xref: /onnv-gate/usr/src/lib/libc/sparcv9/fp/__quad_mag64.s (revision 0:68f95e015346) 
1*0Sstevel@tonic-gate!
2*0Sstevel@tonic-gate! Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
3*0Sstevel@tonic-gate! Use is subject to license terms.
4*0Sstevel@tonic-gate!
5*0Sstevel@tonic-gate! CDDL HEADER START
6*0Sstevel@tonic-gate!
7*0Sstevel@tonic-gate! The contents of this file are subject to the terms of the
8*0Sstevel@tonic-gate! Common Development and Distribution License, Version 1.0 only
9*0Sstevel@tonic-gate! (the "License").  You may not use this file except in compliance
10*0Sstevel@tonic-gate! with the License.
11*0Sstevel@tonic-gate!
12*0Sstevel@tonic-gate! You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
13*0Sstevel@tonic-gate! or http://www.opensolaris.org/os/licensing.
14*0Sstevel@tonic-gate! See the License for the specific language governing permissions
15*0Sstevel@tonic-gate! and limitations under the License.
16*0Sstevel@tonic-gate!
17*0Sstevel@tonic-gate! When distributing Covered Code, include this CDDL HEADER in each
18*0Sstevel@tonic-gate! file and include the License file at usr/src/OPENSOLARIS.LICENSE.
19*0Sstevel@tonic-gate! If applicable, add the following below this CDDL HEADER, with the
20*0Sstevel@tonic-gate! fields enclosed by brackets "[]" replaced with your own identifying
21*0Sstevel@tonic-gate! information: Portions Copyright [yyyy] [name of copyright owner]
22*0Sstevel@tonic-gate!
23*0Sstevel@tonic-gate! CDDL HEADER END
24*0Sstevel@tonic-gate!
25*0Sstevel@tonic-gate
26*0Sstevel@tonic-gate.ident	"%Z%%M%	%I%	%E% SMI"
27*0Sstevel@tonic-gate
28*0Sstevel@tonic-gate! /*
29*0Sstevel@tonic-gate!  * This file contains __quad_mag_add and __quad_mag_sub, the core
30*0Sstevel@tonic-gate!  * of the quad precision add and subtract operations.
31*0Sstevel@tonic-gate!  */
32*0Sstevel@tonic-gate! SPARC V9 version hand-coded in assembly to use 64-bit integer registers
33*0Sstevel@tonic-gate
34*0Sstevel@tonic-gate	.file	"__quad_mag64.s"
35*0Sstevel@tonic-gate
36*0Sstevel@tonic-gate#include <sys/asm_linkage.h>
37*0Sstevel@tonic-gate
38*0Sstevel@tonic-gate! union longdouble {
39*0Sstevel@tonic-gate! 	struct {
40*0Sstevel@tonic-gate! 		unsigned int	msw;
41*0Sstevel@tonic-gate! 		unsigned int	frac2;
42*0Sstevel@tonic-gate! 		unsigned int	frac3;
43*0Sstevel@tonic-gate! 		unsigned int	frac4;
44*0Sstevel@tonic-gate! 	} l;
45*0Sstevel@tonic-gate! 	struct {
46*0Sstevel@tonic-gate! 		unsigned long	msll;
47*0Sstevel@tonic-gate! 		unsigned long	frac;
48*0Sstevel@tonic-gate! 	} ll;
49*0Sstevel@tonic-gate! 	long double	d;
50*0Sstevel@tonic-gate! };
51*0Sstevel@tonic-gate!
52*0Sstevel@tonic-gate! /*
53*0Sstevel@tonic-gate!  * __quad_mag_add(x, y, z, fsr)
54*0Sstevel@tonic-gate!  *
55*0Sstevel@tonic-gate!  * Sets *z = *x + *y, rounded according to the rounding mode in *fsr,
56*0Sstevel@tonic-gate!  * and updates the current exceptions in *fsr.  This routine assumes
57*0Sstevel@tonic-gate!  * *x and *y are finite, with the same sign (i.e., an addition of
58*0Sstevel@tonic-gate!  * magnitudes), |*x| >= |*y|, and *z already has its sign bit set.
59*0Sstevel@tonic-gate!  */
60*0Sstevel@tonic-gate! void
61*0Sstevel@tonic-gate! __quad_mag_add(const union longdouble *x, const union longdouble *y,
62*0Sstevel@tonic-gate! 	union longdouble *z, unsigned int *fsr)
63*0Sstevel@tonic-gate! {
64*0Sstevel@tonic-gate! 	unsigned long	lx, ly, frac, sticky;
65*0Sstevel@tonic-gate! 	unsigned int	ex, ey, round, rm;
66*0Sstevel@tonic-gate! 	int		e, uflo;
67*0Sstevel@tonic-gate!
68*0Sstevel@tonic-gate! 	/* get the leading significand double-words and exponents */
69*0Sstevel@tonic-gate! 	ex = (x->ll.msll >> 48) & 0x7fff;
70*0Sstevel@tonic-gate! 	lx = x->ll.msll & ~0xffff000000000000ul;
71*0Sstevel@tonic-gate! 	if (ex == 0)
72*0Sstevel@tonic-gate! 		ex = 1;
73*0Sstevel@tonic-gate! 	else
74*0Sstevel@tonic-gate! 		lx |= 0x0001000000000000ul;
75*0Sstevel@tonic-gate!
76*0Sstevel@tonic-gate! 	ey = (y->ll.msll >> 48) & 0x7fff;
77*0Sstevel@tonic-gate! 	ly = y->ll.msll & ~0xffff000000000000ul;
78*0Sstevel@tonic-gate! 	if (ey == 0)
79*0Sstevel@tonic-gate! 		ey = 1;
80*0Sstevel@tonic-gate! 	else
81*0Sstevel@tonic-gate! 		ly |= 0x0001000000000000ul;
82*0Sstevel@tonic-gate!
