On Tue, Jan 09, 2001 at 09:41:53PM +0100, Frank Wille wrote:
> Ignatios Souvatzis wrote:
> 
> > > article in the Amiga Format magazine. This mentioned some kind of bug
> > > in the 68060. They also had a program called NoBypass which apparently
> > > fixed this bug. Now I'm able to run the 68060 at almost full speed.
> > > Could NetBSD be affected in my case by the same bug ? The assembler
> > > program can be found on AmiNet in the 68060fix.lha archive.
> > 
> > Hm. I'll check what this about.
> 
> The bug will only occur after a very stupid combination of read
> and write instructions, while reading the same memory address
> twice, AFAIR. No assembler programmer and most good compilers will
> never create such a sequence, so it's not very likely that this is
> the reason of John's MMU problems, IMHO.
> 
> I had this bug though, one year ago, while porting Quake for
> AmigaOS/68k with the vbcc compiler. It happened on both, 060 Rev.1
> and 060 Rev.5 CPUs (CSPPC in both cases), although Motorola stated

Rev 5? what IS Rev 5? (can you look onto the chip, what mask revision it
claims? 

(Only 0, 1, 2 where ever documented, and even them only after much begging
that I did).

> that only Rev.1 CPUs could be affected. Changing bit 5 in the PCR
> fixed it.

That one. As I remember it, it is writing thhe memory location and reading it,
while triggering a couple of other forwarding paths through the pipelines
at the same time. Yes, PCR bit 5 is switched on by NetBSD already... without
that, newer gcc (2.7.2.3) could not compile itself "back then". I had to 
build another, last kernel on the 030 machine.

	-is