From 9c3c571c0c51bc11717b795d800b6523ff4ccfd8 Mon Sep 17 00:00:00 2001 From: Nick Mathewson Date: Mon, 3 Dec 2012 14:50:48 -0500 Subject: [PATCH] Add fallback implementations for curve25519: curve25519_donna This is copied from Adam Langley's curve25519-donna package, as of commit 09427c9cab32075c06c3487aa01628030e1c5ae7. --- src/ext/README | 5 +- src/ext/curve25519_donna/README | 44 ++ .../curve25519_donna/curve25519-donna-c64.c | 421 ++++++++++ src/ext/curve25519_donna/curve25519-donna.c | 724 ++++++++++++++++++ 4 files changed, 1193 insertions(+), 1 deletion(-) create mode 100644 src/ext/curve25519_donna/README create mode 100644 src/ext/curve25519_donna/curve25519-donna-c64.c create mode 100644 src/ext/curve25519_donna/curve25519-donna.c diff --git a/src/ext/README b/src/ext/README index 8c850bef66..cd23f29496 100644 --- a/src/ext/README +++ b/src/ext/README @@ -36,4 +36,7 @@ tor_queue.h sys/queue.h, and the ones that do have diverged in incompatible ways. (CIRCLEQ or no CIRCLEQ? SIMPLQ or STAILQ?) - +curve25519_donna/*.c + + A copy of Adam Langley's curve25519-donna mostly-portable + implementations of curve25519. diff --git a/src/ext/curve25519_donna/README b/src/ext/curve25519_donna/README new file mode 100644 index 0000000000..9f77bd7d95 --- /dev/null +++ b/src/ext/curve25519_donna/README @@ -0,0 +1,44 @@ +See http://code.google.com/p/curve25519-donna/ for details. + +BUILDING: + +If you run `make`, two .a archives will be built, similar to djb's curve25519 +code. Alternatively, read on: + +The C implementation is contained within curve25519-donna.c. It has no external +dependancies and is BSD licenced. You can copy/include/link it directly in with +your program. Recommended C flags: -O2 + +The x86-64 bit implementation is contained within curve25519-donna-x86-64.c and +curve25519-donna-x86-64.s. Build like this: + +% cpp curve25519-donna-x86-64.s > curve25519-donna-x86-64.s.pp +% as -o curve25519-donna-x86-64.s.o curve25519-donna-x86-64.s.pp +% gcc -O2 -c curve25519-donna-x86-64.c + +Then the two .o files can be linked in + +USAGE: + +The usage is exactly the same as djb's code (as described at +http://cr.yp.to/ecdh.html) expect that the function is called curve25519_donna. + +In short, + +To generate a private key, generate 32 random bytes and: + + mysecret[0] &= 248; + mysecret[31] &= 127; + mysecret[31] |= 64; + +To generate the public key, just do + + static const uint8_t basepoint[32] = {9}; + curve25519_donna(mypublic, mysecret, basepoint); + +To generate an agreed key do: + uint8_t shared_key[32]; + curve25519_donna(shared_key, mysecret, theirpublic); + +And hash the shared_key with a cryptographic hash function before using. + diff --git a/src/ext/curve25519_donna/curve25519-donna-c64.c b/src/ext/curve25519_donna/curve25519-donna-c64.c new file mode 100644 index 0000000000..4f9dcc05ec --- /dev/null +++ b/src/ext/curve25519_donna/curve25519-donna-c64.c @@ -0,0 +1,421 @@ +/* Copyright 2008, Google Inc. + * All rights reserved. + * + * Code released into the public domain. + * + * curve25519-donna: Curve25519 elliptic curve, public key function + * + * http://code.google.com/p/curve25519-donna/ + * + * Adam Langley + * + * Derived from public domain C code by Daniel J. Bernstein + * + * More information about curve25519 can be found here + * http://cr.yp.to/ecdh.html + * + * djb's sample implementation of curve25519 is written in a special assembly + * language called qhasm and uses the floating point registers. + * + * This is, almost, a clean room reimplementation from the curve25519 paper. It + * uses many of the tricks described therein. Only the crecip function is taken + * from the sample implementation. + */ + +#include +#include + +typedef uint8_t u8; +typedef uint64_t limb; +typedef limb felem[5]; +// This is a special gcc mode for 128-bit integers. It's implemented on 64-bit +// platforms only as far as I know. +typedef unsigned uint128_t __attribute__((mode(TI))); + +#undef force_inline +#define force_inline __attribute__((always_inline)) + +/* Sum two numbers: output += in */ +static void force_inline +fsum(limb *output, const limb *in) { + output[0] += in[0]; + output[1] += in[1]; + output[2] += in[2]; + output[3] += in[3]; + output[4] += in[4]; +} + +/* Find the difference of two numbers: output = in - output + * (note the order of the arguments!) + * + * Assumes that out[i] < 2**52 + * On return, out[i] < 2**55 + */ +static void force_inline +fdifference_backwards(felem out, const felem in) { + /* 152 is 19 << 3 */ + static const limb two54m152 = (((limb)1) << 54) - 152; + static const limb two54m8 = (((limb)1) << 54) - 8; + + out[0] = in[0] + two54m152 - out[0]; + out[1] = in[1] + two54m8 - out[1]; + out[2] = in[2] + two54m8 - out[2]; + out[3] = in[3] + two54m8 - out[3]; + out[4] = in[4] + two54m8 - out[4]; +} + +/* Multiply a number by a scalar: output = in * scalar */ +static void force_inline +fscalar_product(felem output, const felem in, const limb scalar) { + uint128_t a; + + a = ((uint128_t) in[0]) * scalar; + output[0] = ((limb)a) & 0x7ffffffffffff; + + a = ((uint128_t) in[1]) * scalar + ((limb) (a >> 51)); + output[1] = ((limb)a) & 0x7ffffffffffff; + + a = ((uint128_t) in[2]) * scalar + ((limb) (a >> 51)); + output[2] = ((limb)a) & 0x7ffffffffffff; + + a = ((uint128_t) in[3]) * scalar + ((limb) (a >> 51)); + output[3] = ((limb)a) & 0x7ffffffffffff; + + a = ((uint128_t) in[4]) * scalar + ((limb) (a >> 51)); + output[4] = ((limb)a) & 0x7ffffffffffff; + + output[0] += (a >> 51) * 19; +} + +/* Multiply two numbers: output = in2 * in + * + * output must be distinct to both inputs. The inputs are reduced coefficient + * form, the output is not. + * + * Assumes that in[i] < 2**55 and likewise for in2. + * On return, output[i] < 2**52 + */ +static void force_inline +fmul(felem output, const felem in2, const felem in) { + uint128_t t[5]; + limb r0,r1,r2,r3,r4,s0,s1,s2,s3,s4,c; + + r0 = in[0]; + r1 = in[1]; + r2 = in[2]; + r3 = in[3]; + r4 = in[4]; + + s0 = in2[0]; + s1 = in2[1]; + s2 = in2[2]; + s3 = in2[3]; + s4 = in2[4]; + + t[0] = ((uint128_t) r0) * s0; + t[1] = ((uint128_t) r0) * s1 + ((uint128_t) r1) * s0; + t[2] = ((uint128_t) r0) * s2 + ((uint128_t) r2) * s0 + ((uint128_t) r1) * s1; + t[3] = ((uint128_t) r0) * s3 + ((uint128_t) r3) * s0 + ((uint128_t) r1) * s2 + ((uint128_t) r2) * s1; + t[4] = ((uint128_t) r0) * s4 + ((uint128_t) r4) * s0 + ((uint128_t) r3) * s1 + ((uint128_t) r1) * s3 + ((uint128_t) r2) * s2; + + r4 *= 19; + r1 *= 19; + r2 *= 19; + r3 *= 19; + + t[0] += ((uint128_t) r4) * s1 + ((uint128_t) r1) * s4 + ((uint128_t) r2) * s3 + ((uint128_t) r3) * s2; + t[1] += ((uint128_t) r4) * s2 + ((uint128_t) r2) * s4 + ((uint128_t) r3) * s3; + t[2] += ((uint128_t) r4) * s3 + ((uint128_t) r3) * s4; + t[3] += ((uint128_t) r4) * s4; + + r0 = (limb)t[0] & 0x7ffffffffffff; c = (limb)(t[0] >> 51); + t[1] += c; r1 = (limb)t[1] & 0x7ffffffffffff; c = (limb)(t[1] >> 51); + t[2] += c; r2 = (limb)t[2] & 0x7ffffffffffff; c = (limb)(t[2] >> 51); + t[3] += c; r3 = (limb)t[3] & 0x7ffffffffffff; c = (limb)(t[3] >> 51); + t[4] += c; r4 = (limb)t[4] & 0x7ffffffffffff; c = (limb)(t[4] >> 51); + r0 += c * 19; c = r0 >> 51; r0 = r0 & 0x7ffffffffffff; + r1 += c; c = r1 >> 51; r1 = r1 & 0x7ffffffffffff; + r2 += c; + + output[0] = r0; + output[1] = r1; + output[2] = r2; + output[3] = r3; + output[4] = r4; +} + +static void force_inline +fsquare_times(felem output, const felem in, limb count) { + uint128_t t[5]; + limb r0,r1,r2,r3,r4,c; + limb d0,d1,d2,d4,d419; + + r0 = in[0]; + r1 = in[1]; + r2 = in[2]; + r3 = in[3]; + r4 = in[4]; + + do { + d0 = r0 * 2; + d1 = r1 * 2; + d2 = r2 * 2 * 19; + d419 = r4 * 19; + d4 = d419 * 2; + + t[0] = ((uint128_t) r0) * r0 + ((uint128_t) d4) * r1 + (((uint128_t) d2) * (r3 )); + t[1] = ((uint128_t) d0) * r1 + ((uint128_t) d4) * r2 + (((uint128_t) r3) * (r3 * 19)); + t[2] = ((uint128_t) d0) * r2 + ((uint128_t) r1) * r1 + (((uint128_t) d4) * (r3 )); + t[3] = ((uint128_t) d0) * r3 + ((uint128_t) d1) * r2 + (((uint128_t) r4) * (d419 )); + t[4] = ((uint128_t) d0) * r4 + ((uint128_t) d1) * r3 + (((uint128_t) r2) * (r2 )); + + r0 = (limb)t[0] & 0x7ffffffffffff; c = (limb)(t[0] >> 51); + t[1] += c; r1 = (limb)t[1] & 0x7ffffffffffff; c = (limb)(t[1] >> 51); + t[2] += c; r2 = (limb)t[2] & 0x7ffffffffffff; c = (limb)(t[2] >> 51); + t[3] += c; r3 = (limb)t[3] & 0x7ffffffffffff; c = (limb)(t[3] >> 51); + t[4] += c; r4 = (limb)t[4] & 0x7ffffffffffff; c = (limb)(t[4] >> 51); + r0 += c * 19; c = r0 >> 51; r0 = r0 & 0x7ffffffffffff; + r1 += c; c = r1 >> 51; r1 = r1 & 0x7ffffffffffff; + r2 += c; + } while(--count); + + output[0] = r0; + output[1] = r1; + output[2] = r2; + output[3] = r3; + output[4] = r4; +} + +/* Take a little-endian, 32-byte number and expand it into polynomial form */ +static void +fexpand(limb *output, const u8 *in) { + output[0] = *((const uint64_t *)(in)) & 0x7ffffffffffff; + output[1] = (*((const uint64_t *)(in+6)) >> 3) & 0x7ffffffffffff; + output[2] = (*((const uint64_t *)(in+12)) >> 6) & 0x7ffffffffffff; + output[3] = (*((const uint64_t *)(in+19)) >> 1) & 0x7ffffffffffff; + output[4] = (*((const uint64_t *)(in+24)) >> 12) & 0x7ffffffffffff; +} + +/* Take a fully reduced polynomial form number and contract it into a + * little-endian, 32-byte array + */ +static void +fcontract(u8 *output, const felem input) { + uint128_t t[5]; + + t[0] = input[0]; + t[1] = input[1]; + t[2] = input[2]; + t[3] = input[3]; + t[4] = input[4]; + + t[1] += t[0] >> 51; t[0] &= 0x7ffffffffffff; + t[2] += t[1] >> 51; t[1] &= 0x7ffffffffffff; + t[3] += t[2] >> 51; t[2] &= 0x7ffffffffffff; + t[4] += t[3] >> 51; t[3] &= 0x7ffffffffffff; + t[0] += 19 * (t[4] >> 51); t[4] &= 0x7ffffffffffff; + + t[1] += t[0] >> 51; t[0] &= 0x7ffffffffffff; + t[2] += t[1] >> 51; t[1] &= 0x7ffffffffffff; + t[3] += t[2] >> 51; t[2] &= 0x7ffffffffffff; + t[4] += t[3] >> 51; t[3] &= 0x7ffffffffffff; + t[0] += 19 * (t[4] >> 51); t[4] &= 0x7ffffffffffff; + + /* now t is between 0 and 2^255-1, properly carried. */ + /* case 1: between 0 and 2^255-20. case 2: between 