diff options
Diffstat (limited to 'mac-x86/crypto/bn/bn-586.S')
| -rw-r--r-- | mac-x86/crypto/bn/bn-586.S | 1379 |
1 files changed, 1379 insertions, 0 deletions
diff --git a/mac-x86/crypto/bn/bn-586.S b/mac-x86/crypto/bn/bn-586.S new file mode 100644 index 0000000..34cf56f --- /dev/null +++ b/mac-x86/crypto/bn/bn-586.S @@ -0,0 +1,1379 @@ +#if defined(__i386__) +.file "src/crypto/bn/asm/bn-586.S" +.text +.globl _bn_mul_add_words +.private_extern _bn_mul_add_words +.align 4 +_bn_mul_add_words: +L_bn_mul_add_words_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + + xorl %esi,%esi + movl 20(%esp),%edi + movl 28(%esp),%ecx + movl 24(%esp),%ebx + andl $4294967288,%ecx + movl 32(%esp),%ebp + pushl %ecx + jz L000maw_finish +.align 4,0x90 +L001maw_loop: + # Round 0 + movl (%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl (%edi),%eax + adcl $0,%edx + movl %eax,(%edi) + movl %edx,%esi + # Round 4 + movl 4(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 4(%edi),%eax + adcl $0,%edx + movl %eax,4(%edi) + movl %edx,%esi + # Round 8 + movl 8(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 8(%edi),%eax + adcl $0,%edx + movl %eax,8(%edi) + movl %edx,%esi + # Round 12 + movl 12(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 12(%edi),%eax + adcl $0,%edx + movl %eax,12(%edi) + movl %edx,%esi + # Round 16 + movl 16(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 16(%edi),%eax + adcl $0,%edx + movl %eax,16(%edi) + movl %edx,%esi + # Round 20 + movl 20(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 20(%edi),%eax + adcl $0,%edx + movl %eax,20(%edi) + movl %edx,%esi + # Round 24 + movl 24(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 24(%edi),%eax + adcl $0,%edx + movl %eax,24(%edi) + movl %edx,%esi + # Round 28 + movl 28(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 28(%edi),%eax + adcl $0,%edx + movl %eax,28(%edi) + movl %edx,%esi + + subl $8,%ecx + leal 32(%ebx),%ebx + leal 32(%edi),%edi + jnz L001maw_loop +L000maw_finish: + movl 32(%esp),%ecx + andl $7,%ecx + jnz L002maw_finish2 + jmp L003maw_end +L002maw_finish2: + # Tail Round 0 + movl (%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl (%edi),%eax + adcl $0,%edx + decl %ecx + movl %eax,(%edi) + movl %edx,%esi + jz L003maw_end + # Tail Round 1 + movl 4(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 4(%edi),%eax + adcl $0,%edx + decl %ecx + movl %eax,4(%edi) + movl %edx,%esi + jz L003maw_end + # Tail Round 2 + movl 8(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 8(%edi),%eax + adcl $0,%edx + decl %ecx + movl %eax,8(%edi) + movl %edx,%esi + jz L003maw_end + # Tail Round 3 + movl 12(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 12(%edi),%eax + adcl $0,%edx + decl %ecx + movl %eax,12(%edi) + movl %edx,%esi + jz L003maw_end + # Tail Round 4 + movl 16(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 16(%edi),%eax + adcl $0,%edx + decl %ecx + movl %eax,16(%edi) + movl %edx,%esi + jz L003maw_end + # Tail Round 5 + movl 20(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 20(%edi),%eax + adcl $0,%edx + decl %ecx + movl %eax,20(%edi) + movl %edx,%esi + jz L003maw_end + # Tail Round 6 + movl 24(%ebx),%eax + mull %ebp + addl %esi,%eax + adcl $0,%edx + addl 24(%edi),%eax + adcl $0,%edx + movl %eax,24(%edi) + movl %edx,%esi +L003maw_end: + movl %esi,%eax + popl %ecx + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +.globl _bn_mul_words +.private_extern _bn_mul_words +.align 4 +_bn_mul_words: +L_bn_mul_words_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + + xorl %esi,%esi + movl 20(%esp),%edi + movl 24(%esp),%ebx + movl 28(%esp),%ebp + movl 32(%esp),%ecx + andl $4294967288,%ebp + jz L004mw_finish +L005mw_loop: + # Round 0 + movl (%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,(%edi) + movl %edx,%esi + # Round 4 + movl 4(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,4(%edi) + movl %edx,%esi + # Round 8 + movl 8(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,8(%edi) + movl %edx,%esi + # Round 12 + movl 12(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,12(%edi) + movl %edx,%esi + # Round 16 + movl 16(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,16(%edi) + movl %edx,%esi + # Round 20 + movl 20(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,20(%edi) + movl %edx,%esi + # Round 24 + movl 24(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,24(%edi) + movl %edx,%esi + # Round 28 + movl 28(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,28(%edi) + movl %edx,%esi + + addl $32,%ebx + addl $32,%edi + subl $8,%ebp + jz L004mw_finish + jmp L005mw_loop +L004mw_finish: + movl 28(%esp),%ebp + andl $7,%ebp + jnz L006mw_finish2 + jmp L007mw_end +L006mw_finish2: + # Tail Round 0 + movl (%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,(%edi) + movl %edx,%esi + decl %ebp + jz L007mw_end + # Tail Round 1 + movl 4(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,4(%edi) + movl %edx,%esi + decl %ebp + jz L007mw_end + # Tail Round 2 + movl 8(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,8(%edi) + movl %edx,%esi + decl %ebp + jz L007mw_end + # Tail Round 3 + movl 12(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,12(%edi) + movl %edx,%esi + decl %ebp + jz L007mw_end + # Tail Round 4 + movl 16(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,16(%edi) + movl %edx,%esi + decl %ebp + jz L007mw_end + # Tail Round 5 + movl 20(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,20(%edi) + movl %edx,%esi + decl %ebp + jz L007mw_end + # Tail Round 6 + movl 24(%ebx),%eax + mull %ecx + addl %esi,%eax + adcl $0,%edx + movl %eax,24(%edi) + movl %edx,%esi +L007mw_end: + movl %esi,%eax + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +.globl _bn_sqr_words +.private_extern _bn_sqr_words +.align 4 +_bn_sqr_words: +L_bn_sqr_words_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + + movl 20(%esp),%esi + movl 24(%esp),%edi + movl 28(%esp),%ebx + andl $4294967288,%ebx + jz L008sw_finish +L009sw_loop: + # Round 0 + movl (%edi),%eax + mull %eax + movl %eax,(%esi) + movl %edx,4(%esi) + # Round 4 + movl 4(%edi),%eax + mull %eax + movl %eax,8(%esi) + movl %edx,12(%esi) + # Round 8 + movl 8(%edi),%eax + mull %eax + movl %eax,16(%esi) + movl %edx,20(%esi) + # Round 12 + movl 12(%edi),%eax + mull %eax + movl %eax,24(%esi) + movl %edx,28(%esi) + # Round 16 + movl 16(%edi),%eax + mull %eax + movl %eax,32(%esi) + movl %edx,36(%esi) + # Round 20 + movl 20(%edi),%eax + mull %eax + movl %eax,40(%esi) + movl %edx,44(%esi) + # Round 24 + movl 24(%edi),%eax + mull %eax + movl %eax,48(%esi) + movl %edx,52(%esi) + # Round 28 + movl 28(%edi),%eax + mull %eax + movl %eax,56(%esi) + movl %edx,60(%esi) + + addl $32,%edi + addl $64,%esi + subl $8,%ebx + jnz L009sw_loop +L008sw_finish: + movl 28(%esp),%ebx + andl $7,%ebx + jz L010sw_end + # Tail Round 0 + movl (%edi),%eax + mull %eax + movl %eax,(%esi) + decl %ebx + movl %edx,4(%esi) + jz L010sw_end + # Tail Round 1 + movl 4(%edi),%eax + mull %eax + movl %eax,8(%esi) + decl %ebx + movl %edx,12(%esi) + jz L010sw_end + # Tail Round 2 + movl 8(%edi),%eax + mull %eax + movl %eax,16(%esi) + decl %ebx + movl %edx,20(%esi) + jz L010sw_end + # Tail Round 3 + movl 12(%edi),%eax + mull %eax + movl %eax,24(%esi) + decl %ebx + movl %edx,28(%esi) + jz L010sw_end + # Tail Round 4 + movl 16(%edi),%eax + mull %eax + movl %eax,32(%esi) + decl %ebx + movl %edx,36(%esi) + jz L010sw_end + # Tail Round 5 + movl 20(%edi),%eax + mull %eax + movl %eax,40(%esi) + decl %ebx + movl %edx,44(%esi) + jz L010sw_end + # Tail Round 6 + movl 24(%edi),%eax + mull %eax + movl %eax,48(%esi) + movl %edx,52(%esi) +L010sw_end: + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +.globl _bn_div_words +.private_extern _bn_div_words +.align 4 +_bn_div_words: +L_bn_div_words_begin: + movl 4(%esp),%edx + movl 8(%esp),%eax + movl 12(%esp),%ecx + divl %ecx + ret +.globl _bn_add_words +.private_extern _bn_add_words +.align 4 +_bn_add_words: +L_bn_add_words_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + + movl 20(%esp),%ebx + movl 24(%esp),%esi + movl 28(%esp),%edi + movl 32(%esp),%ebp + xorl %eax,%eax + andl $4294967288,%ebp + jz L011aw_finish +L012aw_loop: + # Round 0 + movl (%esi),%ecx + movl (%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + # Round 1 + movl 4(%esi),%ecx + movl 4(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,4(%ebx) + # Round 2 + movl 8(%esi),%ecx + movl 8(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,8(%ebx) + # Round 3 + movl 12(%esi),%ecx + movl 12(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,12(%ebx) + # Round 4 + movl 16(%esi),%ecx + movl 16(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,16(%ebx) + # Round 5 + movl 20(%esi),%ecx + movl 20(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,20(%ebx) + # Round 6 + movl 24(%esi),%ecx + movl 24(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,24(%ebx) + # Round 7 + movl 28(%esi),%ecx + movl 28(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,28(%ebx) + + addl $32,%esi + addl $32,%edi + addl $32,%ebx + subl $8,%ebp + jnz L012aw_loop +L011aw_finish: + movl 32(%esp),%ebp + andl $7,%ebp + jz L013aw_end + # Tail Round 0 + movl (%esi),%ecx + movl (%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,(%ebx) + jz L013aw_end + # Tail Round 1 + movl 4(%esi),%ecx + movl 4(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,4(%ebx) + jz L013aw_end + # Tail Round 2 + movl 8(%esi),%ecx + movl 8(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,8(%ebx) + jz L013aw_end + # Tail Round 3 + movl 12(%esi),%ecx + movl 12(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,12(%ebx) + jz L013aw_end + # Tail Round 4 + movl 16(%esi),%ecx + movl 16(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,16(%ebx) + jz L013aw_end + # Tail Round 5 + movl 20(%esi),%ecx + movl 20(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,20(%ebx) + jz L013aw_end + # Tail Round 6 + movl 24(%esi),%ecx + movl 