[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: Fast thread-local storage for OpenGL drivers



> Your source code says:
> 
>  * Should be able to be reduced to:
>  *      movl %gs:__foo_dispatch ntpoff, %eax
>  *      jmpl *(%eax)
> 
> but that's not really a different model, just another way
> of accessing LE model variables.  This is simply not
> implemented in the current gcc i386 backend (yet?).

That is exactly the situation.

> If this actually makes any noticable difference in
> performance, I guess you should ask the i386 gcc folks
> to implement it.  If you are generating the code yourself,
> that doesn't matter anyway.

People seem to think it is a sufficient difference to be worth doing.
Jakub Jelinek and Richard Henderson are the people who have worked on the
GCC code in question.  IIRC, someone took a brief crack at making the
direct-from-%gs access work; it turned out to be more than an afternoon's
quick hack, and has not been a scheduled top priority for busy people with
other deadlines.





[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]