OLD | NEW |
1 #if defined(__x86_64__) | 1 #if defined(__x86_64__) |
2 .text | 2 .text |
3 | 3 |
4 .extern OPENSSL_ia32cap_P | 4 .extern OPENSSL_ia32cap_P |
5 .hidden OPENSSL_ia32cap_P | 5 .hidden OPENSSL_ia32cap_P |
6 .globl sha256_block_data_order | 6 .globl sha256_block_data_order |
7 .hidden sha256_block_data_order | 7 .hidden sha256_block_data_order |
8 .type sha256_block_data_order,@function | 8 .type sha256_block_data_order,@function |
9 .align 16 | 9 .align 16 |
10 sha256_block_data_order: | 10 sha256_block_data_order: |
11 leaq OPENSSL_ia32cap_P(%rip),%r11 | 11 leaq OPENSSL_ia32cap_P(%rip),%r11 |
12 movl 0(%r11),%r9d | 12 movl 0(%r11),%r9d |
13 movl 4(%r11),%r10d | 13 movl 4(%r11),%r10d |
14 movl 8(%r11),%r11d | 14 movl 8(%r11),%r11d |
15 andl $1073741824,%r9d | |
16 andl $268435968,%r10d | |
17 orl %r9d,%r10d | |
18 cmpl $1342177792,%r10d | |
19 je .Lavx_shortcut | |
20 testl $512,%r10d | 15 testl $512,%r10d |
21 jnz .Lssse3_shortcut | 16 jnz .Lssse3_shortcut |
22 pushq %rbx | 17 pushq %rbx |
23 pushq %rbp | 18 pushq %rbp |
24 pushq %r12 | 19 pushq %r12 |
25 pushq %r13 | 20 pushq %r13 |
26 pushq %r14 | 21 pushq %r14 |
27 pushq %r15 | 22 pushq %r15 |
28 movq %rsp,%r11 | 23 movq %rsp,%r11 |
29 shlq $4,%rdx | 24 shlq $4,%rdx |
(...skipping 2809 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
2839 movq (%rsi),%r15 | 2834 movq (%rsi),%r15 |
2840 movq 8(%rsi),%r14 | 2835 movq 8(%rsi),%r14 |
2841 movq 16(%rsi),%r13 | 2836 movq 16(%rsi),%r13 |
2842 movq 24(%rsi),%r12 | 2837 movq 24(%rsi),%r12 |
2843 movq 32(%rsi),%rbp | 2838 movq 32(%rsi),%rbp |
2844 movq 40(%rsi),%rbx | 2839 movq 40(%rsi),%rbx |
2845 leaq 48(%rsi),%rsp | 2840 leaq 48(%rsi),%rsp |
2846 .Lepilogue_ssse3: | 2841 .Lepilogue_ssse3: |
2847 .byte 0xf3,0xc3 | 2842 .byte 0xf3,0xc3 |
2848 .size sha256_block_data_order_ssse3,.-sha256_block_data_order_ssse3 | 2843 .size sha256_block_data_order_ssse3,.-sha256_block_data_order_ssse3 |
2849 .type sha256_block_data_order_avx,@function | |
2850 .align 64 | |
2851 sha256_block_data_order_avx: | |
2852 .Lavx_shortcut: | |
2853 pushq %rbx | |
2854 pushq %rbp | |
2855 pushq %r12 | |
2856 pushq %r13 | |
2857 pushq %r14 | |
2858 pushq %r15 | |
2859 movq %rsp,%r11 | |
2860 shlq $4,%rdx | |
2861 subq $96,%rsp | |
2862 leaq (%rsi,%rdx,4),%rdx | |
2863 andq $-64,%rsp | |
2864 movq %rdi,64+0(%rsp) | |
2865 movq %rsi,64+8(%rsp) | |
2866 movq %rdx,64+16(%rsp) | |
2867 movq %r11,64+24(%rsp) | |
2868 .Lprologue_avx: | |
2869 | |
2870 vzeroupper | |
2871 movl 0(%rdi),%eax | |
2872 movl 4(%rdi),%ebx | |
2873 movl 8(%rdi),%ecx | |
2874 movl 12(%rdi),%edx | |
2875 movl 16(%rdi),%r8d | |
2876 movl 20(%rdi),%r9d | |
2877 movl 24(%rdi),%r10d | |
2878 movl 28(%rdi),%r11d | |
2879 vmovdqa K256+512+32(%rip),%xmm8 | |
2880 vmovdqa K256+512+64(%rip),%xmm9 | |
2881 jmp .Lloop_avx | |
2882 .align 16 | |
2883 .Lloop_avx: | |
2884 vmovdqa K256+512(%rip),%xmm7 | |
2885 vmovdqu 0(%rsi),%xmm0 | |
2886 vmovdqu 16(%rsi),%xmm1 | |
2887 vmovdqu 32(%rsi),%xmm2 | |
2888 vmovdqu 48(%rsi),%xmm3 | |
2889 vpshufb %xmm7,%xmm0,%xmm0 | |
2890 leaq K256(%rip),%rbp | |
2891 vpshufb %xmm7,%xmm1,%xmm1 | |
2892 vpshufb %xmm7,%xmm2,%xmm2 | |
2893 vpaddd 0(%rbp),%xmm0,%xmm4 | |
2894 vpshufb %xmm7,%xmm3,%xmm3 | |
2895 vpaddd 32(%rbp),%xmm1,%xmm5 | |
2896 vpaddd 64(%rbp),%xmm2,%xmm6 | |
2897 vpaddd 96(%rbp),%xmm3,%xmm7 | |
2898 vmovdqa %xmm4,0(%rsp) | |
2899 movl %eax,%r14d | |
2900 vmovdqa %xmm5,16(%rsp) | |
2901 movl %ebx,%edi | |
2902 vmovdqa %xmm6,32(%rsp) | |
2903 xorl %ecx,%edi | |
2904 vmovdqa %xmm7,48(%rsp) | |
2905 movl %r8d,%r13d | |
2906 jmp .Lavx_00_47 | |
2907 | |
2908 .align 16 | |
2909 .Lavx_00_47: | |
2910 subq $-128,%rbp | |
2911 vpalignr $4,%xmm0,%xmm1,%xmm4 | |
2912 shrdl $14,%r13d,%r13d | |
2913 movl %r14d,%eax | |
2914 movl %r9d,%r12d | |
2915 vpalignr $4,%xmm2,%xmm3,%xmm7 | |
2916 shrdl $9,%r14d,%r14d | |
