dialects: (linalg) support tiling of linalg.generic ops over tensors

[dipo101]

Part of #6140, which tracks Linalg tiling support missing relative to upstream MLIR. This covers the tensor-based tiling checkbox.

Tiling a linalg.generic over memrefs slices each operand with a memref.subview and writes results through that view. Tensors have value semantics, so a tile is a new value that has to be written back explicitly and carried out of the loops it was computed in. That difference is what this adds.

For an op over tensors, tiling now produces:

%C = scf.for %5 = %0 to %1 step %3 iter_args(%6 = %B) -> (tensor<4x4xf32>) {
  %7 = scf.for %8 = %0 to %2 step %4 iter_args(%9 = %6) -> (tensor<4x4xf32>) {
    %10 = tensor.extract_slice %A[%5, %8] [2, 2] [1, 1] ...
    %11 = tensor.extract_slice %9[%5, %8] [2, 2] [1, 1] ...
    %12 = linalg.generic ins(%10) outs(%11) ... -> tensor<2x2xf32>
    %13 = tensor.insert_slice %12 into %9[%5, %8] [2, 2] [1, 1] ...
    scf.yield %13 : tensor<4x4xf32>
  }
  scf.yield %7 : tensor<4x4xf32>
}

The outermost loop carries the original output tensor and each nested loop carries the enclosing loop's block argument, so the output tile is extracted from the value accumulated so far rather than from the original. The outermost loop's results replace the results of the original op.

Only scf.for is generated. xDSL has no scf.forall, so the parallel form upstream also supports is out of scope here.

Commits

1. dialects: (tensor) add mixed static/dynamic builders to extract_slice — tensor.extract_slice could only be built with fully static offsets, sizes and strides, so it could not express a slice taken at a loop induction variable. Adds get() and infer_result_type(), mirroring the existing memref.SubviewOp pair.
2. dialects: (linalg) thread loop-carried values through tile loops — gives _build_tile_loops an optional iter_args parameter.
3. dialects: (linalg) slice tensor operands with tensor.extract_slice — widens the operand type and picks the slicing op based on it.
4. dialects: (linalg) write tiles back into the tensors carried by tile loops — adds the tensor.insert_slice write-back and the operand-carrying yields.
5. dialects: (linalg) tile linalg.generic ops over tensors — stops rejecting tensor operands during analysis.

Commits 2 to 4 are each inert on their own, since tensor operands are still rejected until the last one. They are kept separate because each is a distinct change to explain.

Behaviour that is unchanged

Tiling memrefs carries nothing, writes nothing back and yields nothing, so its generated IR is byte-identical to before. The existing filecheck expectations are untouched.

An op that reads a tensor and writes a memref has no results, so nothing is carried and the output is still written through its subview. That case used to be rejected and now tiles; it has a test.

Newly rejected

An op with both memref and tensor outputs is rejected. Such an op verifies today but has fewer results than outputs, and the two kinds of output would have to be written back differently within one loop nest.

Tests

• filecheck: both dimensions tiled, one dimension tiled, and tensor input with memref output
• filecheck rejection: mixed memref and tensor outputs
• unit: the loop-carried threading, the slicing op chosen per operand type, the new rejections, and the extract_slice builders

Open questions for review

• Is rejecting mixed memref-and-tensor outputs the right call, or should that be supported?
• tensor.ExtractSliceOp has no verify_ checking that the dynamic index markers match the operand count, unlike memref.SubviewOp. This is pre-existing and deliberately left out of this PR, but happy to add it here if preferred.
• ExtractSliceOp.infer_result_type takes no reduce_rank flag, unlike the memref one, since nothing needs it yet.
• The emitted IR has not been diffed against mlir-opt output, as the MLIR-conversion tests are opt-in.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 86.95%. Comparing base (68e9b73) to head (fce5d67).

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #6362   +/-   ##
=======================================
  Coverage   86.94%   86.95%           
=======================================
  Files         440      440           
  Lines       65983    65999   +16     
  Branches     7493     7493           
=======================================
+ Hits        57368    57387   +19     
+ Misses       7037     7036    -1     
+ Partials     1578     1576    -2     

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[superlopuh]

This is great! This PR is too large to review IMO, and you've already broken it down nicely, so let's iterate in smaller PRs. I'll mark this PR as draft, can you please open a new one with just the first commit to start with?

[dipo101]

Thanks! Opened #6366 with just the first commit (the tensor.extract_slice builders). I'll keep this PR as the draft overview and open the following ones as each lands.