bionic: Do not use <sys/atomics.h> for platform code.
We're going to modify the __atomic_xxx implementation to provide
full memory barriers, to avoid problems for NDK machine code that
link to these functions.
First step is to remove their usage from our platform code.
We now use inlined versions of the same functions for a slight
performance boost.
+ remove obsolete atomics_x86.c (was never compiled)
NOTE: This improvement was benchmarked on various devices.
Comparing a pthread mutex lock + atomic increment + unlock
we get:
- ARMv7 emulator, running on a 2.4 GHz Xeon:
before: 396 ns after: 288 ns
- x86 emulator in KVM mode on same machine:
before: 27 ns after: 27 ns
- Google Nexus S, in ARMv7 mode (single-core):
before: 82 ns after: 76 ns
- Motorola Xoom, in ARMv7 mode (multi-core):
before: 121 ns after: 120 ns
The code has also been rebuilt in ARMv5TE mode for correctness.
Change-Id: Ic1dc72b173d59b2e7af901dd70d6a72fb2f64b17
diff --git a/libc/bionic/atomics_x86.c b/libc/bionic/atomics_x86.c
deleted file mode 100644
index fd60f4f..0000000
--- a/libc/bionic/atomics_x86.c
+++ /dev/null
@@ -1,95 +0,0 @@
-/*
- * Copyright (C) 2008 The Android Open Source Project
- * All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions
- * are met:
- * * Redistributions of source code must retain the above copyright
- * notice, this list of conditions and the following disclaimer.
- * * Redistributions in binary form must reproduce the above copyright
- * notice, this list of conditions and the following disclaimer in
- * the documentation and/or other materials provided with the
- * distribution.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
- * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
- * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
- * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
- * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
- * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
- * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
- * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
- * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
- * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
- * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
- * SUCH DAMAGE.
- */
-#include <sys/atomics.h>
-
-#define FUTEX_SYSCALL 240
-#define FUTEX_WAIT 0
-#define FUTEX_WAKE 1
-
-int __futex_wait(volatile void *ftx, int val)
-{
- int ret;
- asm volatile (
- "int $0x80;"
- : "=a" (ret)
- : "0" (FUTEX_SYSCALL),
- "b" (ftx),
- "c" (FUTEX_WAIT),
- "d" (val),
- "S" (0)
- );
- return ret;
-}
-
-int __futex_wake(volatile void *ftx, int count)
-{
- int ret;
- asm volatile (
- "int $0x80;"
- : "=a" (ret)
- : "0" (FUTEX_SYSCALL),
- "b" (ftx),
- "c" (FUTEX_WAKE),
- "d" (count)
- );
- return ret;
-}
-
-int __atomic_cmpxchg(int old, int new, volatile int* addr) {
- int xchg;
- asm volatile (
- "lock;"
- "cmpxchg %%ecx, (%%edx);"
- "setne %%al;"
- : "=a" (xchg)
- : "a" (old),
- "c" (new),
- "d" (addr)
- );
- return xchg;
-}
-
-int __atomic_swap(int new, volatile int* addr) {
- int old;
- asm volatile (
- "lock;"
- "xchg %%ecx, (%%edx);"
- : "=c" (old)
- : "c" (new),
- "d" (addr)
- );
- return old;
-}
-
-int __atomic_dec(volatile int* addr) {
- int old;
- do {
- old = *addr;
- } while (atomic_cmpxchg(old, old-1, addr));
- return old;
-}
diff --git a/libc/bionic/pthread.c b/libc/bionic/pthread.c
index b893a12..b56822f 100644
--- a/libc/bionic/pthread.c
+++ b/libc/bionic/pthread.c
@@ -692,7 +692,7 @@
goto Exit;
}
}
- while ( __atomic_cmpxchg( flags, flags | PTHREAD_ATTR_FLAG_DETACHED,
+ while ( __bionic_cmpxchg( flags, flags | PTHREAD_ATTR_FLAG_DETACHED,
(volatile int*)&thread->attr.flags ) != 0 );
Exit:
pthread_mutex_unlock(&gThreadListLock);
@@ -926,17 +926,17 @@
int shared = mutex->value & MUTEX_SHARED_MASK;
/*
* The common case is an unlocked mutex, so we begin by trying to
- * change the lock's state from 0 to 1. __atomic_cmpxchg() returns 0
+ * change the lock's state from 0 to 1. __bionic_cmpxchg() returns 0
* if it made the swap successfully. If the result is nonzero, this
* lock is already held by another thread.
