[Bug rtl-optimization/64110] [5 Regression] ICE: Max. number of generated reload insns per insn is achieved (90)

[jakub at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64110

--- Comment #14 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Vladimir Makarov from comment #13)
> (In reply to Jakub Jelinek from comment #11)
> > Looking at the generated assembly, I see there:
> >         movw    %di, -80(%rbp)
> >         vpbroadcastw    -80(%rbp), %ymm1
> >         vmovdqa %ymm1, -80(%rbp)
> > I'd have expected
> >         vmovd   %edi, %xmm1
> >         vpbroadcastw    %xmm1, %ymm1
> >         vmovdqa %ymm1, -80(%rbp)
> > Is there some bug in the *vec_dupv16hi pattern that LRA can't use the !r
> > alternative?  %di should be valid nonimmediate_operand in HImode.  For e.g.
> 
> I know about the problem.  It is coming from a different treatment of ! in
> LRA and reload.  I am going to work on this but as the fix would be in very
> sensitive LRA part. I am afraid to fix it before my vacation.  So probably
> it will be solved only in January.

Np, and thanks a lot.

The ! has been added by me for PR63594, so it isn't there from the era when
i?86 backend was using reload.  If there is a better way to express that RA
should prefer to use memory or xmm register and only use r constraint if it
already is in a r register and doesn't need to be reloaded, I can use that. 
Whether it is ?, ??? or something else.  ! description in gcc docs just fitted
most what I wanted...