diff options
Diffstat (limited to 'crypto/bn/asm/co86unix.cpp')
| -rw-r--r-- | crypto/bn/asm/co86unix.cpp | 1315 |
1 files changed, 1315 insertions, 0 deletions
diff --git a/crypto/bn/asm/co86unix.cpp b/crypto/bn/asm/co86unix.cpp new file mode 100644 index 0000000000..fa80b14046 --- /dev/null +++ b/crypto/bn/asm/co86unix.cpp @@ -0,0 +1,1315 @@ +/* Run the C pre-processor over this file with one of the following defined + * ELF - elf object files, + * OUT - a.out object files, + * BSDI - BSDI style a.out object files + * SOL - Solaris style elf + */ + +#define TYPE(a,b) .type a,b +#define SIZE(a,b) .size a,b + +#if defined(OUT) || defined(BSDI) +#define bn_mul_comba8 _bn_mul_comba8 +#define bn_mul_comba4 _bn_mul_comba4 +#define bn_sqr_comba8 _bn_sqr_comba8 +#define bn_sqr_comba4 _bn_sqr_comba4 + +#endif + +#ifdef OUT +#define OK 1 +#define ALIGN 4 +#endif + +#ifdef BSDI +#define OK 1 +#define ALIGN 4 +#undef SIZE +#undef TYPE +#define SIZE(a,b) +#define TYPE(a,b) +#endif + +#if defined(ELF) || defined(SOL) +#define OK 1 +#define ALIGN 16 +#endif + +#ifndef OK +You need to define one of +ELF - elf systems - linux-elf, NetBSD and DG-UX +OUT - a.out systems - linux-a.out and FreeBSD +SOL - solaris systems, which are elf with strange comment lines +BSDI - a.out with a very primative version of as. +#endif + +/* Let the Assembler begin :-) */ + /* Don't even think of reading this code */ + /* It was automatically generated by bn-586.pl */ + /* Which is a perl program used to generate the x86 assember for */ + /* any of elf, a.out, BSDI,Win32, or Solaris */ + /* eric <eay@cryptsoft.com> */ + + .file "bn-586.s" + .version "01.01" +gcc2_compiled.: +.text + .align ALIGN +.globl bn_mul_comba8 + TYPE(bn_mul_comba8,@function) +bn_mul_comba8: + pushl %esi + movl 12(%esp), %esi + pushl %edi + movl 20(%esp), %edi + pushl %ebp + pushl %ebx + xorl %ebx, %ebx + movl (%esi), %eax + xorl %ecx, %ecx + movl (%edi), %edx + /* ################## Calculate word 0 */ + xorl %ebp, %ebp + /* mul a[0]*b[0] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + movl (%edi), %edx + adcl $0, %ebp + movl %ebx, (%eax) + movl 4(%esi), %eax + /* saved r[0] */ + /* ################## Calculate word 1 */ + xorl %ebx, %ebx + /* mul a[1]*b[0] */ + mull %edx + addl %eax, %ecx + movl (%esi), %eax + adcl %edx, %ebp + movl 4(%edi), %edx + adcl $0, %ebx + /* mul a[0]*b[1] */ + mull %edx + addl %eax, %ecx + movl 20(%esp), %eax + adcl %edx, %ebp + movl (%edi), %edx + adcl $0, %ebx + movl %ecx, 4(%eax) + movl 8(%esi), %eax + /* saved r[1] */ + /* ################## Calculate word 2 */ + xorl %ecx, %ecx + /* mul a[2]*b[0] */ + mull %edx + addl %eax, %ebp + movl 4(%esi), %eax + adcl %edx, %ebx + movl 4(%edi), %edx + adcl $0, %ecx + /* mul a[1]*b[1] */ + mull %edx + addl %eax, %ebp + movl (%esi), %eax + adcl %edx, %ebx + movl 8(%edi), %edx + adcl $0, %ecx + /* mul a[0]*b[2] */ + mull %edx + addl %eax, %ebp + movl 20(%esp), %eax + adcl %edx, %ebx + movl (%edi), %edx + adcl $0, %ecx + movl %ebp, 