visaljs 0.0.2 → 0.0.4

package/runtime/third_party/boringssl/crypto/fipsmodule/mldsa/mldsa.cc.inc

@@ -0,0 +1,3323 @@
+// Copyright 2014 The BoringSSL 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.
+
+#include <openssl/base.h>
+
+#include <memory>
+
+#include <assert.h>
+#include <stdlib.h>
+
+#include <openssl/bytestring.h>
+#include <openssl/mem.h>
+#include <openssl/rand.h>
+
+#include "../../internal.h"
+#include "../../mem_internal.h"
+#include "../bcm_interface.h"
+#include "../keccak/internal.h"
+
+
+using namespace bssl;
+
+#if defined(BORINGSSL_FIPS)
+
+DEFINE_STATIC_ONCE(g_mldsa_keygen_self_test_once)
+DEFINE_STATIC_ONCE(g_mldsa_sign_self_test_once)
+DEFINE_STATIC_ONCE(g_mldsa_verify_self_test_once)
+
+#endif
+
+namespace mldsa {
+namespace {
+
+namespace fips {
+void ensure_keygen_self_test();
+void ensure_sign_self_test();
+void ensure_verify_self_test();
+}  // namespace fips
+
+constexpr int kDegree = 256;
+constexpr int kRhoBytes = 32;
+constexpr int kSigmaBytes = 64;
+constexpr int kKBytes = 32;
+constexpr int kTrBytes = 64;
+constexpr int kMuBytes = 64;
+constexpr int kRhoPrimeBytes = 64;
+
+// 2^23 - 2^13 + 1
+constexpr uint32_t kPrime = 8380417;
+// Inverse of -kPrime modulo 2^32
+constexpr uint32_t kPrimeNegInverse = 4236238847;
+constexpr int kDroppedBits = 13;
+constexpr uint32_t kHalfPrime = (kPrime - 1) / 2;
+// 256^-1 ⋅ R² mod kPrime. This is not 256^-1 in Montgomery form, but 256^-1
+// doubly-converted to Montgomery form.
+constexpr uint32_t kInverseDegreeDoubleMontgomery = 41978;
+
+// Constants that vary depending on ML-DSA size.
+//
+// These are implemented as templates which take the K parameter to distinguish
+// the ML-DSA sizes.
+
+template <int K>
+constexpr size_t public_key_bytes() {
+  if constexpr (K == 6) {
+    return MLDSA65_PUBLIC_KEY_BYTES;
+  } else if constexpr (K == 8) {
+    return MLDSA87_PUBLIC_KEY_BYTES;
+  } else if constexpr (K == 4) {
+    return MLDSA44_PUBLIC_KEY_BYTES;
+  }
+}
+
+template <int K>
+constexpr size_t signature_bytes() {
+  if constexpr (K == 6) {
+    return MLDSA65_SIGNATURE_BYTES;
+  } else if constexpr (K == 8) {
+    return MLDSA87_SIGNATURE_BYTES;
+  } else if constexpr (K == 4) {
+    return MLDSA44_SIGNATURE_BYTES;
+  }
+}
+
+template <int K>
+constexpr int tau() {
+  if constexpr (K == 6) {
+    return 49;
+  } else if constexpr (K == 8) {
+    return 60;
+  } else if constexpr (K == 4) {
+    return 39;
+  }
+}
+
+template <int K>
+constexpr int lambda_bytes() {
+  if constexpr (K == 6) {
+    return 192 / 8;
+  } else if constexpr (K == 8) {
+    return 256 / 8;
+  } else if constexpr (K == 4) {
+    return 128 / 8;
+  }
+}
+
+template <int K>
+constexpr int gamma1_bits() {
+  if constexpr (K == 6 || K == 8) {
+    return 19;
+  } else if constexpr (K == 4) {
+    return 17;
+  }
+}
+
+template <int K>
+constexpr int gamma1() {
+  return 1 << gamma1_bits<K>();
+}
+
+template <int K>
+constexpr int scalar_le_gamma1_bytes() {
+  return ((gamma1_bits<K>() + 1) * kDegree) / 8;
+}
+
+template <int K>
+constexpr uint32_t w1_coeffs_bits() {
+  if constexpr (K == 6 || K == 8) {
+    return 4;
+  } else if constexpr (K == 4) {
+    return 6;
+  }
+}
+
+template <int K>
+constexpr uint32_t w1_scalar_bytes() {
+  return (w1_coeffs_bits<K>() * kDegree) / 8;
+}
+
+template <int K>
+constexpr uint32_t w1_bytes() {
+  return w1_scalar_bytes<K>() * K;
+}
+
+template <int K>
+constexpr uint32_t prime_minus_one_over_gamma2() {
+  if constexpr (K == 6 || K == 8) {
+    return 32;
+  } else if constexpr (K == 4) {
+    return 88;
+  }
+}
+
+template <int K>
+constexpr uint32_t gamma2() {
+  return (kPrime - 1) / prime_minus_one_over_gamma2<K>();
+}
+
+template <int K>
+constexpr int beta() {
+  if constexpr (K == 6) {
+    return 196;
+  } else if constexpr (K == 8) {
+    return 120;
+  } else if constexpr (K == 4) {
+    return 78;
+  }
+}
+
+template <int K>
+constexpr int omega() {
+  if constexpr (K == 6) {
+    return 55;
+  } else if constexpr (K == 8) {
+    return 75;
+  } else if constexpr (K == 4) {
+    return 80;
+  }
+}
+
+template <int K>
+constexpr int eta() {
+  if constexpr (K == 6) {
+    return 4;
+  } else if constexpr (K == 8 || K == 4) {
+    return 2;
+  }
+}
+
+template <int K>
+constexpr int plus_minus_eta_bitlen() {
+  if constexpr (K == 6) {
+    return 4;
+  } else if constexpr (K == 8 || K == 4) {
+    return 3;
+  }
+}
+
+// Fundamental types.
+
+struct scalar {
+  uint32_t c[kDegree];
+};
+
+template <int K>
+struct vector {
+  scalar v[K];
+};
+
+template <int K, int L>
+struct matrix {
+  scalar v[K][L];
+};
+
+/* Arithmetic */
+
+// This bit of Python will be referenced in some of the following comments:
+//
+// q = 8380417
+// # Inverse of -q modulo 2^32
+// q_neg_inverse = 4236238847
+// # 2^64 modulo q
+// montgomery_square = 2365951
+//
+// def bitreverse(i):
+//     ret = 0
+//     for n in range(8):
+//         bit = i & 1
+//         ret <<= 1
+//         ret |= bit
+//         i >>= 1
+//     return ret
+//
+// def montgomery_reduce(x):
+//     a = (x * q_neg_inverse) % 2**32
+//     b = x + a * q
+//     assert b & 0xFFFF_FFFF == 0
+//     c = b >> 32
+//     assert c < q
+//     return c
+//
+// def montgomery_transform(x):
+//     return montgomery_reduce(x * montgomery_square)
+
+// kNTTRootsMontgomery = [
+//   montgomery_transform(pow(1753, bitreverse(i), q)) for i in range(256)
+// ]
+static const uint32_t kNTTRootsMontgomery[256] = {
+    4193792, 25847,   5771523, 7861508, 237124,  7602457, 7504169, 466468,
+    1826347, 2353451, 8021166, 6288512, 3119733, 5495562, 3111497, 2680103,
+    2725464, 1024112, 7300517, 3585928, 7830929, 7260833, 2619752, 6271868,
+    6262231, 4520680, 6980856, 5102745, 1757237, 8360995, 4010497, 280005,
+    2706023, 95776,   3077325, 3530437, 6718724, 4788269, 5842901, 3915439,
+    4519302, 5336701, 3574422, 5512770, 3539968, 8079950, 2348700, 7841118,
+    6681150, 6736599, 3505694, 4558682, 3507263, 6239768, 6779997, 3699596,
+    811944,  531354,  954230,  3881043, 3900724, 5823537, 2071892, 5582638,
+    4450022, 6851714, 4702672, 5339162, 6927966, 3475950, 2176455, 6795196,
+    7122806, 1939314, 4296819, 7380215, 5190273, 5223087, 4747489, 126922,
+    3412210, 7396998, 2147896, 2715295, 5412772, 4686924, 7969390, 5903370,
+    7709315, 7151892, 8357436, 7072248, 7998430, 1349076, 1852771, 6949987,
+    5037034, 264944,  508951,  3097992, 44288,   7280319, 904516,  3958618,
+    4656075, 8371839, 1653064, 5130689, 2389356, 8169440, 759969,  7063561,
+    189548,  4827145, 3159746, 6529015, 5971092, 8202977, 1315589, 1341330,
+    1285669, 6795489, 7567685, 6940675, 5361315, 4499357, 4751448, 3839961,
+    2091667, 3407706, 2316500, 3817976, 5037939, 2244091, 5933984, 4817955,
+    266997,  2434439, 7144689, 3513181, 4860065, 4621053, 7183191, 5187039,
+    900702,  1859098, 909542,  819034,  495491,  6767243, 8337157, 7857917,
+    7725090, 5257975, 2031748, 3207046, 4823422, 7855319, 7611795, 4784579,
+    342297,  286988,  5942594, 4108315, 3437287, 5038140, 1735879, 203044,
+    2842341, 2691481, 5790267, 1265009, 4055324, 1247620, 2486353, 1595974,
+    4613401, 1250494, 2635921, 4832145, 5386378, 1869119, 1903435, 7329447,
+    7047359, 1237275, 5062207, 6950192, 7929317, 1312455, 3306115, 6417775,
+    7100756, 1917081, 5834105, 7005614, 1500165, 777191,  2235880, 3406031,
+    7838005, 5548557, 6709241, 6533464, 5796124, 4656147, 594136,  4603424,
+    6366809, 2432395, 2454455, 8215696, 1957272, 3369112, 185531,  7173032,
+    5196991, 162844,  1616392, 3014001, 810149,  1652634, 4686184, 6581310,
+    5341501, 3523897, 3866901, 269760,  2213111, 7404533, 1717735, 472078,
+    7953734, 1723600, 6577327, 1910376, 6712985, 7276084, 8119771, 4546524,
+    5441381, 6144432, 7959518, 6094090, 183443,  7403526, 1612842, 4834730,
+    7826001, 3919660, 8332111, 7018208, 3937738, 1400424, 7534263, 1976782};
+
+// Reduces x mod kPrime in constant time, where 0 <= x < 2*kPrime.
+inline uint32_t reduce_once(uint32_t x) {
+  declassify_assert(x < 2 * kPrime);
+  // return x < kPrime ? x : x - kPrime;
+  const uint32_t subtracted = x - kPrime;
+  uint32_t mask = 0u - (subtracted >> 31);
+  // Although this is a constant-time select, we omit a value barrier here.
+  // Value barriers impede auto-vectorization (likely because it forces the
+  // value to transit through a general-purpose register). This is a difference
+  // of 1.4x to 1.5x in signing performance.
+  //
+  // We usually add value barriers to selects because Clang turns consecutive
+  // selects with the same condition into a branch instead of CMOV/CSEL. This
+  // condition does not occur in ML-DSA, so omitting it seems to be generally
+  // safe. However, see `coefficient_from_nibble`.
+  return (mask & x) | (~mask & subtracted);
+}
+
+// reduce_from_u32 returns `x mod kPrime`. `x` can be any `uint32_t`.
+inline uint32_t reduce_from_u32(uint32_t x) {
+  // kPrime is 2^23 - 2^13 + 1. We first split `x` into the upper 9 bits and the
+  // lower 23 bits:
+  uint32_t hi = x >> 23, lo = x & ((uint32_t{1} << 23) - 1);
+  // 2^23 = 2^13 - 1 mod kPrime
+  //    x = hi * 2^23 + lo
+  //      = hi * 2^13 - hi + lo mod kPrime
+  //
+  // This value does not overflow and is below 2 * kPrime. We can `reduce_once`.
+  static constexpr uint32_t kMaxHi = UINT32_MAX >> 23;
+  static constexpr uint32_t kMaxLo = (uint32_t{1} << 23) - 1;
+  static_assert((kMaxHi << 13) - kMaxHi + kMaxLo < 2 * kPrime);
+  return reduce_once((hi << 13) - hi + lo);
+}
+
+// Returns the absolute value in constant time, interpreting the high bit as a
+// sign bit.
+inline uint32_t abs_signed(uint32_t x) {
+  // return is_negative(x) ? -x : x;
+  uint32_t mask = 0u - (x >> 31);
+  return constant_time_select_32(mask, 0u - x, x);
+}
+
+// Returns the absolute value modulo kPrime.
+inline uint32_t abs_mod_prime(uint32_t x) {
+  declassify_assert(x < kPrime);
+  // return x <= kHalfPrime ? x : kPrime - x;
+  uint32_t mask = x - kHalfPrime - 1;
+  mask = 0u - (mask >> 31);
+  return constant_time_select_32(mask, x, kPrime - x);
+}
+
+// Returns the maximum of two values in constant time. Each value must be less
+// than kPrime.
+inline uint32_t maximum_reduced(uint32_t x, uint32_t y) {
+  declassify_assert(x < kPrime);
+  declassify_assert(y < kPrime);
+  // return x < y ? y : x;
+  uint32_t mask = x - y;
+  mask = 0u - (mask >> 31);
+  return constant_time_select_32(mask, y, x);
+}
+
+inline uint32_t mod_sub(uint32_t a, uint32_t b) {
+  declassify_assert(a < kPrime);
+  declassify_assert(b < kPrime);
+  uint32_t r = a - b;
+  // return r < 0 ? r + kPrime : r;
+  uint32_t mask = 0u - (r >> 31);
+  // See `reduce_once` for why this does not have a value barrier.
+  return (mask & (r + kPrime)) | (~mask & r);
+}
+
+inline void scalar_add(scalar *out, const scalar *lhs, const scalar *rhs) {
+  for (int i = 0; i < kDegree; i++) {
+    out->c[i] = reduce_once(lhs->c[i] + rhs->c[i]);
+  }
+}
+
+inline void scalar_sub(scalar *out, const scalar *lhs, const scalar *rhs) {
+  for (int i = 0; i < kDegree; i++) {
+    out->c[i] = mod_sub(lhs->c[i], rhs->c[i]);
+  }
+}
+
+// reduce_montgomery_partial returns a number equivalent to x * R^-1 mod kPrime.
+// The output is less than 2 * kPrime. `x` must be at most kPrime * 2^32.
+inline uint32_t reduce_montgomery_partial(uint64_t x) {
+  declassify_assert(x <= ((uint64_t)kPrime << 32));
+  uint64_t a = (uint32_t)x * kPrimeNegInverse;
+  uint64_t b = x + a * kPrime;
+  declassify_assert((b & 0xffffffff) == 0);
+  return b >> 32;
+}
+
+// reduce_montgomery returns x * R^-1 mod kPrime, fully reduced. `x` must be at
+// most kPrime * 2^32.
+inline uint32_t reduce_montgomery(uint64_t x) {
+  return reduce_once(reduce_montgomery_partial(x));
+}
+
+// Montgomery-multiply two scalars in the number theoretically transformed
+// state. Each term gains an extra R^-1 term, which is expected to be canceled
+// by `scalar_inverse_ntt_montgomery`.
+inline void scalar_mult_montgomery(scalar *out, const scalar *lhs,
+                                   const scalar *rhs) {
+  for (int i = 0; i < kDegree; i++) {
+    out->c[i] = reduce_montgomery((uint64_t)lhs->c[i] * (uint64_t)rhs->c[i]);
+  }
+}
+
+// In place number theoretic transform of a given scalar.
+//
+// FIPS 204, Algorithm 41 (`NTT`).
+inline void scalar_ntt(scalar *s) {
+  // Manually unrolled loop to maximize vectorization.
+  //
+  // Intermediate values of `s` are not fully reduced. `ITER` writes to every
+  // element of `s` exactly once. Each write adds less than 2*kPrime to some
+  // value from the previous iteration. After eight iterations, each element
+  // will be less than 17*kPrime. This fits in `uint32_t`, so we never overflow.
+#define ITER(step, offset)                                                    \
+  {                                                                           \
+    int k = 0;                                                                \
+    for (int i = 0; i < step; i++) {                                          \
+      const uint32_t step_root = kNTTRootsMontgomery[step + i];               \
+      for (int j = k; j < k + offset; j++) {                                  \
+        uint32_t even = s->c[j];                                              \
+        /* `reduce_montgomery_partial` works on values up to kPrime*R.        \
+         * `step_root < kPrime` because it's static data. `s->c[...] < R`     \
+         * because it's 32-bit. */                                            \
+        uint32_t odd = reduce_montgomery_partial(uint64_t{step_root} *        \
+                                                 uint64_t{s->c[j + offset]}); \
+        s->c[j] = even + odd;                                                 \
+        s->c[j + offset] = even + 2 * kPrime - odd;                           \
+      }                                                                       \
+      k += 2 * offset;                                                        \
+    }                                                                         \
+  }
+
+  ITER(1, 128)
+  ITER(2, 64)
+  ITER(4, 32)
+  ITER(8, 16)
+  ITER(16, 8)
+  ITER(32, 4)
+  ITER(64, 2)
+  ITER(128, 1)
+  static_assert(kDegree == 256);
+#undef ITER
+
+  // Now `s` is correct but the elements are not fully reduced. Reduce them.
+  for (int i = 0; i < kDegree; i++) {
+    s->c[i] = reduce_from_u32(s->c[i]);
+  }
+}
+
+// In place inverse number theoretic transform of a given scalar, but with each
+// term multiplied by an extra factor of R. This is function is intended to be
+// used with products that have gone through `reduce_montgomery` one extra time,
+// so the extra R^-1 term is canceled out.
+//
+// FIPS 204, Algorithm 42 (`NTT^-1`).
+inline void scalar_inverse_ntt_montgomery(scalar *s) {
+  // Manually unrolled loop to maximize vectorization.
+  //
+  // Intermediate values of `s` are not fully reduced. `ITER` writes
+  // to every element of `s` exactly once. Each write either adds two values
+  // from the previous iteration, or is at most 2 * kPrime. After eight
+  // iterations, each element will be less than 2⁸ * kPrime = 256 * kPrime.
+  // This fits in `uint32_t`, so we never overflow.
