Home | History | Annotate | Line # | Download | only in rs6000
ibm-ldouble.c revision 1.1.1.1.8.2
      1 /* 128-bit long double support routines for Darwin.
      2    Copyright (C) 1993-2013 Free Software Foundation, Inc.
      3 
      4 This file is part of GCC.
      5 
      6 GCC is free software; you can redistribute it and/or modify it under
      7 the terms of the GNU General Public License as published by the Free
      8 Software Foundation; either version 3, or (at your option) any later
      9 version.
     10 
     11 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
     12 WARRANTY; without even the implied warranty of MERCHANTABILITY or
     13 FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
     14 for more details.
     15 
     16 Under Section 7 of GPL version 3, you are granted additional
     17 permissions described in the GCC Runtime Library Exception, version
     18 3.1, as published by the Free Software Foundation.
     19 
     20 You should have received a copy of the GNU General Public License and
     21 a copy of the GCC Runtime Library Exception along with this program;
     22 see the files COPYING3 and COPYING.RUNTIME respectively.  If not, see
     23 <http://www.gnu.org/licenses/>.  */
     24 
     25 
     26 /* Implementations of floating-point long double basic arithmetic
     27    functions called by the IBM C compiler when generating code for
     28    PowerPC platforms.  In particular, the following functions are
     29    implemented: __gcc_qadd, __gcc_qsub, __gcc_qmul, and __gcc_qdiv.
     30    Double-double algorithms are based on the paper "Doubled-Precision
     31    IEEE Standard 754 Floating-Point Arithmetic" by W. Kahan, February 26,
     32    1987.  An alternative published reference is "Software for
     33    Doubled-Precision Floating-Point Computations", by Seppo Linnainmaa,
     34    ACM TOMS vol 7 no 3, September 1981, pages 272-283.  */
     35 
     36 /* Each long double is made up of two IEEE doubles.  The value of the
     37    long double is the sum of the values of the two parts.  The most
     38    significant part is required to be the value of the long double
     39    rounded to the nearest double, as specified by IEEE.  For Inf
     40    values, the least significant part is required to be one of +0.0 or
     41    -0.0.  No other requirements are made; so, for example, 1.0 may be
     42    represented as (1.0, +0.0) or (1.0, -0.0), and the low part of a
     43    NaN is don't-care.
     44 
     45    This code currently assumes the most significant double is in
     46    the lower numbered register or lower addressed memory.  */
     47 
     48 #if defined (__MACH__) || defined (__powerpc__) || defined (_AIX)
     49 
     50 #define fabs(x) __builtin_fabs(x)
     51 #define isless(x, y) __builtin_isless (x, y)
     52 #define inf() __builtin_inf()
     53 
     54 #define unlikely(x) __builtin_expect ((x), 0)
     55 
     56 #define nonfinite(a) unlikely (! isless (fabs (a), inf ()))
     57 
     58 /* Define ALIASNAME as a strong alias for NAME.  */
     59 # define strong_alias(name, aliasname) _strong_alias(name, aliasname)
     60 # define _strong_alias(name, aliasname) \
     61   extern __typeof (name) aliasname __attribute__ ((alias (#name)));
     62 
     63 /* All these routines actually take two long doubles as parameters,
     64    but GCC currently generates poor code when a union is used to turn
     65    a long double into a pair of doubles.  */
     66 
     67 long double __gcc_qadd (double, double, double, double);
     68 long double __gcc_qsub (double, double, double, double);
     69 long double __gcc_qmul (double, double, double, double);
