WIP adding new test for L2 reg file store

crates/cpu/src/next_pc.rs

@@ -20,7 +20,7 @@ use crate::{
         CpuConfig, CpuConfigFetchWidth, CpuConfigMaxFetchesInFlight, CpuConfigRobSize,
         PhantomConstCpuConfig, TwiceCpuConfigFetchWidth,
     },
-    util::array_vec::ArrayVec,
+    util::{LFSR31, array_vec::ArrayVec},
 };
 use fayalite::{
     int::{UIntInRange, UIntInRangeInclusive, UIntInRangeInclusiveType, UIntInRangeType},
@@ -2955,33 +2955,9 @@ impl BTBEntry {
     }
 }
 
-#[hdl]
-struct LFSR31 {
-    // MSB is always zero, 32 bits makes it easier to manipulate
-    state: UInt<32>,
-}
-
 impl SimValueDefault for LFSR31 {
-    #[hdl]
     fn sim_value_default(self) -> SimValue<Self> {
-        #[hdl(sim)]
+        Self::new()
-        Self { state: 1u32 }
-    }
-}
-
-impl LFSR31 {
-    fn next(this: &mut SimValue<Self>) -> u32 {
-        let state = this.state.as_int();
-        let state = if state == 0 {
-            1u32
-        } else {
-            // a maximal-length 31-bit LFSR
-            let lsb = ((state >> 30) ^ (state >> 27)) & 1;
-            let msb = (state << 1) & ((1 << 31) - 1);
-            lsb | msb
-        };
-        *this.state = state.into();
-        state
     }
 }
 
@@ -2995,7 +2971,7 @@ impl BranchTargetBuffer {
     const LOG2_SIZE: usize = 4;
     const SIZE: usize = 1 << Self::LOG2_SIZE;
     fn next_index_to_replace(this: &mut SimValue<Self>) -> usize {
-        LFSR31::next(&mut this.next_index_to_replace_lfsr) as usize % Self::SIZE
+        LFSR31::next_sim(&mut this.next_index_to_replace_lfsr) as usize % Self::SIZE
     }
 }
 

crates/cpu/src/rename_execute_retire.rs

@@ -17,7 +17,7 @@ use crate::{
     register::PRegValue,
     rename_execute_retire::to_unit_interfaces::ExecuteToUnitInterfaces,
     unit::{UnitKind, UnitMOp},
-    util::array_vec::ArrayVec,
+    util::{LFSR31, array_vec::ArrayVec},
 };
 use fayalite::{
     int::UIntInRangeInclusiveType,
@@ -227,6 +227,7 @@ impl<C: PhantomConstCpuConfig> SimValueDefault for RenameExecuteRetireDebugState
             next_pc_canceling,
             unit_canceling,
             l1_reg_file,
+            lfsr: _,
             per_insn_timeline,
         } = self;
         let empty_string = SimOnlyValue::new(String::new());
@@ -239,6 +240,7 @@ impl<C: PhantomConstCpuConfig> SimValueDefault for RenameExecuteRetireDebugState
             next_pc_canceling: zeroed(next_pc_canceling),
             unit_canceling: zeroed(unit_canceling),
             l1_reg_file: zeroed(l1_reg_file),
+            lfsr: LFSR31::new(),
             per_insn_timeline: SimValue::from_array_elements(
                 per_insn_timeline,
                 (0..per_insn_timeline.len()).map(|_| empty_string.clone()),
@@ -570,7 +572,7 @@ struct RobEntryDebugState<C: PhantomConstGet<CpuConfig>> {
     mop: MOpInstance<RenamedMOp<C>>,
     mop_in_unit_state: SimOnlyMOpInUnitState,
     is_speculative: Bool,
-    finished: HdlOption<NextPcPredictorOp<C>>,
+    output: HdlOption<NextPcPredictorOp<C>>,
     caused_cancel: HdlOption<UnitCausedCancel<C>>,
 }
 
@@ -581,7 +583,7 @@ impl<C: PhantomConstCpuConfig> SimValueDefault for RobEntryDebugState<C> {
             mop,
             mop_in_unit_state: _,
             is_speculative: _,
-            finished,
+            output,
             caused_cancel,
         } = self;
         #[hdl(sim)]
@@ -589,8 +591,8 @@ impl<C: PhantomConstCpuConfig> SimValueDefault for RobEntryDebugState<C> {
             mop: zeroed(mop),
             mop_in_unit_state: SimOnlyValue::default(),
             is_speculative: false,
-            finished: #[hdl(sim)]
+            output: #[hdl(sim)]
-            finished.HdlNone(),
+            output.HdlNone(),
             caused_cancel: #[hdl(sim)]
             caused_cancel.HdlNone(),
         }
@@ -602,7 +604,7 @@ struct RobEntry<C: PhantomConstCpuConfig> {
     mop: SimValue<MOpInstance<RenamedMOp<C>>>,
     mop_in_unit_state: MOpInUnitState,
     is_speculative: bool,
-    finished: Option<SimValue<NextPcPredictorOp<C>>>,
+    output: Option<SimValue<NextPcPredictorOp<C>>>,
     caused_cancel: Option<SimValue<UnitCausedCancel<C>>>,
 }
 
@@ -612,7 +614,7 @@ impl<C: PhantomConstCpuConfig> RobEntry<C> {
             mop,
             mop_in_unit_state: MOpInUnitState::NotYetEnqueued,
             is_speculative: true,
-            finished: None,
+            output: None,
             caused_cancel: None,
         }
     }
@@ -637,7 +639,7 @@ impl<C: PhantomConstCpuConfig> RobEntry<C> {
             mop,
             mop_in_unit_state,
             is_speculative,
-            finished,
+            output,
             caused_cancel,
         } = self;
         let ret_ty = RobEntryDebugState[config];
@@ -646,7 +648,7 @@ impl<C: PhantomConstCpuConfig> RobEntry<C> {
             mop,
             mop_in_unit_state: SimOnlyValue::new(*mop_in_unit_state),
             is_speculative,
-            finished: finished.into_sim_value_with_type(ret_ty.finished),
+            output: output.into_sim_value_with_type(ret_ty.output),
             caused_cancel: caused_cancel.into_sim_value_with_type(ret_ty.caused_cancel),
         }
     }
@@ -978,9 +980,15 @@ const SimOnlyString: SimOnlyString = SimOnlyString::TYPE;
 #[hdl(get(|c| c.rob_size.get().next_power_of_two()))]
 type PerInsnTimelineLen<C: PhantomConstGet<CpuConfig>> = DynSize;
 
