This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
loop optimizations, mark 12
- To: egcs-patches at cygnus dot com
- Subject: loop optimizations, mark 12
- From: Richard Henderson <rth at cygnus dot com>
- Date: Thu, 25 Jun 1998 08:46:49 -0700
- Reply-To: Richard Henderson <rth at cygnus dot com>
Two things of note in this revision:
* PA problem should be fixed.
* The expected benefit of any one giv is updated after combinations
are made. We re-sort based on the adjustment and continue. This
helps cases like
[base+0] [base+0x2000] [base+0x4000] [base+0x6000] [base+0x8000]
[base+0xa000] [base+0xe000] [base+0x10000] [base+0x12000]
[base+0x14000] [base+0x16000]
combine to
[g1-0x8000] [g1-0x6000] [g1-0x4000] [g1-0x2000] [g1+0]
[g1+0x2000] [g1+0x4000] [g1+0x6000]
[g2+0] [g2+0x2000] [g2+0x4000]
where before the last three wound up not being combined due
to the illusionary large benefit expected for one of the
intermediate givs.
The benefit ought to be more noticable on machines with a
narrower range of offsets -- pa fp comes to mind.
r~
d-loop-12.gz