Lee Kai Xuan / mbed-os

Important changes to repositories hosted on mbed.com

Mbed hosted mercurial repositories are deprecated and are due to be permanently deleted in July 2026.

To keep a copy of this software download the repository Zip archive or clone locally using Mercurial.

It is also possible to export all your personal repositories from the account settings page.

Fork of mbed-os by erkin yucel

features/mbedtls/inc/mbedtls/bn_mul.h@1:3deb71413561, 2017-07-20 (annotated)

Committer:
xuaner
Date:
Thu Jul 20 14:26:57 2017 +0000
Revision:
1:3deb71413561
Parent:
0:f269e3021894 
mbed_os

Who changed what in which revision?

UserRevisionLine numberNew contents of line
elessair 0:f269e3021894 1 /**
elessair 0:f269e3021894 2 * \file bn_mul.h
elessair 0:f269e3021894 3 *
elessair 0:f269e3021894 4 * \brief Multi-precision integer library
elessair 0:f269e3021894 5 *
elessair 0:f269e3021894 6 * Copyright (C) 2006-2015, ARM Limited, All Rights Reserved
elessair 0:f269e3021894 7 * SPDX-License-Identifier: Apache-2.0
elessair 0:f269e3021894 8 *
elessair 0:f269e3021894 9 * Licensed under the Apache License, Version 2.0 (the "License"); you may
elessair 0:f269e3021894 10 * not use this file except in compliance with the License.
elessair 0:f269e3021894 11 * You may obtain a copy of the License at
elessair 0:f269e3021894 12 *
elessair 0:f269e3021894 13 * http://www.apache.org/licenses/LICENSE-2.0
elessair 0:f269e3021894 14 *
elessair 0:f269e3021894 15 * Unless required by applicable law or agreed to in writing, software
elessair 0:f269e3021894 16 * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT
elessair 0:f269e3021894 17 * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
elessair 0:f269e3021894 18 * See the License for the specific language governing permissions and
elessair 0:f269e3021894 19 * limitations under the License.
elessair 0:f269e3021894 20 *
elessair 0:f269e3021894 21 * This file is part of mbed TLS (https://tls.mbed.org)
elessair 0:f269e3021894 22 */
elessair 0:f269e3021894 23 /*
elessair 0:f269e3021894 24 * Multiply source vector [s] with b, add result
elessair 0:f269e3021894 25 * to destination vector [d] and set carry c.
elessair 0:f269e3021894 26 *
elessair 0:f269e3021894 27 * Currently supports:
elessair 0:f269e3021894 28 *
elessair 0:f269e3021894 29 * . IA-32 (386+) . AMD64 / EM64T
elessair 0:f269e3021894 30 * . IA-32 (SSE2) . Motorola 68000
elessair 0:f269e3021894 31 * . PowerPC, 32-bit . MicroBlaze
elessair 0:f269e3021894 32 * . PowerPC, 64-bit . TriCore
elessair 0:f269e3021894 33 * . SPARC v8 . ARM v3+
elessair 0:f269e3021894 34 * . Alpha . MIPS32
elessair 0:f269e3021894 35 * . C, longlong . C, generic
elessair 0:f269e3021894 36 */
elessair 0:f269e3021894 37 #ifndef MBEDTLS_BN_MUL_H
elessair 0:f269e3021894 38 #define MBEDTLS_BN_MUL_H
elessair 0:f269e3021894 39
elessair 0:f269e3021894 40 #include "bignum.h"
elessair 0:f269e3021894 41
elessair 0:f269e3021894 42 #if defined(MBEDTLS_HAVE_ASM)
elessair 0:f269e3021894 43
elessair 0:f269e3021894 44 #ifndef asm
elessair 0:f269e3021894 45 #define asm __asm
elessair 0:f269e3021894 46 #endif
elessair 0:f269e3021894 47
elessair 0:f269e3021894 48 /* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */
elessair 0:f269e3021894 49 #if defined(__GNUC__) && \
elessair 0:f269e3021894 50 ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 )
elessair 0:f269e3021894 51 #if defined(__i386__)
elessair 0:f269e3021894 52
elessair 0:f269e3021894 53 #define MULADDC_INIT \
elessair 0:f269e3021894 54 asm( \
elessair 0:f269e3021894 55 "movl %%ebx, %0 \n\t" \
elessair 0:f269e3021894 56 "movl %5, %%esi \n\t" \
elessair 0:f269e3021894 57 "movl %6, %%edi \n\t" \
elessair 0:f269e3021894 58 "movl %7, %%ecx \n\t" \
elessair 0:f269e3021894 59 "movl %8, %%ebx \n\t"
elessair 0:f269e3021894 60
elessair 0:f269e3021894 61 #define MULADDC_CORE \
elessair 0:f269e3021894 62 "lodsl \n\t" \
elessair 0:f269e3021894 63 "mull %%ebx \n\t" \
elessair 0:f269e3021894 64 "addl %%ecx, %%eax \n\t" \
elessair 0:f269e3021894 65 "adcl $0, %%edx \n\t" \
elessair 0:f269e3021894 66 "addl (%%edi), %%eax \n\t" \
elessair 0:f269e3021894 67 "adcl $0, %%edx \n\t" \
elessair 0:f269e3021894 68 "movl %%edx, %%ecx \n\t" \
elessair 0:f269e3021894 69 "stosl \n\t"
elessair 0:f269e3021894 70
elessair 0:f269e3021894 71 #if defined(MBEDTLS_HAVE_SSE2)
elessair 0:f269e3021894 72
elessair 0:f269e3021894 73 #define MULADDC_HUIT \
elessair 0:f269e3021894 74 "movd %%ecx, %%mm1 \n\t" \
elessair 0:f269e3021894 75 "movd %%ebx, %%mm0 \n\t" \
elessair 0:f269e3021894 76 "movd (%%edi), %%mm3 \n\t" \
elessair 0:f269e3021894 77 "paddq %%mm3, %%mm1 \n\t" \
elessair 0:f269e3021894 78 "movd (%%esi), %%mm2 \n\t" \
elessair 0:f269e3021894 79 "pmuludq %%mm0, %%mm2 \n\t" \
elessair 0:f269e3021894 80 "movd 4(%%esi), %%mm4 \n\t" \
elessair 0:f269e3021894 81 "pmuludq %%mm0, %%mm4 \n\t" \
elessair 0:f269e3021894 82 "movd 8(%%esi), %%mm6 \n\t" \
elessair 0:f269e3021894 83 "pmuludq %%mm0, %%mm6 \n\t" \
elessair 0:f269e3021894 84 "movd 12(%%esi), %%mm7 \n\t" \
elessair 0:f269e3021894 85 "pmuludq %%mm0, %%mm7 \n\t" \
elessair 0:f269e3021894 86 "paddq %%mm2, %%mm1 \n\t" \
elessair 0:f269e3021894 87 "movd 4(%%edi), %%mm3 \n\t" \
elessair 0:f269e3021894 88 "paddq %%mm4, %%mm3 \n\t" \
elessair 0:f269e3021894 89 "movd 8(%%edi), %%mm5 \n\t" \
elessair 0:f269e3021894 90 "paddq %%mm6, %%mm5 \n\t" \
elessair 0:f269e3021894 91 "movd 12(%%edi), %%mm4 \n\t" \
