1 |
3 |
xianfeng |
/*
|
2 |
|
|
* Basic four-word fraction declaration and manipulation.
|
3 |
|
|
*
|
4 |
|
|
* When adding quadword support for 32 bit machines, we need
|
5 |
|
|
* to be a little careful as double multiply uses some of these
|
6 |
|
|
* macros: (in op-2.h)
|
7 |
|
|
* _FP_MUL_MEAT_2_wide() uses _FP_FRAC_DECL_4, _FP_FRAC_WORD_4,
|
8 |
|
|
* _FP_FRAC_ADD_4, _FP_FRAC_SRS_4
|
9 |
|
|
* _FP_MUL_MEAT_2_gmp() uses _FP_FRAC_SRS_4 (and should use
|
10 |
|
|
* _FP_FRAC_DECL_4: it appears to be broken and is not used
|
11 |
|
|
* anywhere anyway. )
|
12 |
|
|
*
|
13 |
|
|
* I've now fixed all the macros that were here from the sparc64 code.
|
14 |
|
|
* [*none* of the shift macros were correct!] -- PMM 02/1998
|
15 |
|
|
*
|
16 |
|
|
* The only quadword stuff that remains to be coded is:
|
17 |
|
|
* 1) the conversion to/from ints, which requires
|
18 |
|
|
* that we check (in op-common.h) that the following do the right thing
|
19 |
|
|
* for quadwords: _FP_TO_INT(Q,4,r,X,rsz,rsg), _FP_FROM_INT(Q,4,X,r,rs,rt)
|
20 |
|
|
* 2) multiply, divide and sqrt, which require:
|
21 |
|
|
* _FP_MUL_MEAT_4_*(R,X,Y), _FP_DIV_MEAT_4_*(R,X,Y), _FP_SQRT_MEAT_4(R,S,T,X,q),
|
22 |
|
|
* This also needs _FP_MUL_MEAT_Q and _FP_DIV_MEAT_Q to be defined to
|
23 |
|
|
* some suitable _FP_MUL_MEAT_4_* macros in sfp-machine.h.
|
24 |
|
|
* [we're free to choose whatever FP_MUL_MEAT_4_* macros we need for
|
25 |
|
|
* these; they are used nowhere else. ]
|
26 |
|
|
*/
|
27 |
|
|
|
28 |
|
|
#define _FP_FRAC_DECL_4(X) _FP_W_TYPE X##_f[4]
|
29 |
|
|
#define _FP_FRAC_COPY_4(D,S) \
|
30 |
|
|
(D##_f[0] = S##_f[0], D##_f[1] = S##_f[1], \
|
31 |
|
|
D##_f[2] = S##_f[2], D##_f[3] = S##_f[3])
|
32 |
|
|
/* The _FP_FRAC_SET_n(X,I) macro is intended for use with another
|
33 |
|
|
* macro such as _FP_ZEROFRAC_n which returns n comma separated values.
|
34 |
|
|
* The result is that we get an expansion of __FP_FRAC_SET_n(X,I0,I1,I2,I3)
|
35 |
|
|
* which just assigns the In values to the array X##_f[].
