Revert back to the standard form for ccall which supports LazyLibrary improvements

We had initially planned a breaking change design for LazyLibrary, which
this package was updated for, but later we realized a better design
which allows going back to the original version of this code. The
previous form is still very poorly supported in the compiler and was
still rather buggy internally.

Revert "Replace `__init__` with `OncePerProcess` on supported Julia versions (#316)"
This reverts commit dbcb5d2.

Revert "Make FFTW relocatable for PackageCompiler"
This reverts commit e96270a.

Author: vtjnash

Project.toml

@@ -5,6 +5,7 @@ version = "1.10.0-dev"
 [deps]
 AbstractFFTs = "621f4979-c628-5d54-868e-fcf4e3e8185c"
 FFTW_jll = "f5851436-0d7a-5f13-b9de-f02708fd171a"
+Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 LinearAlgebra = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
 MKL_jll = "856f044c-d86e-5d09-b602-aeab76dc8ba7"
 Preferences = "21216c6a-2e73-6563-6e65-726566657250"
@@ -14,6 +15,7 @@ Reexport = "189a3867-3050-52da-a836-e630ba90ab69"
 AbstractFFTs = "1.5"
 Aqua = "0.8"
 FFTW_jll = "3.3.9"
+Libdl = "<0.0.1, 1.6"
 LinearAlgebra = "<0.0.1, 1"
 MKL_jll = "2019.0.117, 2020, 2021, 2022, 2023, 2024, 2025"
 Preferences = "1.2"

src/FFTW.jl

@@ -16,48 +16,70 @@ export dct, idct, dct!, idct!, plan_dct, plan_idct, plan_dct!, plan_idct!
 
 include("providers.jl")
 
-function initialize_library_paths()
-    # If someone is trying to set the provider via the old environment variable, warn them that they
-    # should instead use `set_provider!()` instead.
+function check_env()
     if haskey(ENV, "JULIA_FFTW_PROVIDER")
         Base.depwarn("JULIA_FFTW_PROVIDER is deprecated; use FFTW.set_provider!() instead", :JULIA_FFTW_PROVIDER)
     end
+end
 
-    # Hook FFTW threads up to our partr runtime, and re-assign the
-    # libfftw3{,f} refs at runtime, since we may have relocated and
-    # changed the path to the library since the last time we precompiled.
+if VERSION >= v"1.11.0"
+# This can be be deleted once FFTW_jll is upgraded to the real lazy jll code, to get the real benefits of this mess
+mutable struct FakeLazyLibrary
+    reallibrary::Symbol
+    on_load_callback
+    @atomic h::Ptr{Cvoid}
+end
+import Libdl: LazyLibrary, dlopen
+function dlopen(lib::FakeLazyLibrary)
+    h = @atomic :monotonic lib.h
+    h != C_NULL && return h
+    @lock fftwlock begin
+        h = @atomic :monotonic lib.h
+        h != C_NULL && return h
+        h = dlopen(getglobal(FFTW, lib.reallibrary))
+        lib.on_load_callback()
+        @atomic :release lib.h = h
+    end
+    return h
+end
+function fftw_init_check()
+    check_env()
     @static if fftw_provider == "fftw"
-        libfftw3_path[] = FFTW_jll.libfftw3_path
-        libfftw3f_path[] = FFTW_jll.libfftw3f_path
         fftw_init_threads()
     end
-    @static if fftw_provider == "mkl"
-        libfftw3_path[] = MKL_jll.libmkl_rt_path
-        libfftw3f_path[] = MKL_jll.libmkl_rt_path
-    end
-    return nothing
 end
 
