feat: add a sha256 implementation which is optimized for unconstrained runtime

.github/workflows/benchmark.yml

@@ -21,8 +21,10 @@ jobs:
 
       - name: Install bb
         run: |
-          npm install -g bbup
-          bbup -nv 0.36.0
+          curl -L https://raw.githubusercontent.com/AztecProtocol/aztec-packages/refs/heads/master/barretenberg/bbup/install | bash
+          echo "$HOME/.bb/" >> $GITHUB_PATH
+          ~/.bb/bbup -nv 0.36.0
+          sudo apt install -y libc++-dev
 
       - name: Build Noir benchmark programs
         run: nargo export
@@ -34,7 +36,7 @@ jobs:
 
       - name: Compare gates reports
         id: gates_diff
-        uses: noir-lang/noir-gates-diff@7e4ddaa91c69380f15ccba514eac17bc7432a8cc
+        uses: noir-lang/noir-gates-diff@dbe920a8dcc3370af4be4f702ca9cef29317bec1
         with:
           report: gates_report.json
           summaryQuantile: 0.9 # only display the 10% most significant circuit size diffs in the summary (defaults to 20%)

src/lib.nr

@@ -2,3 +2,4 @@ mod sha256;
 
 pub use sha256::digest;
 pub use sha256::sha256_var;
+

src/sha256.nr

@@ -1,51 +1,17 @@
 use std::hash::sha256_compression;
 use std::runtime::is_unconstrained;
 
+use constants::{
+    BLOCK_BYTE_PTR, BLOCK_SIZE, HASH, INITIAL_STATE, INT_BLOCK, INT_BLOCK_SIZE, INT_SIZE,
+    INT_SIZE_PTR, MSG_BLOCK, MSG_SIZE_PTR, STATE, TWO_POW_16, TWO_POW_24, TWO_POW_32, TWO_POW_8,
+};
+
+mod constants;
 mod tests;
 
 // Implementation of SHA-256 mapping a byte array of variable length to
 // 32 bytes.
 
-// A message block is up to 64 bytes taken from the input.
-global BLOCK_SIZE: u32 = 64;
-
-// The first index in the block where the 8 byte message size will be written.
-global MSG_SIZE_PTR: u32 = 56;
-
-// Size of the message block when packed as 4-byte integer array.
-global INT_BLOCK_SIZE: u32 = 16;
-
-// A `u32` integer consists of 4 bytes.
-global INT_SIZE: u32 = 4;
-
-// Index of the integer in the `INT_BLOCK` where the length is written.
-global INT_SIZE_PTR: u32 = MSG_SIZE_PTR / INT_SIZE;
-
-// Magic numbers for bit shifting.
-// Works with actual bit shifting as well as the compiler turns them into * and /
-// but circuit execution appears to be 10% faster this way.
-global TWO_POW_8: u32 = 256;
-global TWO_POW_16: u32 = TWO_POW_8 * 256;
-global TWO_POW_24: u32 = TWO_POW_16 * 256;
-global TWO_POW_32: u64 = TWO_POW_24 as u64 * 256;
-
-// Index of a byte in a 64 byte block; ie. 0..=63
-type BLOCK_BYTE_PTR = u32;
-
-// The foreign function to compress blocks works on 16 pieces of 4-byte integers, instead of 64 bytes.
-type INT_BLOCK = [u32; INT_BLOCK_SIZE];
-
-// A message block is a slice of the original message of a fixed size,
-// potentially padded with zeros, with neighbouring 4 bytes packed into integers.
-type MSG_BLOCK = INT_BLOCK;
-
-// The hash is 32 bytes.
-type HASH = [u8; 32];
-
-// The state accumulates the blocks.
-// Its overall size is the same as the `HASH`.
-type STATE = [u32; 8];
-
 // Deprecated in favour of `sha256_var`
 // docs:start:sha256
 pub fn sha256<let N: u32>(input: [u8; N]) -> HASH
@@ -63,94 +29,132 @@ pub fn digest<let N: u32>(msg: [u8; N]) -> HASH {
 // Variable size SHA-256 hash
 pub fn sha256_var<let N: u32>(msg: [u8; N], message_size: u64) -> HASH {
     let message_size = message_size as u32;
-    let num_blocks = N / BLOCK_SIZE;
-    let mut msg_block: MSG_BLOCK = [0; INT_BLOCK_SIZE];
-    // Intermediate hash, starting with the canonical initial value
-    let mut h: STATE = [
-        1779033703, 3144134277, 1013904242, 2773480762, 1359893119, 2600822924, 528734635,
-        1541459225,
-    ];
-    // Pointer into msg_block on a 64 byte scale
-    let mut msg_byte_ptr = 0;
-    for i in 0..num_blocks {
-        let msg_start = BLOCK_SIZE * i;
-        let (new_msg_block, new_msg_byte_ptr) =
-            unsafe { build_msg_block(msg, message_size, msg_start) };
+    assert(message_size <= N);
 
