Only run kernel state passes for kernels.

Author: maleadt

src/irgen.jl

@@ -726,11 +726,6 @@ function lower_kernel_state!(fun::LLVM.Function)
     changed = false
 
     # check if we even need a kernel state argument
-    if !job.source.kernel
-        # only kernels have had a kernel state argument added
-        # XXX: for consistency, also add the state to non-kernel compilation jobs?
-        return false
-    end
     state = kernel_state_type(job)
     if state === Nothing
         return false

src/optim.jl

@@ -182,10 +182,12 @@ function optimize!(@nospecialize(job::CompilerJob), mod::LLVM.Module)
 
         add!(pm, FunctionPass("LowerGCFrame", lower_gc_frame!))
 
-        # GC lowering is the last pass that may introduce calls to the runtime library,
-        # and thus additional uses of the kernel state.
-        add!(pm, FunctionPass("LowerKernelState", lower_kernel_state!))
-        add!(pm, ModulePass("CleanupKernelState", cleanup_kernel_state!))
+        if job.source.kernel
+            # GC lowering is the last pass that may introduce calls to the runtime library,
+            # and thus additional uses of the kernel state intrinsic.
+            add!(pm, FunctionPass("LowerKernelState", lower_kernel_state!))
+            add!(pm, ModulePass("CleanupKernelState", cleanup_kernel_state!))
+        end
 
         # remove dead uses of ptls
         aggressive_dce!(pm)
@@ -222,7 +224,9 @@ function optimize!(@nospecialize(job::CompilerJob), mod::LLVM.Module)
     ModulePassManager() do pm
         addTargetPasses!(pm, tm, triple)
 
-        dead_arg_elimination!(pm)   # parent doesn't use return value --> ret void
+        # - remove unused kernel state arguments
+        # - simplify function calls that don't use the returned value
+        dead_arg_elimination!(pm)
 
         run!(pm, mod)
     end