JitDump Profiling with Wasmtime

Date: 2026-03-21

Overview

Wasmtime supports JitDump profiling (--profile jitdump), which integrates with Linux perf to provide instruction-level profiling of JIT-compiled WebAssembly code. This document describes the mechanism, workflow, and characteristics.

Mechanism

When --profile jitdump is enabled, wasmtime writes a binary jit-<pid>.dump file containing JIT-compiled code regions with their memory addresses, sizes, and symbol names. This file follows the JitDump format recognized by Linux perf.

The dump file includes actual machine code bytes for every JIT-compiled function, enabling perf annotate to disassemble functions and attribute sample counts to individual instructions.

Workflow

1. Record

perf record -k mono wado run --profile jitdump prog.wado

Or with cargo during development:

perf record -k mono cargo run --release --bin wado -- run --profile jitdump prog.wado

2. Inject JIT symbols

perf inject --jit -i perf.data -o perf.jit.data

This reads the jit-<pid>.dump file and creates a new perf data file with JIT symbol information embedded.

3. Report (function-level)

perf report -i perf.jit.data

Shows which functions consume the most samples. Look for wasm[1]::function[N]::name entries — these are user Wasm functions compiled from Wado source.

4. Annotate (instruction-level)

perf annotate -i perf.jit.data -s <function_name>

For example:

perf annotate -i perf.jit.data -s run
perf annotate -i perf.jit.data -s inflate_raw_ex

This disassembles the function and shows per-instruction sample percentages. Use this to identify hot loops within inlined code.

Symbol Naming

Symbols include full monomorphization detail:

wasm[1]::function[78]::Status^Deserialize::deserialize<JsonDeserializer>
wasm[1]::function[48]::deflate_with_level

Each function has both a long form (wasm[1]::function[N]::name) and a short alias (name).

Characteristics

• Runtime overhead: ~1% — measurements are not significantly distorted
• Output size: 550–615 KB per dump file (contains actual machine code bytes)
• Platform: Linux only (requires perf)
• Granularity: Instruction-level when combined with perf annotate
• CM-async compatibility: Works correctly

When to Use

JitDump is most valuable when function-level profiling is insufficient. Wado's compiler aggressively inlines functions, so hot functions are often large inlined blobs. Function-level attribution would only tell you "run is hot" — not which inlined callee within run is the bottleneck. JitDump + perf annotate shows instruction-level sample attribution within the inlined function body.

Cleanup

JitDump creates jit-<pid>.dump files in the current working directory. Remove after analysis:

rm -f jit-*.dump perf.data perf.jit.data

Comparison with Other Wasmtime Profilers

See research-wasmtime-profiler-characteristics.md for a detailed comparison of all three wasmtime profiling modes (guest, jitdump, perfmap).