OLD | NEW |
1 #if defined(__x86_64__) | 1 #if defined(__x86_64__) |
2 .text | 2 .text |
3 | 3 |
4 | 4 |
5 .globl _sha256_block_data_order | 5 .globl _sha256_block_data_order |
6 .private_extern _sha256_block_data_order | 6 .private_extern _sha256_block_data_order |
7 | 7 |
8 .p2align 4 | 8 .p2align 4 |
9 _sha256_block_data_order: | 9 _sha256_block_data_order: |
10 leaq _OPENSSL_ia32cap_P(%rip),%r11 | 10 leaq _OPENSSL_ia32cap_P(%rip),%r11 |
11 movl 0(%r11),%r9d | 11 movl 0(%r11),%r9d |
12 movl 4(%r11),%r10d | 12 movl 4(%r11),%r10d |
13 movl 8(%r11),%r11d | 13 movl 8(%r11),%r11d |
14 andl $1073741824,%r9d | |
15 andl $268435968,%r10d | |
16 orl %r9d,%r10d | |
17 cmpl $1342177792,%r10d | |
18 je L$avx_shortcut | |
19 testl $512,%r10d | 14 testl $512,%r10d |
20 jnz L$ssse3_shortcut | 15 jnz L$ssse3_shortcut |
21 pushq %rbx | 16 pushq %rbx |
22 pushq %rbp | 17 pushq %rbp |
23 pushq %r12 | 18 pushq %r12 |
24 pushq %r13 | 19 pushq %r13 |
25 pushq %r14 | 20 pushq %r14 |
26 pushq %r15 | 21 pushq %r15 |
27 movq %rsp,%r11 | 22 movq %rsp,%r11 |
28 shlq $4,%rdx | 23 shlq $4,%rdx |
(...skipping 2809 matching lines...) Expand 10 before | Expand all | Expand 10 after Loading... |
2838 movq (%rsi),%r15 | 2833 movq (%rsi),%r15 |
2839 movq 8(%rsi),%r14 | 2834 movq 8(%rsi),%r14 |
2840 movq 16(%rsi),%r13 | 2835 movq 16(%rsi),%r13 |
2841 movq 24(%rsi),%r12 | 2836 movq 24(%rsi),%r12 |
2842 movq 32(%rsi),%rbp | 2837 movq 32(%rsi),%rbp |
2843 movq 40(%rsi),%rbx | 2838 movq 40(%rsi),%rbx |
2844 leaq 48(%rsi),%rsp | 2839 leaq 48(%rsi),%rsp |
2845 L$epilogue_ssse3: | 2840 L$epilogue_ssse3: |
2846 .byte 0xf3,0xc3 | 2841 .byte 0xf3,0xc3 |
2847 | 2842 |
2848 | |
2849 .p2align 6 | |
2850 sha256_block_data_order_avx: | |
2851 L$avx_shortcut: | |
2852 pushq %rbx | |
2853 pushq %rbp | |
2854 pushq %r12 | |
2855 pushq %r13 | |
2856 pushq %r14 | |
2857 pushq %r15 | |
2858 movq %rsp,%r11 | |
2859 shlq $4,%rdx | |
2860 subq $96,%rsp | |
2861 leaq (%rsi,%rdx,4),%rdx | |
2862 andq $-64,%rsp | |
2863 movq %rdi,64+0(%rsp) | |
2864 movq %rsi,64+8(%rsp) | |
2865 movq %rdx,64+16(%rsp) | |
2866 movq %r11,64+24(%rsp) | |
2867 L$prologue_avx: | |
2868 | |
2869 vzeroupper | |
2870 movl 0(%rdi),%eax | |
2871 movl 4(%rdi),%ebx | |
2872 movl 8(%rdi),%ecx | |
2873 movl 12(%rdi),%edx | |
2874 movl 16(%rdi),%r8d | |
2875 movl 20(%rdi),%r9d | |
2876 movl 24(%rdi),%r10d | |
2877 movl 28(%rdi),%r11d | |
2878 vmovdqa K256+512+32(%rip),%xmm8 | |
2879 vmovdqa K256+512+64(%rip),%xmm9 | |
2880 jmp L$loop_avx | |
2881 .p2align 4 | |
2882 L$loop_avx: | |
2883 vmovdqa K256+512(%rip),%xmm7 | |
2884 vmovdqu 0(%rsi),%xmm0 | |
2885 vmovdqu 16(%rsi),%xmm1 | |
2886 vmovdqu 32(%rsi),%xmm2 | |
2887 vmovdqu 48(%rsi),%xmm3 | |
2888 vpshufb %xmm7,%xmm0,%xmm0 | |
2889 leaq K256(%rip),%rbp | |
2890 vpshufb %xmm7,%xmm1,%xmm1 | |
2891 vpshufb %xmm7,%xmm2,%xmm2 | |
2892 vpaddd 0(%rbp),%xmm0,%xmm4 | |
2893 vpshufb %xmm7,%xmm3,%xmm3 | |
2894 vpaddd 32(%rbp),%xmm1,%xmm5 | |
2895 vpaddd 64(%rbp),%xmm2,%xmm6 | |
2896 vpaddd 96(%rbp),%xmm3,%xmm7 | |
2897 vmovdqa %xmm4,0(%rsp) | |
2898 movl %eax,%r14d | |
2899 vmovdqa %xmm5,16(%rsp) | |
2900 movl %ebx,%edi | |
2901 vmovdqa %xmm6,32(%rsp) | |
2902 xorl %ecx,%edi | |
2903 vmovdqa %xmm7,48(%rsp) | |
2904 movl %r8d,%r13d | |
2905 jmp L$avx_00_47 | |
2906 | |
2907 .p2align 4 | |
2908 L$avx_00_47: | |
2909 subq $-128,%rbp | |
2910 vpalignr $4,%xmm0,%xmm1,%xmm4 | |
2911 shrdl $14,%r13d,%r13d | |
2912 movl %r14d,%eax | |
2913 movl %r9d,%r12d | |
2914 vpalignr $4,%xmm2,%xmm3,%xmm7 | |
2915 shrdl $9,%r14d,%r14d | |
2916 xorl %r8d,%r13d | |
2917 xorl %r10d,%r12d | |
2918 vpsrld $7,%xmm4,%xmm6 | |
2919 shrdl $5,%r13d,%r13d | |
2920 xorl %eax,%r14d | |
2921 andl %r8d,%r12d | |
2922 vpaddd %xmm7,%xmm0,%xmm0 | |
2923 xorl %r8d,%r13d | |
2924 addl 0(%rsp),%r11d | |
2925 movl %eax,%r15d | |