+#[hdl]
+struct RenameDelayedEntry {
+    mop: MOpInstance<MOp>,
+    started_l2_store: Bool,
+}
+
 #[hdl(no_static)]
 pub struct RenameExecuteRetireDebugState<C: PhantomConstGet<CpuConfig>> {
-    rename_delayed: ArrayVec<MOpInstance<MOp>, TwiceCpuConfigFetchWidth<C>>,
+    rename_delayed: ArrayVec<RenameDelayedEntry, TwiceCpuConfigFetchWidth<C>>,
     rename_table: RenameTableDebugState<C>,
     retire_rename_table: RenameTableDebugState<C>,
     rob: ReorderBufferDebugState<C>,
@@ -990,18 +998,20 @@ pub struct RenameExecuteRetireDebugState<C: PhantomConstGet<CpuConfig>> {
         ArrayType<HdlOption<PRegValue>, CpuConfig2PowOutRegNumWidth<C>>,
         CpuConfigUnitCount<C>,
     >,
+    lfsr: LFSR31,
     per_insn_timeline: ArrayType<SimOnlyString, PerInsnTimelineLen<C>>,
 }
 
 #[derive(Debug)]
 struct RenameExecuteRetireState<C: PhantomConstCpuConfig> {
-    rename_delayed: VecDeque<SimValue<MOpInstance<MOp>>>,
+    rename_delayed: VecDeque<SimValue<RenameDelayedEntry>>,
     rename_table: RenameTable<C>,
     retire_rename_table: RenameTable<C>,
     rob: ReorderBuffer<C>,
     next_pc_canceling: Option<NextPcCancelingState>,
     unit_canceling: Box<[bool]>,
     l1_reg_file: Box<[Box<[Option<SimValue<PRegValue>>]>]>,
+    lfsr: SimValue<LFSR31>,
     l2_reg_file_unit_index: usize,
     config: C,
 }
@@ -1020,6 +1030,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                 CpuConfigUnitCount[config]
             ]
             .into_boxed_slice(),
+            lfsr: LFSR31::new(),
             l2_reg_file_unit_index: config
                 .get()
                 .units
@@ -1044,8 +1055,8 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             let masked_id = rob.mop.id.as_int() as usize & mask;
             **retval[masked_id] = fmt::from_fn(|f| {
                 f.write_str(rob.mop_in_unit_state.debug_str())?;
-                if rob.finished.is_some() {
+                if rob.output.is_some() {
-                    f.write_str("(finished)")?;
+                    f.write_str("(output)")?;
                 }
                 if rob.caused_cancel.is_some() {
                     f.write_str("(caused cancel)")?;
@@ -1079,16 +1090,17 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
         state_for_debug: Expr<RenameExecuteRetireDebugState<C>>,
     ) {
         let Self {
-            ref rename_delayed,
+            rename_delayed,
-            ref rename_table,
+            rename_table,
-            ref retire_rename_table,
+            retire_rename_table,
-            ref rob,
+            rob,
-            ref next_pc_canceling,
+            next_pc_canceling,
-            ref unit_canceling,
+            unit_canceling,
-            ref l1_reg_file,
+            l1_reg_file,
+            lfsr,
             l2_reg_file_unit_index: _,
             config: _,
-        } = *self;
+        } = self;
         sim.write(
             state_for_debug,
             #[hdl(sim)]
@@ -1109,6 +1121,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                         SimValue::from_array_elements(state_for_debug.ty().l1_reg_file.element(), v)
                     }),
                 ),
+                lfsr,
                 per_insn_timeline: self.per_insn_timeline(),
             },
         )
@@ -1133,7 +1146,10 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                         .HdlSome
                         .from_iter_sim(
                             zeroed(MOpInstance[MOp]),
-                            self.rename_delayed.iter().chain(self.rob.unrenamed()),
+                            self.rename_delayed
+                                .iter()
+                                .map(|v| &v.mop)
+                                .chain(self.rob.unrenamed()),
                         )
                         .ok()
                         .expect("known to fit"),
@@ -1242,13 +1258,33 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
     #[hdl]
     fn try_rename(
         &mut self,
-        insn: SimValue<MOpInstance<MOp>>,
+        entry: SimValue<RenameDelayedEntry>,
-    ) -> Result<(), SimValue<MOpInstance<MOp>>> {
+    ) -> Result<(), SimValue<RenameDelayedEntry>> {
+        #[hdl(sim)]
+        let RenameDelayedEntry {
+            mop: insn,
+            started_l2_store,
+        } = entry;
+        println!("try_rename: insn: {insn:?}");
         if self.rob.unrenamed_len() >= self.config.get().rob_size.get() {
-            return Err(insn);
+            println!("try_rename: unrenamed_len >= rob_size");
+            return Err(
+                #[hdl(sim)]
+                RenameDelayedEntry {
+                    mop: insn,
+                    started_l2_store,
+                },
+            );
         }
         if self.rob.renamed_len() >= self.config.get().rob_size.get() {
-            return Err(insn);
+            println!("try_rename: renamed_len >= rob_size");
+            return Err(
+                #[hdl(sim)]
+                RenameDelayedEntry {
+                    mop: insn,
+                    started_l2_store,
+                },
+            );
         }
         let unit_kind = UnitMOp::kind_sim(&insn.mop);
         #[hdl(sim)]
@@ -1323,15 +1359,30 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             );
         };
         if space_available == 0 {
-            return Err(insn);
+            println!("try_rename: space_available = 0");
+            return Err(
+                #[hdl(sim)]
+                RenameDelayedEntry {
+                    mop: insn,
+                    started_l2_store,
+                },
+            );
         }
         let Some(out_reg_num) = out_reg_num else {
-            return if self.space_available_for_unit(self.l2_reg_file_unit_index) > 0
+            println!("try_rename: out_reg_num = None");
+            return if !*started_l2_store
+                && self.space_available_for_unit(self.l2_reg_file_unit_index) > 0
                 && let Some(l2_reg_index) = self.find_free_l2_reg()
             {
                 todo!("maybe start a L2 register file store");
             } else {
-                Err(insn)
+                Err(
+                    #[hdl(sim)]
+                    RenameDelayedEntry {
+                        mop: insn,
+                        started_l2_store,
+                    },
+                )
             };
         };
         let out_reg_num_sim = UnitOutRegNum[self.config].new_sim(out_reg_num);
@@ -1408,7 +1459,14 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                 );
                 Ok(())
             } else {
-                Err(insn)
+                println!("try_rename: l2 reg file has no space and/or has no free output regs");
+                Err(
+                    #[hdl(sim)]
+                    RenameDelayedEntry {
+                        mop: insn,
+                        started_l2_store,
+                    },
+                )
             };
         }
         let mop = UnitMOp::with_transformed_move_op_sim(
@@ -1567,15 +1625,15 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             mop: _,
             mop_in_unit_state,
