.file "aesni-x86.s" .text .globl aesni_encrypt .type aesni_encrypt,@function .align 16 aesni_encrypt: .L_aesni_encrypt_begin: movl 4(%esp),%eax movl 12(%esp),%edx movups (%eax),%xmm0 movl 240(%edx),%ecx movl 8(%esp),%eax movups (%edx),%xmm3 movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 .L000enc1_loop: aesenc %xmm4,%xmm0 decl %ecx movups (%edx),%xmm4 leal 16(%edx),%edx jnz .L000enc1_loop aesenclast %xmm4,%xmm0 movups %xmm0,(%eax) ret .size aesni_encrypt,.-.L_aesni_encrypt_begin .globl aesni_decrypt .type aesni_decrypt,@function .align 16 aesni_decrypt: .L_aesni_decrypt_begin: movl 4(%esp),%eax movl 12(%esp),%edx movups (%eax),%xmm0 movl 240(%edx),%ecx movl 8(%esp),%eax movups (%edx),%xmm3 movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 .L001dec1_loop: aesdec %xmm4,%xmm0 decl %ecx movups (%edx),%xmm4 leal 16(%edx),%edx jnz .L001dec1_loop aesdeclast %xmm4,%xmm0 movups %xmm0,(%eax) ret .size aesni_decrypt,.-.L_aesni_decrypt_begin .type _aesni_encrypt3,@function .align 16 _aesni_encrypt3: movups (%edx),%xmm3 shrl $1,%ecx movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 pxor %xmm3,%xmm1 pxor %xmm3,%xmm2 jmp .L002enc3_loop .align 16 .L002enc3_loop: aesenc %xmm4,%xmm0 movups (%edx),%xmm3 aesenc %xmm4,%xmm1 decl %ecx aesenc %xmm4,%xmm2 movups 16(%edx),%xmm4 aesenc %xmm3,%xmm0 leal 32(%edx),%edx aesenc %xmm3,%xmm1 aesenc %xmm3,%xmm2 jnz .L002enc3_loop aesenc %xmm4,%xmm0 movups (%edx),%xmm3 aesenc %xmm4,%xmm1 aesenc %xmm4,%xmm2 aesenclast %xmm3,%xmm0 aesenclast %xmm3,%xmm1 aesenclast %xmm3,%xmm2 ret .size _aesni_encrypt3,.-_aesni_encrypt3 .type _aesni_decrypt3,@function .align 16 _aesni_decrypt3: movups (%edx),%xmm3 shrl $1,%ecx movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 pxor %xmm3,%xmm1 pxor %xmm3,%xmm2 jmp .L003dec3_loop .align 16 .L003dec3_loop: aesdec %xmm4,%xmm0 movups (%edx),%xmm3 aesdec %xmm4,%xmm1 decl %ecx aesdec %xmm4,%xmm2 movups 16(%edx),%xmm4 aesdec %xmm3,%xmm0 leal 32(%edx),%edx aesdec %xmm3,%xmm1 aesdec %xmm3,%xmm2 jnz .L003dec3_loop aesdec %xmm4,%xmm0 movups (%edx),%xmm3 aesdec %xmm4,%xmm1 aesdec %xmm4,%xmm2 aesdeclast %xmm3,%xmm0 aesdeclast %xmm3,%xmm1 aesdeclast %xmm3,%xmm2 ret .size _aesni_decrypt3,.