xref: /onnv-gate/usr/src/lib/libc/sparc/crt/hwmuldiv.s (revision 7298:b69e27387f74)
10Sstevel@tonic-gate/*
20Sstevel@tonic-gate * CDDL HEADER START
30Sstevel@tonic-gate *
40Sstevel@tonic-gate * The contents of this file are subject to the terms of the
5*7298SMark.J.Nelson@Sun.COM * Common Development and Distribution License (the "License").
6*7298SMark.J.Nelson@Sun.COM * You may not use this file except in compliance with the License.
70Sstevel@tonic-gate *
80Sstevel@tonic-gate * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
90Sstevel@tonic-gate * or http://www.opensolaris.org/os/licensing.
100Sstevel@tonic-gate * See the License for the specific language governing permissions
110Sstevel@tonic-gate * and limitations under the License.
120Sstevel@tonic-gate *
130Sstevel@tonic-gate * When distributing Covered Code, include this CDDL HEADER in each
140Sstevel@tonic-gate * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
150Sstevel@tonic-gate * If applicable, add the following below this CDDL HEADER, with the
160Sstevel@tonic-gate * fields enclosed by brackets "[]" replaced with your own identifying
170Sstevel@tonic-gate * information: Portions Copyright [yyyy] [name of copyright owner]
180Sstevel@tonic-gate *
190Sstevel@tonic-gate * CDDL HEADER END
200Sstevel@tonic-gate */
210Sstevel@tonic-gate/*
22544Sjwadams * Copyright 2005 Sun Microsystems, Inc.  All rights reserved.
23544Sjwadams * Use is subject to license terms.
240Sstevel@tonic-gate */
250Sstevel@tonic-gate
26*7298SMark.J.Nelson@Sun.COM	.file	"hwmuldiv.s"
270Sstevel@tonic-gate
280Sstevel@tonic-gate#include <sys/asm_linkage.h>
290Sstevel@tonic-gate
300Sstevel@tonic-gate/*
310Sstevel@tonic-gate * Versions of .mul .umul .div .udiv .rem .urem written using
320Sstevel@tonic-gate * appropriate SPARC V8 instructions.
330Sstevel@tonic-gate */
340Sstevel@tonic-gate
350Sstevel@tonic-gate	ENTRY(.mul)
360Sstevel@tonic-gate	smul	%o0, %o1, %o0
370Sstevel@tonic-gate	rd	%y, %o1
380Sstevel@tonic-gate	sra	%o0, 31, %o2
390Sstevel@tonic-gate	retl
400Sstevel@tonic-gate	cmp	%o1, %o2	! return with Z set if %y == (%o0 >> 31)
410Sstevel@tonic-gate	SET_SIZE(.mul)
420Sstevel@tonic-gate
430Sstevel@tonic-gate	ENTRY(.umul)
440Sstevel@tonic-gate	umul	%o0, %o1, %o0
450Sstevel@tonic-gate	rd	%y, %o1
460Sstevel@tonic-gate	retl
470Sstevel@tonic-gate	tst	%o1		! return with Z set if high order bits are zero
480Sstevel@tonic-gate	SET_SIZE(.umul)
490Sstevel@tonic-gate
500Sstevel@tonic-gate	ENTRY(.div)
510Sstevel@tonic-gate	sra	%o0, 31, %o2
520Sstevel@tonic-gate	wr	%g0, %o2, %y
530Sstevel@tonic-gate	nop
540Sstevel@tonic-gate	nop
550Sstevel@tonic-gate	nop
560Sstevel@tonic-gate	sdivcc	%o0, %o1, %o0
570Sstevel@tonic-gate	bvs,a	1f
580Sstevel@tonic-gate	xnor	%o0, %g0, %o0	! Corbett Correction Factor
590Sstevel@tonic-gate1:	retl
600Sstevel@tonic-gate	nop
610Sstevel@tonic-gate	SET_SIZE(.div)
620Sstevel@tonic-gate
630Sstevel@tonic-gate	ENTRY(.udiv)
640Sstevel@tonic-gate	wr	%g0, %g0, %y
650Sstevel@tonic-gate	nop
660Sstevel@tonic-gate	nop
670Sstevel@tonic-gate	retl
680Sstevel@tonic-gate	udiv	%o0, %o1, %o0
690Sstevel@tonic-gate	SET_SIZE(.udiv)
700Sstevel@tonic-gate
