Merge pull request #4175 from laytan/riscv-compiler

Support RISC-V for the compiler itself
Merge pull request #4176 from Feoramund/fix-context-error-msg
2026-07-05 19:21:37 -07:00 · 2024-09-02 00:20:04 +01:00 · 2024-09-02 00:18:15 +01:00 · 2024-09-01 17:50:50 -04:00 · 2024-09-01 21:42:47 +02:00 · 2024-09-01 17:22:58 +02:00
952 changed files with 126864 additions and 29157 deletions
@@ -1,2 +1,6 @@
 *.odin linguist-language=Odin
-* text=auto
+* text=auto
+
+# These files must always have *nix line-endings
+Makefile text eol=lf
+*.sh text eol=lf
@@ -2,147 +2,157 @@ name: CI
 on: [push, pull_request, workflow_dispatch]

 jobs:
-  build_linux:
-    name: Ubuntu Build, Check, and Test
+  build_netbsd:
+    name: NetBSD Build, Check, and Test
+    runs-on: ubuntu-latest
+    env:
+      PKGSRC_BRANCH: 2024Q2
+    steps:
+    - uses: actions/checkout@v4
+    - name: Build, Check, and Test
+      timeout-minutes: 15
+      uses: vmactions/netbsd-vm@v1
+      with:
+        release: "10.0"
+        envs: PKGSRC_BRANCH
+        usesh: true
+        copyback: false
+        prepare: |
+          PKG_PATH="https://cdn.NetBSD.org/pub/pkgsrc/packages/NetBSD/amd64/$(uname -r | cut -d_ -f1)_${PKGSRC_BRANCH}/All" /usr/sbin/pkg_add pkgin
+          pkgin -y in gmake git bash python311 llvm clang
+          ln -s /usr/pkg/bin/python3.11 /usr/bin/python3
+        run: |
+          set -e -x
+          git config --global --add safe.directory $(pwd)
+          gmake release
+          ./odin version
+          ./odin report
+          gmake -C vendor/stb/src
+          gmake -C vendor/cgltf/src
+          gmake -C vendor/miniaudio/src
+          ./odin check examples/all -vet -strict-style -disallow-do -target:netbsd_amd64
+          ./odin check examples/all -vet -strict-style -disallow-do -target:netbsd_arm64
+          ./odin test tests/core/normal.odin -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+          ./odin test tests/core/speed.odin -file -all-packages -vet -strict-style -disallow-do -o:speed -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+          ./odin test tests/vendor -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+          (cd tests/issues; ./run.sh)
+  build_freebsd:
+    name: FreeBSD Build, Check, and Test
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v1
-      - name: Download LLVM
+    - uses: actions/checkout@v4
+    - name: Build, Check, and Test
+      timeout-minutes: 15
+      uses: vmactions/freebsd-vm@v1
+      with:
+        usesh: true
+        copyback: false
+        prepare: |
+          pkg install -y gmake git bash python3 libxml2 llvm17
+        run: |
+          # `set -e` is needed for test failures to register. https://github.com/vmactions/freebsd-vm/issues/72
+          set -e -x
+          git config --global --add safe.directory $(pwd)
+          gmake release
+          ./odin version
+          ./odin report
+          gmake -C vendor/stb/src
+          gmake -C vendor/cgltf/src
+          gmake -C vendor/miniaudio/src
+          ./odin check examples/all -vet -strict-style -disallow-do -target:freebsd_amd64
+          ./odin test tests/core/normal.odin -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+          ./odin test tests/core/speed.odin -file -all-packages -vet -strict-style -disallow-do -o:speed -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+          ./odin test tests/vendor -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+          (cd tests/issues; ./run.sh)
+  ci:
+    strategy:
+      fail-fast: false
+      matrix:
+        # MacOS 13 runs on Intel, 14 runs on ARM
+        os: [ubuntu-latest, macos-13, macos-14]
+    runs-on: ${{ matrix.os }}
+    name: ${{ matrix.os == 'macos-14' && 'MacOS ARM' || (matrix.os == 'macos-13' && 'MacOS Intel' || 'Ubuntu') }} Build, Check, and Test
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Download LLVM (Linux)
+        if: matrix.os == 'ubuntu-latest'
        run: |
          wget https://apt.llvm.org/llvm.sh
          chmod +x llvm.sh
          sudo ./llvm.sh 17
          echo "/usr/lib/llvm-17/bin" >> $GITHUB_PATH
-      - name: build odin
-        run: ./build_odin.sh release
-      - name: Odin version
-        run: ./odin version
-        timeout-minutes: 1
-      - name: Odin report
-        run: ./odin report
-        timeout-minutes: 1
-      - name: Odin check
-        run: ./odin check examples/demo -vet
-        timeout-minutes: 10
-      - name: Odin run
-        run: ./odin run examples/demo
-        timeout-minutes: 10
-      - name: Odin run -debug
-        run: ./odin run examples/demo -debug
-        timeout-minutes: 10
-      - name: Odin check examples/all
-        run: ./odin check examples/all -strict-style
-        timeout-minutes: 10
-      - name: Core library tests
+
+      - name: Download LLVM (MacOS Intel)
+        if: matrix.os == 'macos-13'
        run: |
-          cd tests/core
-          make
-        timeout-minutes: 10
-      - name: Vendor library tests
-        run: |
-          cd tests/vendor
-          make
-        timeout-minutes: 10
-      - name: Odin internals tests
-        run: |
-          cd tests/internal
-          make
-        timeout-minutes: 10
-      - name: Odin check examples/all for Linux i386
-        run: ./odin check examples/all -vet -strict-style -target:linux_i386
-        timeout-minutes: 10
-      - name: Odin check examples/all for Linux arm64
-        run: ./odin check examples/all -vet -strict-style -target:linux_arm64
-        timeout-minutes: 10
-      - name: Odin check examples/all for FreeBSD amd64
-        run: ./odin check examples/all -vet -strict-style -target:freebsd_amd64
-        timeout-minutes: 10
-      - name: Odin check examples/all for OpenBSD amd64
-        run: ./odin check examples/all -vet -strict-style -target:openbsd_amd64
-        timeout-minutes: 10
-  build_macOS:
-    name: MacOS Build, Check, and Test
-    runs-on: macos-latest
-    steps:
-      - uses: actions/checkout@v1
-      - name: Download LLVM, and setup PATH
-        run: |
-          brew install llvm@17
+          brew install llvm@17 lua@5.4
          echo "/usr/local/opt/llvm@17/bin" >> $GITHUB_PATH
-      - name: build odin
-        run: ./build_odin.sh release
-      - name: Odin version
-        run: ./odin version
-        timeout-minutes: 1
-      - name: Odin report
-        run: ./odin report
-        timeout-minutes: 1
-      - name: Odin check
-        run: ./odin check examples/demo -vet
-        timeout-minutes: 10
-      - name: Odin run
-        run: ./odin run examples/demo
-        timeout-minutes: 10
-      - name: Odin run -debug
-        run: ./odin run examples/demo -debug
-        timeout-minutes: 10
-      - name: Odin check examples/all
-        run: ./odin check examples/all -strict-style
-        timeout-minutes: 10
-      - name: Core library tests
+
+      - name: Download LLVM (MacOS ARM)
+        if: matrix.os == 'macos-14'
        run: |
-          cd tests/core
-          make
-        timeout-minutes: 10
-      - name: Odin internals tests
-        run: |
-          cd tests/internal
-          make
-        timeout-minutes: 10
-  build_macOS_arm:
-    name: MacOS ARM Build, Check, and Test
-    runs-on: macos-14 # This is an arm/m1 runner.
-    steps:
-      - uses: actions/checkout@v1
-      - name: Download LLVM and setup PATH
-        run: |
-          brew install llvm@17
+          brew install llvm@17 wasmtime lua@5.4
          echo "/opt/homebrew/opt/llvm@17/bin" >> $GITHUB_PATH
-      - name: build odin
+
+      - name: Build Odin
        run: ./build_odin.sh release
      - name: Odin version
        run: ./odin version
-        timeout-minutes: 1
      - name: Odin report
        run: ./odin report
-        timeout-minutes: 1
+      - name: Compile needed Vendor
+        run: |
+          make -C vendor/stb/src
+          make -C vendor/cgltf/src
+          make -C vendor/miniaudio/src
      - name: Odin check
        run: ./odin check examples/demo -vet
-        timeout-minutes: 10
      - name: Odin run
        run: ./odin run examples/demo
-        timeout-minutes: 10
      - name: Odin run -debug
        run: ./odin run examples/demo -debug
-        timeout-minutes: 10
      - name: Odin check examples/all
        run: ./odin check examples/all -strict-style
-        timeout-minutes: 10
-      - name: Core library tests
+      - name: Normal Core library tests
+        run: ./odin test tests/core/normal.odin -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+      - name: Optimized Core library tests
+        run: ./odin test tests/core/speed.odin -o:speed -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+      - name: Vendor library tests
+        run: ./odin test tests/vendor -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+      - name: Internals tests
+        run: ./odin test tests/internal -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+      - name: GitHub Issue tests
        run: |
-          cd tests/core
-          make
-        timeout-minutes: 10
-      - name: Odin internals tests
+          cd tests/issues
+          ./run.sh
+
+      - name: Odin check examples/all for Linux i386
+        run: ./odin check examples/all -vet -strict-style -disallow-do -target:linux_i386
+        if: matrix.os == 'ubuntu-latest'
+      - name: Odin check examples/all for Linux arm64
+        run: ./odin check examples/all -vet -strict-style -disallow-do -target:linux_arm64
+        if: matrix.os == 'ubuntu-latest'
+      - name: Odin check examples/all for FreeBSD amd64
+        run: ./odin check examples/all -vet -strict-style -disallow-do -target:freebsd_amd64
+        if: matrix.os == 'ubuntu-latest'
+      - name: Odin check examples/all for OpenBSD amd64
+        run: ./odin check examples/all -vet -strict-style -disallow-do -target:openbsd_amd64
+        if: matrix.os == 'ubuntu-latest'
+
+      - name: Run demo on WASI WASM32
        run: |
-          cd tests/internal
-          make
-        timeout-minutes: 10
+          ./odin build examples/demo -target:wasi_wasm32 -vet -strict-style -disallow-do -out:demo.wasm
+          wasmtime ./demo.wasm
+        if: matrix.os == 'macos-14'
+
  build_windows:
    name: Windows Build, Check, and Test
    runs-on: windows-2022
+    timeout-minutes: 15
    steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@v4
      - name: build Odin
        shell: cmd
        run: |
@@ -150,72 +160,113 @@ jobs:
          ./build.bat 1
      - name: Odin version
        run: ./odin version
-        timeout-minutes: 1
      - name: Odin report
        run: ./odin report
-        timeout-minutes: 1
      - name: Odin check
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
          odin check examples/demo -vet
-        timeout-minutes: 10
      - name: Odin run
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
          odin run examples/demo
-        timeout-minutes: 10
      - name: Odin run -debug
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
-          odin run examples/demo -debug
-        timeout-minutes: 10
+          odin run examples/demo -debug -vet -strict-style -disallow-do
      - name: Odin check examples/all
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
-          odin check examples/all -strict-style
-        timeout-minutes: 10
+          odin check examples/all -vet -strict-style -disallow-do
      - name: Core library tests
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
-          cd tests\core
-          call build.bat
-        timeout-minutes: 10
+          odin test tests/core/normal.odin -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
+      - name: Optimized core library tests
+        shell: cmd
+        run: |
+          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
+          odin test tests/core/speed.odin -o:speed -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
      - name: Vendor library tests
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
-          cd tests\vendor
-          call build.bat
-        timeout-minutes: 10
+          copy vendor\lua\5.4\windows\*.dll .
+          odin test tests/vendor -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
      - name: Odin internals tests
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
-          cd tests\internal
-          call build.bat
-        timeout-minutes: 10
+          odin test tests/internal -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true
      - name: Odin documentation tests
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
          cd tests\documentation
-          rem call build.bat
-        timeout-minutes: 10
+          call build.bat
      - name: core:math/big tests
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
          cd tests\core\math\big
          call build.bat
-        timeout-minutes: 10
      - name: Odin check examples/all for Windows 32bits
        shell: cmd
        run: |
          call "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Auxiliary\Build\vcvars64.bat
          odin check examples/all -strict-style -target:windows_i386
-        timeout-minutes: 10
+
+  build_linux_riscv64:
+    runs-on: ubuntu-latest
+    name: Linux riscv64 (emulated) Build, Check and Test
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Download LLVM (Linux)
+        run: |
+          wget https://apt.llvm.org/llvm.sh
+          chmod +x llvm.sh
+          sudo ./llvm.sh 18
+          echo "/usr/lib/llvm-18/bin" >> $GITHUB_PATH
+
+      - name: Build Odin
+        run: ./build_odin.sh release
+
+      - name: Odin version
+        run: ./odin version
+
+      - name: Odin report
+        run: ./odin report
+
+      - name: Compile needed Vendor
+        run: |
+          make -C vendor/stb/src
+          make -C vendor/cgltf/src
+          make -C vendor/miniaudio/src
+
+      - name: Odin check
+        run: ./odin check examples/all -target:linux_riscv64 -vet -strict-style -disallow-do
+
+      - name: Install riscv64 toolchain and qemu
+        run: sudo apt-get install -y qemu-user qemu-user-static gcc-12-riscv64-linux-gnu libc6-riscv64-cross
+
+      - name: Odin run
+        run: ./odin run examples/demo -vet -strict-style -disallow-do -target:linux_riscv64 -extra-linker-flags:"-fuse-ld=/usr/bin/riscv64-linux-gnu-gcc-12 -static -Wl,-static"
+
+      - name: Odin run -debug
+        run: ./odin run examples/demo -debug -vet -strict-style -disallow-do -target:linux_riscv64 -extra-linker-flags:"-fuse-ld=/usr/bin/riscv64-linux-gnu-gcc-12 -static -Wl,-static"
+
+      - name: Normal Core library tests
+        run: ./odin test tests/core/normal.odin -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true -target:linux_riscv64 -extra-linker-flags:"-fuse-ld=/usr/bin/riscv64-linux-gnu-gcc-12 -static -Wl,-static"
+
+      - name: Optimized Core library tests
+        run: ./odin test tests/core/speed.odin -o:speed -file -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true -target:linux_riscv64 -extra-linker-flags:"-fuse-ld=/usr/bin/riscv64-linux-gnu-gcc-12 -static -Wl,-static"
+
+      - name: Internals tests
+        run: ./odin test tests/internal -all-packages -vet -strict-style -disallow-do -define:ODIN_TEST_FANCY=false -define:ODIN_TEST_FAIL_ON_BAD_MEMORY=true -target:linux_riscv64 -extra-linker-flags:"-fuse-ld=/usr/bin/riscv64-linux-gnu-gcc-12 -static -Wl,-static"
@@ -11,7 +11,7 @@ jobs:
    if: github.repository == 'odin-lang/Odin'
    runs-on: windows-2022
    steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@v4
      - name: build Odin
        shell: cmd
        run: |
@@ -45,13 +45,13 @@ jobs:
    if: github.repository == 'odin-lang/Odin'
    runs-on: ubuntu-latest
    steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@v4
      - name: (Linux) Download LLVM
        run: |
          wget https://apt.llvm.org/llvm.sh
          chmod +x llvm.sh
-          sudo ./llvm.sh 17
-          echo "/usr/lib/llvm-17/bin" >> $GITHUB_PATH
+          sudo ./llvm.sh 18
+          echo "/usr/lib/llvm-18/bin" >> $GITHUB_PATH
      - name: build odin
        run: make nightly
      - name: Odin run
@@ -61,7 +61,6 @@ jobs:
          mkdir dist
          cp odin dist
          cp LICENSE dist
-          cp libLLVM* dist
          cp -r shared dist
          cp -r base dist
          cp -r core dist
@@ -77,13 +76,13 @@ jobs:
  build_macos:
    name: MacOS Build
    if: github.repository == 'odin-lang/Odin'
-    runs-on: macos-latest
+    runs-on: macos-13
    steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@v4
      - name: Download LLVM and setup PATH
        run: |
-          brew install llvm@17 dylibbundler
-          echo "/usr/local/opt/llvm@17/bin" >> $GITHUB_PATH
+          brew install llvm@18 dylibbundler
+          echo "/usr/local/opt/llvm@18/bin" >> $GITHUB_PATH
      - name: build odin
        # These -L makes the linker prioritize system libraries over LLVM libraries, this is mainly to
        # not link with libunwind bundled with LLVM but link with libunwind on the system.
@@ -113,11 +112,11 @@ jobs:
    if: github.repository == 'odin-lang/Odin'
    runs-on: macos-14 # ARM machine
    steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@v4
      - name: Download LLVM and setup PATH
        run: |
-          brew install llvm@17 dylibbundler
-          echo "/opt/homebrew/opt/llvm@17/bin" >> $GITHUB_PATH
+          brew install llvm@18 dylibbundler
+          echo "/opt/homebrew/opt/llvm@18/bin" >> $GITHUB_PATH
      - name: build odin
        # These -L makes the linker prioritize system libraries over LLVM libraries, this is mainly to
        # not link with libunwind bundled with LLVM but link with libunwind on the system.
@@ -146,16 +145,16 @@ jobs:
    runs-on: [ubuntu-latest]
    needs: [build_windows, build_macos, build_macos_arm, build_ubuntu]
    steps:
-      - uses: actions/checkout@v1
+      - uses: actions/checkout@v4
      - uses: actions/setup-python@v2
        with:
          python-version: '3.8.x'

-      - name: Install B2 CLI
+      - name: Install B2 SDK
        shell: bash
        run: |
          python -m pip install --upgrade pip
-          pip install --upgrade b2
+          pip install --upgrade b2sdk

      - name: Display Python version
        run: python -c "import sys; print(sys.version)"
@@ -188,24 +187,9 @@ jobs:
          BUCKET: ${{ secrets.B2_BUCKET }}
          DAYS_TO_KEEP: ${{ secrets.B2_DAYS_TO_KEEP }}
        run: |
-          echo Authorizing B2 account
-          b2 authorize-account "$APPID" "$APPKEY"
-
-          echo Uploading artifcates to B2
-          chmod +x ./ci/upload_create_nightly.sh
-          ./ci/upload_create_nightly.sh "$BUCKET" windows-amd64 windows_artifacts/
-          ./ci/upload_create_nightly.sh "$BUCKET" ubuntu-amd64 ubuntu_artifacts/dist.zip
-          ./ci/upload_create_nightly.sh "$BUCKET" macos-amd64 macos_artifacts/dist.zip
-          ./ci/upload_create_nightly.sh "$BUCKET" macos-arm64 macos_arm_artifacts/dist.zip
-
-          echo Deleting old artifacts in B2
-          python3 ci/delete_old_binaries.py "$BUCKET" "$DAYS_TO_KEEP"
-
-          echo Creating nightly.json
-          python3 ci/create_nightly_json.py "$BUCKET" > nightly.json
-
-          echo Uploading nightly.json
-          b2 upload-file "$BUCKET" nightly.json nightly.json
-
-          echo Clear B2 account info
-          b2 clear-account
+          python3 ci/nightly.py artifact windows-amd64 windows_artifacts/
+          python3 ci/nightly.py artifact ubuntu-amd64 ubuntu_artifacts/dist.zip
+          python3 ci/nightly.py artifact macos-amd64 macos_artifacts/dist.zip
+          python3 ci/nightly.py artifact macos-arm64 macos_arm_artifacts/dist.zip
+          python3 ci/nightly.py prune
+          python3 ci/nightly.py json
@@ -17,41 +17,12 @@
 [Rr]eleases/
 x64/
 x86/
+!/core/simd/x86
 bld/
 [Bb]in/
 [Oo]bj/
 [Ll]og/
 ![Cc]ore/[Ll]og/
-tests/documentation/verify/
-tests/documentation/all.odin-doc
-tests/internal/test_map
-tests/internal/test_pow
-tests/internal/test_rtti
-tests/core/test_core_compress
-tests/core/test_core_container
-tests/core/test_core_filepath
-tests/core/test_core_fmt
-tests/core/test_core_i18n
-tests/core/test_core_image
-tests/core/test_core_libc
-tests/core/test_core_match
-tests/core/test_core_math
-tests/core/test_core_net
-tests/core/test_core_os_exit
-tests/core/test_core_reflect
-tests/core/test_core_strings
-tests/core/test_crypto
-tests/core/test_hash
-tests/core/test_hxa
-tests/core/test_json
-tests/core/test_linalg_glsl_math
-tests/core/test_noise
-tests/core/test_varint
-tests/core/test_xml
-tests/core/test_core_slice
-tests/core/test_core_thread
-tests/core/test_core_runtime
-tests/vendor/vendor_botan
 # Visual Studio 2015 cache/options directory
 .vs/
 # Visual Studio Code options directory
@@ -299,7 +270,7 @@ bin/
 # - Linux/MacOS
 odin
 !odin/
-odin.dSYM
+**/*.dSYM
 *.bin
 demo.bin
 libLLVM*.so*
@@ -318,4 +289,6 @@ build.sh
 !core/debug/

 # RAD debugger project file
-*.raddbg
+*.raddbg
+
+misc/featuregen/featuregen
@@ -1,4 +1,4 @@
-all: debug
+all: default

 demo:
 	./odin run examples/demo/demo.odin -file
@@ -6,12 +6,18 @@ demo:
 report:
 	./odin report

+default:
+	PROGRAM=make ./build_odin.sh # debug
+
 debug:
 	./build_odin.sh debug

 release:
 	./build_odin.sh release

+release-native:
+	./build_odin.sh release-native
+
 release_native:
 	./build_odin.sh release-native

@@ -76,9 +76,9 @@ Answers to common questions about Odin.

 Documentation for all the official packages part of the [core](https://pkg.odin-lang.org/core/) and [vendor](https://pkg.odin-lang.org/vendor/) library collections.

-#### [The Odin Wiki](https://github.com/odin-lang/Odin/wiki)
+#### [Odin Documentation](https://odin-lang.org/docs/)

-A wiki maintained by the Odin community.
+Documentation for the Odin language itself.

 #### [Odin Discord](https://discord.gg/sVBPHEv)

@@ -126,3 +126,5 @@ clamp :: proc(value, minimum, maximum: T) -> T ---

 soa_zip :: proc(slices: ...) -> #soa[]Struct ---
 soa_unzip :: proc(value: $S/#soa[]$E) -> (slices: ...) ---
+
+unreachable :: proc() -> ! ---
@@ -38,9 +38,12 @@ count_leading_zeros  :: proc(x: $T) -> T where type_is_integer(T) || type_is_sim
 reverse_bits         :: proc(x: $T) -> T where type_is_integer(T) || type_is_simd_vector(T) ---
 byte_swap            :: proc(x: $T) -> T where type_is_integer(T) || type_is_float(T) ---

-overflow_add :: proc(lhs, rhs: $T) -> (T, bool) ---
-overflow_sub :: proc(lhs, rhs: $T) -> (T, bool) ---
-overflow_mul :: proc(lhs, rhs: $T) -> (T, bool) ---
+overflow_add :: proc(lhs, rhs: $T) -> (T, bool) where type_is_integer(T) #optional_ok ---
+overflow_sub :: proc(lhs, rhs: $T) -> (T, bool) where type_is_integer(T) #optional_ok ---
+overflow_mul :: proc(lhs, rhs: $T) -> (T, bool) where type_is_integer(T) #optional_ok ---
+
+saturating_add :: proc(lhs, rhs: $T) -> T where type_is_integer(T) ---
+saturating_sub :: proc(lhs, rhs: $T) -> T where type_is_integer(T) ---

 sqrt :: proc(x: $T) -> T where type_is_float(T) || (type_is_simd_vector(T) && type_is_float(type_elem_type(T))) ---

@@ -73,6 +76,8 @@ expect :: proc(val, expected_val: T) -> T ---

 // Linux and Darwin Only
 syscall :: proc(id: uintptr, args: ..uintptr) -> uintptr ---
+// FreeBSD, NetBSD, et cetera
+syscall_bsd :: proc(id: uintptr, args: ..uintptr) -> (uintptr, bool) ---


 // Atomics
@@ -167,17 +172,23 @@ type_is_matrix           :: proc($T: typeid) -> bool ---

 type_has_nil :: proc($T: typeid) -> bool ---

+type_is_matrix_row_major    :: proc($T: typeid) -> bool where type_is_matrix(T) ---
+type_is_matrix_column_major :: proc($T: typeid) -> bool where type_is_matrix(T) ---
+
 type_is_specialization_of :: proc($T, $S: typeid) -> bool ---

-type_is_variant_of :: proc($U, $V: typeid) -> bool where type_is_union(U) ---
-type_union_tag_type :: proc($T: typeid) -> typeid where type_is_union(T) ---
-type_union_tag_offset :: proc($T: typeid) -> uintptr where type_is_union(T) ---
-type_union_base_tag_value :: proc($T: typeid) -> int where type_is_union(U) ---
-type_union_variant_count :: proc($T: typeid) -> int where type_is_union(T) ---
-type_variant_type_of :: proc($T: typeid, $index: int) -> typeid where type_is_union(T) ---
-type_variant_index_of :: proc($U, $V: typeid) -> int where type_is_union(U) ---
+type_is_variant_of        :: proc($U, $V: typeid)          -> bool    where type_is_union(U) ---
+type_union_tag_type       :: proc($T: typeid)              -> typeid  where type_is_union(T) ---
+type_union_tag_offset     :: proc($T: typeid)              -> uintptr where type_is_union(T) ---
+type_union_base_tag_value :: proc($T: typeid)              -> int     where type_is_union(U) ---
+type_union_variant_count  :: proc($T: typeid)              -> int     where type_is_union(T) ---
+type_variant_type_of      :: proc($T: typeid, $index: int) -> typeid  where type_is_union(T) ---
+type_variant_index_of     :: proc($U, $V: typeid)          -> int     where type_is_union(U) ---

-type_has_field :: proc($T: typeid, $name: string) -> bool ---
+type_bit_set_elem_type       :: proc($T: typeid) -> typeid where type_is_bit_set(T) ---
+type_bit_set_underlying_type :: proc($T: typeid) -> typeid where type_is_bit_set(T) ---
+
+type_has_field  :: proc($T: typeid, $name: string) -> bool ---
 type_field_type :: proc($T: typeid, $name: string) -> typeid ---

 type_proc_parameter_count :: proc($T: typeid) -> int where type_is_proc(T) ---
@@ -186,7 +197,8 @@ type_proc_return_count    :: proc($T: typeid) -> int where type_is_proc(T) ---
 type_proc_parameter_type  :: proc($T: typeid, index: int) -> typeid where type_is_proc(T) ---
 type_proc_return_type     :: proc($T: typeid, index: int) -> typeid where type_is_proc(T) ---

-type_struct_field_count :: proc($T: typeid) -> int where type_is_struct(T) ---
+type_struct_field_count          :: proc($T: typeid) -> int  where type_is_struct(T) ---
+type_struct_has_implicit_padding :: proc($T: typeid) -> bool where type_is_struct(T) ---

 type_polymorphic_record_parameter_count :: proc($T: typeid) -> typeid ---
 type_polymorphic_record_parameter_value :: proc($T: typeid, index: int) -> $V ---
@@ -207,14 +219,21 @@ type_map_cell_info :: proc($T: typeid)           -> ^runtime.Map_Cell_Info ---
 type_convert_variants_to_pointers :: proc($T: typeid) -> typeid where type_is_union(T) ---
 type_merge :: proc($U, $V: typeid) -> typeid where type_is_union(U), type_is_union(V) ---

+type_has_shared_fields :: proc($U, $V: typeid) -> bool typeid where type_is_struct(U), type_is_struct(V) ---
+
 constant_utf16_cstring :: proc($literal: string) -> [^]u16 ---

+constant_log2 :: proc($v: $T) -> T where type_is_integer(T) ---
+
 // SIMD related
 simd_add  :: proc(a, b: #simd[N]T) -> #simd[N]T ---
 simd_sub  :: proc(a, b: #simd[N]T) -> #simd[N]T ---
 simd_mul  :: proc(a, b: #simd[N]T) -> #simd[N]T ---
 simd_div  :: proc(a, b: #simd[N]T) -> #simd[N]T where type_is_float(T) ---

+simd_saturating_add  :: proc(a, b: #simd[N]T) -> #simd[N]T where type_is_integer(T) ---
+simd_saturating_sub  :: proc(a, b: #simd[N]T) -> #simd[N]T where type_is_integer(T) ---
+
 // Keeps Odin's Behaviour
 // (x << y) if y <= mask else 0
 simd_shl :: proc(a: #simd[N]T, b: #simd[N]Unsigned_Integer) -> #simd[N]T ---
@@ -225,9 +244,6 @@ simd_shr :: proc(a: #simd[N]T, b: #simd[N]Unsigned_Integer) -> #simd[N]T ---
 simd_shl_masked :: proc(a: #simd[N]T, b: #simd[N]Unsigned_Integer) -> #simd[N]T ---
 simd_shr_masked :: proc(a: #simd[N]T, b: #simd[N]Unsigned_Integer) -> #simd[N]T ---

-simd_add_sat :: proc(a, b: #simd[N]T) -> #simd[N]T ---
-simd_sub_sat :: proc(a, b: #simd[N]T) -> #simd[N]T ---
-
 simd_bit_and     :: proc(a, b: #simd[N]T) -> #simd[N]T ---
 simd_bit_or      :: proc(a, b: #simd[N]T) -> #simd[N]T ---
 simd_bit_xor     :: proc(a, b: #simd[N]T) -> #simd[N]T ---
@@ -256,13 +272,28 @@ simd_lanes_ge :: proc(a, b: #simd[N]T) -> #simd[N]Integer ---
 simd_extract :: proc(a: #simd[N]T, idx: uint) -> T ---
 simd_replace :: proc(a: #simd[N]T, idx: uint, elem: T) -> #simd[N]T ---

-simd_reduce_add_ordered :: proc(a: #simd[N]T) -> T ---
-simd_reduce_mul_ordered :: proc(a: #simd[N]T) -> T ---
-simd_reduce_min         :: proc(a: #simd[N]T) -> T ---
-simd_reduce_max         :: proc(a: #simd[N]T) -> T ---
-simd_reduce_and         :: proc(a: #simd[N]T) -> T ---
-simd_reduce_or          :: proc(a: #simd[N]T) -> T ---
-simd_reduce_xor         :: proc(a: #simd[N]T) -> T ---
+simd_reduce_add_ordered :: proc(a: #simd[N]T) -> T where type_is_integer(T) || type_is_float(T)---
+simd_reduce_mul_ordered :: proc(a: #simd[N]T) -> T where type_is_integer(T) || type_is_float(T)---
+simd_reduce_min         :: proc(a: #simd[N]T) -> T where type_is_integer(T) || type_is_float(T)---
+simd_reduce_max         :: proc(a: #simd[N]T) -> T where type_is_integer(T) || type_is_float(T)---
+simd_reduce_and         :: proc(a: #simd[N]T) -> T where type_is_integer(T) || type_is_float(T)---
+simd_reduce_or          :: proc(a: #simd[N]T) -> T where type_is_integer(T) || type_is_float(T)---
+simd_reduce_xor         :: proc(a: #simd[N]T) -> T where type_is_integer(T) || type_is_float(T)---
+
+simd_reduce_any         :: proc(a: #simd[N]T) -> T where type_is_boolean(T) ---
+simd_reduce_all         :: proc(a: #simd[N]T) -> T where type_is_boolean(T) ---
+
+
+simd_gather       :: proc(ptr: #simd[N]rawptr, val: #simd[N]T, mask: #simd[N]U) -> #simd[N]T where type_is_integer(U) || type_is_boolean(U) ---
+simd_scatter      :: proc(ptr: #simd[N]rawptr, val: #simd[N]T, mask: #simd[N]U)              where type_is_integer(U) || type_is_boolean(U) ---
+
+simd_masked_load  :: proc(ptr: rawptr, val: #simd[N]T, mask: #simd[N]U) -> #simd[N]T where type_is_integer(U) || type_is_boolean(U) ---
+simd_masked_store :: proc(ptr: rawptr, val: #simd[N]T, mask: #simd[N]U)              where type_is_integer(U) || type_is_boolean(U) ---
+
+simd_masked_expand_load    :: proc(ptr: rawptr, val: #simd[N]T, mask: #simd[N]U) -> #simd[N]T where type_is_integer(U) || type_is_boolean(U) ---
+simd_masked_compress_store :: proc(ptr: rawptr, val: #simd[N]T, mask: #simd[N]U)              where type_is_integer(U) || type_is_boolean(U) ---
+
+

 simd_shuffle :: proc(a, b: #simd[N]T, indices: ..int) -> #simd[len(indices)]T ---
 simd_select  :: proc(cond: #simd[N]boolean_or_integer, true, false: #simd[N]T) -> #simd[N]T ---
@@ -276,12 +307,22 @@ simd_nearest :: proc(a: #simd[N]any_float) -> #simd[N]any_float ---

 simd_to_bits :: proc(v: #simd[N]T) -> #simd[N]Integer where size_of(T) == size_of(Integer), type_is_unsigned(Integer) ---

-// equivalent a swizzle with descending indices, e.g. reserve(a, 3, 2, 1, 0)
-simd_reverse :: proc(a: #simd[N]T) -> #simd[N]T ---
+// equivalent to a swizzle with descending indices, e.g. reserve(a, 3, 2, 1, 0)
+simd_lanes_reverse :: proc(a: #simd[N]T) -> #simd[N]T ---

-simd_rotate_left  :: proc(a: #simd[N]T, $offset: int) -> #simd[N]T ---
-simd_rotate_right :: proc(a: #simd[N]T, $offset: int) -> #simd[N]T ---
+simd_lanes_rotate_left  :: proc(a: #simd[N]T, $offset: int) -> #simd[N]T ---
+simd_lanes_rotate_right :: proc(a: #simd[N]T, $offset: int) -> #simd[N]T ---

+// Checks if the current target supports the given target features.
+//
+// Takes a constant comma-seperated string (eg: "sha512,sse4.1"), or a procedure type which has either
+// `@(require_target_feature)` or `@(enable_target_feature)` as its input and returns a boolean indicating
+// if all listed features are supported.
+has_target_feature :: proc($test: $T) -> bool where type_is_string(T) || type_is_proc(T) ---
+
+
+// Returns the value of the procedure where `x` must be a call expression
+procedure_of :: proc(x: $T) -> T where type_is_proc(T) ---

 // WASM targets only
 wasm_memory_grow :: proc(index, delta: uintptr) -> int ---
@@ -293,9 +334,9 @@ wasm_memory_size :: proc(index: uintptr)        -> int ---
 // 0 - indicates that the thread blocked and then was woken up
 // 1 - the loaded value from `ptr` did not match `expected`, the thread did not block
 // 2 - the thread blocked, but the timeout
-@(enable_target_feature="atomics")
+@(require_target_feature="atomics")
 wasm_memory_atomic_wait32   :: proc(ptr: ^u32, expected: u32, timeout_ns: i64) -> u32 ---
-@(enable_target_feature="atomics")
+@(require_target_feature="atomics")
 wasm_memory_atomic_notify32 :: proc(ptr: ^u32, waiters: u32) -> (waiters_woken_up: u32) ---

 // x86 Targets (i386, amd64)
@@ -66,7 +66,7 @@ Type_Info_Named :: struct {
 	name: string,
 	base: ^Type_Info,
 	pkg:  string,
-	loc:  Source_Code_Location,
+	loc:  ^Source_Code_Location,
 }
 Type_Info_Integer    :: struct {signed: bool, endianness: Platform_Endianness}
 Type_Info_Rune       :: struct {}
@@ -112,23 +112,32 @@ Type_Info_Parameters :: struct { // Only used for procedures parameters and resu
 }
 Type_Info_Tuple :: Type_Info_Parameters // Will be removed eventually

-Type_Info_Struct :: struct {
-	types:        []^Type_Info,
-	names:        []string,
-	offsets:      []uintptr,
-	usings:       []bool,
-	tags:         []string,
-	is_packed:    bool,
-	is_raw_union: bool,
-	is_no_copy:   bool,
-	custom_align: bool,
+Type_Info_Struct_Flags :: distinct bit_set[Type_Info_Struct_Flag; u8]
+Type_Info_Struct_Flag :: enum u8 {
+	packed    = 0,
+	raw_union = 1,
+	no_copy   = 2,
+	align     = 3,
+}

