xref: /netbsd-src/sys/arch/m68k/fpsp/decbin.sa (revision ae9172d6cd9432a6a1a56760d86b32c57a66c39c)
1*	$NetBSD: decbin.sa,v 1.2 1994/10/26 07:48:59 cgd Exp $
2
3*	MOTOROLA MICROPROCESSOR & MEMORY TECHNOLOGY GROUP
4*	M68000 Hi-Performance Microprocessor Division
5*	M68040 Software Package
6*
7*	M68040 Software Package Copyright (c) 1993, 1994 Motorola Inc.
8*	All rights reserved.
9*
10*	THE SOFTWARE is provided on an "AS IS" basis and without warranty.
11*	To the maximum extent permitted by applicable law,
12*	MOTOROLA DISCLAIMS ALL WARRANTIES WHETHER EXPRESS OR IMPLIED,
13*	INCLUDING IMPLIED WARRANTIES OF MERCHANTABILITY OR FITNESS FOR A
14*	PARTICULAR PURPOSE and any warranty against infringement with
15*	regard to the SOFTWARE (INCLUDING ANY MODIFIED VERSIONS THEREOF)
16*	and any accompanying written materials.
17*
18*	To the maximum extent permitted by applicable law,
19*	IN NO EVENT SHALL MOTOROLA BE LIABLE FOR ANY DAMAGES WHATSOEVER
20*	(INCLUDING WITHOUT LIMITATION, DAMAGES FOR LOSS OF BUSINESS
21*	PROFITS, BUSINESS INTERRUPTION, LOSS OF BUSINESS INFORMATION, OR
22*	OTHER PECUNIARY LOSS) ARISING OF THE USE OR INABILITY TO USE THE
23*	SOFTWARE.  Motorola assumes no responsibility for the maintenance
24*	and support of the SOFTWARE.
25*
26*	You are hereby granted a copyright license to use, modify, and
27*	distribute the SOFTWARE so long as this entire notice is retained
28*	without alteration in any modified and/or redistributed versions,
29*	and that such modified versions are clearly identified as such.
30*	No licenses are granted by implication, estoppel or otherwise
31*	under any patents or trademarks of Motorola, Inc.
32
33*
34*	decbin.sa 3.3 12/19/90
35*
36*	Description: Converts normalized packed bcd value pointed to by
37*	register A6 to extended-precision value in FP0.
38*
39*	Input: Normalized packed bcd value in ETEMP(a6).
40*
41*	Output:	Exact floating-point representation of the packed bcd value.
42*
43*	Saves and Modifies: D2-D5
44*
45*	Speed: The program decbin takes ??? cycles to execute.
46*
47*	Object Size:
48*
49*	External Reference(s): None.
50*
51*	Algorithm:
52*	Expected is a normal bcd (i.e. non-exceptional; all inf, zero,
53*	and NaN operands are dispatched without entering this routine)
54*	value in 68881/882 format at location ETEMP(A6).
55*
56*	A1.	Convert the bcd exponent to binary by successive adds and muls.
57*	Set the sign according to SE. Subtract 16 to compensate
58*	for the mantissa which is to be interpreted as 17 integer
59*	digits, rather than 1 integer and 16 fraction digits.
60*	Note: this operation can never overflow.
61*
62*	A2. Convert the bcd mantissa to binary by successive
63*	adds and muls in FP0. Set the sign according to SM.
64*	The mantissa digits will be converted with the decimal point
65*	assumed following the least-significant digit.
66*	Note: this operation can never overflow.
67*
68*	A3. Count the number of leading/trailing zeros in the
69*	bcd string.  If SE is positive, count the leading zeros;
70*	if negative, count the trailing zeros.  Set the adjusted
71*	exponent equal to the exponent from A1 and the zero count
72*	added if SM = 1 and subtracted if SM = 0.  Scale the
73*	mantissa the equivalent of forcing in the bcd value:
74*
75*	SM = 0	a non-zero digit in the integer position
76*	SM = 1	a non-zero digit in Mant0, lsd of the fraction
77*
78*	this will insure that any value, regardless of its
79*	representation (ex. 0.1E2, 1E1, 10E0, 100E-1), is converted
80*	consistently.