+#define ITER(step, offset)                                                  \
+  {                                                                         \
+    int k = 0;                                                              \
+    for (int i = 0; i < step; i++) {                                        \
+      const uint32_t step_root =                                            \
+          kPrime - kNTTRootsMontgomery[step + (step - 1 - i)];              \
+      for (int j = k; j < k + offset; j++) {                                \
+        uint32_t even = s->c[j];                                            \
+        uint32_t odd = s->c[j + offset];                                    \
+        s->c[j] = odd + even;                                               \
+        /* `reduce_montgomery_partial` works on values up to kPrime*R.      \
+         * `step_root < kPrime` because it's static data. The other term is \
+         * less than R because it's 32-bit, and does not overflow because   \
+         * `odd` is at most 128 * kPrime per the above. */                  \
+        s->c[j + offset] = reduce_montgomery_partial(                       \
+            uint64_t{step_root} * uint64_t{128 * kPrime + even - odd});     \
+      }                                                                     \
+      k += 2 * offset;                                                      \
+    }                                                                       \
+  }
+
+  ITER(128, 1)
+  ITER(64, 2)
+  ITER(32, 4)
+  ITER(16, 8)
+  ITER(8, 16)
+  ITER(4, 32)
+  ITER(2, 64)
+  ITER(1, 128)
+  static_assert(kDegree == 256);
+#undef ITER
+
+  // `kInverseDegreeDoubleMontgomery` has been double-converted to Montgomery
+  // form, so the result is an extra R term in the final result. This also fully
+  // reduces the partially-reduced values above.
+  for (int i = 0; i < kDegree; i++) {
+    s->c[i] = reduce_montgomery(uint64_t{s->c[i]} *
+                                uint64_t{kInverseDegreeDoubleMontgomery});
+  }
+}
+
+template <int X>
+inline void vector_zero(vector<X> *out) {
+  OPENSSL_memset(out, 0, sizeof(*out));
+}
+
+template <int X>
+inline void vector_add(vector<X> *out, const vector<X> *lhs,
+                       const vector<X> *rhs) {
+  for (int i = 0; i < X; i++) {
+    scalar_add(&out->v[i], &lhs->v[i], &rhs->v[i]);
+  }
+}
+
+template <int X>
+inline void vector_sub(vector<X> *out, const vector<X> *lhs,
+                       const vector<X> *rhs) {
+  for (int i = 0; i < X; i++) {
+    scalar_sub(&out->v[i], &lhs->v[i], &rhs->v[i]);
+  }
+}
+
+template <int X>
+inline void vector_mult_scalar_montgomery(vector<X> *out, const vector<X> *lhs,
+                                          const scalar *rhs) {
+  for (int i = 0; i < X; i++) {
+    scalar_mult_montgomery(&out->v[i], &lhs->v[i], rhs);
+  }
+}
+
+template <int X>
+inline void vector_ntt(vector<X> *a) {
+  for (int i = 0; i < X; i++) {
+    scalar_ntt(&a->v[i]);
+  }
+}
+
+template <int X>
+inline void vector_inverse_ntt_montgomery(vector<X> *a) {
+  for (int i = 0; i < X; i++) {
+    scalar_inverse_ntt_montgomery(&a->v[i]);
+  }
+}
+
+template <int K, int L>
+inline void matrix_mult_montgomery(vector<K> *out, const matrix<K, L> *m,
+                                   const vector<L> *a) {
+  vector_zero(out);
+  for (int i = 0; i < K; i++) {
+    for (int j = 0; j < L; j++) {
+      scalar product;
+      scalar_mult_montgomery(&product, &m->v[i][j], &a->v[j]);
+      scalar_add(&out->v[i], &out->v[i], &product);
+    }
+  }
+}
+
+/* Rounding & hints */
+
+// FIPS 204, Algorithm 35 (`Power2Round`).
+inline void power2_round(uint32_t *r1, uint32_t *r0, uint32_t r) {
+  *r1 = r >> kDroppedBits;
+  *r0 = r - (*r1 << kDroppedBits);
+
+  uint32_t r0_adjusted = mod_sub(*r0, 1 << kDroppedBits);
+  uint32_t r1_adjusted = *r1 + 1;
+
+  // Mask is set iff r0 > 2^(dropped_bits - 1).
+  crypto_word_t mask =
+      constant_time_lt_w((uint32_t)(1 << (kDroppedBits - 1)), *r0);
+  // r0 = mask ? r0_adjusted : r0
+  *r0 = constant_time_select_32(mask, r0_adjusted, *r0);
+  // r1 = mask ? r1_adjusted : r1
+  *r1 = constant_time_select_32(mask, r1_adjusted, *r1);
+}
+
+// Scale back previously rounded value.
+inline void scale_power2_round(uint32_t *out, uint32_t r1) {
+  // Pre-condition: 0 ≤ r1 ≤ 2^10 - 1
+  assert(r1 < (1u << 10));
+
+  *out = r1 << kDroppedBits;
+
+  // Post-condition: 0 ≤ out ≤ 2^23 - 2^13 = kPrime - 1
+  assert(*out < kPrime);
+}
+
+// FIPS 204, Algorithm 37 (`HighBits`).
+template <int K>
+inline uint32_t high_bits(uint32_t x) {
+  // Reference description (given 0 <= x < q):
+  //
+  // ```
+  // int32_t r0 = x mod+- (2 * gamma2);
+  // if (x - r0 == q - 1) {
+  //   return 0;
+  // } else {
+  //   return (x - r0) / (2 * gamma2);
+  // }
+  // ```
+  //
+  uint32_t r1 = (x + 127) >> 7;
+  if constexpr (prime_minus_one_over_gamma2<K>() == 32) {
+    // Below is the formula taken from the reference implementation.
+    //
+    // Here, Gamma2 == 2^18 - 2^8
+    // This returns ((ceil(x / 2^7) ⋅ (2^10 + 1) + 2^21) / 2^22) mod 2^4
+    r1 = (r1 * 1025 + (1 << 21)) >> 22;
+    r1 &= 15;
+  } else if constexpr (prime_minus_one_over_gamma2<K>() == 88) {
+    // 1488/2^24 is close enough to 1/1488 so that r1 becomes x/(2 gamma2)
+    // rounded down.
+    r1 = (r1 * 11275 + (1 << 23)) >> 24;
+
+    // For corner-case r1 = (Q-1)/(2 gamma2) = 44, we have to set r1=0.
+    r1 ^= ((uint32_t)(((int32_t)(43 - r1)) >> 31)) & r1;
+  }
+  return r1;
+}
+
+// FIPS 204, Algorithm 36 (`Decompose`).
+template <int K>
+inline void decompose(uint32_t *r1, int32_t *r0, uint32_t r) {
+  *r1 = high_bits<K>(r);
+
+  *r0 = r;
+  *r0 -= *r1 * 2 * (int32_t)gamma2<K>();
+  *r0 -= (((int32_t)kHalfPrime - *r0) >> 31) & (int32_t)kPrime;
+}
+
+// FIPS 204, Algorithm 38 (`LowBits`).
+template <int K>
+inline int32_t low_bits(uint32_t x) {
+  uint32_t r1;
+  int32_t r0;
+  decompose<K>(&r1, &r0, x);
+  return r0;
+}
+
+// FIPS 204, Algorithm 39 (`MakeHint`).
+//
+// In the spec this takes two arguments, z and r, and is called with
+//   z = -ct0
+//   r = w - cs2 + ct0
+//
+// It then computes HighBits (algorithm 37) of z and z+r. But z+r is just w -
+// cs2, so this takes three arguments and saves an addition.
+template <int K>
+inline int32_t make_hint(uint32_t ct0, uint32_t cs2, uint32_t w) {
+  uint32_t r_plus_z = mod_sub(w, cs2);
+  uint32_t r = reduce_once(r_plus_z + ct0);
+  return high_bits<K>(r) != high_bits<K>(r_plus_z);
+}
+
+// FIPS 204, Algorithm 40 (`UseHint`).
+template <int K>
+inline uint32_t use_hint_vartime(uint32_t h, uint32_t r) {
+  uint32_t r1;
+  int32_t r0;
+  decompose<K>(&r1, &r0, r);
+
+  if (h) {
+    if constexpr (prime_minus_one_over_gamma2<K>() == 32) {
+      if (r0 > 0) {
+        // (Q-1)/(2 gamma2) = m = 16, thus `mod m` in the spec turns into
+        // `& 15`.
+        return (r1 + 1) & 15;
+      } else {
+        return (r1 - 1) & 15;
+      }
+    } else {
+      // m = 44
+      static_assert(prime_minus_one_over_gamma2<K>() == 88);
+      if (r0 > 0) {
+        if (r1 == 43) {
+          return 0;
+        } else {
+          return r1 + 1;
+        }
+      } else {
+        if (r1 == 0) {
+          return 43;
+        } else {
+          return r1 - 1;
+        }
+      }
+    }
+  }
+  return r1;
+}
+
+inline void scalar_power2_round(scalar *s1, scalar *s0, const scalar *s) {
+  for (int i = 0; i < kDegree; i++) {
+    power2_round(&s1->c[i], &s0->c[i], s->c[i]);
+  }
+}
+
+inline void scalar_scale_power2_round(scalar *out, const scalar *in) {
+  for (int i = 0; i < kDegree; i++) {
+    scale_power2_round(&out->c[i], in->c[i]);
+  }
+}
+
+template <int K>
+inline void scalar_high_bits(scalar *out, const scalar *in) {
+  for (int i = 0; i < kDegree; i++) {
+    out->c[i] = high_bits<K>(in->c[i]);
+  }
+}
+
+template <int K>
+inline void scalar_low_bits(scalar *out, const scalar *in) {
+  for (int i = 0; i < kDegree; i++) {
+    out->c[i] = low_bits<K>(in->c[i]);
+  }
+}
+
+inline void scalar_max(uint32_t *max, const scalar *s) {
+  for (int i = 0; i < kDegree; i++) {
+    uint32_t abs = abs_mod_prime(s->c[i]);
+    *max = maximum_reduced(*max, abs);
+  }
+}
+
+inline void scalar_max_signed(uint32_t *max, const scalar *s) {
+  for (int i = 0; i < kDegree; i++) {
+    uint32_t abs = abs_signed(s->c[i]);
+    *max = maximum_reduced(*max, abs);
+  }
+}
+
+template <int K>
+inline void scalar_make_hint(scalar *out, const scalar *ct0, const scalar *cs2,
+                             const scalar *w) {
+  for (int i = 0; i < kDegree; i++) {
+    out->c[i] = make_hint<K>(ct0->c[i], cs2->c[i], w->c[i]);
+  }
+}
+
+template <int K>
+inline void scalar_use_hint_vartime(scalar *out, const scalar *h,
+                                    const scalar *r) {
+  for (int i = 0; i < kDegree; i++) {
+    out->c[i] = use_hint_vartime<K>(h->c[i], r->c[i]);
+  }
+}
+
+template <int X>
+inline void vector_power2_round(vector<X> *t1, vector<X> *t0,
+                                const vector<X> *t) {
+  for (int i = 0; i < X; i++) {
+    scalar_power2_round(&t1->v[i], &t0->v[i], &t->v[i]);
+  }
+}
+
+template <int X>
+inline void vector_scale_power2_round(vector<X> *out, const vector<X> *in) {
+  for (int i = 0; i < X; i++) {
+    scalar_scale_power2_round(&out->v[i], &in->v[i]);
+  }
+}
+
+template <int K>
+inline void vector_high_bits(vector<K> *out, const vector<K> *in) {
+  for (int i = 0; i < K; i++) {
+    scalar_high_bits<K>(&out->v[i], &in->v[i]);
+  }
+}
+
+template <int K>
+inline void vector_low_bits(vector<K> *out, const vector<K> *in) {
+  for (int i = 0; i < K; i++) {
+    scalar_low_bits<K>(&out->v[i], &in->v[i]);
+  }
+}
+
+template <int X>
+inline uint32_t vector_max(const vector<X> *a) {
+  uint32_t max = 0;
+  for (int i = 0; i < X; i++) {
+    scalar_max(&max, &a->v[i]);
+  }
+  return max;
+}
+
+template <int X>
+inline uint32_t vector_max_signed(const vector<X> *a) {
+  uint32_t max = 0;
+  for (int i = 0; i < X; i++) {
+    scalar_max_signed(&max, &a->v[i]);
+  }
+  return max;
+}
+
+// The input vector contains only zeroes and ones.
+template <int X>
+inline size_t vector_count_ones(const vector<X> *a) {
+  size_t count = 0;
+  for (int i = 0; i < X; i++) {
+    for (int j = 0; j < kDegree; j++) {
+      count += a->v[i].c[j];
+    }
+  }
+  return count;
+}
+
+template <int K>
+inline void vector_make_hint(vector<K> *out, const vector<K> *ct0,
+                             const vector<K> *cs2, const vector<K> *w) {
+  for (int i = 0; i < K; i++) {
+    scalar_make_hint<K>(&out->v[i], &ct0->v[i], &cs2->v[i], &w->v[i]);
+  }
+}
+
+template <int K>
+inline void vector_use_hint_vartime(vector<K> *out, const vector<K> *h,
+                                    const vector<K> *r) {
+  for (int i = 0; i < K; i++) {
+    scalar_use_hint_vartime<K>(&out->v[i], &h->v[i], &r->v[i]);
+  }
+}
+
+/* Bit packing */
+
+template <int bits>
+inline void scalar_encode(uint8_t *out, const scalar *s);
+
+// FIPS 204, Algorithm 16 (`SimpleBitPack`). Specialized to bitlen(b) = 4.
+template <>
+inline void scalar_encode<4>(uint8_t out[128], const scalar *s) {
+  // Every two elements lands on a byte boundary.
+  static_assert(kDegree % 2 == 0, "kDegree must be a multiple of 2");
+  for (int i = 0; i < kDegree / 2; i++) {
+    uint32_t a = s->c[2 * i];
+    uint32_t b = s->c[2 * i + 1];
+    declassify_assert(a < 16);
+    declassify_assert(b < 16);
+    out[i] = a | (b << 4);
+  }
+}
+
+// FIPS 204, Algorithm 16 (`SimpleBitPack`). Specialized to bitlen(b) = 6.
+template <>
+inline void scalar_encode<6>(uint8_t out[192], const scalar *s) {
+  // Every four elements lands on a byte boundary.
+  static_assert(kDegree % 4 == 0, "kDegree must be a multiple of 4");
+  for (int i = 0; i < kDegree / 4; i++) {
+    uint32_t a = s->c[4 * i];
+    uint32_t b = s->c[4 * i + 1];
+    uint32_t c = s->c[4 * i + 2];
+    uint32_t d = s->c[4 * i + 3];
+    declassify_assert(a < 64);
+    declassify_assert(b < 64);
+    declassify_assert(c < 64);
+    declassify_assert(d < 64);
+    out[3 * i] = a | (b << 6);
+    out[3 * i + 1] = (b >> 2) | (c << 4);
+    out[3 * i + 2] = (c >> 4) | (d << 2);
+  }
+}
+
+// FIPS 204, Algorithm 16 (`SimpleBitPack`). Specialized to bitlen(b) = 10.
+template <>
+inline void scalar_encode<10>(uint8_t out[320], const scalar *s) {
+  // Every four elements lands on a byte boundary.
+  static_assert(kDegree % 4 == 0, "kDegree must be a multiple of 4");
+  for (int i = 0; i < kDegree / 4; i++) {
+    uint32_t a = s->c[4 * i];
+    uint32_t b = s->c[4 * i + 1];
+    uint32_t c = s->c[4 * i + 2];
+    uint32_t d = s->c[4 * i + 3];
+    declassify_assert(a < 1024);
+    declassify_assert(b < 1024);
+    declassify_assert(c < 1024);
+    declassify_assert(d < 1024);
+    out[5 * i] = (uint8_t)a;
+    out[5 * i + 1] = (uint8_t)((a >> 8) | (b << 2));
+    out[5 * i + 2] = (uint8_t)((b >> 6) | (c << 4));
+    out[5 * i + 3] = (uint8_t)((c >> 4) | (d << 6));
+    out[5 * i + 4] = (uint8_t)(d >> 2);
+  }
+}
+
+template <int bits, uint32_t max>
+inline void scalar_encode_signed(uint8_t *out, const scalar *s);
+
+// FIPS 204, Algorithm 17 (`BitPack`). Specialized to bitlen(a+b) = 4 and b = 4.
+template <>
+inline void scalar_encode_signed<4, 4>(uint8_t out[128], const scalar *s) {
+  // Every two elements lands on a byte boundary.
+  static_assert(kDegree % 2 == 0, "kDegree must be a multiple of 2");
+  for (int i = 0; i < kDegree / 2; i++) {
+    uint32_t a = mod_sub(4, s->c[2 * i]);
+    uint32_t b = mod_sub(4, s->c[2 * i + 1]);
+    declassify_assert(a < 16);
+    declassify_assert(b < 16);
+    out[i] = a | (b << 4);
+  }
+}
+
+// FIPS 204, Algorithm 17 (`BitPack`). Specialized to bitlen(a+b) = 3 and b = 2.
+template <>
+inline void scalar_encode_signed<3, 2>(uint8_t out[96], const scalar *s) {
+  static_assert(kDegree % 8 == 0, "kDegree must be a multiple of 8");
+  for (int i = 0; i < kDegree / 8; i++) {
+    uint32_t a = mod_sub(2, s->c[8 * i]);
+    uint32_t b = mod_sub(2, s->c[8 * i + 1]);
+    uint32_t c = mod_sub(2, s->c[8 * i + 2]);
+    uint32_t d = mod_sub(2, s->c[8 * i + 3]);
+    uint32_t e = mod_sub(2, s->c[8 * i + 4]);
+    uint32_t f = mod_sub(2, s->c[8 * i + 5]);
+    uint32_t g = mod_sub(2, s->c[8 * i + 6]);
+    uint32_t h = mod_sub(2, s->c[8 * i + 7]);
+    uint32_t v = (h << 21) | (g << 18) | (f << 15) | (e << 12) | (d << 9) |
+                 (c << 6) | (b << 3) | a;
+    uint8_t v_bytes[sizeof(v)];
+    CRYPTO_store_u32_le(v_bytes, v);
+    OPENSSL_memcpy(&out[i * 3], v_bytes, 3);
+  }
+}
+
+// FIPS 204, Algorithm 17 (`BitPack`). Specialized to bitlen(a+b) = 13 and b =
+// 2^12.
+template <>
+inline void scalar_encode_signed<13, (1u << 12)>(uint8_t out[416],
+                                                 const scalar *s) {
+  static const uint32_t kMax = 1u << 12;
+  // Every two elements lands on a byte boundary.
