Important changes to repositories hosted on mbed.com
Mbed hosted mercurial repositories are deprecated and are due to be permanently deleted in July 2026.
To keep a copy of this software download the repository Zip archive or clone locally using Mercurial.
It is also possible to export all your personal repositories from the account settings page.
wolfcrypt/src/sha256.c@16:048e5e270a58, 2019-11-19 (annotated)
- Committer:
- sPymbed
- Date:
- Tue Nov 19 14:32:16 2019 +0000
- Revision:
- 16:048e5e270a58
- Parent:
- 15:117db924cf7c
working ssl
Who changed what in which revision?
User | Revision | Line number | New contents of line |
---|---|---|---|
wolfSSL | 15:117db924cf7c | 1 | /* sha256.c |
wolfSSL | 15:117db924cf7c | 2 | * |
wolfSSL | 15:117db924cf7c | 3 | * Copyright (C) 2006-2017 wolfSSL Inc. |
wolfSSL | 15:117db924cf7c | 4 | * |
wolfSSL | 15:117db924cf7c | 5 | * This file is part of wolfSSL. |
wolfSSL | 15:117db924cf7c | 6 | * |
wolfSSL | 15:117db924cf7c | 7 | * wolfSSL is free software; you can redistribute it and/or modify |
wolfSSL | 15:117db924cf7c | 8 | * it under the terms of the GNU General Public License as published by |
wolfSSL | 15:117db924cf7c | 9 | * the Free Software Foundation; either version 2 of the License, or |
wolfSSL | 15:117db924cf7c | 10 | * (at your option) any later version. |
wolfSSL | 15:117db924cf7c | 11 | * |
wolfSSL | 15:117db924cf7c | 12 | * wolfSSL is distributed in the hope that it will be useful, |
wolfSSL | 15:117db924cf7c | 13 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
wolfSSL | 15:117db924cf7c | 14 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
wolfSSL | 15:117db924cf7c | 15 | * GNU General Public License for more details. |
wolfSSL | 15:117db924cf7c | 16 | * |
wolfSSL | 15:117db924cf7c | 17 | * You should have received a copy of the GNU General Public License |
wolfSSL | 15:117db924cf7c | 18 | * along with this program; if not, write to the Free Software |
wolfSSL | 15:117db924cf7c | 19 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335, USA |
wolfSSL | 15:117db924cf7c | 20 | */ |
wolfSSL | 15:117db924cf7c | 21 | |
wolfSSL | 15:117db924cf7c | 22 | |
wolfSSL | 15:117db924cf7c | 23 | /* code submitted by raphael.huck@efixo.com */ |
wolfSSL | 15:117db924cf7c | 24 | |
wolfSSL | 15:117db924cf7c | 25 | #ifdef HAVE_CONFIG_H |
wolfSSL | 15:117db924cf7c | 26 | #include <config.h> |
wolfSSL | 15:117db924cf7c | 27 | #endif |
wolfSSL | 15:117db924cf7c | 28 | |
wolfSSL | 15:117db924cf7c | 29 | #include <wolfssl/wolfcrypt/settings.h> |
wolfSSL | 15:117db924cf7c | 30 | |
wolfSSL | 15:117db924cf7c | 31 | #if !defined(NO_SHA256) && !defined(WOLFSSL_ARMASM) |
wolfSSL | 15:117db924cf7c | 32 | |
wolfSSL | 15:117db924cf7c | 33 | #if defined(HAVE_FIPS) && \ |
wolfSSL | 15:117db924cf7c | 34 | defined(HAVE_FIPS_VERSION) && (HAVE_FIPS_VERSION >= 2) |
wolfSSL | 15:117db924cf7c | 35 | |
wolfSSL | 15:117db924cf7c | 36 | /* set NO_WRAPPERS before headers, use direct internal f()s not wrappers */ |
wolfSSL | 15:117db924cf7c | 37 | #define FIPS_NO_WRAPPERS |
wolfSSL | 15:117db924cf7c | 38 | |
wolfSSL | 15:117db924cf7c | 39 | #ifdef USE_WINDOWS_API |
wolfSSL | 15:117db924cf7c | 40 | #pragma code_seg(".fipsA$d") |
wolfSSL | 15:117db924cf7c | 41 | #pragma const_seg(".fipsB$d") |
wolfSSL | 15:117db924cf7c | 42 | #endif |
wolfSSL | 15:117db924cf7c | 43 | #endif |
wolfSSL | 15:117db924cf7c | 44 | |
wolfSSL | 15:117db924cf7c | 45 | #include <wolfssl/wolfcrypt/sha256.h> |
wolfSSL | 15:117db924cf7c | 46 | #include <wolfssl/wolfcrypt/error-crypt.h> |
wolfSSL | 15:117db924cf7c | 47 | #include <wolfssl/wolfcrypt/cpuid.h> |
wolfSSL | 15:117db924cf7c | 48 | |
wolfSSL | 15:117db924cf7c | 49 | /* fips wrapper calls, user can call direct */ |
wolfSSL | 15:117db924cf7c | 50 | #if defined(HAVE_FIPS) && \ |
wolfSSL | 15:117db924cf7c | 51 | (!defined(HAVE_FIPS_VERSION) || (HAVE_FIPS_VERSION < 2)) |
wolfSSL | 15:117db924cf7c | 52 | |
wolfSSL | 15:117db924cf7c | 53 | int wc_InitSha256(wc_Sha256* sha) |
wolfSSL | 15:117db924cf7c | 54 | { |
wolfSSL | 15:117db924cf7c | 55 | if (sha == NULL) { |
wolfSSL | 15:117db924cf7c | 56 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 57 | } |
wolfSSL | 15:117db924cf7c | 58 | return InitSha256_fips(sha); |
wolfSSL | 15:117db924cf7c | 59 | } |
wolfSSL | 15:117db924cf7c | 60 | int wc_InitSha256_ex(wc_Sha256* sha, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 61 | { |
wolfSSL | 15:117db924cf7c | 62 | (void)heap; |
wolfSSL | 15:117db924cf7c | 63 | (void)devId; |
wolfSSL | 15:117db924cf7c | 64 | if (sha == NULL) { |
wolfSSL | 15:117db924cf7c | 65 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 66 | } |
wolfSSL | 15:117db924cf7c | 67 | return InitSha256_fips(sha); |
wolfSSL | 15:117db924cf7c | 68 | } |
wolfSSL | 15:117db924cf7c | 69 | int wc_Sha256Update(wc_Sha256* sha, const byte* data, word32 len) |
wolfSSL | 15:117db924cf7c | 70 | { |
wolfSSL | 15:117db924cf7c | 71 | if (sha == NULL || (data == NULL && len > 0)) { |
wolfSSL | 15:117db924cf7c | 72 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 73 | } |
wolfSSL | 15:117db924cf7c | 74 | |
wolfSSL | 15:117db924cf7c | 75 | if (data == NULL && len == 0) { |
wolfSSL | 15:117db924cf7c | 76 | /* valid, but do nothing */ |
wolfSSL | 15:117db924cf7c | 77 | return 0; |
wolfSSL | 15:117db924cf7c | 78 | } |
wolfSSL | 15:117db924cf7c | 79 | |
wolfSSL | 15:117db924cf7c | 80 | return Sha256Update_fips(sha, data, len); |
wolfSSL | 15:117db924cf7c | 81 | } |
wolfSSL | 15:117db924cf7c | 82 | int wc_Sha256Final(wc_Sha256* sha, byte* out) |
wolfSSL | 15:117db924cf7c | 83 | { |
wolfSSL | 15:117db924cf7c | 84 | if (sha == NULL || out == NULL) { |
wolfSSL | 15:117db924cf7c | 85 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 86 | } |
wolfSSL | 15:117db924cf7c | 87 | return Sha256Final_fips(sha, out); |
wolfSSL | 15:117db924cf7c | 88 | } |
wolfSSL | 15:117db924cf7c | 89 | void wc_Sha256Free(wc_Sha256* sha) |
wolfSSL | 15:117db924cf7c | 90 | { |
wolfSSL | 15:117db924cf7c | 91 | (void)sha; |
wolfSSL | 15:117db924cf7c | 92 | /* Not supported in FIPS */ |
wolfSSL | 15:117db924cf7c | 93 | } |
wolfSSL | 15:117db924cf7c | 94 | |
wolfSSL | 15:117db924cf7c | 95 | #else /* else build without fips, or for FIPS v2 */ |
wolfSSL | 15:117db924cf7c | 96 | |
wolfSSL | 15:117db924cf7c | 97 | |
wolfSSL | 15:117db924cf7c | 98 | #if defined(WOLFSSL_TI_HASH) |
wolfSSL | 15:117db924cf7c | 99 | /* #include <wolfcrypt/src/port/ti/ti-hash.c> included by wc_port.c */ |
wolfSSL | 15:117db924cf7c | 100 | #else |
wolfSSL | 15:117db924cf7c | 101 | |
wolfSSL | 15:117db924cf7c | 102 | #include <wolfssl/wolfcrypt/logging.h> |
wolfSSL | 15:117db924cf7c | 103 | |
wolfSSL | 15:117db924cf7c | 104 | #ifdef NO_INLINE |
wolfSSL | 15:117db924cf7c | 105 | #include <wolfssl/wolfcrypt/misc.h> |
wolfSSL | 15:117db924cf7c | 106 | #else |
wolfSSL | 15:117db924cf7c | 107 | #define WOLFSSL_MISC_INCLUDED |
wolfSSL | 15:117db924cf7c | 108 | #include <wolfcrypt/src/misc.c> |
wolfSSL | 15:117db924cf7c | 109 | #endif |
wolfSSL | 15:117db924cf7c | 110 | |
wolfSSL | 15:117db924cf7c | 111 | |
wolfSSL | 15:117db924cf7c | 112 | #if defined(USE_INTEL_SPEEDUP) |
wolfSSL | 15:117db924cf7c | 113 | #define HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 114 | |
wolfSSL | 15:117db924cf7c | 115 | #if defined(__GNUC__) && ((__GNUC__ < 4) || \ |
wolfSSL | 15:117db924cf7c | 116 | (__GNUC__ == 4 && __GNUC_MINOR__ <= 8)) |
wolfSSL | 15:117db924cf7c | 117 | #define NO_AVX2_SUPPORT |
wolfSSL | 15:117db924cf7c | 118 | #endif |
wolfSSL | 15:117db924cf7c | 119 | #if defined(__clang__) && ((__clang_major__ < 3) || \ |
wolfSSL | 15:117db924cf7c | 120 | (__clang_major__ == 3 && __clang_minor__ <= 5)) |
wolfSSL | 15:117db924cf7c | 121 | #define NO_AVX2_SUPPORT |
wolfSSL | 15:117db924cf7c | 122 | #elif defined(__clang__) && defined(NO_AVX2_SUPPORT) |
wolfSSL | 15:117db924cf7c | 123 | #undef NO_AVX2_SUPPORT |
wolfSSL | 15:117db924cf7c | 124 | #endif |
wolfSSL | 15:117db924cf7c | 125 | |
wolfSSL | 15:117db924cf7c | 126 | #define HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 127 | #ifndef NO_AVX2_SUPPORT |
wolfSSL | 15:117db924cf7c | 128 | #define HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 129 | #endif |
wolfSSL | 15:117db924cf7c | 130 | #endif /* USE_INTEL_SPEEDUP */ |
wolfSSL | 15:117db924cf7c | 131 | |
wolfSSL | 15:117db924cf7c | 132 | #if defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 133 | #define HAVE_INTEL_RORX |
wolfSSL | 15:117db924cf7c | 134 | #endif |
wolfSSL | 15:117db924cf7c | 135 | |
wolfSSL | 15:117db924cf7c | 136 | |
wolfSSL | 15:117db924cf7c | 137 | #if !defined(WOLFSSL_PIC32MZ_HASH) && !defined(STM32_HASH_SHA2) && \ |
wolfSSL | 15:117db924cf7c | 138 | (!defined(WOLFSSL_IMX6_CAAM) || defined(NO_IMX6_CAAM_HASH)) |
wolfSSL | 15:117db924cf7c | 139 | static int InitSha256(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 140 | { |
wolfSSL | 15:117db924cf7c | 141 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 142 | |
wolfSSL | 15:117db924cf7c | 143 | if (sha256 == NULL) |
wolfSSL | 15:117db924cf7c | 144 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 145 | |
wolfSSL | 15:117db924cf7c | 146 | XMEMSET(sha256->digest, 0, sizeof(sha256->digest)); |
wolfSSL | 15:117db924cf7c | 147 | sha256->digest[0] = 0x6A09E667L; |
wolfSSL | 15:117db924cf7c | 148 | sha256->digest[1] = 0xBB67AE85L; |
wolfSSL | 15:117db924cf7c | 149 | sha256->digest[2] = 0x3C6EF372L; |
wolfSSL | 15:117db924cf7c | 150 | sha256->digest[3] = 0xA54FF53AL; |
wolfSSL | 15:117db924cf7c | 151 | sha256->digest[4] = 0x510E527FL; |
wolfSSL | 15:117db924cf7c | 152 | sha256->digest[5] = 0x9B05688CL; |
wolfSSL | 15:117db924cf7c | 153 | sha256->digest[6] = 0x1F83D9ABL; |
wolfSSL | 15:117db924cf7c | 154 | sha256->digest[7] = 0x5BE0CD19L; |
wolfSSL | 15:117db924cf7c | 155 | |
wolfSSL | 15:117db924cf7c | 156 | sha256->buffLen = 0; |
wolfSSL | 15:117db924cf7c | 157 | sha256->loLen = 0; |
wolfSSL | 15:117db924cf7c | 158 | sha256->hiLen = 0; |
wolfSSL | 15:117db924cf7c | 159 | |
wolfSSL | 15:117db924cf7c | 160 | return ret; |
wolfSSL | 15:117db924cf7c | 161 | } |
wolfSSL | 15:117db924cf7c | 162 | #endif |
wolfSSL | 15:117db924cf7c | 163 | |
wolfSSL | 15:117db924cf7c | 164 | |
wolfSSL | 15:117db924cf7c | 165 | /* Hardware Acceleration */ |
wolfSSL | 15:117db924cf7c | 166 | #if defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 167 | |
wolfSSL | 15:117db924cf7c | 168 | /* in case intel instructions aren't available, plus we need the K[] global */ |
wolfSSL | 15:117db924cf7c | 169 | #define NEED_SOFT_SHA256 |
wolfSSL | 15:117db924cf7c | 170 | |
wolfSSL | 15:117db924cf7c | 171 | /***** |
wolfSSL | 15:117db924cf7c | 172 | Intel AVX1/AVX2 Macro Control Structure |
wolfSSL | 15:117db924cf7c | 173 | |
wolfSSL | 15:117db924cf7c | 174 | #define HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 175 | #define HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 176 | |
wolfSSL | 15:117db924cf7c | 177 | #define HAVE_INTEL_RORX |
wolfSSL | 15:117db924cf7c | 178 | |
wolfSSL | 15:117db924cf7c | 179 | |
wolfSSL | 15:117db924cf7c | 180 | int InitSha256(wc_Sha256* sha256) { |
wolfSSL | 15:117db924cf7c | 181 | Save/Recover XMM, YMM |
wolfSSL | 15:117db924cf7c | 182 | ... |
wolfSSL | 15:117db924cf7c | 183 | } |
wolfSSL | 15:117db924cf7c | 184 | |
wolfSSL | 15:117db924cf7c | 185 | #if defined(HAVE_INTEL_AVX1)|| defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 186 | Transform_Sha256(); Function prototype |
wolfSSL | 15:117db924cf7c | 187 | #else |
wolfSSL | 15:117db924cf7c | 188 | Transform_Sha256() { } |
wolfSSL | 15:117db924cf7c | 189 | int Sha256Final() { |
wolfSSL | 15:117db924cf7c | 190 | Save/Recover XMM, YMM |
wolfSSL | 15:117db924cf7c | 191 | ... |
wolfSSL | 15:117db924cf7c | 192 | } |
wolfSSL | 15:117db924cf7c | 193 | #endif |
wolfSSL | 15:117db924cf7c | 194 | |
wolfSSL | 15:117db924cf7c | 195 | #if defined(HAVE_INTEL_AVX1)|| defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 196 | #if defined(HAVE_INTEL_RORX |
wolfSSL | 15:117db924cf7c | 197 | #define RND with rorx instuction |
wolfSSL | 15:117db924cf7c | 198 | #else |
wolfSSL | 15:117db924cf7c | 199 | #define RND |
wolfSSL | 15:117db924cf7c | 200 | #endif |
wolfSSL | 15:117db924cf7c | 201 | #endif |
wolfSSL | 15:117db924cf7c | 202 | |
wolfSSL | 15:117db924cf7c | 203 | #if defined(HAVE_INTEL_AVX1) |
wolfSSL | 15:117db924cf7c | 204 | |
wolfSSL | 15:117db924cf7c | 205 | #define XMM Instructions/inline asm |
wolfSSL | 15:117db924cf7c | 206 | |
wolfSSL | 15:117db924cf7c | 207 | int Transform_Sha256() { |
wolfSSL | 15:117db924cf7c | 208 | Stitched Message Sched/Round |
wolfSSL | 15:117db924cf7c | 209 | } |
wolfSSL | 15:117db924cf7c | 210 | |
wolfSSL | 15:117db924cf7c | 211 | #elif defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 212 | |
wolfSSL | 15:117db924cf7c | 213 | #define YMM Instructions/inline asm |
wolfSSL | 15:117db924cf7c | 214 | |
wolfSSL | 15:117db924cf7c | 215 | int Transform_Sha256() { |
wolfSSL | 15:117db924cf7c | 216 | More granural Stitched Message Sched/Round |
wolfSSL | 15:117db924cf7c | 217 | } |
wolfSSL | 15:117db924cf7c | 218 | |
wolfSSL | 15:117db924cf7c | 219 | #endif |
wolfSSL | 15:117db924cf7c | 220 | |
wolfSSL | 15:117db924cf7c | 221 | */ |
wolfSSL | 15:117db924cf7c | 222 | |
wolfSSL | 15:117db924cf7c | 223 | /* Each platform needs to query info type 1 from cpuid to see if aesni is |
wolfSSL | 15:117db924cf7c | 224 | * supported. Also, let's setup a macro for proper linkage w/o ABI conflicts |
wolfSSL | 15:117db924cf7c | 225 | */ |
wolfSSL | 15:117db924cf7c | 226 | |
wolfSSL | 15:117db924cf7c | 227 | /* #if defined(HAVE_INTEL_AVX1/2) at the tail of sha256 */ |
wolfSSL | 15:117db924cf7c | 228 | static int Transform_Sha256(wc_Sha256* sha256); |
wolfSSL | 15:117db924cf7c | 229 | #if defined(HAVE_INTEL_AVX1) |
wolfSSL | 15:117db924cf7c | 230 | static int Transform_Sha256_AVX1(wc_Sha256 *sha256); |
wolfSSL | 15:117db924cf7c | 231 | static int Transform_Sha256_AVX1_Len(wc_Sha256* sha256, word32 len); |
wolfSSL | 15:117db924cf7c | 232 | #endif |
wolfSSL | 15:117db924cf7c | 233 | #if defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 234 | static int Transform_Sha256_AVX2(wc_Sha256 *sha256); |
wolfSSL | 15:117db924cf7c | 235 | static int Transform_Sha256_AVX2_Len(wc_Sha256* sha256, word32 len); |
wolfSSL | 15:117db924cf7c | 236 | #ifdef HAVE_INTEL_RORX |
wolfSSL | 15:117db924cf7c | 237 | static int Transform_Sha256_AVX1_RORX(wc_Sha256 *sha256); |
wolfSSL | 15:117db924cf7c | 238 | static int Transform_Sha256_AVX1_RORX_Len(wc_Sha256* sha256, word32 len); |
wolfSSL | 15:117db924cf7c | 239 | static int Transform_Sha256_AVX2_RORX(wc_Sha256 *sha256); |
wolfSSL | 15:117db924cf7c | 240 | static int Transform_Sha256_AVX2_RORX_Len(wc_Sha256* sha256, word32 len); |
wolfSSL | 15:117db924cf7c | 241 | #endif |
wolfSSL | 15:117db924cf7c | 242 | #endif |
wolfSSL | 15:117db924cf7c | 243 | static int (*Transform_Sha256_p)(wc_Sha256* sha256); |
wolfSSL | 15:117db924cf7c | 244 | /* = _Transform_Sha256 */ |
wolfSSL | 15:117db924cf7c | 245 | static int (*Transform_Sha256_Len_p)(wc_Sha256* sha256, word32 len); |
wolfSSL | 15:117db924cf7c | 246 | /* = NULL */ |
wolfSSL | 15:117db924cf7c | 247 | static int transform_check = 0; |
wolfSSL | 15:117db924cf7c | 248 | static word32 intel_flags; |
wolfSSL | 15:117db924cf7c | 249 | #define XTRANSFORM(S) (*Transform_Sha256_p)((S)) |
wolfSSL | 15:117db924cf7c | 250 | #define XTRANSFORM_LEN(S, L) (*Transform_Sha256_Len_p)((S),(L)) |
wolfSSL | 15:117db924cf7c | 251 | |
wolfSSL | 15:117db924cf7c | 252 | static void Sha256_SetTransform(void) |
wolfSSL | 15:117db924cf7c | 253 | { |
wolfSSL | 15:117db924cf7c | 254 | |
wolfSSL | 15:117db924cf7c | 255 | if (transform_check) |
wolfSSL | 15:117db924cf7c | 256 | return; |
wolfSSL | 15:117db924cf7c | 257 | |
wolfSSL | 15:117db924cf7c | 258 | intel_flags = cpuid_get_flags(); |
wolfSSL | 15:117db924cf7c | 259 | |
wolfSSL | 15:117db924cf7c | 260 | #ifdef HAVE_INTEL_AVX2 |
wolfSSL | 15:117db924cf7c | 261 | if (IS_INTEL_AVX2(intel_flags)) { |
wolfSSL | 15:117db924cf7c | 262 | #ifdef HAVE_INTEL_RORX |
wolfSSL | 15:117db924cf7c | 263 | if (IS_INTEL_BMI2(intel_flags)) { |
wolfSSL | 15:117db924cf7c | 264 | Transform_Sha256_p = Transform_Sha256_AVX2_RORX; |
wolfSSL | 15:117db924cf7c | 265 | Transform_Sha256_Len_p = Transform_Sha256_AVX2_RORX_Len; |
wolfSSL | 15:117db924cf7c | 266 | } |
wolfSSL | 15:117db924cf7c | 267 | else |
wolfSSL | 15:117db924cf7c | 268 | #endif |
wolfSSL | 15:117db924cf7c | 269 | if (1) |
wolfSSL | 15:117db924cf7c | 270 | { |
wolfSSL | 15:117db924cf7c | 271 | Transform_Sha256_p = Transform_Sha256_AVX2; |
wolfSSL | 15:117db924cf7c | 272 | Transform_Sha256_Len_p = Transform_Sha256_AVX2_Len; |
wolfSSL | 15:117db924cf7c | 273 | } |
wolfSSL | 15:117db924cf7c | 274 | #ifdef HAVE_INTEL_RORX |
wolfSSL | 15:117db924cf7c | 275 | else { |
wolfSSL | 15:117db924cf7c | 276 | Transform_Sha256_p = Transform_Sha256_AVX1_RORX; |
wolfSSL | 15:117db924cf7c | 277 | Transform_Sha256_Len_p = Transform_Sha256_AVX1_RORX_Len; |
wolfSSL | 15:117db924cf7c | 278 | } |
wolfSSL | 15:117db924cf7c | 279 | #endif |
wolfSSL | 15:117db924cf7c | 280 | } |
wolfSSL | 15:117db924cf7c | 281 | else |
wolfSSL | 15:117db924cf7c | 282 | #endif |
wolfSSL | 15:117db924cf7c | 283 | #ifdef HAVE_INTEL_AVX1 |
wolfSSL | 15:117db924cf7c | 284 | if (IS_INTEL_AVX1(intel_flags)) { |
wolfSSL | 15:117db924cf7c | 285 | Transform_Sha256_p = Transform_Sha256_AVX1; |
wolfSSL | 15:117db924cf7c | 286 | Transform_Sha256_Len_p = Transform_Sha256_AVX1_Len; |
wolfSSL | 15:117db924cf7c | 287 | } |
wolfSSL | 15:117db924cf7c | 288 | else |
wolfSSL | 15:117db924cf7c | 289 | #endif |
wolfSSL | 15:117db924cf7c | 290 | { |
wolfSSL | 15:117db924cf7c | 291 | Transform_Sha256_p = Transform_Sha256; |
wolfSSL | 15:117db924cf7c | 292 | Transform_Sha256_Len_p = NULL; |
wolfSSL | 15:117db924cf7c | 293 | } |
wolfSSL | 15:117db924cf7c | 294 | |
wolfSSL | 15:117db924cf7c | 295 | transform_check = 1; |
wolfSSL | 15:117db924cf7c | 296 | } |
wolfSSL | 15:117db924cf7c | 297 | |
wolfSSL | 15:117db924cf7c | 298 | int wc_InitSha256_ex(wc_Sha256* sha256, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 299 | { |
wolfSSL | 15:117db924cf7c | 300 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 301 | if (sha256 == NULL) |
wolfSSL | 15:117db924cf7c | 302 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 303 | |
wolfSSL | 15:117db924cf7c | 304 | sha256->heap = heap; |
wolfSSL | 15:117db924cf7c | 305 | |
wolfSSL | 15:117db924cf7c | 306 | ret = InitSha256(sha256); |
wolfSSL | 15:117db924cf7c | 307 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 308 | return ret; |
wolfSSL | 15:117db924cf7c | 309 | |
wolfSSL | 15:117db924cf7c | 310 | /* choose best Transform function under this runtime environment */ |
wolfSSL | 15:117db924cf7c | 311 | Sha256_SetTransform(); |
wolfSSL | 15:117db924cf7c | 312 | |
wolfSSL | 15:117db924cf7c | 313 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA256) |
wolfSSL | 15:117db924cf7c | 314 | ret = wolfAsync_DevCtxInit(&sha256->asyncDev, |
wolfSSL | 15:117db924cf7c | 315 | WOLFSSL_ASYNC_MARKER_SHA256, sha256->heap, devId); |
wolfSSL | 15:117db924cf7c | 316 | #else |
wolfSSL | 15:117db924cf7c | 317 | (void)devId; |
wolfSSL | 15:117db924cf7c | 318 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 319 | |
wolfSSL | 15:117db924cf7c | 320 | return ret; |
wolfSSL | 15:117db924cf7c | 321 | } |
wolfSSL | 15:117db924cf7c | 322 | |
wolfSSL | 15:117db924cf7c | 323 | #elif defined(FREESCALE_LTC_SHA) |
wolfSSL | 15:117db924cf7c | 324 | int wc_InitSha256_ex(wc_Sha256* sha256, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 325 | { |
wolfSSL | 15:117db924cf7c | 326 | (void)heap; |
wolfSSL | 15:117db924cf7c | 327 | (void)devId; |
wolfSSL | 15:117db924cf7c | 328 | |
wolfSSL | 15:117db924cf7c | 329 | LTC_HASH_Init(LTC_BASE, &sha256->ctx, kLTC_Sha256, NULL, 0); |
wolfSSL | 15:117db924cf7c | 330 | |
wolfSSL | 15:117db924cf7c | 331 | return 0; |
wolfSSL | 15:117db924cf7c | 332 | } |
wolfSSL | 15:117db924cf7c | 333 | |
wolfSSL | 15:117db924cf7c | 334 | #elif defined(FREESCALE_MMCAU_SHA) |
wolfSSL | 15:117db924cf7c | 335 | |
wolfSSL | 15:117db924cf7c | 336 | #ifdef FREESCALE_MMCAU_CLASSIC_SHA |
wolfSSL | 15:117db924cf7c | 337 | #include "cau_api.h" |
wolfSSL | 15:117db924cf7c | 338 | #else |
wolfSSL | 15:117db924cf7c | 339 | #include "fsl_mmcau.h" |
wolfSSL | 15:117db924cf7c | 340 | #endif |
wolfSSL | 15:117db924cf7c | 341 | |
wolfSSL | 15:117db924cf7c | 342 | #define XTRANSFORM(S) Transform_Sha256((S)) |
wolfSSL | 15:117db924cf7c | 343 | #define XTRANSFORM_LEN(S,L) Transform_Sha256_Len((S),(L)) |
wolfSSL | 15:117db924cf7c | 344 | |
wolfSSL | 15:117db924cf7c | 345 | int wc_InitSha256_ex(wc_Sha256* sha256, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 346 | { |
wolfSSL | 15:117db924cf7c | 347 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 348 | |
wolfSSL | 15:117db924cf7c | 349 | (void)heap; |
wolfSSL | 15:117db924cf7c | 350 | (void)devId; |
wolfSSL | 15:117db924cf7c | 351 | |
wolfSSL | 15:117db924cf7c | 352 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 353 | if (ret != 0) { |
wolfSSL | 15:117db924cf7c | 354 | return ret; |
wolfSSL | 15:117db924cf7c | 355 | } |
wolfSSL | 15:117db924cf7c | 356 | #ifdef FREESCALE_MMCAU_CLASSIC_SHA |
wolfSSL | 15:117db924cf7c | 357 | cau_sha256_initialize_output(sha256->digest); |
wolfSSL | 15:117db924cf7c | 358 | #else |
wolfSSL | 15:117db924cf7c | 359 | MMCAU_SHA256_InitializeOutput((uint32_t*)sha256->digest); |
wolfSSL | 15:117db924cf7c | 360 | #endif |
wolfSSL | 15:117db924cf7c | 361 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 362 | |
wolfSSL | 15:117db924cf7c | 363 | sha256->buffLen = 0; |
wolfSSL | 15:117db924cf7c | 364 | sha256->loLen = 0; |
wolfSSL | 15:117db924cf7c | 365 | sha256->hiLen = 0; |
wolfSSL | 15:117db924cf7c | 366 | |
wolfSSL | 15:117db924cf7c | 367 | return ret; |
wolfSSL | 15:117db924cf7c | 368 | } |
wolfSSL | 15:117db924cf7c | 369 | |
wolfSSL | 15:117db924cf7c | 370 | static int Transform_Sha256(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 371 | { |
wolfSSL | 15:117db924cf7c | 372 | int ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 373 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 374 | #ifdef FREESCALE_MMCAU_CLASSIC_SHA |
wolfSSL | 15:117db924cf7c | 375 | cau_sha256_hash_n((byte*)sha256->buffer, 1, sha256->digest); |
wolfSSL | 15:117db924cf7c | 376 | #else |
wolfSSL | 15:117db924cf7c | 377 | MMCAU_SHA256_HashN((byte*)sha256->buffer, 1, sha256->digest); |
wolfSSL | 15:117db924cf7c | 378 | #endif |
wolfSSL | 15:117db924cf7c | 379 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 380 | } |
wolfSSL | 15:117db924cf7c | 381 | return ret; |
