2 * linux/arch/arm/vfp/vfpsingle.c
4 * This code is derived in part from John R. Housers softfloat library, which
5 * carries the following notice:
7 * ===========================================================================
8 * This C source file is part of the SoftFloat IEC/IEEE Floating-point
9 * Arithmetic Package, Release 2.
11 * Written by John R. Hauser. This work was made possible in part by the
12 * International Computer Science Institute, located at Suite 600, 1947 Center
13 * Street, Berkeley, California 94704. Funding was partially provided by the
14 * National Science Foundation under grant MIP-9311980. The original version
15 * of this code was written as part of a project to build a fixed-point vector
16 * processor in collaboration with the University of California at Berkeley,
17 * overseen by Profs. Nelson Morgan and John Wawrzynek. More information
18 * is available through the web page `http://HTTP.CS.Berkeley.EDU/~jhauser/
19 * arithmetic/softfloat.html'.
21 * THIS SOFTWARE IS DISTRIBUTED AS IS, FOR FREE. Although reasonable effort
22 * has been made to avoid it, THIS SOFTWARE MAY CONTAIN FAULTS THAT WILL AT
23 * TIMES RESULT IN INCORRECT BEHAVIOR. USE OF THIS SOFTWARE IS RESTRICTED TO
24 * PERSONS AND ORGANIZATIONS WHO CAN AND WILL TAKE FULL RESPONSIBILITY FOR ANY
25 * AND ALL LOSSES, COSTS, OR OTHER PROBLEMS ARISING FROM ITS USE.
27 * Derivative works are acceptable, even for commercial purposes, so long as
28 * (1) they include prominent notice that the work is derivative, and (2) they
29 * include prominent notice akin to these three paragraphs for those parts of
30 * this code that are retained.
31 * ===========================================================================
33 #include <linux/kernel.h>
34 #include <linux/bitops.h>
36 #include <asm/div64.h>
37 #include <asm/ptrace.h>
43 static struct vfp_single vfp_single_default_qnan = {
46 .significand = VFP_SINGLE_SIGNIFICAND_QNAN,
49 static void vfp_single_dump(const char *str, struct vfp_single *s)
51 pr_debug("VFP: %s: sign=%d exponent=%d significand=%08x\n",
52 str, s->sign != 0, s->exponent, s->significand);
55 static void vfp_single_normalise_denormal(struct vfp_single *vs)
57 int bits = 31 - fls(vs->significand);
59 vfp_single_dump("normalise_denormal: in", vs);
62 vs->exponent -= bits - 1;
63 vs->significand <<= bits;
66 vfp_single_dump("normalise_denormal: out", vs);
70 #define vfp_single_normaliseround(sd,vsd,fpscr,except,func) __vfp_single_normaliseround(sd,vsd,fpscr,except)
71 u32 __vfp_single_normaliseround(int sd, struct vfp_single *vs, u32 fpscr, u32 exceptions)
73 u32 vfp_single_normaliseround(int sd, struct vfp_single *vs, u32 fpscr, u32 exceptions, const char *func)
76 u32 significand, incr, rmode;
77 int exponent, shift, underflow;
79 vfp_single_dump("pack: in", vs);
82 * Infinities and NaNs are a special case.
84 if (vs->exponent == 255 && (vs->significand == 0 || exceptions))
90 if (vs->significand == 0) {
95 exponent = vs->exponent;
96 significand = vs->significand;
99 * Normalise first. Note that we shift the significand up to
100 * bit 31, so we have VFP_SINGLE_LOW_BITS + 1 below the least
103 shift = 32 - fls(significand);
104 if (shift < 32 && shift) {
106 significand <<= shift;
110 vs->exponent = exponent;
111 vs->significand = significand;
112 vfp_single_dump("pack: normalised", vs);
118 underflow = exponent < 0;
120 significand = vfp_shiftright32jamming(significand, -exponent);
123 vs->exponent = exponent;
124 vs->significand = significand;
125 vfp_single_dump("pack: tiny number", vs);
127 if (!(significand & ((1 << (VFP_SINGLE_LOW_BITS + 1)) - 1)))
132 * Select rounding increment.
135 rmode = fpscr & FPSCR_RMODE_MASK;
137 if (rmode == FPSCR_ROUND_NEAREST) {
138 incr = 1 << VFP_SINGLE_LOW_BITS;
139 if ((significand & (1 << (VFP_SINGLE_LOW_BITS + 1))) == 0)
141 } else if (rmode == FPSCR_ROUND_TOZERO) {
143 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vs->sign != 0))
144 incr = (1 << (VFP_SINGLE_LOW_BITS + 1)) - 1;
146 pr_debug("VFP: rounding increment = 0x%08x\n", incr);
149 * Is our rounding going to overflow?
