http://gcc.gnu.org/bugzilla/show_bug.cgi?id=54971
--- Comment #13 from Jakub Jelinek <jakub at gcc dot gnu.org> 2012-10-29
08:36:03 UTC ---
So, beyond the creation of new debug only accesses for whole struct writes into
hole if there aren't too many holes, I wonder if SRA doesn't have
infrastructure to do aggregate assignment propagation (which could help with
the rest of the -Os -m32 issues on the committed testcase, but even for code
generation on say):
struct A { int a, b, c, d, e, f, g, h; } z;
struct B { struct A a, b, c, d, e, f, g, h; } x, y;
void
foo (void)
{
struct A a = { 1, 2, 3, 4, 5, 6, 7, 8 };
struct A b = a;
struct A c = b;
struct A d = c;
struct A e = d;
z = e;
}
void
bar (void)
{
struct B a = x;
struct B b = a;
struct B c = b;
struct B d = c;
struct B e = d;
y = e;
}
Here, with -Os both routines result in terrible inefficient code, as total
scalarization is not performed and even for these simple cases where there is
one whole aggregate store and one whole aggregate read that is dominated by the
store SRA (nor any other optimization pass, but IMHO SRA has best
infrastructure for that) doesn't attempt to optimize by doing just y = x; (and
b = x; c = x; d = x; e = x; that would be DCEd away). With -O2 only the second
routine generates terrible code. While this testcase is artificial, the
checked in testcase shows at least one level of extra aggregate assignment
happens e.g. with compound literals.