wolfSSL | 15:117db924cf7c | 382 | } |
wolfSSL | 15:117db924cf7c | 383 | |
wolfSSL | 15:117db924cf7c | 384 | #elif defined(WOLFSSL_PIC32MZ_HASH) |
wolfSSL | 15:117db924cf7c | 385 | #include <wolfssl/wolfcrypt/port/pic32/pic32mz-crypt.h> |
wolfSSL | 15:117db924cf7c | 386 | |
wolfSSL | 15:117db924cf7c | 387 | #elif defined(STM32_HASH_SHA2) |
wolfSSL | 15:117db924cf7c | 388 | |
wolfSSL | 15:117db924cf7c | 389 | /* Supports CubeMX HAL or Standard Peripheral Library */ |
wolfSSL | 15:117db924cf7c | 390 | |
wolfSSL | 15:117db924cf7c | 391 | int wc_InitSha256_ex(wc_Sha256* sha256, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 392 | { |
wolfSSL | 15:117db924cf7c | 393 | if (sha256 == NULL) |
wolfSSL | 15:117db924cf7c | 394 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 395 | |
wolfSSL | 15:117db924cf7c | 396 | (void)devId; |
wolfSSL | 15:117db924cf7c | 397 | (void)heap; |
wolfSSL | 15:117db924cf7c | 398 | |
wolfSSL | 15:117db924cf7c | 399 | wc_Stm32_Hash_Init(&sha256->stmCtx); |
wolfSSL | 15:117db924cf7c | 400 | return 0; |
wolfSSL | 15:117db924cf7c | 401 | } |
wolfSSL | 15:117db924cf7c | 402 | |
wolfSSL | 15:117db924cf7c | 403 | int wc_Sha256Update(wc_Sha256* sha256, const byte* data, word32 len) |
wolfSSL | 15:117db924cf7c | 404 | { |
wolfSSL | 15:117db924cf7c | 405 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 406 | |
wolfSSL | 15:117db924cf7c | 407 | if (sha256 == NULL || (data == NULL && len > 0)) { |
wolfSSL | 15:117db924cf7c | 408 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 409 | } |
wolfSSL | 15:117db924cf7c | 410 | |
wolfSSL | 15:117db924cf7c | 411 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 412 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 413 | ret = wc_Stm32_Hash_Update(&sha256->stmCtx, |
wolfSSL | 15:117db924cf7c | 414 | HASH_AlgoSelection_SHA256, data, len); |
wolfSSL | 15:117db924cf7c | 415 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 416 | } |
wolfSSL | 15:117db924cf7c | 417 | return ret; |
wolfSSL | 15:117db924cf7c | 418 | } |
wolfSSL | 15:117db924cf7c | 419 | |
wolfSSL | 15:117db924cf7c | 420 | int wc_Sha256Final(wc_Sha256* sha256, byte* hash) |
wolfSSL | 15:117db924cf7c | 421 | { |
wolfSSL | 15:117db924cf7c | 422 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 423 | |
wolfSSL | 15:117db924cf7c | 424 | if (sha256 == NULL || hash == NULL) { |
wolfSSL | 15:117db924cf7c | 425 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 426 | } |
wolfSSL | 15:117db924cf7c | 427 | |
wolfSSL | 15:117db924cf7c | 428 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 429 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 430 | ret = wc_Stm32_Hash_Final(&sha256->stmCtx, |
wolfSSL | 15:117db924cf7c | 431 | HASH_AlgoSelection_SHA256, hash, WC_SHA256_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 432 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 433 | } |
wolfSSL | 15:117db924cf7c | 434 | |
wolfSSL | 15:117db924cf7c | 435 | (void)wc_InitSha256(sha256); /* reset state */ |
wolfSSL | 15:117db924cf7c | 436 | |
wolfSSL | 15:117db924cf7c | 437 | return ret; |
wolfSSL | 15:117db924cf7c | 438 | } |
wolfSSL | 15:117db924cf7c | 439 | |
wolfSSL | 15:117db924cf7c | 440 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_HASH) |
wolfSSL | 15:117db924cf7c | 441 | /* functions defined in wolfcrypt/src/port/caam/caam_sha256.c */ |
wolfSSL | 15:117db924cf7c | 442 | #else |
wolfSSL | 15:117db924cf7c | 443 | #define NEED_SOFT_SHA256 |
wolfSSL | 15:117db924cf7c | 444 | |
wolfSSL | 15:117db924cf7c | 445 | int wc_InitSha256_ex(wc_Sha256* sha256, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 446 | { |
wolfSSL | 15:117db924cf7c | 447 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 448 | if (sha256 == NULL) |
wolfSSL | 15:117db924cf7c | 449 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 450 | |
wolfSSL | 15:117db924cf7c | 451 | sha256->heap = heap; |
wolfSSL | 15:117db924cf7c | 452 | |
wolfSSL | 15:117db924cf7c | 453 | ret = InitSha256(sha256); |
wolfSSL | 15:117db924cf7c | 454 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 455 | return ret; |
wolfSSL | 15:117db924cf7c | 456 | |
wolfSSL | 15:117db924cf7c | 457 | #ifdef WOLFSSL_SMALL_STACK_CACHE |
wolfSSL | 15:117db924cf7c | 458 | sha256->W = NULL; |
wolfSSL | 15:117db924cf7c | 459 | #endif |
wolfSSL | 15:117db924cf7c | 460 | |
wolfSSL | 15:117db924cf7c | 461 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA256) |
wolfSSL | 15:117db924cf7c | 462 | ret = wolfAsync_DevCtxInit(&sha256->asyncDev, |
wolfSSL | 15:117db924cf7c | 463 | WOLFSSL_ASYNC_MARKER_SHA256, sha256->heap, devId); |
wolfSSL | 15:117db924cf7c | 464 | #else |
wolfSSL | 15:117db924cf7c | 465 | (void)devId; |
wolfSSL | 15:117db924cf7c | 466 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 467 | |
wolfSSL | 15:117db924cf7c | 468 | return ret; |
wolfSSL | 15:117db924cf7c | 469 | } |
wolfSSL | 15:117db924cf7c | 470 | #endif /* End Hardware Acceleration */ |
wolfSSL | 15:117db924cf7c | 471 | |
wolfSSL | 15:117db924cf7c | 472 | #ifdef NEED_SOFT_SHA256 |
wolfSSL | 15:117db924cf7c | 473 | |
wolfSSL | 15:117db924cf7c | 474 | static const ALIGN32 word32 K[64] = { |
wolfSSL | 15:117db924cf7c | 475 | 0x428A2F98L, 0x71374491L, 0xB5C0FBCFL, 0xE9B5DBA5L, 0x3956C25BL, |
wolfSSL | 15:117db924cf7c | 476 | 0x59F111F1L, 0x923F82A4L, 0xAB1C5ED5L, 0xD807AA98L, 0x12835B01L, |
wolfSSL | 15:117db924cf7c | 477 | 0x243185BEL, 0x550C7DC3L, 0x72BE5D74L, 0x80DEB1FEL, 0x9BDC06A7L, |
wolfSSL | 15:117db924cf7c | 478 | 0xC19BF174L, 0xE49B69C1L, 0xEFBE4786L, 0x0FC19DC6L, 0x240CA1CCL, |
wolfSSL | 15:117db924cf7c | 479 | 0x2DE92C6FL, 0x4A7484AAL, 0x5CB0A9DCL, 0x76F988DAL, 0x983E5152L, |
wolfSSL | 15:117db924cf7c | 480 | 0xA831C66DL, 0xB00327C8L, 0xBF597FC7L, 0xC6E00BF3L, 0xD5A79147L, |
wolfSSL | 15:117db924cf7c | 481 | 0x06CA6351L, 0x14292967L, 0x27B70A85L, 0x2E1B2138L, 0x4D2C6DFCL, |
wolfSSL | 15:117db924cf7c | 482 | 0x53380D13L, 0x650A7354L, 0x766A0ABBL, 0x81C2C92EL, 0x92722C85L, |
wolfSSL | 15:117db924cf7c | 483 | 0xA2BFE8A1L, 0xA81A664BL, 0xC24B8B70L, 0xC76C51A3L, 0xD192E819L, |
wolfSSL | 15:117db924cf7c | 484 | 0xD6990624L, 0xF40E3585L, 0x106AA070L, 0x19A4C116L, 0x1E376C08L, |
wolfSSL | 15:117db924cf7c | 485 | 0x2748774CL, 0x34B0BCB5L, 0x391C0CB3L, 0x4ED8AA4AL, 0x5B9CCA4FL, |
wolfSSL | 15:117db924cf7c | 486 | 0x682E6FF3L, 0x748F82EEL, 0x78A5636FL, 0x84C87814L, 0x8CC70208L, |
wolfSSL | 15:117db924cf7c | 487 | 0x90BEFFFAL, 0xA4506CEBL, 0xBEF9A3F7L, 0xC67178F2L |
wolfSSL | 15:117db924cf7c | 488 | }; |
wolfSSL | 15:117db924cf7c | 489 | |
wolfSSL | 15:117db924cf7c | 490 | #define Ch(x,y,z) ((z) ^ ((x) & ((y) ^ (z)))) |
wolfSSL | 15:117db924cf7c | 491 | #define Maj(x,y,z) ((((x) | (y)) & (z)) | ((x) & (y))) |
wolfSSL | 15:117db924cf7c | 492 | #define R(x, n) (((x) & 0xFFFFFFFFU) >> (n)) |
wolfSSL | 15:117db924cf7c | 493 | |
wolfSSL | 15:117db924cf7c | 494 | #define S(x, n) rotrFixed(x, n) |
wolfSSL | 15:117db924cf7c | 495 | #define Sigma0(x) (S(x, 2) ^ S(x, 13) ^ S(x, 22)) |
wolfSSL | 15:117db924cf7c | 496 | #define Sigma1(x) (S(x, 6) ^ S(x, 11) ^ S(x, 25)) |
wolfSSL | 15:117db924cf7c | 497 | #define Gamma0(x) (S(x, 7) ^ S(x, 18) ^ R(x, 3)) |
wolfSSL | 15:117db924cf7c | 498 | #define Gamma1(x) (S(x, 17) ^ S(x, 19) ^ R(x, 10)) |
wolfSSL | 15:117db924cf7c | 499 | |
wolfSSL | 15:117db924cf7c | 500 | #define a(i) S[(0-i) & 7] |
wolfSSL | 15:117db924cf7c | 501 | #define b(i) S[(1-i) & 7] |
wolfSSL | 15:117db924cf7c | 502 | #define c(i) S[(2-i) & 7] |
wolfSSL | 15:117db924cf7c | 503 | #define d(i) S[(3-i) & 7] |
wolfSSL | 15:117db924cf7c | 504 | #define e(i) S[(4-i) & 7] |
wolfSSL | 15:117db924cf7c | 505 | #define f(i) S[(5-i) & 7] |
wolfSSL | 15:117db924cf7c | 506 | #define g(i) S[(6-i) & 7] |
wolfSSL | 15:117db924cf7c | 507 | #define h(i) S[(7-i) & 7] |
wolfSSL | 15:117db924cf7c | 508 | |
wolfSSL | 15:117db924cf7c | 509 | #define RND(j) \ |
wolfSSL | 15:117db924cf7c | 510 | t0 = h(j) + Sigma1(e(j)) + Ch(e(j), f(j), g(j)) + K[i+j] + W[i+j]; \ |
wolfSSL | 15:117db924cf7c | 511 | t1 = Sigma0(a(j)) + Maj(a(j), b(j), c(j)); \ |
wolfSSL | 15:117db924cf7c | 512 | d(j) += t0; \ |
wolfSSL | 15:117db924cf7c | 513 | h(j) = t0 + t1 |
wolfSSL | 15:117db924cf7c | 514 | |
wolfSSL | 15:117db924cf7c | 515 | #ifndef XTRANSFORM |
wolfSSL | 15:117db924cf7c | 516 | #define XTRANSFORM(S) Transform_Sha256((S)) |
wolfSSL | 15:117db924cf7c | 517 | #define XTRANSFORM_LEN(S,L) Transform_Sha256_Len((S),(L)) |
wolfSSL | 15:117db924cf7c | 518 | #endif |
wolfSSL | 15:117db924cf7c | 519 | |
wolfSSL | 15:117db924cf7c | 520 | static int Transform_Sha256(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 521 | { |
wolfSSL | 15:117db924cf7c | 522 | word32 S[8], t0, t1; |
wolfSSL | 15:117db924cf7c | 523 | int i; |
wolfSSL | 15:117db924cf7c | 524 | |
wolfSSL | 15:117db924cf7c | 525 | #ifdef WOLFSSL_SMALL_STACK_CACHE |
wolfSSL | 15:117db924cf7c | 526 | word32* W = sha256->W; |
wolfSSL | 15:117db924cf7c | 527 | if (W == NULL) { |
wolfSSL | 15:117db924cf7c | 528 | W = (word32*)XMALLOC(sizeof(word32) * WC_SHA256_BLOCK_SIZE, NULL, |
wolfSSL | 15:117db924cf7c | 529 | DYNAMIC_TYPE_RNG); |
wolfSSL | 15:117db924cf7c | 530 | if (W == NULL) |
wolfSSL | 15:117db924cf7c | 531 | return MEMORY_E; |
wolfSSL | 15:117db924cf7c | 532 | sha256->W = W; |
wolfSSL | 15:117db924cf7c | 533 | } |
wolfSSL | 15:117db924cf7c | 534 | #elif defined(WOLFSSL_SMALL_STACK) |
wolfSSL | 15:117db924cf7c | 535 | word32* W; |
wolfSSL | 15:117db924cf7c | 536 | W = (word32*)XMALLOC(sizeof(word32) * WC_SHA256_BLOCK_SIZE, NULL, |
wolfSSL | 15:117db924cf7c | 537 | DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 538 | if (W == NULL) |
wolfSSL | 15:117db924cf7c | 539 | return MEMORY_E; |
wolfSSL | 15:117db924cf7c | 540 | #else |
wolfSSL | 15:117db924cf7c | 541 | word32 W[WC_SHA256_BLOCK_SIZE]; |
wolfSSL | 15:117db924cf7c | 542 | #endif |
wolfSSL | 15:117db924cf7c | 543 | |
wolfSSL | 15:117db924cf7c | 544 | /* Copy context->state[] to working vars */ |
wolfSSL | 15:117db924cf7c | 545 | for (i = 0; i < 8; i++) |
wolfSSL | 15:117db924cf7c | 546 | S[i] = sha256->digest[i]; |
wolfSSL | 15:117db924cf7c | 547 | |
wolfSSL | 15:117db924cf7c | 548 | for (i = 0; i < 16; i++) |
wolfSSL | 15:117db924cf7c | 549 | W[i] = sha256->buffer[i]; |
wolfSSL | 15:117db924cf7c | 550 | |
wolfSSL | 15:117db924cf7c | 551 | for (i = 16; i < WC_SHA256_BLOCK_SIZE; i++) |
wolfSSL | 15:117db924cf7c | 552 | W[i] = Gamma1(W[i-2]) + W[i-7] + Gamma0(W[i-15]) + W[i-16]; |
wolfSSL | 15:117db924cf7c | 553 | |
wolfSSL | 15:117db924cf7c | 554 | #ifdef USE_SLOW_SHA256 |
wolfSSL | 15:117db924cf7c | 555 | /* not unrolled - ~2k smaller and ~25% slower */ |
wolfSSL | 15:117db924cf7c | 556 | for (i = 0; i < WC_SHA256_BLOCK_SIZE; i += 8) { |
wolfSSL | 15:117db924cf7c | 557 | int j; |
wolfSSL | 15:117db924cf7c | 558 | for (j = 0; j < 8; j++) { /* braces needed here for macros {} */ |
wolfSSL | 15:117db924cf7c | 559 | RND(j); |
wolfSSL | 15:117db924cf7c | 560 | } |
wolfSSL | 15:117db924cf7c | 561 | } |
wolfSSL | 15:117db924cf7c | 562 | #else |
wolfSSL | 15:117db924cf7c | 563 | /* partially loop unrolled */ |
wolfSSL | 15:117db924cf7c | 564 | for (i = 0; i < WC_SHA256_BLOCK_SIZE; i += 8) { |
wolfSSL | 15:117db924cf7c | 565 | RND(0); RND(1); RND(2); RND(3); |
wolfSSL | 15:117db924cf7c | 566 | RND(4); RND(5); RND(6); RND(7); |
wolfSSL | 15:117db924cf7c | 567 | } |
wolfSSL | 15:117db924cf7c | 568 | #endif /* USE_SLOW_SHA256 */ |
wolfSSL | 15:117db924cf7c | 569 | |
wolfSSL | 15:117db924cf7c | 570 | /* Add the working vars back into digest state[] */ |
wolfSSL | 15:117db924cf7c | 571 | for (i = 0; i < 8; i++) { |
wolfSSL | 15:117db924cf7c | 572 | sha256->digest[i] += S[i]; |
wolfSSL | 15:117db924cf7c | 573 | } |
wolfSSL | 15:117db924cf7c | 574 | |
wolfSSL | 15:117db924cf7c | 575 | #if defined(WOLFSSL_SMALL_STACK) && !defined(WOLFSSL_SMALL_STACK_CACHE) |
wolfSSL | 15:117db924cf7c | 576 | XFREE(W, NULL, DYNAMIC_TYPE_TMP_BUFFER); |
wolfSSL | 15:117db924cf7c | 577 | #endif |
wolfSSL | 15:117db924cf7c | 578 | return 0; |
wolfSSL | 15:117db924cf7c | 579 | } |
wolfSSL | 15:117db924cf7c | 580 | #endif |
wolfSSL | 15:117db924cf7c | 581 | /* End wc_ software implementation */ |
wolfSSL | 15:117db924cf7c | 582 | |
wolfSSL | 15:117db924cf7c | 583 | |
wolfSSL | 15:117db924cf7c | 584 | #ifdef XTRANSFORM |
wolfSSL | 15:117db924cf7c | 585 | |
wolfSSL | 15:117db924cf7c | 586 | static WC_INLINE void AddLength(wc_Sha256* sha256, word32 len) |
wolfSSL | 15:117db924cf7c | 587 | { |
wolfSSL | 15:117db924cf7c | 588 | word32 tmp = sha256->loLen; |
wolfSSL | 15:117db924cf7c | 589 | if ((sha256->loLen += len) < tmp) |
wolfSSL | 15:117db924cf7c | 590 | sha256->hiLen++; /* carry low to high */ |
wolfSSL | 15:117db924cf7c | 591 | } |
wolfSSL | 15:117db924cf7c | 592 | |
wolfSSL | 15:117db924cf7c | 593 | static WC_INLINE int Sha256Update(wc_Sha256* sha256, const byte* data, word32 len) |
wolfSSL | 15:117db924cf7c | 594 | { |
wolfSSL | 15:117db924cf7c | 595 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 596 | byte* local; |
wolfSSL | 15:117db924cf7c | 597 | |
wolfSSL | 15:117db924cf7c | 598 | if (sha256 == NULL || (data == NULL && len > 0)) { |
wolfSSL | 15:117db924cf7c | 599 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 600 | } |
wolfSSL | 15:117db924cf7c | 601 | |
wolfSSL | 15:117db924cf7c | 602 | if (data == NULL && len == 0) { |
wolfSSL | 15:117db924cf7c | 603 | /* valid, but do nothing */ |
wolfSSL | 15:117db924cf7c | 604 | return 0; |
wolfSSL | 15:117db924cf7c | 605 | } |
wolfSSL | 15:117db924cf7c | 606 | |
wolfSSL | 15:117db924cf7c | 607 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA256) |
wolfSSL | 15:117db924cf7c | 608 | if (sha256->asyncDev.marker == WOLFSSL_ASYNC_MARKER_SHA256) { |
wolfSSL | 15:117db924cf7c | 609 | #if defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 610 | return IntelQaSymSha256(&sha256->asyncDev, NULL, data, len); |
wolfSSL | 15:117db924cf7c | 611 | #endif |
wolfSSL | 15:117db924cf7c | 612 | } |
wolfSSL | 15:117db924cf7c | 613 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 614 | |
wolfSSL | 15:117db924cf7c | 615 | /* do block size increments */ |
wolfSSL | 15:117db924cf7c | 616 | local = (byte*)sha256->buffer; |
wolfSSL | 15:117db924cf7c | 617 | |
wolfSSL | 15:117db924cf7c | 618 | /* check that internal buffLen is valid */ |
wolfSSL | 15:117db924cf7c | 619 | if (sha256->buffLen >= WC_SHA256_BLOCK_SIZE) |
wolfSSL | 15:117db924cf7c | 620 | return BUFFER_E; |
wolfSSL | 15:117db924cf7c | 621 | |
wolfSSL | 15:117db924cf7c | 622 | if (sha256->buffLen > 0) { |
wolfSSL | 15:117db924cf7c | 623 | word32 add = min(len, WC_SHA256_BLOCK_SIZE - sha256->buffLen); |
wolfSSL | 15:117db924cf7c | 624 | XMEMCPY(&local[sha256->buffLen], data, add); |
wolfSSL | 15:117db924cf7c | 625 | |
wolfSSL | 15:117db924cf7c | 626 | sha256->buffLen += add; |
wolfSSL | 15:117db924cf7c | 627 | data += add; |
wolfSSL | 15:117db924cf7c | 628 | len -= add; |
wolfSSL | 15:117db924cf7c | 629 | |
wolfSSL | 15:117db924cf7c | 630 | if (sha256->buffLen == WC_SHA256_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 631 | #if defined(LITTLE_ENDIAN_ORDER) && !defined(FREESCALE_MMCAU_SHA) |
wolfSSL | 15:117db924cf7c | 632 | #if defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 633 | if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags)) |
wolfSSL | 15:117db924cf7c | 634 | #endif |
wolfSSL | 15:117db924cf7c | 635 | { |
wolfSSL | 15:117db924cf7c | 636 | ByteReverseWords(sha256->buffer, sha256->buffer, |
wolfSSL | 15:117db924cf7c | 637 | WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 638 | } |
wolfSSL | 15:117db924cf7c | 639 | #endif |
wolfSSL | 15:117db924cf7c | 640 | ret = XTRANSFORM(sha256); |
wolfSSL | 15:117db924cf7c | 641 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 642 | AddLength(sha256, WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 643 | sha256->buffLen = 0; |
wolfSSL | 15:117db924cf7c | 644 | } |
wolfSSL | 15:117db924cf7c | 645 | else |
wolfSSL | 15:117db924cf7c | 646 | len = 0; |
wolfSSL | 15:117db924cf7c | 647 | } |
wolfSSL | 15:117db924cf7c | 648 | } |
wolfSSL | 15:117db924cf7c | 649 | |
wolfSSL | 15:117db924cf7c | 650 | #if defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 651 | if (Transform_Sha256_Len_p != NULL) { |
wolfSSL | 15:117db924cf7c | 652 | word32 blocksLen = len & ~(WC_SHA256_BLOCK_SIZE-1); |
wolfSSL | 15:117db924cf7c | 653 | |
wolfSSL | 15:117db924cf7c | 654 | if (blocksLen > 0) { |
wolfSSL | 15:117db924cf7c | 655 | AddLength(sha256, blocksLen); |
wolfSSL | 15:117db924cf7c | 656 | sha256->data = data; |
wolfSSL | 15:117db924cf7c | 657 | /* Byte reversal performed in function if required. */ |
wolfSSL | 15:117db924cf7c | 658 | XTRANSFORM_LEN(sha256, blocksLen); |
wolfSSL | 15:117db924cf7c | 659 | data += blocksLen; |
wolfSSL | 15:117db924cf7c | 660 | len -= blocksLen; |
wolfSSL | 15:117db924cf7c | 661 | } |
wolfSSL | 15:117db924cf7c | 662 | } |
wolfSSL | 15:117db924cf7c | 663 | else |
wolfSSL | 15:117db924cf7c | 664 | #endif |
wolfSSL | 15:117db924cf7c | 665 | #if !defined(LITTLE_ENDIAN_ORDER) || defined(FREESCALE_MMCAU_SHA) || \ |
wolfSSL | 15:117db924cf7c | 666 | defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 667 | { |
wolfSSL | 15:117db924cf7c | 668 | word32 blocksLen = len & ~(WC_SHA256_BLOCK_SIZE-1); |
wolfSSL | 15:117db924cf7c | 669 | |
wolfSSL | 15:117db924cf7c | 670 | AddLength(sha256, blocksLen); |
wolfSSL | 15:117db924cf7c | 671 | while (len >= WC_SHA256_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 672 | XMEMCPY(local, data, WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 673 | |
wolfSSL | 15:117db924cf7c | 674 | data += WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 675 | len -= WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 676 | |
wolfSSL | 15:117db924cf7c | 677 | /* Byte reversal performed in function if required. */ |
wolfSSL | 15:117db924cf7c | 678 | ret = XTRANSFORM(sha256); |
wolfSSL | 15:117db924cf7c | 679 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 680 | break; |
wolfSSL | 15:117db924cf7c | 681 | } |
wolfSSL | 15:117db924cf7c | 682 | } |
wolfSSL | 15:117db924cf7c | 683 | #else |
wolfSSL | 15:117db924cf7c | 684 | { |
wolfSSL | 15:117db924cf7c | 685 | word32 blocksLen = len & ~(WC_SHA256_BLOCK_SIZE-1); |
wolfSSL | 15:117db924cf7c | 686 | |
wolfSSL | 15:117db924cf7c | 687 | AddLength(sha256, blocksLen); |
wolfSSL | 15:117db924cf7c | 688 | while (len >= WC_SHA256_BLOCK_SIZE) { |
wolfSSL | 15:117db924cf7c | 689 | XMEMCPY(local, data, WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 690 | |
wolfSSL | 15:117db924cf7c | 691 | data += WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 692 | len -= WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 693 | |
wolfSSL | 15:117db924cf7c | 694 | ByteReverseWords(sha256->buffer, sha256->buffer, |
wolfSSL | 15:117db924cf7c | 695 | WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 696 | ret = XTRANSFORM(sha256); |
wolfSSL | 15:117db924cf7c | 697 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 698 | break; |
wolfSSL | 15:117db924cf7c | 699 | } |
wolfSSL | 15:117db924cf7c | 700 | } |
wolfSSL | 15:117db924cf7c | 701 | #endif |
wolfSSL | 15:117db924cf7c | 702 | |
wolfSSL | 15:117db924cf7c | 703 | if (len > 0) { |
wolfSSL | 15:117db924cf7c | 704 | XMEMCPY(local, data, len); |
wolfSSL | 15:117db924cf7c | 705 | sha256->buffLen = len; |
wolfSSL | 15:117db924cf7c | 706 | } |
wolfSSL | 15:117db924cf7c | 707 | |
wolfSSL | 15:117db924cf7c | 708 | return ret; |
wolfSSL | 15:117db924cf7c | 709 | } |
wolfSSL | 15:117db924cf7c | 710 | |
wolfSSL | 15:117db924cf7c | 711 | int wc_Sha256Update(wc_Sha256* sha256, const byte* data, word32 len) |
wolfSSL | 15:117db924cf7c | 712 | { |
wolfSSL | 15:117db924cf7c | 713 | return Sha256Update(sha256, data, len); |
wolfSSL | 15:117db924cf7c | 714 | } |
wolfSSL | 15:117db924cf7c | 715 | |
wolfSSL | 15:117db924cf7c | 716 | static WC_INLINE int Sha256Final(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 717 | { |
wolfSSL | 15:117db924cf7c | 718 | |
wolfSSL | 15:117db924cf7c | 719 | int ret; |
wolfSSL | 15:117db924cf7c | 720 | byte* local = (byte*)sha256->buffer; |
wolfSSL | 15:117db924cf7c | 721 | |
wolfSSL | 15:117db924cf7c | 722 | if (sha256 == NULL) { |
wolfSSL | 15:117db924cf7c | 723 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 724 | } |
wolfSSL | 15:117db924cf7c | 725 | |
wolfSSL | 15:117db924cf7c | 726 | AddLength(sha256, sha256->buffLen); /* before adding pads */ |
wolfSSL | 15:117db924cf7c | 727 | local[sha256->buffLen++] = 0x80; /* add 1 */ |
wolfSSL | 15:117db924cf7c | 728 | |
wolfSSL | 15:117db924cf7c | 729 | /* pad with zeros */ |
wolfSSL | 15:117db924cf7c | 730 | if (sha256->buffLen > WC_SHA256_PAD_SIZE) { |
wolfSSL | 15:117db924cf7c | 731 | XMEMSET(&local[sha256->buffLen], 0, |
wolfSSL | 15:117db924cf7c | 732 | WC_SHA256_BLOCK_SIZE - sha256->buffLen); |
wolfSSL | 15:117db924cf7c | 733 | sha256->buffLen += WC_SHA256_BLOCK_SIZE - sha256->buffLen; |
wolfSSL | 15:117db924cf7c | 734 | |
wolfSSL | 15:117db924cf7c | 735 | { |
wolfSSL | 15:117db924cf7c | 736 | #if defined(LITTLE_ENDIAN_ORDER) && !defined(FREESCALE_MMCAU_SHA) |
wolfSSL | 15:117db924cf7c | 737 | #if defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 738 | if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags)) |
wolfSSL | 15:117db924cf7c | 739 | #endif |
wolfSSL | 15:117db924cf7c | 740 | { |
wolfSSL | 15:117db924cf7c | 741 | ByteReverseWords(sha256->buffer, sha256->buffer, |
wolfSSL | 15:117db924cf7c | 742 | WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 743 | } |
wolfSSL | 15:117db924cf7c | 744 | #endif |
wolfSSL | 15:117db924cf7c | 745 | } |
wolfSSL | 15:117db924cf7c | 746 | |
wolfSSL | 15:117db924cf7c | 747 | ret = XTRANSFORM(sha256); |
wolfSSL | 15:117db924cf7c | 748 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 749 | return ret; |
wolfSSL | 15:117db924cf7c | 750 | |
wolfSSL | 15:117db924cf7c | 751 | sha256->buffLen = 0; |
wolfSSL | 15:117db924cf7c | 752 | } |
wolfSSL | 15:117db924cf7c | 753 | XMEMSET(&local[sha256->buffLen], 0, WC_SHA256_PAD_SIZE - sha256->buffLen); |
wolfSSL | 15:117db924cf7c | 754 | |
wolfSSL | 15:117db924cf7c | 755 | /* put lengths in bits */ |
wolfSSL | 15:117db924cf7c | 756 | sha256->hiLen = (sha256->loLen >> (8 * sizeof(sha256->loLen) - 3)) + |
wolfSSL | 15:117db924cf7c | 757 | (sha256->hiLen << 3); |
wolfSSL | 15:117db924cf7c | 758 | sha256->loLen = sha256->loLen << 3; |
wolfSSL | 15:117db924cf7c | 759 | |
wolfSSL | 15:117db924cf7c | 760 | /* store lengths */ |
wolfSSL | 15:117db924cf7c | 761 | #if defined(LITTLE_ENDIAN_ORDER) && !defined(FREESCALE_MMCAU_SHA) |
wolfSSL | 15:117db924cf7c | 762 | #if defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 763 | if (!IS_INTEL_AVX1(intel_flags) && !IS_INTEL_AVX2(intel_flags)) |
