Re: [Patch/cfgexpand]: also consider assembler_name to call expand_main_function

[Tristan Gingold <gingold@adacore.com>]

On Mar 20, 2012, at 6:17 PM, Jan Hubicka wrote:

>> On Tue, 20 Mar 2012, Tristan Gingold wrote:
>> 
>>> 
>>> On Mar 15, 2012, at 10:37 AM, Richard Guenther wrote:
>>> 
>>>> On Wed, 14 Mar 2012, Tristan Gingold wrote:
>>> [?]
>>> 
>>>> 
>>>> Well.  To make this work in LTO the "main" function (thus, the program
>>>> entry point) should be marked at cgraph level and all users of
>>>> MAIN_NAME_P should instead check a flag on the cgraph node.
>>>> 
>>>>> Will write a predicate in tree.[ch].
>>>> 
>>>> Please instead transition "main-ness" to the graph.
> 
> Yep, I also agree that it is something cgraph code should care about instead of
> random placess across the whole middle-end.
>>> diff --git a/gcc/cfgexpand.c b/gcc/cfgexpand.c
>>> index bd21169..7a7a774 100644
>>> --- a/gcc/cfgexpand.c
>>> +++ b/gcc/cfgexpand.c
>>> @@ -4513,9 +4513,8 @@ gimple_expand_cfg (void)
>>> 
>>>   /* If this function is `main', emit a call to `__main'
>>>      to run global initializers, etc.  */
>>> -  if (DECL_NAME (current_function_decl)
>>> -      && MAIN_NAME_P (DECL_NAME (current_function_decl))
>>> -      && DECL_FILE_SCOPE_P (current_function_decl))
>>> +  if (DECL_FILE_SCOPE_P (current_function_decl)
>>> +      && cgraph_main_function_p (cgraph_get_node (current_function_decl)))
>>>     expand_main_function ();
>> 
>> The DECL_FILE_SCOPE_P check is redundant, please remove them everywhere
>> you call cgraph_main_function_p.  I suppose returning false if the
>> cgraph node is NULL in cgraph_main_function_p would be good.
> 
> How do we handle the cases before cgraph is built with this approach?

Only front-end code need to check wether a function is main before they add
it in cgraph.  As each front-end should know which function is main, this is
not an issue for them.

>>> +/* Return true iff NODE is the main function (main in C).  */
>>> +static inline bool
>>> +cgraph_main_function_p (struct cgraph_node *node)
>>> +{
>>> +  return node->local.main_function;
>> 
>> node && node->local.main_function
> 
> Well, cgraph strategy is ito ICE when NODE is NULL :)
> We could have cgraph_main_function_decl_p wrapper that does the NULL handling, but I still don't
> see how this helps - i.e. when you don't have cgraph node you don't have info whether function
> is main or not, so you should not even try to ask.
> In what cases we ICE here?

We don't ICE here - as long as graph_main_function_p is called after front-end.

>>> +}
>>> +
>>> /* Walk all functions with body defined.  */
>>> #define FOR_EACH_FUNCTION_WITH_GIMPLE_BODY(node) \
>>>    for ((node) = cgraph_first_function_with_gimple_body (); (node); \
>>> diff --git a/gcc/cgraphunit.c b/gcc/cgraphunit.c
>>> index 516f187..4a59f63 100644
>>> --- a/gcc/cgraphunit.c
>>> +++ b/gcc/cgraphunit.c
>>> @@ -346,6 +346,10 @@ cgraph_finalize_function (tree decl, bool nested)
>>>   notice_global_symbol (decl);
>>>   node->local.finalized = true;
>>>   node->lowered = DECL_STRUCT_FUNCTION (decl)->cfg != NULL;
>>> +  node->local.main_function =
>>> +    DECL_FILE_SCOPE_P (decl)
>>> +    && ((!DECL_ASSEMBLER_NAME_SET_P (decl) && MAIN_NAME_P (DECL_NAME (decl)))
>>> +	||decl_assembler_name_equal (decl, main_identifier_node));
>> 
>> If we finalize a function we should always create an assembler name,
>> thus I'd change the above to
>> 
>>  node->local.main_function = decl_assembler_name_equal (decl, 
>> main_identifier_node);
>> 
>> btw, decl_assembler_name_equal doesn't seem to remove target-specific
>> mangling - do some OSes "mangle" main differently (I'm thinking of
>> leading underscores or complete renames)?  Thus, I guess the
>> targets might want to be able to provide the main_identifier_assember_name
>> you use here.
> 
> Yes, name function is mangled, i.e. it is _main on djgpp as long as I remember.
> This is why we have the main_identifier_node to go through the mandling procedure.


USER_LABEL_PREFIX is handled by decl_assembler_name_equal.

One way to simplify that is to change the NESTED argument of cgraph_finalize_function
to LEVEL, which could be either main, top or nested.  With this mechanism, every
front-end will explicitly tell to the middle-end which function is the main entry point.

Thoughts ?

Tristan.

whic