rename_execute_retire: generate l2 stores earlier to make more space in units to increase throughput

crates/cpu/src/rename_execute_retire.rs

@@ -581,6 +581,15 @@ impl MOpInUnitState {
             Self::FinishedAndOrCausedCancel => None,
         }
     }
+    fn is_completed(self) -> bool {
+        match self {
+            Self::NotYetEnqueued => false,
+            Self::InputsNotReadySpeculative { .. } => false,
+            Self::InputsReady { .. } => false,
+            Self::OutputReady { .. } => false,
+            Self::FinishedAndOrCausedCancel => true,
+        }
+    }
 }
 
 impl fmt::Debug for MOpInUnitState {
@@ -1213,7 +1222,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
     fn space_available_for_unit(&self, unit_index: usize) -> usize {
         let mut retval = self.config.get().unit_max_in_flight(unit_index);
         for renamed in self.rob.renamed() {
-            if renamed.unit_index == unit_index {
+            if renamed.unit_index == unit_index && !renamed.mop_in_unit_state.is_completed() {
                 let Some(v) = NonZero::new(retval.get() - 1) else {
                     return 0;
                 };
@@ -1457,11 +1466,21 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             );
         }
         let Some(out_reg_num) = out_reg_num else {
-            println!("try_rename: out_reg_num = None");
-            if self.space_available_for_unit(self.l2_reg_file_unit_index) > 0
-                && self.rename_table.used_unit_out_reg_count(unit_index)
-                    >= (1 << self.config.get().out_reg_num_width)
-                && let Some(l2_reg_index) = self.find_free_l2_reg()
+            let l2_space_available = self.space_available_for_unit(self.l2_reg_file_unit_index);
+            let used_unit_out_reg_count = self.rename_table.used_unit_out_reg_count(unit_index);
+            let unit_out_reg_count = 1 << self.config.get().out_reg_num_width;
+            let threshold_count = (unit_out_reg_count * 3 / 4).clamp(1, unit_out_reg_count);
+            let free_l2_reg = self.find_free_l2_reg();
+            println!(
+                "try_rename: out_reg_num = None \
+                l2_space_available={l2_space_available} \
+                used_unit_out_reg_count={used_unit_out_reg_count} \
+                threshold_count={threshold_count} \
+                free_l2_reg={free_l2_reg:?}",
+            );
+            if l2_space_available > 0
+                && used_unit_out_reg_count >= threshold_count
+                && let Some(l2_reg_index) = free_l2_reg
             {
                 println!("try_rename: start L2 store");
                 let reg_to_free = LFSR31::next_sim(&mut self.lfsr) as usize

crates/cpu/tests/rename_execute_retire.rs

@@ -4331,7 +4331,7 @@ fn test_rename_execute_retire_slow_loop() {
     };
     sim.write_clock(sim.io().cd.clk, false);
     sim.write_reset(sim.io().cd.rst, true);
-    for cycle in 0..600 {
+    for cycle in 0..350 {
         sim.advance_time(SimDuration::from_nanos(500));
         println!("clock tick: {cycle}");
         sim.write_clock(sim.io().cd.clk, true);