[ruby-dev:12385] Re: NaN, Infinity (Bug?)

わたなべです。

ちと調べてみました。

GOTO Kentaro <[e-mail]> writes:

:In message "[ruby-talk:12132] Re: NaN, Infinity (Bug?)"
:    on 01/03/06, Dave Thomas <[e-mail]> writes:
:>I just ran this again on my box here (ruby 1.6.2 (2001-02-02) [i686-linux]):
:>
:> a = -1.0  		# => -1.0
:> a.nan?    		# => false
:> a = Math.log(a)	# => NaN
:> a.nan?                 # => true
:>
:>So, there seems to be a difference between a Linux box and a
:>Windows/cygwin box. Anyone any ideas?
:
:I ran:
:
:  #include <math.h>
:  main(){printf("%f\n", log(-1.0));}
:
:on my Cygwin box with `gcc version 2.95.2 19991024 (release)' and got:
:
:  -Inf
:
:I think this is wrong.  
:Should the configure detect this incompatibility?

Cygwinは内部的にnewlibという名前のlibcを使ってますが、
log()は実はwrapperで、下請けの__ieee754_log()の結果を
変更してNaNを-Infにしてます。
コメントにもそう書いてあるので意図的なものみたい。

newlib/libm/math/w_log.c:
>RETURNS
>Normally, returns the calculated value.  When <[x]> is zero, the
>returned value is <<-HUGE_VAL>> and <<errno>> is set to <<ERANGE>>.
>When <[x]> is negative, the returned value is <<-HUGE_VAL>> and
><<errno>> is set to <<EDOM>>.  You can control the error behavior via
><<matherr>>.

どうしようかなあ？とりあえずこれでcommitしてみます。

#if defined __CYGWIN__
#define log(x) ((x) < 0.0 ? ((x) - (x)) / 0.0 : log(x))
#endif

-- 
わたなべひろふみ