| default rel |
| %define XMMWORD |
| %define YMMWORD |
| %define ZMMWORD |
| section .text code align=64 |
| |
| EXTERN OPENSSL_ia32cap_P |
| global aes_hw_encrypt |
| |
| ALIGN 16 |
| aes_hw_encrypt: |
| movups xmm2,XMMWORD[rcx] |
| mov eax,DWORD[240+r8] |
| movups xmm0,XMMWORD[r8] |
| movups xmm1,XMMWORD[16+r8] |
| lea r8,[32+r8] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_1: |
| DB 102,15,56,220,209 |
| dec eax |
| movups xmm1,XMMWORD[r8] |
| lea r8,[16+r8] |
| jnz NEAR $L$oop_enc1_1 |
| DB 102,15,56,221,209 |
| pxor xmm0,xmm0 |
| pxor xmm1,xmm1 |
| movups XMMWORD[rdx],xmm2 |
| pxor xmm2,xmm2 |
| DB 0F3h,0C3h ;repret |
| |
| |
| global aes_hw_decrypt |
| |
| ALIGN 16 |
| aes_hw_decrypt: |
| movups xmm2,XMMWORD[rcx] |
| mov eax,DWORD[240+r8] |
| movups xmm0,XMMWORD[r8] |
| movups xmm1,XMMWORD[16+r8] |
| lea r8,[32+r8] |
| xorps xmm2,xmm0 |
| $L$oop_dec1_2: |
| DB 102,15,56,222,209 |
| dec eax |
| movups xmm1,XMMWORD[r8] |
| lea r8,[16+r8] |
| jnz NEAR $L$oop_dec1_2 |
| DB 102,15,56,223,209 |
| pxor xmm0,xmm0 |
| pxor xmm1,xmm1 |
| movups XMMWORD[rdx],xmm2 |
| pxor xmm2,xmm2 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_encrypt2: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| movups xmm0,XMMWORD[32+rcx] |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| add rax,16 |
| |
| $L$enc_loop2: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$enc_loop2 |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,221,208 |
| DB 102,15,56,221,216 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_decrypt2: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| movups xmm0,XMMWORD[32+rcx] |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| add rax,16 |
| |
| $L$dec_loop2: |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,222,208 |
| DB 102,15,56,222,216 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$dec_loop2 |
| |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,223,208 |
| DB 102,15,56,223,216 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_encrypt3: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| xorps xmm4,xmm0 |
| movups xmm0,XMMWORD[32+rcx] |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| add rax,16 |
| |
| $L$enc_loop3: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$enc_loop3 |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,221,208 |
| DB 102,15,56,221,216 |
| DB 102,15,56,221,224 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_decrypt3: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| xorps xmm4,xmm0 |
| movups xmm0,XMMWORD[32+rcx] |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| add rax,16 |
| |
| $L$dec_loop3: |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,222,208 |
| DB 102,15,56,222,216 |
| DB 102,15,56,222,224 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$dec_loop3 |
| |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| DB 102,15,56,223,208 |
| DB 102,15,56,223,216 |
| DB 102,15,56,223,224 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_encrypt4: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| xorps xmm4,xmm0 |
| xorps xmm5,xmm0 |
| movups xmm0,XMMWORD[32+rcx] |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| DB 0x0f,0x1f,0x00 |
| add rax,16 |
| |
| $L$enc_loop4: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$enc_loop4 |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,221,208 |
| DB 102,15,56,221,216 |
| DB 102,15,56,221,224 |
| DB 102,15,56,221,232 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_decrypt4: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| xorps xmm4,xmm0 |
| xorps xmm5,xmm0 |
| movups xmm0,XMMWORD[32+rcx] |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| DB 0x0f,0x1f,0x00 |
| add rax,16 |
| |
| $L$dec_loop4: |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,222,208 |
| DB 102,15,56,222,216 |
| DB 102,15,56,222,224 |
| DB 102,15,56,222,232 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$dec_loop4 |
| |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| DB 102,15,56,223,208 |
| DB 102,15,56,223,216 |
| DB 102,15,56,223,224 |
| DB 102,15,56,223,232 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_encrypt6: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| pxor xmm3,xmm0 |
| pxor xmm4,xmm0 |
| DB 102,15,56,220,209 |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| DB 102,15,56,220,217 |
| pxor xmm5,xmm0 |
| pxor xmm6,xmm0 |
| DB 102,15,56,220,225 |
| pxor xmm7,xmm0 |
| movups xmm0,XMMWORD[rax*1+rcx] |
| add rax,16 |
| jmp NEAR $L$enc_loop6_enter |
| ALIGN 16 |
| $L$enc_loop6: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| $L$enc_loop6_enter: |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$enc_loop6 |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,15,56,221,208 |
| DB 102,15,56,221,216 |
| DB 102,15,56,221,224 |
| DB 102,15,56,221,232 |
| DB 102,15,56,221,240 |
| DB 102,15,56,221,248 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_decrypt6: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| pxor xmm3,xmm0 |
| pxor xmm4,xmm0 |
| DB 102,15,56,222,209 |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| DB 102,15,56,222,217 |
| pxor xmm5,xmm0 |
| pxor xmm6,xmm0 |
| DB 102,15,56,222,225 |
| pxor xmm7,xmm0 |
| movups xmm0,XMMWORD[rax*1+rcx] |
| add rax,16 |
| jmp NEAR $L$dec_loop6_enter |
| ALIGN 16 |
| $L$dec_loop6: |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| $L$dec_loop6_enter: |
| DB 102,15,56,222,233 |
| DB 102,15,56,222,241 |
| DB 102,15,56,222,249 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,222,208 |
| DB 102,15,56,222,216 |
| DB 102,15,56,222,224 |
| DB 102,15,56,222,232 |
| DB 102,15,56,222,240 |
| DB 102,15,56,222,248 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$dec_loop6 |
| |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| DB 102,15,56,222,241 |
| DB 102,15,56,222,249 |
| DB 102,15,56,223,208 |
| DB 102,15,56,223,216 |
| DB 102,15,56,223,224 |
| DB 102,15,56,223,232 |
| DB 102,15,56,223,240 |
| DB 102,15,56,223,248 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_encrypt8: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| pxor xmm4,xmm0 |
| pxor xmm5,xmm0 |
| pxor xmm6,xmm0 |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| DB 102,15,56,220,209 |
| pxor xmm7,xmm0 |
| pxor xmm8,xmm0 |
| DB 102,15,56,220,217 |
| pxor xmm9,xmm0 |
| movups xmm0,XMMWORD[rax*1+rcx] |
| add rax,16 |
| jmp NEAR $L$enc_loop8_inner |
| ALIGN 16 |
| $L$enc_loop8: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| $L$enc_loop8_inner: |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| $L$enc_loop8_enter: |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| DB 102,68,15,56,220,192 |
| DB 102,68,15,56,220,200 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$enc_loop8 |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| DB 102,15,56,221,208 |
| DB 102,15,56,221,216 |
| DB 102,15,56,221,224 |
| DB 102,15,56,221,232 |
| DB 102,15,56,221,240 |
| DB 102,15,56,221,248 |
| DB 102,68,15,56,221,192 |
| DB 102,68,15,56,221,200 |
| DB 0F3h,0C3h ;repret |
| |
| |
| ALIGN 16 |
| _aesni_decrypt8: |
| movups xmm0,XMMWORD[rcx] |
| shl eax,4 |
| movups xmm1,XMMWORD[16+rcx] |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm0 |
| pxor xmm4,xmm0 |
| pxor xmm5,xmm0 |
| pxor xmm6,xmm0 |
| lea rcx,[32+rax*1+rcx] |
| neg rax |
| DB 102,15,56,222,209 |
| pxor xmm7,xmm0 |
| pxor xmm8,xmm0 |
| DB 102,15,56,222,217 |
| pxor xmm9,xmm0 |
| movups xmm0,XMMWORD[rax*1+rcx] |
| add rax,16 |
| jmp NEAR $L$dec_loop8_inner |
| ALIGN 16 |
| $L$dec_loop8: |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| $L$dec_loop8_inner: |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| DB 102,15,56,222,241 |
| DB 102,15,56,222,249 |
| DB 102,68,15,56,222,193 |
| DB 102,68,15,56,222,201 |
| $L$dec_loop8_enter: |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,222,208 |
| DB 102,15,56,222,216 |
| DB 102,15,56,222,224 |
| DB 102,15,56,222,232 |
| DB 102,15,56,222,240 |
| DB 102,15,56,222,248 |
| DB 102,68,15,56,222,192 |
| DB 102,68,15,56,222,200 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$dec_loop8 |
