[Webkit-unassigned] [Bug 74170] DFG's interpretation of rare case profiles should be frequency-based not count-based

[bugzilla-daemon at webkit.org]

https://bugs.webkit.org/show_bug.cgi?id=74170





--- Comment #4 from Filip Pizlo <fpizlo at apple.com>  2011-12-09 13:24:56 PST ---
It turns out that this is necessary, but insufficient to precisely catch operations that often take slow path.  The early phase of the program's execution is often not representative of how the program will behave in steady-state.  Hence if we try to make an operation safe by switching it to use a double addition instead of an integer addition, for example, just because 1/2 of the initial executions took slow path then we may run the risk that in steady state that number will actually be <1/10.

So we need a rule like:

Make the operation safe only when:
(It has taken slow path some minimum number of times (probably close to 100),
 and It has taken slow path more than some fraction of total executions (probably 1/10))
or The speculation failures that have led to recompilation being triggered were caused by this operation not being safe.

In this bug, I'll only solve the first two parts, and fix the last part in another bug.

-- 
Configure bugmail: https://bugs.webkit.org/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are the assignee for the bug.