math/Makefile.am, symm/Makefile.am: Use `--no-install' on oddball tests.

[catacomb] / math / fgoldi.c

/* -*-c-*-
 *
 * Arithmetic in the Goldilocks field GF(2^448 - 2^224 - 1)
 *
 * (c) 2017 Straylight/Edgeware
 */

/*----- Licensing notice --------------------------------------------------*
 *
 * This file is part of Catacomb.
 *
 * Catacomb is free software; you can redistribute it and/or modify
 * it under the terms of the GNU Library General Public License as
 * published by the Free Software Foundation; either version 2 of the
 * License, or (at your option) any later version.
 *
 * Catacomb is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU Library General Public License for more details.
 *
 * You should have received a copy of the GNU Library General Public
 * License along with Catacomb; if not, write to the Free
 * Software Foundation, Inc., 59 Temple Place - Suite 330, Boston,
 * MA 02111-1307, USA.
 */

/*----- Header files ------------------------------------------------------*/

#include "config.h"

#include "ct.h"
#include "fgoldi.h"

/*----- Basic setup -------------------------------------------------------*
 *
 * Let φ = 2^224; then p = φ^2 - φ - 1, and, in GF(p), we have φ^2 = φ + 1
 * (hence the name).
 */

typedef fgoldi_piece piece;

#if FGOLDI_IMPL == 28
/* We represent an element of GF(p) as 16 28-bit signed integer pieces x_i:
 * x = SUM_{0<=i<16} x_i 2^(28i).
 */

                       typedef  int64  dblpiece;
typedef uint32 upiece; typedef uint64 udblpiece;
#define PIECEWD(i) 28
#define NPIECE 16
#define P p28

#define B28 0x10000000u
#define B27 0x08000000u
#define M28 0x0fffffffu
#define M27 0x07ffffffu
#define M32 0xffffffffu

#elif FGOLDI_IMPL == 12
/* We represent an element of GF(p) as 40 signed integer pieces x_i: x =
 * SUM_{0<=i<40} x_i 2^ceil(224i/20).  Pieces i with i == 0 (mod 5) are 12
 * bits wide; the others are 11 bits wide, so they form eight groups of 56
 * bits.
 */

                       typedef  int32  dblpiece;
typedef uint16 upiece; typedef uint32 udblpiece;
#define PIECEWD(i) ((i)%5 ? 11 : 12)
#define NPIECE 40
#define P p12

#define B12 0x1000u
#define B11 0x0800u
#define B10 0x0400u
#define M12 0xfffu
#define M11 0x7ffu
#define M10 0x3ffu
#define M8 0xffu

#endif

/*----- Debugging machinery -----------------------------------------------*/

#if defined(FGOLDI_DEBUG) || defined(TEST_RIG)

#include <stdio.h>

#include "mp.h"
#include "mptext.h"

static mp *get_pgoldi(void)
{
  mp *p = MP_NEW, *t = MP_NEW;

  p = mp_setbit(p, MP_ZERO, 448);
  t = mp_setbit(t, MP_ZERO, 224);
  p = mp_sub(p, p, t);
  p = mp_sub(p, p, MP_ONE);
  mp_drop(t);
  return (p);
}

DEF_FDUMP(fdump, piece, PIECEWD, NPIECE, 56, get_pgoldi())

#endif

/*----- Loading and storing -----------------------------------------------*/

/* --- @fgoldi_load@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to store the result
 *              @const octet xv[56]@ = source to read
 *
 * Returns:     ---
 *
 * Use:         Reads an element of %$\gf{2^{448} - 2^{224} - 19}$% in
 *              external representation from @xv@ and stores it in @z@.
 *
 *              External representation is little-endian base-256.  Some
 *              elements have multiple encodings, which are not produced by
 *              correct software; use of noncanonical encodings is not an
 *              error, and toleration of them is considered a performance
 *              feature.
 */

void fgoldi_load(fgoldi *z, const octet xv[56])
{
#if FGOLDI_IMPL == 28

  unsigned i;
  uint32 xw[14];
  piece b, c;

  /* First, read the input value as words. */
  for (i = 0; i < 14; i++) xw[i] = LOAD32_L(xv + 4*i);

  /* Extract unsigned 28-bit pieces from the words. */
  z->P[ 0] = (xw[ 0] >> 0)&M28;
  z->P[ 7] = (xw[ 6] >> 4)&M28;
  z->P[ 8] = (xw[ 7] >> 0)&M28;
  z->P[15] = (xw[13] >> 4)&M28;
  for (i = 1; i < 7; i++) {
    z->P[i + 0] = ((xw[i + 0] << (4*i)) | (xw[i - 1] >> (32 - 4*i)))&M28;
    z->P[i + 8] = ((xw[i + 7] << (4*i)) | (xw[i + 6] >> (32 - 4*i)))&M28;
  }

  /* Convert the nonnegative pieces into a balanced signed representation, so
   * each piece ends up in the interval |z_i| <= 2^27.  For each piece, if
   * its top bit is set, lend a bit leftwards; in the case of z_15, reduce
   * this bit by adding it onto z_0 and z_8, since this is the φ^2 bit, and
   * φ^2 = φ + 1.  We delay this carry until after all of the pieces have
   * been balanced.  If we don't do this, then we have to do a more expensive
   * test for nonzeroness to decide whether to lend a bit leftwards rather
   * than just testing a single bit.
   *
   * Note that we don't try for a canonical representation here: both upper
   * and lower bounds are achievable.
   */
  b = z->P[15]&B27; z->P[15] -= b << 1; c = b >> 27;
  for (i = NPIECE - 1; i--; )
    { b = z->P[i]&B27; z->P[i] -= b << 1; z->P[i + 1] += b >> 27; }
  z->P[0] += c; z->P[8] += c;

#elif FGOLDI_IMPL == 12

  unsigned i, j, n, w, b;
  uint32 a;
  int c;

  /* First, convert the bytes into nonnegative pieces. */
  for (i = j = a = n = 0, w = PIECEWD(0); i < 56; i++) {
    a |= (uint32)xv[i] << n; n += 8;
    if (n >= w) {
      z->P[j++] = a&MASK(w);
      a >>= w; n -= w; w = PIECEWD(j);
    }
  }

  /* Convert the nonnegative pieces into a balanced signed representation, so
   * each piece ends up in the interval |z_i| <= 2^11 + 1.
   */
  b = z->P[39]&B10; z->P[39] -= b << 1; c = b >> 10;
  for (i = NPIECE - 1; i--; ) {
    w = PIECEWD(i) - 1;
    b = z->P[i]&BIT(w);
    z->P[i] -= b << 1;
    z->P[i + 1] += b >> w;
  }
  z->P[0] += c; z->P[20] += c;