             is_speculative: _,
-            finished,
+            output,
             caused_cancel,
         } = rob;
-        assert!(finished.is_none());
+        assert!(output.is_none());
         assert!(caused_cancel.is_none());
         let l1_reg = &mut self.l1_reg_file[unit_index][out_reg_index];
         assert!(l1_reg.is_none());
         *l1_reg = Some(dest_value);
-        *finished = Some(predictor_op);
+        *output = Some(predictor_op);
         *mop_in_unit_state = mop_in_unit_state
             .after_output_ready()
             .expect("should be valid state for output to become ready");
@@ -1596,7 +1654,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             mop: _,
             mop_in_unit_state,
             is_speculative: _,
-            finished,
+            output,
             caused_cancel,
         } = self.rob.renamed_by_id_mut(&id);
         assert!(caused_cancel.is_none());
@@ -1616,7 +1674,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             panic!(
                 "MOp {id:?} made an invalid attempt to finish/cause a cancel:\n\
                     mop_in_unit_state={mop_in_unit_state:?}\n\
-                    finished={finished:?}\n\
+                    output={output:?}\n\
                     caused_cancel={caused_cancel:?}"
             );
         }
@@ -1635,13 +1693,20 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                 )
         }
     }
+    #[hdl]
     fn handle_from_post_decode(&mut self, insns: &[SimValue<MOpInstance<MOp>>]) {
-        if insns.is_empty() {
+        if self.is_canceling() {
+            assert!(insns.is_empty());
             return;
         }
-        assert!(!self.is_canceling());
         for insn in insns {
-            self.rename_delayed.push_back(insn.clone());
+            self.rename_delayed.push_back(
+                #[hdl(sim)]
+                RenameDelayedEntry {
+                    mop: insn,
+                    started_l2_store: false,
+                },
+            );
         }
         for _ in 0..CpuConfigFetchWidth[self.config] {
             let Some(insn) = self.rename_delayed.pop_front() else {
@@ -1666,6 +1731,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             next_pc_canceling,
             unit_canceling: _,
             l1_reg_file: _,
+            lfsr: _,
             l2_reg_file_unit_index: _,
             config: _,
         } = self;
@@ -1695,7 +1761,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                     mop: _,
                     mop_in_unit_state: MOpInUnitState::FinishedAndOrCausedCancel,
                     is_speculative: _,
-                    finished,
+                    output,
                     caused_cancel,
                 } = renamed_entry
                 {
@@ -1703,7 +1769,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                         // only the part before the cancel needs to be ready
                         break;
                     }
-                    assert!(finished.is_some());
+                    assert!(output.is_some());
                 } else {
                     // group isn't ready
                     return retval;
@@ -1740,7 +1806,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                         config: self.config,
                     };
                     for renamed in renamed_entries {
-                        let Some(finished) = &renamed.finished else {
+                        let Some(output) = &renamed.output else {
                             unreachable!();
                         };
                         #[hdl(sim)]
@@ -1748,7 +1814,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
                             call_stack_op,
                             cond_br_taken,
                             config: _,
-                        } = finished;
+                        } = output;
                         #[hdl(sim)]
                         if let CallStackOp::None = &unrenamed_op.call_stack_op {
                             unrenamed_op.call_stack_op = call_stack_op.clone();
@@ -1826,7 +1892,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             mop: _,
             mop_in_unit_state,
             is_speculative: _,
-            finished: _,
+            output: _,
             caused_cancel,
         } in renamed_entries
         {
@@ -1879,7 +1945,7 @@ impl<C: PhantomConstCpuConfig> RenameExecuteRetireState<C> {
             mop: _,
             mop_in_unit_state: MOpInUnitState::FinishedAndOrCausedCancel,
             is_speculative: _,
-            finished: _,
+            output: _,
             caused_cancel: Some(caused_cancel),
         }) = first_renamed
             && !*caused_cancel.cancel_after_retire
@@ -1989,10 +2055,10 @@ async fn rename_execute_retire_run(
                         mop: _,
                         mop_in_unit_state,
                         is_speculative: _,
-                        finished,
+                        output,
                         caused_cancel,
                     } = state.rob.renamed_by_id_mut(&enqueue.mop.id);
-                    assert!(finished.is_none());
+                    assert!(output.is_none());
                     assert!(caused_cancel.is_none());
                     *mop_in_unit_state = mop_in_unit_state
                         .after_enqueue()
@@ -2006,10 +2072,10 @@ async fn rename_execute_retire_run(
                     mop: _,
                     mop_in_unit_state,
                     is_speculative: _,
-                    finished,
+                    output,
                     caused_cancel,
                 } = state.rob.renamed_by_id_mut(&inputs_ready.mop.id);
-                assert!(finished.is_none());
+                assert!(output.is_none());
                 assert!(caused_cancel.is_none());
                 *mop_in_unit_state = mop_in_unit_state
                     .with_inputs_ready()
@@ -2024,11 +2090,9 @@ async fn rename_execute_retire_run(
                     mop: _,
                     mop_in_unit_state,
                     is_speculative: _,
-                    finished,
+                    output: _,
-                    caused_cancel,
+                    caused_cancel: _,
                 } = state.rob.renamed_by_id_mut(&is_no_longer_speculative.id);
-                assert!(finished.is_none());
-                assert!(caused_cancel.is_none());
                 *mop_in_unit_state = mop_in_unit_state
                     .without_speculative()
                     .expect("UnitMOpIsNoLongerSpeculative is known to be valid");
@@ -2043,10 +2107,9 @@ async fn rename_execute_retire_run(
                         mop: _,
                         mop_in_unit_state,
                         is_speculative: _,
-                        finished,
+                        output: _,
                         caused_cancel,
                     } = state.rob.renamed_by_id_mut(&id);
-                    assert!(finished.is_none());
                     assert!(caused_cancel.is_none());
                     *mop_in_unit_state = mop_in_unit_state
                         .with_cant_cause_cancel()