83*0Sstevel@tonic-gate! 	/* prenormalize y */
84*0Sstevel@tonic-gate! 	e = (int) ex - (int) ey;
85*0Sstevel@tonic-gate! 	round = sticky = 0;
86*0Sstevel@tonic-gate! 	if (e >= 114) {
87*0Sstevel@tonic-gate! 		frac = x->ll.frac;
88*0Sstevel@tonic-gate! 		sticky = ly | y->ll.frac;
89*0Sstevel@tonic-gate! 	} else {
90*0Sstevel@tonic-gate! 		frac = y->ll.frac;
91*0Sstevel@tonic-gate! 		if (e >= 64) {
92*0Sstevel@tonic-gate! 			sticky = frac & 0x7ffffffffffffffful;
93*0Sstevel@tonic-gate! 			round = frac >> 63;
94*0Sstevel@tonic-gate! 			frac = ly;
95*0Sstevel@tonic-gate! 			ly = 0;
96*0Sstevel@tonic-gate! 			e -= 64;
97*0Sstevel@tonic-gate! 		}
98*0Sstevel@tonic-gate! 		if (e) {
99*0Sstevel@tonic-gate! 			sticky |= round | (frac & ((1ul << (e - 1)) - 1));
100*0Sstevel@tonic-gate! 			round = (frac >> (e - 1)) & 1;
101*0Sstevel@tonic-gate! 			frac = (frac >> e) | (ly << (64 - e));
102*0Sstevel@tonic-gate! 			ly >>= e;
103*0Sstevel@tonic-gate! 		}
104*0Sstevel@tonic-gate!
105*0Sstevel@tonic-gate! 		/* add, propagating carries */
106*0Sstevel@tonic-gate! 		frac += x->ll.frac;
107*0Sstevel@tonic-gate! 		lx += ly;
108*0Sstevel@tonic-gate! 		if (frac < x->ll.frac)
109*0Sstevel@tonic-gate! 			lx++;
110*0Sstevel@tonic-gate!
111*0Sstevel@tonic-gate! 		/* postnormalize */
112*0Sstevel@tonic-gate! 		if (lx >= 0x0002000000000000ul) {
113*0Sstevel@tonic-gate! 			sticky |= round;
114*0Sstevel@tonic-gate! 			round = frac & 1;
115*0Sstevel@tonic-gate! 			frac = (frac >> 1) | (lx << 63);
116*0Sstevel@tonic-gate! 			lx >>= 1;
117*0Sstevel@tonic-gate! 			ex++;
118*0Sstevel@tonic-gate! 		}
119*0Sstevel@tonic-gate! 	}
120*0Sstevel@tonic-gate!
121*0Sstevel@tonic-gate! 	/* keep track of whether the result before rounding is tiny */
122*0Sstevel@tonic-gate! 	uflo = (lx < 0x0001000000000000ul);
123*0Sstevel@tonic-gate!
124*0Sstevel@tonic-gate! 	/* get the rounding mode, fudging directed rounding modes
125*0Sstevel@tonic-gate! 	   as though the result were positive */
126*0Sstevel@tonic-gate! 	rm = *fsr >> 30;
127*0Sstevel@tonic-gate! 	if (z->l.msw)
128*0Sstevel@tonic-gate! 		rm ^= (rm >> 1);
129*0Sstevel@tonic-gate!
130*0Sstevel@tonic-gate! 	/* see if we need to round */
131*0Sstevel@tonic-gate! 	if (round | sticky) {
132*0Sstevel@tonic-gate! 		*fsr |= FSR_NXC;
133*0Sstevel@tonic-gate!
134*0Sstevel@tonic-gate! 		/* round up if necessary */
135*0Sstevel@tonic-gate! 		if (rm == FSR_RP || (rm == FSR_RN && round &&
136*0Sstevel@tonic-gate! 			(sticky || (frac & 1)))) {
137*0Sstevel@tonic-gate! 			if (++frac == 0)
138*0Sstevel@tonic-gate! 				if (++lx >= 0x0002000000000000ul) {
139*0Sstevel@tonic-gate! 					lx >>= 1;
140*0Sstevel@tonic-gate! 					ex++;
141*0Sstevel@tonic-gate! 				}
142*0Sstevel@tonic-gate! 		}
143*0Sstevel@tonic-gate! 	}
144*0Sstevel@tonic-gate!
145*0Sstevel@tonic-gate! 	/* check for overflow */
146*0Sstevel@tonic-gate! 	if (ex >= 0x7fff) {
147*0Sstevel@tonic-gate! 		/* store the default overflowed result */
148*0Sstevel@tonic-gate! 		*fsr |= FSR_OFC | FSR_NXC;
149*0Sstevel@tonic-gate! 		if (rm == FSR_RN || rm == FSR_RP) {
150*0Sstevel@tonic-gate! 			z->l.msw |= 0x7fff0000;
151*0Sstevel@tonic-gate! 			z->l.frac2 = 0;
152*0Sstevel@tonic-gate! 			z->ll.frac = 0;
153*0Sstevel@tonic-gate! 		} else {
154*0Sstevel@tonic-gate! 			z->l.msw |= 0x7ffeffff;
155*0Sstevel@tonic-gate! 			z->l.frac2 = 0xffffffff;
156*0Sstevel@tonic-gate! 			z->ll.frac = 0xfffffffffffffffful;
157*0Sstevel@tonic-gate! 		}
158*0Sstevel@tonic-gate! 	} else {
159*0Sstevel@tonic-gate! 		/* store the result */
160*0Sstevel@tonic-gate! 		if (lx >= 0x0001000000000000ul)
161*0Sstevel@tonic-gate! 			z->l.msw |= (ex << 16);
162*0Sstevel@tonic-gate! 		z->l.msw |= (lx >> 32) & 0xffff;
163*0Sstevel@tonic-gate! 		z->l.frac2 = (lx & 0xffffffff);
164*0Sstevel@tonic-gate! 		z->ll.frac = frac;
165*0Sstevel@tonic-gate!
