soft-fp/op-1.h

   1 /* Software floating-point emulation.
   2    Basic one-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 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_SRS_1(X,N,sz)  __FP_FRAC_SRS_1(X##_f, N, sz)
  44
  45 #define __FP_FRAC_SRS_1(X,N,sz)                                         \
  46    (X = (X >> (N) | (__builtin_constant_p(N) && (N) == 1                \
  47                      ? X & 1 : (X << (_FP_W_TYPE_SIZE - (N))) != 0)))
  48
  49 #define _FP_FRAC_ADD_1(R,X,Y)   (R##_f = X##_f + Y##_f)
  50 #define _FP_FRAC_SUB_1(R,X,Y)   (R##_f = X##_f - Y##_f)
  51 #define _FP_FRAC_DEC_1(X,Y)     (X##_f -= Y##_f)
  52 #define _FP_FRAC_CLZ_1(z, X)    __FP_CLZ(z, X##_f)
  53
  54 /* Predicates */
  55 #define _FP_FRAC_NEGP_1(X)      ((_FP_WS_TYPE)X##_f < 0)
  56 #define _FP_FRAC_ZEROP_1(X)     (X##_f == 0)
  57 #define _FP_FRAC_OVERP_1(fs,X)  (X##_f & _FP_OVERFLOW_##fs)
  58 #define _FP_FRAC_CLEAR_OVERP_1(fs,X)    (X##_f &= ~_FP_OVERFLOW_##fs)
  59 #define _FP_FRAC_EQ_1(X, Y)     (X##_f == Y##_f)
  60 #define _FP_FRAC_GE_1(X, Y)     (X##_f >= Y##_f)
  61 #define _FP_FRAC_GT_1(X, Y)     (X##_f > Y##_f)
  62
  63 #define _FP_ZEROFRAC_1          0
  64 #define _FP_MINFRAC_1           1
  65 #define _FP_MAXFRAC_1           (~(_FP_WS_TYPE)0)
  66
  67 /*
  68  * Unpack the raw bits of a native fp value.  Do not classify or
  69  * normalize the data.
  70  */
  71
  72 #define _FP_UNPACK_RAW_1(fs, X, val)                            \
  73   do {                                                          \
  74     union _FP_UNION_##fs _flo; _flo.flt = (val);                \
  75                                                                 \
  76     X##_f = _flo.bits.frac;                                     \
  77     X##_e = _flo.bits.exp;                                      \
  78     X##_s = _flo.bits.sign;                                     \
  79   } while (0)
  80
  81 #define _FP_UNPACK_RAW_1_P(fs, X, val)                          \
  82   do {                                                          \
  83     union _FP_UNION_##fs *_flo =                                \
  84       (union _FP_UNION_##fs *)(val);                            \
  85                                                                 \
  86     X##_f = _flo->bits.frac;                                    \
  87     X##_e = _flo->bits.exp;                                     \
  88     X##_s = _flo->bits.sign;                                    \
  89   } while (0)
  90
  91 /*
  92  * Repack the raw bits of a native fp value.
  93  */
  94
  95 #define _FP_PACK_RAW_1(fs, val, X)                              \
  96   do {                                                          \
  97     union _FP_UNION_##fs _flo;                                  \
  98                                                                 \
  99     _flo.bits.frac = X##_f;                                     \
 100     _flo.bits.exp  = X##_e;                                     \
 101     _flo.bits.sign = X##_s;                                     \
 102                                                                 \
 103     (val) = _flo.flt;                                           \
 104   } while (0)
 105
 106 #define _FP_PACK_RAW_1_P(fs, val, X)                            \
 107   do {                                                          \
 108     union _FP_UNION_##fs *_flo =                                \
 109       (union _FP_UNION_##fs *)(val);                            \
 110                                                                 \
 111     _flo->bits.frac = X##_f;                                    \
 112     _flo->bits.exp  = X##_e;                                    \
 113     _flo->bits.sign = X##_s;                                    \
 114   } while (0)
 115
 116
 117 /*
 118  * Multiplication algorithms:
 119  */
 120
 121 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
 122    multiplication immediately.  */
 123
 124 #define _FP_MUL_MEAT_1_imm(wfracbits, R, X, Y)                          \
 125   do {                                                                  \
 126     R##_f = X##_f * Y##_f;                                              \
 127     /* Normalize since we know where the msb of the multiplicands       \
 128        were (bit B), we know that the msb of the of the product is      \
 129        at either 2B or 2B-1.  */                                        \
 130     _FP_FRAC_SRS_1(R, wfracbits-1, 2*wfracbits);                        \
 131   } while (0)
 132
 133 /* Given a 1W * 1W => 2W primitive, do the extended multiplication.  */
 134
 135 #define _FP_MUL_MEAT_1_wide(wfracbits, R, X, Y, doit)                   \
 136   do {                                                                  \
