diff --git a/base/boot.jl b/base/boot.jl
index d0a686e761bf5..8440465167315 100644
--- a/base/boot.jl
+++ b/base/boot.jl
@@ -147,8 +147,6 @@ export
     # constants
     nothing, Main
 
-const AnyVector = Array{Any,1}
-
 abstract type Number end
 abstract type Real     <: Number end
 abstract type AbstractFloat <: Real end
diff --git a/base/coreimg.jl b/base/coreimg.jl
index 8b552d51252e7..10fc79a2947e5 100644
--- a/base/coreimg.jl
+++ b/base/coreimg.jl
@@ -16,6 +16,7 @@ isdefined(Main, :Base) || ((::Type{T})(arg) where {T} = convert(T, arg)::T)
 function return_type end
 
 ## Load essential files and libraries
+include("pair.jl")
 include("essentials.jl")
 include("ctypes.jl")
 include("generator.jl")
@@ -25,7 +26,6 @@ include("options.jl")
 # core operations & types
 include("promotion.jl")
 include("tuple.jl")
-include("pair.jl")
 include("traits.jl")
 include("range.jl")
 include("expr.jl")
diff --git a/base/dict.jl b/base/dict.jl
index cd41a725d669f..73c78165ae2b0 100644
--- a/base/dict.jl
+++ b/base/dict.jl
@@ -595,81 +595,5 @@ function filter!(f, d::Union{ObjectIdDict,Dict})
     return d
 end
 
-struct ImmutableDict{K,V} <: Associative{K,V}
-    parent::ImmutableDict{K,V}
-    key::K
-    value::V
-    ImmutableDict{K,V}() where {K,V} = new() # represents an empty dictionary
-    ImmutableDict{K,V}(key, value) where {K,V} = (empty = new(); new(empty, key, value))
-    ImmutableDict{K,V}(parent::ImmutableDict, key, value) where {K,V} = new(parent, key, value)
-end
-
-"""
-    ImmutableDict
-
-ImmutableDict is a Dictionary implemented as an immutable linked list,
-which is optimal for small dictionaries that are constructed over many individual insertions
-Note that it is not possible to remove a value, although it can be partially overridden and hidden
-by inserting a new value with the same key
-
-    ImmutableDict(KV::Pair)
-
-Create a new entry in the Immutable Dictionary for the key => value pair
-
- - use `(key => value) in dict` to see if this particular combination is in the properties set
- - use `get(dict, key, default)` to retrieve the most recent value for a particular key
-
-"""
-ImmutableDict
-ImmutableDict(KV::Pair{K,V}) where {K,V} = ImmutableDict{K,V}(KV[1], KV[2])
-ImmutableDict(t::ImmutableDict{K,V}, KV::Pair) where {K,V} = ImmutableDict{K,V}(t, KV[1], KV[2])
-
-function in(key_value::Pair, dict::ImmutableDict, valcmp=(==))
-    key, value = key_value
-    while isdefined(dict, :parent)
-        if dict.key == key
-            valcmp(value, dict.value) && return true
-        end
-        dict = dict.parent
-    end
-    return false
-end
-
-function haskey(dict::ImmutableDict, key)
-    while isdefined(dict, :parent)
-        dict.key == key && return true
-        dict = dict.parent
-    end
-    return false
-end
-
-function getindex(dict::ImmutableDict, key)
-    while isdefined(dict, :parent)
-        dict.key == key && return dict.value
-        dict = dict.parent
-    end
-    throw(KeyError(key))
-end
-function get(dict::ImmutableDict, key, default)
-    while isdefined(dict, :parent)
-        dict.key == key && return dict.value
-        dict = dict.parent
-    end
-    return default
-end
-
-# this actually defines reverse iteration (e.g. it should not be used for merge/copy/filter type operations)
-start(t::ImmutableDict) = t
-next(::ImmutableDict{K,V}, t) where {K,V} = (Pair{K,V}(t.key, t.value), t.parent)
-done(::ImmutableDict, t) = !isdefined(t, :parent)
-length(t::ImmutableDict) = count(x->true, t)
-isempty(t::ImmutableDict) = done(t, start(t))
-function similar(t::ImmutableDict)
-    while isdefined(t, :parent)
-        t = t.parent
-    end
-    return t
-end
-
 _similar_for{P<:Pair}(c::Dict, ::Type{P}, itr, isz) = similar(c, P)
 _similar_for(c::Associative, T, itr, isz) = throw(ArgumentError("for Associatives, similar requires an element type of Pair;\n  if calling map, consider a comprehension instead"))
diff --git a/base/distributed/messages.jl b/base/distributed/messages.jl
index 17a447a154d1f..bdca28c77ec30 100644
--- a/base/distributed/messages.jl
+++ b/base/distributed/messages.jl
@@ -38,17 +38,17 @@ null_id(id) =  id == RRID(0, 0)
 struct CallMsg{Mode} <: AbstractMsg
     f::Function
     args::Tuple
-    kwargs::Array
+    kwargs::Base.KWDict
 end
 struct CallWaitMsg <: AbstractMsg
     f::Function
     args::Tuple
-    kwargs::Array
+    kwargs::Base.KWDict
 end
 struct RemoteDoMsg <: AbstractMsg
     f::Function
     args::Tuple
-    kwargs::Array
+    kwargs::Base.KWDict
 end
 struct ResultMsg <: AbstractMsg
     value::Any
diff --git a/base/essentials.jl b/base/essentials.jl
index c5322273ba4a5..0e4299f7032a5 100644
--- a/base/essentials.jl
+++ b/base/essentials.jl
@@ -336,7 +336,9 @@ function vector_any(xs::ANY...)
     a
 end
 
