https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85459
Richard Biener <rguenth at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Priority|P3 |P2
Status|UNCONFIRMED |NEW
Last reconfirmed| |2018-12-20
CC| |rguenth at gcc dot gnu.org
Ever confirmed|0 |1
--- Comment #7 from Richard Biener <rguenth at gcc dot gnu.org> ---
So it's
if (TREE_CODE (ref) != MEM_REF
|| TREE_CODE (TREE_OPERAND (ref, 0)) != ADDR_EXPR)
return false;
tree mem = TREE_OPERAND (TREE_OPERAND (ref, 0), 0);
if (TYPE_MAIN_VARIANT (TREE_TYPE (ref))
!= TYPE_MAIN_VARIANT (TREE_TYPE (mem)))
return true;
specifically the TYPE_MAIN_VARIANT check. This is probably too conservative
given it labels writing a.i as MEM[&a + 4] as VCE.
I wonder if for the purpose of total scalarization we can ignore VCEs
on aggregate copies in case we do not need to re-materialize the thing?
Not having looked at the actual testcase here but guessing from the
testcase added by the above change and memcpy inlining now generating
char[] typed load/stores this is what prevents total scalarization in
too many cases.
