Merge pull request #285 from JuliaIO/kms/parsers-jl

Parse floats using Parsers.jl (fixes #284)

Author: kmsquire

Manifest.toml

@@ -25,6 +25,12 @@ uuid = "d6f4376e-aef5-505a-96c1-9c027394607a"
 [[Mmap]]
 uuid = "a63ad114-7e13-5084-954f-fe012c677804"
 
+[[Parsers]]
+deps = ["Dates", "Test"]
+git-tree-sha1 = "eaed2db080700f1013f0fc05667ecb2a082265a1"
+uuid = "69de0a69-1ddd-5017-9359-2bf0b02dc9f0"
+version = "0.3.5"
+
 [[Printf]]
 deps = ["Unicode"]
 uuid = "de0858da-6303-5e67-8744-51eddeeeb8d7"

Project.toml

@@ -5,6 +5,7 @@ uuid = "682c06a0-de6a-54ab-a142-c8b1cf79cde6"
 Dates = "ade2ca70-3891-5945-98fb-dc099432e06a"
 Distributed = "8ba89e20-285c-5b6f-9357-94700520ee1b"
 Mmap = "a63ad114-7e13-5084-954f-fe012c677804"
+Parsers = "69de0a69-1ddd-5017-9359-2bf0b02dc9f0"
 Sockets = "6462fe0b-24de-5631-8697-dd941f90decc"
 Test = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
 Unicode = "4ec0a83e-493e-50e2-b9ac-8f72acf5a8f5"

src/Parser.jl

@@ -2,6 +2,7 @@ module Parser  # JSON
 
 using Mmap
 using ..Common
+import Parsers
 
 include("pushvector.jl")
 
@@ -27,7 +28,6 @@ end
 # it is convenient to access MemoryParserState like a Vector{UInt8} to avoid copies
 Base.@propagate_inbounds Base.getindex(state::MemoryParserState, i::Int) = codeunit(state.utf8, i)
 Base.length(state::MemoryParserState) = sizeof(state.utf8)
-Base.unsafe_convert(::Type{Ptr{UInt8}}, state::MemoryParserState) = Base.unsafe_convert(Ptr{UInt8}, state.utf8)
 
 mutable struct StreamingParserState{T <: IO} <: ParserState
     io::T
@@ -317,12 +317,21 @@ end
 Parse a float from the given bytes vector, starting at `from` and ending at the
 byte before `to`. Bytes enclosed should all be ASCII characters.
 """
-function float_from_bytes(bytes, from::Int, to::Int)
-    # The ccall is not ideal (Base.tryparse would be better), but it actually
-    # makes an 2× difference to performance
-    hasvalue, val = ccall(:jl_try_substrtod, Tuple{Bool, Float64},
-            (Ptr{UInt8}, Csize_t, Csize_t), bytes, from - 1, to - from + 1)
-    hasvalue ? val : nothing
+float_from_bytes(bytes::PushVector, from::Int, to::Int) = _float_from_bytes(bytes.v, from, to)
+float_from_bytes(bytes::MemoryParserState, from::Int, to::Int) = _float_from_bytes(bytes.utf8, from, to)
+
+function _float_from_bytes(bytes, from::Int, to::Int)::Union{Float64,Nothing}
+    # Would like to use tryparse, but we want it to consume the full input,
+    # and the version in Parsers does not do this.
+
+    # return Parsers.tryparse(Float64, @view bytes.utf8[from:to])
+
+    len = to - from + 1
+    x, code, vpos, vlen, tlen = Parsers.xparse(Float64, bytes, from, to, Parsers.OPTIONS)
+    if !Parsers.ok(code) || vlen < len
+        return nothing
+    end
+    return x::Float64
 end
 
 """

src/pushvector.jl

@@ -9,7 +9,6 @@ end
 # Default length of 20 should be enough to never need to grow in most cases
 PushVector{T}() where {T} = PushVector(Vector{T}(undef, 20), 0)
 
-Base.unsafe_convert(::Type{Ptr{UInt8}}, v::PushVector) = pointer(v.v)
 Base.length(v::PushVector) = v.l
 Base.size(v::PushVector) = (v.l,)
 @inline function Base.getindex(v::PushVector, i)