9f58ba6a086c6fad7717954a5c2b8cd4805be117
[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) 1997,1998,1999,2006 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 Lesser General Public
12    License as published by the Free Software Foundation; either
13    version 2.1 of the 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    Lesser General Public License for more details.
19
20    You should have received a copy of the GNU Lesser General Public
21    License along with the GNU C Library; if not, write to the Free
22    Software Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA
23    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_SRST_1(X,S,N,sz)       __FP_FRAC_SRST_1(X##_f, S, N, sz)
44 #define _FP_FRAC_SRS_1(X,N,sz)  __FP_FRAC_SRS_1(X##_f, N, sz)
45
46 #define __FP_FRAC_SRST_1(X,S,N,sz)                      \
47 do {                                                    \
48   S = (__builtin_constant_p(N) && (N) == 1              \
49        ? X & 1 : (X << (_FP_W_TYPE_SIZE - (N))) != 0);  \
50   X = X >> (N);                                         \
51 } while (0)
52
53 #define __FP_FRAC_SRS_1(X,N,sz)                                         \
54    (X = (X >> (N) | (__builtin_constant_p(N) && (N) == 1                \
55                      ? X & 1 : (X << (_FP_W_TYPE_SIZE - (N))) != 0)))
56
57 #define _FP_FRAC_ADD_1(R,X,Y)   (R##_f = X##_f + Y##_f)
58 #define _FP_FRAC_SUB_1(R,X,Y)   (R##_f = X##_f - Y##_f)
59 #define _FP_FRAC_DEC_1(X,Y)     (X##_f -= Y##_f)
60 #define _FP_FRAC_CLZ_1(z, X)    __FP_CLZ(z, X##_f)
61
62 /* Predicates */
63 #define _FP_FRAC_NEGP_1(X)      ((_FP_WS_TYPE)X##_f < 0)
64 #define _FP_FRAC_ZEROP_1(X)     (X##_f == 0)
65 #define _FP_FRAC_OVERP_1(fs,X)  (X##_f & _FP_OVERFLOW_##fs)
66 #define _FP_FRAC_CLEAR_OVERP_1(fs,X)    (X##_f &= ~_FP_OVERFLOW_##fs)
67 #define _FP_FRAC_EQ_1(X, Y)     (X##_f == Y##_f)
68 #define _FP_FRAC_GE_1(X, Y)     (X##_f >= Y##_f)
69 #define _FP_FRAC_GT_1(X, Y)     (X##_f > Y##_f)
70
71 #define _FP_ZEROFRAC_1          0
72 #define _FP_MINFRAC_1           1
73 #define _FP_MAXFRAC_1           (~(_FP_WS_TYPE)0)
74
75 /*
76  * Unpack the raw bits of a native fp value.  Do not classify or
77  * normalize the data.
78  */
79
80 #define _FP_UNPACK_RAW_1(fs, X, val)                            \
81   do {                                                          \
82     union _FP_UNION_##fs _flo; _flo.flt = (val);                \
83                                                                 \
84     X##_f = _flo.bits.frac;                                     \
85     X##_e = _flo.bits.exp;                                      \
86     X##_s = _flo.bits.sign;                                     \
87   } while (0)
88
89 #define _FP_UNPACK_RAW_1_P(fs, X, val)                          \
90   do {                                                          \
91     union _FP_UNION_##fs *_flo =                                \
92       (union _FP_UNION_##fs *)(val);                            \
93                                                                 \
94     X##_f = _flo->bits.frac;                                    \
95     X##_e = _flo->bits.exp;                                     \
96     X##_s = _flo->bits.sign;                                    \
97   } while (0)
98
99 /*
100  * Repack the raw bits of a native fp value.