2926 vpsrld $3,%xmm4,%xmm7 | |
2927 xorl %r10d,%r12d | |
2928 shrdl $11,%r14d,%r14d | |
2929 xorl %ebx,%r15d | |
2930 vpslld $14,%xmm4,%xmm5 | |
2931 addl %r12d,%r11d | |
2932 shrdl $6,%r13d,%r13d | |
2933 andl %r15d,%edi | |
2934 vpxor %xmm6,%xmm7,%xmm4 | |
2935 xorl %eax,%r14d | |
2936 addl %r13d,%r11d | |
2937 xorl %ebx,%edi | |
2938 vpshufd $250,%xmm3,%xmm7 | |
2939 shrdl $2,%r14d,%r14d | |
2940 addl %r11d,%edx | |
2941 addl %edi,%r11d | |
2942 vpsrld $11,%xmm6,%xmm6 | |
2943 movl %edx,%r13d | |
2944 addl %r11d,%r14d | |
2945 shrdl $14,%r13d,%r13d | |
2946 vpxor %xmm5,%xmm4,%xmm4 | |
2947 movl %r14d,%r11d | |
2948 movl %r8d,%r12d | |
2949 shrdl $9,%r14d,%r14d | |
2950 vpslld $11,%xmm5,%xmm5 | |
2951 xorl %edx,%r13d | |
2952 xorl %r9d,%r12d | |
2953 shrdl $5,%r13d,%r13d | |
2954 vpxor %xmm6,%xmm4,%xmm4 | |
2955 xorl %r11d,%r14d | |
2956 andl %edx,%r12d | |
2957 xorl %edx,%r13d | |
2958 vpsrld $10,%xmm7,%xmm6 | |
2959 addl 4(%rsp),%r10d | |
2960 movl %r11d,%edi | |
2961 xorl %r9d,%r12d | |
2962 vpxor %xmm5,%xmm4,%xmm4 | |
2963 shrdl $11,%r14d,%r14d | |
2964 xorl %eax,%edi | |
2965 addl %r12d,%r10d | |
2966 vpsrlq $17,%xmm7,%xmm7 | |
2967 shrdl $6,%r13d,%r13d | |
2968 andl %edi,%r15d | |
2969 xorl %r11d,%r14d | |
2970 vpaddd %xmm4,%xmm0,%xmm0 | |
2971 addl %r13d,%r10d | |
2972 xorl %eax,%r15d | |
2973 shrdl $2,%r14d,%r14d | |
2974 vpxor %xmm7,%xmm6,%xmm6 | |
2975 addl %r10d,%ecx | |
2976 addl %r15d,%r10d | |
2977 movl %ecx,%r13d | |
2978 vpsrlq $2,%xmm7,%xmm7 | |
2979 addl %r10d,%r14d | |
2980 shrdl $14,%r13d,%r13d | |
2981 movl %r14d,%r10d | |
2982 vpxor %xmm7,%xmm6,%xmm6 | |
2983 movl %edx,%r12d | |
2984 shrdl $9,%r14d,%r14d | |
2985 xorl %ecx,%r13d | |
2986 vpshufb %xmm8,%xmm6,%xmm6 | |
2987 xorl %r8d,%r12d | |
2988 shrdl $5,%r13d,%r13d | |
2989 xorl %r10d,%r14d | |
2990 vpaddd %xmm6,%xmm0,%xmm0 | |
2991 andl %ecx,%r12d | |
2992 xorl %ecx,%r13d | |
2993 addl 8(%rsp),%r9d | |
2994 vpshufd $80,%xmm0,%xmm7 | |
2995 movl %r10d,%r15d | |
2996 xorl %r8d,%r12d | |
2997 shrdl $11,%r14d,%r14d | |
2998 vpsrld $10,%xmm7,%xmm6 | |
2999 xorl %r11d,%r15d | |
3000 addl %r12d,%r9d | |
3001 shrdl $6,%r13d,%r13d | |
3002 vpsrlq $17,%xmm7,%xmm7 | |
3003 andl %r15d,%edi | |
3004 xorl %r10d,%r14d | |
3005 addl %r13d,%r9d | |
3006 vpxor %xmm7,%xmm6,%xmm6 | |
3007 xorl %r11d,%edi | |
3008 shrdl $2,%r14d,%r14d | |
3009 addl %r9d,%ebx | |
3010 vpsrlq $2,%xmm7,%xmm7 | |
3011 addl %edi,%r9d | |
3012 movl %ebx,%r13d | |
3013 addl %r9d,%r14d | |
3014 vpxor %xmm7,%xmm6,%xmm6 | |
3015 shrdl $14,%r13d,%r13d | |
3016 movl %r14d,%r9d | |
3017 movl %ecx,%r12d | |
3018 vpshufb %xmm9,%xmm6,%xmm6 | |
3019 shrdl $9,%r14d,%r14d | |
3020 xorl %ebx,%r13d | |
3021 xorl %edx,%r12d | |
3022 vpaddd %xmm6,%xmm0,%xmm0 | |
3023 shrdl $5,%r13d,%r13d | |
3024 xorl %r9d,%r14d | |
3025 andl %ebx,%r12d | |
3026 vpaddd 0(%rbp),%xmm0,%xmm6 | |
3027 xorl %ebx,%r13d | |
3028 addl 12(%rsp),%r8d | |
3029 movl %r9d,%edi | |
3030 xorl %edx,%r12d | |
3031 shrdl $11,%r14d,%r14d | |
3032 xorl %r10d,%edi | |
3033 addl %r12d,%r8d | |
3034 shrdl $6,%r13d,%r13d | |
3035 andl %edi,%r15d | |
3036 xorl %r9d,%r14d | |
3037 addl %r13d,%r8d | |
3038 xorl %r10d,%r15d | |
3039 shrdl $2,%r14d,%r14d | |
3040 addl %r8d,%eax | |
3041 addl %r15d,%r8d | |
3042 movl %eax,%r13d | |
3043 addl %r8d,%r14d | |
3044 vmovdqa %xmm6,0(%rsp) | |
3045 vpalignr $4,%xmm1,%xmm2,%xmm4 | |
3046 shrdl $14,%r13d,%r13d | |
3047 movl %r14d,%r8d | |
3048 movl %ebx,%r12d | |
3049 vpalignr $4,%xmm3,%xmm0,%xmm7 | |
3050 shrdl $9,%r14d,%r14d | |
3051 xorl %eax,%r13d | |
3052 xorl %ecx,%r12d | |
3053 vpsrld $7,%xmm4,%xmm6 | |
3054 shrdl $5,%r13d,%r13d | |
3055 xorl %r8d,%r14d | |
3056 andl %eax,%r12d | |
3057 vpaddd %xmm7,%xmm1,%xmm1 | |
3058 xorl %eax,%r13d | |
3059 addl 16(%rsp),%edx | |
3060 movl %r8d,%r15d | |
3061 vpsrld $3,%xmm4,%xmm7 | |
