1
   2
   3
   4
   5
   6
   7
   8
   9
  10
  11
  12
  13
  14
  15
  16
  17
  18
  19
  20
  21
  22
  23
  24
  25
  26
  27
  28
  29
  30
  31
  32
  33
  34
  35
  36
  37
  38
  39
  40
  41
  42
  43
  44
  45
  46
  47
  48
  49
  50
  51
  52
  53
  54
  55
  56
  57
  58
  59
  60
  61
  62
  63
  64
  65
  66
  67
  68
  69
  70
  71
  72
  73
  74
  75
  76
  77
  78
  79
  80
  81
  82
  83
  84
  85
  86
  87
  88
  89
  90
  91
  92
  93
  94
  95
  96
  97
  98
  99
 100
 101
 102
 103
 104
 105
 106
 107
 108
 109
 110
 111
 112
 113
 114
 115
 116
 117
 118
 119
 120
 121
 122
 123
 124
 125
 126
 127
 128
 129
 130
 131
 132
 133
 134
 135
 136
 137
 138
 139
 140
 141
 142
 143
 144
 145
 146
 147
 148
 149
 150
 151
 152
 153
 154
 155
 156
 157
 158
 159
 160
 161
 162
 163
 164
 165
 166
 167
 168
 169
 170
 171
 172
 173
 174
 175
 176
 177
 178
 179
 180
 181
 182
 183
 184
 185
 186
 187
 188
 189
 190
 191
 192
 193
 194
 195
 196
 197
 198
 199
 200
 201
 202
 203
 204
 205
 206
 207
 208
 209
 210
 211
 212
 213
 214
 215
 216
 217
 218
 219
 220
 221
 222
 223
 224
 225
 226
 227
 228
 229
 230
 231
 232
 233
 234
 235
 236
 237
 238
 239
 240
 241
 242
 243
 244
 245
 246
 247
 248
 249
 250
 251
 252
 253
 254
 255
 256
 257
 258
 259
 260
 261
 262
 263
 264
 265
 266
 267
 268
 269
 270
 271
 272
 273
 274
 275
 276
 277
 278
 279
 280
 281
 282
 283
 284
 285
 286
 287
 288
 289
 290
 291
 292
 293
 294
 295
 296
 297
 298
 299
 300
 301
 302
 303
 304
 305
 306
 307
 308
 309
 310
 311
 312
 313
 314
 315
 316
 317
 318
 319
 320
 321
 322
 323
 324
 325
 326
 327
 328
 329
 330
 331
 332
 333
 334
 335
 336
 337
 338
 339
 340
 341
 342
 343
 344
 345
 346
 347
 348
 349
 350
 351
 352
 353
 354
 355
 356
 357
 358
 359
 360
 361
 362
 363
 364
 365
 366
 367
 368
 369
 370
 371
 372
 373
 374
 375
 376
 377
 378
 379
 380
 381
 382
 383
 384
 385
 386
 387
 388
 389
 390
 391
 392
 393
 394
 395
 396
 397
 398
 399
 400
 401
 402
 403
 404
 405
 406
 407
 408
 409
 410
 411
 412
 413
 414
 415
 416
 417
 418
 419
 420
 421
 422
 423
 424
 425
 426
 427
 428
 429
 430
 431
 432
 433
 434
 435
 436
 437
 438
 439
 440
 441
 442
 443
 444
 445
 446
 447
 448
 449
 450
 451
 452
 453
 454
 455
 456
 457
 458
 459
 460
 461
 462
 463
 464
 465
 466
 467
 468
 469
 470
 471
 472
 473
 474
 475
 476
 477
 478
 479
 480
 481
 482
 483
 484
 485
 486
 487
 488
 489
 490
 491
 492
 493
 494
 495
 496
 497
 498
 499
 500
 501
 502
 503
 504
 505
 506
 507
 508
 509
 510
 511
 512
 513
 514
 515
 516
 517
 518
 519
 520
 521
 522
 523
 524
 525
 526
 527
 528
 529
 530
 531
 532
 533
 534
 535
 536
 537
 538
 539
 540
 541
 542
 543
 544
 545
 546
 547
 548
 549
 550
 551
 552
 553
 554
 555
 556
 557
 558
 559
 560
 561
 562
 563
 564
 565
 566
 567
 568
 569
 570
 571
 572
 573
 574
 575
 576
 577
 578
 579
 580
 581
 582
 583
 584
 585
 586
 587
 588
 589
 590
 591
 592
 593
 594
 595
 596
 597
 598
 599
 600
 601
 602
 603
 604
 605
 606
 607
 608
 609
 610
 611
 612
 613
 614
 615
 616
 617
 618
 619
 620
 621
 622
 623
 624
 625
 626
 627
 628
 629
 630
 631
 632
 633
 634
 635
 636
 637
 638
 639
 640
 641
 642
 643
 644
 645
 646
 647
 648
 649
 650
 651
 652
 653
 654
 655
 656
 657
 658
 659
 660
 661
 662
 663
 664
 665
 666
 667
 668
 669
 670
 671
 672
 673
 674
 675
 676
 677
 678
 679
 680
 681
 682
 683
 684
 685
 686
 687
 688
 689
 690
 691
 692
 693
 694
 695
 696
 697
 698
 699
 700
 701
 702
 703
 704
 705
 706
 707
 708
 709
 710
 711
 712
 713
 714
 715
 716
 717
 718
 719
 720
 721
 722
 723
 724
 725
 726
 727
 728
 729
 730
 731
 732
 733
 734
 735
 736
 737
 738
 739
 740
 741
 742
 743
 744
 745
 746
 747
 748
 749
 750
 751
 752
 753
 754
 755
 756
 757
 758
 759
 760
 761
 762
 763
 764
 765
 766
 767
 768
 769
 770
 771
 772
 773
 774
 775
 776
 777
 778
 779
 780
 781
 782
 783
 784
 785
 786
 787
 788
 789
 790
 791
 792
 793
 794
 795
 796
 797
 798
 799
 800
 801
 802
 803
 804
 805
 806
 807
 808
 809
 810
 811
 812
 813
 814
 815
 816
 817
 818
 819
 820
 821
 822
 823
 824
 825
 826
 827
 828
 829
 830
 831
 832
 833
 834
 835
 836
 837
 838
 839
 840
 841
 842
 843
 844
 845
 846
 847
 848
 849
 850
 851
 852
 853
 854
 855
 856
 857
 858
 859
 860
 861
 862
 863
 864
 865
 866
 867
 868
 869
 870
 871
 872
 873
 874
 875
 876
 877
 878
 879
 880
 881
 882
 883
 884
 885
 886
 887
 888
 889
 890
 891
 892
 893
 894
 895
 896
 897
 898
 899
 900
 901
 902
 903
 904
 905
 906
 907
 908
 909
 910
 911
 912
 913
 914
 915
 916
 917
 918
 919
 920
 921
 922
 923
 924
 925
 926
 927
 928
 929
 930
 931
 932
 933
 934
 935
 936
 937
 938
 939
 940
 941
 942
 943
 944
 945
 946
 947
 948
 949
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
1089
1090
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
//! The official Rust implementation of the [BLAKE3] cryptographic hash
//! function.
//!
//! # Examples
//!
//! ```
//! # fn main() -> Result<(), Box<dyn std::error::Error>> {
//! // Hash an input all at once.
//! let hash1 = blake3::hash(b"foobarbaz");
//!
//! // Hash an input incrementally.
//! let mut hasher = blake3::Hasher::new();
//! hasher.update(b"foo");
//! hasher.update(b"bar");
//! hasher.update(b"baz");
//! let hash2 = hasher.finalize();
//! assert_eq!(hash1, hash2);
//!
//! // Extended output. OutputReader also implements Read and Seek.
//! # #[cfg(feature = "std")] {
//! let mut output = [0; 1000];
//! let mut output_reader = hasher.finalize_xof();
//! output_reader.fill(&mut output);
//! assert_eq!(&output[..32], hash1.as_bytes());
//! # }
//!
//! // Print a hash as hex.
//! println!("{}", hash1.to_hex());
//! # Ok(())
//! # }
//! ```
//!
//! # Cargo Features
//!
//! The `rayon` feature provides [Rayon]-based multi-threading, in particular
//! the [`join::RayonJoin`] type for use with [`Hasher::update_with_join`]. It
//! is disabled by default, but enabled for [docs.rs].
//!
//! The `neon` feature enables ARM NEON support. Currently there is no runtime
//! CPU feature detection for NEON, so you must only enable this feature for
//! targets that are known to have NEON support. In particular, some ARMv7
//! targets support NEON, and some don't.
//!
//! The `std` feature (enabled by default) is required for implementations of
//! the [`Write`] and [`Seek`] traits, and also for runtime CPU feature
//! detection. If this feature is disabled, the only way to use the SIMD
//! implementations in this crate is to enable the corresponding instruction
//! sets statically for the entire build, with e.g. `RUSTFLAGS="-C
//! target-cpu=native"`. The resulting binary will not be portable to other
//! machines.
//!
//! [BLAKE3]: https://blake3.io
//! [Rayon]: https://github.com/rayon-rs/rayon
//! [`join::RayonJoin`]: join/enum.RayonJoin.html
//! [`Hasher::update_with_join`]: struct.Hasher.html#method.update_with_join
//! [docs.rs]: https://docs.rs/
//! [`Write`]: https://doc.rust-lang.org/std/io/trait.Write.html
//! [`Seek`]: https://doc.rust-lang.org/std/io/trait.Seek.html

