mldsa_gh 0.1.0

data/ext/mldsa_gh_native/vendor/mldsa-native/mldsa/src/poly.c

@@ -0,0 +1,1066 @@
+/*
+ * Copyright (c) The mldsa-native project authors
+ * Copyright (c) The mlkem-native project authors
+ * SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+ */
+
+/* References
+ * ==========
+ *
+ * - [FIPS204]
+ *   FIPS 204 Module-Lattice-Based Digital Signature Standard
+ *   National Institute of Standards and Technology
+ *   https://csrc.nist.gov/pubs/fips/204/final
+ *
+ * - [REF]
+ *   CRYSTALS-Dilithium reference implementation
+ *   Bai, Ducas, Kiltz, Lepoint, Lyubashevsky, Schwabe, Seiler, Stehlé
+ *   https://github.com/pq-crystals/dilithium/tree/master/ref
+ */
+
+#include "poly.h"
+
+#include "common.h"
+#include "ct.h"
+#include "debug.h"
+#include "reduce.h"
+#include "rounding.h"
+#include "symmetric.h"
+
+#if !defined(MLD_CONFIG_MULTILEVEL_NO_SHARED)
+#include "zetas.inc"
+
+MLD_INTERNAL_API
+void mld_poly_reduce(mld_poly *a)
+{
+  unsigned int i;
+  mld_assert_bound(a->coeffs, MLDSA_N, INT32_MIN, MLD_REDUCE32_DOMAIN_MAX);
+
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N)
+    invariant(forall(k0, i, MLDSA_N, a->coeffs[k0] == loop_entry(*a).coeffs[k0]))
+    invariant(array_bound(a->coeffs, 0, i, -MLD_REDUCE32_RANGE_MAX, MLD_REDUCE32_RANGE_MAX))
+    decreases(MLDSA_N - i))
+  {
+    a->coeffs[i] = mld_reduce32(a->coeffs[i]);
+  }
+
+  mld_assert_bound(a->coeffs, MLDSA_N, -MLD_REDUCE32_RANGE_MAX,
+                   MLD_REDUCE32_RANGE_MAX);
+}
+
+MLD_STATIC_TESTABLE void mld_poly_caddq_c(mld_poly *a)
+__contract__(
+  requires(memory_no_alias(a, sizeof(mld_poly)))
+  requires(array_abs_bound(a->coeffs, 0, MLDSA_N, MLDSA_Q))
+  assigns(memory_slice(a, sizeof(mld_poly)))
+  ensures(array_bound(a->coeffs, 0, MLDSA_N, 0, MLDSA_Q))
+)
+{
+  unsigned int i;
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N)
+    invariant(forall(k0, i, MLDSA_N, a->coeffs[k0] == loop_entry(*a).coeffs[k0]))
+    invariant(array_bound(a->coeffs, 0, i, 0, MLDSA_Q))
+    decreases(MLDSA_N - i)
+    )
+  {
+    a->coeffs[i] = mld_caddq(a->coeffs[i]);
+  }
+
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+}
+
+MLD_INTERNAL_API
+void mld_poly_caddq(mld_poly *a)
+{
+#if defined(MLD_USE_NATIVE_POLY_CADDQ)
+  int ret;
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+  ret = mld_poly_caddq_native(a->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+    return;
+  }
+#endif /* MLD_USE_NATIVE_POLY_CADDQ */
+  mld_poly_caddq_c(a);
+}
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API) || !defined(MLD_CONFIG_NO_SIGN_API) || \
+    defined(MLD_CONFIG_REDUCE_RAM) || defined(MLD_UNIT_TEST)
+/* Reference: We use destructive version (output=first input) to avoid
+ *            reasoning about aliasing in the CBMC specification */
+MLD_INTERNAL_API
+void mld_poly_add(mld_poly *r, const mld_poly *b)
+{
+  unsigned int i;
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    assigns(i, memory_slice(r, sizeof(mld_poly)))
+    invariant(i <= MLDSA_N)
+    invariant(forall(k0, i, MLDSA_N, r->coeffs[k0] == loop_entry(*r).coeffs[k0]))
+    invariant(forall(k1, 0, i, r->coeffs[k1] == loop_entry(*r).coeffs[k1] + b->coeffs[k1]))
+    invariant(forall(k2, 0, i, r->coeffs[k2] < MLD_REDUCE32_DOMAIN_MAX))
+    invariant(forall(k2, 0, i, r->coeffs[k2] >= INT32_MIN))
+    decreases(MLDSA_N - i)
+  )
+  {
+    r->coeffs[i] = r->coeffs[i] + b->coeffs[i];
+  }
+}
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API || !MLD_CONFIG_NO_SIGN_API || \
+          MLD_CONFIG_REDUCE_RAM || MLD_UNIT_TEST */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_NO_VERIFY_API)
+/* Reference: We use destructive version (output=first input) to avoid
+ *            reasoning about aliasing in the CBMC specification */
+MLD_INTERNAL_API
+void mld_poly_sub(mld_poly *r, const mld_poly *b)
+{
+  unsigned int i;
+  mld_assert_abs_bound(b->coeffs, MLDSA_N, MLDSA_Q);
+  mld_assert_abs_bound(r->coeffs, MLDSA_N, MLDSA_Q);
+
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N)
+    invariant(array_bound(r->coeffs, 0, i, INT32_MIN, MLD_REDUCE32_DOMAIN_MAX))
+    invariant(forall(k0, i, MLDSA_N, r->coeffs[k0] == loop_entry(*r).coeffs[k0]))
+    decreases(MLDSA_N - i)
+  )
+  {
+    r->coeffs[i] = r->coeffs[i] - b->coeffs[i];
+  }
+
+  mld_assert_bound(r->coeffs, MLDSA_N, INT32_MIN, MLD_REDUCE32_DOMAIN_MAX);
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API || !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_INTERNAL_API
+void mld_poly_shiftl(mld_poly *a)
+{
+  unsigned int i;
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, 1 << 10);
+
+  for (i = 0; i < MLDSA_N; i++)
+  __loop__(
+    invariant(i <= MLDSA_N)
+    invariant(array_bound(a->coeffs, 0, i, 0, MLDSA_Q))
+    invariant(forall(k0, i, MLDSA_N, a->coeffs[k0] == loop_entry(*a).coeffs[k0]))
+    decreases(MLDSA_N - i))
+  {
+    /* Reference: uses a left shift by MLDSA_D which is undefined behaviour in
+     * C90/C99
+     */
+    a->coeffs[i] *= (1 << MLDSA_D);
+  }
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+}
+#endif /* !MLD_CONFIG_NO_VERIFY_API */
+
+static MLD_INLINE int32_t mld_fqmul(int32_t a, int32_t b)
+__contract__(
+  requires(b > -MLDSA_Q_HALF && b < MLDSA_Q_HALF)
+  ensures(return_value > -MLD_FQMUL_BOUND && return_value < MLD_FQMUL_BOUND)
+)
+{
+  /* Bounds: We argue in mld_montgomery_reduce() that the result
+   * of Montgomery reduction is < MLDSA_Q if the input is smaller
+   * than 2^31 * MLDSA_Q in absolute value. Indeed, we have:
+   *
+   *    |a * b|   = |a| * |b|
+   *              < 2^31 * MLDSA_Q_HALF
+   *              < 2^31 * MLDSA_Q
+   *
+   * So the output is < MLDSA_Q < MLD_FQMUL_BOUND.
