r210925 - Adds a Pragma spelling for attributes to tablegen and makes use of it for loop

[Aaron Ballman]

On Thu, Nov 20, 2014 at 4:24 PM, Douglas Gregor <dgregor at apple.com> wrote:
> [Dropping Tyler, who is no longer at this address]
>
> On Nov 20, 2014, at 1:12 PM, Aaron Ballman <aaron at aaronballman.com> wrote:
>
> On Thu, Nov 20, 2014 at 3:59 PM, Douglas Gregor <dgregor at apple.com> wrote:
>
>
> On Jun 13, 2014, at 10:57 AM, Tyler Nowicki <tnowicki at apple.com> wrote:
> Modified: cfe/trunk/include/clang/Sema/AttributeList.h
> URL:
> http://llvm.org/viewvc/llvm-project/cfe/trunk/include/clang/Sema/AttributeList.h?rev=210925&r1=210924&r2=210925&view=diff
> ==============================================================================
> --- cfe/trunk/include/clang/Sema/AttributeList.h (original)
> +++ cfe/trunk/include/clang/Sema/AttributeList.h Fri Jun 13 12:57:25 2014
> @@ -80,7 +80,9 @@ public:
>    /// __declspec(...)
>    AS_Declspec,
>    /// __ptr16, alignas(...), etc.
> -    AS_Keyword
> +    AS_Keyword,
> +    /// #pragma ...
> +    AS_Pragma
>  };
>
>
> Amusingly, this new enumerator overflows the bitfield used to store the
> Syntax value:
>
>  /// The number of expression arguments this attribute has.
>  /// The expressions themselves are stored after the object.
>  unsigned NumArgs : 15;
>
>  /// Corresponds to the Syntax enum.
>  unsigned SyntaxUsed : 2;
>
>
> Yoiks! Good catch, Doug. It seems we need a bit more test coverage. ;-)
>
>
> Yet more amusingly, extending this bit-ield to three bits breaks things, as
> does replacing the one source of AS_Syntax attributes with AS_GNU :)
>
>
> I will investigate, but what "things" does it break when you extend
> the bit field to three bits? I modified NumArgs : 15 (down one from
> 16) and SyntaxUsed : 3 (up one from 2), and all tests passed for me
> with MSVC on Windows (not that test coverage is the end-all, be-all).
>
>
> … and it was apparently a bogeyman. Moving to three bits doesn’t actually
> break anything.

Phew!

Unfortunately, testing this specifically is kind of hard to do since
(UB issues aside), it would only manifest itself by turning a valid
attribute into an invalid one, which is kind of hard to test for
because only a few places worry about invalid attributes, and none of
them can be triggered by a pragma-based attribute yet. If someone has
suggestions, I would love to add a test for this. However, it's
important enough to fix that I'm willing to commit a change without a
test.

~Aaron