diff options
Diffstat (limited to 'sysdeps/x86/fpu/bits/mathinline.h')
-rw-r--r-- | sysdeps/x86/fpu/bits/mathinline.h | 650 |
1 files changed, 1 insertions, 649 deletions
diff --git a/sysdeps/x86/fpu/bits/mathinline.h b/sysdeps/x86/fpu/bits/mathinline.h index 0ff1aa4cec..91ece8dfb8 100644 --- a/sysdeps/x86/fpu/bits/mathinline.h +++ b/sysdeps/x86/fpu/bits/mathinline.h @@ -1,5 +1,5 @@ /* Inline math functions for i387 and SSE. - Copyright (C) 1995-2016 Free Software Foundation, Inc. + Copyright (C) 1995-2018 Free Software Foundation, Inc. This file is part of the GNU C Library. The GNU C Library is free software; you can redistribute it and/or @@ -26,364 +26,6 @@ # define __MATH_INLINE __extern_always_inline #endif - -#if defined __USE_ISOC99 && defined __GNUC__ && __GNUC__ >= 2 -/* GCC 2.97 and up have builtins that actually can be used. */ -# if !__GNUC_PREREQ (2,97) -/* ISO C99 defines some macros to perform unordered comparisons. The - ix87 FPU supports this with special opcodes and we should use them. - These must not be inline functions since we have to be able to handle - all floating-point types. */ -# undef isgreater -# undef isgreaterequal -# undef isless -# undef islessequal -# undef islessgreater -# undef isunordered -# ifdef __i686__ -/* For the PentiumPro and more recent processors we can provide - better code. */ -# define isgreater(x, y) \ - ({ register char __result; \ - __asm__ ("fucomip %%st(1), %%st; seta %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st"); \ - __result; }) -# define isgreaterequal(x, y) \ - ({ register char __result; \ - __asm__ ("fucomip %%st(1), %%st; setae %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st"); \ - __result; }) - -# define isless(x, y) \ - ({ register char __result; \ - __asm__ ("fucomip %%st(1), %%st; seta %%al" \ - : "=a" (__result) : "u" (x), "t" (y) : "cc", "st"); \ - __result; }) - -# define islessequal(x, y) \ - ({ register char __result; \ - __asm__ ("fucomip %%st(1), %%st; setae %%al" \ - : "=a" (__result) : "u" (x), "t" (y) : "cc", "st"); \ - __result; }) - -# define islessgreater(x, y) \ - ({ register char __result; \ - __asm__ ("fucomip %%st(1), %%st; setne %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st"); \ - __result; }) - -# define isunordered(x, y) \ - ({ register char __result; \ - __asm__ ("fucomip %%st(1), %%st; setp %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st"); \ - __result; }) -# else -/* This is the dumb, portable code for i386 and above. */ -# define isgreater(x, y) \ - ({ register char __result; \ - __asm__ ("fucompp; fnstsw; testb $0x45, %%ah; setz %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st", "st(1)"); \ - __result; }) - -# define isgreaterequal(x, y) \ - ({ register char __result; \ - __asm__ ("fucompp; fnstsw; testb $0x05, %%ah; setz %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st", "st(1)"); \ - __result; }) - -# define isless(x, y) \ - ({ register char __result; \ - __asm__ ("fucompp; fnstsw; testb $0x45, %%ah; setz %%al" \ - : "=a" (__result) : "u" (x), "t" (y) : "cc", "st", "st(1)"); \ - __result; }) - -# define islessequal(x, y) \ - ({ register char __result; \ - __asm__ ("fucompp; fnstsw; testb $0x05, %%ah; setz %%al" \ - : "=a" (__result) : "u" (x), "t" (y) : "cc", "st", "st(1)"); \ - __result; }) - -# define islessgreater(x, y) \ - ({ register char __result; \ - __asm__ ("fucompp; fnstsw; testb $0x44, %%ah; setz %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st", "st(1)"); \ - __result; }) - -# define isunordered(x, y) \ - ({ register char __result; \ - __asm__ ("fucompp; fnstsw; sahf; setp %%al" \ - : "=a" (__result) : "u" (y), "t" (x) : "cc", "st", "st(1)"); \ - __result; }) -# endif /* __i686__ */ -# endif /* GCC 2.97 */ - -/* The gcc, version 2.7 or below, has problems with all this inlining - code. So disable it for this version of the compiler. */ -# if __GNUC_PREREQ (2, 8) -__BEGIN_NAMESPACE_C99 - -/* Test for negative number. Used in the signbit() macro. */ -__MATH_INLINE int -__NTH (__signbitf (float __x)) -{ -# ifdef __SSE2_MATH__ - int __m; - __asm ("pmovmskb %1, %0" : "=r" (__m) : "x" (__x)); - return (__m & 0x8) != 0; -# else - __extension__ union { float __f; int __i; } __u = { __f: __x }; - return __u.__i < 0; -# endif -} -__MATH_INLINE int -__NTH (__signbit (double __x)) -{ -# ifdef __SSE2_MATH__ - int __m; - __asm ("pmovmskb %1, %0" : "=r" (__m) : "x" (__x)); - return (__m & 0x80) != 0; -# else - __extension__ union { double __d; int __i[2]; } __u = { __d: __x }; - return __u.__i[1] < 0; -# endif -} -__MATH_INLINE int -__NTH (__signbitl (long double __x)) -{ - __extension__ union { long double __l; int __i[3]; } __u = { __l: __x }; - return (__u.__i[2] & 0x8000) != 0; -} - -__END_NAMESPACE_C99 -# endif -#endif - - -/* The gcc, version 2.7 or below, has problems with all this inlining - code. So disable it for this version of the compiler. */ -#if __GNUC_PREREQ (2, 8) -# if !__GNUC_PREREQ (3, 4) && !defined __NO_MATH_INLINES \ - && defined __OPTIMIZE__ -/* GCC 3.4 introduced builtins for all functions below, so - there's no need to define any of these inline functions. */ - -# ifdef __USE_ISOC99 -__BEGIN_NAMESPACE_C99 - -/* Round to nearest integer. */ -# ifdef __SSE_MATH__ -__MATH_INLINE long int -__NTH (lrintf (float __x)) -{ - long int __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("cvtss2si %1, %0" : "=r" (__res) : "xm" (__x)); - return __res; -} -# endif -# ifdef __SSE2_MATH__ -__MATH_INLINE long int -__NTH (lrint (double __x)) -{ - long int __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("cvtsd2si %1, %0" : "=r" (__res) : "xm" (__x)); - return __res; -} -# endif -# ifdef __x86_64__ -__extension__ -__MATH_INLINE long long int -__NTH (llrintf (float __x)) -{ - long long int __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("cvtss2si %1, %0" : "=r" (__res) : "xm" (__x)); - return __res; -} -__extension__ -__MATH_INLINE long long int -__NTH (llrint (double __x)) -{ - long long int __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("cvtsd2si %1, %0" : "=r" (__res) : "xm" (__x)); - return __res; -} -# endif - -# if defined __FINITE_MATH_ONLY__ && __FINITE_MATH_ONLY__ > 0 \ - && defined __SSE2_MATH__ -/* Determine maximum of two values. */ -__MATH_INLINE float -__NTH (fmaxf (float __x, float __y)) -{ -# ifdef __AVX__ - float __res; - __asm ("vmaxss %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); - return __res; -# else - __asm ("maxss %1, %0" : "+x" (__x) : "xm" (__y)); - return __x; -# endif -} -__MATH_INLINE double -__NTH (fmax (double __x, double __y)) -{ -# ifdef __AVX__ - float __res; - __asm ("vmaxsd %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); - return __res; -# else - __asm ("maxsd %1, %0" : "+x" (__x) : "xm" (__y)); - return __x; -# endif -} - -/* Determine minimum of two values. */ -__MATH_INLINE float -__NTH (fminf (float __x, float __y)) -{ -# ifdef __AVX__ - float __res; - __asm ("vminss %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); - return __res; -# else - __asm ("minss %1, %0" : "+x" (__x) : "xm" (__y)); - return __x; -# endif -} -__MATH_INLINE double -__NTH (fmin (double __x, double __y)) -{ -# ifdef __AVX__ - float __res; - __asm ("vminsd %2, %1, %0" : "=x" (__res) : "x" (x), "xm" (__y)); - return __res; -# else - __asm ("minsd %1, %0" : "+x" (__x) : "xm" (__y)); - return __x; -# endif -} -# endif - -__END_NAMESPACE_C99 -# endif - -# if defined __SSE4_1__ && defined __SSE2_MATH__ -# if defined __USE_XOPEN_EXTENDED || defined __USE_ISOC99 -__BEGIN_NAMESPACE_C99 - -/* Round to nearest integer. */ -__MATH_INLINE double -__NTH (rint (double __x)) -{ - double __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("roundsd $4, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} -__MATH_INLINE float -__NTH (rintf (float __x)) -{ - float __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("roundss $4, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} - -# ifdef __USE_ISOC99 -/* Round to nearest integer without raising inexact exception. */ -__MATH_INLINE double -__NTH (nearbyint (double __x)) -{ - double __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("roundsd $0xc, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} -__MATH_INLINE float -__NTH (nearbyintf (float __x)) -{ - float __res; - /* Mark as volatile since the result is dependent on the state of - the SSE control register (the rounding mode). Otherwise GCC might - remove these assembler instructions since it does not know about - the rounding mode change and cannot currently be told. */ - __asm __volatile__ ("roundss $0xc, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} -# endif - -__END_NAMESPACE_C99 -# endif - -__BEGIN_NAMESPACE_STD -/* Smallest integral value not less than X. */ -__MATH_INLINE double -__NTH (ceil (double __x)) -{ - double __res; - __asm ("roundsd $2, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} -__END_NAMESPACE_STD - -__BEGIN_NAMESPACE_C99 -__MATH_INLINE float -__NTH (ceilf (float __x)) -{ - float __res; - __asm ("roundss $2, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} -__END_NAMESPACE_C99 - -__BEGIN_NAMESPACE_STD -/* Largest integer not greater than X. */ -__MATH_INLINE double -__NTH (floor (double __x)) -{ - double __res; - __asm ("roundsd $1, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} -__END_NAMESPACE_STD - -__BEGIN_NAMESPACE_C99 -__MATH_INLINE float -__NTH (floorf (float __x)) -{ - float __res; - __asm ("roundss $1, %1, %0" : "=x" (__res) : "xm" (__x)); - return __res; -} -__END_NAMESPACE_C99 -# endif -# endif -#endif - /* Disable x87 inlines when -fpmath=sse is passed and also when we're building on x86_64. Older gcc (gcc-3.2 for example) does not define __SSE2_MATH__ for x86_64. */ @@ -532,50 +174,6 @@ __END_NAMESPACE_C99 /* __FAST_MATH__ is defined by gcc -ffast-math. */ # ifdef __FAST_MATH__ -# ifdef __USE_GNU -# define __sincos_code \ - register long double __cosr; \ - register long double __sinr; \ - register unsigned int __swtmp; \ - __asm __volatile__ \ - ("fsincos\n\t" \ - "fnstsw %w2\n\t" \ - "testl $0x400, %2\n\t" \ - "jz 1f\n\t" \ - "fldpi\n\t" \ - "fadd %%st(0)\n\t" \ - "fxch %%st(1)\n\t" \ - "2: fprem1\n\t" \ - "fnstsw %w2\n\t" \ - "testl $0x400, %2\n\t" \ - "jnz 2b\n\t" \ - "fstp %%st(1)\n\t" \ - "fsincos\n\t" \ - "1:" \ - : "=t" (__cosr), "=u" (__sinr), "=a" (__swtmp) : "0" (__x)); \ - *__sinx = __sinr; \ - *__cosx = __cosr - -__MATH_INLINE void -__NTH (__sincos (double __x, double *__sinx, double *__cosx)) -{ - __sincos_code; -} - -__MATH_INLINE void -__NTH (__sincosf (float __x, float *__sinx, float *__cosx)) -{ - __sincos_code; -} - -__MATH_INLINE void -__NTH (__sincosl (long double __x, long double *__sinx, long double *__cosx)) -{ - __sincos_code; -} -# endif - - /* Optimized inline implementation, sometimes with reduced precision and/or argument range. */ @@ -631,50 +229,9 @@ __inline_mathcodeNP_ (long double, __expl, __x, return __builtin_expl (__x)) __inline_mathcodeNP (exp, __x, __exp_code) __inline_mathcodeNP_ (long double, __expl, __x, __exp_code) # endif - - -# if !