diff --git a/.github/actions/manylinux/action.yaml b/.github/actions/manylinux/action.yaml
index 9dfae726..cd23f9c2 100644
--- a/.github/actions/manylinux/action.yaml
+++ b/.github/actions/manylinux/action.yaml
@@ -45,6 +45,7 @@ runs:
 
         uv pip install ${CARGO_TARGET_DIR}/wheels/orjson*.whl
 
+        export PYTHONMALLOC="debug"
         pytest -v test
         ./integration/run thread
         ./integration/run http
diff --git a/.github/workflows/artifact.yaml b/.github/workflows/artifact.yaml
index d48a4cc7..19225724 100644
--- a/.github/workflows/artifact.yaml
+++ b/.github/workflows/artifact.yaml
@@ -2,8 +2,9 @@ name: artifact
 on: push
 env:
   CARGO_UNSTABLE_SPARSE_REGISTRY: "true"
+  FORCE_COLOR: "1"
   PIP_DISABLE_PIP_VERSION_CHECK: "1"
-  RUST_TOOLCHAIN: "nightly-2025-01-07"
+  RUST_TOOLCHAIN: "nightly-2025-04-15"
   UNSAFE_PYO3_BUILD_FREE_THREADED: "1"
   UNSAFE_PYO3_SKIP_VERSION_CHECK: "1"
   UV_LINK_MODE: "copy"
@@ -60,18 +61,19 @@ jobs:
     - run: ./integration/run typestubs
 
     - name: Store sdist
-      if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
         name: orjson_sdist
         path: dist
         overwrite: true
         retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
 
   manylinux_amd64:
     runs-on: ubuntu-24.04
     container:
-      image: fedora:42
+      image: fedora:rawhide
     timeout-minutes: 10
     strategy:
       fail-fast: false
@@ -83,17 +85,19 @@ jobs:
           { interpreter: 'python3.11', compatibility: "manylinux_2_17", publish: true },
           { interpreter: 'python3.10', compatibility: "manylinux_2_17", publish: true },
           { interpreter: 'python3.9', compatibility: "manylinux_2_17", publish: true },
-          { interpreter: 'python3.8', compatibility: "manylinux_2_17", publish: true },
         ]
     env:
       CARGO_TARGET_DIR: "/tmp/orjson"
       CC: "clang"
-      CFLAGS: "-Os -fstrict-aliasing -fno-plt -flto=full -emit-llvm"
+      CFLAGS: "-O2 -fstrict-aliasing -fno-plt -emit-llvm"
       LDFLAGS: "-fuse-ld=lld -Wl,-plugin-opt=also-emit-llvm -Wl,--as-needed -Wl,-zrelro,-znow"
       RUSTFLAGS: "-C linker=clang -C link-arg=-fuse-ld=lld -C linker-plugin-lto -C lto=fat -C link-arg=-Wl,-zrelro,-znow -Z mir-opt-level=4 -Z threads=4 -D warnings"
       VENV: ".venv"
     steps:
 
+      - name: CPU info
+        run: cat /proc/cpuinfo
+
       - run: dnf install --setopt=install_weak_deps=false -y git
 
       - uses: actions/checkout@v4
@@ -103,12 +107,12 @@ jobs:
         with:
           arch: "x86_64"
           interpreter: "${{ matrix.python.interpreter }}"
-          features: "avx512,unstable-simd,yyjson"
+          features: "avx512,yyjson"
           compatibility: "${{ matrix.python.compatibility }}"
           publish: "${{ matrix.python.publish }}"
 
       - name: Store wheels
-        if: "startsWith(github.ref, 'refs/tags/') && inputs.publish == true"
+        if: matrix.python.publish == true
         uses: actions/upload-artifact@v4
         with:
           name: "orjson_manylinux_amd64_${{ matrix.python.interpreter }}_${{ matrix.python.compatibility }}"
@@ -116,10 +120,21 @@ jobs:
           overwrite: true
           retention-days: 1
 
+      - name: Debug
+        env:
+          CARGO_TARGET_DIR: "/tmp/orjson"
+          ORJSON_FEATURES: "avx512,yyjson"
+          PYTHON: "${{ matrix.python.interpreter }}"
+          TARGET: "x86_64-unknown-linux-gnu"
+        run: |
+          export PATH="$PWD/.venv:$HOME/.cargo/bin:$PATH"
+          source .venv/bin/activate
+          script/debug
+
   manylinux_aarch64:
     runs-on: ubuntu-24.04-arm
     container:
-      image: fedora:42
+      image: fedora:rawhide
     timeout-minutes: 10
     strategy:
       fail-fast: false
@@ -131,17 +146,19 @@ jobs:
           { interpreter: 'python3.11', compatibility: "manylinux_2_17", publish: true },
           { interpreter: 'python3.10', compatibility: "manylinux_2_17", publish: true },
           { interpreter: 'python3.9', compatibility: "manylinux_2_17", publish: true },
-          { interpreter: 'python3.8', compatibility: "manylinux_2_17", publish: true },
         ]
     env:
       CARGO_TARGET_DIR: "/tmp/orjson"
       CC: "clang"
-      CFLAGS: "-Os -fstrict-aliasing -fno-plt -flto=full -emit-llvm"
+      CFLAGS: "-O2 -fstrict-aliasing -fno-plt -emit-llvm"
       LDFLAGS: "-fuse-ld=lld -Wl,-plugin-opt=also-emit-llvm -Wl,--as-needed -Wl,-zrelro,-znow"
       RUSTFLAGS: "-C linker=clang -C link-arg=-fuse-ld=lld -C linker-plugin-lto -C lto=fat -C link-arg=-Wl,-zrelro,-znow -Z mir-opt-level=4 -Z threads=4 -D warnings"
       VENV: ".venv"
     steps:
 
+      - name: CPU info
+        run: cat /proc/cpuinfo
+
       - run: dnf install --setopt=install_weak_deps=false -y git
 
       - uses: actions/checkout@v4
@@ -151,12 +168,12 @@ jobs:
         with:
           arch: "aarch64"
           interpreter: "${{ matrix.python.interpreter }}"
-          features: "unstable-simd,yyjson"
+          features: "generic_simd,yyjson"
           compatibility: "${{ matrix.python.compatibility }}"
           publish: "${{ matrix.python.publish }}"
 
       - name: Store wheels
-        if: "startsWith(github.ref, 'refs/tags/') && inputs.publish == true"
+        if: matrix.python.publish == true
         uses: actions/upload-artifact@v4
         with:
           name: "orjson_manylinux_aarch64_${{ matrix.python.interpreter }}_${{ matrix.python.compatibility }}"
@@ -164,9 +181,20 @@ jobs:
           overwrite: true
           retention-days: 1
 
+      - name: Debug
+        env:
+          CARGO_TARGET_DIR: "/tmp/orjson"
+          ORJSON_FEATURES: "generic_simd,yyjson"
+          PYTHON: "${{ matrix.python.interpreter }}"
+          TARGET: "aarch64-unknown-linux-gnu"
+        run: |
+          export PATH="$PWD/.venv:$HOME/.cargo/bin:$PATH"
+          source .venv/bin/activate
+          script/debug
+
   manylinux_cross:
     runs-on: ubuntu-24.04
-    timeout-minutes: 20
+    timeout-minutes: 10
     strategy:
       fail-fast: false
       matrix:
@@ -176,33 +204,32 @@ jobs:
           { interpreter: 'python3.11', abi: 'cp311-cp311' },
           { interpreter: 'python3.10', abi: 'cp310-cp310' },
           { interpreter: 'python3.9', abi: 'cp39-cp39' },
-          { interpreter: 'python3.8', abi: 'cp38-cp38' },
         ]
         target: [
           {
             arch: 'i686',
-            cflags: '-Os -flto -fstrict-aliasing',
-            features: 'unstable-simd,yyjson',
+            cflags: '-Os -fstrict-aliasing',
+            features: 'yyjson',
             rustflags: '-Z mir-opt-level=4 -C lto=fat -D warnings',
             target: 'i686-unknown-linux-gnu',
           },
           {
             arch: 'armv7',
-            cflags: '-Os -flto=full -fstrict-aliasing',
-            features: 'yyjson', # no SIMD
+            cflags: '-Os -fstrict-aliasing',
+            features: 'yyjson',
             rustflags: '-Z mir-opt-level=4 -C lto=fat -D warnings -C opt-level=s',
             target: 'armv7-unknown-linux-gnueabihf',
           },
           {
             arch: 'ppc64le',
-            cflags: '-Os -flto=full -fstrict-aliasing',
-            features: 'unstable-simd,yyjson',
+            cflags: '-Os -fstrict-aliasing',
+            features: 'generic_simd,yyjson',
             rustflags: '-Z mir-opt-level=4 -C lto=fat -D warnings',
             target: 'powerpc64le-unknown-linux-gnu',
           },
           {
             arch: 's390x',
-            cflags: '-Os -flto=full -fstrict-aliasing -march=z10',
+            cflags: '-Os -fstrict-aliasing -march=z10',
             features: 'yyjson',
             rustflags: '-Z mir-opt-level=4 -C lto=fat -D warnings -C target-cpu=z10',
             target: 's390x-unknown-linux-gnu',
@@ -227,17 +254,18 @@ jobs:
         target: "${{ matrix.target.target }}"
         rust-toolchain: "${{ env.RUST_TOOLCHAIN }}"
         rustup-components: rust-src
-        manylinux: auto
+        manylinux: manylinux_2_17
         args: --release --strip --out=dist --features=${{ matrix.target.features }} -i ${{ matrix.python.interpreter }}
 
     - name: Store wheels
-      if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
         name: "orjson_manylinux_${{ matrix.target.arch }}_${{ matrix.python.interpreter }}"
         path: dist
         overwrite: true
         retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
 
   musllinux_amd64:
     runs-on: ubuntu-24.04
@@ -246,22 +274,21 @@ jobs:
       fail-fast: false
       matrix:
         python: [
-          { version: '3.13' },
-          { version: '3.12' },
-          { version: '3.11' },
-          { version: '3.10' },
-          { version: '3.9' },
-          { version: '3.8' },
+          { version: '3.13', pytest: '1' },
+          { version: '3.12', pytest: '1' },
+          { version: '3.11', pytest: '0' },
+          { version: '3.10', pytest: '0' },
+          { version: '3.9', pytest: '0' },
         ]
         platform:
           - target: x86_64-unknown-linux-musl
             arch: x86_64
             platform: linux/amd64
-            features: avx512,unstable-simd,unwind,yyjson
+            features: avx512,unwind,yyjson
           - target: i686-unknown-linux-musl
             arch: i686
             platform: linux/386
-            features: unstable-simd,unwind,yyjson
+            features: unwind,yyjson
     steps:
     - uses: actions/checkout@v4
 
@@ -274,7 +301,7 @@ jobs:
       uses: PyO3/maturin-action@v1
       env:
         CC: "gcc"
-        CFLAGS: "-Os"
+        CFLAGS: "-O2"
         LDFLAGS: "-Wl,--as-needed"
         RUSTFLAGS: "-C lto=fat -Z mir-opt-level=4 -Z threads=2 -D warnings -C target-feature=-crt-static"
       with:
@@ -284,17 +311,10 @@ jobs:
         manylinux: musllinux_1_2
         args: --release --strip --out=dist --features=${{ matrix.platform.features }} -i python${{ matrix.python.version }}
 
-    - name: QEMU
-      if: matrix.platform.arch != 'x86_64'
-      uses: docker/setup-qemu-action@v3
-      with:
-        image: tonistiigi/binfmt:qemu-v8.1.5
-        platforms: ${{ matrix.platform.platform }}
-
     - name: Test
       uses: addnab/docker-run-action@v3
       with:
-        image: quay.io/pypa/musllinux_1_2_${{ matrix.platform.arch }}:latest
+        image: "quay.io/pypa/musllinux_1_2_${{ matrix.platform.arch }}:2025.04.19-1"
         options: -v ${{ github.workspace }}:/io -w /io
         run: |
           apk add tzdata
@@ -303,18 +323,23 @@ jobs:
 
           python${{ matrix.python.version }} -m venv venv
           venv/bin/pip install -U pip wheel
-          venv/bin/pip install -r test/requirements.txt
           venv/bin/pip install orjson --no-index --find-links dist/ --force-reinstall
-          venv/bin/python -m pytest -v test
+
+          # segfault on starting pytest after January 2025 on 3.11 and older; artifact works fine
+          if [ ${{ matrix.python.pytest }} == '1' ]; then
+            venv/bin/pip install -r test/requirements.txt
+            PYTHONMALLOC="debug" venv/bin/python -m pytest -v test
+          fi
 
     - name: Store wheels
-      if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
         name: orjson_musllinux_${{ matrix.platform.arch }}_${{ matrix.python.version }}
         path: dist
         overwrite: true
         retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
 
   musllinux_aarch64:
     runs-on: ubuntu-24.04-arm
@@ -328,17 +353,16 @@ jobs:
           { version: '3.11' },
           { version: '3.10' },
           { version: '3.9' },
-          { version: '3.8' },
         ]
         platform:
           - target: aarch64-unknown-linux-musl
             arch: aarch64
             platform: linux/arm64
-            features: unstable-simd,unwind,yyjson
+            features: generic_simd,unwind,yyjson
           - target: armv7-unknown-linux-musleabihf
             arch: armv7l
             platform: linux/arm/v7
-            features: unstable-simd,yyjson
+            features: yyjson
     steps:
     - uses: actions/checkout@v4
 
@@ -351,7 +375,7 @@ jobs:
       uses: PyO3/maturin-action@v1
       env:
         CC: "gcc"
-        CFLAGS: "-Os"
+        CFLAGS: "-O2"
         LDFLAGS: "-Wl,--as-needed"
         RUSTFLAGS: "-C lto=fat -Z mir-opt-level=4 -Z threads=2 -D warnings -C target-feature=-crt-static"
       with:
@@ -364,7 +388,7 @@ jobs:
     - name: Test
       uses: addnab/docker-run-action@v3
       with:
-        image: quay.io/pypa/musllinux_1_2_${{ matrix.platform.arch }}:latest
+        image: "quay.io/pypa/musllinux_1_2_${{ matrix.platform.arch }}:2025.04.19-1"
         options: -v ${{ github.workspace }}:/io -w /io
         run: |
           apk add tzdata
@@ -375,18 +399,97 @@ jobs:
           venv/bin/pip install -U pip wheel
           venv/bin/pip install -r test/requirements.txt
           venv/bin/pip install orjson --no-index --find-links dist/ --force-reinstall
+          export PYTHONMALLOC="debug"
           venv/bin/python -m pytest -v test
 
     - name: Store wheels
-      if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
         name: orjson_musllinux_${{ matrix.platform.arch }}_${{ matrix.python.version }}
         path: dist
         overwrite: true
         retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
 
   macos_aarch64:
+    runs-on: macos-15
+    timeout-minutes: 10
+    strategy:
+      fail-fast: false
+      matrix:
+        python: [
+          { version: '3.13', macosx_target: "15.0" },
+          { version: '3.12', macosx_target: "15.0" },
+          { version: '3.11', macosx_target: "15.0" },
+        ]
+    env:
+      CC: "clang"
+      LDFLAGS: "-Wl,--as-needed"
+      CFLAGS: "-O2 -fstrict-aliasing -fno-plt -mcpu=apple-m1 -mtune=generic"
+      RUSTFLAGS: "-Z mir-opt-level=4 -Z threads=3 -D warnings"
+      PATH: "/Users/runner/work/orjson/orjson/.venv/bin:/Users/runner/.cargo/bin:/usr/local/opt/curl/bin:/usr/local/bin:/usr/local/sbin:/Users/runner/bin:/Library/Frameworks/Python.framework/Versions/Current/bin:/usr/bin:/bin:/usr/sbin:/sbin"
+    steps:
+
+    - name: CPU info
+      run: sysctl -a | grep brand
+
+    - uses: actions/checkout@v4
+
+    - uses: actions/setup-python@v5
+      with:
+        python-version: "${{ matrix.python.version }}"
+
+    - uses: dtolnay/rust-toolchain@master
+      with:
+        toolchain: "${{ env.RUST_TOOLCHAIN }}"
+        targets: "aarch64-apple-darwin"
+        components: "rust-src"
+
+    - name: Build environment
+      run: |
+        cargo fetch --target aarch64-apple-darwin &
+
+        export PATH=$HOME/.cargo/bin:$HOME/.local/bin:$PATH
+
+        curl -LsSf https://astral.sh/uv/install.sh | sh
+        uv venv --python python${{ matrix.python.version }}
+        uv pip install --upgrade "maturin>=1,<2" -r test/requirements.txt -r integration/requirements.txt
+
+        mkdir .cargo
+        cp ci/config.toml .cargo/config.toml
+
+    - name: maturin
+      run: |
+        export PATH=$HOME/.cargo/bin:$HOME/.local/bin:$PATH
+
+        MACOSX_DEPLOYMENT_TARGET="${{ matrix.python.macosx_target }}" \
+        PYO3_CROSS_LIB_DIR=$(python -c "import sysconfig;print(sysconfig.get_config_var('LIBDIR'))") \
+        maturin build --release --strip \
+          --features=generic_simd,yyjson \
+          --interpreter python${{ matrix.python.version }} \
+          --target=aarch64-apple-darwin
+        uv pip install target/wheels/orjson*.whl
+
+    - run: pytest -v test
+      env:
+        PYTHONMALLOC: "debug"
+
+    - run: source .venv/bin/activate && ./integration/run thread
+    - run: source .venv/bin/activate && ./integration/run http
+    - run: source .venv/bin/activate && ./integration/run init
+
+    - name: Store wheels
+      uses: actions/upload-artifact@v4
+      with:
+        name: orjson_macos_aarch64_${{ matrix.python.version }}
+        path: target/wheels
+        overwrite: true
+        retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
+
+  macos_universal2_aarch64:
     runs-on: macos-15
     timeout-minutes: 10
     strategy:
@@ -400,14 +503,17 @@ jobs:
         ]
     env:
       CC: "clang"
-      CFLAGS: "-Os -fstrict-aliasing -flto=full"
+      CFLAGS: "-O2 -fstrict-aliasing"
       LDFLAGS: "-Wl,--as-needed"
-      CFLAGS_x86_64_apple_darwin: "-O2 -fstrict-aliasing -flto=full -march=x86-64-v2 -mtune=generic"
-      CFLAGS_aarch64_apple_darwin: "-O2 -fstrict-aliasing -flto=full -mcpu=apple-m1 -mtune=generic"
+      CFLAGS_x86_64_apple_darwin: "-O2 -fstrict-aliasing -fno-plt -march=x86-64-v2 -mtune=generic"
+      CFLAGS_aarch64_apple_darwin: "-O2 -fstrict-aliasing -fno-plt -mcpu=apple-m1 -mtune=generic"
       RUSTFLAGS: "-C lto=fat -Z mir-opt-level=4 -Z threads=3 -D warnings"
       PATH: "/Users/runner/work/orjson/orjson/.venv/bin:/Users/runner/.cargo/bin:/usr/local/opt/curl/bin:/usr/local/bin:/usr/local/sbin:/Users/runner/bin:/Library/Frameworks/Python.framework/Versions/Current/bin:/usr/bin:/bin:/usr/sbin:/sbin"
     steps:
 
+    - name: CPU info
+      run: sysctl -a | grep brand
+
     - uses: actions/checkout@v4
 
     - uses: actions/setup-python@v5
@@ -440,7 +546,7 @@ jobs:
         MACOSX_DEPLOYMENT_TARGET="${{ matrix.python.macosx_target }}" \
         PYO3_CROSS_LIB_DIR=$(python -c "import sysconfig;print(sysconfig.get_config_var('LIBDIR'))") \
         maturin build --release --strip \
-          --features=unstable-simd,yyjson \
+          --features=generic_simd,yyjson \
           --interpreter python${{ matrix.python.version }} \
           --target=universal2-apple-darwin
         uv pip install target/wheels/orjson*.whl
@@ -454,15 +560,16 @@ jobs:
     - run: source .venv/bin/activate && ./integration/run init
 
     - name: Store wheels
-      if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
         name: orjson_universal2_aarch64_${{ matrix.python.version }}
         path: target/wheels
         overwrite: true
         retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
 
-  macos_amd64:
+  macos_universal2_amd64:
     runs-on: macos-13
     timeout-minutes: 10
     strategy:
@@ -470,18 +577,20 @@ jobs:
       matrix:
         python: [
           { version: '3.9', macosx_target: "10.15" },
-          { version: '3.8', macosx_target: "10.15" },
         ]
     env:
       CC: "clang"
-      CFLAGS: "-Os -fstrict-aliasing -flto=full"
+      CFLAGS: "-O2 -fstrict-aliasing"
       LDFLAGS: "-Wl,--as-needed"
-      CFLAGS_x86_64_apple_darwin: "-O2 -fstrict-aliasing -flto=full -march=x86-64-v2 -mtune=generic"
-      CFLAGS_aarch64_apple_darwin: "-O2 -fstrict-aliasing -flto=full -mcpu=apple-m1 -mtune=generic"
+      CFLAGS_x86_64_apple_darwin: "-O2 -fstrict-aliasing -fno-plt -march=x86-64-v2 -mtune=generic"
+      CFLAGS_aarch64_apple_darwin: "-O2 -fstrict-aliasing -fno-plt -mcpu=apple-m1 -mtune=generic"
       RUSTFLAGS: "-C lto=fat -Z mir-opt-level=4 -Z threads=3 -D warnings"
       PATH: "/Users/runner/work/orjson/orjson/.venv/bin:/Users/runner/.cargo/bin:/usr/local/opt/curl/bin:/usr/local/bin:/usr/local/sbin:/Users/runner/bin:/Library/Frameworks/Python.framework/Versions/Current/bin:/usr/bin:/bin:/usr/sbin:/sbin"
     steps:
 
+    - name: CPU info
+      run: sysctl -a | grep brand
+
     - uses: actions/checkout@v4
 
     - uses: actions/setup-python@v5
@@ -514,7 +623,7 @@ jobs:
         MACOSX_DEPLOYMENT_TARGET="${{ matrix.python.macosx_target }}" \
         PYO3_CROSS_LIB_DIR=$(python -c "import sysconfig;print(sysconfig.get_config_var('LIBDIR'))") \
         maturin build --release --strip \
-          --features=unstable-simd,yyjson \
+          --features=generic_simd,yyjson \
           --interpreter python${{ matrix.python.version }} \
           --target=universal2-apple-darwin
         uv pip install target/wheels/orjson*.whl
@@ -528,16 +637,17 @@ jobs:
     - run: source .venv/bin/activate && ./integration/run init
 
     - name: Store wheels
-      if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
         name: orjson_universal2_amd64_${{ matrix.python.version }}
         path: target/wheels
         overwrite: true
         retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
 
-  windows:
-    runs-on: windows-2022
+  windows_amd64:
+    runs-on: windows-2025
     timeout-minutes: 10
     strategy:
       fail-fast: false
@@ -548,18 +658,21 @@ jobs:
           { version: '3.11' },
           { version: '3.10' },
           { version: '3.9' },
-          { version: '3.8' },
         ]
         platform: [
-          { arch: "x64", target: "x86_64-pc-windows-msvc" },
-          { arch: "x86", target: "i686-pc-windows-msvc" },
+          { arch: "x64", target: "x86_64-pc-windows-msvc", features: "avx512,yyjson" },
+          { arch: "x86", target: "i686-pc-windows-msvc", features: "yyjson" },
         ]
     env:
-      CFLAGS: "-Os"
+      CFLAGS: "-O2"
       LDFLAGS: "-Wl,--as-needed"
       RUSTFLAGS: "-C lto=fat -Z mir-opt-level=4 -D warnings"
     steps:
 
+    - name: CPU info
+      shell: pwsh
+      run: Get-WmiObject -Class Win32_Processor -ComputerName. | Select-Object -Property Name, NumberOfCores, NumberOfLogicalProcessors
+
     - uses: actions/checkout@v4
 
     - uses: actions/setup-python@v5
@@ -578,14 +691,81 @@ jobs:
         cargo fetch --target "${{ matrix.platform.target }}" &
 
         python.exe -m pip install --upgrade pip "maturin>=1,<2" wheel
-        python.exe -m pip install -r test\requirements.txt -r integration\requirements.txt
+        python.exe -m pip install -r test\requirements.txt
+
+        mkdir .cargo
+        cp ci\config.toml .cargo\config.toml
+
+    - name: maturin
+      run: |
+        maturin.exe build --release --strip --features="${{ matrix.platform.features }}" --target="${{ matrix.platform.target }}"
+        python.exe -m pip install orjson --no-index --find-links target\wheels
+
+    - run: python.exe -m pytest -s -rxX -v test
+      env:
+        PYTHONMALLOC: "debug"
+
+    - name: Store wheels
+      uses: actions/upload-artifact@v4
+      with:
+        name: orjson_windows_amd64_${{ matrix.platform.arch }}_${{ matrix.python.version }}
+        path: target\wheels
+        overwrite: true
+        retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
+
+  windows_aarch64:
+    runs-on: windows-11-arm
+    timeout-minutes: 10
+    strategy:
+      fail-fast: false
+      matrix:
+        python: [
+          { version: '3.13' },
+          { version: '3.12' },
+          { version: '3.11' },
+        ]
+    env:
+      CFLAGS: "-O2"
+      LDFLAGS: "-Wl,--as-needed"
+      RUSTFLAGS: "-Z mir-opt-level=4 -D warnings"
+      TARGET: "aarch64-pc-windows-msvc"
+    steps:
+
+    - name: CPU info
+      shell: pwsh
+      run: Get-WmiObject -Class Win32_Processor -ComputerName. | Select-Object -Property Name, NumberOfCores, NumberOfLogicalProcessors
+
+    - uses: actions/checkout@v4
+
+    - uses: actions/setup-python@v5
+      with:
+        python-version: "${{ matrix.python.version }}"
+        architecture: "arm64"
+
+    # from maturin
+    - shell: pwsh
+      run: |
+        Invoke-WebRequest -Uri "https://static.rust-lang.org/rustup/dist/$env:TARGET/rustup-init.exe" -OutFile rustup-init.exe
+        .\rustup-init.exe --default-toolchain "$env:RUST_TOOLCHAIN-$env:TARGET" --profile minimal --component rust-src -y
+        "$env:USERPROFILE\.cargo\bin" | Out-File -Append -Encoding ascii $env:GITHUB_PATH
+        "CARGO_HOME=$env:USERPROFILE\.cargo" | Out-File -Append -Encoding ascii $env:GITHUB_ENV
+
+    - name: Build environment
+      run: |
+        cargo fetch --target "$" &
+
+        python.exe -m sysconfig
+        python.exe -m pip install --upgrade pip "maturin>=1,<2" wheel
+        python.exe -m pip install -r test\requirements.txt
 
         mkdir .cargo
         cp ci\config.toml .cargo\config.toml
 
     - name: maturin
       run: |
-        maturin.exe build --release --strip --features=unstable-simd,yyjson --target="${{ matrix.platform.target }}"
+        maturin.exe build --release --strip --features=generic_simd --target="$env:TARGET"
         python.exe -m pip install orjson --no-index --find-links target\wheels
 
     - run: python.exe -m pytest -s -rxX -v test
@@ -593,36 +773,40 @@ jobs:
         PYTHONMALLOC: "debug"
 
     - name: Store wheels
-      if: "startsWith(github.ref, 'refs/tags/')"
       uses: actions/upload-artifact@v4
       with:
-        name: orjson_windows_${{ matrix.platform.arch }}_${{ matrix.python.version }}
+        name: orjson_windows_aarch64_${{ matrix.python.version }}
         path: target\wheels
         overwrite: true
         retention-days: 1
+        if-no-files-found: "error"
+        compression-level: 0
 
   pypi:
     name: PyPI
     runs-on: ubuntu-24.04
     timeout-minutes: 10
-    if: "startsWith(github.ref, 'refs/tags/')"
     needs: [
       macos_aarch64,
-      macos_amd64,
+      macos_universal2_aarch64,
+      macos_universal2_amd64,
       manylinux_aarch64,
       manylinux_amd64,
       manylinux_cross,
       musllinux_aarch64,
       musllinux_amd64,
       sdist,
-      windows,
+      windows_aarch64,
+      windows_amd64,
     ]
     environment:
-      name: pypi
+      name: PyPI
       url: https://pypi.org/p/orjson
     permissions:
       id-token: write
     steps:
+      - uses: actions/checkout@v4
+
       - uses: actions/download-artifact@v4
         with:
           merge-multiple: true
@@ -633,9 +817,12 @@ jobs:
 
       - uses: actions/setup-python@v5
         with:
-          python-version: "3.12"
+          python-version: "3.13"
+
+      - run: ./script/check-pypi dist
 
       - name: Publish distribution to PyPI
+        if: "startsWith(github.ref, 'refs/tags/')"
         uses: pypa/gh-action-pypi-publish@release/v1
         with:
           attestations: true
diff --git a/.github/workflows/lint.yaml b/.github/workflows/lint.yaml
index e97ed4b5..c2c90a06 100644
--- a/.github/workflows/lint.yaml
+++ b/.github/workflows/lint.yaml
@@ -1,14 +1,17 @@
 name: lint
 on: push
+env:
+  FORCE_COLOR: "1"
+  PIP_DISABLE_PIP_VERSION_CHECK: "1"
 jobs:
   lint:
     runs-on: ubuntu-24.04
     steps:
     - uses: actions/setup-python@v5
       with:
-        python-version: "3.12"
+        python-version: "3.13"
     - run: curl https://sh.rustup.rs -sSf | sh -s -- --default-toolchain=stable --profile=default -y
-    - run: pip install -U ruff==0.7.1 mypy==1.13.0
+    - run: pip install -U "ruff>=0.11,<0.12" "mypy==1.15.0"
 
     - uses: actions/checkout@v4
 
diff --git a/.github/workflows/debug.yaml b/.github/workflows/unusual.yaml
similarity index 78%
rename from .github/workflows/debug.yaml
rename to .github/workflows/unusual.yaml
index 0111dd40..dc9adda2 100644
--- a/.github/workflows/debug.yaml
+++ b/.github/workflows/unusual.yaml
@@ -1,26 +1,27 @@
-name: debug
+name: unusual
 on: push
+env:
+  CARGO_UNSTABLE_SPARSE_REGISTRY: "true"
+  FORCE_COLOR: "1"
+  PIP_DISABLE_PIP_VERSION_CHECK: "1"
 jobs:
 
