So you like Applesoft ? And you think you can write an action game with it ? Or maybe a science program ? Yes, you can ... will it be fast ? ... Probably not ...
In order to do that, I'm using [AppleWin](https://github.com/AppleWin/AppleWin), an Apple II emulator that has a cycle counting/difference feature. What I do is set a breakpoint within the Applesoft ``NEWSTT`` routine in ``$D801``. The ``NEWSTT`` routine is responsible for checking if there's a (new) statement to process, either on the same line (then, separated with a colon ``:``) or on a new line. In ``$D801`` a new line has been detected and is about to be executed (although there's first a check to see if ``TRACE`` is on and so if it's needed to print on the screen the line number being executed). So, except for a check here and there, setting a breakpoint in ``$D801`` will count the cycles needed to execute a whole line. It gives a good indication of the speed needed and can be used as a base for cycle counts comparisons.
So, we are going to compare code snippets speed. For example, is it faster to divide a number by 2 or to multiply it by 0.5 ? To make sure we don't enter some special cases where values of ``zero`` are treated differently, we first initiate some variables, usually in line 10. The code we actually want to test will be in line 20 most of the time, while line 30 will be a simple ``END`` statement. ``END`` is not necessary normally to end a program but remember that the breakpoint in ``$D801`` only occurs when a **new line** is found, that's why we must finish our code with an ``END`` statement, on a new line
This is faster as line 20 took only 3236 cycles, a difference of **723 cycles** ! (and you already have a first technique to increase speed, I'll explain it later).
Notice that both snippets have the exact same result: variable ``C`` now holds the value ``9`` (which is 18 divided by 2 or 18 multiplied by 0.5).
All our snippets will have the same final effect, otherwise we would not be comparing fairly. For example, the first snippet used a variable assignment in line 20 (``C=A/B``). For the second snippet, it's important we use another variable assignment (``C=A*B``) because we want to compare the speed of the multiplication and the speed of the division. If we had used ``A*B`` with a statement like ``PRINT A*B`` or ``K=PEEK(A*B)`` or ``HTAB A*B``, the cycles taken to handle the statement would disturb our measure and we would be comparing apples and oranges.
It is also important that we did not use ``A=A*B``: even though it's a variable assignment, we would be reusing ``A`` and it has an impact on speed. If we want to reuse ``A`` then we need to do it in both snippets.
The actual difference of **723 cycles** does not really matter. What is important is that the second snippet **actually runs** faster. Actual speed depends on several other factors which will be explained in this article. &
* The cycles count on this page are only an indication of the speed of the code we want to "benchmark".
* The exact cycle count is **not** what matters.
* **Comparison** of cycles count is what we're studying.
* Smaller cycles counts are faster and are considered as a technique to apply whenever possible.
* Sometimes, if you're not careful, using a technique explained here could be **slower** if you don't pay attention to other factors. If that's the case, it will be explained.
1. [Use variables as placeholders for constant values](#1-use-variables-as-placeholders-for-constant-values): accessing a known value in a variable is faster than deciphering values in code.
2. [Declare your most used variables first](#2-declare-your-most-used-variables-first): create and/or reference the variables you're going to use the most as soon as possible
4. [Use addition instead of multiplication by 2](#4-use-addition-instead-of-multiplication-by-2): addition of the same variable twice is faster than multiplying the variable by 2
5. [Addition is faster than subtraction](#5-addition-is-faster-than-subtraction): avoid subtraction whenever possible but don't use negative constants.
Let's consider a code like ``K=PEEK(49152)`` (this code gets the ASCII code of the last key pressed, plus 128 if the keyboard probe has not been reset and store it in variable ``K``).
When this code is run , the Applesoft parser will perform the following:
1. search for a "real/float" variable named ``K``, and create one if needed
2. encountering the ``=`` sign, the parser knows that an expression will be evaluated and attributed to variable ``K``
3. the expression in this case is a memory read request (``PEEK``)
4. the parser will then collate the memory location by evaluating what's between the parenthesis (it could be a formula involving other variables for instance). In this case it will just read the number, character by character:
5. Collating these, results in ``4 9 1 5 2`` as 5 ASCII characters. These represent, for us, humans, a decimal number but not yet for Applesoft.
