[Bug c/63307] [4.9/5 Regression] Cilk+ breaks -fcompare-debug bootstrap

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63307

--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(In reply to Igor Zamyatin from comment #1)
> Would like to ask here first - will something like following be ok:

> +typedef struct
> +{
> +  tree parm;
> +  tree arg;
> +} decl_pair;
> +
> +static vec<decl_pair> vec_arglist;

Just use struct cilk_decl_pair, no need for typedef in C++ here, and
try to avoid ODR issues.  Also, why a static variable?  Then you supposedly
would need to GTY handle it etc., which is IMHO undesirable.

>  static bool
>  wrapper_parm_cb (const void *key0, void **val0, void *data)
>  {
> -  struct wrapper_data *wd = (struct wrapper_data *) data;
>    tree arg = * (tree *)&key0;
>    tree val = (tree)*val0;
>    tree parm;
> +  decl_pair dp;
>  
>    if (val == error_mark_node || val == arg)
>      return true;
> @@ -370,25 +379,48 @@ wrapper_parm_cb (const void *key0, void **val0, void
> *data)
>      }
>    else
>      parm = val;
> -  TREE_CHAIN (parm) = wd->parms;
> -  wd->parms = parm;
> -  wd->argtypes = tree_cons (NULL_TREE, TREE_TYPE (parm), wd->argtypes); 
> -  wd->arglist = tree_cons (NULL_TREE, arg, wd->arglist); 
> +
> +  dp.parm = parm;
> +  dp.arg = arg;
> +  vec_arglist.safe_push(dp);

Formatting, missing space before (.  But more importantly, the diagnostics
still be in random order.  So, I'd strongly suggest to move also the
diagnostics and ADDR_EXPR building etc. into a loop that walks the sorted
vector.

>    return true;
>  }
>  
>  /* This function is used to build a wrapper of a certain type.  */
>  
> +static int
> +compare_decls (const void *a, const void *b)
> +{
> +    const decl_pair* t1 = (const decl_pair*) a;
> +    const decl_pair* t2 = (const decl_pair*) b;
> +
> +    return DECL_UID(t1->arg) > DECL_UID(t2->arg);

Formatting.  Space before *, not after, space before (, indentation 2 columns
rather than 4.

> +}
> +
>  static void
>  build_wrapper_type (struct wrapper_data *wd)
>  {
> +  unsigned int j;
> +  decl_pair * c;
>    wd->arglist = NULL_TREE;
>    wd->parms = NULL_TREE;
>    wd->argtypes = void_list_node;
>  
> -  pointer_map_traverse (wd->decl_map, wrapper_parm_cb, wd);
> +  vec_arglist.create (0);
> +  pointer_map_traverse (wd->decl_map, wrapper_parm_cb, NULL);
>    gcc_assert (wd->type != CILK_BLOCK_FOR);
>  
> +  vec_arglist.qsort(compare_decls);

Formatting.

> +
> +  FOR_EACH_VEC_ELT (vec_arglist, j, c)
> +    {
> +      TREE_CHAIN (c->parm) = wd->parms;
> +      wd->parms = c->parm;
> +      wd->argtypes = tree_cons (NULL_TREE, TREE_TYPE (c->parm),
> wd->argtypes);
> +      wd->arglist = tree_cons (NULL_TREE, c->arg, wd->arglist);
> +    }

As I said earlier, I'd do diagnostics and ADDR_EXPR creation in this loop too.

> +  vec_arglist.release();

Formatting.  You could use auto_vec, perhaps with some stack allocated initial
buffer if you think say 16 vector elements would be typically enough.

Also, what about all the remaining 3 callbacks that create or may create decls
and have the same problem?  for_local_cb, wrapper_local_cb and
declare_one_free_variable.