Re: [PATCH v2 06/14] LoongArch: Generic <math.h> and soft-fp Routines

[Joseph Myers <joseph@codesourcery.com>]

On Tue, 4 Jan 2022, Adhemerval Zanella via Libc-alpha wrote:

> This is being replicated already on aarch64 and mips64, maybe we could
> add on a generic folder and use it instead.

We have sysdeps/ieee754/soft-fp/ for something similar, but that's only 
intended to be used via Implies in cases where software floating point 
with no exceptions / rounding modes support is in use and so the generic 
round-to-odd function implementations would be incorrect.

However, particular functions in that directory can be used via #include 
from elsewhere in sysdeps if appropriate - whether because of the lack of 
another generic implementation to use (as in the case of ldbl-128 sqrtl) 
or because it's more efficient (see how sysdeps/mips/ieee754/s_fmal.c 
includes sysdeps/ieee754/soft-fp/s_fmal.c because long double is always 
implemented in software for MIPS64).  So you could add sqrtl to that 
directory then #include it from individual architectures.  Or you could 
have e.g. ldbl-128-soft-fp that can go in sysdeps when appropriate, with 
the option to use it also for cases like MIPS fmal where the soft-fp 
implementation is more efficient than using the ldbl-128 version when all 
the underlying arithmetic operations used are themselves soft-fp 
operations.

-- 
Joseph S. Myers
joseph@codesourcery.com