I don't understand the point about tail call optimizations not being a part of C... Sure, but you can just rewrite the tail call into a loop and get the exact effect, no?
In some sense the reverse of what you said would be a better strategy: rewrite loops as tail calls. Doing that as part of the CPS conversion, and combining with TCO and a garbage collector, gives you a simple non-optimizing compiler for functional languages (but theoretically can be used for imperative languages like C, hence the conversion of a loop to a tail call).
def even(x):
return x == 0 ? True : odd(x - 1)
def odd(x):
return x == 0 ? False : even(x - 1)
print(even(17))
Such is the case e.g. in Erlang/BEAM.
According to the sysv abi, every function has to have a stack frame; so for strict conformance (e.g. for unwinding the stack), you should have a stack frame for every function. In practice, this is kinda dumb. Gcc/clang will do TCO (and won't even generate a stack frame if you pass -fomit-frame-pointer). There's no particular reason why the generated code should abide internally by the sysv abi, only that it be able to call external code written according to said abi (and pass callbacks to said external code, but you'd need thunks for that anyway to account for closures).
I don't see how that would have an effect on TCO, as the storing or not of the FP (beyond in a register) is not part of the calling convention AFAIK. - Only externally linked functions have to conform, so I can understand if TCO does not apply to mutually recursive external functions. Perhaps with LTO it can? - Even an external function that calls itself in tail-position should be TCO-able under -fno-omit-frame-pointer. The compiler reuses the same frame and generates a jump instead of a full function call.
Sysv abi[1], section 3.2.2:
> In addition to registers, each function has a frame on the run-time stack
So it is part of the calling convention. But again, this is IMO pedantry that doesn't really matter.
> Even an external function that calls itself in tail-position should be TCO-able under -fno-omit-frame-pointer. The compiler reuses the same frame and generates a jump instead of a full function call
Yes, I agree. Just interpreting what I think he meant.
1. https://raw.githubusercontent.com/wiki/hjl-tools/x86-psABI/x...