     70 long double __gcc_qdiv (double, double, double, double);
     71 
     72 #if defined __ELF__ && defined SHARED \
     73     && (defined __powerpc64__ || !(defined __linux__ || defined __gnu_hurd__))
     74 /* Provide definitions of the old symbol names to satisfy apps and
     75    shared libs built against an older libgcc.  To access the _xlq
     76    symbols an explicit version reference is needed, so these won't
     77    satisfy an unadorned reference like _xlqadd.  If dot symbols are
     78    not needed, the assembler will remove the aliases from the symbol
     79    table.  */
     80 __asm__ (".symver __gcc_qadd,_xlqadd (at) GCC_3.4\n\t"
     81 	 ".symver __gcc_qsub,_xlqsub (at) GCC_3.4\n\t"
     82 	 ".symver __gcc_qmul,_xlqmul (at) GCC_3.4\n\t"
     83 	 ".symver __gcc_qdiv,_xlqdiv (at) GCC_3.4\n\t"
     84 	 ".symver .__gcc_qadd,._xlqadd (at) GCC_3.4\n\t"
     85 	 ".symver .__gcc_qsub,._xlqsub (at) GCC_3.4\n\t"
     86 	 ".symver .__gcc_qmul,._xlqmul (at) GCC_3.4\n\t"
     87 	 ".symver .__gcc_qdiv,._xlqdiv (at) GCC_3.4");
     88 #endif
     89 
     90 typedef union
     91 {
     92   long double ldval;
     93   double dval[2];
     94 } longDblUnion;
     95 
     96 /* Add two 'long double' values and return the result.	*/
     97 long double
     98 __gcc_qadd (double a, double aa, double c, double cc)
     99 {
    100   longDblUnion x;
    101   double z, q, zz, xh;
    102 
    103   z = a + c;
    104 
    105   if (nonfinite (z))
    106     {
    107       z = cc + aa + c + a;
    108       if (nonfinite (z))
    109 	return z;
    110       x.dval[0] = z;  /* Will always be DBL_MAX.  */
    111       zz = aa + cc;
    112       if (fabs(a) > fabs(c))
    113 	x.dval[1] = a - z + c + zz;
    114       else
    115 	x.dval[1] = c - z + a + zz;
    116     }
    117   else
    118     {
    119       q = a - z;
    120       zz = q + c + (a - (q + z)) + aa + cc;
    121 
    122       /* Keep -0 result.  */
    123       if (zz == 0.0)
    124 	return z;
    125 
    126       xh = z + zz;
    127       if (nonfinite (xh))
    128 	return xh;
    129 
    130       x.dval[0] = xh;
    131       x.dval[1] = z - xh + zz;
    132     }
    133   return x.ldval;
    134 }
    135 
    136 long double
    137 __gcc_qsub (double a, double b, double c, double d)
    138 {
    139   return __gcc_qadd (a, b, -c, -d);
    140 }
    141 
    142 #ifdef __NO_FPRS__
    143 static double fmsub (double, double, double);
    144 #endif
    145 
    146 long double
    147 __gcc_qmul (double a, double b, double c, double d)
    148 {
    149   longDblUnion z;
    150   double t, tau, u, v, w;
    151 
    152   t = a * c;			/* Highest order double term.  */
    153 
    154   if (unlikely (t == 0)		/* Preserve -0.  */
    155       || nonfinite (t))
    156     return t;
    157 
    158   /* Sum terms of two highest orders. */
    159 
    160   /* Use fused multiply-add to get low part of a * c.  */
    161 #ifndef __NO_FPRS__
    162   asm ("fmsub %0,%1,%2,%3" : "=f"(tau) : "f"(a), "f"(c), "f"(t));
    163 #else
    164   tau = fmsub (a, c, t);
    165 #endif
    166   v = a*d;
    167   w = b*c;
    168   tau += v + w;	    /* Add in other second-order terms.	 */
    169   u = t + tau;
    170 
    171   /* Construct long double result.  */
    172   if (nonfinite (u))
    173     return u;
    174   z.dval[0] = u;
    175   z.dval[1] = (t - u) + tau;
    176   return z.ldval;
    177 }
    178 
    179 long double
    180 __gcc_qdiv (double a, double b, double c, double d)
    181 {
    182   longDblUnion z;
    183   double s, sigma, t, tau, u, v, w;
    184 
    185   t = a / c;                    /* highest order double term */
    186 