+  static_assert(kDegree % 8 == 0, "kDegree must be a multiple of 8");
+  for (int i = 0; i < kDegree / 8; i++) {
+    uint32_t a = mod_sub(kMax, s->c[8 * i]);
+    uint32_t b = mod_sub(kMax, s->c[8 * i + 1]);
+    uint32_t c = mod_sub(kMax, s->c[8 * i + 2]);
+    uint32_t d = mod_sub(kMax, s->c[8 * i + 3]);
+    uint32_t e = mod_sub(kMax, s->c[8 * i + 4]);
+    uint32_t f = mod_sub(kMax, s->c[8 * i + 5]);
+    uint32_t g = mod_sub(kMax, s->c[8 * i + 6]);
+    uint32_t h = mod_sub(kMax, s->c[8 * i + 7]);
+    declassify_assert(a < (1u << 13));
+    declassify_assert(b < (1u << 13));
+    declassify_assert(c < (1u << 13));
+    declassify_assert(d < (1u << 13));
+    declassify_assert(e < (1u << 13));
+    declassify_assert(f < (1u << 13));
+    declassify_assert(g < (1u << 13));
+    declassify_assert(h < (1u << 13));
+    a |= b << 13;
+    a |= c << 26;
+    c >>= 6;
+    c |= d << 7;
+    c |= e << 20;
+    e >>= 12;
+    e |= f << 1;
+    e |= g << 14;
+    e |= h << 27;
+    h >>= 5;
+    CRYPTO_store_u32_le(&out[13 * i], a);
+    CRYPTO_store_u32_le(&out[13 * i + 4], c);
+    CRYPTO_store_u32_le(&out[13 * i + 8], e);
+    out[13 * i + 12] = static_cast<uint8_t>(h);
+  }
+}
+
+// FIPS 204, Algorithm 17 (`BitPack`). Specialized to bitlen(a+b) = 20 and b =
+// 2^19.
+template <>
+inline void scalar_encode_signed<20, (1u << 19)>(uint8_t out[640],
+                                                 const scalar *s) {
+  static const uint32_t kMax = 1u << 19;
+  // Every two elements lands on a byte boundary.
+  static_assert(kDegree % 4 == 0, "kDegree must be a multiple of 4");
+  for (int i = 0; i < kDegree / 4; i++) {
+    uint32_t a = mod_sub(kMax, s->c[4 * i]);
+    uint32_t b = mod_sub(kMax, s->c[4 * i + 1]);
+    uint32_t c = mod_sub(kMax, s->c[4 * i + 2]);
+    uint32_t d = mod_sub(kMax, s->c[4 * i + 3]);
+    declassify_assert(a < (1u << 20));
+    declassify_assert(b < (1u << 20));
+    declassify_assert(c < (1u << 20));
+    declassify_assert(d < (1u << 20));
+    a |= b << 20;
+    b >>= 12;
+    b |= c << 8;
+    b |= d << 28;
+    d >>= 4;
+    CRYPTO_store_u32_le(&out[10 * i], a);
+    CRYPTO_store_u32_le(&out[10 * i + 4], b);
+    CRYPTO_store_u16_le(&out[10 * i + 8], static_cast<uint16_t>(d));
+  }
+}
+
+// FIPS 204, Algorithm 17 (`BitPack`). Specialized to bitlen(a+b) = 18 and b =
+// 2^17.
+template <>
+inline void scalar_encode_signed<18, (1u << 17)>(uint8_t out[576],
+                                                 const scalar *s) {
+  static const uint32_t kMax = 1u << 17;
+  static_assert(kDegree % 4 == 0, "kDegree must be a multiple of 4");
+  for (int i = 0; i < kDegree / 4; i++) {
+    uint32_t a = mod_sub(kMax, s->c[4 * i]);
+    uint32_t b = mod_sub(kMax, s->c[4 * i + 1]);
+    uint32_t c = mod_sub(kMax, s->c[4 * i + 2]);
+    uint32_t d = mod_sub(kMax, s->c[4 * i + 3]);
+    declassify_assert(a < (1u << 18));
+    declassify_assert(b < (1u << 18));
+    declassify_assert(c < (1u << 18));
+    declassify_assert(d < (1u << 18));
+    out[9 * i] = (uint8_t)a;
+    out[9 * i + 1] = (uint8_t)(a >> 8);
+    out[9 * i + 2] = (uint8_t)(a >> 16) | (uint8_t)(b << 2);
+    out[9 * i + 3] = (uint8_t)(b >> 6);
+    out[9 * i + 4] = (uint8_t)(b >> 14) | (uint8_t)(c << 4);
+    out[9 * i + 5] = (uint8_t)(c >> 4);
+    out[9 * i + 6] = (uint8_t)(c >> 12) | (uint8_t)(d << 6);
+    out[9 * i + 7] = (uint8_t)(d >> 2);
+    out[9 * i + 8] = (uint8_t)(d >> 10);
+  }
+}
+
+template <int bits>
+inline void scalar_decode(scalar *out, const uint8_t *in);
+
+// FIPS 204, Algorithm 18 (`SimpleBitUnpack`). Specialized for bitlen(b) == 10.
+template <>
+inline void scalar_decode<10>(scalar *out, const uint8_t in[320]) {
+  static_assert(kDegree % 4 == 0, "kDegree must be a multiple of 4");
+  for (int i = 0; i < kDegree / 4; i++) {
+    uint32_t v = CRYPTO_load_u32_le(&in[5 * i]);
+    out->c[4 * i] = v & 0x3ff;
+    out->c[4 * i + 1] = (v >> 10) & 0x3ff;
+    out->c[4 * i + 2] = (v >> 20) & 0x3ff;
+    out->c[4 * i + 3] = (v >> 30) | (((uint32_t)in[5 * i + 4]) << 2);
+  }
+}
+
+template <int bits, uint32_t max>
+inline int scalar_decode_signed(scalar *out, const uint8_t *in);
+
+// FIPS 204, Algorithm 19 (`BitUnpack`). Specialized to bitlen(a+b) = 4 and b =
+// 4.
+template <>
+inline int scalar_decode_signed<4, 4>(scalar *out, const uint8_t in[128]) {
+  static_assert(kDegree % 8 == 0, "kDegree must be a multiple of 8");
+  for (int i = 0; i < kDegree / 8; i++) {
+    uint32_t v = CRYPTO_load_u32_le(&in[4 * i]);
+    // None of the nibbles may be >= 9. So if the MSB of any nibble is set, none
+    // of the other bits may be set. First, select all the MSBs.
+    const uint32_t msbs = v & 0x88888888u;
+    // For each nibble where the MSB is set, form a mask of all the other bits.
+    const uint32_t mask = (msbs >> 1) | (msbs >> 2) | (msbs >> 3);
+    // A nibble is only out of range in the case of invalid input, in which case
+    // it is okay to leak the value.
+    if (constant_time_declassify_int((mask & v) != 0)) {
+      return 0;
+    }
+
+    out->c[i * 8] = mod_sub(4, v & 15);
+    out->c[i * 8 + 1] = mod_sub(4, (v >> 4) & 15);
+    out->c[i * 8 + 2] = mod_sub(4, (v >> 8) & 15);
+    out->c[i * 8 + 3] = mod_sub(4, (v >> 12) & 15);
+    out->c[i * 8 + 4] = mod_sub(4, (v >> 16) & 15);
+    out->c[i * 8 + 5] = mod_sub(4, (v >> 20) & 15);
+    out->c[i * 8 + 6] = mod_sub(4, (v >> 24) & 15);
+    out->c[i * 8 + 7] = mod_sub(4, v >> 28);
+  }
+  return 1;
+}
+
+// FIPS 204, Algorithm 19 (`BitUnpack`). Specialized to bitlen(a+b) = 3 and b =
+// 2.
+template <>
+inline int scalar_decode_signed<3, 2>(scalar *out, const uint8_t in[96]) {
+  uint32_t v;
+  uint8_t v_bytes[sizeof(v)] = {0};
+  static_assert(kDegree % 8 == 0, "kDegree must be a multiple of 8");
+  for (int i = 0; i < kDegree / 8; i++) {
+    OPENSSL_memcpy(v_bytes, &in[3 * i], 3);
+    v = CRYPTO_load_u32_le(v_bytes);
+    // v contains 8, 3-bit values in the lower 24 bits. None of the values may
+    // be >= 5. So if the MSB of any triple is set, none of the other bits may
+    // be set. First, select all the MSBs.
+    const uint32_t msbs = v & 000044444444u;
+    // For each triple where the MSB is set, form a mask of all the other bits.
+    const uint32_t mask = (msbs >> 1) | (msbs >> 2);
+    // A triple is only out of range in the case of invalid input, in which case
+    // it is okay to leak the value.
+    if (constant_time_declassify_int((mask & v) != 0)) {
+      return 0;
+    }
+
+    out->c[i * 8 + 0] = mod_sub(2, (v >> 0) & 7);
+    out->c[i * 8 + 1] = mod_sub(2, (v >> 3) & 7);
+    out->c[i * 8 + 2] = mod_sub(2, (v >> 6) & 7);
+    out->c[i * 8 + 3] = mod_sub(2, (v >> 9) & 7);
+    out->c[i * 8 + 4] = mod_sub(2, (v >> 12) & 7);
+    out->c[i * 8 + 5] = mod_sub(2, (v >> 15) & 7);
+    out->c[i * 8 + 6] = mod_sub(2, (v >> 18) & 7);
+    out->c[i * 8 + 7] = mod_sub(2, v >> 21);
+  }
+  return 1;
+}
+
+// FIPS 204, Algorithm 19 (`BitUnpack`). Specialized to bitlen(a+b) = 13 and b =
+// 2^12.
+template <>
+inline int scalar_decode_signed<13, (1u << 12)>(scalar *out,
+                                                const uint8_t in[416]) {
+  static const uint32_t kMax = 1u << 12;
+  static const uint32_t k13Bits = (1u << 13) - 1;
+  static const uint32_t k7Bits = (1u << 7) - 1;
+
+  static_assert(kDegree % 8 == 0, "kDegree must be a multiple of 8");
+  for (int i = 0; i < kDegree / 8; i++) {
+    uint32_t a = CRYPTO_load_u32_le(&in[13 * i]);
+    uint32_t b = CRYPTO_load_u32_le(&in[13 * i + 4]);
+    uint32_t c = CRYPTO_load_u32_le(&in[13 * i + 8]);
+    uint8_t d = in[13 * i + 12];
+
+    // It's not possible for a 13-bit number to be out of range when the max is
+    // 2^12.
+    out->c[i * 8] = mod_sub(kMax, a & k13Bits);
+    out->c[i * 8 + 1] = mod_sub(kMax, (a >> 13) & k13Bits);
+    out->c[i * 8 + 2] = mod_sub(kMax, (a >> 26) | ((b & k7Bits) << 6));
+    out->c[i * 8 + 3] = mod_sub(kMax, (b >> 7) & k13Bits);
+    out->c[i * 8 + 4] = mod_sub(kMax, (b >> 20) | ((c & 1) << 12));
+    out->c[i * 8 + 5] = mod_sub(kMax, (c >> 1) & k13Bits);
+    out->c[i * 8 + 6] = mod_sub(kMax, (c >> 14) & k13Bits);
+    out->c[i * 8 + 7] = mod_sub(kMax, (c >> 27) | ((uint32_t)d) << 5);
+  }
+  return 1;
+}
+
+// FIPS 204, Algorithm 19 (`BitUnpack`). Specialized to bitlen(a+b) = 18 and b =
+// 2^17.
+template <>
+inline int scalar_decode_signed<18, (1u << 17)>(scalar *out,
+                                                const uint8_t in[576]) {
+  static const uint32_t kMax = 1u << 17;
+
+  static_assert(kDegree % 4 == 0, "kDegree must be a multiple of 4");
+  for (int i = 0; i < kDegree / 4; i++) {
+    uint32_t a = uint32_t{in[9 * i]} | (uint32_t{in[9 * i + 1]} << 8) |
+                 ((uint32_t{in[9 * i + 2]} & 0x3) << 16);
+    uint32_t b = (uint32_t{in[9 * i + 2]} >> 2) |
+                 (uint32_t{in[9 * i + 3]} << 6) |
+                 ((uint32_t{in[9 * i + 4]} & 0xf) << 14);
+    uint32_t c = (uint32_t{in[9 * i + 4]} >> 4) |
+                 (uint32_t{in[9 * i + 5]} << 4) |
+                 ((uint32_t{in[9 * i + 6]} & 0x3f) << 12);
+    uint32_t d = (uint32_t{in[9 * i + 6]} >> 6) |
+                 (uint32_t{in[9 * i + 7]} << 2) |
+                 (uint32_t{in[9 * i + 8]} << 10);
+
+    out->c[i * 4] = mod_sub(kMax, a);
+    out->c[i * 4 + 1] = mod_sub(kMax, b);
+    out->c[i * 4 + 2] = mod_sub(kMax, c);
+    out->c[i * 4 + 3] = mod_sub(kMax, d);
+  }
+  return 1;
+}
+
+// FIPS 204, Algorithm 19 (`BitUnpack`). Specialized to bitlen(a+b) = 20 and b =
+// 2^19.
+template <>
+inline int scalar_decode_signed<20, (1u << 19)>(scalar *out,
+                                                const uint8_t in[640]) {
+  static const uint32_t kMax = 1u << 19;
+  static const uint32_t k20Bits = (1u << 20) - 1;
+
+  static_assert(kDegree % 4 == 0, "kDegree must be a multiple of 4");
+  for (int i = 0; i < kDegree / 4; i++) {
+    uint32_t a = CRYPTO_load_u32_le(&in[10 * i]);
+    uint32_t b = CRYPTO_load_u32_le(&in[10 * i + 4]);
+    uint16_t c = CRYPTO_load_u16_le(&in[10 * i + 8]);
+
+    // It's not possible for a 20-bit number to be out of range when the max is
+    // 2^19.
+    out->c[i * 4] = mod_sub(kMax, a & k20Bits);
+    out->c[i * 4 + 1] = mod_sub(kMax, (a >> 20) | ((b & 0xff) << 12));
+    out->c[i * 4 + 2] = mod_sub(kMax, (b >> 8) & k20Bits);
+    out->c[i * 4 + 3] = mod_sub(kMax, (b >> 28) | ((uint32_t)c) << 4);
+  }
+  return 1;
+}
+
+
+
+/* Expansion functions */
+
+// FIPS 204, Algorithm 30 (`RejNTTPoly`).
+//
+// Rejection samples a Keccak stream to get uniformly distributed elements. This
+// is used for matrix expansion and only operates on public inputs.
+inline void scalar_from_keccak_block_vartime(scalar *out, int *done,
+                                             const uint8_t block[168]) {
+  // kStep is the number of loop iterations to do at once. May need to become
+  // platform dependent at some point. Must be a compile-time constant - if we
+  // ever need to runtime dispatch on it, it has to be made a template argument.
+  const size_t kStep = 2;
+
+  // Branchless sampling as long as there is enough space in the output buffer
+  // even if all sampled numbers are < `kPrime`.
+  size_t i = 0;
+  for (; i <= size_t{168 - 3 * kStep} && *done <= int{kDegree - kStep};
+       i += 3 * kStep) {
+    uint32_t d[kStep];
+    for (size_t k = 0; k < kStep; ++k) {
+      size_t bi = i + 3 * k;
+      d[k] = (uint32_t)block[bi] | ((uint32_t)block[bi + 1] << 8) |
+             (((uint32_t)block[bi + 2] & 0x7f) << 16);
+    }
+    for (size_t k = 0; k < kStep; ++k) {
+      out->c[*done] = d[k];
+      *done += d[k] < kPrime;
+    }
+  }
+
+  // Branchful "tail".
+  for (; i < 168 && *done < kDegree; i += 3) {
+    // FIPS 204, Algorithm 14 (`CoeffFromThreeBytes`).
+    uint32_t value = (uint32_t)block[i] | ((uint32_t)block[i + 1] << 8) |
+                     (((uint32_t)block[i + 2] & 0x7f) << 16);
+    if (value < kPrime) {
+      out->c[(*done)++] = value;
+    }
+  }
+}
+
+// FIPS 204, Algorithm 30 (`RejNTTPoly`).
+//
+// Rejection samples a Keccak stream to get uniformly distributed elements. This
+// is used for matrix expansion and only operates on public inputs.
+inline void scalar_from_keccak_vartime(scalar *out,
+                                       BORINGSSL_keccak_st *keccak_ctx) {
+  assert(keccak_ctx->squeeze_offset == 0);
+  assert(keccak_ctx->rate_bytes == 168);
+  static_assert(168 % 3 == 0, "block and coefficient boundaries do not align");
+
+  int done = 0;
+  while (done < kDegree) {
+    uint8_t block[168];
+    BORINGSSL_keccak_squeeze(keccak_ctx, block, sizeof(block));
+    scalar_from_keccak_block_vartime(out, &done, block);
+  }
+}
+
+#if defined(HAVE_KECCAK_X2)
+inline void scalar_from_keccak_vartime_x2(scalar *out[2],
+                                          BORINGSSL_keccak_st keccak_ctx[2]) {
+  assert(keccak_ctx[0].squeeze_offset == 0);
+  assert(keccak_ctx[0].rate_bytes == 168);
+  assert(keccak_ctx[1].squeeze_offset == 0);
+  assert(keccak_ctx[1].rate_bytes == 168);
+  static_assert(168 % 3 == 0, "block and coefficient boundaries do not align");
+
+  int done[2] = {0, 0};
+  while (done[0] < kDegree || done[1] < kDegree) {
+    uint8_t block[2][168];
+    uint8_t *blocks[] = {block[0], block[1]};
+    BORINGSSL_keccak_squeeze_x2(keccak_ctx, blocks, sizeof(block[0]));
+    scalar_from_keccak_block_vartime(out[0], &done[0], block[0]);
+    scalar_from_keccak_block_vartime(out[1], &done[1], block[1]);
+  }
+}
+#endif
+
+#if defined(HAVE_KECCAK_X4)
+inline void scalar_from_keccak_vartime_x4(scalar *out[4],
+                                          BORINGSSL_keccak_st keccak_ctx[]) {
+  for (int i = 0; i < 4; i++) {
+    assert(keccak_ctx[i].squeeze_offset == 0);
+    assert(keccak_ctx[i].rate_bytes == 168);
+  }
+  static_assert(168 % 3 == 0, "block and coefficient boundaries do not align");
+
+  int done[4] = {0, 0, 0, 0};
+  while (done[0] < kDegree || done[1] < kDegree || done[2] < kDegree ||
+         done[3] < kDegree) {
+    uint8_t block[4][168];
+    uint8_t *blocks[] = {block[0], block[1], block[2], block[3]};
+    BORINGSSL_keccak_squeeze_x4(keccak_ctx, blocks, sizeof(block[0]));
+    scalar_from_keccak_block_vartime(out[0], &done[0], block[0]);
+    scalar_from_keccak_block_vartime(out[1], &done[1], block[1]);
+    scalar_from_keccak_block_vartime(out[2], &done[2], block[2]);
+    scalar_from_keccak_block_vartime(out[3], &done[3], block[3]);
+  }
+}
+#endif
+
+template <int ETA>
+inline bool coefficient_from_nibble(uint32_t nibble, uint32_t *result);
+
+template <>
+inline bool coefficient_from_nibble<4>(uint32_t nibble, uint32_t *result) {
+  // Knowing bounds on `nibble` seems to tempt some versions of Clang to emit
+  // a branch, if we don't have a barrier in `mod_sub`.
