WIP implementing unit_base

crates/cpu/src/config.rs

@@ -1,8 +1,11 @@
 // SPDX-License-Identifier: LGPL-3.0-or-later
 // See Notices.txt for copyright information
 use crate::{
-    instruction::{PRegNum, UnitNum, UnitOutRegNum, CONST_ZERO_UNIT_NUM},
-    unit::{unit_base::UnitForwardingInfo, UnitCancelInput, UnitKind, UnitMOp, UnitOutputWrite},
+    instruction::{MOpTrait, PRegNum, UnitNum, UnitOutRegNum, CONST_ZERO_UNIT_NUM},
+    unit::{
+        unit_base::{UnitForwardingInfo, UnitToRegAlloc},
+        UnitCancelInput, UnitKind, UnitMOp, UnitOutputWrite,
+    },
 };
 use fayalite::prelude::*;
 use std::num::NonZeroUsize;
@@ -95,4 +98,20 @@ impl CpuConfig {
             .max_in_flight
             .unwrap_or(self.default_unit_max_in_flight)
     }
+    pub fn unit_to_reg_alloc<
+        MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>, SrcRegWidth = DynSize>,
+        ExtraOut: Type,
+    >(
+        &self,
+        mop_ty: MOp,
+        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()]
+    }
 }

crates/cpu/src/instruction.rs

@@ -1,7 +1,11 @@
 // SPDX-License-Identifier: LGPL-3.0-or-later
 // See Notices.txt for copyright information
 use crate::{unit::UnitMOp, util::range_u32_len};
-use fayalite::{expr::ops::ArrayLiteral, intern::Interned, prelude::*};
+use fayalite::{
+    expr::ops::{ArrayLiteral, ExprPartialEq},
+    intern::Interned,
+    prelude::*,
+};
 use std::{fmt, marker::PhantomData, ops::Range};
 
 pub mod power_isa;
@@ -12,10 +16,26 @@ pub trait MOpTrait: Type {
     type SrcRegWidth: Size;
     fn dest_reg_ty(self) -> Self::DestReg;
     fn dest_reg(input: impl ToExpr<Type = Self>) -> Expr<Self::DestReg>;
+    fn src_reg_width(self) -> <Self::SrcRegWidth as Size>::SizeType;
+    fn src_reg_ty(self) -> UIntType<Self::SrcRegWidth> {
+        UInt[self.src_reg_width()]
+    }
+    fn src_regs_ty(self) -> Array<UIntType<Self::SrcRegWidth>, { COMMON_MOP_SRC_LEN }> {
+        Array[self.src_reg_ty()][ConstUsize::<{ COMMON_MOP_SRC_LEN }>]
+    }
     fn for_each_src_reg(
         input: impl ToExpr<Type = Self>,
         f: &mut impl FnMut(Expr<UIntType<Self::SrcRegWidth>>, usize),
     );
+    fn connect_src_regs(
+        input: impl ToExpr<Type = Self>,
+        src_regs: impl ToExpr<Type = Array<UIntType<Self::SrcRegWidth>, { COMMON_MOP_SRC_LEN }>>,
+    ) {
+        let src_regs = src_regs.to_expr();
+        Self::for_each_src_reg(input.to_expr(), &mut |src_reg, index| {
+            connect(src_regs[index], src_reg);
+        });
+    }
     fn mapped_ty<NewDestReg: Type, NewSrcRegWidth: Size>(
         self,
         new_dest_reg: NewDestReg,
@@ -67,6 +87,9 @@ impl<T: CommonMOpTrait> MOpTrait for T {
     fn dest_reg(input: impl ToExpr<Type = Self>) -> Expr<Self::DestReg> {
         T::common_mop(input).dest
     }
+    fn src_reg_width(self) -> <Self::SrcRegWidth as Size>::SizeType {
+        self.common_mop_ty().src.element().width
+    }
     fn for_each_src_reg(
         input: impl ToExpr<Type = Self>,
         f: &mut impl FnMut(Expr<UIntType<Self::SrcRegWidth>>, usize),
@@ -129,13 +152,23 @@ pub enum OutputIntegerMode {
     SignExt8,
 }
 
