From 024a85ec724039c1f82ced433ea07d010fe7a066 Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?H=C3=A5vard=20Anda=20Estensen?= <haavard.ae@gmail.com>
Date: Fri, 6 Dec 2024 12:10:25 +0100
Subject: [PATCH] perf: use precomputed zero hashes in process_subtree

---
 ssz-rs/Cargo.toml                     |  2 ++
 ssz-rs/src/merkleization/merkleize.rs | 20 +++++++++++++++++++-
 2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/ssz-rs/Cargo.toml b/ssz-rs/Cargo.toml
index 36f11cea..ec56eb38 100644
--- a/ssz-rs/Cargo.toml
+++ b/ssz-rs/Cargo.toml
@@ -19,6 +19,8 @@ serde = ["dep:serde", "alloy-primitives/serde"]
 
 [dependencies]
 bitvec = { version = "1.0.0", default-features = false, features = ["alloc"] }
+lazy_static = "1.5"
+once_cell = "1.20"
 rayon = "1.10"
 ssz_rs_derive = { path = "../ssz-rs-derive", version = "0.9.0" }
 sha2 = { version = "0.9.8", default-features = false }
diff --git a/ssz-rs/src/merkleization/merkleize.rs b/ssz-rs/src/merkleization/merkleize.rs
index 7efb8899..54ccd54a 100644
--- a/ssz-rs/src/merkleization/merkleize.rs
+++ b/ssz-rs/src/merkleization/merkleize.rs
@@ -7,6 +7,7 @@ use crate::{
 };
 #[cfg(feature = "serde")]
 use alloy_primitives::hex::FromHex;
+use once_cell::sync::Lazy;
 
 // The generalized index for the root of the "decorated" type in any Merkleized type that supports
 // decoration.
@@ -25,6 +26,13 @@ pub trait HashTreeRoot {
     }
 }
 
+/// Precomputed hash of two 32-byte zero arrays concatenated.
+static ZERO_LEAF_HASH: Lazy<[u8; BYTES_PER_CHUNK]> = Lazy::new(|| {
+    let left_zero = [0u8; BYTES_PER_CHUNK];
+    let right_zero = [0u8; BYTES_PER_CHUNK];
+    hash_chunks(&left_zero, &right_zero)
+});
+
 // Ensures `buffer` can be exactly broken up into `BYTES_PER_CHUNK` chunks of bytes
 // via padding any partial chunks at the end of `buffer`
 pub fn pack_bytes(buffer: &mut Vec<u8>) {
@@ -464,8 +472,18 @@ fn process_subtree(buffer: &mut [u8], size: usize) {
         let left = &buffer[i * BYTES_PER_CHUNK..(i + 1) * BYTES_PER_CHUNK];
         let right = &buffer[(i + 1) * BYTES_PER_CHUNK..(i + 2) * BYTES_PER_CHUNK];
 
+        // Check if both left and right are all zero
+        let left_is_zero = left.iter().all(|&byte| byte == 0);
+        let right_is_zero = right.iter().all(|&byte| byte == 0);
+
         // Compute parent hash
-        let hash = hash_chunks(left, right);
+        let hash = if left_is_zero && right_is_zero {
+            // Use the precomputed zero hash
+            *ZERO_LEAF_HASH
+        } else {
+            // Compute the hash normally
+            hash_chunks(left, right)
+        };
 
         // Store at parent position (i/2)
         // SAFETY: checked subtraction is unnecessary, as i >= 1; qed