aboutsummaryrefslogtreecommitdiffstats
path: root/arch/m68k/math-emu/multi_arith.h
blob: 02251e5afd891f83bcfc304ed8f34a975f2cb809 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
/* multi_arith.h: multi-precision integer arithmetic functions, needed
   to do extended-precision floating point.

   (c) 1998 David Huggins-Daines.

   Somewhat based on arch/alpha/math-emu/ieee-math.c, which is (c)
   David Mosberger-Tang.

   You may copy, modify, and redistribute this file under the terms of
   the GNU General Public License, version 2, or any later version, at
   your convenience. */

/* Note:

   These are not general multi-precision math routines.  Rather, they
   implement the subset of integer arithmetic that we need in order to
   multiply, divide, and normalize 128-bit unsigned mantissae.  */

#ifndef MULTI_ARITH_H
#define MULTI_ARITH_H

#if 0	/* old code... */

/* Unsigned only, because we don't need signs to multiply and divide. */
typedef unsigned int int128[4];

/* Word order */
enum {
	MSW128,
	NMSW128,
	NLSW128,
	LSW128
};

/* big-endian */
#define LO_WORD(ll) (((unsigned int *) &ll)[1])
#define HI_WORD(ll) (((unsigned int *) &ll)[0])

/* Convenience functions to stuff various integer values into int128s */

static inline void zero128(int128 a)
{
	a[LSW128] = a[NLSW128] = a[NMSW128] = a[MSW128] = 0;
}

/* Human-readable word order in the arguments */
static inline void set128(unsigned int i3, unsigned int i2, unsigned int i1,
			  unsigned int i0, int128 a)
{
	a[LSW128] = i0;
	a[NLSW128] = i1;
	a[NMSW128] = i2;
	a[MSW128] = i3;
}

/* Convenience functions (for testing as well) */
static inline void int64_to_128(unsigned long long src, int128 dest)
{
	dest[LSW128] = (unsigned int) src;
	dest[NLSW128] = src >> 32;
	dest[NMSW128] = dest[MSW128] = 0;
}

static inline void int128_to_64(const int128 src, unsigned long long *dest)
{
	*dest = src[LSW128] | (long long) src[NLSW128] << 32;
}

static inline void put_i128(const int128 a)
{
	printk("%08x %08x %08x %08x\n", a[MSW128], a[NMSW128],
	       a[NLSW128], a[LSW128]);
}

/* Internal shifters:

   Note that these are only good for 0 < count < 32.
 */

static inline void _lsl128(unsigned int count, int128 a)
{
	a[MSW128] = (a[MSW128] << count) | (a[NMSW128] >> (32 - count));
	a[NMSW128] = (a[NMSW128] << count) | (a[NLSW128] >> (32 - count));
	a[NLSW128] = (a[NLSW128] << count) | (a[LSW128] >> (32 - count));
	a[LSW128] <<= count;
}

static inline void _lsr128(unsigned int count, int128 a)
{
	a[LSW128] = (a[LSW128] >> count) | (a[NLSW128] << (32 - count));
	a[NLSW128] = (a[NLSW128] >> count) | (a[NMSW128] << (32 - count));
	a[NMSW128] = (a[NMSW128] >> count) | (a[MSW128] << (32 - count));
	a[MSW128] >>= count;
}

/* Should be faster, one would hope */

static inline void lslone128(int128 a)
{
	asm volatile ("lsl.l #1,%0\n"
		      "roxl.l #1,%1\n"
		      "roxl.l #1,%2\n"
		      "roxl.l #1,%3\n"
		      :
		      "=d" (a[LSW128]),
		      "=d"(a[NLSW128]),
		      "=d"(a[NMSW128]),
		      "=d"(a[MSW128])
		      :
		      "0"(a[LSW128]),
		      "1"(a[NLSW128]),
		      "2"(a[NMSW128]),
		      "3"(a[MSW128]));
}

static inline void lsrone128(int128 a)
{
	asm volatile ("lsr.l #1,%0\n"
		      "roxr.l #1,%1\n"
		      "roxr.l #1,%2\n"
		      "roxr.l #1,%3\n"
		      :
		      "=d" (a[MSW128]),
		      "=d"(a[NMSW128]),
		      "=d"(a[NLSW128]),
		      "=d"(a[LSW128])
		      :
		      "0"(a[MSW128]),
		      "1"(a[NMSW128]),
		      "2"(a[NLSW128]),
		      "3"(a[LSW128]));
}

