stan/jdk-24
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

8320052: Zero: Use __atomic built-ins for atomic RMW operations

Browse Source 
Reviewed-by: dholmes, eastigeevich
This commit is contained in:
Aleksey Shipilev 2023-11-20 19:01:20 +00:00
parent 30d895383e
commit 020c9007f8
4 changed files with 12 additions and 211 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

172
src/hotspot/os_cpu/bsd_zero/atomic_bsd_zero.hpp
View File

@ -31,135 +31,6 @@
// Implementation of class atomic
#ifdef M68K
/*
* __m68k_cmpxchg
*
* Atomically store newval in *ptr if *ptr is equal to oldval for user space.
* Returns newval on success and oldval if no exchange happened.
* This implementation is processor specific and works on
* 68020 68030 68040 and 68060.
*
* It will not work on ColdFire, 68000 and 68010 since they lack the CAS
* instruction.
* Using a kernelhelper would be better for arch complete implementation.
*
*/
static inline int __m68k_cmpxchg(int oldval, int newval, volatile int *ptr) {
int ret;
__asm __volatile ("cas%.l %0,%2,%1"
: "=d" (ret), "+m" (*(ptr))
: "d" (newval), "0" (oldval));
return ret;
}
/* Perform an atomic compare and swap: if the current value of `*PTR'
is OLDVAL, then write NEWVAL into `*PTR'. Return the contents of
`*PTR' before the operation.*/
static inline int m68k_compare_and_swap(int newval,
volatile int *ptr,
int oldval) {
for (;;) {
int prev = *ptr;
if (prev != oldval)
return prev;
if (__m68k_cmpxchg (prev, newval, ptr) == newval)
// Success.
return prev;
// We failed even though prev == oldval. Try again.
}
}
/* Atomically add an int to memory. */
static inline int m68k_add_then_fetch(int add_value, volatile int *ptr) {
for (;;) {
// Loop until success.
int prev = *ptr;
if (__m68k_cmpxchg (prev, prev + add_value, ptr) == prev + add_value)
return prev + add_value;
}
}
/* Atomically write VALUE into `*PTR' and returns the previous
contents of `*PTR'. */
static inline int m68k_lock_test_and_set(int newval, volatile int *ptr) {
for (;;) {
// Loop until success.
int prev = *ptr;
if (__m68k_cmpxchg (prev, newval, ptr) == prev)
return prev;
}
}
#endif // M68K
#ifdef ARM
/*
* __kernel_cmpxchg
*
* Atomically store newval in *ptr if *ptr is equal to oldval for user space.
* Return zero if *ptr was changed or non-zero if no exchange happened.
* The C flag is also set if *ptr was changed to allow for assembly
* optimization in the calling code.
*
*/
typedef int (__kernel_cmpxchg_t)(int oldval, int newval, volatile int *ptr);
#define __kernel_cmpxchg (*(__kernel_cmpxchg_t *) 0xffff0fc0)
/* Perform an atomic compare and swap: if the current value of `*PTR'
is OLDVAL, then write NEWVAL into `*PTR'. Return the contents of
`*PTR' before the operation.*/
static inline int arm_compare_and_swap(int newval,
volatile int *ptr,
int oldval) {
for (;;) {
int prev = *ptr;
if (prev != oldval)
return prev;
if (__kernel_cmpxchg (prev, newval, ptr) == 0)
// Success.
return prev;
// We failed even though prev == oldval. Try again.
}
}
/* Atomically add an int to memory. */
static inline int arm_add_then_fetch(int add_value, volatile int *ptr) {
for (;;) {
// Loop until a __kernel_cmpxchg succeeds.
int prev = *ptr;
if (__kernel_cmpxchg (prev, prev + add_value, ptr) == 0)
return prev + add_value;
}
}
/* Atomically write VALUE into `*PTR' and returns the previous
contents of `*PTR'. */
static inline int arm_lock_test_and_set(int newval, volatile int *ptr) {
for (;;) {
// Loop until a __kernel_cmpxchg succeeds.
int prev = *ptr;
if (__kernel_cmpxchg (prev, newval, ptr) == 0)
return prev;
}
}
#endif // ARM
template<size_t byte_size>
struct Atomic::PlatformAdd {
template<typename D, typename I>
@ -178,17 +49,9 @@ inline D Atomic::PlatformAdd<4>::add_then_fetch(D volatile* dest, I add_value,
STATIC_ASSERT(4 == sizeof(I));
STATIC_ASSERT(4 == sizeof(D));
#ifdef ARM
return add_using_helper<int>(arm_add_then_fetch, dest, add_value);
#else
#ifdef M68K
return add_using_helper<int>(m68k_add_then_fetch, dest, add_value);
#else
D res = __atomic_add_fetch(dest, add_value, __ATOMIC_RELEASE);
FULL_MEM_BARRIER;
return res;
#endif // M68K
#endif // ARM
}
template<>
@ -209,26 +72,10 @@ inline T Atomic::PlatformXchg<4>::operator()(T volatile* dest,
T exchange_value,
atomic_memory_order order) const {
STATIC_ASSERT(4 == sizeof(T));
#ifdef ARM
return xchg_using_helper<int>(arm_lock_test_and_set, dest, exchange_value);
#else
#ifdef M68K
return xchg_using_helper<int>(m68k_lock_test_and_set, dest, exchange_value);
#else
// __sync_lock_test_and_set is a bizarrely named atomic exchange
// operation. Note that some platforms only support this with the
// limitation that the only valid value to store is the immediate
// constant 1. There is a test for this in JNI_CreateJavaVM().
T result = __sync_lock_test_and_set (dest, exchange_value);
// All atomic operations are expected to be full memory barriers
// (see atomic.hpp). However, __sync_lock_test_and_set is not
// a full memory barrier, but an acquire barrier. Hence, this added
// barrier. Some platforms (notably ARM) have peculiarities with
// their barrier implementations, delegate it to OrderAccess.
OrderAccess::fence();
FULL_MEM_BARRIER;
T result = __atomic_exchange_n(dest, exchange_value, __ATOMIC_RELAXED);
FULL_MEM_BARRIER;
return result;
#endif // M68K
#endif // ARM
}
template<>
@ -237,8 +84,9 @@ inline T Atomic::PlatformXchg<8>::operator()(T volatile* dest,
T exchange_value,
atomic_memory_order order) const {
STATIC_ASSERT(8 == sizeof(T));
T result = __sync_lock_test_and_set (dest, exchange_value);
OrderAccess::fence();
FULL_MEM_BARRIER;
T result = __atomic_exchange_n(dest, exchange_value, __ATOMIC_RELAXED);
FULL_MEM_BARRIER;
return result;
}
@ -253,20 +101,12 @@ inline T Atomic::PlatformCmpxchg<4>::operator()(T volatile* dest,
T exchange_value,
atomic_memory_order order) const {
STATIC_ASSERT(4 == sizeof(T));
#ifdef ARM
return cmpxchg_using_helper<int>(arm_compare_and_swap, dest, compare_value, exchange_value);
#else
#ifdef M68K
return cmpxchg_using_helper<int>(m68k_compare_and_swap, dest, compare_value, exchange_value);
#else
T value = compare_value;
FULL_MEM_BARRIER;
__atomic_compare_exchange(dest, &value, &exchange_value, /*weak*/false,
__ATOMIC_RELAXED, __ATOMIC_RELAXED);
FULL_MEM_BARRIER;
return value;
#endif // M68K
#endif // ARM
}
template<>

