Re: [PATCH 4/5] aarch/fpu: use generic sqrt, fma functions

[Vineet Gupta via Libc-alpha <libc-alpha@sourceware.org>]

On 6/1/20 9:38 AM, Adhemerval Zanella via Libc-alpha wrote:
> 
> 
> On 29/05/2020 23:00, Vineet Gupta wrote:
>> Signed-off-by: Vineet Gupta <vgupta@synopsys.com>
> 
> LGTM, some comments below.
> 
>> -#include <math_private.h>
>> -#include <libm-alias-finite.h>
>> -
>> -double
>> -__ieee754_sqrt (double d)
>> -{
>> -  return __builtin_sqrt (d);
>> -}
>> -libm_alias_finite (__ieee754_sqrt, __sqrt)
> 
> Ok.

How is one to test aarch64 port with hard-float. build-many-glibc doesn't have a
hf variant and hacking one didn't do the right thing either.

>> diff --git a/sysdeps/aarch64/fpu/math-use-builtins.h b/sysdeps/aarch64/fpu/math-use-builtins.h
>> new file mode 100644
>> index 000000000000..52f0a0dad6dd
>> --- /dev/null
>> +++ b/sysdeps/aarch64/fpu/math-use-builtins.h
>> @@ -0,0 +1,70 @@
>> +/* Using math gcc builtins instead of generic implementation.  aarch64 version.
>> +   Copyright (C) 2019-2020 Free Software Foundation, Inc.
> 
> I think it should be just 2020 here.

Fixed.

>> +
>> +/* Define these macros to 1 to use __builtin_xyz instead of the
>> +   generic implementation.  */
>> +#define USE_NEARBYINT_BUILTIN 0
>> +#define USE_NEARBYINTF_BUILTIN 0
>> +#define USE_NEARBYINTL_BUILTIN 0
>> +#define USE_NEARBYINTF128_BUILTIN 0
> 
> Since we are adding this new file for aarch64, we could also enable it fo
> nearbyint{f} and remove sysdeps/aarch64/fpu/s_nearbyint{f}.c as well.

OK, but only

+#define USE_NEARBYINT_BUILTIN 1
+#define USE_NEARBYINTF_BUILTIN 1

The other 2 are not defined currently. If they are, then I prefer they be enabled
as a separate commit for bisectability.


>> +#define USE_RINT_BUILTIN 0
>> +#define USE_RINTF_BUILTIN 0
>> +#define USE_RINTL_BUILTIN 0
>> +#define USE_RINTF128_BUILTIN 0
> 
> Ditto.

OK and again only RINT and RINTF.
Also I don't understand one thing. Both the generic and aarch64 code have this

float
__rintf (float x)
...
libm_alias_float (__rint, rint)

The alias arg 1 __rint seems to lack suffix 'f' ?


>> +
>> +#define USE_FLOOR_BUILTIN 0
>> +#define USE_FLOORF_BUILTIN 0

Again FLOOR, FLOORF only

>> +#define USE_FLOORL_BUILTIN 0
>> +#define USE_FLOORF128_BUILTIN 0
> 
> Ditto.

Ditto


>> +
>> +#define USE_CEIL_BUILTIN 0
>> +#define USE_CEILF_BUILTIN 0
>> +#define USE_CEILL_BUILTIN 0
>> +#define USE_CEILF128_BUILTIN 0
> 
> Ditto.

Ditto

> 
>> +
>> +#define USE_TRUNC_BUILTIN 0
>> +#define USE_TRUNCF_BUILTIN 0
>> +#define USE_TRUNCL_BUILTIN 0
>> +#define USE_TRUNCF128_BUILTIN 0
> 
> Ditto.

Ditto

>> +
>> +#define USE_ROUND_BUILTIN 0
>> +#define USE_ROUNDF_BUILTIN 0
>> +#define USE_ROUNDL_BUILTIN 0
>> +#define USE_ROUNDF128_BUILTIN 0
> 
> Ditto.

Ditto

>> +
>> +#define USE_COPYSIGNL_BUILTIN 1
>> +#if __GNUC_PREREQ (7, 0)
>> +# define USE_COPYSIGNF128_BUILTIN 1
>> +#else
>> +# define USE_COPYSIGNF128_BUILTIN 0
>> +#endif
>> +
> 
> It should be described in commit message as well (although generated
> instruction are essentially the same).

Well this one is even more special as it was already using the same code, except
that the math-use-builtins.h was generic vs. aarch64 specific one.

Thx
-Vineet