-  debug:
-    runs-on: ubuntu-22.04
+  unusual:
+    runs-on: ubuntu-24.04
     strategy:
       fail-fast: false
       matrix:
         profile: [
           { rust: "1.82", features: "" },
-          { rust: "1.82", features: "--features=yyjson" },
-          { rust: "nightly-2025-01-07", features: "--features=avx512,yyjson,unstable-simd" },
+          { rust: "nightly-2025-04-15", features: "--features=generic_simd" },
         ]
         python: [
           { version: '3.13' },
           { version: '3.12' },
-          { version: '3.8' },
+          { version: '3.9' },
         ]
     env:
-      CC: "gcc"
-      CFLAGS: "-Os -fstrict-aliasing"
-      LDFLAGS: "-Wl,--as-needed"
+      ORJSON_DISABLE_YYJSON: "1"
       CARGO_UNSTABLE_SPARSE_REGISTRY: "true"
       UNSAFE_PYO3_SKIP_VERSION_CHECK: "1"
     steps:
@@ -54,7 +55,6 @@ jobs:
       timeout-minutes: 2
 
     - run: ./integration/run http
-      if: matrix.python.version != '3.13'  # alpha 5 httpx type issue
       timeout-minutes: 2
 
     - run: ./integration/run init
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 34f7f14a..dd2910d3 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,32 @@
 # Changelog
 
+## 3.10.18
+
+### Fixed
+
+- Fix incorrect escaping of the vertical tabulation character. This was
+introduced in 3.10.17.
+
+
+## 3.10.17
+
+### Changed
+
+- Publish PyPI Windows aarch64/arm64 wheels.
+- ABI compatibility with CPython 3.14 alpha 7.
+- Fix incompatibility running on Python 3.13 using WASM.
+
+
+## 3.10.16
+
+### Changed
+
+- Improve performance of serialization on amd64 machines with AVX-512.
+- ABI compatibility with CPython 3.14 alpha 6.
+- Drop support for Python 3.8.
+- Publish additional PyPI wheels for macOS that target only aarch64, macOS 15,
+and recent Python.
+
 
 ## 3.10.15
 
diff --git a/Cargo.lock b/Cargo.lock
index a198f82e..03fdf346 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -25,9 +25,9 @@ dependencies = [
 
 [[package]]
 name = "cc"
-version = "1.2.1"
+version = "1.2.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "fd9de9f2205d5ef3fd67e685b0df337994ddd4495e2a28d185500d0e1edfea47"
+checksum = "04da6a0d40b948dfc4fa8f5bbf402b0fc1a64a28dbf7d12ffd683550f2c1b63a"
 dependencies = [
  "shlex",
 ]
@@ -40,9 +40,9 @@ checksum = "baf1de4339761588bc0619e3cbc0120ee582ebb74b53b4efbf79117bd2da40fd"
 
 [[package]]
 name = "compact_str"
-version = "0.8.1"
+version = "0.9.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3b79c4069c6cad78e2e0cdfcbd26275770669fb39fd308a752dc110e83b9af32"
+checksum = "3fdb1325a1cece981e8a296ab8f0f9b63ae357bd0784a9faaf548cc7b480707a"
 dependencies = [
  "castaway",
  "cfg-if",
@@ -55,9 +55,9 @@ dependencies = [
 
 [[package]]
 name = "crunchy"
-version = "0.2.2"
+version = "0.2.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "7a81dae078cea95a014a339291cec439d2f232ebe854a9d672b796c6afafa9b7"
+checksum = "43da5946c66ffcc7745f48db692ffbb10a83bfe0afd96235c5c2a4fb23994929"
 
 [[package]]
 name = "encoding_rs"
@@ -76,9 +76,9 @@ checksum = "07e28edb80900c19c28f1072f2e8aeca7fa06b23cd4169cefe1af5aa3260783f"
 
 [[package]]
 name = "half"
-version = "2.4.1"
+version = "2.6.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "6dd08c532ae367adf81c312a4580bc67f1d0fe8bc9c460520283f4c0ff277888"
+checksum = "459196ed295495a68f7d7fe1d84f6c4b7ff0e21fe3017b2f283c6fac3ad803c9"
 dependencies = [
  "cfg-if",
  "crunchy",
@@ -86,9 +86,9 @@ dependencies = [
 
 [[package]]
 name = "itoa"
-version = "1.0.14"
+version = "1.0.15"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "d75a2a4b1b190afb6f5425f10f6a8f959d2ea0b9c2b1d79553551850539e4674"
+checksum = "4a5f13b858c8d314ee3e8f639011f7ccefe71f97f96e50151fb991f267928e2c"
 
 [[package]]
 name = "itoap"
@@ -98,20 +98,32 @@ checksum = "9028f49264629065d057f340a86acb84867925865f73bbf8d47b4d149a7e88b8"
 
 [[package]]
 name = "jiff"
-version = "0.1.21"
+version = "0.2.10"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "ed0ce60560149333a8e41ca7dc78799c47c5fd435e2bc18faf6a054382eec037"
+checksum = "5a064218214dc6a10fbae5ec5fa888d80c45d611aba169222fc272072bf7aef6"
 dependencies = [
+ "jiff-static",
  "portable-atomic",
  "portable-atomic-util",
  "serde",
 ]
 
+[[package]]
+name = "jiff-static"
+version = "0.2.10"
+source = "registry+https://github.com/rust-lang/crates.io-index"
+checksum = "199b7932d97e325aff3a7030e141eafe7f2c6268e1d1b24859b753a627f45254"
+dependencies = [
+ "proc-macro2",
+ "quote",
+ "syn",
+]
+
 [[package]]
 name = "libc"
-version = "0.2.169"
+version = "0.2.172"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "b5aba8db14291edd000dfcc4d620c7ebfb122c613afb886ca8803fa4e128a20a"
+checksum = "d750af042f7ef4f724306de029d18836c26c1765a54a6a3f094cbd23a7267ffa"
 
 [[package]]
 name = "memchr"
@@ -121,13 +133,13 @@ checksum = "78ca9ab1a0babb1e7d5695e3530886289c18cf2f87ec19a575a0abdce112e3a3"
 
 [[package]]
 name = "once_cell"
-version = "1.20.2"
+version = "1.21.3"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "1261fe7e33c73b354eab43b1273a57c8f967d0391e80353e51f764ac02cf6775"
+checksum = "42f5e15c9953c5e4ccceeb2e7382a716482c34515315f7b03532b8b4e8393d2d"
 
 [[package]]
 name = "orjson"
-version = "3.10.15"
+version = "3.10.18"
 dependencies = [
  "associative-cache",
  "bytecount",
@@ -154,9 +166,9 @@ dependencies = [
 
 [[package]]
 name = "portable-atomic"
-version = "1.10.0"
+version = "1.11.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "280dc24453071f1b63954171985a0b0d30058d287960968b9b2aca264c8d4ee6"
+checksum = "350e9b48cbc6b0e028b0473b114454c6316e57336ee184ceab6e53f72c178b3e"
 
 [[package]]
 name = "portable-atomic-util"
@@ -169,9 +181,9 @@ dependencies = [
 
 [[package]]
 name = "proc-macro2"
-version = "1.0.92"
+version = "1.0.95"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "37d3544b3f2748c54e147655edb5025752e2303145b5aefb3c3ea2c78b973bb0"
+checksum = "02b3e5e68a3a1a02aad3ec490a98007cbc13c37cbe84a3cd7b8e406d76e7f778"
 dependencies = [
  "unicode-ident",
 ]
@@ -194,39 +206,39 @@ dependencies = [
 
 [[package]]
 name = "quote"
-version = "1.0.38"
+version = "1.0.40"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "0e4dccaaaf89514f546c693ddc140f729f958c247918a13380cccc6078391acc"
+checksum = "1885c039570dc00dcb4ff087a89e185fd56bae234ddc7f056a945bf36467248d"
 dependencies = [
  "proc-macro2",
 ]
 
 [[package]]
 name = "rustversion"
-version = "1.0.19"
+version = "1.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f7c45b9784283f1b2e7fb61b42047c2fd678ef0960d4f6f1eba131594cc369d4"
+checksum = "eded382c5f5f786b989652c49544c4877d9f015cc22e145a5ea8ea66c2921cd2"
 
 [[package]]
 name = "ryu"
-version = "1.0.18"
+version = "1.0.20"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f3cb5ba0dc43242ce17de99c180e96db90b235b8a9fdc9543c96d2209116bd9f"
+checksum = "28d3b2b1366ec20994f1fd18c3c594f05c5dd4bc44d8bb0c1c632c8d6829481f"
 
 [[package]]
 name = "serde"
-version = "1.0.217"
+version = "1.0.219"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "02fc4265df13d6fa1d00ecff087228cc0a2b5f3c0e87e258d8b94a156e984c70"
+checksum = "5f0e2c6ed6606019b4e29e69dbaba95b11854410e5347d525002456dbbb786b6"
 dependencies = [
  "serde_derive",
 ]
 
 [[package]]
 name = "serde_derive"
-version = "1.0.217"
+version = "1.0.219"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "5a9bf7cf98d04a2b28aead066b7496853d4779c9cc183c440dbac457641e19a0"
+checksum = "5b0276cf7f2c73365f7157c8123c21cd9a50fbbd844757af28ca1f5925fc2a00"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -235,9 +247,9 @@ dependencies = [
 
 [[package]]
 name = "serde_json"
-version = "1.0.135"
+version = "1.0.140"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "2b0d7ba2887406110130a978386c4e1befb98c674b4fba677954e4db976630d9"
+checksum = "20068b6e96dc6c9bd23e01df8827e6c7e1f2fddd43c21810382803c136b99373"
 dependencies = [
  "itoa",
  "memchr",
@@ -259,9 +271,9 @@ checksum = "e3a9fe34e3e7a50316060351f37187a3f546bce95496156754b601a5fa71b76e"
 
 [[package]]
 name = "smallvec"
-version = "1.13.2"
+version = "1.15.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3c5e1a9a646d36c3599cd173a41282daf47c44583ad367b8e6837255952e5c67"
+checksum = "8917285742e9f3e1683f0a9c4e6b57960b7314d0b08d30d1ecd426713ee2eee9"
 
 [[package]]
 name = "static_assertions"
@@ -271,9 +283,9 @@ checksum = "a2eb9349b6444b326872e140eb1cf5e7c522154d69e7a0ffb0fb81c06b37543f"
 
 [[package]]
 name = "syn"
-version = "2.0.95"
+version = "2.0.101"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "46f71c0377baf4ef1cc3e3402ded576dccc315800fbc62dfc7fe04b009773b4a"
+checksum = "8ce2b7fc941b3a24138a0a7cf8e858bfc6a992e7978a068a5c760deb0ed43caf"
 dependencies = [
  "proc-macro2",
  "quote",
@@ -282,15 +294,15 @@ dependencies = [
 
 [[package]]
 name = "target-lexicon"
-version = "0.13.1"
+version = "0.13.2"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "dc12939a1c9b9d391e0b7135f72fd30508b73450753e28341fed159317582a77"
+checksum = "e502f78cdbb8ba4718f566c418c52bc729126ffd16baee5baa718cf25dd5a69a"
 
 [[package]]
 name = "unicode-ident"
-version = "1.0.14"
+version = "1.0.18"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "adb9e6ca4f869e1180728b7950e35922a7fc6397f7b641499e8f3ef06e50dc83"
+checksum = "5a5f39404a5da50712a4c1eecf25e90dd62b613502b7e925fd4e4d19b5c96512"
 
 [[package]]
 name = "unwinding"
@@ -303,9 +315,9 @@ dependencies = [
 
 [[package]]
 name = "uuid"
-version = "1.11.0"
+version = "1.16.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "f8c5f0a0af699448548ad1a2fbf920fb4bee257eae39953ba95cb84891a0446a"
+checksum = "458f7a779bf54acc9f347480ac654f68407d3aab21269a6e3c9f922acd9e2da9"
 
 [[package]]
 name = "version_check"
diff --git a/Cargo.toml b/Cargo.toml
index bd3cf92a..7b0d8a86 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "orjson"
-version = "3.10.15"
+version = "3.10.18"
 authors = ["ijl <ijl@mailbox.org>"]
 description = "Fast, correct Python JSON library supporting dataclasses, datetimes, and numpy"
 edition = "2021"
@@ -30,10 +30,7 @@ crate-type = ["cdylib"]
 [features]
 default = []
 
-# Use SIMD intrinsics. This requires Rust on the nightly channel.
-unstable-simd = []
-
-# Include runtime-detected functions that use AVX512VL. Requires unstable-simd and amd64.
+# Include runtime-detected functions that use AVX512VL.
 avx512 = []
 
 # Avoid bundling libgcc on musl.
@@ -44,6 +41,7 @@ unwind = ["unwinding"]
 yyjson = []
 
 # Features detected by build.rs. Do not specify.
+generic_simd = []
 inline_int = []
 intrinsics = []
 optimize = []
@@ -51,12 +49,12 @@ optimize = []
 [dependencies]
 associative-cache = { version = "2", default-features = false }
 bytecount = { version = "^0.6.7", default-features = false, features = ["runtime-dispatch-simd"] }
-compact_str = { version = "0.8", default-features = false, features = ["serde"] }
+compact_str = { version = "0.9", default-features = false, features = ["serde"] }
 encoding_rs = { version = "0.8", default-features = false }
 half = { version = "2", default-features = false, features = ["std"] }
 itoa = { version = "1", default-features = false }
 itoap = { version = "1", default-features = false, features = ["std", "simd"] }
-jiff = { version = "^0.1", default-features = false, features = ["alloc"] }
+jiff = { version = "^0.2", default-features = false, features = ["alloc"] }
 once_cell = { version = "1", default-features = false, features = ["alloc", "race"] }
 pyo3-ffi = { path = "include/pyo3/pyo3-ffi", default-features = false, features = ["extension-module"]}
 ryu = { version = "1", default-features = false }
@@ -64,12 +62,12 @@ serde = { version = "1", default-features = false }
 serde_json = { version = "1", default-features = false, features = ["std", "float_roundtrip"] }
 simdutf8 = { version = "0.1", default-features = false, features = ["std", "public_imp", "aarch64_neon"] }
 smallvec = { version = "^1.11", default-features = false, features = ["union", "write"] }
-unwinding = { version = "^0.2.4", default-features = false, features = ["unwinder"], optional = true }
+unwinding = { version = "=0.2.5", default-features = false, features = ["unwinder"], optional = true }
 uuid = { version = "1", default-features = false }
 xxhash-rust = { version = "^0.8", default-features = false, features = ["xxh3"] }
 
 [build-dependencies]
-cc = { version = "=1.2.1" } # aarch64-linux-musl link errors
+cc = { version = "1" }
 pyo3-build-config = { path = "include/pyo3/pyo3-build-config" }
 version_check = { version = "0.9" }
 
diff --git a/README.md b/README.md
index 63fc3776..543e85fe 100644
--- a/README.md
+++ b/README.md
@@ -21,15 +21,17 @@ RFC 8259 ("The JavaScript Object Notation (JSON) Data Interchange Format").
 Reading from and writing to files, line-delimited JSON files, and so on is
 not provided by the library.
 
-orjson supports CPython 3.8, 3.9, 3.10, 3.11, 3.12, 3.13, and 3.14.
+orjson supports CPython 3.9, 3.10, 3.11, 3.12, 3.13, and 3.14.
 
-It distributes amd64/x86_64, i686/x86, aarch64/armv8, arm7, POWER/ppc64le,
-and s390x wheels for Linux, amd64 and aarch64 wheels for macOS, and amd64
-and i686/x86 wheels for Windows.
+It distributes amd64/x86_64/x64, i686/x86, aarch64/arm64/armv8, arm7,
+ppc64le/POWER8, and s390x wheels for Linux, amd64 and aarch64 wheels
+for macOS, and amd64, i686, and aarch64 wheels for Windows.
 
 orjson does not and will not support PyPy, embedded Python builds for
 Android/iOS, or PEP 554 subinterpreters.
 
+orjson may support PEP 703 free-threading when it is stable.
+
 Releases follow semantic versioning and serializing a new object type
 without an opt-in flag is considered a breaking change.
 
@@ -1078,7 +1080,7 @@ It benefits from also having a C build environment to compile a faster
 deserialization backend. See this project's `manylinux_2_28` builds for an
 example using clang and LTO.
 
-The project's own CI tests against `nightly-2025-01-07` and stable 1.72. It
+The project's own CI tests against `nightly-2025-04-15` and stable 1.82. It
 is prudent to pin the nightly version because that channel can introduce
 breaking changes. There is a significant performance benefit to using
 nightly.
@@ -1094,10 +1096,11 @@ The source distribution on PyPI contains all dependencies' source and can be
 built without network access. The file can be downloaded from
 `https://files.pythonhosted.org/packages/source/o/orjson/orjson-${version}.tar.gz`.
 
-orjson's tests are included in the source distribution on PyPI. The
-requirements to run the tests are specified in `test/requirements.txt`. The
-tests should be run as part of the build. It can be run with
-`pytest -q test`.
+orjson's tests are included in the source distribution on PyPI. The tests
+require only `pytest`. There are optional packages such as `pytz` and `numpy`
+listed in `test/requirements.txt` and used in ~10% of tests. Not having these
+dependencies causes the tests needing them to skip. Tests can be run
+with `pytest -q test`.
 
 ## License
 
diff --git a/bench/requirements.txt b/bench/requirements.txt
index cd245e70..e7710b74 100644
--- a/bench/requirements.txt
+++ b/bench/requirements.txt
@@ -1,6 +1,6 @@
-memory-profiler; python_version<"3.13"
-pandas; python_version<"3.13"
+memory-profiler; python_version<"3.14"
+pandas; python_version<"3.14"
 pytest-benchmark
 pytest-random-order
-seaborn; python_version<"3.13"
+seaborn; python_version<"3.14"
 tabulate
diff --git a/build.rs b/build.rs
index 1de74831..43538c0c 100644
--- a/build.rs
+++ b/build.rs
@@ -19,7 +19,6 @@ fn main() {
     println!("cargo:rustc-check-cfg=cfg(Py_3_12)");
     println!("cargo:rustc-check-cfg=cfg(Py_3_13)");
     println!("cargo:rustc-check-cfg=cfg(Py_3_14)");
-    println!("cargo:rustc-check-cfg=cfg(Py_3_8)");
     println!("cargo:rustc-check-cfg=cfg(Py_3_9)");
     println!("cargo:rustc-check-cfg=cfg(Py_GIL_DISABLED)");
 
@@ -28,6 +27,9 @@ fn main() {
         println!("{cfg}");
     }
 
+    #[allow(unused_variables)]
+    let is_64_bit_python = matches!(python_config.pointer_width, Some(64));
+
     if let Some(true) = version_check::supports_feature("core_intrinsics") {
         println!("cargo:rustc-cfg=feature=\"intrinsics\"");
     }
@@ -37,22 +39,7 @@ fn main() {
     }
 
     #[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
-    if env::var("ORJSON_DISABLE_SIMD").is_err() {
-        // auto build unstable SIMD on nightly
-        if let Some(true) = version_check::supports_feature("portable_simd") {
-            println!("cargo:rustc-cfg=feature=\"unstable-simd\"");
-        }
-        // auto build AVX512 on x86-64-v4 or supporting native targets
-        #[cfg(all(target_arch = "x86_64", target_feature = "avx512vl"))]
-        if let Some(true) = version_check::supports_feature("stdarch_x86_avx512") {
-            if env::var("ORJSON_DISABLE_AVX512").is_err() {
-                println!("cargo:rustc-cfg=feature=\"avx512\"");
-            }
-        }
-    }
-
-    #[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
-    if matches!(python_config.pointer_width, Some(64)) {
+    if is_64_bit_python {
         println!("cargo:rustc-cfg=feature=\"inline_int\"");
     }
 
diff --git a/ci/config.toml b/ci/config.toml
index 0823eb5a..8039ce14 100644
--- a/ci/config.toml
+++ b/ci/config.toml
@@ -1,6 +1,7 @@
 [unstable]
 build-std = ["core", "std", "alloc", "proc_macro", "panic_abort"]
 build-std-features = ["panic_immediate_abort"]
+trim-paths = true
 
 [target.x86_64-apple-darwin]
 linker = "clang"
@@ -8,4 +9,4 @@ rustflags = ["-C", "target-cpu=x86-64-v2", "-Z", "tune-cpu=generic"]
 
 [target.aarch64-apple-darwin]
 linker = "clang"
-rustflags = ["-C", "target-cpu=apple-m1"]
+rustflags = ["-C", "target-cpu=apple-m1", "-Z", "tune-cpu=generic"]
diff --git a/include/pyo3/pyo3-ffi/src/cpython/longobject.rs b/include/pyo3/pyo3-ffi/src/cpython/longobject.rs
index 45acaae5..6063575e 100644
--- a/include/pyo3/pyo3-ffi/src/cpython/longobject.rs
+++ b/include/pyo3/pyo3-ffi/src/cpython/longobject.rs
@@ -61,6 +61,18 @@ extern "C" {
         is_signed: c_int,
     ) -> *mut PyObject;
 
+    #[cfg(Py_3_13)]
+    #[cfg_attr(PyPy, link_name = "_PyPyLong_AsByteArrayO")]
+    pub fn _PyLong_AsByteArray(
+        v: *mut PyLongObject,
+        bytes: *mut c_uchar,
+        n: size_t,
+        little_endian: c_int,
+        is_signed: c_int,
+        with_exceptions: c_int,
+    ) -> c_int;
+
+    #[cfg(not(Py_3_13))]
     #[cfg_attr(PyPy, link_name = "_PyPyLong_AsByteArrayO")]
     pub fn _PyLong_AsByteArray(
         v: *mut PyLongObject,
diff --git a/include/pyo3/pyo3-ffi/src/cpython/tupleobject.rs b/include/pyo3/pyo3-ffi/src/cpython/tupleobject.rs
index 9616d437..c06609fb 100644
--- a/include/pyo3/pyo3-ffi/src/cpython/tupleobject.rs
+++ b/include/pyo3/pyo3-ffi/src/cpython/tupleobject.rs
@@ -5,6 +5,8 @@ use crate::pyport::Py_ssize_t;
 #[repr(C)]
 pub struct PyTupleObject {
     pub ob_base: PyVarObject,
+    #[cfg(Py_3_14)]
+    pub ob_hash: crate::Py_hash_t,
     pub ob_item: [*mut PyObject; 1],
 }
 
diff --git a/include/pyo3/pyo3-ffi/src/cpython/unicodeobject.rs b/include/pyo3/pyo3-ffi/src/cpython/unicodeobject.rs
index fae626b8..72da46c1 100644
--- a/include/pyo3/pyo3-ffi/src/cpython/unicodeobject.rs
+++ b/include/pyo3/pyo3-ffi/src/cpython/unicodeobject.rs
@@ -119,9 +119,12 @@ where
 
 #[cfg(not(GraalPy))]
 const STATE_INTERNED_INDEX: usize = 0;
-#[cfg(not(GraalPy))]
+#[cfg(all(not(GraalPy), not(Py_3_14)))]
 const STATE_INTERNED_WIDTH: u8 = 2;
 
+#[cfg(all(not(GraalPy), Py_3_14))]
+const STATE_INTERNED_WIDTH: u8 = 16;
+
 #[cfg(not(GraalPy))]
 const STATE_KIND_INDEX: usize = STATE_INTERNED_WIDTH as usize;
 #[cfg(not(GraalPy))]
diff --git a/include/yyjson/yyjson.c b/include/yyjson/yyjson.c
index 803c2f56..1031182d 100644
--- a/include/yyjson/yyjson.c
+++ b/include/yyjson/yyjson.c
@@ -5265,7 +5265,7 @@ static_inline bool read_string(u8 **ptr,
     
     u8 *cur = *ptr;
     u8 **end = ptr;
-    u8 *src = ++cur, *dst, *pos;
+    u8 *src = ++cur, *dst;
     u16 hi, lo;
     u32 uni, tmp;
     
@@ -5332,7 +5332,6 @@ static_inline bool read_string(u8 **ptr,
          consecutively. We process the byte sequences of the same length in each
          loop, which is more friendly to branch prediction.
          */
-        pos = src;
 #if YYJSON_DISABLE_UTF8_VALIDATION
         while (true) repeat8({
             if (likely((*src & 0xF0) == 0xE0)) src += 3;
@@ -5550,7 +5549,6 @@ static_inline bool read_string(u8 **ptr,
     
 copy_utf8:
     if (*src & 0x80) { /* non-ASCII character */
-        pos = src;
         uni = byte_load_4(src);
 #if YYJSON_DISABLE_UTF8_VALIDATION
         while (true) repeat4({
diff --git a/integration/requirements.txt b/integration/requirements.txt
index 2f47d3a7..71e82086 100644
--- a/integration/requirements.txt
+++ b/integration/requirements.txt
@@ -1,3 +1,3 @@
 flask;sys_platform!="win"
 gunicorn;sys_platform!="win"
-httpx==0.27.2;sys_platform!="win"
+httpx==0.28.1;sys_platform!="win"
diff --git a/pyproject.toml b/pyproject.toml
index e9790e35..701765cf 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,8 +1,8 @@
 [project]
 name = "orjson"
-version = "3.10.15"
+version = "3.10.18"
 repository = "https://github.com/ijl/orjson"
-requires-python = ">=3.8"
+requires-python = ">=3.9"
 classifiers = [
     "Development Status :: 5 - Production/Stable",
     "Intended Audience :: Developers",
@@ -12,7 +12,6 @@ classifiers = [
     "Operating System :: Microsoft :: Windows",
     "Operating System :: POSIX :: Linux",
     "Programming Language :: Python :: 3",
-    "Programming Language :: Python :: 3.8",
     "Programming Language :: Python :: 3.9",
     "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
@@ -47,7 +46,7 @@ include = [
 
 [tool.ruff]
 line-length = 88
-target-version = "py38"
+target-version = "py39"
 
 [tool.ruff.lint]
 select = [
@@ -62,7 +61,7 @@ ignore = [
 known-first-party = ["orjson"]
 
 [tool.mypy]
-python_version = "3.8"
+python_version = "3.9"
 
 [[tool.mypy.overrides]]
 module = ["dateutil", "pytz"]
diff --git a/requirements.txt b/requirements.txt
index f898acc2..1e477a76 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -2,5 +2,5 @@
 -r integration/requirements.txt
 -r test/requirements.txt
 maturin>=1,<2
-mypy==1.13.0
-ruff>=0.8,<9
+mypy==1.15.0
+ruff>=0.11,<0.12
diff --git a/script/cargo b/script/cargo
new file mode 100755
index 00000000..3e20029b
--- /dev/null
+++ b/script/cargo
@@ -0,0 +1,8 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+export UNSAFE_PYO3_BUILD_FREE_THREADED=1
+export UNSAFE_PYO3_SKIP_VERSION_CHECK=1
+
+RUSTFLAGS="-C panic=abort -Z panic_abort_tests" cargo "$@" --target="${TARGET:-x86_64-unknown-linux-gnu}"
diff --git a/script/check-pypi b/script/check-pypi
new file mode 100755
index 00000000..54f60216
--- /dev/null
+++ b/script/check-pypi
@@ -0,0 +1,86 @@
+#!/usr/bin/env python3
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+import sys
+from pathlib import Path
+
+import tomllib
+
+dist = sys.argv[1]
+
+pyproject_doc = tomllib.loads(Path("pyproject.toml").read_text(encoding="utf-8"))
+pyproject_version = pyproject_doc["project"]["version"]
+
+prefix = f"orjson-{pyproject_version}"
+
+abis = (
+    "cp39-cp39",
+    "cp310-cp310",
+    "cp311-cp311",
+    "cp312-cp312",
+    "cp313-cp313",
+)
+
+per_abi_tags = (
+    "macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2",
+    "manylinux_2_17_aarch64.manylinux2014_aarch64",
+    "manylinux_2_17_armv7l.manylinux2014_armv7l",
+    "manylinux_2_17_ppc64le.manylinux2014_ppc64le",
+    "manylinux_2_17_s390x.manylinux2014_s390x",
+    "manylinux_2_17_x86_64.manylinux2014_x86_64",
+    "manylinux_2_17_i686.manylinux2014_i686",
+    "musllinux_1_2_aarch64",
+    "musllinux_1_2_armv7l",
+    "musllinux_1_2_i686",
+    "musllinux_1_2_x86_64",
+    "win32",
+    "win_amd64",
+)
+
+wheels_matrix = set()
+
+# orjson-3.10.15-cp310-cp310-macosx_10_15_x86_64.macosx_11_0_arm64.macosx_10_15_universal2.whl
+for abi in abis:
+    for tag in per_abi_tags:
+        wheels_matrix.add(f"{prefix}-{abi}-{tag}.whl")
+
+wheels_unique = {
+    f"{prefix}-cp311-cp311-macosx_15_0_arm64.whl",
+    f"{prefix}-cp311-cp311-win_arm64.whl",
+    f"{prefix}-cp312-cp312-macosx_15_0_arm64.whl",
+    f"{prefix}-cp312-cp312-win_arm64.whl",
+    f"{prefix}-cp313-cp313-macosx_15_0_arm64.whl",
+    f"{prefix}-cp313-cp313-win_arm64.whl",
+}
+
+wheels_expected = wheels_matrix | wheels_unique
+
+wheels_queued = set(
+    (str(each).replace(f"{dist}/", "") for each in Path(dist).glob("*.whl"))
+)
+
+exit_code = 0
+
+# sdist
+sdist_path = Path(f"{dist}/{prefix}.tar.gz")
+if sdist_path.exists():
+    print(f"sdist present\n")
+else:
+    exit_code = 1
+    print(f"Missing sdist:\n{sdist_path}\n")
+
+# whl
+if wheels_expected == wheels_queued:
+    print(f"Wheels as expected, {len(wheels_queued)} total\n")
+else:
+    exit_code = 1
+
+    missing = "\n".join(sorted(wheels_expected - wheels_queued))
+    if missing:
+        print(f"Missing wheels:\n{missing}\n")
+
+    additional = "\n".join(sorted(wheels_queued - wheels_expected))
+    if additional:
+        print(f"Unexpected wheels:\n{additional}\n")
+
+sys.exit(exit_code)
diff --git a/script/debug b/script/debug
new file mode 100755
index 00000000..c5d8f4d5
--- /dev/null
+++ b/script/debug
@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+
+set -eou pipefail
+
+rm -rf .cargo
+rm -f ${CARGO_TARGET_DIR}/wheels/*.whl
+
+export UNSAFE_PYO3_BUILD_FREE_THREADED=1
+export UNSAFE_PYO3_SKIP_VERSION_CHECK=1
+
+export CC="${CC:-clang}"
+export LD="${LD:-lld}"
+export TARGET="${TARGET:-x86_64-unknown-linux-gnu}"
+export CARGO_TARGET_DIR="${CARGO_TARGET_DIR:-target}"
+export ORJSON_FEATURES="${ORJSON_FEATURES:-yyjson}"
+
+export CFLAGS="-Os -fstrict-aliasing"
+
+export RUSTFLAGS="-C panic=unwind -C linker=${CC} -C link-arg=-fuse-ld=${LD}"
+
+maturin build --profile=dev --target=${TARGET} --features="${ORJSON_FEATURES}" --interpreter "${PYTHON}"
+
+uv pip install ${CARGO_TARGET_DIR}/wheels/*.whl
+
+pytest -v test
+
+mkdir .cargo
+cp ci/config.toml .cargo
diff --git a/script/develop b/script/develop
index b92a22e9..fa1c66b8 100755
--- a/script/develop
+++ b/script/develop
@@ -12,7 +12,7 @@ export CARGO_TARGET_DIR="${CARGO_TARGET_DIR:-target}"
 
 echo "CC: ${CC}, LD: ${LD}, LD_LIBRARY_PATH: ${LD_LIBRARY_PATH}"
 