wolfSSL | 15:117db924cf7c | 764 | #endif |
wolfSSL | 15:117db924cf7c | 765 | { |
wolfSSL | 15:117db924cf7c | 766 | ByteReverseWords(sha256->buffer, sha256->buffer, |
wolfSSL | 15:117db924cf7c | 767 | WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 768 | } |
wolfSSL | 15:117db924cf7c | 769 | #endif |
wolfSSL | 15:117db924cf7c | 770 | /* ! length ordering dependent on digest endian type ! */ |
wolfSSL | 15:117db924cf7c | 771 | XMEMCPY(&local[WC_SHA256_PAD_SIZE], &sha256->hiLen, sizeof(word32)); |
wolfSSL | 15:117db924cf7c | 772 | XMEMCPY(&local[WC_SHA256_PAD_SIZE + sizeof(word32)], &sha256->loLen, |
wolfSSL | 15:117db924cf7c | 773 | sizeof(word32)); |
wolfSSL | 15:117db924cf7c | 774 | |
wolfSSL | 15:117db924cf7c | 775 | #if defined(FREESCALE_MMCAU_SHA) || defined(HAVE_INTEL_AVX1) || \ |
wolfSSL | 15:117db924cf7c | 776 | defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 777 | /* Kinetis requires only these bytes reversed */ |
wolfSSL | 15:117db924cf7c | 778 | #if defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 779 | if (IS_INTEL_AVX1(intel_flags) || IS_INTEL_AVX2(intel_flags)) |
wolfSSL | 15:117db924cf7c | 780 | #endif |
wolfSSL | 15:117db924cf7c | 781 | { |
wolfSSL | 15:117db924cf7c | 782 | ByteReverseWords( |
wolfSSL | 15:117db924cf7c | 783 | &sha256->buffer[WC_SHA256_PAD_SIZE / sizeof(word32)], |
wolfSSL | 15:117db924cf7c | 784 | &sha256->buffer[WC_SHA256_PAD_SIZE / sizeof(word32)], |
wolfSSL | 15:117db924cf7c | 785 | 2 * sizeof(word32)); |
wolfSSL | 15:117db924cf7c | 786 | } |
wolfSSL | 15:117db924cf7c | 787 | #endif |
wolfSSL | 15:117db924cf7c | 788 | |
wolfSSL | 15:117db924cf7c | 789 | return XTRANSFORM(sha256); |
wolfSSL | 15:117db924cf7c | 790 | } |
wolfSSL | 15:117db924cf7c | 791 | |
wolfSSL | 15:117db924cf7c | 792 | int wc_Sha256FinalRaw(wc_Sha256* sha256, byte* hash) |
wolfSSL | 15:117db924cf7c | 793 | { |
wolfSSL | 15:117db924cf7c | 794 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 795 | word32 digest[WC_SHA256_DIGEST_SIZE / sizeof(word32)]; |
wolfSSL | 15:117db924cf7c | 796 | #endif |
wolfSSL | 15:117db924cf7c | 797 | |
wolfSSL | 15:117db924cf7c | 798 | if (sha256 == NULL || hash == NULL) { |
wolfSSL | 15:117db924cf7c | 799 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 800 | } |
wolfSSL | 15:117db924cf7c | 801 | |
wolfSSL | 15:117db924cf7c | 802 | #ifdef LITTLE_ENDIAN_ORDER |
wolfSSL | 15:117db924cf7c | 803 | ByteReverseWords((word32*)digest, (word32*)sha256->digest, |
wolfSSL | 15:117db924cf7c | 804 | WC_SHA256_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 805 | XMEMCPY(hash, digest, WC_SHA256_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 806 | #else |
wolfSSL | 15:117db924cf7c | 807 | XMEMCPY(hash, sha256->digest, WC_SHA256_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 808 | #endif |
wolfSSL | 15:117db924cf7c | 809 | |
wolfSSL | 15:117db924cf7c | 810 | return 0; |
wolfSSL | 15:117db924cf7c | 811 | } |
wolfSSL | 15:117db924cf7c | 812 | |
wolfSSL | 15:117db924cf7c | 813 | int wc_Sha256Final(wc_Sha256* sha256, byte* hash) |
wolfSSL | 15:117db924cf7c | 814 | { |
wolfSSL | 15:117db924cf7c | 815 | int ret; |
wolfSSL | 15:117db924cf7c | 816 | |
wolfSSL | 15:117db924cf7c | 817 | if (sha256 == NULL || hash == NULL) { |
wolfSSL | 15:117db924cf7c | 818 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 819 | } |
wolfSSL | 15:117db924cf7c | 820 | |
wolfSSL | 15:117db924cf7c | 821 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA256) |
wolfSSL | 15:117db924cf7c | 822 | if (sha256->asyncDev.marker == WOLFSSL_ASYNC_MARKER_SHA256) { |
wolfSSL | 15:117db924cf7c | 823 | #if defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 824 | return IntelQaSymSha256(&sha256->asyncDev, hash, NULL, |
wolfSSL | 15:117db924cf7c | 825 | WC_SHA256_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 826 | #endif |
wolfSSL | 15:117db924cf7c | 827 | } |
wolfSSL | 15:117db924cf7c | 828 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 829 | |
wolfSSL | 15:117db924cf7c | 830 | ret = Sha256Final(sha256); |
wolfSSL | 15:117db924cf7c | 831 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 832 | return ret; |
wolfSSL | 15:117db924cf7c | 833 | |
wolfSSL | 15:117db924cf7c | 834 | #if defined(LITTLE_ENDIAN_ORDER) |
wolfSSL | 15:117db924cf7c | 835 | ByteReverseWords(sha256->digest, sha256->digest, WC_SHA256_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 836 | #endif |
wolfSSL | 15:117db924cf7c | 837 | XMEMCPY(hash, sha256->digest, WC_SHA256_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 838 | |
wolfSSL | 15:117db924cf7c | 839 | return InitSha256(sha256); /* reset state */ |
wolfSSL | 15:117db924cf7c | 840 | } |
wolfSSL | 15:117db924cf7c | 841 | |
wolfSSL | 15:117db924cf7c | 842 | #endif /* XTRANSFORM */ |
wolfSSL | 15:117db924cf7c | 843 | |
wolfSSL | 15:117db924cf7c | 844 | |
wolfSSL | 15:117db924cf7c | 845 | #if defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 846 | |
wolfSSL | 15:117db924cf7c | 847 | #define _LOAD_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 848 | "movl (%[sha256]), %%r8d \n\t" \ |
wolfSSL | 15:117db924cf7c | 849 | "movl 4(%[sha256]), %%r9d \n\t" \ |
wolfSSL | 15:117db924cf7c | 850 | "movl 8(%[sha256]), %%r10d\n\t" \ |
wolfSSL | 15:117db924cf7c | 851 | "movl 12(%[sha256]), %%r11d\n\t" \ |
wolfSSL | 15:117db924cf7c | 852 | "movl 16(%[sha256]), %%r12d\n\t" \ |
wolfSSL | 15:117db924cf7c | 853 | "movl 20(%[sha256]), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 854 | "movl 24(%[sha256]), %%r14d\n\t" \ |
wolfSSL | 15:117db924cf7c | 855 | "movl 28(%[sha256]), %%r15d\n\t" |
wolfSSL | 15:117db924cf7c | 856 | |
wolfSSL | 15:117db924cf7c | 857 | #define _STORE_ADD_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 858 | "addl %%r8d , (%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 859 | "addl %%r9d , 4(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 860 | "addl %%r10d, 8(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 861 | "addl %%r11d, 12(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 862 | "addl %%r12d, 16(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 863 | "addl %%r13d, 20(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 864 | "addl %%r14d, 24(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 865 | "addl %%r15d, 28(%[sha256])\n\t" |
wolfSSL | 15:117db924cf7c | 866 | |
wolfSSL | 15:117db924cf7c | 867 | #define _ADD_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 868 | "addl (%[sha256]), %%r8d \n\t" \ |
wolfSSL | 15:117db924cf7c | 869 | "addl 4(%[sha256]), %%r9d \n\t" \ |
wolfSSL | 15:117db924cf7c | 870 | "addl 8(%[sha256]), %%r10d\n\t" \ |
wolfSSL | 15:117db924cf7c | 871 | "addl 12(%[sha256]), %%r11d\n\t" \ |
wolfSSL | 15:117db924cf7c | 872 | "addl 16(%[sha256]), %%r12d\n\t" \ |
wolfSSL | 15:117db924cf7c | 873 | "addl 20(%[sha256]), %%r13d\n\t" \ |
wolfSSL | 15:117db924cf7c | 874 | "addl 24(%[sha256]), %%r14d\n\t" \ |
wolfSSL | 15:117db924cf7c | 875 | "addl 28(%[sha256]), %%r15d\n\t" |
wolfSSL | 15:117db924cf7c | 876 | |
wolfSSL | 15:117db924cf7c | 877 | #define _STORE_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 878 | "movl %%r8d , (%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 879 | "movl %%r9d , 4(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 880 | "movl %%r10d, 8(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 881 | "movl %%r11d, 12(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 882 | "movl %%r12d, 16(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 883 | "movl %%r13d, 20(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 884 | "movl %%r14d, 24(%[sha256])\n\t" \ |
wolfSSL | 15:117db924cf7c | 885 | "movl %%r15d, 28(%[sha256])\n\t" |
wolfSSL | 15:117db924cf7c | 886 | |
wolfSSL | 15:117db924cf7c | 887 | #define LOAD_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 888 | _LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 889 | |
wolfSSL | 15:117db924cf7c | 890 | #define STORE_ADD_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 891 | _STORE_ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 892 | |
wolfSSL | 15:117db924cf7c | 893 | #define ADD_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 894 | _ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 895 | |
wolfSSL | 15:117db924cf7c | 896 | #define STORE_DIGEST() \ |
wolfSSL | 15:117db924cf7c | 897 | _STORE_DIGEST() |
wolfSSL | 15:117db924cf7c | 898 | |
wolfSSL | 15:117db924cf7c | 899 | |
wolfSSL | 15:117db924cf7c | 900 | #define S_0 %r8d |
wolfSSL | 15:117db924cf7c | 901 | #define S_1 %r9d |
wolfSSL | 15:117db924cf7c | 902 | #define S_2 %r10d |
wolfSSL | 15:117db924cf7c | 903 | #define S_3 %r11d |
wolfSSL | 15:117db924cf7c | 904 | #define S_4 %r12d |
wolfSSL | 15:117db924cf7c | 905 | #define S_5 %r13d |
wolfSSL | 15:117db924cf7c | 906 | #define S_6 %r14d |
wolfSSL | 15:117db924cf7c | 907 | #define S_7 %r15d |
wolfSSL | 15:117db924cf7c | 908 | |
wolfSSL | 15:117db924cf7c | 909 | #define L1 "%%edx" |
wolfSSL | 15:117db924cf7c | 910 | #define L2 "%%ecx" |
wolfSSL | 15:117db924cf7c | 911 | #define L3 "%%eax" |
wolfSSL | 15:117db924cf7c | 912 | #define L4 "%%ebx" |
wolfSSL | 15:117db924cf7c | 913 | #define WK "%%rsp" |
wolfSSL | 15:117db924cf7c | 914 | |
wolfSSL | 15:117db924cf7c | 915 | #define WORK_REGS "eax", "ebx", "ecx", "edx" |
wolfSSL | 15:117db924cf7c | 916 | #define STATE_REGS "r8","r9","r10","r11","r12","r13","r14","r15" |
wolfSSL | 15:117db924cf7c | 917 | #define XMM_REGS "xmm0", "xmm1", "xmm2", "xmm3", "xmm4", "xmm5", "xmm6", \ |
wolfSSL | 15:117db924cf7c | 918 | "xmm7", "xmm8", "xmm9", "xmm10", "xmm11", "xmm12", "xmm13" |
wolfSSL | 15:117db924cf7c | 919 | |
wolfSSL | 15:117db924cf7c | 920 | #if defined(HAVE_INTEL_RORX) |
wolfSSL | 15:117db924cf7c | 921 | #define RND_STEP_RORX_0_1(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 922 | /* L3 = f */ \ |
wolfSSL | 15:117db924cf7c | 923 | "movl %" #f ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 924 | /* L2 = e>>>11 */ \ |
wolfSSL | 15:117db924cf7c | 925 | "rorx $11, %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 926 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 927 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 928 | |
wolfSSL | 15:117db924cf7c | 929 | #define RND_STEP_RORX_0_2(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 930 | /* L2 = (e>>>6) ^ (e>>>11) */ \ |
wolfSSL | 15:117db924cf7c | 931 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 932 | /* L3 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 933 | "xorl %" #g ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 934 | /* L1 = e>>>25 */ \ |
wolfSSL | 15:117db924cf7c | 935 | "rorx $25, %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 936 | |
wolfSSL | 15:117db924cf7c | 937 | #define RND_STEP_RORX_0_3(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 938 | /* L3 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 939 | "andl %" #e ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 940 | /* L1 = Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 941 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 942 | /* L2 = a>>>13 */ \ |
wolfSSL | 15:117db924cf7c | 943 | "rorx $13, %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 944 | |
wolfSSL | 15:117db924cf7c | 945 | #define RND_STEP_RORX_0_4(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 946 | /* h += Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 947 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 948 | /* L1 = a>>>2 */ \ |
wolfSSL | 15:117db924cf7c | 949 | "rorx $2, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 950 | /* L3 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 951 | "xorl %" #g ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 952 | |
wolfSSL | 15:117db924cf7c | 953 | #define RND_STEP_RORX_0_5(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 954 | /* L2 = (a>>>2) ^ (a>>>13) */ \ |
wolfSSL | 15:117db924cf7c | 955 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 956 | /* L1 = a>>>22 */ \ |
wolfSSL | 15:117db924cf7c | 957 | "rorx $22, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 958 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 959 | "addl " L3 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 960 | |
wolfSSL | 15:117db924cf7c | 961 | #define RND_STEP_RORX_0_6(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 962 | /* L1 = Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 963 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 964 | /* L3 = b */ \ |
wolfSSL | 15:117db924cf7c | 965 | "movl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 966 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 967 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 968 | |
wolfSSL | 15:117db924cf7c | 969 | #define RND_STEP_RORX_0_7(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 970 | /* L3 = a ^ b */ \ |
wolfSSL | 15:117db924cf7c | 971 | "xorl %" #a ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 972 | /* h += Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 973 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 974 | /* L4 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 975 | "andl " L3 ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 976 | |
wolfSSL | 15:117db924cf7c | 977 | #define RND_STEP_RORX_0_8(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 978 | /* L4 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 979 | "xorl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 980 | /* L1 = d>>>6 (= e>>>6 next RND) */ \ |
wolfSSL | 15:117db924cf7c | 981 | "rorx $6, %" #d ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 982 | /* h += Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 983 | "addl " L4 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 984 | |
wolfSSL | 15:117db924cf7c | 985 | #define RND_STEP_RORX_1_1(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 986 | /* L4 = f */ \ |
wolfSSL | 15:117db924cf7c | 987 | "movl %" #f ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 988 | /* L2 = e>>>11 */ \ |
wolfSSL | 15:117db924cf7c | 989 | "rorx $11, %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 990 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 991 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 992 | |
wolfSSL | 15:117db924cf7c | 993 | #define RND_STEP_RORX_1_2(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 994 | /* L2 = (e>>>6) ^ (e>>>11) */ \ |
wolfSSL | 15:117db924cf7c | 995 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 996 | /* L4 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 997 | "xorl %" #g ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 998 | /* L1 = e>>>25 */ \ |
wolfSSL | 15:117db924cf7c | 999 | "rorx $25, %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1000 | |
wolfSSL | 15:117db924cf7c | 1001 | #define RND_STEP_RORX_1_3(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1002 | /* L4 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 1003 | "andl %" #e ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1004 | /* L1 = Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1005 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1006 | /* L2 = a>>>13 */ \ |
wolfSSL | 15:117db924cf7c | 1007 | "rorx $13, %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1008 | |
wolfSSL | 15:117db924cf7c | 1009 | #define RND_STEP_RORX_1_4(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1010 | /* h += Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1011 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1012 | /* L1 = a>>>2 */ \ |
wolfSSL | 15:117db924cf7c | 1013 | "rorx $2, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1014 | /* L4 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1015 | "xorl %" #g ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1016 | |
wolfSSL | 15:117db924cf7c | 1017 | #define RND_STEP_RORX_1_5(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1018 | /* L2 = (a>>>2) ^ (a>>>13) */ \ |
wolfSSL | 15:117db924cf7c | 1019 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1020 | /* L1 = a>>>22 */ \ |
wolfSSL | 15:117db924cf7c | 1021 | "rorx $22, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1022 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1023 | "addl " L4 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1024 | |
wolfSSL | 15:117db924cf7c | 1025 | #define RND_STEP_RORX_1_6(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1026 | /* L1 = Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 1027 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1028 | /* L4 = b */ \ |
wolfSSL | 15:117db924cf7c | 1029 | "movl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1030 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1031 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1032 | |
wolfSSL | 15:117db924cf7c | 1033 | #define RND_STEP_RORX_1_7(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1034 | /* L4 = a ^ b */ \ |
wolfSSL | 15:117db924cf7c | 1035 | "xorl %" #a ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1036 | /* h += Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 1037 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1038 | /* L3 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 1039 | "andl " L4 ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1040 | |
wolfSSL | 15:117db924cf7c | 1041 | #define RND_STEP_RORX_1_8(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1042 | /* L3 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1043 | "xorl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1044 | /* L1 = d>>>6 (= e>>>6 next RND) */ \ |
wolfSSL | 15:117db924cf7c | 1045 | "rorx $6, %" #d ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1046 | /* h += Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1047 | "addl " L3 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1048 | |
wolfSSL | 15:117db924cf7c | 1049 | #define _RND_RORX_X_0(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1050 | /* L1 = e>>>6 */ \ |
wolfSSL | 15:117db924cf7c | 1051 | "rorx $6, %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1052 | /* L2 = e>>>11 */ \ |
wolfSSL | 15:117db924cf7c | 1053 | "rorx $11, %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1054 | /* Prev RND: h += Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1055 | "addl " L3 ", %" #a "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1056 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 1057 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1058 | /* L3 = f */ \ |
wolfSSL | 15:117db924cf7c | 1059 | "movl %" #f ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1060 | /* L2 = (e>>>6) ^ (e>>>11) */ \ |
wolfSSL | 15:117db924cf7c | 1061 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1062 | /* L3 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 1063 | "xorl %" #g ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1064 | /* L1 = e>>>25 */ \ |
wolfSSL | 15:117db924cf7c | 1065 | "rorx $25, %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1066 | /* L1 = Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1067 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1068 | /* L3 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 1069 | "andl %" #e ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1070 | /* h += Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1071 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1072 | /* L1 = a>>>2 */ \ |
wolfSSL | 15:117db924cf7c | 1073 | "rorx $2, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1074 | /* L2 = a>>>13 */ \ |
wolfSSL | 15:117db924cf7c | 1075 | "rorx $13, %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1076 | /* L3 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1077 | "xorl %" #g ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1078 | /* L2 = (a>>>2) ^ (a>>>13) */ \ |
wolfSSL | 15:117db924cf7c | 1079 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1080 | /* L1 = a>>>22 */ \ |
wolfSSL | 15:117db924cf7c | 1081 | "rorx $22, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1082 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1083 | "addl " L3 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1084 | /* L1 = Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 1085 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1086 | /* L3 = b */ \ |
wolfSSL | 15:117db924cf7c | 1087 | "movl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1088 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1089 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1090 | /* L3 = a ^ b */ \ |
wolfSSL | 15:117db924cf7c | 1091 | "xorl %" #a ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1092 | /* L4 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 1093 | "andl " L3 ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1094 | /* h += Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 1095 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1096 | /* L4 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1097 | "xorl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1098 | |
wolfSSL | 15:117db924cf7c | 1099 | #define _RND_RORX_X_1(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1100 | /* L1 = e>>>6 */ \ |
wolfSSL | 15:117db924cf7c | 1101 | "rorx $6, %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1102 | /* L2 = e>>>11 */ \ |
wolfSSL | 15:117db924cf7c | 1103 | "rorx $11, %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1104 | /* Prev RND: h += Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1105 | "addl " L4 ", %" #a "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1106 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 1107 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1108 | /* L4 = f */ \ |
wolfSSL | 15:117db924cf7c | 1109 | "movl %" #f ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1110 | /* L2 = (e>>>6) ^ (e>>>11) */ \ |
wolfSSL | 15:117db924cf7c | 1111 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1112 | /* L4 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 1113 | "xorl %" #g ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1114 | /* L1 = e>>>25 */ \ |
wolfSSL | 15:117db924cf7c | 1115 | "rorx $25, %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1116 | /* L1 = Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1117 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1118 | /* L4 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 1119 | "andl %" #e ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1120 | /* h += Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1121 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1122 | /* L1 = a>>>2 */ \ |
wolfSSL | 15:117db924cf7c | 1123 | "rorx $2, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1124 | /* L2 = a>>>13 */ \ |