2^255-19 and 2^255-1. */ + + t[0] += 19; + + t[1] += t[0] >> 51; t[0] &= 0x7ffffffffffff; + t[2] += t[1] >> 51; t[1] &= 0x7ffffffffffff; + t[3] += t[2] >> 51; t[2] &= 0x7ffffffffffff; + t[4] += t[3] >> 51; t[3] &= 0x7ffffffffffff; + t[0] += 19 * (t[4] >> 51); t[4] &= 0x7ffffffffffff; + + /* now between 19 and 2^255-1 in both cases, and offset by 19. */ + + t[0] += 0x8000000000000 - 19; + t[1] += 0x8000000000000 - 1; + t[2] += 0x8000000000000 - 1; + t[3] += 0x8000000000000 - 1; + t[4] += 0x8000000000000 - 1; + + /* now between 2^255 and 2^256-20, and offset by 2^255. */ + + t[1] += t[0] >> 51; t[0] &= 0x7ffffffffffff; + t[2] += t[1] >> 51; t[1] &= 0x7ffffffffffff; + t[3] += t[2] >> 51; t[2] &= 0x7ffffffffffff; + t[4] += t[3] >> 51; t[3] &= 0x7ffffffffffff; + t[4] &= 0x7ffffffffffff; + + *((uint64_t *)(output)) = t[0] | (t[1] << 51); + *((uint64_t *)(output+8)) = (t[1] >> 13) | (t[2] << 38); + *((uint64_t *)(output+16)) = (t[2] >> 26) | (t[3] << 25); + *((uint64_t *)(output+24)) = (t[3] >> 39) | (t[4] << 12); +} + +/* Input: Q, Q', Q-Q' + * Output: 2Q, Q+Q' + * + * x2 z3: long form + * x3 z3: long form + * x z: short form, destroyed + * xprime zprime: short form, destroyed + * qmqp: short form, preserved + */ +static void +fmonty(limb *x2, limb *z2, /* output 2Q */ + limb *x3, limb *z3, /* output Q + Q' */ + limb *x, limb *z, /* input Q */ + limb *xprime, limb *zprime, /* input Q' */ + const limb *qmqp /* input Q - Q' */) { + limb origx[5], origxprime[5], zzz[5], xx[5], zz[5], xxprime[5], + zzprime[5], zzzprime[5]; + + memcpy(origx, x, 5 * sizeof(limb)); + fsum(x, z); + fdifference_backwards(z, origx); // does x - z + + memcpy(origxprime, xprime, sizeof(limb) * 5); + fsum(xprime, zprime); + fdifference_backwards(zprime, origxprime); + fmul(xxprime, xprime, z); + fmul(zzprime, x, zprime); + memcpy(origxprime, xxprime, sizeof(limb) * 5); + fsum(xxprime, zzprime); + fdifference_backwards(zzprime, origxprime); + fsquare_times(x3, xxprime, 1); + fsquare_times(zzzprime, zzprime, 1); + fmul(z3, zzzprime, qmqp); + + fsquare_times(xx, x, 1); + fsquare_times(zz, z, 1); + fmul(x2, xx, zz); + fdifference_backwards(zz, xx); // does zz = xx - zz + fscalar_product(zzz, zz, 121665); + fsum(zzz, xx); + fmul(z2, zz, zzz); +} + +// ----------------------------------------------------------------------------- +// Maybe swap the contents of two limb arrays (@a and @b), each @len elements +// long. Perform the swap iff @swap is non-zero. +// +// This function performs the swap without leaking any side-channel +// information. +// ----------------------------------------------------------------------------- +static void +swap_conditional(limb a[5], limb b[5], limb iswap) { + unsigned i; + const limb swap = -iswap; + + for (i = 0; i < 5; ++i) { + const limb x = swap & (a[i] ^ b[i]); + a[i] ^= x; + b[i] ^= x; + } +} + +/* Calculates nQ where Q is the x-coordinate of a point on the curve + * + * resultx/resultz: the x coordinate of the resulting curve point (short form) + * n: a little endian, 32-byte number + * q: a point of the curve (short form) + */ +static void +cmult(limb *resultx, limb *resultz, const u8 *n, const limb *q) { + limb a[5] = {0}, b[5] = {1}, c[5] = {1}, d[5] = {0}; + limb *nqpqx = a, *nqpqz = b, *nqx = c, *nqz = d, *t; + limb e[5] = {0}, f[5] = {1}, g[5] = {0}, h[5] = {1}; + limb *nqpqx2 = e, *nqpqz2 = f, *nqx2 = g, *nqz2 = h; + + unsigned i, j; + + memcpy(nqpqx, q, sizeof(limb) * 5); + + for (i = 0; i < 32; ++i) { + u8 byte = n[31 - i]; + for (j = 0; j < 8; ++j) { + const limb bit = byte >> 7; + + swap_conditional(nqx, nqpqx, bit); + swap_conditional(nqz, nqpqz, bit); + fmonty(nqx2, nqz2, + nqpqx2, nqpqz2, + nqx, nqz, + nqpqx, nqpqz, + q); + swap_conditional(nqx2, nqpqx2, bit); + swap_conditional(nqz2, nqpqz2, bit); + + t = nqx; + nqx = nqx2; + nqx2 = t; + t = nqz; + nqz = nqz2; + nqz2 = t; + t = nqpqx; + nqpqx = nqpqx2; + nqpqx2 = t; + t = nqpqz; + nqpqz = nqpqz2; + nqpqz2 = t; + + byte <<= 1; + } + } + + memcpy(resultx, nqx, sizeof(limb) * 5); + memcpy(resultz, nqz, sizeof(limb) * 5); +} + + +// ----------------------------------------------------------------------------- +// Shamelessly copied from djb's code, tightened a little +// ----------------------------------------------------------------------------- +static void +crecip(felem out, const felem z) { + felem a,t0,b,c; + + /* 2 */ fsquare_times(a, z, 1); // a = 2 + /* 8 */ fsquare_times(t0, a, 2); + /* 9 */ fmul(b, t0, z); // b = 9 + /* 11 */ fmul(a, b, a); // a = 11 + /* 22 */ fsquare_times(t0, a, 1); + /* 2^5 - 2^0 = 31 */ fmul(b, t0, b); + /* 2^10 - 2^5 */ fsquare_times(t0, b, 5); + /* 2^10 - 2^0 */ fmul(b, t0, b); + /* 2^20 - 2^10 */ fsquare_times(t0, b, 10); + /* 2^20 - 2^0 */ fmul(c, t0, b); + /* 2^40 - 2^20 */ fsquare_times(t0, c, 20); + /* 2^40 - 2^0 */ fmul(t0, t0, c); + /* 2^50 - 2^10 */ fsquare_times(t0, t0, 10); + /* 2^50 - 2^0 */ fmul(b, t0, b); + /* 2^100 - 2^50 */ fsquare_times(t0, b, 50); + /* 2^100 - 2^0 */ fmul(c, t0, b); + /* 2^200 - 2^100 */ fsquare_times(t0, c, 