xref: /netbsd-src/sys/lib/libkern/arch/sparc64/divrem.m4 (revision 946379e7b37692fc43f68eb0d1c10daa0a7f3b6c)
1/*	$NetBSD: divrem.m4,v 1.5 2005/12/11 12:24:45 christos Exp $	*/
2
3/*
4 * Copyright (c) 1992, 1993
5 *	The Regents of the University of California.  All rights reserved.
6 *
7 * This software was developed by the Computer Systems Engineering group
8 * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
9 * contributed to Berkeley.
10 *
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted provided that the following conditions
13 * are met:
14 * 1. Redistributions of source code must retain the above copyright
15 *    notice, this list of conditions and the following disclaimer.
16 * 2. Redistributions in binary form must reproduce the above copyright
17 *    notice, this list of conditions and the following disclaimer in the
18 *    documentation and/or other materials provided with the distribution.
19 * 3. Neither the name of the University nor the names of its contributors
20 *    may be used to endorse or promote products derived from this software
21 *    without specific prior written permission.
22 *
23 * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 * SUCH DAMAGE.
34 *
35 * from: Header: divrem.m4,v 1.4 92/06/25 13:23:57 torek Exp
36 */
37
38#include <machine/asm.h>
39#include <machine/trap.h>
40
41/*
42 * Division and remainder, from Appendix E of the Sparc Version 8
43 * Architecture Manual, with fixes from Gordon Irlam.
44 */
45
46#if defined(LIBC_SCCS)
47	RCSID("$NetBSD: divrem.m4,v 1.5 2005/12/11 12:24:45 christos Exp $")
48#endif
49
50/*
51 * Input: dividend and divisor in %o0 and %o1 respectively.
52 *
53 * m4 parameters:
54 *  NAME	name of function to generate
55 *  OP		OP=div => %o0 / %o1; OP=rem => %o0 % %o1
56 *  S		S=true => signed; S=false => unsigned
57 *
58 * Algorithm parameters:
59 *  N		how many bits per iteration we try to get (4)
60 *  WORDSIZE	total number of bits (32)
61 *
62 * Derived constants:
63 *  TWOSUPN	2^N, for label generation (m4 exponentiation currently broken)
64 *  TOPBITS	number of bits in the top `decade' of a number
65 *
66 * Important variables:
67 *  Q		the partial quotient under development (initially 0)
68 *  R		the remainder so far, initially the dividend
69 *  ITER	number of main division loop iterations required;
70 *		equal to ceil(log2(quotient) / N).  Note that this
71 *		is the log base (2^N) of the quotient.
72 *  V		the current comparand, initially divisor*2^(ITER*N-1)
73 *
74 * Cost:
75 *  Current estimate for non-large dividend is
76 *	ceil(log2(quotient) / N) * (10 + 7N/2) + C
77 *  A large dividend is one greater than 2^(31-TOPBITS) and takes a
78 *  different path, as the upper bits of the quotient must be developed
79 *  one bit at a time.
80 */
81
82define(N, `4')
83define(TWOSUPN, `16')
84define(WORDSIZE, `32')
85define(TOPBITS, eval(WORDSIZE - N*((WORDSIZE-1)/N)))
86
87define(dividend, `%o0')
88define(divisor, `%o1')
89define(Q, `%o2')
90define(R, `%o3')
91define(ITER, `%o4')
92define(V, `%o5')
93
94/* m4 reminder: ifelse(a,b,c,d) => if a is b, then c, else d */
95define(T, `%g1')
96define(SC, `%g5')
97ifelse(S, `true', `define(SIGN, `%g6')')
98
99/*
100 * This is the recursive definition for developing quotient digits.
101 *
102 * Parameters:
103 *  $1	the current depth, 1 <= $1 <= N
104 *  $2	the current accumulation of quotient bits
105 *  N	max depth
106 *
107 * We add a new bit to $2 and either recurse or insert the bits in
108 * the quotient.  R, Q, and V are inputs and outputs as defined above;
109 * the condition codes are expected to reflect the input R, and are
110 * modified to reflect the output R.
111 */
112define(DEVELOP_QUOTIENT_BITS,
113`	! depth $1, accumulated bits $2
114	bl	L.$1.eval(TWOSUPN+$2)
115	srl	V,1,V
116	! remainder is positive
117	subcc	R,V,R
118	ifelse($1, N,
119	`	b	9f
120		add	Q, ($2*2+1), Q
121	', `	DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2+1)')')
122L.$1.eval(TWOSUPN+$2):
123	! remainder is negative
124	addcc	R,V,R
125	ifelse($1, N,
126	`	b	9f
127		add	Q, ($2*2-1), Q
128	', `	DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2-1)')')
129	ifelse($1, 1, `9:')')
130
131FUNC(NAME)
132ifelse(S, `true',
133`	! compute sign of result; if neither is negative, no problem
134	orcc	divisor, dividend, %g0	! either negative?
