1/* 2 * BK Id: SCCS/s.op-4.h 1.5 05/17/01 18:14:23 cort 3 */ 4/* 5 * Basic four-word fraction declaration and manipulation. 6 * 7 * When adding quadword support for 32 bit machines, we need 8 * to be a little careful as double multiply uses some of these 9 * macros: (in op-2.h) 10 * _FP_MUL_MEAT_2_wide() uses _FP_FRAC_DECL_4, _FP_FRAC_WORD_4, 11 * _FP_FRAC_ADD_4, _FP_FRAC_SRS_4 12 * _FP_MUL_MEAT_2_gmp() uses _FP_FRAC_SRS_4 (and should use 13 * _FP_FRAC_DECL_4: it appears to be broken and is not used 14 * anywhere anyway. ) 15 * 16 * I've now fixed all the macros that were here from the sparc64 code. 17 * [*none* of the shift macros were correct!] -- PMM 02/1998 18 * 19 * The only quadword stuff that remains to be coded is: 20 * 1) the conversion to/from ints, which requires 21 * that we check (in op-common.h) that the following do the right thing 22 * for quadwords: _FP_TO_INT(Q,4,r,X,rsz,rsg), _FP_FROM_INT(Q,4,X,r,rs,rt) 23 * 2) multiply, divide and sqrt, which require: 24 * _FP_MUL_MEAT_4_*(R,X,Y), _FP_DIV_MEAT_4_*(R,X,Y), _FP_SQRT_MEAT_4(R,S,T,X,q), 25 * This also needs _FP_MUL_MEAT_Q and _FP_DIV_MEAT_Q to be defined to 26 * some suitable _FP_MUL_MEAT_4_* macros in sfp-machine.h. 27 * [we're free to choose whatever FP_MUL_MEAT_4_* macros we need for 28 * these; they are used nowhere else. ] 29 */ 30 31#define _FP_FRAC_DECL_4(X) _FP_W_TYPE X##_f[4] 32#define _FP_FRAC_COPY_4(D,S) \ 33 (D##_f[0] = S##_f[0], D##_f[1] = S##_f[1], \ 34 D##_f[2] = S##_f[2], D##_f[3] = S##_f[3]) 35/* The _FP_FRAC_SET_n(X,I) macro is intended for use with another 36 * macro such as _FP_ZEROFRAC_n which returns n comma separated values. 37 * The result is that we get an expansion of __FP_FRAC_SET_n(X,I0,I1,I2,I3) 38 * which just assigns the In values to the array X##_f[]. 39 * This is why the number of parameters doesn't appear to match 40 * at first glance... -- PMM 41 */ 42#define _FP_FRAC_SET_4(X,I) __FP_FRAC_SET_4(X, I) 43#define _FP_FRAC_HIGH_4(X) (X##_f[3]) 44#define _FP_FRAC_LOW_4(X) (X##_f[0]) 45#define _FP_FRAC_WORD_4(X,w) (X##_f[w]) 46 47#define _FP_FRAC_SLL_4(X,N) \ 48 do { \ 49 _FP_I_TYPE _up, _down, _skip, _i; \ 50 _skip = (N) / _FP_W_TYPE_SIZE; \ 51 _up = (N) % _FP_W_TYPE_SIZE; \ 52 _down = _FP_W_TYPE_SIZE - _up; \ 53 for (_i = 3; _i > _skip; --_i) \ 54 X##_f[_i] = X##_f[_i-_skip] << _up | X##_f[_i-_skip-1] >> _down; \ 55/* bugfixed: was X##_f[_i] <<= _up; -- PMM 02/1998 */ \ 56 X##_f[_i] = X##_f[0] << _up; \ 57 