1 use crate::leb128::{self, max_leb128_len};
2 use crate::serialize::{self, Decoder as _, Encoder as _};
3 use std::convert::TryInto;
5 use std::io::{self, Write};
6 use std::mem::MaybeUninit;
10 // -----------------------------------------------------------------------------
12 // -----------------------------------------------------------------------------
19 pub fn new() -> Encoder {
20 Encoder { data: vec![] }
24 pub fn position(&self) -> usize {
28 pub fn finish(self) -> Vec<u8> {
33 macro_rules! write_leb128 {
34 ($enc:expr, $value:expr, $int_ty:ty, $fun:ident) => {{
35 const MAX_ENCODED_LEN: usize = max_leb128_len!($int_ty);
36 let old_len = $enc.data.len();
38 if MAX_ENCODED_LEN > $enc.data.capacity() - old_len {
39 $enc.data.reserve(MAX_ENCODED_LEN);
42 // SAFETY: The above check and `reserve` ensures that there is enough
43 // room to write the encoded value to the vector's internal buffer.
45 let buf = &mut *($enc.data.as_mut_ptr().add(old_len)
46 as *mut [MaybeUninit<u8>; MAX_ENCODED_LEN]);
47 let encoded = leb128::$fun(buf, $value);
48 $enc.data.set_len(old_len + encoded.len());
53 /// A byte that [cannot occur in UTF8 sequences][utf8]. Used to mark the end of a string.
54 /// This way we can skip validation and still be relatively sure that deserialization
55 /// did not desynchronize.
57 /// [utf8]: https://en.wikipedia.org/w/index.php?title=UTF-8&oldid=1058865525#Codepage_layout
58 const STR_SENTINEL: u8 = 0xC1;
60 impl serialize::Encoder for Encoder {
62 fn emit_usize(&mut self, v: usize) {
63 write_leb128!(self, v, usize, write_usize_leb128)
67 fn emit_u128(&mut self, v: u128) {
68 write_leb128!(self, v, u128, write_u128_leb128);
72 fn emit_u64(&mut self, v: u64) {
73 write_leb128!(self, v, u64, write_u64_leb128);
77 fn emit_u32(&mut self, v: u32) {
78 write_leb128!(self, v, u32, write_u32_leb128);
82 fn emit_u16(&mut self, v: u16) {
83 self.data.extend_from_slice(&v.to_le_bytes());
87 fn emit_u8(&mut self, v: u8) {
92 fn emit_isize(&mut self, v: isize) {
93 write_leb128!(self, v, isize, write_isize_leb128)
97 fn emit_i128(&mut self, v: i128) {
98 write_leb128!(self, v, i128, write_i128_leb128)
102 fn emit_i64(&mut self, v: i64) {
103 write_leb128!(self, v, i64, write_i64_leb128)
107 fn emit_i32(&mut self, v: i32) {
108 write_leb128!(self, v, i32, write_i32_leb128)
112 fn emit_i16(&mut self, v: i16) {
113 self.data.extend_from_slice(&v.to_le_bytes());
117 fn emit_i8(&mut self, v: i8) {
118 self.emit_u8(v as u8);
122 fn emit_bool(&mut self, v: bool) {
123 self.emit_u8(if v { 1 } else { 0 });
127 fn emit_f64(&mut self, v: f64) {
128 let as_u64: u64 = v.to_bits();
129 self.emit_u64(as_u64);
133 fn emit_f32(&mut self, v: f32) {
134 let as_u32: u32 = v.to_bits();
135 self.emit_u32(as_u32);
139 fn emit_char(&mut self, v: char) {
140 self.emit_u32(v as u32);
144 fn emit_str(&mut self, v: &str) {
145 self.emit_usize(v.len());
146 self.emit_raw_bytes(v.as_bytes());
147 self.emit_u8(STR_SENTINEL);
151 fn emit_raw_bytes(&mut self, s: &[u8]) {
152 self.data.extend_from_slice(s);
156 pub type FileEncodeResult = Result<usize, io::Error>;
158 // `FileEncoder` encodes data to file via fixed-size buffer.
