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.
mbed-os/features/mbedtls/inc/mbedtls/bn_mul.h@0:02dd72d1d465, 2018-04-24 (annotated)
- Committer:
- borlanic
- Date:
- Tue Apr 24 11:45:18 2018 +0000
- Revision:
- 0:02dd72d1d465
BaBoRo_test2 - backup 1
Who changed what in which revision?
| User | Revision | Line number | New contents of line |
|---|---|---|---|
| borlanic | 0:02dd72d1d465 | 1 | /** |
| borlanic | 0:02dd72d1d465 | 2 | * \file bn_mul.h |
| borlanic | 0:02dd72d1d465 | 3 | * |
| borlanic | 0:02dd72d1d465 | 4 | * \brief Multi-precision integer library |
| borlanic | 0:02dd72d1d465 | 5 | */ |
| borlanic | 0:02dd72d1d465 | 6 | /* |
| borlanic | 0:02dd72d1d465 | 7 | * Copyright (C) 2006-2015, ARM Limited, All Rights Reserved |
| borlanic | 0:02dd72d1d465 | 8 | * SPDX-License-Identifier: Apache-2.0 |
| borlanic | 0:02dd72d1d465 | 9 | * |
| borlanic | 0:02dd72d1d465 | 10 | * Licensed under the Apache License, Version 2.0 (the "License"); you may |
| borlanic | 0:02dd72d1d465 | 11 | * not use this file except in compliance with the License. |
| borlanic | 0:02dd72d1d465 | 12 | * You may obtain a copy of the License at |
| borlanic | 0:02dd72d1d465 | 13 | * |
| borlanic | 0:02dd72d1d465 | 14 | * http://www.apache.org/licenses/LICENSE-2.0 |
| borlanic | 0:02dd72d1d465 | 15 | * |
| borlanic | 0:02dd72d1d465 | 16 | * Unless required by applicable law or agreed to in writing, software |
| borlanic | 0:02dd72d1d465 | 17 | * distributed under the License is distributed on an "AS IS" BASIS, WITHOUT |
| borlanic | 0:02dd72d1d465 | 18 | * WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| borlanic | 0:02dd72d1d465 | 19 | * See the License for the specific language governing permissions and |
| borlanic | 0:02dd72d1d465 | 20 | * limitations under the License. |
| borlanic | 0:02dd72d1d465 | 21 | * |
| borlanic | 0:02dd72d1d465 | 22 | * This file is part of mbed TLS (https://tls.mbed.org) |
| borlanic | 0:02dd72d1d465 | 23 | */ |
| borlanic | 0:02dd72d1d465 | 24 | /* |
| borlanic | 0:02dd72d1d465 | 25 | * Multiply source vector [s] with b, add result |
| borlanic | 0:02dd72d1d465 | 26 | * to destination vector [d] and set carry c. |
| borlanic | 0:02dd72d1d465 | 27 | * |
| borlanic | 0:02dd72d1d465 | 28 | * Currently supports: |
| borlanic | 0:02dd72d1d465 | 29 | * |
| borlanic | 0:02dd72d1d465 | 30 | * . IA-32 (386+) . AMD64 / EM64T |
| borlanic | 0:02dd72d1d465 | 31 | * . IA-32 (SSE2) . Motorola 68000 |
| borlanic | 0:02dd72d1d465 | 32 | * . PowerPC, 32-bit . MicroBlaze |
| borlanic | 0:02dd72d1d465 | 33 | * . PowerPC, 64-bit . TriCore |
| borlanic | 0:02dd72d1d465 | 34 | * . SPARC v8 . ARM v3+ |
| borlanic | 0:02dd72d1d465 | 35 | * . Alpha . MIPS32 |
| borlanic | 0:02dd72d1d465 | 36 | * . C, longlong . C, generic |
| borlanic | 0:02dd72d1d465 | 37 | */ |
| borlanic | 0:02dd72d1d465 | 38 | #ifndef MBEDTLS_BN_MUL_H |
| borlanic | 0:02dd72d1d465 | 39 | #define MBEDTLS_BN_MUL_H |
| borlanic | 0:02dd72d1d465 | 40 | |
| borlanic | 0:02dd72d1d465 | 41 | #include "bignum.h" |
| borlanic | 0:02dd72d1d465 | 42 | |
| borlanic | 0:02dd72d1d465 | 43 | #if defined(MBEDTLS_HAVE_ASM) |
| borlanic | 0:02dd72d1d465 | 44 | |
| borlanic | 0:02dd72d1d465 | 45 | #ifndef asm |
| borlanic | 0:02dd72d1d465 | 46 | #define asm __asm |
| borlanic | 0:02dd72d1d465 | 47 | #endif |
| borlanic | 0:02dd72d1d465 | 48 | |
| borlanic | 0:02dd72d1d465 | 49 | /* armcc5 --gnu defines __GNUC__ but doesn't support GNU's extended asm */ |
| borlanic | 0:02dd72d1d465 | 50 | #if defined(__GNUC__) && \ |
| borlanic | 0:02dd72d1d465 | 51 | ( !defined(__ARMCC_VERSION) || __ARMCC_VERSION >= 6000000 ) |
| borlanic | 0:02dd72d1d465 | 52 | #if defined(__i386__) |
| borlanic | 0:02dd72d1d465 | 53 | |
| borlanic | 0:02dd72d1d465 | 54 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 55 | asm( \ |
| borlanic | 0:02dd72d1d465 | 56 | "movl %%ebx, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 57 | "movl %5, %%esi \n\t" \ |
| borlanic | 0:02dd72d1d465 | 58 | "movl %6, %%edi \n\t" \ |
| borlanic | 0:02dd72d1d465 | 59 | "movl %7, %%ecx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 60 | "movl %8, %%ebx \n\t" |
| borlanic | 0:02dd72d1d465 | 61 | |
| borlanic | 0:02dd72d1d465 | 62 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 63 | "lodsl \n\t" \ |
| borlanic | 0:02dd72d1d465 | 64 | "mull %%ebx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 65 | "addl %%ecx, %%eax \n\t" \ |
| borlanic | 0:02dd72d1d465 | 66 | "adcl $0, %%edx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 67 | "addl (%%edi), %%eax \n\t" \ |
| borlanic | 0:02dd72d1d465 | 68 | "adcl $0, %%edx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 69 | "movl %%edx, %%ecx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 70 | "stosl \n\t" |
| borlanic | 0:02dd72d1d465 | 71 | |
| borlanic | 0:02dd72d1d465 | 72 | #if defined(MBEDTLS_HAVE_SSE2) |
| borlanic | 0:02dd72d1d465 | 73 | |
| borlanic | 0:02dd72d1d465 | 74 | #define MULADDC_HUIT \ |
| borlanic | 0:02dd72d1d465 | 75 | "movd %%ecx, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 76 | "movd %%ebx, %%mm0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 77 | "movd (%%edi), %%mm3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 78 | "paddq %%mm3, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 79 | "movd (%%esi), %%mm2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 80 | "pmuludq %%mm0, %%mm2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 81 | "movd 4(%%esi), %%mm4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 82 | "pmuludq %%mm0, %%mm4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 83 | "movd 8(%%esi), %%mm6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 84 | "pmuludq %%mm0, %%mm6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 85 | "movd 12(%%esi), %%mm7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 86 | "pmuludq %%mm0, %%mm7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 87 | "paddq %%mm2, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 88 | "movd 4(%%edi), %%mm3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 89 | "paddq %%mm4, %%mm3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 90 | "movd 8(%%edi), %%mm5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 91 | "paddq %%mm6, %%mm5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 92 | "movd 12(%%edi), %%mm4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 93 | "paddq %%mm4, %%mm7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 94 | "movd %%mm1, (%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 95 | "movd 16(%%esi), %%mm2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 96 | "pmuludq %%mm0, %%mm2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 97 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 98 | "movd 20(%%esi), %%mm4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 99 | "pmuludq %%mm0, %%mm4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 100 | "paddq %%mm3, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 101 | "movd 24(%%esi), %%mm6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 102 | "pmuludq %%mm0, %%mm6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 103 | "movd %%mm1, 4(%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 