summaryrefslogtreecommitdiff
path: root/absl/random/internal/randen_hwaes.cc
blob: 6b82d1d07ad8ef57689ce3ec9e126b86d30dbe5d (plain)
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
// Copyright 2017 The Abseil Authors.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
//      https://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.

// HERMETIC NOTE: The randen_hwaes target must not introduce duplicate
// symbols from arbitrary system and other headers, since it may be built
// with different flags from other targets, using different levels of
// optimization, potentially introducing ODR violations.

#include "absl/random/internal/randen_hwaes.h"

#include <cstdint>
#include <cstring>

#include "absl/random/internal/platform.h"

// ABSL_HAVE_ATTRIBUTE
#if !defined(ABSL_HAVE_ATTRIBUTE)
#ifdef __has_attribute
#define ABSL_HAVE_ATTRIBUTE(x) __has_attribute(x)
#else
#define ABSL_HAVE_ATTRIBUTE(x) 0
#endif
#endif

#if ABSL_HAVE_ATTRIBUTE(always_inline) || \
    (defined(__GNUC__) && !defined(__clang__))
#define ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE \
  __attribute__((always_inline))
#elif defined(_MSC_VER)
// We can achieve something similar to attribute((always_inline)) with MSVC by
// using the __forceinline keyword, however this is not perfect. MSVC is
// much less aggressive about inlining, and even with the __forceinline keyword.
#define ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE __forceinline
#else
#define ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE
#endif

// ABSL_ATTRIBUTE_FLATTEN enables much more aggressive inlining within
// the indicated function.
#undef ABSL_ATTRIBUTE_FLATTEN
#if ABSL_HAVE_ATTRIBUTE(flatten) || (defined(__GNUC__) && !defined(__clang__))
#define ABSL_ATTRIBUTE_FLATTEN __attribute__((flatten))
#else
#define ABSL_ATTRIBUTE_FLATTEN
#endif

// ABSL_RANDEN_HWAES_IMPL indicates whether this file will contain
// a hardware accelerated implementation of randen, or whether it
// will contain stubs that exit the process.
#if defined(ABSL_ARCH_X86_64) || defined(ABSL_ARCH_X86_32)
// The platform.h directives are sufficient to indicate whether
// we should build accelerated implementations for x86.
#if (ABSL_HAVE_ACCELERATED_AES || ABSL_RANDOM_INTERNAL_AES_DISPATCH)
#define ABSL_RANDEN_HWAES_IMPL 1
#endif
#elif defined(ABSL_ARCH_PPC)
// The platform.h directives are sufficient to indicate whether
// we should build accelerated implementations for PPC.
//
// NOTE: This has mostly been tested on 64-bit Power variants,
// and not embedded cpus such as powerpc32-8540
#if ABSL_HAVE_ACCELERATED_AES
#define ABSL_RANDEN_HWAES_IMPL 1
#endif
#elif defined(ABSL_ARCH_ARM) || defined(ABSL_ARCH_AARCH64)
// ARM is somewhat more complicated. We might support crypto natively...
#if ABSL_HAVE_ACCELERATED_AES || \
    (defined(__ARM_NEON) && defined(__ARM_FEATURE_CRYPTO))
#define ABSL_RANDEN_HWAES_IMPL 1

#elif ABSL_RANDOM_INTERNAL_AES_DISPATCH && !defined(__APPLE__) && \
    (defined(__GNUC__) && __GNUC__ > 4 || __GNUC__ == 4 && __GNUC_MINOR__ > 9)
// ...or, on GCC, we can use an ASM directive to
// instruct the assember to allow crypto instructions.
#define ABSL_RANDEN_HWAES_IMPL 1
#define ABSL_RANDEN_HWAES_IMPL_CRYPTO_DIRECTIVE 1
#endif
#else
// HWAES is unsupported by these architectures / platforms:
//   __myriad2__
//   __mips__
//
// Other architectures / platforms are unknown.
//
// See the Abseil documentation on supported macros at:
// https://abseil.io/docs/cpp/platforms/macros
#endif

#if !defined(ABSL_RANDEN_HWAES_IMPL)
// No accelerated implementation is supported.
// The RandenHwAes functions are stubs that print an error and exit.

