grpc 1.81.1 → 1.85.0

data/third_party/boringssl-with-bazel/crypto/fipsmodule/ec/p256.cc.inc

@@ -1,4 +1,6 @@
 // Copyright 2020 The BoringSSL Authors
+// Copyright 2014-2016 The OpenSSL Project Authors. All Rights Reserved.
+// Copyright (c) 2014, Intel Corporation. All Rights Reserved.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -12,90 +14,347 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-// An implementation of the NIST P-256 elliptic curve point multiplication.
-// 256-bit Montgomery form for 64 and 32-bit. Field operations are generated by
-// Fiat, which lives in //third_party/fiat.
-
 #include <openssl/base.h>
 
 #include <openssl/bn.h>
 #include <openssl/ec.h>
-#include <openssl/err.h>
 #include <openssl/mem.h>
 
 #include <assert.h>
-#include <string.h>
+
+#include <iterator>
 
 #include "../../internal.h"
 #include "../delocate.h"
-#include "./internal.h"
+#include "internal.h"
 
-#if defined(BORINGSSL_HAS_UINT128)
-#include "../../../third_party/fiat/p256_64.h"
-#elif defined(OPENSSL_64_BIT)
-#include "../../../third_party/fiat/p256_64_msvc.h"
-#else
-#include "../../../third_party/fiat/p256_32.h"
-#endif
+#include "p256_internal.h"
 
+using namespace bssl;
 
-// utility functions, handwritten
-
-#if defined(OPENSSL_64_BIT)
-#define FIAT_P256_NLIMBS 4
-typedef uint64_t fiat_p256_limb_t;
-typedef uint64_t fiat_p256_felem[FIAT_P256_NLIMBS];
-static const fiat_p256_felem fiat_p256_one = {0x1, 0xffffffff00000000,
-                                              0xffffffffffffffff, 0xfffffffe};
-#else  // 64BIT; else 32BIT
-#define FIAT_P256_NLIMBS 8
-typedef uint32_t fiat_p256_limb_t;
-typedef uint32_t fiat_p256_felem[FIAT_P256_NLIMBS];
-static const fiat_p256_felem fiat_p256_one = {
-    0x1, 0x0, 0x0, 0xffffffff, 0xffffffff, 0xffffffff, 0xfffffffe, 0x0};
-#endif  // 64BIT
-
-
-static fiat_p256_limb_t fiat_p256_nz(
-    const fiat_p256_limb_t in1[FIAT_P256_NLIMBS]) {
-  fiat_p256_limb_t ret;
-  fiat_p256_nonzero(&ret, in1);
-  return ret;
+static void fiat_p256_opp_conditional(fiat_p256_felem x, crypto_word_t c) {
+  alignas(32) fiat_p256_felem n;
+  fiat_p256_opp(n, x);
+  for (size_t i = 0; i < P256_LIMBS; ++i) {
+    x[i] = br_cmov(c, n[i], x[i]);
+  }
 }
 
-static void fiat_p256_copy(fiat_p256_limb_t out[FIAT_P256_NLIMBS],
-                           const fiat_p256_limb_t in1[FIAT_P256_NLIMBS]) {
-  for (size_t i = 0; i < FIAT_P256_NLIMBS; i++) {
-    out[i] = in1[i];
+// bit returns the `i`th bit in `in`.
+static crypto_word_t bit(const uint8_t in[32], int i) {
+  if (i < 0 || i >= 256) {
+    return 0;
   }
+  return (in[i >> 3] >> (i & 7)) & 1;
 }
 
-static void fiat_p256_cmovznz(fiat_p256_limb_t out[FIAT_P256_NLIMBS],
-                              fiat_p256_limb_t t,
-                              const fiat_p256_limb_t z[FIAT_P256_NLIMBS],
-                              const fiat_p256_limb_t nz[FIAT_P256_NLIMBS]) {
-  fiat_p256_selectznz(out, !!t, z, nz);
+// point multiplication
+static void fiat_p256_select_point_16(fiat_p256_felem dst[3],
+                                      const fiat_p256_felem src[16][3],
+                                      size_t i) {
+  OPENSSL_memset(dst, 0, 3 * sizeof(fiat_p256_felem));
+  OPENSSL_CLANG_PRAGMA("clang loop unroll_count(4)")
+  for (size_t j = 0; j < 16; j++) {
+    constant_time_conditional_memxor(dst, &src[j], 3 * sizeof(fiat_p256_felem),
+                                     constant_time_eq_w(i, j));
+  }
 }
 
-static void fiat_p256_from_words(fiat_p256_felem out,
-                                 const BN_ULONG in[32 / sizeof(BN_ULONG)]) {
-  // Typically, |BN_ULONG| and |fiat_p256_limb_t| will be the same type, but on
-  // 64-bit platforms without |uint128_t|, they are different. However, on
-  // little-endian systems, |uint64_t[4]| and |uint32_t[8]| have the same
-  // layout.
-  OPENSSL_memcpy(out, in, 32);
+// Precompute multiples of `p`. p_pre_comp[i] is (i+1) * `p`.
+static void p256_point_mul(fiat_p256_felem out[3], const fiat_p256_felem p[3],
+                           const uint8_t s[32]) {
+  alignas(32) fiat_p256_felem p_pre_comp[16][3];
+  OPENSSL_memcpy(p_pre_comp[0], p, sizeof(p_pre_comp[0]));
+  for (size_t j = 2; j <= 16; ++j) {
+    if (j & 1) {
+      p256_point_add_vartime_if_doubling((uintptr_t)p_pre_comp[j - 1],
+                                         (uintptr_t)p_pre_comp[j - 2],
+                                         (uintptr_t)p_pre_comp[0]);
+    } else {
+      p256_point_double((uintptr_t)p_pre_comp[j - 1],
+                        (uintptr_t)p_pre_comp[(j - 1) / 2]);
+    }
+  }
+  alignas(32) fiat_p256_felem ret[3];
+  bool ret_is_zero = true;
+
+  for (size_t i = 51; i < 52; i--) {
+    if (!ret_is_zero) {
+      for (size_t k = 4; k < 5; k--) {
+        p256_point_double((uintptr_t)ret, (uintptr_t)ret);
+      }
+    }
+
+    crypto_word_t bits = 0;
+    OPENSSL_CLANG_PRAGMA("clang loop unroll(full)")
+    for (size_t k = 5; k < 6; k--) {
+      bits |= bit(s, i * 5 - 1 + k) << k;
+    }
+    crypto_word_t sign, digit;
+    ec_GFp_nistp_recode_scalar_bits(&sign, &digit, bits);
+
+    alignas(32) fiat_p256_felem t[3];
+    fiat_p256_select_point_16(t, p_pre_comp, digit - 1);
+    fiat_p256_opp_conditional(t[1], sign);
+
+    if (!ret_is_zero) {
+      p256_point_add_vartime_if_doubling((uintptr_t)ret, (uintptr_t)ret,
+                                         (uintptr_t)t);
+    } else {
+      OPENSSL_memcpy(ret, t, sizeof(ret));
+      ret_is_zero = false;
+    }
+  }
+  // The first loop iteration will initialize `ret`.
+  assert(!ret_is_zero);
+
+  OPENSSL_memcpy(out, ret, sizeof(ret));
 }
 
