chore: checkpoint before Python removal

vendor/aws-lc-sys/aws-lc/crypto/poly1305/internal.h

@@ -0,0 +1,37 @@
+// Copyright (c) 2016, Google Inc.
+// SPDX-License-Identifier: ISC
+
+#ifndef OPENSSL_HEADER_POLY1305_INTERNAL_H
+#define OPENSSL_HEADER_POLY1305_INTERNAL_H
+
+#include <openssl/base.h>
+#include <openssl/poly1305.h>
+
+#include "../internal.h"
+
+#if defined(__cplusplus)
+extern "C" {
+#endif
+
+static inline struct poly1305_state_st *poly1305_aligned_state(
+    poly1305_state *state) {
+  return align_pointer(state, 64);
+}
+
+#if defined(OPENSSL_ARM) && !defined(OPENSSL_NO_ASM) && !defined(OPENSSL_APPLE)
+#define OPENSSL_POLY1305_NEON
+
+void CRYPTO_poly1305_init_neon(poly1305_state *state, const uint8_t key[32]);
+
+void CRYPTO_poly1305_update_neon(poly1305_state *state, const uint8_t *in,
+                                 size_t in_len);
+
+void CRYPTO_poly1305_finish_neon(poly1305_state *state, uint8_t mac[16]);
+#endif
+
+
+#if defined(__cplusplus)
+}  // extern C
+#endif
+
+#endif  // OPENSSL_HEADER_POLY1305_INTERNAL_H

vendor/aws-lc-sys/aws-lc/crypto/poly1305/poly1305.c

@@ -0,0 +1,299 @@
+// Copyright (c) 2014, Google Inc.
+// SPDX-License-Identifier: ISC
+
+// This implementation of poly1305 is by Andrew Moon
+// (https://github.com/floodyberry/poly1305-donna) and released as public
+// domain.
+
+#include <openssl/poly1305.h>
+
+#include <string.h>
+
+#include "internal.h"
+#include "../internal.h"
+#include "../fipsmodule/cpucap/internal.h"
+
+
+#if !defined(BORINGSSL_HAS_UINT128) || !defined(OPENSSL_X86_64)
+
+static uint64_t mul32x32_64(uint32_t a, uint32_t b) { return (uint64_t)a * b; }
+
+struct poly1305_state_st {
+  uint32_t r0, r1, r2, r3, r4;
+  uint32_t s1, s2, s3, s4;
+  uint32_t h0, h1, h2, h3, h4;
+  uint8_t buf[16];
+  size_t buf_used;
+  uint8_t key[16];
+};
+
+OPENSSL_STATIC_ASSERT(
+    sizeof(struct poly1305_state_st) + 63 <= sizeof(poly1305_state),
+    _poly1305_state_isn_t_large_enough_to_hold_aligned_poly1305_state_st)
+
+
+// poly1305_blocks updates |state| given some amount of input data. This
+// function may only be called with a |len| that is not a multiple of 16 at the
+// end of the data. Otherwise the input must be buffered into 16 byte blocks.
+static void poly1305_update(struct poly1305_state_st *state, const uint8_t *in,
+                            size_t len) {
+  uint32_t t0, t1, t2, t3;
+  uint64_t t[5];
+  uint32_t b;
+  uint64_t c;
+  size_t j;
+  uint8_t mp[16];
+
+  if (len < 16) {
+    goto poly1305_donna_atmost15bytes;
+  }
+
+poly1305_donna_16bytes:
+  t0 = CRYPTO_load_u32_le(in);
+  t1 = CRYPTO_load_u32_le(in + 4);
+  t2 = CRYPTO_load_u32_le(in + 8);
+  t3 = CRYPTO_load_u32_le(in + 12);
+
+  in += 16;
+  len -= 16;
+
+  state->h0 += t0 & 0x3ffffff;
+  state->h1 += ((((uint64_t)t1 << 32) | t0) >> 26) & 0x3ffffff;
+  state->h2 += ((((uint64_t)t2 << 32) | t1) >> 20) & 0x3ffffff;
+  state->h3 += ((((uint64_t)t3 << 32) | t2) >> 14) & 0x3ffffff;
+  state->h4 += (t3 >> 8) | (1 << 24);
+
+poly1305_donna_mul:
+  t[0] = mul32x32_64(state->h0, state->r0) + mul32x32_64(state->h1, state->s4) +
+         mul32x32_64(state->h2, state->s3) + mul32x32_64(state->h3, state->s2) +
+         mul32x32_64(state->h4, state->s1);
+  t[1] = mul32x32_64(state->h0, state->r1) + mul32x32_64(state->h1, state->r0) +
+         mul32x32_64(state->h2, state->s4) + mul32x32_64(state->h3, state->s3) +
+         mul32x32_64(state->h4, state->s2);
+  t[2] = mul32x32_64(state->h0, state->r2) + mul32x32_64(state->h1, state->r1) +
+         mul32x32_64(state->h2, state->r0) + mul32x32_64(state->h3, state->s4) +
+         mul32x32_64(state->h4, state->s3);
+  t[3] = mul32x32_64(state->h0, state->r3) + mul32x32_64(state->h1, state->r2) +
+         mul32x32_64(state->h2, state->r1) + mul32x32_64(state->h3, state->r0) +
+         mul32x32_64(state->h4, state->s4);
+  t[4] = mul32x32_64(state->h0, state->r4) + mul32x32_64(state->h1, state->r3) +
+         mul32x32_64(state->h2, state->r2) + mul32x32_64(state->h3, state->r1) +
+         mul32x32_64(state->h4, state->r0);
+
+  state->h0 = (uint32_t)t[0] & 0x3ffffff;
+  c = (t[0] >> 26);
+  t[1] += c;
+  state->h1 = (uint32_t)t[1] & 0x3ffffff;
+  b = (uint32_t)(t[1] >> 26);
+  t[2] += b;
+  state->h2 = (uint32_t)t[2] & 0x3ffffff;
+  b = (uint32_t)(t[2] >> 26);
+  t[3] += b;
+  state->h3 = (uint32_t)t[3] & 0x3ffffff;
+  b = (uint32_t)(t[3] >> 26);
+  t[4] += b;
+  state->h4 = (uint32_t)t[4] & 0x3ffffff;
+  b = (uint32_t)(t[4] >> 26);
+  state->h0 += b * 5;
+
+  if (len >= 16) {
+    goto poly1305_donna_16bytes;
+  }
+
+// final bytes
+poly1305_donna_atmost15bytes:
+  if (!len) {
+    return;
+  }
+
+  for (j = 0; j < len; j++) {
+    mp[j] = in[j];
+  }
+  mp[j++] = 1;
+  for (; j < 16; j++) {
+    mp[j] = 0;
+  }
+  len = 0;
+
+  t0 = CRYPTO_load_u32_le(mp + 0);
+  t1 = CRYPTO_load_u32_le(mp + 4);
+  t2 = CRYPTO_load_u32_le(mp + 8);
+  t3 = CRYPTO_load_u32_le(mp + 12);
+
+  state->h0 += t0 & 0x3ffffff;
+  state->h1 += ((((uint64_t)t1 << 32) | t0) >> 26) & 0x3ffffff;
+  state->h2 += ((((uint64_t)t2 << 32) | t1) >> 20) & 0x3ffffff;
+  state->h3 += ((((uint64_t)t3 << 32) | t2) >> 14) & 0x3ffffff;
+  state->h4 += (t3 >> 8);
+
+  goto poly1305_donna_mul;
+}
+
+void CRYPTO_poly1305_init(poly1305_state *statep, const uint8_t key[32]) {
+  struct poly1305_state_st *state = poly1305_aligned_state(statep);
+  uint32_t t0, t1, t2, t3;
+
+#if defined(OPENSSL_POLY1305_NEON)
+  if (CRYPTO_is_NEON_capable()) {
+    CRYPTO_poly1305_init_neon(statep, key);
+    return;
+  }
+#endif
+
+  t0 = CRYPTO_load_u32_le(key + 0);
+  t1 = CRYPTO_load_u32_le(key + 4);
+  t2 = CRYPTO_load_u32_le(key + 8);
+  t3 = CRYPTO_load_u32_le(key + 12);
+
+  // precompute multipliers
+  state->r0 = t0 & 0x3ffffff;
+  t0 >>= 26;
+  t0 |= t1 << 6;
+  state->r1 = t0 & 0x3ffff03;
+  t1 >>= 20;
+  t1 |= t2 << 12;
+  state->r2 = t1 & 0x3ffc0ff;
+  t2 >>= 14;
+  t2 |= t3 << 18;
+  state->r3 = t2 & 0x3f03fff;
+  t3 >>= 8;
+  state->r4 = t3 & 0x00fffff;
+
+  state->s1 = state->r1 * 5;
+  state->s2 = state->r2 * 5;
+  state->s3 = state->r3 * 5;
+  state->s4 = state->r4 * 5;
+
+  // init state
+  state->h0 = 0;
+  state->h1 = 0;
+  state->h2 = 0;
+  state->h3 = 0;
+  state->h4 = 0;
+
+  state->buf_used = 0;
+  OPENSSL_memcpy(state->key, key + 16, sizeof(state->key));
+}
+
+void CRYPTO_poly1305_update(poly1305_state *statep, const uint8_t *in,
+                            size_t in_len) {
+  struct poly1305_state_st *state = poly1305_aligned_state(statep);
+
+  // Work around a C language bug. See https://crbug.com/1019588.