/* Generalized 128-bit shifters:

   These bit-shift to a multiple of 32, then move whole longwords.  */

static inline void lsl128(unsigned int count, int128 a)
{
	int wordcount, i;

	if (count % 32)
		_lsl128(count % 32, a);

	if (0 == (wordcount = count / 32))
		return;

	/* argh, gak, endian-sensitive */
	for (i = 0; i < 4 - wordcount; i++) {
		a[i] = a[i + wordcount];
	}
	for (i = 3; i >= 4 - wordcount; --i) {
		a[i] = 0;
	}
}

static inline void lsr128(unsigned int count, int128 a)
{
	int wordcount, i;

	if (count % 32)
		_lsr128(count % 32, a);

	if (0 == (wordcount = count / 32))
		return;

	for (i = 3; i >= wordcount; --i) {
		a[i] = a[i - wordcount];
	}
	for (i = 0; i < wordcount; i++) {
		a[i] = 0;
	}
}

static inline int orl128(int a, int128 b)
{
	b[LSW128] |= a;
}

static inline int btsthi128(const int128 a)
{
	return a[MSW128] & 0x80000000;
}

/* test bits (numbered from 0 = LSB) up to and including "top" */
static inline int bftestlo128(int top, const int128 a)
{
	int r = 0;

	if (top > 31)
		r |= a[LSW128];
	if (top > 63)
		r |= a[NLSW128];
	if (top > 95)
		r |= a[NMSW128];

	r |= a[3 - (top / 32)] & ((1 << (top % 32 + 1)) - 1);

	return (r != 0);
}

/* Aargh.  We need these because GCC is broken */
/* FIXME: do them in assembly, for goodness' sake! */
static inline void mask64(int pos, unsigned long long *mask)
{
	*mask = 0;

	if (pos < 32) {
		LO_WORD(*mask) = (1 << pos) - 1;
		return;
	}
	LO_WORD(*mask) = -1;
	HI_WORD(*mask) = (1 << (pos - 32)) - 1;
}

static inline void bset64(int pos, unsigned long long *dest)
{
	/* This conditional will be optimized away.  Thanks, GCC! */
	if (pos < 32)
		asm volatile ("bset %1,%0":"=m"
			      (LO_WORD(*dest)):"id"(pos));
	else
		asm volatile ("bset %1,%0":"=m"
			      (HI_WORD(*dest)):"id"(pos - 32));
}

static inline int btst64(int pos, unsigned long long dest)
{
	if (pos < 32)
		return (0 != (LO_WORD(dest) & (1 << pos)));
	else
		return (0 != (HI_WORD(dest) & (1 << (pos - 32))));
}

static inline void lsl64(int count, unsigned long long *dest)
{
	if (count < 32) {
		HI_WORD(*dest) = (HI_WORD(*dest) << count)
		    | (LO_WORD(*dest) >> count);
		LO_WORD(*dest) <<= count;
		return;
	}
	count -= 32;
	HI_WORD(*dest) = LO_WORD(*dest) << count;
	LO_WORD(*dest) = 0;
}

static inline void lsr64(int count, unsigned long long *dest)
{
	if (count < 32) {
		LO_WORD(*dest) = (LO_WORD(*dest) >> count)
		    | (HI_WORD(*dest) << (32 - count));
		HI_WORD(*dest) >>= count;
		return;
	}
	count -= 32;
	LO_WORD(*dest) = HI_WORD(*dest) >> count;
	HI_WORD(*dest) = 0;
}
#endif

static inline void fp_denormalize(struct fp_ext *reg, unsigned int cnt)
{
	reg->exp += cnt;