2917 xorl %r8d,%r13d | |
2918 xorl %r10d,%r12d | |
2919 vpsrld $7,%xmm4,%xmm6 | |
2920 shrdl $5,%r13d,%r13d | |
2921 xorl %eax,%r14d | |
2922 andl %r8d,%r12d | |
2923 vpaddd %xmm7,%xmm0,%xmm0 | |
2924 xorl %r8d,%r13d | |
2925 addl 0(%rsp),%r11d | |
2926 movl %eax,%r15d | |
2927 vpsrld $3,%xmm4,%xmm7 | |
2928 xorl %r10d,%r12d | |
2929 shrdl $11,%r14d,%r14d | |
2930 xorl %ebx,%r15d | |
2931 vpslld $14,%xmm4,%xmm5 | |
2932 addl %r12d,%r11d | |
2933 shrdl $6,%r13d,%r13d | |
2934 andl %r15d,%edi | |
2935 vpxor %xmm6,%xmm7,%xmm4 | |
2936 xorl %eax,%r14d | |
2937 addl %r13d,%r11d | |
2938 xorl %ebx,%edi | |
2939 vpshufd $250,%xmm3,%xmm7 | |
2940 shrdl $2,%r14d,%r14d | |
2941 addl %r11d,%edx | |
2942 addl %edi,%r11d | |
2943 vpsrld $11,%xmm6,%xmm6 | |
2944 movl %edx,%r13d | |
2945 addl %r11d,%r14d | |
2946 shrdl $14,%r13d,%r13d | |
2947 vpxor %xmm5,%xmm4,%xmm4 | |
2948 movl %r14d,%r11d | |
2949 movl %r8d,%r12d | |
2950 shrdl $9,%r14d,%r14d | |
2951 vpslld $11,%xmm5,%xmm5 | |
2952 xorl %edx,%r13d | |
2953 xorl %r9d,%r12d | |
2954 shrdl $5,%r13d,%r13d | |
2955 vpxor %xmm6,%xmm4,%xmm4 | |
2956 xorl %r11d,%r14d | |
2957 andl %edx,%r12d | |
2958 xorl %edx,%r13d | |
2959 vpsrld $10,%xmm7,%xmm6 | |
2960 addl 4(%rsp),%r10d | |
2961 movl %r11d,%edi | |
2962 xorl %r9d,%r12d | |
2963 vpxor %xmm5,%xmm4,%xmm4 | |
2964 shrdl $11,%r14d,%r14d | |
2965 xorl %eax,%edi | |
2966 addl %r12d,%r10d | |
2967 vpsrlq $17,%xmm7,%xmm7 | |
2968 shrdl $6,%r13d,%r13d | |
2969 andl %edi,%r15d | |
2970 xorl %r11d,%r14d | |
2971 vpaddd %xmm4,%xmm0,%xmm0 | |
2972 addl %r13d,%r10d | |
2973 xorl %eax,%r15d | |
2974 shrdl $2,%r14d,%r14d | |
2975 vpxor %xmm7,%xmm6,%xmm6 | |
2976 addl %r10d,%ecx | |
2977 addl %r15d,%r10d | |
2978 movl %ecx,%r13d | |
2979 vpsrlq $2,%xmm7,%xmm7 | |
2980 addl %r10d,%r14d | |
2981 shrdl $14,%r13d,%r13d | |
2982 movl %r14d,%r10d | |
2983 vpxor %xmm7,%xmm6,%xmm6 | |
2984 movl %edx,%r12d | |
2985 shrdl $9,%r14d,%r14d | |
2986 xorl %ecx,%r13d | |
2987 vpshufb %xmm8,%xmm6,%xmm6 | |
2988 xorl %r8d,%r12d | |
2989 shrdl $5,%r13d,%r13d | |
2990 xorl %r10d,%r14d | |
2991 vpaddd %xmm6,%xmm0,%xmm0 | |
2992 andl %ecx,%r12d | |
2993 xorl %ecx,%r13d | |
2994 addl 8(%rsp),%r9d | |
2995 vpshufd $80,%xmm0,%xmm7 | |
2996 movl %r10d,%r15d | |
2997 xorl %r8d,%r12d | |
2998 shrdl $11,%r14d,%r14d | |
2999 vpsrld $10,%xmm7,%xmm6 | |
3000 xorl %r11d,%r15d | |
3001 addl %r12d,%r9d | |
3002 shrdl $6,%r13d,%r13d | |
3003 vpsrlq $17,%xmm7,%xmm7 | |
3004 andl %r15d,%edi | |
3005 xorl %r10d,%r14d | |
3006 addl %r13d,%r9d | |
3007 vpxor %xmm7,%xmm6,%xmm6 | |
3008 xorl %r11d,%edi | |
3009 shrdl $2,%r14d,%r14d | |
3010 addl %r9d,%ebx | |
3011 vpsrlq $2,%xmm7,%xmm7 | |
3012 addl %edi,%r9d | |
3013 movl %ebx,%r13d | |
3014 addl %r9d,%r14d | |
3015 vpxor %xmm7,%xmm6,%xmm6 | |
3016 shrdl $14,%r13d,%r13d | |
3017 movl %r14d,%r9d | |
3018 movl %ecx,%r12d | |
3019 vpshufb %xmm9,%xmm6,%xmm6 | |
3020 shrdl $9,%r14d,%r14d | |
3021 xorl %ebx,%r13d | |
3022 xorl %edx,%r12d | |
3023 vpaddd %xmm6,%xmm0,%xmm0 | |
3024 shrdl $5,%r13d,%r13d | |
3025 xorl %r9d,%r14d | |
3026 andl %ebx,%r12d | |
3027 vpaddd 0(%rbp),%xmm0,%xmm6 | |
3028 xorl %ebx,%r13d | |
3029 addl 12(%rsp),%r8d | |
3030 movl %r9d,%edi | |
3031 xorl %edx,%r12d | |
3032 shrdl $11,%r14d,%r14d | |
3033 xorl %r10d,%edi | |
3034 addl %r12d,%r8d | |
3035 shrdl $6,%r13d,%r13d | |
3036 andl %edi,%r15d | |
3037 xorl %r9d,%r14d | |
3038 addl %r13d,%r8d | |
3039 xorl %r10d,%r15d | |
3040 shrdl $2,%r14d,%r14d | |
3041 addl %r8d,%eax | |
3042 addl %r15d,%r8d | |
3043 movl %eax,%r13d | |
3044 addl %r8d,%r14d | |
3045 vmovdqa %xmm6,0(%rsp) | |
3046 vpalignr $4,%xmm1,%xmm2,%xmm4 | |
3047 shrdl $14,%r13d,%r13d | |
3048 movl %r14d,%r8d | |
3049 movl %ebx,%r12d | |
3050 vpalignr $4,%xmm3,%xmm0,%xmm7 | |
3051 shrdl $9,%r14d,%r14d | |
3052 xorl %eax,%r13d | |
3053 xorl %ecx,%r12d | |
3054 vpsrld $7,%xmm4,%xmm6 | |
