Ensure SubArray's unsafe_getindex doesn't check bounds

[mbauman]

This resolves the core issue behind the regressions in SubArray's performance as reported in #5117. Unfortunately, without #7799, the surface syntax isn't very nice since it's a ton of unsafe_ function calls.

Using the test case from #5117 (comment) :

• Julia pre-10525 (benchmark as written, with @inbounds)

follow up: mykern( 2:(siz - 3), s, a, x, y )
   9.484 milliseconds (157 allocations: 10893 bytes)
follow up: mykern( 2:(siz - 3), s1, a, x1, y1 )
  68.632 milliseconds (16353 allocations: 718 KB)
follow up: mykern( 3:(siz - 2), s2, a, x2, y2 )
  16.753 milliseconds (6 allocations: 192 bytes)
  15.493 milliseconds (6 allocations: 192 bytes)
  17.036 milliseconds (6 allocations: 192 bytes)

• This branch (benchmark rewritten to use unsafe_getindex/unsafe_setindex!):

follow up: mykern( 2:(siz - 3), s, a, x, y )
  10.046 milliseconds (158 allocations: 11469 bytes)
follow up: mykern( 2:(siz - 3), s1, a, x1, y1 )
  70.425 milliseconds (19097 allocations: 827 KB)
follow up: mykern( 3:(siz - 2), s2, a, x2, y2 )
  14.263 milliseconds (6 allocations: 192 bytes)
  16.470 milliseconds (6 allocations: 192 bytes)
  16.711 milliseconds (6 allocations: 192 bytes)

[mbauman]

JULIA_TESTFULL for subarray passes without a hitch, as well.

[timholy]

Nice find! Clear evidence that skipping bounds checks basically needs to propagate downwards.

Just to check, are you saying that to get this performance you need to modify the mykern script to explicitly use unsafe_getindex(s, i) rather than s[i]?

[mbauman]

Just to check, are you saying that to get this performance you need to modify the mykern script to explicitly use unsafe_getindex(s, i) rather than s[i]?

Yes, exactly.

[timholy]

Very good. Clearly this is a nice step forward.

Can you also ("temporarily") turn off the explicit bounds checking for getindex? Let the underlying array still check bounds, of course. I think changing all s[i] to unsafe_getindex(s, i) is just too ugly for user code.

[mbauman]

What, you don't think it's reasonable to ask people to rewrite:

@inbounds s[i] = a * (x[i+1] - x[i]) / (y[i+1] - y[i])

to

unsafe_setindex!(s, a * (unsafe_getindex(x, i+1) - unsafe_getindex(x, i)) / (unsafe_getindex(y, i+1) - unsafe_getindex(y, i)), i)

for version 0.4? And then back again later? The latter is so much clearer and more explicit! ;-)

I'll merge this now and tackle the change in semantics separately, with the hope that it'll be reverted someday.