elessair 0:f269e3021894 92 "paddq %%mm4, %%mm7 \n\t" \
elessair 0:f269e3021894 93 "movd %%mm1, (%%edi) \n\t" \
elessair 0:f269e3021894 94 "movd 16(%%esi), %%mm2 \n\t" \
elessair 0:f269e3021894 95 "pmuludq %%mm0, %%mm2 \n\t" \
elessair 0:f269e3021894 96 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 97 "movd 20(%%esi), %%mm4 \n\t" \
elessair 0:f269e3021894 98 "pmuludq %%mm0, %%mm4 \n\t" \
elessair 0:f269e3021894 99 "paddq %%mm3, %%mm1 \n\t" \
elessair 0:f269e3021894 100 "movd 24(%%esi), %%mm6 \n\t" \
elessair 0:f269e3021894 101 "pmuludq %%mm0, %%mm6 \n\t" \
elessair 0:f269e3021894 102 "movd %%mm1, 4(%%edi) \n\t" \
elessair 0:f269e3021894 103 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 104 "movd 28(%%esi), %%mm3 \n\t" \
elessair 0:f269e3021894 105 "pmuludq %%mm0, %%mm3 \n\t" \
elessair 0:f269e3021894 106 "paddq %%mm5, %%mm1 \n\t" \
elessair 0:f269e3021894 107 "movd 16(%%edi), %%mm5 \n\t" \
elessair 0:f269e3021894 108 "paddq %%mm5, %%mm2 \n\t" \
elessair 0:f269e3021894 109 "movd %%mm1, 8(%%edi) \n\t" \
elessair 0:f269e3021894 110 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 111 "paddq %%mm7, %%mm1 \n\t" \
elessair 0:f269e3021894 112 "movd 20(%%edi), %%mm5 \n\t" \
elessair 0:f269e3021894 113 "paddq %%mm5, %%mm4 \n\t" \
elessair 0:f269e3021894 114 "movd %%mm1, 12(%%edi) \n\t" \
elessair 0:f269e3021894 115 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 116 "paddq %%mm2, %%mm1 \n\t" \
elessair 0:f269e3021894 117 "movd 24(%%edi), %%mm5 \n\t" \
elessair 0:f269e3021894 118 "paddq %%mm5, %%mm6 \n\t" \
elessair 0:f269e3021894 119 "movd %%mm1, 16(%%edi) \n\t" \
elessair 0:f269e3021894 120 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 121 "paddq %%mm4, %%mm1 \n\t" \
elessair 0:f269e3021894 122 "movd 28(%%edi), %%mm5 \n\t" \
elessair 0:f269e3021894 123 "paddq %%mm5, %%mm3 \n\t" \
elessair 0:f269e3021894 124 "movd %%mm1, 20(%%edi) \n\t" \
elessair 0:f269e3021894 125 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 126 "paddq %%mm6, %%mm1 \n\t" \
elessair 0:f269e3021894 127 "movd %%mm1, 24(%%edi) \n\t" \
elessair 0:f269e3021894 128 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 129 "paddq %%mm3, %%mm1 \n\t" \
elessair 0:f269e3021894 130 "movd %%mm1, 28(%%edi) \n\t" \
elessair 0:f269e3021894 131 "addl $32, %%edi \n\t" \
elessair 0:f269e3021894 132 "addl $32, %%esi \n\t" \
elessair 0:f269e3021894 133 "psrlq $32, %%mm1 \n\t" \
elessair 0:f269e3021894 134 "movd %%mm1, %%ecx \n\t"
elessair 0:f269e3021894 135
elessair 0:f269e3021894 136 #define MULADDC_STOP \
elessair 0:f269e3021894 137 "emms \n\t" \
elessair 0:f269e3021894 138 "movl %4, %%ebx \n\t" \
elessair 0:f269e3021894 139 "movl %%ecx, %1 \n\t" \
elessair 0:f269e3021894 140 "movl %%edi, %2 \n\t" \
elessair 0:f269e3021894 141 "movl %%esi, %3 \n\t" \
elessair 0:f269e3021894 142 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 143 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 144 : "eax", "ecx", "edx", "esi", "edi" \
elessair 0:f269e3021894 145 );
elessair 0:f269e3021894 146
elessair 0:f269e3021894 147 #else
elessair 0:f269e3021894 148
elessair 0:f269e3021894 149 #define MULADDC_STOP \
elessair 0:f269e3021894 150 "movl %4, %%ebx \n\t" \
elessair 0:f269e3021894 151 "movl %%ecx, %1 \n\t" \
elessair 0:f269e3021894 152 "movl %%edi, %2 \n\t" \
elessair 0:f269e3021894 153 "movl %%esi, %3 \n\t" \
elessair 0:f269e3021894 154 : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 155 : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 156 : "eax", "ecx", "edx", "esi", "edi" \
elessair 0:f269e3021894 157 );
elessair 0:f269e3021894 158 #endif /* SSE2 */
elessair 0:f269e3021894 159 #endif /* i386 */
elessair 0:f269e3021894 160
elessair 0:f269e3021894 161 #if defined(__amd64__) || defined (__x86_64__)
elessair 0:f269e3021894 162
elessair 0:f269e3021894 163 #define MULADDC_INIT \
elessair 0:f269e3021894 164 asm( \
elessair 0:f269e3021894 165 "xorq %%r8, %%r8 \n\t"
elessair 0:f269e3021894 166
elessair 0:f269e3021894 167 #define MULADDC_CORE \
elessair 0:f269e3021894 168 "movq (%%rsi), %%rax \n\t" \
elessair 0:f269e3021894 169 "mulq %%rbx \n\t" \
elessair 0:f269e3021894 170 "addq $8, %%rsi \n\t" \
elessair 0:f269e3021894 171 "addq %%rcx, %%rax \n\t" \
elessair 0:f269e3021894 172 "movq %%r8, %%rcx \n\t" \
elessair 0:f269e3021894 173 "adcq $0, %%rdx \n\t" \
elessair 0:f269e3021894 174 "nop \n\t" \
elessair 0:f269e3021894 175 "addq %%rax, (%%rdi) \n\t" \
elessair 0:f269e3021894 176 "adcq %%rdx, %%rcx \n\t" \
elessair 0:f269e3021894 177 "addq $8, %%rdi \n\t"
elessair 0:f269e3021894 178
elessair 0:f269e3021894 179 #define MULADDC_STOP \
elessair 0:f269e3021894 180 : "+c" (c), "+D" (d), "+S" (s) \
elessair 0:f269e3021894 181 : "b" (b) \
elessair 0:f269e3021894 182 : "rax", "rdx", "r8" \
elessair 0:f269e3021894 183 );
elessair 0:f269e3021894 184
elessair 0:f269e3021894 185 #endif /* AMD64 */
elessair 0:f269e3021894 186
elessair 0:f269e3021894 187 #if defined(__mc68020__) || defined(__mcpu32__)
elessair 0:f269e3021894 188
elessair 0:f269e3021894 189 #define MULADDC_INIT \
elessair 0:f269e3021894 190 asm( \
elessair 0:f269e3021894 191 "movl %3, %%a2 \n\t" \
elessair 0:f269e3021894 192 "movl %4, %%a3 \n\t" \
elessair 0:f269e3021894 193 "movl %5, %%d3 \n\t" \
elessair 0:f269e3021894 194 "movl %6, %%d2 \n\t" \
elessair 0:f269e3021894 195 "moveq #0, %%d0 \n\t"
elessair 0:f269e3021894 196
elessair 0:f269e3021894 197 #define MULADDC_CORE \
