WIP workstealing scheduler + refactor for easier experimentation

base/schedulers/CDLL.jl

@@ -202,9 +202,15 @@
     return node !== nothing && !is_special(node)
 end
 
+const _PADDING_TUPLE = ntuple(zero, 15)
 mutable struct ConcurrentDoublyLinkedList{T}
-    @atomic header::Union{Node{T}, Nothing}
+    @atomic header::Union{Node{T}, Nothing} # 8 bytes
+    padding::NTuple{15,UInt64} # 120 bytes
     @atomic trailer::Union{Node{T}, Nothing}
+    padding2::NTuple{15,UInt64}
+    function ConcurrentDoublyLinkedList{T}(header::Union{Node{T}, Nothing}, trailer::Union{Node{T}, Nothing}) where {T}
+        new{T}(header, _PADDING_TUPLE, trailer, _PADDING_TUPLE)
+    end
 end
 
 function ConcurrentDoublyLinkedList{T}() where {T}
@@ -257,3 +263,3 @@
 
 const Queue = CDLL
 

base/schedulers/workstealing.jl

@@ -74,6 +74,17 @@ function attempt_steal!()
             end
         end
     end
+    for i in 1:(nt) # Try to steal from other threads round robin
+        t = QueueModule.steal!(queue_for(Int(i))) #TODO: Change types of things to avoid the convert
+        if t !== nothing
+            if ccall(:jl_set_task_tid, Cint, (Any, Cint), t, tid-1) == 0
+                push!(queue_for(tid), t)
+                ccall(:jl_wakeup_thread, Cvoid, (Int16,), (Threads.threadid(t) - 1) % Int16)
+            else
+                return t
+            end
+        end
+    end
     return nothing
 end