mldsa_gh 0.1.0

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

@@ -0,0 +1,910 @@
+/*
+ * 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_kl.h"
+
+#include "ct.h"
+#include "debug.h"
+#include "rounding.h"
+#include "symmetric.h"
+
+/* Parameter set namespacing
+ * This is to facilitate building multiple instances
+ * of mldsa-native (e.g. with varying parameter sets)
+ * within a single compilation unit. */
+#define mld_rej_eta MLD_ADD_PARAM_SET(mld_rej_eta)
+#define mld_rej_eta_c MLD_ADD_PARAM_SET(mld_rej_eta_c)
+#define mld_poly_decompose_c MLD_ADD_PARAM_SET(mld_poly_decompose_c)
+#define mld_poly_use_hint_c MLD_ADD_PARAM_SET(mld_poly_use_hint_c)
+#define mld_polyz_unpack_c MLD_ADD_PARAM_SET(mld_polyz_unpack_c)
+/* End of parameter set namespacing */
+
+
+#if !defined(MLD_CONFIG_NO_SIGN_API)
+MLD_STATIC_TESTABLE
+void mld_poly_decompose_c(mld_poly *a1, mld_poly *a0)
+__contract__(
+  requires(memory_no_alias(a1,  sizeof(mld_poly)))
+  requires(memory_no_alias(a0, sizeof(mld_poly)))
+  requires(array_bound(a0->coeffs, 0, MLDSA_N, 0, MLDSA_Q))
+  assigns(memory_slice(a1, sizeof(mld_poly)))
+  assigns(memory_slice(a0, sizeof(mld_poly)))
+  ensures(array_bound(a1->coeffs, 0, MLDSA_N, 0, (MLDSA_Q-1)/(2*MLDSA_GAMMA2)))
+  ensures(array_abs_bound(a0->coeffs, 0, MLDSA_N, MLDSA_GAMMA2+1))
+)
+{
+  unsigned int i;
+  mld_assert_bound(a0->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(array_bound(a0->coeffs, i, MLDSA_N, 0, MLDSA_Q))
+    invariant(array_bound(a1->coeffs, 0, i, 0, (MLDSA_Q-1)/(2*MLDSA_GAMMA2)))
+    invariant(array_abs_bound(a0->coeffs, 0, i, MLDSA_GAMMA2+1))
+    decreases(MLDSA_N - i)
+  )
+  {
+    mld_decompose(&a0->coeffs[i], &a1->coeffs[i], a0->coeffs[i]);
+  }
+
+  mld_assert_abs_bound(a0->coeffs, MLDSA_N, MLDSA_GAMMA2 + 1);
+  mld_assert_bound(a1->coeffs, MLDSA_N, 0, (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2));
+}
+
+MLD_INTERNAL_API
+void mld_poly_decompose(mld_poly *a1, mld_poly *a0)
+{
+#if defined(MLD_USE_NATIVE_POLY_DECOMPOSE_88) && MLD_CONFIG_PARAMETER_SET == 44
+  int ret;
+  mld_assert_bound(a0->coeffs, MLDSA_N, 0, MLDSA_Q);
+  ret = mld_poly_decompose_88_native(a1->coeffs, a0->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_abs_bound(a0->coeffs, MLDSA_N, MLDSA_GAMMA2 + 1);
+    mld_assert_bound(a1->coeffs, MLDSA_N, 0,
+                     (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2));
+    return;
+  }
+#elif defined(MLD_USE_NATIVE_POLY_DECOMPOSE_32) && \
+    (MLD_CONFIG_PARAMETER_SET == 65 || MLD_CONFIG_PARAMETER_SET == 87)
+  int ret;
+  mld_assert_bound(a0->coeffs, MLDSA_N, 0, MLDSA_Q);
+  ret = mld_poly_decompose_32_native(a1->coeffs, a0->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_abs_bound(a0->coeffs, MLDSA_N, MLDSA_GAMMA2 + 1);
+    mld_assert_bound(a1->coeffs, MLDSA_N, 0,
+                     (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2));
+    return;
+  }
+#endif /* !(MLD_USE_NATIVE_POLY_DECOMPOSE_88 && MLD_CONFIG_PARAMETER_SET ==    \
+          44) && MLD_USE_NATIVE_POLY_DECOMPOSE_32 && (MLD_CONFIG_PARAMETER_SET \
+          == 65 || MLD_CONFIG_PARAMETER_SET == 87) */
+  mld_poly_decompose_c(a1, a0);
+}
+
+#endif /* !MLD_CONFIG_NO_SIGN_API */
+
+#if !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_STATIC_TESTABLE void mld_poly_use_hint_c(mld_poly *a, const mld_poly *h)
+__contract__(
+  requires(memory_no_alias(a, sizeof(mld_poly)))
+  requires(memory_no_alias(h, sizeof(mld_poly)))
+  requires(array_bound(a->coeffs, 0, MLDSA_N, 0, MLDSA_Q))
+  requires(array_bound(h->coeffs, 0, MLDSA_N, 0, 2))
+  assigns(memory_slice(a, sizeof(mld_poly)))
+  ensures(array_bound(a->coeffs, 0, MLDSA_N, 0, (MLDSA_Q-1)/(2*MLDSA_GAMMA2)))
+)
+{
+  unsigned int i;
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+  mld_assert_bound(h->coeffs, MLDSA_N, 0, 2);
+
+  for (i = 0; i < MLDSA_N; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N)
+    invariant(array_bound(a->coeffs, 0, i, 0, (MLDSA_Q-1)/(2*MLDSA_GAMMA2)))
+    invariant(array_bound(a->coeffs, i, MLDSA_N, 0, MLDSA_Q))
+    decreases(MLDSA_N - i)
+  )
+  {
+    a->coeffs[i] = mld_use_hint(a->coeffs[i], h->coeffs[i]);
+  }
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2));
+}
+
+MLD_INTERNAL_API
+void mld_poly_use_hint(mld_poly *a, const mld_poly *h)
+{
+#if defined(MLD_USE_NATIVE_POLY_USE_HINT_88) && MLD_CONFIG_PARAMETER_SET == 44
+  int ret;
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+  mld_assert_bound(h->coeffs, MLDSA_N, 0, 2);
+  ret = mld_poly_use_hint_88_native(a->coeffs, h->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_bound(a->coeffs, MLDSA_N, 0, (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2));
+    return;
+  }
+#elif defined(MLD_USE_NATIVE_POLY_USE_HINT_32) && \
+    (MLD_CONFIG_PARAMETER_SET == 65 || MLD_CONFIG_PARAMETER_SET == 87)
+  int ret;
+  mld_assert_bound(a->coeffs, MLDSA_N, 0, MLDSA_Q);
+  mld_assert_bound(h->coeffs, MLDSA_N, 0, 2);
+  ret = mld_poly_use_hint_32_native(a->coeffs, h->coeffs);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_bound(a->coeffs, MLDSA_N, 0, (MLDSA_Q - 1) / (2 * MLDSA_GAMMA2));
+    return;
+  }
+#endif /* !(MLD_USE_NATIVE_POLY_USE_HINT_88 && MLD_CONFIG_PARAMETER_SET == 44) \
+          && MLD_USE_NATIVE_POLY_USE_HINT_32 && (MLD_CONFIG_PARAMETER_SET ==   \
+          65 || MLD_CONFIG_PARAMETER_SET == 87) */
+  mld_poly_use_hint_c(a, h);
+}
+#endif /* !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API)
+/**
+ * Sample uniformly random coefficients in [-MLDSA_ETA, MLDSA_ETA] 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.
