One simple SIMD optimization (#402)

* Adding some SIMD.

* SSE2 version.

* Reformat.

---------

Co-authored-by: Daniel Lemire <dlemire@lemire.me>

Author: lemire

include/ada/common_defs.h

@@ -285,4 +285,15 @@ namespace ada {
     if (!(COND)) __builtin_unreachable(); \
   } while (0)
 #endif
+
+#if defined(__SSE2__) || defined(__x86_64__) || defined(__x86_64) || \
+    (defined(_M_AMD64) || defined(_M_X64) ||                         \
+     (defined(_M_IX86_FP) && _M_IX86_FP == 2))
+#define ADA_SSE2 1
+#endif
+
+#if defined(__aarch64__) || defined(_M_ARM64)
+#define ADA_NEON 1
+#endif
+
 #endif  // ADA_COMMON_DEFS_H

include/ada/unicode.h

@@ -71,7 +71,7 @@ std::string to_unicode(std::string_view input);
  * @attention The has_tabs_or_newline function is a bottleneck and it is simple
  * enough that compilers like GCC can 'autovectorize it'.
  */
-ada_really_inline constexpr bool has_tabs_or_newline(
+ada_really_inline bool has_tabs_or_newline(
     std::string_view user_input) noexcept;
 
 /**

src/unicode.cpp

@@ -8,6 +8,11 @@ ADA_PUSH_DISABLE_ALL_WARNINGS
 ADA_POP_DISABLE_WARNINGS
 
 #include <algorithm>
+#if ADA_NEON
+#include <arm_neon.h>
+#elif ADA_SSE2
+#include <emmintrin.h>
+#endif
 
 namespace ada::unicode {
 
@@ -39,8 +44,58 @@ constexpr bool to_lower_ascii(char* input, size_t length) noexcept {
   }
   return non_ascii == 0;
 }
-
-ada_really_inline constexpr bool has_tabs_or_newline(
+#if ADA_NEON
+ada_really_inline bool has_tabs_or_newline(
+    std::string_view user_input) noexcept {
+  size_t i = 0;
+  const uint8x16_t mask1 = vmovq_n_u8('\r');
+  const uint8x16_t mask2 = vmovq_n_u8('\n');
+  const uint8x16_t mask3 = vmovq_n_u8('\t');
+  uint8x16_t running{0};
+  for (; i + 15 < user_input.size(); i += 16) {
+    uint8x16_t word = vld1q_u8((const uint8_t*)user_input.data() + i);
+    running = vorrq_u8(vorrq_u8(running, vorrq_u8(vceqq_u8(word, mask1),
+                                                  vceqq_u8(word, mask2))),
+                       vceqq_u8(word, mask3));
+  }
+  if (i < user_input.size()) {
+    uint8_t buffer[16];
+    memcpy(buffer, user_input.data() + i, user_input.size() - i);
+    uint8x16_t word = vld1q_u8((const uint8_t*)user_input.data() + i);
+    running = vorrq_u8(vorrq_u8(running, vorrq_u8(vceqq_u8(word, mask1),
+                                                  vceqq_u8(word, mask2))),
+                       vceqq_u8(word, mask3));
+  }
+  return vmaxvq_u8(running) != 0;
+}
+#elif ADA_SSE2
+ada_really_inline bool has_tabs_or_newline(
+    std::string_view user_input) noexcept {
+  size_t i = 0;
+  const __m128i mask1 = _mm_set1_epi8('\r');
+  const __m128i mask2 = _mm_set1_epi8('\n');
+  const __m128i mask3 = _mm_set1_epi8('\t');
+  __m128i running{0};
+  for (; i + 15 < user_input.size(); i += 16) {
+    __m128i word = _mm_loadu_si128((const __m128i*)(user_input.data() + i));
+    running = _mm_or_si128(
+        _mm_or_si128(running, _mm_or_si128(_mm_cmpeq_epi8(word, mask1),
+                                           _mm_cmpeq_epi8(word, mask2))),
+        _mm_cmpeq_epi8(word, mask3));
+  }
+  if (i < user_input.size()) {
+    uint8_t buffer[16];
+    memcpy(buffer, user_input.data() + i, user_input.size() - i);
+    __m128i word = _mm_loadu_si128((const __m128i*)buffer);
+    running = _mm_or_si128(
+        _mm_or_si128(running, _mm_or_si128(_mm_cmpeq_epi8(word, mask1),
+                                           _mm_cmpeq_epi8(word, mask2))),
+        _mm_cmpeq_epi8(word, mask3));
+  }
+  return _mm_movemask_epi8(running) != 0;
+}
+#else
+ada_really_inline bool has_tabs_or_newline(
     std::string_view user_input) noexcept {
   auto has_zero_byte = [](uint64_t v) {
     return ((v - 0x0101010101010101) & ~(v)&0x8080808080808080);
@@ -71,6 +126,7 @@ ada_really_inline constexpr bool has_tabs_or_newline(
   }
   return running;
 }
+#endif
 
 // A forbidden host code point is U+0000 NULL, U+0009 TAB, U+000A LF, U+000D CR,
 // U+0020 SPACE, U+0023 (#), U+002F (/), U+003A (:), U+003C (<), U+003E (>),