3055 shrdl $5,%r13d,%r13d | |
3056 xorl %r8d,%r14d | |
3057 andl %eax,%r12d | |
3058 vpaddd %xmm7,%xmm1,%xmm1 | |
3059 xorl %eax,%r13d | |
3060 addl 16(%rsp),%edx | |
3061 movl %r8d,%r15d | |
3062 vpsrld $3,%xmm4,%xmm7 | |
3063 xorl %ecx,%r12d | |
3064 shrdl $11,%r14d,%r14d | |
3065 xorl %r9d,%r15d | |
3066 vpslld $14,%xmm4,%xmm5 | |
3067 addl %r12d,%edx | |
3068 shrdl $6,%r13d,%r13d | |
3069 andl %r15d,%edi | |
3070 vpxor %xmm6,%xmm7,%xmm4 | |
3071 xorl %r8d,%r14d | |
3072 addl %r13d,%edx | |
3073 xorl %r9d,%edi | |
3074 vpshufd $250,%xmm0,%xmm7 | |
3075 shrdl $2,%r14d,%r14d | |
3076 addl %edx,%r11d | |
3077 addl %edi,%edx | |
3078 vpsrld $11,%xmm6,%xmm6 | |
3079 movl %r11d,%r13d | |
3080 addl %edx,%r14d | |
3081 shrdl $14,%r13d,%r13d | |
3082 vpxor %xmm5,%xmm4,%xmm4 | |
3083 movl %r14d,%edx | |
3084 movl %eax,%r12d | |
3085 shrdl $9,%r14d,%r14d | |
3086 vpslld $11,%xmm5,%xmm5 | |
3087 xorl %r11d,%r13d | |
3088 xorl %ebx,%r12d | |
3089 shrdl $5,%r13d,%r13d | |
3090 vpxor %xmm6,%xmm4,%xmm4 | |
3091 xorl %edx,%r14d | |
3092 andl %r11d,%r12d | |
3093 xorl %r11d,%r13d | |
3094 vpsrld $10,%xmm7,%xmm6 | |
3095 addl 20(%rsp),%ecx | |
3096 movl %edx,%edi | |
3097 xorl %ebx,%r12d | |
3098 vpxor %xmm5,%xmm4,%xmm4 | |
3099 shrdl $11,%r14d,%r14d | |
3100 xorl %r8d,%edi | |
3101 addl %r12d,%ecx | |
3102 vpsrlq $17,%xmm7,%xmm7 | |
3103 shrdl $6,%r13d,%r13d | |
3104 andl %edi,%r15d | |
3105 xorl %edx,%r14d | |
3106 vpaddd %xmm4,%xmm1,%xmm1 | |
3107 addl %r13d,%ecx | |
3108 xorl %r8d,%r15d | |
3109 shrdl $2,%r14d,%r14d | |
3110 vpxor %xmm7,%xmm6,%xmm6 | |
3111 addl %ecx,%r10d | |
3112 addl %r15d,%ecx | |
3113 movl %r10d,%r13d | |
3114 vpsrlq $2,%xmm7,%xmm7 | |
3115 addl %ecx,%r14d | |
3116 shrdl $14,%r13d,%r13d | |
3117 movl %r14d,%ecx | |
3118 vpxor %xmm7,%xmm6,%xmm6 | |
3119 movl %r11d,%r12d | |
3120 shrdl $9,%r14d,%r14d | |
3121 xorl %r10d,%r13d | |
3122 vpshufb %xmm8,%xmm6,%xmm6 | |
3123 xorl %eax,%r12d | |
3124 shrdl $5,%r13d,%r13d | |
3125 xorl %ecx,%r14d | |
3126 vpaddd %xmm6,%xmm1,%xmm1 | |
3127 andl %r10d,%r12d | |
3128 xorl %r10d,%r13d | |
3129 addl 24(%rsp),%ebx | |
3130 vpshufd $80,%xmm1,%xmm7 | |
3131 movl %ecx,%r15d | |
3132 xorl %eax,%r12d | |
3133 shrdl $11,%r14d,%r14d | |
3134 vpsrld $10,%xmm7,%xmm6 | |
3135 xorl %edx,%r15d | |
3136 addl %r12d,%ebx | |
3137 shrdl $6,%r13d,%r13d | |
3138 vpsrlq $17,%xmm7,%xmm7 | |
3139 andl %r15d,%edi | |
3140 xorl %ecx,%r14d | |
3141 addl %r13d,%ebx | |
3142 vpxor %xmm7,%xmm6,%xmm6 | |
3143 xorl %edx,%edi | |
3144 shrdl $2,%r14d,%r14d | |
3145 addl %ebx,%r9d | |
3146 vpsrlq $2,%xmm7,%xmm7 | |
3147 addl %edi,%ebx | |
3148 movl %r9d,%r13d | |
3149 addl %ebx,%r14d | |
3150 vpxor %xmm7,%xmm6,%xmm6 | |
3151 shrdl $14,%r13d,%r13d | |
3152 movl %r14d,%ebx | |
3153 movl %r10d,%r12d | |
3154 vpshufb %xmm9,%xmm6,%xmm6 | |
3155 shrdl $9,%r14d,%r14d | |
3156 xorl %r9d,%r13d | |
3157 xorl %r11d,%r12d | |
3158 vpaddd %xmm6,%xmm1,%xmm1 | |
3159 shrdl $5,%r13d,%r13d | |
3160 xorl %ebx,%r14d | |
3161 andl %r9d,%r12d | |
3162 vpaddd 32(%rbp),%xmm1,%xmm6 | |
3163 xorl %r9d,%r13d | |
3164 addl 28(%rsp),%eax | |
3165 movl %ebx,%edi | |
3166 xorl %r11d,%r12d | |
3167 shrdl $11,%r14d,%r14d | |
3168 xorl %ecx,%edi | |
3169 addl %r12d,%eax | |
3170 shrdl $6,%r13d,%r13d | |
3171 andl %edi,%r15d | |
3172 xorl %ebx,%r14d | |
3173 addl %r13d,%eax | |
3174 xorl %ecx,%r15d | |
3175 shrdl $2,%r14d,%r14d | |
3176 addl %eax,%r8d | |
3177 addl %r15d,%eax | |
3178 movl %r8d,%r13d | |
3179 addl %eax,%r14d | |
3180 vmovdqa %xmm6,16(%rsp) | |
3181 vpalignr $4,%xmm2,%xmm3,%xmm4 | |
3182 shrdl $14,%r13d,%r13d | |
3183 movl %r14d,%eax | |
3184 movl %r9d,%r12d | |
3185 vpalignr $4,%xmm0,%xmm1,%xmm7 | |
3186 shrdl $9,%r14d,%r14d | |
3187 xorl %r8d,%r13d | |
3188 xorl %r10d,%r12d | |
3189 vpsrld $7,%xmm4,%xmm6 | |
3190 shrdl $5,%r13d,%r13d | |
3191 xorl %eax,%r14d | |
3192 andl %r8d,%r12d | |