+ *
+ * @return Number of sampled coefficients. Can be smaller than target if not
+ *         enough random bytes were given.
+ */
+
+/* Reference: `mld_rej_eta()` 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. */
+#if MLDSA_ETA == 2
+/*
+ * Sampling 256 coefficients mod 15 using rejection sampling from 4 bits.
+ * Expected number of required bytes: (256 * (16/15))/2 = 136.5 bytes.
+ * We sample 1 block (=136 bytes) of SHAKE256_RATE output initially.
+ * Sampling 2 blocks initially results in slightly worse performance.
+ */
+#define MLD_POLY_UNIFORM_ETA_NBLOCKS 1
+#elif MLDSA_ETA == 4
+/*
+ * Sampling 256 coefficients mod 9 using rejection sampling from 4 bits.
+ * Expected number of required bytes: (256 * (16/9))/2 = 227.5 bytes.
+ * We sample 2 blocks (=272 bytes) of SHAKE256_RATE output initially.
+ */
+#define MLD_POLY_UNIFORM_ETA_NBLOCKS 2
+#else /* MLDSA_ETA == 4 */
+#error "Invalid value of MLDSA_ETA"
+#endif /* MLDSA_ETA != 2 && MLDSA_ETA != 4 */
+
+MLD_STATIC_TESTABLE unsigned int mld_rej_eta_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_ETA_NBLOCKS * MLD_STREAM256_BLOCKBYTES))
+  requires(memory_no_alias(a, sizeof(int32_t) * target))
+  requires(memory_no_alias(buf, buflen))
+  requires(array_abs_bound(a, 0, offset, MLDSA_ETA + 1))
+  assigns(memory_slice(a, sizeof(int32_t) * target))
+  ensures(offset <= return_value && return_value <= target)
+  ensures(array_abs_bound(a, 0, return_value, MLDSA_ETA + 1))
+)
+{
+  unsigned int ctr, pos;
+  int t_valid;
+  uint32_t t0, t1;
+  mld_assert_abs_bound(a, offset, MLDSA_ETA + 1);
+  ctr = offset;
+  pos = 0;
+  while (ctr < target && pos < buflen)
+  __loop__(
+    invariant(offset <= ctr && ctr <= target && pos <= buflen)
+    invariant(array_abs_bound(a, 0, ctr, MLDSA_ETA + 1))
+    decreases(buflen - pos)
+  )
+  {
+    t0 = buf[pos] & 0x0F;
+    t1 = buf[pos++] >> 4;
+
+    /* Constant time: The inputs and outputs to the rejection sampling are
+     * secret. However, it is fine to leak which coefficients have been
+     * rejected. For constant-time testing, we declassify the result of
+     * the comparison.