6. These 5 characters will then be converted to a real number (using a format known as binary floating-point format)
7. Then, the real number is converted to an integer value (because ``PEEK`` expects a 2-bytes integer)
8. Once this has been done, the value in the appropriate location is read, converted from byte to a binary floating-point value and attributed to variable K
The bottleneck here are the steps 4-6. Building a integer representing a memory location from characters is long.
It is probable that your game will need to read the keyboard regularly. Why do you have to repeat steps 4-6 every time you need to get the last key pressed ? Fortunately, there's a workaround.
It is actually faster for the Applesoft parser to locate a variable in memory and use its value than to "recreate it from scratch". So, all you need to do is save in a variable the value you want to repeatedly use.
line 20 here takes 7128 cycles, that's a difference of **4825 cycles** ! This is ***HUGE*** especially when it's a statement that's going to be executed every time the main game loop cycles !
this line 20 only takes **390** more cycles. This is because ``0`` is only 1 character, while ``49152`` is 5 characters. But anyway, even if the difference is not that important, it's faster.
Should you convert all your constants to variables ? My advice is yes, particularly for the constants used in loops or repeatedly. Among those are:
* Values you might use constantly (often powers of 2) like ``4``, ``8``, ``16``, ``32``, ``64``, ``128`` and ``256`` ... or maybe their lower limits like ``3``, ``7``, ``15``, ``31``, ``63``, ``127`` and ``255``
* Other values you will certainly use like ``0``, ``1`` and ``2``. I like to put these in variables ``Z``, ``U`` ("unit(ary)") and ``T`` (as in "two")
* the screen limits: think of ``VTAB 24``, ``HTAB 40``, ``SCRN(39,39)``, ``HPLOT 279,159`` or their upper boundaries like ``40``, ``280`` ``160`` and ``192``.
* loops' low and high limits: ``0``, ``1`` up to ``9`` , ``10`` or ``19`` and ``20``, etc. Think of ``FOR I=0 TO ...`` or ``FOR I=1 TO ...``
Whatever the value, whether it's an integer or a real <sup>(*)</sup>, this rule will **always** speed up your code, except if you're not careful about the next technique ...
* right after the program's last line is an area pointed by ``VARTAB`` (in zero-page vector ``$69-$6A`` -- decimal ``105-106``) where all the real, integer and string variables are defined and stored <sup>(*)</sup>. It's also where references to "functions" created by ``DEF FN`` are stored.
* just after that area is another area, pointed by the vector ``ARYTAB`` (in ``$6B-$6``, decimal ``107-108``) where all the arrays are stored.
<sup>(*) string variables are not stored in that area, but pointers to their values are stored there. The actual values of string variables being either in the program code itself or in a special area after the array storage area</sup>
* if it's not found, it means it's a new variable.
* If the code where it appears is a variable assignment, then the appropriate space (variable name, type, array indices, value) is reserved at the top of the memory pile where all variables of the same type reside (optionally moving the ``ARYTAB`` area up if a new real/integer/string/function variable needs to be declared).
* If it's not a variable assignment, then the variable type's default value is referenced for next step but the variable IS NOT created.
* if the variable already exists, its value is referenced for the next step
As you see, numeric (float/real and integer) variables, string variables and arrays are stored in several different ways but they all share one thing in common: once a variable is encountered and once its type has been determined, the Applesoft parser will search for the variable in one of the two memory locations in the same way: from one end to the other.
This means that variables are not "ordered" by their names ... It means that, in memory, variable Z might be stored/referenced before variable A... It also means that the time spent to look for a variable depends on how soon it was found in the code. How much time ? Let's find out.
Let's create a variable ``A`` and another named ``Z`` with equal values, then let's print the value of variable ``A`` and then in a second snippet, the value of variable ``Z``.
