Tilelang Cpp Style

TileLang C++ Style

Use this skill when C++ style decisions matter in TileLang. For broad policy changes, review the full guide in docs/developer_guide/cpp_style.md.

Baseline

TileLang C++ should be readable to TVM contributors. Follow TVM/Google C++ style unless a TileLang runtime template, generated kernel, backend shim, or external API adapter has a concrete reason to match another convention.

Do not apply broad style churn to 3rdparty/ or generated/template code unless the task explicitly targets those files.

Before writing new C++:

• Identify whether the file is TileLang-owned compiler/runtime code, a src/tl_templates runtime template, a CUDA/HIP/Metal shim, or an external API adapter. Owned compiler/runtime code follows this skill. Templates and shims may follow the generated or external interface they model.
• Prefer the nearest local pattern in the same subsystem (op, layout, transform, backend, cuda/op, rocm/op, etc.) over inventing a new abstraction.
• Keep style-only edits separate from behavior changes unless the style cleanup is needed to make the behavioral change reviewable.

Naming Checklist

• File names: lower_snake_case.
• Namespaces: lower_snake_case.
• Types, classes, structs, ObjectRefs, ObjectNodes: PascalCase.
• Object nodes: PascalCaseNode; ObjectRefs: PascalCase.
• Public functions and methods: PascalCase.
• Private/protected methods: PascalCase_.
• Inherited TVM visitor, mutator, and codegen hooks keep upstream names such as VisitExpr_, VisitStmt_, VisitStmtDefault_, and InitFuncState_.
• Boolean helpers: Is/Has/Can + PascalCase.
• Parameters and local variables: lower_snake_case.
• Private/protected data members: lower_snake_case_.
• Constants and enum values: kPascalCase.
• Macros: UPPER_SNAKE_CASE.

Prefer new C++ APIs like:

Layout MakeLinearLayout(Array<PrimExpr> shape);
bool IsSharedBuffer(const Buffer& buffer);
Stmt Lower(const LowerArgs& args, arith::Analyzer* analyzer) const;

Avoid adding new lowerCamelCase helpers or ambiguous context names like const LowerArgs& T.

API Boundaries And Ownership

Prefer concrete type declarations over auto when the type is short or the value crosses an API boundary. auto is fine for obvious pattern checks, iterators, lambdas, Downcast<T>(), and node.as<T>() results where spelling the type would obscure the logic.

Pass TVM object handles and non-trivial inputs by const& unless the callee will store the value. If a constructor or helper stores the value, take it by value and std::move into the field.

Mark query methods and helpers const whenever possible. For visitor classes, keep mutable state explicit in private fields and prefer short static entry points such as Collect, Rewrite, or Analyze that hide the visitor object.

Use TVM Array, Map, Optional, and ObjectRefs at FFI/API boundaries. Internal analysis code may use std::vector, std::unordered_map, and std::unordered_set when that is simpler, but convert back at the boundary.

Use TVM_DLL only for functions and constructors that need cross-translation unit or extension visibility, following nearby exported APIs. Keep file-local helpers in anonymous namespaces or private methods.

TVM Object Conventions

Treat Stmt, PrimExpr, Buffer, Var, SBlock, Layout, and related types as TVM ObjectRef handles. Use handles across function boundaries and for stored state.

Keep raw *Node pointers local to visitor callbacks, pattern checks, and CopyOnWrite() mutation sites. Convert callback nodes with GetRef<T>(op) if the value must be passed elsewhere or retained.

Use .same_as(other) for identity comparisons between handles. Use TVM structural equality utilities for structural comparisons. Avoid .get() == other.get() outside narrow adapter code.

For identity maps/sets keyed by handles, use TVM pointer hash/equality helpers:

using BufferSet = std::unordered_set<Buffer, ObjectPtrHash, ObjectPtrEqual>;
using VarMap = std::unordered_map<Var, PrimExpr, ObjectPtrHash, ObjectPtrEqual>;

Use Optional<T> for nullable ObjectRef results. Return an undefined Optional<T> or ObjectRef for "not found" cases instead of retaining raw node pointers or relying on unrelated sentinel state.

Construct ObjectRefs with make_object<Node>(), populate reflected fields, then assign data_ = std::move(node). For Clone(), a shallow make_object<Node>(*this) is fine only when embedded mutable ObjectRefs do not need independent copies; clone nested mutable operators explicitly when needed.

When mutating existing IR/ObjectRefs, call CopyOnWrite() on a handle. Do not mutate raw callback nodes directly.

Symbolic Arithmetic And Analyzer Use

Do not force a PrimExpr to a constant integer if symbolic arithmetic can carry the logic. Keeping expressions symbolic preserves dynamic shape support.

When a constant is truly required, extract it as int64_t (as_const_int, IntImmNode::value, or an equivalent helper), check the failure path, and rebuild constants with the intended dtype (Integer, IntImm(expr->dtype, v), or make_const(dtype, v)). Avoid narrowing to int unless the target API requires it and the range is checked.

Use a populated arith::Analyzer for Simplify, CanProve, CanProveEqual, bounds, and modular reasoning. Bind loop/shape variable ranges before asking the analyzer to prove facts. If a function accepts an optional analyzer pointer, create a local fallback only for simple self-contained reasoning.

Be careful mixing int32 and int64 in analyzer-sensitive expressions. In layout/swizzle code, preserve the native dtype of the expression spine when casts would prevent simplification.

When constructing constants, prefer dtype-aware helpers:

PrimExpr zero = make_const(buffer->dtype, 0);
PrimExpr 1);
PrimExpr same_dtype = IntImm(expr->dtype, value);

ObjectNode Fields

For TVM-style ObjectNode classes, public reflected fields should use lower_snake_case without a trailing underscore. Use trailing underscores for private/protected implementation state.

