fix(ir): accept TileType input in tensor.gather conversion

[YunjiQin]

Summary

tensor.gather previously crashed in convert_tensor_to_tile_ops with CHECK(input_tensor_type) when an upstream conversion had already lowered its input/index to TileType (e.g. when a preceding tensor.create had been rewritten to tile.create).

This PR makes the converter accept either TensorType or TileType for both input and index:

• Extract shape/dtype from whichever type the arg carries.
• Route the per-row materialization through a new emit_load_or_slice helper: it emits tile.load when the source is still a TensorType (original behavior, valid at any rank) and tile.slice when an upstream pass has already produced a TileType (only reached on rank-2 paths today, because FlattenTileNdTo2D rejects tile.slice on >2D tiles).
• The rank-3 paths therefore keep emitting tile.load, so the rank-3 lowering is bit-identical to before.

Test plan

• tests/ut/ir/transforms/test_convert_tensor_to_tile_ops.py::TestConvertGatherOp — UTs pass
• tests/st/runtime/ops/test_gather.py::TestGatherIndex rank-3 cases — pre-compile path no longer trips FlattenTileNdTo2D: tile.slice is not supported on >2D tiles; system-tests / dist-system-tests CI lanes green
• CI: clang-tidy / pre-commit / codegen-tests / pypto-lib-model / system-tests / system-tests-a5sim / unit-tests (ubuntu+macos) / dist-system-tests — all green

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro

Run ID: c1fd0f09-846b-4921-8fff-22ac198e0f99

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The PR refactors the tensor.gather lowering strategy from per-row tile loads to preloaded Vec tiles with per-iteration slicing. Input and index operands are now bridged upfront to Vec tile memory, then sliced per loop iteration instead of loaded individually. All four gather cases and their corresponding test expectations are updated accordingly.

Changes

Gather Operation Lowering Refactoring

Layer / File(s)	Summary
Gather input space bridging and conversion refactoring src/ir/transforms/op_conversion_registry.cpp	gather_input_reqs map bridges both input and index to Vec tiles. The conversion lambda now type-checks for TileType instead of TensorType, derives shapes and dtype from bridged tiles, and removes the previous load_kwargs initialization. Registration passes std::move(gather_input_reqs) to apply bridging.
Gather lowering cases refactored to tile.slice src/ir/transforms/op_conversion_registry.cpp	All four rank/dim cases (rank==2 norm_dim==1, rank==3 norm_dim==2, rank==3 norm_dim==0, rank==3 norm_dim==1) replace per-iteration tile.load with tile.slice on the preloaded Vec tiles. Reshapes are applied where tile.slice produces higher-rank results before calling the existing gather math.
Test expectations for new gather lowering strategy tests/ut/ir/transforms/test_convert_tensor_to_tile_ops.py	Test comments updated to describe the new structure: single upfront pl.load of full input/index tensors into Vec buffers, then per-iteration pl.tile.slice to extract row/index tiles before pl.tile.gather. Expected IR assertions match the new lowering flow.

Estimated code review effort

🎯 3 (Moderate) | ⏱️ ~20 minutes

Poem

A gathered wish on tiled ground,
Where slices dance in circles 'round,
The bunny loaded, bridged with care,
Each gather hop through tile-sliced air! 🐰✨

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Title check	✅ Passed	The title accurately summarizes the main change: the tensor.gather conversion now accepts TileType input in addition to TensorType, fixing a CHECK failure.
Description check	✅ Passed	The description clearly explains the problem (CHECK crash when input/index are already TileType), the solution (accept both types and route through emit_load_or_slice), and provides comprehensive test coverage details.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[gemini-code-assist]

Code Review

This pull request updates the tensor.gather operation conversion to bridge input and index tensors to MemorySpace::Vec tiles, replacing row-by-row tile.load operations with tile.slice on the pre-loaded tiles. Unit tests have been updated to reflect these changes. Feedback is provided to use INTERNAL_CHECK_SPAN instead of CHECK for verifying internal invariants when a Span is available, ensuring source location information is included in error messages.