[tree-ssa] Preliminary analysis on POOMA slow downs (was Re: [tree-ssa] Speed up constant propagation)

[Diego Novillo]

[ moved from gcc-patches@ ]

On Tue, 2003-06-17 at 07:38, Richard Guenther wrote:

> Its src/Field/tests/Cartesian.cpp which you dont have, because thats in my
> local tree ;) A preprocessed file is at
>
> http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/Cartesian.ii.gz
> 
Thanks Richard.  There are a few interesting things in this program.

It exposes quite clearly the performance problems with type-based
aliasing.  If you compile the test with -ftree-points-to=andersen you
should see ~18% speed up in compilation time.  The type-based analyzer
was creating so many alias relations that it creates a lot more work for
the SSA passes:

			Type-based aliasing	Points-to analysis

 tree alias analysis   :  26.63 (10%) usr       19.03 ( 9%) usr
 tree PHI insertion    :  47.80 (18%) usr       15.38 ( 7%) usr
 tree SSA rewrite      :   6.56 ( 2%) usr        2.15 ( 1%) usr
 tree SSA other        :   4.97 ( 2%) usr        4.46 ( 2%) usr
 tree SSA to normal    :   7.15 ( 3%) usr        3.28 ( 1%) usr
 tree CCP              :  10.17 ( 4%) usr        2.84 ( 1%) usr
 tree COPYPROP         :   3.57 ( 1%) usr        3.20 ( 1%) usr
 tree DCE              :   2.51 ( 1%) usr        1.22 ( 1%) usr

 total SSA time        : 109.36 (40%) usr       51,56 (23%) usr
 total compile time    : 268.80       usr      220.32       usr

I'm playing with the idea of removing the alias analysis we do by
default.  Chris Lattner mentioned it at the Symposium and it's
definitely something worth trying.  The fact is that currently we don't
use aliasing information for anything but removing some additional dead
code in some cases.  It may have other repercussions, so I need to give
it some more thought first.

The program also exposes some other differences.  Here, I'm going to
compare tree-ssa with PTA enabled vs mainline.  The listing only
includes passes common with mainline where tree-ssa is noticeably
different, so the TOTAL line doesn't add up (all times are USR secs):

                          mainline   tree-ssa   % slowdown

varconst              :   0.12         9.30        7650%
global CSE            :   2.48        29.65        1095%
integration           :   2.93        22.00         650%
global alloc          :   3.40        19.31         468%
garbage collection    :   1.70         7.65         350%
scheduling 2          :   1.22         3.11         155%
loop analysis         :   3.04         6.41         110%
parser                :   7.00        10.43          49%
name lookup           :   6.28         7.52          20%
branch prediction     :   4.08         2.49         -39%
jump                  :   3.62         1.94         -46%
reload CSE regs       :   1.92         0.89         -53%
local alloc           :   6.81         3.12         -54%
expand                :  15.06         5.18         -65%
combiner              :   4.59         1.04         -77%
TOTAL                 :  92.60       220.32         138%

I'm a bit surprised that the inliner is taking such a big hit.  What
exactly is the varconst pass?  We are giving more work to the garbage
collector, not entirely surprising.  I still haven't examined the other
passes.

The results are a mixed bag.  The SSA passes seem to be helping a few
RTL passes, particularly combine, but we get such a tremendous hit from
the other passes that the end result is more than 2 times slower.

Anyway, lots of things to do.  Clearly.


Diego.