On 2026-01-28 18:48, Andrew Pinski wrote:
+/* In the late (waccess3) pass run, addresses to union members may have been
+ replaced to optimize accesses. This may result in waccess seeing the wrong
+ union member and come to a wrong conclusion about its size. For now, just
+ bail out when we see the possibility of such a situation. This could in
+ future walk through the different types and see if there's a union member at
+ the same address that matches the size of the access. */
+static inline bool
+aliasing_union_addr_p (tree ptr, bool early)
+{
+ if (early)
+ return false;
+
+ /* Thread through the chain of definitions to arrive at an address
+ expression. */
Isn't at this point one look back enough to get the ADDR_EXPR? Or
there cases where more than one is needed?
Do you have a testcase for those included?
I was trying to be conservative but you're right, for this case there
need only be one look back.
+ while (ptr && TREE_CODE (ptr) == SSA_NAME)
+ {
+ gimple *stmt = SSA_NAME_DEF_STMT (ptr);
+ if (gimple_code (stmt) == GIMPLE_ASSIGN)
if (is_a <gassign *> (stmt))
+ ptr = gimple_assign_rhs1 (stmt);
This brings up we looking past a in "a + b". I am not sure that is
correct; yes we are looking for the ADDR_EXPR here though.
Maybe gimple_assign_single_p is better than looking at all assigns.
Agreed.
+ else
+ ptr = NULL;
Why not just return false here?
Indeed, yes.
+ }
+
+ if (ptr && TREE_CODE (ptr) == ADDR_EXPR)
And then this becomes:
if (TREE_CODE (ptr) != ADDR_EXPR)
return false;
Ack, I'll update and send v2.
I think overall the idea here is solid, just the walk back I am
questioning if needed.
Thanks, I was looking for assurance that I'm heading in the right
direction. Thanks to your questions, I realized I should maybe try
experimenting with strncpy as well; I have a feeling this will trip
builtins too, not just general functions with the access attribute.
I'll incorporate that into v2.
Thanks,
Sid