-	equal: Equal_Proc, // set only when the struct has .Comparable set but does not have .Simple_Compare set
+Type_Info_Struct :: struct {
+	// Slice these with `field_count`
+	types:   [^]^Type_Info `fmt:"v,field_count"`,
+	names:   [^]string     `fmt:"v,field_count"`,
+	offsets: [^]uintptr    `fmt:"v,field_count"`,
+	usings:  [^]bool       `fmt:"v,field_count"`,
+	tags:    [^]string     `fmt:"v,field_count"`,
+
+	field_count: i32,
+
+	flags: Type_Info_Struct_Flags,

 	// These are only set iff this structure is an SOA structure
 	soa_kind:      Type_Info_Struct_Soa_Kind,
+	soa_len:       i32,
 	soa_base_type: ^Type_Info,
-	soa_len:       int,
+
+	equal: Equal_Proc, // set only when the struct has .Comparable set but does not have .Simple_Compare set
 }
 Type_Info_Union :: struct {
 	variants:     []^Type_Info,
@@ -142,9 +151,9 @@ Type_Info_Union :: struct {
 	shared_nil:   bool,
 }
 Type_Info_Enum :: struct {
-	base:      ^Type_Info,
-	names:     []string,
-	values:    []Type_Info_Enum_Value,
+	base:   ^Type_Info,
+	names:  []string,
+	values: []Type_Info_Enum_Value,
 }
 Type_Info_Map :: struct {
 	key:      ^Type_Info,
@@ -187,11 +196,12 @@ Type_Info_Soa_Pointer :: struct {
 }
 Type_Info_Bit_Field :: struct {
 	backing_type: ^Type_Info,
-	names:        []string,
-	types:        []^Type_Info,
-	bit_sizes:    []uintptr,
-	bit_offsets:  []uintptr,
-	tags:         []string,
+	names:        [^]string     `fmt:"v,field_count"`,
+	types:        [^]^Type_Info `fmt:"v,field_count"`,
+	bit_sizes:    [^]uintptr    `fmt:"v,field_count"`,
+	bit_offsets:  [^]uintptr    `fmt:"v,field_count"`,
+	tags:         [^]string     `fmt:"v,field_count"`,
+	field_count:  int,
 }

 Type_Info_Flag :: enum u8 {
@@ -273,14 +283,14 @@ Typeid_Kind :: enum u8 {
 }
 #assert(len(Typeid_Kind) < 32)

-// Typeid_Bit_Field :: bit_field #align(align_of(uintptr)) {
-// 	index:    8*size_of(uintptr) - 8,
-// 	kind:     5, // Typeid_Kind
-// 	named:    1,
-// 	special:  1, // signed, cstring, etc
-// 	reserved: 1,
-// }
-// #assert(size_of(Typeid_Bit_Field) == size_of(uintptr));
+Typeid_Bit_Field :: bit_field uintptr {
+	index:    uintptr     | 8*size_of(uintptr) - 8,
+	kind:     Typeid_Kind | 5, // Typeid_Kind
+	named:    bool        | 1,
+	special:  bool        | 1, // signed, cstring, etc
+	reserved: bool        | 1,
+}
+#assert(size_of(Typeid_Bit_Field) == size_of(uintptr))

 // NOTE(bill): only the ones that are needed (not all types)
 // This will be set by the compiler
@@ -299,6 +309,8 @@ when ODIN_OS == .Windows {
 		Thread_Detach  = 3,
 	}
 	dll_forward_reason: DLL_Forward_Reason
+
+	dll_instance: rawptr
 }

 // IMPORTANT NOTE(bill): Must be in this order (as the compiler relies upon it)
@@ -397,11 +409,34 @@ Logger :: struct {
 	options:      Logger_Options,
 }

+
+Random_Generator_Mode :: enum {
+	Read,
+	Reset,
+	Query_Info,
+}
+
+Random_Generator_Query_Info_Flag :: enum u32 {
+	Cryptographic,
+	Uniform,
+	External_Entropy,
+	Resettable,
+}
+Random_Generator_Query_Info :: distinct bit_set[Random_Generator_Query_Info_Flag; u32]
+
+Random_Generator_Proc :: #type proc(data: rawptr, mode: Random_Generator_Mode, p: []byte)
+
+Random_Generator :: struct {
+	procedure: Random_Generator_Proc,
+	data:      rawptr,
+}
+
 Context :: struct {
 	allocator:              Allocator,
 	temp_allocator:         Allocator,
 	assertion_failure_proc: Assertion_Failure_Proc,
 	logger:                 Logger,
+	random_generator:       Random_Generator,

 	user_ptr:   rawptr,
 	user_index: int,
@@ -470,6 +505,15 @@ Raw_Soa_Pointer :: struct {
 	index: int,
 }

+Raw_Complex32     :: struct {real, imag: f16}
+Raw_Complex64     :: struct {real, imag: f32}
+Raw_Complex128    :: struct {real, imag: f64}
+Raw_Quaternion64  :: struct {imag, jmag, kmag: f16, real: f16}
+Raw_Quaternion128 :: struct {imag, jmag, kmag: f32, real: f32}
+Raw_Quaternion256 :: struct {imag, jmag, kmag: f64, real: f64}
+Raw_Quaternion64_Vector_Scalar  :: struct {vector: [3]f16, scalar: f16}
+Raw_Quaternion128_Vector_Scalar :: struct {vector: [3]f32, scalar: f32}
+Raw_Quaternion256_Vector_Scalar :: struct {vector: [3]f64, scalar: f64}


 /*
@@ -482,8 +526,11 @@ Raw_Soa_Pointer :: struct {
 		Essence,
 		FreeBSD,
 		OpenBSD,
+		NetBSD,
+		Haiku,
 		WASI,
 		JS,
+		Orca,
 		Freestanding,
 	}
 */
@@ -499,15 +546,29 @@ Odin_OS_Type :: type_of(ODIN_OS)
 		arm64,
 		wasm32,
 		wasm64p32,
+		riscv64,
 	}
 */
 Odin_Arch_Type :: type_of(ODIN_ARCH)

+Odin_Arch_Types :: bit_set[Odin_Arch_Type]
+
+ALL_ODIN_ARCH_TYPES :: Odin_Arch_Types{
+	.amd64,
+	.i386,
+	.arm32,
+	.arm64,
+	.wasm32,
+	.wasm64p32,
+	.riscv64,
+}
+
 /*
 	// Defined internally by the compiler
 	Odin_Build_Mode_Type :: enum int {
 		Executable,
 		Dynamic,
+		Static,
 		Object,
 		Assembly,
 		LLVM_IR,
@@ -525,6 +586,22 @@ Odin_Build_Mode_Type :: type_of(ODIN_BUILD_MODE)
 */
 Odin_Endian_Type :: type_of(ODIN_ENDIAN)

+Odin_OS_Types :: bit_set[Odin_OS_Type]
+
+ALL_ODIN_OS_TYPES :: Odin_OS_Types{
+	.Windows,
+	.Darwin,
+	.Linux,
+	.Essence,
+	.FreeBSD,
+	.OpenBSD,
+	.NetBSD,
+	.Haiku,
+	.WASI,
+	.JS,
+	.Orca,
+	.Freestanding,
+}

 /*
 	// Defined internally by the compiler
@@ -542,12 +619,25 @@ Odin_Platform_Subtarget_Type :: type_of(ODIN_PLATFORM_SUBTARGET)
 		Memory  = 1,
 		Thread  = 2,
 	}
-	Odin_Sanitizer_Flags :: distinct bitset[Odin_Sanitizer_Flag; u32]
+	Odin_Sanitizer_Flags :: distinct bit_set[Odin_Sanitizer_Flag; u32]

 	ODIN_SANITIZER_FLAGS // is a constant
 */
 Odin_Sanitizer_Flags :: type_of(ODIN_SANITIZER_FLAGS)

+/*
+	// Defined internally by the compiler
+	Odin_Optimization_Mode :: enum int {
+		None       = -1,
+		Minimal    =  0,
+		Size       =  1,
+		Speed      =  2,
+		Aggressive =  3,
+	}
+
+	ODIN_OPTIMIZATION_MODE // is a constant
+*/
+Odin_Optimization_Mode :: type_of(ODIN_OPTIMIZATION_MODE)

 /////////////////////////////
 // Init Startup Procedures //
@@ -683,13 +773,20 @@ __init_context :: proc "contextless" (c: ^Context) {

 	c.logger.procedure = default_logger_proc
 	c.logger.data = nil
+
+	c.random_generator.procedure = default_random_generator_proc
+	c.random_generator.data = nil
 }

 default_assertion_failure_proc :: proc(prefix, message: string, loc: Source_Code_Location) -> ! {
+	default_assertion_contextless_failure_proc(prefix, message, loc)
+}
+
+default_assertion_contextless_failure_proc :: proc "contextless" (prefix, message: string, loc: Source_Code_Location) -> ! {
 	when ODIN_OS == .Freestanding {
 		// Do nothing
 	} else {
-		when !ODIN_DISABLE_ASSERT {
+		when ODIN_OS != .Orca && !ODIN_DISABLE_ASSERT {
 			print_caller_location(loc)
 			print_string(" ")
 		}
@@ -698,7 +795,18 @@ default_assertion_failure_proc :: proc(prefix, message: string, loc: Source_Code
 			print_string(": ")
 			print_string(message)
 		}
-		print_byte('\n')
+
+		when ODIN_OS == .Orca {
+			assert_fail(
+				cstring(raw_data(loc.file_path)),
+				cstring(raw_data(loc.procedure)),
+				loc.line,
+				"",
+				cstring(raw_data(orca_stderr_buffer[:orca_stderr_buffer_idx])),
+			)
+		} else {
+			print_byte('\n')
+		}
 	}
 	trap()
 }
@@ -40,7 +40,7 @@ copy_slice :: proc "contextless" (dst, src: $T/[]$E) -> int {
 	}
 	return n
 }
-// `copy_from_string` is a built-in procedure that copies elements from a source slice `src` to a destination string `dst`.
+// `copy_from_string` is a built-in procedure that copies elements from a source string `src` to a destination slice `dst`.
 // The source and destination may overlap. Copy returns the number of elements copied, which will be the minimum
 // of len(src) and len(dst).
 //
@@ -53,7 +53,7 @@ copy_from_string :: proc "contextless" (dst: $T/[]$E/u8, src: $S/string) -> int
 	}
 	return n
 }
-// `copy` is a built-in procedure that copies elements from a source slice `src` to a destination slice/string `dst`.
+// `copy` is a built-in procedure that copies elements from a source slice/string `src` to a destination slice `dst`.
 // The source and destination may overlap. Copy returns the number of elements copied, which will be the minimum
 // of len(src) and len(dst).
@builtin
@@ -65,7 +65,7 @@ copy :: proc{copy_slice, copy_from_string}
 // with the old value, and reducing the length of the dynamic array by 1.
 //
 // Note: This is an O(1) operation.
-// Note: If you the elements to remain in their order, use `ordered_remove`.
+// Note: If you want the elements to remain in their order, use `ordered_remove`.
 // Note: If the index is out of bounds, this procedure will panic.
@builtin
 unordered_remove :: proc(array: ^$D/[dynamic]$T, index: int, loc := #caller_location) #no_bounds_check {
@@ -79,7 +79,7 @@ unordered_remove :: proc(array: ^$D/[dynamic]$T, index: int, loc := #caller_loca
 // `ordered_remove` removed the element at the specified `index` whilst keeping the order of the other elements.
 //
 // Note: This is an O(N) operation.
-// Note: If you the elements do not have to remain in their order, prefer `unordered_remove`.
+// Note: If the elements do not have to remain in their order, prefer `unordered_remove`.
 // Note: If the index is out of bounds, this procedure will panic.
@builtin
 ordered_remove :: proc(array: ^$D/[dynamic]$T, index: int, loc := #caller_location) #no_bounds_check {
@@ -163,21 +163,43 @@ pop_front_safe :: proc "contextless" (array: ^$T/[dynamic]$E) -> (res: E, ok: bo

 // `clear` will set the length of a passed dynamic array or map to `0`
@builtin
-clear :: proc{clear_dynamic_array, clear_map}
+clear :: proc{
+	clear_dynamic_array,
+	clear_map,
+
+	clear_soa_dynamic_array,
+}

 // `reserve` will try to reserve memory of a passed dynamic array or map to the requested element count (setting the `cap`).
@builtin
-reserve :: proc{reserve_dynamic_array, reserve_map}
+reserve :: proc{
+	reserve_dynamic_array,
+	reserve_map,
+
+	reserve_soa,
+}

@builtin
-non_zero_reserve :: proc{non_zero_reserve_dynamic_array}
+non_zero_reserve :: proc{
+	non_zero_reserve_dynamic_array,
+
+	non_zero_reserve_soa,
+}

 // `resize` will try to resize memory of a passed dynamic array to the requested element count (setting the `len`, and possibly `cap`).
@builtin
-resize :: proc{resize_dynamic_array}
+resize :: proc{
+	resize_dynamic_array,
+
+	resize_soa,
+}

@builtin
-non_zero_resize :: proc{non_zero_resize_dynamic_array}
+non_zero_resize :: proc{
+	non_zero_resize_dynamic_array,
+
+	non_zero_resize_soa,
+}

 // Shrinks the capacity of a dynamic array or map down to the current length, or the given capacity.
@builtin
@@ -268,7 +290,7 @@ new_clone :: proc(data: $T, allocator := context.allocator, loc := #caller_locat
 	return
 }

-DEFAULT_RESERVE_CAPACITY :: 16
+DEFAULT_DYNAMIC_ARRAY_CAPACITY :: 8

@(require_results)
 make_aligned :: proc($T: typeid/[]$E, #any_int len: int, alignment: int, allocator := context.allocator, loc := #caller_location) -> (T, Allocator_Error) #optional_allocator_error {
@@ -295,7 +317,7 @@ make_slice :: proc($T: typeid/[]$E, #any_int len: int, allocator := context.allo
 // Note: Prefer using the procedure group `make`.
@(builtin, require_results)
 make_dynamic_array :: proc($T: typeid/[dynamic]$E, allocator := context.allocator, loc := #caller_location) -> (T, Allocator_Error) #optional_allocator_error {
-	return make_dynamic_array_len_cap(T, 0, DEFAULT_RESERVE_CAPACITY, allocator, loc)
+	return make_dynamic_array_len_cap(T, 0, 0, allocator, loc)
 }
 // `make_dynamic_array_len` allocates and initializes a dynamic array. Like `new`, the first argument is a type, not a value.
 // Unlike `new`, `make`'s return value is the same as the type of its argument, not a pointer to it.
@@ -311,17 +333,24 @@ make_dynamic_array_len :: proc($T: typeid/[dynamic]$E, #any_int len: int, alloca
 // Note: Prefer using the procedure group `make`.
@(builtin, require_results)
 make_dynamic_array_len_cap :: proc($T: typeid/[dynamic]$E, #any_int len: int, #any_int cap: int, allocator := context.allocator, loc := #caller_location) -> (array: T, err: Allocator_Error) #optional_allocator_error {
-	make_dynamic_array_error_loc(loc, len, cap)
-	array.allocator = allocator // initialize allocator before just in case it fails to allocate any memory
-	data := mem_alloc_bytes(size_of(E)*cap, align_of(E), allocator, loc) or_return
-	s := Raw_Dynamic_Array{raw_data(data), len, cap, allocator}
-	if data == nil && size_of(E) != 0 {
-		s.len, s.cap = 0, 0
-	}
-	array = transmute(T)s
+	err = _make_dynamic_array_len_cap((^Raw_Dynamic_Array)(&array), size_of(E), align_of(E), len, cap, allocator, loc)
 	return
 }
-// `make_map` allocates and initializes a dynamic array. Like `new`, the first argument is a type, not a value.
+
+@(require_results)
+_make_dynamic_array_len_cap :: proc(array: ^Raw_Dynamic_Array, size_of_elem, align_of_elem: int, #any_int len: int, #any_int cap: int, allocator := context.allocator, loc := #caller_location) -> (err: Allocator_Error) {
+	make_dynamic_array_error_loc(loc, len, cap)
+	array.allocator = allocator // initialize allocator before just in case it fails to allocate any memory
+	data := mem_alloc_bytes(size_of_elem*cap, align_of_elem, allocator, loc) or_return
+	use_zero := data == nil && size_of_elem != 0
+	array.data = raw_data(data)
+	array.len = 0 if use_zero else len
+	array.cap = 0 if use_zero else cap
+	array.allocator = allocator
+	return
+}
+
+// `make_map` allocates and initializes a map. Like `new`, the first argument is a type, not a value.
 // Unlike `new`, `make`'s return value is the same as the type of its argument, not a pointer to it.
 //
 // Note: Prefer using the procedure group `make`.
@@ -333,7 +362,7 @@ make_map :: proc($T: typeid/map[$K]$E, #any_int capacity: int = 1<<MAP_MIN_LOG2_
 	err = reserve_map(&m, capacity, loc)
 	return
 }
-// `make_multi_pointer` allocates and initializes a dynamic array. Like `new`, the first argument is a type, not a value.
+// `make_multi_pointer` allocates and initializes a multi-pointer. Like `new`, the first argument is a type, not a value.
 // Unlike `new`, `make`'s return value is the same as the type of its argument, not a pointer to it.
 //
 // This is "similar" to doing `raw_data(make([]E, len, allocator))`.
@@ -364,6 +393,11 @@ make :: proc{
 	make_dynamic_array_len_cap,
 	make_map,
 	make_multi_pointer,
+
+	make_soa_slice,
+	make_soa_dynamic_array,
+	make_soa_dynamic_array_len,
+	make_soa_dynamic_array_len_cap,
 }


@@ -383,7 +417,7 @@ clear_map :: proc "contextless" (m: ^$T/map[$K]$V) {
 //
 // Note: Prefer the procedure group `reserve`
@builtin
-reserve_map :: proc(m: ^$T/map[$K]$V, capacity: int, loc := #caller_location) -> Allocator_Error {
+reserve_map :: proc(m: ^$T/map[$K]$V, #any_int capacity: int, loc := #caller_location) -> Allocator_Error {
 	return __dynamic_map_reserve((^Raw_Map)(m), map_info(T), uint(capacity), loc) if m != nil else nil
 }

@@ -413,106 +447,103 @@ delete_key :: proc(m: ^$T/map[$K]$V, key: K) -> (deleted_key: K, deleted_value:
 	return
 }

-_append_elem :: #force_inline proc(array: ^$T/[dynamic]$E, arg: E, should_zero: bool, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+_append_elem :: #force_inline proc(array: ^Raw_Dynamic_Array, size_of_elem, align_of_elem: int, arg_ptr: rawptr, should_zero: bool, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
 	if array == nil {
-		return 0, nil
+		return
 	}
-	when size_of(E) == 0 {
-		array := (^Raw_Dynamic_Array)(array)
-		array.len += 1
-		return 1, nil
-	} else {
-		if cap(array) < len(array)+1 {
-			cap := 2 * cap(array) + max(8, 1)

-			// do not 'or_return' here as it could be a partial success
-			if should_zero {
-				err = reserve(array, cap, loc)
-			} else {
-				err = non_zero_reserve(array, cap, loc) 
-			}
-		}
-		if cap(array)-len(array) > 0 {
-			a := (^Raw_Dynamic_Array)(array)
-			when size_of(E) != 0 {
-				data := ([^]E)(a.data)
-				assert(data != nil, loc=loc)
-				data[a.len] = arg
-			}
-			a.len += 1
-			return 1, err
-		}
-		return 0, err
+	if array.cap < array.len+1 {
+		// Same behavior as _append_elems but there's only one arg, so we always just add DEFAULT_DYNAMIC_ARRAY_CAPACITY.
+		cap := 2 * array.cap + DEFAULT_DYNAMIC_ARRAY_CAPACITY
+
+		// do not 'or_return' here as it could be a partial success
+		err = _reserve_dynamic_array(array, size_of_elem, align_of_elem, cap, should_zero, loc)
 	}
+	if array.cap-array.len > 0 {
+		data := ([^]byte)(array.data)
+		assert(data != nil, loc=loc)
+		data = data[array.len*size_of_elem:]
+		intrinsics.mem_copy_non_overlapping(data, arg_ptr, size_of_elem)
+		array.len += 1
+		n = 1
+	}
+	return
 }

@builtin
 append_elem :: proc(array: ^$T/[dynamic]$E, #no_broadcast arg: E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
-	return _append_elem(array, arg, true, loc=loc)
+	when size_of(E) == 0 {
+		(^Raw_Dynamic_Array)(array).len += 1
+		return 1, nil
+	} else {
+		arg := arg
+		return _append_elem((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), &arg, true, loc=loc)
+	}
 }

@builtin
 non_zero_append_elem :: proc(array: ^$T/[dynamic]$E, #no_broadcast arg: E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
-	return _append_elem(array, arg, false, loc=loc)
+	when size_of(E) == 0 {
+		(^Raw_Dynamic_Array)(array).len += 1
+		return 1, nil
+	} else {
+		arg := arg
+		return _append_elem((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), &arg, false, loc=loc)
+	}
 }

-_append_elems :: #force_inline proc(array: ^$T/[dynamic]$E, should_zero: bool, loc := #caller_location, args: ..E) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+_append_elems :: #force_inline proc(array: ^Raw_Dynamic_Array, size_of_elem, align_of_elem: int, should_zero: bool, loc := #caller_location, args: rawptr, arg_len: int) -> (n: int, err: Allocator_Error) #optional_allocator_error {
 	if array == nil {
 		return 0, nil
 	}

-	arg_len := len(args)
 	if arg_len <= 0 {
 		return 0, nil
 	}

-	when size_of(E) == 0 {
-		array := (^Raw_Dynamic_Array)(array)
-		array.len += arg_len
-		return arg_len, nil
-	} else {
-		if cap(array) < len(array)+arg_len {
-			cap := 2 * cap(array) + max(8, arg_len)
+	if array.cap < array.len+arg_len {
+		cap := 2 * array.cap + max(DEFAULT_DYNAMIC_ARRAY_CAPACITY, arg_len)

-			// do not 'or_return' here as it could be a partial success
-			if should_zero {
-				err = reserve(array, cap, loc)
-			} else {
-				err = non_zero_reserve(array, cap, loc)
-			}
-		}
-		arg_len = min(cap(array)-len(array), arg_len)
-		if arg_len > 0 {
-			a := (^Raw_Dynamic_Array)(array)
-			when size_of(E) != 0 {
-				data := ([^]E)(a.data)
-				assert(data != nil, loc=loc)
-				intrinsics.mem_copy(&data[a.len], raw_data(args), size_of(E) * arg_len)
-			}
-			a.len += arg_len
-		}
-		return arg_len, err
+		// do not 'or_return' here as it could be a partial success
+		err = _reserve_dynamic_array(array, size_of_elem, align_of_elem, cap, should_zero, loc)
 	}
+	arg_len := arg_len
+	arg_len = min(array.cap-array.len, arg_len)
+	if arg_len > 0 {
+		data := ([^]byte)(array.data)
+		assert(data != nil, loc=loc)
+		data = data[array.len*size_of_elem:]
+		intrinsics.mem_copy(data, args, size_of_elem * arg_len) // must be mem_copy (overlapping)
+		array.len += arg_len
+	}
+	return arg_len, err
 }

@builtin
 append_elems :: proc(array: ^$T/[dynamic]$E, #no_broadcast args: ..E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
-	return _append_elems(array, true, loc, ..args)
+	when size_of(E) == 0 {
+		a := (^Raw_Dynamic_Array)(array)
+		a.len += len(args)
+		return len(args), nil
+	} else {
+		return _append_elems((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), true, loc, raw_data(args), len(args))
+	}
 }

@builtin
 non_zero_append_elems :: proc(array: ^$T/[dynamic]$E, #no_broadcast args: ..E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
-	return _append_elems(array, false, loc, ..args)
+	when size_of(E) == 0 {
+		a := (^Raw_Dynamic_Array)(array)
+		a.len += len(args)
+		return len(args), nil
+	} else {
+		return _append_elems((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), false, loc, raw_data(args), len(args))
+	}
 }

 // The append_string built-in procedure appends a string to the end of a [dynamic]u8 like type
 _append_elem_string :: proc(array: ^$T/[dynamic]$E/u8, arg: $A/string, should_zero: bool, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
-	args := transmute([]E)arg
-	if should_zero { 
-		return append_elems(array, ..args, loc=loc)
-	} else {
-		return non_zero_append_elems(array, ..args, loc=loc)
-	}
+	return _append_elems((^Raw_Dynamic_Array)(array), 1, 1, should_zero, loc, raw_data(arg), len(arg))
 }

@builtin
@@ -540,8 +571,23 @@ append_string :: proc(array: ^$T/[dynamic]$E/u8, args: ..string, loc := #caller_
 }

 // The append built-in procedure appends elements to the end of a dynamic array
-@builtin append :: proc{append_elem, append_elems, append_elem_string}
-@builtin non_zero_append :: proc{non_zero_append_elem, non_zero_append_elems, non_zero_append_elem_string}
+@builtin append :: proc{
+	append_elem,
+	append_elems,
+	append_elem_string,
+
+	append_soa_elem,
+	append_soa_elems,
+}
+
+@builtin non_zero_append :: proc{
+	non_zero_append_elem,
+	non_zero_append_elems,
+	non_zero_append_elem_string,
+
+	non_zero_append_soa_elem,
+	non_zero_append_soa_elems,
+}


@builtin
@@ -636,7 +682,7 @@ assign_at_elem :: proc(array: ^$T/[dynamic]$E, index: int, arg: E, loc := #calle


@builtin
-assign_at_elems :: proc(array: ^$T/[dynamic]$E, index: int, args: ..E, loc := #caller_location) -> (ok: bool, err: Allocator_Error) #no_bounds_check #optional_allocator_error {
+assign_at_elems :: proc(array: ^$T/[dynamic]$E, index: int, #no_broadcast args: ..E, loc := #caller_location) -> (ok: bool, err: Allocator_Error) #no_bounds_check #optional_allocator_error {
 	new_size := index + len(args)
 	if len(args) == 0 {
 		ok = true
@@ -686,11 +732,10 @@ clear_dynamic_array :: proc "contextless" (array: ^$T/[dynamic]$E) {
 // `reserve_dynamic_array` will try to reserve memory of a passed dynamic array or map to the requested element count (setting the `cap`).
 //
 // Note: Prefer the procedure group `reserve`.
-_reserve_dynamic_array :: #force_inline proc(array: ^$T/[dynamic]$E, capacity: int, should_zero: bool, loc := #caller_location) -> Allocator_Error {
-	if array == nil {
+_reserve_dynamic_array :: #force_inline proc(a: ^Raw_Dynamic_Array, size_of_elem, align_of_elem: int, capacity: int, should_zero: bool, loc := #caller_location) -> Allocator_Error {
+	if a == nil {
 		return nil
 	}
-	a := (^Raw_Dynamic_Array)(array)

 	if capacity <= a.cap {
 		return nil
@@ -701,15 +746,15 @@ _reserve_dynamic_array :: #force_inline proc(array: ^$T/[dynamic]$E, capacity: i
 	}
 	assert(a.allocator.procedure != nil)

-	old_size  := a.cap * size_of(E)
-	new_size  := capacity * size_of(E)
+	old_size  := a.cap * size_of_elem
+	new_size  := capacity * size_of_elem
 	allocator := a.allocator

 	new_data: []byte
 	if should_zero {
-		new_data = mem_resize(a.data, old_size, new_size, align_of(E), allocator, loc) or_return
+		new_data = mem_resize(a.data, old_size, new_size, align_of_elem, allocator, loc) or_return
 	} else {
-		new_data = non_zero_mem_resize(a.data, old_size, new_size, align_of(E), allocator, loc) or_return
+		new_data = non_zero_mem_resize(a.data, old_size, new_size, align_of_elem, allocator, loc) or_return
 	}
 	if new_data == nil && new_size > 0 {
 		return .Out_Of_Memory
@@ -721,27 +766,24 @@ _reserve_dynamic_array :: #force_inline proc(array: ^$T/[dynamic]$E, capacity: i
 }

@builtin
-reserve_dynamic_array :: proc(array: ^$T/[dynamic]$E, capacity: int, loc := #caller_location) -> Allocator_Error {
-	return _reserve_dynamic_array(array, capacity, true, loc)
+reserve_dynamic_array :: proc(array: ^$T/[dynamic]$E, #any_int capacity: int, loc := #caller_location) -> Allocator_Error {
+	return _reserve_dynamic_array((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), capacity, true, loc)
 }

@builtin
-non_zero_reserve_dynamic_array :: proc(array: ^$T/[dynamic]$E, capacity: int, loc := #caller_location) -> Allocator_Error {
-	return _reserve_dynamic_array(array, capacity, false, loc)
+non_zero_reserve_dynamic_array :: proc(array: ^$T/[dynamic]$E, #any_int capacity: int, loc := #caller_location) -> Allocator_Error {
+	return _reserve_dynamic_array((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), capacity, false, loc)
 }

-// `resize_dynamic_array` will try to resize memory of a passed dynamic array or map to the requested element count (setting the `len`, and possibly `cap`).
-//
-// Note: Prefer the procedure group `resize`
-_resize_dynamic_array :: #force_inline proc(array: ^$T/[dynamic]$E, length: int, should_zero: bool, loc := #caller_location) -> Allocator_Error {
-	if array == nil {
+
+_resize_dynamic_array :: #force_inline proc(a: ^Raw_Dynamic_Array, size_of_elem, align_of_elem: int, length: int, should_zero: bool, loc := #caller_location) -> Allocator_Error {
+	if a == nil {
 		return nil
 	}
-	a := (^Raw_Dynamic_Array)(array)

 	if length <= a.cap {
 		if should_zero && a.len < length {
-			intrinsics.mem_zero(([^]E)(a.data)[a.len:], (length-a.len)*size_of(E))
+			intrinsics.mem_zero(([^]byte)(a.data)[a.len*size_of_elem:], (length-a.len)*size_of_elem)
 		}
 		a.len = max(length, 0)
 		return nil
@@ -752,15 +794,15 @@ _resize_dynamic_array :: #force_inline proc(array: ^$T/[dynamic]$E, length: int,
 	}
 	assert(a.allocator.procedure != nil)

-	old_size  := a.cap * size_of(E)
-	new_size  := length * size_of(E)
+	old_size  := a.cap  * size_of_elem
+	new_size  := length * size_of_elem
 	allocator := a.allocator

 	new_data : []byte
 	if should_zero {
-		new_data = mem_resize(a.data, old_size, new_size, align_of(E), allocator, loc) or_return
+		new_data = mem_resize(a.data, old_size, new_size, align_of_elem, allocator, loc) or_return
 	} else {
-		new_data = non_zero_mem_resize(a.data, old_size, new_size, align_of(E), allocator, loc) or_return
+		new_data = non_zero_mem_resize(a.data, old_size, new_size, align_of_elem, allocator, loc) or_return
 	}
 	if new_data == nil && new_size > 0 {
 		return .Out_Of_Memory
@@ -772,14 +814,17 @@ _resize_dynamic_array :: #force_inline proc(array: ^$T/[dynamic]$E, length: int,
 	return nil
 }

+// `resize_dynamic_array` will try to resize memory of a passed dynamic array or map to the requested element count (setting the `len`, and possibly `cap`).
+//
+// Note: Prefer the procedure group `resize`
@builtin
-resize_dynamic_array :: proc(array: ^$T/[dynamic]$E, length: int, loc := #caller_location) -> Allocator_Error {
-	return _resize_dynamic_array(array, length, true, loc=loc)
+resize_dynamic_array :: proc(array: ^$T/[dynamic]$E, #any_int length: int, loc := #caller_location) -> Allocator_Error {
+	return _resize_dynamic_array((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), length, true, loc=loc)
 }

@builtin
-non_zero_resize_dynamic_array :: proc(array: ^$T/[dynamic]$E, length: int, loc := #caller_location) -> Allocator_Error {
-	return _resize_dynamic_array(array, length, false, loc=loc)
+non_zero_resize_dynamic_array :: proc(array: ^$T/[dynamic]$E, #any_int length: int, loc := #caller_location) -> Allocator_Error {
+	return _resize_dynamic_array((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), length, false, loc=loc)
 }

 /*
@@ -794,10 +839,13 @@ non_zero_resize_dynamic_array :: proc(array: ^$T/[dynamic]$E, length: int, loc :
 	Note: Prefer the procedure group `shrink`
 */
 shrink_dynamic_array :: proc(array: ^$T/[dynamic]$E, new_cap := -1, loc := #caller_location) -> (did_shrink: bool, err: Allocator_Error) {
-	if array == nil {
+	return _shrink_dynamic_array((^Raw_Dynamic_Array)(array), size_of(E), align_of(E), new_cap, loc)
+}
+
+_shrink_dynamic_array :: proc(a: ^Raw_Dynamic_Array, size_of_elem, align_of_elem: int, new_cap := -1, loc := #caller_location) -> (did_shrink: bool, err: Allocator_Error) {
+	if a == nil {
 		return
 	}
-	a := (^Raw_Dynamic_Array)(array)

 	new_cap := new_cap if new_cap >= 0 else a.len

@@ -810,10 +858,10 @@ shrink_dynamic_array :: proc(array: ^$T/[dynamic]$E, new_cap := -1, loc := #call
 	}
 	assert(a.allocator.procedure != nil)

-	old_size := a.cap * size_of(E)
-	new_size := new_cap * size_of(E)
+	old_size := a.cap * size_of_elem
+	new_size := new_cap * size_of_elem

-	new_data := mem_resize(a.data, old_size, new_size, align_of(E), a.allocator, loc) or_return
+	new_data := mem_resize(a.data, old_size, new_size, align_of_elem, a.allocator, loc) or_return

 	a.data = raw_data(new_data)
 	a.len = min(new_cap, a.len)
@@ -900,3 +948,30 @@ unimplemented :: proc(message := "", loc := #caller_location) -> ! {
 	}
 	p("not yet implemented", message, loc)
 }
+
+
+@builtin
+@(disabled=ODIN_DISABLE_ASSERT)
+assert_contextless :: proc "contextless" (condition: bool, message := "", loc := #caller_location) {
+	if !condition {
+		// NOTE(bill): This is wrapped in a procedure call
+		// to improve performance to make the CPU not
+		// execute speculatively, making it about an order of
+		// magnitude faster
+		@(cold)
+		internal :: proc "contextless" (message: string, loc: Source_Code_Location) {
+			default_assertion_contextless_failure_proc("runtime assertion", message, loc)
+		}
+		internal(message, loc)
+	}
+}
+
+@builtin
+panic_contextless :: proc "contextless" (message: string, loc := #caller_location) -> ! {
+	default_assertion_contextless_failure_proc("panic", message, loc)
+}
+
+@builtin
+unimplemented_contextless :: proc "contextless" (message := "", loc := #caller_location) -> ! {
+	default_assertion_contextless_failure_proc("not yet implemented", message, loc)
+}
@@ -55,7 +55,7 @@ raw_soa_footer_slice :: proc(array: ^$T/#soa[]$E) -> (footer: ^Raw_SOA_Footer_Sl
 	if array == nil {
 		return nil
 	}
-	field_count := uintptr(intrinsics.type_struct_field_count(E))
+	field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))
 	footer = (^Raw_SOA_Footer_Slice)(uintptr(array) + field_count*size_of(rawptr))
 	return
 }
@@ -64,12 +64,7 @@ raw_soa_footer_dynamic_array :: proc(array: ^$T/#soa[dynamic]$E) -> (footer: ^Ra
 	if array == nil {
 		return nil
 	}
-	field_count: uintptr
-	when intrinsics.type_is_array(E) {
-		field_count = len(E)
-	} else {
-		field_count = uintptr(intrinsics.type_struct_field_count(E))
-	}
+	field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))
 	footer = (^Raw_SOA_Footer_Dynamic_Array)(uintptr(array) + field_count*size_of(rawptr))
 	return
 }
@@ -98,11 +93,11 @@ make_soa_aligned :: proc($T: typeid/#soa[]$E, length: int, alignment: int, alloc
 	ti = type_info_base(ti)
 	si := &ti.variant.(Type_Info_Struct)