+   */
+  return mld_montgomery_reduce((int64_t)a * (int64_t)b);
+}
+
+/* mld_ntt_butterfly_block()
+ *
+ * Computes a block CT butterflies with a fixed twiddle factor,
+ * using Montgomery multiplication.
+ *
+ * Parameters:
+ * - r: Pointer to base of polynomial (_not_ the base of butterfly block)
+ * - zeta: Twiddle factor to use for the butterfly. This must be in
+ *         Montgomery form and signed canonical.
+ * - start: Offset to the beginning of the butterfly block
+ * - len: Index difference between coefficients subject to a butterfly
+ * - bound: Ghost variable describing coefficient bound: Prior to `start`,
+ *          coefficients must be bound by `bound + MLDSA_Q`. Post `start`,
+ *          they must be bound by `bound`.
+ * When this function returns, output coefficients in the index range
+ * [start, start+2*len) have bound bumped to `bound + MLDSA_Q`.
+ * Example:
+ * - start=8, len=4
+ *   This would compute the following four butterflies
+ *          8     --    12
+ *             9    --     13
+ *                10   --     14
+ *                   11   --     15
+ * - start=4, len=2
+ *   This would compute the following two butterflies
+ *          4 -- 6
+ *             5 -- 7
+ */
+
+/* Reference: Embedded in `ntt()` in the reference implementation @[REF]. */
+static MLD_INLINE void mld_ntt_butterfly_block(int32_t r[MLDSA_N],
+                                               const int32_t zeta,
+                                               const unsigned start,
+                                               const unsigned len,
+                                               const uint32_t bound)
+__contract__(
+  requires(start < MLDSA_N)
+  requires(1 <= len && len <= MLDSA_N / 2 && start + 2 * len <= MLDSA_N)
+  requires(0 <= bound && bound < INT32_MAX - MLD_FQMUL_BOUND)
+  requires(-MLDSA_Q_HALF < zeta && zeta < MLDSA_Q_HALF)
+  requires(memory_no_alias(r, sizeof(int32_t) * MLDSA_N))
+  requires(array_abs_bound(r, 0, start, bound + MLD_FQMUL_BOUND))
+  requires(array_abs_bound(r, start, MLDSA_N, bound))
+  assigns(memory_slice(r, sizeof(int32_t) * MLDSA_N))
+  ensures(array_abs_bound(r, 0, start + 2*len, bound + MLD_FQMUL_BOUND))
+  ensures(array_abs_bound(r, start + 2 * len, MLDSA_N, bound)))
+{
+  /* `bound` is a ghost variable only needed in the CBMC specification */
+  unsigned j;
+  ((void)bound);
+  for (j = start; j < start + len; j++)
+  __loop__(
+    invariant(start <= j && j <= start + len)
+    /*
+     * Coefficients are updated in strided pairs, so the bounds for the
+     * intermediate states alternate twice between the old and new bound
+     */
+    invariant(array_abs_bound(r, 0,           j,           bound + MLD_FQMUL_BOUND))
+    invariant(array_abs_bound(r, j,           start + len, bound))
+    invariant(array_abs_bound(r, start + len, j + len,     bound + MLD_FQMUL_BOUND))
+    invariant(array_abs_bound(r, j + len,     MLDSA_N,     bound))
+    decreases(start + len - j))
+  {
+    int32_t t;
+    t = mld_fqmul(r[j + len], zeta);
+    r[j + len] = r[j] - t;
+    r[j] = r[j] + t;
+  }
+}
+
+/* mld_ntt_layer()
+ *
+ * Compute one layer of forward NTT
+ *
+ * Parameters:
+ * - r:     Pointer to base of polynomial
+ * - layer: Indicates which layer is being applied.
+ */
+
+/* Reference: Embedded in `ntt()` in the reference implementation @[REF]. */
+static MLD_INLINE void mld_ntt_layer(int32_t r[MLDSA_N], const unsigned layer)
+__contract__(
+  requires(memory_no_alias(r, sizeof(int32_t) * MLDSA_N))
+  requires(1 <= layer && layer <= 8)
+  requires(array_abs_bound(r, 0, MLDSA_N, layer * MLD_FQMUL_BOUND))
+  assigns(memory_slice(r, sizeof(int32_t) * MLDSA_N))
+  ensures(array_abs_bound(r, 0, MLDSA_N, (layer + 1) * MLD_FQMUL_BOUND)))
+{
+  unsigned start, k, len;
+  /* Twiddle factors for layer n are at indices 2^(n-1)..2^n-1. */
+  k = 1u << (layer - 1);
+  len = (unsigned)MLDSA_N >> layer;
+  for (start = 0; start < MLDSA_N; start += 2 * len)
+  __loop__(
+    invariant(start < MLDSA_N + 2 * len)
+    invariant(k <= MLDSA_N)
+    invariant(2 * len * k == start + MLDSA_N)
+    invariant(array_abs_bound(r, 0, start, layer * MLD_FQMUL_BOUND + MLD_FQMUL_BOUND))
+    invariant(array_abs_bound(r, start, MLDSA_N, layer * MLD_FQMUL_BOUND))
+    decreases(MLDSA_N - start))
+  {
+    int32_t zeta = mld_zetas[k++];
+    mld_ntt_butterfly_block(r, zeta, start, len, layer * MLD_FQMUL_BOUND);
+  }
+}
+
+MLD_STATIC_TESTABLE void mld_poly_ntt_c(mld_poly *a)
+__contract__(
+  requires(memory_no_alias(a, sizeof(mld_poly)))
+  requires(array_abs_bound(a->coeffs, 0, MLDSA_N, MLDSA_Q))
+  assigns(memory_slice(a, sizeof(mld_poly)))
+  ensures(array_abs_bound(a->coeffs, 0, MLDSA_N, MLD_NTT_BOUND))
+)
+{
+  unsigned int layer;
+  int32_t *r;
+
+
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+  r = a->coeffs;
+
+  for (layer = 1; layer < 9; layer++)
+  __loop__(
+    invariant(1 <= layer && layer <= 9)
+    invariant(array_abs_bound(r, 0, MLDSA_N, layer * MLD_FQMUL_BOUND))
+    decreases(9 - layer)
+  )
+  {
+    mld_ntt_layer(r, layer);
+  }
+
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLD_NTT_BOUND);
+}
+
+MLD_INTERNAL_API
+void mld_poly_ntt(mld_poly *a)
+{
+#if defined(MLD_USE_NATIVE_NTT)
+  int ret;
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+  ret = mld_ntt_native(a->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_abs_bound(a->coeffs, MLDSA_N, MLD_NTT_BOUND);
+    return;
+  }
+#endif /* MLD_USE_NATIVE_NTT */
+  mld_poly_ntt_c(a);
+}
+
+/**
+ * Scale a field element by mont/256, i.e., perform Montgomery multiplication
+ * by mont^2/256.