#include <cstdio>
#include <cstdlib>

namespace absl {
namespace random_internal {

// No accelerated implementation.
bool HasRandenHwAesImplementation() { return false; }

// NOLINTNEXTLINE
const void* RandenHwAes::GetKeys() {
  // Attempted to dispatch to an unsupported dispatch target.
  const int d = ABSL_RANDOM_INTERNAL_AES_DISPATCH;
  fprintf(stderr, "AES Hardware detection failed (%d).\n", d);
  exit(1);
  return nullptr;
}

// NOLINTNEXTLINE
void RandenHwAes::Absorb(const void*, void*) {
  // Attempted to dispatch to an unsupported dispatch target.
  const int d = ABSL_RANDOM_INTERNAL_AES_DISPATCH;
  fprintf(stderr, "AES Hardware detection failed (%d).\n", d);
  exit(1);
}

// NOLINTNEXTLINE
void RandenHwAes::Generate(const void*, void*) {
  // Attempted to dispatch to an unsupported dispatch target.
  const int d = ABSL_RANDOM_INTERNAL_AES_DISPATCH;
  fprintf(stderr, "AES Hardware detection failed (%d).\n", d);
  exit(1);
}

}  // namespace random_internal
}  // namespace absl

#else  // defined(ABSL_RANDEN_HWAES_IMPL)
//
// Accelerated implementations are supported.
// We need the per-architecture includes and defines.
//

#include "absl/random/internal/randen_traits.h"

// ABSL_FUNCTION_ALIGN32 defines a 32-byte alignment attribute
// for the functions in this file.
//
// NOTE: Determine whether we actually have any wins from ALIGN32
// using microbenchmarks. If not, remove.
#undef ABSL_FUNCTION_ALIGN32
#if ABSL_HAVE_ATTRIBUTE(aligned) || (defined(__GNUC__) && !defined(__clang__))
#define ABSL_FUNCTION_ALIGN32 __attribute__((aligned(32)))
#else
#define ABSL_FUNCTION_ALIGN32
#endif

// TARGET_CRYPTO defines a crypto attribute for each architecture.
//
// NOTE: Evaluate whether we should eliminate ABSL_TARGET_CRYPTO.
#if (defined(__clang__) || defined(__GNUC__))
#if defined(ABSL_ARCH_X86_64) || defined(ABSL_ARCH_X86_32)
#define ABSL_TARGET_CRYPTO __attribute__((target("aes")))
#elif defined(ABSL_ARCH_PPC)
#define ABSL_TARGET_CRYPTO __attribute__((target("crypto")))
#else
#define ABSL_TARGET_CRYPTO
#endif
#else
#define ABSL_TARGET_CRYPTO
#endif

#if defined(ABSL_ARCH_PPC)
// NOTE: Keep in mind that PPC can operate in little-endian or big-endian mode,
// however the PPC altivec vector registers (and thus the AES instructions)
// always operate in big-endian mode.

#include <altivec.h>
// <altivec.h> #defines vector __vector; in C++, this is bad form.
#undef vector

// Rely on the PowerPC AltiVec vector operations for accelerated AES
// instructions. GCC support of the PPC vector types is described in:
// https://gcc.gnu.org/onlinedocs/gcc-4.9.0/gcc/PowerPC-AltiVec_002fVSX-Built-in-Functions.html
//
// Already provides operator^=.
using Vector128 = __vector unsigned long long;  // NOLINT(runtime/int)

namespace {

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128
ReverseBytes(const Vector128& v) {
  // Reverses the bytes of the vector.
  const __vector unsigned char perm = {15, 14, 13, 12, 11, 10, 9, 8,
                                       7,  6,  5,  4,  3,  2,  1, 0};
  return vec_perm(v, v, perm);
}

// WARNING: these load/store in native byte order. It is OK to load and then
// store an unchanged vector, but interpreting the bits as a number or input
// to AES will have undefined results.
inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128
Vector128Load(const void* ABSL_RANDOM_INTERNAL_RESTRICT from) {
  return vec_vsx_ld(0, reinterpret_cast<const Vector128*>(from));
}

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void
Vector128Store(const Vector128& v, void* ABSL_RANDOM_INTERNAL_RESTRICT to) {
  vec_vsx_st(v, 0, reinterpret_cast<Vector128*>(to));
}

