Recently, a few UPC test programs failed to compile
due to mis-matches of parameters in a prototype and its
corresponding function definition. The mis-match was
based upon the apparent inequality of UPC layout qualifiers
(blocking factors).
UPC blocking factors are integer constants. They are
recorded in a hash table indexed by the type tree node
that they correspond to.
Currently, the test for equality of blocking factors
tests only the pointer to the tree node defining the constant.
All blocking factors are recorded as "sizetype" type'd nodes.
Given that integer constants are hashed by type/value, it seemed
safe to assume that a given blocking factor would map to
a single tree node due to the underlying hash method that is used
when integral constants are created.
Is it valid to assume that pointer equality is sufficient
to ensure that two integer constants are equal as long
as their type and values are equal?
The bug that we ran into occurred because a garbage collection
pass was run between the point that the function prototype
tree node was created and the point at which the function declaration
was processed. The garbage collector decided that the integer
constant representing the blocking factor was no longer in use,
because it had not been marked.
In fact, the integer constant was needed because it appeared
in the blocking factor hash table, but not via a direct pointer.
Rather it was referenced by nature of the fact that
the blocking factor hash table referenced the integer constant
that is mapped in the integer constant hash table.
Here's a rough diagram:
tree (type) -> [ blocking factor hash ] -> tree (integer constant)
tree (integer constant) -> [ integer constant hash ] {unique map}
-> tree (integer constant)
When the garbage collector deleted the entry from the
"integer constant hash", it forced a new integer constant tree
node to be created for the same (type, value) integral constant
blocking factor.
One easy way to address the current issue is to call
tree_int_cst_equal() if the integer constant tree pointers
do not match:
if ((c1 != c2) && !tree_int_cst_equal (c1, c2))
/* integer constants aren't equal. */
This may be necessary if 'int_cst_hash_table' is viewed as
a cache rather than a persistent, unique mapping.
Another approach, would be to somehow mark the node
in int_cst_hash_table as in use when the blocking factor
hash table is traversed by the garbage collector, or
to add logic the hash table delete function associated
with int_cst_hash_table; to dis-allow the delete if the
integer constant is present in the UPC blocking factor
hash table.
To effect this change in a modular way probably the hash table
delete function associated with 'int_cst_hash_table' would have
to be daisy-chained, where the UPC blocking factor check is made
first. The difficulty with implementing the daisy chaining is that
int_cst_hash_table needs to exist before the UPC-related initialization
code is run. One way to handle this might be yet another language
hook, called from the code that creates 'int_cst_hash_table'.
That seems overly complex.
For reference, the current blocking factor mapping table
is created as follows:
upc_block_factor_for_type = htab_create_ggc (512, tree_map_hash,
tree_map_eq, 0);
Summary:
1. Is it valid to assume that pointer equality is sufficient
to compare two integer constants for equality as long as they
have identical type and value?
2. Should 'int_cst_hash_table' be viewed as a cache, where
the mapping of a given (type, value) integer constant may
vary over time?
3. If the answer to 1. is "yes" and the answer to 2. is "no"
then what is the recommended way to ensure that nodes in
'int_cst_hash_table' are not removed if the integer constant
is being referenced via the 'upc_block_factor_for_type'
hash table?
thanks,
- Gary