*/
- if (__atomic_cmpxchg(shared|0, shared|1, &mutex->value ) != 0) {
+ if (__bionic_cmpxchg(shared|0, shared|1, &mutex->value ) != 0) {
/*
* We want to go to sleep until the mutex is available, which
* requires promoting it to state 2. We need to swap in the new
* state value and then wait until somebody wakes us up.
*
- * __atomic_swap() returns the previous value. We swap 2 in and
+ * __bionic_swap() returns the previous value. We swap 2 in and
* see if we got zero back; if so, we have acquired the lock. If
* not, another thread still holds the lock and we wait again.
*
@@ -947,7 +947,7 @@
* that the mutex is in state 2 when we go to sleep on it, which
* guarantees a wake-up call.
*/
- while (__atomic_swap(shared|2, &mutex->value ) != (shared|0))
+ while (__bionic_swap(shared|2, &mutex->value ) != (shared|0))
__futex_wait_ex(&mutex->value, shared, shared|2, 0);
}
ANDROID_MEMBAR_FULL();
@@ -967,10 +967,10 @@
/*
* The mutex state will be 1 or (rarely) 2. We use an atomic decrement
- * to release the lock. __atomic_dec() returns the previous value;
+ * to release the lock. __bionic_atomic_dec() returns the previous value;
* if it wasn't 1 we have to do some additional work.
*/
- if (__atomic_dec(&mutex->value) != (shared|1)) {
+ if (__bionic_atomic_dec(&mutex->value) != (shared|1)) {
/*
* Start by releasing the lock. The decrement changed it from
* "contended lock" to "uncontended lock", which means we still
@@ -1158,7 +1158,7 @@
/* Handle common case first */
if ( __likely(mtype == MUTEX_TYPE_NORMAL) )
{
- if (__atomic_cmpxchg(shared|0, shared|1, &mutex->value) == 0) {
+ if (__bionic_cmpxchg(shared|0, shared|1, &mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
@@ -1256,13 +1256,13 @@
if ( __likely(mtype == MUTEX_TYPE_NORMAL) )
{
/* fast path for uncontended lock */
- if (__atomic_cmpxchg(shared|0, shared|1, &mutex->value) == 0) {
+ if (__bionic_cmpxchg(shared|0, shared|1, &mutex->value) == 0) {
ANDROID_MEMBAR_FULL();
return 0;
}
/* loop while needed */
- while (__atomic_swap(shared|2, &mutex->value) != (shared|0)) {
+ while (__bionic_swap(shared|2, &mutex->value) != (shared|0)) {
if (__timespec_to_absolute(&ts, &abstime, clock) < 0)
return EBUSY;
@@ -1431,7 +1431,7 @@
long oldval = cond->value;
long newval = ((oldval - COND_COUNTER_INCREMENT) & COND_COUNTER_MASK)
| flags;
- if (__atomic_cmpxchg(oldval, newval, &cond->value) == 0)
+ if (__bionic_cmpxchg(oldval, newval, &cond->value) == 0)
break;
}
diff --git a/libc/bionic/semaphore.c b/libc/bionic/semaphore.c
index 96819ae..9bc8412 100644
--- a/libc/bionic/semaphore.c
+++ b/libc/bionic/semaphore.c
@@ -174,7 +174,7 @@
new = SEMCOUNT_DECREMENT(old);
}
- while (__atomic_cmpxchg((int)(old|shared),
+ while (__bionic_cmpxchg((int)(old|shared),
(int)(new|shared),
(volatile int *)pvalue) != 0);
return ret;
@@ -198,7 +198,7 @@
new = SEMCOUNT_DECREMENT(old);
}
- while (__atomic_cmpxchg((int)(old|shared),
+ while (__bionic_cmpxchg((int)(old|shared),
(int)(new|shared),
(volatile int *)pvalue) != 0);
@@ -235,7 +235,7 @@
else
new = SEMCOUNT_INCREMENT(old);
}
- while ( __atomic_cmpxchg((int)(old|shared),
+ while ( __bionic_cmpxchg((int)(old|shared),
(int)(new|shared),
(volatile int*)pvalue) != 0);
diff --git a/libc/private/bionic_atomic_arm.h b/libc/private/bionic_atomic_arm.h
new file mode 100644
index 0000000..275c1c9
--- /dev/null
+++ b/libc/private/bionic_atomic_arm.h
@@ -0,0 +1,284 @@
+/*
+ * Copyright (C) 2011 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef BIONIC_ATOMIC_ARM_H
+#define BIONIC_ATOMIC_ARM_H
+
+#include <machine/cpu-features.h>
+
+/* Some of the harware instructions used below are not available in Thumb-1
+ * mode (they are if you build in ARM or Thumb-2 mode though). To solve this
+ * problem, we're going to use the same technique than libatomics_ops,
+ * which is to temporarily switch to ARM, do the operation, then switch
+ * back to Thumb-1.