151 if ((significand + incr) < significand) {
153 significand = (significand >> 1) | (significand & 1);
156 vs->exponent = exponent;
157 vs->significand = significand;
158 vfp_single_dump("pack: overflow", vs);
163 * If any of the low bits (which will be shifted out of the
164 * number) are non-zero, the result is inexact.
166 if (significand & ((1 << (VFP_SINGLE_LOW_BITS + 1)) - 1))
167 exceptions |= FPSCR_IXC;
177 if (exponent >= 254) {
178 exceptions |= FPSCR_OFC | FPSCR_IXC;
181 vs->significand = 0x7fffffff;
183 vs->exponent = 255; /* infinity */
187 if (significand >> (VFP_SINGLE_LOW_BITS + 1) == 0)
189 if (exponent || significand > 0x80000000)
192 exceptions |= FPSCR_UFC;
193 vs->exponent = exponent;
194 vs->significand = significand >> 1;
198 vfp_single_dump("pack: final", vs);
200 s32 d = vfp_single_pack(vs);
202 pr_debug("VFP: %s: d(s%d)=%08x exceptions=%08x\n", func,
205 vfp_put_float(d, sd);
212 * Propagate the NaN, setting exceptions if it is signalling.
213 * 'n' is always a NaN. 'm' may be a number, NaN or infinity.
216 vfp_propagate_nan(struct vfp_single *vsd, struct vfp_single *vsn,
217 struct vfp_single *vsm, u32 fpscr)
219 struct vfp_single *nan;
222 tn = vfp_single_type(vsn);
225 tm = vfp_single_type(vsm);
227 if (fpscr & FPSCR_DEFAULT_NAN)
229 * Default NaN mode - always returns a quiet NaN
231 nan = &vfp_single_default_qnan;
234 * Contemporary mode - select the first signalling
235 * NAN, or if neither are signalling, the first
238 if (tn == VFP_SNAN || (tm != VFP_SNAN && tn == VFP_QNAN))
243 * Make the NaN quiet.
245 nan->significand |= VFP_SINGLE_SIGNIFICAND_QNAN;
251 * If one was a signalling NAN, raise invalid operation.
253 return tn == VFP_SNAN || tm == VFP_SNAN ? FPSCR_IOC : VFP_NAN_FLAG;
258 * Extended operations
260 static u32 vfp_single_fabs(int sd, int unused, s32 m, u32 fpscr)
262 vfp_put_float(vfp_single_packed_abs(m), sd);
266 static u32 vfp_single_fcpy(int sd, int unused, s32 m, u32 fpscr)
268 vfp_put_float(m, sd);
272 static u32 vfp_single_fneg(int sd, int unused, s32 m, u32 fpscr)
274 vfp_put_float(vfp_single_packed_negate(m), sd);
278 static const u16 sqrt_oddadjust[] = {
279 0x0004, 0x0022, 0x005d, 0x00b1, 0x011d, 0x019f, 0x0236, 0x02e0,
280 0x039c, 0x0468, 0x0545, 0x0631, 0x072b, 0x0832, 0x0946, 0x0a67
283 static const u16 sqrt_evenadjust[] = {
284 0x0a2d, 0x08af, 0x075a, 0x0629, 0x051a, 0x0429, 0x0356, 0x029e,
285 0x0200, 0x0179, 0x0109, 0x00af, 0x0068, 0x0034, 0x0012, 0x0002
288 u32 vfp_estimate_sqrt_significand(u32 exponent, u32 significand)
293 if ((significand & 0xc0000000) != 0x40000000) {
294 printk(KERN_WARNING "VFP: estimate_sqrt: invalid significand\n");
297 a = significand << 1;
298 index = (a >> 27) & 15;
300 z = 0x4000 + (a >> 17) - sqrt_oddadjust[index];
301 z = ((a / z) << 14) + (z << 15);
304 z = 0x8000 + (a >> 17) - sqrt_evenadjust[index];
306 z = (z >= 0x20000) ? 0xffff8000 : (z << 15);
311 u64 v = (u64)a << 31;
317 static u32 vfp_single_fsqrt(int sd, int unused, s32 m, u32 fpscr)
319 struct vfp_single vsm, vsd;
322 vfp_single_unpack(&vsm, m);
323 tm = vfp_single_type(&vsm);
324 if (tm & (VFP_NAN|VFP_INFINITY)) {
325 struct vfp_single *vsp = &vsd;
328 ret = vfp_propagate_nan(vsp, &vsm, NULL, fpscr);
329 else if (vsm.sign == 0) {
335 vsp = &vfp_single_default_qnan;
338 vfp_put_float(vfp_single_pack(vsp), sd);
343 * sqrt(+/- 0) == +/- 0
349 * Normalise a denormalised number
351 if (tm & VFP_DENORMAL)
352 vfp_single_normalise_denormal(&vsm);
360 vfp_single_dump("sqrt", &vsm);
363 * Estimate the square root.
