1/*
2 * Copyright 2013-2016 The OpenSSL Project Authors. All Rights Reserved.
3 * Copyright (c) 2012, Intel Corporation. All Rights Reserved.
4 *
5 * Licensed under the OpenSSL license (the "License"). You may not use
6 * this file except in compliance with the License. You can obtain a copy
7 * in the file LICENSE in the source distribution or at
8 * https://www.openssl.org/source/license.html
9 *
10 * Originally written by Shay Gueron (1, 2), and Vlad Krasnov (1)
11 * (1) Intel Corporation, Israel Development Center, Haifa, Israel
12 * (2) University of Haifa, Israel
13 */
14
15#ifndef OPENSSL_HEADER_BN_RSAZ_EXP_H
16#define OPENSSL_HEADER_BN_RSAZ_EXP_H
17
18#include <openssl/bn.h>
19#include <openssl/cpu.h>
20
21#include "internal.h"
22
23#if defined(__cplusplus)
24extern "C" {
25#endif
26
27#if !defined(OPENSSL_NO_ASM) && defined(OPENSSL_X86_64)
28#define RSAZ_ENABLED
29
30
31// RSAZ_1024_mod_exp_avx2 sets |result| to |base_norm| raised to |exponent|
32// modulo |m_norm|. |base_norm| must be fully-reduced and |exponent| must have
33// the high bit set (it is 1024 bits wide). |RR| and |k0| must be |RR| and |n0|,
34// respectively, extracted from |m_norm|'s |BN_MONT_CTX|. |storage_words| is a
35// temporary buffer that must be aligned to |MOD_EXP_CTIME_MIN_CACHE_LINE_WIDTH|
36// bytes.
37void RSAZ_1024_mod_exp_avx2(BN_ULONG result[16], const BN_ULONG base_norm[16],
38 const BN_ULONG exponent[16],
39 const BN_ULONG m_norm[16], const BN_ULONG RR[16],
40 BN_ULONG k0,
41 BN_ULONG storage_words[MOD_EXP_CTIME_STORAGE_LEN]);
42
43OPENSSL_INLINE int rsaz_avx2_capable(void) {
44 const uint32_t *cap = OPENSSL_ia32cap_get();
45 return (cap[2] & (1 << 5)) != 0; // AVX2
46}
47
48OPENSSL_INLINE int rsaz_avx2_preferred(void) {
49 const uint32_t *cap = OPENSSL_ia32cap_get();
50 static const uint32_t kBMI2AndADX = (1 << 8) | (1 << 19);
51 if ((cap[2] & kBMI2AndADX) == kBMI2AndADX) {
52 // If BMI2 and ADX are available, x86_64-mont5.pl is faster.
53 return 0;
54 }
55 return (cap[2] & (1 << 5)) != 0; // AVX2
56}
57
58
59// Assembly functions.
60
61// RSAZ represents 1024-bit integers using unsaturated 29-bit limbs stored in
62// 64-bit integers. This requires 36 limbs but padded up to 40.
63//
64// See crypto/bn/asm/rsaz-avx2.pl for further details.
65
66// rsaz_1024_norm2red_avx2 converts |norm| from |BIGNUM| to RSAZ representation
67// and writes the result to |red|.
68void rsaz_1024_norm2red_avx2(BN_ULONG red[40], const BN_ULONG norm[16]);
69
70// rsaz_1024_mul_avx2 computes |a| * |b| mod |n| and writes the result to |ret|.
71// Inputs and outputs are in Montgomery form, using RSAZ's representation. |k|
72// is -|n|^-1 mod 2^64 or |n0| from |BN_MONT_CTX|.
73void rsaz_1024_mul_avx2(BN_ULONG ret[40], const BN_ULONG a[40],
74 const BN_ULONG b[40], const BN_ULONG n[40], BN_ULONG k);
75
76// rsaz_1024_mul_avx2 computes |a|^(2*|count|) mod |n| and writes the result to
77// |ret|. Inputs and outputs are in Montgomery form, using RSAZ's
78// representation. |k| is -|n|^-1 mod 2^64 or |n0| from |BN_MONT_CTX|.
79void rsaz_1024_sqr_avx2(BN_ULONG ret[40], const BN_ULONG a[40],
80 const BN_ULONG n[40], BN_ULONG k, int count);
81
82// rsaz_1024_scatter5_avx2 stores |val| at index |i| of |tbl|. |i| must be
83// positive and at most 31. Note the table only uses 18 |BN_ULONG|s per entry
84// instead of 40. It packs two 29-bit limbs into each |BN_ULONG| and only stores
85// 36 limbs rather than the padded 40.
86void rsaz_1024_scatter5_avx2(BN_ULONG tbl[32 * 18], const BN_ULONG val[40],
87 int i);
88
89// rsaz_1024_gather5_avx2 loads index |i| of |tbl| and writes it to |val|.
90void rsaz_1024_gather5_avx2(BN_ULONG val[40], const BN_ULONG tbl[32 * 18],
91 int i);
92
93// rsaz_1024_red2norm_avx2 converts |red| from RSAZ to |BIGNUM| representation
94// and writes the result to |norm|.
95void rsaz_1024_red2norm_avx2(BN_ULONG norm[16], const BN_ULONG red[40]);
96
97
98#endif // !OPENSSL_NO_ASM && OPENSSL_X86_64
99
100#if defined(__cplusplus)
101} // extern "C"
102#endif
103
104#endif // OPENSSL_HEADER_BN_RSAZ_EXP_H
105