+  if (in_len == 0) {
+    return;
+  }
+
+#if defined(OPENSSL_POLY1305_NEON)
+  if (CRYPTO_is_NEON_capable()) {
+    CRYPTO_poly1305_update_neon(statep, in, in_len);
+    return;
+  }
+#endif
+
+  if (state->buf_used) {
+    size_t todo = 16 - state->buf_used;
+    if (todo > in_len) {
+      todo = in_len;
+    }
+    for (size_t i = 0; i < todo; i++) {
+      state->buf[state->buf_used + i] = in[i];
+    }
+    state->buf_used += todo;
+    in_len -= todo;
+    in += todo;
+
+    if (state->buf_used == 16) {
+      poly1305_update(state, state->buf, 16);
+      state->buf_used = 0;
+    }
+  }
+
+  if (in_len >= 16) {
+    size_t todo = in_len & ~0xf;
+    poly1305_update(state, in, todo);
+    in += todo;
+    in_len &= 0xf;
+  }
+
+  if (in_len) {
+    for (size_t i = 0; i < in_len; i++) {
+      state->buf[i] = in[i];
+    }
+    state->buf_used = in_len;
+  }
+}
+
+void CRYPTO_poly1305_finish(poly1305_state *statep, uint8_t mac[16]) {
+  struct poly1305_state_st *state = poly1305_aligned_state(statep);
+  uint32_t g0, g1, g2, g3, g4;
+  uint32_t b, nb;
+
+#if defined(OPENSSL_POLY1305_NEON)
+  if (CRYPTO_is_NEON_capable()) {
+    CRYPTO_poly1305_finish_neon(statep, mac);
+    return;
+  }
+#endif
+
+  if (state->buf_used) {
+    poly1305_update(state, state->buf, state->buf_used);
+  }
+
+  b = state->h0 >> 26;
+  state->h0 = state->h0 & 0x3ffffff;
+  state->h1 += b;
+  b = state->h1 >> 26;
+  state->h1 = state->h1 & 0x3ffffff;
+  state->h2 += b;
+  b = state->h2 >> 26;
+  state->h2 = state->h2 & 0x3ffffff;
+  state->h3 += b;
+  b = state->h3 >> 26;
+  state->h3 = state->h3 & 0x3ffffff;
+  state->h4 += b;
+  b = state->h4 >> 26;
+  state->h4 = state->h4 & 0x3ffffff;
+  state->h0 += b * 5;
+
+  g0 = state->h0 + 5;
+  b = g0 >> 26;
+  g0 &= 0x3ffffff;
+  g1 = state->h1 + b;
+  b = g1 >> 26;
+  g1 &= 0x3ffffff;
+  g2 = state->h2 + b;
+  b = g2 >> 26;
+  g2 &= 0x3ffffff;
+  g3 = state->h3 + b;
+  b = g3 >> 26;
+  g3 &= 0x3ffffff;
+  g4 = state->h4 + b - (1 << 26);
+
+  b = (g4 >> 31) - 1;
+  nb = ~b;
+  state->h0 = (state->h0 & nb) | (g0 & b);
+  state->h1 = (state->h1 & nb) | (g1 & b);
+  state->h2 = (state->h2 & nb) | (g2 & b);
+  state->h3 = (state->h3 & nb) | (g3 & b);
+  state->h4 = (state->h4 & nb) | (g4 & b);
+
+  uint64_t f0 = ((state->h0) | (state->h1 << 26)) +
+                (uint64_t)CRYPTO_load_u32_le(&state->key[0]);
+  uint64_t f1 = ((state->h1 >> 6) | (state->h2 << 20)) +
+                (uint64_t)CRYPTO_load_u32_le(&state->key[4]);
+  uint64_t f2 = ((state->h2 >> 12) | (state->h3 << 14)) +
+                (uint64_t)CRYPTO_load_u32_le(&state->key[8]);
+  uint64_t f3 = ((state->h3 >> 18) | (state->h4 << 8)) +
+                (uint64_t)CRYPTO_load_u32_le(&state->key[12]);
+
+  CRYPTO_store_u32_le(&mac[0], (uint32_t)f0);
+  f1 += (f0 >> 32);
+  CRYPTO_store_u32_le(&mac[4], (uint32_t)f1);
+  f2 += (f1 >> 32);
+  CRYPTO_store_u32_le(&mac[8], (uint32_t)f2);
+  f3 += (f2 >> 32);
+  CRYPTO_store_u32_le(&mac[12], (uint32_t)f3);
+}
+
+#endif  // !BORINGSSL_HAS_UINT128 || !OPENSSL_X86_64

vendor/aws-lc-sys/aws-lc/crypto/poly1305/poly1305_arm.c

@@ -0,0 +1,296 @@
+// Copyright (c) 2014, Google Inc.
+// SPDX-License-Identifier: ISC
+
+// This implementation was taken from the public domain, neon2 version in
+// SUPERCOP by D. J. Bernstein and Peter Schwabe.
+
+#include <openssl/poly1305.h>
+
+#include <string.h>
+
+#include "../internal.h"
+#include "internal.h"
+
+
+#if defined(OPENSSL_POLY1305_NEON)
+
+typedef struct {
+  uint32_t v[12];  // for alignment; only using 10
+} fe1305x2;
+
+#define addmulmod openssl_poly1305_neon2_addmulmod
+#define blocks openssl_poly1305_neon2_blocks
+
+extern void addmulmod(fe1305x2 *r, const fe1305x2 *x, const fe1305x2 *y,
+                      const fe1305x2 *c);
+
+extern int blocks(fe1305x2 *h, const fe1305x2 *precomp, const uint8_t *in,
+                  size_t inlen);
+
+static void freeze(fe1305x2 *r) {
+  int i;
+
+  uint32_t x0 = r->v[0];
+  uint32_t x1 = r->v[2];
+  uint32_t x2 = r->v[4];
+  uint32_t x3 = r->v[6];
+  uint32_t x4 = r->v[8];
+  uint32_t y0;
+  uint32_t y1;
+  uint32_t y2;
+  uint32_t y3;
+  uint32_t y4;
+  uint32_t swap;
+
+  for (i = 0; i < 3; ++i) {
+    x1 += x0 >> 26;
+    x0 &= 0x3ffffff;
+    x2 += x1 >> 26;
+    x1 &= 0x3ffffff;
+    x3 += x2 >> 26;
+    x2 &= 0x3ffffff;
+    x4 += x3 >> 26;
+    x3 &= 0x3ffffff;
+    x0 += 5 * (x4 >> 26);
+    x4 &= 0x3ffffff;
+  }
+
+  y0 = x0 + 5;
+  y1 = x1 + (y0 >> 26);
+  y0 &= 0x3ffffff;
+  y2 = x2 + (y1 >> 26);
+  y1 &= 0x3ffffff;
+  y3 = x3 + (y2 >> 26);
+  y2 &= 0x3ffffff;
+  y4 = x4 + (y3 >> 26);
+  y3 &= 0x3ffffff;
+  swap = -(y4 >> 26);
+  y4 &= 0x3ffffff;
+
+  y0 ^= x0;
+  y1 ^= x1;
+  y2 ^= x2;
+  y3 ^= x3;
+  y4 ^= x4;
+
+  y0 &= swap;
+  y1 &= swap;
+  y2 &= swap;
+  y3 &= swap;
+  y4 &= swap;
+
+  y0 ^= x0;
+  y1 ^= x1;
+  y2 ^= x2;
+  y3 ^= x3;
+  y4 ^= x4;
+
+  r->v[0] = y0;
+  r->v[2] = y1;
+  r->v[4] = y2;
+  r->v[6] = y3;
+  r->v[8] = y4;
+}
+
+static void store32(uint8_t out[4], uint32_t v) { OPENSSL_memcpy(out, &v, 4); }
+
+// load32 exists to avoid breaking strict aliasing rules in
+// fe1305x2_frombytearray.