-static void fiat_p256_from_generic(fiat_p256_felem out, const EC_FELEM *in) {
-  fiat_p256_from_words(out, in->words);
+// point_mul_public
+#include "./p256_table.h"
+
+static void ec_GFp_nistp256_point_mul_public(const EC_GROUP *group,
+                                             EC_JACOBIAN *r,
+                                             const EC_SCALAR *gs,
+                                             const EC_JACOBIAN *p,
+                                             const EC_SCALAR *ps) {
+  const uint8_t *g_scalar = (uint8_t *)gs->words;
+  int8_t p_wNAF[257] = {0};
+
+  alignas(32) fiat_p256_felem p_pre_comp[1 << (4 - 1)][3];
+  OPENSSL_memcpy(p_pre_comp[0][2], &p->Z, 32);
+  if (!p256_point_iszero((uintptr_t)p_pre_comp[0])) {
+    ec_compute_wNAF(group, p_wNAF, ps, /*bits=*/256, /*w=*/4);
+    // Precompute multiples of `p`. p_pre_comp[i] is (2*i+1) * `p`.
+    OPENSSL_memcpy(p_pre_comp[0][0], &p->X, 32);
+    OPENSSL_memcpy(p_pre_comp[0][1], &p->Y, 32);
+    alignas(32) fiat_p256_felem p2[3];
+    p256_point_double((uintptr_t)p2, (uintptr_t)p_pre_comp[0]);
+    for (size_t i = 1; i < std::size(p_pre_comp); i++) {
+      p256_point_add_nz_nz_neq((uintptr_t)p_pre_comp[i],
+                               (uintptr_t)p_pre_comp[i - 1], (uintptr_t)p2);
+    }
+  }
+
+  alignas(32) fiat_p256_felem ret[3] = {};
+  bool ret_is_zero = true;  // Save some point operations, avoid 0+Q
+  for (int i = 256; i >= 0; i--) {
+    if (!ret_is_zero) {
+      p256_point_double((uintptr_t)ret, (uintptr_t)ret);
+    }
+
+    if (i <= 31) {
+      OPENSSL_CLANG_PRAGMA("clang loop unroll(full)")
+      for (size_t j = 1; j < 2; j--) {
+        crypto_word_t bits = 0;
+        OPENSSL_CLANG_PRAGMA("clang loop unroll(full)")
+        for (size_t k = 3; k < 4; k--) {
+          bits |= bit(g_scalar, i + j * 32 + k * 64) << k;
+        }
+        if (bits != 0) {
+          if (!ret_is_zero) {
+            alignas(32) fiat_p256_felem t[3];
+            fiat_p256_set_one(t[2]);
+            OPENSSL_memcpy(t, fiat_p256_g_pre_comp[j][bits - 1], 64);
+            p256_point_add_affinenz_conditional_vartime_if_doubling(
+                (uintptr_t)ret, (uintptr_t)ret, (uintptr_t)t, 1);
+            ret_is_zero = p256_point_iszero((uintptr_t)ret);
+          } else {
+            OPENSSL_memcpy(ret, fiat_p256_g_pre_comp[j][bits - 1],
+                           sizeof(fiat_p256_g_pre_comp[j][bits - 1]));
+            fiat_p256_set_one(ret[2]);
+            ret_is_zero = false;
+          }
+        }
+      }
+    }
+
+    int digit = p_wNAF[i];
+    if (digit != 0) {
+      assert(digit & 1);
+      size_t idx = (size_t)(digit < 0 ? (-digit) >> 1 : digit >> 1);
+      fiat_p256_felem t[3];
+      OPENSSL_memcpy(t, p_pre_comp[idx], sizeof(t));
+      if (digit < 0) {
+        fiat_p256_opp(t[1], t[1]);
+      }
+      if (!ret_is_zero) {
+        p256_point_add_vartime_if_doubling((uintptr_t)ret, (uintptr_t)ret,
+                                           (uintptr_t)t);
+        ret_is_zero = p256_point_iszero((uintptr_t)ret);
+      } else {
+        OPENSSL_memcpy(ret, t, sizeof(ret));
+        ret_is_zero = false;
+      }
+    }
+  }
+
+  OPENSSL_memcpy(&r->X, ret[0], 32);
+  OPENSSL_memcpy(&r->Y, ret[1], 32);
+  OPENSSL_memcpy(&r->Z, ret[2], 32);
 }
 
-static void fiat_p256_to_generic(EC_FELEM *out, const fiat_p256_felem in) {
-  // See |fiat_p256_from_words|.
-  OPENSSL_memcpy(out->words, in, 32);
+// p256_point_mul_base (small and full)
+
+// fiat_p256_select_point_affine selects the `i`th point from a precomputation
+// table and copies it to `dst`. If `n<=i`, the output is (0, 0).
+static void fiat_p256_select_point_affine(fiat_p256_felem dst[2],
+                                          const fiat_p256_felem src[/*n*/][2],
+                                          size_t i, size_t n) {
+  OPENSSL_memset(dst, 0, 2 * sizeof(fiat_p256_felem));
+  OPENSSL_CLANG_PRAGMA("clang loop unroll_count(8)")
+  for (size_t j = 0; j < n; j++) {
+    constant_time_conditional_memxor(dst, &src[j], 2 * sizeof(fiat_p256_felem),
+                                     constant_time_eq_w(i, j));
+  }
+}
+
+static void fiat_p256_conditional_zero_or_one(fiat_p256_felem out,
+                                              crypto_word_t c) {
+  fiat_p256_set_one(out);
+  crypto_word_t mask = value_barrier_w(~constant_time_is_zero_w(c));
+  for (size_t i = 0; i < P256_LIMBS; ++i) {
+    out[i] &= mask;
+  }
 }
 