24(%edi),%edx + addl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + addl %edx,%ecx + adcl $0,%eax + movl %ecx,24(%ebx) +L013aw_end: + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +.globl _bn_sub_words +.private_extern _bn_sub_words +.align 4 +_bn_sub_words: +L_bn_sub_words_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + + movl 20(%esp),%ebx + movl 24(%esp),%esi + movl 28(%esp),%edi + movl 32(%esp),%ebp + xorl %eax,%eax + andl $4294967288,%ebp + jz L014aw_finish +L015aw_loop: + # Round 0 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + # Round 1 + movl 4(%esi),%ecx + movl 4(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,4(%ebx) + # Round 2 + movl 8(%esi),%ecx + movl 8(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,8(%ebx) + # Round 3 + movl 12(%esi),%ecx + movl 12(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,12(%ebx) + # Round 4 + movl 16(%esi),%ecx + movl 16(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,16(%ebx) + # Round 5 + movl 20(%esi),%ecx + movl 20(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,20(%ebx) + # Round 6 + movl 24(%esi),%ecx + movl 24(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,24(%ebx) + # Round 7 + movl 28(%esi),%ecx + movl 28(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,28(%ebx) + + addl $32,%esi + addl $32,%edi + addl $32,%ebx + subl $8,%ebp + jnz L015aw_loop +L014aw_finish: + movl 32(%esp),%ebp + andl $7,%ebp + jz L016aw_end + # Tail Round 0 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,(%ebx) + jz L016aw_end + # Tail Round 1 + movl 4(%esi),%ecx + movl 4(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,4(%ebx) + jz L016aw_end + # Tail Round 2 + movl 8(%esi),%ecx + movl 8(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,8(%ebx) + jz L016aw_end + # Tail Round 3 + movl 12(%esi),%ecx + movl 12(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,12(%ebx) + jz L016aw_end + # Tail Round 4 + movl 16(%esi),%ecx + movl 16(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,16(%ebx) + jz L016aw_end + # Tail Round 5 + movl 20(%esi),%ecx + movl 20(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,20(%ebx) + jz L016aw_end + # Tail Round 6 + movl 24(%esi),%ecx + movl 24(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,24(%ebx) +L016aw_end: + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +.globl _bn_sub_part_words +.private_extern _bn_sub_part_words +.align 4 +_bn_sub_part_words: +L_bn_sub_part_words_begin: + pushl %ebp + pushl %ebx + pushl %esi + pushl %edi + + movl 20(%esp),%ebx + movl 24(%esp),%esi + movl 28(%esp),%edi + movl 32(%esp),%ebp + xorl %eax,%eax + andl $4294967288,%ebp + jz L017aw_finish +L018aw_loop: + # Round 0 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + # Round 1 + movl 4(%esi),%ecx + movl 4(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,4(%ebx) + # Round 2 + movl 8(%esi),%ecx + movl 8(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,8(%ebx) + # Round 3 + movl 12(%esi),%ecx + movl 12(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,12(%ebx) + # Round 4 + movl 16(%esi),%ecx + movl 16(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,16(%ebx) + # Round 5 + movl 20(%esi),%ecx + movl 20(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,20(%ebx) + # Round 6 + movl 24(%esi),%ecx + movl 