Floating Point-The Results Are In!

[Steven M. Schultz]

Hi -

> From: Johnny Billquist <bqt at Update.UU.SE>
> I wouldn't be *that* surprised by these results. For instance, I believe
> that longs are implemented with FP. And I wouldn't be surprised if a few
> FP ops were sneaked in to compute some stuff that aren't immediately
> appearant.

	It is true that _some_ long arithmetic is done using FP.  The long
	divide is done that way (at least in 2BSD, I've not looked at V7 yet)
	because it is much much less code to convert the operands to FP, do
	the divide, and then convert the result back (the alternative is
	about two pages of code).  Different CPUs handle a fault during a
	double word push to the stack differently, this was a real difficult
	problem to track down and fix.  If during the FP instruction 
	"movfi   fr0,-(sp)" the stackpointer becomes invalid some PDP-11 CPUs
	handle the fault differently.  See 2.11BSD update #150 for the details.

	The C compiler itself did NOT generate FP unless the operands were
	explicitly FP (float or double).  Most C code was 'int' or 'char *'
	and no FP code was needed or used for that.

	FP instructions would be clustered together where the libc.a routines
	were loaded.  The 'ldiv' and 'lrem' routines would have several FP
	instructions close to each other but the rest of the program would
	have very few.  A program such as 'adb' would have a few FP instructions
	in the routines that display the FP registers.  Oh - there's a bug 
	dating back to V7 in adb.  The FP registers for a traced/running 
	process do not display correctly (using adb on a core file works fine).
	Fixed in 2.11 (see update #405) ;-)

	Steven Schultz


Received: (from major at localhost)
	by minnie.cs.adfa.oz.au (8.8.8/8.8.8) id GAA22952
	for pups-liszt; Mon, 11 May 1998 06:42:17 +1000 (EST)