Cache included ASTs in memory: a BOSL2 render goes 67ms to 7.5ms

[revarbat]

Parsing include <BOSL2/std.scad> is ~55 ms, which was 82% of evaluating a small script against it — and it happened on every render, of every script, even though the library never changed. An editor re-rendering paid it each time; BOSL2's docs build paid it once per example, a thousand times over.

Files are now parsed once and shared (openscad_cpp_parser#7).

	before	after
evaluate() with BOSL2	67.2 ms	7.5 ms (9×)
BOSL2's 909-test suite	95.3 s	39.1 s (2.4×)

What changed here

The evaluator takes the borrowed statement list the cache produces, which needed a resolveUseScopes() overload for nodes it does not own. Evaluator/EvalContext/Compiler now carry the run's ScopeTable, since a node no longer carries its own scope — the parser PR explains why it cannot.

A Closure also captures its defining scope now. It used to read it back off its FunctionLiteral node, which works only while the node holds one, and a closure can be called at generate time with no EvalContext left to look it up through.

The bug worth knowing about

childCtx() and callCtx() build a fresh EvalContext field by field. The version of this that forgot to carry scopeTable made every node's scope read back as null. Name resolution then quietly fell back to the enclosing scope, so a module body calling a builtin of its own name found itself — BOSL2's module _cube(...) cube(...); wrapper recursed to the depth limit.

All 1149 tests passed with that bug. BOSL2 caught it. There is now a test that checks each derived context directly; removing the assignment makes it fail (verified).

Verified

• 1153 C++ tests, 648 parser tests, 31 binding tests
• BOSL2's 909 regression tests
• Cache holds one entry per file and replaces on edit, so repeated saves don't accumulate ASTs — tested

🤖 Generated with Claude Code