[Fwd: Re: [sldev] SSE2 / Ouch]

[Dzonatas]

Actually, I'm not fond of the C++ auto-initialized global constructors. 
There is no way to tell in source code when they are executed and in 
what order.  (See: VWR-225 and 
https://wiki.secondlife.com/wiki/Bug_triage/2007-06-11 where such 
movement to fix it went stale)

They are good for debug code, but that is about it. All intentional 
globals should be constructed explicitly within main() or likewise; 
otherwise, it is a chaotic startup. Even so, I still doubt there 
actually is stray SSE code unless someone enabled "whole program 
optimization", which was off.

Most of this is beyond what the original contract wanted me to do. We 
did research it to find out what other changes need to be made. The 
heuristic step added in made it so we could avoid a lot of other changes.

SO.....

What's left is that there is the possibility that LLV4Matrix4 globals 
being initialized with SSE code on non-SSE CPUs. It looks like VWR-225 
would have helped avoid this, but I was being told that I wasted their 
time. (...hunger strike continues)


Nicholaz Beresford wrote:
>
> Dzon says:
> > From what I read on the jira issue and in these mails, it sounds like
> > the gSysCPU.hasSSE() doesn't return the correct result for AMD
> > processors.
>
> From the way I read it (the JIRA and what Dirk says), there's code
> executed through constructors of global objects, long before the
> crt startup code even thinks about calling main()/WinMain().
>
> Either that, or you rely on gSysCPU being istantiated globally
> before those other global or static objects ...
>
>
>
> Nick
> _______________________________________________
> Click here to unsubscribe or manage your list subscription:
> /index.html
>
>

-- 
Power to Change the Void