[REFACTOR][IR] Unify StructInfo and Type

[tqchen]

Summary

• unify Relax's former StructInfo surface into the Type vocabulary and Expr.ty storage path
• remove leftover DependentTypeNode and legacy OpNode::op_type storage
• keep base Type nullable while concrete Relax/DTensor type refs are non-nullable
• clean stale StructInfo/TensorStructInfo/sinfo vocabulary in Python/docs and distributed-op macros
• address Gemini follow-ups for parser annotations, BlockBuilder docstring, and Adreno TensorType cast audit

Validation

• cmake --build build --parallel 4
• PYTHONPATH=$(pwd)/python:$(pwd)/.local/python TVM_LIBRARY_PATH=$(pwd)/build python3 -m pytest -q tests/python/ir/test_ir_type.py tests/python/relax/test_type.py tests/python/relax/test_analysis_type_analysis.py tests/python/relax/test_tvmscript_parser.py tests/python/relax/test_tvmscript_printer_relax.py tests/python/relax/test_expr.py tests/python/relax/distributed/test_distributed_dtensor_type.py tests/python/relax/distributed/test_distributed_tvmscript_parser.py tests/python/relax/distributed/test_distributed_tvmscript_printer.py -> 717 passed, 1 xfailed
• rg stale vocabulary audit over include/src/python/tests/docs -> no matches
• git diff --check
• pre-commit run --files $(git diff --name-only --diff-filter=ACMRT)

Note: this build directory emits ninja: warning: premature end of file; recovering; the final build still completed successfully after broad rebuild.

[gemini-code-assist]

Code Review

This pull request refactors the Relax dialect in TVM by replacing the StructInfo concept with a unified, richer dependent Type system across C++ source files, headers, Python bindings, and documentation. The review feedback highlights several critical and medium-severity issues, including potential null pointer dereferences in dependent_type.cc, type_functor.cc, type.h, and distributed/type.cc due to missing definition checks. Additionally, there are compilation risks identified in dependent_type.cc regarding the use of the potentially undefined TVM_FFI_ICHECK_GE macro, which should be replaced with the standard TVM_FFI_ICHECK macro.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Critical Bug: Potential null pointer dereference. MatchType<ShapeType>(shape) is called on shape before checking if shape is defined on line 121. If shape is null (undefined), MatchType will dereference it and cause a segmentation fault.

Please check shape.defined() first before calling MatchType.

Suggested change

	ffi::Optional<ShapeType> shape_ty = MatchType<ShapeType>(shape);
	TVM_FFI_ICHECK(shape_ty) << "We expect shape to contain pre-set shape type";
	TVM_FFI_ICHECK(shape.defined()) << "Must provide a shape in this constructor";
	TVM_FFI_ICHECK(shape.defined()) << "Must provide a shape in this constructor";
	ffi::Optional<ShapeType> shape_ty = MatchType<ShapeType>(shape);
	TVM_FFI_ICHECK(shape_ty) << "We expect shape to contain pre-set shape type";

[gemini-code-assist]

Critical Bug: Potential empty optional dereference. If op->params is not defined (e.g., for an opaque function type), params remains undefined (std::nullopt). If the return type ret changes, the code enters the else branch and calls params.value(), which will crash.

We should check if params is defined, and if not, construct an opaque function type using FuncType::OpaqueFunc.

  } else {
    TVM_FFI_ICHECK(ret.defined()) << "FuncType must contain ret";
    if (params.defined()) {
      return FuncType(params.value(), ret, op->purity, op->span);
    } else {
      return FuncType::OpaqueFunc(ret, op->purity, op->span);
    }
  }

[gemini-code-assist]

Defensive Programming: Potential null pointer dereference. If this->shape.value()->ty is not defined (null), Downcast<ShapeType> will fail or crash. We should defensively check if ty is defined and is of type ShapeTypeNode before accessing its values.

    if (!shape.defined() || !shape.value()->ty.defined()) return {};
    if (const auto* shape_ty = shape.value()->ty.as<ShapeTypeNode>()) {
      return shape_ty->values;
    }
    return {};

[gemini-code-assist]

Compilation Risk: TVM_FFI_ICHECK_GE might not be defined in the TVM FFI headers. It is safer and more consistent to use the standard TVM_FFI_ICHECK macro with a comparison operator.

Suggested change

	TVM_FFI_ICHECK_GE(ndim, -1) << "ndim of ShapeType must be >= -1, but got " << ndim;
	TVM_FFI_ICHECK(ndim >= -1) << "ndim of ShapeType must be >= -1, but got " << ndim;

[gemini-code-assist]

Compilation Risk: TVM_FFI_ICHECK_GE might not be defined in the TVM FFI headers. It is safer and more consistent to use the standard TVM_FFI_ICHECK macro with a comparison operator.

Suggested change

	TVM_FFI_ICHECK_GE(ndim, -1) << "ndim of TensorType must be >= -1, but got " << ndim;
	TVM_FFI_ICHECK(ndim >= -1) << "ndim of TensorType must be >= -1, but got " << ndim;

[gemini-code-assist]

Defensive Programming: Potential null pointer dereference. If device_mesh or placement is null (not defined), dereferencing them will cause a segmentation fault. We should defensively check if they are defined first.

  TVM_FFI_CHECK(device_mesh.defined(), ValueError) << "device_mesh must be defined";
  TVM_FFI_CHECK(placement.defined(), ValueError) << "placement must be defined";
  TVM_FFI_CHECK_EQ(device_mesh->shape.size(), placement->dim_specs.size(), ValueError)
      << "The device mesh and placement must have the same dimension size";