366 vsd.exponent = ((vsm.exponent - 127) >> 1) + 127;
367 vsd.significand = vfp_estimate_sqrt_significand(vsm.exponent, vsm.significand) + 2;
369 vfp_single_dump("sqrt estimate", &vsd);
374 if ((vsd.significand & VFP_SINGLE_LOW_BITS_MASK) <= 5) {
375 if (vsd.significand < 2) {
376 vsd.significand = 0xffffffff;
380 vsm.significand <<= !(vsm.exponent & 1);
381 term = (u64)vsd.significand * vsd.significand;
382 rem = ((u64)vsm.significand << 32) - term;
384 pr_debug("VFP: term=%016llx rem=%016llx\n", term, rem);
387 vsd.significand -= 1;
388 rem += ((u64)vsd.significand << 1) | 1;
390 vsd.significand |= rem != 0;
393 vsd.significand = vfp_shiftright32jamming(vsd.significand, 1);
395 return vfp_single_normaliseround(sd, &vsd, fpscr, 0, "fsqrt");
404 static u32 vfp_compare(int sd, int signal_on_qnan, s32 m, u32 fpscr)
409 d = vfp_get_float(sd);
410 if (vfp_single_packed_exponent(m) == 255 && vfp_single_packed_mantissa(m)) {
411 ret |= FPSCR_C | FPSCR_V;
412 if (signal_on_qnan || !(vfp_single_packed_mantissa(m) & (1 << (VFP_SINGLE_MANTISSA_BITS - 1))))
414 * Signalling NaN, or signalling on quiet NaN
419 if (vfp_single_packed_exponent(d) == 255 && vfp_single_packed_mantissa(d)) {
420 ret |= FPSCR_C | FPSCR_V;
421 if (signal_on_qnan || !(vfp_single_packed_mantissa(d) & (1 << (VFP_SINGLE_MANTISSA_BITS - 1))))
423 * Signalling NaN, or signalling on quiet NaN
429 if (d == m || vfp_single_packed_abs(d | m) == 0) {
433 ret |= FPSCR_Z | FPSCR_C;
434 } else if (vfp_single_packed_sign(d ^ m)) {
438 if (vfp_single_packed_sign(d))
440 * d is negative, so d < m
445 * d is positive, so d > m
448 } else if ((vfp_single_packed_sign(d) != 0) ^ (d < m)) {
453 } else if ((vfp_single_packed_sign(d) != 0) ^ (d > m)) {
463 static u32 vfp_single_fcmp(int sd, int unused, s32 m, u32 fpscr)
465 return vfp_compare(sd, 0, m, fpscr);
468 static u32 vfp_single_fcmpe(int sd, int unused, s32 m, u32 fpscr)
470 return vfp_compare(sd, 1, m, fpscr);
473 static u32 vfp_single_fcmpz(int sd, int unused, s32 m, u32 fpscr)
475 return vfp_compare(sd, 0, 0, fpscr);
478 static u32 vfp_single_fcmpez(int sd, int unused, s32 m, u32 fpscr)
480 return vfp_compare(sd, 1, 0, fpscr);
483 static u32 vfp_single_fcvtd(int dd, int unused, s32 m, u32 fpscr)
485 struct vfp_single vsm;
486 struct vfp_double vdd;
490 vfp_single_unpack(&vsm, m);
492 tm = vfp_single_type(&vsm);
495 * If we have a signalling NaN, signal invalid operation.
498 exceptions = FPSCR_IOC;
500 if (tm & VFP_DENORMAL)
501 vfp_single_normalise_denormal(&vsm);
504 vdd.significand = (u64)vsm.significand << 32;
507 * If we have an infinity or NaN, the exponent must be 2047.
