Allow sorting of tuples of numbers

[mcabbott]

This allows sort to work on arrays of tuples, motivated by partialsortperm(x; dims)-like things in FluxML/NNlib.jl#353

It also adds the simplest implementation of sortperm using that. Not sure if this is too crude for the package to want to keep? Needs tests, if it is wanted.

[maleadt]

Using zero for an unused value seems wrong, as noticed here it doesn't work for all types, and it's also unnecessary since it just represents an unused value (but we need to do this in order to avoid deadlocks).

What about:

julia> f(i, a) = i ? @inbounds(a[]) : Ref{Int}()[]
f (generic function with 2 methods)

julia> f(false, [1])
140526431453856

julia> f(true, [1])
1

julia> @code_llvm f(true, [1])
;  @ REPL[11]:1 within `f`
define i64 @julia_f_235(i8 zeroext %0, {}* nonnull align 16 dereferenceable(40) %1) #0 {
top:
  %2 = and i8 %0, 1
  %.not = icmp eq i8 %2, 0
  br i1 %.not, label %L28, label %L23

L23:                                              ; preds = %top
; ┌ @ abstractarray.jl:1218 within `getindex`
; │┌ @ abstractarray.jl:1245 within `_getindex`
; ││┌ @ array.jl:839 within `getindex`
     %3 = bitcast {}* %1 to i64**
     %4 = load i64*, i64** %3, align 8
     %5 = load i64, i64* %4, align 8
; └└└
  ret i64 %5

L28:                                              ; preds = %top
  ret i64 undef
}

Bit ugly, but does result in the expected undef.

[mcabbott]

Ok, I did not look closely enough to figure out what the zero was or wasn't being used for. If unused, then a generic undef does sound better.

[maleadt]

Rebased. Removed the sortperm since that has been introduced by another PR. Also replaced the one/zero calls to the above Ref trick since those values are unused anyway.

Outdated.

[maleadt]

Turns out we can just do the naive thing (on 1.7, at least): define a variable conditionally and use it under the same condition so that LLVM can optimize the throw(UndefVarError) away. Let's see how CI things about this.

[mcabbott]

Thanks for picking this up!

[codecov]

Codecov Report

Merging #1196 (5a9611f) into master (bf26270) will increase coverage by 0.00%.
The diff coverage is 0.00%.

@@           Coverage Diff           @@
##           master    #1196   +/-   ##
=======================================
  Coverage   78.94%   78.95%           
=======================================
  Files         119      119           
  Lines        8650     8649    -1     
=======================================
  Hits         6829     6829           
+ Misses       1821     1820    -1     

Impacted Files	Coverage Δ
src/sorting.jl	24.14% <0.00%> (+0.05%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 94ba745...5a9611f. Read the comment docs.