-export CFLAGS="-Os -fstrict-aliasing -fno-plt -flto=full -emit-llvm"
+export CFLAGS="-O2 -fstrict-aliasing -fno-plt -emit-llvm"
 export LDFLAGS="-fuse-ld=${LD} -Wl,-plugin-opt=also-emit-llvm -Wl,--as-needed -Wl,-zrelro,-znow"
 export RUSTFLAGS="-C linker=${CC} -C link-arg=-fuse-ld=${LD} -C linker-plugin-lto -C lto=fat -C link-arg=-Wl,-zrelro,-znow -Z mir-opt-level=4 -Z threads=8"
 
diff --git a/script/lint b/script/lint
index e8e03b28..4def3a3f 100755
--- a/script/lint
+++ b/script/lint
@@ -4,7 +4,8 @@ set -eou pipefail
 
 to_lint="./bench/*.py ./pysrc/orjson/__init__.pyi ./test/*.py script/pydataclass
 script/pysort script/pynumpy script/pynonstr script/pycorrectness script/graph integration/init
-integration/wsgi.py integration/typestubs.py integration/thread script/check-version"
+integration/wsgi.py integration/typestubs.py integration/thread script/check-version
+script/check-pypi"
 
 ruff check ${to_lint} --fix
 ruff format ${to_lint}
diff --git a/src/alloc.rs b/src/alloc.rs
index 4c54938d..71f29998 100644
--- a/src/alloc.rs
+++ b/src/alloc.rs
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use std::alloc::{GlobalAlloc, Layout};
-use std::ffi::c_void;
+use core::alloc::{GlobalAlloc, Layout};
+use core::ffi::c_void;
 
 struct PyMemAllocator {}
 
@@ -13,19 +13,19 @@ unsafe impl Sync for PyMemAllocator {}
 unsafe impl GlobalAlloc for PyMemAllocator {
     #[inline]
     unsafe fn alloc(&self, layout: Layout) -> *mut u8 {
-        unsafe { pyo3_ffi::PyMem_Malloc(layout.size()) as *mut u8 }
+        unsafe { pyo3_ffi::PyMem_Malloc(layout.size()).cast::<u8>() }
     }
 
     #[inline]
     unsafe fn dealloc(&self, ptr: *mut u8, _layout: Layout) {
-        unsafe { pyo3_ffi::PyMem_Free(ptr as *mut c_void) }
+        unsafe { pyo3_ffi::PyMem_Free(ptr.cast::<c_void>()) }
     }
 
     #[inline]
     unsafe fn alloc_zeroed(&self, layout: Layout) -> *mut u8 {
         unsafe {
             let len = layout.size();
-            let ptr = pyo3_ffi::PyMem_Malloc(len) as *mut u8;
+            let ptr = pyo3_ffi::PyMem_Malloc(len).cast::<u8>();
             core::ptr::write_bytes(ptr, 0, len);
             ptr
         }
@@ -33,6 +33,6 @@ unsafe impl GlobalAlloc for PyMemAllocator {
 
     #[inline]
     unsafe fn realloc(&self, ptr: *mut u8, _layout: Layout, new_size: usize) -> *mut u8 {
-        unsafe { pyo3_ffi::PyMem_Realloc(ptr as *mut c_void, new_size) as *mut u8 }
+        unsafe { pyo3_ffi::PyMem_Realloc(ptr.cast::<c_void>(), new_size).cast::<u8>() }
     }
 }
diff --git a/src/deserialize/backend/json.rs b/src/deserialize/backend/json.rs
index 4a962b0e..5cb6bb37 100644
--- a/src/deserialize/backend/json.rs
+++ b/src/deserialize/backend/json.rs
@@ -128,16 +128,7 @@ impl<'de> Visitor<'de> for JsonValue {
         while let Some(key) = map.next_key::<Cow<str>>()? {
             let pykey = get_unicode_key(&key);
             let pyval = map.next_value_seed(self)?;
-            let _ = unsafe {
-                pyo3_ffi::_PyDict_SetItem_KnownHash(
-                    dict_ptr,
-                    pykey,
-                    pyval.as_ptr(),
-                    str_hash!(pykey),
-                )
-            };
-            reverse_pydict_incref!(pykey);
-            reverse_pydict_incref!(pyval.as_ptr());
+            pydict_setitem!(dict_ptr, pykey, pyval.as_ptr());
         }
         Ok(nonnull!(dict_ptr))
     }
diff --git a/src/deserialize/backend/yyjson.rs b/src/deserialize/backend/yyjson.rs
index 9840e183..24d247d7 100644
--- a/src/deserialize/backend/yyjson.rs
+++ b/src/deserialize/backend/yyjson.rs
@@ -1,10 +1,16 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use crate::deserialize::pyobject::*;
+use crate::deserialize::pyobject::{
+    get_unicode_key, parse_f64, parse_false, parse_i64, parse_none, parse_true, parse_u64,
+};
 use crate::deserialize::DeserializeError;
-use crate::ffi::yyjson::*;
+use crate::ffi::yyjson::{
+    yyjson_doc, yyjson_doc_free, yyjson_read_err, yyjson_read_opts, yyjson_val, YYJSON_READ_SUCCESS,
+};
 use crate::str::unicode_from_str;
 use crate::typeref::{yyjson_init, YYJSON_ALLOC, YYJSON_BUFFER_SIZE};
+use crate::util::usize_to_isize;
+
 use core::ffi::c_char;
 use core::ptr::{null, null_mut, NonNull};
 use std::borrow::Cow;
@@ -49,12 +55,14 @@ fn unsafe_yyjson_is_ctn(val: *mut yyjson_val) -> bool {
     unsafe { (*val).tag as u8 & 0b00000110 == 0b00000110 }
 }
 
+#[allow(clippy::cast_ptr_alignment)]
 fn unsafe_yyjson_get_next_container(val: *mut yyjson_val) -> *mut yyjson_val {
-    unsafe { ((val as *mut u8).add((*val).uni.ofs)) as *mut yyjson_val }
+    unsafe { (val.cast::<u8>().add((*val).uni.ofs)).cast::<yyjson_val>() }
 }
 
+#[allow(clippy::cast_ptr_alignment)]
 fn unsafe_yyjson_get_next_non_container(val: *mut yyjson_val) -> *mut yyjson_val {
-    unsafe { ((val as *mut u8).add(YYJSON_VAL_SIZE)) as *mut yyjson_val }
+    unsafe { (val.cast::<u8>().add(YYJSON_VAL_SIZE)).cast::<yyjson_val>() }
 }
 
 pub(crate) fn deserialize(
@@ -91,16 +99,16 @@ pub(crate) fn deserialize(
             unsafe { yyjson_doc_free(doc) };
             Ok(pyval)
         } else if is_yyjson_tag!(val, TAG_ARRAY) {
-            let pyval = nonnull!(ffi!(PyList_New(unsafe_yyjson_get_len(val) as isize)));
+            let pyval = nonnull!(ffi!(PyList_New(usize_to_isize(unsafe_yyjson_get_len(val)))));
             if unsafe_yyjson_get_len(val) > 0 {
                 populate_yy_array(pyval.as_ptr(), val);
             }
             unsafe { yyjson_doc_free(doc) };
             Ok(pyval)
         } else {
-            let pyval = nonnull!(ffi!(_PyDict_NewPresized(
-                unsafe_yyjson_get_len(val) as isize
-            )));
+            let pyval = nonnull!(ffi!(_PyDict_NewPresized(usize_to_isize(
+                unsafe_yyjson_get_len(val)
+            ))));
             if unsafe_yyjson_get_len(val) > 0 {
                 populate_yy_object(pyval.as_ptr(), val);
             }
@@ -157,7 +165,7 @@ impl ElementType {
 #[inline(always)]
 fn parse_yy_string(elem: *mut yyjson_val) -> NonNull<pyo3_ffi::PyObject> {
     nonnull!(unicode_from_str(str_from_slice!(
-        (*elem).uni.str_ as *const u8,
+        (*elem).uni.str_.cast::<u8>(),
         unsafe_yyjson_get_len(elem)
     )))
 }
@@ -192,20 +200,22 @@ fn populate_yy_array(list: *mut pyo3_ffi::PyObject, elem: *mut yyjson_val) {
         let len = unsafe_yyjson_get_len(elem);
         assume!(len >= 1);
         let mut next = unsafe_yyjson_get_first(elem);
-        let mut dptr = (*(list as *mut pyo3_ffi::PyListObject)).ob_item;
+        let mut dptr = (*list.cast::<pyo3_ffi::PyListObject>()).ob_item;
 
         for _ in 0..len {
             let val = next;
             if unlikely!(unsafe_yyjson_is_ctn(val)) {
                 next = unsafe_yyjson_get_next_container(val);
                 if is_yyjson_tag!(val, TAG_ARRAY) {
-                    let pyval = ffi!(PyList_New(unsafe_yyjson_get_len(val) as isize));
+                    let pyval = ffi!(PyList_New(usize_to_isize(unsafe_yyjson_get_len(val))));
                     append_to_list!(dptr, pyval);
                     if unsafe_yyjson_get_len(val) > 0 {
                         populate_yy_array(pyval, val);
                     }
                 } else {
-                    let pyval = ffi!(_PyDict_NewPresized(unsafe_yyjson_get_len(val) as isize));
+                    let pyval = ffi!(_PyDict_NewPresized(usize_to_isize(unsafe_yyjson_get_len(
+                        val
+                    ))));
                     append_to_list!(dptr, pyval);
                     if unsafe_yyjson_get_len(val) > 0 {
                         populate_yy_object(pyval, val);
@@ -230,12 +240,6 @@ fn populate_yy_array(list: *mut pyo3_ffi::PyObject, elem: *mut yyjson_val) {
     }
 }
 
-macro_rules! add_to_dict {
-    ($dict:expr, $pykey:expr, $pyval:expr) => {
-        unsafe { pyo3_ffi::_PyDict_SetItem_KnownHash($dict, $pykey, $pyval, str_hash!($pykey)) }
-    };
-}
-
 #[inline(never)]
 fn populate_yy_object(dict: *mut pyo3_ffi::PyObject, elem: *mut yyjson_val) {
     unsafe {
@@ -247,7 +251,7 @@ fn populate_yy_object(dict: *mut pyo3_ffi::PyObject, elem: *mut yyjson_val) {
             let val = next_val;
             let pykey = {
                 let key_str = str_from_slice!(
-                    (*next_key).uni.str_ as *const u8,
+                    (*next_key).uni.str_.cast::<u8>(),
                     unsafe_yyjson_get_len(next_key)
                 );
                 get_unicode_key(key_str)
@@ -256,18 +260,16 @@ fn populate_yy_object(dict: *mut pyo3_ffi::PyObject, elem: *mut yyjson_val) {
                 next_key = unsafe_yyjson_get_next_container(val);
                 next_val = next_key.add(1);
                 if is_yyjson_tag!(val, TAG_ARRAY) {
-                    let pyval = ffi!(PyList_New(unsafe_yyjson_get_len(val) as isize));
-                    add_to_dict!(dict, pykey, pyval);
-                    reverse_pydict_incref!(pykey);
-                    reverse_pydict_incref!(pyval);
+                    let pyval = ffi!(PyList_New(usize_to_isize(unsafe_yyjson_get_len(val))));
+                    pydict_setitem!(dict, pykey, pyval);
                     if unsafe_yyjson_get_len(val) > 0 {
                         populate_yy_array(pyval, val);
                     }
                 } else {
-                    let pyval = ffi!(_PyDict_NewPresized(unsafe_yyjson_get_len(val) as isize));
-                    add_to_dict!(dict, pykey, pyval);
-                    reverse_pydict_incref!(pykey);
-                    reverse_pydict_incref!(pyval);
+                    let pyval = ffi!(_PyDict_NewPresized(usize_to_isize(unsafe_yyjson_get_len(
+                        val
+                    ))));
+                    pydict_setitem!(dict, pykey, pyval);
                     if unsafe_yyjson_get_len(val) > 0 {
                         populate_yy_object(pyval, val);
                     }
@@ -286,9 +288,7 @@ fn populate_yy_object(dict: *mut pyo3_ffi::PyObject, elem: *mut yyjson_val) {
                     ElementType::Array => unreachable_unchecked!(),
                     ElementType::Object => unreachable_unchecked!(),
                 };
-                add_to_dict!(dict, pykey, pyval.as_ptr());
-                reverse_pydict_incref!(pykey);
-                reverse_pydict_incref!(pyval.as_ptr());
+                pydict_setitem!(dict, pykey, pyval.as_ptr());
             }
         }
     }
diff --git a/src/deserialize/cache.rs b/src/deserialize/cache.rs
index f5357893..20848a3b 100644
--- a/src/deserialize/cache.rs
+++ b/src/deserialize/cache.rs
@@ -15,11 +15,11 @@ unsafe impl Sync for CachedKey {}
 impl CachedKey {
     pub fn new(ptr: *mut pyo3_ffi::PyObject) -> CachedKey {
         CachedKey {
-            ptr: ptr as *mut c_void,
+            ptr: ptr.cast::<c_void>(),
         }
     }
     pub fn get(&mut self) -> *mut pyo3_ffi::PyObject {
-        let ptr = self.ptr as *mut pyo3_ffi::PyObject;
+        let ptr = self.ptr.cast::<pyo3_ffi::PyObject>();
         debug_assert!(ffi!(Py_REFCNT(ptr)) >= 1);
         ffi!(Py_INCREF(ptr));
         ptr
@@ -28,7 +28,7 @@ impl CachedKey {
 
 impl Drop for CachedKey {
     fn drop(&mut self) {
-        ffi!(Py_DECREF(self.ptr as *mut pyo3_ffi::PyObject));
+        ffi!(Py_DECREF(self.ptr.cast::<pyo3_ffi::PyObject>()));
     }
 }
 
@@ -36,9 +36,3 @@ pub type KeyMap =
     AssociativeCache<u64, CachedKey, Capacity2048, HashDirectMapped, RoundRobinReplacement>;
 
 pub static mut KEY_MAP: OnceCell<KeyMap> = OnceCell::new();
-
-#[inline(always)]
-pub fn cache_hash(key: &[u8]) -> u64 {
-    assume!(key.len() <= 64);
-    xxhash_rust::xxh3::xxh3_64(key)
-}
diff --git a/src/deserialize/deserializer.rs b/src/deserialize/deserializer.rs
index b5a94f3a..3a269649 100644
--- a/src/deserialize/deserializer.rs
+++ b/src/deserialize/deserializer.rs
@@ -21,7 +21,7 @@ pub fn deserialize(
         }
     }
 
-    let buffer_str = unsafe { std::str::from_utf8_unchecked(buffer) };
+    let buffer_str = unsafe { core::str::from_utf8_unchecked(buffer) };
 
     crate::deserialize::backend::deserialize(buffer_str)
 }
diff --git a/src/deserialize/pyobject.rs b/src/deserialize/pyobject.rs
index e486c916..2809b662 100644
--- a/src/deserialize/pyobject.rs
+++ b/src/deserialize/pyobject.rs
@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use crate::deserialize::cache::*;
+use crate::deserialize::cache::{CachedKey, KEY_MAP};
 use crate::str::{hash_str, unicode_from_str};
 use crate::typeref::{FALSE, NONE, TRUE};
 use core::ptr::NonNull;
@@ -12,7 +12,8 @@ pub fn get_unicode_key(key_str: &str) -> *mut pyo3_ffi::PyObject {
         hash_str(pyob);
         pyob
     } else {
-        let hash = cache_hash(key_str.as_bytes());
+        assume!(key_str.len() <= 64);
+        let hash = xxhash_rust::xxh3::xxh3_64(key_str.as_bytes());
         unsafe {
             let entry = KEY_MAP
                 .get_mut()
diff --git a/src/deserialize/utf8.rs b/src/deserialize/utf8.rs
index 2d2842c6..9e993330 100644
--- a/src/deserialize/utf8.rs
+++ b/src/deserialize/utf8.rs
@@ -1,9 +1,10 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
 use crate::deserialize::DeserializeError;
-use crate::ffi::*;
+use crate::ffi::{PyBytes_AS_STRING, PyBytes_GET_SIZE, PyMemoryView_GET_BUFFER};
 use crate::str::unicode_to_str;
 use crate::typeref::{BYTEARRAY_TYPE, BYTES_TYPE, MEMORYVIEW_TYPE, STR_TYPE};
+use crate::util::isize_to_usize;
 use crate::util::INVALID_STR;
 use core::ffi::c_char;
 use std::borrow::Cow;
@@ -24,7 +25,7 @@ fn is_valid_utf8(buf: &[u8]) -> bool {
 
 #[cfg(target_arch = "aarch64")]
 fn is_valid_utf8(buf: &[u8]) -> bool {
-    simdutf8::basic::from_utf8(buf).is_ok()
+    unsafe { simdutf8::basic::imp::aarch64::neon::validate_utf8(buf).is_ok() }
 }
 
 #[cfg(not(any(target_arch = "x86_64", target_arch = "aarch64")))]
@@ -40,8 +41,8 @@ pub fn read_input_to_buf(
     if is_type!(obj_type_ptr, BYTES_TYPE) {
         buffer = unsafe {
             core::slice::from_raw_parts(
-                PyBytes_AS_STRING(ptr) as *const u8,
-                PyBytes_GET_SIZE(ptr) as usize,
+                PyBytes_AS_STRING(ptr).cast::<u8>(),
+                isize_to_usize(PyBytes_GET_SIZE(ptr)),
             )
         };
         if !is_valid_utf8(buffer) {
@@ -62,7 +63,7 @@ pub fn read_input_to_buf(
             )));
         }
         buffer = unsafe {
-            core::slice::from_raw_parts((*membuf).buf as *const u8, (*membuf).len as usize)
+            core::slice::from_raw_parts((*membuf).buf as *const u8, isize_to_usize((*membuf).len))
         };
         if !is_valid_utf8(buffer) {
             return Err(DeserializeError::invalid(Cow::Borrowed(INVALID_STR)));
@@ -71,7 +72,7 @@ pub fn read_input_to_buf(
         buffer = unsafe {
             core::slice::from_raw_parts(
                 ffi!(PyByteArray_AsString(ptr)) as *const u8,
-                ffi!(PyByteArray_Size(ptr)) as usize,
+                isize_to_usize(ffi!(PyByteArray_Size(ptr))),
             )
         };
         if !is_valid_utf8(buffer) {
diff --git a/src/ffi/buffer.rs b/src/ffi/buffer.rs
index cd05081a..a4ae4016 100644
--- a/src/ffi/buffer.rs
+++ b/src/ffi/buffer.rs
@@ -1,7 +1,7 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
 use core::ffi::c_int;
-use pyo3_ffi::*;
+use pyo3_ffi::{PyObject, PyVarObject, Py_buffer, Py_hash_t, Py_ssize_t};
 
 #[repr(C)]
 pub struct _PyManagedBufferObject {
diff --git a/src/ffi/fragment.rs b/src/ffi/fragment.rs
index e890fc35..9a7fe098 100644
--- a/src/ffi/fragment.rs
+++ b/src/ffi/fragment.rs
@@ -6,7 +6,11 @@ use core::ffi::c_char;
 use std::sync::atomic::{AtomicIsize, AtomicU32, AtomicU64};
 
 use core::ptr::null_mut;
-use pyo3_ffi::*;
+use pyo3_ffi::{
+    PyErr_SetObject, PyExc_TypeError, PyObject, PyTuple_GET_ITEM, PyTypeObject, PyType_Ready,
+    PyType_Type, PyUnicode_FromStringAndSize, PyVarObject, Py_DECREF, Py_INCREF, Py_SIZE,
+    Py_TPFLAGS_DEFAULT,
+};
 
 // https://docs.python.org/3/c-api/typeobj.html#typedef-examples
 
@@ -21,7 +25,7 @@ pub struct Fragment {
     #[cfg(Py_GIL_DISABLED)]
     pub _padding: u16,
     #[cfg(Py_GIL_DISABLED)]
-    pub ob_mutex: PyMutex,
+    pub ob_mutex: pyo3_ffi::PyMutex,
     #[cfg(Py_GIL_DISABLED)]
     pub ob_gc_bits: u8,
     #[cfg(Py_GIL_DISABLED)]
@@ -41,7 +45,7 @@ fn raise_args_exception() {
     unsafe {
         let msg = "orjson.Fragment() takes exactly 1 positional argument";
         let err_msg =
-            PyUnicode_FromStringAndSize(msg.as_ptr() as *const c_char, msg.len() as isize);
+            PyUnicode_FromStringAndSize(msg.as_ptr().cast::<c_char>(), msg.len() as isize);
         PyErr_SetObject(PyExc_TypeError, err_msg);
         Py_DECREF(err_msg);
     };
@@ -68,7 +72,7 @@ pub unsafe extern "C" fn orjson_fragment_tp_new(
                 #[cfg(Py_GIL_DISABLED)]
                 _padding: 0,
                 #[cfg(Py_GIL_DISABLED)]
-                ob_mutex: PyMutex::new(),
+                ob_mutex: pyo3_ffi::PyMutex::new(),
                 #[cfg(Py_GIL_DISABLED)]
                 ob_gc_bits: 0,
                 #[cfg(Py_GIL_DISABLED)]
@@ -80,7 +84,7 @@ pub unsafe extern "C" fn orjson_fragment_tp_new(
                 ob_type: crate::typeref::FRAGMENT_TYPE,
                 contents: contents,
             });
-            Box::into_raw(obj) as *mut PyObject
+            Box::into_raw(obj).cast::<PyObject>()
         }
     }
 }
@@ -90,16 +94,18 @@ pub unsafe extern "C" fn orjson_fragment_tp_new(
 #[cfg_attr(feature = "optimize", optimize(size))]
 pub unsafe extern "C" fn orjson_fragment_dealloc(object: *mut PyObject) {
     unsafe {
-        Py_DECREF((*(object as *mut Fragment)).contents);
-        std::alloc::dealloc(object as *mut u8, std::alloc::Layout::new::<Fragment>());
+        Py_DECREF((*object.cast::<Fragment>()).contents);
+        std::alloc::dealloc(object.cast::<u8>(), core::alloc::Layout::new::<Fragment>());
     }
 }
 
 #[cfg(Py_GIL_DISABLED)]
-const FRAGMENT_TP_FLAGS: AtomicU64 = AtomicU64::new(Py_TPFLAGS_DEFAULT | Py_TPFLAGS_IMMUTABLETYPE);
+const FRAGMENT_TP_FLAGS: AtomicU64 =
+    AtomicU64::new(Py_TPFLAGS_DEFAULT | pyo3_ffi::Py_TPFLAGS_IMMUTABLETYPE);
 
 #[cfg(all(Py_3_10, not(Py_GIL_DISABLED)))]
-const FRAGMENT_TP_FLAGS: core::ffi::c_ulong = Py_TPFLAGS_DEFAULT | Py_TPFLAGS_IMMUTABLETYPE;
+const FRAGMENT_TP_FLAGS: core::ffi::c_ulong =
+    Py_TPFLAGS_DEFAULT | pyo3_ffi::Py_TPFLAGS_IMMUTABLETYPE;
 
 #[cfg(not(Py_3_10))]
 const FRAGMENT_TP_FLAGS: core::ffi::c_ulong = Py_TPFLAGS_DEFAULT;
@@ -117,7 +123,7 @@ pub unsafe extern "C" fn orjson_fragmenttype_new() -> *mut PyTypeObject {
                     #[cfg(Py_GIL_DISABLED)]
                     _padding: 0,
                     #[cfg(Py_GIL_DISABLED)]
-                    ob_mutex: PyMutex::new(),
+                    ob_mutex: pyo3_ffi::PyMutex::new(),
                     #[cfg(Py_GIL_DISABLED)]
                     ob_gc_bits: 0,
                     #[cfg(Py_GIL_DISABLED)]
@@ -128,11 +134,11 @@ pub unsafe extern "C" fn orjson_fragmenttype_new() -> *mut PyTypeObject {
                     ob_refcnt: pyo3_ffi::PyObjectObRefcnt { ob_refcnt: 0 },
                     #[cfg(not(Py_3_12))]
                     ob_refcnt: 0,
-                    ob_type: core::ptr::addr_of_mut!(PyType_Type),
+                    ob_type: &raw mut PyType_Type,
                 },
                 ob_size: 0,
             },
-            tp_name: "orjson.Fragment\0".as_ptr() as *const c_char,
+            tp_name: c"orjson.Fragment".as_ptr(),
             tp_basicsize: core::mem::size_of::<Fragment>() as isize,
             tp_itemsize: 0,
             tp_dealloc: Some(orjson_fragment_dealloc),
@@ -151,8 +157,6 @@ pub unsafe extern "C" fn orjson_fragmenttype_new() -> *mut PyTypeObject {
             tp_vectorcall: None,
             tp_version_tag: 0,
             tp_weaklist: null_mut(),
-            #[cfg(not(Py_3_9))]
-            tp_print: None,
             tp_vectorcall_offset: 0,
             tp_getattr: None,
             tp_setattr: None,
diff --git a/src/ffi/long.rs b/src/ffi/long.rs
index 7aa49a67..438143be 100644
--- a/src/ffi/long.rs
+++ b/src/ffi/long.rs
@@ -40,37 +40,37 @@ pub struct PyLongObject {
 #[cfg(Py_3_12)]
 #[inline(always)]
 pub fn pylong_is_unsigned(ptr: *mut pyo3_ffi::PyObject) -> bool {
-    unsafe { (*(ptr as *mut PyLongObject)).long_value.lv_tag & SIGN_MASK == 0 }
+    unsafe { (*ptr.cast::<PyLongObject>()).long_value.lv_tag & SIGN_MASK == 0 }
 }
 
 #[cfg(not(Py_3_12))]
 #[inline(always)]
 pub fn pylong_is_unsigned(ptr: *mut pyo3_ffi::PyObject) -> bool {
-    unsafe { (*(ptr as *mut pyo3_ffi::PyVarObject)).ob_size > 0 }
+    unsafe { (*ptr.cast::<pyo3_ffi::PyVarObject>()).ob_size > 0 }
 }
 
 #[cfg(all(Py_3_12, feature = "inline_int"))]
 #[inline(always)]
 pub fn pylong_fits_in_i32(ptr: *mut pyo3_ffi::PyObject) -> bool {
-    unsafe { (*(ptr as *mut PyLongObject)).long_value.lv_tag < (2 << NON_SIZE_BITS) }
+    unsafe { (*ptr.cast::<PyLongObject>()).long_value.lv_tag < (2 << NON_SIZE_BITS) }
 }
 
 #[cfg(all(not(Py_3_12), feature = "inline_int"))]
 #[inline(always)]
 pub fn pylong_fits_in_i32(ptr: *mut pyo3_ffi::PyObject) -> bool {
-    unsafe { isize::abs((*(ptr as *mut pyo3_ffi::PyVarObject)).ob_size) == 1 }
+    unsafe { isize::abs((*ptr.cast::<pyo3_ffi::PyVarObject>()).ob_size) == 1 }
 }
 
 #[cfg(all(Py_3_12, feature = "inline_int"))]
 #[inline(always)]
 pub fn pylong_is_zero(ptr: *mut pyo3_ffi::PyObject) -> bool {
-    unsafe { (*(ptr as *mut PyLongObject)).long_value.lv_tag & SIGN_MASK == SIGN_ZERO }
+    unsafe { (*ptr.cast::<PyLongObject>()).long_value.lv_tag & SIGN_MASK == SIGN_ZERO }
 }
 
 #[cfg(all(not(Py_3_12), feature = "inline_int"))]
 #[inline(always)]
 pub fn pylong_is_zero(ptr: *mut pyo3_ffi::PyObject) -> bool {
-    unsafe { (*(ptr as *mut pyo3_ffi::PyVarObject)).ob_size == 0 }
+    unsafe { (*ptr.cast::<pyo3_ffi::PyVarObject>()).ob_size == 0 }
 }
 
