You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Dispatch generated template registries through a class index instead of linear == and instanceof chains. With 422 page models the last model cost about 1.1 µs per request;
every model now resolves in a few nanoseconds. Subclasses of open models still fall back
to the previous instanceof chain. The new registry also compiles faster: ReAI's compileJava for 344 renderers fell from 10.8 s to 6.3 s.
Write the message usage manifest whenever a catalog exists, so generateMessages=false
keeps build-wide dead-key detection. A module can now render templates from a catalog
whose typed factories another module generates instead of compiling a second factory class.
The module-local unused check that previously ran at compile time is replaced by the
build-wide thimMessageUsageCheck.
Record the measurements behind these changes, and the buffer and text-encoding ideas
that were rejected, in the performance audit.