+  *result = mod_sub(4, value_barrier_u32(nibble));
+  return constant_time_declassify_int(nibble < 9);
+}
+
+template <>
+inline bool coefficient_from_nibble<2>(uint32_t nibble, uint32_t *result) {
+  // Knowing bounds on `nibble` seems to tempt some versions of Clang to emit
+  // a branch, if we don't have a barrier in `mod_sub`.
+  // Constant time "nibble % 5".
+  uint32_t rem = nibble - 5 * ((205 * nibble) >> 10);
+  *result = mod_sub(2, value_barrier_u32(rem));
+  return constant_time_declassify_int(nibble < 15);
+}
+
+template <int ETA>
+inline void scalar_uniform_block(scalar *out, int *done,
+                                 const uint8_t block[136]) {
+  for (size_t i = 0; i < 136 && *done < kDegree; ++i) {
+    uint32_t t0 = block[i] & 0x0F;
+    uint32_t t1 = block[i] >> 4;
+    // FIPS 204, Algorithm 15 (`CoefFromHalfByte`). Although both the input
+    // and output here are secret, it is OK to leak when we rejected a byte.
+    // Individual bytes of the SHAKE-256 stream are (indistinguishable from)
+    // independent of each other and the original seed, so leaking information
+    // about the rejected bytes does not reveal the input or output.
+    *done += coefficient_from_nibble<ETA>(t0, &out->c[*done]);
+    if (*done < kDegree) {
+      *done += coefficient_from_nibble<ETA>(t1, &out->c[*done]);
+    }
+  }
+}
+
+// FIPS 204, Algorithm 31 (`RejBoundedPoly`).
+template <int ETA>
+inline void scalar_uniform(scalar *out, BORINGSSL_keccak_st *keccak_ctx) {
+  assert(keccak_ctx->squeeze_offset == 0);
+  assert(keccak_ctx->rate_bytes == 136);
+
+  int done = 0;
+  while (done < kDegree) {
+    uint8_t block[136];
+    BORINGSSL_keccak_squeeze(keccak_ctx, block, sizeof(block));
+    scalar_uniform_block<ETA>(out, &done, block);
+  }
+}
+
+#if defined(HAVE_KECCAK_X2)
+template <int ETA>
+inline void scalar_uniform_x2(scalar *out[2],
+                              BORINGSSL_keccak_st keccak_ctx[2]) {
+  assert(keccak_ctx[0].squeeze_offset == 0);
+  assert(keccak_ctx[0].rate_bytes == 136);
+  assert(keccak_ctx[1].squeeze_offset == 0);
+  assert(keccak_ctx[1].rate_bytes == 136);
+
+  int done[2] = {0, 0};
+  while (done[0] < kDegree || done[1] < kDegree) {
+    uint8_t block[2][136];
+    uint8_t *blocks[] = {block[0], block[1]};
+    BORINGSSL_keccak_squeeze_x2(keccak_ctx, blocks, sizeof(block[0]));
+    scalar_uniform_block<ETA>(out[0], &done[0], block[0]);
+    scalar_uniform_block<ETA>(out[1], &done[1], block[1]);
+  }
+}
+#endif
+
+#if defined(HAVE_KECCAK_X4)
+template <int ETA>
+inline void scalar_uniform_x4(scalar *out[4],
+                              BORINGSSL_keccak_st keccak_ctx[]) {
+  for (int i = 0; i < 4; i++) {
+    assert(keccak_ctx[i].squeeze_offset == 0);
+    assert(keccak_ctx[i].rate_bytes == 136);
+  }
+
+  int done[4] = {0, 0, 0, 0};
+  while (done[0] < kDegree || done[1] < kDegree || done[2] < kDegree ||
+         done[3] < kDegree) {
+    uint8_t block[4][136];
+    uint8_t *blocks[] = {block[0], block[1], block[2], block[3]};
+    BORINGSSL_keccak_squeeze_x4(keccak_ctx, blocks, sizeof(block[0]));
+    scalar_uniform_block<ETA>(out[0], &done[0], block[0]);
+    scalar_uniform_block<ETA>(out[1], &done[1], block[1]);
+    scalar_uniform_block<ETA>(out[2], &done[2], block[2]);
+    scalar_uniform_block<ETA>(out[3], &done[3], block[3]);
+  }
+}
+#endif
+
+// FIPS 204, Algorithm 34 (`ExpandMask`), but just a single step.
+template <int K>
+inline void scalar_sample_mask(scalar *out,
+                               const uint8_t rho_prime[kRhoPrimeBytes],
+                               size_t index) {
+  uint8_t buf[scalar_le_gamma1_bytes<K>()];
+  struct BORINGSSL_keccak_st ctx;
+  BORINGSSL_keccak_init(&ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&ctx, rho_prime, kRhoPrimeBytes);
+  uint8_t index_buf[2];
+  index_buf[0] = static_cast<uint8_t>(index);
+  index_buf[1] = static_cast<uint8_t>(index >> 8);
+  BORINGSSL_keccak_absorb(&ctx, index_buf, 2);
+  BORINGSSL_keccak_squeeze(&ctx, buf, sizeof(buf));
+  scalar_decode_signed<gamma1_bits<K>() + 1, gamma1<K>()>(out, buf);
+}
+
+// FIPS 204, Algorithm 29 (`SampleInBall`).
+inline void scalar_sample_in_ball_vartime(scalar *out, const uint8_t *seed,
+                                          int len, int tau) {
+  BORINGSSL_keccak_st keccak_ctx;
+  BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&keccak_ctx, seed, len);
+  assert(keccak_ctx.squeeze_offset == 0);
+  assert(keccak_ctx.rate_bytes == 136);
+
+  uint8_t block[136];
+  BORINGSSL_keccak_squeeze(&keccak_ctx, block, sizeof(block));
+
+  uint64_t signs = CRYPTO_load_u64_le(block);
+  int offset = 8;
+  // SampleInBall implements a Fisher–Yates shuffle, which unavoidably leaks
+  // where the zeros are by memory access pattern. Although this leak happens
+  // before bad signatures are rejected, this is safe. See
+  // https://boringssl-review.googlesource.com/c/boringssl/+/67747/comment/8d8f01ac_70af3f21/
+  CONSTTIME_DECLASSIFY(block + offset, sizeof(block) - offset);
+
+  OPENSSL_memset(out, 0, sizeof(*out));
+  for (size_t i = kDegree - tau; i < kDegree; i++) {
+    size_t byte;
+    for (;;) {
+      if (offset == 136) {
+        BORINGSSL_keccak_squeeze(&keccak_ctx, block, sizeof(block));
+        // See above.
+        CONSTTIME_DECLASSIFY(block, sizeof(block));
+        offset = 0;
+      }
+
+      byte = block[offset++];
+      if (byte <= i) {
+        break;
+      }
+    }
+
+    out->c[i] = out->c[byte];
+    out->c[byte] = mod_sub(1, 2 * (signs & 1));
+    signs >>= 1;
+  }
+}
+
+template <int K, int L>
+inline void matrix_expand_step(scalar *out, const uint8_t rho[kRhoBytes],
+                               int index) {
+  BORINGSSL_keccak_st keccak_ctx;
+  BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake128);
+  BORINGSSL_keccak_absorb(&keccak_ctx, rho, kRhoBytes);
+  uint8_t index_buf[2];
+  index_buf[0] = static_cast<uint8_t>(index % L);
+  index_buf[1] = static_cast<uint8_t>(index / L);
+  BORINGSSL_keccak_absorb(&keccak_ctx, index_buf, 2);
+  scalar_from_keccak_vartime(out, &keccak_ctx);
+}
+
+#if defined(HAVE_KECCAK_X2)
+template <int K, int L>
+inline void matrix_expand_step_x2(matrix<K, L> *out,
+                                  const uint8_t rho[kRhoBytes], int index) {
+  BORINGSSL_keccak_st keccak_ctx[2];
+  BORINGSSL_keccak_init(&keccak_ctx[0], boringssl_shake128);
+  BORINGSSL_keccak_init(&keccak_ctx[1], boringssl_shake128);
+  BORINGSSL_keccak_absorb(&keccak_ctx[0], rho, kRhoBytes);
+  BORINGSSL_keccak_absorb(&keccak_ctx[1], rho, kRhoBytes);
+  uint8_t index_buf[2];
+  index_buf[0] = static_cast<uint8_t>(index % L);
+  index_buf[1] = static_cast<uint8_t>(index / L);
+  BORINGSSL_keccak_absorb(&keccak_ctx[0], index_buf, 2);
+  index_buf[0] = static_cast<uint8_t>((index + 1) % L);
+  index_buf[1] = static_cast<uint8_t>((index + 1) / L);
+  BORINGSSL_keccak_absorb(&keccak_ctx[1], index_buf, 2);
+
+  scalar *outs[] = {&out->v[index / L][index % L],
+                    &out->v[(index + 1) / L][(index + 1) % L]};
+  scalar_from_keccak_vartime_x2(outs, keccak_ctx);
+}
+#endif
+
+#if defined(HAVE_KECCAK_X4)
+template <int K, int L>
+inline void matrix_expand_step_x4(matrix<K, L> *out,
+                                  const uint8_t rho[kRhoBytes], int index) {
+  BORINGSSL_keccak_st keccak_ctx[4];
+  for (int k = 0; k < 4; k++) {
+    BORINGSSL_keccak_init(&keccak_ctx[k], boringssl_shake128);
+    BORINGSSL_keccak_absorb(&keccak_ctx[k], rho, kRhoBytes);
+    uint8_t index_buf[2];
+    index_buf[0] = static_cast<uint8_t>((index + k) % L);
+    index_buf[1] = static_cast<uint8_t>((index + k) / L);
+    BORINGSSL_keccak_absorb(&keccak_ctx[k], index_buf, 2);
+  }
+
+  scalar *outs[] = {&out->v[index / L][index % L],
+                    &out->v[(index + 1) / L][(index + 1) % L],
+                    &out->v[(index + 2) / L][(index + 2) % L],
+                    &out->v[(index + 3) / L][(index + 3) % L]};
+  scalar_from_keccak_vartime_x4(outs, keccak_ctx);
+}
+#endif
+
+template <int K, int L>
+inline void matrix_expand(matrix<K, L> *out, const uint8_t rho[kRhoBytes]) {
+  static_assert(K <= 0x100, "K must fit in 8 bits");
+  static_assert(L <= 0x100, "L must fit in 8 bits");
+
+  int index = 0;
+#if defined(HAVE_KECCAK_X4)
+  if (BORINGSSL_have_keccak_x4()) {
+    for (; index + 4 <= K * L; index += 4) {
+      matrix_expand_step_x4<K, L>(out, rho, index);
+    }
+  }
+#endif
+#if defined(HAVE_KECCAK_X2)
+  for (; index + 2 <= K * L; index += 2) {
+    matrix_expand_step_x2(out, rho, index);
+  }
+#endif
+
+  for (; index < K * L; index++) {
+    matrix_expand_step<K, L>(&out->v[index / L][index % L], rho, index);
+  }
+}
+
+template <int K>
+inline void vector_expand_short_step(scalar *out,
+                                     const uint8_t sigma[kSigmaBytes],
+                                     int index) {
+  BORINGSSL_keccak_st keccak_ctx;
+  BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&keccak_ctx, sigma, kSigmaBytes);
+  uint8_t index_buf[2];
+  index_buf[0] = static_cast<uint8_t>(index);
+  index_buf[1] = 0;
+  BORINGSSL_keccak_absorb(&keccak_ctx, index_buf, 2);
+  scalar_uniform<eta<K>()>(out, &keccak_ctx);
+}
+
+#if defined(HAVE_KECCAK_X2)
+template <int K>
+inline void vector_expand_short_step_x2(scalar *out0, scalar *out1,
+                                        const uint8_t sigma[kSigmaBytes],
+                                        int index) {
+  BORINGSSL_keccak_st keccak_ctx[2];
+  BORINGSSL_keccak_init(&keccak_ctx[0], boringssl_shake256);
+  BORINGSSL_keccak_init(&keccak_ctx[1], boringssl_shake256);
+  BORINGSSL_keccak_absorb(&keccak_ctx[0], sigma, kSigmaBytes);
+  BORINGSSL_keccak_absorb(&keccak_ctx[1], sigma, kSigmaBytes);
+  uint8_t index_buf[2];
+  index_buf[0] = static_cast<uint8_t>(index);
+  index_buf[1] = 0;
+  BORINGSSL_keccak_absorb(&keccak_ctx[0], index_buf, 2);
+  index_buf[0] = static_cast<uint8_t>(index + 1);
+  index_buf[1] = 0;
+  BORINGSSL_keccak_absorb(&keccak_ctx[1], index_buf, 2);
+
+  scalar *outs[] = {out0, out1};
+  scalar_uniform_x2<eta<K>()>(outs, keccak_ctx);
+}
+#endif
+
+#if defined(HAVE_KECCAK_X4)
+template <int K>
+inline void vector_expand_short_step_x4(scalar *out0, scalar *out1,
+                                        scalar *out2, scalar *out3,
+                                        const uint8_t sigma[kSigmaBytes],
+                                        int index) {
+  BORINGSSL_keccak_st keccak_ctx[4];
+  for (int k = 0; k < 4; k++) {
+    BORINGSSL_keccak_init(&keccak_ctx[k], boringssl_shake256);
+    BORINGSSL_keccak_absorb(&keccak_ctx[k], sigma, kSigmaBytes);
+    uint8_t index_buf[2];
+    index_buf[0] = static_cast<uint8_t>(index + k);
+    index_buf[1] = 0;
+    BORINGSSL_keccak_absorb(&keccak_ctx[k], index_buf, 2);
+  }
+
+  scalar *outs[] = {out0, out1, out2, out3};
+  scalar_uniform_x4<eta<K>()>(outs, keccak_ctx);
+}
+#endif
+
+// FIPS 204, Algorithm 33 (`ExpandS`).
+template <int K, int L>
+inline void vector_expand_short(vector<L> *s1, vector<K> *s2,
+                                const uint8_t sigma[kSigmaBytes]) {
+  static_assert(K <= 0x100, "K must fit in 8 bits");
+  static_assert(L <= 0x100, "L must fit in 8 bits");
+  static_assert(K + L <= 0x100, "K+L must fit in 8 bits");
+
+  scalar *all_scalars[K + L];
+  for (int i = 0; i < L; i++) {
+    all_scalars[i] = &s1->v[i];
+  }
+  for (int i = 0; i < K; i++) {
+    all_scalars[L + i] = &s2->v[i];
+  }
+
+  int index = 0;
+#if defined(HAVE_KECCAK_X4)
+  if (BORINGSSL_have_keccak_x4()) {
+    for (; index + 4 <= K + L; index += 4) {
+      vector_expand_short_step_x4<K>(all_scalars[index], all_scalars[index + 1],
+                                     all_scalars[index + 2],
+                                     all_scalars[index + 3], sigma, index);
+    }
+  }
+#endif
+#if defined(HAVE_KECCAK_X2)
+  for (; index + 2 <= K + L; index += 2) {
+    vector_expand_short_step_x2<K>(all_scalars[index], all_scalars[index + 1],
+                                   sigma, index);
+  }
+#endif
+
+  for (; index < K + L; index++) {
+    vector_expand_short_step<K>(all_scalars[index], sigma, index);
+  }
+}
+
+// FIPS 204, Algorithm 34 (`ExpandMask`).
+template <int K, int L>
+inline void vector_expand_mask(vector<L> *out,
+                               const uint8_t rho_prime[kRhoPrimeBytes],
+                               size_t kappa) {
+  assert(kappa + L <= 0x10000);
+
+  for (int i = 0; i < L; i++) {
+    size_t index = kappa + i;
+    scalar_sample_mask<K>(&out->v[i], rho_prime, index);
+  }
+}
+
+/* Encoding */
+
+// FIPS 204, Algorithm 16 (`SimpleBitPack`).
+//
+// Encodes an entire vector into 32*K*`bits` bytes. Note that since 256
+// (kDegree) is divisible by 8, the individual vector entries will always fill a
+// whole number of bytes, so we do not need to worry about bit packing here.
+template <int bits, int K>
+inline void vector_encode(uint8_t *out, const vector<K> *a) {
+  for (int i = 0; i < K; i++) {
+    scalar_encode<bits>(out + i * bits * kDegree / 8, &a->v[i]);
+  }
+}
+
+// FIPS 204, Algorithm 18 (`SimpleBitUnpack`).
+template <int bits, int K>
+inline void vector_decode(vector<K> *out, const uint8_t *in) {
+  for (int i = 0; i < K; i++) {
+    scalar_decode<bits>(&out->v[i], in + i * bits * kDegree / 8);
+  }
+}
+
+// FIPS 204, Algorithm 17 (`BitPack`).
+//
+// Encodes an entire vector into 32*L*`bits` bytes. Note that since 256
+// (kDegree) is divisible by 8, the individual vector entries will always fill a
+// whole number of bytes, so we do not need to worry about bit packing here.
+template <int bits, uint32_t max, int X>
+inline void vector_encode_signed(uint8_t *out, const vector<X> *a) {
+  for (int i = 0; i < X; i++) {
+    scalar_encode_signed<bits, max>(out + i * bits * kDegree / 8, &a->v[i]);
+  }
+}
+
+// `vector_encode_signed_ntt` behaves like `vector_encode_signed` but takes its
+// input in NTT form. This is only used by the testing-only `skEncode` and
+// `skDecode` functions and is not performance-sensitive.
