[Mlir-commits] [mlir] [MLIR][Transforms] add eliminate-explicit-rounding pass (PR #93443)

[Ivy Zhang]

crazydemo wrote:

> Re the whole attribute thing above, how about this:
> 
> 1. Add fast-math flags to `arith.truncf` and `arith.extf` if they don't have them already. While they aren't present in LLVM, that's fine, we're allowed extensions
> 2. Define `contract` for `trunf` and `extf` to allow this fold
> 3. In the various promotion passes, set "contract" on the truncations and extensions in addition to their existing fastmath (maybe make that a default-on option)
> 4. Define a folder or canonicalizer on `arith.extf` and `arith.truncf` that collapses
>    ```mlir
>    %1 = arith.extf fastmath(contract) %0 : T to U
>    %2 = arith.truncf fastmath(contract) %1 : U to T
>    ```
>    
>    
>        
>          
>        
>    
>          
>        
>    
>        
>      
>    to just `%0` for all `T` and `U`

Thanks for your suggestions. Currently, this PR and your comment are functionally consistent; the main difference lies in the form of the pass and attribute. I think we can proceed with merging this PR for now.

https://github.com/llvm/llvm-project/pull/93443