-        if msg_start < message_size {
-            msg_block = new_msg_block;
+    if std::runtime::is_unconstrained() {
+        // SAfety: SHA256 is running as an unconstrained function.
+        unsafe {
+            __sha256_var(msg, message_size)
         }
+    } else {
+        let mut msg_block: MSG_BLOCK = [0; INT_BLOCK_SIZE];
+        // Intermediate hash, starting with the canonical initial value
+        let mut h: STATE = INITIAL_STATE;
+        // Pointer into msg_block on a 64 byte scale
+        let mut msg_byte_ptr = 0;
+        let num_blocks = N / BLOCK_SIZE;
+        for i in 0..num_blocks {
+            let msg_start = BLOCK_SIZE * i;
+            let (new_msg_block, new_msg_byte_ptr) =
+                unsafe { build_msg_block(msg, message_size, msg_start) };
+
+            if msg_start < message_size {
+                msg_block = new_msg_block;
+            }
 
-        if !is_unconstrained() {
             // Verify the block we are compressing was appropriately constructed
             let new_msg_byte_ptr = verify_msg_block(msg, message_size, msg_block, msg_start);
             if msg_start < message_size {
                 msg_byte_ptr = new_msg_byte_ptr;
             }
-        } else if msg_start < message_size {
-            msg_byte_ptr = new_msg_byte_ptr;
-        }
 
-        // If the block is filled, compress it.
-        // An un-filled block is handled after this loop.
-        if (msg_start < message_size) & (msg_byte_ptr == BLOCK_SIZE) {
-            h = sha256_compression(msg_block, h);
+            // If the block is filled, compress it.
+            // An un-filled block is handled after this loop.
+            if (msg_start < message_size) & (msg_byte_ptr == BLOCK_SIZE) {
+                h = sha256_compression(msg_block, h);
+            }
         }
-    }
 
-    let modulo = N % BLOCK_SIZE;
-    // Handle setup of the final msg block.
-    // This case is only hit if the msg is less than the block size,
-    // or our message cannot be evenly split into blocks.
-    if modulo != 0 {
-        let msg_start = BLOCK_SIZE * num_blocks;
-        let (new_msg_block, new_msg_byte_ptr) =
-            unsafe { build_msg_block(msg, message_size, msg_start) };
+        let modulo = N % BLOCK_SIZE;
+        // Handle setup of the final msg block.
+        // This case is only hit if the msg is less than the block size,
+        // or our message cannot be evenly split into blocks.
+        if modulo != 0 {
+            let msg_start = BLOCK_SIZE * num_blocks;
+            let (new_msg_block, new_msg_byte_ptr) =
+                unsafe { build_msg_block(msg, message_size, msg_start) };
 
-        if msg_start < message_size {
-            msg_block = new_msg_block;
-        }
+            if msg_start < message_size {
+                msg_block = new_msg_block;
+            }
 