 #[cfg(all(Py_3_12, feature = "inline_int"))]
@@ -78,9 +78,9 @@ pub fn pylong_is_zero(ptr: *mut pyo3_ffi::PyObject) -> bool {
 pub fn pylong_get_inline_value(ptr: *mut pyo3_ffi::PyObject) -> i64 {
     unsafe {
         if pylong_is_unsigned(ptr) {
-            (*(ptr as *mut PyLongObject)).long_value.ob_digit as i64
+            i64::from((*ptr.cast::<PyLongObject>()).long_value.ob_digit)
         } else {
-            -((*(ptr as *mut PyLongObject)).long_value.ob_digit as i64)
+            -i64::from((*ptr.cast::<PyLongObject>()).long_value.ob_digit)
         }
     }
 }
@@ -89,7 +89,7 @@ pub fn pylong_get_inline_value(ptr: *mut pyo3_ffi::PyObject) -> i64 {
 #[inline(always)]
 pub fn pylong_get_inline_value(ptr: *mut pyo3_ffi::PyObject) -> i64 {
     unsafe {
-        (*(ptr as *mut pyo3_ffi::PyVarObject)).ob_size as i64
-            * (*(ptr as *mut PyLongObject)).ob_digit as i64
+        (*ptr.cast::<pyo3_ffi::PyVarObject>()).ob_size as i64
+            * i64::from((*ptr.cast::<PyLongObject>()).ob_digit)
     }
 }
diff --git a/src/lib.rs b/src/lib.rs
index c738e506..21f72d2c 100644
--- a/src/lib.rs
+++ b/src/lib.rs
@@ -3,18 +3,57 @@
 #![cfg_attr(feature = "avx512", feature(stdarch_x86_avx512, avx512_target_feature))]
 #![cfg_attr(feature = "intrinsics", feature(core_intrinsics))]
 #![cfg_attr(feature = "optimize", feature(optimize_attribute))]
-#![cfg_attr(feature = "unstable-simd", feature(portable_simd))]
-#![allow(internal_features)] // core_intrinsics
-#![allow(non_camel_case_types)]
-#![allow(static_mut_refs)]
-#![allow(unknown_lints)] // internal_features
-#![allow(unused_unsafe)]
+#![cfg_attr(feature = "generic_simd", feature(portable_simd))]
+#![allow(clippy::absolute_paths)]
+#![allow(clippy::allow_attributes)]
+#![allow(clippy::allow_attributes_without_reason)]
+#![allow(clippy::arbitrary_source_item_ordering)]
+#![allow(clippy::decimal_literal_representation)]
+#![allow(clippy::doc_markdown)]
+#![allow(clippy::explicit_iter_loop)]
+#![allow(clippy::if_not_else)]
+#![allow(clippy::implicit_return)]
+#![allow(clippy::inline_always)]
+#![allow(clippy::let_underscore_untyped)]
+#![allow(clippy::missing_assert_message)]
+#![allow(clippy::missing_docs_in_private_items)]
+#![allow(clippy::missing_inline_in_public_items)]
+#![allow(clippy::missing_panics_doc)]
 #![allow(clippy::missing_safety_doc)]
+#![allow(clippy::multiple_unsafe_ops_per_block)]
 #![allow(clippy::needless_lifetimes)]
+#![allow(clippy::question_mark_used)]
+#![allow(clippy::redundant_else)]
 #![allow(clippy::redundant_field_names)]
-#![allow(clippy::uninlined_format_args)] // MSRV 1.66
+#![allow(clippy::renamed_function_params)]
+#![allow(clippy::semicolon_outside_block)]
+#![allow(clippy::single_call_fn)]
+#![allow(clippy::undocumented_unsafe_blocks)]
+#![allow(clippy::unreachable)]
+#![allow(clippy::unreadable_literal)]
+#![allow(clippy::unusual_byte_groupings)]
+#![allow(clippy::unwrap_in_result)]
 #![allow(clippy::upper_case_acronyms)]
 #![allow(clippy::zero_prefixed_literal)]
+#![allow(internal_features)] // core_intrinsics
+#![allow(non_camel_case_types)]
+#![allow(static_mut_refs)]
+#![allow(unknown_lints)] // internal_features
+#![allow(unused_unsafe)]
+#![warn(clippy::ptr_arg)]
+#![warn(clippy::ptr_as_ptr)]
+#![warn(clippy::ptr_cast_constness)]
+#![warn(clippy::ptr_eq)]
+#![warn(clippy::redundant_allocation)]
+#![warn(clippy::redundant_clone)]
+#![warn(clippy::redundant_locals)]
+#![warn(clippy::redundant_slicing)]
+#![warn(clippy::size_of_ref)]
+#![warn(clippy::std_instead_of_core)]
+#![warn(clippy::trivially_copy_pass_by_ref)]
+#![warn(clippy::unnecessary_semicolon)]
+#![warn(clippy::unnecessary_wraps)]
+#![warn(clippy::zero_ptr)]
 
 #[cfg(feature = "unwind")]
 extern crate unwinding;
@@ -31,7 +70,15 @@ mod str;
 mod typeref;
 
 use core::ffi::{c_char, c_int, c_void};
-use pyo3_ffi::*;
+use pyo3_ffi::{
+    PyCFunction_NewEx, PyErr_SetObject, PyLong_AsLong, PyLong_FromLongLong, PyMethodDef,
+    PyMethodDefPointer, PyModuleDef, PyModuleDef_HEAD_INIT, PyModuleDef_Slot, PyObject,
+    PyTuple_GET_ITEM, PyTuple_New, PyTuple_SET_ITEM, PyUnicode_FromStringAndSize,
+    PyUnicode_InternFromString, PyVectorcall_NARGS, Py_DECREF, Py_SIZE, Py_ssize_t, METH_KEYWORDS,
+    METH_O,
+};
+
+use crate::util::{isize_to_usize, usize_to_isize};
 
 #[allow(unused_imports)]
 use core::ptr::{null, null_mut, NonNull};
@@ -39,32 +86,32 @@ use core::ptr::{null, null_mut, NonNull};
 #[cfg(Py_3_13)]
 macro_rules! add {
     ($mptr:expr, $name:expr, $obj:expr) => {
-        PyModule_Add($mptr, $name.as_ptr() as *const c_char, $obj);
+        pyo3_ffi::PyModule_Add($mptr, $name.as_ptr(), $obj);
     };
 }
 
 #[cfg(all(Py_3_10, not(Py_3_13)))]
 macro_rules! add {
     ($mptr:expr, $name:expr, $obj:expr) => {
-        PyModule_AddObjectRef($mptr, $name.as_ptr() as *const c_char, $obj);
+        pyo3_ffi::PyModule_AddObjectRef($mptr, $name.as_ptr(), $obj);
     };
 }
 
 #[cfg(not(Py_3_10))]
 macro_rules! add {
     ($mptr:expr, $name:expr, $obj:expr) => {
-        PyModule_AddObject($mptr, $name.as_ptr() as *const c_char, $obj);
+        pyo3_ffi::PyModule_AddObject($mptr, $name.as_ptr(), $obj);
     };
 }
 
 macro_rules! opt {
     ($mptr:expr, $name:expr, $opt:expr) => {
         #[cfg(all(not(target_os = "windows"), target_pointer_width = "64"))]
-        PyModule_AddIntConstant($mptr, $name.as_ptr() as *const c_char, $opt as i64);
+        pyo3_ffi::PyModule_AddIntConstant($mptr, $name.as_ptr(), i64::from($opt));
         #[cfg(all(not(target_os = "windows"), target_pointer_width = "32"))]
-        PyModule_AddIntConstant($mptr, $name.as_ptr() as *const c_char, $opt as i32);
+        pyo3_ffi::PyModule_AddIntConstant($mptr, $name.as_ptr(), $opt as i32);
         #[cfg(target_os = "windows")]
-        PyModule_AddIntConstant($mptr, $name.as_ptr() as *const c_char, $opt as i32);
+        pyo3_ffi::PyModule_AddIntConstant($mptr, $name.as_ptr(), $opt as i32);
     };
 }
 
@@ -75,19 +122,21 @@ macro_rules! opt {
 pub unsafe extern "C" fn orjson_init_exec(mptr: *mut PyObject) -> c_int {
     unsafe {
         typeref::init_typerefs();
+
         {
             let version = env!("CARGO_PKG_VERSION");
             let pyversion = PyUnicode_FromStringAndSize(
-                version.as_ptr() as *const c_char,
-                version.len() as isize,
+                version.as_ptr().cast::<c_char>(),
+                usize_to_isize(version.len()),
             );
-            add!(mptr, "__version__\0", pyversion);
+            add!(mptr, c"__version__", pyversion);
         }
+
         {
-            let dumps_doc = "dumps(obj, /, default=None, option=None)\n--\n\nSerialize Python objects to JSON.\0";
+            let dumps_doc = c"dumps(obj, /, default=None, option=None)\n--\n\nSerialize Python objects to JSON.";
 
             let wrapped_dumps = PyMethodDef {
-                ml_name: "dumps\0".as_ptr() as *const c_char,
+                ml_name: c"dumps".as_ptr(),
                 ml_meth: PyMethodDefPointer {
                     #[cfg(Py_3_10)]
                     PyCFunctionFastWithKeywords: dumps,
@@ -95,65 +144,57 @@ pub unsafe extern "C" fn orjson_init_exec(mptr: *mut PyObject) -> c_int {
                     _PyCFunctionFastWithKeywords: dumps,
                 },
                 ml_flags: pyo3_ffi::METH_FASTCALL | METH_KEYWORDS,
-                ml_doc: dumps_doc.as_ptr() as *const c_char,
+                ml_doc: dumps_doc.as_ptr(),
             };
 
             let func = PyCFunction_NewEx(
                 Box::into_raw(Box::new(wrapped_dumps)),
                 null_mut(),
-                PyUnicode_InternFromString("orjson\0".as_ptr() as *const c_char),
+                PyUnicode_InternFromString(c"orjson".as_ptr()),
             );
-            add!(mptr, "dumps\0", func);
+            add!(mptr, c"dumps", func);
         }
 
         {
-            let loads_doc = "loads(obj, /)\n--\n\nDeserialize JSON to Python objects.\0";
+            let loads_doc = c"loads(obj, /)\n--\n\nDeserialize JSON to Python objects.";
 
             let wrapped_loads = PyMethodDef {
-                ml_name: "loads\0".as_ptr() as *const c_char,
+                ml_name: c"loads".as_ptr(),
                 ml_meth: PyMethodDefPointer { PyCFunction: loads },
                 ml_flags: METH_O,
-                ml_doc: loads_doc.as_ptr() as *const c_char,
+                ml_doc: loads_doc.as_ptr(),
             };
             let func = PyCFunction_NewEx(
                 Box::into_raw(Box::new(wrapped_loads)),
                 null_mut(),
-                PyUnicode_InternFromString("orjson\0".as_ptr() as *const c_char),
+                PyUnicode_InternFromString(c"orjson".as_ptr()),
             );
-            add!(mptr, "loads\0", func);
+            add!(mptr, c"loads", func);
         }
 
-        add!(mptr, "Fragment\0", typeref::FRAGMENT_TYPE as *mut PyObject);
+        add!(mptr, c"Fragment", typeref::FRAGMENT_TYPE.cast::<PyObject>());
 
-        opt!(mptr, "OPT_APPEND_NEWLINE\0", opt::APPEND_NEWLINE);
-        opt!(mptr, "OPT_INDENT_2\0", opt::INDENT_2);
-        opt!(mptr, "OPT_NAIVE_UTC\0", opt::NAIVE_UTC);
-        opt!(mptr, "OPT_NON_STR_KEYS\0", opt::NON_STR_KEYS);
-        opt!(mptr, "OPT_OMIT_MICROSECONDS\0", opt::OMIT_MICROSECONDS);
+        opt!(mptr, c"OPT_APPEND_NEWLINE", opt::APPEND_NEWLINE);
+        opt!(mptr, c"OPT_INDENT_2", opt::INDENT_2);
+        opt!(mptr, c"OPT_NAIVE_UTC", opt::NAIVE_UTC);
+        opt!(mptr, c"OPT_NON_STR_KEYS", opt::NON_STR_KEYS);
+        opt!(mptr, c"OPT_OMIT_MICROSECONDS", opt::OMIT_MICROSECONDS);
         opt!(
             mptr,
-            "OPT_PASSTHROUGH_DATACLASS\0",
+            c"OPT_PASSTHROUGH_DATACLASS",
             opt::PASSTHROUGH_DATACLASS
         );
-        opt!(
-            mptr,
-            "OPT_PASSTHROUGH_DATETIME\0",
-            opt::PASSTHROUGH_DATETIME
-        );
-        opt!(
-            mptr,
-            "OPT_PASSTHROUGH_SUBCLASS\0",
-            opt::PASSTHROUGH_SUBCLASS
-        );
-        opt!(mptr, "OPT_SERIALIZE_DATACLASS\0", opt::SERIALIZE_DATACLASS);
-        opt!(mptr, "OPT_SERIALIZE_NUMPY\0", opt::SERIALIZE_NUMPY);
-        opt!(mptr, "OPT_SERIALIZE_UUID\0", opt::SERIALIZE_UUID);
-        opt!(mptr, "OPT_SORT_KEYS\0", opt::SORT_KEYS);
-        opt!(mptr, "OPT_STRICT_INTEGER\0", opt::STRICT_INTEGER);
-        opt!(mptr, "OPT_UTC_Z\0", opt::UTC_Z);
-
-        add!(mptr, "JSONDecodeError\0", typeref::JsonDecodeError);
-        add!(mptr, "JSONEncodeError\0", typeref::JsonEncodeError);
+        opt!(mptr, c"OPT_PASSTHROUGH_DATETIME", opt::PASSTHROUGH_DATETIME);
+        opt!(mptr, c"OPT_PASSTHROUGH_SUBCLASS", opt::PASSTHROUGH_SUBCLASS);
+        opt!(mptr, c"OPT_SERIALIZE_DATACLASS", opt::SERIALIZE_DATACLASS);
+        opt!(mptr, c"OPT_SERIALIZE_NUMPY", opt::SERIALIZE_NUMPY);
+        opt!(mptr, c"OPT_SERIALIZE_UUID", opt::SERIALIZE_UUID);
+        opt!(mptr, c"OPT_SORT_KEYS", opt::SORT_KEYS);
+        opt!(mptr, c"OPT_STRICT_INTEGER", opt::STRICT_INTEGER);
+        opt!(mptr, c"OPT_UTC_Z", opt::UTC_Z);
+
+        add!(mptr, c"JSONDecodeError", typeref::JsonDecodeError);
+        add!(mptr, c"JSONEncodeError", typeref::JsonEncodeError);
 
         0
     }
@@ -174,18 +215,18 @@ pub unsafe extern "C" fn PyInit_orjson() -> *mut PyModuleDef {
     unsafe {
         let mod_slots: Box<[PyModuleDef_Slot; PYMODULEDEF_LEN]> = Box::new([
             PyModuleDef_Slot {
-                slot: Py_mod_exec,
+                slot: pyo3_ffi::Py_mod_exec,
                 value: orjson_init_exec as *mut c_void,
             },
             #[cfg(Py_3_12)]
             PyModuleDef_Slot {
-                slot: Py_mod_multiple_interpreters,
-                value: Py_MOD_MULTIPLE_INTERPRETERS_NOT_SUPPORTED,
+                slot: pyo3_ffi::Py_mod_multiple_interpreters,
+                value: pyo3_ffi::Py_MOD_MULTIPLE_INTERPRETERS_NOT_SUPPORTED,
             },
             #[cfg(Py_3_13)]
             PyModuleDef_Slot {
-                slot: Py_mod_gil,
-                value: Py_MOD_GIL_USED,
+                slot: pyo3_ffi::Py_mod_gil,
+                value: pyo3_ffi::Py_MOD_GIL_USED,
             },
             PyModuleDef_Slot {
                 slot: 0,
@@ -195,11 +236,11 @@ pub unsafe extern "C" fn PyInit_orjson() -> *mut PyModuleDef {
 
         let init = Box::new(PyModuleDef {
             m_base: PyModuleDef_HEAD_INIT,
-            m_name: "orjson\0".as_ptr() as *const c_char,
+            m_name: c"orjson".as_ptr(),
             m_doc: null(),
             m_size: 0,
             m_methods: null_mut(),
-            m_slots: Box::into_raw(mod_slots) as *mut PyModuleDef_Slot,
+            m_slots: Box::into_raw(mod_slots).cast::<PyModuleDef_Slot>(),
             m_traverse: None,
             m_clear: None,
             m_free: None,
@@ -214,19 +255,20 @@ pub unsafe extern "C" fn PyInit_orjson() -> *mut PyModuleDef {
 #[inline(never)]
 #[cfg_attr(feature = "optimize", optimize(size))]
 fn raise_loads_exception(err: deserialize::DeserializeError) -> *mut PyObject {
-    let pos = err.pos();
-    let msg = err.message;
-    let doc = match err.data {
-        Some(as_str) => unsafe {
-            PyUnicode_FromStringAndSize(as_str.as_ptr() as *const c_char, as_str.len() as isize)
-        },
-        None => {
-            use_immortal!(crate::typeref::EMPTY_UNICODE)
-        }
-    };
     unsafe {
+        let pos = err.pos();
+        let msg = err.message;
+        let doc = match err.data {
+            Some(as_str) => PyUnicode_FromStringAndSize(
+                as_str.as_ptr().cast::<c_char>(),
+                usize_to_isize(as_str.len()),
+            ),
+            None => {
+                use_immortal!(crate::typeref::EMPTY_UNICODE)
+            }
+        };
         let err_msg =
-            PyUnicode_FromStringAndSize(msg.as_ptr() as *const c_char, msg.len() as isize);
+            PyUnicode_FromStringAndSize(msg.as_ptr().cast::<c_char>(), usize_to_isize(msg.len()));
         let args = PyTuple_New(3);
         let pos = PyLong_FromLongLong(pos);
         PyTuple_SET_ITEM(args, 0, err_msg);
@@ -235,7 +277,7 @@ fn raise_loads_exception(err: deserialize::DeserializeError) -> *mut PyObject {
         PyErr_SetObject(typeref::JsonDecodeError, args);
         debug_assert!(ffi!(Py_REFCNT(args)) <= 2);
         Py_DECREF(args);
-    };
+    }
     null_mut()
 }
 
@@ -245,11 +287,11 @@ fn raise_loads_exception(err: deserialize::DeserializeError) -> *mut PyObject {
 fn raise_dumps_exception_fixed(msg: &str) -> *mut PyObject {
     unsafe {
         let err_msg =
-            PyUnicode_FromStringAndSize(msg.as_ptr() as *const c_char, msg.len() as isize);
+            PyUnicode_FromStringAndSize(msg.as_ptr().cast::<c_char>(), usize_to_isize(msg.len()));
         PyErr_SetObject(typeref::JsonEncodeError, err_msg);
         debug_assert!(ffi!(Py_REFCNT(err_msg)) <= 2);
         Py_DECREF(err_msg);
-    };
+    }
     null_mut()
 }
 
@@ -259,20 +301,20 @@ fn raise_dumps_exception_fixed(msg: &str) -> *mut PyObject {
 #[cfg(Py_3_12)]
 fn raise_dumps_exception_dynamic(err: &str) -> *mut PyObject {
     unsafe {
-        let cause_exc: *mut PyObject = PyErr_GetRaisedException();
+        let cause_exc: *mut PyObject = pyo3_ffi::PyErr_GetRaisedException();
 
         let err_msg =
-            PyUnicode_FromStringAndSize(err.as_ptr() as *const c_char, err.len() as isize);
+            PyUnicode_FromStringAndSize(err.as_ptr().cast::<c_char>(), usize_to_isize(err.len()));
         PyErr_SetObject(typeref::JsonEncodeError, err_msg);
         debug_assert!(ffi!(Py_REFCNT(err_msg)) <= 2);
         Py_DECREF(err_msg);
 
         if !cause_exc.is_null() {
-            let exc: *mut PyObject = PyErr_GetRaisedException();
-            PyException_SetCause(exc, cause_exc);
-            PyErr_SetRaisedException(exc);
+            let exc: *mut PyObject = pyo3_ffi::PyErr_GetRaisedException();
+            pyo3_ffi::PyException_SetCause(exc, cause_exc);
+            pyo3_ffi::PyErr_SetRaisedException(exc);
         }
-    };
+    }
     null_mut()
 }
 
@@ -285,30 +327,30 @@ fn raise_dumps_exception_dynamic(err: &str) -> *mut PyObject {
         let mut cause_tp: *mut PyObject = null_mut();
         let mut cause_val: *mut PyObject = null_mut();
         let mut cause_traceback: *mut PyObject = null_mut();
-        PyErr_Fetch(&mut cause_tp, &mut cause_val, &mut cause_traceback);
+        pyo3_ffi::PyErr_Fetch(&mut cause_tp, &mut cause_val, &mut cause_traceback);
 
         let err_msg =
-            PyUnicode_FromStringAndSize(err.as_ptr() as *const c_char, err.len() as isize);
+            PyUnicode_FromStringAndSize(err.as_ptr().cast::<c_char>(), usize_to_isize(err.len()));
         PyErr_SetObject(typeref::JsonEncodeError, err_msg);
         debug_assert!(ffi!(Py_REFCNT(err_msg)) == 2);
         Py_DECREF(err_msg);
         let mut tp: *mut PyObject = null_mut();
         let mut val: *mut PyObject = null_mut();
         let mut traceback: *mut PyObject = null_mut();
-        PyErr_Fetch(&mut tp, &mut val, &mut traceback);
-        PyErr_NormalizeException(&mut tp, &mut val, &mut traceback);
+        pyo3_ffi::PyErr_Fetch(&mut tp, &mut val, &mut traceback);
+        pyo3_ffi::PyErr_NormalizeException(&mut tp, &mut val, &mut traceback);
 
         if !cause_tp.is_null() {
-            PyErr_NormalizeException(&mut cause_tp, &mut cause_val, &mut cause_traceback);
-            PyException_SetCause(val, cause_val);
+            pyo3_ffi::PyErr_NormalizeException(&mut cause_tp, &mut cause_val, &mut cause_traceback);
+            pyo3_ffi::PyException_SetCause(val, cause_val);
             Py_DECREF(cause_tp);
         }
         if !cause_traceback.is_null() {
             Py_DECREF(cause_traceback);
         }
 
-        PyErr_Restore(tp, val, traceback);
-    };
+        pyo3_ffi::PyErr_Restore(tp, val, traceback);
+    }
     null_mut()
 }
 
@@ -331,7 +373,7 @@ pub unsafe extern "C" fn dumps(
         let mut default: Option<NonNull<PyObject>> = None;
         let mut optsptr: Option<NonNull<PyObject>> = None;
 
-        let num_args = PyVectorcall_NARGS(nargs as usize);
+        let num_args = PyVectorcall_NARGS(isize_to_usize(nargs));
         if unlikely!(num_args == 0) {
             return raise_dumps_exception_fixed(
                 "dumps() missing 1 required positional argument: 'obj'",
@@ -346,14 +388,14 @@ pub unsafe extern "C" fn dumps(
         if unlikely!(!kwnames.is_null()) {
             for i in 0..=Py_SIZE(kwnames).saturating_sub(1) {
                 let arg = PyTuple_GET_ITEM(kwnames, i as Py_ssize_t);
-                if arg == typeref::DEFAULT {
+                if core::ptr::eq(arg, typeref::DEFAULT) {
                     if unlikely!(num_args & 2 == 2) {
                         return raise_dumps_exception_fixed(
                             "dumps() got multiple values for argument: 'default'",
                         );
                     }
                     default = Some(NonNull::new_unchecked(*args.offset(num_args + i)));
-                } else if arg == typeref::OPTION {
+                } else if core::ptr::eq(arg, typeref::OPTION) {
                     if unlikely!(num_args & 3 == 3) {
                         return raise_dumps_exception_fixed(
                             "dumps() got multiple values for argument: 'option'",
@@ -371,16 +413,19 @@ pub unsafe extern "C" fn dumps(
         let mut optsbits: i32 = 0;
         if unlikely!(optsptr.is_some()) {
             let opts = optsptr.unwrap();
-            if (*opts.as_ptr()).ob_type == typeref::INT_TYPE {
-                optsbits = PyLong_AsLong(optsptr.unwrap().as_ptr()) as i32;
+            if core::ptr::eq((*opts.as_ptr()).ob_type, typeref::INT_TYPE) {
+                #[allow(clippy::cast_possible_truncation)]
+                let tmp = PyLong_AsLong(optsptr.unwrap().as_ptr()) as i32; // stmt_expr_attributes
+                optsbits = tmp;
                 if unlikely!(!(0..=opt::MAX_OPT).contains(&optsbits)) {
                     return raise_dumps_exception_fixed("Invalid opts");
                 }
-            } else if unlikely!(opts.as_ptr() != typeref::NONE) {
+            } else if unlikely!(!core::ptr::eq(opts.as_ptr(), typeref::NONE)) {
                 return raise_dumps_exception_fixed("Invalid opts");
             }
         }
 
+        #[allow(clippy::cast_sign_loss)]
         match crate::serialize::serialize(*args, default, optsbits as opt::Opt) {
             Ok(val) => val.as_ptr(),
             Err(err) => raise_dumps_exception_dynamic(err.as_str()),
diff --git a/src/opt.rs b/src/opt.rs
index 4ceff66b..fb0930b0 100644
--- a/src/opt.rs
+++ b/src/opt.rs
@@ -24,6 +24,7 @@ pub const SORT_OR_NON_STR_KEYS: Opt = SORT_KEYS | NON_STR_KEYS;
 pub const NOT_PASSTHROUGH: Opt =
     !(PASSTHROUGH_DATETIME | PASSTHROUGH_DATACLASS | PASSTHROUGH_SUBCLASS);
 
+#[allow(clippy::cast_possible_wrap)]
 pub const MAX_OPT: i32 = (APPEND_NEWLINE
     | INDENT_2
     | NAIVE_UTC
diff --git a/src/serialize/buffer.rs b/src/serialize/buffer.rs
index 894ee837..8b2bf73d 100644
--- a/src/serialize/buffer.rs
+++ b/src/serialize/buffer.rs
@@ -23,7 +23,7 @@ impl SmallFixedBuffer {
     pub unsafe fn as_mut_slice(&mut self) -> &mut [u8] {
         unsafe {
             core::slice::from_raw_parts_mut(
-                (core::ptr::addr_of_mut!(self.bytes) as *mut u8).add(self.idx),
+                (&raw mut self.bytes).cast::<u8>().add(self.idx),
                 BUFFER_LENGTH - self.idx,
             )
         }
@@ -39,10 +39,7 @@ impl SmallFixedBuffer {
     pub fn push(&mut self, value: u8) {
         debug_assert!(self.idx + 1 < BUFFER_LENGTH);
         unsafe {
-            core::ptr::write(
-                (core::ptr::addr_of_mut!(self.bytes) as *mut u8).add(self.idx),
-                value,
-            );
+            core::ptr::write((&raw mut self.bytes).cast::<u8>().add(self.idx), value);
             self.idx += 1;
         };
     }
@@ -53,7 +50,7 @@ impl SmallFixedBuffer {
         unsafe {
             core::ptr::copy_nonoverlapping(
                 slice.as_ptr(),
-                (core::ptr::addr_of_mut!(self.bytes) as *mut u8).add(self.idx),
+                (&raw mut self.bytes).cast::<u8>().add(self.idx),
                 slice.len(),
             );
             self.idx += slice.len();
@@ -62,7 +59,7 @@ impl SmallFixedBuffer {
 
     #[inline]
     pub fn as_ptr(&self) -> *const u8 {
-        core::ptr::addr_of!(self.bytes) as *const u8
+        (&raw const self.bytes).cast::<u8>()
     }
 
     #[inline]
diff --git a/src/serialize/error.rs b/src/serialize/error.rs
index f1329a30..49b6d626 100644
--- a/src/serialize/error.rs
+++ b/src/serialize/error.rs
@@ -22,10 +22,10 @@ pub enum SerializeError {
     UnsupportedType(NonNull<pyo3_ffi::PyObject>),
 }
 
-impl std::fmt::Display for SerializeError {
+impl core::fmt::Display for SerializeError {
     #[cold]
     #[cfg_attr(feature = "optimize", optimize(size))]
-    fn fmt(&self, f: &mut std::fmt::Formatter) -> std::fmt::Result {
+    fn fmt(&self, f: &mut core::fmt::Formatter) -> core::fmt::Result {
         match *self {
             SerializeError::DatetimeLibraryUnsupported => write!(
                 f,
@@ -63,7 +63,7 @@ impl std::fmt::Display for SerializeError {
             SerializeError::UnsupportedType(ptr) => {
                 let name =
                     unsafe { CStr::from_ptr((*ob_type!(ptr.as_ptr())).tp_name).to_string_lossy() };
-                write!(f, "Type is not JSON serializable: {}", name)
+                write!(f, "Type is not JSON serializable: {name}")
             }
         }
     }
diff --git a/src/serialize/mod.rs b/src/serialize/mod.rs
index f044bcd2..a0b03357 100644
--- a/src/serialize/mod.rs
+++ b/src/serialize/mod.rs
@@ -6,6 +6,6 @@ mod obtype;
 mod per_type;
 mod serializer;
 mod state;
-mod writer;
+pub mod writer;
 
 pub use serializer::serialize;
diff --git a/src/serialize/per_type/dataclass.rs b/src/serialize/per_type/dataclass.rs
index a5dc707a..370ee9ad 100644
--- a/src/serialize/per_type/dataclass.rs
+++ b/src/serialize/per_type/dataclass.rs
@@ -8,6 +8,7 @@ use crate::str::unicode_to_str;
 use crate::typeref::{
     DATACLASS_FIELDS_STR, DICT_STR, FIELD_TYPE, FIELD_TYPE_STR, SLOTS_STR, STR_TYPE,
 };
+use crate::util::isize_to_usize;
 
 use serde::ser::{Serialize, SerializeMap, Serializer};
 
@@ -88,7 +89,7 @@ impl Serialize for DataclassFastSerializer {
     where
         S: Serializer,
     {
-        let len = ffi!(Py_SIZE(self.ptr)) as usize;
+        let len = isize_to_usize(ffi!(Py_SIZE(self.ptr)));
         if unlikely!(len == 0) {
             return ZeroDictSerializer::new().serialize(serializer);
         }
@@ -100,7 +101,7 @@ impl Serialize for DataclassFastSerializer {
 
         pydict_next!(self.ptr, &mut pos, &mut next_key, &mut next_value);
 
-        for _ in 0..ffi!(Py_SIZE(self.ptr)) as usize {
+        for _ in 0..len {
             let key = next_key;
             let value = next_value;
 
@@ -114,7 +115,7 @@ impl Serialize for DataclassFastSerializer {
                 let tmp = unicode_to_str(key);
                 if unlikely!(tmp.is_none()) {
                     err!(SerializeError::InvalidStr)
-                };
+                }
                 tmp.unwrap()
             };
             if unlikely!(key_as_str.as_bytes()[0] == b'_') {
@@ -158,7 +159,7 @@ impl Serialize for DataclassFallbackSerializer {
         let fields = ffi!(PyObject_GetAttr(self.ptr, DATACLASS_FIELDS_STR));
         debug_assert!(ffi!(Py_REFCNT(fields)) >= 2);
         ffi!(Py_DECREF(fields));
-        let len = ffi!(Py_SIZE(fields)) as usize;
+        let len = isize_to_usize(ffi!(Py_SIZE(fields)));
         if unlikely!(len == 0) {
             return ZeroDictSerializer::new().serialize(serializer);
         }
@@ -170,7 +171,7 @@ impl Serialize for DataclassFallbackSerializer {
 
         pydict_next!(fields, &mut pos, &mut next_key, &mut next_value);
 
-        for _ in 0..ffi!(Py_SIZE(fields)) as usize {
+        for _ in 0..len {
             let attr = next_key;
             let field = next_value;
 
@@ -179,7 +180,7 @@ impl Serialize for DataclassFallbackSerializer {
             let field_type = ffi!(PyObject_GetAttr(field, FIELD_TYPE_STR));
             debug_assert!(ffi!(Py_REFCNT(field_type)) >= 2);
             ffi!(Py_DECREF(field_type));
-            if unsafe { field_type as *mut pyo3_ffi::PyTypeObject != FIELD_TYPE } {
+            if unsafe { !core::ptr::eq(field_type.cast::<pyo3_ffi::PyTypeObject>(), FIELD_TYPE) } {
                 continue;
             }
 