crates/cpu/src/util.rs

@@ -280,3 +280,33 @@ impl<T: Type, N: Size> Rotate<usize> for SimValue<ArrayType<T, N>> {
         retval
     }
 }
+
+#[hdl]
+pub struct LFSR31 {
+    // MSB is always zero, 32 bits makes it easier to manipulate
+    state: UInt<32>,
+}
+
+impl LFSR31 {
+    #[hdl]
+    pub fn new() -> SimValue<Self> {
+        #[hdl(sim)]
+        Self { state: 1u32 }
+    }
+}
+
+impl LFSR31 {
+    pub fn next_sim(this: &mut SimValue<Self>) -> u32 {
+        let state = this.state.as_int();
+        let state = if state == 0 {
+            1u32
+        } else {
+            // a maximal-length 31-bit LFSR
+            let lsb = ((state >> 30) ^ (state >> 27)) & 1;
+            let msb = (state << 1) & ((1 << 31) - 1);
+            lsb | msb
+        };
+        *this.state = state.into();
+        state
+    }
+}

crates/cpu/tests/rename_execute_retire.rs

@@ -8,9 +8,10 @@ use cpu::{
     },
     instruction::{
         AddSubMOp, AluBranchMOp, AluCommonMOp, BranchMOp, COMMON_MOP_SRC_LEN, CommonMOp,
-        CompareMOp, CompareMode, ConditionMode, L2RegisterFileMOp, LoadMOp, LoadStoreCommonMOp,
+        CompareMOp, CompareMode, ConditionMode, L2RegNum, L2RegisterFileMOp, LoadMOp,
-        LoadStoreConversion, LoadStoreMOp, LoadStoreWidth, MOp, MOpDestReg, MOpRegNum, MOpTrait,
+        LoadStoreCommonMOp, LoadStoreConversion, LoadStoreMOp, LoadStoreWidth, MOp, MOpDestReg,
-        MoveRegMOp, OutputIntegerMode, PRegNum, StoreMOp, UnitNum,
+        MOpRegNum, MOpTrait, MoveRegMOp, OutputIntegerMode, PRegNum, ReadL2RegMOp, StoreMOp,
+        UnitNum, WriteL2RegMOp,
     },
     next_pc::CallStackOp,
     register::{PRegFlags, PRegFlagsPowerISA, PRegValue},
@@ -291,7 +292,34 @@ impl InsnsBuilder {
                         .cast_to_static::<SInt<_>>(),
                     true,
                     ConditionMode.SGt(),
+                    true,
+                    true,
                     false,
+                    false,
+                )]
+            },
+        ));
+    }
+    fn power_isa_beq(&mut self, target: InsnsBuilderLabel) {
+        let pc = self.pc;
+        self.add_insn(Insn::new_lazy(
+            4,
+            format!("beq {}", self.labels[target.0].name),
+            move |labels| {
+                [BranchMOp::branch_cond_ctr(
+                    MOpDestReg::new_sim(&[], &[]),
+                    [
+                        MOpRegNum::power_isa_cr_reg_imm(0),
+                        MOpRegNum::const_zero(),
+                        MOpRegNum::const_zero(),
+                    ],
+                    labels[target.0]
+                        .pc()
+                        .wrapping_sub(pc)
+                        .cast_to_static::<SInt<_>>(),
+                    false,
+                    ConditionMode.Eq(),
+                    true,
                     true,
                     false,
                     false,
@@ -843,6 +871,7 @@ fn mock_next_pc<#[hdl(skip)] MI: MakeInsns>(config: PhantomConst<CpuConfig>) {
                 state: Cell::new(0),
             };
             let insns = MI::make_insns();
+            dbg!(&insns);
             sim.resettable(
                 cd,
                 async |mut sim| {
@@ -1272,7 +1301,12 @@ impl MockMemory {
         src_values: &[SimValue<PRegValue>; COMMON_MOP_SRC_LEN],
         is_speculative: bool,
     ) -> Result<SimValue<PRegValue>, AddressCantBeSpeculativelyAccessed> {
-        #[hdl(sim)]
+        println!("MockMemory::run_mop: {:#x}: {:?}", mop.pc.as_int(), mop.mop);
+        println!(
+            "<- {}{src_values:?}",
+            if is_speculative { "(speculative) " } else { "" },
+        );
+        let retval = #[hdl(sim)]
         match &mop.mop {
             LoadStoreMOp::<_, _>::Load(mop) => {
                 #[hdl(sim)]
@@ -1323,13 +1357,11 @@ impl MockMemory {
                         }
                     }
                 };
-                Ok(
                 #[hdl(sim)]
                 PRegValue {
                     int_fp: loaded,
                     flags: PRegFlags::zeroed_sim(),
-                    },
+                }
-                )
             }
             LoadStoreMOp::<_, _>::Store(mop) => {
                 #[hdl(sim)]
@@ -1363,25 +1395,16 @@ impl MockMemory {
                         }
                     }
                 }
-                Ok(PRegValue::zeroed_sim())
+                PRegValue::zeroed_sim()
-            }
             }
+        };
+        println!("-> {retval:?}");
+        Ok(retval)
     }
 }
 
-#[hdl(no_static)]
-struct MockL2RegFileDebugState {}
-
-#[derive(Debug, Default)]
-struct MockL2RegFile {}
-
 trait MockExecutionStateTrait: Default {
     type DebugState: BundleType;
-    fn try_get_l2_reg_file(&mut self) -> Option<&mut MockL2RegFile>;
-    fn get_l2_reg_file(&mut self) -> &mut MockL2RegFile {
-        self.try_get_l2_reg_file()
-            .expect("no MockL2RegFile available in this unit")
-    }
     fn debug_state_ty() -> Self::DebugState;
     fn zeroed_debug_state() -> SimValue<Self::DebugState>;
     fn debug_state(&self) -> SimValue<Self::DebugState>;
@@ -1732,46 +1755,17 @@ trait MockExecutionStateTrait: Default {
                 );
             }
             UnitMOp::<_, _, _>::TransformedMove(mop) => {
-                let l2_reg_file = self.get_l2_reg_file();
+                panic!(
-                #[hdl(sim)]
+                    "mock_unit can't execute L2RegisterFile MOp, that needs mock_l2_reg_file_unit: {mop:#?}"
-                match mop {
+                );
-                    L2RegisterFileMOp::<_, _>::ReadL2Reg(mop) => {
-                        todo!("implement ReadL2Reg")
-                    }
-                    L2RegisterFileMOp::<_, _>::WriteL2Reg(mop) => {
-                        todo!("implement WriteL2Reg")
-                    }
-                }
             }
             UnitMOp::<_, _, _>::Unknown => unreachable!(),
         }
     }
 }
 
-impl MockExecutionStateTrait for MockL2RegFile {
-    type DebugState = MockL2RegFileDebugState;
-    fn try_get_l2_reg_file(&mut self) -> Option<&mut MockL2RegFile> {
-        Some(self)
-    }
-    fn debug_state_ty() -> Self::DebugState {
-        MockL2RegFileDebugState
-    }
-    fn zeroed_debug_state() -> SimValue<Self::DebugState> {
-        zeroed(Self::debug_state_ty())
-    }
-    #[hdl]
-    fn debug_state(&self) -> SimValue<Self::DebugState> {
-        let Self {} = self;
-        #[hdl(sim)]
-        MockL2RegFileDebugState {}
-    }
-}
-
 impl MockExecutionStateTrait for () {
     type DebugState = ();
-    fn try_get_l2_reg_file(&mut self) -> Option<&mut MockL2RegFile> {
-        None
-    }
     fn debug_state_ty() -> Self::DebugState {
         ()
     }
@@ -2228,6 +2222,542 @@ fn mock_unit<#[hdl(skip)] E: MockExecutionStateTrait>(
     }
 }
 
