[mirror_ubuntu-bionic-kernel.git] / arch / alpha / lib / divide.S

/* SPDX-License-Identifier: GPL-2.0 */
/*
 * arch/alpha/lib/divide.S
 *
 * (C) 1995 Linus Torvalds
 *
 * Alpha division..
 */

/*
 * The alpha chip doesn't provide hardware division, so we have to do it
 * by hand.  The compiler expects the functions
 *
 *	__divqu: 64-bit unsigned long divide
 *	__remqu: 64-bit unsigned long remainder
 *	__divqs/__remqs: signed 64-bit
 *	__divlu/__remlu: unsigned 32-bit
 *	__divls/__remls: signed 32-bit
 *
 * These are not normal C functions: instead of the normal
 * calling sequence, these expect their arguments in registers
 * $24 and $25, and return the result in $27. Register $28 may
 * be clobbered (assembly temporary), anything else must be saved. 
 *
 * In short: painful.
 *
 * This is a rather simple bit-at-a-time algorithm: it's very good
 * at dividing random 64-bit numbers, but the more usual case where
 * the divisor is small is handled better by the DEC algorithm
 * using lookup tables. This uses much less memory, though, and is
 * nicer on the cache.. Besides, I don't know the copyright status
 * of the DEC code.
 */

/*
 * My temporaries:
 *	$0 - current bit
 *	$1 - shifted divisor
 *	$2 - modulus/quotient
 *
 *	$23 - return address
 *	$24 - dividend
 *	$25 - divisor
 *
 *	$27 - quotient/modulus
 *	$28 - compare status
 */

#include <asm/export.h>
#define halt .long 0

/*
 * Select function type and registers
 */
#define mask	$0
#define divisor	$1
#define compare $28
#define tmp1	$3
#define tmp2	$4

#ifdef DIV
#define DIV_ONLY(x,y...) x,##y
#define MOD_ONLY(x,y...)
#define func(x) __div##x
#define modulus $2
#define quotient $27
#define GETSIGN(x) xor $24,$25,x
#define STACK 48
#else
#define DIV_ONLY(x,y...)
#define MOD_ONLY(x,y...) x,##y
#define func(x) __rem##x
#define modulus $27
#define quotient $2
#define GETSIGN(x) bis $24,$24,x
#define STACK 32
#endif

/*
 * For 32-bit operations, we need to extend to 64-bit
 */
#ifdef INTSIZE
#define ufunction func(lu)
#define sfunction func(l)
#define LONGIFY(x) zapnot x,15,x
#define SLONGIFY(x) addl x,0,x
#else
#define ufunction func(qu)
#define sfunction func(q)
#define LONGIFY(x)
#define SLONGIFY(x)
#endif

.set noat
.align	3
.globl	ufunction
.ent	ufunction
ufunction:
	subq	$30,STACK,$30
	.frame	$30,STACK,$23
	.prologue 0

7:	stq	$1, 0($30)
	bis	$25,$25,divisor
	stq	$2, 8($30)
	bis	$24,$24,modulus
	stq	$0,16($30)
	bis	$31,$31,quotient
	LONGIFY(divisor)
	stq	tmp1,24($30)
	LONGIFY(modulus)
	bis	$31,1,mask
	DIV_ONLY(stq tmp2,32($30))
	beq	divisor, 9f			/* div by zero */

#ifdef INTSIZE
	/*
	 * shift divisor left, using 3-bit shifts for
	 * 32-bit divides as we can't overflow. Three-bit
	 * shifts will result in looping three times less
	 * here, but can result in two loops more later.
	 * Thus using a large shift isn't worth it (and
	 * s8add pairs better than a sll..)
	 */
1:	cmpult	divisor,modulus,compare
	s8addq	divisor,$31,divisor
	s8addq	mask,$31,mask
	bne	compare,1b
#else
1:	cmpult	divisor,modulus,compare
	blt     divisor, 2f
	addq	divisor,divisor,divisor
	addq	mask,mask,mask
	bne	compare,1b
	unop
#endif

	/* ok, start to go right again.. */
2:	DIV_ONLY(addq quotient,mask,tmp2)
	srl	mask,1,mask
	cmpule	divisor,modulus,compare
	subq	modulus,divisor,tmp1
	DIV_ONLY(cmovne compare,tmp2,quotient)
	srl	divisor,1,divisor
	cmovne	compare,tmp1,modulus
	bne	mask,2b

