xref: /csrg-svn/lib/libc/sparc/gen/divrem.m4 (revision 61167)
154390Storek/*
261166Sbostic * Copyright (c) 1992, 1993
361166Sbostic *	The Regents of the University of California.  All rights reserved.
454390Storek *
554390Storek * This software was developed by the Computer Systems Engineering group
654390Storek * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
754390Storek * contributed to Berkeley.
854390Storek *
954390Storek * %sccs.include.redist.c%
1054390Storek *
1154390Storek * from: $Header: divrem.m4,v 1.4 92/06/25 13:23:57 torek Exp $
1254390Storek */
1354390Storek
1454390Storek/*
1554390Storek * Division and remainder, from Appendix E of the Sparc Version 8
1654390Storek * Architecture Manual, with fixes from Gordon Irlam.
1754390Storek */
1854390Storek
1954390Storek#if defined(LIBC_SCCS) && !defined(lint)
20*61167Sbostic	.asciz "@(#)divrem.m4	8.1 (Berkeley) 06/04/93"
2154390Storek#endif /* LIBC_SCCS and not lint */
2254390Storek
2354390Storek/*
2454390Storek * Input: dividend and divisor in %o0 and %o1 respectively.
2554390Storek *
2654390Storek * m4 parameters:
2754390Storek *  NAME	name of function to generate
2854390Storek *  OP		OP=div => %o0 / %o1; OP=rem => %o0 % %o1
2954390Storek *  S		S=true => signed; S=false => unsigned
3054390Storek *
3154390Storek * Algorithm parameters:
3254390Storek *  N		how many bits per iteration we try to get (4)
3354390Storek *  WORDSIZE	total number of bits (32)
3454390Storek *
3554390Storek * Derived constants:
3654390Storek *  TWOSUPN	2^N, for label generation (m4 exponentiation currently broken)
3754390Storek *  TOPBITS	number of bits in the top `decade' of a number
3854390Storek *
3954390Storek * Important variables:
4054390Storek *  Q		the partial quotient under development (initially 0)
4154390Storek *  R		the remainder so far, initially the dividend
4254390Storek *  ITER	number of main division loop iterations required;
4354390Storek *		equal to ceil(log2(quotient) / N).  Note that this
4454390Storek *		is the log base (2^N) of the quotient.
4554390Storek *  V		the current comparand, initially divisor*2^(ITER*N-1)
4654390Storek *
4754390Storek * Cost:
4854390Storek *  Current estimate for non-large dividend is
4954390Storek *	ceil(log2(quotient) / N) * (10 + 7N/2) + C
5054390Storek *  A large dividend is one greater than 2^(31-TOPBITS) and takes a
5154390Storek *  different path, as the upper bits of the quotient must be developed
5254390Storek *  one bit at a time.
5354390Storek */
5454390Storek
5554390Storekdefine(N, `4')
5654390Storekdefine(TWOSUPN, `16')
5754390Storekdefine(WORDSIZE, `32')
5854390Storekdefine(TOPBITS, eval(WORDSIZE - N*((WORDSIZE-1)/N)))
5954390Storek
6054390Storekdefine(dividend, `%o0')
6154390Storekdefine(divisor, `%o1')
6254390Storekdefine(Q, `%o2')
6354390Storekdefine(R, `%o3')
6454390Storekdefine(ITER, `%o4')
6554390Storekdefine(V, `%o5')
6654390Storek
6754390Storek/* m4 reminder: ifelse(a,b,c,d) => if a is b, then c, else d */
6854390Storekdefine(T, `%g1')
6954390Storekdefine(SC, `%g7')
7054390Storekifelse(S, `true', `define(SIGN, `%g6')')
7154390Storek
7254390Storek/*
7354390Storek * This is the recursive definition for developing quotient digits.