-function as_kwargs(xs::Union{AbstractArray,Associative})
+as_kwargs(xs::Union{AbstractArray,Associative}) = collect_as_kwargs(xs)
+
+function collect_as_kwargs(xs::Union{AbstractArray,Associative})
     n = length(xs)
     to = Vector{Any}(n*2)
     i = 1
@@ -369,3 +371,84 @@ call obsolete versions of a function `f`.
 `f` directly, and the type of the result cannot be inferred by the compiler.)
 """
 invokelatest(f, args...) = Core._apply_latest(f, args)
+
+struct ImmutableDict{K,V} <: Associative{K,V}
+    parent::ImmutableDict{K,V}
+    key::K
+    value::V
+    ImmutableDict{K,V}() where {K,V} = new() # represents an empty dictionary
+    ImmutableDict{K,V}(key, value) where {K,V} = (empty = new(); new(empty, key, value))
+    ImmutableDict{K,V}(parent::ImmutableDict, key, value) where {K,V} = new(parent, key, value)
+end
+
+const KWDict = ImmutableDict{Symbol,Any}
+const EmptyKWDict = KWDict()
+
+as_kwargs(xs::KWDict) = xs
+
+"""
+    ImmutableDict
+
+ImmutableDict is a Dictionary implemented as an immutable linked list,
+which is optimal for small dictionaries that are constructed over many individual insertions
+Note that it is not possible to remove a value, although it can be partially overridden and hidden
+by inserting a new value with the same key
+
+    ImmutableDict(KV::Pair)
+
+Create a new entry in the Immutable Dictionary for the key => value pair
+
+ - use `(key => value) in dict` to see if this particular combination is in the properties set
+ - use `get(dict, key, default)` to retrieve the most recent value for a particular key
+
+"""
+ImmutableDict
+ImmutableDict(KV::Pair{K,V}) where {K,V} = ImmutableDict{K,V}(KV[1], KV[2])
+ImmutableDict(t::ImmutableDict{K,V}, KV::Pair) where {K,V} = ImmutableDict{K,V}(t, KV[1], KV[2])
+
+function in(key_value::Pair, dict::ImmutableDict, valcmp=(==))
+    key, value = key_value
+    while isdefined(dict, :parent)
+        if dict.key == key
+            valcmp(value, dict.value) && return true
+        end
+        dict = dict.parent
+    end
+    return false
+end
+
+function haskey(dict::ImmutableDict, key)
+    while isdefined(dict, :parent)
+        dict.key == key && return true
+        dict = dict.parent
+    end
+    return false
+end
+
+function getindex(dict::ImmutableDict, key)
+    while isdefined(dict, :parent)
+        dict.key == key && return dict.value
+        dict = dict.parent
+    end
+    throw(KeyError(key))
+end
+function get(dict::ImmutableDict, key, default)
+    while isdefined(dict, :parent)
+        dict.key == key && return dict.value
+        dict = dict.parent
+    end
+    return default
+end
+
+# this actually defines reverse iteration (e.g. it should not be used for merge/copy/filter type operations)
+start(t::ImmutableDict) = t
+next(::ImmutableDict{K,V}, t) where {K,V} = (Pair{K,V}(t.key, t.value), t.parent)
+done(::ImmutableDict, t) = !isdefined(t, :parent)
+length(t::ImmutableDict) = count(x->true, t)
+isempty(t::ImmutableDict) = done(t, start(t))
+function similar(t::ImmutableDict)
+    while isdefined(t, :parent)
+        t = t.parent
+    end
+    return t
+end
diff --git a/base/methodshow.jl b/base/methodshow.jl
index c9f304a509c9e..a258b9df7b627 100644
--- a/base/methodshow.jl
+++ b/base/methodshow.jl
@@ -67,7 +67,7 @@ function arg_decl_parts(m::Method)
 end
 