#![cfg_attr(not(feature = "std"), no_std)]

#[cfg(test)]
mod test;

// The guts module is for incremental use cases like the `bao` crate that need
// to explicitly compute chunk and parent chaining values. It is semi-stable
// and likely to keep working, but largely undocumented and not intended for
// widespread use.
#[doc(hidden)]
pub mod guts;

// The platform module is pub for benchmarks only. It is not stable.
#[doc(hidden)]
pub mod platform;

// Platform-specific implementations of the compression function. These
// BLAKE3-specific cfg flags are set in build.rs.
#[cfg(blake3_avx2_rust)]
#[path = "rust_avx2.rs"]
mod avx2;
#[cfg(blake3_avx2_ffi)]
#[path = "ffi_avx2.rs"]
mod avx2;
#[cfg(blake3_avx512_ffi)]
#[path = "ffi_avx512.rs"]
mod avx512;
#[cfg(feature = "neon")]
#[path = "ffi_neon.rs"]
mod neon;
mod portable;
#[cfg(blake3_sse2_rust)]
#[path = "rust_sse2.rs"]
mod sse2;
#[cfg(blake3_sse2_ffi)]
#[path = "ffi_sse2.rs"]
mod sse2;
#[cfg(blake3_sse41_rust)]
#[path = "rust_sse41.rs"]
mod sse41;
#[cfg(blake3_sse41_ffi)]
#[path = "ffi_sse41.rs"]
mod sse41;

pub mod traits;

pub mod join;

use arrayref::{array_mut_ref, array_ref};
use arrayvec::{ArrayString, ArrayVec};
use core::cmp;
use core::fmt;
use join::{Join, SerialJoin};
use platform::{Platform, MAX_SIMD_DEGREE, MAX_SIMD_DEGREE_OR_2};

/// The number of bytes in a [`Hash`](struct.Hash.html), 32.
pub const OUT_LEN: usize = 32;

/// The number of bytes in a key, 32.
pub const KEY_LEN: usize = 32;

// These constants are pub for incremental use cases like `bao`, as well as
// tests and benchmarks. Most callers should not need them.
#[doc(hidden)]
pub const BLOCK_LEN: usize = 64;
#[doc(hidden)]
pub const CHUNK_LEN: usize = 1024;
#[doc(hidden)]
pub const MAX_DEPTH: usize = 54; // 2^54 * CHUNK_LEN = 2^64

// While iterating the compression function within a chunk, the CV is
// represented as words, to avoid doing two extra endianness conversions for
// each compression in the portable implementation. But the hash_many interface
// needs to hash both input bytes and parent nodes, so its better for its
// output CVs to be represented as bytes.
type CVWords = [u32; 8];
type CVBytes = [u8; 32]; // little-endian

const IV: &CVWords = &[
    0x6A09E667, 0xBB67AE85, 0x3C6EF372, 0xA54FF53A, 0x510E527F, 0x9B05688C, 0x1F83D9AB, 0x5BE0CD19,
];

const MSG_SCHEDULE: [[usize; 16]; 7] = [
    [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15],
    [2, 6, 3, 10, 7, 0, 4, 13, 1, 11, 12, 5, 9, 14, 15, 8],
    [3, 4, 10, 12, 13, 2, 7, 14, 6, 5, 9, 0, 11, 15, 8, 1],
    [10, 7, 12, 9, 14, 3, 13, 15, 4, 0, 11, 2, 5, 8, 1, 6],
    [12, 13, 9, 11, 15, 10, 14, 8, 7, 2, 5, 3, 0, 1, 6, 4],
    [9, 14, 11, 5, 8, 12, 15, 1, 13, 3, 0, 10, 2, 6, 4, 7],
    [11, 15, 5, 0, 1, 9, 8, 6, 14, 10, 2, 12, 3, 4, 7, 13],
];

// These are the internal flags that we use to domain separate root/non-root,
// chunk/parent, and chunk beginning/middle/end. These get set at the high end
// of the block flags word in the compression function, so their values start
// high and go down.
const CHUNK_START: u8 = 1 << 0;
const CHUNK_END: u8 = 1 << 1;
const PARENT: u8 = 1 << 2;
const ROOT: u8 = 1 << 3;
const KEYED_HASH: u8 = 1 << 4;
const DERIVE_KEY_CONTEXT: u8 = 1 << 5;
const DERIVE_KEY_MATERIAL: u8 = 1 << 6;

#[inline]
fn counter_low(counter: u64) -> u32 {
    counter as u32
}

#[inline]
fn counter_high(counter: u64) -> u32 {
    (counter >> 32) as u32
}

/// An output of the default size, 32 bytes, which provides constant-time
/// equality checking.
///
/// `Hash` implements [`From`] and [`Into`] for `[u8; 32]`, and it provides an
/// explicit [`as_bytes`] method returning `&[u8; 32]`. However, byte arrays
/// and slices don't provide constant-time equality checking, which is often a
/// security requirement in software that handles private data. `Hash` doesn't
/// implement [`Deref`] or [`AsRef`], to avoid situations where a type
/// conversion happens implicitly and the constant-time property is
/// accidentally lost.
///
/// `Hash` provides the [`to_hex`] method for converting to hexadecimal. It
/// doesn't directly support converting from hexadecimal, but here's an example
/// of doing that with the [`hex`] crate:
///
/// ```
/// # fn main() -> Result<(), Box<dyn std::error::Error>> {
/// use std::convert::TryInto;
///
/// let hash_hex = "d74981efa70a0c880b8d8c1985d075dbcbf679b99a5f9914e5aaf96b831a9e24";
/// let hash_bytes = hex::decode(hash_hex)?;
/// let hash_array: [u8; blake3::OUT_LEN] = hash_bytes[..].try_into()?;
/// let hash: blake3::Hash = hash_array.into();
/// # Ok(())
/// # }
/// ```
///
/// [`From`]: https://doc.rust-lang.org/std/convert/trait.From.html
/// [`Into`]: https://doc.rust-lang.org/std/convert/trait.Into.html
/// [`as_bytes`]: #method.as_bytes
/// [`Deref`]: https://doc.rust-lang.org/stable/std/ops/trait.Deref.html
/// [`AsRef`]: https://doc.rust-lang.org/std/convert/trait.AsRef.html
/// [`to_hex`]: #method.to_hex
/// [`hex`]: https://crates.io/crates/hex
#[derive(Clone, Copy, Hash)]
pub struct Hash([u8; OUT_LEN]);

