diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 2043990fc79..d9ce6ebe0f0 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -4475,6 +4475,8 @@ combine_instruction(opt_ctx& ctx, aco_ptr<Instruction>& instr)
             instr = std::move(mad);
          }
          instr->definitions[0] = add_instr->definitions[0];
+         instr->definitions[0].setPrecise(add_instr->definitions[0].isPrecise() ||
+                                          mul_instr->definitions[0].isPrecise());
 
          /* mark this ssa_def to be re-checked for profitability and literals */
          ctx.mad_infos.emplace_back(std::move(add_instr), mul_instr->definitions[0].tempId());
diff --git a/src/amd/compiler/tests/test_optimizer.cpp b/src/amd/compiler/tests/test_optimizer.cpp
index 043602d8626..205c1ef65e7 100644
--- a/src/amd/compiler/tests/test_optimizer.cpp
+++ b/src/amd/compiler/tests/test_optimizer.cpp
@@ -1583,7 +1583,7 @@ BEGIN_TEST(optimize.mad_mix.fma.precision)
       Temp b16 = inputs[4];
 
       /* the optimization is precise for 32-bit on GFX9 */
-      //~gfx9! v1: %res0 = v_fma_mix_f32 lo(%a16), %b, %c
+      //~gfx9! v1: (precise)%res0 = v_fma_mix_f32 lo(%a16), %b, %c
       //~gfx10! v1: (precise)%res0_tmp = v_fma_mix_f32 lo(%a16), %b, -0
       //~gfx10! v1: %res0 = v_add_f32 %res0_tmp, %c
       //! p_unit_test 0, %res0