-// fiat_p256_inv_square calculates |out| = |in|^{-2}
+#if defined(OPENSSL_SMALL)
+
+#ifdef __clang__
+__attribute__((noinline)) // Do not inline into unrolled loop below.
+#endif
+static void fiat_p256_select_point_affine_15(
+    fiat_p256_felem dst[2], const fiat_p256_felem src[/*n*/][2], size_t i) {
+  fiat_p256_select_point_affine(dst, src, i, 15); // Intended to be inlined.
+}
+
+static void p256_point_mul_base(fiat_p256_felem ret[3], const uint8_t s[32]) {
+  bool ret_is_zero = true;  // Save two point operations in the first round.
+  alignas(32) fiat_p256_felem t[3];
+  fiat_p256_set_one(t[2]);
+  for (size_t i = 31; i < 32; i--) {
+    if (!ret_is_zero) {
+      p256_point_double((uintptr_t)ret, (uintptr_t)ret);
+    }
+    OPENSSL_CLANG_PRAGMA("clang loop unroll(full)")
+    for (size_t j = 1; j < 2; j--) {
+      crypto_word_t bits = 0;
+      OPENSSL_CLANG_PRAGMA("clang loop unroll(full)")
+      for (size_t k = 3; k < 4; k--) {
+        bits |= bit(s, i + 32 * j + 64 * k) << k;
+      }
+      fiat_p256_select_point_affine_15(t, fiat_p256_g_pre_comp[j], bits - 1);
+
+      if (!ret_is_zero) {
+        p256_point_add_affinenz_conditional_vartime_if_doubling(
+            (uintptr_t)ret, (uintptr_t)ret, (uintptr_t)t, (uintptr_t)bits);
+      } else {
+        OPENSSL_memcpy(ret, t, sizeof(t));
+        fiat_p256_conditional_zero_or_one(ret[2], bits);
+        ret_is_zero = false;
+      }
+    }
+  }
+
+  // The first loop iteration will initialize `ret`.
+  assert(!ret_is_zero);
+}
+#else  // defined(OPENSSL_SMALL)
+
+/*
+ * Originally written by Shay Gueron (1, 2), and Vlad Krasnov (1)
+ * (1) Intel Corporation, Israel Development Center, Haifa, Israel
+ * (2) University of Haifa, Israel
+ *
+ * Reference:
+ * S.Gueron and V.Krasnov, "Fast Prime Field Elliptic Curve Cryptography with
+ *                          256 Bit Primes"
+ */
+
+// Precomputed tables for the default generator
+typedef fiat_p256_felem PRECOMP256_ROW[64][2];
+#include "p256-nistz-table.h"
+
+#ifdef __clang__
+__attribute__((noinline)) // Do not inline into unrolled loop below.
+#endif
+static void fiat_p256_select_point_affine_16(fiat_p256_felem dst[2],
+                                             const fiat_p256_felem src[16][2],
+                                             size_t i) {
+  fiat_p256_select_point_affine(dst, src, i, 16);
+}
+
+#ifdef __clang__
+__attribute__((noinline)) // Do not inline into unrolled loop below.
+#endif
+static void fiat_p256_select_point_affine_64(fiat_p256_felem dst[2],
+                                             const fiat_p256_felem src[64][2],
+                                             size_t i) {
+  fiat_p256_select_point_affine(dst, src, i, 64);
+}
+
+// See `ec_GFp_nistp_recode_scalar_bits` in util.c for details
+static crypto_word_t booth_recode_w7(crypto_word_t in) {
+  crypto_word_t s, d;
+  s = ~((in >> 7) - 1);
+  d = (1 << 8) - in - 1;
+  d = (d & s) | (in & ~s);
+  d = (d >> 1) + (d & 1);
+  return (d << 1) + (s & 1);
+}
+
+static void p256_point_mul_base(fiat_p256_felem ret[3], const uint8_t s[32]) {
+  bool ret_is_zero = true;
+  alignas(32) fiat_p256_felem t[3];
+  fiat_p256_set_one(t[2]);
+  for (size_t i = 36; i < 37; i--) {
+    // Load 7-bit windows, plus one bit below the window for `booth_recode_w7`,
+    // i.e. 8 bits from 7 * i - 1 to 7 * (i + 1).
+    constexpr size_t kMask = (1 << (7 + 1)) - 1;
+    crypto_word_t wvalue;
+    if (i == 0) {
+      wvalue = (s[0] << 1) & kMask;
+    } else {
+      size_t first_bit = 7 * i - 1;
+      size_t idx = first_bit / 8;
+      // The window may span two bytes.
+      wvalue =
+          s[idx] | (static_cast<crypto_word_t>(idx < 31 ? s[idx + 1] : 0) << 8);
+      wvalue = (wvalue >> (first_bit % 8)) & kMask;
+    }
+    wvalue = booth_recode_w7(wvalue);
+    if (i == 36) {
+      // The last window has only 4 bits instead of the full 7.
+      declassify_assert((wvalue >> 1) <= 16);
+      fiat_p256_select_point_affine_16(t, ecp_nistz256_precomputed[i],
+                                       (wvalue >> 1) - 1);
+    } else {
+      fiat_p256_select_point_affine_64(t, ecp_nistz256_precomputed[i],
+                                       (wvalue >> 1) - 1);
+    }
+    fiat_p256_opp_conditional(t[1], wvalue & 1);
+
+    if (!ret_is_zero) {
+      p256_point_add_affinenz_conditional_vartime_if_doubling(
+          (uintptr_t)ret, (uintptr_t)ret, (uintptr_t)t, wvalue >> 1);
+    } else {
+      OPENSSL_memcpy(ret, t, sizeof(t));
+      fiat_p256_conditional_zero_or_one(ret[2], wvalue >> 1);
+      ret_is_zero = false;
+    }
+  }
+
+  // The first loop iteration will initialize `ret`.
+  assert(!ret_is_zero);
+}
+#endif  // !OPENSSL_SMALL
+
+
+// FIELD-ELEMENT INVERSION
+
+
+// fiat_p256_inv_square calculates `out` = `in`^{-2}
 //
 // Based on Fermat's Little Theorem:
 //   a^p = a (mod p)
@@ -106,310 +365,182 @@ static void fiat_p256_inv_square(fiat_p256_felem out,
   // This implements the addition chain described in
   // https://briansmith.org/ecc-inversion-addition-chains-01#p256_field_inversion
   fiat_p256_felem x2, x3, x6, x12, x15, x30, x32;
-  fiat_p256_square(x2, in);   // 2^2 - 2^1
-  fiat_p256_mul(x2, x2, in);  // 2^2 - 2^0
+  p256_coord_sqr(x2, in);      // 2^2 - 2^1
+  p256_coord_mul(x2, x2, in);  // 2^2 - 2^0
 
-  fiat_p256_square(x3, x2);   // 2^3 - 2^1
-  fiat_p256_mul(x3, x3, in);  // 2^3 - 2^0
+  p256_coord_sqr(x3, x2);      // 2^3 - 2^1
+  p256_coord_mul(x3, x3, in);  // 2^3 - 2^0
 
-  fiat_p256_square(x6, x3);
+  p256_coord_sqr(x6, x3);
   for (int i = 1; i < 3; i++) {
-    fiat_p256_square(x6, x6);
-  }                           // 2^6 - 2^3
-  fiat_p256_mul(x6, x6, x3);  // 2^6 - 2^0
+    p256_coord_sqr(x6, x6);  // 2^6 - 2^3
+  }
+  p256_coord_mul(x6, x6, x3);  // 2^6 - 2^0
 
-  fiat_p256_square(x12, x6);
+  p256_coord_sqr(x12, x6);
   for (int i = 1; i < 6; i++) {
-    fiat_p256_square(x12, x12);
-  }                             // 2^12 - 2^6
-  fiat_p256_mul(x12, x12, x6);  // 2^12 - 2^0
+    p256_coord_sqr(x12, x12);  // 2^12 - 2^6
+  }
+  p256_coord_mul(x12, x12, x6);  // 2^12 - 2^0
 
-  fiat_p256_square(x15, x12);
+  p256_coord_sqr(x15, x12);
   for (int i = 1; i < 3; i++) {
-    fiat_p256_square(x15, x15);
-  }                             // 2^15 - 2^3
-  fiat_p256_mul(x15, x15, x3);  // 2^15 - 2^0
+    p256_coord_sqr(x15, x15);  // 2^15 - 2^3
+  }
+  p256_coord_mul(x15, x15, x3);  // 2^15 - 2^0
 
-  fiat_p256_square(x30, x15);
+  p256_coord_sqr(x30, x15);
   for (int i = 1; i < 15; i++) {
-    fiat_p256_square(x30, x30);
-  }                              // 2^30 - 2^15
-  fiat_p256_mul(x30, x30, x15);  // 2^30 - 2^0
+    p256_coord_sqr(x30, x30);  // 2^30 - 2^15
+  }
+  p256_coord_mul(x30, x30, x15);  // 2^30 - 2^0
 
-  fiat_p256_square(x32, x30);
-  fiat_p256_square(x32, x32);   // 2^32 - 2^2
-  fiat_p256_mul(x32, x32, x2);  // 2^32 - 2^0
+  p256_coord_sqr(x32, x30);
+  p256_coord_sqr(x32, x32);      // 2^32 - 2^2
+  p256_coord_mul(x32, x32, x2);  // 2^32 - 2^0
 
   fiat_p256_felem ret;
-  fiat_p256_square(ret, x32);
+  p256_coord_sqr(ret, x32);
   for (int i = 1; i < 31 + 1; i++) {
-    fiat_p256_square(ret, ret);
-  }                             // 2^64 - 2^32
-  fiat_p256_mul(ret, ret, in);  // 2^64 - 2^32 + 2^0
+    p256_coord_sqr(ret, ret);  // 2^64 - 2^32
+  }
+  p256_coord_mul(ret, ret, in);  // 2^64 - 2^32 + 2^0
 
   for (int i = 0; i < 96 + 32; i++) {
-    fiat_p256_square(ret, ret);
-  }                              // 2^192 - 2^160 + 2^128
-  fiat_p256_mul(ret, ret, x32);  // 2^192 - 2^160 + 2^128 + 2^32 - 2^0
+    p256_coord_sqr(ret, ret);  // 2^192 - 2^160 + 2^128
+  }
+  p256_coord_mul(ret, ret, x32);  // 2^192 - 2^160 + 2^128 + 2^32 - 2^0
 
   for (int i = 0; i < 32; i++) {
-    fiat_p256_square(ret, ret);
-  }                              // 2^224 - 2^192 + 2^160 + 2^64 - 2^32
-  fiat_p256_mul(ret, ret, x32);  // 2^224 - 2^192 + 2^160 + 2^64 - 2^0
+    p256_coord_sqr(ret, ret);  // 2^224 - 2^192 + 2^160 + 2^64 - 2^32
+  }
+  p256_coord_mul(ret, ret, x32);  // 2^224 - 2^192 + 2^160 + 2^64 - 2^0
 
   for (int i = 0; i < 30; i++) {
-    fiat_p256_square(ret, ret);
-  }                              // 2^254 - 2^222 + 2^190 + 2^94 - 2^30
-  fiat_p256_mul(ret, ret, x30);  // 2^254 - 2^222 + 2^190 + 2^94 - 2^0
+    p256_coord_sqr(ret, ret);  // 2^254 - 2^222 + 2^190 + 2^94 - 2^30
+  }
+  p256_coord_mul(ret, ret, x30);  // 2^254 - 2^222 + 2^190 + 2^94 - 2^0
 