__GNUC_PREREQ (3, 5) -__inline_mathcodeNP (tan, __x, \ - register long double __value; \ - register long double __value2 __attribute__ ((__unused__)); \ - __asm __volatile__ \ - ("fptan" \ - : "=t" (__value2), "=u" (__value) : "0" (__x)); \ - return __value) -# endif # endif /* __FAST_MATH__ */ -# if __GNUC_PREREQ (3, 4) -__inline_mathcodeNP2_ (long double, __atan2l, __y, __x, - return __builtin_atan2l (__y, __x)) -# else -# define __atan2_code \ - register long double __value; \ - __asm __volatile__ \ - ("fpatan" \ - : "=t" (__value) : "0" (__x), "u" (__y) : "st(1)"); \ - return __value -# ifdef __FAST_MATH__ -__inline_mathcodeNP2 (atan2, __y, __x, __atan2_code) -# endif -__inline_mathcodeNP2_ (long double, __atan2l, __y, __x, __atan2_code) -# endif - - -# if defined __FAST_MATH__ && !__GNUC_PREREQ (3, 5) -__inline_mathcodeNP2 (fmod, __x, __y, \ - register long double __value; \ - __asm __volatile__ \ - ("1: fprem\n\t" \ - "fnstsw %%ax\n\t" \ - "sahf\n\t" \ - "jp 1b" \ - : "=t" (__value) : "0" (__x), "u" (__y) : "ax", "cc"); \ - return __value) -# endif - - # ifdef __FAST_MATH__ # if !__GNUC_PREREQ (3,3) __inline_mathopNP (sqrt, "fsqrt") @@ -697,28 +254,6 @@ __inline_mathop (fabs, "fabs") __inline_mathop_ (long double, __fabsl, "fabs") # endif -# ifdef __FAST_MATH__ -# if !__GNUC_PREREQ (3, 4) -/* The argument range of this inline version is reduced. */ -__inline_mathopNP (sin, "fsin") -/* The argument range of this inline version is reduced. */ -__inline_mathopNP (cos, "fcos") - -__inline_mathop_declNP (log, "fldln2; fxch; fyl2x", "0" (__x) : "st(1)") -# endif - -# if !__GNUC_PREREQ (3, 5) -__inline_mathop_declNP (log10, "fldlg2; fxch; fyl2x", "0" (__x) : "st(1)") - -__inline_mathcodeNP (asin, __x, return __atan2l (__x, __libc_sqrtl (1.0 - __x * __x))) -__inline_mathcodeNP (acos, __x, return __atan2l (__libc_sqrtl (1.0 - __x * __x), __x)) -# endif - -# if !__GNUC_PREREQ (3, 4) -__inline_mathop_declNP (atan, "fld1; fpatan", "0" (__x) : "st(1)") -# endif -# endif /* __FAST_MATH__ */ - __inline_mathcode_ (long double, __sgn1l, __x, \ __extension__ union { long double __xld; unsigned int __xi[3]; } __n = \ { __xld: __x }; \ @@ -743,57 +278,6 @@ __inline_mathcodeNP (tanh, __x, \ return __exm1 / (__exm1 + 2.0) * __sgn1l (-__x)) # endif -__inline_mathcodeNP (floor, __x, \ - register long double __value; \ - register int __ignore; \ - unsigned short int __cw; \ - unsigned short int __cwtmp; \ - __asm __volatile ("fnstcw %3\n\t" \ - "movzwl %3, %1\n\t" \ - "andl $0xf3ff, %1\n\t" \ - "orl $0x0400, %1\n\t" /* rounding down */ \ - "movw %w1, %2\n\t" \ - "fldcw %2\n\t" \ - "frndint\n\t" \ - "fldcw %3" \ - : "=t" (__value), "=&q" (__ignore), "=m" (__cwtmp), \ - "=m" (__cw) \ - : "0" (__x)); \ - return __value) - -__inline_mathcodeNP (ceil, __x, \ - register long double __value; \ - register int __ignore; \ - unsigned short int __cw; \ - unsigned short int __cwtmp; \ - __asm __volatile ("fnstcw %3\n\t" \ - "movzwl %3, %1\n\t" \ - "andl $0xf3ff, %1\n\t" \ - "orl $0x0800, %1\n\t" /* rounding up */ \ - "movw %w1, %2\n\t" \ - "fldcw %2\n\t" \ - "frndint\n\t" \ - "fldcw %3" \ - : "=t" (__value), "=&q" (__ignore), "=m" (__cwtmp), \ - "=m" (__cw) \ - : "0" (__x)); \ - return __value) - -# ifdef __FAST_MATH__ -# define __ldexp_code \ - register long double __value; \ - __asm __volatile__ \ - ("fscale" \ - : "=t" (__value) : "0" (__x), "u" ((long double) __y)); \ - return __value - -__MATH_INLINE double -__NTH (ldexp (double __x, int __y)) -{ - __ldexp_code; -} -# endif - /* Optimized versions for some non-standardized functions. */ # ifdef __USE_ISOC99 @@ -801,25 +285,6 @@ __NTH (ldexp (double __x, int __y)) # ifdef __FAST_MATH__ __inline_mathcodeNP (expm1, __x, __expm1_code) -/* We cannot rely on M_SQRT being defined. So we do it for ourself - here. */ -# define __M_SQRT2 1.41421356237309504880L /* sqrt(2) */ - -# if !