core: just in time compilation

[xNaCly]

Overview

The JIT compiler makes use of the plugin package. It compiles functions, either anonymous or named, and does so by generating go code, compiling this code into a go plugin and loading the compiled function from the plugin into the running process. Thus potentially improving performance for long-running or processing intensive Sophia scripts.

A rudimentary overview of the idea behind the JIT compiler in conjunction with the aforementioned plugin package can be found, here.

Current Performance improvements

All tests done with the currently supported features in examples/jit.phia and hyperfine without warmup runs:

hyperfine "./sophia examples/jit.phia" "./sophia -jit=0 examples/jit.phia"

An improvement prefixed with - is to be read as a decline in performance.

2024-04-11: Started

Benchmarks for commit f9bf294

Calls	Jit enabled	Jit disabled	Improvement (ms)	Improvement
100k	208ms	40ms	-168ms	-5x
200k	205ms	74ms	-131ms	-2.76x
300k	210ms	112ms	-98ms	-1.93x
400k	224ms	148ms	-76ms	-1.51x
500k	238ms	185ms	-53ms	-1.29x
600k	242ms	220ms	-22ms	-1.10x
700k	250ms	257ms	7ms	1.03x
800k	260ms	291ms	31ms	1.12x
900k	269ms	329ms	60ms	1.22x
1mio	276ms	362ms	86ms	1.32x
10mio	1.048s	3.618s	2,570ms (2.57s)	3.45x
50mio	4.461s	18.034s	13,573ms (13.573s)	4.04x

The break-even point when comparing the JIT and the AST walker lays at around 700k iterations.

2024-04-11: Paralleled compiler invocation

Benchmarks for commit eb3efbb

Calls	Jit enabled	Jit disabled	Improvement (ms)	Improvement
100000	38.6ms	38.1ms	-0.50	1.01x
200000	62.2ms	74.4ms	12.20	1.2x
300000	82.7ms	111ms	28.30	1.34x
400000	104.4ms	147.9ms	43.50	1.42x
500000	127.7ms	184.8ms	57.10	1.45x
600000	196.3ms	218.8ms	22.50	1.11x
700000	205.7ms	255.6ms	49.90	1.24x
800000	214.5ms	292.6ms	78.10	1.36x
900000	223.9ms	327.5ms	103.60	1.47x
1000000	230.6ms	364.3ms	133.70	1.58x
10000000	1,004ms	3,615ms	2,611.00	3.6x
50000000	4,430ms	18,130ms	13,700.00	4.09x

Comparing for function calls < 1 million:

Paralleling the compiler invocation improved the performance in comparison to the previous commit for all function call amounts. The best improvements were made for less function calls, such as 100k calls which saw an improvement of previously 200ms to now 38ms, the same can be seen for 200k, 300k, 400k and 500k iterations which were all around 100-150ms faster than they were in the previous commit.

2024-04-14: Arithmetics

Benchmarks for commit bbd91b5

With using all newly supported features:

;; jit currently supports
(fun t [a b] 
    (let add (+ a b))
    (let sub (- a b))
    (let mul (* a b))
    (let div (/ a b))
    (let mod (% a b))
    [add sub mul div mod])

(for [n] 100_000
    (t n n))

Calls	Jit enabled	Jit disabled	Improvement (ms)	Improvement
100,000.00	86.3	95.3	9.00	1.10x
200,000.00	151.8	187.6	35.80	1.23x
300,000.00	239.5	279.5	40.00	1.16x
400,000.00	266.4	370.6	104.20	1.39x
500,000.00	286.5	467.3	180.80	1.63x
600,000.00	316.5	558.8	242.30	1.76x
700,000.00	345.5	651.6	306.10	1.88x
800,000.00	366.2	745.1	378.90	2.03x
900,000.00	390	834.7	444.70	2.14x
1,000,000.00	412	931	519.00	2.25x
10,000,000.00	2,632	9,255	6,623.00	3.51x
50,000,000.00	12,450	46,350	33,900.00	3.72x

Progress

• constants
  • floats
  • strings
  • booleans
• variables
• arrays
• template strings
• arithmetic operators
  • addition
  • subtraction
  • multiplication
  • division
  • modulus
  • negation
• comparison operators
  • greater than
  • less than
  • equals
• boolean operators
  • and
  • or
  • negation
• function calls
• lambda functions
• function definition
• for loops
• if statements
• index
  • array
  • object
• match / enhanced switch
• merge
  • array
  • string
• return

Downsides

Downsides of the proposed JIT compiler include:

• the plugin package supports: Linux, macOS, BSD, not windows
• compiling a go plugin requires the go compiler tool chain to be present on the system the JIT operates on