2 * ====================================================
3 * Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved.
5 * Developed at SunPro, a Sun Microsystems, Inc. business.
6 * Permission to use, copy, modify, and distribute this
7 * software is freely granted, provided that this notice
9 * ====================================================
13 * from: @(#)fdlibm.h 5.1 93/09/24
16 #ifndef _MATH_PRIVATE_H_
17 #define _MATH_PRIVATE_H_
21 #include <sys/types.h>
24 #include <get-rounding-mode.h>
26 /* Gather machine dependent _Floatn support. */
27 #include <bits/floatn.h>
29 /* The original fdlibm code used statements like:
30 n0 = ((*(int*)&one)>>29)^1; * index of high word *
31 ix0 = *(n0+(int*)&x); * high word of x *
32 ix1 = *((1-n0)+(int*)&x); * low word of x *
33 to dig two 32 bit words out of the 64 bit IEEE floating point
34 value. That is non-ANSI, and, moreover, the gcc instruction
35 scheduler gets it wrong. We instead use the following macros.
36 Unlike the original code, we determine the endianness at compile
37 time, not at run time; I don't see much benefit to selecting
38 endianness at run time. */
40 /* A union which permits us to convert between a double and two 32 bit
43 #if __FLOAT_WORD_ORDER == __BIG_ENDIAN
54 } ieee_double_shape_type
;
58 #if __FLOAT_WORD_ORDER == __LITTLE_ENDIAN
69 } ieee_double_shape_type
;
73 /* Get two 32 bit ints from a double. */
75 #define EXTRACT_WORDS(ix0,ix1,d) \
77 ieee_double_shape_type ew_u; \
79 (ix0) = ew_u.parts.msw; \
80 (ix1) = ew_u.parts.lsw; \
83 /* Get the more significant 32 bit int from a double. */
86 # define GET_HIGH_WORD(i,d) \
88 ieee_double_shape_type gh_u; \
90 (i) = gh_u.parts.msw; \
94 /* Get the less significant 32 bit int from a double. */
97 # define GET_LOW_WORD(i,d) \
99 ieee_double_shape_type gl_u; \
101 (i) = gl_u.parts.lsw; \
105 /* Get all in one, efficient on 64-bit machines. */
106 #ifndef EXTRACT_WORDS64
107 # define EXTRACT_WORDS64(i,d) \
109 ieee_double_shape_type gh_u; \
115 /* Set a double from two 32 bit ints. */
117 # define INSERT_WORDS(d,ix0,ix1) \
119 ieee_double_shape_type iw_u; \
120 iw_u.parts.msw = (ix0); \
121 iw_u.parts.lsw = (ix1); \
126 /* Get all in one, efficient on 64-bit machines. */
127 #ifndef INSERT_WORDS64
128 # define INSERT_WORDS64(d,i) \
130 ieee_double_shape_type iw_u; \
136 /* Set the more significant 32 bits of a double from an int. */
137 #ifndef SET_HIGH_WORD
138 #define SET_HIGH_WORD(d,v) \
140 ieee_double_shape_type sh_u; \
142 sh_u.parts.msw = (v); \
147 /* Set the less significant 32 bits of a double from an int. */
149 # define SET_LOW_WORD(d,v) \
151 ieee_double_shape_type sl_u; \
153 sl_u.parts.lsw = (v); \
158 /* A union which permits us to convert between a float and a 32 bit
165 } ieee_float_shape_type
;
167 /* Get a 32 bit int from a float. */
168 #ifndef GET_FLOAT_WORD
169 # define GET_FLOAT_WORD(i,d) \
171 ieee_float_shape_type gf_u; \
177 /* Set a float from a 32 bit int. */
178 #ifndef SET_FLOAT_WORD
179 # define SET_FLOAT_WORD(d,i) \
181 ieee_float_shape_type sf_u; \
187 /* We need to guarantee an expansion of name when building
188 ldbl-128 files as another type (e.g _Float128). */
189 #define mathx_hidden_def(name) hidden_def(name)
191 /* Get long double macros from a separate header. */
