On Mon, 2 Jun 2025 20:22:28 GMT, Laurent Bourgès <lbour...@openjdk.org> wrote:

>> PR to replace the use of sun.misc.Unsafe memory access methods in the Marlin 
>> rasterizer with FFM.
>> 
>> I broke this up into the following commits. The bulk of the work is in the 
>> first two:
>> 
>> 1. Encapsulate all off-heap access in OffHeapArray -- All of the memory 
>> allocation and management was already done in the OffHeapArray class, but 
>> the users of OffHeapArray, primarily Renderer and RendererNoAA, called 
>> Unsafe methods directly. I encapsulated all of the calls to Unsafe in 
>> OffHeapArray. The main impact on calling code is that the base address is no 
>> longer accessible or needed. Instead, the `(put|get)(Byte|Int)` methods take 
>> only an offset. This commit was straight refactoring with no behavioral 
>> changes.
>> 2. Initial FFM implementation -- I changed the memory management and access 
>> methods to use FFM. Each OffHeap array uses a shared Arena to manage the 
>> single memory segment allocated at construction time. The resize method 
>> creates a new Arena and memory segment, copying the data from the old and 
>> then closing it
>> 3. Set `used` to 0 in `dispose()` -- While testing and instrumenting the 
>> code, I discovered that the Renderer dispose methods resize the edges array 
>> back to the default size without clearing the "used" field. The used field 
>> will be cleared before the next time it is accessed, but clearing it in 
>> dispose allows optimizing resize to not copy any data.
>> 4. Remove '--sun-misc-unsafe-memory-access=allow' from test and app 
>> execution, since it is no longer needed. This also enables `-Werror` for the 
>> `javafx.graphics` module.
>> 5. ~~Temporary debug prints that will be removed before making this "rfr"~~
>> 
>> Additional commits address review comments.
>
> Few mlnor comments. It looks good to me !

If it helps at this point, we did an initial prototype in the sandbox: 
https://github.com/openjdk/jfx-sandbox/tree/Marlin_FFM with several commits and 
reverts, so might need to go through them. What @bourgesl and I weren't sure 
about was which `Arena` to use. I thought `Confined` was more suitable than 
`Shared` because there's no multithreaded access.

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

PR Comment: https://git.openjdk.org/jfx/pull/1814#issuecomment-2937611540

Reply via email to