| |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| DB 102,15,56,222,241 |
| DB 102,15,56,222,249 |
| DB 102,68,15,56,222,193 |
| DB 102,68,15,56,222,201 |
| DB 102,15,56,223,208 |
| DB 102,15,56,223,216 |
| DB 102,15,56,223,224 |
| DB 102,15,56,223,232 |
| DB 102,15,56,223,240 |
| DB 102,15,56,223,248 |
| DB 102,68,15,56,223,192 |
| DB 102,68,15,56,223,200 |
| DB 0F3h,0C3h ;repret |
| |
| global aes_hw_ecb_encrypt |
| |
| ALIGN 16 |
| aes_hw_ecb_encrypt: |
| mov QWORD[8+rsp],rdi ;WIN64 prologue |
| mov QWORD[16+rsp],rsi |
| mov rax,rsp |
| $L$SEH_begin_aes_hw_ecb_encrypt: |
| mov rdi,rcx |
| mov rsi,rdx |
| mov rdx,r8 |
| mov rcx,r9 |
| mov r8,QWORD[40+rsp] |
| |
| |
| lea rsp,[((-88))+rsp] |
| movaps XMMWORD[rsp],xmm6 |
| movaps XMMWORD[16+rsp],xmm7 |
| movaps XMMWORD[32+rsp],xmm8 |
| movaps XMMWORD[48+rsp],xmm9 |
| $L$ecb_enc_body: |
| and rdx,-16 |
| jz NEAR $L$ecb_ret |
| |
| mov eax,DWORD[240+rcx] |
| movups xmm0,XMMWORD[rcx] |
| mov r11,rcx |
| mov r10d,eax |
| test r8d,r8d |
| jz NEAR $L$ecb_decrypt |
| |
| cmp rdx,0x80 |
| jb NEAR $L$ecb_enc_tail |
| |
| movdqu xmm2,XMMWORD[rdi] |
| movdqu xmm3,XMMWORD[16+rdi] |
| movdqu xmm4,XMMWORD[32+rdi] |
| movdqu xmm5,XMMWORD[48+rdi] |
| movdqu xmm6,XMMWORD[64+rdi] |
| movdqu xmm7,XMMWORD[80+rdi] |
| movdqu xmm8,XMMWORD[96+rdi] |
| movdqu xmm9,XMMWORD[112+rdi] |
| lea rdi,[128+rdi] |
| sub rdx,0x80 |
| jmp NEAR $L$ecb_enc_loop8_enter |
| ALIGN 16 |
| $L$ecb_enc_loop8: |
| movups XMMWORD[rsi],xmm2 |
| mov rcx,r11 |
| movdqu xmm2,XMMWORD[rdi] |
| mov eax,r10d |
| movups XMMWORD[16+rsi],xmm3 |
| movdqu xmm3,XMMWORD[16+rdi] |
| movups XMMWORD[32+rsi],xmm4 |
| movdqu xmm4,XMMWORD[32+rdi] |
| movups XMMWORD[48+rsi],xmm5 |
| movdqu xmm5,XMMWORD[48+rdi] |
| movups XMMWORD[64+rsi],xmm6 |
| movdqu xmm6,XMMWORD[64+rdi] |
| movups XMMWORD[80+rsi],xmm7 |
| movdqu xmm7,XMMWORD[80+rdi] |
| movups XMMWORD[96+rsi],xmm8 |
| movdqu xmm8,XMMWORD[96+rdi] |
| movups XMMWORD[112+rsi],xmm9 |
| lea rsi,[128+rsi] |
| movdqu xmm9,XMMWORD[112+rdi] |
| lea rdi,[128+rdi] |
| $L$ecb_enc_loop8_enter: |
| |
| call _aesni_encrypt8 |
| |
| sub rdx,0x80 |
| jnc NEAR $L$ecb_enc_loop8 |
| |
| movups XMMWORD[rsi],xmm2 |
| mov rcx,r11 |
| movups XMMWORD[16+rsi],xmm3 |
| mov eax,r10d |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| movups XMMWORD[80+rsi],xmm7 |
| movups XMMWORD[96+rsi],xmm8 |
| movups XMMWORD[112+rsi],xmm9 |
| lea rsi,[128+rsi] |
| add rdx,0x80 |
| jz NEAR $L$ecb_ret |
| |
| $L$ecb_enc_tail: |
| movups xmm2,XMMWORD[rdi] |
| cmp rdx,0x20 |
| jb NEAR $L$ecb_enc_one |
| movups xmm3,XMMWORD[16+rdi] |
| je NEAR $L$ecb_enc_two |
| movups xmm4,XMMWORD[32+rdi] |
| cmp rdx,0x40 |
| jb NEAR $L$ecb_enc_three |
| movups xmm5,XMMWORD[48+rdi] |
| je NEAR $L$ecb_enc_four |
| movups xmm6,XMMWORD[64+rdi] |
| cmp rdx,0x60 |
| jb NEAR $L$ecb_enc_five |
| movups xmm7,XMMWORD[80+rdi] |
| je NEAR $L$ecb_enc_six |
| movdqu xmm8,XMMWORD[96+rdi] |
| xorps xmm9,xmm9 |
| call _aesni_encrypt8 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| movups XMMWORD[80+rsi],xmm7 |
| movups XMMWORD[96+rsi],xmm8 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_enc_one: |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_3: |
| DB 102,15,56,220,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_enc1_3 |
| DB 102,15,56,221,209 |
| movups XMMWORD[rsi],xmm2 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_enc_two: |
| call _aesni_encrypt2 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_enc_three: |
| call _aesni_encrypt3 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_enc_four: |
| call _aesni_encrypt4 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_enc_five: |
| xorps xmm7,xmm7 |
| call _aesni_encrypt6 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_enc_six: |
| call _aesni_encrypt6 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| movups XMMWORD[80+rsi],xmm7 |
| jmp NEAR $L$ecb_ret |
| |
| ALIGN 16 |
| $L$ecb_decrypt: |
| cmp rdx,0x80 |
| jb NEAR $L$ecb_dec_tail |
| |
| movdqu xmm2,XMMWORD[rdi] |
| movdqu xmm3,XMMWORD[16+rdi] |
| movdqu xmm4,XMMWORD[32+rdi] |
| movdqu xmm5,XMMWORD[48+rdi] |
| movdqu xmm6,XMMWORD[64+rdi] |
| movdqu xmm7,XMMWORD[80+rdi] |
| movdqu xmm8,XMMWORD[96+rdi] |
| movdqu xmm9,XMMWORD[112+rdi] |
| lea rdi,[128+rdi] |
| sub rdx,0x80 |
| jmp NEAR $L$ecb_dec_loop8_enter |
| ALIGN 16 |
| $L$ecb_dec_loop8: |
| movups XMMWORD[rsi],xmm2 |
| mov rcx,r11 |
| movdqu xmm2,XMMWORD[rdi] |
| mov eax,r10d |
| movups XMMWORD[16+rsi],xmm3 |
| movdqu xmm3,XMMWORD[16+rdi] |
| movups XMMWORD[32+rsi],xmm4 |
| movdqu xmm4,XMMWORD[32+rdi] |
| movups XMMWORD[48+rsi],xmm5 |
| movdqu xmm5,XMMWORD[48+rdi] |
| movups XMMWORD[64+rsi],xmm6 |
| movdqu xmm6,XMMWORD[64+rdi] |
| movups XMMWORD[80+rsi],xmm7 |
| movdqu xmm7,XMMWORD[80+rdi] |
| movups XMMWORD[96+rsi],xmm8 |
| movdqu xmm8,XMMWORD[96+rdi] |
| movups XMMWORD[112+rsi],xmm9 |
| lea rsi,[128+rsi] |
| movdqu xmm9,XMMWORD[112+rdi] |
| lea rdi,[128+rdi] |
| $L$ecb_dec_loop8_enter: |
| |
| call _aesni_decrypt8 |
| |
| movups xmm0,XMMWORD[r11] |
| sub rdx,0x80 |
| jnc NEAR $L$ecb_dec_loop8 |
| |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| mov rcx,r11 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm3,xmm3 |
| mov eax,r10d |
| movups XMMWORD[32+rsi],xmm4 |
| pxor xmm4,xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| pxor xmm5,xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| pxor xmm6,xmm6 |
| movups XMMWORD[80+rsi],xmm7 |
| pxor xmm7,xmm7 |
| movups XMMWORD[96+rsi],xmm8 |
| pxor xmm8,xmm8 |
| movups XMMWORD[112+rsi],xmm9 |
| pxor xmm9,xmm9 |
| lea rsi,[128+rsi] |
| add rdx,0x80 |
| jz NEAR $L$ecb_ret |
| |
| $L$ecb_dec_tail: |
| movups xmm2,XMMWORD[rdi] |
| cmp rdx,0x20 |
| jb NEAR $L$ecb_dec_one |
| movups xmm3,XMMWORD[16+rdi] |
| je NEAR $L$ecb_dec_two |
| movups xmm4,XMMWORD[32+rdi] |
| cmp rdx,0x40 |
| jb NEAR $L$ecb_dec_three |
| movups xmm5,XMMWORD[48+rdi] |
| je NEAR $L$ecb_dec_four |
| movups xmm6,XMMWORD[64+rdi] |
| cmp rdx,0x60 |
| jb NEAR $L$ecb_dec_five |
| movups xmm7,XMMWORD[80+rdi] |
| je NEAR $L$ecb_dec_six |
| movups xmm8,XMMWORD[96+rdi] |
| movups xmm0,XMMWORD[rcx] |
| xorps xmm9,xmm9 |
| call _aesni_decrypt8 |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm3,xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| pxor xmm4,xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| pxor xmm5,xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| pxor xmm6,xmm6 |
| movups XMMWORD[80+rsi],xmm7 |
| pxor xmm7,xmm7 |
| movups XMMWORD[96+rsi],xmm8 |
| pxor xmm8,xmm8 |
| pxor xmm9,xmm9 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_dec_one: |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_dec1_4: |
| DB 102,15,56,222,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_dec1_4 |
| DB 102,15,56,223,209 |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_dec_two: |
| call _aesni_decrypt2 |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm3,xmm3 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_dec_three: |
| call _aesni_decrypt3 |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm3,xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| pxor xmm4,xmm4 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_dec_four: |
| call _aesni_decrypt4 |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm3,xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| pxor xmm4,xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| pxor xmm5,xmm5 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_dec_five: |
| xorps xmm7,xmm7 |
| call _aesni_decrypt6 |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm3,xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| pxor xmm4,xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| pxor xmm5,xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| pxor xmm6,xmm6 |
| pxor xmm7,xmm7 |
| jmp NEAR $L$ecb_ret |
| ALIGN 16 |
| $L$ecb_dec_six: |
| call _aesni_decrypt6 |
| movups XMMWORD[rsi],xmm2 |
| pxor xmm2,xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm3,xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| pxor xmm4,xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| pxor xmm5,xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| pxor xmm6,xmm6 |
