Re: [PATCH v3] LoongArch:Implement 128-bit floating point functions in gcc.

[Xi Ruoyao <xry111@xry111.site>]

On Fri, 2023-08-18 at 15:05 +0800, Xi Ruoyao via Gcc-patches wrote:
> On Fri, 2023-08-18 at 14:58 +0800, Xi Ruoyao via Gcc-patches wrote:
> > On Fri, 2023-08-18 at 14:39 +0800, chenxiaolong wrote:
> > > 在 2023-08-17四的 15:08 +0000，Joseph Myers写道：
> > > > On Thu, 17 Aug 2023, Xi Ruoyao via Gcc-patches wrote:
> > > > 
> > > > > So I guess we just need
> > > > > 
> > > > > builtin_define ("__builtin_fabsq=__builtin_fabsf128");
> > > > > builtin_define ("__builtin_nanq=__builtin_nanf128");
> > > > > 
> > > > > etc. to map the "q" builtins to "f128" builtins if we really need
> > > > > the
> > > > > "q" builtins.
> > > > > 
> > > > > Joseph: the problem here is many customers of LoongArch CPUs wish
> > > > > to
> > > > > compile their old code with minimal change.  Is it acceptable to
> > > > > add
> > > > > these builtin_define's like rs6000-c.cc?  Note "a new architecture"
> > > > > does
> > > > > not mean we'll only compile post-C2x-era programs onto it.
> > > > 
> > > > The powerpc support for __float128 started in GCC 6, predating the
> > > > support 
> > > > for _FloatN type names, built-in functions etc. in GCC 7 - that's
> > > > why 
> > > > there's such backwards compatibility support there.  That name only
> > > > exists 
> > > > on a few architectures.
> > > > 
> > > > If people really want to compile code using the old __float128 names
> > > > for 
> > > > LoongArch I suppose you could have such #defines, but it would be
> > > > better 
> > > > for people to make their code use the standard names (as supported
> > > > from 
> > > > GCC 7 onwards, though only from GCC 13 in C++) and then put
> > > > backwards 
> > > > compatibility in their code for using the __float128 names if they
> > > > want to 
> > > > support the type with older GCC (GCC 6 or before for C; GCC 12 or
> > > > before 
> > > > for C++) on x86_64 / i386 / powerpc / ia64.  Such backwards
> > > > compatibility 
> > > > in user code is more likely to be relevant for C++ than for C, given
> > > > how 
> > > > the C++ support was added to GCC much more recently.  (Note: I
> > > > haven't 
> > > > checked when other compilers added support for the _Float128 name or
> > > > associated built-in functions, whether for C or for C++, which might
> > > > also 
> > > > affect when user code wants such compatibility.)
> > > > 
> > > Thank you for your valuable comments. On the LoongArch architecture,
> > > the "__float128" type is associated with float128_type_node and the "q"
> > > suffix function is mapped to the "f128" function. This allows
> > > compatibility with both "__float128" and "_Float128" types in the GCC
> > > compiler. The new code is modified as follows:
> > >   Add the following to the loongarch-builtins.c file:
> > > +lang_hooks.types.register_builtin_type (float128_type_node,
> > > "__float128");
> > >   Add the following to the loongarch-c.c file:
> > > +builtin_define ("__builtin_fabsq=__builtin_fabsf128");
> > > +builtin_define ("__builtin_copysignq=__builtin_copysignf128");
> > > +builtin_define ("__builtin_nanq=__builtin_nanf128");
> > > +builtin_define ("__builtin_nansq=__builtin_nansf128");
> > > +builtin_define ("__builtin_infq=__builtin_inff128");
> > > +builtin_define ("__builtin_huge_valq=__builtin_huge_valf128");
> > > 
> > >  The regression tests of the six functions were added without problems.
> > > However, the implementation of the __builtin_nansq() function does not
> > > get the result we want. The questions are as follows:
> > >  x86_64:
> > >     _Float128 ret=__builtin_nansf128("NAN");
> > > 
> > >     compiled to (with gcc test.c -O2 ):
> > >                 .cfi_offset 1, -8
> > >         bl      %plt(__builtin_nansf128)
> > >         ..
> > >  LoongArch:
> > >     _Float128 ret=__builtin_nansf128("NAN");
> > >       compiled to (with gcc test.c -O2 ):
> > >         .cfi_offset 1, -8
> > >         bl      %plt(__builtin_nansf128)
> > 
> > It seems wrong.  It should be "bl %plt(nansf128)" instead, without the
> > __builtin_ prefix so the implementation in libm (from Glibc) will be
> > used instead.  AFAIK __builtin_nan and __builtin_nans are rarely called
> > with a non-empty tagp so it's not worthy to inline the implementation
> > for non-empty tagp here.
> > 
> > The same issue happens on x86_64:
> > 
> > call    __builtin_nansf128@PLT
> > 
> > __builtin_nanf128 compiles correct:
> > 
> > call    nanf128@PLT
> > 
> > I'll see if there is a ticket in https://gcc.gnu.org/bugzilla.  If not
> > I'll create one.

https://gcc.gnu.org/PR111058

> Alright, Glibc does not have a "nansf128" function yet.  Actually there
> is even no "nans" function for the plain double type.  So even a plain
> __builtin_nans("114") won't work too.
> 
> If we'll fix this, we need to do it in a generic, target-independent way
> (i. e. fix it all at once for all targets).
> 
> So for now, and for LoongArch specific code, the proper thing to do is
> aliasing float128_type_node as __float128 and the six
> __builtin_define's.
> 
> Please commit them to trunk if regression test passes.  You need to also
> add LoongArch as a target supporting __float128 in extend.texi.

-- 
Xi Ruoyao <xry111@xry111.site>
School of Aerospace Science and Technology, Xidian University