WIP adding fetch::l1_i_cache

crates/cpu/src/config.rs

@@ -37,6 +37,8 @@ pub struct CpuConfig {
     pub max_branches_per_fetch: NonZeroUsize,
     pub max_fetches_in_flight: NonZeroUsize,
     pub log2_fetch_width_in_bytes: u8,
+    pub log2_cache_line_size_in_bytes: u8,
+    pub log2_l1_i_cache_line_count: u8,
     /// default value for [`UnitConfig::max_in_flight`]
     pub default_unit_max_in_flight: NonZeroUsize,
     pub rob_size: NonZeroUsize,
@@ -63,6 +65,8 @@ impl CpuConfig {
         v
     };
     pub const DEFAULT_LOG2_FETCH_WIDTH_IN_BYTES: u8 = 3;
+    pub const DEFAULT_LOG2_CACHE_LINE_SIZE_IN_BYTES: u8 = 6;
+    pub const DEFAULT_LOG2_L1_I_CACHE_LINE_COUNT: u8 = 8;
     pub const DEFAULT_UNIT_MAX_IN_FLIGHT: NonZeroUsize = {
         let Some(v) = NonZeroUsize::new(8) else {
             unreachable!();
@@ -77,6 +81,8 @@ impl CpuConfig {
             max_branches_per_fetch: Self::DEFAULT_MAX_BRANCHES_PER_FETCH,
             max_fetches_in_flight: Self::DEFAULT_MAX_FETCHES_IN_FLIGHT,
             log2_fetch_width_in_bytes: Self::DEFAULT_LOG2_FETCH_WIDTH_IN_BYTES,
+            log2_cache_line_size_in_bytes: Self::DEFAULT_LOG2_CACHE_LINE_SIZE_IN_BYTES,
+            log2_l1_i_cache_line_count: Self::DEFAULT_LOG2_L1_I_CACHE_LINE_COUNT,
             default_unit_max_in_flight: Self::DEFAULT_UNIT_MAX_IN_FLIGHT,
             rob_size,
         }
@@ -141,6 +147,37 @@ impl CpuConfig {
             .checked_shl(self.log2_fetch_width_in_bytes.into())
             .expect("log2_fetch_width_in_bytes is too big")
     }
+    pub fn cache_line_size_in_bytes(&self) -> usize {
+        1usize
+            .checked_shl(self.log2_cache_line_size_in_bytes.into())
+            .expect("log2_cache_line_size_in_bytes is too big")
+    }
+    pub fn log2_fetches_per_cache_line(&self) -> usize {
+        self.log2_cache_line_size_in_bytes
+            .checked_sub(self.log2_fetch_width_in_bytes)
+            .expect("cache line size in bytes must not be smaller than fetch width in bytes")
+            .into()
+    }
+    pub fn fetches_per_cache_line(&self) -> usize {
+        self.log2_fetches_per_cache_line()
+            .try_into()
+            .ok()
+            .and_then(|v| 1usize.checked_shl(v))
+            .expect("log2_fetches_per_cache_line is too big")
+    }
+    pub fn l1_i_cache_line_count(&self) -> usize {
+        1usize
+            .checked_shl(self.log2_l1_i_cache_line_count.into())
+            .expect("log2_l1_i_cache_line_count is too big")
+    }
+    pub fn log2_l1_i_cache_size_in_bytes(&self) -> usize {
+        self.log2_l1_i_cache_line_count as usize + self.log2_cache_line_size_in_bytes as usize
+    }
+    pub fn l1_i_cache_size_in_bytes(&self) -> usize {
+        1usize
+            .checked_shl(self.log2_l1_i_cache_size_in_bytes() as _)
+            .expect("L1 I-Cache is too big")
+    }
 }
 
 #[hdl(get(|c| c.fetch_width.get()))]
@@ -161,6 +198,30 @@ pub type CpuConfigLog2FetchWidthInBytes<C: PhantomConstGet<CpuConfig>> = DynSize
 #[hdl(get(|c| c.fetch_width_in_bytes()))]
 pub type CpuConfigFetchWidthInBytes<C: PhantomConstGet<CpuConfig>> = DynSize;
 
+#[hdl(get(|c| c.log2_fetches_per_cache_line()))]
+pub type CpuConfigLog2FetchesPerCacheLine<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.fetches_per_cache_line()))]
+pub type CpuConfigFetchesPerCacheLine<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.log2_cache_line_size_in_bytes.into()))]
+pub type CpuConfigLog2CacheLineSizeInBytes<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.cache_line_size_in_bytes()))]
+pub type CpuConfigCacheLineSizeInBytes<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.log2_l1_i_cache_line_count.into()))]
+pub type CpuConfigLog2L1ICacheLineCount<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.l1_i_cache_line_count()))]
+pub type CpuConfigL1ICacheLineCount<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.log2_l1_i_cache_size_in_bytes()))]
+pub type CpuConfigLog2L1ICacheSizeInBytes<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl(get(|c| c.l1_i_cache_size_in_bytes()))]
+pub type CpuConfigL1ICacheSizeInBytes<C: PhantomConstGet<CpuConfig>> = DynSize;
+
 #[hdl(get(|c| c.rob_size.get()))]
 pub type CpuConfigRobSize<C: PhantomConstGet<CpuConfig>> = DynSize;
 

