diff --git a/src/dataset.jl b/src/dataset.jl
index e409129..543f832 100644
--- a/src/dataset.jl
+++ b/src/dataset.jl
@@ -22,7 +22,7 @@ cdf = CDFDataset("data.cdf")
 ```
 """
 function CDFDataset(filename)
-    fname = String(filename)
+    fname = String(filename)::String
     # `open(f, name, mode) do` form: routes through varargs splatting (`_apply_iterate`) which `juliac --trim` can't resolve.
     io = open(fname, "r")
     try
diff --git a/src/loading/variable.jl b/src/loading/variable.jl
index 809f61d..90a0786 100644
--- a/src/loading/variable.jl
+++ b/src/loading/variable.jl
@@ -65,7 +65,7 @@ function _majority_swap!(data::Ptr{UInt8}, esz::Int, rdims::Vector{Int}, nrec::I
     end
     sstrides[k+1] = record_bytes
     chunk = clamp((1 << 20) ÷ record_bytes, 1, nrec)
-    dims = [rdims; chunk]
+    dims = push!(copy(rdims), chunk)
     temp = Vector{UInt8}(undef, chunk * record_bytes)
     GC.@preserve temp for r0 in 0:chunk:(nrec-1)
         n = min(chunk, nrec - r0)
@@ -89,7 +89,7 @@ function _copy_subblock!(dst::Ptr{UInt8}, src::Ptr{UInt8}, esz::Int, rdims::Vect
         s *= rdims[i]
     end
     sstrides[k+1] = s
-    dims = [length.(ranges); nrec]
+    dims = push!(length.(ranges), nrec)
     _copy_block!(dst, src + off, esz, dims, sstrides)
     return
 end
@@ -104,7 +104,7 @@ function DiskArrays.readblock!(var::CDFVariable{T,N}, dest::AbstractArray{T}, ra
     rdims = Int[Base.front(var.dims)...]
     other_ranges = UnitRange{Int}[Int(first(r)):Int(last(r)) for r in Base.front(ranges)]
     rec_range = Int(first(ranges[end])):Int(last(ranges[end]))
-    GC.@preserve dest _readblock!(Ptr{UInt8}(pointer(dest)), sizeof(T), rdims, other_ranges, rec_range, var.vdr, var.parentdataset)
+    GC.@preserve dest _readblock!(Ptr{UInt8}(_ptr(dest)), sizeof(T), rdims, other_ranges, rec_range, var.vdr, var.parentdataset)
     is_big_endian_encoding(var) && _byte_swap!(dest)
     return dest
 end
diff --git a/src/parsing.jl b/src/parsing.jl
index cef79f8..3453843 100644
--- a/src/parsing.jl
+++ b/src/parsing.jl
@@ -98,5 +98,9 @@ function _swap_words!(p::Ptr{U}, n::Int) where {U}
     end
     return
 end
+# `pointer` goes through `cconvert`, so when `T` is not inferred any new
+# `cconvert(::Type{Ptr{X}}, ::Vector)` method (e.g. LibGit2's) invalidates the caller.
+_ptr(a::Array{T}) where {T} = Base.unsafe_convert(Ptr{T}, a)
+
 _byte_swap!(data::Array{T}) where {T} =
-    GC.@preserve data _byte_swap!(Ptr{UInt8}(pointer(data)), sizeof(data), _swap_unit(T))
+    GC.@preserve data _byte_swap!(Ptr{UInt8}(_ptr(data)), sizeof(data), _swap_unit(T))
diff --git a/src/records/aedr.jl b/src/records/aedr.jl
index 7800d27..675eb11 100644
--- a/src/records/aedr.jl
+++ b/src/records/aedr.jl
@@ -37,7 +37,7 @@ end
 
 function load_attribute_data(::Type{T}, buffer::Vector{UInt8}, pos, NumElems, needs_byte_swap) where {T}
     data = Vector{T}(undef, NumElems)
-    dst_ptr = pointer(data)
+    dst_ptr = _ptr(data)
     src_ptr = convert(Ptr{T}, pointer(buffer, pos))
     unsafe_copyto!(dst_ptr, src_ptr, NumElems)
     needs_byte_swap && _byte_swap!(data)