3062 xorl %ecx,%r12d | |
3063 shrdl $11,%r14d,%r14d | |
3064 xorl %r9d,%r15d | |
3065 vpslld $14,%xmm4,%xmm5 | |
3066 addl %r12d,%edx | |
3067 shrdl $6,%r13d,%r13d | |
3068 andl %r15d,%edi | |
3069 vpxor %xmm6,%xmm7,%xmm4 | |
3070 xorl %r8d,%r14d | |
3071 addl %r13d,%edx | |
3072 xorl %r9d,%edi | |
3073 vpshufd $250,%xmm0,%xmm7 | |
3074 shrdl $2,%r14d,%r14d | |
3075 addl %edx,%r11d | |
3076 addl %edi,%edx | |
3077 vpsrld $11,%xmm6,%xmm6 | |
3078 movl %r11d,%r13d | |
3079 addl %edx,%r14d | |
3080 shrdl $14,%r13d,%r13d | |
3081 vpxor %xmm5,%xmm4,%xmm4 | |
3082 movl %r14d,%edx | |
3083 movl %eax,%r12d | |
3084 shrdl $9,%r14d,%r14d | |
3085 vpslld $11,%xmm5,%xmm5 | |
3086 xorl %r11d,%r13d | |
3087 xorl %ebx,%r12d | |
3088 shrdl $5,%r13d,%r13d | |
3089 vpxor %xmm6,%xmm4,%xmm4 | |
3090 xorl %edx,%r14d | |
3091 andl %r11d,%r12d | |
3092 xorl %r11d,%r13d | |
3093 vpsrld $10,%xmm7,%xmm6 | |
3094 addl 20(%rsp),%ecx | |
3095 movl %edx,%edi | |
3096 xorl %ebx,%r12d | |
3097 vpxor %xmm5,%xmm4,%xmm4 | |
3098 shrdl $11,%r14d,%r14d | |
3099 xorl %r8d,%edi | |
3100 addl %r12d,%ecx | |
3101 vpsrlq $17,%xmm7,%xmm7 | |
3102 shrdl $6,%r13d,%r13d | |
3103 andl %edi,%r15d | |
3104 xorl %edx,%r14d | |
3105 vpaddd %xmm4,%xmm1,%xmm1 | |
3106 addl %r13d,%ecx | |
3107 xorl %r8d,%r15d | |
3108 shrdl $2,%r14d,%r14d | |
3109 vpxor %xmm7,%xmm6,%xmm6 | |
3110 addl %ecx,%r10d | |
3111 addl %r15d,%ecx | |
3112 movl %r10d,%r13d | |
3113 vpsrlq $2,%xmm7,%xmm7 | |
3114 addl %ecx,%r14d | |
3115 shrdl $14,%r13d,%r13d | |
3116 movl %r14d,%ecx | |
3117 vpxor %xmm7,%xmm6,%xmm6 | |
3118 movl %r11d,%r12d | |
3119 shrdl $9,%r14d,%r14d | |
3120 xorl %r10d,%r13d | |
3121 vpshufb %xmm8,%xmm6,%xmm6 | |
3122 xorl %eax,%r12d | |
3123 shrdl $5,%r13d,%r13d | |
3124 xorl %ecx,%r14d | |
3125 vpaddd %xmm6,%xmm1,%xmm1 | |
3126 andl %r10d,%r12d | |
3127 xorl %r10d,%r13d | |
3128 addl 24(%rsp),%ebx | |
3129 vpshufd $80,%xmm1,%xmm7 | |
3130 movl %ecx,%r15d | |
3131 xorl %eax,%r12d | |
3132 shrdl $11,%r14d,%r14d | |
3133 vpsrld $10,%xmm7,%xmm6 | |
3134 xorl %edx,%r15d | |
3135 addl %r12d,%ebx | |
3136 shrdl $6,%r13d,%r13d | |
3137 vpsrlq $17,%xmm7,%xmm7 | |
3138 andl %r15d,%edi | |
3139 xorl %ecx,%r14d | |
3140 addl %r13d,%ebx | |
3141 vpxor %xmm7,%xmm6,%xmm6 | |
3142 xorl %edx,%edi | |
3143 shrdl $2,%r14d,%r14d | |
3144 addl %ebx,%r9d | |
3145 vpsrlq $2,%xmm7,%xmm7 | |
3146 addl %edi,%ebx | |
3147 movl %r9d,%r13d | |
3148 addl %ebx,%r14d | |
3149 vpxor %xmm7,%xmm6,%xmm6 | |
3150 shrdl $14,%r13d,%r13d | |
3151 movl %r14d,%ebx | |
3152 movl %r10d,%r12d | |
3153 vpshufb %xmm9,%xmm6,%xmm6 | |
3154 shrdl $9,%r14d,%r14d | |
3155 xorl %r9d,%r13d | |
3156 xorl %r11d,%r12d | |
3157 vpaddd %xmm6,%xmm1,%xmm1 | |
3158 shrdl $5,%r13d,%r13d | |
3159 xorl %ebx,%r14d | |
3160 andl %r9d,%r12d | |
3161 vpaddd 32(%rbp),%xmm1,%xmm6 | |
3162 xorl %r9d,%r13d | |
3163 addl 28(%rsp),%eax | |
3164 movl %ebx,%edi | |
3165 xorl %r11d,%r12d | |
3166 shrdl $11,%r14d,%r14d | |
3167 xorl %ecx,%edi | |
3168 addl %r12d,%eax | |
3169 shrdl $6,%r13d,%r13d | |
3170 andl %edi,%r15d | |
3171 xorl %ebx,%r14d | |
3172 addl %r13d,%eax | |
3173 xorl %ecx,%r15d | |
3174 shrdl $2,%r14d,%r14d | |
3175 addl %eax,%r8d | |
3176 addl %r15d,%eax | |
3177 movl %r8d,%r13d | |
3178 addl %eax,%r14d | |
3179 vmovdqa %xmm6,16(%rsp) | |
3180 vpalignr $4,%xmm2,%xmm3,%xmm4 | |
3181 shrdl $14,%r13d,%r13d | |
3182 movl %r14d,%eax | |
3183 movl %r9d,%r12d | |
3184 vpalignr $4,%xmm0,%xmm1,%xmm7 | |
3185 shrdl $9,%r14d,%r14d | |
3186 xorl %r8d,%r13d | |
3187 xorl %r10d,%r12d | |
3188 vpsrld $7,%xmm4,%xmm6 | |
3189 shrdl $5,%r13d,%r13d | |
3190 xorl %eax,%r14d | |
3191 andl %r8d,%r12d | |
3192 vpaddd %xmm7,%xmm2,%xmm2 | |
3193 xorl %r8d,%r13d | |
3194 addl 32(%rsp),%r11d | |
3195 movl %eax,%r15d | |
3196 vpsrld $3,%xmm4,%xmm7 | |
3197 xorl %r10d,%r12d | |