crates/cpu/src/fetch.rs

@@ -0,0 +1,670 @@
+// SPDX-License-Identifier: LGPL-3.0-or-later
+// See Notices.txt for copyright information
+
+use crate::{
+    config::{
+        CpuConfig, CpuConfigCacheLineSizeInBytes, CpuConfigFetchWidthInBytes,
+        CpuConfigFetchesPerCacheLine, CpuConfigL1ICacheLineCount, CpuConfigLog2FetchWidthInBytes,
+        CpuConfigLog2FetchesPerCacheLine, CpuConfigLog2L1ICacheLineCount,
+        CpuConfigMaxFetchesInFlight, PhantomConstCpuConfig,
+    },
+    next_pc::{
+        FETCH_BLOCK_ID_WIDTH, NextPcToFetchInterface, NextPcToFetchInterfaceInner, ResetStatus,
+        ResetSteps, SimValueDefault,
+    },
+    util::array_vec::ArrayVec,
+};
+use fayalite::{
+    int::UIntInRangeType,
+    memory::{ReadWriteStruct, memory_addr_width, splat_mask},
+    prelude::*,
+    util::ready_valid::ReadyValid,
+};
+use std::{collections::VecDeque, ops::Range};
+
+#[hdl]
+pub enum MemoryOperationKind {
+    Read,
+    Write,
+}
+
+#[hdl(no_static)]
+pub struct MemoryOperationStart<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    pub kind: MemoryOperationKind,
+    pub addr: UInt<64>,
+    pub write_data: ArrayType<UInt<8>, CpuConfigFetchWidthInBytes<C>>,
+    pub config: C,
+}
+
+#[hdl]
+pub enum MemoryOperationErrorKind {
+    Generic,
+}
+
+#[hdl]
+pub enum MemoryOperationFinishKind {
+    Success(MemoryOperationKind),
+    Error(MemoryOperationErrorKind),
+}
+
+#[hdl(no_static)]
+pub struct MemoryOperationFinish<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    pub kind: MemoryOperationFinishKind,
+    pub read_data: ArrayType<UInt<8>, CpuConfigFetchWidthInBytes<C>>,
+    pub config: C,
+}
+
+#[hdl(no_static)]
+pub struct MemoryInterface<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    pub start: ReadyValid<MemoryOperationStart<C>>,
+    #[hdl(flip)]
+    pub finish: ReadyValid<MemoryOperationFinish<C>>,
+    pub config: C,
+}
+
+#[hdl(no_static)]
+struct CacheLine<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    data: ArrayType<UInt<8>, CpuConfigCacheLineSizeInBytes<C>>,
+    addr: HdlOption<UIntType<CacheLineTagAddrWidth<C>>>,
+    config: C,
+}
+
+impl<C: PhantomConstCpuConfig> SimValueDefault for CacheLine<C> {
+    #[hdl]
+    fn sim_value_default(self) -> SimValue<Self> {
+        let Self { data, addr, config } = self;
+        #[hdl(sim)]
+        Self {
+            data: data.sim_value_default(),
+            addr: addr.sim_value_default(),
+            config,
+        }
+    }
+}
+
+#[hdl(get(|c| 64usize.saturating_sub(c.log2_l1_i_cache_size_in_bytes())))]
+type CacheLineTagAddrWidth<C: PhantomConstGet<CpuConfig>> = DynSize;
+
+#[hdl]
+enum CacheLookupState {
+    Empty,
+    NotStarted,
+    CanceledWhileReadingCache,
+    ReadingCache,
+    CacheMissCanceled,
+    CacheMiss,
+    CleaningUpErrorCanceled,
+    CleaningUpError,
+    ErrorCleanUpDone,
+}
+
+impl CacheLookupState {
+    #[hdl]
+    fn is_empty(this: &SimValue<Self>) -> bool {
+        #[hdl(sim)]
+        match this {
+            Self::Empty => true,
+            _ => false,
+        }
+    }
+    fn is_canceled(this: &SimValue<Self>) -> bool {
+        Self::try_to_canceled(this).is_none()
+    }
+    /// if `this` is canceled return `None`,
+    /// otherwise return `Some(canceled)` where `canceled` is the canceled equivalent
+    #[hdl]
+    fn try_to_canceled(this: &SimValue<Self>) -> Option<SimValue<Self>> {
+        #[hdl(sim)]
+        match this {
+            Self::Empty
+            | Self::CanceledWhileReadingCache
+            | Self::CacheMissCanceled
+            | Self::CleaningUpErrorCanceled => None,
+            Self::NotStarted => Some(
+                #[hdl(sim)]
+                CacheLookupState.Empty(),
+            ),
+            Self::ReadingCache => Some(
+                #[hdl(sim)]
+                CacheLookupState.CanceledWhileReadingCache(),
+            ),
+            Self::CacheMiss => Some(
+                #[hdl(sim)]
+                CacheLookupState.CacheMissCanceled(),
+            ),
+            Self::CleaningUpError => Some(
+                #[hdl(sim)]
+                CacheLookupState.CleaningUpErrorCanceled(),
+            ),
+            Self::ErrorCleanUpDone => Some(
+                #[hdl(sim)]
+                CacheLookupState.Empty(),
+            ),
+            Self::Unknown => unreachable!(),
+        }
+    }
+    #[hdl]
+    fn to_canceled(this: &SimValue<Self>) -> SimValue<Self> {
+        Self::try_to_canceled(this).unwrap_or_else(|| this.clone())
+    }
+}
+
+#[hdl(no_static)]
+struct CacheMissFillIndexes<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    next_index: UIntInRangeType<ConstUsize<0>, CpuConfigFetchesPerCacheLine<C>>,
+    config: C,
+}
+
+#[hdl(no_static)]
+struct FetchQueueEntry<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    start_pc: UInt<64>,
