Reduce allocations

[mk6i]

@crocodele In this branch AST nodes are retrieved from a stack-allocated array instead of dynamic allocation. When I run the benchmarks on the branch, the JsonPathPhpExtensionBenchmark beats NativePhpBenchmark on benchMediumDataset. Do you get the same results?

[mk6i]

Would having opcache enabled close the performance gap between the implementations, thus rendering the extension moot? IIRC we would need to run this with PHP-FPM to test this out. (It's been like 5 years since I last developed with PHP, so I might be wrong.)

[crocodele]

When I run the benchmarks on the branch, the JsonPathPhpExtensionBenchmark beats NativePhpBenchmark on benchMediumDataset. Do you get the same results?

Hmm, no, I get pretty much the same numbers on refactor/reduce_alloc as on main. 🤔

The NativePhpBenchmark performance seems to depend on the system though to some extent. On my main machine I get values around 240 microseconds for benchMediumDataset, on another machine around 360 microseconds, which is slightly less than JsonPathPhpExtensionBenchmark on both machines.

Would having opcache enabled close the performance gap between the implementations, thus rendering the extension moot?

That's a good point. I ran the benchmarks again with opcache.enable_cli = 1. In the medium benchmark it cut 45% and 15% off the execution time for the two tested PHP libraries. The extension is still a lot faster though. No improvement from OPCache in the tiny or huge benchmarks.

[mk6i]

The NativePhpBenchmark performance seems to depend on the system though to some extent. On my main machine I get values around 240 microseconds for benchMediumDataset, on another machine around 360 microseconds, which is slightly less than JsonPathPhpExtensionBenchmark on both machines.

Interesting, did you try with both PHP 7.4 and 8?

[mk6i]

@crocodele I reverted the buggy dynamic memory pool. See commit message in 7a313cd for a summary of the changes.