#endif
}

/* --- @fgoldi_store@ --- *
 *
 * Arguments:   @octet zv[56]@ = where to write the result
 *              @const fgoldi *x@ = the field element to write
 *
 * Returns:     ---
 *
 * Use:         Stores a field element in the given octet vector in external
 *              representation.  A canonical encoding is always stored.
 */

void fgoldi_store(octet zv[56], const fgoldi *x)
{
#if FGOLDI_IMPL == 28

  piece y[NPIECE], yy[NPIECE], c, d;
  uint32 u, v;
  mask32 m;
  unsigned i;

  for (i = 0; i < NPIECE; i++) y[i] = x->P[i];

  /* First, propagate the carries.  By the end of this, we'll have all of the
   * the pieces canonically sized and positive, and maybe there'll be
   * (signed) carry out.  The carry c is in { -1, 0, +1 }, and the remaining
   * value will be in the half-open interval [0, φ^2).  The whole represented
   * value is then y + φ^2 c.
   *
   * Assume that we start out with |y_i| <= 2^30.  We start off by cutting
   * off and reducing the carry c_15 from the topmost piece, y_15.  This
   * leaves 0 <= y_15 < 2^28; and we'll have |c_15| <= 4.  We'll add this
   * onto y_0 and y_8, and propagate the carries.  It's very clear that we'll
   * end up with |y + (φ + 1) c_15 - φ^2/2| << φ^2.
   *
   * Here, the y_i are signed, so we must be cautious about bithacking them.
   */
  c = ASR(piece, y[15], 28); y[15] = (upiece)y[15]&M28; y[8] += c;
  for (i = 0; i < NPIECE; i++)
    { y[i] += c; c = ASR(piece, y[i], 28); y[i] = (upiece)y[i]&M28; }

  /* Now we have a slightly fiddly job to do.  If c = +1, or if c = 0 and
   * y >= p, then we should subtract p from the whole value; if c = -1 then
   * we should add p; and otherwise we should do nothing.
   *
   * But conditional behaviour is bad, m'kay.  So here's what we do instead.
   *
   * The first job is to sort out what we wanted to do.  If c = -1 then we
   * want to (a) invert the constant addend and (b) feed in a carry-in;
   * otherwise, we don't.
   */
  m = SIGN(c)&M28;
  d = m&1;

  /* Now do the addition/subtraction.  Remember that all of the y_i are
   * nonnegative, so shifting and masking are safe and easy.
   */
      d += y[0] + (1 ^ m); yy[0] = d&M28; d >>= 28;
  for (i = 1; i < 8; i++)
    { d += y[i] +      m;  yy[i] = d&M28; d >>= 28; }
      d += y[8] + (1 ^ m); yy[8] = d&M28; d >>= 28;
  for (i = 9; i < 16; i++)
    { d += y[i] +      m;  yy[i] = d&M28; d >>= 28; }

  /* The final carry-out is in d; since we only did addition, and the y_i are
   * nonnegative, then d is in { 0, 1 }.  We want to keep y', rather than y,
   * if (a) c /= 0 (in which case we know that the old value was
   * unsatisfactory), or (b) if d = 1 (in which case, if c = 0, we know that
   * the subtraction didn't cause a borrow, so we must be in the case where
   * p <= y < φ^2.
   */
  m = NONZEROP(c) | ~NONZEROP(d - 1);
  for (i = 0; i < NPIECE; i++) y[i] = (yy[i]&m) | (y[i]&~m);

  /* Extract 32-bit words from the value. */
  for (i = 0; i < 7; i++) {
    u = ((y[i + 0] >> (4*i)) | ((uint32)y[i + 1] << (28 - 4*i)))&M32;
    v = ((y[i + 8] >> (4*i)) | ((uint32)y[i + 9] << (28 - 4*i)))&M32;
    STORE32_L(zv + 4*i,      u);
    STORE32_L(zv + 4*i + 28, v);
  }

#elif FGOLDI_IMPL == 12

  piece y[NPIECE], yy[NPIECE], c, d;
  uint32 a;
  mask32 m, mm;
  unsigned i, j, n, w;

  for (i = 0; i < NPIECE; i++) y[i] = x->P[i];

  /* First, propagate the carries.  By the end of this, we'll have all of the
   * the pieces canonically sized and positive, and maybe there'll be
   * (signed) carry out.  The carry c is in { -1, 0, +1 }, and the remaining
   * value will be in the half-open interval [0, φ^2).  The whole represented
   * value is then y + φ^2 c.
   *
   * Assume that we start out with |y_i| <= 2^14.  We start off by cutting
   * off and reducing the carry c_39 from the topmost piece, y_39.  This
   * leaves 0 <= y_39 < 2^11; and we'll have |c_39| <= 16.  We'll add this
   * onto y_0 and y_20, and propagate the carries.  It's very clear that
   * we'll end up with |y + (φ + 1) c_39 - φ^2/2| << φ^2.
   *
   * Here, the y_i are signed, so we must be cautious about bithacking them.
   */
  c = ASR(piece, y[39], 11); y[39] = (piece)y[39]&M11; y[20] += c;
  for (i = 0; i < NPIECE; i++) {
    w = PIECEWD(i); m = (1 << w) - 1;
    y[i] += c; c = ASR(piece, y[i], w); y[i] = (upiece)y[i]&m;
  }

  /* Now we have a slightly fiddly job to do.  If c = +1, or if c = 0 and
   * y >= p, then we should subtract p from the whole value; if c = -1 then
   * we should add p; and otherwise we should do nothing.
   *
   * But conditional behaviour is bad, m'kay.  So here's what we do instead.
   *
   * The first job is to sort out what we wanted to do.  If c = -1 then we
   * want to (a) invert the constant addend and (b) feed in a carry-in;
   * otherwise, we don't.
   */
  mm = SIGN(c);
  d = m&1;

  /* Now do the addition/subtraction.  Remember that all of the y_i are
   * nonnegative, so shifting and masking are safe and easy.
   */
    d += y[ 0] + (1 ^ (mm&M12)); yy[ 0] = d&M12; d >>= 12;
  for (i = 1; i < 20; i++) {
    w = PIECEWD(i); m = MASK(w);
    d += y[ i] +      (mm&m);    yy[ i] = d&m;   d >>= w;
  }
    d += y[20] + (1 ^ (mm&M12)); yy[20] = d&M12; d >>= 12;
  for (i = 21; i < 40; i++) {
    w = PIECEWD(i); m = MASK(w);
    d += y[ i] +      (mm&m);    yy[ i] = d&m;   d >>= w;
  }

