--- a/modules/fdlibm/src/math_private.h +++ b/modules/fdlibm/src/math_private.h @@ -30,8 +30,13 @@ * Adapted from https://github.com/freebsd/freebsd-src/search?q=__double_t */ +#if defined __FLT_EVAL_METHOD__ && (__FLT_EVAL_METHOD__ == 2) +typedef long double __double_t; +#else typedef double __double_t; +#endif typedef __double_t double_t; +typedef float __float_t; /* * The original fdlibm code used statements like: @@ -630,6 +634,53 @@ return ((double)(x + 0x1.8p52) - 0x1.8p52); } +static inline float +rnintf(__float_t x) +{ + /* + * As for rnint(), except we could just call that to handle the + * extra precision case, usually without losing efficiency. + */ + return ((float)(x + 0x1.8p23F) - 0x1.8p23F); +} + +#ifdef LDBL_MANT_DIG +/* + * The complications for extra precision are smaller for rnintl() since it + * can safely assume that the rounding precision has been increased from + * its default to FP_PE on x86. We don't exploit that here to get small + * optimizations from limiting the rangle to double. We just need it for + * the magic number to work with long doubles. ld128 callers should use + * rnint() instead of this if possible. ld80 callers should prefer + * rnintl() since for amd64 this avoids swapping the register set, while + * for i386 it makes no difference (assuming FP_PE), and for other arches + * it makes little difference. + */ + +static inline long double +rnintl(long double x) +{ + /* The WRAPPED__CONCAT() macro below is required for non-FreeBSD targets + which don't have a multi-level CONCAT macro implementation. On those + targets the hexadecimal floating-point values being created don't expand + properly resulting in code that cannot be compiled. + + The extra level provided by this macro should not affect FreeBSD, should + this code be used there. + + See the following for more details: + + https://gcc.gnu.org/onlinedocs/gcc-3.0.1/cpp_3.html#SEC32 + https://sources.debian.org/src/glibc/2.32-3/misc/sys/cdefs.h/ + https://github.com/freebsd/freebsd-src/blob/main/sys/sys/cdefs.h + */ + #define WRAPPED__CONCAT(x,y) __CONCAT(x,y) + + return (x + WRAPPED__CONCAT(0x1.8p, LDBL_MANT_DIG) / 2 - + WRAPPED__CONCAT(0x1.8p, LDBL_MANT_DIG) / 2); +} +#endif /* LDBL_MANT_DIG */ + /* * irint() and i64rint() give the same result as casting to their integer * return type provided their arg is a floating point integer. They can @@ -646,6 +697,39 @@ #define irint(x) ((int)(x)) #endif +#define i64rint(x) ((int64_t)(x)) /* only needed for ld128 so not opt. */ + +#if defined(__i386__) && defined(__GNUCLIKE_ASM) +static __inline int +irintf(float x) +{ + int n; + + __asm("fistl %0" : "=m" (n) : "t" (x)); + return (n); +} + +static __inline int +irintd(double x) +{ + int n; + + __asm("fistl %0" : "=m" (n) : "t" (x)); + return (n); +} +#endif + +#if (defined(__amd64__) || defined(__i386__)) && defined(__GNUCLIKE_ASM) +static __inline int +irintl(long double x) +{ + int n; + + __asm("fistl %0" : "=m" (n) : "t" (x)); + return (n); +} +#endif + #ifdef DEBUG #if defined(__amd64__) || defined(__i386__) #define breakpoint() asm("int $3")