[Bug tree-optimization/92283] [10 Regression] 454.calculix miscomparison since r276645 with -O2 -march=znver2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92283

--- Comment #10 from Richard Biener <rguenth at gcc dot gnu.org> ---
Adding

!GCC$ unroll 0

before line 848 or adding a call to an empty function after the loop nest
(after 857) fixes the miscompare.  GIMPLE level difference with the function
call
is one missed invariant motion of a load from a function parameter (iperturb)
but assembly level diff is ~7000 lines :/  The iperturb change can be
compensated
for in the source which still fixes the miscompare (but assembly level
difference
doesn't improve much).

The odd thing is that the differences aren't _too_ big so [stack] memory
corruption seems unlikely(?).