wolfSSL-TLS13-Beta - wolfSSL 3.11.1 for TLS1.3 beta

Users » wolfSSL » Code » wolfSSL-TLS13-Beta

wolfSSL 3.11.1 for TLS1.3 beta

wolfcrypt/src/sha512.c@0:d92f9d21154c, 2015-06-26 (annotated)

Committer:: wolfSSL
Date:: Fri Jun 26 00:39:20 2015 +0000
Revision:: 0:d92f9d21154c

wolfSSL 3.6.0

Who changed what in which revision?

User	Revision	Line number	New contents of line
wolfSSL	0:d92f9d21154c	1	/* sha512.c
wolfSSL	0:d92f9d21154c	2	*
wolfSSL	0:d92f9d21154c	3	* Copyright (C) 2006-2015 wolfSSL Inc.
wolfSSL	0:d92f9d21154c	4	*
wolfSSL	0:d92f9d21154c	5	* This file is part of wolfSSL. (formerly known as CyaSSL)
wolfSSL	0:d92f9d21154c	6	*
wolfSSL	0:d92f9d21154c	7	* wolfSSL is free software; you can redistribute it and/or modify
wolfSSL	0:d92f9d21154c	8	* it under the terms of the GNU General Public License as published by
wolfSSL	0:d92f9d21154c	9	* the Free Software Foundation; either version 2 of the License, or
wolfSSL	0:d92f9d21154c	10	* (at your option) any later version.
wolfSSL	0:d92f9d21154c	11	*
wolfSSL	0:d92f9d21154c	12	* wolfSSL is distributed in the hope that it will be useful,
wolfSSL	0:d92f9d21154c	13	* but WITHOUT ANY WARRANTY; without even the implied warranty of
wolfSSL	0:d92f9d21154c	14	* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
wolfSSL	0:d92f9d21154c	15	* GNU General Public License for more details.
wolfSSL	0:d92f9d21154c	16	*
wolfSSL	0:d92f9d21154c	17	* You should have received a copy of the GNU General Public License
wolfSSL	0:d92f9d21154c	18	* along with this program; if not, write to the Free Software
wolfSSL	0:d92f9d21154c	19	* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301, USA
wolfSSL	0:d92f9d21154c	20	*/
wolfSSL	0:d92f9d21154c	21
wolfSSL	0:d92f9d21154c	22	#ifdef HAVE_CONFIG_H
wolfSSL	0:d92f9d21154c	23	#include <config.h>
wolfSSL	0:d92f9d21154c	24	#endif
wolfSSL	0:d92f9d21154c	25
wolfSSL	0:d92f9d21154c	26	#include <wolfssl/wolfcrypt/settings.h>
wolfSSL	0:d92f9d21154c	27	#include <wolfssl/wolfcrypt/sha512.h>
wolfSSL	0:d92f9d21154c	28
wolfSSL	0:d92f9d21154c	29	#ifdef WOLFSSL_SHA512
wolfSSL	0:d92f9d21154c	30
wolfSSL	0:d92f9d21154c	31	#ifdef HAVE_FIPS
wolfSSL	0:d92f9d21154c	32	int wc_InitSha512(Sha512* sha)
wolfSSL	0:d92f9d21154c	33	{
wolfSSL	0:d92f9d21154c	34	return InitSha512_fips(sha);
wolfSSL	0:d92f9d21154c	35	}
wolfSSL	0:d92f9d21154c	36
wolfSSL	0:d92f9d21154c	37
wolfSSL	0:d92f9d21154c	38	int wc_Sha512Update(Sha512* sha, const byte* data, word32 len)
wolfSSL	0:d92f9d21154c	39	{
wolfSSL	0:d92f9d21154c	40	return Sha512Update_fips(sha, data, len);
wolfSSL	0:d92f9d21154c	41	}
wolfSSL	0:d92f9d21154c	42
wolfSSL	0:d92f9d21154c	43
wolfSSL	0:d92f9d21154c	44	int wc_Sha512Final(Sha512* sha, byte* out)
wolfSSL	0:d92f9d21154c	45	{
wolfSSL	0:d92f9d21154c	46	return Sha512Final_fips(sha, out);
wolfSSL	0:d92f9d21154c	47	}
wolfSSL	0:d92f9d21154c	48
wolfSSL	0:d92f9d21154c	49
wolfSSL	0:d92f9d21154c	50	int wc_Sha512Hash(const byte* data, word32 len, byte* out)
wolfSSL	0:d92f9d21154c	51	{
wolfSSL	0:d92f9d21154c	52	return Sha512Hash(data, len, out);
wolfSSL	0:d92f9d21154c	53	}
wolfSSL	0:d92f9d21154c	54
wolfSSL	0:d92f9d21154c	55	#if defined(WOLFSSL_SHA384) \|\| defined(HAVE_AESGCM)
wolfSSL	0:d92f9d21154c	56
wolfSSL	0:d92f9d21154c	57	int wc_InitSha384(Sha384* sha)
wolfSSL	0:d92f9d21154c	58	{
wolfSSL	0:d92f9d21154c	59	return InitSha384_fips(sha);
wolfSSL	0:d92f9d21154c	60	}
wolfSSL	0:d92f9d21154c	61
wolfSSL	0:d92f9d21154c	62
wolfSSL	0:d92f9d21154c	63	int wc_Sha384Update(Sha384* sha, const byte* data, word32 len)
wolfSSL	0:d92f9d21154c	64	{
wolfSSL	0:d92f9d21154c	65	return Sha384Update_fips(sha, data, len);
wolfSSL	0:d92f9d21154c	66	}
wolfSSL	0:d92f9d21154c	67
wolfSSL	0:d92f9d21154c	68
wolfSSL	0:d92f9d21154c	69	int wc_Sha384Final(Sha384* sha, byte* out)
wolfSSL	0:d92f9d21154c	70	{
wolfSSL	0:d92f9d21154c	71	return Sha384Final_fips(sha, out);
wolfSSL	0:d92f9d21154c	72	}
wolfSSL	0:d92f9d21154c	73
wolfSSL	0:d92f9d21154c	74
wolfSSL	0:d92f9d21154c	75	int wc_Sha384Hash(const byte* data, word32 len, byte* out)
wolfSSL	0:d92f9d21154c	76	{
wolfSSL	0:d92f9d21154c	77	return Sha384Hash(data, len, out);
wolfSSL	0:d92f9d21154c	78	}
wolfSSL	0:d92f9d21154c	79	#endif /* WOLFSSL_SHA384 */
wolfSSL	0:d92f9d21154c	80	#else /* else build without using fips */
wolfSSL	0:d92f9d21154c	81	#include <wolfssl/wolfcrypt/logging.h>
wolfSSL	0:d92f9d21154c	82	#include <wolfssl/wolfcrypt/error-crypt.h>
wolfSSL	0:d92f9d21154c	83
wolfSSL	0:d92f9d21154c	84	#ifdef NO_INLINE
wolfSSL	0:d92f9d21154c	85	#include <wolfssl/wolfcrypt/misc.h>
wolfSSL	0:d92f9d21154c	86	#else
wolfSSL	0:d92f9d21154c	87	#include <wolfcrypt/src/misc.c>
wolfSSL	0:d92f9d21154c	88	#endif
wolfSSL	0:d92f9d21154c	89
wolfSSL	0:d92f9d21154c	90
wolfSSL	0:d92f9d21154c	91	#ifndef WOLFSSL_HAVE_MIN
wolfSSL	0:d92f9d21154c	92	#define WOLFSSL_HAVE_MIN
wolfSSL	0:d92f9d21154c	93
wolfSSL	0:d92f9d21154c	94	static INLINE word32 min(word32 a, word32 b)
wolfSSL	0:d92f9d21154c	95	{
wolfSSL	0:d92f9d21154c	96	return a > b ? b : a;
wolfSSL	0:d92f9d21154c	97	}
wolfSSL	0:d92f9d21154c	98
wolfSSL	0:d92f9d21154c	99	#endif /* WOLFSSL_HAVE_MIN */
wolfSSL	0:d92f9d21154c	100
wolfSSL	0:d92f9d21154c	101	#if defined(USE_INTEL_SPEEDUP)
wolfSSL	0:d92f9d21154c	102	#define HAVE_INTEL_AVX1
wolfSSL	0:d92f9d21154c	103	#define HAVE_INTEL_AVX2
wolfSSL	0:d92f9d21154c	104	#endif
wolfSSL	0:d92f9d21154c	105
wolfSSL	0:d92f9d21154c	106	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	107	/* #define DEBUG_XMM */
wolfSSL	0:d92f9d21154c	108	#endif
wolfSSL	0:d92f9d21154c	109
wolfSSL	0:d92f9d21154c	110	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	111	#define HAVE_INTEL_RORX
wolfSSL	0:d92f9d21154c	112	/* #define DEBUG_YMM */
wolfSSL	0:d92f9d21154c	113	#endif
wolfSSL	0:d92f9d21154c	114
wolfSSL	0:d92f9d21154c	115	/*****
wolfSSL	0:d92f9d21154c	116	Intel AVX1/AVX2 Macro Control Structure
wolfSSL	0:d92f9d21154c	117
wolfSSL	0:d92f9d21154c	118	#if defined(HAVE_INteL_SPEEDUP)
wolfSSL	0:d92f9d21154c	119	#define HAVE_INTEL_AVX1
wolfSSL	0:d92f9d21154c	120	#define HAVE_INTEL_AVX2
wolfSSL	0:d92f9d21154c	121	#endif
wolfSSL	0:d92f9d21154c	122
wolfSSL	0:d92f9d21154c	123	int InitSha512(Sha512* sha512) {
wolfSSL	0:d92f9d21154c	124	Save/Recover XMM, YMM
wolfSSL	0:d92f9d21154c	125	...
wolfSSL	0:d92f9d21154c	126
wolfSSL	0:d92f9d21154c	127	Check Intel AVX cpuid flags
wolfSSL	0:d92f9d21154c	128	}
wolfSSL	0:d92f9d21154c	129
wolfSSL	0:d92f9d21154c	130	#if defined(HAVE_INTEL_AVX1)\|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	131	Transform_AVX1() ; # Function prototype
wolfSSL	0:d92f9d21154c	132	Transform_AVX2() ; #
wolfSSL	0:d92f9d21154c	133	#endif
wolfSSL	0:d92f9d21154c	134
wolfSSL	0:d92f9d21154c	135	_Transform() { # Native Transform Function body
wolfSSL	0:d92f9d21154c	136
wolfSSL	0:d92f9d21154c	137	}
wolfSSL	0:d92f9d21154c	138
wolfSSL	0:d92f9d21154c	139	int Sha512Update() {
wolfSSL	0:d92f9d21154c	140	Save/Recover XMM, YMM
wolfSSL	0:d92f9d21154c	141	...
wolfSSL	0:d92f9d21154c	142	}
wolfSSL	0:d92f9d21154c	143
wolfSSL	0:d92f9d21154c	144	int Sha512Final() {
wolfSSL	0:d92f9d21154c	145	Save/Recover XMM, YMM
wolfSSL	0:d92f9d21154c	146	...
wolfSSL	0:d92f9d21154c	147	}
wolfSSL	0:d92f9d21154c	148
wolfSSL	0:d92f9d21154c	149
wolfSSL	0:d92f9d21154c	150	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	151
wolfSSL	0:d92f9d21154c	152	XMM Instructions/inline asm Definitions
wolfSSL	0:d92f9d21154c	153
wolfSSL	0:d92f9d21154c	154	#endif
wolfSSL	0:d92f9d21154c	155
wolfSSL	0:d92f9d21154c	156	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	157
wolfSSL	0:d92f9d21154c	158	YMM Instructions/inline asm Definitions
wolfSSL	0:d92f9d21154c	159
wolfSSL	0:d92f9d21154c	160	#endif
wolfSSL	0:d92f9d21154c	161
wolfSSL	0:d92f9d21154c	162	#if defnied(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	163
wolfSSL	0:d92f9d21154c	164	int Transform_AVX1() {
wolfSSL	0:d92f9d21154c	165	Stitched Message Sched/Round
wolfSSL	0:d92f9d21154c	166	}
wolfSSL	0:d92f9d21154c	167
wolfSSL	0:d92f9d21154c	168	#endif
wolfSSL	0:d92f9d21154c	169
wolfSSL	0:d92f9d21154c	170	#if defnied(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	171
wolfSSL	0:d92f9d21154c	172	int Transform_AVX2() {
wolfSSL	0:d92f9d21154c	173	Stitched Message Sched/Round
wolfSSL	0:d92f9d21154c	174	}
wolfSSL	0:d92f9d21154c	175	#endif
wolfSSL	0:d92f9d21154c	176
wolfSSL	0:d92f9d21154c	177
wolfSSL	0:d92f9d21154c	178	*/
wolfSSL	0:d92f9d21154c	179
wolfSSL	0:d92f9d21154c	180	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	181
wolfSSL	0:d92f9d21154c	182
wolfSSL	0:d92f9d21154c	183	/* Each platform needs to query info type 1 from cpuid to see if aesni is
wolfSSL	0:d92f9d21154c	184	* supported. Also, let's setup a macro for proper linkage w/o ABI conflicts
wolfSSL	0:d92f9d21154c	185	*/
wolfSSL	0:d92f9d21154c	186
wolfSSL	0:d92f9d21154c	187	#ifndef _MSC_VER
wolfSSL	0:d92f9d21154c	188	#define cpuid(reg, leaf, sub)\
wolfSSL	0:d92f9d21154c	189	__asm__ __volatile__ ("cpuid":\
wolfSSL	0:d92f9d21154c	190	"=a" (reg[0]), "=b" (reg[1]), "=c" (reg[2]), "=d" (reg[3]) :\
wolfSSL	0:d92f9d21154c	191	"a" (leaf), "c"(sub));
wolfSSL	0:d92f9d21154c	192
wolfSSL	0:d92f9d21154c	193	#define XASM_LINK(f) asm(f)
wolfSSL	0:d92f9d21154c	194	#else
wolfSSL	0:d92f9d21154c	195
wolfSSL	0:d92f9d21154c	196	#include <intrin.h>
wolfSSL	0:d92f9d21154c	197	#define cpuid(a,b) __cpuid((int*)a,b)
wolfSSL	0:d92f9d21154c	198
wolfSSL	0:d92f9d21154c	199	#define XASM_LINK(f)
wolfSSL	0:d92f9d21154c	200
wolfSSL	0:d92f9d21154c	201	#endif /* _MSC_VER */
wolfSSL	0:d92f9d21154c	202
wolfSSL	0:d92f9d21154c	203	#define EAX 0
wolfSSL	0:d92f9d21154c	204	#define EBX 1
wolfSSL	0:d92f9d21154c	205	#define ECX 2
wolfSSL	0:d92f9d21154c	206	#define EDX 3
wolfSSL	0:d92f9d21154c	207
wolfSSL	0:d92f9d21154c	208	#define CPUID_AVX1 0x1
wolfSSL	0:d92f9d21154c	209	#define CPUID_AVX2 0x2
wolfSSL	0:d92f9d21154c	210	#define CPUID_RDRAND 0x4
wolfSSL	0:d92f9d21154c	211	#define CPUID_RDSEED 0x8
wolfSSL	0:d92f9d21154c	212	#define CPUID_BMI2 0x10 /* MULX, RORX */
wolfSSL	0:d92f9d21154c	213
wolfSSL	0:d92f9d21154c	214	#define IS_INTEL_AVX1 (cpuid_flags&CPUID_AVX1)
wolfSSL	0:d92f9d21154c	215	#define IS_INTEL_AVX2 (cpuid_flags&CPUID_AVX2)
wolfSSL	0:d92f9d21154c	216	#define IS_INTEL_BMI2 (cpuid_flags&CPUID_BMI2)
wolfSSL	0:d92f9d21154c	217	#define IS_INTEL_RDRAND (cpuid_flags&CPUID_RDRAND)
wolfSSL	0:d92f9d21154c	218	#define IS_INTEL_RDSEED (cpuid_flags&CPUID_RDSEED)
wolfSSL	0:d92f9d21154c	219
wolfSSL	0:d92f9d21154c	220	static word32 cpuid_check = 0 ;
wolfSSL	0:d92f9d21154c	221	static word32 cpuid_flags = 0 ;
wolfSSL	0:d92f9d21154c	222
wolfSSL	0:d92f9d21154c	223	static word32 cpuid_flag(word32 leaf, word32 sub, word32 num, word32 bit) {
wolfSSL	0:d92f9d21154c	224	int got_intel_cpu=0;
wolfSSL	0:d92f9d21154c	225	unsigned int reg[5];
wolfSSL	0:d92f9d21154c	226
wolfSSL	0:d92f9d21154c	227	reg[4] = '\0' ;
wolfSSL	0:d92f9d21154c	228	cpuid(reg, 0, 0);
wolfSSL	0:d92f9d21154c	229	if(memcmp((char *)&(reg[EBX]), "Genu", 4) == 0 &&
wolfSSL	0:d92f9d21154c	230	memcmp((char *)&(reg[EDX]), "ineI", 4) == 0 &&
wolfSSL	0:d92f9d21154c	231	memcmp((char *)&(reg[ECX]), "ntel", 4) == 0) {
wolfSSL	0:d92f9d21154c	232	got_intel_cpu = 1;
wolfSSL	0:d92f9d21154c	233	}
wolfSSL	0:d92f9d21154c	234	if (got_intel_cpu) {
wolfSSL	0:d92f9d21154c	235	cpuid(reg, leaf, sub);
wolfSSL	0:d92f9d21154c	236	return((reg[num]>>bit)&0x1) ;
wolfSSL	0:d92f9d21154c	237	}
wolfSSL	0:d92f9d21154c	238	return 0 ;
wolfSSL	0:d92f9d21154c	239	}
wolfSSL	0:d92f9d21154c	240
wolfSSL	0:d92f9d21154c	241	#define CHECK_SHA512 0x1
wolfSSL	0:d92f9d21154c	242	#define CHECK_SHA384 0x2
wolfSSL	0:d92f9d21154c	243
wolfSSL	0:d92f9d21154c	244	static int set_cpuid_flags(int sha) {
wolfSSL	0:d92f9d21154c	245	if((cpuid_check & sha) ==0) {
wolfSSL	0:d92f9d21154c	246	if(cpuid_flag(1, 0, ECX, 28)){ cpuid_flags \|= CPUID_AVX1 ;}
wolfSSL	0:d92f9d21154c	247	if(cpuid_flag(7, 0, EBX, 5)){ cpuid_flags \|= CPUID_AVX2 ; }
wolfSSL	0:d92f9d21154c	248	if(cpuid_flag(7, 0, EBX, 8)) { cpuid_flags \|= CPUID_BMI2 ; }
wolfSSL	0:d92f9d21154c	249	if(cpuid_flag(1, 0, ECX, 30)){ cpuid_flags \|= CPUID_RDRAND ; }
wolfSSL	0:d92f9d21154c	250	if(cpuid_flag(7, 0, EBX, 18)){ cpuid_flags \|= CPUID_RDSEED ; }
wolfSSL	0:d92f9d21154c	251	cpuid_check \|= sha ;
wolfSSL	0:d92f9d21154c	252	return 0 ;
wolfSSL	0:d92f9d21154c	253	}
wolfSSL	0:d92f9d21154c	254	return 1 ;
wolfSSL	0:d92f9d21154c	255	}
wolfSSL	0:d92f9d21154c	256
wolfSSL	0:d92f9d21154c	257
wolfSSL	0:d92f9d21154c	258	/* #if defined(HAVE_INTEL_AVX1/2) at the tail of sha512 */
wolfSSL	0:d92f9d21154c	259
wolfSSL	0:d92f9d21154c	260	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	261	static int Transform_AVX1(Sha512 *sha512) ;
wolfSSL	0:d92f9d21154c	262	#endif
wolfSSL	0:d92f9d21154c	263
wolfSSL	0:d92f9d21154c	264	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	265	static int Transform_AVX2(Sha512 *sha512) ;
wolfSSL	0:d92f9d21154c	266
wolfSSL	0:d92f9d21154c	267	#if defined(HAVE_INTEL_AVX1) && defined(HAVE_INTEL_AVX2) && defined(HAVE_INTEL_RORX)
wolfSSL	0:d92f9d21154c	268	static int Transform_AVX1_RORX(Sha512 *sha512) ;
wolfSSL	0:d92f9d21154c	269	#endif
wolfSSL	0:d92f9d21154c	270
wolfSSL	0:d92f9d21154c	271	#endif
wolfSSL	0:d92f9d21154c	272
wolfSSL	0:d92f9d21154c	273	static int _Transform(Sha512 *sha512) ;
wolfSSL	0:d92f9d21154c	274
wolfSSL	0:d92f9d21154c	275	static int (Transform_p)(Sha512 sha512) = _Transform ;
wolfSSL	0:d92f9d21154c	276
wolfSSL	0:d92f9d21154c	277	#define Transform(sha512) (*Transform_p)(sha512)
wolfSSL	0:d92f9d21154c	278
