drop AbstractGPUInterpreter

Author: vchuravy

src/driver.jl

@@ -162,11 +162,10 @@ const __llvm_initialized = Ref(false)
         entry = functions(ir)[entry_fn]
     end
 
-    # finalize the current module. this needs to happen before linking deferred modules,
-    # since those modules have been finalized themselves, and we don't want to re-finalize.
+    # finalize the current module.
     entry = finish_module!(job, ir, entry)
 
-    # deferred code generation
+    # rewrite "gpuc.lookup" for deferred code generation
     run_optimization_for_deferred = false
     if haskey(functions(ir), "gpuc.lookup")
         run_optimization_for_deferred = true
@@ -230,13 +229,13 @@ const __llvm_initialized = Ref(false)
                     end
                     replace_uses!(call, fptr)
                 end
-                unsafe_delete!(LLVM.parent(call), call)
+                erase!( call)
             end
         end
 
         # all deferred compilations should have been resolved
         @compiler_assert isempty(uses(dyn_marker)) job
-        unsafe_delete!(ir, dyn_marker)
+        erase!(dyn_marker)
     end
 
     if libraries

src/irgen.jl

@@ -91,6 +91,8 @@ function irgen(@nospecialize(job::CompilerJob))
         compiled[mi] = (; ci, func=safe_name(func), specfunc=safe_name(specfunc))
     end
 
+    # TODO: Should we rewrite gpuc.lookup here?
+
     # minimal required optimization
     @timeit_debug to "rewrite" begin
         if job.config.kernel && needs_byval(job)
@@ -106,9 +108,16 @@ function irgen(@nospecialize(job::CompilerJob))
             end
         end
 
-        # internalize all functions and, but keep exported global variables.
+        # internalize all functions, but keep exported global variables.
         linkage!(entry, LLVM.API.LLVMExternalLinkage)
         preserved_gvs = String[LLVM.name(entry)]
+        for mi in keys(compiled)
+            # delay internalizing of deferred calls since
+            # gpuc.lookup is not yet rewriten.
+            mi == job.source && continue
+            _, _, specfunc = compiled[mi]
+            push!(preserved_gvs, specfunc) # this could be deleted if we rewrite gpuc.lookup earlier
+        end
         for gvar in globals(mod)
             push!(preserved_gvs, LLVM.name(gvar))
         end

src/jlgen.jl

@@ -319,8 +319,7 @@ else
     get_method_table_view(world::UInt, mt::MTType) = OverlayMethodTable(world, mt)
 end
 
-abstract type AbstractGPUInterpreter <: CC.AbstractInterpreter end
-struct GPUInterpreter <: AbstractGPUInterpreter
+struct GPUInterpreter <: CC.AbstractInterpreter
     world::UInt
     method_table::GPUMethodTableView
 
@@ -446,7 +445,7 @@ struct DeferredCallInfo <: CC.CallInfo
 end
 
 # recognize calls to gpuc.deferred and save DeferredCallInfo metadata
-function CC.abstract_call_known(interp::AbstractGPUInterpreter, @nospecialize(f),
+function CC.abstract_call_known(interp::GPUInterpreter, @nospecialize(f),
                                 arginfo::CC.ArgInfo, si::CC.StmtInfo, sv::CC.AbsIntState,
                                 max_methods::Int = CC.get_max_methods(interp, f, sv))
     (; fargs, argtypes) = arginfo
@@ -519,7 +518,7 @@ function find_deferred_edges(ir::CC.IRCode)
 end
 
 if VERSION >= v"1.11.0-"
-function CC.ipo_dataflow_analysis!(interp::AbstractGPUInterpreter, ir::CC.IRCode,
+function CC.ipo_dataflow_analysis!(interp::GPUInterpreter, ir::CC.IRCode,
                                    caller::CC.InferenceResult)
     edges = find_deferred_edges(ir)
     if !isempty(edges)
@@ -530,7 +529,7 @@ function CC.ipo_dataflow_analysis!(interp::AbstractGPUInterpreter, ir::CC.IRCode
 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,
+function CC.finish(interp::GPUInterpreter, opt::CC.OptimizationState, ir::CC.IRCode,
                    caller::CC.InferenceResult)
     edges = find_deferred_edges(ir)
     if !isempty(edges)

test/native_tests.jl

@@ -171,6 +171,8 @@ end
         # smoke test
         job, _ = Native.create_job(eval(kernel), (Int64,))
 
+        # TODO: Add a `kernel=true` test
+
         ci, rt = only(GPUCompiler.code_typed(job))
         @test rt === Ptr{Cvoid}