[PATCH] D49186: [SelectionDAG] add support for float16 libcalls

[Eli Friedman via Phabricator via llvm-commits]

efriedma added a comment.

Needs testcases.

Note that we generally can't accept patches which don't affect any in-tree target: they're impossible to test.



================
Comment at: include/llvm/CodeGen/RuntimeLibcalls.def:202
 HANDLE_LIBCALL(SIN_PPCF128, "sinl")
+HANDLE_LIBCALL(COS_F16, "cosh")
 HANDLE_LIBCALL(COS_F32, "cosf")
----------------
cosh is the hyperbolic cosine, which is not what you want here.


================
Comment at: lib/CodeGen/TargetLoweringBase.cpp:1153
+    TransformToType[MVT::f16] = MVT::i16;
+    ValueTypeActions.setTypeAction(MVT::f16, TypeSoftenFloat);
   }
----------------
This change isn't acceptable; you can't modify the behavior of other targets to call functions which don't exist.


Repository:
  rL LLVM

https://reviews.llvm.org/D49186