glibc/sysdeps/x86_64/fpu/math_private.h

200 lines
6.0 KiB
C
Raw Normal View History

[BZ #3306] 2007-03-27 Jakub Jelinek <jakub@redhat.com> [BZ #3306] * math/math_private.h (math_opt_barrier, math_force_eval): Define. * sysdeps/i386/fpu/math_private.h: New file. * sysdeps/x86_64/fpu/math_private.h: New file. * math/s_nexttowardf.c (__nexttowardf): Use math_opt_barrier and math_force_eval macros. Use "+m" constraint on asm rather than "=m" and "m". * math/s_nextafter.c (__nextafter): Likewise. * sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c (__nexttoward): Likewise. * sysdeps/ieee754/flt-32/s_nextafterf.c (__nextafterf): Likewise. * sysdeps/ieee754/ldbl-128/s_nexttoward.c (__nexttoward): Likewise. * sysdeps/ieee754/ldbl-96/s_nexttoward.c (__nexttoward): Likewise. * sysdeps/i386/fpu/s_nextafterl.c (__nextafterl): Use math_opt_barrier and math_force_eval macros. * sysdeps/ieee754/ldbl-128/s_nextafterl.c (__nextafterl): Likewise. * sysdeps/ieee754/ldbl-96/s_nextafterl.c (__nextafterl): Likewise. * sysdeps/i386/fpu/s_nexttoward.c: Include float.h. (__nexttoward): Use math_opt_barrier and math_force_eval macros. Use "+m" constraint on asm rather than "=m" and "m". Only use asm to force double result if FLT_EVAL_METHOD is 2. * sysdeps/i386/fpu/s_nexttowardf.c: Include float.h. (__nexttowardf): Use math_opt_barrier and math_force_eval macros. Use "+m" constraint on asm rather than "=m" and "m". Only use asm to force double result if FLT_EVAL_METHOD is not 0. * sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c: Include float.h. (__nexttowardf): Use math_opt_barrier and math_force_eval macros. If FLT_EVAL_METHOD is not 0, force x to float using asm. * sysdeps/ieee754/ldbl-opt/s_nexttowardfd.c: Include float.h. (__nldbl_nexttowardf): Use math_opt_barrier and math_force_eval macros. If FLT_EVAL_METHOD is not 0, force x to float using asm. * sysdeps/ieee754/ldbl-96/s_nexttowardf.c: Include float.h. (__nexttowardf): Use math_opt_barrier and math_force_eval macros. If FLT_EVAL_METHOD is not 0, force x to float using asm. * math/bug-nextafter.c (zero, inf): New variables. (main): Add new tests. * math/bug-nexttoward.c (zero, inf): New variables. (main): Add new tests.
2007-04-16 22:41:42 +02:00
#ifndef _MATH_PRIVATE_H
#define math_opt_barrier(x) \
({ __typeof(x) __x; \
if (sizeof (x) <= sizeof (double)) \
__asm ("" : "=x" (__x) : "0" (x)); \
else \
__asm ("" : "=t" (__x) : "0" (x)); \
__x; })
#define math_force_eval(x) \
do \
{ \
if (sizeof (x) <= sizeof (double)) \
__asm __volatile ("" : : "x" (x)); \
else \
__asm __volatile ("" : : "f" (x)); \
} \
while (0)
#include <math/math_private.h>
/* We can do a few things better on x86-64. */
/* Direct movement of float into integer register. */
#undef EXTRACT_WORDS64
#define EXTRACT_WORDS64(i,d) \
do { \
long int i_; \
asm ("movd %1, %0" : "=rm" (i_) : "x" (d)); \
(i) = i_; \
} while (0)
/* And the reverse. */
#undef INSERT_WORDS64
#define INSERT_WORDS64(d,i) \
do { \
long int i_ = i; \
asm ("movd %1, %0" : "=x" (d) : "rm" (i_)); \
} while (0)
/* Direct movement of float into integer register. */
#undef GET_FLOAT_WORD
#define GET_FLOAT_WORD(i,d) \
do { \
int i_; \
asm ("movd %1, %0" : "=rm" (i_) : "x" (d)); \
(i) = i_; \
} while (0)
/* And the reverse. */
#undef SET_FLOAT_WORD
#define SET_FLOAT_WORD(d,i) \
do { \
