1 /*
2  * fp_util.S
3  *
4  * Copyright Roman Zippel, 1997.  All rights reserved.
5  *
6  * Redistribution and use in source and binary forms, with or without
7  * modification, are permitted provided that the following conditions
8  * are met:
9  * 1. Redistributions of source code must retain the above copyright
10  *    notice, and the entire permission notice in its entirety,
11  *    including the disclaimer of warranties.
12  * 2. Redistributions in binary form must reproduce the above copyright
13  *    notice, this list of conditions and the following disclaimer in the
14  *    documentation and/or other materials provided with the distribution.
15  * 3. The name of the author may not be used to endorse or promote
16  *    products derived from this software without specific prior
17  *    written permission.
18  *
19  * ALTERNATIVELY, this product may be distributed under the terms of
20  * the GNU General Public License, in which case the provisions of the GPL are
21  * required INSTEAD OF the above restrictions.  (This clause is
22  * necessary due to a potential bad interaction between the GPL and
23  * the restrictions contained in a BSD-style copyright.)
24  *
25  * THIS SOFTWARE IS PROVIDED ``AS IS'' AND ANY EXPRESS OR IMPLIED
26  * WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES
27  * OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
28  * DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY DIRECT,
29  * INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
30  * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
31  * SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
32  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
33  * STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
34  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED
35  * OF THE POSSIBILITY OF SUCH DAMAGE.
36  */
37 
38 #include "fp_emu.h"
39 
40 /*
41  * Here are lots of conversion and normalization functions mainly
42  * used by fp_scan.S
43  * Note that these functions are optimized for "normal" numbers,
44  * these are handled first and exit as fast as possible, this is
45  * especially important for fp_normalize_ext/fp_conv_ext2ext, as
46  * it's called very often.
47  * The register usage is optimized for fp_scan.S and which register
48  * is currently at that time unused, be careful if you want change
49  * something here. %d0 and %d1 is always usable, sometimes %d2 (or
50  * only the lower half) most function have to return the %a0
51  * unmodified, so that the caller can immediately reuse it.
52  */
53 
54 	.globl	fp_ill, fp_end
55 
56 	| exits from fp_scan:
57 	| illegal instruction
58 fp_ill:
59 	printf	,"fp_illegal\n"
60 	rts
61 	| completed instruction
62 fp_end:
63 	tst.l	(TASK_MM-8,%a2)
64 	jmi	1f
65 	tst.l	(TASK_MM-4,%a2)
66 	jmi	1f
67 	tst.l	(TASK_MM,%a2)
68 	jpl	2f
69 1:	printf	,"oops:%p,%p,%p\n",3,%a2@(TASK_MM-8),%a2@(TASK_MM-4),%a2@(TASK_MM)
70 2:	clr.l	%d0
71 	rts
72 
73 	.globl	fp_conv_long2ext, fp_conv_single2ext
74 	.globl	fp_conv_double2ext, fp_conv_ext2ext
75 	.globl	fp_normalize_ext, fp_normalize_double
76 	.globl	fp_normalize_single, fp_normalize_single_fast
77 	.globl	fp_conv_ext2double, fp_conv_ext2single
78 	.globl	fp_conv_ext2long, fp_conv_ext2short
79 	.globl	fp_conv_ext2byte
80 	.globl	fp_finalrounding_single, fp_finalrounding_single_fast
81 	.globl	fp_finalrounding_double
82 	.globl	fp_finalrounding, fp_finaltest, fp_final
83 
84 /*
85  * First several conversion functions from a source operand
86  * into the extended format. Note, that only fp_conv_ext2ext
87  * normalizes the number and is always called after the other
88  * conversion functions, which only move the information into
89  * fp_ext structure.
90  */
91 
92 	| fp_conv_long2ext:
93 	|
94 	| args:	%d0 = source (32-bit long)
95 	|	%a0 = destination (ptr to struct fp_ext)
96 
97 fp_conv_long2ext:
98 	printf	PCONV,"l2e: %p -> %p(",2,%d0,%a0
99 	clr.l	%d1			| sign defaults to zero
100 	tst.l	%d0
101 	jeq	fp_l2e_zero		| is source zero?
102 	jpl	1f			| positive?
103 	moveq	#1,%d1
104 	neg.l	%d0
105 1:	swap	%d1
106 	move.w	#0x3fff+31,%d1
107 	move.l	%d1,(%a0)+		| set sign / exp
108 	move.l	%d0,(%a0)+		| set mantissa
109 	clr.l	(%a0)
110 	subq.l	#8,%a0			| restore %a0
111 	printx	PCONV,%a0@
112 	printf	PCONV,")\n"
113 	rts
114 	| source is zero
115 fp_l2e_zero:
116 	clr.l	(%a0)+
117 	clr.l	(%a0)+
118 	clr.l	(%a0)
119 	subq.l	#8,%a0
120 	printx	PCONV,%a0@
121 	printf	PCONV,")\n"
122 	rts
123 
124 	| fp_conv_single2ext
125 	| args:	%d0 = source (single-precision fp value)
126 	|	%a0 = dest (struct fp_ext *)
127 
128 fp_conv_single2ext:
129 	printf	PCONV,"s2e: %p -> %p(",2,%d0,%a0
130 	move.l	%d0,%d1
131 	lsl.l	#8,%d0			| shift mantissa
132 	lsr.l	#8,%d1			| exponent / sign
133 	lsr.l	#7,%d1
134 	lsr.w	#8,%d1
135 	jeq	fp_s2e_small		| zero / denormal?
136 	cmp.w	#0xff,%d1		| NaN / Inf?
137 	jeq	fp_s2e_large
138 	bset	#31,%d0			| set explizit bit
139 	add.w	#0x3fff-0x7f,%d1	| re-bias the exponent.
140 9:	move.l	%d1,(%a0)+		| fp_ext.sign, fp_ext.exp
141 	move.l	%d0,(%a0)+		| high lword of fp_ext.mant
142 	clr.l	(%a0)			| low lword = 0
143 	subq.l	#8,%a0
144 	printx	PCONV,%a0@
145 	printf	PCONV,")\n"
146 	rts
147 	| zeros and denormalized
148 fp_s2e_small:
149 	| exponent is zero, so explizit bit is already zero too
150 	tst.l	%d0
151 	jeq	9b
152 	move.w	#0x4000-0x7f,%d1
153 	jra	9b
154 	| infinities and NAN
155 fp_s2e_large:
156 	bclr	#31,%d0			| clear explizit bit
157 	move.w	#0x7fff,%d1
158 	jra	9b
159 
160 fp_conv_double2ext:
161 #ifdef FPU_EMU_DEBUG
162 	getuser.l %a1@(0),%d0,fp_err_ua2,%a1
163 	getuser.l %a1@(4),%d1,fp_err_ua2,%a1
164 	printf	PCONV,"d2e: %p%p -> %p(",3,%d0,%d1,%a0
165 #endif
166 	getuser.l (%a1)+,%d0,fp_err_ua2,%a1
167 	move.l	%d0,%d1
168 	lsl.l	#8,%d0			| shift high mantissa
169 	lsl.l	#3,%d0
170 	lsr.l	#8,%d1			| exponent / sign
171 	lsr.l	#7,%d1
172 	lsr.w	#5,%d1
173 	jeq	fp_d2e_small		| zero / denormal?
174 	cmp.w	#0x7ff,%d1		| NaN / Inf?
175 	jeq	fp_d2e_large
176 	bset	#31,%d0			| set explizit bit
177 	add.w	#0x3fff-0x3ff,%d1	| re-bias the exponent.