160 // When encoding large amounts of data to a file, using `FileEncoder` may be
161 // preferred over using `Encoder` to encode to a `Vec`, and then writing the
162 // `Vec` to file, as the latter uses as much memory as there is encoded data,
163 // while the former uses the fixed amount of memory allocated to the buffer.
164 // `FileEncoder` also has the advantage of not needing to reallocate as data
165 // is appended to it, but the disadvantage of requiring more error handling,
166 // which has some runtime overhead.
167 pub struct FileEncoder {
168 // The input buffer. For adequate performance, we need more control over
169 // buffering than `BufWriter` offers. If `BufWriter` ever offers a raw
170 // buffer access API, we can use it, and remove `buf` and `buffered`.
171 buf: Box<[MaybeUninit<u8>]>,
175 // This is used to implement delayed error handling, as described in the
176 // comment on `trait Encoder`.
177 res: Result<(), io::Error>,
181 pub fn new<P: AsRef<Path>>(path: P) -> io::Result<Self> {
182 const DEFAULT_BUF_SIZE: usize = 8192;
183 FileEncoder::with_capacity(path, DEFAULT_BUF_SIZE)
186 pub fn with_capacity<P: AsRef<Path>>(path: P, capacity: usize) -> io::Result<Self> {
187 // Require capacity at least as large as the largest LEB128 encoding
188 // here, so that we don't have to check or handle this on every write.
189 assert!(capacity >= max_leb128_len());
191 // Require capacity small enough such that some capacity checks can be
192 // done using guaranteed non-overflowing add rather than sub, which
193 // shaves an instruction off those code paths (on x86 at least).
194 assert!(capacity <= usize::MAX - max_leb128_len());
196 let file = File::create(path)?;
199 buf: Box::new_uninit_slice(capacity),
208 pub fn position(&self) -> usize {
209 // Tracking position this way instead of having a `self.position` field
210 // means that we don't have to update the position on every write call.
211 self.flushed + self.buffered
214 pub fn flush(&mut self) {
215 // This is basically a copy of `BufWriter::flush`. If `BufWriter` ever
216 // offers a raw buffer access API, we can use it, and remove this.
218 /// Helper struct to ensure the buffer is updated after all the writes
219 /// are complete. It tracks the number of written bytes and drains them
220 /// all from the front of the buffer when dropped.
221 struct BufGuard<'a> {
222 buffer: &'a mut [u8],
223 encoder_buffered: &'a mut usize,
224 encoder_flushed: &'a mut usize,
228 impl<'a> BufGuard<'a> {
230 buffer: &'a mut [u8],
231 encoder_buffered: &'a mut usize,
232 encoder_flushed: &'a mut usize,
234 assert_eq!(buffer.len(), *encoder_buffered);
235 Self { buffer, encoder_buffered, encoder_flushed, flushed: 0 }
238 /// The unwritten part of the buffer
239 fn remaining(&self) -> &[u8] {
240 &self.buffer[self.flushed..]
243 /// Flag some bytes as removed from the front of the buffer
244 fn consume(&mut self, amt: usize) {
248 /// true if all of the bytes have been written
249 fn done(&self) -> bool {
250 self.flushed >= *self.encoder_buffered
254 impl Drop for BufGuard<'_> {
256 if self.flushed > 0 {
258 *self.encoder_flushed += *self.encoder_buffered;
259 *self.encoder_buffered = 0;
261 self.buffer.copy_within(self.flushed.., 0);
262 *self.encoder_flushed += self.flushed;
263 *self.encoder_buffered -= self.flushed;
269 // If we've already had an error, do nothing. It'll get reported after
270 // `finish` is called.
