DAY 13. More Program Control


novice intermediate advanced expert

Self-Modifying Code

Assembly already has the advantage over high-level languages in terms of speed and efficiency, but another plus is the ability of programs to overwrite themselves. As you are well aware, you can define and read/write a variable in the program code:
    LD     A, (data)
    ADD    A, C
    ADC    A, B
    LD     (data), A
    RET
data:    .DB    35
Notice that a variable is just a label with some numbers after it, which is pretty much identical to any other label you have in the program. Therefore, any label can be treated as a variable, even if it was never intended to be one.

    LD    A, 26
    LD    (Blegh), A
    RET
Blegh:
    CP    7
    RET
What happens here? The LD (Blegh), A will change the instruction CP 7 to something else entirely. Taking a lower-level perspective, FE07C9 is changed to 1A07C9. This new op code sequence represents
    LD     A, (DE)    ; $1A
    RLCA             ; $07
    RET              ; $C9
The entire operation of the program has been totally changed! The code has modified itself, or less reflexively, self-modifying code.

An arbitrary example like this shows SMC's primary disadvantage: really hard to interpret code. The main purpose of SMC is to extend the processor's limited capabilities. Example: When using an index register, the offset is a constant number. With SMC, that offset can be altered.

IX_Offset    .EQU    $+2
    INC    (IX + 0)      ; $DD $34 $00
    LD     A, (IX_Offset)
    ADD    A, 3
    LD     (IX_Offset), A
Or you can change the target of a JR or DJNZ.
    LD     A, B
    ADD    A, A
    LD     (Jump), A
Jump    .EQU    $+1
    JR     $00
    INC    IX    ; JR $00
    INC    IX    ; JR $02
    INC    IX    ; JR $04
    INC    IX    ; JR $06
    INC    IX    ; JR $08
    INC    IX    ; JR $0A
It can also be used to save a register faster than the stack:
    LD     (save_a), A
    LD     (save_hl), HL

    ; Do some math on A and HL ...

save_a     .EQU    $+1
    LD     A, $00

save_hl    .EQU    $+1
    LD     HL, $0000

A PUSH without a POP

When you use RET, the processor pops the top value of the stack into PC. Supposing that you pushed a value and followed it with a RET, you would jump to the address corresponding to that value.

Where would this be useful? Assume a scenario wherein a main module jumps to one of about twenty routines. Once any one of these routines is finished, they should return to a specific address in the main module referenced by the label Start. This would have to be done by placing a JP Start at the end of each routine, which would come out to 60 bytes of code.
On the other hand, if the main module were to push the value of Start before jumping to a routine, then the routine would only need a RET to return. This would save 40 bytes of code.

Start:
    LD    HL, Start
    PUSH    HL
    . . .
    JP    Z, Routine01
    . . .
    JP    Z, Routine02
    . . .
    JP    C, Routine03
    . . .
    ; This is for exiting the program
    POP    AF    ; Remove Start from the stack
    RET

Routine01:
    ; Do stuff
    RET    ; Go back to Start

Routine02:
    ; Do stuff
    RET    ; Go back to Start

Routine03:
    ; Do stuff
    RET    ; Go back to Start

Lookup Tables

A lookup table is just an array, that all. But where a lookup table differs from your ordinary array is that a LUT contains pre-calculated data that is not intended to change. Sometimes you have to perform a function that cannot be calculated easily or at all. The use of a LUT lets you get around this. It may also be used for trivial functions where speed is a necessity.

To construct a LUT, first identify the function's domain (the range of possible input values). For each domain value, calculate the result of the function, format it accordingly, and enter it into the LUT.

Example: a LUT used to calculate sin(x), 0° <= x < 90° (in 8.8 fixed-point format):
Element Value Element Value Element Value
0 $0000 1 $0004 2 $0009
3 $000D 4 $0012 5 $0016
6 $001B 7 $001F 8 $0024
9 $0028 10 $002C 11 $0031
12 $0035 13 $003A 14 $003E
15 $0042 16 $0047 17 $004B
18 $004F 19 $0053 20 $0058
21 $005C 22 $0060 23 $0064
24 $0068 25 $006C 26 $0070
27 $0074 28 $0078 29 $007C
30 $0080 31 $0084 32 $0088
33 $008B 34 $008F 35 $0093
36 $0096 37 $009A 38 $009E
39 $00A1 40 $00A5 41 $00A8
42 $00AB 43 $00AF 44 $00B2
45 $00B5 46 $00B8 47 $00BB
48 $00BE 49 $00C1 50 $00C4
51 $00C7 52 $00CA 53 $00CC
54 $00CF 55 $00D2 56 $00D4
57 $00D7 58 $00D9 59 $00DB
60 $00DE 61 $00E0 62 $00E2
63 $00E4 64 $00E6 65 $00E8
66 $00EA 67 $00EC 68 $00ED
69 $00EF 70 $00F1 71 $00F2
72 $00F3 73 $00F5 74 $00F6
75 $00F7 76 $00F8 77 $00F9
78 $00FA 79 $00FB 80 $00FC
81 $00FD 82 $00FE 83 $00FE
84 $00FF 85 $00FF 86 $00FF
87 $0100 88 $0100 89 $0100
The input value must now be transformed into the value it represents in the LUT. This is done as with any other array.