-_aesni_decrypt3 .type _aesni_encrypt4,@function .align 16 _aesni_encrypt4: movups (%edx),%xmm3 movups 16(%edx),%xmm4 shrl $1,%ecx leal 32(%edx),%edx pxor %xmm3,%xmm0 pxor %xmm3,%xmm1 pxor %xmm3,%xmm2 pxor %xmm3,%xmm7 jmp .L004enc3_loop .align 16 .L004enc3_loop: aesenc %xmm4,%xmm0 movups (%edx),%xmm3 aesenc %xmm4,%xmm1 decl %ecx aesenc %xmm4,%xmm2 aesenc %xmm4,%xmm7 movups 16(%edx),%xmm4 aesenc %xmm3,%xmm0 leal 32(%edx),%edx aesenc %xmm3,%xmm1 aesenc %xmm3,%xmm2 aesenc %xmm3,%xmm7 jnz .L004enc3_loop aesenc %xmm4,%xmm0 movups (%edx),%xmm3 aesenc %xmm4,%xmm1 aesenc %xmm4,%xmm2 aesenc %xmm4,%xmm7 aesenclast %xmm3,%xmm0 aesenclast %xmm3,%xmm1 aesenclast %xmm3,%xmm2 aesenclast %xmm3,%xmm7 ret .size _aesni_encrypt4,.-_aesni_encrypt4 .type _aesni_decrypt4,@function .align 16 _aesni_decrypt4: movups (%edx),%xmm3 movups 16(%edx),%xmm4 shrl $1,%ecx leal 32(%edx),%edx pxor %xmm3,%xmm0 pxor %xmm3,%xmm1 pxor %xmm3,%xmm2 pxor %xmm3,%xmm7 jmp .L005dec3_loop .align 16 .L005dec3_loop: aesdec %xmm4,%xmm0 movups (%edx),%xmm3 aesdec %xmm4,%xmm1 decl %ecx aesdec %xmm4,%xmm2 aesdec %xmm4,%xmm7 movups 16(%edx),%xmm4 aesdec %xmm3,%xmm0 leal 32(%edx),%edx aesdec %xmm3,%xmm1 aesdec %xmm3,%xmm2 aesdec %xmm3,%xmm7 jnz .L005dec3_loop aesdec %xmm4,%xmm0 movups (%edx),%xmm3 aesdec %xmm4,%xmm1 aesdec %xmm4,%xmm2 aesdec %xmm4,%xmm7 aesdeclast %xmm3,%xmm0 aesdeclast %xmm3,%xmm1 aesdeclast %xmm3,%xmm2 aesdeclast %xmm3,%xmm7 ret .size _aesni_decrypt4,.-_aesni_decrypt4 .globl aesni_ecb_encrypt .type aesni_ecb_encrypt,@function .align 16 aesni_ecb_encrypt: .L_aesni_ecb_encrypt_begin: pushl %ebp pushl %ebx pushl %esi pushl %edi movl 20(%esp),%esi movl 24(%esp),%edi movl 28(%esp),%eax movl 32(%esp),%edx movl 36(%esp),%ecx cmpl $16,%eax jb .L006ecb_ret andl $-16,%eax testl %ecx,%ecx movl 240(%edx),%ecx movl %edx,%ebp movl %ecx,%ebx jz .L007ecb_decrypt subl $64,%eax jbe .L008ecb_enc_tail jmp .L009ecb_enc_loop3 .align 16 .L009ecb_enc_loop3: movups (%esi),%xmm0 movups 16(%esi),%xmm1 movups 32(%esi),%xmm2 call _aesni_encrypt3 subl $48,%eax leal 48(%esi),%esi leal 48(%edi),%edi movups %xmm0,-48(%edi) movl %ebp,%edx movups %xmm1,-32(%edi) movl %ebx,%ecx movups %xmm2,-16(%edi) ja .L009ecb_enc_loop3 .L008ecb_enc_tail: addl $64,%eax jz .L006ecb_ret cmpl $16,%eax movups (%esi),%xmm0 je .L010ecb_enc_one cmpl $32,%eax movups 16(%esi),%xmm1 je .L011ecb_enc_two cmpl $48,%eax movups 32(%esi),%xmm2 je .L012ecb_enc_three movups 48(%esi),%xmm7 call _aesni_encrypt4 movups %xmm0,(%edi) movups %xmm1,16(%edi) movups %xmm2,32(%edi) movups %xmm7,48(%edi) jmp .L006ecb_ret .align 16 .L010ecb_enc_one: movups (%edx),%xmm3 movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 .L013enc1_loop: aesenc %xmm4,%xmm0 decl %ecx movups (%edx),%xmm4 leal 16(%edx),%edx jnz .L013enc1_loop aesenclast %xmm4,%xmm0 movups %xmm0,(%edi) jmp .L006ecb_ret .align 16 .L011ecb_enc_two: call _aesni_encrypt3 movups %xmm0,(%edi) movups %xmm1,16(%edi) jmp .L006ecb_ret .align 16 .L012ecb_enc_three: call _aesni_encrypt3 movups %xmm0,(%edi) movups %xmm1,16(%edi) movups %xmm2,32(%edi) jmp .L006ecb_ret .align 16 .L007ecb_decrypt: subl $64,%eax jbe .L014ecb_dec_tail jmp .L015ecb_dec_loop3 .align 16 .L015ecb_dec_loop3: movups (%esi),%xmm0 movups 16(%esi),%xmm1 movups 32(%esi),%xmm2 call _aesni_decrypt3 subl $48,%eax leal 48(%esi),%esi leal 48(%edi),%edi movups %xmm0,-48(%edi) movl %ebp,%edx movups %xmm1,-32(%edi) movl %ebx,%ecx movups %xmm2,-16(%edi) ja .L015ecb_dec_loop3 .L014ecb_dec_tail: addl $64,%eax jz .L006ecb_ret cmpl $16,%eax movups (%esi),%xmm0 je .L016ecb_dec_one cmpl $32,%eax movups 16(%esi),%xmm1 je .L017ecb_dec_two cmpl $48,%eax movups 32(%esi),%xmm2 je .L018ecb_dec_three movups 48(%esi),%xmm7 call _aesni_decrypt4 movups %xmm0,(%edi) movups %xmm1,16(%edi) movups %xmm2,32(%edi) movups %xmm7,48(%edi) jmp .L006ecb_ret .align 16 .L016ecb_dec_one: movups (%edx),%xmm3 movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 .L019dec1_loop: aesdec %xmm4,%xmm0 decl %ecx movups (%edx),%xmm4 leal 16(%edx),%edx jnz .L019dec1_loop aesdeclast %xmm4,%xmm0 movups %xmm0,(%edi) jmp .L006ecb_ret .align 16 .L017ecb_dec_two: call _aesni_decrypt3 movups %xmm0,(%edi) movups %xmm1,16(%edi) jmp .L006ecb_ret .align 16 .L018ecb_dec_three: call _aesni_decrypt3 movups %xmm0,(%edi) movups %xmm1,16(%edi) movups %xmm2,32(%edi) .L006ecb_ret: popl %edi popl %esi popl %ebx popl %ebp ret .size aesni_ecb_encrypt,.-.L_aesni_ecb_encrypt_begin .globl aesni_cbc_encrypt .type aesni_cbc_encrypt,@function .align 16 aesni_cbc_encrypt: .L_aesni_cbc_encrypt_begin: pushl %ebp pushl %ebx pushl %esi pushl %edi movl 20(%esp),%esi movl 24(%esp),%edi movl 28(%esp),%eax movl 32(%esp),%edx testl %eax,%eax movl 36(%esp),%ebp jz .L020cbc_ret cmpl $0,40(%esp) movups (%ebp),%xmm5 movl 240(%edx),%ecx movl %edx,%ebp movl %ecx,%ebx je .L021cbc_decrypt movaps %xmm5,%xmm0 cmpl $16,%eax jb .L022cbc_enc_tail subl $16,%eax jmp .L023cbc_enc_loop .align 16 .L023cbc_enc_loop: movups (%esi),%xmm5 leal 16(%esi),%esi pxor %xmm5,%xmm0 movups (%edx),%xmm3 movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 .L024enc1_loop: aesenc %xmm4,%xmm0 decl %ecx movups (%edx),%xmm4 leal 16(%edx),%edx jnz .L024enc1_loop aesenclast %xmm4,%xmm0 subl $16,%eax leal 16(%edi),%edi movl %ebx,%ecx