[Bug debug/99830] [11 Regression] ICE: in lra_eliminate_regs_1, at lra-eliminations.c:659 with -O2 -fno-expensive-optimizations -fno-split-wide-types -g

["segher at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=99830

--- Comment #14 from Segher Boessenkool <segher at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #13)
> Seems the exact spot where the clobber is optimized away is e.g. when
> simplify_and_const_int_1 (SImode, (ashift:SI (subreg:SI (and:TI (clobber:TI
> (const_int 0 [0])) (const_int 255 [0xff])) 0) (const_int 16 [0x10])), 255);
> is called.
> It calls nonzero_bits, nonzero_bits sees VARYING << 16 and so returns
> 0xffff0000,
>   /* Turn off all bits in the constant that are known to already be zero.
>      Thus, if the AND isn't needed at all, we will have CONSTOP ==
> NONZERO_BITS
>      which is tested below.  */
> 
>   constop &= nonzero;
> 
>   /* If we don't have any bits left, return zero.  */
>   if (constop == 0)
>     return const0_rtx;
> 
> So, are you suggesting that in all such spots we need to test side_effects_p
> and punt?

Yes, you need to do check side_effects_p *everywhere* you can potentially
remove a side effect.  This is not specific to combine, even.

> Note, simplify_and_const_int_1 already starts with:
>   if (GET_CODE (varop) == CLOBBER)
>     return NULL_RTX;
> so it would need to use
>   if (side_effects_p (varop))
>     return NULL_RTX;
> instead.

Yeah.  This no longer disallows a VOIDmode clobber, but we should not see
those here anyway.

You'll need the same change a few lines later, btw:

  varop = force_to_mode (varop, mode, constop, 0);

  /* If VAROP is a CLOBBER, we will fail so return it.  */
  if (GET_CODE (varop) == CLOBBER)
    return varop;

(you only need that second one, even, force_to_mode immediately returns
its arg if it is a clobber).