// One round of AES. "round_key" is a public constant for breaking the
// symmetry of AES (ensures previously equal columns differ afterwards).
inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128
AesRound(const Vector128& state, const Vector128& round_key) {
  return Vector128(__builtin_crypto_vcipher(state, round_key));
}

// Enables native loads in the round loop by pre-swapping.
inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void
SwapEndian(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state) {
  using absl::random_internal::RandenTraits;
  constexpr size_t kLanes = 2;
  constexpr size_t kFeistelBlocks = RandenTraits::kFeistelBlocks;

  for (uint32_t branch = 0; branch < kFeistelBlocks; ++branch) {
    const Vector128 v = ReverseBytes(Vector128Load(state + kLanes * branch));
    Vector128Store(v, state + kLanes * branch);
  }
}

}  // namespace

#elif defined(ABSL_ARCH_ARM) || defined(ABSL_ARCH_AARCH64)

// This asm directive will cause the file to be compiled with crypto extensions
// whether or not the cpu-architecture supports it.
#if ABSL_RANDEN_HWAES_IMPL_CRYPTO_DIRECTIVE
asm(".arch_extension  crypto\n");

// Override missing defines.
#if !defined(__ARM_NEON)
#define __ARM_NEON 1
#endif

#if !defined(__ARM_FEATURE_CRYPTO)
#define __ARM_FEATURE_CRYPTO 1
#endif

#endif

// Rely on the ARM NEON+Crypto advanced simd types, defined in <arm_neon.h>.
// uint8x16_t is the user alias for underlying __simd128_uint8_t type.
// http://infocenter.arm.com/help/topic/com.arm.doc.ihi0073a/IHI0073A_arm_neon_intrinsics_ref.pdf
//
// <arm_neon> defines the following
//
// typedef __attribute__((neon_vector_type(16))) uint8_t uint8x16_t;
// typedef __attribute__((neon_vector_type(16))) int8_t int8x16_t;
// typedef __attribute__((neon_polyvector_type(16))) int8_t poly8x16_t;
//
// vld1q_v
// vst1q_v
// vaeseq_v
// vaesmcq_v
#include <arm_neon.h>

// Already provides operator^=.
using Vector128 = uint8x16_t;

namespace {

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128
Vector128Load(const void* ABSL_RANDOM_INTERNAL_RESTRICT from) {
  return vld1q_u8(reinterpret_cast<const uint8_t*>(from));
}

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void
Vector128Store(const Vector128& v, void* ABSL_RANDOM_INTERNAL_RESTRICT to) {
  vst1q_u8(reinterpret_cast<uint8_t*>(to), v);
}

// One round of AES. "round_key" is a public constant for breaking the
// symmetry of AES (ensures previously equal columns differ afterwards).
inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128
AesRound(const Vector128& state, const Vector128& round_key) {
  // It is important to always use the full round function - omitting the
  // final MixColumns reduces security [https://eprint.iacr.org/2010/041.pdf]
  // and does not help because we never decrypt.
  //
  // Note that ARM divides AES instructions differently than x86 / PPC,
  // And we need to skip the first AddRoundKey step and add an extra
  // AddRoundKey step to the end. Lucky for us this is just XOR.
  return vaesmcq_u8(vaeseq_u8(state, uint8x16_t{})) ^ round_key;
}

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void
SwapEndian(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT) {}

}  // namespace

#elif defined(ABSL_ARCH_X86_64) || defined(ABSL_ARCH_X86_32)
// On x86 we rely on the aesni instructions
#include <wmmintrin.h>

namespace {

// Vector128 class is only wrapper for __m128i, benchmark indicates that it's
// faster than using __m128i directly.
class Vector128 {
 public:
  // Convert from/to intrinsics.
  inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE explicit Vector128(
      const __m128i& Vector128)
      : data_(Vector128) {}

  inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE __m128i data() const {
    return data_;
  }

  inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128& operator^=(
      const Vector128& other) {
    data_ = _mm_xor_si128(data_, other.data());
    return *this;
  }

 private:
  __m128i data_;
};

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128
Vector128Load(const void* ABSL_RANDOM_INTERNAL_RESTRICT from) {
  return Vector128(_mm_load_si128(reinterpret_cast<const __m128i*>(from)));
}

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void
Vector128Store(const Vector128& v, void* ABSL_RANDOM_INTERNAL_RESTRICT to) {
  _mm_store_si128(reinterpret_cast<__m128i * ABSL_RANDOM_INTERNAL_RESTRICT>(to),
                  v.data());
}