192 #include <math_ldbl.h>
194 /* Include function declarations for each floating-point. */
195 #define _Mdouble_ double
197 #include <math_private_calls.h>
201 #define _Mdouble_ float
203 #define __MATH_DECLARING_FLOAT
204 #include <math_private_calls.h>
205 #undef __MATH_DECLARING_FLOAT
209 #define _Mdouble_ long double
211 #define __MATH_DECLARING_LONG_DOUBLE
212 #include <math_private_calls.h>
213 #undef __MATH_DECLARING_LONG_DOUBLE
217 #if __HAVE_DISTINCT_FLOAT128
218 # define _Mdouble_ _Float128
220 # define __MATH_DECLARING_FLOATN
221 # include <math_private_calls.h>
222 # undef __MATH_DECLARING_FLOATN
227 #if __HAVE_DISTINCT_FLOAT128
229 /* __builtin_isinf_sign is broken in GCC < 7 for float128. */
230 # if ! __GNUC_PREREQ (7, 0)
231 # include <ieee754_float128.h>
233 __isinff128 (_Float128 x
)
236 GET_FLOAT128_WORDS64 (hx
, lx
, x
);
237 lx
|= (hx
& 0x7fffffffffffffffLL
) ^ 0x7fff000000000000LL
;
239 return ~(lx
>> 63) & (hx
>> 62);
243 extern inline _Float128
244 fabsf128 (_Float128 x
)
246 return __builtin_fabsf128 (x
);
252 /* Prototypes for functions of the IBM Accurate Mathematical Library. */
253 extern double __exp1 (double __x
, double __xx
, double __error
);
254 extern double __sin (double __x
);
255 extern double __cos (double __x
);
256 extern int __branred (double __x
, double *__a
, double *__aa
);
257 extern void __doasin (double __x
, double __dx
, double __v
[]);
258 extern void __dubsin (double __x
, double __dx
, double __v
[]);
259 extern void __dubcos (double __x
, double __dx
, double __v
[]);
260 extern double __halfulp (double __x
, double __y
);
261 extern double __sin32 (double __x
, double __res
, double __res1
);
262 extern double __cos32 (double __x
, double __res
, double __res1
);
263 extern double __mpsin (double __x
, double __dx
, bool __range_reduce
);
264 extern double __mpcos (double __x
, double __dx
, bool __range_reduce
);
265 extern double __slowexp (double __x
);
266 extern double __slowpow (double __x
, double __y
, double __z
);
267 extern void __docos (double __x
, double __dx
, double __v
[]);
269 #ifndef math_opt_barrier
270 # define math_opt_barrier(x) \
271 ({ __typeof (x) __x = (x); __asm ("" : "+m" (__x)); __x; })
272 # define math_force_eval(x) \
273 ({ __typeof (x) __x = (x); __asm __volatile__ ("" : : "m" (__x)); })
276 /* math_narrow_eval reduces its floating-point argument to the range
277 and precision of its semantic type. (The original evaluation may
278 still occur with excess range and precision, so the result may be
279 affected by double rounding.) */
280 #if FLT_EVAL_METHOD == 0
281 # define math_narrow_eval(x) (x)
283 # if FLT_EVAL_METHOD == 1
284 # define excess_precision(type) __builtin_types_compatible_p (type, float)
286 # define excess_precision(type) (__builtin_types_compatible_p (type, float) \
287 || __builtin_types_compatible_p (type, \
290 # define math_narrow_eval(x) \
292 __typeof (x) math_narrow_eval_tmp = (x); \
293 if (excess_precision (__typeof (math_narrow_eval_tmp))) \
294 __asm__ ("" : "+m" (math_narrow_eval_tmp)); \
295 math_narrow_eval_tmp; \
299 #define fabs_tg(x) __MATH_TG ((x), (__typeof (x)) __builtin_fabs, (x))
301 /* These must be function-like macros because some __MATH_TG
