Simplification and refactoring to restore proof speed and stability.

rod-chapman · mkannwischer · commit 7819c17a1c42 · 2025-11-05T10:47:03.000+08:00
1. Weaken post-condition and loop invariant in polyvecl_add(). The stonger
   post-condition was unncessary.

2. Simplify polyvec_matrix_expand(). Small performance loss here since
   batched_seeds[] is (re-) initialized every time. This is bit slower
   but removes a loop statement entirely.

3. Refactor polyvec_pointwise_acc_montgomery() by splitting core
   "sum of products" calculation into a distinct local function
   mld_pointwise_sum_of_products(). Add proof of the latter.

Proof time for parameter set 87 now 4 minutes (real-time) and
40 minutes (user time) with 64 cores on an r7g instance.

Signed-off-by: Rod Chapman &lt;rodchap@amazon.com&gt;
diff --git a/mldsa/src/polyvec.c b/mldsa/src/polyvec.c
@@ -45,20 +45,12 @@ void mld_polyvec_matrix_expand(mld_polyvecl mat[MLDSA_K],
    * of the same parent object.
    */
 
-  MLD_ALIGN uint8_t seed_ext[4][MLD_ALIGN_UP(MLDSA_SEEDBYTES + 2)];
-
-  for (j = 0; j < 4; j++)
-  __loop__(
-    assigns(j, object_whole(seed_ext))
-    invariant(j <= 4)
-  )
-  {
-    mld_memcpy(seed_ext[j], rho, MLDSA_SEEDBYTES);
-  }
+  MLD_ALIGN uint8_t single_seed[MLD_ALIGN_UP(MLDSA_SEEDBYTES + 2)];
+  MLD_ALIGN uint8_t batched_seeds[4][MLD_ALIGN_UP(MLDSA_SEEDBYTES + 2)];
   /* Sample 4 matrix entries a time. */
   for (i = 0; i < (MLDSA_K * MLDSA_L / 4) * 4; i += 4)
   __loop__(
-    assigns(i, j, object_whole(seed_ext), memory_slice(mat, MLDSA_K * sizeof(mld_polyvecl)))
+    assigns(i, j, object_whole(batched_seeds), memory_slice(mat, MLDSA_K * sizeof(mld_polyvecl)))
     invariant(i <= (MLDSA_K * MLDSA_L / 4) * 4 && i % 4 == 0)
     /* vectors 0 .. i / MLDSA_L are completely sampled */
     invariant(forall(k1, 0, i / MLDSA_L, forall(l1, 0, MLDSA_L,
@@ -70,28 +62,31 @@ void mld_polyvec_matrix_expand(mld_polyvecl mat[MLDSA_K],
   {
     for (j = 0; j < 4; j++)
     __loop__(
-      assigns(j, object_whole(seed_ext))
+      assigns(j, object_whole(batched_seeds))
       invariant(j <= 4)
     )
     {
       uint8_t x = (uint8_t)((i + j) / MLDSA_L);
       uint8_t y = (uint8_t)((i + j) % MLDSA_L);
 
-      seed_ext[j][MLDSA_SEEDBYTES + 0] = y;
-      seed_ext[j][MLDSA_SEEDBYTES + 1] = x;
+      mld_memcpy(batched_seeds[j], rho, MLDSA_SEEDBYTES);
+      batched_seeds[j][MLDSA_SEEDBYTES + 0] = y;
+      batched_seeds[j][MLDSA_SEEDBYTES + 1] = x;
     }
 
     mld_poly_uniform_4x(&mat[i / MLDSA_L].vec[i % MLDSA_L],
                         &mat[(i + 1) / MLDSA_L].vec[(i + 1) % MLDSA_L],
                         &mat[(i + 2) / MLDSA_L].vec[(i + 2) % MLDSA_L],
                         &mat[(i + 3) / MLDSA_L].vec[(i + 3) % MLDSA_L],
-                        seed_ext);
+                        batched_seeds);
   }
 