-	field_count := uintptr(intrinsics.type_struct_field_count(E))
+	field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))

 	total_size := 0
 	for i in 0..<field_count {
-		type := si.types[i].variant.(Type_Info_Pointer).elem
+		type := si.types[i].variant.(Type_Info_Multi_Pointer).elem
 		total_size += type.size * length
 		total_size = align_forward_int(total_size, max_align)
 	}
@@ -126,7 +121,7 @@ make_soa_aligned :: proc($T: typeid/#soa[]$E, length: int, alignment: int, alloc
 	data := uintptr(&array)
 	offset := 0
 	for i in 0..<field_count {
-		type := si.types[i].variant.(Type_Info_Pointer).elem
+		type := si.types[i].variant.(Type_Info_Multi_Pointer).elem

 		offset = align_forward_int(offset, max_align)

@@ -147,7 +142,7 @@ make_soa_slice :: proc($T: typeid/#soa[]$E, length: int, allocator := context.al
@(builtin, require_results)
 make_soa_dynamic_array :: proc($T: typeid/#soa[dynamic]$E, allocator := context.allocator, loc := #caller_location) -> (array: T, err: Allocator_Error) #optional_allocator_error {
 	context.allocator = allocator
-	reserve_soa(&array, DEFAULT_RESERVE_CAPACITY, loc) or_return
+	reserve_soa(&array, 0, loc) or_return
 	return array, nil
 }

@@ -187,8 +182,28 @@ resize_soa :: proc(array: ^$T/#soa[dynamic]$E, length: int, loc := #caller_locat
 	return nil
 }

+@builtin
+non_zero_resize_soa :: proc(array: ^$T/#soa[dynamic]$E, length: int, loc := #caller_location) -> Allocator_Error {
+	if array == nil {
+		return nil
+	}
+	non_zero_reserve_soa(array, length, loc) or_return
+	footer := raw_soa_footer(array)
+	footer.len = length
+	return nil
+}
+
@builtin
 reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, loc := #caller_location) -> Allocator_Error {
+	return _reserve_soa(array, capacity, true, loc)
+}
+
+@builtin
+non_zero_reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, loc := #caller_location) -> Allocator_Error {
+	return _reserve_soa(array, capacity, false, loc)
+}
+
+_reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, zero_memory: bool, loc := #caller_location) -> Allocator_Error {
 	if array == nil {
 		return nil
 	}
@@ -213,12 +228,7 @@ reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, loc := #caller_lo
 	ti = type_info_base(ti)
 	si := &ti.variant.(Type_Info_Struct)

-	field_count: uintptr
-	when intrinsics.type_is_array(E) {
-		field_count = len(E)
-	} else {
-		field_count = uintptr(intrinsics.type_struct_field_count(E))
-	}
+	field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))
 	assert(footer.cap == old_cap)

 	old_size := 0
@@ -226,7 +236,7 @@ reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, loc := #caller_lo

 	max_align :: align_of(E)
 	for i in 0..<field_count {
-		type := si.types[i].variant.(Type_Info_Pointer).elem
+		type := si.types[i].variant.(Type_Info_Multi_Pointer).elem

 		old_size += type.size * old_cap
 		new_size += type.size * capacity
@@ -238,7 +248,7 @@ reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, loc := #caller_lo
 	old_data := (^rawptr)(array)^

 	new_bytes := array.allocator.procedure(
-		array.allocator.data, .Alloc, new_size, max_align,
+		array.allocator.data, .Alloc if zero_memory else .Alloc_Non_Zeroed, new_size, max_align,
 		nil, old_size, loc,
 	) or_return
 	new_data := raw_data(new_bytes)
@@ -249,7 +259,7 @@ reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, loc := #caller_lo
 	old_offset := 0
 	new_offset := 0
 	for i in 0..<field_count {
-		type := si.types[i].variant.(Type_Info_Pointer).elem
+		type := si.types[i].variant.(Type_Info_Multi_Pointer).elem

 		old_offset = align_forward_int(old_offset, max_align)
 		new_offset = align_forward_int(new_offset, max_align)
@@ -273,15 +283,26 @@ reserve_soa :: proc(array: ^$T/#soa[dynamic]$E, capacity: int, loc := #caller_lo
 	return nil
 }

+
@builtin
-append_soa_elem :: proc(array: ^$T/#soa[dynamic]$E, arg: E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+append_soa_elem :: proc(array: ^$T/#soa[dynamic]$E, #no_broadcast arg: E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+	return _append_soa_elem(array, true, arg, loc)
+}
+
+@builtin
+non_zero_append_soa_elem :: proc(array: ^$T/#soa[dynamic]$E, #no_broadcast arg: E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+	return _append_soa_elem(array, false, arg, loc)
+}
+
+_append_soa_elem :: proc(array: ^$T/#soa[dynamic]$E, zero_memory: bool, #no_broadcast arg: E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
 	if array == nil {
 		return 0, nil
 	}

 	if cap(array) <= len(array) + 1 {
-		cap := 2 * cap(array) + 8
-		err = reserve_soa(array, cap, loc) // do not 'or_return' here as it could be a partial success
+		// Same behavior as append_soa_elems but there's only one arg, so we always just add DEFAULT_DYNAMIC_ARRAY_CAPACITY.
+		cap := 2 * cap(array) + DEFAULT_DYNAMIC_ARRAY_CAPACITY
+		err = _reserve_soa(array, cap, zero_memory, loc) // do not 'or_return' here as it could be a partial success
 	}

 	footer := raw_soa_footer(array)
@@ -290,12 +311,7 @@ append_soa_elem :: proc(array: ^$T/#soa[dynamic]$E, arg: E, loc := #caller_locat
 		ti := type_info_of(T)
 		ti = type_info_base(ti)
 		si := &ti.variant.(Type_Info_Struct)
-		field_count: uintptr
-		when intrinsics.type_is_array(E) {
-			field_count = len(E)
-		} else {
-			field_count = uintptr(intrinsics.type_struct_field_count(E))
-		}
+		field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))

 		data := (^rawptr)(array)^

@@ -307,7 +323,7 @@ append_soa_elem :: proc(array: ^$T/#soa[dynamic]$E, arg: E, loc := #caller_locat

 		max_align :: align_of(E)
 		for i in 0..<field_count {
-			type := si.types[i].variant.(Type_Info_Pointer).elem
+			type := si.types[i].variant.(Type_Info_Multi_Pointer).elem

 			soa_offset  = align_forward_int(soa_offset, max_align)
 			item_offset = align_forward_int(item_offset, type.align)
@@ -326,7 +342,17 @@ append_soa_elem :: proc(array: ^$T/#soa[dynamic]$E, arg: E, loc := #caller_locat
 }

@builtin
-append_soa_elems :: proc(array: ^$T/#soa[dynamic]$E, args: ..E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+append_soa_elems :: proc(array: ^$T/#soa[dynamic]$E, #no_broadcast args: ..E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+	return _append_soa_elems(array, true, args=args, loc=loc)
+}
+
+@builtin
+non_zero_append_soa_elems :: proc(array: ^$T/#soa[dynamic]$E, #no_broadcast args: ..E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
+	return _append_soa_elems(array, false, args=args, loc=loc)
+}
+
+
+_append_soa_elems :: proc(array: ^$T/#soa[dynamic]$E, zero_memory: bool, #no_broadcast args: []E, loc := #caller_location) -> (n: int, err: Allocator_Error) #optional_allocator_error {
 	if array == nil {
 		return
 	}
@@ -337,8 +363,8 @@ append_soa_elems :: proc(array: ^$T/#soa[dynamic]$E, args: ..E, loc := #caller_l
 	}

 	if cap(array) <= len(array)+arg_len {
-		cap := 2 * cap(array) + max(8, arg_len)
-		err = reserve_soa(array, cap, loc) // do not 'or_return' here as it could be a partial success
+		cap := 2 * cap(array) + max(DEFAULT_DYNAMIC_ARRAY_CAPACITY, arg_len)
+		err = _reserve_soa(array, cap, zero_memory, loc) // do not 'or_return' here as it could be a partial success
 	}
 	arg_len = min(cap(array)-len(array), arg_len)

@@ -347,7 +373,7 @@ append_soa_elems :: proc(array: ^$T/#soa[dynamic]$E, args: ..E, loc := #caller_l
 		ti := type_info_of(typeid_of(T))
 		ti = type_info_base(ti)
 		si := &ti.variant.(Type_Info_Struct)
-		field_count := uintptr(intrinsics.type_struct_field_count(E))
+		field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))

 		data := (^rawptr)(array)^

@@ -358,7 +384,7 @@ append_soa_elems :: proc(array: ^$T/#soa[dynamic]$E, args: ..E, loc := #caller_l

 		max_align :: align_of(E)
 		for i in 0..<field_count {
-			type := si.types[i].variant.(Type_Info_Pointer).elem
+			type := si.types[i].variant.(Type_Info_Multi_Pointer).elem

 			soa_offset  = align_forward_int(soa_offset, max_align)
 			item_offset = align_forward_int(item_offset, type.align)
@@ -389,7 +415,8 @@ append_soa :: proc{


 delete_soa_slice :: proc(array: $T/#soa[]$E, allocator := context.allocator, loc := #caller_location) -> Allocator_Error {
-	when intrinsics.type_struct_field_count(E) != 0 {
+	field_count :: len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E)
+	when field_count != 0 {
 		array := array
 		ptr := (^rawptr)(&array)^
 		free(ptr, allocator, loc) or_return
@@ -398,7 +425,8 @@ delete_soa_slice :: proc(array: $T/#soa[]$E, allocator := context.allocator, loc
 }

 delete_soa_dynamic_array :: proc(array: $T/#soa[dynamic]$E, loc := #caller_location) -> Allocator_Error {
-	when intrinsics.type_struct_field_count(E) != 0 {
+	field_count :: len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E)
+	when field_count != 0 {
 		array := array
 		ptr := (^rawptr)(&array)^
 		footer := raw_soa_footer(&array)
@@ -416,7 +444,8 @@ delete_soa :: proc{


 clear_soa_dynamic_array :: proc(array: ^$T/#soa[dynamic]$E) {
-	when intrinsics.type_struct_field_count(E) != 0 {
+	field_count :: len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E)
+	when field_count != 0 {
 		footer := raw_soa_footer(array)
 		footer.len = 0
 	}
@@ -438,12 +467,7 @@ into_dynamic_soa :: proc(array: $T/#soa[]$E) -> #soa[dynamic]E {
 		allocator = nil_allocator(),
 	}

-	field_count: uintptr
-	when intrinsics.type_is_array(E) {
-		field_count = len(E)
-	} else {
-		field_count = uintptr(intrinsics.type_struct_field_count(E))
-	}
+	field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))

 	array := array
 	dynamic_data := ([^]rawptr)(&d)[:field_count]
@@ -467,16 +491,11 @@ unordered_remove_soa :: proc(array: ^$T/#soa[dynamic]$E, index: int, loc := #cal
 		ti = type_info_base(ti)
 		si := &ti.variant.(Type_Info_Struct)

-		field_count: uintptr
-		when intrinsics.type_is_array(E) {
-			field_count = len(E)
-		} else {
-			field_count = uintptr(intrinsics.type_struct_field_count(E))
-		}
+		field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))

 		data := uintptr(array)
 		for i in 0..<field_count {
-			type := si.types[i].variant.(Type_Info_Pointer).elem
+			type := si.types[i].variant.(Type_Info_Multi_Pointer).elem

 			offset := rawptr((^uintptr)(data)^ + uintptr(index*type.size))
 			final := rawptr((^uintptr)(data)^ + uintptr((len(array)-1)*type.size))
@@ -500,16 +519,11 @@ ordered_remove_soa :: proc(array: ^$T/#soa[dynamic]$E, index: int, loc := #calle
 		ti = type_info_base(ti)
 		si := &ti.variant.(Type_Info_Struct)

-		field_count: uintptr
-		when intrinsics.type_is_array(E) {
-			field_count = len(E)
-		} else {
-			field_count = uintptr(intrinsics.type_struct_field_count(E))
-		}
+		field_count := uintptr(len(E) when intrinsics.type_is_array(E) else intrinsics.type_struct_field_count(E))

 		data := uintptr(array)
 		for i in 0..<field_count {
-			type := si.types[i].variant.(Type_Info_Pointer).elem
+			type := si.types[i].variant.(Type_Info_Multi_Pointer).elem

 			offset := (^uintptr)(data)^ + uintptr(index*type.size)
 			length := type.size*(len(array) - index - 1)
@@ -6,6 +6,9 @@ when ODIN_DEFAULT_TO_NIL_ALLOCATOR {
 } else when ODIN_DEFAULT_TO_PANIC_ALLOCATOR {
 	default_allocator_proc :: panic_allocator_proc
 	default_allocator :: panic_allocator
+} else when ODIN_OS != .Orca && (ODIN_ARCH == .wasm32 || ODIN_ARCH == .wasm64p32) {
+	default_allocator :: default_wasm_allocator
+	default_allocator_proc :: wasm_allocator_proc
 } else {
 	default_allocator :: heap_allocator
 	default_allocator_proc :: heap_allocator_proc
@@ -1,8 +1,8 @@
 package runtime

 nil_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
-                               size, alignment: int,
-                               old_memory: rawptr, old_size: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
+                           size, alignment: int,
+                           old_memory: rawptr, old_size: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
 	switch mode {
 	case .Alloc, .Alloc_Non_Zeroed:
 		return nil, .Out_Of_Memory
@@ -12,6 +12,8 @@ Memory_Block :: struct {
 	capacity:  uint,
 }

+// NOTE: This is a growing arena that is only used for the default temp allocator.
+// For your own growing arena needs, prefer `Arena` from `core:mem/virtual`.
 Arena :: struct {
 	backing_allocator:  Allocator,
 	curr_block:         ^Memory_Block,
@@ -127,7 +129,7 @@ arena_alloc :: proc(arena: ^Arena, size, alignment: uint, loc := #caller_locatio
 	return
 }

-// `arena_init` will initialize the arena with a usuable block.
+// `arena_init` will initialize the arena with a usable block.
 // This procedure is not necessary to use the Arena as the default zero as `arena_alloc` will set things up if necessary
@(require_results)
 arena_init :: proc(arena: ^Arena, size: uint, backing_allocator: Allocator, loc := #caller_location) -> Allocator_Error {
@@ -1,7 +1,7 @@
 package runtime

 DEFAULT_TEMP_ALLOCATOR_BACKING_SIZE: int : #config(DEFAULT_TEMP_ALLOCATOR_BACKING_SIZE, 4 * Megabyte)
-NO_DEFAULT_TEMP_ALLOCATOR: bool : ODIN_OS == .Freestanding || ODIN_OS == .JS || ODIN_DEFAULT_TO_NIL_ALLOCATOR
+NO_DEFAULT_TEMP_ALLOCATOR: bool : ODIN_OS == .Freestanding || ODIN_DEFAULT_TO_NIL_ALLOCATOR

 when NO_DEFAULT_TEMP_ALLOCATOR {
 	Default_Temp_Allocator :: struct {}
@@ -157,7 +157,7 @@ __dynamic_map_get // dynamic map calls
 __dynamic_map_set // dynamic map calls


-## Dynamic literals ([dymamic]T and map[K]V) (can be disabled with -no-dynamic-literals)
+## Dynamic literals ([dynamic]T and map[K]V) (can be disabled with -no-dynamic-literals)

 __dynamic_array_reserve
 __dynamic_array_append
@@ -577,7 +577,7 @@ map_grow_dynamic :: proc "odin" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_Inf


@(require_results)
-map_reserve_dynamic :: proc "odin" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_Info, new_capacity: uintptr, loc := #caller_location) -> Allocator_Error {
+map_reserve_dynamic :: #force_no_inline proc "odin" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_Info, new_capacity: uintptr, loc := #caller_location) -> Allocator_Error {
 	@(require_results)
 	ceil_log2 :: #force_inline proc "contextless" (x: uintptr) -> uintptr {
 		z := intrinsics.count_leading_zeros(x)
@@ -641,7 +641,7 @@ map_reserve_dynamic :: proc "odin" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_


@(require_results)
-map_shrink_dynamic :: proc "odin" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_Info, loc := #caller_location) -> (did_shrink: bool, err: Allocator_Error) {
+map_shrink_dynamic :: #force_no_inline proc "odin" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_Info, loc := #caller_location) -> (did_shrink: bool, err: Allocator_Error) {
 	if m.allocator.procedure == nil {
 		m.allocator = context.allocator
 	}
@@ -688,7 +688,7 @@ map_shrink_dynamic :: proc "odin" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_I
 }

@(require_results)
-map_free_dynamic :: proc "odin" (m: Raw_Map, info: ^Map_Info, loc := #caller_location) -> Allocator_Error {
+map_free_dynamic :: #force_no_inline proc "odin" (m: Raw_Map, info: ^Map_Info, loc := #caller_location) -> Allocator_Error {
 	ptr := rawptr(map_data(m))
 	size := int(map_total_allocation_size(uintptr(map_cap(m)), info))
 	err := mem_free_with_size(ptr, size, m.allocator, loc)
@@ -700,7 +700,7 @@ map_free_dynamic :: proc "odin" (m: Raw_Map, info: ^Map_Info, loc := #caller_loc
 }

@(require_results)
-map_lookup_dynamic :: proc "contextless" (m: Raw_Map, #no_alias info: ^Map_Info, k: uintptr) -> (index: uintptr, ok: bool) {
+map_lookup_dynamic :: #force_no_inline proc "contextless" (m: Raw_Map, #no_alias info: ^Map_Info, k: uintptr) -> (index: uintptr, ok: bool) {
 	if map_len(m) == 0 {
 		return 0, false
 	}
@@ -723,7 +723,7 @@ map_lookup_dynamic :: proc "contextless" (m: Raw_Map, #no_alias info: ^Map_Info,
 	}
 }
@(require_results)
-map_exists_dynamic :: proc "contextless" (m: Raw_Map, #no_alias info: ^Map_Info, k: uintptr) -> (ok: bool) {
+map_exists_dynamic :: #force_no_inline proc "contextless" (m: Raw_Map, #no_alias info: ^Map_Info, k: uintptr) -> (ok: bool) {
 	if map_len(m) == 0 {
 		return false
 	}
@@ -749,7 +749,7 @@ map_exists_dynamic :: proc "contextless" (m: Raw_Map, #no_alias info: ^Map_Info,


@(require_results)
-map_erase_dynamic :: #force_inline proc "contextless" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_Info, k: uintptr) -> (old_k, old_v: uintptr, ok: bool) {
+map_erase_dynamic :: #force_no_inline proc "contextless" (#no_alias m: ^Raw_Map, #no_alias info: ^Map_Info, k: uintptr) -> (old_k, old_v: uintptr, ok: bool) {
 	index := map_lookup_dynamic(m^, info, k) or_return
 	ks, vs, hs, _, _ := map_kvh_data_dynamic(m^, info)
 	hs[index] |= TOMBSTONE_MASK
@@ -1,5 +1,5 @@
 //+private
-//+build linux, darwin, freebsd, openbsd, haiku
+//+build linux, darwin, freebsd, openbsd, netbsd, haiku
 //+no-instrumentation
 package runtime

@@ -34,6 +34,9 @@ when ODIN_BUILD_MODE == .Dynamic {
 		} else when ODIN_OS == .Darwin && ODIN_ARCH == .arm64 {
 			@require foreign import entry "entry_unix_no_crt_darwin_arm64.asm"
 			SYS_exit :: 1
+		} else when ODIN_ARCH == .riscv64 {
+			@require foreign import entry "entry_unix_no_crt_riscv64.asm"
+			SYS_exit :: 93
 		}
 		@(link_name="_start_odin", linkage="strong", require)
 		_start_odin :: proc "c" (argc: i32, argv: [^]cstring) -> ! {
@@ -0,0 +1,10 @@
+.text
+
+.globl _start
+
+_start:
+	ld a0, 0(sp)
+	addi a1, sp, 8
+	addi sp, sp, ~15
+	call _start_odin
+	ebreak
@@ -6,15 +6,34 @@ package runtime
 import "base:intrinsics"

 when !ODIN_TEST && !ODIN_NO_ENTRY_POINT {
-	@(link_name="_start", linkage="strong", require, export)
-	_start :: proc "c" () {
-		context = default_context()
-		#force_no_inline _startup_runtime()
-		intrinsics.__entry_point()
+	when ODIN_OS == .Orca {
+		@(linkage="strong", require, export)
+		oc_on_init :: proc "c" () {
+			context = default_context()
+			#force_no_inline _startup_runtime()
+			intrinsics.__entry_point()
+		}
+		@(linkage="strong", require, export)
+		oc_on_terminate :: proc "c" () {
+			context = default_context()
+			#force_no_inline _cleanup_runtime()
+		}
+	} else {
+		@(link_name="_start", linkage="strong", require, export)
+		_start :: proc "c" () {
+			context = default_context()
+
+			when ODIN_OS == .WASI {
+				_wasi_setup_args()
+			}
+
+			#force_no_inline _startup_runtime()
+			intrinsics.__entry_point()
+		}
+		@(link_name="_end", linkage="strong", require, export)
+		_end :: proc "c" () {
+			context = default_context()
+			#force_no_inline _cleanup_runtime()
+		}
 	}
-	@(link_name="_end", linkage="strong", require, export)
-	_end :: proc "c" () {
-		context = default_context()
-		#force_no_inline _cleanup_runtime()
-	}
-}
+}
@@ -10,8 +10,9 @@ when ODIN_BUILD_MODE == .Dynamic {
 	DllMain :: proc "system" (hinstDLL: rawptr, fdwReason: u32, lpReserved: rawptr) -> b32 {
 		context = default_context()

-		// Populate Windows DLL-specific global
+		// Populate Windows DLL-specific globals
 		dll_forward_reason = DLL_Forward_Reason(fdwReason)
+		dll_instance       = hinstDLL

 		switch dll_forward_reason {
 		case .Process_Attach:
@@ -4,6 +4,8 @@ package runtime
 bounds_trap :: proc "contextless" () -> ! {
 	when ODIN_OS == .Windows {
 		windows_trap_array_bounds()
+	} else when ODIN_OS == .Orca {
+		abort_ext("", "", 0, "bounds trap")
 	} else {
 		trap()
 	}
@@ -13,12 +15,15 @@ bounds_trap :: proc "contextless" () -> ! {
 type_assertion_trap :: proc "contextless" () -> ! {
 	when ODIN_OS == .Windows {
 		windows_trap_type_assertion()
+	} else when ODIN_OS == .Orca {
+		abort_ext("", "", 0, "type assertion trap")
 	} else {
 		trap()
 	}
 }


+@(disabled=ODIN_NO_BOUNDS_CHECK)
 bounds_check_error :: proc "contextless" (file: string, line, column: i32, index, count: int) {
 	if uint(index) < uint(count) {
 		return
@@ -61,6 +66,7 @@ multi_pointer_slice_handle_error :: proc "contextless" (file: string, line, colu
 }


+@(disabled=ODIN_NO_BOUNDS_CHECK)
 multi_pointer_slice_expr_error :: proc "contextless" (file: string, line, column: i32, lo, hi: int) {
 	if lo <= hi {
 		return
@@ -68,6 +74,7 @@ multi_pointer_slice_expr_error :: proc "contextless" (file: string, line, column
 	multi_pointer_slice_handle_error(file, line, column, lo, hi)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 slice_expr_error_hi :: proc "contextless" (file: string, line, column: i32, hi: int, len: int) {
 	if 0 <= hi && hi <= len {
 		return
@@ -75,6 +82,7 @@ slice_expr_error_hi :: proc "contextless" (file: string, line, column: i32, hi:
 	slice_handle_error(file, line, column, 0, hi, len)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 slice_expr_error_lo_hi :: proc "contextless" (file: string, line, column: i32, lo, hi: int, len: int) {
 	if 0 <= lo && lo <= len && lo <= hi && hi <= len {
 		return
@@ -82,6 +90,7 @@ slice_expr_error_lo_hi :: proc "contextless" (file: string, line, column: i32, l
 	slice_handle_error(file, line, column, lo, hi, len)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 dynamic_array_expr_error :: proc "contextless" (file: string, line, column: i32, low, high, max: int) {
 	if 0 <= low && low <= high && high <= max {
 		return
@@ -102,6 +111,7 @@ dynamic_array_expr_error :: proc "contextless" (file: string, line, column: i32,
 }


+@(disabled=ODIN_NO_BOUNDS_CHECK)
 matrix_bounds_check_error :: proc "contextless" (file: string, line, column: i32, row_index, column_index, row_count, column_count: int) {
 	if uint(row_index) < uint(row_count) &&
 	   uint(column_index) < uint(column_count) {
@@ -224,6 +234,7 @@ when ODIN_NO_RTTI {
 }


+@(disabled=ODIN_NO_BOUNDS_CHECK)
 make_slice_error_loc :: #force_inline proc "contextless" (loc := #caller_location, len: int) {
 	if 0 <= len {
 		return
@@ -239,6 +250,7 @@ make_slice_error_loc :: #force_inline proc "contextless" (loc := #caller_locatio
 	handle_error(loc, len)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 make_dynamic_array_error_loc :: #force_inline proc "contextless" (loc := #caller_location, len, cap: int) {
 	if 0 <= len && len <= cap {
 		return
@@ -256,6 +268,7 @@ make_dynamic_array_error_loc :: #force_inline proc "contextless" (loc := #caller
 	handle_error(loc, len, cap)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 make_map_expr_error_loc :: #force_inline proc "contextless" (loc := #caller_location, cap: int) {
 	if 0 <= cap {
 		return
@@ -274,19 +287,22 @@ make_map_expr_error_loc :: #force_inline proc "contextless" (loc := #caller_loca



-
+@(disabled=ODIN_NO_BOUNDS_CHECK)
 bounds_check_error_loc :: #force_inline proc "contextless" (loc := #caller_location, index, count: int) {
 	bounds_check_error(loc.file_path, loc.line, loc.column, index, count)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 slice_expr_error_hi_loc :: #force_inline proc "contextless" (loc := #caller_location, hi: int, len: int) {
 	slice_expr_error_hi(loc.file_path, loc.line, loc.column, hi, len)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 slice_expr_error_lo_hi_loc :: #force_inline proc "contextless" (loc := #caller_location, lo, hi: int, len: int) {
 	slice_expr_error_lo_hi(loc.file_path, loc.line, loc.column, lo, hi, len)
 }

+@(disabled=ODIN_NO_BOUNDS_CHECK)
 dynamic_array_expr_error_loc :: #force_inline proc "contextless" (loc := #caller_location, low, high, max: int) {
 	dynamic_array_expr_error(loc.file_path, loc.line, loc.column, low, high, max)
 }
@@ -19,12 +19,15 @@ heap_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
 	// the pointer we return to the user.
 	//

-	aligned_alloc :: proc(size, alignment: int, old_ptr: rawptr = nil, zero_memory := true) -> ([]byte, Allocator_Error) {
+	aligned_alloc :: proc(size, alignment: int, old_ptr: rawptr, old_size: int, zero_memory := true) -> ([]byte, Allocator_Error) {
 		a := max(alignment, align_of(rawptr))
 		space := size + a - 1

 		allocated_mem: rawptr
-		if old_ptr != nil {
+
+		force_copy := old_ptr != nil && a > align_of(rawptr)
+
+		if !force_copy && old_ptr != nil {
 			original_old_ptr := ([^]rawptr)(old_ptr)[-1]
 			allocated_mem = heap_resize(original_old_ptr, space+size_of(rawptr))
 		} else {
@@ -36,12 +39,19 @@ heap_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
 		aligned_ptr := (ptr - 1 + uintptr(a)) & -uintptr(a)
 		diff := int(aligned_ptr - ptr)
 		if (size + diff) > space || allocated_mem == nil {
+			aligned_free(old_ptr)
+			aligned_free(allocated_mem)
 			return nil, .Out_Of_Memory
 		}

 		aligned_mem = rawptr(aligned_ptr)
 		([^]rawptr)(aligned_mem)[-1] = allocated_mem

+		if force_copy {
+			mem_copy_non_overlapping(aligned_mem, old_ptr, old_size)
+			aligned_free(old_ptr)
+		}
+
 		return byte_slice(aligned_mem, size), nil
 	}

@@ -53,10 +63,10 @@ heap_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,

 	aligned_resize :: proc(p: rawptr, old_size: int, new_size: int, new_alignment: int, zero_memory := true) -> (new_memory: []byte, err: Allocator_Error) {
 		if p == nil {
-			return nil, nil
+			return aligned_alloc(new_size, new_alignment, nil, old_size, zero_memory)
 		}

-		new_memory = aligned_alloc(new_size, new_alignment, p, zero_memory) or_return
+		new_memory = aligned_alloc(new_size, new_alignment, p, old_size, zero_memory) or_return

 		// NOTE: heap_resize does not zero the new memory, so we do it
 		if zero_memory && new_size > old_size {
@@ -68,7 +78,7 @@ heap_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,

 	switch mode {
 	case .Alloc, .Alloc_Non_Zeroed:
-		return aligned_alloc(size, alignment, nil, mode == .Alloc)
+		return aligned_alloc(size, alignment, nil, 0, mode == .Alloc)

 	case .Free:
 		aligned_free(old_memory)
@@ -77,9 +87,6 @@ heap_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
 		return nil, .Mode_Not_Implemented

 	case .Resize, .Resize_Non_Zeroed:
-		if old_memory == nil {
-			return aligned_alloc(size, alignment, nil, mode == .Resize)
-		}
 		return aligned_resize(old_memory, old_size, size, alignment, mode == .Resize)

 	case .Query_Features:
@@ -97,14 +104,14 @@ heap_allocator_proc :: proc(allocator_data: rawptr, mode: Allocator_Mode,
 }


-heap_alloc :: proc(size: int, zero_memory := true) -> rawptr {
+heap_alloc :: proc "contextless" (size: int, zero_memory := true) -> rawptr {
 	return _heap_alloc(size, zero_memory)
 }

-heap_resize :: proc(ptr: rawptr, new_size: int) -> rawptr {
+heap_resize :: proc "contextless" (ptr: rawptr, new_size: int) -> rawptr {
 	return _heap_resize(ptr, new_size)
 }

-heap_free :: proc(ptr: rawptr) {
+heap_free :: proc "contextless" (ptr: rawptr) {
 	_heap_free(ptr)
 }
@@ -0,0 +1,29 @@
+//+build orca
+//+private
+package runtime
+
+foreign {
+	@(link_name="malloc")   _orca_malloc   :: proc "c" (size: int) -> rawptr ---
+	@(link_name="calloc")   _orca_calloc   :: proc "c" (num, size: int) -> rawptr ---
+	@(link_name="free")     _orca_free     :: proc "c" (ptr: rawptr) ---
+	@(link_name="realloc")  _orca_realloc  :: proc "c" (ptr: rawptr, size: int) -> rawptr ---
+}
+
+_heap_alloc :: proc "contextless" (size: int, zero_memory := true) -> rawptr {
+	if size <= 0 {
+		return nil
+	}
+	if zero_memory {
+		return _orca_calloc(1, size)
+	} else {
+		return _orca_malloc(size)
+	}
+}
+
+_heap_resize :: proc "contextless" (ptr: rawptr, new_size: int) -> rawptr {
+	return _orca_realloc(ptr, new_size)
+}
+
+_heap_free :: proc "contextless" (ptr: rawptr) {
+	_orca_free(ptr)
+}
@@ -2,14 +2,17 @@
 //+private
 package runtime

-_heap_alloc :: proc(size: int, zero_memory := true) -> rawptr {
+_heap_alloc :: proc "contextless" (size: int, zero_memory := true) -> rawptr {
+	context = default_context()
 	unimplemented("base:runtime 'heap_alloc' procedure is not supported on this platform")
 }

-_heap_resize :: proc(ptr: rawptr, new_size: int) -> rawptr {
+_heap_resize :: proc "contextless" (ptr: rawptr, new_size: int) -> rawptr {
+	context = default_context()
 	unimplemented("base:runtime 'heap_resize' procedure is not supported on this platform")
 }

-_heap_free :: proc(ptr: rawptr) {
+_heap_free :: proc "contextless" (ptr: rawptr) {
+	context = default_context()
 	unimplemented("base:runtime 'heap_free' procedure is not supported on this platform")
-}
+}
@@ -1,4 +1,4 @@
-//+build linux, darwin, freebsd, openbsd, haiku
+//+build linux, darwin, freebsd, openbsd, netbsd, haiku
 //+private
 package runtime

@@ -16,7 +16,7 @@ foreign libc {
 	@(link_name="realloc")  _unix_realloc  :: proc(ptr: rawptr, size: int) -> rawptr ---
 }

-_heap_alloc :: proc(size: int, zero_memory := true) -> rawptr {
+_heap_alloc :: proc "contextless" (size: int, zero_memory := true) -> rawptr {
 	if size <= 0 {
 		return nil
 	}
@@ -27,12 +27,12 @@ _heap_alloc :: proc(size: int, zero_memory := true) -> rawptr {
 	}
 }

-_heap_resize :: proc(ptr: rawptr, new_size: int) -> rawptr {
+_heap_resize :: proc "contextless" (ptr: rawptr, new_size: int) -> rawptr {
 	// NOTE: _unix_realloc doesn't guarantee new memory will be zeroed on
 	// POSIX platforms. Ensure your caller takes this into account.
 	return _unix_realloc(ptr, new_size)
 }

-_heap_free :: proc(ptr: rawptr) {
+_heap_free :: proc "contextless" (ptr: rawptr) {
 	_unix_free(ptr)
 }
@@ -14,11 +14,11 @@ foreign kernel32 {
 	HeapFree       :: proc(hHeap: rawptr, dwFlags: u32, lpMem: rawptr) -> b32 ---
 }

-_heap_alloc :: proc(size: int, zero_memory := true) -> rawptr {
+_heap_alloc :: proc "contextless" (size: int, zero_memory := true) -> rawptr {
 	HEAP_ZERO_MEMORY :: 0x00000008
 	return HeapAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY if zero_memory else 0, uint(size))
 }
-_heap_resize :: proc(ptr: rawptr, new_size: int) -> rawptr {
+_heap_resize :: proc "contextless" (ptr: rawptr, new_size: int) -> rawptr {
 	if new_size == 0 {
 		_heap_free(ptr)
 		return nil
@@ -30,7 +30,7 @@ _heap_resize :: proc(ptr: rawptr, new_size: int) -> rawptr {
 	HEAP_ZERO_MEMORY :: 0x00000008
 	return HeapReAlloc(GetProcessHeap(), HEAP_ZERO_MEMORY, ptr, uint(new_size))
 }
-_heap_free :: proc(ptr: rawptr) {
+_heap_free :: proc "contextless" (ptr: rawptr) {
 	if ptr == nil {
 		return
 	}
@@ -1,3 +1,4 @@
+//+vet !cast
 package runtime

 import "base:intrinsics"
@@ -7,10 +8,9 @@ IS_WASM :: ODIN_ARCH == .wasm32 || ODIN_ARCH == .wasm64p32

@(private)
 RUNTIME_LINKAGE :: "strong" when (
-	(ODIN_USE_SEPARATE_MODULES || 
+	ODIN_USE_SEPARATE_MODULES || 
 	ODIN_BUILD_MODE == .Dynamic ||
-	!ODIN_NO_CRT) &&
-	!IS_WASM) else "internal"
+	!ODIN_NO_CRT) else "internal"
 RUNTIME_REQUIRE :: false // !ODIN_TILDE

@(private)
@@ -29,7 +29,7 @@ is_power_of_two_int :: #force_inline proc "contextless" (x: int) -> bool {
 	return (x & (x-1)) == 0
 }

-align_forward_int :: #force_inline proc(ptr, align: int) -> int {
+align_forward_int :: #force_inline proc "odin" (ptr, align: int) -> int {
 	assert(is_power_of_two_int(align))

 	p := ptr
@@ -40,6 +40,24 @@ align_forward_int :: #force_inline proc(ptr, align: int) -> int {
 	return p
 }