104 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 105 | "movd 28(%%esi), %%mm3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 106 | "pmuludq %%mm0, %%mm3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 107 | "paddq %%mm5, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 108 | "movd 16(%%edi), %%mm5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 109 | "paddq %%mm5, %%mm2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 110 | "movd %%mm1, 8(%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 111 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 112 | "paddq %%mm7, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 113 | "movd 20(%%edi), %%mm5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 114 | "paddq %%mm5, %%mm4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 115 | "movd %%mm1, 12(%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 116 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 117 | "paddq %%mm2, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 118 | "movd 24(%%edi), %%mm5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 119 | "paddq %%mm5, %%mm6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 120 | "movd %%mm1, 16(%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 121 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 122 | "paddq %%mm4, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 123 | "movd 28(%%edi), %%mm5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 124 | "paddq %%mm5, %%mm3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 125 | "movd %%mm1, 20(%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 126 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 127 | "paddq %%mm6, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 128 | "movd %%mm1, 24(%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 129 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 130 | "paddq %%mm3, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 131 | "movd %%mm1, 28(%%edi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 132 | "addl $32, %%edi \n\t" \ |
| borlanic | 0:02dd72d1d465 | 133 | "addl $32, %%esi \n\t" \ |
| borlanic | 0:02dd72d1d465 | 134 | "psrlq $32, %%mm1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 135 | "movd %%mm1, %%ecx \n\t" |
| borlanic | 0:02dd72d1d465 | 136 | |
| borlanic | 0:02dd72d1d465 | 137 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 138 | "emms \n\t" \ |
| borlanic | 0:02dd72d1d465 | 139 | "movl %4, %%ebx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 140 | "movl %%ecx, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 141 | "movl %%edi, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 142 | "movl %%esi, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 143 | : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 144 | : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 145 | : "eax", "ecx", "edx", "esi", "edi" \ |
| borlanic | 0:02dd72d1d465 | 146 | ); |
| borlanic | 0:02dd72d1d465 | 147 | |
| borlanic | 0:02dd72d1d465 | 148 | #else |
| borlanic | 0:02dd72d1d465 | 149 | |
| borlanic | 0:02dd72d1d465 | 150 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 151 | "movl %4, %%ebx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 152 | "movl %%ecx, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 153 | "movl %%edi, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 154 | "movl %%esi, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 155 | : "=m" (t), "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 156 | : "m" (t), "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 157 | : "eax", "ecx", "edx", "esi", "edi" \ |
| borlanic | 0:02dd72d1d465 | 158 | ); |
| borlanic | 0:02dd72d1d465 | 159 | #endif /* SSE2 */ |
| borlanic | 0:02dd72d1d465 | 160 | #endif /* i386 */ |
| borlanic | 0:02dd72d1d465 | 161 | |
| borlanic | 0:02dd72d1d465 | 162 | #if defined(__amd64__) || defined (__x86_64__) |
| borlanic | 0:02dd72d1d465 | 163 | |
| borlanic | 0:02dd72d1d465 | 164 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 165 | asm( \ |
| borlanic | 0:02dd72d1d465 | 166 | "xorq %%r8, %%r8 \n\t" |
| borlanic | 0:02dd72d1d465 | 167 | |
| borlanic | 0:02dd72d1d465 | 168 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 169 | "movq (%%rsi), %%rax \n\t" \ |
| borlanic | 0:02dd72d1d465 | 170 | "mulq %%rbx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 171 | "addq $8, %%rsi \n\t" \ |
| borlanic | 0:02dd72d1d465 | 172 | "addq %%rcx, %%rax \n\t" \ |
| borlanic | 0:02dd72d1d465 | 173 | "movq %%r8, %%rcx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 174 | "adcq $0, %%rdx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 175 | "nop \n\t" \ |
| borlanic | 0:02dd72d1d465 | 176 | "addq %%rax, (%%rdi) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 177 | "adcq %%rdx, %%rcx \n\t" \ |
| borlanic | 0:02dd72d1d465 | 178 | "addq $8, %%rdi \n\t" |
| borlanic | 0:02dd72d1d465 | 179 | |
| borlanic | 0:02dd72d1d465 | 180 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 181 | : "+c" (c), "+D" (d), "+S" (s) \ |
| borlanic | 0:02dd72d1d465 | 182 | : "b" (b) \ |
| borlanic | 0:02dd72d1d465 | 183 | : "rax", "rdx", "r8" \ |
| borlanic | 0:02dd72d1d465 | 184 | ); |
| borlanic | 0:02dd72d1d465 | 185 | |
| borlanic | 0:02dd72d1d465 | 186 | #endif /* AMD64 */ |
| borlanic | 0:02dd72d1d465 | 187 | |
| borlanic | 0:02dd72d1d465 | 188 | #if defined(__mc68020__) || defined(__mcpu32__) |
| borlanic | 0:02dd72d1d465 | 189 | |
| borlanic | 0:02dd72d1d465 | 190 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 191 | asm( \ |
| borlanic | 0:02dd72d1d465 | 192 | "movl %3, %%a2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 193 | "movl %4, %%a3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 194 | "movl %5, %%d3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 195 | "movl %6, %%d2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 196 | "moveq #0, %%d0 \n\t" |
| borlanic | 0:02dd72d1d465 | 197 | |
| borlanic | 0:02dd72d1d465 | 198 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 199 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 200 | "mulul %%d2, %%d4:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 201 | "addl %%d3, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 202 | "addxl %%d0, %%d4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 203 | "moveq #0, %%d3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 204 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 205 | "addxl %%d4, %%d3 \n\t" |
| borlanic | 0:02dd72d1d465 | 206 | |
| borlanic | 0:02dd72d1d465 | 207 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 208 | "movl %%d3, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 209 | "movl %%a3, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 210 | "movl %%a2, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 211 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 212 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 213 | : "d0", "d1", "d2", "d3", "d4", "a2", "a3" \ |
| borlanic | 0:02dd72d1d465 | 214 | ); |
| borlanic | 0:02dd72d1d465 | 215 | |
| borlanic | 0:02dd72d1d465 | 216 | #define MULADDC_HUIT \ |