+#[hdl(no_static)]
+struct MockL2RegFileOpDebugState<C: PhantomConstGet<CpuConfig>> {
+    mop: MOpInstance<L2RegisterFileMOp<PRegNum<C>, PRegNum<C>>>,
+    src_values: HdlOption<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
+    dest_value: HdlOption<PRegValue>,
+    sent_cant_cause_cancel: Bool,
+    sent_output_ready: Bool,
+    config: C,
+}
+
+#[derive(Debug)]
+struct MockL2RegFileOp<C: PhantomConstCpuConfig> {
+    mop: SimValue<MOpInstance<L2RegisterFileMOp<PRegNum<C>, PRegNum<C>>>>,
+    src_values: Option<[SimValue<PRegValue>; COMMON_MOP_SRC_LEN]>,
+    dest_value: Option<SimValue<PRegValue>>,
+    sent_cant_cause_cancel: bool,
+    sent_output_ready: bool,
+    config: C,
+}
+
+impl<C: PhantomConstCpuConfig> MockL2RegFileOp<C> {
+    #[hdl]
+    fn debug_state(&self) -> SimValue<MockL2RegFileOpDebugState<C>> {
+        let Self {
+            mop,
+            src_values,
+            dest_value,
+            sent_cant_cause_cancel,
+            sent_output_ready,
+            config,
+        } = self;
+        #[hdl(sim)]
+        MockL2RegFileOpDebugState::<_> {
+            mop,
+            src_values,
+            dest_value,
+            sent_cant_cause_cancel,
+            sent_output_ready,
+            config,
+        }
+    }
+    fn id(&self) -> &SimValue<MOpId> {
+        &self.mop.id
+    }
+}
+
+#[hdl(get(|_| L2RegNum.l2_reg_count()))]
+type L2RegFileSize<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(no_static)]
+struct MockL2RegFileUnitDebugState<C: PhantomConstGet<CpuConfig>> {
+    ops: ArrayVec<MockL2RegFileOpDebugState<C>, CpuConfigMaxUnitMaxInFlight<C>>,
+    l2_regs: ArrayType<PRegValue, L2RegFileSize<C>>,
+    config: C,
+}
+
+struct MockL2RegFileUnitState<C: PhantomConstCpuConfig> {
+    ops: VecDeque<MockL2RegFileOp<C>>,
+    l2_regs: Box<[SimValue<PRegValue>]>,
+    config: C,
+}
+
+impl<C: PhantomConstCpuConfig> MockL2RegFileUnitState<C> {
+    fn new(config: C) -> Self {
+        Self {
+            ops: VecDeque::new(),
+            l2_regs: vec![PRegValue::zeroed_sim(); L2RegFileSize[config]].into_boxed_slice(),
+            config,
+        }
+    }
+    #[hdl]
+    fn debug_state(&self) -> SimValue<MockL2RegFileUnitDebugState<C>> {
+        let Self {
+            ops,
+            l2_regs,
+            config,
+        } = self;
+        let ret_ty = MockL2RegFileUnitDebugState[*config];
+        #[hdl(sim)]
+        MockL2RegFileUnitDebugState::<_> {
+            ops: ret_ty
+                .ops
+                .from_iter_sim(
+                    zeroed(ret_ty.ops.element()),
+                    ops.iter().map(MockL2RegFileOp::debug_state),
+                )
+                .ok()
+                .expect("known to fit"),
+            l2_regs,
+            config,
+        }
+    }
+    fn try_op_by_id(&self, id: &SimValue<MOpId>) -> Option<&MockL2RegFileOp<C>> {
+        self.ops.iter().find(|op| op.id() == id)
+    }
+    fn try_op_by_id_mut(&mut self, id: &SimValue<MOpId>) -> Option<&mut MockL2RegFileOp<C>> {
+        self.ops.iter_mut().find(|op| op.id() == id)
+    }
+    #[track_caller]
+    fn op_by_id(&self, id: &SimValue<MOpId>) -> &MockL2RegFileOp<C> {
+        match self.try_op_by_id(id) {
+            Some(retval) => retval,
+            None => panic!("can't find MockL2RegFileOp with id: {id:?}"),
+        }
+    }
+    #[track_caller]
+    fn op_by_id_mut(&mut self, id: &SimValue<MOpId>) -> &mut MockL2RegFileOp<C> {
+        match self.try_op_by_id_mut(id) {
+            Some(retval) => retval,
+            None => panic!("can't find MockL2RegFileOp with id: {id:?}"),
+        }
+    }
+    #[hdl]
+    fn step(&mut self) {
+        for op in &mut self.ops {
+            let MockL2RegFileOp {
+                mop,
+                src_values,
+                dest_value,
+                sent_cant_cause_cancel: _,
+                sent_output_ready: _,
+                config: _,
+            } = op;
+            #[hdl(sim)]
+            match &mop.mop {
+                L2RegisterFileMOp::<_, _>::ReadL2Reg(mop) => {
+                    #[hdl(sim)]
+                    let ReadL2RegMOp::<_, _> { common } = mop;
+                    if src_values.is_some() && dest_value.is_none() {
+                        *dest_value = Some(self.l2_regs[L2RegNum::value_sim(&common.imm)].clone());
+                    }
+                }
+                L2RegisterFileMOp::<_, _>::WriteL2Reg(mop) => {
+                    #[hdl(sim)]
+                    let WriteL2RegMOp::<_, _> { common } = mop;
+                    if let Some(src_values) = src_values
+                        && dest_value.is_none()
+                    {
+                        self.l2_regs[L2RegNum::value_sim(&common.imm)] = src_values[0].clone();
+                        *dest_value = Some(PRegValue::zeroed_sim());
+                    }
+                    if dest_value.is_none() {
+                        // we can't run following reads yet.
+                        break;
+                    }
+                }
+            }
+        }
+    }
+    #[hdl]
+    fn handle_enqueue(&mut self, enqueue: SimValue<UnitEnqueue<C>>) {
+        #[hdl(sim)]
+        let UnitEnqueue::<_> { mop, config: _ } = enqueue;
+        #[hdl(sim)]
+        let MOpInstance::<_> {
+            fetch_block_id,
+            id,
+            pc,
+            predicted_next_pc,
+            size_in_bytes,
+            is_first_mop_in_insn,
+            is_last_mop_in_insn,
+            mop,
+        } = mop;
+        let mop = #[hdl(sim)]
+        match &mop {
+            RenamedMOp::<_>::TransformedMove(mop) => mop,
+            _ => {
+                panic!("MockL2RegFileUnitState can only handle L2RegisterFile MOps, got: {mop:#?}");
+            }
+        };
+        let mop = #[hdl(sim)]
+        MOpInstance::<_> {
+            fetch_block_id,
+            id,
+            pc,
+            predicted_next_pc,
+            size_in_bytes,
+            is_first_mop_in_insn,
+            is_last_mop_in_insn,
+            mop,
+        };
+        self.ops.push_back(MockL2RegFileOp {
+            mop,
+            src_values: None,
+            dest_value: None,
+            sent_cant_cause_cancel: false,
+            sent_output_ready: false,
+            config: self.config,
+        });
+    }
+    #[hdl]
+    fn handle_inputs_ready(&mut self, inputs_ready: SimValue<UnitInputsReady<C>>) {
+        #[hdl(sim)]
+        let UnitInputsReady::<_> {
+            mop,
+            config: _,
+            src_values,
+        } = inputs_ready;
+        let MockL2RegFileOp {
+            mop: _,
+            src_values: op_src_values,
+            dest_value: _,
+            sent_cant_cause_cancel: _,
+            sent_output_ready: _,
+            config: _,
+        } = self.op_by_id_mut(&mop.id);
+        assert!(op_src_values.is_none());
+        *op_src_values = Some(SimValue::into_value(src_values));