+static uint32_t load32(const uint8_t t[4]) {
+  uint32_t tmp;
+  OPENSSL_memcpy(&tmp, t, sizeof(tmp));
+  return tmp;
+}
+
+static void fe1305x2_tobytearray(uint8_t r[16], fe1305x2 *x) {
+  uint32_t x0 = x->v[0];
+  uint32_t x1 = x->v[2];
+  uint32_t x2 = x->v[4];
+  uint32_t x3 = x->v[6];
+  uint32_t x4 = x->v[8];
+
+  x1 += x0 >> 26;
+  x0 &= 0x3ffffff;
+  x2 += x1 >> 26;
+  x1 &= 0x3ffffff;
+  x3 += x2 >> 26;
+  x2 &= 0x3ffffff;
+  x4 += x3 >> 26;
+  x3 &= 0x3ffffff;
+
+  store32(r, x0 + (x1 << 26));
+  store32(r + 4, (x1 >> 6) + (x2 << 20));
+  store32(r + 8, (x2 >> 12) + (x3 << 14));
+  store32(r + 12, (x3 >> 18) + (x4 << 8));
+}
+
+static void fe1305x2_frombytearray(fe1305x2 *r, const uint8_t *x, size_t xlen) {
+  size_t i;
+  uint8_t t[17];
+
+  for (i = 0; (i < 16) && (i < xlen); i++) {
+    t[i] = x[i];
+  }
+  xlen -= i;
+  x += i;
+  t[i++] = 1;
+  for (; i < 17; i++) {
+    t[i] = 0;
+  }
+
+  r->v[0] = 0x3ffffff & load32(t);
+  r->v[2] = 0x3ffffff & (load32(t + 3) >> 2);
+  r->v[4] = 0x3ffffff & (load32(t + 6) >> 4);
+  r->v[6] = 0x3ffffff & (load32(t + 9) >> 6);
+  r->v[8] = load32(t + 13);
+
+  if (xlen) {
+    for (i = 0; (i < 16) && (i < xlen); i++) {
+      t[i] = x[i];
+    }
+    t[i++] = 1;
+    for (; i < 17; i++) {
+      t[i] = 0;
+    }
+
+    r->v[1] = 0x3ffffff & load32(t);
+    r->v[3] = 0x3ffffff & (load32(t + 3) >> 2);
+    r->v[5] = 0x3ffffff & (load32(t + 6) >> 4);
+    r->v[7] = 0x3ffffff & (load32(t + 9) >> 6);
+    r->v[9] = load32(t + 13);
+  } else {
+    r->v[1] = r->v[3] = r->v[5] = r->v[7] = r->v[9] = 0;
+  }
+}
+
+static const alignas(16) fe1305x2 zero;
+
+struct poly1305_state_st {
+  uint8_t data[sizeof(fe1305x2[5]) + 128];
+  uint8_t buf[32];
+  size_t buf_used;
+  uint8_t key[16];
+};
+
+OPENSSL_STATIC_ASSERT(
+    sizeof(struct poly1305_state_st) + 63 <= sizeof(poly1305_state),
+    _poly1305_state_isn_t_large_enough_to_hold_aligned_poly1305_state_st)
+
+void CRYPTO_poly1305_init_neon(poly1305_state *state, const uint8_t key[32]) {
+  struct poly1305_state_st *st = poly1305_aligned_state(state);
+  fe1305x2 *const r = (fe1305x2 *)(st->data + (15 & (-(int)st->data)));
+  fe1305x2 *const h = r + 1;
+  fe1305x2 *const c = h + 1;
+  fe1305x2 *const precomp = c + 1;
+
+  r->v[1] = r->v[0] = 0x3ffffff & load32(key);
+  r->v[3] = r->v[2] = 0x3ffff03 & (load32(key + 3) >> 2);
+  r->v[5] = r->v[4] = 0x3ffc0ff & (load32(key + 6) >> 4);
+  r->v[7] = r->v[6] = 0x3f03fff & (load32(key + 9) >> 6);
+  r->v[9] = r->v[8] = 0x00fffff & (load32(key + 12) >> 8);
+
+  for (size_t j = 0; j < 10; j++) {
+    h->v[j] = 0;  // XXX: should fast-forward a bit
+  }
+
+  addmulmod(precomp, r, r, &zero);                  // precompute r^2
+  addmulmod(precomp + 1, precomp, precomp, &zero);  // precompute r^4
+
+  OPENSSL_memcpy(st->key, key + 16, 16);
+  st->buf_used = 0;
+}
+
+void CRYPTO_poly1305_update_neon(poly1305_state *state, const uint8_t *in,
+                                 size_t in_len) {
+  struct poly1305_state_st *st = poly1305_aligned_state(state);
+  fe1305x2 *const r = (fe1305x2 *)(st->data + (15 & (-(int)st->data)));
+  fe1305x2 *const h = r + 1;
+  fe1305x2 *const c = h + 1;
+  fe1305x2 *const precomp = c + 1;
+
+  if (st->buf_used) {
+    size_t todo = 32 - st->buf_used;
+    if (todo > in_len) {
+      todo = in_len;
+    }
+    for (size_t i = 0; i < todo; i++) {
+      st->buf[st->buf_used + i] = in[i];
+    }
+    st->buf_used += todo;
+    in_len -= todo;
+    in += todo;
+
+    if (st->buf_used == sizeof(st->buf) && in_len) {
+      addmulmod(h, h, precomp, &zero);
+      fe1305x2_frombytearray(c, st->buf, sizeof(st->buf));
+      for (size_t i = 0; i < 10; i++) {
+        h->v[i] += c->v[i];
+      }
+      st->buf_used = 0;
+    }
+  }
+
+  while (in_len > 32) {
+    size_t tlen = 1048576;
+    if (in_len < tlen) {
+      tlen = in_len;
+    }
+    tlen -= blocks(h, precomp, in, tlen);
+    in_len -= tlen;
+    in += tlen;
+  }
+
+  if (in_len) {
+    for (size_t i = 0; i < in_len; i++) {
+      st->buf[i] = in[i];
+    }
+    st->buf_used = in_len;
+  }
+}
+
+void CRYPTO_poly1305_finish_neon(poly1305_state *state, uint8_t mac[16]) {
+  struct poly1305_state_st *st = poly1305_aligned_state(state);
+  fe1305x2 *const r = (fe1305x2 *)(st->data + (15 & (-(int)st->data)));
+  fe1305x2 *const h = r + 1;
+  fe1305x2 *const c = h + 1;
+  fe1305x2 *const precomp = c + 1;
+
+  addmulmod(h, h, precomp, &zero);
+
+  if (st->buf_used > 16) {
+    fe1305x2_frombytearray(c, st->buf, st->buf_used);
+    precomp->v[1] = r->v[1];
+    precomp->v[3] = r->v[3];
+    precomp->v[5] = r->v[5];
+    precomp->v[7] = r->v[7];
+    precomp->v[9] = r->v[9];
+    addmulmod(h, h, precomp, c);
+  } else if (st->buf_used > 0) {
+    fe1305x2_frombytearray(c, st->buf, st->buf_used);
+    r->v[1] = 1;
+    r->v[3] = 0;
+    r->v[5] = 0;
+    r->v[7] = 0;
+    r->v[9] = 0;
+    addmulmod(h, h, r, c);
+  }
+
+  h->v[0] += h->v[1];
+  h->v[2] += h->v[3];
+  h->v[4] += h->v[5];
+  h->v[6] += h->v[7];
+  h->v[8] += h->v[9];
+  freeze(h);
+
+  fe1305x2_frombytearray(c, st->key, 16);
+  c->v[8] ^= (1 << 24);
+
+  h->v[0] += c->v[0];
+  h->v[2] += c->v[2];
+  h->v[4] += c->v[4];
+  h->v[6] += c->v[6];
+  h->v[8] += c->v[8];
+  fe1305x2_tobytearray(mac, h);
+}
+
+#endif  // OPENSSL_POLY1305_NEON

vendor/aws-lc-sys/aws-lc/crypto/poly1305/poly1305_test.cc

@@ -0,0 +1,102 @@
+// Copyright (c) 2015, Google Inc.
