starlark: optimized dispatch of built-in methods

[adonovan]

This PR is a sketch of an alternative approach to supporting faster dispatch of calls to built-in methods by avoiding the allocation ordinarily performed by Builtin.BindReceiver when a method is accessed and then immediately called as in x.f().

The bound Builtin is still created for g=x.f; g(), or for values of x that do not implement the new HasBuiltinMethods interface.

Details:

• Adds HasBuiltinMethods interface implemented by all the interpreter's built-in types.

• Builtin now has two constructors: the legacy NewBuiltin, and a new one (NewBuiltinMethod) that takes a function of this form

  func(thread *Thread, name string, recv Value, args Tuple, kwargs []Tuple) (Value, error)
  -----------------------
  avoiding the need to allocate a Builtin with BindReceiver. The legacy one wraps the newone, possibly at an unacceptable allocation cost since it calls BindReceiver on every legacy call. The recv argument is nil for non-method calls.

• x.f(...) is compiled to recv, callable = ATTR_METHOD(x, "f") res = CALL(recv, callable, ...) with a special CALL flag indicating that a receiver is provided because we used ATTR_METHOD. If callable is an unbound Builtin, the recv operand is used; otherwise the Builtin supplies the receiver.

WORK IN PROGRESS. Some parts are slop.

[adonovan]

Here's the benchstat delta. The big gains are on very unnatural microbenchmarks, so it seems like a lot of trouble for little gain.

   Benchmark                                  |  HEAD~1  ( sec/op ) |  HEAD  ( sec/op ) |    Time Delta     |  HEAD~1  ( allocs/op ) |  HEAD  ( allocs/op ) |   Allocs Delta
  --------------------------------------------|---------------------|-------------------|-------------------|------------------------|----------------------|-------------------
    Starlark/bench_builtin_method-8           |      107.43 µs      |     93.65 µs      |      -12.83%      |          2001          |         1001         |      -49.98%
    Starlark/bench_string_startswith-8        |      125.40 ns      |     93.78 ns      |      -25.22%      |           3            |          1           |      -66.67%
    Starlark/bench_issubset_duplicate_same-8  |      25.99 µs       |     24.74 µs      |      -4.80%       |           6            |          5           |      -16.67%
    Starlark/bench_issubset_unique_same-8     |      29.49 µs       |     28.83 µs      |      -2.22%       |           8            |          7           |      -12.50%
    Starlark/bench_dict_equal-8               |      25.93 µs       |     25.57 µs      |      -1.37%       |           0            |          0           |         ~
    Starlark/bench_calling-8                  |      176.0 µs       |     178.7 µs      |   ~ ( p=0.151 )   |          303           |         303          |         ~
    Starlark/bench_mix-8                      |      60.28 µs       |     61.88 µs      |      +2.66%       |          483           |         483          |         ~
    Starlark/bench_bigint-8                   |      105.4 µs       |     110.5 µs      |      +4.85%       |          5006          |         5006         |         ~
    Starlark/bench_gauss-8                    |      5.341 ms       |     5.534 ms      |   ~ ( p=0.222 )   |         132320         |        132320        |         ~
    Program/compile-8                         |      152.8 µs       |     152.0 µs      |   ~ ( p=0.841 )   |          1596          |         1596         |         ~
    Program/read-8                            |      43.15 µs       |     42.33 µs      |   ~ ( p=0.421 )   |           5            |          5           |         ~
   Geomean across all benchmarks              |                     |                   |      -0.40%       |                        |                      |      -5.45%