+  mld_memcpy(single_seed, rho, MLDSA_SEEDBYTES);
+
   /* For MLDSA_K=6, MLDSA_L=5, process the last two entries individually */
   while (i < MLDSA_K * MLDSA_L)
   __loop__(
-    assigns(i, object_whole(seed_ext), memory_slice(mat, MLDSA_K * sizeof(mld_polyvecl)))
+    assigns(i, object_whole(single_seed), memory_slice(mat, MLDSA_K * sizeof(mld_polyvecl)))
     invariant(i <= MLDSA_K * MLDSA_L)
     /* vectors 0 .. i / MLDSA_L are completely sampled */
     invariant(forall(k1, 0, i / MLDSA_L, forall(l1, 0, MLDSA_L,
@@ -105,27 +100,31 @@ void mld_polyvec_matrix_expand(mld_polyvecl mat[MLDSA_K],
     uint8_t y = (uint8_t)(i % MLDSA_L);
     mld_poly *this_poly = &mat[i / MLDSA_L].vec[i % MLDSA_L];
 
-    seed_ext[0][MLDSA_SEEDBYTES + 0] = y;
-    seed_ext[0][MLDSA_SEEDBYTES + 1] = x;
+    single_seed[MLDSA_SEEDBYTES + 0] = y;
+    single_seed[MLDSA_SEEDBYTES + 1] = x;
 
-    mld_poly_uniform(this_poly, seed_ext[0]);
+    mld_poly_uniform(this_poly, single_seed);
     i++;
   }
 
   /*
    * The public matrix is generated in NTT domain. If the native backend
-   * uses a custom order in NTT domain, permute A accordingly.
+   * uses a custom order in NTT domain, permute A accordingly. This does
+   * not affect the bounds on the coefficients, so we ignore this for CBMC
+   * to simplify proof.
    */
+#ifndef CBMC
   for (i = 0; i < MLDSA_K; i++)
   {
     for (j = 0; j < MLDSA_L; j++)
     {
       mld_poly_permute_bitrev_to_custom(mat[i].vec[j].coeffs);
     }
   }
+#endif /* !CBMC */
 
   /* @[FIPS204, Section 3.6.3] Destruction of intermediate values. */
-  mld_zeroize(seed_ext, sizeof(seed_ext));
+  mld_zeroize(single_seed, sizeof(single_seed));
 }
 
 MLD_INTERNAL_API
@@ -219,7 +218,6 @@ void mld_polyvecl_add(mld_polyvecl *u, const mld_polyvecl *v)
     invariant(i <= MLDSA_L)
     invariant(forall(k0, i, MLDSA_L,
               forall(k1, 0, MLDSA_N, u->vec[k0].coeffs[k1] == loop_entry(*u).vec[k0].coeffs[k1])))
-    invariant(forall(k4, 0, i, forall(k5, 0, MLDSA_N, u->vec[k4].coeffs[k5] == loop_entry(*u).vec[k4].coeffs[k5] + v->vec[k4].coeffs[k5])))
     invariant(forall(k6, 0, i, array_bound(u->vec[k6].coeffs, 0, MLDSA_N, INT32_MIN, REDUCE32_DOMAIN_MAX)))
   )
   {
@@ -287,87 +285,131 @@ void mld_polyvecl_pointwise_poly_montgomery(mld_polyvecl *r, const mld_poly *a,
   mld_assert_abs_bound_2d(r->vec, MLDSA_L, MLDSA_N, MLDSA_Q);
 }
 
