[llvm] [AMDGPU] Add additional coverage tests for llvm.amdgcn.tanh, NFC (PR #202864)

[Matt Arsenault via llvm-commits]

================
@@ -106,6 +106,106 @@ define amdgpu_kernel void @tanh_undef_f32(ptr addrspace(1) %out) #1 {
   ret void
 }
 
+define void @tanh_f32_v(ptr addrspace(1) %out, float %src) {
----------------
arsenm wrote:

Use the return value instead of store to pointer. Also the scalar test should be changed to use a scalar return type with a shader calling convention, and not use amdgpu_kernel 

https://github.com/llvm/llvm-project/pull/202864