16
src/hotspot/os_cpu/bsd_zero/os_bsd_zero.cpp
View File

@ -311,22 +311,6 @@ extern "C" {
}
};
/////////////////////////////////////////////////////////////////////////////
// Implementations of atomic operations not supported by processors.
// -- http://gcc.gnu.org/onlinedocs/gcc-4.2.1/gcc/Atomic-Builtins.html
#ifndef _LP64
extern "C" {
long long unsigned int __sync_val_compare_and_swap_8(
volatile void *ptr,
long long unsigned int oldval,
long long unsigned int newval) {
ShouldNotCallThis();
return 0; // silence compiler warnings
}
};
#endif // !_LP64
#ifndef PRODUCT
void os::verify_stack_alignment() {
}

19
src/hotspot/os_cpu/linux_zero/atomic_linux_zero.hpp
View File

@ -71,17 +71,9 @@ inline T Atomic::PlatformXchg<4>::operator()(T volatile* dest,
T exchange_value,
atomic_memory_order order) const {
STATIC_ASSERT(4 == sizeof(T));
// __sync_lock_test_and_set is a bizarrely named atomic exchange
// operation. Note that some platforms only support this with the
// limitation that the only valid value to store is the immediate
// constant 1. There is a test for this in JNI_CreateJavaVM().
T result = __sync_lock_test_and_set (dest, exchange_value);
// All atomic operations are expected to be full memory barriers
// (see atomic.hpp). However, __sync_lock_test_and_set is not
// a full memory barrier, but an acquire barrier. Hence, this added
// barrier. Some platforms (notably ARM) have peculiarities with
// their barrier implementations, delegate it to OrderAccess.
OrderAccess::fence();
FULL_MEM_BARRIER;
T result = __atomic_exchange_n(dest, exchange_value, __ATOMIC_RELAXED);
FULL_MEM_BARRIER;
return result;
}
@ -91,8 +83,9 @@ inline T Atomic::PlatformXchg<8>::operator()(T volatile* dest,
T exchange_value,
atomic_memory_order order) const {
STATIC_ASSERT(8 == sizeof(T));
T result = __sync_lock_test_and_set (dest, exchange_value);
OrderAccess::fence();
FULL_MEM_BARRIER;
T result = __atomic_exchange_n(dest, exchange_value, __ATOMIC_RELAXED);
FULL_MEM_BARRIER;
return result;
}

16
src/hotspot/os_cpu/linux_zero/os_linux_zero.cpp
View File

@ -471,22 +471,6 @@ extern "C" {
}
};
/////////////////////////////////////////////////////////////////////////////
// Implementations of atomic operations not supported by processors.
// -- http://gcc.gnu.org/onlinedocs/gcc-4.2.1/gcc/Atomic-Builtins.html
#ifndef _LP64
extern "C" {
long long unsigned int __sync_val_compare_and_swap_8(
volatile void *ptr,
long long unsigned int oldval,
long long unsigned int newval) {
ShouldNotCallThis();
return 0; // silence compiler warnings
}
};
#endif // !_LP64
#ifndef PRODUCT
void os::verify_stack_alignment() {
}