+template <int bits, uint32_t max, int X>
+inline void vector_encode_signed_ntt(uint8_t *out, const vector<X> *a_ntt) {
+  vector<X> a = *a_ntt;
+  vector_inverse_ntt_montgomery(&a);
+  // `vector_inverse_ntt_montgomery` adds an extra factor of R, which we must
+  // cancel with a Montgomery reduction.
+  for (int i = 0; i < X; i++) {
+    for (int j = 0; j < kDegree; j++) {
+      a.v[i].c[j] = reduce_montgomery(a.v[i].c[j]);
+    }
+  }
+
+  vector_encode_signed<bits, max>(out, &a);
+}
+
+// FIPS 204, Algorithm 19 (`BitUnpack`).
+template <int bits, uint32_t max, int X>
+inline int vector_decode_signed(vector<X> *out, const uint8_t *in) {
+  for (int i = 0; i < X; i++) {
+    if (!scalar_decode_signed<bits, max>(&out->v[i],
+                                         in + i * bits * kDegree / 8)) {
+      return 0;
+    }
+  }
+  return 1;
+}
+
+// FIPS 204, Algorithm 28 (`w1Encode`).
+template <int K>
+inline void w1_encode(uint8_t out[w1_bytes<K>()], const vector<K> *w1) {
+  vector_encode<w1_coeffs_bits<K>()>(out, w1);
+}
+
+// FIPS 204, Algorithm 20 (`HintBitPack`).
+template <int K>
+inline void hint_bit_pack(uint8_t out[omega<K>() + K], const vector<K> *h) {
+  OPENSSL_memset(out, 0, omega<K>() + K);
+  int index = 0;
+  for (int i = 0; i < K; i++) {
+    for (int j = 0; j < kDegree; j++) {
+      if (h->v[i].c[j]) {
+        // h must have at most omega<K>() non-zero coefficients.
+        BSSL_CHECK(index < omega<K>());
+        out[index++] = j;
+      }
+    }
+    out[omega<K>() + i] = index;
+  }
+}
+
+// FIPS 204, Algorithm 21 (`HintBitUnpack`).
+template <int K>
+inline int hint_bit_unpack(vector<K> *h, const uint8_t in[omega<K>() + K]) {
+  vector_zero(h);
+  int index = 0;
+  for (int i = 0; i < K; i++) {
+    const int limit = in[omega<K>() + i];
+    if (limit < index || limit > omega<K>()) {
+      return 0;
+    }
+
+    int last = -1;
+    while (index < limit) {
+      int byte = in[index++];
+      if (last >= 0 && byte <= last) {
+        return 0;
+      }
+      last = byte;
+      static_assert(kDegree == 256,
+                    "kDegree must be 256 for this write to be in bounds");
+      h->v[i].c[byte] = 1;
+    }
+  }
+  for (; index < omega<K>(); index++) {
+    if (in[index] != 0) {
+      return 0;
+    }
+  }
+  return 1;
+}
+
+template <int K>
+struct public_key {
+  uint8_t rho[kRhoBytes];
+  vector<K> t1;
+  // Pre-cached value(s).
+  uint8_t public_key_hash[kTrBytes];
+};
+
+template <int K, int L>
+struct private_key {
+  public_key<K> pub;
+  uint8_t k[kKBytes];
+  vector<L> s1_ntt;
+  vector<K> s2_ntt;
+  vector<K> t0_ntt;
+};
+
+template <int K, int L>
+struct signature {
+  uint8_t c_tilde[2 * lambda_bytes<K>()];
+  vector<L> z;
+  vector<K> h;
+};
+
+// FIPS 204, Algorithm 22 (`pkEncode`).
+template <int K>
+inline int mldsa_marshal_public_key(CBB *out, const public_key<K> *pub) {
+  if (!CBB_add_bytes(out, pub->rho, sizeof(pub->rho))) {
+    return 0;
+  }
+
+  uint8_t *vectork_output;
+  if (!CBB_add_space(out, &vectork_output, 320 * K)) {
+    return 0;
+  }
+  vector_encode<10>(vectork_output, &pub->t1);
+
+  return 1;
+}
+
+// FIPS 204, Algorithm 23 (`pkDecode`).
+template <int K>
+inline int mldsa_parse_public_key(public_key<K> *pub, CBS *in) {
+  const CBS orig_in = *in;
+
+  if (!CBS_copy_bytes(in, pub->rho, sizeof(pub->rho))) {
+    return 0;
+  }
+
+  CBS t1_bytes;
+  if (!CBS_get_bytes(in, &t1_bytes, 320 * K) || CBS_len(in) != 0) {
+    return 0;
+  }
+  vector_decode<10>(&pub->t1, CBS_data(&t1_bytes));
+
+  // Compute pre-cached values.
+  BORINGSSL_keccak(pub->public_key_hash, sizeof(pub->public_key_hash),
+                   CBS_data(&orig_in), CBS_len(&orig_in), boringssl_shake256);
+
+  return 1;
+}
+
+// FIPS 204, Algorithm 24 (`skEncode`).
+template <int K, int L>
+inline int mldsa_marshal_private_key(CBB *out, const private_key<K, L> *priv) {
+  if (!CBB_add_bytes(out, priv->pub.rho, sizeof(priv->pub.rho)) ||
+      !CBB_add_bytes(out, priv->k, sizeof(priv->k)) ||
+      !CBB_add_bytes(out, priv->pub.public_key_hash,
+                     sizeof(priv->pub.public_key_hash))) {
+    return 0;
+  }
+
+  constexpr size_t scalar_bytes =
+      (kDegree * plus_minus_eta_bitlen<K>() + 7) / 8;
+  uint8_t *vectorl_output;
+  if (!CBB_add_space(out, &vectorl_output, scalar_bytes * L)) {
+    return 0;
+  }
+  vector_encode_signed_ntt<plus_minus_eta_bitlen<K>(), eta<K>()>(vectorl_output,
+                                                                 &priv->s1_ntt);
+
+  uint8_t *s2_output;
+  if (!CBB_add_space(out, &s2_output, scalar_bytes * K)) {
+    return 0;
+  }
+  vector_encode_signed_ntt<plus_minus_eta_bitlen<K>(), eta<K>()>(s2_output,
+                                                                 &priv->s2_ntt);
+
+  uint8_t *t0_output;
+  if (!CBB_add_space(out, &t0_output, 416 * K)) {
+    return 0;
+  }
+  vector_encode_signed_ntt<13, (1 << 12)>(t0_output, &priv->t0_ntt);
+
+  return 1;
+}
+
+// FIPS 204, Algorithm 25 (`skDecode`). This is only used for testing. The
+// supported external way to construct ML-DSA keys is to use the input seed.
+template <int K, int L>
+inline int mldsa_parse_private_key(private_key<K, L> *priv, CBS *in) {
+  CBS public_key_hash, s1_bytes, s2_bytes, t0_bytes;
+  constexpr size_t scalar_bytes =
+      (kDegree * plus_minus_eta_bitlen<K>() + 7) / 8;
+  if (!CBS_copy_bytes(in, priv->pub.rho, sizeof(priv->pub.rho)) ||
+      !CBS_copy_bytes(in, priv->k, sizeof(priv->k)) ||
+      !CBS_get_bytes(in, &public_key_hash, kTrBytes) ||
+      !CBS_get_bytes(in, &s1_bytes, scalar_bytes * L) ||
+      !vector_decode_signed<plus_minus_eta_bitlen<K>(), eta<K>()>(
+          &priv->s1_ntt, CBS_data(&s1_bytes)) ||
+      !CBS_get_bytes(in, &s2_bytes, scalar_bytes * K) ||
+      !vector_decode_signed<plus_minus_eta_bitlen<K>(), eta<K>()>(
+          &priv->s2_ntt, CBS_data(&s2_bytes)) ||
+      !CBS_get_bytes(in, &t0_bytes, 416 * K)) {
+    return 0;
+  }
+
+  // Compute `t1`, which is not in the `skDecode` input. This also computes
+  // `t0_ntt` and converts `s1_ntt` and `s2_ntt` to NTT form.
+  uint8_t unused[public_key_bytes<K>()];
+  if (!mldsa_finish_keygen(unused, priv)) {
+    return 0;
+  }
+
+  // As a side effect of computing `t1`, we also compute `t0` and
+  // `public_key_hash`. Check they match the received bytes.
+  uint8_t t0_computed[416 * K];
+  vector_encode_signed_ntt<13, (1 << 12)>(t0_computed, &priv->t0_ntt);
+  if (!CBS_mem_equal(&public_key_hash, priv->pub.public_key_hash,
+                     sizeof(priv->pub.public_key_hash)) ||
+      !CBS_mem_equal(&t0_bytes, t0_computed, sizeof(t0_computed))) {
+    return 0;
+  }
+
+  return 1;
+}
+
+// FIPS 204, Algorithm 26 (`sigEncode`).
+template <int K, int L>
+inline int mldsa_marshal_signature(CBB *out, const signature<K, L> *sign) {
+  if (!CBB_add_bytes(out, sign->c_tilde, sizeof(sign->c_tilde))) {
+    return 0;
+  }
+
+  uint8_t *vectorl_output;
+  if (!CBB_add_space(out, &vectorl_output, scalar_le_gamma1_bytes<K>() * L)) {
+    return 0;
+  }
+  vector_encode_signed<gamma1_bits<K>() + 1, gamma1<K>()>(vectorl_output,
+                                                          &sign->z);
+
+  uint8_t *hint_output;
+  if (!CBB_add_space(out, &hint_output, omega<K>() + K)) {
+    return 0;
+  }
+  hint_bit_pack(hint_output, &sign->h);
+
+  return 1;
+}
+
+// FIPS 204, Algorithm 27 (`sigDecode`).
+template <int K, int L>
+inline int mldsa_parse_signature(signature<K, L> *sign, CBS *in) {
+  CBS z_bytes;
+  CBS hint_bytes;
+  if (!CBS_copy_bytes(in, sign->c_tilde, sizeof(sign->c_tilde)) ||
+      !CBS_get_bytes(in, &z_bytes, scalar_le_gamma1_bytes<K>() * L) ||
+      // Note: Decoding b+1 bits into (-2^b, 2^b] cannot fail.
+      !vector_decode_signed<gamma1_bits<K>() + 1, gamma1<K>()>(
+          &sign->z, CBS_data(&z_bytes)) ||
+      !CBS_get_bytes(in, &hint_bytes, omega<K>() + K) ||
+      !hint_bit_unpack(&sign->h, CBS_data(&hint_bytes))) {
+    return 0;
+  };
+
+  return 1;
+}
+
+// FIPS 204, Algorithm 6 (`ML-DSA.KeyGen_internal`), steps 3 and 5–11.
+// Returns 1 on success and 0 on failure.
+//
+// On input, `priv->s1_ntt` and `priv->s2_ntt` are not yet in NTT form. They are
+// converted to NTT form by this function.
+template <int K, int L>
+inline int mldsa_finish_keygen(
+    uint8_t out_encoded_public_key[public_key_bytes<K>()],
+    private_key<K, L> *priv) {
+  // Intermediate values, allocated on the heap to allow use when there is a
+  // limited amount of stack.
+  struct Values {
+    enum { kAllowUniquePtr = true };
+    matrix<K, L> a_ntt;
+    vector<K> t;
+  };
+  auto values = MakeUnique<Values>();
+  if (values == nullptr) {
+    return 0;
+  }
+
+  // Step 3.
+  matrix_expand(&values->a_ntt, priv->pub.rho);
+
+  // Step 5. Note that, on input, `s1_ntt` and `s2_ntt` are not yet in NTT form,
+  // but this function is responsible for converting them to NTT.
+  vector_ntt(&priv->s1_ntt);
+  matrix_mult_montgomery(&values->t, &values->a_ntt, &priv->s1_ntt);
+  vector_inverse_ntt_montgomery(&values->t);
+  vector_add(&values->t, &values->t, &priv->s2_ntt);
+  vector_ntt(&priv->s2_ntt);
+
+  // Step 6-7. We store t0 in NTT form.
+  vector_power2_round(&priv->pub.t1, &priv->t0_ntt, &values->t);
+  vector_ntt(&priv->t0_ntt);
+  // t1 is public.
+  CONSTTIME_DECLASSIFY(&priv->pub.t1, sizeof(priv->pub.t1));
+
+  // Step 8.
+  CBB cbb;
+  CBB_init_fixed(&cbb, out_encoded_public_key, public_key_bytes<K>());
+  if (!mldsa_marshal_public_key(&cbb, &priv->pub)) {
+    return 0;
+  }
+  assert(CBB_len(&cbb) == public_key_bytes<K>());
+
+  // Step 9-11.
+  BORINGSSL_keccak(priv->pub.public_key_hash, sizeof(priv->pub.public_key_hash),
+                   out_encoded_public_key, public_key_bytes<K>(),
+                   boringssl_shake256);
+
+  return 1;
+}
+
+// FIPS 204, Algorithm 6 (`ML-DSA.KeyGen_internal`). Returns 1 on success and 0
+// on failure.
+template <int K, int L>
+inline int mldsa_generate_key_external_entropy_no_self_test(
+    uint8_t out_encoded_public_key[public_key_bytes<K>()],
+    private_key<K, L> *priv, const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  // Step 1-2.
+  uint8_t augmented_entropy[MLDSA_SEED_BYTES + 2];
+  OPENSSL_memcpy(augmented_entropy, entropy, MLDSA_SEED_BYTES);
+  // The k and l parameters are appended to the seed.
+  augmented_entropy[MLDSA_SEED_BYTES] = K;
+  augmented_entropy[MLDSA_SEED_BYTES + 1] = L;
+  uint8_t expanded_seed[kRhoBytes + kSigmaBytes + kKBytes];
+  BORINGSSL_keccak(expanded_seed, sizeof(expanded_seed), augmented_entropy,
+                   sizeof(augmented_entropy), boringssl_shake256);
+  const uint8_t *const rho = expanded_seed;
+  const uint8_t *const sigma = expanded_seed + kRhoBytes;
+  const uint8_t *const k = expanded_seed + kRhoBytes + kSigmaBytes;
+  // rho is public.
+  CONSTTIME_DECLASSIFY(rho, kRhoBytes);
+  OPENSSL_memcpy(priv->pub.rho, rho, sizeof(priv->pub.rho));
+  OPENSSL_memcpy(priv->k, k, sizeof(priv->k));
+  // Steps 4. This is independent of A (step 3) and can be done first.
+  // `mldsa_finish_keygen` will convert `s1_ntt` and `s2_ntt` into NTT from.
+  vector_expand_short(&priv->s1_ntt, &priv->s2_ntt, sigma);
+  // Steps 3 and 5-11.
+  return mldsa_finish_keygen(out_encoded_public_key, priv);
+}
+
+template <int K, int L>
+inline int mldsa_generate_key_external_entropy(
+    uint8_t out_encoded_public_key[public_key_bytes<K>()],
+    private_key<K, L> *priv, const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  fips::ensure_keygen_self_test();
+  return mldsa_generate_key_external_entropy_no_self_test(
+      out_encoded_public_key, priv, entropy);
+}
+
+// FIPS 204, Algorithm 7 (`ML-DSA.Sign_internal`), using a pre-computed mu.
+// Returns 1 on success and 0 on failure.
+template <int K, int L>
+inline int mldsa_sign_mu_no_self_test(
+    uint8_t out_encoded_signature[signature_bytes<K>()],
+    const private_key<K, L> *priv, const uint8_t mu[kMuBytes],
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  uint8_t rho_prime[kRhoPrimeBytes];
+  BORINGSSL_keccak_st keccak_ctx;
+  BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&keccak_ctx, priv->k, sizeof(priv->k));
+  BORINGSSL_keccak_absorb(&keccak_ctx, randomizer,
+                          BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES);
+  BORINGSSL_keccak_absorb(&keccak_ctx, mu, kMuBytes);
+  BORINGSSL_keccak_squeeze(&keccak_ctx, rho_prime, kRhoPrimeBytes);
+
+  // Intermediate values, allocated on the heap to allow use when there is a
+  // limited amount of stack.
+  struct Values {
+    enum { kAllowUniquePtr = true };
+    signature<K, L> sign;
+    matrix<K, L> a_ntt;
+    vector<L> y;
+    vector<K> w;
+    vector<K> w1;
+    vector<L> cs1;
+    vector<K> cs2;
+  };
+  auto values = MakeUnique<Values>();
+  if (values == nullptr) {
+    return 0;
+  }
+  matrix_expand(&values->a_ntt, priv->pub.rho);
+
+  // kappa must not exceed 2**16/L = 13107. But the probability of it
+  // exceeding even 1000 iterations is vanishingly small.
+  for (size_t kappa = 0;; kappa += L) {
+    vector_expand_mask<K, L>(&values->y, rho_prime, kappa);
+
+    vector<L> *y_ntt = &values->cs1;
+    OPENSSL_memcpy(y_ntt, &values->y, sizeof(*y_ntt));
+    vector_ntt(y_ntt);
+
+    matrix_mult_montgomery(&values->w, &values->a_ntt, y_ntt);
+    vector_inverse_ntt_montgomery(&values->w);
+
+    vector_high_bits(&values->w1, &values->w);
+    uint8_t w1_encoded[w1_bytes<K>()];
+    w1_encode(w1_encoded, &values->w1);
+
+    BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake256);
+    BORINGSSL_keccak_absorb(&keccak_ctx, mu, kMuBytes);
+    BORINGSSL_keccak_absorb(&keccak_ctx, w1_encoded, w1_bytes<K>());
+    BORINGSSL_keccak_squeeze(&keccak_ctx, values->sign.c_tilde,
+                             2 * lambda_bytes<K>());
+
+    scalar c_ntt;
+    scalar_sample_in_ball_vartime(&c_ntt, values->sign.c_tilde,
+                                  sizeof(values->sign.c_tilde), tau<K>());
+    scalar_ntt(&c_ntt);
+
+    vector_mult_scalar_montgomery(&values->cs1, &priv->s1_ntt, &c_ntt);
+    vector_inverse_ntt_montgomery(&values->cs1);
+    vector_mult_scalar_montgomery(&values->cs2, &priv->s2_ntt, &c_ntt);
+    vector_inverse_ntt_montgomery(&values->cs2);
+
+    vector_add(&values->sign.z, &values->y, &values->cs1);
+
+    vector<K> *r0 = &values->w1;
+    vector_sub(r0, &values->w, &values->cs2);
+    vector_low_bits(r0, r0);
+
+    // Leaking the fact that a signature was rejected is fine as the next
+    // attempt at a signature will be (indistinguishable from) independent of
+    // this one. Note, however, that we additionally leak which of the two
+    // branches rejected the signature. Section 5.5 of
+    // https://pq-crystals.org/dilithium/data/dilithium-specification-round3.pdf
+    // describes this leak as OK. Note we leak less than what is described by
+    // the paper; we do not reveal which coefficient violated the bound, and
+    // we hide which of the `z_max` or `r0_max` bound failed. See also
+    // https://boringssl-review.googlesource.com/c/boringssl/+/67747/comment/2bbab0fa_d241d35a/
+    uint32_t z_max = vector_max(&values->sign.z);
+    uint32_t r0_max = vector_max_signed(r0);
+    if (constant_time_declassify_w(
+            constant_time_ge_w(z_max, gamma1<K>() - beta<K>()) |
+            constant_time_ge_w(r0_max, gamma2<K>() - beta<K>()))) {
+#if defined(BORINGSSL_FIPS_BREAK_TESTS)
+      // In order to show that our self-tests trigger both restart cases in
+      // this loop, printf-logging is added when built in break-test mode.
