Preserve names of arrays and iteration variables across generated boundary

[timholy]

Fixes #74

[codecov-io]

Codecov Report

Merging #75 into master will increase coverage by 0.05%.
The diff coverage is 83.33%.

@@            Coverage Diff             @@
##           master      #75      +/-   ##
==========================================
+ Coverage   93.23%   93.28%   +0.05%     
==========================================
  Files          27       27              
  Lines        2808     2815       +7     
==========================================
+ Hits         2618     2626       +8     
+ Misses        190      189       -1     

Impacted Files	Coverage Δ
src/reconstruct_loopset.jl	92.41% <73.33%> (ø)
src/condense_loopset.jl	91.42% <100.00%> (+0.10%)	⬆️
src/memory_ops_common.jl	89.23% <0.00%> (ø)
src/determinestrategy.jl	93.97% <0.00%> (+0.04%)	⬆️
src/vectorizationbase_extensions.jl	75.00% <0.00%> (+17.85%)	⬆️

---

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 07e71a1...7b43d46. Read the comment docs.

[chriselrod]

This is fantastic!

As a quick comment, I avoided making any of the structs parametric because I wanted to move towards everything being concretely typed as quickly as possible, although I haven't really dived into how well I'm actually achieving that. Regardless, gensym is actually much slower than a dynamic dispatch, so this should speed up compilation if anything...

[timholy]

If we add enough @nospecialize statements everywhere, concrete typing probably shouldn't matter. (You don't have any reason other than compile performance to want that, right?)

[chriselrod]

Yes, compile performance is the only reason. Any general advice on using @nospecialize?

[timholy]

I'll submit a PR. Best approach is to analyze the number of specializations per method; that tells you how many times they've been compiled. Whole goal is to avoid recompilation.

[timholy]

Not as straightforward as I hoped, it may be a Julia bug. JuliaLang/julia#35131