Canonicalizers for doubly strided ops such as npu.dma_cpy_nd

[newling]

Doubly strided ops have, for both source and target, and for all of offsets/sizes/strides, 2 fields:

1. a vector of dynamic Values
2. a vector of static integers

For example it might be

1. dynamic = {v0}
2. static = {16, kDynamic, 32, 64}

The size of dynamic is always exactly equal to the number of appearances of kDynamic in the static vector. This pass does the following: it checks if any of the Values in dynamic are actually MLIR constants, and if they are then it removes the Value from the dynamic vector and updates the corresponding index in static. So for example if v0 above is actually

%v0 = arith.constant 6 : index

then this canonicalization updates dynamic/static to

1. dynamic = {}
2. static = {16, 6, 32, 64}

[MaheshRavishankar]

There is a canonicalizer upstream here that does this, but it is a bit hard-coded for tensor.extract_slice like op that have a single offset , size and stride. It should be split up to handle each of the offset , size and stride separately. Then you could reuse that here. Maybe just use the same style here to be consistent

[jtuyls]

Nice, should help make IR more compact and readable! Would be great to do for all doubly strided ops in one go I think.

[MaheshRavishankar]

Actually, if your wait for this PR (llvm/llvm-project#104488) it might get even easier

[newling]

@MaheshRavishankar thanks for the pointer, I've changed to use the same style for now.
@jtuyls ok, will get this working for all the doubly strided ops

[jtuyls]

LGTM, just a few small nits

[jtuyls]

Suggested change

	// Simplified from upstream MLIR's foldDynamicIndexList:
	/// Simplified from upstream MLIR's foldDynamicIndexList:

Nit, but IREE uses /// for function/class comments.

[jtuyls]

Suggested change

	// Based on upstream MLIR's
	// OpWithOffsetSizesAndStridesConstantArgumentFolder
	/// Based on upstream MLIR's
	/// OpWithOffsetSizesAndStridesConstantArgumentFolder

Nit, but IREE uses /// for function/class comments.

[jtuyls]

Suggested change

	// Build a NpuDmaCpyNdOp with mixed static and dynamic entries and target
	// and source BD IDs.
	/// Build a NpuDmaCpyNdOp with mixed static and dynamic entries and target
	/// and source BD IDs.

[newling]

LGTM, just a few small nits

Ok I'll use /// by default in future. As this particular file uses both /// and // quite a lot I'm not going to update this PR.