| borlanic | 0:02dd72d1d465 | 217 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 218 | "mulul %%d2, %%d4:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 219 | "addxl %%d3, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 220 | "addxl %%d0, %%d4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 221 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 222 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 223 | "mulul %%d2, %%d3:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 224 | "addxl %%d4, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 225 | "addxl %%d0, %%d3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 226 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 227 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 228 | "mulul %%d2, %%d4:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 229 | "addxl %%d3, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 230 | "addxl %%d0, %%d4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 231 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 232 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 233 | "mulul %%d2, %%d3:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 234 | "addxl %%d4, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 235 | "addxl %%d0, %%d3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 236 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 237 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 238 | "mulul %%d2, %%d4:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 239 | "addxl %%d3, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 240 | "addxl %%d0, %%d4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 241 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 242 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 243 | "mulul %%d2, %%d3:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 244 | "addxl %%d4, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 245 | "addxl %%d0, %%d3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 246 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 247 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 248 | "mulul %%d2, %%d4:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 249 | "addxl %%d3, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 250 | "addxl %%d0, %%d4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 251 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 252 | "movel %%a2@+, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 253 | "mulul %%d2, %%d3:%%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 254 | "addxl %%d4, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 255 | "addxl %%d0, %%d3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 256 | "addl %%d1, %%a3@+ \n\t" \ |
| borlanic | 0:02dd72d1d465 | 257 | "addxl %%d0, %%d3 \n\t" |
| borlanic | 0:02dd72d1d465 | 258 | |
| borlanic | 0:02dd72d1d465 | 259 | #endif /* MC68000 */ |
| borlanic | 0:02dd72d1d465 | 260 | |
| borlanic | 0:02dd72d1d465 | 261 | #if defined(__powerpc64__) || defined(__ppc64__) |
| borlanic | 0:02dd72d1d465 | 262 | |
| borlanic | 0:02dd72d1d465 | 263 | #if defined(__MACH__) && defined(__APPLE__) |
| borlanic | 0:02dd72d1d465 | 264 | |
| borlanic | 0:02dd72d1d465 | 265 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 266 | asm( \ |
| borlanic | 0:02dd72d1d465 | 267 | "ld r3, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 268 | "ld r4, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 269 | "ld r5, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 270 | "ld r6, %6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 271 | "addi r3, r3, -8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 272 | "addi r4, r4, -8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 273 | "addic r5, r5, 0 \n\t" |
| borlanic | 0:02dd72d1d465 | 274 | |
| borlanic | 0:02dd72d1d465 | 275 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 276 | "ldu r7, 8(r3) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 277 | "mulld r8, r7, r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 278 | "mulhdu r9, r7, r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 279 | "adde r8, r8, r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 280 | "ld r7, 8(r4) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 281 | "addze r5, r9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 282 | "addc r8, r8, r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 283 | "stdu r8, 8(r4) \n\t" |
| borlanic | 0:02dd72d1d465 | 284 | |
| borlanic | 0:02dd72d1d465 | 285 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 286 | "addze r5, r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 287 | "addi r4, r4, 8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 288 | "addi r3, r3, 8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 289 | "std r5, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 290 | "std r4, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 291 | "std r3, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 292 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 293 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 294 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| borlanic | 0:02dd72d1d465 | 295 | ); |
| borlanic | 0:02dd72d1d465 | 296 | |
| borlanic | 0:02dd72d1d465 | 297 | |
| borlanic | 0:02dd72d1d465 | 298 | #else /* __MACH__ && __APPLE__ */ |
| borlanic | 0:02dd72d1d465 | 299 | |
| borlanic | 0:02dd72d1d465 | 300 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 301 | asm( \ |
| borlanic | 0:02dd72d1d465 | 302 | "ld %%r3, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 303 | "ld %%r4, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 304 | "ld %%r5, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 305 | "ld %%r6, %6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 306 | "addi %%r3, %%r3, -8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 307 | "addi %%r4, %%r4, -8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 308 | "addic %%r5, %%r5, 0 \n\t" |
| borlanic | 0:02dd72d1d465 | 309 | |
| borlanic | 0:02dd72d1d465 | 310 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 311 | "ldu %%r7, 8(%%r3) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 312 | "mulld %%r8, %%r7, %%r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 313 | "mulhdu %%r9, %%r7, %%r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 314 | "adde %%r8, %%r8, %%r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 315 | "ld %%r7, 8(%%r4) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 316 | "addze %%r5, %%r9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 317 | "addc %%r8, %%r8, %%r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 318 | "stdu %%r8, 8(%%r4) \n\t" |
| borlanic | 0:02dd72d1d465 | 319 | |
| borlanic | 0:02dd72d1d465 | 320 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 321 | "addze %%r5, %%r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 322 | "addi %%r4, %%r4, 8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 323 | "addi %%r3, %%r3, 8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 324 | "std %%r5, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 325 | "std %%r4, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 326 | "std %%r3, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 327 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 328 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 329 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| borlanic | 0:02dd72d1d465 | 330 | ); |
