On Thu, Nov 11, 2021 at 10:45 AM Jan Hubicka via Gcc-patches
<gcc-patc...@gcc.gnu.org> wrote:
>
> > > >
> > > > I think the patch causes the following on x86_64-linux-gnu:
> > > > FAIL: gfortran.dg/inline_matmul_17.f90 -O scan-tree-dump-times
> > > > optimized "matmul_r4" 2
> > >
> > > I get that failure even with d70ef65692f (from before the patches
> > > I committed today).
> >
> > Sorry, you are right, it's one revision before:
> > d70ef65692fced7ab72e0aceeff7407e5a34d96d
> >
> > Honza, can you please take a look?
> The test looks for matmul_r4 calls which we now optimize out in fre1.
> This is because alias info is better now
>
> afunc (&__var_5_mma);
> _188 = __var_5_mma.dim[0].ubound;
> _189 = __var_5_mma.dim[0].lbound;
> _190 = _188 - _189;
> _191 = _190 + 1;
> _192 = MAX_EXPR <_191, 0>;
> _193 = (real(kind=4)) _192;
> _194 = __var_5_mma.dim[1].ubound;
> _195 = __var_5_mma.dim[1].lbound;
> _196 = _194 - _195;
> _197 = _196 + 1;
> _198 = MAX_EXPR <_197, 0>;
> _199 = (real(kind=4)) _198;
> _200 = _193 * _199;
> _201 = _200 * 3.0e+0;
> if (_201 <= 1.0e+9)
> goto <bb 76>; [INV]
> else
> goto <bb 87>; [INV]
> <bb 76> :
> c = {};
>
>
> afunc (&__var_5_mma);
> c = {};
>
> Now afunc writes to __var_5_mma only indirectly so I think it is correct that
> we optimize the conditional out.
>
> Easy fix would be to add -fno-ipa-modref, but perhaps someone with
> better understanding of Fortran would help me to improve the testcase so
> the calls to matmul_r4 remains reachable?
I think the two matmul_r4 cases were missed optimizations before so just
changing the expected number of calls to zero is the correct fix here. Indeed
we can now statically determine the matrices are not large and so only
keep the inline copy.
Richard.
>
> Honza