3193 vpaddd %xmm7,%xmm2,%xmm2 | |
3194 xorl %r8d,%r13d | |
3195 addl 32(%rsp),%r11d | |
3196 movl %eax,%r15d | |
3197 vpsrld $3,%xmm4,%xmm7 | |
3198 xorl %r10d,%r12d | |
3199 shrdl $11,%r14d,%r14d | |
3200 xorl %ebx,%r15d | |
3201 vpslld $14,%xmm4,%xmm5 | |
3202 addl %r12d,%r11d | |
3203 shrdl $6,%r13d,%r13d | |
3204 andl %r15d,%edi | |
3205 vpxor %xmm6,%xmm7,%xmm4 | |
3206 xorl %eax,%r14d | |
3207 addl %r13d,%r11d | |
3208 xorl %ebx,%edi | |
3209 vpshufd $250,%xmm1,%xmm7 | |
3210 shrdl $2,%r14d,%r14d | |
3211 addl %r11d,%edx | |
3212 addl %edi,%r11d | |
3213 vpsrld $11,%xmm6,%xmm6 | |
3214 movl %edx,%r13d | |
3215 addl %r11d,%r14d | |
3216 shrdl $14,%r13d,%r13d | |
3217 vpxor %xmm5,%xmm4,%xmm4 | |
3218 movl %r14d,%r11d | |
3219 movl %r8d,%r12d | |
3220 shrdl $9,%r14d,%r14d | |
3221 vpslld $11,%xmm5,%xmm5 | |
3222 xorl %edx,%r13d | |
3223 xorl %r9d,%r12d | |
3224 shrdl $5,%r13d,%r13d | |
3225 vpxor %xmm6,%xmm4,%xmm4 | |
3226 xorl %r11d,%r14d | |
3227 andl %edx,%r12d | |
3228 xorl %edx,%r13d | |
3229 vpsrld $10,%xmm7,%xmm6 | |
3230 addl 36(%rsp),%r10d | |
3231 movl %r11d,%edi | |
3232 xorl %r9d,%r12d | |
3233 vpxor %xmm5,%xmm4,%xmm4 | |
3234 shrdl $11,%r14d,%r14d | |
3235 xorl %eax,%edi | |
3236 addl %r12d,%r10d | |
3237 vpsrlq $17,%xmm7,%xmm7 | |
3238 shrdl $6,%r13d,%r13d | |
3239 andl %edi,%r15d | |
3240 xorl %r11d,%r14d | |
3241 vpaddd %xmm4,%xmm2,%xmm2 | |
3242 addl %r13d,%r10d | |
3243 xorl %eax,%r15d | |
3244 shrdl $2,%r14d,%r14d | |
3245 vpxor %xmm7,%xmm6,%xmm6 | |
3246 addl %r10d,%ecx | |
3247 addl %r15d,%r10d | |
3248 movl %ecx,%r13d | |
3249 vpsrlq $2,%xmm7,%xmm7 | |
3250 addl %r10d,%r14d | |
3251 shrdl $14,%r13d,%r13d | |
3252 movl %r14d,%r10d | |
3253 vpxor %xmm7,%xmm6,%xmm6 | |
3254 movl %edx,%r12d | |
3255 shrdl $9,%r14d,%r14d | |
3256 xorl %ecx,%r13d | |
3257 vpshufb %xmm8,%xmm6,%xmm6 | |
3258 xorl %r8d,%r12d | |
3259 shrdl $5,%r13d,%r13d | |
3260 xorl %r10d,%r14d | |
3261 vpaddd %xmm6,%xmm2,%xmm2 | |
3262 andl %ecx,%r12d | |
3263 xorl %ecx,%r13d | |
3264 addl 40(%rsp),%r9d | |
3265 vpshufd $80,%xmm2,%xmm7 | |
3266 movl %r10d,%r15d | |
3267 xorl %r8d,%r12d | |
3268 shrdl $11,%r14d,%r14d | |
3269 vpsrld $10,%xmm7,%xmm6 | |
3270 xorl %r11d,%r15d | |
3271 addl %r12d,%r9d | |
3272 shrdl $6,%r13d,%r13d | |
3273 vpsrlq $17,%xmm7,%xmm7 | |
3274 andl %r15d,%edi | |
3275 xorl %r10d,%r14d | |
3276 addl %r13d,%r9d | |
3277 vpxor %xmm7,%xmm6,%xmm6 | |
3278 xorl %r11d,%edi | |
3279 shrdl $2,%r14d,%r14d | |
3280 addl %r9d,%ebx | |
3281 vpsrlq $2,%xmm7,%xmm7 | |
3282 addl %edi,%r9d | |
3283 movl %ebx,%r13d | |
3284 addl %r9d,%r14d | |
3285 vpxor %xmm7,%xmm6,%xmm6 | |
3286 shrdl $14,%r13d,%r13d | |
3287 movl %r14d,%r9d | |
3288 movl %ecx,%r12d | |
3289 vpshufb %xmm9,%xmm6,%xmm6 | |
3290 shrdl $9,%r14d,%r14d | |
3291 xorl %ebx,%r13d | |
3292 xorl %edx,%r12d | |
3293 vpaddd %xmm6,%xmm2,%xmm2 | |
3294 shrdl $5,%r13d,%r13d | |
3295 xorl %r9d,%r14d | |
3296 andl %ebx,%r12d | |
3297 vpaddd 64(%rbp),%xmm2,%xmm6 | |
3298 xorl %ebx,%r13d | |
3299 addl 44(%rsp),%r8d | |
3300 movl %r9d,%edi | |
3301 xorl %edx,%r12d | |
3302 shrdl $11,%r14d,%r14d | |
3303 xorl %r10d,%edi | |
3304 addl %r12d,%r8d | |
3305 shrdl $6,%r13d,%r13d | |
3306 andl %edi,%r15d | |
3307 xorl %r9d,%r14d | |
3308 addl %r13d,%r8d | |
3309 xorl %r10d,%r15d | |
3310 shrdl $2,%r14d,%r14d | |
3311 addl %r8d,%eax | |
3312 addl %r15d,%r8d | |
3313 movl %eax,%r13d | |
3314 addl %r8d,%r14d | |
3315 vmovdqa %xmm6,32(%rsp) | |
3316 vpalignr $4,%xmm3,%xmm0,%xmm4 | |
3317 shrdl $14,%r13d,%r13d | |
3318 movl %r14d,%r8d | |
3319 movl %ebx,%r12d | |
3320 vpalignr $4,%xmm1,%xmm2,%xmm7 | |
3321 shrdl $9,%r14d,%r14d | |
3322 xorl %eax,%r13d | |
3323 xorl %ecx,%r12d | |
3324 vpsrld $7,%xmm4,%xmm6 | |
3325 shrdl $5,%r13d,%r13d | |
3326 xorl %r8d,%r14d | |
3327 andl %eax,%r12d | |
3328 vpaddd %xmm7,%xmm3,%xmm3 | |
3329 xorl %eax,%r13d | |