  /* The final carry-out is in d; since we only did addition, and the y_i are
   * nonnegative, then d is in { 0, 1 }.  We want to keep y', rather than y,
   * if (a) c /= 0 (in which case we know that the old value was
   * unsatisfactory), or (b) if d = 1 (in which case, if c = 0, we know that
   * the subtraction didn't cause a borrow, so we must be in the case where
   * p <= y < φ^2.
   */
  m = NONZEROP(c) | ~NONZEROP(d - 1);
  for (i = 0; i < NPIECE; i++) y[i] = (yy[i]&m) | (y[i]&~m);

  /* Convert that back into octets. */
  for (i = j = a = n = 0; i < NPIECE; i++) {
    a |= (uint32)y[i] << n; n += PIECEWD(i);
    while (n >= 8) { zv[j++] = a&M8; a >>= 8; n -= 8; }
  }

#endif
}

/* --- @fgoldi_set@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to write the result
 *              @int a@ = a small-ish constant
 *
 * Returns:     ---
 *
 * Use:         Sets @z@ to equal @a@.
 */

void fgoldi_set(fgoldi *x, int a)
{
  unsigned i;

  x->P[0] = a;
  for (i = 1; i < NPIECE; i++) x->P[i] = 0;
}

/*----- Basic arithmetic --------------------------------------------------*/

/* --- @fgoldi_add@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@ or @y@)
 *              @const fgoldi *x, *y@ = two operands
 *
 * Returns:     ---
 *
 * Use:         Set @z@ to the sum %$x + y$%.
 */

void fgoldi_add(fgoldi *z, const fgoldi *x, const fgoldi *y)
{
  unsigned i;
  for (i = 0; i < NPIECE; i++) z->P[i] = x->P[i] + y->P[i];
}

/* --- @fgoldi_sub@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@ or @y@)
 *              @const fgoldi *x, *y@ = two operands
 *
 * Returns:     ---
 *
 * Use:         Set @z@ to the difference %$x - y$%.
 */

void fgoldi_sub(fgoldi *z, const fgoldi *x, const fgoldi *y)
{
  unsigned i;
  for (i = 0; i < NPIECE; i++) z->P[i] = x->P[i] - y->P[i];
}

/* --- @fgoldi_neg@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@)
 *              @const fgoldi *x@ = an operand
 *
 * Returns:     ---
 *
 * Use:         Set @z = -x@.
 */

void fgoldi_neg(fgoldi *z, const fgoldi *x)
{
  unsigned i;
  for (i = 0; i < NPIECE; i++) z->P[i] = -x->P[i];
}

/*----- Constant-time utilities -------------------------------------------*/

/* --- @fgoldi_pick2@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@ or @y@)
 *              @const fgoldi *x, *y@ = two operands
 *              @uint32 m@ = a mask
 *
 * Returns:     ---
 *
 * Use:         If @m@ is zero, set @z = y@; if @m@ is all-bits-set, then set
 *              @z = x@.  If @m@ has some other value, then scramble @z@ in
 *              an unhelpful way.
 */

void fgoldi_pick2(fgoldi *z, const fgoldi *x, const fgoldi *y, uint32 m)
{
  mask32 mm = FIX_MASK32(m);
  unsigned i;
  for (i = 0; i < NPIECE; i++) z->P[i] = PICK2(x->P[i], y->P[i], mm);
}

/* --- @fgoldi_pickn@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result
 *              @const fgoldi *v@ = a table of entries
 *              @size_t n@ = the number of entries in @v@
 *              @size_t i@ = an index
 *
 * Returns:     ---
 *
 * Use:         If @0 <= i < n < 32@ then set @z = v[i]@.  If @n >= 32@ then
 *              do something unhelpful; otherwise, if @i >= n@ then set @z@
 *              to zero.
 */

void fgoldi_pickn(fgoldi *z, const fgoldi *v, size_t n, size_t i)
{
  uint32 b = (uint32)1 << (31 - i);
  mask32 m;
  unsigned j;

  for (j = 0; j < NPIECE; j++) z->P[j] = 0;
  while (n--) {
    m = SIGN(b);
    for (j = 0; j < NPIECE; j++) CONDPICK(z->P[j], v->P[j], m);
    v++; b <<= 1;
  }
}

/* --- @fgoldi_condswap@ --- *
 *
 * Arguments:   @fgoldi *x, *y@ = two operands
 *              @uint32 m@ = a mask
 *
 * Returns:     ---
 *
 * Use:         If @m@ is zero, do nothing; if @m@ is all-bits-set, then
 *              exchange @x@ and @y@.  If @m@ has some other value, then
 *              scramble @x@ and @y@ in an unhelpful way.
 */

void fgoldi_condswap(fgoldi *x, fgoldi *y, uint32 m)
{
  unsigned i;
  mask32 mm = FIX_MASK32(m);

  for (i = 0; i < NPIECE; i++) CONDSWAP(x->P[i], y->P[i], mm);
}

/* --- @fgoldi_condneg@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@)
 *              @const fgoldi *x@ = an operand
 *              @uint32 m@ = a mask
 *
 * Returns:     ---
 *
 * Use:         If @m@ is zero, set @z = x@; if @m@ is all-bits-set, then set
 *              @z = -x@.  If @m@ has some other value then scramble @z@ in
 *              an unhelpful way.
 */

void fgoldi_condneg(fgoldi *z, const fgoldi *x, uint32 m)
{
#ifdef NEG_TWOC
  mask32 m_xor = FIX_MASK32(m);
  piece m_add = m&1;
# define CONDNEG(x) (((x) ^ m_xor) + m_add)
#else
  int s = PICK2(-1, +1, m);
# define CONDNEG(x) (s*(x))
#endif

  unsigned i;
  for (i = 0; i < NPIECE; i++) z->P[i] = CONDNEG(x->P[i]);

