RTS Trick

From NESdev Wiki
Revision as of 19:39, 23 January 2014 by Tepples (talk | contribs) (toning down was a good idea; I'll try to start compromising)
Jump to navigationJump to search

The RTS Trick is one implementation of jump tables, which are a neat way to avoid long, complicated branching code to subroutine calls. It is easier to implement than a pile of branches, reads well, and saves bytes. In fact, the syscall dispatcher for the Apple IIGS toolbox (JSL $E10000) uses this same approach.

When to use the RTS Trick

The RTS Trick is especially useful for anything that can be clearly expressed as a finite-state machine.

Sometimes a module or data format will have a large set of subroutines that will be invoked based on the value of a variable or a data read. One example would be a textbox module for an RPG game. In addition to the mundane displaying of letters you will likely need subroutines to perform various textbox-related tasks. Some examples might be:

  • clear the textbox
  • freeze the textbox (to wait for a button press, for example)
  • insert a pause
  • pluralize words (that is, add or omit an "s" based on a number in a case like: 1 *piece of gold/43 pieces of gold)
  • play a sound
  • close the textbox (end the dialogue)
  • update a variable (set flags, add items to inventory, etc)
  • pop up a yes-no box and prompt the player for a response
  • etc..

Another example would be a music engine. Your song data will have values that represent note values and note lengths, but your data format will also have opcodes that invoke subroutines to perform different tasks. For example:

  • set a segno, or loop point (possibly more than one, so you can nest loops)
  • dal segno, or loop
  • alter the pitch of a playing note
  • change tempo
  • silence a channel or otherwise set volume
  • set volume envelope
  • set sweep envelope
  • set square duty
  • apply vibrato
  • etc.

The more opcode-invoked subroutines you have at your disposal, the more unwieldy branching code will become:

<source lang="6502">

   lda (data_pointer), y   ;read a byte from the text data stream
   bmi textbox_opcodes  ;let's say a value >= #$80 indicates an opcode
                        ;in our imaginary text data format
   ;...do ordinary text stuff
   

textbox_opcodes:

   cmp #$FF
   bne not_end
   jsr tb_end_dialogue
   jmp end_of_opcodes

not_end:

   cmp #$FE
   bne not_clear
   jsr tb_clear_dialogue
   jmp end_of_opcodes

not_clear:

   cmp #$FD
   bne not_pluralize
   jsr tb_pluralize_word
   jmp end_of_opcodes

not_pluralize:

   cmp #$FC
   bne not_add_inventory
   jsr tb_add_inventory
   jmp end_of_opcodes

not_add_inventory:

   ...
    
   etc...

</source>

In the above code example, we read a byte from our data stream and then determine whether that byte represents a character of text or a control character. If it's an opcode, we have to check which opcode. If it FF? no. well is it FE? no. well is it FD? crap. how about FC? The more opcodes you have, the more checks you have to do, and the more painful it will be to rearrange if you decide later to swap the values around. This is where jump tables can help, and the RTS Trick is one efficient way to make a jump table on 6502.

About JSR and RTS

Before talking about the RTS Trick, we need to make clear how JSR and RTS work. When you call a subroutine using JSR, two things happen:

  • the address of the last byte of the JSR (that is, the next instruction minus 1) is pushed onto the stack
  • the program counter jumps to the subroutine indicated.

When the program encounters an RTS instruction, this happens:

  • an address is popped off the stack.
  • the program counter jumps to this address + 1
$C0E0: 20 00 80     JSR $8000
$C0E3: A2 00        LDX #$00
$C0E5: A0 00        LDY #$00

$8000: A9 0F        LDA #$0F
$8002: 8D 15 40     STA $4015
$8005: 60           RTS

Let's run through this code.

  • The Program Counter is at $C0E0.
  • We encounter a JSR instruction. The address of the next instruction - 1 is pushed onto the stack. The next instruction is $C0E3. $C0E3 - 1 is $C0E2. $C0, then $E2 are pushed onto the stack (note the order). Then we jump to $8000.
  • We load #$0F into the accumulator.
  • We store the accumulator in $4015.
  • We encounter an RTS instruction. Two bytes are popped off the stack. In this case, those two values will be E2 and C0. The program counter jumps to this address + 1. $C0E2 + 1 is $C0E3. The Program Counter jumps to $C0E3, which is the instruction after our original JSR.
  • We continue on with our program as normal from $C0E3.

The part we need to take note of for the RTS Trick is the fact that RTS expects the stack to have a destination address MINUS 1. For the rest of this article, let's call this a return address. The two byte return address should be positioned on the stack so that the low byte is popped first, then the high byte. That's all we need to know: Destination address minus 1. Low byte popped first.

Using RTS as a jump table

We said above that we want to avoid having a long chain of comparisons and branches, and that the RTS Trick will help us do that. So what is the RTS Trick?

The RTS Trick consists of a lookup table of return addresses. You read a return address from this table, push it onto the stack, and then RTS. It's really, really simple to understand with an example. Let's use our textbox example above. What do we need?