+// SPDX-License-Identifier: ISC
+
+#include <stdio.h>
+#include <string.h>
+
+#include <vector>
+
+#include <gtest/gtest.h>
+
+#include <openssl/poly1305.h>
+
+#include "../internal.h"
+#include "../test/file_test.h"
+#include "../test/test_util.h"
+
+
+static void TestSIMD(unsigned excess, const std::vector<uint8_t> &key,
+                     const std::vector<uint8_t> &in,
+                     const std::vector<uint8_t> &mac) {
+  poly1305_state state;
+  CRYPTO_poly1305_init(&state, key.data());
+
+  size_t done = 0;
+
+  // Feed 16 bytes in. Some implementations begin in non-SIMD mode and upgrade
+  // on-demand. Stress the upgrade path.
+  size_t todo = 16;
+  if (todo > in.size()) {
+    todo = in.size();
+  }
+  CRYPTO_poly1305_update(&state, in.data(), todo);
+  done += todo;
+
+  for (;;) {
+    // Feed 128 + |excess| bytes to test SIMD mode.
+    if (done + 128 + excess > in.size()) {
+      break;
+    }
+    CRYPTO_poly1305_update(&state, in.data() + done, 128 + excess);
+    done += 128 + excess;
+
+    // Feed |excess| bytes to ensure SIMD mode can handle short inputs.
+    if (done + excess > in.size()) {
+      break;
+    }
+    CRYPTO_poly1305_update(&state, in.data() + done, excess);
+    done += excess;
+  }
+
+  // Consume the remainder and finish.
+  CRYPTO_poly1305_update(&state, in.data() + done, in.size() - done);
+
+  uint8_t out[16];
+  CRYPTO_poly1305_finish(&state, out);
+  EXPECT_EQ(Bytes(out), Bytes(mac)) << "SIMD pattern " << excess << " failed.";
+}
+
+TEST(Poly1305Test, TestVectors) {
+  FileTestGTest("crypto/poly1305/poly1305_tests.txt", [](FileTest *t) {
+    std::vector<uint8_t> key, in, mac;
+    ASSERT_TRUE(t->GetBytes(&key, "Key"));
+    ASSERT_TRUE(t->GetBytes(&in, "Input"));
+    ASSERT_TRUE(t->GetBytes(&mac, "MAC"));
+    ASSERT_EQ(32u, key.size());
+    ASSERT_EQ(16u, mac.size());
+
+    // Test single-shot operation.
+    poly1305_state state;
+    CRYPTO_poly1305_init(&state, key.data());
+    CRYPTO_poly1305_update(&state, in.data(), in.size());
+    uint8_t out[16];
+    CRYPTO_poly1305_finish(&state, out);
+    EXPECT_EQ(Bytes(out), Bytes(mac)) << "Single-shot Poly1305 failed.";
+
+    // Test streaming byte-by-byte.
+    CRYPTO_poly1305_init(&state, key.data());
+    for (size_t i = 0; i < in.size(); i++) {
+      CRYPTO_poly1305_update(&state, &in[i], 1);
+    }
+    CRYPTO_poly1305_finish(&state, out);
+    EXPECT_EQ(Bytes(out), Bytes(mac)) << "Streaming Poly1305 failed.";
+
+    // Test |CRYPTO_poly1305_init| and |CRYPTO_poly1305_finish| work on
+    // unaligned values.
+    alignas(8) uint8_t unaligned_key[32 + 1];
+    OPENSSL_memcpy(unaligned_key + 1, key.data(), 32);
+    CRYPTO_poly1305_init(&state, unaligned_key + 1);
+    CRYPTO_poly1305_update(&state, in.data(), in.size());
+    alignas(8) uint8_t unaligned_out[16 + 1];
+    CRYPTO_poly1305_finish(&state, unaligned_out + 1);
+    EXPECT_EQ(Bytes(unaligned_out + 1, 16), Bytes(mac))
+        << "Unaligned Poly1305 failed.";
+
+    // Test SIMD stress patterns. OpenSSL's AVX2 assembly needs a multiple of
+    // four blocks, so test up to three blocks of excess.
+    TestSIMD(0, key, in, mac);
+    TestSIMD(16, key, in, mac);
+    TestSIMD(32, key, in, mac);
+    TestSIMD(48, key, in, mac);
+  });
+}

vendor/aws-lc-sys/aws-lc/crypto/poly1305/poly1305_vec.c

@@ -0,0 +1,833 @@
+// Copyright (c) 2014, Google Inc.
+// SPDX-License-Identifier: ISC
+
+// This implementation of poly1305 is by Andrew Moon
+// (https://github.com/floodyberry/poly1305-donna) and released as public
+// domain. It implements SIMD vectorization based on the algorithm described in
+// http://cr.yp.to/papers.html#neoncrypto. Unrolled to 2 powers, i.e. 64 byte
+// block size
+
+#include <openssl/poly1305.h>
+
+#include "../internal.h"
+
+
+#if defined(BORINGSSL_HAS_UINT128) && defined(OPENSSL_X86_64)
+
+#include <emmintrin.h>
+
+typedef __m128i xmmi;
+
+static const alignas(16) uint32_t poly1305_x64_sse2_message_mask[4] = {
+    (1 << 26) - 1, 0, (1 << 26) - 1, 0};
+static const alignas(16) uint32_t poly1305_x64_sse2_5[4] = {5, 0, 5, 0};
+static const alignas(16) uint32_t poly1305_x64_sse2_1shl128[4] = {
+    (1 << 24), 0, (1 << 24), 0};
+
+static inline uint128_t add128(uint128_t a, uint128_t b) { return a + b; }
+
+static inline uint128_t add128_64(uint128_t a, uint64_t b) { return a + b; }
+
+static inline uint128_t mul64x64_128(uint64_t a, uint64_t b) {
+  return (uint128_t)a * b;
+}
+
+static inline uint64_t lo128(uint128_t a) { return (uint64_t)a; }
+
+static inline uint64_t shr128(uint128_t v, const int shift) {
+  return (uint64_t)(v >> shift);
+}
+
+static inline uint64_t shr128_pair(uint64_t hi, uint64_t lo, const int shift) {
+  return (uint64_t)((((uint128_t)hi << 64) | lo) >> shift);
+}
+
+typedef struct poly1305_power_t {
+  union {
+    xmmi v;
+    uint64_t u[2];
+    uint32_t d[4];
+  } R20, R21, R22, R23, R24, S21, S22, S23, S24;
+} poly1305_power;
+
+typedef struct poly1305_state_internal_t {
+  poly1305_power P[2]; /* 288 bytes, top 32 bit halves unused = 144
+                          bytes of free storage */
+  union {
+    xmmi H[5];  //  80 bytes
+    uint64_t HH[10];
+  } u;
+  // uint64_t r0,r1,r2;       [24 bytes]
+  // uint64_t pad0,pad1;      [16 bytes]
+  uint64_t started;        //   8 bytes
+  uint64_t leftover;       //   8 bytes
+  uint8_t buffer[64];      //  64 bytes
+} poly1305_state_internal; /* 448 bytes total + 63 bytes for
+                              alignment = 511 bytes raw */
+
+OPENSSL_STATIC_ASSERT(
+    sizeof(struct poly1305_state_internal_t) + 63 <= sizeof(poly1305_state),
+    _poly1305_state_isn_t_large_enough_to_hold_aligned_poly1305_state_internal_t)
+
+static inline poly1305_state_internal *poly1305_aligned_state(
+    poly1305_state *state) {
+  return (poly1305_state_internal *)(((uint64_t)state + 63) & ~63);
+}
+
+static inline size_t poly1305_min(size_t a, size_t b) {
+  return (a < b) ? a : b;
+}
+
+void CRYPTO_poly1305_init(poly1305_state *state, const uint8_t key[32]) {
+  poly1305_state_internal *st = poly1305_aligned_state(state);
+  poly1305_power *p;
