Re: attribute data structure rewrite

[Geoffrey Keating <geoffk@apple.com>]

[-- Attachment #1: Type: text/plain, Size: 2579 bytes --]


On 23/09/2004, at 5:47 PM, Gabriel Dos Reis wrote:

> Geoffrey Keating <geoffk@apple.com> writes:
>
> | As a precondition for other compile speed improvements, I've been
> | rewriting the way that GCC represents attributes.  The result is a
> | medium-size patch that touches every language and many targets (a
> | surprising number of targets don't define any special attributes at
> | all).
> |
> | The basic principle is that DECL_ATTRIBUTES and TREE_ATTRIBUTES now
> | point to an attribute_list:
> |
> | /* A structure representing 'attributes' on a DECL or TYPE node.
> |     Each attribute has a NAME (an IDENTIFIER_NODE) and possibly a
> | VALUE. */
> | struct one_attribute GTY(())
> | {
> |    tree name;
> |    tree value;
> | };
> |
> | /* A counted list of attributes.  */
> | struct attribute_list_s GTY(())
> | {
> |    attribute_count n_attributes;
> |    /* There are 16 bits free here.  */
> |    struct one_attribute GTY((length ("%h.n_attributes"))) attribs[1];
> | };
> |
> | This has many benefits, mostly flowing from the fact that it's not a
> | TREE_LIST any more.
> |
> | 1. Do we think this would be acceptable for stage 3, or should I make
> | a branch?
>
>
> I would argue that such a change should be accepted, at this phase of
> stage 3; now.  As:
>   (1) it addresses compile-time/memory regression;
>   (2) it is no less interesting no less useful that the various cleanup
>       we've been having, in particular in the C++ front-end.
>
>
> I would even take a further step and ask an official position about
> how we how attributes to play with language rules.
>
>
> Currently, the compiler fails on any of the following fails,
> complaining about  non-"integral constant expression"-ness, which is
> pure nonense.
>
>     const int N = 4;
>     struct S {
>       enum { M = 4 };
>       float x __attibute__((__aligned__(N)));
>       float x __attibute__((__aligned__(M)));
>     };
>
>    template<typename T, int N>
>       struct buffer {
>          typedef char U[sizeof(T)];
>          U data __attribute__((__aligned__(N)));
>      };
>
> They should be accepted.

That's a very good point.  I noticed from looking at the C front-end 
that an attribute can have as a 'value' any of:

1. NULL
2. an IDENTIFIER
3. an expression

 From that, I suspect that what's happening here is that 'N' is an 
identifier, so that rule matches in preference to treating 'N' as an 
expression and evaluating it.

I would really like to avoid case (2), or at least make sure that you 
can tell which of (2) or (3) any particular attribute should have.

[-- Attachment #2: smime.p7s --]
[-- Type: application/pkcs7-signature, Size: 2408 bytes --]