+is_power_of_two_uint :: #force_inline proc "contextless" (x: uint) -> bool {
+	if x <= 0 {
+		return false
+	}
+	return (x & (x-1)) == 0
+}
+
+align_forward_uint :: #force_inline proc "odin" (ptr, align: uint) -> uint {
+	assert(is_power_of_two_uint(align))
+
+	p := ptr
+	modulo := p & (align-1)
+	if modulo != 0 {
+		p += align - modulo
+	}
+	return p
+}
+
 is_power_of_two_uintptr :: #force_inline proc "contextless" (x: uintptr) -> bool {
 	if x <= 0 {
 		return false
@@ -47,7 +65,7 @@ is_power_of_two_uintptr :: #force_inline proc "contextless" (x: uintptr) -> bool
 	return (x & (x-1)) == 0
 }

-align_forward_uintptr :: #force_inline proc(ptr, align: uintptr) -> uintptr {
+align_forward_uintptr :: #force_inline proc "odin" (ptr, align: uintptr) -> uintptr {
 	assert(is_power_of_two_uintptr(align))

 	p := ptr
@@ -58,6 +76,18 @@ align_forward_uintptr :: #force_inline proc(ptr, align: uintptr) -> uintptr {
 	return p
 }

+is_power_of_two :: proc {
+	is_power_of_two_int,
+	is_power_of_two_uint,
+	is_power_of_two_uintptr,
+}
+
+align_forward :: proc {
+	align_forward_int,
+	align_forward_uint,
+	align_forward_uintptr,
+}
+
 mem_zero :: proc "contextless" (data: rawptr, len: int) -> rawptr {
 	if data == nil {
 		return nil
@@ -453,7 +483,7 @@ quaternion256_ne :: #force_inline proc "contextless" (a, b: quaternion256) -> bo
 string_decode_rune :: #force_inline proc "contextless" (s: string) -> (rune, int) {
 	// NOTE(bill): Duplicated here to remove dependency on package unicode/utf8

-	@static accept_sizes := [256]u8{
+	@(static, rodata) accept_sizes := [256]u8{
 		0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, // 0x00-0x0f
 		0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, // 0x10-0x1f
 		0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, 0xf0, // 0x20-0x2f
@@ -474,7 +504,7 @@ string_decode_rune :: #force_inline proc "contextless" (s: string) -> (rune, int
 	}
 	Accept_Range :: struct {lo, hi: u8}

-	@static accept_ranges := [5]Accept_Range{
+	@(static, rodata) accept_ranges := [5]Accept_Range{
 		{0x80, 0xbf},
 		{0xa0, 0xbf},
 		{0x80, 0x9f},
@@ -612,21 +642,24 @@ abs_quaternion256 :: #force_inline proc "contextless" (x: quaternion256) -> f64


 quo_complex32 :: proc "contextless" (n, m: complex32) -> complex32 {
-	e, f: f16
+	nr, ni := f32(real(n)), f32(imag(n))
+	mr, mi := f32(real(m)), f32(imag(m))

-	if abs(real(m)) >= abs(imag(m)) {
-		ratio := imag(m) / real(m)
-		denom := real(m) + ratio*imag(m)
-		e = (real(n) + imag(n)*ratio) / denom
-		f = (imag(n) - real(n)*ratio) / denom
+	e, f: f32
+
+	if abs(mr) >= abs(mi) {
+		ratio := mi / mr
+		denom := mr + ratio*mi
+		e = (nr + ni*ratio) / denom
+		f = (ni - nr*ratio) / denom
 	} else {
-		ratio := real(m) / imag(m)
-		denom := imag(m) + ratio*real(m)
-		e = (real(n)*ratio + imag(n)) / denom
-		f = (imag(n)*ratio - real(n)) / denom
+		ratio := mr / mi
+		denom := mi + ratio*mr
+		e = (nr*ratio + ni) / denom
+		f = (ni*ratio - nr) / denom
 	}

-	return complex(e, f)
+	return complex(f16(e), f16(f))
 }


@@ -667,15 +700,15 @@ quo_complex128 :: proc "contextless" (n, m: complex128) -> complex128 {
 }

 mul_quaternion64 :: proc "contextless" (q, r: quaternion64) -> quaternion64 {
-	q0, q1, q2, q3 := real(q), imag(q), jmag(q), kmag(q)
-	r0, r1, r2, r3 := real(r), imag(r), jmag(r), kmag(r)
+	q0, q1, q2, q3 := f32(real(q)), f32(imag(q)), f32(jmag(q)), f32(kmag(q))
+	r0, r1, r2, r3 := f32(real(r)), f32(imag(r)), f32(jmag(r)), f32(kmag(r))

 	t0 := r0*q0 - r1*q1 - r2*q2 - r3*q3
 	t1 := r0*q1 + r1*q0 - r2*q3 + r3*q2
 	t2 := r0*q2 + r1*q3 + r2*q0 - r3*q1
 	t3 := r0*q3 - r1*q2 + r2*q1 + r3*q0

-	return quaternion(w=t0, x=t1, y=t2, z=t3)
+	return quaternion(w=f16(t0), x=f16(t1), y=f16(t2), z=f16(t3))
 }

 mul_quaternion128 :: proc "contextless" (q, r: quaternion128) -> quaternion128 {
@@ -703,8 +736,8 @@ mul_quaternion256 :: proc "contextless" (q, r: quaternion256) -> quaternion256 {
 }

 quo_quaternion64 :: proc "contextless" (q, r: quaternion64) -> quaternion64 {
-	q0, q1, q2, q3 := real(q), imag(q), jmag(q), kmag(q)
-	r0, r1, r2, r3 := real(r), imag(r), jmag(r), kmag(r)
+	q0, q1, q2, q3 := f32(real(q)), f32(imag(q)), f32(jmag(q)), f32(kmag(q))
+	r0, r1, r2, r3 := f32(real(r)), f32(imag(r)), f32(jmag(r)), f32(kmag(r))

 	invmag2 := 1.0 / (r0*r0 + r1*r1 + r2*r2 + r3*r3)

@@ -713,7 +746,7 @@ quo_quaternion64 :: proc "contextless" (q, r: quaternion64) -> quaternion64 {
 	t2 := (r0*q2 - r1*q3 - r2*q0 + r3*q1) * invmag2
 	t3 := (r0*q3 + r1*q2 + r2*q1 - r3*q0) * invmag2

-	return quaternion(w=t0, x=t1, y=t2, z=t3)
+	return quaternion(w=f16(t0), x=f16(t1), y=f16(t2), z=f16(t3))
 }

 quo_quaternion128 :: proc "contextless" (q, r: quaternion128) -> quaternion128 {
@@ -801,6 +834,10 @@ truncsfhf2 :: proc "c" (value: f32) -> __float16 {
 	}
 }

+@(link_name="__aeabi_d2h", linkage=RUNTIME_LINKAGE, require=RUNTIME_REQUIRE)
+aeabi_d2h :: proc "c" (value: f64) -> __float16 {
+	return truncsfhf2(f32(value))
+}

@(link_name="__truncdfhf2", linkage=RUNTIME_LINKAGE, require=RUNTIME_REQUIRE)
 truncdfhf2 :: proc "c" (value: f64) -> __float16 {
@@ -841,9 +878,6 @@ extendhfsf2 :: proc "c" (value: __float16) -> f32 {

@(link_name="__floattidf", linkage=RUNTIME_LINKAGE, require=RUNTIME_REQUIRE)
 floattidf :: proc "c" (a: i128) -> f64 {
-when IS_WASM {
-	return 0
-} else {
 	DBL_MANT_DIG :: 53
 	if a == 0 {
 		return 0.0
@@ -883,14 +917,10 @@ when IS_WASM {
 	fb[0] = u32(a)                           // mantissa-low
 	return transmute(f64)fb
 }
-}


@(link_name="__floattidf_unsigned", linkage=RUNTIME_LINKAGE, require=RUNTIME_REQUIRE)
 floattidf_unsigned :: proc "c" (a: u128) -> f64 {
-when IS_WASM {
-	return 0
-} else {
 	DBL_MANT_DIG :: 53
 	if a == 0 {
 		return 0.0
@@ -928,7 +958,6 @@ when IS_WASM {
 	fb[0] = u32(a)                           // mantissa-low
 	return transmute(f64)fb
 }
-}



@@ -978,26 +1007,44 @@ modti3 :: proc "c" (a, b: i128) -> i128 {
 	bn := (b ~ s_b) - s_b

 	r: u128 = ---
-	_ = udivmod128(transmute(u128)an, transmute(u128)bn, &r)
-	return (transmute(i128)r ~ s_a) - s_a
+	_ = udivmod128(u128(an), u128(bn), &r)
+	return (i128(r) ~ s_a) - s_a
 }


@(link_name="__divmodti4", linkage=RUNTIME_LINKAGE, require=RUNTIME_REQUIRE)
 divmodti4 :: proc "c" (a, b: i128, rem: ^i128) -> i128 {
-	u := udivmod128(transmute(u128)a, transmute(u128)b, cast(^u128)rem)
-	return transmute(i128)u
+	s_a := a >> (128 - 1) // -1 if negative or 0
+	s_b := b >> (128 - 1)
+	an := (a ~ s_a) - s_a // absolute
+	bn := (b ~ s_b) - s_b
+
+	s_b   ~= s_a // quotient sign
+	u_s_b := u128(s_b)
+	u_s_a := u128(s_a)
+
+	r: u128 = ---
+	u := i128((udivmodti4(u128(an), u128(bn), &r) ~ u_s_b) - u_s_b) // negate if negative
+	rem^ = i128((r ~ u_s_a) - u_s_a)
+	return u
 }

@(link_name="__divti3", linkage=RUNTIME_LINKAGE, require=RUNTIME_REQUIRE)
 divti3 :: proc "c" (a, b: i128) -> i128 {
-	u := udivmodti4(transmute(u128)a, transmute(u128)b, nil)
-	return transmute(i128)u
+	s_a := a >> (128 - 1) // -1 if negative or 0
+	s_b := b >> (128 - 1)
+	an := (a ~ s_a) - s_a // absolute
+	bn := (b ~ s_b) - s_b
+
+	s_a   ~= s_b // quotient sign
+	u_s_a := u128(s_a)
+
+	return i128((udivmodti4(u128(an), u128(bn), nil) ~ u_s_a) - u_s_a) // negate if negative
 }


@(link_name="__fixdfti", linkage=RUNTIME_LINKAGE, require=RUNTIME_REQUIRE)
-fixdfti :: proc(a: u64) -> i128 {
+fixdfti :: proc "c" (a: u64) -> i128 {
 	significandBits :: 52
 	typeWidth       :: (size_of(u64)*8)
 	exponentBits    :: (typeWidth - significandBits - 1)
@@ -1042,19 +1089,17 @@ fixdfti :: proc(a: u64) -> i128 {
 __write_bits :: proc "contextless" (dst, src: [^]byte, offset: uintptr, size: uintptr) {
 	for i in 0..<size {
 		j := offset+i
-		the_bit := byte((src[i/8]) & (1<<(i&7)) != 0)
-		b := the_bit<<(j&7)
-		dst[j/8] &~= b
-		dst[j/8] |=  b
+		the_bit := byte((src[i>>3]) & (1<<(i&7)) != 0)
+		dst[j>>3] &~=       1<<(j&7)
+		dst[j>>3]  |= the_bit<<(j&7)
 	}
 }

 __read_bits :: proc "contextless" (dst, src: [^]byte, offset: uintptr, size: uintptr) {
 	for j in 0..<size {
 		i := offset+j
-		the_bit := byte((src[i/8]) & (1<<(i&7)) != 0)
-		b := the_bit<<(j&7)
-		dst[j/8] &~= b
-		dst[j/8] |=  b
+		the_bit := byte((src[i>>3]) & (1<<(i&7)) != 0)
+		dst[j>>3] &~=       1<<(j&7)
+		dst[j>>3]  |= the_bit<<(j&7)
 	}
-}
+}
@@ -1,4 +1,4 @@
-//+build freebsd, openbsd
+//+build freebsd, openbsd, netbsd
 //+private
 package runtime

@@ -9,7 +9,11 @@ foreign libc {
 	@(link_name="write")
 	_unix_write :: proc(fd: i32, buf: rawptr, size: int) -> int ---

-	__error :: proc() -> ^i32 ---
+	when ODIN_OS == .NetBSD {
+		@(link_name="__errno") __error :: proc() -> ^i32 ---
+	} else {
+		__error :: proc() -> ^i32 ---
+	}
 }

 _stderr_write :: proc "contextless" (data: []byte) -> (int, _OS_Errno) {
@@ -5,11 +5,24 @@ package runtime
 import "base:intrinsics"

 _stderr_write :: proc "contextless" (data: []byte) -> (int, _OS_Errno) {
-	WRITE  :: 0x2000004
 	STDERR :: 2
-	ret := intrinsics.syscall(WRITE, STDERR, uintptr(raw_data(data)), uintptr(len(data)))
-	if ret < 0 {
-		return 0, _OS_Errno(-ret)
+	when ODIN_NO_CRT {
+		WRITE  :: 0x2000004
+		ret := intrinsics.syscall(WRITE, STDERR, uintptr(raw_data(data)), uintptr(len(data)))
+		if ret < 0 {
+			return 0, _OS_Errno(-ret)
+		}
+		return int(ret), 0
+	} else {
+		foreign {
+			write   :: proc(handle: i32, buffer: [^]byte, count: uint) -> int ---
+			__error :: proc() -> ^i32 ---
+		}
+
+		if ret := write(STDERR, raw_data(data), len(data)); ret >= 0 {
+			return int(ret), 0
+		}
+
+		return 0, _OS_Errno(__error()^)
 	}
-	return int(ret), 0
 }
@@ -12,6 +12,8 @@ _stderr_write :: proc "contextless" (data: []byte) -> (int, _OS_Errno) {
 		SYS_write :: uintptr(4)
 	} else when ODIN_ARCH == .arm32 {
 		SYS_write :: uintptr(4)
+	} else when ODIN_ARCH == .riscv64 {
+		SYS_write :: uintptr(64)
 	}

 	stderr :: 2
@@ -0,0 +1,43 @@
+//+build orca
+//+private
+package runtime
+
+import "base:intrinsics"
+
+// Constants allowing to specify the level of logging verbosity.
+log_level :: enum u32 {
+	// Only errors are logged.
+	ERROR = 0,
+	// Only warnings and errors are logged.
+	WARNING = 1,
+	// All messages are logged.
+	INFO = 2,
+	COUNT = 3,
+}
+
+@(default_calling_convention="c", link_prefix="oc_")
+foreign {
+	abort_ext   :: proc(file: cstring, function: cstring, line: i32, fmt: cstring, #c_vararg args: ..any) -> ! ---
+	assert_fail :: proc(file: cstring, function: cstring, line: i32, src: cstring, fmt: cstring, #c_vararg args: ..any) -> ! ---
+	log_ext     :: proc(level: log_level, function: cstring, file: cstring, line: i32, fmt: cstring, #c_vararg args: ..any) ---
+}
+
+// NOTE: This is all pretty gross, don't look.
+
+// WASM is single threaded so this should be fine.
+orca_stderr_buffer:     [4096]byte
+orca_stderr_buffer_idx: int
+
+_stderr_write :: proc "contextless" (data: []byte) -> (int, _OS_Errno) {
+	for b in data {
+		orca_stderr_buffer[orca_stderr_buffer_idx] = b
+		orca_stderr_buffer_idx += 1
+
+		if b == '\n' || orca_stderr_buffer_idx == len(orca_stderr_buffer)-1 {
+			log_ext(.ERROR, "", "", 0, cstring(raw_data(orca_stderr_buffer[:orca_stderr_buffer_idx])))
+			orca_stderr_buffer_idx = 0
+		}
+	}
+
+	return len(data), 0
+}
@@ -2,10 +2,54 @@
 //+private
 package runtime

-import "core:sys/wasm/wasi"
+foreign import wasi "wasi_snapshot_preview1"
+
+@(default_calling_convention="contextless")
+foreign wasi {
+	fd_write :: proc(
+		fd: i32,
+		iovs: [][]byte,
+		n: ^uint,
+	) -> u16 ---
+
+	@(private="file")
+	args_sizes_get :: proc(
+		num_of_args:  ^uint,
+		size_of_args: ^uint,
+	) -> u16 ---
+
+	@(private="file")
+	args_get :: proc(
+		argv:     [^]cstring,
+		argv_buf: [^]byte,
+	) -> u16 ---
+}

 _stderr_write :: proc "contextless" (data: []byte) -> (int, _OS_Errno) {
-	data := (wasi.ciovec_t)(data)
-	n, err := wasi.fd_write(1, {data})
+	n: uint
+	err := fd_write(1, {data}, &n)
 	return int(n), _OS_Errno(err)
 }
+
+_wasi_setup_args :: proc() {
+	num_of_args, size_of_args: uint
+	if errno := args_sizes_get(&num_of_args, &size_of_args); errno != 0 {
+		return
+	}
+
+	err: Allocator_Error
+	if args__, err = make([]cstring, num_of_args); err != nil {
+		return
+	}
+
+	args_buf: []byte
+	if args_buf, err = make([]byte, size_of_args); err != nil {
+		delete(args__)
+		return
+	}
+
+	if errno := args_get(raw_data(args__), raw_data(args_buf)); errno != 0 {
+		delete(args__)
+		delete(args_buf)
+	}
+}
@@ -6,7 +6,7 @@ _INTEGER_DIGITS :: "0123456789abcdefghijklmnopqrstuvwxyz"
 _INTEGER_DIGITS_VAR := _INTEGER_DIGITS

 when !ODIN_NO_RTTI {
-	print_any_single :: proc "contextless" (arg: any) {
+	print_any_single :: #force_no_inline proc "contextless" (arg: any) {
 		x := arg
 		if x.data == nil {
 			print_string("nil")
@@ -72,7 +72,7 @@ when !ODIN_NO_RTTI {
 			print_string("<invalid-value>")
 		}
 	}
-	println_any :: proc "contextless" (args: ..any) {
+	println_any :: #force_no_inline proc "contextless" (args: ..any) {
 		context = default_context()
 		loop: for arg, i in args {
 			assert(arg.id != nil)
@@ -122,12 +122,12 @@ encode_rune :: proc "contextless" (c: rune) -> ([4]u8, int) {
 	return buf, 4
 }

-print_string :: proc "contextless" (str: string) -> (n: int) {
+print_string :: #force_no_inline proc "contextless" (str: string) -> (n: int) {
 	n, _ = stderr_write(transmute([]byte)str)
 	return
 }

-print_strings :: proc "contextless" (args: ..string) -> (n: int) {
+print_strings :: #force_no_inline proc "contextless" (args: ..string) -> (n: int) {
 	for str in args {
 		m, err := stderr_write(transmute([]byte)str)
 		n += m
@@ -138,12 +138,12 @@ print_strings :: proc "contextless" (args: ..string) -> (n: int) {
 	return
 }

-print_byte :: proc "contextless" (b: byte) -> (n: int) {
+print_byte :: #force_no_inline proc "contextless" (b: byte) -> (n: int) {
 	n, _ = stderr_write([]byte{b})
 	return
 }

-print_encoded_rune :: proc "contextless" (r: rune) {
+print_encoded_rune :: #force_no_inline proc "contextless" (r: rune) {
 	print_byte('\'')

 	switch r {
@@ -170,7 +170,7 @@ print_encoded_rune :: proc "contextless" (r: rune) {
 	print_byte('\'')
 }

-print_rune :: proc "contextless" (r: rune) -> int #no_bounds_check {
+print_rune :: #force_no_inline proc "contextless" (r: rune) -> int #no_bounds_check {
 	RUNE_SELF :: 0x80

 	if r < RUNE_SELF {
@@ -183,7 +183,7 @@ print_rune :: proc "contextless" (r: rune) -> int #no_bounds_check {
 }


-print_u64 :: proc "contextless" (x: u64) #no_bounds_check {
+print_u64 :: #force_no_inline proc "contextless" (x: u64) #no_bounds_check {
 	a: [129]byte
 	i := len(a)
 	b := u64(10)
@@ -198,7 +198,7 @@ print_u64 :: proc "contextless" (x: u64) #no_bounds_check {
 }


-print_i64 :: proc "contextless" (x: i64) #no_bounds_check {
+print_i64 :: #force_no_inline proc "contextless" (x: i64) #no_bounds_check {
 	b :: i64(10)

 	u := x
@@ -223,25 +223,29 @@ print_uint    :: proc "contextless" (x: uint)    { print_u64(u64(x)) }
 print_uintptr :: proc "contextless" (x: uintptr) { print_u64(u64(x)) }
 print_int     :: proc "contextless" (x: int)     { print_i64(i64(x)) }

-print_caller_location :: proc "contextless" (loc: Source_Code_Location) {
+print_caller_location :: #force_no_inline proc "contextless" (loc: Source_Code_Location) {
 	print_string(loc.file_path)
 	when ODIN_ERROR_POS_STYLE == .Default {
 		print_byte('(')
 		print_u64(u64(loc.line))
-		print_byte(':')
-		print_u64(u64(loc.column))
+		if loc.column != 0 {
+			print_byte(':')
+			print_u64(u64(loc.column))
+		}
 		print_byte(')')
 	} else when ODIN_ERROR_POS_STYLE == .Unix {
 		print_byte(':')
 		print_u64(u64(loc.line))
-		print_byte(':')
-		print_u64(u64(loc.column))
+		if loc.column != 0 {
+			print_byte(':')
+			print_u64(u64(loc.column))
+		}
 		print_byte(':')
 	} else {
 		#panic("unhandled ODIN_ERROR_POS_STYLE")
 	}
 }
-print_typeid :: proc "contextless" (id: typeid) {
+print_typeid :: #force_no_inline proc "contextless" (id: typeid) {
 	when ODIN_NO_RTTI {
 		if id == nil {
 			print_string("nil")
@@ -257,7 +261,9 @@ print_typeid :: proc "contextless" (id: typeid) {
 		}
 	}
 }
-print_type :: proc "contextless" (ti: ^Type_Info) {
+
+@(optimization_mode="favor_size")
+print_type :: #force_no_inline proc "contextless" (ti: ^Type_Info) {
 	if ti == nil {
 		print_string("nil")
 		return
@@ -395,15 +401,16 @@ print_type :: proc "contextless" (ti: ^Type_Info) {
 		}

 		print_string("struct ")
-		if info.is_packed    { print_string("#packed ") }
-		if info.is_raw_union { print_string("#raw_union ") }
-		if info.custom_align {
+		if .packed    in info.flags { print_string("#packed ") }
+		if .raw_union in info.flags { print_string("#raw_union ") }
+		if .no_copy   in info.flags { print_string("#no_copy ") }
+		if .align in info.flags {
 			print_string("#align(")
 			print_u64(u64(ti.align))
 			print_string(") ")
 		}
 		print_byte('{')
-		for name, i in info.names {
+		for name, i in info.names[:info.field_count] {
 			if i > 0 { print_string(", ") }
 			print_string(name)
 			print_string(": ")
@@ -463,7 +470,7 @@ print_type :: proc "contextless" (ti: ^Type_Info) {
 		print_string("bit_field ")
 		print_type(info.backing_type)
 		print_string(" {")
-		for name, i in info.names {
+		for name, i in info.names[:info.field_count] {
 			if i > 0 { print_string(", ") }
 			print_string(name)
 			print_string(": ")
@@ -25,13 +25,19 @@ when ODIN_NO_CRT && ODIN_OS == .Windows {
 		RtlMoveMemory(dst, src, len)
 		return dst
 	}
-} else when ODIN_NO_CRT || (ODIN_ARCH == .wasm32 || ODIN_ARCH == .wasm64p32) {
+} else when ODIN_NO_CRT || (ODIN_OS != .Orca && (ODIN_ARCH == .wasm32 || ODIN_ARCH == .wasm64p32)) {
+	// NOTE: on wasm, calls to these procs are generated (by LLVM) with type `i32` instead of `int`.
+	//
+	// NOTE: `#any_int` is also needed, because calls that we generate (and package code)
+	//       will be using `int` and need to be converted.
+	int_t :: i32 when ODIN_ARCH == .wasm64p32 else int
+
 	@(link_name="memset", linkage="strong", require)
-	memset :: proc "c" (ptr: rawptr, val: i32, len: int) -> rawptr {
+	memset :: proc "c" (ptr: rawptr, val: i32, #any_int len: int_t) -> rawptr {
 		if ptr != nil && len != 0 {
 			b := byte(val)
 			p := ([^]byte)(ptr)
-			for i := 0; i < len; i += 1 {
+			for i := int_t(0); i < len; i += 1 {
 				p[i] = b
 			}
 		}
@@ -39,10 +45,10 @@ when ODIN_NO_CRT && ODIN_OS == .Windows {
 	}

 	@(link_name="bzero", linkage="strong", require)
-	bzero :: proc "c" (ptr: rawptr, len: int) -> rawptr {
+	bzero :: proc "c" (ptr: rawptr, #any_int len: int_t) -> rawptr {
 		if ptr != nil && len != 0 {
 			p := ([^]byte)(ptr)
-			for i := 0; i < len; i += 1 {
+			for i := int_t(0); i < len; i += 1 {
 				p[i] = 0
 			}
 		}
@@ -50,7 +56,7 @@ when ODIN_NO_CRT && ODIN_OS == .Windows {
 	}

 	@(link_name="memmove", linkage="strong", require)
-	memmove :: proc "c" (dst, src: rawptr, len: int) -> rawptr {
+	memmove :: proc "c" (dst, src: rawptr, #any_int len: int_t) -> rawptr {
 		d, s := ([^]byte)(dst), ([^]byte)(src)
 		if d == s || len == 0 {
 			return dst
@@ -63,7 +69,7 @@ when ODIN_NO_CRT && ODIN_OS == .Windows {
 		}

 		if s > d && uintptr(s)-uintptr(d) < uintptr(len) {
-			for i := 0; i < len; i += 1 {
+			for i := int_t(0); i < len; i += 1 {
 				d[i] = s[i]
 			}
 			return dst
@@ -71,10 +77,10 @@ when ODIN_NO_CRT && ODIN_OS == .Windows {
 		return memcpy(dst, src, len)
 	}
 	@(link_name="memcpy", linkage="strong", require)
-	memcpy :: proc "c" (dst, src: rawptr, len: int) -> rawptr {
+	memcpy :: proc "c" (dst, src: rawptr, #any_int len: int_t) -> rawptr {
 		d, s := ([^]byte)(dst), ([^]byte)(src)
 		if d != s {
-			for i := 0; i < len; i += 1 {
+			for i := int_t(0); i < len; i += 1 {
 				d[i] = s[i]
 			}
 		}
@@ -92,4 +98,4 @@ when ODIN_NO_CRT && ODIN_OS == .Windows {
 		}
 		return ptr
 	}
-}
+}
@@ -52,3 +52,24 @@ udivti3 :: proc "c" (la, ha, lb, hb: u64) -> u128 {
 	b.lo, b.hi = lb, hb
 	return udivmodti4(a.all, b.all, nil)
 }
+
+@(link_name="__lshrti3", linkage="strong")
+__lshrti3 :: proc "c" (la, ha: u64, b: u32) -> i128 {
+	bits :: size_of(u32)*8
+
+	input, result: ti_int
+	input.lo = la
+	input.hi = ha
+
+	if b & bits != 0 {
+		result.hi = 0
+		result.lo = input.hi >> (b - bits)
+	} else if b == 0 {
+		return input.all
+	} else {
+		result.hi = input.hi >> b
+		result.lo = (input.hi << (bits - b)) | (input.lo >> b)
+	}
+
+	return result.all
+}
@@ -0,0 +1,127 @@
+package runtime
+
+import "base:intrinsics"
+
+@(require_results)
+random_generator_read_bytes :: proc(rg: Random_Generator, p: []byte) -> bool {
+	if rg.procedure != nil {
+		rg.procedure(rg.data, .Read, p)
+		return true
+	}
+	return false
+}
+
+@(require_results)
+random_generator_read_ptr :: proc(rg: Random_Generator, p: rawptr, len: uint) -> bool {
+	if rg.procedure != nil {
+		rg.procedure(rg.data, .Read, ([^]byte)(p)[:len])
+		return true
+	}
+	return false
+}
+
+@(require_results)
+random_generator_query_info :: proc(rg: Random_Generator) -> (info: Random_Generator_Query_Info) {
+	if rg.procedure != nil {
+		rg.procedure(rg.data, .Query_Info, ([^]byte)(&info)[:size_of(info)])
+	}
+	return
+}
+
+
+random_generator_reset_bytes :: proc(rg: Random_Generator, p: []byte) {
+	if rg.procedure != nil {
+		rg.procedure(rg.data, .Reset, p)
+	}
+}
+
+random_generator_reset_u64 :: proc(rg: Random_Generator, p: u64) {
+	if rg.procedure != nil {
+		p := p
+		rg.procedure(rg.data, .Reset, ([^]byte)(&p)[:size_of(p)])
+	}
+}
+
+
+Default_Random_State :: struct {
+	state: u64,
+	inc:   u64,
+}
+
+default_random_generator_proc :: proc(data: rawptr, mode: Random_Generator_Mode, p: []byte) {
+	@(require_results)
+	read_u64 :: proc "contextless" (r: ^Default_Random_State) -> u64 {
+		old_state := r.state
+		r.state = old_state * 6364136223846793005 + (r.inc|1)
+		xor_shifted := (((old_state >> 59) + 5) ~ old_state) * 12605985483714917081
+		rot := (old_state >> 59)
+		return (xor_shifted >> rot) | (xor_shifted << ((-rot) & 63))
+	}
+
+	@(thread_local)
+	global_rand_seed: Default_Random_State
+
+	init :: proc "contextless" (r: ^Default_Random_State, seed: u64) {
+		seed := seed
+		if seed == 0 {
+			seed = u64(intrinsics.read_cycle_counter())
+		}
+		r.state = 0
+		r.inc = (seed << 1) | 1
+		_ = read_u64(r)
+		r.state += seed
+		_ = read_u64(r)
+	}
+
+	r: ^Default_Random_State = ---
+	if data == nil {
+		r = &global_rand_seed
+	} else {
+		r = cast(^Default_Random_State)data
+	}
+
+	switch mode {
+	case .Read:
+		if r.state == 0 && r.inc == 0 {
+			init(r, 0)
+		}
+
+		switch len(p) {
+		case size_of(u64):
+			// Fast path for a 64-bit destination.
+			intrinsics.unaligned_store((^u64)(raw_data(p)), read_u64(r))
+		case:
+			// All other cases.
+			pos := i8(0)
+			val := u64(0)
+			for &v in p {
+				if pos == 0 {
+					val = read_u64(r)
+					pos = 7
+				}
+				v = byte(val)
+				val >>= 8
+				pos -= 1
+			}
+		}
+
+	case .Reset:
+		seed: u64
+		mem_copy_non_overlapping(&seed, raw_data(p), min(size_of(seed), len(p)))
+		init(r, seed)
+
+	case .Query_Info:
+		if len(p) != size_of(Random_Generator_Query_Info) {
+			return
+		}
+		info := (^Random_Generator_Query_Info)(raw_data(p))
+		info^ += {.Uniform, .Resettable}
+	}
+}
+
+default_random_generator :: proc "contextless" (state: ^Default_Random_State = nil) -> Random_Generator {
+	return {
+		procedure = default_random_generator_proc,
+		data = state,
+	}
+}
@@ -0,0 +1,34 @@
+package runtime
+
+Thread_Local_Cleaner :: #type proc "odin" ()
+
+@(private="file")
+thread_local_cleaners: [8]Thread_Local_Cleaner
+
+// Add a procedure that will be run at the end of a thread for the purpose of
+// deallocating state marked as `thread_local`.
+//
+// Intended to be called in an `init` procedure of a package with
+// dynamically-allocated memory that is stored in `thread_local` variables.
+add_thread_local_cleaner :: proc "contextless" (p: Thread_Local_Cleaner) {
+	for &v in thread_local_cleaners {
+		if v == nil {
+			v = p
+			return
+		}
+	}
+	panic_contextless("There are no more thread-local cleaner slots available.")
+}
+
+// Run all of the thread-local cleaner procedures.
+//
+// Intended to be called by the internals of a threading API at the end of a
+// thread's lifetime.
+run_thread_local_cleaners :: proc "odin" () {
+	for p in thread_local_cleaners {
+		if p == nil {
+			break
+		}
+		p()
+	}
+}
@@ -58,7 +58,7 @@ udivmod128 :: proc "c" (a, b: u128, rem: ^u128) -> u128 {
 			return u128(n[high] >> _ctz(d[high]))
 		}

-		sr = transmute(u32)(i32(_clz(d[high])) - i32(_clz(n[high])))
+		sr = u32(i32(_clz(d[high])) - i32(_clz(n[high])))
 		if sr > U64_BITS - 2 {
 			if rem != nil {
 				rem^ = a
@@ -107,7 +107,7 @@ udivmod128 :: proc "c" (a, b: u128, rem: ^u128) -> u128 {
 				r[low]  = n[high] >> (sr - U64_BITS)
 			}
 		} else {
-			sr = transmute(u32)(i32(_clz(d[high])) - i32(_clz(n[high])))
+			sr = u32(i32(_clz(d[high])) - i32(_clz(n[high])))

 			if sr > U64_BITS - 1 {
 				if rem != nil {
@@ -143,7 +143,7 @@ udivmod128 :: proc "c" (a, b: u128, rem: ^u128) -> u128 {
 		r_all = transmute(u128)r
 		s := i128(b - r_all - 1) >> (U128_BITS - 1)
 		carry = u32(s & 1)
-		r_all -= b & transmute(u128)s
+		r_all -= b & u128(s)
 		r = transmute([2]u64)r_all
 	}