wolfSSL	0:d92f9d21154c	279	static void set_Transform(void) {
wolfSSL	0:d92f9d21154c	280	if(set_cpuid_flags(CHECK_SHA512)) return ;
wolfSSL	0:d92f9d21154c	281
wolfSSL	0:d92f9d21154c	282	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	283	if(IS_INTEL_AVX2 && IS_INTEL_BMI2){
wolfSSL	0:d92f9d21154c	284	Transform_p = Transform_AVX1_RORX; return ;
wolfSSL	0:d92f9d21154c	285	Transform_p = Transform_AVX2 ;
wolfSSL	0:d92f9d21154c	286	/* for avoiding warning,"not used" */
wolfSSL	0:d92f9d21154c	287	}
wolfSSL	0:d92f9d21154c	288	#endif
wolfSSL	0:d92f9d21154c	289	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	290	Transform_p = ((IS_INTEL_AVX1) ? Transform_AVX1 : _Transform) ; return ;
wolfSSL	0:d92f9d21154c	291	#endif
wolfSSL	0:d92f9d21154c	292	Transform_p = _Transform ; return ;
wolfSSL	0:d92f9d21154c	293	}
wolfSSL	0:d92f9d21154c	294
wolfSSL	0:d92f9d21154c	295	#else
wolfSSL	0:d92f9d21154c	296	#define Transform(sha512) _Transform(sha512)
wolfSSL	0:d92f9d21154c	297	#endif
wolfSSL	0:d92f9d21154c	298
wolfSSL	0:d92f9d21154c	299	/* Dummy for saving MM_REGs on behalf of Transform */
wolfSSL	0:d92f9d21154c	300	/* #if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	301	#define SAVE_XMM_YMM __asm__ volatile("orq %%r8, %%r8":::\
wolfSSL	0:d92f9d21154c	302	"%ymm0","%ymm1","%ymm2","%ymm3","%ymm4","%ymm5","%ymm6","%ymm7","%ymm8","%ymm9","%ymm10","%ymm11",\
wolfSSL	0:d92f9d21154c	303	"%ymm12","%ymm13","%ymm14","%ymm15")
wolfSSL	0:d92f9d21154c	304	*/
wolfSSL	0:d92f9d21154c	305	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	306	#define SAVE_XMM_YMM __asm__ volatile("orq %%r8, %%r8":::\
wolfSSL	0:d92f9d21154c	307	"xmm0","xmm1","xmm2","xmm3","xmm4","xmm5","xmm6","xmm7","xmm8","xmm9","xmm10","xmm11","xmm12","xmm13","xmm14","xmm15")
wolfSSL	0:d92f9d21154c	308	#else
wolfSSL	0:d92f9d21154c	309	#define SAVE_XMM_YMM
wolfSSL	0:d92f9d21154c	310	#endif
wolfSSL	0:d92f9d21154c	311
wolfSSL	0:d92f9d21154c	312	#if defined(HAVE_INTEL_AVX1)\|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	313
wolfSSL	0:d92f9d21154c	314	#include <string.h>
wolfSSL	0:d92f9d21154c	315
wolfSSL	0:d92f9d21154c	316	#endif /* defined(HAVE_INTEL_AVX1)\|\| defined(HAVE_INTEL_AVX2) */
wolfSSL	0:d92f9d21154c	317
wolfSSL	0:d92f9d21154c	318
wolfSSL	0:d92f9d21154c	319	#if defined(HAVE_INTEL_RORX)
wolfSSL	0:d92f9d21154c	320	#define ROTR(func, bits, x) \
wolfSSL	0:d92f9d21154c	321	word64 func(word64 x) { word64 ret ;\
wolfSSL	0:d92f9d21154c	322	__asm__ ("rorx $"#bits", %1, %0\n\t":"=r"(ret):"r"(x):) ;\
wolfSSL	0:d92f9d21154c	323	return ret ;\
wolfSSL	0:d92f9d21154c	324	}
wolfSSL	0:d92f9d21154c	325
wolfSSL	0:d92f9d21154c	326	static INLINE ROTR(rotrFixed64_28, 28, x)
wolfSSL	0:d92f9d21154c	327	static INLINE ROTR(rotrFixed64_34, 34, x)
wolfSSL	0:d92f9d21154c	328	static INLINE ROTR(rotrFixed64_39, 39, x)
wolfSSL	0:d92f9d21154c	329	static INLINE ROTR(rotrFixed64_14, 14, x)
wolfSSL	0:d92f9d21154c	330	static INLINE ROTR(rotrFixed64_18, 18, x)
wolfSSL	0:d92f9d21154c	331	static INLINE ROTR(rotrFixed64_41, 41, x)
wolfSSL	0:d92f9d21154c	332
wolfSSL	0:d92f9d21154c	333	#define S0_RORX(x) (rotrFixed64_28(x)^rotrFixed64_34(x)^rotrFixed64_39(x))
wolfSSL	0:d92f9d21154c	334	#define S1_RORX(x) (rotrFixed64_14(x)^rotrFixed64_18(x)^rotrFixed64_41(x))
wolfSSL	0:d92f9d21154c	335	#endif
wolfSSL	0:d92f9d21154c	336
wolfSSL	0:d92f9d21154c	337	#if defined(HAVE_BYTEREVERSE64) && !defined(HAVE_INTEL_AVX1) && !defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	338	#define ByteReverseWords64(out, in, size) ByteReverseWords64_1(out, size)
wolfSSL	0:d92f9d21154c	339	#define ByteReverseWords64_1(buf, size)\
wolfSSL	0:d92f9d21154c	340	{ unsigned int i ;\
wolfSSL	0:d92f9d21154c	341	for(i=0; i< size/sizeof(word64); i++){\
wolfSSL	0:d92f9d21154c	342	__asm__ volatile("bswapq %0":"+r"(buf[i])::) ;\
wolfSSL	0:d92f9d21154c	343	}\
wolfSSL	0:d92f9d21154c	344	}
wolfSSL	0:d92f9d21154c	345	#endif
wolfSSL	0:d92f9d21154c	346
wolfSSL	0:d92f9d21154c	347
wolfSSL	0:d92f9d21154c	348	int wc_InitSha512(Sha512* sha512)
wolfSSL	0:d92f9d21154c	349	{
wolfSSL	0:d92f9d21154c	350	sha512->digest[0] = W64LIT(0x6a09e667f3bcc908);
wolfSSL	0:d92f9d21154c	351	sha512->digest[1] = W64LIT(0xbb67ae8584caa73b);
wolfSSL	0:d92f9d21154c	352	sha512->digest[2] = W64LIT(0x3c6ef372fe94f82b);
wolfSSL	0:d92f9d21154c	353	sha512->digest[3] = W64LIT(0xa54ff53a5f1d36f1);
wolfSSL	0:d92f9d21154c	354	sha512->digest[4] = W64LIT(0x510e527fade682d1);
wolfSSL	0:d92f9d21154c	355	sha512->digest[5] = W64LIT(0x9b05688c2b3e6c1f);
wolfSSL	0:d92f9d21154c	356	sha512->digest[6] = W64LIT(0x1f83d9abfb41bd6b);
wolfSSL	0:d92f9d21154c	357	sha512->digest[7] = W64LIT(0x5be0cd19137e2179);
wolfSSL	0:d92f9d21154c	358
wolfSSL	0:d92f9d21154c	359	sha512->buffLen = 0;
wolfSSL	0:d92f9d21154c	360	sha512->loLen = 0;
wolfSSL	0:d92f9d21154c	361	sha512->hiLen = 0;
wolfSSL	0:d92f9d21154c	362
wolfSSL	0:d92f9d21154c	363	#if defined(HAVE_INTEL_AVX1)\|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	364	set_Transform() ; /* choose best Transform function under this runtime environment */
wolfSSL	0:d92f9d21154c	365	#endif
wolfSSL	0:d92f9d21154c	366
wolfSSL	0:d92f9d21154c	367	return 0 ;
wolfSSL	0:d92f9d21154c	368	}
wolfSSL	0:d92f9d21154c	369
wolfSSL	0:d92f9d21154c	370
wolfSSL	0:d92f9d21154c	371	static const word64 K512[80] = {
wolfSSL	0:d92f9d21154c	372	W64LIT(0x428a2f98d728ae22), W64LIT(0x7137449123ef65cd),
wolfSSL	0:d92f9d21154c	373	W64LIT(0xb5c0fbcfec4d3b2f), W64LIT(0xe9b5dba58189dbbc),
wolfSSL	0:d92f9d21154c	374	W64LIT(0x3956c25bf348b538), W64LIT(0x59f111f1b605d019),
wolfSSL	0:d92f9d21154c	375	W64LIT(0x923f82a4af194f9b), W64LIT(0xab1c5ed5da6d8118),
wolfSSL	0:d92f9d21154c	376	W64LIT(0xd807aa98a3030242), W64LIT(0x12835b0145706fbe),
wolfSSL	0:d92f9d21154c	377	W64LIT(0x243185be4ee4b28c), W64LIT(0x550c7dc3d5ffb4e2),
wolfSSL	0:d92f9d21154c	378	W64LIT(0x72be5d74f27b896f), W64LIT(0x80deb1fe3b1696b1),
wolfSSL	0:d92f9d21154c	379	W64LIT(0x9bdc06a725c71235), W64LIT(0xc19bf174cf692694),
wolfSSL	0:d92f9d21154c	380	W64LIT(0xe49b69c19ef14ad2), W64LIT(0xefbe4786384f25e3),
wolfSSL	0:d92f9d21154c	381	W64LIT(0x0fc19dc68b8cd5b5), W64LIT(0x240ca1cc77ac9c65),
wolfSSL	0:d92f9d21154c	382	W64LIT(0x2de92c6f592b0275), W64LIT(0x4a7484aa6ea6e483),
wolfSSL	0:d92f9d21154c	383	W64LIT(0x5cb0a9dcbd41fbd4), W64LIT(0x76f988da831153b5),
wolfSSL	0:d92f9d21154c	384	W64LIT(0x983e5152ee66dfab), W64LIT(0xa831c66d2db43210),
wolfSSL	0:d92f9d21154c	385	W64LIT(0xb00327c898fb213f), W64LIT(0xbf597fc7beef0ee4),
wolfSSL	0:d92f9d21154c	386	W64LIT(0xc6e00bf33da88fc2), W64LIT(0xd5a79147930aa725),
wolfSSL	0:d92f9d21154c	387	W64LIT(0x06ca6351e003826f), W64LIT(0x142929670a0e6e70),
wolfSSL	0:d92f9d21154c	388	W64LIT(0x27b70a8546d22ffc), W64LIT(0x2e1b21385c26c926),
wolfSSL	0:d92f9d21154c	389	W64LIT(0x4d2c6dfc5ac42aed), W64LIT(0x53380d139d95b3df),
wolfSSL	0:d92f9d21154c	390	W64LIT(0x650a73548baf63de), W64LIT(0x766a0abb3c77b2a8),
wolfSSL	0:d92f9d21154c	391	W64LIT(0x81c2c92e47edaee6), W64LIT(0x92722c851482353b),
wolfSSL	0:d92f9d21154c	392	W64LIT(0xa2bfe8a14cf10364), W64LIT(0xa81a664bbc423001),
wolfSSL	0:d92f9d21154c	393	W64LIT(0xc24b8b70d0f89791), W64LIT(0xc76c51a30654be30),
wolfSSL	0:d92f9d21154c	394	W64LIT(0xd192e819d6ef5218), W64LIT(0xd69906245565a910),
wolfSSL	0:d92f9d21154c	395	W64LIT(0xf40e35855771202a), W64LIT(0x106aa07032bbd1b8),
wolfSSL	0:d92f9d21154c	396	W64LIT(0x19a4c116b8d2d0c8), W64LIT(0x1e376c085141ab53),
wolfSSL	0:d92f9d21154c	397	W64LIT(0x2748774cdf8eeb99), W64LIT(0x34b0bcb5e19b48a8),
wolfSSL	0:d92f9d21154c	398	W64LIT(0x391c0cb3c5c95a63), W64LIT(0x4ed8aa4ae3418acb),
wolfSSL	0:d92f9d21154c	399	W64LIT(0x5b9cca4f7763e373), W64LIT(0x682e6ff3d6b2b8a3),
wolfSSL	0:d92f9d21154c	400	W64LIT(0x748f82ee5defb2fc), W64LIT(0x78a5636f43172f60),
wolfSSL	0:d92f9d21154c	401	W64LIT(0x84c87814a1f0ab72), W64LIT(0x8cc702081a6439ec),
wolfSSL	0:d92f9d21154c	402	W64LIT(0x90befffa23631e28), W64LIT(0xa4506cebde82bde9),
wolfSSL	0:d92f9d21154c	403	W64LIT(0xbef9a3f7b2c67915), W64LIT(0xc67178f2e372532b),
wolfSSL	0:d92f9d21154c	404	W64LIT(0xca273eceea26619c), W64LIT(0xd186b8c721c0c207),
wolfSSL	0:d92f9d21154c	405	W64LIT(0xeada7dd6cde0eb1e), W64LIT(0xf57d4f7fee6ed178),
wolfSSL	0:d92f9d21154c	406	W64LIT(0x06f067aa72176fba), W64LIT(0x0a637dc5a2c898a6),
wolfSSL	0:d92f9d21154c	407	W64LIT(0x113f9804bef90dae), W64LIT(0x1b710b35131c471b),
wolfSSL	0:d92f9d21154c	408	W64LIT(0x28db77f523047d84), W64LIT(0x32caab7b40c72493),
wolfSSL	0:d92f9d21154c	409	W64LIT(0x3c9ebe0a15c9bebc), W64LIT(0x431d67c49c100d4c),
wolfSSL	0:d92f9d21154c	410	W64LIT(0x4cc5d4becb3e42b6), W64LIT(0x597f299cfc657e2a),
wolfSSL	0:d92f9d21154c	411	W64LIT(0x5fcb6fab3ad6faec), W64LIT(0x6c44198c4a475817)
wolfSSL	0:d92f9d21154c	412	};
wolfSSL	0:d92f9d21154c	413
wolfSSL	0:d92f9d21154c	414
wolfSSL	0:d92f9d21154c	415
wolfSSL	0:d92f9d21154c	416	#define blk0(i) (W[i] = sha512->buffer[i])
wolfSSL	0:d92f9d21154c	417
wolfSSL	0:d92f9d21154c	418	#define blk2(i) (W[i&15]+=s1(W[(i-2)&15])+W[(i-7)&15]+s0(W[(i-15)&15]))
wolfSSL	0:d92f9d21154c	419
wolfSSL	0:d92f9d21154c	420	#define Ch(x,y,z) (z^(x&(y^z)))
wolfSSL	0:d92f9d21154c	421	#define Maj(x,y,z) ((x&y)\|(z&(x\|y)))
wolfSSL	0:d92f9d21154c	422
wolfSSL	0:d92f9d21154c	423	#define a(i) T[(0-i)&7]
wolfSSL	0:d92f9d21154c	424	#define b(i) T[(1-i)&7]
wolfSSL	0:d92f9d21154c	425	#define c(i) T[(2-i)&7]
wolfSSL	0:d92f9d21154c	426	#define d(i) T[(3-i)&7]
wolfSSL	0:d92f9d21154c	427	#define e(i) T[(4-i)&7]
wolfSSL	0:d92f9d21154c	428	#define f(i) T[(5-i)&7]
wolfSSL	0:d92f9d21154c	429	#define g(i) T[(6-i)&7]
wolfSSL	0:d92f9d21154c	430	#define h(i) T[(7-i)&7]
wolfSSL	0:d92f9d21154c	431
wolfSSL	0:d92f9d21154c	432	#define S0(x) (rotrFixed64(x,28)^rotrFixed64(x,34)^rotrFixed64(x,39))
wolfSSL	0:d92f9d21154c	433	#define S1(x) (rotrFixed64(x,14)^rotrFixed64(x,18)^rotrFixed64(x,41))
wolfSSL	0:d92f9d21154c	434	#define s0(x) (rotrFixed64(x,1)^rotrFixed64(x,8)^(x>>7))
wolfSSL	0:d92f9d21154c	435	#define s1(x) (rotrFixed64(x,19)^rotrFixed64(x,61)^(x>>6))
wolfSSL	0:d92f9d21154c	436
wolfSSL	0:d92f9d21154c	437	#define R(i) h(i)+=S1(e(i))+Ch(e(i),f(i),g(i))+K[i+j]+(j?blk2(i):blk0(i));\
wolfSSL	0:d92f9d21154c	438	d(i)+=h(i);h(i)+=S0(a(i))+Maj(a(i),b(i),c(i))
wolfSSL	0:d92f9d21154c	439
wolfSSL	0:d92f9d21154c	440	#define blk384(i) (W[i] = sha384->buffer[i])
wolfSSL	0:d92f9d21154c	441
wolfSSL	0:d92f9d21154c	442	#define R2(i) h(i)+=S1(e(i))+Ch(e(i),f(i),g(i))+K[i+j]+(j?blk2(i):blk384(i));\
wolfSSL	0:d92f9d21154c	443	d(i)+=h(i);h(i)+=S0(a(i))+Maj(a(i),b(i),c(i))
wolfSSL	0:d92f9d21154c	444
wolfSSL	0:d92f9d21154c	445	static int _Transform(Sha512* sha512)
wolfSSL	0:d92f9d21154c	446	{
wolfSSL	0:d92f9d21154c	447	const word64* K = K512;
wolfSSL	0:d92f9d21154c	448
wolfSSL	0:d92f9d21154c	449	word32 j;
wolfSSL	0:d92f9d21154c	450	word64 T[8];
wolfSSL	0:d92f9d21154c	451
wolfSSL	0:d92f9d21154c	452
wolfSSL	0:d92f9d21154c	453	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	454	word64* W;
wolfSSL	0:d92f9d21154c	455	W = (word64) XMALLOC(sizeof(word64) 16, NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	456	if (W == NULL)
wolfSSL	0:d92f9d21154c	457	return MEMORY_E;
wolfSSL	0:d92f9d21154c	458	#else
wolfSSL	0:d92f9d21154c	459	word64 W[16];
wolfSSL	0:d92f9d21154c	460	#endif
wolfSSL	0:d92f9d21154c	461
wolfSSL	0:d92f9d21154c	462	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	463	XMEMCPY(T, sha512->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	464
wolfSSL	0:d92f9d21154c	465	#ifdef USE_SLOW_SHA2
wolfSSL	0:d92f9d21154c	466	/* over twice as small, but 50% slower */
wolfSSL	0:d92f9d21154c	467	/* 80 operations, not unrolled */
wolfSSL	0:d92f9d21154c	468	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	469	int m;
wolfSSL	0:d92f9d21154c	470	for (m = 0; m < 16; m++) { /* braces needed here for macros {} */
wolfSSL	0:d92f9d21154c	471	R(m);
wolfSSL	0:d92f9d21154c	472	}
wolfSSL	0:d92f9d21154c	473	}
wolfSSL	0:d92f9d21154c	474	#else
wolfSSL	0:d92f9d21154c	475	/* 80 operations, partially loop unrolled */
wolfSSL	0:d92f9d21154c	476	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	477	R( 0); R( 1); R( 2); R( 3);
wolfSSL	0:d92f9d21154c	478	R( 4); R( 5); R( 6); R( 7);
wolfSSL	0:d92f9d21154c	479	R( 8); R( 9); R(10); R(11);
wolfSSL	0:d92f9d21154c	480	R(12); R(13); R(14); R(15);
wolfSSL	0:d92f9d21154c	481	}
wolfSSL	0:d92f9d21154c	482	#endif /* USE_SLOW_SHA2 */
wolfSSL	0:d92f9d21154c	483
wolfSSL	0:d92f9d21154c	484	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	485
wolfSSL	0:d92f9d21154c	486	sha512->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	487	sha512->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	488	sha512->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	489	sha512->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	490	sha512->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	491	sha512->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	492	sha512->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	493	sha512->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	494
wolfSSL	0:d92f9d21154c	495	/* Wipe variables */
wolfSSL	0:d92f9d21154c	496	ForceZero(W, sizeof(word64) * 16);
wolfSSL	0:d92f9d21154c	497	ForceZero(T, sizeof(T));
wolfSSL	0:d92f9d21154c	498
wolfSSL	0:d92f9d21154c	499	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	500	XFREE(W, NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	501	#endif
wolfSSL	0:d92f9d21154c	502
wolfSSL	0:d92f9d21154c	503	return 0;
wolfSSL	0:d92f9d21154c	504	}
wolfSSL	0:d92f9d21154c	505
wolfSSL	0:d92f9d21154c	506