impl Hash {
    /// The bytes of the `Hash`. Note that byte arrays don't provide
    /// constant-time equality checking, so if  you need to compare hashes,
    /// prefer the `Hash` type.
    #[inline]
    pub fn as_bytes(&self) -> &[u8; OUT_LEN] {
        &self.0
    }

    /// The hexadecimal encoding of the `Hash`. The returned [`ArrayString`] is
    /// a fixed size and doesn't allocate memory on the heap. Note that
    /// [`ArrayString`] doesn't provide constant-time equality checking, so if
    /// you need to compare hashes, prefer the `Hash` type.
    ///
    /// [`ArrayString`]: https://docs.rs/arrayvec/0.5.1/arrayvec/struct.ArrayString.html
    pub fn to_hex(&self) -> ArrayString<[u8; 2 * OUT_LEN]> {
        let mut s = ArrayString::new();
        let table = b"0123456789abcdef";
        for &b in self.0.iter() {
            s.push(table[(b >> 4) as usize] as char);
            s.push(table[(b & 0xf) as usize] as char);
        }
        s
    }
}

impl From<[u8; OUT_LEN]> for Hash {
    #[inline]
    fn from(bytes: [u8; OUT_LEN]) -> Self {
        Self(bytes)
    }
}

impl From<Hash> for [u8; OUT_LEN] {
    #[inline]
    fn from(hash: Hash) -> Self {
        hash.0
    }
}

/// This implementation is constant-time.
impl PartialEq for Hash {
    #[inline]
    fn eq(&self, other: &Hash) -> bool {
        constant_time_eq::constant_time_eq_32(&self.0, &other.0)
    }
}

/// This implementation is constant-time.
impl PartialEq<[u8; OUT_LEN]> for Hash {
    #[inline]
    fn eq(&self, other: &[u8; OUT_LEN]) -> bool {
        constant_time_eq::constant_time_eq_32(&self.0, other)
    }
}

impl Eq for Hash {}

impl fmt::Debug for Hash {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        // Formatting field as `&str` to reduce code size since the `Debug`
        // dynamic dispatch table for `&str` is likely needed elsewhere already,
        // but that for `ArrayString<[u8; 64]>` is not.
        let hex = self.to_hex();
        let hex: &str = hex.as_str();

        f.debug_tuple("Hash").field(&hex).finish()
    }
}

// Each chunk or parent node can produce either a 32-byte chaining value or, by
// setting the ROOT flag, any number of final output bytes. The Output struct
// captures the state just prior to choosing between those two possibilities.
#[derive(Clone)]
struct Output {
    input_chaining_value: CVWords,
    block: [u8; 64],
    block_len: u8,
    counter: u64,
    flags: u8,
    platform: Platform,
}

impl Output {
    fn chaining_value(&self) -> CVBytes {
        let mut cv = self.input_chaining_value;
        self.platform.compress_in_place(
            &mut cv,
            &self.block,
            self.block_len,
            self.counter,
            self.flags,
        );
        platform::le_bytes_from_words_32(&cv)
    }

    fn root_hash(&self) -> Hash {
        debug_assert_eq!(self.counter, 0);
        let mut cv = self.input_chaining_value;
        self.platform
            .compress_in_place(&mut cv, &self.block, self.block_len, 0, self.flags | ROOT);
        Hash(platform::le_bytes_from_words_32(&cv))
    }

    fn root_output_block(&self) -> [u8; 2 * OUT_LEN] {
        self.platform.compress_xof(
            &self.input_chaining_value,
            &self.block,
            self.block_len,
            self.counter,
            self.flags | ROOT,
        )
    }
}

#[derive(Clone)]
struct ChunkState {
    cv: CVWords,
    chunk_counter: u64,
    buf: [u8; BLOCK_LEN],
    buf_len: u8,
    blocks_compressed: u8,
    flags: u8,
    platform: Platform,
}

impl ChunkState {
    fn new(key: &CVWords, chunk_counter: u64, flags: u8, platform: Platform) -> Self {
        Self {
            cv: *key,
            chunk_counter,
            buf: [0; BLOCK_LEN],
            buf_len: 0,
            blocks_compressed: 0,
            flags,
            platform,
        }
    }

    fn len(&self) -> usize {
        BLOCK_LEN * self.blocks_compressed as usize + self.buf_len as usize
    }

    fn fill_buf(&mut self, input: &mut &[u8]) {
        let want = BLOCK_LEN - self.buf_len as usize;
        let take = cmp::min(want, input.len());
        self.buf[self.buf_len as usize..][..take].copy_from_slice(&input[..take]);
        self.buf_len += take as u8;
        *input = &input[take..];
    }

    fn start_flag(&self) -> u8 {
        if self.blocks_compressed == 0 {
            CHUNK_START
        } else {
            0
        }
    }

    // Try to avoid buffering as much as possible, by compressing directly from
    // the input slice when full blocks are available.
    fn update(&mut self, mut input: &[u8]) -> &mut Self {
        if self.buf_len > 0 {
            self.fill_buf(&mut input);
            if !input.is_empty() {
                debug_assert_eq!(self.buf_len as usize, BLOCK_LEN);
                let block_flags = self.flags | self.start_flag(); // borrowck
                self.platform.compress_in_place(
                    &mut self.cv,
                    &self.buf,
                    BLOCK_LEN as u8,
                    self.chunk_counter,
                    block_flags,
                );
                self.buf_len = 0;
                self.buf = [0; BLOCK_LEN];
                self.blocks_compressed += 1;
            }
        }

        while input.len() > BLOCK_LEN {
            debug_assert_eq!(self.buf_len, 0);
            let block_flags = self.flags | self.start_flag(); // borrowck
            self.platform.compress_in_place(
                &mut self.cv,
                array_ref!(input, 0, BLOCK_LEN),
                BLOCK_LEN as u8,
                self.chunk_counter,
                block_flags,
            );
            self.blocks_compressed += 1;
            input = &input[BLOCK_LEN..];
        }

        self.fill_buf(&mut input);
        debug_assert!(input.is_empty());
        debug_assert!(self.len() <= CHUNK_LEN);
        self
    }

    fn output(&self) -> Output {
        let block_flags = self.flags | self.start_flag() | CHUNK_END;
        Output {
            input_chaining_value: self.cv,
            block: self.buf,
            block_len: self.buf_len,
            counter: self.chunk_counter,
            flags: block_flags,
            platform: self.platform,
        }
    }
}

// Don't derive(Debug), because the state may be secret.
impl fmt::Debug for ChunkState {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        f.debug_struct("ChunkState")
            .field("len", &self.len())
            .field("chunk_counter", &self.chunk_counter)
            .field("flags", &self.flags)
            .field("platform", &self.platform)
            .finish()
    }
}

// IMPLEMENTATION NOTE
// ===================
// The recursive function compress_subtree_wide(), implemented below, is the
// basis of high-performance BLAKE3. We use it both for all-at-once hashing,
// and for the incremental input with Hasher (though we have to be careful with
// subtree boundaries in the incremental case). compress_subtree_wide() applies
// several optimizations at the same time:
// - Multi-threading with Rayon.
// - Parallel chunk hashing with SIMD.
// - Parallel parent hashing with SIMD. Note that while SIMD chunk hashing
//   maxes out at MAX_SIMD_DEGREE*CHUNK_LEN, parallel parent hashing continues
//   to benefit from larger inputs, because more levels of the tree benefit can
//   use full-width SIMD vectors for parent hashing. Without parallel parent
//   hashing, we lose about 10% of overall throughput on AVX2 and AVX-512.