+impl ExprPartialEq<Self> for OutputIntegerMode {
+    fn cmp_eq(lhs: Expr<Self>, rhs: Expr<Self>) -> Expr<Bool> {
+        lhs.cast_to_bits().cmp_eq(rhs.cast_to_bits())
+    }
+
+    fn cmp_ne(lhs: Expr<Self>, rhs: Expr<Self>) -> Expr<Bool> {
+        lhs.cast_to_bits().cmp_ne(rhs.cast_to_bits())
+    }
+}
+
 pub const MOP_IMM_WIDTH: usize = 34;
 pub const MOP_MIN_REG_WIDTH: usize = 8;
 pub const COMMON_MOP_SRC_LEN: usize = 3;
 pub const COMMON_MOP_MIN_SRC_LEN_WITH_FULL_IMM: usize = 2;
 pub const COMMON_MOP_IMM_LOW_WIDTH: usize = CommonMOpWithMaxSrcCount::IMM_WIDTH - 1;
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct CommonMOp<PrefixPad: KnownSize, DestReg: Type, SrcRegWidth: Size, SrcCount: KnownSize> {
     pub prefix_pad: UIntType<PrefixPad>,
     pub dest: DestReg,
@@ -181,7 +214,7 @@ impl<PrefixPad: KnownSize, DestReg: Type, SrcRegWidth: Size, SrcCount: KnownSize
     }
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct CommonMOpImmParts<ImmInSrcCount: Size> {
     // fields must be in this exact order
     pub imm_low: UInt<{ COMMON_MOP_IMM_LOW_WIDTH }>,
@@ -410,6 +443,9 @@ macro_rules! mop_enum {
                 }
                 dest_reg
             }
+            fn src_reg_width(self) -> <Self::SrcRegWidth as Size>::SizeType {
+                self.$FirstVariant.src_reg_width()
+            }
             #[hdl]
             fn for_each_src_reg(
                 input: impl ToExpr<Type = Self>,
@@ -458,7 +494,7 @@ pub(crate) use mop_enum;
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> AluCommonMOp<NewDestReg, NewSrcRegWidth, SrcCount>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct AluCommonMOp<DestReg: Type, SrcRegWidth: Size, SrcCount: KnownSize> {
         #[common]
         pub common: CommonMOp<ConstUsize<0>, DestReg, SrcRegWidth, SrcCount>,
@@ -468,7 +504,7 @@ common_mop_struct! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> AddSubMOp<NewDestReg, NewSrcRegWidth, SrcCount>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct AddSubMOp<DestReg: Type, SrcRegWidth: Size, SrcCount: KnownSize> {
         #[common]
         pub alu_common: AluCommonMOp<DestReg, SrcRegWidth, SrcCount>,
@@ -481,7 +517,7 @@ common_mop_struct! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> LogicalMOp<NewDestReg, NewSrcRegWidth>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct LogicalMOp<DestReg: Type, SrcRegWidth: Size> {
         #[common]
         pub alu_common: AluCommonMOp<DestReg, SrcRegWidth, ConstUsize<2>>,
@@ -491,7 +527,7 @@ common_mop_struct! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> BranchMOp<NewDestReg, NewSrcRegWidth>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct BranchMOp<DestReg: Type, SrcRegWidth: Size> {
         #[common]
         pub alu_common: AluCommonMOp<DestReg, SrcRegWidth, ConstUsize<2>>,
@@ -510,7 +546,7 @@ mop_enum! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> ReadL2RegMOp<NewDestReg, NewSrcRegWidth>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct ReadL2RegMOp<DestReg: Type, SrcRegWidth: Size> {
         #[common]
         pub common: CommonMOp<ConstUsize<1>, DestReg, SrcRegWidth, ConstUsize<0>>,
@@ -519,7 +555,7 @@ common_mop_struct! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> WriteL2RegMOp<NewDestReg, NewSrcRegWidth>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct WriteL2RegMOp<DestReg: Type, SrcRegWidth: Size> {
         #[common]
         pub common: CommonMOp<ConstUsize<1>, DestReg, SrcRegWidth, ConstUsize<1>>,
@@ -536,7 +572,7 @@ mop_enum! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> LoadStoreCommonMOp<NewDestReg, NewSrcRegWidth, SrcCount>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct LoadStoreCommonMOp<DestReg: Type, SrcRegWidth: Size, SrcCount: KnownSize> {
         #[common]
         pub common: CommonMOp<ConstUsize<1>, DestReg, SrcRegWidth, SrcCount>,
@@ -545,7 +581,7 @@ common_mop_struct! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> LoadMOp<NewDestReg, NewSrcRegWidth>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct LoadMOp<DestReg: Type, SrcRegWidth: Size> {
         #[common]
         pub load_store_common: LoadStoreCommonMOp<DestReg, SrcRegWidth, ConstUsize<1>>,
@@ -554,7 +590,7 @@ common_mop_struct! {
 
 common_mop_struct! {
     #[mapped(<NewDestReg, NewSrcRegWidth> StoreMOp<NewDestReg, NewSrcRegWidth>)]
-    #[hdl]
+    #[hdl(cmp_eq)]
     pub struct StoreMOp<DestReg: Type, SrcRegWidth: Size> {
         #[common]
         pub load_store_common: LoadStoreCommonMOp<DestReg, SrcRegWidth, ConstUsize<2>>,
@@ -569,7 +605,7 @@ mop_enum! {
     }
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 /// there may be more than one unit of a given kind, so UnitNum is not the same as UnitKind.
 /// zero is used for built-in constants, such as the zero register
 pub struct UnitNum<Width: Size> {
@@ -617,12 +653,12 @@ impl<Width: Size> UnitNum<Width> {
 
 pub const CONST_ZERO_UNIT_NUM: usize = 0;
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct UnitOutRegNum<Width: Size> {
     pub value: UIntType<Width>,
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 /// Physical Register Number -- registers in the CPU's backend
 pub struct PRegNum<UnitNumWidth: Size, OutRegNumWidth: Size> {
     pub unit_num: UnitNum<UnitNumWidth>,
@@ -643,7 +679,7 @@ impl<UnitNumWidth: Size, OutRegNumWidth: Size> PRegNum<UnitNumWidth, OutRegNumWi
     }
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 /// µOp Register Number -- register in a micro-operation before register renaming
 #[doc(alias = "UOpRegNum")] // help you find it in the docs if you mis-spell it
 #[doc(alias = "\u{B5}OpRegNum")] // micro sign
@@ -681,7 +717,7 @@ impl MOpRegNum {
         Self::CONST_ZERO_REG_NUM + 1..Self::SPECIAL_REG_NUMS.start;
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 /// all the registers this instruction will write to, they are all renamed to the same physical register.
 pub struct MOpDestReg {
     /// some instructions have multiple destination registers, e.g. x86 div

crates/cpu/src/instruction/power_isa.rs

@@ -4,22 +4,22 @@ use crate::{instruction::MOpRegNum, util::range_u32_nth_or_panic};
 use fayalite::prelude::*;
 use std::ops::Range;
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct PowerIsaRegNum {
     pub value: UInt<5>,
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct PowerIsaFRegNum {
     pub value: UInt<5>,
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct PowerIsaCrFieldNum {
     pub value: UInt<3>,
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct PowerIsaCrBitNum {
     pub cr_field: PowerIsaCrFieldNum,
     pub bit_in_field: UInt<2>,

