add atomics builtins for globals, memory, and setonce

Adds an AtomicMemory{T} type (GenericMemory{:atomic,T,Core.CPU}) for
doing element-wise atomic operations on an array (with a lock
per-element if required).

Adds the full complement of atomic operations for global variables.

Also add a *setonce set of functions for atomically setting a value, but
only if it is unset. The complementary operation (Atomic `_unsetindex!`
for AtomicMemory) is not yet provided.

Author: vtjnash

base/Base.jl

@@ -91,6 +91,36 @@ function replaceproperty!(x, f::Symbol, expected, desired, success_order::Symbol
     val = desired isa ty ? desired : convert(ty, desired)
     return Core.replacefield!(x, f, expected, val, success_order, fail_order)
 end
+function setpropertyonce!(x, f::Symbol, desired, success_order::Symbol=:not_atomic, fail_order::Symbol=success_order)
+    @inline
+    ty = fieldtype(typeof(x), f)
+    val = desired isa ty ? desired : convert(ty, desired)
+    return Core.setfieldonce!(x, f, val, success_order, fail_order)
+end
+
+function swapproperty!(x::Module, f::Symbol, v, order::Symbol=:not_atomic)
+    @inline
+    ty = Core.get_binding_type(x, f)
+    val = v isa ty ? v : convert(ty, v)
+    return Core.swapglobal!(x, f, val, order)
+end
+function modifyproperty!(x::Module, f::Symbol, op, v, order::Symbol=:not_atomic)
+    @inline
+    return Core.modifyglobal!(x, f, op, v, order)
+end
+function replaceproperty!(x::Module, f::Symbol, expected, desired, success_order::Symbol=:not_atomic, fail_order::Symbol=success_order)
+    @inline
+    ty = Core.get_binding_type(x, f)
+    val = desired isa ty ? desired : convert(ty, desired)
+    return Core.replaceglobal!(x, f, expected, val, success_order, fail_order)
+end
+function setpropertyonce!(x::Module, f::Symbol, desired, success_order::Symbol=:not_atomic, fail_order::Symbol=success_order)
+    @inline
+    ty = Core.get_binding_type(x, f)
+    val = desired isa ty ? desired : convert(ty, desired)
+    return Core.setglobalonce!(x, f, val, success_order, fail_order)
+end
+
 
 convert(::Type{Any}, Core.@nospecialize x) = x
 convert(::Type{T}, x::T) where {T} = x

base/boot.jl

@@ -65,7 +65,7 @@
 #end
 
 # struct GenericMemoryRef{kind::Symbol, T, AS::AddrSpace}
-#    mem::Memory{kind, T, AS}
+#    mem::GenericMemory{kind, T, AS}
 #    data::Ptr{Cvoid} # make this GenericPtr{addrspace, Cvoid}
 #end
 
@@ -191,8 +191,8 @@ export
     Tuple, Type, UnionAll, TypeVar, Union, Nothing, Cvoid,
     AbstractArray, DenseArray, NamedTuple, Pair,
     # special objects
-    Function, Method, Array, Memory, MemoryRef, GenericMemory, GenericMemoryRef,
-    Module, Symbol, Task, UndefInitializer, undef, WeakRef, VecElement,
+    Function, Method, Module, Symbol, Task, UndefInitializer, undef, WeakRef, VecElement,
+    Array, Memory, MemoryRef, AtomicMemory, AtomicMemoryRef, GenericMemory, GenericMemoryRef,
     # numeric types
     Number, Real, Integer, Bool, Ref, Ptr,
     AbstractFloat, Float16, Float32, Float64,
@@ -209,10 +209,10 @@ export
     # AST representation
     Expr, QuoteNode, LineNumberNode, GlobalRef,
     # object model functions
-    fieldtype, getfield, setfield!, swapfield!, modifyfield!, replacefield!,
+    fieldtype, getfield, setfield!, swapfield!, modifyfield!, replacefield!, setfieldonce!,
     nfields, throw, tuple, ===, isdefined, eval,
     # access to globals
-    getglobal, setglobal!,
+    getglobal, setglobal!, swapglobal!, modifyglobal!, replaceglobal!, setglobalonce!,
     # ifelse, sizeof    # not exported, to avoid conflicting with Base
     # type reflection
     <:, typeof, isa, typeassert,
@@ -516,22 +516,24 @@ const undef = UndefInitializer()
 (self::Type{GenericMemory{kind,T,addrspace}})(::UndefInitializer, d::NTuple{1,Int}) where {T,kind,addrspace} = self(undef, getfield(d,1))
 # empty vector constructor
 (self::Type{GenericMemory{kind,T,addrspace}})() where {T,kind,addrspace} = self(undef, 0)
-# copy constructors
 
