### Author Topic: The Optimization Compilation  (Read 72781 times)

0 Members and 1 Guest are viewing this topic.

#### squidgetx

• Food.
• CoT Emeritus
• LV10 31337 u53r (Next: 2000)
• Posts: 1881
• Rating: +503/-17
• rawr.
##### The Optimization Compilation
« on: December 28, 2010, 12:17:53 pm »
Optimization Compilation: The Quick & Dirty Guide to Optimizing in Axe

This is intended to be a guide for every Axe programmer with general optimization

For demonstration purposes, I will always use the variable A. Of course, you can do these optimizations with any other var or memory address. Most opts are either in code boxes or on bullet points.

Part I: The BASIC Stuff
Many of the tricks we've learned from BASIC can also be applied to Axe. However, not all of them are applicable. Let's take a look:
Boolean Stuff: As with BASIC, we can optimize
Code: [Select]
If VAR?0 to simply
Code: [Select]
If VAR
However, be warned that this may not always work with compound statements (If A and B). Because the logic operators are bitwise operations, the statement "2 or 0" will return 0. This means that unless you are certain that A and B are always boolean values (1 or 0), you need to do If A?0 and (B?0) instead of If A and B

Another random thing is to ALWAYS CLOSE YOUR QUOTES AND PARENTHESES. Unlike BASIC, the store arrow does not close them for you. Unlike BASIC, leaving them out does NOT improve the size of your program; it is unrelated to optimization but I figured I'd just stick it here since it is a good coding habit and makes code more readable.

Pre-evaluating expressions: Especially in games that heavily reference arrays throughout a section of code, it is often good both for speed and memory to pre-evaluate expressions that do not change throughout the loop. Look at this code for drawing a scrolling 16x16 tilemapper with X and Y positions in pixels:
Code: [Select]
For(A,0,11)For(B,0,7)If {Y/8+B*16+(X/8)+A+L1}Pt-On(A*8-(X^8),B*8-(Y^8),{Y/8+B*32+(X/8)+A}*8+Pic0EndEndEndThere is a HUGE speed gain from simply preevaluating some of the expressions before entering the loop:
Code: [Select]
X/8->EY/8->FX^8->IY^8->JFor(A,0,7)For(B,0,11)If {F+B*16+E+A+L1}->CPt-On(A*8-I,B*8-J,C*8+Pic0EndEndEnd
Pixel Testing

Pixel testing can be a mean and nasty cycle stealer from many programs. But never fear, it can be optimized...a lot. Remember that we have access to the screen buffer in L6.

If you are pixel testing a constant pixel, like pxl-Test(20,20), you can more than halve the speed of this command with the following optimization:
Code: [Select]
{20*12+L6+2}^^re4 This optimization relies on the fact that the numbers can basically be pre-computed: use the following formula to derive the numbers you should use:
Code: [Select]
{Y*12+L6+(X/8)}e(X^8) So for another example, the command pxl-Test(8,1) becomes {12+L6}e1.

The speed gain from this is so great that you can even still save (although not as much) even with a variable Y value. How you treat the constant X value remains the same as before, but simply substitute in your variable Y value in the above code. So for example, pxl-Test(31,Y) becomes {Y*12+L6+3}e7.

Part II: General
Equality Checks:
Code: [Select]
If A=EXP optimizes to
Code: [Select]
!If A xor EXPNote that you should use !If A – EXP if it is an optimized addition/subtraction (See Optimized Math)
Code: [Select]
If A=EXP and (B=EXP) optimizes to
Code: [Select]
!If A-EXP + (B-EXP) where + is the 16bit 'or' operator.
Now, if you are checking the same variable for more than one possible expression, then it yields a greater optimization to do this:
Code: [Select]
If A=EXP1 or (A=EXP2) to
Code: [Select]
If inData(A,Data(EXP1,EXP2,0)) You just have to make sure that you take care of the 0 case first, since this will return a non-zero value if the variable=0 Also, as Quigibo pointed out, this only works with constant, 8bit values.