101  */
102
103 #define _FP_PACK_RAW_1(fs, val, X)                              \
104   do {                                                          \
105     union _FP_UNION_##fs _flo;                                  \
106                                                                 \
107     _flo.bits.frac = X##_f;                                     \
108     _flo.bits.exp  = X##_e;                                     \
109     _flo.bits.sign = X##_s;                                     \
110                                                                 \
111     (val) = _flo.flt;                                           \
112   } while (0)
113
114 #define _FP_PACK_RAW_1_P(fs, val, X)                            \
115   do {                                                          \
116     union _FP_UNION_##fs *_flo =                                \
117       (union _FP_UNION_##fs *)(val);                            \
118                                                                 \
119     _flo->bits.frac = X##_f;                                    \
120     _flo->bits.exp  = X##_e;                                    \
121     _flo->bits.sign = X##_s;                                    \
122   } while (0)
123
124
125 /*
126  * Multiplication algorithms:
127  */
128
129 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
130    multiplication immediately.  */
131
132 #define _FP_MUL_MEAT_1_imm(wfracbits, R, X, Y)                          \
133   do {                                                                  \
134     R##_f = X##_f * Y##_f;                                              \
135     /* Normalize since we know where the msb of the multiplicands       \
136        were (bit B), we know that the msb of the of the product is      \
137        at either 2B or 2B-1.  */                                        \
138     _FP_FRAC_SRS_1(R, wfracbits-1, 2*wfracbits);                        \
139   } while (0)
140
141 /* Given a 1W * 1W => 2W primitive, do the extended multiplication.  */
142
143 #define _FP_MUL_MEAT_1_wide(wfracbits, R, X, Y, doit)                   \
144   do {                                                                  \
145     _FP_W_TYPE _Z_f0, _Z_f1;                                            \
146     doit(_Z_f1, _Z_f0, X##_f, Y##_f);                                   \
147     /* Normalize since we know where the msb of the multiplicands       \
148        were (bit B), we know that the msb of the of the product is      \
149        at either 2B or 2B-1.  */                                        \
150     _FP_FRAC_SRS_2(_Z, wfracbits-1, 2*wfracbits);                       \
151     R##_f = _Z_f0;                                                      \
152   } while (0)
153
154 /* Finally, a simple widening multiply algorithm.  What fun!  */
155
156 #define _FP_MUL_MEAT_1_hard(wfracbits, R, X, Y)                         \
157   do {                                                                  \
158     _FP_W_TYPE _xh, _xl, _yh, _yl, _z_f0, _z_f1, _a_f0, _a_f1;          \
159                                                                         \
160     /* split the words in half */                                       \
161     _xh = X##_f >> (_FP_W_TYPE_SIZE/2);                                 \
162     _xl = X##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);         \
163     _yh = Y##_f >> (_FP_W_TYPE_SIZE/2);                                 \
164     _yl = Y##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);         \
165                                                                         \
166     /* multiply the pieces */                                           \
167     _z_f0 = _xl * _yl;                                                  \
168     _a_f0 = _xh * _yl;                                                  \
169     _a_f1 = _xl * _yh;                                                  \
170     _z_f1 = _xh * _yh;                                                  \
171                                                                         \
172     /* reassemble into two full words */                                \
173     if ((_a_f0 += _a_f1) < _a_f1)                                       \
174       _z_f1 += (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2);                    \
175     _a_f1 = _a_f0 >> (_FP_W_TYPE_SIZE/2);                               \
176     _a_f0 = _a_f0 << (_FP_W_TYPE_SIZE/2);                               \
177     _FP_FRAC_ADD_2(_z, _z, _a);                                         \
178                                                                         \
179     /* normalize */                                                     \
180     _FP_FRAC_SRS_2(_z, wfracbits - 1, 2*wfracbits);                     \
181     R##_f = _z_f0;                                                      \
182   } while (0)
183
184
185 /*
186  * Division algorithms:
187  */
188
189 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
190    division immediately.  Give this macro either _FP_DIV_HELP_imm for
191    C primitives or _FP_DIV_HELP_ldiv for the ISO function.  Which you
192    choose will depend on what the compiler does with divrem4.  */
193
194 #define _FP_DIV_MEAT_1_imm(fs, R, X, Y, doit)           \
