
212 unsigned filepos = ARFile.tell();
215 // Get the data and its size either from the
216 // member's in-memory data or directly from the file.
217 size_t fSize = member.getSize();
218 const char *data = (const char*)member.getData();
219 MemoryBuffer *mFile = 0;
221 OwningPtr<MemoryBuffer> File;
222 if (error_code ec = MemoryBuffer::getFile(member.getPath().c_str(), File)) {
224 *ErrMsg = ec.message();
228 data = mFile->getBufferStart();
229 fSize = mFile->getBufferSize();
232 // Now that we have the data in memory, update the
233 // symbol table if it's a bitcode file.
234 if (CreateSymbolTable && member.isBitcode()) {
235 std::vector<std::string> symbols;
236 std::string FullMemberName = archPath.str() + "(" + member.getPath().str()
239 GetBitcodeSymbols(data, fSize, FullMemberName, Context, symbols, ErrMsg);
241 // If the bitcode parsed successfully
243 for (std::vector<std::string>::iterator SI = symbols.begin(),
244 SE = symbols.end(); SI != SE; ++SI) {
246 std::pair<SymTabType::iterator,bool> Res =
247 symTab.insert(std::make_pair(*SI,filepos));
250 symTabSize += SI->length() +
251 numVbrBytes(SI->length()) +
252 numVbrBytes(filepos);
255 // We don't need this module any more.
260 *ErrMsg = "Can't parse bitcode member: " + member.getPath().str()
268 // Compute the fields of the header
269 ArchiveMemberHeader Hdr;
270 bool writeLongName = fillHeader(member,Hdr,hdrSize,TruncateNames);
272 // Write header to archive file
273 ARFile.write((char*)&Hdr, sizeof(Hdr));
275 // Write the long filename if its long
277 ARFile.write(member.getPath().str().data(),
278 member.getPath().str().length());
281 // Write the (possibly compressed) member's content to the file.
282 ARFile.write(data,fSize);
284 // Make sure the member is an even length
285 if ((ARFile.tell() & 1) == 1)
286 ARFile << ARFILE_PAD;
288 // Close the mapped file if it was opened
293 // Write out the LLVM symbol table as an archive member to the file.
295 Archive::writeSymbolTable(raw_ostream& ARFile) {
297 // Construct the symbol table's header
298 ArchiveMemberHeader Hdr;
300 memcpy(Hdr.name,ARFILE_LLVM_SYMTAB_NAME,16);
301 uint64_t secondsSinceEpoch = sys::TimeValue::now().toEpochTime();
303 sprintf(buffer, "%-8o", 0644);
304 memcpy(Hdr.mode,buffer,8);
305 sprintf(buffer, "%-6u", sys::Process::GetCurrentUserId());
306 memcpy(Hdr.uid,buffer,6);
307 sprintf(buffer, "%-6u", sys::Process::GetCurrentGroupId());
308 memcpy(Hdr.gid,buffer,6);
309 sprintf(buffer,"%-12u", unsigned(secondsSinceEpoch));
310 memcpy(Hdr.date,buffer,12);
311 sprintf(buffer,"%-10u",symTabSize);
312 memcpy(Hdr.size,buffer,10);
315 ARFile.write((char*)&Hdr, sizeof(Hdr));
318 // Save the starting position of the symbol tables data content.
319 unsigned startpos = ARFile.tell();
322 // Write out the symbols sequentially
323 for ( Archive::SymTabType::iterator I = symTab.begin(), E = symTab.end();
326 // Write out the file index
327 writeInteger(I->second, ARFile);
328 // Write out the length of the symbol
329 writeInteger(I->first.length(), ARFile);
330 // Write out the symbol
331 ARFile.write(I->first.data(), I->first.length());
335 // Now that we're done with the symbol table, get the ending file position
336 unsigned endpos = ARFile.tell();
339 // Make sure that the amount we wrote is what we pre-computed. This is
340 // critical for file integrity purposes.
341 assert(endpos - startpos == symTabSize && "Invalid symTabSize computation");
343 // Make sure the symbol table is even sized
344 if (symTabSize % 2 != 0 )
345 ARFile << ARFILE_PAD;
348 // Write the entire archive to the file specified when the archive was created.
349 // This writes to a temporary file first. Options are for creating a symbol
350 // table, flattening the file names (no directories, 15 chars max) and
351 // compressing each archive member.
