[Bug middle-end/78468] [7 regression] libgomp.c/reduction-10.c and many more FAIL

[ebotcazou at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=78468

Eric Botcazou <ebotcazou at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
          Component|libgomp                     |middle-end
           Severity|normal                      |major

--- Comment #21 from Eric Botcazou <ebotcazou at gcc dot gnu.org> ---
> It does not break anything that wasn't broken before.  The Sparc backend was
> just _lucky_ that the allocation code in the middlend was _broken_. 

Yes, it does, it breaks tests that have been working for a decade!  

> The patch fixes the bug that causes dynamic stack allocation to overestimate
> the needed space on the stack most of the time.  To do this, it uses
> information available from elsewhere in the middleend.
>
> It turns out that the backend (or middlend, depends on the point of view)
> lies about the alignment of VIRTUAL_STACK_DYNAMIC_REGNUM.  There may be
> _other_ users users of that value that fail to do their job because they
> think the stored alignment is correct.  Such users may do worse things than
> wasting some stack space - we may just have not noticed them yet.
> 
> So, there is _another_ bug in the backends (or the middleend) that needs to
> be fixed.  It's not "one fix instead of another" - there are two bugs that
> need two separate fixes.

Then let's fix them, please.

> It is also better to make the broken backends respect the STACK_BOUNDARY
> alignment of VIRTUAL_STACK_DYNAMIC_REGNUM than make them tell the middleend
> the real alignment:

But these back-ends were not broken before your patch!  It's your patch that
broke the interface between the middle-end and the back-end without notice.

> Making the middleend force STACK_BOUNDARY alignment of
> VIRTUAL_STACK_DYNAMIC_REGNUM is also possible, but without any information
> from the backend we'd have to generate run time code to do that alignment. 
> So, why not just use STACK_DYNAMIC_OFFSET et. al. to just guarantee that the
> alignment is right?

Because we have 50 architectures in the tree; are you going to audit them all
and make the necessary adjustments?  Just to save a couple of bytes?