Re: [PATCH] Fix ICE in expand_debug_expr on CONJ_EXPR (PR debug/41717)

[Richard Guenther <richard.guenther@gmail.com>]

On Fri, Oct 16, 2009 at 1:23 AM, Jakub Jelinek <jakub@redhat.com> wrote:
> Hi!
>
> CONJ_EXPR wasn't handled by expand_debug_expr, this patch handles at least
> the easy cases.
>
> Also, I've noticed that through a SUBREG with outer mode MODE_INT and inner
> mode e.g. floating mem_loc_descriptor could try to represent e.g. floating
> point multiplication via DW_OP_mul (integer operation) etc.  And, the code
> is prepared to handle only lowpart subregs, not arbitrary ones.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?

Ok.

Thanks,
Richard.

> 2009-10-15  Jakub Jelinek  <jakub@redhat.com>
>
>        PR debug/41717
>        * cfgexpand.c (expand_debug_expr): Handle CONJ_EXPR.
>        * dwarf2out.c (mem_loc_descriptor): Don't handle
>        POST_INT/POST_DEC/POST_MODIFY like SUBREG.  For SUBREG
>        punt if it is not lowpart subreg or if inner mode isn't
>        MODE_INT.
>
>        * gcc.dg/debug/pr41717.c: New test.
>
> --- gcc/cfgexpand.c.jj  2009-10-15 22:07:43.000000000 +0200
> +++ gcc/cfgexpand.c     2009-10-15 23:14:21.000000000 +0200
> @@ -2869,6 +2869,21 @@ expand_debug_expr (tree exp)
>        op1 = gen_rtx_CONST (GET_MODE_INNER (mode), op1);
>       return gen_rtx_CONCAT (mode, op0, op1);
>
> +    case CONJ_EXPR:
> +      if (GET_CODE (op0) == CONCAT)
> +       return gen_rtx_CONCAT (mode, XEXP (op0, 0),
> +                              gen_rtx_NEG (GET_MODE_INNER (mode),
> +                                           XEXP (op0, 1)));
> +      if (MEM_P (op0))
> +       {
> +         enum machine_mode imode = GET_MODE_INNER (mode);
> +         rtx re = adjust_address_nv (op0, imode, 0);
> +         rtx im = adjust_address_nv (op0, imode, GET_MODE_SIZE (imode));
> +         im = gen_rtx_NEG (imode, im);
> +         return gen_rtx_CONCAT (mode, re, im);
> +       }
> +      return NULL;
> +
>     case ADDR_EXPR:
>       op0 = expand_debug_expr (TREE_OPERAND (exp, 0));
>       if (!op0 || !MEM_P (op0))
> --- gcc/dwarf2out.c.jj  2009-10-15 22:07:43.000000000 +0200
> +++ gcc/dwarf2out.c     2009-10-15 23:14:21.000000000 +0200
> @@ -12894,10 +12894,7 @@ mem_loc_descriptor (rtx rtl, enum machin
>     case POST_INC:
>     case POST_DEC:
>     case POST_MODIFY:
> -      /* POST_INC and POST_DEC can be handled just like a SUBREG.  So we
> -        just fall into the SUBREG code.  */
> -
> -      /* ... fall through ...  */
> +      return mem_loc_descriptor (XEXP (rtl, 0), mode, initialized);
>
>     case SUBREG:
>       /* The case of a subreg may arise when we have a local (register)
> @@ -12905,9 +12902,13 @@ mem_loc_descriptor (rtx rtl, enum machin
>         up an entire register.  For now, just assume that it is
>         legitimate to make the Dwarf info refer to the whole register which
>         contains the given subreg.  */
> -      rtl = XEXP (rtl, 0);
> +      if (!subreg_lowpart_p (rtl))
> +       break;
> +      rtl = SUBREG_REG (rtl);
>       if (GET_MODE_SIZE (GET_MODE (rtl)) > DWARF2_ADDR_SIZE)
>        break;
> +      if (GET_MODE_CLASS (GET_MODE (rtl)) != MODE_INT)
> +       break;
>       mem_loc_result = mem_loc_descriptor (rtl, mode, initialized);
>       break;
>
> --- gcc/testsuite/gcc.dg/debug/pr41717.c.jj     2009-10-15 23:13:24.000000000 +0200
> +++ gcc/testsuite/gcc.dg/debug/pr41717.c        2009-10-15 23:12:52.000000000 +0200
> @@ -0,0 +1,10 @@
> +/* PR debug/41717 */
> +/* { dg-do compile } */
> +
> +void
> +foo (void)
> +{
> +  _Complex float v[1], w;
> +  v[1] = 0.0f + 0.8fi;
> +  w = __builtin_conjf (v[1] * v[1]);
> +}
>
>        Jakub
>