| 1 |
use crate::{ |
| 2 |
engine::{general_purpose::INVALID_VALUE, DecodeEstimate, DecodeMetadata, DecodePaddingMode}, |
| 3 |
DecodeError, PAD_BYTE, |
| 4 |
}; |
| 5 |
|
| 6 |
// decode logic operates on chunks of 8 input bytes without padding |
| 7 |
const INPUT_CHUNK_LEN: usize = 8; |
| 8 |
const DECODED_CHUNK_LEN: usize = 6; |
| 9 |
|
| 10 |
// we read a u64 and write a u64, but a u64 of input only yields 6 bytes of output, so the last |
| 11 |
// 2 bytes of any output u64 should not be counted as written to (but must be available in a |
| 12 |
// slice). |
| 13 |
const DECODED_CHUNK_SUFFIX: usize = 2; |
| 14 |
|
| 15 |
// how many u64's of input to handle at a time |
| 16 |
const CHUNKS_PER_FAST_LOOP_BLOCK: usize = 4; |
| 17 |
|
| 18 |
const INPUT_BLOCK_LEN: usize = CHUNKS_PER_FAST_LOOP_BLOCK * INPUT_CHUNK_LEN; |
| 19 |
|
| 20 |
// includes the trailing 2 bytes for the final u64 write |
| 21 |
const DECODED_BLOCK_LEN: usize = |
| 22 |
CHUNKS_PER_FAST_LOOP_BLOCK * DECODED_CHUNK_LEN + DECODED_CHUNK_SUFFIX; |
| 23 |
|
| 24 |
#[doc(hidden)] |
| 25 |
pub struct GeneralPurposeEstimate { |
| 26 |
/// Total number of decode chunks, including a possibly partial last chunk |
| 27 |
num_chunks: usize, |
| 28 |
decoded_len_estimate: usize, |
| 29 |
} |
| 30 |
|
| 31 |
impl GeneralPurposeEstimate { |
| 32 |
pub(crate) fn new(encoded_len: usize) -> Self { |
| 33 |
// Formulas that won't overflow |
| 34 |
Self { |
| 35 |
num_chunks: encoded_len / INPUT_CHUNK_LEN |
| 36 |
+ (encoded_len % INPUT_CHUNK_LEN > 0) as usize, |
| 37 |
decoded_len_estimate: (encoded_len / 4 + (encoded_len % 4 > 0) as usize) * 3, |
| 38 |
} |
| 39 |
} |
| 40 |
} |
| 41 |
|
| 42 |
impl DecodeEstimate for GeneralPurposeEstimate { |
| 43 |
fn decoded_len_estimate(&self) -> usize { |
| 44 |
self.decoded_len_estimate |
| 45 |
} |
| 46 |
} |
| 47 |
|
| 48 |
/// Helper to avoid duplicating num_chunks calculation, which is costly on short inputs. |
| 49 |
/// Returns the decode metadata, or an error. |
| 50 |
// We're on the fragile edge of compiler heuristics here. If this is not inlined, slow. If this is |
| 51 |
// inlined(always), a different slow. plain ol' inline makes the benchmarks happiest at the moment, |
| 52 |
// but this is fragile and the best setting changes with only minor code modifications. |
| 53 |
#[inline] |
| 54 |
pub(crate) fn decode_helper( |
| 55 |
input: &[u8], |
| 56 |
estimate: GeneralPurposeEstimate, |
| 57 |
output: &mut [u8], |
| 58 |
decode_table: &[u8; 256], |
| 59 |
decode_allow_trailing_bits: bool, |
| 60 |
padding_mode: DecodePaddingMode, |
| 61 |
) -> Result<DecodeMetadata, DecodeError> { |
| 62 |
let remainder_len = input.len() % INPUT_CHUNK_LEN; |
| 63 |
|
| 64 |
// Because the fast decode loop writes in groups of 8 bytes (unrolled to |
| 65 |
// CHUNKS_PER_FAST_LOOP_BLOCK times 8 bytes, where possible) and outputs 8 bytes at a time (of |
| 66 |
// which only 6 are valid data), we need to be sure that we stop using the fast decode loop |