wolfSSL	0:d92f9d21154c	507	static INLINE void AddLength(Sha512* sha512, word32 len)
wolfSSL	0:d92f9d21154c	508	{
wolfSSL	0:d92f9d21154c	509	word32 tmp = sha512->loLen;
wolfSSL	0:d92f9d21154c	510	if ( (sha512->loLen += len) < tmp)
wolfSSL	0:d92f9d21154c	511	sha512->hiLen++; /* carry low to high */
wolfSSL	0:d92f9d21154c	512	}
wolfSSL	0:d92f9d21154c	513
wolfSSL	0:d92f9d21154c	514	int wc_Sha512Update(Sha512* sha512, const byte* data, word32 len)
wolfSSL	0:d92f9d21154c	515	{
wolfSSL	0:d92f9d21154c	516	/* do block size increments */
wolfSSL	0:d92f9d21154c	517	byte* local = (byte*)sha512->buffer;
wolfSSL	0:d92f9d21154c	518	SAVE_XMM_YMM ; /* for Intel AVX */
wolfSSL	0:d92f9d21154c	519
wolfSSL	0:d92f9d21154c	520	while (len) {
wolfSSL	0:d92f9d21154c	521	word32 add = min(len, SHA512_BLOCK_SIZE - sha512->buffLen);
wolfSSL	0:d92f9d21154c	522	XMEMCPY(&local[sha512->buffLen], data, add);
wolfSSL	0:d92f9d21154c	523
wolfSSL	0:d92f9d21154c	524	sha512->buffLen += add;
wolfSSL	0:d92f9d21154c	525	data += add;
wolfSSL	0:d92f9d21154c	526	len -= add;
wolfSSL	0:d92f9d21154c	527
wolfSSL	0:d92f9d21154c	528	if (sha512->buffLen == SHA512_BLOCK_SIZE) {
wolfSSL	0:d92f9d21154c	529	int ret;
wolfSSL	0:d92f9d21154c	530	#if defined(LITTLE_ENDIAN_ORDER)
wolfSSL	0:d92f9d21154c	531	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	532	if(!IS_INTEL_AVX1 && !IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	533	#endif
wolfSSL	0:d92f9d21154c	534	ByteReverseWords64(sha512->buffer, sha512->buffer,
wolfSSL	0:d92f9d21154c	535	SHA512_BLOCK_SIZE);
wolfSSL	0:d92f9d21154c	536	#endif
wolfSSL	0:d92f9d21154c	537	ret = Transform(sha512);
wolfSSL	0:d92f9d21154c	538	if (ret != 0)
wolfSSL	0:d92f9d21154c	539	return ret;
wolfSSL	0:d92f9d21154c	540
wolfSSL	0:d92f9d21154c	541	AddLength(sha512, SHA512_BLOCK_SIZE);
wolfSSL	0:d92f9d21154c	542	sha512->buffLen = 0;
wolfSSL	0:d92f9d21154c	543	}
wolfSSL	0:d92f9d21154c	544	}
wolfSSL	0:d92f9d21154c	545	return 0;
wolfSSL	0:d92f9d21154c	546	}
wolfSSL	0:d92f9d21154c	547
wolfSSL	0:d92f9d21154c	548
wolfSSL	0:d92f9d21154c	549	int wc_Sha512Final(Sha512* sha512, byte* hash)
wolfSSL	0:d92f9d21154c	550	{
wolfSSL	0:d92f9d21154c	551	byte* local = (byte*)sha512->buffer;
wolfSSL	0:d92f9d21154c	552	int ret;
wolfSSL	0:d92f9d21154c	553
wolfSSL	0:d92f9d21154c	554	SAVE_XMM_YMM ; /* for Intel AVX */
wolfSSL	0:d92f9d21154c	555	AddLength(sha512, sha512->buffLen); /* before adding pads */
wolfSSL	0:d92f9d21154c	556
wolfSSL	0:d92f9d21154c	557	local[sha512->buffLen++] = 0x80; /* add 1 */
wolfSSL	0:d92f9d21154c	558
wolfSSL	0:d92f9d21154c	559	/* pad with zeros */
wolfSSL	0:d92f9d21154c	560	if (sha512->buffLen > SHA512_PAD_SIZE) {
wolfSSL	0:d92f9d21154c	561	XMEMSET(&local[sha512->buffLen], 0, SHA512_BLOCK_SIZE -sha512->buffLen);
wolfSSL	0:d92f9d21154c	562	sha512->buffLen += SHA512_BLOCK_SIZE - sha512->buffLen;
wolfSSL	0:d92f9d21154c	563	#if defined(LITTLE_ENDIAN_ORDER)
wolfSSL	0:d92f9d21154c	564	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	565	if(!IS_INTEL_AVX1 && !IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	566	#endif
wolfSSL	0:d92f9d21154c	567	ByteReverseWords64(sha512->buffer,sha512->buffer,SHA512_BLOCK_SIZE);
wolfSSL	0:d92f9d21154c	568	#endif
wolfSSL	0:d92f9d21154c	569	ret = Transform(sha512);
wolfSSL	0:d92f9d21154c	570	if (ret != 0)
wolfSSL	0:d92f9d21154c	571	return ret;
wolfSSL	0:d92f9d21154c	572
wolfSSL	0:d92f9d21154c	573	sha512->buffLen = 0;
wolfSSL	0:d92f9d21154c	574	}
wolfSSL	0:d92f9d21154c	575	XMEMSET(&local[sha512->buffLen], 0, SHA512_PAD_SIZE - sha512->buffLen);
wolfSSL	0:d92f9d21154c	576
wolfSSL	0:d92f9d21154c	577	/* put lengths in bits */
wolfSSL	0:d92f9d21154c	578	sha512->hiLen = (sha512->loLen >> (8*sizeof(sha512->loLen) - 3)) +
wolfSSL	0:d92f9d21154c	579	(sha512->hiLen << 3);
wolfSSL	0:d92f9d21154c	580	sha512->loLen = sha512->loLen << 3;
wolfSSL	0:d92f9d21154c	581
wolfSSL	0:d92f9d21154c	582	/* store lengths */
wolfSSL	0:d92f9d21154c	583	#if defined(LITTLE_ENDIAN_ORDER)
wolfSSL	0:d92f9d21154c	584	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	585	if(!IS_INTEL_AVX1 && !IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	586	#endif
wolfSSL	0:d92f9d21154c	587	ByteReverseWords64(sha512->buffer, sha512->buffer, SHA512_PAD_SIZE);
wolfSSL	0:d92f9d21154c	588	#endif
wolfSSL	0:d92f9d21154c	589	/* ! length ordering dependent on digest endian type ! */
wolfSSL	0:d92f9d21154c	590
wolfSSL	0:d92f9d21154c	591	sha512->buffer[SHA512_BLOCK_SIZE / sizeof(word64) - 2] = sha512->hiLen;
wolfSSL	0:d92f9d21154c	592	sha512->buffer[SHA512_BLOCK_SIZE / sizeof(word64) - 1] = sha512->loLen;
wolfSSL	0:d92f9d21154c	593	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	594	if(IS_INTEL_AVX1 \|\| IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	595	ByteReverseWords64(&(sha512->buffer[SHA512_BLOCK_SIZE / sizeof(word64) - 2]),
wolfSSL	0:d92f9d21154c	596	&(sha512->buffer[SHA512_BLOCK_SIZE / sizeof(word64) - 2]),
wolfSSL	0:d92f9d21154c	597	SHA512_BLOCK_SIZE - SHA512_PAD_SIZE);
wolfSSL	0:d92f9d21154c	598	#endif
wolfSSL	0:d92f9d21154c	599	ret = Transform(sha512);
wolfSSL	0:d92f9d21154c	600	if (ret != 0)
wolfSSL	0:d92f9d21154c	601	return ret;
wolfSSL	0:d92f9d21154c	602
wolfSSL	0:d92f9d21154c	603	#ifdef LITTLE_ENDIAN_ORDER
wolfSSL	0:d92f9d21154c	604	ByteReverseWords64(sha512->digest, sha512->digest, SHA512_DIGEST_SIZE);
wolfSSL	0:d92f9d21154c	605	#endif
wolfSSL	0:d92f9d21154c	606	XMEMCPY(hash, sha512->digest, SHA512_DIGEST_SIZE);
wolfSSL	0:d92f9d21154c	607
wolfSSL	0:d92f9d21154c	608	return wc_InitSha512(sha512); /* reset state */
wolfSSL	0:d92f9d21154c	609	}
wolfSSL	0:d92f9d21154c	610
wolfSSL	0:d92f9d21154c	611
wolfSSL	0:d92f9d21154c	612	int wc_Sha512Hash(const byte* data, word32 len, byte* hash)
wolfSSL	0:d92f9d21154c	613	{
wolfSSL	0:d92f9d21154c	614	int ret = 0;
wolfSSL	0:d92f9d21154c	615	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	616	Sha512* sha512;
wolfSSL	0:d92f9d21154c	617	#else
wolfSSL	0:d92f9d21154c	618	Sha512 sha512[1];
wolfSSL	0:d92f9d21154c	619	#endif
wolfSSL	0:d92f9d21154c	620
wolfSSL	0:d92f9d21154c	621	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	622	sha512 = (Sha512*)XMALLOC(sizeof(Sha512), NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	623	if (sha512 == NULL)
wolfSSL	0:d92f9d21154c	624	return MEMORY_E;
wolfSSL	0:d92f9d21154c	625	#endif
wolfSSL	0:d92f9d21154c	626
wolfSSL	0:d92f9d21154c	627	if ((ret = wc_InitSha512(sha512)) != 0) {
wolfSSL	0:d92f9d21154c	628	WOLFSSL_MSG("InitSha512 failed");
wolfSSL	0:d92f9d21154c	629	}
wolfSSL	0:d92f9d21154c	630	else if ((ret = wc_Sha512Update(sha512, data, len)) != 0) {
wolfSSL	0:d92f9d21154c	631	WOLFSSL_MSG("Sha512Update failed");
wolfSSL	0:d92f9d21154c	632	}
wolfSSL	0:d92f9d21154c	633	else if ((ret = wc_Sha512Final(sha512, hash)) != 0) {
wolfSSL	0:d92f9d21154c	634	WOLFSSL_MSG("Sha512Final failed");
wolfSSL	0:d92f9d21154c	635	}
wolfSSL	0:d92f9d21154c	636
wolfSSL	0:d92f9d21154c	637	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	638	XFREE(sha512, NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	639	#endif
wolfSSL	0:d92f9d21154c	640
wolfSSL	0:d92f9d21154c	641	return ret;
wolfSSL	0:d92f9d21154c	642	}
wolfSSL	0:d92f9d21154c	643
wolfSSL	0:d92f9d21154c	644	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	645
wolfSSL	0:d92f9d21154c	646	#define Rx_1(i) h(i)+=S1(e(i))+Ch(e(i),f(i),g(i))+K[i+j] + W_X[i] ;
wolfSSL	0:d92f9d21154c	647	#define Rx_2(i) d(i)+=h(i);
wolfSSL	0:d92f9d21154c	648	#define Rx_3(i) h(i)+=S0(a(i))+Maj(a(i),b(i),c(i));
wolfSSL	0:d92f9d21154c	649
wolfSSL	0:d92f9d21154c	650	#if defined(HAVE_INTEL_RORX)
wolfSSL	0:d92f9d21154c	651	#define Rx_RORX_1(i) h(i)+=S1_RORX(e(i))+Ch(e(i),f(i),g(i))+K[i+j] + W_X[i] ;
wolfSSL	0:d92f9d21154c	652	#define Rx_RORX_2(i) d(i)+=h(i);
wolfSSL	0:d92f9d21154c	653	#define Rx_RORX_3(i) h(i)+=S0_RORX(a(i))+Maj(a(i),b(i),c(i));
wolfSSL	0:d92f9d21154c	654	#endif
wolfSSL	0:d92f9d21154c	655
wolfSSL	0:d92f9d21154c	656	#endif
wolfSSL	0:d92f9d21154c	657
wolfSSL	0:d92f9d21154c	658	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	659	#define Ry_1(i, w) h(i)+=S1(e(i))+Ch(e(i),f(i),g(i))+K[i+j] + w ;
wolfSSL	0:d92f9d21154c	660	#define Ry_2(i, w) d(i)+=h(i);
wolfSSL	0:d92f9d21154c	661	#define Ry_3(i, w) h(i)+=S0(a(i))+Maj(a(i),b(i),c(i));
wolfSSL	0:d92f9d21154c	662	#endif
wolfSSL	0:d92f9d21154c	663
wolfSSL	0:d92f9d21154c	664	#if defined(HAVE_INTEL_AVX1) /* inline Assember for Intel AVX1 instructions */
wolfSSL	0:d92f9d21154c	665	#if defined(DEBUG_XMM)
wolfSSL	0:d92f9d21154c	666
wolfSSL	0:d92f9d21154c	667	#define SAVE_REG(i) __asm__ volatile("vmovdqu %%xmm"#i", %0 \n\t":"=m"(reg[i][0])::XMM_REGs);
wolfSSL	0:d92f9d21154c	668	#define RECV_REG(i) __asm__ volatile("vmovdqu %0, %%xmm"#i" \n\t"::"m"(reg[i][0]):XMM_REGs);
wolfSSL	0:d92f9d21154c	669
wolfSSL	0:d92f9d21154c	670	#define _DUMP_REG(REG, name)\
wolfSSL	0:d92f9d21154c	671	{ word64 buf[16] ;word64 reg[16][2];int k ;\
wolfSSL	0:d92f9d21154c	672	SAVE_REG(0); SAVE_REG(1); SAVE_REG(2); SAVE_REG(3); SAVE_REG(4); \
wolfSSL	0:d92f9d21154c	673	SAVE_REG(5); SAVE_REG(6); SAVE_REG(7);SAVE_REG(8); SAVE_REG(9); SAVE_REG(10);\
wolfSSL	0:d92f9d21154c	674	SAVE_REG(11); SAVE_REG(12); SAVE_REG(13); SAVE_REG(14); SAVE_REG(15); \
wolfSSL	0:d92f9d21154c	675	__asm__ volatile("vmovdqu %%"#REG", %0 \n\t":"=m"(buf[0])::XMM_REGs);\
wolfSSL	0:d92f9d21154c	676	printf(" "#name":\t") ; for(k=0; k<2; k++) printf("%016lx.", (word64)(buf[k])); printf("\n") ; \
wolfSSL	0:d92f9d21154c	677	RECV_REG(0); RECV_REG(1); RECV_REG(2); RECV_REG(3); RECV_REG(4);\
wolfSSL	0:d92f9d21154c	678	RECV_REG(5); RECV_REG(6); RECV_REG(7); RECV_REG(8); RECV_REG(9);\
wolfSSL	0:d92f9d21154c	679	RECV_REG(10); RECV_REG(11); RECV_REG(12); RECV_REG(13); RECV_REG(14); RECV_REG(15);\
wolfSSL	0:d92f9d21154c	680	}
wolfSSL	0:d92f9d21154c	681
wolfSSL	0:d92f9d21154c	682	#define DUMP_REG(REG) _DUMP_REG(REG, #REG)
wolfSSL	0:d92f9d21154c	683	#define PRINTF(fmt, ...)
wolfSSL	0:d92f9d21154c	684
wolfSSL	0:d92f9d21154c	685	#else
wolfSSL	0:d92f9d21154c	686
wolfSSL	0:d92f9d21154c	687	#define DUMP_REG(REG)
wolfSSL	0:d92f9d21154c	688	#define PRINTF(fmt, ...)
wolfSSL	0:d92f9d21154c	689
wolfSSL	0:d92f9d21154c	690	#endif
wolfSSL	0:d92f9d21154c	691
wolfSSL	0:d92f9d21154c	692	#define _MOVE_to_REG(xymm, mem) __asm__ volatile("vmovdqu %0, %%"#xymm" "\
wolfSSL	0:d92f9d21154c	693	:: "m"(mem):XMM_REGs) ;
wolfSSL	0:d92f9d21154c	694	#define _MOVE_to_MEM(mem,i, xymm) __asm__ volatile("vmovdqu %%"#xymm", %0" :\
wolfSSL	0:d92f9d21154c	695	"=m"(mem[i]),"=m"(mem[i+1]),"=m"(mem[i+2]),"=m"(mem[i+3])::XMM_REGs) ;
wolfSSL	0:d92f9d21154c	696	#define _MOVE(dest, src) __asm__ volatile("vmovdqu %%"#src", %%"\
wolfSSL	0:d92f9d21154c	697	#dest" ":::XMM_REGs) ;
wolfSSL	0:d92f9d21154c	698
wolfSSL	0:d92f9d21154c	699	#define _S_TEMP(dest, src, bits, temp) __asm__ volatile("vpsrlq $"#bits", %%"\
wolfSSL	0:d92f9d21154c	700	#src", %%"#dest"\n\tvpsllq $64-"#bits", %%"#src", %%"#temp"\n\tvpor %%"\
wolfSSL	0:d92f9d21154c	701	#temp",%%"#dest", %%"#dest" ":::XMM_REGs) ;
wolfSSL	0:d92f9d21154c	702	#define _AVX1_R(dest, src, bits) __asm__ volatile("vpsrlq $"#bits", %%"\
wolfSSL	0:d92f9d21154c	703	#src", %%"#dest" ":::XMM_REGs) ;
wolfSSL	0:d92f9d21154c	704	#define _XOR(dest, src1, src2) __asm__ volatile("vpxor %%"#src1", %%"\
wolfSSL	0:d92f9d21154c	705	#src2", %%"#dest" ":::XMM_REGs) ;
wolfSSL	0:d92f9d21154c	706	#define _OR(dest, src1, src2) __asm__ volatile("vpor %%"#src1", %%"\
wolfSSL	0:d92f9d21154c	707	#src2", %%"#dest" ":::XMM_REGs) ;
wolfSSL	0:d92f9d21154c	708	#define _ADD(dest, src1, src2) __asm__ volatile("vpaddq %%"#src1", %%"\
wolfSSL	0:d92f9d21154c	709	#src2", %%"#dest" ":::XMM_REGs) ;
wolfSSL	0:d92f9d21154c	710	#define _ADD_MEM(dest, src1, mem) __asm__ volatile("vpaddq %0, %%"#src1", %%"\
wolfSSL	0:d92f9d21154c	711	#dest" "::"m"(mem):XMM_REGs) ;
wolfSSL	0:d92f9d21154c	712
wolfSSL	0:d92f9d21154c	713	#define MOVE_to_REG(xymm, mem) _MOVE_to_REG(xymm, mem)
wolfSSL	0:d92f9d21154c	714	#define MOVE_to_MEM(mem, i, xymm) _MOVE_to_MEM(mem, i, xymm)
wolfSSL	0:d92f9d21154c	715	#define MOVE(dest, src) _MOVE(dest, src)
wolfSSL	0:d92f9d21154c	716
wolfSSL	0:d92f9d21154c	717	#define XOR(dest, src1, src2) _XOR(dest, src1, src2)
wolfSSL	0:d92f9d21154c	718	#define OR(dest, src1, src2) _OR(dest, src1, src2)
wolfSSL	0:d92f9d21154c	719	#define ADD(dest, src1, src2) _ADD(dest, src1, src2)
wolfSSL	0:d92f9d21154c	720
wolfSSL	0:d92f9d21154c	721	#define S_TMP(dest, src, bits, temp) _S_TEMP(dest, src, bits, temp);
wolfSSL	0:d92f9d21154c	722	#define AVX1_S(dest, src, bits) S_TMP(dest, src, bits, S_TEMP)
wolfSSL	0:d92f9d21154c	723	#define AVX1_R(dest, src, bits) _AVX1_R(dest, src, bits)
wolfSSL	0:d92f9d21154c	724
wolfSSL	0:d92f9d21154c	725	#define Init_Mask(mask) \
wolfSSL	0:d92f9d21154c	726	__asm__ volatile("vmovdqu %0, %%xmm1\n\t"::"m"(mask):"%xmm1") ;
wolfSSL	0:d92f9d21154c	727
wolfSSL	0:d92f9d21154c	728	#define _W_from_buff1(w, buff, xmm) \
wolfSSL	0:d92f9d21154c	729	/* X0..3(xmm4..7), W[0..15] = sha512->buffer[0.15]; */\
wolfSSL	0:d92f9d21154c	730	__asm__ volatile("vmovdqu %1, %%"#xmm"\n\t"\
wolfSSL	0:d92f9d21154c	731	"vpshufb %%xmm1, %%"#xmm", %%"#xmm"\n\t"\
wolfSSL	0:d92f9d21154c	732	"vmovdqu %%"#xmm", %0"\
wolfSSL	0:d92f9d21154c	733	:"=m"(w): "m"(buff):"%xmm0") ;
wolfSSL	0:d92f9d21154c	734
wolfSSL	0:d92f9d21154c	735	#define W_from_buff1(w, buff, xmm) _W_from_buff1(w, buff, xmm)
wolfSSL	0:d92f9d21154c	736
wolfSSL	0:d92f9d21154c	737	#define W_from_buff(w, buff)\
