WIP perf(builtin): min, max, mean, median fast paths

[thevilledev]

Motivation

The min, max, mean, and median builtin functions were using reflection for every element in the input array, calling reflect.Value.Index(i).Interface() in a loop. This causes significant allocations (100+ per call for 100 elements) because each .Interface() call triggers heap allocations.

Changes

Avoid reflection and per-element allocations for common typed slices ([]int, []float64, []any) by adding type-switch fast paths that iterate directly without calling reflect.Value.Interface().

Added benchmarks:

• Benchmark_min
• Benchmark_max
• Benchmark_mean
• Benchmark_median

Benchstat results against master branch:

cpu: Apple M1
          │  master.out  │               fix.out               │
          │    sec/op    │   sec/op     vs base                │
_min-8      2850.5n ± 0%   427.7n ± 1%  -85.00% (p=0.000 n=10)
_max-8      2877.5n ± 1%   443.2n ± 0%  -84.60% (p=0.000 n=10)
_mean-8     2140.0n ± 1%   227.2n ± 0%  -89.38% (p=0.000 n=10)
_median-8    4.801µ ± 1%   1.360µ ± 1%  -71.67% (p=0.000 n=10)
geomean      3.030µ        491.9n       -83.76%

          │  master.out  │               fix.out                │
          │     B/op     │     B/op      vs base                │
_min-8      1648.00 ± 0%     48.00 ± 0%  -97.09% (p=0.000 n=10)
_max-8      1648.00 ± 0%     48.00 ± 0%  -97.09% (p=0.000 n=10)
_mean-8     1656.00 ± 0%     56.00 ± 0%  -96.62% (p=0.000 n=10)
_median-8   4.391Ki ± 0%   2.047Ki ± 0%  -53.38% (p=0.000 n=10)
geomean     2.071Ki          128.2       -93.95%

          │  master.out  │              fix.out               │
          │  allocs/op   │ allocs/op   vs base                │
_min-8      102.000 ± 0%   2.000 ± 0%  -98.04% (p=0.000 n=10)
_max-8      102.000 ± 0%   2.000 ± 0%  -98.04% (p=0.000 n=10)
_mean-8     103.000 ± 0%   3.000 ± 0%  -97.09% (p=0.000 n=10)
_median-8    211.00 ± 0%   11.00 ± 0%  -94.79% (p=0.000 n=10)
geomean       122.6        3.390       -97.24%

Further comments

Other typed slices (e.g., []int32, []uint64, custom numeric types) fall back to the original reflection-based path. Open to feedback whether we should include more typed slices here.

Similar optimizations could potentially be applied to other builtins like sum.

[antonmedv]

Nice! I like such optimizations!

[thevilledev]

Thanks! So much fun :)

Needs more tests and fine-tuning, but I'll mark it ready once there 🤞