crates/cpu/src/reg_alloc.rs

@@ -311,7 +311,7 @@ pub fn reg_alloc(config: &CpuConfig) {
             SourceLocation::caller(),
         );
         connect(dyn_unit.cd(unit), cd);
-        let unit_input_insn = dyn_unit.input_insn(unit);
+        let unit_to_reg_alloc = dyn_unit.unit_to_reg_alloc(unit);
         // TODO: handle assigning multiple instructions to a unit at a time
         let assign_to_unit_at_once = NonZeroUsize::new(1).unwrap();
         // TODO: handle retiring multiple instructions from a unit at a time
@@ -333,8 +333,8 @@ pub fn reg_alloc(config: &CpuConfig) {
         );
         connect(unit_free_regs_tracker.alloc_out[0].ready, false);
         connect(
-            unit_input_insn.data,
-            Expr::ty(unit_input_insn).data.HdlNone(),
+            unit_to_reg_alloc.input_insn.data,
+            Expr::ty(unit_to_reg_alloc.input_insn).data.HdlNone(),
         );
         for fetch_index in 0..config.fetch_width.get() {
             #[hdl]
@@ -343,7 +343,7 @@ pub fn reg_alloc(config: &CpuConfig) {
                 connect(available_units[fetch_index][unit_index], false);
             }
             #[hdl]
-            if !unit_input_insn.ready {
+            if !unit_to_reg_alloc.input_insn.ready {
                 // must come after to override connects in loops above
                 connect(available_units[fetch_index][unit_index], false);
             }
@@ -356,11 +356,11 @@ pub fn reg_alloc(config: &CpuConfig) {
                     if let HdlSome(renamed_mop) =
                         HdlOption::and_then(renamed_mops[fetch_index], |v| dyn_unit.extract_mop(v))
                     {
-                        connect(unit_input_insn.data, HdlSome(renamed_mop));
+                        connect(unit_to_reg_alloc.input_insn.data, HdlSome(renamed_mop));
                     } else {
                         connect(
-                            unit_input_insn.data,
-                            HdlSome(Expr::ty(unit_input_insn).data.HdlSome.uninit()),
+                            unit_to_reg_alloc.input_insn.data,
+                            HdlSome(Expr::ty(unit_to_reg_alloc.input_insn).data.HdlSome.uninit()),
                         );
                         // FIXME: add hdl_assert(cd.clk, false.to_expr(), "");
                     }
@@ -387,7 +387,7 @@ pub fn reg_alloc(config: &CpuConfig) {
         }
         // TODO: connect outputs to other units
         connect(
-            dyn_unit.unit_forwarding_info(unit),
+            unit_to_reg_alloc.unit_forwarding_info,
             #[hdl]
             UnitForwardingInfo::<_, _, _> {
                 unit_output_writes: repeat(
@@ -397,10 +397,9 @@ pub fn reg_alloc(config: &CpuConfig) {
                 _phantom: PhantomData,
             },
         );
-        connect(dyn_unit.output(unit).ready, false);
         // TODO: handle cancellation
         connect(
-            dyn_unit.cancel_input(unit).data,
+            unit_to_reg_alloc.cancel_input,
             HdlOption[config.unit_cancel_input()].HdlNone(),
         );
     }

crates/cpu/src/register.rs

@@ -8,7 +8,7 @@ pub enum FlagsMode {
     X86(PRegFlagsX86),
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct PRegFlagsPowerISA {}
 
 impl PRegFlagsPowerISA {
@@ -56,7 +56,7 @@ impl PRegFlagsPowerISA {
     }
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct PRegFlagsX86 {}
 
 impl PRegFlagsX86 {
@@ -100,7 +100,7 @@ impl PRegFlagsX86 {
     }
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 /// this is *not* the same as any particular ISA's flags register,
 /// on PowerISA it is a combination of some bits from XER with a single 4-bit CR field.
 ///
@@ -138,7 +138,7 @@ impl PRegFlags {
     }
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 /// Unit output register's value -- a combination of an integer/fp register
 /// and flags register and CR field.
 ///

crates/cpu/src/unit.rs

@@ -7,13 +7,12 @@ use crate::{
         mop_enum, AluBranchMOp, L2RegisterFileMOp, LoadStoreMOp, MOpTrait, UnitOutRegNum,
     },
     register::PRegValue,
-    unit::unit_base::UnitForwardingInfo,
+    unit::unit_base::UnitToRegAlloc,
 };
 use fayalite::{
     bundle::{Bundle, BundleType},
     intern::{Intern, Interned},
     prelude::*,
-    util::ready_valid::ReadyValid,
 };
 
 pub mod alu_branch;
@@ -143,19 +142,19 @@ all_units! {
     }
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct UnitResultCompleted<ExtraOut> {
     pub value: PRegValue,
     pub extra_out: ExtraOut,
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct UnitOutputWrite<OutRegNumWidth: Size> {
     pub which: UnitOutRegNum<OutRegNumWidth>,
     pub value: PRegValue,
 }
 
-#[hdl]
+#[hdl(cmp_eq)]
 pub struct TrapData {
     // TODO
 }
@@ -166,13 +165,25 @@ pub enum UnitResult<ExtraOut> {
     Trap(TrapData),
 }
 
+impl<ExtraOut: Type> UnitResult<ExtraOut> {
+    pub fn extra_out_ty(self) -> ExtraOut {
+        self.Completed.extra_out
+    }
+}
+
 #[hdl]
 pub struct UnitOutput<OutRegNumWidth: Size, ExtraOut> {
     pub which: UnitOutRegNum<OutRegNumWidth>,
     pub result: UnitResult<ExtraOut>,
 }
 
