Loading...
1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef __ASM_ARM_DIV64
3#define __ASM_ARM_DIV64
4
5#include <linux/types.h>
6#include <asm/compiler.h>
7
8/*
9 * The semantics of __div64_32() are:
10 *
11 * uint32_t __div64_32(uint64_t *n, uint32_t base)
12 * {
13 * uint32_t remainder = *n % base;
14 * *n = *n / base;
15 * return remainder;
16 * }
17 *
18 * In other words, a 64-bit dividend with a 32-bit divisor producing
19 * a 64-bit result and a 32-bit remainder. To accomplish this optimally
20 * we override the generic version in lib/div64.c to call our __do_div64
21 * assembly implementation with completely non standard calling convention
22 * for arguments and results (beware).
23 */
24
25#ifdef __ARMEB__
26#define __xh "r0"
27#define __xl "r1"
28#else
29#define __xl "r0"
30#define __xh "r1"
31#endif
32
33static inline uint32_t __div64_32(uint64_t *n, uint32_t base)
34{
35 register unsigned int __base asm("r4") = base;
36 register unsigned long long __n asm("r0") = *n;
37 register unsigned long long __res asm("r2");
38 register unsigned int __rem asm(__xh);
39 asm( __asmeq("%0", __xh)
40 __asmeq("%1", "r2")
41 __asmeq("%2", "r0")
42 __asmeq("%3", "r4")
43 "bl __do_div64"
44 : "=r" (__rem), "=r" (__res)
45 : "r" (__n), "r" (__base)
46 : "ip", "lr", "cc");
47 *n = __res;
48 return __rem;
49}
50#define __div64_32 __div64_32
51
52#if !defined(CONFIG_AEABI)
53
54/*
55 * In OABI configurations, some uses of the do_div function
56 * cause gcc to run out of registers. To work around that,
57 * we can force the use of the out-of-line version for
58 * configurations that build a OABI kernel.
59 */
60#define do_div(n, base) __div64_32(&(n), base)
61
62#else
63
64/*
65 * gcc versions earlier than 4.0 are simply too problematic for the
66 * __div64_const32() code in asm-generic/div64.h. First there is
67 * gcc PR 15089 that tend to trig on more complex constructs, spurious
68 * .global __udivsi3 are inserted even if none of those symbols are
69 * referenced in the generated code, and those gcc versions are not able
70 * to do constant propagation on long long values anyway.
71 */
72
73#define __div64_const32_is_OK (__GNUC__ >= 4)
74
75static inline uint64_t __arch_xprod_64(uint64_t m, uint64_t n, bool bias)
76{
77 unsigned long long res;
78 register unsigned int tmp asm("ip") = 0;
79
80 if (!bias) {
81 asm ( "umull %Q0, %R0, %Q1, %Q2\n\t"
82 "mov %Q0, #0"
83 : "=&r" (res)
84 : "r" (m), "r" (n)
85 : "cc");
86 } else if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
87 res = m;
88 asm ( "umlal %Q0, %R0, %Q1, %Q2\n\t"
89 "mov %Q0, #0"
90 : "+&r" (res)
91 : "r" (m), "r" (n)
92 : "cc");
93 } else {
94 asm ( "umull %Q0, %R0, %Q2, %Q3\n\t"
95 "cmn %Q0, %Q2\n\t"
96 "adcs %R0, %R0, %R2\n\t"
97 "adc %Q0, %1, #0"
98 : "=&r" (res), "+&r" (tmp)
99 : "r" (m), "r" (n)
100 : "cc");
101 }
102
103 if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
104 asm ( "umlal %R0, %Q0, %R1, %Q2\n\t"
105 "umlal %R0, %Q0, %Q1, %R2\n\t"
106 "mov %R0, #0\n\t"
107 "umlal %Q0, %R0, %R1, %R2"
108 : "+&r" (res)
109 : "r" (m), "r" (n)
110 : "cc");
111 } else {
112 asm ( "umlal %R0, %Q0, %R2, %Q3\n\t"
113 "umlal %R0, %1, %Q2, %R3\n\t"
114 "mov %R0, #0\n\t"
115 "adds %Q0, %1, %Q0\n\t"
116 "adc %R0, %R0, #0\n\t"
117 "umlal %Q0, %R0, %R2, %R3"
118 : "+&r" (res), "+&r" (tmp)
119 : "r" (m), "r" (n)
120 : "cc");
121 }
122
123 return res;
124}
125#define __arch_xprod_64 __arch_xprod_64
126
127#include <asm-generic/div64.h>
128
129#endif
130
131#endif
1/* SPDX-License-Identifier: GPL-2.0 */
2#ifndef __ASM_ARM_DIV64
3#define __ASM_ARM_DIV64
4
5#include <linux/types.h>
6#include <asm/compiler.h>
7
8/*
9 * The semantics of __div64_32() are:
10 *
11 * uint32_t __div64_32(uint64_t *n, uint32_t base)
12 * {
13 * uint32_t remainder = *n % base;
14 * *n = *n / base;
15 * return remainder;
16 * }
17 *
18 * In other words, a 64-bit dividend with a 32-bit divisor producing
19 * a 64-bit result and a 32-bit remainder. To accomplish this optimally
20 * we override the generic version in lib/div64.c to call our __do_div64
21 * assembly implementation with completely non standard calling convention
22 * for arguments and results (beware).