+ *
+ * This results in two 'bx' jumps, just like a normal function call, but
+ * everything is kept inlined, avoids loading or computing the function's
+ * address, and prevents a little I-cache trashing too.
+ *
+ * However, it is highly recommended to avoid compiling any C library source
+ * file that use these functions in Thumb-1 mode.
+ *
+ * Define three helper macros to implement this:
+ */
+#if defined(__thumb__) && !defined(__thumb2__)
+# define __ATOMIC_SWITCH_TO_ARM \
+ "adr r3, 5f\n" \
+ "bx r3\n" \
+ ".align\n" \
+ ".arm\n" \
+ "5:\n"
+/* note: the leading \n below is intentional */
+# define __ATOMIC_SWITCH_TO_THUMB \
+ "\n" \
+ "adr r3, 6f\n" \
+ "bx r3\n" \
+ ".thumb" \
+ "6:\n"
+
+# define __ATOMIC_CLOBBERS "r3" /* list of clobbered registers */
+
+/* Warn the user that ARM mode should really be preferred! */
+# warning Rebuilding this source file in ARM mode is highly recommended for performance!!
+
+#else
+# define __ATOMIC_SWITCH_TO_ARM /* nothing */
+# define __ATOMIC_SWITCH_TO_THUMB /* nothing */
+# define __ATOMIC_CLOBBERS /* nothing */
+#endif
+
+
+/* Define a full memory barrier, this is only needed if we build the
+ * platform for a multi-core device. For the record, using a 'dmb'
+ * instruction on a Nexus One device can take up to 180 ns even if
+ * it is completely un-necessary on this device.
+ *
+ * NOTE: This is where the platform and NDK headers atomic headers are
+ * going to diverge. With the NDK, we don't know if the generated
+ * code is going to run on a single or multi-core device, so we
+ * need to be cautious.
+ *
+ * Fortunately, we can use the kernel helper function that is
+ * mapped at address 0xffff0fa0 in all user process, and that
+ * provides a device-specific barrier operation.
+ *
+ * I.e. on single-core devices, the helper immediately returns,
+ * on multi-core devices, it uses "dmb" or any other means to
+ * perform a full-memory barrier.
+ *
+ * There are three cases to consider for the platform:
+ *
+ * - multi-core ARMv7-A => use the 'dmb' hardware instruction
+ * - multi-core ARMv6 => use the coprocessor
+ * - single core ARMv5TE/6/7 => do not use any hardware barrier
+ */
+#if defined(ANDROID_SMP) && ANDROID_SMP == 1
+
+/* Sanity check, multi-core is only supported starting from ARMv6 */
+# if __ARM_ARCH__ < 6
+# error ANDROID_SMP should not be set to 1 for an ARM architecture less than 6
+# endif
+
+# ifdef __ARM_HAVE_DMB
+/* For ARMv7-A, we can use the 'dmb' instruction directly */
+__ATOMIC_INLINE__ void
+__bionic_memory_barrier(void)
+{
+ /* Note: we always build in ARM or Thumb-2 on ARMv7-A, so don't
+ * bother with __ATOMIC_SWITCH_TO_ARM */
+ __asm__ __volatile__ ( "dmb" : : : "memory" );
+}
+# else /* !__ARM_HAVE_DMB */
+/* Otherwise, i.e. for multi-core ARMv6, we need to use the coprocessor,
+ * which requires the use of a general-purpose register, which is slightly
+ * less efficient.
+ */
+__ATOMIC_INLINE__ void
+__bionic_memory_barrier(void)
+{
+ __asm__ __volatile__ (
+ __SWITCH_TO_ARM
+ "mcr p15, 0, %0, c7, c10, 5"
+ __SWITCH_TO_THUMB
+ : : "r" (0) : __ATOMIC_CLOBBERS "memory");
+}
+# endif /* !__ARM_HAVE_DMB */
+#else /* !ANDROID_SMP */
+__ATOMIC_INLINE__ void
+__bionic_memory_barrier(void)
+{
+ /* A simple compiler barrier */
+ __asm__ __volatile__ ( "" : : : "memory" );
+}
+#endif /* !ANDROID_SMP */
+
+/* Compare-and-swap, without any explicit barriers. Note that this functions
+ * returns 0 on success, and 1 on failure. The opposite convention is typically
+ * used on other platforms.
+ *
+ * There are two cases to consider:
+ *
+ * - ARMv6+ => use LDREX/STREX instructions
+ * - < ARMv6 => use kernel helper function mapped at 0xffff0fc0
+ *
+ * LDREX/STREX are only available starting from ARMv6
+ */
+#ifdef __ARM_HAVE_LDREX_STREX
+__ATOMIC_INLINE__ int
+__bionic_cmpxchg(int32_t old_value, int32_t new_value, volatile int32_t* ptr)
+{
+ int32_t prev, status;
+ do {
+ __asm__ __volatile__ (
+ __ATOMIC_SWITCH_TO_ARM
+ "ldrex %0, [%3]\n"
+ "mov %1, #0\n"
+ "teq %0, %4\n"
+#ifdef __thumb2__
+ "it eq\n"
+#endif
+ "strexeq %1, %5, [%3]"
+ __ATOMIC_SWITCH_TO_THUMB
+ : "=&r" (prev), "=&r" (status), "+m"(*ptr)
+ : "r" (ptr), "Ir" (old_value), "r" (new_value)
+ : __ATOMIC_CLOBBERS "cc");
+ } while (__builtin_expect(status != 0, 0));
+ return prev != old_value;
+}
+# else /* !__ARM_HAVE_LDREX_STREX */
+
+/* Use the handy kernel helper function mapped at 0xffff0fc0 */
+typedef int (kernel_cmpxchg)(int32_t, int32_t, volatile int32_t *);
+
+__ATOMIC_INLINE__ int
+__kernel_cmpxchg(int32_t old_value, int32_t new_value, volatile int32_t* ptr)
+{
+ /* Note: the kernel function returns 0 on success too */
+ return (*(kernel_cmpxchg *)0xffff0fc0)(old_value, new_value, ptr);
+}
+
+__ATOMIC_INLINE__ int
+__bionic_cmpxchg(int32_t old_value, int32_t new_value, volatile int32_t* ptr)
+{
+ return __kernel_cmpxchg(old_value, new_value, ptr);
+}
+#endif /* !__ARM_HAVE_LDREX_STREX */
+
+/* Swap operation, without any explicit barriers.
+ * There are again two similar cases to consider:
+ *
+ * ARMv6+ => use LDREX/STREX
+ * < ARMv6 => use SWP instead.
