On Wed, 10 Sep 2025 22:05:26 GMT, David Beaumont <d...@openjdk.org> wrote:
>> Summary: Add two new methods to ImageReader to make SystemModuleReader more
>> performant.
>>
>> Analysis of benchmarks shows that when the vast majority of resources
>> (~11,000 in the Perfstartup-SwingSet-G1 benchmark) tested for in
>> SystemModuleReader do NOT exist, performance is degraded compared to this
>> code prior to the refactoring in JDK-8360037.
>>
>> The current refactoring of ImageReader has everything going through a single
>> "findNode()" method for simplest possible encapsulation, but while this is
>> functionally correct, it's not tuned for testing for the non-existence of
>> resources.
>>
>> In particular:
>> 1. SystemModuleReader only requests resources (i.e. things in the jimage
>> file with paths *not* starting /modules/ or /packages/). This means
>> findNode() does two look-ups for the resource, the first of which will
>> always fail.
>> 2. The containsResource() logic doesn't need to create and cache nodes in
>> ImageReader, it can just check for the presence of an ImageLocation
>> corresponding to a resource.
>>
>> Thus two new methods are added to resolve these cases:
>> * findResourceNode(module, path)
>> * containsResource(module, path)
>>
>> Their API takes module and path separately so as to not be confusable with
>> findNode(nodename).
>>
>> However care must be taken to prevent these methods being fooled into
>> returning non-resource entries (this was possible before the refactoring by
>> using module names like "modules" or "packages") so new tests have been
>> added.
>
> David Beaumont has updated the pull request incrementally with one additional
> commit since the last revision:
>
> Found additional place where new API can be used.
src/java.base/share/classes/jdk/internal/jimage/ImageReader.java line 379:
> 377: // exist, so it skips checking the nodes cache and only
> checks for
> 378: // an ImageLocation.
> 379: if (moduleName.contains("/") ||
> resourcePath.startsWith("/")) {
The fast-path guard here and above might improve if they were done before
synchronizing, e.g. dropping the modifier and putting a `synchonized (this)`
block around the rest of the method.
Putting the likely cheaper `resourcePath.startsWith` first might be a small
win, too, depending on input/benchmark.
-------------
PR Review Comment: https://git.openjdk.org/jdk/pull/27203#discussion_r2338193115