8(%eax) + movl 12(%esi), %eax + /* saved r[2] */ + /* ################## Calculate word 3 */ + xorl %ebp, %ebp + /* mul a[3]*b[0] */ + mull %edx + addl %eax, %ebx + movl 8(%esi), %eax + adcl %edx, %ecx + movl 4(%edi), %edx + adcl $0, %ebp + /* mul a[2]*b[1] */ + mull %edx + addl %eax, %ebx + movl 4(%esi), %eax + adcl %edx, %ecx + movl 8(%edi), %edx + adcl $0, %ebp + /* mul a[1]*b[2] */ + mull %edx + addl %eax, %ebx + movl (%esi), %eax + adcl %edx, %ecx + movl 12(%edi), %edx + adcl $0, %ebp + /* mul a[0]*b[3] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + movl (%edi), %edx + adcl $0, %ebp + movl %ebx, 12(%eax) + movl 16(%esi), %eax + /* saved r[3] */ + /* ################## Calculate word 4 */ + xorl %ebx, %ebx + /* mul a[4]*b[0] */ + mull %edx + addl %eax, %ecx + movl 12(%esi), %eax + adcl %edx, %ebp + movl 4(%edi), %edx + adcl $0, %ebx + /* mul a[3]*b[1] */ + mull %edx + addl %eax, %ecx + movl 8(%esi), %eax + adcl %edx, %ebp + movl 8(%edi), %edx + adcl $0, %ebx + /* mul a[2]*b[2] */ + mull %edx + addl %eax, %ecx + movl 4(%esi), %eax + adcl %edx, %ebp + movl 12(%edi), %edx + adcl $0, %ebx + /* mul a[1]*b[3] */ + mull %edx + addl %eax, %ecx + movl (%esi), %eax + adcl %edx, %ebp + movl 16(%edi), %edx + adcl $0, %ebx + /* mul a[0]*b[4] */ + mull %edx + addl %eax, %ecx + movl 20(%esp), %eax + adcl %edx, %ebp + movl (%edi), %edx + adcl $0, %ebx + movl %ecx, 16(%eax) + movl 20(%esi), %eax + /* saved r[4] */ + /* ################## Calculate word 5 */ + xorl %ecx, %ecx + /* mul a[5]*b[0] */ + mull %edx + addl %eax, %ebp + movl 16(%esi), %eax + adcl %edx, %ebx + movl 4(%edi), %edx + adcl $0, %ecx + /* mul a[4]*b[1] */ + mull %edx + addl %eax, %ebp + movl 12(%esi), %eax + adcl %edx, %ebx + movl 8(%edi), %edx + adcl $0, %ecx + /* mul a[3]*b[2] */ + mull %edx + addl %eax, %ebp + movl 8(%esi), %eax + adcl %edx, %ebx + movl 12(%edi), %edx + adcl $0, %ecx + /* mul a[2]*b[3] */ + mull %edx + addl %eax, %ebp + movl 4(%esi), %eax + adcl %edx, %ebx + movl 16(%edi), %edx + adcl $0, %ecx + /* mul a[1]*b[4] */ + mull %edx + addl %eax, %ebp + movl (%esi), %eax + adcl %edx, %ebx + movl 20(%edi), %edx + adcl $0, %ecx + /* mul a[0]*b[5] */ + mull %edx + addl %eax, %ebp + movl 20(%esp), %eax + adcl %edx, %ebx + movl (%edi), %edx + adcl $0, %ecx + movl %ebp, 20(%eax) + movl 24(%esi), %eax + /* saved r[5] */ + /* ################## Calculate word 6 */ + xorl %ebp, %ebp + /* mul a[6]*b[0] */ + mull %edx + addl %eax, %ebx + movl 20(%esi), %eax + adcl %edx, %ecx + movl 4(%edi), %edx + adcl $0, %ebp + /* mul a[5]*b[1] */ + mull %edx + addl %eax, %ebx + movl 16(%esi), %eax + adcl %edx, %ecx + movl 8(%edi), %edx + adcl $0, %ebp + /* mul a[4]*b[2] */ + mull %edx + addl %eax, %ebx + movl 12(%esi), %eax + adcl %edx, %ecx + movl 12(%edi), %edx + adcl $0, %ebp + /* mul a[3]*b[3] */ + mull %edx + addl %eax, %ebx + movl 8(%esi), %eax + adcl %edx, %ecx + movl 16(%edi), %edx + adcl $0, %ebp + /* mul a[2]*b[4] */ + mull %edx + addl %eax, %ebx + movl 4(%esi), %eax + adcl %edx, %ecx + movl 20(%edi), %edx + adcl $0, %ebp + /* mul a[1]*b[5] */ + mull %edx + addl %eax, %ebx + movl (%esi), %eax + adcl %edx, %ecx + movl 24(%edi), %edx + adcl $0, %ebp + /* mul a[0]*b[6] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + movl (%edi), %edx + adcl $0, %ebp + movl %ebx, 24(%eax) + movl 28(%esi), %eax + /* saved r[6] */ + /* ################## Calculate word 7 */ + xorl %ebx, %ebx + /* mul a[7]*b[0] */ + mull %edx + addl %eax, %ecx + movl 24(%esi), %eax + adcl %edx, %ebp + movl 4(%edi), %edx + adcl $0, %ebx + /* mul a[6]*b[1] */ + mull %edx + addl %eax, %ecx + movl 20(%esi), %eax + adcl %edx, %ebp + movl 8(%edi), %edx + adcl $0, %ebx + /* mul a[5]*b[2] */ + mull %edx + addl %eax, %ecx + movl 16(%esi), %eax + adcl %edx, %ebp + movl 12(%edi), %edx + adcl $0, %ebx + /* mul a[4]*b[3] */ + mull %edx + addl %eax, %ecx + movl 12(%esi), %eax + adcl %edx, %ebp + movl 16(%edi), %edx + adcl $0, %ebx + /* mul a[3]*b[4] */ + mull %edx + addl %eax, %ecx + movl 8(%esi), %eax + adcl %edx, %ebp + movl 20(%edi), %edx + adcl $0, %ebx + /* mul a[2]*b[5] */ + mull %edx + addl %eax, %ecx + movl 4(%esi), %eax + adcl %edx, %ebp + movl 24(%edi), %edx + adcl $0, %ebx + /* mul a[1]*b[6] */ + mull %edx + addl %eax, %ecx + movl (%esi), %eax + adcl %edx, %ebp + movl 28(%edi), %edx + adcl $0, %ebx + /* mul a[0]*b[7] */ + mull %edx + addl %eax, %ecx + movl 20(%esp), %eax + adcl %edx, %ebp + movl 4(%edi), %edx + adcl $0, %ebx + movl %ecx, 28(%eax) + movl 28(%esi), %eax + /* saved r[7] */ + /* ################## Calculate word 8 */ + xorl %ecx, %ecx + /* mul a[7]*b[1] */ + mull %edx + addl %eax, %ebp + movl 24(%esi), %eax + adcl %edx, %ebx + movl 8(%edi), %edx + adcl $0, %ecx + /* mul a[6]*b[2] */ + mull %edx + addl %eax, %ebp + movl 20(%esi), %eax + adcl %edx, %ebx + movl 12(%edi), %edx + adcl $0, %ecx + /* mul a[5]*b[3] */ + mull %edx + addl %eax, %ebp + movl 16(%esi), %eax + adcl %edx, %ebx + movl 16(%edi), %edx + adcl $0, %ecx + /* mul a[4]*b[4] */ + mull %edx + addl %eax, %ebp + movl 12(%esi), %eax + adcl %edx, %ebx + movl 20(%edi), %edx + adcl $0, %ecx + /* mul a[3]*b[5] */ + mull %edx + addl %eax, %ebp + movl 8(%esi), %eax + adcl %edx, %ebx + movl 24(%edi), %edx + adcl $0, %ecx + /* mul a[2]*b[6] */ + mull %edx + addl %eax, %ebp + movl 4(%esi), %eax + adcl %edx, %ebx + movl 28(%edi), %edx + adcl $0, %ecx + /* mul a[1]*b[7] */ + mull %edx + addl %eax, %ebp + movl 20(%esp), %eax + adcl %edx, %ebx + movl 8(%edi), %edx + adcl $0, %ecx + movl %ebp, 32(%eax) + movl 28(%esi), %eax + /* saved r[8] */ + /* ################## Calculate word 9 */ + xorl %ebp, %ebp + /* mul a[7]*b[2] */ + mull %edx + addl %eax, %ebx + movl 24(%esi), %eax + adcl %edx, %ecx + movl 12(%edi), %edx + adcl $0, %ebp + /* mul a[6]*b[3] */ + mull %edx + addl %eax, %ebx + movl 20(%esi), %eax + adcl %edx, %ecx + movl 16(%edi), %edx + adcl $0, %ebp + /* mul a[5]*b[4] */ + mull %edx + addl %eax, %ebx + movl 16(%esi), %eax + adcl %edx, %ecx + movl 20(%edi), %edx + adcl $0, %ebp + /* mul a[4]*b[5] */ + mull %edx + addl %eax, %ebx + movl 12(%esi), %eax + adcl %edx, %ecx + movl 24(%edi), %edx + adcl $0, %ebp + /* mul a[3]*b[6] */ + mull %edx + addl %eax, %ebx + movl 8(%esi), %eax + adcl %edx, %ecx + movl 28(%edi), %edx + adcl $0, %ebp + /* mul a[2]*b[7] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + movl 12(%edi), %edx + adcl $0, %ebp + movl %ebx, 36(%eax) + movl 28(%esi), %eax + /* saved r[9] */ + /* ################## Calculate word 10 */ + xorl %ebx, %ebx + /* mul a[7]*b[3] */ + mull %edx + addl %eax, %ecx + movl 24(%esi), %eax + adcl %edx, %ebp + movl 16(%edi), %edx + adcl $0, %ebx + /* mul a[6]*b[4] */ + mull %edx + addl %eax, %ecx + movl 20(%esi), %eax + adcl %edx, %ebp + movl 20(%edi), %edx + adcl $0, %ebx + /* mul a[5]*b[5] */ + mull %edx + addl %eax, %ecx + movl 16(%esi), %eax + adcl %edx, %ebp + movl 24(%edi), %edx + adcl $0, %ebx + /* mul a[4]*b[6] */ + mull %edx + addl %eax, %ecx + movl 12(%esi), %eax + adcl %edx, %ebp + movl 28(%edi), %edx + adcl $0, %ebx + /* mul a[3]*b[7] */ + mull %edx + addl %eax, %ecx + movl 20(%esp), %eax + adcl %edx, %ebp + movl 16(%edi), %edx + adcl $0, %ebx + movl %ecx, 40(%eax) + movl 28(%esi), %eax + /* saved r[10] */ + /* ################## Calculate word 11 */ + xorl %ecx, %ecx + /* mul a[7]*b[4] */ + mull %edx + addl %eax, %ebp + movl 24(%esi), %eax + adcl %edx, %ebx + movl 20(%edi), %edx + adcl $0, %ecx + /* mul a[6]*b[5] */ + mull %edx + addl %eax, %ebp + movl 20(%esi), %eax + adcl %edx, %ebx + movl 24(%edi), %edx + adcl $0, %ecx + /* mul a[5]*b[6] */ + mull %edx + addl %eax, %ebp + movl 16(%esi), %eax + adcl %edx, %ebx + movl 28(%edi), %edx + adcl $0, %ecx + /* mul a[4]*b[7] */ + mull %edx + addl %eax, %ebp + movl 20(%esp), %eax + adcl %edx, %ebx + movl 20(%edi), %edx + adcl $0, %ecx + movl %ebp, 44(%eax) + movl 28(%esi), %eax + /* saved r[11] */ + /* ################## Calculate word 12 */ + xorl %ebp, %ebp + /* mul a[7]*b[5] */ + mull %edx + addl %eax, %ebx + movl 24(%esi), %eax + adcl %edx, %ecx + movl 24(%edi), %edx + adcl $0, %ebp + /* mul a[6]*b[6] */ + mull %edx + addl %eax, %ebx + movl 20(%esi), %eax + adcl %edx, %ecx + movl 28(%edi), %edx + adcl $0, %ebp + /* mul a[5]*b[7] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + movl 24(%edi), %edx + adcl $0, %ebp + movl %ebx, 48(%eax) + movl 28(%esi), %eax + /* saved r[12] */ + /* ################## Calculate word 13 */ + xorl %ebx, %ebx + /* mul a[7]*b[6] */ + mull %edx + addl %eax, %ecx + movl 24(%esi), %eax + adcl %edx, %ebp + movl 28(%edi), %edx + adcl $0, %ebx + /* mul a[6]*b[7] */ + mull %edx + addl %eax, %ecx + movl 20(%esp), %eax + adcl %edx, %ebp + movl 28(%edi), %edx + adcl $0, %ebx + movl %ecx, 52(%eax) + movl 28(%esi), %eax + /* saved r[13] */ + /* ################## Calculate word 14 */ + xorl %ecx, %ecx + /* mul a[7]*b[7] */ + mull %edx + addl %eax, %ebp + movl 20(%esp), %eax + adcl %edx, %ebx + adcl $0, %ecx + movl %ebp, 56(%eax) + /* saved r[14] */ + /* save r[15] */ + movl %ebx, 60(%eax) + popl %ebx + popl %ebp + popl %edi + popl %esi + ret +.bn_mul_comba8_end: + SIZE(bn_mul_comba8,.bn_mul_comba8_end-bn_mul_comba8) +.ident "desasm.pl" +.text + .align ALIGN +.globl bn_mul_comba4 + TYPE(bn_mul_comba4,@function) +bn_mul_comba4: + pushl %esi + movl 12(%esp), %esi + pushl %edi + movl 20(%esp), %edi + pushl %ebp + pushl %ebx + xorl %ebx, %ebx + movl (%esi), %eax + xorl %ecx, %ecx + movl (%edi), %edx + /* ################## Calculate word 0 */ + xorl %ebp, %ebp + /* mul a[0]*b[0] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + movl (%edi), %edx + adcl $0, %ebp + movl %ebx, (%eax) + movl 4(%esi), %eax + /* saved r[0] */ + /* ################## Calculate word 1 */ + xorl %ebx, %ebx + /* mul a[1]*b[0] */ + mull %edx + addl %eax, %ecx + movl (%esi), %eax + adcl %edx, %ebp + movl 4(%edi), %edx + adcl $0, %ebx + /* mul a[0]*b[1] */ + mull %edx + addl %eax, %ecx + movl 20(%esp), %eax + adcl %edx, %ebp + movl (%edi), %edx + adcl $0, %ebx + movl %ecx, 4(%eax) + movl 8(%esi), %eax + /* saved r[1] */ + /* ################## Calculate word 2 */ + xorl %ecx, %ecx + /* mul a[2]*b[0] */ + mull %edx + addl %eax, %ebp + movl 4(%esi), %eax + adcl %edx, %ebx + movl 4(%edi), %edx + adcl $0, %ecx + /* mul a[1]*b[1] */ + mull %edx + addl %eax, %ebp + movl (%esi), %eax + adcl %edx, %ebx + movl 8(%edi), %edx + adcl $0, %ecx + /* mul a[0]*b[2] */ + mull %edx + addl %eax, %ebp + movl 20(%esp), %eax + adcl %edx, %ebx + movl (%edi), %edx + adcl $0, %ecx + movl %ebp, 8(%eax) + movl 12(%esi), %eax + /* saved r[2] */ + /* ################## Calculate word 3 */ + xorl %ebp, %ebp + /* mul a[3]*b[0] */ + mull %edx + addl %eax, %ebx + movl 8(%esi), %eax + adcl %edx, %ecx + movl 4(%edi), %edx + adcl $0, %ebp + /* mul a[2]*b[1] */ + mull %edx + addl %eax, %ebx + movl 4(%esi), %eax + adcl %edx, %ecx + movl 8(%edi), %edx + adcl $0, %ebp + /* mul a[1]*b[2] */ + mull %edx + addl %eax, %ebx + movl (%esi), %eax + adcl %edx, %ecx + movl 12(%edi), %edx + adcl $0, %ebp + /* mul a[0]*b[3] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + movl 4(%edi), %edx + adcl $0, %ebp + movl %ebx, 12(%eax) + movl 12(%esi), %eax + /* saved r[3] */ + /* ################## Calculate word 4 */ + xorl %ebx, %ebx + /* mul a[3]*b[1] */ + mull %edx + addl %eax, %ecx + movl 8(%esi), %eax + adcl %edx, %ebp + movl 8(%edi), %edx + adcl $0, %ebx + /* mul a[2]*b[2] */ + mull %edx + addl %eax, %ecx + movl 4(%esi), %eax + adcl %edx, %ebp + movl 12(%edi), %edx + adcl $0, %ebx + /* mul a[1]*b[3] */ + mull %edx + addl %eax, %ecx + movl 20(%esp), %eax + adcl %edx, %ebp + movl 8(%edi), %edx + adcl $0, %ebx + movl %ecx, 16(%eax) + movl 12(%esi), %eax + /* saved r[4] */ + /* ################## Calculate word 5 */ + xorl %ecx, %ecx + /* mul a[3]*b[2] */ + mull %edx + addl %eax, %ebp + movl 8(%esi), %eax + adcl %edx, %ebx + movl 12(%edi), %edx + adcl $0, %ecx + /* mul a[2]*b[3] */ + mull %edx + addl %eax, %ebp + movl 20(%esp), %eax + adcl %edx, %ebx + movl 12(%edi), %edx + adcl $0, %ecx + movl %ebp, 20(%eax) + movl 12(%esi), %eax + /* saved r[5] */ + /* ################## Calculate word 6 */ + xorl %ebp, %ebp + /* mul a[3]*b[3] */ + mull %edx + addl %eax, %ebx + movl 20(%esp), %eax + adcl %edx, %ecx + adcl $0, %ebp + movl %ebx, 24(%eax) + /* saved r[6] */ + /* save r[7] */ + movl %ecx, 28(%eax) + popl %ebx + popl %ebp + popl %edi + popl %esi + ret +.bn_mul_comba4_end: + SIZE(bn_mul_comba4,.bn_mul_comba4_end-bn_mul_comba4) +.ident "desasm.pl" +.text + .align ALIGN +.globl bn_sqr_comba8 + TYPE(bn_sqr_comba8,@function) +bn_sqr_comba8: + pushl %esi + pushl %edi + pushl %ebp + pushl %ebx + movl 20(%esp), %edi + movl 24(%esp), %esi + xorl %ebx, %ebx + xorl %ecx, %ecx + movl (%esi), %eax + /* ############### Calculate word 0 */ + xorl %ebp, %ebp + /* sqr a[0]*a[0] */ + mull %eax + addl %eax, %ebx + adcl %edx, %ecx + movl (%esi), %edx + adcl $0, %ebp + movl %ebx, (%edi) + movl 4(%esi), %eax + /* saved r[0] */ + /* ############### Calculate word 1 */ + xorl %ebx, %ebx + /* sqr a[1]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 8(%esi), %eax + adcl $0, %ebx + movl %ecx, 4(%edi) + movl (%esi), %edx + /* saved r[1] */ + /* ############### Calculate word 2 */ + xorl %ecx, %ecx + /* sqr a[2]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 4(%esi), %eax + adcl $0, %ecx + /* sqr a[1]*a[1] */ + mull %eax + addl %eax, %ebp + adcl %edx, %ebx + movl (%esi), %edx + adcl $0, %ecx + movl %ebp, 8(%edi) + movl 12(%esi), %eax + /* saved r[2] */ + /* ############### Calculate word 3 */ + xorl %ebp, %ebp + /* sqr a[3]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 8(%esi), %eax + adcl $0, %ebp + movl 4(%esi), %edx + /* sqr a[2]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 16(%esi), %eax + adcl $0, %ebp + movl %ebx, 12(%edi) + movl (%esi), %edx + /* saved r[3] */ + /* ############### Calculate word 4 */ + xorl %ebx, %ebx + /* sqr a[4]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 12(%esi), %eax + adcl $0, %ebx + movl 4(%esi), %edx + /* sqr a[3]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 8(%esi), %eax + adcl $0, %ebx + /* sqr a[2]*a[2] */ + mull %eax + addl %eax, %ecx + adcl %edx, %ebp + movl (%esi), %edx + adcl $0, %ebx + movl %ecx, 16(%edi) + movl 20(%esi), %eax + /* saved r[4] */ + /* ############### Calculate word 5 */ + xorl %ecx, %ecx + /* sqr a[5]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 16(%esi), %eax + adcl $0, %ecx + movl 4(%esi), %edx + /* sqr a[4]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 12(%esi), %eax + adcl $0, %ecx + movl 8(%esi), %edx + /* sqr a[3]*a[2] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 24(%esi), %eax + adcl $0, %ecx + movl %ebp, 20(%edi) + movl (%esi), %edx + /* saved r[5] */ + /* ############### Calculate word 6 */ + xorl %ebp, %ebp + /* sqr a[6]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 20(%esi), %eax + adcl $0, %ebp + movl 4(%esi), %edx + /* sqr a[5]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 16(%esi), %eax + adcl $0, %ebp + movl 8(%esi), %edx + /* sqr