[Qemu-devel] [PATCH v2 08/13] qemu-thread: optimize spin_lock for uncontended locks

[Emilio G. Cota]

This way we can acquire the lock with xchg+test, instead of test+xchg+test.
Most spinlocks should be uncontended so this should result in a ne
performance gain.

Before:
  4ad957:   eb 09   jmp4ad962 
  4ad959:   0f 1f 80 00 00 00 00nopl   0x0(%rax)
  4ad960:   f3 90   pause
  4ad962:   8b 03   mov(%rbx),%eax
  4ad964:   85 c0   test   %eax,%eax
  4ad966:   75 f8   jne4ad960 
  4ad968:   89 f8   mov%edi,%eax
  4ad96a:   87 03   xchg   %eax,(%rbx)
  4ad96c:   85 c0   test   %eax,%eax
  4ad96e:   75 f2   jne4ad962 

After:
  4ad980:   89 f8   mov%edi,%eax
  4ad982:   87 03   xchg   %eax,(%rbx)
  4ad984:   85 c0   test   %eax,%eax
  4ad986:   74 12   je 4ad99a 
  4ad988:   0f 1f 84 00 00 00 00nopl   0x0(%rax,%rax,1)
  4ad98f:   00
  4ad990:   8b 03   mov(%rbx),%eax
  4ad992:   85 c0   test   %eax,%eax
  4ad994:   74 ea   je 4ad980 
  4ad996:   f3 90   pause
  4ad998:   eb f6   jmp4ad990 

Signed-off-by: Emilio G. Cota 
---
 include/qemu/thread.h | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/include/qemu/thread.h b/include/qemu/thread.h
index 599965e..e2af57c 100644
--- a/include/qemu/thread.h
+++ b/include/qemu/thread.h
@@ -74,11 +74,11 @@ static inline void qemu_spin_init(QemuSpin *spin)
 
 static inline void qemu_spin_lock(QemuSpin *spin)
 {
-do {
+while (atomic_xchg(&spin->value, true)) {
 while (atomic_read(&spin->value)) {
 cpu_relax();
 }
-} while (atomic_xchg(&spin->value, true));
+}
 }
 
 static inline int qemu_spin_trylock(QemuSpin *spin)
-- 
2.5.0

Re: [Qemu-devel] [PATCH v2 08/13] qemu-thread: optimize spin_lock for uncontended locks

[Richard Henderson]

On 04/07/2016 10:32 AM, Emilio G. Cota wrote:
>  static inline void qemu_spin_lock(QemuSpin *spin)
>  {
> -do {
> +while (atomic_xchg(&spin->value, true)) {
>  while (atomic_read(&spin->value)) {
>  cpu_relax();
>  }
> -} while (atomic_xchg(&spin->value, true));
> +}
>  }

And merge this one as well, please.  It's a good improvement, but there's
little point to keeping these separate.


r~