#undef CONDNEG
}

/*----- Multiplication ----------------------------------------------------*/

#if FGOLDI_IMPL == 28

/* Let B = 2^63 - 1 be the largest value such that +B and -B can be
 * represented in a double-precision piece.  On entry, it must be the case
 * that |X_i| <= M <= B - 2^27 for some M.  If this is the case, then, on
 * exit, we will have |Z_i| <= 2^27 + M/2^27.
 */
#define CARRY_REDUCE(z, x) do {                                         \
  dblpiece _t[NPIECE], _c;                                              \
  unsigned _i;                                                          \
                                                                        \
  /* Bias the input pieces.  This keeps the carries and so on centred   \
   * around zero rather than biased positive.                           \
   */                                                                   \
  for (_i = 0; _i < NPIECE; _i++) _t[_i] = (x)[_i] + B27;               \
                                                                        \
  /* Calculate the reduced pieces.  Careful with the bithacking. */     \
  _c = ASR(dblpiece, _t[15], 28);                                       \
  (z)[0] = (dblpiece)((udblpiece)_t[0]&M28) - B27 + _c;                 \
  for (_i = 1; _i < NPIECE; _i++) {                                     \
    (z)[_i] = (dblpiece)((udblpiece)_t[_i]&M28) - B27 +                 \
      ASR(dblpiece, _t[_i - 1], 28);                                    \
  }                                                                     \
  (z)[8] += _c;                                                         \
} while (0)

#elif FGOLDI_IMPL == 12

static void carry_reduce(dblpiece x[NPIECE])
{
  /* Initial bounds: we assume |x_i| < 2^31 - 2^27. */

  unsigned i, j;
  dblpiece c;

  /* The result is nearly canonical, because we do sequential carry
   * propagation, because smaller processors are more likely to prefer the
   * smaller working set than the instruction-level parallelism.
   *
   * Start at x_37; truncate it to 10 bits, and propagate the carry to x_38.
   * Truncate x_38 to 10 bits, and add the carry onto x_39.  Truncate x_39 to
   * 10 bits, and add the carry onto x_0 and x_20.  And so on.
   *
   * Once we reach x_37 for the second time, we start with |x_37| <= 2^10.
   * The carry into x_37 is at most 2^21; so the carry out into x_38 has
   * magnitude at most 2^10.  In turn, |x_38| <= 2^10 before the carry, so is
   * now no more than 2^11 in magnitude, and the carry out into x_39 is at
   * most 1.  This leaves |x_39| <= 2^10 + 1 after carry propagation.
   *
   * Be careful with the bit hacking because the quantities involved are
   * signed.
   */

  /* For each piece, we bias it so that floor division (as done by an
   * arithmetic right shift) and modulus (as done by bitwise-AND) does the
   * right thing.
   */
#define CARRY(i, wd, b, m) do {                                         \
  x[i] += (b);                                                          \
  c = ASR(dblpiece, x[i], (wd));                                        \
  x[i] = (dblpiece)((udblpiece)x[i]&(m)) - (b);                         \
} while (0)

                             {             CARRY(37, 11, B10, M11);      }
                             { x[38] += c; CARRY(38, 11, B10, M11);      }
                             { x[39] += c; CARRY(39, 11, B10, M11);      }
                               x[20] += c;
  for (i = 0; i < 35; ) {
                             {  x[i] += c; CARRY( i, 12, B11, M12); i++; }
    for (j = i + 4; i < j; ) {  x[i] += c; CARRY( i, 11, B10, M11); i++; }
  }
                             {  x[i] += c; CARRY( i, 12, B11, M12); i++; }
  while (i < 39)             {  x[i] += c; CARRY( i, 11, B10, M11); i++; }
  x[39] += c;
}

#endif

/* --- @fgoldi_mulconst@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@)
 *              @const fgoldi *x@ = an operand
 *              @long a@ = a small-ish constant; %$|a| < 2^{20}$%.
 *
 * Returns:     ---
 *
 * Use:         Set @z@ to the product %$a x$%.
 */

void fgoldi_mulconst(fgoldi *z, const fgoldi *x, long a)
{
  unsigned i;
  dblpiece zz[NPIECE], aa = a;

  for (i = 0; i < NPIECE; i++) zz[i] = aa*x->P[i];
#if FGOLDI_IMPL == 28
  CARRY_REDUCE(z->P, zz);
#elif FGOLDI_IMPL == 12
  carry_reduce(zz);
  for (i = 0; i < NPIECE; i++) z->P[i] = zz[i];
#endif
}

/* --- @fgoldi_mul@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@ or @y@)
 *              @const fgoldi *x, *y@ = two operands
 *
 * Returns:     ---
 *
 * Use:         Set @z@ to the product %$x y$%.
 */

void fgoldi_mul(fgoldi *z, const fgoldi *x, const fgoldi *y)
{
  dblpiece zz[NPIECE], u[NPIECE];
  piece ab[NPIECE/2], cd[NPIECE/2];
  const piece
    *a = x->P + NPIECE/2, *b = x->P,
    *c = y->P + NPIECE/2, *d = y->P;
  unsigned i, j;

#if FGOLDI_IMPL == 28

#  define M(x,i, y,j) ((dblpiece)(x)[i]*(y)[j])

#elif FGOLDI_IMPL == 12

  static const unsigned short off[39] = {
      0,  12,  23,  34,  45,  56,  68,  79,  90, 101,
    112, 124, 135, 146, 157, 168, 180, 191, 202, 213,
    224, 236, 247, 258, 269, 280, 292, 303, 314, 325,
    336, 348, 359, 370, 381, 392, 404, 415, 426
  };

#define M(x,i, y,j)                                                     \
  (((dblpiece)(x)[i]*(y)[j]) << (off[i] + off[j] - off[(i) + (j)]))

#endif

  /* Behold the magic.
   *
   * Write x = a φ + b, and y = c φ + d.  Then x y = a c φ^2 +
   * (a d + b c) φ + b d.  Karatsuba and Ofman observed that a d + b c =
   * (a + b) (c + d) - a c - b d, saving a multiplication, and Hamburg chose
   * the prime p so that φ^2 = φ + 1.  So
   *
   *    x y = ((a + b) (c + d) - b d) φ + a c + b d
   */

  for (i = 0; i < NPIECE; i++) zz[i] = 0;

  /* Our first job will be to calculate (1 - φ) b d, and write the result
   * into z.  As we do this, an interesting thing will happen.  Write
   * b d = u φ + v; then (1 - φ) b d = u φ + v - u φ^2 - v φ = (1 - φ) v - u.
   * So, what we do is to write the product end-swapped and negated, and then
   * we'll subtract the (negated, remember) high half from the low half.
   */
  for (i = 0; i < NPIECE/2; i++) {
    for (j = 0; j < NPIECE/2 - i; j++)
      zz[i + j + NPIECE/2] -= M(b,i, d,j);
    for (; j < NPIECE/2; j++)
      zz[i + j - NPIECE/2] -= M(b,i, d,j);
  }
  for (i = 0; i < NPIECE/2; i++)
    zz[i] -= zz[i + NPIECE/2];