509 if (tm & (VFP_INFINITY|VFP_NAN)) {
512 vdd.significand |= VFP_DOUBLE_SIGNIFICAND_QNAN;
514 } else if (tm & VFP_ZERO)
517 vdd.exponent = vsm.exponent + (1023 - 127);
519 return vfp_double_normaliseround(dd, &vdd, fpscr, exceptions, "fcvtd");
522 vfp_put_double(vfp_double_pack(&vdd), dd);
526 static u32 vfp_single_fuito(int sd, int unused, s32 m, u32 fpscr)
528 struct vfp_single vs;
531 vs.exponent = 127 + 31 - 1;
532 vs.significand = (u32)m;
534 return vfp_single_normaliseround(sd, &vs, fpscr, 0, "fuito");
537 static u32 vfp_single_fsito(int sd, int unused, s32 m, u32 fpscr)
539 struct vfp_single vs;
541 vs.sign = (m & 0x80000000) >> 16;
542 vs.exponent = 127 + 31 - 1;
543 vs.significand = vs.sign ? -m : m;
545 return vfp_single_normaliseround(sd, &vs, fpscr, 0, "fsito");
548 static u32 vfp_single_ftoui(int sd, int unused, s32 m, u32 fpscr)
550 struct vfp_single vsm;
551 u32 d, exceptions = 0;
552 int rmode = fpscr & FPSCR_RMODE_MASK;
555 vfp_single_unpack(&vsm, m);
556 vfp_single_dump("VSM", &vsm);
559 * Do we have a denormalised number?
561 tm = vfp_single_type(&vsm);
562 if (tm & VFP_DENORMAL)
563 exceptions |= FPSCR_IDC;
568 if (vsm.exponent >= 127 + 32) {
569 d = vsm.sign ? 0 : 0xffffffff;
570 exceptions = FPSCR_IOC;
571 } else if (vsm.exponent >= 127 - 1) {
572 int shift = 127 + 31 - vsm.exponent;
576 * 2^0 <= m < 2^32-2^8
578 d = (vsm.significand << 1) >> shift;
579 rem = vsm.significand << (33 - shift);
581 if (rmode == FPSCR_ROUND_NEAREST) {
585 } else if (rmode == FPSCR_ROUND_TOZERO) {
587 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vsm.sign != 0)) {
591 if ((rem + incr) < rem) {
595 exceptions |= FPSCR_IOC;
600 exceptions |= FPSCR_IOC;
602 exceptions |= FPSCR_IXC;
605 if (vsm.exponent | vsm.significand) {
606 exceptions |= FPSCR_IXC;
607 if (rmode == FPSCR_ROUND_PLUSINF && vsm.sign == 0)
609 else if (rmode == FPSCR_ROUND_MINUSINF && vsm.sign) {
611 exceptions |= FPSCR_IOC;
616 pr_debug("VFP: ftoui: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
618 vfp_put_float(d, sd);
623 static u32 vfp_single_ftouiz(int sd, int unused, s32 m, u32 fpscr)
625 return vfp_single_ftoui(sd, unused, m, FPSCR_ROUND_TOZERO);
628 static u32 vfp_single_ftosi(int sd, int unused, s32 m, u32 fpscr)
630 struct vfp_single vsm;
631 u32 d, exceptions = 0;
632 int rmode = fpscr & FPSCR_RMODE_MASK;
635 vfp_single_unpack(&vsm, m);
636 vfp_single_dump("VSM", &vsm);
639 * Do we have a denormalised number?
