CPU: Work around Eigen tanh miscompilation for now

This will be revisited if a performant tanh is needed

src/ngraph/runtime/cpu/emitter.cpp

@@ -1620,16 +1620,18 @@ void Emitter::EMITTER_DECL(EmitTanh)
         (dynamic_pointer_cast<const TensorViewType>(n->get_arguments().at(0)->get_value_type()))
         ->get_element_type();
 
+    // Eigen's generic_fast_tanh_float<float> is currently miscompiled by Clang/LLVM
+    // so we fall-back to std::tanh
+    // TODO: Implement our own internal fast/approximate tanh if this actually gets used
+    // by models
     TU +=
         "    {\n"
-        "        auto arg0 = call_frame->get_tensor_view_data<" + element_type_names[TI(et)] + ">(" +
-        to_string(inputs[0].get_index()) + ");\n"
-        "        auto out  = call_frame->get_tensor_view_data<" + element_type_names[TI(et)] + ">(" +
-        to_string(outputs[0].get_index()) + ");\n"
-        "        EigenArray1d<" + element_type_names[TI(et)] + ">(out, "
-        EIGEN_VECTOR_FORMAT(outputs[0].get_layout<DenseTensorViewLayout>()->get_size()) ") =\n"
-        "        EigenArray1d<" + element_type_names[TI(et)] + ">(arg0, "
-        EIGEN_VECTOR_FORMAT(inputs[0].get_layout<DenseTensorViewLayout>()->get_size()) ").tanh();\n"
+        "        auto& arg0 = call_frame->get_parameterized_tensor_view<" + element_type_names[TI(et)] + ">(" +
+        to_string(inputs[0].get_index()) + ")->get_vector();\n"
+        "        auto& out  = call_frame->get_parameterized_tensor_view<" + element_type_names[TI(et)] + ">(" +
+        to_string(outputs[0].get_index()) + ")->get_vector();\n"
+        "        std::transform(arg0.begin(), arg0.end(), out.begin(), [](" + element_type_names[TI(et)] +
+        "::type x) -> " + element_type_names[TI(et)] + "::type { return std::tanh(x); });\n"
         "    }\n";
 }
 

src/ngraph/runtime/cpu/external_function.cpp

@@ -210,6 +210,8 @@ void ExternalFunction::compile(FunctionMap& function_map)
     TU += R"(// Generated by the NGraph CPU backend
 #include <memory>
 #include <vector>
+#include <algorithm>
+#include <cmath>
 
 #include <Eigen/Dense>