some improvements to summarysize, fixes #32881 (#32886)

- 0-field mutable structs take 1 word
- take uniqueness into account for Strings
- include union selector bytes for Arrays

base/atomics.jl

@@ -334,7 +334,7 @@ inttype(::Type{Float32}) = Int32
 inttype(::Type{Float64}) = Int64
 
 
-gc_alignment(::Type{T}) where {T} = ccall(:jl_alignment, Cint, (Csize_t,), sizeof(T))
+import ..Base.gc_alignment
 
 # All atomic operations have acquire and/or release semantics, depending on
 # whether the load or store values. Most of the time, this is what one wants

base/reflection.jl

@@ -337,6 +337,9 @@ function datatype_alignment(dt::DataType)
     return Int(alignment & 0x1FF)
 end
 
+gc_alignment(sz::Integer) = Int(ccall(:jl_alignment, Cint, (Csize_t,), sz))
+gc_alignment(T::Type) = gc_alignment(Core.sizeof(T))
+
 """
     Base.datatype_haspadding(dt::DataType) -> Bool
 

base/summarysize.jl

@@ -74,14 +74,25 @@ end
     end
     if isa(obj, UnionAll) || isa(obj, Union)
         # black-list of items that don't have a Core.sizeof
-        return 2 * sizeof(Int)
+        sz = 2 * sizeof(Int)
+    else
+        sz = Core.sizeof(obj)
     end
-    return Core.sizeof(obj)
+    if sz == 0
+        # 0-field mutable structs are not unique
+        return gc_alignment(0)
+    end
+    return sz
 end
 
 (::SummarySize)(obj::Symbol) = 0
 (::SummarySize)(obj::SummarySize) = 0
-(::SummarySize)(obj::String) = Core.sizeof(Int) + Core.sizeof(obj)
+
+function (ss::SummarySize)(obj::String)
+    key = ccall(:jl_value_ptr, Ptr{Cvoid}, (Any,), obj)
+    haskey(ss.seen, key) ? (return 0) : (ss.seen[key] = true)
+    return Core.sizeof(Int) + Core.sizeof(obj)
+end
 
 function (ss::SummarySize)(obj::DataType)
     key = pointer_from_objref(obj)
@@ -108,7 +119,12 @@ function (ss::SummarySize)(obj::Array)
     datakey = unsafe_convert(Ptr{Cvoid}, obj)
     if !haskey(ss.seen, datakey)
         ss.seen[datakey] = true
-        size += Core.sizeof(obj)
+        dsize = Core.sizeof(obj)
+        if isbitsunion(eltype(obj))
+            # add 1 union selector byte for each element
+            dsize += length(obj)
+        end
+        size += dsize
         if !isbitstype(eltype(obj)) && !isempty(obj)
             push!(ss.frontier_x, obj)
             push!(ss.frontier_i, 1)

test/misc.jl

@@ -194,6 +194,15 @@ let A = zeros(1000), B = reshape(A, (1,1000))
     @test summarysize(A) > sizeof(A)
 end
 
+# issue #32881
+mutable struct S32881; end
+let s = "abc"
+    @test summarysize([s,s]) < summarysize(["abc","xyz"])
+end
+@test summarysize(Vector{Union{Nothing,Missing}}(undef, 16)) < summarysize(Vector{Union{Nothing,Missing}}(undef, 32))
+@test summarysize(Vector{Nothing}(undef, 16)) == summarysize(Vector{Nothing}(undef, 32))
+@test summarysize(S32881()) == sizeof(Int)
+
 # issue #13021
 let ex = try
     Main.x13021 = 0