24(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,24(%ebx) + # Round 7 + movl 28(%esi),%ecx + movl 28(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,28(%ebx) + + addl $32,%esi + addl $32,%edi + addl $32,%ebx + subl $8,%ebp + jnz L018aw_loop +L017aw_finish: + movl 32(%esp),%ebp + andl $7,%ebp + jz L019aw_end + # Tail Round 0 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + addl $4,%esi + addl $4,%edi + addl $4,%ebx + decl %ebp + jz L019aw_end + # Tail Round 1 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + addl $4,%esi + addl $4,%edi + addl $4,%ebx + decl %ebp + jz L019aw_end + # Tail Round 2 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + addl $4,%esi + addl $4,%edi + addl $4,%ebx + decl %ebp + jz L019aw_end + # Tail Round 3 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + addl $4,%esi + addl $4,%edi + addl $4,%ebx + decl %ebp + jz L019aw_end + # Tail Round 4 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + addl $4,%esi + addl $4,%edi + addl $4,%ebx + decl %ebp + jz L019aw_end + # Tail Round 5 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + addl $4,%esi + addl $4,%edi + addl $4,%ebx + decl %ebp + jz L019aw_end + # Tail Round 6 + movl (%esi),%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + addl $4,%esi + addl $4,%edi + addl $4,%ebx +L019aw_end: + cmpl $0,36(%esp) + je L020pw_end + movl 36(%esp),%ebp + cmpl $0,%ebp + je L020pw_end + jge L021pw_pos + # pw_neg + movl $0,%edx + subl %ebp,%edx + movl %edx,%ebp + andl $4294967288,%ebp + jz L022pw_neg_finish +L023pw_neg_loop: + # dl<0 Round 0 + movl $0,%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,(%ebx) + # dl<0 Round 1 + movl $0,%ecx + movl 4(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,4(%ebx) + # dl<0 Round 2 + movl $0,%ecx + movl 8(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,8(%ebx) + # dl<0 Round 3 + movl $0,%ecx + movl 12(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,12(%ebx) + # dl<0 Round 4 + movl $0,%ecx + movl 16(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,16(%ebx) + # dl<0 Round 5 + movl $0,%ecx + movl 20(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,20(%ebx) + # dl<0 Round 6 + movl $0,%ecx + movl 24(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,24(%ebx) + # dl<0 Round 7 + movl $0,%ecx + movl 28(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,28(%ebx) + + addl $32,%edi + addl $32,%ebx + subl $8,%ebp + jnz L023pw_neg_loop +L022pw_neg_finish: + movl 36(%esp),%edx + movl $0,%ebp + subl %edx,%ebp + andl $7,%ebp + jz L020pw_end + # dl<0 Tail Round 0 + movl $0,%ecx + movl (%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,(%ebx) + jz L020pw_end + # dl<0 Tail Round 1 + movl $0,%ecx + movl 4(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,4(%ebx) + jz L020pw_end + # dl<0 Tail Round 2 + movl $0,%ecx + movl 8(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,8(%ebx) + jz L020pw_end + # dl<0 Tail Round 3 + movl $0,%ecx + movl 12(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,12(%ebx) + jz L020pw_end + # dl<0 Tail Round 4 + movl $0,%ecx + movl 16(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,16(%ebx) + jz L020pw_end + # dl<0 Tail Round 5 + movl $0,%ecx + movl 20(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + decl %ebp + movl %ecx,20(%ebx) + jz L020pw_end + # dl<0 