Add new deferred compilation mechanism.

Author: vchuravy

src/driver.jl

@@ -39,6 +39,41 @@ function JuliaContext(f; kwargs...)
 end
 
 
+## deferred compilation
+
+function var"gpuc.deferred" end
+
+# old, deprecated mechanism slated for removal once Enzyme is updated to the new intrinsic
+begin
+    # primitive mechanism for deferred compilation, for implementing CUDA dynamic parallelism.
+    # this could both be generalized (e.g. supporting actual function calls, instead of
+    # returning a function pointer), and be integrated with the nonrecursive codegen.
+    const deferred_codegen_jobs = Dict{Int, Any}()
+
+    # We make this function explicitly callable so that we can drive OrcJIT's
+    # lazy compilation from, while also enabling recursive compilation.
+    Base.@ccallable Ptr{Cvoid} function deferred_codegen(ptr::Ptr{Cvoid})
+        ptr
+    end
+
+    @generated function deferred_codegen(::Val{ft}, ::Val{tt}) where {ft,tt}
+        id = length(deferred_codegen_jobs) + 1
+        deferred_codegen_jobs[id] = (; ft, tt)
+        # don't bother looking up the method instance, as we'll do so again during codegen
+        # using the world age of the parent.
+        #
+        # this also works around an issue on <1.10, where we don't know the world age of
+        # generated functions so use the current world counter, which may be too new
+        # for the world we're compiling for.
+
+        quote
+            # TODO: add an edge to this method instance to support method redefinitions
+            ccall("extern deferred_codegen", llvmcall, Ptr{Cvoid}, (Int,), $id)
+        end
+    end
+end
+
+
 ## compiler entrypoint
 
 export compile
@@ -127,33 +162,6 @@ function codegen(output::Symbol, @nospecialize(job::CompilerJob); toplevel::Bool
     error("Unknown compilation output $output")
 end
 
-# primitive mechanism for deferred compilation, for implementing CUDA dynamic parallelism.
-# this could both be generalized (e.g. supporting actual function calls, instead of
-# returning a function pointer), and be integrated with the nonrecursive codegen.
-const deferred_codegen_jobs = Dict{Int, Any}()
-
-# We make this function explicitly callable so that we can drive OrcJIT's
-# lazy compilation from, while also enabling recursive compilation.
-Base.@ccallable Ptr{Cvoid} function deferred_codegen(ptr::Ptr{Cvoid})
-    ptr
-end
-
-@generated function deferred_codegen(::Val{ft}, ::Val{tt}) where {ft,tt}
-    id = length(deferred_codegen_jobs) + 1
-    deferred_codegen_jobs[id] = (; ft, tt)
-    # don't bother looking up the method instance, as we'll do so again during codegen
-    # using the world age of the parent.
-    #
-    # this also works around an issue on <1.10, where we don't know the world age of
-    # generated functions so use the current world counter, which may be too new
-    # for the world we're compiling for.
-
-    quote
-        # TODO: add an edge to this method instance to support method redefinitions
-        ccall("extern deferred_codegen", llvmcall, Ptr{Cvoid}, (Int,), $id)
-    end
-end
-
 const __llvm_initialized = Ref(false)
 
 @locked function emit_llvm(@nospecialize(job::CompilerJob); toplevel::Bool,
@@ -183,6 +191,77 @@ const __llvm_initialized = Ref(false)
     entry = finish_module!(job, ir, entry)
 
