Experiment with moving the task pointer to thread local storage

[pcwalton]

We currently thread the task pointer through all Rust functions, which increases register pressure and makes calls slightly more expensive. It may well be cheaper to put it in thread-local storage. We should experiment with this.

[BrendanEich]

SpiderMonkey, the new IonMonkey repo in particular but I hope it can be landed early, has an effort to move the JSContext *cx near-universal first parameter to TLS. Two bugs of interest:

https://bugzilla.mozilla.org/show_bug.cgi?id=659404 Measure TLS cost on platforms/compilers of interest
https://bugzilla.mozilla.org/show_bug.cgi?id=659241 IonMonkey: Stop passing JSContext * everywhere

/be

[BrendanEich]

https://bugzilla.mozilla.org/show_bug.cgi?id=659404 might be of interest.

/be

[pcwalton]

This is working as of commit 9ba8617. Leaving this open while I work on removing a bunch of redundant rust_task traffic.

[pcwalton]

This is needed for stack growth to work; bumping to 0.1 blocker.

[pcwalton]

Enough of this is implemented to get stack growth to work; demoting from 0.1 blocker status.

[marijnh]

Does this mean we could drop the task argument to all functions and just make the runtime get it from tls when it needs it?

[pcwalton]

Yes indeed :) It's a big win. Check out how it works in upcall_malloc (I think I put it there for testing).