-#[hdl]
+impl<OutRegNumWidth: Size, ExtraOut: Type> UnitOutput<OutRegNumWidth, ExtraOut> {
+    pub fn extra_out_ty(self) -> ExtraOut {
+        self.result.extra_out_ty()
+    }
+}
+
+#[hdl(cmp_eq)]
 pub struct UnitCancelInput<OutRegNumWidth: Size> {
     pub which: UnitOutRegNum<OutRegNumWidth>,
 }
@@ -197,19 +208,10 @@ pub trait UnitTrait:
 
     fn module(&self) -> Interned<Module<Self::Type>>;
 
-    fn input_insn(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<Self::MOp>>;
-
-    fn cancel_input(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<UnitCancelInput<DynSize>>>;
-
-    fn unit_forwarding_info(
+    fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitForwardingInfo<DynSize, DynSize, DynSize>>;
-
-    fn output(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ReadyValid<UnitOutput<DynSize, Self::ExtraOut>>>;
+    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>>;
 
     fn cd(&self, this: Expr<Self::Type>) -> Expr<ClockDomain>;
 
@@ -266,26 +268,11 @@ impl UnitTrait for DynUnit {
         self.unit.module()
     }
 
-    fn input_insn(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<Self::MOp>> {
-        self.unit.input_insn(this)
-    }
-
-    fn cancel_input(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<UnitCancelInput<DynSize>>> {
-        self.unit.cancel_input(this)
-    }
-
-    fn unit_forwarding_info(
+    fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitForwardingInfo<DynSize, DynSize, DynSize>> {
-        self.unit.unit_forwarding_info(this)
-    }
-
-    fn output(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ReadyValid<UnitOutput<DynSize, Self::ExtraOut>>> {
-        self.unit.output(this)
+    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>> {
+        self.unit.unit_to_reg_alloc(this)
     }
 
     fn cd(&self, this: Expr<Self::Type>) -> Expr<ClockDomain> {
@@ -332,26 +319,13 @@ impl<T: UnitTrait + Clone + std::hash::Hash + Eq> UnitTrait for DynUnitWrapper<T
         self.0.module().canonical().intern_sized()
     }
 
-    fn input_insn(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<Self::MOp>> {
-        Expr::from_bundle(Expr::as_bundle(self.0.input_insn(Expr::from_bundle(this))))
-    }
-
-    fn cancel_input(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<UnitCancelInput<DynSize>>> {
-        self.0.cancel_input(Expr::from_bundle(this))
-    }
-
-    fn unit_forwarding_info(
+    fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitForwardingInfo<DynSize, DynSize, DynSize>> {
-        self.0.unit_forwarding_info(Expr::from_bundle(this))
-    }
-
-    fn output(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ReadyValid<UnitOutput<DynSize, Self::ExtraOut>>> {
-        Expr::from_bundle(Expr::as_bundle(self.0.output(Expr::from_bundle(this))))
+    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>> {
+        Expr::from_bundle(Expr::as_bundle(
+            self.0.unit_to_reg_alloc(Expr::from_bundle(this)),
+        ))
     }
 
     fn cd(&self, this: Expr<Self::Type>) -> Expr<ClockDomain> {

crates/cpu/src/unit/alu_branch.rs

@@ -5,14 +5,13 @@ use crate::{
     config::CpuConfig,
     instruction::{AluBranchMOp, UnitOutRegNum},
     unit::{
-        unit_base::{unit_base, UnitForwardingInfo},
-        DynUnit, DynUnitWrapper, UnitCancelInput, UnitKind, UnitMOp, UnitOutput, UnitTrait,
+        unit_base::{unit_base, UnitToRegAlloc},
+        DynUnit, DynUnitWrapper, UnitKind, UnitMOp, UnitTrait,
     },
 };
 use fayalite::{
     intern::{Intern, Interned},
     prelude::*,
-    util::ready_valid::ReadyValid,
 };
 
 #[hdl_module]
@@ -20,30 +19,30 @@ pub fn alu_branch(config: &CpuConfig, unit_index: usize) {
     #[hdl]
     let cd: ClockDomain = m.input();
     #[hdl]
-    let input_insn: ReadyValid<AluBranchMOp<UnitOutRegNum<DynSize>, DynSize>> =
-        m.input(ReadyValid[AluBranchMOp[config.unit_out_reg_num()][config.p_reg_num_width()]]);
-    #[hdl]
-    let unit_forwarding_info: UnitForwardingInfo<DynSize, DynSize, DynSize> =
-        m.input(config.unit_forwarding_info());
-    #[hdl]
-    let cancel_input: ReadyValid<UnitCancelInput<DynSize>> =
-        m.input(ReadyValid[config.unit_cancel_input()]);
-    #[hdl]
-    let output: ReadyValid<UnitOutput<DynSize, ()>> =
-        m.output(ReadyValid[UnitOutput[config.out_reg_num_width][()]]);
+    let unit_to_reg_alloc: UnitToRegAlloc<
+        AluBranchMOp<UnitOutRegNum<DynSize>, DynSize>,
+        (),
+        DynSize,
+        DynSize,
+        DynSize,
+    > = m.output(config.unit_to_reg_alloc(
+        AluBranchMOp[config.unit_out_reg_num()][config.p_reg_num_width()],
+        (),
+    ));
     #[hdl]
     let unit_base = instance(unit_base(
         config,
         unit_index,
-        Expr::ty(input_insn).data.HdlSome,
+        Expr::ty(unit_to_reg_alloc).input_insn.data.HdlSome,
+        (),
     ));
-    connect(unit_base.input_insn, input_insn);
+    connect(unit_to_reg_alloc, unit_base.unit_to_reg_alloc);
     connect(unit_base.cd, cd);
-    connect(unit_base.unit_forwarding_info, unit_forwarding_info);
-    connect(unit_base.cancel_input, cancel_input);
-    // TODO: finish
-    connect(unit_base.ready_mop.ready, true);
-    connect(output.data, Expr::ty(output.data).HdlNone());
+    connect(unit_base.execute_start.ready, true); // TODO: finish
+    connect(
+        unit_base.execute_end,
+        Expr::ty(unit_base.execute_end).HdlNone(),
+    ); // TODO: finish
 }
 