     # deferred code generation
+    if haskey(functions(ir), "gpuc.lookup")
+        dyn_marker = functions(ir)["gpuc.lookup"]
+
+        # gpuc.deferred is lowered to a gpuc.lookup foreigncall, so we need to extract the
+        # target method instance from the LLVM IR
+        # TODO: drive deferred compilation from the Julia IR instead
+        function find_base_object(val)
+            while true
+                if val isa ConstantExpr && (opcode(val) == LLVM.API.LLVMIntToPtr ||
+                                            opcode(val) == LLVM.API.LLVMBitCast ||
+                                            opcode(val) == LLVM.API.LLVMAddrSpaceCast)
+                    val = first(operands(val))
+                elseif val isa LLVM.IntToPtrInst ||
+                       val isa LLVM.BitCastInst ||
+                       val isa LLVM.AddrSpaceCastInst
+                    val = first(operands(val))
+                elseif val isa LLVM.LoadInst
+                    # In 1.11+ we no longer embed integer constants directly.
+                    gv = first(operands(val))
+                    if gv isa LLVM.GlobalValue
+                        val = LLVM.initializer(gv)
+                        continue
+                    end
+                    break
+                else
+                    break
+                end
+            end
+            return val
+        end
+
+        worklist = Dict{Any, Vector{LLVM.CallInst}}()
+        for use in uses(dyn_marker)
+            # decode the call
+            call = user(use)::LLVM.CallInst
+            dyn_mi_inst = find_base_object(operands(call)[1])
+            @compiler_assert isa(dyn_mi_inst, LLVM.ConstantInt) job
+            dyn_mi = Base.unsafe_pointer_to_objref(
+                convert(Ptr{Cvoid}, convert(Int, dyn_mi_inst)))
+            push!(get!(worklist, dyn_mi, LLVM.CallInst[]), call)
+        end
+
+        for dyn_mi in keys(worklist)
+            dyn_fn_name = compiled[dyn_mi].specfunc
+            dyn_fn = functions(ir)[dyn_fn_name]
+
+            # insert a pointer to the function everywhere the entry is used
+            T_ptr = convert(LLVMType, Ptr{Cvoid})
+            for call in worklist[dyn_mi]
+                @dispose builder=IRBuilder() begin
+                    position!(builder, call)
+                    fptr = if LLVM.version() >= v"17"
+                        T_ptr = LLVM.PointerType()
+                        bitcast!(builder, dyn_fn, T_ptr)
+                    elseif VERSION >= v"1.12.0-DEV.225"
+                        T_ptr = LLVM.PointerType(LLVM.Int8Type())
+                        bitcast!(builder, dyn_fn, T_ptr)
+                    else
+                        ptrtoint!(builder, dyn_fn, T_ptr)
+                    end
+                    replace_uses!(call, fptr)
+                end
+                unsafe_delete!(LLVM.parent(call), call)
+            end
+        end
+
+        # all deferred compilations should have been resolved
+        @compiler_assert isempty(uses(dyn_marker)) job
+        unsafe_delete!(ir, dyn_marker)
+    end
+    ## old, deprecated implementation
     has_deferred_jobs = toplevel && !only_entry && haskey(functions(ir), "deferred_codegen")
     jobs = Dict{CompilerJob, String}(job => entry_fn)
     if has_deferred_jobs
@@ -194,7 +273,6 @@ const __llvm_initialized = Ref(false)
             changed = false
 
             # find deferred compiler
-            # TODO: recover this information earlier, from the Julia IR
             worklist = Dict{CompilerJob, Vector{LLVM.CallInst}}()
             for use in uses(dyn_marker)
                 # decode the call

src/irgen.jl

@@ -80,6 +80,12 @@ function irgen(@nospecialize(job::CompilerJob))
     compiled[job.source] =
         (; compiled[job.source].ci, func, specfunc)
 
+    for mi in keys(compiled)
+        mi == job.source && continue
+        ci, func, specfunc = compiled[mi]
+        compiled[mi] = (; ci, func=safe_name(func), specfunc=safe_name(specfunc))
+    end
+
     # minimal required optimization
     @timeit_debug to "rewrite" begin
         if job.config.kernel && needs_byval(job)

src/jlgen.jl

@@ -1,6 +1,5 @@
 # Julia compiler integration
 
-
 ## world age lookups
 
 # `tls_world_age` should be used to look up the current world age. in most cases, this is
@@ -12,6 +11,7 @@ else
     tls_world_age() = ccall(:jl_get_tls_world_age, UInt, ())
 end
 
+
 ## looking up method instances
 
 export methodinstance, generic_methodinstance
@@ -159,6 +159,7 @@ end
 
 
 ## code instance cache
+
 const HAS_INTEGRATED_CACHE = VERSION >= v"1.11.0-DEV.1552"
 
 if !HAS_INTEGRATED_CACHE
@@ -318,7 +319,8 @@ else
     get_method_table_view(world::UInt, mt::MTType) = OverlayMethodTable(world, mt)
 end
 