-if VERSION >= v"1.12.0-beta1.29"
-    const initialize_library_paths_once = OncePerProcess{Nothing}() do
-        initialize_library_paths()
-        return
-    end
-    function libfftw3()
-        initialize_library_paths_once()
-        return libfftw3_path[]
-    end
-    function libfftw3f()
-        initialize_library_paths_once()
-        return libfftw3f_path[]
-    end
+@static if fftw_provider == "fftw"
+    import FFTW_jll: libfftw3 as libfftw3_no_init,
+                     libfftw3f as libfftw3f_no_init
+elseif fftw_provider == "mkl"
+    import MKL_jll: libmkl_rt as libfftw3_no_init,
+                    libmkl_rt as libfftw3f_no_init
+end
+const libfftw3 = FakeLazyLibrary(:libfftw3_no_init, fftw_init_check, C_NULL)
+const libfftw3f = FakeLazyLibrary(:libfftw3f_no_init, fftw_init_check, C_NULL)
+
 else
-    function __init__()
-        initialize_library_paths()
+function __init__()
+    # If someone is trying to set the provider via the old environment variable, warn them that they
+    # should instead use `set_provider!()` instead.
+    check_env()
+
+    global libfftw3
+    global libfftw3f
+    # Hook FFTW threads up to our partr runtime
+    @static if fftw_provider == "fftw"
+        libfftw3 = FFTW_jll.libfftw3_path
+        libfftw3f = FFTW_jll.libfftw3f_path
+        fftw_init_threads()
+    end
+    @static if fftw_provider == "mkl"
+        libfftw3 = MKL_jll.libmkl_rt_path
+        libfftw3f = MKL_jll.libmkl_rt_path
     end
-    libfftw3() = libfftw3_path[]
-    libfftw3f() = libfftw3f_path[]
 end
+end
+
 
 # most FFTW calls other than fftw_execute should be protected by a lock to be thread-safe
 const fftwlock = ReentrantLock()

src/fft.jl

@@ -55,8 +55,11 @@ function plan_r2r end
 
 ## FFT: Implement fft by calling fftw.
 
-const version = VersionNumber(split(unsafe_string(cglobal(
-    (:fftw_version,libfftw3()), UInt8)), ['-', ' '])[2])
+# TODO: this is dangerous since it captures runtime data from the compile machine
+# n.b. write this as a function to avoid Julia bugs with the runtime cglobal implementation
+get_version() = VersionNumber(split(unsafe_string(cglobal(
+    (:fftw_version,libfftw3_no_init), UInt8)), ['-', ' '])[2])
+const version = get_version()
 
 ## Direction of FFT
 
@@ -141,32 +144,32 @@ alignment_of(A::FakeArray) = Int32(0)
 @exclusive function export_wisdom(fname::AbstractString)
     f = ccall(:fopen, Ptr{Cvoid}, (Cstring,Cstring), fname, :w)
     systemerror("could not open wisdom file $fname for writing", f == C_NULL)
-    ccall((:fftw_export_wisdom_to_file,libfftw3()), Cvoid, (Ptr{Cvoid},), f)
+    ccall((:fftw_export_wisdom_to_file,libfftw3), Cvoid, (Ptr{Cvoid},), f)
     ccall(:fputs, Int32, (Ptr{UInt8},Ptr{Cvoid}), " "^256, f) # no NUL, hence no Cstring
-    ccall((:fftwf_export_wisdom_to_file,libfftw3f()), Cvoid, (Ptr{Cvoid},), f)
+    ccall((:fftwf_export_wisdom_to_file,libfftw3f), Cvoid, (Ptr{Cvoid},), f)
     ccall(:fclose, Cvoid, (Ptr{Cvoid},), f)
 end
 
 @exclusive function import_wisdom(fname::AbstractString)
     f = ccall(:fopen, Ptr{Cvoid}, (Cstring,Cstring), fname, :r)
     systemerror("could not open wisdom file $fname for reading", f == C_NULL)
-    if ccall((:fftw_import_wisdom_from_file,libfftw3()),Int32,(Ptr{Cvoid},),f)==0||
-       ccall((:fftwf_import_wisdom_from_file,libfftw3f()),Int32,(Ptr{Cvoid},),f)==0
+    if ccall((:fftw_import_wisdom_from_file,libfftw3),Int32,(Ptr{Cvoid},),f)==0||
+       ccall((:fftwf_import_wisdom_from_file,libfftw3f),Int32,(Ptr{Cvoid},),f)==0
         error("failed to import wisdom from $fname")
     end
     ccall(:fclose, Cvoid, (Ptr{Cvoid},), f)
 end
 