3330 addl 48(%rsp),%edx | |
3331 movl %r8d,%r15d | |
3332 vpsrld $3,%xmm4,%xmm7 | |
3333 xorl %ecx,%r12d | |
3334 shrdl $11,%r14d,%r14d | |
3335 xorl %r9d,%r15d | |
3336 vpslld $14,%xmm4,%xmm5 | |
3337 addl %r12d,%edx | |
3338 shrdl $6,%r13d,%r13d | |
3339 andl %r15d,%edi | |
3340 vpxor %xmm6,%xmm7,%xmm4 | |
3341 xorl %r8d,%r14d | |
3342 addl %r13d,%edx | |
3343 xorl %r9d,%edi | |
3344 vpshufd $250,%xmm2,%xmm7 | |
3345 shrdl $2,%r14d,%r14d | |
3346 addl %edx,%r11d | |
3347 addl %edi,%edx | |
3348 vpsrld $11,%xmm6,%xmm6 | |
3349 movl %r11d,%r13d | |
3350 addl %edx,%r14d | |
3351 shrdl $14,%r13d,%r13d | |
3352 vpxor %xmm5,%xmm4,%xmm4 | |
3353 movl %r14d,%edx | |
3354 movl %eax,%r12d | |
3355 shrdl $9,%r14d,%r14d | |
3356 vpslld $11,%xmm5,%xmm5 | |
3357 xorl %r11d,%r13d | |
3358 xorl %ebx,%r12d | |
3359 shrdl $5,%r13d,%r13d | |
3360 vpxor %xmm6,%xmm4,%xmm4 | |
3361 xorl %edx,%r14d | |
3362 andl %r11d,%r12d | |
3363 xorl %r11d,%r13d | |
3364 vpsrld $10,%xmm7,%xmm6 | |
3365 addl 52(%rsp),%ecx | |
3366 movl %edx,%edi | |
3367 xorl %ebx,%r12d | |
3368 vpxor %xmm5,%xmm4,%xmm4 | |
3369 shrdl $11,%r14d,%r14d | |
3370 xorl %r8d,%edi | |
3371 addl %r12d,%ecx | |
3372 vpsrlq $17,%xmm7,%xmm7 | |
3373 shrdl $6,%r13d,%r13d | |
3374 andl %edi,%r15d | |
3375 xorl %edx,%r14d | |
3376 vpaddd %xmm4,%xmm3,%xmm3 | |
3377 addl %r13d,%ecx | |
3378 xorl %r8d,%r15d | |
3379 shrdl $2,%r14d,%r14d | |
3380 vpxor %xmm7,%xmm6,%xmm6 | |
3381 addl %ecx,%r10d | |
3382 addl %r15d,%ecx | |
3383 movl %r10d,%r13d | |
3384 vpsrlq $2,%xmm7,%xmm7 | |
3385 addl %ecx,%r14d | |
3386 shrdl $14,%r13d,%r13d | |
3387 movl %r14d,%ecx | |
3388 vpxor %xmm7,%xmm6,%xmm6 | |
3389 movl %r11d,%r12d | |
3390 shrdl $9,%r14d,%r14d | |
3391 xorl %r10d,%r13d | |
3392 vpshufb %xmm8,%xmm6,%xmm6 | |
3393 xorl %eax,%r12d | |
3394 shrdl $5,%r13d,%r13d | |
3395 xorl %ecx,%r14d | |
3396 vpaddd %xmm6,%xmm3,%xmm3 | |
3397 andl %r10d,%r12d | |
3398 xorl %r10d,%r13d | |
3399 addl 56(%rsp),%ebx | |
3400 vpshufd $80,%xmm3,%xmm7 | |
3401 movl %ecx,%r15d | |
3402 xorl %eax,%r12d | |
3403 shrdl $11,%r14d,%r14d | |
3404 vpsrld $10,%xmm7,%xmm6 | |
3405 xorl %edx,%r15d | |
3406 addl %r12d,%ebx | |
3407 shrdl $6,%r13d,%r13d | |
3408 vpsrlq $17,%xmm7,%xmm7 | |
3409 andl %r15d,%edi | |
3410 xorl %ecx,%r14d | |
3411 addl %r13d,%ebx | |
3412 vpxor %xmm7,%xmm6,%xmm6 | |
3413 xorl %edx,%edi | |
3414 shrdl $2,%r14d,%r14d | |
3415 addl %ebx,%r9d | |
3416 vpsrlq $2,%xmm7,%xmm7 | |
3417 addl %edi,%ebx | |
3418 movl %r9d,%r13d | |
3419 addl %ebx,%r14d | |
3420 vpxor %xmm7,%xmm6,%xmm6 | |
3421 shrdl $14,%r13d,%r13d | |
3422 movl %r14d,%ebx | |
3423 movl %r10d,%r12d | |
3424 vpshufb %xmm9,%xmm6,%xmm6 | |
3425 shrdl $9,%r14d,%r14d | |
3426 xorl %r9d,%r13d | |
3427 xorl %r11d,%r12d | |
3428 vpaddd %xmm6,%xmm3,%xmm3 | |
3429 shrdl $5,%r13d,%r13d | |
3430 xorl %ebx,%r14d | |
3431 andl %r9d,%r12d | |
3432 vpaddd 96(%rbp),%xmm3,%xmm6 | |
3433 xorl %r9d,%r13d | |
3434 addl 60(%rsp),%eax | |
3435 movl %ebx,%edi | |
3436 xorl %r11d,%r12d | |
3437 shrdl $11,%r14d,%r14d | |
3438 xorl %ecx,%edi | |
3439 addl %r12d,%eax | |
3440 shrdl $6,%r13d,%r13d | |
3441 andl %edi,%r15d | |
3442 xorl %ebx,%r14d | |
3443 addl %r13d,%eax | |
3444 xorl %ecx,%r15d | |
3445 shrdl $2,%r14d,%r14d | |
3446 addl %eax,%r8d | |
3447 addl %r15d,%eax | |
3448 movl %r8d,%r13d | |
3449 addl %eax,%r14d | |
3450 vmovdqa %xmm6,48(%rsp) | |
3451 cmpb $0,131(%rbp) | |
3452 jne .Lavx_00_47 | |
3453 shrdl $14,%r13d,%r13d | |
3454 movl %r14d,%eax | |
3455 movl %r9d,%r12d | |
3456 shrdl $9,%r14d,%r14d | |
3457 xorl %r8d,%r13d | |
3458 xorl %r10d,%r12d | |
3459 shrdl $5,%r13d,%r13d | |
3460 xorl %eax,%r14d | |
3461 andl %r8d,%r12d | |
3462 xorl %r8d,%r13d | |
3463 addl 0(%rsp),%r11d | |
3464 movl %eax,%r15d | |
3465 xorl %r10d,%r12d | |
3466 shrdl $11,%r14d,%r14d | |
3467 xorl %ebx,%r15d | |
3468 addl %r12d,%r11d | |
3469 shrdl $6,%r13d,%r13d | |