 function kwarg_decl(m::Method, kwtype::DataType)
-    sig = rewrap_unionall(Tuple{kwtype, Core.AnyVector, unwrap_unionall(m.sig).parameters...}, m.sig)
+    sig = rewrap_unionall(Tuple{kwtype, Any, unwrap_unionall(m.sig).parameters...}, m.sig)
     kwli = ccall(:jl_methtable_lookup, Any, (Any, Any, UInt), kwtype.name.mt, sig, typemax(UInt))
     if kwli !== nothing
         kwli = kwli::Method
diff --git a/base/sysimg.jl b/base/sysimg.jl
index acfb4c7b68ee4..2e5a04122e3ba 100644
--- a/base/sysimg.jl
+++ b/base/sysimg.jl
@@ -45,6 +45,7 @@ if false
 end
 
 ## Load essential files and libraries
+include("pair.jl")
 include("essentials.jl")
 include("ctypes.jl")
 include("base.jl")
@@ -55,7 +56,6 @@ include("options.jl")
 # core operations & types
 include("promotion.jl")
 include("tuple.jl")
-include("pair.jl")
 include("traits.jl")
 include("range.jl")
 include("twiceprecision.jl")
diff --git a/src/julia-syntax.scm b/src/julia-syntax.scm
index ac378d93a5a14..0f5e2f18b3f84 100644
--- a/src/julia-syntax.scm
+++ b/src/julia-syntax.scm
@@ -374,8 +374,6 @@
   (or (number? x) (string? x) (char? x) (and (pair? x) (memq (car x) '(quote inert)))
       (eq? x 'true) (eq? x 'false)))
 