+ *
+ * Input is expected to have absolute value smaller than 256 * MLDSA_Q. Output
+ * has absolute value smaller than MLD_INTT_BOUND.
+ *
+ * @param a Field element to be scaled.
+ */
+static MLD_INLINE int32_t mld_fqscale(int32_t a)
+__contract__(
+  requires(a > -256*MLDSA_Q && a < 256*MLDSA_Q)
+  ensures(return_value > -MLD_INTT_BOUND && return_value < MLD_INTT_BOUND)
+)
+{
+  /* check-magic: 41978 == pow(2,64-8,MLDSA_Q) */
+  const int32_t f = 41978;
+  /* Bounds: MLD_INTT_BOUND is MLDSA_Q, so the bounds reasoning is just
+   * a special case of that in mld_fqmul(). */
+  return mld_montgomery_reduce((int64_t)a * f);
+}
+
+/* Reference: Embedded into `invntt_tomont()` in the reference implementation
+ * @[REF] */
+static MLD_INLINE void mld_invntt_layer(int32_t r[MLDSA_N], unsigned layer)
+__contract__(
+  requires(memory_no_alias(r, sizeof(int32_t) * MLDSA_N))
+  requires(1 <= layer && layer <= 8)
+  requires(array_abs_bound(r, 0, MLDSA_N, (MLDSA_N >> layer) * MLDSA_Q))
+  assigns(memory_slice(r, sizeof(int32_t) * MLDSA_N))
+  ensures(array_abs_bound(r, 0, MLDSA_N, (MLDSA_N >> (layer - 1)) * MLDSA_Q)))
+{
+  unsigned start, k, len;
+  len = (unsigned)MLDSA_N >> layer;
+  k = (1u << layer) - 1;
+  for (start = 0; start < MLDSA_N; start += 2 * len)
+  __loop__(
+    invariant(start <= MLDSA_N && k <= 255)
+    invariant(2 * len * k + start == 2 * MLDSA_N - 2 * len)
+    invariant(array_abs_bound(r, 0, start, (MLDSA_N >> (layer - 1)) * MLDSA_Q))
+    invariant(array_abs_bound(r, start, MLDSA_N, (MLDSA_N >> layer) * MLDSA_Q))
+    decreases(MLDSA_N - start))
+  {
+    unsigned j;
+    int32_t zeta = -mld_zetas[k--];
+
+    /* The bound `(MLDSA_N >> (layer - 1)) * MLDSA_Q` is loose enough to
+     * cover both the input bound `(MLDSA_N >> layer) * MLDSA_Q`
+     * (for layers >= 1) and the fqmul output bound `MLD_FQMUL_BOUND`
+     * (which is < 2 * MLDSA_Q <= (MLDSA_N >> (layer - 1)) * MLDSA_Q). */
+    for (j = start; j < start + len; j++)
+    __loop__(
+      invariant(start <= j && j <= start + len)
+      invariant(array_abs_bound(r, 0, start, (MLDSA_N >> (layer - 1)) * MLDSA_Q))
+      invariant(array_abs_bound(r, start, j, (MLDSA_N >> (layer - 1)) * MLDSA_Q))
+      invariant(array_abs_bound(r, j, start + len, (MLDSA_N >> layer) * MLDSA_Q))
+      invariant(array_abs_bound(r, start + len, j + len, (MLDSA_N >> (layer - 1)) * MLDSA_Q))
+      invariant(array_abs_bound(r, j + len, MLDSA_N, (MLDSA_N >> layer) * MLDSA_Q))
+      decreases(start + len - j))
+    {
+      int32_t t = r[j];
+      r[j] = t + r[j + len];
+      r[j + len] = t - r[j + len];
+      r[j + len] = mld_fqmul(r[j + len], zeta);
+    }
+  }
+}
+
+MLD_STATIC_TESTABLE void mld_poly_invntt_tomont_c(mld_poly *a)
+__contract__(
+  requires(memory_no_alias(a, sizeof(mld_poly)))
+  requires(array_abs_bound(a->coeffs, 0, MLDSA_N, MLDSA_Q))
+  assigns(memory_slice(a, sizeof(mld_poly)))
+  ensures(array_abs_bound(a->coeffs, 0, MLDSA_N, MLD_INTT_BOUND))
+)
+{
+  unsigned int layer, j;
+  int32_t *r;
+
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+
+  r = a->coeffs;
+  for (layer = 8; layer >= 1; layer--)
+  __loop__(
+    invariant(layer <= 8)
+    /* Absolute bounds increase from 1Q before layer 8 */
+    /* up to 256Q after layer 1                        */
+    invariant(array_abs_bound(r, 0, MLDSA_N, (MLDSA_N >> layer) * MLDSA_Q))
+    decreases(layer))
+  {
+    mld_invntt_layer(r, layer);
+  }
+
+  /* Coefficient bounds are now at 256Q. We now scale by mont / 256,
+   * i.e., compute the Montgomery multiplication by mont^2 / 256.
+   * mont corrects the mont^-1  factor introduced in the basemul.
+   * 1/256 performs that scaling of the inverse NTT.
