Post #2926564

sometimes GCC is pretty clever. it recognized that i’m calling atan2f element-wise on a SIMD vector type and uses the libmvec function instead of plain libm accordingly, reducing a lot of expensive scalar extraction and function call instructions to just this one clang is not that clever. it has a __builtin_elementwise_atan2 but that just uses the non-vectorized version. it also has no clean way to do masked shuffles because its shuffle builtin only takes scalar indices