[Bug target/91204] [10 Regression] ICE in expand_expr_real_2, at expr.c:9215 with -O3

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91204

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Jakub Jelinek from comment #3)
> --- gcc/optabs.c.jj	2019-07-15 10:53:10.743205405 +0200
> +++ gcc/optabs.c	2019-07-19 00:38:20.271852242 +0200
> @@ -2972,6 +2972,17 @@ expand_unop (machine_mode mode, optab un
>        return target;
>      }
>  
> +  /* Emit ~op0 as op0 ^ -1.  */
> +  if (unoptab == one_cmpl_optab
> +      && (SCALAR_INT_MODE_P (mode) || GET_MODE_CLASS (mode) ==
> MODE_VECTOR_INT)
> +      && optab_handler (xor_optab, mode) != CODE_FOR_nothing)
> +    {
> +      temp = expand_binop (mode, xor_optab, op0, CONSTM1_RTX (mode),
> +			   target, unsignedp, OPTAB_DIRECT);
> +      if (temp)
> +	return temp;
> +    }
> +
>    if (optab_to_code (unoptab) == NEG)
>      {
>        /* Try negating floating point values by flipping the sign bit.  */
> 
> seems to work for me.  Or of course something similar can be done in
> config/i386/mmx.md, basically copy the sse.md one_cmpl<mode>2 pattern to
> mmx.md with TARGET_MMX_WITH_SSE and MMXMODEI iterator.

Doing it in the middle-end is better IMHO.  But yeah, we're somewhat sloppy
with vector optabs checks in match.pd where it seems "obvious" targets need
to have support...