How fast is local memory? the doc doesn't say much

What I actually experienced, is:

When compiler makes use of the local memory, for example, register spilling, the kernel stops working.

Well, now getting this kind of reply, I’m more and more disappointed with your compiler guys. If you continue hiding important specs like this, your arrogance will be your bane.

You’re already way behind Intel, without OUR help, you won’t catch up, ever.

We do appreciate any help you provide. How about providing a repro case where register spilling stops the kernel from working?

Paulius

Thanks for your appreciating. I’ll provide a repro case for the next release.

The reasoning is:
In current state of CUDA 1.0, reverse engineering for specs and writing ptx from scratch is more productive than reporting bugs.

Well, I apologize for being impatient.
But since my kernel is due today, and it’s unlikely you’ll release 1.1 before that, reporting bugs won’t help me anyway.