Targets & cross-compilation
Marrow generates code through QBE, which supports three architectures: amd64 (x86_64), arm64 (aarch64) and rv64 (riscv64). Marrow adds operating-system layers on top, so a target is an <arch>-<os> pair:
| Target | --nostdlib |
Status |
|---|---|---|
x86_64-linux |
yes | The reference platform; every test runs here. |
aarch64-linux |
yes | Assembled and linked; run under qemu-user in CI. |
riscv64-linux |
yes | Assembled and linked; run under qemu-user in CI. |
aarch64-macos |
yes | Assembled in CI, with a native macOS job configured. Experimental: never run by the author. |
x86_64-macos |
yes | As above. |
x86_64-windows |
yes | Assembled and linked into a PE file against kernel32, advapi32 and ws2_32. Experimental: never run by the author. |
32-bit ARM is not supported because QBE has no 32-bit backend.
--nostdlib: programs without libc
Section titled “--nostdlib: programs without libc”With --nostdlib the program talks to the operating system itself. The compiler appends a small runtime to the generated assembly:
__marrow_syscall(n, a, b, c, d, e, f), the one primitive everything is built on. On Linux it is the raw system call; on macOS it is the BSD call (errors are turned into-errno); on Windowsnindexes a table of Win32 functions._start, on Linux, so the executable needs neither crt0 nor libc.__marrow_ctx_switch, the context switch used by tasks.
The standard library is written against that primitive and nothing else (see the platform layer), so the whole of it, including networking, TLS and HTTPS, works without libc.
--target: building for another machine
Section titled “--target: building for another machine”marrow --nostdlib --target aarch64-linux app.mwFor a target other than the host, marrow stops after producing app.s (QBE assembly plus the runtime). You assemble and link it with the toolchain of your choice. Examples with LLVM:
# Linux on 64-bit ARM, freestandingclang --target=aarch64-linux-gnu -nostdlib -static -fuse-ld=lld app.s -o appqemu-aarch64 ./app
# Linux on RISC-Vmarrow --nostdlib --target riscv64-linux app.mwclang --target=riscv64-linux-gnu -nostdlib -static -fuse-ld=lld app.s -o app
# macOS (build the object on any machine, link on a Mac)marrow --target aarch64-macos app.mwclang --target=aarch64-apple-macos -c app.s -o app.o
# Windowsmarrow --target x86_64-windows app.mwclang --target=x86_64-pc-windows-gnu -c app.s -o app.o# link with kernel32, advapi32 and ws2_32 (-lws2_32 with MinGW)How the standard library picks a platform
Section titled “How the standard library picks a platform”The os.mw module contains @import("platform/{target}"). The compiler replaces {target} with the target’s name (linux_x86_64, macos_aarch64, windows_x86_64…), so a different file of std/platform/ is compiled in for each target. See the platform layer for the contract those files implement.
What is and isn’t verified
Section titled “What is and isn’t verified”Linux x86_64 is exercised by the full test suite on every change. The CI configuration also runs the end-to-end programs natively on Linux aarch64, macOS and Windows runners, and under qemu-user for aarch64 and riscv64, and assembles every target. The macOS and Windows layers were written from the platform documentation and have only been assembled and linked, not executed, so treat them as experimental until those CI jobs have run green.