[RFA] Fix various regressions and kernel build failure due to adjust-alignment issue

[Jakub Jelinek]

On Tue, Jun 09, 2020 at 09:18:25AM -0600, Jeff Law wrote:
> On Tue, 2020-06-09 at 16:59 +0200, Jakub Jelinek wrote:
> > On Tue, Jun 09, 2020 at 08:54:47AM -0600, Jeff Law via Gcc-patches wrote:
> > > > While technically OK the issue is that it does not solve the x86 issue
> > > > which with incoming stack alignment < 8 bytes does not perform
> > > > dynamic re-alignment to align 'long long' variables appropriately.
> > > > Currently the x86 backend thinks it can fixup by lowering alignment
> > > > via LOCAL_DECL_ALIGNMENT but this is a latent wrong-code issue
> > > > because the larger alignment is present in the IL for a long (previously
> > > > RTL expansion, now adjust-aligment) time and your patch makes that
> > > > wrong info last forever (or alternatively cause dynamic stack alignment
> > > > which we do _not_ want to perform here).
> > > I've never looked at the dynamic realignment stuff -- is there a good reason why
> > > we don't dynamically realign when we've got a local with a requested alignment? 
> > > Isn't that a huge oversight in the whole concept of dynamic realignment?
> > 
> > It is quite expensive and long long/double uses are pervasive, so we don't
> > want to realign just because of that.  If we do dynamic realignment for
> > other reasons, there is no reason not to align the long long/double
> Right, but if there's an explicit alignment from the user, don't we need to honor
> that?

Sure.  Do we ignore that?  For user alignment we have DECL_USER_ALIGN bit...

	Jakub