MLIR

Multi-Level IR Compiler Framework

MLIR : Language Server Protocol

This document describes the tools and utilities related to supporting LSP IDE language extensions for the MLIR textual assembly format. An LSP language extension is generally comprised of two components; a language client and a language server. A language client is a piece of code that interacts with the IDE that you are using, such as VSCode. A language server acts as the backend for queries that the client may want to perform, such as “Find Definition”, “Find References”, etc.

MLIR LSP Language Server : mlir-lsp-server 

MLIR provides an implementation of an LSP language server in the form of the mlir-lsp-server tool. This tool interacts with the MLIR C++ API to support rich language queries, such as “Find Definition”.

Supporting custom dialects and passes 

mlir-lsp-server, like many other MLIR based tools, relies on having the appropriate dialects registered to be able to parse in the custom assembly formats used in the textual .mlir files. The mlir-lsp-server found within the main MLIR repository provides support for all of the upstream MLIR dialects and passes. Downstream and out-of-tree users will need to provide a custom mlir-lsp-server executable that registers the entities that they are interested in. The implementation of mlir-lsp-server is provided as a library, making it easy for downstream users to register their dialect/passes and simply call into the main implementation. A simple example is shown below:

#include "mlir/Tools/mlir-lsp-server/MlirLspServerMain.h"

int main(int argc, char **argv) {
  mlir::DialectRegistry registry;
  registerMyDialects(registry);
  registerMyPasses();
  return mlir::failed(mlir::MlirLspServerMain(argc, argv, registry));
}

Features 

This section details a few of the features that the MLIR language server provides. The screenshots are shown in VSCode , but the exact feature set available will depend on your editor client.

Diagnostics 

The language server runs actively runs verification on the IR as you type, showing any generate diagnostics in-place.

IMG

Cross-references 

Cross references allow for navigating the use/def chains of SSA values (i.e. operation results and block arguments), Symbols , and Blocks.

Find definition 

Jump to the definition of the IR entity under the cursor. A few examples are shown below:

  • SSA Values

SSA

  • Symbol References

Symbols

The definition of an operation will also take into account the source location attached, allowing for navigating into the source file that generated the operation.

External Locations

Find references 

Show all references of the IR entity under the cursor.

IMG

Hover 

Hover over an IR entity to see more information about it. The exact information displayed is dependent on the type of IR entity under the cursor. For example, hovering over an Operation may show its generic format.

IMG

The language server will also inform the editor about the structure of symbol tables within the IR. This allows for jumping directly to the definition of a symbol, such as a func, within the file.

IMG

Design 

The design of mlir-lsp-server is largely comprised of three different components:

  • Communication and Transport (via JSON-RPC)
  • Language Server Protocol
  • MLIR Language Server

Index Map Example

Communication and Transport 

mlir-lsp-server communicates with the language client via JSON-RPC over stdin/stdout. In the code, this is the JSONTransport class. This class knows nothing about the Language Server Protocol, it only knows that JSON-RPC messages are coming in and JSON-RPC messages are going out. The handling of incoming and outgoing LSP messages is left to the MessageHandler class. This class routes incoming messages to handlers in the Language Server Protocol layer for interpretation, and packages outgoing messages for transport. This class also has limited knowledge of the LSP, and only has information about the three main classes of messages: notifications, calls, and replies.

Language Server Protocol 

LSPServer handles the interpretation of the finer LSP details. This class registers handlers for LSP messages and then forwards to the MLIR Language Server for processing. The intent of this component is to hold all of the necessary glue when communicating from the MLIR world to the LSP world. In most cases, the LSP message handlers simply forward to the MLIR Language Server. In some cases however, the impedance mismatch between the two requires more complicated glue code.

MLIR Language Server 

MLIRServer provides the internal MLIR-based implementation of all of LSP queries. This is the class that directly interacts with the MLIR C++ API, including parsing .mlir text files, running passes, etc.

Editor Plugins 

LSP Language plugins are available for many popular editors, and in principle mlir-lsp-server should work with any of them, though feature set and interface may vary. Below are a set of plugins that are known to work:

Visual Studio Code 

Provides MLIR language IDE features for VS code:

  • Syntax highlighting for .mlir files and mlir markdown blocks
  • go-to-definition and cross references
  • Detailed information when hovering over IR entities
  • Outline and navigation of symbols and symbol tables
  • Live parser and verifier diagnostics

Setup 

This extension requires the mlir-lsp-server language server . If not found in your path, you must specify the path of the server in the settings of this extension.

Contributing 

This extension is actively developed within the LLVM monorepo , at mlir/utils/vscode. As such, contributions should follow the normal LLVM guidelines , with code reviews sent to phabricator .

When developing or deploying this extension within the LLVM monorepo, a few extra setup steps are required:

  • Copy mlir/utils/textmate/mlir.json to the extension directory and rename to grammar.json.
  • Copy https://mlir.llvm.org//LogoAssets/logo/PNG/full_color/mlir-identity-03.png to the extension directory and rename to icon.png.

Please follow the existing code style when contributing to the extension, we recommend to run npm run format before sending a patch.