 #[derive(Debug, Copy, Clone, PartialEq, Eq, Hash)]
@@ -75,7 +74,7 @@ impl UnitTrait for AluBranch {
     }
 
     fn mop_ty(&self) -> Self::MOp {
-        self.module.io_ty().input_insn.data.HdlSome
+        self.module.io_ty().unit_to_reg_alloc.mop_ty()
     }
 
     fn unit_kind(&self) -> UnitKind {
@@ -93,26 +92,11 @@ impl UnitTrait for AluBranch {
         self.module
     }
 
-    fn input_insn(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<Self::MOp>> {
-        this.input_insn
-    }
-
-    fn cancel_input(&self, this: Expr<Self::Type>) -> Expr<ReadyValid<UnitCancelInput<DynSize>>> {
-        this.cancel_input
-    }
-
-    fn unit_forwarding_info(
+    fn unit_to_reg_alloc(
         &self,
         this: Expr<Self::Type>,
-    ) -> Expr<UnitForwardingInfo<DynSize, DynSize, DynSize>> {
-        this.unit_forwarding_info
-    }
-
-    fn output(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ReadyValid<UnitOutput<DynSize, Self::ExtraOut>>> {
-        this.output
+    ) -> Expr<UnitToRegAlloc<Self::MOp, Self::ExtraOut, DynSize, DynSize, DynSize>> {
+        this.unit_to_reg_alloc
     }
 
     fn cd(&self, this: Expr<Self::Type>) -> Expr<ClockDomain> {

crates/cpu/src/unit/unit_base.rs

@@ -3,9 +3,9 @@
 
 use crate::{
     config::CpuConfig,
-    instruction::{MOpTrait, UnitOutRegNum, COMMON_MOP_SRC_LEN},
+    instruction::{MOpTrait, PRegNum, UnitOutRegNum, COMMON_MOP_SRC_LEN},
     register::PRegValue,
-    unit::{UnitCancelInput, UnitOutputWrite},
+    unit::{UnitCancelInput, UnitOutput, UnitOutputWrite},
     util::tree_reduce::tree_reduce,
 };
 use fayalite::{module::wire_with_loc, prelude::*, ty::StaticType, util::ready_valid::ReadyValid};
@@ -18,113 +18,439 @@ pub struct UnitForwardingInfo<UnitNumWidth: Size, OutRegNumWidth: Size, UnitCoun
 }
 
 #[hdl]
-pub struct ReadyMOp<MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>>> {
+pub struct UnitToRegAlloc<
+    MOp: Type,
+    ExtraOut: Type,
+    UnitNumWidth: Size,
+    OutRegNumWidth: Size,
+    UnitCount: Size,
+> {
+    #[hdl(flip)]
+    pub unit_forwarding_info: UnitForwardingInfo<UnitNumWidth, OutRegNumWidth, UnitCount>,
+    #[hdl(flip)]
+    pub input_insn: ReadyValid<MOp>,
+    #[hdl(flip)]
+    pub cancel_input: HdlOption<UnitCancelInput<OutRegNumWidth>>,
+    pub output: HdlOption<UnitOutput<OutRegNumWidth, ExtraOut>>,
+}
+
+impl<MOp: Type, ExtraOut: Type, UnitNumWidth: Size, OutRegNumWidth: Size, UnitCount: Size>
+    UnitToRegAlloc<MOp, ExtraOut, UnitNumWidth, OutRegNumWidth, UnitCount>
+{
+    pub fn mop_ty(self) -> MOp {
+        self.input_insn.data.HdlSome
+    }
+    pub fn extra_out_ty(self) -> ExtraOut {
+        self.output.HdlSome.extra_out_ty()
+    }
+}
+
+#[hdl]
+pub struct ExecuteStart<MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>>> {
     pub mop: MOp,
     pub src_values: Array<PRegValue, { COMMON_MOP_SRC_LEN }>,
 }
 