elessair 0:f269e3021894 198 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 199 "mulul %%d2, %%d4:%%d1 \n\t" \
elessair 0:f269e3021894 200 "addl %%d3, %%d1 \n\t" \
elessair 0:f269e3021894 201 "addxl %%d0, %%d4 \n\t" \
elessair 0:f269e3021894 202 "moveq #0, %%d3 \n\t" \
elessair 0:f269e3021894 203 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 204 "addxl %%d4, %%d3 \n\t"
elessair 0:f269e3021894 205
elessair 0:f269e3021894 206 #define MULADDC_STOP \
elessair 0:f269e3021894 207 "movl %%d3, %0 \n\t" \
elessair 0:f269e3021894 208 "movl %%a3, %1 \n\t" \
elessair 0:f269e3021894 209 "movl %%a2, %2 \n\t" \
elessair 0:f269e3021894 210 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 211 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 212 : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \
elessair 0:f269e3021894 213 );
elessair 0:f269e3021894 214
elessair 0:f269e3021894 215 #define MULADDC_HUIT \
elessair 0:f269e3021894 216 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 217 "mulul %%d2, %%d4:%%d1 \n\t" \
elessair 0:f269e3021894 218 "addxl %%d3, %%d1 \n\t" \
elessair 0:f269e3021894 219 "addxl %%d0, %%d4 \n\t" \
elessair 0:f269e3021894 220 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 221 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 222 "mulul %%d2, %%d3:%%d1 \n\t" \
elessair 0:f269e3021894 223 "addxl %%d4, %%d1 \n\t" \
elessair 0:f269e3021894 224 "addxl %%d0, %%d3 \n\t" \
elessair 0:f269e3021894 225 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 226 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 227 "mulul %%d2, %%d4:%%d1 \n\t" \
elessair 0:f269e3021894 228 "addxl %%d3, %%d1 \n\t" \
elessair 0:f269e3021894 229 "addxl %%d0, %%d4 \n\t" \
elessair 0:f269e3021894 230 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 231 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 232 "mulul %%d2, %%d3:%%d1 \n\t" \
elessair 0:f269e3021894 233 "addxl %%d4, %%d1 \n\t" \
elessair 0:f269e3021894 234 "addxl %%d0, %%d3 \n\t" \
elessair 0:f269e3021894 235 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 236 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 237 "mulul %%d2, %%d4:%%d1 \n\t" \
elessair 0:f269e3021894 238 "addxl %%d3, %%d1 \n\t" \
elessair 0:f269e3021894 239 "addxl %%d0, %%d4 \n\t" \
elessair 0:f269e3021894 240 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 241 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 242 "mulul %%d2, %%d3:%%d1 \n\t" \
elessair 0:f269e3021894 243 "addxl %%d4, %%d1 \n\t" \
elessair 0:f269e3021894 244 "addxl %%d0, %%d3 \n\t" \
elessair 0:f269e3021894 245 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 246 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 247 "mulul %%d2, %%d4:%%d1 \n\t" \
elessair 0:f269e3021894 248 "addxl %%d3, %%d1 \n\t" \
elessair 0:f269e3021894 249 "addxl %%d0, %%d4 \n\t" \
elessair 0:f269e3021894 250 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 251 "movel %%a2@+, %%d1 \n\t" \
elessair 0:f269e3021894 252 "mulul %%d2, %%d3:%%d1 \n\t" \
elessair 0:f269e3021894 253 "addxl %%d4, %%d1 \n\t" \
elessair 0:f269e3021894 254 "addxl %%d0, %%d3 \n\t" \
elessair 0:f269e3021894 255 "addl %%d1, %%a3@+ \n\t" \
elessair 0:f269e3021894 256 "addxl %%d0, %%d3 \n\t"
elessair 0:f269e3021894 257
elessair 0:f269e3021894 258 #endif /* MC68000 */
elessair 0:f269e3021894 259
elessair 0:f269e3021894 260 #if defined(__powerpc64__) || defined(__ppc64__)
elessair 0:f269e3021894 261
elessair 0:f269e3021894 262 #if defined(__MACH__) && defined(__APPLE__)
elessair 0:f269e3021894 263
elessair 0:f269e3021894 264 #define MULADDC_INIT \
elessair 0:f269e3021894 265 asm( \
elessair 0:f269e3021894 266 "ld r3, %3 \n\t" \
elessair 0:f269e3021894 267 "ld r4, %4 \n\t" \
elessair 0:f269e3021894 268 "ld r5, %5 \n\t" \
elessair 0:f269e3021894 269 "ld r6, %6 \n\t" \
elessair 0:f269e3021894 270 "addi r3, r3, -8 \n\t" \
elessair 0:f269e3021894 271 "addi r4, r4, -8 \n\t" \
elessair 0:f269e3021894 272 "addic r5, r5, 0 \n\t"
elessair 0:f269e3021894 273
elessair 0:f269e3021894 274 #define MULADDC_CORE \
elessair 0:f269e3021894 275 "ldu r7, 8(r3) \n\t" \
elessair 0:f269e3021894 276 "mulld r8, r7, r6 \n\t" \
elessair 0:f269e3021894 277 "mulhdu r9, r7, r6 \n\t" \
elessair 0:f269e3021894 278 "adde r8, r8, r5 \n\t" \
elessair 0:f269e3021894 279 "ld r7, 8(r4) \n\t" \
elessair 0:f269e3021894 280 "addze r5, r9 \n\t" \
elessair 0:f269e3021894 281 "addc r8, r8, r7 \n\t" \
elessair 0:f269e3021894 282 "stdu r8, 8(r4) \n\t"
elessair 0:f269e3021894 283
elessair 0:f269e3021894 284 #define MULADDC_STOP \
elessair 0:f269e3021894 285 "addze r5, r5 \n\t" \
elessair 0:f269e3021894 286 "addi r4, r4, 8 \n\t" \
elessair 0:f269e3021894 287 "addi r3, r3, 8 \n\t" \
elessair 0:f269e3021894 288 "std r5, %0 \n\t" \
elessair 0:f269e3021894 289 "std r4, %1 \n\t" \
elessair 0:f269e3021894 290 "std r3, %2 \n\t" \
elessair 0:f269e3021894 291 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 292 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 293 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
elessair 0:f269e3021894 294 );
elessair 0:f269e3021894 295
elessair 0:f269e3021894 296
elessair 0:f269e3021894 297 #else /* __MACH__ && __APPLE__ */
elessair 0:f269e3021894 298
elessair 0:f269e3021894 299 #define MULADDC_INIT \
elessair 0:f269e3021894 300 asm( \
elessair 0:f269e3021894 301 "ld %%r3, %3 \n\t" \
elessair 0:f269e3021894 302 "ld %%r4, %4 \n\t" \
