Bithacks: Difference between revisions
Runic Rain (talk | contribs) m (alternatives) |
Runic Rain (talk | contribs) (eh) |
||
Line 315: | Line 315: | ||
== Convert hex to decimal (16 bits) - with tables == | == Convert hex to decimal (16 bits) - with tables == | ||
''1069 bytes / 79 cycles'' | ''1069 bytes / 79 cycles'' | ||
<br> | |||
<u>inputs:</u> A | <u>inputs:</u> A | ||
<br> | <br> | ||
Line 361: | Line 362: | ||
== Convert hex to decimal (16 bits) - without tables == | == Convert hex to decimal (16 bits) - without tables == | ||
''161 bytes / 217 cycles'' | ''161 bytes / 217 cycles'' | ||
<br> | |||
<u>inputs:</u> A | <u>inputs:</u> A | ||
<br> | <br> | ||
Line 396: | Line 398: | ||
!bit #= !bit*2 | !bit #= !bit*2 | ||
endwhile | endwhile | ||
</pre> | |||
== DP As Extra Index Register == | |||
''12 bytes / 16 cycles'' | |||
<br> | |||
<u>inputs:</u> !offset, !pointer, !value | |||
<br> | |||
<u>outputs:</u> (none) | |||
<pre> | |||
; this one is more instructive than exact, in terms of cycles and bytes. | |||
; in general it's around +9 cycles to restore/set DP | |||
; inside a tight loop using this trick can save 2 cycles per usage or more | |||
LDA #!pointer | |||
TCD | |||
LDX #!offset | |||
LDA #!value | |||
STA $00,x ; write value to pointer + offset | |||
</pre> | |||
== RAM As Extra Index Register == | |||
''13 bytes / 20 cycles'' | |||
<br> | |||
<u>inputs:</u> !offset, !pointer, !value | |||
<br> | |||
<u>outputs:</u> (none) | |||
<pre> | |||
; this one is more instructive than exact, in terms of cycles and bytes. | |||
; in general it's around +7 cycles to setup long pointer bank in RAM (and it must be in DP) | |||
; slower code that aims to squeeze extra speed out of having the extra index available | |||
LDA #!pointer | |||
STA !ptr+1 | |||
LDX #!offset | |||
LDA #!value | |||
STA [!ptr],x ; write value to pointer + offset | |||
</pre> | </pre> | ||
[[Category:ASM]] | [[Category:ASM]] |
Revision as of 00:30, 23 September 2024
Bithacks are optimization tricks that utilize information in bits and bit manipulation
to accomplish their tasks. Usually they work in a slightly non-obvious way, (the most famous being the fast inverse sqrt), and bit manipulation in general is harder on the 65c816. To that end, here is a collection of some useful tricks.
Note: cycle counts are intended to be a worst case measure.
See also: Useful Code Snippets
Math Bithacks
Signed Division By 2
7 bytes / 8 cycles
inputs: A
outputs: A
CMP #$80 ROR BPL + ADC #$00 +
note: Rounds toward zero.
Arithmetic Shift Right
3 bytes / 4 cycles
inputs: A
outputs: A
CMP #$80 ROR
note: This is similar to division by 2, but rounds toward negative infinity.
