xref: /openbmc/linux/include/math-emu/op-2.h (revision ba61bb17)
1 /* Software floating-point emulation.
2    Basic two-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 #ifndef __MATH_EMU_OP_2_H__
26 #define __MATH_EMU_OP_2_H__
27 
28 #define _FP_FRAC_DECL_2(X)	_FP_W_TYPE X##_f0 = 0, X##_f1 = 0
29 #define _FP_FRAC_COPY_2(D,S)	(D##_f0 = S##_f0, D##_f1 = S##_f1)
30 #define _FP_FRAC_SET_2(X,I)	__FP_FRAC_SET_2(X, I)
31 #define _FP_FRAC_HIGH_2(X)	(X##_f1)
32 #define _FP_FRAC_LOW_2(X)	(X##_f0)
33 #define _FP_FRAC_WORD_2(X,w)	(X##_f##w)
34 
35 #define _FP_FRAC_SLL_2(X,N)						\
36   do {									\
37     if ((N) < _FP_W_TYPE_SIZE)						\
38       {									\
39 	if (__builtin_constant_p(N) && (N) == 1) 			\
40 	  {								\
41 	    X##_f1 = X##_f1 + X##_f1 + (((_FP_WS_TYPE)(X##_f0)) < 0);	\
42 	    X##_f0 += X##_f0;						\
43 	  }								\
44 	else								\
45 	  {								\
46 	    X##_f1 = X##_f1 << (N) | X##_f0 >> (_FP_W_TYPE_SIZE - (N));	\
47 	    X##_f0 <<= (N);						\
48 	  }								\
49       }									\
50     else								\
51       {									\
52 	X##_f1 = X##_f0 << ((N) - _FP_W_TYPE_SIZE);			\
53 	X##_f0 = 0;							\
54       }									\
55   } while (0)
56 
57 #define _FP_FRAC_SRL_2(X,N)						\
58   do {									\
59     if ((N) < _FP_W_TYPE_SIZE)						\
60       {									\
61 	X##_f0 = X##_f0 >> (N) | X##_f1 << (_FP_W_TYPE_SIZE - (N));	\
62 	X##_f1 >>= (N);							\
63       }									\
64     else								\
65       {									\
66 	X##_f0 = X##_f1 >> ((N) - _FP_W_TYPE_SIZE);			\
67 	X##_f1 = 0;							\
68       }									\
69   } while (0)
70 
71 /* Right shift with sticky-lsb.  */
72 #define _FP_FRAC_SRS_2(X,N,sz)						\
73   do {									\
74     if ((N) < _FP_W_TYPE_SIZE)						\
75       {									\
76 	X##_f0 = (X##_f1 << (_FP_W_TYPE_SIZE - (N)) | X##_f0 >> (N) |	\
77 		  (__builtin_constant_p(N) && (N) == 1			\
78 		   ? X##_f0 & 1						\
79 		   : (X##_f0 << (_FP_W_TYPE_SIZE - (N))) != 0));	\
80 	X##_f1 >>= (N);							\
81       }									\
82     else								\
83       {									\
84 	X##_f0 = (X##_f1 >> ((N) - _FP_W_TYPE_SIZE) |			\
85 		(((X##_f1 << (2*_FP_W_TYPE_SIZE - (N))) | X##_f0) != 0)); \
86 	X##_f1 = 0;							\
87       }									\
88   } while (0)
89 
90 #define _FP_FRAC_ADDI_2(X,I)	\
91   __FP_FRAC_ADDI_2(X##_f1, X##_f0, I)
92 
93 #define _FP_FRAC_ADD_2(R,X,Y)	\
94   __FP_FRAC_ADD_2(R##_f1, R##_f0, X##_f1, X##_f0, Y##_f1, Y##_f0)
95 
96 #define _FP_FRAC_SUB_2(R,X,Y)	\
97   __FP_FRAC_SUB_2(R##_f1, R##_f0, X##_f1, X##_f0, Y##_f1, Y##_f0)
98 
99 #define _FP_FRAC_DEC_2(X,Y)	\
100   __FP_FRAC_DEC_2(X##_f1, X##_f0, Y##_f1, Y##_f0)
101 
102 #define _FP_FRAC_CLZ_2(R,X)	\
103   do {				\
104     if (X##_f1)			\
105       __FP_CLZ(R,X##_f1);	\
106     else 			\
107     {				\
108       __FP_CLZ(R,X##_f0);	\
109       R += _FP_W_TYPE_SIZE;	\
110     }				\
111   } while(0)
112 