| borlanic | 0:02dd72d1d465 | 331 | |
| borlanic | 0:02dd72d1d465 | 332 | #endif /* __MACH__ && __APPLE__ */ |
| borlanic | 0:02dd72d1d465 | 333 | |
| borlanic | 0:02dd72d1d465 | 334 | #elif defined(__powerpc__) || defined(__ppc__) /* end PPC64/begin PPC32 */ |
| borlanic | 0:02dd72d1d465 | 335 | |
| borlanic | 0:02dd72d1d465 | 336 | #if defined(__MACH__) && defined(__APPLE__) |
| borlanic | 0:02dd72d1d465 | 337 | |
| borlanic | 0:02dd72d1d465 | 338 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 339 | asm( \ |
| borlanic | 0:02dd72d1d465 | 340 | "lwz r3, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 341 | "lwz r4, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 342 | "lwz r5, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 343 | "lwz r6, %6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 344 | "addi r3, r3, -4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 345 | "addi r4, r4, -4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 346 | "addic r5, r5, 0 \n\t" |
| borlanic | 0:02dd72d1d465 | 347 | |
| borlanic | 0:02dd72d1d465 | 348 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 349 | "lwzu r7, 4(r3) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 350 | "mullw r8, r7, r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 351 | "mulhwu r9, r7, r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 352 | "adde r8, r8, r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 353 | "lwz r7, 4(r4) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 354 | "addze r5, r9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 355 | "addc r8, r8, r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 356 | "stwu r8, 4(r4) \n\t" |
| borlanic | 0:02dd72d1d465 | 357 | |
| borlanic | 0:02dd72d1d465 | 358 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 359 | "addze r5, r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 360 | "addi r4, r4, 4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 361 | "addi r3, r3, 4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 362 | "stw r5, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 363 | "stw r4, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 364 | "stw r3, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 365 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 366 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 367 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| borlanic | 0:02dd72d1d465 | 368 | ); |
| borlanic | 0:02dd72d1d465 | 369 | |
| borlanic | 0:02dd72d1d465 | 370 | #else /* __MACH__ && __APPLE__ */ |
| borlanic | 0:02dd72d1d465 | 371 | |
| borlanic | 0:02dd72d1d465 | 372 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 373 | asm( \ |
| borlanic | 0:02dd72d1d465 | 374 | "lwz %%r3, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 375 | "lwz %%r4, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 376 | "lwz %%r5, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 377 | "lwz %%r6, %6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 378 | "addi %%r3, %%r3, -4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 379 | "addi %%r4, %%r4, -4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 380 | "addic %%r5, %%r5, 0 \n\t" |
| borlanic | 0:02dd72d1d465 | 381 | |
| borlanic | 0:02dd72d1d465 | 382 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 383 | "lwzu %%r7, 4(%%r3) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 384 | "mullw %%r8, %%r7, %%r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 385 | "mulhwu %%r9, %%r7, %%r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 386 | "adde %%r8, %%r8, %%r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 387 | "lwz %%r7, 4(%%r4) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 388 | "addze %%r5, %%r9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 389 | "addc %%r8, %%r8, %%r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 390 | "stwu %%r8, 4(%%r4) \n\t" |
| borlanic | 0:02dd72d1d465 | 391 | |
| borlanic | 0:02dd72d1d465 | 392 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 393 | "addze %%r5, %%r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 394 | "addi %%r4, %%r4, 4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 395 | "addi %%r3, %%r3, 4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 396 | "stw %%r5, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 397 | "stw %%r4, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 398 | "stw %%r3, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 399 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 400 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 401 | : "r3", "r4", "r5", "r6", "r7", "r8", "r9" \ |
| borlanic | 0:02dd72d1d465 | 402 | ); |
| borlanic | 0:02dd72d1d465 | 403 | |
| borlanic | 0:02dd72d1d465 | 404 | #endif /* __MACH__ && __APPLE__ */ |
| borlanic | 0:02dd72d1d465 | 405 | |
| borlanic | 0:02dd72d1d465 | 406 | #endif /* PPC32 */ |
| borlanic | 0:02dd72d1d465 | 407 | |
| borlanic | 0:02dd72d1d465 | 408 | /* |
| borlanic | 0:02dd72d1d465 | 409 | * The Sparc(64) assembly is reported to be broken. |
| borlanic | 0:02dd72d1d465 | 410 | * Disable it for now, until we're able to fix it. |
| borlanic | 0:02dd72d1d465 | 411 | */ |
| borlanic | 0:02dd72d1d465 | 412 | #if 0 && defined(__sparc__) |
| borlanic | 0:02dd72d1d465 | 413 | #if defined(__sparc64__) |
| borlanic | 0:02dd72d1d465 | 414 | |
| borlanic | 0:02dd72d1d465 | 415 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 416 | asm( \ |
| borlanic | 0:02dd72d1d465 | 417 | "ldx %3, %%o0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 418 | "ldx %4, %%o1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 419 | "ld %5, %%o2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 420 | "ld %6, %%o3 \n\t" |
| borlanic | 0:02dd72d1d465 | 421 | |
| borlanic | 0:02dd72d1d465 | 422 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 423 | "ld [%%o0], %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 424 | "inc 4, %%o0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 425 | "ld [%%o1], %%o5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 426 | "umul %%o3, %%o4, %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 427 | "addcc %%o4, %%o2, %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 428 | "rd %%y, %%g1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 429 | "addx %%g1, 0, %%g1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 430 | "addcc %%o4, %%o5, %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 431 | "st %%o4, [%%o1] \n\t" \ |
| borlanic | 0:02dd72d1d465 | 432 | "addx %%g1, 0, %%o2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 433 | "inc 4, %%o1 \n\t" |
| borlanic | 0:02dd72d1d465 | 434 | |
| borlanic | 0:02dd72d1d465 | 435 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 436 | "st %%o2, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 437 | "stx %%o1, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 438 | "stx %%o0, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 439 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 440 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 441 | : "g1", "o0", "o1", "o2", "o3", "o4", \ |