3198 shrdl $11,%r14d,%r14d | |
3199 xorl %ebx,%r15d | |
3200 vpslld $14,%xmm4,%xmm5 | |
3201 addl %r12d,%r11d | |
3202 shrdl $6,%r13d,%r13d | |
3203 andl %r15d,%edi | |
3204 vpxor %xmm6,%xmm7,%xmm4 | |
3205 xorl %eax,%r14d | |
3206 addl %r13d,%r11d | |
3207 xorl %ebx,%edi | |
3208 vpshufd $250,%xmm1,%xmm7 | |
3209 shrdl $2,%r14d,%r14d | |
3210 addl %r11d,%edx | |
3211 addl %edi,%r11d | |
3212 vpsrld $11,%xmm6,%xmm6 | |
3213 movl %edx,%r13d | |
3214 addl %r11d,%r14d | |
3215 shrdl $14,%r13d,%r13d | |
3216 vpxor %xmm5,%xmm4,%xmm4 | |
3217 movl %r14d,%r11d | |
3218 movl %r8d,%r12d | |
3219 shrdl $9,%r14d,%r14d | |
3220 vpslld $11,%xmm5,%xmm5 | |
3221 xorl %edx,%r13d | |
3222 xorl %r9d,%r12d | |
3223 shrdl $5,%r13d,%r13d | |
3224 vpxor %xmm6,%xmm4,%xmm4 | |
3225 xorl %r11d,%r14d | |
3226 andl %edx,%r12d | |
3227 xorl %edx,%r13d | |
3228 vpsrld $10,%xmm7,%xmm6 | |
3229 addl 36(%rsp),%r10d | |
3230 movl %r11d,%edi | |
3231 xorl %r9d,%r12d | |
3232 vpxor %xmm5,%xmm4,%xmm4 | |
3233 shrdl $11,%r14d,%r14d | |
3234 xorl %eax,%edi | |
3235 addl %r12d,%r10d | |
3236 vpsrlq $17,%xmm7,%xmm7 | |
3237 shrdl $6,%r13d,%r13d | |
3238 andl %edi,%r15d | |
3239 xorl %r11d,%r14d | |
3240 vpaddd %xmm4,%xmm2,%xmm2 | |
3241 addl %r13d,%r10d | |
3242 xorl %eax,%r15d | |
3243 shrdl $2,%r14d,%r14d | |
3244 vpxor %xmm7,%xmm6,%xmm6 | |
3245 addl %r10d,%ecx | |
3246 addl %r15d,%r10d | |
3247 movl %ecx,%r13d | |
3248 vpsrlq $2,%xmm7,%xmm7 | |
3249 addl %r10d,%r14d | |
3250 shrdl $14,%r13d,%r13d | |
3251 movl %r14d,%r10d | |
3252 vpxor %xmm7,%xmm6,%xmm6 | |
3253 movl %edx,%r12d | |
3254 shrdl $9,%r14d,%r14d | |
3255 xorl %ecx,%r13d | |
3256 vpshufb %xmm8,%xmm6,%xmm6 | |
3257 xorl %r8d,%r12d | |
3258 shrdl $5,%r13d,%r13d | |
3259 xorl %r10d,%r14d | |
3260 vpaddd %xmm6,%xmm2,%xmm2 | |
3261 andl %ecx,%r12d | |
3262 xorl %ecx,%r13d | |
3263 addl 40(%rsp),%r9d | |
3264 vpshufd $80,%xmm2,%xmm7 | |
3265 movl %r10d,%r15d | |
3266 xorl %r8d,%r12d | |
3267 shrdl $11,%r14d,%r14d | |
3268 vpsrld $10,%xmm7,%xmm6 | |
3269 xorl %r11d,%r15d | |
3270 addl %r12d,%r9d | |
3271 shrdl $6,%r13d,%r13d | |
3272 vpsrlq $17,%xmm7,%xmm7 | |
3273 andl %r15d,%edi | |
3274 xorl %r10d,%r14d | |
3275 addl %r13d,%r9d | |
3276 vpxor %xmm7,%xmm6,%xmm6 | |
3277 xorl %r11d,%edi | |
3278 shrdl $2,%r14d,%r14d | |
3279 addl %r9d,%ebx | |
3280 vpsrlq $2,%xmm7,%xmm7 | |
3281 addl %edi,%r9d | |
3282 movl %ebx,%r13d | |
3283 addl %r9d,%r14d | |
3284 vpxor %xmm7,%xmm6,%xmm6 | |
3285 shrdl $14,%r13d,%r13d | |
3286 movl %r14d,%r9d | |
3287 movl %ecx,%r12d | |
3288 vpshufb %xmm9,%xmm6,%xmm6 | |
3289 shrdl $9,%r14d,%r14d | |
3290 xorl %ebx,%r13d | |
3291 xorl %edx,%r12d | |
3292 vpaddd %xmm6,%xmm2,%xmm2 | |
3293 shrdl $5,%r13d,%r13d | |
3294 xorl %r9d,%r14d | |
3295 andl %ebx,%r12d | |
3296 vpaddd 64(%rbp),%xmm2,%xmm6 | |
3297 xorl %ebx,%r13d | |
3298 addl 44(%rsp),%r8d | |
3299 movl %r9d,%edi | |
3300 xorl %edx,%r12d | |
3301 shrdl $11,%r14d,%r14d | |
3302 xorl %r10d,%edi | |
3303 addl %r12d,%r8d | |
3304 shrdl $6,%r13d,%r13d | |
3305 andl %edi,%r15d | |
3306 xorl %r9d,%r14d | |
3307 addl %r13d,%r8d | |
3308 xorl %r10d,%r15d | |
3309 shrdl $2,%r14d,%r14d | |
3310 addl %r8d,%eax | |
3311 addl %r15d,%r8d | |
3312 movl %eax,%r13d | |
3313 addl %r8d,%r14d | |
3314 vmovdqa %xmm6,32(%rsp) | |
3315 vpalignr $4,%xmm3,%xmm0,%xmm4 | |
3316 shrdl $14,%r13d,%r13d | |
3317 movl %r14d,%r8d | |
3318 movl %ebx,%r12d | |
3319 vpalignr $4,%xmm1,%xmm2,%xmm7 | |
3320 shrdl $9,%r14d,%r14d | |
3321 xorl %eax,%r13d | |
3322 xorl %ecx,%r12d | |
3323 vpsrld $7,%xmm4,%xmm6 | |
3324 shrdl $5,%r13d,%r13d | |
3325 xorl %r8d,%r14d | |
3326 andl %eax,%r12d | |
3327 vpaddd %xmm7,%xmm3,%xmm3 | |
3328 xorl %eax,%r13d | |
3329 addl 48(%rsp),%edx | |
3330 movl %r8d,%r15d | |
3331 vpsrld $3,%xmm4,%xmm7 | |
3332 xorl %ecx,%r12d | |
3333 shrdl $11,%r14d,%r14d | |