641 tm = vfp_single_type(&vsm);
642 if (vfp_single_type(&vsm) & VFP_DENORMAL)
643 exceptions |= FPSCR_IDC;
647 exceptions |= FPSCR_IOC;
648 } else if (vsm.exponent >= 127 + 32) {
650 * m >= 2^31-2^7: invalid
655 exceptions |= FPSCR_IOC;
656 } else if (vsm.exponent >= 127 - 1) {
657 int shift = 127 + 31 - vsm.exponent;
660 /* 2^0 <= m <= 2^31-2^7 */
661 d = (vsm.significand << 1) >> shift;
662 rem = vsm.significand << (33 - shift);
664 if (rmode == FPSCR_ROUND_NEAREST) {
668 } else if (rmode == FPSCR_ROUND_TOZERO) {
670 } else if ((rmode == FPSCR_ROUND_PLUSINF) ^ (vsm.sign != 0)) {
674 if ((rem + incr) < rem && d < 0xffffffff)
676 if (d > 0x7fffffff + (vsm.sign != 0)) {
677 d = 0x7fffffff + (vsm.sign != 0);
678 exceptions |= FPSCR_IOC;
680 exceptions |= FPSCR_IXC;
686 if (vsm.exponent | vsm.significand) {
687 exceptions |= FPSCR_IXC;
688 if (rmode == FPSCR_ROUND_PLUSINF && vsm.sign == 0)
690 else if (rmode == FPSCR_ROUND_MINUSINF && vsm.sign)
695 pr_debug("VFP: ftosi: d(s%d)=%08x exceptions=%08x\n", sd, d, exceptions);
697 vfp_put_float((s32)d, sd);
702 static u32 vfp_single_ftosiz(int sd, int unused, s32 m, u32 fpscr)
704 return vfp_single_ftosi(sd, unused, m, FPSCR_ROUND_TOZERO);
707 static struct op fops_ext[32] = {
708 [FEXT_TO_IDX(FEXT_FCPY)] = { vfp_single_fcpy, 0 },
709 [FEXT_TO_IDX(FEXT_FABS)] = { vfp_single_fabs, 0 },
710 [FEXT_TO_IDX(FEXT_FNEG)] = { vfp_single_fneg, 0 },
711 [FEXT_TO_IDX(FEXT_FSQRT)] = { vfp_single_fsqrt, 0 },
712 [FEXT_TO_IDX(FEXT_FCMP)] = { vfp_single_fcmp, OP_SCALAR },
713 [FEXT_TO_IDX(FEXT_FCMPE)] = { vfp_single_fcmpe, OP_SCALAR },
714 [FEXT_TO_IDX(FEXT_FCMPZ)] = { vfp_single_fcmpz, OP_SCALAR },
715 [FEXT_TO_IDX(FEXT_FCMPEZ)] = { vfp_single_fcmpez, OP_SCALAR },
716 [FEXT_TO_IDX(FEXT_FCVT)] = { vfp_single_fcvtd, OP_SCALAR|OP_DD },
717 [FEXT_TO_IDX(FEXT_FUITO)] = { vfp_single_fuito, OP_SCALAR },
718 [FEXT_TO_IDX(FEXT_FSITO)] = { vfp_single_fsito, OP_SCALAR },
719 [FEXT_TO_IDX(FEXT_FTOUI)] = { vfp_single_ftoui, OP_SCALAR },
720 [FEXT_TO_IDX(FEXT_FTOUIZ)] = { vfp_single_ftouiz, OP_SCALAR },
721 [FEXT_TO_IDX(FEXT_FTOSI)] = { vfp_single_ftosi, OP_SCALAR },
722 [FEXT_TO_IDX(FEXT_FTOSIZ)] = { vfp_single_ftosiz, OP_SCALAR },
730 vfp_single_fadd_nonnumber(struct vfp_single *vsd, struct vfp_single *vsn,
731 struct vfp_single *vsm, u32 fpscr)
733 struct vfp_single *vsp;
737 tn = vfp_single_type(vsn);
738 tm = vfp_single_type(vsm);
740 if (tn & tm & VFP_INFINITY) {
742 * Two infinities. Are they different signs?
744 if (vsn->sign ^ vsm->sign) {
746 * different signs -> invalid
748 exceptions = FPSCR_IOC;
749 vsp = &vfp_single_default_qnan;
752 * same signs -> valid
756 } else if (tn & VFP_INFINITY && tm & VFP_NUMBER) {
758 * One infinity and one number -> infinity
763 * 'n' is a NaN of some type
765 return vfp_propagate_nan(vsd, vsn, vsm, fpscr);
772 vfp_single_add(struct vfp_single *vsd, struct vfp_single *vsn,
773 struct vfp_single *vsm, u32 fpscr)
777 if (vsn->significand & 0x80000000 ||
778 vsm->significand & 0x80000000) {
779 pr_info("VFP: bad FP values in %s\n", __func__);
780 vfp_single_dump("VSN", vsn);
781 vfp_single_dump("VSM", vsm);
785 * Ensure that 'n' is the largest magnitude number. Note that
786 * if 'n' and 'm' have equal exponents, we do not swap them.
787 * This ensures that NaN propagation works correctly.
789 if (vsn->exponent < vsm->exponent) {
790 struct vfp_single *t = vsn;
796 * Is 'n' an infinity or a NaN? Note that 'm' may be a number,
797 * infinity or a NaN here.
799 if (vsn->exponent == 255)
800 return vfp_single_fadd_nonnumber(vsd, vsn, vsm, fpscr);
803 * We have two proper numbers, where 'vsn' is the larger magnitude.
805 * Copy 'n' to 'd' before doing the arithmetic.
810 * Align both numbers.
812 exp_diff = vsn->exponent - vsm->exponent;
813 m_sig = vfp_shiftright32jamming(vsm->significand, exp_diff);
816 * If the signs are different, we are really subtracting.
