Re: regclass fix 2

[Jan Hubicka <jh at suse dot cz>]

Hi,
I've been thinking for some time about the regclass costs and updating them for
keeping the costs of union.  This don't seems to be possible.  If I write
template with two different alternative that first specifies instruction inside
first half of union class, while second in second half, I exploit the knowledge
that the union class is union class.

The regclass costs are computed as minimum over all alternatives (this is an
optimistics guess) and simply since there is no alternative mentioning the
union class, the cost is naturally high.

Of course I can once the costs is computed take a look at the classes union
class consist of and compute the maximum, but thats exactly what I do at
the end, when I construct class preferences and doing so sooner is just
waste of time.

So I tend to believe that solution I have is sane and there is no real need
to update the calculations to keep the "union property".

What do you think?
Honza