Moved to sysdeps/sparc/sparc32/soft-fp/q_sub.c.
[kopensolaris-gnu/glibc.git] / soft-fp / op-1.h
1 /* Software floating-point emulation.
2    Basic one-word fraction declaration and manipulation.
3    Copyright (C) 1997,1998,1999 Free Software Foundation, Inc.
4    This file is part of the GNU C Library.
5    Contributed by Richard Henderson (rth@cygnus.com),
6                   Jakub Jelinek (jj@ultra.linux.cz),
7                   David S. Miller (davem@redhat.com) and
8                   Peter Maydell (pmaydell@chiark.greenend.org.uk).
9
10    The GNU C Library is free software; you can redistribute it and/or
11    modify it under the terms of the GNU Library General Public License as
12    published by the Free Software Foundation; either version 2 of the
13    License, or (at your option) any later version.
14
15    The GNU C Library is distributed in the hope that it will be useful,
16    but WITHOUT ANY WARRANTY; without even the implied warranty of
17    MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
18    Library General Public License for more details.
19
20    You should have received a copy of the GNU Library General Public
21    License along with the GNU C Library; see the file COPYING.LIB.  If
22    not, write to the Free Software Foundation, Inc.,
23    59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.  */
24
25 #define _FP_FRAC_DECL_1(X)      _FP_W_TYPE X##_f
26 #define _FP_FRAC_COPY_1(D,S)    (D##_f = S##_f)
27 #define _FP_FRAC_SET_1(X,I)     (X##_f = I)
28 #define _FP_FRAC_HIGH_1(X)      (X##_f)
29 #define _FP_FRAC_LOW_1(X)       (X##_f)
30 #define _FP_FRAC_WORD_1(X,w)    (X##_f)
31
32 #define _FP_FRAC_ADDI_1(X,I)    (X##_f += I)
33 #define _FP_FRAC_SLL_1(X,N)                     \
34   do {                                          \
35     if (__builtin_constant_p(N) && (N) == 1)    \
36       X##_f += X##_f;                           \
37     else                                        \
38       X##_f <<= (N);                            \
39   } while (0)
40 #define _FP_FRAC_SRL_1(X,N)     (X##_f >>= N)
41
42 /* Right shift with sticky-lsb.  */
43 #define _FP_FRAC_SRS_1(X,N,sz)  __FP_FRAC_SRS_1(X##_f, N, sz)
44
45 #define __FP_FRAC_SRS_1(X,N,sz)                                         \
46    (X = (X >> (N) | (__builtin_constant_p(N) && (N) == 1                \
47                      ? X & 1 : (X << (_FP_W_TYPE_SIZE - (N))) != 0)))
48
49 #define _FP_FRAC_ADD_1(R,X,Y)   (R##_f = X##_f + Y##_f)
50 #define _FP_FRAC_SUB_1(R,X,Y)   (R##_f = X##_f - Y##_f)
51 #define _FP_FRAC_DEC_1(X,Y)     (X##_f -= Y##_f)
52 #define _FP_FRAC_CLZ_1(z, X)    __FP_CLZ(z, X##_f)
53
54 /* Predicates */
55 #define _FP_FRAC_NEGP_1(X)      ((_FP_WS_TYPE)X##_f < 0)
56 #define _FP_FRAC_ZEROP_1(X)     (X##_f == 0)
57 #define _FP_FRAC_OVERP_1(fs,X)  (X##_f & _FP_OVERFLOW_##fs)
58 #define _FP_FRAC_EQ_1(X, Y)     (X##_f == Y##_f)
59 #define _FP_FRAC_GE_1(X, Y)     (X##_f >= Y##_f)
60 #define _FP_FRAC_GT_1(X, Y)     (X##_f > Y##_f)
61
62 #define _FP_ZEROFRAC_1          0
63 #define _FP_MINFRAC_1           1
64 #define _FP_MAXFRAC_1           (~(_FP_WS_TYPE)0)
65
66 /*
67  * Unpack the raw bits of a native fp value.  Do not classify or
68  * normalize the data.
69  */
70
71 #define _FP_UNPACK_RAW_1(fs, X, val)                            \
72   do {                                                          \
73     union _FP_UNION_##fs _flo; _flo.flt = (val);                \
74                                                                 \
75     X##_f = _flo.bits.frac;                                     \
76     X##_e = _flo.bits.exp;                                      \
77     X##_s = _flo.bits.sign;                                     \
78   } while (0)
79
80 #define _FP_UNPACK_RAW_1_P(fs, X, val)                          \
81   do {                                                          \
82     union _FP_UNION_##fs *_flo =                                \
83       (union _FP_UNION_##fs *)(val);                            \
84                                                                 \
85     X##_f = _flo->bits.frac;                                    \
86     X##_e = _flo->bits.exp;                                     \
87     X##_s = _flo->bits.sign;                                    \
88   } while (0)
89
90 /*
91  * Repack the raw bits of a native fp value.