	switch (cnt) {
	case 0 ... 8:
		reg->lowmant = reg->mant.m32[1] << (8 - cnt);
		reg->mant.m32[1] = (reg->mant.m32[1] >> cnt) |
				   (reg->mant.m32[0] << (32 - cnt));
		reg->mant.m32[0] = reg->mant.m32[0] >> cnt;
		break;
	case 9 ... 32:
		reg->lowmant = reg->mant.m32[1] >> (cnt - 8);
		if (reg->mant.m32[1] << (40 - cnt))
			reg->lowmant |= 1;
		reg->mant.m32[1] = (reg->mant.m32[1] >> cnt) |
				   (reg->mant.m32[0] << (32 - cnt));
		reg->mant.m32[0] = reg->mant.m32[0] >> cnt;
		break;
	case 33 ... 39:
		asm volatile ("bfextu %1{%2,#8},%0" : "=d" (reg->lowmant)
			: "m" (reg->mant.m32[0]), "d" (64 - cnt));
		if (reg->mant.m32[1] << (40 - cnt))
			reg->lowmant |= 1;
		reg->mant.m32[1] = reg->mant.m32[0] >> (cnt - 32);
		reg->mant.m32[0] = 0;
		break;
	case 40 ... 71:
		reg->lowmant = reg->mant.m32[0] >> (cnt - 40);
		if ((reg->mant.m32[0] << (72 - cnt)) || reg->mant.m32[1])
			reg->lowmant |= 1;
		reg->mant.m32[1] = reg->mant.m32[0] >> (cnt - 32);
		reg->mant.m32[0] = 0;
		break;
	default:
		reg->lowmant = reg->mant.m32[0] || reg->mant.m32[1];
		reg->mant.m32[0] = 0;
		reg->mant.m32[1] = 0;
		break;
	}
}

static inline int fp_overnormalize(struct fp_ext *reg)
{
	int shift;

	if (reg->mant.m32[0]) {
		asm ("bfffo %1{#0,#32},%0" : "=d" (shift) : "dm" (reg->mant.m32[0]));
		reg->mant.m32[0] = (reg->mant.m32[0] << shift) | (reg->mant.m32[1] >> (32 - shift));
		reg->mant.m32[1] = (reg->mant.m32[1] << shift);
	} else {
		asm ("bfffo %1{#0,#32},%0" : "=d" (shift) : "dm" (reg->mant.m32[1]));
		reg->mant.m32[0] = (reg->mant.m32[1] << shift);
		reg->mant.m32[1] = 0;
		shift += 32;
	}

	return shift;
}

static inline int fp_addmant(struct fp_ext *dest, struct fp_ext *src)
{
	int carry;

	/* we assume here, gcc only insert move and a clr instr */
	asm volatile ("add.b %1,%0" : "=d,g" (dest->lowmant)
		: "g,d" (src->lowmant), "0,0" (dest->lowmant));
	asm volatile ("addx.l %1,%0" : "=d" (dest->mant.m32[1])
		: "d" (src->mant.m32[1]), "0" (dest->mant.m32[1]));
	asm volatile ("addx.l %1,%0" : "=d" (dest->mant.m32[0])
		: "d" (src->mant.m32[0]), "0" (dest->mant.m32[0]));
	asm volatile ("addx.l %0,%0" : "=d" (carry) : "0" (0));

	return carry;
}

static inline int fp_addcarry(struct fp_ext *reg)
{
	if (++reg->exp == 0x7fff) {
		if (reg->mant.m64)
			fp_set_sr(FPSR_EXC_INEX2);
		reg->mant.m64 = 0;
		fp_set_sr(FPSR_EXC_OVFL);
		return 0;
	}
	reg->lowmant = (reg->mant.m32[1] << 7) | (reg->lowmant ? 1 : 0);
	reg->mant.m32[1] = (reg->mant.m32[1] >> 1) |
			   (reg->mant.m32[0] << 31);
	reg->mant.m32[0] = (reg->mant.m32[0] >> 1) | 0x80000000;

	return 1;
}

static inline void fp_submant(struct fp_ext *dest, struct fp_ext *src1,
			      struct fp_ext *src2)
{
	/* we assume here, gcc only insert move and a clr instr */
	asm volatile ("sub.b %1,%0" : "=d,g" (dest->lowmant)
		: "g,d" (src2->lowmant), "0,0" (src1->lowmant));
	asm volatile ("subx.l %1,%0" : "=d" (dest->mant.m32[1])
		: "d" (src2->mant.m32[1]), "0" (src1->mant.m32[1]));
	asm volatile ("subx.l %1,%0" : "=d" (dest->mant.m32[0])
		: "d" (src2->mant.m32[0]), "0" (src1->mant.m32[0]));
}