178 9:	move.l	%d1,(%a0)+		| fp_ext.sign, fp_ext.exp
179 	move.l	%d0,(%a0)+
180 	getuser.l (%a1)+,%d0,fp_err_ua2,%a1
181 	move.l	%d0,%d1
182 	lsl.l	#8,%d0
183 	lsl.l	#3,%d0
184 	move.l	%d0,(%a0)
185 	moveq	#21,%d0
186 	lsr.l	%d0,%d1
187 	or.l	%d1,-(%a0)
188 	subq.l	#4,%a0
189 	printx	PCONV,%a0@
190 	printf	PCONV,")\n"
191 	rts
192 	| zeros and denormalized
193 fp_d2e_small:
194 	| exponent is zero, so explizit bit is already zero too
195 	tst.l	%d0
196 	jeq	9b
197 	move.w	#0x4000-0x3ff,%d1
198 	jra	9b
199 	| infinities and NAN
200 fp_d2e_large:
201 	bclr	#31,%d0			| clear explizit bit
202 	move.w	#0x7fff,%d1
203 	jra	9b
204 
205 	| fp_conv_ext2ext:
206 	| originally used to get longdouble from userspace, now it's
207 	| called before arithmetic operations to make sure the number
208 	| is normalized [maybe rename it?].
209 	| args:	%a0 = dest (struct fp_ext *)
210 	| returns 0 in %d0 for a NaN, otherwise 1
211 
212 fp_conv_ext2ext:
213 	printf	PCONV,"e2e: %p(",1,%a0
214 	printx	PCONV,%a0@
215 	printf	PCONV,"), "
216 	move.l	(%a0)+,%d0
217 	cmp.w	#0x7fff,%d0		| Inf / NaN?
218 	jeq	fp_e2e_large
219 	move.l	(%a0),%d0
220 	jpl	fp_e2e_small		| zero / denorm?
221 	| The high bit is set, so normalization is irrelevant.
222 fp_e2e_checkround:
223 	subq.l	#4,%a0
224 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
225 	move.b	(%a0),%d0
226 	jne	fp_e2e_round
227 #endif
228 	printf	PCONV,"%p(",1,%a0
229 	printx	PCONV,%a0@
230 	printf	PCONV,")\n"
231 	moveq	#1,%d0
232 	rts
233 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
234 fp_e2e_round:
235 	fp_set_sr FPSR_EXC_INEX2
236 	clr.b	(%a0)
237 	move.w	(FPD_RND,FPDATA),%d2
238 	jne	fp_e2e_roundother	| %d2 == 0, round to nearest
239 	tst.b	%d0			| test guard bit
240 	jpl	9f			| zero is closer
241 	btst	#0,(11,%a0)		| test lsb bit
242 	jne	fp_e2e_doroundup	| round to infinity
243 	lsl.b	#1,%d0			| check low bits
244 	jeq	9f			| round to zero
245 fp_e2e_doroundup:
246 	addq.l	#1,(8,%a0)
247 	jcc	9f
248 	addq.l	#1,(4,%a0)
249 	jcc	9f
250 	move.w	#0x8000,(4,%a0)
251 	addq.w	#1,(2,%a0)
252 9:	printf	PNORM,"%p(",1,%a0
253 	printx	PNORM,%a0@
254 	printf	PNORM,")\n"
255 	rts
256 fp_e2e_roundother:
257 	subq.w	#2,%d2
258 	jcs	9b			| %d2 < 2, round to zero
259 	jhi	1f			| %d2 > 2, round to +infinity
260 	tst.b	(1,%a0)			| to -inf
261 	jne	fp_e2e_doroundup	| negative, round to infinity
262 	jra	9b			| positive, round to zero
263 1:	tst.b	(1,%a0)			| to +inf
264 	jeq	fp_e2e_doroundup	| positive, round to infinity
265 	jra	9b			| negative, round to zero
266 #endif
267 	| zeros and subnormals:
268 	| try to normalize these anyway.
269 fp_e2e_small:
270 	jne	fp_e2e_small1		| high lword zero?
271 	move.l	(4,%a0),%d0
272 	jne	fp_e2e_small2
273 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
274 	clr.l	%d0
275 	move.b	(-4,%a0),%d0
276 	jne	fp_e2e_small3
277 #endif
278 	| Genuine zero.
279 	clr.w	-(%a0)
280 	subq.l	#2,%a0
281 	printf	PNORM,"%p(",1,%a0
282 	printx	PNORM,%a0@
283 	printf	PNORM,")\n"
284 	moveq	#1,%d0
285 	rts
286 	| definitely subnormal, need to shift all 64 bits
287 fp_e2e_small1:
288 	bfffo	%d0{#0,#32},%d1
289 	move.w	-(%a0),%d2
290 	sub.w	%d1,%d2
291 	jcc	1f
292 	| Pathologically small, denormalize.
293 	add.w	%d2,%d1
294 	clr.w	%d2
295 1:	move.w	%d2,(%a0)+
296 	move.w	%d1,%d2
297 	jeq	fp_e2e_checkround
298 	| fancy 64-bit double-shift begins here
299 	lsl.l	%d2,%d0
300 	move.l	%d0,(%a0)+
301 	move.l	(%a0),%d0
302 	move.l	%d0,%d1
303 	lsl.l	%d2,%d0
304 	move.l	%d0,(%a0)
305 	neg.w	%d2
306 	and.w	#0x1f,%d2
307 	lsr.l	%d2,%d1
308 	or.l	%d1,-(%a0)
309 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
310 fp_e2e_extra1:
311 	clr.l	%d0
312 	move.b	(-4,%a0),%d0
313 	neg.w	%d2
314 	add.w	#24,%d2
315 	jcc	1f
316 	clr.b	(-4,%a0)
317 	lsl.l	%d2,%d0
318 	or.l	%d0,(4,%a0)
319 	jra	fp_e2e_checkround
320 1:	addq.w	#8,%d2
321 	lsl.l	%d2,%d0
322 	move.b	%d0,(-4,%a0)
323 	lsr.l	#8,%d0
324 	or.l	%d0,(4,%a0)
325 #endif
326 	jra	fp_e2e_checkround
327 	| pathologically small subnormal
328 fp_e2e_small2:
329 	bfffo	%d0{#0,#32},%d1
330 	add.w	#32,%d1
331 	move.w	-(%a0),%d2
332 	sub.w	%d1,%d2
333 	jcc	1f
334 	| Beyond pathologically small, denormalize.
335 	add.w	%d2,%d1
336 	clr.w	%d2
337 1:	move.w	%d2,(%a0)+
338 	ext.l	%d1
339 	jeq	fp_e2e_checkround
340 	clr.l	(4,%a0)
341 	sub.w	#32,%d2
342 	jcs	1f
343 	lsl.l	%d1,%d0			| lower lword needs only to be shifted
344 	move.l	%d0,(%a0)		| into the higher lword
345 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
346 	clr.l	%d0
347 	move.b	(-4,%a0),%d0
348 	clr.b	(-4,%a0)
349 	neg.w	%d1
350 	add.w	#32,%d1
351 	bfins	%d0,(%a0){%d1,#8}
352 #endif
353 	jra	fp_e2e_checkround
354 1:	neg.w	%d1			| lower lword is splitted between
355 	bfins	%d0,(%a0){%d1,#32}	| higher and lower lword
356 #ifndef CONFIG_M68KFPU_EMU_EXTRAPREC
357 	jra	fp_e2e_checkround
358 #else
359 	move.w	%d1,%d2
360 	jra	fp_e2e_extra1
361 	| These are extremely small numbers, that will mostly end up as zero
362 	| anyway, so this is only important for correct rounding.
363 fp_e2e_small3:
364 	bfffo	%d0{#24,#8},%d1
365 	add.w	#40,%d1
366 	move.w	-(%a0),%d2
367 	sub.w	%d1,%d2
368 	jcc	1f
369 	| Pathologically small, denormalize.