wolfSSL | 15:117db924cf7c | 1125 | "rorx $13, %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1126 | /* L4 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1127 | "xorl %" #g ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1128 | /* L2 = (a>>>2) ^ (a>>>13) */ \ |
wolfSSL | 15:117db924cf7c | 1129 | "xorl " L1 ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1130 | /* L1 = a>>>22 */ \ |
wolfSSL | 15:117db924cf7c | 1131 | "rorx $22, %" #a ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1132 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1133 | "addl " L4 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1134 | /* L1 = Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 1135 | "xorl " L2 ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1136 | /* L4 = b */ \ |
wolfSSL | 15:117db924cf7c | 1137 | "movl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1138 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1139 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1140 | /* L4 = a ^ b */ \ |
wolfSSL | 15:117db924cf7c | 1141 | "xorl %" #a ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1142 | /* L2 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 1143 | "andl " L4 ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1144 | /* h += Sigma0(a) */ \ |
wolfSSL | 15:117db924cf7c | 1145 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1146 | /* L3 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1147 | "xorl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1148 | |
wolfSSL | 15:117db924cf7c | 1149 | |
wolfSSL | 15:117db924cf7c | 1150 | #define RND_RORX_X_0(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1151 | _RND_RORX_X_0(a,b,c,d,e,f,g,h,i) |
wolfSSL | 15:117db924cf7c | 1152 | #define RND_RORX_X_1(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1153 | _RND_RORX_X_1(a,b,c,d,e,f,g,h,i) |
wolfSSL | 15:117db924cf7c | 1154 | |
wolfSSL | 15:117db924cf7c | 1155 | #define RND_RORX_X4(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1156 | RND_RORX_X_0(a,b,c,d,e,f,g,h,i+0) \ |
wolfSSL | 15:117db924cf7c | 1157 | RND_RORX_X_1(h,a,b,c,d,e,f,g,i+1) \ |
wolfSSL | 15:117db924cf7c | 1158 | RND_RORX_X_0(g,h,a,b,c,d,e,f,i+2) \ |
wolfSSL | 15:117db924cf7c | 1159 | RND_RORX_X_1(f,g,h,a,b,c,d,e,i+3) |
wolfSSL | 15:117db924cf7c | 1160 | |
wolfSSL | 15:117db924cf7c | 1161 | #endif /* HAVE_INTEL_RORX */ |
wolfSSL | 15:117db924cf7c | 1162 | |
wolfSSL | 15:117db924cf7c | 1163 | #define RND_STEP_0_1(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1164 | /* L1 = e>>>14 */ \ |
wolfSSL | 15:117db924cf7c | 1165 | "rorl $14, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1166 | |
wolfSSL | 15:117db924cf7c | 1167 | #define RND_STEP_0_2(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1168 | /* L3 = b */ \ |
wolfSSL | 15:117db924cf7c | 1169 | "movl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1170 | /* L2 = f */ \ |
wolfSSL | 15:117db924cf7c | 1171 | "movl %" #f ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1172 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 1173 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1174 | /* L2 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 1175 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1176 | |
wolfSSL | 15:117db924cf7c | 1177 | #define RND_STEP_0_3(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1178 | /* L1 = (e>>>14) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1179 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1180 | /* L2 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 1181 | "andl %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1182 | |
wolfSSL | 15:117db924cf7c | 1183 | #define RND_STEP_0_4(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1184 | /* L1 = ((e>>>14) ^ e) >>> 5 */ \ |
wolfSSL | 15:117db924cf7c | 1185 | "rorl $5, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1186 | /* L2 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1187 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1188 | /* L1 = (((e>>>14) ^ e) >>> 5) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1189 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1190 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1191 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1192 | |
wolfSSL | 15:117db924cf7c | 1193 | #define RND_STEP_0_5(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1194 | /* L1 = ((((e>>>14) ^ e) >>> 5) ^ e) >>> 6 */ \ |
wolfSSL | 15:117db924cf7c | 1195 | "rorl $6, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1196 | /* L3 = a ^ b (= b ^ c of next RND) */ \ |
wolfSSL | 15:117db924cf7c | 1197 | "xorl %" #a ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1198 | /* h = h + w_k + Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1199 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1200 | /* L2 = a */ \ |
wolfSSL | 15:117db924cf7c | 1201 | "movl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1202 | |
wolfSSL | 15:117db924cf7c | 1203 | #define RND_STEP_0_6(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1204 | /* L3 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 1205 | "andl " L3 ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1206 | /* L2 = a>>>9 */ \ |
wolfSSL | 15:117db924cf7c | 1207 | "rorl $9, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1208 | /* L2 = (a>>>9) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1209 | "xorl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1210 | /* L1 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1211 | "xorl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1212 | |
wolfSSL | 15:117db924cf7c | 1213 | #define RND_STEP_0_7(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1214 | /* L2 = ((a>>>9) ^ a) >>> 11 */ \ |
wolfSSL | 15:117db924cf7c | 1215 | "rorl $11, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1216 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1217 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1218 | /* L2 = (((a>>>9) ^ a) >>> 11) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1219 | "xorl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1220 | /* h = h + w_k + Sigma1(e) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1221 | "addl " L4 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1222 | |
wolfSSL | 15:117db924cf7c | 1223 | #define RND_STEP_0_8(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1224 | /* L2 = ((((a>>>9) ^ a) >>> 11) ^ a) >>> 2 */ \ |
wolfSSL | 15:117db924cf7c | 1225 | "rorl $2, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1226 | /* L1 = d (e of next RND) */ \ |
wolfSSL | 15:117db924cf7c | 1227 | "movl %" #d ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1228 | /* h = h + w_k + Sigma1(e) Sigma0(a) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1229 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1230 | |
wolfSSL | 15:117db924cf7c | 1231 | #define RND_STEP_1_1(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1232 | /* L1 = e>>>14 */ \ |
wolfSSL | 15:117db924cf7c | 1233 | "rorl $14, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1234 | |
wolfSSL | 15:117db924cf7c | 1235 | #define RND_STEP_1_2(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1236 | /* L3 = b */ \ |
wolfSSL | 15:117db924cf7c | 1237 | "movl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1238 | /* L2 = f */ \ |
wolfSSL | 15:117db924cf7c | 1239 | "movl %" #f ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1240 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 1241 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1242 | /* L2 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 1243 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1244 | |
wolfSSL | 15:117db924cf7c | 1245 | #define RND_STEP_1_3(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1246 | /* L1 = (e>>>14) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1247 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1248 | /* L2 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 1249 | "andl %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1250 | |
wolfSSL | 15:117db924cf7c | 1251 | #define RND_STEP_1_4(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1252 | /* L1 = ((e>>>14) ^ e) >>> 5 */ \ |
wolfSSL | 15:117db924cf7c | 1253 | "rorl $5, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1254 | /* L2 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1255 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1256 | /* L1 = (((e>>>14) ^ e) >>> 5) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1257 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1258 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1259 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1260 | |
wolfSSL | 15:117db924cf7c | 1261 | #define RND_STEP_1_5(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1262 | /* L1 = ((((e>>>14) ^ e) >>> 5) ^ e) >>> 6 */ \ |
wolfSSL | 15:117db924cf7c | 1263 | "rorl $6, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1264 | /* L4 = a ^ b (= b ^ c of next RND) */ \ |
wolfSSL | 15:117db924cf7c | 1265 | "xorl %" #a ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1266 | /* h = h + w_k + Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1267 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1268 | /* L2 = a */ \ |
wolfSSL | 15:117db924cf7c | 1269 | "movl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1270 | |
wolfSSL | 15:117db924cf7c | 1271 | #define RND_STEP_1_6(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1272 | /* L3 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 1273 | "andl " L4 ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1274 | /* L2 = a>>>9 */ \ |
wolfSSL | 15:117db924cf7c | 1275 | "rorl $9, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1276 | /* L2 = (a>>>9) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1277 | "xorl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1278 | /* L1 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1279 | "xorl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1280 | |
wolfSSL | 15:117db924cf7c | 1281 | #define RND_STEP_1_7(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1282 | /* L2 = ((a>>>9) ^ a) >>> 11 */ \ |
wolfSSL | 15:117db924cf7c | 1283 | "rorl $11, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1284 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1285 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1286 | /* L2 = (((a>>>9) ^ a) >>> 11) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1287 | "xorl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1288 | /* h = h + w_k + Sigma1(e) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1289 | "addl " L3 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1290 | |
wolfSSL | 15:117db924cf7c | 1291 | #define RND_STEP_1_8(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1292 | /* L2 = ((((a>>>9) ^ a) >>> 11) ^ a) >>> 2 */ \ |
wolfSSL | 15:117db924cf7c | 1293 | "rorl $2, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1294 | /* L1 = d (e of next RND) */ \ |
wolfSSL | 15:117db924cf7c | 1295 | "movl %" #d ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1296 | /* h = h + w_k + Sigma1(e) Sigma0(a) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1297 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1298 | |
wolfSSL | 15:117db924cf7c | 1299 | #define _RND_ALL_0(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1300 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 1301 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1302 | /* L2 = f */ \ |
wolfSSL | 15:117db924cf7c | 1303 | "movl %" #f ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1304 | /* L3 = b */ \ |
wolfSSL | 15:117db924cf7c | 1305 | "movl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1306 | /* L2 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 1307 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1308 | /* L1 = e>>>14 */ \ |
wolfSSL | 15:117db924cf7c | 1309 | "rorl $14, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1310 | /* L2 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 1311 | "andl %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1312 | /* L1 = (e>>>14) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1313 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1314 | /* L2 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1315 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1316 | /* L1 = ((e>>>14) ^ e) >>> 5 */ \ |
wolfSSL | 15:117db924cf7c | 1317 | "rorl $5, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1318 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1319 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1320 | /* L1 = (((e>>>14) ^ e) >>> 5) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1321 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1322 | /* L3 = a ^ b */ \ |
wolfSSL | 15:117db924cf7c | 1323 | "xorl %" #a ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1324 | /* L1 = ((((e>>>14) ^ e) >>> 5) ^ e) >>> 6 */ \ |
wolfSSL | 15:117db924cf7c | 1325 | "rorl $6, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1326 | /* L2 = a */ \ |
wolfSSL | 15:117db924cf7c | 1327 | "movl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1328 | /* h = h + w_k + Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1329 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1330 | /* L2 = a>>>9 */ \ |
wolfSSL | 15:117db924cf7c | 1331 | "rorl $9, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1332 | /* L3 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 1333 | "andl " L3 ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1334 | /* L2 = (a>>>9) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1335 | "xorl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1336 | /* L1 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1337 | "xorl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1338 | /* L2 = ((a>>>9) ^ a) >>> 11 */ \ |
wolfSSL | 15:117db924cf7c | 1339 | "rorl $11, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1340 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1341 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1342 | /* L2 = (((a>>>9) ^ a) >>> 11) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1343 | "xorl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1344 | /* h = h + w_k + Sigma1(e) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1345 | "addl " L4 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1346 | /* L2 = ((((a>>>9) ^ a) >>> 11) ^ a) >>> 2 */ \ |
wolfSSL | 15:117db924cf7c | 1347 | "rorl $2, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1348 | /* L1 = d (e of next RND) */ \ |
wolfSSL | 15:117db924cf7c | 1349 | "movl %" #d ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1350 | /* h = h + w_k + Sigma1(e) Sigma0(a) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1351 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1352 | |
wolfSSL | 15:117db924cf7c | 1353 | #define _RND_ALL_1(a,b,c,d,e,f,g,h,i) \ |
wolfSSL | 15:117db924cf7c | 1354 | /* h += w_k */ \ |
wolfSSL | 15:117db924cf7c | 1355 | "addl (" #i ")*4(" WK "), %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1356 | /* L2 = f */ \ |
wolfSSL | 15:117db924cf7c | 1357 | "movl %" #f ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1358 | /* L3 = b */ \ |
wolfSSL | 15:117db924cf7c | 1359 | "movl %" #b ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1360 | /* L2 = f ^ g */ \ |
wolfSSL | 15:117db924cf7c | 1361 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1362 | /* L1 = e>>>14 */ \ |
wolfSSL | 15:117db924cf7c | 1363 | "rorl $14, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1364 | /* L2 = (f ^ g) & e */ \ |
wolfSSL | 15:117db924cf7c | 1365 | "andl %" #e ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1366 | /* L1 = (e>>>14) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1367 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1368 | /* L2 = Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1369 | "xorl %" #g ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1370 | /* L1 = ((e>>>14) ^ e) >>> 5 */ \ |
wolfSSL | 15:117db924cf7c | 1371 | "rorl $5, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1372 | /* h += Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1373 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1374 | /* L1 = (((e>>>14) ^ e) >>> 5) ^ e */ \ |
wolfSSL | 15:117db924cf7c | 1375 | "xorl %" #e ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1376 | /* L3 = a ^ b */ \ |
wolfSSL | 15:117db924cf7c | 1377 | "xorl %" #a ", " L4 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1378 | /* L1 = ((((e>>>14) ^ e) >>> 5) ^ e) >>> 6 */ \ |
wolfSSL | 15:117db924cf7c | 1379 | "rorl $6, " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1380 | /* L2 = a */ \ |
wolfSSL | 15:117db924cf7c | 1381 | "movl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1382 | /* h = h + w_k + Sigma1(e) */ \ |
wolfSSL | 15:117db924cf7c | 1383 | "addl " L1 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1384 | /* L2 = a>>>9 */ \ |
wolfSSL | 15:117db924cf7c | 1385 | "rorl $9, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1386 | /* L3 = (a ^ b) & (b ^ c) */ \ |
wolfSSL | 15:117db924cf7c | 1387 | "andl " L4 ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1388 | /* L2 = (a>>>9) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1389 | "xorl %" #a", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1390 | /* L1 = Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1391 | "xorl %" #b ", " L3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1392 | /* L2 = ((a>>>9) ^ a) >>> 11 */ \ |
wolfSSL | 15:117db924cf7c | 1393 | "rorl $11, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1394 | /* d += h + w_k + Sigma1(e) + Ch(e,f,g) */ \ |
wolfSSL | 15:117db924cf7c | 1395 | "addl %" #h ", %" #d "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1396 | /* L2 = (((a>>>9) ^ a) >>> 11) ^ a */ \ |
wolfSSL | 15:117db924cf7c | 1397 | "xorl %" #a ", " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1398 | /* h = h + w_k + Sigma1(e) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1399 | "addl " L3 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1400 | /* L2 = ((((a>>>9) ^ a) >>> 11) ^ a) >>> 2 */ \ |
wolfSSL | 15:117db924cf7c | 1401 | "rorl $2, " L2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1402 | /* L1 = d (e of next RND) */ \ |
wolfSSL | 15:117db924cf7c | 1403 | "movl %" #d ", " L1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1404 | /* h = h + w_k + Sigma1(e) Sigma0(a) + Ch(e,f,g) + Maj(a,b,c) */ \ |
wolfSSL | 15:117db924cf7c | 1405 | "addl " L2 ", %" #h "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1406 | |
wolfSSL | 15:117db924cf7c | 1407 | |
wolfSSL | 15:117db924cf7c | 1408 | #define RND_ALL_0(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1409 | _RND_ALL_0(a, b, c, d, e, f, g, h, i) |
wolfSSL | 15:117db924cf7c | 1410 | #define RND_ALL_1(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1411 | _RND_ALL_1(a, b, c, d, e, f, g, h, i) |
wolfSSL | 15:117db924cf7c | 1412 | |
wolfSSL | 15:117db924cf7c | 1413 | #define RND_ALL_4(a, b, c, d, e, f, g, h, i) \ |
wolfSSL | 15:117db924cf7c | 1414 | RND_ALL_0(a, b, c, d, e, f, g, h, i+0) \ |
wolfSSL | 15:117db924cf7c | 1415 | RND_ALL_1(h, a, b, c, d, e, f, g, i+1) \ |
wolfSSL | 15:117db924cf7c | 1416 | RND_ALL_0(g, h, a, b, c, d, e, f, i+2) \ |
wolfSSL | 15:117db924cf7c | 1417 | RND_ALL_1(f, g, h, a, b, c, d, e, i+3) |
wolfSSL | 15:117db924cf7c | 1418 | |
wolfSSL | 15:117db924cf7c | 1419 | #endif /* defined(HAVE_INTEL_AVX1) || defined(HAVE_INTEL_AVX2) */ |
wolfSSL | 15:117db924cf7c | 1420 | |
wolfSSL | 15:117db924cf7c | 1421 | #if defined(HAVE_INTEL_AVX1) /* inline Assember for Intel AVX1 instructions */ |
wolfSSL | 15:117db924cf7c | 1422 | |
wolfSSL | 15:117db924cf7c | 1423 | #define _VPALIGNR(op1, op2, op3, op4) \ |
wolfSSL | 15:117db924cf7c | 1424 | "vpalignr $" #op4", %" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1425 | #define VPALIGNR(op1, op2, op3, op4) \ |
wolfSSL | 15:117db924cf7c | 1426 | _VPALIGNR(op1, op2, op3, op4) |
wolfSSL | 15:117db924cf7c | 1427 | #define _VPADDD(op1, op2, op3) \ |
wolfSSL | 15:117db924cf7c | 1428 | "vpaddd %" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1429 | #define VPADDD(op1, op2, op3) \ |
wolfSSL | 15:117db924cf7c | 1430 | _VPADDD(op1, op2, op3) |
wolfSSL | 15:117db924cf7c | 1431 | #define _VPSRLD(op1, op2, op3) \ |
wolfSSL | 15:117db924cf7c | 1432 | "vpsrld $" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1433 | #define VPSRLD(op1, op2, op3) \ |
wolfSSL | 15:117db924cf7c | 1434 | _VPSRLD(op1, op2, op3) |
wolfSSL | 15:117db924cf7c | 1435 | #define _VPSRLQ(op1, op2, op3) \ |
wolfSSL | 15:117db924cf7c | 1436 | "vpsrlq $" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1437 | #define VPSRLQ(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1438 | _VPSRLQ(op1,op2,op3) |
wolfSSL | 15:117db924cf7c | 1439 | #define _VPSLLD(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1440 | "vpslld $" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1441 | #define VPSLLD(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1442 | _VPSLLD(op1,op2,op3) |
wolfSSL | 15:117db924cf7c | 1443 | #define _VPOR(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1444 | "vpor %" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1445 | #define VPOR(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1446 | _VPOR(op1,op2,op3) |
wolfSSL | 15:117db924cf7c | 1447 | #define _VPXOR(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1448 | "vpxor %" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1449 | #define VPXOR(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1450 | _VPXOR(op1,op2,op3) |
wolfSSL | 15:117db924cf7c | 1451 | #define _VPSHUFD(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1452 | "vpshufd $" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1453 | #define VPSHUFD(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1454 | _VPSHUFD(op1,op2,op3) |
wolfSSL | 15:117db924cf7c | 1455 | #define _VPSHUFB(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1456 | "vpshufb %" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1457 | #define VPSHUFB(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1458 | _VPSHUFB(op1,op2,op3) |
wolfSSL | 15:117db924cf7c | 1459 | #define _VPSLLDQ(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1460 | "vpslldq $" #op3", %" #op2", %" #op1"\n\t" |
wolfSSL | 15:117db924cf7c | 1461 | #define VPSLLDQ(op1,op2,op3) \ |
wolfSSL | 15:117db924cf7c | 1462 | _VPSLLDQ(op1,op2,op3) |
wolfSSL | 15:117db924cf7c | 1463 | |
