From 88505ebf3d298aab3c7af15d33cf664f86ee7213 Mon Sep 17 00:00:00 2001
From: I <1091761+wx257osn2@users.noreply.github.com>
Date: Sun, 4 Jun 2023 09:35:04 +0900
Subject: [PATCH] implement has_tabs_or_newline in AVX2

---
 include/ada/common_defs.h |  4 ++++
 src/unicode.cpp           | 30 ++++++++++++++++++++++++++++++
 2 files changed, 34 insertions(+)

diff --git a/include/ada/common_defs.h b/include/ada/common_defs.h
index 8aacdb56d..c5f4f59b8 100644
--- a/include/ada/common_defs.h
+++ b/include/ada/common_defs.h
@@ -286,6 +286,10 @@ namespace ada {
   } while (0)
 #endif
 
+#if defined(__AVX2__)
+#define ADA_AVX2 1
+#endif
+
 #if defined(__SSE2__) || defined(__x86_64__) || defined(__x86_64) || \
     (defined(_M_AMD64) || defined(_M_X64) ||                         \
      (defined(_M_IX86_FP) && _M_IX86_FP == 2))
diff --git a/src/unicode.cpp b/src/unicode.cpp
index e4c4ea827..6c7293c13 100644
--- a/src/unicode.cpp
+++ b/src/unicode.cpp
@@ -10,6 +10,8 @@ ADA_POP_DISABLE_WARNINGS
 #include <algorithm>
 #if ADA_NEON
 #include <arm_neon.h>
+#elif ADA_AVX2
+#include <immintrin.h>
 #elif ADA_SSE2
 #include <emmintrin.h>
 #endif
@@ -68,6 +70,34 @@ ada_really_inline bool has_tabs_or_newline(
   }
   return vmaxvq_u8(running) != 0;
 }
+#elif ADA_AVX2
+ada_really_inline bool has_tabs_or_newline(
+    std::string_view user_input) noexcept {
+  size_t i = 0;
+  const __m256i mask1 = _mm256_set1_epi8('\r');
+  const __m256i mask2 = _mm256_set1_epi8('\n');
+  const __m256i mask3 = _mm256_set1_epi8('\t');
+  __m256i running{0};
+  for (; i + 31 < user_input.size(); i += 32) {
+    __m256i word = _mm256_loadu_si256((const __m256i*)(user_input.data() + i));
+    running = _mm256_or_si256(
+        _mm256_or_si256(running,
+                        _mm256_or_si256(_mm256_cmpeq_epi8(word, mask1),
+                                        _mm256_cmpeq_epi8(word, mask2))),
+        _mm256_cmpeq_epi8(word, mask3));
+  }
+  if (i < user_input.size()) {
+    alignas(32) uint8_t buffer[32]{};
+    memcpy(buffer, user_input.data() + i, user_input.size() - i);
+    __m256i word = _mm256_load_si256((const __m256i*)buffer);
+    running = _mm256_or_si256(
+        _mm256_or_si256(running,
+                        _mm256_or_si256(_mm256_cmpeq_epi8(word, mask1),
+                                        _mm256_cmpeq_epi8(word, mask2))),
+        _mm256_cmpeq_epi8(word, mask3));
+  }
+  return _mm256_movemask_epi8(running) != 0;
+}
 #elif ADA_SSE2
 ada_really_inline bool has_tabs_or_newline(
     std::string_view user_input) noexcept {