| 67 |
// soon enough that there will always be 2 more bytes of valid data written after that loop. |
| 68 |
let trailing_bytes_to_skip = match remainder_len { |
| 69 |
// if input is a multiple of the chunk size, ignore the last chunk as it may have padding, |
| 70 |
// and the fast decode logic cannot handle padding |
| 71 |
0 => INPUT_CHUNK_LEN, |
| 72 |
// 1 and 5 trailing bytes are illegal: can't decode 6 bits of input into a byte |
| 73 |
1 | 5 => { |
| 74 |
// trailing whitespace is so common that it's worth it to check the last byte to |
| 75 |
// possibly return a better error message |
| 76 |
if let Some(b) = input.last() { |
| 77 |
if *b != PAD_BYTE && decode_table[*b as usize] == INVALID_VALUE { |
| 78 |
return Err(DecodeError::InvalidByte(input.len() - 1, *b)); |
| 79 |
} |
| 80 |
} |
| 81 |
|
| 82 |
return Err(DecodeError::InvalidLength); |
| 83 |
} |
| 84 |
// This will decode to one output byte, which isn't enough to overwrite the 2 extra bytes |
| 85 |
// written by the fast decode loop. So, we have to ignore both these 2 bytes and the |
| 86 |
// previous chunk. |
| 87 |
2 => INPUT_CHUNK_LEN + 2, |
| 88 |
// If this is 3 un-padded chars, then it would actually decode to 2 bytes. However, if this |
| 89 |
// is an erroneous 2 chars + 1 pad char that would decode to 1 byte, then it should fail |
| 90 |
// with an error, not panic from going past the bounds of the output slice, so we let it |
| 91 |
// use stage 3 + 4. |
| 92 |
3 => INPUT_CHUNK_LEN + 3, |
| 93 |
// This can also decode to one output byte because it may be 2 input chars + 2 padding |
| 94 |
// chars, which would decode to 1 byte. |
| 95 |
4 => INPUT_CHUNK_LEN + 4, |
| 96 |
// Everything else is a legal decode len (given that we don't require padding), and will |
| 97 |
// decode to at least 2 bytes of output. |
| 98 |
_ => remainder_len, |
| 99 |
}; |
| 100 |
|
| 101 |
// rounded up to include partial chunks |
| 102 |
let mut remaining_chunks = estimate.num_chunks; |
| 103 |
|
| 104 |
let mut input_index = 0; |
| 105 |
let mut output_index = 0; |
| 106 |
|
| 107 |
{ |
| 108 |
let length_of_fast_decode_chunks = input.len().saturating_sub(trailing_bytes_to_skip); |
| 109 |
|
| 110 |
// Fast loop, stage 1 |
| 111 |
// manual unroll to CHUNKS_PER_FAST_LOOP_BLOCK of u64s to amortize slice bounds checks |
| 112 |
if let Some(max_start_index) = length_of_fast_decode_chunks.checked_sub(INPUT_BLOCK_LEN) { |
| 113 |
while input_index <= max_start_index { |
| 114 |
let input_slice = &input[input_index..(input_index + INPUT_BLOCK_LEN)]; |
| 115 |
let output_slice = &mut output[output_index..(output_index + DECODED_BLOCK_LEN)]; |
| 116 |
|
| 117 |
decode_chunk( |
| 118 |
&input_slice[0..], |
| 119 |
input_index, |
| 120 |
decode_table, |
| 121 |
&mut output_slice[0..], |
| 122 |
)?; |
| 123 |
decode_chunk( |
| 124 |
&input_slice[8..], |
| 125 |
input_index + 8, |
| 126 |
decode_table, |
