Rework our hash functions a bit for significant better performance:
* Rename int_hash to u32_hash to mirror u64_hash.
* Make pair_int_hash call u64_hash instead of multiple u32_hash()es.
* Implement MurmurHash3's fmix32 and fmix64 for u32_hash and u64_hash.
On my machine, this speeds up u32_hash by 20%, u64_hash by ~290%, and
pair_int_hash by ~260%.
We lose the property that an input of 0 results in something that is not
0. I've experimented with an offset to both hash functions, but it
resulted in a measurable performance degradation for u64_hash. If
there's a good use case for 0 not to result in 0, we can always add in
that offset as a countermeasure in the future.
Replace the ScopePusher RAII class (which performed scope analysis
in its destructor chain during parsing) with a two-phase approach:
1. ScopeCollector builds a tree of ScopeRecord nodes during parsing
via RAII ScopeHandle objects. It records declarations, identifier
references, and flags, but does not resolve anything.
2. After parsing completes, ScopeCollector::analyze() walks the tree
bottom-up and performs all resolution: propagate eval/with
poisoning, resolve identifiers to locals/globals/arguments, hoist
functions (Annex B.3.3), and build FunctionScopeData.
Key design decisions:
- ScopeRecord::ast_node is a RefPtr<ScopeNode> to prevent
use-after-free when synthesize_binding_pattern re-parses an
expression as a binding pattern (the original parse's scope records
survive with stale AST node pointers).
- Parser::scope_collector() returns the override collector if set
(for synthesize_binding_pattern's nested parser), ensuring all
scope operations route to the outer parser's scope tree.
- FunctionNode::local_variables_names() delegates to its body's
ScopeNode rather than copying at parse time, since analysis runs
after parsing.