elessair 0:f269e3021894 303 "ld %%r5, %5 \n\t" \
elessair 0:f269e3021894 304 "ld %%r6, %6 \n\t" \
elessair 0:f269e3021894 305 "addi %%r3, %%r3, -8 \n\t" \
elessair 0:f269e3021894 306 "addi %%r4, %%r4, -8 \n\t" \
elessair 0:f269e3021894 307 "addic %%r5, %%r5, 0 \n\t"
elessair 0:f269e3021894 308
elessair 0:f269e3021894 309 #define MULADDC_CORE \
elessair 0:f269e3021894 310 "ldu %%r7, 8(%%r3) \n\t" \
elessair 0:f269e3021894 311 "mulld %%r8, %%r7, %%r6 \n\t" \
elessair 0:f269e3021894 312 "mulhdu %%r9, %%r7, %%r6 \n\t" \
elessair 0:f269e3021894 313 "adde %%r8, %%r8, %%r5 \n\t" \
elessair 0:f269e3021894 314 "ld %%r7, 8(%%r4) \n\t" \
elessair 0:f269e3021894 315 "addze %%r5, %%r9 \n\t" \
elessair 0:f269e3021894 316 "addc %%r8, %%r8, %%r7 \n\t" \
elessair 0:f269e3021894 317 "stdu %%r8, 8(%%r4) \n\t"
elessair 0:f269e3021894 318
elessair 0:f269e3021894 319 #define MULADDC_STOP \
elessair 0:f269e3021894 320 "addze %%r5, %%r5 \n\t" \
elessair 0:f269e3021894 321 "addi %%r4, %%r4, 8 \n\t" \
elessair 0:f269e3021894 322 "addi %%r3, %%r3, 8 \n\t" \
elessair 0:f269e3021894 323 "std %%r5, %0 \n\t" \
elessair 0:f269e3021894 324 "std %%r4, %1 \n\t" \
elessair 0:f269e3021894 325 "std %%r3, %2 \n\t" \
elessair 0:f269e3021894 326 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 327 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 328 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
elessair 0:f269e3021894 329 );
elessair 0:f269e3021894 330
elessair 0:f269e3021894 331 #endif /* __MACH__ && __APPLE__ */
elessair 0:f269e3021894 332
elessair 0:f269e3021894 333 #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */
elessair 0:f269e3021894 334
elessair 0:f269e3021894 335 #if defined(__MACH__) && defined(__APPLE__)
elessair 0:f269e3021894 336
elessair 0:f269e3021894 337 #define MULADDC_INIT \
elessair 0:f269e3021894 338 asm( \
elessair 0:f269e3021894 339 "lwz r3, %3 \n\t" \
elessair 0:f269e3021894 340 "lwz r4, %4 \n\t" \
elessair 0:f269e3021894 341 "lwz r5, %5 \n\t" \
elessair 0:f269e3021894 342 "lwz r6, %6 \n\t" \
elessair 0:f269e3021894 343 "addi r3, r3, -4 \n\t" \
elessair 0:f269e3021894 344 "addi r4, r4, -4 \n\t" \
elessair 0:f269e3021894 345 "addic r5, r5, 0 \n\t"
elessair 0:f269e3021894 346
elessair 0:f269e3021894 347 #define MULADDC_CORE \
elessair 0:f269e3021894 348 "lwzu r7, 4(r3) \n\t" \
elessair 0:f269e3021894 349 "mullw r8, r7, r6 \n\t" \
elessair 0:f269e3021894 350 "mulhwu r9, r7, r6 \n\t" \
elessair 0:f269e3021894 351 "adde r8, r8, r5 \n\t" \
elessair 0:f269e3021894 352 "lwz r7, 4(r4) \n\t" \
elessair 0:f269e3021894 353 "addze r5, r9 \n\t" \
elessair 0:f269e3021894 354 "addc r8, r8, r7 \n\t" \
elessair 0:f269e3021894 355 "stwu r8, 4(r4) \n\t"
elessair 0:f269e3021894 356
elessair 0:f269e3021894 357 #define MULADDC_STOP \
elessair 0:f269e3021894 358 "addze r5, r5 \n\t" \
elessair 0:f269e3021894 359 "addi r4, r4, 4 \n\t" \
elessair 0:f269e3021894 360 "addi r3, r3, 4 \n\t" \
elessair 0:f269e3021894 361 "stw r5, %0 \n\t" \
elessair 0:f269e3021894 362 "stw r4, %1 \n\t" \
elessair 0:f269e3021894 363 "stw r3, %2 \n\t" \
elessair 0:f269e3021894 364 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 365 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 366 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
elessair 0:f269e3021894 367 );
elessair 0:f269e3021894 368
elessair 0:f269e3021894 369 #else /* __MACH__ && __APPLE__ */
elessair 0:f269e3021894 370
elessair 0:f269e3021894 371 #define MULADDC_INIT \
elessair 0:f269e3021894 372 asm( \
elessair 0:f269e3021894 373 "lwz %%r3, %3 \n\t" \
elessair 0:f269e3021894 374 "lwz %%r4, %4 \n\t" \
elessair 0:f269e3021894 375 "lwz %%r5, %5 \n\t" \
elessair 0:f269e3021894 376 "lwz %%r6, %6 \n\t" \
elessair 0:f269e3021894 377 "addi %%r3, %%r3, -4 \n\t" \
elessair 0:f269e3021894 378 "addi %%r4, %%r4, -4 \n\t" \
elessair 0:f269e3021894 379 "addic %%r5, %%r5, 0 \n\t"
elessair 0:f269e3021894 380
elessair 0:f269e3021894 381 #define MULADDC_CORE \
elessair 0:f269e3021894 382 "lwzu %%r7, 4(%%r3) \n\t" \
elessair 0:f269e3021894 383 "mullw %%r8, %%r7, %%r6 \n\t" \
elessair 0:f269e3021894 384 "mulhwu %%r9, %%r7, %%r6 \n\t" \
elessair 0:f269e3021894 385 "adde %%r8, %%r8, %%r5 \n\t" \
elessair 0:f269e3021894 386 "lwz %%r7, 4(%%r4) \n\t" \
elessair 0:f269e3021894 387 "addze %%r5, %%r9 \n\t" \
elessair 0:f269e3021894 388 "addc %%r8, %%r8, %%r7 \n\t" \
elessair 0:f269e3021894 389 "stwu %%r8, 4(%%r4) \n\t"
elessair 0:f269e3021894 390
elessair 0:f269e3021894 391 #define MULADDC_STOP \
elessair 0:f269e3021894 392 "addze %%r5, %%r5 \n\t" \
elessair 0:f269e3021894 393 "addi %%r4, %%r4, 4 \n\t" \
elessair 0:f269e3021894 394 "addi %%r3, %%r3, 4 \n\t" \
elessair 0:f269e3021894 395 "stw %%r5, %0 \n\t" \
elessair 0:f269e3021894 396 "stw %%r4, %1 \n\t" \
elessair 0:f269e3021894 397 "stw %%r3, %2 \n\t" \
elessair 0:f269e3021894 398 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 399 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 400 : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \
elessair 0:f269e3021894 401 );
elessair 0:f269e3021894 402
elessair 0:f269e3021894 403 #endif /* __MACH__ && __APPLE__ */
elessair 0:f269e3021894 404
elessair 0:f269e3021894 405 #endif /* PPC32 */
elessair 0:f269e3021894 406
elessair 0:f269e3021894 407 /*
elessair 0:f269e3021894 408 * The Sparc(64) assembly is reported to be broken.
elessair 0:f269e3021894 409 * Disable it for now, until we're able to fix it.
