WIP adding rename_execute_retire

crates/cpu/Cargo.toml

@@ -33,3 +33,6 @@ hex-literal.workspace = true
 regex = "1.12.2"
 sha2.workspace = true
 which.workspace = true
+
+[lints.rust]
+unexpected_cfgs = { level = "warn", check-cfg = ['cfg(todo)'] }

crates/cpu/src/config.rs

@@ -1,13 +1,7 @@
 // SPDX-License-Identifier: LGPL-3.0-or-later
 // See Notices.txt for copyright information
-use crate::{
+use crate::{instruction::CONST_ZERO_UNIT_NUM, unit::UnitKind};
-    instruction::{CONST_ZERO_UNIT_NUM, MOpTrait, PRegNum, RenamedMOp, UnitNum, UnitOutRegNum},
+use fayalite::{expr::HdlPartialOrdImpl, prelude::*};
-    unit::{
-        UnitCancelInput, UnitKind, UnitOutputWrite,
-        unit_base::{UnitForwardingInfo, UnitToRegAlloc},
-    },
-};
-use fayalite::prelude::*;
 use serde::{Deserialize, Serialize};
 use std::num::NonZeroUsize;
 
@@ -101,54 +95,20 @@ impl CpuConfig {
     pub fn unit_num_width(&self) -> usize {
         UInt::range(CONST_ZERO_UNIT_NUM..self.non_const_unit_nums().end).width()
     }
-    pub fn unit_num(&self) -> UnitNum<DynSize> {
-        UnitNum[self.unit_num_width()]
-    }
-    pub fn unit_out_reg_num(&self) -> UnitOutRegNum<DynSize> {
-        UnitOutRegNum[self.out_reg_num_width]
-    }
-    pub fn p_reg_num(&self) -> PRegNum<DynSize, DynSize> {
-        PRegNum[self.unit_num_width()][self.out_reg_num_width]
-    }
     pub fn p_reg_num_width(&self) -> usize {
         self.unit_num_width() + self.out_reg_num_width
     }
-    pub fn renamed_mop_in_unit(&self) -> RenamedMOp<UnitOutRegNum<DynSize>, DynSize> {
-        RenamedMOp[self.unit_out_reg_num()][self.p_reg_num_width()]
-    }
-    pub fn unit_output_write(&self) -> UnitOutputWrite<DynSize> {
-        UnitOutputWrite[self.out_reg_num_width]
-    }
-    pub fn unit_output_writes(&self) -> Array<HdlOption<UnitOutputWrite<DynSize>>> {
-        Array[HdlOption[self.unit_output_write()]][self.non_const_unit_nums().len()]
-    }
-    pub fn unit_cancel_input(&self) -> UnitCancelInput<DynSize> {
-        UnitCancelInput[self.out_reg_num_width]
-    }
-    pub fn unit_forwarding_info(&self) -> UnitForwardingInfo<DynSize, DynSize, DynSize> {
-        UnitForwardingInfo[self.unit_num_width()][self.out_reg_num_width]
-            [self.non_const_unit_nums().len()]
-    }
     pub fn unit_max_in_flight(&self, unit_index: usize) -> NonZeroUsize {
         self.units[unit_index]
             .max_in_flight
             .unwrap_or(self.default_unit_max_in_flight)
     }
-    pub fn unit_to_reg_alloc<
+    /// the maximum of all [`unit_max_in_flight()`][Self::unit_max_in_flight()]
-        MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>, SrcRegWidth = DynSize>,
+    pub fn max_unit_max_in_flight(&self) -> NonZeroUsize {
-        ExtraOut: Type,
+        (0..self.units.len())
-    >(
+            .map(|unit_index| self.unit_max_in_flight(unit_index))
-        &self,
+            .max()
-        mop_ty: MOp,
+            .unwrap_or(self.default_unit_max_in_flight)
-        extra_out_ty: ExtraOut,
-    ) -> UnitToRegAlloc<MOp, ExtraOut, DynSize, DynSize, DynSize> {
-        assert_eq!(
-            mop_ty.dest_reg_ty(),
-            self.unit_out_reg_num(),
-            "inconsistent types",
-        );
-        UnitToRegAlloc[mop_ty][extra_out_ty][self.unit_num_width()][self.out_reg_num_width]
-            [self.non_const_unit_nums().len()]
     }
     pub fn fetch_width_in_bytes(&self) -> usize {
         1usize
@@ -188,6 +148,21 @@ impl CpuConfig {
     }
 }
 
+#[hdl(get(|c| c.out_reg_num_width))]
+pub type CpuConfigOutRegNumWidth<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.unit_num_width()))]
+pub type CpuConfigUnitNumWidth<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.p_reg_num_width()))]
+pub type CpuConfigPRegNumWidth<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| 1 << c.out_reg_num_width))]
+pub type CpuConfig2PowOutRegNumWidth<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.units.len()))]
+pub type CpuConfigUnitCount<C: PhantomConstGet<CpuConfig>> = DynSize;
+
 #[hdl(get(|c| c.fetch_width.get()))]
 pub type CpuConfigFetchWidth<C: PhantomConstGet<CpuConfig>> = DynSize;
 
@@ -236,6 +211,10 @@ pub type CpuConfigL1ICacheMaxMissesInFlight<C: PhantomConstGet<CpuConfig>> = Dyn
 #[hdl(get(|c| c.rob_size.get()))]
 pub type CpuConfigRobSize<C: PhantomConstGet<CpuConfig>> = DynSize;
 
+/// the maximum of all [`unit_max_in_flight()`][CpuConfig::unit_max_in_flight()]
+#[hdl(get(|c| c.max_unit_max_in_flight().get()))]
+pub type CpuConfigMaxUnitMaxInFlight<C: PhantomConstGet<CpuConfig>> = DynSize;
+
 pub trait PhantomConstCpuConfig:
     PhantomConstGet<CpuConfig>
     + Into<PhantomConst<CpuConfig>>
@@ -243,6 +222,7 @@ pub trait PhantomConstCpuConfig:
     + Type
     + ToSimValue<Type = Self>
     + ToExpr<Type = Self>
+    + HdlPartialOrdImpl<Self>
 {
 }
 

crates/cpu/src/instruction/power_isa.rs

@@ -192,6 +192,13 @@ impl MOpRegNum {
         power_isa_gpr_or_zero_reg
     }
     #[hdl]
+    pub fn power_isa_gpr_or_zero_reg_imm(index: usize) -> Expr<Self> {
+        #[hdl]
+        Self {
+            value: Self::power_isa_gpr_or_zero_reg_num(index).cast_to_static::<UInt<_>>(),
+        }
+    }
+    #[hdl]
     pub fn power_isa_gpr_or_zero_reg_sim(reg_num: &SimValue<UInt<5>>) -> SimValue<Self> {
         #[hdl(sim)]
         Self {

crates/cpu/src/lib.rs

@@ -7,7 +7,9 @@ pub mod instruction;
 pub mod main_memory_and_io;
 pub mod next_pc;
 pub mod powerisa_instructions_xml;
+#[cfg(todo)]
 pub mod reg_alloc;
 pub mod register;
+pub mod rename_execute_retire;
 pub mod unit;
 pub mod util;

crates/cpu/src/unit.rs

@@ -2,7 +2,7 @@
 // See Notices.txt for copyright information
 
 use crate::{
-    config::CpuConfig,
+    config::{CpuConfig, PhantomConstCpuConfig},
     instruction::{
         AluBranchMOp, LoadStoreMOp, MOp, MOpDestReg, MOpInto, MOpRegNum, MOpTrait,
         MOpVariantVisitOps, MOpVariantVisitor, MOpVisitVariants, RenamedMOp, UnitOutRegNum,
@@ -32,7 +32,7 @@ macro_rules! all_units {
             $(
                 $(#[transformed_move $($transformed_move:tt)*])?
                 #[create_dyn_unit_fn = $create_dyn_unit_fn:expr]
-                #[extract = $extract:ident]
+                #[extract($extract:ident, $extract_sim:ident, $extract_sim_ref:ident, $extract_sim_mut:ident)]
                 $(#[$variant_meta:meta])*
                 $Unit:ident($Op:ty),
             )*
@@ -48,7 +48,7 @@ macro_rules! all_units {
         }
 
         impl $UnitKind {
-            pub fn unit(self, config: &CpuConfig, unit_index: usize) -> DynUnit {
+            pub fn unit(self, config: PhantomConst<CpuConfig>, unit_index: usize) -> DynUnit {
                 match self {
                     $($UnitKind::$Unit => $create_dyn_unit_fn(config, unit_index),)*
                 }
@@ -82,6 +82,7 @@ macro_rules! all_units {
         }
 
