locking/atomic/x86: Introduce arch_atomic64_try_cmpxchg to x86_32
Subsystems (2)
PATCH v2 1/4 locking/atomic/x86: Introduce arch_atomic64_try_cmpxchg to x86_32
Introduce arch_atomic64_try_cmpxchg for 32-bit targets to use optimized target specific implementation instead of a generic one. This implementation eliminates dual-word compare after cmpxchg8b instruction and improves generated asm code from:
2273: f0 0f c7 0f lock cmpxchg8b (%edi)
2277: 8b 74 24 2c mov 0x2c(%esp),%esi
227b: 89 d3 mov %edx,%ebx
227d: 89 c2 mov %eax,%edx
227f: 89 5c 24 10 mov %ebx,0x10(%esp)
2283: 8b 7c 24 30 mov 0x30(%esp),%edi
2287: 89 44 24 1c mov %eax,0x1c(%esp)
228b: 31 f2 xor %esi,%edx
228d: 89 d0 mov %edx,%eax
228f: 89 da mov %ebx,%edx
2291: 31 fa xor %edi,%edx
2293: 09 d0 or %edx,%eax
2295: 0f 85 a5 00 00 00 jne 2340 <...>to:
2270: f0 0f c7 0f lock cmpxchg8b (%edi)
2274: 0f 85 a6 00 00 00 jne 2320 <...>--- arch/x86/include/asm/atomic64_32.h | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-)
arch/x86/include/asm/atomic64_32.h +8 -2
@@ -61,12 +61,18 @@ ATOMIC64_DECL(add_unless);#undef __ATOMIC64_DECL#undef ATOMIC64_EXPORT -static __always_inline s64 arch_atomic64_cmpxchg(atomic64_t *v, s64 o, s64 n) +static __always_inline s64 arch_atomic64_cmpxchg(atomic64_t *v, s64 old, s64 new) { -return arch_cmpxchg64 (&v->counter, o, n); +return arch_cmpxchg64 (&v->counter, old, new); }#define arch_atomic64_cmpxchg arch_atomic64_cmpxchg +static __always_inline bool arch_atomic64_try_cmpxchg(atomic64_t *v, s64 *old, s64 new) +{ +return arch_try_cmpxchg64 (&v->counter, old, new); +} +#define arch_atomic64_try_cmpxchg arch_atomic64_try_cmpxchg + static __always_inline s64 arch_atomic64_xchg(atomic64_t *v, s64 n) {s64 o;
Signature · 1 line
2.42.0
PATCH v2 2/4 locking/atomic/x86: Introduce arch_atomic64_read_tearable to x86_32
Introduce arch_atomic64_read_tearable for 32-bit targets to load the value from atomic64_t location in a non-atomic way. The read might be torn, but can safely be consumed by the 64-bit compare-and-swap loop.
--- arch/x86/include/asm/atomic64_32.h | 14 ++++++++++++++ 1 file changed, 14 insertions(+)
arch/x86/include/asm/atomic64_32.h +14 -0
@@ -14,6 +14,20 @@ typedef struct {#define ATOMIC64_INIT (val) { (val) } +/* +* This function is intended to load the value from atomic64_t +* location in a non-atomic way. The read might be torn, but can +* safely be consumed by the 64-bit compare-and-swap loop. +*/ +static __always_inline s64 arch_atomic64_read_tearable(const atomic64_t *v) +{ +/* +* See the comment in arch_atomic_read() on why we use +* __READ_ONCE() instead of READ_ONCE_NOCHECK() here. +*/ +return __READ_ONCE (v->counter); +} +#define __ATOMIC64_DECL (sym) void atomic64_##sym(atomic64_t *, ...)#ifndef ATOMIC64_EXPORT #define ATOMIC64_DECL_ONE __ATOMIC64_DECL
Signature · 1 line
2.42.0
PATCH v2 3/4 locking/atomic/x86: Rewrite x86_32 arch_atomic64_{,fetch}_{and,or,xor}() functions
In reply to Uros BizjakRewrite x86_32 arch_atomic64_{,fetch}_{and,or,xor}() functions to use arch_atomic64_try_cmpxchg. This implementation avoids one extra trip through the cmpxchg loop.
The value preload before the cmpxchg loop does not need to be atomic. Use arch_atomic64_read_tearable(v) to load the value from atomic_t location in a non-atomic way.
