#php-parser-comparison

Speed comparison of PHP parsers, run automatically in CI every 12 hours.

Each parser walks the same corpus — a freshly cloned Laravel framework with all Composer dependencies installed (src/ + vendor/) — and parses every .php file. Each tool runs 5 times and the average wall-clock time is reported, along with the peak memory (resident set size) of a single run.

#Parsers

#Latest results

Each run produces two tables — every parser pinned to a single core, vs all runner cores available.

#Single core (taskset -c 0)

Rank | Parser                        | Avg (5 runs) | Peak mem | vs slowest
   1 | nikic/php-parser (v5)         |    31121 ms |  211 MB |       1.0x
   2 | z7zmey/php-parser             |     6266 ms |   41 MB |       5.0x
   3 | rectorphp/php-parser-in-go    |     5900 ms |   95 MB |       5.3x
   4 | halleck45/go-php-parser       |     2450 ms |   75 MB |      12.7x
   5 | ext-ast                       |     1620 ms |   99 MB |      19.2x
   6 | mago-syntax (single-threaded) |     1201 ms |   73 MB |      25.9x

#All cores

Rank | Parser                        | Avg (5 runs) | Peak mem | vs slowest
   1 | nikic/php-parser (v5)         |    30574 ms |  211 MB |       1.0x
   2 | z7zmey/php-parser             |     4726 ms |   39 MB |       6.5x
   3 | rectorphp/php-parser-in-go    |     4718 ms |   73 MB |       6.5x
   4 | halleck45/go-php-parser       |     1700 ms |   65 MB |      18.0x
   5 | ext-ast                       |     1645 ms |   99 MB |      18.6x
   6 | mago-syntax (parallel)        |      561 ms |  159 MB |      54.5x

#Sample AST dump

Separate CI jobs (dump-*) parse one small fixture — sample-class.php, a ~25-line class — with each parser and print the resulting node tree to that job's Summary page. This shows the shape of each parser's output (each emits a different format) without wading through the full corpus. Run locally with make dump in any subproject.

Timings come from shared GitHub-hosted runners — good for rough ranking, not precise benchmarking. Live numbers appear in every run's Summary page.

Core count matters. The ubuntu-latest standard runner has only 4 vCPUs (16 GB RAM). How each parser reacts to extra cores:

• mago-syntax (parallel) — the only one that actually parses files across cores. Scales ~2.1x (1201→561 ms) and stays fastest in absolute terms.
• nikic, ext-ast — single-threaded PHP. Single-core and all-core numbers match.
• halleck45, z7zmey, rectorphp/php-parser-in-go — parse sequentially, but the Go runtime (GC, scheduler, sysmon) uses extra cores anyway, so pinning to one core (taskset -c 0) slows them down. The speedup tracks GOMAXPROCS, not the workload — neither does any parallel parsing:
  • halleck45 gains the most (~1.4x: 2450→1700 ms) — Go + cgo around an embedded PHP, so more runtime/allocation work to offload.
  • z7zmey is pure Go with less heap churn, so its gain is smaller (~1.3x: 6266→4726 ms).
  • rectorphp/php-parser-in-go shares the z7zmey lineage and behaves similarly (~1.25x: 5900→4718 ms).

Memory. The Go parsers are the leanest (z7zmey 39–41 MB, halleck45 65–75 MB, rectorphp/php-parser-in-go 73–95 MB); the PHP tools carry the interpreter's footprint (nikic ~211 MB, ext-ast ~99 MB). mago-syntax is tiny single-threaded (73 MB) but jumps to 159 MB in parallel mode — rayon buffering file contents across worker threads is the memory price of its speed.

Absolute numbers reflect a noisy-neighbour VM, not bare metal; only the relative ranking is meaningful, and even that can shift with runner contention.