[Mlir-commits] [mlir] [mlir][arith] Fold addi(x, not(x)) -> -1 (PR #212272)

[Victor Perez]

victor-eds wrote:

> How about generalize these and use `^ -1` hoisting:
> 
> ```ocaml
> addi(not(x), y) -> not(subi(x, y))
> addi(x, not(y)) -> not(subi(y, x))
> subi(not(x), y) -> not(addi(x, y))
> ```
> 
> and for special case when `x == y` next iteration will fold it to `-1`:
> 
> ```ocaml
> not(subi(x, x)) -> not(0) -> -1
> not(subi(x, x)) -> not(0) -> -1
> ```
> 
> Using hoisting unlocks more optimization DRRs like:
> 
> ```ocaml
> addi(subi(not(x), c0), c1) -> 
>   addi(not(addi(x, c0)), c1) ->
>     not(subi(addi(x, c0), c1)) -> not(addi(x, c0 - c1))
> ```

Thanks for the suggestion! I like it: hoisting the `^ -1` outward unlocks the further optimizations like the `addi(subi(not(x), c0), c1)` chain.

The one caveat is that those hoisting rewrites create new operations: `addi(not(x), y) -> not(subi(x, y))` has to materialize a `subi` and a `xori`. So they belong as canonicalization patterns, not in the folder. What this PR adds is a pure fold, which will be applied in more contexts.

I'd suggest we land this fold as-is and do the `^ -1` hoisting as a follow-up canonicalization set.

Sound good?

https://github.com/llvm/llvm-project/pull/212272