100); + /* 2^200 - 2^0 */ fmul(t0, t0, c); + /* 2^250 - 2^50 */ fsquare_times(t0, t0, 50); + /* 2^250 - 2^0 */ fmul(t0, t0, b); + /* 2^255 - 2^5 */ fsquare_times(t0, t0, 5); + /* 2^255 - 21 */ fmul(out, t0, a); +} + +int +curve25519_donna(u8 *mypublic, const u8 *secret, const u8 *basepoint) { + limb bp[5], x[5], z[5], zmone[5]; + uint8_t e[32]; + int i; + + for (i = 0;i < 32;++i) e[i] = secret[i]; + e[0] &= 248; + e[31] &= 127; + e[31] |= 64; + + fexpand(bp, basepoint); + cmult(x, z, e, bp); + crecip(zmone, z); + fmul(z, x, zmone); + fcontract(mypublic, z); + return 0; +} diff --git a/src/ext/curve25519_donna/curve25519-donna.c b/src/ext/curve25519_donna/curve25519-donna.c new file mode 100644 index 0000000000..d4b1b1e276 --- /dev/null +++ b/src/ext/curve25519_donna/curve25519-donna.c @@ -0,0 +1,724 @@ +/* Copyright 2008, Google Inc. + * All rights reserved. + * + * Redistribution and use in source and binary forms, with or without + * modification, are permitted provided that the following conditions are + * met: + * + * * Redistributions of source code must retain the above copyright + * notice, this list of conditions and the following disclaimer. + * * Redistributions in binary form must reproduce the above + * copyright notice, this list of conditions and the following disclaimer + * in the documentation and/or other materials provided with the + * distribution. + * * Neither the name of Google Inc. nor the names of its + * contributors may be used to endorse or promote products derived from + * this software without specific prior written permission. + * + * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS + * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT + * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR + * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT + * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, + * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT + * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, + * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY + * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT + * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE + * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. + * + * curve25519-donna: Curve25519 elliptic curve, public key function + * + * http://code.google.com/p/curve25519-donna/ + * + * Adam Langley + * + * Derived from public domain C code by Daniel J. Bernstein + * + * More information about curve25519 can be found here + * http://cr.yp.to/ecdh.html + * + * djb's sample implementation of curve25519 is written in a special assembly + * language called qhasm and uses the floating point registers. + * + * This is, almost, a clean room reimplementation from the curve25519 paper. It + * uses many of the tricks described therein. Only the crecip function is taken + * from the sample implementation. + */ + +#include +#include + +typedef uint8_t u8; +typedef int32_t s32; +typedef int64_t limb; + +/* Field element representation: + * + * Field elements are written as an array of signed, 64-bit limbs, least + * significant first. The value of the field element is: + * x[0] + 2^26·x[1] + x^51·x[2] + 2^102·x[3] + ... + * + * i.e. the limbs are 26, 25, 26, 25, ... bits wide. + */ + +/* Sum two numbers: output += in */ +static void fsum(limb *output, const limb *in) { + unsigned i; + for (i = 0; i < 10; i += 2) { + output[0+i] = (output[0+i] + in[0+i]); + output[1+i] = (output[1+i] + in[1+i]); + } +} + +/* Find the difference of two numbers: output = in - output + * (note the order of the arguments!) + */ +static void fdifference(limb *output, const limb *in) { + unsigned i; + for (i = 0; i < 10; ++i) { + output[i] = (in[i] - output[i]); + } +} + +/* Multiply a number by a scalar: output = in * scalar */ +static void fscalar_product(limb *output, const limb *in, const limb scalar) { + unsigned i; + for (i = 0; i < 10; ++i) { + output[i] = in[i] * scalar; + } +} + +/* Multiply two numbers: output = in2 * in + * + * output must be distinct to both inputs. The inputs are reduced coefficient + * form, the output is not. + */ +static void fproduct(limb *output, const limb *in2, const limb *in) { + output[0] = ((limb) ((s32) in2[0])) * ((s32) in[0]); + output[1] = ((limb) ((s32) in2[0])) * ((s32) in[1]) + + ((limb) ((s32) in2[1])) * ((s32) in[0]); + output[2] = 2 * ((limb) ((s32) in2[1])) * ((s32) in[1]) + + ((limb) ((s32) in2[0])) * ((s32) in[2]) + + ((limb) ((s32) in2[2])) * ((s32) in[0]); + output[3] = ((limb) ((s32) in2[1])) * ((s32) in[2]) + + ((limb) ((s32) in2[2])) * ((s32) in[1]) + + ((limb) ((s32) in2[0])) * ((s32) in[3]) + + ((limb) ((s32) in2[3])) * ((s32) in[0]); + output[4] = ((limb) ((s32) in2[2])) * ((s32) in[2]) + + 2 * (((limb) ((s32) in2[1])) * ((s32) in[3]) + + ((limb) ((s32) in2[3])) * ((s32) in[1])) + + ((limb) ((s32) in2[0])) * ((s32) in[4]) + + ((limb) ((s32) in2[4])) * ((s32) in[0]); + output[5] = ((limb) ((s32) in2[2])) * ((s32) in[3]) + + ((limb) ((s32) in2[3])) * ((s32) in[2]) + + ((limb) ((s32) in2[1])) * ((s32) in[4]) + + ((limb) ((s32) in2[4])) * ((s32) in[1]) + + ((limb) ((s32) in2[0])) * ((s32) in[5]) + + ((limb) ((s32) in2[5])) * ((s32) in[0]); + output[6] = 