perf: SIMD percent_encode against the character-set bitmap

benchmarks/percent_encode.cpp

@@ -243,6 +243,47 @@ static void C0Control(benchmark::State& state) {
 }
 BENCHMARK(C0Control);
 
+std::string long_examples[] = {
+    "connect timeout=10 application name=myapp server=db host internal "
+    "database=production analytics read preference=secondary preferred "
+    "ssl=true retry writes=true w=majority max pool size=50",
+    "ref=web twc ao gbl adsinfo utm source=twc utm medium=cpc "
+    "utm campaign=brand awareness q4 2024 utm content=banner 300x250 "
+    "utm term=weather forecast today gclid=Cj0KCQiA3Y ABhCnARIsAK",
+};
+
+double long_examples_bytes = []() -> double {
+  size_t bytes{0};
+  for (std::string& url_string : long_examples) {
+    bytes += url_string.size();
+  }
+  return double(bytes);
+}();
+
+static void LongFragment(benchmark::State& state) {
+  for (auto _ : state) {
+    for (std::string& url_string : long_examples) {
+      benchmark::DoNotOptimize(ada::unicode::percent_encode(
+          url_string, ada::character_sets::FRAGMENT_PERCENT_ENCODE));
+    }
+  }
+  state.counters["speed"] = benchmark::Counter(
+      long_examples_bytes, benchmark::Counter::kIsIterationInvariantRate);
+}
+BENCHMARK(LongFragment);
+
+static void LongQuery(benchmark::State& state) {
+  for (auto _ : state) {
+    for (std::string& url_string : long_examples) {
+      benchmark::DoNotOptimize(ada::unicode::percent_encode(
+          url_string, ada::character_sets::QUERY_PERCENT_ENCODE));
+    }
+  }
+  state.counters["speed"] = benchmark::Counter(
+      long_examples_bytes, benchmark::Counter::kIsIterationInvariantRate);
+}
+BENCHMARK(LongQuery);
+
 int main(int argc, char** argv) {
 #if defined(ADA_RUST_VERSION)
   benchmark::AddCustomContext("rust version ", ADA_RUST_VERSION);

src/CMakeLists.txt

@@ -7,7 +7,7 @@ target_include_directories(ada-include-source INTERFACE $<BUILD_INTERFACE:${CMAK
 add_library(ada-source INTERFACE)
 target_sources(ada-source INTERFACE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}>/ada.cpp)
 target_link_libraries(ada-source INTERFACE ada-include-source)
-add_library(ada ada.cpp)
+add_library(ada ada.cpp unicode_percent_encode.cpp)
 target_compile_features(ada PUBLIC cxx_std_20)
 target_include_directories(ada PRIVATE $<BUILD_INTERFACE:${CMAKE_CURRENT_SOURCE_DIR}> )
 target_include_directories(ada PUBLIC "$<BUILD_INTERFACE:${PROJECT_SOURCE_DIR}/include>")
@@ -60,6 +60,10 @@ if(ADA_TESTING)
   target_compile_definitions(ada PRIVATE ADA_TESTING=1)
 endif()
 
+# Keep the SIMD percent-encode kernel in its own TU so the unity ada.cpp
+# inlining budget for setters matches main.
+target_compile_definitions(ada PRIVATE ADA_PERCENT_ENCODE_SIMD_SEPARATE_TU=1)
+
 if(ADA_INCLUDE_URL_PATTERN)
   target_compile_definitions(ada PRIVATE ADA_INCLUDE_URL_PATTERN=1)
 else()

src/ada.cpp

@@ -1,6 +1,9 @@
 #include "ada.h"
 #include "checkers.cpp"
 #include "unicode.cpp"
+#if !defined(ADA_PERCENT_ENCODE_SIMD_SEPARATE_TU)
+#include "unicode_percent_encode.cpp"
+#endif
 #include "serializers.cpp"
 #include "implementation.cpp"
 #include "helpers.cpp"

src/unicode.cpp

@@ -590,6 +590,9 @@ std::string form_urlencoded_decode(const std::string_view input) {
   return out;
 }
 
+void percent_encode_suffix(const char* p, const char* end,
+                           const uint8_t character_set[], std::string& out);
+
 std::string percent_encode(const std::string_view input,
                            const uint8_t character_set[]) {
   auto pointer = std::ranges::find_if(input, [character_set](const char c) {
@@ -604,15 +607,18 @@ std::string percent_encode(const std::string_view input,
   result.reserve(input.length());  // in the worst case, percent encoding might
                                    // produce 3 characters.
   result.append(input.substr(0, std::distance(input.begin(), pointer)));
-
-  for (; pointer != input.end(); pointer++) {
-    if (character_sets::bit_at(character_set, *pointer)) {
-      result.append(character_sets::hex + uint8_t(*pointer) * 4, 3);
-    } else {
-      result += *pointer;
+  if (static_cast<size_t>(input.end() - pointer) >= 48) {
+    percent_encode_suffix(&*pointer, input.data() + input.size(), character_set,
+                          result);
+  } else {
+    for (; pointer != input.end(); pointer++) {
+      if (character_sets::bit_at(character_set, *pointer)) {
+        result.append(character_sets::hex + uint8_t(*pointer) * 4, 3);
+      } else {
+        result += *pointer;
+      }
     }
   }
-
   return result;
 }
 
@@ -651,6 +657,13 @@ bool percent_encode(const std::string_view input, const uint8_t character_set[],
   return true;
 }
 
+template bool percent_encode<true>(std::string_view input,
+                                   const uint8_t character_set[],
+                                   std::string& out);
+template bool percent_encode<false>(std::string_view input,
+                                    const uint8_t character_set[],
+                                    std::string& out);
+
 bool to_ascii(std::optional<std::string>& out, const std::string_view plain,
               size_t first_percent) {
   std::string percent_decoded_buffer;
@@ -675,11 +688,16 @@ std::string percent_encode(const std::string_view input,
   // NOLINTNEXTLINE(bugprone-suspicious-stringview-data-usage)
   out.append(input.data(), index);
   auto pointer = input.begin() + index;
-  for (; pointer != input.end(); pointer++) {
-    if (character_sets::bit_at(character_set, *pointer)) {
-      out.append(character_sets::hex + uint8_t(*pointer) * 4, 3);
-    } else {
-      out += *pointer;
+  if (static_cast<size_t>(input.end() - pointer) >= 48) {
+    percent_encode_suffix(&*pointer, input.data() + input.size(), character_set,
+                          out);
+  } else {
+    for (; pointer != input.end(); pointer++) {
+      if (character_sets::bit_at(character_set, *pointer)) {
+        out.append(character_sets::hex + uint8_t(*pointer) * 4, 3);
+      } else {
+        out += *pointer;
+      }
     }
   }
   return out;