424693 - js_ComputeFilename accounts for 10-12% of date-format-tofte

Status: RESOLVED FIXED

(Core :: JavaScript Engine, defect)

Description

[Robert Sayre]

This test uses eval, and the filename/line# calculations are expensive there.

Comment 1

[Igor Bukanov]

Judging by the code, I think the culprit is js_PCToLineNumber. It walks through all source notes of the script containing the eval to find the note that contains the given pc. A simple fix would be to embed the line number into JSOP_EVAL. 

Comment 2

[Igor Bukanov]

Attachment: v1

The patch introduces JSOP_LINENO and always generates the new bytecode after JSOP_EVAL. I use this independent bytecode and don't extend JSOP_EVAL with an extra 2 bytes to store lineno to avoid extra checks in the shared implementation of JSOP_CALL and JSOP_EVAL.

Comment 3

[Igor Bukanov]

Attachment: v1 for real

Comment 4

[Mike Shaver (:shaver -- probably not reading bugmail closely)]

Comment on attachment 311302 [details] [diff] [review]
v1 for real

r=shaver, indeed.  (The comment above JSXDR_BYTECODE_VERSION should really be changed, since we want to increment the second term, to decrement the expression's value!)

Comment 5

[Igor Bukanov]

Comment on attachment 311302 [details] [diff] [review]
v1 for real

Asking for approval: this is a safe optimization to avoid expensive calculations of the line number based on script's pc pointer.

Comment 6

[Mike Beltzner [:beltzner, not reading bugmail]]

Comment on attachment 311302 [details] [diff] [review]
v1 for real

a1.9b5=beltzner

Comment 7

[Igor Bukanov]

I checked in the patch from the comment 4 to the trunk:

http://bonsai.mozilla.org/cvsquery.cgi?module=PhoenixTinderbox&branch=HEAD&cvsroot=%2Fcvsroot&date=explicit&mindate=1206377895&maxdate=1206378079&who=igor%25mir2.org

Comment 8

[Igor Bukanov]

Attachment: using C, not C++ comments

In the checked in patch I have used // FALL THROUGH. This patch fixes this. I will check it in when the tree will be green.