Start bench at a stride of 128

CMakeLists.txt

@@ -84,9 +84,6 @@ check_cxx_source_compiles("int main(){}" HAS_VERSION_SCRIPT FAIL_REGEX
                           "warning:")
 cmake_pop_check_state()
 
-if(CMAKE_SYSTEM_PROCESSOR STREQUAL x86_64)
-  add_compile_options(-mavx)
-endif()
 if(CMAKE_SYSTEM_NAME STREQUAL Linux)
   add_link_options(LINKER:--gc-sections)
 endif()

src/parser3.h

@@ -396,27 +396,26 @@ inline PRESERVE_NONE WeaselJsonStatus n_string(Parser3 *self, char *buf,
   MUSTTAIL return Parser3::keepGoing(self, buf, bufEnd);
 }
 
-inline PRESERVE_NONE WeaselJsonStatus n_string2(Parser3 *self, char *buf,
-                                                char *bufEnd) {
+template <class V>
+PRESERVE_NONE WeaselJsonStatus n_string2_impl(Parser3 *self, char *buf,
+                                              char *bufEnd) {
   const auto before = buf;
 
   // Advance buf to the first "non-normal" character
   for (;;) {
-    constexpr int kStride = 64;
-    if (bufEnd - buf < kStride) [[unlikely]] {
+    if (bufEnd - buf < V::lanes) [[unlikely]] {
       while (buf != bufEnd &&
              tables.stringByteMeaning[uint8_t(*buf)] == Tables::NORMAL) {
         ++buf;
       }
       break;
     }
-    using V = simd<int8_t, kStride>;
     auto v = V{(int8_t *)buf};
     int normal =
         (v != V::splat('"') & v != V::splat('\\') & v >= V::splat(0x20))
             .count_leading_nonzero_lanes();
     buf += normal;
-    if (normal < kStride) {
+    if (normal < V::lanes) {
       break;
     }
   }
@@ -485,6 +484,24 @@ inline PRESERVE_NONE WeaselJsonStatus n_string2(Parser3 *self, char *buf,
   }
 }
 
+template WeaselJsonStatus
+n_string2_impl<simd<int8_t, 64, sse::Simd_x86_SSE>>(Parser3 *, char *, char *);
+
+template __attribute__((target("avx2"))) WeaselJsonStatus
+n_string2_impl<simd<int8_t, 64, sse::Simd_x86_AVX2>>(Parser3 *, char *, char *);
+
+__attribute__((target("default"))) inline PRESERVE_NONE WeaselJsonStatus
+n_string2(Parser3 *self, char *buf, char *bufEnd) {
+  MUSTTAIL return n_string2_impl<simd<int8_t, 64, sse::Simd_x86_SSE>>(self, buf,
+                                                                      bufEnd);
+}
+
+__attribute__((target("avx2"))) inline PRESERVE_NONE WeaselJsonStatus
+n_string2(Parser3 *self, char *buf, char *bufEnd) {
+  MUSTTAIL return n_string2_impl<simd<int8_t, 64, sse::Simd_x86_AVX2>>(
+      self, buf, bufEnd);
+}
+
 inline PRESERVE_NONE WeaselJsonStatus n_string_following_escape(Parser3 *self,
                                                                 char *buf,
                                                                 char *bufEnd) {

src/test.cpp

@@ -262,7 +262,7 @@ TEST_CASE("bench3") {
   bench.batch(json.size());
   bench.unit("byte");
   auto *parser = WeaselJsonParser_create(1024, &c, nullptr);
-  for (size_t stride = 1; stride <= json.size(); stride *= 2) {
+  for (size_t stride = 128; stride <= json.size(); stride *= 2) {
     bench.run("parser3 (stride: " + std::to_string(stride) + ")", [&]() {
       auto copy = json;
       WeaselJsonParser_reset(parser);