353 Archive::writeToDisk(bool CreateSymbolTable, bool TruncateNames, bool Compress,
356 // Make sure they haven't opened up the file, not loaded it,
357 // but are now trying to write it which would wipe out the file.
358 if (members.empty() && mapfile && mapfile->getBufferSize() > 8) {
360 *ErrMsg = "Can't write an archive not opened for writing";
364 // Create a temporary file to store the archive in
365 SmallString<128> TempArchivePath;
367 if (error_code ec = sys::fs::unique_file("%%-%%-%%-%%" + archPath.str(),
368 ArchFD, TempArchivePath)) {
369 if (ErrMsg) *ErrMsg = ec.message();
373 // Make sure the temporary gets removed if we crash
374 sys::RemoveFileOnSignal(sys::Path(TempArchivePath.str()));
376 // Create archive file for output.
377 raw_fd_ostream ArchiveFile(ArchFD, true);
379 // If we're creating a symbol table, reset it now
380 if (CreateSymbolTable) {
385 // Write magic string to archive.
386 ArchiveFile << ARFILE_MAGIC;
388 // Loop over all member files, and write them out. Note that this also
389 // builds the symbol table, symTab.
390 for (MembersList::iterator I = begin(), E = end(); I != E; ++I) {
391 if (writeMember(*I, ArchiveFile, CreateSymbolTable,
392 TruncateNames, Compress, ErrMsg)) {
395 sys::fs::remove(TempArchivePath.str(), existed);
400 // Close archive file.
403 // Write the symbol table
404 if (CreateSymbolTable) {
405 // At this point we have written a file that is a legal archive but it
406 // doesn't have a symbol table in it. To aid in faster reading and to
407 // ensure compatibility with other archivers we need to put the symbol
408 // table first in the file. Unfortunately, this means mapping the file
409 // we just wrote back in and copying it to the destination file.
410 SmallString<128> TempArchiveWithSymbolTablePath;
412 // Map in the archive we just wrote.
414 OwningPtr<MemoryBuffer> arch;
415 if (error_code ec = MemoryBuffer::getFile(TempArchivePath.c_str(), arch)) {
417 *ErrMsg = ec.message();
420 const char* base = arch->getBufferStart();
422 // Open another temporary file in order to avoid invalidating the
424 if (error_code ec = sys::fs::unique_file("%%-%%-%%-%%" + archPath.str(),
425 ArchFD, TempArchiveWithSymbolTablePath)) {
426 if (ErrMsg) *ErrMsg = ec.message();
429 sys::RemoveFileOnSignal(sys::Path(TempArchiveWithSymbolTablePath.str()));
431 raw_fd_ostream FinalFile(ArchFD, true);
433 // Write the file magic number
434 FinalFile << ARFILE_MAGIC;
436 // If there is a foreign symbol table, put it into the file now. Most
437 // ar(1) implementations require the symbol table to be first but llvm-ar
438 // can deal with it being after a foreign symbol table. This ensures
439 // compatibility with other ar(1) implementations as well as allowing the
440 // archive to store both native .o and LLVM .bc files, both indexed.
442 if (writeMember(*foreignST, FinalFile, false, false, false, ErrMsg)) {
445 sys::fs::remove(TempArchiveWithSymbolTablePath.str(), existed);
450 // Put out the LLVM symbol table now.
451 writeSymbolTable(FinalFile);
453 // Copy the temporary file contents being sure to skip the file's magic
455 FinalFile.write(base + sizeof(ARFILE_MAGIC)-1,
456 arch->getBufferSize()-sizeof(ARFILE_MAGIC)+1);
462 // Move the final file over top of TmpArchive
463 if (error_code ec = sys::fs::rename(TempArchiveWithSymbolTablePath.str(),
464 TempArchivePath.str())) {
465 if (ErrMsg) *ErrMsg = ec.message();
470 // Before we replace the actual archive, we need to forget all the
471 // members, since they point to data in that old archive. We need to do
472 // this because we cannot replace an open file on Windows.
475 if (error_code ec = sys::fs::rename(TempArchivePath.str(),
477 if (ErrMsg) *ErrMsg = ec.message();
481 // Set correct read and write permissions after temporary file is moved
482 // to final destination path.
483 if (archPath.makeReadableOnDisk(ErrMsg))
485 if (archPath.makeWriteableOnDisk(ErrMsg))