Bug 35362 – Parsing of CSS rgb() values is slow

• Component: New Bugs
• Status: RESOLVED FIXED

[Stephen White]

Parsing of CSS rgb() values is slow

[Stephen White]

When running the "Complex Graphics" test from the Peacekeeper suite, upwards of 25% of CPU time is taken parsing CSS rgb() values.  Implementing a simple ad-hoc parser (as is done for hex color values) would go a long way to improving performance on this test.

[Stephen White]

Created attachment 49447 [details]
First draft

[Stephen White]

Re: patch1:

This functionality could be moved up into the other flavor of parseColor() on line 278. The layout tests seem ok with it, but I wasn't sure if this is correct.

[Simon Fraser (smfr)]

*** Bug 23110 has been marked as a duplicate of this bug. ***

[Simon Fraser (smfr)]

You should just do rgba() as well.

[Oliver Hunt]

Comment on attachment 49447 [details]
First draft

String doesn't guarantee null termination, you should instead pass in a length parameter, then you can do:
const UChar* end = *str + length;
and the *c pieces become
c != end

[Stephen White]

Created attachment 49597 [details]
Patch

[Stephen White]

Ok, fixed.  Also changed to use isCSSWhitespace(), since it doesn't include vertical tab, and isASCIISpace() does.

[Stephen White]

(In reply to comment #5)
> You should just do rgba() as well.

Unfortunately, rgba() requires parsing of floats (for the alpha), which is a bit trickier than ints.  I'll give it a shot but I might end up doing another patch for it if that's ok.

[Oliver Hunt]

Comment on attachment 49597 [details]
Patch

I would be interested in what you get for perf numbers if you replace isCSSWhitespace with a simplified ' ' and '\t' only check -- on the basis we are wanting to get the maximum perf out of the normal case.

Also my reading of this seems to indicate that this will accept
rgb(1,1,1)<random text>

[Oliver Hunt]

(In reply to comment #9)
> (In reply to comment #5)
> > You should just do rgba() as well.
> 
> Unfortunately, rgba() requires parsing of floats (for the alpha), which is a
> bit trickier than ints.  I'll give it a shot but I might end up doing another
> patch for it if that's ok.

I think it's important to understand you don't need to parse all valid inputs, you only need the fast path to handle the common cases, eg. 0.[0-9]{1,5} say

[Darin Adler]

(In reply to comment #10)
> I would be interested in what you get for perf numbers if you replace
> isCSSWhitespace with a simplified ' ' and '\t' only check -- on the basis we
> are wanting to get the maximum perf out of the normal case.

I would think space only. I expect that very few cases have tabs.

[Stephen White]

Created attachment 49602 [details]
Patch

[Stephen White]

(In reply to comment #10)
> (From update of attachment 49597 [details])
> I would be interested in what you get for perf numbers if you replace
> isCSSWhitespace with a simplified ' ' and '\t' only check -- on the basis we
> are wanting to get the maximum perf out of the normal case.

Ok, I gave it a try.  The difference was negligible:  +1.75% one run, -1.8% the
next, so within the variance of this particular test, which is about +/-2%. 
(Note that this test is +52.7% faster overall with my fix).

> Also my reading of this seems to indicate that this will accept
> rgb(1,1,1)<random text>

Fixed.  Note that I'm not accepting trailing whitespace either, which I figure is the
common case (no whitespace after the paren).

[Darin Adler]

Comment on attachment 49602 [details]
Patch

> +static inline bool parseInt(const UChar** str, const UChar* end, UChar terminator, int* value)

For in/out arguments we normally use references rather than pointers. Thus "str" and "value" would typically be references.

> +    const UChar *c = *str;
> +    int v = 0;

The "*" here is in the wrong place, should be next to the UChar.

We normally use words to name local variables, not letters.

> +        const UChar* c = name.characters() + 4;

We normally use words to name local variables, not letters.

> +        int r = 0, g = 0, b = 0;

We normally define variables on separate lines rather than all on one line.

There's no need to initialize these variables. I suggest not doing so.

I'm going to say r=me, despite these concerns since they are all stylistic issues.

[Stephen White]

(In reply to comment #15)
> (From update of attachment 49602 [details])
> > +static inline bool parseInt(const UChar** str, const UChar* end, UChar terminator, int* value)
> 
> For in/out arguments we normally use references rather than pointers. Thus
> "str" and "value" would typically be references.
> 
> > +    const UChar *c = *str;
> > +    int v = 0;
> 
> The "*" here is in the wrong place, should be next to the UChar.
> 
> We normally use words to name local variables, not letters.
> 
> > +        const UChar* c = name.characters() + 4;
> 
> We normally use words to name local variables, not letters.
> 
> > +        int r = 0, g = 0, b = 0;
> 
> We normally define variables on separate lines rather than all on one line.
> 
> There's no need to initialize these variables. I suggest not doing so.
> 
> I'm going to say r=me, despite these concerns since they are all stylistic
> issues.

Fixed.

(I wasn't sure if you meant "this is ok, but fix these before landing", "fix and re-upload" or "can land as is", so I fixed and re-uploaded just to be safe).

[Stephen White]

Created attachment 49612 [details]
Patch

[Stephen White]

Committed r55308: <http://trac.webkit.org/changeset/55308>