| movups XMMWORD[80+rsi],xmm7 |
| pxor xmm7,xmm7 |
| |
| $L$ecb_ret: |
| xorps xmm0,xmm0 |
| pxor xmm1,xmm1 |
| movaps xmm6,XMMWORD[rsp] |
| movaps XMMWORD[rsp],xmm0 |
| movaps xmm7,XMMWORD[16+rsp] |
| movaps XMMWORD[16+rsp],xmm0 |
| movaps xmm8,XMMWORD[32+rsp] |
| movaps XMMWORD[32+rsp],xmm0 |
| movaps xmm9,XMMWORD[48+rsp] |
| movaps XMMWORD[48+rsp],xmm0 |
| lea rsp,[88+rsp] |
| $L$ecb_enc_ret: |
| mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| mov rsi,QWORD[16+rsp] |
| DB 0F3h,0C3h ;repret |
| $L$SEH_end_aes_hw_ecb_encrypt: |
| global aes_hw_ccm64_encrypt_blocks |
| |
| ALIGN 16 |
| aes_hw_ccm64_encrypt_blocks: |
| mov QWORD[8+rsp],rdi ;WIN64 prologue |
| mov QWORD[16+rsp],rsi |
| mov rax,rsp |
| $L$SEH_begin_aes_hw_ccm64_encrypt_blocks: |
| mov rdi,rcx |
| mov rsi,rdx |
| mov rdx,r8 |
| mov rcx,r9 |
| mov r8,QWORD[40+rsp] |
| mov r9,QWORD[48+rsp] |
| |
| |
| lea rsp,[((-88))+rsp] |
| movaps XMMWORD[rsp],xmm6 |
| movaps XMMWORD[16+rsp],xmm7 |
| movaps XMMWORD[32+rsp],xmm8 |
| movaps XMMWORD[48+rsp],xmm9 |
| $L$ccm64_enc_body: |
| mov eax,DWORD[240+rcx] |
| movdqu xmm6,XMMWORD[r8] |
| movdqa xmm9,XMMWORD[$L$increment64] |
| movdqa xmm7,XMMWORD[$L$bswap_mask] |
| |
| shl eax,4 |
| mov r10d,16 |
| lea r11,[rcx] |
| movdqu xmm3,XMMWORD[r9] |
| movdqa xmm2,xmm6 |
| lea rcx,[32+rax*1+rcx] |
| DB 102,15,56,0,247 |
| sub r10,rax |
| jmp NEAR $L$ccm64_enc_outer |
| ALIGN 16 |
| $L$ccm64_enc_outer: |
| movups xmm0,XMMWORD[r11] |
| mov rax,r10 |
| movups xmm8,XMMWORD[rdi] |
| |
| xorps xmm2,xmm0 |
| movups xmm1,XMMWORD[16+r11] |
| xorps xmm0,xmm8 |
| xorps xmm3,xmm0 |
| movups xmm0,XMMWORD[32+r11] |
| |
| $L$ccm64_enc2_loop: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$ccm64_enc2_loop |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| paddq xmm6,xmm9 |
| dec rdx |
| DB 102,15,56,221,208 |
| DB 102,15,56,221,216 |
| |
| lea rdi,[16+rdi] |
| xorps xmm8,xmm2 |
| movdqa xmm2,xmm6 |
| movups XMMWORD[rsi],xmm8 |
| DB 102,15,56,0,215 |
| lea rsi,[16+rsi] |
| jnz NEAR $L$ccm64_enc_outer |
| |
| pxor xmm0,xmm0 |
| pxor xmm1,xmm1 |
| pxor xmm2,xmm2 |
| movups XMMWORD[r9],xmm3 |
| pxor xmm3,xmm3 |
| pxor xmm8,xmm8 |
| pxor xmm6,xmm6 |
| movaps xmm6,XMMWORD[rsp] |
| movaps XMMWORD[rsp],xmm0 |
| movaps xmm7,XMMWORD[16+rsp] |
| movaps XMMWORD[16+rsp],xmm0 |
| movaps xmm8,XMMWORD[32+rsp] |
| movaps XMMWORD[32+rsp],xmm0 |
| movaps xmm9,XMMWORD[48+rsp] |
| movaps XMMWORD[48+rsp],xmm0 |
| lea rsp,[88+rsp] |
| $L$ccm64_enc_ret: |
| mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| mov rsi,QWORD[16+rsp] |
| DB 0F3h,0C3h ;repret |
| $L$SEH_end_aes_hw_ccm64_encrypt_blocks: |
| global aes_hw_ccm64_decrypt_blocks |
| |
| ALIGN 16 |
| aes_hw_ccm64_decrypt_blocks: |
| mov QWORD[8+rsp],rdi ;WIN64 prologue |
| mov QWORD[16+rsp],rsi |
| mov rax,rsp |
| $L$SEH_begin_aes_hw_ccm64_decrypt_blocks: |
| mov rdi,rcx |
| mov rsi,rdx |
| mov rdx,r8 |
| mov rcx,r9 |
| mov r8,QWORD[40+rsp] |
| mov r9,QWORD[48+rsp] |
| |
| |
| lea rsp,[((-88))+rsp] |
| movaps XMMWORD[rsp],xmm6 |
| movaps XMMWORD[16+rsp],xmm7 |
| movaps XMMWORD[32+rsp],xmm8 |
| movaps XMMWORD[48+rsp],xmm9 |
| $L$ccm64_dec_body: |
| mov eax,DWORD[240+rcx] |
| movups xmm6,XMMWORD[r8] |
| movdqu xmm3,XMMWORD[r9] |
| movdqa xmm9,XMMWORD[$L$increment64] |
| movdqa xmm7,XMMWORD[$L$bswap_mask] |
| |
| movaps xmm2,xmm6 |
| mov r10d,eax |
| mov r11,rcx |
| DB 102,15,56,0,247 |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_5: |
| DB 102,15,56,220,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_enc1_5 |
| DB 102,15,56,221,209 |
| shl r10d,4 |
| mov eax,16 |
| movups xmm8,XMMWORD[rdi] |
| paddq xmm6,xmm9 |
| lea rdi,[16+rdi] |
| sub rax,r10 |
| lea rcx,[32+r10*1+r11] |
| mov r10,rax |
| jmp NEAR $L$ccm64_dec_outer |
| ALIGN 16 |
| $L$ccm64_dec_outer: |
| xorps xmm8,xmm2 |
| movdqa xmm2,xmm6 |
| movups XMMWORD[rsi],xmm8 |
| lea rsi,[16+rsi] |
| DB 102,15,56,0,215 |
| |
| sub rdx,1 |
| jz NEAR $L$ccm64_dec_break |
| |
| movups xmm0,XMMWORD[r11] |
| mov rax,r10 |
| movups xmm1,XMMWORD[16+r11] |
| xorps xmm8,xmm0 |
| xorps xmm2,xmm0 |
| xorps xmm3,xmm8 |
| movups xmm0,XMMWORD[32+r11] |
| jmp NEAR $L$ccm64_dec2_loop |
| ALIGN 16 |
| $L$ccm64_dec2_loop: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$ccm64_dec2_loop |
| movups xmm8,XMMWORD[rdi] |
| paddq xmm6,xmm9 |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,221,208 |
| DB 102,15,56,221,216 |
| lea rdi,[16+rdi] |
| jmp NEAR $L$ccm64_dec_outer |
| |
| ALIGN 16 |
| $L$ccm64_dec_break: |
| |
| mov eax,DWORD[240+r11] |
| movups xmm0,XMMWORD[r11] |
| movups xmm1,XMMWORD[16+r11] |
| xorps xmm8,xmm0 |
| lea r11,[32+r11] |
| xorps xmm3,xmm8 |
| $L$oop_enc1_6: |
| DB 102,15,56,220,217 |
| dec eax |
| movups xmm1,XMMWORD[r11] |
| lea r11,[16+r11] |
| jnz NEAR $L$oop_enc1_6 |
| DB 102,15,56,221,217 |
| pxor xmm0,xmm0 |
| pxor xmm1,xmm1 |
| pxor xmm2,xmm2 |
| movups XMMWORD[r9],xmm3 |
| pxor xmm3,xmm3 |
| pxor xmm8,xmm8 |
| pxor xmm6,xmm6 |
| movaps xmm6,XMMWORD[rsp] |
| movaps XMMWORD[rsp],xmm0 |
| movaps xmm7,XMMWORD[16+rsp] |
| movaps XMMWORD[16+rsp],xmm0 |
| movaps xmm8,XMMWORD[32+rsp] |
| movaps XMMWORD[32+rsp],xmm0 |
| movaps xmm9,XMMWORD[48+rsp] |
| movaps XMMWORD[48+rsp],xmm0 |
| lea rsp,[88+rsp] |
| $L$ccm64_dec_ret: |
| mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| mov rsi,QWORD[16+rsp] |
| DB 0F3h,0C3h ;repret |
| $L$SEH_end_aes_hw_ccm64_decrypt_blocks: |
| global aes_hw_ctr32_encrypt_blocks |
| |
| ALIGN 16 |
| aes_hw_ctr32_encrypt_blocks: |
| mov QWORD[8+rsp],rdi ;WIN64 prologue |
| mov QWORD[16+rsp],rsi |
| mov rax,rsp |
| $L$SEH_begin_aes_hw_ctr32_encrypt_blocks: |
| mov rdi,rcx |
| mov rsi,rdx |
| mov rdx,r8 |
| mov rcx,r9 |
| mov r8,QWORD[40+rsp] |
| |
| |
| |
| cmp rdx,1 |
| jne NEAR $L$ctr32_bulk |
| |
| |
| |
| movups xmm2,XMMWORD[r8] |
| movups xmm3,XMMWORD[rdi] |
| mov edx,DWORD[240+rcx] |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_7: |
| DB 102,15,56,220,209 |
| dec edx |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_enc1_7 |
| DB 102,15,56,221,209 |
| pxor xmm0,xmm0 |
| pxor xmm1,xmm1 |
| xorps xmm2,xmm3 |
| pxor xmm3,xmm3 |
| movups XMMWORD[rsi],xmm2 |
| xorps xmm2,xmm2 |
| jmp NEAR $L$ctr32_epilogue |
| |
| ALIGN 16 |
| $L$ctr32_bulk: |
| lea r11,[rsp] |
| |
| push rbp |
| |
| sub rsp,288 |
| and rsp,-16 |
| movaps XMMWORD[(-168)+r11],xmm6 |
| movaps XMMWORD[(-152)+r11],xmm7 |
| movaps XMMWORD[(-136)+r11],xmm8 |
| movaps XMMWORD[(-120)+r11],xmm9 |
| movaps XMMWORD[(-104)+r11],xmm10 |
| movaps XMMWORD[(-88)+r11],xmm11 |
| movaps XMMWORD[(-72)+r11],xmm12 |
| movaps XMMWORD[(-56)+r11],xmm13 |
| movaps XMMWORD[(-40)+r11],xmm14 |
| movaps XMMWORD[(-24)+r11],xmm15 |
| $L$ctr32_body: |
| |
| |
| |
| |
| movdqu xmm2,XMMWORD[r8] |
| movdqu xmm0,XMMWORD[rcx] |
| mov r8d,DWORD[12+r8] |
| pxor xmm2,xmm0 |
| mov ebp,DWORD[12+rcx] |
| movdqa XMMWORD[rsp],xmm2 |
| bswap r8d |
| movdqa xmm3,xmm2 |
| movdqa xmm4,xmm2 |
| movdqa xmm5,xmm2 |
| movdqa XMMWORD[64+rsp],xmm2 |
| movdqa XMMWORD[80+rsp],xmm2 |
| movdqa XMMWORD[96+rsp],xmm2 |
| mov r10,rdx |
| movdqa XMMWORD[112+rsp],xmm2 |
| |
| lea rax,[1+r8] |
| lea rdx,[2+r8] |
| bswap eax |
| bswap edx |
| xor eax,ebp |
| xor edx,ebp |
| DB 102,15,58,34,216,3 |
| lea rax,[3+r8] |
| movdqa XMMWORD[16+rsp],xmm3 |
| DB 102,15,58,34,226,3 |
| bswap eax |
| mov rdx,r10 |
| lea r10,[4+r8] |
| movdqa XMMWORD[32+rsp],xmm4 |
| xor eax,ebp |
| bswap r10d |
| DB 102,15,58,34,232,3 |
| xor r10d,ebp |
| movdqa XMMWORD[48+rsp],xmm5 |
| lea r9,[5+r8] |
| mov DWORD[((64+12))+rsp],r10d |
| bswap r9d |
| lea r10,[6+r8] |
| mov eax,DWORD[240+rcx] |
| xor r9d,ebp |
| bswap r10d |
| mov DWORD[((80+12))+rsp],r9d |
| xor r10d,ebp |
| lea r9,[7+r8] |
| mov DWORD[((96+12))+rsp],r10d |
| bswap r9d |
| lea r10,[OPENSSL_ia32cap_P] |
| mov r10d,DWORD[4+r10] |
| xor r9d,ebp |
| and r10d,71303168 |
| mov DWORD[((112+12))+rsp],r9d |
| |
| movups xmm1,XMMWORD[16+rcx] |
| |
| movdqa xmm6,XMMWORD[64+rsp] |
| movdqa xmm7,XMMWORD[80+rsp] |
| |
| cmp rdx,8 |
| jb NEAR $L$ctr32_tail |
| |
| sub rdx,6 |
| cmp r10d,4194304 |
| je NEAR $L$ctr32_6x |
