TileMapLayer memory leak when using Bucket Tool

[Silvanuz]

Tested versions

Reproducible in v4.4.1.stable.official

System information

Godot v4.4.1.stable - Windows 11 (build 26100) - Multi-window, 2 monitors - Vulkan (Forward+) - dedicated NVIDIA GeForce RTX 4060 (NVIDIA; 32.0.15.7283) - AMD Ryzen 5 7600X 6-Core Processor (12 threads)

Issue description

When working with the Engine for a few hours without closing it, I noticed an huge RAM usage:

I noticed that the more I work on my levels with tilesets, the higher the RAM usage becomes.
So I found out that if you edit a tileset frequently with the bucket tool, the RAM keeps increasing but no longer decreases.

With terrain:
https://github.com/user-attachments/assets/14d84fe9-fbee-422a-973f-0a7d1c6847c4

Single Tile:
https://github.com/user-attachments/assets/236f785d-0b1e-4ee6-877f-dc6adc86214f

Steps to reproduce

• Create a TileMapLayer with a fresh Tileset
• Add a physic layer and a terrain set
• Setup the Terrain Set (Bitmask, Collisions)
• Draw a block on the tilemap
• Choose the Bucket tool and click repeatedly on the block

MRP for the exact Setup!

Minimal reproduction project (MRP)

MRP Tilemap memleak.zip

[precup]

I can confirm this is happening. Closing and reopening the scene in question seems to release the memory, there doesn't seem to be a permanent leak, but the memory use does seem to increase without bound while using the bucket tool. I went from < 1 GB of memory used to 3+ GB by doing spam clicking the bucket tool for a little while.

Godot v4.4.1.stable (49a5bc7) - Windows 10 (build 19045) - Multi-window, 3 monitors - Vulkan (Forward+) - dedicated NVIDIA GeForce RTX 2080 Ti (NVIDIA; 31.0.15.3713) - AMD Ryzen 9 3900X 12-Core Processor (24 threads)

[chocola-mint]

I've been looking into this, and I've found something quite worrying about the TileMapLayerEditor in general.

With the same MRP, pick the bucket tool, select a tile to draw, and just keep moving your cursor across the grid. The memory leak happens even without actually performing the operation.

godot-tilemap-layer-editor-memory-leak.mp4

And it works with the brush tool as well!

Perhaps there might be a memory leak with draw_string?

(For other contributors interested in investigating, Visual Studio's heap profiler can be useful here)

(Branch is 4.5-beta)

[chocola-mint]

A quick test reveals that the actual "memory leak" here is just, lots and lots of redundant do/undo methods being recorded by the bucket tool.

The culprit here is TileMapLayerEditorTilesPlugin::_draw_bucket_fill, which inserts tiles to the output even if the tile isn't actually changed, and a tentative fix was able to successfully address the memory leak.

The same issue may technically affect other tools though, so I'll work on a more thorough fix.

[groud]

A quick test reveals that the actual "memory leak" here is just, lots and lots of redundant do/undo methods being recorded by the bucket tool.

Oh I was wondering. Had a quick look at the code and couldn't find what could cause that. The fix is likely a bit easy here (checking if values are different before adding them to the list).

But the bad news is that, if that's the main issue, I don't think we'll be able to optimize undo-redo here much, aside from getting rid of the oldest changes in the UndoRedo stack.