9:	ldq	$1, 0($30)
	ldq	$2, 8($30)
	ldq	$0,16($30)
	ldq	tmp1,24($30)
	DIV_ONLY(ldq tmp2,32($30))
	addq	$30,STACK,$30
	ret	$31,($23),1
	.end	ufunction
EXPORT_SYMBOL(ufunction)

/*
 * Uhh.. Ugly signed division. I'd rather not have it at all, but
 * it's needed in some circumstances. There are different ways to
 * handle this, really. This does:
 * 	-a / b = a / -b = -(a / b)
 *	-a % b = -(a % b)
 *	a % -b = a % b
 * which is probably not the best solution, but at least should
 * have the property that (x/y)*y + (x%y) = x.
 */
.align 3
.globl	sfunction
.ent	sfunction
sfunction:
	subq	$30,STACK,$30
	.frame	$30,STACK,$23
	.prologue 0
	bis	$24,$25,$28
	SLONGIFY($28)
	bge	$28,7b
	stq	$24,0($30)
	subq	$31,$24,$28
	stq	$25,8($30)
	cmovlt	$24,$28,$24	/* abs($24) */
	stq	$23,16($30)
	subq	$31,$25,$28
	stq	tmp1,24($30)
	cmovlt	$25,$28,$25	/* abs($25) */
	unop
	bsr	$23,ufunction
	ldq	$24,0($30)
	ldq	$25,8($30)
	GETSIGN($28)
	subq	$31,$27,tmp1
	SLONGIFY($28)
	ldq	$23,16($30)
	cmovlt	$28,tmp1,$27
	ldq	tmp1,24($30)
	addq	$30,STACK,$30
	ret	$31,($23),1
	.end	sfunction
EXPORT_SYMBOL(sfunction)
Commit	Line	Data
b2441318	1	/* SPDX-License-Identifier: GPL-2.0 */
1da177e4 LT	2	/*
	3	* arch/alpha/lib/divide.S
	4	*
	5	* (C) 1995 Linus Torvalds
	6	*
	7	* Alpha division..
	8	*/
	9
	10	/*
	11	* The alpha chip doesn't provide hardware division, so we have to do it
	12	* by hand. The compiler expects the functions
	13	*
	14	* __divqu: 64-bit unsigned long divide
	15	* __remqu: 64-bit unsigned long remainder
	16	* __divqs/__remqs: signed 64-bit
	17	* __divlu/__remlu: unsigned 32-bit
	18	* __divls/__remls: signed 32-bit
	19	*
	20	* These are not normal C functions: instead of the normal
	21	* calling sequence, these expect their arguments in registers
	22	* $24 and $25, and return the result in $27. Register $28 may
	23	* be clobbered (assembly temporary), anything else must be saved.
	24	*
	25	* In short: painful.
	26	*
	27	* This is a rather simple bit-at-a-time algorithm: it's very good
	28	* at dividing random 64-bit numbers, but the more usual case where
	29	* the divisor is small is handled better by the DEC algorithm
	30	* using lookup tables. This uses much less memory, though, and is
	31	* nicer on the cache.. Besides, I don't know the copyright status
	32	* of the DEC code.
	33	*/
	34
	35	/*
	36	* My temporaries:
	37	* $0 - current bit
	38	* $1 - shifted divisor
	39	* $2 - modulus/quotient
	40	*
	41	* $23 - return address
	42	* $24 - dividend
	43	* $25 - divisor
	44	*
	45	* $27 - quotient/modulus
	46	* $28 - compare status
	47	*/
	48
00fc0e0d	49	#include <asm/export.h>
1da177e4 LT	50	#define halt .long 0
	51
	52	/*
	53	* Select function type and registers
	54	*/
	55	#define mask $0
	56	#define divisor $1
	57	#define compare $28
	58	#define tmp1 $3
	59	#define tmp2 $4
	60
	61	#ifdef DIV
	62	#define DIV_ONLY(x,y...) x,##y
	63	#define MOD_ONLY(x,y...)
	64	#define func(x) __div##x
	65	#define modulus $2
	66	#define quotient $27
	67	#define GETSIGN(x) xor $24,$25,x
	68	#define STACK 48
	69	#else
	70	#define DIV_ONLY(x,y...)
	71	#define MOD_ONLY(x,y...) x,##y
	72	#define func(x) __rem##x
	73	#define modulus $27