| |
| lea rcx,[128+rcx] |
| sub rdx,2 |
| jmp NEAR $L$ctr32_loop8 |
| |
| ALIGN 16 |
| $L$ctr32_6x: |
| shl eax,4 |
| mov r10d,48 |
| bswap ebp |
| lea rcx,[32+rax*1+rcx] |
| sub r10,rax |
| jmp NEAR $L$ctr32_loop6 |
| |
| ALIGN 16 |
| $L$ctr32_loop6: |
| add r8d,6 |
| movups xmm0,XMMWORD[((-48))+r10*1+rcx] |
| DB 102,15,56,220,209 |
| mov eax,r8d |
| xor eax,ebp |
| DB 102,15,56,220,217 |
| DB 0x0f,0x38,0xf1,0x44,0x24,12 |
| lea eax,[1+r8] |
| DB 102,15,56,220,225 |
| xor eax,ebp |
| DB 0x0f,0x38,0xf1,0x44,0x24,28 |
| DB 102,15,56,220,233 |
| lea eax,[2+r8] |
| xor eax,ebp |
| DB 102,15,56,220,241 |
| DB 0x0f,0x38,0xf1,0x44,0x24,44 |
| lea eax,[3+r8] |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[((-32))+r10*1+rcx] |
| xor eax,ebp |
| |
| DB 102,15,56,220,208 |
| DB 0x0f,0x38,0xf1,0x44,0x24,60 |
| lea eax,[4+r8] |
| DB 102,15,56,220,216 |
| xor eax,ebp |
| DB 0x0f,0x38,0xf1,0x44,0x24,76 |
| DB 102,15,56,220,224 |
| lea eax,[5+r8] |
| xor eax,ebp |
| DB 102,15,56,220,232 |
| DB 0x0f,0x38,0xf1,0x44,0x24,92 |
| mov rax,r10 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| movups xmm0,XMMWORD[((-16))+r10*1+rcx] |
| |
| call $L$enc_loop6 |
| |
| movdqu xmm8,XMMWORD[rdi] |
| movdqu xmm9,XMMWORD[16+rdi] |
| movdqu xmm10,XMMWORD[32+rdi] |
| movdqu xmm11,XMMWORD[48+rdi] |
| movdqu xmm12,XMMWORD[64+rdi] |
| movdqu xmm13,XMMWORD[80+rdi] |
| lea rdi,[96+rdi] |
| movups xmm1,XMMWORD[((-64))+r10*1+rcx] |
| pxor xmm8,xmm2 |
| movaps xmm2,XMMWORD[rsp] |
| pxor xmm9,xmm3 |
| movaps xmm3,XMMWORD[16+rsp] |
| pxor xmm10,xmm4 |
| movaps xmm4,XMMWORD[32+rsp] |
| pxor xmm11,xmm5 |
| movaps xmm5,XMMWORD[48+rsp] |
| pxor xmm12,xmm6 |
| movaps xmm6,XMMWORD[64+rsp] |
| pxor xmm13,xmm7 |
| movaps xmm7,XMMWORD[80+rsp] |
| movdqu XMMWORD[rsi],xmm8 |
| movdqu XMMWORD[16+rsi],xmm9 |
| movdqu XMMWORD[32+rsi],xmm10 |
| movdqu XMMWORD[48+rsi],xmm11 |
| movdqu XMMWORD[64+rsi],xmm12 |
| movdqu XMMWORD[80+rsi],xmm13 |
| lea rsi,[96+rsi] |
| |
| sub rdx,6 |
| jnc NEAR $L$ctr32_loop6 |
| |
| add rdx,6 |
| jz NEAR $L$ctr32_done |
| |
| lea eax,[((-48))+r10] |
| lea rcx,[((-80))+r10*1+rcx] |
| neg eax |
| shr eax,4 |
| jmp NEAR $L$ctr32_tail |
| |
| ALIGN 32 |
| $L$ctr32_loop8: |
| add r8d,8 |
| movdqa xmm8,XMMWORD[96+rsp] |
| DB 102,15,56,220,209 |
| mov r9d,r8d |
| movdqa xmm9,XMMWORD[112+rsp] |
| DB 102,15,56,220,217 |
| bswap r9d |
| movups xmm0,XMMWORD[((32-128))+rcx] |
| DB 102,15,56,220,225 |
| xor r9d,ebp |
| nop |
| DB 102,15,56,220,233 |
| mov DWORD[((0+12))+rsp],r9d |
| lea r9,[1+r8] |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| movups xmm1,XMMWORD[((48-128))+rcx] |
| bswap r9d |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| xor r9d,ebp |
| DB 0x66,0x90 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| mov DWORD[((16+12))+rsp],r9d |
| lea r9,[2+r8] |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| DB 102,68,15,56,220,192 |
| DB 102,68,15,56,220,200 |
| movups xmm0,XMMWORD[((64-128))+rcx] |
| bswap r9d |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| xor r9d,ebp |
| DB 0x66,0x90 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| mov DWORD[((32+12))+rsp],r9d |
| lea r9,[3+r8] |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| movups xmm1,XMMWORD[((80-128))+rcx] |
| bswap r9d |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| xor r9d,ebp |
| DB 0x66,0x90 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| mov DWORD[((48+12))+rsp],r9d |
| lea r9,[4+r8] |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| DB 102,68,15,56,220,192 |
| DB 102,68,15,56,220,200 |
| movups xmm0,XMMWORD[((96-128))+rcx] |
| bswap r9d |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| xor r9d,ebp |
| DB 0x66,0x90 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| mov DWORD[((64+12))+rsp],r9d |
| lea r9,[5+r8] |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| movups xmm1,XMMWORD[((112-128))+rcx] |
| bswap r9d |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| xor r9d,ebp |
| DB 0x66,0x90 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| mov DWORD[((80+12))+rsp],r9d |
| lea r9,[6+r8] |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| DB 102,68,15,56,220,192 |
| DB 102,68,15,56,220,200 |
| movups xmm0,XMMWORD[((128-128))+rcx] |
| bswap r9d |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| xor r9d,ebp |
| DB 0x66,0x90 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| mov DWORD[((96+12))+rsp],r9d |
| lea r9,[7+r8] |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| movups xmm1,XMMWORD[((144-128))+rcx] |
| bswap r9d |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| xor r9d,ebp |
| movdqu xmm10,XMMWORD[rdi] |
| DB 102,15,56,220,232 |
| mov DWORD[((112+12))+rsp],r9d |
| cmp eax,11 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| DB 102,68,15,56,220,192 |
| DB 102,68,15,56,220,200 |
| movups xmm0,XMMWORD[((160-128))+rcx] |
| |
| jb NEAR $L$ctr32_enc_done |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| movups xmm1,XMMWORD[((176-128))+rcx] |
| |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| DB 102,68,15,56,220,192 |
| DB 102,68,15,56,220,200 |
| movups xmm0,XMMWORD[((192-128))+rcx] |
| je NEAR $L$ctr32_enc_done |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| movups xmm1,XMMWORD[((208-128))+rcx] |
| |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| DB 102,68,15,56,220,192 |
| DB 102,68,15,56,220,200 |
| movups xmm0,XMMWORD[((224-128))+rcx] |
| jmp NEAR $L$ctr32_enc_done |
| |
| ALIGN 16 |
| $L$ctr32_enc_done: |
| movdqu xmm11,XMMWORD[16+rdi] |
| pxor xmm10,xmm0 |
| movdqu xmm12,XMMWORD[32+rdi] |
| pxor xmm11,xmm0 |
| movdqu xmm13,XMMWORD[48+rdi] |
| pxor xmm12,xmm0 |
| movdqu xmm14,XMMWORD[64+rdi] |
| pxor xmm13,xmm0 |
| movdqu xmm15,XMMWORD[80+rdi] |
| pxor xmm14,xmm0 |
| pxor xmm15,xmm0 |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| DB 102,68,15,56,220,201 |
| movdqu xmm1,XMMWORD[96+rdi] |
| lea rdi,[128+rdi] |
| |
| DB 102,65,15,56,221,210 |
| pxor xmm1,xmm0 |
| movdqu xmm10,XMMWORD[((112-128))+rdi] |
| DB 102,65,15,56,221,219 |
| pxor xmm10,xmm0 |
| movdqa xmm11,XMMWORD[rsp] |
| DB 102,65,15,56,221,228 |
| DB 102,65,15,56,221,237 |
| movdqa xmm12,XMMWORD[16+rsp] |
| movdqa xmm13,XMMWORD[32+rsp] |
| DB 102,65,15,56,221,246 |
| DB 102,65,15,56,221,255 |
| movdqa xmm14,XMMWORD[48+rsp] |
| movdqa xmm15,XMMWORD[64+rsp] |
| DB 102,68,15,56,221,193 |
| movdqa xmm0,XMMWORD[80+rsp] |
| movups xmm1,XMMWORD[((16-128))+rcx] |
| DB 102,69,15,56,221,202 |
| |
| movups XMMWORD[rsi],xmm2 |
| movdqa xmm2,xmm11 |
| movups XMMWORD[16+rsi],xmm3 |
| movdqa xmm3,xmm12 |
| movups XMMWORD[32+rsi],xmm4 |
| movdqa xmm4,xmm13 |
| movups XMMWORD[48+rsi],xmm5 |
| movdqa xmm5,xmm14 |
| movups XMMWORD[64+rsi],xmm6 |
| movdqa xmm6,xmm15 |
| movups XMMWORD[80+rsi],xmm7 |
| movdqa xmm7,xmm0 |
| movups XMMWORD[96+rsi],xmm8 |
| movups XMMWORD[112+rsi],xmm9 |
| lea rsi,[128+rsi] |
| |
| sub rdx,8 |
| jnc NEAR $L$ctr32_loop8 |
| |
| add rdx,8 |
| jz NEAR $L$ctr32_done |
| lea rcx,[((-128))+rcx] |
| |
| $L$ctr32_tail: |
| |
| |
| lea rcx,[16+rcx] |
| cmp rdx,4 |
| jb NEAR $L$ctr32_loop3 |
| je NEAR $L$ctr32_loop4 |
| |
| |
| shl eax,4 |
| movdqa xmm8,XMMWORD[96+rsp] |
| pxor xmm9,xmm9 |
| |
| movups xmm0,XMMWORD[16+rcx] |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| lea rcx,[((32-16))+rax*1+rcx] |
| neg rax |
| DB 102,15,56,220,225 |
| add rax,16 |
| movups xmm10,XMMWORD[rdi] |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| movups xmm11,XMMWORD[16+rdi] |
| movups xmm12,XMMWORD[32+rdi] |
| DB 102,15,56,220,249 |
| DB 102,68,15,56,220,193 |
| |
| call $L$enc_loop8_enter |
| |
| movdqu xmm13,XMMWORD[48+rdi] |
| pxor xmm2,xmm10 |
| movdqu xmm10,XMMWORD[64+rdi] |
| pxor xmm3,xmm11 |
| movdqu XMMWORD[rsi],xmm2 |
| pxor xmm4,xmm12 |
| movdqu XMMWORD[16+rsi],xmm3 |
| pxor xmm5,xmm13 |
| movdqu XMMWORD[32+rsi],xmm4 |
| pxor xmm6,xmm10 |
| movdqu XMMWORD[48+rsi],xmm5 |
| movdqu XMMWORD[64+rsi],xmm6 |
| cmp rdx,6 |
| jb NEAR $L$ctr32_done |
| |
| movups xmm11,XMMWORD[80+rdi] |
| xorps xmm7,xmm11 |
| movups XMMWORD[80+rsi],xmm7 |
| je NEAR $L$ctr32_done |
| |
| movups xmm12,XMMWORD[96+rdi] |