         mop_enum! {
+            #[debug(where $TransformedMoveOp: crate::instruction::MOpDebug)]
             #[impl_mop_into = false]
             #[hdl]
             $(#[$enum_meta])*
@@ -112,6 +113,15 @@ macro_rules! all_units {
                 }
                 unit_kind
             }
+            #[hdl]
+            $vis fn kind_sim(expr: &SimValue<Self>) -> UnitKind {
+                #![allow(unreachable_patterns)]
+                #[hdl(sim)]
+                match expr {
+                    $(Self::$Unit(_) => $UnitKind::$Unit,)*
+                    _ => unreachable!(),
+                }
+            }
             $(
                 #[hdl]
                 $vis fn $extract(expr: impl ToExpr<Type = Self>) -> Expr<HdlOption<$Op>> {
@@ -126,6 +136,34 @@ macro_rules! all_units {
                     }
                     $extract
                 }
+                #[hdl]
+                $vis fn $extract_sim(expr: impl ToSimValue<Type = Self>) -> Option<SimValue<$Op>> {
+                    let expr = expr.into_sim_value();
+                    #[hdl(sim)]
+                    if let Self::$Unit(v) = expr {
+                        Some(v)
+                    } else {
+                        None
+                    }
+                }
+                #[hdl]
+                $vis fn $extract_sim_ref(expr: &SimValue<Self>) -> Option<&SimValue<$Op>> {
+                    #[hdl(sim)]
+                    if let Self::$Unit(v) = expr {
+                        Some(v)
+                    } else {
+                        None
+                    }
+                }
+                #[hdl]
+                $vis fn $extract_sim_mut(expr: &mut SimValue<Self>) -> Option<&mut SimValue<$Op>> {
+                    #[hdl(sim)]
+                    if let Self::$Unit(v) = expr {
+                        Some(v)
+                    } else {
+                        None
+                    }
+                }
             )*
             $vis fn with_transformed_move_op_ty<T>(self, new_transformed_move_op_ty: T) -> $UnitMOpEnum<$DestReg, $SrcRegWidth, T>
             where
@@ -196,6 +234,45 @@ macro_rules! all_units {
                 }
                 with_transformed_move_op
             }
+            #[hdl]
+            $vis fn try_with_transformed_move_op_sim<T, E>(
+                this: impl ToSimValue<Type = Self>,
+                new_transformed_move_op_ty: T,
+                f: impl FnOnce(SimValue<$TransformedMoveOp>) -> Result<SimValue<$UnitMOpEnum<$DestReg, $SrcRegWidth, T>>, E>,
+            ) -> Result<SimValue<$UnitMOpEnum<$DestReg, $SrcRegWidth, T>>, E>
+            where
+                T: MOpTrait<DestReg = $DestReg, SrcRegWidth = $SrcRegWidth>,
+                $TransformedMoveOp: MOpTrait<DestReg = $DestReg, SrcRegWidth = $SrcRegWidth>,
+            {
+                #![allow(unreachable_patterns)]
+                let this = this.into_sim_value();
+                let new_ty = this.ty().with_transformed_move_op_ty(new_transformed_move_op_ty);
+                #[hdl(sim)]
+                match this {
+                    $(Self::$BeforeUnit(unit) => Ok(
+                        #[hdl(sim)]
+                        new_ty.$BeforeUnit(unit)
+                    ),)*
+                    Self::$TransformedMove(unit) => f(unit),
+                    $(Self::$AfterUnit(unit) => Ok(
+                        #[hdl(sim)]
+                        new_ty.$AfterUnit(unit)
+                    ),)*
+                    _ => unreachable!(),
+                }
+            }
+            $vis fn with_transformed_move_op_sim<T>(
+                this: impl ToSimValue<Type = Self>,
+                new_transformed_move_op_ty: T,
+                f: impl FnOnce(SimValue<$TransformedMoveOp>) -> SimValue<$UnitMOpEnum<$DestReg, $SrcRegWidth, T>>,
+            ) -> SimValue<$UnitMOpEnum<$DestReg, $SrcRegWidth, T>>
+            where
+                T: MOpTrait<DestReg = $DestReg, SrcRegWidth = $SrcRegWidth>,
+                $TransformedMoveOp: MOpTrait<DestReg = $DestReg, SrcRegWidth = $SrcRegWidth>,
+            {
+                let Ok::<_, std::convert::Infallible>(retval) = Self::try_with_transformed_move_op_sim(this, new_transformed_move_op_ty, move |v| Ok(f(v)));
+                retval
+            }
         }
 
         const _: () = {
@@ -254,14 +331,14 @@ all_units! {
         })] TransformedMoveOp: Type
     > {
         #[create_dyn_unit_fn = |config, unit_index| alu_branch::AluBranch::new(config, unit_index).to_dyn()]
-        #[extract = alu_branch_mop]
+        #[extract(alu_branch_mop, alu_branch_mop_sim, alu_branch_mop_sim_ref, alu_branch_mop_sim_mut)]
         AluBranch(AluBranchMOp<DestReg, SrcRegWidth>),
         #[transformed_move]
         #[create_dyn_unit_fn = |config, unit_index| todo!()]
-        #[extract = transformed_move_mop]
+        #[extract(transformed_move_mop, transformed_move_mop_sim, transformed_move_mop_sim_ref, transformed_move_mop_sim_mut)]
         TransformedMove(TransformedMoveOp),
         #[create_dyn_unit_fn = |config, unit_index| todo!()]
-        #[extract = load_store_mop]
+        #[extract(load_store_mop, load_store_mop_sim, load_store_mop_sim_ref, load_store_mop_sim_mut)]
         LoadStore(LoadStoreMOp<DestReg, SrcRegWidth>),
     }
 }
@@ -277,9 +354,9 @@ pub struct UnitResultCompleted<ExtraOut> {
     pub extra_out: ExtraOut,
 }
 
-#[hdl(cmp_eq)]
+#[hdl(cmp_eq, no_static)]
-pub struct UnitOutputWrite<OutRegNumWidth: Size> {
+pub struct UnitOutputWrite<C: PhantomConstGet<CpuConfig>> {
-    pub which: UnitOutRegNum<OutRegNumWidth>,
+    pub which: UnitOutRegNum<C>,
     pub value: PRegValue,
 }
 
@@ -300,21 +377,21 @@ impl<ExtraOut: Type> UnitResult<ExtraOut> {
     }
 }
 
-#[hdl]
+#[hdl(no_static)]
-pub struct UnitOutput<OutRegNumWidth: Size, ExtraOut> {
+pub struct UnitOutput<C: PhantomConstGet<CpuConfig>, ExtraOut> {
-    pub which: UnitOutRegNum<OutRegNumWidth>,
+    pub which: UnitOutRegNum<C>,
     pub result: UnitResult<ExtraOut>,
 }
 
-impl<OutRegNumWidth: Size, ExtraOut: Type> UnitOutput<OutRegNumWidth, ExtraOut> {
+impl<C: PhantomConstCpuConfig, ExtraOut: Type> UnitOutput<C, ExtraOut> {
     pub fn extra_out_ty(self) -> ExtraOut {
         self.result.extra_out_ty()
     }
 }
 
-#[hdl(cmp_eq)]
+#[hdl(cmp_eq, no_static)]
-pub struct UnitCancelInput<OutRegNumWidth: Size> {
+pub struct UnitCancelInput<C: PhantomConstGet<CpuConfig>> {
-    pub which: UnitOutRegNum<OutRegNumWidth>,
+    pub which: UnitOutRegNum<C>,
 }
 
 pub trait UnitTrait:
@@ -332,7 +409,7 @@ pub trait UnitTrait:
 
     fn extract_mop(
         &self,
-        mop: Expr<RenamedMOp<UnitOutRegNum<DynSize>, DynSize>>,
+        mop: Expr<RenamedMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>>,
     ) -> Expr<HdlOption<Self::MOp>>;
 
     fn module(&self) -> Interned<Module<Self::Type>>;
@@ -340,7 +417,7 @@ pub trait UnitTrait:
     fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>>;
+    ) -> Expr<UnitToRegAlloc<PhantomConst<CpuConfig>, Self::MOp, Self::ExtraOut>>;
 
     fn cd(&self, this: Expr<Self::Type>) -> Expr<ClockDomain>;
 
@@ -390,7 +467,7 @@ impl UnitTrait for DynUnit {
 
     fn extract_mop(
         &self,
-        mop: Expr<RenamedMOp<UnitOutRegNum<DynSize>, DynSize>>,
+        mop: Expr<RenamedMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>>,
     ) -> Expr<HdlOption<Self::MOp>> {
         self.unit.extract_mop(mop)
     }
@@ -402,7 +479,7 @@ impl UnitTrait for DynUnit {
     fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>> {
+    ) -> Expr<UnitToRegAlloc<PhantomConst<CpuConfig>, Self::MOp, Self::ExtraOut>> {
         self.unit.unit_to_reg_alloc(this)
     }
 
