Added clang inline helper

[WillAyd]

gcc on macOS I think executes clang more often than not, and it appears we didn't have a inline helper defined for that. This is ported from CYTHON_INLINE which essentially does the same in cythonized files, save the static declaration

This might add a bit of build time to macOS (I think partially explains why it's a few minutes faster than the rest) but makes it more similar in performance to the others

Running the full benchmark suite to see where this may help

[jbrockmendel]

makes sense, looking forward to the perf stats

[bashtage]

Doesn't clang define __GNUC__?

[WillAyd]

From what I understand clang defines it but the semantics of the inline keyword are different

[bashtage]

Have you checked whether binaries produced before this patch and with this patch are different? I dont' see why __attribute__ ((__unused__)) would make a difference to the compiler's choice to inline a function. I think it suppressed some warnings.

[WillAyd]

I tried running the entire benchmark suite but got stuck at 30% somewhere over night, so re-ran just for indexing this morning. Here are the results:

       before           after         ratio
     [8ff2ebd9]       [77a96005]
     <axis-cleanup>       <clang-inline>
-      5.05±0.2ms      4.31±0.03ms     0.85  indexing_engines.ObjectEngineIndexing.time_get_loc('non_monotonic')
-         179±2μs          151±7μs     0.85  indexing.NumericSeriesIndexing.time_getitem_scalar(<class 'pandas.core.indexes.numeric.Int64Index'>, 'nonunique_monotonic_inc')
-        24.4±1ms       17.1±0.5ms     0.70  indexing.NumericSeriesIndexing.time_loc_scalar(<class 'pandas.core.indexes.numeric.UInt64Index'>, 'nonunique_monotonic_inc')
-        44.4±3ms         26.0±2ms     0.59  indexing.InsertColumns.time_assign_with_setitem

SOME BENCHMARKS HAVE CHANGED SIGNIFICANTLY.
PERFORMANCE INCREASED.

So I think a measurable boost in some places

[jbrockmendel]

any firm conclusions about the perf impact?

[WillAyd]

Yea looks to give a boost to indexing operations. I didn't run the entire test suite because this inline helper is already generated by Cython so already used in anything but our custom C code

[WillAyd]

Planning to merge at the end of the day tomorrow unless any other comments @pandas-dev/pandas-core

[bashtage]

That is part of this, but the MSVC inline could become static inline.

Do any of these compilers have different behavior for vendor-specific keywords rather than C99 static inline?

[bashtage]

Should

https://github.com/pandas-dev/pandas/pull/30990/files#diff-b3aa9f22c2ca2e89ba79863c5701aafcL18

be first?

GNUC says: "If you are writing a header file to be included in ISO C90 programs, write inline instead of inline. See Alternate Keywords."

[WillAyd]

Not sure, but I don't know if worth straying from what Cython generates for this in any case

[jreback]

i guess this can't hurt.