81*
82*	A4. Calculate the factor 10^exp in FP1 using a table of
83*	10^(2^n) values.  To reduce the error in forming factors
84*	greater than 10^27, a directed rounding scheme is used with
85*	tables rounded to RN, RM, and RP, according to the table
86*	in the comments of the pwrten section.
87*
88*	A5. Form the final binary number by scaling the mantissa by
89*	the exponent factor.  This is done by multiplying the
90*	mantissa in FP0 by the factor in FP1 if the adjusted
91*	exponent sign is positive, and dividing FP0 by FP1 if
92*	it is negative.
93*
94*	Clean up and return.  Check if the final mul or div resulted
95*	in an inex2 exception.  If so, set inex1 in the fpsr and
96*	check if the inex1 exception is enabled.  If so, set d7 upper
97*	word to $0100.  This will signal unimp.sa that an enabled inex1
98*	exception occured.  Unimp will fix the stack.
99*
100
101DECBIN    IDNT    2,1 Motorola 040 Floating Point Software Package
102
103	section	8
104
105	include	fpsp.h
106
107*
108*	PTENRN, PTENRM, and PTENRP are arrays of powers of 10 rounded
109*	to nearest, minus, and plus, respectively.  The tables include
110*	10**{1,2,4,8,16,32,64,128,256,512,1024,2048,4096}.  No rounding
111*	is required until the power is greater than 27, however, all
112*	tables include the first 5 for ease of indexing.
113*
114	xref	PTENRN
115	xref	PTENRM
116	xref	PTENRP
117
118RTABLE	dc.b	0,0,0,0
119	dc.b	2,3,2,3
120	dc.b	2,3,3,2
121	dc.b	3,2,2,3
122
123	xdef	decbin
124	xdef	calc_e
125	xdef	pwrten
126	xdef	calc_m
127	xdef	norm
128	xdef	ap_st_z
129	xdef	ap_st_n
130*
131FNIBS	equ	7
132FSTRT	equ	0
133*
134ESTRT	equ	4
135EDIGITS equ	2
136*
137* Constants in single precision
138FZERO 	dc.l	$00000000
139FONE 	dc.l	$3F800000
140FTEN 	dc.l	$41200000
141
142TEN	equ	10
143
144*
145decbin:
146	fmove.l	#0,FPCR		;clr real fpcr
147	movem.l	d2-d5,-(a7)
148*
149* Calculate exponent:
150*  1. Copy bcd value in memory for use as a working copy.
151*  2. Calculate absolute value of exponent in d1 by mul and add.
152*  3. Correct for exponent sign.
153*  4. Subtract 16 to compensate for interpreting the mant as all integer digits.
154*     (i.e., all digits assumed left of the decimal point.)
155*
156* Register usage:
157*
158*  calc_e:
159*	(*)  d0: temp digit storage
160*	(*)  d1: accumulator for binary exponent
161*	(*)  d2: digit count
162*	(*)  d3: offset pointer
163*	( )  d4: first word of bcd
164*	( )  a0: pointer to working bcd value
165*	( )  a6: pointer to original bcd value
166*	(*)  FP_SCR1: working copy of original bcd value
167*	(*)  L_SCR1: copy of original exponent word
168*
169calc_e:
170	move.l	#EDIGITS,d2	;# of nibbles (digits) in fraction part
171	moveq.l	#ESTRT,d3	;counter to pick up digits
172	lea.l	FP_SCR1(a6),a0	;load tmp bcd storage address
173	move.l	ETEMP(a6),(a0)	;save input bcd value
174	move.l	ETEMP_HI(a6),4(a0) ;save words 2 and 3
175	move.l	ETEMP_LO(a6),8(a0) ;and work with these
176	move.l	(a0),d4		;get first word of bcd
177	clr.l	d1		;zero d1 for accumulator
178e_gd:
179	mulu.l	#TEN,d1		;mul partial product by one digit place
180	bfextu	d4{d3:4},d0	;get the digit and zero extend into d0
181	add.l	d0,d1		;d1 = d1 + d0
182	addq.b	#4,d3		;advance d3 to the next digit
183	dbf.w	d2,e_gd		;if we have used all 3 digits, exit loop
184	btst	#30,d4		;get SE
185	beq.b	e_pos		;don't negate if pos
186	neg.l	d1		;negate before subtracting
187e_pos:
188	sub.l	#16,d1		;sub to compensate for shift of mant
189	bge.b	e_save		;if still pos, do not neg
190	neg.l	d1		;now negative, make pos and set SE
191	or.l	#$40000000,d4	;set SE in d4,
192	or.l	#$40000000,(a0)	;and in working bcd
193e_save:
194	move.l	d1,L_SCR1(a6)	;save exp in memory
195*
196*
197* Calculate mantissa:
198*  1. Calculate absolute value of mantissa in fp0 by mul and add.
