[PATCH v4 08/12] math: Add math-use-builtinds-fmax.h

[Adhemerval Zanella]

On 06/12/2021 18:11, Joseph Myers wrote:
> On Thu, 2 Dec 2021, Adhemerval Zanella via Libc-alpha wrote:
> 
>> +#if __HAVE_FLOAT128
>> +# define USE_BUILTIN_F128 , _Float128 : USE_FMAXF128_BUILTIN
>> +# define BUILTIN_F128     , _Float128 :__builtin_fmaxf128
>> +#else
>> +# define USE_BUILTIN_F128
>> +# define BUILTIN_F128
>> +#endif
>> +
>> +#define USE_BUILTIN(X, Y)                   \
>> +  _Generic((X),                             \
>> +	   float       : USE_FMAXF_BUILTIN, \
>> +	   double      : USE_FMAX_BUILTIN,  \
>> +	   long double : USE_FMAXL_BUILTIN  \
>> +	   USE_BUILTIN_F128)
>> +
>> +#define BUILTIN(X, Y)                       \
>> +  _Generic((X),                             \
>> +	   float       : __builtin_fmaxf,   \
>> +	   double      : __builtin_fmax,    \
>> +	   long double : __builtin_fmaxl    \
>> +	   BUILTIN_F128)                    \
>> +  (X, Y)
> 
> You should not hardcode cases for different types in type-generic 
> templates.  This would fail if we add support for _Float16 functions in 
> future, for example.
> 
> Use M_SUF to call built-in functions (possibly via defining M_FMAX in 
> math-type-macros.h, similar to the other M_* macros there.
> 
> Do something similar for USE_*_BUILTIN (I suggest defining M_USE_BUILTIN 
> in the per-type headers used by math-type-macros.h, so you can then call 
> M_USE_BUILTIN (FMAX).

Indeed it is cleaner, I don't see the need of just M_FMAX.

> 
> You'll probably need to use #if with the M_USE_BUILTIN call (which should 
> be OK after the above changes), because GCC only added __builtin_fmaxfN 
> functions in commit ee5fd23a481f510528e00f4c988ed0e6a71218c2 (GCC 8 and 
> later) but older GCC versions are supported for building for some 
> architectures with _Float128 support, and if __builtin_fmaxf128 gets 
> referenced inside if (0) with older GCC it would be an undeclared 
> identifier.
>

Right, but currently we use declare_mgen_alias for fmax/fmin, so I am not
sure it we really need to take care if __builtin_fmaxfN support.

For _Float128 I will need to check if this does incur in any issue, I 
am currently building some older gcc toolchain to check it.