195   do {                                                  \
196     _FP_W_TYPE _q, _r;                                  \
197     X##_f <<= (X##_f < Y##_f                            \
198                ? R##_e--, _FP_WFRACBITS_##fs            \
199                : _FP_WFRACBITS_##fs - 1);               \
200     doit(_q, _r, X##_f, Y##_f);                         \
201     R##_f = _q | (_r != 0);                             \
202   } while (0)
203
204 /* GCC's longlong.h defines a 2W / 1W => (1W,1W) primitive udiv_qrnnd
205    that may be useful in this situation.  This first is for a primitive
206    that requires normalization, the second for one that does not.  Look
207    for UDIV_NEEDS_NORMALIZATION to tell which your machine needs.  */
208
209 #define _FP_DIV_MEAT_1_udiv_norm(fs, R, X, Y)                           \
210   do {                                                                  \
211     _FP_W_TYPE _nh, _nl, _q, _r, _y;                                    \
212                                                                         \
213     /* Normalize Y -- i.e. make the most significant bit set.  */       \
214     _y = Y##_f << _FP_WFRACXBITS_##fs;                                  \
215                                                                         \
216     /* Shift X op correspondingly high, that is, up one full word.  */  \
217     if (X##_f < Y##_f)                                                  \
218       {                                                                 \
219         R##_e--;                                                        \
220         _nl = 0;                                                        \
221         _nh = X##_f;                                                    \
222       }                                                                 \
223     else                                                                \
224       {                                                                 \
225         _nl = X##_f << (_FP_W_TYPE_SIZE - 1);                           \
226         _nh = X##_f >> 1;                                               \
227       }                                                                 \
228                                                                         \
229     udiv_qrnnd(_q, _r, _nh, _nl, _y);                                   \
230     R##_f = _q | (_r != 0);                                             \
231   } while (0)
232
233 #define _FP_DIV_MEAT_1_udiv(fs, R, X, Y)                \
234   do {                                                  \
235     _FP_W_TYPE _nh, _nl, _q, _r;                        \
236     if (X##_f < Y##_f)                                  \
237       {                                                 \
238         R##_e--;                                        \
239         _nl = X##_f << _FP_WFRACBITS_##fs;              \
240         _nh = X##_f >> _FP_WFRACXBITS_##fs;             \
241       }                                                 \
242     else                                                \
243       {                                                 \
244         _nl = X##_f << (_FP_WFRACBITS_##fs - 1);        \
245         _nh = X##_f >> (_FP_WFRACXBITS_##fs + 1);       \
246       }                                                 \
247     udiv_qrnnd(_q, _r, _nh, _nl, Y##_f);                \
248     R##_f = _q | (_r != 0);                             \
249   } while (0)
250   
251   
252 /*
253  * Square root algorithms:
254  * We have just one right now, maybe Newton approximation
255  * should be added for those machines where division is fast.
256  */
257  
258 #define _FP_SQRT_MEAT_1(R, S, T, X, q)                  \
259   do {                                                  \
260     while (q != _FP_WORK_ROUND)                         \
261       {                                                 \
262         T##_f = S##_f + q;                              \
263         if (T##_f <= X##_f)                             \
264           {                                             \
265             S##_f = T##_f + q;                          \
266             X##_f -= T##_f;                             \
267             R##_f += q;                                 \
268           }                                             \
269         _FP_FRAC_SLL_1(X, 1);                           \
270         q >>= 1;                                        \
271       }                                                 \
272     if (X##_f)                                          \
273       {                                                 \
274         if (S##_f < X##_f)                              \
275           R##_f |= _FP_WORK_ROUND;                      \
276         R##_f |= _FP_WORK_STICKY;                       \
277       }                                                 \
278   } while (0)
279
280 /*
281  * Assembly/disassembly for converting to/from integral types.  
282  * No shifting or overflow handled here.
283  */
284
285 #define _FP_FRAC_ASSEMBLE_1(r, X, rsize)        (r = X##_f)
286 #define _FP_FRAC_DISASSEMBLE_1(X, r, rsize)     (X##_f = r)
287
288
289 /*
290  * Convert FP values between word sizes
291  */
292
293 #define _FP_FRAC_COPY_1_1(D, S)         (D##_f = S##_f)