Don't unconditionally link and trigger rdc compilation.

Author: maleadt

src/compiler/driver.jl

@@ -122,6 +122,15 @@ function codegen(target::Symbol, job::CompilerJob;
         end
 
         kernel_fn = LLVM.name(kernel)
+
+        if libraries
+            # linking the device run-time library requires use of the CUDA linker,
+            # which in turn switches compilation to device relocatable code (-rdc) mode.
+            #
+            # even if not doing any actual calls that need -rdc (i.e., calls to the run-time
+            # library), this significantly hurts performance, so don't do it unconditionally
+            need_libcudadevrt = !isempty(decls(ir))
+        end
     end
 
     # dynamic parallelism
@@ -227,12 +236,16 @@ function codegen(target::Symbol, job::CompilerJob;
             jit_options[CUDAdrv.GENERATE_DEBUG_INFO] = true
         end
 
-        # link the CUDA device library
-        @timeit to[] "linking" begin
-            linker = CUDAdrv.CuLink(jit_options)
-            CUDAdrv.add_file!(linker, libcudadevrt, CUDAdrv.LIBRARY)
-            CUDAdrv.add_data!(linker, kernel_fn, asm)
-            image = CUDAdrv.complete(linker)
+        if libraries && need_libcudadevrt
+            # link the CUDA device library
+            @timeit to[] "linking" begin
+                linker = CUDAdrv.CuLink(jit_options)
+                CUDAdrv.add_file!(linker, libcudadevrt, CUDAdrv.LIBRARY)
+                CUDAdrv.add_data!(linker, kernel_fn, asm)
+                image = CUDAdrv.complete(linker)
+            end
+        else
+            image = asm
         end
 
         @timeit to[] "compilation" begin