+  uint64_t r0, r1, r2;
+  uint64_t t0, t1;
+
+  // clamp key
+  t0 = CRYPTO_load_u64_le(key + 0);
+  t1 = CRYPTO_load_u64_le(key + 8);
+  r0 = t0 & 0xffc0fffffff;
+  t0 >>= 44;
+  t0 |= t1 << 20;
+  r1 = t0 & 0xfffffc0ffff;
+  t1 >>= 24;
+  r2 = t1 & 0x00ffffffc0f;
+
+  // store r in un-used space of st->P[1]
+  p = &st->P[1];
+  p->R20.d[1] = (uint32_t)(r0);
+  p->R20.d[3] = (uint32_t)(r0 >> 32);
+  p->R21.d[1] = (uint32_t)(r1);
+  p->R21.d[3] = (uint32_t)(r1 >> 32);
+  p->R22.d[1] = (uint32_t)(r2);
+  p->R22.d[3] = (uint32_t)(r2 >> 32);
+
+  // store pad
+  p->R23.d[1] = CRYPTO_load_u32_le(key + 16);
+  p->R23.d[3] = CRYPTO_load_u32_le(key + 20);
+  p->R24.d[1] = CRYPTO_load_u32_le(key + 24);
+  p->R24.d[3] = CRYPTO_load_u32_le(key + 28);
+
+  // H = 0
+  st->u.H[0] = _mm_setzero_si128();
+  st->u.H[1] = _mm_setzero_si128();
+  st->u.H[2] = _mm_setzero_si128();
+  st->u.H[3] = _mm_setzero_si128();
+  st->u.H[4] = _mm_setzero_si128();
+
+  st->started = 0;
+  st->leftover = 0;
+}
+
+static void poly1305_first_block(poly1305_state_internal *st,
+                                 const uint8_t *m) {
+  const xmmi MMASK = _mm_load_si128((const xmmi *)poly1305_x64_sse2_message_mask);
+  const xmmi FIVE = _mm_load_si128((const xmmi *)poly1305_x64_sse2_5);
+  const xmmi HIBIT = _mm_load_si128((const xmmi *)poly1305_x64_sse2_1shl128);
+  xmmi T5, T6;
+  poly1305_power *p;
+  uint128_t d[3];
+  uint64_t r0, r1, r2;
+  uint64_t r20, r21, r22, s22;
+  uint64_t pad0, pad1;
+  uint64_t c;
+  uint64_t i;
+
+  // pull out stored info
+  p = &st->P[1];
+
+  r0 = ((uint64_t)p->R20.d[3] << 32) | (uint64_t)p->R20.d[1];
+  r1 = ((uint64_t)p->R21.d[3] << 32) | (uint64_t)p->R21.d[1];
+  r2 = ((uint64_t)p->R22.d[3] << 32) | (uint64_t)p->R22.d[1];
+  pad0 = ((uint64_t)p->R23.d[3] << 32) | (uint64_t)p->R23.d[1];
+  pad1 = ((uint64_t)p->R24.d[3] << 32) | (uint64_t)p->R24.d[1];
+
+  // compute powers r^2,r^4
+  r20 = r0;
+  r21 = r1;
+  r22 = r2;
+  for (i = 0; i < 2; i++) {
+    s22 = r22 * (5 << 2);
+
+    d[0] = add128(mul64x64_128(r20, r20), mul64x64_128(r21 * 2, s22));
+    d[1] = add128(mul64x64_128(r22, s22), mul64x64_128(r20 * 2, r21));
+    d[2] = add128(mul64x64_128(r21, r21), mul64x64_128(r22 * 2, r20));
+
+    r20 = lo128(d[0]) & 0xfffffffffff;
+    c = shr128(d[0], 44);
+    d[1] = add128_64(d[1], c);
+    r21 = lo128(d[1]) & 0xfffffffffff;
+    c = shr128(d[1], 44);
+    d[2] = add128_64(d[2], c);
+    r22 = lo128(d[2]) & 0x3ffffffffff;
+    c = shr128(d[2], 42);
+    r20 += c * 5;
+    c = (r20 >> 44);
+    r20 = r20 & 0xfffffffffff;
+    r21 += c;
+
+    p->R20.v = _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)(r20)&0x3ffffff),
+                                 _MM_SHUFFLE(1, 0, 1, 0));
+    p->R21.v = _mm_shuffle_epi32(
+        _mm_cvtsi32_si128((uint32_t)((r20 >> 26) | (r21 << 18)) & 0x3ffffff),
+        _MM_SHUFFLE(1, 0, 1, 0));
+    p->R22.v =
+        _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)((r21 >> 8)) & 0x3ffffff),
+                          _MM_SHUFFLE(1, 0, 1, 0));
+    p->R23.v = _mm_shuffle_epi32(
+        _mm_cvtsi32_si128((uint32_t)((r21 >> 34) | (r22 << 10)) & 0x3ffffff),
+        _MM_SHUFFLE(1, 0, 1, 0));
+    p->R24.v = _mm_shuffle_epi32(_mm_cvtsi32_si128((uint32_t)((r22 >> 16))),
+                                 _MM_SHUFFLE(1, 0, 1, 0));
+    p->S21.v = _mm_mul_epu32(p->R21.v, FIVE);
+    p->S22.v = _mm_mul_epu32(p->R22.v, FIVE);
+    p->S23.v = _mm_mul_epu32(p->R23.v, FIVE);
+    p->S24.v = _mm_mul_epu32(p->R24.v, FIVE);
+    p--;
+  }
+
+  // put saved info back
+  p = &st->P[1];
+  p->R20.d[1] = (uint32_t)(r0);
+  p->R20.d[3] = (uint32_t)(r0 >> 32);
+  p->R21.d[1] = (uint32_t)(r1);
+  p->R21.d[3] = (uint32_t)(r1 >> 32);
+  p->R22.d[1] = (uint32_t)(r2);
+  p->R22.d[3] = (uint32_t)(r2 >> 32);
+  p->R23.d[1] = (uint32_t)(pad0);
+  p->R23.d[3] = (uint32_t)(pad0 >> 32);
+  p->R24.d[1] = (uint32_t)(pad1);
+  p->R24.d[3] = (uint32_t)(pad1 >> 32);
+
+  // H = [Mx,My]
+  T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 0)),
+                          _mm_loadl_epi64((const xmmi *)(m + 16)));
+  T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 8)),
+                          _mm_loadl_epi64((const xmmi *)(m + 24)));
+  st->u.H[0] = _mm_and_si128(MMASK, T5);
+  st->u.H[1] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+  T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
+  st->u.H[2] = _mm_and_si128(MMASK, T5);
+  st->u.H[3] = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+  st->u.H[4] = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
+}
+
+static void poly1305_blocks(poly1305_state_internal *st, const uint8_t *m,
+                            size_t bytes) {
+  const xmmi MMASK = _mm_load_si128((const xmmi *)poly1305_x64_sse2_message_mask);
+  const xmmi FIVE = _mm_load_si128((const xmmi *)poly1305_x64_sse2_5);
+  const xmmi HIBIT = _mm_load_si128((const xmmi *)poly1305_x64_sse2_1shl128);
+
+  poly1305_power *p;
+  xmmi H0, H1, H2, H3, H4;
+  xmmi T0, T1, T2, T3, T4, T5, T6;
+  xmmi M0, M1, M2, M3, M4;
+  xmmi C1, C2;
+
+  H0 = st->u.H[0];
+  H1 = st->u.H[1];
+  H2 = st->u.H[2];
+  H3 = st->u.H[3];
+  H4 = st->u.H[4];
+
+  while (bytes >= 64) {
+    // H *= [r^4,r^4]
+    p = &st->P[0];
+    T0 = _mm_mul_epu32(H0, p->R20.v);
+    T1 = _mm_mul_epu32(H0, p->R21.v);
+    T2 = _mm_mul_epu32(H0, p->R22.v);
+    T3 = _mm_mul_epu32(H0, p->R23.v);
+    T4 = _mm_mul_epu32(H0, p->R24.v);
+    T5 = _mm_mul_epu32(H1, p->S24.v);
+    T6 = _mm_mul_epu32(H1, p->R20.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H2, p->S23.v);
+    T6 = _mm_mul_epu32(H2, p->S24.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H3, p->S22.v);
+    T6 = _mm_mul_epu32(H3, p->S23.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H4, p->S21.v);
+    T6 = _mm_mul_epu32(H4, p->S22.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H1, p->R21.v);
+    T6 = _mm_mul_epu32(H1, p->R22.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H2, p->R20.v);
+    T6 = _mm_mul_epu32(H2, p->R21.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H3, p->S24.v);