3334 xorl %r9d,%r15d | |
3335 vpslld $14,%xmm4,%xmm5 | |
3336 addl %r12d,%edx | |
3337 shrdl $6,%r13d,%r13d | |
3338 andl %r15d,%edi | |
3339 vpxor %xmm6,%xmm7,%xmm4 | |
3340 xorl %r8d,%r14d | |
3341 addl %r13d,%edx | |
3342 xorl %r9d,%edi | |
3343 vpshufd $250,%xmm2,%xmm7 | |
3344 shrdl $2,%r14d,%r14d | |
3345 addl %edx,%r11d | |
3346 addl %edi,%edx | |
3347 vpsrld $11,%xmm6,%xmm6 | |
3348 movl %r11d,%r13d | |
3349 addl %edx,%r14d | |
3350 shrdl $14,%r13d,%r13d | |
3351 vpxor %xmm5,%xmm4,%xmm4 | |
3352 movl %r14d,%edx | |
3353 movl %eax,%r12d | |
3354 shrdl $9,%r14d,%r14d | |
3355 vpslld $11,%xmm5,%xmm5 | |
3356 xorl %r11d,%r13d | |
3357 xorl %ebx,%r12d | |
3358 shrdl $5,%r13d,%r13d | |
3359 vpxor %xmm6,%xmm4,%xmm4 | |
3360 xorl %edx,%r14d | |
3361 andl %r11d,%r12d | |
3362 xorl %r11d,%r13d | |
3363 vpsrld $10,%xmm7,%xmm6 | |
3364 addl 52(%rsp),%ecx | |
3365 movl %edx,%edi | |
3366 xorl %ebx,%r12d | |
3367 vpxor %xmm5,%xmm4,%xmm4 | |
3368 shrdl $11,%r14d,%r14d | |
3369 xorl %r8d,%edi | |
3370 addl %r12d,%ecx | |
3371 vpsrlq $17,%xmm7,%xmm7 | |
3372 shrdl $6,%r13d,%r13d | |
3373 andl %edi,%r15d | |
3374 xorl %edx,%r14d | |
3375 vpaddd %xmm4,%xmm3,%xmm3 | |
3376 addl %r13d,%ecx | |
3377 xorl %r8d,%r15d | |
3378 shrdl $2,%r14d,%r14d | |
3379 vpxor %xmm7,%xmm6,%xmm6 | |
3380 addl %ecx,%r10d | |
3381 addl %r15d,%ecx | |
3382 movl %r10d,%r13d | |
3383 vpsrlq $2,%xmm7,%xmm7 | |
3384 addl %ecx,%r14d | |
3385 shrdl $14,%r13d,%r13d | |
3386 movl %r14d,%ecx | |
3387 vpxor %xmm7,%xmm6,%xmm6 | |
3388 movl %r11d,%r12d | |
3389 shrdl $9,%r14d,%r14d | |
3390 xorl %r10d,%r13d | |
3391 vpshufb %xmm8,%xmm6,%xmm6 | |
3392 xorl %eax,%r12d | |
3393 shrdl $5,%r13d,%r13d | |
3394 xorl %ecx,%r14d | |
3395 vpaddd %xmm6,%xmm3,%xmm3 | |
3396 andl %r10d,%r12d | |
3397 xorl %r10d,%r13d | |
3398 addl 56(%rsp),%ebx | |
3399 vpshufd $80,%xmm3,%xmm7 | |
3400 movl %ecx,%r15d | |
3401 xorl %eax,%r12d | |
3402 shrdl $11,%r14d,%r14d | |
3403 vpsrld $10,%xmm7,%xmm6 | |
3404 xorl %edx,%r15d | |
3405 addl %r12d,%ebx | |
3406 shrdl $6,%r13d,%r13d | |
3407 vpsrlq $17,%xmm7,%xmm7 | |
3408 andl %r15d,%edi | |
3409 xorl %ecx,%r14d | |
3410 addl %r13d,%ebx | |
3411 vpxor %xmm7,%xmm6,%xmm6 | |
3412 xorl %edx,%edi | |
3413 shrdl $2,%r14d,%r14d | |
3414 addl %ebx,%r9d | |
3415 vpsrlq $2,%xmm7,%xmm7 | |
3416 addl %edi,%ebx | |
3417 movl %r9d,%r13d | |
3418 addl %ebx,%r14d | |
3419 vpxor %xmm7,%xmm6,%xmm6 | |
3420 shrdl $14,%r13d,%r13d | |
3421 movl %r14d,%ebx | |
3422 movl %r10d,%r12d | |
3423 vpshufb %xmm9,%xmm6,%xmm6 | |
3424 shrdl $9,%r14d,%r14d | |
3425 xorl %r9d,%r13d | |
3426 xorl %r11d,%r12d | |
3427 vpaddd %xmm6,%xmm3,%xmm3 | |
3428 shrdl $5,%r13d,%r13d | |
3429 xorl %ebx,%r14d | |
3430 andl %r9d,%r12d | |
3431 vpaddd 96(%rbp),%xmm3,%xmm6 | |
3432 xorl %r9d,%r13d | |
3433 addl 60(%rsp),%eax | |
3434 movl %ebx,%edi | |
3435 xorl %r11d,%r12d | |
3436 shrdl $11,%r14d,%r14d | |
3437 xorl %ecx,%edi | |
3438 addl %r12d,%eax | |
3439 shrdl $6,%r13d,%r13d | |
3440 andl %edi,%r15d | |
3441 xorl %ebx,%r14d | |
3442 addl %r13d,%eax | |
3443 xorl %ecx,%r15d | |
3444 shrdl $2,%r14d,%r14d | |
3445 addl %eax,%r8d | |
3446 addl %r15d,%eax | |
3447 movl %r8d,%r13d | |
3448 addl %eax,%r14d | |
3449 vmovdqa %xmm6,48(%rsp) | |
3450 cmpb $0,131(%rbp) | |
3451 jne L$avx_00_47 | |
3452 shrdl $14,%r13d,%r13d | |
3453 movl %r14d,%eax | |
3454 movl %r9d,%r12d | |
3455 shrdl $9,%r14d,%r14d | |
3456 xorl %r8d,%r13d | |
3457 xorl %r10d,%r12d | |
3458 shrdl $5,%r13d,%r13d | |
3459 xorl %eax,%r14d | |
3460 andl %r8d,%r12d | |
3461 xorl %r8d,%r13d | |
3462 addl 0(%rsp),%r11d | |
3463 movl %eax,%r15d | |
3464 xorl %r10d,%r12d | |
3465 shrdl $11,%r14d,%r14d | |
3466 xorl %ebx,%r15d | |
3467 addl %r12d,%r11d | |
3468 shrdl $6,%r13d,%r13d | |
3469 andl %r15d,%edi | |
3470 xorl %eax,%r14d | |
3471 addl %r13d,%r11d | |