166*0Sstevel@tonic-gate! 		/* if the pre-rounded result was tiny and underflow trapping
167*0Sstevel@tonic-gate! 		   is enabled, simulate underflow */
168*0Sstevel@tonic-gate! 		if (uflo && (*fsr & FSR_UFM))
169*0Sstevel@tonic-gate! 			*fsr |= FSR_UFC;
170*0Sstevel@tonic-gate! 	}
171*0Sstevel@tonic-gate! }
172*0Sstevel@tonic-gate
173*0Sstevel@tonic-gate	ENTRY(__quad_mag_add)
174*0Sstevel@tonic-gate	save	%sp,-SA(MINFRAME),%sp
175*0Sstevel@tonic-gate
176*0Sstevel@tonic-gate	sethi	%hi(0xffff0000),%g1
177*0Sstevel@tonic-gate	sllx	%g1,32,%g1		! g1 = 0xffff000000000000
178*0Sstevel@tonic-gate
179*0Sstevel@tonic-gate	sethi	%hi(0x7fff),%l7
180*0Sstevel@tonic-gate	or	%l7,%lo(0x7fff),%l7	! l7 = 0x7fff
181*0Sstevel@tonic-gate
182*0Sstevel@tonic-gate	ldx	[%i0],%o0
183*0Sstevel@tonic-gate	srlx	%o0,48,%l0
184*0Sstevel@tonic-gate	andcc	%l0,%l7,%l0		! l0 = ex
185*0Sstevel@tonic-gate	beq,pn	%icc,1f
186*0Sstevel@tonic-gate	andn	%o0,%g1,%o0		! o0 = lx
187*0Sstevel@tonic-gate	ba,pt	%icc,2f
188*0Sstevel@tonic-gate	sub	%o0,%g1,%o0
189*0Sstevel@tonic-gate1:
190*0Sstevel@tonic-gate	mov	1,%l0
191*0Sstevel@tonic-gate2:
192*0Sstevel@tonic-gate
193*0Sstevel@tonic-gate	ldx	[%i1],%o1
194*0Sstevel@tonic-gate	srlx	%o1,48,%l1
195*0Sstevel@tonic-gate	andcc	%l1,%l7,%l1		! l1 = ey
196*0Sstevel@tonic-gate	beq,pn	%icc,1f
197*0Sstevel@tonic-gate	andn	%o1,%g1,%o1		! o1 = ly
198*0Sstevel@tonic-gate	ba,pt	%icc,2f
199*0Sstevel@tonic-gate	sub	%o1,%g1,%o1
200*0Sstevel@tonic-gate1:
201*0Sstevel@tonic-gate	mov	1,%l1
202*0Sstevel@tonic-gate2:
203*0Sstevel@tonic-gate
204*0Sstevel@tonic-gate	sub	%l0,%l1,%l1		! l1 = e = ex - ey
205*0Sstevel@tonic-gate	cmp	%l1,114			! see if we need to prenormalize
206*0Sstevel@tonic-gate	bge,pn	%icc,1f
207*0Sstevel@tonic-gate	mov	0,%l6			! l6 = round
208*0Sstevel@tonic-gate	mov	0,%o7			! o7 = sticky
209*0Sstevel@tonic-gate	cmp	%l1,64
210*0Sstevel@tonic-gate	bl,pt	%icc,3f
211*0Sstevel@tonic-gate	ldx	[%i1+8],%o2		! o2 = frac
212*0Sstevel@tonic-gate	sllx	%o2,1,%o7		! lop off high order bit
213*0Sstevel@tonic-gate	srlx	%o2,63,%l6
214*0Sstevel@tonic-gate	mov	%o1,%o2
215*0Sstevel@tonic-gate	mov	0,%o1
216*0Sstevel@tonic-gate	sub	%l1,64,%l1
217*0Sstevel@tonic-gate3:
218*0Sstevel@tonic-gate	tst	%l1
219*0Sstevel@tonic-gate	beq,pn	%icc,4f
220*0Sstevel@tonic-gate	sub	%l1,1,%l2
221*0Sstevel@tonic-gate	mov	1,%o3
222*0Sstevel@tonic-gate	sllx	%o3,%l2,%o3
223*0Sstevel@tonic-gate	sub	%o3,1,%o3
224*0Sstevel@tonic-gate	and	%o3,%o2,%o3
225*0Sstevel@tonic-gate	or	%o3,%l6,%o3
226*0Sstevel@tonic-gate	or	%o7,%o3,%o7
227*0Sstevel@tonic-gate	srlx	%o2,%l2,%o4
228*0Sstevel@tonic-gate	and	%o4,1,%l6
229*0Sstevel@tonic-gate	srlx	%o2,%l1,%o2
230*0Sstevel@tonic-gate	mov	64,%l3
231*0Sstevel@tonic-gate	sub	%l3,%l1,%l3
232*0Sstevel@tonic-gate	sllx	%o1,%l3,%o5
233*0Sstevel@tonic-gate	or	%o2,%o5,%o2
234*0Sstevel@tonic-gate	srlx	%o1,%l1,%o1
235*0Sstevel@tonic-gate4:
236*0Sstevel@tonic-gate	ldx	[%i0+8],%o3
237*0Sstevel@tonic-gate	add	%o2,%o3,%o2		! add, propagating carry
238*0Sstevel@tonic-gate	cmp	%o2,%o3
239*0Sstevel@tonic-gate	bgeu,pt %xcc,5f
240*0Sstevel@tonic-gate	add	%o0,%o1,%o0
241*0Sstevel@tonic-gate	add	%o0,1,%o0
242*0Sstevel@tonic-gate5:
243*0Sstevel@tonic-gate	srlx	%o0,49,%o5		! if sum carried out, postnormalize
244*0Sstevel@tonic-gate	tst	%o5
245*0Sstevel@tonic-gate	beq,pt	%icc,2f
246*0Sstevel@tonic-gate	nop
247*0Sstevel@tonic-gate	or	%o7,%l6,%o7
248*0Sstevel@tonic-gate	and	%o2,1,%l6
249*0Sstevel@tonic-gate	srlx	%o2,1,%o2
250*0Sstevel@tonic-gate	sllx	%o0,63,%o3
251*0Sstevel@tonic-gate	or	%o2,%o3,%o2
252*0Sstevel@tonic-gate	srlx	%o0,1,%o0
253*0Sstevel@tonic-gate	ba,pt	%icc,2f
254*0Sstevel@tonic-gate	add	%l0,1,%l0
255*0Sstevel@tonic-gate1:
256*0Sstevel@tonic-gate	ldx	[%i0+8],%o2		! (full prenormalization shift case)
257*0Sstevel@tonic-gate	ldx	[%i1+8],%o3
258*0Sstevel@tonic-gate	or	%o1,%o3,%o7
259*0Sstevel@tonic-gate2:
260*0Sstevel@tonic-gate
261*0Sstevel@tonic-gate	add	%o0,%g1,%o1		! see if sum is tiny
262*0Sstevel@tonic-gate	srlx	%o1,63,%l2		! l2 = uflo
263*0Sstevel@tonic-gate