wolfSSL | 15:117db924cf7c | 1464 | #define MsgSched(X0,X1,X2,X3,a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1465 | RND_STEP_0_1(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1466 | VPALIGNR (XTMP1, X1, X0, 4) /* XTMP1 = W[-15] */ \ |
wolfSSL | 15:117db924cf7c | 1467 | VPALIGNR (XTMP0, X3, X2, 4) /* XTMP0 = W[-7] */ \ |
wolfSSL | 15:117db924cf7c | 1468 | RND_STEP_0_2(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1469 | RND_STEP_0_3(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1470 | VPSRLD (XTMP2, XTMP1, 7) /* XTMP2 = W[-15] >> 7 */ \ |
wolfSSL | 15:117db924cf7c | 1471 | VPSLLD (XTMP3, XTMP1, 25) /* XTEMP3 = W[-15] << (32-7) */ \ |
wolfSSL | 15:117db924cf7c | 1472 | RND_STEP_0_4(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1473 | RND_STEP_0_5(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1474 | VPSRLD (XTMP4, XTMP1, 18) /* XTEMP4 = W[-15] >> 18 */ \ |
wolfSSL | 15:117db924cf7c | 1475 | VPSLLD (XTMP5, XTMP1, 14) /* XTEMP5 = W[-15] << (32-18) */ \ |
wolfSSL | 15:117db924cf7c | 1476 | RND_STEP_0_6(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1477 | RND_STEP_0_7(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1478 | VPOR (XTMP2, XTMP3, XTMP2) /* XTMP2 = W[-15] >>> 7 */ \ |
wolfSSL | 15:117db924cf7c | 1479 | VPOR (XTMP4, XTMP5, XTMP4) /* XTMP4 = W[-15] >>> 18 */ \ |
wolfSSL | 15:117db924cf7c | 1480 | RND_STEP_0_8(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1481 | RND_STEP_1_1(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1482 | RND_STEP_1_2(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1483 | VPSRLD (XTMP5, XTMP1, 3) /* XTMP4 = W[-15] >> 3 */ \ |
wolfSSL | 15:117db924cf7c | 1484 | VPXOR (XTMP2, XTMP4, XTMP2) \ |
wolfSSL | 15:117db924cf7c | 1485 | /* XTMP2 = W[-15] MY_ROR 7 ^ W[-15] MY_ROR 18 */ \ |
wolfSSL | 15:117db924cf7c | 1486 | RND_STEP_1_3(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1487 | RND_STEP_1_4(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1488 | VPXOR (XTMP1, XTMP5, XTMP2) /* XTMP1 = s0 */ \ |
wolfSSL | 15:117db924cf7c | 1489 | VPSHUFD (XTMP2, X3, 0b11111010) /* XTMP2 = W[-2] {BBAA}*/ \ |
wolfSSL | 15:117db924cf7c | 1490 | RND_STEP_1_5(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1491 | RND_STEP_1_6(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1492 | VPSRLD (XTMP4, XTMP2, 10) /* XTMP4 = W[-2] >> 10 {BBAA} */ \ |
wolfSSL | 15:117db924cf7c | 1493 | VPSRLQ (XTMP3, XTMP2, 19) /* XTMP3 = W[-2] MY_ROR 19 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1494 | RND_STEP_1_7(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1495 | RND_STEP_1_8(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1496 | RND_STEP_0_1(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1497 | VPSRLQ (XTMP2, XTMP2, 17) /* XTMP2 = W[-2] MY_ROR 17 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1498 | VPADDD (XTMP0, XTMP0, X0) \ |
wolfSSL | 15:117db924cf7c | 1499 | RND_STEP_0_2(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1500 | RND_STEP_0_3(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1501 | RND_STEP_0_4(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1502 | VPXOR (XTMP2, XTMP3, XTMP2) \ |
wolfSSL | 15:117db924cf7c | 1503 | VPADDD (XTMP0, XTMP0, XTMP1) /* XTMP0 = W[-16] + W[-7] + s0 */ \ |
wolfSSL | 15:117db924cf7c | 1504 | RND_STEP_0_5(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1505 | VPXOR (XTMP4, XTMP4, XTMP2) /* XTMP4 = s1 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1506 | RND_STEP_0_6(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1507 | VPSHUFB (XTMP4, XTMP4, SHUF_00BA) /* XTMP4 = s1 {00BA} */ \ |
wolfSSL | 15:117db924cf7c | 1508 | RND_STEP_0_7(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1509 | VPADDD (XTMP0, XTMP0, XTMP4) /* XTMP0 = {..., ..., W[1], W[0]} */ \ |
wolfSSL | 15:117db924cf7c | 1510 | RND_STEP_0_8(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1511 | RND_STEP_1_1(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1512 | VPSHUFD (XTMP2, XTMP0, 0b01010000) /* XTMP2 = W[-2] {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 1513 | RND_STEP_1_2(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1514 | VPSRLQ (XTMP4, XTMP2, 17) /* XTMP4 = W[-2] MY_ROR 17 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1515 | VPSRLQ (XTMP3, XTMP2, 19) /* XTMP3 = W[-2] MY_ROR 19 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1516 | RND_STEP_1_3(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1517 | RND_STEP_1_4(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1518 | VPSRLD (XTMP5, XTMP2, 10) /* XTMP5 = W[-2] >> 10 {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 1519 | VPXOR (XTMP4, XTMP3, XTMP4) \ |
wolfSSL | 15:117db924cf7c | 1520 | RND_STEP_1_5(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1521 | RND_STEP_1_6(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1522 | VPXOR (XTMP5, XTMP4, XTMP5) /* XTMP5 = s1 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1523 | RND_STEP_1_7(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1524 | VPSHUFB (XTMP5, XTMP5, SHUF_DC00) /* XTMP5 = s1 {DC00} */ \ |
wolfSSL | 15:117db924cf7c | 1525 | RND_STEP_1_8(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1526 | VPADDD (X0, XTMP5, XTMP0) /* X0 = {W[3], W[2], W[1], W[0]} */ |
wolfSSL | 15:117db924cf7c | 1527 | |
wolfSSL | 15:117db924cf7c | 1528 | #if defined(HAVE_INTEL_RORX) |
wolfSSL | 15:117db924cf7c | 1529 | |
wolfSSL | 15:117db924cf7c | 1530 | #define MsgSched_RORX(X0,X1,X2,X3,a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1531 | RND_STEP_RORX_0_1(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1532 | VPALIGNR (XTMP0, X3, X2, 4) \ |
wolfSSL | 15:117db924cf7c | 1533 | VPALIGNR (XTMP1, X1, X0, 4) /* XTMP1 = W[-15] */ \ |
wolfSSL | 15:117db924cf7c | 1534 | RND_STEP_RORX_0_2(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1535 | RND_STEP_RORX_0_3(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1536 | VPSRLD (XTMP2, XTMP1, 7) \ |
wolfSSL | 15:117db924cf7c | 1537 | VPSLLD (XTMP3, XTMP1, 25) /* VPSLLD (XTMP3, XTMP1, (32-7)) */ \ |
wolfSSL | 15:117db924cf7c | 1538 | RND_STEP_RORX_0_4(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1539 | RND_STEP_RORX_0_5(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1540 | VPSRLD (XTMP4, XTMP1, 3) /* XTMP4 = W[-15] >> 3 */ \ |
wolfSSL | 15:117db924cf7c | 1541 | VPOR (XTMP3, XTMP3, XTMP2) /* XTMP1 = W[-15] MY_ROR 7 */ \ |
wolfSSL | 15:117db924cf7c | 1542 | RND_STEP_RORX_0_6(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1543 | RND_STEP_RORX_0_7(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1544 | RND_STEP_RORX_0_8(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1545 | \ |
wolfSSL | 15:117db924cf7c | 1546 | RND_STEP_RORX_1_1(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1547 | VPSRLD (XTMP2, XTMP1,18) \ |
wolfSSL | 15:117db924cf7c | 1548 | RND_STEP_RORX_1_2(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1549 | VPSLLD (XTMP1, XTMP1, 14) /* VPSLLD (XTMP1, XTMP1, (32-18)) */ \ |
wolfSSL | 15:117db924cf7c | 1550 | RND_STEP_RORX_1_3(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1551 | VPXOR (XTMP3, XTMP3, XTMP1) \ |
wolfSSL | 15:117db924cf7c | 1552 | RND_STEP_RORX_1_4(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1553 | VPXOR (XTMP3, XTMP3, XTMP2) \ |
wolfSSL | 15:117db924cf7c | 1554 | /* XTMP1 = W[-15] MY_ROR 7 ^ W[-15] MY_ROR 18 */ \ |
wolfSSL | 15:117db924cf7c | 1555 | RND_STEP_RORX_1_5(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1556 | VPSHUFD (XTMP2, X3, 0b11111010) /* XTMP2 = W[-2] {BBAA}*/ \ |
wolfSSL | 15:117db924cf7c | 1557 | RND_STEP_RORX_1_6(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1558 | VPXOR (XTMP1, XTMP3, XTMP4) /* XTMP1 = s0 */ \ |
wolfSSL | 15:117db924cf7c | 1559 | RND_STEP_RORX_1_7(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1560 | VPSRLD (XTMP4, XTMP2, 10) /* XTMP4 = W[-2] >> 10 {BBAA} */ \ |
wolfSSL | 15:117db924cf7c | 1561 | RND_STEP_RORX_1_8(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1562 | \ |
wolfSSL | 15:117db924cf7c | 1563 | RND_STEP_RORX_0_1(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1564 | VPSRLQ (XTMP3, XTMP2, 19) /* XTMP3 = W[-2] MY_ROR 19 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1565 | RND_STEP_RORX_0_2(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1566 | VPSRLQ (XTMP2, XTMP2, 17) /* XTMP2 = W[-2] MY_ROR 17 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1567 | VPADDD (XTMP0, XTMP0, X0) \ |
wolfSSL | 15:117db924cf7c | 1568 | RND_STEP_RORX_0_3(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1569 | VPADDD (XTMP0, XTMP0, XTMP1) /* XTMP0 = W[-16] + W[-7] + s0 */ \ |
wolfSSL | 15:117db924cf7c | 1570 | RND_STEP_RORX_0_4(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1571 | VPXOR (XTMP2, XTMP2, XTMP3) \ |
wolfSSL | 15:117db924cf7c | 1572 | RND_STEP_RORX_0_5(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1573 | VPXOR (XTMP4, XTMP4, XTMP2) /* XTMP4 = s1 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1574 | RND_STEP_RORX_0_6(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1575 | VPSHUFB (XTMP4, XTMP4, SHUF_00BA) /* XTMP4 = s1 {00BA} */ \ |
wolfSSL | 15:117db924cf7c | 1576 | RND_STEP_RORX_0_7(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1577 | VPADDD (XTMP0, XTMP0, XTMP4) /* XTMP0 = {..., ..., W[1], W[0]} */ \ |
wolfSSL | 15:117db924cf7c | 1578 | RND_STEP_RORX_0_8(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1579 | \ |
wolfSSL | 15:117db924cf7c | 1580 | RND_STEP_RORX_1_1(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1581 | VPSHUFD (XTMP2, XTMP0, 0b01010000) /* XTMP2 = W[-2] {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 1582 | RND_STEP_RORX_1_2(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1583 | VPSRLD (XTMP5, XTMP2, 10) /* XTMP5 = W[-2] >> 10 {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 1584 | RND_STEP_RORX_1_3(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1585 | VPSRLQ (XTMP3, XTMP2, 19) /* XTMP3 = W[-2] MY_ROR 19 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1586 | RND_STEP_RORX_1_4(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1587 | VPSRLQ (XTMP2, XTMP2, 17) /* XTMP2 = W[-2] MY_ROR 17 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1588 | RND_STEP_RORX_1_5(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1589 | VPXOR (XTMP2, XTMP2, XTMP3) \ |
wolfSSL | 15:117db924cf7c | 1590 | RND_STEP_RORX_1_6(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1591 | VPXOR (XTMP5, XTMP5, XTMP2) /* XTMP5 = s1 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1592 | RND_STEP_RORX_1_7(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1593 | VPSHUFB (XTMP5, XTMP5, SHUF_DC00) /* XTMP5 = s1 {DC00} */ \ |
wolfSSL | 15:117db924cf7c | 1594 | RND_STEP_RORX_1_8(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1595 | VPADDD (X0, XTMP5, XTMP0) /* X0 = {W[3], W[2], W[1], W[0]} */ |
wolfSSL | 15:117db924cf7c | 1596 | |
wolfSSL | 15:117db924cf7c | 1597 | #endif /* HAVE_INTEL_RORX */ |
wolfSSL | 15:117db924cf7c | 1598 | |
wolfSSL | 15:117db924cf7c | 1599 | |
wolfSSL | 15:117db924cf7c | 1600 | #define _W_K_from_buff(X0, X1, X2, X3, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 1601 | "# X0, X1, X2, X3 = W[0..15]\n\t" \ |
wolfSSL | 15:117db924cf7c | 1602 | "vmovdqu (%%rax), %" #X0 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1603 | "vmovdqu 16(%%rax), %" #X1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1604 | VPSHUFB(X0, X0, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 1605 | VPSHUFB(X1, X1, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 1606 | "vmovdqu 32(%%rax), %" #X2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1607 | "vmovdqu 48(%%rax), %" #X3 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1608 | VPSHUFB(X2, X2, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 1609 | VPSHUFB(X3, X3, BYTE_FLIP_MASK) |
wolfSSL | 15:117db924cf7c | 1610 | |
wolfSSL | 15:117db924cf7c | 1611 | #define W_K_from_buff(X0, X1, X2, X3, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 1612 | _W_K_from_buff(X0, X1, X2, X3, BYTE_FLIP_MASK) |
wolfSSL | 15:117db924cf7c | 1613 | |
wolfSSL | 15:117db924cf7c | 1614 | |
wolfSSL | 15:117db924cf7c | 1615 | #define _SET_W_K_XFER_4(i) \ |
wolfSSL | 15:117db924cf7c | 1616 | "vpaddd (" #i "*4)+ 0+%[K], %%xmm0, %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 1617 | "vpaddd (" #i "*4)+16+%[K], %%xmm1, %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 1618 | "vmovdqu %%xmm4, (" WK ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 1619 | "vmovdqu %%xmm5, 16(" WK ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 1620 | "vpaddd (" #i "*4)+32+%[K], %%xmm2, %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 1621 | "vpaddd (" #i "*4)+48+%[K], %%xmm3, %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 1622 | "vmovdqu %%xmm6, 32(" WK ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 1623 | "vmovdqu %%xmm7, 48(" WK ")\n\t" |
wolfSSL | 15:117db924cf7c | 1624 | |
wolfSSL | 15:117db924cf7c | 1625 | #define SET_W_K_XFER_4(i) \ |
wolfSSL | 15:117db924cf7c | 1626 | _SET_W_K_XFER_4(i) |
wolfSSL | 15:117db924cf7c | 1627 | |
wolfSSL | 15:117db924cf7c | 1628 | |
wolfSSL | 15:117db924cf7c | 1629 | static const ALIGN32 word64 mSHUF_00BA[] = |
wolfSSL | 15:117db924cf7c | 1630 | { 0x0b0a090803020100, 0xFFFFFFFFFFFFFFFF }; /* shuffle xBxA -> 00BA */ |
wolfSSL | 15:117db924cf7c | 1631 | static const ALIGN32 word64 mSHUF_DC00[] = |
wolfSSL | 15:117db924cf7c | 1632 | { 0xFFFFFFFFFFFFFFFF, 0x0b0a090803020100 }; /* shuffle xDxC -> DC00 */ |
wolfSSL | 15:117db924cf7c | 1633 | static const ALIGN32 word64 mBYTE_FLIP_MASK[] = |
wolfSSL | 15:117db924cf7c | 1634 | { 0x0405060700010203, 0x0c0d0e0f08090a0b }; |
wolfSSL | 15:117db924cf7c | 1635 | |
wolfSSL | 15:117db924cf7c | 1636 | #define _Init_Masks(mask1, mask2, mask3) \ |
wolfSSL | 15:117db924cf7c | 1637 | "vmovdqa %[FLIP], %" #mask1 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1638 | "vmovdqa %[SHUF00BA], %" #mask2 "\n\t" \ |
wolfSSL | 15:117db924cf7c | 1639 | "vmovdqa %[SHUFDC00], %" #mask3 "\n\t" |
wolfSSL | 15:117db924cf7c | 1640 | |
wolfSSL | 15:117db924cf7c | 1641 | #define Init_Masks(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) \ |
wolfSSL | 15:117db924cf7c | 1642 | _Init_Masks(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) |
wolfSSL | 15:117db924cf7c | 1643 | |
wolfSSL | 15:117db924cf7c | 1644 | #define X0 %xmm0 |
wolfSSL | 15:117db924cf7c | 1645 | #define X1 %xmm1 |
wolfSSL | 15:117db924cf7c | 1646 | #define X2 %xmm2 |
wolfSSL | 15:117db924cf7c | 1647 | #define X3 %xmm3 |
wolfSSL | 15:117db924cf7c | 1648 | |
wolfSSL | 15:117db924cf7c | 1649 | #define XTMP0 %xmm4 |
wolfSSL | 15:117db924cf7c | 1650 | #define XTMP1 %xmm5 |
wolfSSL | 15:117db924cf7c | 1651 | #define XTMP2 %xmm6 |
wolfSSL | 15:117db924cf7c | 1652 | #define XTMP3 %xmm7 |
wolfSSL | 15:117db924cf7c | 1653 | #define XTMP4 %xmm8 |
wolfSSL | 15:117db924cf7c | 1654 | #define XTMP5 %xmm9 |
wolfSSL | 15:117db924cf7c | 1655 | #define XFER %xmm10 |
wolfSSL | 15:117db924cf7c | 1656 | |
wolfSSL | 15:117db924cf7c | 1657 | #define SHUF_00BA %xmm11 /* shuffle xBxA -> 00BA */ |
wolfSSL | 15:117db924cf7c | 1658 | #define SHUF_DC00 %xmm12 /* shuffle xDxC -> DC00 */ |
wolfSSL | 15:117db924cf7c | 1659 | #define BYTE_FLIP_MASK %xmm13 |
wolfSSL | 15:117db924cf7c | 1660 | |
wolfSSL | 15:117db924cf7c | 1661 | |
wolfSSL | 15:117db924cf7c | 1662 | SHA256_NOINLINE static int Transform_Sha256_AVX1(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 1663 | { |
wolfSSL | 15:117db924cf7c | 1664 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 1665 | |
wolfSSL | 15:117db924cf7c | 1666 | "subq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1667 | |
wolfSSL | 15:117db924cf7c | 1668 | "leaq 32(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1669 | Init_Masks(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) |
wolfSSL | 15:117db924cf7c | 1670 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1671 | |
wolfSSL | 15:117db924cf7c | 1672 | W_K_from_buff(X0, X1, X2, X3, BYTE_FLIP_MASK) |
wolfSSL | 15:117db924cf7c | 1673 | |
wolfSSL | 15:117db924cf7c | 1674 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1675 | "movl %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 1676 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1677 | |
wolfSSL | 15:117db924cf7c | 1678 | SET_W_K_XFER_4(0) |
wolfSSL | 15:117db924cf7c | 1679 | MsgSched(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1680 | MsgSched(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1681 | MsgSched(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1682 | MsgSched(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1683 | |
wolfSSL | 15:117db924cf7c | 1684 | SET_W_K_XFER_4(16) |
wolfSSL | 15:117db924cf7c | 1685 | MsgSched(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1686 | MsgSched(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1687 | MsgSched(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1688 | MsgSched(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1689 | |
wolfSSL | 15:117db924cf7c | 1690 | SET_W_K_XFER_4(32) |
wolfSSL | 15:117db924cf7c | 1691 | MsgSched(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1692 | MsgSched(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1693 | MsgSched(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1694 | MsgSched(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1695 | |
wolfSSL | 15:117db924cf7c | 1696 | SET_W_K_XFER_4(48) |
wolfSSL | 15:117db924cf7c | 1697 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1698 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1699 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1700 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1701 | |
wolfSSL | 15:117db924cf7c | 1702 | STORE_ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1703 | |
wolfSSL | 15:117db924cf7c | 1704 | "addq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1705 | |
wolfSSL | 15:117db924cf7c | 1706 | : |
wolfSSL | 15:117db924cf7c | 1707 | : [FLIP] "m" (mBYTE_FLIP_MASK[0]), |
wolfSSL | 15:117db924cf7c | 1708 | [SHUF00BA] "m" (mSHUF_00BA[0]), |
wolfSSL | 15:117db924cf7c | 1709 | [SHUFDC00] "m" (mSHUF_DC00[0]), |
wolfSSL | 15:117db924cf7c | 1710 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 1711 | [K] "m" (K) |
wolfSSL | 15:117db924cf7c | 1712 | : WORK_REGS, STATE_REGS, XMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 1713 | ); |
wolfSSL | 15:117db924cf7c | 1714 | |
wolfSSL | 15:117db924cf7c | 1715 | return 0; |
wolfSSL | 15:117db924cf7c | 1716 | } |
wolfSSL | 15:117db924cf7c | 1717 | |
wolfSSL | 15:117db924cf7c | 1718 | SHA256_NOINLINE static int Transform_Sha256_AVX1_Len(wc_Sha256* sha256, |
wolfSSL | 15:117db924cf7c | 1719 | word32 len) |
wolfSSL | 15:117db924cf7c | 1720 | { |
wolfSSL | 15:117db924cf7c | 1721 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 1722 | |
wolfSSL | 15:117db924cf7c | 1723 | "subq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1724 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1725 | |
wolfSSL | 15:117db924cf7c | 1726 | Init_Masks(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) |
wolfSSL | 15:117db924cf7c | 1727 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1728 | |
wolfSSL | 15:117db924cf7c | 1729 | "# Start of loop processing a block\n" |
wolfSSL | 15:117db924cf7c | 1730 | "1:\n\t" |
wolfSSL | 15:117db924cf7c | 1731 | |
wolfSSL | 15:117db924cf7c | 1732 | W_K_from_buff(X0, X1, X2, X3, BYTE_FLIP_MASK) |
wolfSSL | 15:117db924cf7c | 1733 | |
wolfSSL | 15:117db924cf7c | 1734 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1735 | "movl %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 1736 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1737 | |
wolfSSL | 15:117db924cf7c | 1738 | SET_W_K_XFER_4(0) |
wolfSSL | 15:117db924cf7c | 1739 | MsgSched(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1740 | MsgSched(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1741 | MsgSched(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1742 | MsgSched(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1743 | |
wolfSSL | 15:117db924cf7c | 1744 | SET_W_K_XFER_4(16) |
wolfSSL | 15:117db924cf7c | 1745 | MsgSched(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1746 | MsgSched(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1747 | MsgSched(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1748 | MsgSched(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1749 | |
wolfSSL | 15:117db924cf7c | 1750 | SET_W_K_XFER_4(32) |
wolfSSL | 15:117db924cf7c | 1751 | MsgSched(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1752 | MsgSched(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1753 | MsgSched(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1754 | MsgSched(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1755 | |
wolfSSL | 15:117db924cf7c | 1756 | SET_W_K_XFER_4(48) |
wolfSSL | 15:117db924cf7c | 1757 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1758 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1759 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1760 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1761 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1762 | |
wolfSSL | 15:117db924cf7c | 1763 | ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1764 | |
wolfSSL | 15:117db924cf7c | 1765 | "addq $64, %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1766 | "subl $64, %[len]\n\t" |
wolfSSL | 15:117db924cf7c | 1767 | |
wolfSSL | 15:117db924cf7c | 1768 | STORE_DIGEST() |
wolfSSL | 15:117db924cf7c | 1769 | |
wolfSSL | 15:117db924cf7c | 1770 | "movq %%rax, 120(%[sha256])\n\t" |
wolfSSL | 15:117db924cf7c | 1771 | "jnz 1b\n\t" |
wolfSSL | 15:117db924cf7c | 1772 | |
wolfSSL | 15:117db924cf7c | 1773 | "addq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1774 | |
wolfSSL | 15:117db924cf7c | 1775 | : |
wolfSSL | 15:117db924cf7c | 1776 | : [FLIP] "m" (mBYTE_FLIP_MASK[0]), |
wolfSSL | 15:117db924cf7c | 1777 | [SHUF00BA] "m" (mSHUF_00BA[0]), |
wolfSSL | 15:117db924cf7c | 1778 | [SHUFDC00] "m" (mSHUF_DC00[0]), |
wolfSSL | 15:117db924cf7c | 1779 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 1780 | [len] "r" (len), |
wolfSSL | 15:117db924cf7c | 1781 | [K] "m" (K) |
wolfSSL | 15:117db924cf7c | 1782 | : WORK_REGS, STATE_REGS, XMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 1783 | ); |
wolfSSL | 15:117db924cf7c | 1784 | |
wolfSSL | 15:117db924cf7c | 1785 | return 0; |
wolfSSL | 15:117db924cf7c | 1786 | } |
wolfSSL | 15:117db924cf7c | 1787 | #endif /* HAVE_INTEL_AVX1 */ |
wolfSSL | 15:117db924cf7c | 1788 | |
wolfSSL | 15:117db924cf7c | 1789 | #if defined(HAVE_INTEL_AVX2) && defined(HAVE_INTEL_RORX) |