+    fetch_block_id: UInt<{ FETCH_BLOCK_ID_WIDTH }>,
+    state: CacheLookupState,
+    fill_indexes_to_start: HdlOption<CacheMissFillIndexes<C>>,
+    fill_indexes_to_finish: HdlOption<CacheMissFillIndexes<C>>,
+    config: C,
+}
+
+impl<C: PhantomConstCpuConfig> FetchQueueEntry<C> {
+    fn is_empty(this: &SimValue<Self>) -> bool {
+        CacheLookupState::is_empty(&this.state)
+    }
+}
+
+impl<C: PhantomConstCpuConfig> SimValueDefault for FetchQueueEntry<C> {
+    #[hdl]
+    fn sim_value_default(self) -> SimValue<Self> {
+        let Self {
+            start_pc: _,
+            fetch_block_id,
+            state: _,
+            fill_indexes_to_start,
+            fill_indexes_to_finish,
+            config,
+        } = self;
+        #[hdl(sim)]
+        Self {
+            start_pc: 0u64,
+            fetch_block_id: fetch_block_id.zero(),
+            state: #[hdl(sim)]
+            CacheLookupState.Empty(),
+            fill_indexes_to_start: #[hdl(sim)]
+            fill_indexes_to_start.HdlNone(),
+            fill_indexes_to_finish: #[hdl(sim)]
+            fill_indexes_to_finish.HdlNone(),
+            config,
+        }
+    }
+}
+
+#[hdl(no_static)]
+struct L1ICacheState<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    queue: ArrayVec<FetchQueueEntry<C>, CpuConfigMaxFetchesInFlight<C>>,
+    config: C,
+}
+
+#[derive(Clone, Debug)]
+struct L1ICacheStateSim<C: PhantomConstCpuConfig> {
+    queue: VecDeque<SimValue<FetchQueueEntry<C>>>,
+    config: C,
+}
+
+#[hdl(no_static)]
+struct WriteBackStep<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    cache_addr: UIntInRangeType<ConstUsize<0>, CpuConfigL1ICacheLineCount<C>>,
+    data: CacheLine<C>,
+    mask: AsMask<CacheLine<C>>,
+}
+
+#[hdl(no_static)]
+struct SplitAddr<C: PhantomConstGet<CpuConfig> + PhantomConstCpuConfig> {
+    byte_in_fetch_block: UIntType<CpuConfigLog2FetchWidthInBytes<C>>,
+    fetch_block_in_cache_line: UIntType<CpuConfigLog2FetchesPerCacheLine<C>>,
+    cache_line_index: UIntType<CpuConfigLog2L1ICacheLineCount<C>>,
+    tag: UIntType<CacheLineTagAddrWidth<C>>,
+}
+
+impl<C: PhantomConstCpuConfig> SplitAddr<C> {
+    fn byte_in_fetch_block_bit_range(self) -> Range<usize> {
+        0..self.byte_in_fetch_block.width()
+    }
+    fn fetch_block_in_cache_line_bit_range(self) -> Range<usize> {
+        let start = self.byte_in_fetch_block_bit_range().end;
+        start..(start + self.fetch_block_in_cache_line.width())
+    }
+    fn cache_line_index_bit_range(self) -> Range<usize> {
+        let start = self.fetch_block_in_cache_line_bit_range().end;
+        start..(start + self.cache_line_index.width())
+    }
+    fn tag_bit_range(self) -> Range<usize> {
+        self.cache_line_index_bit_range().end..64
+    }
+    #[hdl]
+    fn split_addr_sim(self, addr: impl ToSimValueWithType<UInt<64>>) -> SimValue<Self> {
+        let addr = addr.into_sim_value_with_type(UInt::<64>::new_static());
+        #[hdl(sim)]
+        Self {
+            byte_in_fetch_block: addr[self.byte_in_fetch_block_bit_range()],
+            fetch_block_in_cache_line: addr[self.fetch_block_in_cache_line_bit_range()],
+            cache_line_index: addr[self.cache_line_index_bit_range()],
+            tag: addr[self.tag_bit_range()],
+        }
+    }
+    #[hdl]
+    fn addr_sim(this: impl ToSimValue<Type = Self>) -> SimValue<UInt<64>> {
+        #[hdl(sim)]
+        let Self {
+            byte_in_fetch_block,
+            fetch_block_in_cache_line,
+            cache_line_index,
+            tag,
+        } = this;
+        todo!()
+    }
+}
+
+impl<C: PhantomConstCpuConfig> L1ICacheStateSim<C> {
+    #[hdl]
+    fn try_start_memory_operation(&mut self) -> Option<SimValue<MemoryOperationStart<C>>> {
+        for entry in &mut self.queue {
+            #[hdl(sim)]
+            let FetchQueueEntry::<_> {
+                start_pc,
+                fetch_block_id: _,
+                state,
+                fill_indexes_to_start,
+                fill_indexes_to_finish: _,
+                config: _,
+            } = entry;
+            #[hdl(sim)]
+            match state {
+                CacheLookupState::Empty
+                | CacheLookupState::NotStarted
+                | CacheLookupState::CanceledWhileReadingCache
+                | CacheLookupState::ReadingCache
+                | CacheLookupState::CleaningUpErrorCanceled
+                | CacheLookupState::CleaningUpError
+                | CacheLookupState::ErrorCleanUpDone => continue,
+                CacheLookupState::CacheMissCanceled => false,
+                CacheLookupState::CacheMiss => false,
+                CacheLookupState::Unknown => unreachable!(),
+            }
+            #[hdl(sim)]
+            if let HdlSome(indexes) = fill_indexes_to_start {
+                let end_index = indexes.next_index.ty().end();
+                let next_index = *indexes.next_index + 1;
