diff --git a/docs/ispc.rst b/docs/ispc.rst
index 9464dcde..3ac57e86 100644
--- a/docs/ispc.rst
+++ b/docs/ispc.rst
@@ -4743,13 +4743,13 @@ have a declaration like:
   };
 
 Because ``varying`` types have size that depends on the size of the gang of
-program instances, ``ispc`` prohibits any varying types from being used in
-parameters to functions with the ``export`` qualifier.  (``ispc`` also
-prohibits passing structures that themselves have varying types as members,
-etc.)  Thus, all datatypes that are shared with the application must have
-the ``uniform`` or ``soa`` rate qualifier applied to them.  (See `Use
-"Structure of Arrays" Layout When Possible`_ in the Performance Guide for
-more discussion of how to load vectors of SOA data from the application.)
+program instances, ``ispc`` has restrictrictions on using varying types in
+parameters to functions with the ``export`` qualifier.  ``ispc `` prohibits
+parameters to exported functions to have varying type unless the parameter is
+of pointer type.  (That is, ``varying float`` isn't allowed, but ``varying float * uniform``
+(uniform pointer to varying float) is permitted.)  Care must be taken 
+by the programmer to ensure that the data being accessed through any 
+pointers to varying data has the correct organization. 
 
 Similarly, ``struct`` types shared with the application can also have
 embedded pointers.
@@ -4770,6 +4770,30 @@ On the ``ispc`` side, the corresponding ``struct`` declaration is:
       float * uniform foo, * uniform bar;
   };
 
+If a pointer to a varying ``struct`` type appears in an exported function,
+the generated header file will have a definition like (for 8-wide SIMD):
+
+::
+
+  // C/C++ code
+  struct Node {
+    int count[8];
+    float pos[3][8];
+  };
+
+
+In the case of multiple target compilation, ``ispc`` will generate multiple
+header files and a "general" header file with definitions for multiple sizes.
+Any pointers to varyings in exported functions will be rewritten as ``void *``.
+At runtime, the ``ispc`` dispatch mechanism will cast these pointers to the appropriate
+types.  Programmers can
+provide C/C++ code can with a mechanism to determine the gang width used 
+at runtime by ``ispc`` by creating an exported function that simply 
+returns the value of ``programCount``.  An example of such a function
+is provided in the file ``examples/util/util.isph`` included in the ``ispc`` 
+distribution.   
+
+
 There is one subtlety related to data layout to be aware of: ``ispc``
 stores ``uniform`` short-vector types in memory with their first element at
 the machine's natural vector alignment (i.e. 16 bytes for a target that is
diff --git a/examples/perfbench/perfbench.cpp b/examples/perfbench/perfbench.cpp
index 04e72bd9..1defffe7 100644
--- a/examples/perfbench/perfbench.cpp
+++ b/examples/perfbench/perfbench.cpp
@@ -69,6 +69,7 @@ static PerfTest tests[] = {
     { xyzSumAOS, "serial", ispc::xyzSumAOSStdlib, "ispc", "AOS vector element sum (stdlib swizzle)" },
     { xyzSumAOS, "serial", ispc::xyzSumAOSNoCoalesce, "ispc", "AOS vector element sum (no coalescing)" },
     { xyzSumSOA, "serial", ispc::xyzSumSOA, "ispc", "SOA vector element sum" },
+    { xyzSumSOA, "serial", (FuncType *) ispc::xyzSumVarying, "ispc", "Varying vector element sum" },
     { ispc::gathers, "gather", ispc::loads, "vector load", "Memory reads" },
     { ispc::scatters, "scatter", ispc::stores, "vector store", "Memory writes" },
 };
diff --git a/examples/perfbench/perfbench.ispc b/examples/perfbench/perfbench.ispc
index 38fe6cee..54c3137a 100644
--- a/examples/perfbench/perfbench.ispc
+++ b/examples/perfbench/perfbench.ispc
@@ -104,6 +104,50 @@ export void xyzSumSOA(uniform float array[], uniform int count,
     result[2] = reduce_add(zsum);
 }
 
+export void xyzSumVarying(varying float array[], uniform int count,
+                          uniform float zeros[], uniform float result[]) {
+    float xsum = 0, ysum = 0, zsum = 0;
+    varying float * uniform ap = array;
+    assert(programCount <= 8);
+
+    for (uniform int i = 0; i < count/3; i += 8) {
+      if (programCount == 4) {
+        float x0 = ap[0];
+        float y0 = ap[2];
+        float z0 = ap[4];
+
+        xsum += x0;
+        ysum += y0;
+        zsum += z0;
+
+        float x1 = ap[1];
+        float y1 = ap[3];
+        float z1 = ap[5];
+
+        xsum += x1;
+        ysum += y1;
+        zsum += z1;
+
+        ap += 6;
+      }
+      else {
+        // programCount == 8
+        float x = ap[0];
+        float y = ap[1];
+        float z = ap[2];
+
+        xsum += x;
+        ysum += y;
+        zsum += z;
+
+        ap += 3;
+      }
+    }
+    result[0] = reduce_add(xsum);
+    result[1] = reduce_add(ysum);
+    result[2] = reduce_add(zsum);
+}
+
 export void gathers(uniform float array[], uniform int count,
                     uniform float zeros[], uniform float result[]) {
     float sum = 0;