Slow generation: HandlebarTemplateEngine rebuilds Handlebars and reparses templates on every render

Summary

HandlebarTemplateEngine.getRendered(...) creates a new Handlebars instance and recompiles the template (and every partial it references) on every call. Profiling shows the jknack ANTLR-based lexer accounting for ~65% of CPU during code generation for a typical OpenAPI specification.

How to reproduce

Generate code with the handlebars engine (the default for v3) against any sufficiently large spec. The cost grows roughly with (templates_rendered) × (partials_per_template) because every render repeats the full Handlebars setup + parse.

Profile

Java Flight Recorder captured during a representative warm run shows:

• ~65% of in-method CPU under org.antlr.v4.runtime.atn.LexerATNSimulator.closure(...) — jknack's parser.
• Each getRendered allocates a new Handlebars, re-registers all helpers, and reparses the entry template, which in turn re-parses every {{> partial }} it transitively includes.

Why it matters

Generating a server-side stub project with ~75 controllers / ~180 output files takes ~50–60 s on the affected code path. Once compiled templates are cached for the lifetime of a single generation, the same workload completes in ~12 s — a ~4.5× speed-up with no API change and no spec-side change.

Precedent

#12313 / its accompanying fix established the same pattern in another hotspot (AbstractJavaCodegen.toModelName() was called millions of times and was cached for a 3× improvement). The fix proposed here applies the same idea to template parsing.

Caveat / why a naive cache breaks output

jknack's stock ConcurrentMapTemplateCache cannot be used as-is for handlebars partials in prettyPrint(true) mode. Partial.merge wraps the partial's TemplateSource in an anonymous class whose equals/hashCode delegate only to the underlying source (filename + lastModified), but whose content() returns the partial body re-indented by the include site's leading whitespace. The same partial included at two different indents therefore collides on lookup and the first-compiled indent wins for every subsequent include — observable as silently shifted whitespace in generated output. See related jknack issues #401 and #708.

The proposed fix carries a small replacement cache (~95 LOC) keyed on (filename, content), which preserves the speed-up while keeping per-indent renders correct.

Proposed fix

A PR will follow that:

1. Holds a single Handlebars instance per HandlebarTemplateEngine (and so per generation), registering helpers once.
2. Caches compiled top-level Templates per templateFile.
3. Installs an IndentAwareTemplateCache to make handlebars' own partial cache safe under prettyPrint(true).

No public API changes. No template changes. No swagger-codegen-generators change required.

Start at HandlebarTemplateEngine.getRendered(...) and inspect how the Handlebars instance, helpers, templates, and partials are created for each render. Review the proposed IndentAwareTemplateCache behavior and verify both compiled-template reuse and correct prettyPrint(true) output at different include indents. Done means preserving output with no public API or template changes while achieving the stated generation speed-up.