fix: don't compile with -march=native by default

.github/workflows/main.yml

@@ -5,20 +5,23 @@ on: [push]
 jobs:
   build:
     strategy:
+      fail-fast: false
       matrix:
         os: [ubuntu-latest]
-        lua_version: [5.1, 5.2, 5.3, 5.4, luajit-2.0.5, luajit-2.1.0-beta3]
+        lua_version: [5.1, 5.2, 5.3, 5.4, luajit-2.0, luajit-2.1]
 
     runs-on: ${{ matrix.os }}
 
     steps:
-    - uses: actions/checkout@v1
+    - uses: actions/checkout@v4
+      with:
+        persist-credentials: false
 
-    - uses: leafo/gh-actions-lua@v8.0.0
+    - uses: leafo/gh-actions-lua@v13
       with:
         luaVersion: ${{ matrix.lua_version }}
 
-    - uses: leafo/gh-actions-luarocks@v4.0.0
+    - uses: leafo/gh-actions-luarocks@v6
 
     - name: Setup rocks depended
       run: |
@@ -29,5 +32,35 @@ jobs:
     - name: Build
       run: luarocks make
 
+    - name: Check the built module is portable
+      shell: bash
+      run: |
+        set -euo pipefail
+        # Anything above the x86-64 baseline (SSE2) makes the module die with
+        # SIGILL on older CPUs, and nothing else catches that until someone
+        # loads it there. The recorded compile flags are the exact check; the
+        # disassembly is the backstop for instructions that arrive by other
+        # means, such as intrinsics behind RAPIDJSON_SSE42.
+        flags=$(grep -h '^CXX_FLAGS' build.luarocks/CMakeFiles/*/flags.make)
+        if grep -qE -- '-m(sse[34]|ssse3|avx|fma|bmi|popcnt|aes|pclmul|f16c|abm|lzcnt|movbe|adx|sha)' <<<"$flags" ||
+           grep -oE -- '-march=[^[:space:]]+' <<<"$flags" | grep -qvE -- '-march=x86-64(-v1)?$'; then
+            echo "$flags"
+            echo "the module is built for a specific CPU, not the x86-64 baseline" >&2
+            exit 1
+        fi
+
+        module=build.luarocks/rapidjson.so
+        test -f "$module"
+        beyond=$(objdump -d --no-show-raw-insn "$module" \
+            | grep -oE '^[[:space:]]+[0-9a-f]+:[[:space:]]+[a-z][a-z0-9]*' \
+            | awk '{print $2}' | sort -u \
+            | grep -E '^(v[a-z][a-z0-9]*|k[a-z][a-z0-9]*|popcnt|lzcnt|tzcnt|crc32|movbe|adcx|adox|prefetchw|andn|bextr|bls[imr]|bzhi|mulx|pdep|pext|rorx|sarx|shlx|shrx|aes[a-z]*|pclmul[a-z]*|sha1[a-z]*|sha256[a-z]*|pcmp[ei]str[im]|pcmpgtq|ptest|pblendw|pblendvb|blendp[sd]|blendvp[sd]|dpp[sd]|pmov[sz]x[a-z]*|pmulld|pmuldq|packusdw|phminposuw|mpsadbw|movntdqa|round[sp][sd]|insertps|extractps|phadd[a-z]*|phsub[a-z]*|pmaddubsw|pmulhrsw|pabs[bwd]|palignr|pshufb|psign[bwd]|lddqu|movddup|movs[hl]dup|addsubp[sd]|haddp[sd]|hsubp[sd]|pm(in|ax)(s[bd]|u[dw])|rdrand|rdseed|rdpid|xbegin|xend|xabort|xtest|rd[fg]sbase|wr[fg]sbase|tile[a-z]*|ldtilecfg|sttilecfg|clwb|clflushopt|movdir[a-z0-9]*|serialize)$' || true)
+        if [ -n "$beyond" ]; then
+            printf '%s\n' "$beyond" | awk 'NR <= 6' | tr '\n' ' '
+            echo
+            echo "rapidjson.so uses instructions above the x86-64 baseline" >&2
+            exit 1
+        fi
+
     - name: Test
       run: busted

.github/workflows/release.yml

@@ -0,0 +1,65 @@
+name: Release
+
+on:
+  push:
+    branches:
+      - "master"
+    paths:
+      - 'rockspec/**'
+
+permissions:
+  contents: write
+
+jobs:
+  release:
+    name: Release
+    runs-on: ubuntu-latest
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v4
+        with:
+          persist-credentials: false
+
+      - name: Install Lua
+        uses: leafo/gh-actions-lua@v13
+
+      - name: Install Luarocks
+        uses: leafo/gh-actions-luarocks@v6
+
+      - name: Extract release name
+        id: release_env
+        shell: bash
+        env:
+          COMMIT_MESSAGE: ${{ github.event.head_commit.message }}
+        run: |
+          title="${COMMIT_MESSAGE}"
+          # tolerate the " (#N)" a squash merge appends to the title
+          re="^feat: release v([^[:space:]]+)"
+          if [[ $title =~ $re ]]; then
+              echo "version=v${BASH_REMATCH[1]}" >> "$GITHUB_OUTPUT"
+              echo "version_without_v=${BASH_REMATCH[1]}" >> "$GITHUB_OUTPUT"
+          else
+              echo "commit message must be 'feat: release v<version>'"
+              exit 1
+          fi
+
+      - name: Create release
+        env:
+          GH_TOKEN: ${{ secrets.GITHUB_TOKEN }}
+          VERSION: ${{ steps.release_env.outputs.version }}
+        run: |
+          # keep this retryable: a failed upload below is re-run with the
+          # release already created
+          if gh release view "${VERSION}" >/dev/null 2>&1; then
+              echo "release ${VERSION} already exists"
+          else
+              gh release create "${VERSION}" --title "${VERSION}" --generate-notes
+          fi
+
+      - name: Upload to luarocks
+        env:
+          LUAROCKS_TOKEN: ${{ secrets.LUAROCKS_TOKEN }}
+          VERSION: ${{ steps.release_env.outputs.version_without_v }}
+        run: |
+          luarocks install dkjson
+          luarocks upload "rockspec/api7-lua-rapidjson-${VERSION}-0.rockspec" --api-key="${LUAROCKS_TOKEN}"

CMakeLists.txt

@@ -32,10 +32,17 @@ if(UNIX)
         set(LINK_FLAGS "-shared")
     endif(APPLE)
     add_compile_options(-g -Wall -fPIC)
-    include(CheckCXXCompilerFlag)
-    CHECK_CXX_COMPILER_FLAG("-march=native" COMPILER_SUPPORTS_ARCH_NATIVE)
-    if (COMPILER_SUPPORTS_ARCH_NATIVE)
-        add_compile_options(-march=native)
+    # -march=native bakes the build host's CPU features into rapidjson.so, so a
+    # module built on a modern machine dies with SIGILL on an older one. Keep it
+    # opt-in: builds that get distributed (packages, images, luarocks) must stay
+    # portable.
+    option(LUA_RAPIDJSON_ARCH_NATIVE "Optimize for the build host CPU; produces a non-portable binary" OFF)
+    if (LUA_RAPIDJSON_ARCH_NATIVE)
+        include(CheckCXXCompilerFlag)
+        CHECK_CXX_COMPILER_FLAG("-march=native" COMPILER_SUPPORTS_ARCH_NATIVE)
+        if (COMPILER_SUPPORTS_ARCH_NATIVE)
+            add_compile_options(-march=native)
+        endif()
     endif()
 else(UNIX)
     if(WIN32)