-        if !is_unconstrained() {
             let new_msg_byte_ptr = verify_msg_block(msg, message_size, msg_block, msg_start);
             if msg_start < message_size {
                 msg_byte_ptr = new_msg_byte_ptr;
                 verify_msg_block_padding(msg_block, msg_byte_ptr);
             }
-        } else if msg_start < message_size {
-            msg_byte_ptr = new_msg_byte_ptr;
         }
+
+        // If we had modulo == 0 then it means the last block was full,
+        // and we can reset the pointer to zero to overwrite it.
+        if msg_byte_ptr == BLOCK_SIZE {
+            msg_byte_ptr = 0;
+        }
+
+        // Pad the rest such that we have a [u32; 2] block at the end representing the length
+        // of the message, and a block of 1 0 ... 0 following the message (i.e. [1 << 7, 0, ..., 0]).
+        // Here we rely on the fact that everything beyond the available input is set to 0.
+        let index = msg_byte_ptr / INT_SIZE;
+        msg_block[index] = set_item_byte_then_zeros(msg_block[index], msg_byte_ptr, 1 << 7);
+
+        msg_byte_ptr = msg_byte_ptr + 1;
+        let last_block = msg_block;
+
+        // If we don't have room to write the size, compress the block and reset it.
+        if msg_byte_ptr > MSG_SIZE_PTR {
+            h = sha256_compression(msg_block, h);
+            // `attach_len_to_msg_block` will zero out everything after the `msg_byte_ptr`.
+            msg_byte_ptr = 0;
+        }
+
+        msg_block = unsafe { attach_len_to_msg_block(msg_block, msg_byte_ptr, message_size) };
+
+        verify_msg_len(msg_block, last_block, msg_byte_ptr, message_size);
+
+        hash_final_block(msg_block, h)
     }
+}
 
-    // If we had modulo == 0 then it means the last block was full,
-    // and we can reset the pointer to zero to overwrite it.
-    if msg_byte_ptr == BLOCK_SIZE {
-        msg_byte_ptr = 0;
+// Variable size SHA-256 hash
+unconstrained fn __sha256_var<let N: u32>(msg: [u8; N], message_size: u32) -> HASH {
+    let num_full_blocks = message_size / BLOCK_SIZE;
+    // Intermediate hash, starting with the canonical initial value
+    let mut h: STATE = INITIAL_STATE;
+    // Pointer into msg_block on a 64 byte scale
+    for i in 0..num_full_blocks {
+        let (msg_block, _) = build_msg_block(msg, message_size, BLOCK_SIZE * i);
+        h = sha256_compression(msg_block, h);
     }
 
+    // Handle setup of the final msg block.
+    // This case is only hit if the msg is less than the block size,
+    // or our message cannot be evenly split into blocks.
+    let modulo = message_size % BLOCK_SIZE;
+    let (mut msg_block, mut msg_byte_ptr): (INT_BLOCK, u32) = if modulo != 0 {
+        let msg_start = BLOCK_SIZE * num_full_blocks;
+        let (new_msg_block, new_msg_byte_ptr) = build_msg_block(msg, message_size, msg_start);
+
+        (new_msg_block, new_msg_byte_ptr)
+    } else {
+        // If we had modulo == 0 then it means the last block was full,
+        // and we can reset the pointer to zero to overwrite it.
+        ([0; INT_BLOCK_SIZE], 0)
+    };
+
     // Pad the rest such that we have a [u32; 2] block at the end representing the length
     // of the message, and a block of 1 0 ... 0 following the message (i.e. [1 << 7, 0, ..., 0]).
     // Here we rely on the fact that everything beyond the available input is set to 0.
-    msg_block = update_block_item(
-        msg_block,
-        msg_byte_ptr,
-        |msg_item| set_item_byte_then_zeros(msg_item, msg_byte_ptr, 1 << 7),
-    );
-    msg_byte_ptr = msg_byte_ptr + 1;
-    let last_block = msg_block;
+    let index = msg_byte_ptr / INT_SIZE;
+    msg_block[index] = set_item_byte_then_zeros(msg_block[index], msg_byte_ptr, 1 << 7);
 
     // If we don't have room to write the size, compress the block and reset it.
-    if msg_byte_ptr > MSG_SIZE_PTR {
-        h = sha256_compression(msg_block, h);
+    let (h, mut msg_byte_ptr): (STATE, u32) = if msg_byte_ptr >= MSG_SIZE_PTR {
         // `attach_len_to_msg_block` will zero out everything after the `msg_byte_ptr`.
-        msg_byte_ptr = 0;
-    }
-
-    msg_block = unsafe { attach_len_to_msg_block(msg_block, msg_byte_ptr, message_size) };
-
-    if !is_unconstrained() {
-        verify_msg_len(msg_block, last_block, msg_byte_ptr, message_size);
-    }
+        (sha256_compression(msg_block, h), 0)
+    } else {
+        (h, msg_byte_ptr + 1)
+    };
+    msg_block = attach_len_to_msg_block(msg_block, msg_byte_ptr, message_size);
 
     hash_final_block(msg_block, h)
 }
@@ -332,17 +336,6 @@ fn verify_msg_block_equals_last(
     }
 }
 
