Re: [tree-ssa] decls in tree-ssa (fwd)

[Jan Hubicka <jh at suse dot cz>]

> 
> On Monday, February 10, 2003, at 03:36  PM, Diego Novillo wrote:
> 
> >On Mon, 10 Feb 2003, Dan Nicolaescu wrote:
> >
> >>
> >>One more difference between the code generated by the tree-ssa
> >>compiler: it inlines fewer mem* functions.
> >>
> >Yes, because the estimate that relates number of trees to number
> >of RTL insns is wrong for GIMPLE.  Dan Berlin mentioned that he
> >wanted to increase it, I don't know if he ever got around to
> >doing it.
> >
> 
> No, not yet.
> I can do it when i get home.
> 
> Dan, if you want to test before then, simply change INSNS_PER_STMT in 
> tree-inline.c to 3, rather than 10.
> 
> The actual average insns per GIMPLE statement was 3.something or other 
> over all the code i ran a statistic gatherer through.

But in case you are speaking about memcpy calls, the problem is probably
the fact that i386.md expands memcpy inline only for constant arguments
that are hidden by gimplification...

Concerning the inlining limits, with the unit-at-time scheme I was
playing around with the idea of topologically sorting the CFG and then
compiling functions in order.  When function happends to have very few
RTL insns, I can mark it inline.  When it's address is not taken, I
think the compilation can be easilly canceled.  That can give better
estimate and the slowdown won't be too bad as we do extra compilation
only of the functions that are small (perhaps...)

Honza
> 
> 
> >Diego.