Re: [PATCH 8/9] Add TARGET_ADDR_SPACE_ZERO_ADDRESS_VALID

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

On Wed, Oct 14, 2015 at 11:19 AM, Richard Biener
<richard.guenther@gmail.com> wrote:
> On Tue, Oct 13, 2015 at 10:59 PM, Richard Henderson <rth@redhat.com> wrote:
>> On 10/14/2015 02:49 AM, Jeff Law wrote:
>>>
>>> The problem here is we don't know what address space the *0 is going to
>>> hit,
>>> right?
>>
>>
>> Correct, not before we do the walk of stmt to see what's present.
>>
>>> Isn't that also an issue for code generation as well?
>>
>>
>> What sort of problem are you thinking of?  I haven't seen one yet.
>
> The actual dereference of course has a properly address-space qualified zero.
>
> Only your walking depends on operand_equal_p to treat different address-space
> zero addresses as equal (which they are of course not ...):
>
>
> int
> operand_equal_p (const_tree arg0, const_tree arg1, unsigned int flags)
> {
> ...
>   /* Check equality of integer constants before bailing out due to
>      precision differences.  */
>   if (TREE_CODE (arg0) == INTEGER_CST && TREE_CODE (arg1) == INTEGER_CST)
>     {
>       /* Address of INTEGER_CST is not defined; check that we did not forget
>          to drop the OEP_ADDRESS_OF/OEP_CONSTANT_ADDRESS_OF flags.  */
>       gcc_checking_assert (!(flags
>                              & (OEP_ADDRESS_OF | OEP_CONSTANT_ADDRESS_OF)));
>       return tree_int_cst_equal (arg0, arg1);
>     }
>
> but only later we do
>
>       /* We cannot consider pointers to different address space equal.  */
>       if (POINTER_TYPE_P (TREE_TYPE (arg0))
>                           && POINTER_TYPE_P (TREE_TYPE (arg1))
>           && (TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (arg0)))
>               != TYPE_ADDR_SPACE (TREE_TYPE (TREE_TYPE (arg1)))))
>         return 0;
>
> So "fixing" that would make the walker only look for default
> address-space zero dereferences.
>
> I think we need to fix operand_equal_p anyway because 0 is clearly not
> equal to 0 (only if
> they convert to the same literal)

I think you could trigger bogus CSE of dereferences of literal addresses
from different address-spaces.

Richard.

> Richard.
>
>
>>
>> r~