Simple Math Stuff
• If you can, always put the constants last in your expressions:CONST+A to A+CONST
• Use two byte vars over one byte vars or nibbles if you can afford the free RAM; it will be slightly smaller and faster. (but only by a little)
Note: As far as I can tell the speed difference is negligible. Unless you’re raycasting or something
• Avoid parentheses as much as possible when writing expressions. Use Axe's left-to-right order of operations to your advantage. This is in the documentation.
• A quick way to determine the sign of a value (better than >>0) is EXP//32768 It will return -1 if the value is negative, and 0 if the value is 0 or positive
• The new ++ -- arguments are a smaller and faster way of doing {EXP}+/-1->{EXP}. In the old form, EXP had to be evaluated twice. Now, it only needs to be evaluated once.
• One random trick is if you need to initialize a 1-byte variable to 0, the line and 0->{EXP} will be smaller than 0->{EXP}

Optimized Math

Some operations are hard-coded optimized versions that don’t use the usual arithmetic operations. A complete list of them can be found in the following spoiler which I stole from Runer112’s handy list of command speed/size
Spoiler For Optimized Math:

;Optimized Math
;-----------------------------------------------
p_Add0:            0 bytes      0 cycles      +0            Results in no compiled code

p_Add1:            1 byte      6 cycles      +1            Tied for the smallest way to change a value in Axe

p_Add2:            2 bytes      12 cycles      +2

p_Add3:            3 bytes      18 cycles      +3

p_Add254:         3 bytes      16 cycles      +254

p_Add255:         2 bytes      10 cycles      +255

p_Add256:         1 byte      4 cycles      +256            The absolute smallest and fastest way to change a value in Axe

p_Add257:         2 bytes      10 cycles      +257

p_Add258:         3 bytes      16 cycles      +258

p_Add510:         4 bytes      20 cycles      +510

p_Add511:         3 bytes      14 cycles      +511

p_Add512:         2 bytes      8 cycles      +512

p_Add513:         3 bytes      14 cycles      +513

p_Add514:         4 bytes      20 cycles      +514

p_Add767:         4 bytes      18 cycles      +767

p_Add768:         3 bytes      12 cycles      +768

p_Add769:         4 bytes      18 cycles      +769

p_Add1024:         4 bytes      16 cycles      +1024

p_Sub0:            0 bytes      0 cycles      -0            Results in no compiled code

p_Sub1:            1 byte      6 cycles      -1            Tied for the smallest way to change a value in Axe

p_Sub2:            2 bytes      12 cycles      -2

p_Sub3:            3 bytes      18 cycles      -3

p_Sub254:         3 bytes      16 cycles      -254

p_Sub255:         2 bytes      10 cycles      -255

p_Sub256:         1 byte      4 cycles      -256            Also the absolute smallest and fastest way to change a value in Axe

p_Sub257:         2 bytes      10 cycles      -257

p_Sub258:         3 bytes      16 cycles      -258

p_Sub510:         4 bytes      20 cycles      -510

p_Sub511:         3 bytes      14 cycles      -511

p_Sub512:         2 bytes      8 cycles      -512

p_Sub513:         3 bytes      14 cycles      -513

p_Sub514:         4 bytes      20 cycles      -514

p_Sub767:         4 bytes      18 cycles      -767

p_Sub768:         3 bytes      12 cycles      -768

p_Sub769:         4 bytes      18 cycles      -769

p_Sub1024:         4 bytes      16 cycles      -1024

p_Mul0:            3 bytes      10 cycles      *0            Same as loading the constant 0

p_Mul1:            0 bytes      0 cycles      *1            Results in no compiled code

p_MulN1:         6 bytes      24 cycles      *?1            Same as p_IntNeg

p_Mul2:            1 byte      11 cycles      *2            Tied for the smallest way to change a value in Axe