+     */
+#if MLDSA_ETA == 2
+    t_valid = t0 < 15;
+    MLD_CT_TESTING_DECLASSIFY(&t_valid, sizeof(int));
+    if (t_valid) /* t0 < 15 */
+    {
+      t0 = t0 - (205 * t0 >> 10) * 5;
+      a[ctr++] = 2 - (int32_t)t0;
+    }
+    t_valid = t1 < 15;
+    MLD_CT_TESTING_DECLASSIFY(&t_valid, sizeof(int));
+    if (t_valid && ctr < target) /* t1 < 15 */
+    {
+      t1 = t1 - (205 * t1 >> 10) * 5;
+      a[ctr++] = 2 - (int32_t)t1;
+    }
+#elif MLDSA_ETA == 4
+    t_valid = t0 < 9;
+    MLD_CT_TESTING_DECLASSIFY(&t_valid, sizeof(int));
+    if (t_valid) /* t0 < 9 */
+    {
+      a[ctr++] = 4 - (int32_t)t0;
+    }
+    t_valid = t1 < 9; /* t1 < 9 */
+    MLD_CT_TESTING_DECLASSIFY(&t_valid, sizeof(int));
+    if (t_valid && ctr < target)
+    {
+      a[ctr++] = 4 - (int32_t)t1;
+    }
+#else /* MLDSA_ETA == 4 */
+#error "Invalid value of MLDSA_ETA"
+#endif /* MLDSA_ETA != 2 && MLDSA_ETA != 4 */
+  }
+
+  mld_assert_abs_bound(a, ctr, MLDSA_ETA + 1);
+
+  return ctr;
+}
+
+static unsigned int mld_rej_eta(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_ETA_NBLOCKS * MLD_STREAM256_BLOCKBYTES))
+  requires(memory_no_alias(a, sizeof(int32_t) * target))
+  requires(memory_no_alias(buf, buflen))
+  requires(array_abs_bound(a, 0, offset, MLDSA_ETA + 1))
+  assigns(memory_slice(a, sizeof(int32_t) * target))
+  ensures(offset <= return_value && return_value <= target)
+  ensures(array_abs_bound(a, 0, return_value, MLDSA_ETA + 1))
+)
+{
+#if MLDSA_ETA == 2 && defined(MLD_USE_NATIVE_REJ_UNIFORM_ETA2)
+  int ret;
+  mld_assert_abs_bound(a, offset, MLDSA_ETA + 1);
+  if (offset == 0)
+  {
+    ret = mld_rej_uniform_eta2_native(a, target, buf, buflen);
+    if (ret != MLD_NATIVE_FUNC_FALLBACK)
+    {
+      unsigned res = (unsigned)ret;
+      mld_assert_abs_bound(a, res, MLDSA_ETA + 1);
+      return res;
+    }
+  }
+#elif MLDSA_ETA == 4 && defined(MLD_USE_NATIVE_REJ_UNIFORM_ETA4)
+  int ret;
+  mld_assert_abs_bound(a, offset, MLDSA_ETA + 1);
+  if (offset == 0)
+  {
+    ret = mld_rej_uniform_eta4_native(a, target, buf, buflen);
+    if (ret != MLD_NATIVE_FUNC_FALLBACK)
+    {
+      unsigned res = (unsigned)ret;
+      mld_assert_abs_bound(a, res, MLDSA_ETA + 1);
+      return res;
+    }
+  }
+#endif /* !(MLDSA_ETA == 2 && MLD_USE_NATIVE_REJ_UNIFORM_ETA2) && MLDSA_ETA == \
+          4 && MLD_USE_NATIVE_REJ_UNIFORM_ETA4 */
+
+  return mld_rej_eta_c(a, target, offset, buf, buflen);
+}
+
+#if !defined(MLD_CONFIG_SERIAL_FIPS202_ONLY)
+MLD_INTERNAL_API
+void mld_poly_uniform_eta_4x(mld_poly *r0, mld_poly *r1, mld_poly *r2,
+                             mld_poly *r3, const uint8_t seed[MLDSA_CRHBYTES],
+                             uint8_t nonce0, uint8_t nonce1, uint8_t nonce2,
+                             uint8_t nonce3)
+{
+  /* Temporary buffers for XOF output before rejection sampling */
+  MLD_ALIGN uint8_t buf[4][MLD_ALIGN_UP(MLD_POLY_UNIFORM_ETA_NBLOCKS *
+                                        MLD_STREAM256_BLOCKBYTES)];
+
+  MLD_ALIGN uint8_t extseed[4][MLD_ALIGN_UP(MLDSA_CRHBYTES + 2)];
+
+  /* Tracks the number of coefficients we have already sampled */
+  unsigned ctr[4];
+  mld_xof256_x4_ctx state;
+  unsigned buflen;
+
+  mld_memcpy(extseed[0], seed, MLDSA_CRHBYTES);
+  mld_memcpy(extseed[1], seed, MLDSA_CRHBYTES);
+  mld_memcpy(extseed[2], seed, MLDSA_CRHBYTES);
+  mld_memcpy(extseed[3], seed, MLDSA_CRHBYTES);
+  extseed[0][MLDSA_CRHBYTES] = nonce0;
+  extseed[1][MLDSA_CRHBYTES] = nonce1;
+  extseed[2][MLDSA_CRHBYTES] = nonce2;
+  extseed[3][MLDSA_CRHBYTES] = nonce3;
+  extseed[0][MLDSA_CRHBYTES + 1] = 0;
+  extseed[1][MLDSA_CRHBYTES + 1] = 0;
+  extseed[2][MLDSA_CRHBYTES + 1] = 0;
+  extseed[3][MLDSA_CRHBYTES + 1] = 0;
+
+  mld_xof256_x4_init(&state);
+  mld_xof256_x4_absorb(&state, extseed, MLDSA_CRHBYTES + 2);
+
+  /*
+   * Initially, squeeze heuristic number of MLD_POLY_UNIFORM_ETA_NBLOCKS.
+   * This should generate the coefficients with high probability.
+   */
+  mld_xof256_x4_squeezeblocks(buf, MLD_POLY_UNIFORM_ETA_NBLOCKS, &state);
+  buflen = MLD_POLY_UNIFORM_ETA_NBLOCKS * MLD_STREAM256_BLOCKBYTES;
+
+  ctr[0] = mld_rej_eta(r0->coeffs, MLDSA_N, 0, buf[0], buflen);
+  ctr[1] = mld_rej_eta(r1->coeffs, MLDSA_N, 0, buf[1], buflen);
+  ctr[2] = mld_rej_eta(r2->coeffs, MLDSA_N, 0, buf[2], buflen);
+  ctr[3] = mld_rej_eta(r3->coeffs, MLDSA_N, 0, buf[3], buflen);
+
+  /*
+   * So long as not all entries have been generated, squeeze
+   * one more block at a time until we're done.