+#if defined(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L4) && \
+    MLD_CONFIG_PARAMETER_SET == 44
+
 MLD_INTERNAL_API
 void mld_polyvecl_pointwise_acc_montgomery(mld_poly *w, const mld_polyvecl *u,
                                            const mld_polyvecl *v)
 {
-#if defined(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L4) && \
-    MLD_CONFIG_PARAMETER_SET == 44
   /* TODO: proof */
   mld_assert_bound_2d(u->vec, MLDSA_L, MLDSA_N, 0, MLDSA_Q);
   mld_assert_abs_bound_2d(v->vec, MLDSA_L, MLDSA_N, MLD_NTT_BOUND);
   mld_polyvecl_pointwise_acc_montgomery_l4_native(
       w->coeffs, (const int32_t(*)[MLDSA_N])u->vec,
       (const int32_t(*)[MLDSA_N])v->vec);
   mld_assert_abs_bound(w->coeffs, MLDSA_N, MLDSA_Q);
+}
+
 #elif defined(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L5) && \
     MLD_CONFIG_PARAMETER_SET == 65
+
+void mld_polyvecl_pointwise_acc_montgomery(mld_poly *w, const mld_polyvecl *u,
+                                           const mld_polyvecl *v)
+{
   /* TODO: proof */
   mld_assert_bound_2d(u->vec, MLDSA_L, MLDSA_N, 0, MLDSA_Q);
   mld_assert_abs_bound_2d(v->vec, MLDSA_L, MLDSA_N, MLD_NTT_BOUND);
   mld_polyvecl_pointwise_acc_montgomery_l5_native(
       w->coeffs, (const int32_t(*)[MLDSA_N])u->vec,
       (const int32_t(*)[MLDSA_N])v->vec);
   mld_assert_abs_bound(w->coeffs, MLDSA_N, MLDSA_Q);
+}
+
 #elif defined(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L7) && \
     MLD_CONFIG_PARAMETER_SET == 87
+void mld_polyvecl_pointwise_acc_montgomery(mld_poly *w, const mld_polyvecl *u,
+                                           const mld_polyvecl *v)
+{
   /* TODO: proof */
   mld_assert_bound_2d(u->vec, MLDSA_L, MLDSA_N, 0, MLDSA_Q);
   mld_assert_abs_bound_2d(v->vec, MLDSA_L, MLDSA_N, MLD_NTT_BOUND);
   mld_polyvecl_pointwise_acc_montgomery_l7_native(
       w->coeffs, (const int32_t(*)[MLDSA_N])u->vec,
       (const int32_t(*)[MLDSA_N])v->vec);
   mld_assert_abs_bound(w->coeffs, MLDSA_N, MLDSA_Q);
-#else  /* !(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L4 && \
-          MLD_CONFIG_PARAMETER_SET == 44) &&                       \
-          !(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L5 && \
-          MLD_CONFIG_PARAMETER_SET == 65) &&                       \
-          MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L7 &&   \
-          MLD_CONFIG_PARAMETER_SET == 87 */
-  unsigned int i, j;
-  mld_assert_bound_2d(u->vec, MLDSA_L, MLDSA_N, 0, MLDSA_Q);
-  mld_assert_abs_bound_2d(v->vec, MLDSA_L, MLDSA_N, MLD_NTT_BOUND);
-  /* The first input is bounded by [0, Q-1] inclusive
-   * The second input is bounded by [-9Q+1, 9Q-1] inclusive . Hence, we can
+}
+
+#else /* !(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L4 && \
+         MLD_CONFIG_PARAMETER_SET == 44) &&                       \
+         !(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L5 && \
+         MLD_CONFIG_PARAMETER_SET == 65) &&                       \
+         MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L7 &&   \
+         MLD_CONFIG_PARAMETER_SET == 87 */
+
+#define mld_pointwise_sum_of_products \
+  MLD_NAMESPACE_KL(mld_pointwise_sum_of_products)
+static int64_t mld_pointwise_sum_of_products(const mld_polyvecl *u,
+                                             const mld_polyvecl *v,
+                                             unsigned int i)
+__contract__(
+  requires(memory_no_alias(u, sizeof(mld_polyvecl)))
+  requires(memory_no_alias(v, sizeof(mld_polyvecl)))
+  requires(i < MLDSA_N)
+  requires(forall(l0, 0, MLDSA_L,
+                  array_bound(u->vec[l0].coeffs, 0, MLDSA_N, 0, MLDSA_Q)))
+  requires(forall(l1, 0, MLDSA_L,
+    array_abs_bound(v->vec[l1].coeffs, 0, MLDSA_N, MLD_NTT_BOUND)))
+  ensures(return_value >= -(int64_t) MLDSA_L*(MLDSA_Q - 1)*(MLD_NTT_BOUND - 1))
+  ensures(return_value <=  (int64_t) MLDSA_L*(MLDSA_Q - 1)*(MLD_NTT_BOUND - 1))
+)
+{
+  /* Input vector u is bounded by [0, Q-1] inclusive
+   * Input vector v is bounded by [-9Q+1, 9Q-1] inclusive . Hence, we can
    * safely accumulate in 64-bits without intermediate reductions as
    * MLDSA_L * (MLD_NTT_BOUND-1) * (Q-1) < INT64_MAX
    *
    * The worst case is ML-DSA-87: 7 * (9Q-1) * (Q-1) < 2**52
    * (and likewise for negative values)
    */
 