p_Mul3:            4 bytes      30 cycles      *3

p_Mul4:            2 bytes      22 cycles      *4

p_Mul5:            5 bytes      41 cycles      *5

p_Mul6:            5 bytes      41 cycles      *6

p_Mul7:            6 bytes      52 cycles      *7

p_Mul8:            3 bytes      33 cycles      *8

p_Mul9:            6 bytes      52 cycles      *9

p_Mul10:         6 bytes      52 cycles      *10

p_Mul12:         6 bytes      52 cycles      *12

p_Mul16:         4 bytes      44 cycles      *16

p_Mul32:         5 bytes      55 cycles      *32

p_Mul64:         5 bytes      144 cycles      *64

p_Mul128:         5 bytes      170 cycles      *128

p_Mul255:         6 bytes      31 cycles      *255

p_Mul256:         3 bytes      11 cycles      *256

p_Mul257:         3 bytes      12 cycles      *257

p_Mul258:         4 bytes      23 cycles      *258

p_Mul260:         5 bytes      34 cycles      *260

p_Mul264:         6 bytes      45 cycles      *264

p_Mul512:         4 bytes      22 cycles      *512

p_Mul513:         6 bytes      37 cycles      *513

p_Mul514:         4 bytes      23 cycles      *514

p_Mul516:         5 bytes      34 cycles      *516

p_Mul520:         6 bytes      45 cycles      *520

p_Mul768:         6 bytes      23 cycles      *768

p_Mul1024:         5 bytes      33 cycles      *1024

p_Mul1028:         5 bytes      34 cycles      *1028

p_Mul1032:         6 bytes      45 cycles      *1032

p_Mul2048:         6 bytes      44 cycles      *2048

p_Mul2056:         6 bytes      45 cycles      *2056

p_Mul4096:         5 bytes      290 cycles      *4096

p_Mul8192:         5 bytes      314 cycles      *8192

p_Mul16384:         5 bytes      338 cycles      *16384

p_Mul32768:         6 bytes      24 cycles      *32768

p_Mul65535:         6 bytes      24 cycles      *65535            Same as p_MulN1

p_Div0:            3 bytes      10 cycles      /0            Same as loading the constant 65535

p_Div1:            0 bytes      0 cycles      /1            Results in no compiled code

p_Div2:            4 bytes      16 cycles      /2

p_Div10:         3 bytes      ~1896 cycles      /10            n*3+1878 cycles, n=number of set bits in result

p_Div128:         5 bytes      27 cycles      /128

p_Div256:         3 bytes      11 cycles      /256

p_Div512:         5 bytes      19 cycles      /512

p_Div32768:         5 bytes      27 cycles      /32768

p_SDiv0:         3 bytes      10 cycles      //0            **NOT MATHEMATICALLY CORRECT** Same as loading the constant 65535

p_SDiv1:         0 bytes      0 cycles      //1            Results in no compiled code

p_SDiv2:         4 bytes      16 cycles      //2

p_SDiv64:         6 bytes      38 cycles      //64

p_SDiv128:         4 bytes      23 cycles      //128

p_SDiv256:         5 bytes      20 cycles      //256

p_SDiv512:         6 bytes      38 cycles      //512

p_SDiv16384:         6 bytes      38 cycles      //16384

p_SDiv32768:         3 bytes      26 cycles      //32768

p_Mod1:            3 bytes      10 cycles      ^1            Same as loading the constant 0

