Link the runtime earlier, and allow re-optimizing it.

Author: maleadt

src/driver.jl

@@ -171,6 +171,11 @@ const __llvm_initialized = Ref(false)
             # target-specific libraries
             undefined_fns = LLVM.name.(decls(ir))
             @timeit_debug to "target libraries" link_libraries!(job, ir, undefined_fns)
+
+            # GPU run-time library
+            if any(fn -> fn in runtime_fns, undefined_fns)
+                @timeit_debug to "runtime library" link_library!(ir, runtime)
+            end
         end
     end
 
@@ -263,20 +268,6 @@ const __llvm_initialized = Ref(false)
             entry = functions(ir)[entry_fn]
         end
 
-        if libraries
-            # GPU run-time library
-            #
-            # we do this late for multiple reasons:
-            # - the runtime library is already optimized, so we don't want to re-optimize
-            # - if `malloc(...) = 0`, the consequent stores are reduced to a trap, which
-            #   results in e.g. every `box` function just trapping. this breaks our test
-            #   suite, which runs without malloc, but expects actual code being generated.
-            undefined_fns = LLVM.name.(decls(ir))
-            if any(fn -> fn in runtime_fns, undefined_fns)
-                @timeit_debug to "runtime library" link_library!(ir, runtime)
-            end
-        end
-
         if ccall(:jl_is_debugbuild, Cint, ()) == 1
             @timeit_debug to "verification" verify(ir)
         end

test/util.jl

@@ -35,7 +35,9 @@ end
 module TestRuntime
     # dummy methods
     signal_exception() = return
-    malloc(sz) = C_NULL
+    # HACK: if malloc returns 0 or traps, all calling functions (like jl_box_*)
+    #       get reduced to a trap, which really messes with our test suite.
+    malloc(sz) = Ptr{Cvoid}(Int(0xDEADBEEF))
     report_oom(sz) = return
     report_exception(ex) = return
     report_exception_name(ex) = return