faster xor and != for Bool

[stevengj]

See discourse. Before this change:

julia> b = rand(Bool, 150); i = UInt8.(b);

julia> f(a) = reduce(xor, a); g(a) = reduce(!=, a);

julia> @btime f($b); @btime g($b);
  84.023 ns (0 allocations: 0 bytes)
  92.179 ns (0 allocations: 0 bytes)

julia> @btime f($i); @btime g($i);
  14.540 ns (0 allocations: 0 bytes)
  97.243 ns (0 allocations: 0 bytes)

After this change:

julia> @btime f($b); @btime g($b);
  14.557 ns (0 allocations: 0 bytes)
  14.647 ns (0 allocations: 0 bytes)

[stevengj]

Weirdly, for UInt8, !== is much faster than !=, and it is also faster than !== for Bool:

julia> h(a) = reduce(!==, a);

julia> @btime h($b); @btime h($i);
  89.135 ns (0 allocations: 0 bytes)
  8.933 ns (0 allocations: 0 bytes)

even though @code_native (!==)(true,false) and @code_native (!==)(0x01,0x02) have identical output. What gives?

[KristofferC]

Probably specialization heuristics. That's usually the case when the `code_´ doesn't tell the truth about what code end up being run.

[vtjnash]

Ah, I found some potential codegen optimizations to be able to this automatically–and they made things worse :P

I think we used to have this optimization, but had to remove it because we don't zero-initialize arrays (#30350). This PR possibly just finds a code path that we don't sanitize and codegen correctly.