Fix sliding-window sequence length mask detection

[umangyadav]

Summary

• validate that sliding-window masks reference an i32 current-sequence-length block argument
• preserve and reconcile sequence-length clip bounds when folding sliding-window and KV-cache masks
• add regression coverage for invalid operands and mismatched sequence lengths or clip bounds

Test plan

• ninja rocmlir-opt
• TosaToRock attention lit tests (12/12 passed)
• git clang-format --diff origin/develop
• ninja check-rocmlir (11,802 passed; one unrelated tuningRunner-gemm.py infrastructure failure because the build directory was not found)

[Copilot]

Pull request overview

This PR tightens rocMLIR’s TOSA→Rock attention mask folding by ensuring sliding-window mask detection is anchored to a validated i32 current-sequence-length block argument, and by reconciling (or rejecting) mismatched sequence-length clamp bounds when combining sliding-window and KV-cache masking.

Changes:

• Enhance sliding-window mask detection to require the non-constant operand resolves to an i32 block argument (currentSeqLen), and to extract any surrounding clip bounds.
• Reconcile sliding-window and KV-cache currentSeqLen + clip bounds after all select masks are analyzed, making results independent of select nesting order.
• Add a regression test covering a non-sliding-window “wrong operand” case and a no-KV-cache sliding-window case that must preserve the validated clip.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
mlir/lib/Conversion/TosaToRock/TosaToRock.cpp	Adds validated sliding-window seq-len resolution + clip tracking, and reconciles it with KV-cache-derived seq-len/clip constraints.
mlir/test/Conversion/TosaToRock/tosa-to-rock-attention-sliding-window-neg.mlir	Adds regression coverage for correct sliding-window adoption of clipped currentSeqLen, and for rejecting masks with an unrelated operand.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[rocmlir-pr-reviewer]

Verdict: APPROVE -- submitted as COMMENT (automated reviews are advisory)  ·  Findings: 0 (0 Critical, 0 Major, 0 Minor)

---

Scope

Fixes sliding-window sequence-length mask detection in the TosaToRock attention rewrite (mlir/lib/Conversion/TosaToRock/TosaToRock.cpp) and adds Lit coverage (mlir/test/Conversion/TosaToRock/tosa-to-rock-attention-sliding-window-neg.mlir). Previously trySlidingWindowPattern returned only a window size without checking that the non-constant add operand is really an i32 currentSeqLen block argument, so an unrelated greater(x - const, col) could be misclassified. The PR validates that operand, threads through detected clip bounds, and reconciles the KV-cache vs. sliding-window seq-len block arg and clips after all masks are analyzed.

Findings

No blocking issues found.

Notes

• Correctness spot-checks pass: the aggregate initializer at TosaToRock.cpp:2622 matches the updated SeqLenMaskResult field order at TosaToRock.cpp:2110; SlidingWindowResult/slidingWindowSeqLen are consumed only at the analyze site (2588) and the reconciliation block (2660-2675). The new sameSeqLenBlockArg and clip storage mirror the existing KV-cache design (both store the resolved block arg plus optional clip bounds), and the reconciliation re-resolves both operands through seqLenSkip before comparing, so it cannot dereference a null value.
• Conservative-failure behavior is intentional and safe: when the KV-cache and sliding-window masks reference different block args or different clip bounds, getSeqLenMask returns failure() and abandons the whole rock.attention rewrite rather than emitting a possibly-incorrect op.
• Minor (non-blocking, not a checklist item): the file is named ...-neg.mlir, but the first function sliding_window_no_kvcache is a positive case (it does detect a sliding-window mask and adopts its seq-len); only the second function is a true negative. Naming is slightly misleading but harmless.

CI status

No non-self CI failures. C/C++ premerge checks, Python format and lint checks, and Python performance script tests all pass; the in-progress review check is this auto-review pipeline and is expected.

[codecov]

Codecov Report

❌ Patch coverage is 91.48936% with 4 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
mlir/lib/Conversion/TosaToRock/TosaToRock.cpp	91.49%	0 Missing and 4 partials ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #2436      +/-   ##
===========================================
+ Coverage    82.57%   83.57%   +1.00%     
===========================================
  Files          120      120              
  Lines        42852    42977     +125     
  Branches      7110     7150      +40     
===========================================
+ Hits         35381    35914     +533     
+ Misses        4815     4492     -323     
+ Partials      2656     2571      -85     

Flag	Coverage Δ
gfx120x	83.42% <91.49%> (+0.90%)	⬆️
gfx950	83.38% <91.49%> (+1.04%)	⬆️
mfma	83.42% <91.49%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

Files with missing lines	Coverage Δ
mlir/lib/Conversion/TosaToRock/TosaToRock.cpp	79.39% <91.49%> (+0.98%)	⬆️

... and 32 files with indirect coverage changes

🚀 New features to boost your workflow:

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

[bogdan-petkovic]

LGTM