+   */
+  buflen = MLD_STREAM256_BLOCKBYTES;
+  while (ctr[0] < MLDSA_N || ctr[1] < MLDSA_N || ctr[2] < MLDSA_N ||
+         ctr[3] < MLDSA_N)
+  __loop__(
+    assigns(ctr, state, memory_slice(r0, sizeof(mld_poly)),
+            memory_slice(r1, sizeof(mld_poly)), memory_slice(r2, sizeof(mld_poly)),
+            memory_slice(r3, sizeof(mld_poly)), object_whole(buf[0]),
+            object_whole(buf[1]), object_whole(buf[2]),
+            object_whole(buf[3]))
+    invariant(ctr[0] <= MLDSA_N && ctr[1] <= MLDSA_N)
+    invariant(ctr[2] <= MLDSA_N && ctr[3] <= MLDSA_N)
+    invariant(array_abs_bound(r0->coeffs, 0, ctr[0], MLDSA_ETA + 1))
+    invariant(array_abs_bound(r1->coeffs, 0, ctr[1], MLDSA_ETA + 1))
+    invariant(array_abs_bound(r2->coeffs, 0, ctr[2], MLDSA_ETA + 1))
+    invariant(array_abs_bound(r3->coeffs, 0, ctr[3], MLDSA_ETA + 1)))
+  {
+    mld_xof256_x4_squeezeblocks(buf, 1, &state);
+    ctr[0] = mld_rej_eta(r0->coeffs, MLDSA_N, ctr[0], buf[0], buflen);
+    ctr[1] = mld_rej_eta(r1->coeffs, MLDSA_N, ctr[1], buf[1], buflen);
+    ctr[2] = mld_rej_eta(r2->coeffs, MLDSA_N, ctr[2], buf[2], buflen);
+    ctr[3] = mld_rej_eta(r3->coeffs, MLDSA_N, ctr[3], buf[3], buflen);
+  }
+
+  mld_xof256_x4_release(&state);
+
+  mld_assert_abs_bound(r0->coeffs, MLDSA_N, MLDSA_ETA + 1);
+  mld_assert_abs_bound(r1->coeffs, MLDSA_N, MLDSA_ETA + 1);
+  mld_assert_abs_bound(r2->coeffs, MLDSA_N, MLDSA_ETA + 1);
+  mld_assert_abs_bound(r3->coeffs, MLDSA_N, MLDSA_ETA + 1);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(buf, sizeof(buf));
+  mld_zeroize(extseed, sizeof(extseed));
+}
+#else  /* !MLD_CONFIG_SERIAL_FIPS202_ONLY */
+
+MLD_INTERNAL_API
+void mld_poly_uniform_eta(mld_poly *r, const uint8_t seed[MLDSA_CRHBYTES],
+                          uint8_t nonce)
+{
+  /* Temporary buffer for XOF output before rejection sampling */
+  MLD_ALIGN uint8_t
+      buf[MLD_POLY_UNIFORM_ETA_NBLOCKS * MLD_STREAM256_BLOCKBYTES];
+  MLD_ALIGN uint8_t extseed[MLDSA_CRHBYTES + 2];
+
+  /* Tracks the number of coefficients we have already sampled */
+  unsigned ctr;
+  mld_xof256_ctx state;
+  unsigned buflen;
+
+  mld_memcpy(extseed, seed, MLDSA_CRHBYTES);
+  extseed[MLDSA_CRHBYTES] = nonce;
+  extseed[MLDSA_CRHBYTES + 1] = 0;
+
+  mld_xof256_init(&state);
+  mld_xof256_absorb_once(&state, extseed, MLDSA_CRHBYTES + 2);
+
+  /*
+   * Initially, squeeze heuristic number of MLD_POLY_UNIFORM_ETA_NBLOCKS.
+   * This should generate the coefficients with high probability.
+   */
+  mld_xof256_squeezeblocks(buf, MLD_POLY_UNIFORM_ETA_NBLOCKS, &state);
+  buflen = MLD_POLY_UNIFORM_ETA_NBLOCKS * MLD_STREAM256_BLOCKBYTES;
+
+  ctr = mld_rej_eta(r->coeffs, MLDSA_N, 0, buf, buflen);
+
+  /*
+   * So long as not all entries have been generated, squeeze
+   * one more block at a time until we're done.
+   */
+  buflen = MLD_STREAM256_BLOCKBYTES;
+  while (ctr < MLDSA_N)
+  __loop__(
+    assigns(ctr, object_whole(&state),
+      object_whole(buf), memory_slice(r, sizeof(mld_poly)))
+    invariant(ctr <= MLDSA_N)
+    invariant(state.pos <= SHAKE256_RATE)
+    invariant(array_abs_bound(r->coeffs, 0, ctr, MLDSA_ETA + 1)))
+  {
+    mld_xof256_squeezeblocks(buf, 1, &state);
+    ctr = mld_rej_eta(r->coeffs, MLDSA_N, ctr, buf, buflen);
+  }
+
+  mld_xof256_release(&state);
+
+  mld_assert_abs_bound(r->coeffs, MLDSA_N, MLDSA_ETA + 1);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(buf, sizeof(buf));
+  mld_zeroize(extseed, sizeof(extseed));
+}
+#endif /* MLD_CONFIG_SERIAL_FIPS202_ONLY */
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API)
+#define MLD_POLY_UNIFORM_GAMMA1_NBLOCKS                       \
+  ((MLDSA_POLYZ_PACKEDBYTES + MLD_STREAM256_BLOCKBYTES - 1) / \
+   MLD_STREAM256_BLOCKBYTES)
+
+#if MLD_CONFIG_PARAMETER_SET == 65 ||              \
+    defined(MLD_CONFIG_SERIAL_FIPS202_ONLY) ||     \
+    defined(MLD_CONFIG_REDUCE_RAM) || defined(MLD_UNIT_TEST)
+MLD_INTERNAL_API
+void mld_poly_uniform_gamma1(mld_poly *a, const uint8_t seed[MLDSA_CRHBYTES],
+                             uint16_t nonce)
+{
+  MLD_ALIGN uint8_t