-  fiat_p256_square(ret, ret);
-  fiat_p256_square(out, ret);  // 2^256 - 2^224 + 2^192 + 2^96 - 2^2
+  p256_coord_sqr(ret, ret);
+  p256_coord_sqr(out, ret);  // 2^256 - 2^224 + 2^192 + 2^96 - 2^2
 }
 
-// Group operations
-// ----------------
-//
-// Building on top of the field operations we have the operations on the
-// elliptic curve group itself. Points on the curve are represented in Jacobian
-// coordinates.
-//
-// Both operations were transcribed to Coq and proven to correspond to naive
-// implementations using Affine coordinates, for all suitable fields.  In the
-// Coq proofs, issues of constant-time execution and memory layout (aliasing)
-// conventions were not considered. Specification of affine coordinates:
-// <https://github.com/mit-plv/fiat-crypto/blob/79f8b5f39ed609339f0233098dee1a3c4e6b3080/src/Spec/WeierstrassCurve.v#L28>
-// As a sanity check, a proof that these points form a commutative group:
-// <https://github.com/mit-plv/fiat-crypto/blob/79f8b5f39ed609339f0233098dee1a3c4e6b3080/src/Curves/Weierstrass/AffineProofs.v#L33>
-
-// fiat_p256_point_double calculates 2*(x_in, y_in, z_in)
-//
-// The method is taken from:
-//   http://hyperelliptic.org/EFD/g1p/auto-shortw-jacobian-3.html#doubling-dbl-2001-b
-//
-// Coq transcription and correctness proof:
-// <https://github.com/mit-plv/fiat-crypto/blob/79f8b5f39ed609339f0233098dee1a3c4e6b3080/src/Curves/Weierstrass/Jacobian.v#L93>
-// <https://github.com/mit-plv/fiat-crypto/blob/79f8b5f39ed609339f0233098dee1a3c4e6b3080/src/Curves/Weierstrass/Jacobian.v#L201>
-//
-// Outputs can equal corresponding inputs, i.e., x_out == x_in is allowed.
-// while x_out == y_in is not (maybe this works, but it's not tested).
-static void fiat_p256_point_double(fiat_p256_felem x_out, fiat_p256_felem y_out,
-                                   fiat_p256_felem z_out,
-                                   const fiat_p256_felem x_in,
-                                   const fiat_p256_felem y_in,
-                                   const fiat_p256_felem z_in) {
-  fiat_p256_felem delta, gamma, beta, ftmp, ftmp2, tmptmp, alpha, fourbeta;
-  // delta = z^2
-  fiat_p256_square(delta, z_in);
-  // gamma = y^2
-  fiat_p256_square(gamma, y_in);
-  // beta = x*gamma
-  fiat_p256_mul(beta, x_in, gamma);
-
-  // alpha = 3*(x-delta)*(x+delta)
-  fiat_p256_sub(ftmp, x_in, delta);
-  fiat_p256_add(ftmp2, x_in, delta);
-
-  fiat_p256_add(tmptmp, ftmp2, ftmp2);
-  fiat_p256_add(ftmp2, ftmp2, tmptmp);
-  fiat_p256_mul(alpha, ftmp, ftmp2);
-
-  // x' = alpha^2 - 8*beta
-  fiat_p256_square(x_out, alpha);
-  fiat_p256_add(fourbeta, beta, beta);
-  fiat_p256_add(fourbeta, fourbeta, fourbeta);
-  fiat_p256_add(tmptmp, fourbeta, fourbeta);
-  fiat_p256_sub(x_out, x_out, tmptmp);
-
-  // z' = (y + z)^2 - gamma - delta
-  fiat_p256_add(delta, gamma, delta);
-  fiat_p256_add(ftmp, y_in, z_in);
-  fiat_p256_square(z_out, ftmp);
-  fiat_p256_sub(z_out, z_out, delta);
-
-  // y' = alpha*(4*beta - x') - 8*gamma^2
-  fiat_p256_sub(y_out, fourbeta, x_out);
-  fiat_p256_add(gamma, gamma, gamma);
-  fiat_p256_square(gamma, gamma);
-  fiat_p256_mul(y_out, alpha, y_out);
-  fiat_p256_add(gamma, gamma, gamma);
-  fiat_p256_sub(y_out, y_out, gamma);
-}
 
-// fiat_p256_point_add calculates (x1, y1, z1) + (x2, y2, z2)
-//
-// The method is taken from:
-//   http://hyperelliptic.org/EFD/g1p/auto-shortw-jacobian-3.html#addition-add-2007-bl,
-// adapted for mixed addition (z2 = 1, or z2 = 0 for the point at infinity).
-//
-// Coq transcription and correctness proof:
-// <https://github.com/mit-plv/fiat-crypto/blob/79f8b5f39ed609339f0233098dee1a3c4e6b3080/src/Curves/Weierstrass/Jacobian.v#L135>
-// <https://github.com/mit-plv/fiat-crypto/blob/79f8b5f39ed609339f0233098dee1a3c4e6b3080/src/Curves/Weierstrass/Jacobian.v#L205>
-//
-// This function includes a branch for checking whether the two input points
-// are equal, (while not equal to the point at infinity). This case never
-// happens during single point multiplication, so there is no timing leak for
-// ECDH or ECDSA signing.
-static void fiat_p256_point_add(fiat_p256_felem x3, fiat_p256_felem y3,
-                                fiat_p256_felem z3, const fiat_p256_felem x1,
-                                const fiat_p256_felem y1,
-                                const fiat_p256_felem z1, const int mixed,
-                                const fiat_p256_felem x2,
-                                const fiat_p256_felem y2,
-                                const fiat_p256_felem z2) {
-  fiat_p256_felem x_out, y_out, z_out;
-  fiat_p256_limb_t z1nz = fiat_p256_nz(z1);
-  fiat_p256_limb_t z2nz = fiat_p256_nz(z2);
-
-  // z1z1 = z1z1 = z1**2
-  fiat_p256_felem z1z1;
-  fiat_p256_square(z1z1, z1);
-
-  fiat_p256_felem u1, s1, two_z1z2;
-  if (!mixed) {
-    // z2z2 = z2**2
-    fiat_p256_felem z2z2;
-    fiat_p256_square(z2z2, z2);
-
-    // u1 = x1*z2z2
-    fiat_p256_mul(u1, x1, z2z2);
-
-    // two_z1z2 = (z1 + z2)**2 - (z1z1 + z2z2) = 2z1z2
-    fiat_p256_add(two_z1z2, z1, z2);
-    fiat_p256_square(two_z1z2, two_z1z2);
-    fiat_p256_sub(two_z1z2, two_z1z2, z1z1);
-    fiat_p256_sub(two_z1z2, two_z1z2, z2z2);
-
-    // s1 = y1 * z2**3
-    fiat_p256_mul(s1, z2, z2z2);
-    fiat_p256_mul(s1, s1, y1);
-  } else {
-    // We'll assume z2 = 1 (special case z2 = 0 is handled later).
-
-    // u1 = x1*z2z2
-    fiat_p256_copy(u1, x1);
-    // two_z1z2 = 2z1z2
-    fiat_p256_add(two_z1z2, z1, z1);
-    // s1 = y1 * z2**3
-    fiat_p256_copy(s1, y1);
-  }
+// Arithmetic modulo curve order
 
