Re: [PATCH] Don't make all MEM_REFs with TREE_READONLY arguments TREE_READONLY (PR sanitizer/64336)

[Richard Biener]

On Wed, 17 Dec 2014, Jakub Jelinek wrote:

> Hi!
> 
> MEM_REF (the only tcc_reference code with 2 operands) has TREE_READONLY set
> whenever all the arguments are TREE_READONLY, which is wrong, if the
> pointer/reference is read-only, it doesn't say anything about whether
> what it points to is also read-only.
> 
> Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
> trunk?

Thinking about this some more I think that we should instead do sth like

  read_only = 1;
  side_effects = TREE_SIDE_EFFECTS (t);

  if (code == MEM_REF)
   {
     if (TREE_CODE (arg0) == ADDR_EXPR)
       {
         tree t = arg0;
         PROCESS_ARG (0);
       }
     else
       read_only = 0;
   }
 else
   {
     PROCESS_ARG (0);
     PROCESS_ARG (1);
   }

as otherwise read_only will be wrong for MEM_REFs of const decls?

That is, in your patch PROCESS_ARG (0) will still do bogus stuff
for MEM_REFs.

Thanks,
Richard.

> I'll work on TREE_THIS_VOLATILE for MEM_REF/TARGET_MEM_REF later on, as
> a follow-up.
> 
> 2014-12-17  Jakub Jelinek  <ja...@redhat.com>
> 
>       PR sanitizer/64336
>       * tree.c (build2_stat): Don't set TREE_READONLY on MEM_REF
>       even when the arguments are TREE_READONLY.
> 
> --- gcc/tree.c.jj     2014-12-15 10:36:23.000000000 +0100
> +++ gcc/tree.c        2014-12-17 10:48:07.216729576 +0100
> @@ -4352,7 +4352,7 @@ build2_stat (enum tree_code code, tree t
>       arguments are as well.  */
>    constant = (TREE_CODE_CLASS (code) == tcc_comparison
>             || TREE_CODE_CLASS (code) == tcc_binary);
> -  read_only = 1;
> +  read_only = code != MEM_REF;
>    side_effects = TREE_SIDE_EFFECTS (t);
>  
>    PROCESS_ARG (0);
> 
>       Jakub
> 
> 

-- 
Richard Biener <rguent...@suse.de>
SUSE LINUX GmbH, GF: Felix Imendoerffer, Jane Smithard, Jennifer Guild,
Dilip Upmanyu, Graham Norton HRB 21284 (AG Nuernberg)