https://gcc.gnu.org/bugzilla/show_bug.cgi?id=106089
kargl at gcc dot gnu.org changed:
What |Removed |Added
----------------------------------------------------------------------------
Status|UNCONFIRMED |NEW
CC| |kargl at gcc dot gnu.org
Last reconfirmed| |2022-07-02
Ever confirmed|0 |1
Priority|P3 |P4
Keywords|diagnostic |
--- Comment #3 from kargl at gcc dot gnu.org ---
Removed 'diagnostic' keyword. This is much worse than just a bad diagnostic.
Consider a much shorter testcase.
program foo
integer, allocatable :: i(:)
i = [1,1]
if (any(i /= 1)) stop
end program foo
% gfortran11 -Wall -c a.f90 |& grep -i warn
Warning: 'i.offset' is used uninitialized [-Wuninitialized]
Warning: 'i.dim[0].lbound' is used uninitialized [-Wuninitialized]
Warning: 'i.dim[0].ubound' is used uninitialized [-Wuninitialized]
Warning: 'i.dim[0].lbound' may be used uninitialized [-Wmaybe-uninitialized]
Warning: 'i.dim[0].ubound' may be used uninitialized [-Wmaybe-uninitialized]
Warning: 'i.dim[0].ubound' may be used uninitialized [-Wmaybe-uninitialized]
Warning: 'i.dim[0].lbound' may be used uninitialized [-Wmaybe-uninitialized]
With the patch that follows, I get
% gfcx -Wall -c a.f90 |& grep -i warn
Warning: 'i.offset' is used uninitialized [-Wuninitialized]
Warning: 'i.dim[0].lbound' is used uninitialized [-Wuninitialized]
Warning: 'i.dim[0].ubound' is used uninitialized [-Wuninitialized]
so 4 of the rogue warnings are no longer issued.
diff --git a/gcc/fortran/trans-array.cc b/gcc/fortran/trans-array.cc
index 05134952db4..793e6a21e6d 100644
--- a/gcc/fortran/trans-array.cc
+++ b/gcc/fortran/trans-array.cc
@@ -10734,27 +10734,27 @@ gfc_alloc_allocatable_for_assignment (gfc_loopinfo
*loop,
/* If the lhs shape is not the same as the rhs jump to setting the
bounds and doing the reallocation....... */
- for (n = 0; n < expr1->rank; n++)
+ if (expr1->shape)
{
- /* Check the shape. */
- lbound = gfc_conv_descriptor_lbound_get (desc, gfc_rank_cst[n]);
- ubound = gfc_conv_descriptor_ubound_get (desc, gfc_rank_cst[n]);
- tmp = fold_build2_loc (input_location, MINUS_EXPR,
- gfc_array_index_type,
- loop->to[n], loop->from[n]);
- tmp = fold_build2_loc (input_location, PLUS_EXPR,
- gfc_array_index_type,
- tmp, lbound);
- tmp = fold_build2_loc (input_location, MINUS_EXPR,
- gfc_array_index_type,
- tmp, ubound);
- cond = fold_build2_loc (input_location, NE_EXPR,
- logical_type_node,
- tmp, gfc_index_zero_node);
- tmp = build3_v (COND_EXPR, cond,
- build1_v (GOTO_EXPR, jump_label1),
- build_empty_stmt (input_location));
- gfc_add_expr_to_block (&fblock, tmp);
+ for (n = 0; n < expr1->rank; n++)
+ {
+ /* Check the shape. */
+ lbound = gfc_conv_descriptor_lbound_get (desc, gfc_rank_cst[n]);
+ ubound = gfc_conv_descriptor_ubound_get (desc, gfc_rank_cst[n]);
+ tmp = fold_build2_loc (input_location, MINUS_EXPR,
+ gfc_array_index_type,
+ loop->to[n], loop->from[n]);
+ tmp = fold_build2_loc (input_location, PLUS_EXPR,
+ gfc_array_index_type, tmp, lbound);
+ tmp = fold_build2_loc (input_location, MINUS_EXPR,
+ gfc_array_index_type, tmp, ubound);
+ cond = fold_build2_loc (input_location, NE_EXPR,
+ logical_type_node, tmp, gfc_index_zero_node);
+ tmp = build3_v (COND_EXPR, cond,
+ build1_v (GOTO_EXPR, jump_label1),
+ build_empty_stmt (input_location));
+ gfc_add_expr_to_block (&fblock, tmp);
+ }
}
/* ...else if the element lengths are not the same also go to
So, what's the problem. The scalarizer is broken, which was
originally written before (re)allocation of the LHS was introduced
to Fortran. The above is not valid Fortran 95. The LHS and RHS
of the assignment must be conformable. This means that, if an
error had not been emitted, then the LHS and RHS have the same
array descriptor. This leads to two problems. The first is
the scalarizer appears to write the descriptors into the intermediate
representation (IR).
% gfortran11 -Wall -c a.f90 -fdump-tree-original
% more a.f90.005t.original (NOte inline annotations)
void foo ()
{
struct array01_integer(kind=4) i;
i.data = 0B;
i.dtype = {.elem_len=4, .rank=1, .type=1};
{
*** This is the descriptor for the RHS
integer(kind=4)[0:] * restrict D.3943;
integer(kind=8) D.3944;
integer(kind=8) D.3945;
integer(kind=8) D.3946;
static integer(kind=4) A.0[2] = {1, 1};
integer(kind=8) D.3949;
*** end RHS
*** This is the descriptor for the LHS
D.3943 = (integer(kind=4)[0:] * restrict) i.data;
D.3944 = i.offset;
D.3945 = i.dim[0].lbound;
D.3946 = i.dim[0].ubound;
*** end LHS
D.3943 is a validate expression. D.3944, D.3945, and D.3946
are the cause of 3 warnings; and in fact, these are the
three that remain after my patch above is applied. Due to
the complexity of scalarization, I have been unable to find
where the code is added to the IR.
Now, when (re)allocation on assignment was introduced, it seems
the checking assumed that LHS had already been allocated (i.e.,
it has a proper descriptor). So, the descriptor for the LHS was
used to build the information need to determine if reallocation
was need. Fortunately, an initially unallocated allocated has its
expr->shape pointer set to NULL. I use this in the above patch
to block the use of the LHS descriptor.
