perf(libsignal): cache the verify-side Edwards derivations per sender key

wacore/libsignal/src/core/curve.rs

@@ -237,6 +237,82 @@ use curve25519_dalek::edwards::CompressedEdwardsY;
 use curve25519_dalek::scalar::Scalar;
 use std::sync::OnceLock;
 
+use curve25519_dalek::edwards::EdwardsPoint;
+use curve25519_dalek::montgomery::MontgomeryPoint;
+
+/// A verifying view of a [`PublicKey`] that caches the per-key XEdDSA
+/// derivations: the Montgomery-to-Edwards conversion and its compressed
+/// encoding, per sign bit (the bit travels in each signature, and a given
+/// signer always produces the same one). Repeat verifications of one key,
+/// e.g. every incoming message under a group sender key, skip two field
+/// inversions per signature.
+///
+/// Clones carry initialized entries, so a warmed instance stays warm through
+/// the per-use clones of whatever memoizes it.
+/// `(-A, A_compressed)` for one sign bit of a verifying key.
+type PreparedEdwards = (EdwardsPoint, [u8; 32]);
+
+#[derive(Clone)]
+pub struct PreparedVerifyingKey {
+    mont: [u8; 32],
+    /// Per sign bit; `None` when the key has no Edwards form for that bit
+    /// (such a signature can never verify).
+    cached: [OnceLock<Option<PreparedEdwards>>; 2],
+}
+
+impl PreparedVerifyingKey {
+    pub fn new(key: &PublicKey) -> Self {
+        let PublicKeyData::DjbPublicKey(mont) = key.key;
+        Self {
+            mont,
+            cached: [OnceLock::new(), OnceLock::new()],
+        }
+    }
+
+    fn entry(&self, sign: u8) -> Option<&(EdwardsPoint, [u8; 32])> {
+        self.cached[usize::from(sign & 1)]
+            .get_or_init(|| {
+                MontgomeryPoint(self.mont)
+                    .to_edwards(sign)
+                    .map(|point| (-point, point.compress().to_bytes()))
+            })
+            .as_ref()
+    }
+
+    pub fn verify_signature(&self, message: &[u8], signature: &[u8]) -> bool {
+        self.verify_signature_for_multipart_message(&[message], signature)
+    }
+
+    pub fn verify_signature_for_multipart_message(
+        &self,
+        message: &[&[u8]],
+        signature: &[u8],
+    ) -> bool {
+        let Ok(signature) = <&[u8; 64]>::try_from(signature) else {
+            return false;
+        };
+        let sign = (signature[63] & 0b1000_0000_u8) >> 7;
+        let Some((minus_cap_a, cap_a_bytes)) = self.entry(sign) else {
+            return false;
+        };
+        curve25519::verify_signature_prepared(minus_cap_a, cap_a_bytes, message, signature)
+    }
+}
+
+impl From<&PublicKey> for PreparedVerifyingKey {
+    fn from(key: &PublicKey) -> Self {
+        Self::new(key)
+    }
+}
+
+impl std::fmt::Debug for PreparedVerifyingKey {
+    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
+        f.debug_struct("PreparedVerifyingKey")
+            .field("mont", &hex::encode(self.mont))
+            .finish_non_exhaustive()
+    }
+}
+
 /// Cached data for XEdDSA signing.
 /// This avoids an expensive scalar multiplication on every signature
 /// and caches the scalar representation to avoid repeated modular reduction.
@@ -511,6 +587,72 @@ mod tests {
         rand::make_rng::<rand::rngs::StdRng>()
     }
 
+    /// The cached verifier must agree with the plain path on every input:
+    /// valid signatures under both sign bits, corrupted signatures and
+    /// messages, garbage keys, and wrong-length signatures.
+    #[test]
+    fn prepared_verifier_matches_plain_verify() {
+        let mut csprng = rng();
+        let message: &[u8] = b"mensagem para verificar";
+
+        let mut seen_signs = [false; 2];
+        for _ in 0..32 {
+            let keypair = KeyPair::generate(&mut csprng);
+            let prepared = PreparedVerifyingKey::new(&keypair.public_key);
+            let signature = keypair
+                .calculate_signature(message, &mut csprng)
+                .expect("sign");
+            seen_signs[usize::from(signature[63] >> 7)] = true;
+
+            assert!(keypair.public_key.verify_signature(message, &signature));
+            assert!(prepared.verify_signature(message, &signature));
+
+            // Corrupted signature and corrupted message reject on both paths.
+            let mut bad_sig = signature;
+            bad_sig[7] ^= 0x40;
+            assert!(!keypair.public_key.verify_signature(message, &bad_sig));
+            assert!(!prepared.verify_signature(message, &bad_sig));
+            assert!(!keypair.public_key.verify_signature(b"outra", &signature));
+            assert!(!prepared.verify_signature(b"outra", &signature));
+
+            // Flipped sign bit must agree between paths too.
+            let mut flipped = signature;
+            flipped[63] ^= 0x80;
+            assert_eq!(
+                keypair.public_key.verify_signature(message, &flipped),
+                prepared.verify_signature(message, &flipped)
+            );
+
+            // Wrong-length signatures reject on both paths.
+            assert!(
+                !keypair
+                    .public_key
+                    .verify_signature(message, &signature[..63])
+            );
+            assert!(!prepared.verify_signature(message, &signature[..63]));
+
+            // A warmed CLONE must verify identically (the memoized instance
+            // hands out its state by reference, but clones must stay correct).
+            assert!(prepared.clone().verify_signature(message, &signature));
+        }
+        assert!(
+            seen_signs[0] && seen_signs[1],
+            "corpus must exercise both signature sign bits, got {seen_signs:?}"
+        );
+
+        // Garbage key bytes: both paths must reject the same way without panicking.
+        let mut garbage = [0u8; 33];
+        garbage[0] = 0x05;
+        garbage[1..].fill(0xFF);
+        let bad_key = PublicKey::deserialize(&garbage).expect("type-tagged bytes parse");
+        let prepared_bad = PreparedVerifyingKey::new(&bad_key);
+        let some_sig = [0x11u8; 64];
+        assert_eq!(
+            bad_key.verify_signature(message, &some_sig),
+            prepared_bad.verify_signature(message, &some_sig)
+        );
+    }
+
     #[test]
     fn test_signature_with_lazy_cache() {
         let mut csprng = rng();