// pub for benchmarks
#[doc(hidden)]
#[derive(Clone, Copy)]
pub enum IncrementCounter {
    Yes,
    No,
}

impl IncrementCounter {
    #[inline]
    fn yes(&self) -> bool {
        match self {
            IncrementCounter::Yes => true,
            IncrementCounter::No => false,
        }
    }
}

// The largest power of two less than or equal to `n`, used for left_len()
// immediately below, and also directly in Hasher::update().
fn largest_power_of_two_leq(n: usize) -> usize {
    ((n / 2) + 1).next_power_of_two()
}

// Given some input larger than one chunk, return the number of bytes that
// should go in the left subtree. This is the largest power-of-2 number of
// chunks that leaves at least 1 byte for the right subtree.
fn left_len(content_len: usize) -> usize {
    debug_assert!(content_len > CHUNK_LEN);
    // Subtract 1 to reserve at least one byte for the right side.
    let full_chunks = (content_len - 1) / CHUNK_LEN;
    largest_power_of_two_leq(full_chunks) * CHUNK_LEN
}

// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE chunks at the same time
// on a single thread. Write out the chunk chaining values and return the
// number of chunks hashed. These chunks are never the root and never empty;
// those cases use a different codepath.
fn compress_chunks_parallel(
    input: &[u8],
    key: &CVWords,
    chunk_counter: u64,
    flags: u8,
    platform: Platform,
    out: &mut [u8],
) -> usize {
    debug_assert!(!input.is_empty(), "empty chunks below the root");
    debug_assert!(input.len() <= MAX_SIMD_DEGREE * CHUNK_LEN);

    let mut chunks_exact = input.chunks_exact(CHUNK_LEN);
    let mut chunks_array = ArrayVec::<[&[u8; CHUNK_LEN]; MAX_SIMD_DEGREE]>::new();
    for chunk in &mut chunks_exact {
        chunks_array.push(array_ref!(chunk, 0, CHUNK_LEN));
    }
    platform.hash_many(
        &chunks_array,
        key,
        chunk_counter,
        IncrementCounter::Yes,
        flags,
        CHUNK_START,
        CHUNK_END,
        out,
    );

    // Hash the remaining partial chunk, if there is one. Note that the empty
    // chunk (meaning the empty message) is a different codepath.
    let chunks_so_far = chunks_array.len();
    if !chunks_exact.remainder().is_empty() {
        let counter = chunk_counter + chunks_so_far as u64;
        let mut chunk_state = ChunkState::new(key, counter, flags, platform);
        chunk_state.update(chunks_exact.remainder());
        *array_mut_ref!(out, chunks_so_far * OUT_LEN, OUT_LEN) =
            chunk_state.output().chaining_value();
        chunks_so_far + 1
    } else {
        chunks_so_far
    }
}

// Use SIMD parallelism to hash up to MAX_SIMD_DEGREE parents at the same time
// on a single thread. Write out the parent chaining values and return the
// number of parents hashed. (If there's an odd input chaining value left over,
// return it as an additional output.) These parents are never the root and
// never empty; those cases use a different codepath.
fn compress_parents_parallel(
    child_chaining_values: &[u8],
    key: &CVWords,
    flags: u8,
    platform: Platform,
    out: &mut [u8],
) -> usize {
    debug_assert_eq!(child_chaining_values.len() % OUT_LEN, 0, "wacky hash bytes");
    let num_children = child_chaining_values.len() / OUT_LEN;
    debug_assert!(num_children >= 2, "not enough children");
    debug_assert!(num_children <= 2 * MAX_SIMD_DEGREE_OR_2, "too many");

    let mut parents_exact = child_chaining_values.chunks_exact(BLOCK_LEN);
    // Use MAX_SIMD_DEGREE_OR_2 rather than MAX_SIMD_DEGREE here, because of
    // the requirements of compress_subtree_wide().
    let mut parents_array = ArrayVec::<[&[u8; BLOCK_LEN]; MAX_SIMD_DEGREE_OR_2]>::new();
    for parent in &mut parents_exact {
        parents_array.push(array_ref!(parent, 0, BLOCK_LEN));
    }
    platform.hash_many(
        &parents_array,
        key,
        0, // Parents always use counter 0.
        IncrementCounter::No,
        flags | PARENT,
        0, // Parents have no start flags.
        0, // Parents have no end flags.
        out,
    );

    // If there's an odd child left over, it becomes an output.
    let parents_so_far = parents_array.len();
    if !parents_exact.remainder().is_empty() {
        out[parents_so_far * OUT_LEN..][..OUT_LEN].copy_from_slice(parents_exact.remainder());
        parents_so_far + 1
    } else {
        parents_so_far
    }
}

// The wide helper function returns (writes out) an array of chaining values
// and returns the length of that array. The number of chaining values returned
// is the dyanmically detected SIMD degree, at most MAX_SIMD_DEGREE. Or fewer,
// if the input is shorter than that many chunks. The reason for maintaining a
// wide array of chaining values going back up the tree, is to allow the
// implementation to hash as many parents in parallel as possible.
//
// As a special case when the SIMD degree is 1, this function will still return
// at least 2 outputs. This guarantees that this function doesn't perform the
// root compression. (If it did, it would use the wrong flags, and also we
// wouldn't be able to implement exendable ouput.) Note that this function is
// not used when the whole input is only 1 chunk long; that's a different
// codepath.
//
// Why not just have the caller split the input on the first update(), instead
// of implementing this special rule? Because we don't want to limit SIMD or
// multi-threading parallelism for that update().
fn compress_subtree_wide<J: Join>(
    input: &[u8],
    key: &CVWords,
    chunk_counter: u64,
    flags: u8,
    platform: Platform,
    out: &mut [u8],
) -> usize {
    // Note that the single chunk case does *not* bump the SIMD degree up to 2
    // when it is 1. This allows Rayon the option of multi-threading even the
    // 2-chunk case, which can help performance on smaller platforms.
    if input.len() <= platform.simd_degree() * CHUNK_LEN {
        return compress_chunks_parallel(input, key, chunk_counter, flags, platform, out);
    }

    // With more than simd_degree chunks, we need to recurse. Start by dividing
    // the input into left and right subtrees. (Note that this is only optimal
    // as long as the SIMD degree is a power of 2. If we ever get a SIMD degree
    // of 3 or something, we'll need a more complicated strategy.)
    debug_assert_eq!(platform.simd_degree().count_ones(), 1, "power of 2");
    let (left, right) = input.split_at(left_len(input.len()));
    let right_chunk_counter = chunk_counter + (left.len() / CHUNK_LEN) as u64;