113 /* Predicates */
114 #define _FP_FRAC_NEGP_2(X)	((_FP_WS_TYPE)X##_f1 < 0)
115 #define _FP_FRAC_ZEROP_2(X)	((X##_f1 | X##_f0) == 0)
116 #define _FP_FRAC_OVERP_2(fs,X)	(_FP_FRAC_HIGH_##fs(X) & _FP_OVERFLOW_##fs)
117 #define _FP_FRAC_CLEAR_OVERP_2(fs,X)	(_FP_FRAC_HIGH_##fs(X) &= ~_FP_OVERFLOW_##fs)
118 #define _FP_FRAC_EQ_2(X, Y)	(X##_f1 == Y##_f1 && X##_f0 == Y##_f0)
119 #define _FP_FRAC_GT_2(X, Y)	\
120   (X##_f1 > Y##_f1 || (X##_f1 == Y##_f1 && X##_f0 > Y##_f0))
121 #define _FP_FRAC_GE_2(X, Y)	\
122   (X##_f1 > Y##_f1 || (X##_f1 == Y##_f1 && X##_f0 >= Y##_f0))
123 
124 #define _FP_ZEROFRAC_2		0, 0
125 #define _FP_MINFRAC_2		0, 1
126 #define _FP_MAXFRAC_2		(~(_FP_WS_TYPE)0), (~(_FP_WS_TYPE)0)
127 
128 /*
129  * Internals
130  */
131 
132 #define __FP_FRAC_SET_2(X,I1,I0)	(X##_f0 = I0, X##_f1 = I1)
133 
134 #define __FP_CLZ_2(R, xh, xl)	\
135   do {				\
136     if (xh)			\
137       __FP_CLZ(R,xh);		\
138     else 			\
139     {				\
140       __FP_CLZ(R,xl);		\
141       R += _FP_W_TYPE_SIZE;	\
142     }				\
143   } while(0)
144 
145 #if 0
146 
147 #ifndef __FP_FRAC_ADDI_2
148 #define __FP_FRAC_ADDI_2(xh, xl, i)	\
149   (xh += ((xl += i) < i))
150 #endif
151 #ifndef __FP_FRAC_ADD_2
152 #define __FP_FRAC_ADD_2(rh, rl, xh, xl, yh, yl)	\
153   (rh = xh + yh + ((rl = xl + yl) < xl))
154 #endif
155 #ifndef __FP_FRAC_SUB_2
156 #define __FP_FRAC_SUB_2(rh, rl, xh, xl, yh, yl)	\
157   (rh = xh - yh - ((rl = xl - yl) > xl))
158 #endif
159 #ifndef __FP_FRAC_DEC_2
160 #define __FP_FRAC_DEC_2(xh, xl, yh, yl)	\
161   do {					\
162     UWtype _t = xl;			\
163     xh -= yh + ((xl -= yl) > _t);	\
164   } while (0)
165 #endif
166 
167 #else
168 
169 #undef __FP_FRAC_ADDI_2
170 #define __FP_FRAC_ADDI_2(xh, xl, i)	add_ssaaaa(xh, xl, xh, xl, 0, i)
171 #undef __FP_FRAC_ADD_2
172 #define __FP_FRAC_ADD_2			add_ssaaaa
173 #undef __FP_FRAC_SUB_2
174 #define __FP_FRAC_SUB_2			sub_ddmmss
175 #undef __FP_FRAC_DEC_2
176 #define __FP_FRAC_DEC_2(xh, xl, yh, yl)	sub_ddmmss(xh, xl, xh, xl, yh, yl)
177 
178 #endif
179 
180 /*
181  * Unpack the raw bits of a native fp value.  Do not classify or
182  * normalize the data.
183  */
184 
185 #define _FP_UNPACK_RAW_2(fs, X, val)			\
186   do {							\
187     union _FP_UNION_##fs _flo; _flo.flt = (val);	\
188 							\
189     X##_f0 = _flo.bits.frac0;				\
190     X##_f1 = _flo.bits.frac1;				\
191     X##_e  = _flo.bits.exp;				\
192     X##_s  = _flo.bits.sign;				\
193   } while (0)
194 
195 #define _FP_UNPACK_RAW_2_P(fs, X, val)			\
196   do {							\
197     union _FP_UNION_##fs *_flo =			\
198       (union _FP_UNION_##fs *)(val);			\
199 							\
200     X##_f0 = _flo->bits.frac0;				\
201     X##_f1 = _flo->bits.frac1;				\
202     X##_e  = _flo->bits.exp;				\
203     X##_s  = _flo->bits.sign;				\
204   } while (0)
205 
206 
207 /*
208  * Repack the raw bits of a native fp value.
