http://gcc.gnu.org/bugzilla/show_bug.cgi?id=33135
Oleg Endo <olegendo at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |olegendo at gcc dot gnu.org
--- Comment #1 from Oleg Endo <olegendo at gcc dot gnu.org> 2012-07-15 20:08:49
UTC ---
(In reply to comment #0)
> The documentation of -ffinite-math-only states
> "This option should never be turned on by any '-O' option"
>
> but for sh's -mieee it is said "the default is set to -ffinite-math-only"
> (-mieee is unset by default)
>
> as a result the -ffinite-math-only is turned on by default for the sh for all
> optimisation levels.
>
> on chapter 2 it was reenforced that: ... "*By default*, it (gcc) will act as
> the compiler for a hosted implementation" ... and a conforming hosted
> implementation
> supports the whole standard.
>
> So either the documentation is false or infinite float values should be
> supported (by default).
> The ability to perform optimum floating point code is preserved by explicitly
> setting it.
The SH -mieee option documentation is wrong.
-ffinite-math-only is not enabled by default (maybe it used to be some time
ago?). Thus, -mieee is enabled by default (and -mno-ieee has no effect).
Moreover, the documentation part
"... getting IEEE-conforming results for comparisons of NANs / infinities
incurs extra overhead in every floating-point comparison..."
is confusing, because this is not what the compiler actually does. If I
observe correctly, the only cases where two fcmp instructions are generated are
when doing '>=' or '<=' comparisons.
I would suggest to make the -mieee option control only the FPU comparisons and
not make it change the flag_finite_math_only. Also support for -mno-ieee
should be added.
