internal/task: use a lock-free queue #2290
Conversation
|
Oops, i meant FIFO not LIFO. Also xtensa seems to still need work? |
|
Size impact statistics:
|
src/runtime/arch_avr.go
Outdated
| @@ -14,3 +17,14 @@ func align(ptr uintptr) uintptr { | |||
| } | |||
|
|
|||
| func getCurrentStackPointer() uintptr | |||
|
|
|||
| //export __sync_val_compare_and_swap_2 | |||
| func __sync_val_compare_and_swap_2(ptr *uint32, expected, desired uint32) uint32 { | |||
There was a problem hiding this comment.
The same goes for AVR.
Note that atomics are somewhat broken on AVR. If you see weird behavior on AVR, it could be the backend that is emitting invalid instructions.
For more details, see: https://reviews.llvm.org/D97127
I am not intending to do anything specific right now, but this might be useful once we want to implement multicore support for something.
Right now:
The Push operation can also be safely called concurrently. . . but only if the tasks are pinned safely (which is currently an issue, but this is more of a future thing). |
niaow
force-pushed
the
atomic-queue-2
branch
from
6b7d551 to
27813be
Compare
|
This has been rebased now that #2307 is merged |
|
I still need to test this fully on AVR though |
This changes task.Stack and task.Queue to use atomic compare-and-swap operations. With this change, they can also be executed safely from multiple threads/cores. Additionally, the queue is actually LIFO now (not sure what I was thinking before).
|
Seems like this is perhaps not working correctly on the HiFive1b? I tried several times, and also ran some other test code successfully. |
|
@aykevl If this is correct, would you be willing to merge it? |
|
I think we can close this now? |
This changes task.Stack and task.Queue to use atomic compare-and-swap operations.
With this change, Push can also be executed safely from multiple threads/cores.
Additionally, the queue is actually LIFO now (not sure what I was thinking before).
Not entirely sure about this yet.