2 * (((limb) ((s32) in2[3])) * ((s32) in[3]) + + ((limb) ((s32) in2[1])) * ((s32) in[5]) + + ((limb) ((s32) in2[5])) * ((s32) in[1])) + + ((limb) ((s32) in2[2])) * ((s32) in[4]) + + ((limb) ((s32) in2[4])) * ((s32) in[2]) + + ((limb) ((s32) in2[0])) * ((s32) in[6]) + + ((limb) ((s32) in2[6])) * ((s32) in[0]); + output[7] = ((limb) ((s32) in2[3])) * ((s32) in[4]) + + ((limb) ((s32) in2[4])) * ((s32) in[3]) + + ((limb) ((s32) in2[2])) * ((s32) in[5]) + + ((limb) ((s32) in2[5])) * ((s32) in[2]) + + ((limb) ((s32) in2[1])) * ((s32) in[6]) + + ((limb) ((s32) in2[6])) * ((s32) in[1]) + + ((limb) ((s32) in2[0])) * ((s32) in[7]) + + ((limb) ((s32) in2[7])) * ((s32) in[0]); + output[8] = ((limb) ((s32) in2[4])) * ((s32) in[4]) + + 2 * (((limb) ((s32) in2[3])) * ((s32) in[5]) + + ((limb) ((s32) in2[5])) * ((s32) in[3]) + + ((limb) ((s32) in2[1])) * ((s32) in[7]) + + ((limb) ((s32) in2[7])) * ((s32) in[1])) + + ((limb) ((s32) in2[2])) * ((s32) in[6]) + + ((limb) ((s32) in2[6])) * ((s32) in[2]) + + ((limb) ((s32) in2[0])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[0]); + output[9] = ((limb) ((s32) in2[4])) * ((s32) in[5]) + + ((limb) ((s32) in2[5])) * ((s32) in[4]) + + ((limb) ((s32) in2[3])) * ((s32) in[6]) + + ((limb) ((s32) in2[6])) * ((s32) in[3]) + + ((limb) ((s32) in2[2])) * ((s32) in[7]) + + ((limb) ((s32) in2[7])) * ((s32) in[2]) + + ((limb) ((s32) in2[1])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[1]) + + ((limb) ((s32) in2[0])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[0]); + output[10] = 2 * (((limb) ((s32) in2[5])) * ((s32) in[5]) + + ((limb) ((s32) in2[3])) * ((s32) in[7]) + + ((limb) ((s32) in2[7])) * ((s32) in[3]) + + ((limb) ((s32) in2[1])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[1])) + + ((limb) ((s32) in2[4])) * ((s32) in[6]) + + ((limb) ((s32) in2[6])) * ((s32) in[4]) + + ((limb) ((s32) in2[2])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[2]); + output[11] = ((limb) ((s32) in2[5])) * ((s32) in[6]) + + ((limb) ((s32) in2[6])) * ((s32) in[5]) + + ((limb) ((s32) in2[4])) * ((s32) in[7]) + + ((limb) ((s32) in2[7])) * ((s32) in[4]) + + ((limb) ((s32) in2[3])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[3]) + + ((limb) ((s32) in2[2])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[2]); + output[12] = ((limb) ((s32) in2[6])) * ((s32) in[6]) + + 2 * (((limb) ((s32) in2[5])) * ((s32) in[7]) + + ((limb) ((s32) in2[7])) * ((s32) in[5]) + + ((limb) ((s32) in2[3])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[3])) + + ((limb) ((s32) in2[4])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[4]); + output[13] = ((limb) ((s32) in2[6])) * ((s32) in[7]) + + ((limb) ((s32) in2[7])) * ((s32) in[6]) + + ((limb) ((s32) in2[5])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[5]) + + ((limb) ((s32) in2[4])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[4]); + output[14] = 2 * (((limb) ((s32) in2[7])) * ((s32) in[7]) + + ((limb) ((s32) in2[5])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[5])) + + ((limb) ((s32) in2[6])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[6]); + output[15] = ((limb) ((s32) in2[7])) * ((s32) in[8]) + + ((limb) ((s32) in2[8])) * ((s32) in[7]) + + ((limb) ((s32) in2[6])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[6]); + output[16] = ((limb) ((s32) in2[8])) * ((s32) in[8]) + + 2 * (((limb) ((s32) in2[7])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[7])); + output[17] = ((limb) ((s32) in2[8])) * ((s32) in[9]) + + ((limb) ((s32) in2[9])) * ((s32) in[8]); + output[18] = 2 * ((limb) ((s32) in2[9])) * ((s32) in[9]); +} + +/* Reduce a long form to a short form by taking the input mod 2^255 - 19. */ +static void freduce_degree(limb *output) { + /* Each of these shifts and adds ends up multiplying the value by 19. */ + output[8] += output[18] << 4; + output[8] += output[18] << 1; + output[8] += output[18]; + output[7] += output[17] << 4; + output[7] += output[17] << 1; + output[7] += output[17]; + output[6] += output[16] << 4; + output[6] += output[16] << 1; + output[6] += output[16]; + output[5] += output[15] << 4; + output[5] += output[15] << 1; + output[5] += output[15]; + output[4] += output[14] << 4; + output[4] += output[14] << 1; + output[4] += output[14]; + output[3] += output[13] << 4; + output[3] += output[13] << 1; + output[3] += output[13]; + output[2] += output[12] << 4; + output[2] += output[12] << 1; + output[2] += output[12]; + output[1] += output[11] << 4; + output[1] += output[11] << 1; + output[1] += output[11]; + output[0] += output[10] << 4; + output[0] += output[10] << 1; + output[0] += output[10]; +} + +#if (-1 & 3) != 3 +#error "This code only works on a two's complement system" +#endif + +/* return v / 2^26, using only shifts and adds. */ +static inline limb +div_by_2_26(const limb v) +{ + /* High word of v; no shift needed*/ + const uint32_t