http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51165
Aldy Hernandez <aldyh at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |rguenth at gcc dot gnu.org,
| |rth at gcc dot gnu.org
--- Comment #5 from Aldy Hernandez <aldyh at gcc dot gnu.org> 2012-01-11
15:07:15 UTC ---
The main problem here is that we are using ptr_deref_may_alias_global_p() to
determine if a dereferenced address escapes whereas we were previously using
is_call_clobbered() which understood VAR_DECLs, not just SSA_NAMEs.
In requires_barrier() we call ptr_deref_may_alias_global_p() to determine if
the address of `lala' below will escape:
struct large { int x[100]; };
extern struct large foobie (void) __attribute__((transaction_safe));
int asdf;
int f()
{
struct large lala;
struct large lacopy = foobie();
__transaction_atomic {
lala = lacopy; <-- STORE SHOULD BE TRXN/THREAD LOCAL
}
return lala.x[asdf];
}
Before the fix to PR tree-optimization/43572, we used is_call_clobbered() which
returned false for `lala', and everything worked fine. However, we are now
using ptr_deref_may_alias_global_p() which only understands SSA_NAMEs, and
`lala' is a VAR_DECL.
Mr. Guenther (or Mr. Henderson), what is the recommended course of action here?
As an aside (for the PR record), requires_barrier() also needs to handle
ARRAY_REF.
