https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92621
Tobias Burnus <burnus at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |burnus at gcc dot gnu.org
--- Comment #19 from Tobias Burnus <burnus at gcc dot gnu.org> ---
Created attachment 51289
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=51289&action=edit
Patch - Free memory in the callee
The attached patch frees the memory in the callee, however:
* integer, allocatable – those are passed as 'integer **dummy'
I believe those should be passed as CFI descriptor
-> runs into the assert.
* In the Fortran caller, the freeing is not needed if the the implementation is
in Fortran as the callee frees this. In principle, something like the following
should work (trans-expr; there is one case for scalar and one for arrays):
@@ -6541 +6548,3 @@ gfc_conv_procedure_call
- && fsym->attr.intent == INTENT_OUT)
+ && fsym->attr.intent == INTENT_OUT
+ && (!sym->attr.is_bind_c
+ || sym->attr.if_source != IFSRC_DECL))
However, that does not work – and shows that the gfortran → CFI conversion is
completely broken:
void * cfi.0;
if (cfi.0 != 0B)
{
__builtin_free (cfi.0);
That's uninitialized memory :-(
* Also broken is that the data type for BIND(C) uses the Fortran array
descriptor and not the CFI one – it is currently only used as pointer and later
cast to void* but still that's pretty bad :-(
* * *
I have still to look at José's patches.
* * *
In any case, I think that the TYPE handling needs to be fixed first, followed
by moving _gfortran_gfc_desc_to_cfi_desc (and cfi→gfc) from libgfortran to the
front end to generate proper context-aware in-line code.