// One round of AES. "round_key" is a public constant for breaking the
// symmetry of AES (ensures previously equal columns differ afterwards).
inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE Vector128
AesRound(const Vector128& state, const Vector128& round_key) {
  // It is important to always use the full round function - omitting the
  // final MixColumns reduces security [https://eprint.iacr.org/2010/041.pdf]
  // and does not help because we never decrypt.
  return Vector128(_mm_aesenc_si128(state.data(), round_key.data()));
}

inline ABSL_TARGET_CRYPTO ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE void
SwapEndian(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT) {}

}  // namespace

#endif

namespace {

// u64x2 is a 128-bit, (2 x uint64_t lanes) struct used to store
// the randen_keys.
struct alignas(16) u64x2 {
  constexpr u64x2(uint64_t hi, uint64_t lo)
#if defined(ABSL_ARCH_PPC)
      // This has been tested with PPC running in little-endian mode;
      // We byte-swap the u64x2 structure from little-endian to big-endian
      // because altivec always runs in big-endian mode.
      : v{__builtin_bswap64(hi), __builtin_bswap64(lo)} {
#else
      : v{lo, hi} {
#endif
  }

  constexpr bool operator==(const u64x2& other) const {
    return v[0] == other.v[0] && v[1] == other.v[1];
  }

  constexpr bool operator!=(const u64x2& other) const {
    return !(*this == other);
  }

  uint64_t v[2];
};  // namespace

#ifdef __clang__
#pragma clang diagnostic push
#pragma clang diagnostic ignored "-Wunknown-pragmas"
#endif

// At this point, all of the platform-specific features have been defined /
// implemented.
//
// REQUIRES: using u64x2 = ...
// REQUIRES: using Vector128 = ...
// REQUIRES: Vector128 Vector128Load(void*) {...}
// REQUIRES: void Vector128Store(Vector128, void*) {...}
// REQUIRES: Vector128 AesRound(Vector128, Vector128) {...}
// REQUIRES: void SwapEndian(uint64_t*) {...}
//
// PROVIDES: absl::random_internal::RandenHwAes::Absorb
// PROVIDES: absl::random_internal::RandenHwAes::Generate

// RANDen = RANDom generator or beetroots in Swiss German.
// 'Strong' (well-distributed, unpredictable, backtracking-resistant) random
// generator, faster in some benchmarks than std::mt19937_64 and pcg64_c32.
//
// High-level summary:
// 1) Reverie (see "A Robust and Sponge-Like PRNG with Improved Efficiency") is
//    a sponge-like random generator that requires a cryptographic permutation.
//    It improves upon "Provably Robust Sponge-Based PRNGs and KDFs" by
//    achieving backtracking resistance with only one Permute() per buffer.
//
// 2) "Simpira v2: A Family of Efficient Permutations Using the AES Round
//    Function" constructs up to 1024-bit permutations using an improved
//    Generalized Feistel network with 2-round AES-128 functions. This Feistel
//    block shuffle achieves diffusion faster and is less vulnerable to
//    sliced-biclique attacks than the Type-2 cyclic shuffle.
//
// 3) "Improving the Generalized Feistel" and "New criterion for diffusion
//    property" extends the same kind of improved Feistel block shuffle to 16
//    branches, which enables a 2048-bit permutation.
//
// We combine these three ideas and also change Simpira's subround keys from
// structured/low-entropy counters to digits of Pi.

// Randen constants.
using absl::random_internal::RandenTraits;
constexpr size_t kStateBytes = RandenTraits::kStateBytes;
constexpr size_t kCapacityBytes = RandenTraits::kCapacityBytes;
constexpr size_t kFeistelBlocks = RandenTraits::kFeistelBlocks;
constexpr size_t kFeistelRounds = RandenTraits::kFeistelRounds;
constexpr size_t kFeistelFunctions = RandenTraits::kFeistelFunctions;

// Independent keys (272 = 2.1 KiB) for the first AES subround of each function.
constexpr size_t kKeys = kFeistelRounds * kFeistelFunctions;

// INCLUDE keys.
#include "absl/random/internal/randen-keys.inc"

static_assert(kKeys == kRoundKeys, "kKeys and kRoundKeys must be equal");
static_assert(round_keys[kKeys - 1] != u64x2(0, 0),
              "Too few round_keys initializers");