+    }
+    #[hdl]
+    fn handle_mop_is_no_longer_speculative(
+        &mut self,
+        is_no_longer_speculative: SimValue<UnitMOpIsNoLongerSpeculative<C>>,
+    ) {
+        #[hdl(sim)]
+        let UnitMOpIsNoLongerSpeculative::<_> { id, config: _ } = is_no_longer_speculative;
+        let MockL2RegFileOp {
+            mop: _,
+            src_values: _,
+            dest_value: _,
+            sent_cant_cause_cancel: _,
+            sent_output_ready: _,
+            config: _,
+        } = self.op_by_id_mut(&id);
+    }
+    #[hdl]
+    fn peek_mop_cant_cause_cancel(&self) -> SimValue<HdlOption<UnitMOpCantCauseCancel<C>>> {
+        let ret_ty = HdlOption[UnitMOpCantCauseCancel[self.config]];
+        for MockL2RegFileOp {
+            mop,
+            src_values: _,
+            dest_value: _,
+            sent_cant_cause_cancel,
+            sent_output_ready: _,
+            config: _,
+        } in &self.ops
+        {
+            if *sent_cant_cause_cancel {
+                continue;
+            }
+            // L2 reg file ops can never cause cancels
+            return #[hdl(sim)]
+            ret_ty.HdlSome(
+                #[hdl(sim)]
+                UnitMOpCantCauseCancel::<_> {
+                    id: &mop.id,
+                    config: self.config,
+                },
+            );
+        }
+        #[hdl(sim)]
+        ret_ty.HdlNone()
+    }
+    #[hdl]
+    fn handle_mop_cant_cause_cancel(
+        &mut self,
+        cant_cause_cancel: SimValue<UnitMOpCantCauseCancel<C>>,
+    ) {
+        #[hdl(sim)]
+        let UnitMOpCantCauseCancel::<_> { id, config: _ } = cant_cause_cancel;
+        let op = self.op_by_id_mut(&id);
+        assert!(!op.sent_cant_cause_cancel);
+        op.sent_cant_cause_cancel = true;
+    }
+    #[hdl]
+    fn peek_output_ready(&self) -> SimValue<HdlOption<UnitOutputReady<C>>> {
+        let ret_ty = HdlOption[UnitOutputReady[self.config]];
+        for MockL2RegFileOp {
+            mop,
+            src_values: _,
+            dest_value,
+            sent_cant_cause_cancel: _,
+            sent_output_ready,
+            config: _,
+        } in &self.ops
+        {
+            if *sent_output_ready {
+                continue;
+            }
+            if let Some(dest_value) = dest_value {
+                return #[hdl(sim)]
+                ret_ty.HdlSome(
+                    #[hdl(sim)]
+                    UnitOutputReady::<_> {
+                        id: &mop.id,
+                        dest_value,
+                        predictor_op: #[hdl(sim)]
+                        NextPcPredictorOp::<_> {
+                            call_stack_op: #[hdl(sim)]
+                            CallStackOp.None(),
+                            cond_br_taken: #[hdl(sim)]
+                            HdlNone(),
+                            config: self.config,
+                        },
+                    },
+                );
+            }
+        }
+        #[hdl(sim)]
+        ret_ty.HdlNone()
+    }
+    #[hdl]
+    fn handle_output_ready(&mut self, output_ready: SimValue<UnitOutputReady<C>>) {
+        #[hdl(sim)]
+        let UnitOutputReady::<_> {
+            id,
+            dest_value: _,
+            predictor_op: _,
+        } = output_ready;
+        let op = self.op_by_id_mut(&id);
+        assert!(!op.sent_output_ready);
+        op.sent_output_ready = true;
+    }
+    #[hdl]
+    fn peek_finish_cause_cancel(&self) -> SimValue<HdlOption<UnitFinishCauseCancel<C>>> {
+        let ret_ty = HdlOption[UnitFinishCauseCancel[self.config]];
+        if let Some(MockL2RegFileOp {
+            mop,
+            src_values: _,
+            dest_value,
+            sent_cant_cause_cancel: _,
+            sent_output_ready: _,
+            config: _,
+        }) = self.ops.front()
+        {
+            if dest_value.is_some() {
+                return #[hdl(sim)]
+                ret_ty.HdlSome(
+                    #[hdl(sim)]
+                    UnitFinishCauseCancel::<_> {
+                        id: &mop.id,
+                        caused_cancel: #[hdl(sim)]
+                        (ret_ty.HdlSome.caused_cancel).HdlNone(),
+                        config: self.config,
+                    },
+                );
+            }
+        }
+        #[hdl(sim)]
+        ret_ty.HdlNone()
+    }
+    #[hdl]
+    fn handle_finish_cause_cancel(
+        &mut self,
+        finish_cause_cancel: SimValue<UnitFinishCauseCancel<C>>,
+    ) {
+        #[hdl(sim)]
+        let UnitFinishCauseCancel::<_> {
+            id,
+            caused_cancel: _,
+            config: _,
+        } = finish_cause_cancel;
+        let Some(op) = self.ops.pop_front() else {
+            unreachable!();
+        };
+        assert_eq!(*op.id(), id);
+    }
+    fn cancel_all(&mut self) {
+        let Self {
+            ops,
+            l2_regs: _,
+            config: _,
+        } = self;
+        ops.clear();
+    }
+}
+
+#[hdl_module(extern)]
+fn mock_l2_reg_file_unit(config: PhantomConst<CpuConfig>, unit_index: usize) {
+    assert!(
+        config
+            .get()
+            .units
+            .iter()
+            .enumerate()
+            .all(|(i, unit)| (unit_index == i) == (unit.kind == UnitKind::TransformedMove)),
+        "unit index {unit_index} must be the only L2 register file unit: {config:#?}",
+    );
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let from_execute: ExecuteToUnitInterface<PhantomConst<CpuConfig>> =
+        m.input(ExecuteToUnitInterface[config]);
+    #[hdl]
+    let debug_state: MockL2RegFileUnitDebugState<PhantomConst<CpuConfig>> =
+        m.output(MockL2RegFileUnitDebugState[config]);
+    m.register_clock_for_past(cd.clk);
+    m.extern_module_simulation_fn(
+        (cd, from_execute, debug_state, config),
+        async |args, mut sim| {
+            let (cd, from_execute, debug_state, config) = args;
+            sim.resettable(
+                cd,
+                async |mut sim| {
+                    #[hdl]
+                    let ExecuteToUnitInterface::<_> {
+                        enqueue,
+                        inputs_ready: _,
+                        is_no_longer_speculative: _,
+                        cant_cause_cancel,
+                        output_ready,
+                        finish_cause_cancel,
+                        unit_outputs_ready: _,
+                        cancel_all,
+                        config: _,
+                    } = from_execute;
+                    sim.write(enqueue.ready, false).await;
+                    sim.write(cancel_all.ready, false).await;
+                    sim.write(
+                        cant_cause_cancel,
+                        #[hdl(sim)]
+                        (cant_cause_cancel.ty()).HdlNone(),
+                    )
+                    .await;
+                    sim.write(
+                        output_ready,
