fix(ir): handle mixed tensor store-load bridges

[lwDavid]

Summary

Fix ExpandMixedKernel for mixed V/C kernels where cross-core data transfer happens through tile.store -> tile.load
instead of an explicit tile.move.

Changes

• detect implicit tensor bridges in mixed kernels
• lower V->C tensor bridges to tpush/tpop
• match bridge producer/consumer by SSA name instead of Var*
• reuse existing boundary-move data structures and emission logic
• add A2A3 regression coverage for the tensor-bridge case

Fixes #965

[coderabbitai]

📝 Walkthrough

Walkthrough

This PR implements tensor-bridge pattern detection in the mixed-kernel expansion pass to identify tile.store and tile.load producer-consumer pairs crossing core affinities (VECTOR ↔ CUBE), emitting tpush/tpop instructions for cross-core transfers instead of free tile aliasing.

Changes

Cohort / File(s)	Summary
Tensor-Bridge Pattern Detection & Emission src/ir/transforms/expand_mixed_kernel_pass.cpp	Added CollectTensorStoreLoadBoundaries() to detect store-load chains via name-hint matching and recursive IterArg resolution; introduced GetTensorBridgeDirection() to infer CVDirection for VECTOR↔CUBE transitions; extended BuildCoreBody() signature with tensor_bridge_pushes/tensor_bridge_pops maps; added emit_push()/emit_pop() helpers and early dispatch for tensor-bridge statements to emit tpush/tpop instead of tile.move; propagated new boundary maps through nested MIXED compound statements.
Tensor-Bridge Functional Test tests/ut/ir/transforms/test_expand_mixed_kernel_a2a3.py	Added test_v2c_tensor_bridge_store_load_uses_push_pop_on_a2a3() constructing a MixedTensorBridge program with intermediate tensor store-load crossing VECTOR→CUBE; asserts generated AIC/AIV contain tpush_to_aic/tpop_from_aiv; verifies absence of _nz/_zn substrings and bridge-referencing tile.load/tile.store lines.

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~50 minutes

Possibly related PRs

• feat(ir): support cross-core transfer view on Ascend910B #896: Modifies BuildCoreBody/AIV push-path logic and same expand_mixed_kernel_pass.cpp file for cross-core tile transfer handling.
• feat(codegen): reorder tpop chains for hardware pipe ordering and improve cross-core protocol #693: Extends cross-core tpop/tpush/tfree handling and threads tpop state through BuildCoreBody/ExpandMixedFunction.
• fix(runtime): stabilize mixed-kernel split AIV flow #951: Modifies mixed-kernel expansion and A2A3 cross-core transfer logic with related test coverage.

Suggested reviewers

• lyfne123
• Hzfengsy

Poem

🐰 Whiskers twitch with joy ~

A bridge of tensors spans the cores,
From VECTOR dreams to CUBE's great door—
With push and pop, we flow so right,
No free aliases in our sight! ✨

🚥 Pre-merge checks | ✅ 2 | ❌ 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 (2 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title 'fix(ir): handle mixed tensor store-load bridges' directly and concisely summarizes the main change—handling implicit tensor bridges in mixed kernels via store-load patterns.
Description check	✅ Passed	The pull request description directly addresses the changeset, explaining the fix for ExpandMixedKernel to handle implicit tensor bridges in mixed V/C kernels.

_{✏️ 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 implements a mechanism to handle tensor store/load boundaries during mixed kernel expansion, ensuring that Vector-to-Cube (V->C) tensor bridges are correctly transformed into tpush and tpop operations rather than simple tile aliases. The changes include refactoring BuildCoreBody to use modular emit_push and emit_pop helpers and adding logic to collect tensor store/load pairs. The review feedback correctly identifies a technical issue where the code attempts to access the name_ member directly from an ExprPtr base class; it is recommended to use GetKind() and static_pointer_cast to ensure both correctness and better performance compared to dynamic_pointer_cast.

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against the current code and only fix it if needed.

Inline comments:
In `@src/ir/transforms/expand_mixed_kernel_pass.cpp`:
- Around line 229-236: The function ResolveTensorBridgeInputVar currently checks
for Var before IterArg causing IterArg (which derives from Var) to be returned
prematurely; update ResolveTensorBridgeInputVar to check for
std::dynamic_pointer_cast<const IterArg> first and, if present, recurse on
iter_arg->initValue_, otherwise then check for Var and return var.get(), so
loop/while-carried tensor bridges resolve to their originating tensor via
initValue_.
- Around line 295-299: The loop currently skips producers with multiple
consumers, which silently drops multi-load bridges; instead, when
producer_candidates.size() != 1 create a single shared boundary for that
producer and map push_boundaries[producer_stmt] to it and map every
consumer_stmt in producer_candidates to pop_boundaries with the same boundary so
all cross-core loads use a single transfer (avoid per-side private allocations).
Update the loop over candidates (and the handling of producer_candidates) to
handle the multi-consumer case by constructing one boundary and assigning it to
producer_stmt and to each consumer_stmt rather than continuing/ignoring; use the
existing symbols producer_stmt, producer_candidates, consumer_stmt, boundary,
push_boundaries, and pop_boundaries to implement this.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: fb2c9f81-a008-4b04-afdb-4aa7cf3da812

📥 Commits

Reviewing files that changed from the base of the PR and between 3e680d1 and c8bdcc0.

📒 Files selected for processing (2)

• src/ir/transforms/expand_mixed_kernel_pass.cpp
• tests/ut/ir/transforms/test_expand_mixed_kernel_a2a3.py

[lwDavid]

The issue can be addressed by removing redundant lines in frontend kernel. Convert this PR to draft.

[lyfne123]

I think it not a issue of expandkernle