 137     _FP_W_TYPE _Z_f0, _Z_f1;                                            \
 138     doit(_Z_f1, _Z_f0, X##_f, Y##_f);                                   \
 139     /* Normalize since we know where the msb of the multiplicands       \
 140        were (bit B), we know that the msb of the of the product is      \
 141        at either 2B or 2B-1.  */                                        \
 142     _FP_FRAC_SRS_2(_Z, wfracbits-1, 2*wfracbits);                       \
 143     R##_f = _Z_f0;                                                      \
 144   } while (0)
 145
 146 /* Finally, a simple widening multiply algorithm.  What fun!  */
 147
 148 #define _FP_MUL_MEAT_1_hard(wfracbits, R, X, Y)                         \
 149   do {                                                                  \
 150     _FP_W_TYPE _xh, _xl, _yh, _yl, _z_f0, _z_f1, _a_f0, _a_f1;          \
 151                                                                         \
 152     /* split the words in half */                                       \
 153     _xh = X##_f >> (_FP_W_TYPE_SIZE/2);                                 \
 154     _xl = X##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);         \
 155     _yh = Y##_f >> (_FP_W_TYPE_SIZE/2);                                 \
 156     _yl = Y##_f & (((_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2)) - 1);         \
 157                                                                         \
 158     /* multiply the pieces */                                           \
 159     _z_f0 = _xl * _yl;                                                  \
 160     _a_f0 = _xh * _yl;                                                  \
 161     _a_f1 = _xl * _yh;                                                  \
 162     _z_f1 = _xh * _yh;                                                  \
 163                                                                         \
 164     /* reassemble into two full words */                                \
 165     if ((_a_f0 += _a_f1) < _a_f1)                                       \
 166       _z_f1 += (_FP_W_TYPE)1 << (_FP_W_TYPE_SIZE/2);                    \
 167     _a_f1 = _a_f0 >> (_FP_W_TYPE_SIZE/2);                               \
 168     _a_f0 = _a_f0 << (_FP_W_TYPE_SIZE/2);                               \
 169     _FP_FRAC_ADD_2(_z, _z, _a);                                         \
 170                                                                         \
 171     /* normalize */                                                     \
 172     _FP_FRAC_SRS_2(_z, wfracbits - 1, 2*wfracbits);                     \
 173     R##_f = _z_f0;                                                      \
 174   } while (0)
 175
 176
 177 /*
 178  * Division algorithms:
 179  */
 180
 181 /* Basic.  Assuming the host word size is >= 2*FRACBITS, we can do the
 182    division immediately.  Give this macro either _FP_DIV_HELP_imm for
 183    C primitives or _FP_DIV_HELP_ldiv for the ISO function.  Which you
 184    choose will depend on what the compiler does with divrem4.  */
 185
 186 #define _FP_DIV_MEAT_1_imm(fs, R, X, Y, doit)           \
 187   do {                                                  \
 188     _FP_W_TYPE _q, _r;                                  \
 189     X##_f <<= (X##_f < Y##_f                            \
 190                ? R##_e--, _FP_WFRACBITS_##fs            \
 191                : _FP_WFRACBITS_##fs - 1);               \
 192     doit(_q, _r, X##_f, Y##_f);                         \
 193     R##_f = _q | (_r != 0);                             \
 194   } while (0)
 195
 196 /* GCC's longlong.h defines a 2W / 1W => (1W,1W) primitive udiv_qrnnd
 197    that may be useful in this situation.  This first is for a primitive
 198    that requires normalization, the second for one that does not.  Look
 199    for UDIV_NEEDS_NORMALIZATION to tell which your machine needs.  */
 200
 201 #define _FP_DIV_MEAT_1_udiv_norm(fs, R, X, Y)                           \
 202   do {                                                                  \
 203     _FP_W_TYPE _nh, _nl, _q, _r, _y;                                    \
 204                                                                         \
 205     /* Normalize Y -- i.e. make the most significant bit set.  */       \
 206     _y = Y##_f << _FP_WFRACXBITS_##fs;                                  \
 207                                                                         \
 208     /* Shift X op correspondingly high, that is, up one full word.  */  \
 209     if (X##_f < Y##_f)                                                  \
 210       {                                                                 \
 211         R##_e--;                                                        \