@@ -0,0 +1,871 @@
+//+build wasm32, wasm64p32
+package runtime
+
+import "base:intrinsics"
+
+/*
+Port of emmalloc, modified for use in Odin.
+
+Invariants:
+	- Per-allocation header overhead is 8 bytes, smallest allocated payload
+	  amount is 8 bytes, and a multiple of 4 bytes.
+	- Acquired memory blocks are subdivided into disjoint regions that lie
+	  next to each other.
+	- A region is either in used or free.
+	  Used regions may be adjacent, and a used and unused region
+	  may be adjacent, but not two unused ones - they would be
+	  merged.
+	- Memory allocation takes constant time, unless the alloc needs to wasm_memory_grow()
+	  or memory is very close to being exhausted.
+	- Free and used regions are managed inside "root regions", which are slabs
+	  of memory acquired via wasm_memory_grow().
+	- Memory retrieved using wasm_memory_grow() can not be given back to the OS.
+	  Therefore, frees are internal to the allocator.
+
+Copyright (c) 2010-2014 Emscripten authors, see AUTHORS file.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in
+all copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
+THE SOFTWARE.
+*/
+
+WASM_Allocator :: struct #no_copy {
+	// The minimum alignment of allocations.
+	alignment: uint,
+	// A region that contains as payload a single forward linked list of pointers to
+	// root regions of each disjoint region blocks.
+	list_of_all_regions: ^Root_Region,
+	// For each of the buckets, maintain a linked list head node. The head node for each
+	// free region is a sentinel node that does not actually represent any free space, but
+	// the sentinel is used to avoid awkward testing against (if node == freeRegionHeadNode)
+	// when adding and removing elements from the linked list, i.e. we are guaranteed that
+	// the sentinel node is always fixed and there, and the actual free region list elements
+	// start at free_region_buckets[i].next each.
+	free_region_buckets: [NUM_FREE_BUCKETS]Region,
+	// A bitmask that tracks the population status for each of the 64 distinct memory regions:
+	// a zero at bit position i means that the free list bucket i is empty. This bitmask is
+	// used to avoid redundant scanning of the 64 different free region buckets: instead by
+	// looking at the bitmask we can find in constant time an index to a free region bucket
+	// that contains free memory of desired size.
+	free_region_buckets_used: BUCKET_BITMASK_T,
+	// Because wasm memory can only be allocated in pages of 64k at a time, we keep any
+	// spilled/unused bytes that are left from the allocated pages here, first using this
+	// when bytes are needed.
+	spill: []byte,
+	// Mutex for thread safety, only used if the target feature "atomics" is enabled.
+	mu: Mutex_State,
+}
+
+// Not required to be called, called on first allocation otherwise.
+wasm_allocator_init :: proc(a: ^WASM_Allocator, alignment: uint = 8) {
+	assert(is_power_of_two(alignment), "alignment must be a power of two")
+	assert(alignment > 4, "alignment must be more than 4")
+
+	a.alignment = alignment
+
+	for i in 0..<NUM_FREE_BUCKETS {
+		a.free_region_buckets[i].next = &a.free_region_buckets[i]
+		a.free_region_buckets[i].prev = a.free_region_buckets[i].next
+	}
+
+	if !claim_more_memory(a, 3*size_of(Region)) {
+		panic("wasm_allocator: initial memory could not be allocated")
+	}
+}
+
+global_default_wasm_allocator_data: WASM_Allocator
+
+default_wasm_allocator :: proc() -> Allocator {
+	return wasm_allocator(&global_default_wasm_allocator_data)
+}
+
+wasm_allocator :: proc(a: ^WASM_Allocator) -> Allocator {
+	return {
+		data      = a,
+		procedure = wasm_allocator_proc,
+	}
+}
+
+wasm_allocator_proc :: proc(a: rawptr, mode: Allocator_Mode, size, alignment: int, old_memory: rawptr, old_size: int, loc := #caller_location) -> ([]byte, Allocator_Error) {
+	a := (^WASM_Allocator)(a)
+	if a == nil {
+		a = &global_default_wasm_allocator_data
+	}
+
+	if a.alignment == 0 {
+		wasm_allocator_init(a)
+	}
+
+	switch mode {
+	case .Alloc:
+		ptr := aligned_alloc(a, uint(alignment), uint(size), loc)
+		if ptr == nil {
+			return nil, .Out_Of_Memory
+		}
+		intrinsics.mem_zero(ptr, size)
+		return ([^]byte)(ptr)[:size], nil
+
+	case .Alloc_Non_Zeroed:
+		ptr := aligned_alloc(a, uint(alignment), uint(size), loc)
+		if ptr == nil {
+			return nil, .Out_Of_Memory
+		}
+		return ([^]byte)(ptr)[:size], nil
+
+	case .Resize:
+		ptr := aligned_realloc(a, old_memory, uint(alignment), uint(size), loc)
+		if ptr == nil {
+			return nil, .Out_Of_Memory
+		}
+
+		bytes := ([^]byte)(ptr)[:size]
+
+		if size > old_size {
+			new_region := raw_data(bytes[old_size:])
+			intrinsics.mem_zero(new_region, size - old_size)
+		}
+
+		return bytes, nil
+
+	case .Resize_Non_Zeroed:
+		ptr := aligned_realloc(a, old_memory, uint(alignment), uint(size), loc)
+		if ptr == nil {
+			return nil, .Out_Of_Memory
+		}
+		return ([^]byte)(ptr)[:size], nil
+
+	case .Free:
+		free(a, old_memory, loc)
+		return nil, nil
+
+	case .Free_All, .Query_Info:
+		return nil, .Mode_Not_Implemented
+
+	case .Query_Features:
+		set := (^Allocator_Mode_Set)(old_memory)
+		if set != nil {
+			set^ = {.Alloc, .Alloc_Non_Zeroed, .Free, .Resize, .Resize_Non_Zeroed, .Query_Features }
+		}
+		return nil, nil
+	}
+
+	unreachable()
+}
+
+// Returns the allocated size of the allocator (both free and used).
+// If `nil` is given, the global allocator is used.
+wasm_allocator_size :: proc(a: ^WASM_Allocator = nil) -> (size: uint) {
+	a := a
+	if a == nil {
+		a = &global_default_wasm_allocator_data
+	}
+
+	lock(a)
+	defer unlock(a)
+
+	root := a.list_of_all_regions
+	for root != nil {
+		size += uint(uintptr(root.end_ptr) - uintptr(root))
+		root = root.next
+	}
+
+	size += len(a.spill)
+
+	return
+}
+
+// Returns the amount of free memory on the allocator.
+// If `nil` is given, the global allocator is used.
+wasm_allocator_free_space :: proc(a: ^WASM_Allocator = nil) -> (free: uint) {
+	a := a
+	if a == nil {
+		a = &global_default_wasm_allocator_data
+	}
+
+	lock(a)
+	defer unlock(a)
+
+	bucket_index: u64 = 0
+	bucket_mask := a.free_region_buckets_used
+
+	for bucket_mask != 0 {
+		index_add := intrinsics.count_trailing_zeros(bucket_mask)
+		bucket_index += index_add
+		bucket_mask >>= index_add
+		for free_region := a.free_region_buckets[bucket_index].next; free_region != &a.free_region_buckets[bucket_index]; free_region = free_region.next {
+			free += free_region.size - REGION_HEADER_SIZE
+		}
+		bucket_index += 1
+		bucket_mask >>= 1
+	}
+
+	free += len(a.spill)
+
+	return
+}
+
+@(private="file")
+NUM_FREE_BUCKETS :: 64
+@(private="file")
+BUCKET_BITMASK_T :: u64
+
+// Dynamic memory is subdivided into regions, in the format
+
+// <size:u32> ..... <size:u32> | <size:u32> ..... <size:u32> | <size:u32> ..... <size:u32> | .....
+
+// That is, at the bottom and top end of each memory region, the size of that region is stored. That allows traversing the
+// memory regions backwards and forwards. Because each allocation must be at least a multiple of 4 bytes, the lowest two bits of
+// each size field is unused. Free regions are distinguished by used regions by having the FREE_REGION_FLAG bit present
+// in the size field. I.e. for free regions, the size field is odd, and for used regions, the size field reads even.
+@(private="file")
+FREE_REGION_FLAG :: 0x1
+
+// Attempts to alloc more than this many bytes would cause an overflow when calculating the size of a region,
+// therefore allocations larger than this are short-circuited immediately on entry.
+@(private="file")
+MAX_ALLOC_SIZE :: 0xFFFFFFC7
+
+// A free region has the following structure:
+// <size:uint> <prevptr> <nextptr> ... <size:uint>
+
+@(private="file")
+Region :: struct {
+	size: uint,
+	prev, next: ^Region,
+	_at_the_end_of_this_struct_size: uint,
+}
+
+// Each memory block starts with a Root_Region at the beginning.
+// The Root_Region specifies the size of the region block, and forms a linked
+// list of all Root_Regions in the program, starting with `list_of_all_regions`
+// below.
+@(private="file")
+Root_Region :: struct {
+	size:    u32,
+	next:    ^Root_Region,
+	end_ptr: ^byte,
+}
+
+@(private="file")
+Mutex_State :: enum u32 {
+	Unlocked = 0,
+	Locked   = 1,
+	Waiting  = 2,
+}
+
+@(private="file")
+lock :: proc(a: ^WASM_Allocator) {
+	when intrinsics.has_target_feature("atomics") {
+		@(cold)
+		lock_slow :: proc(a: ^WASM_Allocator, curr_state: Mutex_State) {
+			new_state := curr_state // Make a copy of it
+
+			spin_lock: for spin in 0..<i32(100) {
+				state, ok := intrinsics.atomic_compare_exchange_weak_explicit(&a.mu, .Unlocked, new_state, .Acquire, .Consume)
+				if ok {
+					return
+				}
+
+				if state == .Waiting {
+					break spin_lock
+				}
+
+				for i := min(spin+1, 32); i > 0; i -= 1 {
+					intrinsics.cpu_relax()
+				}
+			}
+
+			// Set just in case 100 iterations did not do it
+			new_state = .Waiting
+
+			for {
+				if intrinsics.atomic_exchange_explicit(&a.mu, .Waiting, .Acquire) == .Unlocked {
+					return
+				}
+
+				ret := intrinsics.wasm_memory_atomic_wait32((^u32)(&a.mu), u32(new_state), -1)
+				assert(ret != 0)
+				intrinsics.cpu_relax()
+			}
+		}
+
+
+		if v := intrinsics.atomic_exchange_explicit(&a.mu, .Locked, .Acquire); v != .Unlocked {
+			lock_slow(a, v)
+		}
+	}
+}
+
+@(private="file")
+unlock :: proc(a: ^WASM_Allocator) {
+	when intrinsics.has_target_feature("atomics") {
+		@(cold)
+		unlock_slow :: proc(a: ^WASM_Allocator) {
+			for {
+				s := intrinsics.wasm_memory_atomic_notify32((^u32)(&a.mu), 1)
+				if s >= 1 {
+					return
+				}
+			}
+		}
+
+		switch intrinsics.atomic_exchange_explicit(&a.mu, .Unlocked, .Release) {
+		case .Unlocked:
+			unreachable()
+		case .Locked:
+		// Okay
+		case .Waiting:
+			unlock_slow(a)
+		}
+	}
+}
+
+@(private="file")
+assert_locked :: proc(a: ^WASM_Allocator) {
+	when intrinsics.has_target_feature("atomics") {
+		assert(intrinsics.atomic_load(&a.mu) != .Unlocked)
+	}
+}
+
+@(private="file")
+has_alignment_uintptr :: proc(ptr: uintptr, #any_int alignment: uintptr) -> bool {
+	return ptr & (alignment-1) == 0
+}
+
+@(private="file")
+has_alignment_uint :: proc(ptr: uint, alignment: uint) -> bool {
+	return ptr & (alignment-1) == 0
+}
+
+@(private="file")
+has_alignment :: proc {
+	has_alignment_uintptr,
+	has_alignment_uint,
+}
+
+@(private="file")
+REGION_HEADER_SIZE :: 2*size_of(uint)
+
+@(private="file")
+SMALLEST_ALLOCATION_SIZE :: 2*size_of(rawptr)
+
+// Subdivide regions of free space into distinct circular doubly linked lists, where each linked list
+// represents a range of free space blocks. The following function compute_free_list_bucket() converts
+// an allocation size to the bucket index that should be looked at.
+#assert(NUM_FREE_BUCKETS == 64, "Following function is tailored specifically for the NUM_FREE_BUCKETS == 64 case")
+@(private="file")
+compute_free_list_bucket :: proc(size: uint) -> uint {
+	if size < 128 { return (size >> 3) - 1 }
+
+	clz := intrinsics.count_leading_zeros(i32(size))
+	bucket_index: i32 = ((clz > 19) \
+		?     110 - (clz<<2) + ((i32)(size >> (u32)(29-clz)) ~ 4) \
+		: min( 71 - (clz<<1) + ((i32)(size >> (u32)(30-clz)) ~ 2), NUM_FREE_BUCKETS-1))
+
+	assert(bucket_index >= 0)
+	assert(bucket_index < NUM_FREE_BUCKETS)
+	return uint(bucket_index)
+}
+
+@(private="file")
+prev_region :: proc(region: ^Region) -> ^Region {
+	prev_region_size := ([^]uint)(region)[-1]
+	prev_region_size  = prev_region_size & ~uint(FREE_REGION_FLAG)
+	return (^Region)(uintptr(region)-uintptr(prev_region_size))
+}
+
+@(private="file")
+next_region :: proc(region: ^Region) -> ^Region {
+	return (^Region)(uintptr(region)+uintptr(region.size))
+}
+
+@(private="file")
+region_ceiling_size :: proc(region: ^Region) -> uint {
+	return ([^]uint)(uintptr(region)+uintptr(region.size))[-1]
+}
+
+@(private="file")
+region_is_free :: proc(r: ^Region) -> bool {
+	return region_ceiling_size(r) & FREE_REGION_FLAG >= 1
+}
+
+@(private="file")
+region_is_in_use :: proc(r: ^Region) -> bool {
+	return r.size == region_ceiling_size(r)
+}
+
+@(private="file")
+region_payload_start_ptr :: proc(r: ^Region) -> [^]byte {
+	return ([^]byte)(r)[size_of(uint):]
+}
+
+@(private="file")
+region_payload_end_ptr :: proc(r: ^Region) -> [^]byte {
+	return ([^]byte)(r)[r.size-size_of(uint):]
+}
+
+@(private="file")
+create_used_region :: proc(ptr: rawptr, size: uint) {
+	assert(has_alignment(uintptr(ptr), size_of(uint)))
+	assert(has_alignment(size, size_of(uint)))
+	assert(size >= size_of(Region))
+
+	uptr := ([^]uint)(ptr)
+	uptr[0] = size
+	uptr[size/size_of(uint)-1] = size
+}
+
+@(private="file")
+create_free_region :: proc(ptr: rawptr, size: uint) {
+	assert(has_alignment(uintptr(ptr), size_of(uint)))
+	assert(has_alignment(size, size_of(uint)))
+	assert(size >= size_of(Region))
+
+	free_region := (^Region)(ptr)
+	free_region.size = size
+	([^]uint)(ptr)[size/size_of(uint)-1] = size | FREE_REGION_FLAG
+}
+
+@(private="file")
+prepend_to_free_list :: proc(region: ^Region, prepend_to: ^Region) {
+	assert(region_is_free(region))
+	region.next = prepend_to
+	region.prev = prepend_to.prev
+	prepend_to.prev = region
+	region.prev.next = region
+}
+
+@(private="file")
+unlink_from_free_list :: proc(region: ^Region) {
+	assert(region_is_free(region))
+	region.prev.next = region.next
+	region.next.prev = region.prev
+}
+
+@(private="file")
+link_to_free_list :: proc(a: ^WASM_Allocator, free_region: ^Region) {
+	assert(free_region.size >= size_of(Region))
+	bucket_index := compute_free_list_bucket(free_region.size-REGION_HEADER_SIZE)
+	free_list_head := &a.free_region_buckets[bucket_index]
+	free_region.prev = free_list_head
+	free_region.next = free_list_head.next
+	free_list_head.next = free_region
+	free_region.next.prev = free_region
+	a.free_region_buckets_used |= BUCKET_BITMASK_T(1) << bucket_index
+}
+
+@(private="file")
+claim_more_memory :: proc(a: ^WASM_Allocator, num_bytes: uint) -> bool {
+
+	PAGE_SIZE :: 64 * 1024
+
+	page_alloc :: proc(page_count: int) -> []byte {
+		prev_page_count := intrinsics.wasm_memory_grow(0, uintptr(page_count))
+		if prev_page_count < 0 { return nil }
+
+		ptr := ([^]byte)(uintptr(prev_page_count) * PAGE_SIZE)
+		return ptr[:page_count * PAGE_SIZE]
+	}
+
+	alloc :: proc(a: ^WASM_Allocator, num_bytes: uint) -> (bytes: [^]byte) #no_bounds_check {
+		if uint(len(a.spill)) >= num_bytes {
+			bytes = raw_data(a.spill[:num_bytes])
+			a.spill = a.spill[num_bytes:]
+			return
+		}
+
+		pages := int((num_bytes / PAGE_SIZE) + 1)
+		allocated := page_alloc(pages)
+		if allocated == nil { return nil }
+
+		// If the allocated memory is a direct continuation of the spill from before,
+		// we can just extend the spill.
+		spill_end := uintptr(raw_data(a.spill)) + uintptr(len(a.spill))
+		if spill_end == uintptr(raw_data(allocated)) {
+			raw_spill := (^Raw_Slice)(&a.spill)
+			raw_spill.len += len(allocated)
+		} else {
+			// Otherwise, we have to "waste" the previous spill.
+			// Now this is probably uncommon, and will only happen if another code path
+			// is also requesting pages.
+			a.spill = allocated
+		}
+
+		bytes = raw_data(a.spill)
+		a.spill = a.spill[num_bytes:]
+		return
+	}
+
+	num_bytes := num_bytes
+	num_bytes  = align_forward(num_bytes, a.alignment)
+
+	start_ptr := alloc(a, uint(num_bytes))
+	if start_ptr == nil { return false }
+
+	assert(has_alignment(uintptr(start_ptr), align_of(uint)))
+	end_ptr := start_ptr[num_bytes:]
+
+	end_sentinel_region := (^Region)(end_ptr[-size_of(Region):])
+	create_used_region(end_sentinel_region, size_of(Region))
+
+	// If we are the sole user of wasm_memory_grow(), it will feed us continuous/consecutive memory addresses - take advantage
+	// of that if so: instead of creating two disjoint memory regions blocks, expand the previous one to a larger size.
+	prev_alloc_end_address := a.list_of_all_regions != nil ? a.list_of_all_regions.end_ptr : nil
+	if start_ptr == prev_alloc_end_address {
+		prev_end_sentinel := prev_region((^Region)(start_ptr))
+		assert(region_is_in_use(prev_end_sentinel))
+		prev_region := prev_region(prev_end_sentinel)
+
+		a.list_of_all_regions.end_ptr = end_ptr
+
+		// Two scenarios, either the last region of the previous block was in use, in which case we need to create
+		// a new free region in the newly allocated space; or it was free, in which case we can extend that region
+		// to cover a larger size.
+		if region_is_free(prev_region) {
+			new_free_region_size := uint(uintptr(end_sentinel_region) - uintptr(prev_region))
+			unlink_from_free_list(prev_region)
+			create_free_region(prev_region, new_free_region_size)
+			link_to_free_list(a, prev_region)
+			return true
+		}
+
+		start_ptr = start_ptr[-size_of(Region):]
+	} else {
+		create_used_region(start_ptr, size_of(Region))
+
+		new_region_block := (^Root_Region)(start_ptr)
+		new_region_block.next = a.list_of_all_regions
+		new_region_block.end_ptr = end_ptr
+		a.list_of_all_regions = new_region_block
+		start_ptr = start_ptr[size_of(Region):]
+	}
+
+	create_free_region(start_ptr, uint(uintptr(end_sentinel_region)-uintptr(start_ptr)))
+	link_to_free_list(a, (^Region)(start_ptr))
+	return true
+}
+
+@(private="file")
+validate_alloc_size :: proc(size: uint) -> uint {
+	#assert(size_of(uint) >= size_of(uintptr))
+	#assert(size_of(uint)  % size_of(uintptr) == 0)
+
+	// NOTE: emmalloc aligns this forward on pointer size, but I think that is a mistake and will
+	// do bad on wasm64p32.
+
+	validated_size := size > SMALLEST_ALLOCATION_SIZE ? align_forward(size, size_of(uint)) : SMALLEST_ALLOCATION_SIZE
+	assert(validated_size >= size) // Assert we haven't wrapped.
+
+	return validated_size
+}
+
+@(private="file")
+allocate_memory :: proc(a: ^WASM_Allocator, alignment: uint, size: uint, loc := #caller_location) -> rawptr {
+
+	attempt_allocate :: proc(a: ^WASM_Allocator, free_region: ^Region, alignment, size: uint) -> rawptr {
+		assert_locked(a)
+		free_region := free_region
+
+		payload_start_ptr := uintptr(region_payload_start_ptr(free_region))
+		payload_start_ptr_aligned := align_forward(payload_start_ptr, uintptr(alignment))
+		payload_end_ptr := uintptr(region_payload_end_ptr(free_region))
+
+		if payload_start_ptr_aligned + uintptr(size) > payload_end_ptr {
+			return nil
+		}
+
+		// We have enough free space, so the memory allocation will be made into this region. Remove this free region
+		// from the list of free regions: whatever slop remains will be later added back to the free region pool.
+		unlink_from_free_list(free_region)
+
+		// Before we proceed further, fix up the boundary between this and the preceding region,
+		// so that the boundary between the two regions happens at a right spot for the payload to be aligned.
+		if payload_start_ptr != payload_start_ptr_aligned {
+			prev := prev_region(free_region)
+			assert(region_is_in_use(prev))
+			region_boundary_bump_amount := payload_start_ptr_aligned - payload_start_ptr
+			new_this_region_size := free_region.size - uint(region_boundary_bump_amount)
+			create_used_region(prev, prev.size + uint(region_boundary_bump_amount))
+			free_region = (^Region)(uintptr(free_region) + region_boundary_bump_amount)
+			free_region.size = new_this_region_size
+		}
+
+		// Next, we need to decide whether this region is so large that it should be split into two regions,
+		// one representing the newly used memory area, and at the high end a remaining leftover free area.
+		// This splitting to two is done always if there is enough space for the high end to fit a region.
+		// Carve 'size' bytes of payload off this region. So,
+		// [sz prev next sz]
+		// becomes
+		// [sz payload sz] [sz prev next sz]
+		if size_of(Region) + REGION_HEADER_SIZE + size <= free_region.size {
+			new_free_region := (^Region)(uintptr(free_region) + REGION_HEADER_SIZE + uintptr(size))
+			create_free_region(new_free_region, free_region.size - size - REGION_HEADER_SIZE)
+			link_to_free_list(a, new_free_region)
+			create_used_region(free_region, size + REGION_HEADER_SIZE)
+		} else {
+			// There is not enough space to split the free memory region into used+free parts, so consume the whole
+			// region as used memory, not leaving a free memory region behind.
+			// Initialize the free region as used by resetting the ceiling size to the same value as the size at bottom.
+			([^]uint)(uintptr(free_region) + uintptr(free_region.size))[-1] = free_region.size
+		}
+
+		return rawptr(uintptr(free_region) + size_of(uint))
+	}
+
+	assert_locked(a)
+	assert(is_power_of_two(alignment))
+	assert(size <= MAX_ALLOC_SIZE, "allocation too big", loc=loc)
+
+	alignment := alignment
+	alignment  = max(alignment, a.alignment)
+
+	size := size
+	size  = validate_alloc_size(size)
+
+	// Attempt to allocate memory starting from smallest bucket that can contain the required amount of memory.
+	// Under normal alignment conditions this should always be the first or second bucket we look at, but if
+	// performing an allocation with complex alignment, we may need to look at multiple buckets.
+	bucket_index := compute_free_list_bucket(size)
+	bucket_mask := a.free_region_buckets_used >> bucket_index
+
+	// Loop through each bucket that has free regions in it, based on bits set in free_region_buckets_used bitmap.
+	for bucket_mask != 0 {
+		index_add := intrinsics.count_trailing_zeros(bucket_mask)
+		bucket_index += uint(index_add)
+		bucket_mask >>= index_add
+		assert(bucket_index <= NUM_FREE_BUCKETS-1)
+		assert(a.free_region_buckets_used & (BUCKET_BITMASK_T(1) << bucket_index) > 0)
+
+		free_region := a.free_region_buckets[bucket_index].next
+		assert(free_region != nil)
+		if free_region != &a.free_region_buckets[bucket_index] {
+			ptr := attempt_allocate(a, free_region, alignment, size)
+			if ptr != nil {
+				return ptr
+			}
+
+			// We were not able to allocate from the first region found in this bucket, so penalize
+			// the region by cycling it to the end of the doubly circular linked list. (constant time)
+			// This provides a randomized guarantee that when performing allocations of size k to a
+			// bucket of [k-something, k+something] range, we will not always attempt to satisfy the
+			// allocation from the same available region at the front of the list, but we try each
+			// region in turn.
+			unlink_from_free_list(free_region)
+			prepend_to_free_list(free_region, &a.free_region_buckets[bucket_index])
+			// But do not stick around to attempt to look at other regions in this bucket - move
+			// to search the next populated bucket index if this did not fit. This gives a practical
+			// "allocation in constant time" guarantee, since the next higher bucket will only have
+			// regions that are all of strictly larger size than the requested allocation. Only if
+			// there is a difficult alignment requirement we may fail to perform the allocation from
+			// a region in the next bucket, and if so, we keep trying higher buckets until one of them
+			// works.
+			bucket_index += 1
+			bucket_mask >>= 1
+		} else {
+			// This bucket was not populated after all with any regions,
+			// but we just had a stale bit set to mark a populated bucket.
+			// Reset the bit to update latest status so that we do not
+			// redundantly look at this bucket again.
+			a.free_region_buckets_used &~= BUCKET_BITMASK_T(1) << bucket_index
+			bucket_mask ~= 1
+		}
+
+		assert((bucket_index == NUM_FREE_BUCKETS && bucket_mask == 0) || (bucket_mask == a.free_region_buckets_used >> bucket_index))
+	}
+
+	// None of the buckets were able to accommodate an allocation. If this happens we are almost out of memory.
+	// The largest bucket might contain some suitable regions, but we only looked at one region in that bucket, so
+	// as a last resort, loop through more free regions in the bucket that represents the largest allocations available.
+	// But only if the bucket representing largest allocations available is not any of the first thirty buckets,
+	// these represent allocatable areas less than <1024 bytes - which could be a lot of scrap.
+	// In such case, prefer to claim more memory right away.
+	largest_bucket_index := NUM_FREE_BUCKETS - 1 - intrinsics.count_leading_zeros(a.free_region_buckets_used)
+	// free_region will be null if there is absolutely no memory left. (all buckets are 100% used)
+	free_region := a.free_region_buckets_used > 0 ? a.free_region_buckets[largest_bucket_index].next : nil
+	// The 30 first free region buckets cover memory blocks < 2048 bytes, so skip looking at those here (too small)
+	if a.free_region_buckets_used >> 30 > 0 {
+		// Look only at a constant number of regions in this bucket max, to avoid bad worst case behavior.
+		// If this many regions cannot find free space, we give up and prefer to claim more memory instead.
+		max_regions_to_try_before_giving_up :: 99
+		num_tries_left := max_regions_to_try_before_giving_up
+		for ; free_region != &a.free_region_buckets[largest_bucket_index] && num_tries_left > 0; num_tries_left -= 1 {
+			ptr := attempt_allocate(a, free_region, alignment, size)
+			if ptr != nil {
+				return ptr
+			}
+			free_region = free_region.next
+		}
+	}
+
+	// We were unable to find a free memory region. Must claim more memory!
+	num_bytes_to_claim := size+size_of(Region)*3
+	if alignment > a.alignment {
+		num_bytes_to_claim += alignment
+	}
+	success := claim_more_memory(a, num_bytes_to_claim)
+	if (success) {
+		// Try allocate again with the newly available memory.
+		return allocate_memory(a, alignment, size)
+	}
+
+	// also claim_more_memory failed, we are really really constrained :( As a last resort, go back to looking at the
+	// bucket we already looked at above, continuing where the above search left off - perhaps there are
+	// regions we overlooked the first time that might be able to satisfy the allocation.
+	if free_region != nil {
+		for free_region != &a.free_region_buckets[largest_bucket_index] {
+			ptr := attempt_allocate(a, free_region, alignment, size)
+			if ptr != nil {
+				return ptr
+			}
+			free_region = free_region.next
+		}
+	}
+
+	// Fully out of memory.
+	return nil
+}
+
+@(private="file")
+aligned_alloc :: proc(a: ^WASM_Allocator, alignment, size: uint, loc := #caller_location) -> rawptr {
+	lock(a)
+	defer unlock(a)
+
+	return allocate_memory(a, alignment, size, loc)
+}
+
+@(private="file")
+free :: proc(a: ^WASM_Allocator, ptr: rawptr, loc := #caller_location) {
+	if ptr == nil {
+		return
+	}
+
+	region_start_ptr := uintptr(ptr) - size_of(uint)
+	region := (^Region)(region_start_ptr)
+	assert(has_alignment(region_start_ptr, size_of(uint)))
+
+	lock(a)
+	defer unlock(a)
+
+	size := region.size
+	assert(region_is_in_use(region), "double free or corrupt region", loc=loc)
+
+	prev_region_size_field := ([^]uint)(region)[-1]
+	prev_region_size := prev_region_size_field & ~uint(FREE_REGION_FLAG)
+	if prev_region_size_field != prev_region_size {
+		prev_region := (^Region)(uintptr(region) - uintptr(prev_region_size))
+		unlink_from_free_list(prev_region)
+		region_start_ptr = uintptr(prev_region)
+		size += prev_region_size
+	}
+
+	next_reg := next_region(region)
+	size_at_end := (^uint)(region_payload_end_ptr(next_reg))^
+	if next_reg.size != size_at_end {
+		unlink_from_free_list(next_reg)
+		size += next_reg.size
+	}
+
+	create_free_region(rawptr(region_start_ptr), size)
+	link_to_free_list(a, (^Region)(region_start_ptr))
+}
+
+@(private="file")
+aligned_realloc :: proc(a: ^WASM_Allocator, ptr: rawptr, alignment, size: uint, loc := #caller_location) -> rawptr {
+
+	attempt_region_resize :: proc(a: ^WASM_Allocator, region: ^Region, size: uint) -> bool {
+		lock(a)
+		defer unlock(a)
+
+		// First attempt to resize this region, if the next region that follows this one
+		// is a free region.
+		next_reg := next_region(region)
+		next_region_end_ptr := uintptr(next_reg) + uintptr(next_reg.size)
+		size_at_ceiling := ([^]uint)(next_region_end_ptr)[-1]
+		if next_reg.size != size_at_ceiling { // Next region is free?
+			assert(region_is_free(next_reg))
+			new_next_region_start_ptr := uintptr(region) + uintptr(size)
+			assert(has_alignment(new_next_region_start_ptr, size_of(uint)))
+			// Next region does not shrink to too small size?
+			if new_next_region_start_ptr + size_of(Region) <= next_region_end_ptr {
+				unlink_from_free_list(next_reg)
+				create_free_region(rawptr(new_next_region_start_ptr), uint(next_region_end_ptr - new_next_region_start_ptr))
+				link_to_free_list(a, (^Region)(new_next_region_start_ptr))
+				create_used_region(region, uint(new_next_region_start_ptr - uintptr(region)))
+				return true
+			}
+			// If we remove the next region altogether, allocation is satisfied?
+			if new_next_region_start_ptr <= next_region_end_ptr {
+				unlink_from_free_list(next_reg)
+				create_used_region(region, region.size + next_reg.size)
+				return true
+			}
+		} else {
+			// Next region is an used region - we cannot change its starting address. However if we are shrinking the
+			// size of this region, we can create a new free region between this and the next used region.
+			if size + size_of(Region) <= region.size {
+				free_region_size := region.size - size
+				create_used_region(region, size)
+				free_region := (^Region)(uintptr(region) + uintptr(size))
+				create_free_region(free_region, free_region_size)
+				link_to_free_list(a, free_region)
+				return true
+			} else if size <= region.size {
+				// Caller was asking to shrink the size, but due to not being able to fit a full Region in the shrunk
+				// area, we cannot actually do anything. This occurs if the shrink amount is really small. In such case,
+				// just call it success without doing any work.
+				return true
+			}
+		}
+
+		return false
+	}
+
+	if ptr == nil {
+		return aligned_alloc(a, alignment, size, loc)
+	}
+
+	if size == 0 {
+		free(a, ptr, loc)
+		return nil
+	}
+
+	if size > MAX_ALLOC_SIZE {
+		return nil
+	}
+
+	assert(is_power_of_two(alignment))
+	assert(has_alignment(uintptr(ptr), alignment), "realloc on different alignment than original allocation", loc=loc)
+
+	size := size
+	size  = validate_alloc_size(size)
+
+	region := (^Region)(uintptr(ptr) - size_of(uint))
+
+	// Attempt an in-place resize.
+	if attempt_region_resize(a, region, size + REGION_HEADER_SIZE) {
+		return ptr
+	}
+
+	// Can't do it in-place, allocate new region and copy over.
+	newptr := aligned_alloc(a, alignment, size, loc)
+	if newptr != nil {
+		intrinsics.mem_copy(newptr, ptr, min(size, region.size - REGION_HEADER_SIZE))
+		free(a, ptr, loc=loc)
+	}
+
+	return newptr
+}
@@ -48,6 +48,9 @@ if "%2" == "1" (
 set odin_version_raw="dev-%curr_year%-%curr_month%"

 set compiler_flags= -nologo -Oi -TP -fp:precise -Gm- -MP -FC -EHsc- -GR- -GF
+rem Parse source code as utf-8 even on shift-jis and other codepages
+rem See https://learn.microsoft.com/en-us/cpp/build/reference/utf-8-set-source-and-executable-character-sets-to-utf-8?view=msvc-170
+set compiler_flags= %compiler_flags% /utf-8
 set compiler_defines= -DODIN_VERSION_RAW=\"%odin_version_raw%\"

 if not exist .git\ goto skip_git_hash
@@ -111,7 +114,10 @@ call build_vendor.bat
 if %errorlevel% neq 0 goto end_of_build

 rem If the demo doesn't run for you and your CPU is more than a decade old, try -microarch:native
-if %release_mode% EQU 0 odin run examples/demo -- Hellope World
+if %release_mode% EQU 0 odin run examples/demo -vet -strict-style -- Hellope World
+
+rem Many non-compiler devs seem to run debug build but don't realize.
+if %release_mode% EQU 0 echo: & echo Debug compiler built. Note: run "build.bat release" if you want a faster, release mode compiler.

 del *.obj > NUL 2> NUL

@@ -2,7 +2,6 @@
 set -eu

 : ${CPPFLAGS=}
-: ${CXX=clang++}
 : ${CXXFLAGS=}
 : ${LDFLAGS=}
 : ${LLVM_CONFIG=}
@@ -24,19 +23,29 @@ error() {
 	exit 1
 }

+# Brew advises people not to add llvm to their $PATH, so try and use brew to find it.
+if [ -z "$LLVM_CONFIG" ] &&  [ -n "$(command -v brew)" ]; then
+    if   [ -n "$(command -v $(brew --prefix llvm@18)/bin/llvm-config)" ]; then LLVM_CONFIG="$(brew --prefix llvm@18)/bin/llvm-config"
+    elif [ -n "$(command -v $(brew --prefix llvm@17)/bin/llvm-config)" ]; then LLVM_CONFIG="$(brew --prefix llvm@17)/bin/llvm-config"
+    elif [ -n "$(command -v $(brew --prefix llvm@14)/bin/llvm-config)" ]; then LLVM_CONFIG="$(brew --prefix llvm@14)/bin/llvm-config"
+    fi
+fi
+
 if [ -z "$LLVM_CONFIG" ]; then
 	# darwin, linux, openbsd
-	if   [ -n "$(command -v llvm-config-17)" ]; then LLVM_CONFIG="llvm-config-17"
+	if   [ -n "$(command -v llvm-config-18)" ]; then LLVM_CONFIG="llvm-config-18"
+	elif [ -n "$(command -v llvm-config-17)" ]; then LLVM_CONFIG="llvm-config-17"
 	elif [ -n "$(command -v llvm-config-14)" ]; then LLVM_CONFIG="llvm-config-14"
 	elif [ -n "$(command -v llvm-config-13)" ]; then LLVM_CONFIG="llvm-config-13"
 	elif [ -n "$(command -v llvm-config-12)" ]; then LLVM_CONFIG="llvm-config-12"
 	elif [ -n "$(command -v llvm-config-11)" ]; then LLVM_CONFIG="llvm-config-11"
 	# freebsd
-	elif [ -n "$(command -v llvm-config17)" ]; then  LLVM_CONFIG="llvm-config-17"
-	elif [ -n "$(command -v llvm-config14)" ]; then  LLVM_CONFIG="llvm-config-14"
-	elif [ -n "$(command -v llvm-config13)" ]; then  LLVM_CONFIG="llvm-config-13"
-	elif [ -n "$(command -v llvm-config12)" ]; then  LLVM_CONFIG="llvm-config-12"
-	elif [ -n "$(command -v llvm-config11)" ]; then  LLVM_CONFIG="llvm-config-11"
+	elif [ -n "$(command -v llvm-config18)" ]; then  LLVM_CONFIG="llvm-config18"
+	elif [ -n "$(command -v llvm-config17)" ]; then  LLVM_CONFIG="llvm-config17"
+	elif [ -n "$(command -v llvm-config14)" ]; then  LLVM_CONFIG="llvm-config14"
+	elif [ -n "$(command -v llvm-config13)" ]; then  LLVM_CONFIG="llvm-config13"
+	elif [ -n "$(command -v llvm-config12)" ]; then  LLVM_CONFIG="llvm-config12"
+	elif [ -n "$(command -v llvm-config11)" ]; then  LLVM_CONFIG="llvm-config11"
 	# fallback
 	elif [ -n "$(command -v llvm-config)" ]; then LLVM_CONFIG="llvm-config"
 	else
@@ -44,37 +53,57 @@ if [ -z "$LLVM_CONFIG" ]; then
 	fi
 fi

+if [ -x "$(which clang++)" ]; then
+	: ${CXX="clang++"}
+elif [ -x "$($LLVM_CONFIG --bindir)/clang++" ]; then
+	: ${CXX=$($LLVM_CONFIG --bindir)/clang++}
+else
+	error "No clang++ command found. Set CXX to proceed."
+fi
+
 LLVM_VERSION="$($LLVM_CONFIG --version)"
 LLVM_VERSION_MAJOR="$(echo $LLVM_VERSION | awk -F. '{print $1}')"
 LLVM_VERSION_MINOR="$(echo $LLVM_VERSION | awk -F. '{print $2}')"
 LLVM_VERSION_PATCH="$(echo $LLVM_VERSION | awk -F. '{print $3}')"