    LD     H, 0
    LD     L, A
    LD     DE, sine_table
    ADD    HL, DE
    LD     A, (HL)
    INC    HL
    LD     H, (HL)
    LD     L, A
    RET
sine_table:
    ; The lookup table
.DW    $0000, $0004, $0009, $000D, $0012, $0016, $001B, $001F, $0024
.DW    $0028, $002C, $0031, $0035, $003A, $003E, $0042, $0047, $004B
.DW    $004F, $0053, $0058, $005C, $0060, $0064, $0068, $006C, $0070
.DW    $0074, $0078, $007C, $0080, $0084, $0088, $008B, $008F, $0093
.DW    $0096, $009A, $009E, $00A1, $00A5, $00A8, $00AB, $00AF, $00B2
.DW    $00B5, $00B8, $00BB, $00BE, $00C1, $00C4, $00C7, $00CA, $00CC
.DW    $00CF, $00D2, $00D4, $00D7, $00D9, $00DB, $00DE, $00E0, $00E2
.DW    $00E4, $00E6, $00E8, $00EA, $00EC, $00ED, $00EF, $00F1, $00F2
.DW    $00F3, $00F5, $00F6, $00F7, $00F8, $00F9, $00FA, $00FB, $00FC
.DW    $00FD, $00FE, $00FE, $00FF, $00FF, $00FF, $0100, $0100, $0100
The most prohibitive drawback to using lookup tables is their giant size, but there's nothing that can be done about that (actually, in the case of trigonometry, you could use the symmetry of the sine function to have a LUT with only the entries for one-quarter of a circle).

Vector Tables

A vector table is a lookup table where each table entry is a 16-bit address. The input value is converted to an address, then a jump to that address is done. This is how the switch() type of control structure works. If you want to relate a value to a subroutine, then a vector table will do it faster and easier than a large, confusing CP/JP block.

Here is an example vector table:

VectTbl:
    .DW     ClearScreen
    .DW     PutSprite
    .DW     DrawLine
    .DW     EndPrgm
    .DW     BlackScreen
The elements of the vector table are accessed just as a lookup table
    LD     H, 0
    LD     L, A
    LD     HL, HL
    LD     DE, VectTbl
    ADD    HL, DE

    LD     A, (HL)
    INC    HL
    LD     H, (HL)
    LD     L, A
    JP     (HL)

Jump Tables

Similar to vector tables, jump tables work by holding entire jump instructions instead of just addresses.
JumpTbl:
    JP    ClearScreen
    JP    PutSprite
    JP    DrawLine
    JP    EndPrgm
    JP    BlackScreen
To call or jump to a routine in the jump table, you use an address of
JumpTbl + 3 * n
where n is the number of the routine you want. Supposing you wanted to run DrawLine, then you would use
CALL JumpTbl + 3 * 2
Jump tables are very useful for shells (the b_call addresses are part of one big jump table (or vector table, does it really make a difference?)). Suppose you made this new shell called "StratocumulusOS" (little joke there for the boys at United-TI :D) that was so freakin' awesome it made MirageOS look like the dog it is (well that remains to be seen ;)), and all its library functions were referenced through a jump table that the shell stored at, oh, let's say $8000.
Address Jump
$8000 JP $9321
$8003 JP $9356
$8006 JP $9400
$8009 JP $9436
$800C JP $945A
$800F JP $949F
Now if someone makes a program for StratocumulusOS and uses the libraries
    CALL    stratocumulus_routines + 3 * 3
This will call $8009 which will then jump to $9436. But, one month later you release StratocumulusOS v1.1 which, owing to a truly brilliant optimization on your part, has a much smaller library size; as well, you store the routines somewhere else (but the jump table is still stored at $8000):
Address Jump
$8000 JP $A58E
$8003 JP $A59C
$8006 JP $A5E2
$8009 JP $A602
$800C JP $A631
$800F JP $A66B
When that original v1.0 program is run under v1.1, then that CALL $8009 will still jump to the right routine.

If you don't mind the size difference, you can also use a jump table as a replacement for a vector table. In this case, you have to multiply A by three (since each jump is three bytes in size).

Relocation

For reasons of security, the TI-83 Plus will crash if PC is ever $C000 or greater. This is a mega-bitch, because it means the executable part of a program can never exceed 8811 bytes. You can get a partial reprieve by copying code to the scrap RAM areas. But if this code references an address within itself with 16-bit addresses, you have to relocate them.
    LD     A, 40
Loop:
    INC    (HL)
    INC    HL
    DEC    A
    JP     NZ, Loop
Say that this piece of code were to be relocated to $9900. The problem is that when assembled, TASM will determine the value of Loop in the JP as relative to $9D95, when it should be relative to $9900. You have to handle situations like this. It's easy: first you subtract $9D95 from the value of Loop, giving the offset from $9D95. Then you add the offset to $9900 to get the correct address. This is the fixed code:
    LD     A, 40
Loop:
    INC    (HL)
    INC    HL
    DEC    A
    JP     NZ, Loop - $9D95 + $9900
If there need to be a lot of relocations you would do well to make some macros.



This is part of Learn TI-83 Plus Assembly In 28 Days
Copyright (c) 2002, 2003, 2004 Sean McLaughlin
See the file gfdl.html for copying conditions