  /* Next, we add on a c.  There are no surprises here. */
  for (i = 0; i < NPIECE/2; i++)
    for (j = 0; j < NPIECE/2; j++)
      zz[i + j] += M(a,i, c,j);

  /* Now, calculate a + b and c + d. */
  for (i = 0; i < NPIECE/2; i++)
    { ab[i] = a[i] + b[i]; cd[i] = c[i] + d[i]; }

  /* Finally (for the multiplication) we must add on (a + b) (c + d) φ.
   * Write (a + b) (c + d) as u φ + v; then we actually want u φ^2 + v φ =
   * v φ + (1 + φ) u.  We'll store u in a temporary place and add it on
   * twice.
   */
  for (i = 0; i < NPIECE; i++) u[i] = 0;
  for (i = 0; i < NPIECE/2; i++) {
    for (j = 0; j < NPIECE/2 - i; j++)
      zz[i + j + NPIECE/2] += M(ab,i, cd,j);
    for (; j < NPIECE/2; j++)
      u[i + j - NPIECE/2] += M(ab,i, cd,j);
  }
  for (i = 0; i < NPIECE/2; i++)
    { zz[i] += u[i]; zz[i + NPIECE/2] += u[i]; }

#undef M

#if FGOLDI_IMPL == 28
  /* That wraps it up for the multiplication.  Let's figure out some bounds.
   * Fortunately, Karatsuba is a polynomial identity, so all of the pieces
   * end up the way they'd be if we'd done the thing the easy way, which
   * simplifies the analysis.  Suppose we started with |x_i|, |y_i| <= 9/5
   * 2^28.  The overheads in the result are given by the coefficients of
   *
   *    ((u^16 - 1)/(u - 1))^2 mod u^16 - u^8 - 1
   *
   * the greatest of which is 38.  So |z_i| <= 38*81/25*2^56 < 2^63.
   *
   * Anyway, a round of `CARRY_REDUCE' will leave us with |z_i| < 2^27 +
   * 2^36; and a second round will leave us with |z_i| < 2^27 + 512.
   */
  for (i = 0; i < 2; i++) CARRY_REDUCE(zz, zz);
  for (i = 0; i < NPIECE; i++) z->P[i] = zz[i];
#elif FGOLDI_IMPL == 12
  carry_reduce(zz);
  for (i = 0; i < NPIECE; i++) z->P[i] = zz[i];
#endif
}

/* --- @fgoldi_sqr@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@ or @y@)
 *              @const fgoldi *x@ = an operand
 *
 * Returns:     ---
 *
 * Use:         Set @z@ to the square %$x^2$%.
 */

void fgoldi_sqr(fgoldi *z, const fgoldi *x)
{
#if FGOLDI_IMPL == 28

  dblpiece zz[NPIECE], u[NPIECE];
  piece ab[NPIECE];
  const piece *a = x->P + NPIECE/2, *b = x->P;
  unsigned i, j;

#  define M(x,i, y,j) ((dblpiece)(x)[i]*(y)[j])

  /* The magic is basically the same as `fgoldi_mul' above.  We write
   * x = a φ + b and use Karatsuba and the special prime shape.  This time,
   * we have
   *
   *    x^2 = ((a + b)^2 - b^2) φ + a^2 + b^2
   */

  for (i = 0; i < NPIECE; i++) zz[i] = 0;

  /* Our first job will be to calculate (1 - φ) b^2, and write the result
   * into z.  Again, this interacts pleasantly with the prime shape.
   */
  for (i = 0; i < NPIECE/4; i++) {
    zz[2*i + NPIECE/2] -= M(b,i, b,i);
    for (j = i + 1; j < NPIECE/2 - i; j++)
      zz[i + j + NPIECE/2] -= 2*M(b,i, b,j);
    for (; j < NPIECE/2; j++)
      zz[i + j - NPIECE/2] -= 2*M(b,i, b,j);
  }
  for (; i < NPIECE/2; i++) {
    zz[2*i - NPIECE/2] -= M(b,i, b,i);
    for (j = i + 1; j < NPIECE/2; j++)
      zz[i + j - NPIECE/2] -= 2*M(b,i, b,j);
  }
  for (i = 0; i < NPIECE/2; i++)
    zz[i] -= zz[i + NPIECE/2];

  /* Next, we add on a^2.  There are no surprises here. */
  for (i = 0; i < NPIECE/2; i++) {
    zz[2*i] += M(a,i, a,i);
    for (j = i + 1; j < NPIECE/2; j++)
      zz[i + j] += 2*M(a,i, a,j);
  }

  /* Now, calculate a + b. */
  for (i = 0; i < NPIECE/2; i++)
    ab[i] = a[i] + b[i];

  /* Finally (for the multiplication) we must add on (a + b)^2 φ.
   * Write (a + b)^2 as u φ + v; then we actually want (u + v) φ + u.  We'll
   * store u in a temporary place and add it on twice.
   */
  for (i = 0; i < NPIECE; i++) u[i] = 0;
  for (i = 0; i < NPIECE/4; i++) {
    zz[2*i + NPIECE/2] += M(ab,i, ab,i);
    for (j = i + 1; j < NPIECE/2 - i; j++)
      zz[i + j + NPIECE/2] += 2*M(ab,i, ab,j);
    for (; j < NPIECE/2; j++)
      u[i + j - NPIECE/2] += 2*M(ab,i, ab,j);
  }
  for (; i < NPIECE/2; i++) {
    u[2*i - NPIECE/2] += M(ab,i, ab,i);
    for (j = i + 1; j < NPIECE/2; j++)
      u[i + j - NPIECE/2] += 2*M(ab,i, ab,j);
  }
  for (i = 0; i < NPIECE/2; i++)
    { zz[i] += u[i]; zz[i + NPIECE/2] += u[i]; }

#undef M

  /* Finally, carrying. */
  for (i = 0; i < 2; i++) CARRY_REDUCE(zz, zz);
  for (i = 0; i < NPIECE; i++) z->P[i] = zz[i];

#elif FGOLDI_IMPL == 12
  fgoldi_mul(z, x, x);
#endif
}

/*----- More advanced operations ------------------------------------------*/

/* --- @fgoldi_inv@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@)
 *              @const fgoldi *x@ = an operand
 *
 * Returns:     ---
 *
 * Use:         Stores in @z@ the multiplicative inverse %$x^{-1}$%.  If
 *              %$x = 0$% then @z@ is set to zero.  This is considered a
 *              feature.
 */

void fgoldi_inv(fgoldi *z, const fgoldi *x)
{
  fgoldi t, u;
  unsigned i;