+   * The reduced value is bounded by  MLD_INTT_BOUND in absolute
+   * value.*/
+  for (j = 0; j < MLDSA_N; ++j)
+  __loop__(
+    invariant(j <= MLDSA_N)
+    invariant(array_abs_bound(r, 0, j, MLD_INTT_BOUND))
+    invariant(array_abs_bound(r, j, MLDSA_N, MLDSA_N * MLDSA_Q))
+    decreases(MLDSA_N - j)
+  )
+  {
+    r[j] = mld_fqscale(r[j]);
+  }
+
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLD_INTT_BOUND);
+}
+
+
+MLD_INTERNAL_API
+void mld_poly_invntt_tomont(mld_poly *a)
+{
+#if defined(MLD_USE_NATIVE_INTT)
+  int ret;
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+  ret = mld_intt_native(a->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_abs_bound(a->coeffs, MLDSA_N, MLD_INTT_BOUND);
+    return;
+  }
+#endif /* MLD_USE_NATIVE_INTT */
+  mld_poly_invntt_tomont_c(a);
+}
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_NO_VERIFY_API) || \
+    defined(MLD_CONFIG_REDUCE_RAM) || defined(MLD_UNIT_TEST)
+MLD_STATIC_TESTABLE void mld_poly_pointwise_montgomery_c(mld_poly *a,
+                                                         const mld_poly *b)
+__contract__(
+  requires(memory_no_alias(a, sizeof(mld_poly)))
+  requires(memory_no_alias(b, sizeof(mld_poly)))
+  requires(array_abs_bound(a->coeffs, 0, MLDSA_N, MLD_NTT_BOUND))
+  requires(array_abs_bound(b->coeffs, 0, MLDSA_N, MLD_NTT_BOUND))
+  assigns(memory_slice(a, sizeof(mld_poly)))
+  ensures(array_abs_bound(a->coeffs, 0, MLDSA_N, MLDSA_Q))
+)
+{
+  unsigned int i;
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLD_NTT_BOUND);
+  mld_assert_abs_bound(b->coeffs, MLDSA_N, MLD_NTT_BOUND);
+
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N)
+    invariant(array_abs_bound(a->coeffs, 0, i, MLDSA_Q))
+    invariant(array_abs_bound(a->coeffs, i, MLDSA_N, MLD_NTT_BOUND))
+    decreases(MLDSA_N - i)
+  )
+  {
+    a->coeffs[i] = mld_montgomery_reduce((int64_t)a->coeffs[i] * b->coeffs[i]);
+  }
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+}
+
+MLD_INTERNAL_API
+void mld_poly_pointwise_montgomery(mld_poly *a, const mld_poly *b)
+{
+#if defined(MLD_USE_NATIVE_POINTWISE_MONTGOMERY)
+  int ret;
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLD_NTT_BOUND);
+  mld_assert_abs_bound(b->coeffs, MLDSA_N, MLD_NTT_BOUND);
+  ret = mld_poly_pointwise_montgomery_native(a->coeffs, b->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_Q);
+    return;
+  }
+#endif /* MLD_USE_NATIVE_POINTWISE_MONTGOMERY */
+  mld_poly_pointwise_montgomery_c(a, b);
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API || !MLD_CONFIG_NO_VERIFY_API || \
+          MLD_CONFIG_REDUCE_RAM || MLD_UNIT_TEST */
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API)
+MLD_INTERNAL_API
+void mld_poly_power2round(mld_poly *a1, mld_poly *a0, const mld_poly *a)
+{
+  unsigned int i;
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    assigns(i, memory_slice(a0, sizeof(mld_poly)), memory_slice(a1, sizeof(mld_poly)))
+    invariant(i <= MLDSA_N)
+    invariant(forall(k0, i, MLDSA_N, a->coeffs[k0] == loop_entry(*a).coeffs[k0]))
+    invariant(array_bound(a0->coeffs, 0, i, -(MLD_2_POW_D/2)+1, (MLD_2_POW_D/2)+1))
+    invariant(array_bound(a1->coeffs, 0, i, 0, ((MLDSA_Q - 1) / MLD_2_POW_D) + 1))
+    decreases(MLDSA_N - i)
+  )
+  {
+    mld_power2round(&a0->coeffs[i], &a1->coeffs[i], a->coeffs[i]);
+  }
+
+  mld_assert_bound(a0->coeffs, MLDSA_N, -(MLD_2_POW_D / 2) + 1,
+                   (MLD_2_POW_D / 2) + 1);
+  mld_assert_bound(a1->coeffs, MLDSA_N, 0, ((MLDSA_Q - 1) / MLD_2_POW_D) + 1);
+}
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API */
+
+#ifndef MLD_POLY_UNIFORM_NBLOCKS
+#define MLD_POLY_UNIFORM_NBLOCKS \
+  ((768 + MLD_STREAM128_BLOCKBYTES - 1) / MLD_STREAM128_BLOCKBYTES)
+#endif
+/* Reference: `mld_rej_uniform()` in the reference implementation @[REF].
+ *            - Our signature differs from the reference implementation
+ *              in that it adds the offset and always expects the base of the
+ *              target buffer. This avoids shifting the buffer base in the
+ *              caller, which appears tricky to reason about. */
+MLD_STATIC_TESTABLE unsigned int mld_rej_uniform_c(int32_t *a,
+                                                   unsigned int target,
+                                                   unsigned int offset,
+                                                   const uint8_t *buf,
+                                                   unsigned int buflen)
+__contract__(
+  requires(offset <= target && target <= MLDSA_N)
+  requires(buflen <= (MLD_POLY_UNIFORM_NBLOCKS * MLD_STREAM128_BLOCKBYTES) && buflen % 3 == 0)
+  requires(memory_no_alias(a, sizeof(int32_t) * target))
+  requires(memory_no_alias(buf, buflen))
+  requires(array_bound(a, 0, offset, 0, MLDSA_Q))
+  assigns(memory_slice(a, sizeof(int32_t) * target))
+  ensures(offset <= return_value && return_value <= target)
+  ensures(array_bound(a, 0, return_value, 0, MLDSA_Q))
+)
+{
+  unsigned int ctr, pos;
+  uint32_t t;
+  mld_assert_bound(a, offset, 0, MLDSA_Q);
+
+  ctr = offset;
+  pos = 0;
+  /* pos + 3 cannot overflow due to the assumption
+  buflen <= (MLD_POLY_UNIFORM_NBLOCKS * MLD_STREAM128_BLOCKBYTES) */
+  while (ctr < target && pos + 3 <= buflen)
+  __loop__(
+    invariant(offset <= ctr && ctr <= target && pos <= buflen)
+    invariant(array_bound(a, 0, ctr, 0, MLDSA_Q))
+    decreases(buflen - pos))
+  {
+    t = buf[pos++];
+    t |= (uint32_t)buf[pos++] << 8;
+    t |= (uint32_t)buf[pos++] << 16;
+    t &= 0x7FFFFF;
+
+    if (t < MLDSA_Q)
+    {
+      a[ctr++] = (int32_t)t;
+    }
+  }
+
+  mld_assert_bound(a, ctr, 0, MLDSA_Q);
+
+  return ctr;
+}
+/**
+ * Sample uniformly random coefficients in [0, MLDSA_Q-1] by performing
+ * rejection sampling on an array of random bytes.