370 	add.w	%d2,%d1
371 	clr.w	%d2
372 1:	move.w	%d2,(%a0)+
373 	ext.l	%d1
374 	jeq	fp_e2e_checkround
375 	cmp.w	#8,%d1
376 	jcs	2f
377 1:	clr.b	(-4,%a0)
378 	sub.w	#64,%d1
379 	jcs	1f
380 	add.w	#24,%d1
381 	lsl.l	%d1,%d0
382 	move.l	%d0,(%a0)
383 	jra	fp_e2e_checkround
384 1:	neg.w	%d1
385 	bfins	%d0,(%a0){%d1,#8}
386 	jra	fp_e2e_checkround
387 2:	lsl.l	%d1,%d0
388 	move.b	%d0,(-4,%a0)
389 	lsr.l	#8,%d0
390 	move.b	%d0,(7,%a0)
391 	jra	fp_e2e_checkround
392 #endif
393 1:	move.l	%d0,%d1			| lower lword is splitted between
394 	lsl.l	%d2,%d0			| higher and lower lword
395 	move.l	%d0,(%a0)
396 	move.l	%d1,%d0
397 	neg.w	%d2
398 	add.w	#32,%d2
399 	lsr.l	%d2,%d0
400 	move.l	%d0,-(%a0)
401 	jra	fp_e2e_checkround
402 	| Infinities and NaNs
403 fp_e2e_large:
404 	move.l	(%a0)+,%d0
405 	jne	3f
406 1:	tst.l	(%a0)
407 	jne	4f
408 	moveq	#1,%d0
409 2:	subq.l	#8,%a0
410 	printf	PCONV,"%p(",1,%a0
411 	printx	PCONV,%a0@
412 	printf	PCONV,")\n"
413 	rts
414 	| we have maybe a NaN, shift off the highest bit
415 3:	lsl.l	#1,%d0
416 	jeq	1b
417 	| we have a NaN, clear the return value
418 4:	clrl	%d0
419 	jra	2b
420 
421 
422 /*
423  * Normalization functions.  Call these on the output of general
424  * FP operators, and before any conversion into the destination
425  * formats. fp_normalize_ext has always to be called first, the
426  * following conversion functions expect an already normalized
427  * number.
428  */
429 
430 	| fp_normalize_ext:
431 	| normalize an extended in extended (unpacked) format, basically
432 	| it does the same as fp_conv_ext2ext, additionally it also does
433 	| the necessary postprocessing checks.
434 	| args:	%a0 (struct fp_ext *)
435 	| NOTE: it does _not_ modify %a0/%a1 and the upper word of %d2
436 
437 fp_normalize_ext:
438 	printf	PNORM,"ne: %p(",1,%a0
439 	printx	PNORM,%a0@
440 	printf	PNORM,"), "
441 	move.l	(%a0)+,%d0
442 	cmp.w	#0x7fff,%d0		| Inf / NaN?
443 	jeq	fp_ne_large
444 	move.l	(%a0),%d0
445 	jpl	fp_ne_small		| zero / denorm?
446 	| The high bit is set, so normalization is irrelevant.
447 fp_ne_checkround:
448 	subq.l	#4,%a0
449 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
450 	move.b	(%a0),%d0
451 	jne	fp_ne_round
452 #endif
453 	printf	PNORM,"%p(",1,%a0
454 	printx	PNORM,%a0@
455 	printf	PNORM,")\n"
456 	rts
457 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
458 fp_ne_round:
459 	fp_set_sr FPSR_EXC_INEX2
460 	clr.b	(%a0)
461 	move.w	(FPD_RND,FPDATA),%d2
462 	jne	fp_ne_roundother	| %d2 == 0, round to nearest
463 	tst.b	%d0			| test guard bit
464 	jpl	9f			| zero is closer
465 	btst	#0,(11,%a0)		| test lsb bit
466 	jne	fp_ne_doroundup		| round to infinity
467 	lsl.b	#1,%d0			| check low bits
468 	jeq	9f			| round to zero
469 fp_ne_doroundup:
470 	addq.l	#1,(8,%a0)
471 	jcc	9f
472 	addq.l	#1,(4,%a0)
473 	jcc	9f
474 	addq.w	#1,(2,%a0)
475 	move.w	#0x8000,(4,%a0)
476 9:	printf	PNORM,"%p(",1,%a0
477 	printx	PNORM,%a0@
478 	printf	PNORM,")\n"
479 	rts
480 fp_ne_roundother:
481 	subq.w	#2,%d2
482 	jcs	9b			| %d2 < 2, round to zero
483 	jhi	1f			| %d2 > 2, round to +infinity
484 	tst.b	(1,%a0)			| to -inf
485 	jne	fp_ne_doroundup		| negative, round to infinity
486 	jra	9b			| positive, round to zero
487 1:	tst.b	(1,%a0)			| to +inf
488 	jeq	fp_ne_doroundup		| positive, round to infinity
489 	jra	9b			| negative, round to zero
490 #endif
491 	| Zeros and subnormal numbers
492 	| These are probably merely subnormal, rather than "denormalized"
493 	|  numbers, so we will try to make them normal again.
494 fp_ne_small:
495 	jne	fp_ne_small1		| high lword zero?
496 	move.l	(4,%a0),%d0
497 	jne	fp_ne_small2
498 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
499 	clr.l	%d0
500 	move.b	(-4,%a0),%d0
501 	jne	fp_ne_small3
502 #endif
503 	| Genuine zero.
504 	clr.w	-(%a0)
505 	subq.l	#2,%a0
506 	printf	PNORM,"%p(",1,%a0
507 	printx	PNORM,%a0@
508 	printf	PNORM,")\n"
509 	rts
510 	| Subnormal.
511 fp_ne_small1:
512 	bfffo	%d0{#0,#32},%d1
513 	move.w	-(%a0),%d2
514 	sub.w	%d1,%d2
515 	jcc	1f
516 	| Pathologically small, denormalize.
517 	add.w	%d2,%d1
518 	clr.w	%d2
519 	fp_set_sr FPSR_EXC_UNFL
520 1:	move.w	%d2,(%a0)+
521 	move.w	%d1,%d2
522 	jeq	fp_ne_checkround
523 	| This is exactly the same 64-bit double shift as seen above.
524 	lsl.l	%d2,%d0
525 	move.l	%d0,(%a0)+
526 	move.l	(%a0),%d0
527 	move.l	%d0,%d1
528 	lsl.l	%d2,%d0
529 	move.l	%d0,(%a0)
530 	neg.w	%d2
531 	and.w	#0x1f,%d2
532 	lsr.l	%d2,%d1
533 	or.l	%d1,-(%a0)
534 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
535 fp_ne_extra1:
536 	clr.l	%d0
537 	move.b	(-4,%a0),%d0
538 	neg.w	%d2
539 	add.w	#24,%d2
540 	jcc	1f
541 	clr.b	(-4,%a0)
542 	lsl.l	%d2,%d0
543 	or.l	%d0,(4,%a0)
544 	jra	fp_ne_checkround
545 1:	addq.w	#8,%d2
546 	lsl.l	%d2,%d0
547 	move.b	%d0,(-4,%a0)
548 	lsr.l	#8,%d0
549 	or.l	%d0,(4,%a0)
550 #endif
551 	jra	fp_ne_checkround
552 	| May or may not be subnormal, if so, only 32 bits to shift.
553 fp_ne_small2:
554 	bfffo	%d0{#0,#32},%d1
555 	add.w	#32,%d1
556 	move.w	-(%a0),%d2
557 	sub.w	%d1,%d2
558 	jcc	1f
559 	| Beyond pathologically small, denormalize.
560 	add.w	%d2,%d1
561 	clr.w	%d2
562 	fp_set_sr FPSR_EXC_UNFL
563 1:	move.w	%d2,(%a0)+
564 	ext.l	%d1
565 	jeq	fp_ne_checkround
566 	clr.l	(4,%a0)
567 	sub.w	#32,%d1
568 	jcs	1f
569 	lsl.l	%d1,%d0			| lower lword needs only to be shifted
570 	move.l	%d0,(%a0)		| into the higher lword
571 #ifdef CONFIG_M68KFPU_EMU_EXTRAPREC
572 	clr.l	%d0
573 	move.b	(-4,%a0),%d0
574 	clr.b	(-4,%a0)
575 	neg.w	%d1
576 	add.w	#32,%d1
577 	bfins	%d0,(%a0){%d1,#8}
578 #endif
579 	jra	fp_ne_checkround
580 1:	neg.w	%d1			| lower lword is splitted between
581 	bfins	%d0,(%a0){%d1,#32}	| higher and lower lword
582 #ifndef CONFIG_M68KFPU_EMU_EXTRAPREC
583 	jra	fp_ne_checkround
584 #else
585 	move.w	%d1,%d2
586 	jra	fp_ne_extra1
587 	| These are extremely small numbers, that will mostly end up as zero
588 	| anyway, so this is only important for correct rounding.
