https://gcc.gnu.org/bugzilla/show_bug.cgi?id=116873
Bug ID: 116873
Summary: esra does not remove unused store sometimes
Product: gcc
Version: 15.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: enhancement
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: pinskia at gcc dot gnu.org
Target Milestone: ---
Take:
```
struct Foo {
int t;
int x;
};
int f(struct Foo a, struct Foo b, int c)
{
struct Foo d;
struct Foo e;
struct Foo *l;
if (c) l = &a; else l = &b;
d = *l;
e = *l;
return d.t + e.x;
}
```
After esra we get:
```
# l_3 = PHI <&a(2), &b(3)>
d = MEM[(const struct Foo &)l_3];
d$t_6 = MEM[(const struct Foo &)l_3].t;
e = MEM[(const struct Foo &)l_3];
e$x_5 = MEM[(const struct Foo &)l_3].x;
_1 = d$t_6;
_2 = e$x_5;
_10 = _1 + _2;
d ={v} {CLOBBER(eos)};
e ={v} {CLOBBER(eos)};
return _10;
```
BUT the store to d and e are dea. If we full scalarize d and e, the stores are
removed (that is use d.x and e.t).
Shouldn't we just mark d and e to be removed too?
Yes DSE does remove them later on but I don't see why SRA could not handle them
if fields are unused.
