diff --git a/src/nouveau/compiler/nak/from_nir.rs b/src/nouveau/compiler/nak/from_nir.rs
index 11e5533689a..918d527ba67 100644
--- a/src/nouveau/compiler/nak/from_nir.rs
+++ b/src/nouveau/compiler/nak/from_nir.rs
@@ -1959,6 +1959,13 @@ impl<'a> ShaderFromNir<'a> {
                 di += 1;
             }
         }
+
+        if self.sm.sm() < 50 {
+            // TODO: texbar should be created by calc_instr_deps() and
+            // should be less conservative than textures_left=0.
+            // See the old pass: NVC0LegalizePostRA::insertTextureBarriers
+            b.push_op(OpTexDepBar { textures_left: 0 });
+        }
         self.set_ssa(tex.def.as_def(), nir_dst);
     }
 
diff --git a/src/nouveau/compiler/nak/ir.rs b/src/nouveau/compiler/nak/ir.rs
index ef6378fa3f2..7befadd74f2 100644
--- a/src/nouveau/compiler/nak/ir.rs
+++ b/src/nouveau/compiler/nak/ir.rs
@@ -5816,6 +5816,25 @@ impl DisplayOp for OpBar {
 }
 impl_display_for_op!(OpBar);
 
+/// Instruction only used on Kepler(A|B).
+/// Kepler has explicit dependency tracking for texture loads.
+/// When a texture load is executed, it is put on some kind of FIFO queue
+/// for later execution.
+/// Before the results of a texture are used we need to wait on the queue,
+/// texdepbar waits until the queue has at most `textures_left` elements.
+#[repr(C)]
+#[derive(SrcsAsSlice, DstsAsSlice)]
+pub struct OpTexDepBar {
+    pub textures_left: i8,
+}
+
+impl DisplayOp for OpTexDepBar {
+    fn fmt_op(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "texdepbar {}", self.textures_left)
+    }
+}
+impl_display_for_op!(OpTexDepBar);
+
 #[repr(C)]
 #[derive(SrcsAsSlice, DstsAsSlice)]
 pub struct OpCS2R {
@@ -6535,6 +6554,7 @@ pub enum Op {
     Exit(OpExit),
     WarpSync(OpWarpSync),
     Bar(OpBar),
+    TexDepBar(OpTexDepBar),
     CS2R(OpCS2R),
     Isberd(OpIsberd),
     Kill(OpKill),
@@ -6699,6 +6719,7 @@ impl Op {
 
             // Miscellaneous ops
             Op::Bar(_)
+            | Op::TexDepBar(_)
             | Op::CS2R(_)
             | Op::Isberd(_)
             | Op::Kill(_)
@@ -7089,6 +7110,7 @@ impl Instr {
             | Op::Exit(_)
             | Op::WarpSync(_)
             | Op::Bar(_)
+            | Op::TexDepBar(_)
             | Op::RegOut(_)
             | Op::Out(_)
             | Op::OutFinal(_)
diff --git a/src/nouveau/compiler/nak/opt_instr_sched_common.rs b/src/nouveau/compiler/nak/opt_instr_sched_common.rs
index ef54bc7c38a..113243b99ee 100644
--- a/src/nouveau/compiler/nak/opt_instr_sched_common.rs
+++ b/src/nouveau/compiler/nak/opt_instr_sched_common.rs
@@ -193,9 +193,12 @@ pub fn side_effect_type(op: &Op) -> SideEffect {
         Op::Out(_) | Op::OutFinal(_) => SideEffect::Barrier,
 
         // Miscellaneous ops
-        Op::Bar(_) | Op::CS2R(_) | Op::Isberd(_) | Op::Kill(_) | Op::S2R(_) => {
-            SideEffect::Barrier
-        }
+        Op::Bar(_)
+        | Op::TexDepBar(_)
+        | Op::CS2R(_)
+        | Op::Isberd(_)
+        | Op::Kill(_)
+        | Op::S2R(_) => SideEffect::Barrier,
         Op::PixLd(_) | Op::Nop(_) | Op::Vote(_) => SideEffect::None,
 
         // Virtual ops
@@ -282,6 +285,7 @@ pub fn estimate_variable_latency(sm: u8, op: &Op) -> u32 {
 
         // Miscellaneous ops
         Op::Bar(_)
+        | Op::TexDepBar(_)
         | Op::CS2R(_)
         | Op::Isberd(_)
         | Op::Kill(_)