LLVM Explorer is a browser-only compiler explorer for C, C++, LLVM IR, and MLIR in JupyterLab, JupyterLite, and a standalone app. No kernel or remote compiler is needed.

The web app opens the standalone explorer. Select Try in Jupyter to use the explorer alongside C23 and C++23 notebooks.

Select About in the explorer for project credits and the guide below. The same guide is available as LLVM Explorer guide.md in JupyterLite.

LLVM Explorer compiles and inspects C, C++, LLVM IR, and MLIR in your browser, with a shared workbench for standalone use, JupyterLab, and JupyterLite. No kernel or remote compiler is needed.

Anutosh Bhat’s work on WasmBolt brought the LLVM compiler tools into the browser using WebAssembly. LLVM Explorer’s compiler runtime builds on that work, released under the MIT License. We reuse WasmBolt’s compiler module, LLVM lifecycle adaptations, and compilation pipeline, with an adapted build recipe. LLVM Explorer adds the shared React/Lumino workbench, compiler worker service, and JupyterLab, JupyterLite, and standalone integrations. WasmBolt’s original copyright and license notice are included in our distributions.

The compiler itself is provided by LLVM/Clang and built for WebAssembly using Emscripten.

Our notebook environment uses xeus-cpp, CppInterOp, and jupyterlite-xeus, with browser packages from emscripten-forge. Their work makes interactive C and C++ notebooks possible without a server. The Lite build preserves package license notices alongside the kernel assets.

Install and run locally:

pip install llvm-explorer
llvm-explorer

The command opens the explorer in your browser. Try in Jupyter opens the bundled JupyterLite environment with C23 and C++23 notebooks in the same tab. Everything is served from your computer; compilation and notebook execution stay in the browser. No JupyterLab installation, Node, or compiler build is required.

Keep the terminal open while using LLVM Explorer; press Ctrl+C to stop serving it. Use llvm-explorer --no-browser to print the address, or llvm-explorer --port 8001 to choose another port. Your browser saves sessions separately for each address.

If you have JupyterLab 4.6 or later, the same package also provides its extension. Restart JupyterLab and select Open LLVM Explorer in the launcher or command palette. The npm package includes the compiler and shared workbench for applications that supply their own Lumino host.

Edit a function, choose a language, target, and optimization level, then select Compile or press Ctrl/Cmd+Enter. Each compilation generates all outputs for the selected language and target. Switching output tabs does not recompile. You can keep editing during compilation. Cancel stops the compiler; the next compile reloads it.

• C/C++: AST, LLVM IR before passes, optimized IR, analysis, function graphs, assembly, and object.
• LLVM IR: validated input IR, optimized IR, analysis, function graphs, assembly, and object.
• MLIR: transformed MLIR and operation graph.
• WebAssembly target with C/C++ or LLVM IR: also a linked Wasm module and its metadata.

The workbench starts with source beside output. Only outputs for the selected language and target are shown. The tools along the bottom open when selected; select the active tool again to fold it away, with the mouse or Enter/Space. Compilation errors open Diagnostics automatically. Narrow windows stack the panes; widening the window restores the saved arrangement.

Every compilation includes diagnostics, recorded commands, raw streams, timings, and generated files. A failed stage preserves successful independent outputs. Expand Build details in Diagnostics for stage status and timings. Select a diagnostic to jump to its source location. Text views provide line numbers, search, copy, and download. Find or Ctrl/Cmd+F searches the focused source or text output; Ctrl/Cmd+Enter compiles from either editor. In the source editor, Escape, then Tab moves focus out without inserting indentation. Graphs have function selection, zoom, fit, and DOT/SVG downloads. The Wasm inspector lists size, imports, exports, and function signatures without executing the module.

Assembly hides compiler metadata by default, retaining code, data, and their directives. Uncheck Hide metadata to see it all. Copy uses the displayed text; Download always saves the original file.

The status strip shows download progress in MB, then preparation and compilation activity. Diagnostics lists the individual compiler downloads and any loading failure. Progress counts decoded asset bytes against their packaged sizes; compressed network transfers may be smaller. Asset verification requires HTTPS or a local server on localhost.

The defaults are C++23, WebAssembly, and O2. C23 and O0–O3 are available. The packaged LLVM runtime reports WebAssembly, x86-64, and AArch64 backends. Native targets produce inspection artifacts with Clang built-in headers only; the C/C++ system headers are for WebAssembly.

Use Pipelines to set LLVM optimization, analysis, and MLIR passes. An empty LLVM pipeline follows the optimization level, for example default<O2>. Frontend semantics and backend code generation also use that level. The first IR view has LLVM optimization passes disabled. Analysis defaults to dominator trees and loops; MLIR defaults to builtin.module(canonicalize,cse). LLVM inputs with incompatible target triples or layouts report an error instead of being silently retargeted. Changing language switches an untouched example to that language; edited source is preserved. Reset example, beside the source filename, replaces it with that language's example and can be undone. Pipelines shows only the fields relevant to the selected language.

Run reuses a current Wasm module, or builds it first if source or options have changed. Compilation and inspection never execute generated code. The separate runner worker initializes on the first Run. Choose an exported function and enter its scalar arguments in the Run pane. Supported signatures are i32 or f64 returns with zero, one, or two matching arguments, and void(). Unsupported signatures remain visible. Pointer and aggregate values are not supported. A compatible selection survives a rebuild; otherwise the runner prefers supported main, then a sole callable export. Ambiguous exports require selection. main(i32, i32) receives argc = 0 and a null argv.

The pane shows the return value, stdout, stderr, status, and errors. Repeated calls retain module state. Reset execution, Stop, timeout, traps, and module replacement discard the runner without losing compilation artifacts. The default execution timeout is 10 seconds, configurable under Execution settings in the Run pane; it starts after initialization. NaN is a valid return value. Execution requires WebAssembly; native targets remain available for inspection.

