#222 introduced a major performance regression

[csobrinho]

Hi Shigeru, #222 and 1719817 added a major regression to the CtClass loading. I profiled the code and now 75-80% of all the cpu time is wasted on "urlConnection.getInputStream()".

We saw a regression of at least 20x when the jar is big enough (120k classes, 400 MB jar).

• with 3.23.1-GA, processing of all classes takes ~200s
• with 3.25.0-GA, processing of all classes takes ~4200s

[TimvdLippe]

Would it be possible to revert the commit for now? Seems like the performance regression is a lot more severe than the memory leak.

[chibash]

I've reverted the commit.

[csobrinho]

Awesome @chibash. Any chance that commit could be cherrypicked into 3.25.0-GA as 3.25.1-GA or a new release be pushed? Thanks

[chibash]

It should be soon. by the end of September?

[csobrinho]

Sounds good, thanks!

[csobrinho]

Some more details. I tested 3.26.0-GA and it's faster (Thank you!) but still 10x slower than 3.22.0-GA.

When the JarClassPath.jarEntries is really big, the ArrayList.contains call inside the find will be O(n) and this will be very slow if done several thousand times.

In the profile picture we can see that almost all the time taken by ClassPool.openClassfile is consumed by List.contains.

If we only need to check if the class is there or not, then an HashSet will be much better and faster with O(1) checks. Also insertion order and index of the element are not used and necessary.

This made our compilation jump from 3m back to the original 20s in 3.22.0-GA.

Please be so kind and merge this as 3.26.1-GA.
Thank you!