elf-x86 target successfully running in a freestanding i686 kernel environment

[Shagedoorn1]

elf-x86 target successfully running in a freestanding i686 kernel environment

I'm building my own hobby OS, Phonon. It is a 32 bit x86 OS.

I've been interested in C3 for a while, and noticed that the support matrix in the README lists ELF freestanding x86 as untested.

So I figured: let's just try it.

It works!

The initial test was deliberately simple: a C3 function that calls the kernel's existing printk() and prints:

hello from C3!

Although, given that printk is a printf clone that ultimately outputs to both VGA memory and serial, perhaps calling this a "simple print test" is underselling how much kernel machinery is actually involved.

My build scheme is very simple: turn every source file into an object file, then throw all the objects at the linker. No special project structure required.

It took a little bit of experimentation to get c3c to emit exactly the object files I wanted. Eventually I ended up with:

C3       := c3c
C3FLAGS := \
    --target elf-x86 \
    --use-stdlib=no \
    --emit-stdlib=no \
    -g0

$(BUILD_DIR)/%.o: %.c3
    @New-Item -ItemType Directory -Force -Path "$(dir $@)" | Out-Null
    @Write-Host "C3  $<"
    @& "$(C3)" compile-only "$<" $(C3FLAGS) --obj-out "$(dir $@)" --emit-only "$(basename $(notdir $<))"

The resulting files are normal ELF32 relocatable objects, which I can feed directly to i686-elf-ld alongside object files produced by GCC and GAS. The linker resolves everything without any special treatment.

C <> C3 works too. The C3 test function calls printk(), which is defined in C, and a C function successfully calls the exported C3 function.

The test executes correctly under both QEMU and VirtualBox, and I've inspected the resulting kernel with both IDA and Ghidra. The generated code looks completely sane and healthy.

For example, the C3 function is just a tiny bit of 32-bit x86 code:

lea     eax, "hello from C3!\n"
call    printk
ret

The generated object has the expected ELF32header and normal R_386_* relocations.

So while I obviously haven't tested the entire C3 feature set on this target, basic freestanding ELF32 compilation, C <> C3, linking and execution appear to work quite nicely.

The compiler does still print:
WARNING! This architecture is unsupported.
which is quite funny, considering the resulting code is now happily running inside my kernel. :)

Screenshots