Runtime code relocation by Cadaver
----------------------------------

0. Introduction

Starting from Metal Warrior 4, Covert Bitops games have loaded some code during
runtime, often to implement title screens or story scripting, or enemy logic
that is not required at every point in the game. This way, the amount of all-
time resident code in the "main load" or "engine" is kept smaller.

Metal Warrior 4 and Hessian both employed a 2 KB fixed-address buffer for the
loadable code; only one such code chunk would be in memory at once and would
always be fully overwritten when the next chunk was loaded. In this case, the
code in each chunk could be pre-assembled to the fixed address.

To avoid having to define fixed maximum sizes for various data and to be able
to utilize free memory better, Hessian already used dynamic memory allocation
for resource data like level maps and sprites, but from Steel Ranger onward,
also code is considered a dynamically allocated resource. Because now the code
load address is not known beforehand, or previously loaded code can move in the
memory as resource chunks are loaded or unloaded, runtime code relocation
becomes necessary.


1. Relocation algorithm

The principle of relocation is simple, as long as some limitations are kept.

It only requires recognizing 3-byte long instructions, which contain a 16-bit
absolute address. If the address is a reference to the loadable code itself, or 
data it has included in the same resource chunk, it must be relocated. Loadable 
code may also access the variables and routines of the resident engine code, or 
the C64's hardware registers, and these accesses need to be left alone. Handling 
this properly requires just a memory range check: if the address is within the 
dynamic allocation area, it should be relocated.

There must also be a way to detect the end of the code. The BRK instruction
($00) was chosen for simplicity.

When a new piece of loadable code is loaded into memory for the first time, it
uses the addresses it was initially assembled to. In Covert Bitops games $8000 
is used, which is guaranteed to be inside the dynamic allocation area.

The relocation algorithm becomes:

- Get next instruction, until BRK encountered
- Is it a 3-byte long instruction?
- If yes, check the address. Is it within the dynamic allocation area?
- If yes, add a delta value to the address to match the code's current load 
  address


2. Limitations

The code to be relocated must be unambiguous, ie. tricks that use for example
the BIT instruction to skip the next instruction can not be allowed, as this 
gives multiple interpretations of the instructions.

Including data tables that contain addresses (either 16-bit, or separated into
8-bit low and high tables) is a problem, as they can not be detected for 
relocation by just going through code. The algorithm simply chooses to not 
support this.

This can be circumvented somewhat, as the dynamic allocation system supports 
querying the current address of any "object" inside a resource chunk. The
loadable code can define extra objects it needs, for example text strings, that
it can use when executing.

In Covert Bitops games, loadable code is handcrafted in assembly to correspond 
to the resource chunk format, as some extra data like the total data size and
number of resource objects are needed. See examples from the Steel Ranger demo
source code and the c64gameframework example game.


3. Optimizations

A naive implementation would use a 256-byte table to store the length of every
instruction.

In practice, regularities in the legal instruction set can be observed. 
Instructions ?c - ?f are always 3 bytes, and instructions ?4 - ?7 are always 2 
bytes. For the remaining instructions, 4 lengths can be packed into one byte, so 
the final table becomes only 32 bytes.


4. The code

This is adapted from file.s in the c64gameframework project. Two zeropage
pointers (ptr - the code to relocate, and relocDelta - the relocation offset to 
add) are used, and the instruction length table is below the code.

CodeRelocLoop:  ldy #$00
                lda (ptr),y                     ;Read instruction
                beq CodeRelocDone               ;BRK - done
                lsr
                lsr
                lsr
                bcc LookupLength
                and #$01                        ;Opcodes xc - xf always 3 bytes
                ora #$02                        ;Opcodes x4 - x7 always 2 bytes
                bne HasLength
LookupLength:   tax
                lda (ptr),y
                and #$03
                tay
                lda instrLenTbl,x               ;4 lengths packed into one byte
DecodeLength:   dey
                bmi DecodeLengthDone
                lsr                             ;Shift until has the wanted
                lsr
                bpl DecodeLength
DecodeLengthDone:
                and #$03
HasLength:      cmp #$03                        ;3 byte long instructions need
                bne NotAbsolute                 ;relocation
                ldy #$02
                lda (ptr),y                     ;Read absolute address highbyte
                cmp #>fileAreaStart             ;If not reference to self, skip
                bcc NoRelocation
                cmp #>fileAreaEnd
                bcs NoRelocation
                dey
                lda (ptr),y                     ;Add relocation offset to the
                adc relocDelta                  ;address
                sta (ptr),y
                iny
                lda (ptr),y
                adc relocDelta+1
                sta (ptr),y
NoRelocation:   lda #$03
NotAbsolute:    clc                             ;Proceed to next instruction
                adc ptr                         ;(length of current in A)
                sta ptr
                bcc CodeRelocLoop
                inc ptr+1
                bne CodeRelocLoop
CodeRelocDone:  rts

instrLenTbl:    dc.b $99,$99,$9a,$dd,$9b,$99,$9a,$dd
                dc.b $99,$99,$9a,$dd,$99,$99,$9a,$dd
                dc.b $aa,$99,$da,$dd,$aa,$99,$9a,$dd
                dc.b $aa,$99,$9a,$dd,$aa,$99,$9a,$dd