|
36 |
|
|
* This is why the number of parameters doesn't appear to match
|
37 |
|
|
* at first glance... -- PMM
|
38 |
|
|
*/
|
39 |
|
|
#define _FP_FRAC_SET_4(X,I) __FP_FRAC_SET_4(X, I)
|
40 |
|
|
#define _FP_FRAC_HIGH_4(X) (X##_f[3])
|
41 |
|
|
#define _FP_FRAC_LOW_4(X) (X##_f[0])
|
42 |
|
|
#define _FP_FRAC_WORD_4(X,w) (X##_f[w])
|
43 |
|
|
|
44 |
|
|
#define _FP_FRAC_SLL_4(X,N) \
|
45 |
|
|
do { \
|
46 |
|
|
_FP_I_TYPE _up, _down, _skip, _i; \
|
47 |
|
|
_skip = (N) / _FP_W_TYPE_SIZE; \
|
48 |
|
|
_up = (N) % _FP_W_TYPE_SIZE; \
|
49 |
|
|
_down = _FP_W_TYPE_SIZE - _up; \
|
50 |
|
|
for (_i = 3; _i > _skip; --_i) \
|
51 |
|
|
X##_f[_i] = X##_f[_i-_skip] << _up | X##_f[_i-_skip-1] >> _down; \
|
52 |
|
|
/* bugfixed: was X##_f[_i] <<= _up; -- PMM 02/1998 */ \
|
53 |
|
|
X##_f[_i] = X##_f[0] << _up; \
|
54 |
|
|
for (--_i; _i >= 0; --_i) \
|
55 |
|
|
X##_f[_i] = 0; \
|
56 |
|
|
} while (0)
|
57 |
|
|
|
58 |
|
|
/* This one was broken too */
|
59 |
|
|
#define _FP_FRAC_SRL_4(X,N) \
|
60 |
|
|
do { \
|
61 |
|
|
_FP_I_TYPE _up, _down, _skip, _i; \
|
62 |
|
|
_skip = (N) / _FP_W_TYPE_SIZE; \
|
63 |
|
|
_down = (N) % _FP_W_TYPE_SIZE; \
|
64 |
|
|
_up = _FP_W_TYPE_SIZE - _down; \
|
65 |
|
|
for (_i = 0; _i < 3-_skip; ++_i) \
|
66 |
|
|
X##_f[_i] = X##_f[_i+_skip] >> _down | X##_f[_i+_skip+1] << _up; \
|
67 |
|
|
X##_f[_i] = X##_f[3] >> _down; \
|
68 |
|
|
for (++_i; _i < 4; ++_i) \
|
69 |
|
|
X##_f[_i] = 0; \
|
70 |
|
|
} while (0)
|
71 |
|
|
|
72 |
|
|
|
73 |
|
|
/* Right shift with sticky-lsb.
|
74 |
|
|
* What this actually means is that we do a standard right-shift,
|
75 |
|
|
* but that if any of the bits that fall off the right hand side
|
76 |
|
|
* were one then we always set the LSbit.
|
77 |
|
|
*/
|
78 |
|
|
#define _FP_FRAC_SRS_4(X,N,size) \
|
79 |
|
|
do { \
|
80 |
|
|
_FP_I_TYPE _up, _down, _skip, _i; \
|
81 |
|
|
_FP_W_TYPE _s; \
|
82 |
|
|
_skip = (N) / _FP_W_TYPE_SIZE; \
|
83 |
|
|
_down = (N) % _FP_W_TYPE_SIZE; \
|
84 |
|
|
_up = _FP_W_TYPE_SIZE - _down; \
|
85 |
|
|
for (_s = _i = 0; _i < _skip; ++_i) \
|
86 |
|
|
_s |= X##_f[_i]; \
|
87 |
|
|
_s |= X##_f[_i] << _up; \
|
88 |
|
|
/* s is now != 0 if we want to set the LSbit */ \
|
89 |
|
|
for (_i = 0; _i < 3-_skip; ++_i) \
|
90 |
|
|
X##_f[_i] = X##_f[_i+_skip] >> _down | X##_f[_i+_skip+1] << _up; \
|
91 |
|
|
X##_f[_i] = X##_f[3] >> _down; \
|
92 |
|
|
for (++_i; _i < 4; ++_i) \
|
93 |
|
|
X##_f[_i] = 0; \
|
94 |
|
|
/* don't fix the LSB until the very end when we're sure f[0] is stable */ \
|
95 |
|
|
X##_f[0] |= (_s != 0); \
|
96 |
|
|
} while (0)
|
97 |
|
|
|
98 |
|
|
#define _FP_FRAC_ADD_4(R,X,Y) \
|
99 |
|
|
