| 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 |
static 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 |
#ifdef CONFIG_CC_OPTIMIZE_FOR_PERFORMANCE |
| 56 |
static __always_inline |
| 57 |
#else |
| 58 |
static inline |
| 59 |
#endif |
| 60 |
uint64_t __arch_xprod_64(uint64_t m, uint64_t n, bool bias) |
| 61 |
{ |
| 62 |
unsigned long long res; |
| 63 |
register unsigned int tmp asm("ip") = 0; |
| 64 |
bool no_ovf = __builtin_constant_p(m) && |
| 65 |
((m >> 32) + (m & 0xffffffff) < 0x100000000); |
| 66 |
|
| 67 |
if (!bias) { |
| 68 |
asm ( "umull %Q0, %R0, %Q1, %Q2\n\t" |
| 69 |
"mov %Q0, #0" |
| 70 |
: "=&r" (res) |
| 71 |
: "r" (m), "r" (n) |
| 72 |
: "cc"); |
| 73 |
} else if (no_ovf) { |
| 74 |
res = m; |
| 75 |
asm ( "umlal %Q0, %R0, %Q1, %Q2\n\t" |
| 76 |
"mov %Q0, #0" |
| 77 |
: "+&r" (res) |
| 78 |
: "r" (m), "r" (n) |
| 79 |
: "cc"); |
| 80 |
} else { |
| 81 |
asm ( "umull %Q0, %R0, %Q2, %Q3\n\t" |
| 82 |
"cmn %Q0, %Q2\n\t" |
| 83 |
"adcs %R0, %R0, %R2\n\t" |
| 84 |
"adc %Q0, %1, #0" |
| 85 |
: "=&r" (res), "+&r" (tmp) |
| 86 |
: "r" (m), "r" (n) |
| 87 |
: "cc"); |
| 88 |
} |
| 89 |
|
| 90 |
if (no_ovf) { |
| 91 |
asm ( "umlal %R0, %Q0, %R1, %Q2\n\t" |
| 92 |
"umlal %R0, %Q0, %Q1, %R2\n\t" |
| 93 |
"mov %R0, #0\n\t" |
| 94 |
"umlal %Q0, %R0, %R1, %R2" |
| 95 |
: "+&r" (res) |
| 96 |
: "r" (m), "r" (n) |
| 97 |
: "cc"); |
| 98 |
} else { |
| 99 |
asm ( "umlal %R0, %Q0, %R2, %Q3\n\t" |
| 100 |
"umlal %R0, %1, %Q2, %R3\n\t" |
| 101 |
"mov %R0, #0\n\t" |
| 102 |
"adds %Q0, %1, %Q0\n\t" |
| 103 |
"adc %R0, %R0, #0\n\t" |
| 104 |
"umlal %Q0, %R0, %R2, %R3" |
| 105 |
: "+&r" (res), "+&r" (tmp) |
| 106 |
: "r" (m), "r" (n) |
| 107 |
: "cc"); |
| 108 |
} |
| 109 |
|
| 110 |
return res; |
| 111 |
} |
| 112 |
#define __arch_xprod_64 __arch_xprod_64 |
| 113 |
|
| 114 |
#include <asm-generic/div64.h> |
| 115 |
|
| 116 |
#endif |
| 117 |
|
| 118 |
#endif |