589 fp_ne_small3:
590 	bfffo	%d0{#24,#8},%d1
591 	add.w	#40,%d1
592 	move.w	-(%a0),%d2
593 	sub.w	%d1,%d2
594 	jcc	1f
595 	| Pathologically small, denormalize.
596 	add.w	%d2,%d1
597 	clr.w	%d2
598 1:	move.w	%d2,(%a0)+
599 	ext.l	%d1
600 	jeq	fp_ne_checkround
601 	cmp.w	#8,%d1
602 	jcs	2f
603 1:	clr.b	(-4,%a0)
604 	sub.w	#64,%d1
605 	jcs	1f
606 	add.w	#24,%d1
607 	lsl.l	%d1,%d0
608 	move.l	%d0,(%a0)
609 	jra	fp_ne_checkround
610 1:	neg.w	%d1
611 	bfins	%d0,(%a0){%d1,#8}
612 	jra	fp_ne_checkround
613 2:	lsl.l	%d1,%d0
614 	move.b	%d0,(-4,%a0)
615 	lsr.l	#8,%d0
616 	move.b	%d0,(7,%a0)
617 	jra	fp_ne_checkround
618 #endif
619 	| Infinities and NaNs, again, same as above.
620 fp_ne_large:
621 	move.l	(%a0)+,%d0
622 	jne	3f
623 1:	tst.l	(%a0)
624 	jne	4f
625 2:	subq.l	#8,%a0
626 	printf	PNORM,"%p(",1,%a0
627 	printx	PNORM,%a0@
628 	printf	PNORM,")\n"
629 	rts
630 	| we have maybe a NaN, shift off the highest bit
631 3:	move.l	%d0,%d1
632 	lsl.l	#1,%d1
633 	jne	4f
634 	clr.l	(-4,%a0)
635 	jra	1b
636 	| we have a NaN, test if it is signaling
637 4:	bset	#30,%d0
638 	jne	2b
639 	fp_set_sr FPSR_EXC_SNAN
640 	move.l	%d0,(-4,%a0)
641 	jra	2b
642 
643 	| these next two do rounding as per the IEEE standard.
644 	| values for the rounding modes appear to be:
645 	| 0:	Round to nearest
646 	| 1:	Round to zero
647 	| 2:	Round to -Infinity
648 	| 3:	Round to +Infinity
649 	| both functions expect that fp_normalize was already
650 	| called (and extended argument is already normalized
651 	| as far as possible), these are used if there is different
652 	| rounding precision is selected and before converting
653 	| into single/double
654 
655 	| fp_normalize_double:
656 	| normalize an extended with double (52-bit) precision
657 	| args:	 %a0 (struct fp_ext *)
658 
659 fp_normalize_double:
660 	printf	PNORM,"nd: %p(",1,%a0
661 	printx	PNORM,%a0@
662 	printf	PNORM,"), "
663 	move.l	(%a0)+,%d2
664 	tst.w	%d2
665 	jeq	fp_nd_zero		| zero / denormalized
666 	cmp.w	#0x7fff,%d2
667 	jeq	fp_nd_huge		| NaN / infinitive.
668 	sub.w	#0x4000-0x3ff,%d2	| will the exponent fit?
669 	jcs	fp_nd_small		| too small.
670 	cmp.w	#0x7fe,%d2
671 	jcc	fp_nd_large		| too big.
672 	addq.l	#4,%a0
673 	move.l	(%a0),%d0		| low lword of mantissa
674 	| now, round off the low 11 bits.
675 fp_nd_round:
676 	moveq	#21,%d1
677 	lsl.l	%d1,%d0			| keep 11 low bits.
678 	jne	fp_nd_checkround	| Are they non-zero?
679 	| nothing to do here
680 9:	subq.l	#8,%a0
681 	printf	PNORM,"%p(",1,%a0
682 	printx	PNORM,%a0@
683 	printf	PNORM,")\n"
684 	rts
685 	| Be careful with the X bit! It contains the lsb
686 	| from the shift above, it is needed for round to nearest.
687 fp_nd_checkround:
688 	fp_set_sr FPSR_EXC_INEX2	| INEX2 bit
689 	and.w	#0xf800,(2,%a0)		| clear bits 0-10
690 	move.w	(FPD_RND,FPDATA),%d2	| rounding mode
691 	jne	2f			| %d2 == 0, round to nearest
692 	tst.l	%d0			| test guard bit
693 	jpl	9b			| zero is closer
694 	| here we test the X bit by adding it to %d2
695 	clr.w	%d2			| first set z bit, addx only clears it
696 	addx.w	%d2,%d2			| test lsb bit
697 	| IEEE754-specified "round to even" behaviour.  If the guard
698 	| bit is set, then the number is odd, so rounding works like
699 	| in grade-school arithmetic (i.e. 1.5 rounds to 2.0)
700 	| Otherwise, an equal distance rounds towards zero, so as not
701 	| to produce an odd number.  This is strange, but it is what
702 	| the standard says.
703 	jne	fp_nd_doroundup		| round to infinity
704 	lsl.l	#1,%d0			| check low bits
705 	jeq	9b			| round to zero
706 fp_nd_doroundup:
707 	| round (the mantissa, that is) towards infinity
708 	add.l	#0x800,(%a0)
709 	jcc	9b			| no overflow, good.
710 	addq.l	#1,-(%a0)		| extend to high lword
711 	jcc	1f			| no overflow, good.
712 	| Yow! we have managed to overflow the mantissa.  Since this
713 	| only happens when %d1 was 0xfffff800, it is now zero, so
714 	| reset the high bit, and increment the exponent.
715 	move.w	#0x8000,(%a0)
716 	addq.w	#1,-(%a0)
717 	cmp.w	#0x43ff,(%a0)+		| exponent now overflown?
718 	jeq	fp_nd_large		| yes, so make it infinity.
719 1:	subq.l	#4,%a0
720 	printf	PNORM,"%p(",1,%a0
721 	printx	PNORM,%a0@
722 	printf	PNORM,")\n"
723 	rts
724 2:	subq.w	#2,%d2
725 	jcs	9b			| %d2 < 2, round to zero
726 	jhi	3f			| %d2 > 2, round to +infinity
727 	| Round to +Inf or -Inf.  High word of %d2 contains the
728 	| sign of the number, by the way.
729 	swap	%d2			| to -inf
730 	tst.b	%d2
731 	jne	fp_nd_doroundup		| negative, round to infinity
732 	jra	9b			| positive, round to zero
733 3:	swap	%d2			| to +inf
734 	tst.b	%d2
735 	jeq	fp_nd_doroundup		| positive, round to infinity
736 	jra	9b			| negative, round to zero
737 	| Exponent underflow.  Try to make a denormal, and set it to
738 	| the smallest possible fraction if this fails.
739 fp_nd_small:
740 	fp_set_sr FPSR_EXC_UNFL		| set UNFL bit
741 	move.w	#0x3c01,(-2,%a0)	| 2**-1022
742 	neg.w	%d2			| degree of underflow
743 	cmp.w	#32,%d2			| single or double shift?
744 	jcc	1f
745 	| Again, another 64-bit double shift.
746 	move.l	(%a0),%d0
747 	move.l	%d0,%d1
748 	lsr.l	%d2,%d0
749 	move.l	%d0,(%a0)+
750 	move.l	(%a0),%d0
751 	lsr.l	%d2,%d0
752 	neg.w	%d2
753 	add.w	#32,%d2
754 	lsl.l	%d2,%d1
755 	or.l	%d1,%d0
756 	move.l	(%a0),%d1
757 	move.l	%d0,(%a0)
758 	| Check to see if we shifted off any significant bits
759 	lsl.l	%d2,%d1
760 	jeq	fp_nd_round		| Nope, round.
