Are we Loong yet?

The patents expired a while ago, but they were also on some absolutely horrible instructions. They covered the load-word-left and load-word-right instructions. These were generated in pairs to do unaligned loads by doing a full-word load and shifting/masking the result into part of an architecture. They’re annoying because:

• The compiler must emit them for any load that might be unaligned.
• They keep an extra rename register alive during the load if you do a naïve implementation.
• To efficiently implement them, you typically want to fuse them into a single micro-op, which means that they are just a really inefficient encoding of an unaligned load.
• You need to implement them in the unfused case, because you may take a page fault between the two, so you have to do the shifting and masking in the LSU, though that’s a fairly small overhead. But the need to make the intermediate state architecturally visible does add complexity.
• For any load that doesn’t span a cache-line boundary, it’s easier to just do the load directly.
• For loads that do span a cahce-line (or, worse, page) boundary it’s not that hard to just lock two cache lines then do the load. And since it doesn’t need to be atomic you can do it sequentially, move the state, and not need to propagate the results to the architectural register file.

Just not doing these instructions makes a lot more sense on a complex microarchitecture.

It'd be nice to have an overview of what the emojis mean up front