movl %ebp,%edx movups %xmm0,-16(%edi) jnc .L023cbc_enc_loop addl $16,%eax jnz .L022cbc_enc_tail movaps %xmm0,%xmm5 jmp .L020cbc_ret .L022cbc_enc_tail: movl %eax,%ecx .long 2767451785 movl $16,%ecx subl %eax,%ecx xorl %eax,%eax .long 2868115081 leal -16(%edi),%edi movl %ebx,%ecx movl %edi,%esi movl %ebp,%edx jmp .L023cbc_enc_loop .align 16 .L021cbc_decrypt: subl $64,%eax jbe .L025cbc_dec_tail jmp .L026cbc_dec_loop3 .align 16 .L026cbc_dec_loop3: movups (%esi),%xmm0 movups 16(%esi),%xmm1 movups 32(%esi),%xmm2 movaps %xmm0,%xmm6 movaps %xmm1,%xmm7 call _aesni_decrypt3 subl $48,%eax leal 48(%esi),%esi leal 48(%edi),%edi pxor %xmm5,%xmm0 pxor %xmm6,%xmm1 movups -16(%esi),%xmm5 pxor %xmm7,%xmm2 movups %xmm0,-48(%edi) movl %ebx,%ecx movups %xmm1,-32(%edi) movl %ebp,%edx movups %xmm2,-16(%edi) ja .L026cbc_dec_loop3 .L025cbc_dec_tail: addl $64,%eax jz .L020cbc_ret movups (%esi),%xmm0 cmpl $16,%eax movaps %xmm0,%xmm6 jbe .L027cbc_dec_one movups 16(%esi),%xmm1 cmpl $32,%eax movaps %xmm1,%xmm7 jbe .L028cbc_dec_two movups 32(%esi),%xmm2 cmpl $48,%eax jbe .L029cbc_dec_three movups 48(%esi),%xmm7 call _aesni_decrypt4 movups 16(%esi),%xmm3 movups 32(%esi),%xmm4 pxor %xmm5,%xmm0 pxor %xmm6,%xmm1 movups 48(%esi),%xmm5 movups %xmm0,(%edi) pxor %xmm3,%xmm2 pxor %xmm4,%xmm7 movups %xmm1,16(%edi) movups %xmm2,32(%edi) movaps %xmm7,%xmm0 leal 48(%edi),%edi jmp .L030cbc_dec_tail_collected .L027cbc_dec_one: movups (%edx),%xmm3 movups 16(%edx),%xmm4 leal 32(%edx),%edx pxor %xmm3,%xmm0 .L031dec1_loop: aesdec %xmm4,%xmm0 decl %ecx movups (%edx),%xmm4 leal 16(%edx),%edx jnz .L031dec1_loop aesdeclast %xmm4,%xmm0 pxor %xmm5,%xmm0 movaps %xmm6,%xmm5 jmp .L030cbc_dec_tail_collected .L028cbc_dec_two: call _aesni_decrypt3 pxor %xmm5,%xmm0 pxor %xmm6,%xmm1 movups %xmm0,(%edi) movaps %xmm1,%xmm0 movaps %xmm7,%xmm5 leal 16(%edi),%edi jmp .L030cbc_dec_tail_collected .L029cbc_dec_three: call _aesni_decrypt3 pxor %xmm5,%xmm0 pxor %xmm6,%xmm1 pxor %xmm7,%xmm2 movups %xmm0,(%edi) movups %xmm1,16(%edi) movaps %xmm2,%xmm0 movups 32(%esi),%xmm5 leal 32(%edi),%edi .L030cbc_dec_tail_collected: andl $15,%eax jnz .L032cbc_dec_tail_partial movups %xmm0,(%edi) jmp .L020cbc_ret .L032cbc_dec_tail_partial: movl %esp,%ebp subl $16,%esp andl $-16,%esp movaps %xmm0,(%esp) movl %esp,%esi movl %eax,%ecx .long 2767451785 movl %ebp,%esp .L020cbc_ret: movl 36(%esp),%ebp movups %xmm5,(%ebp) popl %edi popl %esi popl %ebx popl %ebp ret .size aesni_cbc_encrypt,.