[Mlir-commits] [mlir] [mlir][vector] Add mask elimination transform (PR #99314)

[Andrzej Warzyński]

banach-space wrote:

> * Could we move the trivial pattern-matching cases to the folder/canonicalizer?

+1 IIUC, Diego is asking for more code re-use. In particular, if similar logic is already available somewhere in folder/conanicalization "landscape", it should be re-used here:
```cpp
struct UnknownMaskDim {
    size_t position;
    Value dimSize;
};

// Return `Failure` if at least one Mask dim is statically known not to be "all true". Otherwise,
// returns a vector of `UnknownMaskDim` yet to be analysed to see whether these are "all-true" lanes.
// If the return value is an empty vector, all Mask dims are statically known to be "all-true".
FailureOr<SmallVector<UnknownMaskDim>> getPossiblyTrueMaskDim() {
  SmallVector<UnknownMaskDim> unknownDims;
  for (auto [i, dimSize] : llvm::enumerate(createMaskOp.getOperands())) {
    if (auto intSize = getConstantIntValue(dimSize)) {
      // Mask not all-true for this dim.
      if (maskTypeDimScalableFlags[i] || intSize < maskTypeDimSizes[i])
        return failure();
    } else if (auto vscaleMultiplier = getConstantVscaleMultiplier(dimSize)) {
      // Mask not all-true for this dim.
      if (vscaleMultiplier < maskTypeDimSizes[i])
        return failure();
    } else {
      // Unknown (without further analysis).
      unknownDims.push_back(UnknownMaskDim{i, dimSize});
    }
  }
  return unknownDims;
}
```

>From what I can tell, you can wrap this into an utilility funciton ,e.g. 
```cpp
SmallVector<>
```

> * Could we use `constant_mask` instead of `arith.constant`? This is not enforced in any way but we've been using the former to represent vector masks (they are kind of redundant, actually).

+1 to this suggestion. @dcaballe , are you OK to implement this as a follow-up?

https://github.com/llvm/llvm-project/pull/99314