RFR: 8374377: PNGImageDecoder Slow For 8-bit PNGs [v3]

[Daniel Gredler]

On Sat, 3 Jan 2026 04:31:15 GMT, Jeremy Wood <jwood at openjdk.org> wrote:

>> test/jdk/sun/awt/image/png/PngImageDecoder8BitTest.java line 174:
>> 
>>> (failed to retrieve contents of file, check the PR for context)
>> Am I correct in assuming that both models end up using `PNGImageDecoder` under the covers? If so, won't `expected` and `actual` always match, even if there is a bug in `PNGImageDecoder`? I wonder if it would be better to keep the original `BufferedImage` around (the one we draw on), use it as `expected`, and compare it to the two model-generated images.
>
>> Am I correct in assuming that both models end up using PNGImageDecoder under the covers?
> 
> It'd make a lot of sense if they shared the same code, but no. It is my understanding we have two separate decoders. ImageIO uses the com.sun.imageio.plugins.png.PNGImageReader , and this PR modifies the sun.awt.image.PNGImageDecoder.
> 
> The fact that the PNGImageDecoder is a little slower than the ImageIO classes came to my attention because I was comparing the two decoding models, and the older sun.awt classes [appear to be faster](https://docs.google.com/spreadsheets/d/1SoiqnDPSVALb4xraq5hBIGAQLrOQTzA-XuVDohZbCJs/edit?usp=sharing) than the newer ImageIO classes -- except for this one case. This PR will fix this discrepancy.
> 
> (Separately: I've submitted a few other PRs that similarly focus on the older sun.awt decoding classes. See [JDK-8357034](https://github.com/openjdk/jdk/pull/25264) , [JDK-8356320](https://github.com/openjdk/jdk/pull/25076) , [JDK-8356137](https://github.com/openjdk/jdk/pull/25044) , [JDK-8351913](https://github.com/openjdk/jdk/pull/24271) . They generally compare the two models against each other for correctness. I also visually inspected the results at the time to triple-check, but in all of those cases ImageIO was already "doing the right thing". I guess I've made it my goal to bring the older sun.awt decoding classes up-to-date.)
> 
>> I wonder if it would be better to keep the original BufferedImage around (the one we draw on), use it as expected, and compare it to the two model-generated images.
> 
> Sure. I updated this test so it avoids decoding with ImageIO.

I had no idea, thanks for clarifying!

-------------

PR Review Comment: https://git.openjdk.org/jdk/pull/29004#discussion_r2661180031