-if [ $LLVM_VERSION_MAJOR -lt 11 ] ||
-	([ $LLVM_VERSION_MAJOR -gt 14 ] && [ $LLVM_VERSION_MAJOR -lt 17 ]); then
-	error "Invalid LLVM version $LLVM_VERSION: must be 11, 12, 13, 14 or 17"
+if [ $LLVM_VERSION_MAJOR -lt 11 ] || ([ $LLVM_VERSION_MAJOR -gt 14 ] && [ $LLVM_VERSION_MAJOR -lt 17 ]) || [ $LLVM_VERSION_MAJOR -gt 18 ]; then
+	error "Invalid LLVM version $LLVM_VERSION: must be 11, 12, 13, 14, 17 or 18"
 fi

 case "$OS_NAME" in
 Darwin)
 	if [ "$OS_ARCH" = "arm64" ]; then
-		if [ $LLVM_VERSION_MAJOR -lt 13 ] || [ $LLVM_VERSION_MAJOR -gt 17 ]; then
-			error "Darwin Arm64 requires LLVM 13, 14 or 17"
+		if [ $LLVM_VERSION_MAJOR -lt 13 ]; then
+			error "Invalid LLVM version $LLVM_VERSION: Darwin Arm64 requires LLVM 13, 14, 17 or 18"
 		fi
 	fi

-	CXXFLAGS="$CXXFLAGS $($LLVM_CONFIG --cxxflags --ldflags)"
+	darwin_sysroot=
+	if [ $(which xcrun) ]; then
+		darwin_sysroot="--sysroot $(xcrun --sdk macosx --show-sdk-path)"
+	elif [[ -e "/Library/Developer/CommandLineTools/SDKs/MacOSX.sdk" ]]; then
+		darwin_sysroot="--sysroot /Library/Developer/CommandLineTools/SDKs/MacOSX.sdk"
+	else
+		echo "Warning: MacOSX.sdk not found."
+	fi
+
+	CXXFLAGS="$CXXFLAGS $($LLVM_CONFIG --cxxflags --ldflags) ${darwin_sysroot}"
 	LDFLAGS="$LDFLAGS -liconv -ldl -framework System -lLLVM"
 	;;
 FreeBSD)
 	CXXFLAGS="$CXXFLAGS $($LLVM_CONFIG --cxxflags --ldflags)"
 	LDFLAGS="$LDFLAGS $($LLVM_CONFIG --libs core native --system-libs)"
 	;;
+NetBSD)
+	CXXFLAGS="$CXXFLAGS $($LLVM_CONFIG --cxxflags --ldflags)"
+	LDFLAGS="$LDFLAGS $($LLVM_CONFIG --libs core native --system-libs)"
+	;;
 Linux)
 	CXXFLAGS="$CXXFLAGS $($LLVM_CONFIG --cxxflags --ldflags)"
 	LDFLAGS="$LDFLAGS -ldl $($LLVM_CONFIG --libs core native --system-libs --libfiles)"
 	# Copy libLLVM*.so into current directory for linking
 	# NOTE: This is needed by the Linux release pipeline!
-	cp $(readlink -f $($LLVM_CONFIG --libfiles)) ./
+	# cp $(readlink -f $($LLVM_CONFIG --libfiles)) ./
 	LDFLAGS="$LDFLAGS -Wl,-rpath=\$ORIGIN"
 	;;
 OpenBSD)
@@ -123,12 +152,17 @@ build_odin() {
 }

 run_demo() {
-	./odin run examples/demo/demo.odin -file -- Hellope World
+	if [ $# -eq 0 ] || [ "$1" = "debug" ]; then
+		./odin run examples/demo -vet -strict-style -- Hellope World
+	fi
 }

 if [ $# -eq 0 ]; then
 	build_odin debug
 	run_demo
+
+	: ${PROGRAM:=$0}
+	printf "\nDebug compiler built. Note: run \"$PROGRAM release\" or \"$PROGRAM release-native\" if you want a faster, release mode compiler.\n"
 elif [ $# -eq 1 ]; then
 	case $1 in
 	report)
@@ -1,51 +0,0 @@
-import subprocess
-import sys
-import json
-import datetime
-import urllib.parse
-import sys
-
-def main():
-    files_by_date = {}
-    bucket = sys.argv[1]
-
-    files_lines = execute_cli(f"b2 ls --long {bucket} nightly").split("\n")
-    for x in files_lines:
-        parts = x.split(" ", 1)
-        if parts[0]:
-            json_str = execute_cli(f"b2 get-file-info {parts[0]}")
-            data = json.loads(json_str)
-            name = remove_prefix(data['fileName'], "nightly/")
-            url = f"https://f001.backblazeb2.com/file/{bucket}/nightly/{urllib.parse.quote_plus(name)}"
-            sha1 = data['contentSha1']
-            size = int(data['size'])
-            ts = int(data['fileInfo']['src_last_modified_millis'])
-            date = datetime.datetime.fromtimestamp(ts/1000).strftime('%Y-%m-%d')
-            
-            if date not in files_by_date.keys():
-                files_by_date[date] = []
-
-            files_by_date[date].append({
-                                            'name': name,
-                                            'url': url,
-                                            'sha1': sha1,
-                                            'sizeInBytes': size,
-                                         })
-
-    now = datetime.datetime.utcnow().isoformat()
-
-    print(json.dumps({
-                        'last_updated' : now,
-                        'files': files_by_date
-                     }, sort_keys=True, indent=4))
-
-def remove_prefix(text, prefix):
-    return text[text.startswith(prefix) and len(prefix):]
-
-def execute_cli(command):
-    sb = subprocess.Popen(command, shell=True, stdout=subprocess.PIPE)
-    return sb.stdout.read().decode("utf-8");
-
-if __name__ == '__main__':
-    sys.exit(main())
-
@@ -1,34 +0,0 @@
-import subprocess
-import sys
-import json
-import datetime
-import urllib.parse
-import sys
-
-def main():
-    files_by_date = {}
-    bucket = sys.argv[1]
-    days_to_keep = int(sys.argv[2])
-    print(f"Looking for binaries to delete older than {days_to_keep} days")
-
-    files_lines = execute_cli(f"b2 ls --long --versions {bucket} nightly").split("\n")
-    for x in files_lines:
-        parts = [y for y in x.split(' ') if y]
-
-        if parts and parts[0]:
-            date = datetime.datetime.strptime(parts[2], '%Y-%m-%d').replace(hour=0, minute=0, second=0, microsecond=0)
-            now = datetime.datetime.utcnow().replace(hour=0, minute=0, second=0, microsecond=0)
-            delta = now - date
-
-            if delta.days > days_to_keep:
-                print(f'Deleting {parts[5]}')
-                execute_cli(f'b2 delete-file-version {parts[0]}')
-
-
-def execute_cli(command):
-    sb = subprocess.Popen(command, shell=True, stdout=subprocess.PIPE)
-    return sb.stdout.read().decode("utf-8");
-
-if __name__ == '__main__':
-    sys.exit(main())
-
@@ -0,0 +1,140 @@
+import os
+import sys
+from zipfile  import ZipFile, ZIP_DEFLATED
+from b2sdk.v2 import InMemoryAccountInfo, B2Api
+from datetime import datetime
+import json
+
+UPLOAD_FOLDER = "nightly/"
+
+info   = InMemoryAccountInfo()
+b2_api = B2Api(info)
+application_key_id = os.environ['APPID']
+application_key    = os.environ['APPKEY']
+bucket_name        = os.environ['BUCKET']
+days_to_keep       = os.environ['DAYS_TO_KEEP']
+
+def auth() -> bool:
+	try:
+		realm = b2_api.account_info.get_realm()
+		return True # Already authenticated
+	except:
+		pass        # Not yet authenticated
+
+	err = b2_api.authorize_account("production", application_key_id, application_key)
+	return err == None
+
+def get_bucket():
+	if not auth(): sys.exit(1)
+	return b2_api.get_bucket_by_name(bucket_name)
+
+def remove_prefix(text: str, prefix: str) -> str:
+	return text[text.startswith(prefix) and len(prefix):]
+
+def create_and_upload_artifact_zip(platform: str, artifact: str) -> int:
+	now = datetime.utcnow().replace(hour=0, minute=0, second=0, microsecond=0)
+	destination_zip_name = "odin-{}-nightly+{}.zip".format(platform, now.strftime("%Y-%m-%d"))
+
+	source_zip_name = artifact
+	if not artifact.endswith(".zip"):
+		print(f"Creating archive {destination_zip_name} from {artifact} and uploading to {bucket_name}")
+
+		source_zip_name = destination_zip_name
+		with ZipFile(source_zip_name, mode='w', compression=ZIP_DEFLATED, compresslevel=9) as z:
+			for root, directory, filenames in os.walk(artifact):
+				for file in filenames:
+					file_path = os.path.join(root, file)
+					zip_path  = os.path.join("dist", os.path.relpath(file_path, artifact))
+					z.write(file_path, zip_path)
+
+		if not os.path.exists(source_zip_name):
+			print(f"Error: Newly created ZIP archive {source_zip_name} not found.")
+			return 1
+
+	print("Uploading {} to {}".format(source_zip_name, UPLOAD_FOLDER + destination_zip_name))
+	bucket = get_bucket()
+	res = bucket.upload_local_file(
+		source_zip_name,                   # Local file to upload
+		"nightly/" + destination_zip_name, # B2 destination path
+	)
+	return 0
+
+def prune_artifacts():
+	print(f"Looking for binaries to delete older than {days_to_keep} days")
+
+	bucket = get_bucket()
+	for file, _ in bucket.ls(UPLOAD_FOLDER, latest_only=False):
+		# Timestamp is in milliseconds
+		date  = datetime.fromtimestamp(file.upload_timestamp / 1_000.0).replace(hour=0, minute=0, second=0, microsecond=0)
+		now   = datetime.utcnow().replace(hour=0, minute=0, second=0, microsecond=0)
+		delta = now - date
+
+		if delta.days > int(days_to_keep):
+			print("Deleting {}".format(file.file_name))
+			file.delete()
+
+	return 0
+
+def update_nightly_json():
+	print(f"Updating nightly.json with files {days_to_keep} days or newer")
+
+	files_by_date = {}
+
+	bucket = get_bucket()
+
+	for file, _ in bucket.ls(UPLOAD_FOLDER, latest_only=True):
+		# Timestamp is in milliseconds
+		date = datetime.fromtimestamp(file.upload_timestamp / 1_000.0).replace(hour=0, minute=0, second=0, microsecond=0).strftime('%Y-%m-%d')
+		name = remove_prefix(file.file_name, UPLOAD_FOLDER)
+		sha1 = file.content_sha1
+		size = file.size
+		url  = bucket.get_download_url(file.file_name)
+
+		if date not in files_by_date.keys():
+			files_by_date[date] = []
+
+		files_by_date[date].append({
+			'name':        name,
+			'url':         url,
+			'sha1':        sha1,
+			'sizeInBytes': size,
+		})
+
+	now = datetime.utcnow().isoformat()
+
+	nightly = json.dumps({
+		'last_updated' : now,
+		'files': files_by_date
+	}, sort_keys=True, indent=4, ensure_ascii=False).encode('utf-8')
+
+	res = bucket.upload_bytes(
+		nightly,        # JSON bytes
+		"nightly.json", # B2 destination path
+	)
+	return 0
+
+if __name__ == "__main__":
+	if len(sys.argv) == 1:
+		print("Usage: {} <verb> [arguments]".format(sys.argv[0]))
+		print("\tartifact <platform prefix> <artifact path>\n\t\tCreates and uploads a platform artifact zip.")
+		print("\tprune\n\t\tDeletes old artifacts from bucket")
+		print("\tjson\n\t\tUpdate and upload nightly.json")
+		sys.exit(1)
+	else:
+		command = sys.argv[1].lower()
+		if command == "artifact":
+			if len(sys.argv) != 4:
+				print("Usage: {} artifact <platform prefix> <artifact path>".format(sys.argv[0]))
+				print("Error: Expected artifact command to be given platform prefix and artifact path.\n")
+				sys.exit(1)
+
+			res = create_and_upload_artifact_zip(sys.argv[2], sys.argv[3])
+			sys.exit(res)
+
+		elif command == "prune":
+			res = prune_artifacts()
+			sys.exit(res)
+
+		elif command == "json":
+			res = update_nightly_json()
+			sys.exit(res)
@@ -1,25 +0,0 @@
-#!/bin/bash
-
-set -e
-
-bucket=$1
-platform=$2
-artifact=$3
-
-now=$(date +'%Y-%m-%d')
-filename="odin-$platform-nightly+$now.zip"
-
-echo "Creating archive $filename from $artifact and uploading to $bucket"
-
-# If this is already zipped up (done before artifact upload to keep permissions in tact), just move it.
-if [ "${artifact: -4}" == ".zip" ]
-then
-	echo "Artifact already a zip"
-	mkdir -p "output"
-	mv "$artifact" "output/$filename"
-else
-	echo "Artifact needs to be zipped"
-	7z a -bd "output/$filename" -r "$artifact"
-fi
-
-b2 upload-file --noProgress "$bucket" "output/$filename" "nightly/$filename"
@@ -29,12 +29,12 @@ MIN_READ_BUFFER_SIZE :: 16
@(private)
 DEFAULT_MAX_CONSECUTIVE_EMPTY_READS :: 128

-reader_init :: proc(b: ^Reader, rd: io.Reader, size: int = DEFAULT_BUF_SIZE, allocator := context.allocator) {
+reader_init :: proc(b: ^Reader, rd: io.Reader, size: int = DEFAULT_BUF_SIZE, allocator := context.allocator, loc := #caller_location) {
 	size := size
 	size = max(size, MIN_READ_BUFFER_SIZE)
 	reader_reset(b, rd)
 	b.buf_allocator = allocator
-	b.buf = make([]byte, size, allocator)
+	b.buf = make([]byte, size, allocator, loc)
 }

 reader_init_with_buf :: proc(b: ^Reader, rd: io.Reader, buf: []byte) {
@@ -81,7 +81,7 @@ _reader_read_new_chunk :: proc(b: ^Reader) -> io.Error {
 	for i := b.max_consecutive_empty_reads; i > 0; i -= 1 {
 		n, err := io.read(b.rd, b.buf[b.w:])
 		if n < 0 {
-			return .Negative_Read
+			return err if err != nil else .Negative_Read
 		}
 		b.w += n
 		if err != nil {
@@ -189,7 +189,7 @@ reader_read :: proc(b: ^Reader, p: []byte) -> (n: int, err: io.Error) {
 		if len(p) >= len(b.buf) {
 			n, b.err = io.read(b.rd, p)
 			if n < 0 {
-				return 0, .Negative_Read
+				return 0, b.err if b.err != nil else .Negative_Read
 			}

 			if n > 0 {
@@ -202,7 +202,7 @@ reader_read :: proc(b: ^Reader, p: []byte) -> (n: int, err: io.Error) {
 		b.r, b.w = 0, 0
 		n, b.err = io.read(b.rd, b.buf)
 		if n < 0 {
-			return 0, .Negative_Read
+			return 0, b.err if b.err != nil else .Negative_Read
 		}
 		if n == 0 {
 			return 0, _reader_consume_err(b)
@@ -290,7 +290,7 @@ reader_write_to :: proc(b: ^Reader, w: io.Writer) -> (n: i64, err: io.Error) {
 	write_buf :: proc(b: ^Reader, w: io.Writer) -> (i64, io.Error) {
 		n, err := io.write(w, b.buf[b.r:b.w])
 		if n < 0 {
-			return 0, .Negative_Write
+			return 0, err if err != nil else .Negative_Write
 		}
 		b.r += n
 		return i64(n), err
@@ -95,6 +95,10 @@ writer_write :: proc(b: ^Writer, p: []byte) -> (n: int, err: io.Error) {
 		m: int
 		if writer_buffered(b) == 0 {
 			m, b.err = io.write(b.wr, p)
+			if m < 0 && b.err == nil {
+				b.err = .Negative_Write
+				break
+			}
 		} else {
 			m = copy(b.buf[b.n:], p)
 			b.n += m
@@ -27,19 +27,19 @@ Read_Op :: enum i8 {
 }


-buffer_init :: proc(b: ^Buffer, buf: []byte) {
-	resize(&b.buf, len(buf))
+buffer_init :: proc(b: ^Buffer, buf: []byte, loc := #caller_location) {
+	resize(&b.buf, len(buf), loc=loc)
 	copy(b.buf[:], buf)
 }

-buffer_init_string :: proc(b: ^Buffer, s: string) {
-	resize(&b.buf, len(s))
+buffer_init_string :: proc(b: ^Buffer, s: string, loc := #caller_location) {
+	resize(&b.buf, len(s), loc=loc)
 	copy(b.buf[:], s)
 }

-buffer_init_allocator :: proc(b: ^Buffer, len, cap: int, allocator := context.allocator) {
+buffer_init_allocator :: proc(b: ^Buffer, len, cap: int, allocator := context.allocator, loc := #caller_location) {
 	if b.buf == nil {
-		b.buf = make([dynamic]byte, len, cap, allocator)
+		b.buf = make([dynamic]byte, len, cap, allocator, loc)
 		return
 	}

@@ -96,28 +96,28 @@ buffer_truncate :: proc(b: ^Buffer, n: int) {
 }

@(private)
-_buffer_try_grow :: proc(b: ^Buffer, n: int) -> (int, bool) {
+_buffer_try_grow :: proc(b: ^Buffer, n: int, loc := #caller_location) -> (int, bool) {
 	if l := len(b.buf); n <= cap(b.buf)-l {
-		resize(&b.buf, l+n)
+		resize(&b.buf, l+n, loc=loc)
 		return l, true
 	}
 	return 0, false
 }

@(private)
-_buffer_grow :: proc(b: ^Buffer, n: int) -> int {
+_buffer_grow :: proc(b: ^Buffer, n: int, loc := #caller_location) -> int {
 	m := buffer_length(b)
 	if m == 0 && b.off != 0 {
 		buffer_reset(b)
 	}
-	if i, ok := _buffer_try_grow(b, n); ok {
+	if i, ok := _buffer_try_grow(b, n, loc=loc); ok {
 		return i
 	}

 	if b.buf == nil && n <= SMALL_BUFFER_SIZE {
 		// Fixes #2756 by preserving allocator if already set on Buffer via init_buffer_allocator
-		reserve(&b.buf, SMALL_BUFFER_SIZE)
-		resize(&b.buf, n)
+		reserve(&b.buf, SMALL_BUFFER_SIZE, loc=loc)
+		resize(&b.buf, n, loc=loc)
 		return 0
 	}

@@ -127,31 +127,34 @@ _buffer_grow :: proc(b: ^Buffer, n: int) -> int {
 	} else if c > max(int) - c - n {
 		panic("bytes.Buffer: too large")
 	} else {
-		resize(&b.buf, 2*c + n)
+		resize(&b.buf, 2*c + n, loc=loc)
 		copy(b.buf[:], b.buf[b.off:])
 	}
 	b.off = 0
-	resize(&b.buf, m+n)
+	resize(&b.buf, m+n, loc=loc)
 	return m
 }

-buffer_grow :: proc(b: ^Buffer, n: int) {
+buffer_grow :: proc(b: ^Buffer, n: int, loc := #caller_location) {
 	if n < 0 {
 		panic("bytes.buffer_grow: negative count")
 	}
-	m := _buffer_grow(b, n)
-	resize(&b.buf, m)
+	m := _buffer_grow(b, n, loc=loc)
+	resize(&b.buf, m, loc=loc)
 }

-buffer_write_at :: proc(b: ^Buffer, p: []byte, offset: int) -> (n: int, err: io.Error) {
+buffer_write_at :: proc(b: ^Buffer, p: []byte, offset: int, loc := #caller_location) -> (n: int, err: io.Error) {
+	if len(p) == 0 {
+		return 0, nil
+	}
 	b.last_read = .Invalid
 	if offset < 0 {
 		err = .Invalid_Offset
 		return
 	}
-	_, ok := _buffer_try_grow(b, offset+len(p))
+	_, ok := _buffer_try_grow(b, offset+len(p), loc=loc)
 	if !ok {
-		_ = _buffer_grow(b, offset+len(p))
+		_ = _buffer_grow(b, offset+len(p), loc=loc)
 	}
 	if len(b.buf) <= offset {
 		return 0, .Short_Write
@@ -160,47 +163,47 @@ buffer_write_at :: proc(b: ^Buffer, p: []byte, offset: int) -> (n: int, err: io.
 }


-buffer_write :: proc(b: ^Buffer, p: []byte) -> (n: int, err: io.Error) {
+buffer_write :: proc(b: ^Buffer, p: []byte, loc := #caller_location) -> (n: int, err: io.Error) {
 	b.last_read = .Invalid
-	m, ok := _buffer_try_grow(b, len(p))
+	m, ok := _buffer_try_grow(b, len(p), loc=loc)
 	if !ok {
-		m = _buffer_grow(b, len(p))
+		m = _buffer_grow(b, len(p), loc=loc)
 	}
 	return copy(b.buf[m:], p), nil
 }

-buffer_write_ptr :: proc(b: ^Buffer, ptr: rawptr, size: int) -> (n: int, err: io.Error) {
-	return buffer_write(b, ([^]byte)(ptr)[:size])
+buffer_write_ptr :: proc(b: ^Buffer, ptr: rawptr, size: int, loc := #caller_location) -> (n: int, err: io.Error) {
+	return buffer_write(b, ([^]byte)(ptr)[:size], loc=loc)
 }

-buffer_write_string :: proc(b: ^Buffer, s: string) -> (n: int, err: io.Error) {
+buffer_write_string :: proc(b: ^Buffer, s: string, loc := #caller_location) -> (n: int, err: io.Error) {
 	b.last_read = .Invalid
-	m, ok := _buffer_try_grow(b, len(s))
+	m, ok := _buffer_try_grow(b, len(s), loc=loc)
 	if !ok {
-		m = _buffer_grow(b, len(s))
+		m = _buffer_grow(b, len(s), loc=loc)
 	}
 	return copy(b.buf[m:], s), nil
 }

-buffer_write_byte :: proc(b: ^Buffer, c: byte) -> io.Error {
+buffer_write_byte :: proc(b: ^Buffer, c: byte, loc := #caller_location) -> io.Error {
 	b.last_read = .Invalid
-	m, ok := _buffer_try_grow(b, 1)
+	m, ok := _buffer_try_grow(b, 1, loc=loc)
 	if !ok {
-		m = _buffer_grow(b, 1)
+		m = _buffer_grow(b, 1, loc=loc)
 	}
 	b.buf[m] = c
 	return nil
 }

-buffer_write_rune :: proc(b: ^Buffer, r: rune) -> (n: int, err: io.Error) {
+buffer_write_rune :: proc(b: ^Buffer, r: rune, loc := #caller_location) -> (n: int, err: io.Error) {
 	if r < utf8.RUNE_SELF {
-		buffer_write_byte(b, byte(r))
+		buffer_write_byte(b, byte(r), loc=loc)
 		return 1, nil
 	}
 	b.last_read = .Invalid
-	m, ok := _buffer_try_grow(b, utf8.UTF_MAX)
+	m, ok := _buffer_try_grow(b, utf8.UTF_MAX, loc=loc)
 	if !ok {
-		m = _buffer_grow(b, utf8.UTF_MAX)
+		m = _buffer_grow(b, utf8.UTF_MAX, loc=loc)
 	}
 	res: [4]byte
 	res, n = utf8.encode_rune(r)
@@ -246,10 +249,13 @@ buffer_read_ptr :: proc(b: ^Buffer, ptr: rawptr, size: int) -> (n: int, err: io.
 }

 buffer_read_at :: proc(b: ^Buffer, p: []byte, offset: int) -> (n: int, err: io.Error) {
+	if len(p) == 0 {
+		return 0, nil
+	}
 	b.last_read = .Invalid

 	if uint(offset) >= len(b.buf) {
-		err = .Invalid_Offset
+		err = .EOF
 		return
 	}
 	n = copy(p, b.buf[offset:])
@@ -310,6 +316,27 @@ buffer_unread_rune :: proc(b: ^Buffer) -> io.Error {
 	return nil
 }

+buffer_seek :: proc(b: ^Buffer, offset: i64, whence: io.Seek_From) -> (i64, io.Error) {
+	abs: i64
+	switch whence {
+	case .Start:
+		abs = offset
+	case .Current:
+		abs = i64(b.off) + offset
+	case .End:
+		abs = i64(len(b.buf)) + offset
+	case:
+		return 0, .Invalid_Whence
+	}
+
+	abs_int := int(abs)
+	if abs_int < 0 {
+		return 0, .Invalid_Offset
+	}
+	b.last_read = .Invalid
+	b.off = abs_int
+	return abs, nil
+}

 buffer_read_bytes :: proc(b: ^Buffer, delim: byte) -> (line: []byte, err: io.Error) {
 	i := index_byte(b.buf[b.off:], delim)
@@ -359,7 +386,7 @@ buffer_read_from :: proc(b: ^Buffer, r: io.Reader) -> (n: i64, err: io.Error) #n
 		resize(&b.buf, i)
 		m, e := io.read(r, b.buf[i:cap(b.buf)])
 		if m < 0 {
-			err = .Negative_Read
+			err = e if e != nil else .Negative_Read
 			return
 		}

@@ -395,14 +422,17 @@ _buffer_proc :: proc(stream_data: rawptr, mode: io.Stream_Mode, p: []byte, offse
 		return io._i64_err(buffer_write(b, p))
 	case .Write_At:
 		return io._i64_err(buffer_write_at(b, p, int(offset)))
+	case .Seek:
+		n, err = buffer_seek(b, offset, whence)
+		return
 	case .Size:
-		n = i64(buffer_capacity(b))
+		n = i64(buffer_length(b))
 		return
 	case .Destroy:
 		buffer_destroy(b)
 		return
 	case .Query:
-		return io.query_utility({.Read, .Read_At, .Write, .Write_At, .Size, .Destroy})
+		return io.query_utility({.Read, .Read_At, .Write, .Write_At, .Seek, .Size, .Destroy, .Query})
 	}
 	return 0, .Empty
 }
@@ -1,9 +1,38 @@
 package bytes

+import "base:intrinsics"
 import "core:mem"
+import "core:simd"
 import "core:unicode"
 import "core:unicode/utf8"

+when ODIN_ARCH == .amd64 && intrinsics.has_target_feature("avx2") {
+	@(private)
+	SCANNER_INDICES_256 : simd.u8x32 : {
+		0,  1,  2,  3,  4,  5,  6,  7,
+		8,  9, 10, 11, 12, 13, 14, 15,
+		16, 17, 18, 19, 20, 21, 22, 23,
+		24, 25, 26, 27, 28, 29, 30, 31,
+	}
+	@(private)
+	SCANNER_SENTINEL_MAX_256: simd.u8x32 : u8(0x00)
+	@(private)
+	SCANNER_SENTINEL_MIN_256: simd.u8x32 : u8(0xff)
+	@(private)
+	SIMD_REG_SIZE_256 :: 32
+}
+@(private)
+SCANNER_INDICES_128 : simd.u8x16 : {
+	0,  1,  2,  3,  4,  5,  6,  7,
+	8,  9, 10, 11, 12, 13, 14, 15,
+}
+@(private)
+SCANNER_SENTINEL_MAX_128: simd.u8x16 : u8(0x00)
+@(private)
+SCANNER_SENTINEL_MIN_128: simd.u8x16 : u8(0xff)
+@(private)
+SIMD_REG_SIZE_128 :: 16
+
 clone :: proc(s: []byte, allocator := context.allocator, loc := #caller_location) -> []byte {
 	c := make([]byte, len(s), allocator, loc)
 	copy(c, s)
@@ -293,28 +322,279 @@ split_after_iterator :: proc(s: ^[]byte, sep: []byte) -> ([]byte, bool) {
 	return _split_iterator(s, sep, len(sep))
 }

+/*
+Scan a slice of bytes for a specific byte.

-index_byte :: proc(s: []byte, c: byte) -> int {
-	for i := 0; i < len(s); i += 1 {
+This procedure safely handles slices of any length, including empty slices.
+
+Inputs:
+- data: A slice of bytes.
+- c: The byte to search for.
+
+Returns:
+- index: The index of the byte `c`, or -1 if it was not found.
+*/
+index_byte :: proc(s: []byte, c: byte) -> (index: int) #no_bounds_check {
+	i, l := 0, len(s)
+
+	// Guard against small strings.  On modern systems, it is ALWAYS
+	// worth vectorizing assuming there is a hardware vector unit, and
+	// the data size is large enough.
+	if l < SIMD_REG_SIZE_128 {
+		for /**/; i < l; i += 1 {
+			if s[i] == c {
+				return i
+			}
+		}
+		return -1
+	}
+
+	c_vec: simd.u8x16 = c
+	when !simd.IS_EMULATED {
+		// Note: While this is something that could also logically take
+		// advantage of AVX512, the various downclocking and power
+		// consumption related woes make premature to have a dedicated
+		// code path.
+		when ODIN_ARCH == .amd64 && intrinsics.has_target_feature("avx2") {
+			c_vec_256: simd.u8x32 = c
+
+			s_vecs: [4]simd.u8x32 = ---
+			c_vecs: [4]simd.u8x32 = ---
+			m_vec: [4]u8 = ---
+
+			// Scan 128-byte chunks, using 256-bit SIMD.
+			for nr_blocks := l / (4 * SIMD_REG_SIZE_256); nr_blocks > 0; nr_blocks -= 1 {
+				#unroll for j in 0..<4 {
+					s_vecs[j] = intrinsics.unaligned_load(cast(^simd.u8x32)raw_data(s[i+j*SIMD_REG_SIZE_256:]))
+					c_vecs[j] = simd.lanes_eq(s_vecs[j], c_vec_256)
+					m_vec[j] = simd.reduce_or(c_vecs[j])
+				}
+				if m_vec[0] | m_vec[1] | m_vec[2] | m_vec[3] > 0 {
+					#unroll for j in 0..<4 {
+						if m_vec[j] > 0 {
+							sel := simd.select(c_vecs[j], SCANNER_INDICES_256, SCANNER_SENTINEL_MIN_256)
+							off := simd.reduce_min(sel)
+							return i + j * SIMD_REG_SIZE_256 + int(off)
+						}
+					}
+				}
+
+				i += 4 * SIMD_REG_SIZE_256
+			}
+
+			// Scan 64-byte chunks, using 256-bit SIMD.
+			for nr_blocks := (l - i) / (2 * SIMD_REG_SIZE_256); nr_blocks > 0; nr_blocks -= 1 {
+				#unroll for j in 0..<2 {
+					s_vecs[j] = intrinsics.unaligned_load(cast(^simd.u8x32)raw_data(s[i+j*SIMD_REG_SIZE_256:]))
+					c_vecs[j] = simd.lanes_eq(s_vecs[j], c_vec_256)
+					m_vec[j] = simd.reduce_or(c_vecs[j])
+				}
+				if m_vec[0] | m_vec[1] > 0 {
+					#unroll for j in 0..<2 {
+						if m_vec[j] > 0 {
+							sel := simd.select(c_vecs[j], SCANNER_INDICES_256, SCANNER_SENTINEL_MIN_256)
+							off := simd.reduce_min(sel)
+							return i + j * SIMD_REG_SIZE_256 + int(off)
+						}
+					}
+				}
+
+				i += 2 * SIMD_REG_SIZE_256
+			}
+		} else {
+			s_vecs: [4]simd.u8x16 = ---
+			c_vecs: [4]simd.u8x16 = ---
+			m_vecs: [4]u8 = ---
+
+			// Scan 64-byte chunks, using 128-bit SIMD.
+			for nr_blocks := l / (4 * SIMD_REG_SIZE_128); nr_blocks > 0; nr_blocks -= 1 {
+				#unroll for j in 0..<4 {
+					s_vecs[j]= intrinsics.unaligned_load(cast(^simd.u8x16)raw_data(s[i+j*SIMD_REG_SIZE_128:]))
+					c_vecs[j] = simd.lanes_eq(s_vecs[j], c_vec)
+					m_vecs[j] = simd.reduce_or(c_vecs[j])
+				}
+				if m_vecs[0] | m_vecs[1] | m_vecs[2] | m_vecs[3] > 0 {
+					#unroll for j in 0..<4 {
+						if m_vecs[j] > 0 {
+							sel := simd.select(c_vecs[j], SCANNER_INDICES_128, SCANNER_SENTINEL_MIN_128)
+							off := simd.reduce_min(sel)
+							return i + j * SIMD_REG_SIZE_128 + int(off)
+						}
+					}
+				}
+
+				i += 4 * SIMD_REG_SIZE_128
+			}
+		}
+	}
+
+	// Scan the remaining SIMD register sized chunks.
+	//
+	// Apparently LLVM does ok with 128-bit SWAR, so this path is also taken
+	// on potato targets.  Scanning more at a time when LLVM is emulating SIMD
+	// likely does not buy much, as all that does is increase GP register
+	// pressure.
+	for nr_blocks := (l - i) / SIMD_REG_SIZE_128; nr_blocks > 0; nr_blocks -= 1 {
+		s0 := intrinsics.unaligned_load(cast(^simd.u8x16)raw_data(s[i:]))
+		c0 := simd.lanes_eq(s0, c_vec)
+		if simd.reduce_or(c0) > 0 {
+			sel := simd.select(c0, SCANNER_INDICES_128, SCANNER_SENTINEL_MIN_128)
+			off := simd.reduce_min(sel)
+			return i + int(off)
+		}
+
+		i += SIMD_REG_SIZE_128
+	}
+
+	// Scan serially for the remainder.
+	for /**/; i < l; i += 1 {
 		if s[i] == c {
 			return i
 		}
 	}
+
 	return -1
 }

