modf.S revision 1.5 1 1.5 joerg /* $NetBSD: modf.S,v 1.5 2013/09/12 15:36:16 joerg Exp $ */
2 1.1 eeh
3 1.1 eeh /*
4 1.1 eeh * Copyright (c) 1992, 1993
5 1.1 eeh * The Regents of the University of California. All rights reserved.
6 1.1 eeh *
7 1.1 eeh * This software was developed by the Computer Systems Engineering group
8 1.1 eeh * at Lawrence Berkeley Laboratory under DARPA contract BG 91-66 and
9 1.1 eeh * contributed to Berkeley.
10 1.1 eeh *
11 1.1 eeh * Redistribution and use in source and binary forms, with or without
12 1.1 eeh * modification, are permitted provided that the following conditions
13 1.1 eeh * are met:
14 1.1 eeh * 1. Redistributions of source code must retain the above copyright
15 1.1 eeh * notice, this list of conditions and the following disclaimer.
16 1.1 eeh * 2. Redistributions in binary form must reproduce the above copyright
17 1.1 eeh * notice, this list of conditions and the following disclaimer in the
18 1.1 eeh * documentation and/or other materials provided with the distribution.
19 1.4 agc * 3. Neither the name of the University nor the names of its contributors
20 1.1 eeh * may be used to endorse or promote products derived from this software
21 1.1 eeh * without specific prior written permission.
22 1.1 eeh *
23 1.1 eeh * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
24 1.1 eeh * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
25 1.1 eeh * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
26 1.1 eeh * ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
27 1.1 eeh * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
28 1.1 eeh * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
29 1.1 eeh * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
30 1.1 eeh * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
31 1.1 eeh * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
32 1.1 eeh * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
33 1.1 eeh * SUCH DAMAGE.
34 1.1 eeh *
35 1.1 eeh * from: Header: modf.s,v 1.3 92/06/20 00:00:54 torek Exp
36 1.1 eeh */
37 1.1 eeh
38 1.1 eeh #include <machine/asm.h>
39 1.1 eeh #if defined(LIBC_SCCS) && !defined(lint)
40 1.1 eeh #if 0
41 1.1 eeh .asciz "@(#)modf.s 8.1 (Berkeley) 6/4/93"
42 1.1 eeh #else
43 1.5 joerg RCSID("$NetBSD: modf.S,v 1.5 2013/09/12 15:36:16 joerg Exp $")
44 1.1 eeh #endif
45 1.1 eeh #endif /* LIBC_SCCS and not lint */
46 1.1 eeh
47 1.1 eeh #include <machine/fsr.h>
48 1.1 eeh
49 1.1 eeh /*
50 1.1 eeh * double modf(double val, double *iptr)
51 1.1 eeh *
52 1.1 eeh * Returns the fractional part of `val', storing the integer part of
53 1.1 eeh * `val' in *iptr. Both *iptr and the return value have the same sign
54 1.1 eeh * as `val'.
55 1.1 eeh *
56 1.1 eeh * Method:
57 1.1 eeh *
58 1.1 eeh * We use the fpu's normalization hardware to compute the integer portion
59 1.1 eeh * of the double precision argument. Sun IEEE double precision numbers
60 1.1 eeh * have 52 bits of mantissa, 11 bits of exponent, and one bit of sign,
61 1.1 eeh * with the sign occupying bit 31 of word 0, and the exponent bits 30:20
62 1.1 eeh * of word 0. Thus, values >= 2^52 are by definition integers.
63 1.1 eeh *
64 1.1 eeh * If we take a value that is in the range [+0..2^52) and add 2^52, all
65 1.1 eeh * of the fractional bits fall out and all of the integer bits are summed
66 1.1 eeh * with 2^52. If we then subtract 2^52, we get those integer bits back.
67 1.1 eeh * This must be done with rounding set to `towards 0' or `towards -inf'.
68 1.1 eeh * `Toward -inf' fails when the value is 0 (we get -0 back)....
69 1.1 eeh *
70 1.1 eeh * Note that this method will work anywhere, but is machine dependent in
71 1.1 eeh * various aspects.
72 1.1 eeh *
73 1.1 eeh * Stack usage:
74 1.2 eeh * 4@[%fp + BIAS - 4] saved %fsr
75 1.2 eeh * 4@[%fp + BIAS - 8] new %fsr with rounding set to `towards 0'
76 1.2 eeh * 8@[%fp + BIAS - 16] space for moving between %i and %f registers
77 1.1 eeh * Register usage:
78 1.3 eeh * %f0:f1 double val;
79 1.1 eeh * %l0 scratch
80 1.1 eeh * %l1 sign bit (0x80000000)
81 1.3 eeh * %i1 double *iptr;
82 1.1 eeh * %f2:f3 `magic number' 2^52, in fpu registers
83 1.1 eeh * %f4:f5 double v, in fpu registers
84 1.3 eeh * %f6:f7 double temp.
