summaryrefslogtreecommitdiff
path: root/sysdeps/sparc/divrem.m4
blob: 665abf11ae348ad1443aaac941157b5ceab69be9 (plain)
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
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
/*
 * Division and remainder, from Appendix E of the Sparc Version 8
 * Architecture Manual, with fixes from Gordon Irlam.
 */

/*
 * Input: dividend and divisor in %o0 and %o1 respectively.
 *
 * m4 parameters:
 *  NAME	name of function to generate
 *  OP		OP=div => %o0 / %o1; OP=rem => %o0 % %o1
 *  S		S=true => signed; S=false => unsigned
 *
 * Algorithm parameters:
 *  N		how many bits per iteration we try to get (4)
 *  WORDSIZE	total number of bits (32)
 *
 * Derived constants:
 *  TOPBITS	number of bits in the top `decade' of a number
 *
 * Important variables:
 *  Q		the partial quotient under development (initially 0)
 *  R		the remainder so far, initially the dividend
 *  ITER	number of main division loop iterations required;
 *		equal to ceil(log2(quotient) / N).  Note that this
 *		is the log base (2^N) of the quotient.
 *  V		the current comparand, initially divisor*2^(ITER*N-1)
 *
 * Cost:
 *  Current estimate for non-large dividend is
 *	ceil(log2(quotient) / N) * (10 + 7N/2) + C
 *  A large dividend is one greater than 2^(31-TOPBITS) and takes a
 *  different path, as the upper bits of the quotient must be developed
 *  one bit at a time.
 */

define(N, `4')dnl
define(WORDSIZE, `32')dnl
define(TOPBITS, eval(WORDSIZE - N*((WORDSIZE-1)/N)))dnl
dnl
define(dividend, `%o0')dnl
define(divisor, `%o1')dnl
define(Q, `%o2')dnl
define(R, `%o3')dnl
define(ITER, `%o4')dnl
define(V, `%o5')dnl
dnl
dnl m4 reminder: ifelse(a,b,c,d) => if a is b, then c, else d
define(T, `%g1')dnl
define(SC, `%g7')dnl
ifelse(S, `true', `define(SIGN, `%g6')')dnl

dnl
dnl This is the recursive definition for developing quotient digits.
dnl
dnl Parameters:
dnl  $1	the current depth, 1 <= $1 <= N
dnl  $2	the current accumulation of quotient bits
dnl  N	max depth
dnl
dnl We add a new bit to $2 and either recurse or insert the bits in
dnl the quotient.  R, Q, and V are inputs and outputs as defined above;
dnl the condition codes are expected to reflect the input R, and are
dnl modified to reflect the output R.
dnl
define(DEVELOP_QUOTIENT_BITS,
`	! depth $1, accumulated bits $2
	bl	L.$1.eval(2**N+$2)
	srl	V,1,V
	! remainder is positive
	subcc	R,V,R
	ifelse($1, N,
	`	b	9f
		add	Q, ($2*2+1), Q
	', `	DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2+1)')')
L.$1.eval(2**N+$2):
	! remainder is negative
	addcc	R,V,R
	ifelse($1, N,
	`	b	9f
		add	Q, ($2*2-1), Q
	', `	DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2-1)')')
	ifelse($1, 1, `9:')')dnl

#include "sysdep.h"
#ifdef __linux__
#include <asm/traps.h>
#else
#ifdef __svr4__
#include <sys/trap.h>
#else
#include <machine/trap.h>
#endif
#endif

FUNC(NAME)
ifelse(S, `true',
`	! compute sign of result; if neither is negative, no problem
	orcc	divisor, dividend, %g0	! either negative?
	bge	2f			! no, go do the divide
ifelse(OP, `div',
`	xor	divisor, dividend, SIGN	! compute sign in any case',
`	mov	dividend, SIGN		! sign of remainder matches dividend')
	tst	divisor
	bge	1f
	tst	dividend
	! divisor is definitely negative; dividend might also be negative
	bge	2f			! if dividend not negative...
	sub	%g0, divisor, divisor	! in any case, make divisor nonneg
1:	! dividend is negative, divisor is nonnegative
	sub	%g0, dividend, dividend	! make dividend nonnegative
2:
')
	! Ready to divide.  Compute size of quotient; scale comparand.
	orcc	divisor, %g0, V
	bne	1f
	mov	dividend, R