@@ -187,7 +188,7 @@ impl Serialize for DataclassFallbackSerializer {
                 let tmp = unicode_to_str(attr);
                 if unlikely!(tmp.is_none()) {
                     err!(SerializeError::InvalidStr)
-                };
+                }
                 tmp.unwrap()
             };
             if key_as_str.as_bytes()[0] == b'_' {
@@ -200,7 +201,7 @@ impl Serialize for DataclassFallbackSerializer {
             let pyvalue = PyObjectSerializer::new(value, self.state, self.default);
 
             map.serialize_key(key_as_str).unwrap();
-            map.serialize_value(&pyvalue)?
+            map.serialize_value(&pyvalue)?;
         }
         map.end()
     }
diff --git a/src/serialize/per_type/datetime.rs b/src/serialize/per_type/datetime.rs
index 634c40fc..0cad0ad9 100644
--- a/src/serialize/per_type/datetime.rs
+++ b/src/serialize/per_type/datetime.rs
@@ -1,12 +1,12 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use crate::opt::*;
+use crate::opt::{Opt, OMIT_MICROSECONDS};
 use crate::serialize::buffer::SmallFixedBuffer;
 use crate::serialize::error::SerializeError;
 use crate::serialize::per_type::datetimelike::{DateTimeError, DateTimeLike, Offset};
-#[cfg(Py_3_9)]
-use crate::typeref::ZONEINFO_TYPE;
-use crate::typeref::{CONVERT_METHOD_STR, DST_STR, NORMALIZE_METHOD_STR, UTCOFFSET_METHOD_STR};
+use crate::typeref::{
+    CONVERT_METHOD_STR, DST_STR, NORMALIZE_METHOD_STR, UTCOFFSET_METHOD_STR, ZONEINFO_TYPE,
+};
 use serde::ser::{Serialize, Serializer};
 
 macro_rules! write_double_digit {
@@ -55,13 +55,19 @@ impl Date {
         }
         buf.push(b'-');
         {
-            let month = ffi!(PyDateTime_GET_MONTH(self.ptr)) as u32;
-            write_double_digit!(buf, month);
+            let val_py = ffi!(PyDateTime_GET_MONTH(self.ptr));
+            debug_assert!(val_py >= 0);
+            #[allow(clippy::cast_sign_loss)]
+            let val = val_py as u32;
+            write_double_digit!(buf, val);
         }
         buf.push(b'-');
         {
-            let day = ffi!(PyDateTime_GET_DAY(self.ptr)) as u32;
-            write_double_digit!(buf, day);
+            let val_py = ffi!(PyDateTime_GET_DAY(self.ptr));
+            debug_assert!(val_py >= 0);
+            #[allow(clippy::cast_sign_loss)]
+            let val = val_py as u32;
+            write_double_digit!(buf, val);
         }
     }
 }
@@ -95,7 +101,7 @@ impl Time {
 
     #[inline(never)]
     pub fn write_buf(&self, buf: &mut SmallFixedBuffer) -> Result<(), TimeError> {
-        if unsafe { (*(self.ptr as *mut pyo3_ffi::PyDateTime_Time)).hastzinfo == 1 } {
+        if unsafe { (*self.ptr.cast::<pyo3_ffi::PyDateTime_Time>()).hastzinfo == 1 } {
             return Err(TimeError::HasTimezone);
         }
         let hour = ffi!(PyDateTime_TIME_GET_HOUR(self.ptr)) as u8;
@@ -122,7 +128,7 @@ impl Serialize for Time {
         let mut buf = SmallFixedBuffer::new();
         if self.write_buf(&mut buf).is_err() {
             err!(SerializeError::DatetimeLibraryUnsupported)
-        };
+        }
         serializer.serialize_unit_struct(str_from_slice!(buf.as_ptr(), buf.len()))
     }
 }
@@ -144,7 +150,11 @@ impl DateTime {
 macro_rules! pydatetime_get {
     ($fn: ident, $pyfn: ident, $ty: ident) => {
         fn $fn(&self) -> $ty {
-            ffi!($pyfn(self.ptr)) as $ty
+            let ret = ffi!($pyfn(self.ptr));
+            debug_assert!(ret >= 0);
+            #[allow(clippy::cast_sign_loss)]
+            let ret2 = ret as $ty; // stmt_expr_attributes
+            ret2
         }
     };
 }
@@ -163,9 +173,10 @@ impl DateTimeLike for DateTime {
     }
 
     fn has_tz(&self) -> bool {
-        unsafe { (*(self.ptr as *mut pyo3_ffi::PyDateTime_DateTime)).hastzinfo == 1 }
+        unsafe { (*(self.ptr.cast::<pyo3_ffi::PyDateTime_DateTime>())).hastzinfo == 1 }
     }
 
+    #[inline(never)]
     fn slow_offset(&self) -> Result<Offset, DateTimeError> {
         let tzinfo = ffi!(PyDateTime_DATE_GET_TZINFO(self.ptr));
         if ffi!(PyObject_HasAttr(tzinfo, CONVERT_METHOD_STR)) == 1 {
@@ -202,13 +213,13 @@ impl DateTimeLike for DateTime {
         }
     }
 
-    #[cfg(Py_3_9)]
+    #[inline]
     fn offset(&self) -> Result<Offset, DateTimeError> {
         if !self.has_tz() {
             Ok(Offset::default())
         } else {
             let tzinfo = ffi!(PyDateTime_DATE_GET_TZINFO(self.ptr));
-            if unsafe { ob_type!(tzinfo) == ZONEINFO_TYPE } {
+            if unsafe { core::ptr::eq(ob_type!(tzinfo), ZONEINFO_TYPE) } {
                 // zoneinfo
                 let py_offset = call_method!(tzinfo, UTCOFFSET_METHOD_STR, self.ptr);
                 let offset = Offset {
@@ -222,15 +233,6 @@ impl DateTimeLike for DateTime {
             }
         }
     }
-
-    #[cfg(not(Py_3_9))]
-    fn offset(&self) -> Result<Offset, DateTimeError> {
-        if !self.has_tz() {
-            Ok(Offset::default())
-        } else {
-            self.slow_offset()
-        }
-    }
 }
 
 impl Serialize for DateTime {
diff --git a/src/serialize/per_type/datetimelike.rs b/src/serialize/per_type/datetimelike.rs
index 2e24479b..17c3d144 100644
--- a/src/serialize/per_type/datetimelike.rs
+++ b/src/serialize/per_type/datetimelike.rs
@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use crate::opt::*;
+use crate::opt::{Opt, NAIVE_UTC, OMIT_MICROSECONDS, UTC_Z};
 
 use crate::serialize::buffer::SmallFixedBuffer;
 
@@ -60,7 +60,7 @@ pub trait DateTimeLike {
     /// Is the object time-zone aware?
     fn has_tz(&self) -> bool;
 
-    //// python3.8 or below implementation of offset()
+    //// Non-zoneinfo implementation of offset()
     fn slow_offset(&self) -> Result<Offset, DateTimeError>;
 
     /// The offset of the timezone.
diff --git a/src/serialize/per_type/default.rs b/src/serialize/per_type/default.rs
index 6c9e1d47..519dc6c4 100644
--- a/src/serialize/per_type/default.rs
+++ b/src/serialize/per_type/default.rs
@@ -35,11 +35,14 @@ impl<'a> Serialize for DefaultSerializer<'a> {
                     core::ptr::null_mut() as *mut pyo3_ffi::PyObject
                 ));
                 #[cfg(Py_3_10)]
+                #[allow(clippy::cast_sign_loss)]
+                let nargs = ffi!(PyVectorcall_NARGS(1)) as usize;
+                #[cfg(Py_3_10)]
                 let default_obj = unsafe {
                     pyo3_ffi::PyObject_Vectorcall(
                         callable.as_ptr(),
-                        core::ptr::addr_of!(self.previous.ptr),
-                        pyo3_ffi::PyVectorcall_NARGS(1) as usize,
+                        &raw const self.previous.ptr,
+                        nargs,
                         core::ptr::null_mut(),
                     )
                 };
diff --git a/src/serialize/per_type/dict.rs b/src/serialize/per_type/dict.rs
index b5e6cb3b..e2e04f62 100644
--- a/src/serialize/per_type/dict.rs
+++ b/src/serialize/per_type/dict.rs
@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use crate::opt::*;
+use crate::opt::{NON_STR_KEYS, NOT_PASSTHROUGH, SORT_KEYS, SORT_OR_NON_STR_KEYS};
 use crate::serialize::buffer::SmallFixedBuffer;
 use crate::serialize::error::SerializeError;
 use crate::serialize::obtype::{pyobject_to_obtype, ObType};
@@ -15,6 +15,7 @@ use crate::serialize::serializer::PyObjectSerializer;
 use crate::serialize::state::SerializerState;
 use crate::str::{unicode_to_str, unicode_to_str_via_ffi};
 use crate::typeref::{STR_TYPE, TRUE, VALUE_STR};
+use crate::util::isize_to_usize;
 use compact_str::CompactString;
 use core::ptr::NonNull;
 use serde::ser::{Serialize, SerializeMap, Serializer};
@@ -73,16 +74,17 @@ impl Serialize for DictGenericSerializer {
             ZeroDictSerializer::new().serialize(serializer)
         } else if likely!(opt_disabled!(self.state.opts(), SORT_OR_NON_STR_KEYS)) {
             unsafe {
-                core::mem::transmute::<&DictGenericSerializer, &Dict>(self).serialize(serializer)
+                (*(core::ptr::from_ref::<DictGenericSerializer>(self)).cast::<Dict>())
+                    .serialize(serializer)
             }
         } else if opt_enabled!(self.state.opts(), NON_STR_KEYS) {
             unsafe {
-                core::mem::transmute::<&DictGenericSerializer, &DictNonStrKey>(self)
+                (*(core::ptr::from_ref::<DictGenericSerializer>(self)).cast::<DictNonStrKey>())
                     .serialize(serializer)
             }
         } else {
             unsafe {
-                core::mem::transmute::<&DictGenericSerializer, &DictSortedKey>(self)
+                (*(core::ptr::from_ref::<DictGenericSerializer>(self)).cast::<DictSortedKey>())
                     .serialize(serializer)
             }
         }
@@ -223,7 +225,7 @@ impl Serialize for Dict {
 
         let mut map = serializer.serialize_map(None).unwrap();
 
-        let len = ffi!(Py_SIZE(self.ptr)) as usize;
+        let len = isize_to_usize(ffi!(Py_SIZE(self.ptr)));
         assume!(len > 0);
 
         for _ in 0..len {
@@ -241,7 +243,7 @@ impl Serialize for Dict {
                 let tmp = unicode_to_str(key);
                 if unlikely!(tmp.is_none()) {
                     err!(SerializeError::InvalidStr)
-                };
+                }
                 tmp.unwrap()
             };
 
@@ -271,7 +273,7 @@ impl Serialize for DictSortedKey {
 
         pydict_next!(self.ptr, &mut pos, &mut next_key, &mut next_value);
 
-        let len = ffi!(Py_SIZE(self.ptr)) as usize;
+        let len = isize_to_usize(ffi!(Py_SIZE(self.ptr)));
         assume!(len > 0);
 
         let mut items: SmallVec<[(&str, *mut pyo3_ffi::PyObject); 8]> =
@@ -283,7 +285,7 @@ impl Serialize for DictSortedKey {
 
             pydict_next!(self.ptr, &mut pos, &mut next_key, &mut next_value);
 
-            if unlikely!(unsafe { ob_type!(key) != STR_TYPE }) {
+            if unlikely!(unsafe { !core::ptr::eq(ob_type!(key), STR_TYPE) }) {
                 err!(SerializeError::KeyMustBeStr)
             }
             let data = unicode_to_str(key);
@@ -327,6 +329,7 @@ fn non_str_str_subclass(key: *mut pyo3_ffi::PyObject) -> Result<CompactString, S
     }
 }
 
+#[allow(clippy::unnecessary_wraps)]
 #[inline(never)]
 fn non_str_date(key: *mut pyo3_ffi::PyObject) -> Result<CompactString, SerializeError> {
     let mut buf = SmallFixedBuffer::new();
@@ -364,6 +367,7 @@ fn non_str_time(
     Ok(CompactString::from(key_as_str))
 }
 
+#[allow(clippy::unnecessary_wraps)]
 #[inline(never)]
 fn non_str_uuid(key: *mut pyo3_ffi::PyObject) -> Result<CompactString, SerializeError> {
     let mut buf = SmallFixedBuffer::new();
@@ -372,6 +376,7 @@ fn non_str_uuid(key: *mut pyo3_ffi::PyObject) -> Result<CompactString, Serialize
     Ok(CompactString::from(key_as_str))
 }
 
+#[allow(clippy::unnecessary_wraps)]
 #[cold]
 #[inline(never)]
 fn non_str_float(key: *mut pyo3_ffi::PyObject) -> Result<CompactString, SerializeError> {
@@ -383,6 +388,7 @@ fn non_str_float(key: *mut pyo3_ffi::PyObject) -> Result<CompactString, Serializ
     }
 }
 
+#[allow(clippy::unnecessary_wraps)]
 #[inline(never)]
 fn non_str_int(key: *mut pyo3_ffi::PyObject) -> Result<CompactString, SerializeError> {
     let ival = ffi!(PyLong_AsLongLong(key));
@@ -417,7 +423,7 @@ impl DictNonStrKey {
         match pyobject_to_obtype(key, opts) {
             ObType::None => Ok(CompactString::const_new("null")),
             ObType::Bool => {
-                if unsafe { key == TRUE } {
+                if unsafe { core::ptr::eq(key, TRUE) } {
                     Ok(CompactString::const_new("true"))
                 } else {
                     Ok(CompactString::const_new("false"))
@@ -464,7 +470,7 @@ impl Serialize for DictNonStrKey {
 
         let opts = self.state.opts() & NOT_PASSTHROUGH;
 
-        let len = ffi!(Py_SIZE(self.ptr)) as usize;
+        let len = isize_to_usize(ffi!(Py_SIZE(self.ptr)));
         assume!(len > 0);
 
         let mut items: SmallVec<[(CompactString, *mut pyo3_ffi::PyObject); 8]> =
diff --git a/src/serialize/per_type/fragment.rs b/src/serialize/per_type/fragment.rs
index c3b2e74c..c8afa3b8 100644
--- a/src/serialize/per_type/fragment.rs
+++ b/src/serialize/per_type/fragment.rs
@@ -4,6 +4,7 @@ use crate::ffi::{Fragment, PyBytes_AS_STRING, PyBytes_GET_SIZE};
 use crate::serialize::error::SerializeError;
 use crate::str::unicode_to_str;
 use crate::typeref::{BYTES_TYPE, STR_TYPE};
+use crate::util::isize_to_usize;
 
 use serde::ser::{Serialize, Serializer};
 
@@ -28,14 +29,14 @@ impl Serialize for FragmentSerializer {
     {
         let buffer: &[u8];
         unsafe {
-            let fragment: *mut Fragment = self.ptr as *mut Fragment;
+            let fragment: *mut Fragment = self.ptr.cast::<Fragment>();
             let ob_type = ob_type!((*fragment).contents);
-            if ob_type == BYTES_TYPE {
+            if core::ptr::eq(ob_type, BYTES_TYPE) {
                 buffer = core::slice::from_raw_parts(
-                    PyBytes_AS_STRING((*fragment).contents) as *const u8,
-                    PyBytes_GET_SIZE((*fragment).contents) as usize,
+                    PyBytes_AS_STRING((*fragment).contents).cast::<u8>(),
+                    isize_to_usize(PyBytes_GET_SIZE((*fragment).contents)),
                 );
-            } else if ob_type == STR_TYPE {
+            } else if core::ptr::eq(ob_type, STR_TYPE) {
                 let uni = unicode_to_str((*fragment).contents);
                 if unlikely!(uni.is_none()) {
                     err!(SerializeError::InvalidStr)
diff --git a/src/serialize/per_type/int.rs b/src/serialize/per_type/int.rs
index 75778cba..00ef9475 100644
--- a/src/serialize/per_type/int.rs
+++ b/src/serialize/per_type/int.rs
@@ -34,28 +34,41 @@ impl Serialize for IntSerializer {
             if crate::ffi::pylong_is_zero(self.ptr) {
                 return serializer.serialize_bytes(b"0");
             }
-            let is_signed = !crate::ffi::pylong_is_unsigned(self.ptr) as i32;
+            let is_signed = i32::from(!crate::ffi::pylong_is_unsigned(self.ptr));
             if crate::ffi::pylong_fits_in_i32(self.ptr) {
                 if is_signed == 0 {
+                    #[allow(clippy::cast_sign_loss)]
                     serializer.serialize_u64(crate::ffi::pylong_get_inline_value(self.ptr) as u64)
                 } else {
                     serializer.serialize_i64(crate::ffi::pylong_get_inline_value(self.ptr))
                 }
             } else {
                 let mut buffer: [u8; 8] = [0; 8];
+
+                #[cfg(not(Py_3_13))]
+                let ret = pyo3_ffi::_PyLong_AsByteArray(
+                    self.ptr.cast::<pyo3_ffi::PyLongObject>(),
+                    buffer.as_mut_ptr().cast::<core::ffi::c_uchar>(),
+                    8,
+                    1,
+                    is_signed,
+                );
+                #[cfg(Py_3_13)]
                 let ret = pyo3_ffi::_PyLong_AsByteArray(
-                    self.ptr as *mut pyo3_ffi::PyLongObject,
-                    buffer.as_mut_ptr() as *mut core::ffi::c_uchar,
+                    self.ptr.cast::<pyo3_ffi::PyLongObject>(),
+                    buffer.as_mut_ptr().cast::<core::ffi::c_uchar>(),
                     8,
                     1,
                     is_signed,
+                    0,
                 );
                 if unlikely!(ret == -1) {
+                    #[cfg(not(Py_3_13))]
                     ffi!(PyErr_Clear());
                     err!(SerializeError::Integer64Bits)
                 }
                 if is_signed == 0 {
-                    let val = core::mem::transmute::<[u8; 8], u64>(buffer);
+                    let val = u64::from_ne_bytes(buffer);
                     if unlikely!(opt_enabled!(self.opts, STRICT_INTEGER))
                         && val > STRICT_INT_MAX as u64
                     {
@@ -63,7 +76,7 @@ impl Serialize for IntSerializer {
                     }
                     serializer.serialize_u64(val)
                 } else {
-                    let val = core::mem::transmute::<[u8; 8], i64>(buffer);
+                    let val = i64::from_ne_bytes(buffer);
                     if unlikely!(opt_enabled!(self.opts, STRICT_INTEGER))
                         && !(STRICT_INT_MIN..=STRICT_INT_MAX).contains(&val)
                     {
diff --git a/src/serialize/per_type/list.rs b/src/serialize/per_type/list.rs
index c866d21d..505840e2 100644
--- a/src/serialize/per_type/list.rs
+++ b/src/serialize/per_type/list.rs
@@ -9,7 +9,8 @@ use crate::serialize::per_type::{
 };
 use crate::serialize::serializer::PyObjectSerializer;
 use crate::serialize::state::SerializerState;
-use crate::typeref::*;
+use crate::typeref::{LIST_TYPE, TUPLE_TYPE};
+use crate::util::isize_to_usize;
 
 use core::ptr::NonNull;
 use serde::ser::{Serialize, SerializeSeq, Serializer};
@@ -49,8 +50,8 @@ impl ListTupleSerializer {
             is_type!(ob_type!(ptr), LIST_TYPE)
                 || is_subclass_by_flag!(tp_flags!(ob_type!(ptr)), Py_TPFLAGS_LIST_SUBCLASS)
         );
-        let data_ptr = unsafe { (*(ptr as *mut pyo3_ffi::PyListObject)).ob_item };
-        let len = ffi!(Py_SIZE(ptr)) as usize;
+        let data_ptr = unsafe { (*ptr.cast::<pyo3_ffi::PyListObject>()).ob_item };
+        let len = isize_to_usize(ffi!(Py_SIZE(ptr)));
         Self {
             data_ptr: data_ptr,
             len: len,
@@ -68,8 +69,8 @@ impl ListTupleSerializer {
             is_type!(ob_type!(ptr), TUPLE_TYPE)
                 || is_subclass_by_flag!(tp_flags!(ob_type!(ptr)), Py_TPFLAGS_TUPLE_SUBCLASS)
         );
-        let data_ptr = unsafe { (*(ptr as *mut pyo3_ffi::PyTupleObject)).ob_item.as_ptr() };
-        let len = ffi!(Py_SIZE(ptr)) as usize;
+        let data_ptr = unsafe { (*ptr.cast::<pyo3_ffi::PyTupleObject>()).ob_item.as_ptr() };
+        let len = isize_to_usize(ffi!(Py_SIZE(ptr)));
         Self {
             data_ptr: data_ptr,
             len: len,
diff --git a/src/serialize/per_type/numpy.rs b/src/serialize/per_type/numpy.rs
index 0844292b..f888029f 100644
--- a/src/serialize/per_type/numpy.rs
+++ b/src/serialize/per_type/numpy.rs
@@ -1,5 +1,6 @@
-use crate::opt::*;
+// SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
+use crate::opt::Opt;
 use crate::serialize::buffer::SmallFixedBuffer;
 use crate::serialize::error::SerializeError;
 use crate::serialize::per_type::{
@@ -7,10 +8,11 @@ use crate::serialize::per_type::{
 };
 use crate::serialize::serializer::PyObjectSerializer;
 use crate::typeref::{load_numpy_types, ARRAY_STRUCT_STR, DESCR_STR, DTYPE_STR, NUMPY_TYPES};
+use crate::util::isize_to_usize;
 use core::ffi::{c_char, c_int, c_void};
 use jiff::civil::DateTime;
 use jiff::Timestamp;
-use pyo3_ffi::*;
+use pyo3_ffi::{PyObject, PyTypeObject, Py_intptr_t, Py_ssize_t};
 use serde::ser::{self, Serialize, SerializeSeq, Serializer};
 use std::fmt;
 
@@ -36,7 +38,7 @@ impl<'a> Serialize for NumpySerializer<'a> {
         match NumpyArray::new(self.previous.ptr, self.previous.state.opts()) {
             Ok(val) => val.serialize(serializer),
             Err(PyArrayError::Malformed) => err!(SerializeError::NumpyMalformed),
-            Err(PyArrayError::NotContiguous) | Err(PyArrayError::UnsupportedDataType)
+            Err(PyArrayError::NotContiguous | PyArrayError::UnsupportedDataType)
                 if self.previous.default.is_some() =>
             {
                 DefaultSerializer::new(self.previous).serialize(serializer)
@@ -67,19 +69,19 @@ pub fn is_numpy_scalar(ob_type: *mut PyTypeObject) -> bool {
         false
     } else {
         let scalar_types = unsafe { numpy_types.unwrap().as_ref() };
-        ob_type == scalar_types.float64
-            || ob_type == scalar_types.float32
-            || ob_type == scalar_types.float16
-            || ob_type == scalar_types.int64
-            || ob_type == scalar_types.int16
-            || ob_type == scalar_types.int32
-            || ob_type == scalar_types.int8
-            || ob_type == scalar_types.uint64
-            || ob_type == scalar_types.uint32
-            || ob_type == scalar_types.uint8
-            || ob_type == scalar_types.uint16
-            || ob_type == scalar_types.bool_
-            || ob_type == scalar_types.datetime64
+        core::ptr::eq(ob_type, scalar_types.float64)
+            || core::ptr::eq(ob_type, scalar_types.float32)
+            || core::ptr::eq(ob_type, scalar_types.float16)
+            || core::ptr::eq(ob_type, scalar_types.int64)
+            || core::ptr::eq(ob_type, scalar_types.int16)
+            || core::ptr::eq(ob_type, scalar_types.int32)
+            || core::ptr::eq(ob_type, scalar_types.int8)
+            || core::ptr::eq(ob_type, scalar_types.uint64)
+            || core::ptr::eq(ob_type, scalar_types.uint32)
+            || core::ptr::eq(ob_type, scalar_types.uint8)
+            || core::ptr::eq(ob_type, scalar_types.uint16)
+            || core::ptr::eq(ob_type, scalar_types.bool_)
+            || core::ptr::eq(ob_type, scalar_types.datetime64)
     }
 }
 
@@ -90,7 +92,7 @@ pub fn is_numpy_array(ob_type: *mut PyTypeObject) -> bool {
         false
     } else {
         let scalar_types = unsafe { numpy_types.unwrap().as_ref() };
-        unsafe { ob_type == scalar_types.array }
+        unsafe { core::ptr::eq(ob_type, scalar_types.array) }
     }
 }
 
@@ -193,7 +195,11 @@ impl NumpyArray {
     #[cfg_attr(feature = "optimize", optimize(size))]
     pub fn new(ptr: *mut PyObject, opts: Opt) -> Result<Self, PyArrayError> {
         let capsule = ffi!(PyObject_GetAttr(ptr, ARRAY_STRUCT_STR));
-        let array = unsafe { (*(capsule as *mut PyCapsule)).pointer as *mut PyArrayInterface };
+        let array = unsafe {
+            (*capsule.cast::<PyCapsule>())
+                .pointer
+                .cast::<PyArrayInterface>()
+        };
         if unsafe { (*array).two != 2 } {
             ffi!(Py_DECREF(capsule));
             Err(PyArrayError::Malformed)
@@ -204,6 +210,8 @@ impl NumpyArray {
             ffi!(Py_DECREF(capsule));
             Err(PyArrayError::NotNativeEndian)
         } else {
+            debug_assert!(unsafe { (*array).nd >= 0 });
+            #[allow(clippy::cast_sign_loss)]
             let num_dimensions = unsafe { (*array).nd as usize };
             if num_dimensions == 0 {
                 ffi!(Py_DECREF(capsule));
@@ -220,7 +228,7 @@ impl NumpyArray {
                         position: vec![0; num_dimensions],
                         children: Vec::with_capacity(num_dimensions),
                         depth: 0,
-                        capsule: capsule as *mut PyCapsule,
+                        capsule: capsule.cast::<PyCapsule>(),
                         kind: kind,
                         opts,
                     };
@@ -252,15 +260,15 @@ impl NumpyArray {
     fn build(&mut self) {
         if self.depth < self.dimensions() - 1 {
             for i in 0..self.shape()[self.depth] {
-                let mut position: Vec<isize> = self.position.to_vec();
+                let mut position: Vec<isize> = self.position.clone();
                 position[self.depth] = i;
                 let num_children: usize = if self.depth < self.dimensions() - 2 {
-                    self.shape()[self.depth + 1] as usize
+                    isize_to_usize(self.shape()[self.depth + 1])
                 } else {
                     0
                 };
                 self.children
-                    .push(self.child_from_parent(position, num_children))
+                    .push(self.child_from_parent(position, num_children));
             }
         }
     }
@@ -278,27 +286,30 @@ impl NumpyArray {
     }
 
     fn num_items(&self) -> usize {
-        self.shape()[self.shape().len() - 1] as usize
+        isize_to_usize(self.shape()[self.shape().len() - 1])
     }
 
     fn dimensions(&self) -> usize {
-        unsafe { (*self.array).nd as usize }
+        #[allow(clippy::cast_sign_loss)]
+        unsafe {
+            (*self.array).nd as usize
+        }
     }
 
     fn shape(&self) -> &[isize] {
-        slice!((*self.array).shape as *const isize, self.dimensions())
+        slice!((*self.array).shape.cast_const(), self.dimensions())
     }
 
     fn strides(&self) -> &[isize] {
-        slice!((*self.array).strides as *const isize, self.dimensions())
+        slice!((*self.array).strides.cast_const(), self.dimensions())
     }
 }
 
 impl Drop for NumpyArray {
     fn drop(&mut self) {
         if self.depth == 0 {
-            ffi!(Py_DECREF(self.array as *mut pyo3_ffi::PyObject));
-            ffi!(Py_DECREF(self.capsule as *mut pyo3_ffi::PyObject));
+            ffi!(Py_DECREF(self.array.cast::<PyObject>()));
+            ffi!(Py_DECREF(self.capsule.cast::<PyObject>()));
         }
     }
 }
@@ -322,55 +333,55 @@ impl Serialize for NumpyArray {
         } else {
             match self.kind {
                 ItemType::F64 => {
-                    NumpyF64Array::new(slice!(self.data() as *const f64, self.num_items()))
+                    NumpyF64Array::new(slice!(self.data().cast::<f64>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::F32 => {
-                    NumpyF32Array::new(slice!(self.data() as *const f32, self.num_items()))
+                    NumpyF32Array::new(slice!(self.data().cast::<f32>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::F16 => {
-                    NumpyF16Array::new(slice!(self.data() as *const u16, self.num_items()))
+                    NumpyF16Array::new(slice!(self.data().cast::<u16>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::U64 => {
-                    NumpyU64Array::new(slice!(self.data() as *const u64, self.num_items()))
+                    NumpyU64Array::new(slice!(self.data().cast::<u64>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::U32 => {
-                    NumpyU32Array::new(slice!(self.data() as *const u32, self.num_items()))
+                    NumpyU32Array::new(slice!(self.data().cast::<u32>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::U16 => {
-                    NumpyU16Array::new(slice!(self.data() as *const u16, self.num_items()))
+                    NumpyU16Array::new(slice!(self.data().cast::<u16>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::U8 => {
-                    NumpyU8Array::new(slice!(self.data() as *const u8, self.num_items()))
+                    NumpyU8Array::new(slice!(self.data().cast::<u8>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::I64 => {
-                    NumpyI64Array::new(slice!(self.data() as *const i64, self.num_items()))
+                    NumpyI64Array::new(slice!(self.data().cast::<i64>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::I32 => {
-                    NumpyI32Array::new(slice!(self.data() as *const i32, self.num_items()))
+                    NumpyI32Array::new(slice!(self.data().cast::<i32>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::I16 => {
-                    NumpyI16Array::new(slice!(self.data() as *const i16, self.num_items()))
+                    NumpyI16Array::new(slice!(self.data().cast::<i16>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::I8 => {
-                    NumpyI8Array::new(slice!(self.data() as *const i8, self.num_items()))
+                    NumpyI8Array::new(slice!(self.data().cast::<i8>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::BOOL => {
-                    NumpyBoolArray::new(slice!(self.data() as *const u8, self.num_items()))
+                    NumpyBoolArray::new(slice!(self.data().cast::<u8>(), self.num_items()))
                         .serialize(serializer)
                 }
                 ItemType::DATETIME64(unit) => NumpyDatetime64Array::new(
-                    slice!(self.data() as *const i64, self.num_items()),
+                    slice!(self.data().cast::<i64>(), self.num_items()),
                     unit,
                     self.opts,
                 )
@@ -624,7 +635,7 @@ impl Serialize for DataTypeU16 {
     where
         S: Serializer,
     {
-        serializer.serialize_u32(self.obj as u32)
+        serializer.serialize_u32(u32::from(self.obj))
     }
 }
 
@@ -747,7 +758,7 @@ impl Serialize for DataTypeI16 {
     where
         S: Serializer,
     {
-        serializer.serialize_i32(self.obj as i32)
+        serializer.serialize_i32(i32::from(self.obj))
     }
 }
 
