On 4 Mar 2003, Christian Groessler wrote:

> when I try to boot a -current kernel, I get early in the boot process
> a kernel panic: (this is typed from the screen)
>
> Kernel Illegal instruction Trap
> trap type 2, code = 0x0, v = 0x0
> kernel program counter = 0x37ac

Me too, on a Quadra 630. I guess you're working on an '040 Mac, too?

> Using the program counter as a hint I disassembled the kernel
> executable and the problem seems to be in the m68k_make_fpu_idle_frame
> function.
>
> 0x37ac seems to be the "fnop" after the "fmovecrx #15,%fp0" in
> sys/arch/m68k/m68k/switch_subr.s. (maybe the fmovecrx is the problem?)

I bet! fmovecrx is emulated on the 68040. Not good.

> With the following change the kernel boots.

Taking a hint from fpsp, I tried this, instead...

Index: switch_subr.s
===================================================================
RCS file: /cvsroot/src/sys/arch/m68k/m68k/switch_subr.s,v
retrieving revision 1.2
diff -u -r1.2 switch_subr.s
--- switch_subr.s	2003/01/17 23:18:28	1.2
+++ switch_subr.s	2003/03/05 02:59:33
@@ -519,8 +519,8 @@
 	frestore %sp@		| Effectively `resets' the FPU
 	fnop

-	/* Loading '0.0' from the constant rom will change FPU to "idle". */
-	fmovecrx #15,%fp0
+	/* Loading '0.0' will change FPU to "idle". */
+	fmove.x	#0,%fp0
 	fnop

 	/* Save the resulting idle frame into the buffer */

With that, the emuspeed regression tests pass, and it can display
grayscale images with "xwud". (I disassembled "xwud", and it does
indeed use floating point.)

BTW, I also build the whole kernel without -msoft-float...

Index: Makefile.mac68k
===================================================================
RCS file: /cvsroot/src/sys/arch/mac68k/conf/Makefile.mac68k,v
retrieving revision 1.85
diff -u -r1.85 Makefile.mac68k
--- Makefile.mac68k	2002/11/26 18:51:15	1.85
+++ Makefile.mac68k	2003/03/05 03:21:54
@@ -32,7 +32,7 @@
 ## (2) compile settings
 ##
 CPPFLAGS+=	-Dmac68k
-CFLAGS+=	-msoft-float
+#CFLAGS+=	-msoft-float
 AFLAGS+=	-x assembler-with-cpp -traditional-cpp

 ##

as it seems silly to me to hamper the compiler that way, when the
kernel is already riddled with floating point assembler anyhow.

Frederick