761 	bset	#0,%d0			| Yes, so set the "sticky bit".
762 	jra	fp_nd_round		| Now, round.
763 	| Another 64-bit single shift and store
764 1:	sub.w	#32,%d2
765 	cmp.w	#32,%d2			| Do we really need to shift?
766 	jcc	2f			| No, the number is too small.
767 	move.l	(%a0),%d0
768 	clr.l	(%a0)+
769 	move.l	%d0,%d1
770 	lsr.l	%d2,%d0
771 	neg.w	%d2
772 	add.w	#32,%d2
773 	| Again, check to see if we shifted off any significant bits.
774 	tst.l	(%a0)
775 	jeq	1f
776 	bset	#0,%d0			| Sticky bit.
777 1:	move.l	%d0,(%a0)
778 	lsl.l	%d2,%d1
779 	jeq	fp_nd_round
780 	bset	#0,%d0
781 	jra	fp_nd_round
782 	| Sorry, the number is just too small.
783 2:	clr.l	(%a0)+
784 	clr.l	(%a0)
785 	moveq	#1,%d0			| Smallest possible fraction,
786 	jra	fp_nd_round		| round as desired.
787 	| zero and denormalized
788 fp_nd_zero:
789 	tst.l	(%a0)+
790 	jne	1f
791 	tst.l	(%a0)
792 	jne	1f
793 	subq.l	#8,%a0
794 	printf	PNORM,"%p(",1,%a0
795 	printx	PNORM,%a0@
796 	printf	PNORM,")\n"
797 	rts				| zero.  nothing to do.
798 	| These are not merely subnormal numbers, but true denormals,
799 	| i.e. pathologically small (exponent is 2**-16383) numbers.
800 	| It is clearly impossible for even a normal extended number
801 	| with that exponent to fit into double precision, so just
802 	| write these ones off as "too darn small".
803 1:	fp_set_sr FPSR_EXC_UNFL		| Set UNFL bit
804 	clr.l	(%a0)
805 	clr.l	-(%a0)
806 	move.w	#0x3c01,-(%a0)		| i.e. 2**-1022
807 	addq.l	#6,%a0
808 	moveq	#1,%d0
809 	jra	fp_nd_round		| round.
810 	| Exponent overflow.  Just call it infinity.
811 fp_nd_large:
812 	move.w	#0x7ff,%d0
813 	and.w	(6,%a0),%d0
814 	jeq	1f
815 	fp_set_sr FPSR_EXC_INEX2
816 1:	fp_set_sr FPSR_EXC_OVFL
817 	move.w	(FPD_RND,FPDATA),%d2
818 	jne	3f			| %d2 = 0 round to nearest
819 1:	move.w	#0x7fff,(-2,%a0)
820 	clr.l	(%a0)+
821 	clr.l	(%a0)
822 2:	subq.l	#8,%a0
823 	printf	PNORM,"%p(",1,%a0
824 	printx	PNORM,%a0@
825 	printf	PNORM,")\n"
826 	rts
827 3:	subq.w	#2,%d2
828 	jcs	5f			| %d2 < 2, round to zero
829 	jhi	4f			| %d2 > 2, round to +infinity
830 	tst.b	(-3,%a0)		| to -inf
831 	jne	1b
832 	jra	5f
833 4:	tst.b	(-3,%a0)		| to +inf
834 	jeq	1b
835 5:	move.w	#0x43fe,(-2,%a0)
836 	moveq	#-1,%d0
837 	move.l	%d0,(%a0)+
838 	move.w	#0xf800,%d0
839 	move.l	%d0,(%a0)
840 	jra	2b
841 	| Infinities or NaNs
842 fp_nd_huge:
843 	subq.l	#4,%a0
844 	printf	PNORM,"%p(",1,%a0
845 	printx	PNORM,%a0@
846 	printf	PNORM,")\n"
847 	rts
848 
849 	| fp_normalize_single:
850 	| normalize an extended with single (23-bit) precision
851 	| args:	 %a0 (struct fp_ext *)
852 
853 fp_normalize_single:
854 	printf	PNORM,"ns: %p(",1,%a0
855 	printx	PNORM,%a0@
856 	printf	PNORM,") "
857 	addq.l	#2,%a0
858 	move.w	(%a0)+,%d2
859 	jeq	fp_ns_zero		| zero / denormalized
860 	cmp.w	#0x7fff,%d2
861 	jeq	fp_ns_huge		| NaN / infinitive.
862 	sub.w	#0x4000-0x7f,%d2	| will the exponent fit?
863 	jcs	fp_ns_small		| too small.
864 	cmp.w	#0xfe,%d2
865 	jcc	fp_ns_large		| too big.
866 	move.l	(%a0)+,%d0		| get high lword of mantissa
867 fp_ns_round:
868 	tst.l	(%a0)			| check the low lword
869 	jeq	1f
870 	| Set a sticky bit if it is non-zero.  This should only
871 	| affect the rounding in what would otherwise be equal-
872 	| distance situations, which is what we want it to do.
873 	bset	#0,%d0
874 1:	clr.l	(%a0)			| zap it from memory.
875 	| now, round off the low 8 bits of the hi lword.
876 	tst.b	%d0			| 8 low bits.
877 	jne	fp_ns_checkround	| Are they non-zero?
878 	| nothing to do here
879 	subq.l	#8,%a0
880 	printf	PNORM,"%p(",1,%a0
881 	printx	PNORM,%a0@
882 	printf	PNORM,")\n"
883 	rts
884 fp_ns_checkround:
885 	fp_set_sr FPSR_EXC_INEX2	| INEX2 bit
886 	clr.b	-(%a0)			| clear low byte of high lword
887 	subq.l	#3,%a0
888 	move.w	(FPD_RND,FPDATA),%d2	| rounding mode
889 	jne	2f			| %d2 == 0, round to nearest
890 	tst.b	%d0			| test guard bit
891 	jpl	9f			| zero is closer
892 	btst	#8,%d0			| test lsb bit
893 	| round to even behaviour, see above.
894 	jne	fp_ns_doroundup		| round to infinity
895 	lsl.b	#1,%d0			| check low bits
896 	jeq	9f			| round to zero
897 fp_ns_doroundup:
898 	| round (the mantissa, that is) towards infinity
899 	add.l	#0x100,(%a0)
900 	jcc	9f			| no overflow, good.
901 	| Overflow.  This means that the %d1 was 0xffffff00, so it
902 	| is now zero.  We will set the mantissa to reflect this, and
903 	| increment the exponent (checking for overflow there too)
904 	move.w	#0x8000,(%a0)
905 	addq.w	#1,-(%a0)
906 	cmp.w	#0x407f,(%a0)+		| exponent now overflown?
907 	jeq	fp_ns_large		| yes, so make it infinity.
908 9:	subq.l	#4,%a0
909 	printf	PNORM,"%p(",1,%a0
910 	printx	PNORM,%a0@
911 	printf	PNORM,")\n"
912 	rts
913 	| check nondefault rounding modes
914 2:	subq.w	#2,%d2
915 	jcs	9b			| %d2 < 2, round to zero
916 	jhi	3f			| %d2 > 2, round to +infinity
917 	tst.b	(-3,%a0)		| to -inf
918 	jne	fp_ns_doroundup		| negative, round to infinity
919 	jra	9b			| positive, round to zero
920 3:	tst.b	(-3,%a0)		| to +inf
921 	jeq	fp_ns_doroundup		| positive, round to infinity
922 	jra	9b			| negative, round to zero
923 	| Exponent underflow.  Try to make a denormal, and set it to
924 	| the smallest possible fraction if this fails.
925 fp_ns_small:
926 	fp_set_sr FPSR_EXC_UNFL		| set UNFL bit
927 	move.w	#0x3f81,(-2,%a0)	| 2**-126
928 	neg.w	%d2			| degree of underflow
929 	cmp.w	#32,%d2			| single or double shift?
