ENH: Implement string comparison ufuncs (or almost) (#21041)

* ENH: Implement string comparison ufuncs (or almost)

This makes all comparison operators and ufuncs work on strings
using the ufunc machinery.
It requires a half-manual "ufunc" to keep supporting void comparisons
and especially `np.compare_chararrays` (that one may have a bit more
overhead now).

In general the new code should be much faster, and has a lot of easier
optimization potential.  It is also much simpler since it can outsource
some complexities to the ufunc/iterator machinery.

This further fixes a couple of bugs with byte-swapped strings.

The backward compatibility related change is that using the normal
ufunc machinery means that string comparisons between string and
unicode now give a `FutureWarning` (instead of just False).

* MAINT: Do not use C99 tagged struct init in C++

C++ does not like it (at least not before C++20)... GCC and clang
don't seem to mind, but MSVC seems to.

* BENCH: Add basic string comparison benchmarks

* DOC,STY: Fixup string-comparisons comments based on review

Thanks to Marten's comments, a few clarfications and slight fixups.

* ENH: Use `memcmp` because it may be faster for the byte case

* TST: Improve string and unicode comparison tests.

* MAINT: Use switch statement based on review

As suggested be Serge.

Co-authored-by: Serge Guelton <serge.guelton@telecom-bretagne.eu>

* TST: Make unicode byte-swap test slightly more concrete

The issue is that the `view` needs to use native byte-order, so
just ensure native byte-order for the view, and then do another cast
to get it right.

* BUG: Add `np.compare_chararrays` to test and fix typo

* TST: Add test for empty string comparisons

* TST: Fixup string test based on martens review

* MAINT: Move definitions back into string_ufuncs.h

* MAINT: Use enum class for comparison operator templating

This removes the need for a dynamic (or static) assert in the
switch statement.

* Template version of add_loop to avoid redundant code

* STY: Fixup style, two spaces, error is -1

* STY: Small `string_ufuncs.cpp` fixups based on Serge's review

* MAINT: Fix merge conflict (ensure_dtype_nbo was removed)

Co-authored-by: Serge Guelton <serge.guelton@telecom-bretagne.eu>

1 files changed, 45 insertions, 0 deletions

diff --git a/benchmarks/benchmarks/bench_strings.py b/benchmarks/benchmarks/bench_strings.py
new file mode 100644
index 000000000..e500d7f3f
--- /dev/null
+++ b/benchmarks/benchmarks/bench_strings.py
@@ -0,0 +1,45 @@
+from __future__ import absolute_import, division, print_function
+
+from .common import Benchmark
+
+import numpy as np
+import operator
+
+
+_OPERATORS = {
+    '==': operator.eq,
+    '!=': operator.ne,
+    '<': operator.lt,
+    '<=': operator.le,
+    '>': operator.gt,
+    '>=': operator.ge,
+}
+
+
+class StringComparisons(Benchmark):
+    # Basic string comparison speed tests
+    params = [
+        [100, 10000, (1000, 20)],
+        ['U', 'S'],
+        [True, False],
+        ['==', '!=', '<', '<=', '>', '>=']]
+    param_names = ['shape', 'dtype', 'contig', 'operator']
+    int64 = np.dtype(np.int64)
+
+    def setup(self, shape, dtype, contig, operator):
+        self.arr = np.arange(np.prod(shape)).astype(dtype).reshape(shape)
+        self.arr_identical = self.arr.copy()
+        self.arr_different = self.arr[::-1].copy()
+
+        if not contig:
+            self.arr = self.arr[..., ::2]
+            self.arr_identical = self.arr_identical[..., ::2]
+            self.arr_different = self.arr_different[..., ::2]
+
+        self.operator = _OPERATORS[operator]
+
+    def time_compare_identical(self, shape, dtype, contig, operator):
+        self.operator(self.arr, self.arr_identical)
+
+    def time_compare_different(self, shape, dtype, contig, operator):
+        self.operator(self.arr, self.arr_different)