Coalesce writes in SVGPathByteStreamBuilder

Coalesce writes in SVGPathByteStreamBuilder

Add a helper class CoalescingBuffer, that collects the data for a segment
into a local (stack) buffer, and then writes the entire stack buffer in
one go to the SVGPathByteStream.

This reduces codesize (on GCC 4.8) by roughly 50% (closePath being the
only method that grows).

Committed: https://src.chromium.org/viewvc/blink?view=rev&revision=166523

[fs, 2014-02-05 16:43:00]

Crazy stuff(?). PTAL.

(Some quick perf runs shows that this also speeds things up by slightly less
than 10% - but that likely varies by compiler and target, so just mentioning it
in passing.)

[Stephen Chennney, 2014-02-05 16:46:16]

Actually, really good idea. LGTM.

[fs, 2014-02-05 17:20:34]

The CQ bit was checked by fs@opera.com

[commit-bot: I haz the power, 2014-02-05 17:21:07]

CQ is trying da patch. Follow status at
https://chromium-status.appspot.com/cq/fs@opera.com/149913004/1

[commit-bot: I haz the power, 2014-02-05 18:15:45]

Change committed as 166523

[Chris Evans, 2014-02-05 18:54:48]

https://codereview.chromium.org/149913004/diff/1/Source/core/svg/SVGPathByteS...
File Source/core/svg/SVGPathByteStreamBuilder.cpp (right):

https://codereview.chromium.org/149913004/diff/1/Source/core/svg/SVGPathByteS...
Source/core/svg/SVGPathByteStreamBuilder.cpp:41:
m_byteStream->append(m_bytes[i]);
Drive-by: I was looking at this recently and noticed another optimization we
could do.

We could call Vector::ensureCapacity(m_currentOffset) and then do a raw memcpy()
to Vector::end(). This will eliminate a vector length check per-append.

This path is super hot on some of the SVG/ benchmarks.

[fs, 2014-02-06 08:12:29]

On 2014/02/05 18:54:48, Chris Evans wrote:
>
https://codereview.chromium.org/149913004/diff/1/Source/core/svg/SVGPathByteS...
> File Source/core/svg/SVGPathByteStreamBuilder.cpp (right):
> 
>
https://codereview.chromium.org/149913004/diff/1/Source/core/svg/SVGPathByteS...
> Source/core/svg/SVGPathByteStreamBuilder.cpp:41:
> m_byteStream->append(m_bytes[i]);
> Drive-by: I was looking at this recently and noticed another optimization we
> could do.
> 
> We could call Vector::ensureCapacity(m_currentOffset) and then do a raw
memcpy()
> to Vector::end().

Yepp, that would be nice, but the Vector interface would need some extensions to
achieve a sequence like:

v.reserveCapacity(v.size() + totalByteCount);
memcpy(v.end(), bytes, itemByteCount);
...more memcpys...
v.setSize(v.size() + totalByteCount); <<-- "Missing" AFAICS

Unless one wants to jinx the size of course...

> This will eliminate a vector length check per-append.

Using the append(const T*, size_t) version does eliminate the per-byte
length-check too, but there was no noticeable improvement in doing so (in my
measurements).