ERROR: UndefVarError: ##op#5321 not defined

[zsoerenm]

If I increment an index of one of the variables, I will get the error ##op#5321 not defined
MWE:

function complex_mul_with_index_offset(a_re, a_im, b_re, b_im, c_re, c_im)
    @avx for i = 1:length(a_re) - 1
        c_re[i] = b_re[i] * a_re[i + 1] - b_im[i] * a_im[i + 1]
        c_im[i] = b_re[i] * a_im[i + 1] + b_im[i] * a_re[i + 1]
    end
end
a_re = ones(Int, 64); a_im = copy(a_re); b_re = copy(a_re); b_im = copy(a_re); c_im = copy(a_re); c_re = copy(a_re)

julia> complex_mul_with_index_offset(a_re, a_im, b_re, b_im, c_re, c_im)
ERROR: UndefVarError: ##op#5321 not defined
Stacktrace:
 [1] sum_with_index_offset(::Array{Int64,1}, ::Array{Int64,1}, ::Array{Int64,1}, ::Array{Int64,1}, ::Array{Int64,1}, ::Array{Int64,1}) at /home/zorn/.julia/packages/LoopVectorization/2jh99/src/reconstruct_loopset.jl:260

This does work, however:

function complex_mul_with_index_offset1(a_re, a_im, b_re, b_im, c_re, c_im)
    @avx for i = 1:length(a_re) - 1
        c_re[i] = b_re[i] * a_re[i + 1] - b_im[i] * a_im[i + 1]
    end
end

[chriselrod]

Interesting problem, thanks.

@_avx instead of @avx also works. @avx calls a generated function (which can use extra type information to make better decisions), while @_avx just transforms the loops (making a few default assumptions about what types there probably are).

There seems to be a mistranslation of the loop as it gets passed to the generated function, but also a problem in the elimination of redundant loads, currently getting confused by the "i + 1" indexing (which is why a_re[i + 1] and a_im[i+1] only once, as in your second example, works).

I'm still looking into what exactly is going wrong, but I hope to clean this up and fix it fairly soon.

[chriselrod]

I fixed this and added your example to the testsuite.