amzn
Evaluate whether switching `float` to little endian in Ion 1.1 is worthwhile
Our int encodings (including primitives) have switched to little endian in Ion 1.1 for a performance boost. Floats also rely on the CPU's native endianness, but the current version of the spec leaves them big endian.
For reference, here's the assembly generated by the conversion to LE bytes and to BE bytes on x86_64:
Depending on your perspective, that's either a trivial difference or it's 100% more instructions! 😛
It does look pretty trivial (unless, perhaps, you're encoding lots of floats in a tight loop). We should see if there's a significant performance difference in Java (and maybe some other languages) to make a more informed decision.
Per team discussion, we're leaning little endian for consistency. Java's ByteBuffer has an API for swapping endianness.
@zslayton, I'm going to close this as completed. Feel free to re-open if you disagree.