264*0Sstevel@tonic-gate	ld	[%i3],%i4		! get the rounding mode
265*0Sstevel@tonic-gate	srl	%i4,30,%l3		! l3 = rm
266*0Sstevel@tonic-gate	ld	[%i2],%l4		! l4 = z->l.msw
267*0Sstevel@tonic-gate	tst	%l4
268*0Sstevel@tonic-gate	beq,pn	%icc,1f
269*0Sstevel@tonic-gate	srl	%l3,1,%l5
270*0Sstevel@tonic-gate	xor	%l3,%l5,%l3
271*0Sstevel@tonic-gate1:
272*0Sstevel@tonic-gate
273*0Sstevel@tonic-gate	orcc	%o7,%l6,%g0		! see if we need to round
274*0Sstevel@tonic-gate	beq,pn	%xcc,1f
275*0Sstevel@tonic-gate	andcc	%l3,1,%g0
276*0Sstevel@tonic-gate	or	%i4,1,%i4
277*0Sstevel@tonic-gate	bne,pn	%icc,1f
278*0Sstevel@tonic-gate	tst	%l3
279*0Sstevel@tonic-gate	bne,pn	%icc,2f
280*0Sstevel@tonic-gate	tst	%l6
281*0Sstevel@tonic-gate	beq,pn	%icc,1f
282*0Sstevel@tonic-gate	and	%o2,1,%o3
283*0Sstevel@tonic-gate	orcc	%o3,%o7,%g0
284*0Sstevel@tonic-gate	beq,pn	%xcc,1f
285*0Sstevel@tonic-gate	nop
286*0Sstevel@tonic-gate2:
287*0Sstevel@tonic-gate	addcc	%o2,1,%o2		! round up and check for carry out
288*0Sstevel@tonic-gate	bne,pt	%xcc,1f
289*0Sstevel@tonic-gate	nop
290*0Sstevel@tonic-gate	add	%o0,1,%o0
291*0Sstevel@tonic-gate	srlx	%o0,49,%o1
292*0Sstevel@tonic-gate	tst	%o1
293*0Sstevel@tonic-gate	beq,pt	%icc,1f
294*0Sstevel@tonic-gate	nop
295*0Sstevel@tonic-gate	srlx	%o0,1,%o0
296*0Sstevel@tonic-gate	add	%l0,1,%l0
297*0Sstevel@tonic-gate1:
298*0Sstevel@tonic-gate
299*0Sstevel@tonic-gate	cmp	%l0,%l7			! check for overflow
300*0Sstevel@tonic-gate	bge,pn	%icc,1f
301*0Sstevel@tonic-gate	addcc	%o0,%g1,%g0
302*0Sstevel@tonic-gate	bl,pn	%xcc,2f
303*0Sstevel@tonic-gate	sll	%l0,16,%l1
304*0Sstevel@tonic-gate	or	%l4,%l1,%l4
305*0Sstevel@tonic-gate2:
306*0Sstevel@tonic-gate	sllx	%o0,16,%o1
307*0Sstevel@tonic-gate	srlx	%o1,48,%o1
308*0Sstevel@tonic-gate	or	%l4,%o1,%l4
309*0Sstevel@tonic-gate	st	%l4,[%i2]
310*0Sstevel@tonic-gate	st	%o0,[%i2+4]
311*0Sstevel@tonic-gate	stx	%o2,[%i2+8]
312*0Sstevel@tonic-gate	tst	%l2			! see if we need to raise underflow
313*0Sstevel@tonic-gate	beq,pt	%icc,3f
314*0Sstevel@tonic-gate	srl	%i4,23,%i5
315*0Sstevel@tonic-gate	andcc	%i5,4,%i5
316*0Sstevel@tonic-gate	ba,pt	%icc,3f
317*0Sstevel@tonic-gate	or	%i4,%i5,%i4
318*0Sstevel@tonic-gate
319*0Sstevel@tonic-gate1:
320*0Sstevel@tonic-gate	andcc	%l3,1,%g0
321*0Sstevel@tonic-gate	bne,pn	%icc,2f
322*0Sstevel@tonic-gate	or	%i4,9,%i4		! overflow
323*0Sstevel@tonic-gate	sll	%l7,16,%l7		! 7fff00000...
324*0Sstevel@tonic-gate	or	%l4,%l7,%l4
325*0Sstevel@tonic-gate	st	%l4,[%i2]
326*0Sstevel@tonic-gate	st	%g0,[%i2+4]
327*0Sstevel@tonic-gate	ba,pt	%icc,3f
328*0Sstevel@tonic-gate	stx	%g0,[%i2+8]
329*0Sstevel@tonic-gate2:
330*0Sstevel@tonic-gate	mov	-1,%o0			! 7ffeffff...
331*0Sstevel@tonic-gate	sll	%l7,16,%l7
332*0Sstevel@tonic-gate	add	%o0,%l7,%l7
333*0Sstevel@tonic-gate	or	%l4,%l7,%l4
334*0Sstevel@tonic-gate	st	%l4,[%i2]
335*0Sstevel@tonic-gate	st	%o0,[%i2+4]
336*0Sstevel@tonic-gate	stx	%o0,[%i2+8]
337*0Sstevel@tonic-gate
338*0Sstevel@tonic-gate3:
339*0Sstevel@tonic-gate	st	%i4,[%i3]
340*0Sstevel@tonic-gate	ret
341*0Sstevel@tonic-gate	restore
342*0Sstevel@tonic-gate
343*0Sstevel@tonic-gate	SET_SIZE(__quad_mag_add)
344*0Sstevel@tonic-gate
345*0Sstevel@tonic-gate! /*
346*0Sstevel@tonic-gate!  * __quad_mag_sub(x, y, z, fsr)
347*0Sstevel@tonic-gate!  *
348*0Sstevel@tonic-gate!  * Sets *z = *x - *y, rounded according to the rounding mode in *fsr,
349*0Sstevel@tonic-gate!  * and updates the current exceptions in *fsr.  This routine assumes
350*0Sstevel@tonic-gate!  * *x and *y are finite, with opposite signs (i.e., a subtraction of
351*0Sstevel@tonic-gate!  * magnitudes), |*x| >= |*y|, and *z already has its sign bit set.
352*0Sstevel@tonic-gate!  */
353*0Sstevel@tonic-gate! void
354*0Sstevel@tonic-gate! __quad_mag_sub(const union longdouble *x, const union longdouble *y,
355*0Sstevel@tonic-gate! 	union longdouble *z, unsigned int *fsr)
356*0Sstevel@tonic-gate! {
357*0Sstevel@tonic-gate! 	unsigned long	lx, ly, frac, sticky;
358*0Sstevel@tonic-gate! 	unsigned int	ex, ey, gr, borrow, rm;
359*0Sstevel@tonic-gate! 	int		e;
360*0Sstevel@tonic-gate!