@@ -445,7 +522,7 @@ impl<T: UnitTrait + Clone + std::hash::Hash + Eq> UnitTrait for DynUnitWrapper<T
 
     fn extract_mop(
         &self,
-        mop: Expr<RenamedMOp<UnitOutRegNum<DynSize>, DynSize>>,
+        mop: Expr<RenamedMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>>,
     ) -> Expr<HdlOption<Self::MOp>> {
         Expr::from_enum(Expr::as_enum(self.0.extract_mop(mop)))
     }
@@ -457,7 +534,7 @@ impl<T: UnitTrait + Clone + std::hash::Hash + Eq> UnitTrait for DynUnitWrapper<T
     fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>> {
+    ) -> Expr<UnitToRegAlloc<PhantomConst<CpuConfig>, Self::MOp, Self::ExtraOut>> {
         Expr::from_bundle(Expr::as_bundle(
             self.0.unit_to_reg_alloc(Expr::from_bundle(this)),
         ))

crates/cpu/src/unit/alu_branch.rs

@@ -19,16 +19,13 @@ use crate::{
     },
 };
 use fayalite::{
-    intern::{Intern, Interned},
+    intern::Interned, module::wire_with_loc, prelude::*, util::ready_valid::ReadyValid,
-    module::wire_with_loc,
-    prelude::*,
-    util::ready_valid::ReadyValid,
 };
 use std::{collections::HashMap, ops::RangeTo};
 
 #[hdl]
 fn add_sub<SrcCount: KnownSize>(
-    mop: Expr<AddSubMOp<UnitOutRegNum<DynSize>, DynSize, SrcCount>>,
+    mop: Expr<AddSubMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize, SrcCount>>,
     pc: Expr<UInt<64>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
@@ -245,7 +242,7 @@ fn add_sub<SrcCount: KnownSize>(
 
 #[hdl]
 fn logical_flags(
-    mop: Expr<LogicalFlagsMOp<UnitOutRegNum<DynSize>, DynSize>>,
+    mop: Expr<LogicalFlagsMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
 ) -> Expr<UnitResultCompleted<()>> {
@@ -259,7 +256,7 @@ fn logical_flags(
 
 #[hdl]
 fn logical(
-    mop: Expr<LogicalMOp<UnitOutRegNum<DynSize>, DynSize, ConstUsize<2>>>,
+    mop: Expr<LogicalMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize, ConstUsize<2>>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
 ) -> Expr<UnitResultCompleted<()>> {
@@ -273,7 +270,7 @@ fn logical(
 
 #[hdl]
 fn logical_i(
-    mop: Expr<LogicalMOp<UnitOutRegNum<DynSize>, DynSize, ConstUsize<1>>>,
+    mop: Expr<LogicalMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize, ConstUsize<1>>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
 ) -> Expr<UnitResultCompleted<()>> {
@@ -287,7 +284,7 @@ fn logical_i(
 
 #[hdl]
 fn shift_rotate(
-    mop: Expr<ShiftRotateMOp<UnitOutRegNum<DynSize>, DynSize>>,
+    mop: Expr<ShiftRotateMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
 ) -> Expr<UnitResultCompleted<()>> {
@@ -301,7 +298,7 @@ fn shift_rotate(
 
 #[hdl]
 fn compare<SrcCount: KnownSize>(
-    mop: Expr<CompareMOp<UnitOutRegNum<DynSize>, DynSize, SrcCount>>,
+    mop: Expr<CompareMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize, SrcCount>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
 ) -> Expr<UnitResultCompleted<()>> {
@@ -315,7 +312,7 @@ fn compare<SrcCount: KnownSize>(
 
 #[hdl]
 fn branch<SrcCount: KnownSize>(
-    mop: Expr<BranchMOp<UnitOutRegNum<DynSize>, DynSize, SrcCount>>,
+    mop: Expr<BranchMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize, SrcCount>>,
     pc: Expr<UInt<64>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
@@ -330,7 +327,7 @@ fn branch<SrcCount: KnownSize>(
 
 #[hdl]
 fn read_special(
-    mop: Expr<ReadSpecialMOp<UnitOutRegNum<DynSize>, DynSize>>,
+    mop: Expr<ReadSpecialMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>>,
     pc: Expr<UInt<64>>,
     flags_mode: Expr<FlagsMode>,
     src_values: Expr<Array<PRegValue, { COMMON_MOP_SRC_LEN }>>,
@@ -344,20 +341,18 @@ fn read_special(
 }
 
 #[hdl_module]
-pub fn alu_branch(config: &CpuConfig, unit_index: usize) {
+pub fn alu_branch(config: PhantomConst<CpuConfig>, unit_index: usize) {
     #[hdl]
     let cd: ClockDomain = m.input();
     #[hdl]
     let unit_to_reg_alloc: UnitToRegAlloc<
-        AluBranchMOp<UnitOutRegNum<DynSize>, DynSize>,
+        PhantomConst<CpuConfig>,
+        AluBranchMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>,
         (),
-        DynSize,
+    > = m.output(
-        DynSize,
+        UnitToRegAlloc[config][AluBranchMOp[UnitOutRegNum[config]][config.get().p_reg_num_width()]]
-        DynSize,
+            [()],
-    > = m.output(config.unit_to_reg_alloc(
+    );
-        AluBranchMOp[config.unit_out_reg_num()][config.p_reg_num_width()],
-        (),
-    ));
     #[hdl]
     let global_state: GlobalState = m.input();
 
@@ -375,10 +370,11 @@ pub fn alu_branch(config: &CpuConfig, unit_index: usize) {
     #[hdl]
     if let HdlSome(execute_start) = ReadyValid::firing_data(unit_base.execute_start) {
         #[hdl]
-        let ExecuteStart::<_> {
+        let ExecuteStart::<_, _> {
             mop,
             pc,
             src_values,
+            config: _,
         } = execute_start;
         #[hdl]
         match mop {
@@ -580,14 +576,14 @@ pub fn alu_branch(config: &CpuConfig, unit_index: usize) {
 
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
 pub struct AluBranch {
-    config: Interned<CpuConfig>,
+    config: PhantomConst<CpuConfig>,
     module: Interned<Module<alu_branch>>,
 }
 
 impl AluBranch {
-    pub fn new(config: &CpuConfig, unit_index: usize) -> Self {
+    pub fn new(config: PhantomConst<CpuConfig>, unit_index: usize) -> Self {
         Self {
-            config: config.intern(),
+            config,
             module: alu_branch(config, unit_index),
         }
     }
@@ -596,7 +592,7 @@ impl AluBranch {
 impl UnitTrait for AluBranch {
     type Type = alu_branch;
     type ExtraOut = ();
-    type MOp = AluBranchMOp<UnitOutRegNum<DynSize>, DynSize>;
+    type MOp = AluBranchMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>;
 
     fn ty(&self) -> Self::Type {
         self.module.io_ty()
@@ -616,7 +612,7 @@ impl UnitTrait for AluBranch {
 
     fn extract_mop(
         &self,
-        mop: Expr<RenamedMOp<UnitOutRegNum<DynSize>, DynSize>>,
+        mop: Expr<RenamedMOp<UnitOutRegNum<PhantomConst<CpuConfig>>, DynSize>>,
     ) -> Expr<HdlOption<Self::MOp>> {
         UnitMOp::alu_branch_mop(mop)
     }
@@ -628,7 +624,7 @@ impl UnitTrait for AluBranch {
     fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>> {
+    ) -> Expr<UnitToRegAlloc<PhantomConst<CpuConfig>, Self::MOp, Self::ExtraOut>> {
         this.unit_to_reg_alloc
     }
 

crates/cpu/src/unit/unit_base.rs

@@ -2,7 +2,7 @@
 // See Notices.txt for copyright information
 
 use crate::{
-    config::CpuConfig,
+    config::{CpuConfig, CpuConfigUnitCount, PhantomConstCpuConfig},
     instruction::{COMMON_MOP_SRC_LEN, MOpTrait, PRegNum, UnitNum, UnitOutRegNum},
     register::PRegValue,
     unit::{UnitCancelInput, UnitOutput, UnitOutputWrite},
@@ -15,13 +15,11 @@ use fayalite::{
     ty::StaticType,
     util::ready_valid::ReadyValid,
 };
-use std::marker::PhantomData;
 