271 if self.res.is_err() {
275 let mut guard = BufGuard::new(
276 unsafe { MaybeUninit::slice_assume_init_mut(&mut self.buf[..self.buffered]) },
281 while !guard.done() {
282 match self.file.write(guard.remaining()) {
284 self.res = Err(io::Error::new(
285 io::ErrorKind::WriteZero,
286 "failed to write the buffered data",
290 Ok(n) => guard.consume(n),
291 Err(ref e) if e.kind() == io::ErrorKind::Interrupted => {}
301 fn capacity(&self) -> usize {
306 fn write_one(&mut self, value: u8) {
307 // We ensure this during `FileEncoder` construction.
308 debug_assert!(self.capacity() >= 1);
310 let mut buffered = self.buffered;
312 if std::intrinsics::unlikely(buffered >= self.capacity()) {
317 // SAFETY: The above check and `flush` ensures that there is enough
318 // room to write the input to the buffer.
320 *MaybeUninit::slice_as_mut_ptr(&mut self.buf).add(buffered) = value;
323 self.buffered = buffered + 1;
327 fn write_all(&mut self, buf: &[u8]) {
328 let capacity = self.capacity();
329 let buf_len = buf.len();
331 if std::intrinsics::likely(buf_len <= capacity) {
332 let mut buffered = self.buffered;
334 if std::intrinsics::unlikely(buf_len > capacity - buffered) {
339 // SAFETY: The above check and `flush` ensures that there is enough
340 // room to write the input to the buffer.
342 let src = buf.as_ptr();
343 let dst = MaybeUninit::slice_as_mut_ptr(&mut self.buf).add(buffered);
344 ptr::copy_nonoverlapping(src, dst, buf_len);
347 self.buffered = buffered + buf_len;
349 self.write_all_unbuffered(buf);
353 fn write_all_unbuffered(&mut self, mut buf: &[u8]) {
354 // If we've already had an error, do nothing. It'll get reported after
355 // `finish` is called.
356 if self.res.is_err() {
360 if self.buffered > 0 {
364 // This is basically a copy of `Write::write_all` but also updates our
365 // `self.flushed`. It's necessary because `Write::write_all` does not
366 // return the number of bytes written when an error is encountered, and
367 // without that, we cannot accurately update `self.flushed` on error.
368 while !buf.is_empty() {
369 match self.file.write(buf) {
371 self.res = Err(io::Error::new(
372 io::ErrorKind::WriteZero,
373 "failed to write whole buffer",
381 Err(ref e) if e.kind() == io::ErrorKind::Interrupted => {}
390 pub fn finish(mut self) -> Result<usize, io::Error> {
393 let res = std::mem::replace(&mut self.res, Ok(()));
394 res.map(|()| self.position())
398 impl Drop for FileEncoder {
400 // Likely to be a no-op, because `finish` should have been called and
401 // it also flushes. But do it just in case.
402 let _result = self.flush();
406 macro_rules! file_encoder_write_leb128 {
407 ($enc:expr, $value:expr, $int_ty:ty, $fun:ident) => {{
408 const MAX_ENCODED_LEN: usize = max_leb128_len!($int_ty);
410 // We ensure this during `FileEncoder` construction.
411 debug_assert!($enc.capacity() >= MAX_ENCODED_LEN);
413 let mut buffered = $enc.buffered;
415 // This can't overflow. See assertion in `FileEncoder::with_capacity`.
416 if std::intrinsics::unlikely(buffered + MAX_ENCODED_LEN > $enc.capacity()) {
421 // SAFETY: The above check and flush ensures that there is enough
422 // room to write the encoded value to the buffer.
