* Copyright (c) 1988 by Sozobon, Limited.  Author: Johann Ruegg
*
* Permission is granted to anyone to use this software for any purpose
* on any computer system, and to redistribute it freely, with the
* following restrictions:
* 1) No charge may be made other than reasonable charges for reproduction.
* 2) Modified versions must be clearly marked as such.
* 3) The authors are not responsible for any harmful consequences
*    of using this software, even if they result from defects in it.
*
*
* MODIFICATIONS:
*
* D.W.Brooks	Mar 89	Implemented rounding, inlined one-call subroutines,
*			and several other speedups
*
* Rounding in the ambiguous case (tail byte = 0x80) uses round-to-even, as
* Knuth insists (TAOCP3SNA).
*
*	fpsub
*
	.globl	fpsub
	.globl	_fpsub
fpsub:
_fpsub:
	eor.b	#$80,11(sp)	* negate B, fall through (can produce -0.0)
*
*	fpadd
*
	.globl	fpadd
	.globl	_fpadd
fpadd:
_fpadd:
	move.l	d3,a1		* save d3
	move.l	d4,a2		* and d4
	move.l	4(sp),d0	* First arg
	move.l	8(sp),d1	* Second arg
	move.l	d0,d4		* Keeps the result's sign
	moveq.l	#$7f,d3

	move.b	d0,d2		* Compare signs only
	eor.b	d1,d2
	bpl	same_sign
* different signs
	and.b	d3,d0		* Force positive
	and.b	d3,d1

	move.l	d0,d2		* note d2 is copy of d0
	cmp.b	d0,d1		* Compare magnitudes
	bne	sk1
	cmp.l	d0,d1
sk1:
	ble	norev1
	move.l	d1,d0		* swap d0/d1...d0 was saved in d2...
	move.l	d2,d1
	move.b	d0,d2		* ...now d2.b is copy of new d0.b
	not.b	d4		* Invert sign
norev1:
* Begin unsigned subtract. Here we have (d0-d1), d0 >= d1, both positive.
* d2.b has result exp.  Subtract 0.0 works without needing a special case.

	move.b	d0,d3
	sub.b	d1,d3		* diff of exps (in low byte)

	clr.b	d0		* Extract mantissas
	clr.b	d1

	cmp.b	#25,d3
	bgt	out		* Nonsignificant addend: d0 is correct mantissa
	lsr.l	d3,d1		* Shift uses d3 mod 64
	sub.l	d1,d0

	beq	out0		* Zero: all done!
	bmi	outsub		* Already normalized
* normalize loop
nloop:
	subq.b	#1,d2
	ble	underfl
	add.l	d0,d0
	bpl	nloop

outsub:
	add.l	#$80,d0		* Force a rounding step.  This can't overflow.
				* (rum-ti-rum...can it?)
*
* Common exit point for checking rounding after adding 0x80.  If the tail
* was not 0x80, it's already the correct result.  Otherwise, the tail is now
* 0, and we implement the round-to-even.
*
outround:
	tst.b	d0
	bne	out
	and.w	#$FE00,d0
*
* Common exit point.  d0 has the mantissa, d2 the exponent and d4 the sign.
*
out:
	move.b	d2,d0
	and.b	#$80,d4
	or.b	d4,d0		* fix sign
*
* Come here when d0 has the result
*
out0:
	move.l	a1,d3		* restore
	move.l	a2,d4
	rts

* underflow
underfl:
	moveq.l	#0,d0
	bra	out0
*
* same signs: do add.
*
same_sign:
	and.b	d3,d0		* Force both positive
	and.b	d3,d1

	move.l	d0,d2
	cmp.b	d0,d1		* Compare magnitudes
	bne	sk2
	cmp.l	d0,d1
sk2:
	ble	norev2
	move.l	d1,d0		* Swap d0/d1...
	move.l	d2,d1
	move.b	d0,d2		* ...and d2.b is a copy of d0.b
norev2:
* Begin unsigned add: (d0+d1), d0 >= d1
	sub.b	d1,d0
	move.b	d0,d3		* diff of exps

	clr.b	d0		* Extract mantissas
	clr.b	d1

	cmp.b	#25,d3
	bge	out		* Nonsignificant addend: d0 has result mantissa
	lsr.l	d3,d1
	add.l	d1,d0
	bcc	around
	roxr.l	#1,d0		* Mantissa overflowed the word
	addq.b	#1,d2
	bmi	oflo		* Exponent too hot to handle

around:				* Round the fraction.
	add.l	#$80,d0		* This rounds unless d0.b was 0x80
	bcc	outround	* No overflow, but do final rounding
	roxr.l	#1,d0		* Actually loads 0x800000nn
	addq.b	#1,d2		* Compensate
	bpl	out		* Ok if no exponent overflow

oflo:
	moveq.l	#$ffffffff,d0	* Handle exponent overflow, maintaining sign
	moveq.l	#$7f,d2
	bra	out

	.end