This is the mail archive of the
gcc@gcc.gnu.org
mailing list for the GCC project.
Re: Inlining and estimate_num_insns
Steven Bosscher <stevenb@suse.de> wrote:
>> In the end we surely want to watch CiSBE and SPEC testers.
>
> Maybe so, but your timings already show this is pretty unacceptable.
I strongly object this. Benchmarks show that we are doing *much* worse at
inlining in 4.0, and we are seeing bad regressions in code generation because
of that. Richard's patches are simply restoring the 3.4 behaviour and fixing
this *bad* regression. If compile time goes high, then sorry, it's because of
other things. Maybe tree-ssa, who knows. If we got good compile times because
an optimization was accidentally disabled, keeping it disabled is not the good
thing to do.
-O3 means we want to try to inline everything, and that will cost something.
Keeping the mainline broken in this regard because we save some compilation
time at -O3 is nonsense. Otherwise, please give me -O4 and I'll let you use
this broken -O3 which does nothing. And I am only half joking here. If saving
some compilation time at -O3 and losing so much in code generation is that
important to you, then please allow me to have a
flag -finline-functions-non-broken, and -O4.
Personally, I find Richard's patches so correct in fixing this big regression
which I am surprised they find so much opposition. I kindly ask you to
reconsider your position.
Giovanni Bajo