__FP_FRAC_ADD_4(R##_f[3], R##_f[2], R##_f[1], R##_f[0], \
|
100 |
|
|
X##_f[3], X##_f[2], X##_f[1], X##_f[0], \
|
101 |
|
|
Y##_f[3], Y##_f[2], Y##_f[1], Y##_f[0])
|
102 |
|
|
|
103 |
|
|
#define _FP_FRAC_SUB_4(R,X,Y) \
|
104 |
|
|
__FP_FRAC_SUB_4(R##_f[3], R##_f[2], R##_f[1], R##_f[0], \
|
105 |
|
|
X##_f[3], X##_f[2], X##_f[1], X##_f[0], \
|
106 |
|
|
Y##_f[3], Y##_f[2], Y##_f[1], Y##_f[0])
|
107 |
|
|
|
108 |
|
|
#define _FP_FRAC_ADDI_4(X,I) \
|
109 |
|
|
__FP_FRAC_ADDI_4(X##_f[3], X##_f[2], X##_f[1], X##_f[0], I)
|
110 |
|
|
|
111 |
|
|
#define _FP_ZEROFRAC_4 0,0,0,0
|
112 |
|
|
#define _FP_MINFRAC_4 0,0,0,1
|
113 |
|
|
|
114 |
|
|
#define _FP_FRAC_ZEROP_4(X) ((X##_f[0] | X##_f[1] | X##_f[2] | X##_f[3]) == 0)
|
115 |
|
|
#define _FP_FRAC_NEGP_4(X) ((_FP_WS_TYPE)X##_f[3] < 0)
|
116 |
|
|
#define _FP_FRAC_OVERP_4(fs,X) (X##_f[0] & _FP_OVERFLOW_##fs)
|
117 |
|
|
|
118 |
|
|
#define _FP_FRAC_EQ_4(X,Y) \
|
119 |
|
|
(X##_f[0] == Y##_f[0] && X##_f[1] == Y##_f[1] \
|
120 |
|
|
&& X##_f[2] == Y##_f[2] && X##_f[3] == Y##_f[3])
|
121 |
|
|
|
122 |
|
|
#define _FP_FRAC_GT_4(X,Y) \
|
123 |
|
|
(X##_f[3] > Y##_f[3] || \
|
124 |
|
|
(X##_f[3] == Y##_f[3] && (X##_f[2] > Y##_f[2] || \
|
125 |
|
|
(X##_f[2] == Y##_f[2] && (X##_f[1] > Y##_f[1] || \
|
126 |
|
|
(X##_f[1] == Y##_f[1] && X##_f[0] > Y##_f[0]) \
|
127 |
|
|
)) \
|
128 |
|
|
)) \
|
129 |
|
|
)
|
130 |
|
|
|
131 |
|
|
#define _FP_FRAC_GE_4(X,Y) \
|
132 |
|
|
(X##_f[3] > Y##_f[3] || \
|
133 |
|
|
(X##_f[3] == Y##_f[3] && (X##_f[2] > Y##_f[2] || \
|
134 |
|
|
(X##_f[2] == Y##_f[2] && (X##_f[1] > Y##_f[1] || \
|
135 |
|
|
(X##_f[1] == Y##_f[1] && X##_f[0] >= Y##_f[0]) \
|
136 |
|
|
)) \
|
137 |
|
|
)) \
|
138 |
|
|
)
|
139 |
|
|
|
140 |
|
|
|
141 |
|
|
#define _FP_FRAC_CLZ_4(R,X) \
|
142 |
|
|
do { \
|
143 |
|
|
if (X##_f[3]) \
|
144 |
|
|
{ \
|
145 |
|
|
__FP_CLZ(R,X##_f[3]); \
|
146 |
|
|
} \
|
147 |
|
|
else if (X##_f[2]) \
|
148 |
|
|
{ \
|
149 |
|
|
__FP_CLZ(R,X##_f[2]); \
|
150 |
|
|
R += _FP_W_TYPE_SIZE; \
|
151 |
|
|
} \
|
152 |
|
|
else if (X##_f[1]) \
|
153 |
|
|
{ \
|
154 |
|
|
__FP_CLZ(R,X##_f[2]); \
|
155 |
|
|
R += _FP_W_TYPE_SIZE*2; \
|
156 |
|
|
} \
|
157 |
|
|
else \
|
158 |
|
|
{ \
|
159 |
|
|
__FP_CLZ(R,X##_f[0]); \
|
160 |
|
|
R += _FP_W_TYPE_SIZE*3; \
|
161 |
|
|
} \
|
162 |
|
|
} while(0)
|
163 |
|
|
|
164 |
|
|
|
165 |
|
|
#define _FP_UNPACK_RAW_4(fs, X, val) \
|
166 |
|
|
do { \
|
167 |
|
|
union _FP_UNION_##fs _flo; _flo.flt = (val); \
|
168 |
|
|
X##_f[0] = _flo.bits.frac0; \
|
169 |
|
|
X##_f[1] = _flo.bits.frac1; \
|
170 |
|
|
X##_f[2] = _flo.bits.frac2; \
|
171 |
|
|
X##_f[3] = _flo.bits.frac3; \
|
172 |
|
|
X##_e = _flo.bits.exp; \