+                        #[hdl(sim)]
+                        (output_ready.ty()).HdlNone(),
+                    )
+                    .await;
+                    sim.write(
+                        finish_cause_cancel,
+                        #[hdl(sim)]
+                        (finish_cause_cancel.ty()).HdlNone(),
+                    )
+                    .await;
+                    sim.write(
+                        debug_state,
+                        #[hdl(sim)]
+                        MockL2RegFileUnitDebugState::<_> {
+                            ops: zeroed(debug_state.ty().ops),
+                            l2_regs: zeroed(debug_state.ty().l2_regs),
+                            config,
+                        },
+                    )
+                    .await;
+                },
+                |sim, ()| run_fn(cd, from_execute, debug_state, config, sim),
+            )
+            .await;
+        },
+    );
+    #[hdl]
+    async fn run_fn(
+        cd: Expr<ClockDomain>,
+        from_execute: Expr<ExecuteToUnitInterface<PhantomConst<CpuConfig>>>,
+        debug_state: Expr<MockL2RegFileUnitDebugState<PhantomConst<CpuConfig>>>,
+        config: PhantomConst<CpuConfig>,
+        mut sim: ExternModuleSimulationState,
+    ) {
+        let mut state = MockL2RegFileUnitState::new(config);
+        loop {
+            {
+                #[hdl]
+                let ExecuteToUnitInterface::<_> {
+                    enqueue,
+                    inputs_ready: _,
+                    is_no_longer_speculative: _,
+                    cant_cause_cancel,
+                    output_ready,
+                    finish_cause_cancel,
+                    unit_outputs_ready: _,
+                    cancel_all,
+                    config: _,
+                } = from_execute;
+                sim.write(enqueue.ready, true).await;
+                sim.write(cancel_all.ready, true).await;
+                sim.write(cant_cause_cancel, state.peek_mop_cant_cause_cancel())
+                    .await;
+                sim.write(output_ready, state.peek_output_ready()).await;
+                sim.write(finish_cause_cancel, state.peek_finish_cause_cancel())
+                    .await;
+            }
+            sim.write(debug_state, state.debug_state()).await;
+            sim.wait_for_clock_edge(cd.clk).await;
+            {
+                #[hdl]
+                let ExecuteToUnitInterface::<_> {
+                    enqueue,
+                    inputs_ready,
+                    is_no_longer_speculative,
+                    cant_cause_cancel,
+                    output_ready,
+                    finish_cause_cancel,
+                    unit_outputs_ready,
+                    cancel_all,
+                    config: _,
+                } = from_execute;
+                if sim.read_past_bool(enqueue.ready, cd.clk).await {
+                    #[hdl(sim)]
+                    if let HdlSome(enqueue) = sim.read_past(enqueue.data, cd.clk).await {
+                        state.handle_enqueue(enqueue);
+                    }
+                }
+                #[hdl(sim)]
+                if let HdlSome(inputs_ready) = sim.read_past(inputs_ready, cd.clk).await {
+                    state.handle_inputs_ready(inputs_ready);
+                }
+                #[hdl(sim)]
+                if let HdlSome(is_no_longer_speculative) =
+                    sim.read_past(is_no_longer_speculative, cd.clk).await
+                {
+                    state.handle_mop_is_no_longer_speculative(is_no_longer_speculative);
+                }
+                if sim.read_past_bool(unit_outputs_ready, cd.clk).await {
+                    #[hdl(sim)]
+                    if let HdlSome(cant_cause_cancel) =
+                        sim.read_past(cant_cause_cancel, cd.clk).await
+                    {
+                        state.handle_mop_cant_cause_cancel(cant_cause_cancel);
+                    }
+                    #[hdl(sim)]
+                    if let HdlSome(output_ready) = sim.read_past(output_ready, cd.clk).await {
+                        state.handle_output_ready(output_ready);
+                    }
+                    #[hdl(sim)]
+                    if let HdlSome(finish_cause_cancel) =
+                        sim.read_past(finish_cause_cancel, cd.clk).await
+                    {
+                        state.handle_finish_cause_cancel(finish_cause_cancel);
+                    }
+                }
+                if sim.read_past_bool(cancel_all.ready, cd.clk).await {
+                    #[hdl(sim)]
+                    if let HdlSome(cancel_all) = sim.read_past(cancel_all.data, cd.clk).await {
+                        let () = *cancel_all;
+                        state.cancel_all();
+                    }
+                }
+            }
+            state.step();
+        }
+    }
+}
+
 #[hdl(no_static)]
 struct MockLoadStoreOpDebugState<C: PhantomConstGet<CpuConfig>> {
     mop: MOpInstance<LoadStoreMOp<PRegNum<C>, PRegNum<C>>>,
@@ -2897,7 +3427,8 @@ fn rename_execute_retire_test_harness<#[hdl(skip)] MI: MakeInsns>(config: Phanto
         .into_iter()
         .enumerate()
     {
-        if config.get().units[unit_index].kind == UnitKind::LoadStore {
+        match config.get().units[unit_index].kind {
+            UnitKind::LoadStore => {
                 let mock_unit = instance_with_loc(
                     &dut.ty().to_units.unit_field_name(unit_index),
                     mock_load_store_unit::<MI>(config, unit_index),
@@ -2909,20 +3440,26 @@ fn rename_execute_retire_test_harness<#[hdl(skip)] MI: MakeInsns>(config: Phanto
                 if !mock_unit.all_outputs_written {
                     connect(all_outputs_written, false);
                 }
-        } else {
+            }
-            let mock_unit_module = if is_the_l2_reg_file_unit(config, unit_index) {
+            UnitKind::TransformedMove => {
-                mock_unit::<MockL2RegFile>(config, unit_index)
-            } else {
-                mock_unit::<()>(config, unit_index)
-            };
                 let mock_unit = instance_with_loc(
                     &dut.ty().to_units.unit_field_name(unit_index),
-                mock_unit_module,
+                    mock_l2_reg_file_unit(config, unit_index),
                     SourceLocation::caller(),
                 );
                 connect(mock_unit.cd, cd);
                 connect(mock_unit.from_execute, to_unit);
             }
+            UnitKind::AluBranch => {
+                let mock_unit = instance_with_loc(
+                    &dut.ty().to_units.unit_field_name(unit_index),
+                    mock_unit::<()>(config, unit_index),
+                    SourceLocation::caller(),
+                );
+                connect(mock_unit.cd, cd);
+                connect(mock_unit.from_execute, to_unit);
+            }
+        }
     }
 }
 