85 1.1 eeh */
86 1.1 eeh
87 1.1 eeh .align 8
88 1.1 eeh Lmagic:
89 1.1 eeh .word 0x43300000 ! sign = 0, exponent = 52 + 1023, mantissa = 0
90 1.1 eeh .word 0 ! (i.e., .double 0r4503599627370496e+00)
91 1.1 eeh
92 1.1 eeh L0:
93 1.1 eeh .word 0 ! 0.0
94 1.1 eeh .word 0
95 1.1 eeh
96 1.1 eeh ENTRY(modf)
97 1.2 eeh save %sp, -CC64FSZ-16, %sp
98 1.1 eeh
99 1.1 eeh /*
100 1.3 eeh * First, compute v = abs(val)
101 1.3 eeh */
102 1.3 eeh fabsd %f0, %f4 ! %f4:f5 = v
103 1.3 eeh fcmped %fcc1, %f0, %f4 ! %fcc1 = (val == abs(val))
104 1.5 joerg #ifdef __PIC__
105 1.1 eeh PICCY_SET(Lmagic, %l0, %o7)
106 1.1 eeh ldd [%l0], %f2
107 1.1 eeh #else
108 1.1 eeh sethi %hi(Lmagic), %l0
109 1.1 eeh ldd [%l0 + %lo(Lmagic)], %f2
110 1.1 eeh #endif
111 1.1 eeh
112 1.1 eeh /*
113 1.1 eeh * Is %f4:f5 >= %f2:f3 ? If so, it is all integer bits.
114 1.1 eeh * It is probably less, though.
115 1.1 eeh */
116 1.1 eeh fcmped %f4, %f2
117 1.1 eeh fbuge Lbig ! if >= (or unordered), go out
118 1.1 eeh nop
119 1.1 eeh
120 1.1 eeh /*
121 1.1 eeh * v < 2^52, so add 2^52, then subtract 2^52, but do it all
122 1.1 eeh * with rounding set towards zero. We leave any enabled
123 1.1 eeh * traps enabled, but change the rounding mode. This might
124 1.1 eeh * not be so good. Oh well....
125 1.1 eeh */
126 1.3 eeh st %fsr, [%fp + BIAS - 4] ! %l5 = current FSR mode
127 1.1 eeh set FSR_RD, %l3 ! %l3 = rounding direction mask
128 1.2 eeh ld [%fp + BIAS - 4], %l5
129 1.1 eeh set FSR_RD_RZ << FSR_RD_SHIFT, %l4
130 1.1 eeh andn %l5, %l3, %l6
131 1.1 eeh or %l6, %l4, %l6 ! round towards zero, please
132 1.1 eeh and %l5, %l3, %l5 ! save original rounding mode
133 1.2 eeh st %l6, [%fp + BIAS - 8]
134 1.2 eeh ld [%fp + BIAS - 8], %fsr
135 1.1 eeh
136 1.1 eeh faddd %f4, %f2, %f4 ! %f4:f5 += 2^52
137 1.1 eeh fsubd %f4, %f2, %f4 ! %f4:f5 -= 2^52
138 1.1 eeh
139 1.1 eeh /*
140 1.1 eeh * Restore %fsr, but leave exceptions accrued.
141 1.1 eeh */
142 1.2 eeh st %fsr, [%fp + BIAS - 4]
143 1.2 eeh ld [%fp + BIAS - 4], %l6
144 1.1 eeh andn %l6, %l3, %l6 ! %l6 = %fsr & ~FSR_RD;
145 1.1 eeh or %l5, %l6, %l5 ! %l5 |= %l6;
146 1.2 eeh st %l5, [%fp + BIAS - 4]
147 1.3 eeh ld [%fp + BIAS - 4], %fsr ! restore %fsr, leaving accrued stuff
148 1.1 eeh
149 1.1 eeh /*
150 1.1 eeh * Now insert the original sign in %f4:f5.
151 1.3 eeh * %fcc1 should still have the reults of (val == abs(val))
152 1.3 eeh * from above, so we use a conditional move on %fcc1 to:
153 1.3 eeh *
154 1.3 eeh * %f4 = (val == abs(val)) ? %f4 : -%f4
155 1.3 eeh *
156 1.1 eeh */
157 1.3 eeh fnegd %f4, %f6
158 1.3 eeh fmovdnz %fcc1, %f6, %f4
159 1.1 eeh 1:
160 1.1 eeh
161 1.1 eeh /*
162 1.1 eeh * The value in %f4:f5 is now the integer portion of the original
163 1.3 eeh * argument. We need to store this in *ival (%i1), subtract it
164 1.3 eeh * from the original value argument (%d0), and return the result.
165 1.1 eeh */
166 1.3 eeh std %f4, [%i1] ! *ival = %f4:f5;
167 1.1 eeh fsubd %f0, %f4, %f0 ! %f0:f1 -= %f4:f5;
168 1.1 eeh ret
169 1.1 eeh restore
170 1.1 eeh
171 1.1 eeh Lbig:
172 1.1 eeh /*
173 1.1 eeh * We get here if the original comparison of %f4:f5 (v) to
174 1.1 eeh * %f2:f3 (2^52) came out `greater or unordered'. In this
175 1.1 eeh * case the integer part is the original value, and the
176 1.1 eeh * fractional part is 0.
177 1.1 eeh */
178 1.5 joerg #ifdef __PIC__
179 1.1 eeh PICCY_SET(L0, %l0, %o7)
180 1.3 eeh std %f0, [%i1] ! *ival = val;
181 1.1 eeh ldd [%l0], %f0 ! return 0.0;
182 1.1 eeh #else
183 1.1 eeh sethi %hi(L0), %l0
184 1.3 eeh std %f0, [%i1] ! *ival = val;
185 1.1 eeh ldd [%l0 + %lo(L0)], %f0 ! return 0.0;
186 1.1 eeh #endif
187 1.1 eeh ret
188 1.1 eeh restore
189 1.3 eeh
190