Problem
Regex predicates ((identifier =~ /foo.*/)) deserialize their DFA from module bytes on every evaluation, inside the VM's match loop:
crates/plotnik-vm/src/engine/vm.rs, evaluate_predicate → plotnik_bytecode::deserialize_dfa(regex_bytes) per call.
A regex predicate inside a quantified pattern over a large file deserializes the same DFA thousands of times.
Approach
Deserialize once and reuse:
- Simplest: a lazy per-regex-index cache on the VM (
Vec<OnceCell<DFA>> sized to the regex table), built on first use per execution.
- Better: deserialize zero-copy at
Module load — the crate already ensures 64-byte alignment for embedded bytecode (include_query_aligned), and regex_automata DFAs support zero-copy deserialization from aligned bytes. This may be a lifetime exercise rather than a copy.
Acceptance
- A trace/benchmark shows one deserialization per distinct regex per execution (or zero at exec time with the load-time approach).
- No behavior change — existing predicate tests pass.
Problem
Regex predicates (
(identifier =~ /foo.*/)) deserialize their DFA from module bytes on every evaluation, inside the VM's match loop:crates/plotnik-vm/src/engine/vm.rs,evaluate_predicate→plotnik_bytecode::deserialize_dfa(regex_bytes)per call.A regex predicate inside a quantified pattern over a large file deserializes the same DFA thousands of times.
Approach
Deserialize once and reuse:
Vec<OnceCell<DFA>>sized to the regex table), built on first use per execution.Moduleload — the crate already ensures 64-byte alignment for embedded bytecode (include_query_aligned), andregex_automataDFAs support zero-copy deserialization from aligned bytes. This may be a lifetime exercise rather than a copy.Acceptance