Re: [hatari-devel] Possible regression, Audio Sculpture

[Toni Wilen <twilen@xxxxxxxxxx>]

>> Hi,
>>
>> Well that was quick :) Yes, it works with the latest source!
>>
>> /Troed
>>
> 
> Possibly the fastest bug ever fixed after being reported on the ML :)

I finished the fix about 3 hours before it was reported :)

Reason for this stupid bug was error in my cpu tester when validating
68000 (and 68010) PC field in bus/address error field. It was basically
ignored completely.

Address errors were fixed yesterday. Read bus errors were fixed yearlier
today, prefetch bus errors are partially fixed (new logic needed, some
addressing modes advance internal PC early, some late..). Write bus
errors also need fixing but they most likely will get mostly
automatically fixed after prefetch bus errors get fixed. Many 68000 bus
error stack frame fields was mostly wrong until winter 2019 so there was
no danger of something breaking :)

btw, I also did softfloat FPU testing few weeks ago: basic FPU
instructions are confirmed fully correct (assuming tester actually
worked as I expected it to work!), only packed to extended conversions
have some tiny rounding issues (64-bit mantissa can have +-1 difference
in some situations. So close...)

68888x-only logarithmic and trigonometric functions also appear to work
correctly but obviously results are not 100% identical (and no one
expects them to be 100% identical). Not even 68040/060 FPSP return
identical results when compared to 6888x.

Only remaining problematic instructions seem to be FREM and FMOD, some
input values return very different results (even exponent part is very
different). FREM and FMOD seem to use original softfloat algorithm (x87
I guess) with few tweaks but Motorola apparently uses different
algorithm. This needs reimplementation using FPSP algorithm.