summaryrefslogtreecommitdiffstats
path: root/src/include/port/atomics/arch-ppc.h
blob: a82ae38c1d6e8d04ca7e22f757ba9ff4580bef57 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
/*-------------------------------------------------------------------------
 *
 * arch-ppc.h
 *	  Atomic operations considerations specific to PowerPC
 *
 * Portions Copyright (c) 1996-2020, PostgreSQL Global Development Group
 * Portions Copyright (c) 1994, Regents of the University of California
 *
 * NOTES:
 *
 * src/include/port/atomics/arch-ppc.h
 *
 *-------------------------------------------------------------------------
 */

#if defined(__GNUC__)

/*
 * lwsync orders loads with respect to each other, and similarly with stores.
 * But a load can be performed before a subsequent store, so sync must be used
 * for a full memory barrier.
 */
#define pg_memory_barrier_impl()	__asm__ __volatile__ ("sync" : : : "memory")
#define pg_read_barrier_impl()		__asm__ __volatile__ ("lwsync" : : : "memory")
#define pg_write_barrier_impl()		__asm__ __volatile__ ("lwsync" : : : "memory")
#endif

#define PG_HAVE_ATOMIC_U32_SUPPORT
typedef struct pg_atomic_uint32
{
	volatile uint32 value;
} pg_atomic_uint32;

/* 64bit atomics are only supported in 64bit mode */
#if SIZEOF_VOID_P >= 8
#define PG_HAVE_ATOMIC_U64_SUPPORT
typedef struct pg_atomic_uint64
{
	volatile uint64 value pg_attribute_aligned(8);
} pg_atomic_uint64;

#endif

/*
 * This mimics gcc __atomic_compare_exchange_n(..., __ATOMIC_SEQ_CST), but
 * code generation differs at the end.  __atomic_compare_exchange_n():
 *  100:	isync
 *  104:	mfcr    r3
 *  108:	rlwinm  r3,r3,3,31,31
 *  10c:	bne     120 <.eb+0x10>
 *  110:	clrldi  r3,r3,63
 *  114:	addi    r1,r1,112
 *  118:	blr
 *  11c:	nop
 *  120:	clrldi  r3,r3,63
 *  124:	stw     r9,0(r4)
 *  128:	addi    r1,r1,112
 *  12c:	blr
 *
 * This:
 *   f0:	isync
 *   f4:	mfcr    r9
 *   f8:	rldicl. r3,r9,35,63
 *   fc:	bne     104 <.eb>
 *  100:	stw     r10,0(r4)
 *  104:	addi    r1,r1,112
 *  108:	blr
 *
 * This implementation may or may not have materially different performance.
 * It's not exploiting the fact that cr0 still holds the relevant comparison
 * bits, set during the __asm__.  One could fix that by moving more code into
 * the __asm__.  (That would remove the freedom to eliminate dead stores when
 * the caller ignores "expected", but few callers do.)
 *
 * Recognizing constant "newval" would be superfluous, because there's no
 * immediate-operand version of stwcx.
 */
#define PG_HAVE_ATOMIC_COMPARE_EXCHANGE_U32
static inline bool
pg_atomic_compare_exchange_u32_impl(volatile pg_atomic_uint32 *ptr,
									uint32 *expected, uint32 newval)
{
	uint32 found;
	uint32 condition_register;
	bool ret;

#ifdef HAVE_I_CONSTRAINT__BUILTIN_CONSTANT_P
	if (__builtin_constant_p(*expected) &&
		(int32) *expected <= PG_INT16_MAX &&
		(int32) *expected >= PG_INT16_MIN)
		__asm__ __volatile__(
			"	sync				\n"
			"	lwarx   %0,0,%5		\n"
			"	cmpwi   %0,%3		\n"
			"	bne     $+12		\n"		/* branch to isync */
			"	stwcx.  %4,0,%5		\n"
			"	bne     $-16		\n"		/* branch to lwarx */
			"	isync				\n"
			"	mfcr    %1          \n"
:			"=&r"(found), "=r"(condition_register), "+m"(ptr->value)
:			"i"(*expected), "r"(newval), "r"(&ptr->value)
:			"memory", "cc");
	else
#endif
		__asm__ __volatile__(
			"	sync				\n"
			"	lwarx   %0,0,%5		\n"
			"	cmpw    %0,%3		\n"
			"	bne     $+12		\n"		/* branch to isync */
			"	stwcx.  %4,0,%5		\n"
			"	bne     $-16		\n"		/* branch to lwarx */
			"	isync				\n"
			"	mfcr    %1          \n"
:			"=&r"(found), "=r"(condition_register), "+m"(ptr->value)
:			"r"(*expected), "r"(newval), "r"(&ptr->value)
:			"memory", "cc");

	ret = (condition_register >> 29) & 1;	/* test eq bit of cr0 */
	if (!ret)
		*expected = found;
	return ret;
}

/*
 * This mirrors gcc __sync_fetch_and_add().
 *
 * Like tas(), use constraint "=&b" to avoid allocating r0.
 */
#define PG_HAVE_ATOMIC_FETCH_ADD_U32
static inline uint32
pg_atomic_fetch_add_u32_impl(volatile pg_atomic_uint32 *ptr, int32 add_)
{
	uint32 _t;
	uint32 res;