209  */
210 
211 #define _FP_PACK_RAW_2(fs, val, X)			\
212   do {							\
213     union _FP_UNION_##fs _flo;				\
214 							\
215     _flo.bits.frac0 = X##_f0;				\
216     _flo.bits.frac1 = X##_f1;				\
217     _flo.bits.exp   = X##_e;				\
218     _flo.bits.sign  = X##_s;				\
219 							\
220     (val) = _flo.flt;					\
221   } while (0)
222 
223 #define _FP_PACK_RAW_2_P(fs, val, X)			\
224   do {							\
225     union _FP_UNION_##fs *_flo =			\
226       (union _FP_UNION_##fs *)(val);			\
227 							\
228     _flo->bits.frac0 = X##_f0;				\
229     _flo->bits.frac1 = X##_f1;				\
230     _flo->bits.exp   = X##_e;				\
231     _flo->bits.sign  = X##_s;				\
232   } while (0)
233 
234 
235 /*
236  * Multiplication algorithms:
237  */
238 
239 /* Given a 1W * 1W => 2W primitive, do the extended multiplication.  */
240 
241 #define _FP_MUL_MEAT_2_wide(wfracbits, R, X, Y, doit)			\
242   do {									\
243     _FP_FRAC_DECL_4(_z); _FP_FRAC_DECL_2(_b); _FP_FRAC_DECL_2(_c);	\
244 									\
245     doit(_FP_FRAC_WORD_4(_z,1), _FP_FRAC_WORD_4(_z,0), X##_f0, Y##_f0);	\
246     doit(_b_f1, _b_f0, X##_f0, Y##_f1);					\
247     doit(_c_f1, _c_f0, X##_f1, Y##_f0);					\
248     doit(_FP_FRAC_WORD_4(_z,3), _FP_FRAC_WORD_4(_z,2), X##_f1, Y##_f1);	\
249 									\
250     __FP_FRAC_ADD_3(_FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
251 		    _FP_FRAC_WORD_4(_z,1), 0, _b_f1, _b_f0,		\
252 		    _FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
253 		    _FP_FRAC_WORD_4(_z,1));				\
254     __FP_FRAC_ADD_3(_FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
255 		    _FP_FRAC_WORD_4(_z,1), 0, _c_f1, _c_f0,		\
256 		    _FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
257 		    _FP_FRAC_WORD_4(_z,1));				\
258 									\
259     /* Normalize since we know where the msb of the multiplicands	\
260        were (bit B), we know that the msb of the of the product is	\
261        at either 2B or 2B-1.  */					\
262     _FP_FRAC_SRS_4(_z, wfracbits-1, 2*wfracbits);			\
263     R##_f0 = _FP_FRAC_WORD_4(_z,0);					\
264     R##_f1 = _FP_FRAC_WORD_4(_z,1);					\
265   } while (0)
266 
267 /* Given a 1W * 1W => 2W primitive, do the extended multiplication.
