https://gcc.gnu.org/bugzilla/show_bug.cgi?id=108621
--- Comment #5 from Tobias Burnus <burnus at gcc dot gnu.org> ---
The warning itself is bogus (false positive in the middle end).
I get:
Warning: ‘f.dim[idx.1_32].lbound’ may be used uninitialized
[-Wmaybe-uninitialized]
If I now look at the 021t.ssa dump, I see:
f.data = 0B;
...
_1 = f.data;
cfi.0.base_addr = _1;
...
_2 = cfi.0.base_addr;
if (_2 != 0B)
goto <bb 3>; [INV]
else
goto <bb 6>; [INV]
...
<bb 4> :
_3 = f.dim[idx.1_32].lbound;
...
<bb 6> :
fun (&cfi.0);
The basic block <bb 4> is the only place where idx.1_32 gets used - but as
f.data == NULL, we directly jump to <bb 6>. Seemingly, the range/value
propagation from f.data to cfi.0.base_addr did not work. — Thus and probably
unsurprisingly, the warning is gone once optimization has been turned on (-Og,
-Os or -O1 are sufficient).
→ This is now tracked in the new PR middle-end/108906
* * *
Still, we can do better in the FE by producing less code when we know that the
dummy argument is 'intent(out)'. That's what the lightly tested patch below
does.
(Note: It fixes the testcase of comment 0, but when changing the intent, e.g.,
to 'intent(inout)' the bogus warning will re-appear.)
TODO: I think something similar needs to be done for allocatable +
'intent(out)', except that we still need to handle the DEALLOCATE in the
caller.
--- a/gcc/fortran/trans-expr.cc
+++ b/gcc/fortran/trans-expr.cc
@@ -5675,3 +5675,3 @@ gfc_conv_gfc_desc_to_cfi_desc (gfc_se *parmse, gfc_expr
*e, gfc_symbol *fsym)
- if (e->rank == 0)
+ if (e->rank == 0 && (!fsym->attr.pointer || fsym->attr.intent !=
INTENT_OUT))
{
@@ -5680,3 +5680,3 @@ gfc_conv_gfc_desc_to_cfi_desc (gfc_se *parmse, gfc_expr
*e, gfc_symbol *fsym)
}
- else
+ else if (!fsym->attr.pointer || fsym->attr.intent != INTENT_OUT)
{
@@ -5697,2 +5697,5 @@ gfc_conv_gfc_desc_to_cfi_desc (gfc_se *parmse, gfc_expr
*e, gfc_symbol *fsym)
+ if (fsym->attr.pointer && fsym->attr.intent == INTENT_OUT)
+ goto done;
+
/* When allocatable + intent out, free the cfi descriptor. */