| xorps xmm8,xmm12 |
| movups XMMWORD[96+rsi],xmm8 |
| jmp NEAR $L$ctr32_done |
| |
| ALIGN 32 |
| $L$ctr32_loop4: |
| DB 102,15,56,220,209 |
| lea rcx,[16+rcx] |
| dec eax |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| movups xmm1,XMMWORD[rcx] |
| jnz NEAR $L$ctr32_loop4 |
| DB 102,15,56,221,209 |
| DB 102,15,56,221,217 |
| movups xmm10,XMMWORD[rdi] |
| movups xmm11,XMMWORD[16+rdi] |
| DB 102,15,56,221,225 |
| DB 102,15,56,221,233 |
| movups xmm12,XMMWORD[32+rdi] |
| movups xmm13,XMMWORD[48+rdi] |
| |
| xorps xmm2,xmm10 |
| movups XMMWORD[rsi],xmm2 |
| xorps xmm3,xmm11 |
| movups XMMWORD[16+rsi],xmm3 |
| pxor xmm4,xmm12 |
| movdqu XMMWORD[32+rsi],xmm4 |
| pxor xmm5,xmm13 |
| movdqu XMMWORD[48+rsi],xmm5 |
| jmp NEAR $L$ctr32_done |
| |
| ALIGN 32 |
| $L$ctr32_loop3: |
| DB 102,15,56,220,209 |
| lea rcx,[16+rcx] |
| dec eax |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| movups xmm1,XMMWORD[rcx] |
| jnz NEAR $L$ctr32_loop3 |
| DB 102,15,56,221,209 |
| DB 102,15,56,221,217 |
| DB 102,15,56,221,225 |
| |
| movups xmm10,XMMWORD[rdi] |
| xorps xmm2,xmm10 |
| movups XMMWORD[rsi],xmm2 |
| cmp rdx,2 |
| jb NEAR $L$ctr32_done |
| |
| movups xmm11,XMMWORD[16+rdi] |
| xorps xmm3,xmm11 |
| movups XMMWORD[16+rsi],xmm3 |
| je NEAR $L$ctr32_done |
| |
| movups xmm12,XMMWORD[32+rdi] |
| xorps xmm4,xmm12 |
| movups XMMWORD[32+rsi],xmm4 |
| |
| $L$ctr32_done: |
| xorps xmm0,xmm0 |
| xor ebp,ebp |
| pxor xmm1,xmm1 |
| pxor xmm2,xmm2 |
| pxor xmm3,xmm3 |
| pxor xmm4,xmm4 |
| pxor xmm5,xmm5 |
| movaps xmm6,XMMWORD[((-168))+r11] |
| movaps XMMWORD[(-168)+r11],xmm0 |
| movaps xmm7,XMMWORD[((-152))+r11] |
| movaps XMMWORD[(-152)+r11],xmm0 |
| movaps xmm8,XMMWORD[((-136))+r11] |
| movaps XMMWORD[(-136)+r11],xmm0 |
| movaps xmm9,XMMWORD[((-120))+r11] |
| movaps XMMWORD[(-120)+r11],xmm0 |
| movaps xmm10,XMMWORD[((-104))+r11] |
| movaps XMMWORD[(-104)+r11],xmm0 |
| movaps xmm11,XMMWORD[((-88))+r11] |
| movaps XMMWORD[(-88)+r11],xmm0 |
| movaps xmm12,XMMWORD[((-72))+r11] |
| movaps XMMWORD[(-72)+r11],xmm0 |
| movaps xmm13,XMMWORD[((-56))+r11] |
| movaps XMMWORD[(-56)+r11],xmm0 |
| movaps xmm14,XMMWORD[((-40))+r11] |
| movaps XMMWORD[(-40)+r11],xmm0 |
| movaps xmm15,XMMWORD[((-24))+r11] |
| movaps XMMWORD[(-24)+r11],xmm0 |
| movaps XMMWORD[rsp],xmm0 |
| movaps XMMWORD[16+rsp],xmm0 |
| movaps XMMWORD[32+rsp],xmm0 |
| movaps XMMWORD[48+rsp],xmm0 |
| movaps XMMWORD[64+rsp],xmm0 |
| movaps XMMWORD[80+rsp],xmm0 |
| movaps XMMWORD[96+rsp],xmm0 |
| movaps XMMWORD[112+rsp],xmm0 |
| mov rbp,QWORD[((-8))+r11] |
| |
| lea rsp,[r11] |
| |
| $L$ctr32_epilogue: |
| mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| mov rsi,QWORD[16+rsp] |
| DB 0F3h,0C3h ;repret |
| |
| $L$SEH_end_aes_hw_ctr32_encrypt_blocks: |
| global aes_hw_xts_encrypt |
| |
| ALIGN 16 |
| aes_hw_xts_encrypt: |
| mov QWORD[8+rsp],rdi ;WIN64 prologue |
| mov QWORD[16+rsp],rsi |
| mov rax,rsp |
| $L$SEH_begin_aes_hw_xts_encrypt: |
| mov rdi,rcx |
| mov rsi,rdx |
| mov rdx,r8 |
| mov rcx,r9 |
| mov r8,QWORD[40+rsp] |
| mov r9,QWORD[48+rsp] |
| |
| |
| |
| lea r11,[rsp] |
| |
| push rbp |
| |
| sub rsp,272 |
| and rsp,-16 |
| movaps XMMWORD[(-168)+r11],xmm6 |
| movaps XMMWORD[(-152)+r11],xmm7 |
| movaps XMMWORD[(-136)+r11],xmm8 |
| movaps XMMWORD[(-120)+r11],xmm9 |
| movaps XMMWORD[(-104)+r11],xmm10 |
| movaps XMMWORD[(-88)+r11],xmm11 |
| movaps XMMWORD[(-72)+r11],xmm12 |
| movaps XMMWORD[(-56)+r11],xmm13 |
| movaps XMMWORD[(-40)+r11],xmm14 |
| movaps XMMWORD[(-24)+r11],xmm15 |
| $L$xts_enc_body: |
| movups xmm2,XMMWORD[r9] |
| mov eax,DWORD[240+r8] |
| mov r10d,DWORD[240+rcx] |
| movups xmm0,XMMWORD[r8] |
| movups xmm1,XMMWORD[16+r8] |
| lea r8,[32+r8] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_8: |
| DB 102,15,56,220,209 |
| dec eax |
| movups xmm1,XMMWORD[r8] |
| lea r8,[16+r8] |
| jnz NEAR $L$oop_enc1_8 |
| DB 102,15,56,221,209 |
| movups xmm0,XMMWORD[rcx] |
| mov rbp,rcx |
| mov eax,r10d |
| shl r10d,4 |
| mov r9,rdx |
| and rdx,-16 |
| |
| movups xmm1,XMMWORD[16+r10*1+rcx] |
| |
| movdqa xmm8,XMMWORD[$L$xts_magic] |
| movdqa xmm15,xmm2 |
| pshufd xmm9,xmm2,0x5f |
| pxor xmm1,xmm0 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm10,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm10,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm11,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm11,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm12,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm12,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm13,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm13,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm15 |
| psrad xmm9,31 |
| paddq xmm15,xmm15 |
| pand xmm9,xmm8 |
| pxor xmm14,xmm0 |
| pxor xmm15,xmm9 |
| movaps XMMWORD[96+rsp],xmm1 |
| |
| sub rdx,16*6 |
| jc NEAR $L$xts_enc_short |
| |
| mov eax,16+96 |
| lea rcx,[32+r10*1+rbp] |
| sub rax,r10 |
| movups xmm1,XMMWORD[16+rbp] |
| mov r10,rax |
| lea r8,[$L$xts_magic] |
| jmp NEAR $L$xts_enc_grandloop |
| |
| ALIGN 32 |
| $L$xts_enc_grandloop: |
| movdqu xmm2,XMMWORD[rdi] |
| movdqa xmm8,xmm0 |
| movdqu xmm3,XMMWORD[16+rdi] |
| pxor xmm2,xmm10 |
| movdqu xmm4,XMMWORD[32+rdi] |
| pxor xmm3,xmm11 |
| DB 102,15,56,220,209 |
| movdqu xmm5,XMMWORD[48+rdi] |
| pxor xmm4,xmm12 |
| DB 102,15,56,220,217 |
| movdqu xmm6,XMMWORD[64+rdi] |
| pxor xmm5,xmm13 |
| DB 102,15,56,220,225 |
| movdqu xmm7,XMMWORD[80+rdi] |
| pxor xmm8,xmm15 |
| movdqa xmm9,XMMWORD[96+rsp] |
| pxor xmm6,xmm14 |
| DB 102,15,56,220,233 |
| movups xmm0,XMMWORD[32+rbp] |
| lea rdi,[96+rdi] |
| pxor xmm7,xmm8 |
| |
| pxor xmm10,xmm9 |
| DB 102,15,56,220,241 |
| pxor xmm11,xmm9 |
| movdqa XMMWORD[rsp],xmm10 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[48+rbp] |
| pxor xmm12,xmm9 |
| |
| DB 102,15,56,220,208 |
| pxor xmm13,xmm9 |
| movdqa XMMWORD[16+rsp],xmm11 |
| DB 102,15,56,220,216 |
| pxor xmm14,xmm9 |
| movdqa XMMWORD[32+rsp],xmm12 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| pxor xmm8,xmm9 |
| movdqa XMMWORD[64+rsp],xmm14 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| movups xmm0,XMMWORD[64+rbp] |
| movdqa XMMWORD[80+rsp],xmm8 |
| pshufd xmm9,xmm15,0x5f |
| jmp NEAR $L$xts_enc_loop6 |
| ALIGN 32 |
| $L$xts_enc_loop6: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[((-64))+rax*1+rcx] |
| add rax,32 |
| |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| movups xmm0,XMMWORD[((-80))+rax*1+rcx] |
| jnz NEAR $L$xts_enc_loop6 |
| |
| movdqa xmm8,XMMWORD[r8] |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| DB 102,15,56,220,209 |
| paddq xmm15,xmm15 |
| psrad xmm14,31 |
| DB 102,15,56,220,217 |
| pand xmm14,xmm8 |
| movups xmm10,XMMWORD[rbp] |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| pxor xmm15,xmm14 |
| movaps xmm11,xmm10 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[((-64))+rcx] |
| |
| movdqa xmm14,xmm9 |
| DB 102,15,56,220,208 |
| paddd xmm9,xmm9 |
| pxor xmm10,xmm15 |
| DB 102,15,56,220,216 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| pand xmm14,xmm8 |
| movaps xmm12,xmm11 |
| DB 102,15,56,220,240 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| DB 102,15,56,220,248 |
| movups xmm0,XMMWORD[((-48))+rcx] |
| |
| paddd xmm9,xmm9 |
| DB 102,15,56,220,209 |
| pxor xmm11,xmm15 |
| psrad xmm14,31 |
| DB 102,15,56,220,217 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| movdqa XMMWORD[48+rsp],xmm13 |
| pxor xmm15,xmm14 |
| DB 102,15,56,220,241 |
| movaps xmm13,xmm12 |
| movdqa xmm14,xmm9 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[((-32))+rcx] |
| |
| paddd xmm9,xmm9 |
| DB 102,15,56,220,208 |
| pxor xmm12,xmm15 |
| psrad xmm14,31 |
| DB 102,15,56,220,216 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| pxor xmm15,xmm14 |
| movaps xmm14,xmm13 |
| DB 102,15,56,220,248 |
| |
| movdqa xmm0,xmm9 |
| paddd xmm9,xmm9 |
| DB 102,15,56,220,209 |
| pxor xmm13,xmm15 |
| psrad xmm0,31 |
| DB 102,15,56,220,217 |
| paddq xmm15,xmm15 |
| pand xmm0,xmm8 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| pxor xmm15,xmm0 |
| movups xmm0,XMMWORD[rbp] |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[16+rbp] |
| |
