[Bug fortran/50981] [4.4/4.5/4.6/4.7 Regression] Wrong-code for scalarizing ELEMENTAL call with absent OPTIONAL argument

[burnus at gcc dot gnu.org]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50981

Tobias Burnus <burnus at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
   Target Milestone|---                         |4.4.7

--- Comment #2 from Tobias Burnus <burnus at gcc dot gnu.org> 2011-11-03 17:12:15 UTC ---
(In reply to comment #1)
> No bootstraped compiler at hand; does this work?

Does not help. (Nit: in trans-array.c "se.want..." instead of "se->".)

Now the segfaults already for "ff" where the "b" argument is present. The
segfault occurs at
    ELEMENTAL REAL(KIND=8) FUNCTION gg(a,b)
      ...
      IF(PRESENT(b)) THEN
        b1=b  ! <<< here

While "ff" now looks reasonable:

ff (real(kind=8) & restrict a, integer(kind=4) * b)
    integer(kind=4) * D.1747;
    D.1747 = b;
          val.0 = gg (&ac[S.1 + -1], D.1747) + val.0;

one has an issue with the call:
        rr[S.5 + -1] = ff (&aa[S.5 + -1], 1);

as it is probably not a good idea to dereference the literal "1", unless you
really want to access the memory address 0x1 ...

 * * *

Works:
   4.4.5 20100627 (prerelease) [gcc-4_4-branch revision 161471]
Fails:
   4.4.5 20100627 (prerelease) [gcc-4_4-branch revision 161472]

Unless I made a mistake with building those versions, the regression is caused
by:

Rev. 161472 (PR fortran/43841 and PR 43843).
http://gcc.gnu.org/viewcvs?root=gcc&view=rev&rev=161472

        * trans-expr.c (gfc_conv_expr): Supply an address expression for
        GFC_SS_REFERENCE.
        (gfc_conv_expr_reference): Call gfc_conv_expr and return for
        GFC_SS_REFERENCE.
        * trans-array.c (gfc_add_loop_ss_code): Store the value rather
        than the address of a GFC_SS_REFERENCE.
        * trans.h : Change comment on GFC_SS_REFERENCE.