92  */
93
94 #define _FP_PACK_RAW_1(fs, val, X)                              \
95   do {                                                          \
96     union _FP_UNION_##fs _flo;                                  \
97                                                                 \
98     _flo.bits.frac = X##_f;                                     \
99     _flo.bits.exp  = X##_e;                                     \
100     _flo.bits.sign = X##_s;                                     \
101                                                                 \
102     (val) = _flo.flt;                                           \
103   } while (0)
104
105 #define _FP_PACK_RAW_1_P(fs, val, X)                            \
106   do {                                                          \
107     union _FP_UNION_##fs *_flo =                                \
108       (union _FP_UNION_##fs *)(val);                            \
109                                                                 \
110     _flo->bits.frac = X##_f;                                    \
111     _flo->bits.exp  = X##_e;                                    \
112     _flo->bits.sign = X##_s;                                    \
113   } while (0)
114
115
116 /*
117  * Multiplication algorithms:
118  */
119
120 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
121    multiplication immediately.  */
122
123 #define _FP_MUL_MEAT_1_imm(wfracbits, R, X, Y)                          \
124   do {                                                                  \
125     R##_f = X##_f * Y##_f;                                              \
126     /* Normalize since we know where the msb of the multiplicands       \
127        were (bit B), we know that the msb of the of the product is      \
128        at either 2B or 2B-1.  */                                        \
129     _FP_FRAC_SRS_1(R, wfracbits-1, 2*wfracbits);                        \
130   } while (0)
131
132 /* Given a 1W * 1W => 2W primitive, do the extended multiplication.  */
133
134 #define _FP_MUL_MEAT_1_wide(wfracbits, R, X, Y, doit)                   \
135   do {                                                                  \
136     _FP_W_TYPE _Z_f0, _Z_f1;                                            \
137     doit(_Z_f1, _Z_f0, X##_f, Y##_f);                                   \
138     /* Normalize since we know where the msb of the multiplicands       \
139        were (bit B), we know that the msb of the of the product is      \
140        at either 2B or 2B-1.  */                                        \
141     _FP_FRAC_SRS_2(_Z, wfracbits-1, 2*wfracbits);                       \
142     R##_f = _Z_f0;                                                      \
143   } while (0)
144
145 /* Finally, a simple widening multiply algorithm.  What fun!  */
146
147 #define _FP_MUL_MEAT_1_hard(wfracbits, R, X, Y)                         \
148   do {                                                                  \
149     _FP_W_TYPE _xh, _xl, _yh, _yl, _z_f0, _z_f1, _a_f0, _a_f1;          \
150                                                                         \
151     /* split the words in half */                                       \
152     _xh = X##_f >> (_FP_W_TYPE_SIZE/2);                                 \
153     _xl = X##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);         \
154     _yh = Y##_f >> (_FP_W_TYPE_SIZE/2);                                 \
155     _yl = Y##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);         \
156                                                                         \
157     /* multiply the pieces */                                           \
158     _z_f0 = _xl * _yl;                                                  \
159     _a_f0 = _xh * _yl;                                                  \
160     _a_f1 = _xl * _yh;                                                  \
161     _z_f1 = _xh * _yh;                                                  \
162                                                                         \
163     /* reassemble into two full words */                                \
164     if ((_a_f0 += _a_f1) < _a_f1)                                       \
165       _z_f1 += (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2);                    \
166     _a_f1 = _a_f0 >> (_FP_W_TYPE_SIZE/2);                               \
167     _a_f0 = _a_f0 << (_FP_W_TYPE_SIZE/2);                               \
168     _FP_FRAC_ADD_2(_z, _z, _a);                                         \
169                                                                         \
170     /* normalize */                                                     \
171     _FP_FRAC_SRS_2(_z, wfracbits - 1, 2*wfracbits);                     \
172     R##_f = _z_f0;                                                      \
173   } while (0)
174
175
176 /*
177  * Division algorithms:
178  */
179
180 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
181    division immediately.  Give this macro either _FP_DIV_HELP_imm for
182    C primitives or _FP_DIV_HELP_ldiv for the ISO function.  Which you
183    choose will depend on what the compiler does with divrem4.  */
184
185 #define _FP_DIV_MEAT_1_imm(fs, R, X, Y, doit)           \
186   do {                                                  \
187     _FP_W_TYPE _q, _r;                                  \
188     X##_f <<= (X##_f < Y##_f                            \
189                ? R##_e--, _FP_WFRACBITS_##fs            \
190                : _FP_WFRACBITS_##fs - 1);               \
191     doit(_q, _r, X##_f, Y##_f);                         \
192     R##_f = _q | (_r != 0);                             \
193   } while (0)
194
195 /* GCC's longlong.h defines a 2W / 1W => (1W,1W) primitive udiv_qrnnd