-// Apply a function on the block item which the pointer indicates.
-fn update_block_item<Env>(
-    mut msg_block: MSG_BLOCK,
-    msg_byte_ptr: BLOCK_BYTE_PTR,
-    f: fn[Env](u32) -> u32,
-) -> MSG_BLOCK {
-    let i = msg_byte_ptr / INT_SIZE;
-    msg_block[i] = f(msg_block[i]);
-    msg_block
-}
-
 // Set the rightmost `zeros` number of bytes to 0.
 fn set_item_zeros(item: u32, zeros: u8) -> u32 {
     lshift8(rshift8(item, zeros), zeros)
@@ -517,3 +510,13 @@ fn hash_final_block(msg_block: MSG_BLOCK, mut state: STATE) -> HASH {
     out_h
 }
 
+mod equivalence_test {
+
+    #[test]
+    fn test_implementations_agree(msg: [u8; 100], message_size: u64) {
+        let message_size = message_size % 100;
+        let unconstrained_sha = unsafe { super::__sha256_var(msg, message_size as u32) };
+        let sha = super::sha256_var(msg, message_size);
+        assert_eq(sha, unconstrained_sha);
+    }
+}

src/sha256/constants.nr

@@ -0,0 +1,42 @@
+// A message block is up to 64 bytes taken from the input.
+pub(crate) global BLOCK_SIZE: u32 = 64;
+
+// The first index in the block where the 8 byte message size will be written.
+pub(crate) global MSG_SIZE_PTR: u32 = 56;
+
+// Size of the message block when packed as 4-byte integer array.
+pub(crate) global INT_BLOCK_SIZE: u32 = 16;
+
+// A `u32` integer consists of 4 bytes.
+pub(crate) global INT_SIZE: u32 = 4;
+
+// Index of the integer in the `INT_BLOCK` where the length is written.
+pub(crate) global INT_SIZE_PTR: u32 = MSG_SIZE_PTR / INT_SIZE;
+
+// Magic numbers for bit shifting.
+// Works with actual bit shifting as well as the compiler turns them into * and /
+// but circuit execution appears to be 10% faster this way.
+pub(crate) global TWO_POW_8: u32 = 256;
+pub(crate) global TWO_POW_16: u32 = TWO_POW_8 * 256;
+pub(crate) global TWO_POW_24: u32 = TWO_POW_16 * 256;
+pub(crate) global TWO_POW_32: u64 = TWO_POW_24 as u64 * 256;
+
+// Index of a byte in a 64 byte block; ie. 0..=63
+pub(crate) type BLOCK_BYTE_PTR = u32;
+
+// The foreign function to compress blocks works on 16 pieces of 4-byte integers, instead of 64 bytes.
+pub(crate) type INT_BLOCK = [u32; INT_BLOCK_SIZE];
+
+// A message block is a slice of the original message of a fixed size,
+// potentially padded with zeros, with neighbouring 4 bytes packed into integers.
+pub(crate) type MSG_BLOCK = INT_BLOCK;
+
+// The hash is 32 bytes.
+pub(crate) type HASH = [u8; 32];
+
+// The state accumulates the blocks.
+// Its overall size is the same as the `HASH`.
+pub(crate) type STATE = [u32; 8];
+
+pub(crate) global INITIAL_STATE: STATE =
+    [1779033703, 3144134277, 1013904242, 2773480762, 1359893119, 2600822924, 528734635, 1541459225];

src/sha256/tests.nr

@@ -1,7 +1,8 @@
 use super::{
-    attach_len_to_msg_block, build_msg_block, byte_into_item, get_item_byte, INT_BLOCK, make_item,
+    attach_len_to_msg_block, build_msg_block, byte_into_item, get_item_byte, make_item,
     set_item_byte_then_zeros, set_item_zeros, sha256, sha256_var,
 };
+use super::constants::INT_BLOCK;
 
 #[export]
 fn test_sha256_1(input: [u8; 1], len: u64) -> [u8; 32] {