Vectorize find-like algorithms for Clang for more types

[Alex Guteniev (AlexGuteniev)]

Resolves #5479.

⚙️ Optimization

• Use _Is_same_and_builtin_trivially_equality_comparable in another place. Make that part common, as it repeats.
• Adjust _Could_compare_equal_to_value_type to return true for the same types. This also handles std::byte concisely.
• Adjust casts in vectorized dispatch to use _Find_arg_cast. Which only widens or truncates integers, and uses _Bit_cast for the rest. Also makes pointers non-special.

That's it!

🏁 Benchmark

The benchmark shows that:

• The changes indeed dispatch to vectorized implementation, and
• Clang auto-vectorization either does not apply, or is less efficient than manual vectorization

Only find_and_count ought to be enough. It is sufficient to show that the dispatch works.

To see that other algorithms (search_n, replace, remove, remove_copy) also benefit from the change, use their existing benchmark, and set CXXFLAGS=/arch:AVX2 /D_USE_STD_VECTOR_ALGORITHMS=0 to imitate non-optimization on integer types, and then set CXXFLAGS=/arch:AVX2 /D_USE_STD_VECTOR_ALGORITHMS=1 to see the optimization results. The newly vectorized types should perform the same, as corresponding integer types.

⏱️ Benchmark results

It was run with set CXXFLAGS=/arch:AVX2 to compete with Clang auto-vectorization fairly: the algorithms detect AVX2 at runtime.

Benchmark	Before	After	Speedup
bm<point, not_highly_aligned_allocator, Op::FindSized>/8021/3056	749 ns	157 ns	4.77
bm<point, not_highly_aligned_allocator, Op::FindSized>/63/62	23.1 ns	3.66 ns	6.31
bm<point, not_highly_aligned_allocator, Op::FindSized>/31/30	7.58 ns	2.24 ns	3.38
bm<point, not_highly_aligned_allocator, Op::FindSized>/15/14	3.76 ns	1.72 ns	2.19
bm<point, not_highly_aligned_allocator, Op::FindSized>/7/6	1.81 ns	1.76 ns	1.03
bm<point, not_highly_aligned_allocator, Op::Count>/8021/3056	1174 ns	326 ns	3.60
bm<point, not_highly_aligned_allocator, Op::Count>/63/62	10.2 ns	4.03 ns	2.53
bm<point, not_highly_aligned_allocator, Op::Count>/31/30	5.58 ns	3.13 ns	1.78
bm<point, not_highly_aligned_allocator, Op::Count>/15/14	3.14 ns	2.89 ns	1.09
bm<point, not_highly_aligned_allocator, Op::Count>/7/6	2.20 ns	2.89 ns	0.76

[Stephan T. Lavavej (StephanTLavavej)]

Thanks! 😻 I really like the removal of all of the 64-bit/32-bit special cases for pointers, and the is_same generalization that supersedes checking for byte. I pushed a conflict-free merge with main and trivial comment cleanups.

I had to think for a moment why T{'0'} and T{'1'} were cromulent for the aggregate, but decided that it didn't need code changes or DMIs.

[Alex Guteniev (AlexGuteniev)]

I had to think for a moment why T{'0'} and T{'1'} were cromulent for the aggregate, but decided that it didn't need code changes or DMIs.

Ah, I initially added the template parameter to pass obviously-cromulent values, but in benchmark results it looked ugly and long, it made table cells multiline, so I abandoned that, preferring neat benchmark results over clarity, but didn't undo the template parameters change.

[Stephan T. Lavavej (StephanTLavavej)]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[Stephan T. Lavavej (StephanTLavavej)]

🎉 😻 🐈