// Number of uint64_t lanes per 128-bit vector;
constexpr size_t kLanes = 2;

// Block shuffles applies a shuffle to the entire state between AES rounds.
// Improved odd-even shuffle from "New criterion for diffusion property".
inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE ABSL_TARGET_CRYPTO void
BlockShuffle(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state) {
  static_assert(kFeistelBlocks == 16, "Expecting 16 FeistelBlocks.");

  constexpr size_t shuffle[kFeistelBlocks] = {7,  2, 13, 4,  11, 8,  3, 6,
                                              15, 0, 9,  10, 1,  14, 5, 12};

  // The fully unrolled loop without the memcpy improves the speed by about
  // 30% over the equivalent loop.
  const Vector128 v0 = Vector128Load(state + kLanes * shuffle[0]);
  const Vector128 v1 = Vector128Load(state + kLanes * shuffle[1]);
  const Vector128 v2 = Vector128Load(state + kLanes * shuffle[2]);
  const Vector128 v3 = Vector128Load(state + kLanes * shuffle[3]);
  const Vector128 v4 = Vector128Load(state + kLanes * shuffle[4]);
  const Vector128 v5 = Vector128Load(state + kLanes * shuffle[5]);
  const Vector128 v6 = Vector128Load(state + kLanes * shuffle[6]);
  const Vector128 v7 = Vector128Load(state + kLanes * shuffle[7]);
  const Vector128 w0 = Vector128Load(state + kLanes * shuffle[8]);
  const Vector128 w1 = Vector128Load(state + kLanes * shuffle[9]);
  const Vector128 w2 = Vector128Load(state + kLanes * shuffle[10]);
  const Vector128 w3 = Vector128Load(state + kLanes * shuffle[11]);
  const Vector128 w4 = Vector128Load(state + kLanes * shuffle[12]);
  const Vector128 w5 = Vector128Load(state + kLanes * shuffle[13]);
  const Vector128 w6 = Vector128Load(state + kLanes * shuffle[14]);
  const Vector128 w7 = Vector128Load(state + kLanes * shuffle[15]);

  Vector128Store(v0, state + kLanes * 0);
  Vector128Store(v1, state + kLanes * 1);
  Vector128Store(v2, state + kLanes * 2);
  Vector128Store(v3, state + kLanes * 3);
  Vector128Store(v4, state + kLanes * 4);
  Vector128Store(v5, state + kLanes * 5);
  Vector128Store(v6, state + kLanes * 6);
  Vector128Store(v7, state + kLanes * 7);
  Vector128Store(w0, state + kLanes * 8);
  Vector128Store(w1, state + kLanes * 9);
  Vector128Store(w2, state + kLanes * 10);
  Vector128Store(w3, state + kLanes * 11);
  Vector128Store(w4, state + kLanes * 12);
  Vector128Store(w5, state + kLanes * 13);
  Vector128Store(w6, state + kLanes * 14);
  Vector128Store(w7, state + kLanes * 15);
}

// Feistel round function using two AES subrounds. Very similar to F()
// from Simpira v2, but with independent subround keys. Uses 17 AES rounds
// per 16 bytes (vs. 10 for AES-CTR). Computing eight round functions in
// parallel hides the 7-cycle AESNI latency on HSW. Note that the Feistel
// XORs are 'free' (included in the second AES instruction).
inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE ABSL_TARGET_CRYPTO const
    u64x2*
    FeistelRound(uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state,
                 const u64x2* ABSL_RANDOM_INTERNAL_RESTRICT keys) {
  static_assert(kFeistelBlocks == 16, "Expecting 16 FeistelBlocks.");

