[LLVMdev] Vectorizer using Instruction, not opcodes

[David Tweed]

Hi Renato,

On 5 February 2013 11:03, David Tweed <david.tweed at arm.com<mailto:david.tweed at arm.com>> wrote:
since the more instructions there
are the more an out-of-order CPU can put them into otherwise unused slots. I
can't think of a way of figuring out such a mapping other than empirically.


> Given the amount of uncertainty on these OOO guesses, I don't think we can get anything worth trying, even empirically. The noise will always outweigh the signal.

> You can normally save a few cycles on a static micro-benchmark and think you're in control, but you normally don't evaluate the exact impact that the same guess had on other benchmarks or real code. Since this is auto-vectorization, it'll be hard to
>  compare, but more so to evaluate the overall impact on the rest.

That may be the case. Just to be clear I was thinking of something incredibly simple along the lines of (cost) ^ (min(1-alpha*noOfINstructions,floorTheshold)) for some values of alpha and floorThreshold. I'll probably have a look at this when I next get some discretionary time, but that wont' be for a couple of months. Having said that, it may turn out that it doesn't really matter in that the more vector instructions you've got the higher the likelihood it will have a lower cost than a scalar version.
ave

-- IMPORTANT NOTICE: The contents of this email and any attachments are confidential and may also be privileged. If you are not the intended recipient, please notify the sender immediately and do not disclose the contents to any other person, use it for any purpose, or store or copy the information in any medium. Thank you.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20130205/cd754f4e/attachment.html>