Google
 

Trailing-Edge - PDP-10 Archives - decus_20tap1_198111 - decus/20-0003/zerovr.mac
There is 1 other file named zerovr.mac in the archive. Click here to see a list.
	title zerovr - module to set floating point error results to 0

comment %

********************************************
WARNING: This routine uses KL instructions
********************************************

This routine is designed to be called at interrupt level to clear out
the result of an operation that overflowed, underflowed, etc.  It
sets the result to zero.  The easiest way to use it is:

  psidefine(6,1,zerovr);
  psidefine(7,1,zerovr);
  psienable(6);
  psienable(7);

This will result in all arithmetic errors turning into zero.  If you
want to print messages, count errors, etc., you should write your
own interrupt handler, which would then call this one to zero out the
result.  Note that your procedure should accept at least 5 parameters,
and pass them on to this when it calls it.  If you only want to worry
about floating point errors, consider just activating channel 7.

If the overhead involved in Pascal interrupt handling is too much for
you (e.g. you expect problems in an inner loop), you should instead
explicitly test for errors and do your own recovery in the program.
E.g.
   x := x/y;
   if aritherr then x := 0;

Aritherr can be defined as follows:
	  twoseg 400000
	  entry arithe
	  p==17
  arithe: setzm 1(p)	;assume no error
	  jfcl 17,.+2	;test for any error
	  popj p,	;none
	  aos 1(p)	;error - return true
	  popj p,
	  end

The jfcl 17, tests for all error bits.  You can of course choose to
test for only certain ones.  If use this method you should obviously
disable the usual Pascal arithmetic error trap.  Either run with
/NOCHECK or (*$C-*), which will disable all checking, or if you
prefer to get the usual bounds checking, etc., just do
    psidisable(6);  psidisable(7)
to turn off the arithmetic traps.

When you call psidefine to set up zerovr, zerovr will replace the
default Pascal arithmetic trap, so you do not need to (indeed should
not) disable channels 6 and 7 when you use zerovr.

For the ultimate in arithmetic error processing, see the module
CALFOR, which will allow you to call the Fortran arithmetic error
handler.
%

entry zerovr

	twoseg 400000

	a==1
	b==2
	c==3
	d==4
	e==5
	f==6
	p==17

;where called by the interrupt procedure, arguments are set up:
;	c - old PC
;	e - where the saved old PC is
;	f - where the saved AC's are

zerovr:	move a,-1(c)	;interrupted instruction
	move b,(c)	;next instruction
	ldb d,[point 9,b,8] ;op code
	cain d,(<jfcl>_-9) ;unless next instruction is jfcl
	jrst isjfcl	;in which case we may not want to do anything
jfclOK:	hrrzs (e)	;clear error bits from PC if not jfcl
;the place we have to clear depends upon the operation that blew up
;so the following code uses a table with one entry for each opcode,
;specifying a routine to use.
	ldb b,[point 9,a,8] ;op code of one that blew
	cail b,110	;if not 110-377
	caile b,377
	popj p,		;also ignore
	subi b,110	;get offset into dispatch table
	adjbp b,[point 6,zertab,5] ;get pointer to routine code
	ldb c,b		;c _ routine code
	jrst @zerdis(c)	;go to routine

;this code is called when the next instruction after the one that
;blew up is a jfcl.  The idea is that if the user is about to test
;for an error bit, we should let him handle it himself.  But we first
;check to be sure the error is one of the ones he is testing for.
isjfcl:	ldb d,[point 4,b,12] ;look at the ac field of the jfcl - bits to test
	lsh d,^D32	;align with bits in PC
	tdnn d,c	;do we have any of the bits the jfcl is going to test?
	jrst jfclOK	;no - ignore the jfcl
	and c,d		;clear out the other bits
	hllm c,(e)	;in saved PC, so he doesn't get any he can't handle
	popj p,		;yes - let the user handle it

;This is a dispatch table of routines to go to for various kinds of
;instruction.
zerdis:	clrn		;noop
	clra		;clear one AC
	clra2		;clear two AC's
	clra4		;clear four AC's
	clrm		;clear memory
	clrm2		;clear two memory loc's
	clram		;clear AC and memory
	clra2m		;clear two AC's and memory