This takes 27898 cycles. That's a difference of 34 cycles. It looks **insignificant** and, as such, **it is** ! but it has an impact on all the other techniques I'm gonna teach you.
Let's have another example. Now this time, we will declare 26 different variables named from ``A`` to ``Z`` and see the cycles count difference when accessing the first one or the last one declared.
Line 20 took 20241 cycles. Second snippet is identical except we access variable ``Z`` instead of variable ``A``. You'll notice that the values of these two variables are identical to eliminate the possible fact that different values are handled with different speeds.
This took 21026 cycles. The difference is **only** 785 cycles. Let's be honest, it's not gigantic.
But ! Wait ! Remember that snippet in the section [Use variables as placeholders for constant values](#use-variables-as-placeholders-for-constant-values) where we handled value ``0`` ?
It had a difference of 390 cycles just by replacing a hardcoded/constant value of ``0`` with a variable name. It would mean that if we're not careful, we might lose the advantage we took for granted.
Let me rephrase this: imagine if Z was holding a value you need to use **OFTEN** ... myself I like to put **zero** in Z because it's obviously a good variable name for such a value ...
Let's see that with two other snippets. Snippet #1 will declare ``Z`` first, snippet #2 will declare ``Z`` last and snippet #3 will not use ``Z`` but a hardcoded value of ``0``
Line 20 took 20672 cycles, a difference of only 431 cycles with the first snippet where we use ``Z=0`` as the first declared variable, but also it's 354 cycles **faster** than the version where ``Z=0`` is declared last ! Thus, negating any interest in replacing ``0`` with a variable if it's not declared in time !
Your most used variables should be declared first. In fact **you should have a line in your code where all these variables are declared/created before doing anything else**, otherwise you might inadvertently create a variable. The most common error being to display the instructions or a splash screen for the game and then wait for a keypress with something like ``GET K$``, as ``K$`` might be your very first declared variable !
So which variables should you declare first ? and with many variables to declare, how do you know if it's best to use a variable or an actual value ? It depends on many factors.
It's best to declare the variables used in your main game loop first. Most common variables and constants are possibly:
* previous player position (like ``OX,OY`` although you should prefer single-character variables like ``A,B`` or ``V,W``, more about that later)
* loop counters (like ``I,J``) as used in ``FOR/NEXT`` loops or other loops
* ``49152``, memory location to read a key and ``49168`` to clear the keyboard strobe (but more about that later)
* expected ASCII+128 values (``201``, ``202``, ``203`` & ``204`` are for I/J/K/L which are 4 directions keys on EVERY latin keyboard around the world), maybe ``160`` for space bar, etc.
* I like to use single variables for very common values like ``Z`` for ``0``, ``U`` (unit(ary)) for ``1`` and ``T`` for ``2`` ... it depends if you need these or not ...
* a variable to hold an energy meter (``E`` ?) or a score (``S`` ?)
* player speed (horizontal, vertical)
* a shape rotation ?
* enemies positions + previous cycle positions
* missiles/bullets positions
* etc.
Once you know which variables you use in your main game loop, you need to consider the following:
In this snippet, ``X`` is incremented and checked against a maximum limit. In the extreme case where ``X`` exceeds the limit, its value is set to that limit.
This is the kind of code that typically happens when drawing a moving object on the screen.
If you consider that line 20 is part of the main loop, then ``X`` is referenced 3 times:
The limit variable ``M`` is referenced only once during the comparison, then a second time when the comparison result is true. As the comparison will probably be false most of the time, you can consider that ``X`` is referenced 3 times, while ``M`` only once. But even if the comparison was true most of the time, ``X`` would still be referenced more often than ``M``. Obviously ``X`` should be declared before ``M``.
The same kind of process should be made with the variable ``U``. Should it be declared before ``M`` ? With these two snippets, ``U`` is referenced only once, whereas ``M`` could be referenced twice when ``X>M`` ... but it's probable that ``U`` (placeholder for the constant ``1``) is used elsewhere in the main game loop, while ``M`` has not many other uses than to check X-coordinates maximum limit ... so ``U`` will probably be more efficiently referenced if declared before ``M``.