930 	jcc	2f
931 	| a 32-bit shift.
932 	move.l	(%a0),%d0
933 	move.l	%d0,%d1
934 	lsr.l	%d2,%d0
935 	move.l	%d0,(%a0)+
936 	| Check to see if we shifted off any significant bits.
937 	neg.w	%d2
938 	add.w	#32,%d2
939 	lsl.l	%d2,%d1
940 	jeq	1f
941 	bset	#0,%d0			| Sticky bit.
942 	| Check the lower lword
943 1:	tst.l	(%a0)
944 	jeq	fp_ns_round
945 	clr	(%a0)
946 	bset	#0,%d0			| Sticky bit.
947 	jra	fp_ns_round
948 	| Sorry, the number is just too small.
949 2:	clr.l	(%a0)+
950 	clr.l	(%a0)
951 	moveq	#1,%d0			| Smallest possible fraction,
952 	jra	fp_ns_round		| round as desired.
953 	| Exponent overflow.  Just call it infinity.
954 fp_ns_large:
955 	tst.b	(3,%a0)
956 	jeq	1f
957 	fp_set_sr FPSR_EXC_INEX2
958 1:	fp_set_sr FPSR_EXC_OVFL
959 	move.w	(FPD_RND,FPDATA),%d2
960 	jne	3f			| %d2 = 0 round to nearest
961 1:	move.w	#0x7fff,(-2,%a0)
962 	clr.l	(%a0)+
963 	clr.l	(%a0)
964 2:	subq.l	#8,%a0
965 	printf	PNORM,"%p(",1,%a0
966 	printx	PNORM,%a0@
967 	printf	PNORM,")\n"
968 	rts
969 3:	subq.w	#2,%d2
970 	jcs	5f			| %d2 < 2, round to zero
971 	jhi	4f			| %d2 > 2, round to +infinity
972 	tst.b	(-3,%a0)		| to -inf
973 	jne	1b
974 	jra	5f
975 4:	tst.b	(-3,%a0)		| to +inf
976 	jeq	1b
977 5:	move.w	#0x407e,(-2,%a0)
978 	move.l	#0xffffff00,(%a0)+
979 	clr.l	(%a0)
980 	jra	2b
981 	| zero and denormalized
982 fp_ns_zero:
983 	tst.l	(%a0)+
984 	jne	1f
985 	tst.l	(%a0)
986 	jne	1f
987 	subq.l	#8,%a0
988 	printf	PNORM,"%p(",1,%a0
989 	printx	PNORM,%a0@
990 	printf	PNORM,")\n"
991 	rts				| zero.  nothing to do.
992 	| These are not merely subnormal numbers, but true denormals,
993 	| i.e. pathologically small (exponent is 2**-16383) numbers.
994 	| It is clearly impossible for even a normal extended number
995 	| with that exponent to fit into single precision, so just
996 	| write these ones off as "too darn small".
997 1:	fp_set_sr FPSR_EXC_UNFL		| Set UNFL bit
998 	clr.l	(%a0)
999 	clr.l	-(%a0)
1000 	move.w	#0x3f81,-(%a0)		| i.e. 2**-126
1001 	addq.l	#6,%a0
1002 	moveq	#1,%d0
1003 	jra	fp_ns_round		| round.
1004 	| Infinities or NaNs
1005 fp_ns_huge:
1006 	subq.l	#4,%a0
1007 	printf	PNORM,"%p(",1,%a0
1008 	printx	PNORM,%a0@
1009 	printf	PNORM,")\n"
1010 	rts
1011 
1012 	| fp_normalize_single_fast:
1013 	| normalize an extended with single (23-bit) precision
1014 	| this is only used by fsgldiv/fsgdlmul, where the
1015 	| operand is not completly normalized.
1016 	| args:	 %a0 (struct fp_ext *)
1017 
1018 fp_normalize_single_fast:
1019 	printf	PNORM,"nsf: %p(",1,%a0
1020 	printx	PNORM,%a0@
1021 	printf	PNORM,") "
1022 	addq.l	#2,%a0
1023 	move.w	(%a0)+,%d2
1024 	cmp.w	#0x7fff,%d2
1025 	jeq	fp_nsf_huge		| NaN / infinitive.
1026 	move.l	(%a0)+,%d0		| get high lword of mantissa
1027 fp_nsf_round:
1028 	tst.l	(%a0)			| check the low lword
1029 	jeq	1f
1030 	| Set a sticky bit if it is non-zero.  This should only
1031 	| affect the rounding in what would otherwise be equal-
1032 	| distance situations, which is what we want it to do.
1033 	bset	#0,%d0
1034 1:	clr.l	(%a0)			| zap it from memory.
1035 	| now, round off the low 8 bits of the hi lword.
1036 	tst.b	%d0			| 8 low bits.
1037 	jne	fp_nsf_checkround	| Are they non-zero?
1038 	| nothing to do here
1039 	subq.l	#8,%a0
1040 	printf	PNORM,"%p(",1,%a0
1041 	printx	PNORM,%a0@
1042 	printf	PNORM,")\n"
1043 	rts
1044 fp_nsf_checkround:
1045 	fp_set_sr FPSR_EXC_INEX2	| INEX2 bit
1046 	clr.b	-(%a0)			| clear low byte of high lword
1047 	subq.l	#3,%a0
1048 	move.w	(FPD_RND,FPDATA),%d2	| rounding mode
1049 	jne	2f			| %d2 == 0, round to nearest
1050 	tst.b	%d0			| test guard bit
1051 	jpl	9f			| zero is closer
1052 	btst	#8,%d0			| test lsb bit
1053 	| round to even behaviour, see above.
1054 	jne	fp_nsf_doroundup		| round to infinity
1055 	lsl.b	#1,%d0			| check low bits
1056 	jeq	9f			| round to zero
1057 fp_nsf_doroundup:
1058 	| round (the mantissa, that is) towards infinity
1059 	add.l	#0x100,(%a0)
1060 	jcc	9f			| no overflow, good.
1061 	| Overflow.  This means that the %d1 was 0xffffff00, so it
1062 	| is now zero.  We will set the mantissa to reflect this, and
1063 	| increment the exponent (checking for overflow there too)
1064 	move.w	#0x8000,(%a0)
1065 	addq.w	#1,-(%a0)
1066 	cmp.w	#0x407f,(%a0)+		| exponent now overflown?
1067 	jeq	fp_nsf_large		| yes, so make it infinity.
1068 9:	subq.l	#4,%a0
1069 	printf	PNORM,"%p(",1,%a0
1070 	printx	PNORM,%a0@
1071 	printf	PNORM,")\n"
1072 	rts
1073 	| check nondefault rounding modes
1074 2:	subq.w	#2,%d2
1075 	jcs	9b			| %d2 < 2, round to zero
1076 	jhi	3f			| %d2 > 2, round to +infinity
1077 	tst.b	(-3,%a0)		| to -inf
1078 	jne	fp_nsf_doroundup	| negative, round to infinity
1079 	jra	9b			| positive, round to zero
1080 3:	tst.b	(-3,%a0)		| to +inf
1081 	jeq	fp_nsf_doroundup		| positive, round to infinity
1082 	jra	9b			| negative, round to zero
1083 	| Exponent overflow.  Just call it infinity.