elessair 0:f269e3021894 410 */
elessair 0:f269e3021894 411 #if 0 && defined(__sparc__)
elessair 0:f269e3021894 412 #if defined(__sparc64__)
elessair 0:f269e3021894 413
elessair 0:f269e3021894 414 #define MULADDC_INIT \
elessair 0:f269e3021894 415 asm( \
elessair 0:f269e3021894 416 "ldx %3, %%o0 \n\t" \
elessair 0:f269e3021894 417 "ldx %4, %%o1 \n\t" \
elessair 0:f269e3021894 418 "ld %5, %%o2 \n\t" \
elessair 0:f269e3021894 419 "ld %6, %%o3 \n\t"
elessair 0:f269e3021894 420
elessair 0:f269e3021894 421 #define MULADDC_CORE \
elessair 0:f269e3021894 422 "ld [%%o0], %%o4 \n\t" \
elessair 0:f269e3021894 423 "inc 4, %%o0 \n\t" \
elessair 0:f269e3021894 424 "ld [%%o1], %%o5 \n\t" \
elessair 0:f269e3021894 425 "umul %%o3, %%o4, %%o4 \n\t" \
elessair 0:f269e3021894 426 "addcc %%o4, %%o2, %%o4 \n\t" \
elessair 0:f269e3021894 427 "rd %%y, %%g1 \n\t" \
elessair 0:f269e3021894 428 "addx %%g1, 0, %%g1 \n\t" \
elessair 0:f269e3021894 429 "addcc %%o4, %%o5, %%o4 \n\t" \
elessair 0:f269e3021894 430 "st %%o4, [%%o1] \n\t" \
elessair 0:f269e3021894 431 "addx %%g1, 0, %%o2 \n\t" \
elessair 0:f269e3021894 432 "inc 4, %%o1 \n\t"
elessair 0:f269e3021894 433
elessair 0:f269e3021894 434 #define MULADDC_STOP \
elessair 0:f269e3021894 435 "st %%o2, %0 \n\t" \
elessair 0:f269e3021894 436 "stx %%o1, %1 \n\t" \
elessair 0:f269e3021894 437 "stx %%o0, %2 \n\t" \
elessair 0:f269e3021894 438 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 439 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 440 : "g1", "o0", "o1", "o2", "o3", "o4", \
elessair 0:f269e3021894 441 "o5" \
elessair 0:f269e3021894 442 );
elessair 0:f269e3021894 443
elessair 0:f269e3021894 444 #else /* __sparc64__ */
elessair 0:f269e3021894 445
elessair 0:f269e3021894 446 #define MULADDC_INIT \
elessair 0:f269e3021894 447 asm( \
elessair 0:f269e3021894 448 "ld %3, %%o0 \n\t" \
elessair 0:f269e3021894 449 "ld %4, %%o1 \n\t" \
elessair 0:f269e3021894 450 "ld %5, %%o2 \n\t" \
elessair 0:f269e3021894 451 "ld %6, %%o3 \n\t"
elessair 0:f269e3021894 452
elessair 0:f269e3021894 453 #define MULADDC_CORE \
elessair 0:f269e3021894 454 "ld [%%o0], %%o4 \n\t" \
elessair 0:f269e3021894 455 "inc 4, %%o0 \n\t" \
elessair 0:f269e3021894 456 "ld [%%o1], %%o5 \n\t" \
elessair 0:f269e3021894 457 "umul %%o3, %%o4, %%o4 \n\t" \
elessair 0:f269e3021894 458 "addcc %%o4, %%o2, %%o4 \n\t" \
elessair 0:f269e3021894 459 "rd %%y, %%g1 \n\t" \
elessair 0:f269e3021894 460 "addx %%g1, 0, %%g1 \n\t" \
elessair 0:f269e3021894 461 "addcc %%o4, %%o5, %%o4 \n\t" \
elessair 0:f269e3021894 462 "st %%o4, [%%o1] \n\t" \
elessair 0:f269e3021894 463 "addx %%g1, 0, %%o2 \n\t" \
elessair 0:f269e3021894 464 "inc 4, %%o1 \n\t"
elessair 0:f269e3021894 465
elessair 0:f269e3021894 466 #define MULADDC_STOP \
elessair 0:f269e3021894 467 "st %%o2, %0 \n\t" \
elessair 0:f269e3021894 468 "st %%o1, %1 \n\t" \
elessair 0:f269e3021894 469 "st %%o0, %2 \n\t" \
elessair 0:f269e3021894 470 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 471 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 472 : "g1", "o0", "o1", "o2", "o3", "o4", \
elessair 0:f269e3021894 473 "o5" \
elessair 0:f269e3021894 474 );
elessair 0:f269e3021894 475
elessair 0:f269e3021894 476 #endif /* __sparc64__ */
elessair 0:f269e3021894 477 #endif /* __sparc__ */
elessair 0:f269e3021894 478
elessair 0:f269e3021894 479 #if defined(__microblaze__) || defined(microblaze)
elessair 0:f269e3021894 480
elessair 0:f269e3021894 481 #define MULADDC_INIT \
elessair 0:f269e3021894 482 asm( \
elessair 0:f269e3021894 483 "lwi r3, %3 \n\t" \
elessair 0:f269e3021894 484 "lwi r4, %4 \n\t" \
elessair 0:f269e3021894 485 "lwi r5, %5 \n\t" \
elessair 0:f269e3021894 486 "lwi r6, %6 \n\t" \
elessair 0:f269e3021894 487 "andi r7, r6, 0xffff \n\t" \
elessair 0:f269e3021894 488 "bsrli r6, r6, 16 \n\t"
elessair 0:f269e3021894 489
elessair 0:f269e3021894 490 #define MULADDC_CORE \
elessair 0:f269e3021894 491 "lhui r8, r3, 0 \n\t" \
elessair 0:f269e3021894 492 "addi r3, r3, 2 \n\t" \
elessair 0:f269e3021894 493 "lhui r9, r3, 0 \n\t" \
elessair 0:f269e3021894 494 "addi r3, r3, 2 \n\t" \
elessair 0:f269e3021894 495 "mul r10, r9, r6 \n\t" \
elessair 0:f269e3021894 496 "mul r11, r8, r7 \n\t" \
elessair 0:f269e3021894 497 "mul r12, r9, r7 \n\t" \
elessair 0:f269e3021894 498 "mul r13, r8, r6 \n\t" \
elessair 0:f269e3021894 499 "bsrli r8, r10, 16 \n\t" \
elessair 0:f269e3021894 500 "bsrli r9, r11, 16 \n\t" \
elessair 0:f269e3021894 501 "add r13, r13, r8 \n\t" \
elessair 0:f269e3021894 502 "add r13, r13, r9 \n\t" \
elessair 0:f269e3021894 503 "bslli r10, r10, 16 \n\t" \
elessair 0:f269e3021894 504 "bslli r11, r11, 16 \n\t" \
elessair 0:f269e3021894 505 "add r12, r12, r10 \n\t" \
elessair 0:f269e3021894 506 "addc r13, r13, r0 \n\t" \
elessair 0:f269e3021894 507 "add r12, r12, r11 \n\t" \
elessair 0:f269e3021894 508 "addc r13, r13, r0 \n\t" \
elessair 0:f269e3021894 509 "lwi r10, r4, 0 \n\t" \
elessair 0:f269e3021894 510 "add r12, r12, r10 \n\t" \
elessair 0:f269e3021894 511 "addc r13, r13, r0 \n\t" \
elessair 0:f269e3021894 512 "add r12, r12, r5 \n\t" \
elessair 0:f269e3021894 513 "addc r5, r13, r0 \n\t" \
elessair 0:f269e3021894 514 "swi r12, r4, 0 \n\t" \
elessair 0:f269e3021894 515 "addi r4, r4, 4 \n\t"
elessair 0:f269e3021894 516
elessair 0:f269e3021894 517 #define MULADDC_STOP \
elessair 0:f269e3021894 518 "swi r5, %0 \n\t" \
elessair 0:f269e3021894 519 "swi r4, %1 \n\t" \
elessair 0:f269e3021894 520 "swi r3, %2 \n\t" \
elessair 0:f269e3021894 521 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 522 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 523 : "r3", "r4" "r5", "r6", "r7", "r8", \
elessair 0:f269e3021894 524 "r9", "r10", "r11", "r12", "r13" \
elessair 0:f269e3021894 525 );
elessair 0:f269e3021894 526
elessair 0:f269e3021894 527 #endif /* MicroBlaze */
elessair 0:f269e3021894 528
elessair 0:f269e3021894 529 #if defined(__tricore__)
elessair 0:f269e3021894 530
elessair 0:f269e3021894 531 #define MULADDC_INIT \
elessair 0:f269e3021894 532 asm( \
elessair 0:f269e3021894 533 "ld.a %%a2, %3 \n\t" \
elessair 0:f269e3021894 534 "ld.a %%a3, %4 \n\t" \
elessair 0:f269e3021894 535 "ld.w %%d4, %5 \n\t" \
elessair 0:f269e3021894 536 "ld.w %%d1, %6 \n\t" \
elessair 0:f269e3021894 537 "xor %%d5, %%d5 \n\t"
elessair 0:f269e3021894 538
elessair 0:f269e3021894 539 #define MULADDC_CORE \
elessair 0:f269e3021894 540 "ld.w %%d0, [%%a2+] \n\t" \
elessair 0:f269e3021894 541 "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \
elessair 0:f269e3021894 542 "ld.w %%d0, [%%a3] \n\t" \
elessair 0:f269e3021894 543 "addx %%d2, %%d2, %%d0 \n\t" \
elessair 0:f269e3021894 544 "addc %%d3, %%d3, 0 \n\t" \
elessair 0:f269e3021894 545 "mov %%d4, %%d3 \n\t" \
elessair 0:f269e3021894 546 "st.w [%%a3+], %%d2 \n\t"
elessair 0:f269e3021894 547
elessair 0:f269e3021894 548 #define MULADDC_STOP \
elessair 0:f269e3021894 549 "st.w %0, %%d4 \n\t" \
elessair 0:f269e3021894 550 "st.a %1, %%a3 \n\t" \
elessair 0:f269e3021894 551 "st.a %2, %%a2 \n\t" \
elessair 0:f269e3021894 552 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 553 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 554 : "d0", "d1", "e2", "d4", "a2", "a3" \
elessair 0:f269e3021894 555 );
elessair 0:f269e3021894 556
elessair 0:f269e3021894 557 #endif /* TriCore */
elessair 0:f269e3021894 558
elessair 0:f269e3021894 559 /*
elessair 0:f269e3021894 560 * gcc -O0 by default uses r7 for the frame pointer, so it complains about our
elessair 0:f269e3021894 561 * use of r7 below, unless -fomit-frame-pointer is passed. Unfortunately,
elessair 0:f269e3021894 562 * passing that option is not easy when building with yotta.