| 127 |
&mut output_slice[6..], |
| 128 |
)?; |
| 129 |
decode_chunk( |
| 130 |
&input_slice[16..], |
| 131 |
input_index + 16, |
| 132 |
decode_table, |
| 133 |
&mut output_slice[12..], |
| 134 |
)?; |
| 135 |
decode_chunk( |
| 136 |
&input_slice[24..], |
| 137 |
input_index + 24, |
| 138 |
decode_table, |
| 139 |
&mut output_slice[18..], |
| 140 |
)?; |
| 141 |
|
| 142 |
input_index += INPUT_BLOCK_LEN; |
| 143 |
output_index += DECODED_BLOCK_LEN - DECODED_CHUNK_SUFFIX; |
| 144 |
remaining_chunks -= CHUNKS_PER_FAST_LOOP_BLOCK; |
| 145 |
} |
| 146 |
} |
| 147 |
|
| 148 |
// Fast loop, stage 2 (aka still pretty fast loop) |
| 149 |
// 8 bytes at a time for whatever we didn't do in stage 1. |
| 150 |
if let Some(max_start_index) = length_of_fast_decode_chunks.checked_sub(INPUT_CHUNK_LEN) { |
| 151 |
while input_index < max_start_index { |
| 152 |
decode_chunk( |
| 153 |
&input[input_index..(input_index + INPUT_CHUNK_LEN)], |
| 154 |
input_index, |
| 155 |
decode_table, |
| 156 |
&mut output |
| 157 |
[output_index..(output_index + DECODED_CHUNK_LEN + DECODED_CHUNK_SUFFIX)], |
| 158 |
)?; |
| 159 |
|
| 160 |
output_index += DECODED_CHUNK_LEN; |
| 161 |
input_index += INPUT_CHUNK_LEN; |
| 162 |
remaining_chunks -= 1; |
| 163 |
} |
| 164 |
} |
| 165 |
} |
| 166 |
|
| 167 |
// Stage 3 |
| 168 |
// If input length was such that a chunk had to be deferred until after the fast loop |
| 169 |
// because decoding it would have produced 2 trailing bytes that wouldn't then be |
| 170 |
// overwritten, we decode that chunk here. This way is slower but doesn't write the 2 |
| 171 |
// trailing bytes. |
| 172 |
// However, we still need to avoid the last chunk (partial or complete) because it could |
| 173 |
// have padding, so we always do 1 fewer to avoid the last chunk. |
| 174 |
for _ in 1..remaining_chunks { |
| 175 |
decode_chunk_precise( |
| 176 |
&input[input_index..], |
| 177 |
input_index, |
| 178 |
decode_table, |
| 179 |
&mut output[output_index..(output_index + DECODED_CHUNK_LEN)], |
| 180 |
)?; |
| 181 |
|
| 182 |
input_index += INPUT_CHUNK_LEN; |
| 183 |
output_index += DECODED_CHUNK_LEN; |
| 184 |
} |
| 185 |
|
| 186 |
// always have one more (possibly partial) block of 8 input |
| 187 |
debug_assert!(input.len() - input_index > 1 || input.is_empty()); |
| 188 |
debug_assert!(input.len() - input_index <= 8); |
| 189 |
|
| 190 |
super::decode_suffix::decode_suffix( |
| 191 |
input, |
| 192 |
input_index, |
| 193 |
output, |
| 194 |
output_index, |
| 195 |
decode_table, |
| 196 |
decode_allow_trailing_bits, |
| 197 |
padding_mode, |
| 198 |
) |
| 199 |
} |
| 200 |
|
| 201 |
/// Decode 8 bytes of input into 6 bytes of output. 8 bytes of output will be written, but only the |
| 202 |
/// first 6 of those contain meaningful data. |
| 203 |
/// |
| 204 |
/// `input` is the bytes to decode, of which the first 8 bytes will be processed. |
| 205 |
/// `index_at_start_of_input` is the offset in the overall input (used for reporting errors |
