[llvm] [AMDGPU] Narrow an int to fp source that fits in half its width (PR #223107)

[Matt Arsenault via llvm-commits]

================
@@ -1406,6 +1406,41 @@ define amdgpu_kernel void @s_sint_to_fp_32_sign_bits_to_f32(ptr addrspace(1) %ou
   ret void
 }
 
+define amdgpu_kernel void @s_sint_to_fp_trunc_nsw_i64_to_f32(ptr addrspace(1) %out, i64 %in) #0 {
----------------
arsenm wrote:

Maybe move these to a separate test file, this is a more specific optimization. Also test 16/32 bit cases?

https://github.com/llvm/llvm-project/pull/223107