Faster ensure_string_array

[WillAyd]

Benchmarks are still using 0.29.x, but I think this also prevents a regression from upgrading to 3.X as noted in #55179

| Change   | Before [f4f598fb] <main>   | After [b12de6c8] <faster-ensure-string>   |   Ratio | Benchmark (Parameter)                                                          |
|----------|----------------------------|-------------------------------------------|---------|--------------------------------------------------------------------------------|
| -        | 1.78±0.02ms                | 1.42±0.01ms                               |    0.8  | strings.Construction.time_construction('categorical_series', 'string[python]') |
| -        | 981±10μs                   | 694±6μs                                   |    0.71 | strings.Construction.time_construction('frame', 'string[python]')              |
| -        | 860±7μs                    | 436±4μs                                   |    0.51 | strings.Construction.time_construction('series', 'string[python]')             |

SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY.
PERFORMANCE INCREASED.

[WillAyd]

Also added a change to map_infer_mask which yields this improvement:

| Change   | Before [f4f598fb] <main>   | After [7a823afa] <faster-ensure-string>   |   Ratio | Benchmark (Parameter)                             |
|----------|----------------------------|-------------------------------------------|---------|---------------------------------------------------|
| -        | 11.7±0.6ms                 | 10.6±0.04ms                               |    0.91 | strings.Methods.time_isalnum('str')               |
| -        | 7.76±1ms                   | 7.04±0.03ms                               |    0.91 | strings.Methods.time_isnumeric('str')             |
| -        | 28.6±2ms                   | 26.0±0.5ms                                |    0.91 | strings.Methods.time_pad('string[pyarrow]')       |
| -        | 7.37±0.4ms                 | 6.62±0.05ms                               |    0.9  | strings.Methods.time_isspace('str')               |
| -        | 14.1±0.4ms                 | 12.7±0.2ms                                |    0.9  | strings.Methods.time_len('str')                   |
| -        | 11.4±0.6ms                 | 10.3±0.05ms                               |    0.9  | strings.Methods.time_startswith('string[python]') |
| -        | 8.11±0.9ms                 | 7.11±0.03ms                               |    0.88 | strings.Methods.time_isdigit('str')               |
| -        | 8.06±0.7ms                 | 6.96±0.03ms                               |    0.86 | strings.Methods.time_isdecimal('str')             |
| -        | 7.32±0.9ms                 | 6.30±0.02ms                               |    0.86 | strings.Methods.time_isdigit('string[python]')    |
| -        | 14.8±2ms                   | 12.5±0.1ms                                |    0.85 | strings.Methods.time_slice('string[python]')      |
| -        | 7.75±1ms                   | 6.32±0.03ms                               |    0.82 | strings.Methods.time_isnumeric('string[python]')  |
| -        | 15.1±2ms                   | 11.9±0.09ms                               |    0.79 | strings.Methods.time_slice('str')                 |

SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY.
PERFORMANCE INCREASED.

Generally seems like the pattern of:

cdef object arr = ...
cdef Py_ssize_t i = len(foo)
for i in range(i):
    ... = arr[i]

Gets slower from 0.29->3.x, but was pretty inefficient to begin with. By declaring as an array outside of the loop it should help our code base

[rhshadrach]

lgtm

[rhshadrach]

Thanks @WillAyd

[jbrockmendel]

can this check be done outside the loop?

[scoder]

Personally, I'd rather spell this for i, val in enumerate(arr):. Probably worth comparing if this gives the same performance.

[rhshadrach]

This patch may have induced a performance regression. If it was a necessary behavior change, this may have been expected and everything is okay.

Please check the links below. If any ASVs are parameterized, the combinations of parameters that a regression has been detected for appear as subbullets.

• libs.InferDtype.time_infer_dtype
  • dtype='np-object'
• libs.InferDtype.time_infer_dtype_skipna
  • dtype='np-object'

Subsequent benchmarks may have skipped some commits. The link below lists the commits that are between the two benchmark runs where the regression was identified.

28f274c...2a08b05