[PATCH] D93377: [Clang] Add __ibm128 type to represent ppc_fp128

[Jonathan Wakely via Phabricator via cfe-commits]

jwakely added inline comments.


================
Comment at: clang/lib/Parse/ParseExprCXX.cpp:2245
+  case tok::kw___ibm128:
+    DS.SetTypeSpecType(DeclSpec::TST_ibm128, Loc, PrevSpec, DiagID, Policy);
+    break;
----------------
hubert.reinterpretcast wrote:
> jwakely wrote:
> > hubert.reinterpretcast wrote:
> > > qiucf wrote:
> > > > hubert.reinterpretcast wrote:
> > > > > qiucf wrote:
> > > > > > hubert.reinterpretcast wrote:
> > > > > > > Not sure what the best method is to implement this, but `long double` and `__ibm128` are the same type for GCC when `-mabi=ibmlongdouble` is in effect.
> > > > > > Seems clang is also different from GCC under `-mabi=ieeelongdouble`? I saw `__float128` and `long double` are the same for GCC but not for clang.
> > > > > Have you checked whether the new libstdc++ for which this support is being added needs the GCC behaviour to work properly?
> > > > > 
> > > > > The GCC behaviour allows the following to be compiled without introducing novel overload resolution tiebreakers:
> > > > > ```
> > > > > void f(__float128);
> > > > > void f(__ibm128);
> > > > > void f(int);
> > > > > 
> > > > > long double ld;
> > > > > 
> > > > > int main() { f(ld); }
> > > > > ```
> > > > As I saw both GCC and clang have error for ambiguous `operator<<` for:
> > > > 
> > > > ```
> > > >     std::cout << "long double:\n";
> > > >     std::cout << std::numeric_limits<long double>::max() << std::endl;
> > > >     std::cout << std::numeric_limits<long double>::min() << std::endl;
> > > > 
> > > >     std::cout << "__float128:\n";
> > > >     std::cout << std::numeric_limits<__float128>::max() << std::endl;
> > > >     std::cout << std::numeric_limits<__float128>::min() << std::endl;
> > > > 
> > > >     std::cout << "__ibm128:\n";
> > > >     std::cout << std::numeric_limits<__ibm128>::max() << std::endl;
> > > >     std::cout << std::numeric_limits<__ibm128>::min() << std::endl;
> > > > ```
> > > @jwakely, are the overload resolution errors expected? @qiucf, are you sure you have a sufficiently new libstdc++?
> > > @jwakely, are the overload resolution errors expected? 
> > 
> > Yes. Iostreams support `long double` but not `__float128`, unless that happens to be the same type as `long double` (due to a `-mabi=ieeelongdouble` option).
> Meaning that Clang's `__float128` iosteams support (with libstdc++) would diverge from GCC.
> 
> For example, Clang reports the call below as ambiguous even with `-mabi=ieeelongdouble`:
> ```
> void f(double);
> void f(long double);
> 
> void g(__float128 f128) { f(f128); }
> ```
> 
> https://godbolt.org/z/dhYEKa
> 
> Insofar as the user experience goes, is this lack of iostreams support for `__float128` (even with `-mabi=ieeelongdouble`) within the realm of the intended design of libstdc++?
The lack of iostreams support for `__float128` is the intended design.

On power we support `float`, `double` and three types of `long double`. If `__float128` is a distinct type from all those `long double` types it won't work.

GCC on power defines `__float128` as a macro expanding to `__ieee128`, so it is the same as one of the `long double` types.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D93377/new/

https://reviews.llvm.org/D93377