-const Memory{T} = GenericMemory{:not_atomic, T, CPU}
-const MemoryRef{T} = GenericMemoryRef{:not_atomic, T, CPU}
 GenericMemoryRef(mem::GenericMemory) = memoryref(mem)
 GenericMemoryRef(ref::GenericMemoryRef, i::Integer) = memoryref(ref, Int(i), @_boundscheck)
 GenericMemoryRef(mem::GenericMemory, i::Integer) = memoryref(memoryref(mem), Int(i), @_boundscheck)
-MemoryRef(mem::Memory) = memoryref(mem)
-MemoryRef(ref::MemoryRef, i::Integer) = memoryref(ref, Int(i), @_boundscheck)
-MemoryRef(mem::Memory, i::Integer) = memoryref(memoryref(mem), Int(i), @_boundscheck)
-MemoryRef{T}(mem::Memory{T}) where {T} = memoryref(mem)
-MemoryRef{T}(ref::MemoryRef{T}, i::Integer) where {T} = memoryref(ref, Int(i), @_boundscheck)
-MemoryRef{T}(mem::Memory{T}, i::Integer) where {T} = memoryref(memoryref(mem), Int(i), @_boundscheck)
+GenericMemoryRef{kind,<:Any,AS}(mem::GenericMemory{kind,<:Any,AS}) where {kind,AS} = memoryref(mem)
+GenericMemoryRef{kind,<:Any,AS}(ref::GenericMemoryRef{kind,<:Any,AS}, i::Integer) where {kind,AS} = memoryref(ref, Int(i), @_boundscheck)
+GenericMemoryRef{kind,<:Any,AS}(mem::GenericMemory{kind,<:Any,AS}, i::Integer) where {kind,AS}  = memoryref(memoryref(mem), Int(i), @_boundscheck)
+GenericMemoryRef{kind,T,AS}(mem::GenericMemory{kind,T,AS}) where {kind,T,AS}  = memoryref(mem)
+GenericMemoryRef{kind,T,AS}(ref::GenericMemoryRef{kind,T,AS}, i::Integer) where {kind,T,AS} = memoryref(ref, Int(i), @_boundscheck)
+GenericMemoryRef{kind,T,AS}(mem::GenericMemory{kind,T,AS}, i::Integer) where {kind,T,AS} = memoryref(memoryref(mem), Int(i), @_boundscheck)
+
+const Memory{T} = GenericMemory{:not_atomic, T, CPU}
+const MemoryRef{T} = GenericMemoryRef{:not_atomic, T, CPU}
+const AtomicMemory{T} = GenericMemory{:atomic, T, CPU}
+const AtomicMemoryRef{T} = GenericMemoryRef{:atomic, T, CPU}
 
 # construction helpers for Array
-new_as_memoryref(self::Type{GenericMemoryRef{isatomic,T,addrspace}}, m::Int) where {T,isatomic,addrspace} = memoryref(fieldtype(self, :mem)(undef, m))
+new_as_memoryref(self::Type{GenericMemoryRef{kind,T,addrspace}}, m::Int) where {T,kind,addrspace} = memoryref(fieldtype(self, :mem)(undef, m))
 