-#[hdl]
+#[hdl(no_static)]
-pub struct UnitForwardingInfo<UnitNumWidth: Size, OutRegNumWidth: Size, UnitCount: Size> {
+pub struct UnitForwardingInfo<C: PhantomConstGet<CpuConfig>> {
-    pub unit_output_writes: ArrayType<HdlOption<UnitOutputWrite<OutRegNumWidth>>, UnitCount>,
+    pub unit_output_writes: ArrayType<HdlOption<UnitOutputWrite<C>>, CpuConfigUnitCount<C>>,
-    pub unit_reg_frees: ArrayType<HdlOption<UnitOutRegNum<OutRegNumWidth>>, UnitCount>,
+    pub unit_reg_frees: ArrayType<HdlOption<UnitOutRegNum<C>>, CpuConfigUnitCount<C>>,
-    pub _phantom: PhantomData<UnitNumWidth>,
 }
 
 #[hdl]
@@ -30,26 +28,18 @@ pub struct UnitInput<MOp: Type> {
     pub pc: UInt<64>,
 }
 
-#[hdl]
+#[hdl(no_static)]
-pub struct UnitToRegAlloc<
+pub struct UnitToRegAlloc<C: PhantomConstGet<CpuConfig>, MOp: Type, ExtraOut: Type> {
-    MOp: Type,
-    ExtraOut: Type,
-    UnitNumWidth: Size,
-    OutRegNumWidth: Size,
-    UnitCount: Size,
-> {
     #[hdl(flip)]
-    pub unit_forwarding_info: UnitForwardingInfo<UnitNumWidth, OutRegNumWidth, UnitCount>,
+    pub unit_forwarding_info: UnitForwardingInfo<C>,
     #[hdl(flip)]
     pub input: ReadyValid<UnitInput<MOp>>,
     #[hdl(flip)]
-    pub cancel_input: HdlOption<UnitCancelInput<OutRegNumWidth>>,
+    pub cancel_input: HdlOption<UnitCancelInput<C>>,
-    pub output: HdlOption<UnitOutput<OutRegNumWidth, ExtraOut>>,
+    pub output: HdlOption<UnitOutput<C, ExtraOut>>,
 }
 
-impl<MOp: Type, ExtraOut: Type, UnitNumWidth: Size, OutRegNumWidth: Size, UnitCount: Size>
+impl<C: PhantomConstCpuConfig, MOp: Type, ExtraOut: Type> UnitToRegAlloc<C, MOp, ExtraOut> {
-    UnitToRegAlloc<MOp, ExtraOut, UnitNumWidth, OutRegNumWidth, UnitCount>
-{
     pub fn mop_ty(self) -> MOp {
         self.input.data.HdlSome.mop
     }
@@ -58,16 +48,20 @@ impl<MOp: Type, ExtraOut: Type, UnitNumWidth: Size, OutRegNumWidth: Size, UnitCo
     }
 }
 
-#[hdl]
+#[hdl(no_static)]
-pub struct ExecuteStart<MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>>> {
+pub struct ExecuteStart<
+    C: PhantomConstGet<CpuConfig>,
+    MOp: Type + MOpTrait<DestReg = UnitOutRegNum<C>>,
+> {
     pub mop: MOp,
     pub pc: UInt<64>,
     pub src_values: Array<PRegValue, { COMMON_MOP_SRC_LEN }>,
+    pub config: C,
 }
 
-#[hdl]
+#[hdl(no_static)]
-pub struct ExecuteEnd<OutRegNumWidth: Size, ExtraOut> {
+pub struct ExecuteEnd<C: PhantomConstGet<CpuConfig>, ExtraOut> {
-    pub unit_output: UnitOutput<OutRegNumWidth, ExtraOut>,
+    pub unit_output: UnitOutput<C, ExtraOut>,
 }
 
 #[hdl]
@@ -240,10 +234,10 @@ impl InFlightOpsSummary<DynSize> {
 
 #[hdl_module]
 pub fn unit_base<
-    MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>, SrcRegWidth = DynSize>,
+    MOp: Type + MOpTrait<DestReg = UnitOutRegNum<PhantomConst<CpuConfig>>, SrcRegWidth = DynSize>,
     ExtraOut: Type,
 >(
-    config: &CpuConfig,
+    config: PhantomConst<CpuConfig>,
     unit_index: usize,
     mop_ty: MOp,
     extra_out_ty: ExtraOut,
@@ -251,17 +245,18 @@ pub fn unit_base<
     #[hdl]
     let cd: ClockDomain = m.input();
     #[hdl]
-    let unit_to_reg_alloc: UnitToRegAlloc<MOp, ExtraOut, DynSize, DynSize, DynSize> =
+    let unit_to_reg_alloc: UnitToRegAlloc<PhantomConst<CpuConfig>, MOp, ExtraOut> =
-        m.output(config.unit_to_reg_alloc(mop_ty, extra_out_ty));
+        m.output(UnitToRegAlloc[config][mop_ty][extra_out_ty]);
     #[hdl]
-    let execute_start: ReadyValid<ExecuteStart<MOp>> = m.output(ReadyValid[ExecuteStart[mop_ty]]);
+    let execute_start: ReadyValid<ExecuteStart<PhantomConst<CpuConfig>, MOp>> =
+        m.output(ReadyValid[ExecuteStart[config][mop_ty]]);
     #[hdl]
-    let execute_end: HdlOption<ExecuteEnd<DynSize, ExtraOut>> =
+    let execute_end: HdlOption<ExecuteEnd<PhantomConst<CpuConfig>, ExtraOut>> =
-        m.input(HdlOption[ExecuteEnd[config.out_reg_num_width][extra_out_ty]]);
+        m.input(HdlOption[ExecuteEnd[config][extra_out_ty]]);
 
     connect(execute_start.data, execute_start.ty().data.HdlNone());
 
-    let max_in_flight = config.unit_max_in_flight(unit_index).get();
+    let max_in_flight = config.get().unit_max_in_flight(unit_index).get();
     let in_flight_op_ty = InFlightOp[mop_ty];
     #[hdl]
     let in_flight_ops = reg_builder()
@@ -279,16 +274,15 @@ pub fn unit_base<
     );
 
     #[hdl]
-    let UnitForwardingInfo::<_, _, _> {
+    let UnitForwardingInfo::<_> {
         unit_output_writes,
         unit_reg_frees,
-        _phantom: _,
     } = unit_to_reg_alloc.unit_forwarding_info;
     #[hdl]
     let read_src_regs = wire(mop_ty.src_regs_ty());
     connect(
         read_src_regs,
-        repeat(config.p_reg_num().const_zero().cast_to_bits(), ConstUsize),
+        repeat(PRegNum[config].const_zero().cast_to_bits(), ConstUsize),
     );
     #[hdl]
     let read_src_values = wire();
@@ -297,7 +291,7 @@ pub fn unit_base<
     let input_src_regs = wire(mop_ty.src_regs_ty());
     connect(
         input_src_regs,
-        repeat(config.p_reg_num().const_zero().cast_to_bits(), ConstUsize),
+        repeat(PRegNum[config].const_zero().cast_to_bits(), ConstUsize),
     );
     #[hdl]
     let input_src_regs_valid = wire();
@@ -309,7 +303,7 @@ pub fn unit_base<
                 Bool,
                 SourceLocation::caller(),
             );
-            mem.depth(1 << config.out_reg_num_width);
+            mem.depth(1 << config.get().out_reg_num_width);
             mem
         })
         .collect();
@@ -319,11 +313,11 @@ pub fn unit_base<
             PRegValue,
             SourceLocation::caller(),
         );
-        unit_output_regs.depth(1 << config.out_reg_num_width);
+        unit_output_regs.depth(1 << config.get().out_reg_num_width);
 
         for src_index in 0..COMMON_MOP_SRC_LEN {
             let read_port = unit_output_regs.new_read_port();
-            let p_reg_num = read_src_regs[src_index].cast_bits_to(config.p_reg_num());
+            let p_reg_num = read_src_regs[src_index].cast_bits_to(PRegNum[config]);
             connect_any(read_port.addr, p_reg_num.unit_out_reg.value);
             connect(read_port.en, false);
             connect(read_port.clk, cd.clk);
@@ -336,7 +330,7 @@ pub fn unit_base<
 