+ *
+ * @param[out] a      Pointer to output array (allocated).
+ * @param      target Requested number of coefficients to sample.
+ * @param      offset Number of coefficients already sampled.
+ * @param[in]  buf    Array of random bytes to sample from.
+ * @param      buflen Length of array of random bytes (must be multiple of 3).
+ *
+ * @return Number of sampled coefficients. Can be smaller than len if not
+ *         enough random bytes were given.
+ */
+
+/* Reference: `mld_rej_uniform()` in the reference implementation @[REF].
+ *            - Our signature differs from the reference implementation
+ *              in that it adds the offset and always expects the base of the
+ *              target buffer. This avoids shifting the buffer base in the
+ *              caller, which appears tricky to reason about. */
+static unsigned int mld_rej_uniform(int32_t *a, unsigned int target,
+                                    unsigned int offset, const uint8_t *buf,
+                                    unsigned int buflen)
+__contract__(
+  requires(offset <= target && target <= MLDSA_N)
+  requires(buflen <= (MLD_POLY_UNIFORM_NBLOCKS * MLD_STREAM128_BLOCKBYTES) && buflen % 3 == 0)
+  requires(memory_no_alias(a, sizeof(int32_t) * target))
+  requires(memory_no_alias(buf, buflen))
+  requires(array_bound(a, 0, offset, 0, MLDSA_Q))
+  assigns(memory_slice(a, sizeof(int32_t) * target))
+  ensures(offset <= return_value && return_value <= target)
+  ensures(array_bound(a, 0, return_value, 0, MLDSA_Q))
+)
+{
+#if defined(MLD_USE_NATIVE_REJ_UNIFORM)
+  int ret;
+  mld_assert_bound(a, offset, 0, MLDSA_Q);
+  if (offset == 0)
+  {
+    ret = mld_rej_uniform_native(a, target, buf, buflen);
+    if (ret != MLD_NATIVE_FUNC_FALLBACK)
+    {
+      unsigned res = (unsigned)ret;
+      mld_assert_bound(a, res, 0, MLDSA_Q);
+      return res;
+    }
+  }
+#endif /* MLD_USE_NATIVE_REJ_UNIFORM */
+
+  return mld_rej_uniform_c(a, target, offset, buf, buflen);
+}
+
+/* Reference: poly_uniform() in the reference implementation @[REF].
+ *           - Simplified from reference by removing buffer tail handling
+ *             since buflen % 3 = 0 always holds true (MLD_STREAM128_BLOCKBYTES
+ * = 168).
+ *           - Modified rej_uniform interface to track offset directly.
+ *           - Pass nonce packed in the extended seed array instead of a third
+ *             argument.
+ * */
+MLD_INTERNAL_API
+void mld_poly_uniform(mld_poly *a, const uint8_t seed[MLDSA_SEEDBYTES + 2])
+{
+  unsigned int ctr;
+  unsigned int buflen = MLD_POLY_UNIFORM_NBLOCKS * MLD_STREAM128_BLOCKBYTES;
+  MLD_ALIGN uint8_t buf[MLD_POLY_UNIFORM_NBLOCKS * MLD_STREAM128_BLOCKBYTES];
+  mld_xof128_ctx state;
+
+  mld_xof128_init(&state);
+  mld_xof128_absorb_once(&state, seed, MLDSA_SEEDBYTES + 2);
+  mld_xof128_squeezeblocks(buf, MLD_POLY_UNIFORM_NBLOCKS, &state);
+
+  ctr = mld_rej_uniform(a->coeffs, MLDSA_N, 0, buf, buflen);
+  buflen = MLD_STREAM128_BLOCKBYTES;
+  while (ctr < MLDSA_N)
+  __loop__(
+    assigns(ctr, state, memory_slice(a, sizeof(mld_poly)), object_whole(buf))
+    invariant(ctr <= MLDSA_N)
+    invariant(array_bound(a->coeffs, 0, ctr, 0, MLDSA_Q))
+    invariant(state.pos <= SHAKE128_RATE)
+  )
+  {
+    mld_xof128_squeezeblocks(buf, 1, &state);
+    ctr = mld_rej_uniform(a->coeffs, MLDSA_N, ctr, buf, buflen);
+  }
+  mld_xof128_release(&state);
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(buf, sizeof(buf));
+}
+
+#if !defined(MLD_CONFIG_SERIAL_FIPS202_ONLY) && \
+    (!defined(MLD_CONFIG_REDUCE_RAM) || defined(MLD_UNIT_TEST))
+MLD_INTERNAL_API
+void mld_poly_uniform_4x(mld_poly *vec0, mld_poly *vec1, mld_poly *vec2,
+                         mld_poly *vec3,
+                         uint8_t seed[4][MLD_ALIGN_UP(MLDSA_SEEDBYTES + 2)])
+{
+  /* Temporary buffers for XOF output before rejection sampling */
+  MLD_ALIGN uint8_t
+      buf[4][MLD_ALIGN_UP(MLD_POLY_UNIFORM_NBLOCKS * MLD_STREAM128_BLOCKBYTES)];
+
+  /* Tracks the number of coefficients we have already sampled */
+  unsigned ctr[4];
+  mld_xof128_x4_ctx state;
+  unsigned buflen;
+
+  mld_xof128_x4_init(&state);
+  mld_xof128_x4_absorb(&state, seed, MLDSA_SEEDBYTES + 2);
+
+  /*
+   * Initially, squeeze heuristic number of MLD_POLY_UNIFORM_NBLOCKS.
+   * This should generate the matrix entries with high probability.
+   */
+
+  mld_xof128_x4_squeezeblocks(buf, MLD_POLY_UNIFORM_NBLOCKS, &state);
+  buflen = MLD_POLY_UNIFORM_NBLOCKS * MLD_STREAM128_BLOCKBYTES;
+  ctr[0] = mld_rej_uniform(vec0->coeffs, MLDSA_N, 0, buf[0], buflen);
+  ctr[1] = mld_rej_uniform(vec1->coeffs, MLDSA_N, 0, buf[1], buflen);
+  ctr[2] = mld_rej_uniform(vec2->coeffs, MLDSA_N, 0, buf[2], buflen);
+  ctr[3] = mld_rej_uniform(vec3->coeffs, MLDSA_N, 0, buf[3], buflen);
+
+  /*
+   * So long as not all matrix entries have been generated, squeeze
+   * one more block a time until we're done.