3472 xorl %ebx,%edi | |
3473 shrdl $2,%r14d,%r14d | |
3474 addl %r11d,%edx | |
3475 addl %edi,%r11d | |
3476 movl %edx,%r13d | |
3477 addl %r11d,%r14d | |
3478 shrdl $14,%r13d,%r13d | |
3479 movl %r14d,%r11d | |
3480 movl %r8d,%r12d | |
3481 shrdl $9,%r14d,%r14d | |
3482 xorl %edx,%r13d | |
3483 xorl %r9d,%r12d | |
3484 shrdl $5,%r13d,%r13d | |
3485 xorl %r11d,%r14d | |
3486 andl %edx,%r12d | |
3487 xorl %edx,%r13d | |
3488 addl 4(%rsp),%r10d | |
3489 movl %r11d,%edi | |
3490 xorl %r9d,%r12d | |
3491 shrdl $11,%r14d,%r14d | |
3492 xorl %eax,%edi | |
3493 addl %r12d,%r10d | |
3494 shrdl $6,%r13d,%r13d | |
3495 andl %edi,%r15d | |
3496 xorl %r11d,%r14d | |
3497 addl %r13d,%r10d | |
3498 xorl %eax,%r15d | |
3499 shrdl $2,%r14d,%r14d | |
3500 addl %r10d,%ecx | |
3501 addl %r15d,%r10d | |
3502 movl %ecx,%r13d | |
3503 addl %r10d,%r14d | |
3504 shrdl $14,%r13d,%r13d | |
3505 movl %r14d,%r10d | |
3506 movl %edx,%r12d | |
3507 shrdl $9,%r14d,%r14d | |
3508 xorl %ecx,%r13d | |
3509 xorl %r8d,%r12d | |
3510 shrdl $5,%r13d,%r13d | |
3511 xorl %r10d,%r14d | |
3512 andl %ecx,%r12d | |
3513 xorl %ecx,%r13d | |
3514 addl 8(%rsp),%r9d | |
3515 movl %r10d,%r15d | |
3516 xorl %r8d,%r12d | |
3517 shrdl $11,%r14d,%r14d | |
3518 xorl %r11d,%r15d | |
3519 addl %r12d,%r9d | |
3520 shrdl $6,%r13d,%r13d | |
3521 andl %r15d,%edi | |
3522 xorl %r10d,%r14d | |
3523 addl %r13d,%r9d | |
3524 xorl %r11d,%edi | |
3525 shrdl $2,%r14d,%r14d | |
3526 addl %r9d,%ebx | |
3527 addl %edi,%r9d | |
3528 movl %ebx,%r13d | |
3529 addl %r9d,%r14d | |
3530 shrdl $14,%r13d,%r13d | |
3531 movl %r14d,%r9d | |
3532 movl %ecx,%r12d | |
3533 shrdl $9,%r14d,%r14d | |
3534 xorl %ebx,%r13d | |
3535 xorl %edx,%r12d | |
3536 shrdl $5,%r13d,%r13d | |
3537 xorl %r9d,%r14d | |
3538 andl %ebx,%r12d | |
3539 xorl %ebx,%r13d | |
3540 addl 12(%rsp),%r8d | |
3541 movl %r9d,%edi | |
3542 xorl %edx,%r12d | |
3543 shrdl $11,%r14d,%r14d | |
3544 xorl %r10d,%edi | |
3545 addl %r12d,%r8d | |
3546 shrdl $6,%r13d,%r13d | |
3547 andl %edi,%r15d | |
3548 xorl %r9d,%r14d | |
3549 addl %r13d,%r8d | |
3550 xorl %r10d,%r15d | |
3551 shrdl $2,%r14d,%r14d | |
3552 addl %r8d,%eax | |
3553 addl %r15d,%r8d | |
3554 movl %eax,%r13d | |
3555 addl %r8d,%r14d | |
3556 shrdl $14,%r13d,%r13d | |
3557 movl %r14d,%r8d | |
3558 movl %ebx,%r12d | |
3559 shrdl $9,%r14d,%r14d | |
3560 xorl %eax,%r13d | |
3561 xorl %ecx,%r12d | |
3562 shrdl $5,%r13d,%r13d | |
3563 xorl %r8d,%r14d | |
3564 andl %eax,%r12d | |
3565 xorl %eax,%r13d | |
3566 addl 16(%rsp),%edx | |
3567 movl %r8d,%r15d | |
3568 xorl %ecx,%r12d | |
3569 shrdl $11,%r14d,%r14d | |
3570 xorl %r9d,%r15d | |
3571 addl %r12d,%edx | |
3572 shrdl $6,%r13d,%r13d | |
3573 andl %r15d,%edi | |
3574 xorl %r8d,%r14d | |
3575 addl %r13d,%edx | |
3576 xorl %r9d,%edi | |
3577 shrdl $2,%r14d,%r14d | |
3578 addl %edx,%r11d | |
3579 addl %edi,%edx | |
3580 movl %r11d,%r13d | |
3581 addl %edx,%r14d | |
3582 shrdl $14,%r13d,%r13d | |
3583 movl %r14d,%edx | |
3584 movl %eax,%r12d | |
3585 shrdl $9,%r14d,%r14d | |
3586 xorl %r11d,%r13d | |
3587 xorl %ebx,%r12d | |
3588 shrdl $5,%r13d,%r13d | |
3589 xorl %edx,%r14d | |
3590 andl %r11d,%r12d | |
3591 xorl %r11d,%r13d | |
3592 addl 20(%rsp),%ecx | |
3593 movl %edx,%edi | |
3594 xorl %ebx,%r12d | |
3595 shrdl $11,%r14d,%r14d | |
3596 xorl %r8d,%edi | |
3597 addl %r12d,%ecx | |
3598 shrdl $6,%r13d,%r13d | |
3599 andl %edi,%r15d | |
3600 xorl %edx,%r14d | |
3601 addl %r13d,%ecx | |
3602 xorl %r8d,%r15d | |
3603 shrdl $2,%r14d,%r14d | |
3604 addl %ecx,%r10d | |
3605 addl %r15d,%ecx | |
3606 movl %r10d,%r13d | |
3607 addl %ecx,%r14d | |
3608 shrdl $14,%r13d,%r13d | |
3609 movl %r14d,%ecx | |
3610 movl %r11d,%r12d | |
3611 shrdl $9,%r14d,%r14d | |
3612 xorl %r10d,%r13d | |
3613 xorl %eax,%r12d | |
3614 shrdl $5,%r13d,%r13d | |
3615 xorl %ecx,%r14d | |