23 */
24static inline uint32_t __div64_32(uint64_t *n, uint32_t base)
25{
26 register unsigned int __base asm("r4") = base;
27 register unsigned long long __n asm("r0") = *n;
28 register unsigned long long __res asm("r2");
29 unsigned int __rem;
30 asm( __asmeq("%0", "r0")
31 __asmeq("%1", "r2")
32 __asmeq("%2", "r4")
33 "bl __do_div64"
34 : "+r" (__n), "=r" (__res)
35 : "r" (__base)
36 : "ip", "lr", "cc");
37 __rem = __n >> 32;
38 *n = __res;
39 return __rem;
40}
41#define __div64_32 __div64_32
42
43#if !defined(CONFIG_AEABI)
44
45/*
46 * In OABI configurations, some uses of the do_div function
47 * cause gcc to run out of registers. To work around that,
48 * we can force the use of the out-of-line version for
49 * configurations that build a OABI kernel.
50 */
51#define do_div(n, base) __div64_32(&(n), base)
52
53#else
54
55/*
56 * gcc versions earlier than 4.0 are simply too problematic for the
57 * __div64_const32() code in asm-generic/div64.h. First there is
58 * gcc PR 15089 that tend to trig on more complex constructs, spurious
59 * .global __udivsi3 are inserted even if none of those symbols are
60 * referenced in the generated code, and those gcc versions are not able
61 * to do constant propagation on long long values anyway.
62 */
63
64#define __div64_const32_is_OK (__GNUC__ >= 4)
65
66static inline uint64_t __arch_xprod_64(uint64_t m, uint64_t n, bool bias)
67{
68 unsigned long long res;
69 register unsigned int tmp asm("ip") = 0;
70
71 if (!bias) {
72 asm ( "umull %Q0, %R0, %Q1, %Q2\n\t"
73 "mov %Q0, #0"
74 : "=&r" (res)
75 : "r" (m), "r" (n)
76 : "cc");
77 } else if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
78 res = m;
79 asm ( "umlal %Q0, %R0, %Q1, %Q2\n\t"
80 "mov %Q0, #0"
81 : "+&r" (res)
82 : "r" (m), "r" (n)
83 : "cc");
84 } else {
85 asm ( "umull %Q0, %R0, %Q2, %Q3\n\t"
86 "cmn %Q0, %Q2\n\t"
87 "adcs %R0, %R0, %R2\n\t"
88 "adc %Q0, %1, #0"
89 : "=&r" (res), "+&r" (tmp)
90 : "r" (m), "r" (n)
91 : "cc");
92 }
93
94 if (!(m & ((1ULL << 63) | (1ULL << 31)))) {
95 asm ( "umlal %R0, %Q0, %R1, %Q2\n\t"
96 "umlal %R0, %Q0, %Q1, %R2\n\t"
97 "mov %R0, #0\n\t"
98 "umlal %Q0, %R0, %R1, %R2"
99 : "+&r" (res)
100 : "r" (m), "r" (n)
101 : "cc");
102 } else {
103 asm ( "umlal %R0, %Q0, %R2, %Q3\n\t"
104 "umlal %R0, %1, %Q2, %R3\n\t"
105 "mov %R0, #0\n\t"
106 "adds %Q0, %1, %Q0\n\t"
107 "adc %R0, %R0, #0\n\t"
108 "umlal %Q0, %R0, %R2, %R3"
109 : "+&r" (res), "+&r" (tmp)
110 : "r" (m), "r" (n)
111 : "cc");
112 }
113
114 return res;
115}
116#define __arch_xprod_64 __arch_xprod_64
117
118#include <asm-generic/div64.h>
119
120#endif
121
122#endif