| 206 |
/// accurately) |
| 207 |
/// `decode_table` is the lookup table for the particular base64 alphabet. |
| 208 |
/// `output` will have its first 8 bytes overwritten, of which only the first 6 are valid decoded |
| 209 |
/// data. |
| 210 |
// yes, really inline (worth 30-50% speedup) |
| 211 |
#[inline(always)] |
| 212 |
fn decode_chunk( |
| 213 |
input: &[u8], |
| 214 |
index_at_start_of_input: usize, |
| 215 |
decode_table: &[u8; 256], |
| 216 |
output: &mut [u8], |
| 217 |
) -> Result<(), DecodeError> { |
| 218 |
let morsel = decode_table[input[0] as usize]; |
| 219 |
if morsel == INVALID_VALUE { |
| 220 |
return Err(DecodeError::InvalidByte(index_at_start_of_input, input[0])); |
| 221 |
} |
| 222 |
let mut accum = (morsel as u64) << 58; |
| 223 |
|
| 224 |
let morsel = decode_table[input[1] as usize]; |
| 225 |
if morsel == INVALID_VALUE { |
| 226 |
return Err(DecodeError::InvalidByte( |
| 227 |
index_at_start_of_input + 1, |
| 228 |
input[1], |
| 229 |
)); |
| 230 |
} |
| 231 |
accum |= (morsel as u64) << 52; |
| 232 |
|
| 233 |
let morsel = decode_table[input[2] as usize]; |
| 234 |
if morsel == INVALID_VALUE { |
| 235 |
return Err(DecodeError::InvalidByte( |
| 236 |
index_at_start_of_input + 2, |
| 237 |
input[2], |
| 238 |
)); |
| 239 |
} |
| 240 |
accum |= (morsel as u64) << 46; |
| 241 |
|
| 242 |
let morsel = decode_table[input[3] as usize]; |
| 243 |
if morsel == INVALID_VALUE { |
| 244 |
return Err(DecodeError::InvalidByte( |
| 245 |
index_at_start_of_input + 3, |
| 246 |
input[3], |
| 247 |
)); |
| 248 |
} |
| 249 |
accum |= (morsel as u64) << 40; |
| 250 |
|
| 251 |
let morsel = decode_table[input[4] as usize]; |
| 252 |
if morsel == INVALID_VALUE { |
| 253 |
return Err(DecodeError::InvalidByte( |
| 254 |
index_at_start_of_input + 4, |
| 255 |
input[4], |
| 256 |
)); |
| 257 |
} |
| 258 |
accum |= (morsel as u64) << 34; |
| 259 |
|
| 260 |
let morsel = decode_table[input[5] as usize]; |
| 261 |
if morsel == INVALID_VALUE { |
| 262 |
return Err(DecodeError::InvalidByte( |
| 263 |
index_at_start_of_input + 5, |
| 264 |
input[5], |
| 265 |
)); |
| 266 |
} |
| 267 |
accum |= (morsel as u64) << 28; |
| 268 |
|
| 269 |
let morsel = decode_table[input[6] as usize]; |
| 270 |
if morsel == INVALID_VALUE { |
| 271 |
return Err(DecodeError::InvalidByte( |
| 272 |
index_at_start_of_input + 6, |
| 273 |
input[6], |
| 274 |
)); |
| 275 |
} |
| 276 |
accum |= (morsel as u64) << 22; |
| 277 |
|
| 278 |
let morsel = decode_table[input[7] as usize]; |
| 279 |
if morsel == INVALID_VALUE { |
| 280 |
return Err(DecodeError::InvalidByte( |
| 281 |
index_at_start_of_input + 7, |
| 282 |
input[7], |
| 283 |
)); |
| 284 |
} |
| 285 |
accum |= (morsel as u64) << 16; |
| 286 |
|
| 287 |
write_u64(output, accum); |
| 288 |
|
| 289 |
Ok(()) |
| 290 |
} |
| 291 |
|
| 292 |
/// Decode an 8-byte chunk, but only write the 6 bytes actually decoded instead of including 2 |