elessair 0:f269e3021894 563 *
elessair 0:f269e3021894 564 * On the other hand, -fomit-frame-pointer is implied by any -Ox options with
elessair 0:f269e3021894 565 * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by
elessair 0:f269e3021894 566 * clang and armcc5 under the same conditions).
elessair 0:f269e3021894 567 *
elessair 0:f269e3021894 568 * So, only use the optimized assembly below for optimized build, which avoids
elessair 0:f269e3021894 569 * the build error and is pretty reasonable anyway.
elessair 0:f269e3021894 570 */
elessair 0:f269e3021894 571 #if defined(__GNUC__) && !defined(__OPTIMIZE__)
elessair 0:f269e3021894 572 #define MULADDC_CANNOT_USE_R7
elessair 0:f269e3021894 573 #endif
elessair 0:f269e3021894 574
elessair 0:f269e3021894 575 #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7)
elessair 0:f269e3021894 576
elessair 0:f269e3021894 577 #if defined(__thumb__) && !defined(__thumb2__)
elessair 0:f269e3021894 578
elessair 0:f269e3021894 579 #define MULADDC_INIT \
elessair 0:f269e3021894 580 asm( \
elessair 0:f269e3021894 581 "ldr r0, %3 \n\t" \
elessair 0:f269e3021894 582 "ldr r1, %4 \n\t" \
elessair 0:f269e3021894 583 "ldr r2, %5 \n\t" \
elessair 0:f269e3021894 584 "ldr r3, %6 \n\t" \
elessair 0:f269e3021894 585 "lsr r7, r3, #16 \n\t" \
elessair 0:f269e3021894 586 "mov r9, r7 \n\t" \
elessair 0:f269e3021894 587 "lsl r7, r3, #16 \n\t" \
elessair 0:f269e3021894 588 "lsr r7, r7, #16 \n\t" \
elessair 0:f269e3021894 589 "mov r8, r7 \n\t"
elessair 0:f269e3021894 590
elessair 0:f269e3021894 591 #define MULADDC_CORE \
elessair 0:f269e3021894 592 "ldmia r0!, {r6} \n\t" \
elessair 0:f269e3021894 593 "lsr r7, r6, #16 \n\t" \
elessair 0:f269e3021894 594 "lsl r6, r6, #16 \n\t" \
elessair 0:f269e3021894 595 "lsr r6, r6, #16 \n\t" \
elessair 0:f269e3021894 596 "mov r4, r8 \n\t" \
elessair 0:f269e3021894 597 "mul r4, r6 \n\t" \
elessair 0:f269e3021894 598 "mov r3, r9 \n\t" \
elessair 0:f269e3021894 599 "mul r6, r3 \n\t" \
elessair 0:f269e3021894 600 "mov r5, r9 \n\t" \
elessair 0:f269e3021894 601 "mul r5, r7 \n\t" \
elessair 0:f269e3021894 602 "mov r3, r8 \n\t" \
elessair 0:f269e3021894 603 "mul r7, r3 \n\t" \
elessair 0:f269e3021894 604 "lsr r3, r6, #16 \n\t" \
elessair 0:f269e3021894 605 "add r5, r5, r3 \n\t" \
elessair 0:f269e3021894 606 "lsr r3, r7, #16 \n\t" \
elessair 0:f269e3021894 607 "add r5, r5, r3 \n\t" \
elessair 0:f269e3021894 608 "add r4, r4, r2 \n\t" \
elessair 0:f269e3021894 609 "mov r2, #0 \n\t" \
elessair 0:f269e3021894 610 "adc r5, r2 \n\t" \
elessair 0:f269e3021894 611 "lsl r3, r6, #16 \n\t" \
elessair 0:f269e3021894 612 "add r4, r4, r3 \n\t" \
elessair 0:f269e3021894 613 "adc r5, r2 \n\t" \
elessair 0:f269e3021894 614 "lsl r3, r7, #16 \n\t" \
elessair 0:f269e3021894 615 "add r4, r4, r3 \n\t" \
elessair 0:f269e3021894 616 "adc r5, r2 \n\t" \
elessair 0:f269e3021894 617 "ldr r3, [r1] \n\t" \
elessair 0:f269e3021894 618 "add r4, r4, r3 \n\t" \
elessair 0:f269e3021894 619 "adc r2, r5 \n\t" \
elessair 0:f269e3021894 620 "stmia r1!, {r4} \n\t"
elessair 0:f269e3021894 621
elessair 0:f269e3021894 622 #define MULADDC_STOP \
elessair 0:f269e3021894 623 "str r2, %0 \n\t" \
elessair 0:f269e3021894 624 "str r1, %1 \n\t" \
elessair 0:f269e3021894 625 "str r0, %2 \n\t" \
elessair 0:f269e3021894 626 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 627 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 628 : "r0", "r1", "r2", "r3", "r4", "r5", \
elessair 0:f269e3021894 629 "r6", "r7", "r8", "r9", "cc" \
elessair 0:f269e3021894 630 );
elessair 0:f269e3021894 631
elessair 0:f269e3021894 632 #else
elessair 0:f269e3021894 633
elessair 0:f269e3021894 634 #define MULADDC_INIT \
elessair 0:f269e3021894 635 asm( \
elessair 0:f269e3021894 636 "ldr r0, %3 \n\t" \
elessair 0:f269e3021894 637 "ldr r1, %4 \n\t" \
elessair 0:f269e3021894 638 "ldr r2, %5 \n\t" \
elessair 0:f269e3021894 639 "ldr r3, %6 \n\t"
elessair 0:f269e3021894 640
elessair 0:f269e3021894 641 #define MULADDC_CORE \
elessair 0:f269e3021894 642 "ldr r4, [r0], #4 \n\t" \
elessair 0:f269e3021894 643 "mov r5, #0 \n\t" \
elessair 0:f269e3021894 644 "ldr r6, [r1] \n\t" \
elessair 0:f269e3021894 645 "umlal r2, r5, r3, r4 \n\t" \
elessair 0:f269e3021894 646 "adds r7, r6, r2 \n\t" \
elessair 0:f269e3021894 647 "adc r2, r5, #0 \n\t" \
elessair 0:f269e3021894 648 "str r7, [r1], #4 \n\t"
elessair 0:f269e3021894 649
elessair 0:f269e3021894 650 #define MULADDC_STOP \
elessair 0:f269e3021894 651 "str r2, %0 \n\t" \
elessair 0:f269e3021894 652 "str r1, %1 \n\t" \
elessair 0:f269e3021894 653 "str r0, %2 \n\t" \
elessair 0:f269e3021894 654 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 655 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 656 : "r0", "r1", "r2", "r3", "r4", "r5", \
elessair 0:f269e3021894 657 "r6", "r7", "cc" \
elessair 0:f269e3021894 658 );
elessair 0:f269e3021894 659
elessair 0:f269e3021894 660 #endif /* Thumb */
elessair 0:f269e3021894 661
elessair 0:f269e3021894 662 #endif /* ARMv3 */
elessair 0:f269e3021894 663
elessair 0:f269e3021894 664 #if defined(__alpha__)
elessair 0:f269e3021894 665
elessair 0:f269e3021894 666 #define MULADDC_INIT \
elessair 0:f269e3021894 667 asm( \
elessair 0:f269e3021894 668 "ldq $1, %3 \n\t" \
elessair 0:f269e3021894 669 "ldq $2, %4 \n\t" \
elessair 0:f269e3021894 670 "ldq $3, %5 \n\t" \
elessair 0:f269e3021894 671 "ldq $4, %6 \n\t"
elessair 0:f269e3021894 672
elessair 0:f269e3021894 673 #define MULADDC_CORE \
elessair 0:f269e3021894 674 "ldq $6, 0($1) \n\t" \
elessair 0:f269e3021894 675 "addq $1, 8, $1 \n\t" \
elessair 0:f269e3021894 676 "mulq $6, $4, $7 \n\t" \
elessair 0:f269e3021894 677 "umulh $6, $4, $6 \n\t" \
elessair 0:f269e3021894 678 "addq $7, $3, $7 \n\t" \