+  int64_t t = 0;
+  unsigned int j;
+  for (j = 0; j < MLDSA_L; j++)
+  __loop__(
+    assigns(j, t)
+    invariant(j <= MLDSA_L)
+    invariant(t >= -(int64_t)j*(MLDSA_Q - 1)*(MLD_NTT_BOUND - 1))
+    invariant(t <= (int64_t)j*(MLDSA_Q - 1)*(MLD_NTT_BOUND - 1))
+  )
+  {
+    const int64_t u64 = (int64_t)u->vec[j].coeffs[i];
+    const int64_t v64 = (int64_t)v->vec[j].coeffs[i];
+    /* Helper assertions for proof efficiency. Do not remove */
+    mld_assert(u64 >= 0 && u64 < MLDSA_Q);
+    mld_assert(v64 > -MLD_NTT_BOUND && v64 < MLD_NTT_BOUND);
+    t += (u64 * v64);
+  }
+  return t;
+}
+
+void mld_polyvecl_pointwise_acc_montgomery(mld_poly *w, const mld_polyvecl *u,
+                                           const mld_polyvecl *v)
+{
+  unsigned int i;
+
+  mld_assert_bound_2d(u->vec, MLDSA_L, MLDSA_N, 0, MLDSA_Q);
+  mld_assert_abs_bound_2d(v->vec, MLDSA_L, MLDSA_N, MLD_NTT_BOUND);
   for (i = 0; i < MLDSA_N; i++)
   __loop__(
-    assigns(i, j, object_whole(w))
+    assigns(i, object_whole(w))
     invariant(i <= MLDSA_N)
     invariant(array_abs_bound(w->coeffs, 0, i, MLDSA_Q))
   )
   {
-    int64_t t = 0;
-    int32_t r;
-    for (j = 0; j < MLDSA_L; j++)
-    __loop__(
-      assigns(j, t)
-      invariant(j <= MLDSA_L)
-      invariant(t >= -(int64_t)j*(MLDSA_Q - 1)*(MLD_NTT_BOUND - 1))
-      invariant(t <= (int64_t)j*(MLDSA_Q - 1)*(MLD_NTT_BOUND - 1))
-    )
-    {
-      t += (int64_t)u->vec[j].coeffs[i] * v->vec[j].coeffs[i];
-    }
-
-    r = mld_montgomery_reduce(t);
-    w->coeffs[i] = r;
+    w->coeffs[i] =
+        mld_montgomery_reduce(mld_pointwise_sum_of_products(u, v, i));
   }
 
   mld_assert_abs_bound(w->coeffs, MLDSA_N, MLDSA_Q);
+}
+
 #endif /* !(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L4 && \
           MLD_CONFIG_PARAMETER_SET == 44) &&                       \
           !(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L5 && \
           MLD_CONFIG_PARAMETER_SET == 65) &&                       \
           !(MLD_USE_NATIVE_POLYVECL_POINTWISE_ACC_MONTGOMERY_L7 && \
           MLD_CONFIG_PARAMETER_SET == 87) */
-}
 