#define fp_mul64(desth, destl, src1, src2) ({				\
	asm ("mulu.l %2,%1:%0" : "=d" (destl), "=d" (desth)		\
		: "g" (src1), "0" (src2));				\
})
#define fp_div64(quot, rem, srch, srcl, div)				\
	asm ("divu.l %2,%1:%0" : "=d" (quot), "=d" (rem)		\
		: "dm" (div), "1" (srch), "0" (srcl))
#define fp_add64(dest1, dest2, src1, src2) ({				\
	asm ("add.l %1,%0" : "=d,dm" (dest2)				\
		: "dm,d" (src2), "0,0" (dest2));			\
	asm ("addx.l %1,%0" : "=d" (dest1)				\
		: "d" (src1), "0" (dest1));				\
})
#define fp_addx96(dest, src) ({						\
	/* we assume here, gcc only insert move and a clr instr */	\
	asm volatile ("add.l %1,%0" : "=d,g" (dest->m32[2])		\
		: "g,d" (temp.m32[1]), "0,0" (dest->m32[2]));		\
	asm volatile ("addx.l %1,%0" : "=d" (dest->m32[1])		\
		: "d" (temp.m32[0]), "0" (dest->m32[1]));		\
	asm volatile ("addx.l %1,%0" : "=d" (dest->m32[0])		\
		: "d" (0), "0" (dest->m32[0]));				\
})
#define fp_sub64(dest, src) ({						\
	asm ("sub.l %1,%0" : "=d,dm" (dest.m32[1])			\
		: "dm,d" (src.m32[1]), "0,0" (dest.m32[1]));		\
	asm ("subx.l %1,%0" : "=d" (dest.m32[0])			\
		: "d" (src.m32[0]), "0" (dest.m32[0]));			\
})
#define fp_sub96c(dest, srch, srcm, srcl) ({				\
	char carry;							\
	asm ("sub.l %1,%0" : "=d,dm" (dest.m32[2])			\
		: "dm,d" (srcl), "0,0" (dest.m32[2]));			\
	asm ("subx.l %1,%0" : "=d" (dest.m32[1])			\
		: "d" (srcm), "0" (dest.m32[1]));			\
	asm ("subx.l %2,%1; scs %0" : "=d" (carry), "=d" (dest.m32[0])	\
		: "d" (srch), "1" (dest.m32[0]));			\
	carry;								\
})

static inline void fp_multiplymant(union fp_mant128 *dest, struct fp_ext *src1,
				   struct fp_ext *src2)
{
	union fp_mant64 temp;

	fp_mul64(dest->m32[0], dest->m32[1], src1->mant.m32[0], src2->mant.m32[0]);
	fp_mul64(dest->m32[2], dest->m32[3], src1->mant.m32[1], src2->mant.m32[1]);

	fp_mul64(temp.m32[0], temp.m32[1], src1->mant.m32[0], src2->mant.m32[1]);
	fp_addx96(dest, temp);

	fp_mul64(temp.m32[0], temp.m32[1], src1->mant.m32[1], src2->mant.m32[0]);
	fp_addx96(dest, temp);
}

static inline void fp_dividemant(union fp_mant128 *dest, struct fp_ext *src,
				 struct fp_ext *div)
{
	union fp_mant128 tmp;
	union fp_mant64 tmp64;
	unsigned long *mantp = dest->m32;
	unsigned long fix, rem, first, dummy;
	int i;

	/* the algorithm below requires dest to be smaller than div,
	   but both have the high bit set */
	if (src->mant.m64 >= div->mant.m64) {
		fp_sub64(src->mant, div->mant);
		*mantp = 1;
	} else
		*mantp = 0;
	mantp++;

	/* basic idea behind this algorithm: we can't divide two 64bit numbers
	   (AB/CD) directly, but we can calculate AB/C0, but this means this
	   quotient is off by C0/CD, so we have to multiply the first result
	   to fix the result, after that we have nearly the correct result
	   and only a few corrections are needed. */

	/* C0/CD can be precalculated, but it's an 64bit division again, but
	   we can make it a bit easier, by dividing first through C so we get
	   10/1D and now only a single shift and the value fits into 32bit. */
	fix = 0x80000000;
	dummy = div->mant.m32[1] / div->mant.m32[0] + 1;
	dummy = (dummy >> 1) | fix;
	fp_div64(fix, dummy, fix, 0, dummy);
	fix--;

	for (i = 0; i < 3; i++, mantp++) {
		if (src->mant.m32[0] == div->mant.m32[0]) {
			fp_div64(first, rem, 0, src->mant.m32[1], div->mant.m32[0]);

			fp_mul64(*mantp, dummy, first, fix);
			*mantp += fix;
		} else {
			fp_div64(first, rem, src->mant.m32[0], src->mant.m32[1], div->mant.m32[0]);