268    Do only 3 multiplications instead of four. This one is for machines
269    where multiplication is much more expensive than subtraction.  */
270 
271 #define _FP_MUL_MEAT_2_wide_3mul(wfracbits, R, X, Y, doit)		\
272   do {									\
273     _FP_FRAC_DECL_4(_z); _FP_FRAC_DECL_2(_b); _FP_FRAC_DECL_2(_c);	\
274     _FP_W_TYPE _d;							\
275     int _c1, _c2;							\
276 									\
277     _b_f0 = X##_f0 + X##_f1;						\
278     _c1 = _b_f0 < X##_f0;						\
279     _b_f1 = Y##_f0 + Y##_f1;						\
280     _c2 = _b_f1 < Y##_f0;						\
281     doit(_d, _FP_FRAC_WORD_4(_z,0), X##_f0, Y##_f0);			\
282     doit(_FP_FRAC_WORD_4(_z,2), _FP_FRAC_WORD_4(_z,1), _b_f0, _b_f1);	\
283     doit(_c_f1, _c_f0, X##_f1, Y##_f1);					\
284 									\
285     _b_f0 &= -_c2;							\
286     _b_f1 &= -_c1;							\
287     __FP_FRAC_ADD_3(_FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
288 		    _FP_FRAC_WORD_4(_z,1), (_c1 & _c2), 0, _d,		\
289 		    0, _FP_FRAC_WORD_4(_z,2), _FP_FRAC_WORD_4(_z,1));	\
290     __FP_FRAC_ADDI_2(_FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
291 		     _b_f0);						\
292     __FP_FRAC_ADDI_2(_FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
293 		     _b_f1);						\
294     __FP_FRAC_DEC_3(_FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
295 		    _FP_FRAC_WORD_4(_z,1),				\
296 		    0, _d, _FP_FRAC_WORD_4(_z,0));			\
297     __FP_FRAC_DEC_3(_FP_FRAC_WORD_4(_z,3),_FP_FRAC_WORD_4(_z,2),	\
298 		    _FP_FRAC_WORD_4(_z,1), 0, _c_f1, _c_f0);		\
299     __FP_FRAC_ADD_2(_FP_FRAC_WORD_4(_z,3), _FP_FRAC_WORD_4(_z,2),	\
300 		    _c_f1, _c_f0,					\
301 		    _FP_FRAC_WORD_4(_z,3), _FP_FRAC_WORD_4(_z,2));	\
302 									\
303     /* Normalize since we know where the msb of the multiplicands	\
304        were (bit B), we know that the msb of the of the product is	\
305        at either 2B or 2B-1.  */					\
306     _FP_FRAC_SRS_4(_z, wfracbits-1, 2*wfracbits);			\
307     R##_f0 = _FP_FRAC_WORD_4(_z,0);					\
308     R##_f1 = _FP_FRAC_WORD_4(_z,1);					\
309   } while (0)
310 
311 #define _FP_MUL_MEAT_2_gmp(wfracbits, R, X, Y)				\
312   do {									\
313     _FP_FRAC_DECL_4(_z);						\
314     _FP_W_TYPE _x[2], _y[2];						\
315     _x[0] = X##_f0; _x[1] = X##_f1;					\
316     _y[0] = Y##_f0; _y[1] = Y##_f1;					\
317 									\
318     mpn_mul_n(_z_f, _x, _y, 2);						\
319 									\
320     /* Normalize since we know where the msb of the multiplicands	\
321        were (bit B), we know that the msb of the of the product is	\
322        at either 2B or 2B-1.  */					\
323     _FP_FRAC_SRS_4(_z, wfracbits-1, 2*wfracbits);			\
324     R##_f0 = _z_f[0];							\
325     R##_f1 = _z_f[1];							\
326   } while (0)
327 
328 /* Do at most 120x120=240 bits multiplication using double floating
329    point multiplication.  This is useful if floating point
330    multiplication has much bigger throughput than integer multiply.
331    It is supposed to work for _FP_W_TYPE_SIZE 64 and wfracbits
332    between 106 and 120 only.
