aboutsummaryrefslogtreecommitdiffstats
path: root/arch/sparc/lib/umul.S
blob: a784720a8a220b4624365673bd810171e087d8ef (plain) (blame)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
/* $Id: umul.S,v 1.4 1996/09/30 02:22:39 davem Exp $
 * umul.S:      This routine was taken from glibc-1.09 and is covered
 *              by the GNU Library General Public License Version 2.
 */


/*
 * Unsigned multiply.  Returns %o0 * %o1 in %o1%o0 (i.e., %o1 holds the
 * upper 32 bits of the 64-bit product).
 *
 * This code optimizes short (less than 13-bit) multiplies.  Short
 * multiplies require 25 instruction cycles, and long ones require
 * 45 instruction cycles.
 *
 * On return, overflow has occurred (%o1 is not zero) if and only if
 * the Z condition code is clear, allowing, e.g., the following:
 *
 *	call	.umul
 *	nop
 *	bnz	overflow	(or tnz)
 */

	.globl .umul
.umul:
	or	%o0, %o1, %o4
	mov	%o0, %y		! multiplier -> Y

	andncc	%o4, 0xfff, %g0	! test bits 12..31 of *both* args
	be	Lmul_shortway	! if zero, can do it the short way
	 andcc	%g0, %g0, %o4	! zero the partial product and clear N and V

	/*
	 * Long multiply.  32 steps, followed by a final shift step.
	 */
	mulscc	%o4, %o1, %o4	! 1
	mulscc	%o4, %o1, %o4	! 2
	mulscc	%o4, %o1, %o4	! 3
	mulscc	%o4, %o1, %o4	! 4
	mulscc	%o4, %o1, %o4	! 5
	mulscc	%o4, %o1, %o4	! 6
	mulscc	%o4, %o1, %o4	! 7
	mulscc	%o4, %o1, %o4	! 8
	mulscc	%o4, %o1, %o4	! 9
	mulscc	%o4, %o1, %o4	! 10
	mulscc	%o4, %o1, %o4	! 11
	mulscc	%o4, %o1, %o4	! 12
	mulscc	%o4, %o1, %o4	! 13
	mulscc	%o4, %o1, %o4	! 14
	mulscc	%o4, %o1, %o4	! 15
	mulscc	%o4, %o1, %o4	! 16
	mulscc	%o4, %o1, %o4	! 17
	mulscc	%o4, %o1, %o4	! 18
	mulscc	%o4, %o1, %o4	! 19
	mulscc	%o4, %o1, %o4	! 20
	mulscc	%o4, %o1, %o4	! 21
	mulscc	%o4, %o1, %o4	! 22
	mulscc	%o4, %o1, %o4	! 23
	mulscc	%o4, %o1, %o4	! 24
	mulscc	%o4, %o1, %o4	! 25
	mulscc	%o4, %o1, %o4	! 26
	mulscc	%o4, %o1, %o4	! 27
	mulscc	%o4, %o1, %o4	! 28
	mulscc	%o4, %o1, %o4	! 29
	mulscc	%o4, %o1, %o4	! 30
	mulscc	%o4, %o1, %o4	! 31
	mulscc	%o4, %o1, %o4	! 32
	mulscc	%o4, %g0, %o4	! final shift


	/*
	 * Normally, with the shift-and-add approach, if both numbers are
	 * positive you get the correct result.  With 32-bit two's-complement
	 * numbers, -x is represented as
	 *
	 *		  x		    32
	 *	( 2  -  ------ ) mod 2  *  2
	 *		   32
	 *		  2
	 *
	 * (the `mod 2' subtracts 1 from 1.bbbb).  To avoid lots of 2^32s,
	 * we can treat this as if the radix point were just to the left
	 * of the sign bit (multiply by 2^32), and get
	 *
	 *	-x  =  (2 - x) mod 2
	 *
	 * Then, ignoring the `mod 2's for convenience:
	 *
	 *   x *  y	= xy
	 *  -x *  y	= 2y - xy
	 *   x * -y	= 2x - xy
	 *  -x * -y	= 4 - 2x - 2y + xy
	 *
	 * For signed multiplies, we subtract (x << 32) from the partial
	 * product to fix this problem for negative multipliers (see mul.s).
	 * Because of the way the shift into the partial product is calculated
	 * (N xor V), this term is automatically removed for the multiplicand,
	 * so we don't have to adjust.
	 *
	 * But for unsigned multiplies, the high order bit wasn't a sign bit,
	 * and the correction is wrong.  So for unsigned multiplies where the
	 * high order bit is one, we end up with xy - (y << 32).  To fix it
	 * we add y << 32.
	 */
#if 0
	tst	%o1
	bl,a	1f		! if %o1 < 0 (high order bit = 1),
	 add	%o4, %o0, %o4	! %o4 += %o0 (add y to upper half)

1:
	rd	%y, %o0		! get lower half of product
	retl
	 addcc	%o4, %g0, %o1	! put upper half in place and set Z for %o1==0
#else
	/* Faster code from tege@sics.se.  */
	sra	%o1, 31, %o2	! make mask from sign bit
	and	%o0, %o2, %o2	! %o2 = 0 or %o0, depending on sign of %o1
	rd	%y, %o0		! get lower half of product
	retl
	 addcc	%o4, %o2, %o1	! add compensation and put upper half in place
#endif

Lmul_shortway:
	/*
	 * Short multiply.  12 steps, followed by a final shift step.
	 * The resulting bits are off by 12 and (32-12) = 20 bit positions,
	 * but there is no problem with %o0 being negative (unlike above),
	 * and overflow is impossible (the answer is at most 24 bits long).
	 */
	mulscc	%o4, %o1, %o4	! 1
	mulscc	%o4, %o1, %o4	! 2
	mulscc	%o4, %o1, %o4	! 3
	mulscc	%o4, %o1, %o4	! 4
	mulscc	%o4, %o1, %o4	! 5
	mulscc	%o4, %o1, %o4	! 6
	mulscc	%o4, %o1, %o4	! 7
	mulscc	%o4, %o1, %o4	! 8
	mulscc	%o4, %o1, %o4	! 9
	mulscc	%o4, %o1, %o4	! 10
	mulscc	%o4, %o1, %o4	! 11
	mulscc	%o4, %o1, %o4	! 12
	mulscc	%o4, %g0, %o4	! final shift

	/*
	 * %o4 has 20 of the bits that should be in the result; %y has
	 * the bottom 12 (as %y's top 12).  That is:
	 *
	 *	  %o4		    %y
	 * +----------------+----------------+
	 * | -12- |   -20-  | -12- |   -20-  |
	 * +------(---------+------)---------+
	 *	   -----result-----
	 *
	 * The 12 bits of %o4 left of the `result' area are all zero;
	 * in fact, all top 20 bits of %o4 are zero.
	 */

	rd	%y, %o5
	sll	%o4, 12, %o0	! shift middle bits left 12
	srl	%o5, 20, %o5	! shift low bits right 20
	or	%o5, %o0, %o0
	retl
	 addcc	%g0, %g0, %o1	! %o1 = zero, and set Z

	.globl	.umul_patch
.umul_patch:
	umul	%o0, %o1, %o0
	retl
	 rd	%y, %o1
	nop