diff --git a/stmt.cpp b/stmt.cpp
index c8293b5c..a9905096 100644
--- a/stmt.cpp
+++ b/stmt.cpp
@@ -1336,10 +1336,14 @@ ForeachStmt::EmitCode(FunctionEmitContext *ctx) const {
     llvm::BasicBlock *bbExit = ctx->CreateBasicBlock("foreach_exit");
 
     llvm::Value *oldMask = ctx->GetInternalMask();
+    llvm::Value *oldFunctionMask = ctx->GetFunctionMask();
 
     ctx->SetDebugPos(pos);
     ctx->StartScope();
 
+    ctx->SetInternalMask(LLVMMaskAllOn);
+    ctx->SetFunctionMask(LLVMMaskAllOn);
+
     // This should be caught during typechecking
     Assert(startExprs.size() == dimVariables.size() && 
            endExprs.size() == dimVariables.size());
@@ -1765,7 +1769,9 @@ ForeachStmt::EmitCode(FunctionEmitContext *ctx) const {
     ///////////////////////////////////////////////////////////////////////////
     // foreach_exit: All done.  Restore the old mask and clean up
     ctx->SetCurrentBasicBlock(bbExit);
+
     ctx->SetInternalMask(oldMask);
+    ctx->SetFunctionMask(oldFunctionMask);
 
     ctx->EndForeach();
     ctx->EndScope();
diff --git a/tests/foreach-mask-1.ispc b/tests/foreach-mask-1.ispc
new file mode 100644
index 00000000..2f462b48
--- /dev/null
+++ b/tests/foreach-mask-1.ispc
@@ -0,0 +1,21 @@
+
+export uniform int width() { return programCount; }
+
+
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    uniform float val[programCount+3];
+    for (uniform int i = 0; i < programCount+3; ++i)
+        val[i] = 0;
+
+    // make sure we reset the func mask in the foreach loop...
+    if ((int)aFOO[programIndex] & 1)
+        foreach (i = 0 ... programCount+3)
+            val[i] += aFOO[i] - 1;
+
+    RET[programIndex] = val[3+programIndex]; 
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = programIndex+3;
+}
diff --git a/tests/foreach-mask.ispc b/tests/foreach-mask.ispc
new file mode 100644
index 00000000..0d01b16a
--- /dev/null
+++ b/tests/foreach-mask.ispc
@@ -0,0 +1,25 @@
+
+export uniform int width() { return programCount; }
+
+
+// make sure we reset the func mask in the foreach loop...
+
+void update(uniform float val[], const uniform float a[]) {
+    foreach (i = 0 ... programCount+3)
+        val[i] += a[i] - 1;
+}
+
+export void f_f(uniform float RET[], uniform float aFOO[]) {
+    uniform float val[programCount+3];
+    for (uniform int i = 0; i < programCount+3; ++i)
+        val[i] = 0;
+
+    if ((int)aFOO[programIndex] & 1)
+        update(val, aFOO);
+
+    RET[programIndex] = val[3+programIndex]; 
+}
+
+export void result(uniform float RET[]) {
+    RET[programIndex] = programIndex+3;
+}