Arithmetic Shift Right, multiple steps
6+n bytes / 6+2n cycles
inputs: A
outputs: A
; signed division by two, n times macro ASR_multi(n) LSR #<n> BIT.b #$80>><n> BEQ ?positive ORA.b #$FF00>><n> ; sign extension ?positive: endmacro ; -1 cycle and +n bytes, but must have N flag set before use macro ASR_multi(n) BMI ?negative LSR #<n> BRA ?end ?negative: LSR #<n> ORA.b #$FF00>><n> ; sign extension ?end: endmacro
Absolute Value
5 bytes / 6 cycles
inputs: A, (N Flag)
outputs: A
macro abs() BPL ?plus EOR #$FF INC ?plus: ; only 3 cycles if branch taken endmacro
Absolute Value (SEC)
4 bytes / 4 cycles
inputs: A, (Carry Set)
outputs: A
; compared to the branching version this is 1 byte smaller ; it's either 2 cycles slower/faster depending on branch taken EOR #$7F ; SEC ; the instant you add this in it becomes worse than the branching version SBC #$7F
Magnitude/Extents Check
~7 bytes / 12 cycles
inputs: A
outputs: (none)
; asks "Is [A] on the zero-side of value [X] or the far side?" ; good for magnitude checks, smaller *AND* faster than alternatives ; NOTE: in the event that it is exactly [X] it will have that value at branch ; doesn't need to be an indexed CMP but is most useful this way ; this can be used to combine the BPL and BMI checks for both signs into one SEC : SBC Extents,x EOR Extents,x BMI .zero_side .far_side: ; do things .zero_side: ; do things Extents: db -$23, $23
Sign Extend
8 bytes / 8 cycles
inputs: 8bit value in A (16-bit mode)
outputs: A
BIT #$0080 : BEQ + ; CMP #$0080 : BCC + ; alternative to above ORA #$FF00 ; on branch taken: additional -2 cycles +
Sign Extend (Branchless)
7 bytes / 8 cycles
inputs: 8bit value in A (16-bit mode)
outputs: A
; Value in A must have high byte clear EOR #$0080 SEC : SBC #$0080
Sign Extend (Unaligned)
11 bytes / 15 cycles
inputs: 8bit value in $10 (pulled as 16-bit)
outputs: A
; REP #$20 LDA $10-1 ; load $10 into A high, and garbage in low AND #$FF00 ; discard garbage BPL + ORA #$00FF + XBA
Clamp Signed (To Constants)
16 bytes/15 cycles
inputs: A
outputs: A
; clamp signed value in A to [min,max] if min/max are signed constants macro clamp_const(min,max) EOR #$80 CMP #$80^<min> : BCS ?+ LDA #$80^<min> ?+ CMP #$80^<max> : BCC ?+ LDA #$80^<max> ?+ EOR #$80 endmacro
Misc. Tricks
As this list grows tricks here will be consolidated into their own sections. Clever optimization tricks that aren't necessarily what someone might personally call a "bithack" are okay here as well!
XCN
12 bytes / 16 cycles
inputs: A
outputs: A
; eXchaNge Nibble without a LUT ASL : ADC #$00 ASL : ADC #$00 ASL : ADC #$00 ASL : ADC #$00
Clear Low Byte of Accumulator
1 byte / 2 cycles
inputs: (none)
outputs: A
; "Trashes" A but clears low byte TDC
Direction/Facing As Index
4 bytes / 6 cycles
inputs: A
outputs: A
; Ever wonder why facing flags are 0=right and 1=left? This is why. It's incredibly cheap. ; The input here is specifically a signed speed, or similar value. ASL ROL AND #$01
Check N Conditions True
n+7 bytes / 2n+7 cycles
inputs: A
outputs: A
; You can test for multiple conditions being true (7 conditions true, at least 5 conditions, etc.) by simply using a counter and rounding to the next power of 2 and test if that bit is set. ; You can also test for "Less than N True", "More than N", etc. with variations. ; This is almost more a coding technique, but it's super helpful, so worth pointing out. ; It can allow you to re-arrange branches of code as independent blocks among other useful things. ; You can also use any RAM instead of A at a small cost. ; Example Test For 5 True Conditions: !Next_Highest_Power_of_2 = $08 !N_True_Target = $05 LDA #!Next_Highest_Power_of_2!-!N_True_Target-1 ; here we set up our rounding, the -1 isn't strictly necessary *most* of the time %TestSomeCondition() BCC + ; here we're going to say our test just returns carry set on true (but it could directly INC inside the code as well) INC + ; ... repeat the above 5 times for different tests N_True_Test: INC ; replace our -1 to bring us up to a full power of 2 if we had enough True AND #!Next_Highest_Power_of_2 BEQ .false .true: ; N Tests were True .false: ; Not exactly N tests were true