+   */
+  buflen = MLD_STREAM128_BLOCKBYTES;
+  while (ctr[0] < MLDSA_N || ctr[1] < MLDSA_N || ctr[2] < MLDSA_N ||
+         ctr[3] < MLDSA_N)
+  __loop__(
+    assigns(ctr, state, object_whole(buf),
+            memory_slice(vec0, sizeof(mld_poly)), memory_slice(vec1, sizeof(mld_poly)),
+            memory_slice(vec2, sizeof(mld_poly)), memory_slice(vec3, sizeof(mld_poly)))
+    invariant(ctr[0] <= MLDSA_N && ctr[1] <= MLDSA_N)
+    invariant(ctr[2] <= MLDSA_N && ctr[3] <= MLDSA_N)
+    invariant(array_bound(vec0->coeffs, 0, ctr[0], 0, MLDSA_Q))
+    invariant(array_bound(vec1->coeffs, 0, ctr[1], 0, MLDSA_Q))
+    invariant(array_bound(vec2->coeffs, 0, ctr[2], 0, MLDSA_Q))
+    invariant(array_bound(vec3->coeffs, 0, ctr[3], 0, MLDSA_Q)))
+  {
+    mld_xof128_x4_squeezeblocks(buf, 1, &state);
+    ctr[0] = mld_rej_uniform(vec0->coeffs, MLDSA_N, ctr[0], buf[0], buflen);
+    ctr[1] = mld_rej_uniform(vec1->coeffs, MLDSA_N, ctr[1], buf[1], buflen);
+    ctr[2] = mld_rej_uniform(vec2->coeffs, MLDSA_N, ctr[2], buf[2], buflen);
+    ctr[3] = mld_rej_uniform(vec3->coeffs, MLDSA_N, ctr[3], buf[3], buflen);
+  }
+  mld_xof128_x4_release(&state);
+
+  mld_assert_bound(vec0->coeffs, MLDSA_N, 0, MLDSA_Q);
+  mld_assert_bound(vec1->coeffs, MLDSA_N, 0, MLDSA_Q);
+  mld_assert_bound(vec2->coeffs, MLDSA_N, 0, MLDSA_Q);
+  mld_assert_bound(vec3->coeffs, MLDSA_N, 0, MLDSA_Q);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(buf, sizeof(buf));
+}
+
+#endif /* !MLD_CONFIG_SERIAL_FIPS202_ONLY && (!MLD_CONFIG_REDUCE_RAM || \
+          MLD_UNIT_TEST) */
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API)
+MLD_INTERNAL_API
+void mld_polyt1_pack(uint8_t r[MLDSA_POLYT1_PACKEDBYTES], const mld_poly *a)
+{
+  unsigned int i;
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, 1 << 10);
+
+  for (i = 0; i < MLDSA_N / 4; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/4)
+    decreases(MLDSA_N / 4 - i))
+  {
+    r[5 * i + 0] = (uint8_t)((a->coeffs[4 * i + 0] >> 0) & 0xFF);
+    r[5 * i + 1] =
+        (uint8_t)(((a->coeffs[4 * i + 0] >> 8) | (a->coeffs[4 * i + 1] << 2)) &
+                  0xFF);
+    r[5 * i + 2] =
+        (uint8_t)(((a->coeffs[4 * i + 1] >> 6) | (a->coeffs[4 * i + 2] << 4)) &
+                  0xFF);
+    r[5 * i + 3] =
+        (uint8_t)(((a->coeffs[4 * i + 2] >> 4) | (a->coeffs[4 * i + 3] << 6)) &
+                  0xFF);
+    r[5 * i + 4] = (uint8_t)((a->coeffs[4 * i + 3] >> 2) & 0xFF);
+  }
+}
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API */
+
+#if !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_INTERNAL_API
+void mld_polyt1_unpack(mld_poly *r, const uint8_t a[MLDSA_POLYT1_PACKEDBYTES])
+{
+  unsigned int i;
+
+  for (i = 0; i < MLDSA_N / 4; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/4)
+    invariant(array_bound(r->coeffs, 0, i*4, 0, 1 << 10))
+    decreases(MLDSA_N / 4 - i))
+  {
+    r->coeffs[4 * i + 0] =
+        ((a[5 * i + 0] >> 0) | ((int32_t)a[5 * i + 1] << 8)) & 0x3FF;
+    r->coeffs[4 * i + 1] =
+        ((a[5 * i + 1] >> 2) | ((int32_t)a[5 * i + 2] << 6)) & 0x3FF;
+    r->coeffs[4 * i + 2] =
+        ((a[5 * i + 2] >> 4) | ((int32_t)a[5 * i + 3] << 4)) & 0x3FF;
+    r->coeffs[4 * i + 3] =
+        ((a[5 * i + 3] >> 6) | ((int32_t)a[5 * i + 4] << 2)) & 0x3FF;
+  }
+
+  mld_assert_bound(r->coeffs, MLDSA_N, 0, 1 << 10);
+}
+#endif /* !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API)
+MLD_INTERNAL_API
+void mld_polyt0_pack(uint8_t r[MLDSA_POLYT0_PACKEDBYTES], const mld_poly *a)
+{
+  unsigned int i;
+  uint32_t t[8];
+
+  mld_assert_bound(a->coeffs, MLDSA_N, -(1 << (MLDSA_D - 1)) + 1,
+                   (1 << (MLDSA_D - 1)) + 1);
+
+  for (i = 0; i < MLDSA_N / 8; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/8)
+    decreases(MLDSA_N / 8 - i))
+  {
+    /* Safety: a->coeffs[i] <= (1 << (MLDSA_D - 1) as they are output of
+     * power2round, hence, these casts are safe. */
+    t[0] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 0]);
+    t[1] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 1]);
+    t[2] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 2]);
+    t[3] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 3]);
+    t[4] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 4]);
+    t[5] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 5]);
+    t[6] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 6]);
+    t[7] = (uint32_t)((1 << (MLDSA_D - 1)) - a->coeffs[8 * i + 7]);
+
+    r[13 * i + 0] = (uint8_t)((t[0]) & 0xFF);
+    r[13 * i + 1] = (uint8_t)((t[0] >> 8) & 0xFF);
+    r[13 * i + 1] |= (uint8_t)((t[1] << 5) & 0xFF);
