Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

This patch adds an inline capacity of 32 to CSSTokenizer's vector of
parser tokens. This improves performance of parsing short pieces of CSS,
for example setting inline styles, as we don't need to allocate memory
for this vector. The capacity of 32 is chosen arbitrarily to accomodate
most uses of inline styles (we reserve an initial capacity of length/3).

This patch makes setting inline transforms to strings that miss the fast
path ~2% faster on my z620. Note that this is not a reliable testing
device, but we still expect that this optimisation improves performance.

BUG=619498
Committed: https://crrev.com/7d18ae343a06b83d460ee4084fa2f5bc0f1e4acf
Cr-Commit-Position: refs/heads/master@{#399426}

[Timothy Loh, 2016-06-13 09:50:16]

Description was changed from

==========
Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

This patch adds an inline capacity of 32 to CSSTokenizer's vector of
parser tokens. This improves performance of parsing short pieces of CSS,
for example setting inline styles, as we don't need to allocate memory
for this vector. The capacity of 32 is chosen arbitrarily to accomodate
most uses of inline styles (we reserve an initial capacity of length/3).

This patch improves performance on the test [XXX] by [XXX]%

BUG=619498
==========

to

==========
Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

This patch adds an inline capacity of 32 to CSSTokenizer's vector of
parser tokens. This improves performance of parsing short pieces of CSS,
for example setting inline styles, as we don't need to allocate memory
for this vector. The capacity of 32 is chosen arbitrarily to accomodate
most uses of inline styles (we reserve an initial capacity of length/3).

This patch makes setting inline transforms to strings that miss the fast
path ~2% faster on my z620. Note that this is not a reliable testing
device, but we still expect that this optimisation improves performance.

BUG=619498
==========

[Timothy Loh, 2016-06-13 09:53:43]

timloh@chromium.org changed reviewers:
+ esprehn@chromium.org, meade@chromium.org

[Timothy Loh, 2016-06-13 09:53:44]

Numbers from my z620 are not as large as I expected but I haven't set it up as a
perf testing machine so the actual change could be bigger or smaller.

[esprehn, 2016-06-13 09:59:45]

On 2016/06/13 at 09:53:44, timloh wrote:
> Numbers from my z620 are not as large as I expected but I haven't set it up as
a perf testing machine so the actual change could be bigger or smaller.

The Z620 can give some very strange results if your benchmark is too fast,
what's your benchmark? You could try running perf+pprof against the
Leaves>Animometer benchmark if you want to check. I think this is good change
either way though, we should land it.

We probably want to do the same optimization for CSSVariableResolver eventually
too btw, it has on stack vectors of tokens.

lgtm

[Timothy Loh, 2016-06-13 10:07:34]

On 2016/06/13 09:59:45, esprehn wrote:
> On 2016/06/13 at 09:53:44, timloh wrote:
> > Numbers from my z620 are not as large as I expected but I haven't set it up
as
> a perf testing machine so the actual change could be bigger or smaller.
> 
> The Z620 can give some very strange results if your benchmark is too fast,
> what's your benchmark? You could try running perf+pprof against the
> Leaves>Animometer benchmark if you want to check. I think this is good change
> either way though, we should land it.

It sets translate on inline style in a loop some numbers of times (takes about
10s). I alternated running with and without the change a few times.

> We probably want to do the same optimization for CSSVariableResolver
eventually
> too btw, it has on stack vectors of tokens.

Yep, I just wanted to make this small since I realise we want to get this in
real soon.

[Timothy Loh, 2016-06-13 10:07:40]

The CQ bit was checked by timloh@chromium.org

[commit-bot: I haz the power, 2016-06-13 10:07:55]

CQ is trying da patch. Follow status at
 https://chromium-cq-status.appspot.com/patch-status/2058403002/1

[Timothy Loh, 2016-06-13 10:08:28]

On 2016/06/13 10:07:34, Timothy Loh wrote:
> On 2016/06/13 09:59:45, esprehn wrote:
> > On 2016/06/13 at 09:53:44, timloh wrote:
> > > Numbers from my z620 are not as large as I expected but I haven't set it
up
> as
> > a perf testing machine so the actual change could be bigger or smaller.
> > 
> > The Z620 can give some very strange results if your benchmark is too fast,
> > what's your benchmark? You could try running perf+pprof against the
> > Leaves>Animometer benchmark if you want to check. I think this is good
change
> > either way though, we should land it.
> 
> It sets translate on inline style in a loop some numbers of times (takes about
> 10s). I alternated running with and without the change a few times.

