diff --git a/artiq/compiler/transforms/artiq_ir_generator.py b/artiq/compiler/transforms/artiq_ir_generator.py
index 518335e2f..aea58f0d5 100644
--- a/artiq/compiler/transforms/artiq_ir_generator.py
+++ b/artiq/compiler/transforms/artiq_ir_generator.py
@@ -1717,7 +1717,7 @@ class ARTIQIRGenerator(algorithm.Visitor):
             if len(node.args) == 1 and len(node.keywords) == 0:
                 arg = self.visit(node.args[0])
                 arg_mu_float = self.append(ir.Arith(ast.Div(loc=None), arg, self.ref_period))
-                arg_mu = self.append(ir.Coerce(arg_mu_float, builtins.TInt64()))
+                arg_mu = self.append(ir.Builtin("round", [arg_mu_float], builtins.TInt64()))
                 return self.append(ir.Builtin("delay_mu", [arg_mu], builtins.TNone()))
             else:
                 assert False
diff --git a/artiq/test/coredevice/test_rtio.py b/artiq/test/coredevice/test_rtio.py
index de1e6c188..3c8fb0131 100644
--- a/artiq/test/coredevice/test_rtio.py
+++ b/artiq/test/coredevice/test_rtio.py
@@ -314,6 +314,19 @@ class Handover(EnvExperiment):
         self.k("t2")
 
 
+class Rounding(EnvExperiment):
+    def build(self):
+        self.setattr_device("core")
+
+    @kernel
+    def run(self):
+        self.core.reset()
+        t1 = now_mu()
+        delay(8*us)
+        t2 = now_mu()
+        self.set_dataset("delta", t2 - t1)
+
+
 class DummyException(Exception):
     pass
 
@@ -434,6 +447,11 @@ class CoredeviceTest(ExperimentCase):
         self.assertEqual(self.dataset_mgr.get("t1") + 1234,
                          self.dataset_mgr.get("t2"))
 
+    def test_rounding(self):
+        self.execute(Rounding)
+        dt = self.dataset_mgr.get("delta")
+        self.assertEqual(dt, 8000)
+
 
 class RPCTiming(EnvExperiment):
     def build(self, repeats=100):