Cycle counting

From NESdev Wiki
Jump to navigationJump to search

It is often useful to delay a specific number of CPU cycles. Timing raster effects or generating PCM audio are some examples that might utilize this. This article outlines a few relevant techniques.

Instruction timings

You can use a comprehensive guide[1][2] as reference for instruction timings, but there are some rules-of-thumb[3] that can help remember most of them:

  • There is a minimum of 2 cycles per instruction.
  • Each byte of memory read or written adds 1 more cycle to the instruction. This includes fetching the instruction, and each byte of its operand, then any memory it references.
  • Indexed instructions which cross a page take 1 extra cycle to adjust the high byte of the effective address first.
  • Read-modify-write instructions perform a dummy write during the "modify" stage and thus take 1 extra cycle.
  • Instructions that push data onto the stack take 1 extra cycle.
  • Instructions that pop data from the stack take 2 extra cycles, since they also need to pre-increment the stack pointer.
  • "Extra" cycles often include an extra read or write that usually does not affect the outcome.


  • SEC - 2 cycles: 1 byte opcode, but has to wait for the 2-cycle minimum.
  • AND #imm - 2 cycles: opcode + operand = 2 bytes. Only affects registers.
  • LDA zp - 3 cycles: opcode + operand + byte fetched from zp.
  • STA abs - 4 cycles: opcode + 2 byte operand + byte written to abs.
  • LDA abs, X - 4 or 5 cycles: opcode + 2 byte operand + read from abs, but if the addition of the X index causes a page crossing it delays 1 extra cycle.
  • ASL zp - 5 cycles: opcode + operand + read from zp + write to zp, but it takes 1 extra cycle to modify the value.
  • LDA (indirect), Y - 5 or 6 cycles: opcode + operand + two reads from zp + read from indirect address. 1 extra cycle if a page is crossed.
  • STA (indirect), Y - 6 cycles: like LDA (indirect) but assumes the worst case of page crossing, so always spends 1 extra read cycle in case the page correction is being applied.
  • PHA - 3 cycles: opcode + stack write, but requires 1 extra cycle to perform the stack operation.
  • RTS - 6 cycles: opcode + two stack reads, but requires 2 extra cycles to perform the stack operations, plus 1 cycle to post-increment the program counter (to compensate for the off-by-1 address pushed by JSR).

Short delays

Here are few ways to create short delays without side effects. As the shortest instruction time is 2 cycles, it is not possible to delay 1 cycle on its own. NOP is essential for 2 cycle delays. 3 cycle delays always take 2 bytes, but usually have some compromise. More options become available as delays become longer.

  • NOP - 2 cycles, 1 byte, no side effects
  • JMP *+3 - 3 cycles, 3 bytes, no side effects
  • Bxx *+2 - 3 cycles, 2 bytes, no side effects but requires a known flag state (e.g. BCC if carry is known to be clear)
  • BIT zp - 3 cycles, 2 bytes, clobbers NVZ but preserves C, reads zp
  • IGN zp - 3 cycles, 2 bytes, only side effect is a read, unofficial instruction
  • NOP, NOP - 4 cycles, 2 bytes
  • NOP, ... - 5 cycles, 3 bytes (... = 3 cycle delay of choice)
  • CLV, BVC *+2 - 5 cycles, 3 bytes, clears V flag (can instead use C flag with CLC, BCC or SEC, BCS)
  • NOP, NOP, NOP - 6 cycles, 4 bytes
  • PHP, PLP - 7 cycles, 2 bytes, modifies 1 byte of stack
  • NOP, NOP, NOP, NOP - 8 cycles, 4 bytes
  • PHP, PLP, NOP - 9 cycles, 3 bytes, modifies 1 byte of stack
  • PHP, CMP zp, PLP - 10 cycles, 4 byes, modifies 1 byte of stack, reads zp
  • PHP, PLP, NOP, NOP - 11 cycles, 4 bytes, modifies 1 byte of stack
  • JSR, RTS - 12 cycles, 3 bytes (if taking advantage of an existing RTS elsewhere), modifies 2 bytes of stack


A clockslide[4] is a sequence of instructions that wastes a small constant amount of cycles plus one cycle per executed byte, no matter whether it's entered on an odd or even address. With official instructions, one can construct a clockslide from CMP instructions: ... C9 C9 C9 C9 C5 EA Disassemble from the start and you get CMP #$C9 CMP #$C9 CMP $EA (6 bytes, 7 cycles). Disassemble one byte in and you get CMP #$C9 CMP #$C5 NOP (5 bytes, 6 cycles). The entry point can be controlled with an indirect jump or the RTS Trick to precisely control raster effect or sample playback timing.

CMP has a side effect of destroying most of the flags, but you can substitute other instructions with the same size and timing that preserve whichever flags/registers you need at the end of the slide. There are unofficial instructions that can avoid altering any state: replace $C9 (CMP) with $89 or $80, which ignores an immediate operand, and replace $C5 with $04, $44, or $64, which ignore a read from the zero page.