			fp_mul64(*mantp, dummy, first, fix);
		}

		fp_mul64(tmp.m32[0], tmp.m32[1], div->mant.m32[0], first - *mantp);
		fp_add64(tmp.m32[0], tmp.m32[1], 0, rem);
		tmp.m32[2] = 0;

		fp_mul64(tmp64.m32[0], tmp64.m32[1], *mantp, div->mant.m32[1]);
		fp_sub96c(tmp, 0, tmp64.m32[0], tmp64.m32[1]);

		src->mant.m32[0] = tmp.m32[1];
		src->mant.m32[1] = tmp.m32[2];

		while (!fp_sub96c(tmp, 0, div->mant.m32[0], div->mant.m32[1])) {
			src->mant.m32[0] = tmp.m32[1];
			src->mant.m32[1] = tmp.m32[2];
			*mantp += 1;
		}
	}
}

#if 0
static inline unsigned int fp_fls128(union fp_mant128 *src)
{
	unsigned long data;
	unsigned int res, off;

	if ((data = src->m32[0]))
		off = 0;
	else if ((data = src->m32[1]))
		off = 32;
	else if ((data = src->m32[2]))
		off = 64;
	else if ((data = src->m32[3]))
		off = 96;
	else
		return 128;

	asm ("bfffo %1{#0,#32},%0" : "=d" (res) : "dm" (data));
	return res + off;
}

static inline void fp_shiftmant128(union fp_mant128 *src, int shift)
{
	unsigned long sticky;

	switch (shift) {
	case 0:
		return;
	case 1:
		asm volatile ("lsl.l #1,%0"
			: "=d" (src->m32[3]) : "0" (src->m32[3]));
		asm volatile ("roxl.l #1,%0"
			: "=d" (src->m32[2]) : "0" (src->m32[2]));
		asm volatile ("roxl.l #1,%0"
			: "=d" (src->m32[1]) : "0" (src->m32[1]));
		asm volatile ("roxl.l #1,%0"
			: "=d" (src->m32[0]) : "0" (src->m32[0]));
		return;
	case 2 ... 31:
		src->m32[0] = (src->m32[0] << shift) | (src->m32[1] >> (32 - shift));
		src->m32[1] = (src->m32[1] << shift) | (src->m32[2] >> (32 - shift));
		src->m32[2] = (src->m32[2] << shift) | (src->m32[3] >> (32 - shift));
		src->m32[3] = (src->m32[3] << shift);
		return;
	case 32 ... 63:
		shift -= 32;
		src->m32[0] = (src->m32[1] << shift) | (src->m32[2] >> (32 - shift));
		src->m32[1] = (src->m32[2] << shift) | (src->m32[3] >> (32 - shift));
		src->m32[2] = (src->m32[3] << shift);
		src->m32[3] = 0;
		return;
	case 64 ... 95:
		shift -= 64;
		src->m32[0] = (src->m32[2] << shift) | (src->m32[3] >> (32 - shift));
		src->m32[1] = (src->m32[3] << shift);
		src->m32[2] = src->m32[3] = 0;
		return;
	case 96 ... 127:
		shift -= 96;
		src->m32[0] = (src->m32[3] << shift);
		src->m32[1] = src->m32[2] = src->m32[3] = 0;
		return;
	case -31 ... -1:
		shift = -shift;
		sticky = 0;
		if (src->m32[3] << (32 - shift))
			sticky = 1;
		src->m32[3] = (src->m32[3] >> shift) | (src->m32[2] << (32 - shift)) | sticky;
		src->m32[2] = (src->m32[2] >> shift) | (src->m32[1] << (32 - shift));
		src->m32[1] = (src->m32[1] >> shift) | (src->m32[0] << (32 - shift));
		src->m32[0] = (src->m32[0] >> shift);
		return;
	case -63 ... -32:
		shift = -shift - 32;
		sticky = 0;
		if ((src->m32[2] << (32 - shift)) || src->m32[3])
			sticky = 1;
		src->m32[3] = (src->m32[2] >> shift) | (src->m32[1] << (32 - shift)) | sticky;
		src->m32[2] = (src->m32[1] >> shift) | (src->m32[0] << (32 - shift));
		src->m32[1] = (src->m32[0] >> shift);
		src->m32[0] = 0;
		return;
	case -95 ... -64:
		shift = -shift - 64;
		sticky = 0;
		if ((src->m32[1] << (32 - shift)) || src->m32[2] || src->m32[3])
			sticky = 1;
		src->m32[3] = (src->m32[1] >> shift) | (src->m32[0] << (32 - shift)) | sticky;
		src->m32[2] = (src->m32[0] >> shift);
		src->m32[1] = src->m32[0] = 0;
		return;
	case -127 ... -96:
		shift = -shift - 96;
		sticky = 0;
		if ((src->m32[0] << (32 - shift)) || src->m32[1] || src->m32[2] || src->m32[3])
			sticky = 1;
		src->m32[3] = (src->m32[0] >> shift) | sticky;
		src->m32[2] = src->m32[1] = src->m32[0] = 0;
		return;
	}