302 implementations macro-expand the function-name argument before
303 concatenating a suffix to it. */
304 #define min_of_type_f() FLT_MIN
305 #define min_of_type_() DBL_MIN
306 #define min_of_type_l() LDBL_MIN
307 #define min_of_type_f128() FLT128_MIN
309 #define min_of_type(x) __MATH_TG ((x), (__typeof (x)) min_of_type_, ())
311 /* If X (which is not a NaN) is subnormal, force an underflow
313 #define math_check_force_underflow(x) \
316 __typeof (x) force_underflow_tmp = (x); \
317 if (fabs_tg (force_underflow_tmp) \
318 < min_of_type (force_underflow_tmp)) \
320 __typeof (force_underflow_tmp) force_underflow_tmp2 \
321 = force_underflow_tmp * force_underflow_tmp; \
322 math_force_eval (force_underflow_tmp2); \
326 /* Likewise, but X is also known to be nonnegative. */
327 #define math_check_force_underflow_nonneg(x) \
330 __typeof (x) force_underflow_tmp = (x); \
331 if (force_underflow_tmp \
332 < min_of_type (force_underflow_tmp)) \
334 __typeof (force_underflow_tmp) force_underflow_tmp2 \
335 = force_underflow_tmp * force_underflow_tmp; \
336 math_force_eval (force_underflow_tmp2); \
340 /* Likewise, for both real and imaginary parts of a complex
342 #define math_check_force_underflow_complex(x) \
345 __typeof (x) force_underflow_complex_tmp = (x); \
346 math_check_force_underflow (__real__ force_underflow_complex_tmp); \
347 math_check_force_underflow (__imag__ force_underflow_complex_tmp); \
351 /* The standards only specify one variant of the fenv.h interfaces.
352 But at least for some architectures we can be more efficient if we
353 know what operations are going to be performed. Therefore we
354 define additional interfaces. By default they refer to the normal
357 static __always_inline
void
358 default_libc_feholdexcept (fenv_t
*e
)
360 (void) __feholdexcept (e
);
363 #ifndef libc_feholdexcept
364 # define libc_feholdexcept default_libc_feholdexcept
366 #ifndef libc_feholdexceptf
367 # define libc_feholdexceptf default_libc_feholdexcept
369 #ifndef libc_feholdexceptl
370 # define libc_feholdexceptl default_libc_feholdexcept
373 static __always_inline
void
374 default_libc_fesetround (int r
)
376 (void) __fesetround (r
);
379 #ifndef libc_fesetround
380 # define libc_fesetround default_libc_fesetround
382 #ifndef libc_fesetroundf
383 # define libc_fesetroundf default_libc_fesetround
385 #ifndef libc_fesetroundl
386 # define libc_fesetroundl default_libc_fesetround
389 static __always_inline
void
390 default_libc_feholdexcept_setround (fenv_t
*e
, int r
)
396 #ifndef libc_feholdexcept_setround
397 # define libc_feholdexcept_setround default_libc_feholdexcept_setround
399 #ifndef libc_feholdexcept_setroundf
400 # define libc_feholdexcept_setroundf default_libc_feholdexcept_setround
402 #ifndef libc_feholdexcept_setroundl
403 # define libc_feholdexcept_setroundl default_libc_feholdexcept_setround
406 #ifndef libc_feholdsetround_53bit
407 # define libc_feholdsetround_53bit libc_feholdsetround
410 #ifndef libc_fetestexcept
411 # define libc_fetestexcept fetestexcept
413 #ifndef libc_fetestexceptf
414 # define libc_fetestexceptf fetestexcept
416 #ifndef libc_fetestexceptl
417 # define libc_fetestexceptl fetestexcept
420 static __always_inline
void
421 default_libc_fesetenv (fenv_t
*e
)
423 (void) __fesetenv (e
);