__GNUC_PREREQ (3, 5) -__inline_mathcodeNP (log1p, __x, \ - register long double __value; \ - if (__fabsl (__x) >= 1.0 - 0.5 * __M_SQRT2) \ - __value = logl (1.0 + __x); \ - else \ - __asm __volatile__ \ - ("fldln2\n\t" \ - "fxch\n\t" \ - "fyl2xp1" \ - : "=t" (__value) : "0" (__x) : "st(1)"); \ - return __value) -# endif - - /* The argument range of the inline version of asinhl is slightly reduced. */ __inline_mathcodeNP (asinh, __x, \ register long double __y = __fabsl (__x); \ @@ -837,126 +302,14 @@ __inline_mathcodeNP (atanh, __x, \ __inline_mathcodeNP2 (hypot, __x, __y, return __libc_sqrtl (__x * __x + __y * __y)) -# if !__GNUC_PREREQ (3, 5) -__inline_mathcodeNP(logb, __x, \ - register long double __value; \ - register long double __junk; \ - __asm __volatile__ \ - ("fxtract\n\t" \ - : "=t" (__junk), "=u" (__value) : "0" (__x)); \ - return __value) -# endif - # endif # endif -# ifdef __USE_ISOC99 -# ifdef __FAST_MATH__ - -# if !__GNUC_PREREQ (3, 5) -__inline_mathop_declNP (log2, "fld1; fxch; fyl2x", "0" (__x) : "st(1)") -# endif - -__MATH_INLINE float -__NTH (ldexpf (float __x, int __y)) -{ - __ldexp_code; -} - -__MATH_INLINE long double -__NTH (ldexpl (long double __x, int __y)) -{ - __ldexp_code; -} - -__inline_mathopNP (rint, "frndint") -# endif /* __FAST_MATH__ */ - -# define __lrint_code \ - long int __lrintres; \ - __asm__ __volatile__ \ - ("fistpl %0" \ - : "=m" (__lrintres) : "t" (__x) : "st"); \ - return __lrintres -__MATH_INLINE long int -__NTH (lrintf (float __x)) -{ - __lrint_code; -} -__MATH_INLINE long int -__NTH (lrint (double __x)) -{ - __lrint_code; -} -__MATH_INLINE long int -__NTH (lrintl (long double __x)) -{ - __lrint_code; -} -# undef __lrint_code - -# define __llrint_code \ - long long int __llrintres; \ - __asm__ __volatile__ \ - ("fistpll %0" \ - : "=m" (__llrintres) : "t" (__x) : "st"); \ - return __llrintres -__extension__ -__MATH_INLINE long long int -__NTH (llrintf (float __x)) -{ - __llrint_code; -} -__extension__ -__MATH_INLINE long long int -__NTH (llrint (double __x)) -{ - __llrint_code; -} -__extension__ -__MATH_INLINE long long int -__NTH (llrintl (long double __x)) -{ - __llrint_code; -} -# undef __llrint_code - -# endif - - -# ifdef __USE_MISC - -# if defined __FAST_MATH__ && !__GNUC_PREREQ (3, 5) -__inline_mathcodeNP2 (drem, __x, __y, \ - register double __value; \ - register int __clobbered; \ - __asm __volatile__ \ - ("1: fprem1\n\t" \ - "fstsw %%ax\n\t" \ - "sahf\n\t" \ - "jp 1b" \ - : "=t" (__value), "=&a" (__clobbered) : "0" (__x), "u" (__y) : "cc"); \ - return __value) -# endif - - -/* This function is used in the `isfinite' macro. */ -__MATH_INLINE int -__NTH (__finite (double __x)) -{ - return (__extension__ - (((((union { double __d; int __i[2]; }) {__d: __x}).__i[1] - | 0x800fffffu) + 1) >> 31)); -} - -# endif /* __USE_MISC */ /* Undefine some of the large macros which are not used anymore. */ -# undef __atan2_code # ifdef __FAST_MATH__ # undef __expm1_code # undef __exp_code -# undef __sincos_code # endif /* __FAST_MATH__ */ # endif /* __NO_MATH_INLINES */ @@ -964,7 +317,6 @@ __NTH (__finite (double __x)) /* This code is used internally in the GNU libc. */ # ifdef __LIBC_INTERNAL_MATH_INLINES -__inline_mathop (__ieee754_sqrt, "fsqrt") __inline_mathcode2_ (long double, __ieee754_atan2l, __y, __x, register long double __value; __asm __volatile__ ("fpatan\n\t" |