Testcase:
#define SIZE 1024
While looking into why manually SRA'd testcase worked better. I found this
interesting problem.
Take:
struct a
{
long long b;
a();
};
a f(a &g, a &h)
{
int i;
a res = g;
for (i = 0; i < SIZE; i++)
res.b += h.b;
return res;
}
at -O2 we change res here to <retval> so we have <retval> in the inner loop.
If we supply -fno-elide-constructors which disables NVR in C++ front-end, we
get good code of doing loop based Copy Prop (SCEV copy prop) as we able to pull
out res.b. The issue is that we say h can alias <retval> but we don't say the
same thing for res though.
--
Summary: NVR in the front-end causes missed optimization later on
(<retval> thought to alias arguments)
Product: gcc
Version: 4.4.0
Status: UNCONFIRMED
Keywords: missed-optimization, alias
Severity: enhancement
Priority: P3
Component: middle-end
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: pinskia at gcc dot gnu dot org
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=36201