  • We need a table of return addresses
  • We need a data format that will allow us to index into this table easily.
  • We need a RTS Trick launcher subroutine (I'll tell you why below. Don't worry, it's super-simple)

Step 1: Return Address Table

This is really easy to make. Assuming we already have a bunch of subroutines:

<source lang="6502"> tb_end_dialogue:

   ;do stuff
   rts
  

tb_clear_textbox:

   ;do stuff
   rts
  

tb_pluralize_word:

   ;check a number and add an "s" if it's > 1
   rts
  

tb_add_inventory:

   ;do stuff
   rts
  

tb_set_flag:

   ;do stuff
   rts
 

tb_insert_dramatic_pause:

   ;do stuff
   rts
  

tb_play_dramatic_sound:

   ;do stuff
   rts
  

tb_set_text_color:

   ;do stuff
   rts
   ... etc

</source>

This is how our return address table will look:

<source lang="6502"> tb_opcode_rts_table:

    .word tb_end_dialogue-1
    .word tb_clear_textbox-1
    .word tb_pluralize_word-1
    .word tb_add_inventory-1
    .word tb_set_flag-1
    .word tb_insert_dramatic_pause-1
    .word tb_play_dramatic_sound-1
    .word tb_set_text_color-1
    .word ...etc

</source>

Finished. That's it. Recall that RTS will expect the address on the stack to be its destination - 1. So we tack a minus one to all of our addresses in the table.

Step 2: Table-friendly Data Format

Our text engine reads a byte from our data stream. It checks to see whether that value represents a letter or an opcode. If it is an opcode, we want to turn that value into an index into our return address table. Indexes start at 0 and go up. The data format we have now, counting down from $FF probably won't be quite as convenient as one that counts up. It doesn't really matter from where. It could start at $80, or $C0, or $E0, or it could start from $00 and reach $1F (such as $0A for a new line). For this example, let's keep it simple and choose $80. Here is our new data format:

$00-$7F: letters (a, b, c, d..etc)
$80-$FF(potentially): opcodes
Opcode list:

$80 = end dialogue
$81 = clear the textbox
$82 = pluralize a word
$83 = add an item to the character's inventory
$84 = set a flag
$85 = insert a pause
$86 = play a sound
$87 = set text color
$88 = add to gold total
$89 = subtract from gold total
$8A ... etc

This Data Format for the opcodes is very table-friendly. It starts at $80 and goes up. When we read a byte from our data stream and determine it to be an opcode, all we have to do is chop off bit7 to turn our value into a table index:

$80 = %10000000
$81 = %10000001
$82 = %10000010
$83 = %10000011
       |
       +----------- if we chop off this bit7, we will get $00, $01, $02, $03, etc.
                    That's easy.  Just AND with $7F (%01111111) 
      

Finished. Step 2 Complete.

Step 3: RTS Trick Launcher subroutine.

This is easy. We need a subroutine that will turn our data value into an index, read a return address from the table, push the values and RTS.

<source lang="6502"> tb_opcode_launcher:

   asl             ;we have a table of addresses, which are two bytes each. double that index.
                   ;incidentally, this chops off bit 7 and stashes it in the carry.
   tax
   lda tb_opcode_rts_table+1, x    ;RTS will expect the low byte to be popped first,
                                   ;so we need to push the high byte first
   pha
   lda tb_opcode_rts_table, x      ;push the low byte
   pha
   rts             ;this rts will launch our opcode subroutine.

</source>

That's it!

Potential Pitfall

One point I would like to stress: your launcher must be a subroutine. The opcode subroutines that you are RTS-launching all end in an RTS of their own. Do you know where they will return? Consider the following two examples:

The Right Way

<source lang="6502"> tb_read_data:

   lda (data_pointer), y
   bmi opcode_stuff    ;if negative (bit7 set), we have an opcode
   ;do normal text stuff here

opcode_stuff:

   jsr tb_opcode_launcher    ;key point!  This JSR will put a return address on the stack!
                             ;When our opcode finishes execution, it will return to the right place!
   iny
   ;continue

</source>

The Wrong Way

<source lang="6502"> tb_read_data:

   lda (data_pointer), y
   bmi opcode_stuff
   ;do normal text stuff

opcode_stuff:

   and #$7F
   asl
   tax
   lda tb_opcode_rts_table+1, x
   pha
   lda tb_opcode_rts_table, x
   pha
   rts ;the opcode return address is on the stack, but our starting point isn't.
       ;When the opcode returns, our game will leave this subroutine entirely
   iny ;and will never get here.

</source>

Let's assume that in both cases, the last INY instruction happens to be located at address space $D333.

In the first example, we JSR to our launcher. The JSR will put $D332 on the stack and jump to our launcher. Our launcher will read an opcode return address from our table and push it onto the stack. Then our launcher will RTS, which will pull the opcode return address off the stack, add one, and jump there. Our opcode will execute and return. Our opcode's RTS will pull our D332 address off the stack, add one and jump there. Good! We made it back to our INY instruction.

In the second example, there is no JSR to our launcher. No values are pushed on the stack to mark our place. We read from the table and push an opcode return address onto the stack and RTS to it. Our opcode executes and returns. Our opcode's RTS will pull two bytes off the stack, but we have no idea what those two bytes are. Usually, this will be the return address of whatever called our launcher, but if you have pushed something else onto the stack, it could return anywhere. We are in big trouble.

So make sure your launcher is wrapped up in a subroutine and called via JSR, or that it is otherwise the last thing in a subroutine, and you will be safe.

Games that use the RTS Trick

  • The Guardian Legend uses the RTS trick in its sound engine for opcodes.
  • Concentration Room has three: one for the menu, one for the state of the board, and one for the state of the AI.
  • Famicom Z-machine has two: one for Z-character decoding (which uses a state machine) and one for instruction decoding.

See also

The Fixed Bit Length Encoding article has an example of the RTS Trick in action.