Opinions on packaging ATLAS (for the x86 architecture)

[Ralf Ertzinger]

Hi.

On Wed, 7 Oct 2009 19:10:28 +0100, Richard W.M. Jones wrote

> If it's possible to write programs and shared library loaders so that
> redetection can be performed mid-execution, then prefer that method
> over one which only detects hardware when the program starts up.

I have no qualms whatsoever with hardware changing between boots.
Network cards, hard disks, CPU features, you name it. But having
CPU features change from one instruction to the other (which the 
above would suggest, correct me if I'm wrong)... how do you suggest
this would work? Testing for the feature before using it (every time?
That should nullify any speedup gained by using the features in
the first place) does not work, because the machine may move between
the test and the instruction (maybe there's a way around this).

Catching SIGILL? What about the data that was in the registers that
are suddenly no longer there?

And it's not only userspace, the kernel can use SSE, too (at least it did,
once, for RAID checksumming at least).