p_Mod2:            5 bytes      20 cycles      ^2

p_Mod4:            6 bytes      22 cycles      ^4

p_Mod8:            6 bytes      22 cycles      ^8

p_Mod16:         6 bytes      22 cycles      ^16

p_Mod32:         6 bytes      22 cycles      ^32

p_Mod64:         6 bytes      22 cycles      ^64

p_Mod128:         4 bytes      15 cycles      ^128

p_Mod256:         2 bytes      7 cycles      ^256

p_Mod512:         4 bytes      15 cycles      ^512

p_Mod1024:         4 bytes      15 cycles      ^1024

p_Mod2048:         4 bytes      15 cycles      ^2048

p_Mod4096:         4 bytes      15 cycles      ^4096

p_Mod8192:         4 bytes      15 cycles      ^8192

p_Mod16384:         4 bytes      15 cycles      ^16384

p_Mod32768:         2 bytes      8 cycles      ^32768

p_EQN512:         9 bytes      44 cycles      =?512

p_EQN256:         8 bytes      40 cycles      =?256

p_EQN2:            8 bytes      40 cycles      =?2

p_EQN1:            7 bytes      36 cycles      =?1

p_EQ0:            7 bytes      36 cycles      =0

p_EQ1:            7 bytes      ~29 cycles      =1            24 cycles if true, 34 cycles if false

p_EQ2:            8 bytes      ~33 cycles      =2            28 cycles if true, 38 cycles if false

p_EQ256:         8 bytes      40 cycles      =256

p_EQ512:         9 bytes      44 cycles      =512

p_NEN512:         9 bytes      ~31 cycles      ??512            33 cycles if true, 28 cycles if false

p_NEN256:         8 bytes      ~27 cycles      ??256            29 cycles if true, 24 cycles if false

p_NEN2:            8 bytes      40 cycles      ??2

p_NEN1:            7 bytes      36 cycles      ??1

p_NE0:            7 bytes      ~23 cycles      ?0            25 cycles if true, 20 cycles if false

p_NE1:            8 bytes      ~27 cycles      ?1            29 cycles if true, 24 cycles if false

p_NE2:            9 bytes      ~31 cycles      ?2            33 cycles if true, 28 cycles if false

p_NE256:         8 bytes      ~27 cycles      ?1            29 cycles if true, 24 cycles if false

p_NE512:         9 bytes      ~31 cycles      ?2            33 cycles if true, 28 cycles if false

p_GE0:            3 bytes      10 cycles      ?0            Same as loading the constant 1

p_GT65535:         3 bytes      10 cycles      >65535            Same as loading the constant 0

p_LE65535:         3 bytes      10 cycles      ?65535            Same as loading the constant 1

p_LT0:            3 bytes      10 cycles      <0            Same as loading the constant 0

p_GE1:            7 bytes      ~23 cycles      ?1            25 cycles if true, 20 cycles if false

p_GT0:            7 bytes      ~23 cycles      >0            25 cycles if true, 20 cycles if false

p_LE0:            7 bytes      36 cycles      ?0

p_LT1:            7 bytes      36 cycles      <1

p_SGE0:            4 bytes      32 cycles      ??0

p_SLT0:            5 bytes      27 cycles      <<0

p_GetBit0:         5 bytes      27 cycles      ee0

p_GetBit1:         6 bytes      38 cycles      ee1

p_GetBit2:         7 bytes      37 cycles      ee2

p_GetBit3:         7 bytes      37 cycles      ee3

p_GetBit4:         7 bytes      37 cycles      ee4

p_GetBit5:         7 bytes      37 cycles      ee5

p_GetBit6:         7 bytes      26 cycles      ee6

p_GetBit7:         6 bytes      22 cycles      ee7

p_GetBit8:         5 bytes      27 cycles      ee8      e0

p_GetBit9:         6 bytes      38 cycles      ee9      e1

p_GetBit10:         7 bytes      37 cycles      ee10      e2

p_GetBit11:         7 bytes      37 cycles      ee11      e3

p_GetBit12:         7 bytes      37 cycles      ee12      e4

p_GetBit13:         7 bytes      37 cycles      ee13      e5

p_GetBit14:         7 bytes      26 cycles      ee14      e6

p_GetBit15:         5 bytes      20 cycles      ee15      e7

Part III: "HL is the Ans of Axe" -Runer112

Like BASIC, Axe also has an Ans-counterpart, a register called HL. It is written to every time you "load" an expression or constant. This yields a multitude of
small optimizations:
-Duplicate Arguments:
Code: [Select]
Text(0,0,PTR) optimizes to
Code: [Select]
Text(0,,PTR)See how that works? When Axe goes to parse the second argument of Text(, it would normally load the second argument into HL and then use that as the Y position of the text. However, in the optimized piece it doesn't find anything to load for the second argument, so it just uses what was already in HL, namely, 0.

