[PATCH] Teach DeadArgElimination not to eliminate return values of functions with 'returned' arguments

[Stephen Lin]

> Well, I'm not advocating doing this, but theoretically if we improved
> CodeGen to generate optimally efficient code around calls to functions
> with 'returned' in all cases (i.e. not just the ARM case where the
> argument and return value are in the same register, and not only in
> the case that the argument fits exactly in one GPR), then I could see
> the possibility of a decent target-independent heuristic for this:
> basically, you would do this if it could consistently reduce the live
> range of a IR value in most/all callers (i.e. the variable would be
> consumed by the function call as its last use, and the return value
> would be used thereafter), so the value does not have to be live
> across the call.

(also, you would want to check that the callee function to be modified
with 'returned' is likely to be able to save the argument either
cheaply or for free; basically the shorter it is and the less calls it
has to opaque functions, the better, I think.)

>
> It wouldn't be guaranteed to work for all targets, though, and it
> would actually impede rather than promote other mid-level optimization
> opportunities usually, so it probably would have to be done in some
> kind of final CodeGenPrepare step with some target-specific knowledge.
> (I'll ignore the issue of whether or not CodeGenPrepare-type passes
> are a good idea.)
>
> Stephen