Optimize assume(...) away? (agaiin)

Optimize assume(...) away? (agaiin)

[Dmitry Antipov]

1) http://lists.gnu.org/archive/html/emacs-devel/2013-10/msg00238.html
2) http://debbugs.gnu.org/cgi/bugreport.cgi?bug=15565

3) Consider non-trivial eassert like:

eassert ((BUFFERP (it->object)
           && IT_CHARPOS (*it) == it->bidi_it.charpos
           && IT_BYTEPOS (*it) == it->bidi_it.bytepos)
          || (STRINGP (it->object)
              && IT_STRING_CHARPOS (*it) == it->bidi_it.charpos
              && IT_STRING_BYTEPOS (*it) == it->bidi_it.bytepos)
          || (CONSP (it->object) && it->method == GET_FROM_STRETCH))

   a) It takes some time to check whether there are side effects here;
   b) If someone change 'struct it' so IT_STRING_CHARPOS (*it)
      becomes a non-trivial function with side effects, it would be
      too annoying (and error-prone) to review all easserts.

Due to 1), 2) and 3) I strongly suggests to redesign eassert
to avoid any assumptions about its argument expression.

Dmitry

Re: Optimize assume(...) away? (agaiin)

[Eli Zaretskii]

> Date: Wed, 09 Oct 2013 20:43:17 +0400
> From: Dmitry Antipov <dmantipov@yandex.ru>
> CC: Stefan Monnier <monnier@IRO.UMontreal.CA>, 
>  Eli Zaretskii <eliz@gnu.org>,
>  Paul Eggert <eggert@cs.ucla.edu>
> 
> 1) http://lists.gnu.org/archive/html/emacs-devel/2013-10/msg00238.html
> 2) http://debbugs.gnu.org/cgi/bugreport.cgi?bug=15565
> 
> 3) Consider non-trivial eassert like:
> 
> eassert ((BUFFERP (it->object)
>            && IT_CHARPOS (*it) == it->bidi_it.charpos
>            && IT_BYTEPOS (*it) == it->bidi_it.bytepos)
>           || (STRINGP (it->object)
>               && IT_STRING_CHARPOS (*it) == it->bidi_it.charpos
>               && IT_STRING_BYTEPOS (*it) == it->bidi_it.bytepos)
>           || (CONSP (it->object) && it->method == GET_FROM_STRETCH))
> 
>    a) It takes some time to check whether there are side effects here;
>    b) If someone change 'struct it' so IT_STRING_CHARPOS (*it)
>       becomes a non-trivial function with side effects, it would be
>       too annoying (and error-prone) to review all easserts.
> 
> Due to 1), 2) and 3) I strongly suggests to redesign eassert
> to avoid any assumptions about its argument expression.

How about simply removing 'assume' from eassert?  

'assume' is primarily an optimization device, and thus important in
production builds of Emacs where optimizations are enabled.  By
contrast, eassert will be defined to nothing in such builds.  So it
sounds like we are at best optimizing a rare, or even non-existing use
case, a.k.a. "premature optimization".

Re: Optimize assume(...) away? (agaiin)

[Paul Eggert]

On 10/09/13 10:21, Eli Zaretskii wrote:
> How about simply removing 'assume' from eassert?  

Yes, that sounds better.  I installed a fix along these lines,
in trunk bzr 114593.  The 'assume' helps performance
in some cases, but not others.  We could have a different
flavor of 'eassert' for when we know 'assume' helps performance
(this was briefly in the trunk as 'eassert_and_assume')
but it's not clear that it's worth the maintenance confusion
to have two flavors of 'assert'.