Bug 15617

Summary:

improve speed of integer conversions

Product:

WebKit

Reporter:

Darin Adler <darin>

Component:

JavaScriptCore

Assignee:

Darin Adler <darin>

Status:

RESOLVED FIXED

Severity:

Enhancement

Priority:

Version:

523.x (Safari 3)

Hardware:

Mac

OS:

OS X 10.4

Attachments:

Description	Flags
patch with change log	none
patch with change log	none
patch with change log (newer version with fixed .exp file)	eric: review+

Darin Adler

Reported 2007-10-22 07:31:19 PDT

I was able to get a 1% speedup on SunSpider by eliminating an unnecessary float/double conversion.

Attachments
patch with change log (4.45 KB, patch) 2007-10-22 07:36 PDT, Darin Adler	no flags	Details Formatted Diff Diff
patch with change log (14.77 KB, patch) 2007-10-22 10:15 PDT, Darin Adler	no flags	Details Formatted Diff Diff
patch with change log (newer version with fixed .exp file) (14.77 KB, patch) 2007-10-22 10:26 PDT, Darin Adler	eric: review+	Details Formatted Diff Diff
Show Obsolete (2) View All Add attachment proposed patch, testcase, etc.

Darin Adler

Comment 1 2007-10-22 07:36:43 PDT

Created attachment 16798 [details] patch with change log

Mark Rowe (bdash)

Comment 2 2007-10-22 07:51:49 PDT

This looks nearly identical to Eric's patch on bug 15607.

Darin Adler

Comment 3 2007-10-22 10:05:00 PDT

I have a better version in the works.

Darin Adler

Comment 4 2007-10-22 10:15:39 PDT

Created attachment 16802 [details] patch with change log

Darin Adler

Comment 5 2007-10-22 10:26:42 PDT

Created attachment 16803 [details] patch with change log (newer version with fixed .exp file)

Eric Seidel (no email)

Comment 6 2007-10-22 10:49:28 PDT

Comment on attachment 16803 [details] patch with change log (newer version with fixed .exp file) Looks great. I assume you ran the tests. Also, (pot calling the kettle black here) I think it would be nice to replace all your float special values 4294967296.0, etc with some nice names (to make the code make sense to someone who didn't just hack on it for a couple hours. :) (Also, nice catch with toInt32SlowCase calling toUInt32. I'm curious if that would ever be "wrong" (such that we could build a test). I guess it just changes what values could be stored in the float, since it's just doing the sign wrap at a different time.

Darin Adler

Comment 7 2007-10-22 10:57:10 PDT

(In reply to comment #6) > (Also, nice catch with > toInt32SlowCase calling toUInt32. I'm curious if that would ever be "wrong" > (such that we could build a test). It wasn't wrong in the sense of "buggy". It was just inefficient for negative numbers.

Darin Adler

Comment 8 2007-10-22 10:58:40 PDT

(In reply to comment #6) > I think it would be nice to replace all your float special values > 4294967296.0, etc with some nice names (to make the code make sense to someone > who didn't just hack on it for a couple hours. :) I'll consider that. I'm worried that if I use const double or const float that it will generate different code, so it's a change I need to make carefully with measurements.

Note You need to log in before you can comment on or make changes to this bug.