perf: improve general performance

[P403n1x87]

We try to use sets for quick lookups and cache the result of some repetitive operations to speed up some operations. We also remove a good deal of asserts embedded within the code that adde extra overhead. Where appropriate, these should be replaced with actual tests.

The current change squeezes about 3/4% performance from a decompile/recompile loop.

Benchmarking

The following round-trip script has been used to benchmark the performance of the proposed changes

from time import monotonic

from bytecode import Bytecode

n = []
for _ in range(30):
    start = monotonic()
    for _ in range(2000):
        Bytecode.from_code(Bytecode.from_code.__code__).to_code()
    n.append(monotonic() - start)


avg = sum(n) / len(n)
sdev = (sum(_**2 for _ in n) / len(n) - avg**2) ** 0.5
print(f"{avg:0.2f} +/- {sdev:0.2f}")

Baseline: 1.10 +/- 0.02
This PR: 1.03 +/- 0.02

[codecov-commenter]

Codecov Report

❌ Patch coverage is 82.89474% with 26 lines in your changes missing coverage. Please review.
✅ Project coverage is 91.49%. Comparing base (7f3ed3e) to head (d5a0515).

Files with missing lines	Patch %	Lines
src/bytecode/concrete.py	83.45%	16 Missing and 6 partials ⚠️
src/bytecode/instr.py	83.33%	3 Missing ⚠️
src/bytecode/cfg.py	0.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #172      +/-   ##
==========================================
- Coverage   95.43%   91.49%   -3.94%     
==========================================
  Files           7        7              
  Lines        2147     2164      +17     
  Branches      481      489       +8     
==========================================
- Hits         2049     1980      -69     
- Misses         56      122      +66     
- Partials       42       62      +20     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[MatthieuDartiailh]

Thanks for starting this, just one small question. Also you may need some convincing for Mypy to be happy with you without all the assertions.

[MatthieuDartiailh]

Out of curiosity do you have a micro-benchmark for this. It is not obvious to me that for a single element a set is faster ?

[P403n1x87]

The core of the performance gains come from caching and avoiding some of the more expensive asserts. These small sets are more for type coherence than anything else. I have the feeling that all the branching that is done at runtime is also a main "offender" performance-wise. It would be great if sys.version_info checks could be done just once at compile time, possibly at the cost of copy-pasting the same code. I'll see if there are any other low-hanging fruits that we can pick

[P403n1x87]

In fact I've got a pretty decent improvement when turning HASJREL etc... into sets. I'll post some benchmarking notes in the description.

[MatthieuDartiailh]

I agree that I would love to have the equivalent of compile time specialization.