7454390Storek *
7554390Storek * Parameters:
7654390Storek *  $1	the current depth, 1 <= $1 <= N
7754390Storek *  $2	the current accumulation of quotient bits
7854390Storek *  N	max depth
7954390Storek *
8054390Storek * We add a new bit to $2 and either recurse or insert the bits in
8154390Storek * the quotient.  R, Q, and V are inputs and outputs as defined above;
8254390Storek * the condition codes are expected to reflect the input R, and are
8354390Storek * modified to reflect the output R.
8454390Storek */
8554390Storekdefine(DEVELOP_QUOTIENT_BITS,
8654390Storek`	! depth $1, accumulated bits $2
8754390Storek	bl	L.$1.eval(TWOSUPN+$2)
8854390Storek	srl	V,1,V
8954390Storek	! remainder is positive
9054390Storek	subcc	R,V,R
9154390Storek	ifelse($1, N,
9254390Storek	`	b	9f
9354390Storek		add	Q, ($2*2+1), Q
9454390Storek	', `	DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2+1)')')
9554390StorekL.$1.eval(TWOSUPN+$2):
9654390Storek	! remainder is negative
9754390Storek	addcc	R,V,R
9854390Storek	ifelse($1, N,
9954390Storek	`	b	9f
10054390Storek		add	Q, ($2*2-1), Q
10154390Storek	', `	DEVELOP_QUOTIENT_BITS(incr($1), `eval(2*$2-1)')')
10254390Storek	ifelse($1, 1, `9:')')
10354390Storek
10454390Storek#include "DEFS.h"
10554390Storek#include <machine/trap.h>
10654390Storek
10754390StorekFUNC(NAME)
10854390Storekifelse(S, `true',
10954390Storek`	! compute sign of result; if neither is negative, no problem
11054390Storek	orcc	divisor, dividend, %g0	! either negative?
11154390Storek	bge	2f			! no, go do the divide
11254390Storek	xor	divisor, dividend, SIGN	! compute sign in any case
11354390Storek	tst	divisor
11454390Storek	bge	1f
11554390Storek	tst	dividend
11654390Storek	! divisor is definitely negative; dividend might also be negative
11754390Storek	bge	2f			! if dividend not negative...
11854390Storek	neg	divisor			! in any case, make divisor nonneg
11954390Storek1:	! dividend is negative, divisor is nonnegative
12054390Storek	neg	dividend		! make dividend nonnegative
12154390Storek2:
12254390Storek')
12354390Storek	! Ready to divide.  Compute size of quotient; scale comparand.
12454390Storek	orcc	divisor, %g0, V
12554390Storek	bnz	1f
12654390Storek	mov	dividend, R
12754390Storek
12854390Storek		! Divide by zero trap.  If it returns, return 0 (about as
12954390Storek		! wrong as possible, but that is what SunOS does...).
13054390Storek		t	ST_DIV0
13154390Storek		retl
13254390Storek		clr	%o0
13354390Storek
13454390Storek1:
13554390Storek	cmp	R, V			! if divisor exceeds dividend, done
13654390Storek	blu	Lgot_result		! (and algorithm fails otherwise)
13754390Storek	clr	Q
13854390Storek	sethi	%hi(1 << (WORDSIZE - TOPBITS - 1)), T
13954390Storek	cmp	R, T
14054390Storek	blu	Lnot_really_big
14154390Storek	clr	ITER
14254390Storek
14354390Storek	! `Here the dividend is >= 2^(31-N) or so.  We must be careful here,
14454390Storek	! as our usual N-at-a-shot divide step will cause overflow and havoc.
14554390Storek	! The number of bits in the result here is N*ITER+SC, where SC <= N.
14654390Storek	! Compute ITER in an unorthodox manner: know we need to shift V into
14754390Storek	! the top decade: so do not even bother to compare to R.'
14854390Storek	1:
14954390Storek		cmp	V, T
15054390Storek		bgeu	3f
15154390Storek		mov	1, SC
15254390Storek		sll	V, N, V
15354390Storek		b	1b
15454390Storek		inc	ITER
15554390Storek
15654390Storek	! Now compute SC.