3470 andl %r15d,%edi | |
3471 xorl %eax,%r14d | |
3472 addl %r13d,%r11d | |
3473 xorl %ebx,%edi | |
3474 shrdl $2,%r14d,%r14d | |
3475 addl %r11d,%edx | |
3476 addl %edi,%r11d | |
3477 movl %edx,%r13d | |
3478 addl %r11d,%r14d | |
3479 shrdl $14,%r13d,%r13d | |
3480 movl %r14d,%r11d | |
3481 movl %r8d,%r12d | |
3482 shrdl $9,%r14d,%r14d | |
3483 xorl %edx,%r13d | |
3484 xorl %r9d,%r12d | |
3485 shrdl $5,%r13d,%r13d | |
3486 xorl %r11d,%r14d | |
3487 andl %edx,%r12d | |
3488 xorl %edx,%r13d | |
3489 addl 4(%rsp),%r10d | |
3490 movl %r11d,%edi | |
3491 xorl %r9d,%r12d | |
3492 shrdl $11,%r14d,%r14d | |
3493 xorl %eax,%edi | |
3494 addl %r12d,%r10d | |
3495 shrdl $6,%r13d,%r13d | |
3496 andl %edi,%r15d | |
3497 xorl %r11d,%r14d | |
3498 addl %r13d,%r10d | |
3499 xorl %eax,%r15d | |
3500 shrdl $2,%r14d,%r14d | |
3501 addl %r10d,%ecx | |
3502 addl %r15d,%r10d | |
3503 movl %ecx,%r13d | |
3504 addl %r10d,%r14d | |
3505 shrdl $14,%r13d,%r13d | |
3506 movl %r14d,%r10d | |
3507 movl %edx,%r12d | |
3508 shrdl $9,%r14d,%r14d | |
3509 xorl %ecx,%r13d | |
3510 xorl %r8d,%r12d | |
3511 shrdl $5,%r13d,%r13d | |
3512 xorl %r10d,%r14d | |
3513 andl %ecx,%r12d | |
3514 xorl %ecx,%r13d | |
3515 addl 8(%rsp),%r9d | |
3516 movl %r10d,%r15d | |
3517 xorl %r8d,%r12d | |
3518 shrdl $11,%r14d,%r14d | |
3519 xorl %r11d,%r15d | |
3520 addl %r12d,%r9d | |
3521 shrdl $6,%r13d,%r13d | |
3522 andl %r15d,%edi | |
3523 xorl %r10d,%r14d | |
3524 addl %r13d,%r9d | |
3525 xorl %r11d,%edi | |
3526 shrdl $2,%r14d,%r14d | |
3527 addl %r9d,%ebx | |
3528 addl %edi,%r9d | |
3529 movl %ebx,%r13d | |
3530 addl %r9d,%r14d | |
3531 shrdl $14,%r13d,%r13d | |
3532 movl %r14d,%r9d | |
3533 movl %ecx,%r12d | |
3534 shrdl $9,%r14d,%r14d | |
3535 xorl %ebx,%r13d | |
3536 xorl %edx,%r12d | |
3537 shrdl $5,%r13d,%r13d | |
3538 xorl %r9d,%r14d | |
3539 andl %ebx,%r12d | |
3540 xorl %ebx,%r13d | |
3541 addl 12(%rsp),%r8d | |
3542 movl %r9d,%edi | |
3543 xorl %edx,%r12d | |
3544 shrdl $11,%r14d,%r14d | |
3545 xorl %r10d,%edi | |
3546 addl %r12d,%r8d | |
3547 shrdl $6,%r13d,%r13d | |
3548 andl %edi,%r15d | |
3549 xorl %r9d,%r14d | |
3550 addl %r13d,%r8d | |
3551 xorl %r10d,%r15d | |
3552 shrdl $2,%r14d,%r14d | |
3553 addl %r8d,%eax | |
3554 addl %r15d,%r8d | |
3555 movl %eax,%r13d | |
3556 addl %r8d,%r14d | |
3557 shrdl $14,%r13d,%r13d | |
3558 movl %r14d,%r8d | |
3559 movl %ebx,%r12d | |
3560 shrdl $9,%r14d,%r14d | |
3561 xorl %eax,%r13d | |
3562 xorl %ecx,%r12d | |
3563 shrdl $5,%r13d,%r13d | |
3564 xorl %r8d,%r14d | |
3565 andl %eax,%r12d | |
3566 xorl %eax,%r13d | |
3567 addl 16(%rsp),%edx | |
3568 movl %r8d,%r15d | |
3569 xorl %ecx,%r12d | |
3570 shrdl $11,%r14d,%r14d | |
3571 xorl %r9d,%r15d | |
3572 addl %r12d,%edx | |
3573 shrdl $6,%r13d,%r13d | |
3574 andl %r15d,%edi | |
3575 xorl %r8d,%r14d | |
3576 addl %r13d,%edx | |
3577 xorl %r9d,%edi | |
3578 shrdl $2,%r14d,%r14d | |
3579 addl %edx,%r11d | |
3580 addl %edi,%edx | |
3581 movl %r11d,%r13d | |
3582 addl %edx,%r14d | |
3583 shrdl $14,%r13d,%r13d | |
3584 movl %r14d,%edx | |
3585 movl %eax,%r12d | |
3586 shrdl $9,%r14d,%r14d | |
3587 xorl %r11d,%r13d | |
3588 xorl %ebx,%r12d | |
3589 shrdl $5,%r13d,%r13d | |
3590 xorl %edx,%r14d | |
3591 andl %r11d,%r12d | |
3592 xorl %r11d,%r13d | |
3593 addl 20(%rsp),%ecx | |
3594 movl %edx,%edi | |
3595 xorl %ebx,%r12d | |
3596 shrdl $11,%r14d,%r14d | |
3597 xorl %r8d,%edi | |
3598 addl %r12d,%ecx | |
3599 shrdl $6,%r13d,%r13d | |
3600 andl %edi,%r15d | |
3601 xorl %edx,%r14d | |
3602 addl %r13d,%ecx | |
3603 xorl %r8d,%r15d | |
3604 shrdl $2,%r14d,%r14d | |
3605 addl %ecx,%r10d | |
3606 addl %r15d,%ecx | |
3607 movl %r10d,%r13d | |
3608 addl %ecx,%r14d | |
3609 shrdl $14,%r13d,%r13d | |
3610 movl %r14d,%ecx | |
3611 movl %r11d,%r12d | |
3612 shrdl $9,%r14d,%r14d | |
3613 xorl %r10d,%r13d | |
3614 xorl %eax,%r12d | |
3615 shrdl $5,%r13d,%r13d | |