+    T6 = _mm_mul_epu32(H3, p->R20.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H4, p->S23.v);
+    T6 = _mm_mul_epu32(H4, p->S24.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H1, p->R23.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(H2, p->R22.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(H3, p->R21.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(H4, p->R20.v);
+    T4 = _mm_add_epi64(T4, T5);
+
+    // H += [Mx,My]*[r^2,r^2]
+    T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 0)),
+                            _mm_loadl_epi64((const xmmi *)(m + 16)));
+    T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 8)),
+                            _mm_loadl_epi64((const xmmi *)(m + 24)));
+    M0 = _mm_and_si128(MMASK, T5);
+    M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+    T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
+    M2 = _mm_and_si128(MMASK, T5);
+    M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+    M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
+
+    p = &st->P[1];
+    T5 = _mm_mul_epu32(M0, p->R20.v);
+    T6 = _mm_mul_epu32(M0, p->R21.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(M1, p->S24.v);
+    T6 = _mm_mul_epu32(M1, p->R20.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(M2, p->S23.v);
+    T6 = _mm_mul_epu32(M2, p->S24.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(M3, p->S22.v);
+    T6 = _mm_mul_epu32(M3, p->S23.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(M4, p->S21.v);
+    T6 = _mm_mul_epu32(M4, p->S22.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(M0, p->R22.v);
+    T6 = _mm_mul_epu32(M0, p->R23.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(M1, p->R21.v);
+    T6 = _mm_mul_epu32(M1, p->R22.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(M2, p->R20.v);
+    T6 = _mm_mul_epu32(M2, p->R21.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(M3, p->S24.v);
+    T6 = _mm_mul_epu32(M3, p->R20.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(M4, p->S23.v);
+    T6 = _mm_mul_epu32(M4, p->S24.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(M0, p->R24.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(M1, p->R23.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(M2, p->R22.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(M3, p->R21.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(M4, p->R20.v);
+    T4 = _mm_add_epi64(T4, T5);
+
+    // H += [Mx,My]
+    T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 32)),
+                            _mm_loadl_epi64((const xmmi *)(m + 48)));
+    T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 40)),
+                            _mm_loadl_epi64((const xmmi *)(m + 56)));
+    M0 = _mm_and_si128(MMASK, T5);
+    M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+    T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
+    M2 = _mm_and_si128(MMASK, T5);
+    M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+    M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
+
+    T0 = _mm_add_epi64(T0, M0);
+    T1 = _mm_add_epi64(T1, M1);
+    T2 = _mm_add_epi64(T2, M2);
+    T3 = _mm_add_epi64(T3, M3);
+    T4 = _mm_add_epi64(T4, M4);
+
+    // reduce
+    C1 = _mm_srli_epi64(T0, 26);
+    C2 = _mm_srli_epi64(T3, 26);
+    T0 = _mm_and_si128(T0, MMASK);
+    T3 = _mm_and_si128(T3, MMASK);
+    T1 = _mm_add_epi64(T1, C1);
+    T4 = _mm_add_epi64(T4, C2);
+    C1 = _mm_srli_epi64(T1, 26);
+    C2 = _mm_srli_epi64(T4, 26);
+    T1 = _mm_and_si128(T1, MMASK);
+    T4 = _mm_and_si128(T4, MMASK);
+    T2 = _mm_add_epi64(T2, C1);
+    T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
+    C1 = _mm_srli_epi64(T2, 26);
+    C2 = _mm_srli_epi64(T0, 26);
+    T2 = _mm_and_si128(T2, MMASK);
+    T0 = _mm_and_si128(T0, MMASK);
+    T3 = _mm_add_epi64(T3, C1);
+    T1 = _mm_add_epi64(T1, C2);
+    C1 = _mm_srli_epi64(T3, 26);
+    T3 = _mm_and_si128(T3, MMASK);
+    T4 = _mm_add_epi64(T4, C1);
+
+    // H = (H*[r^4,r^4] + [Mx,My]*[r^2,r^2] + [Mx,My])
+    H0 = T0;
+    H1 = T1;
+    H2 = T2;
+    H3 = T3;
+    H4 = T4;
+
+    m += 64;
+    bytes -= 64;
+  }
+
+  st->u.H[0] = H0;
+  st->u.H[1] = H1;
+  st->u.H[2] = H2;
+  st->u.H[3] = H3;
+  st->u.H[4] = H4;
+}
+
+static size_t poly1305_combine(poly1305_state_internal *st, const uint8_t *m,
+                               size_t bytes) {
+  const xmmi MMASK = _mm_load_si128((const xmmi *)poly1305_x64_sse2_message_mask);
+  const xmmi HIBIT = _mm_load_si128((const xmmi *)poly1305_x64_sse2_1shl128);
+  const xmmi FIVE = _mm_load_si128((const xmmi *)poly1305_x64_sse2_5);
+
+  poly1305_power *p;
+  xmmi H0, H1, H2, H3, H4;
+  xmmi M0, M1, M2, M3, M4;
+  xmmi T0, T1, T2, T3, T4, T5, T6;
+  xmmi C1, C2;
+
+  uint64_t r0, r1, r2;
+  uint64_t t0, t1, t2, t3, t4;
+  uint64_t c;
+  size_t consumed = 0;
+
+  H0 = st->u.H[0];
+  H1 = st->u.H[1];
+  H2 = st->u.H[2];
+  H3 = st->u.H[3];
+  H4 = st->u.H[4];
+
+  // p = [r^2,r^2]
+  p = &st->P[1];
+
+  if (bytes >= 32) {
+    // H *= [r^2,r^2]
+    T0 = _mm_mul_epu32(H0, p->R20.v);
+    T1 = _mm_mul_epu32(H0, p->R21.v);
+    T2 = _mm_mul_epu32(H0, p->R22.v);
+    T3 = _mm_mul_epu32(H0, p->R23.v);
+    T4 = _mm_mul_epu32(H0, p->R24.v);
+    T5 = _mm_mul_epu32(H1, p->S24.v);
+    T6 = _mm_mul_epu32(H1, p->R20.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H2, p->S23.v);
+    T6 = _mm_mul_epu32(H2, p->S24.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H3, p->S22.v);
+    T6 = _mm_mul_epu32(H3, p->S23.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H4, p->S21.v);
+    T6 = _mm_mul_epu32(H4, p->S22.v);
+    T0 = _mm_add_epi64(T0, T5);
+    T1 = _mm_add_epi64(T1, T6);
+    T5 = _mm_mul_epu32(H1, p->R21.v);
+    T6 = _mm_mul_epu32(H1, p->R22.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H2, p->R20.v);
+    T6 = _mm_mul_epu32(H2, p->R21.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H3, p->S24.v);
+    T6 = _mm_mul_epu32(H3, p->R20.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H4, p->S23.v);
+    T6 = _mm_mul_epu32(H4, p->S24.v);
+    T2 = _mm_add_epi64(T2, T5);