710Sstevel@tonic-gate	ENTRY(.rem)
720Sstevel@tonic-gate	sra	%o0, 31, %o4
730Sstevel@tonic-gate	wr	%o4, %g0, %y
740Sstevel@tonic-gate	nop
750Sstevel@tonic-gate	nop
760Sstevel@tonic-gate	nop
770Sstevel@tonic-gate	sdivcc	%o0, %o1, %o2
780Sstevel@tonic-gate	bvs,a	1f
790Sstevel@tonic-gate	xnor	%o2, %g0, %o2	! Corbett Correction Factor
800Sstevel@tonic-gate1:	smul	%o2, %o1, %o2
810Sstevel@tonic-gate	retl
820Sstevel@tonic-gate	sub	%o0, %o2, %o0
830Sstevel@tonic-gate	SET_SIZE(.rem)
840Sstevel@tonic-gate
850Sstevel@tonic-gate	ENTRY(.urem)
860Sstevel@tonic-gate	wr	%g0, %g0, %y
870Sstevel@tonic-gate	nop
880Sstevel@tonic-gate	nop
890Sstevel@tonic-gate	nop
900Sstevel@tonic-gate	udiv	%o0, %o1, %o2
910Sstevel@tonic-gate	umul	%o2, %o1, %o2
920Sstevel@tonic-gate	retl
930Sstevel@tonic-gate	sub	%o0, %o2, %o0
940Sstevel@tonic-gate	SET_SIZE(.urem)
95544Sjwadams
96544Sjwadams/*
97544Sjwadams * v8plus versions of __{u,}{mul,div,rem}64 compiler support routines
98544Sjwadams */
99544Sjwadams
100544Sjwadams/*
101544Sjwadams * Convert 32-bit arg pairs in %o0:o1 and %o2:%o3 to 64-bit args in %o1 and %o2
102544Sjwadams */
103544Sjwadams#define	ARGS_TO_64				\
104544Sjwadams	sllx	%o0, 32, %o0;			\
105544Sjwadams	srl	%o1, 0, %o1;			\
106544Sjwadams	sllx	%o2, 32, %o2;			\
107544Sjwadams	srl	%o3, 0, %o3;			\
108544Sjwadams	or	%o0, %o1, %o1;			\
109544Sjwadams	or	%o2, %o3, %o2
110544Sjwadams
111544Sjwadams!
112544Sjwadams! division, signed
113544Sjwadams!
114544Sjwadams	ENTRY(__div64)
115544Sjwadams	ARGS_TO_64
116544Sjwadams	sdivx	%o1, %o2, %o1
117544Sjwadams	retl
118544Sjwadams	srax	%o1, 32, %o0
119544Sjwadams	SET_SIZE(__div64)
120544Sjwadams
121544Sjwadams!
122544Sjwadams! division, unsigned
123544Sjwadams!
124544Sjwadams	ENTRY(__udiv64)
125544Sjwadams	ARGS_TO_64
126544Sjwadams	udivx	%o1, %o2, %o1
127544Sjwadams	retl
128544Sjwadams	srax	%o1, 32, %o0
129544Sjwadams	SET_SIZE(__udiv64)
130544Sjwadams
131544Sjwadams!
132544Sjwadams! multiplication, signed and unsigned
133544Sjwadams!
134544Sjwadams	ENTRY(__mul64)
135544Sjwadams	ALTENTRY(__umul64)
136544Sjwadams	ARGS_TO_64
137544Sjwadams	sub	%o1, %o2, %o0	! %o0 = a - b
138544Sjwadams	movrlz	%o0, %g0, %o0	! %o0 = (a < b) ? 0 : a - b
139544Sjwadams	sub	%o1, %o0, %o1	! %o1 = (a < b) ? a : b = min(a, b)
140544Sjwadams	add	%o2, %o0, %o2	! %o2 = (a < b) ? b : a = max(a, b)
141544Sjwadams	mulx	%o1, %o2, %o1	! min(a, b) in "rs1" for early exit
142544Sjwadams	retl
143544Sjwadams	srax	%o1, 32, %o0
144544Sjwadams	SET_SIZE(__mul64)
145544Sjwadams	SET_SIZE(__umul64)
146544Sjwadams
147544Sjwadams!
148544Sjwadams! unsigned remainder
149544Sjwadams!
150544Sjwadams	ENTRY(__urem64)
151544Sjwadams	ARGS_TO_64
152544Sjwadams	udivx	%o1, %o2, %o3
153544Sjwadams	mulx	%o3, %o2, %o3
154544Sjwadams	sub	%o1, %o3, %o1
155544Sjwadams	retl
156544Sjwadams	srax	%o1, 32, %o0
157544Sjwadams	SET_SIZE(__urem64)
158544Sjwadams
159544Sjwadams!
160544Sjwadams! signed remainder
161544Sjwadams!
162544Sjwadams	ENTRY(__rem64)
163544Sjwadams	ARGS_TO_64
164544Sjwadams	sdivx	%o1, %o2, %o3
165544Sjwadams	mulx	%o2, %o3, %o3
166544Sjwadams	sub	%o1, %o3, %o1
167544Sjwadams	retl
168544Sjwadams	srax	%o1, 32, %o0
169544Sjwadams	SET_SIZE(__rem64)
170