wolfSSL | 15:117db924cf7c | 1790 | SHA256_NOINLINE static int Transform_Sha256_AVX1_RORX(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 1791 | { |
wolfSSL | 15:117db924cf7c | 1792 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 1793 | |
wolfSSL | 15:117db924cf7c | 1794 | "subq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1795 | |
wolfSSL | 15:117db924cf7c | 1796 | Init_Masks(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) |
wolfSSL | 15:117db924cf7c | 1797 | "leaq 32(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1798 | W_K_from_buff(X0, X1, X2, X3, BYTE_FLIP_MASK) |
wolfSSL | 15:117db924cf7c | 1799 | |
wolfSSL | 15:117db924cf7c | 1800 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1801 | |
wolfSSL | 15:117db924cf7c | 1802 | SET_W_K_XFER_4(0) |
wolfSSL | 15:117db924cf7c | 1803 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1804 | "rorx $6, %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 1805 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1806 | MsgSched_RORX(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1807 | MsgSched_RORX(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1808 | MsgSched_RORX(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1809 | MsgSched_RORX(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1810 | |
wolfSSL | 15:117db924cf7c | 1811 | SET_W_K_XFER_4(16) |
wolfSSL | 15:117db924cf7c | 1812 | MsgSched_RORX(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1813 | MsgSched_RORX(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1814 | MsgSched_RORX(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1815 | MsgSched_RORX(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1816 | |
wolfSSL | 15:117db924cf7c | 1817 | SET_W_K_XFER_4(32) |
wolfSSL | 15:117db924cf7c | 1818 | MsgSched_RORX(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1819 | MsgSched_RORX(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1820 | MsgSched_RORX(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1821 | MsgSched_RORX(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1822 | |
wolfSSL | 15:117db924cf7c | 1823 | SET_W_K_XFER_4(48) |
wolfSSL | 15:117db924cf7c | 1824 | "xorl " L3 ", " L3 "\n\t" |
wolfSSL | 15:117db924cf7c | 1825 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1826 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1827 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1828 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1829 | /* Prev RND: h += Maj(a,b,c) */ |
wolfSSL | 15:117db924cf7c | 1830 | "addl " L3 ", %%r8d\n\t" |
wolfSSL | 15:117db924cf7c | 1831 | |
wolfSSL | 15:117db924cf7c | 1832 | STORE_ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1833 | |
wolfSSL | 15:117db924cf7c | 1834 | "addq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1835 | |
wolfSSL | 15:117db924cf7c | 1836 | : |
wolfSSL | 15:117db924cf7c | 1837 | : [FLIP] "m" (mBYTE_FLIP_MASK[0]), |
wolfSSL | 15:117db924cf7c | 1838 | [SHUF00BA] "m" (mSHUF_00BA[0]), |
wolfSSL | 15:117db924cf7c | 1839 | [SHUFDC00] "m" (mSHUF_DC00[0]), |
wolfSSL | 15:117db924cf7c | 1840 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 1841 | [K] "m" (K) |
wolfSSL | 15:117db924cf7c | 1842 | : WORK_REGS, STATE_REGS, XMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 1843 | ); |
wolfSSL | 15:117db924cf7c | 1844 | |
wolfSSL | 15:117db924cf7c | 1845 | return 0; |
wolfSSL | 15:117db924cf7c | 1846 | } |
wolfSSL | 15:117db924cf7c | 1847 | |
wolfSSL | 15:117db924cf7c | 1848 | SHA256_NOINLINE static int Transform_Sha256_AVX1_RORX_Len(wc_Sha256* sha256, |
wolfSSL | 15:117db924cf7c | 1849 | word32 len) |
wolfSSL | 15:117db924cf7c | 1850 | { |
wolfSSL | 15:117db924cf7c | 1851 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 1852 | |
wolfSSL | 15:117db924cf7c | 1853 | "subq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1854 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1855 | |
wolfSSL | 15:117db924cf7c | 1856 | Init_Masks(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) |
wolfSSL | 15:117db924cf7c | 1857 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1858 | |
wolfSSL | 15:117db924cf7c | 1859 | "# Start of loop processing a block\n" |
wolfSSL | 15:117db924cf7c | 1860 | "1:\n\t" |
wolfSSL | 15:117db924cf7c | 1861 | |
wolfSSL | 15:117db924cf7c | 1862 | W_K_from_buff(X0, X1, X2, X3, BYTE_FLIP_MASK) |
wolfSSL | 15:117db924cf7c | 1863 | |
wolfSSL | 15:117db924cf7c | 1864 | SET_W_K_XFER_4(0) |
wolfSSL | 15:117db924cf7c | 1865 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1866 | "rorx $6, %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 1867 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 1868 | MsgSched_RORX(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1869 | MsgSched_RORX(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1870 | MsgSched_RORX(X2, X3, X0, X1, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 8) |
wolfSSL | 15:117db924cf7c | 1871 | MsgSched_RORX(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1872 | |
wolfSSL | 15:117db924cf7c | 1873 | SET_W_K_XFER_4(16) |
wolfSSL | 15:117db924cf7c | 1874 | MsgSched_RORX(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1875 | MsgSched_RORX(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1876 | MsgSched_RORX(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1877 | MsgSched_RORX(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1878 | |
wolfSSL | 15:117db924cf7c | 1879 | SET_W_K_XFER_4(32) |
wolfSSL | 15:117db924cf7c | 1880 | MsgSched_RORX(X0, X1, X2, X3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1881 | MsgSched_RORX(X1, X2, X3, X0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1882 | MsgSched_RORX(X2, X3, X0, X1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1883 | MsgSched_RORX(X3, X0, X1, X2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1884 | |
wolfSSL | 15:117db924cf7c | 1885 | SET_W_K_XFER_4(48) |
wolfSSL | 15:117db924cf7c | 1886 | "xorl " L3 ", " L3 "\n\t" |
wolfSSL | 15:117db924cf7c | 1887 | "xorl " L2 ", " L2 "\n\t" |
wolfSSL | 15:117db924cf7c | 1888 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 1889 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 4) |
wolfSSL | 15:117db924cf7c | 1890 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 8) |
wolfSSL | 15:117db924cf7c | 1891 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 1892 | /* Prev RND: h += Maj(a,b,c) */ |
wolfSSL | 15:117db924cf7c | 1893 | "addl " L3 ", %%r8d\n\t" |
wolfSSL | 15:117db924cf7c | 1894 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1895 | |
wolfSSL | 15:117db924cf7c | 1896 | ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 1897 | |
wolfSSL | 15:117db924cf7c | 1898 | "addq $64, %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 1899 | "subl $64, %[len]\n\t" |
wolfSSL | 15:117db924cf7c | 1900 | |
wolfSSL | 15:117db924cf7c | 1901 | STORE_DIGEST() |
wolfSSL | 15:117db924cf7c | 1902 | |
wolfSSL | 15:117db924cf7c | 1903 | "movq %%rax, 120(%[sha256])\n\t" |
wolfSSL | 15:117db924cf7c | 1904 | "jnz 1b\n\t" |
wolfSSL | 15:117db924cf7c | 1905 | |
wolfSSL | 15:117db924cf7c | 1906 | "addq $64, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 1907 | |
wolfSSL | 15:117db924cf7c | 1908 | : |
wolfSSL | 15:117db924cf7c | 1909 | : [FLIP] "m" (mBYTE_FLIP_MASK[0]), |
wolfSSL | 15:117db924cf7c | 1910 | [SHUF00BA] "m" (mSHUF_00BA[0]), |
wolfSSL | 15:117db924cf7c | 1911 | [SHUFDC00] "m" (mSHUF_DC00[0]), |
wolfSSL | 15:117db924cf7c | 1912 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 1913 | [len] "r" (len), |
wolfSSL | 15:117db924cf7c | 1914 | [K] "m" (K) |
wolfSSL | 15:117db924cf7c | 1915 | : WORK_REGS, STATE_REGS, XMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 1916 | ); |
wolfSSL | 15:117db924cf7c | 1917 | |
wolfSSL | 15:117db924cf7c | 1918 | return 0; |
wolfSSL | 15:117db924cf7c | 1919 | } |
wolfSSL | 15:117db924cf7c | 1920 | #endif /* HAVE_INTEL_AVX2 && HAVE_INTEL_RORX */ |
wolfSSL | 15:117db924cf7c | 1921 | |
wolfSSL | 15:117db924cf7c | 1922 | |
wolfSSL | 15:117db924cf7c | 1923 | #if defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 1924 | #define Y0 %ymm0 |
wolfSSL | 15:117db924cf7c | 1925 | #define Y1 %ymm1 |
wolfSSL | 15:117db924cf7c | 1926 | #define Y2 %ymm2 |
wolfSSL | 15:117db924cf7c | 1927 | #define Y3 %ymm3 |
wolfSSL | 15:117db924cf7c | 1928 | |
wolfSSL | 15:117db924cf7c | 1929 | #define YTMP0 %ymm4 |
wolfSSL | 15:117db924cf7c | 1930 | #define YTMP1 %ymm5 |
wolfSSL | 15:117db924cf7c | 1931 | #define YTMP2 %ymm6 |
wolfSSL | 15:117db924cf7c | 1932 | #define YTMP3 %ymm7 |
wolfSSL | 15:117db924cf7c | 1933 | #define YTMP4 %ymm8 |
wolfSSL | 15:117db924cf7c | 1934 | #define YTMP5 %ymm9 |
wolfSSL | 15:117db924cf7c | 1935 | #define YXFER %ymm10 |
wolfSSL | 15:117db924cf7c | 1936 | |
wolfSSL | 15:117db924cf7c | 1937 | #define SHUF_Y_00BA %ymm11 /* shuffle xBxA -> 00BA */ |
wolfSSL | 15:117db924cf7c | 1938 | #define SHUF_Y_DC00 %ymm12 /* shuffle xDxC -> DC00 */ |
wolfSSL | 15:117db924cf7c | 1939 | #define BYTE_FLIP_Y_MASK %ymm13 |
wolfSSL | 15:117db924cf7c | 1940 | |
wolfSSL | 15:117db924cf7c | 1941 | #define YMM_REGS "ymm0", "ymm1", "ymm2", "ymm3", "ymm4", "ymm5", "ymm6", \ |
wolfSSL | 15:117db924cf7c | 1942 | "ymm7", "ymm8", "ymm9", "ymm10", "ymm11", "ymm12", "ymm13" |
wolfSSL | 15:117db924cf7c | 1943 | |
wolfSSL | 15:117db924cf7c | 1944 | #define MsgSched_Y(Y0,Y1,Y2,Y3,a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1945 | RND_STEP_0_1(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1946 | VPALIGNR (YTMP1, Y1, Y0, 4) /* YTMP1 = W[-15] */ \ |
wolfSSL | 15:117db924cf7c | 1947 | VPALIGNR (YTMP0, Y3, Y2, 4) /* YTMP0 = W[-7] */ \ |
wolfSSL | 15:117db924cf7c | 1948 | RND_STEP_0_2(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1949 | RND_STEP_0_3(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1950 | VPSRLD (YTMP2, YTMP1, 7) /* YTMP2 = W[-15] >> 7 */ \ |
wolfSSL | 15:117db924cf7c | 1951 | VPSLLD (YTMP3, YTMP1, 25) /* YTEMP3 = W[-15] << (32-7) */ \ |
wolfSSL | 15:117db924cf7c | 1952 | RND_STEP_0_4(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1953 | RND_STEP_0_5(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1954 | VPSRLD (YTMP4, YTMP1, 18) /* YTEMP4 = W[-15] >> 18 */ \ |
wolfSSL | 15:117db924cf7c | 1955 | VPSLLD (YTMP5, YTMP1, 14) /* YTEMP5 = W[-15] << (32-18) */ \ |
wolfSSL | 15:117db924cf7c | 1956 | RND_STEP_0_6(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1957 | RND_STEP_0_7(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1958 | VPOR (YTMP2, YTMP3, YTMP2) /* YTMP2 = W[-15] >>> 7 */ \ |
wolfSSL | 15:117db924cf7c | 1959 | VPOR (YTMP4, YTMP5, YTMP4) /* YTMP4 = W[-15] >>> 18 */ \ |
wolfSSL | 15:117db924cf7c | 1960 | RND_STEP_0_8(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 1961 | RND_STEP_1_1(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1962 | RND_STEP_1_2(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1963 | VPSRLD (YTMP5, YTMP1, 3) /* YTMP4 = W[-15] >> 3 */ \ |
wolfSSL | 15:117db924cf7c | 1964 | VPXOR (YTMP2, YTMP4, YTMP2) /* YTMP2 = W[-15] >>> 7 ^ W[-15] >>> 18 */ \ |
wolfSSL | 15:117db924cf7c | 1965 | RND_STEP_1_3(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1966 | RND_STEP_1_4(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1967 | VPXOR (YTMP1, YTMP5, YTMP2) /* YTMP1 = s0 */ \ |
wolfSSL | 15:117db924cf7c | 1968 | VPSHUFD (YTMP2, Y3, 0b11111010) /* YTMP2 = W[-2] {BBAA}*/ \ |
wolfSSL | 15:117db924cf7c | 1969 | RND_STEP_1_5(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1970 | RND_STEP_1_6(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1971 | VPSRLD (YTMP4, YTMP2, 10) /* YTMP4 = W[-2] >> 10 {BBAA} */ \ |
wolfSSL | 15:117db924cf7c | 1972 | VPSRLQ (YTMP3, YTMP2, 19) /* YTMP3 = W[-2] MY_ROR 19 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1973 | RND_STEP_1_7(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1974 | RND_STEP_1_8(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 1975 | RND_STEP_0_1(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1976 | VPSRLQ (YTMP2, YTMP2, 17) /* YTMP2 = W[-2] MY_ROR 17 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1977 | VPADDD (YTMP0, YTMP0, Y0) \ |
wolfSSL | 15:117db924cf7c | 1978 | RND_STEP_0_2(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1979 | RND_STEP_0_3(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1980 | RND_STEP_0_4(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1981 | VPXOR (YTMP2, YTMP3, YTMP2) \ |
wolfSSL | 15:117db924cf7c | 1982 | VPADDD (YTMP0, YTMP0, YTMP1) /* YTMP0 = W[-16] + W[-7] + s0 */ \ |
wolfSSL | 15:117db924cf7c | 1983 | RND_STEP_0_5(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1984 | VPXOR (YTMP4, YTMP4, YTMP2) /* YTMP4 = s1 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 1985 | RND_STEP_0_6(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1986 | VPSHUFB (YTMP4, YTMP4, SHUF_Y_00BA) /* YTMP4 = s1 {00BA} */ \ |
wolfSSL | 15:117db924cf7c | 1987 | RND_STEP_0_7(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1988 | VPADDD (YTMP0, YTMP0, YTMP4) /* YTMP0 = {..., ..., W[1], W[0]} */ \ |
wolfSSL | 15:117db924cf7c | 1989 | RND_STEP_0_8(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 1990 | RND_STEP_1_1(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1991 | VPSHUFD (YTMP2, YTMP0, 0b01010000) /* YTMP2 = W[-2] {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 1992 | RND_STEP_1_2(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1993 | VPSRLQ (YTMP4, YTMP2, 17) /* YTMP4 = W[-2] MY_ROR 17 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1994 | VPSRLQ (YTMP3, YTMP2, 19) /* YTMP3 = W[-2] MY_ROR 19 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 1995 | RND_STEP_1_3(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1996 | RND_STEP_1_4(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 1997 | VPSRLD (YTMP5, YTMP2, 10) /* YTMP5 = W[-2] >> 10 {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 1998 | VPXOR (YTMP4, YTMP3, YTMP4) \ |
wolfSSL | 15:117db924cf7c | 1999 | RND_STEP_1_5(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2000 | RND_STEP_1_6(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2001 | VPXOR (YTMP5, YTMP4, YTMP5) /* YTMP5 = s1 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 2002 | RND_STEP_1_7(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2003 | VPSHUFB (YTMP5, YTMP5, SHUF_Y_DC00) /* YTMP5 = s1 {DC00} */ \ |
wolfSSL | 15:117db924cf7c | 2004 | RND_STEP_1_8(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2005 | VPADDD (Y0, YTMP5, YTMP0) /* Y0 = {W[3], W[2], W[1], W[0]} */ |
wolfSSL | 15:117db924cf7c | 2006 | |
wolfSSL | 15:117db924cf7c | 2007 | #if defined(HAVE_INTEL_RORX) |
wolfSSL | 15:117db924cf7c | 2008 | |
wolfSSL | 15:117db924cf7c | 2009 | #define MsgSched_Y_RORX(Y0,Y1,Y2,Y3,a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2010 | RND_STEP_RORX_0_1(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2011 | VPALIGNR (YTMP1, Y1, Y0, 4) /* YTMP1 = W[-15] */ \ |
wolfSSL | 15:117db924cf7c | 2012 | RND_STEP_RORX_0_2(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2013 | VPALIGNR (YTMP0, Y3, Y2, 4) /* YTMP0 = W[-7] */ \ |
wolfSSL | 15:117db924cf7c | 2014 | RND_STEP_RORX_0_3(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2015 | VPSRLD (YTMP2, YTMP1, 7) /* YTMP2 = W[-15] >> 7 */ \ |
wolfSSL | 15:117db924cf7c | 2016 | RND_STEP_RORX_0_4(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2017 | VPSLLD (YTMP3, YTMP1, 25) /* YTEMP3 = W[-15] << (32-7) */ \ |
wolfSSL | 15:117db924cf7c | 2018 | RND_STEP_RORX_0_5(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2019 | VPSRLD (YTMP4, YTMP1, 18) /* YTEMP4 = W[-15] >> 18 */ \ |
wolfSSL | 15:117db924cf7c | 2020 | RND_STEP_RORX_0_6(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2021 | VPSLLD (YTMP5, YTMP1, 14) /* YTEMP5 = W[-15] << (32-18) */ \ |
wolfSSL | 15:117db924cf7c | 2022 | RND_STEP_RORX_0_7(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2023 | VPOR (YTMP2, YTMP2, YTMP3) /* YTMP2 = W[-15] >>> 7 */ \ |
wolfSSL | 15:117db924cf7c | 2024 | RND_STEP_RORX_0_8(a,b,c,d,e,f,g,h,_i) \ |
wolfSSL | 15:117db924cf7c | 2025 | VPOR (YTMP4, YTMP4, YTMP5) /* YTMP4 = W[-15] >>> 18 */ \ |
wolfSSL | 15:117db924cf7c | 2026 | RND_STEP_RORX_1_1(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2027 | VPSRLD (YTMP5, YTMP1, 3) /* YTMP4 = W[-15] >> 3 */ \ |
wolfSSL | 15:117db924cf7c | 2028 | RND_STEP_RORX_1_2(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2029 | VPXOR (YTMP2, YTMP2, YTMP4) /* YTMP2 = W[-15] >>> 7 ^ W[-15] >>> 18 */ \ |
wolfSSL | 15:117db924cf7c | 2030 | RND_STEP_RORX_1_3(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2031 | VPSHUFD (YTMP3, Y3, 0b11111010) /* YTMP2 = W[-2] {BBAA}*/ \ |
wolfSSL | 15:117db924cf7c | 2032 | RND_STEP_RORX_1_4(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2033 | VPXOR (YTMP1, YTMP5, YTMP2) /* YTMP1 = s0 */ \ |
wolfSSL | 15:117db924cf7c | 2034 | RND_STEP_RORX_1_5(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2035 | VPSRLD (YTMP4, YTMP3, 10) /* YTMP4 = W[-2] >> 10 {BBAA} */ \ |
wolfSSL | 15:117db924cf7c | 2036 | RND_STEP_RORX_1_6(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2037 | VPSRLQ (YTMP2, YTMP3, 19) /* YTMP3 = W[-2] MY_ROR 19 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 2038 | RND_STEP_RORX_1_7(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2039 | VPSRLQ (YTMP3, YTMP3, 17) /* YTMP2 = W[-2] MY_ROR 17 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 2040 | RND_STEP_RORX_1_8(h,a,b,c,d,e,f,g,_i+1) \ |
wolfSSL | 15:117db924cf7c | 2041 | VPADDD (YTMP0, YTMP0, Y0) \ |
wolfSSL | 15:117db924cf7c | 2042 | RND_STEP_RORX_0_1(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2043 | VPXOR (YTMP2, YTMP2, YTMP3) \ |
wolfSSL | 15:117db924cf7c | 2044 | RND_STEP_RORX_0_2(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2045 | VPXOR (YTMP4, YTMP4, YTMP2) /* YTMP4 = s1 {xBxA} */ \ |
wolfSSL | 15:117db924cf7c | 2046 | RND_STEP_RORX_0_3(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2047 | VPADDD (YTMP0, YTMP0, YTMP1) /* YTMP0 = W[-16] + W[-7] + s0 */ \ |
wolfSSL | 15:117db924cf7c | 2048 | RND_STEP_RORX_0_4(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2049 | VPSHUFB (YTMP4, YTMP4, SHUF_Y_00BA) /* YTMP4 = s1 {00BA} */ \ |
wolfSSL | 15:117db924cf7c | 2050 | RND_STEP_RORX_0_5(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2051 | VPADDD (YTMP0, YTMP0, YTMP4) /* YTMP0 = {..., ..., W[1], W[0]} */ \ |
wolfSSL | 15:117db924cf7c | 2052 | RND_STEP_RORX_0_6(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2053 | VPSHUFD (YTMP2, YTMP0, 0b01010000) /* YTMP2 = W[-2] {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 2054 | RND_STEP_RORX_0_7(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2055 | RND_STEP_RORX_0_8(g,h,a,b,c,d,e,f,_i+2) \ |
wolfSSL | 15:117db924cf7c | 2056 | VPSRLQ (YTMP4, YTMP2, 17) /* YTMP4 = W[-2] MY_ROR 17 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 2057 | RND_STEP_RORX_1_1(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2058 | VPSRLQ (YTMP3, YTMP2, 19) /* YTMP3 = W[-2] MY_ROR 19 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 2059 | RND_STEP_RORX_1_2(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2060 | VPSRLD (YTMP5, YTMP2, 10) /* YTMP5 = W[-2] >> 10 {DDCC} */ \ |
wolfSSL | 15:117db924cf7c | 2061 | RND_STEP_RORX_1_3(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2062 | VPXOR (YTMP4, YTMP4, YTMP3) \ |
wolfSSL | 15:117db924cf7c | 2063 | RND_STEP_RORX_1_4(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2064 | VPXOR (YTMP5, YTMP5, YTMP4) /* YTMP5 = s1 {xDxC} */ \ |
wolfSSL | 15:117db924cf7c | 2065 | RND_STEP_RORX_1_5(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2066 | RND_STEP_RORX_1_6(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2067 | VPSHUFB (YTMP5, YTMP5, SHUF_Y_DC00) /* YTMP5 = s1 {DC00} */ \ |
wolfSSL | 15:117db924cf7c | 2068 | RND_STEP_RORX_1_7(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2069 | RND_STEP_RORX_1_8(f,g,h,a,b,c,d,e,_i+3) \ |
wolfSSL | 15:117db924cf7c | 2070 | VPADDD (Y0, YTMP5, YTMP0) /* Y0 = {W[3], W[2], W[1], W[0]} */ \ |
wolfSSL | 15:117db924cf7c | 2071 | |
wolfSSL | 15:117db924cf7c | 2072 | #endif /* HAVE_INTEL_RORX */ |
wolfSSL | 15:117db924cf7c | 2073 | |
wolfSSL | 15:117db924cf7c | 2074 | #define _VINSERTI128(op1,op2,op3,op4) \ |
wolfSSL | 15:117db924cf7c | 2075 | "vinserti128 $" #op4 ", %" #op3 ", %" #op2 ", %" #op1 "\n\t" |
wolfSSL | 15:117db924cf7c | 2076 | #define VINSERTI128(op1,op2,op3,op4) \ |
wolfSSL | 15:117db924cf7c | 2077 | _VINSERTI128(op1,op2,op3,op4) |
wolfSSL | 15:117db924cf7c | 2078 | |
wolfSSL | 15:117db924cf7c | 2079 | |
wolfSSL | 15:117db924cf7c | 2080 | #define _LOAD_W_K_LOW(BYTE_FLIP_MASK, reg) \ |
wolfSSL | 15:117db924cf7c | 2081 | "# X0, X1, X2, X3 = W[0..15]\n\t" \ |
wolfSSL | 15:117db924cf7c | 2082 | "vmovdqu (%%" #reg "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 2083 | "vmovdqu 16(%%" #reg "), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 2084 | VPSHUFB(X0, X0, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 2085 | VPSHUFB(X1, X1, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 2086 | "vmovdqu 32(%%" #reg "), %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 2087 | "vmovdqu 48(%%" #reg "), %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 2088 | VPSHUFB(X2, X2, BYTE_FLIP_MASK) \ |
wolfSSL | 15:117db924cf7c | 2089 | VPSHUFB(X3, X3, BYTE_FLIP_MASK) |
wolfSSL | 15:117db924cf7c | 2090 | |
wolfSSL | 15:117db924cf7c | 2091 | #define LOAD_W_K_LOW(BYTE_FLIP_MASK, reg) \ |
wolfSSL | 15:117db924cf7c | 2092 | _LOAD_W_K_LOW(BYTE_FLIP_MASK, reg) |
wolfSSL | 15:117db924cf7c | 2093 | |
wolfSSL | 15:117db924cf7c | 2094 | |
wolfSSL | 15:117db924cf7c | 2095 | #define _LOAD_W_K(BYTE_FLIP_Y_MASK, reg) \ |
wolfSSL | 15:117db924cf7c | 2096 | "# X0, X1, X2, X3 = W[0..15]\n\t" \ |
wolfSSL | 15:117db924cf7c | 2097 | "vmovdqu (%%" #reg "), %%xmm0\n\t" \ |
wolfSSL | 15:117db924cf7c | 2098 | "vmovdqu 16(%%" #reg "), %%xmm1\n\t" \ |
wolfSSL | 15:117db924cf7c | 2099 | "vmovdqu 64(%%" #reg "), %%xmm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 2100 | "vmovdqu 80(%%" #reg "), %%xmm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 2101 | VINSERTI128(Y0, Y0, XTMP0, 1) \ |
wolfSSL | 15:117db924cf7c | 2102 | VINSERTI128(Y1, Y1, XTMP1, 1) \ |
wolfSSL | 15:117db924cf7c | 2103 | VPSHUFB(Y0, Y0, BYTE_FLIP_Y_MASK) \ |
wolfSSL | 15:117db924cf7c | 2104 | VPSHUFB(Y1, Y1, BYTE_FLIP_Y_MASK) \ |
wolfSSL | 15:117db924cf7c | 2105 | "vmovdqu 32(%%" #reg "), %%xmm2\n\t" \ |
wolfSSL | 15:117db924cf7c | 2106 | "vmovdqu 48(%%" #reg "), %%xmm3\n\t" \ |
wolfSSL | 15:117db924cf7c | 2107 | "vmovdqu 96(%%" #reg "), %%xmm6\n\t" \ |
wolfSSL | 15:117db924cf7c | 2108 | "vmovdqu 112(%%" #reg "), %%xmm7\n\t" \ |
wolfSSL | 15:117db924cf7c | 2109 | VINSERTI128(Y2, Y2, XTMP2, 1) \ |
