1 ac0e2db6 2004-04-21 devnull /*
 2 ac0e2db6 2004-04-21 devnull  *	mpvecdigmulsub(mpdigit *b, int n, mpdigit m, mpdigit *p)
 3 ac0e2db6 2004-04-21 devnull  *
 4 ac0e2db6 2004-04-21 devnull  *	p -= b*m
 5 ac0e2db6 2004-04-21 devnull  *
 6 ac0e2db6 2004-04-21 devnull  *	each step look like:
 7 ac0e2db6 2004-04-21 devnull  *		hi,lo = m*b[i]
 8 ac0e2db6 2004-04-21 devnull  *		lo += oldhi + carry
 9 ac0e2db6 2004-04-21 devnull  *		hi += carry
10 ac0e2db6 2004-04-21 devnull  *		p[i] += lo
11 ac0e2db6 2004-04-21 devnull  *		oldhi = hi
12 ac0e2db6 2004-04-21 devnull  *
13 ac0e2db6 2004-04-21 devnull  *	the registers are:
14 ac0e2db6 2004-04-21 devnull  *		hi = DX		- constrained by hardware
15 ac0e2db6 2004-04-21 devnull  *		lo = AX		- constrained by hardware
16 ac0e2db6 2004-04-21 devnull  *		b = SI		- can't be BP
17 ac0e2db6 2004-04-21 devnull  *		p = DI		- can't be BP
18 ac0e2db6 2004-04-21 devnull  *		i = BP
19 ac0e2db6 2004-04-21 devnull  *		n = CX		- constrained by LOOP instr
20 ac0e2db6 2004-04-21 devnull  *		m = BX
21 ac0e2db6 2004-04-21 devnull  *		oldhi = EX
22 ac0e2db6 2004-04-21 devnull  *		
23 ac0e2db6 2004-04-21 devnull  */
24 ac0e2db6 2004-04-21 devnull .text
25 ac0e2db6 2004-04-21 devnull 
26 ac0e2db6 2004-04-21 devnull /* XXX: had to use "-4(%esp)" kludge to get around inability to
27 ac0e2db6 2004-04-21 devnull  *      push/pop without first adjusting %esp.  This may not be
28 ac0e2db6 2004-04-21 devnull  *      as fast as using push/pop (and accessing pushed element
29 ac0e2db6 2004-04-21 devnull  *      with "(%esp)".)
30 ac0e2db6 2004-04-21 devnull  */
31 ac0e2db6 2004-04-21 devnull 
32 ac0e2db6 2004-04-21 devnull .p2align 2,0x90
33 ac0e2db6 2004-04-21 devnull .globl mpvecdigmulsub
34 ac0e2db6 2004-04-21 devnull 	.type mpvecdigmulsub, @function
35 ac0e2db6 2004-04-21 devnull mpvecdigmulsub:
36 ac0e2db6 2004-04-21 devnull 	/* Prelude */
37 ac0e2db6 2004-04-21 devnull 	pushl %ebp
38 ac0e2db6 2004-04-21 devnull 	movl %ebx, -8(%esp)		/* save on stack */
39 ac0e2db6 2004-04-21 devnull 	movl %esi, -12(%esp)
40 ac0e2db6 2004-04-21 devnull 	movl %edi, -16(%esp)
41 ac0e2db6 2004-04-21 devnull 
42 ac0e2db6 2004-04-21 devnull 	movl	8(%esp), %esi		/* b */
43 ac0e2db6 2004-04-21 devnull 	movl	12(%esp), %ecx		/* n */
44 ac0e2db6 2004-04-21 devnull 	movl	16(%esp), %ebx		/* m */
45 ac0e2db6 2004-04-21 devnull 	movl	20(%esp), %edi		/* p */
46 ac0e2db6 2004-04-21 devnull 	xorl	%ebp, %ebp
47 ac0e2db6 2004-04-21 devnull 	movl	%ebp, -4(%esp)
48 ac0e2db6 2004-04-21 devnull _mulsubloop:
49 ac0e2db6 2004-04-21 devnull 	movl	(%esi, %ebp, 4),%eax	/* lo = b[i] */
50 ac0e2db6 2004-04-21 devnull 	mull	%ebx			/* hi, lo = b[i] * m */
51 ac0e2db6 2004-04-21 devnull 	addl	-4(%esp), %eax		/* lo += oldhi */
52 ac0e2db6 2004-04-21 devnull 	jae	_mulsubnocarry1
53 ac0e2db6 2004-04-21 devnull 	incl	%edx			/* hi += carry */
54 ac0e2db6 2004-04-21 devnull _mulsubnocarry1:
55 ac0e2db6 2004-04-21 devnull 	subl	%eax, (%edi, %ebp, 4)
56 ac0e2db6 2004-04-21 devnull 	jae	_mulsubnocarry2
57 ac0e2db6 2004-04-21 devnull 	incl	%edx			/* hi += carry */
58 ac0e2db6 2004-04-21 devnull _mulsubnocarry2:
59 ac0e2db6 2004-04-21 devnull 	movl	%edx, -4(%esp)
60 ac0e2db6 2004-04-21 devnull 	incl	%ebp
61 ac0e2db6 2004-04-21 devnull 	loop	_mulsubloop
62 ac0e2db6 2004-04-21 devnull 	movl	-4(%esp), %eax
63 ac0e2db6 2004-04-21 devnull 	subl	%eax, (%edi, %ebp, 4)
64 ac0e2db6 2004-04-21 devnull 	jae	_mulsubnocarry3
65 ac0e2db6 2004-04-21 devnull 	movl	$-1, %eax
66 ac0e2db6 2004-04-21 devnull 	jmp done
67 ac0e2db6 2004-04-21 devnull 
68 ac0e2db6 2004-04-21 devnull _mulsubnocarry3:
69 ac0e2db6 2004-04-21 devnull 	movl	$1, %eax
70 ac0e2db6 2004-04-21 devnull 
71 ac0e2db6 2004-04-21 devnull done:
72 ac0e2db6 2004-04-21 devnull 	/* Postlude */
73 ac0e2db6 2004-04-21 devnull 	movl -8(%esp), %ebx		/* restore from stack */
74 ac0e2db6 2004-04-21 devnull 	movl -12(%esp), %esi
75 ac0e2db6 2004-04-21 devnull 	movl -16(%esp), %edi
76 ac0e2db6 2004-04-21 devnull 	movl %esp, %ebp
77 ac0e2db6 2004-04-21 devnull 	leave
78 ac0e2db6 2004-04-21 devnull 	ret