+      buf[MLD_POLY_UNIFORM_GAMMA1_NBLOCKS * MLD_STREAM256_BLOCKBYTES];
+  MLD_ALIGN uint8_t extseed[MLDSA_CRHBYTES + 2];
+  mld_xof256_ctx state;
+
+  mld_memcpy(extseed, seed, MLDSA_CRHBYTES);
+  extseed[MLDSA_CRHBYTES] = (uint8_t)(nonce & 0xFF);
+  extseed[MLDSA_CRHBYTES + 1] = (uint8_t)(nonce >> 8);
+
+  mld_xof256_init(&state);
+  mld_xof256_absorb_once(&state, extseed, MLDSA_CRHBYTES + 2);
+
+  mld_xof256_squeezeblocks(buf, MLD_POLY_UNIFORM_GAMMA1_NBLOCKS, &state);
+  mld_polyz_unpack(a, buf);
+
+  mld_xof256_release(&state);
+
+  mld_assert_bound(a->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(buf, sizeof(buf));
+  mld_zeroize(extseed, sizeof(extseed));
+}
+#endif /* MLD_CONFIG_PARAMETER_SET == 65 || MLD_CONFIG_SERIAL_FIPS202_ONLY || \
+          MLD_CONFIG_REDUCE_RAM || MLD_UNIT_TEST */
+
+
+#if !defined(MLD_CONFIG_SERIAL_FIPS202_ONLY) && \
+    (!defined(MLD_CONFIG_REDUCE_RAM) || defined(MLD_UNIT_TEST))
+MLD_INTERNAL_API
+void mld_poly_uniform_gamma1_4x(mld_poly *r0, mld_poly *r1, mld_poly *r2,
+                                mld_poly *r3,
+                                const uint8_t seed[MLDSA_CRHBYTES],
+                                uint16_t nonce0, uint16_t nonce1,
+                                uint16_t nonce2, uint16_t nonce3)
+{
+  /* Temporary buffers for XOF output before rejection sampling */
+  MLD_ALIGN uint8_t buf[4][MLD_ALIGN_UP(MLD_POLY_UNIFORM_GAMMA1_NBLOCKS *
+                                        MLD_STREAM256_BLOCKBYTES)];
+
+  MLD_ALIGN uint8_t extseed[4][MLD_ALIGN_UP(MLDSA_CRHBYTES + 2)];
+
+  /* Tracks the number of coefficients we have already sampled */
+  mld_xof256_x4_ctx state;
+
+  mld_memcpy(extseed[0], seed, MLDSA_CRHBYTES);
+  mld_memcpy(extseed[1], seed, MLDSA_CRHBYTES);
+  mld_memcpy(extseed[2], seed, MLDSA_CRHBYTES);
+  mld_memcpy(extseed[3], seed, MLDSA_CRHBYTES);
+  extseed[0][MLDSA_CRHBYTES] = (uint8_t)(nonce0 & 0xFF);
+  extseed[1][MLDSA_CRHBYTES] = (uint8_t)(nonce1 & 0xFF);
+  extseed[2][MLDSA_CRHBYTES] = (uint8_t)(nonce2 & 0xFF);
+  extseed[3][MLDSA_CRHBYTES] = (uint8_t)(nonce3 & 0xFF);
+  extseed[0][MLDSA_CRHBYTES + 1] = (uint8_t)(nonce0 >> 8);
+  extseed[1][MLDSA_CRHBYTES + 1] = (uint8_t)(nonce1 >> 8);
+  extseed[2][MLDSA_CRHBYTES + 1] = (uint8_t)(nonce2 >> 8);
+  extseed[3][MLDSA_CRHBYTES + 1] = (uint8_t)(nonce3 >> 8);
+
+  mld_xof256_x4_init(&state);
+  mld_xof256_x4_absorb(&state, extseed, MLDSA_CRHBYTES + 2);
+  mld_xof256_x4_squeezeblocks(buf, MLD_POLY_UNIFORM_GAMMA1_NBLOCKS, &state);
+
+  mld_polyz_unpack(r0, buf[0]);
+  mld_polyz_unpack(r1, buf[1]);
+  mld_polyz_unpack(r2, buf[2]);
+  mld_polyz_unpack(r3, buf[3]);
+  mld_xof256_x4_release(&state);
+
+  mld_assert_bound(r0->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+  mld_assert_bound(r1->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+  mld_assert_bound(r2->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+  mld_assert_bound(r3->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(buf, sizeof(buf));
+  mld_zeroize(extseed, sizeof(extseed));
+}
+#endif /* !MLD_CONFIG_SERIAL_FIPS202_ONLY && (!MLD_CONFIG_REDUCE_RAM || \
+          MLD_UNIT_TEST) */
+#endif /* !MLD_CONFIG_NO_SIGN_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_INTERNAL_API
+void mld_poly_challenge(mld_poly *c, const uint8_t seed[MLDSA_CTILDEBYTES])
+{
+  unsigned int i, j, pos;
+  uint64_t signs;
+  uint64_t offset;
+  MLD_ALIGN uint8_t buf[SHAKE256_RATE];
+  mld_shake256ctx state;
+
+  mld_shake256_init(&state);
+  mld_shake256_absorb(&state, seed, MLDSA_CTILDEBYTES);
+  mld_shake256_finalize(&state);
+  mld_shake256_squeeze(buf, SHAKE256_RATE, &state);
+
+  /* Convert the first 8 bytes of buf[] into an unsigned 64-bit value.   */
+  /* Each bit of that dictates the sign of the resulting challenge value */
+  signs = 0;
+  for (i = 0; i < 8; ++i)
+  __loop__(
+    assigns(i, signs)
+    invariant(i <= 8)
+    decreases(8 - i)
+  )
+  {
+    signs |= (uint64_t)buf[i] << 8 * i;
+  }
+  pos = 8;
+
+  mld_memset(c, 0, sizeof(mld_poly));
+
+  for (i = MLDSA_N - MLDSA_TAU; i < MLDSA_N; ++i)