3616 andl %r10d,%r12d | |
3617 xorl %r10d,%r13d | |
3618 addl 24(%rsp),%ebx | |
3619 movl %ecx,%r15d | |
3620 xorl %eax,%r12d | |
3621 shrdl $11,%r14d,%r14d | |
3622 xorl %edx,%r15d | |
3623 addl %r12d,%ebx | |
3624 shrdl $6,%r13d,%r13d | |
3625 andl %r15d,%edi | |
3626 xorl %ecx,%r14d | |
3627 addl %r13d,%ebx | |
3628 xorl %edx,%edi | |
3629 shrdl $2,%r14d,%r14d | |
3630 addl %ebx,%r9d | |
3631 addl %edi,%ebx | |
3632 movl %r9d,%r13d | |
3633 addl %ebx,%r14d | |
3634 shrdl $14,%r13d,%r13d | |
3635 movl %r14d,%ebx | |
3636 movl %r10d,%r12d | |
3637 shrdl $9,%r14d,%r14d | |
3638 xorl %r9d,%r13d | |
3639 xorl %r11d,%r12d | |
3640 shrdl $5,%r13d,%r13d | |
3641 xorl %ebx,%r14d | |
3642 andl %r9d,%r12d | |
3643 xorl %r9d,%r13d | |
3644 addl 28(%rsp),%eax | |
3645 movl %ebx,%edi | |
3646 xorl %r11d,%r12d | |
3647 shrdl $11,%r14d,%r14d | |
3648 xorl %ecx,%edi | |
3649 addl %r12d,%eax | |
3650 shrdl $6,%r13d,%r13d | |
3651 andl %edi,%r15d | |
3652 xorl %ebx,%r14d | |
3653 addl %r13d,%eax | |
3654 xorl %ecx,%r15d | |
3655 shrdl $2,%r14d,%r14d | |
3656 addl %eax,%r8d | |
3657 addl %r15d,%eax | |
3658 movl %r8d,%r13d | |
3659 addl %eax,%r14d | |
3660 shrdl $14,%r13d,%r13d | |
3661 movl %r14d,%eax | |
3662 movl %r9d,%r12d | |
3663 shrdl $9,%r14d,%r14d | |
3664 xorl %r8d,%r13d | |
3665 xorl %r10d,%r12d | |
3666 shrdl $5,%r13d,%r13d | |
3667 xorl %eax,%r14d | |
3668 andl %r8d,%r12d | |
3669 xorl %r8d,%r13d | |
3670 addl 32(%rsp),%r11d | |
3671 movl %eax,%r15d | |
3672 xorl %r10d,%r12d | |
3673 shrdl $11,%r14d,%r14d | |
3674 xorl %ebx,%r15d | |
3675 addl %r12d,%r11d | |
3676 shrdl $6,%r13d,%r13d | |
3677 andl %r15d,%edi | |
3678 xorl %eax,%r14d | |
3679 addl %r13d,%r11d | |
3680 xorl %ebx,%edi | |
3681 shrdl $2,%r14d,%r14d | |
3682 addl %r11d,%edx | |
3683 addl %edi,%r11d | |
3684 movl %edx,%r13d | |
3685 addl %r11d,%r14d | |
3686 shrdl $14,%r13d,%r13d | |
3687 movl %r14d,%r11d | |
3688 movl %r8d,%r12d | |
3689 shrdl $9,%r14d,%r14d | |
3690 xorl %edx,%r13d | |
3691 xorl %r9d,%r12d | |
3692 shrdl $5,%r13d,%r13d | |
3693 xorl %r11d,%r14d | |
3694 andl %edx,%r12d | |
3695 xorl %edx,%r13d | |
3696 addl 36(%rsp),%r10d | |
3697 movl %r11d,%edi | |
3698 xorl %r9d,%r12d | |
3699 shrdl $11,%r14d,%r14d | |
3700 xorl %eax,%edi | |
3701 addl %r12d,%r10d | |
3702 shrdl $6,%r13d,%r13d | |
3703 andl %edi,%r15d | |
3704 xorl %r11d,%r14d | |
3705 addl %r13d,%r10d | |
3706 xorl %eax,%r15d | |
3707 shrdl $2,%r14d,%r14d | |
3708 addl %r10d,%ecx | |
3709 addl %r15d,%r10d | |
3710 movl %ecx,%r13d | |
3711 addl %r10d,%r14d | |
3712 shrdl $14,%r13d,%r13d | |
3713 movl %r14d,%r10d | |
3714 movl %edx,%r12d | |
3715 shrdl $9,%r14d,%r14d | |
3716 xorl %ecx,%r13d | |
3717 xorl %r8d,%r12d | |
3718 shrdl $5,%r13d,%r13d | |
3719 xorl %r10d,%r14d | |
3720 andl %ecx,%r12d | |
3721 xorl %ecx,%r13d | |
3722 addl 40(%rsp),%r9d | |
3723 movl %r10d,%r15d | |
3724 xorl %r8d,%r12d | |
3725 shrdl $11,%r14d,%r14d | |
3726 xorl %r11d,%r15d | |
3727 addl %r12d,%r9d | |
3728 shrdl $6,%r13d,%r13d | |
3729 andl %r15d,%edi | |
3730 xorl %r10d,%r14d | |
3731 addl %r13d,%r9d | |
3732 xorl %r11d,%edi | |
3733 shrdl $2,%r14d,%r14d | |
3734 addl %r9d,%ebx | |
3735 addl %edi,%r9d | |
3736 movl %ebx,%r13d | |
3737 addl %r9d,%r14d | |
3738 shrdl $14,%r13d,%r13d | |
3739 movl %r14d,%r9d | |
3740 movl %ecx,%r12d | |
3741 shrdl $9,%r14d,%r14d | |
3742 xorl %ebx,%r13d | |
3743 xorl %edx,%r12d | |
3744 shrdl $5,%r13d,%r13d | |
3745 xorl %r9d,%r14d | |
3746 andl %ebx,%r12d | |
3747 xorl %ebx,%r13d | |
3748 addl 44(%rsp),%r8d | |
3749 movl %r9d,%edi | |
3750 xorl %edx,%r12d | |
3751 shrdl $11,%r14d,%r14d | |
3752 xorl %r10d,%edi | |
3753 addl %r12d,%r8d | |
3754 shrdl $6,%r13d,%r13d | |
3755 andl %edi,%r15d | |
3756 xorl %r9d,%r14d | |
3757 addl %r13d,%r8d | |
3758 xorl %r10d,%r15d | |
3759 shrdl $2,%r14d,%r14d | |
