| 1 |
734 |
jeremybenn |
/* TILE atomics.
|
| 2 |
|
|
Copyright (C) 2011, 2012
|
| 3 |
|
|
Free Software Foundation, Inc.
|
| 4 |
|
|
Contributed by Walter Lee (walt@tilera.com)
|
| 5 |
|
|
|
| 6 |
|
|
This file is free software; you can redistribute it and/or modify it
|
| 7 |
|
|
under the terms of the GNU General Public License as published by the
|
| 8 |
|
|
Free Software Foundation; either version 3, or (at your option) any
|
| 9 |
|
|
later version.
|
| 10 |
|
|
|
| 11 |
|
|
This file is distributed in the hope that it will be useful, but
|
| 12 |
|
|
WITHOUT ANY WARRANTY; without even the implied warranty of
|
| 13 |
|
|
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
| 14 |
|
|
General Public License for more details.
|
| 15 |
|
|
|
| 16 |
|
|
Under Section 7 of GPL version 3, you are granted additional
|
| 17 |
|
|
permissions described in the GCC Runtime Library Exception, version
|
| 18 |
|
|
3.1, as published by the Free Software Foundation.
|
| 19 |
|
|
|
| 20 |
|
|
You should have received a copy of the GNU General Public License and
|
| 21 |
|
|
a copy of the GCC Runtime Library Exception along with this program;
|
| 22 |
|
|
see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
|
| 23 |
|
|
<http://www.gnu.org/licenses/>. */
|
| 24 |
|
|
|
| 25 |
|
|
#include "system.h"
|
| 26 |
|
|
#include "coretypes.h"
|
| 27 |
|
|
#include "atomic.h"
|
| 28 |
|
|
|
| 29 |
|
|
/* This code should be inlined by the compiler, but for now support
|
| 30 |
|
|
it as out-of-line methods in libgcc. */
|
| 31 |
|
|
|
| 32 |
|
|
static void
|
| 33 |
|
|
pre_atomic_barrier (int model)
|
| 34 |
|
|
{
|
| 35 |
|
|
switch ((enum memmodel) model)
|
| 36 |
|
|
{
|
| 37 |
|
|
case MEMMODEL_RELEASE:
|
| 38 |
|
|
case MEMMODEL_ACQ_REL:
|
| 39 |
|
|
case MEMMODEL_SEQ_CST:
|
| 40 |
|
|
__atomic_thread_fence (model);
|
| 41 |
|
|
break;
|
| 42 |
|
|
default:
|
| 43 |
|
|
break;
|
| 44 |
|
|
}
|
| 45 |
|
|
return;
|
| 46 |
|
|
}
|
| 47 |
|
|
|
| 48 |
|
|
static void
|
| 49 |
|
|
post_atomic_barrier (int model)
|
| 50 |
|
|
{
|
| 51 |
|
|
switch ((enum memmodel) model)
|
| 52 |
|
|
{
|
| 53 |
|
|
case MEMMODEL_ACQUIRE:
|
| 54 |
|
|
case MEMMODEL_ACQ_REL:
|
| 55 |
|
|
case MEMMODEL_SEQ_CST:
|
| 56 |
|
|
__atomic_thread_fence (model);
|
| 57 |
|
|
break;
|
| 58 |
|
|
default:
|
| 59 |
|
|
break;
|
| 60 |
|
|
}
|
| 61 |
|
|
return;
|
| 62 |
|
|
}
|
| 63 |
|
|
|
| 64 |
|
|
#define __unused __attribute__((unused))
|
| 65 |
|
|
|
| 66 |
|
|
/* Provide additional methods not implemented by atomic.h. */
|
| 67 |
|
|
#define atomic_xor(mem, mask) \
|
| 68 |
|
|
__atomic_update_cmpxchg(mem, mask, __old ^ __value)
|
| 69 |
|
|
#define atomic_nand(mem, mask) \
|
| 70 |
|
|
__atomic_update_cmpxchg(mem, mask, ~(__old & __value))
|
| 71 |
|
|
|
| 72 |