wolfSSL	0:d92f9d21154c	738	Init_Mask(mBYTE_FLIP_MASK[0]) ;\
wolfSSL	0:d92f9d21154c	739	W_from_buff1(w[0], buff[0], W_0);\
wolfSSL	0:d92f9d21154c	740	W_from_buff1(w[2], buff[2], W_2);\
wolfSSL	0:d92f9d21154c	741	W_from_buff1(w[4], buff[4], W_4);\
wolfSSL	0:d92f9d21154c	742	W_from_buff1(w[6], buff[6], W_6);\
wolfSSL	0:d92f9d21154c	743	W_from_buff1(w[8], buff[8], W_8);\
wolfSSL	0:d92f9d21154c	744	W_from_buff1(w[10],buff[10],W_10);\
wolfSSL	0:d92f9d21154c	745	W_from_buff1(w[12],buff[12],W_12);\
wolfSSL	0:d92f9d21154c	746	W_from_buff1(w[14],buff[14],W_14);
wolfSSL	0:d92f9d21154c	747
wolfSSL	0:d92f9d21154c	748	static word64 mBYTE_FLIP_MASK[] = { 0x0001020304050607, 0x08090a0b0c0d0e0f } ;
wolfSSL	0:d92f9d21154c	749
wolfSSL	0:d92f9d21154c	750	#define W_I_15 xmm14
wolfSSL	0:d92f9d21154c	751	#define W_I_7 xmm11
wolfSSL	0:d92f9d21154c	752	#define W_I_2 xmm13
wolfSSL	0:d92f9d21154c	753	#define W_I xmm12
wolfSSL	0:d92f9d21154c	754	#define G_TEMP xmm0
wolfSSL	0:d92f9d21154c	755	#define S_TEMP xmm1
wolfSSL	0:d92f9d21154c	756	#define XMM_TEMP0 xmm2
wolfSSL	0:d92f9d21154c	757
wolfSSL	0:d92f9d21154c	758	#define W_0 xmm12
wolfSSL	0:d92f9d21154c	759	#define W_2 xmm3
wolfSSL	0:d92f9d21154c	760	#define W_4 xmm4
wolfSSL	0:d92f9d21154c	761	#define W_6 xmm5
wolfSSL	0:d92f9d21154c	762	#define W_8 xmm6
wolfSSL	0:d92f9d21154c	763	#define W_10 xmm7
wolfSSL	0:d92f9d21154c	764	#define W_12 xmm8
wolfSSL	0:d92f9d21154c	765	#define W_14 xmm9
wolfSSL	0:d92f9d21154c	766
wolfSSL	0:d92f9d21154c	767	#define XMM_REGs
wolfSSL	0:d92f9d21154c	768
wolfSSL	0:d92f9d21154c	769	#define s0_1(dest, src) AVX1_S(dest, src, 1);
wolfSSL	0:d92f9d21154c	770	#define s0_2(dest, src) AVX1_S(G_TEMP, src, 8); XOR(dest, G_TEMP, dest) ;
wolfSSL	0:d92f9d21154c	771	#define s0_3(dest, src) AVX1_R(G_TEMP, src, 7); XOR(dest, G_TEMP, dest) ;
wolfSSL	0:d92f9d21154c	772
wolfSSL	0:d92f9d21154c	773	#define s1_1(dest, src) AVX1_S(dest, src, 19);
wolfSSL	0:d92f9d21154c	774	#define s1_2(dest, src) AVX1_S(G_TEMP, src, 61); XOR(dest, G_TEMP, dest) ;
wolfSSL	0:d92f9d21154c	775	#define s1_3(dest, src) AVX1_R(G_TEMP, src, 6); XOR(dest, G_TEMP, dest) ;
wolfSSL	0:d92f9d21154c	776
wolfSSL	0:d92f9d21154c	777	#define s0_(dest, src) s0_1(dest, src) ; s0_2(dest, src) ; s0_3(dest, src)
wolfSSL	0:d92f9d21154c	778	#define s1_(dest, src) s1_1(dest, src) ; s1_2(dest, src) ; s1_3(dest, src)
wolfSSL	0:d92f9d21154c	779
wolfSSL	0:d92f9d21154c	780	#define Block_xx_1(i) \
wolfSSL	0:d92f9d21154c	781	MOVE_to_REG(W_I_15, W_X[(i-15)&15]) ;\
wolfSSL	0:d92f9d21154c	782	MOVE_to_REG(W_I_7, W_X[(i- 7)&15]) ;\
wolfSSL	0:d92f9d21154c	783
wolfSSL	0:d92f9d21154c	784	#define Block_xx_2(i) \
wolfSSL	0:d92f9d21154c	785	MOVE_to_REG(W_I_2, W_X[(i- 2)&15]) ;\
wolfSSL	0:d92f9d21154c	786	MOVE_to_REG(W_I, W_X[(i)]) ;\
wolfSSL	0:d92f9d21154c	787
wolfSSL	0:d92f9d21154c	788	#define Block_xx_3(i) \
wolfSSL	0:d92f9d21154c	789	s0_ (XMM_TEMP0, W_I_15) ;\
wolfSSL	0:d92f9d21154c	790
wolfSSL	0:d92f9d21154c	791	#define Block_xx_4(i) \
wolfSSL	0:d92f9d21154c	792	ADD(W_I, W_I, XMM_TEMP0) ;\
wolfSSL	0:d92f9d21154c	793	ADD(W_I, W_I, W_I_7) ;\
wolfSSL	0:d92f9d21154c	794
wolfSSL	0:d92f9d21154c	795	#define Block_xx_5(i) \
wolfSSL	0:d92f9d21154c	796	s1_ (XMM_TEMP0, W_I_2) ;\
wolfSSL	0:d92f9d21154c	797
wolfSSL	0:d92f9d21154c	798	#define Block_xx_6(i) \
wolfSSL	0:d92f9d21154c	799	ADD(W_I, W_I, XMM_TEMP0) ;\
wolfSSL	0:d92f9d21154c	800	MOVE_to_MEM(W_X,i, W_I) ;\
wolfSSL	0:d92f9d21154c	801	if(i==0)\
wolfSSL	0:d92f9d21154c	802	MOVE_to_MEM(W_X,16, W_I) ;\
wolfSSL	0:d92f9d21154c	803
wolfSSL	0:d92f9d21154c	804	#define Block_xx_7(i) \
wolfSSL	0:d92f9d21154c	805	MOVE_to_REG(W_I_15, W_X[(i-15)&15]) ;\
wolfSSL	0:d92f9d21154c	806	MOVE_to_REG(W_I_7, W_X[(i- 7)&15]) ;\
wolfSSL	0:d92f9d21154c	807
wolfSSL	0:d92f9d21154c	808	#define Block_xx_8(i) \
wolfSSL	0:d92f9d21154c	809	MOVE_to_REG(W_I_2, W_X[(i- 2)&15]) ;\
wolfSSL	0:d92f9d21154c	810	MOVE_to_REG(W_I, W_X[(i)]) ;\
wolfSSL	0:d92f9d21154c	811
wolfSSL	0:d92f9d21154c	812	#define Block_xx_9(i) \
wolfSSL	0:d92f9d21154c	813	s0_ (XMM_TEMP0, W_I_15) ;\
wolfSSL	0:d92f9d21154c	814
wolfSSL	0:d92f9d21154c	815	#define Block_xx_10(i) \
wolfSSL	0:d92f9d21154c	816	ADD(W_I, W_I, XMM_TEMP0) ;\
wolfSSL	0:d92f9d21154c	817	ADD(W_I, W_I, W_I_7) ;\
wolfSSL	0:d92f9d21154c	818
wolfSSL	0:d92f9d21154c	819	#define Block_xx_11(i) \
wolfSSL	0:d92f9d21154c	820	s1_ (XMM_TEMP0, W_I_2) ;\
wolfSSL	0:d92f9d21154c	821
wolfSSL	0:d92f9d21154c	822	#define Block_xx_12(i) \
wolfSSL	0:d92f9d21154c	823	ADD(W_I, W_I, XMM_TEMP0) ;\
wolfSSL	0:d92f9d21154c	824	MOVE_to_MEM(W_X,i, W_I) ;\
wolfSSL	0:d92f9d21154c	825	if((i)==0)\
wolfSSL	0:d92f9d21154c	826	MOVE_to_MEM(W_X,16, W_I) ;\
wolfSSL	0:d92f9d21154c	827
wolfSSL	0:d92f9d21154c	828	static inline void Block_0_1(word64 *W_X) { Block_xx_1(0) ; }
wolfSSL	0:d92f9d21154c	829	static inline void Block_0_2(word64 *W_X) { Block_xx_2(0) ; }
wolfSSL	0:d92f9d21154c	830	static inline void Block_0_3(void) { Block_xx_3(0) ; }
wolfSSL	0:d92f9d21154c	831	static inline void Block_0_4(void) { Block_xx_4(0) ; }
wolfSSL	0:d92f9d21154c	832	static inline void Block_0_5(void) { Block_xx_5(0) ; }
wolfSSL	0:d92f9d21154c	833	static inline void Block_0_6(word64 *W_X) { Block_xx_6(0) ; }
wolfSSL	0:d92f9d21154c	834	static inline void Block_0_7(word64 *W_X) { Block_xx_7(2) ; }
wolfSSL	0:d92f9d21154c	835	static inline void Block_0_8(word64 *W_X) { Block_xx_8(2) ; }
wolfSSL	0:d92f9d21154c	836	static inline void Block_0_9(void) { Block_xx_9(2) ; }
wolfSSL	0:d92f9d21154c	837	static inline void Block_0_10(void){ Block_xx_10(2) ; }
wolfSSL	0:d92f9d21154c	838	static inline void Block_0_11(void){ Block_xx_11(2) ; }
wolfSSL	0:d92f9d21154c	839	static inline void Block_0_12(word64 *W_X){ Block_xx_12(2) ; }
wolfSSL	0:d92f9d21154c	840
wolfSSL	0:d92f9d21154c	841	static inline void Block_4_1(word64 *W_X) { Block_xx_1(4) ; }
wolfSSL	0:d92f9d21154c	842	static inline void Block_4_2(word64 *W_X) { Block_xx_2(4) ; }
wolfSSL	0:d92f9d21154c	843	static inline void Block_4_3(void) { Block_xx_3(4) ; }
wolfSSL	0:d92f9d21154c	844	static inline void Block_4_4(void) { Block_xx_4(4) ; }
wolfSSL	0:d92f9d21154c	845	static inline void Block_4_5(void) { Block_xx_5(4) ; }
wolfSSL	0:d92f9d21154c	846	static inline void Block_4_6(word64 *W_X) { Block_xx_6(4) ; }
wolfSSL	0:d92f9d21154c	847	static inline void Block_4_7(word64 *W_X) { Block_xx_7(6) ; }
wolfSSL	0:d92f9d21154c	848	static inline void Block_4_8(word64 *W_X) { Block_xx_8(6) ; }
wolfSSL	0:d92f9d21154c	849	static inline void Block_4_9(void) { Block_xx_9(6) ; }
wolfSSL	0:d92f9d21154c	850	static inline void Block_4_10(void){ Block_xx_10(6) ; }
wolfSSL	0:d92f9d21154c	851	static inline void Block_4_11(void){ Block_xx_11(6) ; }
wolfSSL	0:d92f9d21154c	852	static inline void Block_4_12(word64 *W_X){ Block_xx_12(6) ; }
wolfSSL	0:d92f9d21154c	853
wolfSSL	0:d92f9d21154c	854	static inline void Block_8_1(word64 *W_X) { Block_xx_1(8) ; }
wolfSSL	0:d92f9d21154c	855	static inline void Block_8_2(word64 *W_X) { Block_xx_2(8) ; }
wolfSSL	0:d92f9d21154c	856	static inline void Block_8_3(void) { Block_xx_3(8) ; }
wolfSSL	0:d92f9d21154c	857	static inline void Block_8_4(void) { Block_xx_4(8) ; }
wolfSSL	0:d92f9d21154c	858	static inline void Block_8_5(void) { Block_xx_5(8) ; }
wolfSSL	0:d92f9d21154c	859	static inline void Block_8_6(word64 *W_X) { Block_xx_6(8) ; }
wolfSSL	0:d92f9d21154c	860	static inline void Block_8_7(word64 *W_X) { Block_xx_7(10) ; }
wolfSSL	0:d92f9d21154c	861	static inline void Block_8_8(word64 *W_X) { Block_xx_8(10) ; }
wolfSSL	0:d92f9d21154c	862	static inline void Block_8_9(void) { Block_xx_9(10) ; }
wolfSSL	0:d92f9d21154c	863	static inline void Block_8_10(void){ Block_xx_10(10) ; }
wolfSSL	0:d92f9d21154c	864	static inline void Block_8_11(void){ Block_xx_11(10) ; }
wolfSSL	0:d92f9d21154c	865	static inline void Block_8_12(word64 *W_X){ Block_xx_12(10) ; }
wolfSSL	0:d92f9d21154c	866
wolfSSL	0:d92f9d21154c	867	static inline void Block_12_1(word64 *W_X) { Block_xx_1(12) ; }
wolfSSL	0:d92f9d21154c	868	static inline void Block_12_2(word64 *W_X) { Block_xx_2(12) ; }
wolfSSL	0:d92f9d21154c	869	static inline void Block_12_3(void) { Block_xx_3(12) ; }
wolfSSL	0:d92f9d21154c	870	static inline void Block_12_4(void) { Block_xx_4(12) ; }
wolfSSL	0:d92f9d21154c	871	static inline void Block_12_5(void) { Block_xx_5(12) ; }
wolfSSL	0:d92f9d21154c	872	static inline void Block_12_6(word64 *W_X) { Block_xx_6(12) ; }
wolfSSL	0:d92f9d21154c	873	static inline void Block_12_7(word64 *W_X) { Block_xx_7(14) ; }
wolfSSL	0:d92f9d21154c	874	static inline void Block_12_8(word64 *W_X) { Block_xx_8(14) ; }
wolfSSL	0:d92f9d21154c	875	static inline void Block_12_9(void) { Block_xx_9(14) ; }
wolfSSL	0:d92f9d21154c	876	static inline void Block_12_10(void){ Block_xx_10(14) ; }
wolfSSL	0:d92f9d21154c	877	static inline void Block_12_11(void){ Block_xx_11(14) ; }
wolfSSL	0:d92f9d21154c	878	static inline void Block_12_12(word64 *W_X){ Block_xx_12(14) ; }
wolfSSL	0:d92f9d21154c	879
wolfSSL	0:d92f9d21154c	880	#endif
wolfSSL	0:d92f9d21154c	881
wolfSSL	0:d92f9d21154c	882	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	883	static const unsigned long mBYTE_FLIP_MASK_Y[] =
wolfSSL	0:d92f9d21154c	884	{ 0x0001020304050607, 0x08090a0b0c0d0e0f, 0x0001020304050607, 0x08090a0b0c0d0e0f } ;
wolfSSL	0:d92f9d21154c	885
wolfSSL	0:d92f9d21154c	886	#define W_from_buff_Y(buff)\
wolfSSL	0:d92f9d21154c	887	{ /* X0..3(ymm9..12), W_X[0..15] = sha512->buffer[0.15]; */\
wolfSSL	0:d92f9d21154c	888	__asm__ volatile("vmovdqu %0, %%ymm8\n\t"::"m"(mBYTE_FLIP_MASK_Y[0]):YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	889	__asm__ volatile("vmovdqu %0, %%ymm12\n\t"\
wolfSSL	0:d92f9d21154c	890	"vmovdqu %1, %%ymm4\n\t"\
wolfSSL	0:d92f9d21154c	891	"vpshufb %%ymm8, %%ymm12, %%ymm12\n\t"\
wolfSSL	0:d92f9d21154c	892	"vpshufb %%ymm8, %%ymm4, %%ymm4\n\t"\
wolfSSL	0:d92f9d21154c	893	:: "m"(buff[0]), "m"(buff[4]):YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	894	__asm__ volatile("vmovdqu %0, %%ymm5\n\t"\
wolfSSL	0:d92f9d21154c	895	"vmovdqu %1, %%ymm6\n\t"\
wolfSSL	0:d92f9d21154c	896	"vpshufb %%ymm8, %%ymm5, %%ymm5\n\t"\
wolfSSL	0:d92f9d21154c	897	"vpshufb %%ymm8, %%ymm6, %%ymm6\n\t"\
wolfSSL	0:d92f9d21154c	898	:: "m"(buff[8]), "m"(buff[12]):YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	899	}
wolfSSL	0:d92f9d21154c	900
wolfSSL	0:d92f9d21154c	901	#if defined(DEBUG_YMM)
wolfSSL	0:d92f9d21154c	902
wolfSSL	0:d92f9d21154c	903	#define SAVE_REG_Y(i) __asm__ volatile("vmovdqu %%ymm"#i", %0 \n\t":"=m"(reg[i-4][0])::YMM_REGs);
wolfSSL	0:d92f9d21154c	904	#define RECV_REG_Y(i) __asm__ volatile("vmovdqu %0, %%ymm"#i" \n\t"::"m"(reg[i-4][0]):YMM_REGs);
wolfSSL	0:d92f9d21154c	905
wolfSSL	0:d92f9d21154c	906	#define _DUMP_REG_Y(REG, name)\
wolfSSL	0:d92f9d21154c	907	{ word64 buf[16] ;word64 reg[16][2];int k ;\
wolfSSL	0:d92f9d21154c	908	SAVE_REG_Y(4); SAVE_REG_Y(5); SAVE_REG_Y(6); SAVE_REG_Y(7); \
wolfSSL	0:d92f9d21154c	909	SAVE_REG_Y(8); SAVE_REG_Y(9); SAVE_REG_Y(10); SAVE_REG_Y(11); SAVE_REG_Y(12);\
wolfSSL	0:d92f9d21154c	910	SAVE_REG_Y(13); SAVE_REG_Y(14); SAVE_REG_Y(15); \
wolfSSL	0:d92f9d21154c	911	__asm__ volatile("vmovdqu %%"#REG", %0 \n\t":"=m"(buf[0])::YMM_REGs);\
wolfSSL	0:d92f9d21154c	912	printf(" "#name":\t") ; for(k=0; k<4; k++) printf("%016lx.", (word64)buf[k]) ; printf("\n") ; \
wolfSSL	0:d92f9d21154c	913	RECV_REG_Y(4); RECV_REG_Y(5); RECV_REG_Y(6); RECV_REG_Y(7); \
wolfSSL	0:d92f9d21154c	914	RECV_REG_Y(8); RECV_REG_Y(9); RECV_REG_Y(10); RECV_REG_Y(11); RECV_REG_Y(12); \
wolfSSL	0:d92f9d21154c	915	RECV_REG_Y(13); RECV_REG_Y(14); RECV_REG_Y(15);\
wolfSSL	0:d92f9d21154c	916	}
wolfSSL	0:d92f9d21154c	917
wolfSSL	0:d92f9d21154c	918	#define DUMP_REG_Y(REG) _DUMP_REG_Y(REG, #REG)
wolfSSL	0:d92f9d21154c	919	#define DUMP_REG2_Y(REG) _DUMP_REG_Y(REG, #REG)
wolfSSL	0:d92f9d21154c	920	#define PRINTF_Y(fmt, ...)
wolfSSL	0:d92f9d21154c	921
wolfSSL	0:d92f9d21154c	922	#else
wolfSSL	0:d92f9d21154c	923
wolfSSL	0:d92f9d21154c	924	#define DUMP_REG_Y(REG)
wolfSSL	0:d92f9d21154c	925	#define DUMP_REG2_Y(REG)
wolfSSL	0:d92f9d21154c	926	#define PRINTF_Y(fmt, ...)
wolfSSL	0:d92f9d21154c	927
wolfSSL	0:d92f9d21154c	928	#endif
wolfSSL	0:d92f9d21154c	929
wolfSSL	0:d92f9d21154c	930	#define _MOVE_to_REGy(ymm, mem) __asm__ volatile("vmovdqu %0, %%"#ymm" "\
wolfSSL	0:d92f9d21154c	931	:: "m"(mem):YMM_REGs) ;
wolfSSL	0:d92f9d21154c	932	#define _MOVE_to_MEMy(mem,i, ymm) __asm__ volatile("vmovdqu %%"#ymm", %0" \
wolfSSL	0:d92f9d21154c	933	: "=m"(mem[i]),"=m"(mem[i+1]),"=m"(mem[i+2]),"=m"(mem[i+3])::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	934	#define _MOVE_128y(ymm0, ymm1, ymm2, map) __asm__ volatile("vperm2i128 $"\
wolfSSL	0:d92f9d21154c	935	#map", %%"#ymm2", %%"#ymm1", %%"#ymm0" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	936	#define _S_TEMPy(dest, src, bits, temp) \
wolfSSL	0:d92f9d21154c	937	__asm__ volatile("vpsrlq $"#bits", %%"#src", %%"#dest"\n\tvpsllq $64-"#bits\
wolfSSL	0:d92f9d21154c	938	", %%"#src", %%"#temp"\n\tvpor %%"#temp",%%"#dest", %%"#dest" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	939	#define _AVX2_R(dest, src, bits) __asm__ volatile("vpsrlq $"#bits", %%"\
wolfSSL	0:d92f9d21154c	940	#src", %%"#dest" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	941	#define _XORy(dest, src1, src2) __asm__ volatile("vpxor %%"#src1", %%"\
wolfSSL	0:d92f9d21154c	942	#src2", %%"#dest" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	943	#define _ADDy(dest, src1, src2) __asm__ volatile("vpaddq %%"#src1", %%"\