 # checked-multiply intrinsic function for dimensions
 _checked_mul_dims() = 1, false

base/compiler/abstractinterpretation.jl

@@ -1113,12 +1113,12 @@ function const_prop_function_heuristic(interp::AbstractInterpreter, @nospecializ
                 if !still_nothrow || ismutabletype(arrty)
                     return false
                 end
-            elseif ⊑(𝕃ᵢ, arrty, Array)
+            elseif ⊑(𝕃ᵢ, arrty, Array) || ⊑(𝕃ᵢ, arrty, GenericMemory)
                 return false
             end
         elseif istopfunction(f, :iterate)
             itrty = argtypes[2]
-            if ⊑(𝕃ᵢ, itrty, Array)
+            if ⊑(𝕃ᵢ, itrty, Array) || ⊑(𝕃ᵢ, itrty, GenericMemory)
                 return false
             end
         end
@@ -1501,7 +1501,7 @@ function abstract_iteration(interp::AbstractInterpreter, @nospecialize(itft), @n
     # Return Bottom if this is not an iterator.
     # WARNING: Changes to the iteration protocol must be reflected here,
     # this is not just an optimization.
-    # TODO: this doesn't realize that Array, SimpleVector, Tuple, and NamedTuple do not use the iterate protocol
+    # TODO: this doesn't realize that Array, GenericMemory, SimpleVector, Tuple, and NamedTuple do not use the iterate protocol
     stateordonet === Bottom && return AbstractIterationResult(Any[Bottom], AbstractIterationInfo(CallMeta[CallMeta(Bottom, Any, call.effects, info)], true))
     valtype = statetype = Bottom
     ret = Any[]
@@ -2076,8 +2076,8 @@ function abstract_call_known(interp::AbstractInterpreter, @nospecialize(f),
             return abstract_apply(interp, argtypes, si, sv, max_methods)
         elseif f === invoke
             return abstract_invoke(interp, arginfo, si, sv)
-        elseif f === modifyfield!
-            return abstract_modifyfield!(interp, argtypes, si, sv)
+        elseif f === modifyfield! || f === Core.modifyglobal! || f === Core.memoryrefmodify! || f === atomic_pointermodify
+            return abstract_modifyop!(interp, f, argtypes, si, sv)
         elseif f === Core.finalizer
             return abstract_finalizer(interp, argtypes, sv)
         elseif f === applicable

base/compiler/ssair/inlining.jl

@@ -1286,10 +1286,18 @@ function process_simple!(todo::Vector{Pair{Int,Any}}, ir::IRCode, idx::Int, stat
         end
     end
 
-    if (sig.f !== Core.invoke && sig.f !== Core.finalizer && sig.f !== modifyfield!) &&
-        is_builtin(optimizer_lattice(state.interp), sig)
-        # No inlining for builtins (other invoke/apply/typeassert/finalizer)
-        return nothing
+    if is_builtin(optimizer_lattice(state.interp), sig)
+        let f = sig.f
+            if (f !== Core.invoke &&
+                f !== Core.finalizer &&
+                f !== modifyfield! &&
+                f !== Core.modifyglobal! &&
+                f !== Core.memoryrefmodify! &&
+                f !== atomic_pointermodify)
+                # No inlining defined for most builtins (just invoke/apply/typeassert/finalizer), so attempt an early exit for them
+                return nothing
+            end
+        end
     end
 
     # Special case inliners for regular functions
@@ -1571,7 +1579,7 @@ function handle_opaque_closure_call!(todo::Vector{Pair{Int,Any}},
     return nothing
 end
 
-function handle_modifyfield!_call!(ir::IRCode, idx::Int, stmt::Expr, info::ModifyFieldInfo, state::InliningState)
+function handle_modifyop!_call!(ir::IRCode, idx::Int, stmt::Expr, info::ModifyOpInfo, state::InliningState)
     info = info.info
     info isa MethodResultPure && (info = info.info)
     info isa ConstCallInfo && (info = info.call)
@@ -1679,8 +1687,8 @@ function assemble_inline_todo!(ir::IRCode, state::InliningState)
         # handle special cased builtins
         if isa(info, OpaqueClosureCallInfo)
             handle_opaque_closure_call!(todo, ir, idx, stmt, info, flag, sig, state)
-        elseif isa(info, ModifyFieldInfo)
-            handle_modifyfield!_call!(ir, idx, stmt, info, state)
+        elseif isa(info, ModifyOpInfo)
+            handle_modifyop!_call!(ir, idx, stmt, info, state)
         elseif isa(info, InvokeCallInfo)
             handle_invoke_call!(todo, ir, idx, stmt, info, flag, sig, state)
         elseif isa(info, FinalizerInfo)

base/compiler/stmtinfo.jl

@@ -222,13 +222,18 @@ struct FinalizerInfo <: CallInfo
 end
 
 """
-    info::ModifyFieldInfo <: CallInfo
+    info::ModifyOpInfo <: CallInfo
 
-Represents a resolved all of `modifyfield!(obj, name, op, x, [order])`.
-`info.info` wraps the call information of `op(getfield(obj, name), x)`.
+Represents a resolved call of one of:
+ - `modifyfield!(obj, name, op, x, [order])`
+ - `modifyglobal!(mod, var, op, x, order)`
+ - `memoryrefmodify!(memref, op, x, order, boundscheck)`
+ - `Intrinsics.atomic_pointermodify(ptr, op, x, order)`
+
+`info.info` wraps the call information of `op(getval(), x)`.
 """
-struct ModifyFieldInfo <: CallInfo
-    info::CallInfo # the callinfo for the `op(getfield(obj, name), x)` call
+struct ModifyOpInfo <: CallInfo
+    info::CallInfo # the callinfo for the `op(getval(), x)` call
 end
 
 @specialize