a[4]*a[2] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 12(%esi), %eax + adcl $0, %ebp + /* sqr a[3]*a[3] */ + mull %eax + addl %eax, %ebx + adcl %edx, %ecx + movl (%esi), %edx + adcl $0, %ebp + movl %ebx, 24(%edi) + movl 28(%esi), %eax + /* saved r[6] */ + /* ############### Calculate word 7 */ + xorl %ebx, %ebx + /* sqr a[7]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 24(%esi), %eax + adcl $0, %ebx + movl 4(%esi), %edx + /* sqr a[6]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 20(%esi), %eax + adcl $0, %ebx + movl 8(%esi), %edx + /* sqr a[5]*a[2] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 16(%esi), %eax + adcl $0, %ebx + movl 12(%esi), %edx + /* sqr a[4]*a[3] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 28(%esi), %eax + adcl $0, %ebx + movl %ecx, 28(%edi) + movl 4(%esi), %edx + /* saved r[7] */ + /* ############### Calculate word 8 */ + xorl %ecx, %ecx + /* sqr a[7]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 24(%esi), %eax + adcl $0, %ecx + movl 8(%esi), %edx + /* sqr a[6]*a[2] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 20(%esi), %eax + adcl $0, %ecx + movl 12(%esi), %edx + /* sqr a[5]*a[3] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 16(%esi), %eax + adcl $0, %ecx + /* sqr a[4]*a[4] */ + mull %eax + addl %eax, %ebp + adcl %edx, %ebx + movl 8(%esi), %edx + adcl $0, %ecx + movl %ebp, 32(%edi) + movl 28(%esi), %eax + /* saved r[8] */ + /* ############### Calculate word 9 */ + xorl %ebp, %ebp + /* sqr a[7]*a[2] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 24(%esi), %eax + adcl $0, %ebp + movl 12(%esi), %edx + /* sqr a[6]*a[3] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 20(%esi), %eax + adcl $0, %ebp + movl 16(%esi), %edx + /* sqr a[5]*a[4] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 28(%esi), %eax + adcl $0, %ebp + movl %ebx, 36(%edi) + movl 12(%esi), %edx + /* saved r[9] */ + /* ############### Calculate word 10 */ + xorl %ebx, %ebx + /* sqr a[7]*a[3] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 24(%esi), %eax + adcl $0, %ebx + movl 16(%esi), %edx + /* sqr a[6]*a[4] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 20(%esi), %eax + adcl $0, %ebx + /* sqr a[5]*a[5] */ + mull %eax + addl %eax, %ecx + adcl %edx, %ebp + movl 16(%esi), %edx + adcl $0, %ebx + movl %ecx, 40(%edi) + movl 28(%esi), %eax + /* saved r[10] */ + /* ############### Calculate word 11 */ + xorl %ecx, %ecx + /* sqr a[7]*a[4] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 24(%esi), %eax + adcl $0, %ecx + movl 20(%esi), %edx + /* sqr a[6]*a[5] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 28(%esi), %eax + adcl $0, %ecx + movl %ebp, 44(%edi) + movl 20(%esi), %edx + /* saved r[11] */ + /* ############### Calculate word 12 */ + xorl %ebp, %ebp + /* sqr a[7]*a[5] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 24(%esi), %eax + adcl $0, %ebp + /* sqr a[6]*a[6] */ + mull %eax + addl %eax, %ebx + adcl %edx, %ecx + movl 24(%esi), %edx + adcl $0, %ebp + movl %ebx, 48(%edi) + movl 