starlark: built-in types map of *Builtin

[emcfarlane]

NewBuiltin and Sort are called at init. Slightly faster method calls:

benchmark                            old ns/op     new ns/op     delta
BenchmarkStringHash/hard-1-8         3.90          3.65          -6.41%
BenchmarkStringHash/soft-1-8         2.69          2.21          -17.84%
BenchmarkStringHash/hard-2-8         4.52          4.38          -3.10%
BenchmarkStringHash/soft-2-8         3.55          2.68          -24.51%
BenchmarkStringHash/hard-4-8         5.32          5.07          -4.70%
BenchmarkStringHash/soft-4-8         5.32          3.63          -31.77%
BenchmarkStringHash/hard-8-8         7.58          7.03          -7.26%
BenchmarkStringHash/soft-8-8         6.62          5.60          -15.41%
BenchmarkStringHash/hard-16-8        7.00          6.43          -8.14%
BenchmarkStringHash/soft-16-8        13.1          9.44          -27.94%
BenchmarkStringHash/hard-32-8        7.28          7.12          -2.20%
BenchmarkStringHash/soft-32-8        21.7          21.2          -2.30%
BenchmarkStringHash/hard-64-8        9.13          8.75          -4.16%
BenchmarkStringHash/soft-64-8        50.9          54.1          +6.29%
BenchmarkStringHash/hard-128-8       13.0          12.8          -1.54%
BenchmarkStringHash/soft-128-8       114           112           -1.75%
BenchmarkStringHash/hard-256-8       15.7          15.3          -2.55%
BenchmarkStringHash/soft-256-8       241           235           -2.49%
BenchmarkStringHash/hard-512-8       23.7          23.2          -2.11%
BenchmarkStringHash/soft-512-8       492           483           -1.83%
BenchmarkStringHash/hard-1024-8      38.8          38.5          -0.77%
BenchmarkStringHash/soft-1024-8      987           981           -0.61%
BenchmarkHashtable-8                 2152734       2163910       +0.52%
Benchmark/bench_bigint-8             271664        269095        -0.95%
Benchmark/bench_builtin_method-8     392849        354319        -9.81%
Benchmark/bench_calling-8            643982        638527        -0.85%
Benchmark/bench_int-8                124893        123253        -1.31%
Benchmark/bench_range-8              308           301           -2.27%
BenchmarkProgram/read-8              17762         17676         -0.48%
BenchmarkProgram/compile-8           269868        268148        -0.64%
BenchmarkProgram/encode-8            11632         11434         -1.70%
BenchmarkProgram/decode-8            13839         13554         -2.06%
BenchmarkScan-8                      632375        629915        -0.39%
BenchmarkParse-8                     1310575       1305860       -0.36%

[alandonovan]

Thanks for this change. I'm curious: do you actually need faster method calls, or was this just for fun?

The code looks sound. However, I wonder if it's not a little more complex than necessary. Specifically, I don't think we need to optimize the AttrNames operation, which is used only for dir(x), whose performance is not so important, I think. If we revert it to the old AttrNames approach of allocating a new slice and sorting the map keys, then the remainder of the change is the simplification of method closures to avoid the two allocations (one for the unnecessary change-of-signature wrapper, and one for Builtin). We could then replace 'struct builtins' with a simple map[string]*Builtin.

Thoughts?

[emcfarlane]

Just for fun! I noticed the double allocation while trying to implement built ins. I’m happy to go with the simplified *Builtin.

[alandonovan]

I like this even better. Thanks!