Switching from setjmp/longjmp exceptions to another kind

[gbaraldi]

Setjmp/longjmp exceptions are expensive and require several workarounds to work correctly. Things like the phic nodes/upsilon nodes that cause too much rooting. Also LLVM poorly understands the behaviour of it, basically giving up on many things when seeing a return twice function.
Setjmp is also not a free operation, requiring one to store the whole CPU state.

Talking with @vtjnash I think we should work towards moving to exceptions that look a bit more like C++ ones. Which are "zero-cost".

Related issue: #52533

[Keno]

I'm pretty skeptical. If the argument is performance, we should first switch to the version of setjmp that only stores the stack/ip pointer. If the argument is LLVM optimizability, we can switch to LLVM's exception semantics while lowering to our exception semantics. The problem with C++ zero-cost abstraction is that they require correct synchronous unwind, which we haven't been able to provide at all and not for lack of trying. Unless that is addressed (which is a significant undertaking), I don't think this is feasible.

[gbaraldi]

The issue I see is mostly with the phic nodes. We've seen several issues with them before, they are weird GC root like things that aren't proper GC roots, so they don't get optimized as well, leaving objects to have longer lifetimes, while also introducing many volatile stores in the functions because LLVM doesn't understand their use.