It is too risky to rely on the compiler to not insert any operations using the
stack.
Having a stack-setting call in Rust using the cortex-a crate as the first action
in a Rust-only _start() function does not work if you're subsequently using the
stack, because the compiler often inserts the operations to make room on the
stack to prepare a function call BEFORE the call to set the stack, which crashes
the boot process.
Hence, keep on using a small piece of assembly boot code throughout.
- Don't wildcard-import from arch modules. Make it explicit.
- Put translation table code into its own module.
- Put boot code in boot.rs instead of cpu.rs
- Other minor changes, most memory subsystem.
The GCC versions of `objdump` and `nm` seem to have better out-of-the-box
support (for AArch64). Demangling works better, and instructions in objdump are
presented as 4 bytes instead of 4 * 1 bytes, which helps a lot.
Hence, switch to the GCC versions for now until LLVM has caught up.
Detect if we run on Linux. If so, dockerize Makefile targets that need access to
USB devices. If we run on anything else, attempt to execute natively.
This will enable non-Linux Unixes like macOS to run the chainboot target, given
they install the needed Ruby dependencies.
`rustup target add aarch64-unknown-none-softfloat` is now a thing.
It is no longer required to build rust-std on-demand with `xbuild`.
This reduces compile-times significantly.
Andre Richter