+      printf("MLDSA signature restart case 1.\n");
+#endif
+      continue;
+    }
+
+    vector<K> *ct0 = &values->w1;
+    vector_mult_scalar_montgomery(ct0, &priv->t0_ntt, &c_ntt);
+    vector_inverse_ntt_montgomery(ct0);
+    vector_make_hint(&values->sign.h, ct0, &values->cs2, &values->w);
+
+    // See above.
+    uint32_t ct0_max = vector_max(ct0);
+    size_t h_ones = vector_count_ones(&values->sign.h);
+    if (constant_time_declassify_w(constant_time_ge_w(ct0_max, gamma2<K>()) |
+                                   constant_time_lt_w(omega<K>(), h_ones))) {
+#if defined(BORINGSSL_FIPS_BREAK_TESTS)
+      // In order to show that our self-tests trigger both restart cases in
+      // this loop, printf-logging is added when built in break-test mode.
+      printf("MLDSA signature restart case 2.\n");
+#endif
+      continue;
+    }
+
+    // Although computed with the private key, the signature is public.
+    CONSTTIME_DECLASSIFY(values->sign.c_tilde, sizeof(values->sign.c_tilde));
+    CONSTTIME_DECLASSIFY(&values->sign.z, sizeof(values->sign.z));
+    CONSTTIME_DECLASSIFY(&values->sign.h, sizeof(values->sign.h));
+
+    CBB cbb;
+    CBB_init_fixed(&cbb, out_encoded_signature, signature_bytes<K>());
+    if (!mldsa_marshal_signature(&cbb, &values->sign)) {
+      return 0;
+    }
+
+    BSSL_CHECK(CBB_len(&cbb) == signature_bytes<K>());
+    return 1;
+  }
+}
+
+// FIPS 204, Algorithm 7 (`ML-DSA.Sign_internal`), using a pre-computed mu.
+// Returns 1 on success and 0 on failure.
+template <int K, int L>
+inline int mldsa_sign_mu(
+    uint8_t out_encoded_signature[signature_bytes<K>()],
+    const private_key<K, L> *priv, const uint8_t mu[kMuBytes],
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  fips::ensure_sign_self_test();
+  return mldsa_sign_mu_no_self_test(out_encoded_signature, priv, mu,
+                                    randomizer);
+}
+
+// FIPS 204, Algorithm 7 (`ML-DSA.Sign_internal`). Returns 1 on success and 0
+// on failure.
+template <int K, int L>
+inline int mldsa_sign_internal_no_self_test(
+    uint8_t out_encoded_signature[signature_bytes<K>()],
+    const private_key<K, L> *priv, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context_prefix, size_t context_prefix_len,
+    const uint8_t *context, size_t context_len,
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  uint8_t mu[kMuBytes];
+  BORINGSSL_keccak_st keccak_ctx;
+  BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&keccak_ctx, priv->pub.public_key_hash,
+                          sizeof(priv->pub.public_key_hash));
+  BORINGSSL_keccak_absorb(&keccak_ctx, context_prefix, context_prefix_len);
+  BORINGSSL_keccak_absorb(&keccak_ctx, context, context_len);
+  BORINGSSL_keccak_absorb(&keccak_ctx, msg, msg_len);
+  BORINGSSL_keccak_squeeze(&keccak_ctx, mu, kMuBytes);
+
+  return mldsa_sign_mu_no_self_test(out_encoded_signature, priv, mu,
+                                    randomizer);
+}
+
+// FIPS 204, Algorithm 7 (`ML-DSA.Sign_internal`). Returns 1 on success and 0
+// on failure.
+template <int K, int L>
+inline int mldsa_sign_internal(
+    uint8_t out_encoded_signature[signature_bytes<K>()],
+    const private_key<K, L> *priv, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context_prefix, size_t context_prefix_len,
+    const uint8_t *context, size_t context_len,
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  fips::ensure_sign_self_test();
+  return mldsa_sign_internal_no_self_test(
+      out_encoded_signature, priv, msg, msg_len, context_prefix,
+      context_prefix_len, context, context_len, randomizer);
+}
+
+struct prehash_context {
+  BORINGSSL_keccak_st keccak_ctx;
+};
+
+template <int K>
+inline void mldsa_prehash_init(prehash_context *out_prehash_ctx,
+                               const public_key<K> *pub,
+                               const uint8_t *context_prefix,
+                               size_t context_prefix_len,
+                               const uint8_t *context, size_t context_len) {
+  BORINGSSL_keccak_init(&out_prehash_ctx->keccak_ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&out_prehash_ctx->keccak_ctx, pub->public_key_hash,
+                          sizeof(pub->public_key_hash));
+  BORINGSSL_keccak_absorb(&out_prehash_ctx->keccak_ctx, context_prefix,
+                          context_prefix_len);
+  BORINGSSL_keccak_absorb(&out_prehash_ctx->keccak_ctx, context, context_len);
+}
+
+inline void mldsa_prehash_update(prehash_context *inout_prehash_ctx,
+                                 const uint8_t *msg, size_t msg_len) {
+  BORINGSSL_keccak_absorb(&inout_prehash_ctx->keccak_ctx, msg, msg_len);
+}
+
+inline void mldsa_prehash_finalize(uint8_t out_msg_rep[kMuBytes],
+                                   prehash_context *inout_prehash_ctx) {
+  BORINGSSL_keccak_squeeze(&inout_prehash_ctx->keccak_ctx, out_msg_rep,
+                           kMuBytes);
+}
+
+// FIPS 204, Algorithm 8 (`ML-DSA.Verify_internal`), using a pre-computed mu.
+// Returns 1 on success and 0 on failure.
+template <int K, int L>
+inline int mldsa_verify_mu_no_self_test(
+    const public_key<K> *pub,
+    const uint8_t encoded_signature[signature_bytes<K>()],
+    const uint8_t mu[kMuBytes]) {
+  // Intermediate values, allocated on the heap to allow use when there is a
+  // limited amount of stack.
+  struct Values {
+    enum { kAllowUniquePtr = true };
+    signature<K, L> sign;
+    matrix<K, L> a_ntt;
+    vector<L> z_ntt;
+    vector<K> az_ntt;
+    vector<K> ct1_ntt;
+  };
+  auto values = MakeUnique<Values>();
+  if (values == nullptr) {
+    return 0;
+  }
+
+  CBS cbs;
+  CBS_init(&cbs, encoded_signature, signature_bytes<K>());
+  if (!mldsa_parse_signature(&values->sign, &cbs)) {
+    return 0;
+  }
+
+  matrix_expand(&values->a_ntt, pub->rho);
+
+  scalar c_ntt;
+  scalar_sample_in_ball_vartime(&c_ntt, values->sign.c_tilde,
+                                sizeof(values->sign.c_tilde), tau<K>());
+  scalar_ntt(&c_ntt);
+
+  OPENSSL_memcpy(&values->z_ntt, &values->sign.z, sizeof(values->z_ntt));
+  vector_ntt(&values->z_ntt);
+
+  matrix_mult_montgomery(&values->az_ntt, &values->a_ntt, &values->z_ntt);
+
+  vector_scale_power2_round(&values->ct1_ntt, &pub->t1);
+  vector_ntt(&values->ct1_ntt);
+
+  vector_mult_scalar_montgomery(&values->ct1_ntt, &values->ct1_ntt, &c_ntt);
+
+  vector<K> *const w1 = &values->az_ntt;
+  vector_sub(w1, &values->az_ntt, &values->ct1_ntt);
+  vector_inverse_ntt_montgomery(w1);
+
+  vector_use_hint_vartime(w1, &values->sign.h, w1);
+  uint8_t w1_encoded[w1_bytes<K>()];
+  w1_encode(w1_encoded, w1);
+
+  uint8_t c_tilde[2 * lambda_bytes<K>()];
+  BORINGSSL_keccak_st keccak_ctx;
+  BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&keccak_ctx, mu, kMuBytes);
+  BORINGSSL_keccak_absorb(&keccak_ctx, w1_encoded, w1_bytes<K>());
+  BORINGSSL_keccak_squeeze(&keccak_ctx, c_tilde, 2 * lambda_bytes<K>());
+
+  uint32_t z_max = vector_max(&values->sign.z);
+  return z_max < static_cast<uint32_t>(gamma1<K>() - beta<K>()) &&
+         OPENSSL_memcmp(c_tilde, values->sign.c_tilde, 2 * lambda_bytes<K>()) ==
+             0;
+}
+
+// FIPS 204, Algorithm 8 (`ML-DSA.Verify_internal`), using a pre-computed mu.
+// Returns 1 on success and 0 on failure.
+template <int K, int L>
+inline int mldsa_verify_mu(
+    const public_key<K> *pub,
+    const uint8_t encoded_signature[signature_bytes<K>()],
+    const uint8_t mu[kMuBytes]) {
+  fips::ensure_verify_self_test();
+  return mldsa_verify_mu_no_self_test<K, L>(pub, encoded_signature, mu);
+}
+
+// FIPS 204, Algorithm 8 (`ML-DSA.Verify_internal`).
+template <int K, int L>
+inline int mldsa_verify_internal_no_self_test(
+    const public_key<K> *pub,
+    const uint8_t encoded_signature[signature_bytes<K>()], const uint8_t *msg,
+    size_t msg_len, const uint8_t *context_prefix, size_t context_prefix_len,
+    const uint8_t *context, size_t context_len) {
+  uint8_t mu[kMuBytes];
+  BORINGSSL_keccak_st keccak_ctx;
+  BORINGSSL_keccak_init(&keccak_ctx, boringssl_shake256);
+  BORINGSSL_keccak_absorb(&keccak_ctx, pub->public_key_hash,
+                          sizeof(pub->public_key_hash));
+  BORINGSSL_keccak_absorb(&keccak_ctx, context_prefix, context_prefix_len);
+  BORINGSSL_keccak_absorb(&keccak_ctx, context, context_len);
+  BORINGSSL_keccak_absorb(&keccak_ctx, msg, msg_len);
+  BORINGSSL_keccak_squeeze(&keccak_ctx, mu, kMuBytes);
+
+  return mldsa_verify_mu_no_self_test<K, L>(pub, encoded_signature, mu);
+}
+
+// FIPS 204, Algorithm 8 (`ML-DSA.Verify_internal`).
+template <int K, int L>
+inline int mldsa_verify_internal(
+    const public_key<K> *pub,
+    const uint8_t encoded_signature[signature_bytes<K>()], const uint8_t *msg,
+    size_t msg_len, const uint8_t *context_prefix, size_t context_prefix_len,
+    const uint8_t *context, size_t context_len) {
+  fips::ensure_verify_self_test();
+  return mldsa_verify_internal_no_self_test<K, L>(
+      pub, encoded_signature, msg, msg_len, context_prefix, context_prefix_len,
+      context, context_len);
+}
+
+static_assert(sizeof(MLDSA65_private_key) == sizeof(private_key<6, 5>));
+static_assert(alignof(MLDSA65_private_key) == alignof(private_key<6, 5>));
+
+inline const private_key<6, 5> *private_key_from_external_65(
+    const MLDSA65_private_key *external) {
+  return reinterpret_cast<const private_key<6, 5> *>(external);
+}
+inline private_key<6, 5> *private_key_from_external_65(
+    MLDSA65_private_key *external) {
+  return reinterpret_cast<private_key<6, 5> *>(external);
+}
+
+static_assert(sizeof(MLDSA65_public_key) == sizeof(public_key<6>));
+static_assert(alignof(MLDSA65_public_key) == alignof(public_key<6>));
+
+inline const public_key<6> *public_key_from_external_65(
+    const MLDSA65_public_key *external) {
+  return reinterpret_cast<const public_key<6> *>(external);
+}
+inline public_key<6> *public_key_from_external_65(
+    MLDSA65_public_key *external) {
+  return reinterpret_cast<public_key<6> *>(external);
+}
+
+inline prehash_context *prehash_context_from_external_65(
+    MLDSA65_prehash *external) {
+  static_assert(sizeof(MLDSA65_prehash) == sizeof(prehash_context));
+  static_assert(alignof(MLDSA65_prehash) == alignof(prehash_context));
+  return reinterpret_cast<prehash_context *>(external);
+}
+
+static_assert(sizeof(MLDSA87_private_key) == sizeof(private_key<8, 7>));
+static_assert(alignof(MLDSA87_private_key) == alignof(private_key<8, 7>));
+inline const private_key<8, 7> *private_key_from_external_87(
+    const MLDSA87_private_key *external) {
+  return reinterpret_cast<const private_key<8, 7> *>(external);
+}
+inline private_key<8, 7> *private_key_from_external_87(
+    MLDSA87_private_key *external) {
+  return reinterpret_cast<private_key<8, 7> *>(external);
+}
+
+static_assert(sizeof(MLDSA87_public_key) == sizeof(public_key<8>));
+static_assert(alignof(MLDSA87_public_key) == alignof(public_key<8>));
+
+inline const public_key<8> *public_key_from_external_87(
+    const MLDSA87_public_key *external) {
+  return reinterpret_cast<const public_key<8> *>(external);
+}
+inline public_key<8> *public_key_from_external_87(
+    MLDSA87_public_key *external) {
+  return reinterpret_cast<public_key<8> *>(external);
+}
+
+inline prehash_context *prehash_context_from_external_87(
+    MLDSA87_prehash *external) {
+  static_assert(sizeof(MLDSA87_prehash) == sizeof(prehash_context));
+  static_assert(alignof(MLDSA87_prehash) == alignof(prehash_context));
+  return reinterpret_cast<prehash_context *>(external);
+}
+
+static_assert(sizeof(MLDSA44_private_key) == sizeof(private_key<4, 4>));
+static_assert(alignof(MLDSA44_private_key) == alignof(private_key<4, 4>));
+
+inline const private_key<4, 4> *private_key_from_external_44(
+    const MLDSA44_private_key *external) {
+  return reinterpret_cast<const private_key<4, 4> *>(external);
+}
+inline private_key<4, 4> *private_key_from_external_44(
+    MLDSA44_private_key *external) {
+  return reinterpret_cast<private_key<4, 4> *>(external);
+}
+
+static_assert(sizeof(MLDSA44_public_key) == sizeof(public_key<4>));
+static_assert(alignof(MLDSA44_public_key) == alignof(public_key<4>));
+
+inline const public_key<4> *public_key_from_external_44(
+    const MLDSA44_public_key *external) {
+  return reinterpret_cast<const public_key<4> *>(external);
+}
+inline public_key<4> *public_key_from_external_44(
+    MLDSA44_public_key *external) {
+  return reinterpret_cast<public_key<4> *>(external);
+}
+
+inline prehash_context *prehash_context_from_external_44(
+    MLDSA44_prehash *external) {
+  static_assert(sizeof(MLDSA44_prehash) == sizeof(prehash_context));
+  static_assert(alignof(MLDSA44_prehash) == alignof(prehash_context));
+  return reinterpret_cast<prehash_context *>(external);
+}
+
+namespace fips {
+
+#include "fips_known_values.inc"
+
+inline int keygen_self_test() {
+  struct Values {
+    enum { kAllowUniquePtr = true };
+    private_key<6, 5> priv;
+    uint8_t pub_bytes[MLDSA65_PUBLIC_KEY_BYTES];
+    uint8_t priv_bytes[BCM_MLDSA65_PRIVATE_KEY_BYTES];
+  };
+  auto values = MakeUnique<Values>();
+  if (values == nullptr ||
+      !mldsa_generate_key_external_entropy_no_self_test(
+          values->pub_bytes, &values->priv, kGenerateKeyEntropy)) {
+    return 0;
+  }
+
+  CBB cbb;
+  CBB_init_fixed(&cbb, values->priv_bytes, sizeof(values->priv_bytes));
+  if (!mldsa_marshal_private_key(&cbb, &values->priv)) {
+    return 0;
+  }
+
+  if (!BORINGSSL_check_test(kExpectedPublicKey, values->pub_bytes,
+                            "ML-DSA keygen public key") ||
+      !BORINGSSL_check_test(kExpectedPrivateKey, values->priv_bytes,
+                            "ML-DSA keygen private key")) {
+    return 0;
+  }
+
+  return 1;
+}
+
+inline int sign_self_test() {
+  struct Values {
+    enum { kAllowUniquePtr = true };
+    private_key<6, 5> priv;
+    uint8_t pub_bytes[MLDSA65_PUBLIC_KEY_BYTES];
+    uint8_t sig[MLDSA65_SIGNATURE_BYTES];
+  };
+  auto values = MakeUnique<Values>();
+  if (values == nullptr ||
+      !mldsa_generate_key_external_entropy(values->pub_bytes, &values->priv,
+                                           kSignEntropy)) {
+    return 0;
+  }
+
+  const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES] = {};
+
+  // This message triggers the first restart case for signing.
+  uint8_t message[4] = {0};
+  if (!mldsa_sign_internal_no_self_test(values->sig, &values->priv, message,
+                                        sizeof(message), nullptr, 0, nullptr, 0,
+                                        randomizer)) {
+    return 0;
+  }
+  if (!BORINGSSL_check_test(kExpectedCase1Signature, values->sig,
+                            "ML-DSA sign case 1")) {
+    return 0;
+  }
+
+  // This message triggers the second restart case for signing.