@@ -788,7 +799,7 @@ impl Serialize for DataTypeI8 {
     where
         S: Serializer,
     {
-        serializer.serialize_i32(self.obj as i32)
+        serializer.serialize_i32(i32::from(self.obj))
     }
 }
 
@@ -829,7 +840,7 @@ impl Serialize for DataTypeU8 {
     where
         S: Serializer,
     {
-        serializer.serialize_u32(self.obj as u32)
+        serializer.serialize_u32(u32::from(self.obj))
     }
 }
 
@@ -896,33 +907,33 @@ impl Serialize for NumpyScalar {
             let ob_type = ob_type!(self.ptr);
             let scalar_types =
                 unsafe { NUMPY_TYPES.get_or_init(load_numpy_types).unwrap().as_ref() };
-            if ob_type == scalar_types.float64 {
-                (*(self.ptr as *mut NumpyFloat64)).serialize(serializer)
-            } else if ob_type == scalar_types.float32 {
-                (*(self.ptr as *mut NumpyFloat32)).serialize(serializer)
-            } else if ob_type == scalar_types.float16 {
-                (*(self.ptr as *mut NumpyFloat16)).serialize(serializer)
-            } else if ob_type == scalar_types.int64 {
-                (*(self.ptr as *mut NumpyInt64)).serialize(serializer)
-            } else if ob_type == scalar_types.int32 {
-                (*(self.ptr as *mut NumpyInt32)).serialize(serializer)
-            } else if ob_type == scalar_types.int16 {
-                (*(self.ptr as *mut NumpyInt16)).serialize(serializer)
-            } else if ob_type == scalar_types.int8 {
-                (*(self.ptr as *mut NumpyInt8)).serialize(serializer)
-            } else if ob_type == scalar_types.uint64 {
-                (*(self.ptr as *mut NumpyUint64)).serialize(serializer)
-            } else if ob_type == scalar_types.uint32 {
-                (*(self.ptr as *mut NumpyUint32)).serialize(serializer)
-            } else if ob_type == scalar_types.uint16 {
-                (*(self.ptr as *mut NumpyUint16)).serialize(serializer)
-            } else if ob_type == scalar_types.uint8 {
-                (*(self.ptr as *mut NumpyUint8)).serialize(serializer)
-            } else if ob_type == scalar_types.bool_ {
-                (*(self.ptr as *mut NumpyBool)).serialize(serializer)
-            } else if ob_type == scalar_types.datetime64 {
+            if core::ptr::eq(ob_type, scalar_types.float64) {
+                (*(self.ptr.cast::<NumpyFloat64>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.float32) {
+                (*(self.ptr.cast::<NumpyFloat32>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.float16) {
+                (*(self.ptr.cast::<NumpyFloat16>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.int64) {
+                (*(self.ptr.cast::<NumpyInt64>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.int32) {
+                (*(self.ptr.cast::<NumpyInt32>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.int16) {
+                (*(self.ptr.cast::<NumpyInt16>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.int8) {
+                (*(self.ptr.cast::<NumpyInt8>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.uint64) {
+                (*(self.ptr.cast::<NumpyUint64>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.uint32) {
+                (*(self.ptr.cast::<NumpyUint32>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.uint16) {
+                (*(self.ptr.cast::<NumpyUint16>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.uint8) {
+                (*(self.ptr.cast::<NumpyUint8>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.bool_) {
+                (*(self.ptr.cast::<NumpyBool>())).serialize(serializer)
+            } else if core::ptr::eq(ob_type, scalar_types.datetime64) {
                 let unit = NumpyDatetimeUnit::from_pyobject(self.ptr);
-                let obj = &*(self.ptr as *mut NumpyDatetime64);
+                let obj = &*self.ptr.cast::<NumpyDatetime64>();
                 let dt = unit
                     .datetime(obj.value, self.opts)
                     .map_err(NumpyDateTimeError::into_serde_err)?;
@@ -947,7 +958,7 @@ impl Serialize for NumpyInt8 {
     where
         S: Serializer,
     {
-        serializer.serialize_i32(self.value as i32)
+        serializer.serialize_i32(i32::from(self.value))
     }
 }
 
@@ -964,7 +975,7 @@ impl Serialize for NumpyInt16 {
     where
         S: Serializer,
     {
-        serializer.serialize_i32(self.value as i32)
+        serializer.serialize_i32(i32::from(self.value))
     }
 }
 
@@ -1015,7 +1026,7 @@ impl Serialize for NumpyUint8 {
     where
         S: Serializer,
     {
-        serializer.serialize_u32(self.value as u32)
+        serializer.serialize_u32(u32::from(self.value))
     }
 }
 
@@ -1032,7 +1043,7 @@ impl Serialize for NumpyUint16 {
     where
         S: Serializer,
     {
-        serializer.serialize_u32(self.value as u32)
+        serializer.serialize_u32(u32::from(self.value))
     }
 }
 
@@ -1182,7 +1193,7 @@ impl fmt::Display for NumpyDatetimeUnit {
             Self::Attoseconds => "attoseconds",
             Self::Generic => "generic",
         };
-        write!(f, "{}", unit)
+        write!(f, "{unit}")
     }
 }
 
@@ -1196,9 +1207,9 @@ impl NumpyDateTimeError {
     #[cold]
     fn into_serde_err<T: ser::Error>(self) -> T {
         let err = match self {
-            Self::UnsupportedUnit(unit) => format!("unsupported numpy.datetime64 unit: {}", unit),
+            Self::UnsupportedUnit(unit) => format!("unsupported numpy.datetime64 unit: {unit}"),
             Self::Unrepresentable { unit, val } => {
-                format!("unrepresentable numpy.datetime64: {} {}", val, unit)
+                format!("unrepresentable numpy.datetime64: {val} {unit}")
             }
         };
         ser::Error::custom(err)
@@ -1209,7 +1220,7 @@ macro_rules! to_jiff_datetime {
     ($timestamp:expr, $self:expr, $val:expr) => {
         Ok(
             ($timestamp.map_err(|_| NumpyDateTimeError::Unrepresentable {
-                unit: *$self,
+                unit: $self,
                 val: $val,
             })?)
             .to_zoned(jiff::tz::TimeZone::UTC)
@@ -1268,12 +1279,12 @@ impl NumpyDatetimeUnit {
     /// Returns an `Err(NumpyDateTimeError)` if the value is invalid for this unit.
     #[cold]
     #[cfg_attr(feature = "optimize", optimize(size))]
-    fn datetime(&self, val: i64, opts: Opt) -> Result<NumpyDatetime64Repr, NumpyDateTimeError> {
+    fn datetime(self, val: i64, opts: Opt) -> Result<NumpyDatetime64Repr, NumpyDateTimeError> {
         match self {
             Self::Years => Ok(DateTime::new(
                 (val + 1970)
                     .try_into()
-                    .map_err(|_| NumpyDateTimeError::Unrepresentable { unit: *self, val })?,
+                    .map_err(|_| NumpyDateTimeError::Unrepresentable { unit: self, val })?,
                 1,
                 1,
                 0,
@@ -1285,10 +1296,10 @@ impl NumpyDatetimeUnit {
             Self::Months => Ok(DateTime::new(
                 (val / 12 + 1970)
                     .try_into()
-                    .map_err(|_| NumpyDateTimeError::Unrepresentable { unit: *self, val })?,
+                    .map_err(|_| NumpyDateTimeError::Unrepresentable { unit: self, val })?,
                 (val % 12 + 1)
                     .try_into()
-                    .map_err(|_| NumpyDateTimeError::Unrepresentable { unit: *self, val })?,
+                    .map_err(|_| NumpyDateTimeError::Unrepresentable { unit: self, val })?,
                 1,
                 0,
                 0,
@@ -1306,9 +1317,9 @@ impl NumpyDatetimeUnit {
             Self::Milliseconds => to_jiff_datetime!(Timestamp::from_millisecond(val), self, val),
             Self::Microseconds => to_jiff_datetime!(Timestamp::from_microsecond(val), self, val),
             Self::Nanoseconds => {
-                to_jiff_datetime!(Timestamp::from_nanosecond(val as i128), self, val)
+                to_jiff_datetime!(Timestamp::from_nanosecond(i128::from(val)), self, val)
             }
-            _ => Err(NumpyDateTimeError::UnsupportedUnit(*self)),
+            _ => Err(NumpyDateTimeError::UnsupportedUnit(self)),
         }
         .map(|dt| NumpyDatetime64Repr { dt, opts })
     }
@@ -1354,7 +1365,10 @@ pub struct NumpyDatetime64 {
 macro_rules! forward_inner {
     ($meth: ident, $ty: ident) => {
         fn $meth(&self) -> $ty {
-            self.dt.$meth() as $ty
+            debug_assert!(self.dt.$meth() >= 0);
+            #[allow(clippy::cast_sign_loss)]
+            let ret = self.dt.$meth() as $ty; // stmt_expr_attributes
+            ret
         }
     };
 }
@@ -1373,7 +1387,10 @@ impl DateTimeLike for NumpyDatetime64Repr {
     forward_inner!(second, u8);
 
     fn nanosecond(&self) -> u32 {
-        self.dt.subsec_nanosecond() as u32
+        debug_assert!(self.dt.subsec_nanosecond() >= 0);
+        #[allow(clippy::cast_sign_loss)]
+        let ret = self.dt.subsec_nanosecond() as u32; // stmt_expr_attributes
+        ret
     }
 
     fn microsecond(&self) -> u32 {
diff --git a/src/serialize/per_type/pybool.rs b/src/serialize/per_type/pybool.rs
index 6109693b..f6c495cf 100644
--- a/src/serialize/per_type/pybool.rs
+++ b/src/serialize/per_type/pybool.rs
@@ -19,6 +19,6 @@ impl Serialize for BoolSerializer {
     where
         S: Serializer,
     {
-        serializer.serialize_bool(unsafe { self.ptr == crate::typeref::TRUE })
+        serializer.serialize_bool(unsafe { core::ptr::eq(self.ptr, crate::typeref::TRUE) })
     }
 }
diff --git a/src/serialize/per_type/unicode.rs b/src/serialize/per_type/unicode.rs
index ecc8ae6c..23270916 100644
--- a/src/serialize/per_type/unicode.rs
+++ b/src/serialize/per_type/unicode.rs
@@ -26,7 +26,7 @@ impl Serialize for StrSerializer {
             let tmp = unicode_to_str(self.ptr);
             if unlikely!(tmp.is_none()) {
                 err!(SerializeError::InvalidStr)
-            };
+            }
             tmp.unwrap()
         };
         serializer.serialize_str(uni)
diff --git a/src/serialize/per_type/uuid.rs b/src/serialize/per_type/uuid.rs
index f065fd2a..12e5bb78 100644
--- a/src/serialize/per_type/uuid.rs
+++ b/src/serialize/per_type/uuid.rs
@@ -22,16 +22,26 @@ impl UUID {
             // test_uuid_immutable, test_uuid_int
             let py_int = ffi!(PyObject_GetAttr(self.ptr, INT_ATTR_STR));
             ffi!(Py_DECREF(py_int));
-            let buffer: [c_uchar; 16] = [0; 16];
+            let mut buffer: [c_uchar; 16] = [0; 16];
             unsafe {
                 // test_uuid_overflow
+                #[cfg(not(Py_3_13))]
                 pyo3_ffi::_PyLong_AsByteArray(
-                    py_int as *mut pyo3_ffi::PyLongObject,
-                    buffer.as_ptr() as *mut c_uchar,
+                    py_int.cast::<pyo3_ffi::PyLongObject>(),
+                    buffer.as_mut_ptr(),
                     16,
                     1, // little_endian
                     0, // is_signed
-                )
+                );
+                #[cfg(Py_3_13)]
+                pyo3_ffi::_PyLong_AsByteArray(
+                    py_int.cast::<pyo3_ffi::PyLongObject>(),
+                    buffer.as_mut_ptr(),
+                    16,
+                    1, // little_endian
+                    0, // is_signed
+                    0,
+                );
             };
             value = u128::from_le_bytes(buffer);
         }
diff --git a/src/serialize/serializer.rs b/src/serialize/serializer.rs
index 852d31e0..7b9a5de5 100644
--- a/src/serialize/serializer.rs
+++ b/src/serialize/serializer.rs
@@ -27,7 +27,7 @@ pub fn serialize(
         to_writer_pretty(&mut buf, &obj)
     };
     match res {
-        Ok(_) => {
+        Ok(()) => {
             if opt_enabled!(opts, APPEND_NEWLINE) {
                 let _ = buf.write(b"\n");
             }
diff --git a/src/serialize/state.rs b/src/serialize/state.rs
index 91b5b0aa..81094731 100644
--- a/src/serialize/state.rs
+++ b/src/serialize/state.rs
@@ -1,6 +1,6 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use crate::opt::*;
+use crate::opt::Opt;
 
 const RECURSION_SHIFT: usize = 24;
 const RECURSION_MASK: u32 = 255 << RECURSION_SHIFT;
@@ -20,27 +20,27 @@ pub struct SerializerState {
 impl SerializerState {
     #[inline(always)]
     pub fn new(opts: Opt) -> Self {
-        debug_assert!(opts < u16::MAX as u32);
+        debug_assert!(opts < u32::from(u16::MAX));
         Self { state: opts }
     }
 
     #[inline(always)]
-    pub fn opts(&self) -> u32 {
+    pub fn opts(self) -> u32 {
         self.state
     }
 
     #[inline(always)]
-    pub fn recursion_limit(&self) -> bool {
+    pub fn recursion_limit(self) -> bool {
         self.state & RECURSION_MASK == RECURSION_MASK
     }
 
     #[inline(always)]
-    pub fn default_calls_limit(&self) -> bool {
+    pub fn default_calls_limit(self) -> bool {
         self.state & DEFAULT_MASK == DEFAULT_MASK
     }
 
     #[inline(always)]
-    pub fn copy_for_recursive_call(&self) -> Self {
+    pub fn copy_for_recursive_call(self) -> Self {
         let opt = self.state & !RECURSION_MASK;
         let recursion = (((self.state & RECURSION_MASK) >> RECURSION_SHIFT) + 1) << RECURSION_SHIFT;
         Self {
@@ -49,7 +49,7 @@ impl SerializerState {
     }
 
     #[inline(always)]
-    pub fn copy_for_default_call(&self) -> Self {
+    pub fn copy_for_default_call(self) -> Self {
         let opt = self.state & !DEFAULT_MASK;
         let default_calls = (((self.state & DEFAULT_MASK) >> DEFAULT_SHIFT) + 1) << DEFAULT_SHIFT;
         Self {
diff --git a/src/serialize/writer/byteswriter.rs b/src/serialize/writer/byteswriter.rs
index dd121838..b2d13d78 100644
--- a/src/serialize/writer/byteswriter.rs
+++ b/src/serialize/writer/byteswriter.rs
@@ -1,10 +1,8 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use core::ffi::c_char;
+use crate::util::usize_to_isize;
 use core::ptr::NonNull;
-use pyo3_ffi::{
-    PyBytesObject, PyBytes_FromStringAndSize, PyObject, PyVarObject, Py_ssize_t, _PyBytes_Resize,
-};
+use pyo3_ffi::{PyBytesObject, PyBytes_FromStringAndSize, PyObject, PyVarObject, _PyBytes_Resize};
 use std::io::Error;
 
 const BUFFER_LENGTH: usize = 1024;
@@ -21,43 +19,36 @@ impl BytesWriter {
             cap: BUFFER_LENGTH,
             len: 0,
             bytes: unsafe {
-                PyBytes_FromStringAndSize(core::ptr::null_mut(), BUFFER_LENGTH as isize)
-                    as *mut PyBytesObject
+                PyBytes_FromStringAndSize(core::ptr::null_mut(), usize_to_isize(BUFFER_LENGTH))
+                    .cast::<PyBytesObject>()
             },
         }
     }
 
     pub fn bytes_ptr(&mut self) -> NonNull<PyObject> {
-        unsafe { NonNull::new_unchecked(self.bytes as *mut PyObject) }
+        unsafe { NonNull::new_unchecked(self.bytes.cast::<PyObject>()) }
     }
 
     pub fn finish(&mut self) -> NonNull<PyObject> {
         unsafe {
             core::ptr::write(self.buffer_ptr(), 0);
-            (*self.bytes.cast::<PyVarObject>()).ob_size = self.len as Py_ssize_t;
+            (*self.bytes.cast::<PyVarObject>()).ob_size = usize_to_isize(self.len);
             self.resize(self.len);
             self.bytes_ptr()
         }
     }
 
     fn buffer_ptr(&self) -> *mut u8 {
-        unsafe {
-            core::mem::transmute::<*mut [c_char; 1], *mut u8>(core::ptr::addr_of_mut!(
-                (*self.bytes).ob_sval
-            ))
-            .add(self.len)
-        }
+        unsafe { (&raw mut (*self.bytes).ob_sval).cast::<u8>().add(self.len) }
     }
 
     #[inline]
     pub fn resize(&mut self, len: usize) {
         self.cap = len;
         unsafe {
-            #[allow(clippy::unnecessary_cast)]
             _PyBytes_Resize(
-                core::ptr::addr_of_mut!(self.bytes) as *mut *mut PyBytesObject
-                    as *mut *mut PyObject,
-                len as isize,
+                (&raw mut self.bytes).cast::<*mut PyObject>(),
+                usize_to_isize(len),
             );
         }
     }
diff --git a/src/serialize/writer/json.rs b/src/serialize/writer/json.rs
index cbb5fb33..c893d5e5 100644
--- a/src/serialize/writer/json.rs
+++ b/src/serialize/writer/json.rs
@@ -2,7 +2,6 @@
 // This is an adaptation of `src/value/ser.rs` from serde-json.
 
 use crate::serialize::writer::formatter::{CompactFormatter, Formatter, PrettyFormatter};
-use crate::serialize::writer::str::*;
 use crate::serialize::writer::WriteExt;
 use serde::ser::{self, Impossible, Serialize};
 use serde_json::error::{Error, Result};
@@ -572,7 +571,23 @@ macro_rules! reserve_str {
     };
 }
 
-#[cfg(all(feature = "unstable-simd", not(target_arch = "x86_64")))]
+#[cfg(all(target_arch = "x86_64", feature = "avx512"))]
+type StrFormatter = unsafe fn(*mut u8, *const u8, usize) -> usize;
+
+#[cfg(all(target_arch = "x86_64", feature = "avx512"))]
+static mut STR_FORMATTER_FN: StrFormatter =
+    crate::serialize::writer::str::format_escaped_str_impl_sse2_128;
+
+pub fn set_str_formatter_fn() {
+    unsafe {
+        #[cfg(all(target_arch = "x86_64", feature = "avx512"))]
+        if std::is_x86_feature_detected!("avx512vl") {
+            STR_FORMATTER_FN = crate::serialize::writer::str::format_escaped_str_impl_512vl;
+        }
+    }
+}
+
+#[cfg(all(target_arch = "x86_64", not(feature = "avx512")))]
 #[inline(always)]
 fn format_escaped_str<W>(writer: &mut W, value: &str)
 where
@@ -581,7 +596,7 @@ where
     unsafe {
         reserve_str!(writer, value);
 
-        let written = format_escaped_str_impl_generic_128(
+        let written = crate::serialize::writer::str::format_escaped_str_impl_sse2_128(
             writer.as_mut_buffer_ptr(),
             value.as_bytes().as_ptr(),
             value.len(),
@@ -591,11 +606,7 @@ where
     }
 }
 
-#[cfg(all(
-    feature = "unstable-simd",
-    target_arch = "x86_64",
-    not(feature = "avx512")
-))]
+#[cfg(all(target_arch = "x86_64", feature = "avx512"))]
 #[inline(always)]
 fn format_escaped_str<W>(writer: &mut W, value: &str)
 where
@@ -604,7 +615,7 @@ where
     unsafe {
         reserve_str!(writer, value);
 
-        let written = format_escaped_str_impl_sse2_128(
+        let written = STR_FORMATTER_FN(
             writer.as_mut_buffer_ptr(),
             value.as_bytes().as_ptr(),
             value.len(),
@@ -614,34 +625,11 @@ where
     }
 }
 
-#[cfg(all(feature = "unstable-simd", target_arch = "x86_64", feature = "avx512"))]
-#[inline(always)]
-fn format_escaped_str<W>(writer: &mut W, value: &str)
-where
-    W: ?Sized + io::Write + WriteExt,
-{
-    unsafe {
-        reserve_str!(writer, value);
-
-        if std::is_x86_feature_detected!("avx512vl") {
-            let written = format_escaped_str_impl_512vl(
-                writer.as_mut_buffer_ptr(),
-                value.as_bytes().as_ptr(),
-                value.len(),
-            );
-            writer.set_written(written);
-        } else {
-            let written = format_escaped_str_impl_sse2_128(
-                writer.as_mut_buffer_ptr(),
-                value.as_bytes().as_ptr(),
-                value.len(),
-            );
-            writer.set_written(written);
-        };
-    }
-}
-
-#[cfg(all(not(feature = "unstable-simd"), not(target_arch = "x86_64")))]
+#[cfg(all(
+    not(target_arch = "x86_64"),
+    not(feature = "avx512"),
+    feature = "generic_simd"
+))]
 #[inline(always)]
 fn format_escaped_str<W>(writer: &mut W, value: &str)
 where
@@ -650,16 +638,17 @@ where
     unsafe {
         reserve_str!(writer, value);
 
-        let written = format_escaped_str_scalar(
+        let written = crate::serialize::writer::str::format_escaped_str_impl_generic_128(
             writer.as_mut_buffer_ptr(),
             value.as_bytes().as_ptr(),
             value.len(),
         );
+
         writer.set_written(written);
     }
 }
 
-#[cfg(all(not(feature = "unstable-simd"), target_arch = "x86_64"))]
+#[cfg(all(not(target_arch = "x86_64"), not(feature = "generic_simd")))]
 #[inline(always)]
 fn format_escaped_str<W>(writer: &mut W, value: &str)
 where
@@ -668,7 +657,7 @@ where
     unsafe {
         reserve_str!(writer, value);
 
-        let written = format_escaped_str_impl_sse2_128(
+        let written = crate::serialize::writer::str::format_escaped_str_scalar(
             writer.as_mut_buffer_ptr(),
             value.as_bytes().as_ptr(),
             value.len(),
diff --git a/src/serialize/writer/mod.rs b/src/serialize/writer/mod.rs
index 3e945b2c..942e66b4 100644
--- a/src/serialize/writer/mod.rs
+++ b/src/serialize/writer/mod.rs
@@ -6,4 +6,4 @@ mod json;
 mod str;
 
 pub use byteswriter::{BytesWriter, WriteExt};
-pub use json::{to_writer, to_writer_pretty};
+pub use json::{set_str_formatter_fn, to_writer, to_writer_pretty};
diff --git a/src/serialize/writer/str/avx512.rs b/src/serialize/writer/str/avx512.rs
index 983d4a30..02b7d81e 100644
--- a/src/serialize/writer/str/avx512.rs
+++ b/src/serialize/writer/str/avx512.rs
@@ -1,114 +1,89 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use core::mem::transmute;
-
-use super::escape::QUOTE_TAB;
-
 use core::arch::x86_64::{
-    __m256i, _mm256_cmpeq_epu8_mask, _mm256_cmplt_epu8_mask, _mm256_load_si256, _mm256_loadu_si256,
-    _mm256_maskz_loadu_epi8, _mm256_storeu_epi8,
+    _mm256_cmpeq_epu8_mask, _mm256_cmplt_epu8_mask, _mm256_loadu_epi8, _mm256_maskz_loadu_epi8,
+    _mm256_set1_epi8, _mm256_storeu_epi8,
 };
 
-#[repr(C, align(32))]
-struct ConstArray {
-    pub data: [u8; 32],
+macro_rules! splat_mm256 {
+    ($val:expr) => {
+        _mm256_set1_epi8(core::mem::transmute::<u8, i8>($val))
+    };
 }
 
-const BLASH: ConstArray = ConstArray { data: [b'\\'; 32] };
-const QUOTE: ConstArray = ConstArray { data: [b'"'; 32] };
-const X20: ConstArray = ConstArray { data: [32; 32] };
+#[inline(never)]
+#[target_feature(enable = "avx512f,avx512bw,avx512vl,bmi2")]
+pub unsafe fn format_escaped_str_impl_512vl(
+    odst: *mut u8,
+    value_ptr: *const u8,
+    value_len: usize,
+) -> usize {
+    unsafe {
+        const STRIDE: usize = 32;
 
-macro_rules! impl_format_simd_avx512vl {
-    ($dst:expr, $src:expr, $value_len:expr) => {
-        let mut nb: usize = $value_len;
+        let mut dst = odst;
+        let mut src = value_ptr;
+        let mut nb: usize = value_len;
+
+        let blash = splat_mm256!(b'\\');
+        let quote = splat_mm256!(b'"');
+        let x20 = splat_mm256!(32);
 
-        let blash = _mm256_load_si256(BLASH.data.as_ptr() as *const __m256i);
-        let quote = _mm256_load_si256(QUOTE.data.as_ptr() as *const __m256i);
-        let x20 = _mm256_load_si256(X20.data.as_ptr() as *const __m256i);
+        core::ptr::write(dst, b'"');
+        dst = dst.add(1);
 
         unsafe {
             while nb >= STRIDE {
-                let str_vec = _mm256_loadu_si256(transmute::<*const u8, *const __m256i>($src));
+                let str_vec = _mm256_loadu_epi8(src.cast::<i8>());
 
-                _mm256_storeu_epi8($dst as *mut i8, str_vec);
+                _mm256_storeu_epi8(dst.cast::<i8>(), str_vec);
 
                 let mask = _mm256_cmpeq_epu8_mask(str_vec, blash)
                     | _mm256_cmpeq_epu8_mask(str_vec, quote)
                     | _mm256_cmplt_epu8_mask(str_vec, x20);
 
-                if unlikely!(mask > 0) {
+                if mask != 0 {
                     let cn = trailing_zeros!(mask);
-                    $src = $src.add(cn);
-                    $dst = $dst.add(cn);
+                    src = src.add(cn);
+                    dst = dst.add(cn);
                     nb -= cn;
                     nb -= 1;
 
-                    let escape = QUOTE_TAB[*($src) as usize];
-                    $src = $src.add(1);
-
-                    write_escape!(escape, $dst);
-                    $dst = $dst.add(escape.1 as usize);
+                    write_escape!(*(src), dst);
+                    src = src.add(1);
                 } else {
                     nb -= STRIDE;
-                    $dst = $dst.add(STRIDE);
-                    $src = $src.add(STRIDE);
+                    dst = dst.add(STRIDE);
+                    src = src.add(STRIDE);
                 }
             }
 
-            if nb > 0 {
-                loop {
-                    let remainder_mask = !(u32::MAX << nb);
-                    let str_vec = _mm256_maskz_loadu_epi8(remainder_mask, $src as *const i8);
-
-                    _mm256_storeu_epi8($dst as *mut i8, str_vec);
-
-                    let mask = (_mm256_cmpeq_epu8_mask(str_vec, blash)
-                        | _mm256_cmpeq_epu8_mask(str_vec, quote)
-                        | _mm256_cmplt_epu8_mask(str_vec, x20))
-                        & remainder_mask;
-
-                    if unlikely!(mask > 0) {
-                        let cn = trailing_zeros!(mask);
-                        $src = $src.add(cn);
-                        $dst = $dst.add(cn);
-                        nb -= cn;
-                        nb -= 1;
-
-                        let escape = QUOTE_TAB[*($src) as usize];
-                        $src = $src.add(1);
-
-                        write_escape!(escape, $dst);
-                        $dst = $dst.add(escape.1 as usize);
-                    } else {
-                        $dst = $dst.add(nb);
-                        break;
-                    }
-                }
-            }
-        }
-    };
-}
+            loop {
+                let remainder_mask = !(u32::MAX << nb);
+                let str_vec = _mm256_maskz_loadu_epi8(remainder_mask, src.cast::<i8>());
 
-#[inline(never)]
-#[cfg_attr(
-    feature = "avx512",
-    target_feature(enable = "avx512f,avx512bw,avx512vl,bmi2")
-)]
-pub unsafe fn format_escaped_str_impl_512vl(
-    odst: *mut u8,
-    value_ptr: *const u8,
-    value_len: usize,
-) -> usize {
-    unsafe {
-        const STRIDE: usize = 32;
+                _mm256_storeu_epi8(dst.cast::<i8>(), str_vec);
 
-        let mut dst = odst;
-        let mut src = value_ptr;
+                let mask = (_mm256_cmpeq_epu8_mask(str_vec, blash)
+                    | _mm256_cmpeq_epu8_mask(str_vec, quote)
+                    | _mm256_cmplt_epu8_mask(str_vec, x20))
+                    & remainder_mask;
 
-        core::ptr::write(dst, b'"');
-        dst = dst.add(1);
+                if mask != 0 {
+                    let cn = trailing_zeros!(mask);
+                    src = src.add(cn);
+                    dst = dst.add(cn);
+                    nb -= cn;
+                    nb -= 1;
 
-        impl_format_simd_avx512vl!(dst, src, value_len);
+                    write_escape!(*(src), dst);
+                    src = src.add(1);
+                } else {
+                    dst = dst.add(nb);
+                    break;
+                }
+            }
+        }
 
         core::ptr::write(dst, b'"');
         dst = dst.add(1);
diff --git a/src/serialize/writer/str/escape.rs b/src/serialize/writer/str/escape.rs
index b1c8e0e8..915d54fa 100644
--- a/src/serialize/writer/str/escape.rs
+++ b/src/serialize/writer/str/escape.rs
@@ -1,9 +1,25 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 // the constants and SIMD approach are adapted from cloudwego's sonic-rs
 
+#[cfg(feature = "inline_int")]
 macro_rules! write_escape {
-    ($escape:expr, $dst:expr) => {
-        core::ptr::copy_nonoverlapping($escape.0.as_ptr(), $dst, 8);
+    ($byte:expr, $dst:expr) => {
+        debug_assert!($byte < 96);
+        let escape = u64::from_ne_bytes(
+            *crate::serialize::writer::str::escape::QUOTE_TAB.get_unchecked($byte as usize),
+        );
+        core::ptr::write($dst.cast::<u64>(), escape);
+        $dst = $dst.add((escape as usize) >> 56);
+    };
+}
+
+#[cfg(not(feature = "inline_int"))]
+macro_rules! write_escape {
+    ($byte:expr, $dst:expr) => {
+        debug_assert!($byte < 96);
+        let escape = crate::serialize::writer::str::escape::QUOTE_TAB.get_unchecked($byte as usize);
+        core::ptr::copy_nonoverlapping(escape.as_ptr(), $dst, 8);
+        $dst = $dst.add(((*escape.as_ptr().add(7)) as usize));
     };
 }
 
