[Bug middle-end/79720] floating point result different at compile time / runtime

[tkoenig at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79720

Thomas Koenig <tkoenig at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
            Summary|floating point result       |floating point result
                   |depends on optimization     |different at compile time /
                   |level                       |runtime

--- Comment #2 from Thomas Koenig <tkoenig at gcc dot gnu.org> ---
Ah... constant folding happens with -O:

From the .optimized dump:

  a1 = __complex__ (8.3333333289470346016969415359199047088623046875e-11,
2.22222233121084274117421693446605690525075260666199028492e-20);
  a2 = __complex__ (8.3333333289470346016969415359199047088623046875e-11,
2.22222233121084274117421693446605690525075260666199028492e-20);

If constant folding is suppressed by hiding the opportunities to do so via
-fno-inline, this happens:

$ gfortran -O recip-orig.f90 && ./a.out
   (8.333333329E-11,2.222222331E-20)   (8.333333329E-11,2.222222331E-20)
             (0.00000000,0.00000000)
$ gfortran -fno-inline -O recip-orig.f90 && ./a.out
   (8.333333329E-11,2.222222170E-20)   (8.333333329E-11,2.222222331E-20)
       (0.00000000,-1.615587134E-27)
$ gfortran  recip-orig.f90 && ./a.out
   (8.333333329E-11,2.222222170E-20)   (8.333333329E-11,2.222222331E-20)
       (0.00000000,-1.615587134E-27)

-ffp-contract has no effect:

$ gfortran -fno-inline -ffp-contract=off -O recip-orig.f90 && ./a.out
   (8.333333329E-11,2.222222170E-20)   (8.333333329E-11,2.222222331E-20)
       (0.00000000,-1.615587134E-27)
$ gfortran -ffp-contract=off -O recip-orig.f90 && ./a.out
   (8.333333329E-11,2.222222331E-20)   (8.333333329E-11,2.222222331E-20)
             (0.00000000,0.00000000)

So, the question becomes why the contant folding gets a different
result from the runtime evaluation.