@@ -379,10 +379,10 @@ function nested_codegen!(
379379
380380 target = DefaultCompilerTarget ()
381381 params = PrimalCompilerParams (mode)
382- job = CompilerJob (funcspec, CompilerConfig (target, params; kernel = false ), world)
382+ job = CompilerJob (funcspec, CompilerConfig (target, params; kernel = false , libraries = true , toplevel = true , optimize = false , cleanup = false , only_entry = false , validate = false ), world)
383383
384384 GPUCompiler. prepare_job! (job)
385- otherMod, meta = GPUCompiler. emit_llvm (job; libraries = true , toplevel = true , optimize = false , cleanup = false , only_entry = false , validate = false )
385+ otherMod, meta = GPUCompiler. emit_llvm (job)
386386
387387 prepare_llvm (otherMod, job, meta)
388388
@@ -3477,9 +3477,20 @@ function GPUCompiler.codegen(
34773477 if parent_job === nothing
34783478 primal_target = DefaultCompilerTarget ()
34793479 primal_params = PrimalCompilerParams (mode)
3480+ config2 = CompilerConfig (
3481+ primal_target,
3482+ primal_params;
3483+ kernel = false ,
3484+ libraries= true ,
3485+ toplevel= toplevel,
3486+ optimize= false ,
3487+ cleanup= false ,
3488+ only_entry= false ,
3489+ validate= false
3490+ )
34803491 primal_job = CompilerJob (
34813492 primal,
3482- CompilerConfig (primal_target, primal_params; kernel = false ) ,
3493+ config2 ,
34833494 job. world,
34843495 )
34853496 else
@@ -3490,12 +3501,18 @@ function GPUCompiler.codegen(
34903501 parent_job. config. entry_abi,
34913502 parent_job. config. name,
34923503 parent_job. config. always_inline,
3504+ libraries= true ,
3505+ toplevel= toplevel,
3506+ optimize= false ,
3507+ cleanup= false ,
3508+ only_entry= false ,
3509+ validate= false ,
34933510 )
34943511 primal_job = CompilerJob (primal, config2, job. world) # TODO EnzymeInterp params, etc
34953512 end
34963513
34973514 GPUCompiler. prepare_job! (primal_job)
3498- mod, meta = GPUCompiler. emit_llvm (primal_job; libraries = true , toplevel = toplevel, optimize = false , cleanup = false , only_entry = false , validate = false )
3515+ mod, meta = GPUCompiler. emit_llvm (primal_job)
34993516 edges = Any[]
35003517 mod_to_edges[mod] = edges
35013518
0 commit comments