## 3) Use one-letter variables names whenever possible
### 🍎Forget about meaningful variable names
Applesoft only supports variables names made of one or two characters. The first character **must** be a letter, while the second character may be a letter **or** a number.
It is allowed to name variables with more characters but these are ignored, meaning that variable ``ABRACADABRA`` is really stored as ``AB`` and maybe referenced later in the code with that name. It also means that you can't really have a variable named ``A10`` as it will be in fact variable ``A1``.
This behavior is true for any kind of variable (float, integer, string, array of (float/integer/strings), and even ``DEF FN`` which are just referenced as another kind of variable). It means that for every variable type, you can't have more than 26 + 26 * (26+10) variables = 962 different variables names.
The problem is that the Applesoft parser will take 56 cycles for every extra character in a variable name.
```basic
10 A=17
20 END
```
Line 10 takes 2776 cycles. Now, let's use variable ``AB`` instead of ``A``:
```basic
10 AB=17
20 END
```
Line 10 now takes 2832 cycles, that's 56 more cycles. Now, for the extreme example:
```basic
10 ABRACADABRA = 17
20 END
```
Line 10 takes now 3336 cycles, which is 504 cycles slower, which is exactly 56 cycles for each of the extra 9 characters after ``AB``.
Every time a variable is parsed, it takes 56 more cycles to parse a 2-characters variable name than if the variable name was only 1 character. And this is true for every use of the variable: assignment, calculation, print, memory read/write, 6502 subroutine call, etc.
### 🍎Recommendations
- Use one-letter variable names as much as possible in your main loop.
- It means, **yes**, you need to avoid ``OX,OY`` for the previous cycle's ``X,Y`` coordinates.
- If needed, re-use variables names declared before starting the main loop if these variables hold values you don't care about anymore.
- Remember to [declare your most used variables first](#2-declare-your-most-used-variables-first) !
- Use as few two-characters variables names as possible in your main loop, as each 2-characters variable name will take 56 more cycles **just** to parse the name of the variable in the code.
- **NEVER** use more than two-characters variables names in your main loop <sup>(*)</sup>.
<sup>(*) you should apply this rule even if you're not looking for speed as confusion is waiting behind the corner: variable ``LEVEL`` and variable ``LENGTH`` both refer to variable ``LE`` ...</sup>
This is **always** true if what you want to do is ``A=2*B`` and that you use variables and replace hardcoded constants with variables (see section [Use variables as placeholders for constant values](#1-use-variables-as-placeholders-for-constant-values)). If you don't, you might get mitigated results.
Line 20 takes 6795 cycles. Notice how line 10 declares five variables ``A-E``. These variables will be used in the subsequent snippets. Declaring them, even though they're not used, allows us to ignore the extra cycles needed to create a new variable.
Line 20 takes 9072 cycles, which is slower (2277 cycles slower).
Now you might think that storing the result of ``A/B`` would be faster. It's not. Except, maybe if you intend to use that result elsewhere in your code in which case it might be worth to spend those cycles storing a result in a variable.
First snippet demonstrates the speed if you don't care about the result of ``A/B``
Line 20 takes 7090 cycles, it's 295 cycles slower than using directly ``E=C*A/B``.
This second snippet illustrates the speed if the result of ``A/B`` is of any interest and is meant to be reused several other times: it's thus calculated on line 10 and excluded from cycles count.
The addition of ``A+B`` and assignment of the result to variable ``C`` in line 20 takes 2171 cycles.
Now if we had this line instead
```basic
20 C = A-B
```
It would take 2327 cycles, a difference of **156** cycles in favor of addition.
Knowing that, your initial intuition would be to replace subtraction with additions whenever possible. It's easy as all you have to do is to make the second operand negative. Unfortunately ...