wacore/libsignal/src/core/curve/curve25519.rs

@@ -46,6 +46,45 @@ pub struct PrivateKey {
     sign_bit: u8,
 }
 
+/// The XEdDSA verify equation with the per-key derivations precomputed:
+/// `minus_cap_a` is the negated Edwards form of the signer's public key for
+/// the signature's sign bit, and `cap_a_bytes` its compressed encoding. The
+/// single source of truth for verification; `PrivateKey::verify_signature`
+/// derives the inputs per call, while cached verifiers reuse them.
+pub(crate) fn verify_signature_prepared(
+    minus_cap_a: &EdwardsPoint,
+    cap_a_bytes: &[u8; 32],
+    message: &[&[u8]],
+    signature: &[u8; SIGNATURE_LENGTH],
+) -> bool {
+    let mut cap_r = [0u8; 32];
+    cap_r.copy_from_slice(&signature[..32]);
+    let mut s = [0u8; 32];
+    s.copy_from_slice(&signature[32..]);
+    s[31] &= 0b0111_1111_u8;
+    if (s[31] & 0b1110_0000_u8) != 0 {
+        return false;
+    }
+
+    let mut hash = Sha512::new();
+    // Explicitly pass a slice to avoid generating multiple versions of update().
+    hash.update(&cap_r[..]);
+    hash.update(cap_a_bytes);
+    for message_piece in message {
+        hash.update(message_piece);
+    }
+    let h = Scalar::from_bytes_mod_order_wide(&hash.finalize().into());
+
+    let cap_r_check_point = EdwardsPoint::vartime_double_scalar_mul_basepoint(
+        &h,
+        minus_cap_a,
+        &Scalar::from_bytes_mod_order(s),
+    );
+    let cap_r_check = cap_r_check_point.compress();
+
+    bool::from(cap_r_check.as_bytes().ct_eq(&cap_r))
+}
+
 impl PrivateKey {
     /// Computes the cached scalar, Edwards public key, and sign bit from a StaticSecret.
     #[inline]
@@ -256,33 +295,7 @@ impl PrivateKey {
                 None => return false,
             };
         let cap_a = ed_pub_key_point.compress();
-        let mut cap_r = [0u8; 32];
-        cap_r.copy_from_slice(&signature[..32]);
-        let mut s = [0u8; 32];
-        s.copy_from_slice(&signature[32..]);
-        s[31] &= 0b0111_1111_u8;
-        if (s[31] & 0b1110_0000_u8) != 0 {
-            return false;
-        }
-        let minus_cap_a = -ed_pub_key_point;
-
-        let mut hash = Sha512::new();
-        // Explicitly pass a slice to avoid generating multiple versions of update().
-        hash.update(&cap_r[..]);
-        hash.update(cap_a.as_bytes());
-        for message_piece in message {
-            hash.update(message_piece);
-        }
-        let h = Scalar::from_bytes_mod_order_wide(&hash.finalize().into());
-
-        let cap_r_check_point = EdwardsPoint::vartime_double_scalar_mul_basepoint(
-            &h,
-            &minus_cap_a,
-            &Scalar::from_bytes_mod_order(s),
-        );
-        let cap_r_check = cap_r_check_point.compress();
-
-        bool::from(cap_r_check.as_bytes().ct_eq(&cap_r))
+        verify_signature_prepared(&-ed_pub_key_point, cap_a.as_bytes(), message, signature)
     }
 