  // MSVC does a horrible job at unrolling loops.
  // So we unroll the loop by hand to improve the performance.
  const Vector128 s0 = Vector128Load(state + kLanes * 0);
  const Vector128 s1 = Vector128Load(state + kLanes * 1);
  const Vector128 s2 = Vector128Load(state + kLanes * 2);
  const Vector128 s3 = Vector128Load(state + kLanes * 3);
  const Vector128 s4 = Vector128Load(state + kLanes * 4);
  const Vector128 s5 = Vector128Load(state + kLanes * 5);
  const Vector128 s6 = Vector128Load(state + kLanes * 6);
  const Vector128 s7 = Vector128Load(state + kLanes * 7);
  const Vector128 s8 = Vector128Load(state + kLanes * 8);
  const Vector128 s9 = Vector128Load(state + kLanes * 9);
  const Vector128 s10 = Vector128Load(state + kLanes * 10);
  const Vector128 s11 = Vector128Load(state + kLanes * 11);
  const Vector128 s12 = Vector128Load(state + kLanes * 12);
  const Vector128 s13 = Vector128Load(state + kLanes * 13);
  const Vector128 s14 = Vector128Load(state + kLanes * 14);
  const Vector128 s15 = Vector128Load(state + kLanes * 15);

  // Encode even blocks with keys.
  const Vector128 e0 = AesRound(s0, Vector128Load(keys + 0));
  const Vector128 e2 = AesRound(s2, Vector128Load(keys + 1));
  const Vector128 e4 = AesRound(s4, Vector128Load(keys + 2));
  const Vector128 e6 = AesRound(s6, Vector128Load(keys + 3));
  const Vector128 e8 = AesRound(s8, Vector128Load(keys + 4));
  const Vector128 e10 = AesRound(s10, Vector128Load(keys + 5));
  const Vector128 e12 = AesRound(s12, Vector128Load(keys + 6));
  const Vector128 e14 = AesRound(s14, Vector128Load(keys + 7));

  // Encode odd blocks with even output from above.
  const Vector128 o1 = AesRound(e0, s1);
  const Vector128 o3 = AesRound(e2, s3);
  const Vector128 o5 = AesRound(e4, s5);
  const Vector128 o7 = AesRound(e6, s7);
  const Vector128 o9 = AesRound(e8, s9);
  const Vector128 o11 = AesRound(e10, s11);
  const Vector128 o13 = AesRound(e12, s13);
  const Vector128 o15 = AesRound(e14, s15);

  // Store odd blocks. (These will be shuffled later).
  Vector128Store(o1, state + kLanes * 1);
  Vector128Store(o3, state + kLanes * 3);
  Vector128Store(o5, state + kLanes * 5);
  Vector128Store(o7, state + kLanes * 7);
  Vector128Store(o9, state + kLanes * 9);
  Vector128Store(o11, state + kLanes * 11);
  Vector128Store(o13, state + kLanes * 13);
  Vector128Store(o15, state + kLanes * 15);

  return keys + 8;
}

// Cryptographic permutation based via type-2 Generalized Feistel Network.
// Indistinguishable from ideal by chosen-ciphertext adversaries using less than
// 2^64 queries if the round function is a PRF. This is similar to the b=8 case
// of Simpira v2, but more efficient than its generic construction for b=16.
inline ABSL_RANDOM_INTERNAL_ATTRIBUTE_ALWAYS_INLINE ABSL_TARGET_CRYPTO void
Permute(const void* ABSL_RANDOM_INTERNAL_RESTRICT keys,
        uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state) {
  const u64x2* ABSL_RANDOM_INTERNAL_RESTRICT keys128 =
      static_cast<const u64x2*>(keys);

  // (Successfully unrolled; the first iteration jumps into the second half)
#ifdef __clang__
#pragma clang loop unroll_count(2)
#endif
  for (size_t round = 0; round < kFeistelRounds; ++round) {
    keys128 = FeistelRound(state, keys128);
    BlockShuffle(state);
  }
}

}  // namespace

namespace absl {
namespace random_internal {

bool HasRandenHwAesImplementation() { return true; }

const void* ABSL_TARGET_CRYPTO ABSL_FUNCTION_ALIGN32 ABSL_ATTRIBUTE_FLATTEN
RandenHwAes::GetKeys() {
  // Round keys for one AES per Feistel round and branch.
  // The canonical implementation uses first digits of Pi.
  return round_keys;
}