426 #ifndef libc_fesetenv
427 # define libc_fesetenv default_libc_fesetenv
429 #ifndef libc_fesetenvf
430 # define libc_fesetenvf default_libc_fesetenv
432 #ifndef libc_fesetenvl
433 # define libc_fesetenvl default_libc_fesetenv
436 static __always_inline
void
437 default_libc_feupdateenv (fenv_t
*e
)
439 (void) __feupdateenv (e
);
442 #ifndef libc_feupdateenv
443 # define libc_feupdateenv default_libc_feupdateenv
445 #ifndef libc_feupdateenvf
446 # define libc_feupdateenvf default_libc_feupdateenv
448 #ifndef libc_feupdateenvl
449 # define libc_feupdateenvl default_libc_feupdateenv
452 #ifndef libc_feresetround_53bit
453 # define libc_feresetround_53bit libc_feresetround
456 static __always_inline
int
457 default_libc_feupdateenv_test (fenv_t
*e
, int ex
)
459 int ret
= fetestexcept (ex
);
464 #ifndef libc_feupdateenv_test
465 # define libc_feupdateenv_test default_libc_feupdateenv_test
467 #ifndef libc_feupdateenv_testf
468 # define libc_feupdateenv_testf default_libc_feupdateenv_test
470 #ifndef libc_feupdateenv_testl
471 # define libc_feupdateenv_testl default_libc_feupdateenv_test
474 /* Save and set the rounding mode. The use of fenv_t to store the old mode
475 allows a target-specific version of this function to avoid converting the
476 rounding mode from the fpu format. By default we have no choice but to
477 manipulate the entire env. */
479 #ifndef libc_feholdsetround
480 # define libc_feholdsetround libc_feholdexcept_setround
482 #ifndef libc_feholdsetroundf
483 # define libc_feholdsetroundf libc_feholdexcept_setroundf
485 #ifndef libc_feholdsetroundl
486 # define libc_feholdsetroundl libc_feholdexcept_setroundl
489 /* ... and the reverse. */
491 #ifndef libc_feresetround
492 # define libc_feresetround libc_feupdateenv
494 #ifndef libc_feresetroundf
495 # define libc_feresetroundf libc_feupdateenvf
497 #ifndef libc_feresetroundl
498 # define libc_feresetroundl libc_feupdateenvl
501 /* ... and a version that may also discard exceptions. */
503 #ifndef libc_feresetround_noex
504 # define libc_feresetround_noex libc_fesetenv
506 #ifndef libc_feresetround_noexf
507 # define libc_feresetround_noexf libc_fesetenvf
509 #ifndef libc_feresetround_noexl
510 # define libc_feresetround_noexl libc_fesetenvl
514 # define HAVE_RM_CTX 0
518 /* Set/Restore Rounding Modes only when necessary. If defined, these functions
519 set/restore floating point state only if the state needed within the lexical
520 block is different from the current state. This saves a lot of time when
521 the floating point unit is much slower than the fixed point units. */
523 # ifndef libc_feholdsetround_noex_ctx
524 # define libc_feholdsetround_noex_ctx libc_feholdsetround_ctx
526 # ifndef libc_feholdsetround_noexf_ctx
527 # define libc_feholdsetround_noexf_ctx libc_feholdsetroundf_ctx
529 # ifndef libc_feholdsetround_noexl_ctx
530 # define libc_feholdsetround_noexl_ctx libc_feholdsetroundl_ctx
533 # ifndef libc_feresetround_noex_ctx
534 # define libc_feresetround_noex_ctx libc_fesetenv_ctx
536 # ifndef libc_feresetround_noexf_ctx
537 # define libc_feresetround_noexf_ctx libc_fesetenvf_ctx
539 # ifndef libc_feresetround_noexl_ctx
540 # define libc_feresetround_noexl_ctx libc_fesetenvl_ctx
545 /* Default implementation using standard fenv functions.
