Take some steps toward eliminating the last unsafe in ring::polyfill

[briansmith]

Replace a few uses of ChunksFixed with a new ArraySplitMap. ArraySplitMap avoids the need to construct an intermediate array; we would hope the optimizer would handle that, but at least in the past it did not always do so.

The approach here isn't very scalable since the source code of each implementation of ArraySplitMap grows twice as fast as the size of the input array. Maybe we can find a cleverer solution, at least for the larger ones.

[codecov]

Codecov Report

Merging #1732 (6e6b118) into main (75c620a) will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main    #1732      +/-   ##
==========================================
+ Coverage   96.11%   96.12%   +0.01%     
==========================================
  Files         136      137       +1     
  Lines       20521    20548      +27     
  Branches      217      217              
==========================================
+ Hits        19723    19752      +29     
  Misses        760      760              
+ Partials       38       36       -2     

Files	Coverage Δ
src/aead/aes.rs	98.66% <100.00%> (ø)
src/aead/chacha.rs	100.00% <100.00%> (ø)
src/aead/gcm.rs	98.00% <100.00%> (-0.01%)	⬇️
src/aead/gcm/gcm_nohw.rs	76.58% <100.00%> (-0.34%)	⬇️
src/polyfill.rs	100.00% <ø> (ø)
src/polyfill/array_split_map.rs	100.00% <100.00%> (ø)
src/polyfill/chunks_fixed.rs	100.00% <ø> (ø)

... and 2 files with indirect coverage changes

📣 We’re building smart automated test selection to slash your CI/CD build times. Learn more

[briansmith]

@joshlf PTAL. This is mostly to start a discussion. Definitely I am curious about whether there's a simpler way of doing this. I was hoping the array_* standard library facilities would be available by now, but it seems unlikely that they will become stable soon.

[joshlf]

@joshlf PTAL. This is mostly to start a discussion. Definitely I am curious about whether there's a simpler way of doing this. I was hoping the array_* standard library facilities would be available by now, but it seems unlikely that they will become stable soon.

I don't have any strong opinions here, but the approach seems reasonable. Anything that removes unsafes is good in my book 🙂

[briansmith]

For the small arrays that are converted here (the ones for which lines are removed above) this approach seems to optimize well. For the other two larger array sizes (mentioned below), I will post a separate PR that uses different techniques.

[briansmith]

This change ends up being +91 −33 (+58 net) LoC. The future changes to eliminate ChunksFixed comletely are +13 -48 (-35 net) LoC. So it looks like the project to eliminate ChunksFixed, for the purpose of eliminating the one unsafe line, costs about +23 net LoC.

I believe all the changes here don't worsen the code, and slightly improve the callers. The remaining changes to remove ChunksFixed are more debatable. So I'm going to merge this for now and we'll discuss the issues in that PR.