diff --git a/src/nouveau/mme/mme_fermi_builder.c b/src/nouveau/mme/mme_fermi_builder.c
index f5fad851bce..9e716742489 100644
--- a/src/nouveau/mme/mme_fermi_builder.c
+++ b/src/nouveau/mme/mme_fermi_builder.c
@@ -520,12 +520,14 @@ mme_fermi_start_while(struct mme_builder *b)
 
 static void
 mme_fermi_end_while_zero(struct mme_builder *b,
+                         struct mme_cf cf,
                          struct mme_value cond,
                          bool is_zero)
 {
    struct mme_fermi_builder *fb = &b->fermi;
 
-   struct mme_cf cf = mme_fermi_end_cf(b, MME_CF_TYPE_WHILE);
+   if (fb->inst_parts)
+      mme_fermi_new_inst(fb);
 
    int delta = fb->inst_count - cf.start_ip - 2;
    mme_fermi_branch(fb, mme_value_alu_reg(cond), -delta, !is_zero);
@@ -540,13 +542,15 @@ mme_fermi_end_while(struct mme_builder *b,
 {
    assert(op == MME_CMP_OP_EQ);
 
+   struct mme_cf cf = mme_fermi_end_cf(b, MME_CF_TYPE_WHILE);
+
    if (mme_is_zero(x)) {
-      mme_fermi_end_while_zero(b, y, if_true);
+      mme_fermi_end_while_zero(b, cf, y, if_true);
    } else if (mme_is_zero(y)) {
-      mme_fermi_end_while_zero(b, x, if_true);
+      mme_fermi_end_while_zero(b, cf, x, if_true);
    } else {
       struct mme_value tmp = mme_fermi_neq(b, x, y);
-      mme_fermi_end_while_zero(b, tmp, if_true);
+      mme_fermi_end_while_zero(b, cf, tmp, if_true);
       mme_free_reg(b, tmp);
    }
 }
@@ -569,7 +573,7 @@ mme_fermi_end_loop(struct mme_builder *b)
    struct mme_fermi_builder *fb = &b->fermi;
 
    mme_sub_to(b, fb->loop_counter, fb->loop_counter, mme_imm(1));
-   mme_fermi_end_while_zero(b, fb->loop_counter, false);
+   mme_fermi_end_while(b, MME_CMP_OP_EQ, false, fb->loop_counter, mme_zero());
 
    mme_free_reg(b, fb->loop_counter);
    fb->loop_counter = mme_zero();
diff --git a/src/nouveau/mme/tests/mme_builder_test.cpp b/src/nouveau/mme/tests/mme_builder_test.cpp
index 90367d0254b..dbfb9803034 100644
--- a/src/nouveau/mme/tests/mme_builder_test.cpp
+++ b/src/nouveau/mme/tests/mme_builder_test.cpp
@@ -259,3 +259,53 @@ TEST_F(mme_builder_test, merge)
       ASSERT_SIM_DATA(sim);
    }
 }
+
+TEST_F(mme_builder_test, while_ine)
+{
+   static const uint32_t cases[] = { 1, 3, 5, 8 };
+
+   for (auto sim : sims) {
+      mme_builder b;
+      mme_builder_init(&b, sim->devinfo);
+
+      mme_value inc = mme_load(&b);
+      mme_value bound = mme_load(&b);
+
+      mme_value x = mme_mov(&b, mme_zero());
+      mme_value y = mme_mov(&b, mme_zero());
+      mme_value z = mme_mov(&b, mme_zero());
+
+      mme_value i = mme_mov(&b, mme_zero());
+      mme_while(&b, ine, i, bound) {
+         mme_add_to(&b, x, x, mme_imm(1));
+         mme_add_to(&b, y, y, mme_imm(2));
+         mme_add_to(&b, z, z, mme_imm(3));
+         mme_add_to(&b, i, i, inc);
+      }
+
+      sim->mme_store_data(&b, 0, x);
+      sim->mme_store_data(&b, 1, y);
+      sim->mme_store_data(&b, 2, z);
+      sim->mme_store_data(&b, 3, i);
+
+      auto macro = mme_builder_finish_vec(&b);
+
+      for (unsigned i = 0; i < ARRAY_SIZE(cases); i++) {
+         const uint32_t inc = cases[i];
+         const uint32_t count = cases[ARRAY_SIZE(cases) - i];
+         const uint32_t bound = inc * count;
+
+         std::vector<uint32_t> params;
+         params.push_back(inc);
+         params.push_back(bound);
+
+         expected[0] = count * 1;
+         expected[1] = count * 2;
+         expected[2] = count * 3;
+         expected[3] = count * inc;
+
+         sim->run_macro(macro, params);
+         ASSERT_SIM_DATA(sim);
+      }
+   }
+}