	74	#define quotient $2
	75	#define GETSIGN(x) bis $24,$24,x
	76	#define STACK 32
	77	#endif
	78
	79	/*
	80	* For 32-bit operations, we need to extend to 64-bit
	81	*/
	82	#ifdef INTSIZE
	83	#define ufunction func(lu)
	84	#define sfunction func(l)
	85	#define LONGIFY(x) zapnot x,15,x
	86	#define SLONGIFY(x) addl x,0,x
	87	#else
	88	#define ufunction func(qu)
	89	#define sfunction func(q)
	90	#define LONGIFY(x)
	91	#define SLONGIFY(x)
	92	#endif
	93
	94	.set noat
	95	.align 3
	96	.globl ufunction
	97	.ent ufunction
	98	ufunction:
	99	subq $30,STACK,$30
	100	.frame $30,STACK,$23
	101	.prologue 0
	102
	103	7: stq $1, 0($30)
	104	bis $25,$25,divisor
	105	stq $2, 8($30)
	106	bis $24,$24,modulus
	107	stq $0,16($30)
	108	bis $31,$31,quotient
	109	LONGIFY(divisor)
	110	stq tmp1,24($30)
	111	LONGIFY(modulus)
	112	bis $31,1,mask
	113	DIV_ONLY(stq tmp2,32($30))
114	beq divisor, 9f /* div by zero */
115
116	#ifdef INTSIZE
117	/*
118	* shift divisor left, using 3-bit shifts for
119	* 32-bit divides as we can't overflow. Three-bit
120	* shifts will result in looping three times less
121	* here, but can result in two loops more later.
122	* Thus using a large shift isn't worth it (and
123	* s8add pairs better than a sll..)
124	*/
125	1: cmpult divisor,modulus,compare
126	s8addq divisor,$31,divisor
127	s8addq mask,$31,mask
128	bne compare,1b
129	#else
130	1: cmpult divisor,modulus,compare
131	blt divisor, 2f
132	addq divisor,divisor,divisor
133	addq mask,mask,mask
134	bne compare,1b
135	unop
136	#endif
137
138	/* ok, start to go right again.. */
139	2: DIV_ONLY(addq quotient,mask,tmp2)
140	srl mask,1,mask
141	cmpule divisor,modulus,compare
142	subq modulus,divisor,tmp1
143	DIV_ONLY(cmovne compare,tmp2,quotient)
144	srl divisor,1,divisor
145	cmovne compare,tmp1,modulus
146	bne mask,2b
147
148	9: ldq $1, 0($30)
149	ldq $2, 8($30)
150	ldq $0,16($30)
151	ldq tmp1,24($30)
152	DIV_ONLY(ldq tmp2,32($30))
153	addq $30,STACK,$30
154	ret $31,($23),1
155	.end ufunction
00fc0e0d	156	EXPORT_SYMBOL(ufunction)
1da177e4 LT	157
	158	/*
	159	* Uhh.. Ugly signed division. I'd rather not have it at all, but
	160	* it's needed in some circumstances. There are different ways to
	161	* handle this, really. This does:
	162	* -a / b = a / -b = -(a / b)
	163	* -a % b = -(a % b)
	164	* a % -b = a % b
	165	* which is probably not the best solution, but at least should
	166	* have the property that (x/y)*y + (x%y) = x.
	167	*/
	168	.align 3
	169	.globl sfunction
	170	.ent sfunction
	171	sfunction:
	172	subq $30,STACK,$30
	173	.frame $30,STACK,$23
	174	.prologue 0
	175	bis $24,$25,$28
	176	SLONGIFY($28)
	177	bge $28,7b
	178	stq $24,0($30)
	179	subq $31,$24,$28
	180	stq $25,8($30)
	181	cmovlt $24,$28,$24 /* abs($24) */
	182	stq $23,16($30)
	183	subq $31,$25,$28
	184	stq tmp1,24($30)
	185	cmovlt $25,$28,$25 /* abs($25) */
	186	unop
	187	bsr $23,ufunction
	188	ldq $24,0($30)
	189	ldq $25,8($30)
	190	GETSIGN($28)
	191	subq $31,$27,tmp1
	192	SLONGIFY($28)
	193	ldq $23,16($30)
	194	cmovlt $28,tmp1,$27
	195	ldq tmp1,24($30)
	196	addq $30,STACK,$30
	197	ret $31,($23),1
	198	.end sfunction
00fc0e0d	199	EXPORT_SYMBOL(sfunction)