361*0Sstevel@tonic-gate! 	/* get the leading significand double-words and exponents */
362*0Sstevel@tonic-gate! 	ex = (x->ll.msll >> 48) & 0x7fff;
363*0Sstevel@tonic-gate! 	lx = x->ll.msll & ~0xffff000000000000ul;
364*0Sstevel@tonic-gate! 	if (ex == 0)
365*0Sstevel@tonic-gate! 		ex = 1;
366*0Sstevel@tonic-gate! 	else
367*0Sstevel@tonic-gate! 		lx |= 0x0001000000000000ul;
368*0Sstevel@tonic-gate!
369*0Sstevel@tonic-gate! 	ey = (y->ll.msll >> 48) & 0x7fff;
370*0Sstevel@tonic-gate! 	ly = y->ll.msll & ~0xffff000000000000ul;
371*0Sstevel@tonic-gate! 	if (ey == 0)
372*0Sstevel@tonic-gate! 		ey = 1;
373*0Sstevel@tonic-gate! 	else
374*0Sstevel@tonic-gate! 		ly |= 0x0001000000000000ul;
375*0Sstevel@tonic-gate!
376*0Sstevel@tonic-gate! 	/* prenormalize y */
377*0Sstevel@tonic-gate! 	e = (int) ex - (int) ey;
378*0Sstevel@tonic-gate! 	gr = sticky = 0;
379*0Sstevel@tonic-gate! 	if (e > 114) {
380*0Sstevel@tonic-gate! 		sticky = ly | y->ll.frac;
381*0Sstevel@tonic-gate! 		ly = frac = 0;
382*0Sstevel@tonic-gate! 	} else {
383*0Sstevel@tonic-gate! 		frac = y->ll.frac;
384*0Sstevel@tonic-gate! 		if (e >= 64) {
385*0Sstevel@tonic-gate! 			gr = frac >> 62;
386*0Sstevel@tonic-gate! 			sticky = frac << 2;
387*0Sstevel@tonic-gate! 			frac = ly;
388*0Sstevel@tonic-gate! 			ly = 0;
389*0Sstevel@tonic-gate! 			e -= 64;
390*0Sstevel@tonic-gate! 		}
391*0Sstevel@tonic-gate! 		if (e > 1) {
392*0Sstevel@tonic-gate! 			sticky |= gr | (frac & ((1ul << (e - 2)) - 1));
393*0Sstevel@tonic-gate! 			gr = (frac >> (e - 2)) & 3;
394*0Sstevel@tonic-gate! 			frac = (frac >> e) | (ly << (64 - e));
395*0Sstevel@tonic-gate! 			ly >>= e;
396*0Sstevel@tonic-gate! 		} else if (e == 1) {
397*0Sstevel@tonic-gate! 			sticky |= (gr & 1);
398*0Sstevel@tonic-gate! 			gr = (gr >> 1) | ((frac & 1) << 1);
399*0Sstevel@tonic-gate! 			frac = (frac >> 1) | (ly << 63);
400*0Sstevel@tonic-gate! 			ly >>= 1;
401*0Sstevel@tonic-gate! 		}
402*0Sstevel@tonic-gate! 	}
403*0Sstevel@tonic-gate!
404*0Sstevel@tonic-gate! 	/* complement guard, round, and sticky as need be */
405*0Sstevel@tonic-gate! 	gr <<= 1;
406*0Sstevel@tonic-gate! 	if (sticky)
407*0Sstevel@tonic-gate! 		gr |= 1;
408*0Sstevel@tonic-gate! 	gr = (-gr & 7);
409*0Sstevel@tonic-gate!	if (gr)
410*0Sstevel@tonic-gate!		if (++frac == 0)
411*0Sstevel@tonic-gate!			ly++;
412*0Sstevel@tonic-gate!
413*0Sstevel@tonic-gate! 	/* subtract, propagating borrows */
414*0Sstevel@tonic-gate! 	frac = x->ll.frac - frac;
415*0Sstevel@tonic-gate! 	lx -= ly;
416*0Sstevel@tonic-gate! 	if (frac > x->ll.frac)
417*0Sstevel@tonic-gate! 		lx--;
418*0Sstevel@tonic-gate!
419*0Sstevel@tonic-gate! 	/* get the rounding mode */
420*0Sstevel@tonic-gate! 	rm = *fsr >> 30;
421*0Sstevel@tonic-gate!
422*0Sstevel@tonic-gate! 	/* handle zero result */
423*0Sstevel@tonic-gate! 	if (!(lx | frac | gr)) {
424*0Sstevel@tonic-gate! 		z->l.msw = ((rm == FSR_RM)? 0x80000000 : 0);
425*0Sstevel@tonic-gate! 		z->l.frac2 = z->l.frac3 = z->l.frac4 = 0;
426*0Sstevel@tonic-gate! 		return;
427*0Sstevel@tonic-gate! 	}
428*0Sstevel@tonic-gate!
429*0Sstevel@tonic-gate! 	/* postnormalize */
430*0Sstevel@tonic-gate! 	if (lx < 0x0001000000000000ul) {
431*0Sstevel@tonic-gate! 		/* if cancellation occurred or the exponent is 1,
432*0Sstevel@tonic-gate! 		   the result is exact */
433*0Sstevel@tonic-gate! 		if (lx < 0x0000800000000000ul || ex == 1) {
434*0Sstevel@tonic-gate! 			if ((lx | (frac & 0xfffe000000000000ul)) == 0 &&
435*0Sstevel@tonic-gate! 				ex > 64) {
436*0Sstevel@tonic-gate! 				lx = frac;
437*0Sstevel@tonic-gate! 				frac = (unsigned long) gr << 61;
438*0Sstevel@tonic-gate! 				gr = 0;
439*0Sstevel@tonic-gate! 				ex -= 64;
440*0Sstevel@tonic-gate! 			}
441*0Sstevel@tonic-gate! 			while (lx < 0x0001000000000000ul && ex > 1) {
442*0Sstevel@tonic-gate! 				lx = (lx << 1) | (frac >> 63);
443*0Sstevel@tonic-gate! 				frac = (frac << 1) | (gr >> 2);
444*0Sstevel@tonic-gate! 				gr = 0;
445*0Sstevel@tonic-gate! 				ex--;
446*0Sstevel@tonic-gate! 			}
447*0Sstevel@tonic-gate! 			if (lx >= 0x0001000000000000ul)
448*0Sstevel@tonic-gate! 				z->l.msw |= (ex << 16);
449*0Sstevel@tonic-gate! 			z->l.msw |= ((lx >> 32) & 0xffff);
450*0Sstevel@tonic-gate! 			z->l.frac2 = (lx & 0xffffffff);
451*0Sstevel@tonic-gate! 			z->ll.frac = frac;
452*0Sstevel@tonic-gate!
