From 28b8908347c0b49938d90c27970088fbe81c5ec3 Mon Sep 17 00:00:00 2001
From: Nic Barker <contact+github@nicbarker.com>
Date: Wed, 19 Mar 2025 11:30:00 +1300
Subject: [PATCH] Interim commit

---
 clay.h | 134 +++++++++++++++++++++++++++++++++++++++++++++++++++------
 1 file changed, 122 insertions(+), 12 deletions(-)

diff --git a/clay.h b/clay.h
index 2f69878..85b3770 100644
--- a/clay.h
+++ b/clay.h
@@ -1348,22 +1348,132 @@ Clay_ElementId Clay__HashString(Clay_String key, const uint32_t offset, const ui
     return CLAY__INIT(Clay_ElementId) { .id = hash + 1, .offset = offset, .baseId = base + 1, .stringId = key }; // Reserve the hash result of zero as "null id"
 }
 
+#if !defined(CLAY_DISABLE_SIMD) && (defined(__x86_64__) || defined(_M_X64) || defined(_M_AMD64))
+// Rotate left in AVX2 (equivalent to _mm256_rol_epi64 in AVX512)
+static inline __m256i rol64(__m256i x, int r) {
+    return _mm256_or_si256(_mm256_slli_epi64(x, r), _mm256_srli_epi64(x, 64 - r));
+}
+
+// A simple ARX mix function
+static inline void arx_mix(__m256i *a, __m256i *b) {
+    *a = _mm256_add_epi64(*a, *b);
+    *b = _mm256_xor_si256(rol64(*b, 17), *a);
+}
+
+// SIMD ARX hash function (AVX2)
+uint64_t arx_simd_hash(const uint8_t *data, size_t len) {
+    // Pinched these constants from the BLAKE implementation
+    __m256i v0 = _mm256_set1_epi64x(0x6a09e667f3bcc908ULL);
+    __m256i v1 = _mm256_set1_epi64x(0xbb67ae8584caa73bULL);
+    __m256i v2 = _mm256_set1_epi64x(0x3c6ef372fe94f82bULL);
+    __m256i v3 = _mm256_set1_epi64x(0xa54ff53a5f1d36f1ULL);
+
+    uint8_t overflowBuffer[16] = {0};  // Temporary buffer for small inputs
+
+    // Process 32-byte chunks
+    while (len > 0) {
+        __m256i msg;
+        if (len >= 32) {
+            msg = _mm256_loadu_si256((const __m256i *)data);
+            data += 32;
+            len -= 32;
+        } else {
+            memset(overflowBuffer, 0, 16);
+            memcpy(overflowBuffer, data, len);
+            msg = _mm256_loadu_si256((const __m256i *)overflowBuffer);
+            len = 0;
+        }
+
+        v0 = _mm256_xor_si256(v0, msg);
+        arx_mix(&v0, &v1);
+        arx_mix(&v2, &v3);
+
+        // Cross-lane mixing
+        v0 = _mm256_add_epi64(v0, v2);
+        v1 = _mm256_add_epi64(v1, v3);
+    }
+
+    // Final mixing rounds
+    arx_mix(&v0, &v1);
+    arx_mix(&v2, &v3);
+    v0 = _mm256_add_epi64(v0, v2);
+    v1 = _mm256_add_epi64(v1, v3);
+
+    // Extract final hash
+    uint64_t result[4];
+    _mm256_storeu_si256((__m256i *)result, v0);
+
+    return result[0] ^ result[1] ^ result[2] ^ result[3];
+}
+#elif !defined(CLAY_DISABLE_SIMD) && defined(__aarch64__)
+    // Rotate left in NEON (simulating _mm256_rol_epi64)
+    static inline uint64x2_t rol64(uint64x2_t x, int r) {
+        return vorrq_u64(vshlq_n_u64(x, 17), vshrq_n_u64(x, 64 - 17));
+    }
+
+    // A simple ARX mix function
+    static inline void arx_mix(uint64x2_t *a, uint64x2_t *b) {
+        *a = vaddq_u64(*a, *b);
+        *b = veorq_u64(rol64(*b, 17), *a);
+    }
+
+    // SIMD ARX hash function (NEON)
+    uint64_t arx_simd_hash(const uint8_t *data, size_t len) {
+        // Pinched these constants from the BLAKE implementation
+        uint64x2_t v0 = vdupq_n_u64(0x6a09e667f3bcc908ULL);
+        uint64x2_t v1 = vdupq_n_u64(0xbb67ae8584caa73bULL);
+        uint64x2_t v2 = vdupq_n_u64(0x3c6ef372fe94f82bULL);
+        uint64x2_t v3 = vdupq_n_u64(0xa54ff53a5f1d36f1ULL);
+
+        uint8_t overflowBuffer[8] = {0};
+
+        // Process 16-byte chunks
+        while (len > 0) {
+            uint64x2_t msg;
+            if (len > 16) {
+                msg = vld1q_u64((const uint64_t *)data);
+                data += 16;
+                len -= 16;
+            } else if (len > 8) {
+                msg = vcombine_u64(vld1_u64((const uint64_t *)data), vdup_n_u64(0));
+                data += 8;
+                len -= 8;
+            } else {
+                for (int i = 0; i < len; i++) {
+                    overflowBuffer[i] = data[i];
+                }
+                uint8x8_t lower = vld1_u8(overflowBuffer); // Load up to 8 bytes
+                msg = vcombine_u8(lower, vdup_n_u8(0)); // Zero upper 8 bytes
+                len = 0;
+            }
+            v0 = veorq_u64(v0, msg);
+            arx_mix(&v0, &v1);
+            arx_mix(&v2, &v3);
+
+            // Cross-lane mixing
+            v0 = vaddq_u64(v0, v2);
+            v1 = vaddq_u64(v1, v3);
+        }
+
+        // Final mixing rounds
+        arx_mix(&v0, &v1);
+        arx_mix(&v2, &v3);
+        v0 = vaddq_u64(v0, v2);
+        v1 = vaddq_u64(v1, v3);
+
+        // Extract final hash
+        uint64_t result[2];
+        vst1q_u64(result, v0);
+
+        return result[0] ^ result[1];
+    }
+#endif
+
 uint32_t Clay__HashTextWithConfig(Clay_String *text, Clay_TextElementConfig *config) {
     uint32_t hash = 0;
     uintptr_t pointerAsNumber = (uintptr_t)text->chars;
 
-    if (config->hashStringContents) {
-        uint32_t maxLengthToHash = CLAY__MIN(text->length, 256);
-        for (uint32_t i = 0; i < maxLengthToHash; i++) {
-            hash += text->chars[i];
-            hash += (hash << 10);
-            hash ^= (hash >> 6);
-        }
-    } else {
-        hash += pointerAsNumber;
-        hash += (hash << 10);
-        hash ^= (hash >> 6);
-    }
+    hash = arx_simd_hash((const uint8_t *)text->chars, text->length) % UINT32_MAX;
 
     hash += text->length;
     hash += (hash << 10);