 212         _nl = 0;                                                        \
 213         _nh = X##_f;                                                    \
 214       }                                                                 \
 215     else                                                                \
 216       {                                                                 \
 217         _nl = X##_f << (_FP_W_TYPE_SIZE - 1);                           \
 218         _nh = X##_f >> 1;                                               \
 219       }                                                                 \
 220                                                                         \
 221     udiv_qrnnd(_q, _r, _nh, _nl, _y);                                   \
 222     R##_f = _q | (_r != 0);                                             \
 223   } while (0)
 224
 225 #define _FP_DIV_MEAT_1_udiv(fs, R, X, Y)                \
 226   do {                                                  \
 227     _FP_W_TYPE _nh, _nl, _q, _r;                        \
 228     if (X##_f < Y##_f)                                  \
 229       {                                                 \
 230         R##_e--;                                        \
 231         _nl = X##_f << _FP_WFRACBITS_##fs;              \
 232         _nh = X##_f >> _FP_WFRACXBITS_##fs;             \
 233       }                                                 \
 234     else                                                \
 235       {                                                 \
 236         _nl = X##_f << (_FP_WFRACBITS_##fs - 1);        \
 237         _nh = X##_f >> (_FP_WFRACXBITS_##fs + 1);       \
 238       }                                                 \
 239     udiv_qrnnd(_q, _r, _nh, _nl, Y##_f);                \
 240     R##_f = _q | (_r != 0);                             \
 241   } while (0)
 242
 243
 244 /*
 245  * Square root algorithms:
 246  * We have just one right now, maybe Newton approximation
 247  * should be added for those machines where division is fast.
 248  */
 249
 250 #define _FP_SQRT_MEAT_1(R, S, T, X, q)                  \
 251   do {                                                  \
 252     while (q != _FP_WORK_ROUND)                         \
 253       {                                                 \
 254         T##_f = S##_f + q;                              \
 255         if (T##_f <= X##_f)                             \
 256           {                                             \
 257             S##_f = T##_f + q;                          \
 258             X##_f -= T##_f;                             \
 259             R##_f += q;                                 \
 260           }                                             \
 261         _FP_FRAC_SLL_1(X, 1);                           \
 262         q >>= 1;                                        \
 263       }                                                 \
 264     if (X##_f)                                          \
 265       {                                                 \
 266         if (S##_f < X##_f)                              \
 267           R##_f |= _FP_WORK_ROUND;                      \
 268         R##_f |= _FP_WORK_STICKY;                       \
 269       }                                                 \
 270   } while (0)
 271
 272 /*
 273  * Assembly/disassembly for converting to/from integral types.
 274  * No shifting or overflow handled here.
 275  */
 276
 277 #define _FP_FRAC_ASSEMBLE_1(r, X, rsize)        (r = X##_f)
 278 #define _FP_FRAC_DISASSEMBLE_1(X, r, rsize)     (X##_f = r)
 279
 280
 281 /*
 282  * Convert FP values between word sizes
 283  */
 284
 285 #define _FP_FRAC_CONV_1_1(dfs, sfs, D, S)                               \
 286   do {                                                                  \
 287     D##_f = S##_f;                                                      \
 288     if (_FP_WFRACBITS_##sfs > _FP_WFRACBITS_##dfs)                      \
 289       {                                                                 \
 290         if (S##_c != FP_CLS_NAN)                                        \
 291           _FP_FRAC_SRS_1(D, (_FP_WFRACBITS_##sfs-_FP_WFRACBITS_##dfs),  \
 292                          _FP_WFRACBITS_##sfs);                          \
 293         else                                                            \
 294           _FP_FRAC_SRL_1(D, (_FP_WFRACBITS_##sfs-_FP_WFRACBITS_##dfs)); \
 295       }                                                                 \
 296     else                                                                \
 297       D##_f <<= _FP_WFRACBITS_##dfs - _FP_WFRACBITS_##sfs;              \
 298   } while (0)