This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: [tree-ssa] POOMA compile time / memory requirement comparison
- From: law at redhat dot com
- To: Richard Guenther <rguenth at tat dot physik dot uni-tuebingen dot de>
- Cc: Steven Bosscher <stevenb at suse dot de>, Mark Mitchell <mark at codesourcery dot com>, Diego Novillo <dnovillo at redhat dot com>, "gcc at gcc dot gnu dot org" <gcc at gcc dot gnu dot org>
- Date: Mon, 17 May 2004 13:06:44 -0600
- Subject: Re: [tree-ssa] POOMA compile time / memory requirement comparison
- Reply-to: law at redhat dot com
In message <40994706.9020902@tat.physik.uni-tuebingen.de>, Richard Guenther wri
tes:
>Steven Bosscher wrote:
>> On Wednesday 05 May 2004 20:41, Richard Guenther wrote:
>>
>>>Using the suggested -fno-gcse -fno-cse-follow-jumps -fno-cse-skip-blocks
>>>-fno-web, -O2 compilation times improve about 8%. This leaves us with
>>
>>
>> Why not -fno-delete-null-pointer-checks? And -fweb is enabled on mainline
>> and helps tree-ssa iirc, so that should really stay enabled...
>
>-fno-delete-null-pointer-checks doesn't help. It seems to be cheap.
The old RTL version could suck up around 1% on particularly ugly code. We
had limiters which basically turned it off if it looked like it might get
expensive.
It also tried to be smart in that it first identified all the non-null tests
and the registers used in those tests. That allowed it to greatly reduce
the size of the vectors it used during its global propagation phase.
>Yes. I'm confident we're going to gain more than the current 20%
>regression. It just needs people remind others about the regressions.
Yes. I'd like to see us working on having the front-ends start producing
generic directly. We're wasting a fair amount of memory creating front-end
specific nodes, then wasting time converting those into generic/gimple.
jeff