@Peter Huene I started https://github.com/bytecodealliance/wasmtime/pull/2289 -- I'm trying to de-duplicate codegen's unwinding logic, and perhaps remove dependency of gimli crate (in the future?) Next step will be working on new backend. Did you have insights if this approach (UNWIND_INFO
) will work for ARM64 ?
I'll look this over and let you know. I'm not as familiar with Window's ARM64 ABI and the pdata/xdata is quite a bit different (stack-manipulating functions can even be described without having xdata entries provided they follow specific canonical forms and don't have exception handlers).
I put some comments up on the PR. I spent the last hour or two pouring over the Windows ARM64 ABI docs to understand how it differs in describing unwind information from x64. I don't see any inherent problems in the refactoring that would interfere with ARM64 on Windows. My concern is primarily over having to look at every instruction in the function on Windows; for x64 we only ever need to look at the entry block instructions and with ARM64 it's likely to be same provided Cranelift only emits epilogues that have a 1:1 reverse instruction for each prologue instruction.
Yeah, I was wondering about performance too. I'll see if it can be addressed somehow.
Last updated: Dec 23 2024 at 13:07 UTC