[PATCH] D25975: AMDGPU/SI: Make f16 a legal type for VI subtargets

[Tom Stellard via llvm-commits]

tstellarAMD added inline comments.


================
Comment at: include/llvm/IR/IntrinsicsAMDGPU.td:172-182
 def int_amdgcn_ldexp : Intrinsic<
-  [llvm_anyfloat_ty], [LLVMMatchType<0>, llvm_i32_ty], [IntrNoMem]
+  [llvm_anyfloat_ty], [LLVMMatchType<0>, llvm_anyint_ty], [IntrNoMem]
 >;
 
 def int_amdgcn_frexp_mant : Intrinsic<
   [llvm_anyfloat_ty], [LLVMMatchType<0>], [IntrNoMem]
 >;
----------------
Can these changes be done in a different patch?  They change the intrinsic signature, and will require updating any users of this intrinsic in external projects.


================
Comment at: lib/Target/AMDGPU/SIISelLowering.cpp:85-86
     addRegisterClass(MVT::i16, &AMDGPU::SReg_32RegClass);
+    addRegisterClass(MVT::f16, &AMDGPU::SReg_32RegClass);
+    addRegisterClass(MVT::f16, &AMDGPU::VGPR_32RegClass);
+  }
----------------
There is a 1-to-1 mapping between types and register classes, so you can drop the first addRegisterClass(MVT::f16, ... ) call


https://reviews.llvm.org/D25975