What

A set of internal, behavior-preserving optimizations that reduce web application
deployment/startup time. There are no changes to defaults, configuration,
public API, parsing output, or threading model — outputs are identical; only
redundant work and allocations are removed.

Changes

Annotation / JAR scanning (ContextConfig)

• Index the @HandlesTypes types by name (class-file descriptor form for
  annotations, binary name for classes) so matching a scanned class is an O(1)
  lookup instead of a linear scan of all handled types plus a per-annotation
  string conversion.
• Record whether a fragment JAR contains META-INF/resources/ during the
  annotation scan (which already iterates every entry), so processResourceJARs
  no longer re-opens and re-scans each JAR.
• Check the cached default web.xml freshness with File.lastModified() for
  file: URIs instead of opening a URLConnection.
• Use Collections.singleton for the per-fragment annotation merge.

Mapper (Mapper / batch registration)

• Register a context's servlet mappings into the copy-on-write wrapper arrays in
  a single merge per mapping category instead of reallocating the array once per
  mapping (previously O(n²)). Verified equivalent to the old sequential
  insertMap behaviour (sorted order, existing-wins dedup, wildcard nesting)
  over 300k randomized cases.

Web resources

• JarWarResourceSet: build the META-INF bloom filter once when the entries
  are read rather than rebuilding it on every resource lookup.
• AbstractSingleArchiveResourceSet: pre-size the entry map from the JAR entry
  count.
• StandardRoot: hoist a loop invariant in listResources; Collections.addAll
  instead of Arrays.asList wrapper.
• AbstractFileResourceSet / FileResource: file() already resolves the
  canonical path to validate that a file is located under the resource set base,
  then discards it, leaving FileResource.getCanonicalPath() to resolve the same
  file a second time. The body of file() moves to a new validatedFile() that
  returns the file together with that path, and the two resource sets that create
  a FileResource pass it on. file() is kept for its other callers.

Misc internal

• StandardJarScanner: hoist loop-invariant getJarFileURL()/toURI() out of the
  manifest Class-Path loop; Collections.addAll for the class-path URLs.
• Digester.updateAttributes: resolve the class loader once per call.
• FragmentJarScannerCallback: single map lookup for duplicate detection.
• LifecycleBase.fireLifecycleEvent: skip allocating the event when there are no
  listeners.
• AbstractProtocol.init: compute the (built/quoted) name once.
• MbeansDescriptorsIntrospectionSource: Set membership test instead of a
  linear array scan.

Correctness

All changes preserve observable behaviour. The Mapper change was fuzz-checked
against the original sequential implementation (identical arrays and nesting).
TestContextConfigAnnotation is updated to build the new derived index (mirroring
the normal processClasses flow).

The canonical path reuse does not weaken the resource set validation: the checks
in file() are unchanged, and the path is only handed to the FileResource that
those checks just validated. It is null when getAllowLinking() short-circuits
the validation, and nothing is cached across resources or over time, so this is a
per-instance value rather than a canonical path cache. The existing 7-arg
FileResource constructor is retained as a delegate.

Please run at least TestMapper*, TestContextConfig*, the webresources tests,
and TestDefaultServlet*.

Performance

Measured on a build from source (JDK 25), median of 8 start/stop cycles, same
webapp set both sides. The gains scale with application size:

• Mapping-heavy app (~8000 classes, 20000 servlet mappings): server startup
  ~560 ms → ~483 ms (~14%). Measured without the newest CanonicalPath cache, which even improves it more.
• Small/typical apps: low single-digit %, near run-to-run noise.

The canonical path reuse removes one syscall per file resource lookup. On a
resource-heavy deploy (JFR, Windows, ~500 classes under WEB-INF/classes) it
eliminated FileResource.getCanonicalPath() from the profile entirely, 21 of 112
canonicalize0 samples. The effect is smaller on Linux, where path resolution is
considerably cheaper than on Windows.

Changelog entries are included.

Co-Authored-By: Claude Opus 4.8 noreply@anthropic.com