     pub fn derive_public_key_bytes(&self) -> [u8; PUBLIC_KEY_LENGTH] {

wacore/libsignal/src/protocol/group_cipher.rs

@@ -210,9 +210,9 @@ pub async fn group_decrypt(
     }
 
     let signing_key = sender_key_state
-        .signing_key_public()
+        .signing_key_verifier()
         .map_err(|_| SignalProtocolError::InvalidSenderKeySession)?;
-    if !skm.verify_signature(&signing_key)? {
+    if !skm.verify_signature_prepared(signing_key)? {
         return Err(SignalProtocolError::SignatureValidationFailed);
     }
 

wacore/libsignal/src/protocol/mod.rs

@@ -32,7 +32,7 @@ mod state;
 mod storage;
 mod stores;
 mod timestamp;
-pub use crate::core::curve::{CurveError, KeyPair, PrivateKey, PublicKey};
+pub use crate::core::curve::{CurveError, KeyPair, PreparedVerifyingKey, PrivateKey, PublicKey};
 pub use crate::core::{
     Aci, DeviceId, Pni, ProtocolAddress, ServiceId, ServiceIdFixedWidthBinaryBytes, ServiceIdKind,
 };

wacore/libsignal/src/protocol/protocol.rs

@@ -504,6 +504,21 @@ impl SenderKeyMessage {
         Ok(valid)
     }
 
+    /// Like [`Self::verify_signature`], against a cached verifier: the
+    /// per-key Edwards derivations are reused across messages instead of
+    /// recomputed per signature.
+    pub fn verify_signature_prepared(
+        &self,
+        signature_key: &crate::core::curve::PreparedVerifyingKey,
+    ) -> Result<bool> {
+        let valid = signature_key.verify_signature(
+            &self.serialized[..self.serialized.len() - Self::SIGNATURE_LEN],
+            &self.serialized[self.serialized.len() - Self::SIGNATURE_LEN..],
+        );
+
+        Ok(valid)
+    }
+
     #[inline]
     pub fn message_version(&self) -> u8 {
         self.message_version

wacore/libsignal/src/protocol/sender_keys.rs

@@ -177,6 +177,10 @@ pub struct SenderKeyState {
     /// rebuilt lazily after a cold load. If a signing-key setter is ever
     /// added, it must reset this memo.
     signing_key_memo: std::sync::OnceLock<PrivateKey>,
+    /// Receive-side mirror of `signing_key_memo`: cached verifier whose
+    /// Edwards derivations are reused across every incoming message under
+    /// this sender key. Same lifecycle rules as above.
+    verifying_key_memo: std::sync::OnceLock<crate::core::curve::PreparedVerifyingKey>,
 }
 
 // Manual impl with the signing key REDACTED: the protobuf state embeds the
@@ -224,16 +228,22 @@ impl SenderKeyState {
             key.precompute_signing_cache();
             let _ = signing_key_memo.set(key);
         }
+        let verifying_key_memo = std::sync::OnceLock::new();
+        let _ = verifying_key_memo.set(crate::core::curve::PreparedVerifyingKey::new(
+            &signature_key,
+        ));
         Self {
             state,
             signing_key_memo,
+            verifying_key_memo,
         }
     }
 
     pub(crate) fn from_protobuf(state: SenderKeyStateStructure) -> Self {
         Self {
             state,
             signing_key_memo: std::sync::OnceLock::new(),
+            verifying_key_memo: std::sync::OnceLock::new(),
         }
     }
 
@@ -276,6 +286,23 @@ impl SenderKeyState {
         }
     }
 
+    /// Cached verifier for this sender's signing key; the Edwards
+    /// derivations warm on first use and persist with the in-memory state.
+    pub fn signing_key_verifier(
+        &self,
+    ) -> Result<&crate::core::curve::PreparedVerifyingKey, InvalidSenderKeySessionError> {
+        if let Some(verifier) = self.verifying_key_memo.get() {
+            return Ok(verifier);
+        }
+        let verifier = crate::core::curve::PreparedVerifyingKey::new(&self.signing_key_public()?);
+        // Benign race: concurrent firsts compute the same value.
+        let _ = self.verifying_key_memo.set(verifier);
+        Ok(self
+            .verifying_key_memo
+            .get()
+            .expect("set on the line above"))
+    }
+
     pub fn signing_key_private(&self) -> Result<PrivateKey, InvalidSenderKeySessionError> {
         if let Some(key) = self.signing_key_memo.get() {
             return Ok(key.clone());
@@ -639,6 +666,15 @@ mod tests {
                 .has_warm_signing_cache()
         );
 
+        // Verifier memo: seeded at creation, rebuilt lazily after a cold
+        // load, and clones carry it.
+        assert!(state.verifying_key_memo.get().is_some());
+        let cold = SenderKeyState::from_protobuf(state.as_protobuf());
+        assert!(cold.verifying_key_memo.get().is_none());
+        let _ = cold.signing_key_verifier().expect("verifier");
+        assert!(cold.verifying_key_memo.get().is_some());
+        assert!(cold.clone().verifying_key_memo.get().is_some());
+
         // The memoized key still signs correctly.
         let msg = b"skmsg";
         let sig = key.calculate_signature(msg, &mut rng).expect("sign");