Update comment as suggested by tra@google.com.

Eigen/src/Core/arch/GPU/Half.h

diff --git a/Eigen/src/Core/arch/GPU/Half.h b/Eigen/src/Core/arch/GPU/Half.h
index c510d3b..ffd4789 100644
--- a/Eigen/src/Core/arch/GPU/Half.h
+++ b/Eigen/src/Core/arch/GPU/Half.h
@@ -313,9 +313,9 @@
 
 #endif
 
-// The EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC clause is necessary on Mac where we
-// suspect that NVCC is doing device-side preprocessing with clang, but actually
-// compiling with cicc, which is obviously not clang.
+// We need to distinguish ‘clang as the CUDA compiler’ from ‘clang as the host compiler,
+// invoked by NVCC’ (e.g. on MacOS). The former needs to see both host and device implementation
+// of the functions, while the latter can only deal with one one of them.
 #if !defined(EIGEN_HAS_NATIVE_FP16) || (EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC) // Emulate support for half floats
 
 #if EIGEN_COMP_CLANG && defined(EIGEN_CUDACC)

Eigen/src/Core/arch/GPU/PacketMath.h

diff --git a/Eigen/src/Core/arch/GPU/PacketMath.h b/Eigen/src/Core/arch/GPU/PacketMath.h
index d097951..afc0860 100644
--- a/Eigen/src/Core/arch/GPU/PacketMath.h
+++ b/Eigen/src/Core/arch/GPU/PacketMath.h
@@ -102,9 +102,9 @@
   return make_double2(from, from);
 }
 
-// The EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC clause is necessary on Mac where we
-// suspect that NVCC is doing device-side preprocessing with clang, but actually
-// compiling with cicc, which is obviously not clang.
+// We need to distinguish ‘clang as the CUDA compiler’ from ‘clang as the host compiler,
+// invoked by NVCC’ (e.g. on MacOS). The former needs to see both host and device implementation
+// of the functions, while the latter can only deal with one one of them.
 #if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE) || (defined(EIGEN_CUDACC) && EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC)
 namespace {