Very interesting, I thought AHA-1542 cards were meant to be kind of crappy, but they do okay here. I have a couple kicking around. May have to plan to put one in my potential ISA 486DX50 build...... Might be an idea for my 386DX40 board as well, but I was tending toward keeping that "middle of the road"/normal. Think I maybe had a future domain 16bit ISA SCSI of some sort too. Might only have kinda lame and chonky clonky 300 and 80 MB drives..

@rasz_pl
The post is big and things get mixed up.
Repeated the test configuration details on top of every chart.
Gotek floppy drives are working.
Bus-mastering SCSI controllers are the only factor that causes the issue with BL3 L1 cache.
If IDE or non bus-mastering SCSI controllers are used all is good.
So far i tested on the 2 fastest 386 motherboards for obvious reasons. There is a chance both of them exhibit this problem.
In the coming days will check how other motherboards handle things.
ECS Panda 386V and Chaintech 340SCB come on top of the list so far.

@feipoa
Yes, missing characters can happen with several VGA and local storage controllers.
Especially if the two controllers are slot-neighbors. Moving them to specific slots away from each other sometimes can fix things.

So far i didn't have data corruption from the write-access issue.
In fact i thought i escaped it all together with "magical" Promise EIDE Pro, but the recent in-depth testing proved me wrong.
The solution is to switch to mechanical HDDs, or if the preference is to stay with CF cards just increase the ISA bus divider and loose a lot of system performance.

Scraped some pictures from the internet of Winbond based IDE controllers that i think i tried before. Cannot be 100%. Going by memory here.

ROM shadowing
None of the caching controllers handle their own ROM shadowing.
So far only some Adaptec, Bustek, Buslogic, Procom models do it.

Keyboard controllers
I can safely say that no stone was left unturned there.
The problem is clearly the local storage controllers.

Advansys ABP-5140/5142
The best SCSI controller indeed.
So far tested with 2 motherboards - see my note to rasz_pl above.
Will check few more soon. As well as SXL2 processors.
But regardless of what the outcome is, it is unlikely i will be moving to other motherboards.

SIIG i540/542
It is impressive indeed, but does not handle overclocking very well.
Starts to flake beyond 16.67MHz ISA bus.
Didn't test anything yet with DLC/SXL2 CPUs.
Only 386DX and BL3 for now. It is not included in the charts with BL3 CPU because it is unstable or not working at all at 25 and 27.5 MHz.
These tests are big churn. Adding more configurations is borderline unfeasible.
At most i will check how 2-3 other mobos/chipsests handle the BL3 L1 cache issue with SCSIs and check few key controllers of interest with SXL2, but thats it.

SIIG i540/542 and Advansys ABP-5140/5142 are basically the same design/implementation/BIOS_microcodes, yet they behave differently.
i540 is faster at 10MHz, ABP-5140/42 handles overclocking just fine.

@BitWrangler
I always thought that the late AHA-1542 models such as CF/CP are some of the best controllers in their time.
Future Domain in the other hand failed to impress. They are definitely more trouble than the AHA CF/CP.

Jesus tits, that's a big pile of disk controllers. Most of the controllers I had back in the day were either winbond or UMC, with an occasional DTK PTI-217. You seem to have a lot of the more obscure ones.
I think I have a PTI-227B, which is also based on a winbond chip. Is it worth testing?
I also have one of these. It's meant for CD-ROMS. I think it's normally configured for a tertiary drive, but it might be possible to use it with a hard drive using the primary address.
https://www.amazon.com/CREATIVE-TECHNOLOGY-CT … D/dp/B00450I8U0

I think the notion I had about adaptecs being crappy might have come from early Linux SCSI documentation, where the authors were obviously preferring PCI solutions. However, there was in the mid 90s somewhat of a plague of crappy SCSI cards that came "free" with SCSI scanners and other peripheral types.

@anonymous coward
Convincing number of test samples is the main factor, otherwise we are only speculating.
PTI-227B. The B variant didint do well. It was my first attempt to find a counterpart to just 227.

@BitWrangler
Fact is that SCSI controllers are more pretencious in general, so i know what you mean.
But also, a byproduct of the examination was identifying 2-3 of them that are kind of just working and also very fast. So, that changes the story to a point.