Just for fun: Which processors prefer sign-extended loads, and which prefer zero-extended loads?

Just for fun, let’s compare what happens on different processor architectures when you load a value from memory into a larger register. Possible responses are

• Unable: The processor does not support loading data into a larger register.
• Zero-extend: The processor zero-extends the small value to the large value.
• Sign-extend: The processor sign-extends the small value to the large value.
• Either: The processor has separate instructions for sign and zero extension.
• (blank): Does not apply because the processor doesn’t have registers larger than this.

Notes:

• The Alpha AXP with BWX is the only processor on the list where the extension mode changes between zero and sign, depending on size. This is a historical artifact: It inherited the “sign-extend” entry from the “Alpha AXP” row above it, but it turns out that bytes are usually zero-extended, so the BWX extension chose zero-extension for bytes. Words went along with bytes to minimize the weirdness.
• The x86-64 supports both zero-extended 32-to-64 loads and sign-extended 32-to-64 loads, but the zero-extended load offers a more compact encoding.

We can organize the processors by how they choose to extend small values to larger values: