WIP: adding load/store unit

crates/cpu/src/instruction.rs

@@ -19,6 +19,7 @@ use std::{
     borrow::Cow,
     fmt,
     marker::PhantomData,
+    num::NonZeroU64,
     ops::{ControlFlow, Range},
 };
 
@@ -2983,6 +2984,19 @@ pub enum LoadStoreWidth {
     Width64Bit,
 }
 
+impl LoadStoreWidth {
+    #[hdl]
+    pub fn access_byte_size_sim(this: &<Self as Type>::SimValue) -> NonZeroU64 {
+        #[hdl(sim)]
+        match this {
+            Self::Width8Bit => const { NonZeroU64::new(1).unwrap() },
+            Self::Width16Bit => const { NonZeroU64::new(2).unwrap() },
+            Self::Width32Bit => const { NonZeroU64::new(4).unwrap() },
+            Self::Width64Bit => const { NonZeroU64::new(8).unwrap() },
+        }
+    }
+}
+
 #[hdl(cmp_eq)]
 pub enum LoadStoreConversion {
     ZeroExt,

crates/cpu/src/main_memory_and_io.rs

@@ -127,8 +127,47 @@ impl AddressRange {
             },
         }
     }
+    pub const fn is_overlapping(self, other: Self) -> bool {
+        self.contains(other.start().0) || other.contains(self.start().0)
+    }
 }
 
+const _: () = {
+    const N: u64 = 4;
+    const fn is_overlapping_reference(range1: AddressRange, range2: AddressRange) -> bool {
+        // we just need to check values < N since all starts/ends are < N
+        let mut v = 0;
+        while v < N {
+            if range1.contains(v) && range2.contains(v) {
+                return true;
+            }
+            v += 1;
+        }
+        false
+    }
+    let mut start1 = 0;
+    while start1 < N {
+        let mut end1 = 0;
+        while end1 < N {
+            let mut start2 = 0;
+            while start2 < N {
+                let mut end2 = 0;
+                while end2 < N {
+                    let range1 = AddressRange::from_wrapping_range_inclusive(start1..=end1);
+                    let range2 = AddressRange::from_wrapping_range_inclusive(start2..=end2);
+                    assert!(
+                        range1.is_overlapping(range2) == is_overlapping_reference(range1, range2)
+                    );
+                    end2 += 1;
+                }
+                start2 += 1;
+            }
+            end1 += 1;
+        }
+        start1 += 1;
+    }
+};
+
 #[derive(Copy, Clone, Eq, PartialEq, Hash, Debug, serde::Serialize, serde::Deserialize)]
 #[non_exhaustive]
 pub struct MemoryInterfaceConfig {

crates/cpu/src/unit.rs

@@ -8,6 +8,7 @@ use crate::{
         MOpVariantVisitOps, MOpVariantVisitor, MOpVisitVariants, RenamedMOp, mop_enum,
     },
     rename_execute_retire::ExecuteToUnitInterface,
+    unit::load_store::LoadStoreToDCacheInterface,
 };
 use fayalite::{
     bundle::{Bundle, BundleType},
@@ -18,6 +19,7 @@ use serde::{Deserialize, Serialize};
 use std::ops::ControlFlow;
 
 pub mod alu_branch;
+pub mod load_store;
 
 macro_rules! all_units {
     (
@@ -333,7 +335,7 @@ all_units! {
         #[create_dyn_unit_fn = |config, unit_index, filter| todo!()]
         #[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, filter| todo!()]
+        #[create_dyn_unit_fn = |config, unit_index, filter| load_store::LoadStore::new(config, unit_index, filter).to_dyn()]
         #[extract(load_store_mop, load_store_mop_sim, load_store_mop_sim_ref, load_store_mop_sim_mut)]
         LoadStore(LoadStoreMOp<DestReg, SrcReg>),
     }
@@ -387,6 +389,13 @@ impl RenamedMOpFilter for () {
     }
 }
 
+#[derive(Copy, Clone, PartialEq, Eq, Hash, Debug)]
+pub struct UnitIO {
+    pub cd: Option<Expr<ClockDomain>>,
+    pub from_execute: Expr<ExecuteToUnitInterface<PhantomConst<CpuConfig>>>,
+    pub to_d_cache: Option<Expr<LoadStoreToDCacheInterface<PhantomConst<CpuConfig>>>>,
+}
+
 pub trait UnitTrait:
     'static + Send + Sync + std::fmt::Debug + fayalite::intern::SupportsPtrEqWithTypeId
 {
@@ -395,11 +404,7 @@ pub trait UnitTrait:
     fn ty(&self) -> Self::Type;
     fn unit_kind(&self) -> UnitKind;
     fn module(&self) -> Interned<Module<Self::Type>>;
-    fn cd(&self, this: Expr<Self::Type>) -> Option<Expr<ClockDomain>>;
-    fn from_execute(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ExecuteToUnitInterface<PhantomConst<CpuConfig>>>;
+    fn io(&self, this: Expr<Self::Type>) -> UnitIO;
     fn to_dyn(&self) -> DynUnit;
 }
 
@@ -434,15 +439,8 @@ impl UnitTrait for DynUnit {
         self.unit.module()
     }
 
-    fn cd(&self, this: Expr<Self::Type>) -> Option<Expr<ClockDomain>> {
-        self.unit.cd(this)
-    }
-
-    fn from_execute(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ExecuteToUnitInterface<PhantomConst<CpuConfig>>> {
-        self.unit.from_execute(this)
+    fn io(&self, this: Expr<Self::Type>) -> UnitIO {
+        self.unit.io(this)
     }
 
     fn to_dyn(&self) -> DynUnit {
@@ -468,15 +466,8 @@ impl<T: UnitTrait + Clone + std::hash::Hash + Eq> UnitTrait for DynUnitWrapper<T
         self.0.module().canonical().intern_sized()
     }
 
-    fn cd(&self, this: Expr<Self::Type>) -> Option<Expr<ClockDomain>> {
-        self.0.cd(Expr::from_bundle(this))
-    }
-
-    fn from_execute(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ExecuteToUnitInterface<PhantomConst<CpuConfig>>> {
-        self.0.from_execute(Expr::from_bundle(this))
+    fn io(&self, this: Expr<Self::Type>) -> UnitIO {
+        self.0.io(Expr::from_bundle(this))
     }
 
     fn to_dyn(&self) -> DynUnit {

crates/cpu/src/unit/alu_branch.rs

@@ -18,7 +18,7 @@ use crate::{
         ExecuteToUnitInterface, GlobalState, NextPcPredictorOp, RenamedMOp, UnitCausedCancel,
         UnitFinishCauseCancel, UnitInputsReady, UnitOutputReady,
     },
-    unit::{DynUnit, DynUnitWrapper, RenamedMOpFilter, UnitKind, UnitTrait},
+    unit::{DynUnit, DynUnitWrapper, RenamedMOpFilter, UnitIO, UnitKind, UnitTrait},
 };
 use fayalite::{
     expr::CastToImpl,
@@ -1613,15 +1613,12 @@ impl UnitTrait for AluBranch {
         self.module
     }
 
-    fn cd(&self, _this: Expr<Self::Type>) -> Option<Expr<ClockDomain>> {
-        None
-    }
-
-    fn from_execute(
-        &self,
-        this: Expr<Self::Type>,
-    ) -> Expr<ExecuteToUnitInterface<PhantomConst<CpuConfig>>> {
-        this.from_execute
+    fn io(&self, this: Expr<Self::Type>) -> UnitIO {
+        UnitIO {
+            cd: None,
+            from_execute: this.from_execute,
+            to_d_cache: None,
+        }
     }
 
     fn to_dyn(&self) -> DynUnit {

crates/cpu/tests/units_formal.rs

@@ -14,7 +14,7 @@ use cpu::{
         DecodeAndRunSingleInsnInput, DecodeAndRunSingleInsnOutput, DecodeOneInsnInput,
         DecodeOneInsnMaxMOpCount, decode_and_run_single_insn,
     },
-    unit::{RenamedMOpFilter, UnitKind, UnitMOp, UnitTrait},
+    unit::{RenamedMOpFilter, UnitIO, UnitKind, UnitMOp, UnitTrait},
     util::array_vec::ArrayVec,
 };
 use fayalite::{
@@ -91,12 +91,20 @@ fn formal_harness(
             dyn_unit.module(),
             SourceLocation::caller(),
         );
+        let UnitIO {
+            cd: unit_cd,
+            from_execute: unit_from_execute,
+            to_d_cache,
+        } = dyn_unit.io(unit);
+        if let Some(unit_cd) = unit_cd {
+            connect(unit_cd, cd);
+        }
         connect(
-            dyn_unit.from_execute(unit),
+            unit_from_execute,
             ExecuteToUnitInterfaces::unit_fields(decode_and_run.to_units)[unit_index],
         );
-        if let Some(unit_cd) = dyn_unit.cd(unit) {
-            connect(unit_cd, cd);
+        if let Some(to_d_cache) = to_d_cache {
+            unimplemented!("to_d_cache for {unit:?}: {to_d_cache:?}");
         }
     }
     hdl_assert(cd.clk, !decode_and_run.error, "");