@@ -18,101 +34,101 @@ pub const NEED_ESCAPED: [u8; 256] = [
     0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
 ];
 
-pub const QUOTE_TAB: [([u8; 7], u8); 96] = [
-    (*b"\\u0000\0", 6),
-    (*b"\\u0001\0", 6),
-    (*b"\\u0002\0", 6),
-    (*b"\\u0003\0", 6),
-    (*b"\\u0004\0", 6),
-    (*b"\\u0005\0", 6),
-    (*b"\\u0006\0", 6),
-    (*b"\\u0007\0", 6),
-    (*b"\\b\0\0\0\0\0", 2),
-    (*b"\\t\0\0\0\0\0", 2),
-    (*b"\\n\0\0\0\0\0", 2),
-    (*b"\\u000b\0", 6),
-    (*b"\\f\0\0\0\0\0", 2),
-    (*b"\\r\0\0\0\0\0", 2),
-    (*b"\\u000e\0", 6),
-    (*b"\\u000f\0", 6),
-    (*b"\\u0010\0", 6),
-    (*b"\\u0011\0", 6),
-    (*b"\\u0012\0", 6),
-    (*b"\\u0013\0", 6),
-    (*b"\\u0014\0", 6),
-    (*b"\\u0015\0", 6),
-    (*b"\\u0016\0", 6),
-    (*b"\\u0017\0", 6),
-    (*b"\\u0018\0", 6),
-    (*b"\\u0019\0", 6),
-    (*b"\\u001a\0", 6),
-    (*b"\\u001b\0", 6),
-    (*b"\\u001c\0", 6),
-    (*b"\\u001d\0", 6),
-    (*b"\\u001e\0", 6),
-    (*b"\\u001f\0", 6),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    (*b"\\\"\0\0\0\0\0", 2),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    (*b"\\\\\0\0\0\0\0", 2),
-    ([0; 7], 0),
-    ([0; 7], 0),
-    ([0; 7], 0),
+pub const QUOTE_TAB: [[u8; 8]; 96] = [
+    [b'\\', b'u', b'0', b'0', b'0', b'0', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'1', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'2', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'3', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'4', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'5', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'6', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'7', 0, 6],
+    [b'\\', b'b', b'0', b'0', b'0', b'0', 0, 2],
+    [b'\\', b't', b'0', b'0', b'0', b'0', 0, 2],
+    [b'\\', b'n', b'0', b'0', b'0', b'0', 0, 2],
+    [b'\\', b'u', b'0', b'0', b'0', b'b', 0, 6],
+    [b'\\', b'f', b'0', b'0', b'0', b'0', 0, 2],
+    [b'\\', b'r', b'0', b'0', b'0', b'0', 0, 2],
+    [b'\\', b'u', b'0', b'0', b'0', b'e', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'0', b'f', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'0', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'1', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'2', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'3', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'4', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'5', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'6', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'7', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'8', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'9', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'a', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'b', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'c', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'd', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'e', 0, 6],
+    [b'\\', b'u', b'0', b'0', b'1', b'f', 0, 6],
+    [0; 8],
+    [0; 8],
+    [b'\\', b'"', 0, 0, 0, 0, 0, 2],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [0; 8],
+    [b'\\', b'\\', 0, 0, 0, 0, 0, 2],
+    [0; 8],
+    [0; 8],
+    [0; 8],
 ];
diff --git a/src/serialize/writer/str/generic.rs b/src/serialize/writer/str/generic.rs
index e7973f2a..c4e60d10 100644
--- a/src/serialize/writer/str/generic.rs
+++ b/src/serialize/writer/str/generic.rs
@@ -1,46 +1,58 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use super::escape::{NEED_ESCAPED, QUOTE_TAB};
 use core::simd::cmp::{SimdPartialEq, SimdPartialOrd};
 
-macro_rules! impl_format_simd_generic_128 {
-    ($dst:expr, $src:expr, $value_len:expr) => {
-        let last_stride_src = $src.add($value_len).sub(STRIDE);
-        let mut nb: usize = $value_len;
+#[inline(never)]
+#[cfg_attr(target_arch = "aarch64", target_feature(enable = "neon"))]
+pub unsafe fn format_escaped_str_impl_generic_128(
+    odst: *mut u8,
+    value_ptr: *const u8,
+    value_len: usize,
+) -> usize {
+    unsafe {
+        const STRIDE: usize = 16;
+        type StrVector = core::simd::u8x16;
 
-        assume!($value_len >= STRIDE);
+        let mut dst = odst;
+        let mut src = value_ptr;
 
-        const BLASH: StrVector = StrVector::from_array([b'\\'; STRIDE]);
-        const QUOTE: StrVector = StrVector::from_array([b'"'; STRIDE]);
-        const X20: StrVector = StrVector::from_array([32; STRIDE]);
+        core::ptr::write(dst, b'"');
+        dst = dst.add(1);
 
-        unsafe {
-            {
-                while nb >= STRIDE {
-                    let v = StrVector::from_slice(core::slice::from_raw_parts($src, STRIDE));
-                    let mask =
-                        (v.simd_eq(BLASH) | v.simd_eq(QUOTE) | v.simd_lt(X20)).to_bitmask() as u32;
-                    v.copy_to_slice(core::slice::from_raw_parts_mut($dst, STRIDE));
+        if value_len < STRIDE {
+            impl_format_scalar!(dst, src, value_len);
+        } else {
+            let blash: StrVector = StrVector::splat(b'\\');
+            let quote: StrVector = StrVector::splat(b'"');
+            let x20: StrVector = StrVector::splat(32);
 
-                    if unlikely!(mask > 0) {
-                        let cn = trailing_zeros!(mask) as usize;
-                        nb -= cn;
-                        $dst = $dst.add(cn);
-                        $src = $src.add(cn);
-                        nb -= 1;
-                        let escape = QUOTE_TAB[*($src) as usize];
-                        write_escape!(escape, $dst);
-                        $dst = $dst.add(escape.1 as usize);
-                        $src = $src.add(1);
-                    } else {
-                        nb -= STRIDE;
-                        $dst = $dst.add(STRIDE);
-                        $src = $src.add(STRIDE);
+            let last_stride_src = src.add(value_len).sub(STRIDE);
+            let mut nb: usize = value_len;
+
+            unsafe {
+                {
+                    while nb >= STRIDE {
+                        let v = StrVector::from_slice(core::slice::from_raw_parts(src, STRIDE));
+                        let mask = (v.simd_eq(blash) | v.simd_eq(quote) | v.simd_lt(x20))
+                            .to_bitmask() as u32;
+                        v.copy_to_slice(core::slice::from_raw_parts_mut(dst, STRIDE));
+
+                        if mask != 0 {
+                            let cn = trailing_zeros!(mask) as usize;
+                            nb -= cn;
+                            dst = dst.add(cn);
+                            src = src.add(cn);
+                            nb -= 1;
+                            write_escape!(*(src), dst);
+                            src = src.add(1);
+                        } else {
+                            nb -= STRIDE;
+                            dst = dst.add(STRIDE);
+                            src = src.add(STRIDE);
+                        }
                     }
                 }
-            }
 
-            if nb > 0 {
                 let mut scratch: [u8; 32] = [b'a'; 32];
                 let mut v =
                     StrVector::from_slice(core::slice::from_raw_parts(last_stride_src, STRIDE));
@@ -52,56 +64,27 @@ macro_rules! impl_format_simd_generic_128 {
                 let mut scratch_ptr = scratch.as_mut_ptr().add(16 - nb);
                 v = StrVector::from_slice(core::slice::from_raw_parts(scratch_ptr, STRIDE));
                 let mut mask =
-                    (v.simd_eq(BLASH) | v.simd_eq(QUOTE) | v.simd_lt(X20)).to_bitmask() as u32;
+                    (v.simd_eq(blash) | v.simd_eq(quote) | v.simd_lt(x20)).to_bitmask() as u32;
 
-                while nb > 0 {
-                    v.copy_to_slice(core::slice::from_raw_parts_mut($dst, STRIDE));
-                    if unlikely!(mask > 0) {
+                loop {
+                    v.copy_to_slice(core::slice::from_raw_parts_mut(dst, STRIDE));
+                    if mask != 0 {
                         let cn = trailing_zeros!(mask) as usize;
                         nb -= cn;
-                        $dst = $dst.add(cn);
+                        dst = dst.add(cn);
                         scratch_ptr = scratch_ptr.add(cn);
                         nb -= 1;
                         mask >>= cn + 1;
-                        let escape = QUOTE_TAB[*(scratch_ptr) as usize];
-                        write_escape!(escape, $dst);
-                        $dst = $dst.add(escape.1 as usize);
+                        write_escape!(*(scratch_ptr), dst);
                         scratch_ptr = scratch_ptr.add(1);
                         v = StrVector::from_slice(core::slice::from_raw_parts(scratch_ptr, STRIDE));
                     } else {
-                        $dst = $dst.add(nb);
+                        dst = dst.add(nb);
                         break;
                     }
                 }
             }
         }
-    };
-}
-
-#[allow(dead_code)]
-#[inline(never)]
-#[cfg_attr(target_arch = "x86_64", target_feature(enable = "sse2,bmi1"))]
-#[cfg_attr(target_arch = "aarch64", target_feature(enable = "neon"))]
-pub unsafe fn format_escaped_str_impl_generic_128(
-    odst: *mut u8,
-    value_ptr: *const u8,
-    value_len: usize,
-) -> usize {
-    unsafe {
-        const STRIDE: usize = 16;
-        type StrVector = core::simd::u8x16;
-
-        let mut dst = odst;
-        let mut src = value_ptr;
-
-        core::ptr::write(dst, b'"');
-        dst = dst.add(1);
-
-        if value_len < STRIDE {
-            impl_format_scalar!(dst, src, value_len)
-        } else {
-            impl_format_simd_generic_128!(dst, src, value_len);
-        }
 
         core::ptr::write(dst, b'"');
         dst = dst.add(1);
diff --git a/src/serialize/writer/str/mod.rs b/src/serialize/writer/str/mod.rs
index 6b18f066..81d49af8 100644
--- a/src/serialize/writer/str/mod.rs
+++ b/src/serialize/writer/str/mod.rs
@@ -5,25 +5,24 @@ mod escape;
 #[macro_use]
 mod scalar;
 
+#[cfg(all(feature = "generic_simd", not(target_arch = "x86_64")))]
+mod generic;
+
 #[cfg(target_arch = "x86_64")]
 mod sse2;
 
-#[cfg(all(feature = "unstable-simd", target_arch = "x86_64", feature = "avx512"))]
+#[cfg(all(target_arch = "x86_64", feature = "avx512"))]
 mod avx512;
 
-#[cfg(feature = "unstable-simd")]
-mod generic;
-
-#[cfg(all(not(feature = "unstable-simd"), not(target_arch = "x86_64")))]
+#[cfg(all(not(target_arch = "x86_64"), not(feature = "generic_simd")))]
 pub use scalar::format_escaped_str_scalar;
 
-#[allow(unused_imports)]
-#[cfg(feature = "unstable-simd")]
-pub use generic::format_escaped_str_impl_generic_128;
-
-#[cfg(all(feature = "unstable-simd", target_arch = "x86_64", feature = "avx512"))]
+#[cfg(all(target_arch = "x86_64", feature = "avx512"))]
 pub use avx512::format_escaped_str_impl_512vl;
 
 #[allow(unused_imports)]
 #[cfg(target_arch = "x86_64")]
 pub use sse2::format_escaped_str_impl_sse2_128;
+
+#[cfg(all(feature = "generic_simd", not(target_arch = "x86_64")))]
+pub use generic::format_escaped_str_impl_generic_128;
diff --git a/src/serialize/writer/str/scalar.rs b/src/serialize/writer/str/scalar.rs
index 29150809..75cfc9c0 100644
--- a/src/serialize/writer/str/scalar.rs
+++ b/src/serialize/writer/str/scalar.rs
@@ -1,8 +1,5 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-#[cfg(all(not(feature = "unstable-simd"), not(target_arch = "x86_64")))]
-use super::escape::{NEED_ESCAPED, QUOTE_TAB};
-
 macro_rules! impl_format_scalar {
     ($dst:expr, $src:expr, $value_len:expr) => {
         unsafe {
@@ -10,17 +7,16 @@ macro_rules! impl_format_scalar {
                 core::ptr::write($dst, *($src));
                 $src = $src.add(1);
                 $dst = $dst.add(1);
-                if unlikely!(NEED_ESCAPED[*($src.sub(1)) as usize] > 0) {
-                    let escape = QUOTE_TAB[*($src.sub(1)) as usize];
-                    write_escape!(escape, $dst.sub(1));
-                    $dst = $dst.add(escape.1 as usize - 1);
+                if *super::escape::NEED_ESCAPED.get_unchecked(*($src.sub(1)) as usize) != 0 {
+                    $dst = $dst.sub(1);
+                    write_escape!(*($src.sub(1)), $dst);
                 }
             }
         }
     };
 }
 
-#[cfg(all(not(feature = "unstable-simd"), not(target_arch = "x86_64")))]
+#[cfg(all(not(target_arch = "x86_64"), not(feature = "generic_simd")))]
 pub unsafe fn format_escaped_str_scalar(
     odst: *mut u8,
     value_ptr: *const u8,
diff --git a/src/serialize/writer/str/sse2.rs b/src/serialize/writer/str/sse2.rs
index 43def30a..2150d546 100644
--- a/src/serialize/writer/str/sse2.rs
+++ b/src/serialize/writer/str/sse2.rs
@@ -1,9 +1,5 @@
 // SPDX-License-Identifier: Apache-2.0
 
-use super::escape::{NEED_ESCAPED, QUOTE_TAB};
-
-use core::mem::transmute;
-
 use core::arch::x86_64::{
     __m128i, _mm_cmpeq_epi8, _mm_loadu_si128, _mm_movemask_epi8, _mm_or_si128, _mm_set1_epi8,
     _mm_setzero_si128, _mm_storeu_si128, _mm_subs_epu8,
@@ -11,57 +7,70 @@ use core::arch::x86_64::{
 
 macro_rules! splat_mm128 {
     ($val:expr) => {
-        _mm_set1_epi8(transmute::<u8, i8>($val))
+        _mm_set1_epi8(core::mem::transmute::<u8, i8>($val))
     };
 }
 
-macro_rules! impl_format_simd_sse2_128 {
-    ($dst:expr, $src:expr, $value_len:expr) => {
-        let last_stride_src = $src.add($value_len).sub(STRIDE);
-        let mut nb: usize = $value_len;
+#[allow(dead_code)]
+#[expect(clippy::cast_ptr_alignment)]
+#[inline(never)]
+pub unsafe fn format_escaped_str_impl_sse2_128(
+    odst: *mut u8,
+    value_ptr: *const u8,
+    value_len: usize,
+) -> usize {
+    unsafe {
+        const STRIDE: usize = 16;
 
-        assume!($value_len >= STRIDE);
+        let mut dst = odst;
+        let mut src = value_ptr;
 
-        let blash = splat_mm128!(b'\\');
-        let quote = splat_mm128!(b'"');
-        let x20 = splat_mm128!(31);
-        let v0 = _mm_setzero_si128();
+        core::ptr::write(dst, b'"');
+        dst = dst.add(1);
 
-        unsafe {
-            while nb >= STRIDE {
-                let str_vec = _mm_loadu_si128($src as *const __m128i);
+        if value_len < STRIDE {
+            impl_format_scalar!(dst, src, value_len);
+        } else {
+            let blash = splat_mm128!(b'\\');
+            let quote = splat_mm128!(b'"');
+            let x20 = splat_mm128!(31);
+            let v0 = _mm_setzero_si128();
 
-                let mask = _mm_movemask_epi8(_mm_or_si128(
-                    _mm_or_si128(
-                        _mm_cmpeq_epi8(str_vec, blash),
-                        _mm_cmpeq_epi8(str_vec, quote),
-                    ),
-                    _mm_cmpeq_epi8(_mm_subs_epu8(str_vec, x20), v0),
-                )) as u32;
-
-                _mm_storeu_si128($dst as *mut __m128i, str_vec);
-
-                if unlikely!(mask > 0) {
-                    let cn = trailing_zeros!(mask) as usize;
-                    nb -= cn;
-                    $dst = $dst.add(cn);
-                    $src = $src.add(cn);
-                    nb -= 1;
-                    let escape = QUOTE_TAB[*($src) as usize];
-                    write_escape!(escape, $dst);
-                    $dst = $dst.add(escape.1 as usize);
-                    $src = $src.add(1);
-                } else {
-                    nb -= STRIDE;
-                    $dst = $dst.add(STRIDE);
-                    $src = $src.add(STRIDE);
+            let last_stride_src = src.add(value_len).sub(STRIDE);
+            let mut nb: usize = value_len;
+
+            unsafe {
+                while nb >= STRIDE {
+                    let str_vec = _mm_loadu_si128(src.cast::<__m128i>());
+
+                    let mask = _mm_movemask_epi8(_mm_or_si128(
+                        _mm_or_si128(
+                            _mm_cmpeq_epi8(str_vec, blash),
+                            _mm_cmpeq_epi8(str_vec, quote),
+                        ),
+                        _mm_cmpeq_epi8(_mm_subs_epu8(str_vec, x20), v0),
+                    ));
+
+                    _mm_storeu_si128(dst.cast::<__m128i>(), str_vec);
+
+                    if mask != 0 {
+                        let cn = trailing_zeros!(mask) as usize;
+                        nb -= cn;
+                        dst = dst.add(cn);
+                        src = src.add(cn);
+                        nb -= 1;
+                        write_escape!(*(src), dst);
+                        src = src.add(1);
+                    } else {
+                        nb -= STRIDE;
+                        dst = dst.add(STRIDE);
+                        src = src.add(STRIDE);
+                    }
                 }
-            }
 
-            if nb > 0 {
                 let mut scratch: [u8; 32] = [b'a'; 32];
-                let mut str_vec = _mm_loadu_si128(last_stride_src as *const __m128i);
-                _mm_storeu_si128(scratch.as_mut_ptr() as *mut __m128i, str_vec);
+                let mut str_vec = _mm_loadu_si128(last_stride_src.cast::<__m128i>());
+                _mm_storeu_si128(scratch.as_mut_ptr().cast::<__m128i>(), str_vec);
 
                 let mut scratch_ptr = scratch.as_mut_ptr().add(16 - nb);
                 str_vec = _mm_loadu_si128(scratch_ptr as *const __m128i);
@@ -72,54 +81,28 @@ macro_rules! impl_format_simd_sse2_128 {
                         _mm_cmpeq_epi8(str_vec, quote),
                     ),
                     _mm_cmpeq_epi8(_mm_subs_epu8(str_vec, x20), v0),
-                )) as u32;
+                ));
 
-                while nb > 0 {
-                    _mm_storeu_si128($dst as *mut __m128i, str_vec);
+                loop {
+                    _mm_storeu_si128(dst.cast::<__m128i>(), str_vec);
 
-                    if unlikely!(mask > 0) {
+                    if mask != 0 {
                         let cn = trailing_zeros!(mask) as usize;
                         nb -= cn;
-                        $dst = $dst.add(cn);
+                        dst = dst.add(cn);
                         scratch_ptr = scratch_ptr.add(cn);
                         nb -= 1;
                         mask >>= cn + 1;
-                        let escape = QUOTE_TAB[*(scratch_ptr) as usize];
-                        write_escape!(escape, $dst);
-                        $dst = $dst.add(escape.1 as usize);
+                        write_escape!(*(scratch_ptr), dst);
                         scratch_ptr = scratch_ptr.add(1);
                         str_vec = _mm_loadu_si128(scratch_ptr as *const __m128i);
                     } else {
-                        $dst = $dst.add(nb);
+                        dst = dst.add(nb);
                         break;
                     }
                 }
             }
         }
-    };
-}
-
-#[allow(dead_code)]
-#[inline(never)]
-pub unsafe fn format_escaped_str_impl_sse2_128(
-    odst: *mut u8,
-    value_ptr: *const u8,
-    value_len: usize,
-) -> usize {
-    unsafe {
-        const STRIDE: usize = 16;
-
-        let mut dst = odst;
-        let mut src = value_ptr;
-
-        core::ptr::write(dst, b'"');
-        dst = dst.add(1);
-
-        if value_len < STRIDE {
-            impl_format_scalar!(dst, src, value_len)
-        } else {
-            impl_format_simd_sse2_128!(dst, src, value_len);
-        }
 
         core::ptr::write(dst, b'"');
         dst = dst.add(1);
diff --git a/src/str/avx512.rs b/src/str/avx512.rs
index 230a2de1..c52dc21d 100644
--- a/src/str/avx512.rs
+++ b/src/str/avx512.rs
@@ -3,7 +3,7 @@
 use crate::str::pyunicode_new::*;
 
 use core::arch::x86_64::{
-    __m256i, _mm256_and_si256, _mm256_cmpgt_epu8_mask, _mm256_cmpneq_epi8_mask, _mm256_loadu_si256,
+    _mm256_and_si256, _mm256_cmpgt_epu8_mask, _mm256_cmpneq_epi8_mask, _mm256_loadu_epi8,
     _mm256_mask_cmpneq_epi8_mask, _mm256_maskz_loadu_epi8, _mm256_max_epu8, _mm256_set1_epi8,
 };
 
@@ -25,13 +25,13 @@ macro_rules! impl_kind_simd_avx512vl {
 
             let remainder_mask: u32 = !(u32::MAX << remainder);
             let mut str_vec =
-                _mm256_maskz_loadu_epi8(remainder_mask, $buf.as_bytes().as_ptr() as *const i8);
+                _mm256_maskz_loadu_epi8(remainder_mask, $buf.as_bytes().as_ptr().cast::<i8>());
             let sptr = $buf.as_bytes().as_ptr().add(remainder);
 
             for i in 0..num_loops {
                 str_vec = _mm256_max_epu8(
                     str_vec,
-                    _mm256_loadu_si256(sptr.add(STRIDE * i) as *const __m256i),
+                    _mm256_loadu_epi8(sptr.add(STRIDE * i).cast::<i8>()),
                 );
             }
 
@@ -49,7 +49,7 @@ macro_rules! impl_kind_simd_avx512vl {
                     _mm256_and_si256(
                         _mm256_maskz_loadu_epi8(
                             remainder_mask,
-                            $buf.as_bytes().as_ptr() as *const i8
+                            $buf.as_bytes().as_ptr().cast::<i8>()
                         ),
                         multibyte
                     ),
@@ -59,7 +59,7 @@ macro_rules! impl_kind_simd_avx512vl {
                 for i in 0..num_loops {
                     num_chars += popcnt!(_mm256_cmpneq_epi8_mask(
                         _mm256_and_si256(
-                            _mm256_loadu_si256(sptr.add(STRIDE * i) as *const __m256i),
+                            _mm256_loadu_epi8(sptr.add(STRIDE * i).cast::<i8>()),
                             multibyte
                         ),
                         vec_128,
@@ -79,10 +79,7 @@ macro_rules! impl_kind_simd_avx512vl {
 }
 
 #[inline(never)]
-#[cfg_attr(
-    feature = "avx512",
-    target_feature(enable = "avx512f,avx512bw,avx512vl,bmi2")
-)]
+#[target_feature(enable = "avx512f,avx512bw,avx512vl,bmi2")]
 pub unsafe fn create_str_impl_avx512vl(buf: &str) -> *mut pyo3_ffi::PyObject {
     impl_kind_simd_avx512vl!(buf)
 }
@@ -93,10 +90,18 @@ pub fn unicode_from_str(buf: &str) -> *mut pyo3_ffi::PyObject {
         if unlikely!(buf.is_empty()) {
             return use_immortal!(crate::typeref::EMPTY_UNICODE);
         }
+        STR_CREATE_FN(buf)
+    }
+}
+
+pub type StrDeserializer = unsafe fn(&str) -> *mut pyo3_ffi::PyObject;
+
+static mut STR_CREATE_FN: StrDeserializer = super::scalar::str_impl_kind_scalar;
+
+pub fn set_str_create_fn() {
+    unsafe {
         if std::is_x86_feature_detected!("avx512vl") {
-            create_str_impl_avx512vl(buf)
-        } else {
-            super::scalar::unicode_from_str(buf)
+            STR_CREATE_FN = create_str_impl_avx512vl;
         }
     }
 }
diff --git a/src/str/ffi.rs b/src/str/ffi.rs
index 21841a88..4f1c76bd 100644
--- a/src/str/ffi.rs
+++ b/src/str/ffi.rs
@@ -1,7 +1,8 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
+use crate::util::isize_to_usize;
 use core::ffi::c_void;
-use pyo3_ffi::*;
+use pyo3_ffi::{PyASCIIObject, PyCompactUnicodeObject, PyObject, Py_hash_t};
 
 // see unicodeobject.h for documentation
 
@@ -11,12 +12,14 @@ pub fn hash_str(op: *mut PyObject) -> Py_hash_t {
         let data_ptr: *mut c_void = if (*op.cast::<PyASCIIObject>()).compact() == 1
             && (*op.cast::<PyASCIIObject>()).ascii() == 1
         {
-            (op as *mut PyASCIIObject).offset(1) as *mut c_void
+            op.cast::<PyASCIIObject>().offset(1).cast::<c_void>()
         } else {
-            (op as *mut PyCompactUnicodeObject).offset(1) as *mut c_void
+            op.cast::<PyCompactUnicodeObject>()
+                .offset(1)
+                .cast::<c_void>()
         };
         let num_bytes =
-            (*(op as *mut PyASCIIObject)).length * ((*(op as *mut PyASCIIObject)).kind()) as isize;
+            (*op.cast::<PyASCIIObject>()).length * ((*op.cast::<PyASCIIObject>()).kind()) as isize;
         #[cfg(Py_3_14)]
         let hash = pyo3_ffi::Py_HashBuffer(data_ptr, num_bytes);
         #[cfg(not(Py_3_14))]
@@ -29,11 +32,11 @@ pub fn hash_str(op: *mut PyObject) -> Py_hash_t {
 #[inline(never)]
 pub fn unicode_to_str_via_ffi(op: *mut PyObject) -> Option<&'static str> {
     let mut str_size: pyo3_ffi::Py_ssize_t = 0;
-    let ptr = ffi!(PyUnicode_AsUTF8AndSize(op, &mut str_size)) as *const u8;
+    let ptr = ffi!(PyUnicode_AsUTF8AndSize(op, &mut str_size)).cast::<u8>();
     if unlikely!(ptr.is_null()) {
         None
     } else {
-        Some(str_from_slice!(ptr, str_size as usize))
+        Some(str_from_slice!(ptr, isize_to_usize(str_size)))
     }
 }
 
@@ -44,11 +47,11 @@ pub fn unicode_to_str(op: *mut PyObject) -> Option<&'static str> {
             unicode_to_str_via_ffi(op)
         } else if (*op.cast::<PyASCIIObject>()).ascii() == 1 {
             let ptr = op.cast::<PyASCIIObject>().offset(1) as *const u8;
-            let len = (*op.cast::<PyASCIIObject>()).length as usize;
+            let len = isize_to_usize((*op.cast::<PyASCIIObject>()).length);
             Some(str_from_slice!(ptr, len))
         } else if (*op.cast::<PyCompactUnicodeObject>()).utf8_length != 0 {
             let ptr = (*op.cast::<PyCompactUnicodeObject>()).utf8 as *const u8;
-            let len = (*op.cast::<PyCompactUnicodeObject>()).utf8_length as usize;
+            let len = isize_to_usize((*op.cast::<PyCompactUnicodeObject>()).utf8_length);
             Some(str_from_slice!(ptr, len))
         } else {
             unicode_to_str_via_ffi(op)
diff --git a/src/str/mod.rs b/src/str/mod.rs
index 75149e9b..c71ce296 100644
--- a/src/str/mod.rs
+++ b/src/str/mod.rs
@@ -7,9 +7,9 @@ mod pyunicode_new;
 mod scalar;
 
 #[cfg(not(feature = "avx512"))]
-pub use scalar::unicode_from_str;
+pub use scalar::{set_str_create_fn, unicode_from_str};
 
 #[cfg(feature = "avx512")]
-pub use avx512::unicode_from_str;
+pub use avx512::{set_str_create_fn, unicode_from_str};
 
 pub use ffi::*;
diff --git a/src/str/pyunicode_new.rs b/src/str/pyunicode_new.rs
index b5496891..7b1c2df2 100644
--- a/src/str/pyunicode_new.rs
+++ b/src/str/pyunicode_new.rs
@@ -1,15 +1,30 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use pyo3_ffi::{PyASCIIObject, PyCompactUnicodeObject};
+use crate::util::usize_to_isize;
+use pyo3_ffi::{PyASCIIObject, PyCompactUnicodeObject, PyObject};
+
+macro_rules! validate_str {
+    ($ptr:expr) => {
+        #[cfg(not(Py_3_12))]
+        debug_assert!((*($ptr.cast::<PyASCIIObject>())).ready() == 1);
+
+        debug_assert!((*($ptr.cast::<PyASCIIObject>())).compact() == 1);
+        debug_assert!((*($ptr.cast::<PyASCIIObject>())).interned() == 0);
+
+        debug_assert!(ffi!(_PyUnicode_CheckConsistency($ptr.cast::<PyObject>(), 1)) == 1);
+    };
+}
 
 #[inline(never)]
 pub fn pyunicode_ascii(buf: *const u8, num_chars: usize) -> *mut pyo3_ffi::PyObject {
     unsafe {
-        let ptr = ffi!(PyUnicode_New(num_chars as isize, 127));
-        let data_ptr = ptr.cast::<PyASCIIObject>().offset(1) as *mut u8;
+        let ptr = ffi!(PyUnicode_New(usize_to_isize(num_chars), 127));
+        let data_ptr = ptr.cast::<PyASCIIObject>().offset(1).cast::<u8>();
         core::ptr::copy_nonoverlapping(buf, data_ptr, num_chars);
         core::ptr::write(data_ptr.add(num_chars), 0);
-        ptr
+        debug_assert!((*(ptr.cast::<PyASCIIObject>())).ascii() == 1);
+        validate_str!(ptr);
+        ptr.cast::<PyObject>()
     }
 }
 
@@ -17,41 +32,44 @@ pub fn pyunicode_ascii(buf: *const u8, num_chars: usize) -> *mut pyo3_ffi::PyObj
 #[inline(never)]
 pub fn pyunicode_onebyte(buf: &str, num_chars: usize) -> *mut pyo3_ffi::PyObject {
     unsafe {
-        let ptr = ffi!(PyUnicode_New(num_chars as isize, 255));
-        let mut data_ptr = ptr.cast::<PyCompactUnicodeObject>().offset(1) as *mut u8;
+        let ptr = ffi!(PyUnicode_New(usize_to_isize(num_chars), 255));
+        let mut data_ptr = ptr.cast::<PyCompactUnicodeObject>().offset(1).cast::<u8>();
         for each in buf.chars().fuse() {
             core::ptr::write(data_ptr, each as u8);
             data_ptr = data_ptr.offset(1);
         }
         core::ptr::write(data_ptr, 0);
-        ptr
+        validate_str!(ptr);
+        ptr.cast::<PyObject>()
     }
 }
 
