diff --git a/llvm_passes/HipDynMem.cpp b/llvm_passes/HipDynMem.cpp
index b4c0258f7..2fd4c9442 100644
--- a/llvm_passes/HipDynMem.cpp
+++ b/llvm_passes/HipDynMem.cpp
@@ -19,7 +19,6 @@
 
 #include "HipDynMem.h"
 
-#include "llvm/ADT/SmallPtrSet.h"
 #include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/Statistic.h"
 #include "llvm/IR/Function.h"
@@ -44,7 +43,6 @@ using namespace llvm;
 #define SPIR_LOCAL_AS 3
 #define GENERIC_AS 4
 
-typedef llvm::SmallPtrSet<Function *, 16> FSet;
 typedef llvm::SetVector<Function *> OrderedFSet;
 typedef llvm::SmallVector<GlobalVariable *, 8> GVarVec;
 
@@ -91,7 +89,7 @@ class HipDynMemExternReplacePass : public ModulePass {
     }
   }
 
-  static void recursivelyFindDirectUsers(Value *V, FSet &FS) {
+  static void recursivelyFindDirectUsers(Value *V, OrderedFSet &FS) {
     for (auto U : V->users()) {
       Instruction *Inst = dyn_cast<Instruction>(U);
       if (Inst) {
@@ -440,7 +438,7 @@ CloneFunctionInto(NewF, F, VV, CloneFunctionChangeType::GlobalChanges, RI);
 
 
     for (GlobalVariable *GV : GVars) {
-      FSet DirectUserSet;
+      OrderedFSet DirectUserSet;
 
       // first, find functions that directly use the GVar. However, these may be
       // called from other functions, so we need to append the
diff --git a/tests/compiler/CMakeLists.txt b/tests/compiler/CMakeLists.txt
index 85c355d42..2d54704ee 100644
--- a/tests/compiler/CMakeLists.txt
+++ b/tests/compiler/CMakeLists.txt
@@ -202,6 +202,8 @@ add_hipcc_test(TestDeviceVirtualFunctions.hip HIPCC_OPTIONS -c)
 add_shell_test(TestDeadVTableVirtualBase.bash)
 # Program-scope variable init kernels must use their work items (#582).
 add_shell_test(TestFix582VarInitWorkItems.bash)
+# Lowering a module with many shared memory users must be reproducible (#1661).
+add_shell_test(TestFix1661DynMemOrder.bash)
 add_hipcc_test(TestHostSideHIPVectors.hip HIPCC_OPTIONS -fsyntax-only)
 add_hipcc_test(TestAlignAttr.hip HIPCC_OPTIONS -fsyntax-only)
 # Check CHIP_FAST_MATH is set for -ffast-math and preprocessor guards
diff --git a/tests/compiler/TestFix1661DynMemOrder.bash b/tests/compiler/TestFix1661DynMemOrder.bash
new file mode 100644
index 000000000..2d304836f
--- /dev/null
+++ b/tests/compiler/TestFix1661DynMemOrder.bash
@@ -0,0 +1,31 @@
+#!/bin/bash
+# Regression test for CHIP-SPV/chipStar#1661: lowering a module must not depend
+# on allocation addresses, so repeated post-link pipeline runs agree.
+set -eu
+
+OPT="@LLVM_TOOLS_BINARY_DIR@/opt"
+PLUGIN="@CMAKE_BINARY_DIR@/lib/libLLVMHipSpvPasses.so"
+OUT="@CMAKE_CURRENT_BINARY_DIR@/@TEST_NAME@.d"
+
+rm -rf "${OUT}"; mkdir -p "${OUT}"; cd "${OUT}"
+
+# More than 16 users: a small pointer set iterates in insertion order anyway.
+{
+  echo '@__smem = external addrspace(3) global [0 x i32]'
+  for i in $(seq 32); do
+    printf 'define spir_kernel void @k%d(i32 %%v) {\n' "${i}"
+    printf '  store i32 %%v, ptr addrspace(3) @__smem\n  ret void\n}\n'
+  done
+} > in.ll
+
+# Heap addresses only change between processes, so compare several runs.
+for i in $(seq 8); do
+  "${OPT}" -load-pass-plugin "${PLUGIN}" -passes=hip-post-link-passes \
+      in.ll -S -o "run${i}.ll"
+  if ! cmp -s run1.ll "run${i}.ll"; then
+    echo "FAIL: run ${i} lowered the same module differently from run 1"
+    diff run1.ll "run${i}.ll" | head -10
+    exit 1
+  fi
+done
+echo "PASSED"