Re: egcs 1.1 performance report

[Richard Henderson <rth at cygnus dot com>]

On Tue, Aug 25, 1998 at 12:45:07PM -0700, Joe Buck wrote:
> >    145.fpppp         9600       767     12.5       1075      8.93
> 
> Have you looked to see why the code is worse?

Yes.  Haifa sched1 extends the register lifetimes so much that
we end up with 2k more spills in subroutine fpppp.

DaveM and Jeff and I have been discussing the problem of late.
It is not an easy problem to solve.  Dave has started looking
into a new register allocator and in the process has come up
with some ideas for solving these related problems.

It'll be a while before this is adequately addressed, but note
that fpppp is a rather exceptional case -- a function of 10k
insns all in one basic block.


r~