         for src_index in 0..COMMON_MOP_SRC_LEN {
             let read_port = unit_output_regs_valid[unit_index].new_read_port();
-            let p_reg_num = input_src_regs[src_index].cast_bits_to(config.p_reg_num());
+            let p_reg_num = input_src_regs[src_index].cast_bits_to(PRegNum[config]);
             connect_any(read_port.addr, p_reg_num.unit_out_reg.value);
             connect(read_port.en, false);
             connect(read_port.clk, cd.clk);
@@ -367,8 +361,8 @@ pub fn unit_base<
             connect_any(ready_write_port.addr, unit_output_write.which.value);
             connect(ready_write_port.en, true);
             let p_reg_num = #[hdl]
-            PRegNum::<_, _> {
+            PRegNum::<_> {
-                unit_num: config.unit_num().from_index(unit_index),
+                unit_num: UnitNum[config].from_index(unit_index),
                 unit_out_reg: unit_output_write.which,
             };
             for src_index in 0..COMMON_MOP_SRC_LEN {
@@ -399,10 +393,11 @@ pub fn unit_base<
                 execute_start.data,
                 HdlSome(
                     #[hdl]
-                    ExecuteStart::<_> {
+                    ExecuteStart::<_, _> {
                         mop: in_flight_op.mop,
                         pc: in_flight_op.pc,
                         src_values: read_src_values,
+                        config,
                     },
                 ),
             );
@@ -425,7 +420,7 @@ pub fn unit_base<
         let input_mop_src_regs = wire(mop_ty.src_regs_ty());
         connect(
             input_mop_src_regs,
-            repeat(config.p_reg_num().const_zero().cast_to_bits(), ConstUsize),
+            repeat(PRegNum[config].const_zero().cast_to_bits(), ConstUsize),
         );
         MOp::connect_src_regs(mop, input_mop_src_regs);
         let src_ready_flags = wire_with_loc(
@@ -497,7 +492,7 @@ pub fn unit_base<
             );
             connect(
                 src_regs,
-                repeat(config.p_reg_num().const_zero().cast_to_bits(), ConstUsize),
+                repeat(PRegNum[config].const_zero().cast_to_bits(), ConstUsize),
             );
             MOp::connect_src_regs(mop, src_regs);
 
@@ -521,8 +516,8 @@ pub fn unit_base<
                         value: _,
                     } = unit_output_write;
                     let p_reg_num = #[hdl]
-                    PRegNum::<_, _> {
+                    PRegNum::<_> {
-                        unit_num: config.unit_num().from_index(unit_index),
+                        unit_num: UnitNum[config].from_index(unit_index),
                         unit_out_reg,
                     };
                     for src_index in 0..COMMON_MOP_SRC_LEN {

crates/cpu/tests/reg_alloc.rs

@@ -1,5 +1,6 @@
 // SPDX-License-Identifier: LGPL-3.0-or-later
 // See Notices.txt for copyright information
+#![cfg(todo)]
 