Tail Round 6 + movl $0,%ecx + movl 24(%edi),%edx + subl %eax,%ecx + movl $0,%eax + adcl %eax,%eax + subl %edx,%ecx + adcl $0,%eax + movl %ecx,24(%ebx) + jmp L020pw_end +L021pw_pos: + andl $4294967288,%ebp + jz L024pw_pos_finish +L025pw_pos_loop: + # dl>0 Round 0 + movl (%esi),%ecx + subl %eax,%ecx + movl %ecx,(%ebx) + jnc L026pw_nc0 + # dl>0 Round 1 + movl 4(%esi),%ecx + subl %eax,%ecx + movl %ecx,4(%ebx) + jnc L027pw_nc1 + # dl>0 Round 2 + movl 8(%esi),%ecx + subl %eax,%ecx + movl %ecx,8(%ebx) + jnc L028pw_nc2 + # dl>0 Round 3 + movl 12(%esi),%ecx + subl %eax,%ecx + movl %ecx,12(%ebx) + jnc L029pw_nc3 + # dl>0 Round 4 + movl 16(%esi),%ecx + subl %eax,%ecx + movl %ecx,16(%ebx) + jnc L030pw_nc4 + # dl>0 Round 5 + movl 20(%esi),%ecx + subl %eax,%ecx + movl %ecx,20(%ebx) + jnc L031pw_nc5 + # dl>0 Round 6 + movl 24(%esi),%ecx + subl %eax,%ecx + movl %ecx,24(%ebx) + jnc L032pw_nc6 + # dl>0 Round 7 + movl 28(%esi),%ecx + subl %eax,%ecx + movl %ecx,28(%ebx) + jnc L033pw_nc7 + + addl $32,%esi + addl $32,%ebx + subl $8,%ebp + jnz L025pw_pos_loop +L024pw_pos_finish: + movl 36(%esp),%ebp + andl $7,%ebp + jz L020pw_end + # dl>0 Tail Round 0 + movl (%esi),%ecx + subl %eax,%ecx + movl %ecx,(%ebx) + jnc L034pw_tail_nc0 + decl %ebp + jz L020pw_end + # dl>0 Tail Round 1 + movl 4(%esi),%ecx + subl %eax,%ecx + movl %ecx,4(%ebx) + jnc L035pw_tail_nc1 + decl %ebp + jz L020pw_end + # dl>0 Tail Round 2 + movl 8(%esi),%ecx + subl %eax,%ecx + movl %ecx,8(%ebx) + jnc L036pw_tail_nc2 + decl %ebp + jz L020pw_end + # dl>0 Tail Round 3 + movl 12(%esi),%ecx + subl %eax,%ecx + movl %ecx,12(%ebx) + jnc L037pw_tail_nc3 + decl %ebp + jz L020pw_end + # dl>0 Tail Round 4 + movl 16(%esi),%ecx + subl %eax,%ecx + movl %ecx,16(%ebx) + jnc L038pw_tail_nc4 + decl %ebp + jz L020pw_end + # dl>0 Tail Round 5 + movl 20(%esi),%ecx + subl %eax,%ecx + movl %ecx,20(%ebx) + jnc L039pw_tail_nc5 + decl %ebp + jz L020pw_end + # dl>0 Tail Round 6 + movl 24(%esi),%ecx + subl %eax,%ecx + movl %ecx,24(%ebx) + jnc L040pw_tail_nc6 + movl $1,%eax + jmp L020pw_end +L041pw_nc_loop: + movl (%esi),%ecx + movl %ecx,(%ebx) +L026pw_nc0: + movl 4(%esi),%ecx + movl %ecx,4(%ebx) +L027pw_nc1: + movl 8(%esi),%ecx + movl %ecx,8(%ebx) +L028pw_nc2: + movl 12(%esi),%ecx + movl %ecx,12(%ebx) +L029pw_nc3: + movl 16(%esi),%ecx + movl %ecx,16(%ebx) +L030pw_nc4: + movl 20(%esi),%ecx + movl %ecx,20(%ebx) +L031pw_nc5: + movl 24(%esi),%ecx + movl %ecx,24(%ebx) +L032pw_nc6: + movl 28(%esi),%ecx + movl %ecx,28(%ebx) +L033pw_nc7: + + addl $32,%esi + addl $32,%ebx + subl $8,%ebp + jnz L041pw_nc_loop + movl 36(%esp),%ebp + andl $7,%ebp + jz L042pw_nc_end + movl (%esi),%ecx + movl %ecx,(%ebx) +L034pw_tail_nc0: + decl %ebp + jz L042pw_nc_end + movl 4(%esi),%ecx + movl %ecx,4(%ebx) +L035pw_tail_nc1: + decl %ebp + jz L042pw_nc_end + movl 8(%esi),%ecx + movl %ecx,8(%ebx) +L036pw_tail_nc2: + decl %ebp + jz L042pw_nc_end + movl 12(%esi),%ecx + movl %ecx,12(%ebx) +L037pw_tail_nc3: + decl %ebp + jz L042pw_nc_end + movl 16(%esi),%ecx + movl %ecx,16(%ebx) +L038pw_tail_nc4: + decl %ebp + jz L042pw_nc_end + movl 20(%esi),%ecx + movl %ecx,20(%ebx) +L039pw_tail_nc5: + decl %ebp + jz L042pw_nc_end + movl 24(%esi),%ecx + movl %ecx,24(%ebx) +L040pw_tail_nc6: +L042pw_nc_end: + movl $0,%eax +L020pw_end: + popl %edi + popl %esi + popl %ebx + popl %ebp + ret +#endif |