The Terminal runs clang, clang++, opt, llc, wasm-ld, mlir-opt, and dot through the compiler worker. It accepts single and double quotes, backslash escapes, and > / 2> redirection. Quote LLVM pipeline arguments containing angle brackets. Each command runs one tool; shell pipelines and input redirection are not supported. For example:

opt "-passes=print<domtree>" -disable-output optimized.ll 2> tree.txt
dot -Tsvg .square.dot -o square.svg

The current directory is /workspace. Compile fills it with source and generated files. Manual commands can use explicit libraries and compiler flags; their output updates the workspace while completed build artifacts remain unchanged. A new Compile replaces the workspace. Worker recovery retains its latest snapshot. Files provides previews and downloads; select a manually linked .wasm file and choose Use module to run it. Keep files under /workspace to include them in snapshots and the runner's dependencies.

In standalone, Share copies a link with your source and compiler options. New links keep these inputs after #, outside the request sent to the server. Opening a link saves its inputs locally without compiling or running, then removes them from the address so reloading keeps subsequent edits. Binaries, logs, layout, and execution state are excluded. Sharing is unavailable in JupyterLab and JupyterLite, where sessions belong to the host workspace.

Source, pipeline options, output selection, pane sizes, and execution timeout are saved automatically. Jupyter also keeps a browser copy of recent edits for each workspace. Reopening restores the session without compiling. Output is marked out of date when source or options change. Invalid saved state opens a default session with a warning. Compiled artifacts, command files, and running processes are not saved.

Our JupyterLite site includes xeus-cpp 0.10.0, with C23 and C++23 kernels. Open C++ examples.ipynb or C examples.ipynb and choose Run → Run All Cells. The examples use packaged standard library headers, define functions, and reuse state across cells.

The notebook interpreter runs in its own browser worker. It is independent of LLVM Explorer's compiler: code, options, and results are not synchronized between them. Its Clang version also differs from the explorer's LLVM runtime. The first kernel start downloads the interpreter and its libraries. Browser memory limits apply; native processes, native platform APIs, and arbitrary native libraries are unavailable.

These kernels run in the bundled JupyterLite environment. They are not native JupyterLab kernels.

See CONTRIBUTING.md for the complete Pixi setup and build sequence. The compiler is a separate heavyweight build; frontend builds require its generated assets and verify their hashes.

For an already built checkout:

pixi run --as-is jupyter lab
pixi run --as-is jlpm serve
pixi run --as-is jlpm serve:standalone

In JupyterLab or JupyterLite, select Open LLVM Explorer in the launcher or command palette. The Lite testbed is served on port 8080; the standalone preview prints its local address. These commands run in separate terminals.

Built standalone and Lite directories can be served below a URL prefix. Keep each site's compiler assets at their generated relative location. Serve .js as JavaScript, .wasm as application/wasm, and .data as application/octet-stream. Compilation works offline after initialization. MLIR and execution each need their runtime assets loaded before offline use; offline page reload is a separate feature.

• Pure model, request construction, and diagnostic parsing.
• One instance-owned store outside React.
• Lumino commands orchestrate semantic changes and compiler effects.
• Functional React views, with explicit store and CodeMirror bridges.
• A shared Lumino workbench owns workers and view lifecycles, with tab panels inside resizable split panels. Jupyter owns docking of the workbench itself.
• Thin Jupyter and standalone adapters supply shell, persistence, and share URLs.

The shared package entry exports these contracts. Only src/jupyter/ imports JupyterLab packages; the plugin uses llvm-explorer:plugin.

The compiler API returns Result.artifacts, Result.stages, and Result.files. Artifacts carry their build ID, kind, path, and text or binary content. Stages record status, commands, raw streams, diagnostics, and duration. Consumers should select artifacts by kind and account for partial failures. ICompiler.compile accepts stage progress; ICompiler.command returns a stage and replacement workspace snapshot. Options includes the three pipeline fields and the four input languages. IRunner, RunRequest, RunResult, and inspectWasm are exported separately. Treat all returned binary buffers as immutable; transport never detaches buffers already owned by application state.

Command names describe actions, such as CommandIDs.setSource, CommandIDs.compile, and CommandIDs.selectOutput. Their IDs use the same words in kebab case, such as llvm-explorer:set-source. Jupyter registers CommandIDs.open; registerCommands takes an ICommandContext for the shared commands.

WasmFunction.signature is display text, such as i32(i32). WasmFunction.signatureCode is the numeric runner ABI code, or null for an unsupported export. Pass that code as RunRequest.signatureCode.

The runtime is reproduced from WasmBolt with LLVM 23.1.0 and Emscripten 4.0.9. runtime/README.md records its origin, pins, licenses, and build details. Generated compiler/manifest.json records asset sizes and SHA-256 hashes. Browser test attachments record timings and Wasm memory observations.

Each initialized compiler or runner is large and reserves 256 MiB of initial Wasm memory, with memory growth enabled and a 32 MiB stack. Running a program alongside the compiler therefore requires two instances. Browser memory limits still apply. Cancellation releases the worker; a later compile must initialize another. Initialization and runtime failures offer a retry path. Ordinary compiler errors retain diagnostics and raw output.

The optional MLIR driver downloads only when MLIR is used, with the same integrity checks, progress, cancellation, and retry behavior as the core. MLIR exploration uses explicit passes; automatic lowering to an executable, automatic compilation, and notebook synchronization remain outside the explorer.

LLVM Explorer is BSD-3-Clause licensed. The compiler incorporates WasmBolt and other separately licensed software. Required notices are included in runtime/licenses/ and copied into each distribution.