Choose a RISC subset of x86_64 about mijit HOT 5 CLOSED

If we are content always to assemble a SIB byte when accessing memory, the differences between R12 and other registers vanishes. This adds one byte of code per memory access. A good trade?

We would still need to special-case SP-relative and program-relative addressing.

from mijit.

Would it be too terrible just to emit the extra byte when R12 is used?

from mijit.

It wouldn't be terrible, no. It would be hidden in a black box, and it's probably less work than e.g. making separate methods for assembling R12-relative memory accesses. Maybe we should just do that.

Especially since we could do the same for SP-relative addressing. I recently realized that we're going to have to assemble instructions to add constants to SP; they could be unified too if SP is just another register. There are monsters in push RSP and push [RSP], and similarly for pop, but I suppose it's not my problem if callers engage those monsters.

By the way, we could save way more bytes by using 8-bit immediate constants where possible. I've been ignoring that optimization until now, so I should probably rethink what is and is not worth doing. Maybe I should just ignore R12 completely until the ambient ambition level rises.

from mijit.

Based on today's work, I've added multiplication, division and conditional moves to the list of requirements. Conditional moves are needed for maximum and minimum operations, and also for absolute value.

from mijit.

Closing in favour of #8, #9 and #10.

from mijit.