perf: use SIMD for ASCII scan, lowercase, and UTF transcoding

.github/workflows/alpine.yml

@@ -8,6 +8,9 @@ jobs:
     name: Build on Alpine ${{ matrix.arch }}
     runs-on: ubuntu-latest
     strategy:
+      # setup-alpine + qemu can flake on one arch (apk DB under emulation);
+      # do not cancel the others.
+      fail-fast: false
       matrix:
         arch:
           - x86_64
@@ -22,11 +25,15 @@ jobs:
       - name: Install extra dependencies
         run: |
           sudo apt update && sudo apt install -y binfmt-support
-      - name: Install latest Alpine Linux for ${{ matrix.arch }}
+      - name: Install Alpine Linux for ${{ matrix.arch }}
         uses: jirutka/setup-alpine@v1
         with:
           arch: ${{ matrix.arch }}
-          branch: ${{ matrix.arch == 'riscv64' && 'edge' || 'latest-stable' }}
+          # latest-stable is 3.23+ (apk-tools v3). setup-alpine still
+          # bootstraps with a static apk v2.14 binary, so apk add in the
+          # chroot fails with "database: file format is invalid" on some
+          # qemu arches (seen on ppc64le). v3.22 is the last apk v2 release.
+          branch: ${{ matrix.arch == 'riscv64' && 'edge' || 'v3.22' }}
           packages: >
             build-base
             cmake

include/ada/idna/to_ascii.h

@@ -28,8 +28,10 @@ std::string to_ascii(std::string_view ut8_string);
 // https://url.spec.whatwg.org/#forbidden-domain-code-point
 bool contains_forbidden_domain_code_point(std::string_view ascii_string);
 
-bool constexpr is_ascii(std::u32string_view view);
-bool constexpr is_ascii(std::string_view view);
+// Runtime SIMD (SSE2/NEON) with a SWAR fallback. Not constexpr: vector
+// paths cannot be evaluated at compile time.
+bool is_ascii(std::u32string_view view) noexcept;
+bool is_ascii(std::string_view view) noexcept;
 
 }  // namespace ada::idna
 

src/mapping.cpp

@@ -6,6 +6,7 @@
 #include <string>
 
 #include "table_store.hpp"
+#include "simd.hpp"
 #include "mapping_tables.cpp"
 
 namespace ada::idna {
@@ -81,28 +82,7 @@ static size_t utf8_count_codepoints(const uint8_t* ptr) noexcept {
 
 // --- ASCII fast path ---------------------------------------------------------
 void ascii_map(char* input, size_t length) {
-  auto broadcast = [](uint8_t v) -> uint64_t {
-    return 0x101010101010101ull * v;
-  };
-  uint64_t broadcast_80 = broadcast(0x80);
-  uint64_t broadcast_Ap = broadcast(128 - 'A');
-  uint64_t broadcast_Zp = broadcast(128 - 'Z' - 1);
-  size_t i = 0;
-
-  for (; i + 7 < length; i += 8) {
-    uint64_t word{};
-    std::memcpy(&word, input + i, sizeof(word));
-    word ^=
-        (((word + broadcast_Ap) ^ (word + broadcast_Zp)) & broadcast_80) >> 2;
-    std::memcpy(input + i, &word, sizeof(word));
-  }
-  if (i < length) {
-    uint64_t word{};
-    std::memcpy(&word, input + i, length - i);
-    word ^=
-        (((word + broadcast_Ap) ^ (word + broadcast_Zp)) & broadcast_80) >> 2;
-    std::memcpy(input + i, &word, length - i);
-  }
+  (void)simd::ascii_lowercase_is_ascii(input, length);
 }
 
 // Two-pass map: first validate + exact size, then write once (no growth