Skip Dead Code
1-2 bytes / 2-3 cycles
inputs: (none)
outputs: (none)
; If you need to skip just one byte of dead code (due to a hijack or whatever reason) you can use: NOP ; 1 byte, 2 cycles ; But if you need to skip two bytes the most efficient is: ; NOTE: many times WDM is used as a breakpoint for debugging so only do this as a final pass to speed up your code! WDM ; 2 bytes, 2 cycles ; Finally, if you need to skip a large amount of dead code you can use BRA/JMP instead ; JMP is as fast as BRA on the SNES CPU, but will be slightly slower on SA-1, and 1 cycle slower on SPC. So BRA is recommended ; (The extra byte used for JMP in this case doesn't matter) BRA + ; 2 bytes, 3 cycles ; dead code +
Check 3 Conditions
2 bytes / 2 cycles
inputs: A
outputs: (none)
; just the opcode as normal here (not counting the conditions), using any operand that's not immediate (#) ; it's worth noting that you can do up to 3 tests with a single opcode though! ; Just As A Reminder: the V & N flag are set by the *operand* to BIT not the result of the AND! BIT $00 BMI .bit7_set BVS .bit6_set BNE .bit5_set ; assuming #$20 is in $00 .bit7_set: .bit6_set: .bit5_set:
Combine Carry Flag
4 bytes / 8 cycles
inputs: (Flag, On Stack)
outputs: (Carry Flag)
; flag on stack via PHP (8-Bit A if this), etc. ; code that alters Carry Flag PLA : BCS + LSR +
Transfer Carry Flag To Overflow Flag
2 bytes / 2 cycles
inputs: (Carry Flag)
outputs: (Overflow Flag)
ADC #$7F ; #$7FFF for 16-bit
Convert hex to decimal (16 bits) - with tables
1069 bytes / 79 cycles
inputs: A
outputs: Y (lower four digits, one per nybble), !top_digit (ten-thousands digit)
pha and #$FF00 xba asl tax lda.l LowToDecimal+1,x lsr #2 and #$0007 sta !top_digit lda.l HighToDecimal,x sta !tmp pla and #$00FF asl tax lda.l LowToDecimal,x and #$03FF sed adc !tmp bcc + inc !top_digit + cld ... LowToDecimal: !i = 0 while !i < 256 dw floor(!i*256/10000)*$400+$!i ; not a typo !i #= !i+1 endwhile HighToDecimal: !i = 0 while !i < 256 dw $!{i}00 !i #= !i+1 endwhile
Convert hex to decimal (16 bits) - without tables
161 bytes / 217 cycles
inputs: A
outputs: Y (lower four digits, one per nybble), !top_digit (ten-thousands digit)
rep #$30 sed tax stz !top_digit and #$0007 ; bottom three bits are easy tay !bit = 16 while !bit <= 32768 txa bit.w #!bit beq + tya adc.w #$!bit ; not a typo tay if !bit == 8192 bcc + inc !top_digit endif if !bit >= 16384 lda !top_digit adc.w #!bit/10000 sta !top_digit if !bit < 32768 ; carry isn't used after 32768 clc endif endif + !bit #= !bit*2 endwhile
DP As Extra Index Register
12 bytes / 16 cycles
inputs: !offset, !pointer, !value
outputs: (none)
; this one is more instructive than exact, in terms of cycles and bytes. ; in general it's around +9 cycles to restore/set DP ; inside a tight loop using this trick can save 2 cycles per usage or more LDA #!pointer TCD LDX #!offset LDA #!value STA $00,x ; write value to pointer + offset
RAM As Extra Index Register
13 bytes / 20 cycles
inputs: !offset, !pointer, !value
outputs: (none)
; this one is more instructive than exact, in terms of cycles and bytes. ; in general it's around +7 cycles to setup long pointer bank in RAM (and it must be in DP) ; slower code that aims to squeeze extra speed out of having the extra index available LDA #!pointer STA !ptr+1 LDX #!offset LDA #!value STA [!ptr],x ; write value to pointer + offset