x86_64-gf2m.S revision 305153
1/* $FreeBSD: stable/11/secure/lib/libcrypto/amd64/x86_64-gf2m.S 305153 2016-08-31 20:33:59Z jkim $ */ 2/* Do not modify. This file is auto-generated from x86_64-gf2m.pl. */ 3.text 4 5.type _mul_1x1,@function 6.align 16 7_mul_1x1: 8 subq $128+8,%rsp 9 movq $-1,%r9 10 leaq (%rax,%rax,1),%rsi 11 shrq $3,%r9 12 leaq (,%rax,4),%rdi 13 andq %rax,%r9 14 leaq (,%rax,8),%r12 15 sarq $63,%rax 16 leaq (%r9,%r9,1),%r10 17 sarq $63,%rsi 18 leaq (,%r9,4),%r11 19 andq %rbp,%rax 20 sarq $63,%rdi 21 movq %rax,%rdx 22 shlq $63,%rax 23 andq %rbp,%rsi 24 shrq $1,%rdx 25 movq %rsi,%rcx 26 shlq $62,%rsi 27 andq %rbp,%rdi 28 shrq $2,%rcx 29 xorq %rsi,%rax 30 movq %rdi,%rbx 31 shlq $61,%rdi 32 xorq %rcx,%rdx 33 shrq $3,%rbx 34 xorq %rdi,%rax 35 xorq %rbx,%rdx 36 37 movq %r9,%r13 38 movq $0,0(%rsp) 39 xorq %r10,%r13 40 movq %r9,8(%rsp) 41 movq %r11,%r14 42 movq %r10,16(%rsp) 43 xorq %r12,%r14 44 movq %r13,24(%rsp) 45 46 xorq %r11,%r9 47 movq %r11,32(%rsp) 48 xorq %r11,%r10 49 movq %r9,40(%rsp) 50 xorq %r11,%r13 51 movq %r10,48(%rsp) 52 xorq %r14,%r9 53 movq %r13,56(%rsp) 54 xorq %r14,%r10 55 56 movq %r12,64(%rsp) 57 xorq %r14,%r13 58 movq %r9,72(%rsp) 59 xorq %r11,%r9 60 movq %r10,80(%rsp) 61 xorq %r11,%r10 62 movq %r13,88(%rsp) 63 64 xorq %r11,%r13 65 movq %r14,96(%rsp) 66 movq %r8,%rsi 67 movq %r9,104(%rsp) 68 andq %rbp,%rsi 69 movq %r10,112(%rsp) 70 shrq $4,%rbp 71 movq %r13,120(%rsp) 72 movq %r8,%rdi 73 andq %rbp,%rdi 74 shrq $4,%rbp 75 76 movq (%rsp,%rsi,8),%xmm0 77 movq %r8,%rsi 78 andq %rbp,%rsi 79 shrq $4,%rbp 80 movq (%rsp,%rdi,8),%rcx 81 movq %r8,%rdi 82 movq %rcx,%rbx 83 shlq $4,%rcx 84 andq %rbp,%rdi 85 movq (%rsp,%rsi,8),%xmm1 86 shrq $60,%rbx 87 xorq %rcx,%rax 88 pslldq $1,%xmm1 89 movq %r8,%rsi 90 shrq $4,%rbp 91 xorq %rbx,%rdx 92 andq %rbp,%rsi 93 shrq $4,%rbp 94 pxor %xmm1,%xmm0 95 movq (%rsp,%rdi,8),%rcx 96 movq %r8,%rdi 97 movq %rcx,%rbx 98 shlq $12,%rcx 99 andq %rbp,%rdi 100 movq (%rsp,%rsi,8),%xmm1 101 shrq $52,%rbx 102 xorq %rcx,%rax 103 pslldq $2,%xmm1 104 movq %r8,%rsi 105 shrq $4,%rbp 106 xorq %rbx,%rdx 107 andq %rbp,%rsi 108 shrq $4,%rbp 109 pxor %xmm1,%xmm0 110 movq (%rsp,%rdi,8),%rcx 111 movq %r8,%rdi 112 movq %rcx,%rbx 113 shlq $20,%rcx 114 andq %rbp,%rdi 115 movq (%rsp,%rsi,8),%xmm1 116 shrq $44,%rbx 117 xorq %rcx,%rax 118 pslldq $3,%xmm1 119 movq %r8,%rsi 120 shrq $4,%rbp 121 xorq %rbx,%rdx 122 andq %rbp,%rsi 123 shrq $4,%rbp 124 pxor %xmm1,%xmm0 125 movq (%rsp,%rdi,8),%rcx 126 movq %r8,%rdi 127 movq %rcx,%rbx 128 shlq $28,%rcx 129 andq %rbp,%rdi 130 movq (%rsp,%rsi,8),%xmm1 131 shrq $36,%rbx 132 xorq %rcx,%rax 133 pslldq $4,%xmm1 134 movq %r8,%rsi 135 shrq $4,%rbp 136 xorq %rbx,%rdx 137 andq %rbp,%rsi 138 shrq $4,%rbp 139 pxor %xmm1,%xmm0 140 movq (%rsp,%rdi,8),%rcx 141 movq %r8,%rdi 142 movq %rcx,%rbx 143 shlq $36,%rcx 144 andq %rbp,%rdi 145 movq (%rsp,%rsi,8),%xmm1 146 shrq $28,%rbx 147 xorq %rcx,%rax 148 pslldq $5,%xmm1 149 movq %r8,%rsi 150 shrq $4,%rbp 151 xorq %rbx,%rdx 152 andq %rbp,%rsi 153 shrq $4,%rbp 154 pxor %xmm1,%xmm0 155 movq (%rsp,%rdi,8),%rcx 156 movq %r8,%rdi 157 movq %rcx,%rbx 158 shlq $44,%rcx 159 andq %rbp,%rdi 160 movq (%rsp,%rsi,8),%xmm1 161 shrq $20,%rbx 162 xorq %rcx,%rax 163 pslldq $6,%xmm1 164 movq %r8,%rsi 165 shrq $4,%rbp 166 xorq %rbx,%rdx 167 andq %rbp,%rsi 168 shrq $4,%rbp 169 pxor %xmm1,%xmm0 170 movq (%rsp,%rdi,8),%rcx 171 movq %r8,%rdi 172 movq %rcx,%rbx 173 shlq $52,%rcx 174 andq %rbp,%rdi 175 movq (%rsp,%rsi,8),%xmm1 176 shrq $12,%rbx 177 xorq %rcx,%rax 178 pslldq $7,%xmm1 179 movq %r8,%rsi 180 shrq $4,%rbp 181 xorq %rbx,%rdx 182 andq %rbp,%rsi 183 shrq $4,%rbp 184 pxor %xmm1,%xmm0 185 movq (%rsp,%rdi,8),%rcx 186 movq %rcx,%rbx 187 shlq $60,%rcx 188.byte 102,72,15,126,198 189 shrq $4,%rbx 190 xorq %rcx,%rax 191 psrldq $8,%xmm0 192 xorq %rbx,%rdx 193.byte 102,72,15,126,199 194 xorq %rsi,%rax 195 xorq %rdi,%rdx 196 197 addq $128+8,%rsp 198 .byte 0xf3,0xc3 199.Lend_mul_1x1: 200.size _mul_1x1,.-_mul_1x1 201 202.globl bn_GF2m_mul_2x2 203.type bn_GF2m_mul_2x2,@function 204.align 16 205bn_GF2m_mul_2x2: 206 movq OPENSSL_ia32cap_P(%rip),%rax 207 btq $33,%rax 208 jnc .Lvanilla_mul_2x2 209 210.byte 102,72,15,110,198 211.byte 102,72,15,110,201 212.byte 102,72,15,110,210 213.byte 102,73,15,110,216 214 movdqa %xmm0,%xmm4 215 movdqa %xmm1,%xmm5 216.byte 102,15,58,68,193,0 217 pxor %xmm2,%xmm4 218 pxor %xmm3,%xmm5 219.byte 102,15,58,68,211,0 220.byte 102,15,58,68,229,0 221 xorps %xmm0,%xmm4 222 xorps %xmm2,%xmm4 223 movdqa %xmm4,%xmm5 224 pslldq $8,%xmm4 225 psrldq $8,%xmm5 226 pxor %xmm4,%xmm2 227 pxor %xmm5,%xmm0 228 movdqu %xmm2,0(%rdi) 229 movdqu %xmm0,16(%rdi) 230 .byte 0xf3,0xc3 231 232.align 16 233.Lvanilla_mul_2x2: 234 leaq -136(%rsp),%rsp 235 movq %r14,80(%rsp) 236 movq %r13,88(%rsp) 237 movq %r12,96(%rsp) 238 movq %rbp,104(%rsp) 239 movq %rbx,112(%rsp) 240.Lbody_mul_2x2: 241 movq %rdi,32(%rsp) 242 movq %rsi,40(%rsp) 243 movq %rdx,48(%rsp) 244 movq %rcx,56(%rsp) 245 movq %r8,64(%rsp) 246 247 movq $0xf,%r8 248 movq %rsi,%rax 249 movq %rcx,%rbp 250 call _mul_1x1 251 movq %rax,16(%rsp) 252 movq %rdx,24(%rsp) 253 254 movq 48(%rsp),%rax 255 movq 64(%rsp),%rbp 256 call _mul_1x1 257 movq %rax,0(%rsp) 258 movq %rdx,8(%rsp) 259 260 movq 40(%rsp),%rax 261 movq 56(%rsp),%rbp 262 xorq 48(%rsp),%rax 263 xorq 64(%rsp),%rbp 264 call _mul_1x1 265 movq 0(%rsp),%rbx 266 movq 8(%rsp),%rcx 267 movq 16(%rsp),%rdi 268 movq 24(%rsp),%rsi 269 movq 32(%rsp),%rbp 270 271 xorq %rdx,%rax 272 xorq %rcx,%rdx 273 xorq %rbx,%rax 274 movq %rbx,0(%rbp) 275 xorq %rdi,%rdx 276 movq %rsi,24(%rbp) 277 xorq %rsi,%rax 278 xorq %rsi,%rdx 279 xorq %rdx,%rax 280 movq %rdx,16(%rbp) 281 movq %rax,8(%rbp) 282 283 movq 80(%rsp),%r14 284 movq 88(%rsp),%r13 285 movq 96(%rsp),%r12 286 movq 104(%rsp),%rbp 287 movq 112(%rsp),%rbx 288 leaq 136(%rsp),%rsp 289 .byte 0xf3,0xc3 290.Lend_mul_2x2: 291.size bn_GF2m_mul_2x2,.-bn_GF2m_mul_2x2 292.byte 71,70,40,50,94,109,41,32,77,117,108,116,105,112,108,105,99,97,116,105,111,110,32,102,111,114,32,120,56,54,95,54,52,44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111,114,103,62,0 293.align 16 294