+                if next_index < end_index {
+                    *indexes.next_index = next_index;
+                } else {
+                    let ty = fill_indexes_to_start.ty();
+                    *fill_indexes_to_start = #[hdl(sim)]
+                    ty.HdlNone();
+                }
+                let ty = MemoryOperationStart[self.config];
+                return Some(
+                    #[hdl(sim)]
+                    MemoryOperationStart::<_> {
+                        kind: #[hdl(sim)]
+                        MemoryOperationKind.Read(),
+                        addr: todo!(),
+                        write_data: repeat(ty.write_data.element().zero(), ty.write_data.len()),
+                        config: self.config,
+                    },
+                );
+            }
+        }
+        None
+    }
+    /// tries to do a cache fill, only calls the passed in functions when it's waiting on those
+    #[must_use]
+    #[hdl]
+    fn do_cache_fill<'a>(
+        &mut self,
+        try_finish_memory_operation: impl FnOnce() -> Result<&'a SimValue<MemoryOperationFinish<C>>, ()>,
+    ) -> Option<SimValue<WriteBackStep<C>>> {
+        let mut try_finish_memory_operation = Some(try_finish_memory_operation);
+        for entry in &mut self.queue {
+            #[hdl(sim)]
+            let FetchQueueEntry::<_> {
+                start_pc,
+                fetch_block_id,
+                state,
+                fill_indexes_to_start,
+                fill_indexes_to_finish,
+                config,
+            } = entry;
+            let cleaning_up_error = #[hdl(sim)]
+            match state {
+                CacheLookupState::Empty
+                | CacheLookupState::NotStarted
+                | CacheLookupState::CanceledWhileReadingCache
+                | CacheLookupState::ReadingCache
+                | CacheLookupState::ErrorCleanUpDone => continue,
+                CacheLookupState::CacheMissCanceled => false,
+                CacheLookupState::CacheMiss => false,
+                CacheLookupState::CleaningUpErrorCanceled => true,
+                CacheLookupState::CleaningUpError => true,
+                CacheLookupState::Unknown => unreachable!(),
+            };
+            #[hdl(sim)]
+            if let HdlSome(fill_indexes_to_finish) = fill_indexes_to_finish {
+                match try_finish_memory_operation() {
+                    Ok(_) => todo!(),
+                    Err(()) => {}
+                }
+            }
+        }
+        None
+    }
+}
+
+impl<C: PhantomConstCpuConfig> SimValueDefault for L1ICacheState<C> {
+    #[hdl]
+    fn sim_value_default(self) -> SimValue<Self> {
+        let Self { queue, config } = self;
+        #[hdl(sim)]
+        Self {
+            queue: queue.sim_value_default(),
+            config,
+        }
+    }
+}
+
+impl<C: PhantomConstCpuConfig> ResetSteps for L1ICacheState<C> {
+    #[hdl]
+    fn reset_step(this: &mut SimValue<Self>, step: usize) -> ResetStatus {
+        #[hdl(sim)]
+        let Self {
+            // overwritten every cycle, no reset needed
+            queue: _,
+            config: _,
+        } = this;
+        let _ = step;
+        ResetStatus::Done
+    }
+}
+
+#[hdl_module(extern)]
+fn l1_i_cache_impl(config: PhantomConst<CpuConfig>) {
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let memory_interface: MemoryInterface<PhantomConst<CpuConfig>> =
+        m.output(MemoryInterface[config]);
+    #[hdl]
+    let from_next_pc: NextPcToFetchInterface<PhantomConst<CpuConfig>> =
+        m.input(NextPcToFetchInterface[config]);
+    // i_cache_port.clk is externally overridden with cd.clk
+    #[hdl]
+    let i_cache_port: ReadWriteStruct<CacheLine<PhantomConst<CpuConfig>>, DynSize> = m.output(
+        ReadWriteStruct[CacheLine[config]][memory_addr_width(CpuConfigL1ICacheLineCount[config])],
+    );
+    #[hdl]
+    let state_for_debug: L1ICacheState<PhantomConst<CpuConfig>> = m.output(L1ICacheState[config]);
+    m.register_clock_for_past(cd.clk);
+    #[hdl]
+    async fn run(
+        mut sim: ExternModuleSimulationState,
+        cd: Expr<ClockDomain>,
+        memory_interface: Expr<MemoryInterface<PhantomConst<CpuConfig>>>,
+        from_next_pc: Expr<NextPcToFetchInterface<PhantomConst<CpuConfig>>>,
+        i_cache_port: Expr<ReadWriteStruct<CacheLine<PhantomConst<CpuConfig>>, DynSize>>,
+        state_expr: Expr<L1ICacheState<PhantomConst<CpuConfig>>>,
+    ) {
+        let mut state = sim.read(state_expr).await;
+        let config = state.config.ty();
+        let l1_i_cache_line_count = CpuConfigL1ICacheLineCount[config];
+        let cache_line_ty = CacheLine[config];
+        let entry_ty = FetchQueueEntry[config];
+        let max_fetches_in_flight = CpuConfigMaxFetchesInFlight[config];
+        for step in 0usize.. {
