Re: [PATCH 7/9] Refactor math-finite.h and introduce mathcalls-redir.h

[Tulio Magno Quites Machado Filho <tuliom at linux dot ibm dot com>]

Joseph Myers <joseph@codesourcery.com> writes:

> On Wed, 6 Jun 2018, Tulio Magno Quites Machado Filho wrote:
>
>> On a new enough compiler, redirecting the same function twice can cause
>> -Werror=pragmas errors.  This issue may appear when redirecting an ISO C
>> floating point function to a different ABI at the same time that finite
>> redirections are expected.
>
> Does this issue already appear with -mlong-double-64?  If so, it should 
> definitely be separated out from this patch series and have a bug filed in 
> Bugzilla as usual for anything fixing a bug that was user-visible in a 
> release.

Yes, I can reproduce it in our tests if I replace -mlong-double-128 with
-mlong-double-64.
However, this is not user-visible because this warning is ignored in system
headers.

Do you still think a bug report is necessary?

> The approach taken in this patch involves a lot of duplication of both 
> function prototypes and the feature-test-macro conditions on when those 
> functions are declared (duplicating them for functions not included in 
> math-finite.h, when previously only those in math-finite.h had such 
> duplication).  Such duplication is, from experience, fragile and a bad 
> idea; the conditions and declarations are too likely to get out of sync.

Agreed.

> Instead, I'd suggest merging bits/math-finite.h into bits/mathcalls.h as 
> much as possible - have macros such as __MATHCALL_FINITE that are called 
> in bits/mathcalls.h to declare the functions with __*_finite variants, and 
> with __MATHCALL_FINITE defined the same as __MATHCALL when the finite 
> function variants aren't in use.  That way you have just a single 
> redirection for each function.  (You'll need a sysdeps bits/ header for an 
> architecture, i.e. ia64, to declare it doesn't support the _finite 
> functions at all.)

It doesn't work well this way because we have prototypes with extra attributes,
e.g:

__MATHDECL_1 (int, totalorder,, (_Mdouble_ __x, _Mdouble_ __y))
     __attribute__ ((__const__));

So, if I do not define __MATHDECL_1 in bits/math-finite.h, this code would end
up as an invalid code:

__attribute__ ((__const__));

I managed to remove bits/math-redir.h and modified both math.h and
bits/math-finite.h to reuse bits/mathcalls.h.  I had to keep using
__MATH_DECLARE_FINITE and __MATH_DECLARE_NON_FINITE.

Would this be acceptable?

-- 
Tulio Magno