453*0Sstevel@tonic-gate! 			/* if the result is tiny and underflow trapping is
454*0Sstevel@tonic-gate! 			   enabled, simulate underflow */
455*0Sstevel@tonic-gate! 			if (lx < 0x0001000000000000ul && (*fsr & FSR_UFM))
456*0Sstevel@tonic-gate! 				*fsr |= FSR_UFC;
457*0Sstevel@tonic-gate! 			return;
458*0Sstevel@tonic-gate! 		}
459*0Sstevel@tonic-gate!
460*0Sstevel@tonic-gate! 		/* otherwise we only borrowed one place */
461*0Sstevel@tonic-gate! 		lx = (lx << 1) | (frac >> 63);
462*0Sstevel@tonic-gate! 		frac = (frac << 1) | (gr >> 2);
463*0Sstevel@tonic-gate! 		gr &= 3;
464*0Sstevel@tonic-gate! 		ex--;
465*0Sstevel@tonic-gate! 	}
466*0Sstevel@tonic-gate! 	else
467*0Sstevel@tonic-gate! 		gr = (gr >> 1) | (gr & 1);
468*0Sstevel@tonic-gate!
469*0Sstevel@tonic-gate! 	/* fudge directed rounding modes as though the result were positive */
470*0Sstevel@tonic-gate! 	if (z->l.msw)
471*0Sstevel@tonic-gate! 		rm ^= (rm >> 1);
472*0Sstevel@tonic-gate!
473*0Sstevel@tonic-gate! 	/* see if we need to round */
474*0Sstevel@tonic-gate! 	if (gr) {
475*0Sstevel@tonic-gate! 		*fsr |= FSR_NXC;
476*0Sstevel@tonic-gate!
477*0Sstevel@tonic-gate! 		/* round up if necessary */
478*0Sstevel@tonic-gate! 		if (rm == FSR_RP || (rm == FSR_RN && (gr & 2) &&
479*0Sstevel@tonic-gate! 			((gr & 1) || (frac & 1)))) {
480*0Sstevel@tonic-gate! 			if (++frac == 0)
481*0Sstevel@tonic-gate! 				if (++lx >= 0x0002000000000000ul) {
482*0Sstevel@tonic-gate! 					lx >>= 1;
483*0Sstevel@tonic-gate! 					ex++;
484*0Sstevel@tonic-gate! 				}
485*0Sstevel@tonic-gate! 		}
486*0Sstevel@tonic-gate! 	}
487*0Sstevel@tonic-gate!
488*0Sstevel@tonic-gate! 	/* store the result */
489*0Sstevel@tonic-gate! 	z->l.msw |= (ex << 16) | ((lx >> 32) & 0xffff);
490*0Sstevel@tonic-gate! 	z->l.frac2 = (lx & 0xffffffff);
491*0Sstevel@tonic-gate! 	z->ll.frac = frac;
492*0Sstevel@tonic-gate! }
493*0Sstevel@tonic-gate
494*0Sstevel@tonic-gate	ENTRY(__quad_mag_sub)
495*0Sstevel@tonic-gate	save	%sp,-SA(MINFRAME),%sp
496*0Sstevel@tonic-gate
497*0Sstevel@tonic-gate	sethi	%hi(0xffff0000),%g1
498*0Sstevel@tonic-gate	sllx	%g1,32,%g1		! g1 = 0xffff000000000000
499*0Sstevel@tonic-gate
500*0Sstevel@tonic-gate	sethi	%hi(0x7fff),%l7
501*0Sstevel@tonic-gate	or	%l7,%lo(0x7fff),%l7	! l7 = 0x7fff
502*0Sstevel@tonic-gate
503*0Sstevel@tonic-gate	ldx	[%i0],%o0
504*0Sstevel@tonic-gate	srlx	%o0,48,%l0
505*0Sstevel@tonic-gate	andcc	%l0,%l7,%l0		! l0 = ex
506*0Sstevel@tonic-gate	beq,pn	%icc,1f
507*0Sstevel@tonic-gate	andn	%o0,%g1,%o0		! o0 = lx
508*0Sstevel@tonic-gate	ba,pt	%icc,2f
509*0Sstevel@tonic-gate	sub	%o0,%g1,%o0
510*0Sstevel@tonic-gate1:
511*0Sstevel@tonic-gate	mov	1,%l0
512*0Sstevel@tonic-gate2:
513*0Sstevel@tonic-gate
514*0Sstevel@tonic-gate	ldx	[%i1],%o1
515*0Sstevel@tonic-gate	srlx	%o1,48,%l1
516*0Sstevel@tonic-gate	andcc	%l1,%l7,%l1		! l1 = ey
517*0Sstevel@tonic-gate	beq,pn	%icc,1f
518*0Sstevel@tonic-gate	andn	%o1,%g1,%o1		! o1 = ly
519*0Sstevel@tonic-gate	ba,pt	%icc,2f
520*0Sstevel@tonic-gate	sub	%o1,%g1,%o1
521*0Sstevel@tonic-gate1:
522*0Sstevel@tonic-gate	mov	1,%l1
523*0Sstevel@tonic-gate2:
524*0Sstevel@tonic-gate
525*0Sstevel@tonic-gate	sub	%l0,%l1,%l1		! l1 = e = ex - ey
526*0Sstevel@tonic-gate	cmp	%l1,114			! see if we need to prenormalize y
527*0Sstevel@tonic-gate	bg,pn	%icc,1f
528*0Sstevel@tonic-gate	mov	0,%l6			! l6 = gr
529*0Sstevel@tonic-gate	mov	0,%o7			! o7 = sticky
530*0Sstevel@tonic-gate	cmp	%l1,64
531*0Sstevel@tonic-gate	bl,pt	%icc,3f
532*0Sstevel@tonic-gate	ldx	[%i1+8],%o2		! o2 = frac
533*0Sstevel@tonic-gate	srlx	%o2,62,%l6
534*0Sstevel@tonic-gate	sllx	%o2,2,%o7		! lop off top two bits
535*0Sstevel@tonic-gate	mov	%o1,%o2
536*0Sstevel@tonic-gate	mov	0,%o1
537*0Sstevel@tonic-gate	sub	%l1,64,%l1