+            sim.write(state_expr, state).await;
+            sim.wait_for_clock_edge(cd.clk).await;
+            state = sim.read_past(state_expr, cd.clk).await;
+            sim.write(i_cache_port.en, false).await;
+            let mut reset_status = ResetSteps::reset_step(&mut state, step);
+            if step < l1_i_cache_line_count {
+                reset_status = ResetStatus::Working;
+                #[hdl]
+                let ReadWriteStruct::<_, _> {
+                    addr,
+                    en,
+                    clk: _, // externally overridden with cd.clk
+                    rdata: _,
+                    wmode,
+                    wdata,
+                    wmask,
+                } = i_cache_port;
+                sim.write(addr, step.cast_to(addr.ty())).await;
+                sim.write(en, true).await;
+                sim.write(wmode, true).await;
+                sim.write(wdata, CacheLine::sim_value_default(wdata.ty()))
+                    .await;
+                sim.write(wmask, splat_mask(cache_line_ty, true.to_expr()))
+                    .await;
+            }
+            match reset_status {
+                ResetStatus::Done => break,
+                ResetStatus::Working => {}
+            }
+        }
+        sim.write(from_next_pc.cancel.ready, true).await;
+        let mut queue = VecDeque::<SimValue<_>>::with_capacity(max_fetches_in_flight);
+        loop {
+            while queue
+                .pop_back_if(|entry| FetchQueueEntry::is_empty(entry))
+                .is_some()
+            {}
+            while queue
+                .pop_front_if(|entry| FetchQueueEntry::is_empty(entry))
+                .is_some()
+            {}
+            sim.write(
+                from_next_pc.fetch.ready,
+                queue.len() < max_fetches_in_flight,
+            )
+            .await;
+            // TODO
+            state.queue = state
+                .queue
+                .ty()
+                .from_iter_sim(state.queue.ty().element().sim_value_default(), &queue)
+                .expect("known to fit");
+            sim.write(state_expr, state).await;
+            sim.wait_for_clock_edge(cd.clk).await;
+            state = sim.read_past(state_expr, cd.clk).await;
+            // handle cancels before pushing new fetch op
+            if sim.read_past_bool(from_next_pc.cancel.ready, cd.clk).await {
+                #[hdl(sim)]
+                if let HdlSome(in_progress_fetches_to_cancel) =
+                    sim.read_past(from_next_pc.cancel.data, cd.clk).await
+                {
+                    let mut in_progress_fetches_to_cancel = *in_progress_fetches_to_cancel;
+                    // cancel in-progress fetches from newest to oldest
+                    for entry in queue.iter_mut().rev() {
+                        if in_progress_fetches_to_cancel == 0 {
+                            break;
+                        }
+                        if let Some(canceled) = CacheLookupState::try_to_canceled(&entry.state) {
+                            entry.state = canceled;
+                            in_progress_fetches_to_cancel -= 1;
+                        }
+                    }
+                }
+            }
+            while queue
+                .pop_back_if(|entry| FetchQueueEntry::is_empty(entry))
+                .is_some()
+            {}
+            if !sim.read_past_bool(from_next_pc.fetch.ready, cd.clk).await {
+                continue;
+            }
+            // handle pushing new fetch op after handling cancels
+            #[hdl(sim)]
+            if let HdlSome(inner) = sim.read_past(from_next_pc.fetch.data, cd.clk).await {
+                #[hdl(sim)]
+                let NextPcToFetchInterfaceInner {
+                    start_pc,
+                    fetch_block_id,
+                } = &inner;
+                queue.push_back(
+                    #[hdl(sim)]
+                    FetchQueueEntry::<_> {
+                        start_pc,
+                        fetch_block_id,
+                        state: CacheLookupState.NotStarted(),
+                        fill_indexes_to_start: entry_ty.fill_indexes_to_start.HdlNone(),
+                        fill_indexes_to_finish: entry_ty.fill_indexes_to_finish.HdlNone(),
+                        config,
+                    },
+                );
+            }
+        }
+    }
+    m.extern_module_simulation_fn(
+        (
+            cd,
+            memory_interface,
+            from_next_pc,
+            i_cache_port,
+            state_for_debug,
+        ),
+        |(cd, memory_interface, from_next_pc, i_cache_port, state_for_debug), mut sim| async move {
+            let config = memory_interface.ty().config;
+            let cache_line_ty = CacheLine[config];
+            sim.write(i_cache_port.clk, false).await; // externally overridden with cd.clk, so just write a constant here
