Re: [PATCH] [4.0] Fix performance regressions due to inlining

[Richard Guenther <richard dot guenther at gmail dot com>]

On Fri, 4 Mar 2005 13:12:13 +0100, Richard Guenther
<richard.guenther@gmail.com> wrote:
> I'll followup myself with short rationales for the individual patches:

And updated estimated-insns vs. real-code-size graphs for whole gcc
build for 3.4,
4.0 (40b) and 4.0 patched (40a5) with set #2:

http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/gcc-sizes-34.png
http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/gcc-sizes-40b.png
http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/gcc-sizes-40a5.png

after applying patch-set #2 we get less scattering, overall lower
estimate and for
small functions better linearity.  Raw data used for plotting are at
*.txt.  Plotted
is log-log of columns 3 vs. 2.  Estimated insns are those after
inlining (naturally).

Note that these new numbers suggest we could lower our inlining limits
to at most
50% of those of gcc 3.4.  At least the present lowering by 10% will surely _not_
lead to less inlining compared to 3.4.  Even 20% should be pretty safe.

http://www.tat.physik.uni-tuebingen.de/~rguenth/gcc/estimated-sizes.txt

contains data for a set of functions out of gcc with the size
estimates (before inlining)
from 3.4, 4.0 and 4.0 patched with set #2.


Richard.