https://gcc.gnu.org/bugzilla/show_bug.cgi?id=71194
Tobias Burnus <burnus at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |ice-on-valid-code
Status|UNCONFIRMED |NEW
Last reconfirmed| |2016-06-20
CC| |burnus at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #1 from Tobias Burnus <burnus at gcc dot gnu.org> ---
The problem on GCC source level is in gfc_trans_pointer_assignment
(trans-expr.c):
7936 tmp = gfc_create_var (tmp, "ptrtemp");
7937 lse.descriptor_only = 0;
7938 lse.expr = tmp;
7939 lse.direct_byref = 1;
7940 gfc_conv_expr_descriptor (&lse, expr2);
where lse.expr == NULL_TREE after that call.
Looking at the dump (and the source below), w/o bound checking, a "ptrtemp" is
generated and passed to the function:
conv2real (&ptrtemp.12, D.3483);
If one traces the -fcheck=bound case, the gfc_conv_expr_descriptor call leads
to a call to gfc_conv_procedure_call, which does:
if (byref)
...
/* Add the function call to the pre chain. There is no expression. */
gfc_add_expr_to_block (&se->pre, se->expr);
se->expr = NULL_TREE;
Hence, lse.expr == NULL_TREE, which fails later on.
[Side note: If one sets lse.direct_byref to 0, it compiles as code following
immediately after the NULL_TREE sets se->expr and, hence, lse.expr again, but
as "conv2real" is called with "atmp", looking at "ptrtemp" will lead to
randomly wrong results.]