135	bge	2f			! no, go do the divide
136	ifelse(OP, `div',
137		`xor	divisor, dividend, SIGN',
138		`mov	dividend, SIGN')	! compute sign in any case
139	tst	divisor
140	bge	1f
141	tst	dividend
142	! divisor is definitely negative; dividend might also be negative
143	bge	2f			! if dividend not negative...
144	neg	divisor			! in any case, make divisor nonneg
1451:	! dividend is negative, divisor is nonnegative
146	neg	dividend		! make dividend nonnegative
1472:
148')
149	! Ready to divide.  Compute size of quotient; scale comparand.
150	orcc	divisor, %g0, V
151	bnz	1f
152	mov	dividend, R
153
154		! Divide by zero trap.  If it returns, return 0 (about as
155		! wrong as possible, but that is what SunOS does...).
156		t	ST_DIV0
157		retl
158		clr	%o0
159
1601:
161	cmp	R, V			! if divisor exceeds dividend, done
162	blu	Lgot_result		! (and algorithm fails otherwise)
163	clr	Q
164	sethi	%hi(1 << (WORDSIZE - TOPBITS - 1)), T
165	cmp	R, T
166	blu	Lnot_really_big
167	clr	ITER
168
169	! `Here the dividend is >= 2^(31-N) or so.  We must be careful here,
170	! as our usual N-at-a-shot divide step will cause overflow and havoc.
171	! The number of bits in the result here is N*ITER+SC, where SC <= N.
172	! Compute ITER in an unorthodox manner: know we need to shift V into
173	! the top decade: so do not even bother to compare to R.'
174	1:
175		cmp	V, T
176		bgeu	3f
177		mov	1, SC
178		sll	V, N, V
179		b	1b
180		inc	ITER
181
182	! Now compute SC.
183	2:	addcc	V, V, V
184		bcc	Lnot_too_big
185		inc	SC
186
187		! We get here if the divisor overflowed while shifting.
188		! This means that R has the high-order bit set.
189		! Restore V and subtract from R.
190		sll	T, TOPBITS, T	! high order bit
191		srl	V, 1, V		! rest of V
192		add	V, T, V
193		b	Ldo_single_div
194		dec	SC
195
196	Lnot_too_big:
197	3:	cmp	V, R
198		blu	2b
199		nop
200		be	Ldo_single_div
201		nop
202	/* NB: these are commented out in the V8-Sparc manual as well */
203	/* (I do not understand this) */
204	! V > R: went too far: back up 1 step
205	!	srl	V, 1, V
206	!	dec	SC
207	! do single-bit divide steps
208	!
209	! We have to be careful here.  We know that R >= V, so we can do the
210	! first divide step without thinking.  BUT, the others are conditional,
211	! and are only done if R >= 0.  Because both R and V may have the high-
212	! order bit set in the first step, just falling into the regular
213	! division loop will mess up the first time around.
214	! So we unroll slightly...
215	Ldo_single_div:
216		deccc	SC
217		bl	Lend_regular_divide
218		nop
219		sub	R, V, R
220		mov	1, Q
221		b	Lend_single_divloop
222		nop
223	Lsingle_divloop:
224		sll	Q, 1, Q
225		bl	1f
226		srl	V, 1, V
227		! R >= 0
228		sub	R, V, R
229		b	2f
230		inc	Q
231	1:	! R < 0
232		add	R, V, R
233		dec	Q
234	2:
235	Lend_single_divloop:
236		deccc	SC
237		bge	Lsingle_divloop
238		tst	R
239		b,a	Lend_regular_divide
240
241Lnot_really_big:
2421:
243	sll	V, N, V
244	cmp	V, R
245	bleu	1b
246	inccc	ITER
247	be	Lgot_result
248	dec	ITER
249
250	tst	R	! set up for initial iteration
251Ldivloop:
252	sll	Q, N, Q
253	DEVELOP_QUOTIENT_BITS(1, 0)
254Lend_regular_divide:
255	deccc	ITER
256	bge	Ldivloop
257	tst	R
258	bl,a	Lgot_result
259	! non-restoring fixup here (one instruction only!)
260ifelse(OP, `div',
261`	dec	Q
262', `	add	R, divisor, R
263')
264
265Lgot_result:
266ifelse(S, `true',
267`	! check to see if answer should be < 0
268	tst	SIGN
269	bl,a	1f
270	ifelse(OP, `div', `neg Q', `neg R')
2711:')
272	retl
273	ifelse(OP, `div', `mov Q, %o0', `mov R, %o0')
274