Fix libunwind backtraces & Make defaulttraceinfo more generic.

[JohanEngelen]

Libunwind-based druntime implementation for backtraces results in wrong line numbers (very confusing backtrace).

1. Make DefaultTraceInfo smart enough to not strictly require execinfo.
2. Stop using libunwind for musl. Fix libunwind implementation

[JohanEngelen]

@Geod24 The libunwind implementation needs a rework. The testcase that shows broken behavior clearly is dmd/runnable/test17559.d.
On line 70, frame.address += pip.start_ip; is where the backtrace address for symbol:line resolution is stored. But frame.address is zero there, so it will be loaded with the function's entry address. Not so useful. Instead we should use something like unw_get_reg(&cursor, UNW_REG_IP, &ip); to get the address of the return address (after the function call instruction).

Edit: I now have a proper fix for libunwind aswell. Will need to iron out an issue with importC and @nogc nothrow.

[JohanEngelen]

if it works here for LDC CI, I'll submit it upstream.

[JohanEngelen]

I first had it working with importC, but the LLVM header of libunwind is pretty small so let's use the existing hard-coded binaries (easy to check that they are correct, and the additions needed are arch independent).

[kinke]

Nice that you're hunting down the musl issues! 👍 - How's it looking, many failures still remaining or close to the finish line? Just wrt. v1.38.0 final that I'd like to release this weekend, unless you are close and require a bit more time.

[JohanEngelen]

Let's not wait for the musl work (yesterday I ran stdlib tests for the first time, many succeed, but e.g. fiber tests hung and I have not looked into that at all yet). I'm hoping to reach the endpoint with one Github Action that builds using an alpine container (quick to set up in linux), such that we have a fully statically linked musl release.

[Geod24]

Indeed, thanks for doing this! I haven't personally needed Musl for a few years so never got around to fixing the more tedious issues, but I do remember backtraces not being working fully...

[JohanEngelen]

PR is fully finished. Feel free to squash and merge, or postpone until after 1.38 if deemed too risky.

[JohanEngelen]

PR is fully finished. Feel free to squash and merge, or postpone until after 1.38 if deemed too risky.

there is still something wrong with backtraces. One bug I know (need to adjust backtrace address to point to call instruction instead of right after call instruction), but then there is still something buggy giving slightly incorrect backtrace on throwing an exception.

[JohanEngelen]

Nice that you're hunting down the musl issues! 👍 - How's it looking, many failures still remaining or close to the finish line? Just wrt. v1.38.0 final that I'd like to release this weekend, unless you are close and require a bit more time.

To see current status:
https://github.com/ldc-developers/ldc/actions/runs/8976987555/job/24654919738

[JohanEngelen]

PR is fully finished. Feel free to squash and merge, or postpone until after 1.38 if deemed too risky.

there is still something wrong with backtraces. One bug I know (need to adjust backtrace address to point to call instruction instead of right after call instruction), but then there is still something buggy giving slightly incorrect backtrace on throwing an exception.

This is fixed now (needed to adjust IP to point to the call instruction, rather than after the call).

[JohanEngelen]

It turns out we can use this code fragment for libunwind aswell:

ldc/runtime/druntime/src/core/runtime.d

Lines 673 to 722 in c497e0a

	version (LDC) version (Darwin)
	{
	nothrow:
	extern (C)
	{
	enum _URC_NO_REASON = 0;
	enum _URC_END_OF_STACK = 5;
	alias _Unwind_Context_Ptr = void*;
	alias _Unwind_Trace_Fn = int function(_Unwind_Context_Ptr, void*);
	int _Unwind_Backtrace(_Unwind_Trace_Fn, void*);
	ptrdiff_t _Unwind_GetIP(_Unwind_Context_Ptr context);
	}
	// Use our own backtrce() based on _Unwind_Backtrace(), as the former (from
	// execinfo) doesn't seem to handle missing frame pointers too well.
	private int backtrace(void** buffer, int maxSize)
	{
	if (maxSize < 0) return 0;
	struct State
	{
	void** buffer;
	int maxSize;
	int entriesWritten = 0;
	}
	static extern(C) int handler(_Unwind_Context_Ptr context, void* statePtr)
	{
	auto state = cast(State*)statePtr;
	if (state.entriesWritten >= state.maxSize) return _URC_END_OF_STACK;
	auto instructionPtr = _Unwind_GetIP(context);
	if (!instructionPtr) return _URC_END_OF_STACK;
	state.buffer[state.entriesWritten] = cast(void*)instructionPtr;
	++state.entriesWritten;
	return _URC_NO_REASON;
	}
	State state;
	state.buffer = buffer;
	state.maxSize = maxSize;
	_Unwind_Backtrace(&handler, &state);
	return state.entriesWritten;
	}
	}

So we can completely remove backtrace/handler.d. To be submitted in new PR.