-struct GPUInterpreter <: CC.AbstractInterpreter
+abstract type AbstractGPUInterpreter <: CC.AbstractInterpreter end
+struct GPUInterpreter <: AbstractGPUInterpreter
     world::UInt
     method_table::GPUMethodTableView
 
@@ -436,6 +438,112 @@ function CC.concrete_eval_eligible(interp::GPUInterpreter,
 end
 
 
+## deferred compilation
+
+struct DeferredCallInfo <: CC.CallInfo
+    rt::DataType
+    info::CC.CallInfo
+end
+
+# recognize calls to gpuc.deferred and save DeferredCallInfo metadata
+function CC.abstract_call_known(interp::AbstractGPUInterpreter, @nospecialize(f),
+                                arginfo::CC.ArgInfo, si::CC.StmtInfo, sv::CC.AbsIntState,
+                                max_methods::Int = CC.get_max_methods(interp, f, sv))
+    (; fargs, argtypes) = arginfo
+    if f === var"gpuc.deferred"
+        argvec = argtypes[2:end]
+        call = CC.abstract_call(interp, CC.ArgInfo(nothing, argvec), si, sv, max_methods)
+        callinfo = DeferredCallInfo(call.rt, call.info)
+        @static if VERSION < v"1.11.0-"
+            return CC.CallMeta(Ptr{Cvoid}, CC.Effects(), callinfo)
+        else
+            return CC.CallMeta(Ptr{Cvoid}, Union{}, CC.Effects(), callinfo)
+        end
+    end
+    return @invoke CC.abstract_call_known(interp::CC.AbstractInterpreter, f,
+        arginfo::CC.ArgInfo, si::CC.StmtInfo, sv::CC.AbsIntState,
+        max_methods::Int)
+end
+
+# during inlining, refine deferred calls to gpuc.lookup foreigncalls
+const FlagType = VERSION >= v"1.11.0-" ? UInt32 : UInt8
+function CC.handle_call!(todo::Vector{Pair{Int,Any}}, ir::CC.IRCode, idx::CC.Int,
+                         stmt::Expr, info::DeferredCallInfo, flag::FlagType,
+                         sig::CC.Signature, state::CC.InliningState)
+    minfo = info.info
+    results = minfo.results
+    if length(results.matches) != 1
+        return nothing
+    end
+    match = only(results.matches)
+
+    # lookup the target mi with correct edge tracking
+    case = CC.compileable_specialization(match, CC.Effects(), CC.InliningEdgeTracker(state),
+                                         info)
+    @assert case isa CC.InvokeCase
+    @assert stmt.head === :call
+
+    args = Any[
+        "extern gpuc.lookup",
+        Ptr{Cvoid},
+        Core.svec(Any, Any, match.spec_types.parameters[2:end]...), # Must use Any for MethodInstance or ftype
+        0,
+        QuoteNode(:llvmcall),
+        case.invoke,
+        stmt.args[2:end]...
+    ]
+    stmt.head = :foreigncall
+    stmt.args = args
+    return nothing
+end
+
+struct DeferredEdges
+    edges::Vector{MethodInstance}
+end
+
+function find_deferred_edges(ir::CC.IRCode)
+    edges = MethodInstance[]
+    # XXX: can we add this instead in handle_call?
+    for stmt in ir.stmts
+        inst = stmt[:inst]
+        inst isa Expr || continue
+        expr = inst::Expr
+        if expr.head === :foreigncall &&
+            expr.args[1] == "extern gpuc.lookup"
+            deferred_mi = expr.args[6]
+            push!(edges, deferred_mi)
+        end
+    end
+    unique!(edges)
+    return edges
+end
+
+if VERSION >= v"1.11.0-"
+function CC.ipo_dataflow_analysis!(interp::AbstractGPUInterpreter, ir::CC.IRCode,
+                                   caller::CC.InferenceResult)
+    edges = find_deferred_edges(ir)
+    if !isempty(edges)
+        CC.stack_analysis_result!(caller, DeferredEdges(edges))
+    end
+    @invoke CC.ipo_dataflow_analysis!(interp::CC.AbstractInterpreter, ir::CC.IRCode,
+                                      caller::CC.InferenceResult)
+end
+else # v1.10
+# 1.10 doesn't have stack_analysis_result or ipo_dataflow_analysis
+function CC.finish(interp::AbstractGPUInterpreter, opt::CC.OptimizationState, ir::CC.IRCode,
+                   caller::CC.InferenceResult)
+    edges = find_deferred_edges(ir)
+    if !isempty(edges)
+        # HACK: we store the deferred edges in the argescapes field, which is invalid,
+        #       but nobody should be running EA on our results.
+        caller.argescapes = DeferredEdges(edges)
+    end
+    @invoke CC.finish(interp::CC.AbstractInterpreter, opt::CC.OptimizationState,
+                      ir::CC.IRCode, caller::CC.InferenceResult)
+end
+end
+
+
 ## world view of the cache
 using Core.Compiler: WorldView
 