wolfSSL	0:d92f9d21154c	944	#src2", %%"#dest" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	945	#define _BLENDy(map, dest, src1, src2) __asm__ volatile("vpblendd $"#map", %%"\
wolfSSL	0:d92f9d21154c	946	#src1", %%"#src2", %%"#dest" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	947	#define _BLENDQy(map, dest, src1, src2) __asm__ volatile("vblendpd $"#map", %%"\
wolfSSL	0:d92f9d21154c	948	#src1", %%"#src2", %%"#dest" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	949	#define _PERMQy(map, dest, src) __asm__ volatile("vpermq $"#map", %%"\
wolfSSL	0:d92f9d21154c	950	#src", %%"#dest" ":::YMM_REGs) ;
wolfSSL	0:d92f9d21154c	951
wolfSSL	0:d92f9d21154c	952	#define MOVE_to_REGy(ymm, mem) _MOVE_to_REGy(ymm, mem)
wolfSSL	0:d92f9d21154c	953	#define MOVE_to_MEMy(mem, i, ymm) _MOVE_to_MEMy(mem, i, ymm)
wolfSSL	0:d92f9d21154c	954
wolfSSL	0:d92f9d21154c	955	#define MOVE_128y(ymm0, ymm1, ymm2, map) _MOVE_128y(ymm0, ymm1, ymm2, map)
wolfSSL	0:d92f9d21154c	956	#define XORy(dest, src1, src2) _XORy(dest, src1, src2)
wolfSSL	0:d92f9d21154c	957	#define ADDy(dest, src1, src2) _ADDy(dest, src1, src2)
wolfSSL	0:d92f9d21154c	958	#define BLENDy(map, dest, src1, src2) _BLENDy(map, dest, src1, src2)
wolfSSL	0:d92f9d21154c	959	#define BLENDQy(map, dest, src1, src2) _BLENDQy(map, dest, src1, src2)
wolfSSL	0:d92f9d21154c	960	#define PERMQy(map, dest, src) _PERMQy(map, dest, src)
wolfSSL	0:d92f9d21154c	961
wolfSSL	0:d92f9d21154c	962
wolfSSL	0:d92f9d21154c	963	#define S_TMPy(dest, src, bits, temp) _S_TEMPy(dest, src, bits, temp);
wolfSSL	0:d92f9d21154c	964	#define AVX2_S(dest, src, bits) S_TMPy(dest, src, bits, S_TEMPy)
wolfSSL	0:d92f9d21154c	965	#define AVX2_R(dest, src, bits) _AVX2_R(dest, src, bits)
wolfSSL	0:d92f9d21154c	966
wolfSSL	0:d92f9d21154c	967
wolfSSL	0:d92f9d21154c	968	#define FEEDBACK1_to_W_I_2(w_i_2, w_i) MOVE_128y(YMM_TEMP0, w_i, w_i, 0x08) ;\
wolfSSL	0:d92f9d21154c	969	BLENDy(0xf0, w_i_2, YMM_TEMP0, w_i_2) ;
wolfSSL	0:d92f9d21154c	970
wolfSSL	0:d92f9d21154c	971	#define MOVE_W_to_W_I_15(w_i_15, w_0, w_4) BLENDQy(0x1, w_i_15, w_4, w_0) ;\
wolfSSL	0:d92f9d21154c	972	PERMQy(0x39, w_i_15, w_i_15) ;
wolfSSL	0:d92f9d21154c	973	#define MOVE_W_to_W_I_7(w_i_7, w_8, w_12) BLENDQy(0x1, w_i_7, w_12, w_8) ;\
wolfSSL	0:d92f9d21154c	974	PERMQy(0x39, w_i_7, w_i_7) ;
wolfSSL	0:d92f9d21154c	975	#define MOVE_W_to_W_I_2(w_i_2, w_12) BLENDQy(0xc, w_i_2, w_12, w_i_2) ;\
wolfSSL	0:d92f9d21154c	976	PERMQy(0x0e, w_i_2, w_i_2) ;
wolfSSL	0:d92f9d21154c	977
wolfSSL	0:d92f9d21154c	978
wolfSSL	0:d92f9d21154c	979	#define W_I_16y ymm8
wolfSSL	0:d92f9d21154c	980	#define W_I_15y ymm9
wolfSSL	0:d92f9d21154c	981	#define W_I_7y ymm10
wolfSSL	0:d92f9d21154c	982	#define W_I_2y ymm11
wolfSSL	0:d92f9d21154c	983	#define W_Iy ymm12
wolfSSL	0:d92f9d21154c	984	#define G_TEMPy ymm13
wolfSSL	0:d92f9d21154c	985	#define S_TEMPy ymm14
wolfSSL	0:d92f9d21154c	986	#define YMM_TEMP0 ymm15
wolfSSL	0:d92f9d21154c	987	#define YMM_TEMP0x xmm15
wolfSSL	0:d92f9d21154c	988	#define W_I_TEMPy ymm7
wolfSSL	0:d92f9d21154c	989	#define W_K_TEMPy ymm15
wolfSSL	0:d92f9d21154c	990	#define W_K_TEMPx xmm15
wolfSSL	0:d92f9d21154c	991	#define W_0y ymm12
wolfSSL	0:d92f9d21154c	992	#define W_4y ymm4
wolfSSL	0:d92f9d21154c	993	#define W_8y ymm5
wolfSSL	0:d92f9d21154c	994	#define W_12y ymm6
wolfSSL	0:d92f9d21154c	995
wolfSSL	0:d92f9d21154c	996	#define YMM_REGs
wolfSSL	0:d92f9d21154c	997	/* Registers are saved in Sha512Update/Final */
wolfSSL	0:d92f9d21154c	998	/* "%ymm7","%ymm8","%ymm9","%ymm10","%ymm11","%ymm12","%ymm13","%ymm14","%ymm15"*/
wolfSSL	0:d92f9d21154c	999
wolfSSL	0:d92f9d21154c	1000	#define MOVE_15_to_16(w_i_16, w_i_15, w_i_7)\
wolfSSL	0:d92f9d21154c	1001	__asm__ volatile("vperm2i128 $0x01, %%"#w_i_15", %%"#w_i_15", %%"#w_i_15" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1002	__asm__ volatile("vpblendd $0x08, %%"#w_i_15", %%"#w_i_7", %%"#w_i_16" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1003	__asm__ volatile("vperm2i128 $0x01, %%"#w_i_7", %%"#w_i_7", %%"#w_i_15" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1004	__asm__ volatile("vpblendd $0x80, %%"#w_i_15", %%"#w_i_16", %%"#w_i_16" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1005	__asm__ volatile("vpshufd $0x93, %%"#w_i_16", %%"#w_i_16" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1006
wolfSSL	0:d92f9d21154c	1007	#define MOVE_7_to_15(w_i_15, w_i_7)\
wolfSSL	0:d92f9d21154c	1008	__asm__ volatile("vmovdqu %%"#w_i_7", %%"#w_i_15" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1009
wolfSSL	0:d92f9d21154c	1010	#define MOVE_I_to_7(w_i_7, w_i)\
wolfSSL	0:d92f9d21154c	1011	__asm__ volatile("vperm2i128 $0x01, %%"#w_i", %%"#w_i", %%"#w_i_7" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1012	__asm__ volatile("vpblendd $0x01, %%"#w_i_7", %%"#w_i", %%"#w_i_7" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1013	__asm__ volatile("vpshufd $0x39, %%"#w_i_7", %%"#w_i_7" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1014
wolfSSL	0:d92f9d21154c	1015	#define MOVE_I_to_2(w_i_2, w_i)\
wolfSSL	0:d92f9d21154c	1016	__asm__ volatile("vperm2i128 $0x01, %%"#w_i", %%"#w_i", %%"#w_i_2" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1017	__asm__ volatile("vpshufd $0x0e, %%"#w_i_2", %%"#w_i_2" ":::YMM_REGs) ;\
wolfSSL	0:d92f9d21154c	1018
wolfSSL	0:d92f9d21154c	1019	#endif
wolfSSL	0:d92f9d21154c	1020
wolfSSL	0:d92f9d21154c	1021
wolfSSL	0:d92f9d21154c	1022	/* Transform Body */
wolfSSL	0:d92f9d21154c	1023	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	1024
wolfSSL	0:d92f9d21154c	1025	static int Transform_AVX1(Sha512* sha512)
wolfSSL	0:d92f9d21154c	1026	{
wolfSSL	0:d92f9d21154c	1027	const word64* K = K512;
wolfSSL	0:d92f9d21154c	1028	word64 W_X[16+4];
wolfSSL	0:d92f9d21154c	1029	word32 j;
wolfSSL	0:d92f9d21154c	1030	word64 T[8];
wolfSSL	0:d92f9d21154c	1031	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	1032	XMEMCPY(T, sha512->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	1033
wolfSSL	0:d92f9d21154c	1034	W_from_buff(W_X, sha512->buffer) ;
wolfSSL	0:d92f9d21154c	1035	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1036	Rx_1( 0); Block_0_1(W_X); Rx_2( 0); Block_0_2(W_X); Rx_3( 0); Block_0_3();
wolfSSL	0:d92f9d21154c	1037	Rx_1( 1); Block_0_4(); Rx_2( 1); Block_0_5(); Rx_3( 1); Block_0_6(W_X);
wolfSSL	0:d92f9d21154c	1038	Rx_1( 2); Block_0_7(W_X); Rx_2( 2); Block_0_8(W_X); Rx_3( 2); Block_0_9();
wolfSSL	0:d92f9d21154c	1039	Rx_1( 3); Block_0_10();Rx_2( 3); Block_0_11();Rx_3( 3); Block_0_12(W_X);
wolfSSL	0:d92f9d21154c	1040
wolfSSL	0:d92f9d21154c	1041	Rx_1( 4); Block_4_1(W_X); Rx_2( 4); Block_4_2(W_X); Rx_3( 4); Block_4_3();
wolfSSL	0:d92f9d21154c	1042	Rx_1( 5); Block_4_4(); Rx_2( 5); Block_4_5(); Rx_3( 5); Block_4_6(W_X);
wolfSSL	0:d92f9d21154c	1043	Rx_1( 6); Block_4_7(W_X); Rx_2( 6); Block_4_8(W_X); Rx_3( 6); Block_4_9();
wolfSSL	0:d92f9d21154c	1044	Rx_1( 7); Block_4_10();Rx_2( 7); Block_4_11();Rx_3( 7); Block_4_12(W_X);
wolfSSL	0:d92f9d21154c	1045
wolfSSL	0:d92f9d21154c	1046	Rx_1( 8); Block_8_1(W_X); Rx_2( 8); Block_8_2(W_X); Rx_3( 8); Block_8_3();
wolfSSL	0:d92f9d21154c	1047	Rx_1( 9); Block_8_4(); Rx_2( 9); Block_8_5(); Rx_3( 9); Block_8_6(W_X);
wolfSSL	0:d92f9d21154c	1048	Rx_1(10); Block_8_7(W_X); Rx_2(10); Block_8_8(W_X); Rx_3(10); Block_8_9();
wolfSSL	0:d92f9d21154c	1049	Rx_1(11); Block_8_10();Rx_2(11); Block_8_11();Rx_3(11); Block_8_12(W_X);
wolfSSL	0:d92f9d21154c	1050
wolfSSL	0:d92f9d21154c	1051	Rx_1(12); Block_12_1(W_X); Rx_2(12); Block_12_2(W_X); Rx_3(12); Block_12_3();
wolfSSL	0:d92f9d21154c	1052	Rx_1(13); Block_12_4(); Rx_2(13); Block_12_5(); Rx_3(13); Block_12_6(W_X);
wolfSSL	0:d92f9d21154c	1053	Rx_1(14); Block_12_7(W_X); Rx_2(14); Block_12_8(W_X); Rx_3(14); Block_12_9();
wolfSSL	0:d92f9d21154c	1054	Rx_1(15); Block_12_10();Rx_2(15); Block_12_11();Rx_3(15); Block_12_12(W_X);
wolfSSL	0:d92f9d21154c	1055	}
wolfSSL	0:d92f9d21154c	1056
wolfSSL	0:d92f9d21154c	1057	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	1058
wolfSSL	0:d92f9d21154c	1059	sha512->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	1060	sha512->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	1061	sha512->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	1062	sha512->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	1063	sha512->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	1064	sha512->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	1065	sha512->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	1066	sha512->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	1067
wolfSSL	0:d92f9d21154c	1068	/* Wipe variables */
wolfSSL	0:d92f9d21154c	1069	#if !defined(HAVE_INTEL_AVX1)&&!defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1070	XMEMSET(W_X, 0, sizeof(word64) * 16);
wolfSSL	0:d92f9d21154c	1071	#endif
wolfSSL	0:d92f9d21154c	1072	XMEMSET(T, 0, sizeof(T));
wolfSSL	0:d92f9d21154c	1073
wolfSSL	0:d92f9d21154c	1074	return 0;
wolfSSL	0:d92f9d21154c	1075	}
wolfSSL	0:d92f9d21154c	1076
wolfSSL	0:d92f9d21154c	1077	#endif
wolfSSL	0:d92f9d21154c	1078
wolfSSL	0:d92f9d21154c	1079	#if defined(HAVE_INTEL_AVX2) && defined(HAVE_INTEL_AVX1) && defined(HAVE_INTEL_RORX)
wolfSSL	0:d92f9d21154c	1080
wolfSSL	0:d92f9d21154c	1081	static int Transform_AVX1_RORX(Sha512* sha512)
wolfSSL	0:d92f9d21154c	1082	{
wolfSSL	0:d92f9d21154c	1083	const word64* K = K512;
wolfSSL	0:d92f9d21154c	1084	word64 W_X[16+4];
wolfSSL	0:d92f9d21154c	1085	word32 j;
wolfSSL	0:d92f9d21154c	1086	word64 T[8];
wolfSSL	0:d92f9d21154c	1087	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	1088	XMEMCPY(T, sha512->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	1089
wolfSSL	0:d92f9d21154c	1090	W_from_buff(W_X, sha512->buffer) ;
wolfSSL	0:d92f9d21154c	1091	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1092	Rx_RORX_1( 0); Block_0_1(W_X); Rx_RORX_2( 0); Block_0_2(W_X);
wolfSSL	0:d92f9d21154c	1093	Rx_RORX_3( 0); Block_0_3();
wolfSSL	0:d92f9d21154c	1094	Rx_RORX_1( 1); Block_0_4(); Rx_RORX_2( 1); Block_0_5();
wolfSSL	0:d92f9d21154c	1095	Rx_RORX_3( 1); Block_0_6(W_X);
wolfSSL	0:d92f9d21154c	1096	Rx_RORX_1( 2); Block_0_7(W_X); Rx_RORX_2( 2); Block_0_8(W_X);
wolfSSL	0:d92f9d21154c	1097	Rx_RORX_3( 2); Block_0_9();
wolfSSL	0:d92f9d21154c	1098	Rx_RORX_1( 3); Block_0_10();Rx_RORX_2( 3); Block_0_11();
wolfSSL	0:d92f9d21154c	1099	Rx_RORX_3( 3); Block_0_12(W_X);
wolfSSL	0:d92f9d21154c	1100
wolfSSL	0:d92f9d21154c	1101	Rx_RORX_1( 4); Block_4_1(W_X); Rx_RORX_2( 4); Block_4_2(W_X);
wolfSSL	0:d92f9d21154c	1102	Rx_RORX_3( 4); Block_4_3();
wolfSSL	0:d92f9d21154c	1103	Rx_RORX_1( 5); Block_4_4(); Rx_RORX_2( 5); Block_4_5();
wolfSSL	0:d92f9d21154c	1104	Rx_RORX_3( 5); Block_4_6(W_X);
wolfSSL	0:d92f9d21154c	1105	Rx_RORX_1( 6); Block_4_7(W_X); Rx_RORX_2( 6); Block_4_8(W_X);
wolfSSL	0:d92f9d21154c	1106	Rx_RORX_3( 6); Block_4_9();
wolfSSL	0:d92f9d21154c	1107	Rx_RORX_1( 7); Block_4_10();Rx_RORX_2( 7); Block_4_11();
wolfSSL	0:d92f9d21154c	1108	Rx_RORX_3( 7); Block_4_12(W_X);
wolfSSL	0:d92f9d21154c	1109
wolfSSL	0:d92f9d21154c	1110	Rx_RORX_1( 8); Block_8_1(W_X); Rx_RORX_2( 8); Block_8_2(W_X);
wolfSSL	0:d92f9d21154c	1111	Rx_RORX_3( 8); Block_8_3();
wolfSSL	0:d92f9d21154c	1112	Rx_RORX_1( 9); Block_8_4(); Rx_RORX_2( 9); Block_8_5();
wolfSSL	0:d92f9d21154c	1113	Rx_RORX_3( 9); Block_8_6(W_X);
wolfSSL	0:d92f9d21154c	1114	Rx_RORX_1(10); Block_8_7(W_X); Rx_RORX_2(10); Block_8_8(W_X);
wolfSSL	0:d92f9d21154c	1115	Rx_RORX_3(10); Block_8_9();
wolfSSL	0:d92f9d21154c	1116	Rx_RORX_1(11); Block_8_10();Rx_RORX_2(11); Block_8_11();
wolfSSL	0:d92f9d21154c	1117	Rx_RORX_3(11); Block_8_12(W_X);
wolfSSL	0:d92f9d21154c	1118
wolfSSL	0:d92f9d21154c	1119	Rx_RORX_1(12); Block_12_1(W_X); Rx_RORX_2(12); Block_12_2(W_X);
wolfSSL	0:d92f9d21154c	1120	Rx_RORX_3(12); Block_12_3();
wolfSSL	0:d92f9d21154c	1121	Rx_RORX_1(13); Block_12_4(); Rx_RORX_2(13); Block_12_5();
wolfSSL	0:d92f9d21154c	1122	Rx_RORX_3(13); Block_12_6(W_X);
wolfSSL	0:d92f9d21154c	1123	Rx_RORX_1(14); Block_12_7(W_X); Rx_RORX_2(14); Block_12_8(W_X);
wolfSSL	0:d92f9d21154c	1124	Rx_RORX_3(14); Block_12_9();
wolfSSL	0:d92f9d21154c	1125	Rx_RORX_1(15); Block_12_10();Rx_RORX_2(15); Block_12_11();
wolfSSL	0:d92f9d21154c	1126	Rx_RORX_3(15); Block_12_12(W_X);
wolfSSL	0:d92f9d21154c	1127	}
wolfSSL	0:d92f9d21154c	1128	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	1129
wolfSSL	0:d92f9d21154c	1130	sha512->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	1131	sha512->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	1132	sha512->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	1133	sha512->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	1134	sha512->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	1135	sha512->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	1136	sha512->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	1137	sha512->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	1138
wolfSSL	0:d92f9d21154c	1139	/* Wipe variables */
wolfSSL	0:d92f9d21154c	1140	#if !defined(HAVE_INTEL_AVX1)&&!defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1141	XMEMSET(W_X, 0, sizeof(word64) * 16);
wolfSSL	0:d92f9d21154c	1142	#endif
wolfSSL	0:d92f9d21154c	1143	XMEMSET(T, 0, sizeof(T));
wolfSSL	0:d92f9d21154c	1144
wolfSSL	0:d92f9d21154c	1145	return 0;
wolfSSL	0:d92f9d21154c	1146	}
wolfSSL	0:d92f9d21154c	1147	#endif
wolfSSL	0:d92f9d21154c	1148
wolfSSL	0:d92f9d21154c	1149	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1150