199*  2. Correct for mantissa sign.
200*     (i.e., all digits assumed left of the decimal point.)
201*
202* Register usage:
203*
204*  calc_m:
205*	(*)  d0: temp digit storage
206*	(*)  d1: lword counter
207*	(*)  d2: digit count
208*	(*)  d3: offset pointer
209*	( )  d4: words 2 and 3 of bcd
210*	( )  a0: pointer to working bcd value
211*	( )  a6: pointer to original bcd value
212*	(*) fp0: mantissa accumulator
213*	( )  FP_SCR1: working copy of original bcd value
214*	( )  L_SCR1: copy of original exponent word
215*
216calc_m:
217	moveq.l	#1,d1		;word counter, init to 1
218	fmove.s	FZERO,fp0	;accumulator
219*
220*
221*  Since the packed number has a long word between the first & second parts,
222*  get the integer digit then skip down & get the rest of the
223*  mantissa.  We will unroll the loop once.
224*
225	bfextu	(a0){28:4},d0	;integer part is ls digit in long word
226	fadd.b	d0,fp0		;add digit to sum in fp0
227*
228*
229*  Get the rest of the mantissa.
230*
231loadlw:
232	move.l	(a0,d1.L*4),d4	;load mantissa lonqword into d4
233	moveq.l	#FSTRT,d3	;counter to pick up digits
234	moveq.l	#FNIBS,d2	;reset number of digits per a0 ptr
235md2b:
236	fmul.s	FTEN,fp0	;fp0 = fp0 * 10
237	bfextu	d4{d3:4},d0	;get the digit and zero extend
238	fadd.b	d0,fp0		;fp0 = fp0 + digit
239*
240*
241*  If all the digits (8) in that long word have been converted (d2=0),
242*  then inc d1 (=2) to point to the next long word and reset d3 to 0
243*  to initialize the digit offset, and set d2 to 7 for the digit count;
244*  else continue with this long word.
245*
246	addq.b	#4,d3		;advance d3 to the next digit
247	dbf.w	d2,md2b		;check for last digit in this lw
248nextlw:
249	addq.l	#1,d1		;inc lw pointer in mantissa
250	cmp.l	#2,d1		;test for last lw
251	ble	loadlw		;if not, get last one
252
253*
254*  Check the sign of the mant and make the value in fp0 the same sign.
255*
256m_sign:
257	btst	#31,(a0)	;test sign of the mantissa
258	beq.b	ap_st_z		;if clear, go to append/strip zeros
259	fneg.x	fp0		;if set, negate fp0
260
261*
262* Append/strip zeros:
263*
264*  For adjusted exponents which have an absolute value greater than 27*,
265*  this routine calculates the amount needed to normalize the mantissa
266*  for the adjusted exponent.  That number is subtracted from the exp
267*  if the exp was positive, and added if it was negative.  The purpose
268*  of this is to reduce the value of the exponent and the possibility
269*  of error in calculation of pwrten.
270*
271*  1. Branch on the sign of the adjusted exponent.
