This directory contains additional chip and/or compiler designs, built using the same tools from the nand package.
Each module contains all the required pieces to simulate a chip and translate and run VM programs on it.
See each module for instructions.
Four alternative implementations use more or less chip hardware to make programs run faster, or to fit larger programs in ROM:
alt/sp.py adds instructions for pushing/popping values to/from the stack, making programs more compact and efficient.
alt/threaded.py adds lightweight CALL/RTN instructions, enabling a very compact "threaded interpreter" translation, which runs a little slower.
alt/shift.py adds a "shiftr" instruction, and rewrites "push constant 16; call Math.divide" to use it instead; also a more efficient Math.multiply using shiftr.
alt/eight.py is, finally, a smaller CPU, by using an 8-bit ALU and 2 cycles per instruction.
alt/big.py has a single, flat memory space, with maximum RAM and the ability to read data from ROM (and code from RAM.) This is much more flexible and realistic, but adds a cycle to fetch each instruction from the shared memory system. Moving static data to ROM can dramtically improve code size and performance, but because the computer uses character-mode graphics, these metrics don't provide a direct comparison (even if you did port the VM and OS, which I haven't.) This architecture is intended to support more sophisticated languages (e.g. BASIC, Scheme, or Forth), and interactive programming.
These implementations all use the standard CPU, and try to generate more efficient code for it:
alt/lazy.py has a slighty cleverer translator which avoids updating the stack when that's easy to do.
alt/reg.py is a much more ambititous compiler which uses the "registers" at locations 5-12 for transient local variables and expression evaluation, reserving the stack only for subroutine calls and locals that cross them.
alt/reduce.py adds an optimization phase after parsing and before the normal compiler runs, which replaces certain function calls with lower-overhead "reduced" alternatives.
alt/scheme provides a compiler and REPL for the Scheme language (circa R4RS), using the "big" architecture.
| Location | Nands | ROM size | Cycles per frame | Cycles for init |
|---|---|---|---|---|
| project_0*.py | 1,262 | 25,700 | 41,450 | 129,200 |
| alt/sp.py | 1,844 (+46%) | 14,150 (-45%) | 27,440 (-34%) | 76,240 (-41%) |
| alt/threaded.py | 1,549 (+23%) | 8,100 (-68%) | 49,600 (+20%) | 173,750 (+34%) |
| alt/shift.py | 1,311 (+4%) | 26,050 (+1%) | 19,800 (-52%) | same |
| alt/eight.py | 1,032 (-18%) | same | +100% | +100% |
| alt/big.py | 1,448 (+14%) | ? | ? | ? |
| alt/lazy.py | same | 23,650 (-8%) | 37,300 (-10%) | 111,000 (-14%) |
| alt/reg.py | same | 18,200 (-29%) | 12,450 (-70%) | 55,250 (-57%) |
| alt/reduce.py | same | 27,350 (+6.5%) | 20,300 (-51%) | same |
ROM Size is the total number of instructions in ROM when Pong is compiled and translated from the Jack source.
Cycles per frame is the number of cycles to run the first iteration of the Pong game loop.
Specifically, this includes Bat.move and PongGame.moveBall, and not Screen.clearScreen
or Sys.wait.
Cycles for init is the number of cycles from start to reaching call Main.main 0. That is,
the number of cycles to execute all of the common setup code in Sys.init.
The measurements in the table are all produced by alt/compare.py, slightly cleaned up.
Note: reduce.py and shift.py produce similar improvements, because they both optimize the important case of division by a constant equal to 2^n. It's interesting that a similar improvement is achieved with either 5% more chip area or 6% more ROM space, but that comparison isn't entirely fair for various reasons.
Furthermore, because reduce.py's optimizations run so early, they can actually be combined with any of the other implementations for a compound effect. In particular, when reduce.py introduces additional temporary variables and removes function calls, reg.py will handle those changes especially well and that combination comes closest to emulating a "modern" optimizing compiler. Those results aren't shown here.