for (--_i; _i >= 0; --_i) \ 58 X##_f[_i] = 0; \ 59 } while (0) 60 61/* This one was broken too */ 62#define _FP_FRAC_SRL_4(X,N) \ 63 do { \ 64 _FP_I_TYPE _up, _down, _skip, _i; \ 65 _skip = (N) / _FP_W_TYPE_SIZE; \ 66 _down = (N) % _FP_W_TYPE_SIZE; \ 67 _up = _FP_W_TYPE_SIZE - _down; \ 68 for (_i = 0; _i < 3-_skip; ++_i) \ 69 X##_f[_i] = X##_f[_i+_skip] >> _down | X##_f[_i+_skip+1] << _up; \ 70 X##_f[_i] = X##_f[3] >> _down; \ 71 for (++_i; _i < 4; ++_i) \ 72 X##_f[_i] = 0; \ 73 } while (0) 74 75 76/* Right shift with sticky-lsb. 77 * What this actually means is that we do a standard right-shift, 78 * but that if any of the bits that fall off the right hand side 79 * were one then we always set the LSbit. 80 */ 81#define _FP_FRAC_SRS_4(X,N,size) \ 82 do { \ 83 _FP_I_TYPE _up, _down, _skip, _i; \ 84 _FP_W_TYPE _s; \ 85 _skip = (N) / _FP_W_TYPE_SIZE; \ 86 _down = (N) % _FP_W_TYPE_SIZE; \ 87 _up = _FP_W_TYPE_SIZE - _down; \ 88 for (_s = _i = 0; _i < _skip; ++_i) \ 89 _s |= X##_f[_i]; \ 90 _s |= X##_f[_i] << _up; \ 91/* s is now != 0 if we want to set the LSbit */ \ 92 for (_i = 0; _i < 3-_skip; ++_i) \ 93 X##_f[_i] = X##_f[_i+_skip] >> _down | X##_f[_i+_skip+1] << _up; \ 94 X##_f[_i] = X##_f[3] >> _down; \ 95 for (++_i; _i < 4; ++_i) \ 96 X##_f[_i] = 0; \ 97 /* don't fix the LSB until the very end when we're sure f[0] is stable */ \ 98 X##_f[0] |= (_s != 0); \ 99 } while (0) 100 101#define _FP_FRAC_ADD_4(R,X,Y) \ 102 __FP_FRAC_ADD_4(R##_f[3], R##_f[2], R##_f[1], R##_f[0], \ 103 X##_f[3], X##_f[2], X##_f[1], X##_f[0], \ 104 Y##_f[3], Y##_f[2], Y##_f[1], Y##_f[0]) 105 106#define _FP_FRAC_SUB_4(R,X,Y) \ 107 __FP_FRAC_SUB_4(R##_f[3], R##_f[2], R##_f[1], R##_f[0], \ 108 X##_f[3], X##_f[2], X##_f[1], X##_f[0], \ 109 Y##_f[3], Y##_f[2], Y##_f[1], Y##_f[0]) 110 111#define _FP_FRAC_ADDI_4(X,I) \ 112 __FP_FRAC_ADDI_4(X##_f[3], X##_f[2], X##_f[1], X##_f[0], I) 113 114#define _FP_ZEROFRAC_4 0,0,0,0 115#define _FP_MINFRAC_4 0,0,0,1 116 117#define _FP_FRAC_ZEROP_4(X) ((X##_f[0] | X##_f[1] | X##_f[2] | X##_f[3]) == 0) 118#define _FP_FRAC_NEGP_4(X) ((_FP_WS_TYPE)X##_f[3] < 0) 119#define _FP_FRAC_OVERP_4(fs,X) (X##_f[0] & _FP_OVERFLOW_##fs) 120 121#define _FP_FRAC_EQ_4(X,Y) \ 122 (X##_f[0] == Y##_f[0] && X##_f[1] == Y##_f[1] \ 123 && X##_f[2] == Y##_f[2] && X##_f[3] == Y##_f[3]) 124 125#define _FP_FRAC_GT_4(X,Y) \ 126 (X##_f[3] > Y##_f[3] || \ 127 (X##_f[3] == Y##_f[3] && (X##_f[2] > Y##_f[2] || \ 128 (X##_f[2] == Y##_f[2] && (X##_f[1] > Y##_f[1] || \ 129 (X##_f[1] == Y##_f[1] && X##_f[0] > Y##_f[0]) \ 130 )) \ 131 )) \ 132 ) 133 134#define _FP_FRAC_GE_4(X,Y) \ 135 (X##_f[3] > Y##_f[3] || \ 136 (X##_f[3] == Y##_f[3] && (X##_f[2] > Y##_f[2] || \ 137 (X##_f[2] == Y##_f[2] && (X##_f[1] > Y##_f[1] || \ 138 (X##_f[1] == Y##_f[1] && X##_f[0] >= Y##_f[0]) \ 139 )) \ 140 )) \ 141 ) 142 143 144#define _FP_FRAC_CLZ_4(R,X) \ 145 do { \ 146 if (X##_f[3]) \ 147 { \ 148 __FP_CLZ(R,X##_f[3]); \ 149 } \ 150 else if (X##_f[2]) \ 151 { \ 152 __FP_CLZ(R,X##_f[2]); \ 153 R += _FP_W_TYPE_SIZE; \ 154 } \ 155 else if (X##_f[1]) \ 156 { \ 157 __FP_CLZ(R,X##_f[2]); \ 158 R += _FP_W_TYPE_SIZE*2; \ 159 } \ 160 else \ 161 { \ 162 __FP_CLZ(R,X##_f[0]); \ 163 R += _FP_W_TYPE_SIZE*3; \ 164 } \ 165 } while(0) 166 167 168#define _FP_UNPACK_RAW_4(fs, X, val) \ 169 do { \ 170 union _FP_UNION_##fs _flo; _flo.flt = (val); \ 171 X##_f[0] = _flo.bits.frac0; \ 172 X##_f[1] = _flo.bits.frac1; \ 173 X##_f[2] = _flo.bits.frac2; \ 174 X##_f[3] = _flo.bits.frac3; \ 175 X##_e = _flo.bits.exp; \ 176 X##_s = _flo.bits.sign; \ 177 } while (0) 178 179#define _FP_PACK_RAW_4(fs, val, X) \ 180 do { \ 181 union _FP_UNION_##fs _flo; \ 182 _flo.bits.frac0 = X##_f[0]; \ 183 _flo.bits.frac1 = X##_f[1]; \ 184 _flo.bits.frac2 = X##_f[2]; \ 185 _flo.bits.frac3 = X##_f[3]; \ 186 _flo.bits.exp = X##_e; \ 187 _flo.bits.sign = X##_s; \ 188 (val) = _flo.flt; \ 189 } while (0) 190 191 192/* 193 * Internals 194 */ 195 196#define __FP_FRAC_SET_4(X,I3,I2,I1,I0) \ 197 (X##_f[3] = I3, X##_f[2] = I2, X##_f[1] = I1, X##_f[0] = I0) 198 199#ifndef __FP_FRAC_ADD_4 200#define __FP_FRAC_ADD_4(r3,r2,r1,r0,x3,x2,x1,x0,y3,y2,y1,y0) \ 201 (r0 = x0 + y0, \ 202 r1 = x1 + y1 + (r0 < x0), \ 203 r2 = x2 + y2 + (r1 < x1), \ 204 r3 = x3 + y3 + (r2 < x2)) 205#endif 206 207#ifndef __FP_FRAC_SUB_4 208#define __FP_FRAC_SUB_4(r3,r2,r1,r0,x3,x2,x1,x0,y3,y2,y1,y0) \ 209 (r0 = x0 - y0, \ 210 r1 = x1 - y1 - (r0 > x0), \ 211 r2 = x2 - y2 - (r1 > x1), \ 212 r3 = x3 - y3 - (r2 > x2)) 213#endif 214 215#ifndef __FP_FRAC_ADDI_4 216/* I always wanted to be a lisp programmer :-> */ 217#define __FP_FRAC_ADDI_4(x3,x2,x1,x0,i) \ 218 (x3 += ((x2 += ((x1 += ((x0 += i) < x0)) < x1) < x2))) 219#endif 220 221/* Convert FP values between word sizes. This appears to be more 222 * complicated than I'd have expected it to be, so these might be 223 * wrong... These macros are in any case somewhat bogus because they 224 * use information