|
|
#define __atomic_fetch_and_do(type, size, opname) \
|
| 73 |
|
|
type \
|
| 74 |
|
|
__atomic_fetch_##opname##_##size(type* p, type i, int model) \
|
| 75 |
|
|
{ \
|
| 76 |
|
|
pre_atomic_barrier(model); \
|
| 77 |
|
|
type rv = atomic_##opname(p, i); \
|
| 78 |
|
|
post_atomic_barrier(model); \
|
| 79 |
|
|
return rv; \
|
| 80 |
|
|
}
|
| 81 |
|
|
|
| 82 |
|
|
__atomic_fetch_and_do (int, 4, add)
|
| 83 |
|
|
__atomic_fetch_and_do (int, 4, sub)
|
| 84 |
|
|
__atomic_fetch_and_do (int, 4, or)
|
| 85 |
|
|
__atomic_fetch_and_do (int, 4, and)
|
| 86 |
|
|
__atomic_fetch_and_do (int, 4, xor)
|
| 87 |
|
|
__atomic_fetch_and_do (int, 4, nand)
|
| 88 |
|
|
__atomic_fetch_and_do (long long, 8, add)
|
| 89 |
|
|
__atomic_fetch_and_do (long long, 8, sub)
|
| 90 |
|
|
__atomic_fetch_and_do (long long, 8, or)
|
| 91 |
|
|
__atomic_fetch_and_do (long long, 8, and)
|
| 92 |
|
|
__atomic_fetch_and_do (long long, 8, xor)
|
| 93 |
|
|
__atomic_fetch_and_do (long long, 8, nand)
|
| 94 |
|
|
#define __atomic_do_and_fetch(type, size, opname, op) \
|
| 95 |
|
|
type \
|
| 96 |
|
|
__atomic_##opname##_fetch_##size(type* p, type i, int model) \
|
| 97 |
|
|
{ \
|
| 98 |
|
|
pre_atomic_barrier(model); \
|
| 99 |
|
|
type rv = atomic_##opname(p, i) op i; \
|
| 100 |
|
|
post_atomic_barrier(model); \
|
| 101 |
|
|
return rv; \
|
| 102 |
|
|
}
|
| 103 |
|
|
__atomic_do_and_fetch (int, 4, add, +)
|
| 104 |
|
|
__atomic_do_and_fetch (int, 4, sub, -)
|
| 105 |
|
|
__atomic_do_and_fetch (int, 4, or, |)
|
| 106 |
|
|
__atomic_do_and_fetch (int, 4, and, &)
|
| 107 |
|
|
__atomic_do_and_fetch (int, 4, xor, |)
|
| 108 |
|
|
__atomic_do_and_fetch (int, 4, nand, &)
|
| 109 |
|
|
__atomic_do_and_fetch (long long, 8, add, +)
|
| 110 |
|
|
__atomic_do_and_fetch (long long, 8, sub, -)
|
| 111 |
|
|
__atomic_do_and_fetch (long long, 8, or, |)
|
| 112 |
|
|
__atomic_do_and_fetch (long long, 8, and, &)
|
| 113 |
|
|
__atomic_do_and_fetch (long long, 8, xor, |)
|
| 114 |
|
|
__atomic_do_and_fetch (long long, 8, nand, &)
|
| 115 |
|
|
#define __atomic_exchange_methods(type, size) \
|
| 116 |
|
|
bool \
|
| 117 |
|
|
__atomic_compare_exchange_##size(volatile type* ptr, type* oldvalp, \
|
| 118 |
|
|
type newval, bool weak __unused, \
|
| 119 |
|
|
int models, int modelf __unused) \
|
| 120 |
|
|
{ \
|
| 121 |
|
|
type oldval = *oldvalp; \
|
| 122 |
|
|
pre_atomic_barrier(models); \
|
| 123 |
|
|
type retval = atomic_val_compare_and_exchange(ptr, oldval, newval); \
|
| 124 |
|
|
post_atomic_barrier(models); \
|
| 125 |
|
|
bool success = (retval == oldval); \
|
| 126 |
|
|
*oldvalp = retval; \
|
| 127 |
|
|
return success; \
|
| 128 |
|
|
} \
|
| 129 |
|
|
\
|
| 130 |
|
|
type \
|
| 131 |
|
|
__atomic_exchange_##size(volatile type* ptr, type val, int model) \
|
| 132 |
