1 /*
2  * Basic one-word fraction declaration and manipulation.
3  */
4 
5 #define _FP_FRAC_DECL_1(X)	_FP_W_TYPE X##_f
6 #define _FP_FRAC_COPY_1(D,S)	(D##_f = S##_f)
7 #define _FP_FRAC_SET_1(X,I)	(X##_f = I)
8 #define _FP_FRAC_HIGH_1(X)	(X##_f)
9 #define _FP_FRAC_LOW_1(X)	(X##_f)
10 #define _FP_FRAC_WORD_1(X,w)	(X##_f)
11 
12 #define _FP_FRAC_ADDI_1(X,I)	(X##_f += I)
13 #define _FP_FRAC_SLL_1(X,N)			\
14   do {						\
15     if (__builtin_constant_p(N) && (N) == 1)	\
16       X##_f += X##_f;				\
17     else					\
18       X##_f <<= (N);				\
19   } while (0)
20 #define _FP_FRAC_SRL_1(X,N)	(X##_f >>= N)
21 
22 /* Right shift with sticky-lsb.  */
23 #define _FP_FRAC_SRS_1(X,N,sz)	__FP_FRAC_SRS_1(X##_f, N, sz)
24 
25 #define __FP_FRAC_SRS_1(X,N,sz)						\
26    (X = (X >> (N) | (__builtin_constant_p(N) && (N) == 1		\
27 		     ? X & 1 : (X << (_FP_W_TYPE_SIZE - (N))) != 0)))
28 
29 #define _FP_FRAC_ADD_1(R,X,Y)	(R##_f = X##_f + Y##_f)
30 #define _FP_FRAC_SUB_1(R,X,Y)	(R##_f = X##_f - Y##_f)
31 #define _FP_FRAC_CLZ_1(z, X)	__FP_CLZ(z, X##_f)
32 
33 /* Predicates */
34 #define _FP_FRAC_NEGP_1(X)	((_FP_WS_TYPE)X##_f < 0)
35 #define _FP_FRAC_ZEROP_1(X)	(X##_f == 0)
36 #define _FP_FRAC_OVERP_1(fs,X)	(X##_f & _FP_OVERFLOW_##fs)
37 #define _FP_FRAC_EQ_1(X, Y)	(X##_f == Y##_f)
38 #define _FP_FRAC_GE_1(X, Y)	(X##_f >= Y##_f)
39 #define _FP_FRAC_GT_1(X, Y)	(X##_f > Y##_f)
40 
41 #define _FP_ZEROFRAC_1		0
42 #define _FP_MINFRAC_1		1
43 
44 /*
45  * Unpack the raw bits of a native fp value.  Do not classify or
46  * normalize the data.
47  */
48 
49 #define _FP_UNPACK_RAW_1(fs, X, val)				\
50   do {								\
51     union _FP_UNION_##fs _flo; _flo.flt = (val);		\
52 								\
53     X##_f = _flo.bits.frac;					\
54     X##_e = _flo.bits.exp;					\
55     X##_s = _flo.bits.sign;					\
56   } while (0)
57 
58 
59 /*
60  * Repack the raw bits of a native fp value.
61  */
62 
63 #define _FP_PACK_RAW_1(fs, val, X)				\
64   do {								\
65     union _FP_UNION_##fs _flo;					\
66 								\
67     _flo.bits.frac = X##_f;					\
68     _flo.bits.exp  = X##_e;					\
69     _flo.bits.sign = X##_s;					\
70 								\
71     (val) = _flo.flt;						\
72   } while (0)
73 
74 
75 /*
76  * Multiplication algorithms:
77  */
78 
79 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
80    multiplication immediately.  */
81 
82 #define _FP_MUL_MEAT_1_imm(fs, R, X, Y)					\
83   do {									\
84     R##_f = X##_f * Y##_f;						\
85     /* Normalize since we know where the msb of the multiplicands	\
86        were (bit B), we know that the msb of the of the product is	\
87        at either 2B or 2B-1.  */					\
88     _FP_FRAC_SRS_1(R, _FP_WFRACBITS_##fs-1, 2*_FP_WFRACBITS_##fs);	\
89   } while (0)
90 
91 /* Given a 1W * 1W => 2W primitive, do the extended multiplication.  */
92 
93 #define _FP_MUL_MEAT_1_wide(fs, R, X, Y, doit)				\
94   do {									\
95     _FP_W_TYPE _Z_f0, _Z_f1;						\
96     doit(_Z_f1, _Z_f0, X##_f, Y##_f);					\
97     /* Normalize since we know where the msb of the multiplicands	\
98        were (bit B), we know that the msb of the of the product is	\
99        at either 2B or 2B-1.  */					\
100     _FP_FRAC_SRS_2(_Z, _FP_WFRACBITS_##fs-1, 2*_FP_WFRACBITS_##fs);	\
101     R##_f = _Z_f0;							\
102   } while (0)
103 
104 /* Finally, a simple widening multiply algorithm.  What fun!  */
105 
106 #define _FP_MUL_MEAT_1_hard(fs, R, X, Y)				\
107   do {									\
108     _FP_W_TYPE _xh, _xl, _yh, _yl, _z_f0, _z_f1, _a_f0, _a_f1;		\
109 									\
110     /* split the words in half */					\
111     _xh = X##_f >> (_FP_W_TYPE_SIZE/2);					\
112     _xl = X##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);		\
113     _yh = Y##_f >> (_FP_W_TYPE_SIZE/2);					\
114     _yl = Y##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);		\
115 									\
116     /* multiply the pieces */						\
117     _z_f0 = _xl * _yl;							\
118     _a_f0 = _xh * _yl;							\
119     _a_f1 = _xl * _yh;							\
120     _z_f1 = _xh * _yh;							\
121 									\
122     /* reassemble into two full words */				\
