On Mon, May 13, 2019 at 12:52:54PM -0600, Jeff Law wrote:
> On 5/10/19 9:17 AM, Jakub Jelinek wrote:
> > On Fri, May 10, 2019 at 01:46:07PM +0000, Wilco Dijkstra wrote:
> >> And looking at the generated code, emitting a tailcall for this case is
> >> actually a
> >> de-optimization since the large eh_return epilog must be repeated for every
> >> tailcall.
> >
> > On x86_64, the code is shorter with the tail call rather than without.
> >
> > That said, here is actually tested workaround until targets are fixed.
> > Richard or Jeff, do we want this workaround?
> >
> > I don't see why we would need extra testsuite coverage for this, given the
> > number of FAILs or bootstrap issues on targets that are broken.
> >
> > 2019-05-10 Jakub Jelinek <ja...@redhat.com>
> >
> > PR c++/59813
> > * unwind.inc (_Unwind_Resume_or_Rethrow): Add optimize attribute
> > to temporarily avoid tail calls in the function.
> I think my worry would be all the targets that support tail calls, but
> which we can't easily test. ie, we put in the hack, but when do we
> know it can be removed?
>
> Is there any property of the code that we can look at instead of a hack
> like this?
We can test crtl->calls_eh_return if if make sure it is computed early
(before expansion), or we could just compute it in the ok_for_sibcall
hook if we add caching of that per function (check if any bb that doesn't
have successors and is still in GIMPLE IL ends with __builtin_eh_return
call), if not already crtl->calls_eh_return.
I think only targets that define EH_RETURN_STACKADJ_RTX are problematic, and
from those only those that do not use eh_return_internal patterns (e.g. i386
or bfin or riscv have special eh_return_internal patterns that expand
epilogue with not just true/false for sibcall/for_sibcall, but either a 3
state argument or two bools whether it is sibcall/normal/eh_return).
Out of these, aarch64 and rs6000 have been fixed (though, the latter maybe
not for powerpc-darwin yet).
alpha, arc, cr16, csky, frv, m32c, m68k, mmix, msp430, nds32,
nios2, or1k, pa, sh, tilegx, tilepro, visium, xtensa
might be still broken.
Some of those might be even harder to fix (e.g. alpha, pa, mmix to name a
few), because they don't even bother to pass the sibcall/for_sibcall
argument to epilogue expansion.
Jakub
