X-Git-Url: http://erislabs.net/gitweb/?a=blobdiff_plain;f=m4%2Ffrexpf.m4;h=d8f58cd0f5c46d5c9ac1437990bd948dd4f5c6ba;hb=39a489fa27ab3873e0fc0f65844413f46fcb2117;hp=7928011c38c469a3417d30415ff06608ae469ebb;hpb=df8ad3525d27df0b821539df90af26c7fa6d3188;p=gnulib.git diff --git a/m4/frexpf.m4 b/m4/frexpf.m4 index 7928011c3..d8f58cd0f 100644 --- a/m4/frexpf.m4 +++ b/m4/frexpf.m4 @@ -1,4 +1,4 @@ -# frexpf.m4 serial 1 +# frexpf.m4 serial 3 dnl Copyright (C) 2011 Free Software Foundation, Inc. dnl This file is free software; the Free Software Foundation dnl gives unlimited permission to copy and/or distribute it, @@ -9,17 +9,85 @@ AC_DEFUN([gl_FUNC_FREXPF], AC_REQUIRE([gl_MATH_H_DEFAULTS]) AC_REQUIRE([gl_FUNC_FREXP]) - dnl Test whether frexpf() exists. Assume that frexpf(), if it exists, is - dnl defined in the same library as frexp(). - save_LIBS="$LIBS" - LIBS="$LIBS $FREXP_LIBM" - AC_CHECK_FUNCS([frexpf]) - LIBS="$save_LIBS" - if test $ac_cv_func_frexpf = yes; then - FREXPF_LIBM="$FREXP_LIBM" + dnl Test whether frexpf() exists. We cannot assume that frexpf(), if it + dnl exists, is defined in the same library as frexp(). This is not the case + dnl on NetBSD, OpenBSD. + gl_MATHFUNC([frexpf], [float], [(float, int *)]) + if test $gl_cv_func_frexpf_no_libm = yes \ + || test $gl_cv_func_frexpf_in_libm = yes; then + save_LIBS="$LIBS" + LIBS="$LIBS $FREXPF_LIBM" + gl_FUNC_FREXPF_WORKS + LIBS="$save_LIBS" + case "$gl_cv_func_frexpf_works" in + *yes) ;; + *) REPLACE_FREXPF=1 ;; + esac else HAVE_FREXPF=0 + fi + if test $HAVE_FREXPF = 0 || test $REPLACE_FREXPF = 1; then FREXPF_LIBM="$FREXP_LIBM" fi AC_SUBST([FREXPF_LIBM]) ]) + +dnl Test whether frexpf() works also on infinite numbers (this fails e.g. on +dnl IRIX 6.5 and mingw) and on negative zero (this fails e.g. on mingw). +AC_DEFUN([gl_FUNC_FREXPF_WORKS], +[ + AC_REQUIRE([AC_PROG_CC]) + AC_REQUIRE([AC_CANONICAL_HOST]) dnl for cross-compiles + AC_CACHE_CHECK([whether frexpf works], [gl_cv_func_frexpf_works], + [ + AC_RUN_IFELSE( + [AC_LANG_SOURCE([[ +#include +#include +#include +/* HP cc on HP-UX 10.20 has a bug with the constant expression -0.0. + ICC 10.0 has a bug when optimizing the expression -zero. + The expression -FLT_MIN * FLT_MIN does not work when cross-compiling + to PowerPC on MacOS X 10.5. */ +#if defined __hpux || defined __sgi || defined __ICC +static float +compute_minus_zero (void) +{ + return -FLT_MIN * FLT_MIN; +} +# define minus_zero compute_minus_zero () +#else +float minus_zero = -0.0f; +#endif +int main() +{ + int result = 0; + volatile float x; + float zero = 0.0f; + /* Test on infinite numbers. */ + x = 1.0f / zero; + { + int exp; + float y = frexpf (x, &exp); + if (y != x) + result |= 1; + } + /* Test on negative zero. */ + x = minus_zero; + { + int exp; + float y = frexpf (x, &exp); + if (memcmp (&y, &x, sizeof x)) + result |= 2; + } + return result; +}]])], + [gl_cv_func_frexpf_works=yes], + [gl_cv_func_frexpf_works=no], + [case "$host_os" in + irix* | mingw*) gl_cv_func_frexpf_works="guessing no";; + *) gl_cv_func_frexpf_works="guessing yes";; + esac + ]) + ]) +])