git.oblomov.eu Git - linux-2.6/blob - arch/powerpc/math-emu/op-1.h

   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)