[llvm-commits] [patch] Change how we handle mandatory unwind tables

[Duncan Sands]

Hi Rafael,

> I like not having "eh" in the flag name. The main reason for having abi
> mandated unwind tables is so the external tools can cause an unwind at
> any time.

well that fits into the tri-state scheme too, which is a general scheme for
handling codegen options that need to be mentioned in the IR.

> Unfortunately, I do think I need the 4 possible values. What should the
> optimizer do to mark a uwtables that it was able to show doesn't throw?

I think the nounwind flag should continue to exist independently as it does
now: it is not a codegen flag: it describes the semantics of executing the
IR: no exceptions will be unwound.  The fact that it can be defined without
any mention of how code is generated proves that it is not a codegen flag :)
That's not to say that it has no impact on what codegen does: if the code
generators see the nounwind flag they can try to produce more optimal code
by exploiting that.  For example, they can not bother producing unwind tables.

Ciao, Duncan.