28(%esi), %eax + /* saved r[12] */ + /* ############### Calculate word 13 */ + xorl %ebx, %ebx + /* sqr a[7]*a[6] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 28(%esi), %eax + adcl $0, %ebx + movl %ecx, 52(%edi) + /* saved r[13] */ + /* ############### Calculate word 14 */ + xorl %ecx, %ecx + /* sqr a[7]*a[7] */ + mull %eax + addl %eax, %ebp + adcl %edx, %ebx + adcl $0, %ecx + movl %ebp, 56(%edi) + /* saved r[14] */ + movl %ebx, 60(%edi) + popl %ebx + popl %ebp + popl %edi + popl %esi + ret +.bn_sqr_comba8_end: + SIZE(bn_sqr_comba8,.bn_sqr_comba8_end-bn_sqr_comba8) +.ident "desasm.pl" +.text + .align ALIGN +.globl bn_sqr_comba4 + TYPE(bn_sqr_comba4,@function) +bn_sqr_comba4: + pushl %esi + pushl %edi + pushl %ebp + pushl %ebx + movl 20(%esp), %edi + movl 24(%esp), %esi + xorl %ebx, %ebx + xorl %ecx, %ecx + movl (%esi), %eax + /* ############### Calculate word 0 */ + xorl %ebp, %ebp + /* sqr a[0]*a[0] */ + mull %eax + addl %eax, %ebx + adcl %edx, %ecx + movl (%esi), %edx + adcl $0, %ebp + movl %ebx, (%edi) + movl 4(%esi), %eax + /* saved r[0] */ + /* ############### Calculate word 1 */ + xorl %ebx, %ebx + /* sqr a[1]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 8(%esi), %eax + adcl $0, %ebx + movl %ecx, 4(%edi) + movl (%esi), %edx + /* saved r[1] */ + /* ############### Calculate word 2 */ + xorl %ecx, %ecx + /* sqr a[2]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 4(%esi), %eax + adcl $0, %ecx + /* sqr a[1]*a[1] */ + mull %eax + addl %eax, %ebp + adcl %edx, %ebx + movl (%esi), %edx + adcl $0, %ecx + movl %ebp, 8(%edi) + movl 12(%esi), %eax + /* saved r[2] */ + /* ############### Calculate word 3 */ + xorl %ebp, %ebp + /* sqr a[3]*a[0] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 8(%esi), %eax + adcl $0, %ebp + movl 4(%esi), %edx + /* sqr a[2]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebp + addl %eax, %ebx + adcl %edx, %ecx + movl 12(%esi), %eax + adcl $0, %ebp + movl %ebx, 12(%edi) + movl 4(%esi), %edx + /* saved r[3] */ + /* ############### Calculate word 4 */ + xorl %ebx, %ebx + /* sqr a[3]*a[1] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ebx + addl %eax, %ecx + adcl %edx, %ebp + movl 8(%esi), %eax + adcl $0, %ebx + /* sqr a[2]*a[2] */ + mull %eax + addl %eax, %ecx + adcl %edx, %ebp + movl 8(%esi), %edx + adcl $0, %ebx + movl %ecx, 16(%edi) + movl 12(%esi), %eax + /* saved r[4] */ + /* ############### Calculate word 5 */ + xorl %ecx, %ecx + /* sqr a[3]*a[2] */ + mull %edx + addl %eax, %eax + adcl %edx, %edx + adcl $0, %ecx + addl %eax, %ebp + adcl %edx, %ebx + movl 12(%esi), %eax + adcl $0, %ecx + movl %ebp, 20(%edi) + /* saved r[5] */ + /* ############### Calculate word 6 */ + xorl %ebp, %ebp + /* sqr a[3]*a[3] */ + mull %eax + addl %eax, %ebx + adcl %edx, %ecx + adcl $0, %ebp + movl %ebx, 24(%edi) + /* saved r[6] */ + movl %ecx, 28(%edi) + popl %ebx + popl %ebp + popl %edi + popl %esi + ret +.bn_sqr_comba4_end: + SIZE(bn_sqr_comba4,.bn_sqr_comba4_end-bn_sqr_comba4) +.ident "desasm.pl" |