| borlanic | 0:02dd72d1d465 | 442 | "o5" \ |
| borlanic | 0:02dd72d1d465 | 443 | ); |
| borlanic | 0:02dd72d1d465 | 444 | |
| borlanic | 0:02dd72d1d465 | 445 | #else /* __sparc64__ */ |
| borlanic | 0:02dd72d1d465 | 446 | |
| borlanic | 0:02dd72d1d465 | 447 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 448 | asm( \ |
| borlanic | 0:02dd72d1d465 | 449 | "ld %3, %%o0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 450 | "ld %4, %%o1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 451 | "ld %5, %%o2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 452 | "ld %6, %%o3 \n\t" |
| borlanic | 0:02dd72d1d465 | 453 | |
| borlanic | 0:02dd72d1d465 | 454 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 455 | "ld [%%o0], %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 456 | "inc 4, %%o0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 457 | "ld [%%o1], %%o5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 458 | "umul %%o3, %%o4, %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 459 | "addcc %%o4, %%o2, %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 460 | "rd %%y, %%g1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 461 | "addx %%g1, 0, %%g1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 462 | "addcc %%o4, %%o5, %%o4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 463 | "st %%o4, [%%o1] \n\t" \ |
| borlanic | 0:02dd72d1d465 | 464 | "addx %%g1, 0, %%o2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 465 | "inc 4, %%o1 \n\t" |
| borlanic | 0:02dd72d1d465 | 466 | |
| borlanic | 0:02dd72d1d465 | 467 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 468 | "st %%o2, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 469 | "st %%o1, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 470 | "st %%o0, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 471 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 472 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 473 | : "g1", "o0", "o1", "o2", "o3", "o4", \ |
| borlanic | 0:02dd72d1d465 | 474 | "o5" \ |
| borlanic | 0:02dd72d1d465 | 475 | ); |
| borlanic | 0:02dd72d1d465 | 476 | |
| borlanic | 0:02dd72d1d465 | 477 | #endif /* __sparc64__ */ |
| borlanic | 0:02dd72d1d465 | 478 | #endif /* __sparc__ */ |
| borlanic | 0:02dd72d1d465 | 479 | |
| borlanic | 0:02dd72d1d465 | 480 | #if defined(__microblaze__) || defined(microblaze) |
| borlanic | 0:02dd72d1d465 | 481 | |
| borlanic | 0:02dd72d1d465 | 482 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 483 | asm( \ |
| borlanic | 0:02dd72d1d465 | 484 | "lwi r3, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 485 | "lwi r4, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 486 | "lwi r5, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 487 | "lwi r6, %6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 488 | "andi r7, r6, 0xffff \n\t" \ |
| borlanic | 0:02dd72d1d465 | 489 | "bsrli r6, r6, 16 \n\t" |
| borlanic | 0:02dd72d1d465 | 490 | |
| borlanic | 0:02dd72d1d465 | 491 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 492 | "lhui r8, r3, 0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 493 | "addi r3, r3, 2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 494 | "lhui r9, r3, 0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 495 | "addi r3, r3, 2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 496 | "mul r10, r9, r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 497 | "mul r11, r8, r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 498 | "mul r12, r9, r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 499 | "mul r13, r8, r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 500 | "bsrli r8, r10, 16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 501 | "bsrli r9, r11, 16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 502 | "add r13, r13, r8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 503 | "add r13, r13, r9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 504 | "bslli r10, r10, 16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 505 | "bslli r11, r11, 16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 506 | "add r12, r12, r10 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 507 | "addc r13, r13, r0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 508 | "add r12, r12, r11 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 509 | "addc r13, r13, r0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 510 | "lwi r10, r4, 0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 511 | "add r12, r12, r10 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 512 | "addc r13, r13, r0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 513 | "add r12, r12, r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 514 | "addc r5, r13, r0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 515 | "swi r12, r4, 0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 516 | "addi r4, r4, 4 \n\t" |
| borlanic | 0:02dd72d1d465 | 517 | |
| borlanic | 0:02dd72d1d465 | 518 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 519 | "swi r5, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 520 | "swi r4, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 521 | "swi r3, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 522 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 523 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 524 | : "r3", "r4" "r5", "r6", "r7", "r8", \ |
| borlanic | 0:02dd72d1d465 | 525 | "r9", "r10", "r11", "r12", "r13" \ |
| borlanic | 0:02dd72d1d465 | 526 | ); |
| borlanic | 0:02dd72d1d465 | 527 | |
| borlanic | 0:02dd72d1d465 | 528 | #endif /* MicroBlaze */ |
| borlanic | 0:02dd72d1d465 | 529 | |
| borlanic | 0:02dd72d1d465 | 530 | #if defined(__tricore__) |
| borlanic | 0:02dd72d1d465 | 531 | |
| borlanic | 0:02dd72d1d465 | 532 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 533 | asm( \ |
| borlanic | 0:02dd72d1d465 | 534 | "ld.a %%a2, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 535 | "ld.a %%a3, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 536 | "ld.w %%d4, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 537 | "ld.w %%d1, %6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 538 | "xor %%d5, %%d5 \n\t" |
| borlanic | 0:02dd72d1d465 | 539 | |
| borlanic | 0:02dd72d1d465 | 540 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 541 | "ld.w %%d0, [%%a2+] \n\t" \ |
| borlanic | 0:02dd72d1d465 | 542 | "madd.u %%e2, %%e4, %%d0, %%d1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 543 | "ld.w %%d0, [%%a3] \n\t" \ |
| borlanic | 0:02dd72d1d465 | 544 | "addx %%d2, %%d2, %%d0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 545 | "addc %%d3, %%d3, 0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 546 | "mov %%d4, %%d3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 547 | "st.w [%%a3+], %%d2 \n\t" |
| borlanic | 0:02dd72d1d465 | 548 | |
