Xmake Build Optimization

Xmake Build Optimization & Best Practices

A roadmap for making an xmake build faster. Try them in order — each has a different cost/benefit. Measure before and after with XMAKE_PROFILE=perf:tag (or time xmake).

0. Measure first

Don't guess where the time goes:

XMAKE_PROFILE=perf:tag     xmake        # compile vs link vs other
XMAKE_PROFILE=perf:process xmake        # which subprocesses dominate
XMAKE_PROFILE=perf:call    xmake        # expensive Lua functions
XMAKE_PROFILE=trace        xmake f -c   # trace every subprocess in configure
time xmake --rebuild                    # total wall-clock

Profile before each change — "speedups" often turn out to be in the cache-hit, not the fix.

See xmake-troubleshooting for the full profiler reference.

1. Parallelism

xmake -j N                              # N parallel jobs
xmake -j $(nproc)
xmake -j $(sysctl -n hw.ncpu)           # macOS

Default is already ⌈3/2 × ncpu⌉. Override for:

• Huge machines where xmake's default is too conservative.
• Memory-limited builds where you need to reduce parallelism (big LTO links can OOM).

2. Local compile cache (on by default)

xmake f --ccache=y          # default
xmake f --ccache=n          # disable for debugging

Xmake ships a built-in cache since 2.6.6 — faster than external ccache, cross-platform (MSVC too). See xmake-build-cache.

Verify it's working — look for cache compiling.release in the output.

3. Unity builds

Compile multiple source files as one translation unit. Massive speedup (2–5×) for projects with hundreds of small files.

target("mylib")
    add_rules("c++.unity_build", {batchsize = 8})    -- 8 files per unit
    add_rules("c.unity_build")
    add_files("src/*.cpp")

Pitfalls (usual unity-build ones):

• File-local static symbols clash with same-named symbols in another file in the same unit. Rename or exclude.
• Anonymous namespaces collide similarly.
• Order-dependent macros leak across files in the same unit.

Exclude problem files:

add_files("src/*.cpp")
add_files("src/legacy.cpp", {unity_ignored_files = true})

4. Precompiled headers (PCH)

target("app")
    set_pcxxheader("src/pch.hpp")       -- C++ PCH
    -- or:
    set_pcheader("src/pch.h")           -- C PCH
    add_files("src/*.cpp")

Good fit when many files include the same heavy STL/boost/Qt header. PCH is compiled once and reused.

5. Reduce the work

Before optimizing the compiler, remove files you don't need:

• Scope includes correctly. add_includedirs("include", {public = true}) only for the actual public dir — not src/.
• Split targets. Libraries that rarely change stay cached; split noisy code into its own target so it doesn't invalidate the rest.
• Avoid per-file defines. add_defines at target level is parsed once; target:add inside on_load runs per-target but is still cheaper than per-file.
• Glob, don't enumerate. add_files("src/**/*.cpp") lets xmake hash once; listing 500 files explicitly is slower to parse.
• set_policy("build.warning", false) when CI checks warnings — don't re-emit warning diagnostics in the inner loop.

6. Binary package cache

If you use lots of packages, pull precompiled binaries instead of building from source:

xrepo install --precompiled fmt spdlog boost

Or project-level: packages that have prebuilt binaries are used automatically when available. xmake require --info shows whether the precompiled variant exists.

Cache package installs in CI:

# GitHub Actions
- uses: actions/cache@v4
  with:
    path: |
      ~/.xmake/packages
      ~/.xmake/cache
    key: xmake-${{ runner.os }}-${{ hashFiles('xmake.lua') }}

7. Link-time improvements

Linking often dominates build time for big projects.

LTO (release only)

target("app")
    if is_mode("release") then
        set_policy("build.optimization.lto", true)
    end

LTO makes release builds slower but smaller/faster at runtime. Don't enable for debug builds.

Use lld instead of ld

xmake f --ldflags="-fuse-ld=lld"        # Linux/macOS with clang
xmake f --ldflags="-fuse-ld=mold"       # even faster (Linux)

mold and lld can be 5–10× faster than BFD ld on large links.

Split shared libraries

Huge monolithic static libs link slowly. Split into a few shared or static targets with proper add_deps — xmake will only re-link what changed.

Symbol visibility

Hide symbols by default to shrink link work:

set_symbols("hidden")
-- or flags:
add_cxflags("-fvisibility=hidden", "-fvisibility-inlines-hidden")

8. Debug-build-specific speedups

Release isn't the bottleneck for daily dev — debug is.

if is_mode("debug") then
    set_optimize("none")                    -- or "fast" for balanced
    set_symbols("debug")
    add_cxflags("-g1", {force = true})      -- minimal debug info
end

• -g1 instead of -g cuts debug info size drastically.
• Split DWARF: add_cxflags("-gsplit-dwarf") + add_ldflags("-Wl,--gdb-index").
• -fno-inline off if you don't need to step into inlined functions.

9. Distributed and remote caches

Scale beyond one machine:

• Distributed compile (--distcc) — farm out compile jobs to a cluster. See xmake-distributed-compilation.
• Remote cache (--ccache) — share compiled objects across a team/CI. See xmake-build-cache.

Both compose with the local cache and each other. Typical team setup:

xmake service --connect --distcc --ccache

10. Keep xmake's own overhead low

The description-domain parse time and tool detection can dominate tiny projects.

• Don't put I/O in description domain. See xmake-style. Every parse re-runs top-level code.
• Avoid os.iorun / git at top level. Use on_load or a task.
• xmake f once, not per file. Configure caches to .xmake/; don't re-configure inside tight loops.
• Avoid deep add_moduledirs chains during steady-state builds — only needed while iterating on native modules.

Order to try (cheat sheet)

1. XMAKE_PROFILE=perf:tag xmake — know where time goes first.
2. -jN — parallelism.
3. Verify --ccache=y and look for cache hits.
4. Unity build on the biggest target.
5. PCH for common headers.
6. Binary packages (--precompiled, CI cache for ~/.xmake/packages).
7. lld / mold linker.
8. Debug-build-specific flag trimming.
9. LTO (release only).
10. Distcc + remote cache (team/CI scale).

Common anti-patterns

• Optimizing release when dev is debug. Profile the build you actually run.
• Enabling LTO in debug builds. Makes debug slower for no gain.
• Disabling the cache to "get clean times". Fine for benchmarks, not for dev loops.
• Huge monolithic targets. Changes anywhere → rebuild everything → re-link everything.
• Unity build + anonymous namespaces collision. Check unity-friendliness before enabling.
• Turning everything on blindly. LTO + PCH + unity + -g3 can fight each other. Add one knob at a time.

When to branch out

• Cache details (local / remote) → xmake-build-cache
• Multi-machine job pooling → xmake-distributed-compilation
• Compiling on a remote box → xmake-remote-compilation
• Slowness is actually a hang → xmake-troubleshooting
• xmake.lua parsing is slow → xmake-style

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: xmake-io
• Source: xmake-io/xmake-skills
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.