123     if ((_a_f0 += _a_f1) < _a_f1)					\
124       _z_f1 += (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2);			\
125     _a_f1 = _a_f0 >> (_FP_W_TYPE_SIZE/2);				\
126     _a_f0 = _a_f0 << (_FP_W_TYPE_SIZE/2);				\
127     _FP_FRAC_ADD_2(_z, _z, _a);						\
128 									\
129     /* normalize */							\
130     _FP_FRAC_SRS_2(_z, _FP_WFRACBITS_##fs - 1, 2*_FP_WFRACBITS_##fs);	\
131     R##_f = _z_f0;							\
132   } while (0)
133 
134 
135 /*
136  * Division algorithms:
137  */
138 
139 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
140    division immediately.  Give this macro either _FP_DIV_HELP_imm for
141    C primitives or _FP_DIV_HELP_ldiv for the ISO function.  Which you
142    choose will depend on what the compiler does with divrem4.  */
143 
144 #define _FP_DIV_MEAT_1_imm(fs, R, X, Y, doit)		\
145   do {							\
146     _FP_W_TYPE _q, _r;					\
147     X##_f <<= (X##_f < Y##_f				\
148 	       ? R##_e--, _FP_WFRACBITS_##fs		\
149 	       : _FP_WFRACBITS_##fs - 1);		\
150     doit(_q, _r, X##_f, Y##_f);				\
151     R##_f = _q | (_r != 0);				\
152   } while (0)
153 
154 /* GCC's longlong.h defines a 2W / 1W => (1W,1W) primitive udiv_qrnnd
155    that may be useful in this situation.  This first is for a primitive
156    that requires normalization, the second for one that does not.  Look
157    for UDIV_NEEDS_NORMALIZATION to tell which your machine needs.  */
158 
159 #define _FP_DIV_MEAT_1_udiv_norm(fs, R, X, Y)				\
160   do {									\
161     _FP_W_TYPE _nh, _nl, _q, _r;					\
162 									\
163     /* Normalize Y -- i.e. make the most significant bit set.  */	\
164     Y##_f <<= _FP_WFRACXBITS_##fs - 1;					\
165 									\
166     /* Shift X op correspondingly high, that is, up one full word.  */	\
167     if (X##_f <= Y##_f)							\
168       {									\
169 	_nl = 0;							\
170 	_nh = X##_f;							\
171       }									\
172     else								\
173       {									\
174 	R##_e++;							\
175 	_nl = X##_f << (_FP_W_TYPE_SIZE-1);				\
176 	_nh = X##_f >> 1;						\
177       }									\
178     									\
179     udiv_qrnnd(_q, _r, _nh, _nl, Y##_f);				\
180     R##_f = _q | (_r != 0);						\
181   } while (0)
182 
183 #define _FP_DIV_MEAT_1_udiv(fs, R, X, Y)		\
184   do {							\
185     _FP_W_TYPE _nh, _nl, _q, _r;			\
186     if (X##_f < Y##_f)					\
187       {							\
188 	R##_e--;					\
189 	_nl = X##_f << _FP_WFRACBITS_##fs;		\
190 	_nh = X##_f >> _FP_WFRACXBITS_##fs;		\
191       }							\
192     else						\
193       {							\
194 	_nl = X##_f << (_FP_WFRACBITS_##fs - 1);	\
195 	_nh = X##_f >> (_FP_WFRACXBITS_##fs + 1);	\
196       }							\
197     udiv_qrnnd(_q, _r, _nh, _nl, Y##_f);		\
198     R##_f = _q | (_r != 0);				\
199   } while (0)
200 
201 
202 /*
203  * Square root algorithms:
204  * We have just one right now, maybe Newton approximation
205  * should be added for those machines where division is fast.
206  */
207 
208 #define _FP_SQRT_MEAT_1(R, S, T, X, q)			\
209   do {							\
210     while (q)						\
211       {							\
212         T##_f = S##_f + q;				\
213         if (T##_f <= X##_f)				\
214           {						\
215             S##_f = T##_f + q;				\
216             X##_f -= T##_f;				\
217             R##_f += q;					\
218           }						\
219         _FP_FRAC_SLL_1(X, 1);				\
220         q >>= 1;					\
221       }							\
222   } while (0)
223 
224 /*
225  * Assembly/disassembly for converting to/from integral types.
226  * No shifting or overflow handled here.
227  */
228 
229 #define _FP_FRAC_ASSEMBLE_1(r, X, rsize)	(r = X##_f)
230 #define _FP_FRAC_DISASSEMBLE_1(X, r, rsize)	(X##_f = r)
231 
232 
233 /*
234  * Convert FP values between word sizes
235  */
236 
237 #define _FP_FRAC_CONV_1_1(dfs, sfs, D, S)				\
238   do {									\
239     D##_f = S##_f;							\
240     if (_FP_WFRACBITS_##sfs > _FP_WFRACBITS_##dfs)			\
241       _FP_FRAC_SRS_1(D, (_FP_WFRACBITS_##sfs-_FP_WFRACBITS_##dfs),	\
242 		     _FP_WFRACBITS_##sfs);				\
243     else								\
244       D##_f <<= _FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs;		\
245   } while (0)
246