-.L_aesni_cbc_encrypt_begin .type _aesni_set_encrypt_key,@function .align 16 _aesni_set_encrypt_key: testl %eax,%eax jz .L033bad_pointer testl %edx,%edx jz .L033bad_pointer movups (%eax),%xmm0 pxor %xmm4,%xmm4 leal 16(%edx),%edx cmpl $256,%ecx je .L03414rounds cmpl $192,%ecx je .L03512rounds cmpl $128,%ecx jne .L036bad_keybits .align 16 .L03710rounds: movl $9,%ecx movups %xmm0,-16(%edx) aeskeygenassist $1,%xmm0,%xmm1 call .L038key_128_cold aeskeygenassist $2,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $4,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $8,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $16,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $32,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $64,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $128,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $27,%xmm0,%xmm1 call .L039key_128 aeskeygenassist $54,%xmm0,%xmm1 call .L039key_128 movups %xmm0,(%edx) movl %ecx,80(%edx) xorl %eax,%eax ret .align 16 .L039key_128: movups %xmm0,(%edx) leal 16(%edx),%edx .L038key_128_cold: shufps $16,%xmm0,%xmm4 pxor %xmm4,%xmm0 shufps $140,%xmm0,%xmm4 pxor %xmm4,%xmm0 pshufd $255,%xmm1,%xmm1 pxor %xmm1,%xmm0 ret .align 16 .L03512rounds: movq 16(%eax),%xmm2 movl $11,%ecx movups %xmm0,-16(%edx) aeskeygenassist $1,%xmm2,%xmm1 call .L040key_192a_cold aeskeygenassist $2,%xmm2,%xmm1 call .L041key_192b aeskeygenassist $4,%xmm2,%xmm1 call .L042key_192a aeskeygenassist $8,%xmm2,%xmm1 call .L041key_192b aeskeygenassist $16,%xmm2,%xmm1 call .L042key_192a aeskeygenassist $32,%xmm2,%xmm1 call .L041key_192b aeskeygenassist $64,%xmm2,%xmm1 call .L042key_192a aeskeygenassist $128,%xmm2,%xmm1 call .L041key_192b movups %xmm0,(%edx) movl %ecx,48(%edx) xorl %eax,%eax ret .align 16 .L042key_192a: movups %xmm0,(%edx) leal 16(%edx),%edx .align 16 .L040key_192a_cold: movaps %xmm2,%xmm5 .L043key_192b_warm: shufps $16,%xmm0,%xmm4 movaps %xmm2,%xmm3 pxor %xmm4,%xmm0 shufps $140,%xmm0,%xmm4 pslldq $4,%xmm3 pxor %xmm4,%xmm0 pshufd $85,%xmm1,%xmm1 pxor %xmm3,%xmm2 pxor %xmm1,%xmm0 pshufd $255,%xmm0,%xmm3 pxor %xmm3,%xmm2 ret .align 16 .L041key_192b: movaps %xmm0,%xmm3 shufps $68,%xmm0,%xmm5 movups %xmm5,(%edx) shufps $78,%xmm2,%xmm3 movups %xmm3,16(%edx) leal 32(%edx),%edx jmp .L043key_192b_warm .align 16 .L03414rounds: movups 16(%eax),%xmm2 movl $13,%ecx leal 16(%edx),%edx movups %xmm0,-32(%edx) movups %xmm2,-16(%edx) aeskeygenassist $1,%xmm2,%xmm1 call .L044key_256a_cold aeskeygenassist $1,%xmm0,%xmm1 call .L045key_256b aeskeygenassist $2,%xmm2,%xmm1 call .L046key_256a aeskeygenassist $2,%xmm0,%xmm1 call .L045key_256b aeskeygenassist $4,%xmm2,%xmm1 call .L046key_256a aeskeygenassist $4,%xmm0,%xmm1 call .L045key_256b aeskeygenassist $8,%xmm2,%xmm1 call .L046key_256a aeskeygenassist $8,%xmm0,%xmm1 call .L045key_256b aeskeygenassist $16,%xmm2,%xmm1 call .L046key_256a aeskeygenassist $16,%xmm0,%xmm1 call .L045key_256b aeskeygenassist $32,%xmm2,%xmm1 call .L046key_256a aeskeygenassist $32,%xmm0,%xmm1 call .L045key_256b aeskeygenassist $64,%xmm2,%xmm1 call .L046key_256a movups %xmm0,(%edx) movl %ecx,16(%edx) xorl %eax,%eax ret .align 16 .L046key_256a: movups %xmm2,(%edx) leal 16(%edx),%edx .L044key_256a_cold: shufps $16,%xmm0,%xmm4 pxor %xmm4,%xmm0 shufps $140,%xmm0,%xmm4 pxor %xmm4,%xmm0 pshufd $255,%xmm1,%xmm1 pxor %xmm1,%xmm0 ret .align 16 .L045key_256b: movups %xmm0,(%edx) leal 16(%edx),%edx shufps $16,%xmm2,%xmm4 pxor %xmm4,%xmm2 shufps $140,%xmm2,%xmm4 pxor %xmm4,%xmm2 pshufd $170,%xmm1,%xmm1 pxor %xmm1,%xmm2 ret .align 4 .L033bad_pointer: movl $-1,%eax ret .align 4 .L036bad_keybits: movl $-2,%eax ret .size _aesni_set_encrypt_key,.-_aesni_set_encrypt_key .globl aesni_set_encrypt_key .type aesni_set_encrypt_key,@function .align 16 aesni_set_encrypt_key: .L_aesni_set_encrypt_key_begin: movl 4(%esp),%eax movl 8(%esp),%ecx movl 12(%esp),%edx call _aesni_set_encrypt_key ret .size aesni_set_encrypt_key,.-.L_aesni_set_encrypt_key_begin .globl aesni_set_decrypt_key .type aesni_set_decrypt_key,@function .align 16 aesni_set_decrypt_key: .L_aesni_set_decrypt_key_begin: movl 4(%esp),%eax movl 8(%esp),%ecx movl 12(%esp),%edx call _aesni_set_encrypt_key movl 12(%esp),%edx shll $4,%ecx testl %eax,%eax jnz .L047dec_key_ret leal 16(%edx,%ecx,1),%eax movups (%edx),%xmm0 movups (%eax),%xmm1 movups %xmm0,(%eax) movups %xmm1,(%edx) leal 16(%edx),%edx leal -16(%eax),%eax .L048dec_key_inverse: movups (%edx),%xmm0 movups (%eax),%xmm1 aesimc %xmm0,%xmm0 aesimc %xmm1,%xmm1 leal 16(%edx),%edx leal -16(%eax),%eax cmpl %edx,%eax movups %xmm0,16(%eax) movups %xmm1,-16(%edx) ja .L048dec_key_inverse movups (%edx),%xmm0 aesimc %xmm0,%xmm0 movups %xmm0,(%edx) xorl %eax,%eax .L047dec_key_ret: ret .size aesni_set_decrypt_key,.-.L_aesni_set_decrypt_key_begin .byte 65,69,83,32,102,111,114,32,73,110,116,101,108,32,65,69 .byte 83,45,78,73,44,32,67,82,89,80,84,79,71,65,77,83 .byte 32,98,121,32,60,97,112,112,114,111,64,111,112,101,110,115 .byte 115,108,46,111,114,103,62,0