[tuples] New memory/time comparison vs trunk

["Diego Novillo" <dnovillo at google dot com>]

The branch has regressed in terms of memory and compile time.
However, I think the regressions are localized.

I compiled tuples and trunk@137837 with --disable-checking
--enable-gather-detailed-mem-stats.  I ran cc1 and cc1plus over the
preprocessed code of cc1-i-files, FF3D, SPEC2000, TRAMP3D, PR 12850,
PR 15855 and PR 30089.

I had to disable PRE and FRE on both trunk and tuples because trunk
wouldn't bootstrap if I only disabled the same bits that are disabled
in tuples.

The attached is a comparison using the compile and memory figures
emitted by -ftime-report.  The "Before" column corresponds to
trunk@137837, the "After" column corresponds to tuples@138062.

The times and memory figures are aggregated over ALL the files in the
collection I described above.

Some preliminary notes:

- The branch is 25secs slower than mainline (2%).  Out of those 25
seconds, 15.31 seconds are in tree DSE, which is puzzling.  The other
10 seconds seem to be going to RTL DSE1 (3 secs), tree CCP (2 secs)
and various others more evenly distributed.  In fact, if you take away
tree DSE, the tree passes are faster on tuples.

- Memory utilization is tricky to check.  The difference is almost
totally going to the gimplifier and RTL expansion (about 900Kb of the
1.3Mb difference).  These two have trees and gimple coexisting for a
short period of time, particularly in the gimplifier.  However, the
trees that are used in RTL expansion are built and destroyed on the
spot, so that should not affect peak memory use, but the report
doesn't let me distinguish that.

- The rest of the memory utilization difference is mostly in inlining
(240Kb) and SSA update (50Kb).

I think the main focus points should be DSE and trying to get a good
way of measuring the memory utilization differences.  Jan, any
suggestion?


Thanks.  Diego.

Attachment: 20080722-trunk-vs-tuples-timings.txt
Description: Text document