Extend vectorized algorithms to ARM64EC

[StephanTLavavej]

Related to #813.

Depends on #2310.

In #2434, we found that vectorized find and count don't build for ARM64EC due to missing intrinsics. The existing attempt to enable reverse and reverse_copy for ARM64EC was also questionable, as Curtis J Bezault (@cbezault) found a broken SSE intrinsic. I'm going to disable the vectorized algorithms for ARM64EC, so this issue is for followup investigation.

• Can we get the broken SSE intrinsic fixed, so we can re-enable reverse and reverse_copy for ARM64EC?
• Should ARM64 NEON algorithms (Extend vectorized algorithms to ARM64/ARM64EC #813) supersede any attempt to get SSE codepaths working for ARM64EC?

[StephanTLavavej]

The missing instructions I saw were:

error LNK2019: unresolved external symbol BROADCAST128 referenced in function #BROADCAST128$exit_thunk (EC Symbol)
error LNK2019: unresolved external symbol PCMPEQ128 referenced in function #PCMPEQ128$exit_thunk (EC Symbol)
error LNK2019: unresolved external symbol PMOVMSKB128 referenced in function #PMOVMSKB128$exit_thunk (EC Symbol)
error LNK2019: unresolved external symbol POPCNT64 referenced in function #POPCNT64$exit_thunk (EC Symbol)
error LNK2019: unresolved external symbol PSHUFB128 referenced in function #PSHUFB128$exit_thunk (EC Symbol)
error LNK2019: unresolved external symbol PSHUFD128 referenced in function #PSHUFD128$exit_thunk (EC Symbol)

[AlexGuteniev]

Turns out that we miss linking to softintrin.lib.

Stephan T. Lavavej (@StephanTLavavej) however raised additional concern that emulated intrinsics implementation may not perform well.

We need to benchmark on actual hardware.

I think, that benchmarking one of the algorithms is not enough, we need to cover a simple case, and some cases with more tricky instructions where even some real CPUs (older AMDs, low power Intel cores) perform worse than others:

• swap_ranges as very baseline, having only loads and stores. if this fails to be efficient, then the whole thing may be pointless.
• reverse for various element bit widths for "swizzling" intrinsics
  • find_first_of for 32-bit and 64-bit elements can be additional exercise for swizzling
• find as an example of an usual vectorized search
• find_first_of for 8-bit and 16-bit elements for "text" intrinsic
  • additionally search for for 8-bit and 16-bit elements would be a good idea
• replace for 32-bit and 64-bit elements for masked loads/stores.

[AlexGuteniev]

May need to take care of alignment; see #5766

Edit: no, we don't. ARM64 handles misaligned well.

[StephanTLavavej]

Now that we've merged #6084, let's fuse this into #813.