highword = ((uint64_t) v) >> 32; + /* Set to all 1s if v was negative; else set to 0s. */ + const int32_t sign = ((int32_t) highword) >> 31; + /* Set to 0x3ffffff if v was negative; else set to 0. */ + const int32_t roundoff = ((uint32_t) sign) >> 6; + /* Should return v / (1<<26) */ + return (v + roundoff) >> 26; +} + +/* return v / (2^25), using only shifts and adds. */ +static inline limb +div_by_2_25(const limb v) +{ + /* High word of v; no shift needed*/ + const uint32_t highword = ((uint64_t) v) >> 32; + /* Set to all 1s if v was negative; else set to 0s. */ + const int32_t sign = ((int32_t) highword) >> 31; + /* Set to 0x1ffffff if v was negative; else set to 0. */ + const int32_t roundoff = ((uint32_t) sign) >> 7; + /* Should return v / (1<<25) */ + return (v + roundoff) >> 25; +} + +static inline s32 +div_s32_by_2_25(const s32 v) +{ + const s32 roundoff = ((uint32_t)(v >> 31)) >> 7; + return (v + roundoff) >> 25; +} + +/* Reduce all coefficients of the short form input so that |x| < 2^26. + * + * On entry: |output[i]| < 2^62 + */ +static void freduce_coefficients(limb *output) { + unsigned i; + + output[10] = 0; + + for (i = 0; i < 10; i += 2) { + limb over = div_by_2_26(output[i]); + output[i] -= over << 26; + output[i+1] += over; + + over = div_by_2_25(output[i+1]); + output[i+1] -= over << 25; + output[i+2] += over; + } + /* Now |output[10]| < 2 ^ 38 and all other coefficients are reduced. */ + output[0] += output[10] << 4; + output[0] += output[10] << 1; + output[0] += output[10]; + + output[10] = 0; + + /* Now output[1..9] are reduced, and |output[0]| < 2^26 + 19 * 2^38 + * So |over| will be no more than 77825 */ + { + limb over = div_by_2_26(output[0]); + output[0] -= over << 26; + output[1] += over; + } + + /* Now output[0,2..9] are reduced, and |output[1]| < 2^25 + 77825 + * So |over| will be no more than 1. */ + { + /* output[1] fits in 32 bits, so we can use div_s32_by_2_25 here. */ + s32 over32 = div_s32_by_2_25(output[1]); + output[1] -= over32 << 25; + output[2] += over32; + } + + /* Finally, output[0,1,3..9] are reduced, and output[2] is "nearly reduced": + * we have |output[2]| <= 2^26. This is good enough for all of our math, + * but it will require an extra freduce_coefficients before fcontract. */ +} + +/* A helpful wrapper around fproduct: output = in * in2. + * + * output must be distinct to both inputs. The output is reduced degree and + * reduced coefficient. + */ +static void +fmul(limb *output, const limb *in, const limb *in2) { + limb t[19]; + fproduct(t, in, in2); + freduce_degree(t); + freduce_coefficients(t); + memcpy(output, t, sizeof(limb) * 10); +} + +static void fsquare_inner(limb *output, const limb *in) { + output[0] = ((limb) ((s32) in[0])) * ((s32) in[0]); + output[1] = 2 * ((limb) ((s32) in[0])) * ((s32) in[1]); + output[2] = 2 * (((limb) ((s32) in[1])) * ((s32) in[1]) + + ((limb) ((s32) in[0])) * ((s32) in[2])); + output[3] = 2 * (((limb) ((s32) in[1])) * ((s32) in[2]) + + ((limb) ((s32) in[0])) * ((s32) in[3])); + output[4] = ((limb) ((s32) in[2])) * ((s32) in[2]) + + 4 * ((limb) ((s32) in[1])) * ((s32) in[3]) + + 2 * ((limb) ((s32) in[0])) * ((s32) in[4]); + output[5] = 2 * (((limb) ((s32) in[2])) * ((s32) in[3]) + + ((limb) ((s32) in[1])) * ((s32) in[4]) + + ((limb) ((s32) in[0])) * ((s32) in[5])); + output[6] = 2 * (((limb) ((s32) in[3])) * ((s32) in[3]) + + ((limb) ((s32) in[2])) * ((s32) in[4]) + + ((limb) ((s32) in[0])) * ((s32) in[6]) + + 2 * ((limb) ((s32) in[1])) * ((s32) in[5])); + output[7] = 2 * (((limb) ((s32) in[3])) * ((s32) in[4]) + + ((limb) ((s32) in[2])) * ((s32) in[5]) + + ((limb) ((s32) in[1])) * ((s32) in[6]) + + ((limb) ((s32) in[0])) * ((s32) in[7])); + output[8] = ((limb) ((s32) in[4])) * ((s32) in[4]) + + 2 * (((limb) ((s32) in[2])) * ((s32) in[6]) + + ((limb) ((s32) in[0])) * ((s32) in[8]) + + 2 * (((limb) ((s32) in[1])) * ((s32) in[7]) + + ((limb) ((s32) in[3])) * ((s32) in[5]))); + output[9] = 2 * (((limb) ((s32) in[4])) * ((s32) in[5]) + + ((limb) ((s32) in[3])) * ((s32) in[6]) + + ((limb) ((s32) in[2])) * ((s32) in[7]) + + ((limb) ((s32) in[1])) * ((s32) in[8]) + + ((limb) ((s32) in[0])) * ((s32) in[9])); + output[10] = 2 * (((limb) ((s32) in[5])) * ((s32) in[5]) + + ((limb) ((s32) in[4])) * ((s32) in[6]) + + ((limb) ((s32) in[2])) * ((s32) in[8]) + + 2 * (((limb) ((s32) in[3])) * ((s32) in[7]) + + ((limb) ((s32) in[1])) * ((s32) in[9]))); + output[11] = 2 * (((limb) ((s32) in[5])) * ((s32) in[6]) + + ((limb) ((s32) in[4])) * ((s32) in[7]) + + ((limb) ((s32) in[3])) * ((s32) in[8]) + + ((limb) ((s32) in[2])) * ((s32) in[9])); + output[12] = ((limb) ((s32) in[6])) * ((s32) in[6]) + + 2 * (((limb) ((s32) in[4])) * ((s32) in[8]) + + 2 * (((limb) ((s32) in[5])) * ((s32) in[7]) + + ((limb) ((s32) in[3])) * ((s32) in[9]))); + output[13] = 2 * (((limb) ((s32) in[6])) * ((s32) in[7]) + + ((limb) ((s32) in[5])) * ((s32) in[8]) + + ((limb) ((s32) in[4])) * ((s32) in[9])); + output[14] = 2 * (((limb) ((s32) in[7])) * ((s32) in[7]) + + ((limb) ((s32) in[6])) * ((s32) in[8]) + + 