-  // u2 = x2*z1z1
-  fiat_p256_felem u2;
-  fiat_p256_mul(u2, x2, z1z1);
 
-  // h = u2 - u1
-  fiat_p256_felem h;
-  fiat_p256_sub(h, u2, u1);
+static void p256_order_inv0(const EC_GROUP *group, EC_SCALAR *out,
+                            const EC_SCALAR *in) {
+  // table[i] stores a power of `in` corresponding to the matching enum value.
+  enum {
+    // The following indices specify the power in binary.
+    i_1 = 0,
+    i_10,
+    i_11,
+    i_101,
+    i_111,
+    i_1010,
+    i_1111,
+    i_10101,
+    i_101010,
+    i_101111,
+    // The following indices specify 2^N-1, or N ones in a row.
+    i_x6,
+    i_x8,
+    i_x16,
+    i_x32
+  };
+  BN_ULONG table[15][P256_LIMBS];
 
-  fiat_p256_limb_t xneq = fiat_p256_nz(h);
+  // https://briansmith.org/ecc-inversion-addition-chains-01#p256_scalar_inversion
+  // This code uses specialized field arithmetic and saves 12 sqr and 13 mul.
 
-  // z_out = two_z1z2 * h
-  fiat_p256_mul(z_out, h, two_z1z2);
+  // Pre-calculate powers.
+  OPENSSL_memcpy(table[i_1], in->words, P256_LIMBS * sizeof(BN_ULONG));
 
-  // z1z1z1 = z1 * z1z1
-  fiat_p256_felem z1z1z1;
-  fiat_p256_mul(z1z1z1, z1, z1z1);
+  p256_order_sqr(group, table[i_10], table[i_1], 1);
 
-  // s2 = y2 * z1**3
-  fiat_p256_felem s2;
-  fiat_p256_mul(s2, y2, z1z1z1);
+  p256_order_mul(group, table[i_11], table[i_1], table[i_10]);
 
-  // r = (s2 - s1)*2
-  fiat_p256_felem r;
-  fiat_p256_sub(r, s2, s1);
-  fiat_p256_add(r, r, r);
+  p256_order_mul(group, table[i_101], table[i_11], table[i_10]);
 
-  fiat_p256_limb_t yneq = fiat_p256_nz(r);
+  p256_order_mul(group, table[i_111], table[i_101], table[i_10]);
 
-  fiat_p256_limb_t is_nontrivial_double = constant_time_is_zero_w(xneq | yneq) &
-                                          ~constant_time_is_zero_w(z1nz) &
-                                          ~constant_time_is_zero_w(z2nz);
-  if (constant_time_declassify_w(is_nontrivial_double)) {
-    fiat_p256_point_double(x3, y3, z3, x1, y1, z1);
-    return;
-  }
+  p256_order_sqr(group, table[i_1010], table[i_101], 1);
 
-  // I = (2h)**2
-  fiat_p256_felem i;
-  fiat_p256_add(i, h, h);
-  fiat_p256_square(i, i);
-
-  // J = h * I
-  fiat_p256_felem j;
-  fiat_p256_mul(j, h, i);
-
-  // V = U1 * I
-  fiat_p256_felem v;
-  fiat_p256_mul(v, u1, i);
-
-  // x_out = r**2 - J - 2V
-  fiat_p256_square(x_out, r);
-  fiat_p256_sub(x_out, x_out, j);
-  fiat_p256_sub(x_out, x_out, v);
-  fiat_p256_sub(x_out, x_out, v);
-
-  // y_out = r(V-x_out) - 2 * s1 * J
-  fiat_p256_sub(y_out, v, x_out);
-  fiat_p256_mul(y_out, y_out, r);
-  fiat_p256_felem s1j;
-  fiat_p256_mul(s1j, s1, j);
-  fiat_p256_sub(y_out, y_out, s1j);
-  fiat_p256_sub(y_out, y_out, s1j);
-
-  fiat_p256_cmovznz(x_out, z1nz, x2, x_out);
-  fiat_p256_cmovznz(x3, z2nz, x1, x_out);
-  fiat_p256_cmovznz(y_out, z1nz, y2, y_out);
-  fiat_p256_cmovznz(y3, z2nz, y1, y_out);
-  fiat_p256_cmovznz(z_out, z1nz, z2, z_out);
-  fiat_p256_cmovznz(z3, z2nz, z1, z_out);
-}
+  p256_order_mul(group, table[i_1111], table[i_1010], table[i_101]);
 
-#include "./p256_table.h"
+  p256_order_sqr(group, table[i_10101], table[i_1010], 1);
+  p256_order_mul(group, table[i_10101], table[i_10101], table[i_1]);
+
+  p256_order_sqr(group, table[i_101010], table[i_10101], 1);
+
+  p256_order_mul(group, table[i_101111], table[i_101010], table[i_101]);
+
+  p256_order_mul(group, table[i_x6], table[i_101010], table[i_10101]);
+
+  p256_order_sqr(group, table[i_x8], table[i_x6], 2);
+  p256_order_mul(group, table[i_x8], table[i_x8], table[i_11]);
+
+  p256_order_sqr(group, table[i_x16], table[i_x8], 8);
+  p256_order_mul(group, table[i_x16], table[i_x16], table[i_x8]);
+
+  p256_order_sqr(group, table[i_x32], table[i_x16], 16);
+  p256_order_mul(group, table[i_x32], table[i_x32], table[i_x16]);
 
-// fiat_p256_select_point_affine selects the |idx-1|th point from a
-// precomputation table and copies it to out. If |idx| is zero, the output is
-// the point at infinity.
-static void fiat_p256_select_point_affine(
-    const fiat_p256_limb_t idx, size_t size,
-    const fiat_p256_felem pre_comp[/*size*/][2], fiat_p256_felem out[3]) {
-  OPENSSL_memset(out, 0, sizeof(fiat_p256_felem) * 3);
-  for (size_t i = 0; i < size; i++) {
-    fiat_p256_limb_t mismatch = i ^ (idx - 1);
-    fiat_p256_cmovznz(out[0], mismatch, pre_comp[i][0], out[0]);
-    fiat_p256_cmovznz(out[1], mismatch, pre_comp[i][1], out[1]);
+  // Compute `in` raised to the order-2.
+  p256_order_sqr(group, out->words, table[i_x32], 64);
+  p256_order_mul(group, out->words, out->words, table[i_x32]);
+  static const struct {
+    uint8_t p, i;
+  } kChain[27] = {{32, i_x32},    {6, i_101111}, {5, i_111},    {4, i_11},
+                  {5, i_1111},    {5, i_10101},  {4, i_101},    {3, i_101},
+                  {3, i_101},     {5, i_111},    {9, i_101111}, {6, i_1111},
+                  {2, i_1},       {5, i_1},      {6, i_1111},   {5, i_111},
+                  {4, i_111},     {5, i_111},    {5, i_101},    {3, i_11},
+                  {10, i_101111}, {2, i_11},     {5, i_11},     {5, i_11},
+                  {3, i_1},       {7, i_10101},  {6, i_1111}};
+  for (const auto &step : kChain) {
+    p256_order_sqr(group, out->words, out->words, step.p);
+    p256_order_mul(group, out->words, out->words, table[step.i]);
   }
-  fiat_p256_cmovznz(out[2], idx, out[2], fiat_p256_one);
 }
 