| pxor xmm14,xmm15 |
| DB 102,15,56,221,84,36,0 |
| psrad xmm9,31 |
| paddq xmm15,xmm15 |
| DB 102,15,56,221,92,36,16 |
| DB 102,15,56,221,100,36,32 |
| pand xmm9,xmm8 |
| mov rax,r10 |
| DB 102,15,56,221,108,36,48 |
| DB 102,15,56,221,116,36,64 |
| DB 102,15,56,221,124,36,80 |
| pxor xmm15,xmm9 |
| |
| lea rsi,[96+rsi] |
| movups XMMWORD[(-96)+rsi],xmm2 |
| movups XMMWORD[(-80)+rsi],xmm3 |
| movups XMMWORD[(-64)+rsi],xmm4 |
| movups XMMWORD[(-48)+rsi],xmm5 |
| movups XMMWORD[(-32)+rsi],xmm6 |
| movups XMMWORD[(-16)+rsi],xmm7 |
| sub rdx,16*6 |
| jnc NEAR $L$xts_enc_grandloop |
| |
| mov eax,16+96 |
| sub eax,r10d |
| mov rcx,rbp |
| shr eax,4 |
| |
| $L$xts_enc_short: |
| |
| mov r10d,eax |
| pxor xmm10,xmm0 |
| add rdx,16*6 |
| jz NEAR $L$xts_enc_done |
| |
| pxor xmm11,xmm0 |
| cmp rdx,0x20 |
| jb NEAR $L$xts_enc_one |
| pxor xmm12,xmm0 |
| je NEAR $L$xts_enc_two |
| |
| pxor xmm13,xmm0 |
| cmp rdx,0x40 |
| jb NEAR $L$xts_enc_three |
| pxor xmm14,xmm0 |
| je NEAR $L$xts_enc_four |
| |
| movdqu xmm2,XMMWORD[rdi] |
| movdqu xmm3,XMMWORD[16+rdi] |
| movdqu xmm4,XMMWORD[32+rdi] |
| pxor xmm2,xmm10 |
| movdqu xmm5,XMMWORD[48+rdi] |
| pxor xmm3,xmm11 |
| movdqu xmm6,XMMWORD[64+rdi] |
| lea rdi,[80+rdi] |
| pxor xmm4,xmm12 |
| pxor xmm5,xmm13 |
| pxor xmm6,xmm14 |
| pxor xmm7,xmm7 |
| |
| call _aesni_encrypt6 |
| |
| xorps xmm2,xmm10 |
| movdqa xmm10,xmm15 |
| xorps xmm3,xmm11 |
| xorps xmm4,xmm12 |
| movdqu XMMWORD[rsi],xmm2 |
| xorps xmm5,xmm13 |
| movdqu XMMWORD[16+rsi],xmm3 |
| xorps xmm6,xmm14 |
| movdqu XMMWORD[32+rsi],xmm4 |
| movdqu XMMWORD[48+rsi],xmm5 |
| movdqu XMMWORD[64+rsi],xmm6 |
| lea rsi,[80+rsi] |
| jmp NEAR $L$xts_enc_done |
| |
| ALIGN 16 |
| $L$xts_enc_one: |
| movups xmm2,XMMWORD[rdi] |
| lea rdi,[16+rdi] |
| xorps xmm2,xmm10 |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_9: |
| DB 102,15,56,220,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_enc1_9 |
| DB 102,15,56,221,209 |
| xorps xmm2,xmm10 |
| movdqa xmm10,xmm11 |
| movups XMMWORD[rsi],xmm2 |
| lea rsi,[16+rsi] |
| jmp NEAR $L$xts_enc_done |
| |
| ALIGN 16 |
| $L$xts_enc_two: |
| movups xmm2,XMMWORD[rdi] |
| movups xmm3,XMMWORD[16+rdi] |
| lea rdi,[32+rdi] |
| xorps xmm2,xmm10 |
| xorps xmm3,xmm11 |
| |
| call _aesni_encrypt2 |
| |
| xorps xmm2,xmm10 |
| movdqa xmm10,xmm12 |
| xorps xmm3,xmm11 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| lea rsi,[32+rsi] |
| jmp NEAR $L$xts_enc_done |
| |
| ALIGN 16 |
| $L$xts_enc_three: |
| movups xmm2,XMMWORD[rdi] |
| movups xmm3,XMMWORD[16+rdi] |
| movups xmm4,XMMWORD[32+rdi] |
| lea rdi,[48+rdi] |
| xorps xmm2,xmm10 |
| xorps xmm3,xmm11 |
| xorps xmm4,xmm12 |
| |
| call _aesni_encrypt3 |
| |
| xorps xmm2,xmm10 |
| movdqa xmm10,xmm13 |
| xorps xmm3,xmm11 |
| xorps xmm4,xmm12 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| lea rsi,[48+rsi] |
| jmp NEAR $L$xts_enc_done |
| |
| ALIGN 16 |
| $L$xts_enc_four: |
| movups xmm2,XMMWORD[rdi] |
| movups xmm3,XMMWORD[16+rdi] |
| movups xmm4,XMMWORD[32+rdi] |
| xorps xmm2,xmm10 |
| movups xmm5,XMMWORD[48+rdi] |
| lea rdi,[64+rdi] |
| xorps xmm3,xmm11 |
| xorps xmm4,xmm12 |
| xorps xmm5,xmm13 |
| |
| call _aesni_encrypt4 |
| |
| pxor xmm2,xmm10 |
| movdqa xmm10,xmm14 |
| pxor xmm3,xmm11 |
| pxor xmm4,xmm12 |
| movdqu XMMWORD[rsi],xmm2 |
| pxor xmm5,xmm13 |
| movdqu XMMWORD[16+rsi],xmm3 |
| movdqu XMMWORD[32+rsi],xmm4 |
| movdqu XMMWORD[48+rsi],xmm5 |
| lea rsi,[64+rsi] |
| jmp NEAR $L$xts_enc_done |
| |
| ALIGN 16 |
| $L$xts_enc_done: |
| and r9,15 |
| jz NEAR $L$xts_enc_ret |
| mov rdx,r9 |
| |
| $L$xts_enc_steal: |
| movzx eax,BYTE[rdi] |
| movzx ecx,BYTE[((-16))+rsi] |
| lea rdi,[1+rdi] |
| mov BYTE[((-16))+rsi],al |
| mov BYTE[rsi],cl |
| lea rsi,[1+rsi] |
| sub rdx,1 |
| jnz NEAR $L$xts_enc_steal |
| |
| sub rsi,r9 |
| mov rcx,rbp |
| mov eax,r10d |
| |
| movups xmm2,XMMWORD[((-16))+rsi] |
| xorps xmm2,xmm10 |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_10: |
| DB 102,15,56,220,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_enc1_10 |
| DB 102,15,56,221,209 |
| xorps xmm2,xmm10 |
| movups XMMWORD[(-16)+rsi],xmm2 |
| |
| $L$xts_enc_ret: |
| xorps xmm0,xmm0 |
| pxor xmm1,xmm1 |
| pxor xmm2,xmm2 |
| pxor xmm3,xmm3 |
| pxor xmm4,xmm4 |
| pxor xmm5,xmm5 |
| movaps xmm6,XMMWORD[((-168))+r11] |
| movaps XMMWORD[(-168)+r11],xmm0 |
| movaps xmm7,XMMWORD[((-152))+r11] |
| movaps XMMWORD[(-152)+r11],xmm0 |
| movaps xmm8,XMMWORD[((-136))+r11] |
| movaps XMMWORD[(-136)+r11],xmm0 |
| movaps xmm9,XMMWORD[((-120))+r11] |
| movaps XMMWORD[(-120)+r11],xmm0 |
| movaps xmm10,XMMWORD[((-104))+r11] |
| movaps XMMWORD[(-104)+r11],xmm0 |
| movaps xmm11,XMMWORD[((-88))+r11] |
| movaps XMMWORD[(-88)+r11],xmm0 |
| movaps xmm12,XMMWORD[((-72))+r11] |
| movaps XMMWORD[(-72)+r11],xmm0 |
| movaps xmm13,XMMWORD[((-56))+r11] |
| movaps XMMWORD[(-56)+r11],xmm0 |
| movaps xmm14,XMMWORD[((-40))+r11] |
| movaps XMMWORD[(-40)+r11],xmm0 |
| movaps xmm15,XMMWORD[((-24))+r11] |
| movaps XMMWORD[(-24)+r11],xmm0 |
| movaps XMMWORD[rsp],xmm0 |
| movaps XMMWORD[16+rsp],xmm0 |
| movaps XMMWORD[32+rsp],xmm0 |
| movaps XMMWORD[48+rsp],xmm0 |
| movaps XMMWORD[64+rsp],xmm0 |
| movaps XMMWORD[80+rsp],xmm0 |
| movaps XMMWORD[96+rsp],xmm0 |
| mov rbp,QWORD[((-8))+r11] |
| |
| lea rsp,[r11] |
| |
| $L$xts_enc_epilogue: |
| mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| mov rsi,QWORD[16+rsp] |
| DB 0F3h,0C3h ;repret |
| |
| $L$SEH_end_aes_hw_xts_encrypt: |
| global aes_hw_xts_decrypt |
| |
| ALIGN 16 |
| aes_hw_xts_decrypt: |
| mov QWORD[8+rsp],rdi ;WIN64 prologue |
| mov QWORD[16+rsp],rsi |
| mov rax,rsp |
| $L$SEH_begin_aes_hw_xts_decrypt: |
| mov rdi,rcx |
| mov rsi,rdx |
| mov rdx,r8 |
| mov rcx,r9 |
| mov r8,QWORD[40+rsp] |
| mov r9,QWORD[48+rsp] |
| |
| |
| |
| lea r11,[rsp] |
| |
| push rbp |
| |
| sub rsp,272 |
| and rsp,-16 |
| movaps XMMWORD[(-168)+r11],xmm6 |
| movaps XMMWORD[(-152)+r11],xmm7 |
| movaps XMMWORD[(-136)+r11],xmm8 |
| movaps XMMWORD[(-120)+r11],xmm9 |
| movaps XMMWORD[(-104)+r11],xmm10 |
| movaps XMMWORD[(-88)+r11],xmm11 |
| movaps XMMWORD[(-72)+r11],xmm12 |
| movaps XMMWORD[(-56)+r11],xmm13 |
| movaps XMMWORD[(-40)+r11],xmm14 |
| movaps XMMWORD[(-24)+r11],xmm15 |
| $L$xts_dec_body: |
| movups xmm2,XMMWORD[r9] |
| mov eax,DWORD[240+r8] |
| mov r10d,DWORD[240+rcx] |
| movups xmm0,XMMWORD[r8] |
| movups xmm1,XMMWORD[16+r8] |
| lea r8,[32+r8] |
| xorps xmm2,xmm0 |
| $L$oop_enc1_11: |
| DB 102,15,56,220,209 |
| dec eax |
| movups xmm1,XMMWORD[r8] |
| lea r8,[16+r8] |
| jnz NEAR $L$oop_enc1_11 |
| DB 102,15,56,221,209 |
| xor eax,eax |
| test rdx,15 |
| setnz al |
| shl rax,4 |
| sub rdx,rax |
| |
| movups xmm0,XMMWORD[rcx] |
| mov rbp,rcx |
| mov eax,r10d |
| shl r10d,4 |
| mov r9,rdx |
| and rdx,-16 |
| |
| movups xmm1,XMMWORD[16+r10*1+rcx] |
| |
| movdqa xmm8,XMMWORD[$L$xts_magic] |
| movdqa xmm15,xmm2 |
| pshufd xmm9,xmm2,0x5f |
| pxor xmm1,xmm0 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm10,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm10,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm11,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm11,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm12,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm12,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| movdqa xmm13,xmm15 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| pxor xmm13,xmm0 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm15 |
| psrad xmm9,31 |
| paddq xmm15,xmm15 |
| pand xmm9,xmm8 |
| pxor xmm14,xmm0 |
| pxor xmm15,xmm9 |
| movaps XMMWORD[96+rsp],xmm1 |
| |
| sub rdx,16*6 |
| jc NEAR $L$xts_dec_short |
| |
| mov eax,16+96 |
| lea rcx,[32+r10*1+rbp] |
| sub rax,r10 |
| movups xmm1,XMMWORD[16+rbp] |
| mov r10,rax |
| lea r8,[$L$xts_magic] |
| jmp NEAR $L$xts_dec_grandloop |
| |
| ALIGN 32 |
| $L$xts_dec_grandloop: |
| movdqu xmm2,XMMWORD[rdi] |
| movdqa xmm8,xmm0 |
| movdqu xmm3,XMMWORD[16+rdi] |
| pxor xmm2,xmm10 |
| movdqu xmm4,XMMWORD[32+rdi] |
| pxor xmm3,xmm11 |
| DB 102,15,56,222,209 |
| movdqu xmm5,XMMWORD[48+rdi] |
| pxor xmm4,xmm12 |
| DB 102,15,56,222,217 |
| movdqu xmm6,XMMWORD[64+rdi] |
| pxor xmm5,xmm13 |
| DB 102,15,56,222,225 |