 @exclusive function import_system_wisdom()
-    if ccall((:fftw_import_system_wisdom,libfftw3()), Int32, ()) == 0 ||
-       ccall((:fftwf_import_system_wisdom,libfftw3f()), Int32, ()) == 0
+    if ccall((:fftw_import_system_wisdom,libfftw3), Int32, ()) == 0 ||
+       ccall((:fftwf_import_system_wisdom,libfftw3f), Int32, ()) == 0
         error("failed to import system wisdom")
     end
 end
 
 @exclusive function forget_wisdom()
-    ccall((:fftw_forget_wisdom,libfftw3()), Cvoid, ())
-    ccall((:fftwf_forget_wisdom,libfftw3f()), Cvoid, ())
+    ccall((:fftw_forget_wisdom,libfftw3), Cvoid, ())
+    ccall((:fftwf_forget_wisdom,libfftw3f), Cvoid, ())
 end
 
 # Threads
@@ -176,15 +179,15 @@ function _set_num_threads(num_threads::Integer)
     @static if fftw_provider == "mkl"
         _last_num_threads[] = num_threads
     end
-    ccall((:fftw_plan_with_nthreads,libfftw3()), Cvoid, (Int32,), num_threads)
-    ccall((:fftwf_plan_with_nthreads,libfftw3f()), Cvoid, (Int32,), num_threads)
+    ccall((:fftw_plan_with_nthreads,libfftw3), Cvoid, (Int32,), num_threads)
+    ccall((:fftwf_plan_with_nthreads,libfftw3f), Cvoid, (Int32,), num_threads)
 end
 
 @exclusive set_num_threads(num_threads::Integer) = _set_num_threads(num_threads)
 
 function get_num_threads()
     @static if fftw_provider == "fftw"
-        ccall((:fftw_planner_nthreads,libfftw3()), Cint, ())
+        ccall((:fftw_planner_nthreads,libfftw3), Cint, ())
     else
         _last_num_threads[]
     end
@@ -211,9 +214,9 @@ const NO_TIMELIMIT = -1.0 # from fftw3.h
 
 # only call these when fftwlock is held:
 unsafe_set_timelimit(precision::fftwTypeDouble,seconds) =
-    ccall((:fftw_set_timelimit,libfftw3()), Cvoid, (Float64,), seconds)
+    ccall((:fftw_set_timelimit,libfftw3), Cvoid, (Float64,), seconds)
 unsafe_set_timelimit(precision::fftwTypeSingle,seconds) =
-    ccall((:fftwf_set_timelimit,libfftw3f()), Cvoid, (Float64,), seconds)
+    ccall((:fftwf_set_timelimit,libfftw3f), Cvoid, (Float64,), seconds)
 @exclusive set_timelimit(precision, seconds) = unsafe_set_timelimit(precision, seconds)
 
 # Array alignment mod 16:
@@ -234,9 +237,9 @@ unsafe_set_timelimit(precision::fftwTypeSingle,seconds) =
         convert(Int32, convert(Int64, pointer(A)) % 16)
 else
     alignment_of(A::StridedArray{T}) where {T<:fftwDouble} =
-        ccall((:fftw_alignment_of, libfftw3()), Int32, (Ptr{T},), A)
+        ccall((:fftw_alignment_of, libfftw3), Int32, (Ptr{T},), A)
     alignment_of(A::StridedArray{T}) where {T<:fftwSingle} =
-        ccall((:fftwf_alignment_of, libfftw3f()), Int32, (Ptr{T},), A)
+        ccall((:fftwf_alignment_of, libfftw3f), Int32, (Ptr{T},), A)
 end
 
 # FFTWPlan (low-level)
@@ -320,9 +323,9 @@ unsafe_convert(::Type{PlanPtr}, p::FFTWPlan) = p.plan
 
 # these functions should only be called while the fftwlock is held
 unsafe_destroy_plan(@nospecialize(plan::FFTWPlan{<:fftwDouble})) =
-    ccall((:fftw_destroy_plan,libfftw3()), Cvoid, (PlanPtr,), plan)
+    ccall((:fftw_destroy_plan,libfftw3), Cvoid, (PlanPtr,), plan)
 unsafe_destroy_plan(@nospecialize(plan::FFTWPlan{<:fftwSingle})) =
-    ccall((:fftwf_destroy_plan,libfftw3f()), Cvoid, (PlanPtr,), plan)
+    ccall((:fftwf_destroy_plan,libfftw3f), Cvoid, (PlanPtr,), plan)
 
 const deferred_destroy_lock = ReentrantLock() # lock protecting the deferred_destroy_plans list
 const deferred_destroy_plans = FFTWPlan[]
@@ -383,19 +386,19 @@ end
 #################################################################################################
 
 cost(plan::FFTWPlan{<:fftwDouble}) =
-    ccall((:fftw_cost,libfftw3()), Float64, (PlanPtr,), plan)
+    ccall((:fftw_cost,libfftw3), Float64, (PlanPtr,), plan)
 cost(plan::FFTWPlan{<:fftwSingle}) =
-    ccall((:fftwf_cost,libfftw3f()), Float64, (PlanPtr,), plan)
+    ccall((:fftwf_cost,libfftw3f), Float64, (PlanPtr,), plan)
 
 @exclusive function arithmetic_ops(plan::FFTWPlan{<:fftwDouble})
     add, mul, fma = Ref(0.0), Ref(0.0), Ref(0.0)
-    ccall((:fftw_flops,libfftw3()), Cvoid,
+    ccall((:fftw_flops,libfftw3), Cvoid,
           (PlanPtr,Ref{Float64},Ref{Float64},Ref{Float64}), plan, add, mul, fma)
     return (round(Int64, add[]), round(Int64, mul[]), round(Int64, fma[]))
 end
 @exclusive function arithmetic_ops(plan::FFTWPlan{<:fftwSingle})
     add, mul, fma = Ref(0.0), Ref(0.0), Ref(0.0)
-    ccall((:fftwf_flops,libfftw3f()), Cvoid,
+    ccall((:fftwf_flops,libfftw3f), Cvoid,
           (PlanPtr,Ref{Float64},Ref{Float64},Ref{Float64}), plan, add, mul, fma)
     return (round(Int64, add[]), round(Int64, mul[]), round(Int64, fma[]))
 end
@@ -426,9 +429,9 @@ const has_sprint_plan = version >= v"3.3.4" && fftw_provider == "fftw"
 
 @static if has_sprint_plan
     sprint_plan_(plan::FFTWPlan{<:fftwDouble}) =
-        ccall((:fftw_sprint_plan,libfftw3()), Ptr{UInt8}, (PlanPtr,), plan)
+        ccall((:fftw_sprint_plan,libfftw3), Ptr{UInt8}, (PlanPtr,), plan)
     sprint_plan_(plan::FFTWPlan{<:fftwSingle}) =
-        ccall((:fftwf_sprint_plan,libfftw3f()), Ptr{UInt8}, (PlanPtr,), plan)
+        ccall((:fftwf_sprint_plan,libfftw3f), Ptr{UInt8}, (PlanPtr,), plan)
     function sprint_plan(plan::FFTWPlan)
         p = sprint_plan_(plan)
         str = unsafe_string(p)
@@ -510,49 +513,49 @@ _colmajorstrides(p) = ()
 # Execute
 