1084 fp_nsf_large:
1085 	tst.b	(3,%a0)
1086 	jeq	1f
1087 	fp_set_sr FPSR_EXC_INEX2
1088 1:	fp_set_sr FPSR_EXC_OVFL
1089 	move.w	(FPD_RND,FPDATA),%d2
1090 	jne	3f			| %d2 = 0 round to nearest
1091 1:	move.w	#0x7fff,(-2,%a0)
1092 	clr.l	(%a0)+
1093 	clr.l	(%a0)
1094 2:	subq.l	#8,%a0
1095 	printf	PNORM,"%p(",1,%a0
1096 	printx	PNORM,%a0@
1097 	printf	PNORM,")\n"
1098 	rts
1099 3:	subq.w	#2,%d2
1100 	jcs	5f			| %d2 < 2, round to zero
1101 	jhi	4f			| %d2 > 2, round to +infinity
1102 	tst.b	(-3,%a0)		| to -inf
1103 	jne	1b
1104 	jra	5f
1105 4:	tst.b	(-3,%a0)		| to +inf
1106 	jeq	1b
1107 5:	move.w	#0x407e,(-2,%a0)
1108 	move.l	#0xffffff00,(%a0)+
1109 	clr.l	(%a0)
1110 	jra	2b
1111 	| Infinities or NaNs
1112 fp_nsf_huge:
1113 	subq.l	#4,%a0
1114 	printf	PNORM,"%p(",1,%a0
1115 	printx	PNORM,%a0@
1116 	printf	PNORM,")\n"
1117 	rts
1118 
1119 	| conv_ext2int (macro):
1120 	| Generates a subroutine that converts an extended value to an
1121 	| integer of a given size, again, with the appropriate type of
1122 	| rounding.
1123 
1124 	| Macro arguments:
1125 	| s:	size, as given in an assembly instruction.
1126 	| b:	number of bits in that size.
1127 
1128 	| Subroutine arguments:
1129 	| %a0:	source (struct fp_ext *)
1130 
1131 	| Returns the integer in %d0 (like it should)
1132 
1133 .macro conv_ext2int s,b
1134 	.set	inf,(1<<(\b-1))-1	| i.e. MAXINT
1135 	printf	PCONV,"e2i%d: %p(",2,#\b,%a0
1136 	printx	PCONV,%a0@
1137 	printf	PCONV,") "
1138 	addq.l	#2,%a0
1139 	move.w	(%a0)+,%d2		| exponent
1140 	jeq	fp_e2i_zero\b		| zero / denorm (== 0, here)
1141 	cmp.w	#0x7fff,%d2
1142 	jeq	fp_e2i_huge\b		| Inf / NaN
1143 	sub.w	#0x3ffe,%d2
1144 	jcs	fp_e2i_small\b
1145 	cmp.w	#\b,%d2
1146 	jhi	fp_e2i_large\b
1147 	move.l	(%a0),%d0
1148 	move.l	%d0,%d1
1149 	lsl.l	%d2,%d1
1150 	jne	fp_e2i_round\b
1151 	tst.l	(4,%a0)
1152 	jne	fp_e2i_round\b
1153 	neg.w	%d2
1154 	add.w	#32,%d2
1155 	lsr.l	%d2,%d0
1156 9:	tst.w	(-4,%a0)
1157 	jne	1f
1158 	tst.\s	%d0
1159 	jmi	fp_e2i_large\b
1160 	printf	PCONV,"-> %p\n",1,%d0
1161 	rts
1162 1:	neg.\s	%d0
1163 	jeq	1f
1164 	jpl	fp_e2i_large\b
1165 1:	printf	PCONV,"-> %p\n",1,%d0
1166 	rts
1167 fp_e2i_round\b:
1168 	fp_set_sr FPSR_EXC_INEX2	| INEX2 bit
1169 	neg.w	%d2
1170 	add.w	#32,%d2
1171 	.if	\b>16
1172 	jeq	5f
1173 	.endif
1174 	lsr.l	%d2,%d0
1175 	move.w	(FPD_RND,FPDATA),%d2	| rounding mode
1176 	jne	2f			| %d2 == 0, round to nearest
1177 	tst.l	%d1			| test guard bit
1178 	jpl	9b			| zero is closer
1179 	btst	%d2,%d0			| test lsb bit (%d2 still 0)
1180 	jne	fp_e2i_doroundup\b
1181 	lsl.l	#1,%d1			| check low bits
1182 	jne	fp_e2i_doroundup\b
1183 	tst.l	(4,%a0)
1184 	jeq	9b
1185 fp_e2i_doroundup\b:
1186 	addq.l	#1,%d0
1187 	jra	9b
1188 	| check nondefault rounding modes
1189 2:	subq.w	#2,%d2
1190 	jcs	9b			| %d2 < 2, round to zero
1191 	jhi	3f			| %d2 > 2, round to +infinity
1192 	tst.w	(-4,%a0)		| to -inf
1193 	jne	fp_e2i_doroundup\b	| negative, round to infinity
1194 	jra	9b			| positive, round to zero
1195 3:	tst.w	(-4,%a0)		| to +inf
1196 	jeq	fp_e2i_doroundup\b	| positive, round to infinity
1197 	jra	9b	| negative, round to zero
1198 	| we are only want -2**127 get correctly rounded here,
1199 	| since the guard bit is in the lower lword.
1200 	| everything else ends up anyway as overflow.
1201 	.if	\b>16
1202 5:	move.w	(FPD_RND,FPDATA),%d2	| rounding mode
1203 	jne	2b			| %d2 == 0, round to nearest
1204 	move.l	(4,%a0),%d1		| test guard bit
1205 	jpl	9b			| zero is closer
1206 	lsl.l	#1,%d1			| check low bits
1207 	jne	fp_e2i_doroundup\b
1208 	jra	9b
1209 	.endif
1210 fp_e2i_zero\b:
1211 	clr.l	%d0
1212 	tst.l	(%a0)+
1213 	jne	1f
1214 	tst.l	(%a0)
1215 	jeq	3f
1216 1:	subq.l	#4,%a0
1217 	fp_clr_sr FPSR_EXC_UNFL		| fp_normalize_ext has set this bit
1218 fp_e2i_small\b:
1219 	fp_set_sr FPSR_EXC_INEX2
1220 	clr.l	%d0
1221 	move.w	(FPD_RND,FPDATA),%d2	| rounding mode
1222 	subq.w	#2,%d2
1223 	jcs	3f			| %d2 < 2, round to nearest/zero
1224 	jhi	2f			| %d2 > 2, round to +infinity
1225 	tst.w	(-4,%a0)		| to -inf
1226 	jeq	3f
1227 	subq.\s	#1,%d0
1228 	jra	3f
1229 2:	tst.w	(-4,%a0)		| to +inf
1230 	jne	3f
1231 	addq.\s	#1,%d0
1232 3:	printf	PCONV,"-> %p\n",1,%d0
1233 	rts
1234 fp_e2i_large\b:
1235 	fp_set_sr FPSR_EXC_OPERR
1236 	move.\s	#inf,%d0
1237 	tst.w	(-4,%a0)
1238 	jeq	1f
1239 	addq.\s	#1,%d0
1240 1:	printf	PCONV,"-> %p\n",1,%d0
1241 	rts
1242 fp_e2i_huge\b:
1243 	move.\s	(%a0),%d0
1244 	tst.l	(%a0)
1245 	jne	1f
1246 	tst.l	(%a0)
1247 	jeq	fp_e2i_large\b
1248 	| fp_normalize_ext has set this bit already
1249 	| and made the number nonsignaling
1250 1:	fp_tst_sr FPSR_EXC_SNAN
1251 	jne	1f
1252 	fp_set_sr FPSR_EXC_OPERR
1253 1:	printf	PCONV,"-> %p\n",1,%d0
1254 	rts
1255 .endm
1256 
1257 fp_conv_ext2long:
1258 	conv_ext2int l,32
1259 
1260 fp_conv_ext2short:
1261 	conv_ext2int w,16
1262 
1263 fp_conv_ext2byte:
1264 	conv_ext2int b,8
1265 
1266 fp_conv_ext2double:
1267 	jsr	fp_normalize_double