424 &mut *($enc.buf.as_mut_ptr().add(buffered) as *mut [MaybeUninit<u8>; MAX_ENCODED_LEN])
427 let encoded = leb128::$fun(buf, $value);
428 $enc.buffered = buffered + encoded.len();
432 impl serialize::Encoder for FileEncoder {
434 fn emit_usize(&mut self, v: usize) {
435 file_encoder_write_leb128!(self, v, usize, write_usize_leb128)
439 fn emit_u128(&mut self, v: u128) {
440 file_encoder_write_leb128!(self, v, u128, write_u128_leb128)
444 fn emit_u64(&mut self, v: u64) {
445 file_encoder_write_leb128!(self, v, u64, write_u64_leb128)
449 fn emit_u32(&mut self, v: u32) {
450 file_encoder_write_leb128!(self, v, u32, write_u32_leb128)
454 fn emit_u16(&mut self, v: u16) {
455 self.write_all(&v.to_le_bytes());
459 fn emit_u8(&mut self, v: u8) {
464 fn emit_isize(&mut self, v: isize) {
465 file_encoder_write_leb128!(self, v, isize, write_isize_leb128)
469 fn emit_i128(&mut self, v: i128) {
470 file_encoder_write_leb128!(self, v, i128, write_i128_leb128)
474 fn emit_i64(&mut self, v: i64) {
475 file_encoder_write_leb128!(self, v, i64, write_i64_leb128)
479 fn emit_i32(&mut self, v: i32) {
480 file_encoder_write_leb128!(self, v, i32, write_i32_leb128)
484 fn emit_i16(&mut self, v: i16) {
485 self.write_all(&v.to_le_bytes());
489 fn emit_i8(&mut self, v: i8) {
490 self.emit_u8(v as u8);
494 fn emit_bool(&mut self, v: bool) {
495 self.emit_u8(if v { 1 } else { 0 });
499 fn emit_f64(&mut self, v: f64) {
500 let as_u64: u64 = v.to_bits();
501 self.emit_u64(as_u64);
505 fn emit_f32(&mut self, v: f32) {
506 let as_u32: u32 = v.to_bits();
507 self.emit_u32(as_u32);
511 fn emit_char(&mut self, v: char) {
512 self.emit_u32(v as u32);
516 fn emit_str(&mut self, v: &str) {
517 self.emit_usize(v.len());
518 self.emit_raw_bytes(v.as_bytes());
519 self.emit_u8(STR_SENTINEL);
523 fn emit_raw_bytes(&mut self, s: &[u8]) {
528 // -----------------------------------------------------------------------------
530 // -----------------------------------------------------------------------------
532 pub struct Decoder<'a> {
537 impl<'a> Decoder<'a> {
539 pub fn new(data: &'a [u8], position: usize) -> Decoder<'a> {
540 Decoder { data, position }
544 pub fn position(&self) -> usize {
549 pub fn set_position(&mut self, pos: usize) {
554 pub fn advance(&mut self, bytes: usize) {
555 self.position += bytes;
559 macro_rules! read_leb128 {
560 ($dec:expr, $fun:ident) => {{ leb128::$fun($dec.data, &mut $dec.position) }};
563 impl<'a> serialize::Decoder for Decoder<'a> {
565 fn read_u128(&mut self) -> u128 {
566 read_leb128!(self, read_u128_leb128)
570 fn read_u64(&mut self) -> u64 {
571 read_leb128!(self, read_u64_leb128)
575 fn read_u32(&mut self) -> u32 {
576 read_leb128!(self, read_u32_leb128)
580 fn read_u16(&mut self) -> u16 {
581 let bytes = [self.data[self.position], self.data[self.position + 1]];
582 let value = u16::from_le_bytes(bytes);
588 fn read_u8(&mut self) -> u8 {
589 let value = self.data[self.position];
595 fn read_usize(&mut self) -> usize {
596 read_leb128!(self, read_usize_leb128)
600 fn read_i128(&mut self) -> i128 {
601 read_leb128!(self, read_i128_leb128)
605 fn read_i64(&mut self) -> i64 {
606 read_leb128!(self, read_i64_leb128)
610 fn read_i32(&mut self) -> i32 {
611 read_leb128!(self, read_i32_leb128)
615 fn read_i16(&mut self) -> i16 {
616 let bytes = [self.data[self.position], self.data[self.position + 1]];
617 let value = i16::from_le_bytes(bytes);
623 fn read_i8(&mut self) -> i8 {
624 let value = self.data[self.position];
630 fn read_isize(&mut self) -> isize {
631 read_leb128!(self, read_isize_leb128)
635 fn read_bool(&mut self) -> bool {
636 let value = self.read_u8();
641 fn read_f64(&mut self) -> f64 {
642 let bits = self.read_u64();
647 fn read_f32(&mut self) -> f32 {
648 let bits = self.read_u32();
653 fn read_char(&mut self) -> char {
654 let bits = self.read_u32();
655 std::char::from_u32(bits).unwrap()
659 fn read_str(&mut self) -> &'a str {
660 let len = self.read_usize();
661 let sentinel = self.data[self.position + len];
662 assert!(sentinel == STR_SENTINEL);
664 std::str::from_utf8_unchecked(&self.data[self.position..self.position + len])
666 self.position += len + 1;
671 fn read_raw_bytes(&mut self, bytes: usize) -> &'a [u8] {
672 let start = self.position;
673 self.position += bytes;
674 &self.data[start..self.position]
678 // Specializations for contiguous byte sequences follow. The default implementations for slices
679 // encode and decode each element individually. This isn't necessary for `u8` slices when using
680 // opaque encoders and decoders, because each `u8` is unchanged by encoding and decoding.