|
173 |
|
|
X##_s = _flo.bits.sign; \
|
174 |
|
|
} while (0)
|
175 |
|
|
|
176 |
|
|
#define _FP_PACK_RAW_4(fs, val, X) \
|
177 |
|
|
do { \
|
178 |
|
|
union _FP_UNION_##fs _flo; \
|
179 |
|
|
_flo.bits.frac0 = X##_f[0]; \
|
180 |
|
|
_flo.bits.frac1 = X##_f[1]; \
|
181 |
|
|
_flo.bits.frac2 = X##_f[2]; \
|
182 |
|
|
_flo.bits.frac3 = X##_f[3]; \
|
183 |
|
|
_flo.bits.exp = X##_e; \
|
184 |
|
|
_flo.bits.sign = X##_s; \
|
185 |
|
|
(val) = _flo.flt; \
|
186 |
|
|
} while (0)
|
187 |
|
|
|
188 |
|
|
|
189 |
|
|
/*
|
190 |
|
|
* Internals
|
191 |
|
|
*/
|
192 |
|
|
|
193 |
|
|
#define __FP_FRAC_SET_4(X,I3,I2,I1,I0) \
|
194 |
|
|
(X##_f[3] = I3, X##_f[2] = I2, X##_f[1] = I1, X##_f[0] = I0)
|
195 |
|
|
|
196 |
|
|
#ifndef __FP_FRAC_ADD_4
|
197 |
|
|
#define __FP_FRAC_ADD_4(r3,r2,r1,r0,x3,x2,x1,x0,y3,y2,y1,y0) \
|
198 |
|
|
(r0 = x0 + y0, \
|
199 |
|
|
r1 = x1 + y1 + (r0 < x0), \
|
200 |
|
|
r2 = x2 + y2 + (r1 < x1), \
|
201 |
|
|
r3 = x3 + y3 + (r2 < x2))
|
202 |
|
|
#endif
|
203 |
|
|
|
204 |
|
|
#ifndef __FP_FRAC_SUB_4
|
205 |
|
|
#define __FP_FRAC_SUB_4(r3,r2,r1,r0,x3,x2,x1,x0,y3,y2,y1,y0) \
|
206 |
|
|
(r0 = x0 - y0, \
|
207 |
|
|
r1 = x1 - y1 - (r0 > x0), \
|
208 |
|
|
r2 = x2 - y2 - (r1 > x1), \
|
209 |
|
|
r3 = x3 - y3 - (r2 > x2))
|
210 |
|
|
#endif
|
211 |
|
|
|
212 |
|
|
#ifndef __FP_FRAC_ADDI_4
|
213 |
|
|
/* I always wanted to be a lisp programmer :-> */
|
214 |
|
|
#define __FP_FRAC_ADDI_4(x3,x2,x1,x0,i) \
|
215 |
|
|
(x3 += ((x2 += ((x1 += ((x0 += i) < x0)) < x1) < x2)))
|
216 |
|
|
#endif
|
217 |
|
|
|
218 |
|
|
/* Convert FP values between word sizes. This appears to be more
|
219 |
|
|
* complicated than I'd have expected it to be, so these might be
|
220 |
|
|
* wrong... These macros are in any case somewhat bogus because they
|
221 |
|
|
* use information about what various FRAC_n variables look like
|
222 |
|
|
* internally [eg, that 2 word vars are X_f0 and x_f1]. But so do
|
223 |
|
|
* the ones in op-2.h and op-1.h.
|
224 |
|
|
*/
|
225 |
|
|
#define _FP_FRAC_CONV_1_4(dfs, sfs, D, S) \
|
226 |
|
|
do { \
|
227 |
|
|
_FP_FRAC_SRS_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs), \
|
228 |
|
|
_FP_WFRACBITS_##sfs); \
|
229 |
|
|
D##_f = S##_f[0]; \
|
230 |
|
|
} while (0)
|
231 |
|
|
|
232 |
|
|
#define _FP_FRAC_CONV_2_4(dfs, sfs, D, S) \
|
233 |
|
|
do { \
|
234 |
|
|
_FP_FRAC_SRS_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs), \
|
235 |
|
|
_FP_WFRACBITS_##sfs); \
|
236 |
|
|
D##_f0 = S##_f[0]; \
|
237 |
|
|
D##_f1 = S##_f[1]; \
|
238 |
|
|
} while (0)
|
239 |
|
|
|
240 |
|
|
/* Assembly/disassembly for converting to/from integral types.