+    r[13 * i + 2] = (uint8_t)((t[1] >> 3) & 0xFF);
+    r[13 * i + 3] = (uint8_t)((t[1] >> 11) & 0xFF);
+    r[13 * i + 3] |= (uint8_t)((t[2] << 2) & 0xFF);
+    r[13 * i + 4] = (uint8_t)((t[2] >> 6) & 0xFF);
+    r[13 * i + 4] |= (uint8_t)((t[3] << 7) & 0xFF);
+    r[13 * i + 5] = (uint8_t)((t[3] >> 1) & 0xFF);
+    r[13 * i + 6] = (uint8_t)((t[3] >> 9) & 0xFF);
+    r[13 * i + 6] |= (uint8_t)((t[4] << 4) & 0xFF);
+    r[13 * i + 7] = (uint8_t)((t[4] >> 4) & 0xFF);
+    r[13 * i + 8] = (uint8_t)((t[4] >> 12) & 0xFF);
+    r[13 * i + 8] |= (uint8_t)((t[5] << 1) & 0xFF);
+    r[13 * i + 9] = (uint8_t)((t[5] >> 7) & 0xFF);
+    r[13 * i + 9] |= (uint8_t)((t[6] << 6) & 0xFF);
+    r[13 * i + 10] = (uint8_t)((t[6] >> 2) & 0xFF);
+    r[13 * i + 11] = (uint8_t)((t[6] >> 10) & 0xFF);
+    r[13 * i + 11] |= (uint8_t)((t[7] << 3) & 0xFF);
+    r[13 * i + 12] = (uint8_t)((t[7] >> 5) & 0xFF);
+  }
+}
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || defined(MLD_UNIT_TEST)
+MLD_INTERNAL_API
+void mld_polyt0_unpack(mld_poly *r, const uint8_t a[MLDSA_POLYT0_PACKEDBYTES])
+{
+  unsigned int i;
+
+  for (i = 0; i < MLDSA_N / 8; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/8)
+    invariant(array_bound(r->coeffs, 0, i*8, -(1<<(MLDSA_D-1)) + 1, (1<<(MLDSA_D-1)) + 1))
+    decreases(MLDSA_N / 8 - i))
+  {
+    r->coeffs[8 * i + 0] = a[13 * i + 0];
+    r->coeffs[8 * i + 0] |= (int32_t)a[13 * i + 1] << 8;
+    r->coeffs[8 * i + 0] &= 0x1FFF;
+
+    r->coeffs[8 * i + 1] = a[13 * i + 1] >> 5;
+    r->coeffs[8 * i + 1] |= (int32_t)a[13 * i + 2] << 3;
+    r->coeffs[8 * i + 1] |= (int32_t)a[13 * i + 3] << 11;
+    r->coeffs[8 * i + 1] &= 0x1FFF;
+
+    r->coeffs[8 * i + 2] = a[13 * i + 3] >> 2;
+    r->coeffs[8 * i + 2] |= (int32_t)a[13 * i + 4] << 6;
+    r->coeffs[8 * i + 2] &= 0x1FFF;
+
+    r->coeffs[8 * i + 3] = a[13 * i + 4] >> 7;
+    r->coeffs[8 * i + 3] |= (int32_t)a[13 * i + 5] << 1;
+    r->coeffs[8 * i + 3] |= (int32_t)a[13 * i + 6] << 9;
+    r->coeffs[8 * i + 3] &= 0x1FFF;
+
+    r->coeffs[8 * i + 4] = a[13 * i + 6] >> 4;
+    r->coeffs[8 * i + 4] |= (int32_t)a[13 * i + 7] << 4;
+    r->coeffs[8 * i + 4] |= (int32_t)a[13 * i + 8] << 12;
+    r->coeffs[8 * i + 4] &= 0x1FFF;
+
+    r->coeffs[8 * i + 5] = a[13 * i + 8] >> 1;
+    r->coeffs[8 * i + 5] |= (int32_t)a[13 * i + 9] << 7;
+    r->coeffs[8 * i + 5] &= 0x1FFF;
+
+    r->coeffs[8 * i + 6] = a[13 * i + 9] >> 6;
+    r->coeffs[8 * i + 6] |= (int32_t)a[13 * i + 10] << 2;
+    r->coeffs[8 * i + 6] |= (int32_t)a[13 * i + 11] << 10;
+    r->coeffs[8 * i + 6] &= 0x1FFF;
+
+    r->coeffs[8 * i + 7] = a[13 * i + 11] >> 3;
+    r->coeffs[8 * i + 7] |= (int32_t)a[13 * i + 12] << 5;
+    r->coeffs[8 * i + 7] &= 0x1FFF;
+
+    r->coeffs[8 * i + 0] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 0];
+    r->coeffs[8 * i + 1] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 1];
+    r->coeffs[8 * i + 2] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 2];
+    r->coeffs[8 * i + 3] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 3];
+    r->coeffs[8 * i + 4] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 4];
+    r->coeffs[8 * i + 5] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 5];
+    r->coeffs[8 * i + 6] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 6];
+    r->coeffs[8 * i + 7] = (1 << (MLDSA_D - 1)) - r->coeffs[8 * i + 7];
+  }
+
+  mld_assert_bound(r->coeffs, MLDSA_N, -(1 << (MLDSA_D - 1)) + 1,
+                   (1 << (MLDSA_D - 1)) + 1);
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API || MLD_UNIT_TEST */
+
+MLD_STATIC_TESTABLE uint32_t mld_poly_chknorm_c(const mld_poly *a, int32_t B)
+__contract__(
+  requires(memory_no_alias(a, sizeof(mld_poly)))
+  requires(0 <= B && B <= MLDSA_Q - MLD_REDUCE32_RANGE_MAX)
+  requires(array_bound(a->coeffs, 0, MLDSA_N, -MLD_REDUCE32_RANGE_MAX, MLD_REDUCE32_RANGE_MAX))
+  ensures(return_value == 0 || return_value == 0xFFFFFFFF)
+  ensures((return_value == 0) == array_abs_bound(a->coeffs, 0, MLDSA_N, B))
+)
+{
+  unsigned int i;
+  uint32_t t = 0;
+  mld_assert_bound(a->coeffs, MLDSA_N, -MLD_REDUCE32_RANGE_MAX,
+                   MLD_REDUCE32_RANGE_MAX);
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N)
+    invariant(t == 0 || t == 0xFFFFFFFF)
+    invariant((t == 0) == array_abs_bound(a->coeffs, 0, i, B))
+    decreases(MLDSA_N - i)
+  )
+  {
+    /*
+     * Since we know that -MLD_REDUCE32_RANGE_MAX <= a < MLD_REDUCE32_RANGE_MAX,
+     * and B <= MLDSA_Q - MLD_REDUCE32_RANGE_MAX, to check if
+     * -B < (a mod± MLDSA_Q) < B, it suffices to check if -B < a < B.
