[llvm] [SelectionDAG] Not issue TRAP node if naked function (PR #132147)

[Matt Arsenault via llvm-commits]

arsenm wrote:

> I think this PR should have an exception for Wasm target.

I have a couple issues with the current implementation / existence of the TrapOnUnreachable option. First, it's one of the ugly TargetOptions which isn't part of the IR module. In AMDGPUAttributor we try to detect whether any traps are used to enable ABI inputs required for traps on some subtargets, but this option hides them. We can only semi-reliably check for the option there and detect unreachable, as it's not part of the module it depends on consistency of flags between initial optimize and codegen. Plus it assumes later passes will not introduce new unreachables (although if the setting was reliably determinable, we could at least make a conservatively correct check).

Second, this bypasses the ordinary ABI handling you would get for a return. For consistency we should really go through something that looks like LowerReturn for unreachable 

https://github.com/llvm/llvm-project/pull/132147