diff options
Diffstat (limited to 'secure/lib/libcrypto/i386/chacha-x86.S')
| -rw-r--r-- | secure/lib/libcrypto/i386/chacha-x86.S | 741 |
1 files changed, 741 insertions, 0 deletions
diff --git a/secure/lib/libcrypto/i386/chacha-x86.S b/secure/lib/libcrypto/i386/chacha-x86.S new file mode 100644 index 000000000000..261c6010f89a --- /dev/null +++ b/secure/lib/libcrypto/i386/chacha-x86.S @@ -0,0 +1,741 @@ +/* $FreeBSD$ */ +/* Do not modify. This file is auto-generated from chacha-x86.pl. */ +#ifdef PIC +.text +.globl ChaCha20_ctr32 +.type ChaCha20_ctr32,@function +.align 16 +ChaCha20_ctr32: +.L_ChaCha20_ctr32_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + xorl %eax,%eax + cmpl 28(%esp),%eax + je .L000no_data + movl 32(%esp),%esi + movl 36(%esp),%edi + subl $132,%esp + movl (%esi),%eax + movl 4(%esi),%ebx + movl 8(%esi),%ecx + movl 12(%esi),%edx + movl %eax,80(%esp) + movl %ebx,84(%esp) + movl %ecx,88(%esp) + movl %edx,92(%esp) + movl 16(%esi),%eax + movl 20(%esi),%ebx + movl 24(%esi),%ecx + movl 28(%esi),%edx + movl %eax,96(%esp) + movl %ebx,100(%esp) + movl %ecx,104(%esp) + movl %edx,108(%esp) + movl (%edi),%eax + movl 4(%edi),%ebx + movl 8(%edi),%ecx + movl 12(%edi),%edx + subl $1,%eax + movl %eax,112(%esp) + movl %ebx,116(%esp) + movl %ecx,120(%esp) + movl %edx,124(%esp) + jmp .L001entry +.align 16 +.L002outer_loop: + movl %ebx,156(%esp) + movl %eax,152(%esp) + movl %ecx,160(%esp) +.L001entry: + movl $1634760805,%eax + movl $857760878,4(%esp) + movl $2036477234,8(%esp) + movl $1797285236,12(%esp) + movl 84(%esp),%ebx + movl 88(%esp),%ebp + movl 104(%esp),%ecx + movl 108(%esp),%esi + movl 116(%esp),%edx + movl 120(%esp),%edi + movl %ebx,20(%esp) + movl %ebp,24(%esp) + movl %ecx,40(%esp) + movl %esi,44(%esp) + movl %edx,52(%esp) + movl %edi,56(%esp) + movl 92(%esp),%ebx + movl 124(%esp),%edi + movl 112(%esp),%edx + movl 80(%esp),%ebp + movl 96(%esp),%ecx + movl 100(%esp),%esi + addl $1,%edx + movl %ebx,28(%esp) + movl %edi,60(%esp) + movl %edx,112(%esp) + movl $10,%ebx + jmp .L003loop +.align 16 +.L003loop: + addl %ebp,%eax + movl %ebx,128(%esp) + movl %ebp,%ebx + xorl %eax,%edx + roll $16,%edx + addl %edx,%ecx + xorl %ecx,%ebx + movl 52(%esp),%edi + roll $12,%ebx + movl 20(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,(%esp) + roll $8,%edx + movl 4(%esp),%eax + addl %edx,%ecx + movl %edx,48(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + movl %ecx,32(%esp) + roll $16,%edi + movl %ebx,16(%esp) + addl %edi,%esi + movl 40(%esp),%ecx + xorl %esi,%ebp + movl 56(%esp),%edx + roll $12,%ebp + movl 24(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,4(%esp) + roll $8,%edi + movl 8(%esp),%eax + addl %edi,%esi + movl %edi,52(%esp) + xorl %esi,%ebp + addl %ebx,%eax + roll $7,%ebp + xorl %eax,%edx + movl %esi,36(%esp) + roll $16,%edx + movl %ebp,20(%esp) + addl %edx,%ecx + movl 44(%esp),%esi + xorl %ecx,%ebx + movl 60(%esp),%edi + roll $12,%ebx + movl 28(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,8(%esp) + roll $8,%edx + movl 12(%esp),%eax + addl %edx,%ecx + movl %edx,56(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + roll $16,%edi + movl %ebx,24(%esp) + addl %edi,%esi + xorl %esi,%ebp + roll $12,%ebp + movl 20(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,12(%esp) + roll $8,%edi + movl (%esp),%eax + addl %edi,%esi + movl %edi,%edx + xorl %esi,%ebp + addl %ebx,%eax + roll $7,%ebp + xorl %eax,%edx + roll $16,%edx + movl %ebp,28(%esp) + addl %edx,%ecx + xorl %ecx,%ebx + movl 48(%esp),%edi + roll $12,%ebx + movl 24(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,(%esp) + roll $8,%edx + movl 4(%esp),%eax + addl %edx,%ecx + movl %edx,60(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + movl %ecx,40(%esp) + roll $16,%edi + movl %ebx,20(%esp) + addl %edi,%esi + movl 32(%esp),%ecx + xorl %esi,%ebp + movl 52(%esp),%edx + roll $12,%ebp + movl 28(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,4(%esp) + roll $8,%edi + movl 8(%esp),%eax + addl %edi,%esi + movl %edi,48(%esp) + xorl %esi,%ebp + addl %ebx,%eax + roll $7,%ebp + xorl %eax,%edx + movl %esi,44(%esp) + roll $16,%edx + movl %ebp,24(%esp) + addl %edx,%ecx + movl 36(%esp),%esi + xorl %ecx,%ebx + movl 56(%esp),%edi + roll $12,%ebx + movl 16(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,8(%esp) + roll $8,%edx + movl 12(%esp),%eax + addl %edx,%ecx + movl %edx,52(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + roll $16,%edi + movl %ebx,28(%esp) + addl %edi,%esi + xorl %esi,%ebp + movl 48(%esp),%edx + roll $12,%ebp + movl 128(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,12(%esp) + roll $8,%edi + movl (%esp),%eax + addl %edi,%esi + movl %edi,56(%esp) + xorl %esi,%ebp + roll $7,%ebp + decl %ebx + jnz .L003loop + movl 160(%esp),%ebx + addl $1634760805,%eax + addl 80(%esp),%ebp + addl 96(%esp),%ecx + addl 100(%esp),%esi + cmpl $64,%ebx + jb .L004tail + movl 156(%esp),%ebx + addl 112(%esp),%edx + addl 120(%esp),%edi + xorl (%ebx),%eax + xorl 16(%ebx),%ebp + movl %eax,(%esp) + movl 152(%esp),%eax + xorl 32(%ebx),%ecx + xorl 36(%ebx),%esi + xorl 48(%ebx),%edx + xorl 56(%ebx),%edi + movl %ebp,16(%eax) + movl %ecx,32(%eax) + movl %esi,36(%eax) + movl %edx,48(%eax) + movl %edi,56(%eax) + movl 4(%esp),%ebp + movl 8(%esp),%ecx + movl 12(%esp),%esi + movl 20(%esp),%edx + movl 24(%esp),%edi + addl $857760878,%ebp + addl $2036477234,%ecx + addl $1797285236,%esi + addl 84(%esp),%edx + addl 88(%esp),%edi + xorl 4(%ebx),%ebp + xorl 8(%ebx),%ecx + xorl 12(%ebx),%esi + xorl 20(%ebx),%edx + xorl 24(%ebx),%edi + movl %ebp,4(%eax) + movl %ecx,8(%eax) + movl %esi,12(%eax) + movl %edx,20(%eax) + movl %edi,24(%eax) + movl 28(%esp),%ebp + movl 40(%esp),%ecx + movl 44(%esp),%esi + movl 52(%esp),%edx + movl 60(%esp),%edi + addl 92(%esp),%ebp + addl 104(%esp),%ecx + addl 108(%esp),%esi + addl 116(%esp),%edx + addl 124(%esp),%edi + xorl 28(%ebx),%ebp + xorl 40(%ebx),%ecx + xorl 44(%ebx),%esi + xorl 52(%ebx),%edx + xorl 60(%ebx),%edi + leal 64(%ebx),%ebx + movl %ebp,28(%eax) + movl (%esp),%ebp + movl %ecx,40(%eax) + movl 160(%esp),%ecx + movl %esi,44(%eax) + movl %edx,52(%eax) + movl %edi,60(%eax) + movl %ebp,(%eax) + leal 64(%eax),%eax + subl $64,%ecx + jnz .L002outer_loop + jmp .L005done +.L004tail: + addl 112(%esp),%edx + addl 120(%esp),%edi + movl %eax,(%esp) + movl %ebp,16(%esp) + movl %ecx,32(%esp) + movl %esi,36(%esp) + movl %edx,48(%esp) + movl %edi,56(%esp) + movl 4(%esp),%ebp + movl 8(%esp),%ecx + movl 12(%esp),%esi + movl 20(%esp),%edx + movl 24(%esp),%edi + addl $857760878,%ebp + addl $2036477234,%ecx + addl $1797285236,%esi + addl 84(%esp),%edx + addl 88(%esp),%edi + movl %ebp,4(%esp) + movl %ecx,8(%esp) + movl %esi,12(%esp) + movl %edx,20(%esp) + movl %edi,24(%esp) + movl 28(%esp),%ebp + movl 40(%esp),%ecx + movl 44(%esp),%esi + movl 52(%esp),%edx + movl 60(%esp),%edi + addl 92(%esp),%ebp + addl 104(%esp),%ecx + addl 108(%esp),%esi + addl 116(%esp),%edx + addl 124(%esp),%edi + movl %ebp,28(%esp) + movl 156(%esp),%ebp + movl %ecx,40(%esp) + movl 152(%esp),%ecx + movl %esi,44(%esp) + xorl %esi,%esi + movl %edx,52(%esp) + movl %edi,60(%esp) + xorl %eax,%eax + xorl %edx,%edx +.L006tail_loop: + movb (%esi,%ebp,1),%al + movb (%esp,%esi,1),%dl + leal 1(%esi),%esi + xorb %dl,%al + movb %al,-1(%ecx,%esi,1) + decl %ebx + jnz .L006tail_loop +.L005done: + addl $132,%esp +.L000no_data: + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +.size ChaCha20_ctr32,.-.L_ChaCha20_ctr32_begin +.byte 67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54 +.byte 44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32 +.byte 60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111 +.byte 114,103,62,0 +#else +.text +.globl ChaCha20_ctr32 +.type ChaCha20_ctr32,@function +.align 16 +ChaCha20_ctr32: +.L_ChaCha20_ctr32_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + xorl %eax,%eax + cmpl 28(%esp),%eax + je .L000no_data + movl 32(%esp),%esi + movl 36(%esp),%edi + subl $132,%esp + movl (%esi),%eax + movl 4(%esi),%ebx + movl 8(%esi),%ecx + movl 12(%esi),%edx + movl %eax,80(%esp) + movl %ebx,84(%esp) + movl %ecx,88(%esp) + movl %edx,92(%esp) + movl 16(%esi),%eax + movl 20(%esi),%ebx + movl 24(%esi),%ecx + movl 28(%esi),%edx + movl %eax,96(%esp) + movl %ebx,100(%esp) + movl %ecx,104(%esp) + movl %edx,108(%esp) + movl (%edi),%eax + movl 4(%edi),%ebx + movl 8(%edi),%ecx + movl 12(%edi),%edx + subl $1,%eax + movl %eax,112(%esp) + movl %ebx,116(%esp) + movl %ecx,120(%esp) + movl %edx,124(%esp) + jmp .L001entry +.align 16 +.L002outer_loop: + movl %ebx,156(%esp) + movl %eax,152(%esp) + movl %ecx,160(%esp) +.L001entry: + movl $1634760805,%eax + movl $857760878,4(%esp) + movl $2036477234,8(%esp) + movl $1797285236,12(%esp) + movl 84(%esp),%ebx + movl 88(%esp),%ebp + movl 104(%esp),%ecx + movl 108(%esp),%esi + movl 116(%esp),%edx + movl 120(%esp),%edi + movl %ebx,20(%esp) + movl %ebp,24(%esp) + movl %ecx,40(%esp) + movl %esi,44(%esp) + movl %edx,52(%esp) + movl %edi,56(%esp) + movl 92(%esp),%ebx + movl 124(%esp),%edi + movl 112(%esp),%edx + movl 80(%esp),%ebp + movl 96(%esp),%ecx + movl 100(%esp),%esi + addl $1,%edx + movl %ebx,28(%esp) + movl %edi,60(%esp) + movl %edx,112(%esp) + movl $10,%ebx + jmp .L003loop +.align 16 +.L003loop: + addl %ebp,%eax + movl %ebx,128(%esp) + movl %ebp,%ebx + xorl %eax,%edx + roll $16,%edx + addl %edx,%ecx + xorl %ecx,%ebx + movl 52(%esp),%edi + roll $12,%ebx + movl 20(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,(%esp) + roll $8,%edx + movl 4(%esp),%eax + addl %edx,%ecx + movl %edx,48(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + movl %ecx,32(%esp) + roll $16,%edi + movl %ebx,16(%esp) + addl %edi,%esi + movl 40(%esp),%ecx + xorl %esi,%ebp + movl 56(%esp),%edx + roll $12,%ebp + movl 24(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,4(%esp) + roll $8,%edi + movl 8(%esp),%eax + addl %edi,%esi + movl %edi,52(%esp) + xorl %esi,%ebp + addl %ebx,%eax + roll $7,%ebp + xorl %eax,%edx + movl %esi,36(%esp) + roll $16,%edx + movl %ebp,20(%esp) + addl %edx,%ecx + movl 44(%esp),%esi + xorl %ecx,%ebx + movl 60(%esp),%edi + roll $12,%ebx + movl 28(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,8(%esp) + roll $8,%edx + movl 12(%esp),%eax + addl %edx,%ecx + movl %edx,56(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + roll $16,%edi + movl %ebx,24(%esp) + addl %edi,%esi + xorl %esi,%ebp + roll $12,%ebp + movl 20(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,12(%esp) + roll $8,%edi + movl (%esp),%eax + addl %edi,%esi + movl %edi,%edx + xorl %esi,%ebp + addl %ebx,%eax + roll $7,%ebp + xorl %eax,%edx + roll $16,%edx + movl %ebp,28(%esp) + addl %edx,%ecx + xorl %ecx,%ebx + movl 48(%esp),%edi + roll $12,%ebx + movl 24(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,(%esp) + roll $8,%edx + movl 4(%esp),%eax + addl %edx,%ecx + movl %edx,60(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + movl %ecx,40(%esp) + roll $16,%edi + movl %ebx,20(%esp) + addl %edi,%esi + movl 32(%esp),%ecx + xorl %esi,%ebp + movl 52(%esp),%edx + roll $12,%ebp + movl 28(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,4(%esp) + roll $8,%edi + movl 8(%esp),%eax + addl %edi,%esi + movl %edi,48(%esp) + xorl %esi,%ebp + addl %ebx,%eax + roll $7,%ebp + xorl %eax,%edx + movl %esi,44(%esp) + roll $16,%edx + movl %ebp,24(%esp) + addl %edx,%ecx + movl 36(%esp),%esi + xorl %ecx,%ebx + movl 56(%esp),%edi + roll $12,%ebx + movl 16(%esp),%ebp + addl %ebx,%eax + xorl %eax,%edx + movl %eax,8(%esp) + roll $8,%edx + movl 12(%esp),%eax + addl %edx,%ecx + movl %edx,52(%esp) + xorl %ecx,%ebx + addl %ebp,%eax + roll $7,%ebx + xorl %eax,%edi + roll $16,%edi + movl %ebx,28(%esp) + addl %edi,%esi + xorl %esi,%ebp + movl 48(%esp),%edx + roll $12,%ebp + movl 128(%esp),%ebx + addl %ebp,%eax + xorl %eax,%edi + movl %eax,12(%esp) + roll $8,%edi + movl (%esp),%eax + addl %edi,%esi + movl %edi,56(%esp) + xorl %esi,%ebp + roll $7,%ebp + decl %ebx + jnz .L003loop + movl 160(%esp),%ebx + addl $1634760805,%eax + addl 80(%esp),%ebp + addl 96(%esp),%ecx + addl 100(%esp),%esi + cmpl $64,%ebx + jb .L004tail + movl 156(%esp),%ebx + addl 112(%esp),%edx + addl 120(%esp),%edi + xorl (%ebx),%eax + xorl 16(%ebx),%ebp + movl %eax,(%esp) + movl 152(%esp),%eax + xorl 32(%ebx),%ecx + xorl 36(%ebx),%esi + xorl 48(%ebx),%edx + xorl 56(%ebx),%edi + movl %ebp,16(%eax) + movl %ecx,32(%eax) + movl %esi,36(%eax) + movl %edx,48(%eax) + movl %edi,56(%eax) + movl 4(%esp),%ebp + movl 8(%esp),%ecx + movl 12(%esp),%esi + movl 20(%esp),%edx + movl 24(%esp),%edi + addl $857760878,%ebp + addl $2036477234,%ecx + addl $1797285236,%esi + addl 84(%esp),%edx + addl 88(%esp),%edi + xorl 4(%ebx),%ebp + xorl 8(%ebx),%ecx + xorl 12(%ebx),%esi + xorl 20(%ebx),%edx + xorl 24(%ebx),%edi + movl %ebp,4(%eax) + movl %ecx,8(%eax) + movl %esi,12(%eax) + movl %edx,20(%eax) + movl %edi,24(%eax) + movl 28(%esp),%ebp + movl 40(%esp),%ecx + movl 44(%esp),%esi + movl 52(%esp),%edx + movl 60(%esp),%edi + addl 92(%esp),%ebp + addl 104(%esp),%ecx + addl 108(%esp),%esi + addl 116(%esp),%edx + addl 124(%esp),%edi + xorl 28(%ebx),%ebp + xorl 40(%ebx),%ecx + xorl 44(%ebx),%esi + xorl 52(%ebx),%edx + xorl 60(%ebx),%edi + leal 64(%ebx),%ebx + movl %ebp,28(%eax) + movl (%esp),%ebp + movl %ecx,40(%eax) + movl 160(%esp),%ecx + movl %esi,44(%eax) + movl %edx,52(%eax) + movl %edi,60(%eax) + movl %ebp,(%eax) + leal 64(%eax),%eax + subl $64,%ecx + jnz .L002outer_loop + jmp .L005done +.L004tail: + addl 112(%esp),%edx + addl 120(%esp),%edi + movl %eax,(%esp) + movl %ebp,16(%esp) + movl %ecx,32(%esp) + movl %esi,36(%esp) + movl %edx,48(%esp) + movl %edi,56(%esp) + movl 4(%esp),%ebp + movl 8(%esp),%ecx + movl 12(%esp),%esi + movl 20(%esp),%edx + movl 24(%esp),%edi + addl $857760878,%ebp + addl $2036477234,%ecx + addl $1797285236,%esi + addl 84(%esp),%edx + addl 88(%esp),%edi + movl %ebp,4(%esp) + movl %ecx,8(%esp) + movl %esi,12(%esp) + movl %edx,20(%esp) + movl %edi,24(%esp) + movl 28(%esp),%ebp + movl 40(%esp),%ecx + movl 44(%esp),%esi + movl 52(%esp),%edx + movl 60(%esp),%edi + addl 92(%esp),%ebp + addl 104(%esp),%ecx + addl 108(%esp),%esi + addl 116(%esp),%edx + addl 124(%esp),%edi + movl %ebp,28(%esp) + movl 156(%esp),%ebp + movl %ecx,40(%esp) + movl 152(%esp),%ecx + movl %esi,44(%esp) + xorl %esi,%esi + movl %edx,52(%esp) + movl %edi,60(%esp) + xorl %eax,%eax + xorl %edx,%edx +.L006tail_loop: + movb (%esi,%ebp,1),%al + movb (%esp,%esi,1),%dl + leal 1(%esi),%esi + xorb %dl,%al + movb %al,-1(%ecx,%esi,1) + decl %ebx + jnz .L006tail_loop +.L005done: + addl $132,%esp +.L000no_data: + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +.size ChaCha20_ctr32,.-.L_ChaCha20_ctr32_begin +.byte 67,104,97,67,104,97,50,48,32,102,111,114,32,120,56,54 +.byte 44,32,67,82,89,80,84,79,71,65,77,83,32,98,121,32 +.byte 60,97,112,112,114,111,64,111,112,101,110,115,115,108,46,111 +.byte 114,103,62,0 +#endif |