+     *
+     * We prove this to be true using the following CBMC assertions.
+     * a ==> b expressed as !a || b to also allow run-time assertion.
+     */
+    mld_assert(a->coeffs[i] < B || a->coeffs[i] - MLDSA_Q <= -B);
+    mld_assert(a->coeffs[i] > -B || a->coeffs[i] + MLDSA_Q >= B);
+
+    /* Reference: Leaks which coefficient violates the bound via a conditional.
+     * We are more conservative to reduce the number of declassifications in
+     * constant-time testing.
+     */
+
+    /* if (abs(a[i]) >= B) */
+    t |= mld_ct_cmask_neg_i32(B - 1 - mld_ct_abs_i32(a->coeffs[i]));
+  }
+
+  return t;
+}
+
+/* Reference: explicitly checks the bound B to be <= (MLDSA_Q - 1) / 8).
+ * This is unnecessary as it's always a compile-time constant.
+ * We instead model it as a precondition.
+ * Checking the bound is performed using a conditional arguing
+ * that it is okay to leak which coefficient violates the bound (while the
+ * coefficient itself must remain secret).
+ * We instead perform everything in constant-time.
+ * Also it is sufficient to check that it is smaller than
+ * MLDSA_Q - MLD_REDUCE32_RANGE_MAX > (MLDSA_Q - 1) / 8).
+ */
+MLD_INTERNAL_API
+uint32_t mld_poly_chknorm(const mld_poly *a, int32_t B)
+{
+#if defined(MLD_USE_NATIVE_POLY_CHKNORM)
+  int ret;
+  int success;
+  mld_assert_bound(a->coeffs, MLDSA_N, -MLD_REDUCE32_RANGE_MAX,
+                   MLD_REDUCE32_RANGE_MAX);
+  /* The native backend returns 0 if all coefficients are within the bound,
+   * 1 if at least one coefficient exceeds the bound, and
+   * -1 (MLD_NATIVE_FUNC_FALLBACK) if the platform does not have the
+   * required capabilities to run the native function.
+   */
+  ret = mld_poly_chknorm_native(a->coeffs, B);
+
+  success = (ret != MLD_NATIVE_FUNC_FALLBACK);
+  /* Constant-time: It would be fine to leak the return value of chknorm
+   * entirely (as it is fine to leak if any coefficient exceeded the bound or
+   * not). However, it is cleaner to perform declassification in sign.c.
+   * Hence, here we only declassify if the native function returned
+   * MLD_NATIVE_FUNC_FALLBACK or not (which solely depends on system
+   * capabilities).
+   */
+  MLD_CT_TESTING_DECLASSIFY(&success, sizeof(int));
+  if (success)
+  {
+    /* Convert 0 / 1 to 0 / 0xFFFFFFFF here */
+    return mld_ct_cmask_nonzero_u32((uint32_t)ret);
+  }
+#endif /* MLD_USE_NATIVE_POLY_CHKNORM */
+  return mld_poly_chknorm_c(a, B);
+}
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_NO_VERIFY_API)
+#if defined(MLD_CONFIG_MULTILEVEL_WITH_SHARED) || \
+    MLD_CONFIG_PARAMETER_SET == 44
+MLD_INTERNAL_API
+void mld_polyw1_pack_88(uint8_t r[MLDSA_POLYW1_PACKEDBYTES_88],
+                        const mld_poly *a)
+{
+  unsigned int i;
+
+  mld_assert_bound(a->coeffs, MLDSA_N, 0,
+                   (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2_88));
+
+  for (i = 0; i < MLDSA_N / 4; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/4)
+    decreases(MLDSA_N / 4 - i))
+  {
+    r[3 * i + 0] = (uint8_t)((a->coeffs[4 * i + 0]) & 0xFF);
+    r[3 * i + 0] |= (uint8_t)((a->coeffs[4 * i + 1] << 6) & 0xFF);
+    r[3 * i + 1] = (uint8_t)((a->coeffs[4 * i + 1] >> 2) & 0xFF);
+    r[3 * i + 1] |= (uint8_t)((a->coeffs[4 * i + 2] << 4) & 0xFF);
+    r[3 * i + 2] = (uint8_t)((a->coeffs[4 * i + 2] >> 4) & 0xFF);
+    r[3 * i + 2] |= (uint8_t)((a->coeffs[4 * i + 3] << 2) & 0xFF);
+  }
+}
+#endif /* MLD_CONFIG_MULTILEVEL_WITH_SHARED || MLD_CONFIG_PARAMETER_SET == 44 \
+        */
+
+#if defined(MLD_CONFIG_MULTILEVEL_WITH_SHARED) || \
+    (MLD_CONFIG_PARAMETER_SET == 65 || MLD_CONFIG_PARAMETER_SET == 87)
+MLD_INTERNAL_API
+void mld_polyw1_pack_32(uint8_t r[MLDSA_POLYW1_PACKEDBYTES_32],
+                        const mld_poly *a)
+{
+  unsigned int i;
+
+  mld_assert_bound(a->coeffs, MLDSA_N, 0,
+                   (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2_32));
+
+  for (i = 0; i < MLDSA_N / 2; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/2)
+    decreases(MLDSA_N / 2 - i))
+  {
+    r[i] =
+        (uint8_t)((a->coeffs[2 * i + 0] | (a->coeffs[2 * i + 1] << 4)) & 0xFF);
+  }
+}
+#endif /* MLD_CONFIG_MULTILEVEL_WITH_SHARED || MLD_CONFIG_PARAMETER_SET == 65 \
+          || MLD_CONFIG_PARAMETER_SET == 87 */
+#endif /* !MLD_CONFIG_NO_SIGN_API || !MLD_CONFIG_NO_VERIFY_API */
+
+#else  /* !MLD_CONFIG_MULTILEVEL_NO_SHARED */
+MLD_EMPTY_CU(mld_poly)
+#endif /* MLD_CONFIG_MULTILEVEL_NO_SHARED */
+
+/* To facilitate single-compilation-unit (SCU) builds, undefine all macros.
+ * Don't modify by hand -- this is auto-generated by scripts/autogen. */
+#undef MLD_POLY_UNIFORM_NBLOCKS