#define SQRN(z, x, n) do {                                              \
  fgoldi_sqr((z), (x));                                                 \
  for (i = 1; i < (n); i++) fgoldi_sqr((z), (z));                       \
} while (0)

  /* Calculate x^-1 = x^(p - 2) = x^(2^448 - 2^224 - 3), which also handles
   * x = 0 as intended.  The addition chain is home-made.
   */                                   /* step | value */
  fgoldi_sqr(&u, x);                    /*    1 | 2 */
  fgoldi_mul(&t, &u, x);                /*    2 | 3 */
  SQRN(&u, &t, 2);                      /*    4 | 12 */
  fgoldi_mul(&t, &u, &t);               /*    5 | 15 */
  SQRN(&u, &t, 4);                      /*    9 | 240 */
  fgoldi_mul(&u, &u, &t);               /*   10 | 255 = 2^8 - 1 */
  SQRN(&u, &u, 4);                      /*   14 | 2^12 - 16 */
  fgoldi_mul(&t, &u, &t);               /*   15 | 2^12 - 1 */
  SQRN(&u, &t, 12);                     /*   27 | 2^24 - 2^12 */
  fgoldi_mul(&u, &u, &t);               /*   28 | 2^24 - 1 */
  SQRN(&u, &u, 12);                     /*   40 | 2^36 - 2^12 */
  fgoldi_mul(&t, &u, &t);               /*   41 | 2^36 - 1 */
  fgoldi_sqr(&t, &t);                   /*   42 | 2^37 - 2 */
  fgoldi_mul(&t, &t, x);                /*   43 | 2^37 - 1 */
  SQRN(&u, &t, 37);                     /*   80 | 2^74 - 2^37 */
  fgoldi_mul(&u, &u, &t);               /*   81 | 2^74 - 1 */
  SQRN(&u, &u, 37);                     /*  118 | 2^111 - 2^37 */
  fgoldi_mul(&t, &u, &t);               /*  119 | 2^111 - 1 */
  SQRN(&u, &t, 111);                    /*  230 | 2^222 - 2^111 */
  fgoldi_mul(&t, &u, &t);               /*  231 | 2^222 - 1 */
  fgoldi_sqr(&u, &t);                   /*  232 | 2^223 - 2 */
  fgoldi_mul(&u, &u, x);                /*  233 | 2^223 - 1 */
  SQRN(&u, &u, 223);                    /*  456 | 2^446 - 2^223 */
  fgoldi_mul(&t, &u, &t);               /*  457 | 2^446 - 2^222 - 1 */
  SQRN(&t, &t, 2);                      /*  459 | 2^448 - 2^224 - 4 */
  fgoldi_mul(z, &t, x);                 /*  460 | 2^448 - 2^224 - 3 */

#undef SQRN
}

/* --- @fgoldi_quosqrt@ --- *
 *
 * Arguments:   @fgoldi *z@ = where to put the result (may alias @x@ or @y@)
 *              @const fgoldi *x, *y@ = two operands
 *
 * Returns:     Zero if successful, @-1@ if %$x/y$% is not a square.
 *
 * Use:         Stores in @z@ the one of the square roots %$\pm\sqrt{x/y}$%.
 *              If %$x = y = 0% then the result is zero; if %$y = 0$% but %$x
 *              \ne 0$% then the operation fails.  If you wanted a specific
 *              square root then you'll have to pick it yourself.
 */

int fgoldi_quosqrt(fgoldi *z, const fgoldi *x, const fgoldi *y)
{
  fgoldi t, u, v;
  octet xb[56], b0[56];
  int32 rc = -1;
  mask32 m;
  unsigned i;

#define SQRN(z, x, n) do {                                              \
  fgoldi_sqr((z), (x));                                                 \
  for (i = 1; i < (n); i++) fgoldi_sqr((z), (z));                       \
} while (0)

  /* This is, fortunately, significantly easier than the equivalent problem
   * in GF(2^255 - 19), since p == 3 (mod 4).
   *
   * If x/y is square, then so is v = y^2 x/y = x y, and therefore u has
   * order r = (p - 1)/2.  Let w = v^{(p-3)/4}.  Then w^2 = v^{(p-3)/2} =
   * u^{r-1} = 1/v = 1/x y.  Clearly, then, (x w)^2 = x^2/x y = x/y, so x w
   * is one of the square roots we seek.
   *
   * The addition chain, then, is a prefix of the previous one.
   */
  fgoldi_mul(&v, x, y);

  fgoldi_sqr(&u, &v);                   /*    1 | 2 */
  fgoldi_mul(&t, &u, &v);               /*    2 | 3 */
  SQRN(&u, &t, 2);                      /*    4 | 12 */
  fgoldi_mul(&t, &u, &t);               /*    5 | 15 */
  SQRN(&u, &t, 4);                      /*    9 | 240 */
  fgoldi_mul(&u, &u, &t);               /*   10 | 255 = 2^8 - 1 */
  SQRN(&u, &u, 4);                      /*   14 | 2^12 - 16 */
  fgoldi_mul(&t, &u, &t);               /*   15 | 2^12 - 1 */
  SQRN(&u, &t, 12);                     /*   27 | 2^24 - 2^12 */
  fgoldi_mul(&u, &u, &t);               /*   28 | 2^24 - 1 */
  SQRN(&u, &u, 12);                     /*   40 | 2^36 - 2^12 */
  fgoldi_mul(&t, &u, &t);               /*   41 | 2^36 - 1 */
  fgoldi_sqr(&t, &t);                   /*   42 | 2^37 - 2 */
  fgoldi_mul(&t, &t, &v);               /*   43 | 2^37 - 1 */
  SQRN(&u, &t, 37);                     /*   80 | 2^74 - 2^37 */
  fgoldi_mul(&u, &u, &t);               /*   81 | 2^74 - 1 */
  SQRN(&u, &u, 37);                     /*  118 | 2^111 - 2^37 */
  fgoldi_mul(&t, &u, &t);               /*  119 | 2^111 - 1 */
  SQRN(&u, &t, 111);                    /*  230 | 2^222 - 2^111 */
  fgoldi_mul(&t, &u, &t);               /*  231 | 2^222 - 1 */
  fgoldi_sqr(&u, &t);                   /*  232 | 2^223 - 2 */
  fgoldi_mul(&u, &u, &v);               /*  233 | 2^223 - 1 */
  SQRN(&u, &u, 223);                    /*  456 | 2^446 - 2^223 */
  fgoldi_mul(&t, &u, &t);               /*  457 | 2^446 - 2^222 - 1 */