-// fiat_p256_select_point selects the |idx|th point from a precomputation table
-// and copies it to out.
-static void fiat_p256_select_point(const fiat_p256_limb_t idx, size_t size,
-                                   const fiat_p256_felem pre_comp[/*size*/][3],
-                                   fiat_p256_felem out[3]) {
-  OPENSSL_memset(out, 0, sizeof(fiat_p256_felem) * 3);
-  for (size_t i = 0; i < size; i++) {
-    fiat_p256_limb_t mismatch = i ^ idx;
-    fiat_p256_cmovznz(out[0], mismatch, pre_comp[i][0], out[0]);
-    fiat_p256_cmovznz(out[1], mismatch, pre_comp[i][1], out[1]);
-    fiat_p256_cmovznz(out[2], mismatch, pre_comp[i][2], out[2]);
+static int p256_order_mont_inv_vartime(const EC_GROUP *group, EC_SCALAR *out,
+                                       const EC_SCALAR *in) {
+#if !defined(OPENSSL_NO_ASM) && \
+    (defined(OPENSSL_X86_64) || defined(OPENSSL_AARCH64))
+#if defined(OPENSSL_X86_64)
+  if (!CRYPTO_is_AVX_capable()) {
+    // No AVX support; fallback to generic code.
+    return ec_simple_scalar_to_montgomery_inv_vartime(group, out, in);
   }
-}
+#endif
 
-// fiat_p256_get_bit returns the |i|th bit in |in|.
-static crypto_word_t fiat_p256_get_bit(const EC_SCALAR *in, int i) {
-  if (i < 0 || i >= 256) {
+  assert(group->order.N.width == P256_LIMBS);
+  if (!beeu_mod_inverse_vartime(out->words, in->words, group->order.N.d)) {
     return 0;
   }
-#if defined(OPENSSL_64_BIT)
-  static_assert(sizeof(BN_ULONG) == 8, "BN_ULONG was not 64-bit");
-  return (in->words[i >> 6] >> (i & 63)) & 1;
+
+  // The result should be returned in the Montgomery domain.
+  ec_scalar_to_montgomery(group, out, out);
+  return 1;
 #else
-  static_assert(sizeof(BN_ULONG) == 4, "BN_ULONG was not 32-bit");
-  return (in->words[i >> 5] >> (i & 31)) & 1;
+  return ec_simple_scalar_to_montgomery_inv_vartime(group, out, in);
 #endif
 }
 
-// OPENSSL EC_METHOD FUNCTIONS
+
+// OPENSSL EC_METHOD WRAPPERS
+
+
+static void fiat_p256_from_generic(fiat_p256_felem out, const EC_FELEM *in) {
+  OPENSSL_memcpy(out, in->words, 32);
+}
+
+static void fiat_p256_to_generic(EC_FELEM *out, const fiat_p256_felem in) {
+  OPENSSL_memcpy(out->words, in, 32);
+}
 
 // Takes the Jacobian coordinates (X, Y, Z) of a point and returns (X', Y') =
 // (X/Z^2, Y/Z^3).
@@ -426,260 +557,25 @@ static int ec_GFp_nistp256_point_get_affine_coordinates(
   fiat_p256_from_generic(z1, &point->Z);
   fiat_p256_inv_square(z2, z1);
 
-  if (x_out != NULL) {
+  if (x_out != nullptr) {
     fiat_p256_felem x;
     fiat_p256_from_generic(x, &point->X);
-    fiat_p256_mul(x, x, z2);
+    p256_coord_mul(x, x, z2);
     fiat_p256_to_generic(x_out, x);
   }
 
-  if (y_out != NULL) {
+  if (y_out != nullptr) {
     fiat_p256_felem y;
     fiat_p256_from_generic(y, &point->Y);
-    fiat_p256_square(z2, z2);  // z^-4
-    fiat_p256_mul(y, y, z1);   // y * z
-    fiat_p256_mul(y, y, z2);   // y * z^-3
+    p256_coord_sqr(z2, z2);    // z^-4
+    p256_coord_mul(y, y, z1);  // y * z
+    p256_coord_mul(y, y, z2);  // y * z^-3
     fiat_p256_to_generic(y_out, y);
   }
 
