[Bug tree-optimization/113576] [14 regression] 502.gcc_r hangs r14-8223-g1c1853a70f9422169190e65e568dcccbce02d95c

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=113576

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rsandifo at gcc dot gnu.org

--- Comment #11 from Richard Biener <rguenth at gcc dot gnu.org> ---
(In reply to Hongtao Liu from comment #8)
> maybe 
> 
> diff --git a/gcc/fold-const.cc b/gcc/fold-const.cc
> index 1fd957288d4..6d321f9baef 100644
> --- a/gcc/fold-const.cc
> +++ b/gcc/fold-const.cc
> @@ -8035,6 +8035,9 @@ native_encode_vector_part (const_tree expr, unsigned
> char *ptr, int len,
>        unsigned int extract_elts = extract_bytes * elts_per_byte;
>        for (unsigned int i = 0; i < extract_elts; ++i)
>         {
> +         /* Don't encode any bit beyond the range of the vector.  */
> +         if (first_elt + i > count)
> +           break;

Hmm.  I think that VECTOR_CST_ELT should have ICEd for out-of-bound
element queries but it seems to make up elements for us here.  Richard?

But yes, we do

      unsigned int extract_elts = extract_bytes * elts_per_byte;

and since native_encode_* and native_interpret_* operate on bytes we have
difficulties dealing with bit-precision entities with padding.

There's either the possibility to fail encoding when that happens or
do something else.  Note that RTL expansion will do

    case VECTOR_CST:
      {
        tree tmp = NULL_TREE; 
        if (VECTOR_MODE_P (mode))
          return const_vector_from_tree (exp);
        scalar_int_mode int_mode;
        if (is_int_mode (mode, &int_mode))
          {
            tree type_for_mode = lang_hooks.types.type_for_mode (int_mode, 1);
            if (type_for_mode)
              tmp = fold_unary_loc (loc, VIEW_CONVERT_EXPR,
                                    type_for_mode, exp);

which I think should always succeed (otherwise it falls back to expanding
a CTOR).  That means failing to encode/interpret might get into
store_constructor which I think will zero a register destination and thus
fill padding with zeros.

So yeah, something like this looks OK, but I think instead of only
testing against 'count' we should also test against TYPE_VECTOR_SUBPARTS
(that might be variable, so with known_gt).

Would be interesting to see whether this fixes the issue without the
now installed patch.