Faster setindex! for RHSs with inefficient linear indexing

[timholy]

This was a long-standing TODO item, for which we now have the needed infrastructure.

Provides a 7-8x speedup on the last line of this demo:

function testset!(A, X, indexes, n)
    for i = 1:n
        setindex!(A, X, indexes...)
    end
    A
end

A = rand(1000,1001)
B = rand(800,400)
S = sub(B, 2:799, 2:399)

testset!(A, B, (101:900,101:500), 1)
testset!(A, S, (101:898,101:498), 1)

@time 1
@time testset!(A, B, (101:900,101:500), 100)
@time testset!(A, S, (101:898,101:498), 100)

[timholy]

I presume the AppVeyor failure is unrelated?

[tkelman]

Yes, https://ci.appveyor.com/project/StefanKarpinski/julia/build/1.0.2329/job/9fx92an62in0sonp does look like an instance of #10045. Though the backtrace here (and in a few other cases, but not all) is interesting

    From worker 3:       * linalg2              in  65.09 seconds

Please submit a bug report with steps to reproduce this fault, and any error messages that follow (in their entirety). Thanks.
Exception: EXCEPTION_ACCESS_VIOLATION at 0x6e07661a -- DCABS1 at C:\projects\julia\usr\bin\libopenblas.DLL (unknown line)
DCABS1 at C:\projects\julia\usr\bin\libopenblas.DLL (unknown line)
zgemv at C:\projects\julia\usr\bin\libopenblas.DLL (unknown line)
zgemv at C:\projects\julia\usr\bin\libopenblas.DLL (unknown line)
Worker 3 terminated.    From worker 3:       * linalg3              

dcabs1 makes me think the win32 problem could possibly be the same as the 32-bit Linux problem in #10027

[timholy]

Thanks for checking, @tkelman.

[Jutho]

I think this commit had a negative effect on the speed of setindex! as experienced in the /test/perf/cat tests. The time increase is not in cat but also in the manual performance tests, so it is in setindex!(::Array,::Array,::UnitRange{Int},::UnitRange{Int}) call.

[Jutho]

What about writing this as:

Xs = start(X)
@nloops $N i d->(1:length(I_d)) d->(@inbounds offset_{d-1} = offset_d + (unsafe_getindex(I_d, i_d)-1)*stride_d) begin
v, Xs = next(X, Xs)
@inbounds A[offset_0] = v
end

Then fast linear indexing will be used for X if available, and cartesian indexing otherwise.

[timholy]

I'd had the impression that eachindex used a linear index when appropriate, but that is clearly wrong. (And probably reasonable, too.) So your strategy is fine. Want to do it yourself, or shall I?