2 * ((limb) ((s32) in[5])) * ((s32) in[9])); + output[15] = 2 * (((limb) ((s32) in[7])) * ((s32) in[8]) + + ((limb) ((s32) in[6])) * ((s32) in[9])); + output[16] = ((limb) ((s32) in[8])) * ((s32) in[8]) + + 4 * ((limb) ((s32) in[7])) * ((s32) in[9]); + output[17] = 2 * ((limb) ((s32) in[8])) * ((s32) in[9]); + output[18] = 2 * ((limb) ((s32) in[9])) * ((s32) in[9]); +} + +static void +fsquare(limb *output, const limb *in) { + limb t[19]; + fsquare_inner(t, in); + freduce_degree(t); + freduce_coefficients(t); + memcpy(output, t, sizeof(limb) * 10); +} + +/* Take a little-endian, 32-byte number and expand it into polynomial form */ +static void +fexpand(limb *output, const u8 *input) { +#define F(n,start,shift,mask) \ + output[n] = ((((limb) input[start + 0]) | \ + ((limb) input[start + 1]) << 8 | \ + ((limb) input[start + 2]) << 16 | \ + ((limb) input[start + 3]) << 24) >> shift) & mask; + F(0, 0, 0, 0x3ffffff); + F(1, 3, 2, 0x1ffffff); + F(2, 6, 3, 0x3ffffff); + F(3, 9, 5, 0x1ffffff); + F(4, 12, 6, 0x3ffffff); + F(5, 16, 0, 0x1ffffff); + F(6, 19, 1, 0x3ffffff); + F(7, 22, 3, 0x1ffffff); + F(8, 25, 4, 0x3ffffff); + F(9, 28, 6, 0x1ffffff); +#undef F +} + +#if (-32 >> 1) != -16 +#error "This code only works when >> does sign-extension on negative numbers" +#endif + +/* Take a fully reduced polynomial form number and contract it into a + * little-endian, 32-byte array + */ +static void +fcontract(u8 *output, limb *input) { + int i; + int j; + + for (j = 0; j < 2; ++j) { + for (i = 0; i < 9; ++i) { + if ((i & 1) == 1) { + /* This calculation is a time-invariant way to make input[i] positive + by borrowing from the next-larger limb. + */ + const s32 mask = (s32)(input[i]) >> 31; + const s32 carry = -(((s32)(input[i]) & mask) >> 25); + input[i] = (s32)(input[i]) + (carry << 25); + input[i+1] = (s32)(input[i+1]) - carry; + } else { + const s32 mask = (s32)(input[i]) >> 31; + const s32 carry = -(((s32)(input[i]) & mask) >> 26); + input[i] = (s32)(input[i]) + (carry << 26); + input[i+1] = (s32)(input[i+1]) - carry; + } + } + const s32 mask = (s32)(input[9]) >> 31; + const s32 carry = -(((s32)(input[9]) & mask) >> 25); + input[9] = (s32)(input[9]) + (carry << 25); + input[0] = (s32)(input[0]) - (carry * 19); + } + + /* The first borrow-propagation pass above ended with every limb + except (possibly) input[0] non-negative. + + Since each input limb except input[0] is decreased by at most 1 + by a borrow-propagation pass, the second borrow-propagation pass + could only have wrapped around to decrease input[0] again if the + first pass left input[0] negative *and* input[1] through input[9] + were all zero. In that case, input[1] is now 2^25 - 1, and this + last borrow-propagation step will leave input[1] non-negative. + */ + const s32 mask = (s32)(input[0]) >> 31; + const s32 carry = -(((s32)(input[0]) & mask) >> 26); + input[0] = (s32)(input[0]) + (carry << 26); + input[1] = (s32)(input[1]) - carry; + + /* Both passes through the above loop, plus the last 0-to-1 step, are + necessary: if input[9] is -1 and input[0] through input[8] are 0, + negative values will remain in the array until the end. + */ + + input[1] <<= 2; + input[2] <<= 3; + input[3] <<= 5; + input[4] <<= 6; + input[6] <<= 1; + input[7] <<= 3; + input[8] <<= 4; + input[9] <<= 6; +#define F(i, s) \ + output[s+0] |= input[i] & 0xff; \ + output[s+1] = (input[i] >> 8) & 0xff; \ + output[s+2] = (input[i] >> 16) & 0xff; \ + output[s+3] = (input[i] >> 24) & 0xff; + output[0] = 0; + output[16] = 0; + F(0,0); + F(1,3); + F(2,6); + F(3,9); + F(4,12); + F(5,16); + F(6,19); + F(7,22); + F(8,25); + F(9,28); +#undef F +} + +/* Input: Q, Q', Q-Q' + * Output: 2Q, Q+Q' + * + * x2 z3: long form + * x3 z3: long form + * x z: short form, destroyed + * xprime zprime: short form, destroyed + * qmqp: short form, preserved + */ +static void fmonty(limb *x2, limb *z2, /* output 2Q */ + limb *x3, limb *z3, /* output Q + Q' */ + limb *x, limb *z, /* input Q */ + limb *xprime, limb *zprime, /* input Q' */ + const limb *qmqp /* input Q - Q' */) { + limb origx[10], origxprime[10], zzz[19], xx[19], zz[19], xxprime[19], + zzprime[19], zzzprime[19], xxxprime[19]; + + memcpy(origx, x, 10 * sizeof(limb)); + fsum(x, z); + fdifference(z, origx); // does x - z + + memcpy(origxprime, xprime, sizeof(limb) * 10); + fsum(xprime, zprime); + fdifference(zprime, origxprime); + fproduct(xxprime, xprime, z); + fproduct(zzprime, x, zprime); + freduce_degree(xxprime); + freduce_coefficients(xxprime); + freduce_degree(zzprime); + freduce_coefficients(zzprime); + memcpy(origxprime, xxprime, sizeof(limb) * 10); + fsum(xxprime, zzprime); + fdifference(zzprime, origxprime); + fsquare(xxxprime, xxprime); + fsquare(zzzprime, zzprime); + fproduct(zzprime, zzzprime, qmqp); + freduce_degree(zzprime); + freduce_coefficients(zzprime); + memcpy(x3, xxxprime, sizeof(limb) * 10); + memcpy(z3, zzprime, sizeof(limb) * 10); + + fsquare(xx, x); + fsquare(zz, z); + fproduct(x2, xx, zz); + freduce_degree(x2); + freduce_coefficients(x2); + fdifference(zz, xx); // does