    187   if (unlikely (t == 0)		/* Preserve -0.  */
    188       || nonfinite (t))
    189     return t;
    190 
    191   /* Finite nonzero result requires corrections to the highest order
    192      term.  These corrections require the low part of c * t to be
    193      exactly represented in double.  */
    194   if (fabs (a) <= 0x1p-969)
    195     {
    196       a *= 0x1p106;
    197       b *= 0x1p106;
    198       c *= 0x1p106;
    199       d *= 0x1p106;
    200     }
    201 
    202   s = c * t;                    /* (s,sigma) = c*t exactly.  */
    203   w = -(-b + d * t);	/* Written to get fnmsub for speed, but not
    204 			   numerically necessary.  */
    205 
    206   /* Use fused multiply-add to get low part of c * t.	 */
    207 #ifndef __NO_FPRS__
    208   asm ("fmsub %0,%1,%2,%3" : "=f"(sigma) : "f"(c), "f"(t), "f"(s));
    209 #else
    210   sigma = fmsub (c, t, s);
    211 #endif
    212   v = a - s;
    213 
    214   tau = ((v-sigma)+w)/c;   /* Correction to t.  */
    215   u = t + tau;
    216 
    217   /* Construct long double result.  */
    218   if (nonfinite (u))
    219     return u;
    220   z.dval[0] = u;
    221   z.dval[1] = (t - u) + tau;
    222   return z.ldval;
    223 }
    224 
    225 #if defined (_SOFT_DOUBLE) && defined (__LONG_DOUBLE_128__)
    226 
    227 long double __gcc_qneg (double, double);
    228 int __gcc_qeq (double, double, double, double);
    229 int __gcc_qne (double, double, double, double);
    230 int __gcc_qge (double, double, double, double);
    231 int __gcc_qle (double, double, double, double);
    232 long double __gcc_stoq (float);
    233 long double __gcc_dtoq (double);
    234 float __gcc_qtos (double, double);
    235 double __gcc_qtod (double, double);
    236 int __gcc_qtoi (double, double);
    237 unsigned int __gcc_qtou (double, double);
    238 long double __gcc_itoq (int);
    239 long double __gcc_utoq (unsigned int);
    240 
    241 extern int __eqdf2 (double, double);
    242 extern int __ledf2 (double, double);
    243 extern int __gedf2 (double, double);
    244 
    245 /* Negate 'long double' value and return the result.	*/
    246 long double
    247 __gcc_qneg (double a, double aa)
    248 {
    249   longDblUnion x;
    250 
    251   x.dval[0] = -a;
    252   x.dval[1] = -aa;
    253   return x.ldval;
    254 }
    255 
    256 /* Compare two 'long double' values for equality.  */
    257 int
    258 __gcc_qeq (double a, double aa, double c, double cc)
    259 {
    260   if (__eqdf2 (a, c) == 0)
    261     return __eqdf2 (aa, cc);
    262   return 1;
    263 }
    264 
    265 strong_alias (__gcc_qeq, __gcc_qne);
    266 
    267 /* Compare two 'long double' values for less than or equal.  */
    268 int
    269 __gcc_qle (double a, double aa, double c, double cc)
    270 {
    271   if (__eqdf2 (a, c) == 0)
    272     return __ledf2 (aa, cc);
    273   return __ledf2 (a, c);
    274 }
    275 
    276 strong_alias (__gcc_qle, __gcc_qlt);
    277 
    278 /* Compare two 'long double' values for greater than or equal.  */
    279 int
    280 __gcc_qge (double a, double aa, double c, double cc)
    281 {
    282   if (__eqdf2 (a, c) == 0)
    283     return __gedf2 (aa, cc);
    284   return __gedf2 (a, c);
    285 }
    286 
    287 strong_alias (__gcc_qge, __gcc_qgt);
    288 
    289 /* Convert single to long double.  */
    290 long double
    291 __gcc_stoq (float a)
    292 {
    293   longDblUnion x;
    294 
    295   x.dval[0] = (double) a;
    296   x.dval[1] = 0.0;
    297 
    298   return x.ldval;
    299 }
    300 
    301 /* Convert double to long double.  */
    302 long double
    303 __gcc_dtoq (double a)
    304 {
    305   longDblUnion x;
    306 
    307   x.dval[0] = a;
