Over the last few weeks I’ve really been concentrating on shoring up the emulation quality of jsbeeb, mainly by adding test cases for all the undefined opcodes. Thankfully, there are some processor test suites out there and I’ve been able to get them running in jsbeeb as part of the continuous build. It now takes about 40 minutes to run all the tests, but I’m pretty darned sure jsbeeb has an accurate NMOS 6502 emulation.
Most recently I’ve been taking a glance over the BBC Master emulation, both in terms of hardware and the slightly different CMOS 65SC12 chip it used.
With a little help from my friends on the stardot forums I was able to get some tests run on real Masters and compare the output of the few instructions I couldn’t find info for. Most notably the unusual behaviour of the 6502 when in Binary Coded Decimal mode has mostly been fixed in the 65SC12.
To test the emulation quality, Rich Talbot-Watkins suggested I try out a Master-only demo that Tom Walker (of b-em fame) had written1. I fired it up and was both very impressed at the quality of the demo, and pleased that it worked first time!
Tom's impressive Master demo
Then I restarted it and the second time it ran, the sound went all crazy.
A bit of experimentation and it seemed about 30% of the time the sound would go nuts about a minute into the demo, squeaking and missing half the notes.
At first I suspected the sound chip emulation. The emulation was taken from my earlier Sega Master System emulator which happened to use the same chip, and I had recently updated the emulation there to fix up some other problems that Rich had noticed. I’d taken the information from the SMS Power emulation site, and it was pretty accurate, but maybe it didn’t match the chip in the Beeb.
It’s definitely the case that Tom’s b-em emulator and jsbeeb differ on how they latch the register values — maybe that was it? I ended up writing a “b-em” style sound system and running it in parallel with the jsbeeb one and comparing their outputs. They agreed 100%, even when the squeaks were there.
Scratch that as a source of error, then. It was a bit suspect anyway as the problem is intermittent and (somewhat) non-deterministic2.
While instrumenting the sound code I was able to notice a pattern: at the
point the squeaks started, two of the sound channels would be set to
values. This was a handy automatic diagnostic of the problem — it seems
there’s no normal time when the music ought to be this squeaky.
The sound code in the demo is based off of Rich’s “Blurp” music player. As
such he had a pretty good understanding of how it worked, and he was able to
determine what inputs to the sound code would make it choose to play a sound
$020, reading from a particular part of the frequency lookup
Armed with this information I added a breakpoint on reading that frequency table location. That allowed me to see when it was being read, but there wasn’t any obvious reason why: I needed to somehow track to the root cause.
Writing an emulator gives you amazing flexibility in debugging —
I added a circular buffer
of the last 256 values of all the registers after executing
each instruction. That meant when the breakpoint tripped I had a history of
how the code and the values in the
Y and flags registers had
Poring over the history I noticed something odd in the code compared to the version in Rich’s original: there was one comparison with zero, which was followed by a branch-if-less-than — a branch that would never be taken.
PC opcode A X Y ; my comment 8248 CPX #$00 20 02 00 824a BNE $8291 20 02 00 8291 LDX #$00 20 00 00 8293 CMP #$c0 20 00 00 ; >= c0? 8295 BCC $829b 20 00 00 ; no... 829b CMP #$00 20 00 00 ; >= 0 (er?) 829d BCC $82a3 20 00 00 ; no... carry is set
This is deeply suspicious — this wasn’t in Rich’s Blurp code and didn’t make
any sense. On a hunch I reloaded until I got a “working” version where the
sound didn’t squeak and lo and behold: The code at location
#$60. Something was overwriting the code!
Excitedly I breakpointed on memory writes to the
$60 byte at location
$829a and was rewarded with the culprit: a pretty screen-clearing routine
that shutter-style cleared the screen at the point in the demo where the sound
Hooray! Now I knew how the corruption was happening. Next I needed to work out why it was intermittent…
The clear routine writes zero bytes in a shutter-style pattern to the screen
memory. Nothing too unusual there. The screen memory ends at
$7fff in the
Master memory map, and then in the bank
#bfff there’s a selectable
16KB bank which can be ROM or one of four banks of RAM. The sound playing
routines live in one of the RAM banks at location
To apparently save some cycles somewhere, the screen routine doesn’t
explicitly check for overrunning the memory at
$8000: it writes past the end
in some cases. 100% reproducible disaster, you might think. But no: Tom’s
smarter than that and every time he’s about to write to the screen, he swaps
the RAM out of
$8000 and instead pages in some ROM — which of course makes
the writes harmless. Here’s a snippet of the code:
loop: LDA #$0f STA $fe30 ; page in ROM SEI ; disable IRQs LDA #$00 STA ($72),Y ; store zero CLI ; re-enable IRQs LDA $72 CLC ADC #$08 ; move to the next char STA $72 BNE loop ; and keep clearing
The interesting part is where interrupts are disabled: recall I said the sound
playing routine is in one of the RAM banks that gets paged in at
Well, the sound player is driven by interrupts. The main interrupt handler
(which lives down in unpaged memory) pages in the relevant RAM bank when it
realises it needs to handle sound (in response to a timer firing). Critically,
when it returns it doesn’t restore the page that was previously there.
To prevent this being a problem, Tom sensibly disables interrupts while he does his write. However he doesn’t disable interrupts between paging the ROM in and writing so there’s a tiny window of opportunity where, should an IRQ happen, the sound handling bank is left in just before the zero byte is written.
The window is tiny - there’s only 6 cycles between the point of no return
for the store, and the
SEI taking effect. Additionally, it’s only if the
clearing routine is just past the end of the screen when the IRQ happens that
it causes an issue.
Now we understand how, where and why. One last nagging thing to take care of — why is it intermittent? Further instrumentation gives the answer: the timer values that are used to set the interrupt frequency aren’t being deterministically set. Instead their values are loaded in the interrupt routine itself. That means the “start” time of the interrupts depends on how many cycles have passed since the last hard reset. I was able to use this to find the 6-cycle window: by externally varying the timer values at the point at which the first instruction of the demo executes I could find the values that caused it to go wrong.
b-em and other emulators don’t seem to see this issue — I believe as they don’t model the interrupts as precisely in all cases.
The only remaining mystery is why nobody has ever seen the issue on a real BBC Master. The marvellous Ed Spittles has run the demo a number of times on a real Master and never seen the issue. This leaves me feeling that perhaps I’m missing something, but it could just be an emulation artifact (maybe in how long the disc drive takes to load?) that makes the few timer values where it’s a problem more likely.
Any ideas on this gratefully accepted! For now, I’m happy enough with my explanation and can hopefully get on with some of the other things that need attending to.
Big thanks to Rich for putting up with my enormous nocturnal email rambles as I picked my way through all this in addition to giving fantastic advice on what the problem might be. Again, also thanks to Tom Walker for both the demo and for his excellent emulator which was the inspiration for jsbeeb.