<html><head><meta http-equiv="content-type" content="text/html; charset=utf-8"></head><body dir="auto"><div>We are more likely to fix a crash. So I don't think it is firing. The only case this could regress is a potentially trapping operation that doesn't actually trap for reasons that are invisible to the optimizer.</div><div><br></div><div>Adding the extra control flow to guard it is not clear to be a win. We should probably fix the miscompilation first. If we do have a perf regression somewhere, it can be addressed in a subsequent patch.<br><br>Sent from my iPhone</div><div><br>On 2013-06-04, at 11:22, James Molloy <<a href="mailto:james@jamesmolloy.co.uk">james@jamesmolloy.co.uk</a>> wrote:<br><br></div><blockquote type="cite"><div>Hi Rafael,<div><br></div><div>I don't know. Do we know this broken optimization doesn't kick in in the test suite? If it does, we'd regress on performance if we took the conservative way out.</div><div><br></div>
<div>Just my 2c.</div><div><br></div><div>Cheers,</div><div><br></div><div>James<br><br><div class="gmail_quote">On 4 June 2013 15:40, Rafael Espíndola <span dir="ltr"><<a href="mailto:rafael.espindola@gmail.com" target="_blank">rafael.espindola@gmail.com</a>></span> wrote:<br>
<blockquote class="gmail_quote" style="margin:0 0 0 .8ex;border-left:1px #ccc solid;padding-left:1ex"><div class="im">On 4 June 2013 09:49, James Molloy <<a href="mailto:james@jamesmolloy.co.uk">james@jamesmolloy.co.uk</a>> wrote:<br>
> Hi David,<br>
><br>
> Might a more performant solution not be to wrap any hoisted instructions<br>
> that are not safe to expand in a conditional that tests if the loop will be<br>
> entered at least once?<br>
<br>
</div>Is that really worth it? Given that this bug went undetected for years<br>
(it was already broken in r122610), I don't think it would kick often<br>
enough to justify it.<br>
<br>
Cheers,<br>
Rafael<br>
</blockquote></div><br></div>
</div></blockquote></body></html>