The generated code improves from:
1917d5: 31 c9 xor %ecx,%ecx 1917d7: 31 db xor %ebx,%ebx 1917d9: 89 4c 24 3c mov %ecx,0x3c(%esp) 1917dd: 8b 74 24 24 mov 0x24(%esp),%esi 1917e1: 89 c8 mov %ecx,%eax 1917e3: 89 5c 24 34 mov %ebx,0x34(%esp) 1917e7: 8b 7c 24 28 mov 0x28(%esp),%edi 1917eb: 21 ce and %ecx,%esi 1917ed: 89 74 24 4c mov %esi,0x4c(%esp) 1917f1: 21 df and %ebx,%edi 1917f3: 89 de mov %ebx,%esi 1917f5: 89 7c 24 50 mov %edi,0x50(%esp) 1917f9: 8b 54 24 4c mov 0x4c(%esp),%edx 1917fd: 8b 7c 24 2c mov 0x2c(%esp),%edi 191801: 8b 4c 24 50 mov 0x50(%esp),%ecx 191805: 89 d3 mov %edx,%ebx 191807: 89 f2 mov %esi,%edx 191809: f0 0f c7 0f lock cmpxchg8b (%edi) 19180d: 89 c1 mov %eax,%ecx 19180f: 8b 74 24 34 mov 0x34(%esp),%esi 191813: 89 d3 mov %edx,%ebx 191815: 89 44 24 4c mov %eax,0x4c(%esp) 191819: 8b 44 24 3c mov 0x3c(%esp),%eax 19181d: 89 df mov %ebx,%edi 19181f: 89 54 24 44 mov %edx,0x44(%esp) 191823: 89 ca mov %ecx,%edx 191825: 31 de xor %ebx,%esi 191827: 31 c8 xor %ecx,%eax 191829: 09 f0 or %esi,%eax 19182b: 75 ac jne 1917d9 <...>
to:
1912ba: 8b 06 mov (%esi),%eax 1912bc: 8b 56 04 mov 0x4(%esi),%edx 1912bf: 89 44 24 3c mov %eax,0x3c(%esp) 1912c3: 89 c1 mov %eax,%ecx 1912c5: 23 4c 24 34 and 0x34(%esp),%ecx 1912c9: 89 d3 mov %edx,%ebx 1912cb: 23 5c 24 38 and 0x38(%esp),%ebx 1912cf: 89 54 24 40 mov %edx,0x40(%esp) 1912d3: 89 4c 24 2c mov %ecx,0x2c(%esp) 1912d7: 89 5c 24 30 mov %ebx,0x30(%esp) 1912db: 8b 5c 24 2c mov 0x2c(%esp),%ebx 1912df: 8b 4c 24 30 mov 0x30(%esp),%ecx 1912e3: f0 0f c7 0e lock cmpxchg8b (%esi) 1912e7: 0f 85 f3 02 00 00 jne 1915e0 <...>
--- arch/x86/include/asm/atomic64_32.h | 43 +++++++++++++----------------- 1 file changed, 18 insertions(+), 25 deletions(-)
arch/x86/include/asm/atomic64_32.h +18 -25
@@ -215,69 +215,62 @@ static __always_inline s64 arch_atomic64_dec_if_positive(atomic64_t *v) static __always_inline void arch_atomic64_and(s64 i, atomic64_t *v) { -s64 old, c =0 ; +s64 val =arch_atomic64_read_tearable (v); -while ((old =arch_atomic64_cmpxchg (v, c, c & i)) != c) - c = old; +do { }while (!arch_atomic64_try_cmpxchg (v, &val, val & i)); } static __always_inline s64 arch_atomic64_fetch_and(s64 i, atomic64_t *v) { -s64 old, c =0 ; +s64 val =arch_atomic64_read_tearable (v); -while ((old =arch_atomic64_cmpxchg (v, c, c & i)) != c) - c = old; +do { }while (!arch_atomic64_try_cmpxchg (v, &val, val & i)); -return old; +return val; }#define arch_atomic64_fetch_and arch_atomic64_fetch_and static __always_inline void arch_atomic64_or(s64 i, atomic64_t *v) { -s64 old, c =0 ; +s64 val =arch_atomic64_read_tearable (v); -while ((old =arch_atomic64_cmpxchg (v, c, c | i)) != c) - c = old; +do { }while (!arch_atomic64_try_cmpxchg (v, &val, val | i)); } static __always_inline s64 arch_atomic64_fetch_or(s64 i, atomic64_t *v) { -s64 old, c =0 ; +s64 val =arch_atomic64_read_tearable (v); -while ((old =arch_atomic64_cmpxchg (v, c, c | i)) != c) - c = old; +do { }while (!arch_atomic64_try_cmpxchg (v, &val, val | i)); -return old; +return val; }#define arch_atomic64_fetch_or arch_atomic64_fetch_or static __always_inline void arch_atomic64_xor(s64 i, atomic64_t *v) { -s64 old, c =0 ; +s64 val =arch_atomic64_read_tearable (v); -while ((old =arch_atomic64_cmpxchg (v, c, c ^ i)) != c) - c = old; +do { }while (!arch_atomic64_try_cmpxchg (v, &val, val ^ i)); } static __always_inline s64 arch_atomic64_fetch_xor(s64 i, atomic64_t *v) { -s64 old, c =0 ; +s64 val =arch_atomic64_read_tearable (v); -while ((old =arch_atomic64_cmpxchg (v, c, c ^ i)) != c) - c = old; +do { }while (!arch_atomic64_try_cmpxchg (v, &val, val ^ i)); -return old; +return val; }#define arch_atomic64_fetch_xor arch_atomic64_fetch_xor static __always_inline s64 arch_atomic64_fetch_add(s64 i, atomic64_t *v) { -s64 old, c =0 ; +s64 val =arch_atomic64_read_tearable (v); -while ((old =arch_atomic64_cmpxchg (v, c, c + i)) != c) - c = old; +do { }while (!arch_atomic64_try_cmpxchg (v, &val, val + i)); -return old; +return val; }#define arch_atomic64_fetch_add arch_atomic64_fetch_add
Signature · 1 line
2.42.0
PATCH v2 4/4 locking/atomic/x86: Define arch_atomic_sub() family using arch_atomic_add() functions
In reply to Uros BizjakThere is no need to implement arch_atomic_sub() family of inline functions, corresponding macros can be directly implemented using arch_atomic_add() inlines with negated argument.
No functional changes intended.
--- arch/x86/include/asm/atomic.h | 12 ++---------- arch/x86/include/asm/atomic64_64.h | 12 ++---------- 2 files changed, 4 insertions(+), 20 deletions(-)
arch/x86/include/asm/atomic.h +2 -10
@@ -86,11 +86,7 @@ static __always_inline int arch_atomic_add_return(int i, atomic_t *v) }#define arch_atomic_add_return arch_atomic_add_return -static __always_inline int arch_atomic_sub_return(int i, atomic_t *v) -{ -return arch_atomic_add_return (-i, v); -} -#define arch_atomic_sub_return arch_atomic_sub_return +#define arch_atomic_sub_return (i, v) arch_atomic_add_return(-(i), v) static __always_inline int arch_atomic_fetch_add(int i, atomic_t *v) {
@@ -98,11 +94,7 @@ static __always_inline int arch_atomic_fetch_add(int i, atomic_t *v) }#define arch_atomic_fetch_add arch_atomic_fetch_add -static __always_inline int arch_atomic_fetch_sub(int i, atomic_t *v) -{ -return xadd (&v->counter, -i); -} -#define arch_atomic_fetch_sub arch_atomic_fetch_sub +#define arch_atomic_fetch_sub (i, v) arch_atomic_fetch_add(-(i), v) static __always_inline int arch_atomic_cmpxchg(atomic_t *v, int old, int new) {
arch/x86/include/asm/atomic64_64.h +2 -10
@@ -80,11 +80,7 @@ static __always_inline s64 arch_atomic64_add_return(s64 i, atomic64_t *v) }#define arch_atomic64_add_return arch_atomic64_add_return -static __always_inline s64 arch_atomic64_sub_return(s64 i, atomic64_t *v) -{ -return arch_atomic64_add_return (-i, v); -} -#define arch_atomic64_sub_return arch_atomic64_sub_return +#define arch_atomic64_sub_return (i, v) arch_atomic64_add_return(-(i), v) static __always_inline s64 arch_atomic64_fetch_add(s64 i, atomic64_t *v) {
@@ -92,11 +88,7 @@ static __always_inline s64 arch_atomic64_fetch_add(s64 i, atomic64_t *v) }#define arch_atomic64_fetch_add arch_atomic64_fetch_add -static __always_inline s64 arch_atomic64_fetch_sub(s64 i, atomic64_t *v) -{ -return xadd (&v->counter, -i); -} -#define arch_atomic64_fetch_sub arch_atomic64_fetch_sub +#define arch_atomic64_fetch_sub (i, v) arch_atomic64_fetch_add(-(i), v) static __always_inline s64 arch_atomic64_cmpxchg(atomic64_t *v, s64 old, s64 new) {
Signature · 1 line
2.42.0