;nothing
clrn:	popj p,

;clear one AC
clra:	ldb b,[point 4,a,12] ;ac field
	add b,f		;where it is stored
	setzm (b)	;clear ac
	popj p,

;clear AC and AC+1
clra2:  ldb b,[point 4,a,12] ;ac
	move c,b
	add c,f		;relocate to where stored
	setzm (c)
	addi b,1	;next ac
	andi b,17
	add b,f
	setzm (b)
	popj p,

;clear AC to AC+3, handling wraparounds
clra4:	ldb a,[point 4,a,12] ;ac
	move b,a
	add b,f
	setzm (b)
	movei c,3	;three more ac's
clra4l:	addi a,1
	move b,a
	andi b,17
	add b,f
	setzm (b)
	sojg c,clra4l
	popj p,

;clear effective address
clrm:	ldb b,[point 4,a,17] ;index reg
	tlz a,777757	;now clear op code,ac, and index from instruction
	caie b,0	;unless no index reg
	tlo a,c		;modify to use index c
	tlo a,(movei b,);movei
	add b,f		;relocate to addr of index reg used
	move c,(b)	;and get contents
	xct a		;do it - c now has addr of thing to be changed
	caig b,17
	jrst clrac	;if it is an ac, it is special
	setzm (b)	;no - just do it
	popj p,

clrac:	move c,b	;can't garbage b
	add c,f		;relocate into user ac's
	setzm (c)	;clear it
	popj p,

;clear effective address and E+1
clrm2:	pushj p,clrm	;do first one
	hrri b,1(b)	;now go to next
	caig b,17	;and do it
	jrst clrac
	setzm (b)
	popj p,

;clear AC and E
clram:	pushj p,clra
	pushj p,clrm
	popj p,

;clear AC, AC+1, and E
clra2m:	pushj p,clra2
	pushj p,clrm
	popj p,

;The following codes are used in the table of opcode data.  They are
;offsets in the dispatch table above. 
	n==0
	a==1
	a2==2
	a4==3
	m==4
	m2==5
	am==6
	a2m==7

;here we have a table of what to clear for each instruction
; that can cause overflow/underflow, etc.  It simply indicates where
; the results of that instruction go.  This table was made up when I
; was half asleep, so please report any errors in it.  Its accuracy
; is obviously crucial to the behavior of this code.
zertab:	byte (6)a2,a2,a2,a2,a2,a2	;110-115
	byte (6)a4,a4,n,a2,a,n		;116-123
	byte (6)n,m2,a,a,a2,am		;124-131
	byte (6)a,n,n,n,n,n		;132-137
	byte (6)a,a2,m,am,a,a		;140-145
	byte (6)m,am,a,a2,m,am		;146-153
	byte (6)a,a,m,am,a,a2		;154-161
	byte (6)m,am,a,a,m,am		;162-167
	byte (6)a,a2,m,am,a,a		;170-175
	byte (6)m,am,n,n,n,n		;176-203
	byte (6)n,n,n,n,a,a		;204-211
	byte (6)m,m,n,n,n,n		;212-217
	byte (6)a,a,m,am,a2,a2		;220-225
	byte (6)m,a2m,a2,a2,m,a2m	;226-233
	byte (6)a2,a2,m,a2m,n,n		;234-241
	z				;242-247
	z				;250-255
	z				;256-263
	byte (6)n,n,n,n,a,a		;264-271
	byte (6)m,am,a,a,m,am		;272-277
	z				;300-305
	z				;306-313
	z				;314-321
	z				;322-327
	z				;330-335
	byte (6)n,n,a,a,a,a		;336-343
	byte (6)a,a,a,a,m,m		;344-351
	byte (6)m,m,m,m,m,m		;352-357
	byte (6)a,a,a,a,a,a		;360-365
	byte (6)a,a,m,m,m,m		;366-373
	byte (6)m,m,m,m,n,n		;374-377

	end