The author shares an experience with using the sqrtps intrinsic in code while discussing an optimization quirk with clang that results in precision issues between Intel and AMD CPUs. To ensure consistent instruction selection, inline assembly can be used. The author also introduces a workaround using __builtin_constant_p for constant folding optimization. Despite some limitations with GCC, the workaround effectively enhances constant folding. Surprisingly, if the function is marked as noinline, the code is optimized to eliminate branches. The author suggests a potential improvement in the GCC extension for easier optimization. Overall, the content delves deep into optimizing code for precise instruction selection and constant folding.
https://www.neilhenning.dev/posts/yourownconstantfolder/