[Bug tree-optimization/84013] wrong __restrict clique with inline asm operand

["rguenther at suse dot de" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84013

--- Comment #3 from rguenther at suse dot de <rguenther at suse dot de> ---
On Wed, 31 Jan 2018, katsunori.kumatani at gmail dot com wrote:

> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84013
> 
> --- Comment #2 from Katsunori Kumatani <katsunori.kumatani at gmail dot com> ---
> I'm not familiar with tree-ssa-structalias, but it appears to me that the "fix"
> is quite simple? Or am I missing something? Here's the snippet from it, updated
> with my attempt:
> 
>       if (restrict_var)
>         {
>           /* Now look at possible dereferences of ptr.  */
>           imm_use_iterator ui;
>           gimple *use_stmt;
>           bool used = false;
>           FOR_EACH_IMM_USE_STMT (use_stmt, ui, ptr)
>             {
>               if (!gimple_assign_single_p (use_stmt))
>                 {
>                   /* ???  Calls.  */
>                   if (gimple_code (use_stmt) != GIMPLE_ASM)
>                     continue;
> 
>                   gasm *asm_stmt = as_a <gasm *> (use_stmt);
>                   unsigned n = gimple_asm_ninputs (asm_stmt);
>                   for (unsigned i = 0; i < n; i++)
>                     {
>                       tree op = TREE_VALUE (gimple_asm_input_op (asm_stmt, i));
>                       used |= maybe_set_dependence_info (op, ptr, clique,
>                                                          restrict_var,
>                                                          last_ruid);
>                     }
>                   n = gimple_asm_noutputs (asm_stmt);
>                   for (unsigned i = 0; i < n; i++)
>                     {
>                       tree op = TREE_VALUE (gimple_asm_output_op (asm_stmt,
> i));
>                       used |= maybe_set_dependence_info (op, ptr, clique,
>                                                          restrict_var,
>                                                          last_ruid);
>                     }
>                   continue;
>                 }
> 
> 
> Does this not work? Sorry if I am missing something here.

Yes, this should work.  It's not difficult to fix I just thought it
wasn't important to track restrict across asm()s ...  A similar
fix is needed for call return and operands.  Writing the code
a little nicer and more compact should be possible as well - I
just need to think about it somewhat (similar to the code a bit
below we should be able to use walk_stmt_load_store_ops).

I will deal with this for GCC 9 given this isn't a regression.