zz = xx - zz + memset(zzz + 10, 0, sizeof(limb) * 9); + fscalar_product(zzz, zz, 121665); + /* No need to call freduce_degree here: + fscalar_product doesn't increase the degree of its input. */ + freduce_coefficients(zzz); + fsum(zzz, xx); + fproduct(z2, zz, zzz); + freduce_degree(z2); + freduce_coefficients(z2); +} + +/* Conditionally swap two reduced-form limb arrays if 'iswap' is 1, but leave + * them unchanged if 'iswap' is 0. Runs in data-invariant time to avoid + * side-channel attacks. + * + * NOTE that this function requires that 'iswap' be 1 or 0; other values give + * wrong results. Also, the two limb arrays must be in reduced-coefficient, + * reduced-degree form: the values in a[10..19] or b[10..19] aren't swapped, + * and all all values in a[0..9],b[0..9] must have magnitude less than + * INT32_MAX. + */ +static void +swap_conditional(limb a[19], limb b[19], limb iswap) { + unsigned i; + const s32 swap = -iswap; + + for (i = 0; i < 10; ++i) { + const s32 x = swap & ( ((s32)a[i]) ^ ((s32)b[i]) ); + a[i] = ((s32)a[i]) ^ x; + b[i] = ((s32)b[i]) ^ x; + } +} + +/* Calculates nQ where Q is the x-coordinate of a point on the curve + * + * resultx/resultz: the x coordinate of the resulting curve point (short form) + * n: a little endian, 32-byte number + * q: a point of the curve (short form) + */ +static void +cmult(limb *resultx, limb *resultz, const u8 *n, const limb *q) { + limb a[19] = {0}, b[19] = {1}, c[19] = {1}, d[19] = {0}; + limb *nqpqx = a, *nqpqz = b, *nqx = c, *nqz = d, *t; + limb e[19] = {0}, f[19] = {1}, g[19] = {0}, h[19] = {1}; + limb *nqpqx2 = e, *nqpqz2 = f, *nqx2 = g, *nqz2 = h; + + unsigned i, j; + + memcpy(nqpqx, q, sizeof(limb) * 10); + + for (i = 0; i < 32; ++i) { + u8 byte = n[31 - i]; + for (j = 0; j < 8; ++j) { + const limb bit = byte >> 7; + + swap_conditional(nqx, nqpqx, bit); + swap_conditional(nqz, nqpqz, bit); + fmonty(nqx2, nqz2, + nqpqx2, nqpqz2, + nqx, nqz, + nqpqx, nqpqz, + q); + swap_conditional(nqx2, nqpqx2, bit); + swap_conditional(nqz2, nqpqz2, bit); + + t = nqx; + nqx = nqx2; + nqx2 = t; + t = nqz; + nqz = nqz2; + nqz2 = t; + t = nqpqx; + nqpqx = nqpqx2; + nqpqx2 = t; + t = nqpqz; + nqpqz = nqpqz2; + nqpqz2 = t; + + byte <<= 1; + } + } + + memcpy(resultx, nqx, sizeof(limb) * 10); + memcpy(resultz, nqz, sizeof(limb) * 10); +} + +// ----------------------------------------------------------------------------- +// Shamelessly copied from djb's code +// ----------------------------------------------------------------------------- +static void +crecip(limb *out, const limb *z) { + limb z2[10]; + limb z9[10]; + limb z11[10]; + limb z2_5_0[10]; + limb z2_10_0[10]; + limb z2_20_0[10]; + limb z2_50_0[10]; + limb z2_100_0[10]; + limb t0[10]; + limb t1[10]; + int i; + + /* 2 */ fsquare(z2,z); + /* 4 */ fsquare(t1,z2); + /* 8 */ fsquare(t0,t1); + /* 9 */ fmul(z9,t0,z); + /* 11 */ fmul(z11,z9,z2); + /* 22 */ fsquare(t0,z11); + /* 2^5 - 2^0 = 31 */ fmul(z2_5_0,t0,z9); + + /* 2^6 - 2^1 */ fsquare(t0,z2_5_0); + /* 2^7 - 2^2 */ fsquare(t1,t0); + /* 2^8 - 2^3 */ fsquare(t0,t1); + /* 2^9 - 2^4 */ fsquare(t1,t0); + /* 2^10 - 2^5 */ fsquare(t0,t1); + /* 2^10 - 2^0 */ fmul(z2_10_0,t0,z2_5_0); + + /* 2^11 - 2^1 */ fsquare(t0,z2_10_0); + /* 2^12 - 2^2 */ fsquare(t1,t0); + /* 2^20 - 2^10 */ for (i = 2;i < 10;i += 2) { fsquare(t0,t1); fsquare(t1,t0); } + /* 2^20 - 2^0 */ fmul(z2_20_0,t1,z2_10_0); + + /* 2^21 - 2^1 */ fsquare(t0,z2_20_0); + /* 2^22 - 2^2 */ fsquare(t1,t0); + /* 2^40 - 2^20 */ for (i = 2;i < 20;i += 2) { fsquare(t0,t1); fsquare(t1,t0); } + /* 2^40 - 2^0 */ fmul(t0,t1,z2_20_0); + + /* 2^41 - 2^1 */ fsquare(t1,t0); + /* 2^42 - 2^2 */ fsquare(t0,t1); + /* 2^50 - 2^10 */ for (i = 2;i < 10;i += 2) { fsquare(t1,t0); fsquare(t0,t1); } + /* 2^50 - 2^0 */ fmul(z2_50_0,t0,z2_10_0); + + /* 2^51 - 2^1 */ fsquare(t0,z2_50_0); + /* 2^52 - 2^2 */ fsquare(t1,t0); + /* 2^100 - 2^50 */ for (i = 2;i < 50;i += 2) { fsquare(t0,t1); fsquare(t1,t0); } + /* 2^100 - 2^0 */ fmul(z2_100_0,t1,z2_50_0); + + /* 2^101 - 2^1 */ fsquare(t1,z2_100_0); + /* 2^102 - 2^2 */ fsquare(t0,t1); + /* 2^200 - 2^100 */ for (i = 2;i < 100;i += 2) { fsquare(t1,t0); fsquare(t0,t1); } + /* 2^200 - 2^0 */ fmul(t1,t0,z2_100_0); + + /* 2^201 - 2^1 */ fsquare(t0,t1); + /* 2^202 - 2^2 */ fsquare(t1,t0); + /* 2^250 - 2^50 */ for (i = 2;i < 50;i += 2) { fsquare(t0,t1); fsquare(t1,t0); } + /* 2^250 - 2^0 */ fmul(t0,t1,z2_50_0); + + /* 2^251 - 2^1 */ fsquare(t1,t0); + /* 2^252 - 2^2 */ fsquare(t0,t1); + /* 2^253 - 2^3 */ fsquare(t1,t0); + /* 2^254 - 2^4 */ fsquare(t0,t1); + /* 2^255 - 2^5 */ fsquare(t1,t0); + /* 2^255 - 21 */ fmul(out,t1,z11); +} + +int +curve25519_donna(u8 *mypublic, const u8 *secret, const u8 *basepoint) { + limb bp[10], x[10], z[11], zmone[10]; + uint8_t e[32]; + int i; + + for (i = 0; i < 32; ++i) e[i] = secret[i]; + e[0] &= 248; + e[31] &= 127; + e[31] |= 64; + + fexpand(bp, basepoint); + cmult(x, z, e, bp); + crecip(zmone, z); + fmul(z, x, zmone); + freduce_coefficients(z); + fcontract(mypublic, z); + return 0; +}