[RFA] Limit stack usage growth caused by inliner
Jan Hubicka
jh@suse.cz
Tue Nov 7 08:29:00 GMT 2006
> On Nov 5, 2006, at 1:27 PM, Jan Hubicka wrote:
> >this patch is trying to address stack usage problems we run into since
> >we started inlining functions called once.
>
> I like the patch, though, I'd love to see people follow this up with
> yet more measurements and refine the 256 default number. It seems
> too small to me, but, maybe that's just me, I don't have any numbers
> to back my feeling.
Note that the stacks frames of 256 bytes (it is just space consumed by
non-register allocated stuff) are still allowed to grow tenfold, giving
us 2.5Kb stack frame first to be cut in inlining. I tried my usual
testsuite (Tramp3d, Gerald's benchmark, SPEC). For SPEC there is small
speedup on, I believe, Mesa where it is the only application actually
changing. For tramp3d there is code size decrease and no measure
difference in speed, in Gerald's application there are both ups and
downs with no significant change in geometric average, but the benchmark
behave this way to almost any change in inlining decisions.
I would love to hear feedback, definitly.
Honza
More information about the Gcc-patches
mailing list