[llvm-branch-commits] [flang] [mlir] [flang][OpenACC] generate Destroy region to free memory of private and firstprivate if needed (PR #162702)

[Valentin Clement バレンタイン クレメン via llvm-branch-commits]

clementval wrote:

> > Question: I followed the Destroy design from OpenACC.td and also passed the original value. I am wondering if that it is needed
> 
> This issue was previously discussed [here](https://github.com/llvm/llvm-project/pull/156545#discussion_r2317296732), and my reasoning is based on the principles of consistency, generality, and flexibility:
> 
> * Consistency: All recipe regions take "original" as the first argument, ensuring uniformity across the design.
> * Generality: If a dialect does not encode sufficient information within its private variable to destroy it, the necessary type information can be retrieved from the original variable. While this is not crucial for CIR, FIR, and memref, it provides a robust fallback.
> * Flexibility: Including the original variable allows for custom recipe bodies. For example, if a runtime listener call is embedded within the recipe body, it could report both the address of the original variable and the private copy upon destruction. Although this is also a weaker argument, it highlights potential use cases.
> 
> Moreover, the additional argument will be effectively a no-op after recipe materialization if the dialect-specific recipe implementation does not require it, thus incurring no cost. However, should we need to incorporate this in the future, it would involve considerable rework and IR test modifications.
> 
> Since we are revisiting this design decision, I would like to either get buy-in or agree to change it and remove the extra argument now. @erichkeane @clementval

I was more to remove it in the previous PR but I'm fine with either choice. 

https://github.com/llvm/llvm-project/pull/162702