+  __loop__(
+    assigns(i, j, object_whole(buf), state, pos, memory_slice(c, sizeof(mld_poly)), signs)
+    invariant(i >= MLDSA_N - MLDSA_TAU)
+    invariant(i <= MLDSA_N)
+    invariant(pos >= 1)
+    invariant(pos <= SHAKE256_RATE)
+    invariant(array_bound(c->coeffs, 0, MLDSA_N, -1, 2))
+    invariant(state.pos <= SHAKE256_RATE)
+    decreases(MLDSA_N - i)
+  )
+  {
+    /* This loop terminates only probabilistically, hence no decreases
+     * clause. */
+    do
+    __loop__(
+      assigns(j, object_whole(buf), state, pos)
+      invariant(state.pos <= SHAKE256_RATE)
+    )
+    {
+      if (pos >= SHAKE256_RATE)
+      {
+        mld_shake256_squeeze(buf, SHAKE256_RATE, &state);
+        pos = 0;
+      }
+      j = buf[pos++];
+    } while (j > i);
+
+    c->coeffs[i] = c->coeffs[j];
+
+    /* Reference: Compute coefficient value here in two steps to */
+    /* avoid mixing unsigned and signed arithmetic with implicit */
+    /* conversions, and so that CBMC can keep track of ranges    */
+    /* to complete type-safety proof here.                       */
+
+    /* The least-significant bit of signs tells us if we want -1 or +1 */
+    offset = 2 * (signs & 1);
+
+    /* offset has value 0 or 2 here, so (1 - (int32_t) offset) has
+     * value -1 or +1 */
+    c->coeffs[j] = 1 - (int32_t)offset;
+
+    /* Move to the next bit of signs for next time */
+    signs >>= 1;
+  }
+
+  mld_assert_bound(c->coeffs, MLDSA_N, -1, 2);
+  mld_shake256_release(&state);
+
+  /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
+  mld_zeroize(buf, sizeof(buf));
+  mld_zeroize(&signs, sizeof(signs));
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API || !MLD_CONFIG_NO_VERIFY_API */
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API)
+MLD_INTERNAL_API
+void mld_polyeta_pack(uint8_t r[MLDSA_POLYETA_PACKEDBYTES], const mld_poly *a)
+{
+  unsigned int i;
+  uint8_t t[8];
+
+  mld_assert_abs_bound(a->coeffs, MLDSA_N, MLDSA_ETA + 1);
+
+#if MLDSA_ETA == 2
+  for (i = 0; i < MLDSA_N / 8; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/8)
+    decreases(MLDSA_N / 8 - i))
+  {
+    /* The casts are safe since we assume that the coefficients
+     * of a are <= MLDSA_ETA in absolute value. */
+    t[0] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 0]);
+    t[1] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 1]);
+    t[2] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 2]);
+    t[3] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 3]);
+    t[4] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 4]);
+    t[5] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 5]);
+    t[6] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 6]);
+    t[7] = (uint8_t)(MLDSA_ETA - a->coeffs[8 * i + 7]);
+
+    r[3 * i + 0] = (uint8_t)(((t[0] >> 0) | (t[1] << 3) | (t[2] << 6)) & 0xFF);
+    r[3 * i + 1] =
+        (uint8_t)(((t[2] >> 2) | (t[3] << 1) | (t[4] << 4) | (t[5] << 7)) &
+                  0xFF);
+    r[3 * i + 2] = (uint8_t)(((t[5] >> 1) | (t[6] << 2) | (t[7] << 5)) & 0xFF);
+  }
+#elif MLDSA_ETA == 4
+  for (i = 0; i < MLDSA_N / 2; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/2)
+    decreases(MLDSA_N / 2 - i))
+  {
+    /* The casts are safe since we assume that the coefficients
+     * of a are <= MLDSA_ETA in absolute value. */
+    t[0] = (uint8_t)(MLDSA_ETA - a->coeffs[2 * i + 0]);
+    t[1] = (uint8_t)(MLDSA_ETA - a->coeffs[2 * i + 1]);
+    r[i] = (uint8_t)(t[0] | (t[1] << 4));
+  }
+#else /* MLDSA_ETA == 4 */
+#error "Invalid value of MLDSA_ETA"
+#endif /* MLDSA_ETA != 2 && MLDSA_ETA != 4 */
+}
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API */
+
+#if !defined(MLD_CONFIG_NO_KEYPAIR_API) || !defined(MLD_CONFIG_NO_SIGN_API)
+MLD_INTERNAL_API
+void mld_polyeta_unpack(mld_poly *r, const uint8_t a[MLDSA_POLYETA_PACKEDBYTES])
+{
+  unsigned int i;
+
+#if MLDSA_ETA == 2
+  for (i = 0; i < MLDSA_N / 8; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/8)
+    invariant(array_bound(r->coeffs, 0, i*8, -5, MLDSA_ETA + 1))
+    decreases(MLDSA_N / 8 - i))