elessair 0:f269e3021894 679 "cmpult $7, $3, $3 \n\t" \
elessair 0:f269e3021894 680 "ldq $5, 0($2) \n\t" \
elessair 0:f269e3021894 681 "addq $7, $5, $7 \n\t" \
elessair 0:f269e3021894 682 "cmpult $7, $5, $5 \n\t" \
elessair 0:f269e3021894 683 "stq $7, 0($2) \n\t" \
elessair 0:f269e3021894 684 "addq $2, 8, $2 \n\t" \
elessair 0:f269e3021894 685 "addq $6, $3, $3 \n\t" \
elessair 0:f269e3021894 686 "addq $5, $3, $3 \n\t"
elessair 0:f269e3021894 687
elessair 0:f269e3021894 688 #define MULADDC_STOP \
elessair 0:f269e3021894 689 "stq $3, %0 \n\t" \
elessair 0:f269e3021894 690 "stq $2, %1 \n\t" \
elessair 0:f269e3021894 691 "stq $1, %2 \n\t" \
elessair 0:f269e3021894 692 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 693 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 694 : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \
elessair 0:f269e3021894 695 );
elessair 0:f269e3021894 696 #endif /* Alpha */
elessair 0:f269e3021894 697
elessair 0:f269e3021894 698 #if defined(__mips__) && !defined(__mips64)
elessair 0:f269e3021894 699
elessair 0:f269e3021894 700 #define MULADDC_INIT \
elessair 0:f269e3021894 701 asm( \
elessair 0:f269e3021894 702 "lw $10, %3 \n\t" \
elessair 0:f269e3021894 703 "lw $11, %4 \n\t" \
elessair 0:f269e3021894 704 "lw $12, %5 \n\t" \
elessair 0:f269e3021894 705 "lw $13, %6 \n\t"
elessair 0:f269e3021894 706
elessair 0:f269e3021894 707 #define MULADDC_CORE \
elessair 0:f269e3021894 708 "lw $14, 0($10) \n\t" \
elessair 0:f269e3021894 709 "multu $13, $14 \n\t" \
elessair 0:f269e3021894 710 "addi $10, $10, 4 \n\t" \
elessair 0:f269e3021894 711 "mflo $14 \n\t" \
elessair 0:f269e3021894 712 "mfhi $9 \n\t" \
elessair 0:f269e3021894 713 "addu $14, $12, $14 \n\t" \
elessair 0:f269e3021894 714 "lw $15, 0($11) \n\t" \
elessair 0:f269e3021894 715 "sltu $12, $14, $12 \n\t" \
elessair 0:f269e3021894 716 "addu $15, $14, $15 \n\t" \
elessair 0:f269e3021894 717 "sltu $14, $15, $14 \n\t" \
elessair 0:f269e3021894 718 "addu $12, $12, $9 \n\t" \
elessair 0:f269e3021894 719 "sw $15, 0($11) \n\t" \
elessair 0:f269e3021894 720 "addu $12, $12, $14 \n\t" \
elessair 0:f269e3021894 721 "addi $11, $11, 4 \n\t"
elessair 0:f269e3021894 722
elessair 0:f269e3021894 723 #define MULADDC_STOP \
elessair 0:f269e3021894 724 "sw $12, %0 \n\t" \
elessair 0:f269e3021894 725 "sw $11, %1 \n\t" \
elessair 0:f269e3021894 726 "sw $10, %2 \n\t" \
elessair 0:f269e3021894 727 : "=m" (c), "=m" (d), "=m" (s) \
elessair 0:f269e3021894 728 : "m" (s), "m" (d), "m" (c), "m" (b) \
elessair 0:f269e3021894 729 : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \
elessair 0:f269e3021894 730 );
elessair 0:f269e3021894 731
elessair 0:f269e3021894 732 #endif /* MIPS */
elessair 0:f269e3021894 733 #endif /* GNUC */
elessair 0:f269e3021894 734
elessair 0:f269e3021894 735 #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__)
elessair 0:f269e3021894 736
elessair 0:f269e3021894 737 #define MULADDC_INIT \
elessair 0:f269e3021894 738 __asm mov esi, s \
elessair 0:f269e3021894 739 __asm mov edi, d \
elessair 0:f269e3021894 740 __asm mov ecx, c \
elessair 0:f269e3021894 741 __asm mov ebx, b
elessair 0:f269e3021894 742
elessair 0:f269e3021894 743 #define MULADDC_CORE \
elessair 0:f269e3021894 744 __asm lodsd \
elessair 0:f269e3021894 745 __asm mul ebx \
elessair 0:f269e3021894 746 __asm add eax, ecx \
elessair 0:f269e3021894 747 __asm adc edx, 0 \
elessair 0:f269e3021894 748 __asm add eax, [edi] \
elessair 0:f269e3021894 749 __asm adc edx, 0 \
elessair 0:f269e3021894 750 __asm mov ecx, edx \
elessair 0:f269e3021894 751 __asm stosd
elessair 0:f269e3021894 752
elessair 0:f269e3021894 753 #if defined(MBEDTLS_HAVE_SSE2)
elessair 0:f269e3021894 754
elessair 0:f269e3021894 755 #define EMIT __asm _emit
elessair 0:f269e3021894 756
elessair 0:f269e3021894 757 #define MULADDC_HUIT \
elessair 0:f269e3021894 758 EMIT 0x0F EMIT 0x6E EMIT 0xC9 \
elessair 0:f269e3021894 759 EMIT 0x0F EMIT 0x6E EMIT 0xC3 \
elessair 0:f269e3021894 760 EMIT 0x0F EMIT 0x6E EMIT 0x1F \
elessair 0:f269e3021894 761 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
elessair 0:f269e3021894 762 EMIT 0x0F EMIT 0x6E EMIT 0x16 \
elessair 0:f269e3021894 763 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
elessair 0:f269e3021894 764 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \
elessair 0:f269e3021894 765 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
elessair 0:f269e3021894 766 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \
elessair 0:f269e3021894 767 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
elessair 0:f269e3021894 768 EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \
elessair 0:f269e3021894 769 EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \
elessair 0:f269e3021894 770 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
elessair 0:f269e3021894 771 EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \
elessair 0:f269e3021894 772 EMIT 0x0F EMIT 0xD4 EMIT 0xDC \
elessair 0:f269e3021894 773 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \
elessair 0:f269e3021894 774 EMIT 0x0F EMIT 0xD4 EMIT 0xEE \
elessair 0:f269e3021894 775 EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \
elessair 0:f269e3021894 776 EMIT 0x0F EMIT 0xD4 EMIT 0xFC \
elessair 0:f269e3021894 777 EMIT 0x0F EMIT 0x7E EMIT 0x0F \
elessair 0:f269e3021894 778 EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \
elessair 0:f269e3021894 779 EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \
elessair 0:f269e3021894 780 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 781 EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \
elessair 0:f269e3021894 782 EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \
elessair 0:f269e3021894 783 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
elessair 0:f269e3021894 784 EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \
elessair 0:f269e3021894 785 EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \
elessair 0:f269e3021894 786 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \
elessair 0:f269e3021894 787 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 788 EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \
elessair 0:f269e3021894 789 EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \
elessair 0:f269e3021894 790 EMIT 0x0F EMIT 0xD4 EMIT 0xCD \
elessair 0:f269e3021894 791 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \
elessair 0:f269e3021894 792 EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \
elessair 0:f269e3021894 793 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \
elessair 0:f269e3021894 794 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 795 EMIT 0x0F EMIT 0xD4 EMIT 0xCF \
elessair 0:f269e3021894 796 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \
elessair 0:f269e3021894 797 EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \
elessair 0:f269e3021894 798 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \
elessair 0:f269e3021894 799 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 800 EMIT 0x0F EMIT 0xD4 EMIT 0xCA \
elessair 0:f269e3021894 801 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \
elessair 0:f269e3021894 802 EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \
elessair 0:f269e3021894 803 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \
elessair 0:f269e3021894 804 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 805 EMIT 0x0F EMIT 0xD4 EMIT 0xCC \
elessair 0:f269e3021894 806 EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \
elessair 0:f269e3021894 807 EMIT 0x0F EMIT 0xD4 EMIT 0xDD \
elessair 0:f269e3021894 808 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \
elessair 0:f269e3021894 809 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 810 EMIT 0x0F EMIT 0xD4 EMIT 0xCE \
elessair 0:f269e3021894 811 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \
elessair 0:f269e3021894 812 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 813 EMIT 0x0F EMIT 0xD4 EMIT 0xCB \
elessair 0:f269e3021894 814 EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \
elessair 0:f269e3021894 815 EMIT 0x83 EMIT 0xC7 EMIT 0x20 \
elessair 0:f269e3021894 816 EMIT 0x83 EMIT 0xC6 EMIT 0x20 \
elessair 0:f269e3021894 817 EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \
elessair 0:f269e3021894 818 EMIT 0x0F EMIT 0x7E EMIT 0xC9
elessair 0:f269e3021894 819
elessair 0:f269e3021894 820 #define MULADDC_STOP \
elessair 0:f269e3021894 821 EMIT 0x0F EMIT 0x77 \
elessair 0:f269e3021894 822 __asm mov c, ecx \
elessair 0:f269e3021894 823 __asm mov d, edi \
elessair 0:f269e3021894 824 __asm mov s, esi \
elessair 0:f269e3021894 825
elessair 0:f269e3021894 826 #else
elessair 0:f269e3021894 827
elessair 0:f269e3021894 828 #define MULADDC_STOP \
elessair 0:f269e3021894 829 __asm mov c, ecx \
elessair 0:f269e3021894 830 __asm mov d, edi \
elessair 0:f269e3021894 831 __asm mov s, esi \
elessair 0:f269e3021894 832
elessair 0:f269e3021894 833 #endif /* SSE2 */
elessair 0:f269e3021894 834 #endif /* MSVC */
elessair 0:f269e3021894 835
elessair 0:f269e3021894 836 #endif /* MBEDTLS_HAVE_ASM */
elessair 0:f269e3021894 837
elessair 0:f269e3021894 838 #if !defined(MULADDC_CORE)
elessair 0:f269e3021894 839 #if defined(MBEDTLS_HAVE_UDBL)
elessair 0:f269e3021894 840
elessair 0:f269e3021894 841 #define MULADDC_INIT \
elessair 0:f269e3021894 842 { \
elessair 0:f269e3021894 843 mbedtls_t_udbl r; \
elessair 0:f269e3021894 844 mbedtls_mpi_uint r0, r1;
elessair 0:f269e3021894 845
elessair 0:f269e3021894 846 #define MULADDC_CORE \
elessair 0:f269e3021894 847 r = *(s++) * (mbedtls_t_udbl) b; \
elessair 0:f269e3021894 848 r0 = (mbedtls_mpi_uint) r; \
elessair 0:f269e3021894 849 r1 = (mbedtls_mpi_uint)( r >> biL ); \
elessair 0:f269e3021894 850 r0 += c; r1 += (r0 < c); \
elessair 0:f269e3021894 851 r0 += *d; r1 += (r0 < *d); \
elessair 0:f269e3021894 852 c = r1; *(d++) = r0;
elessair 0:f269e3021894 853
elessair 0:f269e3021894 854 #define MULADDC_STOP \
elessair 0:f269e3021894 855 }
elessair 0:f269e3021894 856
elessair 0:f269e3021894 857 #else
elessair 0:f269e3021894 858 #define MULADDC_INIT \
elessair 0:f269e3021894 859 { \
elessair 0:f269e3021894 860 mbedtls_mpi_uint s0, s1, b0, b1; \
elessair 0:f269e3021894 861 mbedtls_mpi_uint r0, r1, rx, ry; \
elessair 0:f269e3021894 862 b0 = ( b << biH ) >> biH; \
elessair 0:f269e3021894 863 b1 = ( b >> biH );
elessair 0:f269e3021894 864
elessair 0:f269e3021894 865 #define MULADDC_CORE \
elessair 0:f269e3021894 866 s0 = ( *s << biH ) >> biH; \
elessair 0:f269e3021894 867 s1 = ( *s >> biH ); s++; \
elessair 0:f269e3021894 868 rx = s0 * b1; r0 = s0 * b0; \
elessair 0:f269e3021894 869 ry = s1 * b0; r1 = s1 * b1; \
elessair 0:f269e3021894 870 r1 += ( rx >> biH ); \
elessair 0:f269e3021894 871 r1 += ( ry >> biH ); \
elessair 0:f269e3021894 872 rx <<= biH; ry <<= biH; \
elessair 0:f269e3021894 873 r0 += rx; r1 += (r0 < rx); \
elessair 0:f269e3021894 874 r0 += ry; r1 += (r0 < ry); \
elessair 0:f269e3021894 875 r0 += c; r1 += (r0 < c); \
elessair 0:f269e3021894 876 r0 += *d; r1 += (r0 < *d); \
elessair 0:f269e3021894 877 c = r1; *(d++) = r0;
elessair 0:f269e3021894 878
elessair 0:f269e3021894 879 #define MULADDC_STOP \
elessair 0:f269e3021894 880 }
elessair 0:f269e3021894 881
elessair 0:f269e3021894 882 #endif /* C (generic) */
elessair 0:f269e3021894 883 #endif /* C (longlong) */
elessair 0:f269e3021894 884
elessair 0:f269e3021894 885 #endif /* bn_mul.h */