1//===-- umodsi3.S - 32-bit unsigned integer modulus -----------------------===//
2//
3// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4// See https://llvm.org/LICENSE.txt for license information.
5// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6//
7//===----------------------------------------------------------------------===//
8//
9// This file implements the __umodsi3 (32-bit unsigned integer modulus)
10// function for the ARM 32-bit architecture.
11//
12//===----------------------------------------------------------------------===//
13
14#include "../assembly.h"
15
16	.syntax unified
17	.text
18	DEFINE_CODE_STATE
19
20@ unsigned int __umodsi3(unsigned int divident, unsigned int divisor)
21@   Calculate and return the remainder of the (unsigned) division.
22
23	.p2align 2
24DEFINE_COMPILERRT_FUNCTION(__umodsi3)
25#if __ARM_ARCH_EXT_IDIV__
26	tst     r1, r1
27	beq     LOCAL_LABEL(divby0)
28	udiv	r2, r0, r1
29	mls 	r0, r2, r1, r0
30	bx  	lr
31#else
32	cmp	r1, #1
33	bcc	LOCAL_LABEL(divby0)
34	ITT(eq)
35	moveq	r0, #0
36	JMPc(lr, eq)
37	cmp	r0, r1
38	IT(cc)
39	JMPc(lr, cc)
40
41	// Implement division using binary long division algorithm.
42	//
43	// r0 is the numerator, r1 the denominator.
44	//
45	// The code before JMP computes the correct shift I, so that
46	// r0 and (r1 << I) have the highest bit set in the same position.
47	// At the time of JMP, ip := .Ldiv0block - 8 * I.
48	// This depends on the fixed instruction size of block.
49	// For ARM mode, this is 8 Bytes, for THUMB mode 10 Bytes.
50	//
51	// block(shift) implements the test-and-update-quotient core.
52	// It assumes (r0 << shift) can be computed without overflow and
53	// that (r0 << shift) < 2 * r1. The quotient is stored in r3.
54
55#  ifdef __ARM_FEATURE_CLZ
56	clz	ip, r0
57	clz	r3, r1
58	// r0 >= r1 implies clz(r0) <= clz(r1), so ip <= r3.
59	sub	r3, r3, ip
60#    if defined(USE_THUMB_2)
61	adr	ip, LOCAL_LABEL(div0block) + 1
62	sub	ip, ip, r3, lsl #1
63#    else
64	adr	ip, LOCAL_LABEL(div0block)
65#    endif
66	sub	ip, ip, r3, lsl #3
67	bx	ip
68#  else
69#    if defined(USE_THUMB_2)
70#    error THUMB mode requires CLZ or UDIV
71#    endif
72	mov	r2, r0
73	adr	ip, LOCAL_LABEL(div0block)
74
75	lsr	r3, r2, #16
76	cmp	r3, r1
77	movhs	r2, r3
78	subhs	ip, ip, #(16 * 8)
79
80	lsr	r3, r2, #8
81	cmp	r3, r1
82	movhs	r2, r3
83	subhs	ip, ip, #(8 * 8)
84
85	lsr	r3, r2, #4
86	cmp	r3, r1
87	movhs	r2, r3
88	subhs	ip, #(4 * 8)
89
90	lsr	r3, r2, #2
91	cmp	r3, r1
92	movhs	r2, r3
93	subhs	ip, ip, #(2 * 8)
94
95	// Last block, no need to update r2 or r3.
96	cmp	r1, r2, lsr #1
97	subls	ip, ip, #(1 * 8)
98
99	JMP(ip)
100#  endif
101
102#define	IMM	#
103
104#define block(shift)                                                           \
105	cmp	r0, r1, lsl IMM shift;                                         \
106	IT(hs);                                                                \
107	WIDE(subhs)	r0, r0, r1, lsl IMM shift
108
109	block(31)
110	block(30)
111	block(29)
112	block(28)
113	block(27)
114	block(26)
115	block(25)
116	block(24)
117	block(23)
118	block(22)
119	block(21)
120	block(20)
121	block(19)
122	block(18)
123	block(17)
124	block(16)
125	block(15)
126	block(14)
127	block(13)
128	block(12)
129	block(11)
130	block(10)
131	block(9)
132	block(8)
133	block(7)
134	block(6)
135	block(5)
136	block(4)
137	block(3)
138	block(2)
139	block(1)
140LOCAL_LABEL(div0block):
141	block(0)
142	JMP(lr)
143#endif // __ARM_ARCH_EXT_IDIV__
144
145LOCAL_LABEL(divby0):
146	mov	r0, #0
147#ifdef __ARM_EABI__
148	b	__aeabi_idiv0
149#else
150	JMP(lr)
151#endif
152
153END_COMPILERRT_FUNCTION(__umodsi3)
154
155NO_EXEC_STACK_DIRECTIVE
156
157