Home | History | Annotate | Line # | Download | only in gen
modf.S revision 1.5
      1 /*	$NetBSD: modf.S,v 1.5 1997/07/16 14:37:41 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. All advertising materials mentioning features or use of this software
     20  *    must display the following acknowledgement:
     21  *	This product includes software developed by the University of
     22  *	California, Berkeley and its contributors.
     23  * 4. Neither the name of the University nor the names of its contributors
     24  *    may be used to endorse or promote products derived from this software
     25  *    without specific prior written permission.
     26  *
     27  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
     28  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
     29  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
     30  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
     31  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
     32  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
     33  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
     34  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
     35  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
     36  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
     37  * SUCH DAMAGE.
     38  *
     39  * from: Header: modf.s,v 1.3 92/06/20 00:00:54 torek Exp
     40  */
     41 
     42 #include <machine/asm.h>
     43 #if defined(LIBC_SCCS) && !defined(lint)
     44 #if 0
     45 	.asciz "@(#)modf.s	8.1 (Berkeley) 6/4/93"
     46 #else
     47 	RCSID("$NetBSD: modf.S,v 1.5 1997/07/16 14:37:41 christos Exp $")
     48 #endif
     49 #endif /* LIBC_SCCS and not lint */
     50 
     51 #include <machine/fsr.h>
     52 
     53 /*
     54  * double modf(double val, double *iptr)
     55  *
     56  * Returns the fractional part of `val', storing the integer part of
     57  * `val' in *iptr.  Both *iptr and the return value have the same sign
     58  * as `val'.
     59  *
     60  * Method:
     61  *
     62  * We use the fpu's normalization hardware to compute the integer portion
     63  * of the double precision argument.  Sun IEEE double precision numbers
     64  * have 52 bits of mantissa, 11 bits of exponent, and one bit of sign,
     65  * with the sign occupying bit 31 of word 0, and the exponent bits 30:20
     66  * of word 0.  Thus, values >= 2^52 are by definition integers.
     67  *
     68  * If we take a value that is in the range [+0..2^52) and add 2^52, all
     69  * of the fractional bits fall out and all of the integer bits are summed
     70  * with 2^52.  If we then subtract 2^52, we get those integer bits back.
     71  * This must be done with rounding set to `towards 0' or `towards -inf'.
     72  * `Toward -inf' fails when the value is 0 (we get -0 back)....
     73  *
     74  * Note that this method will work anywhere, but is machine dependent in
     75  * various aspects.
     76  *
     77  * Stack usage:
     78  *	4@[%fp - 4]	saved %fsr
     79  *	4@[%fp - 8]	new %fsr with rounding set to `towards 0'
     80  *	8@[%fp - 16]	space for moving between %i and %f registers
     81  * Register usage:
     82  *	%i0%i1		double val;
     83  *	%l0		scratch
     84  *	%l1		sign bit (0x80000000)
     85  *	%i2		double *iptr;
     86  *	%f2:f3		`magic number' 2^52, in fpu registers
     87  *	%f4:f5		double v, in fpu registers
     88  */
     89 
     90 	.align	8
     91 Lmagic:
     92 	.word	0x43300000	! sign = 0, exponent = 52 + 1023, mantissa = 0
     93 	.word	0		! (i.e., .double 0r4503599627370496e+00)
     94 
     95 L0:
     96 	.word	0		! 0.0
     97 	.word	0
     98 
     99 ENTRY(modf)
    100 	save	%sp, -64-16, %sp
    101 
    102 	/*
    103 	 * First, compute v = abs(val) by clearing sign bit,
    104 	 * and then set up the fpu registers.  This would be
    105 	 * much easier if we could do alu operations on fpu registers!