+ */
+#ifdef __ARM_HAVE_LDREX_STREX
+__ATOMIC_INLINE__ int32_t
+__bionic_swap(int32_t new_value, volatile int32_t* ptr)
+{
+ int32_t prev, status;
+ do {
+ __asm__ __volatile__ (
+ __ATOMIC_SWITCH_TO_ARM
+ "ldrex %0, [%3]\n"
+ "strex %1, %4, [%3]"
+ __ATOMIC_SWITCH_TO_THUMB
+ : "=&r" (prev), "=&r" (status), "+m" (*ptr)
+ : "r" (ptr), "r" (new_value)
+ : __ATOMIC_CLOBBERS "cc");
+ } while (__builtin_expect(status != 0, 0));
+ return prev;
+}
+#else /* !__ARM_HAVE_LDREX_STREX */
+__ATOMIC_INLINE__ int32_t
+__bionic_swap(int32_t new_value, volatile int32_t* ptr)
+{
+ int32_t prev;
+ /* NOTE: SWP is available in Thumb-1 too */
+ __asm__ __volatile__ ("swp %0, %2, [%3]"
+ : "=&r" (prev), "+m" (*ptr)
+ : "r" (new_value), "r" (ptr)
+ : "cc");
+ return prev;
+}
+#endif /* !__ARM_HAVE_LDREX_STREX */
+
+/* Atomic increment - without any barriers
+ * This returns the old value
+ */
+#ifdef __ARM_HAVE_LDREX_STREX
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_inc(volatile int32_t* ptr)
+{
+ int32_t prev, tmp, status;
+ do {
+ __asm__ __volatile__ (
+ __ATOMIC_SWITCH_TO_ARM
+ "ldrex %0, [%4]\n"
+ "add %1, %0, #1\n"
+ "strex %2, %1, [%4]"
+ __ATOMIC_SWITCH_TO_THUMB
+ : "=&r" (prev), "=&r" (tmp), "=&r" (status), "+m"(*ptr)
+ : "r" (ptr)
+ : __ATOMIC_CLOBBERS "cc");
+ } while (__builtin_expect(status != 0, 0));
+ return prev;
+}
+#else
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_inc(volatile int32_t* ptr)
+{
+ int32_t prev, status;
+ do {
+ prev = *ptr;
+ status = __kernel_cmpxchg(prev, prev+1, ptr);
+ } while (__builtin_expect(status != 0, 0));
+ return prev;
+}
+#endif
+
+/* Atomic decrement - without any barriers
+ * This returns the old value.
+ */
+#ifdef __ARM_HAVE_LDREX_STREX
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_dec(volatile int32_t* ptr)
+{
+ int32_t prev, tmp, status;
+ do {
+ __asm__ __volatile__ (
+ __ATOMIC_SWITCH_TO_ARM
+ "ldrex %0, [%4]\n"
+ "sub %1, %0, #1\n"
+ "strex %2, %1, [%4]"
+ __ATOMIC_SWITCH_TO_THUMB
+ : "=&r" (prev), "=&r" (tmp), "=&r" (status), "+m"(*ptr)
+ : "r" (ptr)
+ : __ATOMIC_CLOBBERS "cc");
+ } while (__builtin_expect(status != 0, 0));
+ return prev;
+}
+#else
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_dec(volatile int32_t* ptr)
+{
+ int32_t prev, status;
+ do {
+ prev = *ptr;
+ status = __kernel_cmpxchg(prev, prev-1, ptr);
+ } while (__builtin_expect(status != 0, 0));
+ return prev;
+}
+#endif
+
+#endif /* SYS_ATOMICS_ARM_H */
diff --git a/libc/private/bionic_atomic_gcc_builtin.h b/libc/private/bionic_atomic_gcc_builtin.h
new file mode 100644
index 0000000..e7c5761
--- /dev/null
+++ b/libc/private/bionic_atomic_gcc_builtin.h
@@ -0,0 +1,62 @@
+/*
+ * Copyright (C) 2011 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef BIONIC_ATOMIC_GCC_BUILTIN_H
+#define BIONIC_ATOMIC_GCC_BUILTIN_H
+
+/* This header file is used by default if we don't have optimized atomic
+ * routines for a given platform. See bionic_atomic_arm.h and
+ * bionic_atomic_x86.h for examples.