272*  2p.(positive exp)
273*   2. Check M16 and the digits in lwords 2 and 3 in decending order.
274*   3. Add one for each zero encountered until a non-zero digit.
275*   4. Subtract the count from the exp.
276*   5. Check if the exp has crossed zero in #3 above; make the exp abs
277*	   and set SE.
278*	6. Multiply the mantissa by 10**count.
279*  2n.(negative exp)
280*   2. Check the digits in lwords 3 and 2 in decending order.
281*   3. Add one for each zero encountered until a non-zero digit.
282*   4. Add the count to the exp.
283*   5. Check if the exp has crossed zero in #3 above; clear SE.
284*   6. Divide the mantissa by 10**count.
285*
286*  *Why 27?  If the adjusted exponent is within -28 < expA < 28, than
287*   any adjustment due to append/strip zeros will drive the resultane
288*   exponent towards zero.  Since all pwrten constants with a power
289*   of 27 or less are exact, there is no need to use this routine to
290*   attempt to lessen the resultant exponent.
291*
292* Register usage:
293*
294*  ap_st_z:
295*	(*)  d0: temp digit storage
296*	(*)  d1: zero count
297*	(*)  d2: digit count
298*	(*)  d3: offset pointer
299*	( )  d4: first word of bcd
300*	(*)  d5: lword counter
301*	( )  a0: pointer to working bcd value
302*	( )  FP_SCR1: working copy of original bcd value
303*	( )  L_SCR1: copy of original exponent word
304*
305*
306* First check the absolute value of the exponent to see if this
307* routine is necessary.  If so, then check the sign of the exponent
308* and do append (+) or strip (-) zeros accordingly.
309* This section handles a positive adjusted exponent.
310*
311ap_st_z:
312	move.l	L_SCR1(a6),d1	;load expA for range test
313	cmp.l	#27,d1		;test is with 27
314	ble.w	pwrten		;if abs(expA) <28, skip ap/st zeros
315	btst	#30,(a0)	;check sign of exp
316	bne.b	ap_st_n		;if neg, go to neg side
317	clr.l	d1		;zero count reg
318	move.l	(a0),d4		;load lword 1 to d4
319	bfextu	d4{28:4},d0	;get M16 in d0
320	bne.b	ap_p_fx		;if M16 is non-zero, go fix exp
321	addq.l	#1,d1		;inc zero count
322	moveq.l	#1,d5		;init lword counter
323	move.l	(a0,d5.L*4),d4	;get lword 2 to d4
324	bne.b	ap_p_cl		;if lw 2 is zero, skip it
325	addq.l	#8,d1		;and inc count by 8
326	addq.l	#1,d5		;inc lword counter
327	move.l	(a0,d5.L*4),d4	;get lword 3 to d4
328ap_p_cl:
329	clr.l	d3		;init offset reg
330	moveq.l	#7,d2		;init digit counter
331ap_p_gd:
332	bfextu	d4{d3:4},d0	;get digit
333	bne.b	ap_p_fx		;if non-zero, go to fix exp
334	addq.l	#4,d3		;point to next digit
335	addq.l	#1,d1		;inc digit counter
336	dbf.w	d2,ap_p_gd	;get next digit
337ap_p_fx:
338	move.l	d1,d0		;copy counter to d2
339	move.l	L_SCR1(a6),d1	;get adjusted exp from memory
340	sub.l	d0,d1		;subtract count from exp
341	bge.b	ap_p_fm		;if still pos, go to pwrten
342	neg.l	d1		;now its neg; get abs
343	move.l	(a0),d4		;load lword 1 to d4
344	or.l	#$40000000,d4	; and set SE in d4
345	or.l	#$40000000,(a0)	; and in memory
346*
347* Calculate the mantissa multiplier to compensate for the striping of
348* zeros from the mantissa.