@@ -584,6 +692,24 @@ function compile_method_instance(@nospecialize(job::CompilerJob))
         error("Cannot compile $(job.source) for world $(job.world); method is only valid in worlds $(job.source.def.primary_world) to $(job.source.def.deleted_world)")
     end
 
+    compiled = IdDict()
+    llvm_mod, outstanding = compile_method_instance(job, compiled)
+    worklist = outstanding
+    while !isempty(worklist)
+        source = pop!(worklist)
+        haskey(compiled, source) && continue
+        job2 = CompilerJob(source, job.config)
+        @debug "Processing..." job2
+        llvm_mod2, outstanding = compile_method_instance(job2, compiled)
+        append!(worklist, outstanding)
+        @assert context(llvm_mod) == context(llvm_mod2)
+        link!(llvm_mod, llvm_mod2)
+    end
+
+    return llvm_mod, compiled
+end
+
+function compile_method_instance(@nospecialize(job::CompilerJob), compiled::IdDict{Any, Any})
     # populate the cache
     interp = get_interpreter(job)
     cache = CC.code_cache(interp)
@@ -594,7 +720,7 @@ function compile_method_instance(@nospecialize(job::CompilerJob))
 
     # create a callback to look-up function in our cache,
     # and keep track of the method instances we needed.
-    method_instances = []
+    method_instances = Any[]
     if Sys.ARCH == :x86 || Sys.ARCH == :x86_64
         function lookup_fun(mi, min_world, max_world)
             push!(method_instances, mi)
@@ -659,7 +785,6 @@ function compile_method_instance(@nospecialize(job::CompilerJob))
     end
 
     # process all compiled method instances
-    compiled = Dict()
     for mi in method_instances
         ci = ci_cache_lookup(cache, mi, job.world, job.world)
         ci === nothing && continue
@@ -696,10 +821,34 @@ function compile_method_instance(@nospecialize(job::CompilerJob))
         compiled[mi] = (; ci, func=llvm_func, specfunc=llvm_specfunc)
     end
 
+    # Collect the deferred edges
+    outstanding = Any[]
+    for mi in method_instances
+        !haskey(compiled, mi) && continue # Equivalent to ci_cache_lookup == nothing
+        ci = compiled[mi].ci
+        @static if VERSION >= v"1.11.0-"
+            edges = CC.traverse_analysis_results(ci) do @nospecialize result
+                return result isa DeferredEdges ? result : return
+            end
+        else
+            edges = ci.argescapes
+            if !(edges isa Union{Nothing, DeferredEdges})
+                edges = nothing
+            end
+        end
+        if edges !== nothing
+            for deferred_mi in (edges::DeferredEdges).edges
+                if !haskey(compiled, deferred_mi)
+                    push!(outstanding, deferred_mi)
+                end
+            end
+        end
+    end
+
     # ensure that the requested method instance was compiled
     @assert haskey(compiled, job.source)
 
-    return llvm_mod, compiled
+    return llvm_mod, outstanding
 end
 
 # partially revert JuliaLangjulia#49391