wolfSSL | 15:117db924cf7c | 2110 | VINSERTI128(Y3, Y3, XTMP3, 1) \ |
wolfSSL | 15:117db924cf7c | 2111 | VPSHUFB(Y2, Y2, BYTE_FLIP_Y_MASK) \ |
wolfSSL | 15:117db924cf7c | 2112 | VPSHUFB(Y3, Y3, BYTE_FLIP_Y_MASK) |
wolfSSL | 15:117db924cf7c | 2113 | |
wolfSSL | 15:117db924cf7c | 2114 | #define LOAD_W_K(BYTE_FLIP_Y_MASK, reg) \ |
wolfSSL | 15:117db924cf7c | 2115 | _LOAD_W_K(BYTE_FLIP_Y_MASK, reg) |
wolfSSL | 15:117db924cf7c | 2116 | |
wolfSSL | 15:117db924cf7c | 2117 | |
wolfSSL | 15:117db924cf7c | 2118 | #define _SET_W_Y_4(i) \ |
wolfSSL | 15:117db924cf7c | 2119 | "vpaddd (" #i "*8)+ 0+%[K], %%ymm0, %%ymm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 2120 | "vpaddd (" #i "*8)+32+%[K], %%ymm1, %%ymm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 2121 | "vmovdqu %%ymm4, (" #i "*8)+ 0(" WK ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 2122 | "vmovdqu %%ymm5, (" #i "*8)+32(" WK ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 2123 | "vpaddd (" #i "*8)+64+%[K], %%ymm2, %%ymm4\n\t" \ |
wolfSSL | 15:117db924cf7c | 2124 | "vpaddd (" #i "*8)+96+%[K], %%ymm3, %%ymm5\n\t" \ |
wolfSSL | 15:117db924cf7c | 2125 | "vmovdqu %%ymm4, (" #i "*8)+64(" WK ")\n\t" \ |
wolfSSL | 15:117db924cf7c | 2126 | "vmovdqu %%ymm5, (" #i "*8)+96(" WK ")\n\t" |
wolfSSL | 15:117db924cf7c | 2127 | |
wolfSSL | 15:117db924cf7c | 2128 | #define SET_W_Y_4(i) \ |
wolfSSL | 15:117db924cf7c | 2129 | _SET_W_Y_4(i) |
wolfSSL | 15:117db924cf7c | 2130 | |
wolfSSL | 15:117db924cf7c | 2131 | |
wolfSSL | 15:117db924cf7c | 2132 | static const ALIGN32 word64 mSHUF_Y_00BA[] = |
wolfSSL | 15:117db924cf7c | 2133 | { 0x0b0a090803020100, 0xFFFFFFFFFFFFFFFF, |
wolfSSL | 15:117db924cf7c | 2134 | 0x0b0a090803020100, 0xFFFFFFFFFFFFFFFF }; /* shuffle xBxA -> 00BA */ |
wolfSSL | 15:117db924cf7c | 2135 | static const ALIGN32 word64 mSHUF_Y_DC00[] = |
wolfSSL | 15:117db924cf7c | 2136 | { 0xFFFFFFFFFFFFFFFF, 0x0b0a090803020100, |
wolfSSL | 15:117db924cf7c | 2137 | 0xFFFFFFFFFFFFFFFF, 0x0b0a090803020100 }; /* shuffle xDxC -> DC00 */ |
wolfSSL | 15:117db924cf7c | 2138 | static const ALIGN32 word64 mBYTE_FLIP_Y_MASK[] = |
wolfSSL | 15:117db924cf7c | 2139 | { 0x0405060700010203, 0x0c0d0e0f08090a0b, |
wolfSSL | 15:117db924cf7c | 2140 | 0x0405060700010203, 0x0c0d0e0f08090a0b }; |
wolfSSL | 15:117db924cf7c | 2141 | |
wolfSSL | 15:117db924cf7c | 2142 | #define _INIT_MASKS_Y(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) \ |
wolfSSL | 15:117db924cf7c | 2143 | "vmovdqa %[FLIP], %" #BYTE_FLIP_MASK "\n\t" \ |
wolfSSL | 15:117db924cf7c | 2144 | "vmovdqa %[SHUF00BA], %" #SHUF_00BA "\n\t" \ |
wolfSSL | 15:117db924cf7c | 2145 | "vmovdqa %[SHUFDC00], %" #SHUF_DC00 "\n\t" |
wolfSSL | 15:117db924cf7c | 2146 | |
wolfSSL | 15:117db924cf7c | 2147 | #define INIT_MASKS_Y(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) \ |
wolfSSL | 15:117db924cf7c | 2148 | _INIT_MASKS_Y(BYTE_FLIP_MASK, SHUF_00BA, SHUF_DC00) |
wolfSSL | 15:117db924cf7c | 2149 | |
wolfSSL | 15:117db924cf7c | 2150 | static const ALIGN32 word32 K256[128] = { |
wolfSSL | 15:117db924cf7c | 2151 | 0x428A2F98L, 0x71374491L, 0xB5C0FBCFL, 0xE9B5DBA5L, |
wolfSSL | 15:117db924cf7c | 2152 | 0x428A2F98L, 0x71374491L, 0xB5C0FBCFL, 0xE9B5DBA5L, |
wolfSSL | 15:117db924cf7c | 2153 | 0x3956C25BL, 0x59F111F1L, 0x923F82A4L, 0xAB1C5ED5L, |
wolfSSL | 15:117db924cf7c | 2154 | 0x3956C25BL, 0x59F111F1L, 0x923F82A4L, 0xAB1C5ED5L, |
wolfSSL | 15:117db924cf7c | 2155 | 0xD807AA98L, 0x12835B01L, 0x243185BEL, 0x550C7DC3L, |
wolfSSL | 15:117db924cf7c | 2156 | 0xD807AA98L, 0x12835B01L, 0x243185BEL, 0x550C7DC3L, |
wolfSSL | 15:117db924cf7c | 2157 | 0x72BE5D74L, 0x80DEB1FEL, 0x9BDC06A7L, 0xC19BF174L, |
wolfSSL | 15:117db924cf7c | 2158 | 0x72BE5D74L, 0x80DEB1FEL, 0x9BDC06A7L, 0xC19BF174L, |
wolfSSL | 15:117db924cf7c | 2159 | 0xE49B69C1L, 0xEFBE4786L, 0x0FC19DC6L, 0x240CA1CCL, |
wolfSSL | 15:117db924cf7c | 2160 | 0xE49B69C1L, 0xEFBE4786L, 0x0FC19DC6L, 0x240CA1CCL, |
wolfSSL | 15:117db924cf7c | 2161 | 0x2DE92C6FL, 0x4A7484AAL, 0x5CB0A9DCL, 0x76F988DAL, |
wolfSSL | 15:117db924cf7c | 2162 | 0x2DE92C6FL, 0x4A7484AAL, 0x5CB0A9DCL, 0x76F988DAL, |
wolfSSL | 15:117db924cf7c | 2163 | 0x983E5152L, 0xA831C66DL, 0xB00327C8L, 0xBF597FC7L, |
wolfSSL | 15:117db924cf7c | 2164 | 0x983E5152L, 0xA831C66DL, 0xB00327C8L, 0xBF597FC7L, |
wolfSSL | 15:117db924cf7c | 2165 | 0xC6E00BF3L, 0xD5A79147L, 0x06CA6351L, 0x14292967L, |
wolfSSL | 15:117db924cf7c | 2166 | 0xC6E00BF3L, 0xD5A79147L, 0x06CA6351L, 0x14292967L, |
wolfSSL | 15:117db924cf7c | 2167 | 0x27B70A85L, 0x2E1B2138L, 0x4D2C6DFCL, 0x53380D13L, |
wolfSSL | 15:117db924cf7c | 2168 | 0x27B70A85L, 0x2E1B2138L, 0x4D2C6DFCL, 0x53380D13L, |
wolfSSL | 15:117db924cf7c | 2169 | 0x650A7354L, 0x766A0ABBL, 0x81C2C92EL, 0x92722C85L, |
wolfSSL | 15:117db924cf7c | 2170 | 0x650A7354L, 0x766A0ABBL, 0x81C2C92EL, 0x92722C85L, |
wolfSSL | 15:117db924cf7c | 2171 | 0xA2BFE8A1L, 0xA81A664BL, 0xC24B8B70L, 0xC76C51A3L, |
wolfSSL | 15:117db924cf7c | 2172 | 0xA2BFE8A1L, 0xA81A664BL, 0xC24B8B70L, 0xC76C51A3L, |
wolfSSL | 15:117db924cf7c | 2173 | 0xD192E819L, 0xD6990624L, 0xF40E3585L, 0x106AA070L, |
wolfSSL | 15:117db924cf7c | 2174 | 0xD192E819L, 0xD6990624L, 0xF40E3585L, 0x106AA070L, |
wolfSSL | 15:117db924cf7c | 2175 | 0x19A4C116L, 0x1E376C08L, 0x2748774CL, 0x34B0BCB5L, |
wolfSSL | 15:117db924cf7c | 2176 | 0x19A4C116L, 0x1E376C08L, 0x2748774CL, 0x34B0BCB5L, |
wolfSSL | 15:117db924cf7c | 2177 | 0x391C0CB3L, 0x4ED8AA4AL, 0x5B9CCA4FL, 0x682E6FF3L, |
wolfSSL | 15:117db924cf7c | 2178 | 0x391C0CB3L, 0x4ED8AA4AL, 0x5B9CCA4FL, 0x682E6FF3L, |
wolfSSL | 15:117db924cf7c | 2179 | 0x748F82EEL, 0x78A5636FL, 0x84C87814L, 0x8CC70208L, |
wolfSSL | 15:117db924cf7c | 2180 | 0x748F82EEL, 0x78A5636FL, 0x84C87814L, 0x8CC70208L, |
wolfSSL | 15:117db924cf7c | 2181 | 0x90BEFFFAL, 0xA4506CEBL, 0xBEF9A3F7L, 0xC67178F2L, |
wolfSSL | 15:117db924cf7c | 2182 | 0x90BEFFFAL, 0xA4506CEBL, 0xBEF9A3F7L, 0xC67178F2L |
wolfSSL | 15:117db924cf7c | 2183 | }; |
wolfSSL | 15:117db924cf7c | 2184 | |
wolfSSL | 15:117db924cf7c | 2185 | SHA256_NOINLINE static int Transform_Sha256_AVX2(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 2186 | { |
wolfSSL | 15:117db924cf7c | 2187 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 2188 | |
wolfSSL | 15:117db924cf7c | 2189 | "subq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2190 | "leaq 32(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2191 | |
wolfSSL | 15:117db924cf7c | 2192 | INIT_MASKS_Y(BYTE_FLIP_MASK, SHUF_Y_00BA, SHUF_Y_DC00) |
wolfSSL | 15:117db924cf7c | 2193 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2194 | |
wolfSSL | 15:117db924cf7c | 2195 | LOAD_W_K_LOW(BYTE_FLIP_MASK, rax) |
wolfSSL | 15:117db924cf7c | 2196 | |
wolfSSL | 15:117db924cf7c | 2197 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2198 | "movl %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 2199 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2200 | |
wolfSSL | 15:117db924cf7c | 2201 | SET_W_Y_4(0) |
wolfSSL | 15:117db924cf7c | 2202 | MsgSched_Y(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 2203 | MsgSched_Y(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 8) |
wolfSSL | 15:117db924cf7c | 2204 | MsgSched_Y(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 16) |
wolfSSL | 15:117db924cf7c | 2205 | MsgSched_Y(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 24) |
wolfSSL | 15:117db924cf7c | 2206 | |
wolfSSL | 15:117db924cf7c | 2207 | SET_W_Y_4(16) |
wolfSSL | 15:117db924cf7c | 2208 | MsgSched_Y(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 32) |
wolfSSL | 15:117db924cf7c | 2209 | MsgSched_Y(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 40) |
wolfSSL | 15:117db924cf7c | 2210 | MsgSched_Y(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 48) |
wolfSSL | 15:117db924cf7c | 2211 | MsgSched_Y(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 56) |
wolfSSL | 15:117db924cf7c | 2212 | |
wolfSSL | 15:117db924cf7c | 2213 | SET_W_Y_4(32) |
wolfSSL | 15:117db924cf7c | 2214 | MsgSched_Y(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 64) |
wolfSSL | 15:117db924cf7c | 2215 | MsgSched_Y(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 72) |
wolfSSL | 15:117db924cf7c | 2216 | MsgSched_Y(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 80) |
wolfSSL | 15:117db924cf7c | 2217 | MsgSched_Y(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 88) |
wolfSSL | 15:117db924cf7c | 2218 | |
wolfSSL | 15:117db924cf7c | 2219 | SET_W_Y_4(48) |
wolfSSL | 15:117db924cf7c | 2220 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 96) |
wolfSSL | 15:117db924cf7c | 2221 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 104) |
wolfSSL | 15:117db924cf7c | 2222 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 112) |
wolfSSL | 15:117db924cf7c | 2223 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 120) |
wolfSSL | 15:117db924cf7c | 2224 | |
wolfSSL | 15:117db924cf7c | 2225 | STORE_ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2226 | |
wolfSSL | 15:117db924cf7c | 2227 | "addq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2228 | |
wolfSSL | 15:117db924cf7c | 2229 | : |
wolfSSL | 15:117db924cf7c | 2230 | : [FLIP] "m" (mBYTE_FLIP_MASK[0]), |
wolfSSL | 15:117db924cf7c | 2231 | [SHUF00BA] "m" (mSHUF_Y_00BA[0]), |
wolfSSL | 15:117db924cf7c | 2232 | [SHUFDC00] "m" (mSHUF_Y_DC00[0]), |
wolfSSL | 15:117db924cf7c | 2233 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 2234 | [K] "m" (K256) |
wolfSSL | 15:117db924cf7c | 2235 | : WORK_REGS, STATE_REGS, YMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 2236 | ); |
wolfSSL | 15:117db924cf7c | 2237 | |
wolfSSL | 15:117db924cf7c | 2238 | return 0; |
wolfSSL | 15:117db924cf7c | 2239 | } |
wolfSSL | 15:117db924cf7c | 2240 | |
wolfSSL | 15:117db924cf7c | 2241 | SHA256_NOINLINE static int Transform_Sha256_AVX2_Len(wc_Sha256* sha256, |
wolfSSL | 15:117db924cf7c | 2242 | word32 len) |
wolfSSL | 15:117db924cf7c | 2243 | { |
wolfSSL | 15:117db924cf7c | 2244 | if ((len & WC_SHA256_BLOCK_SIZE) != 0) { |
wolfSSL | 15:117db924cf7c | 2245 | XMEMCPY(sha256->buffer, sha256->data, WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2246 | Transform_Sha256_AVX2(sha256); |
wolfSSL | 15:117db924cf7c | 2247 | sha256->data += WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2248 | len -= WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2249 | if (len == 0) |
wolfSSL | 15:117db924cf7c | 2250 | return 0; |
wolfSSL | 15:117db924cf7c | 2251 | } |
wolfSSL | 15:117db924cf7c | 2252 | |
wolfSSL | 15:117db924cf7c | 2253 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 2254 | |
wolfSSL | 15:117db924cf7c | 2255 | "subq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2256 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2257 | |
wolfSSL | 15:117db924cf7c | 2258 | INIT_MASKS_Y(BYTE_FLIP_Y_MASK, SHUF_Y_00BA, SHUF_Y_DC00) |
wolfSSL | 15:117db924cf7c | 2259 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2260 | |
wolfSSL | 15:117db924cf7c | 2261 | "# Start of loop processing two blocks\n" |
wolfSSL | 15:117db924cf7c | 2262 | "1:\n\t" |
wolfSSL | 15:117db924cf7c | 2263 | |
wolfSSL | 15:117db924cf7c | 2264 | LOAD_W_K(BYTE_FLIP_Y_MASK, rax) |
wolfSSL | 15:117db924cf7c | 2265 | |
wolfSSL | 15:117db924cf7c | 2266 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2267 | "movl %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 2268 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2269 | |
wolfSSL | 15:117db924cf7c | 2270 | SET_W_Y_4(0) |
wolfSSL | 15:117db924cf7c | 2271 | MsgSched_Y(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 2272 | MsgSched_Y(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 8) |
wolfSSL | 15:117db924cf7c | 2273 | MsgSched_Y(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 16) |
wolfSSL | 15:117db924cf7c | 2274 | MsgSched_Y(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 24) |
wolfSSL | 15:117db924cf7c | 2275 | |
wolfSSL | 15:117db924cf7c | 2276 | SET_W_Y_4(16) |
wolfSSL | 15:117db924cf7c | 2277 | MsgSched_Y(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 32) |
wolfSSL | 15:117db924cf7c | 2278 | MsgSched_Y(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 40) |
wolfSSL | 15:117db924cf7c | 2279 | MsgSched_Y(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 48) |
wolfSSL | 15:117db924cf7c | 2280 | MsgSched_Y(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 56) |
wolfSSL | 15:117db924cf7c | 2281 | |
wolfSSL | 15:117db924cf7c | 2282 | SET_W_Y_4(32) |
wolfSSL | 15:117db924cf7c | 2283 | MsgSched_Y(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 64) |
wolfSSL | 15:117db924cf7c | 2284 | MsgSched_Y(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 72) |
wolfSSL | 15:117db924cf7c | 2285 | MsgSched_Y(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 80) |
wolfSSL | 15:117db924cf7c | 2286 | MsgSched_Y(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 88) |
wolfSSL | 15:117db924cf7c | 2287 | |
wolfSSL | 15:117db924cf7c | 2288 | SET_W_Y_4(48) |
wolfSSL | 15:117db924cf7c | 2289 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 96) |
wolfSSL | 15:117db924cf7c | 2290 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 104) |
wolfSSL | 15:117db924cf7c | 2291 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 112) |
wolfSSL | 15:117db924cf7c | 2292 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 120) |
wolfSSL | 15:117db924cf7c | 2293 | |
wolfSSL | 15:117db924cf7c | 2294 | ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2295 | STORE_DIGEST() |
wolfSSL | 15:117db924cf7c | 2296 | |
wolfSSL | 15:117db924cf7c | 2297 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2298 | "movl %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 2299 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2300 | |
wolfSSL | 15:117db924cf7c | 2301 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 4) |
wolfSSL | 15:117db924cf7c | 2302 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 2303 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 20) |
wolfSSL | 15:117db924cf7c | 2304 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 28) |
wolfSSL | 15:117db924cf7c | 2305 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 36) |
wolfSSL | 15:117db924cf7c | 2306 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 44) |
wolfSSL | 15:117db924cf7c | 2307 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 52) |
wolfSSL | 15:117db924cf7c | 2308 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 60) |
wolfSSL | 15:117db924cf7c | 2309 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 68) |
wolfSSL | 15:117db924cf7c | 2310 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 76) |
wolfSSL | 15:117db924cf7c | 2311 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 84) |
wolfSSL | 15:117db924cf7c | 2312 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 92) |
wolfSSL | 15:117db924cf7c | 2313 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 100) |
wolfSSL | 15:117db924cf7c | 2314 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 108) |
wolfSSL | 15:117db924cf7c | 2315 | RND_ALL_4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 116) |
wolfSSL | 15:117db924cf7c | 2316 | RND_ALL_4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 124) |
wolfSSL | 15:117db924cf7c | 2317 | |
wolfSSL | 15:117db924cf7c | 2318 | ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2319 | |
wolfSSL | 15:117db924cf7c | 2320 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2321 | "addq $128, %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2322 | "subl $128, %[len]\n\t" |
wolfSSL | 15:117db924cf7c | 2323 | |
wolfSSL | 15:117db924cf7c | 2324 | STORE_DIGEST() |
wolfSSL | 15:117db924cf7c | 2325 | |
wolfSSL | 15:117db924cf7c | 2326 | "movq %%rax, 120(%[sha256])\n\t" |
wolfSSL | 15:117db924cf7c | 2327 | "jnz 1b\n\t" |
wolfSSL | 15:117db924cf7c | 2328 | |
wolfSSL | 15:117db924cf7c | 2329 | "addq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2330 | |
wolfSSL | 15:117db924cf7c | 2331 | : |
wolfSSL | 15:117db924cf7c | 2332 | : [FLIP] "m" (mBYTE_FLIP_Y_MASK[0]), |
wolfSSL | 15:117db924cf7c | 2333 | [SHUF00BA] "m" (mSHUF_Y_00BA[0]), |
wolfSSL | 15:117db924cf7c | 2334 | [SHUFDC00] "m" (mSHUF_Y_DC00[0]), |
wolfSSL | 15:117db924cf7c | 2335 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 2336 | [len] "r" (len), |
wolfSSL | 15:117db924cf7c | 2337 | [K] "m" (K256) |
wolfSSL | 15:117db924cf7c | 2338 | : WORK_REGS, STATE_REGS, YMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 2339 | ); |
wolfSSL | 15:117db924cf7c | 2340 | |
wolfSSL | 15:117db924cf7c | 2341 | return 0; |
wolfSSL | 15:117db924cf7c | 2342 | } |
wolfSSL | 15:117db924cf7c | 2343 | |
wolfSSL | 15:117db924cf7c | 2344 | #if defined(HAVE_INTEL_RORX) |
wolfSSL | 15:117db924cf7c | 2345 | SHA256_NOINLINE static int Transform_Sha256_AVX2_RORX(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 2346 | { |
wolfSSL | 15:117db924cf7c | 2347 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 2348 | |
wolfSSL | 15:117db924cf7c | 2349 | "subq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2350 | "leaq 32(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2351 | |
wolfSSL | 15:117db924cf7c | 2352 | INIT_MASKS_Y(BYTE_FLIP_MASK, SHUF_Y_00BA, SHUF_Y_DC00) |
wolfSSL | 15:117db924cf7c | 2353 | LOAD_W_K_LOW(BYTE_FLIP_MASK, rax) |
wolfSSL | 15:117db924cf7c | 2354 | |
wolfSSL | 15:117db924cf7c | 2355 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2356 | |
wolfSSL | 15:117db924cf7c | 2357 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2358 | "rorx $6, %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 2359 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2360 | |
wolfSSL | 15:117db924cf7c | 2361 | SET_W_Y_4(0) |
wolfSSL | 15:117db924cf7c | 2362 | MsgSched_Y_RORX(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 2363 | MsgSched_Y_RORX(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 8) |
wolfSSL | 15:117db924cf7c | 2364 | MsgSched_Y_RORX(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 16) |
wolfSSL | 15:117db924cf7c | 2365 | MsgSched_Y_RORX(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 24) |
wolfSSL | 15:117db924cf7c | 2366 | |
wolfSSL | 15:117db924cf7c | 2367 | SET_W_Y_4(16) |
wolfSSL | 15:117db924cf7c | 2368 | MsgSched_Y_RORX(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 32) |
wolfSSL | 15:117db924cf7c | 2369 | MsgSched_Y_RORX(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 40) |
wolfSSL | 15:117db924cf7c | 2370 | MsgSched_Y_RORX(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 48) |
wolfSSL | 15:117db924cf7c | 2371 | MsgSched_Y_RORX(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 56) |
wolfSSL | 15:117db924cf7c | 2372 | |
wolfSSL | 15:117db924cf7c | 2373 | SET_W_Y_4(32) |
wolfSSL | 15:117db924cf7c | 2374 | MsgSched_Y_RORX(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 64) |
wolfSSL | 15:117db924cf7c | 2375 | MsgSched_Y_RORX(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 72) |
wolfSSL | 15:117db924cf7c | 2376 | MsgSched_Y_RORX(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 80) |
wolfSSL | 15:117db924cf7c | 2377 | MsgSched_Y_RORX(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 88) |
wolfSSL | 15:117db924cf7c | 2378 | |
wolfSSL | 15:117db924cf7c | 2379 | SET_W_Y_4(48) |
wolfSSL | 15:117db924cf7c | 2380 | "xorl " L3 ", " L3 "\n\t" |
wolfSSL | 15:117db924cf7c | 2381 | "xorl " L2 ", " L2 "\n\t" |
wolfSSL | 15:117db924cf7c | 2382 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 96) |
wolfSSL | 15:117db924cf7c | 2383 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 104) |
wolfSSL | 15:117db924cf7c | 2384 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 112) |
wolfSSL | 15:117db924cf7c | 2385 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 120) |
wolfSSL | 15:117db924cf7c | 2386 | /* Prev RND: h += Maj(a,b,c) */ |
wolfSSL | 15:117db924cf7c | 2387 | "addl " L3 ", %%r8d\n\t" |
wolfSSL | 15:117db924cf7c | 2388 | |
wolfSSL | 15:117db924cf7c | 2389 | STORE_ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2390 | |
wolfSSL | 15:117db924cf7c | 2391 | "addq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2392 | |
wolfSSL | 15:117db924cf7c | 2393 | : |
wolfSSL | 15:117db924cf7c | 2394 | : [FLIP] "m" (mBYTE_FLIP_MASK[0]), |
wolfSSL | 15:117db924cf7c | 2395 | [SHUF00BA] "m" (mSHUF_Y_00BA[0]), |
wolfSSL | 15:117db924cf7c | 2396 | [SHUFDC00] "m" (mSHUF_Y_DC00[0]), |
wolfSSL | 15:117db924cf7c | 2397 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 2398 | [K] "m" (K256) |
wolfSSL | 15:117db924cf7c | 2399 | : WORK_REGS, STATE_REGS, YMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 2400 | ); |
wolfSSL | 15:117db924cf7c | 2401 | |
wolfSSL | 15:117db924cf7c | 2402 | return 0; |
wolfSSL | 15:117db924cf7c | 2403 | } |
wolfSSL | 15:117db924cf7c | 2404 | |
wolfSSL | 15:117db924cf7c | 2405 | SHA256_NOINLINE static int Transform_Sha256_AVX2_RORX_Len(wc_Sha256* sha256, |
wolfSSL | 15:117db924cf7c | 2406 | word32 len) |
wolfSSL | 15:117db924cf7c | 2407 | { |
wolfSSL | 15:117db924cf7c | 2408 | if ((len & WC_SHA256_BLOCK_SIZE) != 0) { |
wolfSSL | 15:117db924cf7c | 2409 | XMEMCPY(sha256->buffer, sha256->data, WC_SHA256_BLOCK_SIZE); |
wolfSSL | 15:117db924cf7c | 2410 | Transform_Sha256_AVX2_RORX(sha256); |
wolfSSL | 15:117db924cf7c | 2411 | sha256->data += WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2412 | len -= WC_SHA256_BLOCK_SIZE; |
wolfSSL | 15:117db924cf7c | 2413 | if (len == 0) |
wolfSSL | 15:117db924cf7c | 2414 | return 0; |
wolfSSL | 15:117db924cf7c | 2415 | } |
wolfSSL | 15:117db924cf7c | 2416 | |
wolfSSL | 15:117db924cf7c | 2417 | __asm__ __volatile__ ( |
wolfSSL | 15:117db924cf7c | 2418 | |
wolfSSL | 15:117db924cf7c | 2419 | "subq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2420 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2421 | |
wolfSSL | 15:117db924cf7c | 2422 | INIT_MASKS_Y(BYTE_FLIP_Y_MASK, SHUF_Y_00BA, SHUF_Y_DC00) |
wolfSSL | 15:117db924cf7c | 2423 | LOAD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2424 | |
wolfSSL | 15:117db924cf7c | 2425 | "# Start of loop processing two blocks\n" |
wolfSSL | 15:117db924cf7c | 2426 | "1:\n\t" |
wolfSSL | 15:117db924cf7c | 2427 | |
wolfSSL | 15:117db924cf7c | 2428 | LOAD_W_K(BYTE_FLIP_Y_MASK, rax) |
wolfSSL | 15:117db924cf7c | 2429 | |
wolfSSL | 15:117db924cf7c | 2430 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2431 | "rorx $6, %%r12d, " L1 "\n\t" |
wolfSSL | 15:117db924cf7c | 2432 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2433 | |
wolfSSL | 15:117db924cf7c | 2434 | SET_W_Y_4(0) |
wolfSSL | 15:117db924cf7c | 2435 | MsgSched_Y_RORX(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 0) |
wolfSSL | 15:117db924cf7c | 2436 | MsgSched_Y_RORX(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 8) |
