[Issue 9584] Exceptions in D are ludicrously slow (far worse than Java)

[d-bugmail at puremagic.com]

https://d.puremagic.com/issues/show_bug.cgi?id=9584



--- Comment #20 from Adam D. Ruppe <destructionator at gmail.com> 2014-02-07 20:10:53 PST ---
My pull request there is posix only. A win but not complete without looking at
Windows too.

And I'm not sure I understand the Windows stack trace code. There must be
something lazy initialized in there since the first exception handled is
significantly slower than any subsequent ones, with or without the trace
handler. At first I thought it was loading the debug help library, but that
seems to be done in a static module constructor. Maybe it is deferred anyway,
idk.

But anywho with the trace handler turned on, the second exception clocks in at
about 250 us. With the trace handler turned off, the second exception comes in
at about 30, same ballpark as linux after the change. So we are still looking
at a huge potential improvement here. But I'm not sure where and my windows dev
environment sucks. Maybe we can move the call to m_trace = trace out, into
opApply too, just storing the CONTEXT*. Could work, but then again, the trace
method says it returns a thing to be used by resolve later on, so maybe it is
important to be done here.

idk, gotta experiment with it and I'm out of time tonight. At least the Linux
change seems to be a big success so far.

BTW I can't help but wonder if we can't shave a few more useconds off by
getting the trace info objects from a small reusable pool instead of doing
`new` each time we throw an exception. I think we should get a small, cost-free
(from the user perspective) win there too.

-- 
Configure issuemail: https://d.puremagic.com/issues/userprefs.cgi?tab=email
------- You are receiving this mail because: -------