loop optimizations, mark 9

[Richard Henderson <rth at cygnus dot com>]

I'd left off this patch set for some time, and was reminded of it
while running spec95 last night testing other optimizations.

This is version 9 of the patch; previous versions can be found in
the mailing list archives.

Highlights of this patch are:

  * Tweeks to fold() to handle some of the "interesting" output from
    loop, via expand_mult_add, when expanding the giv increments.

  * A tweek to combine_givs to prefer (r+0),(r+8) over (b+r1),(r2+0),
    as would tend to happen on hosts that support reg+reg addressing.

The result is that, on our good friend RESID, I believe the resulting
code is as good as it can possibly get without GCSE (which ought to 
elimitate two redundancies that normal CSE cannot fix up).


r~

d-loop-9.gz