| 293 |
/// trailing garbage bytes. |
| 294 |
#[inline] |
| 295 |
fn decode_chunk_precise( |
| 296 |
input: &[u8], |
| 297 |
index_at_start_of_input: usize, |
| 298 |
decode_table: &[u8; 256], |
| 299 |
output: &mut [u8], |
| 300 |
) -> Result<(), DecodeError> { |
| 301 |
let mut tmp_buf = [0_u8; 8]; |
| 302 |
|
| 303 |
decode_chunk( |
| 304 |
input, |
| 305 |
index_at_start_of_input, |
| 306 |
decode_table, |
| 307 |
&mut tmp_buf[..], |
| 308 |
)?; |
| 309 |
|
| 310 |
output[0..6].copy_from_slice(&tmp_buf[0..6]); |
| 311 |
|
| 312 |
Ok(()) |
| 313 |
} |
| 314 |
|
| 315 |
#[inline] |
| 316 |
fn write_u64(output: &mut [u8], value: u64) { |
| 317 |
output[..8].copy_from_slice(&value.to_be_bytes()); |
| 318 |
} |
| 319 |
|
| 320 |
#[cfg(test)] |
| 321 |
mod tests { |
| 322 |
use super::*; |
| 323 |
|
| 324 |
use crate::engine::general_purpose::STANDARD; |
| 325 |
|
| 326 |
#[test] |
| 327 |
fn decode_chunk_precise_writes_only_6_bytes() { |
| 328 |
let input = b"Zm9vYmFy"; // "foobar" |
| 329 |
let mut output = [0_u8, 1, 2, 3, 4, 5, 6, 7]; |
| 330 |
|
| 331 |
decode_chunk_precise(&input[..], 0, &STANDARD.decode_table, &mut output).unwrap(); |
| 332 |
assert_eq!(&vec![b'f', b'o', b'o', b'b', b'a', b'r', 6, 7], &output); |
| 333 |
} |
| 334 |
|
| 335 |
#[test] |
| 336 |
fn decode_chunk_writes_8_bytes() { |
| 337 |
let input = b"Zm9vYmFy"; // "foobar" |
| 338 |
let mut output = [0_u8, 1, 2, 3, 4, 5, 6, 7]; |
| 339 |
|
| 340 |
decode_chunk(&input[..], 0, &STANDARD.decode_table, &mut output).unwrap(); |
| 341 |
assert_eq!(&vec![b'f', b'o', b'o', b'b', b'a', b'r', 0, 0], &output); |
| 342 |
} |
| 343 |
|
| 344 |
#[test] |
| 345 |
fn estimate_short_lengths() { |
| 346 |
for (range, (num_chunks, decoded_len_estimate)) in [ |
| 347 |
(0..=0, (0, 0)), |
| 348 |
(1..=4, (1, 3)), |
| 349 |
(5..=8, (1, 6)), |
| 350 |
(9..=12, (2, 9)), |
| 351 |
(13..=16, (2, 12)), |
| 352 |
(17..=20, (3, 15)), |
| 353 |
] { |
| 354 |
for encoded_len in range { |
| 355 |
let estimate = GeneralPurposeEstimate::new(encoded_len); |
| 356 |
assert_eq!(num_chunks, estimate.num_chunks); |
| 357 |
assert_eq!(decoded_len_estimate, estimate.decoded_len_estimate); |
| 358 |
} |
| 359 |
} |
| 360 |
} |
| 361 |
|
| 362 |
#[test] |
| 363 |
fn estimate_via_u128_inflation() { |
| 364 |
// cover both ends of usize |
| 365 |
(0..1000) |
| 366 |
.chain(usize::MAX - 1000..=usize::MAX) |
| 367 |
.for_each(|encoded_len| { |
| 368 |
// inflate to 128 bit type to be able to safely use the easy formulas |
| 369 |
let len_128 = encoded_len as u128; |
| 370 |
|
| 371 |
let estimate = GeneralPurposeEstimate::new(encoded_len); |
| 372 |
assert_eq!( |
| 373 |
((len_128 + (INPUT_CHUNK_LEN - 1) as u128) / (INPUT_CHUNK_LEN as u128)) |
| 374 |
as usize, |
| 375 |
estimate.num_chunks |
| 376 |
); |
| 377 |
assert_eq!( |
| 378 |
((len_128 + 3) / 4 * 3) as usize, |
| 379 |
estimate.decoded_len_estimate |
| 380 |
); |
| 381 |
}) |
| 382 |
} |
| 383 |
} |