wolfSSL	0:d92f9d21154c	1151	#define s0_1y(dest, src) AVX2_S(dest, src, 1);
wolfSSL	0:d92f9d21154c	1152	#define s0_2y(dest, src) AVX2_S(G_TEMPy, src, 8); XORy(dest, G_TEMPy, dest) ;
wolfSSL	0:d92f9d21154c	1153	#define s0_3y(dest, src) AVX2_R(G_TEMPy, src, 7); XORy(dest, G_TEMPy, dest) ;
wolfSSL	0:d92f9d21154c	1154
wolfSSL	0:d92f9d21154c	1155	#define s1_1y(dest, src) AVX2_S(dest, src, 19);
wolfSSL	0:d92f9d21154c	1156	#define s1_2y(dest, src) AVX2_S(G_TEMPy, src, 61); XORy(dest, G_TEMPy, dest) ;
wolfSSL	0:d92f9d21154c	1157	#define s1_3y(dest, src) AVX2_R(G_TEMPy, src, 6); XORy(dest, G_TEMPy, dest) ;
wolfSSL	0:d92f9d21154c	1158
wolfSSL	0:d92f9d21154c	1159	#define s0_y(dest, src) s0_1y(dest, src) ; s0_2y(dest, src) ; s0_3y(dest, src)
wolfSSL	0:d92f9d21154c	1160	#define s1_y(dest, src) s1_1y(dest, src) ; s1_2y(dest, src) ; s1_3y(dest, src)
wolfSSL	0:d92f9d21154c	1161
wolfSSL	0:d92f9d21154c	1162	#define blk384(i) (W[i] = sha384->buffer[i])
wolfSSL	0:d92f9d21154c	1163
wolfSSL	0:d92f9d21154c	1164
wolfSSL	0:d92f9d21154c	1165	#define Block_Y_xx_1(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1166	MOVE_W_to_W_I_15(W_I_15y, w_0, w_4) ;\
wolfSSL	0:d92f9d21154c	1167	MOVE_W_to_W_I_7 (W_I_7y, w_8, w_12) ;\
wolfSSL	0:d92f9d21154c	1168	MOVE_W_to_W_I_2 (W_I_2y, w_12) ;\
wolfSSL	0:d92f9d21154c	1169
wolfSSL	0:d92f9d21154c	1170	#define Block_Y_xx_2(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1171	s0_1y (YMM_TEMP0, W_I_15y) ;\
wolfSSL	0:d92f9d21154c	1172
wolfSSL	0:d92f9d21154c	1173	#define Block_Y_xx_3(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1174	s0_2y (YMM_TEMP0, W_I_15y) ;\
wolfSSL	0:d92f9d21154c	1175
wolfSSL	0:d92f9d21154c	1176	#define Block_Y_xx_4(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1177	s0_3y (YMM_TEMP0, W_I_15y) ;\
wolfSSL	0:d92f9d21154c	1178
wolfSSL	0:d92f9d21154c	1179	#define Block_Y_xx_5(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1180	ADDy(W_I_TEMPy, w_0, YMM_TEMP0) ;\
wolfSSL	0:d92f9d21154c	1181
wolfSSL	0:d92f9d21154c	1182	#define Block_Y_xx_6(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1183	ADDy(W_I_TEMPy, W_I_TEMPy, W_I_7y) ;\
wolfSSL	0:d92f9d21154c	1184	s1_1y (YMM_TEMP0, W_I_2y) ;\
wolfSSL	0:d92f9d21154c	1185
wolfSSL	0:d92f9d21154c	1186	#define Block_Y_xx_7(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1187	s1_2y (YMM_TEMP0, W_I_2y) ;\
wolfSSL	0:d92f9d21154c	1188
wolfSSL	0:d92f9d21154c	1189	#define Block_Y_xx_8(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1190	s1_3y (YMM_TEMP0, W_I_2y) ;\
wolfSSL	0:d92f9d21154c	1191	ADDy(w_0, W_I_TEMPy, YMM_TEMP0) ;\
wolfSSL	0:d92f9d21154c	1192
wolfSSL	0:d92f9d21154c	1193	#define Block_Y_xx_9(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1194	FEEDBACK1_to_W_I_2(W_I_2y, w_0) ;\
wolfSSL	0:d92f9d21154c	1195
wolfSSL	0:d92f9d21154c	1196	#define Block_Y_xx_10(i, w_0, w_4, w_8, w_12) \
wolfSSL	0:d92f9d21154c	1197	s1_1y (YMM_TEMP0, W_I_2y) ;\
wolfSSL	0:d92f9d21154c	1198
wolfSSL	0:d92f9d21154c	1199	#define Block_Y_xx_11(i, w_0, w_4, w_8, w_12) \
wolfSSL	0:d92f9d21154c	1200	s1_2y (YMM_TEMP0, W_I_2y) ;\
wolfSSL	0:d92f9d21154c	1201
wolfSSL	0:d92f9d21154c	1202	#define Block_Y_xx_12(i, w_0, w_4, w_8, w_12)\
wolfSSL	0:d92f9d21154c	1203	s1_3y (YMM_TEMP0, W_I_2y) ;\
wolfSSL	0:d92f9d21154c	1204	ADDy(w_0, W_I_TEMPy, YMM_TEMP0) ;\
wolfSSL	0:d92f9d21154c	1205	MOVE_to_MEMy(w,0, w_4) ;\
wolfSSL	0:d92f9d21154c	1206
wolfSSL	0:d92f9d21154c	1207
wolfSSL	0:d92f9d21154c	1208	static inline void Block_Y_0_1(void) { Block_Y_xx_1(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1209	static inline void Block_Y_0_2(void) { Block_Y_xx_2(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1210	static inline void Block_Y_0_3(void) { Block_Y_xx_3(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1211	static inline void Block_Y_0_4(void) { Block_Y_xx_4(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1212	static inline void Block_Y_0_5(void) { Block_Y_xx_5(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1213	static inline void Block_Y_0_6(void) { Block_Y_xx_6(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1214	static inline void Block_Y_0_7(void) { Block_Y_xx_7(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1215	static inline void Block_Y_0_8(void) { Block_Y_xx_8(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1216	static inline void Block_Y_0_9(void) { Block_Y_xx_9(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1217	static inline void Block_Y_0_10(void){ Block_Y_xx_10(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1218	static inline void Block_Y_0_11(void){ Block_Y_xx_11(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1219	static inline void Block_Y_0_12(word64 *w){ Block_Y_xx_12(0, W_0y, W_4y, W_8y, W_12y) ; }
wolfSSL	0:d92f9d21154c	1220
wolfSSL	0:d92f9d21154c	1221	static inline void Block_Y_4_1(void) { Block_Y_xx_1(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1222	static inline void Block_Y_4_2(void) { Block_Y_xx_2(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1223	static inline void Block_Y_4_3(void) { Block_Y_xx_3(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1224	static inline void Block_Y_4_4(void) { Block_Y_xx_4(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1225	static inline void Block_Y_4_5(void) { Block_Y_xx_5(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1226	static inline void Block_Y_4_6(void) { Block_Y_xx_6(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1227	static inline void Block_Y_4_7(void) { Block_Y_xx_7(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1228	static inline void Block_Y_4_8(void) { Block_Y_xx_8(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1229	static inline void Block_Y_4_9(void) { Block_Y_xx_9(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1230	static inline void Block_Y_4_10(void) { Block_Y_xx_10(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1231	static inline void Block_Y_4_11(void) { Block_Y_xx_11(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1232	static inline void Block_Y_4_12(word64 *w) { Block_Y_xx_12(4, W_4y, W_8y, W_12y, W_0y) ; }
wolfSSL	0:d92f9d21154c	1233
wolfSSL	0:d92f9d21154c	1234	static inline void Block_Y_8_1(void) { Block_Y_xx_1(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1235	static inline void Block_Y_8_2(void) { Block_Y_xx_2(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1236	static inline void Block_Y_8_3(void) { Block_Y_xx_3(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1237	static inline void Block_Y_8_4(void) { Block_Y_xx_4(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1238	static inline void Block_Y_8_5(void) { Block_Y_xx_5(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1239	static inline void Block_Y_8_6(void) { Block_Y_xx_6(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1240	static inline void Block_Y_8_7(void) { Block_Y_xx_7(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1241	static inline void Block_Y_8_8(void) { Block_Y_xx_8(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1242	static inline void Block_Y_8_9(void) { Block_Y_xx_9(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1243	static inline void Block_Y_8_10(void) { Block_Y_xx_10(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1244	static inline void Block_Y_8_11(void) { Block_Y_xx_11(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1245	static inline void Block_Y_8_12(word64 *w) { Block_Y_xx_12(8, W_8y, W_12y, W_0y, W_4y) ; }
wolfSSL	0:d92f9d21154c	1246
wolfSSL	0:d92f9d21154c	1247	static inline void Block_Y_12_1(void) { Block_Y_xx_1(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1248	static inline void Block_Y_12_2(void) { Block_Y_xx_2(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1249	static inline void Block_Y_12_3(void) { Block_Y_xx_3(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1250	static inline void Block_Y_12_4(void) { Block_Y_xx_4(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1251	static inline void Block_Y_12_5(void) { Block_Y_xx_5(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1252	static inline void Block_Y_12_6(void) { Block_Y_xx_6(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1253	static inline void Block_Y_12_7(void) { Block_Y_xx_7(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1254	static inline void Block_Y_12_8(void) { Block_Y_xx_8(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1255	static inline void Block_Y_12_9(void) { Block_Y_xx_9(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1256	static inline void Block_Y_12_10(void) { Block_Y_xx_10(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1257	static inline void Block_Y_12_11(void) { Block_Y_xx_11(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1258	static inline void Block_Y_12_12(word64 *w) { Block_Y_xx_12(12, W_12y, W_0y, W_4y, W_8y) ; }
wolfSSL	0:d92f9d21154c	1259
wolfSSL	0:d92f9d21154c	1260
wolfSSL	0:d92f9d21154c	1261	static int Transform_AVX2(Sha512* sha512)
wolfSSL	0:d92f9d21154c	1262	{
wolfSSL	0:d92f9d21154c	1263	const word64* K = K512;
wolfSSL	0:d92f9d21154c	1264	word64 w[4] ;
wolfSSL	0:d92f9d21154c	1265	word32 j /, k/;
wolfSSL	0:d92f9d21154c	1266	word64 T[8];
wolfSSL	0:d92f9d21154c	1267	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	1268	XMEMCPY(T, sha512->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	1269
wolfSSL	0:d92f9d21154c	1270	W_from_buff_Y(sha512->buffer) ;
wolfSSL	0:d92f9d21154c	1271	MOVE_to_MEMy(w,0, W_0y) ;
wolfSSL	0:d92f9d21154c	1272	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1273	Ry_1( 0, w[0]); Block_Y_0_1(); Ry_2( 0, w[0]); Block_Y_0_2();
wolfSSL	0:d92f9d21154c	1274	Ry_3( 0, w[0]); Block_Y_0_3();
wolfSSL	0:d92f9d21154c	1275	Ry_1( 1, w[1]); Block_Y_0_4(); Ry_2( 1, w[1]); Block_Y_0_5();
wolfSSL	0:d92f9d21154c	1276	Ry_3( 1, w[1]); Block_Y_0_6();
wolfSSL	0:d92f9d21154c	1277	Ry_1( 2, w[2]); Block_Y_0_7(); Ry_2( 2, w[2]); Block_Y_0_8();
wolfSSL	0:d92f9d21154c	1278	Ry_3( 2, w[2]); Block_Y_0_9();
wolfSSL	0:d92f9d21154c	1279	Ry_1( 3, w[3]); Block_Y_0_10();Ry_2( 3, w[3]); Block_Y_0_11();
wolfSSL	0:d92f9d21154c	1280	Ry_3( 3, w[3]); Block_Y_0_12(w);
wolfSSL	0:d92f9d21154c	1281
wolfSSL	0:d92f9d21154c	1282	Ry_1( 4, w[0]); Block_Y_4_1(); Ry_2( 4, w[0]); Block_Y_4_2();
wolfSSL	0:d92f9d21154c	1283	Ry_3( 4, w[0]); Block_Y_4_3();
wolfSSL	0:d92f9d21154c	1284	Ry_1( 5, w[1]); Block_Y_4_4(); Ry_2( 5, w[1]); Block_Y_4_5();
wolfSSL	0:d92f9d21154c	1285	Ry_3( 5, w[1]); Block_Y_4_6();
wolfSSL	0:d92f9d21154c	1286	Ry_1( 6, w[2]); Block_Y_4_7(); Ry_2( 6, w[2]); Block_Y_4_8();
wolfSSL	0:d92f9d21154c	1287	Ry_3( 6, w[2]); Block_Y_4_9();
wolfSSL	0:d92f9d21154c	1288	Ry_1( 7, w[3]); Block_Y_4_10(); Ry_2( 7, w[3]);Block_Y_4_11();
wolfSSL	0:d92f9d21154c	1289	Ry_3( 7, w[3]);Block_Y_4_12(w);
wolfSSL	0:d92f9d21154c	1290
wolfSSL	0:d92f9d21154c	1291	Ry_1( 8, w[0]); Block_Y_8_1(); Ry_2( 8, w[0]); Block_Y_8_2();
wolfSSL	0:d92f9d21154c	1292	Ry_3( 8, w[0]); Block_Y_8_3();
wolfSSL	0:d92f9d21154c	1293	Ry_1( 9, w[1]); Block_Y_8_4(); Ry_2( 9, w[1]); Block_Y_8_5();
wolfSSL	0:d92f9d21154c	1294	Ry_3( 9, w[1]); Block_Y_8_6();
wolfSSL	0:d92f9d21154c	1295	Ry_1(10, w[2]); Block_Y_8_7(); Ry_2(10, w[2]); Block_Y_8_8();
wolfSSL	0:d92f9d21154c	1296	Ry_3(10, w[2]); Block_Y_8_9();
wolfSSL	0:d92f9d21154c	1297	Ry_1(11, w[3]); Block_Y_8_10();Ry_2(11, w[3]); Block_Y_8_11();
wolfSSL	0:d92f9d21154c	1298	Ry_3(11, w[3]); Block_Y_8_12(w);
wolfSSL	0:d92f9d21154c	1299
wolfSSL	0:d92f9d21154c	1300	Ry_1(12, w[0]); Block_Y_12_1(); Ry_2(12, w[0]); Block_Y_12_2();
wolfSSL	0:d92f9d21154c	1301	Ry_3(12, w[0]); Block_Y_12_3();
wolfSSL	0:d92f9d21154c	1302	Ry_1(13, w[1]); Block_Y_12_4(); Ry_2(13, w[1]); Block_Y_12_5();
wolfSSL	0:d92f9d21154c	1303	Ry_3(13, w[1]); Block_Y_12_6();
wolfSSL	0:d92f9d21154c	1304	Ry_1(14, w[2]); Block_Y_12_7(); Ry_2(14, w[2]); Block_Y_12_8();
wolfSSL	0:d92f9d21154c	1305	Ry_3(14, w[2]); Block_Y_12_9();
wolfSSL	0:d92f9d21154c	1306	Ry_1(15, w[3]); Block_Y_12_10();Ry_2(15, w[3]); Block_Y_12_11();
wolfSSL	0:d92f9d21154c	1307	Ry_3(15, w[3]);Block_Y_12_12(w);
wolfSSL	0:d92f9d21154c	1308	}
wolfSSL	0:d92f9d21154c	1309
wolfSSL	0:d92f9d21154c	1310	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	1311
wolfSSL	0:d92f9d21154c	1312	sha512->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	1313	sha512->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	1314	sha512->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	1315	sha512->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	1316	sha512->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	1317	sha512->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	1318	sha512->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	1319	sha512->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	1320
wolfSSL	0:d92f9d21154c	1321	/* Wipe variables */
wolfSSL	0:d92f9d21154c	1322	#if !defined(HAVE_INTEL_AVX1)&&!defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1323	XMEMSET(W, 0, sizeof(word64) * 16);
wolfSSL	0:d92f9d21154c	1324	#endif
wolfSSL	0:d92f9d21154c	1325	XMEMSET(T, 0, sizeof(T));
wolfSSL	0:d92f9d21154c	1326
wolfSSL	0:d92f9d21154c	1327	return 0;
wolfSSL	0:d92f9d21154c	1328	}
wolfSSL	0:d92f9d21154c	1329
wolfSSL	0:d92f9d21154c	1330	#endif
wolfSSL	0:d92f9d21154c	1331
wolfSSL	0:d92f9d21154c	1332
wolfSSL	0:d92f9d21154c	1333	#ifdef WOLFSSL_SHA384
wolfSSL	0:d92f9d21154c	1334
wolfSSL	0:d92f9d21154c	1335	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1336
