<html><head><meta http-equiv="Content-Type" content="text/html charset=utf-8"></head><body style="word-wrap: break-word; -webkit-nbsp-mode: space; -webkit-line-break: after-white-space;" class=""><br class=""><div><blockquote type="cite" class=""><div class="">On Apr 15, 2015, at 2:18 AM, Chandler Carruth <<a href="mailto:chandlerc@google.com" class="">chandlerc@google.com</a>> wrote:</div><br class="Apple-interchange-newline"><div class=""><div dir="ltr" class=""><div class="gmail_quote">On Wed, Apr 15, 2015 at 2:13 AM Nick Lewycky <<a href="mailto:nicholas@mxc.ca" class="">nicholas@mxc.ca</a>> wrote:<br class=""><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"> So the criterial<br class="">
> deciding whether to do the merging for a case is to see whether<br class="">
> merging is beneficial. We want to find out all the important cases for<br class="">
> which gep merging is beneficial.<br class="">
<br class="">
My default assumption is that gep merging is beneficial; the cost<br class="">
metrics all assume that fewer instructions is better, your .bc will be<br class="">
smaller, your optimizations have fewer instructions to look through,<br class="">
some optimizations have limits on how many instructions they will look<br class="">
through, etc. We want to find out all the important cases for which gep<br class="">
merging is harmful, or find out which cases should have gep splitting.<br class="">
<br class="">
Put another way, if having these GEPs split apart is better then why<br class="">
doesn't the optimizer turn a single GEP into two when that's better?<br class="">
Only removing the optimization here can't be the right answer.<br class=""></blockquote><div class=""><br class=""></div><div class="">I agree. As Hal said, I'm pretty worried about blocking merging in an ad-hoc way. It seems really likely to create limitations in every part of the optimizer that wants to look through things.</div><div class=""><br class=""></div><div class="">Unless we have a sufficiently strong heuristic to *split* GEPs, my suspicion is that we don't have the correct heuristic.</div></div></div></div></blockquote><div><br class=""></div><div>Well, Wei’s numbers seem to prove this is not an issue in practice.</div><div>I would rather determine the right heuristic on actual motivating examples than sitting on the performance benefits Wei demonstrated.</div><div><br class=""></div><div>What would it take to convince you (not you in particular Chandler :)) that Wei’s heuristic is good enough for now?</div><div><br class=""></div><div>Would a diff of the binaries showing no regressions on the would test suite for, let say x86_64, arm, aarch64, and PPC, do?</div><div>Note: I am talking about binary diff, no need to run.</div><div><br class=""></div><div>BTW, I assume Wei, that you already done something similar.</div><div>What were the results?</div><div><br class=""></div><div>Cheers,</div><div>-Quentin</div><br class=""><blockquote type="cite" class=""><div class=""><div dir="ltr" class=""><div class="gmail_quote"><div class=""> </div><blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex">
<br class="">
> Quentin mentioned the beneficial addressing mode case I didn't think<br class="">
> of before. And it is possible that I can filter out some testcases<br class="">
> from internal benchmarks. working on it right now.<br class="">
<br class="">
Sounds good! I've heard the addressing mode matching code is a pain, but<br class="">
if we could make it smarter I think that would be better. If you have a<br class="">
testcase where GEP merging causes a loss of performance then you also<br class="">
have a testcase where the addressing mode matcher is missing something,<br class="">
right?<br class="">
_______________________________________________<br class="">
llvm-commits mailing list<br class="">
<a href="mailto:llvm-commits@cs.uiuc.edu" target="_blank" class="">llvm-commits@cs.uiuc.edu</a><br class="">
<a href="http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits" target="_blank" class="">http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits</a><br class="">
</blockquote></div></div>
_______________________________________________<br class="">llvm-commits mailing list<br class=""><a href="mailto:llvm-commits@cs.uiuc.edu" class="">llvm-commits@cs.uiuc.edu</a><br class="">http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits<br class=""></div></blockquote></div><br class=""></body></html>