basic bytecode to machine code compiler (part 3)

[Rouslan Korneychuk]

My compiler now supports the x86-64 instruction set, in addition to x86.
It also generates faster x86 machine code.

Although it's designed to support 64-bit Windows, I have only tested it
on Linux so far, and it doesn't support running with Windows' DEP yet.

It's available at https://github.com/Rouslan/nativecompile

Here are the new benchmark results. The 64-bit code has a better margin
of improvement even though the Python interpreter is already faster in
64-bit mode.


The first test performs a quicksort on a list of 100 numbers, 5000
times. The second calculates all the prime numbers up to 10000000. Each
test is run three times in a row, first with the interpreter, then with
the compiled code.

#### SCRIPT ONE ####

import time
import random
import nativecompile

bcode = compile('''
def quicksort(array):
if len(array) <= 1:
return array
pindex = len(array)//2
pivot = array[pindex]
less = []
greater = []
for i,x in enumerate(array):
if i != pindex:
(less if x <= pivot else greater).append(x)
return quicksort(less) + [pivot] + quicksort(greater)

in_ = list(range(100))
random.seed(346097)
random.shuffle(in_)

t = time.clock()
for x in range(5000):
out = quicksort(in_)
t = time.clock()-t

assert out == sorted(in_)

print('execution time: {}'.format(round(t,10)))
''','<string>','exec')

mcode = nativecompile.compile(bcode)

print('byte code')
for x in range(3): eval(bcode)
print()

print('machine code')
for x in range(3): mcode()
print()

#### X86 OUTPUT ####

byte code
execution time: 1.76
execution time: 1.75
execution time: 1.77

machine code
execution time: 1.43
execution time: 1.43
execution time: 1.43

#### X86-64 OUTPUT ####

byte code
execution time: 1.49
execution time: 1.48
execution time: 1.48

machine code
execution time: 1.17
execution time: 1.17
execution time: 1.17



#### SCRIPT TWO ####

import time
import math
import nativecompile

bcode = compile('''
def primes_list(upto):
nums = [True] * (upto//2-1)

for i in range(3,math.floor(math.sqrt(upto))+1,2):
if nums[i//2-1]:
for j in range(i*3,upto,i*2):
nums[j//2-1] = False

primes = []
for i,n in enumerate(nums):
if n: primes.append((i+1)*2+1)

return primes

t = time.clock()
primes = primes_list(10000000)
t = time.clock()-t

print(primes[-1])
print('execution time: {}'.format(round(t,10)))
''','<string>','exec')

mcode = nativecompile.compile(bcode)

print('byte code')
for x in range(3): eval(bcode)
print()

print('machine code')
for x in range(3): mcode()
print()

#### X86 OUTPUT ####

byte code
9999991
execution time: 3.57
9999991
execution time: 3.34
9999991
execution time: 3.37

machine code
9999991
execution time: 2.73
9999991
execution time: 2.67
9999991
execution time: 2.73

#### X86-64 OUTPUT ####

byte code
9999991
execution time: 3.05
9999991
execution time: 3.24
9999991
execution time: 3.17

machine code
9999991
execution time: 2.15
9999991
execution time: 2.17
9999991
execution time: 2.17

 

[Ulrich Eckhardt]

if i != pindex:
(less if x <= pivot else greater).append(x)

Just curious, is there a reason why you wrote this last line that way
instead of using a "normal" if/else clause?


Cheers!

Uli

 

[Rouslan Korneychuk]

Just curious, is there a reason why you wrote this last line that way
instead of using a "normal" if/else clause?


Cheers!

Uli

No special reason. I just like concise code.