@@ -2977,3 +3514,135 @@ fn test_rename_execute_retire_fibonacci() {
         panic!();
     }
 }
+
+struct SlowLoopInsns;
+
+impl SlowLoopInsns {
+    const CONSTANTS_ADDR: u64 = 0x4000;
+    const CONSTANTS_COUNT: usize = 16;
+    const CONSTANTS_STEP: usize = 8;
+    const LOG2_RESULT_FACTOR: u32 = 2;
+}
+
+impl MakeInsns for SlowLoopInsns {
+    fn make_insns() -> Insns {
+        let mut b = InsnsBuilder::new();
+
+        let slow_loop = b.new_label("slow_loop");
+        b.power_isa_ld(3, 0, MockMemory::IO_ADDR as i16); // load input
+        b.power_isa_addi(1, 0, 0x4000); // setup stack pointer
+        b.power_isa_bl(slow_loop);
+        b.power_isa_std(3, 0, MockMemory::IO_ADDR as i16); // store output
+        let done = b.new_defined_label("done");
+        b.power_isa_b(done);
+
+        b.set_pc(0x1000);
+        b.define_label(slow_loop);
+        b.power_isa_addi(4, 0, 0); // clear sum
+
+        let loop_header = b.new_defined_label("loop_header");
+        b.power_isa_cmpldi(0, 3, 0);
+        let loop_done = b.new_label("loop_done");
+        b.power_isa_beq(loop_done); // if input == 0 goto loop_done
+
+        // long sequence of loads to provoke L2 register file store
+        let start_reg = 5;
+        assert!(
+            start_reg + Self::CONSTANTS_COUNT <= 32,
+            "too many constants to load them all into PowerISA GPRs",
+        );
+        for i in 0..Self::CONSTANTS_COUNT {
+            b.power_isa_ld(
+                start_reg + i,
+                0,
+                (Self::CONSTANTS_ADDR + (Self::CONSTANTS_STEP * i) as u64) as i16,
+            );
+        }
+        for i in 0..Self::CONSTANTS_COUNT {
+            b.power_isa_add(4, 4, start_reg + i);
+        }
+
+        b.power_isa_addi(3, 3, -1);
+        b.power_isa_b(loop_header);
+
+        b.define_label(loop_done);
+        b.power_isa_mr(3, 4);
+        for _ in 0..Self::LOG2_RESULT_FACTOR {
+            b.power_isa_add(3, 3, 3);
+        }
+        b.power_isa_blr(); // return
+
+        b.build()
+    }
+    fn make_load_store_execution_state() -> MockMemory {
+        let expected = 0x0123_4567_89AB_CDEF_u64;
+        let constants: [[u8; Self::CONSTANTS_STEP]; Self::CONSTANTS_COUNT] =
+            std::array::from_fn(|i| {
+                let start_bit_index = i * 64 / Self::CONSTANTS_COUNT;
+                let end_bit_index = (i + 1) * 64 / Self::CONSTANTS_COUNT;
+                let start_bit = 1u64.unbounded_shl(start_bit_index as u32);
+                let end_bit = 1u64.unbounded_shl(end_bit_index as u32);
+                let mask = end_bit.wrapping_sub(start_bit);
+                (expected & mask).to_le_bytes()
+            });
+        let loop_count = 4;
+        MockMemory::new(
+            loop_count,
+            expected * loop_count * 2u64.pow(Self::LOG2_RESULT_FACTOR),
+            [(Self::CONSTANTS_ADDR, constants.as_flattened())],
+        )
+    }
+}
+
+#[hdl]
+#[test]
+fn test_rename_execute_retire_slow_loop() {
+    let _n = SourceLocation::normalize_files_for_tests();
+    let mut config = CpuConfig::new(
+        vec![
+            UnitConfig::new(UnitKind::AluBranch),
+            UnitConfig::new(UnitKind::AluBranch),
+            UnitConfig::new(UnitKind::AluBranch),
+            UnitConfig::new(UnitKind::AluBranch),
+            UnitConfig::new(UnitKind::LoadStore),
+            UnitConfig::new(UnitKind::TransformedMove),
+        ],
+        NonZeroUsize::new(20).unwrap(),
+    );
+    config.fetch_width = NonZeroUsize::new(4).unwrap();
+    let m = rename_execute_retire_test_harness::<SlowLoopInsns>(PhantomConst::new_sized(config));
+    let mut sim = Simulation::new(m);
+    let writer = RcWriter::default();
+    sim.add_trace_writer(VcdWriterDecls::new(writer.clone()));
+    struct DumpVcdOnDrop {
+        writer: Option<RcWriter>,
+    }
+    impl Drop for DumpVcdOnDrop {
+        fn drop(&mut self) {
+            if let Some(mut writer) = self.writer.take() {
+                let vcd = String::from_utf8(writer.take()).unwrap();
+                println!("####### VCD:\n{vcd}\n#######");
+            }
+        }
+    }
+    let mut writer = DumpVcdOnDrop {
+        writer: Some(writer),
+    };
+    sim.write_clock(sim.io().cd.clk, false);
+    sim.write_reset(sim.io().cd.rst, true);
+    for cycle in 0..200 {
+        sim.advance_time(SimDuration::from_nanos(500));
+        println!("clock tick: {cycle}");
+        sim.write_clock(sim.io().cd.clk, true);
+        sim.advance_time(SimDuration::from_nanos(500));
+        sim.write_clock(sim.io().cd.clk, false);
+        sim.write_reset(sim.io().cd.rst, false);
+    }
+    assert!(sim.read_bool(sim.io().all_outputs_written));
+    // FIXME: vcd is just whatever rename_execute_retire does now, which isn't known to be correct
+    let vcd = String::from_utf8(writer.writer.take().unwrap().take()).unwrap();
+    println!("####### VCD:\n{vcd}\n#######");
+    if vcd != include_str!("expected/rename_execute_retire_slow_loop.vcd") {
+        panic!();
+    }
+}