PATCH: [gcc3.5 improvement branch] Very Simple constant propagation
Caroline Tice
ctice@apple.com
Thu Jan 15 23:37:00 GMT 2004
Using "time" and measuring it on the C++ code from PR8361 (which I have
been informed is a
pretty good compile time benchmark), I get
95.750u 7.430s 1:45.27 98.0% 0+0k 56+320io 0pf+0w using gcc 3.4,
-O3 -ffast-math,
and I get
96.000u 7.340s 1:43.75 99.6% 0+0k 0+320io 0pf+0w using my patch,
adding the -fss-const-prop flag
So it appears to have a very small impact on compile time. The impact
on performance will depend a lot
on the code you use it on (although performance should only get better,
never worse). In the case that
prompted this patch, it caused the elimination of a floating point
multiplication from a loop that was
executed many times, which had a major positive impact on performance.
The tree-ssa branch currently exhibits the problem behavior that
prompted this patch. However Jan
Hubicka has suggested a possible alternative approach for tree-ssa to
take.
-- Caroline Tice
On Jan 15, 2004, at 3:14 PM, Steven Bosscher wrote:
> On Friday 16 January 2004 00:00, Caroline Tice wrote:
>> This patch has been tested on an Apple G4 running apple-darwin. It
>> bootstraps and passes DejaGnu tests. Is this okay to commit to the
>> gcc
>> 3.5 improvements branch? Or gcc 3.4?
>
> What are the compile time impacts of this pass, and what is its impact
> on the performance of the resulting code? Also, is this still relevant
> after tree-ssa hits the branch for 3.5 (assuming it does)?.
>
> Gr.
> Steven
>
More information about the Gcc-patches
mailing list