#ifdef HAVE_I_CONSTRAINT__BUILTIN_CONSTANT_P
	if (__builtin_constant_p(add_) &&
		add_ <= PG_INT16_MAX && add_ >= PG_INT16_MIN)
		__asm__ __volatile__(
			"	sync				\n"
			"	lwarx   %1,0,%4		\n"
			"	addi    %0,%1,%3	\n"
			"	stwcx.  %0,0,%4		\n"
			"	bne     $-12		\n"		/* branch to lwarx */
			"	isync				\n"
:			"=&r"(_t), "=&b"(res), "+m"(ptr->value)
:			"i"(add_), "r"(&ptr->value)
:			"memory", "cc");
	else
#endif
		__asm__ __volatile__(
			"	sync				\n"
			"	lwarx   %1,0,%4		\n"
			"	add     %0,%1,%3	\n"
			"	stwcx.  %0,0,%4		\n"
			"	bne     $-12		\n"		/* branch to lwarx */
			"	isync				\n"
:			"=&r"(_t), "=&r"(res), "+m"(ptr->value)
:			"r"(add_), "r"(&ptr->value)
:			"memory", "cc");

	return res;
}

#ifdef PG_HAVE_ATOMIC_U64_SUPPORT

#define PG_HAVE_ATOMIC_COMPARE_EXCHANGE_U64
static inline bool
pg_atomic_compare_exchange_u64_impl(volatile pg_atomic_uint64 *ptr,
									uint64 *expected, uint64 newval)
{
	uint64 found;
	uint32 condition_register;
	bool ret;

	/* Like u32, but s/lwarx/ldarx/; s/stwcx/stdcx/; s/cmpw/cmpd/ */
#ifdef HAVE_I_CONSTRAINT__BUILTIN_CONSTANT_P
	if (__builtin_constant_p(*expected) &&
		(int64) *expected <= PG_INT16_MAX &&
		(int64) *expected >= PG_INT16_MIN)
		__asm__ __volatile__(
			"	sync				\n"
			"	ldarx   %0,0,%5		\n"
			"	cmpdi   %0,%3		\n"
			"	bne     $+12		\n"		/* branch to isync */
			"	stdcx.  %4,0,%5		\n"
			"	bne     $-16		\n"		/* branch to ldarx */
			"	isync				\n"
			"	mfcr    %1          \n"
:			"=&r"(found), "=r"(condition_register), "+m"(ptr->value)
:			"i"(*expected), "r"(newval), "r"(&ptr->value)
:			"memory", "cc");
	else
#endif
		__asm__ __volatile__(
			"	sync				\n"
			"	ldarx   %0,0,%5		\n"
			"	cmpd    %0,%3		\n"
			"	bne     $+12		\n"		/* branch to isync */
			"	stdcx.  %4,0,%5		\n"
			"	bne     $-16		\n"		/* branch to ldarx */
			"	isync				\n"
			"	mfcr    %1          \n"
:			"=&r"(found), "=r"(condition_register), "+m"(ptr->value)
:			"r"(*expected), "r"(newval), "r"(&ptr->value)
:			"memory", "cc");

	ret = (condition_register >> 29) & 1;	/* test eq bit of cr0 */
	if (!ret)
		*expected = found;
	return ret;
}

#define PG_HAVE_ATOMIC_FETCH_ADD_U64
static inline uint64
pg_atomic_fetch_add_u64_impl(volatile pg_atomic_uint64 *ptr, int64 add_)
{
	uint64 _t;
	uint64 res;

	/* Like u32, but s/lwarx/ldarx/; s/stwcx/stdcx/ */
#ifdef HAVE_I_CONSTRAINT__BUILTIN_CONSTANT_P
	if (__builtin_constant_p(add_) &&
		add_ <= PG_INT16_MAX && add_ >= PG_INT16_MIN)
		__asm__ __volatile__(
			"	sync				\n"
			"	ldarx   %1,0,%4		\n"
			"	addi    %0,%1,%3	\n"
			"	stdcx.  %0,0,%4		\n"
			"	bne     $-12		\n"		/* branch to ldarx */
			"	isync				\n"
:			"=&r"(_t), "=&b"(res), "+m"(ptr->value)
:			"i"(add_), "r"(&ptr->value)
:			"memory", "cc");
	else
#endif
		__asm__ __volatile__(
			"	sync				\n"
			"	ldarx   %1,0,%4		\n"
			"	add     %0,%1,%3	\n"
			"	stdcx.  %0,0,%4		\n"
			"	bne     $-12		\n"		/* branch to ldarx */
			"	isync				\n"
:			"=&r"(_t), "=&r"(res), "+m"(ptr->value)
:			"r"(add_), "r"(&ptr->value)
:			"memory", "cc");

	return res;
}

#endif /* PG_HAVE_ATOMIC_U64_SUPPORT */

/* per architecture manual doubleword accesses have single copy atomicity */
#define PG_HAVE_8BYTE_SINGLE_COPY_ATOMICITY