#undef SQRN

  /* Now we must decide whether the answer was right.  We should have z^2 =
   * x/y, so y z^2 = x.
   *
   * The easiest way to compare is to encode.  This isn't as wasteful as it
   * sounds: the hard part is normalizing the representations, which we have
   * to do anyway.
   */
  fgoldi_mul(z, x, &t);
  fgoldi_sqr(&t, z);
  fgoldi_mul(&t, &t, y);
  fgoldi_store(xb, x);
  fgoldi_store(b0, &t);
  m = -ct_memeq(xb, b0, 56);
  rc = PICK2(0, rc, m);
  return (rc);
}

/*----- Test rig ----------------------------------------------------------*/

#ifdef TEST_RIG

#include <mLib/report.h>
#include <mLib/str.h>
#include <mLib/testrig.h>

static void fixdstr(dstr *d)
{
  if (d->len > 56)
    die(1, "invalid length for fgoldi");
  else if (d->len < 56) {
    dstr_ensure(d, 56);
    memset(d->buf + d->len, 0, 56 - d->len);
    d->len = 56;
  }
}

static void cvt_fgoldi(const char *buf, dstr *d)
{
  dstr dd = DSTR_INIT;

  type_hex.cvt(buf, &dd); fixdstr(&dd);
  dstr_ensure(d, sizeof(fgoldi)); d->len = sizeof(fgoldi);
  fgoldi_load((fgoldi *)d->buf, (const octet *)dd.buf);
  dstr_destroy(&dd);
}

static void dump_fgoldi(dstr *d, FILE *fp)
  { fdump(stderr, "???", (const piece *)d->buf); }

static void cvt_fgoldi_ref(const char *buf, dstr *d)
  { type_hex.cvt(buf, d); fixdstr(d); }

static void dump_fgoldi_ref(dstr *d, FILE *fp)
{
  fgoldi x;

  fgoldi_load(&x, (const octet *)d->buf);
  fdump(stderr, "???", x.P);
}

static int eq(const fgoldi *x, dstr *d)
  { octet b[56]; fgoldi_store(b, x); return (memcmp(b, d->buf, 56) == 0); }

static const test_type
  type_fgoldi = { cvt_fgoldi, dump_fgoldi },
  type_fgoldi_ref = { cvt_fgoldi_ref, dump_fgoldi_ref };

#define TEST_UNOP(op)                                                   \
  static int vrf_##op(dstr dv[])                                        \
  {                                                                     \
    fgoldi *x = (fgoldi *)dv[0].buf;                                    \
    fgoldi z, zz;                                                       \
    int ok = 1;                                                         \
                                                                        \
    fgoldi_##op(&z, x);                                                 \
    if (!eq(&z, &dv[1])) {                                              \
      ok = 0;                                                           \
      fprintf(stderr, "failed!\n");                                     \
      fdump(stderr, "x", x->P);                                         \
      fdump(stderr, "calc", z.P);                                       \
      fgoldi_load(&zz, (const octet *)dv[1].buf);                       \
      fdump(stderr, "z", zz.P);                                         \
    }                                                                   \
                                                                        \
    return (ok);                                                        \
  }

TEST_UNOP(sqr)
TEST_UNOP(inv)
TEST_UNOP(neg)

#define TEST_BINOP(op)                                                  \
  static int vrf_##op(dstr dv[])                                        \
  {                                                                     \
    fgoldi *x = (fgoldi *)dv[0].buf, *y = (fgoldi *)dv[1].buf;          \
    fgoldi z, zz;                                                       \
    int ok = 1;                                                         \
                                                                        \
    fgoldi_##op(&z, x, y);                                              \
    if (!eq(&z, &dv[2])) {                                              \
      ok = 0;                                                           \
      fprintf(stderr, "failed!\n");                                     \
      fdump(stderr, "x", x->P);                                         \
      fdump(stderr, "y", y->P);                                         \
      fdump(stderr, "calc", z.P);                                       \
      fgoldi_load(&zz, (const octet *)dv[2].buf);                       \
      fdump(stderr, "z", zz.P);                                         \
    }                                                                   \
                                                                        \
    return (ok);                                                        \
  }

TEST_BINOP(add)
TEST_BINOP(sub)
TEST_BINOP(mul)

static int vrf_mulc(dstr dv[])
{
  fgoldi *x = (fgoldi *)dv[0].buf;
  long a = *(const long *)dv[1].buf;
  fgoldi z, zz;
  int ok = 1;

  fgoldi_mulconst(&z, x, a);
  if (!eq(&z, &dv[2])) {
    ok = 0;
    fprintf(stderr, "failed!\n");
    fdump(stderr, "x", x->P);
    fprintf(stderr, "a = %ld\n", a);
    fdump(stderr, "calc", z.P);
    fgoldi_load(&zz, (const octet *)dv[2].buf);
    fdump(stderr, "z", zz.P);
  }

  return (ok);
}

static int vrf_condneg(dstr dv[])
{
  fgoldi *x = (fgoldi *)dv[0].buf;
  uint32 m = *(uint32 *)dv[1].buf;
  fgoldi z;
  int ok = 1;

  fgoldi_condneg(&z, x, m);
  if (!eq(&z, &dv[2])) {
    ok = 0;
    fprintf(stderr, "failed!\n");
    fdump(stderr, "x", x->P);
    fprintf(stderr, "m = 0x%08lx\n", (unsigned long)m);
    fdump(stderr, "calc z", z.P);
    fgoldi_load(&z, (const octet *)dv[1].buf);
    fdump(stderr, "want z", z.P);
  }

  return (ok);
}

static int vrf_pick2(dstr dv[])
{
  fgoldi *x = (fgoldi *)dv[0].buf, *y = (fgoldi *)dv[1].buf;
  uint32 m = *(uint32 *)dv[2].buf;
  fgoldi z;
  int ok = 1;

  fgoldi_pick2(&z, x, y, m);
  if (!eq(&z, &dv[3])) {
    ok = 0;
    fprintf(stderr, "failed!\n");
    fdump(stderr, "x", x->P);
    fdump(stderr, "y", y->P);
    fprintf(stderr, "m = 0x%08lx\n", (unsigned long)m);
    fdump(stderr, "calc z", z.P);
    fgoldi_load(&z, (const octet *)dv[3].buf);
    fdump(stderr, "want z", z.P);
  }

  return (ok);
}

static int vrf_pickn(dstr dv[])
{
  dstr d = DSTR_INIT;
  fgoldi v[32], z;
  size_t i = *(uint32 *)dv[1].buf, j, n;
  const char *p;
  char *q;
  int ok = 1;