about what various FRAC_n variables look like 225 * internally [eg, that 2 word vars are X_f0 and x_f1]. But so do 226 * the ones in op-2.h and op-1.h. 227 */ 228#define _FP_FRAC_CONV_1_4(dfs, sfs, D, S) \ 229 do { \ 230 _FP_FRAC_SRS_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs), \ 231 _FP_WFRACBITS_##sfs); \ 232 D##_f = S##_f[0]; \ 233 } while (0) 234 235#define _FP_FRAC_CONV_2_4(dfs, sfs, D, S) \ 236 do { \ 237 _FP_FRAC_SRS_4(S, (_FP_WFRACBITS_##sfs - _FP_WFRACBITS_##dfs), \ 238 _FP_WFRACBITS_##sfs); \ 239 D##_f0 = S##_f[0]; \ 240 D##_f1 = S##_f[1]; \ 241 } while (0) 242 243/* Assembly/disassembly for converting to/from integral types. 244 * No shifting or overflow handled here. 245 */ 246/* Put the FP value X into r, which is an integer of size rsize. */ 247#define _FP_FRAC_ASSEMBLE_4(r, X, rsize) \ 248 do { \ 249 if (rsize <= _FP_W_TYPE_SIZE) \ 250 r = X##_f[0]; \ 251 else if (rsize <= 2*_FP_W_TYPE_SIZE) \ 252 { \ 253 r = X##_f[1]; \ 254 r <<= _FP_W_TYPE_SIZE; \ 255 r += X##_f[0]; \ 256 } \ 257 else \ 258 { \ 259 /* I'm feeling lazy so we deal with int == 3words (implausible)*/ \ 260 /* and int == 4words as a single case. */ \ 261 r = X##_f[3]; \ 262 r <<= _FP_W_TYPE_SIZE; \ 263 r += X##_f[2]; \ 264 r <<= _FP_W_TYPE_SIZE; \ 265 r += X##_f[1]; \ 266 r <<= _FP_W_TYPE_SIZE; \ 267 r += X##_f[0]; \ 268 } \ 269 } while (0) 270 271/* "No disassemble Number Five!" */ 272/* move an integer of size rsize into X's fractional part. We rely on 273 * the _f[] array consisting of words of size _FP_W_TYPE_SIZE to avoid 274 * having to mask the values we store into it. 275 */ 276#define _FP_FRAC_DISASSEMBLE_4(X, r, rsize) \ 277 do { \ 278 X##_f[0] = r; \ 279 X##_f[1] = (rsize <= _FP_W_TYPE_SIZE ? 0 : r >> _FP_W_TYPE_SIZE); \ 280 X##_f[2] = (rsize <= 2*_FP_W_TYPE_SIZE ? 0 : r >> 2*_FP_W_TYPE_SIZE); \ 281 X##_f[3] = (rsize <= 3*_FP_W_TYPE_SIZE ? 0 : r >> 3*_FP_W_TYPE_SIZE); \ 282 } while (0); 283 284#define _FP_FRAC_CONV_4_1(dfs, sfs, D, S) \ 285 do { \ 286 D##_f[0] = S##_f; \ 287 D##_f[1] = D##_f[2] = D##_f[3] = 0; \ 288 _FP_FRAC_SLL_4(D, (_FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs)); \ 289 } while (0) 290 291#define _FP_FRAC_CONV_4_2(dfs, sfs, D, S) \ 292 do { \ 293 D##_f[0] = S##_f0; \ 294 D##_f[1] = S##_f1; \ 295 D##_f[2] = D##_f[3] = 0; \ 296 _FP_FRAC_SLL_4(D, (_FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs)); \ 297 } while (0) 298 299#define _FP_SQRT_MEAT_4(R, S, T, X, q) 300