diff --git a/ci.sh b/ci.sh
index 7b35d7c3357..060ac85ea7e 100755
--- a/ci.sh
+++ b/ci.sh
@@ -718,7 +718,8 @@ cmd_msan_install() {
       ["17"]="17.0.6"
       ["18"]="18.1.8"
       ["19"]="19.1.7"
-      ["20"]="20.1.2"
+      ["20"]="20.1.8"
+      ["21"]="21.1.5"
     ) 
     local llvm_tag="${CLANG_VERSION}.0.0"
     if [[ -n "${llvm_tag_by_version["${CLANG_VERSION}"]}" ]]; then
diff --git a/lib/jxl/jxl_test.cc b/lib/jxl/jxl_test.cc
index 7a8ec80a76b..e5923b24231 100644
--- a/lib/jxl/jxl_test.cc
+++ b/lib/jxl/jxl_test.cc
@@ -1427,7 +1427,7 @@ TEST(JxlTest, RoundtripAnimationPatches) {
                         21378u);
   EXPECT_EQ(ppf_out.frames.size(), t.ppf().frames.size());
   // >10 with broken patches; not all patches are detected on borders.
-  EXPECT_SLIGHTLY_BELOW(ButteraugliDistance(t.ppf(), ppf_out), 1.85);
+  EXPECT_SLIGHTLY_BELOW(ButteraugliDistance(t.ppf(), ppf_out), 1.88);
 }
 
 size_t RoundtripJpeg(const std::vector<uint8_t>& jpeg_in, ThreadPool* pool) {
diff --git a/lib/jxl/render_pipeline/stage_write.cc b/lib/jxl/render_pipeline/stage_write.cc
index 2946d48f912..5856613a7d0 100644
--- a/lib/jxl/render_pipeline/stage_write.cc
+++ b/lib/jxl/render_pipeline/stage_write.cc
@@ -50,6 +50,8 @@ using hwy::HWY_NAMESPACE::Mul;
 using hwy::HWY_NAMESPACE::NearestInt;
 using hwy::HWY_NAMESPACE::Or;
 using hwy::HWY_NAMESPACE::Rebind;
+using hwy::HWY_NAMESPACE::RebindToSigned;
+using hwy::HWY_NAMESPACE::RebindToUnsigned;
 using hwy::HWY_NAMESPACE::ShiftLeftSame;
 using hwy::HWY_NAMESPACE::ShiftRightSame;
 using hwy::HWY_NAMESPACE::VFromD;
@@ -204,6 +206,8 @@ template <typename T>
 VFromD<Rebind<T, DF>> MakeUnsigned(VFromD<DF> v, size_t x0, size_t y0,
                                    VFromD<DF> mul) {
   static_assert(std::is_unsigned<T>::value, "T must be an unsigned type");
+  using DI32 = RebindToSigned<DF>;
+  using DU32 = RebindToUnsigned<DF>;
   using DU = Rebind<T, DF>;
   v = Mul(v, mul);
   // TODO(veluca): if constexpr with C++17
@@ -217,7 +221,9 @@ VFromD<Rebind<T, DF>> MakeUnsigned(VFromD<DF> v, size_t x0, size_t y0,
     v = Add(v, dither);
   }
   v = Clamp(Zero(DF()), v, mul);
-  return DemoteTo(DU(), NearestInt(v));
+  VFromD<DI32> ni = NearestInt(v);
+  VFromD<DU32> nu = BitCast(DU32(), ni);
+  return DemoteTo(DU(), nu);
 }
 
 class WriteToOutputStage : public RenderPipelineStage {