On Mon, May 19, 2014 at 12:03 PM, Richard Sandiford
<rdsandif...@googlemail.com> wrote:
> Richard Biener <richard.guent...@gmail.com> writes:
>> On Mon, May 19, 2014 at 11:27 AM, Richard Sandiford
>> <rdsandif...@googlemail.com> wrote:
>>> Richard Biener <richard.guent...@gmail.com> writes:
>>>> On Sat, May 17, 2014 at 10:15 AM, Richard Sandiford
>>>> <rdsandif...@googlemail.com> wrote:
>>>>> The main thing keeping zero-precision wide-ints alive was void_zero_node,
>>>>> a tree used in the C and C++ frontends that has type VOID_TYPE but code
>>>>> INTEGER_CST.
>>>>>
>>>>> Richard B. asked me to replace the INTEGER_CST with a new constant type,
>>>>> here called VOID_CST. Most of it is straight-forward. The one perhaps
>>>>> controversial bit is that C++ uses void_(zero_)node to represent dummy
>>>>> objects when e.g. taking the address of a member function without an
>>>>> associated object. IIUC the node used in this situation needs to be
>>>>> distinct from anything that could occur in user code and therefore
>>>>> couldn't
>>>>> be a simple null pointer.
>>>>>
>>>>> This reaches the gimplifier in cases like
>>>>> g++.old-deja/g++.brendan/operators4.C. I chose to handle it in the
>>>>> gimplifier, since void_zero_node was previously handled there too,
>>>>> although perhaps by accident. If you object strongly to this then
>>>>> I'll need pretty detailed instructions about what to do instead,
>>>>> i.e. exactly which parts of the C++ front end need to be changed
>>>>> in order for dummy objects never to escape.
>>>>
>>>> I suppose it reaches the gimplifier because it's not handled in
>>>> fold-const.c:fold_convert_loc while the INTEGER_CST void_zero_node
>>>> was (through fold_convert_const).
>>>
>>> But like I said, void_zero_node reached the gimplifier too. Try adding:
>>>
>>> gcc_assert (TREE_TYPE (TREE_OPERAND (*expr_p, 0)) != void_type_node
>>> || TREE_CODE (TREE_OPERAND (*expr_p, 0)) != INTEGER_CST);
>>>
>>> to gimplify_conversion and running g++.old-deja/g++.brendan/operators4.C
>>> to see what I mean.
>>>
>>>> That said, only handling (T)void_cst in gimplification looks like
>>>> a hack. If necessary we'd want to treat it as construct-T-with-zero-value
>>>> consistently.
>>>
>>> OK, so just remove the gcc_checking_assert?
>>
>> Which one?
>
> The one I'd added to the gimplifier:
>
> Index: gcc/gimplify.c
> ===================================================================
> --- gcc/gimplify.c 2014-05-15 13:49:24.483656013 +0100
> +++ gcc/gimplify.c 2014-05-16 17:55:17.087700837 +0100
> @@ -1681,7 +1681,15 @@ gimplify_conversion (tree *expr_p)
> /* Then strip away all but the outermost conversion. */
> STRIP_SIGN_NOPS (TREE_OPERAND (*expr_p, 0));
>
> - /* And remove the outermost conversion if it's useless. */
> + /* Support C++-style dummy objects, in which void_zero is
> + cast to a pointer type. We treat these as null pointers. */
> + if (TREE_OPERAND (*expr_p, 0) == void_node)
> + {
> + gcc_checking_assert (POINTER_TYPE_P (TREE_TYPE (*expr_p)));
> + *expr_p = build_int_cst (TREE_TYPE (*expr_p), 0);
> + }
> +
> + /* Remove the outermost conversion if it's useless. */
> if (tree_ssa_useless_type_conversion (*expr_p))
> *expr_p = TREE_OPERAND (*expr_p, 0);
Ah. yes, remove the assert and instead use
*expr_p = build_zero_cost (TREE_TYPE (*expr_p));
>> I'd add VOID_CST handling to fold_convert_const.
>
> But like you say, if that was enough, void_zero_node would never
> have reached the gimplifier, whereas as above it does. I tried adding:
>
> gcc_assert (TREE_CODE (arg1) != VOID_CST);
>
> to fold_convert_const and it doesn't trigger for operators4.C.
> It also doesn't trigger for g++-mike/p10769b.C, which was the
> other main motivating case.
Ok, I see. Thus fine IMHO with doing the above (though I wonder
why we couldn't have done this in the FE specific gimplification
routines for void_zero_node ...?)
Still defer to C++ people for the C++ parts.
Thanks,
Richard.
> Thanks,
> Richard