   return 1;
 }
 
-static void ec_GFp_nistp256_add(const EC_GROUP *group, EC_JACOBIAN *r,
-                                const EC_JACOBIAN *a, const EC_JACOBIAN *b) {
-  fiat_p256_felem x1, y1, z1, x2, y2, z2;
-  fiat_p256_from_generic(x1, &a->X);
-  fiat_p256_from_generic(y1, &a->Y);
-  fiat_p256_from_generic(z1, &a->Z);
-  fiat_p256_from_generic(x2, &b->X);
-  fiat_p256_from_generic(y2, &b->Y);
-  fiat_p256_from_generic(z2, &b->Z);
-  fiat_p256_point_add(x1, y1, z1, x1, y1, z1, 0 /* both Jacobian */, x2, y2,
-                      z2);
-  fiat_p256_to_generic(&r->X, x1);
-  fiat_p256_to_generic(&r->Y, y1);
-  fiat_p256_to_generic(&r->Z, z1);
-}
-
-static void ec_GFp_nistp256_dbl(const EC_GROUP *group, EC_JACOBIAN *r,
-                                const EC_JACOBIAN *a) {
-  fiat_p256_felem x, y, z;
-  fiat_p256_from_generic(x, &a->X);
-  fiat_p256_from_generic(y, &a->Y);
-  fiat_p256_from_generic(z, &a->Z);
-  fiat_p256_point_double(x, y, z, x, y, z);
-  fiat_p256_to_generic(&r->X, x);
-  fiat_p256_to_generic(&r->Y, y);
-  fiat_p256_to_generic(&r->Z, z);
-}
-
-static void ec_GFp_nistp256_point_mul(const EC_GROUP *group, EC_JACOBIAN *r,
-                                      const EC_JACOBIAN *p,
-                                      const EC_SCALAR *scalar) {
-  fiat_p256_felem p_pre_comp[17][3];
-  OPENSSL_memset(&p_pre_comp, 0, sizeof(p_pre_comp));
-  // Precompute multiples.
-  fiat_p256_from_generic(p_pre_comp[1][0], &p->X);
-  fiat_p256_from_generic(p_pre_comp[1][1], &p->Y);
-  fiat_p256_from_generic(p_pre_comp[1][2], &p->Z);
-  for (size_t j = 2; j <= 16; ++j) {
-    if (j & 1) {
-      fiat_p256_point_add(p_pre_comp[j][0], p_pre_comp[j][1], p_pre_comp[j][2],
-                          p_pre_comp[1][0], p_pre_comp[1][1], p_pre_comp[1][2],
-                          0, p_pre_comp[j - 1][0], p_pre_comp[j - 1][1],
-                          p_pre_comp[j - 1][2]);
-    } else {
-      fiat_p256_point_double(p_pre_comp[j][0], p_pre_comp[j][1],
-                             p_pre_comp[j][2], p_pre_comp[j / 2][0],
-                             p_pre_comp[j / 2][1], p_pre_comp[j / 2][2]);
-    }
-  }
-
-  // Set nq to the point at infinity.
-  fiat_p256_felem nq[3] = {{0}, {0}, {0}}, ftmp, tmp[3];
-
-  // Loop over |scalar| msb-to-lsb, incorporating |p_pre_comp| every 5th round.
-  int skip = 1;  // Save two point operations in the first round.
-  for (size_t i = 255; i < 256; i--) {
-    // double
-    if (!skip) {
-      fiat_p256_point_double(nq[0], nq[1], nq[2], nq[0], nq[1], nq[2]);
-    }
-
-    // do other additions every 5 doublings
-    if (i % 5 == 0) {
-      crypto_word_t bits = fiat_p256_get_bit(scalar, i + 4) << 5;
-      bits |= fiat_p256_get_bit(scalar, i + 3) << 4;
-      bits |= fiat_p256_get_bit(scalar, i + 2) << 3;
-      bits |= fiat_p256_get_bit(scalar, i + 1) << 2;
-      bits |= fiat_p256_get_bit(scalar, i) << 1;
-      bits |= fiat_p256_get_bit(scalar, i - 1);
-      crypto_word_t sign, digit;
-      ec_GFp_nistp_recode_scalar_bits(&sign, &digit, bits);
-
-      // select the point to add or subtract, in constant time.
-      fiat_p256_select_point((fiat_p256_limb_t)digit, 17,
-                             (const fiat_p256_felem(*)[3])p_pre_comp, tmp);
-      fiat_p256_opp(ftmp, tmp[1]);  // (X, -Y, Z) is the negative point.
-      fiat_p256_cmovznz(tmp[1], (fiat_p256_limb_t)sign, tmp[1], ftmp);
-
-      if (!skip) {
-        fiat_p256_point_add(nq[0], nq[1], nq[2], nq[0], nq[1], nq[2],
-                            0 /* mixed */, tmp[0], tmp[1], tmp[2]);
-      } else {
-        fiat_p256_copy(nq[0], tmp[0]);
-        fiat_p256_copy(nq[1], tmp[1]);
-        fiat_p256_copy(nq[2], tmp[2]);
-        skip = 0;
-      }
-    }
-  }
-
-  fiat_p256_to_generic(&r->X, nq[0]);
-  fiat_p256_to_generic(&r->Y, nq[1]);
-  fiat_p256_to_generic(&r->Z, nq[2]);
-}
-
-static void ec_GFp_nistp256_point_mul_base(const EC_GROUP *group,
-                                           EC_JACOBIAN *r,
-                                           const EC_SCALAR *scalar) {
-  // Set nq to the point at infinity.
-  fiat_p256_felem nq[3] = {{0}, {0}, {0}}, tmp[3];
-
-  int skip = 1;  // Save two point operations in the first round.
-  for (size_t i = 31; i < 32; i--) {
-    if (!skip) {
-      fiat_p256_point_double(nq[0], nq[1], nq[2], nq[0], nq[1], nq[2]);
-    }
-
-    // First, look 32 bits upwards.
-    crypto_word_t bits = fiat_p256_get_bit(scalar, i + 224) << 3;
-    bits |= fiat_p256_get_bit(scalar, i + 160) << 2;
-    bits |= fiat_p256_get_bit(scalar, i + 96) << 1;
-    bits |= fiat_p256_get_bit(scalar, i + 32);
-    // Select the point to add, in constant time.
-    fiat_p256_select_point_affine((fiat_p256_limb_t)bits, 15,
-                                  fiat_p256_g_pre_comp[1], tmp);
-
-    if (!skip) {
-      fiat_p256_point_add(nq[0], nq[1], nq[2], nq[0], nq[1], nq[2],
-                          1 /* mixed */, tmp[0], tmp[1], tmp[2]);
-    } else {
-      fiat_p256_copy(nq[0], tmp[0]);
-      fiat_p256_copy(nq[1], tmp[1]);
-      fiat_p256_copy(nq[2], tmp[2]);
-      skip = 0;
-    }
-
-    // Second, look at the current position.
-    bits = fiat_p256_get_bit(scalar, i + 192) << 3;
-    bits |= fiat_p256_get_bit(scalar, i + 128) << 2;
-    bits |= fiat_p256_get_bit(scalar, i + 64) << 1;
-    bits |= fiat_p256_get_bit(scalar, i);
-    // Select the point to add, in constant time.
-    fiat_p256_select_point_affine((fiat_p256_limb_t)bits, 15,
-                                  fiat_p256_g_pre_comp[0], tmp);
-    fiat_p256_point_add(nq[0], nq[1], nq[2], nq[0], nq[1], nq[2], 1 /* mixed */,
-                        tmp[0], tmp[1], tmp[2]);
-  }
-
-  fiat_p256_to_generic(&r->X, nq[0]);
-  fiat_p256_to_generic(&r->Y, nq[1]);
-  fiat_p256_to_generic(&r->Z, nq[2]);
-}
-
-static void ec_GFp_nistp256_point_mul_public(const EC_GROUP *group,
-                                             EC_JACOBIAN *r,
-                                             const EC_SCALAR *g_scalar,
-                                             const EC_JACOBIAN *p,
-                                             const EC_SCALAR *p_scalar) {
-#define P256_WSIZE_PUBLIC 4
-  // Precompute multiples of |p|. p_pre_comp[i] is (2*i+1) * |p|.
-  fiat_p256_felem p_pre_comp[1 << (P256_WSIZE_PUBLIC - 1)][3];
-  fiat_p256_from_generic(p_pre_comp[0][0], &p->X);
-  fiat_p256_from_generic(p_pre_comp[0][1], &p->Y);
-  fiat_p256_from_generic(p_pre_comp[0][2], &p->Z);
-  fiat_p256_felem p2[3];
-  fiat_p256_point_double(p2[0], p2[1], p2[2], p_pre_comp[0][0],
-                         p_pre_comp[0][1], p_pre_comp[0][2]);
-  for (size_t i = 1; i < OPENSSL_ARRAY_SIZE(p_pre_comp); i++) {
-    fiat_p256_point_add(p_pre_comp[i][0], p_pre_comp[i][1], p_pre_comp[i][2],
-                        p_pre_comp[i - 1][0], p_pre_comp[i - 1][1],
-                        p_pre_comp[i - 1][2], 0 /* not mixed */, p2[0], p2[1],
-                        p2[2]);
-  }
-
-  // Set up the coefficients for |p_scalar|.
-  int8_t p_wNAF[257];
-  ec_compute_wNAF(group, p_wNAF, p_scalar, 256, P256_WSIZE_PUBLIC);
-
-  // Set |ret| to the point at infinity.
-  int skip = 1;  // Save some point operations.
-  fiat_p256_felem ret[3] = {{0}, {0}, {0}};
-  for (int i = 256; i >= 0; i--) {
-    if (!skip) {
-      fiat_p256_point_double(ret[0], ret[1], ret[2], ret[0], ret[1], ret[2]);
-    }
-
-    // For the |g_scalar|, we use the precomputed table without the
-    // constant-time lookup.
-    if (i <= 31) {
-      // First, look 32 bits upwards.
-      crypto_word_t bits = fiat_p256_get_bit(g_scalar, i + 224) << 3;
-      bits |= fiat_p256_get_bit(g_scalar, i + 160) << 2;
-      bits |= fiat_p256_get_bit(g_scalar, i + 96) << 1;
-      bits |= fiat_p256_get_bit(g_scalar, i + 32);
-      if (bits != 0) {
-        size_t index = (size_t)(bits - 1);
-        fiat_p256_point_add(ret[0], ret[1], ret[2], ret[0], ret[1], ret[2],
-                            1 /* mixed */, fiat_p256_g_pre_comp[1][index][0],
-                            fiat_p256_g_pre_comp[1][index][1],
-                            fiat_p256_one);
-        skip = 0;
-      }
-
-      // Second, look at the current position.
-      bits = fiat_p256_get_bit(g_scalar, i + 192) << 3;
-      bits |= fiat_p256_get_bit(g_scalar, i + 128) << 2;
-      bits |= fiat_p256_get_bit(g_scalar, i + 64) << 1;
-      bits |= fiat_p256_get_bit(g_scalar, i);
-      if (bits != 0) {
-        size_t index = (size_t)(bits - 1);
-        fiat_p256_point_add(ret[0], ret[1], ret[2], ret[0], ret[1], ret[2],
-                            1 /* mixed */, fiat_p256_g_pre_comp[0][index][0],
-                            fiat_p256_g_pre_comp[0][index][1],
-                            fiat_p256_one);
-        skip = 0;
-      }
-    }
-
-    int digit = p_wNAF[i];
-    if (digit != 0) {
-      assert(digit & 1);
-      size_t idx = (size_t)(digit < 0 ? (-digit) >> 1 : digit >> 1);
-      fiat_p256_felem *y = &p_pre_comp[idx][1], tmp;
-      if (digit < 0) {
-        fiat_p256_opp(tmp, p_pre_comp[idx][1]);
-        y = &tmp;
-      }
-      if (!skip) {
-        fiat_p256_point_add(ret[0], ret[1], ret[2], ret[0], ret[1], ret[2],
-                            0 /* not mixed */, p_pre_comp[idx][0], *y,
-                            p_pre_comp[idx][2]);
-      } else {
-        fiat_p256_copy(ret[0], p_pre_comp[idx][0]);
-        fiat_p256_copy(ret[1], *y);
-        fiat_p256_copy(ret[2], p_pre_comp[idx][2]);
-        skip = 0;
-      }
-    }
-  }
-
-  fiat_p256_to_generic(&r->X, ret[0]);
-  fiat_p256_to_generic(&r->Y, ret[1]);
-  fiat_p256_to_generic(&r->Z, ret[2]);
-}
-
 static int ec_GFp_nistp256_cmp_x_coordinate(const EC_GROUP *group,
                                             const EC_JACOBIAN *p,
                                             const EC_SCALAR *r) {
@@ -692,11 +588,11 @@ static int ec_GFp_nistp256_cmp_x_coordinate(const EC_GROUP *group,
   // not.
   fiat_p256_felem Z2_mont;
   fiat_p256_from_generic(Z2_mont, &p->Z);
-  fiat_p256_mul(Z2_mont, Z2_mont, Z2_mont);
+  p256_coord_mul(Z2_mont, Z2_mont, Z2_mont);
 
