diff --git a/artiq/compiler/transforms/llvm_ir_generator.py b/artiq/compiler/transforms/llvm_ir_generator.py
index 61f66599c..084e5ae09 100644
--- a/artiq/compiler/transforms/llvm_ir_generator.py
+++ b/artiq/compiler/transforms/llvm_ir_generator.py
@@ -10,6 +10,7 @@ from llvmlite import ir as ll, binding as llvm
 from ...language import core as language_core
 from .. import types, builtins, ir
 from ..embedding import SpecializedFunction
+from artiq.compiler.targets import RV32GTarget
 
 
 llvoid     = ll.VoidType()
@@ -1097,20 +1098,28 @@ class LLVMIRGenerator:
             if self.target.now_pinning:
                 # Word swap now.old as CPU is little endian
                 # Most significant word is stored in lower address (see generated csr.rs)
-                llnow_raw = self.llbuilder.load(self.llbuiltin("now"), name=insn.name)
-                llnow_lo = self.llbuilder.shl(llnow_raw, ll.Constant(lli64, 32))
-                llnow_hi = self.llbuilder.lshr(llnow_raw, ll.Constant(lli64, 32))
-                return self.llbuilder.or_(llnow_lo, llnow_hi)
+                csr_offset = 2 if isinstance(self.target, RV32GTarget) else 1
+
+                llnow_hiptr = self.llbuilder.bitcast(self.llbuiltin("now"), lli32.as_pointer())
+                llnow_loptr = self.llbuilder.gep(llnow_hiptr, [self.llindex(csr_offset)])
+                llnow_hi = self.llbuilder.load(llnow_hiptr, name="now.hi")
+                llnow_lo = self.llbuilder.load(llnow_loptr, name="now.lo")
+                llzext_hi = self.llbuilder.zext(llnow_hi, lli64)
+                llshifted_hi = self.llbuilder.shl(llzext_hi, ll.Constant(lli64, 32))
+                llzext_lo = self.llbuilder.zext(llnow_lo, lli64)
+                return self.llbuilder.or_(llshifted_hi, llzext_lo)
             else:
                 return self.llbuilder.call(self.llbuiltin("now_mu"), [])
         elif insn.op == "at_mu":
             time, = insn.operands
             lltime = self.map(time)
             if self.target.now_pinning:
+                csr_offset = 2 if isinstance(self.target, RV32GTarget) else 1
+
                 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)])
+                llnow_loptr = self.llbuilder.gep(llnow_hiptr, [self.llindex(csr_offset)])
                 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
@@ -1120,20 +1129,22 @@ class LLVMIRGenerator:
             interval, = insn.operands
             llinterval = self.map(interval)
             if self.target.now_pinning:
-                llnowptr = self.llbuiltin("now")
-                llnow = self.llbuilder.load(llnowptr, name="now.old")
-
                 # Word swap now.old as CPU is little endian
                 # Most significant word is stored in lower address (see generated csr.rs)
-                llnow_lo = self.llbuilder.shl(llnow, ll.Constant(lli64, 32))
-                llnow_hi = self.llbuilder.lshr(llnow, ll.Constant(lli64, 32))
-                llnow = self.llbuilder.or_(llnow_lo, llnow_hi)
+                csr_offset = 2 if isinstance(self.target, RV32GTarget) else 1
+
+                llnow_hiptr = self.llbuilder.bitcast(self.llbuiltin("now"), lli32.as_pointer())
+                llnow_loptr = self.llbuilder.gep(llnow_hiptr, [self.llindex(csr_offset)])
+                llnow_hi = self.llbuilder.load(llnow_hiptr, name="now.hi")
+                llnow_lo = self.llbuilder.load(llnow_loptr, name="now.lo")
+                llzext_hi = self.llbuilder.zext(llnow_hi, lli64)
+                llshifted_hi = self.llbuilder.shl(llzext_hi, ll.Constant(lli64, 32))
+                llzext_lo = self.llbuilder.zext(llnow_lo, lli64)
+                llnow = self.llbuilder.or_(llshifted_hi, llzext_lo)
 
                 lladjusted = self.llbuilder.add(llnow, llinterval, name="now.new")
                 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