+  message[0] = 123;
+  if (!mldsa_sign_internal_no_self_test(values->sig, &values->priv, message,
+                                        sizeof(message), nullptr, 0, nullptr, 0,
+                                        randomizer)) {
+    return 0;
+  }
+  if (!BORINGSSL_check_test(kExpectedCase2Signature, values->sig,
+                            "ML-DSA sign case 2")) {
+    return 0;
+  }
+
+  return 1;
+}
+
+inline int verify_self_test() {
+  struct Values {
+    enum { kAllowUniquePtr = true };
+    private_key<6, 5> priv;
+    uint8_t pub_bytes[MLDSA65_PUBLIC_KEY_BYTES];
+  };
+  auto values = MakeUnique<Values>();
+  if (values == nullptr) {
+    return 0;
+  }
+
+  if (!mldsa_generate_key_external_entropy(values->pub_bytes, &values->priv,
+                                           kSignEntropy)) {
+    return 0;
+  }
+
+  const uint8_t message[4] = {1, 0};
+  if (!mldsa_verify_internal_no_self_test<6, 5>(
+          &values->priv.pub, kExpectedVerifySignature, message, sizeof(message),
+          nullptr, 0, nullptr, 0)) {
+    return 0;
+  }
+
+  return 1;
+}
+
+template <int K, int L>
+inline int check_key(const private_key<K, L> *priv) {
+  uint8_t sig[signature_bytes<K>()];
+  uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES] = {};
+  if (!mldsa_sign_internal_no_self_test(sig, priv, nullptr, 0, nullptr, 0,
+                                        nullptr, 0, randomizer)) {
+    return 0;
+  }
+
+  if (boringssl_fips_break_test("MLDSA_PWCT")) {
+    sig[0] ^= 1;
+  }
+
+  if (!mldsa_verify_internal_no_self_test<K, L>(&priv->pub, sig, nullptr, 0,
+                                                nullptr, 0, nullptr, 0)) {
+    return 0;
+  }
+  return 1;
+}
+
+#if defined(BORINGSSL_FIPS)
+
+void ensure_keygen_self_test() {
+  CRYPTO_once(g_mldsa_keygen_self_test_once_bss_get(), []() {
+    if (!keygen_self_test()) {
+      BORINGSSL_FIPS_abort();
+    }
+  });
+}
+
+void ensure_sign_self_test() {
+  CRYPTO_once(g_mldsa_sign_self_test_once_bss_get(), []() {
+    if (!sign_self_test()) {
+      BORINGSSL_FIPS_abort();
+    }
+  });
+}
+
+void ensure_verify_self_test() {
+  CRYPTO_once(g_mldsa_verify_self_test_once_bss_get(), []() {
+    if (!verify_self_test()) {
+      BORINGSSL_FIPS_abort();
+    }
+  });
+}
+
+#else
+
+void ensure_keygen_self_test() {}
+void ensure_sign_self_test() {}
+void ensure_verify_self_test() {}
+
+#endif
+
+}  // namespace fips
+
+}  // namespace
+}  // namespace mldsa
+
+
+// ML-DSA-65 specific wrappers.
+
+bcm_status bssl::BCM_mldsa65_parse_public_key(MLDSA65_public_key *public_key,
+                                              CBS *in) {
+  return bcm_as_approved_status(mldsa_parse_public_key(
+      mldsa::public_key_from_external_65(public_key), in));
+}
+
+bcm_status bssl::BCM_mldsa65_marshal_private_key(
+    CBB *out, const MLDSA65_private_key *private_key) {
+  return bcm_as_approved_status(mldsa_marshal_private_key(
+      out, mldsa::private_key_from_external_65(private_key)));
+}
+
+bcm_status bssl::BCM_mldsa65_parse_private_key(MLDSA65_private_key *private_key,
+                                               CBS *in) {
+  return bcm_as_approved_status(
+      mldsa_parse_private_key(mldsa::private_key_from_external_65(private_key),
+                              in) &&
+      CBS_len(in) == 0);
+}
+
+bcm_status bssl::BCM_mldsa65_check_key_fips(MLDSA65_private_key *private_key) {
+  return bcm_as_approved_status(
+      mldsa::fips::check_key(mldsa::private_key_from_external_65(private_key)));
+}
+
+// Calls `MLDSA_generate_key_external_entropy` with random bytes from
+// `BCM_rand_bytes`.
+bcm_status bssl::BCM_mldsa65_generate_key(
+    uint8_t out_encoded_public_key[MLDSA65_PUBLIC_KEY_BYTES],
+    uint8_t out_seed[MLDSA_SEED_BYTES], MLDSA65_private_key *out_private_key) {
+  BCM_rand_bytes(out_seed, MLDSA_SEED_BYTES);
+  CONSTTIME_SECRET(out_seed, MLDSA_SEED_BYTES);
+  return BCM_mldsa65_generate_key_external_entropy(out_encoded_public_key,
+                                                   out_private_key, out_seed);
+}
+
+bcm_status bssl::BCM_mldsa65_private_key_from_seed(
+    MLDSA65_private_key *out_private_key,
+    const uint8_t seed[MLDSA_SEED_BYTES]) {
+  uint8_t public_key[MLDSA65_PUBLIC_KEY_BYTES];
+  return BCM_mldsa65_generate_key_external_entropy(public_key, out_private_key,
+                                                   seed);
+}
+
+bcm_status bssl::BCM_mldsa65_generate_key_external_entropy(
+    uint8_t out_encoded_public_key[MLDSA65_PUBLIC_KEY_BYTES],
+    MLDSA65_private_key *out_private_key,
+    const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  return bcm_as_not_approved_status(mldsa_generate_key_external_entropy(
+      out_encoded_public_key,
+      mldsa::private_key_from_external_65(out_private_key), entropy));
+}
+
+bcm_status bssl::BCM_mldsa65_generate_key_fips(
+    uint8_t out_encoded_public_key[MLDSA65_PUBLIC_KEY_BYTES],
+    uint8_t out_seed[MLDSA_SEED_BYTES], MLDSA65_private_key *out_private_key) {
+  if (out_encoded_public_key == nullptr || out_private_key == nullptr) {
+    return bcm_status::failure;
+  }
+  if (BCM_mldsa65_generate_key(out_encoded_public_key, out_seed,
+                               out_private_key) == bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa65_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa65_generate_key_external_entropy_fips(
+    uint8_t out_encoded_public_key[MLDSA65_PUBLIC_KEY_BYTES],
+    MLDSA65_private_key *out_private_key,
+    const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  if (out_encoded_public_key == nullptr || out_private_key == nullptr) {
+    return bcm_status::failure;
+  }
+  if (BCM_mldsa65_generate_key_external_entropy(out_encoded_public_key,
+                                                out_private_key, entropy) ==
+      bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa65_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa65_private_key_from_seed_fips(
+    MLDSA65_private_key *out_private_key,
+    const uint8_t seed[MLDSA_SEED_BYTES]) {
+  uint8_t public_key[MLDSA65_PUBLIC_KEY_BYTES];
+  if (BCM_mldsa65_generate_key_external_entropy(public_key, out_private_key,
+                                                seed) == bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa65_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa65_public_from_private(
+    MLDSA65_public_key *out_public_key,
+    const MLDSA65_private_key *private_key) {
+  const auto *priv = mldsa::private_key_from_external_65(private_key);
+  auto *out_pub = mldsa::public_key_from_external_65(out_public_key);
+  *out_pub = priv->pub;
+  return bcm_status::approved;
+}
+
+const MLDSA65_public_key *bssl::BCM_mldsa65_public_of_private(
+    const MLDSA65_private_key *private_key) {
+  return reinterpret_cast<const MLDSA65_public_key *>(
+      &mldsa::private_key_from_external_65(private_key)->pub);
+}
+
+bcm_status bssl::BCM_mldsa65_sign_internal(
+    uint8_t out_encoded_signature[MLDSA65_SIGNATURE_BYTES],
+    const MLDSA65_private_key *private_key, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context_prefix, size_t context_prefix_len,
+    const uint8_t *context, size_t context_len,
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  return bcm_as_approved_status(mldsa_sign_internal(
+      out_encoded_signature, mldsa::private_key_from_external_65(private_key),
+      msg, msg_len, context_prefix, context_prefix_len, context, context_len,
+      randomizer));
+}
+
+bcm_status bssl::BCM_mldsa65_sign_mu_internal(
+    uint8_t out_encoded_signature[MLDSA65_SIGNATURE_BYTES],
+    const MLDSA65_private_key *private_key,
+    const uint8_t msg_rep[MLDSA_MU_BYTES],
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  return bcm_as_approved_status(mldsa_sign_mu(
+      out_encoded_signature, mldsa::private_key_from_external_65(private_key),
+      msg_rep, randomizer));
+}
+
+// ML-DSA signature in randomized mode, filling the random bytes with
+// `BCM_rand_bytes`.
+bcm_status bssl::BCM_mldsa65_sign(
+    uint8_t out_encoded_signature[MLDSA65_SIGNATURE_BYTES],
+    const MLDSA65_private_key *private_key, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context, size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+  uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES];
+  BCM_rand_bytes(randomizer, sizeof(randomizer));
+  CONSTTIME_SECRET(randomizer, sizeof(randomizer));
+
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  return BCM_mldsa65_sign_internal(
+      out_encoded_signature, private_key, msg, msg_len, context_prefix,
+      sizeof(context_prefix), context, context_len, randomizer);
+}
+
+// ML-DSA pre-hashed API: initializing a pre-hashing context.
+void bssl::BCM_mldsa65_prehash_init(MLDSA65_prehash *out_prehash_ctx,
+                                    const MLDSA65_public_key *public_key,
+                                    const uint8_t *context,
+                                    size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  mldsa_prehash_init(mldsa::prehash_context_from_external_65(out_prehash_ctx),
+                     mldsa::public_key_from_external_65(public_key),
+                     context_prefix, sizeof(context_prefix), context,
+                     context_len);
+}
+
+// ML-DSA pre-hashed API: updating a pre-hashing context with a message chunk.
+void bssl::BCM_mldsa65_prehash_update(MLDSA65_prehash *inout_prehash_ctx,
+                                      const uint8_t *msg, size_t msg_len) {
+  mldsa_prehash_update(
+      mldsa::prehash_context_from_external_65(inout_prehash_ctx), msg, msg_len);
+}
+
+// ML-DSA pre-hashed API: obtaining a message representative to sign.
+void bssl::BCM_mldsa65_prehash_finalize(uint8_t out_msg_rep[MLDSA_MU_BYTES],
+                                        MLDSA65_prehash *inout_prehash_ctx) {
+  mldsa_prehash_finalize(
+      out_msg_rep, mldsa::prehash_context_from_external_65(inout_prehash_ctx));
+}
+
+// ML-DSA pre-hashed API: signing a message representative.
+bcm_status bssl::BCM_mldsa65_sign_message_representative(
+    uint8_t out_encoded_signature[MLDSA65_SIGNATURE_BYTES],
+    const MLDSA65_private_key *private_key,
+    const uint8_t msg_rep[MLDSA_MU_BYTES]) {
+  uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES];
+  BCM_rand_bytes(randomizer, sizeof(randomizer));
+  CONSTTIME_SECRET(randomizer, sizeof(randomizer));
+
+  return bcm_as_approved_status(mldsa_sign_mu(
+      out_encoded_signature, mldsa::private_key_from_external_65(private_key),
+      msg_rep, randomizer));
+}
+
+// ML-DSA pre-hashed API: verifying a message representative.
+bcm_status bssl::BCM_mldsa65_verify_message_representative(
+    const MLDSA65_public_key *public_key,
+    const uint8_t signature[MLDSA65_SIGNATURE_BYTES],
+    const uint8_t msg_rep[MLDSA_MU_BYTES]) {
+  return bcm_as_approved_status(mldsa::mldsa_verify_mu<6, 5>(
+      mldsa::public_key_from_external_65(public_key), signature, msg_rep));
+}
+
+// FIPS 204, Algorithm 3 (`ML-DSA.Verify`).
+bcm_status bssl::BCM_mldsa65_verify(
+    const MLDSA65_public_key *public_key,
+    const uint8_t signature[MLDSA65_SIGNATURE_BYTES], const uint8_t *msg,
+    size_t msg_len, const uint8_t *context, size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  return BCM_mldsa65_verify_internal(public_key, signature, msg, msg_len,
+                                     context_prefix, sizeof(context_prefix),
+                                     context, context_len);
+}
+
+bcm_status bssl::BCM_mldsa65_verify_internal(
+    const MLDSA65_public_key *public_key,
+    const uint8_t encoded_signature[MLDSA65_SIGNATURE_BYTES],
+    const uint8_t *msg, size_t msg_len, const uint8_t *context_prefix,
+    size_t context_prefix_len, const uint8_t *context, size_t context_len) {
+  return bcm_as_approved_status(mldsa::mldsa_verify_internal<6, 5>(
+      mldsa::public_key_from_external_65(public_key), encoded_signature, msg,
+      msg_len, context_prefix, context_prefix_len, context, context_len));
+}
+
+bcm_status bssl::BCM_mldsa65_marshal_public_key(
+    CBB *out, const MLDSA65_public_key *public_key) {
+  return bcm_as_approved_status(mldsa_marshal_public_key(
+      out, mldsa::public_key_from_external_65(public_key)));
+}
+
+int bssl::BCM_mldsa65_public_keys_equal(const MLDSA65_public_key *a,
+                                        const MLDSA65_public_key *b) {
+  auto *a_pub = mldsa::public_key_from_external_65(a);
+  auto *b_pub = mldsa::public_key_from_external_65(b);
+  // It is sufficient to compare `public_key_hash`. When importing a public key,
+  // the hash must be computed. When importing a private key in expanded form
+  // (an internal testing-only API), the hash is provided, but we recompute it
+  // and check for correctness.
+  return OPENSSL_memcmp(a_pub->public_key_hash, b_pub->public_key_hash,
+                        sizeof(a_pub->public_key_hash)) == 0;
+}
+
+
+// ML-DSA-87 specific wrappers.
+
+bcm_status bssl::BCM_mldsa87_parse_public_key(MLDSA87_public_key *public_key,
+                                              CBS *in) {
+  return bcm_as_approved_status(mldsa_parse_public_key(
+      mldsa::public_key_from_external_87(public_key), in));
+}
+
+bcm_status bssl::BCM_mldsa87_marshal_private_key(
+    CBB *out, const MLDSA87_private_key *private_key) {
+  return bcm_as_approved_status(mldsa_marshal_private_key(
+      out, mldsa::private_key_from_external_87(private_key)));
+}
+
+bcm_status bssl::BCM_mldsa87_parse_private_key(MLDSA87_private_key *private_key,
+                                               CBS *in) {
+  return bcm_as_approved_status(
+      mldsa_parse_private_key(mldsa::private_key_from_external_87(private_key),
+                              in) &&
+      CBS_len(in) == 0);
+}
+
+bcm_status bssl::BCM_mldsa87_check_key_fips(MLDSA87_private_key *private_key) {
+  return bcm_as_approved_status(
+      mldsa::fips::check_key(mldsa::private_key_from_external_87(private_key)));
+}
+
+// Calls `MLDSA_generate_key_external_entropy` with random bytes from
+// `BCM_rand_bytes`.
+bcm_status bssl::BCM_mldsa87_generate_key(
+    uint8_t out_encoded_public_key[MLDSA87_PUBLIC_KEY_BYTES],
+    uint8_t out_seed[MLDSA_SEED_BYTES], MLDSA87_private_key *out_private_key) {
+  BCM_rand_bytes(out_seed, MLDSA_SEED_BYTES);
+  CONSTTIME_SECRET(out_seed, MLDSA_SEED_BYTES);
+  return BCM_mldsa87_generate_key_external_entropy(out_encoded_public_key,
+                                                   out_private_key, out_seed);
+}
+
+bcm_status bssl::BCM_mldsa87_private_key_from_seed(
+    MLDSA87_private_key *out_private_key,
+    const uint8_t seed[MLDSA_SEED_BYTES]) {
+  uint8_t public_key[MLDSA87_PUBLIC_KEY_BYTES];
+  return BCM_mldsa87_generate_key_external_entropy(public_key, out_private_key,
+                                                   seed);
+}
+
+bcm_status bssl::BCM_mldsa87_generate_key_external_entropy(
+    uint8_t out_encoded_public_key[MLDSA87_PUBLIC_KEY_BYTES],
+    MLDSA87_private_key *out_private_key,
+    const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  return bcm_as_not_approved_status(mldsa_generate_key_external_entropy(
+      out_encoded_public_key,
+      mldsa::private_key_from_external_87(out_private_key), entropy));
+}
+
+bcm_status bssl::BCM_mldsa87_generate_key_fips(
+    uint8_t out_encoded_public_key[MLDSA87_PUBLIC_KEY_BYTES],
+    uint8_t out_seed[MLDSA_SEED_BYTES], MLDSA87_private_key *out_private_key) {
+  if (out_encoded_public_key == nullptr || out_private_key == nullptr) {
+    return bcm_status::failure;
+  }
+  if (BCM_mldsa87_generate_key(out_encoded_public_key, out_seed,
+                               out_private_key) == bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa87_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa87_generate_key_external_entropy_fips(
+    uint8_t out_encoded_public_key[MLDSA87_PUBLIC_KEY_BYTES],
+    MLDSA87_private_key *out_private_key,
+    const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  if (out_encoded_public_key == nullptr || out_private_key == nullptr) {
+    return bcm_status::failure;
+  }
+  if (BCM_mldsa87_generate_key_external_entropy(out_encoded_public_key,
+                                                out_private_key, entropy) ==
+      bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa87_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa87_private_key_from_seed_fips(
+    MLDSA87_private_key *out_private_key,
+    const uint8_t seed[MLDSA_SEED_BYTES]) {
+  uint8_t public_key[MLDSA87_PUBLIC_KEY_BYTES];
+  if (BCM_mldsa87_generate_key_external_entropy(public_key, out_private_key,
+                                                seed) == bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa87_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa87_public_from_private(
+    MLDSA87_public_key *out_public_key,
+    const MLDSA87_private_key *private_key) {
+  const auto *priv = mldsa::private_key_from_external_87(private_key);
+  auto *out_pub = mldsa::public_key_from_external_87(out_public_key);
+  *out_pub = priv->pub;
+  return bcm_status::approved;
+}
+
+const MLDSA87_public_key *bssl::BCM_mldsa87_public_of_private(
+    const MLDSA87_private_key *private_key) {
+  return reinterpret_cast<const MLDSA87_public_key *>(
+      &mldsa::private_key_from_external_87(private_key)->pub);
+}
+
+bcm_status bssl::BCM_mldsa87_sign_internal(
+    uint8_t out_encoded_signature[MLDSA87_SIGNATURE_BYTES],
+    const MLDSA87_private_key *private_key, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context_prefix, size_t context_prefix_len,
+    const uint8_t *context, size_t context_len,
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  return bcm_as_approved_status(mldsa_sign_internal(
+      out_encoded_signature, mldsa::private_key_from_external_87(private_key),
+      msg, msg_len, context_prefix, context_prefix_len, context, context_len,
+      randomizer));
+}
+
+bcm_status bssl::BCM_mldsa87_sign_mu_internal(
+    uint8_t out_encoded_signature[MLDSA87_SIGNATURE_BYTES],
+    const MLDSA87_private_key *private_key,
+    const uint8_t msg_rep[MLDSA_MU_BYTES],
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  return bcm_as_approved_status(mldsa_sign_mu(
+      out_encoded_signature, mldsa::private_key_from_external_87(private_key),
+      msg_rep, randomizer));
+}
+
+// ML-DSA signature in randomized mode, filling the random bytes with
+// `BCM_rand_bytes`.
