That is indeed a cool trick. I know this argument has been done a million times by now, but I do think modern ISAs have gotten too complex and it would be good if someone made:

1. A simpler ISA that doesn't do all the backward compatible stuff.

2. A LLVM backend targeting the new ISA.

3. For the inevitable complainers who lost the source code 4 decades ago, still haven't gotten around to migrating to a newer system and who positively MUST keep running a binary from the eighties: A recompiler that takes machine code for the older x86(-64) ISA and outputs machine code for the new ISA, possibly with a slight performance penalty.

I bet CPU developers could leverage the circuitry freed up this way quite effectively into more effective branch predictors and bigger caches etc, and the total gains might be worth it even for the recompiled code.