333    Caller guarantees that X and Y has (1LLL << (wfracbits - 1)) set.
334    SETFETZ is a macro which will disable all FPU exceptions and set rounding
335    towards zero,  RESETFE should optionally reset it back.  */
336 
337 #define _FP_MUL_MEAT_2_120_240_double(wfracbits, R, X, Y, setfetz, resetfe)	\
338   do {										\
339     static const double _const[] = {						\
340       /* 2^-24 */ 5.9604644775390625e-08,					\
341       /* 2^-48 */ 3.5527136788005009e-15,					\
342       /* 2^-72 */ 2.1175823681357508e-22,					\
343       /* 2^-96 */ 1.2621774483536189e-29,					\
344       /* 2^28 */ 2.68435456e+08,						\
345       /* 2^4 */ 1.600000e+01,							\
346       /* 2^-20 */ 9.5367431640625e-07,						\
347       /* 2^-44 */ 5.6843418860808015e-14,					\
348       /* 2^-68 */ 3.3881317890172014e-21,					\
349       /* 2^-92 */ 2.0194839173657902e-28,					\
350       /* 2^-116 */ 1.2037062152420224e-35};					\
351     double _a240, _b240, _c240, _d240, _e240, _f240, 				\
352 	   _g240, _h240, _i240, _j240, _k240;					\
353     union { double d; UDItype i; } _l240, _m240, _n240, _o240,			\
354 				   _p240, _q240, _r240, _s240;			\
355     UDItype _t240, _u240, _v240, _w240, _x240, _y240 = 0;			\
356 										\
357     if (wfracbits < 106 || wfracbits > 120)					\
358       abort();									\
359 										\
360     setfetz;									\
361 										\
362     _e240 = (double)(long)(X##_f0 & 0xffffff);					\
363     _j240 = (double)(long)(Y##_f0 & 0xffffff);					\
364     _d240 = (double)(long)((X##_f0 >> 24) & 0xffffff);				\
365     _i240 = (double)(long)((Y##_f0 >> 24) & 0xffffff);				\
366     _c240 = (double)(long)(((X##_f1 << 16) & 0xffffff) | (X##_f0 >> 48));	\
367     _h240 = (double)(long)(((Y##_f1 << 16) & 0xffffff) | (Y##_f0 >> 48));	\
368     _b240 = (double)(long)((X##_f1 >> 8) & 0xffffff);				\
369     _g240 = (double)(long)((Y##_f1 >> 8) & 0xffffff);				\
370     _a240 = (double)(long)(X##_f1 >> 32);					\
371     _f240 = (double)(long)(Y##_f1 >> 32);					\
372     _e240 *= _const[3];								\
373     _j240 *= _const[3];								\
374     _d240 *= _const[2];								\
375     _i240 *= _const[2];								\
376     _c240 *= _const[1];								\
377     _h240 *= _const[1];								\
378     _b240 *= _const[0];								\
379     _g240 *= _const[0];								\
380     _s240.d =							      _e240*_j240;\
381     _r240.d =						_d240*_j240 + _e240*_i240;\
382     _q240.d =				  _c240*_j240 + _d240*_i240 + _e240*_h240;\
383     _p240.d =		    _b240*_j240 + _c240*_i240 + _d240*_h240 + _e240*_g240;\
384     _o240.d = _a240*_j240 + _b240*_i240 + _c240*_h240 + _d240*_g240 + _e240*_f240;\
385     _n240.d = _a240*_i240 + _b240*_h240 + _c240*_g240 + _d240*_f240;		\
386     _m240.d = _a240*_h240 + _b240*_g240 + _c240*_f240;				\
387     _l240.d = _a240*_g240 + _b240*_f240;					\
388     _k240 =   _a240*_f240;							\
389     _r240.d += _s240.d;								\
390     _q240.d += _r240.d;								\
391     _p240.d += _q240.d;								\
392     _o240.d += _p240.d;								\
393     _n240.d += _o240.d;								\
394     _m240.d += _n240.d;								\
395     _l240.d += _m240.d;								\
396     _k240 += _l240.d;								\
397     _s240.d -= ((_const[10]+_s240.d)-_const[10]);				\
398     _r240.d -= ((_const[9]+_r240.d)-_const[9]);					\
399     _q240.d -= ((_const[8]+_q240.d)-_const[8]);					\
400     _p240.d -= ((_const[7]+_p240.d)-_const[7]);					\
401     _o240.d += _const[7];							\
402     _n240.d += _const[6];							\
403     _m240.d += _const[5];							\
404     _l240.d += _const[4];							\
405     if (_s240.d != 0.0) _y240 = 1;						\
406     if (_r240.d != 0.0) _y240 = 1;						\
407     if (_q240.d != 0.0) _y240 = 1;						\
408     if (_p240.d != 0.0) _y240 = 1;						\
409     _t240 = (DItype)_k240;							\
410     _u240 = _l240.i;								\
411     _v240 = _m240.i;								\
412     _w240 = _n240.i;								\
413     _x240 = _o240.i;								\
414     R##_f1 = (_t240 << (128 - (wfracbits - 1)))					\
