1 //! Management of the encoding of LLVM bytecode into rlibs
3 //! This module contains the management of encoding LLVM bytecode into rlibs,
4 //! primarily for the usage in LTO situations. Currently the compiler will
5 //! unconditionally encode LLVM-IR into rlibs regardless of what's happening
6 //! elsewhere, so we currently compress the bytecode via deflate to avoid taking
7 //! up too much space on disk.
9 //! After compressing the bytecode we then have the rest of the format to
10 //! basically deal with various bugs in various archive implementations. The
11 //! format currently is:
13 //! RLIB LLVM-BYTECODE OBJECT LAYOUT
16 //! 0..10 "RUST_OBJECT" encoded in ASCII
17 //! 11..14 format version as little-endian u32
18 //! 15..19 the length of the module identifier string
19 //! 20..n the module identifier string
20 //! n..n+8 size in bytes of deflate compressed LLVM bitcode as
22 //! n+9.. compressed LLVM bitcode
23 //! ? maybe a byte to make this whole thing even length
25 use std::io::{Read, Write};
29 use flate2::read::DeflateDecoder;
30 use flate2::write::DeflateEncoder;
31 use flate2::Compression;
33 // This is the "magic number" expected at the beginning of a LLVM bytecode
35 pub const RLIB_BYTECODE_OBJECT_MAGIC: &[u8] = b"RUST_OBJECT";
37 // The version number this compiler will write to bytecode objects in rlibs
38 pub const RLIB_BYTECODE_OBJECT_VERSION: u8 = 2;
40 pub fn encode(identifier: &str, bytecode: &[u8]) -> Vec<u8> {
41 let mut encoded = Vec::new();
43 // Start off with the magic string
44 encoded.extend_from_slice(RLIB_BYTECODE_OBJECT_MAGIC);
46 // Next up is the version
47 encoded.extend_from_slice(&[RLIB_BYTECODE_OBJECT_VERSION, 0, 0, 0]);
49 // Next is the LLVM module identifier length + contents
50 let identifier_len = identifier.len();
51 encoded.extend_from_slice(&[
52 (identifier_len >> 0) as u8,
53 (identifier_len >> 8) as u8,
54 (identifier_len >> 16) as u8,
55 (identifier_len >> 24) as u8,
57 encoded.extend_from_slice(identifier.as_bytes());
59 // Next is the LLVM module deflate compressed, prefixed with its length. We
60 // don't know its length yet, so fill in 0s
61 let deflated_size_pos = encoded.len();
62 encoded.extend_from_slice(&[0, 0, 0, 0, 0, 0, 0, 0]);
64 let before = encoded.len();
65 DeflateEncoder::new(&mut encoded, Compression::fast()).write_all(bytecode).unwrap();
66 let after = encoded.len();
68 // Fill in the length we reserved space for before
69 let bytecode_len = (after - before) as u64;
70 encoded[deflated_size_pos + 0] = (bytecode_len >> 0) as u8;
71 encoded[deflated_size_pos + 1] = (bytecode_len >> 8) as u8;
72 encoded[deflated_size_pos + 2] = (bytecode_len >> 16) as u8;
73 encoded[deflated_size_pos + 3] = (bytecode_len >> 24) as u8;
74 encoded[deflated_size_pos + 4] = (bytecode_len >> 32) as u8;
75 encoded[deflated_size_pos + 5] = (bytecode_len >> 40) as u8;
76 encoded[deflated_size_pos + 6] = (bytecode_len >> 48) as u8;
77 encoded[deflated_size_pos + 7] = (bytecode_len >> 56) as u8;
79 // If the number of bytes written to the object so far is odd, add a
80 // padding byte to make it even. This works around a crash bug in LLDB
82 if encoded.len() % 2 == 1 {
89 pub struct DecodedBytecode<'a> {
91 encoded_bytecode: &'a [u8],
94 impl<'a> DecodedBytecode<'a> {
95 pub fn new(data: &'a [u8]) -> Result<DecodedBytecode<'a>, &'static str> {
96 if !data.starts_with(RLIB_BYTECODE_OBJECT_MAGIC) {
97 return Err("magic bytecode prefix not found");
99 let data = &data[RLIB_BYTECODE_OBJECT_MAGIC.len()..];
100 if !data.starts_with(&[RLIB_BYTECODE_OBJECT_VERSION, 0, 0, 0]) {
101 return Err("wrong version prefix found in bytecode");
103 let data = &data[4..];
105 return Err("bytecode corrupted");
108 unsafe { u32::from_le(ptr::read_unaligned(data.as_ptr() as *const u32)) as usize };
109 let data = &data[4..];
110 if data.len() < identifier_len {
111 return Err("bytecode corrupted");
113 let identifier = match str::from_utf8(&data[..identifier_len]) {
115 Err(_) => return Err("bytecode corrupted"),
117 let data = &data[identifier_len..];
119 return Err("bytecode corrupted");
122 unsafe { u64::from_le(ptr::read_unaligned(data.as_ptr() as *const u64)) as usize };
123 let data = &data[8..];
124 if data.len() < bytecode_len {
125 return Err("bytecode corrupted");
127 let encoded_bytecode = &data[..bytecode_len];
129 Ok(DecodedBytecode { identifier, encoded_bytecode })
132 pub fn bytecode(&self) -> Vec<u8> {
133 let mut data = Vec::new();
134 DeflateDecoder::new(self.encoded_bytecode).read_to_end(&mut data).unwrap();
138 pub fn identifier(&self) -> &'a str {