 unsafe_execute!(plan::FFTWPlan{<:fftwDouble}) =
-    ccall((:fftw_execute,libfftw3()), Cvoid, (PlanPtr,), plan)
+    ccall((:fftw_execute,libfftw3), Cvoid, (PlanPtr,), plan)
 
 unsafe_execute!(plan::FFTWPlan{<:fftwSingle}) =
-    ccall((:fftwf_execute,libfftw3f()), Cvoid, (PlanPtr,), plan)
+    ccall((:fftwf_execute,libfftw3f), Cvoid, (PlanPtr,), plan)
 
 unsafe_execute!(plan::cFFTWPlan{T},
                 X::StridedArray{T}, Y::StridedArray{T}) where {T<:fftwDouble} =
-    ccall((:fftw_execute_dft,libfftw3()), Cvoid,
+    ccall((:fftw_execute_dft,libfftw3), Cvoid,
           (PlanPtr,Ptr{T},Ptr{T}), plan, X, Y)
 
 unsafe_execute!(plan::cFFTWPlan{T},
                 X::StridedArray{T}, Y::StridedArray{T}) where {T<:fftwSingle} =
-    ccall((:fftwf_execute_dft,libfftw3f()), Cvoid,
+    ccall((:fftwf_execute_dft,libfftw3f), Cvoid,
           (PlanPtr,Ptr{T},Ptr{T}), plan, X, Y)
 
 unsafe_execute!(plan::rFFTWPlan{Float64,FORWARD},
                 X::StridedArray{Float64}, Y::StridedArray{Complex{Float64}}) =
-    ccall((:fftw_execute_dft_r2c,libfftw3()), Cvoid,
+    ccall((:fftw_execute_dft_r2c,libfftw3), Cvoid,
           (PlanPtr,Ptr{Float64},Ptr{Complex{Float64}}), plan, X, Y)
 
 unsafe_execute!(plan::rFFTWPlan{Float32,FORWARD},
                 X::StridedArray{Float32}, Y::StridedArray{Complex{Float32}}) =
-    ccall((:fftwf_execute_dft_r2c,libfftw3f()), Cvoid,
+    ccall((:fftwf_execute_dft_r2c,libfftw3f), Cvoid,
           (PlanPtr,Ptr{Float32},Ptr{Complex{Float32}}), plan, X, Y)
 
 unsafe_execute!(plan::rFFTWPlan{Complex{Float64},BACKWARD},
                 X::StridedArray{Complex{Float64}}, Y::StridedArray{Float64}) =
-    ccall((:fftw_execute_dft_c2r,libfftw3()), Cvoid,
+    ccall((:fftw_execute_dft_c2r,libfftw3), Cvoid,
           (PlanPtr,Ptr{Complex{Float64}},Ptr{Float64}), plan, X, Y)
 
 unsafe_execute!(plan::rFFTWPlan{Complex{Float32},BACKWARD},
                 X::StridedArray{Complex{Float32}}, Y::StridedArray{Float32}) =
-    ccall((:fftwf_execute_dft_c2r,libfftw3f()), Cvoid,
+    ccall((:fftwf_execute_dft_c2r,libfftw3f), Cvoid,
           (PlanPtr,Ptr{Complex{Float32}},Ptr{Float32}), plan, X, Y)
 
 unsafe_execute!(plan::r2rFFTWPlan{T},
                 X::StridedArray{T}, Y::StridedArray{T}) where {T<:fftwDouble} =
-    ccall((:fftw_execute_r2r,libfftw3()), Cvoid,
+    ccall((:fftw_execute_r2r,libfftw3), Cvoid,
           (PlanPtr,Ptr{T},Ptr{T}), plan, X, Y)
 
 unsafe_execute!(plan::r2rFFTWPlan{T},
                 X::StridedArray{T}, Y::StridedArray{T}) where {T<:fftwSingle} =
-    ccall((:fftwf_execute_r2r,libfftw3f()), Cvoid,
+    ccall((:fftwf_execute_r2r,libfftw3f), Cvoid,
           (PlanPtr,Ptr{T},Ptr{T}), plan, X, Y)
 
