[PATCH] PR c/86407 - Add option to ignore fndecl attributes on function pointers

[Joseph Myers]

On Thu, 30 May 2019, Alex Henrie wrote:

> In Wine we need a way to (without warnings) put ms_hook_prologue into
> a macro that is applied to functions, function pointers, and function
> pointer typedefs. It sounds like you're saying that you will not
> accept a patch that silences or splits off warnings about using
> ms_hook_prologue with function pointers and function pointer typedefs.
> So how do you think Wine's problem should be solved?

By actually implementing the required semantics, not by silencing warnings 
that say you're using an attribute in a location for which no semantics 
for that attribute are known.

1. An attribute appertaining to a function pointer (an object whose type 
is pointer to function) is something different from an attribute 
appertaining to the type of the pointed-to function.  For example, the 
"section" attribute appertains to declarations.  If you apply it to the 
declaration of a function, it puts that function in that section.  If you 
apply it to the declaration of an object whose type is pointer to 
function, it puts that object in that section (it does not say what 
section the pointed-to function is in).

2. The "Attribute Syntax" section in the manual describes how to determine 
to what construct an attribute in the GNU syntax appertains.  Please note 
that this is *not* always the same construct to which a C++ / C2x 
attribute in the same place would appertain (I might look at C2x attribute 
implementation and other C2x features implementation around 
August/September, not before then).  You should first be familiar with how 
the C standard defines the syntax for declarators and their corresponding 
types, in order to follow the extensions for attributes included in 
declarators.

3. There are certain cases where GCC is deliberately lax about attributes 
being in the syntactically correct places, which in particular makes it 
possible to change an attribute from being a declaration attribute to one 
on function types without breaking existing source code expecting it to be 
on function declarations.  See the type_required and 
function_type_required members of struct attribute_spec.  Specifying an 
attribute as being a function type attribute like that means it can be 
specified in a position where syntactically it should appertain to the 
declaration of the function pointer object, and be automatically made to 
appertain to the pointed-to type instead - but that only works if all the 
code in GCC that looks to see whether some declaration has the attribute 
present, and acts accordingly, is changed to check for the attribute on 
the type instead.

4. So suppose you have a particular attribute, currently only accepted on 
function declarations, that you wish to be accepted on function types.  
You need to find all the places checking for that attribute on 
declarations within GCC, to implement its semantics, and change them (a) 
to check the type of the declaration instead and (b) to work with the 
appropriate type when there isn't a declaration at all, as for indirect 
calls.  (What's involved in this may depend on whether the code just 
checks the list of attributes on the declaration, or whether there is some 
other bit set on the tree that may or may not be present for types at 
all.)  You also need to change the relevant table of attributes to set 
type_required and function_type_required instead of decl_required.  And 
you need to change the attribute handler accordingly so that it works when 
given a type rather than a declaration.

-- 
Joseph S. Myers
joseph@codesourcery.com