speed up simulation by optimizing SimulationImpl::read_traces

this makes cpu/crates/cpu/tests/next_pc.rs take 56s instead of 168s

crates/fayalite/src/sim.rs

@@ -40,7 +40,7 @@ use crate::{
         OpaqueSimValue, OpaqueSimValueSize, OpaqueSimValueSizeRange, OpaqueSimValueSlice,
         OpaqueSimValueWriter,
     },
-    util::{BitSliceWriteWithBase, DebugAsDisplay, HashMap, HashSet},
+    util::{BitSliceWriteWithBase, DebugAsDisplay, HashMap, HashSet, copy_le_bytes_to_bitslice},
 };
 use bitvec::{bits, order::Lsb0, slice::BitSlice, vec::BitVec, view::BitView};
 use num_bigint::BigInt;
@@ -2198,14 +2198,11 @@ impl SimulationImpl {
                 SimTraceKind::BigUInt { index, ty: _ } | SimTraceKind::BigSInt { index, ty: _ } => {
                     let state = state.unwrap_bits_mut();
                     let bigint = &self.state.big_slots[index];
-                    let mut bytes = bigint.to_signed_bytes_le();
+                    copy_le_bytes_to_bitslice(
-                    bytes.resize(
+                        state,
-                        state.len().div_ceil(8),
+                        &bigint.to_signed_bytes_le(),
-                        if bigint.is_negative() { 0xFF } else { 0 },
+                        bigint.is_negative(),
                     );
-                    let bitslice = BitSlice::<u8, Lsb0>::from_slice(&bytes);
-                    let bitslice = &bitslice[..state.len()];
-                    state.clone_from_bitslice(bitslice);
                 }
                 SimTraceKind::BigBool { index }
                 | SimTraceKind::BigAsyncReset { index }

crates/fayalite/src/util.rs

@@ -41,7 +41,7 @@ pub use misc::{
     os_str_strip_suffix, serialize_to_json_ascii, serialize_to_json_ascii_pretty,
     serialize_to_json_ascii_pretty_with_indent, slice_range, try_slice_range,
 };
-pub(crate) use misc::{InternedStrCompareAsStr, chain};
+pub(crate) use misc::{InternedStrCompareAsStr, chain, copy_le_bytes_to_bitslice};
 
 pub mod job_server;
 pub mod prefix_sum;

crates/fayalite/src/util/misc.rs

@@ -612,3 +612,43 @@ impl std::borrow::Borrow<str> for InternedStrCompareAsStr {
         &self.0
     }
 }
+
+pub(crate) fn copy_le_bytes_to_bitslice(
+    dest: &mut BitSlice<usize, Lsb0>,
+    bytes: &[u8],
+    msb_fill: bool,
+) {
+    let (chunks, remainder) = bytes.as_chunks();
+    let mut filled_to = 0;
+    for (i, chunk) in chunks.iter().enumerate() {
+        if let Some(start_bit_index) = i.checked_mul(usize::BITS as usize)
+            && start_bit_index < dest.len()
+        {
+            let end_bit_index = start_bit_index
+                .saturating_add(usize::BITS as usize)
+                .min(dest.len());
+            let bit_len = end_bit_index - start_bit_index;
+            let chunk = usize::from_le_bytes(*chunk);
+            dest[start_bit_index..end_bit_index].copy_from_bitslice(&chunk.view_bits()[..bit_len]);
+            filled_to = end_bit_index;
+        } else {
+            break;
+        }
+    }
+    if !remainder.is_empty() {
+        if let Some(start_bit_index) = chunks.len().checked_mul(usize::BITS as usize)
+            && start_bit_index < dest.len()
+        {
+            let end_bit_index = start_bit_index
+                .saturating_add(usize::BITS as usize)
+                .min(dest.len());
+            let bit_len = end_bit_index - start_bit_index;
+            let mut chunk = [if msb_fill { !0 } else { 0 }; _];
+            chunk[..remainder.len()].copy_from_slice(remainder);
+            let chunk = usize::from_le_bytes(chunk);
+            dest[start_bit_index..end_bit_index].copy_from_bitslice(&chunk.view_bits()[..bit_len]);
+            filled_to = end_bit_index;
+        }
+    }
+    dest[filled_to..].fill(msb_fill);
+}