    // Make space for the child outputs. Here we use MAX_SIMD_DEGREE_OR_2 to
    // account for the special case of returning 2 outputs when the SIMD degree
    // is 1.
    let mut cv_array = [0; 2 * MAX_SIMD_DEGREE_OR_2 * OUT_LEN];
    let degree = if left.len() == CHUNK_LEN {
        // The "simd_degree=1 and we're at the leaf nodes" case.
        debug_assert_eq!(platform.simd_degree(), 1);
        1
    } else {
        cmp::max(platform.simd_degree(), 2)
    };
    let (left_out, right_out) = cv_array.split_at_mut(degree * OUT_LEN);

    // Recurse! This uses multiple threads if the "rayon" feature is enabled.
    let (left_n, right_n) = J::join(
        || compress_subtree_wide::<J>(left, key, chunk_counter, flags, platform, left_out),
        || compress_subtree_wide::<J>(right, key, right_chunk_counter, flags, platform, right_out),
        left.len(),
        right.len(),
    );

    // The special case again. If simd_degree=1, then we'll have left_n=1 and
    // right_n=1. Rather than compressing them into a single output, return
    // them directly, to make sure we always have at least two outputs.
    debug_assert_eq!(left_n, degree);
    debug_assert!(right_n >= 1 && right_n <= left_n);
    if left_n == 1 {
        out[..2 * OUT_LEN].copy_from_slice(&cv_array[..2 * OUT_LEN]);
        return 2;
    }

    // Otherwise, do one layer of parent node compression.
    let num_children = left_n + right_n;
    compress_parents_parallel(
        &cv_array[..num_children * OUT_LEN],
        key,
        flags,
        platform,
        out,
    )
}

// Hash a subtree with compress_subtree_wide(), and then condense the resulting
// list of chaining values down to a single parent node. Don't compress that
// last parent node, however. Instead, return its message bytes (the
// concatenated chaining values of its children). This is necessary when the
// first call to update() supplies a complete subtree, because the topmost
// parent node of that subtree could end up being the root. It's also necessary
// for extended output in the general case.
//
// As with compress_subtree_wide(), this function is not used on inputs of 1
// chunk or less. That's a different codepath.
fn compress_subtree_to_parent_node<J: Join>(
    input: &[u8],
    key: &CVWords,
    chunk_counter: u64,
    flags: u8,
    platform: Platform,
) -> [u8; BLOCK_LEN] {
    debug_assert!(input.len() > CHUNK_LEN);
    let mut cv_array = [0; MAX_SIMD_DEGREE_OR_2 * OUT_LEN];
    let mut num_cvs =
        compress_subtree_wide::<J>(input, &key, chunk_counter, flags, platform, &mut cv_array);
    debug_assert!(num_cvs >= 2);

    // If MAX_SIMD_DEGREE is greater than 2 and there's enough input,
    // compress_subtree_wide() returns more than 2 chaining values. Condense
    // them into 2 by forming parent nodes repeatedly.
    let mut out_array = [0; MAX_SIMD_DEGREE_OR_2 * OUT_LEN / 2];
    while num_cvs > 2 {
        let cv_slice = &cv_array[..num_cvs * OUT_LEN];
        num_cvs = compress_parents_parallel(cv_slice, key, flags, platform, &mut out_array);
        cv_array[..num_cvs * OUT_LEN].copy_from_slice(&out_array[..num_cvs * OUT_LEN]);
    }
    *array_ref!(cv_array, 0, 2 * OUT_LEN)
}

// Hash a complete input all at once. Unlike compress_subtree_wide() and
// compress_subtree_to_parent_node(), this function handles the 1 chunk case.
// Note that this we use SerialJoin here, so this is always single-threaded.
fn hash_all_at_once(input: &[u8], key: &CVWords, flags: u8) -> Output {
    let platform = Platform::detect();

    // If the whole subtree is one chunk, hash it directly with a ChunkState.
    if input.len() <= CHUNK_LEN {
        return ChunkState::new(key, 0, flags, platform)
            .update(input)
            .output();
    }

    // Otherwise construct an Output object from the parent node returned by
    // compress_subtree_to_parent_node().
    Output {
        input_chaining_value: *key,
        block: compress_subtree_to_parent_node::<SerialJoin>(input, key, 0, flags, platform),
        block_len: BLOCK_LEN as u8,
        counter: 0,
        flags: flags | PARENT,
        platform,
    }
}

/// The default hash function.
///
/// For an incremental version that accepts multiple writes, see [`Hasher::update`].
///
/// This function is always single-threaded. For multi-threading support, see
/// [`Hasher::update_with_join`].
///
/// [`Hasher::update`]: struct.Hasher.html#method.update
/// [`Hasher::update_with_join`]: struct.Hasher.html#method.update_with_join
pub fn hash(input: &[u8]) -> Hash {
    hash_all_at_once(input, IV, 0).root_hash()
}

/// The keyed hash function.
///
/// This is suitable for use as a message authentication code, for
/// example to replace an HMAC instance.
///  In that use case, the constant-time equality checking provided by
/// [`Hash`](struct.Hash.html) is almost always a security requirement, and
/// callers need to be careful not to compare MACs as raw bytes.
///
/// This function is always single-threaded. For multi-threading support, see
/// [`Hasher::update_with_join`].
///
/// [`Hasher::update_with_join`]: struct.Hasher.html#method.update_with_join
pub fn keyed_hash(key: &[u8; KEY_LEN], input: &[u8]) -> Hash {
    let key_words = platform::words_from_le_bytes_32(key);
    hash_all_at_once(input, &key_words, KEYED_HASH).root_hash()
}

/// The key derivation function.
///
/// Given cryptographic key material of any length and a context string of any
/// length, this function outputs a derived subkey of any length. **The context
/// string should be hardcoded, globally unique, and application-specific.** A
/// good default format for such strings is `"[application] [commit timestamp]
/// [purpose]"`, e.g., `"example.com 2019-12-25 16:18:03 session tokens v1"`.
///
/// Key derivation is important when you want to use the same key in multiple
/// algorithms or use cases. Using the same key with different cryptographic
/// algorithms is generally forbidden, and deriving a separate subkey for each
/// use case protects you from bad interactions. Derived keys also mitigate the
/// damage from one part of your application accidentally leaking its key.
///
/// As a rare exception to that general rule, however, it is possible to use
/// `derive_key` itself with key material that you are already using with
/// another algorithm. You might need to do this if you're adding features to
/// an existing application, which does not yet use key derivation internally.
/// However, you still must not share key material with algorithms that forbid
/// key reuse entirely, like a one-time pad.
///
/// Note that BLAKE3 is not a password hash, and **`derive_key` should never be
/// used with passwords.** Instead, use a dedicated password hash like
/// [Argon2]. Password hashes are entirely different from generic hash
/// functions, with opposite design requirements.
///
/// This function is always single-threaded. For multi-threading support, see
/// [`Hasher::update_with_join`].
///
/// [`Hasher::new_derive_key`]: struct.Hasher.html#method.new_derive_key
/// [`Hasher::finalize_xof`]: struct.Hasher.html#method.finalize_xof
/// [Argon2]: https://en.wikipedia.org/wiki/Argon2
/// [`Hasher::update_with_join`]: struct.Hasher.html#method.update_with_join
pub fn derive_key(context: &str, key_material: &[u8], output: &mut [u8]) {
    let context_key = hash_all_at_once(context.as_bytes(), IV, DERIVE_KEY_CONTEXT).root_hash();
    let context_key_words = platform::words_from_le_bytes_32(context_key.as_bytes());
    let inner_output = hash_all_at_once(key_material, &context_key_words, DERIVE_KEY_MATERIAL);
    OutputReader::new(inner_output).fill(output);
}

fn parent_node_output(
    left_child: &CVBytes,
    right_child: &CVBytes,
    key: &CVWords,
    flags: u8,
    platform: Platform,
) -> Output {
    let mut block = [0; BLOCK_LEN];
    block[..32].copy_from_slice(left_child);
    block[32..].copy_from_slice(right_child);
    Output {
        input_chaining_value: *key,
        block,
        block_len: BLOCK_LEN as u8,
        counter: 0,
        flags: flags | PARENT,
        platform,
    }
}

