[PATCH 1/1] Keep the denormal-operand exception masked; modify FE_ALL_EXCEPT accordingly.

[Houder]

On 2018-08-15 17:26, Corinna Vinschen wrote:
> On Aug 15 17:01, Houder wrote:
>> On 2018-08-15 16:54, Corinna Vinschen wrote:
>> > On Aug 15 12:59, J.H. van de Water wrote:
>> > > By excluding the denormal-operand exception from FE_ALL_EXCEPT, it
>> > > will not
>> > > be possible anymore to UNmask this exception by means of the API
>> > > defined by
>> > > /usr/include/fenv.h
>> > >
>> > > Note: terminology has changed since IEEE Std 854-1987; denormalized
>> > > numbers
>> > > are called subnormal numbers nowadays.
>> > >
>> > > This modification has basically been motivated by the fact that it
>> > > is also
>> > > not possible on Linux to manipulate the denormal-operand exception
>> > > by means
>> > > of the interface as defined by /usr/include/fenv.h. This has been
>> > > the state
>> > > of affairs on Linux since 2001 (Andreas Jaeger).
>> > >
>> > > The exceptions required by the standard (IEEE Std 754), in case they
>> > > can be
>> > > supported by the implementation, are:
>> > > FE_INEXACT, FE_UNDERFLOW, FE_OVERFLOW, FE_DIVBYZERO and FE_INVALID.
>> > >
>> > > Although it is allowed to define additional exceptions, there is no
>> > > reason
>> > > to support the "denormal-operand exception" in this case (fenv.h),
>> > > because
>> > > the subnormal numbers can be handled almost as fast the normalized
>> > > numbers
>> > > by the hardware of the x86/x86_64 architecture. Said differently, a
>> > > reason
>> > > to trap on the input of subnormal numbers does not exist. At least
>> > > that is
>> > > what William Kahan and others at Intel asserted around 2000.
>> > > (that is William Kahan of the K-C-S draft, the precursor to the
>> > > standard)
>> > >
>> > > This commit modifies winsup/cygwin/include/fenv.h as follows:
>> > >  - redefines FE_ALL_EXCEPT from 0x3f to 0x3d
>> > >  - removes the definition for FE_DENORMAL
>> > >  - introduces __FE_DENORM (0x2) (enum in Linux also uses __FE_DENORM)
>> > >  - introduces FE_ALL_EXCEPT_X86 (0x3f), i.e. ALL x86/x86_64 FP
>> > > exceptions
>> >
>> > Shouldn't FE_ALL_EXCEPT_X86 be defined locally in fenv.cc only?
>> > I don't see that Linux exports that definition.
>> 
>> Ah, Sorry. Do I have to resubmit my patch? Or is it easy enough for 
>> you to
>> make this modification?
> 
> It's easy enough but I'm still mulling over __FE_DENORM.  The glibc
> fenv.h header defines it, so I guess we should stick to it.  In that
> case it might make sense to revert the original comment and just move
> __FE_ALL_EXCEPT_X86.

... uhm, my intention was to remove FE_DENORMAL from fenv.h, because it
is no longer part of the interface.

I should have defined "a mask" in fenv.cc, that would have enabled me to
initialize the MXCSR register (i.e. mask ALL exceptions).

I defined __FE_ALL_EXCEPT_X86 (and should have defined it as 0x3f).

Basically, there was no reason to define __FE_DENORM (or __FE_DENORMAL),
now that I think it over.

(I needed __FE_DENORM(AL) for testing)

Henri