[RemoveLayoutConversions] Fix backward remat conflict detection

[etiotto]

Port upstream Triton PR #9953: when getConvertBackwardSlice encounters
a value that already has the desired encoding, record the layout for
conflict detection without adding it to the remat slice. Previously the
layout was not recorded, so a conflicting encoding via a different path
went undetected.

Closes #6700

[etiotto]

No performance regressions detected on Grafana:

https://benchmarks.glados.intel.com/d/1pXX4hUSz/microbenchmarks?orgId=1&var-tag=ci&var-tag=client&var-tag=pr-6693&var-bench=All&var-device=Max%201550&var-device=B580&var-compiler=triton&var-backend=All&var-table_old=ci&var-baseline_backend_old=triton-ci-Max%201550&var-target_backend_old=triton-ci-Max%201550&var-okr_device_old=All

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Ports upstream Triton fix to improve backward rematerialization conflict detection by ensuring layouts are recorded even when a value already has the desired encoding, preventing undetected encoding conflicts.

Changes:

• Always record encountered value layouts for conflict detection, even when the value already has the target encoding.
• Adjust slice construction so values with matching encodings are not added to the remat slice.
• Add MLIR regression tests covering conflict scenarios across multiple paths/ops.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 4 comments.

File	Description
third_party/intel/lib/TritonIntelGPUTransforms/Utility.cpp	Updates backward-slice logic to record layouts while avoiding unnecessary slice entries when encodings already match.
lib/Dialect/TritonGPU/Transforms/Utility.cpp	Mirrors the upstream fix in the core TritonGPU utility implementation.
test/TritonIntelGPU/RemoveLayoutConversions/backward-remat-conflict.mlir	Adds regression tests to validate conflict detection and preservation/elimination of convert_layout as appropriate.