    106 	 */
    107 	sethi	%hi(0x80000000), %l1	! sign bit
    108 	andn	%i0, %l1, %l0
    109 	st	%l0, [%fp - 16]
    110 #ifdef PIC
    111 	PICCY_SET(Lmagic, %l0, %o7)
    112 	ldd	[%l0], %f2
    113 #else
    114 	sethi	%hi(Lmagic), %l0
    115 	ldd	[%l0 + %lo(Lmagic)], %f2
    116 #endif
    117 	st	%i1, [%fp - 12]
    118 	ldd	[%fp - 16], %f4		! %f4:f5 = v
    119 
    120 	/*
    121 	 * Is %f4:f5 >= %f2:f3 ?  If so, it is all integer bits.
    122 	 * It is probably less, though.
    123 	 */
    124 	fcmped	%f4, %f2
    125 	nop				! fpop2 delay
    126 	fbuge	Lbig			! if >= (or unordered), go out
    127 	nop
    128 
    129 	/*
    130 	 * v < 2^52, so add 2^52, then subtract 2^52, but do it all
    131 	 * with rounding set towards zero.  We leave any enabled
    132 	 * traps enabled, but change the rounding mode.  This might
    133 	 * not be so good.  Oh well....
    134 	 */
    135 	st	%fsr, [%fp - 4]		! %l5 = current FSR mode
    136 	set	FSR_RD, %l3		! %l3 = rounding direction mask
    137 	ld	[%fp - 4], %l5
    138 	set	FSR_RD_RZ << FSR_RD_SHIFT, %l4
    139 	andn	%l5, %l3, %l6
    140 	or	%l6, %l4, %l6		! round towards zero, please
    141 	and	%l5, %l3, %l5		! save original rounding mode
    142 	st	%l6, [%fp - 8]
    143 	ld	[%fp - 8], %fsr
    144 
    145 	faddd	%f4, %f2, %f4		! %f4:f5 += 2^52
    146 	fsubd	%f4, %f2, %f4		! %f4:f5 -= 2^52
    147 
    148 	/*
    149 	 * Restore %fsr, but leave exceptions accrued.
    150 	 */
    151 	st	%fsr, [%fp - 4]
    152 	ld	[%fp - 4], %l6
    153 	andn	%l6, %l3, %l6		! %l6 = %fsr & ~FSR_RD;
    154 	or	%l5, %l6, %l5		! %l5 |= %l6;
    155 	st	%l5, [%fp - 4]
    156 	ld	[%fp - 4], %fsr		! restore %fsr, leaving accrued stuff
    157 
    158 	/*
    159 	 * Now insert the original sign in %f4:f5.
    160 	 * This is a lot of work, so it is conditional here.
    161 	 */
    162 	btst	%l1, %i0
    163 	be	1f
    164 	nop
    165 	st	%f4, [%fp - 16]
    166 	ld	[%fp - 16], %g1
    167 	or	%l1, %g1, %g1
    168 	st	%g1, [%fp - 16]
    169 	ld	[%fp - 16], %f4
    170 1:
    171 
    172 	/*
    173 	 * The value in %f4:f5 is now the integer portion of the original
    174 	 * argument.  We need to store this in *ival (%i2), subtract it
    175 	 * from the original value argument (%i0:i1), and return the result.
    176 	 */
    177 	std	%f4, [%i2]		! *ival = %f4:f5;
    178 	std	%i0, [%fp - 16]
    179 	ldd	[%fp - 16], %f0		! %f0:f1 = val;
    180 	fsubd	%f0, %f4, %f0		! %f0:f1 -= %f4:f5;
    181 	ret
    182 	restore
    183 
    184 Lbig:
    185 	/*
    186 	 * We get here if the original comparison of %f4:f5 (v) to
    187 	 * %f2:f3 (2^52) came out `greater or unordered'.  In this
    188 	 * case the integer part is the original value, and the
    189 	 * fractional part is 0.
    190 	 */
    191 #ifdef PIC
    192 	PICCY_SET(L0, %l0, %o7)
    193 	std	%f0, [%i2]		! *ival = val;
    194 	ldd	[%l0], %f0		! return 0.0;
    195 #else
    196 	sethi	%hi(L0), %l0
    197 	std	%f0, [%i2]		! *ival = val;
    198 	ldd	[%l0 + %lo(L0)], %f0	! return 0.0;
    199 #endif
    200 	ret
    201 	restore
    202