1268 	printf	PCONV,"e2d: %p(",1,%a0
1269 	printx	PCONV,%a0@
1270 	printf	PCONV,"), "
1271 	move.l	(%a0)+,%d2
1272 	cmp.w	#0x7fff,%d2
1273 	jne	1f
1274 	move.w	#0x7ff,%d2
1275 	move.l	(%a0)+,%d0
1276 	jra	2f
1277 1:	sub.w	#0x3fff-0x3ff,%d2
1278 	move.l	(%a0)+,%d0
1279 	jmi	2f
1280 	clr.w	%d2
1281 2:	lsl.w	#5,%d2
1282 	lsl.l	#7,%d2
1283 	lsl.l	#8,%d2
1284 	move.l	%d0,%d1
1285 	lsl.l	#1,%d0
1286 	lsr.l	#4,%d0
1287 	lsr.l	#8,%d0
1288 	or.l	%d2,%d0
1289 	putuser.l %d0,(%a1)+,fp_err_ua2,%a1
1290 	moveq	#21,%d0
1291 	lsl.l	%d0,%d1
1292 	move.l	(%a0),%d0
1293 	lsr.l	#4,%d0
1294 	lsr.l	#7,%d0
1295 	or.l	%d1,%d0
1296 	putuser.l %d0,(%a1),fp_err_ua2,%a1
1297 #ifdef FPU_EMU_DEBUG
1298 	getuser.l %a1@(-4),%d0,fp_err_ua2,%a1
1299 	getuser.l %a1@(0),%d1,fp_err_ua2,%a1
1300 	printf	PCONV,"%p(%08x%08x)\n",3,%a1,%d0,%d1
1301 #endif
1302 	rts
1303 
1304 fp_conv_ext2single:
1305 	jsr	fp_normalize_single
1306 	printf	PCONV,"e2s: %p(",1,%a0
1307 	printx	PCONV,%a0@
1308 	printf	PCONV,"), "
1309 	move.l	(%a0)+,%d1
1310 	cmp.w	#0x7fff,%d1
1311 	jne	1f
1312 	move.w	#0xff,%d1
1313 	move.l	(%a0)+,%d0
1314 	jra	2f
1315 1:	sub.w	#0x3fff-0x7f,%d1
1316 	move.l	(%a0)+,%d0
1317 	jmi	2f
1318 	clr.w	%d1
1319 2:	lsl.w	#8,%d1
1320 	lsl.l	#7,%d1
1321 	lsl.l	#8,%d1
1322 	bclr	#31,%d0
1323 	lsr.l	#8,%d0
1324 	or.l	%d1,%d0
1325 	printf	PCONV,"%08x\n",1,%d0
1326 	rts
1327 
1328 	| special return addresses for instr that
1329 	| encode the rounding precision in the opcode
1330 	| (e.g. fsmove,fdmove)
1331 
1332 fp_finalrounding_single:
1333 	addq.l	#8,%sp
1334 	jsr	fp_normalize_ext
1335 	jsr	fp_normalize_single
1336 	jra	fp_finaltest
1337 
1338 fp_finalrounding_single_fast:
1339 	addq.l	#8,%sp
1340 	jsr	fp_normalize_ext
1341 	jsr	fp_normalize_single_fast
1342 	jra	fp_finaltest
1343 
1344 fp_finalrounding_double:
1345 	addq.l	#8,%sp
1346 	jsr	fp_normalize_ext
1347 	jsr	fp_normalize_double
1348 	jra	fp_finaltest
1349 
1350 	| fp_finaltest:
1351 	| set the emulated status register based on the outcome of an
1352 	| emulated instruction.
1353 
1354 fp_finalrounding:
1355 	addq.l	#8,%sp
1356 |	printf	,"f: %p\n",1,%a0
1357 	jsr	fp_normalize_ext
1358 	move.w	(FPD_PREC,FPDATA),%d0
1359 	subq.w	#1,%d0
1360 	jcs	fp_finaltest
1361 	jne	1f
1362 	jsr	fp_normalize_single
1363 	jra	2f
1364 1:	jsr	fp_normalize_double
1365 2:|	printf	,"f: %p\n",1,%a0
1366 fp_finaltest:
1367 	| First, we do some of the obvious tests for the exception
1368 	| status byte and condition code bytes of fp_sr here, so that
1369 	| they do not have to be handled individually by every
1370 	| emulated instruction.
1371 	clr.l	%d0
1372 	addq.l	#1,%a0
1373 	tst.b	(%a0)+			| sign
1374 	jeq	1f
1375 	bset	#FPSR_CC_NEG-24,%d0	| N bit
1376 1:	cmp.w	#0x7fff,(%a0)+		| exponent
1377 	jeq	2f
1378 	| test for zero
1379 	moveq	#FPSR_CC_Z-24,%d1
1380 	tst.l	(%a0)+
1381 	jne	9f
1382 	tst.l	(%a0)
1383 	jne	9f
1384 	jra	8f
1385 	| infinitiv and NAN
1386 2:	moveq	#FPSR_CC_NAN-24,%d1
1387 	move.l	(%a0)+,%d2
1388 	lsl.l	#1,%d2			| ignore high bit
1389 	jne	8f
1390 	tst.l	(%a0)
1391 	jne	8f
1392 	moveq	#FPSR_CC_INF-24,%d1
1393 8:	bset	%d1,%d0
1394 9:	move.b	%d0,(FPD_FPSR+0,FPDATA)	| set condition test result
1395 	| move instructions enter here
1396 	| Here, we test things in the exception status byte, and set
1397 	| other things in the accrued exception byte accordingly.
1398 	| Emulated instructions can set various things in the former,
1399 	| as defined in fp_emu.h.
1400 fp_final:
1401 	move.l	(FPD_FPSR,FPDATA),%d0
1402 #if 0
1403 	btst	#FPSR_EXC_SNAN,%d0	| EXC_SNAN
1404 	jne	1f
1405 	btst	#FPSR_EXC_OPERR,%d0	| EXC_OPERR
1406 	jeq	2f
1407 1:	bset	#FPSR_AEXC_IOP,%d0	| set IOP bit
1408 2:	btst	#FPSR_EXC_OVFL,%d0	| EXC_OVFL
1409 	jeq	1f
1410 	bset	#FPSR_AEXC_OVFL,%d0	| set OVFL bit
1411 1:	btst	#FPSR_EXC_UNFL,%d0	| EXC_UNFL
1412 	jeq	1f
1413 	btst	#FPSR_EXC_INEX2,%d0	| EXC_INEX2
1414 	jeq	1f
1415 	bset	#FPSR_AEXC_UNFL,%d0	| set UNFL bit
1416 1:	btst	#FPSR_EXC_DZ,%d0	| EXC_INEX1
1417 	jeq	1f
1418 	bset	#FPSR_AEXC_DZ,%d0	| set DZ bit
1419 1:	btst	#FPSR_EXC_OVFL,%d0	| EXC_OVFL
1420 	jne	1f
1421 	btst	#FPSR_EXC_INEX2,%d0	| EXC_INEX2
1422 	jne	1f
1423 	btst	#FPSR_EXC_INEX1,%d0	| EXC_INEX1
1424 	jeq	2f
1425 1:	bset	#FPSR_AEXC_INEX,%d0	| set INEX bit
1426 2:	move.l	%d0,(FPD_FPSR,FPDATA)
1427 #else
1428 	| same as above, greatly optimized, but untested (yet)
1429 	move.l	%d0,%d2
1430 	lsr.l	#5,%d0
1431 	move.l	%d0,%d1
1432 	lsr.l	#4,%d1
1433 	or.l	%d0,%d1
1434 	and.b	#0x08,%d1
1435 	move.l	%d2,%d0
1436 	lsr.l	#6,%d0
1437 	or.l	%d1,%d0
1438 	move.l	%d2,%d1
1439 	lsr.l	#4,%d1
1440 	or.b	#0xdf,%d1
1441 	and.b	%d1,%d0
1442 	move.l	%d2,%d1
1443 	lsr.l	#7,%d1
1444 	and.b	#0x80,%d1
1445 	or.b	%d1,%d0
1446 	and.b	#0xf8,%d0
1447 	or.b	%d0,%d2
1448 	move.l	%d2,(FPD_FPSR,FPDATA)
1449 #endif
1450 	move.b	(FPD_FPSR+2,FPDATA),%d0
1451 	and.b	(FPD_FPCR+2,FPDATA),%d0
1452 	jeq	1f
1453 	printf	,"send signal!!!\n"
1454 1:	jra	fp_end
1455