 #[inline(never)]
 pub fn pyunicode_twobyte(buf: &str, num_chars: usize) -> *mut pyo3_ffi::PyObject {
     unsafe {
-        let ptr = ffi!(PyUnicode_New(num_chars as isize, 65535));
-        let mut data_ptr = ptr.cast::<PyCompactUnicodeObject>().offset(1) as *mut u16;
+        let ptr = ffi!(PyUnicode_New(usize_to_isize(num_chars), 65535));
+        let mut data_ptr = ptr.cast::<PyCompactUnicodeObject>().offset(1).cast::<u16>();
         for each in buf.chars().fuse() {
             core::ptr::write(data_ptr, each as u16);
             data_ptr = data_ptr.offset(1);
         }
         core::ptr::write(data_ptr, 0);
-        ptr
+        validate_str!(ptr);
+        ptr.cast::<PyObject>()
     }
 }
 
 #[inline(never)]
 pub fn pyunicode_fourbyte(buf: &str, num_chars: usize) -> *mut pyo3_ffi::PyObject {
     unsafe {
-        let ptr = ffi!(PyUnicode_New(num_chars as isize, 1114111));
-        let mut data_ptr = ptr.cast::<PyCompactUnicodeObject>().offset(1) as *mut u32;
+        let ptr = ffi!(PyUnicode_New(usize_to_isize(num_chars), 1114111));
+        let mut data_ptr = ptr.cast::<PyCompactUnicodeObject>().offset(1).cast::<u32>();
         for each in buf.chars().fuse() {
             core::ptr::write(data_ptr, each as u32);
             data_ptr = data_ptr.offset(1);
         }
         core::ptr::write(data_ptr, 0);
-        ptr
+        validate_str!(ptr);
+        ptr.cast::<PyObject>()
     }
 }
diff --git a/src/str/scalar.rs b/src/str/scalar.rs
index 0aa84c7d..12adb634 100644
--- a/src/str/scalar.rs
+++ b/src/str/scalar.rs
@@ -1,10 +1,15 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
-use crate::str::pyunicode_new::*;
-use crate::typeref::EMPTY_UNICODE;
+use crate::str::pyunicode_new::{
+    pyunicode_ascii, pyunicode_fourbyte, pyunicode_onebyte, pyunicode_twobyte,
+};
 
-#[inline(always)]
-pub fn str_impl_kind_scalar(buf: &str, num_chars: usize) -> *mut pyo3_ffi::PyObject {
+#[inline(never)]
+pub fn str_impl_kind_scalar(buf: &str) -> *mut pyo3_ffi::PyObject {
+    let num_chars = bytecount::num_chars(buf.as_bytes());
+    if buf.len() == num_chars {
+        return pyunicode_ascii(buf.as_ptr(), num_chars);
+    }
     unsafe {
         let len = buf.len();
         assume!(len > 0);
@@ -31,15 +36,14 @@ pub fn str_impl_kind_scalar(buf: &str, num_chars: usize) -> *mut pyo3_ffi::PyObj
     }
 }
 
-#[inline(never)]
+#[cfg(not(feature = "avx512"))]
+#[inline(always)]
 pub fn unicode_from_str(buf: &str) -> *mut pyo3_ffi::PyObject {
     if unlikely!(buf.is_empty()) {
-        return use_immortal!(EMPTY_UNICODE);
-    }
-    let num_chars = bytecount::num_chars(buf.as_bytes());
-    if buf.len() == num_chars {
-        pyunicode_ascii(buf.as_ptr(), num_chars)
-    } else {
-        str_impl_kind_scalar(buf, num_chars)
+        return use_immortal!(crate::typeref::EMPTY_UNICODE);
     }
+    str_impl_kind_scalar(buf)
 }
+
+#[cfg(not(feature = "avx512"))]
+pub fn set_str_create_fn() {}
diff --git a/src/typeref.rs b/src/typeref.rs
index f9c15e37..bf925a4d 100644
--- a/src/typeref.rs
+++ b/src/typeref.rs
@@ -1,16 +1,17 @@
 // SPDX-License-Identifier: (Apache-2.0 OR MIT)
 
 use crate::ffi::orjson_fragmenttype_new;
-use core::ffi::c_char;
 #[cfg(feature = "yyjson")]
 use core::ffi::c_void;
+use core::ffi::CStr;
 #[cfg(feature = "yyjson")]
 use core::mem::MaybeUninit;
 use core::ptr::{null_mut, NonNull};
 use once_cell::race::{OnceBool, OnceBox};
-use pyo3_ffi::*;
+
 #[cfg(feature = "yyjson")]
-use std::cell::UnsafeCell;
+use core::cell::UnsafeCell;
+use pyo3_ffi::*;
 
 pub struct NumpyTypes {
     pub array: *mut PyTypeObject,
@@ -58,7 +59,6 @@ pub static mut FRAGMENT_TYPE: *mut PyTypeObject = null_mut();
 
 pub static mut NUMPY_TYPES: OnceBox<Option<NonNull<NumpyTypes>>> = OnceBox::new();
 
-#[cfg(Py_3_9)]
 pub static mut ZONEINFO_TYPE: *mut PyTypeObject = null_mut();
 
 pub static mut UTCOFFSET_METHOD_STR: *mut PyObject = null_mut();
@@ -97,7 +97,7 @@ pub fn yyjson_init() -> Box<YYJSONAlloc> {
     // Using unsafe to ensure allocation happens on the heap without going through the stack
     // so we don't stack overflow in debug mode. Once rust-lang/rust#63291 is stable (Box::new_uninit)
     // we can use that instead.
-    let layout = std::alloc::Layout::new::<YYJSONBuffer>();
+    let layout = core::alloc::Layout::new::<YYJSONBuffer>();
     let buffer = unsafe { Box::from_raw(std::alloc::alloc(layout).cast::<YYJSONBuffer>()) };
     let mut alloc = crate::ffi::yyjson::yyjson_alc {
         malloc: None,
@@ -135,66 +135,60 @@ pub fn init_typerefs() {
 #[cfg_attr(feature = "optimize", optimize(size))]
 fn _init_typerefs_impl() -> bool {
     unsafe {
-        debug_assert!(crate::opt::MAX_OPT < u16::MAX as i32);
+        debug_assert!(crate::opt::MAX_OPT < i32::from(u16::MAX));
 
         assert!(crate::deserialize::KEY_MAP
             .set(crate::deserialize::KeyMap::default())
             .is_ok());
-        FRAGMENT_TYPE = orjson_fragmenttype_new();
-        PyDateTime_IMPORT();
+
+        crate::serialize::writer::set_str_formatter_fn();
+        crate::str::set_str_create_fn();
+
         NONE = Py_None();
         TRUE = Py_True();
         FALSE = Py_False();
         EMPTY_UNICODE = PyUnicode_New(0, 255);
-        STR_TYPE = (*EMPTY_UNICODE).ob_type;
-        BYTES_TYPE = (*PyBytes_FromStringAndSize("".as_ptr() as *const c_char, 0)).ob_type;
 
-        {
-            let bytearray = PyByteArray_FromStringAndSize("".as_ptr() as *const c_char, 0);
-            BYTEARRAY_TYPE = (*bytearray).ob_type;
+        STR_TYPE = &raw mut PyUnicode_Type;
+        BYTES_TYPE = &raw mut PyBytes_Type;
+        DICT_TYPE = &raw mut PyDict_Type;
+        LIST_TYPE = &raw mut PyList_Type;
+        TUPLE_TYPE = &raw mut PyTuple_Type;
+        NONE_TYPE = (*NONE).ob_type;
+        BOOL_TYPE = &raw mut PyBool_Type;
+        INT_TYPE = &raw mut PyLong_Type;
+        FLOAT_TYPE = &raw mut PyFloat_Type;
+        BYTEARRAY_TYPE = &raw mut PyByteArray_Type;
+        MEMORYVIEW_TYPE = &raw mut PyMemoryView_Type;
 
-            let memoryview = PyMemoryView_FromObject(bytearray);
-            MEMORYVIEW_TYPE = (*memoryview).ob_type;
-            Py_DECREF(memoryview);
-            Py_DECREF(bytearray);
-        }
+        PyDateTime_IMPORT();
 
-        DICT_TYPE = (*PyDict_New()).ob_type;
-        LIST_TYPE = (*PyList_New(0)).ob_type;
-        TUPLE_TYPE = (*PyTuple_New(0)).ob_type;
-        NONE_TYPE = (*NONE).ob_type;
-        BOOL_TYPE = (*TRUE).ob_type;
-        INT_TYPE = (*PyLong_FromLongLong(0)).ob_type;
-        FLOAT_TYPE = (*PyFloat_FromDouble(0.0)).ob_type;
         DATETIME_TYPE = look_up_datetime_type();
         DATE_TYPE = look_up_date_type();
         TIME_TYPE = look_up_time_type();
         UUID_TYPE = look_up_uuid_type();
         ENUM_TYPE = look_up_enum_type();
-        FIELD_TYPE = look_up_field_type();
 
-        #[cfg(Py_3_9)]
-        {
-            ZONEINFO_TYPE = look_up_zoneinfo_type();
-        }
+        FRAGMENT_TYPE = orjson_fragmenttype_new();
 
-        INT_ATTR_STR = PyUnicode_InternFromString("int\0".as_ptr() as *const c_char);
-        UTCOFFSET_METHOD_STR = PyUnicode_InternFromString("utcoffset\0".as_ptr() as *const c_char);
-        NORMALIZE_METHOD_STR = PyUnicode_InternFromString("normalize\0".as_ptr() as *const c_char);
-        CONVERT_METHOD_STR = PyUnicode_InternFromString("convert\0".as_ptr() as *const c_char);
-        DST_STR = PyUnicode_InternFromString("dst\0".as_ptr() as *const c_char);
-        DICT_STR = PyUnicode_InternFromString("__dict__\0".as_ptr() as *const c_char);
-        DATACLASS_FIELDS_STR =
-            PyUnicode_InternFromString("__dataclass_fields__\0".as_ptr() as *const c_char);
-        SLOTS_STR = PyUnicode_InternFromString("__slots__\0".as_ptr() as *const c_char);
-        FIELD_TYPE_STR = PyUnicode_InternFromString("_field_type\0".as_ptr() as *const c_char);
-        ARRAY_STRUCT_STR =
-            PyUnicode_InternFromString("__array_struct__\0".as_ptr() as *const c_char);
-        DTYPE_STR = PyUnicode_InternFromString("dtype\0".as_ptr() as *const c_char);
-        DESCR_STR = PyUnicode_InternFromString("descr\0".as_ptr() as *const c_char);
-        VALUE_STR = PyUnicode_InternFromString("value\0".as_ptr() as *const c_char);
-        DEFAULT = PyUnicode_InternFromString("default\0".as_ptr() as *const c_char);
-        OPTION = PyUnicode_InternFromString("option\0".as_ptr() as *const c_char);
+        FIELD_TYPE = look_up_field_type();
+        ZONEINFO_TYPE = look_up_zoneinfo_type();
+
+        INT_ATTR_STR = PyUnicode_InternFromString(c"int".as_ptr());
+        UTCOFFSET_METHOD_STR = PyUnicode_InternFromString(c"utcoffset".as_ptr());
+        NORMALIZE_METHOD_STR = PyUnicode_InternFromString(c"normalize".as_ptr());
+        CONVERT_METHOD_STR = PyUnicode_InternFromString(c"convert".as_ptr());
+        DST_STR = PyUnicode_InternFromString(c"dst".as_ptr());
+        DICT_STR = PyUnicode_InternFromString(c"__dict__".as_ptr());
+        DATACLASS_FIELDS_STR = PyUnicode_InternFromString(c"__dataclass_fields__".as_ptr());
+        SLOTS_STR = PyUnicode_InternFromString(c"__slots__".as_ptr());
+        FIELD_TYPE_STR = PyUnicode_InternFromString(c"_field_type".as_ptr());
+        ARRAY_STRUCT_STR = PyUnicode_InternFromString(c"__array_struct__".as_ptr());
+        DTYPE_STR = PyUnicode_InternFromString(c"dtype".as_ptr());
+        DESCR_STR = PyUnicode_InternFromString(c"descr".as_ptr());
+        VALUE_STR = PyUnicode_InternFromString(c"value".as_ptr());
+        DEFAULT = PyUnicode_InternFromString(c"default".as_ptr());
+        OPTION = PyUnicode_InternFromString(c"option".as_ptr());
         JsonEncodeError = pyo3_ffi::PyExc_TypeError;
         Py_INCREF(JsonEncodeError);
         JsonDecodeError = look_up_json_exc();
@@ -206,15 +200,10 @@ fn _init_typerefs_impl() -> bool {
 #[cfg_attr(feature = "optimize", optimize(size))]
 unsafe fn look_up_json_exc() -> *mut PyObject {
     unsafe {
-        let module = PyImport_ImportModule("json\0".as_ptr() as *const c_char);
+        let module = PyImport_ImportModule(c"json".as_ptr());
         let module_dict = PyObject_GenericGetDict(module, null_mut());
-        let ptr =
-            PyMapping_GetItemString(module_dict, "JSONDecodeError\0".as_ptr() as *const c_char);
-        let res = pyo3_ffi::PyErr_NewException(
-            "orjson.JSONDecodeError\0".as_ptr() as *const c_char,
-            ptr,
-            null_mut(),
-        );
+        let ptr = PyMapping_GetItemString(module_dict, c"JSONDecodeError".as_ptr());
+        let res = pyo3_ffi::PyErr_NewException(c"orjson.JSONDecodeError".as_ptr(), ptr, null_mut());
         Py_DECREF(ptr);
         Py_DECREF(module_dict);
         Py_DECREF(module);
@@ -225,11 +214,14 @@ unsafe fn look_up_json_exc() -> *mut PyObject {
 
 #[cold]
 #[cfg_attr(feature = "optimize", optimize(size))]
-unsafe fn look_up_numpy_type(numpy_module_dict: *mut PyObject, np_type: &str) -> *mut PyTypeObject {
+unsafe fn look_up_numpy_type(
+    numpy_module_dict: *mut PyObject,
+    np_type: &CStr,
+) -> *mut PyTypeObject {
     unsafe {
-        let ptr = PyMapping_GetItemString(numpy_module_dict, np_type.as_ptr() as *const c_char);
+        let ptr = PyMapping_GetItemString(numpy_module_dict, np_type.as_ptr());
         Py_XDECREF(ptr);
-        ptr as *mut PyTypeObject
+        ptr.cast::<PyTypeObject>()
     }
 }
 
@@ -237,27 +229,27 @@ unsafe fn look_up_numpy_type(numpy_module_dict: *mut PyObject, np_type: &str) ->
 #[cfg_attr(feature = "optimize", optimize(size))]
 pub fn load_numpy_types() -> Box<Option<NonNull<NumpyTypes>>> {
     unsafe {
-        let numpy = PyImport_ImportModule("numpy\0".as_ptr() as *const c_char);
+        let numpy = PyImport_ImportModule(c"numpy".as_ptr());
         if numpy.is_null() {
             PyErr_Clear();
             return Box::new(None);
         }
         let numpy_module_dict = PyObject_GenericGetDict(numpy, null_mut());
         let types = Box::new(NumpyTypes {
-            array: look_up_numpy_type(numpy_module_dict, "ndarray\0"),
-            float16: look_up_numpy_type(numpy_module_dict, "half\0"),
-            float32: look_up_numpy_type(numpy_module_dict, "float32\0"),
-            float64: look_up_numpy_type(numpy_module_dict, "float64\0"),
-            int8: look_up_numpy_type(numpy_module_dict, "int8\0"),
-            int16: look_up_numpy_type(numpy_module_dict, "int16\0"),
-            int32: look_up_numpy_type(numpy_module_dict, "int32\0"),
-            int64: look_up_numpy_type(numpy_module_dict, "int64\0"),
-            uint16: look_up_numpy_type(numpy_module_dict, "uint16\0"),
-            uint32: look_up_numpy_type(numpy_module_dict, "uint32\0"),
-            uint64: look_up_numpy_type(numpy_module_dict, "uint64\0"),
-            uint8: look_up_numpy_type(numpy_module_dict, "uint8\0"),
-            bool_: look_up_numpy_type(numpy_module_dict, "bool_\0"),
-            datetime64: look_up_numpy_type(numpy_module_dict, "datetime64\0"),
+            array: look_up_numpy_type(numpy_module_dict, c"ndarray"),
+            float16: look_up_numpy_type(numpy_module_dict, c"half"),
+            float32: look_up_numpy_type(numpy_module_dict, c"float32"),
+            float64: look_up_numpy_type(numpy_module_dict, c"float64"),
+            int8: look_up_numpy_type(numpy_module_dict, c"int8"),
+            int16: look_up_numpy_type(numpy_module_dict, c"int16"),
+            int32: look_up_numpy_type(numpy_module_dict, c"int32"),
+            int64: look_up_numpy_type(numpy_module_dict, c"int64"),
+            uint16: look_up_numpy_type(numpy_module_dict, c"uint16"),
+            uint32: look_up_numpy_type(numpy_module_dict, c"uint32"),
+            uint64: look_up_numpy_type(numpy_module_dict, c"uint64"),
+            uint8: look_up_numpy_type(numpy_module_dict, c"uint8"),
+            bool_: look_up_numpy_type(numpy_module_dict, c"bool_"),
+            datetime64: look_up_numpy_type(numpy_module_dict, c"datetime64"),
         });
         Py_XDECREF(numpy_module_dict);
         Py_XDECREF(numpy);
@@ -269,10 +261,9 @@ pub fn load_numpy_types() -> Box<Option<NonNull<NumpyTypes>>> {
 #[cfg_attr(feature = "optimize", optimize(size))]
 unsafe fn look_up_field_type() -> *mut PyTypeObject {
     unsafe {
-        let module = PyImport_ImportModule("dataclasses\0".as_ptr() as *const c_char);
+        let module = PyImport_ImportModule(c"dataclasses".as_ptr());
         let module_dict = PyObject_GenericGetDict(module, null_mut());
-        let ptr = PyMapping_GetItemString(module_dict, "_FIELD\0".as_ptr() as *const c_char)
-            as *mut PyTypeObject;
+        let ptr = PyMapping_GetItemString(module_dict, c"_FIELD".as_ptr()).cast::<PyTypeObject>();
         Py_DECREF(module_dict);
         Py_DECREF(module);
         ptr
@@ -283,10 +274,9 @@ unsafe fn look_up_field_type() -> *mut PyTypeObject {
 #[cfg_attr(feature = "optimize", optimize(size))]
 unsafe fn look_up_enum_type() -> *mut PyTypeObject {
     unsafe {
-        let module = PyImport_ImportModule("enum\0".as_ptr() as *const c_char);
+        let module = PyImport_ImportModule(c"enum".as_ptr());
         let module_dict = PyObject_GenericGetDict(module, null_mut());
-        let ptr = PyMapping_GetItemString(module_dict, "EnumMeta\0".as_ptr() as *const c_char)
-            as *mut PyTypeObject;
+        let ptr = PyMapping_GetItemString(module_dict, c"EnumMeta".as_ptr()).cast::<PyTypeObject>();
         Py_DECREF(module_dict);
         Py_DECREF(module);
         ptr
@@ -297,10 +287,9 @@ unsafe fn look_up_enum_type() -> *mut PyTypeObject {
 #[cfg_attr(feature = "optimize", optimize(size))]
 unsafe fn look_up_uuid_type() -> *mut PyTypeObject {
     unsafe {
-        let uuid_mod = PyImport_ImportModule("uuid\0".as_ptr() as *const c_char);
+        let uuid_mod = PyImport_ImportModule(c"uuid".as_ptr());
         let uuid_mod_dict = PyObject_GenericGetDict(uuid_mod, null_mut());
-        let uuid =
-            PyMapping_GetItemString(uuid_mod_dict, "NAMESPACE_DNS\0".as_ptr() as *const c_char);
+        let uuid = PyMapping_GetItemString(uuid_mod_dict, c"NAMESPACE_DNS".as_ptr());
         let ptr = (*uuid).ob_type;
         Py_DECREF(uuid);
         Py_DECREF(uuid_mod_dict);
@@ -353,15 +342,13 @@ unsafe fn look_up_time_type() -> *mut PyTypeObject {
     }
 }
 
-#[cfg(Py_3_9)]
 #[cold]
 #[cfg_attr(feature = "optimize", optimize(size))]
 unsafe fn look_up_zoneinfo_type() -> *mut PyTypeObject {
     unsafe {
-        let module = PyImport_ImportModule("zoneinfo\0".as_ptr() as *const c_char);
+        let module = PyImport_ImportModule(c"zoneinfo".as_ptr());
         let module_dict = PyObject_GenericGetDict(module, null_mut());
-        let ptr = PyMapping_GetItemString(module_dict, "ZoneInfo\0".as_ptr() as *const c_char)
-            as *mut PyTypeObject;
+        let ptr = PyMapping_GetItemString(module_dict, c"ZoneInfo".as_ptr()).cast::<PyTypeObject>();
         Py_DECREF(module_dict);
         Py_DECREF(module);
         ptr
diff --git a/src/util.rs b/src/util.rs
index 1a0b630a..ccf425ce 100644
--- a/src/util.rs
+++ b/src/util.rs
@@ -47,7 +47,7 @@ macro_rules! is_subclass_by_flag {
 macro_rules! is_subclass_by_type {
     ($ob_type:expr, $type:ident) => {
         unsafe {
-            (*($ob_type as *mut pyo3_ffi::PyTypeObject))
+            (*($ob_type.cast::<pyo3_ffi::PyTypeObject>()))
                 .ob_base
                 .ob_base
                 .ob_type
@@ -112,7 +112,7 @@ macro_rules! nonnull {
 
 macro_rules! str_from_slice {
     ($ptr:expr, $size:expr) => {
-        unsafe { std::str::from_utf8_unchecked(core::slice::from_raw_parts($ptr, $size as usize)) }
+        unsafe { core::str::from_utf8_unchecked(core::slice::from_raw_parts($ptr, $size as usize)) }
     };
 }
 
@@ -168,7 +168,6 @@ macro_rules! ffi {
     };
 }
 
-#[cfg(Py_3_9)]
 macro_rules! call_method {
     ($obj1:expr, $obj2:expr) => {
         unsafe { pyo3_ffi::PyObject_CallMethodNoArgs($obj1, $obj2) }
@@ -178,29 +177,6 @@ macro_rules! call_method {
     };
 }
 
-#[cfg(not(Py_3_9))]
-macro_rules! call_method {
-    ($obj1:expr, $obj2:expr) => {
-        unsafe {
-            pyo3_ffi::PyObject_CallMethodObjArgs(
-                $obj1,
-                $obj2,
-                core::ptr::null_mut() as *mut pyo3_ffi::PyObject,
-            )
-        }
-    };
-    ($obj1:expr, $obj2:expr, $obj3:expr) => {
-        unsafe {
-            pyo3_ffi::PyObject_CallMethodObjArgs(
-                $obj1,
-                $obj2,
-                $obj3,
-                core::ptr::null_mut() as *mut pyo3_ffi::PyObject,
-            )
-        }
-    };
-}
-
 macro_rules! str_hash {
     ($op:expr) => {
         unsafe { (*$op.cast::<pyo3_ffi::PyASCIIObject>()).hash }
@@ -278,6 +254,28 @@ macro_rules! pydict_next {
     };
 }
 
+macro_rules! pydict_setitem {
+    ($dict:expr, $pykey:expr, $pyval:expr) => {
+        debug_assert!(ffi!(Py_REFCNT($dict)) == 1);
+        debug_assert!(str_hash!($pykey) != -1);
+        #[cfg(not(Py_3_13))]
+        unsafe {
+            let _ = pyo3_ffi::_PyDict_SetItem_KnownHash($dict, $pykey, $pyval, str_hash!($pykey));
+        }
+        #[cfg(Py_3_13)]
+        unsafe {
+            let _ = pyo3_ffi::_PyDict_SetItem_KnownHash_LockHeld(
+                $dict.cast::<pyo3_ffi::PyDictObject>(),
+                $pykey,
+                $pyval,
+                str_hash!($pykey),
+            );
+        }
+        reverse_pydict_incref!($pykey);
+        reverse_pydict_incref!($pyval);
+    };
+}
+
 macro_rules! reserve_minimum {
     ($writer:expr) => {
         $writer.reserve(64);
@@ -337,3 +335,17 @@ macro_rules! unreachable_unchecked {
         unsafe { core::hint::unreachable_unchecked() }
     };
 }
+
+#[inline(always)]
+#[allow(clippy::cast_possible_wrap)]
+pub fn usize_to_isize(val: usize) -> isize {
+    debug_assert!(val < (isize::MAX as usize));
+    val as isize
+}
+
+#[inline(always)]
+#[allow(clippy::cast_sign_loss)]
+pub fn isize_to_usize(val: isize) -> usize {
+    debug_assert!(val >= 0);
+    val as usize
+}
diff --git a/test/requirements.txt b/test/requirements.txt
index 97bee4ca..d5916c84 100644
--- a/test/requirements.txt
+++ b/test/requirements.txt
@@ -1,8 +1,7 @@
-arrow
 faker
 numpy;(platform_machine=="x86_64" or (platform_machine=="aarch64" and sys_platform == "linux")) and python_version<"3.13"
-pendulum;sys_platform=="linux" and platform_machine=="x86_64" and python_version<"3.12"
-time-machine < 2.15;sys_platform=="linux" and platform_machine=="x86_64" and python_version<"3.12"
+pendulum;sys_platform=="linux" and platform_machine=="x86_64" and python_version<"3.14"
 psutil;(sys_platform=="linux" or sys_platform == "macos") and platform_machine=="x86_64" and python_version<"3.13"
 pytest
+python-dateutil >=2,<3;python_version<"3.14"
 pytz
diff --git a/test/test_api.py b/test/test_api.py
index f4078b5b..58e5d4fa 100644
--- a/test/test_api.py
+++ b/test/test_api.py
@@ -79,7 +79,7 @@ def test_loads_recursion_valid_limit_mixed(self):
         loads() recursion limit at limit mixed
         """
         n = LOADS_RECURSION_LIMIT
-        value = b"[" b'{"key":' * n + b'{"key":true}' + b"}" * n + b"]"
+        value = b"".join((b"[", b'{"key":' * n, b'{"key":true}' + b"}" * n, b"]"))
         pytest.raises(orjson.JSONDecodeError, orjson.loads, value)
 
     def test_loads_recursion_valid_excessive_array(self):
@@ -111,7 +111,7 @@ def test_loads_recursion_valid_limit_mixed_pretty(self):
         loads() recursion limit at limit mixed pretty
         """
         n = LOADS_RECURSION_LIMIT
-        value = b"[\n  " b'{"key":' * n + b'{"key":true}' + b"}" * n + b"]"
+        value = b'[\n  {"key":' * n + b'{"key":true}' + b"}" * n + b"]"
         pytest.raises(orjson.JSONDecodeError, orjson.loads, value)
 
     def test_loads_recursion_valid_excessive_array_pretty(self):
diff --git a/test/test_escape.py b/test/test_escape.py
new file mode 100644
index 00000000..30f1c153
--- /dev/null
+++ b/test/test_escape.py
@@ -0,0 +1,126 @@
+# SPDX-License-Identifier: (Apache-2.0 OR MIT)
+
+import orjson
+
+
+def test_issue565():
+    assert (
+        orjson.dumps("\n\r\u000b\f\u001c\u001d\u001e")
+        == b'"\\n\\r\\u000b\\f\\u001c\\u001d\\u001e"'
+    )
+
+
+def test_0x00():
+    assert orjson.dumps("\u0000") == b'"\\u0000"'
+
+
+def test_0x01():
+    assert orjson.dumps("\u0001") == b'"\\u0001"'
+
+
+def test_0x02():
+    assert orjson.dumps("\u0002") == b'"\\u0002"'
+
+
+def test_0x03():
+    assert orjson.dumps("\u0003") == b'"\\u0003"'
+
+
+def test_0x04():
+    assert orjson.dumps("\u0004") == b'"\\u0004"'
+
+
+def test_0x05():
+    assert orjson.dumps("\u0005") == b'"\\u0005"'
+
+
+def test_0x06():
+    assert orjson.dumps("\u0006") == b'"\\u0006"'
+
+
+def test_0x07():
+    assert orjson.dumps("\u0007") == b'"\\u0007"'
+
+
+def test_0x08():
+    assert orjson.dumps("\u0008") == b'"\\b"'
+
+
+def test_0x09():
+    assert orjson.dumps("\u0009") == b'"\\t"'
+
+
+def test_0x0a():
+    assert orjson.dumps("\u000a") == b'"\\n"'
+
+
+def test_0x0b():
+    assert orjson.dumps("\u000b") == b'"\\u000b"'
+
+
+def test_0x0c():
+    assert orjson.dumps("\u000c") == b'"\\f"'
+
+
+def test_0x0d():
+    assert orjson.dumps("\u000d") == b'"\\r"'
+
+
+def test_0x0e():
+    assert orjson.dumps("\u000e") == b'"\\u000e"'
+
+
+def test_0x0f():
+    assert orjson.dumps("\u000f") == b'"\\u000f"'
+
+
+def test_0x10():
+    assert orjson.dumps("\u0010") == b'"\\u0010"'
+
+
+def test_0x11():
+    assert orjson.dumps("\u0011") == b'"\\u0011"'
+
+
+def test_0x12():
+    assert orjson.dumps("\u0012") == b'"\\u0012"'
+
+
+def test_0x13():
+    assert orjson.dumps("\u0013") == b'"\\u0013"'
+
+
+def test_0x14():
+    assert orjson.dumps("\u0014") == b'"\\u0014"'
+
+
+def test_0x15():
+    assert orjson.dumps("\u0015") == b'"\\u0015"'
+
+
+def test_0x16():
+    assert orjson.dumps("\u0016") == b'"\\u0016"'
+
+
+def test_0x17():
+    assert orjson.dumps("\u0017") == b'"\\u0017"'
+
+
+def test_0x18():
+    assert orjson.dumps("\u0018") == b'"\\u0018"'
+
+
+def test_0x19():
+    assert orjson.dumps("\u0019") == b'"\\u0019"'
+
+
+def test_0x1a():
+    assert orjson.dumps("\u001a") == b'"\\u001a"'
+
+
+def test_backslash():
+    assert orjson.dumps("\\") == b'"\\\\"'
+
+
+def test_quote():
+    assert orjson.dumps('"') == b'"\\""'