+    T3 = _mm_add_epi64(T3, T6);
+    T5 = _mm_mul_epu32(H1, p->R23.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(H2, p->R22.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(H3, p->R21.v);
+    T4 = _mm_add_epi64(T4, T5);
+    T5 = _mm_mul_epu32(H4, p->R20.v);
+    T4 = _mm_add_epi64(T4, T5);
+
+    // H += [Mx,My]
+    T5 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 0)),
+                            _mm_loadl_epi64((const xmmi *)(m + 16)));
+    T6 = _mm_unpacklo_epi64(_mm_loadl_epi64((const xmmi *)(m + 8)),
+                            _mm_loadl_epi64((const xmmi *)(m + 24)));
+    M0 = _mm_and_si128(MMASK, T5);
+    M1 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+    T5 = _mm_or_si128(_mm_srli_epi64(T5, 52), _mm_slli_epi64(T6, 12));
+    M2 = _mm_and_si128(MMASK, T5);
+    M3 = _mm_and_si128(MMASK, _mm_srli_epi64(T5, 26));
+    M4 = _mm_or_si128(_mm_srli_epi64(T6, 40), HIBIT);
+
+    T0 = _mm_add_epi64(T0, M0);
+    T1 = _mm_add_epi64(T1, M1);
+    T2 = _mm_add_epi64(T2, M2);
+    T3 = _mm_add_epi64(T3, M3);
+    T4 = _mm_add_epi64(T4, M4);
+
+    // reduce
+    C1 = _mm_srli_epi64(T0, 26);
+    C2 = _mm_srli_epi64(T3, 26);
+    T0 = _mm_and_si128(T0, MMASK);
+    T3 = _mm_and_si128(T3, MMASK);
+    T1 = _mm_add_epi64(T1, C1);
+    T4 = _mm_add_epi64(T4, C2);
+    C1 = _mm_srli_epi64(T1, 26);
+    C2 = _mm_srli_epi64(T4, 26);
+    T1 = _mm_and_si128(T1, MMASK);
+    T4 = _mm_and_si128(T4, MMASK);
+    T2 = _mm_add_epi64(T2, C1);
+    T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
+    C1 = _mm_srli_epi64(T2, 26);
+    C2 = _mm_srli_epi64(T0, 26);
+    T2 = _mm_and_si128(T2, MMASK);
+    T0 = _mm_and_si128(T0, MMASK);
+    T3 = _mm_add_epi64(T3, C1);
+    T1 = _mm_add_epi64(T1, C2);
+    C1 = _mm_srli_epi64(T3, 26);
+    T3 = _mm_and_si128(T3, MMASK);
+    T4 = _mm_add_epi64(T4, C1);
+
+    // H = (H*[r^2,r^2] + [Mx,My])
+    H0 = T0;
+    H1 = T1;
+    H2 = T2;
+    H3 = T3;
+    H4 = T4;
+
+    consumed = 32;
+  }
+
+  // finalize, H *= [r^2,r]
+  r0 = ((uint64_t)p->R20.d[3] << 32) | (uint64_t)p->R20.d[1];
+  r1 = ((uint64_t)p->R21.d[3] << 32) | (uint64_t)p->R21.d[1];
+  r2 = ((uint64_t)p->R22.d[3] << 32) | (uint64_t)p->R22.d[1];
+
+  p->R20.d[2] = (uint32_t)(r0)&0x3ffffff;
+  p->R21.d[2] = (uint32_t)((r0 >> 26) | (r1 << 18)) & 0x3ffffff;
+  p->R22.d[2] = (uint32_t)((r1 >> 8)) & 0x3ffffff;
+  p->R23.d[2] = (uint32_t)((r1 >> 34) | (r2 << 10)) & 0x3ffffff;
+  p->R24.d[2] = (uint32_t)((r2 >> 16));
+  p->S21.d[2] = p->R21.d[2] * 5;
+  p->S22.d[2] = p->R22.d[2] * 5;
+  p->S23.d[2] = p->R23.d[2] * 5;
+  p->S24.d[2] = p->R24.d[2] * 5;
+
+  // H *= [r^2,r]
+  T0 = _mm_mul_epu32(H0, p->R20.v);
+  T1 = _mm_mul_epu32(H0, p->R21.v);
+  T2 = _mm_mul_epu32(H0, p->R22.v);
+  T3 = _mm_mul_epu32(H0, p->R23.v);
+  T4 = _mm_mul_epu32(H0, p->R24.v);
+  T5 = _mm_mul_epu32(H1, p->S24.v);
+  T6 = _mm_mul_epu32(H1, p->R20.v);
+  T0 = _mm_add_epi64(T0, T5);
+  T1 = _mm_add_epi64(T1, T6);
+  T5 = _mm_mul_epu32(H2, p->S23.v);
+  T6 = _mm_mul_epu32(H2, p->S24.v);
+  T0 = _mm_add_epi64(T0, T5);
+  T1 = _mm_add_epi64(T1, T6);
+  T5 = _mm_mul_epu32(H3, p->S22.v);
+  T6 = _mm_mul_epu32(H3, p->S23.v);
+  T0 = _mm_add_epi64(T0, T5);
+  T1 = _mm_add_epi64(T1, T6);
+  T5 = _mm_mul_epu32(H4, p->S21.v);
+  T6 = _mm_mul_epu32(H4, p->S22.v);
+  T0 = _mm_add_epi64(T0, T5);
+  T1 = _mm_add_epi64(T1, T6);
+  T5 = _mm_mul_epu32(H1, p->R21.v);
+  T6 = _mm_mul_epu32(H1, p->R22.v);
+  T2 = _mm_add_epi64(T2, T5);
+  T3 = _mm_add_epi64(T3, T6);
+  T5 = _mm_mul_epu32(H2, p->R20.v);
+  T6 = _mm_mul_epu32(H2, p->R21.v);
+  T2 = _mm_add_epi64(T2, T5);
+  T3 = _mm_add_epi64(T3, T6);
+  T5 = _mm_mul_epu32(H3, p->S24.v);
+  T6 = _mm_mul_epu32(H3, p->R20.v);
+  T2 = _mm_add_epi64(T2, T5);
+  T3 = _mm_add_epi64(T3, T6);
+  T5 = _mm_mul_epu32(H4, p->S23.v);
+  T6 = _mm_mul_epu32(H4, p->S24.v);
+  T2 = _mm_add_epi64(T2, T5);
+  T3 = _mm_add_epi64(T3, T6);
+  T5 = _mm_mul_epu32(H1, p->R23.v);
+  T4 = _mm_add_epi64(T4, T5);
+  T5 = _mm_mul_epu32(H2, p->R22.v);
+  T4 = _mm_add_epi64(T4, T5);
+  T5 = _mm_mul_epu32(H3, p->R21.v);
+  T4 = _mm_add_epi64(T4, T5);
+  T5 = _mm_mul_epu32(H4, p->R20.v);
+  T4 = _mm_add_epi64(T4, T5);
+
+  C1 = _mm_srli_epi64(T0, 26);
+  C2 = _mm_srli_epi64(T3, 26);
+  T0 = _mm_and_si128(T0, MMASK);
+  T3 = _mm_and_si128(T3, MMASK);
+  T1 = _mm_add_epi64(T1, C1);
+  T4 = _mm_add_epi64(T4, C2);
+  C1 = _mm_srli_epi64(T1, 26);
+  C2 = _mm_srli_epi64(T4, 26);
+  T1 = _mm_and_si128(T1, MMASK);
+  T4 = _mm_and_si128(T4, MMASK);
+  T2 = _mm_add_epi64(T2, C1);
+  T0 = _mm_add_epi64(T0, _mm_mul_epu32(C2, FIVE));
+  C1 = _mm_srli_epi64(T2, 26);
+  C2 = _mm_srli_epi64(T0, 26);
+  T2 = _mm_and_si128(T2, MMASK);
+  T0 = _mm_and_si128(T0, MMASK);
+  T3 = _mm_add_epi64(T3, C1);
+  T1 = _mm_add_epi64(T1, C2);
+  C1 = _mm_srli_epi64(T3, 26);
+  T3 = _mm_and_si128(T3, MMASK);
+  T4 = _mm_add_epi64(T4, C1);
+
+  // H = H[0]+H[1]
+  H0 = _mm_add_epi64(T0, _mm_srli_si128(T0, 8));
+  H1 = _mm_add_epi64(T1, _mm_srli_si128(T1, 8));
+  H2 = _mm_add_epi64(T2, _mm_srli_si128(T2, 8));
+  H3 = _mm_add_epi64(T3, _mm_srli_si128(T3, 8));
+  H4 = _mm_add_epi64(T4, _mm_srli_si128(T4, 8));
+
+  t0 = _mm_cvtsi128_si32(H0);
+  c = (t0 >> 26);
+  t0 &= 0x3ffffff;
+  t1 = _mm_cvtsi128_si32(H1) + c;
+  c = (t1 >> 26);
+  t1 &= 0x3ffffff;
+  t2 = _mm_cvtsi128_si32(H2) + c;
+  c = (t2 >> 26);
+  t2 &= 0x3ffffff;
+  t3 = _mm_cvtsi128_si32(H3) + c;
+  c = (t3 >> 26);
+  t3 &= 0x3ffffff;
+  t4 = _mm_cvtsi128_si32(H4) + c;
+  c = (t4 >> 26);
+  t4 &= 0x3ffffff;
+  t0 = t0 + (c * 5);
+  c = (t0 >> 26);
+  t0 &= 0x3ffffff;
+  t1 = t1 + c;
+
+  st->u.HH[0] = ((t0) | (t1 << 26)) & UINT64_C(0xfffffffffff);
+  st->u.HH[1] = ((t1 >> 18) | (t2 << 8) | (t3 << 34)) & UINT64_C(0xfffffffffff);
+  st->u.HH[2] = ((t3 >> 10) | (t4 << 16)) & UINT64_C(0x3ffffffffff);
+
+  return consumed;
+}
+
+void CRYPTO_poly1305_update(poly1305_state *state, const uint8_t *m,
+                            size_t bytes) {
+  poly1305_state_internal *st = poly1305_aligned_state(state);
+  size_t want;
+
+  // Work around a C language bug. See https://crbug.com/1019588.