 use cpu::{
     config::{CpuConfig, UnitConfig},

crates/cpu/tests/rename_execute_retire.rs

@@ -0,0 +1,924 @@
+// SPDX-License-Identifier: LGPL-3.0-or-later
+// See Notices.txt for copyright information
+
+use cpu::{
+    config::{CpuConfig, CpuConfigFetchWidth, PhantomConstCpuConfig, UnitConfig},
+    instruction::{
+        AddSubMOp, AluBranchMOp, BranchMOp, CompareMOp, CompareMode, ConditionMode, LoadMOp,
+        LoadStoreConversion, LoadStoreWidth, MOp, MOpDestReg, MOpRegNum, MOpTrait, MoveRegMOp,
+        OutputIntegerMode, StoreMOp,
+    },
+    rename_execute_retire::{
+        MOpInstance, PostDecodeOutputInterface, RetireToNextPcInterface, rename_execute_retire,
+    },
+    unit::UnitKind,
+    util::array_vec::ArrayVec,
+};
+use fayalite::{
+    prelude::*,
+    sim::vcd::VcdWriterDecls,
+    ty::{OpaqueSimValue, StaticType},
+    util::RcWriter,
+};
+use std::{
+    cell::Cell,
+    collections::{BTreeMap, VecDeque},
+    fmt::{self, Write},
+    mem,
+    num::NonZeroUsize,
+};
+
+fn zeroed<T: Type>(ty: T) -> SimValue<T> {
+    SimValue::from_opaque(
+        ty,
+        OpaqueSimValue::from_bits(UInt::new(ty.canonical().bit_width()).zero()),
+    )
+}
+
+fn signed_hex(v: impl Into<i64>) -> impl fmt::Display {
+    let v = v.into();
+    fmt::from_fn(move |f| {
+        if v < 0 {
+            f.write_char('-')?;
+        }
+        write!(f, "{:#x}", v.unsigned_abs())
+    })
+}
+
+struct RandomState {
+    state: Cell<u64>,
+}
+
+impl RandomState {
+    fn random_u64(&self, key: u32) -> u64 {
+        let state = self.state.replace(self.state.get().wrapping_add(1));
+        // make a pseudo-random number deterministically based on state and key
+        let mut random = state
+            .wrapping_add(1)
+            .wrapping_mul(0x39FF446D8BFB75BB) // random prime
+            .rotate_left(32)
+            .wrapping_mul(0x73161B54984B1C21) // random prime
+            .rotate_right(60);
+        random ^= key as u64;
+        random
+            .wrapping_mul(0x39FF446D8BFB75BB) // random prime
+            .rotate_left(32)
+            .wrapping_mul(0x73161B54984B1C21) // random prime
+            .rotate_right(60)
+    }
+}
+
+const START_PC: u64 = 0x0; // match microwatt's reset pc
+
+#[derive(Clone, Debug)]
+struct Insn<MOps = Vec<SimValue<self::MOp>>> {
+    size_in_bytes: u8,
+    power_isa: String,
+    mops: MOps,
+}
+
+enum LazyMOps {
+    MOps(Vec<SimValue<self::MOp>>),
+    Lazy(Box<dyn FnOnce(&[InsnsBuilderLabelState]) -> Vec<SimValue<self::MOp>>>),
+}
+
+impl Insn<LazyMOps> {
+    fn new<I: IntoIterator<Item: ToSimValue<Type = MOp>>>(
+        size_in_bytes: u8,
+        power_isa: String,
+        mops: I,
+    ) -> Self {
+        Self {
+            size_in_bytes,
+            power_isa,
+            mops: LazyMOps::MOps(mops.into_iter().map(|mop| mop.into_sim_value()).collect()),
+        }
+    }
+    fn new_lazy<I: IntoIterator<Item: ToSimValue<Type = MOp>>>(
+        size_in_bytes: u8,
+        power_isa: String,
+        lazy_mops: impl FnOnce(&[InsnsBuilderLabelState]) -> I + 'static,
+    ) -> Self {
+        Self {
+            size_in_bytes,
+            power_isa,
+            mops: LazyMOps::Lazy(Box::new(|labels| {
+                lazy_mops(labels)
+                    .into_iter()
+                    .map(|mop| mop.into_sim_value())
+                    .collect()
+            })),
+        }
+    }
+}
+
+struct Insns {
+    insns: BTreeMap<u64, Insn>,
+}
+
+impl Insns {
+    fn get_mop_mut(&mut self, pc: u64, mop_index: usize) -> Option<&mut SimValue<MOp>> {
+        self.insns.get_mut(&pc)?.mops.get_mut(mop_index)
+    }
+}
+
+impl fmt::Debug for Insns {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        f.debug_map()
+            .entries(
+                self.insns
+                    .iter()
+                    .map(|(k, v)| (fmt::from_fn(move |f| write!(f, "{k:#x}")), v)),
+            )
+            .finish()
+    }
+}
+
+struct InsnsBuilderLabelState {
+    name: &'static str,
+    pc: Option<u64>,
+    location: &'static std::panic::Location<'static>,
+}
+
+impl InsnsBuilderLabelState {
+    fn pc(&self) -> u64 {
+        match self.pc {
+            Some(pc) => pc,
+            None => panic!("label not defined {}: at: {}", self.name, self.location),
+        }
+    }
+}
+
+#[derive(Clone, Copy, Debug)]
+struct InsnsBuilderLabel(usize);
+
+struct InsnsBuilder {
+    labels: Vec<InsnsBuilderLabelState>,
+    insns: BTreeMap<u64, Insn<LazyMOps>>,
+    pc: u64,
+}
+
+impl InsnsBuilder {
+    fn new() -> Self {
+        Self {
+            labels: Vec::new(),
+            insns: BTreeMap::new(),
+            pc: START_PC,
+        }
+    }
+    #[track_caller]
+    fn new_label(&mut self, name: &'static str) -> InsnsBuilderLabel {
+        let retval = InsnsBuilderLabel(self.labels.len());
+        self.labels.push(InsnsBuilderLabelState {
+            pc: None,
+            name,
+            location: std::panic::Location::caller(),
+        });
+        retval
+    }
+    #[track_caller]
+    fn define_label(&mut self, label: InsnsBuilderLabel) {
+        let label_state = &mut self.labels[label.0];
+        if label_state.pc.is_some() {
+            panic!("label already defined at: {}", label_state.location);
+        }
+        label_state.pc = Some(self.pc);
+        label_state.location = std::panic::Location::caller();
+    }
+    #[track_caller]
+    fn new_defined_label(&mut self, name: &'static str) -> InsnsBuilderLabel {
+        let label = self.new_label(name);
+        self.define_label(label);
+        label
+    }
+    #[track_caller]
+    fn add_insn(&mut self, insn: Insn<LazyMOps>) {
+        let pc = self.pc;
+        let next_pc = pc.wrapping_add(insn.size_in_bytes as u64);
+        if self.insns.insert(self.pc, insn).is_some() {
+            panic!("instruction already stored at {pc:#x}");
+        }
+        self.pc = next_pc;
+    }
+    fn build(self) -> Insns {
+        Insns {
+            insns: self
+                .insns
+                .into_iter()
+                .map(|(pc, insn)| {
+                    let Insn {
+                        size_in_bytes,
+                        power_isa,
+                        mops,
+                    } = insn;
+                    (
+                        pc,
+                        Insn {
+                            size_in_bytes,
+                            power_isa,
+                            mops: match mops {
+                                LazyMOps::MOps(mops) => mops,
+                                LazyMOps::Lazy(f) => f(&self.labels),
+                            },
+                        },
+                    )
+                })
+                .collect(),
+        }
+    }
+    fn set_pc(&mut self, pc: u64) {
+        self.pc = pc;
+    }
+    fn power_isa_b(&mut self, target: InsnsBuilderLabel) {
+        let pc = self.pc;
+        self.add_insn(Insn::new_lazy(
+            4,
+            format!("b {}", self.labels[target.0].name),
+            move |labels| {
+                [BranchMOp::branch_i(
+                    MOpDestReg::new([], []),
+                    MOpRegNum::const_zero().value,
+                    labels[target.0]
+                        .pc()
+                        .wrapping_sub(pc)
+                        .cast_to_static::<SInt<_>>(),
+                    true,
+                    false,
+                    false,
+                )]
+            },
+        ));
+    }
+    fn power_isa_blr(&mut self) {
+        self.add_insn(Insn::new(
+            4,
+            format!("blr"),
+            [BranchMOp::branch_i(
+                MOpDestReg::new([], []),
+                MOpRegNum::power_isa_lr_reg().value,
+                0i8.cast_to_static::<SInt<_>>(),
+                false,
+                false,
+                true,
+            )],
+        ));
+    }
+    fn power_isa_bgt(&mut self, target: InsnsBuilderLabel) {
+        let pc = self.pc;
+        self.add_insn(Insn::new_lazy(
+            4,
+            format!("bgt {}", self.labels[target.0].name),
+            move |labels| {
+                [BranchMOp::branch_cond_ctr(
+                    MOpDestReg::new([], []),
+                    [
+                        MOpRegNum::power_isa_cr_reg_imm(0).value,
+                        MOpRegNum::const_zero().value,
+                        MOpRegNum::const_zero().value,
+                    ],
+                    labels[target.0]
+                        .pc()
+                        .wrapping_sub(pc)
+                        .cast_to_static::<SInt<_>>(),
+                    false,
+                    ConditionMode.SGt(),
+                    false,
+                    true,
+                    false,
+                    false,
+                )]
+            },
+        ));
+    }
+    fn power_isa_bl(&mut self, target: InsnsBuilderLabel) {
+        let pc = self.pc;
+        self.add_insn(Insn::new_lazy(
+            4,
+            format!("bl {}", self.labels[target.0].name),
+            move |labels| {
+                [BranchMOp::branch_i(
+                    MOpDestReg::new([MOpRegNum::power_isa_lr_reg()], []),
+                    MOpRegNum::const_zero().value,
+                    labels[target.0]
+                        .pc()
+                        .wrapping_sub(pc)
+                        .cast_to_static::<SInt<_>>(),
+                    true,
+                    true,
+                    false,
+                )]
+            },
+        ));
+    }
+    fn power_isa_add(&mut self, dest: usize, src1: usize, src2: usize) {
+        self.add_insn(Insn::new(
+            4,
+            format!("add {dest}, {src1}, {src2}"),
+            [AddSubMOp::add_sub(
+                MOpDestReg::new_sim(&[MOpRegNum::power_isa_gpr_reg_num(dest)], &[]),
+                [
+                    MOpRegNum::power_isa_gpr_reg_imm(src1).value,
+                    MOpRegNum::power_isa_gpr_reg_imm(src2).value,
+                    MOpRegNum::const_zero().value,