546 Avoid unnecessary rounding mode changes by first checking the
547 current rounding mode. Note the use of __glibc_unlikely is
548 important for performance. */
550 static __always_inline
void
551 libc_feholdsetround_ctx (struct rm_ctx
*ctx
, int round
)
553 ctx
->updated_status
= false;
555 /* Update rounding mode only if different. */
556 if (__glibc_unlikely (round
!= get_rounding_mode ()))
558 ctx
->updated_status
= true;
559 __fegetenv (&ctx
->env
);
560 __fesetround (round
);
564 static __always_inline
void
565 libc_feresetround_ctx (struct rm_ctx
*ctx
)
567 /* Restore the rounding mode if updated. */
568 if (__glibc_unlikely (ctx
->updated_status
))
569 __feupdateenv (&ctx
->env
);
572 static __always_inline
void
573 libc_feholdsetround_noex_ctx (struct rm_ctx
*ctx
, int round
)
575 /* Save exception flags and rounding mode. */
576 __fegetenv (&ctx
->env
);
578 /* Update rounding mode only if different. */
579 if (__glibc_unlikely (round
!= get_rounding_mode ()))
580 __fesetround (round
);
583 static __always_inline
void
584 libc_feresetround_noex_ctx (struct rm_ctx
*ctx
)
586 /* Restore exception flags and rounding mode. */
587 __fesetenv (&ctx
->env
);
590 # define libc_feholdsetroundf_ctx libc_feholdsetround_ctx
591 # define libc_feholdsetroundl_ctx libc_feholdsetround_ctx
592 # define libc_feresetroundf_ctx libc_feresetround_ctx
593 # define libc_feresetroundl_ctx libc_feresetround_ctx
595 # define libc_feholdsetround_noexf_ctx libc_feholdsetround_noex_ctx
596 # define libc_feholdsetround_noexl_ctx libc_feholdsetround_noex_ctx
597 # define libc_feresetround_noexf_ctx libc_feresetround_noex_ctx
598 # define libc_feresetround_noexl_ctx libc_feresetround_noex_ctx
602 #ifndef libc_feholdsetround_53bit_ctx
603 # define libc_feholdsetround_53bit_ctx libc_feholdsetround_ctx
605 #ifndef libc_feresetround_53bit_ctx
606 # define libc_feresetround_53bit_ctx libc_feresetround_ctx
609 #define SET_RESTORE_ROUND_GENERIC(RM,ROUNDFUNC,CLEANUPFUNC) \
610 struct rm_ctx ctx __attribute__((cleanup (CLEANUPFUNC ## _ctx))); \
611 ROUNDFUNC ## _ctx (&ctx, (RM))
613 /* Set the rounding mode within a lexical block. Restore the rounding mode to
614 the value at the start of the block. The exception mode must be preserved.
615 Exceptions raised within the block must be set in the exception flags.
616 Non-stop mode may be enabled inside the block. */
618 #define SET_RESTORE_ROUND(RM) \
619 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround, libc_feresetround)
620 #define SET_RESTORE_ROUNDF(RM) \
621 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundf, libc_feresetroundf)
622 #define SET_RESTORE_ROUNDL(RM) \
623 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetroundl, libc_feresetroundl)
625 /* Set the rounding mode within a lexical block. Restore the rounding mode to
626 the value at the start of the block. The exception mode must be preserved.
627 Exceptions raised within the block must be discarded, and exception flags
628 are restored to the value at the start of the block.
629 Non-stop mode may be enabled inside the block. */
631 #define SET_RESTORE_ROUND_NOEX(RM) \
632 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround_noex, \
633 libc_feresetround_noex)
634 #define SET_RESTORE_ROUND_NOEXF(RM) \
635 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround_noexf, \
636 libc_feresetround_noexf)
637 #define SET_RESTORE_ROUND_NOEXL(RM) \
638 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround_noexl, \
639 libc_feresetround_noexl)
641 /* Like SET_RESTORE_ROUND, but also set rounding precision to 53 bits. */
642 #define SET_RESTORE_ROUND_53BIT(RM) \
643 SET_RESTORE_ROUND_GENERIC (RM, libc_feholdsetround_53bit, \
644 libc_feresetround_53bit)
646 #endif /* _MATH_PRIVATE_H_ */