   fiat_p256_felem r_Z2;
-  fiat_p256_from_words(r_Z2, r->words);  // r < order < p, so this is valid.
-  fiat_p256_mul(r_Z2, r_Z2, Z2_mont);
+  OPENSSL_memcpy(r_Z2, r->words, 32);  // r < order < p, so this is valid.
+  p256_coord_mul(r_Z2, r_Z2, Z2_mont);
 
   fiat_p256_felem X;
   fiat_p256_from_generic(X, &p->X);
@@ -708,7 +604,7 @@ static int ec_GFp_nistp256_cmp_x_coordinate(const EC_GROUP *group,
 
   // During signing the x coefficient is reduced modulo the group order.
   // Therefore there is a small possibility, less than 1/2^128, that group_order
-  // < p.x < P. in that case we need not only to compare against |r| but also to
+  // < p.x < P. in that case we need not only to compare against `r` but also to
   // compare against r+group_order.
   assert(group->field.N.width == group->order.N.width);
   EC_FELEM tmp;
@@ -717,7 +613,7 @@ static int ec_GFp_nistp256_cmp_x_coordinate(const EC_GROUP *group,
   if (carry == 0 &&
       bn_less_than_words(tmp.words, group->field.N.d, group->field.N.width)) {
     fiat_p256_from_generic(r_Z2, &tmp);
-    fiat_p256_mul(r_Z2, r_Z2, Z2_mont);
+    p256_coord_mul(r_Z2, r_Z2, Z2_mont);
     if (OPENSSL_memcmp(&r_Z2, &X, sizeof(r_Z2)) == 0) {
       return 1;
     }
@@ -726,6 +622,58 @@ static int ec_GFp_nistp256_cmp_x_coordinate(const EC_GROUP *group,
   return 0;
 }
 
+static void ec_GFp_nistp256_point_mul(const EC_GROUP *group, EC_JACOBIAN *r,
+                                      const EC_JACOBIAN *p,
+                                      const EC_SCALAR *scalar) {
+  fiat_p256_felem t[3];
+  fiat_p256_from_generic(t[0], &p->X);
+  fiat_p256_from_generic(t[1], &p->Y);
+  fiat_p256_from_generic(t[2], &p->Z);
+  p256_point_mul(t, t, (uint8_t *)scalar->words);
+
+  fiat_p256_to_generic(&r->X, t[0]);
+  fiat_p256_to_generic(&r->Y, t[1]);
+  fiat_p256_to_generic(&r->Z, t[2]);
+}
+
+static void ec_GFp_nistp256_point_mul_base(const EC_GROUP *group,
+                                           EC_JACOBIAN *r, const EC_SCALAR *s) {
+  alignas(32) fiat_p256_felem ret[3];
+  p256_point_mul_base(ret, (uint8_t *)s->words);
+  fiat_p256_to_generic(&r->X, ret[0]);
+  fiat_p256_to_generic(&r->Y, ret[1]);
+  fiat_p256_to_generic(&r->Z, ret[2]);
+}
+
+static void ec_GFp_nistp256_add(const EC_GROUP *group, EC_JACOBIAN *r,
+                                const EC_JACOBIAN *a, const EC_JACOBIAN *b) {
+  fiat_p256_felem p[3], q[3];
+  fiat_p256_from_generic(p[0], &a->X);
+  fiat_p256_from_generic(p[1], &a->Y);
+  fiat_p256_from_generic(p[2], &a->Z);
+  fiat_p256_from_generic(q[0], &b->X);
+  fiat_p256_from_generic(q[1], &b->Y);
+  fiat_p256_from_generic(q[2], &b->Z);
+  p256_point_add_vartime_if_doubling((uintptr_t)p, (uintptr_t)p, (uintptr_t)q);
+  fiat_p256_to_generic(&r->X, p[0]);
+  fiat_p256_to_generic(&r->Y, p[1]);
+  fiat_p256_to_generic(&r->Z, p[2]);
+}
+
+static void ec_GFp_nistp256_dbl(const EC_GROUP *group, EC_JACOBIAN *r,
+                                const EC_JACOBIAN *a) {
+  fiat_p256_felem p[3];
+  fiat_p256_from_generic(p[0], &a->X);
+  fiat_p256_from_generic(p[1], &a->Y);
+  fiat_p256_from_generic(p[2], &a->Z);
+  p256_point_double((uintptr_t)p, (uintptr_t)p);
+  fiat_p256_to_generic(&r->X, p[0]);
+  fiat_p256_to_generic(&r->Y, p[1]);
+  fiat_p256_to_generic(&r->Z, p[2]);
+}
+
+BSSL_NAMESPACE_BEGIN
+
 DEFINE_METHOD_FUNCTION(EC_METHOD, EC_GFp_nistp256_method) {
   out->point_get_affine_coordinates =
       ec_GFp_nistp256_point_get_affine_coordinates;
@@ -734,16 +682,9 @@ DEFINE_METHOD_FUNCTION(EC_METHOD, EC_GFp_nistp256_method) {
   out->mul = ec_GFp_nistp256_point_mul;
   out->mul_base = ec_GFp_nistp256_point_mul_base;
   out->mul_public = ec_GFp_nistp256_point_mul_public;
-  out->felem_mul = ec_GFp_mont_felem_mul;
-  out->felem_sqr = ec_GFp_mont_felem_sqr;
-  out->felem_to_bytes = ec_GFp_mont_felem_to_bytes;
-  out->felem_from_bytes = ec_GFp_mont_felem_from_bytes;
-  out->felem_reduce = ec_GFp_mont_felem_reduce;
-  // TODO(davidben): This should use the specialized field arithmetic
-  // implementation, rather than the generic one.
-  out->felem_exp = ec_GFp_mont_felem_exp;
-  out->scalar_inv0_montgomery = ec_simple_scalar_inv0_montgomery;
-  out->scalar_to_montgomery_inv_vartime =
-      ec_simple_scalar_to_montgomery_inv_vartime;
+  out->scalar_inv0_montgomery = p256_order_inv0;
+  out->scalar_to_montgomery_inv_vartime = p256_order_mont_inv_vartime;
   out->cmp_x_coordinate = ec_GFp_nistp256_cmp_x_coordinate;
 }
+
+BSSL_NAMESPACE_END