|
|
{ \
|
| 133 |
|
|
pre_atomic_barrier(model); \
|
| 134 |
|
|
type retval = atomic_exchange(ptr, val); \
|
| 135 |
|
|
post_atomic_barrier(model); \
|
| 136 |
|
|
return retval; \
|
| 137 |
|
|
}
|
| 138 |
|
|
__atomic_exchange_methods (int, 4)
|
| 139 |
|
|
__atomic_exchange_methods (long long, 8)
|
| 140 |
|
|
|
| 141 |
|
|
/* Subword methods require the same approach for both TILEPro and
|
| 142 |
|
|
TILE-Gx. We load the background data for the word, insert the
|
| 143 |
|
|
desired subword piece, then compare-and-exchange it into place. */
|
| 144 |
|
|
#define u8 unsigned char
|
| 145 |
|
|
#define u16 unsigned short
|
| 146 |
|
|
#define __atomic_subword_cmpxchg(type, size) \
|
| 147 |
|
|
\
|
| 148 |
|
|
bool \
|
| 149 |
|
|
__atomic_compare_exchange_##size(volatile type* ptr, type* guess, \
|
| 150 |
|
|
type val, bool weak __unused, int models, \
|
| 151 |
|
|
int modelf __unused) \
|
| 152 |
|
|
{ \
|
| 153 |
|
|
pre_atomic_barrier(models); \
|
| 154 |
|
|
unsigned int *p = (unsigned int *)((unsigned long)ptr & ~3UL); \
|
| 155 |
|
|
const int shift = ((unsigned long)ptr & 3UL) * 8; \
|
| 156 |
|
|
const unsigned int valmask = (1 << (sizeof(type) * 8)) - 1; \
|
| 157 |
|
|
const unsigned int bgmask = ~(valmask << shift); \
|
| 158 |
|
|
unsigned int oldword = *p; \
|
| 159 |
|
|
type oldval = (oldword >> shift) & valmask; \
|
| 160 |
|
|
if (__builtin_expect((oldval == *guess), 1)) { \
|
| 161 |
|
|
unsigned int word = (oldword & bgmask) | ((val & valmask) << shift); \
|
| 162 |
|
|
oldword = atomic_val_compare_and_exchange(p, oldword, word); \
|
| 163 |
|
|
oldval = (oldword >> shift) & valmask; \
|
| 164 |
|
|
} \
|
| 165 |
|
|
post_atomic_barrier(models); \
|
| 166 |
|
|
bool success = (oldval == *guess); \
|
| 167 |
|
|
*guess = oldval; \
|
| 168 |
|
|
return success; \
|
| 169 |
|
|
}
|
| 170 |
|
|
__atomic_subword_cmpxchg (u8, 1)
|
| 171 |
|
|
__atomic_subword_cmpxchg (u16, 2)
|
| 172 |
|
|
/* For the atomic-update subword methods, we use the same approach as
|
| 173 |
|
|
above, but we retry until we succeed if the compare-and-exchange
|
| 174 |
|
|
fails. */
|
| 175 |
|
|
#define __atomic_subword(type, proto, top, expr, bottom) \
|
| 176 |
|
|
proto \
|
| 177 |
|
|
{ \
|
| 178 |
|
|
top \
|
| 179 |
|
|
unsigned int *p = (unsigned int *)((unsigned long)ptr & ~3UL); \
|
| 180 |
|
|
const int shift = ((unsigned long)ptr & 3UL) * 8; \
|
| 181 |
|
|
const unsigned int valmask = (1 << (sizeof(type) * 8)) - 1; \
|
| 182 |
|
|
const unsigned int bgmask = ~(valmask << shift); \
|
| 183 |
|
|
unsigned int oldword, xword = *p; \
|
| 184 |
|
|
type val, oldval; \
|
| 185 |
|
|
do { \
|
| 186 |
|
|
oldword = xword; \
|
| 187 |
|
|
oldval = (oldword >> shift) & valmask; \
|
| 188 |
|
|
val = expr; \
|
| 189 |
|
|
