Switch to PrecompileTools.jl.

Project.toml

@@ -9,6 +9,7 @@ InteractiveUtils = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
 LLVM = "929cbde3-209d-540e-8aea-75f648917ca0"
 Libdl = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
 Logging = "56ddb016-857b-54e1-b83d-db4d58db5568"
+PrecompileTools = "aea7be01-6a6a-4083-8856-8a6e6704d82a"
 Preferences = "21216c6a-2e73-6563-6e65-726566657250"
 Scratch = "6c6a2e73-6563-6170-7368-637461726353"
 Serialization = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
@@ -22,10 +23,11 @@ InteractiveUtils = "1"
 LLVM = "8"
 Libdl = "1"
 Logging = "1"
-UUIDs = "1"
+PrecompileTools = "1"
 Preferences = "1"
 Scratch = "1"
 Serialization = "1"
 TOML = "1"
 TimerOutputs = "0.5"
+UUIDs = "1"
 julia = "1.10"

src/GPUCompiler.jl

@@ -16,6 +16,9 @@ using Preferences
 const CC = Core.Compiler
 using Core: MethodInstance, CodeInstance, CodeInfo
 
+compile_cache = nothing # set during __init__()
+const pkgver = Base.pkgversion(GPUCompiler)
+
 include("utils.jl")
 include("mangling.jl")
 
@@ -46,12 +49,6 @@ include("execution.jl")
 include("reflection.jl")
 
 include("precompile.jl")
-_precompile_()
-
-
-
-compile_cache = "" # defined in __init__()
-const pkgver = Base.pkgversion(GPUCompiler)
 
 function __init__()
     STDERR_HAS_COLOR[] = get(stderr, :color, false)

src/precompile.jl

@@ -1,65 +1,41 @@
-const __bodyfunction__ = Dict{Method,Any}()
+using PrecompileTools: @setup_workload, @compile_workload
 
-# Find keyword "body functions" (the function that contains the body
-# as written by the developer, called after all missing keyword-arguments
-# have been assigned values), in a manner that doesn't depend on
-# gensymmed names.
-# `mnokw` is the method that gets called when you invoke it without
-# supplying any keywords.
-function __lookup_kwbody__(mnokw::Method)
-    function getsym(arg)
-        isa(arg, Symbol) && return arg
-        @assert isa(arg, GlobalRef)
-        return arg.name
+@setup_workload begin
+    precompile_module = @eval module $(gensym())
+        using ..GPUCompiler
+
+        module DummyRuntime
+            # dummy methods
+            signal_exception() = return
+            malloc(sz) = C_NULL
+            report_oom(sz) = return
+            report_exception(ex) = return
+            report_exception_name(ex) = return
+            report_exception_frame(idx, func, file, line) = return
+        end
+
+        struct DummyCompilerParams <: AbstractCompilerParams end
+        const DummyCompilerJob = CompilerJob{NativeCompilerTarget, DummyCompilerParams}
+
+        GPUCompiler.runtime_module(::DummyCompilerJob) = DummyRuntime
     end
 
-    f = get(__bodyfunction__, mnokw, nothing)
-    if f === nothing
-        fmod = mnokw.module
-        # The lowered code for `mnokw` should look like
-        #   %1 = mkw(kwvalues..., #self#, args...)
-        #        return %1
-        # where `mkw` is the name of the "active" keyword body-function.
-        ast = Base.uncompressed_ast(mnokw)
-        if isa(ast, Core.CodeInfo) && length(ast.code) >= 2
-            callexpr = ast.code[end-1]
-            if isa(callexpr, Expr) && callexpr.head == :call
-                fsym = callexpr.args[1]
-                if isa(fsym, Symbol)
-                    f = getfield(fmod, fsym)
-                elseif isa(fsym, GlobalRef)
-                    if fsym.mod === Core && fsym.name === :_apply
-                        f = getfield(mnokw.module, getsym(callexpr.args[2]))
-                    elseif fsym.mod === Core && fsym.name === :_apply_iterate
-                        f = getfield(mnokw.module, getsym(callexpr.args[3]))
-                    else
-                        f = getfield(fsym.mod, fsym.name)
-                    end
-                else
-                    f = missing
-                end
-            else
-                f = missing
-            end
-        else
-            f = missing
+    kernel() = nothing
+
+    @compile_workload begin
+        source = methodinstance(typeof(kernel), Tuple{})
+        target = NativeCompilerTarget()
+        params = precompile_module.DummyCompilerParams()
+        config = CompilerConfig(target, params)
+        job = CompilerJob(source, config)
+
+        JuliaContext() do ctx
+            # XXX: on Windows, compiling the GPU runtime leaks GPU code in the native cache,
+            #      so prevent building the runtime library (see JuliaGPU/GPUCompiler.jl#601)
+            GPUCompiler.compile(:asm, job; libraries=false)
         end
-        __bodyfunction__[mnokw] = f
     end
-    return f
-end
 
-function _precompile_()
-    ccall(:jl_generating_output, Cint, ()) == 1 || return nothing
-    @assert precompile(Tuple{typeof(GPUCompiler.assign_args!),Expr,Vector{Any}})
-    @assert precompile(Tuple{typeof(GPUCompiler.lower_unreachable!),LLVM.Function})
-    @assert precompile(Tuple{typeof(GPUCompiler.lower_gc_frame!),LLVM.Function})
-    @assert precompile(Tuple{typeof(GPUCompiler.lower_throw!),LLVM.Module})
-    #@assert precompile(Tuple{typeof(GPUCompiler.split_kwargs),Tuple{},Vector{Symbol},Vararg{Vector{Symbol}, N} where N})
-    # let fbody = try __lookup_kwbody__(which(GPUCompiler.compile, (Symbol,GPUCompiler.CompilerJob,))) catch missing end
-    #     if !ismissing(fbody)
-    #         @assert precompile(fbody, (Bool,Bool,Bool,Bool,Bool,Bool,Bool,typeof(GPUCompiler.compile),Symbol,GPUCompiler.CompilerJob,))
-    #         @assert precompile(fbody, (Bool,Bool,Bool,Bool,Bool,Bool,Bool,typeof(GPUCompiler.compile),Symbol,GPUCompiler.CompilerJob,))
-    #     end
-    # end
+    # reset state that was initialized during precompilation
+    __llvm_initialized[] = false
 end

src/rtlib.jl

@@ -120,6 +120,11 @@ end
 const runtime_lock = ReentrantLock()
 
 @locked function load_runtime(@nospecialize(job::CompilerJob))
+    global compile_cache
+    if compile_cache === nothing    # during precompilation
+        return build_runtime(job)
+    end
+
     lock(runtime_lock) do
         slug = runtime_slug(job)
         if !supports_typed_pointers(context())