Before renaming reflected fields, check whether the field name is FFI-visible or used by Python/debugging surfaces. Do not break compatibility without an explicit migration plan.

For a new FFI-visible Object:

• Node classes use TVM_FFI_DECLARE_OBJECT_INFO or TVM_FFI_DECLARE_OBJECT_INFO_FINAL.
• ObjectRef classes use TVM_FFI_DEFINE_OBJECT_REF_METHODS_NULLABLE unless the handle is intentionally non-nullable.
• RegisterReflection() defines public fields with stable string names.
• A TVM_FFI_STATIC_INIT_BLOCK() registers reflection and global functions.

Treat registered field names and global names as API. Python-facing global names usually follow existing strings such as tl.transform.Simplify, tl.make_linear_layout, or target-specific tl.cuda.transform.*.

Tile Operators And Backend Implementations

New tile operators should follow the TileOperatorNode pattern:

• Parse ffi::Array<PrimExpr> arguments and Map<String, ObjectRef> annotations in the ObjectRef constructor.
• Fill access regions with SetAccessRegions so analysis sees reads/writes.
• Implement Lower, InferLayout, and Clone.
• Register the TIR op with TIR_REGISTER_TL_TILE_OP or a matching TVM_REGISTER_OP variant. Set call effects and printer names consistently.

Target-specific lowering belongs in target directories such as src/cuda/op, src/rocm/op, src/cpu/op, or src/metal/op. Use small target predicates and registration structs (RegisterCopyImpl, RegisterAtomic...Impl, etc.) instead of hardcoding target branches in common code. Put shared fallback logic in src/backend/common or the common op file.

Use target helper APIs (TargetIsCuda, TargetHasAsyncCopy, TargetCudaGetWarpSize, etc.) instead of string-matching targets at call sites.

For annotations, prefer existing attr::k... constants when available. When a string key is required for compatibility, parse it through a small helper and validate type/range close to the parse site.

Passes And Visitors

Transform passes usually use local visitor/mutator classes plus a small public factory:

tvm::transform::Pass MyPass() {
  auto pass_func = [](PrimFunc func, const IRModule& mod,
                      PassContext ctx) -> PrimFunc {
    arith::Analyzer analyzer;
    return MyRewriter::Rewrite(std::move(func), &analyzer);
  };
  return CreatePrimFuncPass(pass_func, 0, "tl.MyPass", {});
}

Register pass factories in TVM_FFI_STATIC_INIT_BLOCK() under tl.transform.* or target-specific namespaces like tl.cuda.transform.*.

Visitor rules:

• Use StmtExprVisitor, StmtExprMutator, StmtMutator, or arith::IRMutatorWithAnalyzer according to whether the pass needs expression traversal, mutation, or analyzer context.
• Provide static entry points such as Collect, Rewrite, Substitute, or Apply; keep mutable traversal state private.
• In callback overrides, call the base visitor/mutator unless intentionally stopping traversal. Convert callback nodes with GetRef<T>(op) when passing or storing the handle.
• Preserve annotations, spans, and read/write metadata when rebuilding IR unless the pass intentionally changes them.

Headers And Includes

• TileLang currently has no separate installed public C++ header tree. Avoid using namespace in any future installed public headers and in headers that behave like shared cross-module interfaces inside the repository.
• Prefer explicit qualifications in shared headers, or narrow aliases when they are part of the API or remove real repetition.
• In .cc files and narrowly included internal src/ helper headers, file-local namespace imports are acceptable when they improve readability. tirx imports are common for IR-heavy code; prefer explicit ffi:: names in core-style code unless FFI helpers are pervasive in the file.
• Include the headers that define the types used by the file.
• Keep implementation-only helpers in .cc files or anonymous namespaces.
• Group includes as paired header, standard library, TVM/third-party, then local TileLang headers.
• Use header guards shaped like TVM_TL_<PATH>_<FILE>_H_ where possible, and close namespaces and guards with comments.

Error Handling And Diagnostics

Use ICHECK/ICHECK_EQ for internal compiler invariants. Use LOG(FATAL) for unsupported or unreachable branches where a streamed message is clearer than a boolean check. Avoid new bare ICHECK(0) without context.

Use CHECK(..., ErrorKind) only for user-facing FFI validation paths where a typed TVM FFI error is intended. Runtime ABI validation should follow existing generated AssertStmt or runtime/error_helpers patterns.

Make messages actionable: include the op, target, buffer name, scope, dtype, shape/range, or annotation key that caused the failure. Prefer messages that explain the violated constraint over "not supported".

Comments And Documentation

Use Doxygen comments for public APIs, ObjectRefs/ObjectNodes, pass factories, and non-obvious fields. For local code, comment invariants and target-specific constraints, not line-by-line restatements.

Good comments in TileLang often explain why a fallback, annotation, dtype cast, or analyzer binding is required. Keep comments short unless they describe a subtle lowering invariant.

Clang-Format And Macro Boundaries

Use clang-format off/on narrowly around dense generated tables, inline asm, or registration blocks only when the formatter makes the code materially worse. Keep the disabled region as small as possible.

Write macros so clang-format can still parse surrounding C++: macro functions should look like calls at use sites, and declaration macros should leave the declaration syntactically complete, including a semicolon when appropriate.

Migration Policy

Style convergence should be incremental:

• New C++ code should follow the style guide.
• Touched code should avoid adding new inconsistencies.
• Mechanical renames should be small and isolated.
• Do not mix broad formatting changes with behavior changes.
• Keep .clang-format changes in a focused PR.

Validate style/documentation edits with:

python3 -m pre_commit run --files <changed-file>...