On Thu, May 07, 1998 at 11:22:58AM -0400, Alex deVries wrote:
> So, I was compiling .99 last night, and my machine hung. I came in this
> morning, and found reams of errors on the screen, all of which were:
> Setting flush to zero for patch.
> Unimplemented exception at 0x004057f0
Still lacking the kernel math code for MIPS I implemented a cruel hack.
If we catch an unimplemented exception and flush-to-zero is cleared, we
set it and retry the instruction. If flush-to-zero is already set, the
problem is beyond what we can deal with ...
The only mysterious thing for me is right now why patch is using floating
point. Whatever, this error message is not directly related to the hang,
it's more telling people that fp results might be inaccurate or wrong.
Anybody with floating point experience and time to write this stuff?
What we need is basically a subset of a coprozessor emulation; there is
example code available in the NetBSD, OpenBSD and MACH MIPS assembler
code; the Sparc64 code is even written in C.
> This is 2.1.91 kernel I got from Ralf. Maybe this is useless, since I'm
> on my way to running .99 anyway.
I'm interested to hear if .99 hangs again for you. All I can say is
that I tested .99 by recompiling libc and that crashed ``No space left
on device'', so it must be good ;-)