 #[hdl]
-struct InFlightOp<MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>>> {
-    pub mop: MOp,
-    pub src_values: Array<HdlOption<PRegValue>, { COMMON_MOP_SRC_LEN }>,
+pub struct ExecuteEnd<OutRegNumWidth: Size, ExtraOut> {
+    pub unit_output: UnitOutput<OutRegNumWidth, ExtraOut>,
+}
+
+#[hdl]
+enum InFlightOpState {
+    Ready,
+    Running,
+    CanceledAndRunning,
+}
+
+impl InFlightOpState {
+    fn ready_next_state(canceling: bool, starting: bool, ending: bool) -> Expr<HdlOption<Self>> {
+        match (canceling, starting, ending) {
+            (false, false, _) => HdlSome(InFlightOpState.Ready()),
+            (false, true, false) => HdlSome(InFlightOpState.Running()),
+            (false, true, true) => HdlNone(),
+            (true, false, _) => HdlNone(),
+            (true, true, false) => HdlSome(InFlightOpState.CanceledAndRunning()),
+            (true, true, true) => HdlNone(),
+        }
+    }
+    fn running_next_state(canceling: bool, _starting: bool, ending: bool) -> Expr<HdlOption<Self>> {
+        match (canceling, ending) {
+            (false, false) => HdlSome(InFlightOpState.Running()),
+            (false, true) => HdlNone(),
+            (true, false) => HdlSome(InFlightOpState.CanceledAndRunning()),
+            (true, true) => HdlNone(),
+        }
+    }
+    fn canceled_and_running_next_state(
+        _canceling: bool,
+        _starting: bool,
+        ending: bool,
+    ) -> Expr<HdlOption<Self>> {
+        if ending {
+            HdlNone()
+        } else {
+            HdlSome(InFlightOpState.CanceledAndRunning())
+        }
+    }
+    /// FIXME: this is working around #[hdl] match not supporting matching values inside structs yet
+    #[hdl]
+    fn connect_next_state(
+        canceling: Expr<Bool>,
+        starting: Expr<Bool>,
+        ending: Expr<Bool>,
+        next_state_fn: fn(canceling: bool, starting: bool, ending: bool) -> Expr<HdlOption<Self>>,
+        next_state: Expr<HdlOption<Self>>,
+    ) {
+        #[hdl]
+        fn recurse<const N: usize>(
+            exprs: &[Expr<Bool>; N],
+            bools: &mut [bool; N],
+            f: &mut impl FnMut(&[bool; N]),
+            arg_index: usize,
+        ) {
+            if arg_index < N {
+                #[hdl]
+                if exprs[arg_index] {
+                    bools[arg_index] = true;
+                    recurse(exprs, bools, f, arg_index + 1);
+                } else {
+                    bools[arg_index] = false;
+                    recurse(exprs, bools, f, arg_index + 1);
+                }
+            } else {
+                f(bools);
+            }
+        }
+        recurse(
+            &[canceling, starting, ending],
+            &mut [false; 3],
+            &mut |&[canceling, starting, ending]| {
+                connect(next_state, next_state_fn(canceling, starting, ending))
+            },
+            0,
+        );
+    }
+}
+
+#[hdl]
+struct InFlightOp<MOp: Type> {
+    state: InFlightOpState,
+    mop: MOp,
+    src_ready_flags: Array<Bool, { COMMON_MOP_SRC_LEN }>,
+}
+
+#[hdl]
+struct InFlightOpsSummary<OpIndexWidth: Size> {
+    empty_op_index: HdlOption<UIntType<OpIndexWidth>>,
+    ready_op_index: HdlOption<UIntType<OpIndexWidth>>,
+}
+
+impl<OpIndexWidth: Size> InFlightOpsSummary<OpIndexWidth> {
+    #[hdl]
+    fn new<MOp: Type>(
+        op_index: usize,
+        op_index_ty: UIntType<OpIndexWidth>,
+        in_flight_op: impl ToExpr<Type = HdlOption<InFlightOp<MOp>>>,
+    ) -> Expr<Self> {
+        let empty_op_index = wire_with_loc(
+            &format!("empty_op_index_{op_index}"),
+            SourceLocation::caller(),
+            HdlOption[op_index_ty],
+        );
+        connect(empty_op_index, HdlOption[op_index_ty].HdlNone());
+        let ready_op_index = wire_with_loc(
+            &format!("ready_op_index_{op_index}"),
+            SourceLocation::caller(),
+            HdlOption[op_index_ty],
+        );
+        connect(ready_op_index, HdlOption[op_index_ty].HdlNone());
+        #[hdl]
+        if let HdlSome(in_flight_op) = in_flight_op {
+            #[hdl]
+            let InFlightOp::<_> {
+                state,
+                mop: _,
+                src_ready_flags,
+            } = in_flight_op;
+            connect(ready_op_index, HdlOption[op_index_ty].HdlNone());
+            #[hdl]
+            match state {
+                InFlightOpState::Ready =>
+                {
+                    #[hdl]
+                    if src_ready_flags.cmp_eq([true; COMMON_MOP_SRC_LEN]) {
+                        connect(ready_op_index, HdlSome(op_index.cast_to(op_index_ty)));
+                    }
+                }
+                InFlightOpState::CanceledAndRunning | InFlightOpState::Running => {}
+            }
+        } else {
+            connect(empty_op_index, HdlSome(op_index.cast_to(op_index_ty)));
+        }
+        #[hdl]
+        InFlightOpsSummary::<_> {
+            empty_op_index,
+            ready_op_index,
+        }
+    }
+    #[hdl]
+    fn combine(l: impl ToExpr<Type = Self>, r: impl ToExpr<Type = Self>) -> Expr<Self> {
+        let l = l.to_expr();
+        let r = r.to_expr();
+        #[hdl]
+        InFlightOpsSummary::<_> {
+            empty_op_index: HdlOption::or(l.empty_op_index, r.empty_op_index),
+            ready_op_index: HdlOption::or(l.ready_op_index, r.ready_op_index),
+        }
+    }
+}
+
+impl InFlightOpsSummary<DynSize> {
+    fn summarize<MOp: Type, MaxInFlight: Size>(
+        in_flight_ops: impl ToExpr<Type = ArrayType<HdlOption<InFlightOp<MOp>>, MaxInFlight>>,
+    ) -> Expr<Self> {
+        let in_flight_ops = in_flight_ops.to_expr();
+        let max_in_flight = Expr::ty(in_flight_ops).len();
+        let index_range = 0..max_in_flight;
+        let index_ty = UInt::range(index_range.clone());
+        tree_reduce(
+            index_range.map(|i| Self::new(i, index_ty, in_flight_ops[i])),
+            Self::combine,
+        )
+        .expect("in_flight_ops is known to have len > 0")
+    }
 }
 