+            sim.resettable(
+                cd,
+                |mut sim: ExternModuleSimulationState| async move {
+                    sim.write(
+                        memory_interface.start.data,
+                        memory_interface.ty().start.data.HdlNone(),
+                    )
+                    .await;
+                    sim.write(memory_interface.finish.ready, false).await;
+                    sim.write(
+                        from_next_pc.next_fetch_block_ids,
+                        from_next_pc.ty().next_fetch_block_ids.HdlNone(),
+                    )
+                    .await;
+                    sim.write(from_next_pc.fetch.ready, false).await;
+                    sim.write(from_next_pc.cancel.ready, false).await;
+                    sim.write(i_cache_port.addr, 0u8.cast_to(i_cache_port.addr.ty()))
+                        .await;
+                    sim.write(i_cache_port.en, false).await;
+                    sim.write(i_cache_port.wmode, false).await;
+                    sim.write(
+                        i_cache_port.wdata,
+                        CacheLine::sim_value_default(cache_line_ty),
+                    )
+                    .await;
+                    sim.write(
+                        i_cache_port.wmask,
+                        splat_mask(cache_line_ty, false.to_expr()),
+                    )
+                    .await;
+                    sim.write(state_for_debug, state_for_debug.ty().sim_value_default())
+                        .await;
+                },
+                |sim, ()| {
+                    run(
+                        sim,
+                        cd,
+                        memory_interface,
+                        from_next_pc,
+                        i_cache_port,
+                        state_for_debug,
+                    )
+                },
+            )
+            .await;
+        },
+    );
+}
+
+/// implements a direct-mapped L1 I-Cache
+#[hdl_module]
+pub fn l1_i_cache(config: PhantomConst<CpuConfig>) {
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let memory_interface: MemoryInterface<PhantomConst<CpuConfig>> =
+        m.output(MemoryInterface[config]);
+    #[hdl]
+    let from_next_pc: NextPcToFetchInterface<PhantomConst<CpuConfig>> =
+        m.input(NextPcToFetchInterface[config]);
+    let cache_line_ty = CacheLine[config];
+    let cache_line_count = CpuConfigL1ICacheLineCount[config];
+    // TODO: convert to memory with single read/write port once semantics
+    // for read/write latencies are properly implemented in the simulator:
+    // https://git.libre-chip.org/libre-chip/fayalite/src/commit/c632e5d570d4763e8e18d764e95b7a9e515ebf99/crates/fayalite/src/sim/compiler.rs#L4774
+    // which depends on:
+    // https://github.com/chipsalliance/firrtl-spec/issues/263
+    #[hdl]
+    let i_cache = reg_builder()
+        .clock_domain(cd)
+        .no_reset(ArrayType[cache_line_ty][cache_line_count]);
+    #[hdl]
+    let l1_i_cache_impl = instance(l1_i_cache_impl(config));
+    connect(l1_i_cache_impl.cd, cd);
+    connect(memory_interface, l1_i_cache_impl.memory_interface);
+    connect(l1_i_cache_impl.from_next_pc, from_next_pc);
+    #[hdl]
+    let ReadWriteStruct::<_, _> {
+        addr,
+        en,
+        clk: _,
+        rdata,
+        wmode,
+        wdata,
+        wmask,
+    } = l1_i_cache_impl.i_cache_port;
+    connect(rdata, rdata.ty().uninit());
+    #[hdl]
+    if en {
+        let i_cache_line = i_cache[addr];
+        #[hdl]
+        if wmode {
+            #[hdl]
+            let CacheLine::<_> {
+                data: wdata_data,
+                addr: wdata_addr,
+                config: _,
+            } = wdata;
+            for ((dest, src), mask) in i_cache_line
+                .data
+                .into_iter()
+                .zip(wdata_data)
+                .zip(wmask.data)
+            {
+                #[hdl]
+                if mask {
+                    connect(dest, src);
+                }
+            }
+            #[hdl]
+            if wmask.addr {
+                connect(i_cache_line.addr, wdata_addr);
+            }
+        } else {
+            connect(rdata, i_cache_line);
+        }
+    }
+}
+
+#[hdl_module]
+pub fn fetch(config: PhantomConst<CpuConfig>) {
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let memory_interface: MemoryInterface<PhantomConst<CpuConfig>> =
+        m.output(MemoryInterface[config]);
+    #[hdl]
+    let from_next_pc: NextPcToFetchInterface<PhantomConst<CpuConfig>> =
+        m.input(NextPcToFetchInterface[config]);
+    #[hdl]
+    let l1_i_cache = instance(l1_i_cache(config));
+    connect(l1_i_cache.cd, cd);
+    connect(memory_interface, l1_i_cache.memory_interface);
+    connect(l1_i_cache.from_next_pc, from_next_pc);
+}

