On Mon, Nov 14, 2016 at 12:04:24AM -0500, Jason Merrill wrote:
> On Wed, Nov 9, 2016 at 7:24 AM, Jakub Jelinek <ja...@redhat.com> wrote:
> > The match.pd hunk is needed, otherwise the generic folding happily folds
> > int arr[2];
> > ...
> > auto [ x, y ] = arr;
> > &x == &arr[0]
> > into 0, because it thinks x and arr are distinct VAR_DECLs. Though, if
> > such comparisons are required to be folded in constexpr contexts under
> > certain conditions, we'd need to handle the DECL_VALUE_EXPRs in constexpr.c
> > somehow.
>
> What do you think of this approach instead?
As Richard said, we'd need to change the 3 other functions too.
And there is additional complication, for OpenMP we defer gimplification of
vars with DECL_VALUE_EXPRs on them, because they often get a different
DECL_VALUE_EXPR. So if optimizers look through DECL_VALUE_EXPR rather than
punt on vars with DECL_VALUE_EXPR, we risk the undesirable value expressions
might leak into the IL.
So I think we want just punt on vars with DECL_VALUE_EXPR (perhaps even in those
4 functions), except for constexpr.c where we perhaps special case the
decomposition vars or something similar.
Jakub