 #[hdl_module]
-pub fn unit_base<MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>>>(
+pub fn unit_base<
+    MOp: Type + MOpTrait<DestReg = UnitOutRegNum<DynSize>, SrcRegWidth = DynSize>,
+    ExtraOut: Type,
+>(
     config: &CpuConfig,
     unit_index: usize,
     mop_ty: MOp,
+    extra_out_ty: ExtraOut,
 ) {
     #[hdl]
     let cd: ClockDomain = m.input();
     #[hdl]
-    let unit_forwarding_info: UnitForwardingInfo<DynSize, DynSize, DynSize> =
-        m.input(config.unit_forwarding_info());
+    let unit_to_reg_alloc: UnitToRegAlloc<MOp, ExtraOut, DynSize, DynSize, DynSize> =
+        m.output(config.unit_to_reg_alloc(mop_ty, extra_out_ty));
     #[hdl]
-    let input_insn: ReadyValid<MOp> = m.input(ReadyValid[mop_ty]);
-    connect(input_insn.ready, false);
+    let execute_start: ReadyValid<ExecuteStart<MOp>> = m.output(ReadyValid[ExecuteStart[mop_ty]]);
     #[hdl]
-    let cancel_input: ReadyValid<UnitCancelInput<DynSize>> =
-        m.input(ReadyValid[config.unit_cancel_input()]);
-    connect(cancel_input.ready, true);
-    #[hdl]
-    let ready_mop: ReadyValid<ReadyMOp<MOp>> = m.output(ReadyValid[ReadyMOp[mop_ty]]);
-    connect(ready_mop.data, Expr::ty(ready_mop.data).HdlNone());
+    let execute_end: HdlOption<ExecuteEnd<DynSize, ExtraOut>> =
+        m.input(HdlOption[ExecuteEnd[config.out_reg_num_width][extra_out_ty]]);
+
+    connect(execute_start.data, Expr::ty(execute_start).data.HdlNone());
+
     let max_in_flight = config.unit_max_in_flight(unit_index).get();
+    let in_flight_op_ty = InFlightOp[mop_ty];
     #[hdl]
-    let in_flight_ops = reg_builder().clock_domain(cd).reset(repeat(
-        HdlOption[InFlightOp[mop_ty]].HdlNone(),
-        max_in_flight,
-    ));
-    let in_flight_op_index_ty = UInt::range(0..max_in_flight);
+    let in_flight_ops = reg_builder()
+        .clock_domain(cd)
+        .reset(repeat(HdlOption[in_flight_op_ty].HdlNone(), max_in_flight));
+
+    let in_flight_ops_summary_value = InFlightOpsSummary::summarize(in_flight_ops);
     #[hdl]
-    let input_index = wire(HdlOption[in_flight_op_index_ty]);
+    let in_flight_ops_summary = wire(Expr::ty(in_flight_ops_summary_value));
+    connect(in_flight_ops_summary, in_flight_ops_summary_value);
+
     connect(
-        input_index,
-        tree_reduce(
-            (0..max_in_flight).map(|i| -> Expr<HdlOption<UInt>> {
-                HdlOption::map(in_flight_ops[i], |_| i.cast_to(in_flight_op_index_ty))
-            }),
-            HdlOption::or,
-        )
-        .expect("max_in_flight is known to be non-zero"),
+        unit_to_reg_alloc.input_insn.ready,
+        HdlOption::is_some(in_flight_ops_summary.empty_op_index),
     );
+    // TODO: connect(execute_start.data, <read_regs>(in_flight_ops_summary.ready_op_index));
+
+    connect(
+        unit_to_reg_alloc.output,
+        Expr::ty(unit_to_reg_alloc.output).HdlNone(),
+    ); // TODO: finish
+
     #[hdl]
-    let input_in_flight_op = wire(HdlOption[InFlightOp[mop_ty]]);
-    connect(input_in_flight_op, Expr::ty(input_in_flight_op).HdlNone());
+    let input_in_flight_op = wire(HdlOption[in_flight_op_ty]);
+    connect(input_in_flight_op, HdlOption[in_flight_op_ty].HdlNone());
     #[hdl]
-    if let HdlSome(mop) = ReadyValid::firing_data(input_insn) {
-        let src_values = wire_with_loc(
-            "input_in_flight_op_src_values",
+    if let HdlSome(mop) = ReadyValid::firing_data(unit_to_reg_alloc.input_insn) {
+        #[hdl]
+        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),
+        );
+        MOp::connect_src_regs(mop, input_mop_src_regs);
+        let src_ready_flags = wire_with_loc(
+            "input_in_flight_op_src_ready_flags",
             SourceLocation::caller(),
             StaticType::TYPE,
         );
-        connect(
-            src_values,
-            [HdlSome(PRegValue::zeroed()); COMMON_MOP_SRC_LEN],
-        );
-        MOp::for_each_src_reg(mop, &mut |src_reg, src_index| {
+        for src_index in 0..COMMON_MOP_SRC_LEN {
+            connect(
+                src_ready_flags[src_index],
+                config
+                    .p_reg_num()
+                    .const_zero()
+                    .cast_to_bits()
+                    .cmp_eq(input_mop_src_regs[src_index]),
+            );
+        }
+        #[hdl]
+        if unit_to_reg_alloc.cancel_input.cmp_ne(HdlSome(
             #[hdl]
-            if config
-                .p_reg_num()
-                .const_zero()
-                .cast_to_bits()
-                .cmp_ne(src_reg)
-            {
-                connect(src_values[src_index], HdlNone());
-            }
-        });
-        connect(
-            input_in_flight_op,
-            HdlSome(
-                #[hdl]
-                InFlightOp::<_> { mop, src_values },
-            ),
-        );
+            UnitCancelInput::<_> {
+                which: MOp::dest_reg(mop),
+            },
+        )) {
+            connect(
+                input_in_flight_op,
+                HdlSome(
+                    #[hdl]
+                    InFlightOp::<_> {
+                        state: InFlightOpState.Ready(),
+                        mop,
+                        src_ready_flags,
+                    },
+                ),
+            );
+        }
+        #[hdl]
+        if let HdlSome(empty_op_index) = in_flight_ops_summary.empty_op_index {
+            connect(in_flight_ops[empty_op_index], input_in_flight_op);
+        }
     }
+
+    #[hdl]
+    let in_flight_op_next_state = wire(Array[HdlOption[InFlightOpState]][max_in_flight]);
+    #[hdl]
+    let in_flight_op_next_src_ready_flags =
+        wire(Array[in_flight_op_ty.src_ready_flags][max_in_flight]);
+    #[hdl]
+    let in_flight_op_canceling = wire(Array[Bool][max_in_flight]);
+    #[hdl]
+    let in_flight_op_execute_starting = wire(Array[Bool][max_in_flight]);
+    #[hdl]