  for (q = dv[0].buf, n = 0; (p = str_qword(&q, 0)) != 0; n++)
    { cvt_fgoldi(p, &d); v[n] = *(fgoldi *)d.buf; }

  fgoldi_pickn(&z, v, n, i);
  if (!eq(&z, &dv[2])) {
    ok = 0;
    fprintf(stderr, "failed!\n");
    for (j = 0; j < n; j++) {
      fprintf(stderr, "v[%2u]", (unsigned)j);
      fdump(stderr, "", v[j].P);
    }
    fprintf(stderr, "i = %u\n", (unsigned)i);
    fdump(stderr, "calc z", z.P);
    fgoldi_load(&z, (const octet *)dv[2].buf);
    fdump(stderr, "want z", z.P);
  }

  dstr_destroy(&d);
  return (ok);
}

static int vrf_condswap(dstr dv[])
{
  fgoldi *x = (fgoldi *)dv[0].buf, *y = (fgoldi *)dv[1].buf;
  fgoldi xx = *x, yy = *y;
  uint32 m = *(uint32 *)dv[2].buf;
  int ok = 1;

  fgoldi_condswap(&xx, &yy, m);
  if (!eq(&xx, &dv[3]) || !eq(&yy, &dv[4])) {
    ok = 0;
    fprintf(stderr, "failed!\n");
    fdump(stderr, "x", x->P);
    fdump(stderr, "y", y->P);
    fprintf(stderr, "m = 0x%08lx\n", (unsigned long)m);
    fdump(stderr, "calc xx", xx.P);
    fdump(stderr, "calc yy", yy.P);
    fgoldi_load(&xx, (const octet *)dv[3].buf);
    fgoldi_load(&yy, (const octet *)dv[4].buf);
    fdump(stderr, "want xx", xx.P);
    fdump(stderr, "want yy", yy.P);
  }

  return (ok);
}

static int vrf_quosqrt(dstr dv[])
{
  fgoldi *x = (fgoldi *)dv[0].buf, *y = (fgoldi *)dv[1].buf;
  fgoldi z, zz;
  int rc;
  int ok = 1;

  if (dv[2].len) { fixdstr(&dv[2]); fixdstr(&dv[3]); }
  rc = fgoldi_quosqrt(&z, x, y);
  if (!dv[2].len ? !rc : (rc || (!eq(&z, &dv[2]) && !eq(&z, &dv[3])))) {
    ok = 0;
    fprintf(stderr, "failed!\n");
    fdump(stderr, "x", x->P);
    fdump(stderr, "y", y->P);
    if (rc) fprintf(stderr, "calc: FAIL\n");
    else fdump(stderr, "calc", z.P);
    if (!dv[2].len)
      fprintf(stderr, "exp: FAIL\n");
    else {
      fgoldi_load(&zz, (const octet *)dv[2].buf);
      fdump(stderr, "z", zz.P);
      fgoldi_load(&zz, (const octet *)dv[3].buf);
      fdump(stderr, "z'", zz.P);
    }
  }

  return (ok);
}

static int vrf_sub_mulc_add_sub_mul(dstr dv[])
{
  fgoldi *u = (fgoldi *)dv[0].buf, *v = (fgoldi *)dv[1].buf,
    *w = (fgoldi *)dv[3].buf, *x = (fgoldi *)dv[4].buf,
    *y = (fgoldi *)dv[5].buf;
  long a = *(const long *)dv[2].buf;
  fgoldi umv, aumv, wpaumv, xmy, z, zz;
  int ok = 1;

  fgoldi_sub(&umv, u, v);
  fgoldi_mulconst(&aumv, &umv, a);
  fgoldi_add(&wpaumv, w, &aumv);
  fgoldi_sub(&xmy, x, y);
  fgoldi_mul(&z, &wpaumv, &xmy);

  if (!eq(&z, &dv[6])) {
    ok = 0;
    fprintf(stderr, "failed!\n");
    fdump(stderr, "u", u->P);
    fdump(stderr, "v", v->P);
    fdump(stderr, "u - v", umv.P);
    fprintf(stderr, "a = %ld\n", a);
    fdump(stderr, "a (u - v)", aumv.P);
    fdump(stderr, "w + a (u - v)", wpaumv.P);
    fdump(stderr, "x", x->P);
    fdump(stderr, "y", y->P);
    fdump(stderr, "x - y", xmy.P);
    fdump(stderr, "(x - y) (w + a (u - v))", z.P);
    fgoldi_load(&zz, (const octet *)dv[6].buf); fdump(stderr, "z", zz.P);
  }

  return (ok);
}

static test_chunk tests[] = {
  { "add", vrf_add, { &type_fgoldi, &type_fgoldi, &type_fgoldi_ref } },
  { "sub", vrf_sub, { &type_fgoldi, &type_fgoldi, &type_fgoldi_ref } },
  { "neg", vrf_neg, { &type_fgoldi, &type_fgoldi_ref } },
  { "condneg", vrf_condneg,
    { &type_fgoldi, &type_uint32, &type_fgoldi_ref } },
  { "mul", vrf_mul, { &type_fgoldi, &type_fgoldi, &type_fgoldi_ref } },
  { "mulconst", vrf_mulc, { &type_fgoldi, &type_long, &type_fgoldi_ref } },
  { "pick2", vrf_pick2,
    { &type_fgoldi, &type_fgoldi, &type_uint32, &type_fgoldi_ref } },
  { "pickn", vrf_pickn,
    { &type_string, &type_uint32, &type_fgoldi_ref } },
  { "condswap", vrf_condswap,
    { &type_fgoldi, &type_fgoldi, &type_uint32,
      &type_fgoldi_ref, &type_fgoldi_ref } },
  { "sqr", vrf_sqr, { &type_fgoldi, &type_fgoldi_ref } },
  { "inv", vrf_inv, { &type_fgoldi, &type_fgoldi_ref } },
  { "quosqrt", vrf_quosqrt,
    { &type_fgoldi, &type_fgoldi, &type_hex, &type_hex } },
  { "sub-mulc-add-sub-mul", vrf_sub_mulc_add_sub_mul,
    { &type_fgoldi, &type_fgoldi, &type_long, &type_fgoldi,
      &type_fgoldi, &type_fgoldi, &type_fgoldi_ref } },
  { 0, 0, { 0 } }
};

int main(int argc, char *argv[])
{
  test_run(argc, argv, tests, SRCDIR "/t/fgoldi");
  return (0);
}

#endif

/*----- That's all, folks -------------------------------------------------*/