Re: [PATCH] C++: target attribute - local decl

[Jason Merrill <jason@redhat.com>]

On 3/4/21 3:19 AM, Martin Liška wrote:
> On 3/2/21 6:57 PM, Jason Merrill wrote:
>> On 3/2/21 5:34 AM, Martin Liška wrote:
>>> On 3/1/21 8:58 PM, Jason Merrill wrote:
>>>> On 3/1/21 11:59 AM, Martin Liška wrote:
>>>>> On 3/1/21 5:36 PM, Jason Merrill wrote:
>>>>>> On 3/1/21 7:43 AM, Martin Liška wrote:
>>>>>>> On 2/22/21 11:53 PM, Jason Merrill wrote:
>>>>>>>> The problem seems to be with the handling of local decls.  If 
>>>>>>>> DECL_LOCAL_DECL_P, you need to look at DECL_LOCAL_DECL_ALIAS to 
>>>>>>>> find the namespace-scope decl.  But then if there is no 
>>>>>>>> preceding namespace-scope declaration, the new decl created by 
>>>>>>>> push_local_extern_decl_alias doesn't have a cgraph node, either. 
>>>>>>>> I guess maybe_function_versions 
>>>>>>>> also needs to look through DECL_LOCAL_DECL_ALIAS.
>>>>>>>
>>>>>>> Ah, I see. Are you sure about the name 'maybe_function_versions'? 
>>>>>>> I can't find it.
>>>>>>
>>>>>> Ah, it's maybe_version_functions, sorry.
>>>>>
>>>>> Thanks, I see the function now.
>>>>> So about your guess:
>>>>>
>>>>>> I guess maybe_function_versions also needs to look through 
>>>>>> DECL_LOCAL_DECL_ALIAS.
>>>>>
>>>>> Do you mean maybe_version_functions's argument 'record' should 
>>>>> depend on DECL_LOCAL_DECL_ALIAS of newdecl/oldddecl
>>>>> (if present)? Or that DECL_FUNCTION_VERSIONED should be set for 
>>>>> DECL_LOCAL_DECL_ALIASes of the newdecl/olddecl
>>>>> function declarations?
>>>>
>>>> The latter.
>>>
>>> I see, but will not help us. Problem is that
>>> #2  0x00000000015d8899 in ix86_get_function_versions_dispatcher 
>>> (decl=0x7ffff7755000) at 
>>> /home/marxin/Programming/gcc/gcc/config/i386/i386-features.c:2862
>>>
>>> is called for a declaration for which
>>>
>>> Breakpoint 5, maybe_version_functions (newdecl=<function_decl 
>>> 0x7ffff7755000 f>, olddecl=<function_decl 0x7ffff7751f00 f>, 
>>> record=false) at /home/marxin/Programming/gcc/gcc/cp/decl.c:1118
>>>
>>> is called with record=false. So that cgraph_node is not created for it.
>>>
>>> Or is a possible solution that get_function_version_dispatcher should 
>>> look through the DECL_LOCAL_DECL_ALIAS?
>>
>> Yes, I was thinking both that function and maybe_version_functions.
>>
>> Jason
>>
> 
> All right, the following patch fixes that.
> 
> Patch can bootstrap on x86_64-linux-gnu and survives regression tests.

> +  record_function_versions (olddecl, olddecl, record);
> +  if (DECL_LOCAL_DECL_P (olddecl)
> +      && DECL_LOCAL_DECL_ALIAS (olddecl) != NULL_TREE)
> +    record_function_versions (olddecl, DECL_LOCAL_DECL_ALIAS (olddecl), true);
>  
> +  record_function_versions (olddecl, newdecl, record);
> +  if (DECL_LOCAL_DECL_P (newdecl)
> +      && DECL_LOCAL_DECL_ALIAS (newdecl) != NULL_TREE)
> +    record_function_versions (olddecl, DECL_LOCAL_DECL_ALIAS (newdecl), true);

Do you really need to register all these pairs?  I was expecting that 
you'd just look through DECL_LOCAL_DECL_ALIAS and then register the pair 
you end up with.  Local decls shouldn't need a cgraph node.

Jason