Adding type info for f16c floating-point type

[ygao]

Summary

Hi,
The following patch adds support for half-precision floating point type supported by the clang front-end.
Native operations on this type is supported by OpenCL and to a limited extent by f16c (conversions are
supported); otherwise it is a storage-only type.
Could someone take a look whether this is good to go in?

• Gao.

[ygao]

Ping.

At least for x86-64, the half-precision floating-point type is mentioned in the ia64-abi in 2.9.2 Place of Emission:
Specifically, the run-time support library should contain type_info objects for the types X, X* and X const*, for every X in: void, std::nullptr_t, bool, wchar_t, char, unsigned char, signed char, short, unsigned short, int, unsigned int, long, unsigned long, long long, unsigned long long, float, double, long double, char16_t, char32_t, and the IEEE 754r decimal and half-precision floating point types.
(reference: http://mentorembedded.github.io/cxx-abi/abi.html#rtti)

[ygao]

Ping again.

Some additional information: Right now, only OpenCL supports this type in terms of language spec (plus
a number of shading languages, but clang does not support them), but AFAIK OpenCL does not support
exceptions. ARM and x86 have this type as an extension: storage-only type, not as argument or return
type. So maybe exception handling is not supported on those targets, i.e., one cannot throw a half-
precision floating point typed object.

It is not clear whether the typeid operator is supposed to work on this type.

__fp16 foo = 1.5f;

#include <typeinfo>
const char *get_type() {
  return typeid(foo).name();
}

Any comments or suggestions?

[ygao]

ping.

Although the type info object is provided by the runtime library, the runtime library relies
on clang to generate such a type object upon seeing a magic __fundamantal_type_info
struct, so one still has to patch the compiler if this type object is supposed to be there.
Which leads back to the question: should a type info object be generated for the 16-bit
floating-point type, so that the typeid operator will work? Any opinion?

(Reference: http://mentorembedded.github.io/cxx-abi/abi.html)

[ygao]

A fourth ping...

[ygao]

A fifth ping...

[rnk]

lgtm

This seems consistent with the ABI document: "Specifically, the run-time support library should contain type_info objects for the types ..., and the IEEE 754r decimal and half-precision floating point types."

lib/CodeGen/CGRTTI.cpp

                                   Context.UnsignedShortTy, Context.IntTy,
                                   Context.UnsignedIntTy, Context.LongTy, 
                                   Context.UnsignedLongTy, Context.LongLongTy, 
-                                  Context.UnsignedLongLongTy, Context.FloatTy,
+                                  Context.UnsignedLongLongTy,
+                                  Context.HalfTy, Context.FloatTy,
                                   Context.DoubleTy, Context.LongDoubleTy,
                                   Context.Char16Ty, Context.Char32Ty };
   for (unsigned i = 0; i < llvm::array_lengthof(FundamentalTypes); ++i)

test/CodeGenCXX/rtti-fundamental.cpp

 // CHECK: @_ZTIPy = unnamed_addr constant
 // CHECK: @_ZTIPKy = unnamed_addr constant
 
+// half
+// CHECK: @_ZTIDh = unnamed_addr constant
+// CHECK: @_ZTIPDh = unnamed_addr constant
+// CHECK: @_ZTIPKDh = unnamed_addr constant
+
 // float
 // CHECK: @_ZTIf = unnamed_addr constant
 // CHECK: @_ZTIPf = unnamed_addr constant