	if (shift < 0 && (src->m32[0] || src->m32[1] || src->m32[2] || src->m32[3]))
		src->m32[3] = 1;
	else
		src->m32[3] = 0;
	src->m32[2] = 0;
	src->m32[1] = 0;
	src->m32[0] = 0;
}
#endif

static inline void fp_putmant128(struct fp_ext *dest, union fp_mant128 *src,
				 int shift)
{
	unsigned long tmp;

	switch (shift) {
	case 0:
		dest->mant.m64 = src->m64[0];
		dest->lowmant = src->m32[2] >> 24;
		if (src->m32[3] || (src->m32[2] << 8))
			dest->lowmant |= 1;
		break;
	case 1:
		asm volatile ("lsl.l #1,%0"
			: "=d" (tmp) : "0" (src->m32[2]));
		asm volatile ("roxl.l #1,%0"
			: "=d" (dest->mant.m32[1]) : "0" (src->m32[1]));
		asm volatile ("roxl.l #1,%0"
			: "=d" (dest->mant.m32[0]) : "0" (src->m32[0]));
		dest->lowmant = tmp >> 24;
		if (src->m32[3] || (tmp << 8))
			dest->lowmant |= 1;
		break;
	case 31:
		asm volatile ("lsr.l #1,%1; roxr.l #1,%0"
			: "=d" (dest->mant.m32[0])
			: "d" (src->m32[0]), "0" (src->m32[1]));
		asm volatile ("roxr.l #1,%0"
			: "=d" (dest->mant.m32[1]) : "0" (src->m32[2]));
		asm volatile ("roxr.l #1,%0"
			: "=d" (tmp) : "0" (src->m32[3]));
		dest->lowmant = tmp >> 24;
		if (src->m32[3] << 7)
			dest->lowmant |= 1;
		break;
	case 32:
		dest->mant.m32[0] = src->m32[1];
		dest->mant.m32[1] = src->m32[2];
		dest->lowmant = src->m32[3] >> 24;
		if (src->m32[3] << 8)
			dest->lowmant |= 1;
		break;
	}
}

#if 0 /* old code... */
static inline int fls(unsigned int a)
{
	int r;

	asm volatile ("bfffo %1{#0,#32},%0"
		      : "=d" (r) : "md" (a));
	return r;
}

/* fls = "find last set" (cf. ffs(3)) */
static inline int fls128(const int128 a)
{
	if (a[MSW128])
		return fls(a[MSW128]);
	if (a[NMSW128])
		return fls(a[NMSW128]) + 32;
	/* XXX: it probably never gets beyond this point in actual
	   use, but that's indicative of a more general problem in the
	   algorithm (i.e. as per the actual 68881 implementation, we
	   really only need at most 67 bits of precision [plus
	   overflow]) so I'm not going to fix it. */
	if (a[NLSW128])
		return fls(a[NLSW128]) + 64;
	if (a[LSW128])
		return fls(a[LSW128]) + 96;
	else
		return -1;
}

static inline int zerop128(const int128 a)
{
	return !(a[LSW128] | a[NLSW128] | a[NMSW128] | a[MSW128]);
}

static inline int nonzerop128(const int128 a)
{
	return (a[LSW128] | a[NLSW128] | a[NMSW128] | a[MSW128]);
}