3616 xorl %ecx,%r14d | |
3617 andl %r10d,%r12d | |
3618 xorl %r10d,%r13d | |
3619 addl 24(%rsp),%ebx | |
3620 movl %ecx,%r15d | |
3621 xorl %eax,%r12d | |
3622 shrdl $11,%r14d,%r14d | |
3623 xorl %edx,%r15d | |
3624 addl %r12d,%ebx | |
3625 shrdl $6,%r13d,%r13d | |
3626 andl %r15d,%edi | |
3627 xorl %ecx,%r14d | |
3628 addl %r13d,%ebx | |
3629 xorl %edx,%edi | |
3630 shrdl $2,%r14d,%r14d | |
3631 addl %ebx,%r9d | |
3632 addl %edi,%ebx | |
3633 movl %r9d,%r13d | |
3634 addl %ebx,%r14d | |
3635 shrdl $14,%r13d,%r13d | |
3636 movl %r14d,%ebx | |
3637 movl %r10d,%r12d | |
3638 shrdl $9,%r14d,%r14d | |
3639 xorl %r9d,%r13d | |
3640 xorl %r11d,%r12d | |
3641 shrdl $5,%r13d,%r13d | |
3642 xorl %ebx,%r14d | |
3643 andl %r9d,%r12d | |
3644 xorl %r9d,%r13d | |
3645 addl 28(%rsp),%eax | |
3646 movl %ebx,%edi | |
3647 xorl %r11d,%r12d | |
3648 shrdl $11,%r14d,%r14d | |
3649 xorl %ecx,%edi | |
3650 addl %r12d,%eax | |
3651 shrdl $6,%r13d,%r13d | |
3652 andl %edi,%r15d | |
3653 xorl %ebx,%r14d | |
3654 addl %r13d,%eax | |
3655 xorl %ecx,%r15d | |
3656 shrdl $2,%r14d,%r14d | |
3657 addl %eax,%r8d | |
3658 addl %r15d,%eax | |
3659 movl %r8d,%r13d | |
3660 addl %eax,%r14d | |
3661 shrdl $14,%r13d,%r13d | |
3662 movl %r14d,%eax | |
3663 movl %r9d,%r12d | |
3664 shrdl $9,%r14d,%r14d | |
3665 xorl %r8d,%r13d | |
3666 xorl %r10d,%r12d | |
3667 shrdl $5,%r13d,%r13d | |
3668 xorl %eax,%r14d | |
3669 andl %r8d,%r12d | |
3670 xorl %r8d,%r13d | |
3671 addl 32(%rsp),%r11d | |
3672 movl %eax,%r15d | |
3673 xorl %r10d,%r12d | |
3674 shrdl $11,%r14d,%r14d | |
3675 xorl %ebx,%r15d | |
3676 addl %r12d,%r11d | |
3677 shrdl $6,%r13d,%r13d | |
3678 andl %r15d,%edi | |
3679 xorl %eax,%r14d | |
3680 addl %r13d,%r11d | |
3681 xorl %ebx,%edi | |
3682 shrdl $2,%r14d,%r14d | |
3683 addl %r11d,%edx | |
3684 addl %edi,%r11d | |
3685 movl %edx,%r13d | |
3686 addl %r11d,%r14d | |
3687 shrdl $14,%r13d,%r13d | |
3688 movl %r14d,%r11d | |
3689 movl %r8d,%r12d | |
3690 shrdl $9,%r14d,%r14d | |
3691 xorl %edx,%r13d | |
3692 xorl %r9d,%r12d | |
3693 shrdl $5,%r13d,%r13d | |
3694 xorl %r11d,%r14d | |
3695 andl %edx,%r12d | |
3696 xorl %edx,%r13d | |
3697 addl 36(%rsp),%r10d | |
3698 movl %r11d,%edi | |
3699 xorl %r9d,%r12d | |
3700 shrdl $11,%r14d,%r14d | |
3701 xorl %eax,%edi | |
3702 addl %r12d,%r10d | |
3703 shrdl $6,%r13d,%r13d | |
3704 andl %edi,%r15d | |
3705 xorl %r11d,%r14d | |
3706 addl %r13d,%r10d | |
3707 xorl %eax,%r15d | |
3708 shrdl $2,%r14d,%r14d | |
3709 addl %r10d,%ecx | |
3710 addl %r15d,%r10d | |
3711 movl %ecx,%r13d | |
3712 addl %r10d,%r14d | |
3713 shrdl $14,%r13d,%r13d | |
3714 movl %r14d,%r10d | |
3715 movl %edx,%r12d | |
3716 shrdl $9,%r14d,%r14d | |
3717 xorl %ecx,%r13d | |
3718 xorl %r8d,%r12d | |
3719 shrdl $5,%r13d,%r13d | |
3720 xorl %r10d,%r14d | |
3721 andl %ecx,%r12d | |
3722 xorl %ecx,%r13d | |
3723 addl 40(%rsp),%r9d | |
3724 movl %r10d,%r15d | |
3725 xorl %r8d,%r12d | |
3726 shrdl $11,%r14d,%r14d | |
3727 xorl %r11d,%r15d | |
3728 addl %r12d,%r9d | |
3729 shrdl $6,%r13d,%r13d | |
3730 andl %r15d,%edi | |
3731 xorl %r10d,%r14d | |
3732 addl %r13d,%r9d | |
3733 xorl %r11d,%edi | |
3734 shrdl $2,%r14d,%r14d | |
3735 addl %r9d,%ebx | |
3736 addl %edi,%r9d | |
3737 movl %ebx,%r13d | |
3738 addl %r9d,%r14d | |
3739 shrdl $14,%r13d,%r13d | |
3740 movl %r14d,%r9d | |
3741 movl %ecx,%r12d | |
3742 shrdl $9,%r14d,%r14d | |
3743 xorl %ebx,%r13d | |
3744 xorl %edx,%r12d | |
3745 shrdl $5,%r13d,%r13d | |
3746 xorl %r9d,%r14d | |
3747 andl %ebx,%r12d | |
3748 xorl %ebx,%r13d | |
3749 addl 44(%rsp),%r8d | |
3750 movl %r9d,%edi | |
3751 xorl %edx,%r12d | |
3752 shrdl $11,%r14d,%r14d | |
3753 xorl %r10d,%edi | |
3754 addl %r12d,%r8d | |
3755 shrdl $6,%r13d,%r13d | |
3756 andl %edi,%r15d | |
3757 xorl %r9d,%r14d | |
3758 addl %r13d,%r8d | |
3759 xorl %r10d,%r15d | |
3760 shrdl $2,%r14d,%r14d | |