|
241 |
|
|
* No shifting or overflow handled here.
|
242 |
|
|
*/
|
243 |
|
|
/* Put the FP value X into r, which is an integer of size rsize. */
|
244 |
|
|
#define _FP_FRAC_ASSEMBLE_4(r, X, rsize) \
|
245 |
|
|
do { \
|
246 |
|
|
if (rsize <= _FP_W_TYPE_SIZE) \
|
247 |
|
|
r = X##_f[0]; \
|
248 |
|
|
else if (rsize <= 2*_FP_W_TYPE_SIZE) \
|
249 |
|
|
{ \
|
250 |
|
|
r = X##_f[1]; \
|
251 |
|
|
r <<= _FP_W_TYPE_SIZE; \
|
252 |
|
|
r += X##_f[0]; \
|
253 |
|
|
} \
|
254 |
|
|
else \
|
255 |
|
|
{ \
|
256 |
|
|
/* I'm feeling lazy so we deal with int == 3words (implausible)*/ \
|
257 |
|
|
/* and int == 4words as a single case. */ \
|
258 |
|
|
r = X##_f[3]; \
|
259 |
|
|
r <<= _FP_W_TYPE_SIZE; \
|
260 |
|
|
r += X##_f[2]; \
|
261 |
|
|
r <<= _FP_W_TYPE_SIZE; \
|
262 |
|
|
r += X##_f[1]; \
|
263 |
|
|
r <<= _FP_W_TYPE_SIZE; \
|
264 |
|
|
r += X##_f[0]; \
|
265 |
|
|
} \
|
266 |
|
|
} while (0)
|
267 |
|
|
|
268 |
|
|
/* "No disassemble Number Five!" */
|
269 |
|
|
/* move an integer of size rsize into X's fractional part. We rely on
|
270 |
|
|
* the _f[] array consisting of words of size _FP_W_TYPE_SIZE to avoid
|
271 |
|
|
* having to mask the values we store into it.
|
272 |
|
|
*/
|
273 |
|
|
#define _FP_FRAC_DISASSEMBLE_4(X, r, rsize) \
|
274 |
|
|
do { \
|
275 |
|
|
X##_f[0] = r; \
|
276 |
|
|
X##_f[1] = (rsize <= _FP_W_TYPE_SIZE ? 0 : r >> _FP_W_TYPE_SIZE); \
|
277 |
|
|
X##_f[2] = (rsize <= 2*_FP_W_TYPE_SIZE ? 0 : r >> 2*_FP_W_TYPE_SIZE); \
|
278 |
|
|
X##_f[3] = (rsize <= 3*_FP_W_TYPE_SIZE ? 0 : r >> 3*_FP_W_TYPE_SIZE); \
|
279 |
|
|
} while (0)
|
280 |
|
|
|
281 |
|
|
#define _FP_FRAC_CONV_4_1(dfs, sfs, D, S) \
|
282 |
|
|
do { \
|
283 |
|
|
D##_f[0] = S##_f; \
|
284 |
|
|
D##_f[1] = D##_f[2] = D##_f[3] = 0; \
|
285 |
|
|
_FP_FRAC_SLL_4(D, (_FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs)); \
|
286 |
|
|
} while (0)
|
287 |
|
|
|
288 |
|
|
#define _FP_FRAC_CONV_4_2(dfs, sfs, D, S) \
|
289 |
|
|
do { \
|
290 |
|
|
D##_f[0] = S##_f0; \
|
291 |
|
|
D##_f[1] = S##_f1; \
|
292 |
|
|
D##_f[2] = D##_f[3] = 0; \
|
293 |
|
|
_FP_FRAC_SLL_4(D, (_FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs)); \
|
294 |
|
|
} while (0)
|
295 |
|
|
|
296 |
|
|
/* FIXME! This has to be written */
|
297 |
|
|
#define _FP_SQRT_MEAT_4(R, S, T, X, q)
|