-// Returns -1 if c is not present
-last_index_byte :: proc(s: []byte, c: byte) -> int {
-	for i := len(s)-1; i >= 0; i -= 1 {
+/*
+Scan a slice of bytes for a specific byte, starting from the end and working
+backwards to the start.
+
+This procedure safely handles slices of any length, including empty slices.
+
+Inputs:
+- data: A slice of bytes.
+- c: The byte to search for.
+
+Returns:
+- index: The index of the byte `c`, or -1 if it was not found.
+*/
+last_index_byte :: proc(s: []byte, c: byte) -> int #no_bounds_check {
+	i := len(s)
+
+	// Guard against small strings.  On modern systems, it is ALWAYS
+	// worth vectorizing assuming there is a hardware vector unit, and
+	// the data size is large enough.
+	if i < SIMD_REG_SIZE_128 {
+		if i > 0 { // Handle s == nil.
+			for /**/; i >= 0; i -= 1 {
+				if s[i] == c {
+					return i
+				}
+			}
+		}
+		return -1
+	}
+
+	c_vec: simd.u8x16 = c
+	when !simd.IS_EMULATED {
+		// Note: While this is something that could also logically take
+		// advantage of AVX512, the various downclocking and power
+		// consumption related woes make premature to have a dedicated
+		// code path.
+		when ODIN_ARCH == .amd64 && intrinsics.has_target_feature("avx2") {
+			c_vec_256: simd.u8x32 = c
+
+			s_vecs: [4]simd.u8x32 = ---
+			c_vecs: [4]simd.u8x32 = ---
+			m_vec: [4]u8 = ---
+
+			// Scan 128-byte chunks, using 256-bit SIMD.
+			for i >= 4 * SIMD_REG_SIZE_256 {
+				i -= 4 * SIMD_REG_SIZE_256
+
+				#unroll for j in 0..<4 {
+					s_vecs[j] = intrinsics.unaligned_load(cast(^simd.u8x32)raw_data(s[i+j*SIMD_REG_SIZE_256:]))
+					c_vecs[j] = simd.lanes_eq(s_vecs[j], c_vec_256)
+					m_vec[j] = simd.reduce_or(c_vecs[j])
+				}
+				if m_vec[0] | m_vec[1] | m_vec[2] | m_vec[3] > 0 {
+					#unroll for j in 0..<4 {
+						if m_vec[3-j] > 0 {
+							sel := simd.select(c_vecs[3-j], SCANNER_INDICES_256, SCANNER_SENTINEL_MAX_256)
+							off := simd.reduce_max(sel)
+							return i + (3-j) * SIMD_REG_SIZE_256 + int(off)
+						}
+					}
+				}
+			}
+
+			// Scan 64-byte chunks, using 256-bit SIMD.
+			for i >= 2 * SIMD_REG_SIZE_256 {
+				i -= 2 * SIMD_REG_SIZE_256
+
+				#unroll for j in 0..<2 {
+					s_vecs[j] = intrinsics.unaligned_load(cast(^simd.u8x32)raw_data(s[i+j*SIMD_REG_SIZE_256:]))
+					c_vecs[j] = simd.lanes_eq(s_vecs[j], c_vec_256)
+					m_vec[j] = simd.reduce_or(c_vecs[j])
+				}
+				if m_vec[0] | m_vec[1] > 0 {
+					#unroll for j in 0..<2 {
+						if m_vec[1-j] > 0 {
+							sel := simd.select(c_vecs[1-j], SCANNER_INDICES_256, SCANNER_SENTINEL_MAX_256)
+							off := simd.reduce_max(sel)
+							return i + (1-j) * SIMD_REG_SIZE_256 + int(off)
+						}
+					}
+				}
+			}
+		} else {
+			s_vecs: [4]simd.u8x16 = ---
+			c_vecs: [4]simd.u8x16 = ---
+			m_vecs: [4]u8 = ---
+
+			// Scan 64-byte chunks, using 128-bit SIMD.
+			for i >= 4 * SIMD_REG_SIZE_128 {
+				i -= 4 * SIMD_REG_SIZE_128
+
+				#unroll for j in 0..<4 {
+					s_vecs[j] = intrinsics.unaligned_load(cast(^simd.u8x16)raw_data(s[i+j*SIMD_REG_SIZE_128:]))
+					c_vecs[j] = simd.lanes_eq(s_vecs[j], c_vec)
+					m_vecs[j] = simd.reduce_or(c_vecs[j])
+				}
+				if m_vecs[0] | m_vecs[1] | m_vecs[2] | m_vecs[3] > 0 {
+					#unroll for j in 0..<4 {
+						if m_vecs[3-j] > 0 {
+							sel := simd.select(c_vecs[3-j], SCANNER_INDICES_128, SCANNER_SENTINEL_MAX_128)
+							off := simd.reduce_max(sel)
+							return i + (3-j) * SIMD_REG_SIZE_128 + int(off)
+						}
+					}
+				}
+			}
+		}
+	}
+
+	// Scan the remaining SIMD register sized chunks.
+	//
+	// Apparently LLVM does ok with 128-bit SWAR, so this path is also taken
+	// on potato targets.  Scanning more at a time when LLVM is emulating SIMD
+	// likely does not buy much, as all that does is increase GP register
+	// pressure.
+	for i >= SIMD_REG_SIZE_128 {
+		i -= SIMD_REG_SIZE_128
+
+		s0 := intrinsics.unaligned_load(cast(^simd.u8x16)raw_data(s[i:]))
+		c0 := simd.lanes_eq(s0, c_vec)
+		if simd.reduce_or(c0) > 0 {
+			sel := simd.select(c0, SCANNER_INDICES_128, SCANNER_SENTINEL_MAX_128)
+			off := simd.reduce_max(sel)
+			return i + int(off)
+		}
+	}
+
+	// Scan serially for the remainder.
+	for i > 0 {
+		i -= 1
 		if s[i] == c {
 			return i
 		}
 	}
+
 	return -1
 }


-
@private PRIME_RABIN_KARP :: 16777619

 index :: proc(s, substr: []byte) -> int {
@@ -1167,3 +1447,28 @@ fields_proc :: proc(s: []byte, f: proc(rune) -> bool, allocator := context.alloc

 	return subslices[:]
 }
+
+// alias returns true iff a and b have a non-zero length, and any part of
+// a overlaps with b.
+alias :: proc "contextless" (a, b: []byte) -> bool {
+	a_len, b_len := len(a), len(b)
+	if a_len == 0 || b_len == 0 {
+		return false
+	}
+
+	a_start, b_start := uintptr(raw_data(a)), uintptr(raw_data(b))
+	a_end, b_end := a_start + uintptr(a_len-1), b_start + uintptr(b_len-1)
+
+	return a_start <= b_end && b_start <= a_end
+}
+
+// alias_inexactly returns true iff a and b have a non-zero length,
+// the base pointer of a and b are NOT equal, and any part of a overlaps
+// with b (ie: `alias(a, b)` with an exception that returns false for
+// `a == b`, `b = a[:len(a)-69]` and similar conditions).
+alias_inexactly :: proc "contextless" (a, b: []byte) -> bool {
+	if raw_data(a) == raw_data(b) {
+		return false
+	}
+	return alias(a, b)
+}
@@ -9,10 +9,11 @@ Reader :: struct {
 	prev_rune: int,    // previous reading index of rune or < 0
 }

-reader_init :: proc(r: ^Reader, s: []byte) {
+reader_init :: proc(r: ^Reader, s: []byte) -> io.Stream {
 	r.s = s
 	r.i = 0
 	r.prev_rune = -1
+	return reader_to_stream(r)
 }

 reader_to_stream :: proc(r: ^Reader) -> (s: io.Stream) {
@@ -33,6 +34,9 @@ reader_size :: proc(r: ^Reader) -> i64 {
 }

 reader_read :: proc(r: ^Reader, p: []byte) -> (n: int, err: io.Error) {
+	if len(p) == 0 {
+		return 0, nil
+	}
 	if r.i >= i64(len(r.s)) {
 		return 0, .EOF
 	}
@@ -42,6 +46,9 @@ reader_read :: proc(r: ^Reader, p: []byte) -> (n: int, err: io.Error) {
 	return
 }
 reader_read_at :: proc(r: ^Reader, p: []byte, off: i64) -> (n: int, err: io.Error) {
+	if len(p) == 0 {
+		return 0, nil
+	}
 	if off < 0 {
 		return 0, .Invalid_Offset
 	}
@@ -97,7 +104,6 @@ reader_unread_rune :: proc(r: ^Reader) -> io.Error {
 	return nil
 }
 reader_seek :: proc(r: ^Reader, offset: i64, whence: io.Seek_From) -> (i64, io.Error) {
-	r.prev_rune = -1
 	abs: i64
 	switch whence {
 	case .Start:
@@ -114,6 +120,7 @@ reader_seek :: proc(r: ^Reader, offset: i64, whence: io.Seek_From) -> (i64, io.E
 		return 0, .Invalid_Offset
 	}
 	r.i = abs
+	r.prev_rune = -1
 	return abs, nil
 }
 reader_write_to :: proc(r: ^Reader, w: io.Writer) -> (n: i64, err: io.Error) {
@@ -1,34 +1,31 @@
 /*
-package demo
+Example:
+	package demo

-import tokenizer "core:c/frontend/tokenizer"
-import preprocessor "core:c/frontend/preprocessor"
-import "core:fmt"
+	import tokenizer "core:c/frontend/tokenizer"
+	import preprocessor "core:c/frontend/preprocessor"
+	import "core:fmt"

-main :: proc() {
-	t := &tokenizer.Tokenizer{};
-	tokenizer.init_defaults(t);
+	main :: proc() {
+		t := &tokenizer.Tokenizer{};
+		tokenizer.init_defaults(t);

-	cpp := &preprocessor.Preprocessor{};
-	cpp.warn, cpp.err = t.warn, t.err;
-	preprocessor.init_lookup_tables(cpp);
-	preprocessor.init_default_macros(cpp);
-	cpp.include_paths = {"my/path/to/include"};
+		cpp := &preprocessor.Preprocessor{};
+		cpp.warn, cpp.err = t.warn, t.err;
+		preprocessor.init_lookup_tables(cpp);
+		preprocessor.init_default_macros(cpp);
+		cpp.include_paths = {"my/path/to/include"};

-	tok := tokenizer.tokenize_file(t, "the/source/file.c", 1);
+		tok := tokenizer.tokenize_file(t, "the/source/file.c", 1);

-	tok = preprocessor.preprocess(cpp, tok);
-	if tok != nil {
-		for t := tok; t.kind != .EOF; t = t.next {
-			fmt.println(t.lit);
+		tok = preprocessor.preprocess(cpp, tok);
+		if tok != nil {
+			for t := tok; t.kind != .EOF; t = t.next {
+				fmt.println(t.lit);
+			}
 		}
+
+		fmt.println("[Done]");
 	}
-
-	fmt.println("[Done]");
-}
 */
-
-
 package c_frontend_tokenizer
-
-
@@ -47,8 +47,8 @@ foreign libc {
 	clogf   :: proc(z: complex_float) -> complex_float ---

 	// 7.3.8 Power and absolute-value functions
-	cabs    :: proc(z: complex_double) -> complex_double ---
-	cabsf   :: proc(z: complex_float) -> complex_float ---
+	cabs    :: proc(z: complex_double) -> double ---
+	cabsf   :: proc(z: complex_float) -> float ---
 	cpow    :: proc(x, y: complex_double) -> complex_double ---
 	cpowf   :: proc(x, y: complex_float) -> complex_float ---
 	csqrt   :: proc(z: complex_double) -> complex_double ---
@@ -40,7 +40,7 @@ when ODIN_OS == .FreeBSD {
 	ERANGE :: 34
 }

-when ODIN_OS == .OpenBSD {
+when ODIN_OS == .OpenBSD || ODIN_OS == .NetBSD {
 	@(private="file")
 	@(default_calling_convention="c")
 	foreign libc {
@@ -102,6 +102,6 @@ when ODIN_OS == .Haiku {
 // read the value, or to produce an lvalue such that you can assign a different
 // error value to errno. To work around this, just expose it as a function like
 // it actually is.
-errno :: #force_inline proc() -> ^int {
+errno :: #force_inline proc "contextless" () -> ^int {
 	return _get_errno()
 }
@@ -32,24 +32,21 @@ when ODIN_OS == .Windows {
 		// the RDX register will contain zero and correctly set the flag to disable
 		// stack unwinding.
 		@(link_name="_setjmp")
-		setjmp  :: proc(env: ^jmp_buf, hack: rawptr = nil) -> int ---
+		setjmp :: proc(env: ^jmp_buf, hack: rawptr = nil) -> int ---
 	}
 } else {
 	@(default_calling_convention="c")
 	foreign libc {
 		// 7.13.1 Save calling environment
-		//
-		// NOTE(dweiler): C11 requires setjmp be a macro, which means it won't
-		// necessarily export a symbol named setjmp but rather _setjmp in the case
-		// of musl, glibc, BSD libc, and msvcrt.
-		@(link_name="_setjmp")
-		setjmp  :: proc(env: ^jmp_buf) -> int ---
+		@(link_name=LSETJMP)
+		setjmp :: proc(env: ^jmp_buf) -> int ---
 	}
 }

@(default_calling_convention="c")
 foreign libc {
 	// 7.13.2 Restore calling environment
+	@(link_name=LLONGJMP)
 	longjmp :: proc(env: ^jmp_buf, val: int) -> ! ---
 }

@@ -64,3 +61,11 @@ foreign libc {
 // The choice of 4096 bytes for storage of this type is more than enough on all
 // relevant platforms.
 jmp_buf :: struct #align(16) { _: [4096]char, }
+
+when ODIN_OS == .NetBSD {
+	@(private) LSETJMP  :: "__setjmp14"
+	@(private) LLONGJMP :: "__longjmp14"
+} else {
+	@(private) LSETJMP  :: "setjmp"
+	@(private) LLONGJMP :: "longjmp"
+}
@@ -34,20 +34,7 @@ when ODIN_OS == .Windows {
 	SIGTERM :: 15
 }

-when ODIN_OS == .Linux || ODIN_OS == .FreeBSD {
-	SIG_ERR  :: rawptr(~uintptr(0))
-	SIG_DFL  :: rawptr(uintptr(0))
-	SIG_IGN  :: rawptr(uintptr(1)) 
-
-	SIGABRT  :: 6
-	SIGFPE   :: 8
-	SIGILL   :: 4
-	SIGINT   :: 2
-	SIGSEGV  :: 11
-	SIGTERM  :: 15
-}
-
-when ODIN_OS == .Darwin {
+when ODIN_OS == .Linux || ODIN_OS == .FreeBSD || ODIN_OS == .Haiku || ODIN_OS == .OpenBSD || ODIN_OS == .NetBSD || ODIN_OS == .Darwin {
 	SIG_ERR  :: rawptr(~uintptr(0))
 	SIG_DFL  :: rawptr(uintptr(0))
 	SIG_IGN  :: rawptr(uintptr(1)) 
@@ -17,6 +17,12 @@ when ODIN_OS == .Windows {

 FILE :: struct {}

+Whence :: enum int {
+	SET = SEEK_SET,
+	CUR = SEEK_CUR,
+	END = SEEK_END,
+}
+
 // MSVCRT compatible.
 when ODIN_OS == .Windows {
 	_IOFBF       :: 0x0000
@@ -83,7 +89,7 @@ when ODIN_OS == .Linux {
 	}
 }

-when ODIN_OS == .OpenBSD {
+when ODIN_OS == .OpenBSD || ODIN_OS == .NetBSD {
 	fpos_t :: distinct i64

 	_IOFBF :: 0
@@ -101,11 +107,15 @@ when ODIN_OS == .OpenBSD {
 	SEEK_CUR :: 1
 	SEEK_END :: 2

+	TMP_MAX :: 308915776
+
 	foreign libc {
-		stderr: ^FILE
-		stdin:  ^FILE
-		stdout: ^FILE
+		__sF: [3]FILE
 	}
+
+	stdin:  ^FILE = &__sF[0]
+	stdout: ^FILE = &__sF[1]
+	stderr: ^FILE = &__sF[2]
 }

 when ODIN_OS == .FreeBSD {
@@ -126,10 +136,12 @@ when ODIN_OS == .FreeBSD {
 	SEEK_CUR :: 1
 	SEEK_END :: 2

+	TMP_MAX :: 308915776
+
 	foreign libc {
-		stderr: ^FILE
-		stdin:  ^FILE
-		stdout: ^FILE
+		@(link_name="__stderrp") stderr: ^FILE
+		@(link_name="__stdinp")  stdin:  ^FILE
+		@(link_name="__stdoutp") stdout: ^FILE
 	}
 }

@@ -193,10 +205,21 @@ when ODIN_OS == .Haiku {
 	}
 }

+when ODIN_OS == .NetBSD {
+	@(private) LRENAME  :: "__posix_rename"
+	@(private) LFGETPOS :: "__fgetpos50"
+	@(private) LFSETPOS :: "__fsetpos50"
+} else {
+	@(private) LRENAME  :: "rename"
+	@(private) LFGETPOS :: "fgetpos"
+	@(private) LFSETPOS :: "fsetpos"
+}
+
@(default_calling_convention="c")
 foreign libc {
 	// 7.21.4 Operations on files
 	remove    :: proc(filename: cstring) -> int ---
+	@(link_name=LRENAME)
 	rename    :: proc(old, new: cstring) -> int ---
 	tmpfile   :: proc() -> ^FILE ---
 	tmpnam    :: proc(s: [^]char) -> [^]char ---
@@ -238,8 +261,10 @@ foreign libc {
 	fwrite    :: proc(ptr: rawptr, size: size_t, nmemb: size_t, stream: ^FILE) -> size_t ---

 	// 7.21.9 File positioning functions
+	@(link_name=LFGETPOS)
 	fgetpos   :: proc(stream: ^FILE, pos: ^fpos_t) -> int ---
-	fseek     :: proc(stream: ^FILE, offset: long, whence: int) -> int ---
+	fseek     :: proc(stream: ^FILE, offset: long, whence: Whence) -> int ---
+	@(link_name=LFSETPOS)
 	fsetpos   :: proc(stream: ^FILE, pos: ^fpos_t) -> int ---
 	ftell     :: proc(stream: ^FILE) -> long ---
 	rewind    :: proc(stream: ^FILE) ---
@@ -286,11 +311,11 @@ to_stream :: proc(file: ^FILE) -> io.Stream {
 				return 0, unknown_or_eof(file)
 			}

-			if fseek(file, long(offset), SEEK_SET) != 0 {
+			if fseek(file, long(offset), .SET) != 0 {
 				return 0, unknown_or_eof(file)
 			}

-			defer fseek(file, long(curr), SEEK_SET)
+			defer fseek(file, long(curr), .SET)

 			n = i64(fread(raw_data(p), size_of(byte), len(p), file))
 			if n == 0 { err = unknown_or_eof(file) }
@@ -305,17 +330,21 @@ to_stream :: proc(file: ^FILE) -> io.Stream {
 				return 0, unknown_or_eof(file)
 			}

-			if fseek(file, long(offset), SEEK_SET) != 0 {
+			if fseek(file, long(offset), .SET) != 0 {
 				return 0, unknown_or_eof(file)
 			}

-			defer fseek(file, long(curr), SEEK_SET)
+			defer fseek(file, long(curr), .SET)

 			n = i64(fwrite(raw_data(p), size_of(byte), len(p), file))
 			if n == 0 { err = unknown_or_eof(file) }

 		case .Seek:
-			if fseek(file, long(offset), int(whence)) != 0 {
+			#assert(int(Whence.SET) == int(io.Seek_From.Start))
+			#assert(int(Whence.CUR) == int(io.Seek_From.Current))
+			#assert(int(Whence.END) == int(io.Seek_From.End))
+
+			if fseek(file, long(offset), Whence(whence)) != 0 {
 				return 0, unknown_or_eof(file)
 			}
 		
@@ -324,9 +353,9 @@ to_stream :: proc(file: ^FILE) -> io.Stream {
 			if curr == -1 {
 				return 0, unknown_or_eof(file)
 			}
-			defer fseek(file, curr, SEEK_SET)
+			defer fseek(file, curr, .SET)

-			if fseek(file, 0, SEEK_END) != 0 {
+			if fseek(file, 0, .END) != 0 {
 				return 0, unknown_or_eof(file)
 			}

@@ -339,7 +368,7 @@ to_stream :: proc(file: ^FILE) -> io.Stream {
 			return 0, .Empty
 		
 		case .Query:
-			return io.query_utility({ .Close, .Flush, .Read, .Read_At, .Write, .Write_At, .Seek, .Size })
+			return io.query_utility({ .Close, .Flush, .Read, .Read_At, .Write, .Write_At, .Seek, .Size, .Query })
 		}
 		return
 	}
@@ -40,10 +40,9 @@ when ODIN_OS == .Linux {
 }


-when ODIN_OS == .Darwin {
+when ODIN_OS == .Darwin || ODIN_OS == .FreeBSD || ODIN_OS == .OpenBSD {
 	RAND_MAX :: 0x7fffffff

-	// GLIBC and MUSL only
 	@(private="file")
 	@(default_calling_convention="c")
 	foreign libc {
@@ -55,6 +54,20 @@ when ODIN_OS == .Darwin {
 	}
 }

+when ODIN_OS == .NetBSD {
+	RAND_MAX :: 0x7fffffff
+
+	@(private="file")
+	@(default_calling_convention="c")
+	foreign libc {
+		__mb_cur_max: size_t
+	}
+
+	MB_CUR_MAX :: #force_inline proc() -> size_t {
+		return __mb_cur_max
+	}
+}
+
 // C does not declare what these values should be, as an implementation is free
 // to use any two distinct values it wants to indicate success or failure.
 // However, nobody actually does and everyone appears to have agreed upon these
@@ -99,7 +112,7 @@ foreign libc {
 	at_quick_exit :: proc(func: proc "c" ()) -> int ---
 	exit          :: proc(status: int) -> ! ---
 	_Exit         :: proc(status: int) -> ! ---
-	getenv        :: proc(name: cstring) -> [^]char ---
+	getenv        :: proc(name: cstring) -> cstring ---
 	quick_exit    :: proc(status: int) -> ! ---
 	system        :: proc(cmd: cstring) -> int ---

@@ -150,4 +163,4 @@ aligned_free :: #force_inline proc "c" (ptr: rawptr) {
 	} else {
 		free(ptr)
 	}
-}
+}
@@ -40,7 +40,7 @@ foreign libc {
 	strtok   :: proc(s1: [^]char, s2: cstring) -> [^]char ---

 	// 7.24.6 Miscellaneous functions
-	strerror :: proc(errnum: int) -> [^]char ---
+	strerror :: proc(errnum: int) -> cstring ---
 	strlen   :: proc(s: cstring) -> size_t ---
 }
 memset :: proc "c" (s: rawptr, c: int, n: size_t) -> rawptr {
@@ -45,35 +45,61 @@ when ODIN_OS == .Windows {
 	}
 }

-when ODIN_OS == .Linux || ODIN_OS == .FreeBSD || ODIN_OS == .Darwin || ODIN_OS == .OpenBSD || ODIN_OS == .Haiku {
+when ODIN_OS == .Linux || ODIN_OS == .FreeBSD || ODIN_OS == .Darwin || ODIN_OS == .OpenBSD || ODIN_OS == .NetBSD || ODIN_OS == .Haiku {
 	@(default_calling_convention="c")
 	foreign libc {
 		// 7.27.2 Time manipulation functions
 		clock        :: proc() -> clock_t ---
+		@(link_name=LDIFFTIME)
 		difftime     :: proc(time1, time2: time_t) -> double ---
+		@(link_name=LMKTIME)
 		mktime       :: proc(timeptr: ^tm) -> time_t ---
+		@(link_name=LTIME)
 		time         :: proc(timer: ^time_t) -> time_t ---
 		timespec_get :: proc(ts: ^timespec, base: int) -> int ---

 		// 7.27.3 Time conversion functions
 		asctime      :: proc(timeptr: ^tm) -> [^]char ---
+		@(link_name=LCTIME)
 		ctime        :: proc(timer: ^time_t) -> [^]char ---
+		@(link_name=LGMTIME)
 		gmtime       :: proc(timer: ^time_t) -> ^tm ---
+		@(link_name=LLOCALTIME)
 		localtime    :: proc(timer: ^time_t) -> ^tm ---
 		strftime     :: proc(s: [^]char, maxsize: size_t, format: cstring, timeptr: ^tm) -> size_t ---
 	}

+	when ODIN_OS == .NetBSD {
+		@(private) LDIFFTIME  :: "__difftime50"
+		@(private) LMKTIME    :: "__mktime50"
+		@(private) LTIME      :: "__time50"
+		@(private) LCTIME     :: "__ctime50"
+		@(private) LGMTIME    :: "__gmtime50"
+		@(private) LLOCALTIME :: "__localtime50"
+	} else {
+		@(private) LDIFFTIME  :: "difftime"
+		@(private) LMKTIME    :: "mktime"
+		@(private) LTIME      :: "time"
+		@(private) LCTIME     :: "ctime"
+		@(private) LGMTIME    :: "gmtime"
+		@(private) LLOCALTIME :: "localtime"
+	}
+
 	when ODIN_OS == .OpenBSD {
 		CLOCKS_PER_SEC :: 100
 	} else {
 		CLOCKS_PER_SEC :: 1000000
 	}

-	TIME_UTC       :: 1
+	TIME_UTC :: 1

-	time_t         :: distinct i64
+	time_t :: distinct i64

-	clock_t        :: long
+	when ODIN_OS == .FreeBSD || ODIN_OS == .NetBSD {
+		clock_t :: distinct int32_t
+	} else {
+		clock_t :: distinct long
+	}

 	timespec :: struct {
 		tv_sec:  time_t,
@@ -22,7 +22,7 @@ when ODIN_OS == .Windows {
 	wctrans_t :: distinct int
 	wctype_t  :: distinct u32

-} else when ODIN_OS == .OpenBSD {
+} else when ODIN_OS == .OpenBSD || ODIN_OS == .NetBSD {
 	wctrans_t :: distinct rawptr
 	wctype_t  :: distinct rawptr

@@ -34,13 +34,13 @@ COMPRESS_OUTPUT_ALLOCATE_MIN :: int(#config(COMPRESS_OUTPUT_ALLOCATE_MIN, 1 << 2
 */
 when size_of(uintptr) == 8 {

-        // For 64-bit platforms, we set the default max buffer size to 4 GiB,
-        // which is GZIP and PKZIP's max payload size.
+	// For 64-bit platforms, we set the default max buffer size to 4 GiB,
+	// which is GZIP and PKZIP's max payload size.
 	COMPRESS_OUTPUT_ALLOCATE_MAX :: int(#config(COMPRESS_OUTPUT_ALLOCATE_MAX, 1 << 32))
 } else {
 	
 	// For 32-bit platforms, we set the default max buffer size to 512 MiB.
-        COMPRESS_OUTPUT_ALLOCATE_MAX :: int(#config(COMPRESS_OUTPUT_ALLOCATE_MAX, 1 << 29))
+	COMPRESS_OUTPUT_ALLOCATE_MAX :: int(#config(COMPRESS_OUTPUT_ALLOCATE_MAX, 1 << 29))
 }


@@ -186,7 +186,7 @@ input_size_from_stream :: proc(z: ^Context_Stream_Input) -> (res: i64, err: Erro

 input_size :: proc{input_size_from_memory, input_size_from_stream}

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_slice_from_memory :: #force_inline proc(z: ^Context_Memory_Input, size: int) -> (res: []u8, err: io.Error) {
 	#no_bounds_check {
 		if len(z.input_data) >= size {
@@ -203,7 +203,7 @@ read_slice_from_memory :: #force_inline proc(z: ^Context_Memory_Input, size: int
 	}
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_slice_from_stream :: #force_inline proc(z: ^Context_Stream_Input, size: int) -> (res: []u8, err: io.Error) {
 	// TODO: REMOVE ALL USE OF context.temp_allocator here
 	// there is literally no need for it
@@ -214,13 +214,13 @@ read_slice_from_stream :: #force_inline proc(z: ^Context_Stream_Input, size: int

 read_slice :: proc{read_slice_from_memory, read_slice_from_stream}

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_data :: #force_inline proc(z: ^$C, $T: typeid) -> (res: T, err: io.Error) {
 	b := read_slice(z, size_of(T)) or_return
 	return (^T)(&b[0])^, nil
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_u8_from_memory :: #force_inline proc(z: ^Context_Memory_Input) -> (res: u8, err: io.Error) {
 	#no_bounds_check {
 		if len(z.input_data) >= 1 {
@@ -232,7 +232,7 @@ read_u8_from_memory :: #force_inline proc(z: ^Context_Memory_Input) -> (res: u8,
 	return 0, .EOF
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_u8_from_stream :: #force_inline proc(z: ^Context_Stream_Input) -> (res: u8, err: io.Error) {
 	b := read_slice_from_stream(z, 1) or_return
 	return b[0], nil
@@ -242,7 +242,7 @@ read_u8 :: proc{read_u8_from_memory, read_u8_from_stream}

 // You would typically only use this at the end of Inflate, to drain bits from the code buffer
 // preferentially.
-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_u8_prefer_code_buffer_lsb :: #force_inline proc(z: ^$C) -> (res: u8, err: io.Error) {
 	if z.num_bits >= 8 {
 		res = u8(read_bits_no_refill_lsb(z, 8))
@@ -257,7 +257,7 @@ read_u8_prefer_code_buffer_lsb :: #force_inline proc(z: ^$C) -> (res: u8, err: i
 	return
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_data_from_memory :: #force_inline proc(z: ^Context_Memory_Input, $T: typeid) -> (res: T, err: io.Error) {
 	size :: size_of(T)

@@ -275,7 +275,7 @@ peek_data_from_memory :: #force_inline proc(z: ^Context_Memory_Input, $T: typeid
 	}
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_data_at_offset_from_memory :: #force_inline proc(z: ^Context_Memory_Input, $T: typeid, #any_int offset: int) -> (res: T, err: io.Error) {
 	size :: size_of(T)

@@ -293,7 +293,7 @@ peek_data_at_offset_from_memory :: #force_inline proc(z: ^Context_Memory_Input,
 	}
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_data_from_stream :: #force_inline proc(z: ^Context_Stream_Input, $T: typeid) -> (res: T, err: io.Error) {
 	size :: size_of(T)

@@ -317,7 +317,7 @@ peek_data_from_stream :: #force_inline proc(z: ^Context_Stream_Input, $T: typeid
 	return res, .None
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_data_at_offset_from_stream :: #force_inline proc(z: ^Context_Stream_Input, $T: typeid, #any_int offset: int) -> (res: T, err: io.Error) {
 	size :: size_of(T)

@@ -352,14 +352,14 @@ peek_data :: proc{peek_data_from_memory, peek_data_from_stream, peek_data_at_off


 // Sliding window read back
-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_back_byte :: #force_inline proc(z: ^$C, offset: i64) -> (res: u8, err: io.Error) {
 	// Look back into the sliding window.
 	return z.output.buf[z.bytes_written - offset], .None
 }

 // Generalized bit reader LSB
-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 refill_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width := i8(48)) {
 	refill := u64(width)
 	b      := u64(0)
@@ -385,7 +385,7 @@ refill_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width :=
 }

 // Generalized bit reader LSB
-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 refill_lsb_from_stream :: proc(z: ^Context_Stream_Input, width := i8(24)) {
 	refill := u64(width)

@@ -414,13 +414,13 @@ refill_lsb_from_stream :: proc(z: ^Context_Stream_Input, width := i8(24)) {
 refill_lsb :: proc{refill_lsb_from_memory, refill_lsb_from_stream}


-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 consume_bits_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width: u8) {
 	z.code_buffer >>= width
 	z.num_bits -= u64(width)
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 consume_bits_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, width: u8) {
 	z.code_buffer >>= width
 	z.num_bits -= u64(width)
@@ -428,7 +428,7 @@ consume_bits_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, wid

 consume_bits_lsb :: proc{consume_bits_lsb_from_memory, consume_bits_lsb_from_stream}

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_bits_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width: u8) -> u32 {
 	if z.num_bits < u64(width) {
 		refill_lsb(z)
@@ -436,7 +436,7 @@ peek_bits_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width:
 	return u32(z.code_buffer &~ (~u64(0) << width))
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_bits_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, width: u8) -> u32 {
 	if z.num_bits < u64(width) {
 		refill_lsb(z)
@@ -446,13 +446,13 @@ peek_bits_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, width:

 peek_bits_lsb :: proc{peek_bits_lsb_from_memory, peek_bits_lsb_from_stream}

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_bits_no_refill_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width: u8) -> u32 {
 	assert(z.num_bits >= u64(width))
 	return u32(z.code_buffer &~ (~u64(0) << width))
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 peek_bits_no_refill_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, width: u8) -> u32 {
 	assert(z.num_bits >= u64(width))
 	return u32(z.code_buffer &~ (~u64(0) << width))
@@ -460,14 +460,14 @@ peek_bits_no_refill_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Inp

 peek_bits_no_refill_lsb :: proc{peek_bits_no_refill_lsb_from_memory, peek_bits_no_refill_lsb_from_stream}

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_bits_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width: u8) -> u32 {
 	k := #force_inline peek_bits_lsb(z, width)
 	#force_inline consume_bits_lsb(z, width)
 	return k
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_bits_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, width: u8) -> u32 {
 	k := peek_bits_lsb(z, width)
 	consume_bits_lsb(z, width)
@@ -476,14 +476,14 @@ read_bits_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, width:

 read_bits_lsb :: proc{read_bits_lsb_from_memory, read_bits_lsb_from_stream}

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_bits_no_refill_lsb_from_memory :: #force_inline proc(z: ^Context_Memory_Input, width: u8) -> u32 {
 	k := #force_inline peek_bits_no_refill_lsb(z, width)
 	#force_inline consume_bits_lsb(z, width)
 	return k
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 read_bits_no_refill_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Input, width: u8) -> u32 {
 	k := peek_bits_no_refill_lsb(z, width)
 	consume_bits_lsb(z, width)
@@ -493,14 +493,14 @@ read_bits_no_refill_lsb_from_stream :: #force_inline proc(z: ^Context_Stream_Inp
 read_bits_no_refill_lsb :: proc{read_bits_no_refill_lsb_from_memory, read_bits_no_refill_lsb_from_stream}


-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 discard_to_next_byte_lsb_from_memory :: proc(z: ^Context_Memory_Input) {
 	discard := u8(z.num_bits & 7)
 	#force_inline consume_bits_lsb(z, discard)
 }


-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 discard_to_next_byte_lsb_from_stream :: proc(z: ^Context_Stream_Input) {
 	discard := u8(z.num_bits & 7)
 	consume_bits_lsb(z, discard)
@@ -1,5 +1,5 @@
 //+build ignore
-package gzip
+package compress_gzip

 /*
 	Copyright 2021 Jeroen van Rijn <nom@duclavier.com>.
@@ -1,4 +1,4 @@
-package gzip
+package compress_gzip

 /*
 	Copyright 2021 Jeroen van Rijn <nom@duclavier.com>.
@@ -5,7 +5,7 @@
 */

 // package shoco is an implementation of the shoco short string compressor
-package shoco
+package compress_shoco

 DEFAULT_MODEL :: Shoco_Model {
 	min_char = 39,
@@ -9,7 +9,7 @@
 */

 // package shoco is an implementation of the shoco short string compressor
-package shoco
+package compress_shoco

 import "base:intrinsics"
 import "core:compress"
@@ -98,7 +98,7 @@ decompress_slice_to_output_buffer :: proc(input: []u8, output: []u8, model := DE
 	validate_model(model) or_return

 	for inp < inp_end {
-		val  := transmute(i8)input[inp]
+		val  := i8(input[inp])
 		mark := int(-1)

 		for val < 0 {
@@ -274,12 +274,9 @@ compress_string_to_buffer :: proc(input: string, output: []u8, model := DEFAULT_
 				out_ptr := raw_data(output[out:])

 				switch pack.bytes_packed {
-				case 4:
-					intrinsics.unaligned_store(transmute(^u32)out_ptr, code)
-				case 2:
-					intrinsics.unaligned_store(transmute(^u16)out_ptr, u16(code))
-				case 1:
-					intrinsics.unaligned_store(transmute(^u8)out_ptr,  u8(code))
+				case 4: intrinsics.unaligned_store((^u32)(out_ptr), code)
+				case 2: intrinsics.unaligned_store((^u16)(out_ptr), u16(code))
+				case 1: intrinsics.unaligned_store( (^u8)(out_ptr),  u8(code))
 				case:
 					return out, .Unknown_Compression_Method
 				}
@@ -1,5 +1,5 @@
 //+build ignore
-package zlib
+package compress_zlib

 /*
 	Copyright 2021 Jeroen van Rijn <nom@duclavier.com>.
@@ -1,5 +1,5 @@
 //+vet !using-param
-package zlib
+package compress_zlib

 /*
 	Copyright 2021 Jeroen van Rijn <nom@duclavier.com>.
@@ -12,6 +12,7 @@ package zlib

 import "core:compress"

+import "base:intrinsics"
 import "core:mem"
 import "core:io"
 import "core:hash"
@@ -120,23 +121,17 @@ Huffman_Table :: struct {
 }

 // Implementation starts here
-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 z_bit_reverse :: #force_inline proc(n: u16, bits: u8) -> (r: u16) {
 	assert(bits <= 16)
-	// NOTE: Can optimize with llvm.bitreverse.i64 or some bit twiddling
-	// by reversing all of the bits and masking out the unneeded ones.
-	r = n
-	r = ((r & 0xAAAA) >>  1) | ((r & 0x5555) << 1)
-	r = ((r & 0xCCCC) >>  2) | ((r & 0x3333) << 2)
-	r = ((r & 0xF0F0) >>  4) | ((r & 0x0F0F) << 4)
-	r = ((r & 0xFF00) >>  8) | ((r & 0x00FF) << 8)
+	r = intrinsics.reverse_bits(n)

 	r >>= (16 - bits)
 	return
 }


-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 grow_buffer :: proc(buf: ^[dynamic]u8) -> (err: compress.Error) {
 	/*
 		That we get here at all means that we didn't pass an expected output size,
@@ -154,7 +149,7 @@ grow_buffer :: proc(buf: ^[dynamic]u8) -> (err: compress.Error) {
 	TODO: Make these return compress.Error.
 */

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 write_byte :: #force_inline proc(z: ^$C, c: u8) -> (err: io.Error) #no_bounds_check {
 	/*
 		Resize if needed.
@@ -173,7 +168,7 @@ write_byte :: #force_inline proc(z: ^$C, c: u8) -> (err: io.Error) #no_bounds_ch
 	return .None
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 repl_byte :: proc(z: ^$C, count: u16, c: u8) -> (err: io.Error) #no_bounds_check {
 	/*
 		TODO(Jeroen): Once we have a magic ring buffer, we can just peek/write into it
@@ -201,7 +196,7 @@ repl_byte :: proc(z: ^$C, count: u16, c: u8) -> (err: io.Error) #no_bounds_check
 	return .None
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 repl_bytes :: proc(z: ^$C, count: u16, distance: u16) -> (err: io.Error) {
 	/*
 		TODO(Jeroen): Once we have a magic ring buffer, we can just peek/write into it
@@ -234,8 +229,8 @@ allocate_huffman_table :: proc(allocator := context.allocator) -> (z: ^Huffman_T
 	return new(Huffman_Table, allocator), nil
 }

-@(optimization_mode="speed")
-build_huffman :: proc(z: ^Huffman_Table, code_lengths: []u8) -> (err: Error) {
+@(optimization_mode="favor_size")
+build_huffman :: #force_no_inline proc(z: ^Huffman_Table, code_lengths: []u8) -> (err: Error) {
 	sizes:     [HUFFMAN_MAX_BITS+1]int
 	next_code: [HUFFMAN_MAX_BITS+1]int

@@ -293,7 +288,7 @@ build_huffman :: proc(z: ^Huffman_Table, code_lengths: []u8) -> (err: Error) {
 	return nil
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 decode_huffman_slowpath :: proc(z: ^$C, t: ^Huffman_Table) -> (r: u16, err: Error) #no_bounds_check {
 	code := u16(compress.peek_bits_lsb(z,16))

@@ -324,7 +319,7 @@ decode_huffman_slowpath :: proc(z: ^$C, t: ^Huffman_Table) -> (r: u16, err: Erro
 	return r, nil
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 decode_huffman :: proc(z: ^$C, t: ^Huffman_Table) -> (r: u16, err: Error) #no_bounds_check {
 	if z.num_bits < 16 {
 		if z.num_bits > 63 {
@@ -344,7 +339,7 @@ decode_huffman :: proc(z: ^$C, t: ^Huffman_Table) -> (r: u16, err: Error) #no_bo
 	return decode_huffman_slowpath(z, t)
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 parse_huffman_block :: proc(z: ^$C, z_repeat, z_offset: ^Huffman_Table) -> (err: Error) #no_bounds_check {
 	#no_bounds_check for {
 		value, e := decode_huffman(z, z_repeat)
@@ -413,7 +408,7 @@ parse_huffman_block :: proc(z: ^$C, z_repeat, z_offset: ^Huffman_Table) -> (err:
 	}
 }

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 inflate_from_context :: proc(using ctx: ^compress.Context_Memory_Input, raw := false, expected_output_size := -1, allocator := context.allocator) -> (err: Error) #no_bounds_check {
 	/*
 		ctx.output must be a bytes.Buffer for now. We'll add a separate implementation that writes to a stream.
@@ -486,7 +481,7 @@ inflate_from_context :: proc(using ctx: ^compress.Context_Memory_Input, raw := f

 // TODO: Check alignment of reserve/resize.

