ClassLoaderFilesResourcePatternResolver.isDeleted() calls
resource.exists() and resource.getURI() inside the loop over the
uploaded files, so both are repeated once per DELETED entry although
neither depends on the entry. Every resource lookup made while the
application context is being built goes through this method - about
1300 times per restart in the application I measured - so a session
that has accumulated 100 deleted files performs around 130 000 file
system lookups where 1300 are enough.
This commit hoists both calls out of the loop: the resource is
inspected at most once per call, on the first DELETED entry, and the
URI comparison then uses the cached value. The method still returns on
the first match and still reports a failing getURI() as an
IllegalStateException, so behaviour is unchanged.
The loop now walks the per-directory entry sets, which is what
ClassLoaderFiles.addAll() and RestartServer already do, instead of the
flattened view added in gh-46289. That view is wrapped in
Collections.unmodifiableSet(), whose iterator is shared JDK code; in a
running application its delegate calls are megamorphic and C2 stops
inlining them, which made restarts with many accumulated files and no
deletions slower than before gh-46289. isDeleted() returns a boolean,
so the visiting order of the entries cannot change its result.
Measured on a Spring Boot application driven through the remote restart
path (restart request to ApplicationReadyEvent), 12 JVMs per variant
and 20 restarts per JVM, medians:
100 source directories, 10 000 entries, 10 deleted: 278 ms -> 200 ms
100 source directories, 10 000 entries, 100 deleted: 891 ms -> 201 ms
100 source directories, 10 000 entries, 500 deleted: 3536 ms -> 200 ms
100 source directories, 50 000 entries, none deleted: 431 ms -> 341 ms
100 source directories, 10 000 entries, none deleted: 202 ms -> 192 ms
no uploaded files: 150 ms -> 148 ms
See gh-51841
Signed-off-by: DongHoon Lee <dhl1924@naver.com>