Use OncePerTask instead of a handmade thread local cache (#123)

Author: Drvi

Project.toml

@@ -1,7 +1,7 @@
 name = "FixedPointDecimals"
 uuid = "fb4d412d-6eee-574d-9565-ede6634db7b0"
 authors = ["Fengyang Wang <[email protected]>", "Curtis Vogt <[email protected]>"]
-version = "0.6.3"
+version = "0.6.4"
 
 [deps]
 BitIntegers = "c3b6d118-76ef-56ca-8cc7-ebb389d030a1"

src/FixedPointDecimals.jl

@@ -116,12 +116,41 @@ const FD = FixedDecimal
 
 include("parse.jl")
 
-function __init__()
-    nt = isdefined(Base.Threads, :maxthreadid) ? Threads.maxthreadid() : Threads.nthreads()
-    # Buffers used in parsing when dealing with BigInts, see _divpow10! in parse.jl
-    resize!(empty!(_BIGINT_10s), nt)
-    resize!(empty!(_BIGINT_Rs), nt)
-    return
+const _BIGINT1 = BigInt(1)
+const _BIGINT2 = BigInt(2)
+const _BIGINT10 = BigInt(10)
+# Adapted from Parsers.jl, see https://github.com/JuliaData/Parsers.jl/pull/195
+@static if isdefined(Base, :OncePerTask)
+    const _get_bigint10s = OncePerTask{BigInt}(() -> BigInt(; nbits=256))
+    const _get_bigintRs = OncePerTask{BigInt}(() -> BigInt(; nbits=256))
+else
+    # N.B This code is not thread safe in the presence of thread migration
+    const _BIGINT_10s = BigInt[] # buffer for "remainders" in _divpow10!, accessed via `access_threaded`
+    const _BIGINT_Rs = BigInt[]  # buffer for "remainders" in _divpow10!, accessed via `access_threaded`
+
+    _get_bigint10s() = access_threaded(() -> (@static VERSION > v"1.5" ? BigInt(; nbits=256) : BigInt()), _BIGINT_10s)
+    _get_bigintRs() = access_threaded(() -> (@static VERSION > v"1.5" ? BigInt(; nbits=256) : BigInt()), _BIGINT_Rs)
+
+    function access_threaded(f, v::Vector)
+        tid = Threads.threadid()
+        0 < tid <= length(v) || _length_assert()
+        if @inbounds isassigned(v, tid)
+            @inbounds x = v[tid]
+        else
+            x = f()
+            @inbounds v[tid] = x
+        end
+        return x
+    end
+    @noinline _length_assert() = @assert false "0 < tid <= v"
+
+    function __init__()
+        nt = isdefined(Base.Threads, :maxthreadid) ? Threads.maxthreadid() : Threads.nthreads()
+        # Buffers used in parsing when dealing with BigInts, see _divpow10! in parse.jl
+        resize!(empty!(_BIGINT_10s), nt)
+        resize!(empty!(_BIGINT_Rs), nt)
+        return
+    end
 end
 
 # Custom widemul implementation to avoid the cost of widening to BigInt.

src/parse.jl

@@ -31,12 +31,6 @@ const OPTIONS_ROUND_THROWS = Parsers.Options(rounding=nothing)
 # TODO: a lookup table per type would be faster
 @inline _shift(n::T, decpos) where {T} = T(10)^decpos * n
 
-const _BIGINT1 = BigInt(1)
-const _BIGINT2 = BigInt(2)
-const _BIGINT10 = BigInt(10)
-const _BIGINT_10s = BigInt[] # buffer for "remainders" in _divpow10!, accessed via `Parsers.access_threaded`
-const _BIGINT_Rs = BigInt[]  # buffer for "remainders" in _divpow10!, accessed via `Parsers.access_threaded`
-
 for T in (Base.BitSigned_types..., Base.BitUnsigned_types...)
     let bytes = Tuple(codeunits(string(typemax(T))))
         # The number of digits an integer of type T can hold
@@ -74,8 +68,8 @@ function _divpow10!(x::T, code, pow, ::RoundingMode{:Throw}) where {T}
 end
 function _divpow10!(x::BigInt, code, pow, ::RoundingMode{:Nearest})
     # adapted from https://github.com/JuliaLang/julia/blob/112554e1a533cebad4cb0daa27df59636405c075/base/div.jl#L217
-    @inbounds r = Parsers.access_threaded(() -> (@static VERSION > v"1.5" ? BigInt(; nbits=256) : BigInt()), _BIGINT_Rs)  # we must not yield here!
-    @inbounds y = Parsers.access_threaded(() -> (@static VERSION > v"1.5" ? BigInt(; nbits=256) : BigInt()), _BIGINT_10s) # we must not yield here!
+    r = _get_bigintRs()  # we must not yield here!
+    y = _get_bigint10s() # we must not yield here!
     Base.GMP.MPZ.set!(y, _BIGINT10)             # y = 10
     Base.GMP.MPZ.pow_ui!(y, pow)                # y = y^pow
     Base.GMP.MPZ.tdiv_qr!(x, r, x, y)           # x, r = divrem(x, y)
@@ -87,15 +81,15 @@ function _divpow10!(x::BigInt, code, pow, ::RoundingMode{:Nearest})
     return x, code
 end
 function _divpow10!(x::BigInt, code, pow, ::RoundingMode{:ToZero})
-    @inbounds y = Parsers.access_threaded(() -> (@static VERSION > v"1.5" ? BigInt(; nbits=256) : BigInt()), _BIGINT_10s) # we must not yield here!
+    y = _get_bigint10s() # we must not yield here!
     Base.GMP.MPZ.set!(y, _BIGINT10) # y = 10
     Base.GMP.MPZ.pow_ui!(y, pow)    # y = y^pow
     Base.GMP.MPZ.tdiv_q!(x, y)      # x = div(x, y)
     return x, code
 end
 
 function _divpow10!(x::BigInt, code, pow, ::RoundingMode{:Throw})
-    @inbounds y = Parsers.access_threaded(() -> (@static VERSION > v"1.5" ? BigInt(; nbits=256) : BigInt()), _BIGINT_10s) # we must not yield here!
+    y = _get_bigint10s() # we must not yield here!
     Base.GMP.MPZ.set!(y, _BIGINT10)   # y = 10
     Base.GMP.MPZ.pow_ui!(y, pow)      # y = y^pow
     Base.GMP.MPZ.tdiv_qr!(x, y, x, y) # x, y = divrem(x, y)