3760 addl %r8d,%eax | |
3761 addl %r15d,%r8d | |
3762 movl %eax,%r13d | |
3763 addl %r8d,%r14d | |
3764 shrdl $14,%r13d,%r13d | |
3765 movl %r14d,%r8d | |
3766 movl %ebx,%r12d | |
3767 shrdl $9,%r14d,%r14d | |
3768 xorl %eax,%r13d | |
3769 xorl %ecx,%r12d | |
3770 shrdl $5,%r13d,%r13d | |
3771 xorl %r8d,%r14d | |
3772 andl %eax,%r12d | |
3773 xorl %eax,%r13d | |
3774 addl 48(%rsp),%edx | |
3775 movl %r8d,%r15d | |
3776 xorl %ecx,%r12d | |
3777 shrdl $11,%r14d,%r14d | |
3778 xorl %r9d,%r15d | |
3779 addl %r12d,%edx | |
3780 shrdl $6,%r13d,%r13d | |
3781 andl %r15d,%edi | |
3782 xorl %r8d,%r14d | |
3783 addl %r13d,%edx | |
3784 xorl %r9d,%edi | |
3785 shrdl $2,%r14d,%r14d | |
3786 addl %edx,%r11d | |
3787 addl %edi,%edx | |
3788 movl %r11d,%r13d | |
3789 addl %edx,%r14d | |
3790 shrdl $14,%r13d,%r13d | |
3791 movl %r14d,%edx | |
3792 movl %eax,%r12d | |
3793 shrdl $9,%r14d,%r14d | |
3794 xorl %r11d,%r13d | |
3795 xorl %ebx,%r12d | |
3796 shrdl $5,%r13d,%r13d | |
3797 xorl %edx,%r14d | |
3798 andl %r11d,%r12d | |
3799 xorl %r11d,%r13d | |
3800 addl 52(%rsp),%ecx | |
3801 movl %edx,%edi | |
3802 xorl %ebx,%r12d | |
3803 shrdl $11,%r14d,%r14d | |
3804 xorl %r8d,%edi | |
3805 addl %r12d,%ecx | |
3806 shrdl $6,%r13d,%r13d | |
3807 andl %edi,%r15d | |
3808 xorl %edx,%r14d | |
3809 addl %r13d,%ecx | |
3810 xorl %r8d,%r15d | |
3811 shrdl $2,%r14d,%r14d | |
3812 addl %ecx,%r10d | |
3813 addl %r15d,%ecx | |
3814 movl %r10d,%r13d | |
3815 addl %ecx,%r14d | |
3816 shrdl $14,%r13d,%r13d | |
3817 movl %r14d,%ecx | |
3818 movl %r11d,%r12d | |
3819 shrdl $9,%r14d,%r14d | |
3820 xorl %r10d,%r13d | |
3821 xorl %eax,%r12d | |
3822 shrdl $5,%r13d,%r13d | |
3823 xorl %ecx,%r14d | |
3824 andl %r10d,%r12d | |
3825 xorl %r10d,%r13d | |
3826 addl 56(%rsp),%ebx | |
3827 movl %ecx,%r15d | |
3828 xorl %eax,%r12d | |
3829 shrdl $11,%r14d,%r14d | |
3830 xorl %edx,%r15d | |
3831 addl %r12d,%ebx | |
3832 shrdl $6,%r13d,%r13d | |
3833 andl %r15d,%edi | |
3834 xorl %ecx,%r14d | |
3835 addl %r13d,%ebx | |
3836 xorl %edx,%edi | |
3837 shrdl $2,%r14d,%r14d | |
3838 addl %ebx,%r9d | |
3839 addl %edi,%ebx | |
3840 movl %r9d,%r13d | |
3841 addl %ebx,%r14d | |
3842 shrdl $14,%r13d,%r13d | |
3843 movl %r14d,%ebx | |
3844 movl %r10d,%r12d | |
3845 shrdl $9,%r14d,%r14d | |
3846 xorl %r9d,%r13d | |
3847 xorl %r11d,%r12d | |
3848 shrdl $5,%r13d,%r13d | |
3849 xorl %ebx,%r14d | |
3850 andl %r9d,%r12d | |
3851 xorl %r9d,%r13d | |
3852 addl 60(%rsp),%eax | |
3853 movl %ebx,%edi | |
3854 xorl %r11d,%r12d | |
3855 shrdl $11,%r14d,%r14d | |
3856 xorl %ecx,%edi | |
3857 addl %r12d,%eax | |
3858 shrdl $6,%r13d,%r13d | |
3859 andl %edi,%r15d | |
3860 xorl %ebx,%r14d | |
3861 addl %r13d,%eax | |
3862 xorl %ecx,%r15d | |
3863 shrdl $2,%r14d,%r14d | |
3864 addl %eax,%r8d | |
3865 addl %r15d,%eax | |
3866 movl %r8d,%r13d | |
3867 addl %eax,%r14d | |
3868 movq 64+0(%rsp),%rdi | |
3869 movl %r14d,%eax | |
3870 | |
3871 addl 0(%rdi),%eax | |
3872 leaq 64(%rsi),%rsi | |
3873 addl 4(%rdi),%ebx | |
3874 addl 8(%rdi),%ecx | |
3875 addl 12(%rdi),%edx | |
3876 addl 16(%rdi),%r8d | |
3877 addl 20(%rdi),%r9d | |
3878 addl 24(%rdi),%r10d | |
3879 addl 28(%rdi),%r11d | |
3880 | |
3881 cmpq 64+16(%rsp),%rsi | |
3882 | |
3883 movl %eax,0(%rdi) | |
3884 movl %ebx,4(%rdi) | |
3885 movl %ecx,8(%rdi) | |
3886 movl %edx,12(%rdi) | |
3887 movl %r8d,16(%rdi) | |
3888 movl %r9d,20(%rdi) | |
3889 movl %r10d,24(%rdi) | |
3890 movl %r11d,28(%rdi) | |
3891 jb L$loop_avx | |
3892 | |
3893 movq 64+24(%rsp),%rsi | |
3894 vzeroupper | |
3895 movq (%rsi),%r15 | |
3896 movq 8(%rsi),%r14 | |
3897 movq 16(%rsi),%r13 | |
3898 movq 24(%rsi),%r12 | |
3899 movq 32(%rsi),%rbp | |
3900 movq 40(%rsi),%rbx | |
3901 leaq 48(%rsi),%rsp | |
3902 L$epilogue_avx: | |
3903 .byte 0xf3,0xc3 | |
3904 | |
3905 #endif | 2843 #endif |
OLD | NEW |