[tensor-descriptor]: Extend support when tensor descriptor created in control flow

[etiotto]

Enhance layout propagation and tensor descriptor lowering to support cases where descriptors or pointers are created within control flow constructs.

[mfrancepillois]

The code LGTM, but I'm not sure to understand why we need this PR if the conversion from tensor_descriptor to block_pointer is one of the first passes to run at the ttir level, how can we have tensor_descriptor ops with different layouts given the encodings are assigned to ops when lowering from the ttir level to the ttgir level.
As I understand it, we are not supposed to have tensor_descriptor ops after translating to ttgir dialect, no?

[kurapov-peter]

BTW FYI, the existing TMA lowering expects the tensor descriptor to always have a layout.

PR WIP

[etiotto]

The code LGTM, but I'm not sure to understand why we need this PR if the conversion from tensor_descriptor to block_pointer is one of the first passes to run at the ttir level, how can we have tensor_descriptor ops with different layouts given the encodings are assigned to ops when lowering from the ttir level to the ttgir level. As I understand it, we are not supposed to have tensor_descriptor ops after translating to ttgir dialect, no?

I have updated the code quite a bit to change the loginc used to do the translation. The pass is now simpler and able to handle tesnor descriptors created in control flow (i.e. a branch) inside a loop.

[etiotto]

No performance degradation in Triton benchmarks: https://github.com/intel-sandbox/applications.python.intel-xpu-backend-for-triton.infrastructure/actions/runs/15142551794/job/42570177577

[Copilot]

Pull Request Overview

Enhance layout propagation and tensor descriptor lowering to support cases where descriptors or pointers are created within control flow constructs.

• Add updateAdvanceOpChain to recursively update chains of AdvanceOp users.
• Refactor rewriteStoreOp to trace back through AdvanceOp chains before creating new MakeTensorPtrOp.
• Rewrite the TensorDesc-to-block-pointer pass to drop old descriptor lookup, always find or create MakeTensorPtrOp, and handle loop-carried block pointer types.

Reviewed Changes

Copilot reviewed 2 out of 5 changed files in this pull request and generated 2 comments.

File	Description
third_party/intel/lib/TritonIntelGPUTransforms/RemoveLayoutConversions.cpp	Added recursive chain update, improved rewriteStoreOp, and added verification asserts after each transformation stage.
third_party/intel/lib/Dialect/Triton/Transforms/TensorDescToBlockPointer.cpp	Removed legacy descriptor lookup, consolidated pointer creation via findOrCreateMakeTensorPtr, replaced offset logic, and updated loop argument types.

Files not reviewed (3)

• test/Triton/Intel/TensorDescToBlockPointer/basic.mlir: Language not supported
• test/Triton/Intel/TensorDescToBlockPointer/loop.mlir: Language not supported
• test/TritonIntelGPU/backward_combine_dpas_dot_layout.mlir: Language not supported

Comments suppressed due to low confidence (1)

third_party/intel/lib/TritonIntelGPUTransforms/RemoveLayoutConversions.cpp:793

• The variable value is undefined in this scope, leading to a compile error. It should reference the store's value (e.g., storeOp.getValue()) or the converted value extracted earlier.

Value dataToStore = getValueAs(value, encoding);

[etiotto]

@whitneywhtsang I have split this pull request and move the part that deals with removing layout conversion for store operation that use a block ptr updated by a tt.advance operation in PR #4277

[whitneywhtsang]

@whitneywhtsang I have split this pull request and move the part that deals with removing layout conversion for store operation that use a block ptr updated by a tt.advance operation in PR #4277

Thanks, this part of the code LGTM.