Conversation
|
Closing again until fixed, seems to deadlock a dmd test. |
|
Ready now, after fixing a dead-lock caused by onOutOfMemoryError. |
MartinNowak
force-pushed
the
spinLock
branch
from
95c6505 to
5738103
Compare
|
updated to resolve merge conflict |
src/core/internal/spinlock.d
Outdated
| { | ||
| for (size_t n; atomicLoad!(MemoryOrder.raw)(val); ++n) | ||
| yield(n); | ||
| if (cas(&val, false, true)) |
There was a problem hiding this comment.
Don't you want to optimistically start with the cas?
There was a problem hiding this comment.
Done, hardly measurable.
|
I was to pull this, but raises enough flags to warrant another pass. |
- the compiler doesn't recognize this as super call and will use dynamic dispatch leading to an infinite loop
- less overhead than pthread_mutex - uses test and test-and-set algorithm with configurable backoff
- always use runLocked which now unlocks on error - check for recursive GC entrance during finalization inFinalizer can be thread local b/c we're finalizing in the same thread that currently holds the lock
- all trace handlers should be @nogc but that's currently out of scope b/c it requires to rewrite all backtrace and demangle code
|
Update to address the review points. |
|
|
||
| void lock() | ||
| { | ||
| if (cas(&val, size_t(0), size_t(1))) |
There was a problem hiding this comment.
I think restructuring this function as a do/while so there's only one cas() might generate tighter code. But not urgent.
There was a problem hiding this comment.
I wanted to avoid accessing the contention field.
|
Auto-merge toggled on |
2 participants
Had to reopen spinlock for GC by MartinNowak · Pull Request #1153 after force pushing, please look the existing review before repeating any questions.