Many years ago I was told that inserting NOPs into IIgs screen updates could actually make things faster, because of the way the different hardware components interact when you're doing a fast copy with shadowed memory (insert a NOP every 8 or 13 PEIs). Somebody attached a logic analyzer to a IIgs and timed it a few different ways, and determined that there is actually no value to adding NOPs:

https://retrocomputing.stackexchange.com/q/228/56

Is there a flaw in their test methodology, or was this actually a myth?

In article <8c43afb7-e8e6-45a5-b3bf-4c0847a2c731n@googlegroups.com>,
fadden <thefadden@gmail.com> wrote:
>Many years ago I was told that inserting NOPs into IIgs screen updates
>could actually make things faster, because of the way the different
>hardware components interact when you're doing a fast copy with shadowed
>memory (insert a NOP every 8 or 13 PEIs). Somebody attached a logic
>analyzer to a IIgs and timed it a few different ways, and determined
>that there is actually no value to adding NOPs:
>
>https://retrocomputing.stackexchange.com/q/228/56
>
>Is there a flaw in their test methodology, or was this actually a myth?

Here is the post:

---
Mythbusted!

With 2021 knowledge of how the FPI/CYA work, this simply didn't make sense. So
I plugged in a logic analyser and timed the Wolfenstein 3D code (NOP every 8
PEI's). I then modified the code to change it to a NOP every 13 PEI's, and
finally to remove the NOPs altogether.

These are the results:

NOP every 8 PEI's - 40 PEI's takes 0.179005ms (4.475125�s/PEI)
NOP every 13 PEI's - 39 PEI's takes 0.17446ms (4.473333�s/PEI)
39 PEI's (No NOPs) takes 0.173758ms (4.445533�s/PEI)
As you can see it is faster to simply issue PEI's without any NOP's.

- Ian Brumby
---

Ian has collected a bunch of data, but he's not fully shared enough to
prove the NOP doesn't help (in my opinion). He probably has that data,
and he's convinced, so he's almost certainly right. It is unfair
to put the burden of proof on Ian to prove the NOP doesn't help.

A detailed waveform of 40+ PEIs (more would be better--long enough to show a
pattern repeating several times) would probably be enough to show NOPs don't
help PEIs. To put the issue to bed, PEA, PHA/STA dloc traces should also be
taken (since each takes different number of clock cycles: 5, 4 vs. 6 for PEI).

Kent

On Thursday, September 9, 2021 at 10:29:33 AM UTC-7, Kent Dickey wrote:
> A detailed waveform of 40+ PEIs (more would be better--long enough to show a
> pattern repeating several times) would probably be enough to show NOPs don't
> help PEIs.

I was thinking more along the lines of a stripped-down game loop that ran for some large number of iterations, with timings measured by pointing a cell phone camera at it and counting elapsed video frames. Less precise, but it can be done without fancy hardware.

Mostly I was hoping somebody had already done this experiment. If they had reached the conclusion that NOPs do help, it then becomes a question of why different experiments reached different conclusions.

I wasn't able to find any evidence while researching the question initially, but before I moved the "accepted answer" check I figured I'd ping the world again to see if this was ever more than a rumor.