mir-borrowck-profiling-notes.md

2018.06.28

HEAD: 9e67309451

Version	Times	Comment
Plain	3.66s
NLL	6.06s
Ratio	1.65	(NLL / plain)
%	0.39%	(NLL - plain) / NLL

• {do_mir_borrowck} is 37%
  • compute_regions is 17%
    • {type_check_internal} is 7%
    • {generate_constraints} is 3%
    • {liveness_of_locals} is 2%
    • {RegionInferenceContext::solve} is 2%
    • {RegionInferenceContext::new} is 1%
  • visit_statement_entry is 7%
  • do_dataflow is 5%
  • visit_terminator_entry is 2%

Looking at hotspots from the perf focus graph, I found some possible routes for further exploration:

• Rewrite places_conflict to not use zipped iterators, or to use a custom iterator that walks back places lazilly
  • the perf annotate results seem to suggest that a lot of time is being spent in the iterator and smallvec machinery; rewriting may or may not help though
• check for duplicate constraints (We are
• liveness is non-trivial percentage (2%)
  • it is currently iterating over the whole graph every time
  • keep a "changed" list?
  • using datafrog would be better, but that might require more careful thought about how to integrate the results
  • still, perhaps worth pursuing
• liveness simulate_block
  • allocates a temporary per basic block; it could be shared across all basic blocks
  • this func is 1% on the profile but it's basically all here
• lazier causal tracking (rust-lang/rust#51710)
  • seems to account for most of generate_constraints
• optimizing canonicalization (rust-lang/rust#48417)
  • 6% of total time
  • but 2% of MIR borrowck time
  • still, a clear win and one that will pay dividends over time