New design for pthread_once (with cancelability); ABI change

1 files changed, 9 insertions, 7 deletions

diff --git a/ptw32_InterlockedCompareExchange.c b/ptw32_InterlockedCompareExchange.c
index 581d19d..5ffc754 100644
--- a/ptw32_InterlockedCompareExchange.c
+++ b/ptw32_InterlockedCompareExchange.c
@@ -204,12 +204,13 @@ ptw32_InterlockedExchange (LPLONG location,
        * it doesn't lock the bus. If an interrupt or context switch
        * occurs between the MOV and the CMPXCHG then the value in
        * 'location' may have changed, in which case we will loop
-       * back to do the MOV again. Because both instructions
-       * reference the same location, they will not be re-ordered
-       * in the pipeline.
+       * back to do the MOV again.
+       *
+       * FIXME! Need memory barriers for the MOV+CMPXCHG combo?
+       *
        * Tests show that this routine has almost identical timing
        * to Win32's InterlockedExchange(), which is much faster than
-       * using the an inlined 'xchg' instruction, so it's probably
+       * using the inlined 'xchg' instruction above, so it's probably
        * doing something similar to this (on UP systems).
        *
        * Can we do without the PUSH/POP instructions?
@@ -245,9 +246,10 @@ L1:	MOV          eax,dword ptr [ecx]
        * it doesn't lock the bus. If an interrupt or context switch
        * occurs between the movl and the cmpxchgl then the value in
        * 'location' may have changed, in which case we will loop
-       * back to do the movl again. Because both instructions
-       * reference the same location, they will not be re-ordered
-       * in the pipeline.
+       * back to do the movl again.
+       *
+       * FIXME! Need memory barriers for the MOV+CMPXCHG combo?
+       *
        * Tests show that this routine has almost identical timing
        * to Win32's InterlockedExchange(), which is much faster than
        * using the an inlined 'xchg' instruction, so it's probably