+ */
+
+__ATOMIC_INLINE__ void
+__bionic_memory_barrier(void)
+{
+ __sync_synchronize();
+}
+
+__ATOMIC_INLINE__ int
+__bionic_cmpxchg(int32_t old_value, int32_t new_value, volatile int32_t* ptr)
+{
+ /* We must return 0 on success */
+ return __sync_bool_compare_and_swap(ptr, old_value, new_value) == 0;
+}
+
+__ATOMIC_INLINE__ int32_t
+__bionic_swap(int32_t new_value, volatile int32_t* ptr)
+{
+ int32_t prev;
+ do {
+ prev = *ptr;
+ status = __sync_val_compare_and_swap(ptr, prev, new_value);
+ } while (status == 0);
+ return prev;
+}
+
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_inc(volatile int32_t* ptr)
+{
+ /* We must return the old value */
+ return __sync_fetch_and_add(ptr, 1);
+}
+
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_dec(volatile int32_t* ptr)
+{
+ /* We must return the old value */
+ return __sync_fetch_and_add(ptr, -1);
+}
+
+#endif /* BIONIC_ATOMIC_GCC_BUILTIN_H */
diff --git a/libc/private/bionic_atomic_inline.h b/libc/private/bionic_atomic_inline.h
index 95766e1..821ad39 100644
--- a/libc/private/bionic_atomic_inline.h
+++ b/libc/private/bionic_atomic_inline.h
@@ -43,62 +43,21 @@
extern "C" {
#endif
-/*
- * Define the full memory barrier for an SMP system. This is
- * platform-specific.
+/* Define __ATOMIC_INLINE__ to control the inlining of all atomics
+ * functions declared here. For a slight performance boost, we want
+ * all of them to be always_inline
*/
+#define __ATOMIC_INLINE__ static __inline__ __attribute__((always_inline))
#ifdef __arm__
-#include <machine/cpu-features.h>
-
-/*
- * For ARMv6K we need to issue a specific MCR instead of the DMB, since
- * that wasn't added until v7. For anything older, SMP isn't relevant.
- * Since we don't have an ARMv6K to test with, we're not going to deal
- * with that now.
- *
- * The DMB instruction is found in the ARM and Thumb2 instruction sets.
- * This will fail on plain 16-bit Thumb.
- */
-#if defined(__ARM_HAVE_DMB)
-# define _ANDROID_MEMBAR_FULL_SMP() \
- do { __asm__ __volatile__ ("dmb" ::: "memory"); } while (0)
+# include <bionic_atomic_arm.h>
+#elif defined(__i386__)
+# include <bionic_atomic_x86.h>
#else
-# define _ANDROID_MEMBAR_FULL_SMP() ARM_SMP_defined_but_no_DMB()
+# include <bionic_atomic_gcc_builtin.h>
#endif
-#elif defined(__i386__) || defined(__x86_64__)
-/*
- * For recent x86, we can use the SSE2 mfence instruction.
- */
-# define _ANDROID_MEMBAR_FULL_SMP() \
- do { __asm__ __volatile__ ("mfence" ::: "memory"); } while (0)
-
-#else
-/*
- * Implementation not defined for this platform. Hopefully we're building
- * in uniprocessor mode.
- */
-# define _ANDROID_MEMBAR_FULL_SMP() SMP_barrier_not_defined_for_platform()
-#endif
-
-
-/*
- * Full barrier. On uniprocessors this is just a compiler reorder barrier,
- * which ensures that the statements appearing above the barrier in the C/C++
- * code will be issued after the statements appearing below the barrier.
- *
- * For SMP this also includes a memory barrier instruction. On an ARM
- * CPU this means that the current core will flush pending writes, wait
- * for pending reads to complete, and discard any cached reads that could
- * be stale. Other CPUs may do less, but the end result is equivalent.