wolfSSL	0:d92f9d21154c	1337	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	1338	static int Transform384_AVX1(Sha384 *sha384) ;
wolfSSL	0:d92f9d21154c	1339	#endif
wolfSSL	0:d92f9d21154c	1340	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1341	static int Transform384_AVX2(Sha384 *sha384) ;
wolfSSL	0:d92f9d21154c	1342	#endif
wolfSSL	0:d92f9d21154c	1343
wolfSSL	0:d92f9d21154c	1344	#if defined(HAVE_INTEL_AVX1) && defined(HAVE_INTEL_AVX2) &&defined(HAVE_INTEL_RORX)
wolfSSL	0:d92f9d21154c	1345	static int Transform384_AVX1_RORX(Sha384 *sha384) ;
wolfSSL	0:d92f9d21154c	1346	#endif
wolfSSL	0:d92f9d21154c	1347
wolfSSL	0:d92f9d21154c	1348	static int _Transform384(Sha384 *sha384) ;
wolfSSL	0:d92f9d21154c	1349	static int (Transform384_p)(Sha384 sha384) = _Transform384 ;
wolfSSL	0:d92f9d21154c	1350
wolfSSL	0:d92f9d21154c	1351	#define Transform384(sha384) (*Transform384_p)(sha384)
wolfSSL	0:d92f9d21154c	1352	static void set_Transform384(void) {
wolfSSL	0:d92f9d21154c	1353	if(set_cpuid_flags(CHECK_SHA384))return ;
wolfSSL	0:d92f9d21154c	1354
wolfSSL	0:d92f9d21154c	1355	#if defined(HAVE_INTEL_AVX1) && !defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1356	Transform384_p = ((IS_INTEL_AVX1) ? Transform384_AVX1 : _Transform384) ;
wolfSSL	0:d92f9d21154c	1357	#elif defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1358	#if defined(HAVE_INTEL_AVX1) && defined(HAVE_INTEL_RORX)
wolfSSL	0:d92f9d21154c	1359	if(IS_INTEL_AVX2 && IS_INTEL_BMI2) { Transform384_p = Transform384_AVX1_RORX ; return ; }
wolfSSL	0:d92f9d21154c	1360	#endif
wolfSSL	0:d92f9d21154c	1361	if(IS_INTEL_AVX2) { Transform384_p = Transform384_AVX2 ; return ; }
wolfSSL	0:d92f9d21154c	1362	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	1363	Transform384_p = ((IS_INTEL_AVX1) ? Transform384_AVX1 : _Transform384) ;
wolfSSL	0:d92f9d21154c	1364	#endif
wolfSSL	0:d92f9d21154c	1365	#else
wolfSSL	0:d92f9d21154c	1366	Transform384_p = ((IS_INTEL_AVX1) ? Transform384_AVX1 : _Transform384) ;
wolfSSL	0:d92f9d21154c	1367	#endif
wolfSSL	0:d92f9d21154c	1368	}
wolfSSL	0:d92f9d21154c	1369
wolfSSL	0:d92f9d21154c	1370	#else
wolfSSL	0:d92f9d21154c	1371	#define Transform384(sha512) _Transform384(sha512)
wolfSSL	0:d92f9d21154c	1372	#endif
wolfSSL	0:d92f9d21154c	1373
wolfSSL	0:d92f9d21154c	1374	int wc_InitSha384(Sha384* sha384)
wolfSSL	0:d92f9d21154c	1375	{
wolfSSL	0:d92f9d21154c	1376	sha384->digest[0] = W64LIT(0xcbbb9d5dc1059ed8);
wolfSSL	0:d92f9d21154c	1377	sha384->digest[1] = W64LIT(0x629a292a367cd507);
wolfSSL	0:d92f9d21154c	1378	sha384->digest[2] = W64LIT(0x9159015a3070dd17);
wolfSSL	0:d92f9d21154c	1379	sha384->digest[3] = W64LIT(0x152fecd8f70e5939);
wolfSSL	0:d92f9d21154c	1380	sha384->digest[4] = W64LIT(0x67332667ffc00b31);
wolfSSL	0:d92f9d21154c	1381	sha384->digest[5] = W64LIT(0x8eb44a8768581511);
wolfSSL	0:d92f9d21154c	1382	sha384->digest[6] = W64LIT(0xdb0c2e0d64f98fa7);
wolfSSL	0:d92f9d21154c	1383	sha384->digest[7] = W64LIT(0x47b5481dbefa4fa4);
wolfSSL	0:d92f9d21154c	1384
wolfSSL	0:d92f9d21154c	1385	sha384->buffLen = 0;
wolfSSL	0:d92f9d21154c	1386	sha384->loLen = 0;
wolfSSL	0:d92f9d21154c	1387	sha384->hiLen = 0;
wolfSSL	0:d92f9d21154c	1388
wolfSSL	0:d92f9d21154c	1389	#if defined(HAVE_INTEL_AVX1)\|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1390	set_Transform384() ;
wolfSSL	0:d92f9d21154c	1391	#endif
wolfSSL	0:d92f9d21154c	1392
wolfSSL	0:d92f9d21154c	1393	return 0;
wolfSSL	0:d92f9d21154c	1394	}
wolfSSL	0:d92f9d21154c	1395
wolfSSL	0:d92f9d21154c	1396	static int _Transform384(Sha384* sha384)
wolfSSL	0:d92f9d21154c	1397	{
wolfSSL	0:d92f9d21154c	1398	const word64* K = K512;
wolfSSL	0:d92f9d21154c	1399
wolfSSL	0:d92f9d21154c	1400	word32 j;
wolfSSL	0:d92f9d21154c	1401	word64 T[8];
wolfSSL	0:d92f9d21154c	1402
wolfSSL	0:d92f9d21154c	1403	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	1404	word64* W;
wolfSSL	0:d92f9d21154c	1405
wolfSSL	0:d92f9d21154c	1406	W = (word64) XMALLOC(sizeof(word64) 16, NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	1407	if (W == NULL)
wolfSSL	0:d92f9d21154c	1408	return MEMORY_E;
wolfSSL	0:d92f9d21154c	1409	#else
wolfSSL	0:d92f9d21154c	1410	word64 W[16];
wolfSSL	0:d92f9d21154c	1411	#endif
wolfSSL	0:d92f9d21154c	1412
wolfSSL	0:d92f9d21154c	1413	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	1414	XMEMCPY(T, sha384->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	1415
wolfSSL	0:d92f9d21154c	1416	#ifdef USE_SLOW_SHA2
wolfSSL	0:d92f9d21154c	1417	/* over twice as small, but 50% slower */
wolfSSL	0:d92f9d21154c	1418	/* 80 operations, not unrolled */
wolfSSL	0:d92f9d21154c	1419	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1420	int m;
wolfSSL	0:d92f9d21154c	1421	for (m = 0; m < 16; m++) { /* braces needed for macros {} */
wolfSSL	0:d92f9d21154c	1422	R2(m);
wolfSSL	0:d92f9d21154c	1423	}
wolfSSL	0:d92f9d21154c	1424	}
wolfSSL	0:d92f9d21154c	1425	#else
wolfSSL	0:d92f9d21154c	1426	/* 80 operations, partially loop unrolled */
wolfSSL	0:d92f9d21154c	1427	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1428	R2( 0); R2( 1); R2( 2); R2( 3);
wolfSSL	0:d92f9d21154c	1429	R2( 4); R2( 5); R2( 6); R2( 7);
wolfSSL	0:d92f9d21154c	1430	R2( 8); R2( 9); R2(10); R2(11);
wolfSSL	0:d92f9d21154c	1431	R2(12); R2(13); R2(14); R2(15);
wolfSSL	0:d92f9d21154c	1432	}
wolfSSL	0:d92f9d21154c	1433	#endif /* USE_SLOW_SHA2 */
wolfSSL	0:d92f9d21154c	1434
wolfSSL	0:d92f9d21154c	1435	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	1436
wolfSSL	0:d92f9d21154c	1437	sha384->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	1438	sha384->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	1439	sha384->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	1440	sha384->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	1441	sha384->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	1442	sha384->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	1443	sha384->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	1444	sha384->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	1445
wolfSSL	0:d92f9d21154c	1446	/* Wipe variables */
wolfSSL	0:d92f9d21154c	1447	XMEMSET(W, 0, sizeof(word64) * 16);
wolfSSL	0:d92f9d21154c	1448	XMEMSET(T, 0, sizeof(T));
wolfSSL	0:d92f9d21154c	1449
wolfSSL	0:d92f9d21154c	1450	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	1451	XFREE(W, NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	1452	#endif
wolfSSL	0:d92f9d21154c	1453
wolfSSL	0:d92f9d21154c	1454	return 0;
wolfSSL	0:d92f9d21154c	1455	}
wolfSSL	0:d92f9d21154c	1456
wolfSSL	0:d92f9d21154c	1457	static INLINE void AddLength384(Sha384* sha384, word32 len)
wolfSSL	0:d92f9d21154c	1458	{
wolfSSL	0:d92f9d21154c	1459	word32 tmp = sha384->loLen;
wolfSSL	0:d92f9d21154c	1460	if ( (sha384->loLen += len) < tmp)
wolfSSL	0:d92f9d21154c	1461	sha384->hiLen++; /* carry low to high */
wolfSSL	0:d92f9d21154c	1462	}
wolfSSL	0:d92f9d21154c	1463
wolfSSL	0:d92f9d21154c	1464	int wc_Sha384Update(Sha384* sha384, const byte* data, word32 len)
wolfSSL	0:d92f9d21154c	1465	{
wolfSSL	0:d92f9d21154c	1466	/* do block size increments */
wolfSSL	0:d92f9d21154c	1467	byte* local = (byte*)sha384->buffer;
wolfSSL	0:d92f9d21154c	1468
wolfSSL	0:d92f9d21154c	1469	SAVE_XMM_YMM ; /* for Intel AVX */
wolfSSL	0:d92f9d21154c	1470
wolfSSL	0:d92f9d21154c	1471	while (len) {
wolfSSL	0:d92f9d21154c	1472	word32 add = min(len, SHA384_BLOCK_SIZE - sha384->buffLen);
wolfSSL	0:d92f9d21154c	1473	XMEMCPY(&local[sha384->buffLen], data, add);
wolfSSL	0:d92f9d21154c	1474
wolfSSL	0:d92f9d21154c	1475	sha384->buffLen += add;
wolfSSL	0:d92f9d21154c	1476	data += add;
wolfSSL	0:d92f9d21154c	1477	len -= add;
wolfSSL	0:d92f9d21154c	1478
wolfSSL	0:d92f9d21154c	1479	if (sha384->buffLen == SHA384_BLOCK_SIZE) {
wolfSSL	0:d92f9d21154c	1480	int ret;
wolfSSL	0:d92f9d21154c	1481
wolfSSL	0:d92f9d21154c	1482	#if defined(LITTLE_ENDIAN_ORDER)
wolfSSL	0:d92f9d21154c	1483	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1484	if(!IS_INTEL_AVX1 && !IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1485	#endif
wolfSSL	0:d92f9d21154c	1486	ByteReverseWords64(sha384->buffer, sha384->buffer,
wolfSSL	0:d92f9d21154c	1487	SHA384_BLOCK_SIZE);
wolfSSL	0:d92f9d21154c	1488	#endif
wolfSSL	0:d92f9d21154c	1489	ret = Transform384(sha384);
wolfSSL	0:d92f9d21154c	1490	if (ret != 0)
wolfSSL	0:d92f9d21154c	1491	return ret;
wolfSSL	0:d92f9d21154c	1492
wolfSSL	0:d92f9d21154c	1493	AddLength384(sha384, SHA384_BLOCK_SIZE);
wolfSSL	0:d92f9d21154c	1494	sha384->buffLen = 0;
wolfSSL	0:d92f9d21154c	1495	}
wolfSSL	0:d92f9d21154c	1496	}
wolfSSL	0:d92f9d21154c	1497	return 0;
wolfSSL	0:d92f9d21154c	1498	}
wolfSSL	0:d92f9d21154c	1499
wolfSSL	0:d92f9d21154c	1500
wolfSSL	0:d92f9d21154c	1501	int wc_Sha384Final(Sha384* sha384, byte* hash)
wolfSSL	0:d92f9d21154c	1502	{
wolfSSL	0:d92f9d21154c	1503	byte* local = (byte*)sha384->buffer;
wolfSSL	0:d92f9d21154c	1504	int ret;
wolfSSL	0:d92f9d21154c	1505
wolfSSL	0:d92f9d21154c	1506	SAVE_XMM_YMM ; /* for Intel AVX */
wolfSSL	0:d92f9d21154c	1507	AddLength384(sha384, sha384->buffLen); /* before adding pads */
wolfSSL	0:d92f9d21154c	1508
wolfSSL	0:d92f9d21154c	1509	local[sha384->buffLen++] = 0x80; /* add 1 */
wolfSSL	0:d92f9d21154c	1510
wolfSSL	0:d92f9d21154c	1511	/* pad with zeros */
wolfSSL	0:d92f9d21154c	1512	if (sha384->buffLen > SHA384_PAD_SIZE) {
wolfSSL	0:d92f9d21154c	1513	XMEMSET(&local[sha384->buffLen], 0, SHA384_BLOCK_SIZE -sha384->buffLen);
wolfSSL	0:d92f9d21154c	1514	sha384->buffLen += SHA384_BLOCK_SIZE - sha384->buffLen;
wolfSSL	0:d92f9d21154c	1515
wolfSSL	0:d92f9d21154c	1516	#if defined(LITTLE_ENDIAN_ORDER)
wolfSSL	0:d92f9d21154c	1517	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1518	if(!IS_INTEL_AVX1 && !IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1519	#endif
wolfSSL	0:d92f9d21154c	1520	ByteReverseWords64(sha384->buffer, sha384->buffer,
wolfSSL	0:d92f9d21154c	1521	SHA384_BLOCK_SIZE);
wolfSSL	0:d92f9d21154c	1522	#endif
wolfSSL	0:d92f9d21154c	1523	ret = Transform384(sha384);
wolfSSL	0:d92f9d21154c	1524	if (ret != 0)
wolfSSL	0:d92f9d21154c	1525	return ret;
wolfSSL	0:d92f9d21154c	1526
wolfSSL	0:d92f9d21154c	1527	sha384->buffLen = 0;
wolfSSL	0:d92f9d21154c	1528	}
wolfSSL	0:d92f9d21154c	1529	XMEMSET(&local[sha384->buffLen], 0, SHA384_PAD_SIZE - sha384->buffLen);
wolfSSL	0:d92f9d21154c	1530
wolfSSL	0:d92f9d21154c	1531	/* put lengths in bits */
wolfSSL	0:d92f9d21154c	1532	sha384->hiLen = (sha384->loLen >> (8*sizeof(sha384->loLen) - 3)) +
wolfSSL	0:d92f9d21154c	1533	(sha384->hiLen << 3);
wolfSSL	0:d92f9d21154c	1534	sha384->loLen = sha384->loLen << 3;
wolfSSL	0:d92f9d21154c	1535
wolfSSL	0:d92f9d21154c	1536	/* store lengths */
wolfSSL	0:d92f9d21154c	1537	#if defined(LITTLE_ENDIAN_ORDER)
wolfSSL	0:d92f9d21154c	1538	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1539	if(!IS_INTEL_AVX1 && !IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1540	#endif
wolfSSL	0:d92f9d21154c	1541	ByteReverseWords64(sha384->buffer, sha384->buffer,
wolfSSL	0:d92f9d21154c	1542	SHA384_BLOCK_SIZE);
wolfSSL	0:d92f9d21154c	1543	#endif
wolfSSL	0:d92f9d21154c	1544	/* ! length ordering dependent on digest endian type ! */
wolfSSL	0:d92f9d21154c	1545	sha384->buffer[SHA384_BLOCK_SIZE / sizeof(word64) - 2] = sha384->hiLen;
wolfSSL	0:d92f9d21154c	1546	sha384->buffer[SHA384_BLOCK_SIZE / sizeof(word64) - 1] = sha384->loLen;
wolfSSL	0:d92f9d21154c	1547	#if defined(HAVE_INTEL_AVX1) \|\| defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1548	if(IS_INTEL_AVX1 \|\| IS_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1549	ByteReverseWords64(&(sha384->buffer[SHA384_BLOCK_SIZE / sizeof(word64) - 2]),
wolfSSL	0:d92f9d21154c	1550	&(sha384->buffer[SHA384_BLOCK_SIZE / sizeof(word64) - 2]),
wolfSSL	0:d92f9d21154c	1551	SHA384_BLOCK_SIZE - SHA384_PAD_SIZE);
wolfSSL	0:d92f9d21154c	1552	#endif
wolfSSL	0:d92f9d21154c	1553	ret = Transform384(sha384);
wolfSSL	0:d92f9d21154c	1554	if (ret != 0)
wolfSSL	0:d92f9d21154c	1555	return ret;
wolfSSL	0:d92f9d21154c	1556
wolfSSL	0:d92f9d21154c	1557	#ifdef LITTLE_ENDIAN_ORDER
wolfSSL	0:d92f9d21154c	1558	ByteReverseWords64(sha384->digest, sha384->digest, SHA384_DIGEST_SIZE);
wolfSSL	0:d92f9d21154c	1559	#endif
wolfSSL	0:d92f9d21154c	1560	XMEMCPY(hash, sha384->digest, SHA384_DIGEST_SIZE);
wolfSSL	0:d92f9d21154c	1561
wolfSSL	0:d92f9d21154c	1562	return wc_InitSha384(sha384); /* reset state */
wolfSSL	0:d92f9d21154c	1563	}
wolfSSL	0:d92f9d21154c	1564
wolfSSL	0:d92f9d21154c	1565
wolfSSL	0:d92f9d21154c	1566	int wc_Sha384Hash(const byte* data, word32 len, byte* hash)
wolfSSL	0:d92f9d21154c	1567	{
wolfSSL	0:d92f9d21154c	1568	int ret = 0;
wolfSSL	0:d92f9d21154c	1569	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	1570	Sha384* sha384;
wolfSSL	0:d92f9d21154c	1571	#else
wolfSSL	0:d92f9d21154c	1572	Sha384 sha384[1];
wolfSSL	0:d92f9d21154c	1573	#endif
wolfSSL	0:d92f9d21154c	1574
wolfSSL	0:d92f9d21154c	1575	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	1576	sha384 = (Sha384*)XMALLOC(sizeof(Sha384), NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	1577	if (sha384 == NULL)
wolfSSL	0:d92f9d21154c	1578	return MEMORY_E;
wolfSSL	0:d92f9d21154c	1579	#endif
wolfSSL	0:d92f9d21154c	1580
wolfSSL	0:d92f9d21154c	1581	if ((ret = wc_InitSha384(sha384)) != 0) {
wolfSSL	0:d92f9d21154c	1582	WOLFSSL_MSG("InitSha384 failed");
wolfSSL	0:d92f9d21154c	1583	}
wolfSSL	0:d92f9d21154c	1584	else if ((ret = wc_Sha384Update(sha384, data, len)) != 0) {
wolfSSL	0:d92f9d21154c	1585	WOLFSSL_MSG("Sha384Update failed");
wolfSSL	0:d92f9d21154c	1586	}
