diff options
Diffstat (limited to 'm4/isfinite.m4')
| -rw-r--r-- | m4/isfinite.m4 | 165 |
1 files changed, 165 insertions, 0 deletions
diff --git a/m4/isfinite.m4 b/m4/isfinite.m4 new file mode 100644 index 000000000..b54b403b8 --- /dev/null +++ b/m4/isfinite.m4 @@ -0,0 +1,165 @@ +# isfinite.m4 serial 13 +dnl Copyright (C) 2007-2013 Free Software Foundation, Inc. +dnl This file is free software; the Free Software Foundation +dnl gives unlimited permission to copy and/or distribute it, +dnl with or without modifications, as long as this notice is preserved. + +AC_DEFUN([gl_ISFINITE], +[ + AC_REQUIRE([gl_MATH_H_DEFAULTS]) + dnl Persuade glibc <math.h> to declare isfinite. + AC_REQUIRE([gl_USE_SYSTEM_EXTENSIONS]) + AC_CHECK_DECLS([isfinite], , , [[#include <math.h>]]) + if test "$ac_cv_have_decl_isfinite" = yes; then + gl_CHECK_MATH_LIB([ISFINITE_LIBM], + [x = isfinite (x) + isfinite ((float) x);]) + if test "$ISFINITE_LIBM" != missing; then + dnl Test whether isfinite() on 'long double' works. + gl_ISFINITEL_WORKS + case "$gl_cv_func_isfinitel_works" in + *yes) ;; + *) ISFINITE_LIBM=missing;; + esac + dnl Also, isfinite() on 'double' does not work on Linux/ia64 (because of + dnl signalling NaNs). But this does not have to be tested, since + dnl isfinite(long double) also does not work in this situation. + fi + fi + if test "$ac_cv_have_decl_isfinite" != yes || + test "$ISFINITE_LIBM" = missing; then + REPLACE_ISFINITE=1 + dnl No libraries are needed to link lib/isfinite.c. + ISFINITE_LIBM= + fi + AC_SUBST([ISFINITE_LIBM]) +]) + +dnl Test whether isfinite() on 'long double' recognizes all numbers which are +dnl neither finite nor infinite. This test fails e.g. on i686, x86_64, ia64, +dnl because of +dnl - pseudo-denormals on x86_64, +dnl - pseudo-zeroes, unnormalized numbers, and pseudo-denormals on i686, +dnl - pseudo-NaN, pseudo-Infinity, pseudo-zeroes, unnormalized numbers, and +dnl pseudo-denormals on ia64. +AC_DEFUN([gl_ISFINITEL_WORKS], +[ + AC_REQUIRE([AC_PROG_CC]) + AC_REQUIRE([gl_BIGENDIAN]) + AC_REQUIRE([gl_LONG_DOUBLE_VS_DOUBLE]) + AC_REQUIRE([AC_CANONICAL_HOST]) dnl for cross-compiles + AC_CACHE_CHECK([whether isfinite(long double) works], [gl_cv_func_isfinitel_works], + [ + AC_RUN_IFELSE([AC_LANG_SOURCE([[ +#include <float.h> +#include <limits.h> +#include <math.h> +#define NWORDS \ + ((sizeof (long double) + sizeof (unsigned int) - 1) / sizeof (unsigned int)) +typedef union { unsigned int word[NWORDS]; long double value; } + memory_long_double; +/* On Irix 6.5, gcc 3.4.3 can't compute compile-time NaN, and needs the + runtime type conversion. */ +#ifdef __sgi +static long double NaNl () +{ + double zero = 0.0; + return zero / zero; +} +#else +# define NaNl() (0.0L / 0.0L) +#endif +int main () +{ + int result = 0; + + { + memory_long_double m; + unsigned int i; + + /* The isfinite macro should be immune against changes in the sign bit and + in the mantissa bits. The xor operation twiddles a bit that can only be + a sign bit or a mantissa bit (since the exponent never extends to + bit 31). */ + m.value = NaNl (); + m.word[NWORDS / 2] ^= (unsigned int) 1 << (sizeof (unsigned int) * CHAR_BIT - 1); + for (i = 0; i < NWORDS; i++) + m.word[i] |= 1; + if (isfinite (m.value)) + result |= 1; + } + +#if ((defined __ia64 && LDBL_MANT_DIG == 64) || (defined __x86_64__ || defined __amd64__) || (defined __i386 || defined __i386__ || defined _I386 || defined _M_IX86 || defined _X86_)) && !HAVE_SAME_LONG_DOUBLE_AS_DOUBLE +/* Representation of an 80-bit 'long double' as an initializer for a sequence + of 'unsigned int' words. */ +# ifdef WORDS_BIGENDIAN +# define LDBL80_WORDS(exponent,manthi,mantlo) \ + { ((unsigned int) (exponent) << 16) | ((unsigned int) (manthi) >> 16), \ + ((unsigned int) (manthi) << 16) | (unsigned int) (mantlo) >> 16), \ + (unsigned int) (mantlo) << 16 \ + } +# else +# define LDBL80_WORDS(exponent,manthi,mantlo) \ + { mantlo, manthi, exponent } +# endif + { /* Quiet NaN. */ + static memory_long_double x = + { LDBL80_WORDS (0xFFFF, 0xC3333333, 0x00000000) }; + if (isfinite (x.value)) + result |= 2; + } + { + /* Signalling NaN. */ + static memory_long_double x = + { LDBL80_WORDS (0xFFFF, 0x83333333, 0x00000000) }; + if (isfinite (x.value)) + result |= 2; + } + /* The isfinite macro should recognize Pseudo-NaNs, Pseudo-Infinities, + Pseudo-Zeroes, Unnormalized Numbers, and Pseudo-Denormals, as defined in + Intel IA-64 Architecture Software Developer's Manual, Volume 1: + Application Architecture. + Table 5-2 "Floating-Point Register Encodings" + Figure 5-6 "Memory to Floating-Point Register Data Translation" + */ + { /* Pseudo-NaN. */ + static memory_long_double x = + { LDBL80_WORDS (0xFFFF, 0x40000001, 0x00000000) }; + if (isfinite (x.value)) + result |= 4; + } + { /* Pseudo-Infinity. */ + static memory_long_double x = + { LDBL80_WORDS (0xFFFF, 0x00000000, 0x00000000) }; + if (isfinite (x.value)) + result |= 8; + } + { /* Pseudo-Zero. */ + static memory_long_double x = + { LDBL80_WORDS (0x4004, 0x00000000, 0x00000000) }; + if (isfinite (x.value)) + result |= 16; + } + { /* Unnormalized number. */ + static memory_long_double x = + { LDBL80_WORDS (0x4000, 0x63333333, 0x00000000) }; + if (isfinite (x.value)) + result |= 32; + } + { /* Pseudo-Denormal. */ + static memory_long_double x = + { LDBL80_WORDS (0x0000, 0x83333333, 0x00000000) }; + if (isfinite (x.value)) + result |= 64; + } +#endif + + return result; +}]])], [gl_cv_func_isfinitel_works=yes], [gl_cv_func_isfinitel_works=no], + [case "$host_cpu" in + # Guess no on ia64, x86_64, i386. + ia64 | x86_64 | i*86) gl_cv_func_isfinitel_works="guessing no";; + *) gl_cv_func_isfinitel_works="guessing yes";; + esac + ]) + ]) +]) |