Bug 90151
| Summary: | Element::attributeChanged could be devirtualized | ||
|---|---|---|---|
| Product: | WebKit | Reporter: | Hajime Morrita <morrita> |
| Component: | DOM | Assignee: | Nobody <webkit-unassigned> |
| Status: | RESOLVED WONTFIX | ||
| Severity: | Normal | CC: | haraken, rniwa |
| Priority: | P2 | ||
| Version: | 528+ (Nightly build) | ||
| Hardware: | Unspecified | ||
| OS: | Unspecified | ||
Hajime Morrita
There are only two overrides (StyledElement and SVGElement) thus easily devirtualized.
| Attachments | ||
|---|---|---|
| Add attachment proposed patch, testcase, etc. |
Hajime Morrita
The benchmark didn't show any speedup. closing.
Kentaro Hara
I've tried a couple of devirtualization, but I couldn't observe any performance improvement. It looks difficult to get a perf win by devirtualization.
Ryosuke Niwa
I think we should devirtualize functions if we can. The gain is small but if small improvements add up. The corollary of this is that virtualizing functions may not cause a significant performance regression by itself but will eventually add up.
Ryosuke Niwa
Also note that most modern x86 processors do indirect branch predictions. If we really wanted to see the perf. improvements for devirtualization, we should be testing on ARM and architectures where such isn't implemented.
Hajime Morrita
This is off-topic but having ARM perf bot will be huge positive leap for us.