wolfSSL	0:d92f9d21154c	1587	else if ((ret = wc_Sha384Final(sha384, hash)) != 0) {
wolfSSL	0:d92f9d21154c	1588	WOLFSSL_MSG("Sha384Final failed");
wolfSSL	0:d92f9d21154c	1589	}
wolfSSL	0:d92f9d21154c	1590
wolfSSL	0:d92f9d21154c	1591	#ifdef WOLFSSL_SMALL_STACK
wolfSSL	0:d92f9d21154c	1592	XFREE(sha384, NULL, DYNAMIC_TYPE_TMP_BUFFER);
wolfSSL	0:d92f9d21154c	1593	#endif
wolfSSL	0:d92f9d21154c	1594
wolfSSL	0:d92f9d21154c	1595	return ret;
wolfSSL	0:d92f9d21154c	1596	}
wolfSSL	0:d92f9d21154c	1597
wolfSSL	0:d92f9d21154c	1598	#if defined(HAVE_INTEL_AVX1)
wolfSSL	0:d92f9d21154c	1599
wolfSSL	0:d92f9d21154c	1600	static int Transform384_AVX1(Sha384* sha384)
wolfSSL	0:d92f9d21154c	1601	{
wolfSSL	0:d92f9d21154c	1602	const word64* K = K512;
wolfSSL	0:d92f9d21154c	1603	word64 W_X[16+4];
wolfSSL	0:d92f9d21154c	1604	word32 j;
wolfSSL	0:d92f9d21154c	1605	word64 T[8];
wolfSSL	0:d92f9d21154c	1606
wolfSSL	0:d92f9d21154c	1607	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	1608	XMEMCPY(T, sha384->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	1609	W_from_buff(W_X, sha384->buffer) ;
wolfSSL	0:d92f9d21154c	1610	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1611	Rx_1( 0); Block_0_1(W_X); Rx_2( 0); Block_0_2(W_X); Rx_3( 0); Block_0_3();
wolfSSL	0:d92f9d21154c	1612	Rx_1( 1); Block_0_4(); Rx_2( 1); Block_0_5(); Rx_3( 1); Block_0_6(W_X);
wolfSSL	0:d92f9d21154c	1613	Rx_1( 2); Block_0_7(W_X); Rx_2( 2); Block_0_8(W_X); Rx_3( 2); Block_0_9();
wolfSSL	0:d92f9d21154c	1614	Rx_1( 3); Block_0_10();Rx_2( 3); Block_0_11();Rx_3( 3); Block_0_12(W_X);
wolfSSL	0:d92f9d21154c	1615
wolfSSL	0:d92f9d21154c	1616	Rx_1( 4); Block_4_1(W_X); Rx_2( 4); Block_4_2(W_X); Rx_3( 4); Block_4_3();
wolfSSL	0:d92f9d21154c	1617	Rx_1( 5); Block_4_4(); Rx_2( 5); Block_4_5(); Rx_3( 5); Block_4_6(W_X);
wolfSSL	0:d92f9d21154c	1618	Rx_1( 6); Block_4_7(W_X); Rx_2( 6); Block_4_8(W_X); Rx_3( 6); Block_4_9();
wolfSSL	0:d92f9d21154c	1619	Rx_1( 7); Block_4_10();Rx_2( 7); Block_4_11();Rx_3( 7); Block_4_12(W_X);
wolfSSL	0:d92f9d21154c	1620
wolfSSL	0:d92f9d21154c	1621	Rx_1( 8); Block_8_1(W_X); Rx_2( 8); Block_8_2(W_X); Rx_3( 8); Block_8_3();
wolfSSL	0:d92f9d21154c	1622	Rx_1( 9); Block_8_4(); Rx_2( 9); Block_8_5(); Rx_3( 9); Block_8_6(W_X);
wolfSSL	0:d92f9d21154c	1623	Rx_1(10); Block_8_7(W_X); Rx_2(10); Block_8_8(W_X); Rx_3(10); Block_8_9();
wolfSSL	0:d92f9d21154c	1624	Rx_1(11); Block_8_10();Rx_2(11); Block_8_11();Rx_3(11); Block_8_12(W_X);
wolfSSL	0:d92f9d21154c	1625
wolfSSL	0:d92f9d21154c	1626	Rx_1(12); Block_12_1(W_X); Rx_2(12); Block_12_2(W_X); Rx_3(12); Block_12_3();
wolfSSL	0:d92f9d21154c	1627	Rx_1(13); Block_12_4(); Rx_2(13); Block_12_5(); Rx_3(13); Block_12_6(W_X);
wolfSSL	0:d92f9d21154c	1628	Rx_1(14); Block_12_7(W_X); Rx_2(14); Block_12_8(W_X); Rx_3(14); Block_12_9();
wolfSSL	0:d92f9d21154c	1629	Rx_1(15); Block_12_10();Rx_2(15); Block_12_11();Rx_3(15); Block_12_12(W_X);
wolfSSL	0:d92f9d21154c	1630	}
wolfSSL	0:d92f9d21154c	1631
wolfSSL	0:d92f9d21154c	1632	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	1633
wolfSSL	0:d92f9d21154c	1634	sha384->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	1635	sha384->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	1636	sha384->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	1637	sha384->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	1638	sha384->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	1639	sha384->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	1640	sha384->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	1641	sha384->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	1642
wolfSSL	0:d92f9d21154c	1643	/* Wipe variables */
wolfSSL	0:d92f9d21154c	1644	#if !defined(HAVE_INTEL_AVX1)&&!defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1645	XMEMSET(W, 0, sizeof(word64) * 16);
wolfSSL	0:d92f9d21154c	1646	#endif
wolfSSL	0:d92f9d21154c	1647	XMEMSET(T, 0, sizeof(T));
wolfSSL	0:d92f9d21154c	1648
wolfSSL	0:d92f9d21154c	1649	return 0;
wolfSSL	0:d92f9d21154c	1650	}
wolfSSL	0:d92f9d21154c	1651
wolfSSL	0:d92f9d21154c	1652	#endif
wolfSSL	0:d92f9d21154c	1653
wolfSSL	0:d92f9d21154c	1654	#if defined(HAVE_INTEL_AVX1) && defined(HAVE_INTEL_AVX2) && defined(HAVE_INTEL_RORX)
wolfSSL	0:d92f9d21154c	1655	static int Transform384_AVX1_RORX(Sha384* sha384)
wolfSSL	0:d92f9d21154c	1656	{
wolfSSL	0:d92f9d21154c	1657	const word64* K = K512;
wolfSSL	0:d92f9d21154c	1658	word64 W_X[16+4];
wolfSSL	0:d92f9d21154c	1659	word32 j;
wolfSSL	0:d92f9d21154c	1660	word64 T[8];
wolfSSL	0:d92f9d21154c	1661
wolfSSL	0:d92f9d21154c	1662	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	1663	XMEMCPY(T, sha384->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	1664
wolfSSL	0:d92f9d21154c	1665	W_from_buff(W_X, sha384->buffer) ;
wolfSSL	0:d92f9d21154c	1666	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1667	Rx_RORX_1( 0); Block_0_1(W_X); Rx_RORX_2( 0);
wolfSSL	0:d92f9d21154c	1668	Block_0_2(W_X); Rx_RORX_3( 0); Block_0_3();
wolfSSL	0:d92f9d21154c	1669	Rx_RORX_1( 1); Block_0_4(); Rx_RORX_2( 1);
wolfSSL	0:d92f9d21154c	1670	Block_0_5(); Rx_RORX_3( 1); Block_0_6(W_X);
wolfSSL	0:d92f9d21154c	1671	Rx_RORX_1( 2); Block_0_7(W_X); Rx_RORX_2( 2);
wolfSSL	0:d92f9d21154c	1672	Block_0_8(W_X); Rx_RORX_3( 2); Block_0_9();
wolfSSL	0:d92f9d21154c	1673	Rx_RORX_1( 3); Block_0_10();Rx_RORX_2( 3);
wolfSSL	0:d92f9d21154c	1674	Block_0_11();Rx_RORX_3( 3); Block_0_12(W_X);
wolfSSL	0:d92f9d21154c	1675
wolfSSL	0:d92f9d21154c	1676	Rx_RORX_1( 4); Block_4_1(W_X); Rx_RORX_2( 4);
wolfSSL	0:d92f9d21154c	1677	Block_4_2(W_X); Rx_RORX_3( 4); Block_4_3();
wolfSSL	0:d92f9d21154c	1678	Rx_RORX_1( 5); Block_4_4(); Rx_RORX_2( 5);
wolfSSL	0:d92f9d21154c	1679	Block_4_5(); Rx_RORX_3( 5); Block_4_6(W_X);
wolfSSL	0:d92f9d21154c	1680	Rx_RORX_1( 6); Block_4_7(W_X); Rx_RORX_2( 6);
wolfSSL	0:d92f9d21154c	1681	Block_4_8(W_X); Rx_RORX_3( 6); Block_4_9();
wolfSSL	0:d92f9d21154c	1682	Rx_RORX_1( 7); Block_4_10();Rx_RORX_2( 7);
wolfSSL	0:d92f9d21154c	1683	Block_4_11();Rx_RORX_3( 7); Block_4_12(W_X);
wolfSSL	0:d92f9d21154c	1684
wolfSSL	0:d92f9d21154c	1685	Rx_RORX_1( 8); Block_8_1(W_X); Rx_RORX_2( 8);
wolfSSL	0:d92f9d21154c	1686	Block_8_2(W_X); Rx_RORX_3( 8); Block_8_3();
wolfSSL	0:d92f9d21154c	1687	Rx_RORX_1( 9); Block_8_4(); Rx_RORX_2( 9);
wolfSSL	0:d92f9d21154c	1688	Block_8_5(); Rx_RORX_3( 9); Block_8_6(W_X);
wolfSSL	0:d92f9d21154c	1689	Rx_RORX_1(10); Block_8_7(W_X); Rx_RORX_2(10);
wolfSSL	0:d92f9d21154c	1690	Block_8_8(W_X); Rx_RORX_3(10); Block_8_9();
wolfSSL	0:d92f9d21154c	1691	Rx_RORX_1(11); Block_8_10();Rx_RORX_2(11);
wolfSSL	0:d92f9d21154c	1692	Block_8_11();Rx_RORX_3(11); Block_8_12(W_X);
wolfSSL	0:d92f9d21154c	1693
wolfSSL	0:d92f9d21154c	1694	Rx_RORX_1(12); Block_12_1(W_X); Rx_RORX_2(12);
wolfSSL	0:d92f9d21154c	1695	Block_12_2(W_X); Rx_RORX_3(12); Block_12_3();
wolfSSL	0:d92f9d21154c	1696	Rx_RORX_1(13); Block_12_4(); Rx_RORX_2(13);
wolfSSL	0:d92f9d21154c	1697	Block_12_5(); Rx_RORX_3(13); Block_12_6(W_X);
wolfSSL	0:d92f9d21154c	1698	Rx_RORX_1(14); Block_12_7(W_X); Rx_RORX_2(14);
wolfSSL	0:d92f9d21154c	1699	Block_12_8(W_X); Rx_RORX_3(14); Block_12_9();
wolfSSL	0:d92f9d21154c	1700	Rx_RORX_1(15); Block_12_10();Rx_RORX_2(15);
wolfSSL	0:d92f9d21154c	1701	Block_12_11();Rx_RORX_3(15); Block_12_12(W_X);
wolfSSL	0:d92f9d21154c	1702	}
wolfSSL	0:d92f9d21154c	1703
wolfSSL	0:d92f9d21154c	1704	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	1705
wolfSSL	0:d92f9d21154c	1706	sha384->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	1707	sha384->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	1708	sha384->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	1709	sha384->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	1710	sha384->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	1711	sha384->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	1712	sha384->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	1713	sha384->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	1714
wolfSSL	0:d92f9d21154c	1715	/* Wipe variables */
wolfSSL	0:d92f9d21154c	1716	#if !defined(HAVE_INTEL_AVX1)&&!defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1717	XMEMSET(W, 0, sizeof(word64) * 16);
wolfSSL	0:d92f9d21154c	1718	#endif
wolfSSL	0:d92f9d21154c	1719	XMEMSET(T, 0, sizeof(T));
wolfSSL	0:d92f9d21154c	1720
wolfSSL	0:d92f9d21154c	1721	return 0;
wolfSSL	0:d92f9d21154c	1722	}
wolfSSL	0:d92f9d21154c	1723	#endif
wolfSSL	0:d92f9d21154c	1724
wolfSSL	0:d92f9d21154c	1725	#if defined(HAVE_INTEL_AVX2)
wolfSSL	0:d92f9d21154c	1726
wolfSSL	0:d92f9d21154c	1727	static int Transform384_AVX2(Sha384* sha384)
wolfSSL	0:d92f9d21154c	1728	{
wolfSSL	0:d92f9d21154c	1729	const word64* K = K512;
wolfSSL	0:d92f9d21154c	1730	word64 w[4] ;
wolfSSL	0:d92f9d21154c	1731	word32 j;
wolfSSL	0:d92f9d21154c	1732	word64 T[8];
wolfSSL	0:d92f9d21154c	1733
wolfSSL	0:d92f9d21154c	1734	/* Copy digest to working vars */
wolfSSL	0:d92f9d21154c	1735	XMEMCPY(T, sha384->digest, sizeof(T));
wolfSSL	0:d92f9d21154c	1736
wolfSSL	0:d92f9d21154c	1737	/* over twice as small, but 50% slower */
wolfSSL	0:d92f9d21154c	1738	/* 80 operations, not unrolled */
wolfSSL	0:d92f9d21154c	1739
wolfSSL	0:d92f9d21154c	1740	W_from_buff_Y(sha384->buffer) ;
wolfSSL	0:d92f9d21154c	1741
wolfSSL	0:d92f9d21154c	1742	MOVE_to_MEMy(w,0, W_0y) ;
wolfSSL	0:d92f9d21154c	1743	for (j = 0; j < 80; j += 16) {
wolfSSL	0:d92f9d21154c	1744	Ry_1( 0, w[0]); Block_Y_0_1(); Ry_2( 0, w[0]);
wolfSSL	0:d92f9d21154c	1745	Block_Y_0_2(); Ry_3( 0, w[0]); Block_Y_0_3();
wolfSSL	0:d92f9d21154c	1746	Ry_1( 1, w[1]); Block_Y_0_4(); Ry_2( 1, w[1]);
wolfSSL	0:d92f9d21154c	1747	Block_Y_0_5(); Ry_3( 1, w[1]); Block_Y_0_6();
wolfSSL	0:d92f9d21154c	1748	Ry_1( 2, w[2]); Block_Y_0_7(); Ry_2( 2, w[2]);
wolfSSL	0:d92f9d21154c	1749	Block_Y_0_8(); Ry_3( 2, w[2]); Block_Y_0_9();
wolfSSL	0:d92f9d21154c	1750	Ry_1( 3, w[3]); Block_Y_0_10();Ry_2( 3, w[3]);
wolfSSL	0:d92f9d21154c	1751	Block_Y_0_11();Ry_3( 3, w[3]); Block_Y_0_12(w);
wolfSSL	0:d92f9d21154c	1752
wolfSSL	0:d92f9d21154c	1753	Ry_1( 4, w[0]); Block_Y_4_1(); Ry_2( 4, w[0]);
wolfSSL	0:d92f9d21154c	1754	Block_Y_4_2(); Ry_3( 4, w[0]); Block_Y_4_3();
wolfSSL	0:d92f9d21154c	1755	Ry_1( 5, w[1]); Block_Y_4_4(); Ry_2( 5, w[1]);
wolfSSL	0:d92f9d21154c	1756	Block_Y_4_5(); Ry_3( 5, w[1]); Block_Y_4_6();
wolfSSL	0:d92f9d21154c	1757	Ry_1( 6, w[2]); Block_Y_4_7(); Ry_2( 6, w[2]);
wolfSSL	0:d92f9d21154c	1758	Block_Y_4_8(); Ry_3( 6, w[2]); Block_Y_4_9();
wolfSSL	0:d92f9d21154c	1759	Ry_1( 7, w[3]); Block_Y_4_10(); Ry_2( 7, w[3]);
wolfSSL	0:d92f9d21154c	1760	Block_Y_4_11(); Ry_3( 7, w[3]);Block_Y_4_12(w);
wolfSSL	0:d92f9d21154c	1761
wolfSSL	0:d92f9d21154c	1762	Ry_1( 8, w[0]); Block_Y_8_1(); Ry_2( 8, w[0]);
wolfSSL	0:d92f9d21154c	1763	Block_Y_8_2(); Ry_3( 8, w[0]); Block_Y_8_3();
wolfSSL	0:d92f9d21154c	1764	Ry_1( 9, w[1]); Block_Y_8_4(); Ry_2( 9, w[1]);
wolfSSL	0:d92f9d21154c	1765	Block_Y_8_5(); Ry_3( 9, w[1]); Block_Y_8_6();
wolfSSL	0:d92f9d21154c	1766	Ry_1(10, w[2]); Block_Y_8_7(); Ry_2(10, w[2]);
wolfSSL	0:d92f9d21154c	1767	Block_Y_8_8(); Ry_3(10, w[2]); Block_Y_8_9();
wolfSSL	0:d92f9d21154c	1768	Ry_1(11, w[3]); Block_Y_8_10();Ry_2(11, w[3]);
wolfSSL	0:d92f9d21154c	1769	Block_Y_8_11();Ry_3(11, w[3]); Block_Y_8_12(w);
wolfSSL	0:d92f9d21154c	1770
wolfSSL	0:d92f9d21154c	1771	Ry_1(12, w[0]); Block_Y_12_1(); Ry_2(12, w[0]);
wolfSSL	0:d92f9d21154c	1772	Block_Y_12_2(); Ry_3(12, w[0]); Block_Y_12_3();
wolfSSL	0:d92f9d21154c	1773	Ry_1(13, w[1]); Block_Y_12_4(); Ry_2(13, w[1]);
wolfSSL	0:d92f9d21154c	1774	Block_Y_12_5(); Ry_3(13, w[1]); Block_Y_12_6();
wolfSSL	0:d92f9d21154c	1775	Ry_1(14, w[2]); Block_Y_12_7(); Ry_2(14, w[2]);
wolfSSL	0:d92f9d21154c	1776	Block_Y_12_8(); Ry_3(14, w[2]); Block_Y_12_9();
wolfSSL	0:d92f9d21154c	1777	Ry_1(15, w[3]); Block_Y_12_10();Ry_2(15, w[3]);
wolfSSL	0:d92f9d21154c	1778	Block_Y_12_11();Ry_3(15, w[3]); Block_Y_12_12(w);
wolfSSL	0:d92f9d21154c	1779	}
wolfSSL	0:d92f9d21154c	1780
wolfSSL	0:d92f9d21154c	1781	/* Add the working vars back into digest */
wolfSSL	0:d92f9d21154c	1782
wolfSSL	0:d92f9d21154c	1783	sha384->digest[0] += a(0);
wolfSSL	0:d92f9d21154c	1784	sha384->digest[1] += b(0);
wolfSSL	0:d92f9d21154c	1785	sha384->digest[2] += c(0);
wolfSSL	0:d92f9d21154c	1786	sha384->digest[3] += d(0);
wolfSSL	0:d92f9d21154c	1787	sha384->digest[4] += e(0);
wolfSSL	0:d92f9d21154c	1788	sha384->digest[5] += f(0);
wolfSSL	0:d92f9d21154c	1789	sha384->digest[6] += g(0);
wolfSSL	0:d92f9d21154c	1790	sha384->digest[7] += h(0);
wolfSSL	0:d92f9d21154c	1791
wolfSSL	0:d92f9d21154c	1792	/* Wipe variables */
wolfSSL	0:d92f9d21154c	1793	XMEMSET(T, 0, sizeof(T));
wolfSSL	0:d92f9d21154c	1794
wolfSSL	0:d92f9d21154c	1795	return 0;
wolfSSL	0:d92f9d21154c	1796	}
wolfSSL	0:d92f9d21154c	1797
wolfSSL	0:d92f9d21154c	1798	#endif
wolfSSL	0:d92f9d21154c	1799
wolfSSL	0:d92f9d21154c	1800	#endif /* WOLFSSL_SHA384 */
wolfSSL	0:d92f9d21154c	1801
wolfSSL	0:d92f9d21154c	1802	#endif /* HAVE_FIPS */
wolfSSL	0:d92f9d21154c	1803
wolfSSL	0:d92f9d21154c	1804	#endif /* WOLFSSL_SHA512 */
wolfSSL	0:d92f9d21154c	1805
wolfSSL	0:d92f9d21154c	1806

Repository toolbox

Export to desktop IDE

Repository details

Type:	Library
Created:	30 May 2017
Imports:	0
Forks:	0
Commits:	14
Dependents:	0
Dependencies:	0
Followers:	1

This repository is Public (Unlisted).

wolfcrypt/src/sha512.c@0:d92f9d21154c, 2015-06-26 (annotated)

Who changed what in which revision?

Repository toolbox

Repository details

Important Information for this Arm website

Access Warning