| borlanic | 0:02dd72d1d465 | 549 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 550 | "st.w %0, %%d4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 551 | "st.a %1, %%a3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 552 | "st.a %2, %%a2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 553 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 554 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 555 | : "d0", "d1", "e2", "d4", "a2", "a3" \ |
| borlanic | 0:02dd72d1d465 | 556 | ); |
| borlanic | 0:02dd72d1d465 | 557 | |
| borlanic | 0:02dd72d1d465 | 558 | #endif /* TriCore */ |
| borlanic | 0:02dd72d1d465 | 559 | |
| borlanic | 0:02dd72d1d465 | 560 | /* |
| borlanic | 0:02dd72d1d465 | 561 | * gcc -O0 by default uses r7 for the frame pointer, so it complains about our |
| borlanic | 0:02dd72d1d465 | 562 | * use of r7 below, unless -fomit-frame-pointer is passed. Unfortunately, |
| borlanic | 0:02dd72d1d465 | 563 | * passing that option is not easy when building with yotta. |
| borlanic | 0:02dd72d1d465 | 564 | * |
| borlanic | 0:02dd72d1d465 | 565 | * On the other hand, -fomit-frame-pointer is implied by any -Ox options with |
| borlanic | 0:02dd72d1d465 | 566 | * x !=0, which we can detect using __OPTIMIZE__ (which is also defined by |
| borlanic | 0:02dd72d1d465 | 567 | * clang and armcc5 under the same conditions). |
| borlanic | 0:02dd72d1d465 | 568 | * |
| borlanic | 0:02dd72d1d465 | 569 | * So, only use the optimized assembly below for optimized build, which avoids |
| borlanic | 0:02dd72d1d465 | 570 | * the build error and is pretty reasonable anyway. |
| borlanic | 0:02dd72d1d465 | 571 | */ |
| borlanic | 0:02dd72d1d465 | 572 | #if defined(__GNUC__) && !defined(__OPTIMIZE__) |
| borlanic | 0:02dd72d1d465 | 573 | #define MULADDC_CANNOT_USE_R7 |
| borlanic | 0:02dd72d1d465 | 574 | #endif |
| borlanic | 0:02dd72d1d465 | 575 | |
| borlanic | 0:02dd72d1d465 | 576 | #if defined(__arm__) && !defined(MULADDC_CANNOT_USE_R7) |
| borlanic | 0:02dd72d1d465 | 577 | |
| borlanic | 0:02dd72d1d465 | 578 | #if defined(__thumb__) && !defined(__thumb2__) |
| borlanic | 0:02dd72d1d465 | 579 | |
| borlanic | 0:02dd72d1d465 | 580 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 581 | asm( \ |
| borlanic | 0:02dd72d1d465 | 582 | "ldr r0, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 583 | "ldr r1, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 584 | "ldr r2, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 585 | "ldr r3, %6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 586 | "lsr r7, r3, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 587 | "mov r9, r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 588 | "lsl r7, r3, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 589 | "lsr r7, r7, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 590 | "mov r8, r7 \n\t" |
| borlanic | 0:02dd72d1d465 | 591 | |
| borlanic | 0:02dd72d1d465 | 592 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 593 | "ldmia r0!, {r6} \n\t" \ |
| borlanic | 0:02dd72d1d465 | 594 | "lsr r7, r6, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 595 | "lsl r6, r6, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 596 | "lsr r6, r6, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 597 | "mov r4, r8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 598 | "mul r4, r6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 599 | "mov r3, r9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 600 | "mul r6, r3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 601 | "mov r5, r9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 602 | "mul r5, r7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 603 | "mov r3, r8 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 604 | "mul r7, r3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 605 | "lsr r3, r6, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 606 | "add r5, r5, r3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 607 | "lsr r3, r7, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 608 | "add r5, r5, r3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 609 | "add r4, r4, r2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 610 | "mov r2, #0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 611 | "adc r5, r2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 612 | "lsl r3, r6, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 613 | "add r4, r4, r3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 614 | "adc r5, r2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 615 | "lsl r3, r7, #16 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 616 | "add r4, r4, r3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 617 | "adc r5, r2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 618 | "ldr r3, [r1] \n\t" \ |
| borlanic | 0:02dd72d1d465 | 619 | "add r4, r4, r3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 620 | "adc r2, r5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 621 | "stmia r1!, {r4} \n\t" |
| borlanic | 0:02dd72d1d465 | 622 | |
| borlanic | 0:02dd72d1d465 | 623 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 624 | "str r2, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 625 | "str r1, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 626 | "str r0, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 627 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 628 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 629 | : "r0", "r1", "r2", "r3", "r4", "r5", \ |
| borlanic | 0:02dd72d1d465 | 630 | "r6", "r7", "r8", "r9", "cc" \ |
| borlanic | 0:02dd72d1d465 | 631 | ); |
| borlanic | 0:02dd72d1d465 | 632 | |
| borlanic | 0:02dd72d1d465 | 633 | #else |
| borlanic | 0:02dd72d1d465 | 634 | |
| borlanic | 0:02dd72d1d465 | 635 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 636 | asm( \ |
| borlanic | 0:02dd72d1d465 | 637 | "ldr r0, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 638 | "ldr r1, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 639 | "ldr r2, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 640 | "ldr r3, %6 \n\t" |
| borlanic | 0:02dd72d1d465 | 641 | |
| borlanic | 0:02dd72d1d465 | 642 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 643 | "ldr r4, [r0], #4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 644 | "mov r5, #0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 645 | "ldr r6, [r1] \n\t" \ |
| borlanic | 0:02dd72d1d465 | 646 | "umlal r2, r5, r3, r4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 647 | "adds r7, r6, r2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 648 | "adc r2, r5, #0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 649 | "str r7, [r1], #4 \n\t" |
| borlanic | 0:02dd72d1d465 | 650 | |
| borlanic | 0:02dd72d1d465 | 651 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 652 | "str r2, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 653 | "str r1, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 654 | "str r0, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 655 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 656 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 657 | : "r0", "r1", "r2", "r3", "r4", "r5", \ |
| borlanic | 0:02dd72d1d465 | 658 | "r6", "r7", "cc" \ |
| borlanic | 0:02dd72d1d465 | 659 | ); |
| borlanic | 0:02dd72d1d465 | 660 | |
| borlanic | 0:02dd72d1d465 | 661 | #endif /* Thumb */ |