415 	     | ((_u240 & 0xffffff) >> ((wfracbits - 1) - 104));			\
416     R##_f0 = ((_u240 & 0xffffff) << (168 - (wfracbits - 1)))			\
417     	     | ((_v240 & 0xffffff) << (144 - (wfracbits - 1)))			\
418     	     | ((_w240 & 0xffffff) << (120 - (wfracbits - 1)))			\
419     	     | ((_x240 & 0xffffff) >> ((wfracbits - 1) - 96))			\
420     	     | _y240;								\
421     resetfe;									\
422   } while (0)
423 
424 /*
425  * Division algorithms:
426  */
427 
428 #define _FP_DIV_MEAT_2_udiv(fs, R, X, Y)				\
429   do {									\
430     _FP_W_TYPE _n_f2, _n_f1, _n_f0, _r_f1, _r_f0, _m_f1, _m_f0;		\
431     if (_FP_FRAC_GT_2(X, Y))						\
432       {									\
433 	_n_f2 = X##_f1 >> 1;						\
434 	_n_f1 = X##_f1 << (_FP_W_TYPE_SIZE - 1) | X##_f0 >> 1;		\
435 	_n_f0 = X##_f0 << (_FP_W_TYPE_SIZE - 1);			\
436       }									\
437     else								\
438       {									\
439 	R##_e--;							\
440 	_n_f2 = X##_f1;							\
441 	_n_f1 = X##_f0;							\
442 	_n_f0 = 0;							\
443       }									\
444 									\
445     /* Normalize, i.e. make the most significant bit of the 		\
446        denominator set. */						\
447     _FP_FRAC_SLL_2(Y, _FP_WFRACXBITS_##fs);				\
448 									\
449     udiv_qrnnd(R##_f1, _r_f1, _n_f2, _n_f1, Y##_f1);			\
450     umul_ppmm(_m_f1, _m_f0, R##_f1, Y##_f0);				\
451     _r_f0 = _n_f0;							\
452     if (_FP_FRAC_GT_2(_m, _r))						\
453       {									\
454 	R##_f1--;							\
455 	_FP_FRAC_ADD_2(_r, Y, _r);					\
456 	if (_FP_FRAC_GE_2(_r, Y) && _FP_FRAC_GT_2(_m, _r))		\
457 	  {								\
458 	    R##_f1--;							\
459 	    _FP_FRAC_ADD_2(_r, Y, _r);					\
460 	  }								\
461       }									\
462     _FP_FRAC_DEC_2(_r, _m);						\
463 									\
464     if (_r_f1 == Y##_f1)						\
465       {									\
466 	/* This is a special case, not an optimization			\
467 	   (_r/Y##_f1 would not fit into UWtype).			\
468 	   As _r is guaranteed to be < Y,  R##_f0 can be either		\
469 	   (UWtype)-1 or (UWtype)-2.  But as we know what kind		\
470 	   of bits it is (sticky, guard, round),  we don't care.	\
471 	   We also don't care what the reminder is,  because the	\
472 	   guard bit will be set anyway.  -jj */			\
473 	R##_f0 = -1;							\
474       }									\
475     else								\
476       {									\
477 	udiv_qrnnd(R##_f0, _r_f1, _r_f1, _r_f0, Y##_f1);		\
478 	umul_ppmm(_m_f1, _m_f0, R##_f0, Y##_f0);			\
479 	_r_f0 = 0;							\
480 	if (_FP_FRAC_GT_2(_m, _r))					\
481 	  {								\
482 	    R##_f0--;							\
483 	    _FP_FRAC_ADD_2(_r, Y, _r);					\
484 	    if (_FP_FRAC_GE_2(_r, Y) && _FP_FRAC_GT_2(_m, _r))		\
485 	      {								\
486 		R##_f0--;						\
487 		_FP_FRAC_ADD_2(_r, Y, _r);				\
488 	      }								\
489 	  }								\
490 	if (!_FP_FRAC_EQ_2(_r, _m))					\
491 	  R##_f0 |= _FP_WORK_STICKY;					\
492       }									\
493   } while (0)
494 
495 
496 #define _FP_DIV_MEAT_2_gmp(fs, R, X, Y)					\
497   do {									\
498     _FP_W_TYPE _x[4], _y[2], _z[4];					\
499     _y[0] = Y##_f0; _y[1] = Y##_f1;					\
500     _x[0] = _x[3] = 0;							\
501     if (_FP_FRAC_GT_2(X, Y))						\
502       {									\
503 	R##_e++;							\
504 	_x[1] = (X##_f0 << (_FP_WFRACBITS_##fs-1 - _FP_W_TYPE_SIZE) |	\
505 		 X##_f1 >> (_FP_W_TYPE_SIZE -				\
506 			    (_FP_WFRACBITS_##fs-1 - _FP_W_TYPE_SIZE)));	\
507 	_x[2] = X##_f1 << (_FP_WFRACBITS_##fs-1 - _FP_W_TYPE_SIZE);	\
508       }									\
509     else								\
510       {									\
511 	_x[1] = (X##_f0 << (_FP_WFRACBITS_##fs - _FP_W_TYPE_SIZE) |	\
512 		 X##_f1 >> (_FP_W_TYPE_SIZE -				\
513 			    (_FP_WFRACBITS_##fs - _FP_W_TYPE_SIZE)));	\
514 	_x[2] = X##_f1 << (_FP_WFRACBITS_##fs - _FP_W_TYPE_SIZE);	\
515       }									\
516 									\
517     (void) mpn_divrem (_z, 0, _x, 4, _y, 2);				\
518     R##_f1 = _z[1];							\
519     R##_f0 = _z[0] | ((_x[0] | _x[1]) != 0);				\
520   } while (0)
521 
522 
523 /*
524  * Square root algorithms:
525  * We have just one right now, maybe Newton approximation
526  * should be added for those machines where division is fast.
