https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120413
--- Comment #5 from Tobias Burnus <burnus at gcc dot gnu.org> ---
ICE FIXED so far for mainline (GCC 16) and GCC 15.
* * *
Reading:
> There is also BIND_EXPR_VARS, dunno if that should be walked instead
> or in addition.
The current code is about adding map clauses for lambda expressions,
excluding locally defined declarations:
—————————————————————————————————————————
/* Used to walk OpenMP target directive body. */
struct omp_target_walk_data
{
...
/* For lambda functions, the __closure object expression of the current
function, and the set of captured variables accessed in target body. */
tree current_closure;
hash_set<tree> closure_vars_accessed;
/* Local variables declared inside a BIND_EXPR, used to filter out such
variables when recording lambda_objects_accessed. */
hash_set<tree> local_decls;
};
...
finish_omp_target_clauses_r (tree *tp, int *walk_subtrees, void *ptr)
{
...
if (TREE_CODE (t) == BIND_EXPR)
{
if (tree block = BIND_EXPR_BLOCK (t))
for (tree var = BLOCK_VARS (block); var; var = DECL_CHAIN (var))
if (!data->local_decls.contains (var))
data->local_decls.add (var);
#if 0 // To add?
for (tree var = BIND_EXPR_VARS (t); var; var = DECL_CHAIN (var))
if (!data->local_decls.contains (var))
data->local_decls.add (var);
#endif
return NULL_TREE;
}
...
if (TREE_TYPE (t) && LAMBDA_TYPE_P (TREE_TYPE (t)))
{
...
if (!data->lambda_objects_accessed.contains (t)
/* Do not prepare to create target maps for locally declared
lambdas or anonymous ones. */
&& !data->local_decls.contains (t)
&& TREE_CODE (t) != TARGET_EXPR)
data->lambda_objects_accessed.add (t);
...
for (hash_set<tree>::iterator i = data.lambda_objects_accessed.begin ();
i != data.lambda_objects_accessed.end (); ++i)
{
...
tree lc = build_omp_clause (loc, OMP_CLAUSE_MAP);
OMP_CLAUSE_SET_MAP_KIND (lc, GOMP_MAP_TO);
OMP_CLAUSE_DECL (lc) = lobj;
...
—————————————————————————————————————————
It seems that the only testcase for the current code is:
libgomp.c++/target-lambda-1.C
—————————————————————————————————————————
template <typename L>
void
omp_target_loop (int begin, int end, L loop)
{
#pragma omp target teams distribute parallel for
for (int i = begin; i < end; i++)
loop (i);
}
...
omp_target_loop (0, N, [=](int i) { data1[i] = val; });
omp_target_loop (0, N, [=](int i) { data2[i] = valref + 1; });
—————————————————————————————————————————
Adding some warning diagnostic shows that during initial parsing:
33 | for (int i = 0; i < len; i++)
that both BIND_EXPR_BLOCK and BIND_EXPR_VARS contain 'i'.
Then for
libgomp.c++/target-lambda-1.C: In instantiation of ‘void omp_target_loop(int,
int, L) [with L = main()::<lambda(int)>]’:
(A) The same for 'i'.
(B) 'lambda_objects_accessed.add' for 'loop'.
And likewise for the other lambda.
=> In this case, BIND_EXPR_BLOCK and BIND_EXPR_VARS contain the same variable.
For the mapping, we get for the first call:
;; Function void omp_target_loop(int, int, L) [with L = main()::<lambda(int)>]
(null)
#pragma omp target map(to:loop [len: 16]) map(alloc:*loop.__data1 [len: 0])
\
map(attach_zero_length_array_section:loop.__data1 [bias:
0])
with
int * const data1 [value-expr: __closure->__data1];
const int val [value-expr: __closure->__val];
* * *
It looks as if BIND_EXPR is primarily created via c-family/c-gimplify.cc's
tree decls, bind;
if (block == NULL_TREE)
decls = NULL_TREE;
else if (TREE_CODE (block) == BLOCK)
decls = BLOCK_VARS (block);
else
{
decls = block;
...
if (decls || block)
{
bind = build3 (BIND_EXPR, void_type_node, decls, body, block);
such that walking BLOCK_VARS shouldn't be required.
Only other use with decl != NULL:
cp/init.cc:4214: controller = build3 (BIND_EXPR, void_type_node, tbase,
NULL_TREE, NULL_TREE);
* * *
→ Hence, IMHO, the current code is fine – as would be using BLOCK_VARS.
