Re: [PATCH v2 6/7] alpha: Fix fesetexceptflag (BZ 30998)

[Bruno Haible <bruno@clisp.org>]

Carlos O'Donell wrote:
> I would expect (taken from hppa code I wrote for that port):
> 
> /* Clear all status bits we care about.  */
> tmp = tmp & ~(excepts & SWCR_STATUS_MASK);
> /* Install the new ones.  */
> tmp |= *flagp & excepts & SWCR_STATUS_MASK;

This code is correct. And, as I wrote in
https://sourceware.org/bugzilla/show_bug.cgi?id=30998#c2
the code with the double-xor is faster, because it uses one less
instruction.

Basically the code you proposed masks the bits to clear
and the bits to set separately. Whereas the code with the double-xor
masks the bits to *change* - a single AND instead of two.

> >    /* Set all the bits that were called for.  */
> > -  tmp = (tmp & ~SWCR_STATUS_MASK) | (*flagp & excepts & SWCR_STATUS_MASK);
> > +  tmp ^= (tmp ^ *flagp) & excepts & SWCR_STATUS_MASK;
> 
> Does this actually work?

Yes: It computes the bits to *change*, then applies the mask, then applies
the changes.

For bits in the mask, it does
        tmp.bit[i] ^= tmp.bit[i] ^ (*flagp).bit[i];
which simplifies to
        tmp.bit[i] = (*flagp).bit[i];

For bits outside the mask, it does
        tmp.bit[i] ^= 0;
i.e. it leaves the bit unchanged.

I'm using this same idiom also in gnulib
https://git.savannah.gnu.org/gitweb/?p=gnulib.git;a=blob;f=lib/fenv-except-state-set.c;h=2035f61e2714126e432c0c1f29450f18f58d81f1;hb=HEAD#l182
and I have verified that it works.

> Assume excepts is FE_INVALID, and *flagp bit 17 is 0.
> Assume currently bit 17 is 1.
> 
> tmp ^ *flagp       => bit 17 is still 1, even though bit 17 in flagp is 0.
> & excepts          => bit 17 is still 1.
> & SWCR_STATUS_MASK => bit 17 is still 1.
> ^=                 => bit 17 is still 1.

In the last step: 1 xor 1 is 0.

Bruno