Full Text Bug Listing

John Engelhart

Reported 2008-09-13 12:52:32 PDT

Very, very slow interactive response for a <input type="search"> field. CPU jumps to 100% for each key press, and can take a few seconds for just a dozen keystrokes. I've condensed down the HTML that reproduces the bug. Tried with nightly r35942 and Mac OS X Safari 4.0 (5528.1, I think the latest developer preview available). I also grabbed a stack sample and included the heavy hitting part in the html comments. I have no idea which of the bits and pieces cause the problem, so the HTML code that I'm including is the result of distilling the problem down in to something minimal that definitely reproduces the issue. The nightly build is noticeably more sluggish than the Safari 4 developer preview, but not by much. --- begin HTML code to reproduce problem --- <!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01//EN" "http://www.w3.org/TR/html4/strict.dtd"> <html lang="en"> <head> <meta http-equiv="Content-Type" content="text/html;charset=utf-8"> <style type="text/css"> .alphaTopAndBottomFade { border-bottom: 1px solid rgba(0,0,0,0.075); border-top: 1px solid rgba(0,0,0,0.2); border-left: 1px solid rgba(0,0,0,0.05); border-right: 1px solid rgba(0,0,0,0.05); -webkit-box-shadow: 0px 0px 1px rgba(0,0,0,0.2); padding-bottom: 0.5em; margin-bottom: 0.5em; background: -webkit-gradient(linear, left top, left bottom, from(rgb(213,221,231)), color-stop(0.33, rgb(255,255,255)), color-stop(0.66, rgb(255,255,255)), to(rgb(240,240,240))); -webkit-border-radius: 5px; -moz-border-radius: 5px; } .table { display: table; } .table > .row { display: table-row; } .table > .row > .cell { display: table-cell; } </style> </head> <body> <div class="content" style=""> <div class="alphaTopAndBottomFade" style="position: relative;"> <div class="table" style="width: 100%; height: 300px; padding: 0.375em 3ex;"><div class="row"> <div class="cell" style="width: 50%; height: 100%;"> <div style="width: 100%; height: 100%; position: relative;"> <div style="margin: 0em 3ex; color: #333; font-family: 'Lucida Grande'; font-weight: normal;"> <div style="font-size: 18px; font-weight: normal;">Title</div> <div style="color: #666; margin-left: 1ex; margin-top: 0.25em;">Subtitle!</div> </div> <div style="font: 11px normal 'Lucida Grande'; color: #555"> <p>Yay for us.. and stuff</p> <p>And some more bits and pieces.</p>   <input type="search" name="req" value="" results="5" placeholder="Search This" autosave="">  </div> </div> </div> </div> </div> </div> </div> </body> </html>

John Engelhart

Comment 1 2008-09-15 16:13:17 PDT

While doing more web page hacking, it's become pretty clear that the following combination causes a huge CPU spike (on a 1.5 GHz g4 powerbook): Some kind of container (div). A webkit transition, preferably one that infinitely loops. And the one that really breaks the bank: Some element in the render hiearchy that is a parent for the animating webkit element must have a background: -webkit-gradient(linear, left top, left bottom, from(rgba(255,255,255,0.25)), to(rgba(0,0,0,1.0))); (or equiv). The use of rgba / alpha in the gradient doesn't seem to matter, all that's needed is the gradient. Hopefully this helps narrow things down.

John Engelhart

Comment 2 2008-09-15 16:14:29 PDT

Crap. Strike my last (comment #1). Pasted to the wrong bug. :(.

John Engelhart

Comment 3 2008-09-19 01:07:28 PDT

I now strongly suspect that this bug is related to a different bug I've just recently filed: bug # 20929. I tested this bugs included snippet and removed the border attributes for .alphaTopAndBottomFade and the problem went away.

Simon Fraser (smfr)

Comment 4 2009-01-03 17:06:57 PST

Bad search field redrawing in the testcase filed as bug 23093. John: if I extract the HTML you posted (with the garbage in the middle removed), typing in the search field seems fine. Please attach an HTML file that shows the problem using the Create New Attachment link above.

John Engelhart

Comment 5 2009-01-03 20:59:53 PST

(In reply to comment #4) > Bad search field redrawing in the testcase filed as bug 23093. > > John: if I extract the HTML you posted (with the garbage in the middle > removed), typing in the search field seems fine. Please attach an HTML file > that shows the problem using the Create New Attachment link above. > I just retested this as requested. I tested it on a late 2008 unibody MacBook (2.0GHz) and a 1.5GHz PPC PowerMac. The problem doesn't really manifest itself in any noticeable fashion on the faster machine. Only on the slower powerbook is it noticeable. Just tried forcing Safari.app to execute under Rosetta and it's slightly noticeable, but still not to the degree that it is on the powerbook. I'm afraid to really see this in action, you'll need to find yourself an old PPC machine. For grins, I tried it on an iPod touch, but it handles the text input in a completely different way, so can't really test it that way unfortunately. IMHO, the machines are fast enough now that you can't really perceive a problem any more. However (and this is just a guy feeling) I think it still points to an underlying performance problem, possibly in the rendering code. Thinking back, I remember digging a little bit more in to this, and I seem to remember coming to the conclusion that it seemed like a case of 'over redrawing', that each keystroke triggered a storm of unnecessary redraws. My initial comment says that getting rid of the -webkit-gradient background causes the problem to go away... does that ring any bells? Something about having a -webkit-gradient in the 'NSView' (sorry, I don't know webkit all that well, but whatever is analogous to the cocoa NSView paradigm) hierarchy that might be causing some over-zealous redraw/refresh? So, just to be clear: The problem still exists on my 1.5GHz PPC PowerBook, but isn't noticeable on a 2.0GHz x86 late 2008 unibody MacBook. Hmm, for grins, I just tried the example HTML unmodified in one table, and commenting out the .alpha... CSS part in another tab on my MacBook, and then typing in both search boxes. Watching MenuMeter's CPU usage, the unmodified version does use a noticeable (but small) amount of CPU, and the commented out version uses an undetectable amount of CPU. I suspect that a Shark.app session between the two versions (unmodified, and .alpha... CSS commented out) would be revealing to someone who knows the code base.

Simon Fraser (smfr)

Comment 6 2009-01-04 10:46:54 PST

If there really is a perf issue, it should be possible to cook up a testcase that shows it even on faster machines. Maybe try lots of gradient divs layered under the search field, or something? However, without a testcase showing the problem in an obvious way, it's not going to get attention.

Summary:	Very slow <input type="search">. 100% CPU usage for each key press.
Product:	WebKit	Reporter:	John Engelhart <john.engelhart>
Component:	Text	Assignee:	Nobody <webkit-unassigned>
Status:	RESOLVED WORKSFORME
Severity:	Normal	CC:	simon.fraser
Priority:	P2
Version:	528+ (Nightly build)
Hardware:	Mac
OS:	OS X 10.5

Bug 20831