It's just a slightly modified version of what you uploaded to 605792

[commit-bot: I haz the power, 2016-06-13 10:52:18]

Description was changed from

==========
Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

This patch adds an inline capacity of 32 to CSSTokenizer's vector of
parser tokens. This improves performance of parsing short pieces of CSS,
for example setting inline styles, as we don't need to allocate memory
for this vector. The capacity of 32 is chosen arbitrarily to accomodate
most uses of inline styles (we reserve an initial capacity of length/3).

This patch makes setting inline transforms to strings that miss the fast
path ~2% faster on my z620. Note that this is not a reliable testing
device, but we still expect that this optimisation improves performance.

BUG=619498
==========

to

==========
Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

This patch adds an inline capacity of 32 to CSSTokenizer's vector of
parser tokens. This improves performance of parsing short pieces of CSS,
for example setting inline styles, as we don't need to allocate memory
for this vector. The capacity of 32 is chosen arbitrarily to accomodate
most uses of inline styles (we reserve an initial capacity of length/3).

This patch makes setting inline transforms to strings that miss the fast
path ~2% faster on my z620. Note that this is not a reliable testing
device, but we still expect that this optimisation improves performance.

BUG=619498
==========

[commit-bot: I haz the power, 2016-06-13 10:52:19]

Committed patchset #1 (id:1)

[commit-bot: I haz the power, 2016-06-13 10:52:24]

CQ bit was unchecked

[commit-bot: I haz the power, 2016-06-13 10:54:40]

Description was changed from

==========
Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

This patch adds an inline capacity of 32 to CSSTokenizer's vector of
parser tokens. This improves performance of parsing short pieces of CSS,
for example setting inline styles, as we don't need to allocate memory
for this vector. The capacity of 32 is chosen arbitrarily to accomodate
most uses of inline styles (we reserve an initial capacity of length/3).

This patch makes setting inline transforms to strings that miss the fast
path ~2% faster on my z620. Note that this is not a reliable testing
device, but we still expect that this optimisation improves performance.

BUG=619498
==========

to

==========
Use an inline capacity for CSSTokenizer's Vector<CSSParserToken>

This patch adds an inline capacity of 32 to CSSTokenizer's vector of
parser tokens. This improves performance of parsing short pieces of CSS,
for example setting inline styles, as we don't need to allocate memory
for this vector. The capacity of 32 is chosen arbitrarily to accomodate
most uses of inline styles (we reserve an initial capacity of length/3).

This patch makes setting inline transforms to strings that miss the fast
path ~2% faster on my z620. Note that this is not a reliable testing
device, but we still expect that this optimisation improves performance.

BUG=619498

Committed: https://crrev.com/7d18ae343a06b83d460ee4084fa2f5bc0f1e4acf
Cr-Commit-Position: refs/heads/master@{#399426}
==========

[commit-bot: I haz the power, 2016-06-13 10:54:41]

Patchset 1 (id:??) landed as
https://crrev.com/7d18ae343a06b83d460ee4084fa2f5bc0f1e4acf
Cr-Commit-Position: refs/heads/master@{#399426}

[esprehn, 2016-06-13 11:56:13]

On 2016/06/13 at 10:08:28, timloh wrote:
> On 2016/06/13 10:07:34, Timothy Loh wrote:
> > On 2016/06/13 09:59:45, esprehn wrote:
> > > On 2016/06/13 at 09:53:44, timloh wrote:
> > > > Numbers from my z620 are not as large as I expected but I haven't set it
up
> > as
> > > a perf testing machine so the actual change could be bigger or smaller.
> > > 
> > > The Z620 can give some very strange results if your benchmark is too fast,
> > > what's your benchmark? You could try running perf+pprof against the
> > > Leaves>Animometer benchmark if you want to check. I think this is good
change
> > > either way though, we should land it.
> > 
> > It sets translate on inline style in a loop some numbers of times (takes
about
> > 10s). I alternated running with and without the change a few times.
> 
> It's just a slightly modified version of what you uploaded to 605792

Btw I see Vector<CSSParserTokenType> m_blockStack; also show up in profiles,
that's block nesting like @media { .foo { } } ? We should probably add a small
inline capacity there too.