681 // Therefore, we can use more efficient implementations that process the entire sequence at once.
683 // Specialize encoding byte slices. This specialization also applies to encoding `Vec<u8>`s, etc.,
684 // since the default implementations call `encode` on their slices internally.
685 impl serialize::Encodable<Encoder> for [u8] {
686 fn encode(&self, e: &mut Encoder) {
687 serialize::Encoder::emit_usize(e, self.len());
688 e.emit_raw_bytes(self);
692 impl serialize::Encodable<FileEncoder> for [u8] {
693 fn encode(&self, e: &mut FileEncoder) {
694 serialize::Encoder::emit_usize(e, self.len());
695 e.emit_raw_bytes(self);
699 // Specialize decoding `Vec<u8>`. This specialization also applies to decoding `Box<[u8]>`s, etc.,
700 // since the default implementations call `decode` to produce a `Vec<u8>` internally.
701 impl<'a> serialize::Decodable<Decoder<'a>> for Vec<u8> {
702 fn decode(d: &mut Decoder<'a>) -> Self {
703 let len = serialize::Decoder::read_usize(d);
704 d.read_raw_bytes(len).to_owned()
708 // An integer that will always encode to 8 bytes.
709 pub struct IntEncodedWithFixedSize(pub u64);
711 impl IntEncodedWithFixedSize {
712 pub const ENCODED_SIZE: usize = 8;
715 impl serialize::Encodable<Encoder> for IntEncodedWithFixedSize {
717 fn encode(&self, e: &mut Encoder) {
718 let _start_pos = e.position();
719 e.emit_raw_bytes(&self.0.to_le_bytes());
720 let _end_pos = e.position();
721 debug_assert_eq!((_end_pos - _start_pos), IntEncodedWithFixedSize::ENCODED_SIZE);
725 impl serialize::Encodable<FileEncoder> for IntEncodedWithFixedSize {
727 fn encode(&self, e: &mut FileEncoder) {
728 let _start_pos = e.position();
729 e.emit_raw_bytes(&self.0.to_le_bytes());
730 let _end_pos = e.position();
731 debug_assert_eq!((_end_pos - _start_pos), IntEncodedWithFixedSize::ENCODED_SIZE);
735 impl<'a> serialize::Decodable<Decoder<'a>> for IntEncodedWithFixedSize {
737 fn decode(decoder: &mut Decoder<'a>) -> IntEncodedWithFixedSize {
738 let _start_pos = decoder.position();
739 let bytes = decoder.read_raw_bytes(IntEncodedWithFixedSize::ENCODED_SIZE);
740 let value = u64::from_le_bytes(bytes.try_into().unwrap());
741 let _end_pos = decoder.position();
742 debug_assert_eq!((_end_pos - _start_pos), IntEncodedWithFixedSize::ENCODED_SIZE);
744 IntEncodedWithFixedSize(value)