196    that may be useful in this situation.  This first is for a primitive
197    that requires normalization, the second for one that does not.  Look
198    for UDIV_NEEDS_NORMALIZATION to tell which your machine needs.  */
199
200 #define _FP_DIV_MEAT_1_udiv_norm(fs, R, X, Y)                           \
201   do {                                                                  \
202     _FP_W_TYPE _nh, _nl, _q, _r, _y;                                    \
203                                                                         \
204     /* Normalize Y -- i.e. make the most significant bit set.  */       \
205     _y = Y##_f << _FP_WFRACXBITS_##fs;                                  \
206                                                                         \
207     /* Shift X op correspondingly high, that is, up one full word.  */  \
208     if (X##_f < Y##_f)                                                  \
209       {                                                                 \
210         R##_e--;                                                        \
211         _nl = 0;                                                        \
212         _nh = X##_f;                                                    \
213       }                                                                 \
214     else                                                                \
215       {                                                                 \
216         _nl = X##_f << (_FP_W_TYPE_SIZE - 1);                           \
217         _nh = X##_f >> 1;                                               \
218       }                                                                 \
219                                                                         \
220     udiv_qrnnd(_q, _r, _nh, _nl, _y);                                   \
221     R##_f = _q | (_r != 0);                                             \
222   } while (0)
223
224 #define _FP_DIV_MEAT_1_udiv(fs, R, X, Y)                \
225   do {                                                  \
226     _FP_W_TYPE _nh, _nl, _q, _r;                        \
227     if (X##_f < Y##_f)                                  \
228       {                                                 \
229         R##_e--;                                        \
230         _nl = X##_f << _FP_WFRACBITS_##fs;              \
231         _nh = X##_f >> _FP_WFRACXBITS_##fs;             \
232       }                                                 \
233     else                                                \
234       {                                                 \
235         _nl = X##_f << (_FP_WFRACBITS_##fs - 1);        \
236         _nh = X##_f >> (_FP_WFRACXBITS_##fs + 1);       \
237       }                                                 \
238     udiv_qrnnd(_q, _r, _nh, _nl, Y##_f);                \
239     R##_f = _q | (_r != 0);                             \
240   } while (0)
241   
242   
243 /*
244  * Square root algorithms:
245  * We have just one right now, maybe Newton approximation
246  * should be added for those machines where division is fast.
247  */
248  
249 #define _FP_SQRT_MEAT_1(R, S, T, X, q)                  \
250   do {                                                  \
251     while (q != _FP_WORK_ROUND)                         \
252       {                                                 \
253         T##_f = S##_f + q;                              \
254         if (T##_f <= X##_f)                             \
255           {                                             \
256             S##_f = T##_f + q;                          \
257             X##_f -= T##_f;                             \
258             R##_f += q;                                 \
259           }                                             \
260         _FP_FRAC_SLL_1(X, 1);                           \
261         q >>= 1;                                        \
262       }                                                 \
263     if (X##_f)                                          \
264       {                                                 \
265         if (S##_f < X##_f)                              \
266           R##_f |= _FP_WORK_ROUND;                      \
267         R##_f |= _FP_WORK_STICKY;                       \
268       }                                                 \
269   } while (0)
270
271 /*
272  * Assembly/disassembly for converting to/from integral types.  
273  * No shifting or overflow handled here.
274  */
275
276 #define _FP_FRAC_ASSEMBLE_1(r, X, rsize)        (r = X##_f)
277 #define _FP_FRAC_DISASSEMBLE_1(X, r, rsize)     (X##_f = r)
278
279
280 /*
281  * Convert FP values between word sizes
282  */
283
284 #define _FP_FRAC_CONV_1_1(dfs, sfs, D, S)                               \
285   do {                                                                  \
286     D##_f = S##_f;                                                      \
287     if (_FP_WFRACBITS_##sfs > _FP_WFRACBITS_##dfs)                      \
288       {                                                                 \
289         if (S##_c != FP_CLS_NAN)                                        \
290           _FP_FRAC_SRS_1(D, (_FP_WFRACBITS_##sfs-_FP_WFRACBITS_##dfs),  \
291                          _FP_WFRACBITS_##sfs);                          \
292         else                                                            \
293           _FP_FRAC_SRL_1(D, (_FP_WFRACBITS_##sfs-_FP_WFRACBITS_##dfs)); \
294       }                                                                 \
295     else                                                                \
296       D##_f <<= _FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs;              \
297   } while (0)