[wcampbell0x2a]

See commits. as they will be easier to review as commits since I moved some code around.

422b79c bench: Use gungraun to easily compare instr counts

• Add criterion for measuring on CI (non-quiet infra). This will hopefully
  lead to good benchmarks to use in merge requests, while keeping
  the regular criterion for wall-time, which is still very usefull since
  callgrind is just estimating the cputime per instruction.
• Copy the existing tests into shared shapes usable by both.

187a48b Keep whole-byte field runs on the byte path

• Fix regression fonud in be_six_bytes_aligned_x128.
• Apply batching logic only if the fields are not whole-width reading.
  For example, on a plain u8 field read with no bit-reads. This uses
  the u64 and shift/mask logic. This was greatly needed to cover
  the regression of be_six_bytes_aligned_x128. This also should work
  over the problem found in perf: read and write a run of adjacent bit fields in one call #677 regarding protocol pipeline issues.
• I benchmarked this over a multi-protocol benchmark (in which we should
  add) and it shows gains (also shows gains in the micro-benchmark
  of be_six_bytes_aligned_x128.

537552b Move plan_bits_run and friends into common mod

6a73d81 test: Add protocol benchmarking to evade branch prediction

• Add testing to confirm the comment above read_bits_uint_msb0 where
  having code that the branch predictor can't help with quickly makes
  the code pattern that we currently have required. The code didn't really
  mention the branch predictor, but I bet that's what they were hitting.

f2f29c2 ci: run benchmark suite in Docker
Keep only the callgrind benchmark, which gives stable instruction
counts. The criterion wall-clock job gave noisy results on CI runners.

[wcampbell0x2a]

@Simon-Calbert-Aerospacelab in #677 you mentioned that pipeline'ing benchmarks are driving the decision you made in src/reader.rs, so I added some benchmarks to prove it. See b2e5de0, as I make some more assumptions on getting some performance back for some common paths.

[wcampbell0x2a]

The benchmark CI I added/changed is failing. But hopefully it's just benches against master which are expected to fail.

[wcampbell0x2a]

I need to update for #684