3761 addl %r8d,%eax | |
3762 addl %r15d,%r8d | |
3763 movl %eax,%r13d | |
3764 addl %r8d,%r14d | |
3765 shrdl $14,%r13d,%r13d | |
3766 movl %r14d,%r8d | |
3767 movl %ebx,%r12d | |
3768 shrdl $9,%r14d,%r14d | |
3769 xorl %eax,%r13d | |
3770 xorl %ecx,%r12d | |
3771 shrdl $5,%r13d,%r13d | |
3772 xorl %r8d,%r14d | |
3773 andl %eax,%r12d | |
3774 xorl %eax,%r13d | |
3775 addl 48(%rsp),%edx | |
3776 movl %r8d,%r15d | |
3777 xorl %ecx,%r12d | |
3778 shrdl $11,%r14d,%r14d | |
3779 xorl %r9d,%r15d | |
3780 addl %r12d,%edx | |
3781 shrdl $6,%r13d,%r13d | |
3782 andl %r15d,%edi | |
3783 xorl %r8d,%r14d | |
3784 addl %r13d,%edx | |
3785 xorl %r9d,%edi | |
3786 shrdl $2,%r14d,%r14d | |
3787 addl %edx,%r11d | |
3788 addl %edi,%edx | |
3789 movl %r11d,%r13d | |
3790 addl %edx,%r14d | |
3791 shrdl $14,%r13d,%r13d | |
3792 movl %r14d,%edx | |
3793 movl %eax,%r12d | |
3794 shrdl $9,%r14d,%r14d | |
3795 xorl %r11d,%r13d | |
3796 xorl %ebx,%r12d | |
3797 shrdl $5,%r13d,%r13d | |
3798 xorl %edx,%r14d | |
3799 andl %r11d,%r12d | |
3800 xorl %r11d,%r13d | |
3801 addl 52(%rsp),%ecx | |
3802 movl %edx,%edi | |
3803 xorl %ebx,%r12d | |
3804 shrdl $11,%r14d,%r14d | |
3805 xorl %r8d,%edi | |
3806 addl %r12d,%ecx | |
3807 shrdl $6,%r13d,%r13d | |
3808 andl %edi,%r15d | |
3809 xorl %edx,%r14d | |
3810 addl %r13d,%ecx | |
3811 xorl %r8d,%r15d | |
3812 shrdl $2,%r14d,%r14d | |
3813 addl %ecx,%r10d | |
3814 addl %r15d,%ecx | |
3815 movl %r10d,%r13d | |
3816 addl %ecx,%r14d | |
3817 shrdl $14,%r13d,%r13d | |
3818 movl %r14d,%ecx | |
3819 movl %r11d,%r12d | |
3820 shrdl $9,%r14d,%r14d | |
3821 xorl %r10d,%r13d | |
3822 xorl %eax,%r12d | |
3823 shrdl $5,%r13d,%r13d | |
3824 xorl %ecx,%r14d | |
3825 andl %r10d,%r12d | |
3826 xorl %r10d,%r13d | |
3827 addl 56(%rsp),%ebx | |
3828 movl %ecx,%r15d | |
3829 xorl %eax,%r12d | |
3830 shrdl $11,%r14d,%r14d | |
3831 xorl %edx,%r15d | |
3832 addl %r12d,%ebx | |
3833 shrdl $6,%r13d,%r13d | |
3834 andl %r15d,%edi | |
3835 xorl %ecx,%r14d | |
3836 addl %r13d,%ebx | |
3837 xorl %edx,%edi | |
3838 shrdl $2,%r14d,%r14d | |
3839 addl %ebx,%r9d | |
3840 addl %edi,%ebx | |
3841 movl %r9d,%r13d | |
3842 addl %ebx,%r14d | |
3843 shrdl $14,%r13d,%r13d | |
3844 movl %r14d,%ebx | |
3845 movl %r10d,%r12d | |
3846 shrdl $9,%r14d,%r14d | |
3847 xorl %r9d,%r13d | |
3848 xorl %r11d,%r12d | |
3849 shrdl $5,%r13d,%r13d | |
3850 xorl %ebx,%r14d | |
3851 andl %r9d,%r12d | |
3852 xorl %r9d,%r13d | |
3853 addl 60(%rsp),%eax | |
3854 movl %ebx,%edi | |
3855 xorl %r11d,%r12d | |
3856 shrdl $11,%r14d,%r14d | |
3857 xorl %ecx,%edi | |
3858 addl %r12d,%eax | |
3859 shrdl $6,%r13d,%r13d | |
3860 andl %edi,%r15d | |
3861 xorl %ebx,%r14d | |
3862 addl %r13d,%eax | |
3863 xorl %ecx,%r15d | |
3864 shrdl $2,%r14d,%r14d | |
3865 addl %eax,%r8d | |
3866 addl %r15d,%eax | |
3867 movl %r8d,%r13d | |
3868 addl %eax,%r14d | |
3869 movq 64+0(%rsp),%rdi | |
3870 movl %r14d,%eax | |
3871 | |
3872 addl 0(%rdi),%eax | |
3873 leaq 64(%rsi),%rsi | |
3874 addl 4(%rdi),%ebx | |
3875 addl 8(%rdi),%ecx | |
3876 addl 12(%rdi),%edx | |
3877 addl 16(%rdi),%r8d | |
3878 addl 20(%rdi),%r9d | |
3879 addl 24(%rdi),%r10d | |
3880 addl 28(%rdi),%r11d | |
3881 | |
3882 cmpq 64+16(%rsp),%rsi | |
3883 | |
3884 movl %eax,0(%rdi) | |
3885 movl %ebx,4(%rdi) | |
3886 movl %ecx,8(%rdi) | |
3887 movl %edx,12(%rdi) | |
3888 movl %r8d,16(%rdi) | |
3889 movl %r9d,20(%rdi) | |
3890 movl %r10d,24(%rdi) | |
3891 movl %r11d,28(%rdi) | |
3892 jb .Lloop_avx | |
3893 | |
3894 movq 64+24(%rsp),%rsi | |
3895 vzeroupper | |
3896 movq (%rsi),%r15 | |
3897 movq 8(%rsi),%r14 | |
3898 movq 16(%rsi),%r13 | |
3899 movq 24(%rsi),%r12 | |
3900 movq 32(%rsi),%rbp | |
3901 movq 40(%rsi),%rbx | |
3902 leaq 48(%rsi),%rsp | |
3903 .Lepilogue_avx: | |
3904 .byte 0xf3,0xc3 | |
3905 .size sha256_block_data_order_avx,.-sha256_block_data_order_avx | |
3906 #endif | 2844 #endif |
OLD | NEW |