+  {
+    r->coeffs[8 * i + 0] = (a[3 * i + 0] >> 0) & 7;
+    r->coeffs[8 * i + 1] = (a[3 * i + 0] >> 3) & 7;
+    r->coeffs[8 * i + 2] = ((a[3 * i + 0] >> 6) | (a[3 * i + 1] << 2)) & 7;
+    r->coeffs[8 * i + 3] = (a[3 * i + 1] >> 1) & 7;
+    r->coeffs[8 * i + 4] = (a[3 * i + 1] >> 4) & 7;
+    r->coeffs[8 * i + 5] = ((a[3 * i + 1] >> 7) | (a[3 * i + 2] << 1)) & 7;
+    r->coeffs[8 * i + 6] = (a[3 * i + 2] >> 2) & 7;
+    r->coeffs[8 * i + 7] = (a[3 * i + 2] >> 5) & 7;
+
+    r->coeffs[8 * i + 0] = MLDSA_ETA - r->coeffs[8 * i + 0];
+    r->coeffs[8 * i + 1] = MLDSA_ETA - r->coeffs[8 * i + 1];
+    r->coeffs[8 * i + 2] = MLDSA_ETA - r->coeffs[8 * i + 2];
+    r->coeffs[8 * i + 3] = MLDSA_ETA - r->coeffs[8 * i + 3];
+    r->coeffs[8 * i + 4] = MLDSA_ETA - r->coeffs[8 * i + 4];
+    r->coeffs[8 * i + 5] = MLDSA_ETA - r->coeffs[8 * i + 5];
+    r->coeffs[8 * i + 6] = MLDSA_ETA - r->coeffs[8 * i + 6];
+    r->coeffs[8 * i + 7] = MLDSA_ETA - r->coeffs[8 * i + 7];
+  }
+#elif MLDSA_ETA == 4
+  for (i = 0; i < MLDSA_N / 2; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/2)
+    invariant(array_bound(r->coeffs, 0, i*2, -11, MLDSA_ETA + 1))
+    decreases(MLDSA_N / 2 - i))
+  {
+    r->coeffs[2 * i + 0] = a[i] & 0x0F;
+    r->coeffs[2 * i + 1] = a[i] >> 4;
+    r->coeffs[2 * i + 0] = MLDSA_ETA - r->coeffs[2 * i + 0];
+    r->coeffs[2 * i + 1] = MLDSA_ETA - r->coeffs[2 * i + 1];
+  }
+#else /* MLDSA_ETA == 4 */
+#error "Invalid value of MLDSA_ETA"
+#endif /* MLDSA_ETA != 2 && MLDSA_ETA != 4 */
+
+  mld_assert_bound(r->coeffs, MLDSA_N, MLD_POLYETA_UNPACK_LOWER_BOUND,
+                   MLDSA_ETA + 1);
+}
+#endif /* !MLD_CONFIG_NO_KEYPAIR_API || !MLD_CONFIG_NO_SIGN_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API)
+MLD_INTERNAL_API
+void mld_polyz_pack(uint8_t r[MLDSA_POLYZ_PACKEDBYTES], const mld_poly *a)
+{
+  unsigned int i;
+  uint32_t t[4];
+
+  mld_assert_bound(a->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+
+#if MLD_CONFIG_PARAMETER_SET == 44
+  for (i = 0; i < MLDSA_N / 4; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/4)
+    decreases(MLDSA_N / 4 - i))
+  {
+    /* Safety: a->coeffs[i] <= MLDSA_GAMMA1, hence, these casts are safe. */
+    t[0] = (uint32_t)(MLDSA_GAMMA1 - a->coeffs[4 * i + 0]);
+    t[1] = (uint32_t)(MLDSA_GAMMA1 - a->coeffs[4 * i + 1]);
+    t[2] = (uint32_t)(MLDSA_GAMMA1 - a->coeffs[4 * i + 2]);
+    t[3] = (uint32_t)(MLDSA_GAMMA1 - a->coeffs[4 * i + 3]);
+
+    r[9 * i + 0] = (uint8_t)((t[0]) & 0xFF);
+    r[9 * i + 1] = (uint8_t)((t[0] >> 8) & 0xFF);
+    r[9 * i + 2] = (uint8_t)((t[0] >> 16) & 0xFF);
+    r[9 * i + 2] |= (uint8_t)((t[1] << 2) & 0xFF);
+    r[9 * i + 3] = (uint8_t)((t[1] >> 6) & 0xFF);
+    r[9 * i + 4] = (uint8_t)((t[1] >> 14) & 0xFF);
+    r[9 * i + 4] |= (uint8_t)((t[2] << 4) & 0xFF);
+    r[9 * i + 5] = (uint8_t)((t[2] >> 4) & 0xFF);
+    r[9 * i + 6] = (uint8_t)((t[2] >> 12) & 0xFF);
+    r[9 * i + 6] |= (uint8_t)((t[3] << 6) & 0xFF);
+    r[9 * i + 7] = (uint8_t)((t[3] >> 2) & 0xFF);
+    r[9 * i + 8] = (uint8_t)((t[3] >> 10) & 0xFF);
+  }
+#else  /* MLD_CONFIG_PARAMETER_SET == 44 */
+  for (i = 0; i < MLDSA_N / 2; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/2)
+    decreases(MLDSA_N / 2 - i))
+  {
+    /* Safety: a->coeffs[i] <= MLDSA_GAMMA1, hence, these casts are safe. */
+    t[0] = (uint32_t)(MLDSA_GAMMA1 - a->coeffs[2 * i + 0]);
+    t[1] = (uint32_t)(MLDSA_GAMMA1 - a->coeffs[2 * i + 1]);
+
+    r[5 * i + 0] = (uint8_t)((t[0]) & 0xFF);
+    r[5 * i + 1] = (uint8_t)((t[0] >> 8) & 0xFF);
+    r[5 * i + 2] = (uint8_t)((t[0] >> 16) & 0xFF);
+    r[5 * i + 2] |= (uint8_t)((t[1] << 4) & 0xFF);
+    r[5 * i + 3] = (uint8_t)((t[1] >> 4) & 0xFF);
+    r[5 * i + 4] = (uint8_t)((t[1] >> 12) & 0xFF);
+  }
+#endif /* MLD_CONFIG_PARAMETER_SET != 44 */
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API */
+
+#if !defined(MLD_CONFIG_NO_SIGN_API) || !defined(MLD_CONFIG_NO_VERIFY_API)
+MLD_STATIC_TESTABLE void mld_polyz_unpack_c(
+    mld_poly *r, const uint8_t a[MLDSA_POLYZ_PACKEDBYTES])
+__contract__(