+                ],
+                0i8.cast_to_static::<SInt<_>>(),
+                OutputIntegerMode.Full64(),
+                false,
+                false,
+                false,
+                false,
+            )],
+        ));
+    }
+    fn power_isa_addi(&mut self, dest: usize, src: usize, imm: i16) {
+        self.add_insn(Insn::new(
+            4,
+            format!("addi {dest}, {src}, {}", signed_hex(imm)),
+            [AddSubMOp::add_sub_i(
+                MOpDestReg::new_sim(&[MOpRegNum::power_isa_gpr_reg_num(dest)], &[]),
+                [
+                    MOpRegNum::power_isa_gpr_or_zero_reg_imm(src).value,
+                    MOpRegNum::const_zero().value,
+                ],
+                imm.cast_to_static::<SInt<_>>(),
+                OutputIntegerMode.Full64(),
+                false,
+                false,
+                false,
+                false,
+            )],
+        ));
+    }
+    fn power_isa_cmpldi(&mut self, dest: usize, src: usize, imm: u16) {
+        self.add_insn(Insn::new(
+            4,
+            format!("cmpldi {dest}, {src}, {imm:#x}"),
+            [CompareMOp::compare_i(
+                MOpDestReg::new_sim(&[MOpRegNum::power_isa_cr_reg_num(dest)], &[]),
+                [MOpRegNum::power_isa_gpr_reg_imm(src).value],
+                imm.cast_to_static::<SInt<_>>(),
+                OutputIntegerMode.Full64(),
+                CompareMode.U64(),
+            )],
+        ));
+    }
+    fn power_isa_ld(&mut self, dest: usize, src: usize, disp: i16) {
+        self.add_insn(Insn::new(
+            4,
+            format!("ld {dest}, {}({src})", signed_hex(disp)),
+            [
+                AddSubMOp::add_sub_i(
+                    MOpDestReg::new_sim(&[MOpRegNum::POWER_ISA_TEMP_REG_NUM], &[]),
+                    [
+                        MOpRegNum::power_isa_gpr_or_zero_reg_imm(src).value,
+                        MOpRegNum::const_zero().value,
+                    ],
+                    disp.cast_to_static::<SInt<_>>(),
+                    OutputIntegerMode.Full64(),
+                    false,
+                    false,
+                    false,
+                    false,
+                ),
+                LoadMOp::load(
+                    MOpDestReg::new_sim(&[MOpRegNum::power_isa_gpr_reg_num(dest)], &[]),
+                    [MOpRegNum::power_isa_temp_reg().value],
+                    LoadStoreWidth.Width64Bit(),
+                    LoadStoreConversion.ZeroExt(),
+                ),
+            ],
+        ));
+    }
+    fn power_isa_std(&mut self, value_src: usize, addr_src: usize, disp: i16) {
+        self.add_insn(Insn::new(
+            4,
+            format!("std {value_src}, {}({addr_src})", signed_hex(disp)),
+            [
+                AddSubMOp::add_sub_i(
+                    MOpDestReg::new_sim(&[MOpRegNum::POWER_ISA_TEMP_REG_NUM], &[]),
+                    [
+                        MOpRegNum::power_isa_gpr_or_zero_reg_imm(addr_src).value,
+                        MOpRegNum::const_zero().value,
+                    ],
+                    disp.cast_to_static::<SInt<_>>(),
+                    OutputIntegerMode.Full64(),
+                    false,
+                    false,
+                    false,
+                    false,
+                ),
+                StoreMOp::store(
+                    MOpDestReg::new_sim(&[], &[]),
+                    [
+                        MOpRegNum::power_isa_temp_reg().value,
+                        MOpRegNum::power_isa_gpr_reg_imm(value_src).value,
+                    ],
+                    LoadStoreWidth.Width64Bit(),
+                    LoadStoreConversion.ZeroExt(),
+                ),
+            ],
+        ));
+    }
+    fn power_isa_mflr(&mut self, dest: usize) {
+        self.add_insn(Insn::new(
+            4,
+            format!("mflr {dest}"),
+            [MoveRegMOp::move_reg(
+                MOpDestReg::new_sim(&[MOpRegNum::power_isa_gpr_reg_num(dest)], &[]),
+                [MOpRegNum::power_isa_lr_reg().value],
+                0i8.cast_to_static::<SInt<_>>(),
+            )],
+        ));
+    }
+    fn power_isa_mtlr(&mut self, src: usize) {
+        self.add_insn(Insn::new(
+            4,
+            format!("mtlr {src}"),
+            [MoveRegMOp::move_reg(
+                MOpDestReg::new_sim(&[MOpRegNum::POWER_ISA_LR_REG_NUM], &[]),
+                [MOpRegNum::power_isa_gpr_reg_imm(src).value],
+                0i8.cast_to_static::<SInt<_>>(),
+            )],
+        ));
+    }
+    fn power_isa_mr(&mut self, dest: usize, src: usize) {
+        self.add_insn(Insn::new(
+            4,
+            format!("mr {dest}, {src}"),
+            [MoveRegMOp::move_reg(
+                MOpDestReg::new_sim(&[MOpRegNum::power_isa_gpr_reg_num(dest)], &[]),
+                [MOpRegNum::power_isa_gpr_reg_imm(src).value],
+                0i8.cast_to_static::<SInt<_>>(),
+            )],
+        ));
+    }
+}
+
+trait MakeInsns {
+    fn make_insns() -> Insns;
+}
+
+struct FibonacciInsns;
+
+impl MakeInsns for FibonacciInsns {
+    fn make_insns() -> Insns {
+        let mut b = InsnsBuilder::new();
+
+        let fib = b.new_label("fib");
+        let fib_small = b.new_label("fib_small");
+        b.power_isa_ld(3, 0, -16); // load input
+        b.power_isa_addi(1, 0, 0x4000); // setup stack pointer
+        b.power_isa_bl(fib);
+        b.power_isa_std(3, 0, -16); // store output
+        let done = b.new_defined_label("done");
+        b.power_isa_b(done);
+
+        b.set_pc(0x1000);
+        b.define_label(fib);
+        b.power_isa_cmpldi(0, 3, 1);
+        b.power_isa_bgt(fib_small); // if input > 1 goto fib_small
+
+        // push stack frame
+        b.power_isa_mflr(0);
+        b.power_isa_addi(1, 1, -16);
+        b.power_isa_std(0, 1, 0);
+        b.power_isa_std(30, 1, 8);
+
+        b.power_isa_addi(30, 3, -2);
+        b.power_isa_addi(3, 3, -1);
+        b.power_isa_bl(fib); // fib(input - 1)
+        b.power_isa_mr(2, 3);
+        b.power_isa_mr(3, 30);
+        b.power_isa_mr(30, 2);
+        b.power_isa_bl(fib); // fib(input - 2)
+        b.power_isa_add(3, 3, 30); // set output to fib(input - 1) + fib(input - 2)
+
+        // pop stack frame and return
+        b.power_isa_ld(0, 1, 0);
+        b.power_isa_ld(30, 1, 8);
+        b.power_isa_addi(1, 1, 16);
+        b.power_isa_mtlr(0);
+        b.power_isa_blr();
+
+        b.define_label(fib_small);
+        b.power_isa_addi(3, 0, 1); // set output to 1
+        b.power_isa_blr(); // return
+
+        b.build()
+    }
+}
+
+#[hdl(no_static)]
+struct MockNextPcDebugState<C: PhantomConstGet<CpuConfig>> {
+    next_pc: UInt<64>,
+    next_mop_index: UInt<8>,
+    next_br_pred_state: BrPredDebugState,
+    next_id: UInt<64>,
+    fetch_block_id: UInt<64>,
+    fetch_queue: ArrayVec<(BrPredDebugState, MOpInstance<MOp>), CpuConfigFetchWidth<C>>,
+    config: C,
+}
+
+#[hdl]
+struct BrPredDebugState {
+    call_stack: ArrayVec<UInt<64>, ConstUsize<{ BrPredState::CALL_STACK_CAPACITY }>>,
+}
+
+#[derive(Copy, Clone)]
+struct BrPredState {
+    call_stack: [u64; Self::CALL_STACK_CAPACITY],
+    call_stack_len: usize,
+}
+
+impl BrPredState {}
+
+impl fmt::Debug for BrPredState {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        let Self {
+            call_stack,
+            call_stack_len,
+        } = self;
+        f.debug_struct("BrPredState")
+            .field(
+                "call_stack",
+                &fmt::from_fn(|f| {
+                    let mut debug_list = f.debug_list();
+                    for pc in &call_stack[..*call_stack_len] {
+                        debug_list.entry(&fmt::from_fn(|f| write!(f, "{pc:#x}")));
+                    }
+                    debug_list.finish()
+                }),
+            )
+            .finish()
+    }
+}
+
+impl BrPredState {
+    const CALL_STACK_CAPACITY: usize = 8;
+    fn new() -> Self {
+        Self {
+            call_stack: [0; _],
+            call_stack_len: 0,
+        }
+    }
+    fn predict_call(&mut self, return_pc: u64) {
+        if self.call_stack_len == Self::CALL_STACK_CAPACITY {
+            self.call_stack.copy_within(1.., 0);
+            *self.call_stack.last_mut().expect("known to be non-zero") = return_pc;
+        } else {
+            self.call_stack[self.call_stack_len] = return_pc;
+            self.call_stack_len += 1;
+        }
+    }
+    fn predict_ret(&mut self) -> Option<u64> {
+        if self.call_stack_len > 0 {
+            self.call_stack_len -= 1;
+            Some(mem::replace(&mut self.call_stack[self.call_stack_len], 0))
+        } else {
+            None
+        }
+    }
+    fn predict_branch<SrcCount: KnownSize>(
+        &mut self,
+        mop: &SimValue<BranchMOp<MOpDestReg, ConstUsize<{ MOpRegNum::WIDTH }>, SrcCount>>,
+        branch_pc: u64,
+        fallthrough_pc: u64,
+    ) -> u64 {
+        if *mop.is_ret {
+            if let Some(retval) = self.predict_ret() {
+                return retval;
+            }
+        }
+        let mut src_regs = [MOpRegNum::CONST_ZERO_REG_NUM; 3];
+        MOpTrait::for_each_src_reg_sim_ref(mop, &mut |reg, index| {
+            src_regs[index] = reg.cast_to_static::<UInt<32>>().as_int();
+        });
+        if src_regs[1] != MOpRegNum::CONST_ZERO_REG_NUM {
+            // indirect branch -- this test doesn't implement predicting them, so just use the fallthrough_pc
+            return fallthrough_pc;
+        }
+        let target_pc = (*mop.common.imm)
+            .cast_to_static::<UInt<64>>()
+            .as_int()
+            .wrapping_add(if *mop.pc_relative { branch_pc } else { 0 });
+        if *mop.is_call {
+            self.predict_call(fallthrough_pc);
+        }
+        if src_regs[0] != MOpRegNum::CONST_ZERO_REG_NUM
+            || src_regs[2] != MOpRegNum::CONST_ZERO_REG_NUM
+        {
+            // conditional branch -- this test just predicts taken for backward branches and not taken for forward branches