crates/cpu/src/lib.rs

@@ -2,6 +2,7 @@
 // See Notices.txt for copyright information
 pub mod config;
 pub mod decoder;
+pub mod fetch;
 pub mod instruction;
 pub mod next_pc;
 pub mod powerisa_instructions_xml;

crates/cpu/src/next_pc.rs

@@ -2719,13 +2719,13 @@ impl SimValueDefault for BranchPredictionState {
 
 #[derive(Copy, Clone, Debug)]
 #[must_use]
-enum ResetStatus {
+pub(crate) enum ResetStatus {
     Done,
     Working,
 }
 
 impl ResetStatus {
-    fn and(self, other: Self) -> Self {
+    pub(crate) fn and(self, other: Self) -> Self {
         match (self, other) {
             (ResetStatus::Done, ResetStatus::Done) => ResetStatus::Done,
             (ResetStatus::Done | ResetStatus::Working, ResetStatus::Working)
@@ -2734,7 +2734,7 @@ impl ResetStatus {
     }
 }
 
-trait SimValueDefault: Type {
+pub(crate) trait SimValueDefault: Type {
     fn sim_value_default(self) -> SimValue<Self>;
 }
 
@@ -2828,7 +2828,7 @@ impl SimValueDefault for WipDecodedInsn {
     }
 }
 
-trait ResetSteps: Type {
+pub(crate) trait ResetSteps: Type {
     fn reset_step(this: &mut SimValue<Self>, step: usize) -> ResetStatus;
 }
 

crates/cpu/src/util/array_vec.rs

@@ -2,6 +2,24 @@
 // See Notices.txt for copyright information
 
 use fayalite::{expr::ops::ExprIndex, int::UIntInRangeInclusiveType, prelude::*};
+use std::fmt;
+
+#[derive(Clone, Debug)]
+pub struct ArrayVecFullError<V, I: Iterator> {
+    pub value: V,
+    pub rest: std::iter::Chain<std::iter::Once<I::Item>, I>,
+}
+
+impl<V, I: Iterator> fmt::Display for ArrayVecFullError<V, I> {
+    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
+        write!(f, "ArrayVec is full")
+    }
+}
+
+impl<V: fmt::Debug, I: Iterator<Item: fmt::Debug> + fmt::Debug> std::error::Error
+    for ArrayVecFullError<V, I>
+{
+}
 
 #[hdl]
 pub type Length<Max: Size> = UIntInRangeInclusiveType<ConstUsize<0>, Max>;
@@ -46,6 +64,29 @@ impl<T: Type, N: Size> ArrayVec<T, N> {
             len: self.elements.len().to_sim_value_with_type(self.len),
         }
     }
+    pub fn from_iter_sim<I: IntoIterator<Item: ToSimValueWithType<T>>>(
+        self,
+        uninit_element: impl ToSimValueWithType<T>,
+        iter: I,
+    ) -> Result<SimValue<Self>, ArrayVecFullError<SimValue<Self>, I::IntoIter>> {
+        let mut value = Self::new_sim(self, uninit_element);
+        let element = self.element();
+        let mut iter = iter.into_iter();
+        for i in 0..self.capacity() {
+            let Some(v) = iter.next() else {
+                break;
+            };
+            value.elements[i] = v.into_sim_value_with_type(element);
+        }
+        if let Some(extra) = iter.next() {
+            Err(ArrayVecFullError {
+                value,
+                rest: std::iter::once(extra).chain(iter),
+            })
+        } else {
+            Ok(value)
+        }
+    }
     pub fn element(self) -> T {
         self.elements.element()
     }