+  requires(memory_no_alias(r, sizeof(mld_poly)))
+  requires(memory_no_alias(a, MLDSA_POLYZ_PACKEDBYTES))
+  assigns(memory_slice(r, sizeof(mld_poly)))
+  ensures(array_bound(r->coeffs, 0, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1))
+)
+{
+  unsigned int i;
+#if MLD_CONFIG_PARAMETER_SET == 44
+  for (i = 0; i < MLDSA_N / 4; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/4)
+    invariant(array_bound(r->coeffs, 0, i*4, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1))
+    decreases(MLDSA_N / 4 - i))
+  {
+    r->coeffs[4 * i + 0] = a[9 * i + 0];
+    r->coeffs[4 * i + 0] |= (int32_t)a[9 * i + 1] << 8;
+    r->coeffs[4 * i + 0] |= (int32_t)a[9 * i + 2] << 16;
+    r->coeffs[4 * i + 0] &= 0x3FFFF;
+
+    r->coeffs[4 * i + 1] = a[9 * i + 2] >> 2;
+    r->coeffs[4 * i + 1] |= (int32_t)a[9 * i + 3] << 6;
+    r->coeffs[4 * i + 1] |= (int32_t)a[9 * i + 4] << 14;
+    r->coeffs[4 * i + 1] &= 0x3FFFF;
+
+    r->coeffs[4 * i + 2] = a[9 * i + 4] >> 4;
+    r->coeffs[4 * i + 2] |= (int32_t)a[9 * i + 5] << 4;
+    r->coeffs[4 * i + 2] |= (int32_t)a[9 * i + 6] << 12;
+    r->coeffs[4 * i + 2] &= 0x3FFFF;
+
+    r->coeffs[4 * i + 3] = a[9 * i + 6] >> 6;
+    r->coeffs[4 * i + 3] |= (int32_t)a[9 * i + 7] << 2;
+    r->coeffs[4 * i + 3] |= (int32_t)a[9 * i + 8] << 10;
+    r->coeffs[4 * i + 3] &= 0x3FFFF;
+
+    r->coeffs[4 * i + 0] = MLDSA_GAMMA1 - r->coeffs[4 * i + 0];
+    r->coeffs[4 * i + 1] = MLDSA_GAMMA1 - r->coeffs[4 * i + 1];
+    r->coeffs[4 * i + 2] = MLDSA_GAMMA1 - r->coeffs[4 * i + 2];
+    r->coeffs[4 * i + 3] = MLDSA_GAMMA1 - r->coeffs[4 * i + 3];
+  }
+#else  /* MLD_CONFIG_PARAMETER_SET == 44 */
+  for (i = 0; i < MLDSA_N / 2; ++i)
+  __loop__(
+    invariant(i <= MLDSA_N/2)
+    invariant(array_bound(r->coeffs, 0, i*2, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1))
+    decreases(MLDSA_N / 2 - i))
+  {
+    r->coeffs[2 * i + 0] = a[5 * i + 0];
+    r->coeffs[2 * i + 0] |= (int32_t)a[5 * i + 1] << 8;
+    r->coeffs[2 * i + 0] |= (int32_t)a[5 * i + 2] << 16;
+    r->coeffs[2 * i + 0] &= 0xFFFFF;
+
+    r->coeffs[2 * i + 1] = a[5 * i + 2] >> 4;
+    r->coeffs[2 * i + 1] |= (int32_t)a[5 * i + 3] << 4;
+    r->coeffs[2 * i + 1] |= (int32_t)a[5 * i + 4] << 12;
+    /* r->coeffs[2*i+1] &= 0xFFFFF; */ /* No effect, since we're anyway at 20
+                                          bits */
+
+    r->coeffs[2 * i + 0] = MLDSA_GAMMA1 - r->coeffs[2 * i + 0];
+    r->coeffs[2 * i + 1] = MLDSA_GAMMA1 - r->coeffs[2 * i + 1];
+  }
+#endif /* MLD_CONFIG_PARAMETER_SET != 44 */
+  mld_assert_bound(r->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+}
+
+MLD_INTERNAL_API
+void mld_polyz_unpack(mld_poly *r, const uint8_t a[MLDSA_POLYZ_PACKEDBYTES])
+{
+#if defined(MLD_USE_NATIVE_POLYZ_UNPACK_17) && MLD_CONFIG_PARAMETER_SET == 44
+  int ret;
+  ret = mld_polyz_unpack_17_native(r->coeffs, a);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_bound(r->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+    return;
+  }
+#elif defined(MLD_USE_NATIVE_POLYZ_UNPACK_19) && \
+    (MLD_CONFIG_PARAMETER_SET == 65 || MLD_CONFIG_PARAMETER_SET == 87)
+  int ret;
+  ret = mld_polyz_unpack_19_native(r->coeffs, a);
+  if (ret == MLD_NATIVE_FUNC_SUCCESS)
+  {
+    mld_assert_bound(r->coeffs, MLDSA_N, -(MLDSA_GAMMA1 - 1), MLDSA_GAMMA1 + 1);
+    return;
+  }
+#endif /* !(MLD_USE_NATIVE_POLYZ_UNPACK_17 && MLD_CONFIG_PARAMETER_SET == 44)  \
+          && MLD_USE_NATIVE_POLYZ_UNPACK_19 && (MLD_CONFIG_PARAMETER_SET == 65 \
+          || MLD_CONFIG_PARAMETER_SET == 87) */
+
+  mld_polyz_unpack_c(r, a);
+}
+#endif /* !MLD_CONFIG_NO_SIGN_API || !MLD_CONFIG_NO_VERIFY_API */
+
+/* To facilitate single-compilation-unit (SCU) builds, undefine all macros. */
+
+/* To facilitate single-compilation-unit (SCU) builds, undefine all macros.
+ * Don't modify by hand -- this is auto-generated by scripts/autogen. */
+#undef mld_rej_eta
+#undef mld_rej_eta_c
+#undef mld_poly_decompose_c
+#undef mld_poly_use_hint_c
+#undef mld_polyz_unpack_c
+#undef MLD_POLY_UNIFORM_ETA_NBLOCKS
+#undef MLD_POLY_UNIFORM_GAMMA1_NBLOCKS