This, extended, yields further optimization opportunities:
-More Omitted Arguments: Not only does loading arguments in commands write to HL, so does loading arguments in normal math and control structure operations. This means that we can omit lots of things, saving more space and speed. For example...
Code: [Select]
If A>3: 1?B becomes
Code: [Select]
If A>3: ?BBut wait! What if you want to use this optimization on an equality check? Normally, If A=1:?B would work, but if we wanted to optimize the equality check as well, it doesn't work quite right. !If A-1 returns zero if true (or rather, if false), so you might then think that it's probably best to skip this particular optimization and go
Code: [Select]
!If A-1:1?B However, although it may seem counterintuitive, the following is actually smaller and faster:
Code: [Select]
!If A-1:+1?B Similarly,
Code: [Select]
If EXP: -1?B:End(with a minus sign) is preferable to
Code: [Select]
If EXP:0?B:End

Code: [Select]
0?B: 0?A becomes
Code: [Select]
 0?B?Anote: When initializing variables that are 2 or less apart, it also yields further optimization to do this:
Code: [Select]
0?A: 1?B: 3?C to
Code: [Select]
0?A+1?B+2?CAn example of HL abuse in a subroutine:
Code: [Select]
sub(LBL,EXP).....Lbl LBL: EXP*2?{L1} to
Code: [Select]
sub(LBL,EXP)....Lbl LBL:*2?{L1]
-Here's another abuse of HL, creating the fastest and smallest loop structure (by Runer 112)
It will execute the loop n times, with A starting at n-1 and decreasing down to 0:

Code: [Select]
nWhile   -1?A  ;Code  AEnd

Moving around a sprite
Check this out: again courtesy of Runer112: (comments by me)
Code: [Select]
:.SMILE:[004466000000817E]->Pic1:DiagnosticOff:0->X->Y:Repeat getkey(15):ClrDraw:getKey(3)-getKey(2)+X //check getkeys for X. But we're not going to store the value just yet...:!If +1                         //First, check if X is negative one; that is, we'll check if X+1=0:+1                             //if so, add one to make the value in HL 1 (!If statements return 0 if true):End:-1                             //subtract 1 since we added 1 earlier:Pt-On(min(,88)?X,getKey(1)-getKey(4)+Y+(=?1)min(,56)?Y,Pic1)    //now we'll use the smallest value of either HL or 88 as the X position. For Y, we'll first handle keypresses, then add 1 (the boolean value of y=-1) if Y=-1. Then we'll use the smallest value of either HL or 56 as the Y position.:DispGraph:End
I can't post every possible abuse of HL here: there are so many ways to use it. Study these examples to see how they work and you can apply it in your own programs.

Much of HL abuse has been obsoleted by Axe's peephole optimizer. However, you can still make use of some of these tricks. While the weird syntax is not always necessary, remember to arrange your code and operations such that the peephole optimizer can target it. I left the section somewhat intact so that you can look through it to understand the concept; the optimizer is useless if your code isn’t written intelligently.

Part IV: Subroutines
Subroutines probably are capable of saving the most space than any other type of optimization. And they are easy to use, too. There are only a couple rules of thumb to follow:
1. If you are rewriting a section of code more than once (and it is more than just one command), best put it into a subroutine.
2. If you aren't, then don't put it in a subroutine.

Routine calls are around 3-5 bytes, and an additional 3 bytes for every argument you load. Using the recursive subroutine feature that saves your arguments (sub(LBLr....) costs you 15 bytes per argument.