+            if target_pc < fallthrough_pc {
+                target_pc
+            } else {
+                fallthrough_pc
+            }
+        } else {
+            target_pc
+        }
+    }
+    #[hdl]
+    fn debug_state(&self) -> SimValue<BrPredDebugState> {
+        let Self {
+            call_stack,
+            call_stack_len,
+        } = self;
+        #[hdl(sim)]
+        BrPredDebugState {
+            call_stack: BrPredDebugState
+                .call_stack
+                .from_iter_sim(0u64, &call_stack[..*call_stack_len])
+                .expect("known to fit"),
+        }
+    }
+}
+
+struct MockNextPcState<'a, C: PhantomConstCpuConfig> {
+    random_state: &'a RandomState,
+    insns: &'a Insns,
+    next_pc: u64,
+    next_mop_index: usize,
+    next_br_pred_state: BrPredState,
+    next_id: u64,
+    fetch_block_id: u64,
+    fetch_queue: VecDeque<(BrPredState, SimValue<MOpInstance<MOp>>)>,
+    config: C,
+}
+
+impl<'a, C: PhantomConstCpuConfig> MockNextPcState<'a, C> {
+    fn new(random_state: &'a RandomState, insns: &'a Insns, config: C) -> Self {
+        Self {
+            random_state,
+            insns,
+            next_pc: START_PC,
+            next_mop_index: 0,
+            next_br_pred_state: BrPredState::new(),
+            next_id: 0,
+            fetch_block_id: 0,
+            fetch_queue: VecDeque::with_capacity(config.get().fetch_width.get()),
+            config,
+        }
+    }
+    #[hdl]
+    fn debug_state(&self) -> SimValue<MockNextPcDebugState<C>> {
+        let Self {
+            random_state: _,
+            insns: _,
+            config,
+            next_pc,
+            next_mop_index,
+            next_br_pred_state,
+            next_id,
+            fetch_block_id,
+            fetch_queue,
+        } = self;
+        let ty = MockNextPcDebugState[*config];
+        #[hdl(sim)]
+        MockNextPcDebugState::<_> {
+            next_pc,
+            next_mop_index: *next_mop_index as u8,
+            next_br_pred_state: next_br_pred_state.debug_state(),
+            next_id,
+            fetch_block_id,
+            fetch_queue: ty
+                .fetch_queue
+                .from_iter_sim(
+                    zeroed(StaticType::TYPE),
+                    fetch_queue
+                        .iter()
+                        .map(|(br_pred_state, mop)| (br_pred_state.debug_state(), mop)),
+                )
+                .expect("known to fit"),
+            config,
+        }
+    }
+    #[hdl]
+    fn try_push_fetch_queue(&mut self) -> Result<(), ()> {
+        if self.fetch_queue.len() >= self.config.get().fetch_width.get() {
+            return Err(());
+        }
+        let insn = self.insns.insns.get(&self.next_pc).ok_or(())?;
+        let fallthrough_pc = self.next_pc.wrapping_add(insn.size_in_bytes.into());
+        let is_last_mop_index = self.next_mop_index + 1 >= insn.mops.len();
+        let mut predicted_next_pc = if is_last_mop_index {
+            fallthrough_pc
+        } else {
+            self.next_pc
+        };
+        let prev_br_pred_state = self.next_br_pred_state;
+        let mop = &insn.mops[self.next_mop_index];
+        #[hdl(sim)]
+        if let MOp::AluBranch(mop) = mop {
+            #[hdl(sim)]
+            match mop {
+                AluBranchMOp::<_, _>::Branch(mop) => {
+                    predicted_next_pc =
+                        self.next_br_pred_state
+                            .predict_branch(mop, self.next_pc, fallthrough_pc);
+                }
+                AluBranchMOp::<_, _>::BranchI(mop) => {
+                    predicted_next_pc =
+                        self.next_br_pred_state
+                            .predict_branch(mop, self.next_pc, fallthrough_pc);
+                }
+                _ => {}
+            }
+        }
+        let mop = #[hdl(sim)]
+        MOpInstance::<_> {
+            fetch_block_id: self.fetch_block_id.cast_to_static::<UInt<_>>(),
+            id: self.next_id.cast_to_static::<UInt<_>>(),
+            pc: self.next_pc,
+            predicted_next_pc,
+            size_in_bytes: insn.size_in_bytes.cast_to_static::<UInt<_>>(),
+            is_first_mop_in_insn: self.next_mop_index == 0,
+            mop,
+        };
+        println!("pushed to fetch queue: {mop:?}");
+        self.fetch_queue.push_back((prev_br_pred_state, mop));
+        if is_last_mop_index {
+            self.next_mop_index = 0;
+            self.next_pc = predicted_next_pc;
+        } else {
+            self.next_mop_index += 1;
+        }
+        Ok(())
+    }
+    fn peek_post_decode_output_insns(
+        &self,
+    ) -> SimValue<ArrayVec<MOpInstance<MOp>, CpuConfigFetchWidth<C>>> {
+        let ty = ArrayVec[MOpInstance[MOp]][CpuConfigFetchWidth[self.config]];
+        let peek_size = self.random_state.random_u64(u32::from_le_bytes(*b"pdoi")) as usize
+            % (ty.capacity() + 1);
+        ty.from_iter_sim(
+            zeroed(MOpInstance[MOp]),
+            self.fetch_queue.iter().map(|(_, v)| v).take(peek_size),
+        )
+        .ok()
+        .expect("known to fit")
+    }
+}
+
+#[hdl_module(extern)]
+fn mock_next_pc<#[hdl(skip)] MI: MakeInsns>(config: PhantomConst<CpuConfig>) {
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let post_decode_output: PostDecodeOutputInterface<PhantomConst<CpuConfig>> =
+        m.output(PostDecodeOutputInterface[config]);
+    #[hdl]
+    let from_retire: RetireToNextPcInterface<PhantomConst<CpuConfig>> =
+        m.input(RetireToNextPcInterface[config]);
+    #[hdl]
+    let debug_state: MockNextPcDebugState<PhantomConst<CpuConfig>> =
+        m.output(MockNextPcDebugState[config]);
+    m.register_clock_for_past(cd.clk);
+    m.extern_module_simulation_fn(
+        (cd, post_decode_output, from_retire, debug_state),
+        |args, mut sim| async move {
+            let (cd, post_decode_output, from_retire, debug_state) = args;
+            // intentionally have a different sequence each time we're reset
+            let random_state = RandomState {
+                state: Cell::new(0),
+            };
+            let insns = MI::make_insns();
+            sim.resettable(
+                cd,
+                async |mut sim| {
+                    sim.write(
+                        post_decode_output.insns,
+                        post_decode_output
+                            .ty()
+                            .insns
+                            .new_sim(zeroed(StaticType::TYPE)),
+                    )
+                    .await;
+                    sim.write(
+                        post_decode_output.cancel.data,
+                        #[hdl(sim)]
+                        HdlNone(),
+                    )
+                    .await;
+                    sim.write(from_retire.inner.ready, false).await;
+                },
+                |sim, ()| {
+                    run_fn(
+                        cd,
+                        post_decode_output,
+                        from_retire,
+                        debug_state,
+                        &random_state,
+                        &insns,
+                        sim,
+                    )
+                },
+            )
+            .await;
+        },
+    );
+    #[hdl]
+    async fn run_fn(
+        cd: Expr<ClockDomain>,
+        post_decode_output: Expr<PostDecodeOutputInterface<PhantomConst<CpuConfig>>>,
+        from_retire: Expr<RetireToNextPcInterface<PhantomConst<CpuConfig>>>,
+        debug_state: Expr<MockNextPcDebugState<PhantomConst<CpuConfig>>>,
+        random_state: &RandomState,
+        insns: &Insns,
+        mut sim: ExternModuleSimulationState,
+    ) {
+        let config = post_decode_output.ty().config;
+        let mut state = MockNextPcState::new(random_state, insns, config);
+        loop {
+            while state.try_push_fetch_queue().is_ok() {}
+            state.fetch_block_id += 1;
+            let post_decode_output_insns = state.peek_post_decode_output_insns();
+            sim.write(post_decode_output.insns, &post_decode_output_insns)
+                .await;
+            sim.write(
+                post_decode_output.cancel.data,
+                #[hdl(sim)]
+                HdlNone(),
+            )
+            .await;
+            sim.write(from_retire.inner.ready, false).await;
+            sim.write(debug_state, state.debug_state()).await;
+            sim.wait_for_clock_edge(cd.clk).await;
+            let ready = *sim.read_past(post_decode_output.ready, cd.clk).await;
+            for _ in ArrayVec::elements_sim_ref(&post_decode_output_insns)
+                .iter()
+                .take(ready)
+            {
+                let Some((prev_br_pred_state, started)) = state.fetch_queue.pop_front() else {
+                    unreachable!();
+                };
+                println!("started: {started:?}\nprev_br_pred_state: {prev_br_pred_state:?}");
+            }
+        }
+    }
+}
+
+#[hdl_module]
+fn rename_execute_retire_test_harness<#[hdl(skip)] MI: MakeInsns>(config: PhantomConst<CpuConfig>) {
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let next_pc = instance(mock_next_pc::<MI>(config));
+    connect(next_pc.cd, cd);
+    #[hdl]
+    let dut = instance(rename_execute_retire(config));
+    connect(dut.cd, cd);
+    connect(dut.from_post_decode, next_pc.post_decode_output);
+    connect(next_pc.from_retire, dut.to_next_pc);
+}
+
+#[hdl]
+#[test]
+fn test_rename_execute_retire_fibonacci() {
+    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::LoadStore),
+            UnitConfig::new(UnitKind::TransformedMove),
+        ],
+        NonZeroUsize::new(20).unwrap(),
+    );
+    config.fetch_width = NonZeroUsize::new(3).unwrap();
+    let m = rename_execute_retire_test_harness::<FibonacciInsns>(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..50 {
+        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);
+    }
+    panic!();
+    // 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_fibonacci.vcd") {
+        panic!();
+    }
+}