Re: [PATCH] Teach VRP to derive ranges from SWITCH_EXPR (fix PR21258)

[Diego Novillo <dnovillo at redhat dot com>]

Richard Guenther wrote on 04/12/07 11:26:

> Well, I did show it does a lot on tree.c and I would expect it do
> even more on fold-const.c.  Let's see what it does in it's current
> limited form on the C++ testcases.

Run it over the usual suspects, please.  I want to see exactly what we
are getting into at what cost.  Perhaps we want to enable this for -Os,
for instance, not -O2.

> Well, I suppose that even happens more rare than producing useful
> ASSERT_EXPRs.  But I'll implement it just to check.  What kind of
> difference would it make for the propagator to know the executed edge?
> (other than maybe not simulating the non-executable ones)

That's a huge difference.  The more edges you discard, the faster you'll
converge.