+bcm_status bssl::BCM_mldsa87_sign(
+    uint8_t out_encoded_signature[MLDSA87_SIGNATURE_BYTES],
+    const MLDSA87_private_key *private_key, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context, size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+  uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES];
+  BCM_rand_bytes(randomizer, sizeof(randomizer));
+  CONSTTIME_SECRET(randomizer, sizeof(randomizer));
+
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  return BCM_mldsa87_sign_internal(
+      out_encoded_signature, private_key, msg, msg_len, context_prefix,
+      sizeof(context_prefix), context, context_len, randomizer);
+}
+
+// ML-DSA pre-hashed API: initializing a pre-hashing context.
+void bssl::BCM_mldsa87_prehash_init(MLDSA87_prehash *out_prehash_ctx,
+                                    const MLDSA87_public_key *public_key,
+                                    const uint8_t *context,
+                                    size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  mldsa_prehash_init(mldsa::prehash_context_from_external_87(out_prehash_ctx),
+                     mldsa::public_key_from_external_87(public_key),
+                     context_prefix, sizeof(context_prefix), context,
+                     context_len);
+}
+
+// ML-DSA pre-hashed API: updating a pre-hashing context with a message chunk.
+void bssl::BCM_mldsa87_prehash_update(MLDSA87_prehash *inout_prehash_ctx,
+                                      const uint8_t *msg, size_t msg_len) {
+  mldsa_prehash_update(
+      mldsa::prehash_context_from_external_87(inout_prehash_ctx), msg, msg_len);
+}
+
+// ML-DSA pre-hashed API: obtaining a message representative to sign.
+void bssl::BCM_mldsa87_prehash_finalize(uint8_t out_msg_rep[MLDSA_MU_BYTES],
+                                        MLDSA87_prehash *inout_prehash_ctx) {
+  mldsa_prehash_finalize(
+      out_msg_rep, mldsa::prehash_context_from_external_87(inout_prehash_ctx));
+}
+
+// ML-DSA pre-hashed API: signing a message representative.
+bcm_status bssl::BCM_mldsa87_sign_message_representative(
+    uint8_t out_encoded_signature[MLDSA87_SIGNATURE_BYTES],
+    const MLDSA87_private_key *private_key,
+    const uint8_t msg_rep[MLDSA_MU_BYTES]) {
+  uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES];
+  BCM_rand_bytes(randomizer, sizeof(randomizer));
+  CONSTTIME_SECRET(randomizer, sizeof(randomizer));
+
+  return bcm_as_approved_status(mldsa_sign_mu(
+      out_encoded_signature, mldsa::private_key_from_external_87(private_key),
+      msg_rep, randomizer));
+}
+
+// ML-DSA pre-hashed API: verifying a message representative.
+bcm_status bssl::BCM_mldsa87_verify_message_representative(
+    const MLDSA87_public_key *public_key,
+    const uint8_t signature[MLDSA87_SIGNATURE_BYTES],
+    const uint8_t msg_rep[MLDSA_MU_BYTES]) {
+  return bcm_as_approved_status(mldsa::mldsa_verify_mu<8, 7>(
+      mldsa::public_key_from_external_87(public_key), signature, msg_rep));
+}
+
+// FIPS 204, Algorithm 3 (`ML-DSA.Verify`).
+bcm_status bssl::BCM_mldsa87_verify(const MLDSA87_public_key *public_key,
+                                    const uint8_t *signature,
+                                    const uint8_t *msg, size_t msg_len,
+                                    const uint8_t *context,
+                                    size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  return BCM_mldsa87_verify_internal(public_key, signature, msg, msg_len,
+                                     context_prefix, sizeof(context_prefix),
+                                     context, context_len);
+}
+
+bcm_status bssl::BCM_mldsa87_verify_internal(
+    const MLDSA87_public_key *public_key,
+    const uint8_t encoded_signature[MLDSA87_SIGNATURE_BYTES],
+    const uint8_t *msg, size_t msg_len, const uint8_t *context_prefix,
+    size_t context_prefix_len, const uint8_t *context, size_t context_len) {
+  return bcm_as_approved_status(mldsa::mldsa_verify_internal<8, 7>(
+      mldsa::public_key_from_external_87(public_key), encoded_signature, msg,
+      msg_len, context_prefix, context_prefix_len, context, context_len));
+}
+
+bcm_status bssl::BCM_mldsa87_marshal_public_key(
+    CBB *out, const MLDSA87_public_key *public_key) {
+  return bcm_as_approved_status(mldsa_marshal_public_key(
+      out, mldsa::public_key_from_external_87(public_key)));
+}
+
+int bssl::BCM_mldsa87_public_keys_equal(const MLDSA87_public_key *a,
+                                        const MLDSA87_public_key *b) {
+  auto *a_pub = mldsa::public_key_from_external_87(a);
+  auto *b_pub = mldsa::public_key_from_external_87(b);
+  // It is sufficient to compare `public_key_hash`. When importing a public key,
+  // the hash must be computed. When importing a private key in expanded form
+  // (an internal testing-only API), the hash is provided, but we recompute it
+  // and check for correctness.
+  return OPENSSL_memcmp(a_pub->public_key_hash, b_pub->public_key_hash,
+                        sizeof(a_pub->public_key_hash)) == 0;
+}
+
+
+// ML-DSA-44 specific wrappers.
+
+bcm_status bssl::BCM_mldsa44_parse_public_key(MLDSA44_public_key *public_key,
+                                              CBS *in) {
+  return bcm_as_approved_status(mldsa_parse_public_key(
+      mldsa::public_key_from_external_44(public_key), in));
+}
+
+bcm_status bssl::BCM_mldsa44_marshal_private_key(
+    CBB *out, const MLDSA44_private_key *private_key) {
+  return bcm_as_approved_status(mldsa_marshal_private_key(
+      out, mldsa::private_key_from_external_44(private_key)));
+}
+
+bcm_status bssl::BCM_mldsa44_parse_private_key(MLDSA44_private_key *private_key,
+                                               CBS *in) {
+  return bcm_as_approved_status(
+      mldsa_parse_private_key(mldsa::private_key_from_external_44(private_key),
+                              in) &&
+      CBS_len(in) == 0);
+}
+
+bcm_status bssl::BCM_mldsa44_check_key_fips(MLDSA44_private_key *private_key) {
+  return bcm_as_approved_status(
+      mldsa::fips::check_key(mldsa::private_key_from_external_44(private_key)));
+}
+
+// Calls `MLDSA_generate_key_external_entropy` with random bytes from
+// `BCM_rand_bytes`.
+bcm_status bssl::BCM_mldsa44_generate_key(
+    uint8_t out_encoded_public_key[MLDSA44_PUBLIC_KEY_BYTES],
+    uint8_t out_seed[MLDSA_SEED_BYTES], MLDSA44_private_key *out_private_key) {
+  BCM_rand_bytes(out_seed, MLDSA_SEED_BYTES);
+  CONSTTIME_SECRET(out_seed, MLDSA_SEED_BYTES);
+  return BCM_mldsa44_generate_key_external_entropy(out_encoded_public_key,
+                                                   out_private_key, out_seed);
+}
+
+bcm_status bssl::BCM_mldsa44_private_key_from_seed(
+    MLDSA44_private_key *out_private_key,
+    const uint8_t seed[MLDSA_SEED_BYTES]) {
+  uint8_t public_key[MLDSA44_PUBLIC_KEY_BYTES];
+  return BCM_mldsa44_generate_key_external_entropy(public_key, out_private_key,
+                                                   seed);
+}
+
+bcm_status bssl::BCM_mldsa44_generate_key_external_entropy(
+    uint8_t out_encoded_public_key[MLDSA44_PUBLIC_KEY_BYTES],
+    MLDSA44_private_key *out_private_key,
+    const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  return bcm_as_not_approved_status(mldsa_generate_key_external_entropy(
+      out_encoded_public_key,
+      mldsa::private_key_from_external_44(out_private_key), entropy));
+}
+
+bcm_status bssl::BCM_mldsa44_generate_key_fips(
+    uint8_t out_encoded_public_key[MLDSA44_PUBLIC_KEY_BYTES],
+    uint8_t out_seed[MLDSA_SEED_BYTES], MLDSA44_private_key *out_private_key) {
+  if (out_encoded_public_key == nullptr || out_private_key == nullptr) {
+    return bcm_status::failure;
+  }
+  if (BCM_mldsa44_generate_key(out_encoded_public_key, out_seed,
+                               out_private_key) == bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa44_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa44_generate_key_external_entropy_fips(
+    uint8_t out_encoded_public_key[MLDSA44_PUBLIC_KEY_BYTES],
+    MLDSA44_private_key *out_private_key,
+    const uint8_t entropy[MLDSA_SEED_BYTES]) {
+  if (out_encoded_public_key == nullptr || out_private_key == nullptr) {
+    return bcm_status::failure;
+  }
+  if (BCM_mldsa44_generate_key_external_entropy(out_encoded_public_key,
+                                                out_private_key, entropy) ==
+      bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa44_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa44_private_key_from_seed_fips(
+    MLDSA44_private_key *out_private_key,
+    const uint8_t seed[MLDSA_SEED_BYTES]) {
+  uint8_t public_key[MLDSA44_PUBLIC_KEY_BYTES];
+  if (BCM_mldsa44_generate_key_external_entropy(public_key, out_private_key,
+                                                seed) == bcm_status::failure) {
+    return bcm_status::failure;
+  }
+  return BCM_mldsa44_check_key_fips(out_private_key);
+}
+
+bcm_status bssl::BCM_mldsa44_public_from_private(
+    MLDSA44_public_key *out_public_key,
+    const MLDSA44_private_key *private_key) {
+  const auto *priv = mldsa::private_key_from_external_44(private_key);
+  auto *out_pub = mldsa::public_key_from_external_44(out_public_key);
+  *out_pub = priv->pub;
+  return bcm_status::approved;
+}
+
+const MLDSA44_public_key *bssl::BCM_mldsa44_public_of_private(
+    const MLDSA44_private_key *private_key) {
+  return reinterpret_cast<const MLDSA44_public_key *>(
+      &mldsa::private_key_from_external_44(private_key)->pub);
+}
+
+bcm_status bssl::BCM_mldsa44_sign_internal(
+    uint8_t out_encoded_signature[MLDSA44_SIGNATURE_BYTES],
+    const MLDSA44_private_key *private_key, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context_prefix, size_t context_prefix_len,
+    const uint8_t *context, size_t context_len,
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  return bcm_as_approved_status(mldsa_sign_internal(
+      out_encoded_signature, mldsa::private_key_from_external_44(private_key),
+      msg, msg_len, context_prefix, context_prefix_len, context, context_len,
+      randomizer));
+}
+
+bcm_status bssl::BCM_mldsa44_sign_mu_internal(
+    uint8_t out_encoded_signature[MLDSA44_SIGNATURE_BYTES],
+    const MLDSA44_private_key *private_key,
+    const uint8_t msg_rep[MLDSA_MU_BYTES],
+    const uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES]) {
+  return bcm_as_approved_status(mldsa_sign_mu(
+      out_encoded_signature, mldsa::private_key_from_external_44(private_key),
+      msg_rep, randomizer));
+}
+
+// ML-DSA signature in randomized mode, filling the random bytes with
+// `BCM_rand_bytes`.
+bcm_status bssl::BCM_mldsa44_sign(
+    uint8_t out_encoded_signature[MLDSA44_SIGNATURE_BYTES],
+    const MLDSA44_private_key *private_key, const uint8_t *msg, size_t msg_len,
+    const uint8_t *context, size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+  uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES];
+  BCM_rand_bytes(randomizer, sizeof(randomizer));
+  CONSTTIME_SECRET(randomizer, sizeof(randomizer));
+
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  return BCM_mldsa44_sign_internal(
+      out_encoded_signature, private_key, msg, msg_len, context_prefix,
+      sizeof(context_prefix), context, context_len, randomizer);
+}
+
+// ML-DSA pre-hashed API: initializing a pre-hashing context.
+void bssl::BCM_mldsa44_prehash_init(MLDSA44_prehash *out_prehash_ctx,
+                                    const MLDSA44_public_key *public_key,
+                                    const uint8_t *context,
+                                    size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  mldsa_prehash_init(mldsa::prehash_context_from_external_44(out_prehash_ctx),
+                     mldsa::public_key_from_external_44(public_key),
+                     context_prefix, sizeof(context_prefix), context,
+                     context_len);
+}
+
+// ML-DSA pre-hashed API: updating a pre-hashing context with a message chunk.
+void bssl::BCM_mldsa44_prehash_update(MLDSA44_prehash *inout_prehash_ctx,
+                                      const uint8_t *msg, size_t msg_len) {
+  mldsa_prehash_update(
+      mldsa::prehash_context_from_external_44(inout_prehash_ctx), msg, msg_len);
+}
+
+// ML-DSA pre-hashed API: obtaining a message representative to sign.
+void bssl::BCM_mldsa44_prehash_finalize(uint8_t out_msg_rep[MLDSA_MU_BYTES],
+                                        MLDSA44_prehash *inout_prehash_ctx) {
+  mldsa_prehash_finalize(
+      out_msg_rep, mldsa::prehash_context_from_external_44(inout_prehash_ctx));
+}
+
+// ML-DSA pre-hashed API: signing a message representative.
+bcm_status bssl::BCM_mldsa44_sign_message_representative(
+    uint8_t out_encoded_signature[MLDSA44_SIGNATURE_BYTES],
+    const MLDSA44_private_key *private_key,
+    const uint8_t msg_rep[MLDSA_MU_BYTES]) {
+  uint8_t randomizer[BCM_MLDSA_SIGNATURE_RANDOMIZER_BYTES];
+  BCM_rand_bytes(randomizer, sizeof(randomizer));
+  CONSTTIME_SECRET(randomizer, sizeof(randomizer));
+
+  return bcm_as_approved_status(mldsa_sign_mu(
+      out_encoded_signature, mldsa::private_key_from_external_44(private_key),
+      msg_rep, randomizer));
+}
+
+// ML-DSA pre-hashed API: verifying a message representative.
+bcm_status bssl::BCM_mldsa44_verify_message_representative(
+    const MLDSA44_public_key *public_key,
+    const uint8_t signature[MLDSA44_SIGNATURE_BYTES],
+    const uint8_t msg_rep[MLDSA_MU_BYTES]) {
+  return bcm_as_approved_status(mldsa::mldsa_verify_mu<4, 4>(
+      mldsa::public_key_from_external_44(public_key), signature, msg_rep));
+}
+
+// FIPS 204, Algorithm 3 (`ML-DSA.Verify`).
+bcm_status bssl::BCM_mldsa44_verify(const MLDSA44_public_key *public_key,
+                                    const uint8_t *signature,
+                                    const uint8_t *msg, size_t msg_len,
+                                    const uint8_t *context,
+                                    size_t context_len) {
+  BSSL_CHECK(context_len <= 255);
+  const uint8_t context_prefix[2] = {0, static_cast<uint8_t>(context_len)};
+  return BCM_mldsa44_verify_internal(public_key, signature, msg, msg_len,
+                                     context_prefix, sizeof(context_prefix),
+                                     context, context_len);
+}
+
+bcm_status bssl::BCM_mldsa44_verify_internal(
+    const MLDSA44_public_key *public_key,
+    const uint8_t encoded_signature[MLDSA44_SIGNATURE_BYTES],
+    const uint8_t *msg, size_t msg_len, const uint8_t *context_prefix,
+    size_t context_prefix_len, const uint8_t *context, size_t context_len) {
+  return bcm_as_approved_status(mldsa::mldsa_verify_internal<4, 4>(
+      mldsa::public_key_from_external_44(public_key), encoded_signature, msg,
+      msg_len, context_prefix, context_prefix_len, context, context_len));
+}
+
+bcm_status bssl::BCM_mldsa44_marshal_public_key(
+    CBB *out, const MLDSA44_public_key *public_key) {
+  return bcm_as_approved_status(mldsa_marshal_public_key(
+      out, mldsa::public_key_from_external_44(public_key)));
+}
+
+int bssl::BCM_mldsa44_public_keys_equal(const MLDSA44_public_key *a,
+                                        const MLDSA44_public_key *b) {
+  auto *a_pub = mldsa::public_key_from_external_44(a);
+  auto *b_pub = mldsa::public_key_from_external_44(b);
+  // It is sufficient to compare `public_key_hash`. When importing a public key,
+  // the hash must be computed. When importing a private key in expanded form
+  // (an internal testing-only API), the hash is provided, but we recompute it
+  // and check for correctness.
+  return OPENSSL_memcmp(a_pub->public_key_hash, b_pub->public_key_hash,
+                        sizeof(a_pub->public_key_hash)) == 0;
+}
+
+int bssl::boringssl_self_test_mldsa() {
+  return mldsa::fips::keygen_self_test() && mldsa::fips::sign_self_test() &&
+         mldsa::fips::verify_self_test();
+}