wolfSSL | 15:117db924cf7c | 2437 | MsgSched_Y_RORX(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 16) |
wolfSSL | 15:117db924cf7c | 2438 | MsgSched_Y_RORX(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 24) |
wolfSSL | 15:117db924cf7c | 2439 | |
wolfSSL | 15:117db924cf7c | 2440 | SET_W_Y_4(16) |
wolfSSL | 15:117db924cf7c | 2441 | MsgSched_Y_RORX(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 32) |
wolfSSL | 15:117db924cf7c | 2442 | MsgSched_Y_RORX(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 40) |
wolfSSL | 15:117db924cf7c | 2443 | MsgSched_Y_RORX(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 48) |
wolfSSL | 15:117db924cf7c | 2444 | MsgSched_Y_RORX(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 56) |
wolfSSL | 15:117db924cf7c | 2445 | |
wolfSSL | 15:117db924cf7c | 2446 | SET_W_Y_4(32) |
wolfSSL | 15:117db924cf7c | 2447 | MsgSched_Y_RORX(Y0, Y1, Y2, Y3, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 64) |
wolfSSL | 15:117db924cf7c | 2448 | MsgSched_Y_RORX(Y1, Y2, Y3, Y0, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 72) |
wolfSSL | 15:117db924cf7c | 2449 | MsgSched_Y_RORX(Y2, Y3, Y0, Y1, S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 80) |
wolfSSL | 15:117db924cf7c | 2450 | MsgSched_Y_RORX(Y3, Y0, Y1, Y2, S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 88) |
wolfSSL | 15:117db924cf7c | 2451 | |
wolfSSL | 15:117db924cf7c | 2452 | SET_W_Y_4(48) |
wolfSSL | 15:117db924cf7c | 2453 | "xorl " L3 ", " L3 "\n\t" |
wolfSSL | 15:117db924cf7c | 2454 | "xorl " L2 ", " L2 "\n\t" |
wolfSSL | 15:117db924cf7c | 2455 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 96) |
wolfSSL | 15:117db924cf7c | 2456 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 104) |
wolfSSL | 15:117db924cf7c | 2457 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 112) |
wolfSSL | 15:117db924cf7c | 2458 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 120) |
wolfSSL | 15:117db924cf7c | 2459 | /* Prev RND: h += Maj(a,b,c) */ |
wolfSSL | 15:117db924cf7c | 2460 | "addl " L3 ", %%r8d\n\t" |
wolfSSL | 15:117db924cf7c | 2461 | "xorl " L2 ", " L2 "\n\t" |
wolfSSL | 15:117db924cf7c | 2462 | |
wolfSSL | 15:117db924cf7c | 2463 | ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2464 | STORE_DIGEST() |
wolfSSL | 15:117db924cf7c | 2465 | |
wolfSSL | 15:117db924cf7c | 2466 | "movl %%r9d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2467 | "xorl " L3 ", " L3 "\n\t" |
wolfSSL | 15:117db924cf7c | 2468 | "xorl %%r10d, " L4 "\n\t" |
wolfSSL | 15:117db924cf7c | 2469 | |
wolfSSL | 15:117db924cf7c | 2470 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 4) |
wolfSSL | 15:117db924cf7c | 2471 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 12) |
wolfSSL | 15:117db924cf7c | 2472 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 20) |
wolfSSL | 15:117db924cf7c | 2473 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 28) |
wolfSSL | 15:117db924cf7c | 2474 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 36) |
wolfSSL | 15:117db924cf7c | 2475 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 44) |
wolfSSL | 15:117db924cf7c | 2476 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 52) |
wolfSSL | 15:117db924cf7c | 2477 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 60) |
wolfSSL | 15:117db924cf7c | 2478 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 68) |
wolfSSL | 15:117db924cf7c | 2479 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 76) |
wolfSSL | 15:117db924cf7c | 2480 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 84) |
wolfSSL | 15:117db924cf7c | 2481 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 92) |
wolfSSL | 15:117db924cf7c | 2482 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 100) |
wolfSSL | 15:117db924cf7c | 2483 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 108) |
wolfSSL | 15:117db924cf7c | 2484 | RND_RORX_X4(S_0, S_1, S_2, S_3, S_4, S_5, S_6, S_7, 116) |
wolfSSL | 15:117db924cf7c | 2485 | RND_RORX_X4(S_4, S_5, S_6, S_7, S_0, S_1, S_2, S_3, 124) |
wolfSSL | 15:117db924cf7c | 2486 | /* Prev RND: h += Maj(a,b,c) */ |
wolfSSL | 15:117db924cf7c | 2487 | "addl " L3 ", %%r8d\n\t" |
wolfSSL | 15:117db924cf7c | 2488 | "movq 120(%[sha256]), %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2489 | |
wolfSSL | 15:117db924cf7c | 2490 | ADD_DIGEST() |
wolfSSL | 15:117db924cf7c | 2491 | |
wolfSSL | 15:117db924cf7c | 2492 | "addq $128, %%rax\n\t" |
wolfSSL | 15:117db924cf7c | 2493 | "subl $128, %[len]\n\t" |
wolfSSL | 15:117db924cf7c | 2494 | |
wolfSSL | 15:117db924cf7c | 2495 | STORE_DIGEST() |
wolfSSL | 15:117db924cf7c | 2496 | |
wolfSSL | 15:117db924cf7c | 2497 | "movq %%rax, 120(%[sha256])\n\t" |
wolfSSL | 15:117db924cf7c | 2498 | "jnz 1b\n\t" |
wolfSSL | 15:117db924cf7c | 2499 | |
wolfSSL | 15:117db924cf7c | 2500 | "addq $512, %%rsp\n\t" |
wolfSSL | 15:117db924cf7c | 2501 | |
wolfSSL | 15:117db924cf7c | 2502 | : |
wolfSSL | 15:117db924cf7c | 2503 | : [FLIP] "m" (mBYTE_FLIP_Y_MASK[0]), |
wolfSSL | 15:117db924cf7c | 2504 | [SHUF00BA] "m" (mSHUF_Y_00BA[0]), |
wolfSSL | 15:117db924cf7c | 2505 | [SHUFDC00] "m" (mSHUF_Y_DC00[0]), |
wolfSSL | 15:117db924cf7c | 2506 | [sha256] "r" (sha256), |
wolfSSL | 15:117db924cf7c | 2507 | [len] "r" (len), |
wolfSSL | 15:117db924cf7c | 2508 | [K] "m" (K256) |
wolfSSL | 15:117db924cf7c | 2509 | : WORK_REGS, STATE_REGS, YMM_REGS, "memory" |
wolfSSL | 15:117db924cf7c | 2510 | ); |
wolfSSL | 15:117db924cf7c | 2511 | |
wolfSSL | 15:117db924cf7c | 2512 | return 0; |
wolfSSL | 15:117db924cf7c | 2513 | } |
wolfSSL | 15:117db924cf7c | 2514 | #endif /* HAVE_INTEL_RORX */ |
wolfSSL | 15:117db924cf7c | 2515 | #endif /* HAVE_INTEL_AVX2 */ |
wolfSSL | 15:117db924cf7c | 2516 | |
wolfSSL | 15:117db924cf7c | 2517 | |
wolfSSL | 15:117db924cf7c | 2518 | #ifdef WOLFSSL_SHA224 |
wolfSSL | 15:117db924cf7c | 2519 | |
wolfSSL | 15:117db924cf7c | 2520 | #ifdef STM32_HASH_SHA2 |
wolfSSL | 15:117db924cf7c | 2521 | |
wolfSSL | 15:117db924cf7c | 2522 | /* Supports CubeMX HAL or Standard Peripheral Library */ |
wolfSSL | 15:117db924cf7c | 2523 | |
wolfSSL | 15:117db924cf7c | 2524 | int wc_InitSha224_ex(wc_Sha224* sha224, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 2525 | { |
wolfSSL | 15:117db924cf7c | 2526 | if (sha224 == NULL) |
wolfSSL | 15:117db924cf7c | 2527 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2528 | |
wolfSSL | 15:117db924cf7c | 2529 | (void)devId; |
wolfSSL | 15:117db924cf7c | 2530 | (void)heap; |
wolfSSL | 15:117db924cf7c | 2531 | |
wolfSSL | 15:117db924cf7c | 2532 | wc_Stm32_Hash_Init(&sha224->stmCtx); |
wolfSSL | 15:117db924cf7c | 2533 | return 0; |
wolfSSL | 15:117db924cf7c | 2534 | } |
wolfSSL | 15:117db924cf7c | 2535 | |
wolfSSL | 15:117db924cf7c | 2536 | int wc_Sha224Update(wc_Sha224* sha224, const byte* data, word32 len) |
wolfSSL | 15:117db924cf7c | 2537 | { |
wolfSSL | 15:117db924cf7c | 2538 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2539 | |
wolfSSL | 15:117db924cf7c | 2540 | if (sha224 == NULL || (data == NULL && len > 0)) { |
wolfSSL | 15:117db924cf7c | 2541 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2542 | } |
wolfSSL | 15:117db924cf7c | 2543 | |
wolfSSL | 15:117db924cf7c | 2544 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 2545 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 2546 | ret = wc_Stm32_Hash_Update(&sha224->stmCtx, |
wolfSSL | 15:117db924cf7c | 2547 | HASH_AlgoSelection_SHA224, data, len); |
wolfSSL | 15:117db924cf7c | 2548 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 2549 | } |
wolfSSL | 15:117db924cf7c | 2550 | return ret; |
wolfSSL | 15:117db924cf7c | 2551 | } |
wolfSSL | 15:117db924cf7c | 2552 | |
wolfSSL | 15:117db924cf7c | 2553 | int wc_Sha224Final(wc_Sha224* sha224, byte* hash) |
wolfSSL | 15:117db924cf7c | 2554 | { |
wolfSSL | 15:117db924cf7c | 2555 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2556 | |
wolfSSL | 15:117db924cf7c | 2557 | if (sha224 == NULL || hash == NULL) { |
wolfSSL | 15:117db924cf7c | 2558 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2559 | } |
wolfSSL | 15:117db924cf7c | 2560 | |
wolfSSL | 15:117db924cf7c | 2561 | ret = wolfSSL_CryptHwMutexLock(); |
wolfSSL | 15:117db924cf7c | 2562 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 2563 | ret = wc_Stm32_Hash_Final(&sha224->stmCtx, |
wolfSSL | 15:117db924cf7c | 2564 | HASH_AlgoSelection_SHA224, hash, WC_SHA224_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 2565 | wolfSSL_CryptHwMutexUnLock(); |
wolfSSL | 15:117db924cf7c | 2566 | } |
wolfSSL | 15:117db924cf7c | 2567 | |
wolfSSL | 15:117db924cf7c | 2568 | (void)wc_InitSha224(sha224); /* reset state */ |
wolfSSL | 15:117db924cf7c | 2569 | |
wolfSSL | 15:117db924cf7c | 2570 | return ret; |
wolfSSL | 15:117db924cf7c | 2571 | } |
wolfSSL | 15:117db924cf7c | 2572 | |
wolfSSL | 15:117db924cf7c | 2573 | #elif defined(WOLFSSL_IMX6_CAAM) && !defined(NO_IMX6_CAAM_HASH) |
wolfSSL | 15:117db924cf7c | 2574 | /* functions defined in wolfcrypt/src/port/caam/caam_sha256.c */ |
wolfSSL | 15:117db924cf7c | 2575 | #else |
wolfSSL | 15:117db924cf7c | 2576 | |
wolfSSL | 15:117db924cf7c | 2577 | #define NEED_SOFT_SHA224 |
wolfSSL | 15:117db924cf7c | 2578 | |
wolfSSL | 15:117db924cf7c | 2579 | |
wolfSSL | 15:117db924cf7c | 2580 | static int InitSha224(wc_Sha224* sha224) |
wolfSSL | 15:117db924cf7c | 2581 | { |
wolfSSL | 15:117db924cf7c | 2582 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2583 | |
wolfSSL | 15:117db924cf7c | 2584 | if (sha224 == NULL) { |
wolfSSL | 15:117db924cf7c | 2585 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2586 | } |
wolfSSL | 15:117db924cf7c | 2587 | |
wolfSSL | 15:117db924cf7c | 2588 | sha224->digest[0] = 0xc1059ed8; |
wolfSSL | 15:117db924cf7c | 2589 | sha224->digest[1] = 0x367cd507; |
wolfSSL | 15:117db924cf7c | 2590 | sha224->digest[2] = 0x3070dd17; |
wolfSSL | 15:117db924cf7c | 2591 | sha224->digest[3] = 0xf70e5939; |
wolfSSL | 15:117db924cf7c | 2592 | sha224->digest[4] = 0xffc00b31; |
wolfSSL | 15:117db924cf7c | 2593 | sha224->digest[5] = 0x68581511; |
wolfSSL | 15:117db924cf7c | 2594 | sha224->digest[6] = 0x64f98fa7; |
wolfSSL | 15:117db924cf7c | 2595 | sha224->digest[7] = 0xbefa4fa4; |
wolfSSL | 15:117db924cf7c | 2596 | |
wolfSSL | 15:117db924cf7c | 2597 | sha224->buffLen = 0; |
wolfSSL | 15:117db924cf7c | 2598 | sha224->loLen = 0; |
wolfSSL | 15:117db924cf7c | 2599 | sha224->hiLen = 0; |
wolfSSL | 15:117db924cf7c | 2600 | |
wolfSSL | 15:117db924cf7c | 2601 | #if defined(HAVE_INTEL_AVX1)|| defined(HAVE_INTEL_AVX2) |
wolfSSL | 15:117db924cf7c | 2602 | /* choose best Transform function under this runtime environment */ |
wolfSSL | 15:117db924cf7c | 2603 | Sha256_SetTransform(); |
wolfSSL | 15:117db924cf7c | 2604 | #endif |
wolfSSL | 15:117db924cf7c | 2605 | |
wolfSSL | 15:117db924cf7c | 2606 | return ret; |
wolfSSL | 15:117db924cf7c | 2607 | } |
wolfSSL | 15:117db924cf7c | 2608 | |
wolfSSL | 15:117db924cf7c | 2609 | #endif |
wolfSSL | 15:117db924cf7c | 2610 | |
wolfSSL | 15:117db924cf7c | 2611 | #ifdef NEED_SOFT_SHA224 |
wolfSSL | 15:117db924cf7c | 2612 | int wc_InitSha224_ex(wc_Sha224* sha224, void* heap, int devId) |
wolfSSL | 15:117db924cf7c | 2613 | { |
wolfSSL | 15:117db924cf7c | 2614 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2615 | |
wolfSSL | 15:117db924cf7c | 2616 | if (sha224 == NULL) |
wolfSSL | 15:117db924cf7c | 2617 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2618 | |
wolfSSL | 15:117db924cf7c | 2619 | sha224->heap = heap; |
wolfSSL | 15:117db924cf7c | 2620 | |
wolfSSL | 15:117db924cf7c | 2621 | ret = InitSha224(sha224); |
wolfSSL | 15:117db924cf7c | 2622 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 2623 | return ret; |
wolfSSL | 15:117db924cf7c | 2624 | |
wolfSSL | 15:117db924cf7c | 2625 | #ifdef WOLFSSL_SMALL_STACK_CACHE |
wolfSSL | 15:117db924cf7c | 2626 | sha224->W = NULL; |
wolfSSL | 15:117db924cf7c | 2627 | #endif |
wolfSSL | 15:117db924cf7c | 2628 | |
wolfSSL | 15:117db924cf7c | 2629 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA224) |
wolfSSL | 15:117db924cf7c | 2630 | ret = wolfAsync_DevCtxInit(&sha224->asyncDev, |
wolfSSL | 15:117db924cf7c | 2631 | WOLFSSL_ASYNC_MARKER_SHA224, sha224->heap, devId); |
wolfSSL | 15:117db924cf7c | 2632 | #else |
wolfSSL | 15:117db924cf7c | 2633 | (void)devId; |
wolfSSL | 15:117db924cf7c | 2634 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 2635 | |
wolfSSL | 15:117db924cf7c | 2636 | return ret; |
wolfSSL | 15:117db924cf7c | 2637 | } |
wolfSSL | 15:117db924cf7c | 2638 | |
wolfSSL | 15:117db924cf7c | 2639 | int wc_Sha224Update(wc_Sha224* sha224, const byte* data, word32 len) |
wolfSSL | 15:117db924cf7c | 2640 | { |
wolfSSL | 15:117db924cf7c | 2641 | int ret; |
wolfSSL | 15:117db924cf7c | 2642 | |
wolfSSL | 15:117db924cf7c | 2643 | if (sha224 == NULL || (data == NULL && len > 0)) { |
wolfSSL | 15:117db924cf7c | 2644 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2645 | } |
wolfSSL | 15:117db924cf7c | 2646 | |
wolfSSL | 15:117db924cf7c | 2647 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA224) |
wolfSSL | 15:117db924cf7c | 2648 | if (sha224->asyncDev.marker == WOLFSSL_ASYNC_MARKER_SHA224) { |
wolfSSL | 15:117db924cf7c | 2649 | #if defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 2650 | return IntelQaSymSha224(&sha224->asyncDev, NULL, data, len); |
wolfSSL | 15:117db924cf7c | 2651 | #endif |
wolfSSL | 15:117db924cf7c | 2652 | } |
wolfSSL | 15:117db924cf7c | 2653 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 2654 | |
wolfSSL | 15:117db924cf7c | 2655 | ret = Sha256Update((wc_Sha256*)sha224, data, len); |
wolfSSL | 15:117db924cf7c | 2656 | |
wolfSSL | 15:117db924cf7c | 2657 | return ret; |
wolfSSL | 15:117db924cf7c | 2658 | } |
wolfSSL | 15:117db924cf7c | 2659 | |
wolfSSL | 15:117db924cf7c | 2660 | int wc_Sha224Final(wc_Sha224* sha224, byte* hash) |
wolfSSL | 15:117db924cf7c | 2661 | { |
wolfSSL | 15:117db924cf7c | 2662 | int ret; |
wolfSSL | 15:117db924cf7c | 2663 | |
wolfSSL | 15:117db924cf7c | 2664 | if (sha224 == NULL || hash == NULL) { |
wolfSSL | 15:117db924cf7c | 2665 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2666 | } |
wolfSSL | 15:117db924cf7c | 2667 | |
wolfSSL | 15:117db924cf7c | 2668 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA224) |
wolfSSL | 15:117db924cf7c | 2669 | if (sha224->asyncDev.marker == WOLFSSL_ASYNC_MARKER_SHA224) { |
wolfSSL | 15:117db924cf7c | 2670 | #if defined(HAVE_INTEL_QA) |
wolfSSL | 15:117db924cf7c | 2671 | return IntelQaSymSha224(&sha224->asyncDev, hash, NULL, |
wolfSSL | 15:117db924cf7c | 2672 | WC_SHA224_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 2673 | #endif |
wolfSSL | 15:117db924cf7c | 2674 | } |
wolfSSL | 15:117db924cf7c | 2675 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 2676 | |
wolfSSL | 15:117db924cf7c | 2677 | ret = Sha256Final((wc_Sha256*)sha224); |
wolfSSL | 15:117db924cf7c | 2678 | if (ret != 0) |
wolfSSL | 15:117db924cf7c | 2679 | return ret; |
wolfSSL | 15:117db924cf7c | 2680 | |
wolfSSL | 15:117db924cf7c | 2681 | #if defined(LITTLE_ENDIAN_ORDER) |
wolfSSL | 15:117db924cf7c | 2682 | ByteReverseWords(sha224->digest, sha224->digest, WC_SHA224_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 2683 | #endif |
wolfSSL | 15:117db924cf7c | 2684 | XMEMCPY(hash, sha224->digest, WC_SHA224_DIGEST_SIZE); |
wolfSSL | 15:117db924cf7c | 2685 | |
wolfSSL | 15:117db924cf7c | 2686 | return InitSha224(sha224); /* reset state */ |
wolfSSL | 15:117db924cf7c | 2687 | } |
wolfSSL | 15:117db924cf7c | 2688 | #endif /* end of SHA224 software implementation */ |
wolfSSL | 15:117db924cf7c | 2689 | |
wolfSSL | 15:117db924cf7c | 2690 | int wc_InitSha224(wc_Sha224* sha224) |
wolfSSL | 15:117db924cf7c | 2691 | { |
wolfSSL | 15:117db924cf7c | 2692 | return wc_InitSha224_ex(sha224, NULL, INVALID_DEVID); |
wolfSSL | 15:117db924cf7c | 2693 | } |
wolfSSL | 15:117db924cf7c | 2694 | |
wolfSSL | 15:117db924cf7c | 2695 | void wc_Sha224Free(wc_Sha224* sha224) |
wolfSSL | 15:117db924cf7c | 2696 | { |
wolfSSL | 15:117db924cf7c | 2697 | if (sha224 == NULL) |
wolfSSL | 15:117db924cf7c | 2698 | return; |
wolfSSL | 15:117db924cf7c | 2699 | |
wolfSSL | 15:117db924cf7c | 2700 | #ifdef WOLFSSL_SMALL_STACK_CACHE |
wolfSSL | 15:117db924cf7c | 2701 | if (sha224->W != NULL) { |
wolfSSL | 15:117db924cf7c | 2702 | XFREE(sha224->W, NULL, DYNAMIC_TYPE_RNG); |
wolfSSL | 15:117db924cf7c | 2703 | sha224->W = NULL; |
wolfSSL | 15:117db924cf7c | 2704 | } |
wolfSSL | 15:117db924cf7c | 2705 | #endif |
wolfSSL | 15:117db924cf7c | 2706 | |
wolfSSL | 15:117db924cf7c | 2707 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA224) |
wolfSSL | 15:117db924cf7c | 2708 | wolfAsync_DevCtxFree(&sha224->asyncDev, WOLFSSL_ASYNC_MARKER_SHA224); |
wolfSSL | 15:117db924cf7c | 2709 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 2710 | } |
wolfSSL | 15:117db924cf7c | 2711 | #endif /* WOLFSSL_SHA224 */ |
wolfSSL | 15:117db924cf7c | 2712 | |
wolfSSL | 15:117db924cf7c | 2713 | |
wolfSSL | 15:117db924cf7c | 2714 | int wc_InitSha256(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 2715 | { |
wolfSSL | 15:117db924cf7c | 2716 | return wc_InitSha256_ex(sha256, NULL, INVALID_DEVID); |
wolfSSL | 15:117db924cf7c | 2717 | } |
wolfSSL | 15:117db924cf7c | 2718 | |
wolfSSL | 15:117db924cf7c | 2719 | void wc_Sha256Free(wc_Sha256* sha256) |
wolfSSL | 15:117db924cf7c | 2720 | { |
wolfSSL | 15:117db924cf7c | 2721 | if (sha256 == NULL) |
wolfSSL | 15:117db924cf7c | 2722 | return; |
wolfSSL | 15:117db924cf7c | 2723 | |
wolfSSL | 15:117db924cf7c | 2724 | #ifdef WOLFSSL_SMALL_STACK_CACHE |
wolfSSL | 15:117db924cf7c | 2725 | if (sha256->W != NULL) { |
wolfSSL | 15:117db924cf7c | 2726 | XFREE(sha256->W, NULL, DYNAMIC_TYPE_RNG); |
wolfSSL | 15:117db924cf7c | 2727 | sha256->W = NULL; |
wolfSSL | 15:117db924cf7c | 2728 | } |
wolfSSL | 15:117db924cf7c | 2729 | #endif |
wolfSSL | 15:117db924cf7c | 2730 | |
wolfSSL | 15:117db924cf7c | 2731 | #if defined(WOLFSSL_ASYNC_CRYPT) && defined(WC_ASYNC_ENABLE_SHA256) |
wolfSSL | 15:117db924cf7c | 2732 | wolfAsync_DevCtxFree(&sha256->asyncDev, WOLFSSL_ASYNC_MARKER_SHA256); |
wolfSSL | 15:117db924cf7c | 2733 | #endif /* WOLFSSL_ASYNC_CRYPT */ |
wolfSSL | 15:117db924cf7c | 2734 | } |
wolfSSL | 15:117db924cf7c | 2735 | |
wolfSSL | 15:117db924cf7c | 2736 | #endif /* !WOLFSSL_TI_HASH */ |
wolfSSL | 15:117db924cf7c | 2737 | #endif /* HAVE_FIPS */ |
wolfSSL | 15:117db924cf7c | 2738 | |
wolfSSL | 15:117db924cf7c | 2739 | |
wolfSSL | 15:117db924cf7c | 2740 | #ifndef WOLFSSL_TI_HASH |
wolfSSL | 15:117db924cf7c | 2741 | #ifdef WOLFSSL_SHA224 |
wolfSSL | 15:117db924cf7c | 2742 | int wc_Sha224GetHash(wc_Sha224* sha224, byte* hash) |
wolfSSL | 15:117db924cf7c | 2743 | { |
wolfSSL | 15:117db924cf7c | 2744 | int ret; |
wolfSSL | 15:117db924cf7c | 2745 | wc_Sha224 tmpSha224; |
wolfSSL | 15:117db924cf7c | 2746 | |
wolfSSL | 15:117db924cf7c | 2747 | if (sha224 == NULL || hash == NULL) |
wolfSSL | 15:117db924cf7c | 2748 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2749 | |
wolfSSL | 15:117db924cf7c | 2750 | ret = wc_Sha224Copy(sha224, &tmpSha224); |
wolfSSL | 15:117db924cf7c | 2751 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 2752 | ret = wc_Sha224Final(&tmpSha224, hash); |
wolfSSL | 15:117db924cf7c | 2753 | wc_Sha224Free(&tmpSha224); |
wolfSSL | 15:117db924cf7c | 2754 | } |
wolfSSL | 15:117db924cf7c | 2755 | return ret; |
wolfSSL | 15:117db924cf7c | 2756 | } |
wolfSSL | 15:117db924cf7c | 2757 | int wc_Sha224Copy(wc_Sha224* src, wc_Sha224* dst) |
wolfSSL | 15:117db924cf7c | 2758 | { |
wolfSSL | 15:117db924cf7c | 2759 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2760 | |
wolfSSL | 15:117db924cf7c | 2761 | if (src == NULL || dst == NULL) |
wolfSSL | 15:117db924cf7c | 2762 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2763 | |
wolfSSL | 15:117db924cf7c | 2764 | XMEMCPY(dst, src, sizeof(wc_Sha224)); |
wolfSSL | 15:117db924cf7c | 2765 | #ifdef WOLFSSL_SMALL_STACK_CACHE |
wolfSSL | 15:117db924cf7c | 2766 | dst->W = NULL; |
wolfSSL | 15:117db924cf7c | 2767 | #endif |
wolfSSL | 15:117db924cf7c | 2768 | |
wolfSSL | 15:117db924cf7c | 2769 | #ifdef WOLFSSL_ASYNC_CRYPT |
wolfSSL | 15:117db924cf7c | 2770 | ret = wolfAsync_DevCopy(&src->asyncDev, &dst->asyncDev); |
wolfSSL | 15:117db924cf7c | 2771 | #endif |
wolfSSL | 15:117db924cf7c | 2772 | |
wolfSSL | 15:117db924cf7c | 2773 | return ret; |
wolfSSL | 15:117db924cf7c | 2774 | } |
wolfSSL | 15:117db924cf7c | 2775 | #endif /* WOLFSSL_SHA224 */ |
wolfSSL | 15:117db924cf7c | 2776 | |
wolfSSL | 15:117db924cf7c | 2777 | int wc_Sha256GetHash(wc_Sha256* sha256, byte* hash) |
wolfSSL | 15:117db924cf7c | 2778 | { |
wolfSSL | 15:117db924cf7c | 2779 | int ret; |
wolfSSL | 15:117db924cf7c | 2780 | wc_Sha256 tmpSha256; |
wolfSSL | 15:117db924cf7c | 2781 | |
wolfSSL | 15:117db924cf7c | 2782 | if (sha256 == NULL || hash == NULL) |
wolfSSL | 15:117db924cf7c | 2783 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2784 | |
wolfSSL | 15:117db924cf7c | 2785 | ret = wc_Sha256Copy(sha256, &tmpSha256); |
wolfSSL | 15:117db924cf7c | 2786 | if (ret == 0) { |
wolfSSL | 15:117db924cf7c | 2787 | ret = wc_Sha256Final(&tmpSha256, hash); |
wolfSSL | 15:117db924cf7c | 2788 | wc_Sha256Free(&tmpSha256); |
wolfSSL | 15:117db924cf7c | 2789 | } |
wolfSSL | 15:117db924cf7c | 2790 | return ret; |
wolfSSL | 15:117db924cf7c | 2791 | } |
wolfSSL | 15:117db924cf7c | 2792 | int wc_Sha256Copy(wc_Sha256* src, wc_Sha256* dst) |
wolfSSL | 15:117db924cf7c | 2793 | { |
wolfSSL | 15:117db924cf7c | 2794 | int ret = 0; |
wolfSSL | 15:117db924cf7c | 2795 | |
wolfSSL | 15:117db924cf7c | 2796 | if (src == NULL || dst == NULL) |
wolfSSL | 15:117db924cf7c | 2797 | return BAD_FUNC_ARG; |
wolfSSL | 15:117db924cf7c | 2798 | |
wolfSSL | 15:117db924cf7c | 2799 | XMEMCPY(dst, src, sizeof(wc_Sha256)); |
wolfSSL | 15:117db924cf7c | 2800 | #ifdef WOLFSSL_SMALL_STACK_CACHE |
wolfSSL | 15:117db924cf7c | 2801 | dst->W = NULL; |
wolfSSL | 15:117db924cf7c | 2802 | #endif |
wolfSSL | 15:117db924cf7c | 2803 | |
wolfSSL | 15:117db924cf7c | 2804 | #ifdef WOLFSSL_ASYNC_CRYPT |
wolfSSL | 15:117db924cf7c | 2805 | ret = wolfAsync_DevCopy(&src->asyncDev, &dst->asyncDev); |
wolfSSL | 15:117db924cf7c | 2806 | #endif |
wolfSSL | 15:117db924cf7c | 2807 | #ifdef WOLFSSL_PIC32MZ_HASH |
wolfSSL | 15:117db924cf7c | 2808 | ret = wc_Pic32HashCopy(&src->cache, &dst->cache); |
wolfSSL | 15:117db924cf7c | 2809 | #endif |
wolfSSL | 15:117db924cf7c | 2810 | |
wolfSSL | 15:117db924cf7c | 2811 | return ret; |
wolfSSL | 15:117db924cf7c | 2812 | } |
wolfSSL | 15:117db924cf7c | 2813 | #endif /* !WOLFSSL_TI_HASH */ |
wolfSSL | 15:117db924cf7c | 2814 | |
wolfSSL | 15:117db924cf7c | 2815 | #endif /* NO_SHA256 */ |
wolfSSL | 15:117db924cf7c | 2816 |