538*0Sstevel@tonic-gate3:
539*0Sstevel@tonic-gate	cmp	%l1,1
540*0Sstevel@tonic-gate	ble,pn	%icc,4f
541*0Sstevel@tonic-gate	sub	%l1,2,%l2		! shift more than one bit
542*0Sstevel@tonic-gate	mov	1,%o3
543*0Sstevel@tonic-gate	sllx	%o3,%l2,%o3
544*0Sstevel@tonic-gate	sub	%o3,1,%o3
545*0Sstevel@tonic-gate	and	%o3,%o2,%o3
546*0Sstevel@tonic-gate	or	%o3,%l6,%o3
547*0Sstevel@tonic-gate	or	%o7,%o3,%o7
548*0Sstevel@tonic-gate	srlx	%o2,%l2,%o4
549*0Sstevel@tonic-gate	and	%o4,3,%l6
550*0Sstevel@tonic-gate	srlx	%o2,%l1,%o2
551*0Sstevel@tonic-gate	mov	64,%l3
552*0Sstevel@tonic-gate	sub	%l3,%l1,%l3
553*0Sstevel@tonic-gate	sllx	%o1,%l3,%o5
554*0Sstevel@tonic-gate	or	%o2,%o5,%o2
555*0Sstevel@tonic-gate	ba,pt	%icc,2f
556*0Sstevel@tonic-gate	srlx	%o1,%l1,%o1
557*0Sstevel@tonic-gate4:
558*0Sstevel@tonic-gate	bne,pn	%icc,2f
559*0Sstevel@tonic-gate	and	%l6,1,%o3		! shift one bit
560*0Sstevel@tonic-gate	or	%o7,%o3,%o7
561*0Sstevel@tonic-gate	and	%o2,1,%o4
562*0Sstevel@tonic-gate	sllx	%o4,1,%o4
563*0Sstevel@tonic-gate	srl	%l6,1,%l6
564*0Sstevel@tonic-gate	or	%l6,%o4,%l6
565*0Sstevel@tonic-gate	srlx	%o2,1,%o2
566*0Sstevel@tonic-gate	sllx	%o1,63,%o5
567*0Sstevel@tonic-gate	or	%o2,%o5,%o2
568*0Sstevel@tonic-gate	ba,pt	%icc,2f
569*0Sstevel@tonic-gate	srlx	%o1,1,%o1
570*0Sstevel@tonic-gate1:
571*0Sstevel@tonic-gate	ldx	[%i1+8],%o3		! (full prenormalization shift case)
572*0Sstevel@tonic-gate	or	%o1,%o3,%o7
573*0Sstevel@tonic-gate	mov	0,%o1
574*0Sstevel@tonic-gate	mov	0,%o2
575*0Sstevel@tonic-gate2:
576*0Sstevel@tonic-gate
577*0Sstevel@tonic-gate	tst	%o7			! complement guard, round, and
578*0Sstevel@tonic-gate	beq,pn	%xcc,1f			! sticky as need be
579*0Sstevel@tonic-gate	sll	%l6,1,%l6
580*0Sstevel@tonic-gate	or	%l6,1,%l6
581*0Sstevel@tonic-gate1:
582*0Sstevel@tonic-gate	subcc	%g0,%l6,%l6
583*0Sstevel@tonic-gate	beq,pn	%icc,1f
584*0Sstevel@tonic-gate	and	%l6,7,%l6
585*0Sstevel@tonic-gate	addcc	%o2,1,%o2
586*0Sstevel@tonic-gate	beq,a,pn %xcc,1f
587*0Sstevel@tonic-gate	add	%o1,1,%o1
588*0Sstevel@tonic-gate1:
589*0Sstevel@tonic-gate
590*0Sstevel@tonic-gate	ldx	[%i0+8],%o3		! subtract, propagating borrows
591*0Sstevel@tonic-gate	sub	%o3,%o2,%o2
592*0Sstevel@tonic-gate	cmp	%o3,%o2
593*0Sstevel@tonic-gate	bgeu,pt	%xcc,5f
594*0Sstevel@tonic-gate	sub	%o0,%o1,%o0
595*0Sstevel@tonic-gate	sub	%o0,1,%o0
596*0Sstevel@tonic-gate5:
597*0Sstevel@tonic-gate
598*0Sstevel@tonic-gate	ld	[%i3],%i4		! get the rounding mode
599*0Sstevel@tonic-gate	srl	%i4,30,%l3		! l3 = rm
600*0Sstevel@tonic-gate
601*0Sstevel@tonic-gate	or	%o0,%o2,%o1		! look for zero result
602*0Sstevel@tonic-gate	orcc	%o1,%l6,%g0
603*0Sstevel@tonic-gate	bne,pt	%xcc,1f
604*0Sstevel@tonic-gate	srl	%l3,1,%l4
605*0Sstevel@tonic-gate	and	%l3,%l4,%l4
606*0Sstevel@tonic-gate	sll	%l4,31,%l4
607*0Sstevel@tonic-gate	st	%l4,[%i2]
608*0Sstevel@tonic-gate	st	%g0,[%i2+4]
609*0Sstevel@tonic-gate	stx	%g0,[%i2+8]
610*0Sstevel@tonic-gate	ret
611*0Sstevel@tonic-gate	restore
612*0Sstevel@tonic-gate
613*0Sstevel@tonic-gate1:
614*0Sstevel@tonic-gate	addcc	%o0,%g1,%g0		! postnormalize
615*0Sstevel@tonic-gate	bl,pt	%xcc,1f
616*0Sstevel@tonic-gate	ld	[%i2],%l4		! l4 = z->l.msw
617*0Sstevel@tonic-gate	and	%l6,1,%l5		! (no cancellation or borrow case)
618*0Sstevel@tonic-gate	srl	%l6,1,%l6
619*0Sstevel@tonic-gate	ba,pt	%icc,2f
620*0Sstevel@tonic-gate	or	%l6,%l5,%l6
621*0Sstevel@tonic-gate1:
622*0Sstevel@tonic-gate	srax	%g1,1,%o7
623*0Sstevel@tonic-gate	addcc	%o0,%o7,%g0
624*0Sstevel@tonic-gate	bl,pn	%xcc,1f
625*0Sstevel@tonic-gate	cmp	%l0,1
626*0Sstevel@tonic-gate	beq,pt	%icc,1f
627*0Sstevel@tonic-gate	srlx	%o2,63,%o3		! borrowed one place
628*0Sstevel@tonic-gate	sllx	%o0,1,%o0
629*0Sstevel@tonic-gate	or	%o0,%o3,%o0
630*0Sstevel@tonic-gate	srl	%l6,2,%o4
