[Bug rtl-optimization/55294] Invalid RTL sharing in lower-subreg

[olegendo at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=55294

--- Comment #5 from Oleg Endo <olegendo at gcc dot gnu.org> ---
(In reply to Andrew Pinski from comment #4)
> (In reply to Andrew Pinski from comment #2)
> > I think this is because:
> > (subreg:SF (reg/v:V2SF 167 [ a ]) 0)))
> > 
> > is invalid to begin with.  Yes we don't reject it but I think we should.
> 
> So this is valid thing these days. Assuming little endian.

Hm?  If this is rejected, then how would one access the individual SF reg of a
V2SF?  I'm actually using this in the fsca /sincossf3 pattern.  The "fsca"
pattern always writes the sin+cos result into a V2SF.  Then the result will be
picked up via

  emit_move_insn (operands[0], gen_rtx_SUBREG (SFmode, fsca, 4));