On Sat, May 11, 2019 at 09:38:22AM +0100, Richard Sandiford wrote: > Jakub Jelinek <ja...@redhat.com> writes: > > Previously, the mere existence of the addressable variables this_context > > and cur_context prevented tail call on the early out > > return _Unwind_RaiseException (exc); > > but since r271013 the tailcall analysis figures that while those two > > variables are there, they aren't touched before the possible tail call > > site, so they can't be really live during the call.
Ah! That explains why this never happened before. > > This does a lot of register saving and restoring, which is not needed but is > > not wrong-code (guess separate shrink wrapping would help here if > > implemented for the target). SWS does not handle eh_return: https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=gcc/shrink-wrap.c;h=57124db92c662bf52efc7ea94c274d9b4e234d04;hb=HEAD#l1780 But, a function calling eh_return will not get *any* shrink-wrapping: https://gcc.gnu.org/git/?p=gcc.git;a=blob;f=gcc/shrink-wrap.c;h=57124db92c662bf52efc7ea94c274d9b4e234d04;hb=HEAD#l658 (it has done this since r179553, when shrink-wrapping was added). Some targets run splitters for their eh_return and evilness like that. It never was documented what the compiler expects the target to do with eh_return, and what the limitations are, etc. > > The only wrong-code is actually the > > add sp, sp, x4 instruction though. The previous instruction restored sp to > > the value it had at the start of the function and then we should just tail > > call. But is that correct? eh_return is supposed to adjust the stack? And, what should other targets do? > FWIW, I agree we should just fix the targets and not even use the > workaround you posted later. It won't be the first time that many > targets have got something wrong due to lack of coverage. eh_return isn't defined or documented very well. That should be fixed... Or maybe eh_return should just be removed. Segher > > --- gcc/config/aarch64/aarch64.c.jj 2019-05-02 12:18:40.004979690 +0200 > > +++ gcc/config/aarch64/aarch64.c 2019-05-09 20:08:00.774718003 +0200 > > @@ -5913,7 +5913,7 @@ aarch64_expand_epilogue (bool for_sibcal > > } > > > > /* Stack adjustment for exception handler. */ > > - if (crtl->calls_eh_return) > > + if (crtl->calls_eh_return && !for_sibcall) > > { > > /* We need to unwind the stack by the offset computed by > > EH_RETURN_STACKADJ_RTX. We have already reset the CFA