### 🍎Adding negative numbers is only slightly faster than subtraction
With the previous example, if `B` is negative, we have
```basic
10 A = 123: B = -85: C = 0
20 C = A+B
30 END
```
Line 20 takes 2307 cycles, which is marginally faster (20 cycles) than subtracting a positive number. Is it always like that ?
### 🍎Avoiding addition: is it worth it ?
Here's a real life example where you might be tempted to add a negative number instead of subtracting a positive number to obtain the same result.
Let's say you're trying to center the contents of ```A$``` on screen, your code will look like
```basic
10 D=2: V=20: A$ = "APPLESOFT: THE NEED FOR SPEED"
20 HTAB V-LEN(A$)/D
30 PRINT A$
40 END
```
Line 20 will take 5846 cycles. Now, if you set `D=-2` and use addition instead of subtraction:
```basic
10 D=-2: V=20: A$ = "APPLESOFT: THE NEED FOR SPEED"
20 HTAB V+LEN(A$)/D
30 PRINT A$
40 END
```
Line 20 now takes 5826 cycles, which is only 20 cycles faster. **AGAIN !**
Even worse: it's likely you'll have to declare in your code the constant ``2`` and assign it to a variable for other purposes. So now your code looks like this:
```basic
10 D=2: V=20: A$ = "APPLESOFT: THE NEED FOR SPEED": E=-2
20 HTAB V-LEN(A$)/D
30 PRINT A$
40 END
```
As ``E`` is still declared last and the order of the other variables is identical to the order in our first snippet, line 20 takes exactly the same cycles count: 5846 cycles.
But if we change line 20 to
```basic
20 HTAB V+LEN(A$)/E
```
Line 20 takes now 5928 cycles ! This is 82 cycles **slower** than subtraction ! Even if we declare ``E`` earlier:
```basic
10 D=2: E=-2: V=20: A$ = "APPLESOFT: THE NEED FOR SPEED"
```
(line 20 is still an addition), this again takes 5928 cycles ! It's only when we declare ``E`` before ``D`` that
```basic
10 E=-2: D=2: V=20: A$ = "APPLESOFT: THE NEED FOR SPEED"
```
we see an improvement: 5894 cycles ! But compared to negation (5846 cycles), it is still slower ! We need to declare ``D`` last to see an advantage (because ``D`` is not used anymore in our snippets).
```basic
10 E=-2: V=20: A$ = "APPLESOFT: THE NEED FOR SPEED": D=2
```
We now have 5826 cycles for line 20, this is just 20 cycles faster than if we had used subtraction.
As you can imagine, inverting the order of declaration of ``E`` and ``D`` is not worth it. ``D`` is now declared last, which might have an impact on speed on other parts of our code where the constant ``2`` is more important than ``-2`` ...
### 🍎When is it worth to substitute subtraction with addition then ?
Adding the negative of a number (instead of subtraction) has rarely a positive impact on speed.
But there are other times when you can substitute negation and addition and gain something in return: in comparisons.
You certainly know that
```basic
IF A > B-C THEN ...
```
can be rewritten as
```basic
IF A+C > B THEN ...
```
This also works with any other comparison operator: ``=``, ``<>``, ``>``, `<`, `>=` and `<=`
Let's just see how much faster it is with a simple example.
```basic
10 A=10: B=20: C=9
20 IF A>B-C THEN D=A
30 END
```
Line 20 took 3253 cycles. Now if we replace line 20 with
```basic
20 IF A+C>B THEN D=A
```
it only takes 3099 cycles, which is 154 cycles faster.
### 🍎Recommendations
* Addition is faster than subtraction, so whenever possible use addition instead of subtraction, unless it means using a negative number for the second operand of the addition <sup>(*)</sup>.
* Because of that restriction, you probably won't be able to substitute subtraction with addition much for general calculations.
* However, you will always improve the speed if the substitution occurs within a comparison.
<sup>(*) Negative constants to avoid subtraction are bad: you would need to declare those before their positive counterparts and you would only win 20 cycles. And more than that, you would probably lose more cycles because your positive constants are declared later.</sup>