/// An incremental hash state that can accept any number of writes.
///
/// In addition to its inherent methods, this type implements several commonly
/// used traits from the [`digest`](https://crates.io/crates/digest) and
/// [`crypto_mac`](https://crates.io/crates/crypto-mac) crates.
///
/// **Performance note:** The [`update`] and [`update_with_join`] methods
/// perform poorly when the caller's input buffer is small. See their method
/// docs below. A 16 KiB buffer is large enough to leverage all currently
/// supported SIMD instruction sets.
///
/// # Examples
///
/// ```
/// # fn main() -> Result<(), Box<dyn std::error::Error>> {
/// // Hash an input incrementally.
/// let mut hasher = blake3::Hasher::new();
/// hasher.update(b"foo");
/// hasher.update(b"bar");
/// hasher.update(b"baz");
/// assert_eq!(hasher.finalize(), blake3::hash(b"foobarbaz"));
///
/// // Extended output. OutputReader also implements Read and Seek.
/// # #[cfg(feature = "std")] {
/// let mut output = [0; 1000];
/// let mut output_reader = hasher.finalize_xof();
/// output_reader.fill(&mut output);
/// assert_eq!(&output[..32], blake3::hash(b"foobarbaz").as_bytes());
/// # }
/// # Ok(())
/// # }
/// ```
///
/// [`update`]: #method.update
/// [`update_with_join`]: #method.update_with_join
#[derive(Clone)]
pub struct Hasher {
    key: CVWords,
    chunk_state: ChunkState,
    // The stack size is MAX_DEPTH + 1 because we do lazy merging. For example,
    // with 7 chunks, we have 3 entries in the stack. Adding an 8th chunk
    // requires a 4th entry, rather than merging everything down to 1, because
    // we don't know whether more input is coming. This is different from how
    // the reference implementation does things.
    cv_stack: ArrayVec<[CVBytes; MAX_DEPTH + 1]>,
}

impl Hasher {
    fn new_internal(key: &CVWords, flags: u8) -> Self {
        Self {
            key: *key,
            chunk_state: ChunkState::new(key, 0, flags, Platform::detect()),
            cv_stack: ArrayVec::new(),
        }
    }

    /// Construct a new `Hasher` for the regular hash function.
    pub fn new() -> Self {
        Self::new_internal(IV, 0)
    }

    /// Construct a new `Hasher` for the keyed hash function. See
    /// [`keyed_hash`].
    ///
    /// [`keyed_hash`]: fn.keyed_hash.html
    pub fn new_keyed(key: &[u8; KEY_LEN]) -> Self {
        let key_words = platform::words_from_le_bytes_32(key);
        Self::new_internal(&key_words, KEYED_HASH)
    }

    /// Construct a new `Hasher` for the key derivation function. See
    /// [`derive_key`]. The context string should be hardcoded, globally
    /// unique, and application-specific.
    ///
    /// [`derive_key`]: fn.derive_key.html
    pub fn new_derive_key(context: &str) -> Self {
        let context_key = hash_all_at_once(context.as_bytes(), IV, DERIVE_KEY_CONTEXT).root_hash();
        let context_key_words = platform::words_from_le_bytes_32(context_key.as_bytes());
        Self::new_internal(&context_key_words, DERIVE_KEY_MATERIAL)
    }

    /// Reset the `Hasher` to its initial state.
    ///
    /// This is functionally the same as overwriting the `Hasher` with a new
    /// one, using the same key or context string if any. However, depending on
    /// how much inlining the optimizer does, moving a `Hasher` might copy its
    /// entire CV stack, most of which is useless uninitialized bytes. This
    /// methods avoids that copy.
    pub fn reset(&mut self) -> &mut Self {
        self.chunk_state = ChunkState::new(
            &self.key,
            0,
            self.chunk_state.flags,
            self.chunk_state.platform,
        );
        self.cv_stack.clear();
        self
    }

    // As described in push_cv() below, we do "lazy merging", delaying merges
    // until right before the next CV is about to be added. This is different
    // from the reference implementation. Another difference is that we aren't
    // always merging 1 chunk at a time. Instead, each CV might represent any
    // power-of-two number of chunks, as long as the smaller-above-larger stack
    // order is maintained. Instead of the "count the trailing 0-bits"
    // algorithm described in the spec, we use a "count the total number of
    // 1-bits" variant that doesn't require us to retain the subtree size of
    // the CV on top of the stack. The principle is the same: each CV that
    // should remain in the stack is represented by a 1-bit in the total number
    // of chunks (or bytes) so far.
    fn merge_cv_stack(&mut self, total_len: u64) {
        let post_merge_stack_len = total_len.count_ones() as usize;
        while self.cv_stack.len() > post_merge_stack_len {
            let right_child = self.cv_stack.pop().unwrap();
            let left_child = self.cv_stack.pop().unwrap();
            let parent_output = parent_node_output(
                &left_child,
                &right_child,
                &self.key,
                self.chunk_state.flags,
                self.chunk_state.platform,
            );
            self.cv_stack.push(parent_output.chaining_value());
        }
    }

    // In reference_impl.rs, we merge the new CV with existing CVs from the
    // stack before pushing it. We can do that because we know more input is
    // coming, so we know none of the merges are root.
    //
    // This setting is different. We want to feed as much input as possible to
    // compress_subtree_wide(), without setting aside anything for the
    // chunk_state. If the user gives us 64 KiB, we want to parallelize over
    // all 64 KiB at once as a single subtree, if at all possible.
    //
    // This leads to two problems:
    // 1) This 64 KiB input might be the only call that ever gets made to
    //    update. In this case, the root node of the 64 KiB subtree would be
    //    the root node of the whole tree, and it would need to be ROOT
    //    finalized. We can't compress it until we know.
    // 2) This 64 KiB input might complete a larger tree, whose root node is
    //    similarly going to be the the root of the whole tree. For example,
    //    maybe we have 196 KiB (that is, 128 + 64) hashed so far. We can't
    //    compress the node at the root of the 256 KiB subtree until we know
    //    how to finalize it.
    //
    // The second problem is solved with "lazy merging". That is, when we're
    // about to add a CV to the stack, we don't merge it with anything first,
    // as the reference impl does. Instead we do merges using the *previous* CV
    // that was added, which is sitting on top of the stack, and we put the new
    // CV (unmerged) on top of the stack afterwards. This guarantees that we
    // never merge the root node until finalize().
    //
    // Solving the first problem requires an additional tool,
    // compress_subtree_to_parent_node(). That function always returns the top
    // *two* chaining values of the subtree it's compressing. We then do lazy
    // merging with each of them separately, so that the second CV will always
    // remain unmerged. (That also helps us support extendable output when
    // we're hashing an input all-at-once.)
    fn push_cv(&mut self, new_cv: &CVBytes, chunk_counter: u64) {
        self.merge_cv_stack(chunk_counter);
        self.cv_stack.push(*new_cv);
    }

