Improve the remove layout in handling loop op.

[chengjunlu]

This PR improves the Intel GPU RemoveLayoutConversions pass to better handle loop-related values during backward/forward rematerialization, aiming to avoid duplicated loop bodies and improve kernel performance (referenced in issue #7604).

chunk_gated_delta_rule_fwd_kernel_h_blockdim64 kernel time 207us vs 235us in #7604

[Copilot]

Pull request overview

This PR improves the Intel GPU RemoveLayoutConversions pass to better handle loop-related values during backward/forward rematerialization, aiming to avoid duplicated loop bodies and improve kernel performance (referenced in issue #7604).

Changes:

• Extend forward propagation to traverse/control loop ops and handle tt.descriptor_store source rewrites.
• Improve forward-propagation candidate collection to account for loop iter-args / loop results (to prevent duplicated loop bodies after rematerialization).
• Add an MLIR regression test reproducing the loop-related inefficiency.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
third_party/intel/lib/TritonIntelGPUTransforms/RemoveLayoutConversions.cpp	Updates rematerialization forward-propagation logic and loop-related candidate selection; adds extra debug logging.
test/TritonIntelGPU/RemoveLayoutConversions/remove_layout_conversions_7625.mlir	Adds a regression test covering loop handling in remove-layout-conversions.

[whitneywhtsang]

The removal of the scf::ForOp/scf::IfOp skip is too broad. The needRewrite path below calls rewriteOp(&op, ...) which is not designed for structured control flow ops — if a ForOp result happens to be in valuesToPropagate, this will attempt to rewrite the loop itself.

Consider keeping the skip for the needRewrite path while still entering regions for the DescriptorStoreOp case.

[chengjunlu]

Add back the guard of the rewriteOp. The value in the valuesToPropagate is chosen conservitively of the element-wise operations. But anyway we have added the protection.

[etiotto]

LGTM.

One optional non-blocker: the assert(newOperand) in the DescriptorStoreOp path (RemoveLayoutConversions.cpp) compiles out in release builds — an if (!newOperand) continue; would be safer. (The dead sliceOps is noted inline.) Neither blocks merge.