fix(structured): correct memref pointer and wraparound lowering

[maochuanjun]

Summary

Fix StructuredToMemref pointer analysis and lowering for rank-1 wraparound, non-unit and zero strides, side-by-side accesses, and masked gather/scatter paths.

The previous implementation was primarily shaped around rank-2 split pointers and unit-stride assumptions. It could compute incorrect offsets for linear or side-by-side wraparound, use invalid mask dimensions, or apply fusion patterns whose address assumptions were not valid.

Problems addressed

• Rank-1 wraparound split pointers were not fully materialized.
• Linear wraparound could calculate an incorrect second segment start or size, especially for non-unit or zero strides.
• Side-by-side wraparound mixed row and column stride calculations.
• Mask dimensions could be shorter than the pointer rank.
• Masked reduce/store fusion could be selected for non-unit-stride pointers, where its address model is invalid.
• Rank-1 gather/scatter lowering needed a tensor-based fallback instead of relying on the problematic memref materialization path.

Changes

• Harden analysis helpers:

  • return nullopt for empty OpFoldResult integer queries;
  • avoid propagating an AndOp restriction through an absent structured dimension;
  • extend mask dimensions to cover the rank of GatherScatterPtr and
    MakeTensorPtr.

• Add explicit linear wraparound support:

  • introduce the WRAP_LINEAR representation;
  • calculate rank-1 split sizes and second-segment starts safely;
  • handle runtime zero stride without invalid division;
  • materialize rank-1 split tensor pointers and casts.

• Correct side-by-side wraparound:

  • derive offsets from the in-row column position and column stride;
  • handle zero row stride explicitly;
  • avoid treating row stride as column displacement.

• Restrict unsafe fusions:

  • apply masked reduction and masked elementwise-store fusion only when the
    rank-1 pointer has a statically unit stride.

• Extend split/gather/scatter lowering:

  • support rank-1 and rank-2 split-pointer rewriting;
  • add linear copy, subview, slice, and masked-store paths;
  • lower rank-1 gather/scatter through tensor.empty, linalg.fill, scf.for, and tensor.insert;
  • retain the existing memref allocation path for higher-rank cases.

Regression coverage

New tests:

• wraparound_linear_stride.mlir

  • rank-1 linear wraparound with non-unit stride.

• wraparound_linear_zero_stride.mlir

  • runtime zero-stride handling, including safe select and divsi paths.

• wraparound_side_by_side_non_unit_stride.mlir

  • side-by-side accesses with non-unit column stride and zero row stride.

• tensor_first_vector_cpu_strided_load.mlir

  • verifies that a stride-2 load does not take the masked-elementwise fusion path.

Updated tests cover:

• rank-1 and rank-2 elementwise stores;
• splat stores;
• static side-by-side wraparound;
• gather/scatter pointer lowering, including loop-contained paths;
• rank-1 masked reduction fallback behavior.

[WuXintong123]

The rank-1 lowering represents the access with exactly two strided memrefs, which is only correct when the sequence wraps at most once. For example, start=8, modulo=10, stride=3, size=8 should access 8,1,4,7,0,3,6,9, but the second memref accesses 1,4,7,10,13,16,19. Please either support repeated wraparound or explicitly reject/prove impossible cases before applying this lowering, and add a regression test covering a second wrap.

[maochuanjun]

Thanks for pointing this out. I have replaced the two-memref rank-1 lowering, which could only represent one wrap, with elementwise lowering.

Also I added regression tests for repeated-wrap loads/stores, negative strides, negative initial offsets, and runtime zero stride.

PTAL,thx.