    /// Add input bytes to the hash state. You can call this any number of
    /// times.
    ///
    /// This method is always single-threaded. For multi-threading support, see
    /// `update_with_join` below.
    ///
    /// Note that the degree of SIMD parallelism that `update` can use is
    /// limited by the size of this input buffer. The 8 KiB buffer currently
    /// used by [`std::io::copy`] is enough to leverage AVX2, for example, but
    /// not enough to leverage AVX-512. A 16 KiB buffer is large enough to
    /// leverage all currently supported SIMD instruction sets.
    ///
    /// [`std::io::copy`]: https://doc.rust-lang.org/std/io/fn.copy.html
    pub fn update(&mut self, input: &[u8]) -> &mut Self {
        self.update_with_join::<SerialJoin>(input)
    }

    /// Add input bytes to the hash state, as with `update`, but potentially
    /// using multi-threading. See the example below, and the
    /// [`join`](join/index.html) module for a more detailed explanation.
    ///
    /// To get any performance benefit from multi-threading, the input buffer
    /// size needs to be very large. As a rule of thumb on x86_64, there is no
    /// benefit to multi-threading inputs less than 128 KiB. Other platforms
    /// have different thresholds, and in general you need to benchmark your
    /// specific use case. Where possible, memory mapping an entire input file
    /// is recommended, to take maximum advantage of multi-threading without
    /// needing to tune a specific buffer size. Where memory mapping is not
    /// possible, good multi-threading performance requires doing IO on a
    /// background thread, to avoid sleeping all your worker threads while the
    /// input buffer is (serially) refilled. This is quite complicated compared
    /// to memory mapping.
    ///
    /// # Example
    ///
    /// ```
    /// // Hash a large input using multi-threading. Note that multi-threading
    /// // comes with some overhead, and it can actually hurt performance for small
    /// // inputs. The meaning of "small" varies, however, depending on the
    /// // platform and the number of threads. (On x86_64, the cutoff tends to be
    /// // around 128 KiB.) You should benchmark your own use case to see whether
    /// // multi-threading helps.
    /// # #[cfg(feature = "rayon")]
    /// # {
    /// # fn some_large_input() -> &'static [u8] { b"foo" }
    /// let input: &[u8] = some_large_input();
    /// let mut hasher = blake3::Hasher::new();
    /// hasher.update_with_join::<blake3::join::RayonJoin>(input);
    /// let hash = hasher.finalize();
    /// # }
    /// ```
    pub fn update_with_join<J: Join>(&mut self, mut input: &[u8]) -> &mut Self {
        // If we have some partial chunk bytes in the internal chunk_state, we
        // need to finish that chunk first.
        if self.chunk_state.len() > 0 {
            let want = CHUNK_LEN - self.chunk_state.len();
            let take = cmp::min(want, input.len());
            self.chunk_state.update(&input[..take]);
            input = &input[take..];
            if !input.is_empty() {
                // We've filled the current chunk, and there's more input
                // coming, so we know it's not the root and we can finalize it.
                // Then we'll proceed to hashing whole chunks below.
                debug_assert_eq!(self.chunk_state.len(), CHUNK_LEN);
                let chunk_cv = self.chunk_state.output().chaining_value();
                self.push_cv(&chunk_cv, self.chunk_state.chunk_counter);
                self.chunk_state = ChunkState::new(
                    &self.key,
                    self.chunk_state.chunk_counter + 1,
                    self.chunk_state.flags,
                    self.chunk_state.platform,
                );
            } else {
                return self;
            }
        }

        // Now the chunk_state is clear, and we have more input. If there's
        // more than a single chunk (so, definitely not the root chunk), hash
        // the largest whole subtree we can, with the full benefits of SIMD and
        // multi-threading parallelism. Two restrictions:
        // - The subtree has to be a power-of-2 number of chunks. Only subtrees
        //   along the right edge can be incomplete, and we don't know where
        //   the right edge is going to be until we get to finalize().
        // - The subtree must evenly divide the total number of chunks up until
        //   this point (if total is not 0). If the current incomplete subtree
        //   is only waiting for 1 more chunk, we can't hash a subtree of 4
        //   chunks. We have to complete the current subtree first.
        // Because we might need to break up the input to form powers of 2, or
        // to evenly divide what we already have, this part runs in a loop.
        while input.len() > CHUNK_LEN {
            debug_assert_eq!(self.chunk_state.len(), 0, "no partial chunk data");
            debug_assert_eq!(CHUNK_LEN.count_ones(), 1, "power of 2 chunk len");
            let mut subtree_len = largest_power_of_two_leq(input.len());
            let count_so_far = self.chunk_state.chunk_counter * CHUNK_LEN as u64;
            // Shrink the subtree_len until it evenly divides the count so far.
            // We know that subtree_len itself is a power of 2, so we can use a
            // bitmasking trick instead of an actual remainder operation. (Note
            // that if the caller consistently passes power-of-2 inputs of the
            // same size, as is hopefully typical, this loop condition will
            // always fail, and subtree_len will always be the full length of
            // the input.)
            //
            // An aside: We don't have to shrink subtree_len quite this much.
            // For example, if count_so_far is 1, we could pass 2 chunks to
            // compress_subtree_to_parent_node. Since we'll get 2 CVs back,
            // we'll still get the right answer in the end, and we might get to
            // use 2-way SIMD parallelism. The problem with this optimization,
            // is that it gets us stuck always hashing 2 chunks. The total
            // number of chunks will remain odd, and we'll never graduate to
            // higher degrees of parallelism. See
            // https://github.com/BLAKE3-team/BLAKE3/issues/69.
            while (subtree_len - 1) as u64 & count_so_far != 0 {
                subtree_len /= 2;
            }
            // The shrunken subtree_len might now be 1 chunk long. If so, hash
            // that one chunk by itself. Otherwise, compress the subtree into a
            // pair of CVs.
            let subtree_chunks = (subtree_len / CHUNK_LEN) as u64;
            if subtree_len <= CHUNK_LEN {
                debug_assert_eq!(subtree_len, CHUNK_LEN);
                self.push_cv(
                    &ChunkState::new(
                        &self.key,
                        self.chunk_state.chunk_counter,
                        self.chunk_state.flags,
                        self.chunk_state.platform,
                    )
                    .update(&input[..subtree_len])
                    .output()
                    .chaining_value(),
                    self.chunk_state.chunk_counter,
                );
            } else {
                // This is the high-performance happy path, though getting here
                // depends on the caller giving us a long enough input.
                let cv_pair = compress_subtree_to_parent_node::<J>(
                    &input[..subtree_len],
                    &self.key,
                    self.chunk_state.chunk_counter,
                    self.chunk_state.flags,
                    self.chunk_state.platform,
                );
                let left_cv = array_ref!(cv_pair, 0, 32);
                let right_cv = array_ref!(cv_pair, 32, 32);
                // Push the two CVs we received into the CV stack in order. Because
                // the stack merges lazily, this guarantees we aren't merging the
                // root.
                self.push_cv(left_cv, self.chunk_state.chunk_counter);
                self.push_cv(
                    right_cv,
                    self.chunk_state.chunk_counter + (subtree_chunks / 2),
                );
            }
            self.chunk_state.chunk_counter += subtree_chunks;
            input = &input[subtree_len..];
        }