| movdqu xmm7,XMMWORD[80+rdi] |
| pxor xmm8,xmm15 |
| movdqa xmm9,XMMWORD[96+rsp] |
| pxor xmm6,xmm14 |
| DB 102,15,56,222,233 |
| movups xmm0,XMMWORD[32+rbp] |
| lea rdi,[96+rdi] |
| pxor xmm7,xmm8 |
| |
| pxor xmm10,xmm9 |
| DB 102,15,56,222,241 |
| pxor xmm11,xmm9 |
| movdqa XMMWORD[rsp],xmm10 |
| DB 102,15,56,222,249 |
| movups xmm1,XMMWORD[48+rbp] |
| pxor xmm12,xmm9 |
| |
| DB 102,15,56,222,208 |
| pxor xmm13,xmm9 |
| movdqa XMMWORD[16+rsp],xmm11 |
| DB 102,15,56,222,216 |
| pxor xmm14,xmm9 |
| movdqa XMMWORD[32+rsp],xmm12 |
| DB 102,15,56,222,224 |
| DB 102,15,56,222,232 |
| pxor xmm8,xmm9 |
| movdqa XMMWORD[64+rsp],xmm14 |
| DB 102,15,56,222,240 |
| DB 102,15,56,222,248 |
| movups xmm0,XMMWORD[64+rbp] |
| movdqa XMMWORD[80+rsp],xmm8 |
| pshufd xmm9,xmm15,0x5f |
| jmp NEAR $L$xts_dec_loop6 |
| ALIGN 32 |
| $L$xts_dec_loop6: |
| DB 102,15,56,222,209 |
| DB 102,15,56,222,217 |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| DB 102,15,56,222,241 |
| DB 102,15,56,222,249 |
| movups xmm1,XMMWORD[((-64))+rax*1+rcx] |
| add rax,32 |
| |
| DB 102,15,56,222,208 |
| DB 102,15,56,222,216 |
| DB 102,15,56,222,224 |
| DB 102,15,56,222,232 |
| DB 102,15,56,222,240 |
| DB 102,15,56,222,248 |
| movups xmm0,XMMWORD[((-80))+rax*1+rcx] |
| jnz NEAR $L$xts_dec_loop6 |
| |
| movdqa xmm8,XMMWORD[r8] |
| movdqa xmm14,xmm9 |
| paddd xmm9,xmm9 |
| DB 102,15,56,222,209 |
| paddq xmm15,xmm15 |
| psrad xmm14,31 |
| DB 102,15,56,222,217 |
| pand xmm14,xmm8 |
| movups xmm10,XMMWORD[rbp] |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| DB 102,15,56,222,241 |
| pxor xmm15,xmm14 |
| movaps xmm11,xmm10 |
| DB 102,15,56,222,249 |
| movups xmm1,XMMWORD[((-64))+rcx] |
| |
| movdqa xmm14,xmm9 |
| DB 102,15,56,222,208 |
| paddd xmm9,xmm9 |
| pxor xmm10,xmm15 |
| DB 102,15,56,222,216 |
| psrad xmm14,31 |
| paddq xmm15,xmm15 |
| DB 102,15,56,222,224 |
| DB 102,15,56,222,232 |
| pand xmm14,xmm8 |
| movaps xmm12,xmm11 |
| DB 102,15,56,222,240 |
| pxor xmm15,xmm14 |
| movdqa xmm14,xmm9 |
| DB 102,15,56,222,248 |
| movups xmm0,XMMWORD[((-48))+rcx] |
| |
| paddd xmm9,xmm9 |
| DB 102,15,56,222,209 |
| pxor xmm11,xmm15 |
| psrad xmm14,31 |
| DB 102,15,56,222,217 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| movdqa XMMWORD[48+rsp],xmm13 |
| pxor xmm15,xmm14 |
| DB 102,15,56,222,241 |
| movaps xmm13,xmm12 |
| movdqa xmm14,xmm9 |
| DB 102,15,56,222,249 |
| movups xmm1,XMMWORD[((-32))+rcx] |
| |
| paddd xmm9,xmm9 |
| DB 102,15,56,222,208 |
| pxor xmm12,xmm15 |
| psrad xmm14,31 |
| DB 102,15,56,222,216 |
| paddq xmm15,xmm15 |
| pand xmm14,xmm8 |
| DB 102,15,56,222,224 |
| DB 102,15,56,222,232 |
| DB 102,15,56,222,240 |
| pxor xmm15,xmm14 |
| movaps xmm14,xmm13 |
| DB 102,15,56,222,248 |
| |
| movdqa xmm0,xmm9 |
| paddd xmm9,xmm9 |
| DB 102,15,56,222,209 |
| pxor xmm13,xmm15 |
| psrad xmm0,31 |
| DB 102,15,56,222,217 |
| paddq xmm15,xmm15 |
| pand xmm0,xmm8 |
| DB 102,15,56,222,225 |
| DB 102,15,56,222,233 |
| pxor xmm15,xmm0 |
| movups xmm0,XMMWORD[rbp] |
| DB 102,15,56,222,241 |
| DB 102,15,56,222,249 |
| movups xmm1,XMMWORD[16+rbp] |
| |
| pxor xmm14,xmm15 |
| DB 102,15,56,223,84,36,0 |
| psrad xmm9,31 |
| paddq xmm15,xmm15 |
| DB 102,15,56,223,92,36,16 |
| DB 102,15,56,223,100,36,32 |
| pand xmm9,xmm8 |
| mov rax,r10 |
| DB 102,15,56,223,108,36,48 |
| DB 102,15,56,223,116,36,64 |
| DB 102,15,56,223,124,36,80 |
| pxor xmm15,xmm9 |
| |
| lea rsi,[96+rsi] |
| movups XMMWORD[(-96)+rsi],xmm2 |
| movups XMMWORD[(-80)+rsi],xmm3 |
| movups XMMWORD[(-64)+rsi],xmm4 |
| movups XMMWORD[(-48)+rsi],xmm5 |
| movups XMMWORD[(-32)+rsi],xmm6 |
| movups XMMWORD[(-16)+rsi],xmm7 |
| sub rdx,16*6 |
| jnc NEAR $L$xts_dec_grandloop |
| |
| mov eax,16+96 |
| sub eax,r10d |
| mov rcx,rbp |
| shr eax,4 |
| |
| $L$xts_dec_short: |
| |
| mov r10d,eax |
| pxor xmm10,xmm0 |
| pxor xmm11,xmm0 |
| add rdx,16*6 |
| jz NEAR $L$xts_dec_done |
| |
| pxor xmm12,xmm0 |
| cmp rdx,0x20 |
| jb NEAR $L$xts_dec_one |
| pxor xmm13,xmm0 |
| je NEAR $L$xts_dec_two |
| |
| pxor xmm14,xmm0 |
| cmp rdx,0x40 |
| jb NEAR $L$xts_dec_three |
| je NEAR $L$xts_dec_four |
| |
| movdqu xmm2,XMMWORD[rdi] |
| movdqu xmm3,XMMWORD[16+rdi] |
| movdqu xmm4,XMMWORD[32+rdi] |
| pxor xmm2,xmm10 |
| movdqu xmm5,XMMWORD[48+rdi] |
| pxor xmm3,xmm11 |
| movdqu xmm6,XMMWORD[64+rdi] |
| lea rdi,[80+rdi] |
| pxor xmm4,xmm12 |
| pxor xmm5,xmm13 |
| pxor xmm6,xmm14 |
| |
| call _aesni_decrypt6 |
| |
| xorps xmm2,xmm10 |
| xorps xmm3,xmm11 |
| xorps xmm4,xmm12 |
| movdqu XMMWORD[rsi],xmm2 |
| xorps xmm5,xmm13 |
| movdqu XMMWORD[16+rsi],xmm3 |
| xorps xmm6,xmm14 |
| movdqu XMMWORD[32+rsi],xmm4 |
| pxor xmm14,xmm14 |
| movdqu XMMWORD[48+rsi],xmm5 |
| pcmpgtd xmm14,xmm15 |
| movdqu XMMWORD[64+rsi],xmm6 |
| lea rsi,[80+rsi] |
| pshufd xmm11,xmm14,0x13 |
| and r9,15 |
| jz NEAR $L$xts_dec_ret |
| |
| movdqa xmm10,xmm15 |
| paddq xmm15,xmm15 |
| pand xmm11,xmm8 |
| pxor xmm11,xmm15 |
| jmp NEAR $L$xts_dec_done2 |
| |
| ALIGN 16 |
| $L$xts_dec_one: |
| movups xmm2,XMMWORD[rdi] |
| lea rdi,[16+rdi] |
| xorps xmm2,xmm10 |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_dec1_12: |
| DB 102,15,56,222,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_dec1_12 |
| DB 102,15,56,223,209 |
| xorps xmm2,xmm10 |
| movdqa xmm10,xmm11 |
| movups XMMWORD[rsi],xmm2 |
| movdqa xmm11,xmm12 |
| lea rsi,[16+rsi] |
| jmp NEAR $L$xts_dec_done |
| |
| ALIGN 16 |
| $L$xts_dec_two: |
| movups xmm2,XMMWORD[rdi] |
| movups xmm3,XMMWORD[16+rdi] |
| lea rdi,[32+rdi] |
| xorps xmm2,xmm10 |
| xorps xmm3,xmm11 |
| |
| call _aesni_decrypt2 |
| |
| xorps xmm2,xmm10 |
| movdqa xmm10,xmm12 |
| xorps xmm3,xmm11 |
| movdqa xmm11,xmm13 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| lea rsi,[32+rsi] |
| jmp NEAR $L$xts_dec_done |
| |
| ALIGN 16 |
| $L$xts_dec_three: |
| movups xmm2,XMMWORD[rdi] |
| movups xmm3,XMMWORD[16+rdi] |
| movups xmm4,XMMWORD[32+rdi] |
| lea rdi,[48+rdi] |
| xorps xmm2,xmm10 |
| xorps xmm3,xmm11 |
| xorps xmm4,xmm12 |
| |
| call _aesni_decrypt3 |
| |
| xorps xmm2,xmm10 |
| movdqa xmm10,xmm13 |
| xorps xmm3,xmm11 |
| movdqa xmm11,xmm14 |
| xorps xmm4,xmm12 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| lea rsi,[48+rsi] |
| jmp NEAR $L$xts_dec_done |
| |
| ALIGN 16 |
| $L$xts_dec_four: |
| movups xmm2,XMMWORD[rdi] |
| movups xmm3,XMMWORD[16+rdi] |
| movups xmm4,XMMWORD[32+rdi] |
| xorps xmm2,xmm10 |
| movups xmm5,XMMWORD[48+rdi] |
| lea rdi,[64+rdi] |
| xorps xmm3,xmm11 |
| xorps xmm4,xmm12 |
| xorps xmm5,xmm13 |
| |
| call _aesni_decrypt4 |
| |
| pxor xmm2,xmm10 |
| movdqa xmm10,xmm14 |
| pxor xmm3,xmm11 |
| movdqa xmm11,xmm15 |
| pxor xmm4,xmm12 |
| movdqu XMMWORD[rsi],xmm2 |
| pxor xmm5,xmm13 |
| movdqu XMMWORD[16+rsi],xmm3 |
| movdqu XMMWORD[32+rsi],xmm4 |
| movdqu XMMWORD[48+rsi],xmm5 |
| lea rsi,[64+rsi] |
| jmp NEAR $L$xts_dec_done |
| |
| ALIGN 16 |
| $L$xts_dec_done: |
| and r9,15 |
| jz NEAR $L$xts_dec_ret |
| $L$xts_dec_done2: |
| mov rdx,r9 |
| mov rcx,rbp |
| mov eax,r10d |
| |
| movups xmm2,XMMWORD[rdi] |
| xorps xmm2,xmm11 |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_dec1_13: |
| DB 102,15,56,222,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_dec1_13 |
| DB 102,15,56,223,209 |
| xorps xmm2,xmm11 |
| movups XMMWORD[rsi],xmm2 |
| |
| $L$xts_dec_steal: |
| movzx eax,BYTE[16+rdi] |
| movzx ecx,BYTE[rsi] |
| lea rdi,[1+rdi] |
| mov BYTE[rsi],al |
| mov BYTE[16+rsi],cl |
| lea rsi,[1+rsi] |
| sub rdx,1 |
| jnz NEAR $L$xts_dec_steal |
| |
| sub rsi,r9 |
| mov rcx,rbp |
| mov eax,r10d |
| |
| movups xmm2,XMMWORD[rsi] |
| xorps xmm2,xmm10 |
| movups xmm0,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+rcx] |
| lea rcx,[32+rcx] |
| xorps xmm2,xmm0 |
| $L$oop_dec1_14: |
| DB 102,15,56,222,209 |
| dec eax |
| movups xmm1,XMMWORD[rcx] |
| lea rcx,[16+rcx] |
| jnz NEAR $L$oop_dec1_14 |
| DB 102,15,56,223,209 |
| xorps xmm2,xmm10 |
| movups XMMWORD[rsi],xmm2 |
| |
| $L$xts_dec_ret: |
| xorps xmm0,xmm0 |
| pxor xmm1,xmm1 |
| pxor xmm2,xmm2 |
| pxor xmm3,xmm3 |
| pxor xmm4,xmm4 |
| pxor xmm5,xmm5 |
| movaps xmm6,XMMWORD[((-168))+r11] |
| movaps XMMWORD[(-168)+r11],xmm0 |
| movaps xmm7,XMMWORD[((-152))+r11] |
