On Mon, Jul 14, 2025 at 05:01:40PM -0400, Bradley Lucier wrote:
> On 7/14/25 16:30, Segher Boessenkool wrote:
> > PRE is run whenever you use -fgcse, to a first order approximation 🙂
> > The biggest contraindication is functions that call setjmp, those will
> > never get PRE done.
>
> On both aarch64 and x86_64 -fno-gcse flag is given as a gcc command line
> argument; the typical call is
Oh! Then you do not get PRE at all of course.
bool
pass_rtl_pre::gate (function *fun)
{
return optimize > 0 && flag_gcse
&& !fun->calls_setjmp
&& optimize_function_for_speed_p (fun)
&& dbg_cnt (pre);
}
No -fgcse -> no PRE.
> /pkgs/gcc-15.1.0/bin/gcc -save-temps -O1 -fexpensive-optimizations -fno-gcse
> --param max-gcse-memory=400000 -Wno-unused -Wno-write-strings
-fno-gcse is the default at -O1 anyway. You need -O2 to get reasonable
optimisation.
> -Wdisabled-optimization -fwrapv -fno-strict-aliasing -fno-trapping-math
-fwrapv is a great way to get slower code, too. Is there something in
your code that does not work without this reality-distorting flag?
Similar for -fno-strict-aliasing.
> So maybe with this combination of flags, PRE is run on aarch64 and not on
> x86_64, and there's no mystery.
It should not run on *any* arch with -fno-gcse, and that is the default
at -O1 anyway!
Segher
