[Cleanup][Frontend] Dead code, style nits, and minor smells in PyTorchSimFrontend

[YWHyuk]

Summary

Checklist of low-priority cleanup items found while reviewing PyTorchSimFrontend/. None are correctness bugs on their own (the real bugs are filed separately in #236, #237, #238, #239, #240). Filing as a single tracking issue for easy batch cleanup.

Dead code

• extension_op.py:116-179 — generate_inner_product_matrix is 63 lines with no call sites anywhere in the repo. Safe to delete.
• extension_op.py:285-288 — sparse_mm_dummy_stonne_outer uses yield, making it a generator function. It is referenced from the wrapper header in mlir_codegen_backend.py but is never iterated, so calling it returns a generator object, not the result of out.copy_. Either drop the yield or remove the function.
• extension_codecache.py:249-251 — w_offset is computed conditionally then unconditionally overwritten with 0. Drop the dead branch (commented-out original expression suggests this is a known hack).
• mlir_codegen_backend.py:317, 1410 — dma_write_counter is initialized but never incremented; only dma_read_counter is bumped (line 1407). If MVOUT is always type MVOUT1, drop dma_write_counter to avoid confusion.

Copy-paste / misleading messages

• extension_codecache.py:137-139 — TileSizeError.__init__ default message is "SPAD overflow occurred.", identical to SpadOverflowError's message (copy-paste). Change it to something like "Tile size constraint violated."
• extension_codecache.py:170, 220 — MLIRCodeCache.load instantiates two FileLock objects against the same lock-file path with a non-locked gap between them. The on-disk file is the actual serialization primitive so there is no race, but folding both with lock: blocks under one acquisition is clearer.
• extension_codecache.py:19-23 and mlir_autotune.py:19-23 — get_write_path and hash_prefix are duplicated verbatim across the two modules (the comment notes circular-import avoidance). Move both to extension_config.py and import from there.

Style / Pythonic nits

• mlir_common.py:20, 25 — import sympy appears twice. Remove the second one.
• mlir_common.py:556 — MLIRMultiDimTile.get_nr_dim is annotated -> str but returns len(self._tile_size), an int. Fix the annotation.
• mlir_scheduling.py:185 — bare except: swallows KeyboardInterrupt / SystemExit. Switch to except Exception:.
• mlir_scheduling.py:63 — revert_group has no return statement, but the caller does node1 = self.revert_group(node1) and then never uses node1 again. Drop the assignment or have revert_group return its argument.
• Mutable default arguments across the codegen stack — none are currently mutated, but they are a long-running foot-gun:
  • extension_codecache.py:155, 270 — arg_attributes=[], tile_size=[]
  • mlir_common.py:203, 1153 — extra_node=dict(), affine_yield=dict()
  • mlir_scheduling.py:310 — origins={}
  • mlir_codegen_backend.py:435 — indirect_dims=[]
    Replace each with None and initialize in the body.
• Active code paths mixing print() with the module logger. Examples:
  • mlir_codegen_backend.py:486
  • extension_config.py:159
  • extension_op.py:176-178, 187-188
  • mlir_caller_codegen.py:199-200
    These bypass log-level filtering. Route through logger (or remove if debug-only).

Duplication (templates) — defer if too invasive

The four conv templates (mlir_conv_template.py, mlir_conv_sb_template.py, mlir_conv_sbs_template.py, mlir_conv_mt_template.py) and the gemm/bmm templates share large copy-pasted blocks. Lifting the common pieces into mlir_conv_common.py / mlir_common.py would shrink the surface, but the templates also intentionally diverge (single-batch vs multi-tile vs subtile-stationary variants), so the dedup needs careful design. Filing as a note rather than an action item.