| borlanic | 0:02dd72d1d465 | 662 | |
| borlanic | 0:02dd72d1d465 | 663 | #endif /* ARMv3 */ |
| borlanic | 0:02dd72d1d465 | 664 | |
| borlanic | 0:02dd72d1d465 | 665 | #if defined(__alpha__) |
| borlanic | 0:02dd72d1d465 | 666 | |
| borlanic | 0:02dd72d1d465 | 667 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 668 | asm( \ |
| borlanic | 0:02dd72d1d465 | 669 | "ldq $1, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 670 | "ldq $2, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 671 | "ldq $3, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 672 | "ldq $4, %6 \n\t" |
| borlanic | 0:02dd72d1d465 | 673 | |
| borlanic | 0:02dd72d1d465 | 674 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 675 | "ldq $6, 0($1) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 676 | "addq $1, 8, $1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 677 | "mulq $6, $4, $7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 678 | "umulh $6, $4, $6 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 679 | "addq $7, $3, $7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 680 | "cmpult $7, $3, $3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 681 | "ldq $5, 0($2) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 682 | "addq $7, $5, $7 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 683 | "cmpult $7, $5, $5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 684 | "stq $7, 0($2) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 685 | "addq $2, 8, $2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 686 | "addq $6, $3, $3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 687 | "addq $5, $3, $3 \n\t" |
| borlanic | 0:02dd72d1d465 | 688 | |
| borlanic | 0:02dd72d1d465 | 689 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 690 | "stq $3, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 691 | "stq $2, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 692 | "stq $1, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 693 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 694 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 695 | : "$1", "$2", "$3", "$4", "$5", "$6", "$7" \ |
| borlanic | 0:02dd72d1d465 | 696 | ); |
| borlanic | 0:02dd72d1d465 | 697 | #endif /* Alpha */ |
| borlanic | 0:02dd72d1d465 | 698 | |
| borlanic | 0:02dd72d1d465 | 699 | #if defined(__mips__) && !defined(__mips64) |
| borlanic | 0:02dd72d1d465 | 700 | |
| borlanic | 0:02dd72d1d465 | 701 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 702 | asm( \ |
| borlanic | 0:02dd72d1d465 | 703 | "lw $10, %3 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 704 | "lw $11, %4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 705 | "lw $12, %5 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 706 | "lw $13, %6 \n\t" |
| borlanic | 0:02dd72d1d465 | 707 | |
| borlanic | 0:02dd72d1d465 | 708 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 709 | "lw $14, 0($10) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 710 | "multu $13, $14 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 711 | "addi $10, $10, 4 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 712 | "mflo $14 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 713 | "mfhi $9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 714 | "addu $14, $12, $14 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 715 | "lw $15, 0($11) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 716 | "sltu $12, $14, $12 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 717 | "addu $15, $14, $15 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 718 | "sltu $14, $15, $14 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 719 | "addu $12, $12, $9 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 720 | "sw $15, 0($11) \n\t" \ |
| borlanic | 0:02dd72d1d465 | 721 | "addu $12, $12, $14 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 722 | "addi $11, $11, 4 \n\t" |
| borlanic | 0:02dd72d1d465 | 723 | |
| borlanic | 0:02dd72d1d465 | 724 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 725 | "sw $12, %0 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 726 | "sw $11, %1 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 727 | "sw $10, %2 \n\t" \ |
| borlanic | 0:02dd72d1d465 | 728 | : "=m" (c), "=m" (d), "=m" (s) \ |
| borlanic | 0:02dd72d1d465 | 729 | : "m" (s), "m" (d), "m" (c), "m" (b) \ |
| borlanic | 0:02dd72d1d465 | 730 | : "$9", "$10", "$11", "$12", "$13", "$14", "$15" \ |
| borlanic | 0:02dd72d1d465 | 731 | ); |
| borlanic | 0:02dd72d1d465 | 732 | |
| borlanic | 0:02dd72d1d465 | 733 | #endif /* MIPS */ |
| borlanic | 0:02dd72d1d465 | 734 | #endif /* GNUC */ |
| borlanic | 0:02dd72d1d465 | 735 | |
| borlanic | 0:02dd72d1d465 | 736 | #if (defined(_MSC_VER) && defined(_M_IX86)) || defined(__WATCOMC__) |
| borlanic | 0:02dd72d1d465 | 737 | |
| borlanic | 0:02dd72d1d465 | 738 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 739 | __asm mov esi, s \ |
| borlanic | 0:02dd72d1d465 | 740 | __asm mov edi, d \ |
| borlanic | 0:02dd72d1d465 | 741 | __asm mov ecx, c \ |
| borlanic | 0:02dd72d1d465 | 742 | __asm mov ebx, b |
| borlanic | 0:02dd72d1d465 | 743 | |
| borlanic | 0:02dd72d1d465 | 744 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 745 | __asm lodsd \ |
| borlanic | 0:02dd72d1d465 | 746 | __asm mul ebx \ |
| borlanic | 0:02dd72d1d465 | 747 | __asm add eax, ecx \ |
| borlanic | 0:02dd72d1d465 | 748 | __asm adc edx, 0 \ |
| borlanic | 0:02dd72d1d465 | 749 | __asm add eax, [edi] \ |
| borlanic | 0:02dd72d1d465 | 750 | __asm adc edx, 0 \ |
| borlanic | 0:02dd72d1d465 | 751 | __asm mov ecx, edx \ |
| borlanic | 0:02dd72d1d465 | 752 | __asm stosd |
| borlanic | 0:02dd72d1d465 | 753 | |
| borlanic | 0:02dd72d1d465 | 754 | #if defined(MBEDTLS_HAVE_SSE2) |
| borlanic | 0:02dd72d1d465 | 755 | |
| borlanic | 0:02dd72d1d465 | 756 | #define EMIT __asm _emit |
| borlanic | 0:02dd72d1d465 | 757 | |
| borlanic | 0:02dd72d1d465 | 758 | #define MULADDC_HUIT \ |
| borlanic | 0:02dd72d1d465 | 759 | EMIT 0x0F EMIT 0x6E EMIT 0xC9 \ |
| borlanic | 0:02dd72d1d465 | 760 | EMIT 0x0F EMIT 0x6E EMIT 0xC3 \ |
| borlanic | 0:02dd72d1d465 | 761 | EMIT 0x0F EMIT 0x6E EMIT 0x1F \ |
| borlanic | 0:02dd72d1d465 | 762 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| borlanic | 0:02dd72d1d465 | 763 | EMIT 0x0F EMIT 0x6E EMIT 0x16 \ |
| borlanic | 0:02dd72d1d465 | 764 | EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ |
| borlanic | 0:02dd72d1d465 | 765 | EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x04 \ |
| borlanic | 0:02dd72d1d465 | 766 | EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ |
| borlanic | 0:02dd72d1d465 | 767 | EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x08 \ |
| borlanic | 0:02dd72d1d465 | 768 | EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ |
| borlanic | 0:02dd72d1d465 | 769 | EMIT 0x0F EMIT 0x6E EMIT 0x7E EMIT 0x0C \ |
| borlanic | 0:02dd72d1d465 | 770 | EMIT 0x0F EMIT 0xF4 EMIT 0xF8 \ |
| borlanic | 0:02dd72d1d465 | 771 | EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ |
| borlanic | 0:02dd72d1d465 | 772 | EMIT 0x0F EMIT 0x6E EMIT 0x5F EMIT 0x04 \ |
| borlanic | 0:02dd72d1d465 | 773 | EMIT 0x0F EMIT 0xD4 EMIT 0xDC \ |