    308   x.dval[1] = 0.0;
    309 
    310   return x.ldval;
    311 }
    312 
    313 /* Convert long double to single.  */
    314 float
    315 __gcc_qtos (double a, double aa __attribute__ ((__unused__)))
    316 {
    317   return (float) a;
    318 }
    319 
    320 /* Convert long double to double.  */
    321 double
    322 __gcc_qtod (double a, double aa __attribute__ ((__unused__)))
    323 {
    324   return a;
    325 }
    326 
    327 /* Convert long double to int.  */
    328 int
    329 __gcc_qtoi (double a, double aa)
    330 {
    331   double z = a + aa;
    332   return (int) z;
    333 }
    334 
    335 /* Convert long double to unsigned int.  */
    336 unsigned int
    337 __gcc_qtou (double a, double aa)
    338 {
    339   double z = a + aa;
    340   return (unsigned int) z;
    341 }
    342 
    343 /* Convert int to long double.  */
    344 long double
    345 __gcc_itoq (int a)
    346 {
    347   return __gcc_dtoq ((double) a);
    348 }
    349 
    350 /* Convert unsigned int to long double.  */
    351 long double
    352 __gcc_utoq (unsigned int a)
    353 {
    354   return __gcc_dtoq ((double) a);
    355 }
    356 
    357 #endif
    358 
    359 #ifdef __NO_FPRS__
    360 
    361 int __gcc_qunord (double, double, double, double);
    362 
    363 extern int __eqdf2 (double, double);
    364 extern int __unorddf2 (double, double);
    365 
    366 /* Compare two 'long double' values for unordered.  */
    367 int
    368 __gcc_qunord (double a, double aa, double c, double cc)
    369 {
    370   if (__eqdf2 (a, c) == 0)
    371     return __unorddf2 (aa, cc);
    372   return __unorddf2 (a, c);
    373 }
    374 
    375 #include "soft-fp/soft-fp.h"
    376 #include "soft-fp/double.h"
    377 #include "soft-fp/quad.h"
    378 
    379 /* Compute floating point multiply-subtract with higher (quad) precision.  */
    380 static double
    381 fmsub (double a, double b, double c)
    382 {
    383     FP_DECL_EX;
    384     FP_DECL_D(A);
    385     FP_DECL_D(B);
    386     FP_DECL_D(C);
    387     FP_DECL_Q(X);
    388     FP_DECL_Q(Y);
    389     FP_DECL_Q(Z);
    390     FP_DECL_Q(U);
    391     FP_DECL_Q(V);
    392     FP_DECL_D(R);
    393     double r;
    394     long double u, x, y, z;
    395 
    396     FP_INIT_ROUNDMODE;
    397     FP_UNPACK_RAW_D (A, a);
    398     FP_UNPACK_RAW_D (B, b);
    399     FP_UNPACK_RAW_D (C, c);
    400 
    401     /* Extend double to quad.  */
    402 #if (2 * _FP_W_TYPE_SIZE) < _FP_FRACBITS_Q
    403     FP_EXTEND(Q,D,4,2,X,A);
    404     FP_EXTEND(Q,D,4,2,Y,B);
    405     FP_EXTEND(Q,D,4,2,Z,C);
    406 #else
    407     FP_EXTEND(Q,D,2,1,X,A);
    408     FP_EXTEND(Q,D,2,1,Y,B);
    409     FP_EXTEND(Q,D,2,1,Z,C);
    410 #endif
    411     FP_PACK_RAW_Q(x,X);
    412     FP_PACK_RAW_Q(y,Y);
    413     FP_PACK_RAW_Q(z,Z);
    414     FP_HANDLE_EXCEPTIONS;
    415 
    416     /* Multiply.  */
    417     FP_INIT_ROUNDMODE;
    418     FP_UNPACK_Q(X,x);
    419     FP_UNPACK_Q(Y,y);
    420     FP_MUL_Q(U,X,Y);
    421     FP_PACK_Q(u,U);
    422     FP_HANDLE_EXCEPTIONS;
    423 
    424     /* Subtract.  */
    425     FP_INIT_ROUNDMODE;
    426     FP_UNPACK_SEMIRAW_Q(U,u);
    427     FP_UNPACK_SEMIRAW_Q(Z,z);
    428     FP_SUB_Q(V,U,Z);
    429 
    430     /* Truncate quad to double.  */
    431 #if (2 * _FP_W_TYPE_SIZE) < _FP_FRACBITS_Q
    432     V_f[3] &= 0x0007ffff;
    433     FP_TRUNC(D,Q,2,4,R,V);
    434 #else
    435     V_f1 &= 0x0007ffffffffffffL;
    436     FP_TRUNC(D,Q,1,2,R,V);
    437 #endif
    438     FP_PACK_SEMIRAW_D(r,R);
    439     FP_HANDLE_EXCEPTIONS;
    440 
    441     return r;
    442 }
    443 
    444 #endif
    445 
    446 #endif
    447