diff --git a/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp b/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp
index cafb31e573e..4a31a2b6527 100644
--- a/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp
+++ b/src/panfrost/bifrost/valhall/test/test-insert-flow.cpp
@@ -32,13 +32,13 @@
    bi_builder *A = bit_builder(mem_ctx); \
    bi_builder *B = bit_builder(mem_ctx); \
    { \
-      bi_builder *b = A; \
+      UNUSED bi_builder *b = A; \
       A->shader->stage = MESA_SHADER_ ## shader_stage; \
       test; \
    } \
    va_insert_flow_control_nops(A->shader); \
    { \
-      bi_builder *b = B; \
+      UNUSED bi_builder *b = B; \
       B->shader->stage = MESA_SHADER_ ## shader_stage; \
       expected; \
    } \
@@ -60,6 +60,10 @@ protected:
    void *mem_ctx;
 };
 
+TEST_F(InsertFlow, PreserveEmptyShader) {
+   CASE(FRAGMENT, {}, {});
+}
+
 TEST_F(InsertFlow, TilebufferWait7) {
    CASE(FRAGMENT, {
         bi_fadd_f32_to(b, bi_register(0), bi_register(0), bi_register(0));
diff --git a/src/panfrost/bifrost/valhall/va_insert_flow.c b/src/panfrost/bifrost/valhall/va_insert_flow.c
index f7e75f2d3ea..1a3e7ed91e0 100644
--- a/src/panfrost/bifrost/valhall/va_insert_flow.c
+++ b/src/panfrost/bifrost/valhall/va_insert_flow.c
@@ -353,6 +353,19 @@ va_discard_before_block(bi_block *block)
    return false;
 }
 
+/*
+ * Test if a program is empty, in the sense of having zero instructions. Empty
+ * shaders get special handling.
+ */
+static bool
+bi_is_empty(bi_context *ctx)
+{
+   bi_foreach_instr_global(ctx, _)
+      return false;
+
+   return true;
+}
+
 /*
  * Given a program with no flow control modifiers, insert NOPs signaling the
  * required flow control. Not much optimization happens here.
@@ -360,6 +373,14 @@ va_discard_before_block(bi_block *block)
 void
 va_insert_flow_control_nops(bi_context *ctx)
 {
+   /* Special case: if a program is empty, leave it empty. In particular, do not
+    * insert NOP.end. There is special handling in the driver for skipping empty
+    * shaders for shader stage. The .end is not necessary and disrupts
+    * optimizations.
+    */
+   if (bi_is_empty(ctx))
+      return;
+
    /* First do dataflow analysis for the scoreboard. This populates I->flow with
     * a bitmap of slots to wait on.
     *