Start benchmarking with tango

[kubouch]

Sets up some benchmarks using tango-bench.

Apart from total compile time, each stage is also benchmarked separately.

Result on my machine (AMD Ryzen 9 5950X)

> target/benchmarks/compiler/benchmarks solo -s 100 --warmup true
def_parse                                           [   3.8 us ...   3.8 us ...   3.9 us ]  stddev:  22.7 ns
def_resolve                                         [ 854.0 ns ... 863.0 ns ... 880.5 ns ]  stddev:   5.6 ns
def_typecheck                                       [ 686.6 ns ... 725.7 ns ... 753.4 ns ]  stddev:  13.5 ns
def_resolve_merge                                   [   1.0 us ...   1.0 us ...   1.1 us ]  stddev:   7.1 ns
def_typecheck_merge                                 [ 692.3 ns ... 765.9 ns ... 789.9 ns ]  stddev:  15.0 ns
def_compile                                         [   4.9 us ...   5.0 us ...   5.0 us ]  stddev:  19.0 ns
def_nu_old                                          [   3.2 us ...   3.3 us ...   3.4 us ]  stddev:  29.6 ns
if_parse                                            [   4.8 us ...   4.8 us ...   4.9 us ]  stddev:  28.0 ns
if_resolve                                          [ 445.6 ns ... 450.4 ns ... 458.1 ns ]  stddev:   2.5 ns
if_typecheck                                        [ 402.8 ns ... 408.8 ns ... 416.7 ns ]  stddev:   3.8 ns
if_resolve_merge                                    [ 567.8 ns ... 575.0 ns ... 605.2 ns ]  stddev:   5.4 ns
if_typecheck_merge                                  [ 426.1 ns ... 433.1 ns ... 442.9 ns ]  stddev:   3.9 ns
if_compile                                          [   5.5 us ...   5.6 us ...   5.7 us ]  stddev:  28.5 ns
if_nu_old                                           [   4.8 us ...   4.9 us ...   5.0 us ]  stddev:  40.5 ns
nu_old_empty                                        [ 317.7 ns ... 325.0 ns ... 333.8 ns ]  stddev:   3.0 ns

[fdncred]

This looks cool. It would be even cooler if we could run the same tests against nu-parser to see how new-nu-parser compares.

[kubouch]

I can try to add nu-parser benchmarks, good idea.

[kubouch]

Oops, so the new compiler appears slower, even in its incomplete implementation. I wonder why. Clearly, the parsing stage is much slower than everything else, that might be due to the constant peek() re-lexing.

[fdncred]

Another item, while we're talking about benchmarks, might be memory consumption since one of the points is to remove spans. Not sure how to get there exactly but it's something to think about.

[kubouch]

Good idea. I'll merge this at least, but let's keep digging.