1/*===-- udivmodsi4.S - 32-bit unsigned integer divide and modulus ---------===//
2 *
3 *                     The LLVM Compiler Infrastructure
4 *
5 * This file is dual licensed under the MIT and the University of Illinois Open
6 * Source Licenses. See LICENSE.TXT for details.
7 *
8 *===----------------------------------------------------------------------===//
9 *
10 * This file implements the __udivmodsi4 (32-bit unsigned integer divide and
11 * modulus) function for the ARM 32-bit architecture.
12 *
13 *===----------------------------------------------------------------------===*/
14
15#include "../assembly.h"
16
17	.syntax unified
18	.text
19
20@ unsigned int __udivmodsi4(unsigned int divident, unsigned int divisor,
21@                           unsigned int *remainder)
22@   Calculate the quotient and remainder of the (unsigned) division.  The return
23@   value is the quotient, the remainder is placed in the variable.
24
25	.p2align 2
26DEFINE_COMPILERRT_FUNCTION(__udivmodsi4)
27#if __ARM_ARCH_EXT_IDIV__
28	tst     r1, r1
29	beq     LOCAL_LABEL(divby0)
30	mov 	r3, r0
31	udiv	r0, r3, r1
32	mls 	r1, r0, r1, r3
33	str 	r1, [r2]
34	bx  	lr
35#else
36	cmp	r1, #1
37	bcc	LOCAL_LABEL(divby0)
38	beq	LOCAL_LABEL(divby1)
39	cmp	r0, r1
40	bcc	LOCAL_LABEL(quotient0)
41	/*
42	 * Implement division using binary long division algorithm.
43	 *
44	 * r0 is the numerator, r1 the denominator.
45	 *
46	 * The code before JMP computes the correct shift I, so that
47	 * r0 and (r1 << I) have the highest bit set in the same position.
48	 * At the time of JMP, ip := .Ldiv0block - 12 * I.
49	 * This depends on the fixed instruction size of block.
50	 * For ARM mode, this is 12 Bytes, for THUMB mode 14 Bytes.
51	 *
52	 * block(shift) implements the test-and-update-quotient core.
53	 * It assumes (r0 << shift) can be computed without overflow and
54	 * that (r0 << shift) < 2 * r1. The quotient is stored in r3.
55	 */
56
57#  ifdef __ARM_FEATURE_CLZ
58	clz	ip, r0
59	clz	r3, r1
60	/* r0 >= r1 implies clz(r0) <= clz(r1), so ip <= r3. */
61	sub	r3, r3, ip
62	adr	ip, LOCAL_LABEL(div0block)
63	sub	ip, ip, r3, lsl #2
64	sub	ip, ip, r3, lsl #3
65	mov	r3, #0
66	bx	ip
67#  else
68	str	r4, [sp, #-8]!
69
70	mov	r4, r0
71	adr	ip, LOCAL_LABEL(div0block)
72
73	lsr	r3, r4, #16
74	cmp	r3, r1
75	movhs	r4, r3
76	subhs	ip, ip, #(16 * 12)
77
78	lsr	r3, r4, #8
79	cmp	r3, r1
80	movhs	r4, r3
81	subhs	ip, ip, #(8 * 12)
82
83	lsr	r3, r4, #4
84	cmp	r3, r1
85	movhs	r4, r3
86	subhs	ip, #(4 * 12)
87
88	lsr	r3, r4, #2
89	cmp	r3, r1
90	movhs	r4, r3
91	subhs	ip, ip, #(2 * 12)
92
93	/* Last block, no need to update r3 or r4. */
94	cmp	r1, r4, lsr #1
95	subls	ip, ip, #(1 * 12)
96
97	ldr	r4, [sp], #8	/* restore r4, we are done with it. */
98	mov	r3, #0
99
100	JMP(ip)
101#  endif
102
103#define	IMM	#
104
105#define block(shift)                                                           \
106	cmp	r0, r1, lsl IMM shift;                                         \
107	ITT(hs);                                                               \
108	WIDE(addhs)	r3, r3, IMM (1 << shift);                              \
109	WIDE(subhs)	r0, r0, r1, lsl IMM shift
110
111	block(31)
112	block(30)
113	block(29)
114	block(28)
115	block(27)
116	block(26)
117	block(25)
118	block(24)
119	block(23)
120	block(22)
121	block(21)
122	block(20)
123	block(19)
124	block(18)
125	block(17)
126	block(16)
127	block(15)
128	block(14)
129	block(13)
130	block(12)
131	block(11)
132	block(10)
133	block(9)
134	block(8)
135	block(7)
136	block(6)
137	block(5)
138	block(4)
139	block(3)
140	block(2)
141	block(1)
142LOCAL_LABEL(div0block):
143	block(0)
144
145	str	r0, [r2]
146	mov	r0, r3
147	JMP(lr)
148
149LOCAL_LABEL(quotient0):
150	str	r0, [r2]
151	mov	r0, #0
152	JMP(lr)
153
154LOCAL_LABEL(divby1):
155	mov	r3, #0
156	str	r3, [r2]
157	JMP(lr)
158#endif /* __ARM_ARCH_EXT_IDIV__ */
159
160LOCAL_LABEL(divby0):
161	mov	r0, #0
162#ifdef __ARM_EABI__
163	b	__aeabi_idiv0
164#else
165	JMP(lr)
166#endif
167
168END_COMPILERRT_FUNCTION(__udivmodsi4)
169