15754390Storek	2:	addcc	V, V, V
15854390Storek		bcc	Lnot_too_big
15954390Storek		inc	SC
16054390Storek
16154390Storek		! We get here if the divisor overflowed while shifting.
16254390Storek		! This means that R has the high-order bit set.
16354390Storek		! Restore V and subtract from R.
16454390Storek		sll	T, TOPBITS, T	! high order bit
16554390Storek		srl	V, 1, V		! rest of V
16654390Storek		add	V, T, V
16754390Storek		b	Ldo_single_div
16854390Storek		dec	SC
16954390Storek
17054390Storek	Lnot_too_big:
17154390Storek	3:	cmp	V, R
17254390Storek		blu	2b
17354390Storek		nop
17454390Storek		be	Ldo_single_div
17554390Storek		nop
17654390Storek	/* NB: these are commented out in the V8-Sparc manual as well */
17754390Storek	/* (I do not understand this) */
17854390Storek	! V > R: went too far: back up 1 step
17954390Storek	!	srl	V, 1, V
18054390Storek	!	dec	SC
18154390Storek	! do single-bit divide steps
18254390Storek	!
18354390Storek	! We have to be careful here.  We know that R >= V, so we can do the
18454390Storek	! first divide step without thinking.  BUT, the others are conditional,
18554390Storek	! and are only done if R >= 0.  Because both R and V may have the high-
18654390Storek	! order bit set in the first step, just falling into the regular
18754390Storek	! division loop will mess up the first time around.
18854390Storek	! So we unroll slightly...
18954390Storek	Ldo_single_div:
19054390Storek		deccc	SC
19154390Storek		bl	Lend_regular_divide
19254390Storek		nop
19354390Storek		sub	R, V, R
19454390Storek		mov	1, Q
19554390Storek		b	Lend_single_divloop
19654390Storek		nop
19754390Storek	Lsingle_divloop:
19854390Storek		sll	Q, 1, Q
19954390Storek		bl	1f
20054390Storek		srl	V, 1, V
20154390Storek		! R >= 0
20254390Storek		sub	R, V, R
20354390Storek		b	2f
20454390Storek		inc	Q
20554390Storek	1:	! R < 0
20654390Storek		add	R, V, R
20754390Storek		dec	Q
20854390Storek	2:
20954390Storek	Lend_single_divloop:
21054390Storek		deccc	SC
21154390Storek		bge	Lsingle_divloop
21254390Storek		tst	R
21354390Storek		b,a	Lend_regular_divide
21454390Storek
21554390StorekLnot_really_big:
21654390Storek1:
21754390Storek	sll	V, N, V
21854390Storek	cmp	V, R
21954390Storek	bleu	1b
22054390Storek	inccc	ITER
22154390Storek	be	Lgot_result
22254390Storek	dec	ITER
22354390Storek
22454390Storek	tst	R	! set up for initial iteration
22554390StorekLdivloop:
22654390Storek	sll	Q, N, Q
22754390Storek	DEVELOP_QUOTIENT_BITS(1, 0)
22854390StorekLend_regular_divide:
22954390Storek	deccc	ITER
23054390Storek	bge	Ldivloop
23154390Storek	tst	R
23254390Storek	bl,a	Lgot_result
23354390Storek	! non-restoring fixup here (one instruction only!)
23454390Storekifelse(OP, `div',
23554390Storek`	dec	Q
23654390Storek', `	add	R, divisor, R
23754390Storek')
23854390Storek
23954390StorekLgot_result:
24054390Storekifelse(S, `true',
24154390Storek`	! check to see if answer should be < 0
24254390Storek	tst	SIGN
24354390Storek	bl,a	1f
24454390Storek	ifelse(OP, `div', `neg Q', `neg R')
24554390Storek1:')
24654390Storek	retl
24754390Storek	ifelse(OP, `div', `mov Q, %o0', `mov R, %o0')
248