From 9cceb5a5be55e7c62162dfd234f6134bf91fda7d Mon Sep 17 00:00:00 2001
From: Mathias Berchtold <mberchtold@gmail.com>
Date: Thu, 3 Sep 2026 16:44:59 +0200
Subject: [PATCH 1/2] Implement SIMD optimizations for tag comparison

Add SIMD support for poly1305 tag comparison
---
 src/chacha20poly1305.c | 17 +++++++++++++++++
 1 file changed, 17 insertions(+)

diff --git a/src/chacha20poly1305.c b/src/chacha20poly1305.c
index 86157dc0..707ae304 100644
--- a/src/chacha20poly1305.c
+++ b/src/chacha20poly1305.c
@@ -10,6 +10,12 @@
 #include "mystdint.h"
 #include <string.h>
 
+#if defined(__wasm_simd128__)
+#include <wasm_simd128.h>
+#elif defined(__SSE4_1__)
+#include <immintrin.h>
+#endif
+
 #define ROL32(x, c) (((x) << (c)) | ((x) >> (32-(c))))
 #define ROR32(x, c) (((x) >> (c)) | ((x) << (32-(c))))
 
@@ -215,6 +221,16 @@ void poly1305(const uint8_t* msg, size_t n, const uint8_t key[32],
 SQLITE_PRIVATE
 int poly1305_tagcmp(const uint8_t tag1[16], const uint8_t tag2[16])
 {
+#if defined(__wasm_simd128__)
+  v128_t tagDifference = wasm_v128_xor(
+    wasm_v128_load(tag1), wasm_v128_load(tag2));
+  return wasm_v128_any_true(tagDifference);
+#elif defined(__SSE4_1__)
+  __m128i tag1_vec = _mm_loadu_si128((const __m128i*)tag1);
+  __m128i tag2_vec = _mm_loadu_si128((const __m128i*)tag2);
+  __m128i tagDifference = _mm_xor_si128(tag1_vec, tag2_vec);
+  return !_mm_testz_si128(tagDifference, tagDifference);
+#else
   uint8_t d = 0;
   d |= tag1[ 0] ^ tag2[ 0];
   d |= tag1[ 1] ^ tag2[ 1];
@@ -233,6 +249,7 @@ int poly1305_tagcmp(const uint8_t tag1[16], const uint8_t tag2[16])
   d |= tag1[14] ^ tag2[14];
   d |= tag1[15] ^ tag2[15];
   return (int) d;
+#endif
 }
 
 /*

From 64d9a8a0cb703eac9575b5e4f5d96277197a0cde Mon Sep 17 00:00:00 2001
From: Mathias Berchtold <mberchtold@gmail.com>
Date: Thu, 3 Sep 2026 17:05:49 +0200
Subject: [PATCH 2/2] added AArch64 (ARMv8) implementation

---
 src/chacha20poly1305.c | 7 +++++++
 1 file changed, 7 insertions(+)

diff --git a/src/chacha20poly1305.c b/src/chacha20poly1305.c
index 707ae304..a48bcdc4 100644
--- a/src/chacha20poly1305.c
+++ b/src/chacha20poly1305.c
@@ -14,6 +14,8 @@
 #include <wasm_simd128.h>
 #elif defined(__SSE4_1__)
 #include <immintrin.h>
+#elif defined(__aarch64__)
+#include <arm_neon.h>
 #endif
 
 #define ROL32(x, c) (((x) << (c)) | ((x) >> (32-(c))))
@@ -230,6 +232,11 @@ int poly1305_tagcmp(const uint8_t tag1[16], const uint8_t tag2[16])
   __m128i tag2_vec = _mm_loadu_si128((const __m128i*)tag2);
   __m128i tagDifference = _mm_xor_si128(tag1_vec, tag2_vec);
   return !_mm_testz_si128(tagDifference, tagDifference);
+#elif defined(__aarch64__)
+  uint8x16_t tag1_vec = vld1q_u8(tag1);
+  uint8x16_t tag2_vec = vld1q_u8(tag2);
+  uint8x16_t tagDifference = veorq_u8(tag1_vec, tag2_vec);
+  return vmaxvq_u8(tagDifference) != 0;
 #else
   uint8_t d = 0;
   d |= tag1[ 0] ^ tag2[ 0];