| movaps XMMWORD[(-152)+r11],xmm0 |
| movaps xmm8,XMMWORD[((-136))+r11] |
| movaps XMMWORD[(-136)+r11],xmm0 |
| movaps xmm9,XMMWORD[((-120))+r11] |
| movaps XMMWORD[(-120)+r11],xmm0 |
| movaps xmm10,XMMWORD[((-104))+r11] |
| movaps XMMWORD[(-104)+r11],xmm0 |
| movaps xmm11,XMMWORD[((-88))+r11] |
| movaps XMMWORD[(-88)+r11],xmm0 |
| movaps xmm12,XMMWORD[((-72))+r11] |
| movaps XMMWORD[(-72)+r11],xmm0 |
| movaps xmm13,XMMWORD[((-56))+r11] |
| movaps XMMWORD[(-56)+r11],xmm0 |
| movaps xmm14,XMMWORD[((-40))+r11] |
| movaps XMMWORD[(-40)+r11],xmm0 |
| movaps xmm15,XMMWORD[((-24))+r11] |
| movaps XMMWORD[(-24)+r11],xmm0 |
| movaps XMMWORD[rsp],xmm0 |
| movaps XMMWORD[16+rsp],xmm0 |
| movaps XMMWORD[32+rsp],xmm0 |
| movaps XMMWORD[48+rsp],xmm0 |
| movaps XMMWORD[64+rsp],xmm0 |
| movaps XMMWORD[80+rsp],xmm0 |
| movaps XMMWORD[96+rsp],xmm0 |
| mov rbp,QWORD[((-8))+r11] |
| |
| lea rsp,[r11] |
| |
| $L$xts_dec_epilogue: |
| mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| mov rsi,QWORD[16+rsp] |
| DB 0F3h,0C3h ;repret |
| |
| $L$SEH_end_aes_hw_xts_decrypt: |
| global aes_hw_ocb_encrypt |
| |
| ALIGN 32 |
| aes_hw_ocb_encrypt: |
| mov QWORD[8+rsp],rdi ;WIN64 prologue |
| mov QWORD[16+rsp],rsi |
| mov rax,rsp |
| $L$SEH_begin_aes_hw_ocb_encrypt: |
| mov rdi,rcx |
| mov rsi,rdx |
| mov rdx,r8 |
| mov rcx,r9 |
| mov r8,QWORD[40+rsp] |
| mov r9,QWORD[48+rsp] |
| |
| |
| |
| lea rax,[rsp] |
| push rbx |
| |
| push rbp |
| |
| push r12 |
| |
| push r13 |
| |
| push r14 |
| |
| lea rsp,[((-160))+rsp] |
| movaps XMMWORD[rsp],xmm6 |
| movaps XMMWORD[16+rsp],xmm7 |
| movaps XMMWORD[32+rsp],xmm8 |
| movaps XMMWORD[48+rsp],xmm9 |
| movaps XMMWORD[64+rsp],xmm10 |
| movaps XMMWORD[80+rsp],xmm11 |
| movaps XMMWORD[96+rsp],xmm12 |
| movaps XMMWORD[112+rsp],xmm13 |
| movaps XMMWORD[128+rsp],xmm14 |
| movaps XMMWORD[144+rsp],xmm15 |
| $L$ocb_enc_body: |
| mov rbx,QWORD[56+rax] |
| mov rbp,QWORD[((56+8))+rax] |
| |
| mov r10d,DWORD[240+rcx] |
| mov r11,rcx |
| shl r10d,4 |
| movups xmm9,XMMWORD[rcx] |
| movups xmm1,XMMWORD[16+r10*1+rcx] |
| |
| movdqu xmm15,XMMWORD[r9] |
| pxor xmm9,xmm1 |
| pxor xmm15,xmm1 |
| |
| mov eax,16+32 |
| lea rcx,[32+r10*1+r11] |
| movups xmm1,XMMWORD[16+r11] |
| sub rax,r10 |
| mov r10,rax |
| |
| movdqu xmm10,XMMWORD[rbx] |
| movdqu xmm8,XMMWORD[rbp] |
| |
| test r8,1 |
| jnz NEAR $L$ocb_enc_odd |
| |
| bsf r12,r8 |
| add r8,1 |
| shl r12,4 |
| movdqu xmm7,XMMWORD[r12*1+rbx] |
| movdqu xmm2,XMMWORD[rdi] |
| lea rdi,[16+rdi] |
| |
| call __ocb_encrypt1 |
| |
| movdqa xmm15,xmm7 |
| movups XMMWORD[rsi],xmm2 |
| lea rsi,[16+rsi] |
| sub rdx,1 |
| jz NEAR $L$ocb_enc_done |
| |
| $L$ocb_enc_odd: |
| lea r12,[1+r8] |
| lea r13,[3+r8] |
| lea r14,[5+r8] |
| lea r8,[6+r8] |
| bsf r12,r12 |
| bsf r13,r13 |
| bsf r14,r14 |
| shl r12,4 |
| shl r13,4 |
| shl r14,4 |
| |
| sub rdx,6 |
| jc NEAR $L$ocb_enc_short |
| jmp NEAR $L$ocb_enc_grandloop |
| |
| ALIGN 32 |
| $L$ocb_enc_grandloop: |
| movdqu xmm2,XMMWORD[rdi] |
| movdqu xmm3,XMMWORD[16+rdi] |
| movdqu xmm4,XMMWORD[32+rdi] |
| movdqu xmm5,XMMWORD[48+rdi] |
| movdqu xmm6,XMMWORD[64+rdi] |
| movdqu xmm7,XMMWORD[80+rdi] |
| lea rdi,[96+rdi] |
| |
| call __ocb_encrypt6 |
| |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| movups XMMWORD[80+rsi],xmm7 |
| lea rsi,[96+rsi] |
| sub rdx,6 |
| jnc NEAR $L$ocb_enc_grandloop |
| |
| $L$ocb_enc_short: |
| add rdx,6 |
| jz NEAR $L$ocb_enc_done |
| |
| movdqu xmm2,XMMWORD[rdi] |
| cmp rdx,2 |
| jb NEAR $L$ocb_enc_one |
| movdqu xmm3,XMMWORD[16+rdi] |
| je NEAR $L$ocb_enc_two |
| |
| movdqu xmm4,XMMWORD[32+rdi] |
| cmp rdx,4 |
| jb NEAR $L$ocb_enc_three |
| movdqu xmm5,XMMWORD[48+rdi] |
| je NEAR $L$ocb_enc_four |
| |
| movdqu xmm6,XMMWORD[64+rdi] |
| pxor xmm7,xmm7 |
| |
| call __ocb_encrypt6 |
| |
| movdqa xmm15,xmm14 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| movups XMMWORD[64+rsi],xmm6 |
| |
| jmp NEAR $L$ocb_enc_done |
| |
| ALIGN 16 |
| $L$ocb_enc_one: |
| movdqa xmm7,xmm10 |
| |
| call __ocb_encrypt1 |
| |
| movdqa xmm15,xmm7 |
| movups XMMWORD[rsi],xmm2 |
| jmp NEAR $L$ocb_enc_done |
| |
| ALIGN 16 |
| $L$ocb_enc_two: |
| pxor xmm4,xmm4 |
| pxor xmm5,xmm5 |
| |
| call __ocb_encrypt4 |
| |
| movdqa xmm15,xmm11 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| |
| jmp NEAR $L$ocb_enc_done |
| |
| ALIGN 16 |
| $L$ocb_enc_three: |
| pxor xmm5,xmm5 |
| |
| call __ocb_encrypt4 |
| |
| movdqa xmm15,xmm12 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| |
| jmp NEAR $L$ocb_enc_done |
| |
| ALIGN 16 |
| $L$ocb_enc_four: |
| call __ocb_encrypt4 |
| |
| movdqa xmm15,xmm13 |
| movups XMMWORD[rsi],xmm2 |
| movups XMMWORD[16+rsi],xmm3 |
| movups XMMWORD[32+rsi],xmm4 |
| movups XMMWORD[48+rsi],xmm5 |
| |
| $L$ocb_enc_done: |
| pxor xmm15,xmm0 |
| movdqu XMMWORD[rbp],xmm8 |
| movdqu XMMWORD[r9],xmm15 |
| |
| xorps xmm0,xmm0 |
| pxor xmm1,xmm1 |
| pxor xmm2,xmm2 |
| pxor xmm3,xmm3 |
| pxor xmm4,xmm4 |
| pxor xmm5,xmm5 |
| movaps xmm6,XMMWORD[rsp] |
| movaps XMMWORD[rsp],xmm0 |
| movaps xmm7,XMMWORD[16+rsp] |
| movaps XMMWORD[16+rsp],xmm0 |
| movaps xmm8,XMMWORD[32+rsp] |
| movaps XMMWORD[32+rsp],xmm0 |
| movaps xmm9,XMMWORD[48+rsp] |
| movaps XMMWORD[48+rsp],xmm0 |
| movaps xmm10,XMMWORD[64+rsp] |
| movaps XMMWORD[64+rsp],xmm0 |
| movaps xmm11,XMMWORD[80+rsp] |
| movaps XMMWORD[80+rsp],xmm0 |
| movaps xmm12,XMMWORD[96+rsp] |
| movaps XMMWORD[96+rsp],xmm0 |
| movaps xmm13,XMMWORD[112+rsp] |
| movaps XMMWORD[112+rsp],xmm0 |
| movaps xmm14,XMMWORD[128+rsp] |
| movaps XMMWORD[128+rsp],xmm0 |
| movaps xmm15,XMMWORD[144+rsp] |
| movaps XMMWORD[144+rsp],xmm0 |
| lea rax,[((160+40))+rsp] |
| $L$ocb_enc_pop: |
| mov r14,QWORD[((-40))+rax] |
| |
| mov r13,QWORD[((-32))+rax] |
| |
| mov r12,QWORD[((-24))+rax] |
| |
| mov rbp,QWORD[((-16))+rax] |
| |
| mov rbx,QWORD[((-8))+rax] |
| |
| lea rsp,[rax] |
| |
| $L$ocb_enc_epilogue: |
| mov rdi,QWORD[8+rsp] ;WIN64 epilogue |
| mov rsi,QWORD[16+rsp] |
| DB 0F3h,0C3h ;repret |
| |
| $L$SEH_end_aes_hw_ocb_encrypt: |
| |
| |
| ALIGN 32 |
| __ocb_encrypt6: |
| pxor xmm15,xmm9 |
| movdqu xmm11,XMMWORD[r12*1+rbx] |
| movdqa xmm12,xmm10 |
| movdqu xmm13,XMMWORD[r13*1+rbx] |
| movdqa xmm14,xmm10 |
| pxor xmm10,xmm15 |
| movdqu xmm15,XMMWORD[r14*1+rbx] |
| pxor xmm11,xmm10 |
| pxor xmm8,xmm2 |
| pxor xmm2,xmm10 |
| pxor xmm12,xmm11 |
| pxor xmm8,xmm3 |
| pxor xmm3,xmm11 |
| pxor xmm13,xmm12 |
| pxor xmm8,xmm4 |
| pxor xmm4,xmm12 |
| pxor xmm14,xmm13 |
| pxor xmm8,xmm5 |
| pxor xmm5,xmm13 |
| pxor xmm15,xmm14 |
| pxor xmm8,xmm6 |
| pxor xmm6,xmm14 |
| pxor xmm8,xmm7 |
| pxor xmm7,xmm15 |
| movups xmm0,XMMWORD[32+r11] |
| |
| lea r12,[1+r8] |
| lea r13,[3+r8] |
| lea r14,[5+r8] |
| add r8,6 |
| pxor xmm10,xmm9 |
| bsf r12,r12 |
| bsf r13,r13 |
| bsf r14,r14 |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| pxor xmm11,xmm9 |
| pxor xmm12,xmm9 |
| DB 102,15,56,220,241 |
| pxor xmm13,xmm9 |
| pxor xmm14,xmm9 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[48+r11] |
| pxor xmm15,xmm9 |
| |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| movups xmm0,XMMWORD[64+r11] |
| shl r12,4 |
| shl r13,4 |
| jmp NEAR $L$ocb_enc_loop6 |
| |
| ALIGN 32 |
| $L$ocb_enc_loop6: |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[rax*1+rcx] |
| add rax,32 |
| |
| DB 102,15,56,220,208 |
| DB 102,15,56,220,216 |
| DB 102,15,56,220,224 |
| DB 102,15,56,220,232 |
| DB 102,15,56,220,240 |
| DB 102,15,56,220,248 |
| movups xmm0,XMMWORD[((-16))+rax*1+rcx] |
| jnz NEAR $L$ocb_enc_loop6 |
| |
| DB 102,15,56,220,209 |
| DB 102,15,56,220,217 |
| DB 102,15,56,220,225 |
| DB 102,15,56,220,233 |
| DB 102,15,56,220,241 |
| DB 102,15,56,220,249 |
| movups xmm1,XMMWORD[16+r11] |
| shl r14,4 |
| |
| DB 102,65,15,56,221,210 |
| movdqu xmm10,XMMWORD[rbx] |
| mov rax,r10 |
| DB 102,65,15,56,221,219 |
| DB 102,65,15,56,221,228 |
| DB 102,65,15,56,221,237 |
| DB 102,65,15,56,221,246 |
| DB 102,65,15,56,221,255 |
| DB 0F3h,0C3h ;repret |
| |
| |
| |
| ALIGN 32 |
| __ocb_encrypt4: |
| pxor xmm15,xmm9 |
| movdqu xmm11,XMMWORD[r12*1+rbx] |
| movdqa xmm12,xmm10 |
| movdqu xmm13,XMMWORD[r13*1+rbx] |
| pxor xmm10,xmm15 |
| pxor xmm11,xmm10 |
| pxor xmm8,xmm2 |
| pxor xmm2,xmm10 |
