Re: [PATCHES] equal() perf tweak

Tom Lane wrote:

>Andrew Dunstan <andrew(at)dunslane(dot)net> writes:
>
>
>>Just a thought - if we are messing with the List definition should we at
>>the same time address the strict aliasing issues arising from Node's
>>multiple personalities (I think it is the main offender).
>>
>>
>
>
>
>>Or is the intention never to do this, or not any time soon?
>>
>>
>
>I have no intention of messing with the Node concept; it's built into
>the backend far too firmly to consider any significant change.
>
>I don't think we understand exactly what we'd have to avoid in order to
>enable strict aliasing, but if it requires getting rid of Node then it
>ain't happening. (I doubt that it does, anyway. I think the issues are
>probably quite localized. The main problem I see is that we don't have
>any trustworthy check to find out everyplace that strict aliasing could
>cause problems.)
>
>
>

*nod* (I made the last point previously :-) )

I don't claim that my understanding is anything like complete.

However, in the case of gcc it might be easier to fix than you think,
without a lot of declaration rearranging, using gcc's new "may_alias"
type attribute, (see
http://gcc.gnu.org/onlinedocs/gcc-3.3.2/gcc/Type-Attributes.html ) which
could probably be added in a very few places (e.g. Node's typedef) to
fix things.

Of course, that doesn't help when vendor foo's compiler starts doing
type-based alias analysis.

Still, to quote "Airplane", that's not important right now.

cheers

andrew