+  if (bytes == 0) {
+    return;
+  }
+
+  // need at least 32 initial bytes to start the accelerated branch
+  if (!st->started) {
+    if ((st->leftover == 0) && (bytes > 32)) {
+      poly1305_first_block(st, m);
+      m += 32;
+      bytes -= 32;
+    } else {
+      want = poly1305_min(32 - st->leftover, bytes);
+      OPENSSL_memcpy(st->buffer + st->leftover, m, want);
+      bytes -= want;
+      m += want;
+      st->leftover += want;
+      if ((st->leftover < 32) || (bytes == 0)) {
+        return;
+      }
+      poly1305_first_block(st, st->buffer);
+      st->leftover = 0;
+    }
+    st->started = 1;
+  }
+
+  // handle leftover
+  if (st->leftover) {
+    want = poly1305_min(64 - st->leftover, bytes);
+    OPENSSL_memcpy(st->buffer + st->leftover, m, want);
+    bytes -= want;
+    m += want;
+    st->leftover += want;
+    if (st->leftover < 64) {
+      return;
+    }
+    poly1305_blocks(st, st->buffer, 64);
+    st->leftover = 0;
+  }
+
+  // process 64 byte blocks
+  if (bytes >= 64) {
+    want = (bytes & ~63);
+    poly1305_blocks(st, m, want);
+    m += want;
+    bytes -= want;
+  }
+
+  if (bytes) {
+    OPENSSL_memcpy(st->buffer + st->leftover, m, bytes);
+    st->leftover += bytes;
+  }
+}
+
+void CRYPTO_poly1305_finish(poly1305_state *state, uint8_t mac[16]) {
+  poly1305_state_internal *st = poly1305_aligned_state(state);
+  size_t leftover = st->leftover;
+  uint8_t *m = st->buffer;
+  uint128_t d[3];
+  uint64_t h0, h1, h2;
+  uint64_t t0, t1;
+  uint64_t g0, g1, g2, c, nc;
+  uint64_t r0, r1, r2, s1, s2;
+  poly1305_power *p;
+
+  if (st->started) {
+    size_t consumed = poly1305_combine(st, m, leftover);
+    leftover -= consumed;
+    m += consumed;
+  }
+
+  // st->HH will either be 0 or have the combined result
+  h0 = st->u.HH[0];
+  h1 = st->u.HH[1];
+  h2 = st->u.HH[2];
+
+  p = &st->P[1];
+  r0 = ((uint64_t)p->R20.d[3] << 32) | (uint64_t)p->R20.d[1];
+  r1 = ((uint64_t)p->R21.d[3] << 32) | (uint64_t)p->R21.d[1];
+  r2 = ((uint64_t)p->R22.d[3] << 32) | (uint64_t)p->R22.d[1];
+  s1 = r1 * (5 << 2);
+  s2 = r2 * (5 << 2);
+
+  if (leftover < 16) {
+    goto poly1305_donna_atmost15bytes;
+  }
+
+poly1305_donna_atleast16bytes:
+  t0 = CRYPTO_load_u64_le(m + 0);
+  t1 = CRYPTO_load_u64_le(m + 8);
+  h0 += t0 & 0xfffffffffff;
+  t0 = shr128_pair(t1, t0, 44);
+  h1 += t0 & 0xfffffffffff;
+  h2 += (t1 >> 24) | ((uint64_t)1 << 40);
+
+poly1305_donna_mul:
+  d[0] = add128(add128(mul64x64_128(h0, r0), mul64x64_128(h1, s2)),
+                mul64x64_128(h2, s1));
+  d[1] = add128(add128(mul64x64_128(h0, r1), mul64x64_128(h1, r0)),
+                mul64x64_128(h2, s2));
+  d[2] = add128(add128(mul64x64_128(h0, r2), mul64x64_128(h1, r1)),
+                mul64x64_128(h2, r0));
+  h0 = lo128(d[0]) & 0xfffffffffff;
+  c = shr128(d[0], 44);
+  d[1] = add128_64(d[1], c);
+  h1 = lo128(d[1]) & 0xfffffffffff;
+  c = shr128(d[1], 44);
+  d[2] = add128_64(d[2], c);
+  h2 = lo128(d[2]) & 0x3ffffffffff;
+  c = shr128(d[2], 42);
+  h0 += c * 5;
+
+  m += 16;
+  leftover -= 16;
+  if (leftover >= 16) {
+    goto poly1305_donna_atleast16bytes;
+  }
+
+// final bytes
+poly1305_donna_atmost15bytes:
+  if (!leftover) {
+    goto poly1305_donna_finish;
+  }
+
+  m[leftover++] = 1;
+  OPENSSL_memset(m + leftover, 0, 16 - leftover);
+  leftover = 16;
+
+  t0 = CRYPTO_load_u64_le(m + 0);
+  t1 = CRYPTO_load_u64_le(m + 8);
+  h0 += t0 & 0xfffffffffff;
+  t0 = shr128_pair(t1, t0, 44);
+  h1 += t0 & 0xfffffffffff;
+  h2 += (t1 >> 24);
+
+  goto poly1305_donna_mul;
+
+poly1305_donna_finish:
+  c = (h0 >> 44);
+  h0 &= 0xfffffffffff;
+  h1 += c;
+  c = (h1 >> 44);
+  h1 &= 0xfffffffffff;
+  h2 += c;
+  c = (h2 >> 42);
+  h2 &= 0x3ffffffffff;
+  h0 += c * 5;
+
+  g0 = h0 + 5;
+  c = (g0 >> 44);
+  g0 &= 0xfffffffffff;
+  g1 = h1 + c;
+  c = (g1 >> 44);
+  g1 &= 0xfffffffffff;
+  g2 = h2 + c - ((uint64_t)1 << 42);
+
+  c = (g2 >> 63) - 1;
+  nc = ~c;
+  h0 = (h0 & nc) | (g0 & c);
+  h1 = (h1 & nc) | (g1 & c);
+  h2 = (h2 & nc) | (g2 & c);
+
+  // pad
+  t0 = ((uint64_t)p->R23.d[3] << 32) | (uint64_t)p->R23.d[1];
+  t1 = ((uint64_t)p->R24.d[3] << 32) | (uint64_t)p->R24.d[1];
+  h0 += (t0 & 0xfffffffffff);
+  c = (h0 >> 44);
+  h0 &= 0xfffffffffff;
+  t0 = shr128_pair(t1, t0, 44);
+  h1 += (t0 & 0xfffffffffff) + c;
+  c = (h1 >> 44);
+  h1 &= 0xfffffffffff;
+  t1 = (t1 >> 24);
+  h2 += (t1)+c;
+
+  CRYPTO_store_u64_le(mac + 0, ((h0) | (h1 << 44)));
+  CRYPTO_store_u64_le(mac + 8, ((h1 >> 20) | (h2 << 24)));
+}
+
+#endif  // BORINGSSL_HAS_UINT128 && OPENSSL_X86_64