 MLD_INTERNAL_API
 uint32_t mld_polyvecl_chknorm(const mld_polyvecl *v, int32_t bound)
diff --git a/mldsa/src/polyvec.h b/mldsa/src/polyvec.h
@@ -93,7 +93,6 @@ __contract__(
   requires(forall(k0, 0, MLDSA_L, forall(k1, 0, MLDSA_N, (int64_t) u->vec[k0].coeffs[k1] + v->vec[k0].coeffs[k1] < REDUCE32_DOMAIN_MAX)))
   requires(forall(k2, 0, MLDSA_L, forall(k3, 0, MLDSA_N, (int64_t) u->vec[k2].coeffs[k3] + v->vec[k2].coeffs[k3] >= INT32_MIN)))
   assigns(object_whole(u))
-  ensures(forall(k4, 0, MLDSA_L, forall(k5, 0, MLDSA_N, u->vec[k4].coeffs[k5] == old(*u).vec[k4].coeffs[k5] + v->vec[k4].coeffs[k5])))
   ensures(forall(k6, 0, MLDSA_L,
                  array_bound(u->vec[k6].coeffs, 0, MLDSA_N, INT32_MIN, REDUCE32_DOMAIN_MAX)))
 );
diff --git a/proofs/cbmc/pointwise_sum_of_products/Makefile b/proofs/cbmc/pointwise_sum_of_products/Makefile
@@ -0,0 +1,55 @@
+# Copyright (c) The mldsa-native project authors
+# SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+
+include ../Makefile_params.common
+
+HARNESS_ENTRY = harness
+HARNESS_FILE = pointwise_sum_of_products_harness
+
+# This should be a unique identifier for this proof, and will appear on the
+# Litani dashboard. It can be human-readable and contain spaces if you wish.
+PROOF_UID = pointwise_sum_of_products
+
+DEFINES +=
+INCLUDES +=
+
+REMOVE_FUNCTION_BODY +=
+UNWINDSET +=
+
+PROOF_SOURCES += $(PROOFDIR)/$(HARNESS_FILE).c
+PROJECT_SOURCES += $(SRCDIR)/mldsa/src/polyvec.c
+
+CHECK_FUNCTION_CONTRACTS=$(MLD_NAMESPACE)mld_pointwise_sum_of_products
+USE_FUNCTION_CONTRACTS=
+APPLY_LOOP_CONTRACTS=on
+USE_DYNAMIC_FRAMES=1
+
+# Disable any setting of EXTERNAL_SAT_SOLVER, and choose SMT backend instead
+EXTERNAL_SAT_SOLVER=
+CBMCFLAGS=--smt2 --slice-formula
+
+FUNCTION_NAME = pointwise_sum_of_products
+
+# If this proof is found to consume huge amounts of RAM, you can set the
+# EXPENSIVE variable. With new enough versions of the proof tools, this will
+# restrict the number of EXPENSIVE CBMC jobs running at once. See the
+# documentation in Makefile.common under the "Job Pools" heading for details.
+# EXPENSIVE = true
+
+# This function is large enough to need...
+CBMC_OBJECT_BITS = 12
+
+# If you require access to a file-local ("static") function or object to conduct
+# your proof, set the following (and do not include the original source file
+# ("mldsa/poly.c") in PROJECT_SOURCES).
+# REWRITTEN_SOURCES = $(PROOFDIR)/<__SOURCE_FILE_BASENAME__>.i
+# include ../Makefile.common
+# $(PROOFDIR)/<__SOURCE_FILE_BASENAME__>.i_SOURCE = $(SRCDIR)/mldsa/poly.c
+# $(PROOFDIR)/<__SOURCE_FILE_BASENAME__>.i_FUNCTIONS = foo bar
+# $(PROOFDIR)/<__SOURCE_FILE_BASENAME__>.i_OBJECTS = baz
+# Care is required with variables on the left-hand side: REWRITTEN_SOURCES must
+# be set before including Makefile.common, but any use of variables on the
+# left-hand side requires those variables to be defined. Hence, _SOURCE,
+# _FUNCTIONS, _OBJECTS is set after including Makefile.common.
+
+include ../Makefile.common
diff --git a/proofs/cbmc/pointwise_sum_of_products/pointwise_sum_of_products_harness.c b/proofs/cbmc/pointwise_sum_of_products/pointwise_sum_of_products_harness.c
@@ -0,0 +1,17 @@
+// Copyright (c) The mldsa-native project authors
+// SPDX-License-Identifier: Apache-2.0 OR ISC OR MIT
+
+#include "polyvec.h"
+
+#define mld_pointwise_sum_of_products \
+  MLD_NAMESPACE_KL(mld_pointwise_sum_of_products)
+int64_t mld_pointwise_sum_of_products(const mld_polyvecl *u,
+                                      const mld_polyvecl *v, unsigned int i);
+
+void harness(void)
+{
+  mld_polyvecl *u, *v;
+  unsigned int i;
+  int64_t r;
+  r = mld_pointwise_sum_of_products(u, v, i);
+}
diff --git a/proofs/cbmc/polyvecl_pointwise_acc_montgomery/Makefile b/proofs/cbmc/polyvecl_pointwise_acc_montgomery/Makefile
@@ -20,7 +20,7 @@ PROOF_SOURCES += $(PROOFDIR)/$(HARNESS_FILE).c
 PROJECT_SOURCES += $(SRCDIR)/mldsa/src/polyvec.c
 
 CHECK_FUNCTION_CONTRACTS=$(MLD_NAMESPACE)polyvecl_pointwise_acc_montgomery
-USE_FUNCTION_CONTRACTS=mld_montgomery_reduce
+USE_FUNCTION_CONTRACTS=mld_montgomery_reduce $(MLD_NAMESPACE)mld_pointwise_sum_of_products
 APPLY_LOOP_CONTRACTS=on
 USE_DYNAMIC_FRAMES=1
 
