Many of our JIT templates end with a jump to the next opcode. This is generally needed to ensure correct control flow... but when the jump is the very last instruction, it's entirely unnecessary. Instead, we can just continue execution into the next instruction.

This removes these jumps for the following platforms (performance results):

• x86_64-pc-windows-msvc (2% faster)
• i686-pc-windows-msvc (3% slower)
• x86_64-apple-darwin (no benchmarks)
• aarch64-unknown-linux-gnu (no benchmarks)
• x86_64-unknown-linux-gnu (2% faster)

(aarch64-apple-darwin needs a bit more work, since the final jump is still made up of 3 instructions. That can come next.)

I'm assuming that the 32-bit Windows result is a fluke. If it's indeed the case that leaving the zero-length jumps in is faster, then we could easily remove this later by commenting out that case in remove_jump. If so, maybe it's an alignment thing and we should just pad with nops instead.

Also, this will probably be even more effective once we have hot/cold splitting of some kind. Most of the templates end with error handling or deoptimization code, and still have to jump over it to get to the next instruction.

• Issue: Improving JIT code quality #115802