[Bug ipa/103585] fatigue2 requires inlining of peridida to work well

[hubicka at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=103585

--- Comment #15 from Jan Hubicka <hubicka at gcc dot gnu.org> ---
We get 47s runtime with -O2 -flto and 53s with -O2
-fno-inline-functions-called-once.

The call sequence is:
  <bb 81> [local count: 109362591]:
  _1656 = (unsigned long) _45;
  _1655 = _1656 + ivtmp.1182_2540;
  _229 = (double *) _1655;
  _1646 = (unsigned long) _35;
  _1645 = _1646 + ivtmp.1182_2540;
  _230 = (double *) _1645;
  _1636 = (unsigned long) _55;
  _1635 = _1636 + ivtmp.1182_2540;
  _231 = (double *) _1635;
  _1628 = (unsigned long) _17;
  _1627 = _1628 + ivtmp.1182_2540;
  _232 = (double *) _1627;
  _1618 = (unsigned long) _64;
  _1617 = _1618 + ivtmp.1182_2540;
  _233 = (double *) _1617;
  _234 = (double *) ivtmp.1181_2551;
  _235 = (double *) ivtmp.1180_2575;
  _236 = (double *) ivtmp.1178_2586;
  _2607 = yield_stress;
  perdida.constprop.isra (&dt, &lambda, &mu, _2607, &r_infinity, &b,
&x_infinity, &gamma, &eta, &plastic_strain_threshold, _229, _230, _231, _232,
_236, _233, _235, _234, &failure_threshold, &crack_closure_parameter);

It is not clear to me why lambda is not replaced.  Howevever for dt it seems to
be:
! Disqualifying parameter number 0 - Dereferences in callers would happen much
more frequently.

I think this is too early since if we SRA all the way down to the original
caller we will avoid all dereferences completely.

Other place for improvement is non-LTO. Here IPA-sra disables itself since it
does not have cost model for cloning (that could be also improved).

Situation could be improved by ipa-modref that may optimize away unused parts
of the array descriptors. ipa-modref however gives up on the fact that perdida
has Fortran i/o and it then gives up on tracking on the descriptors even if the
descritors are never escaping to the i/o.

For this I need to finish the non-escaping analysis. I.e. make difference
between arguments that does not escape in a sense that once function return
they are not saved in global memory and in a sense that they are never passed
down to callee function.

We also may annotate fortran i/o and understand what it does.

So there is still a lot to do.