527  */
528 
529 #define _FP_SQRT_MEAT_2(R, S, T, X, q)			\
530   do {							\
531     while (q)						\
532       {							\
533 	T##_f1 = S##_f1 + q;				\
534 	if (T##_f1 <= X##_f1)				\
535 	  {						\
536 	    S##_f1 = T##_f1 + q;			\
537 	    X##_f1 -= T##_f1;				\
538 	    R##_f1 += q;				\
539 	  }						\
540 	_FP_FRAC_SLL_2(X, 1);				\
541 	q >>= 1;					\
542       }							\
543     q = (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE - 1);		\
544     while (q != _FP_WORK_ROUND)				\
545       {							\
546 	T##_f0 = S##_f0 + q;				\
547 	T##_f1 = S##_f1;				\
548 	if (T##_f1 < X##_f1 || 				\
549 	    (T##_f1 == X##_f1 && T##_f0 <= X##_f0))	\
550 	  {						\
551 	    S##_f0 = T##_f0 + q;			\
552 	    S##_f1 += (T##_f0 > S##_f0);		\
553 	    _FP_FRAC_DEC_2(X, T);			\
554 	    R##_f0 += q;				\
555 	  }						\
556 	_FP_FRAC_SLL_2(X, 1);				\
557 	q >>= 1;					\
558       }							\
559     if (X##_f0 | X##_f1)				\
560       {							\
561 	if (S##_f1 < X##_f1 || 				\
562 	    (S##_f1 == X##_f1 && S##_f0 < X##_f0))	\
563 	  R##_f0 |= _FP_WORK_ROUND;			\
564 	R##_f0 |= _FP_WORK_STICKY;			\
565       }							\
566   } while (0)
567 
568 
569 /*
570  * Assembly/disassembly for converting to/from integral types.
571  * No shifting or overflow handled here.
572  */
573 
574 #define _FP_FRAC_ASSEMBLE_2(r, X, rsize)	\
575   do {						\
576     if (rsize <= _FP_W_TYPE_SIZE)		\
577       r = X##_f0;				\
578     else					\
579       {						\
580 	r = X##_f1;				\
581 	r <<= _FP_W_TYPE_SIZE;			\
582 	r += X##_f0;				\
583       }						\
584   } while (0)
585 
586 #define _FP_FRAC_DISASSEMBLE_2(X, r, rsize)				\
587   do {									\
588     X##_f0 = r;								\
589     X##_f1 = (rsize <= _FP_W_TYPE_SIZE ? 0 : r >> _FP_W_TYPE_SIZE);	\
590   } while (0)
591 
592 /*
593  * Convert FP values between word sizes
594  */
595 
596 #define _FP_FRAC_CONV_1_2(dfs, sfs, D, S)				\
597   do {									\
598     if (S##_c != FP_CLS_NAN)						\
599       _FP_FRAC_SRS_2(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs),	\
600 		     _FP_WFRACBITS_##sfs);				\
601     else								\
602       _FP_FRAC_SRL_2(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs));	\
603     D##_f = S##_f0;							\
604   } while (0)
605 
606 #define _FP_FRAC_CONV_2_1(dfs, sfs, D, S)				\
607   do {									\
608     D##_f0 = S##_f;							\
609     D##_f1 = 0;								\
610     _FP_FRAC_SLL_2(D, (_FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs));	\
611   } while (0)
612 
613 #endif
614