- */
-#if ANDROID_SMP != 0
-# define ANDROID_MEMBAR_FULL() _ANDROID_MEMBAR_FULL_SMP()
-#else
-# define ANDROID_MEMBAR_FULL() \
- do { __asm__ __volatile__ ("" ::: "memory"); } while (0)
-#endif
+#define ANDROID_MEMBAR_FULL __bionic_memory_barrier
#ifdef __cplusplus
} // extern "C"
diff --git a/libc/private/bionic_atomic_x86.h b/libc/private/bionic_atomic_x86.h
new file mode 100644
index 0000000..aca0c4b
--- /dev/null
+++ b/libc/private/bionic_atomic_x86.h
@@ -0,0 +1,88 @@
+/*
+ * Copyright (C) 2011 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#ifndef BIONIC_ATOMIC_X86_H
+#define BIONIC_ATOMIC_X86_H
+
+/* Define a full memory barrier, this is only needed if we build the
+ * platform for a multi-core device.
+ */
+#if defined(ANDROID_SMP) && ANDROID_SMP == 1
+__ATOMIC_INLINE__ void
+__bionic_memory_barrier()
+{
+ __asm__ __volatile__ ( "mfence" : : : "memory" );
+}
+#else
+__ATOMIC_INLINE__ void
+__bionic_memory_barrier()
+{
+ /* A simple compiler barrier */
+ __asm__ __volatile__ ( "" : : : "memory" );
+}
+#endif
+
+/* Compare-and-swap, without any explicit barriers. Note that this function
+ * returns 0 on success, and 1 on failure. The opposite convention is typically
+ * used on other platforms.
+ */
+__ATOMIC_INLINE__ int
+__bionic_cmpxchg(int32_t old_value, int32_t new_value, volatile int32_t* ptr)
+{
+ int32_t prev;
+ __asm__ __volatile__ ("lock; cmpxchgl %1, %2"
+ : "=a" (prev)
+ : "q" (new_value), "m" (*ptr), "0" (old_value)
+ : "memory");
+ return prev != old_value;
+}
+
+
+/* Swap, without any explicit barriers */
+__ATOMIC_INLINE__ int32_t
+__bionic_swap(int32_t new_value, volatile int32_t *ptr)
+{
+ __asm__ __volatile__ ("xchgl %1, %0"
+ : "=r" (new_value)
+ : "m" (*ptr), "0" (new_value)
+ : "memory");
+ return new_value;
+}
+
+/* Atomic increment, without explicit barriers */
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_inc(volatile int32_t *ptr)
+{
+ int increment = 1;
+ __asm__ __volatile__ ("lock; xaddl %0, %1"
+ : "+r" (increment), "+m" (*ptr)
+ : : "memory");
+ /* increment now holds the old value of *ptr */
+ return increment;
+}
+
+/* Atomic decrement, without explicit barriers */
+__ATOMIC_INLINE__ int32_t
+__bionic_atomic_dec(volatile int32_t *ptr)
+{
+ int increment = -1;
+ __asm__ __volatile__ ("lock; xaddl %0, %1"
+ : "+r" (increment), "+m" (*ptr)
+ : : "memory");
+ /* increment now holds the old value of *ptr */
+ return increment;
+}
+
+#endif /* BIONIC_ATOMIC_X86_H */
diff --git a/libstdc++/src/one_time_construction.cpp b/libstdc++/src/one_time_construction.cpp
index 2a44c79..f3d7138 100644
--- a/libstdc++/src/one_time_construction.cpp
+++ b/libstdc++/src/one_time_construction.cpp
@@ -20,11 +20,11 @@
// 6 untouched, wait and return 0
// 1 untouched, return 0
retry:
- if (__atomic_cmpxchg(0, 0x2, gv) == 0) {
+ if (__bionic_cmpxchg(0, 0x2, gv) == 0) {
ANDROID_MEMBAR_FULL();
return 1;
}
- __atomic_cmpxchg(0x2, 0x6, gv); // Indicate there is a waiter
+ __bionic_cmpxchg(0x2, 0x6, gv); // Indicate there is a waiter
__futex_wait(gv, 0x6, NULL);
if(*gv != 1) // __cxa_guard_abort was called, let every thread try since there is no return code for this condition
@@ -39,7 +39,7 @@
// 2 -> 1
// 6 -> 1, and wake
ANDROID_MEMBAR_FULL();
- if (__atomic_cmpxchg(0x2, 0x1, gv) == 0) {
+ if (__bionic_cmpxchg(0x2, 0x1, gv) == 0) {
return;
}