diff --git a/src/amd/compiler/aco_instruction_selection_setup.cpp b/src/amd/compiler/aco_instruction_selection_setup.cpp
index 3aef36e91be..c23f59e30a9 100644
--- a/src/amd/compiler/aco_instruction_selection_setup.cpp
+++ b/src/amd/compiler/aco_instruction_selection_setup.cpp
@@ -884,6 +884,21 @@ setup_nir(isel_context *ctx, nir_shader *nir)
       NIR_PASS(more_algebraic, nir, nir_opt_algebraic);
    }
 
+   /* Do late algebraic optimization to turn add(a, neg(b)) back into
+    * subs, then the mandatory cleanup after algebraic.  Note that it may
+    * produce fnegs, and if so then we need to keep running to squash
+    * fneg(fneg(a)).
+    */
+   bool more_late_algebraic = true;
+   while (more_late_algebraic) {
+      more_late_algebraic = false;
+      NIR_PASS(more_late_algebraic, nir, nir_opt_algebraic_late);
+      NIR_PASS_V(nir, nir_opt_constant_folding);
+      NIR_PASS_V(nir, nir_copy_prop);
+      NIR_PASS_V(nir, nir_opt_dce);
+      NIR_PASS_V(nir, nir_opt_cse);
+   }
+
    /* cleanup passes */
    nir_lower_load_const_to_scalar(nir);
    nir_opt_shrink_load(nir);