crates/cpu/tests/fetch.rs

@@ -0,0 +1,198 @@
+// SPDX-License-Identifier: LGPL-3.0-or-later
+// See Notices.txt for copyright information
+
+use cpu::{
+    config::{CpuConfig, UnitConfig},
+    fetch::{MemoryInterface, fetch},
+    next_pc::NextPcToFetchInterface,
+    unit::UnitKind,
+    util::array_vec::ArrayVec,
+};
+use fayalite::{
+    prelude::*,
+    sim::vcd::VcdWriterDecls,
+    util::{DebugAsDisplay, RcWriter},
+};
+use std::{cell::Cell, collections::VecDeque, num::NonZeroUsize};
+
+const MEMORY_QUEUE_SIZE: usize = 32;
+
+#[hdl]
+struct MemoryQueueEntry {
+    addr: UInt<64>,
+    cycles_left: UInt<8>,
+}
+
+impl MemoryQueueEntry {
+    #[hdl]
+    fn default_sim(self) -> SimValue<Self> {
+        #[hdl(sim)]
+        Self {
+            addr: 0u64,
+            cycles_left: 0u8,
+        }
+    }
+}
+
+#[hdl_module(extern)]
+fn mock_memory(config: PhantomConst<CpuConfig>) {
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let memory_interface: MemoryInterface<PhantomConst<CpuConfig>> =
+        m.input(MemoryInterface[config]);
+    #[hdl]
+    let queue_debug: ArrayVec<MemoryQueueEntry, ConstUsize<{ MEMORY_QUEUE_SIZE }>> = m.output();
+    m.register_clock_for_past(cd.clk);
+    m.extern_module_simulation_fn(
+        (cd, memory_interface, queue_debug),
+        |(cd, memory_interface, queue_debug), mut sim| async move {
+            // intentionally have a different sequence each time we're reset
+            let delay_sequence_index = Cell::new(0);
+            sim.resettable(
+                cd,
+                async |mut sim| {
+                    sim.write(memory_interface.start.ready, false).await;
+                    sim.write(
+                        memory_interface.finish.data,
+                        memory_interface.ty().finish.data.HdlNone(),
+                    )
+                    .await;
+                    sim.write(
+                        queue_debug,
+                        queue_debug.ty().new_sim(MemoryQueueEntry.default_sim()),
+                    )
+                    .await;
+                },
+                |sim, ()| {
+                    run_fn(
+                        cd,
+                        memory_interface,
+                        queue_debug,
+                        &delay_sequence_index,
+                        sim,
+                    )
+                },
+            )
+            .await;
+        },
+    );
+    #[hdl]
+    async fn run_fn(
+        cd: Expr<ClockDomain>,
+        memory_interface: Expr<MemoryInterface<PhantomConst<CpuConfig>>>,
+        queue_debug: Expr<ArrayVec<MemoryQueueEntry, ConstUsize<{ MEMORY_QUEUE_SIZE }>>>,
+        delay_sequence_index: &Cell<u64>,
+        mut sim: ExternModuleSimulationState,
+    ) {
+        let config = memory_interface.config.ty();
+        let mut queue: VecDeque<SimValue<MemoryQueueEntry>> = VecDeque::new();
+        loop {
+            let mut sim_queue = queue_debug.ty().new_sim(MemoryQueueEntry.default_sim());
+            for entry in &queue {
+                ArrayVec::try_push_sim(&mut sim_queue, entry)
+                    .ok()
+                    .expect("queue is known to be small enough");
+            }
+            sim.write(queue_debug, sim_queue).await;
+            // TODO:
+            sim.wait_for_clock_edge(cd.clk).await;
+            println!(
+                "Dump mock memory queue: {:#?}",
+                Vec::from_iter(
+                    queue
+                        .iter()
+                        .map(|v| { DebugAsDisplay(format!("addr={:#x}", v.addr.as_int())) })
+                )
+            );
+        }
+    }
+}
+
+#[hdl_module]
+fn dut(config: PhantomConst<CpuConfig>) {
+    #[hdl]
+    let cd: ClockDomain = m.input();
+    #[hdl]
+    let from_next_pc: NextPcToFetchInterface<PhantomConst<CpuConfig>> =
+        m.input(NextPcToFetchInterface[config]);
+    #[hdl]
+    let fetch = instance(fetch(config));
+    #[hdl]
+    let fetch {
+        cd: fetch_cd,
+        memory_interface: fetch_memory_interface,
+        from_next_pc: fetch_from_next_pc,
+    } = fetch;
+    connect(fetch_cd, cd);
+    connect(fetch_from_next_pc, from_next_pc);
+    #[hdl]
+    let mock_memory = instance(mock_memory(config));
+    #[hdl]
+    let mock_memory {
+        cd: mock_memory_cd,
+        memory_interface: mock_memory_interface,
+        queue_debug: _,
+    } = mock_memory;
+    connect(mock_memory_cd, cd);
+    connect(mock_memory_interface, fetch_memory_interface);
+}
+
+#[test]
+fn test_fetch() {
+    let _n = SourceLocation::normalize_files_for_tests();
+    let mut config = CpuConfig::new(
+        vec![
+            UnitConfig::new(UnitKind::AluBranch),
+            UnitConfig::new(UnitKind::AluBranch),
+        ],
+        NonZeroUsize::new(20).unwrap(),
+    );
+    config.fetch_width = NonZeroUsize::new(2).unwrap();
+    config.log2_fetch_width_in_bytes = 4;
+    config.l1_i_cache_line_count = NonZeroUsize::new(16).unwrap();
+    let m = dut(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),
+    };
+    let from_next_pc_ty = sim.io().from_next_pc.ty();
+    sim.write_clock(sim.io().cd.clk, false);
+    sim.write_reset(sim.io().cd.rst, true);
+    sim.write(
+        sim.io().from_next_pc.cancel.data,
+        from_next_pc_ty.cancel.data.HdlNone(),
+    );
+    sim.write(
+        sim.io().from_next_pc.fetch.data,
+        from_next_pc_ty.fetch.data.HdlNone(),
+    );
+    for cycle in 0..2000 {
+        // TODO: drive from_next_pc
+        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);
+    }
+    // FIXME: vcd is just whatever fetch 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/fetch.vcd") {
+        panic!();
+    }
+}