Porting SBR-Threading

This part does not deal with simple SBR-CALL threading and concentrates on the fast-threading mode using SBR-CALL-ARG threading where it is assumed that the called C-routine will get the BODY-ptr as an argument and the CODE-ptr can be given as an argument via the same way.

Choosing an ARG-register

The ARG-register should be a scratch-register in the local ABI (application binary interface - the description about the structure of call-frames and the registers possibly affected by all into a routine) - if you choose a callee-saved register then it has the effect that the compiler will generate some code in the procedure header to save away that register to the return stack which is absolulty unncessary and possibly counterproductive.

About all cpu ABI do know some scratch register for local computations whose value must not be saved with a procedure and whose value may also not be assumed to be left untouched with an intervening function call. The latter limits the possible use of the value in that register and one should better check if the compiler is clever enough to save value around function calls - there are numourous bugs in compilers for such explicit register allocations.

When looking for a scratch register, the ABI's return register is often a good choice since all the forth words are void-routines. The return-register is generally considered to be touched by function calls, so it falls into the category of a real scratch register. However, the return-register is often the accumulator of traditional cpus, and in many cases that cpu ISA (instruction set architecture - the possible instructions and which registers they can take for operands) must take this register for some instructions to be possible. In most such cases, the displacement- register might be an alternative.

Looking at the current choice, the "%eax" accumulator register is chosen on i386 which actually works, so one does not need to look at the "%edx" register. For the m68k, the return-register "%d0" is not a good choice as the ARG to the routines is always a pointer that points to the actual data - the pointer is used to fetch data with constant displacements, and it is incremented by one cell in many cases - the m68k has some optimized opcodes for these address operations but which work on "%a*" registers only. The compiler would therefore move the "%d0" arg to the scratch "%a0" register for operations. The latter however is not a good choice either, since the compiler does want to use it for indirection function calls like "(*)()" on C level, and does not push it elsewhere. However, the "%a1" register showed to be also a scratch register in this ABI, and it works. Similar decision chains can be made for the powerpc ABI where the "r0" register is used for many hardcoded temporaries, and "r1" is used for the program return stack. Using "r2" succeeds here.

Be aware of the diffences between an ABI and an ISA for a given cpu - in the def-regs section, we do currently one differentiate by cpu-type, which does match with the ISA type. Most cpu makers hand out docs about the recommended ABI for their processor, and in most cases they are overtaken - which leads to a single matching ABI. But this one-to-one relationship can not be assured, see as an example the powerpc ISA which has no dedicated register for the return-stack (it does not even have a dedicated instruction for subroutine calls) and any of the 32 general registers can be used as the stack pointer for an ABI on top of this ISA.

Traditional CPU

A traditional CPU however has a dedicated CALL instruction, a register to save the return address, and a RTS instruction to return-from-subroutine. In this mode, it is best to alias the forth RP and the cpu RP - the current IP in the caller's routine can be deduced by looking at the value in the return-stack. This mode saves these two registers from being needed to be assigned globally.

The RP is declared locally within the USE_CODE_ADDR macro in the C-sources, and the RP value is given as the ARG to this routine (_XE routines in the sources). The current forth-RP is simply the value of the cpu-RP before the call instruction which pushes the IP on the very same stack. Within the _XE routine, the IP can therefore be modified and by using RP[-1] (in general, the offset is indeed just minus one - but most ISAs know some flavour of calls that saves more data to the stack, e.g. call-interrupt-handler).

In this mode, the RP-pointer itself must be left constant since otherwise the CPU will not find its return-adress on the stack. When some _XE routine needs to modify the RP, it must ensure to save the return-address (the IP value), then modify the RP (PUSH, POP, ROOM, DROP), and then restore the return-address as the uppermost value of the real cpu-RP. See the _NEW_RP macros around that handle this problem.

To change the IP alone, one can simply write to the value inside the call-frame - the next RTS will find that value.