349*
350ap_p_fm:
351	move.l	#PTENRN,a1	;get address of power-of-ten table
352	clr.l	d3		;init table index
353	fmove.s	FONE,fp1	;init fp1 to 1
354	moveq.l	#3,d2		;init d2 to count bits in counter
355ap_p_el:
356	asr.l	#1,d0		;shift lsb into carry
357	bcc.b	ap_p_en		;if 1, mul fp1 by pwrten factor
358	fmul.x	(a1,d3),fp1	;mul by 10**(d3_bit_no)
359ap_p_en:
360	add.l	#12,d3		;inc d3 to next rtable entry
361	tst.l	d0		;check if d0 is zero
362	bne.b	ap_p_el		;if not, get next bit
363	fmul.x	fp1,fp0		;mul mantissa by 10**(no_bits_shifted)
364	bra.b	pwrten		;go calc pwrten
365*
366* This section handles a negative adjusted exponent.
367*
368ap_st_n:
369	clr.l	d1		;clr counter
370	moveq.l	#2,d5		;set up d5 to point to lword 3
371	move.l	(a0,d5.L*4),d4	;get lword 3
372	bne.b	ap_n_cl		;if not zero, check digits
373	sub.l	#1,d5		;dec d5 to point to lword 2
374	addq.l	#8,d1		;inc counter by 8
375	move.l	(a0,d5.L*4),d4	;get lword 2
376ap_n_cl:
377	move.l	#28,d3		;point to last digit
378	moveq.l	#7,d2		;init digit counter
379ap_n_gd:
380	bfextu	d4{d3:4},d0	;get digit
381	bne.b	ap_n_fx		;if non-zero, go to exp fix
382	subq.l	#4,d3		;point to previous digit
383	addq.l	#1,d1		;inc digit counter
384	dbf.w	d2,ap_n_gd	;get next digit
385ap_n_fx:
386	move.l	d1,d0		;copy counter to d0
387	move.l	L_SCR1(a6),d1	;get adjusted exp from memory
388	sub.l	d0,d1		;subtract count from exp
389	bgt.b	ap_n_fm		;if still pos, go fix mantissa
390	neg.l	d1		;take abs of exp and clr SE
391	move.l	(a0),d4		;load lword 1 to d4
392	and.l	#$bfffffff,d4	; and clr SE in d4
393	and.l	#$bfffffff,(a0)	; and in memory
394*
395* Calculate the mantissa multiplier to compensate for the appending of
396* zeros to the mantissa.
397*
398ap_n_fm:
399	move.l	#PTENRN,a1	;get address of power-of-ten table
400	clr.l	d3		;init table index
401	fmove.s	FONE,fp1	;init fp1 to 1
402	moveq.l	#3,d2		;init d2 to count bits in counter
403ap_n_el:
404	asr.l	#1,d0		;shift lsb into carry
405	bcc.b	ap_n_en		;if 1, mul fp1 by pwrten factor
406	fmul.x	(a1,d3),fp1	;mul by 10**(d3_bit_no)
407ap_n_en:
408	add.l	#12,d3		;inc d3 to next rtable entry
409	tst.l	d0		;check if d0 is zero
410	bne.b	ap_n_el		;if not, get next bit
411	fdiv.x	fp1,fp0		;div mantissa by 10**(no_bits_shifted)
412*
413*
414* Calculate power-of-ten factor from adjusted and shifted exponent.