/* Addition and subtraction */
/* Do these in "pure" assembly, because "extended" asm is unmanageable
   here */
static inline void add128(const int128 a, int128 b)
{
	/* rotating carry flags */
	unsigned int carry[2];

	carry[0] = a[LSW128] > (0xffffffff - b[LSW128]);
	b[LSW128] += a[LSW128];

	carry[1] = a[NLSW128] > (0xffffffff - b[NLSW128] - carry[0]);
	b[NLSW128] = a[NLSW128] + b[NLSW128] + carry[0];

	carry[0] = a[NMSW128] > (0xffffffff - b[NMSW128] - carry[1]);
	b[NMSW128] = a[NMSW128] + b[NMSW128] + carry[1];

	b[MSW128] = a[MSW128] + b[MSW128] + carry[0];
}

/* Note: assembler semantics: "b -= a" */
static inline void sub128(const int128 a, int128 b)
{
	/* rotating borrow flags */
	unsigned int borrow[2];

	borrow[0] = b[LSW128] < a[LSW128];
	b[LSW128] -= a[LSW128];

	borrow[1] = b[NLSW128] < a[NLSW128] + borrow[0];
	b[NLSW128] = b[NLSW128] - a[NLSW128] - borrow[0];

	borrow[0] = b[NMSW128] < a[NMSW128] + borrow[1];
	b[NMSW128] = b[NMSW128] - a[NMSW128] - borrow[1];

	b[MSW128] = b[MSW128] - a[MSW128] - borrow[0];
}

/* Poor man's 64-bit expanding multiply */
static inline void mul64(unsigned long long a, unsigned long long b, int128 c)
{
	unsigned long long acc;
	int128 acc128;

	zero128(acc128);
	zero128(c);

	/* first the low words */
	if (LO_WORD(a) && LO_WORD(b)) {
		acc = (long long) LO_WORD(a) * LO_WORD(b);
		c[NLSW128] = HI_WORD(acc);
		c[LSW128] = LO_WORD(acc);
	}
	/* Next the high words */
	if (HI_WORD(a) && HI_WORD(b)) {
		acc = (long long) HI_WORD(a) * HI_WORD(b);
		c[MSW128] = HI_WORD(acc);
		c[NMSW128] = LO_WORD(acc);
	}
	/* The middle words */
	if (LO_WORD(a) && HI_WORD(b)) {
		acc = (long long) LO_WORD(a) * HI_WORD(b);
		acc128[NMSW128] = HI_WORD(acc);
		acc128[NLSW128] = LO_WORD(acc);
		add128(acc128, c);
	}
	/* The first and last words */
	if (HI_WORD(a) && LO_WORD(b)) {
		acc = (long long) HI_WORD(a) * LO_WORD(b);
		acc128[NMSW128] = HI_WORD(acc);
		acc128[NLSW128] = LO_WORD(acc);
		add128(acc128, c);
	}
}

/* Note: unsigned */
static inline int cmp128(int128 a, int128 b)
{
	if (a[MSW128] < b[MSW128])
		return -1;
	if (a[MSW128] > b[MSW128])
		return 1;
	if (a[NMSW128] < b[NMSW128])
		return -1;
	if (a[NMSW128] > b[NMSW128])
		return 1;
	if (a[NLSW128] < b[NLSW128])
		return -1;
	if (a[NLSW128] > b[NLSW128])
		return 1;

	return (signed) a[LSW128] - b[LSW128];
}

inline void div128(int128 a, int128 b, int128 c)
{
	int128 mask;

	/* Algorithm:

	   Shift the divisor until it's at least as big as the
	   dividend, keeping track of the position to which we've
	   shifted it, i.e. the power of 2 which we've multiplied it
	   by.

	   Then, for this power of 2 (the mask), and every one smaller
	   than it, subtract the mask from the dividend and add it to
	   the quotient until the dividend is smaller than the raised
	   divisor.  At this point, divide the dividend and the mask
	   by 2 (i.e. shift one place to the right).  Lather, rinse,
	   and repeat, until there are no more powers of 2 left. */

	/* FIXME: needless to say, there's room for improvement here too. */

	/* Shift up */
	/* XXX: since it just has to be "at least as big", we can
	   probably eliminate this horribly wasteful loop.  I will
	   have to prove this first, though */
	set128(0, 0, 0, 1, mask);
	while (cmp128(b, a) < 0 && !btsthi128(b)) {
		lslone128(b);
		lslone128(mask);
	}

	/* Shift down */
	zero128(c);
	do {
		if (cmp128(a, b) >= 0) {
			sub128(b, a);
			add128(mask, c);
		}
		lsrone128(mask);
		lsrone128(b);
	} while (nonzerop128(mask));

	/* The remainder is in a... */
}
#endif

#endif	/* MULTI_ARITH_H */