818 if (vsn->sign ^ vsm->sign) {
819 m_sig = vsn->significand - m_sig;
820 if ((s32)m_sig < 0) {
821 vsd->sign = vfp_sign_negate(vsd->sign);
823 } else if (m_sig == 0) {
824 vsd->sign = (fpscr & FPSCR_RMODE_MASK) ==
825 FPSCR_ROUND_MINUSINF ? 0x8000 : 0;
828 m_sig = vsn->significand + m_sig;
830 vsd->significand = m_sig;
836 vfp_single_multiply(struct vfp_single *vsd, struct vfp_single *vsn, struct vfp_single *vsm, u32 fpscr)
838 vfp_single_dump("VSN", vsn);
839 vfp_single_dump("VSM", vsm);
842 * Ensure that 'n' is the largest magnitude number. Note that
843 * if 'n' and 'm' have equal exponents, we do not swap them.
844 * This ensures that NaN propagation works correctly.
846 if (vsn->exponent < vsm->exponent) {
847 struct vfp_single *t = vsn;
850 pr_debug("VFP: swapping M <-> N\n");
853 vsd->sign = vsn->sign ^ vsm->sign;
856 * If 'n' is an infinity or NaN, handle it. 'm' may be anything.
858 if (vsn->exponent == 255) {
859 if (vsn->significand || (vsm->exponent == 255 && vsm->significand))
860 return vfp_propagate_nan(vsd, vsn, vsm, fpscr);
861 if ((vsm->exponent | vsm->significand) == 0) {
862 *vsd = vfp_single_default_qnan;
865 vsd->exponent = vsn->exponent;
866 vsd->significand = 0;
871 * If 'm' is zero, the result is always zero. In this case,
872 * 'n' may be zero or a number, but it doesn't matter which.
874 if ((vsm->exponent | vsm->significand) == 0) {
876 vsd->significand = 0;
881 * We add 2 to the destination exponent for the same reason as
882 * the addition case - though this time we have +1 from each
885 vsd->exponent = vsn->exponent + vsm->exponent - 127 + 2;
886 vsd->significand = vfp_hi64to32jamming((u64)vsn->significand * vsm->significand);
888 vfp_single_dump("VSD", vsd);
892 #define NEG_MULTIPLY (1 << 0)
893 #define NEG_SUBTRACT (1 << 1)
896 vfp_single_multiply_accumulate(int sd, int sn, s32 m, u32 fpscr, u32 negate, char *func)
898 struct vfp_single vsd, vsp, vsn, vsm;
902 v = vfp_get_float(sn);
903 pr_debug("VFP: s%u = %08x\n", sn, v);
904 vfp_single_unpack(&vsn, v);
905 if (vsn.exponent == 0 && vsn.significand)
906 vfp_single_normalise_denormal(&vsn);
908 vfp_single_unpack(&vsm, m);
909 if (vsm.exponent == 0 && vsm.significand)
910 vfp_single_normalise_denormal(&vsm);
912 exceptions = vfp_single_multiply(&vsp, &vsn, &vsm, fpscr);
913 if (negate & NEG_MULTIPLY)
914 vsp.sign = vfp_sign_negate(vsp.sign);
916 v = vfp_get_float(sd);
917 pr_debug("VFP: s%u = %08x\n", sd, v);
918 vfp_single_unpack(&vsn, v);
919 if (negate & NEG_SUBTRACT)
920 vsn.sign = vfp_sign_negate(vsn.sign);
922 exceptions |= vfp_single_add(&vsd, &vsn, &vsp, fpscr);
924 return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, func);
928 * Standard operations
932 * sd = sd + (sn * sm)
934 static u32 vfp_single_fmac(int sd, int sn, s32 m, u32 fpscr)
936 return vfp_single_multiply_accumulate(sd, sn, m, fpscr, 0, "fmac");
940 * sd = sd - (sn * sm)
942 static u32 vfp_single_fnmac(int sd, int sn, s32 m, u32 fpscr)
944 return vfp_single_multiply_accumulate(sd, sn, m, fpscr, NEG_MULTIPLY, "fnmac");
948 * sd = -sd + (sn * sm)
950 static u32 vfp_single_fmsc(int sd, int sn, s32 m, u32 fpscr)
