[deliverable/linux.git] / arch / alpha / lib / divide.S

/*
 * 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
 */

#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

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