| borlanic | 0:02dd72d1d465 | 774 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x08 \ |
| borlanic | 0:02dd72d1d465 | 775 | EMIT 0x0F EMIT 0xD4 EMIT 0xEE \ |
| borlanic | 0:02dd72d1d465 | 776 | EMIT 0x0F EMIT 0x6E EMIT 0x67 EMIT 0x0C \ |
| borlanic | 0:02dd72d1d465 | 777 | EMIT 0x0F EMIT 0xD4 EMIT 0xFC \ |
| borlanic | 0:02dd72d1d465 | 778 | EMIT 0x0F EMIT 0x7E EMIT 0x0F \ |
| borlanic | 0:02dd72d1d465 | 779 | EMIT 0x0F EMIT 0x6E EMIT 0x56 EMIT 0x10 \ |
| borlanic | 0:02dd72d1d465 | 780 | EMIT 0x0F EMIT 0xF4 EMIT 0xD0 \ |
| borlanic | 0:02dd72d1d465 | 781 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 782 | EMIT 0x0F EMIT 0x6E EMIT 0x66 EMIT 0x14 \ |
| borlanic | 0:02dd72d1d465 | 783 | EMIT 0x0F EMIT 0xF4 EMIT 0xE0 \ |
| borlanic | 0:02dd72d1d465 | 784 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| borlanic | 0:02dd72d1d465 | 785 | EMIT 0x0F EMIT 0x6E EMIT 0x76 EMIT 0x18 \ |
| borlanic | 0:02dd72d1d465 | 786 | EMIT 0x0F EMIT 0xF4 EMIT 0xF0 \ |
| borlanic | 0:02dd72d1d465 | 787 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x04 \ |
| borlanic | 0:02dd72d1d465 | 788 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 789 | EMIT 0x0F EMIT 0x6E EMIT 0x5E EMIT 0x1C \ |
| borlanic | 0:02dd72d1d465 | 790 | EMIT 0x0F EMIT 0xF4 EMIT 0xD8 \ |
| borlanic | 0:02dd72d1d465 | 791 | EMIT 0x0F EMIT 0xD4 EMIT 0xCD \ |
| borlanic | 0:02dd72d1d465 | 792 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x10 \ |
| borlanic | 0:02dd72d1d465 | 793 | EMIT 0x0F EMIT 0xD4 EMIT 0xD5 \ |
| borlanic | 0:02dd72d1d465 | 794 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x08 \ |
| borlanic | 0:02dd72d1d465 | 795 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 796 | EMIT 0x0F EMIT 0xD4 EMIT 0xCF \ |
| borlanic | 0:02dd72d1d465 | 797 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x14 \ |
| borlanic | 0:02dd72d1d465 | 798 | EMIT 0x0F EMIT 0xD4 EMIT 0xE5 \ |
| borlanic | 0:02dd72d1d465 | 799 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x0C \ |
| borlanic | 0:02dd72d1d465 | 800 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 801 | EMIT 0x0F EMIT 0xD4 EMIT 0xCA \ |
| borlanic | 0:02dd72d1d465 | 802 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x18 \ |
| borlanic | 0:02dd72d1d465 | 803 | EMIT 0x0F EMIT 0xD4 EMIT 0xF5 \ |
| borlanic | 0:02dd72d1d465 | 804 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x10 \ |
| borlanic | 0:02dd72d1d465 | 805 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 806 | EMIT 0x0F EMIT 0xD4 EMIT 0xCC \ |
| borlanic | 0:02dd72d1d465 | 807 | EMIT 0x0F EMIT 0x6E EMIT 0x6F EMIT 0x1C \ |
| borlanic | 0:02dd72d1d465 | 808 | EMIT 0x0F EMIT 0xD4 EMIT 0xDD \ |
| borlanic | 0:02dd72d1d465 | 809 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x14 \ |
| borlanic | 0:02dd72d1d465 | 810 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 811 | EMIT 0x0F EMIT 0xD4 EMIT 0xCE \ |
| borlanic | 0:02dd72d1d465 | 812 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x18 \ |
| borlanic | 0:02dd72d1d465 | 813 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 814 | EMIT 0x0F EMIT 0xD4 EMIT 0xCB \ |
| borlanic | 0:02dd72d1d465 | 815 | EMIT 0x0F EMIT 0x7E EMIT 0x4F EMIT 0x1C \ |
| borlanic | 0:02dd72d1d465 | 816 | EMIT 0x83 EMIT 0xC7 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 817 | EMIT 0x83 EMIT 0xC6 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 818 | EMIT 0x0F EMIT 0x73 EMIT 0xD1 EMIT 0x20 \ |
| borlanic | 0:02dd72d1d465 | 819 | EMIT 0x0F EMIT 0x7E EMIT 0xC9 |
| borlanic | 0:02dd72d1d465 | 820 | |
| borlanic | 0:02dd72d1d465 | 821 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 822 | EMIT 0x0F EMIT 0x77 \ |
| borlanic | 0:02dd72d1d465 | 823 | __asm mov c, ecx \ |
| borlanic | 0:02dd72d1d465 | 824 | __asm mov d, edi \ |
| borlanic | 0:02dd72d1d465 | 825 | __asm mov s, esi \ |
| borlanic | 0:02dd72d1d465 | 826 | |
| borlanic | 0:02dd72d1d465 | 827 | #else |
| borlanic | 0:02dd72d1d465 | 828 | |
| borlanic | 0:02dd72d1d465 | 829 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 830 | __asm mov c, ecx \ |
| borlanic | 0:02dd72d1d465 | 831 | __asm mov d, edi \ |
| borlanic | 0:02dd72d1d465 | 832 | __asm mov s, esi \ |
| borlanic | 0:02dd72d1d465 | 833 | |
| borlanic | 0:02dd72d1d465 | 834 | #endif /* SSE2 */ |
| borlanic | 0:02dd72d1d465 | 835 | #endif /* MSVC */ |
| borlanic | 0:02dd72d1d465 | 836 | |
| borlanic | 0:02dd72d1d465 | 837 | #endif /* MBEDTLS_HAVE_ASM */ |
| borlanic | 0:02dd72d1d465 | 838 | |
| borlanic | 0:02dd72d1d465 | 839 | #if !defined(MULADDC_CORE) |
| borlanic | 0:02dd72d1d465 | 840 | #if defined(MBEDTLS_HAVE_UDBL) |
| borlanic | 0:02dd72d1d465 | 841 | |
| borlanic | 0:02dd72d1d465 | 842 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 843 | { \ |
| borlanic | 0:02dd72d1d465 | 844 | mbedtls_t_udbl r; \ |
| borlanic | 0:02dd72d1d465 | 845 | mbedtls_mpi_uint r0, r1; |
| borlanic | 0:02dd72d1d465 | 846 | |
| borlanic | 0:02dd72d1d465 | 847 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 848 | r = *(s++) * (mbedtls_t_udbl) b; \ |
| borlanic | 0:02dd72d1d465 | 849 | r0 = (mbedtls_mpi_uint) r; \ |
| borlanic | 0:02dd72d1d465 | 850 | r1 = (mbedtls_mpi_uint)( r >> biL ); \ |
| borlanic | 0:02dd72d1d465 | 851 | r0 += c; r1 += (r0 < c); \ |
| borlanic | 0:02dd72d1d465 | 852 | r0 += *d; r1 += (r0 < *d); \ |
| borlanic | 0:02dd72d1d465 | 853 | c = r1; *(d++) = r0; |
| borlanic | 0:02dd72d1d465 | 854 | |
| borlanic | 0:02dd72d1d465 | 855 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 856 | } |
| borlanic | 0:02dd72d1d465 | 857 | |
| borlanic | 0:02dd72d1d465 | 858 | #else |
| borlanic | 0:02dd72d1d465 | 859 | #define MULADDC_INIT \ |
| borlanic | 0:02dd72d1d465 | 860 | { \ |
| borlanic | 0:02dd72d1d465 | 861 | mbedtls_mpi_uint s0, s1, b0, b1; \ |
| borlanic | 0:02dd72d1d465 | 862 | mbedtls_mpi_uint r0, r1, rx, ry; \ |
| borlanic | 0:02dd72d1d465 | 863 | b0 = ( b << biH ) >> biH; \ |
| borlanic | 0:02dd72d1d465 | 864 | b1 = ( b >> biH ); |
| borlanic | 0:02dd72d1d465 | 865 | |
| borlanic | 0:02dd72d1d465 | 866 | #define MULADDC_CORE \ |
| borlanic | 0:02dd72d1d465 | 867 | s0 = ( *s << biH ) >> biH; \ |
| borlanic | 0:02dd72d1d465 | 868 | s1 = ( *s >> biH ); s++; \ |
| borlanic | 0:02dd72d1d465 | 869 | rx = s0 * b1; r0 = s0 * b0; \ |
| borlanic | 0:02dd72d1d465 | 870 | ry = s1 * b0; r1 = s1 * b1; \ |
| borlanic | 0:02dd72d1d465 | 871 | r1 += ( rx >> biH ); \ |
| borlanic | 0:02dd72d1d465 | 872 | r1 += ( ry >> biH ); \ |
| borlanic | 0:02dd72d1d465 | 873 | rx <<= biH; ry <<= biH; \ |
| borlanic | 0:02dd72d1d465 | 874 | r0 += rx; r1 += (r0 < rx); \ |
| borlanic | 0:02dd72d1d465 | 875 | r0 += ry; r1 += (r0 < ry); \ |
| borlanic | 0:02dd72d1d465 | 876 | r0 += c; r1 += (r0 < c); \ |
| borlanic | 0:02dd72d1d465 | 877 | r0 += *d; r1 += (r0 < *d); \ |
| borlanic | 0:02dd72d1d465 | 878 | c = r1; *(d++) = r0; |
| borlanic | 0:02dd72d1d465 | 879 | |
| borlanic | 0:02dd72d1d465 | 880 | #define MULADDC_STOP \ |
| borlanic | 0:02dd72d1d465 | 881 | } |
| borlanic | 0:02dd72d1d465 | 882 | |
| borlanic | 0:02dd72d1d465 | 883 | #endif /* C (generic) */ |
| borlanic | 0:02dd72d1d465 | 884 | #endif /* C (longlong) */ |
| borlanic | 0:02dd72d1d465 | 885 | |
| borlanic | 0:02dd72d1d465 | 886 | #endif /* bn_mul.h */ |