952 return vfp_single_multiply_accumulate(sd, sn, m, fpscr, NEG_SUBTRACT, "fmsc");
956 * sd = -sd - (sn * sm)
958 static u32 vfp_single_fnmsc(int sd, int sn, s32 m, u32 fpscr)
960 return vfp_single_multiply_accumulate(sd, sn, m, fpscr, NEG_SUBTRACT | NEG_MULTIPLY, "fnmsc");
966 static u32 vfp_single_fmul(int sd, int sn, s32 m, u32 fpscr)
968 struct vfp_single vsd, vsn, vsm;
970 s32 n = vfp_get_float(sn);
972 pr_debug("VFP: s%u = %08x\n", sn, n);
974 vfp_single_unpack(&vsn, n);
975 if (vsn.exponent == 0 && vsn.significand)
976 vfp_single_normalise_denormal(&vsn);
978 vfp_single_unpack(&vsm, m);
979 if (vsm.exponent == 0 && vsm.significand)
980 vfp_single_normalise_denormal(&vsm);
982 exceptions = vfp_single_multiply(&vsd, &vsn, &vsm, fpscr);
983 return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fmul");
989 static u32 vfp_single_fnmul(int sd, int sn, s32 m, u32 fpscr)
991 struct vfp_single vsd, vsn, vsm;
993 s32 n = vfp_get_float(sn);
995 pr_debug("VFP: s%u = %08x\n", sn, n);
997 vfp_single_unpack(&vsn, n);
998 if (vsn.exponent == 0 && vsn.significand)
999 vfp_single_normalise_denormal(&vsn);
1001 vfp_single_unpack(&vsm, m);
1002 if (vsm.exponent == 0 && vsm.significand)
1003 vfp_single_normalise_denormal(&vsm);
1005 exceptions = vfp_single_multiply(&vsd, &vsn, &vsm, fpscr);
1006 vsd.sign = vfp_sign_negate(vsd.sign);
1007 return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fnmul");
1013 static u32 vfp_single_fadd(int sd, int sn, s32 m, u32 fpscr)
1015 struct vfp_single vsd, vsn, vsm;
1017 s32 n = vfp_get_float(sn);
1019 pr_debug("VFP: s%u = %08x\n", sn, n);
1022 * Unpack and normalise denormals.
1024 vfp_single_unpack(&vsn, n);
1025 if (vsn.exponent == 0 && vsn.significand)
1026 vfp_single_normalise_denormal(&vsn);
1028 vfp_single_unpack(&vsm, m);
1029 if (vsm.exponent == 0 && vsm.significand)
1030 vfp_single_normalise_denormal(&vsm);
1032 exceptions = vfp_single_add(&vsd, &vsn, &vsm, fpscr);
1034 return vfp_single_normaliseround(sd, &vsd, fpscr, exceptions, "fadd");
1040 static u32 vfp_single_fsub(int sd, int sn, s32 m, u32 fpscr)
1043 * Subtraction is addition with one sign inverted.
1045 return vfp_single_fadd(sd, sn, vfp_single_packed_negate(m), fpscr);
1051 static u32 vfp_single_fdiv(int sd, int sn, s32 m, u32 fpscr)
1053 struct vfp_single vsd, vsn, vsm;
1055 s32 n = vfp_get_float(sn);
1058 pr_debug("VFP: s%u = %08x\n", sn, n);
1060 vfp_single_unpack(&vsn, n);
1061 vfp_single_unpack(&vsm, m);
1063 vsd.sign = vsn.sign ^ vsm.sign;
1065 tn = vfp_single_type(&vsn);
1066 tm = vfp_single_type(&vsm);
1081 * If n and m are infinity, the result is invalid
1082 * If n and m are zero, the result is invalid
1084 if (tm & tn & (VFP_INFINITY|VFP_ZERO))
1088 * If n is infinity, the result is infinity
1090 if (tn & VFP_INFINITY)
1094 * If m is zero, raise div0 exception
1100 * If m is infinity, or n is zero, the result is zero
1102 if (tm & VFP_INFINITY || tn & VFP_ZERO)
1105 if (tn & VFP_DENORMAL)
1106 vfp_single_normalise_denormal(&vsn);
1107 if (tm & VFP_DENORMAL)
1108 vfp_single_normalise_denormal(&vsm);
1111 * Ok, we have two numbers, we can perform division.
