diff --git a/src/amd/compiler/aco_insert_delay_alu.cpp b/src/amd/compiler/aco_insert_delay_alu.cpp
index 1b5042839ba..27c0dee998c 100644
--- a/src/amd/compiler/aco_insert_delay_alu.cpp
+++ b/src/amd/compiler/aco_insert_delay_alu.cpp
@@ -307,6 +307,28 @@ handle_block(Program* program, Block& block, delay_ctx& ctx)
    block.instructions.swap(new_instructions);
 }
 
+void
+handle_loop_latch(Program* program, Block& block, delay_ctx& ctx)
+{
+   /* No actual loop. */
+   if (block.linear_preds.size() == 1)
+      return;
+
+   /* The loop header is also the loop latch. */
+   if (block.kind & block_kind_loop_latch)
+      return;
+
+   unsigned i = block.index;
+   while (ctx.program->blocks[i].loop_nest_depth >= block.loop_nest_depth) {
+      Block& latch_block = ctx.program->blocks[i++];
+      if (latch_block.loop_nest_depth == block.loop_nest_depth &&
+          (latch_block.kind & block_kind_loop_latch)) {
+         handle_block(program, latch_block, ctx);
+         break;
+      }
+   }
+}
+
 } /* end namespace */
 
 void
@@ -321,6 +343,12 @@ insert_delay_alu(Program* program)
       if (current.instructions.empty())
          continue;
 
+      /* Handle the loop latch block before the loop body. */
+      if (current.kind & block_kind_loop_header)
+         handle_loop_latch(program, current, ctx);
+      else if (current.kind & block_kind_loop_latch)
+         continue;
+
       handle_block(program, current, ctx);
 
       /* Reset ctx if there is a jump, assuming ALU will be done