int i_ = i; \
asm ("movd %1, %0" : "=x" (d) : "rm" (i_)); \
} while (0)
[BZ #3306] 2007-03-27 Jakub Jelinek <jakub@redhat.com> [BZ #3306] * math/math_private.h (math_opt_barrier, math_force_eval): Define. * sysdeps/i386/fpu/math_private.h: New file. * sysdeps/x86_64/fpu/math_private.h: New file. * math/s_nexttowardf.c (__nexttowardf): Use math_opt_barrier and math_force_eval macros. Use "+m" constraint on asm rather than "=m" and "m". * math/s_nextafter.c (__nextafter): Likewise. * sysdeps/ieee754/ldbl-128ibm/s_nexttoward.c (__nexttoward): Likewise. * sysdeps/ieee754/flt-32/s_nextafterf.c (__nextafterf): Likewise. * sysdeps/ieee754/ldbl-128/s_nexttoward.c (__nexttoward): Likewise. * sysdeps/ieee754/ldbl-96/s_nexttoward.c (__nexttoward): Likewise. * sysdeps/i386/fpu/s_nextafterl.c (__nextafterl): Use math_opt_barrier and math_force_eval macros. * sysdeps/ieee754/ldbl-128/s_nextafterl.c (__nextafterl): Likewise. * sysdeps/ieee754/ldbl-96/s_nextafterl.c (__nextafterl): Likewise. * sysdeps/i386/fpu/s_nexttoward.c: Include float.h. (__nexttoward): Use math_opt_barrier and math_force_eval macros. Use "+m" constraint on asm rather than "=m" and "m". Only use asm to force double result if FLT_EVAL_METHOD is 2. * sysdeps/i386/fpu/s_nexttowardf.c: Include float.h. (__nexttowardf): Use math_opt_barrier and math_force_eval macros. Use "+m" constraint on asm rather than "=m" and "m". Only use asm to force double result if FLT_EVAL_METHOD is not 0. * sysdeps/ieee754/ldbl-128ibm/s_nexttowardf.c: Include float.h. (__nexttowardf): Use math_opt_barrier and math_force_eval macros. If FLT_EVAL_METHOD is not 0, force x to float using asm. * sysdeps/ieee754/ldbl-opt/s_nexttowardfd.c: Include float.h. (__nldbl_nexttowardf): Use math_opt_barrier and math_force_eval macros. If FLT_EVAL_METHOD is not 0, force x to float using asm. * sysdeps/ieee754/ldbl-96/s_nexttowardf.c: Include float.h. (__nexttowardf): Use math_opt_barrier and math_force_eval macros. If FLT_EVAL_METHOD is not 0, force x to float using asm. * math/bug-nextafter.c (zero, inf): New variables. (main): Add new tests. * math/bug-nexttoward.c (zero, inf): New variables. (main): Add new tests.
2007-04-16 22:41:42 +02:00
#endif
2011-10-08 16:18:26 +02:00
#define __isnan(d) \
2011-10-17 22:00:39 +02:00
({ long int __di; EXTRACT_WORDS64 (__di, (double) (d)); \
2011-10-08 16:18:26 +02:00
(__di & 0x7fffffffffffffffl) > 0x7ff0000000000000l; })
#define __isnanf(d) \
({ int __di; GET_FLOAT_WORD (__di, (float) d); \
2011-10-08 16:18:26 +02:00
(__di & 0x7fffffff) > 0x7f800000; })
#define __isinf_ns(d) \
2011-10-17 22:00:39 +02:00
({ long int __di; EXTRACT_WORDS64 (__di, (double) (d)); \
2011-10-08 16:18:26 +02:00
(__di & 0x7fffffffffffffffl) == 0x7ff0000000000000l; })
#define __isinf_nsf(d) \
({ int __di; GET_FLOAT_WORD (__di, (float) d); \
2011-10-08 16:18:26 +02:00
(__di & 0x7fffffff) == 0x7f800000; })
#define __finite(d) \
2011-10-17 22:00:39 +02:00
({ long int __di; EXTRACT_WORDS64 (__di, (double) (d)); \
2011-10-08 16:18:26 +02:00
(__di & 0x7fffffffffffffffl) < 0x7ff0000000000000l; })
#define __finitef(d) \
({ int __di; GET_FLOAT_WORD (__di, (float) d); \
2011-10-08 16:18:26 +02:00
(__di & 0x7fffffff) < 0x7f800000; })
#define __ieee754_sqrt(d) \
({ double __res; \
2011-10-17 22:00:39 +02:00
asm ("sqrtsd %1, %0" : "=x" (__res) : "xm" ((double) (d))); \
