[llvm] r202304 - Add a limit to the heuristic that register allocates instructions in local order.

[Renato Golin]

On 27 February 2014 19:36, Andrew Trick <atrick at apple.com> wrote:
> - A bug in the host compiler on that builder.

I'm using Ubuntu's own 4.8.1, which has been working for months. I
think the chances of the compiler being bad are pretty slim.

What kind of LNT runs do you do on ARM? Is it Darwin only?


> - A recent bug in CodeGen that has been hidden by that heuristic and is now exposed.

I believe this one is the one we should look for. It fails at -O3
only, so it's probably something that a new optimization is doing and
the old heuristics can't cope with.

I did a quick debug session and it failed on something that I remember
seeing last year. The same problem, and it was magically fixed, and if
memory serves me well, it was a register allocation commit that "fixed
it", it might just as well be yours...


> Reverting would certainly only hide whatever the issue is.

Agree.


> 3. I can temporarily disable this change on armv7 using a target hook and file a PR.

If you could have a look at it on an ARM/Linux, we don't need to
disable it temporarily. But Linaro Connect will start this weekend and
will take a whole next week, and I won't be able to look at it until
the other Monday. When I do start, it'll take a while before I can
find out where it's breaking, and get a closer idea of how to fix it.

If not, I think we should disable on ARM, create a PR and assign it to me.

Makes sense?

cheers,
--renato