415*
416* Register usage:
417*
418*  pwrten:
419*	(*)  d0: temp
420*	( )  d1: exponent
421*	(*)  d2: {FPCR[6:5],SM,SE} as index in RTABLE; temp
422*	(*)  d3: FPCR work copy
423*	( )  d4: first word of bcd
424*	(*)  a1: RTABLE pointer
425*  calc_p:
426*	(*)  d0: temp
427*	( )  d1: exponent
428*	(*)  d3: PWRTxx table index
429*	( )  a0: pointer to working copy of bcd
430*	(*)  a1: PWRTxx pointer
431*	(*) fp1: power-of-ten accumulator
432*
433* Pwrten calculates the exponent factor in the selected rounding mode
434* according to the following table:
435*
436*	Sign of Mant  Sign of Exp  Rounding Mode  PWRTEN Rounding Mode
437*
438*	ANY	  ANY	RN	RN
439*
440*	 +	   +	RP	RP
441*	 -	   +	RP	RM
442*	 +	   -	RP	RM
443*	 -	   -	RP	RP
444*
445*	 +	   +	RM	RM
446*	 -	   +	RM	RP
447*	 +	   -	RM	RP
448*	 -	   -	RM	RM
449*
450*	 +	   +	RZ	RM
451*	 -	   +	RZ	RM
452*	 +	   -	RZ	RP
453*	 -	   -	RZ	RP
454*
455*
456pwrten:
457	move.l	USER_FPCR(a6),d3 ;get user's FPCR
458	bfextu	d3{26:2},d2	;isolate rounding mode bits
459	move.l	(a0),d4		;reload 1st bcd word to d4
460	asl.l	#2,d2		;format d2 to be
461	bfextu	d4{0:2},d0	; {FPCR[6],FPCR[5],SM,SE}
462	add.l	d0,d2		;in d2 as index into RTABLE
463	lea.l	RTABLE,a1	;load rtable base
464	move.b	(a1,d2),d0	;load new rounding bits from table
465	clr.l	d3			;clear d3 to force no exc and extended
466	bfins	d0,d3{26:2}	;stuff new rounding bits in FPCR
467	fmove.l	d3,FPCR		;write new FPCR
468	asr.l	#1,d0		;write correct PTENxx table
469	bcc.b	not_rp		;to a1
470	lea.l	PTENRP,a1	;it is RP
471	bra.b	calc_p		;go to init section
472not_rp:
473	asr.l	#1,d0		;keep checking
474	bcc.b	not_rm
475	lea.l	PTENRM,a1	;it is RM
476	bra.b	calc_p		;go to init section
477not_rm:
478	lea.l	PTENRN,a1	;it is RN
479calc_p:
480	move.l	d1,d0		;copy exp to d0;use d0
481	bpl.b	no_neg		;if exp is negative,
482	neg.l	d0		;invert it
483	or.l	#$40000000,(a0)	;and set SE bit
484no_neg:
485	clr.l	d3		;table index
486	fmove.s	FONE,fp1	;init fp1 to 1
487e_loop:
488	asr.l	#1,d0		;shift next bit into carry
489	bcc.b	e_next		;if zero, skip the mul
490	fmul.x	(a1,d3),fp1	;mul by 10**(d3_bit_no)
491e_next:
492	add.l	#12,d3		;inc d3 to next rtable entry
493	tst.l	d0		;check if d0 is zero
494	bne.b	e_loop		;not zero, continue shifting
495*
496*
497*  Check the sign of the adjusted exp and make the value in fp0 the
498*  same sign. If the exp was pos then multiply fp1*fp0;
499*  else divide fp0/fp1.
500*
501* Register Usage:
502*  norm:
503*	( )  a0: pointer to working bcd value
504*	(*) fp0: mantissa accumulator
505*	( ) fp1: scaling factor - 10**(abs(exp))
506*
507norm:
508	btst	#30,(a0)	;test the sign of the exponent
509	beq.b	mul		;if clear, go to multiply
510div:
511	fdiv.x	fp1,fp0		;exp is negative, so divide mant by exp
512	bra.b	end_dec
513mul:
514	fmul.x	fp1,fp0		;exp is positive, so multiply by exp
515*
516*
517* Clean up and return with result in fp0.
518*
519* If the final mul/div in decbin incurred an inex exception,
520* it will be inex2, but will be reported as inex1 by get_op.
521*
522end_dec:
523	fmove.l	FPSR,d0		;get status register
524	bclr.l	#inex2_bit+8,d0	;test for inex2 and clear it
525	fmove.l	d0,FPSR		;return status reg w/o inex2
526	beq.b	no_exc		;skip this if no exc
527	or.l	#inx1a_mask,USER_FPSR(a6) ;set inex1/ainex
528no_exc:
529	movem.l	(a7)+,d2-d5
530	rts
531	end
532