The EVM does not provide subroutines as a primitive. Instead, calls can be synthesized by fetching and pushing the current program counter on the data stack and jumping to the subroutine address; returns can be synthesized by getting the return address to the top of the stack and jumping back to it. These conventions are more costly than necessary.
Facilities to directly support subroutines are provided by all but one of the real and virtual machines programmed by the lead author, including the Burroughs 5000, CDC 7600, IBM 360, DEC PDP 11 and VAX, Motorola 68000, a few generations of Intel silicon, Sun SPARC, UCSD p-Machine, Sun JVM, Wasm, and the sole exception the EVM. In whatever form, these operations provide for
> We also wish to be able to arrange for the splitting up of operations into subsidiary operations. This should be done in such a way that once we have written down how an operation is done we can use it as a subsidiary to any other operation.
> When we wish to start on a subsidiary operation we need only make a note of where we left off the major operation and then apply the first instruction of the subsidiary. When the subsidiary is over we look up the note and continue with the major operation. Each subsidiary operation can end with instructions for this recovery of the note. How is the burying and disinterring of the note to be done? There are of course many ways. One is to keep a list of these notes in one or more standard size delay lines, (1024) with the most recent last. The position of the most recent of these will be kept in a fixed TS, and this reference will be modified every time a subsidiary is started or finished...
We propose to follow Turing's simple concept in our subroutine design, as specified below. Note that this specification is entirely semantic. It constrains only stack usage and control flow and imposes no syntax on code beyond being a sequence of bytes to be executed.
We introduce one more stack into the EVM in addition to the existing `data stack`, which we call the `return stack`. The `return stack` is limited to `1024` items. This stack supports three new instructions for subroutines.
To take full advantage of the performance benefits of simple subroutines we also provide two new static, relative jump functions that take their arguments as immediate data rather then off the stack.
* _If a resulting `PC` to be executed is beyond the last instruction then the opcode is implicitly a `STOP`, which is not an error._
* _Values popped off the `return stack` do not need to be validated, since they are alterable only by `JUMPSUB` and `RETURNSUB`._
* _The description above lays out the semantics of this feature in terms of a `return stack`. But the actual state of the `return stack` is not observable by EVM code or consensus-critical to the protocol. (For example, a node implementor may code `JUMPSUB` to unobservably push `PC` on the `return stack` rather than `PC + 1`, which is allowed so long as `RETURNSUB` observably returns control to the `PC + 1` location.)_
We modeled this design on Moore's 1970 [Forth virtual machine](http://www.ultratechnology.com/4th_1970.pdf). It is a two-stack design – the data stack is supplemented with a return stack to support jumping into and returning from subroutines, as specified above. The return address (Turing's "note") is pushed onto the return stack (Turing's "delay line") when calling, and the stack is popped into the `PC` when returning.
The alternative design is to push the return address and the destination address on the data stack before jumping, and to pop the data stack and jump back to the popped `PC` to return. We prefer the separate return stack because it ensures that the return address cannot be overwritten or mislaid, uses fewer data stack slots, and obviates any need to swap the return address past the arguments or return values on the stack. Crucially, a dynamic jump is not needed to implement subroutine returns`.
The _low_ cost of `JUMPSUB` is justified by needing only about six Go operations to push the return address on the return stack, and decode the immediate two byte destination to the `PC`. The _verylow_ cost of `RETURNSUB` is justified by needing only about three Go operations to pop the return stack into the `PC`. No 256-bit arithmetic or checking for valid destinations is needed. Also, `JUMP` is assigned _mid_, and `JUMPSUB` and JUMPR should be more efficient, as decoding immediate bytes should be cheaper than than converting 32-byte stack items, and the destination address will not need to be checked for either `JUMPSUB` or `RETURNSUB`. Benchmarking will be needed to tell if the costs are well-balanced.
These opcodes reduce the gas costs of both ordinary subroutine calls and low-level optimizations. The savings reported here will of course be less relevant to programs that use a few large subroutines rather than being a factored than in to smaller ones. The choice of gas costs for the new opcodes above does not make a large difference in this analysis, as much of the improvement is due to PUSH and SWAP operations that are no longer needed. Even if `JUMPSUB` cost the same as `JUMP`– 8 gas rather than 5 - a simple subroutine call would still be 52% less costly versus 54%.
These changes are compatible with using [EIP-3337](https://eips.ethereum.org/EIPS/eip-3337) to provide stack frames, by associating a frame with each subroutine.
These changes do introduce new flow control instructions, so any software which does static/dynamic analysis of EVM code needs to be modified accordingly. The `JUMPSUB` semantics are similar to `JUMP` (but jumping to a `BEGINSUB`), whereas the `RETURNSUB` instruction is different, since it can 'land' on any opcode (but the possible destinations can be statically inferred).
If [`EIP-`3779](./eip-3779.md) – Safe Control Flow for the EVM – advances then the requirement on `JUMPSUB` to `abort` if the opcode at `location` is not a `BEGINSUB` will need to be enforced at creation time rather than runtime.
In this example. the JUMPSUB is on the last byte of code. When the subroutine returns, it should hit the 'virtual stop' _after_ the bytecode, and not exit with error
