udivmodsi4.S revision 5d71de26cedae3dafc17449fe0182045c0bd20e8 
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	.p2align 2
21DEFINE_COMPILERRT_FUNCTION(__udivmodsi4)
22#if __ARM_ARCH_EXT_IDIV__
23	tst     r1, r1
24	beq     LOCAL_LABEL(divby0)
25	mov 	r3, r0
26	udiv	r0, r3, r1
27	mls 	r1, r0, r1, r3
28	str 	r1, [r2]
29	bx  	lr
30#else
31	cmp	r1, #1
32	bcc	LOCAL_LABEL(divby0)
33	beq	LOCAL_LABEL(divby1)
34	cmp	r0, r1
35	bcc	LOCAL_LABEL(quotient0)
36	/*
37	 * Implement division using binary long division algorithm.
38	 *
39	 * r0 is the numerator, r1 the denominator.
40	 *
41	 * The code before JMP computes the correct shift I, so that
42	 * r0 and (r1 << I) have the highest bit set in the same position.
43	 * At the time of JMP, ip := .Ldiv0block - 12 * I.
44	 * This depends on the fixed instruction size of block.
45	 *
46	 * block(shift) implements the test-and-update-quotient core.
47	 * It assumes (r0 << shift) can be computed without overflow and
48	 * that (r0 << shift) < 2 * r1. The quotient is stored in r3.
49	 */
50
51#  ifdef __ARM_FEATURE_CLZ
52	clz	ip, r0
53	clz	r3, r1
54	/* r0 >= r1 implies clz(r0) <= clz(r1), so ip <= r3. */
55	sub	r3, r3, ip
56	adr	ip, LOCAL_LABEL(div0block)
57	sub	ip, ip, r3, lsl #2
58	sub	ip, ip, r3, lsl #3
59	mov	r3, #0
60	bx	ip
61#  else
62	str	r4, [sp, #-8]!
63
64	mov	r4, r0
65	adr	ip, LOCAL_LABEL(div0block)
66
67	lsr	r3, r4, #16
68	cmp	r3, r1
69	movhs	r4, r3
70	subhs	ip, ip, #(16 * 12)
71
72	lsr	r3, r4, #8
73	cmp	r3, r1
74	movhs	r4, r3
75	subhs	ip, ip, #(8 * 12)
76
77	lsr	r3, r4, #4
78	cmp	r3, r1
79	movhs	r4, r3
80	subhs	ip, #(4 * 12)
81
82	lsr	r3, r4, #2
83	cmp	r3, r1
84	movhs	r4, r3
85	subhs	ip, ip, #(2 * 12)
86
87	/* Last block, no need to update r3 or r4. */
88	cmp	r1, r4, lsr #1
89	subls	ip, ip, #(1 * 12)
90
91	ldr	r4, [sp], #8	/* restore r4, we are done with it. */
92	mov	r3, #0
93
94	JMP(ip)
95#  endif
96
97#define	IMM	#
98
99#define block(shift) \
100	cmp	r0, r1, lsl IMM shift; \
101	addhs	r3, r3, IMM (1 << shift); \
102	subhs	r0, r0, r1, lsl IMM shift
103
104	block(31)
105	block(30)
106	block(29)
107	block(28)
108	block(27)
109	block(26)
110	block(25)
111	block(24)
112	block(23)
113	block(22)
114	block(21)
115	block(20)
116	block(19)
117	block(18)
118	block(17)
119	block(16)
120	block(15)
121	block(14)
122	block(13)
123	block(12)
124	block(11)
125	block(10)
126	block(9)
127	block(8)
128	block(7)
129	block(6)
130	block(5)
131	block(4)
132	block(3)
133	block(2)
134	block(1)
135LOCAL_LABEL(div0block):
136	block(0)
137
138	str	r0, [r2]
139	mov	r0, r3
140	JMP(lr)
141
142LOCAL_LABEL(quotient0):
143	str	r0, [r2]
144	mov	r0, #0
145	JMP(lr)
146
147LOCAL_LABEL(divby1):
148	mov	r3, #0
149	str	r3, [r2]
150	JMP(lr)
151#endif /* __ARM_ARCH_EXT_IDIV__ */
152
153LOCAL_LABEL(divby0):
154	mov	r0, #0
155#ifdef __ARM_EABI__
156	b	__aeabi_idiv0
157#else
158	JMP(lr)
159#endif
160
161END_COMPILERRT_FUNCTION(__udivmodsi4)
162