// NOLINTNEXTLINE
void ABSL_TARGET_CRYPTO ABSL_FUNCTION_ALIGN32 ABSL_ATTRIBUTE_FLATTEN
RandenHwAes::Absorb(const void* seed_void, void* state_void) {
  uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state =
      reinterpret_cast<uint64_t*>(state_void);
  const uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT seed =
      reinterpret_cast<const uint64_t*>(seed_void);

  constexpr size_t kCapacityBlocks = kCapacityBytes / sizeof(Vector128);
  constexpr size_t kStateBlocks = kStateBytes / sizeof(Vector128);

  static_assert(kCapacityBlocks * sizeof(Vector128) == kCapacityBytes,
                "Not i*V");
  static_assert(kCapacityBlocks == 1, "Unexpected Randen kCapacityBlocks");
  static_assert(kStateBlocks == 16, "Unexpected Randen kStateBlocks");

  Vector128 b1 = Vector128Load(state + kLanes * 1);
  b1 ^= Vector128Load(seed + kLanes * 0);
  Vector128Store(b1, state + kLanes * 1);

  Vector128 b2 = Vector128Load(state + kLanes * 2);
  b2 ^= Vector128Load(seed + kLanes * 1);
  Vector128Store(b2, state + kLanes * 2);

  Vector128 b3 = Vector128Load(state + kLanes * 3);
  b3 ^= Vector128Load(seed + kLanes * 2);
  Vector128Store(b3, state + kLanes * 3);

  Vector128 b4 = Vector128Load(state + kLanes * 4);
  b4 ^= Vector128Load(seed + kLanes * 3);
  Vector128Store(b4, state + kLanes * 4);

  Vector128 b5 = Vector128Load(state + kLanes * 5);
  b5 ^= Vector128Load(seed + kLanes * 4);
  Vector128Store(b5, state + kLanes * 5);

  Vector128 b6 = Vector128Load(state + kLanes * 6);
  b6 ^= Vector128Load(seed + kLanes * 5);
  Vector128Store(b6, state + kLanes * 6);

  Vector128 b7 = Vector128Load(state + kLanes * 7);
  b7 ^= Vector128Load(seed + kLanes * 6);
  Vector128Store(b7, state + kLanes * 7);

  Vector128 b8 = Vector128Load(state + kLanes * 8);
  b8 ^= Vector128Load(seed + kLanes * 7);
  Vector128Store(b8, state + kLanes * 8);

  Vector128 b9 = Vector128Load(state + kLanes * 9);
  b9 ^= Vector128Load(seed + kLanes * 8);
  Vector128Store(b9, state + kLanes * 9);

  Vector128 b10 = Vector128Load(state + kLanes * 10);
  b10 ^= Vector128Load(seed + kLanes * 9);
  Vector128Store(b10, state + kLanes * 10);

  Vector128 b11 = Vector128Load(state + kLanes * 11);
  b11 ^= Vector128Load(seed + kLanes * 10);
  Vector128Store(b11, state + kLanes * 11);

  Vector128 b12 = Vector128Load(state + kLanes * 12);
  b12 ^= Vector128Load(seed + kLanes * 11);
  Vector128Store(b12, state + kLanes * 12);

  Vector128 b13 = Vector128Load(state + kLanes * 13);
  b13 ^= Vector128Load(seed + kLanes * 12);
  Vector128Store(b13, state + kLanes * 13);

  Vector128 b14 = Vector128Load(state + kLanes * 14);
  b14 ^= Vector128Load(seed + kLanes * 13);
  Vector128Store(b14, state + kLanes * 14);

  Vector128 b15 = Vector128Load(state + kLanes * 15);
  b15 ^= Vector128Load(seed + kLanes * 14);
  Vector128Store(b15, state + kLanes * 15);
}

// NOLINTNEXTLINE
void ABSL_TARGET_CRYPTO ABSL_FUNCTION_ALIGN32 ABSL_ATTRIBUTE_FLATTEN
RandenHwAes::Generate(const void* keys, void* state_void) {
  static_assert(kCapacityBytes == sizeof(Vector128), "Capacity mismatch");

  uint64_t* ABSL_RANDOM_INTERNAL_RESTRICT state =
      reinterpret_cast<uint64_t*>(state_void);

  const Vector128 prev_inner = Vector128Load(state);

  SwapEndian(state);

  Permute(keys, state);

  SwapEndian(state);

  // Ensure backtracking resistance.
  Vector128 inner = Vector128Load(state);
  inner ^= prev_inner;
  Vector128Store(inner, state);
}

#ifdef __clang__
#pragma clang diagnostic pop
#endif

}  // namespace random_internal
}  // namespace absl

#endif  // (ABSL_RANDEN_HWAES_IMPL)