Generate reverse for StaticVector types

[OTDE]

Hat tip to @MasonProtter for this neat little trick.

This PR eliminates allocations in SVector containers that hold a single bits eltype and speeds them up by 1000 times the speed of the base branch in some cases:

Benchmarking v = SVector(Tuple(randn(200)))

base branch:

julia> @benchmark reverse($v)
BenchmarkTools.Trial: 
  memory estimate:  486.89 KiB
  allocs estimate:  20699
  --------------
  minimum time:     536.714 μs (0.00% GC)
  median time:      540.254 μs (0.00% GC)
  mean time:        567.529 μs (1.55% GC)
  maximum time:     1.777 ms (65.64% GC)
  --------------
  samples:          8788
  evals/sample:     1

sc/reverse-speedup:

julia> @benchmark reverse($v)
BenchmarkTools.Trial: 
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     31.943 ns (0.00% GC)
  median time:      32.063 ns (0.00% GC)
  mean time:        33.041 ns (0.00% GC)
  maximum time:     98.280 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     994

Benchmarking v = SVector(Char.(rand('a':'z', 200))...):

base branch:

julia> @benchmark reverse($v)
BenchmarkTools.Trial: 
  memory estimate:  87.34 KiB
  allocs estimate:  400
  --------------
  minimum time:     462.796 μs (0.00% GC)
  median time:      465.363 μs (0.00% GC)
  mean time:        479.433 μs (0.49% GC)
  maximum time:     2.088 ms (76.58% GC)
  --------------
  samples:          10000
  evals/sample:     1

sc/reverse-speedup:

julia> @benchmark reverse($v)
BenchmarkTools.Trial: 
  memory estimate:  0 bytes
  allocs estimate:  0
  --------------
  minimum time:     10.548 ns (0.00% GC)
  median time:      11.030 ns (0.00% GC)
  mean time:        11.695 ns (0.00% GC)
  maximum time:     54.831 ns (0.00% GC)
  --------------
  samples:          10000
  evals/sample:     999

For non-bits types, the speedup and allocation saves are closer to 100 times those of the base branch:

Benchmarking v = SVector(("a", :b, randn(198)...)) (which has an eltype of Any):

base branch:

julia> @benchmark reverse($v)
BenchmarkTools.Trial: 
  memory estimate:  485.23 KiB
  allocs estimate:  20694
  --------------
  minimum time:     537.686 μs (0.00% GC)
  median time:      552.485 μs (0.00% GC)
  mean time:        591.481 μs (1.78% GC)
  maximum time:     2.932 ms (54.48% GC)
  --------------
  samples:          8438
  evals/sample:     1

sc/reverse-speedup:

julia> @benchmark reverse($v)
BenchmarkTools.Trial: 
  memory estimate:  4.70 KiB
  allocs estimate:  200
  --------------
  minimum time:     5.252 μs (0.00% GC)
  median time:      5.419 μs (0.00% GC)
  mean time:        5.933 μs (1.44% GC)
  maximum time:     167.364 μs (94.68% GC)
  --------------
  samples:          10000
  evals/sample:     6

Benchmark information:

julia> versioninfo()
Julia Version 1.4.1
Commit 381693d3df* (2020-04-14 17:20 UTC)
Platform Info:
  OS: Linux (x86_64-pc-linux-gnu)
  CPU: Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-8.0.1 (ORCJIT, skylake)
Environment:
  JULIA_EDITOR = atom  -a
  JULIA_NUM_THREADS = 6

[OTDE]

CI failures appear to be unrelated.

[andyferris]

Nice!

I wish the tuple methods in base were a bit more robust, but oh well.

[c42f]

Looks good to me, cheers!

[MasonProtter]

Any thoughts on if this should be attempted to be upstreamed to Base?

[martinholters]

Ref. JuliaLang/julia#16604 (note JuliaLang/julia#16604 (comment)).

[andyferris]

I think you will want if @generated for a Tuple method if you want this accepted in Base. But yes it might be a good idea?