__res; })
#define __ieee754_sqrtf(d) \
({ float __res; \
2011-10-17 22:00:39 +02:00
asm ("sqrtss %1, %0" : "=x" (__res) : "xm" ((float) (d))); \
__res; })
#define __ieee754_sqrtl(d) \
({ long double __res; \
2011-10-17 22:00:39 +02:00
asm ("fsqrt" : "=t" (__res) : "0" ((long double) (d))); \
__res; })
#ifdef __SSE4_1__
# ifndef __rint
# define __rint(d) \
({ double __res; \
asm ("roundsd $4, %1, %0" : "=x" (__res) : "xm" ((double) (d))); \
__res; })
# endif
# ifndef __rintf
# define __rintf(d) \
({ float __res; \
asm ("roundss $4, %1, %0" : "=x" (__res) : "xm" ((float) (d))); \
__res; })
# endif
# ifndef __floor
# define __floor(d) \
({ double __res; \
asm ("roundsd $1, %1, %0" : "=x" (__res) : "xm" ((double) (d))); \
__res; })
# endif
# ifndef __floorf
# define __floorf(d) \
({ float __res; \
asm ("roundss $1, %1, %0" : "=x" (__res) : "xm" ((float) (d))); \
__res; })
# endif
#endif
/* Specialized variants of the <fenv.h> interfaces which only handle
either the FPU or the SSE unit. */
#undef libc_fegetround
#define libc_fegetround() \
({ \
unsigned int mxcsr; \
asm volatile ("stmxcsr %0" : "=m" (*&mxcsr)); \
(mxcsr & 0x6000) >> 3; \
})
#undef libc_fegetroundf
#define libc_fegetroundf() libc_fegetround ()
// #define libc_fegetroundl() fegetround ()
#undef libc_fesetround
#define libc_fesetround(r) \
do { \
unsigned int mxcsr; \
asm ("stmxcsr %0" : "=m" (*&mxcsr)); \
mxcsr = (mxcsr & ~0x6000) | ((r) << 3); \
asm volatile ("ldmxcsr %0" : : "m" (*&mxcsr)); \
} while (0)
#undef libc_fesetroundf
#define libc_fesetroundf(r) libc_fesetround (r)
// #define libc_fesetroundl(r) (void) fesetround (r)
#undef libc_feholdexcept
#define libc_feholdexcept(e) \
do { \
unsigned int mxcsr; \
asm ("stmxcsr %0" : "=m" (*&mxcsr)); \
(e)->__mxcsr = mxcsr; \
mxcsr = (mxcsr | 0x1f80) & ~0x3f; \
asm volatile ("ldmxcsr %0" : : "m" (*&mxcsr)); \
} while (0)
#undef libc_feholdexceptf
#define libc_feholdexceptf(e) libc_feholdexcept (e)
// #define libc_feholdexceptl(e) (void) feholdexcept (e)
#undef libc_feholdexcept_setround
#define libc_feholdexcept_setround(e, r) \
do { \
unsigned int mxcsr; \
asm ("stmxcsr %0" : "=m" (*&mxcsr)); \
(e)->__mxcsr = mxcsr; \
mxcsr = ((mxcsr | 0x1f80) & ~0x603f) | ((r) << 3); \
asm volatile ("ldmxcsr %0" : : "m" (*&mxcsr)); \
} while (0)
#undef libc_feholdexcept_setroundf
#define libc_feholdexcept_setroundf(e, r) libc_feholdexcept_setround (e, r)
// #define libc_feholdexcept_setroundl(e, r) ...
#undef libc_fetestexcept
#define libc_fetestexcept(e) \
({ unsigned int mxcsr; asm volatile ("stmxcsr %0" : "=m" (*&mxcsr)); \
mxcsr & (e) & FE_ALL_EXCEPT; })
#undef libc_fetestexceptf
#define libc_fetestexceptf(e) libc_fetestexcept (e)
// #define libc_fetestexceptl(e) fetestexcept (e)
#undef libc_fesetenv
#define libc_fesetenv(e) \
asm volatile ("ldmxcsr %0" : : "m" ((e)->__mxcsr))
#undef libc_fesetenvf
#define libc_fesetenvf(e) libc_fesetenv (e)
// #define libc_fesetenvl(e) (void) fesetenv (e)
#undef libc_feupdateenv
#define libc_feupdateenv(e) \
do { \
unsigned int mxcsr; \
asm volatile ("stmxcsr %0" : "=m" (*&mxcsr)); \
asm volatile ("ldmxcsr %0" : : "m" ((e)->__mxcsr)); \
__feraiseexcept (mxcsr & FE_ALL_EXCEPT); \
} while (0)
#undef libc_feupdateenvf
#define libc_feupdateenvf(e) libc_feupdateenv (e)
// #define libc_feupdateenvl(e) (void) feupdateenv (e)