[TensorDescriptor] Fallback to TritonRewriteTensorDescriptorToPointer

[whitneywhtsang]

This PR adds fallback logic for tensor descriptor gather and scatter in Triton, integrates the upstream rewrite pass in the pipeline, and updates tests accordingly.

Fixes #4267

[Copilot]

Pull Request Overview

This PR adds fallback logic for tensor descriptor rewrites in Triton, integrates a new rewrite pass in the Python compiler pipeline, and updates tests accordingly.

• Detect and skip modules containing unsupported descriptor_gather/descriptor_scatter ops before converting descriptors.
• Invoke rewrite_tensor_descriptor_to_pointer in compiler.py.
• Add an MLIR test to ensure descriptors aren’t converted when used by descriptor_gather.
• Remove XPU-specific skips in Python unit tests.

Reviewed Changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
third_party/intel/lib/Dialect/Triton/Transforms/TensorDescToBlockPointer.cpp	Add unsupported-op prewalk and skip logic for fallback case
third_party/intel/backend/compiler.py	Register the new rewrite pass in the Triton compiler pipeline
test/Triton/Intel/TensorDescToBlockPointer/invalid.mlir	New invalid.mlir to verify descriptor not rewritten on gather
python/test/unit/language/test_tensor_descriptor.py	Removed XPU skips in several test cases

Comments suppressed due to low confidence (2)

test/Triton/Intel/TensorDescToBlockPointer/invalid.mlir:3

• Consider adding a parallel test for descriptor_scatter to verify the fallback logic handles that unsupported op similarly.

// COM: Test make_tensor_descriptor is not rewritten when it is used by descriptor_gather.

python/test/unit/language/test_tensor_descriptor.py:1332

• Removing the XPU-specific skip may cause failures on XPU devices; either update these tests to support XPU or reintroduce a conditional skip based on actual unsupported features.

if is_xpu():