Tail Call Optimization
This is lesser known and lesser used, but it is still worth sticking here I guess: If you are calling subroutines from a subroutine as the last line, then you can use a process known as Tail Call Optimization to change this:
Code: [Select]
 Lbl A : Stuff : sub(B) : Return to this:
Code: [Select]
 Lbl A: Stuff : Goto B  The Return is not needed because you end up "stealing" subroutine B's return instead of having to return to A and then return again to the main program.

Part V: Speed over Size
These are all optimizations for aggressive speed gain at the expense of size.

Short Circuit Evaluation:
In most cases, it yields a (rather impressive) speed gain to change
Code: [Select]
If EXP1 and EXP2 to
Code: [Select]
If EXP1 : If EXP2 Make sure to have EXP1 (the outside If block) be the expression that is less likely to be true to gain the most speed.

Factoring Constants (and powers of 2); When multiplying and dividing by multiples of powers of 2, it yields an optimization to factor the number first.
Code: [Select]
EXP*96 to
Code: [Select]
EXP*32*3
With powers of two, this is only advantageous with *64 and *128 (Check the auto-optimizations list for the list of numbers that this applies to).

Part VI: Miscellaneous
All the stuff I couldn't fit into another category...

• If you need to draw horizontal or vertical lines, use the Rect() function with a width or height of 1 instead: Line(0,,20,0) to Rect(0,,20,1) This improves both speed and size
• Use ElseIf and DS<( if you can. That's why they're there. They are surprisingly useful.
• The Text() and Bitmap() commands are slow because they are TI's routines. Avoid them if possible. (I know, it's hard not to use Text(...but do your best ).
• Note about the Fill() command: If you are initializing less than 10 bytes of data, it is better to just use ->{}r 5 times instead.
Also, a little-known fact regarding printing text at constant coordinates:
[/list]
Code: [Select]
.Coordinate=Y*256+XText(30*256+20)Text "Stuff"Is 7 bytes smaller than:
Code: [Select]
Text(20,30,"Stuff")The same applies for text drawn to the home screen:
Code: [Select]
.Coordinate=X*256+YOutput(20*256+30)Disp "Stuff"Is smaller than:
Code: [Select]
Output(20,30,"Stuff")
By even more, 8 bytes.
• The new DispGraphClrDraw command is the same as DispGraph:ClrDraw but twice as fast. Use it.
However, remember each new command you use in Axe creates its own subroutine in your program. Thus, (for size reasons) it’s best to avoid using Rect if you've already used Line and Pt-On instead of Pt-Change, (and preinvert the sprite).
Quote from: Quigibo
One major optimization that usually gets ignored is recycling large axe commands.  Axe is not like BASIC and so each command needs its own subroutine to add to your program.  For instance, lets say you use Pt-On() and Pt-Mask() in your code.  Each one has to bring its own 100+ byte subroutine into your program.  But you can probably get away with having just a Pt-Mask routine, recycling it to act like Pt-On by simply adding a 2nd layer to your sprite which is only 8 bytes extra instead of 100ish.  Or you could do the opposite too and only have Pt-On() and Pt-Change() to manually change both buffers at once.  This generally reduces the overall size of the program by a lot when you use the routines only once or very rarely in your code.  And I don't mean calling it rarely, it could be the most used routine in your code, I just mean rarely appears in your source.

...And that's all I've got. Happy coding! Let me know if you have any questions, comments, additions, or corrections and I'll be happy to accommodate you This post will also be updated with any new major optimizations found.

