Talk:RTS Trick: Difference between revisions
m (Reverted edits by 137.175.14.129 (talk) to last revision by Zzo38) |
m (→Indirect vs. RTS: JMP ($XXXX) is not re-entrant) |
||
Line 6: | Line 6: | ||
:* PHA, PHA, RTS requires less bytes than STA, STA, JMP (3 vs. 9). | :* PHA, PHA, RTS requires less bytes than STA, STA, JMP (3 vs. 9). | ||
:--[[User:MetalSlime|MetalSlime]] 00:08, 26 Jun 2009 (PDT) | :--[[User:MetalSlime|MetalSlime]] 00:08, 26 Jun 2009 (PDT) | ||
::JMP ($0200) isn't re-entrant. The RTS/RTI approach uses the stack therefore can be done by mainline and interrupt code freekly. [[User:Blargg|Blargg]] ([[User talk:Blargg|talk]]) 02:12, 23 January 2014 (MST) | |||
== Self-modifying == | == Self-modifying == |
Revision as of 09:12, 23 January 2014
Indirect vs. RTS
Is there an advantage over using JMP ($0200), where $0200 has been loaded from the same kind of jump-table? that's what I wonder, but I'm not gonna count up the cpu cycles needed for either method right now. --Memblers 20:21, 25 Jun 2009 (PDT)
- Not sure. Seems like a pick'em to me. Here are some things that come to mind:
- RTS Trick doesn't require any RAM.
- I personally think the RTS Trick is more readable. If I see a table of pointers in my (or somebody else's) code and they all have a "-1" after them, I immediately know their purpose and how they are used.
- PHA, PHA, RTS requires less bytes than STA, STA, JMP (3 vs. 9).
- --MetalSlime 00:08, 26 Jun 2009 (PDT)
Self-modifying
If you use self modifying code and assure that the table has to start at a page border (and store pointers to the routines, without the -1) then you can use a smaller and faster code:
tb_opcode_launcher_smc: ; bytes, cycles asl ; 1, 2 sta smc+2 ; 3, 4 smc: jmp (tb_opcode_rts_table) ; 3, 5 ; total 7 bytes and 11 cycles
tb_opcode_launcher: ; bytes, cycles asl ; 1, 2 tax ; 1, 2 lda tb_opcode_rts_table+1, x ; 3, 4 pha ; 1, 3 lda tb_opcode_rts_table, x ; 3, 4 pha ; 1, 3 rts ; 1, 6 ; total 11 bytes and 24 cycles
--212.8.208.194 (talk)
- Assuming that by
sta smc+2
you meantsta smc+1
because 6502 is little-endian. But if you're doing any sort of nontrivial work in the NMI or IRQ handler, you would need separate 7-byte self-modifying trampolines in RAM for the main, NMI, and possibly IRQ handlers. And with the NES's 2048 byte RAM, 21 bytes might be a lot, though it's still not as bad as it would be on the Atari 2600. --Tepples (talk) 11:18, 21 May 2013 (MDT)
Fast RTS Trick
You can make a bit faster RTS trick also if all of the jump targets are in one 256-byte page. Alternatively, you can align each jump target to a separate page and then not even use a lookup table. In addition, even when using the full addresses, I prefer to store the high bytes in one table and the low bytes in another table, instead of doing it together. This results in a smaller and faster code. --Zzo38 (talk) 16:45, 8 December 2013 (MST)