+    let in_flight_op_execute_ending = wire(Array[Bool][max_in_flight]);
     for in_flight_op_index in 0..max_in_flight {
+        connect(
+            in_flight_op_next_src_ready_flags[in_flight_op_index],
+            [false; COMMON_MOP_SRC_LEN],
+        );
+        connect(in_flight_op_canceling[in_flight_op_index], false);
+        connect(in_flight_op_execute_starting[in_flight_op_index], false);
+        connect(in_flight_op_execute_ending[in_flight_op_index], false);
         #[hdl]
         if let HdlSome(in_flight_op) = in_flight_ops[in_flight_op_index] {
             #[hdl]
-            if let HdlSome(cancel_input) = ReadyValid::firing_data(cancel_input) {
+            let InFlightOp::<_> {
+                state,
+                mop,
+                src_ready_flags,
+            } = in_flight_op;
+            let which = MOp::dest_reg(mop);
+            let src_regs = wire_with_loc(
+                &format!("in_flight_op_src_regs_{in_flight_op_index}"),
+                SourceLocation::caller(),
+                mop_ty.src_regs_ty(),
+            );
+            connect(
+                src_regs,
+                repeat(config.p_reg_num().const_zero().cast_to_bits(), ConstUsize),
+            );
+            MOp::connect_src_regs(mop, src_regs);
+
+            connect(
+                in_flight_op_next_src_ready_flags[in_flight_op_index],
+                src_ready_flags,
+            );
+            let unit_output_writes = unit_to_reg_alloc.unit_forwarding_info.unit_output_writes;
+            for unit_index in 0..Expr::ty(unit_output_writes).len() {
                 #[hdl]
-                let UnitCancelInput::<_> { which } = cancel_input;
-                #[hdl]
-                if which.value.cmp_eq(MOp::dest_reg(in_flight_op.mop).value) {
-                    // TODO: if it needs extra time to cancel (e.g. still in pipeline), handle that here
-                    connect(
-                        in_flight_ops[in_flight_op_index],
-                        HdlOption[InFlightOp[mop_ty]].HdlNone(),
-                    );
+                if let HdlSome(unit_output_write) = unit_output_writes[unit_index] {
+                    #[hdl]
+                    let UnitOutputWrite::<_> {
+                        which: unit_out_reg,
+                        value: _,
+                    } = unit_output_write;
+                    let p_reg_num = #[hdl]
+                    PRegNum::<_, _> {
+                        unit_num: config.unit_num().from_index(unit_index),
+                        unit_out_reg,
+                    };
+                    for src_index in 0..COMMON_MOP_SRC_LEN {
+                        #[hdl]
+                        if p_reg_num.cast_to_bits().cmp_eq(src_regs[src_index]) {
+                            connect(
+                                in_flight_op_next_src_ready_flags[in_flight_op_index][src_index],
+                                true,
+                            );
+                        }
+                    }
                 }
             }
-            // TODO: finish
-        } else if let HdlSome(input_index) = input_index {
-            connect(input_insn.ready, true);
+
+            connect(
+                in_flight_op_canceling[in_flight_op_index],
+                unit_to_reg_alloc.cancel_input.cmp_eq(HdlSome(
+                    #[hdl]
+                    UnitCancelInput::<_> { which },
+                )),
+            );
+
             #[hdl]
-            if input_index.cmp_eq(in_flight_op_index) {
-                connect(in_flight_ops[in_flight_op_index], input_in_flight_op);
+            if let HdlSome(execute_end) = execute_end {
+                #[hdl]
+                let ExecuteEnd::<_, _> { unit_output } = execute_end;
+                #[hdl]
+                if which.cmp_eq(unit_output.which) {
+                    connect(in_flight_op_execute_ending[in_flight_op_index], true);
+                }
             }
+            #[hdl]
+            if let HdlSome(execute_start) = ReadyValid::firing_data(execute_start) {
+                #[hdl]
+                if which.cmp_eq(MOp::dest_reg(execute_start.mop)) {
+                    connect(in_flight_op_execute_starting[in_flight_op_index], true);
+                }
+            }
+            let connect_next_state = |f| {
+                InFlightOpState::connect_next_state(
+                    in_flight_op_canceling[in_flight_op_index],
+                    in_flight_op_execute_starting[in_flight_op_index],
+                    in_flight_op_execute_ending[in_flight_op_index],
+                    f,
+                    in_flight_op_next_state[in_flight_op_index],
+                );
+            };
+            #[hdl]
+            match state {
+                InFlightOpState::Ready => connect_next_state(InFlightOpState::ready_next_state),
+                InFlightOpState::Running => connect_next_state(InFlightOpState::running_next_state),
+                InFlightOpState::CanceledAndRunning => {
+                    connect_next_state(InFlightOpState::canceled_and_running_next_state);
+                }
+            }
+            #[hdl]
+            if let HdlSome(state) = in_flight_op_next_state[in_flight_op_index] {
+                connect(
+                    in_flight_ops[in_flight_op_index],
+                    HdlSome(
+                        #[hdl]
+                        InFlightOp::<_> {
+                            state,
+                            mop,
+                            src_ready_flags,
+                        },
+                    ),
+                );
+            } else {
+                connect(
+                    in_flight_ops[in_flight_op_index],
+                    HdlOption[in_flight_op_ty].HdlNone(),
+                );
+            }
+        } else {
+            connect(in_flight_op_next_state[in_flight_op_index], HdlNone());
         }
     }
 }