Also special thanks to Quigibo and Runer112 who are the main contributors of all these optimizations

Oh and....666th POST
[/list]
« Last Edit: October 30, 2012, 03:30:31 pm by squidgetx »

#### Munchor

• LV13 Extreme Addict (Next: 9001)
• Posts: 6199
• Rating: +295/-121
• Code Recycler
##### Re: The Optimization Compilation
« Reply #1 on: December 28, 2010, 12:39:21 pm »
I took a while to read this and I am astonished. Very good work. I'm astonished because I learn't lots of new stuff! I knew the basic ones, but the others, more complex will really help me making my programs faster. Thanks a lot.

#### Binder News

• Posts: 785
• Rating: +46/-3
• Zombie of Tomorrow
##### Re: The Optimization Compilation
« Reply #2 on: December 28, 2010, 12:42:05 pm »
EPIC!!! I just HAD to bookmark this.
Spoiler For userbars:

Hacker-in-training!   Z80 Assembly Programmer     Axe Programmer
C++ H4X0R             Java Coder                           I <3 Python!

Perdidisti ludum     Cerebrum non habes

"We are humans first, no matter what."
"Fame is a vapor, popularity an accident, and riches take wings. Only one thing endures, and that is character."
Spoiler For Test Results:

#### Munchor

• LV13 Extreme Addict (Next: 9001)
• Posts: 6199
• Rating: +295/-121
• Code Recycler
##### Re: The Optimization Compilation
« Reply #3 on: December 28, 2010, 12:47:32 pm »
EPIC!!! I just HAD to bookmark this.

Great idea, I copied it to a .txt file, WHICH I WILL NOT release in any way, don't worry. I bookmarked too and rated up.

#### Happybobjr

• James Oldiges
• LV11 Super Veteran (Next: 3000)
• Posts: 2325
• Rating: +128/-20
• Howdy :)
##### Re: The Optimization Compilation
« Reply #4 on: December 28, 2010, 01:03:17 pm »
This is great.

Will this will be updated regularly when ever a new optimization is found?

Part VI: Miscellaneous
All the stuff I couldn't fit into another category...

If you need to draw horizontal or vertical lines, use the Rect() function with a width or height of 1 instead:
Code: [Select]
Line(0,,20,0) to
Code: [Select]
Rect(0,,20,1)
this has speed improvement as well as size right?
« Last Edit: December 28, 2010, 07:58:36 pm by happybobjr »
School: East Central High School

Axe: １.０.0
TI-84 +SE  ||| OS: 2.53 MP (patched) ||| Version: "M"
TI-Nspire    |||  Lent out, and never returned
____________________________________________________________

#### Runer112

• Project Author
• LV11 Super Veteran (Next: 3000)
• Posts: 2289
• Rating: +639/-31
##### Re: The Optimization Compilation
« Reply #5 on: December 28, 2010, 06:11:54 pm »
Part III: "HL is the Ans of Axe" -Runer112

My favorite optimization trick of all time

#### Ikkerens

• LV6 Super Member (Next: 500)
• Posts: 378
• Rating: +28/-9
• JavaScript Magician
##### Re: The Optimization Compilation
« Reply #6 on: December 29, 2010, 07:16:44 am »
I'm feeling slightly... astounded by this...
Remembering the code of splut, I'm missing ALOT of optimizations

Splut for Android [----------]
Paused/halted indefinitely, might be abandoned, our graphic designer quit and the rest of us simply doesn't have the time to work on it...

#### FinaleTI

• Believe in the pony that believes in you!
• CoT Emeritus
• LV10 31337 u53r (Next: 2000)
• Posts: 1830
• Rating: +121/-2
• Believe in the pony that believes in you!
##### Re: The Optimization Compilation
« Reply #7 on: December 29, 2010, 12:03:40 pm »
Now, if you are checking the same variable for more than one possible expression, then it yields a greater optimization to do this:
Code: [Select]
If A=EXP1 or (B=EXP2) to
Code: [Select]
If inData(A,Data(EXP1,EXP2,0)) You just have to make sure that you take care of the 0 case first, since this will return a non-zero value if the variable=0
Shouldn't that first one be
Code: [Select]
If A=EXP1 or (A=EXP2)
Anyways, great guide. I'm definitely printing this out when I can.

