diff --git a/src/amd/compiler/aco_optimizer.cpp b/src/amd/compiler/aco_optimizer.cpp
index 5ac2ebc7335..b0d08774379 100644
--- a/src/amd/compiler/aco_optimizer.cpp
+++ b/src/amd/compiler/aco_optimizer.cpp
@@ -3473,6 +3473,10 @@ can_use_mad_mix(opt_ctx& ctx, aco_ptr<Instruction>& instr)
    if (ctx.program->chip_class < GFX9)
       return false;
 
+   /* v_mad_mix* on GFX9 always flushes denormals for 16-bit inputs/outputs */
+   if (ctx.program->chip_class == GFX9 && ctx.fp_mode.denorm16_64)
+      return false;
+
    switch (instr->opcode) {
    case aco_opcode::v_add_f32:
    case aco_opcode::v_sub_f32:
diff --git a/src/amd/compiler/tests/test_optimizer.cpp b/src/amd/compiler/tests/test_optimizer.cpp
index 3f1da2dbd9a..f4ff665143f 100644
--- a/src/amd/compiler/tests/test_optimizer.cpp
+++ b/src/amd/compiler/tests/test_optimizer.cpp
@@ -1164,6 +1164,8 @@ BEGIN_TEST(optimize.mad_mix.input_conv.basic)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1197,6 +1199,8 @@ BEGIN_TEST(optimize.mad_mix.input_conv.precision)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1248,6 +1252,8 @@ BEGIN_TEST(optimize.mad_mix.input_conv.modifiers)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1341,6 +1347,8 @@ BEGIN_TEST(optimize.mad_mix.output_conv.basic)
       if (!setup_cs("v1 v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp c = inputs[2];
@@ -1381,6 +1389,8 @@ BEGIN_TEST(optimize.mad_mix.output_conv.precision)
       if (!setup_cs("v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a16 = inputs[0];
 
       //! v2b: %res0_tmp = v_mul_f16 %a16, %a16
@@ -1403,6 +1413,8 @@ BEGIN_TEST(optimize.mad_mix.output_conv.modifiers)
       if (!setup_cs("v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp a16 = inputs[2];
@@ -1450,6 +1462,8 @@ BEGIN_TEST(optimize.mad_mix.fma.basic)
       if (!setup_cs("v1 v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp c = inputs[2];
@@ -1502,6 +1516,8 @@ BEGIN_TEST(optimize.mad_mix.fma.precision)
       if (!setup_cs("v1 v1 v1 v2b v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp b = inputs[1];
       Temp c = inputs[2];
@@ -1562,6 +1578,8 @@ BEGIN_TEST(optimize.mad_mix.clamp)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];
 
@@ -1587,6 +1605,8 @@ BEGIN_TEST(optimize.mad_mix.cast)
       if (!setup_cs("v1 v2b", (chip_class)i))
          continue;
 
+      program->blocks[0].fp_mode.denorm16_64 = fp_denorm_flush;
+
       Temp a = inputs[0];
       Temp a16 = inputs[1];