1113 vsd.exponent = vsn.exponent - vsm.exponent + 127 - 1;
1114 vsm.significand <<= 1;
1115 if (vsm.significand <= (2 * vsn.significand)) {
1116 vsn.significand >>= 1;
1120 u64 significand = (u64)vsn.significand << 32;
1121 do_div(significand, vsm.significand);
1122 vsd.significand = significand;
1124 if ((vsd.significand & 0x3f) == 0)
1125 vsd.significand |= ((u64)vsm.significand * vsd.significand != (u64)vsn.significand << 32);
1127 return vfp_single_normaliseround(sd, &vsd, fpscr, 0, "fdiv");
1130 exceptions = vfp_propagate_nan(&vsd, &vsn, &vsm, fpscr);
1132 vfp_put_float(vfp_single_pack(&vsd), sd);
1136 exceptions = vfp_propagate_nan(&vsd, &vsm, &vsn, fpscr);
1141 vsd.significand = 0;
1145 exceptions = FPSCR_DZC;
1148 vsd.significand = 0;
1152 vfp_put_float(vfp_single_pack(&vfp_single_default_qnan), sd);
1156 static struct op fops[16] = {
1157 [FOP_TO_IDX(FOP_FMAC)] = { vfp_single_fmac, 0 },
1158 [FOP_TO_IDX(FOP_FNMAC)] = { vfp_single_fnmac, 0 },
1159 [FOP_TO_IDX(FOP_FMSC)] = { vfp_single_fmsc, 0 },
1160 [FOP_TO_IDX(FOP_FNMSC)] = { vfp_single_fnmsc, 0 },
1161 [FOP_TO_IDX(FOP_FMUL)] = { vfp_single_fmul, 0 },
1162 [FOP_TO_IDX(FOP_FNMUL)] = { vfp_single_fnmul, 0 },
1163 [FOP_TO_IDX(FOP_FADD)] = { vfp_single_fadd, 0 },
1164 [FOP_TO_IDX(FOP_FSUB)] = { vfp_single_fsub, 0 },
1165 [FOP_TO_IDX(FOP_FDIV)] = { vfp_single_fdiv, 0 },
1168 #define FREG_BANK(x) ((x) & 0x18)
1169 #define FREG_IDX(x) ((x) & 7)
1171 u32 vfp_single_cpdo(u32 inst, u32 fpscr)
1173 u32 op = inst & FOP_MASK;
1176 unsigned int sn = vfp_get_sn(inst);
1177 unsigned int sm = vfp_get_sm(inst);
1178 unsigned int vecitr, veclen, vecstride;
1181 vecstride = 1 + ((fpscr & FPSCR_STRIDE_MASK) == FPSCR_STRIDE_MASK);
1183 fop = (op == FOP_EXT) ? &fops_ext[FEXT_TO_IDX(inst)] : &fops[FOP_TO_IDX(op)];
1186 * fcvtsd takes a dN register number as destination, not sN.
1187 * Technically, if bit 0 of dd is set, this is an invalid
1188 * instruction. However, we ignore this for efficiency.
1189 * It also only operates on scalars.
1191 if (fop->flags & OP_DD)
1192 dest = vfp_get_dd(inst);
1194 dest = vfp_get_sd(inst);
1197 * If destination bank is zero, vector length is always '1'.
1198 * ARM DDI0100F C5.1.3, C5.3.2.
1200 if ((fop->flags & OP_SCALAR) || FREG_BANK(dest) == 0)
1203 veclen = fpscr & FPSCR_LENGTH_MASK;
1205 pr_debug("VFP: vecstride=%u veclen=%u\n", vecstride,
1206 (veclen >> FPSCR_LENGTH_BIT) + 1);
1211 for (vecitr = 0; vecitr <= veclen; vecitr += 1 << FPSCR_LENGTH_BIT) {
1212 s32 m = vfp_get_float(sm);
1216 type = fop->flags & OP_DD ? 'd' : 's';
1218 pr_debug("VFP: itr%d (%c%u) = op[%u] (s%u=%08x)\n",
1219 vecitr >> FPSCR_LENGTH_BIT, type, dest, sn,
1222 pr_debug("VFP: itr%d (%c%u) = (s%u) op[%u] (s%u=%08x)\n",
1223 vecitr >> FPSCR_LENGTH_BIT, type, dest, sn,
1224 FOP_TO_IDX(op), sm, m);
1226 except = fop->fn(dest, sn, m, fpscr);
1227 pr_debug("VFP: itr%d: exceptions=%08x\n",
1228 vecitr >> FPSCR_LENGTH_BIT, except);
1230 exceptions |= except;
1233 * CHECK: It appears to be undefined whether we stop when
1234 * we encounter an exception. We continue.
1236 dest = FREG_BANK(dest) + ((FREG_IDX(dest) + vecstride) & 7);
1237 sn = FREG_BANK(sn) + ((FREG_IDX(sn) + vecstride) & 7);
1238 if (FREG_BANK(sm) != 0)
1239 sm = FREG_BANK(sm) + ((FREG_IDX(sm) + vecstride) & 7);