As an example, take a look at this m68k assembler of the IF-execution which knows that directly after the IF-call a target-address follows which points to the resp ELSE. Here we chose the "%a5" register to hold the forth parameter-stack:

0000095e _p4_if_execution_:
     95e:       4a9d            tstl %a5@+
     960:       660a            bnes 96c
     962:       2069 fffc       moveal %a1@(-4),%a0
     966:       2350 fffc       movel %a0@,%a1@(-4)
     96a:       4e75            rts
     96c:       58a9 fffc       addql #4,%a1@(-4)
     970:       4e75            rts

reading it: pop the value and test it, if it is non-zero, jump to 96c which skips the target-address by advancing the IP by one cell. The %a1 register has been given as an argument and points into the real return-stack just below the return-address of this function - the add-instruction on "%a1@(-4)" is therefore identical with "%sp@" and you can read the code as "addql #4,%sp@" to modify the address that the following "rts" will jump to. The same accounts for the false-case - here we just fetch the value at the caller's IP which has the target-address, and then store this value into the IP so that the following "rts" will find the new value and BRANCH to it.

One might be tempted here to tell the compiler to just use "sp@" as the IP value, but this is only true if the compiler did not generate a local variable frame - in which case the "sp" would be well above the address of the return-address. The gcc supports the __builtin_return_address but it is not an lvalue, so it can not be modified nor can one take the address of it, and the other __builtin_frame_address is known to carry wrong values when compiled as -fomit-frame-pointer and the local function did not need a frame for real. Therefore, it is safest to just pass down the sp-value from just before the call to the routine and get the IP as RP[-1]. It just needs to put another opcode before the CALL-part to the _XE routine, representing "movl sp, a1" on m68k.

RISC w/ LR type CPU

The powerpc architecture does no have a dedicated return-stack or instructions to call a subroutine or return from a subroutine. Instead it has a special Link Register (LR) and the subroutine calls are derived from the branch-instructions. The branch-instruction (with the L-bit set), will first copy the next code address to the LR and then jump to the target. The code at the target address will then be responsible to fetch the value from the LR and store it on the stack - in the presented case, this stack is referenced by the "r1" global register. The same is done on return-from-subroutine - the return-address is restored into the LR and then a "blr" = branch to address in link-register is compiled.

Therefore, the setup sequence for a new procedure is quite long and simulates the behaviour of a single CALL-instruction on traditional CISC architectures - the same happens as a simulation for the RTS-instruction. This is good practice but imposes a serious problem: how to define the IP. On traditional ISAs, the IP will always be right above the last good RP value, but in this mode, the return-address is nothing more than a local variable of the called routine.

Furthermore, even very simple functions will actually have a local-frame, where they will store the RP of the previous caller, so that the current cpu-RP value is not actually the pointer to the forth-RP data - again, some assumptions may break here if just adding a displacement or trying to fiddle with that __builtin_frame_address - and an RP-change is even more problematic as not just one cell must be saved before the change, but the whole locals-frame of the caller and the callee routine.

It took several hours of experimenting that there are always cases that it will fail to describe the forth-RP as an alias of the cpu-RP. Instead, it is best to not alias the two, and allocate a separate forth-RP as a global register allocation

It turns out that this makes the setup-code for a colon-routine even shorter and faster as we are not bound to the system ABI - in the example powerpc ABI it became visible that the system ABI did usually save the return-address into a place in the callers stack-frame - that simply means that such RP space must be reserved when calling any of the primitives compiled by the C compiler. Using an extra forth RP, we do not need to deal with these specialties, and keep the setup-code down to the minimum.

Since the forth-RP is an extra independent global register on the powerpc, another question must be answered - where to get the callers IP from. One way might be to give it as an ARG to the routine, in the place of the RP that would be normally pushed down, but it happens to be not necessary - instead of suppling some setup-code in the colon-routine just before the call into the forth-subroutine, the IP arg can be setup within the forth-subroutine by creating a copy of the special LR-register in a general-purpose register where it can be easily accessed and modified. The lead-out code then pastes move-arg-to-lr and branch-via-lr into the C-routine (you may notice some garbage left over in the assembler output since the C compiler will then still compile its own lead-out code containing another branch-via-lr).


To alias the forth-RP and the cpu-RP has the advantage of saving another cpu register for local computations - and on a register-starved ISA like intel-32bit, this happens to an essential requirement for good native speed. The contrary is true for an LR-type RISC cpu where the ABI's locals-frame interferes with the forth-operations on an RP-stack. A CPU with enough cpu general registers and traditional return-stack ISA might choose either way.