-(define empty-vector-any '(call (core AnyVector) 0))
-
 (define (keywords-method-def-expr name sparams argl body isstaged rett)
   (let* ((kargl (cdar argl))  ;; keyword expressions (= k v)
          (pargl (cdr argl))   ;; positional args
@@ -443,7 +441,7 @@
             ;; call mangled(vals..., [rest_kw,] pargs..., [vararg]...)
             (return (call ,mangled
                           ,@(if ordered-defaults keynames vals)
-                          ,@(if (null? restkw) '() (list empty-vector-any))
+                          ,@(if (null? restkw) '() (list '(top EmptyKWDict)))
                           ,@(map arg-name pargl)
                           ,@(if (null? vararg) '()
                                 (list `(... ,(arg-name (car vararg))))))))
@@ -474,20 +472,29 @@
           `((|::|
              ;; if there are optional positional args, we need to be able to reference the function name
              ,(if (any kwarg? pargl) (gensy) UNUSED)
-             (call (core kwftype) ,ftype)) (:: ,kw (core AnyVector)) ,@pargl ,@vararg)
+             (call (core kwftype) ,ftype)) ,kw ,@pargl ,@vararg)
           `(block
             ;; initialize keyword args to their defaults, or set a flag telling
             ;; whether this keyword needs to be set.
             ,@(map (lambda (kwname) `(local ,kwname)) keynames)
-            ,@(map (lambda (name dflt flag)
-                     (if (const-default? dflt)
-                         `(= ,name ,dflt)
-                         `(= ,flag true)))
-                   keynames vals flags)
+            ,@(apply append (map (lambda (name dflt)
+                                   (if (const-default? dflt)
+                                       `((= ,name ,dflt))
+                                       '()))
+                                 keynames vals))
+            ,@(map (lambda (flag) `(= ,flag true))
+                   flags)
             ,@(if (null? restkw) '()
-                  `((= ,rkw ,empty-vector-any)))
+                  `((= ,rkw (top EmptyKWDict))))
+            ,(if (and (not (null? restkw)) (null? keynames))
+                 `(if (call (core isa) ,kw (top KWDict))
+                      (block
+                       (= ,rkw ,kw)  ;; reuse kwarg list for delegation if possible
+                       (symbolicgoto do_call)))
+                 `(if (call (core isa) ,kw (top KWDict))
+                      (= ,kw (call (top collect_as_kwargs) ,kw))))
             ;; for i = 1:(length(kw)>>1)
-            (for (= ,i (: 1 (call (top >>) (call (top length) ,kw) 1)))
+            (for (= ,i (: (call (top >>) (call (top length) ,kw) 1) -1 1))
                  (block
                   ;; ii = i*2 - 1
                   (= ,ii (call (top -) (call (top *) ,i 2) 1))
@@ -523,11 +530,10 @@
                                              rval0)))
                               ;; if kw[ii] == 'k; k = kw[ii+1]::Type; end
                               `(if (comparison ,elt === (quote ,(decl-var k)))
-                                   (block
-                                    (= ,(decl-var k) ,rval)
-                                    ,@(if (not (const-default? (cadr kvf)))
-                                          `((= ,(caddr kvf) false))
-                                          '()))
+                                   (if ,(caddr kvf)
+                                       (block
+                                        (= ,(decl-var k) ,rval)
+                                        (= ,(caddr kvf) false)))
                                    ,else)))
                           (if (null? restkw)
                               ;; if no rest kw, give error for unrecognized
@@ -535,10 +541,10 @@
                                      ,@(if (null? vararg) '()
                                            (list `(... ,(arg-name (car vararg))))))
                               ;; otherwise add to rest keywords
-                              `(foreigncall 'jl_array_ptr_1d_push (core Void) (call (core svec) Any Any)
-                                            ,rkw 0 (tuple ,elt
-                                                          (call (core arrayref) ,kw
-                                                                (call (top +) ,ii 1))) 0))
+                              `(if (call (top haskey) ,rkw ,elt)
+                                   (null)
+                                   (= ,rkw (call (top KWDict) ,rkw ,elt (call (core arrayref) ,kw
+                                                                              (call (top +) ,ii 1))))))
                           (map list vars vals flags))))
             ;; set keywords that weren't present to their default values
             ,@(apply append
@@ -547,6 +553,7 @@
                                 '()
                                 `((if ,flag (= ,name ,dflt)))))
                           keynames vals flags))
+            (symboliclabel do_call)
             ;; finally, call the core function
             (return (call ,mangled
                           ,@keynames
diff --git a/test/keywordargs.jl b/test/keywordargs.jl
index 0ac76fa7bcaaf..bd13515814ac4 100644
--- a/test/keywordargs.jl
+++ b/test/keywordargs.jl
@@ -188,7 +188,7 @@ function test4974(;kwargs...)
     end
 end
 
-@test test4974(a=1) == (2, [(:a, 1)])
+@test test4974(a=1) == (2, Base.KWDict(:a, 1))
 
 # issue #7704, computed keywords
 @test kwf1(1; (:tens, 2)) == 21
@@ -285,3 +285,17 @@ let a = 0
     g21518()(; :kw=>1)
     @test a == 2
 end
+
+# issue #4916 - rest keywords as a dict
+let ks(;xs...) = xs
+    d = ks(a=1, b=2)
+    @test d[:a] == 1
+    @test d[:b] == 2
+    # preserving order
+    @test collect(ks(a=1, b=2)) == [:a=>1, :b=>2]
+    @test collect(ks(b=2, a=1)) == [:b=>2, :a=>1]
+    # deduplication
+    @test collect(ks(b=2, a=1; :b=>3)) == [:a=>1, :b=>3]
+    @test isempty(ks())
+    @test eltype(ks()) <: Pair
+end
diff --git a/test/parse.jl b/test/parse.jl
index 12377dbdbb01e..c7f99d587fb20 100644
--- a/test/parse.jl
+++ b/test/parse.jl
@@ -989,7 +989,7 @@ let f = function (x; kw...)
     g = function (x; a = 2)
             return (x, a)
         end
-    @test f(1) == (1, Any[])
+    @test f(1) == (1, Base.KWDict())
     @test g(1) == (1, 2)
 end