Spoiler For Projects:

My projects haven't been worked on in a while, so they're all on hiatus for the time being. I do hope to eventually return to them in some form or another...

Spoiler For Pokemon TI:
Axe port of Pokemon Red/Blue to the 83+/84+ family. On hold.

Spoiler For Nostalgia:
My big personal project, an original RPG about dimensional travel and a few heroes tasked with saving the world.
Coding-wise, on hold, but I am re-working the story.

Spoiler For Finale's Super Insane Tunnel Pack of Doom:
I will be combining Blur and Collision Course into a single gamepack. On hold.

Spoiler For Nostalgia Origins: Sky's Story:
Prequel to Nostalgia. On hold, especially while the story is re-worked.

#### Munchor

• LV13 Extreme Addict (Next: 9001)
• Posts: 6199
• Rating: +295/-121
• Code Recycler
##### Re: The Optimization Compilation
« Reply #8 on: December 29, 2010, 12:12:38 pm »

#### AngelFish

• Is this my custom title?
• LV12 Extreme Poster (Next: 5000)
• Posts: 3242
• Rating: +270/-27
• I'm a Fishbot
##### Re: The Optimization Compilation
« Reply #9 on: December 29, 2010, 12:16:59 pm »
I'm feeling slightly... astounded by this...
Remembering the code of splut, I'm missing ALOT of optimizations

Alot? :p
« Last Edit: December 29, 2010, 12:17:21 pm by Qwerty.55 »
∂²Ψ    -(2m(V(x)-E)Ψ
---  = -------------
∂x²        ℏ²Ψ

#### Hot_Dog

• CoT Emeritus
• LV11 Super Veteran (Next: 3000)
• Posts: 2994
• Rating: +445/-10
##### Re: The Optimization Compilation
« Reply #10 on: December 29, 2010, 12:19:09 pm »
Wow, I can see you put a lot of work into this.  Great job!

#### Munchor

• LV13 Extreme Addict (Next: 9001)
• Posts: 6199
• Rating: +295/-121
• Code Recycler
##### Re: The Optimization Compilation
« Reply #11 on: December 29, 2010, 12:19:11 pm »

#### squidgetx

• Food.
• CoT Emeritus
• LV10 31337 u53r (Next: 2000)
• Posts: 1881
• Rating: +503/-17
• rawr.
##### Re: The Optimization Compilation
« Reply #12 on: December 29, 2010, 12:36:10 pm »
Lol, nice to see everyone enjoys it, and thanks for pointing out that small error FinaleTI.

#### Quigibo

• The Executioner
• CoT Emeritus
• LV11 Super Veteran (Next: 3000)
• Posts: 2031
• Rating: +1075/-24
##### Re: The Optimization Compilation
« Reply #13 on: December 29, 2010, 02:14:09 pm »
By the way, you can't have non-constant data as part of the Data() command.  And that whole inData() optimization only works for comparisons with numbers between 0-255.  You can optimize it to this always though:

Code: [Select]
!If A-EXP1 [16bit and] (A-EXP2)
___Axe_Parser___
Today the calculator, tomorrow the world!

#### Michael_Lee

• LV9 Veteran (Next: 1337)
• Posts: 1019
• Rating: +124/-9
##### Re: The Optimization Compilation
« Reply #14 on: December 29, 2010, 05:46:06 pm »
Ah...
I finally understand this:
Code: [Select]
0->XWhile +26  Do stuff...  X-2->XEnd
This is incredibly helpful - thanks!
My website: Currently boring.

Projects:
Axe Interpreter
> Core: Done
> Memory: Need write code to add constants.
> Graphics: Rewritten.  Needs to integrate sprites with constants.
> IO: GetKey done.  Need to add mostly homescreen IO stuff.
Croquette:
> Stomping bugs
> Internet version: On hold until I can make my website less boring/broken.