unsigned int word = (oldword & bgmask) | ((val & valmask) << shift); \
|
| 190 |
|
|
xword = atomic_val_compare_and_exchange(p, oldword, word); \
|
| 191 |
|
|
} while (__builtin_expect(xword != oldword, 0)); \
|
| 192 |
|
|
bottom \
|
| 193 |
|
|
}
|
| 194 |
|
|
#define __atomic_subword_fetch(type, funcname, expr, retval) \
|
| 195 |
|
|
__atomic_subword(type, \
|
| 196 |
|
|
type __atomic_ ## funcname(volatile type *ptr, type i, int model), \
|
| 197 |
|
|
pre_atomic_barrier(model);, \
|
| 198 |
|
|
expr, \
|
| 199 |
|
|
post_atomic_barrier(model); return retval;)
|
| 200 |
|
|
__atomic_subword_fetch (u8, fetch_add_1, oldval + i, oldval)
|
| 201 |
|
|
__atomic_subword_fetch (u8, fetch_sub_1, oldval - i, oldval)
|
| 202 |
|
|
__atomic_subword_fetch (u8, fetch_or_1, oldval | i, oldval)
|
| 203 |
|
|
__atomic_subword_fetch (u8, fetch_and_1, oldval & i, oldval)
|
| 204 |
|
|
__atomic_subword_fetch (u8, fetch_xor_1, oldval ^ i, oldval)
|
| 205 |
|
|
__atomic_subword_fetch (u8, fetch_nand_1, ~(oldval & i), oldval)
|
| 206 |
|
|
__atomic_subword_fetch (u16, fetch_add_2, oldval + i, oldval)
|
| 207 |
|
|
__atomic_subword_fetch (u16, fetch_sub_2, oldval - i, oldval)
|
| 208 |
|
|
__atomic_subword_fetch (u16, fetch_or_2, oldval | i, oldval)
|
| 209 |
|
|
__atomic_subword_fetch (u16, fetch_and_2, oldval & i, oldval)
|
| 210 |
|
|
__atomic_subword_fetch (u16, fetch_xor_2, oldval ^ i, oldval)
|
| 211 |
|
|
__atomic_subword_fetch (u16, fetch_nand_2, ~(oldval & i), oldval)
|
| 212 |
|
|
__atomic_subword_fetch (u8, add_fetch_1, oldval + i, val)
|
| 213 |
|
|
__atomic_subword_fetch (u8, sub_fetch_1, oldval - i, val)
|
| 214 |
|
|
__atomic_subword_fetch (u8, or_fetch_1, oldval | i, val)
|
| 215 |
|
|
__atomic_subword_fetch (u8, and_fetch_1, oldval & i, val)
|
| 216 |
|
|
__atomic_subword_fetch (u8, xor_fetch_1, oldval ^ i, val)
|
| 217 |
|
|
__atomic_subword_fetch (u8, nand_fetch_1, ~(oldval & i), val)
|
| 218 |
|
|
__atomic_subword_fetch (u16, add_fetch_2, oldval + i, val)
|
| 219 |
|
|
__atomic_subword_fetch (u16, sub_fetch_2, oldval - i, val)
|
| 220 |
|
|
__atomic_subword_fetch (u16, or_fetch_2, oldval | i, val)
|
| 221 |
|
|
__atomic_subword_fetch (u16, and_fetch_2, oldval & i, val)
|
| 222 |
|
|
__atomic_subword_fetch (u16, xor_fetch_2, oldval ^ i, val)
|
| 223 |
|
|
__atomic_subword_fetch (u16, nand_fetch_2, ~(oldval & i), val)
|
| 224 |
|
|
#define __atomic_subword_lock(type, size) \
|
| 225 |
|
|
\
|
| 226 |
|
|
__atomic_subword(type, \
|
| 227 |
|
|
type __atomic_exchange_##size(volatile type* ptr, type nval, int model), \
|
| 228 |
|
|
pre_atomic_barrier(model);, \
|
| 229 |
|
|
nval, \
|
| 230 |
|
|
post_atomic_barrier(model); return oldval;)
|
| 231 |
|
|
__atomic_subword_lock (u8, 1)
|
| 232 |
|
|
__atomic_subword_lock (u16, 2)
|
