Following on from a PHP 8/JIT benchmark on twitter, I decided to have a look myself.

I've picked an example that I know speeds up really well when reimplementing it in C. I wrote about this RDP algorithm some time ago.

What it does is to take a line of geospatial points (lon/lat coordinates), and simplifies it. It's my go-to example to show raw algorithmic performance, which is probably the best place to use a JIT for non-trivial code. I actually use this in production.

With PHP 7.4:

$ pe 7.4dev; time php -n \
        -dzend_extension=opcache -dopcache.enable=1 -dopcache.enable_cli=1 \
        -dopcache.jit=1235 -dopcache.jit_buffer_size=64M \
        bench-rdp.php 1000
Using array (
  0 => 'RDP',
  1 => 'simplify',
)

real    0m8.778s
user    0m8.630s
sys     0m0.117s

(I realise that the opcache arguments do nothing on the command line here). This runs RDP::simplify (my PHP implementation) 1000 times in about 8 seconds.

With PHP 8.0 and JIT:

$ pe trunk; time php -n \
        -dzend_extension=opcache -dopcache.enable=1 -dopcache.enable_cli=1 \
        -dopcache.jit=1235 -dopcache.jit_buffer_size=64M \
        bench-rdp.php 1000
Using array (
  0 => 'RDP',
  1 => 'simplify',
)

real    0m4.640s
user    0m4.627s
sys     0m0.008s

It jumps from ~8.8s to ~4.6s, a reduction in time of ~4.2s (or 48%), which is pretty good.

Now if I run the same with the geospatial extension which has a C implementation.

With PHP 7.4 and the extension:

$ pe 7.4dev; time php -n -dextension=geospatial \
        -dzend_extension=opcache -dopcache.enable=1 -dopcache.enable_cli=1 \
        -dopcache.jit=1235 -dopcache.jit_buffer_size=64M bench-rdp.php 1000
Using 'rdp_simplify'

real    0m0.695s
user    0m0.675s
sys     0m0.021s

Which gives a reduction in speed compared to PHP 7.4 of ~8.1s (or 92%).

So it looks like the JIT does do some good work for something that's highly optimisable, but still nowhere near what an implementation in C could do.

The code that I used is in this Gist.

This ran on a 4th gen ThinkPad X1 Carbon, making sure my CPU was pinned at its maximum speed of 3.3Ghz. Although I've pasted only one result for each, I did run them several times with very close outcomes.