https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82224
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Keywords| |alias
Status|UNCONFIRMED |ASSIGNED
Last reconfirmed| |2017-09-18
Component|c |tree-optimization
Ever confirmed|0 |1
--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
It's more related to PR77745. With the testcase we are still removing the
stores that change the active union member. They do have the same alias set
unfortunately given they use union s1s2 * in their access path.
The IL that we then up optimizing is
<bb 2> [100.00%] [count: INV]:
q[0].v1.x = 4321;
...
_3 = &q + _2;
_14 = MEM[(struct s1 *)&q].x;
...
if (_15 != 0)
goto <bb 3>; [33.00%] [count: INV]
else
goto <bb 4>; [67.00%] [count: INV]
<bb 3> [33.00%] [count: INV]:
MEM[(struct s2 *)_3].x = 1234;
<bb 4> [100.00%] [count: INV]:
_17 = MEM[(struct s1 *)&q].x;
...
where you can see all loads/stores being done through s1/s2 types which do
_not_ alias.
So we have
q->v1.x = 4321; // make v1 active
if ((&q->v1)->x) // read via pointer to active member, supposedly valid
{
q->v2.x = q->v1.x; // change v2 to be active, same value, optimized away
(&q->v2)->x = 1234; // write via pointer to active member
q->v1.x = q->v2.x; // change v1 to be active, same value, optimized away
}
if ((&q->v1)->x != 1234)
abort ();
if we do not DSE the active member changing stores everything works as
desired.
Note that for GCC it isn't necessary to go through the union type when
changing the active type. Note it definitely is desirable to elide
the load/store changing the active member in the end.
It's going to be tricky to find a solution that doesn't pessimize optimization
when unions are involved. DOM is similarly affected, in the end it needs
sufficiently obfuscated IL to trick our "handle must-aliases conservatively"
code enough.
