[AArch64] FP load balancing pass for Cortex-A57

[James Molloy]

Hi Renato,

http://reviews.llvm.org/D4791

Here's one I made earlier ;) Also known as, forgetting to SVN add my test file. No wonder you asked for "more" tests! :(

I've made the formatting and dox changes you suggested too. Although I'm still not sure what the preferred way to format a lambda inside a std::sort is...

Cheers,

James

-----Original Message-----
From: Renato Golin [mailto:renato.golin at linaro.org] 
Sent: 05 August 2014 16:26
To: James Molloy
Cc: Tim Northover; llvm-commits
Subject: Re: [AArch64] FP load balancing pass for Cortex-A57

On 5 August 2014 16:18, James Molloy <james.molloy at arm.com> wrote:
> Thanks for the feedback! "More tests!" - gotcha.

As usual... :)


> The reason we pick those requiring fixups first is exactly because of that comment. We can't change the parity of those chains, so it is best to look at them first, so the global "parity counter" has taken them into account when we get to the chains that we *can* recolor. This, I think, is the best way to achieve the most balanced parity overall.

Get the largest and hardest ones first. Makes sense.


> Yes, this is on a basic block level. Crossing basic blocks I think would be difficult - as soon as you get a phi node you have multiple incoming blocks to worry about, and you have to worry about execution paths. I could think about it but I'm not sure it'd be easy!

Thought so. Hopefully, if the results here were vectorized, they'll probably be if-converted if they had branches, and thus hopefully on the same basic block or on simpler sequential basic blocks. If that's true, you could (later, of course) try to just peek in the next block and give a try. But that's well in the future, not this patch.

cheers,
--renato

PS: Can you use Phab for the next steps?