diff --git a/artiq/compiler/transforms/llvm_ir_generator.py b/artiq/compiler/transforms/llvm_ir_generator.py
index 2ea7c9e5d..a752f5c51 100644
--- a/artiq/compiler/transforms/llvm_ir_generator.py
+++ b/artiq/compiler/transforms/llvm_ir_generator.py
@@ -149,10 +149,6 @@ class LLVMIRGenerator:
             self.tbaa_tree,
             ll.Constant(lli64, 1)
         ])
-        self.tbaa_now = self.llmodule.add_metadata([
-            ll.MetaDataString(self.llmodule, "timeline position"),
-            self.tbaa_tree
-        ])
 
     def needs_sret(self, lltyp, may_be_large=True):
         if isinstance(lltyp, ll.VoidType):
@@ -1152,20 +1148,30 @@ class LLVMIRGenerator:
             return self.map(insn.operands[0])
         elif insn.op == "now_mu":
             llnow = self.llbuilder.load(self.llbuiltin("now"), name=insn.name)
-            llnow.set_metadata("tbaa", self.tbaa_now)
             return llnow
         elif insn.op == "at_mu":
             time, = insn.operands
-            return self.llbuilder.store(self.map(time), self.llbuiltin("now"))
+            lltime = self.map(time)
+            lltime_hi = self.llbuilder.trunc(self.llbuilder.lshr(lltime, ll.Constant(lli64, 32)), lli32)
+            lltime_lo = self.llbuilder.trunc(lltime, lli32)
+            llnow_hiptr = self.llbuilder.bitcast(self.llbuiltin("now"), lli32.as_pointer())
+            llnow_loptr = self.llbuilder.gep(llnow_hiptr, [self.llindex(1)])
+            llstore_hi = self.llbuilder.store_atomic(lltime_hi, llnow_hiptr, ordering="seq_cst", align=4)
+            llstore_lo = self.llbuilder.store_atomic(lltime_lo, llnow_loptr, ordering="seq_cst", align=4)
+            return llstore_lo
         elif insn.op == "delay_mu":
             interval, = insn.operands
             llnowptr = self.llbuiltin("now")
             llnow = self.llbuilder.load(llnowptr, name="now.old")
-            llnow.set_metadata("tbaa", self.tbaa_now)
             lladjusted = self.llbuilder.add(llnow, self.map(interval), name="now.new")
-            llnowstore = self.llbuilder.store(lladjusted, llnowptr)
-            llnowstore.set_metadata("tbaa", self.tbaa_now)
-            return llnowstore
+
+            lladjusted_hi = self.llbuilder.trunc(self.llbuilder.lshr(lladjusted, ll.Constant(lli64, 32)), lli32)
+            lladjusted_lo = self.llbuilder.trunc(lladjusted, lli32)
+            llnow_hiptr = self.llbuilder.bitcast(llnowptr, lli32.as_pointer())
+            llnow_loptr = self.llbuilder.gep(llnow_hiptr, [self.llindex(1)])
+            llstore_hi = self.llbuilder.store_atomic(lladjusted_hi, llnow_hiptr, ordering="seq_cst", align=4)
+            llstore_lo = self.llbuilder.store_atomic(lladjusted_lo, llnow_loptr, ordering="seq_cst", align=4)
+            return llstore_lo
         elif insn.op == "watchdog_set":
             interval, = insn.operands
             return self.llbuilder.call(self.llbuiltin("watchdog_set"), [self.map(interval)])
diff --git a/artiq/gateware/rtio/cri.py b/artiq/gateware/rtio/cri.py
index c03c2edc1..72acc879b 100644
--- a/artiq/gateware/rtio/cri.py
+++ b/artiq/gateware/rtio/cri.py
@@ -87,18 +87,15 @@ class KernelInitiator(Module, AutoCSR):
 
         # # #
 
-        now_lo_backing = Signal(32)
+        now_hi_backing = Signal(32)
         now = Signal(64, reset_less=True)
         self.sync += [
-            # TODO: fix compiler and make atomic
-            #If(self.now_lo.re, now_lo_backing.eq(self.now_lo.r)),
-            #If(self.now_hi.re, now.eq(Cat(now_lo_backing, self.now_hi.r)))
-            If(self.now_lo.re, now[:32].eq(self.now_lo.r)),
-            If(self.now_hi.re, now[32:].eq(self.now_hi.r))
+            If(self.now_hi.re, now_hi_backing.eq(self.now_hi.r)),
+            If(self.now_lo.re, now.eq(Cat(self.now_lo.r, now_hi_backing)))
         ]
         self.comb += [
-            self.now_lo.w.eq(now[:32]),
-            self.now_hi.w.eq(now[32:])
+            self.now_hi.w.eq(now[32:]),
+            self.now_lo.w.eq(now[:32])
         ]
 
         self.comb += [
diff --git a/conda/artiq-dev/meta.yaml b/conda/artiq-dev/meta.yaml
index 8212a62c1..1544cebf6 100644
--- a/conda/artiq-dev/meta.yaml
+++ b/conda/artiq-dev/meta.yaml
@@ -20,7 +20,7 @@ requirements:
     - microscope
     - binutils-or1k-linux >=2.27
     - llvm-or1k 6.0.0
-    - llvmlite-artiq 0.23.0.dev py35_4
+    - llvmlite-artiq 0.23.0.dev py35_5
     - rust-core-or1k 1.28.0 21
     - openocd 0.10.0 6
     - lit