http://gcc.gnu.org/bugzilla/show_bug.cgi?id=56982
--- Comment #5 from Richard Biener <rguenth at gcc dot gnu.org> 2013-04-17
08:48:33 UTC ---
So the questions are:
- is it desirable that uncprop does anything to SSA_NAME_VAR == NULL phis?
sure - it is all about improving out-of-SSA coalescing opportunities
and avoiding copies
- shouldn't something like that be not performed if current function calls
setjmp (or more narrowly, if there is a returns twice function somewhere in
between the considered setter and user)?
the testcase shows that uncprop extends the lifetime of an SSA name
across a setjmp call - but it can only do so because it's an SSA name.
Which means the testcase is questionable as 'n' is not declared volatile, no?
- what other optimizations might be similarly problematic across returns twice
calls?
every optimization pass that performs hosting. PRE comes to my mind for
if (x)
tem = expr;
setjmp ()
var = expr;
which would happily eliminate the partial redundancy, moving expr to the
else arm of the if () and thus extending the lifetime of 'var' across
the setjmp call.
We do not explicitely model the abnormal control flow for setjmp / longjmp
which is the reason all these issues may appear. So I believe the
correct fix is to either declare the testcase invalid or to model
the abnormal control flow explicitely. Add abnormal edges from all
call sites in the function that may end up calling longjmp _and_ eventually an
abnormal edge from function entry as we can call longjmp from callers
as well (though that may be invalid and thus we do not have to care?).
I don't see an "easy" fix for the issue (well, maybe the specific testcase).
That it happens only after my patch is probably pure luck because of for
example the PRE issue. Testcase for that:
int f (int a, int flag)
{
int tem;
if (flag)
tem = a + 1;
int x = setjmp (env);
int tem2 = a + 1;
if (x)
return tem2;
return tem;
}
validity of course is questionable, but we clearly use tem only on the
normal path and tem2 on the abnormal path. PRE does the transform
I indicated, proper abnormal edges would disable the transform.