        // What remains is 1 chunk or less. Add it to the chunk state.
        debug_assert!(input.len() <= CHUNK_LEN);
        if !input.is_empty() {
            self.chunk_state.update(input);
            // Having added some input to the chunk_state, we know what's in
            // the CV stack won't become the root node, and we can do an extra
            // merge. This simplifies finalize().
            self.merge_cv_stack(self.chunk_state.chunk_counter);
        }

        self
    }

    fn final_output(&self) -> Output {
        // If the current chunk is the only chunk, that makes it the root node
        // also. Convert it directly into an Output. Otherwise, we need to
        // merge subtrees below.
        if self.cv_stack.is_empty() {
            debug_assert_eq!(self.chunk_state.chunk_counter, 0);
            return self.chunk_state.output();
        }

        // If there are any bytes in the ChunkState, finalize that chunk and
        // merge its CV with everything in the CV stack. In that case, the work
        // we did at the end of update() above guarantees that the stack
        // doesn't contain any unmerged subtrees that need to be merged first.
        // (This is important, because if there were two chunk hashes sitting
        // on top of the stack, they would need to merge with each other, and
        // merging a new chunk hash into them would be incorrect.)
        //
        // If there are no bytes in the ChunkState, we'll merge what's already
        // in the stack. In this case it's fine if there are unmerged chunks on
        // top, because we'll merge them with each other. Note that the case of
        // the empty chunk is taken care of above.
        let mut output: Output;
        let mut num_cvs_remaining = self.cv_stack.len();
        if self.chunk_state.len() > 0 {
            debug_assert_eq!(
                self.cv_stack.len(),
                self.chunk_state.chunk_counter.count_ones() as usize,
                "cv stack does not need a merge"
            );
            output = self.chunk_state.output();
        } else {
            debug_assert!(self.cv_stack.len() >= 2);
            output = parent_node_output(
                &self.cv_stack[num_cvs_remaining - 2],
                &self.cv_stack[num_cvs_remaining - 1],
                &self.key,
                self.chunk_state.flags,
                self.chunk_state.platform,
            );
            num_cvs_remaining -= 2;
        }
        while num_cvs_remaining > 0 {
            output = parent_node_output(
                &self.cv_stack[num_cvs_remaining - 1],
                &output.chaining_value(),
                &self.key,
                self.chunk_state.flags,
                self.chunk_state.platform,
            );
            num_cvs_remaining -= 1;
        }
        output
    }

    /// Finalize the hash state and return the [`Hash`](struct.Hash.html) of
    /// the input.
    ///
    /// This method is idempotent. Calling it twice will give the same result.
    /// You can also add more input and finalize again.
    pub fn finalize(&self) -> Hash {
        self.final_output().root_hash()
    }

    /// Finalize the hash state and return an [`OutputReader`], which can
    /// supply any number of output bytes.
    ///
    /// This method is idempotent. Calling it twice will give the same result.
    /// You can also add more input and finalize again.
    ///
    /// [`OutputReader`]: struct.OutputReader.html
    pub fn finalize_xof(&self) -> OutputReader {
        OutputReader::new(self.final_output())
    }
}

// Don't derive(Debug), because the state may be secret.
impl fmt::Debug for Hasher {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        f.debug_struct("Hasher")
            .field("flags", &self.chunk_state.flags)
            .field("platform", &self.chunk_state.platform)
            .finish()
    }
}

impl Default for Hasher {
    #[inline]
    fn default() -> Self {
        Self::new()
    }
}

#[cfg(feature = "std")]
impl std::io::Write for Hasher {
    /// This is equivalent to [`update`](#method.update).
    #[inline]
    fn write(&mut self, input: &[u8]) -> std::io::Result<usize> {
        self.update(input);
        Ok(input.len())
    }

    #[inline]
    fn flush(&mut self) -> std::io::Result<()> {
        Ok(())
    }
}

/// An incremental reader for extended output, returned by
/// [`Hasher::finalize_xof`](struct.Hasher.html#method.finalize_xof).
#[derive(Clone)]
pub struct OutputReader {
    inner: Output,
    position_within_block: u8,
}

impl OutputReader {
    fn new(inner: Output) -> Self {
        Self {
            inner,
            position_within_block: 0,
        }
    }

    /// Fill a buffer with output bytes and advance the position of the
    /// `OutputReader`. This is equivalent to [`Read::read`], except that it
    /// doesn't return a `Result`. Both methods always fill the entire buffer.
    ///
    /// Note that `OutputReader` doesn't buffer output bytes internally, so
    /// calling `fill` repeatedly with a short-length or odd-length slice will
    /// end up performing the same compression multiple times. If you're
    /// reading output in a loop, prefer a slice length that's a multiple of
    /// 64.
    ///
    /// The maximum output size of BLAKE3 is 2<sup>64</sup>-1 bytes. If you try
    /// to extract more than that, for example by seeking near the end and
    /// reading further, the behavior is unspecified.
    ///
    /// [`Read::read`]: #method.read
    pub fn fill(&mut self, mut buf: &mut [u8]) {
        while !buf.is_empty() {
            let block: [u8; BLOCK_LEN] = self.inner.root_output_block();
            let output_bytes = &block[self.position_within_block as usize..];
            let take = cmp::min(buf.len(), output_bytes.len());
            buf[..take].copy_from_slice(&output_bytes[..take]);
            buf = &mut buf[take..];
            self.position_within_block += take as u8;
            if self.position_within_block == BLOCK_LEN as u8 {
                self.inner.counter += 1;
                self.position_within_block = 0;
            }
        }
    }

    /// Return the current read position in the output stream. The position of
    /// a new `OutputReader` starts at 0, and each call to [`fill`] or
    /// [`Read::read`] moves the position forward by the number of bytes read.
    ///
    /// [`fill`]: #method.fill
    /// [`Read::read`]: #method.read
    pub fn position(&self) -> u64 {
        self.inner.counter * BLOCK_LEN as u64 + self.position_within_block as u64
    }

    /// Seek to a new read position in the output stream. This is equivalent to
    /// calling [`Seek::seek`] with [`SeekFrom::Start`], except that it doesn't
    /// return a `Result`.
    ///
    /// [`Seek::seek`]: #method.seek
    /// [`SeekFrom::Start`]: https://doc.rust-lang.org/std/io/enum.SeekFrom.html
    pub fn set_position(&mut self, position: u64) {
        self.position_within_block = (position % BLOCK_LEN as u64) as u8;
        self.inner.counter = position / BLOCK_LEN as u64;
    }
}

// Don't derive(Debug), because the state may be secret.
impl fmt::Debug for OutputReader {
    fn fmt(&self, f: &mut fmt::Formatter) -> fmt::Result {
        f.debug_struct("OutputReader")
            .field("position", &self.position())
            .finish()
    }
}

#[cfg(feature = "std")]
impl std::io::Read for OutputReader {
    #[inline]
    fn read(&mut self, buf: &mut [u8]) -> std::io::Result<usize> {
        self.fill(buf);
        Ok(buf.len())
    }
}

#[cfg(feature = "std")]
impl std::io::Seek for OutputReader {
    fn seek(&mut self, pos: std::io::SeekFrom) -> std::io::Result<u64> {
        let max_position = u64::max_value() as i128;
        let target_position: i128 = match pos {
            std::io::SeekFrom::Start(x) => x as i128,
            std::io::SeekFrom::Current(x) => self.position() as i128 + x as i128,
            std::io::SeekFrom::End(_) => {
                return Err(std::io::Error::new(
                    std::io::ErrorKind::InvalidInput,
                    "seek from end not supported",
                ));
            }
        };
        if target_position < 0 {
            return Err(std::io::Error::new(
                std::io::ErrorKind::InvalidInput,
                "seek before start",
            ));
        }
        self.set_position(cmp::min(target_position, max_position) as u64);
        Ok(self.position())
    }
}