CPU cycles per SVGImage

[Kouhei Ueno]

Hi fs,

Thank you for reviewing and so sorry for confusing you on the CL.

chikamune@ opened up the docs linked on the CL - 

https://docs.google.com/document/d/1nm_CCuNz8Y5w4oeCVxdVvJWJIvYmAN498OrH0koMyFY/edit?resourcekey=0-6FUJvXpd758F0gu7CgNL_Q#heading=h.18tqg1je8u97

https://docs.google.com/document/d/1gLgE8orkxT4K9Ck9EYaDH2VnOlDuYt0v7CU1yjUC2T0/edit#heading=h.4nvl7mk4lsou

Let us explain the context (although you seemed to have correctly guessed it already) - yes SVGImage creation is appearing as a bottleneck on some latency sensitive websites, and we wish to optimize it.

Especially when loading small images - typically seen embedded as a data url, it seems that we are spending too much time initializing the environment for it (page / localframe / parser / etc.) compared to the actual time spent.

> Keeping the `ChromeClient` separate would go a long way from my PoV.

I don't fully follow here but am very curious. Would you elaborate?

> Latency-hiding by pooling may have benefits, but I'd be a bit wary about it when it comes to objects of this complexity (and IIRC they are not trivially small either - probably larger now then when we measured their size some [long, probably] time ago).

I don't disagree. I'm not fully convinced myself if latency hiding by this would make sense, however we were interested in taking the actual performance data / surfacing any breakages on tests.

Thanks,

kouhei