		! Divide by zero trap.  If it returns, return 0 (about as
		! wrong as possible, but that is what SunOS does...).
		ta	ST_DIV0
		retl
		clr	%o0

1:
	cmp	R, V			! if divisor exceeds dividend, done
	blu	Lgot_result		! (and algorithm fails otherwise)
	clr	Q
	sethi	%hi(1 << (WORDSIZE - TOPBITS - 1)), T
	cmp	R, T
	blu	Lnot_really_big
	clr	ITER

	! `Here the dividend is >= 2**(31-N) or so.  We must be careful here,
	! as our usual N-at-a-shot divide step will cause overflow and havoc.
	! The number of bits in the result here is N*ITER+SC, where SC <= N.
	! Compute ITER in an unorthodox manner: know we need to shift V into
	! the top decade: so do not even bother to compare to R.'
	1:
		cmp	V, T
		bgeu	3f
		mov	1, SC
		sll	V, N, V
		b	1b
		add	ITER, 1, ITER

	! Now compute SC.
	2:	addcc	V, V, V
		bcc	Lnot_too_big
		add	SC, 1, SC

		! We get here if the divisor overflowed while shifting.
		! This means that R has the high-order bit set.
		! Restore V and subtract from R.
		sll	T, TOPBITS, T	! high order bit
		srl	V, 1, V		! rest of V
		add	V, T, V
		b	Ldo_single_div
		sub	SC, 1, SC

	Lnot_too_big:
	3:	cmp	V, R
		blu	2b
		nop
		be	Ldo_single_div
		nop
	/* NB: these are commented out in the V8-Sparc manual as well */
	/* (I do not understand this) */
	! V > R: went too far: back up 1 step
	!	srl	V, 1, V
	!	dec	SC
	! do single-bit divide steps
	!
	! We have to be careful here.  We know that R >= V, so we can do the
	! first divide step without thinking.  BUT, the others are conditional,
	! and are only done if R >= 0.  Because both R and V may have the high-
	! order bit set in the first step, just falling into the regular
	! division loop will mess up the first time around.
	! So we unroll slightly...
	Ldo_single_div:
		subcc	SC, 1, SC
		bl	Lend_regular_divide
		nop
		sub	R, V, R
		mov	1, Q
		b	Lend_single_divloop
		nop
	Lsingle_divloop:
		sll	Q, 1, Q
		bl	1f
		srl	V, 1, V
		! R >= 0
		sub	R, V, R
		b	2f
		add	Q, 1, Q
	1:	! R < 0
		add	R, V, R
		sub	Q, 1, Q
	2:
	Lend_single_divloop:
		subcc	SC, 1, SC
		bge	Lsingle_divloop
		tst	R
		b,a	Lend_regular_divide

Lnot_really_big:
1:
	sll	V, N, V
	cmp	V, R
	bleu	1b
	addcc	ITER, 1, ITER
	be	Lgot_result
	sub	ITER, 1, ITER

	tst	R	! set up for initial iteration
Ldivloop:
	sll	Q, N, Q
	DEVELOP_QUOTIENT_BITS(1, 0)
Lend_regular_divide:
	subcc	ITER, 1, ITER
	bge	Ldivloop
	tst	R
	bl,a	Lgot_result
	! non-restoring fixup here (one instruction only!)
ifelse(OP, `div',
`	sub	Q, 1, Q
', `	add	R, divisor, R
')

Lgot_result:
ifelse(S, `true',
`	! check to see if answer should be < 0
	tst	SIGN
	bl,a	1f
	ifelse(OP, `div', `sub %g0, Q, Q', `sub %g0, R, R')
1:')
	retl
	ifelse(OP, `div', `mov Q, %o0', `mov R, %o0')