 # NOTE ON GC (garbage collection):
@@ -649,7 +652,7 @@ for (Tr,Tc,fftw,lib) in ((:Float64,:(Complex{Float64}),"fftw",:libfftw3),
         unsafe_set_timelimit($Tr, timelimit)
         R = isa(region, Tuple) ? region : copy(region)
         dims, howmany = dims_howmany(X, Y, size(X), R)
-        plan = ccall(($(string(fftw,"_plan_guru64_dft")),$lib()),
+        plan = ccall(($(string(fftw,"_plan_guru64_dft")),$lib),
                      PlanPtr,
                      (Int32, Ptr{Int}, Int32, Ptr{Int},
                       Ptr{$Tc}, Ptr{$Tc}, Int32, UInt32),
@@ -669,7 +672,7 @@ for (Tr,Tc,fftw,lib) in ((:Float64,:(Complex{Float64}),"fftw",:libfftw3),
         regionshft = _circshiftmin1(region) # FFTW halves last dim
         unsafe_set_timelimit($Tr, timelimit)
         dims, howmany = dims_howmany(X, Y, size(X), regionshft)
-        plan = ccall(($(string(fftw,"_plan_guru64_dft_r2c")),$lib()),
+        plan = ccall(($(string(fftw,"_plan_guru64_dft_r2c")),$lib),
                      PlanPtr,
                      (Int32, Ptr{Int}, Int32, Ptr{Int},
                       Ptr{$Tr}, Ptr{$Tc}, UInt32),
@@ -689,7 +692,7 @@ for (Tr,Tc,fftw,lib) in ((:Float64,:(Complex{Float64}),"fftw",:libfftw3),
         regionshft = _circshiftmin1(region) # FFTW halves last dim
         unsafe_set_timelimit($Tr, timelimit)
         dims, howmany = dims_howmany(X, Y, size(Y), regionshft)
-        plan = ccall(($(string(fftw,"_plan_guru64_dft_c2r")),$lib()),
+        plan = ccall(($(string(fftw,"_plan_guru64_dft_c2r")),$lib),
                      PlanPtr,
                      (Int32, Ptr{Int}, Int32, Ptr{Int},
                       Ptr{$Tc}, Ptr{$Tr}, UInt32),
@@ -711,7 +714,7 @@ for (Tr,Tc,fftw,lib) in ((:Float64,:(Complex{Float64}),"fftw",:libfftw3),
         knd = fix_kinds(region, kinds)
         unsafe_set_timelimit($Tr, timelimit)
         dims, howmany = dims_howmany(X, Y, size(X), region)
-        plan = ccall(($(string(fftw,"_plan_guru64_r2r")),$lib()),
+        plan = ccall(($(string(fftw,"_plan_guru64_r2r")),$lib),
                      PlanPtr,
                      (Int32, Ptr{Int}, Int32, Ptr{Int},
                       Ptr{$Tr}, Ptr{$Tr}, Ptr{Int32}, UInt32),
@@ -739,7 +742,7 @@ for (Tr,Tc,fftw,lib) in ((:Float64,:(Complex{Float64}),"fftw",:libfftw3),
             howmany[2:3, :] .*= 2
         end
         howmany = [howmany [2,1,1]] # append loop over real/imag parts
-        plan = ccall(($(string(fftw,"_plan_guru64_r2r")),$lib()),
+        plan = ccall(($(string(fftw,"_plan_guru64_r2r")),$lib),
                      PlanPtr,
                      (Int32, Ptr{Int}, Int32, Ptr{Int},
                       Ptr{$Tc}, Ptr{$Tc}, Ptr{Int32}, UInt32),