1 /* @(#)e_fmod.c 1.3 95/01/18 */
2 /*-
3  * ====================================================
4  * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5  *
6  * Developed at SunSoft, a Sun Microsystems, Inc. business.
7  * Permission to use, copy, modify, and distribute this
8  * software is freely granted, provided that this notice
9  * is preserved.
10  * ====================================================
11  */
12 
13 #include <sys/types.h>
14 #include <machine/ieee.h>
15 
16 #include <float.h>
17 #include <math.h>
18 #include <stdint.h>
19 
20 #include "math_private.h"
21 
22 #define	BIAS (LDBL_MAX_EXP - 1)
23 
24 /*
25  * These macros add and remove an explicit integer bit in front of the
26  * fractional mantissa, if the architecture doesn't have such a bit by
27  * default already.
28  */
29 #ifdef LDBL_IMPLICIT_NBIT
30 #define	LDBL_NBIT	0
31 #define	SET_NBIT(hx)	((hx) | (1ULL << LDBL_MANH_SIZE))
32 #define	HFRAC_BITS	EXT_FRACHBITS
33 #else
34 #define	LDBL_NBIT	0x80000000
35 #define	SET_NBIT(hx)	(hx)
36 #define	HFRAC_BITS	(EXT_FRACHBITS - 1)
37 #endif
38 
39 #define	MANL_SHIFT	(EXT_FRACLBITS - 1)
40 
41 static const long double one = 1.0, Zero[] = {0.0, -0.0,};
42 
43 /*
44  * fmodl(x,y)
45  * Return x mod y in exact arithmetic
46  * Method: shift and subtract
47  *
48  * Assumptions:
49  * - The low part of the mantissa fits in a manl_t exactly.
50  * - The high part of the mantissa fits in an int64_t with enough room
51  *   for an explicit integer bit in front of the fractional bits.
52  */
53 long double
54 fmodl(long double x, long double y)
55 {
56 	union {
57 		long double e;
58 		struct ieee_ext bits;
59 	} ux, uy;
60 	int64_t hx,hz;	/* We need a carry bit even if LDBL_MANH_SIZE is 32. */
61 	uint32_t hy;
62 	uint32_t lx,ly,lz;
63 	int ix,iy,n,sx;
64 
65 	ux.e = x;
66 	uy.e = y;
67 	sx = ux.bits.ext_sign;
68 
69     /* purge off exception values */
70 	if((uy.bits.ext_exp|uy.bits.ext_frach|uy.bits.ext_fracl)==0 || /* y=0 */
71 	   (ux.bits.ext_exp == BIAS + LDBL_MAX_EXP) ||	 /* or x not finite */
72 	   (uy.bits.ext_exp == BIAS + LDBL_MAX_EXP &&
73 	    ((uy.bits.ext_frach&~LDBL_NBIT)|uy.bits.ext_fracl)!=0)) /* or y is NaN */
74 	    return (x*y)/(x*y);
75 	if(ux.bits.ext_exp<=uy.bits.ext_exp) {
76 	    if((ux.bits.ext_exp<uy.bits.ext_exp) ||
77 	       (ux.bits.ext_frach<=uy.bits.ext_frach &&
78 		(ux.bits.ext_frach<uy.bits.ext_frach ||
79 		 ux.bits.ext_fracl<uy.bits.ext_fracl))) {
80 		return x;		/* |x|<|y| return x or x-y */
81 	    }
82 	    if(ux.bits.ext_frach==uy.bits.ext_frach &&
83 		ux.bits.ext_fracl==uy.bits.ext_fracl) {
84 		return Zero[sx];	/* |x|=|y| return x*0*/
85 	    }
86 	}
87 
88     /* determine ix = ilogb(x) */
89 	if(ux.bits.ext_exp == 0) {	/* subnormal x */
90 	    ux.e *= 0x1.0p512;
91 	    ix = ux.bits.ext_exp - (BIAS + 512);
92 	} else {
93 	    ix = ux.bits.ext_exp - BIAS;
94 	}
95 
96     /* determine iy = ilogb(y) */
97 	if(uy.bits.ext_exp == 0) {	/* subnormal y */
98 	    uy.e *= 0x1.0p512;
99 	    iy = uy.bits.ext_exp - (BIAS + 512);
100 	} else {
101 	    iy = uy.bits.ext_exp - BIAS;
102 	}
103 
104     /* set up {hx,lx}, {hy,ly} and align y to x */
105 	hx = SET_NBIT(ux.bits.ext_frach);
106 	hy = SET_NBIT(uy.bits.ext_frach);
107 	lx = ux.bits.ext_fracl;
108 	ly = uy.bits.ext_fracl;
109 
110     /* fix point fmod */
111 	n = ix - iy;
112 
113 	while(n--) {
114 	    hz=hx-hy;lz=lx-ly; if(lx<ly) hz -= 1;
115 	    if(hz<0){hx = hx+hx+(lx>>MANL_SHIFT); lx = lx+lx;}
116 	    else {
117 		if ((hz|lz)==0)		/* return sign(x)*0 */
118 		    return Zero[sx];
119 		hx = hz+hz+(lz>>MANL_SHIFT); lx = lz+lz;
120 	    }
121 	}
122 	hz=hx-hy;lz=lx-ly; if(lx<ly) hz -= 1;
123 	if(hz>=0) {hx=hz;lx=lz;}
124 
125     /* convert back to floating value and restore the sign */
126 	if((hx|lx)==0)			/* return sign(x)*0 */
127 	    return Zero[sx];
128 	while(hx<(int64_t)(1ULL<<HFRAC_BITS)) {	/* normalize x */
129 	    hx = hx+hx+(lx>>MANL_SHIFT); lx = lx+lx;
130 	    iy -= 1;
131 	}
132 	ux.bits.ext_frach = hx; /* The mantissa is truncated here if needed. */
133 	ux.bits.ext_fracl = lx;
134 	if (iy < LDBL_MIN_EXP) {
135 	    ux.bits.ext_exp = iy + (BIAS + 512);
136 	    ux.e *= 0x1p-512;
137 	} else {
138 	    ux.bits.ext_exp = iy + BIAS;
139 	}
140 	x = ux.e * one;		/* create necessary signal */
141 	return x;		/* exact output */
142 }
143