631*0Sstevel@tonic-gate	sllx	%o2,1,%o2
632*0Sstevel@tonic-gate	or	%o2,%o4,%o2
633*0Sstevel@tonic-gate	and	%l6,3,%l6
634*0Sstevel@tonic-gate	ba,pt	%icc,2f
635*0Sstevel@tonic-gate	sub	%l0,1,%l0
636*0Sstevel@tonic-gate1:
637*0Sstevel@tonic-gate	srlx	%o2,49,%o3		! cancellation or tiny result
638*0Sstevel@tonic-gate	orcc	%o0,%o3,%g0
639*0Sstevel@tonic-gate	bne,pt	%xcc,1f
640*0Sstevel@tonic-gate	cmp	%l0,64
641*0Sstevel@tonic-gate	ble,pn	%icc,1f
642*0Sstevel@tonic-gate	nop
643*0Sstevel@tonic-gate	mov	%o2,%o0
644*0Sstevel@tonic-gate	sllx	%l6,61,%o2
645*0Sstevel@tonic-gate	mov	0,%l6
646*0Sstevel@tonic-gate	sub	%l0,64,%l0
647*0Sstevel@tonic-gate1:
648*0Sstevel@tonic-gate	addcc	%o0,%g1,%g0		! normalization loop
649*0Sstevel@tonic-gate	bge,pn	%xcc,1f
650*0Sstevel@tonic-gate	cmp	%l0,1
651*0Sstevel@tonic-gate	ble,pn	%icc,1f
652*0Sstevel@tonic-gate	srl	%l6,2,%l6
653*0Sstevel@tonic-gate	srlx	%o2,63,%o3
654*0Sstevel@tonic-gate	sllx	%o0,1,%o0
655*0Sstevel@tonic-gate	or	%o0,%o3,%o0
656*0Sstevel@tonic-gate	sllx	%o2,1,%o2
657*0Sstevel@tonic-gate	or	%o2,%l6,%o2
658*0Sstevel@tonic-gate	ba,pt	%icc,1b
659*0Sstevel@tonic-gate	sub	%l0,1,%l0
660*0Sstevel@tonic-gate1:
661*0Sstevel@tonic-gate	sllx	%o0,16,%o1
662*0Sstevel@tonic-gate	srlx	%o1,48,%l5
663*0Sstevel@tonic-gate	or	%l4,%l5,%l4
664*0Sstevel@tonic-gate	addcc	%o0,%g1,%g0		! see if result is tiny
665*0Sstevel@tonic-gate	bl,pn	%xcc,1f
666*0Sstevel@tonic-gate	sll	%l0,16,%l5
667*0Sstevel@tonic-gate	or	%l4,%l5,%l4
668*0Sstevel@tonic-gate1:
669*0Sstevel@tonic-gate	st	%l4,[%i2]
670*0Sstevel@tonic-gate	st	%o0,[%i2+4]
671*0Sstevel@tonic-gate	bge,pt	%xcc,1f
672*0Sstevel@tonic-gate	stx	%o2,[%i2+8]
673*0Sstevel@tonic-gate	srl	%i4,23,%i5
674*0Sstevel@tonic-gate	andcc	%i5,4,%g0		! see if we need to raise underflow
675*0Sstevel@tonic-gate	beq,pt	%icc,1f
676*0Sstevel@tonic-gate	or	%i4,4,%i4
677*0Sstevel@tonic-gate	st	%i4,[%i3]
678*0Sstevel@tonic-gate1:
679*0Sstevel@tonic-gate	ret
680*0Sstevel@tonic-gate	restore
681*0Sstevel@tonic-gate
682*0Sstevel@tonic-gate2:
683*0Sstevel@tonic-gate	tst	%l4			! fudge directect rounding modes
684*0Sstevel@tonic-gate	beq,pn	%icc,1f
685*0Sstevel@tonic-gate	srl	%l3,1,%l5
686*0Sstevel@tonic-gate	xor	%l3,%l5,%l3
687*0Sstevel@tonic-gate1:
688*0Sstevel@tonic-gate
689*0Sstevel@tonic-gate	tst	%l6			! see if we need to round
690*0Sstevel@tonic-gate	beq,pn	%icc,1f
691*0Sstevel@tonic-gate	or	%i4,1,%i4
692*0Sstevel@tonic-gate	st	%i4,[%i3]
693*0Sstevel@tonic-gate	andcc	%l3,1,%g0
694*0Sstevel@tonic-gate	bne,pn	%icc,1f
695*0Sstevel@tonic-gate	tst	%l3
696*0Sstevel@tonic-gate	bne,pn	%icc,2f
697*0Sstevel@tonic-gate	andcc	%l6,2,%g0
698*0Sstevel@tonic-gate	beq,pn	%icc,1f
699*0Sstevel@tonic-gate	or	%l6,%o2,%o3
700*0Sstevel@tonic-gate	andcc	%o3,1,%o3
701*0Sstevel@tonic-gate	beq,pn	%xcc,1f
702*0Sstevel@tonic-gate	nop
703*0Sstevel@tonic-gate2:
704*0Sstevel@tonic-gate	addcc	%o2,1,%o2		! round up and check for carry
705*0Sstevel@tonic-gate	bne,pt	%xcc,1f
706*0Sstevel@tonic-gate	nop
707*0Sstevel@tonic-gate	add	%o0,1,%o0
708*0Sstevel@tonic-gate	srlx	%o0,49,%o1
709*0Sstevel@tonic-gate	tst	%o1
710*0Sstevel@tonic-gate	beq,pt	%icc,1f
711*0Sstevel@tonic-gate	nop
712*0Sstevel@tonic-gate	srlx	%o0,1,%o0
713*0Sstevel@tonic-gate	add	%l0,1,%l0
714*0Sstevel@tonic-gate1:
715*0Sstevel@tonic-gate
716*0Sstevel@tonic-gate	sllx	%o0,16,%o1
717*0Sstevel@tonic-gate	srlx	%o1,48,%o1
718*0Sstevel@tonic-gate	or	%l4,%o1,%l4
719*0Sstevel@tonic-gate	sll	%l0,16,%l5
720*0Sstevel@tonic-gate	or	%l4,%l5,%l4
721*0Sstevel@tonic-gate	st	%l4,[%i2]
722*0Sstevel@tonic-gate	st	%o0,[%i2+4]
723*0Sstevel@tonic-gate	stx	%o2,[%i2+8]
724*0Sstevel@tonic-gate	ret
725*0Sstevel@tonic-gate	restore
726*0Sstevel@tonic-gate
727*0Sstevel@tonic-gate	SET_SIZE(__quad_mag_sub)
728