Disable libunwind for NuttX

[no1wudi]

NuttX is a resource-constrained system, libunwind is too large for it in terms of both code size and memory usage. So we disable libunwind for NuttX.

[workingjubilee]

@no1wudi why are you building with unwinding enabled?

[no1wudi]

@workingjubilee Previously, I have been using the RISC-V platform for development work related to Rust and NuttX. Recently, when I started testing on the ARM platform, I discovered that the ARM platform references _Unwind_GetIP, which is currently not supported in the NuttX environment. Considering that the size of the Rust code could be reduced by approximately 20% after disabling libunwind, we are currently considering turning it off for the time being.

[workingjubilee]

@no1wudi it's... not supported?

...but this is Arm we're talking about...?

[no1wudi]

@no1wudi it's... not supported?

...but this is Arm we're talking about...?

For ARM platform in NuttX, we are using unwind support from GCC (libgcc.a from arm-none-eabi-gcc), it does not provide an implementation of _Unwind_GetIP.

This branch works, implement _Unwind_GetIP on top of _Unwind_VRS_Get:

backtrace-rs/src/backtrace/libunwind.rs

Lines 214 to 291 in f8cc6ac

	use core::ptr::addr_of_mut;
	// On android and arm, the function `_Unwind_GetIP` and a bunch of
	// others are macros, so we define functions containing the
	// expansion of the macros.
	//
	// TODO: link to the header file that defines these macros, if you
	// can find it. (I, fitzgen, cannot find the header file that some
	// of these macro expansions were originally borrowed from.)
	#[repr(C)]
	enum _Unwind_VRS_Result {
	_UVRSR_OK = 0,
	_UVRSR_NOT_IMPLEMENTED = 1,
	_UVRSR_FAILED = 2,
	}
	#[repr(C)]
	enum _Unwind_VRS_RegClass {
	_UVRSC_CORE = 0,
	_UVRSC_VFP = 1,
	_UVRSC_FPA = 2,
	_UVRSC_WMMXD = 3,
	_UVRSC_WMMXC = 4,
	}
	#[repr(C)]
	enum _Unwind_VRS_DataRepresentation {
	_UVRSD_UINT32 = 0,
	_UVRSD_VFPX = 1,
	_UVRSD_FPAX = 2,
	_UVRSD_UINT64 = 3,
	_UVRSD_FLOAT = 4,
	_UVRSD_DOUBLE = 5,
	}
	type _Unwind_Word = libc::c_uint;
	extern "C" {
	fn _Unwind_VRS_Get(
	ctx: *mut _Unwind_Context,
	klass: _Unwind_VRS_RegClass,
	word: _Unwind_Word,
	repr: _Unwind_VRS_DataRepresentation,
	data: *mut c_void,
	) -> _Unwind_VRS_Result;
	}
	pub unsafe fn _Unwind_GetIP(ctx: *mut _Unwind_Context) -> libc::uintptr_t {
	let mut val: _Unwind_Word = 0;
	let ptr = addr_of_mut!(val);
	let _ = _Unwind_VRS_Get(
	ctx,
	_Unwind_VRS_RegClass::_UVRSC_CORE,
	15,
	_Unwind_VRS_DataRepresentation::_UVRSD_UINT32,
	ptr.cast::<c_void>(),
	);
	(val & !1) as libc::uintptr_t
	}
	// R13 is the stack pointer on arm.
	const SP: _Unwind_Word = 13;
	pub unsafe fn get_sp(ctx: *mut _Unwind_Context) -> libc::uintptr_t {
	let mut val: _Unwind_Word = 0;
	let ptr = addr_of_mut!(val);
	let _ = _Unwind_VRS_Get(
	ctx,
	_Unwind_VRS_RegClass::_UVRSC_CORE,
	SP,
	_Unwind_VRS_DataRepresentation::_UVRSD_UINT32,
	ptr.cast::<c_void>(),
	);
	val as libc::uintptr_t
	}
	// This function also doesn't exist on Android or ARM/Linux, so make it
	// a no-op.
	pub unsafe fn _Unwind_FindEnclosingFunction(pc: *mut c_void) -> *mut c_void {
	pc
	}

However, I have observed that compared to disabling libunwind, the code size increases by approximately 20%. Considering that as an embedded system, we typically analyze call stack information and the like on a PC, it is therefore advisable to disable it in order to conserve code space.

[workingjubilee]

Isn't this what panic_immediate_abort is for?

[no1wudi]

Get it, thank you !

I originally thought that setting -Zbuild-std=std,panic_abort would be sufficient, but I overlooked -Zbuild-std-features=panic_immediate_abort.