-@(optimization_mode="speed")
+@(optimization_mode="favor_size")
 inflate_raw :: proc(z: ^$C, expected_output_size := -1, allocator := context.allocator) -> (err: Error) #no_bounds_check {
 	context.allocator = allocator
 	expected_output_size := expected_output_size
@@ -670,4 +665,4 @@ inflate_from_byte_array_raw :: proc(input: []u8, buf: ^bytes.Buffer, raw := fals
 	return inflate_raw(&ctx, expected_output_size=expected_output_size)
 }

-inflate :: proc{inflate_from_context, inflate_from_byte_array}
+inflate :: proc{inflate_from_context, inflate_from_byte_array}
@@ -5,13 +5,10 @@ The implementation is non-intrusive, and non-recursive.
 */
 package container_avl

-import "base:intrinsics"
-import "base:runtime"
+@(require) import "base:intrinsics"
+@(require) import "base:runtime"
 import "core:slice"

-_ :: intrinsics
-_ :: runtime
-
 // Originally based on the CC0 implementation by Eric Biggers
 // See: https://github.com/ebiggers/avl_tree/

@@ -90,7 +87,7 @@ init_cmp :: proc(
 init_ordered :: proc(
 	t: ^$T/Tree($Value),
 	node_allocator := context.allocator,
-) where intrinsics.type_is_ordered_numeric(Value) {
+) where intrinsics.type_is_ordered(Value) {
 	init_cmp(t, slice.cmp_proc(Value), node_allocator)
 }

@@ -675,4 +672,4 @@ iterator_first :: proc "contextless" (it: ^Iterator($Value)) {
 	if it._cur != nil {
 		it._next = node_next_or_prev_in_order(it._cur, it._direction)
 	}
-}
+}
@@ -1,4 +1,4 @@
-package dynamic_bit_array
+package container_dynamic_bit_array

 import "base:intrinsics"
 import "core:mem"
@@ -210,8 +210,11 @@ set :: proc(ba: ^Bit_Array, #any_int index: uint, set_to: bool = true, allocator

 	ba.max_index = max(idx, ba.max_index)

-	if set_to{ ba.bits[leg_index] |= 1 << uint(bit_index) }
-	else { ba.bits[leg_index] &= ~(1 << uint(bit_index)) }
+	if set_to {
+		ba.bits[leg_index] |=  1 << uint(bit_index)
+	} else {
+		ba.bits[leg_index] &~= 1 << uint(bit_index)
+	}

 	return true
 }
@@ -253,7 +256,7 @@ Inputs:
 - index: Which bit in the array
 */
 unsafe_unset :: proc(b: ^Bit_Array, bit: int) #no_bounds_check {
-	b.bits[bit >> INDEX_SHIFT] &= ~(1 << uint(bit & INDEX_MASK))
+	b.bits[bit >> INDEX_SHIFT] &~= 1 << uint(bit & INDEX_MASK)
 }
 /*
 A helper function to create a Bit Array with optional bias, in case your smallest index is non-zero (including negative).
@@ -1,53 +1,52 @@
-package dynamic_bit_array
-
 /*
-	The Bit Array can be used in several ways:
+The Bit Array can be used in several ways:

-	-- By default you don't need to instantiate a Bit Array:
+- By default you don't need to instantiate a Bit Array:

-		package test
+	package test

-		import "core:fmt"
-		import "core:container/bit_array"
+	import "core:fmt"
+	import "core:container/bit_array"

-		main :: proc() {
-			using bit_array
+	main :: proc() {
+		using bit_array

-			bits: Bit_Array
+		bits: Bit_Array

-			// returns `true`
-			fmt.println(set(&bits, 42))
+		// returns `true`
+		fmt.println(set(&bits, 42))

-			// returns `false`, `false`, because this Bit Array wasn't created to allow negative indices.
-			was_set, was_retrieved := get(&bits, -1)
-			fmt.println(was_set, was_retrieved) 
-			destroy(&bits)
+		// returns `false`, `false`, because this Bit Array wasn't created to allow negative indices.
+		was_set, was_retrieved := get(&bits, -1)
+		fmt.println(was_set, was_retrieved) 
+		destroy(&bits)
+	}
+
+- A Bit Array can optionally allow for negative indices, if the minimum value was given during creation:
+
+	package test
+
+	import "core:fmt"
+	import "core:container/bit_array"
+
+	main :: proc() {
+		Foo :: enum int {
+			Negative_Test = -42,
+			Bar           = 420,
+			Leaves        = 69105,
 		}

-	-- A Bit Array can optionally allow for negative indices, if the mininum value was given during creation:
+		using bit_array

-		package test
+		bits := create(int(max(Foo)), int(min(Foo)))
+		defer destroy(bits)

-		import "core:fmt"
-		import "core:container/bit_array"
-
-		main :: proc() {
-			Foo :: enum int {
-				Negative_Test = -42,
-				Bar           = 420,
-				Leaves        = 69105,
-			}
-
-			using bit_array
-
-			bits := create(int(max(Foo)), int(min(Foo)))
-			defer destroy(bits)
-
-			fmt.printf("Set(Bar):           %v\n",     set(bits, Foo.Bar))
-			fmt.printf("Get(Bar):           %v, %v\n", get(bits, Foo.Bar))
-			fmt.printf("Set(Negative_Test): %v\n",     set(bits, Foo.Negative_Test))
-			fmt.printf("Get(Leaves):        %v, %v\n", get(bits, Foo.Leaves))
-			fmt.printf("Get(Negative_Test): %v, %v\n", get(bits, Foo.Negative_Test))
-			fmt.printf("Freed.\n")
-		}
-*/
+		fmt.printf("Set(Bar):           %v\n",     set(bits, Foo.Bar))
+		fmt.printf("Get(Bar):           %v, %v\n", get(bits, Foo.Bar))
+		fmt.printf("Set(Negative_Test): %v\n",     set(bits, Foo.Negative_Test))
+		fmt.printf("Get(Leaves):        %v, %v\n", get(bits, Foo.Leaves))
+		fmt.printf("Get(Negative_Test): %v, %v\n", get(bits, Foo.Negative_Test))
+		fmt.printf("Freed.\n")
+	}
+*/
+package container_dynamic_bit_array
@@ -0,0 +1,46 @@
+/*
+Package list implements an intrusive doubly-linked list.
+
+An intrusive container requires a `Node` to be embedded in your own structure, like this:
+
+	My_String :: struct {
+		node:  list.Node,
+		value: string,
+	}
+
+Embedding the members of a `list.Node` in your structure with the `using` keyword is also allowed:
+
+	My_String :: struct {
+		using node: list.Node,
+		value: string,
+	}
+
+Here is a full example:
+
+	package test
+	
+	import "core:fmt"
+	import "core:container/intrusive/list"
+	
+	main :: proc() {
+	    l: list.List
+	
+	    one := My_String{value="Hello"}
+	    two := My_String{value="World"}
+	
+	    list.push_back(&l, &one.node)
+	    list.push_back(&l, &two.node)
+	
+	    iter := list.iterator_head(l, My_String, "node")
+	    for s in list.iterate_next(&iter) {
+	        fmt.println(s.value)
+	    }
+	}
+	
+	My_String :: struct {
+	    node:  list.Node,
+	    value: string,
+	}
+
+*/
+package container_intrusive_list
@@ -18,11 +18,18 @@ List :: struct {
 	tail: ^Node,
 }

-
+// The list link you must include in your own structure.
 Node :: struct {
 	prev, next: ^Node,
 }

+/*
+Inserts a new element at the front of the list with O(1) time complexity.
+
+**Inputs**
+- list: The container list
+- node: The node member of the user-defined element structure
+*/
 push_front :: proc "contextless" (list: ^List, node: ^Node) {
 	if list.head != nil {
 		list.head.prev = node
@@ -33,7 +40,13 @@ push_front :: proc "contextless" (list: ^List, node: ^Node) {
 		node.prev, node.next = nil, nil
 	}
 }
+/*
+Inserts a new element at the back of the list with O(1) time complexity.

+**Inputs**
+- list: The container list
+- node: The node member of the user-defined element structure
+*/
 push_back :: proc "contextless" (list: ^List, node: ^Node) {
 	if list.tail != nil {
 		list.tail.next = node
@@ -45,6 +58,13 @@ push_back :: proc "contextless" (list: ^List, node: ^Node) {
 	}
 }

+/*
+Removes an element from a list with O(1) time complexity.
+
+**Inputs**
+- list: The container list
+- node: The node member of the user-defined element structure to be removed
+*/
 remove :: proc "contextless" (list: ^List, node: ^Node) {
 	if node != nil {
 		if node.next != nil {
@@ -61,7 +81,13 @@ remove :: proc "contextless" (list: ^List, node: ^Node) {
 		}
 	}
 }
+/*
+Removes from the given list all elements that satisfy a condition with O(N) time complexity.

+**Inputs**
+- list: The container list
+- to_erase: The condition procedure. It should return `true` if a node should be removed, `false` otherwise
+*/
 remove_by_proc :: proc(list: ^List, to_erase: proc(^Node) -> bool) {
 	for node := list.head; node != nil; {
 		next := node.next
@@ -82,7 +108,13 @@ remove_by_proc :: proc(list: ^List, to_erase: proc(^Node) -> bool) {
 		node = next
 	}
 }
+/*
+Removes from the given list all elements that satisfy a condition with O(N) time complexity.

+**Inputs**
+- list: The container list
+- to_erase: The _contextless_ condition procedure. It should return `true` if a node should be removed, `false` otherwise
+*/
 remove_by_proc_contextless :: proc(list: ^List, to_erase: proc "contextless" (^Node) -> bool) {
 	for node := list.head; node != nil; {
 		next := node.next
@@ -104,12 +136,26 @@ remove_by_proc_contextless :: proc(list: ^List, to_erase: proc "contextless" (^N
 	}
 }

+/*
+Checks whether the given list does not contain any element.

+**Inputs**
+- list: The container list

+**Returns** `true` if `list` is empty, `false` otherwise
+*/
 is_empty :: proc "contextless" (list: ^List) -> bool {
 	return list.head == nil
 }

+/*
+Removes and returns the element at the front of the list with O(1) time complexity.
+
+**Inputs**
+- list: The container list
+
+**Returns** The node member of the user-defined element structure, or `nil` if the list is empty
+*/
 pop_front :: proc "contextless" (list: ^List) -> ^Node {
 	link := list.head
 	if link == nil {
@@ -130,6 +176,14 @@ pop_front :: proc "contextless" (list: ^List) -> ^Node {
 	return link

 }
+/*
+Removes and returns the element at the back of the list with O(1) time complexity.
+
+**Inputs**
+- list: The container list
+
+**Returns** The node member of the user-defined element structure, or `nil` if the list is empty
+*/
 pop_back :: proc "contextless" (list: ^List) -> ^Node {
 	link := list.tail
 	if link == nil {
@@ -151,29 +205,102 @@ pop_back :: proc "contextless" (list: ^List) -> ^Node {
 }


+
 Iterator :: struct($T: typeid) {
 	curr:   ^Node,
 	offset: uintptr,
 }

+/*
+Creates an iterator pointing at the head of the given list. For an example, see `iterate_next`.
+
+**Inputs**
+- list: The container list
+- T: The type of the list's elements
+- field_name: The name of the node field in the `T` structure
+
+**Returns** An iterator pointing at the head of `list`
+
+*/
 iterator_head :: proc "contextless" (list: List, $T: typeid, $field_name: string) -> Iterator(T)
 	where intrinsics.type_has_field(T, field_name),
 	      intrinsics.type_field_type(T, field_name) == Node {
 	return {list.head, offset_of_by_string(T, field_name)}
 }
+/*
+Creates an iterator pointing at the tail of the given list. For an example, see `iterate_prev`.

+**Inputs**
+- list: The container list
+- T: The type of the list's elements
+- field_name: The name of the node field in the `T` structure
+
+**Returns** An iterator pointing at the tail of `list`
+
+*/
 iterator_tail :: proc "contextless" (list: List, $T: typeid, $field_name: string) -> Iterator(T)
 	where intrinsics.type_has_field(T, field_name),
 	      intrinsics.type_field_type(T, field_name) == Node {
 	return {list.tail, offset_of_by_string(T, field_name)}
 }
+/*
+Creates an iterator pointing at the specified node of a list.

+**Inputs**
+- node: a list node
+- T: The type of the list's elements
+- field_name: The name of the node field in the `T` structure
+
+**Returns** An iterator pointing at `node`
+
+*/
 iterator_from_node :: proc "contextless" (node: ^Node, $T: typeid, $field_name: string) -> Iterator(T)
 	where intrinsics.type_has_field(T, field_name),
 	      intrinsics.type_field_type(T, field_name) == Node {
 	return {node, offset_of_by_string(T, field_name)}
 }

+/*
+Retrieves the next element in a list and advances the iterator.
+
+**Inputs**  
+- it: The iterator
+
+**Returns**
+- ptr: The next list element
+- ok: `true` if the element is valid (the iterator could advance), `false` otherwise
+
+Example:
+
+	import "core:fmt"
+	import "core:container/intrusive/list"
+
+	iterate_next_example :: proc() {
+		l: list.List
+
+		one := My_Struct{value=1}
+		two := My_Struct{value=2}
+
+		list.push_back(&l, &one.node)
+		list.push_back(&l, &two.node)
+
+		it := list.iterator_head(l, My_Struct, "node")
+		for num in list.iterate_next(&it) {
+			fmt.println(num.value)
+		}
+	}
+
+	My_Struct :: struct {
+		node : list.Node,
+		value: int,
+	}
+
+Output:
+
+	1
+	2
+
+*/
 iterate_next :: proc "contextless" (it: ^Iterator($T)) -> (ptr: ^T, ok: bool) {
 	node := it.curr
 	if node == nil {
@@ -183,7 +310,47 @@ iterate_next :: proc "contextless" (it: ^Iterator($T)) -> (ptr: ^T, ok: bool) {

 	return (^T)(uintptr(node) - it.offset), true
 }
+/*
+Retrieves the previous element in a list and recede the iterator.

+**Inputs**  
+- it: The iterator
+
+**Returns**
+- ptr: The previous list element
+- ok: `true` if the element is valid (the iterator could recede), `false` otherwise
+
+Example:
+
+	import "core:fmt"
+	import "core:container/intrusive/list"
+
+	iterate_next_example :: proc() {
+		l: list.List
+
+		one := My_Struct{value=1}
+		two := My_Struct{value=2}
+
+		list.push_back(&l, &one.node)
+		list.push_back(&l, &two.node)
+
+		it := list.iterator_tail(l, My_Struct, "node")
+		for num in list.iterate_prev(&it) {
+			fmt.println(num.value)
+		}
+	}
+
+	My_Struct :: struct {
+		node : list.Node,
+		value: int,
+	}
+
+Output:
+
+	2
+	1
+
+*/
 iterate_prev :: proc "contextless" (it: ^Iterator($T)) -> (ptr: ^T, ok: bool) {
 	node := it.curr
 	if node == nil {
@@ -192,4 +359,4 @@ iterate_prev :: proc "contextless" (it: ^Iterator($T)) -> (ptr: ^T, ok: bool) {
 	it.curr = node.prev

 	return (^T)(uintptr(node) - it.offset), true
-}
+}
@@ -70,8 +70,7 @@ set :: proc(c: ^$C/Cache($Key, $Value), key: Key, value: Value) -> runtime.Alloc
 	if c.count == c.capacity {
 		e = c.tail
 		_remove_node(c, e)
-	}
-	else {
+	} else {
 		c.count += 1
 		e = new(Node(Key, Value), c.node_allocator) or_return
 	}
@@ -95,11 +95,11 @@ front_ptr :: proc(q: ^$Q/Queue($T)) -> ^T {
 }

 back :: proc(q: ^$Q/Queue($T)) -> T {
-	idx := (q.offset+uint(q.len))%builtin.len(q.data)
+	idx := (q.offset+uint(q.len - 1))%builtin.len(q.data)
 	return q.data[idx]
 }
 back_ptr :: proc(q: ^$Q/Queue($T)) -> ^T {
-	idx := (q.offset+uint(q.len))%builtin.len(q.data)
+	idx := (q.offset+uint(q.len - 1))%builtin.len(q.data)
 	return &q.data[idx]
 }

@@ -189,7 +189,7 @@ pop_front_safe :: proc(q: ^$Q/Queue($T)) -> (elem: T, ok: bool) {
 	return
 }

-// Push multiple elements to the front of the queue
+// Push multiple elements to the back of the queue
 push_back_elems :: proc(q: ^$Q/Queue($T), elems: ..T) -> (ok: bool, err: runtime.Allocator_Error)  {
 	n := uint(builtin.len(elems))
 	if space(q^) < int(n) {
@@ -241,7 +241,7 @@ clear :: proc(q: ^$Q/Queue($T)) {
 }


-// Internal growinh procedure
+// Internal growing procedure
 _grow :: proc(q: ^$Q/Queue($T), min_capacity: uint = 0) -> runtime.Allocator_Error {
 	new_capacity := max(min_capacity, uint(8), uint(builtin.len(q.data))*2)
 	n := uint(builtin.len(q.data))
@@ -0,0 +1,568 @@
+// This package implements a red-black tree
+package container_rbtree
+
+@(require) import "base:intrinsics"
+@(require) import "base:runtime"
+import "core:slice"
+
+// Originally based on the CC0 implementation from literateprograms.org
+// But with API design mimicking `core:container/avl` for ease of use.
+
+// Direction specifies the traversal direction for a tree iterator.
+Direction :: enum i8 {
+	// Backward is the in-order backwards direction.
+	Backward = -1,
+	// Forward is the in-order forwards direction.
+	Forward  = 1,
+}
+
+Ordering :: slice.Ordering
+
+// Tree is a red-black tree
+Tree :: struct($Key: typeid, $Value: typeid) {
+	// user_data is a parameter that will be passed to the on_remove
+	// callback.
+	user_data: rawptr,
+	// on_remove is an optional callback that can be called immediately
+	// after a node is removed from the tree.
+	on_remove: proc(key: Key, value: Value, user_data: rawptr),
+
+	_root:           ^Node(Key, Value),
+	_node_allocator: runtime.Allocator,
+	_cmp_fn:         proc(Key, Key) -> Ordering,
+	_size:           int,
+}
+
+// Node is a red-black tree node.
+//
+// WARNING: It is unsafe to mutate value if the node is part of a tree
+// if doing so will alter the Node's sort position relative to other
+// elements in the tree.
+Node :: struct($Key: typeid, $Value: typeid) {
+	key:    Key,
+	value:  Value,
+
+	_parent: ^Node(Key, Value),
+	_left:   ^Node(Key, Value),
+	_right:  ^Node(Key, Value),
+	_color:  Color,
+}
+
+// Might store this in the node pointer in the future, but that'll require a decent amount of rework to pass ^^N instead of ^N
+Color :: enum uintptr {Black = 0, Red = 1}
+
+// Iterator is a tree iterator.
+//
+// WARNING: It is unsafe to modify the tree while iterating, except via
+// the iterator_remove method.
+Iterator :: struct($Key: typeid, $Value: typeid) {
+	_tree:        ^Tree(Key, Value),
+	_cur:         ^Node(Key, Value),
+	_next:        ^Node(Key, Value),
+	_direction:   Direction,
+	_called_next: bool,
+}
+
+// init initializes a tree.
+init :: proc {
+	init_ordered,
+	init_cmp,
+}
+
+// init_cmp initializes a tree.
+init_cmp :: proc(t: ^$T/Tree($Key, $Value), cmp_fn: proc(a, b: Key) -> Ordering, node_allocator := context.allocator) {
+	t._root   = nil
+	t._node_allocator = node_allocator
+	t._cmp_fn = cmp_fn
+	t._size = 0
+}
+
+// init_ordered initializes a tree containing ordered keys, with
+// a comparison function that results in an ascending order sort.
+init_ordered :: proc(t: ^$T/Tree($Key, $Value), node_allocator := context.allocator) where intrinsics.type_is_ordered(Key) {
+	init_cmp(t, slice.cmp_proc(Key), node_allocator)
+}
+
+// destroy de-initializes a tree.
+destroy :: proc(t: ^$T/Tree($Key, $Value), call_on_remove: bool = true) {
+	iter := iterator(t, .Forward)
+	for _ in iterator_next(&iter) {
+		iterator_remove(&iter, call_on_remove)
+	}
+}
+
+len :: proc "contextless" (t: ^$T/Tree($Key, $Value)) -> (node_count: int) {
+	return t._size
+}
+
+// first returns the first node in the tree (in-order) or nil iff
+// the tree is empty.
+first :: proc "contextless" (t: ^$T/Tree($Key, $Value)) -> ^Node(Key, Value) {
+	return tree_first_or_last_in_order(t, Direction.Backward)
+}
+
+// last returns the last element in the tree (in-order) or nil iff
+// the tree is empty.
+last :: proc "contextless" (t: ^$T/Tree($Key, $Value)) -> ^Node(Key, Value) {
+	return tree_first_or_last_in_order(t, Direction.Forward)
+}
+
+// find finds the key in the tree, and returns the corresponding node, or nil iff the value is not present.
+find :: proc(t: ^$T/Tree($Key, $Value), key: Key) -> (node: ^Node(Key, Value)) {
+	node = t._root
+	for node != nil {
+		switch t._cmp_fn(key, node.key) {
+		case .Equal:   return node
+		case .Less:    node = node._left
+		case .Greater: node = node._right
+		}
+	}
+	return node
+}
+
+// find_value finds the key in the tree, and returns the corresponding value, or nil iff the value is not present.
+find_value :: proc(t: ^$T/Tree($Key, $Value), key: Key) -> (value: Value, ok: bool) #optional_ok {
+	if n := find(t, key); n != nil {
+		return n.value, true
+	}
+	return
+}
+
+// find_or_insert attempts to insert the value into the tree, and returns
+// the node, a boolean indicating if the value was inserted, and the
+// node allocator error if relevant.  If the value is already present, the existing node is updated.
+find_or_insert :: proc(t: ^$T/Tree($Key, $Value), key: Key, value: Value) -> (n: ^Node(Key, Value), inserted: bool, err: runtime.Allocator_Error) {
+	n_ptr := &t._root
+	for n_ptr^ != nil {
+		n = n_ptr^
+		switch t._cmp_fn(key, n.key) {
+		case .Less:
+			n_ptr = &n._left
+		case .Greater:
+			n_ptr = &n._right
+		case .Equal:
+			return
+		}
+	}
+	_parent := n
+
+	n = new_clone(Node(Key, Value){key=key, value=value, _parent=_parent, _color=.Red}, t._node_allocator) or_return
+	n_ptr^ = n
+	insert_case1(t, n)
+	t._size += 1
+	return n, true, nil
+}
+
+// remove removes a node or value from the tree, and returns true iff the
+// removal was successful.  While the node's value will be left intact,
+// the node itself will be freed via the tree's node allocator.
+remove :: proc {
+	remove_key,
+	remove_node,
+}
+
+// remove_value removes a value from the tree, and returns true iff the
+// removal was successful.  While the node's key + value will be left intact,
+// the node itself will be freed via the tree's node allocator.
+remove_key :: proc(t: ^$T/Tree($Key, $Value), key: Key, call_on_remove := true) -> bool {
+	n := find(t, key)
+	if n == nil {
+		return false // Key not found, nothing to do
+	}
+	return remove_node(t, n, call_on_remove)
+}
+
+// remove_node removes a node from the tree, and returns true iff the
+// removal was successful.  While the node's key + value will be left intact,
+// the node itself will be freed via the tree's node allocator.
+remove_node :: proc(t: ^$T/Tree($Key, $Value), node: ^$N/Node(Key, Value), call_on_remove := true) -> (found: bool) {
+	if node._parent == node || (node._parent == nil && t._root != node) {
+		return false // Don't touch self-parented or dangling nodes.
+	}
+	node := node
+	if node._left != nil && node._right != nil {
+		// Copy key + value from predecessor and delete it instead
+		predecessor := maximum_node(node._left)
+		node.key   = predecessor.key
+		node.value = predecessor.value
+		node = predecessor
+	}
+
+	child := node._right == nil ? node._left : node._right
+	if node_color(node) == .Black {
+		node._color = node_color(child)
+		remove_case1(t, node)
+	}
+	replace_node(t, node, child)
+	if node._parent == nil && child != nil {
+		child._color = .Black // root should be black
+	}
+
+	if call_on_remove && t.on_remove != nil {
+		t.on_remove(node.key, node.value, t.user_data)
+	}
+	free(node, t._node_allocator)
+	t._size -= 1
+	return true
+}
+
+// iterator returns a tree iterator in the specified direction.
+iterator :: proc "contextless" (t: ^$T/Tree($Key, $Value), direction: Direction) -> Iterator(Key, Value) {
+	it: Iterator(Key, Value)
+	it._tree      = cast(^Tree(Key, Value))t
+	it._direction = direction
+
+	iterator_first(&it)
+
+	return it
+}
+
+// iterator_from_pos returns a tree iterator in the specified direction,
+// spanning the range [pos, last] (inclusive).
+iterator_from_pos :: proc "contextless" (t: ^$T/Tree($Key, $Value), pos: ^Node(Key, Value), direction: Direction) -> Iterator(Key, Value) {
+	it: Iterator(Key, Value)
+	it._tree        = transmute(^Tree(Key, Value))t
+	it._direction   = direction
+	it._next        = nil
+	it._called_next = false
+
+	if it._cur = pos; pos != nil {
+		it._next = node_next_or_prev_in_order(it._cur, it._direction)
+	}
+
+	return it
+}
+
+// iterator_get returns the node currently pointed to by the iterator,
+// or nil iff the node has been removed, the tree is empty, or the end
+// of the tree has been reached.
+iterator_get :: proc "contextless" (it: ^$I/Iterator($Key, $Value)) -> ^Node(Key, Value) {
+	return it._cur
+}
+
+// iterator_remove removes the node currently pointed to by the iterator,
+// and returns true iff the removal was successful.  Semantics are the
+// same as the Tree remove.
+iterator_remove :: proc(it: ^$I/Iterator($Key, $Value), call_on_remove: bool = true) -> bool {
+	if it._cur == nil {
+		return false
+	}
+
+	ok := remove_node(it._tree, it._cur , call_on_remove)
+	if ok {
+		it._cur = nil
+	}
+
+	return ok
+}
+
+// iterator_next advances the iterator and returns the (node, true) or
+// or (nil, false) iff the end of the tree has been reached.
+//
+// Note: The first call to iterator_next will return the first node instead
+// of advancing the iterator.
+iterator_next :: proc "contextless" (it: ^$I/Iterator($Key, $Value)) -> (^Node(Key, Value), bool) {
+	// This check is needed so that the first element gets returned from
+	// a brand-new iterator, and so that the somewhat contrived case where
+	// iterator_remove is called before the first call to iterator_next
+	// returns the correct value.
+	if !it._called_next {
+		it._called_next = true
+
+		// There can be the contrived case where iterator_remove is
+		// called before ever calling iterator_next, which needs to be
+		// handled as an actual call to next.
+		//
+		// If this happens it._cur will be nil, so only return the
+		// first value, if it._cur is valid.
+		if it._cur != nil {
+			return it._cur, true
+		}
+	}
+
+	if it._next == nil {
+		return nil, false
+	}
+
+	it._cur = it._next
+	it._next = node_next_or_prev_in_order(it._cur, it._direction)
+
+	return it._cur, true
+}
+
+@(private)
+tree_first_or_last_in_order :: proc "contextless" (t: ^$T/Tree($Key, $Value), direction: Direction) -> ^Node(Key, Value) {
+	first, sign := t._root, i8(direction)
+	if first != nil {
+		for {
+			tmp := node_get_child(first, sign)
+			if tmp == nil {
+				break
+			}
+			first = tmp
+		}
+	}
+	return first
+}
+
+@(private)
+node_get_child :: #force_inline proc "contextless" (n: ^Node($Key, $Value), sign: i8) -> ^Node(Key, Value) {
+	if sign < 0 {
+		return n._left
+	}
+	return n._right
+}
+
+@(private)
+node_next_or_prev_in_order :: proc "contextless" (n: ^Node($Key, $Value), direction: Direction) -> ^Node(Key, Value) {
+	next, tmp: ^Node(Key, Value)
+	sign := i8(direction)
+
+	if next = node_get_child(n, +sign); next != nil {
+		for {
+			tmp = node_get_child(next, -sign)
+			if tmp == nil {
+				break
+			}
+			next = tmp
+		}
+	} else {
+		tmp, next = n, n._parent
+		for next != nil && tmp == node_get_child(next, +sign) {
+			tmp, next = next, next._parent
+		}
+	}
+	return next
+}
+
+@(private)
+iterator_first :: proc "contextless" (it: ^Iterator($Key, $Value)) {
+	// This is private because behavior when the user manually calls
+	// iterator_first followed by iterator_next is unintuitive, since
+	// the first call to iterator_next MUST return the first node
+	// instead of advancing so that `for node in iterator_next(&next)`
+	// works as expected.
+
+	switch it._direction {
+	case .Forward:
+		it._cur = tree_first_or_last_in_order(it._tree, .Backward)
+	case .Backward:
+		it._cur = tree_first_or_last_in_order(it._tree, .Forward)
+	}
+
+	it._next = nil
+	it._called_next = false
+
+	if it._cur != nil {
+		it._next = node_next_or_prev_in_order(it._cur, it._direction)
+	}
+}
+
+@(private)
+grand_parent :: proc(n: ^$N/Node($Key, $Value)) -> (g: ^N) {
+	return n._parent._parent
+}
+
+@(private)
+sibling :: proc(n: ^$N/Node($Key, $Value)) -> (s: ^N) {
+	if n == n._parent._left {
+		return n._parent._right
+		} else {
+			return n._parent._left
+		}
+}
+
+@(private)
+uncle :: proc(n: ^$N/Node($Key, $Value)) -> (u: ^N) {
+	return sibling(n._parent)
+}
+
+@(private)
+rotate__left :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	r := n._right
+	replace_node(t, n, r)
+	n._right = r._left
+	if r._left != nil {
+		r._left._parent = n
+	}
+	r._left   = n
+	n._parent = r
+}
+
+@(private)
+rotate__right :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	l := n._left
+	replace_node(t, n, l)
+	n._left = l._right
+	if l._right != nil {
+		l._right._parent = n
+	}
+	l._right  = n
+	n._parent = l
+}
+
+@(private)
+replace_node :: proc(t: ^$T/Tree($Key, $Value), old_n: ^$N/Node(Key, Value), new_n: ^N) {
+	if old_n._parent == nil {
+		t._root = new_n
+	} else {
+		if (old_n == old_n._parent._left) {
+			old_n._parent._left  = new_n
+		} else {
+			old_n._parent._right = new_n
+		}
+	}
+	if new_n != nil {
+		new_n._parent = old_n._parent
+	}
+}
+
+@(private)
+insert_case1 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if n._parent == nil {
+		n._color = .Black
+	} else {
+		insert_case2(t, n)
+	}
+}
+
+@(private)
+insert_case2 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if node_color(n._parent) == .Black {
+		return // Tree is still valid
+	} else {
+		insert_case3(t, n)
+	}
+}
+
+@(private)
+insert_case3 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if node_color(uncle(n)) == .Red {
+		n._parent._color       = .Black
+		uncle(n)._color       = .Black
+		grand_parent(n)._color = .Red
+		insert_case1(t, grand_parent(n))
+	} else {
+		insert_case4(t, n)
+	}
+}
+
+@(private)
+insert_case4 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	n := n
+	if n == n._parent._right && n._parent == grand_parent(n)._left {
+		rotate__left(t, n._parent)
+		n = n._left
+	} else if n == n._parent._left && n._parent == grand_parent(n)._right {
+		rotate__right(t, n._parent)
+		n = n._right
+	}
+	insert_case5(t, n)
+}
+
+@(private)
+insert_case5 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	n._parent._color = .Black
+	grand_parent(n)._color = .Red
+	if n == n._parent._left && n._parent == grand_parent(n)._left {
+		rotate__right(t, grand_parent(n))
+	} else {
+		rotate__left(t, grand_parent(n))
+	}
+}
+
+// The maximum_node() helper function just walks _right until it reaches the last non-leaf:
+@(private)
+maximum_node :: proc(n: ^$N/Node($Key, $Value)) -> (max_node: ^N) {
+	n := n
+	for n._right != nil {
+		n = n._right
+	}
+	return n
+}
+
+@(private)
+remove_case1 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if n._parent == nil {
+		return
+	} else {
+		remove_case2(t, n)
+	}
+}
+
+@(private)
+remove_case2 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if node_color(sibling(n)) == .Red {
+		n._parent._color = .Red
+		sibling(n)._color = .Black
+		if n == n._parent._left {
+			rotate__left(t, n._parent)
+		} else {
+			rotate__right(t, n._parent)
+		}
+	}
+	remove_case3(t, n)
+}
+
+@(private)
+remove_case3 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if node_color(n._parent) == .Black &&
+		node_color(sibling(n)) == .Black &&
+		node_color(sibling(n)._left) == .Black &&
+		node_color(sibling(n)._right) == .Black {
+			sibling(n)._color = .Red
+			remove_case1(t, n._parent)
+	} else {
+		remove_case4(t, n)
+	}
+}
+
+@(private)
+remove_case4 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if node_color(n._parent) == .Red &&
+		node_color(sibling(n)) == .Black &&
+		node_color(sibling(n)._left) == .Black &&
+		node_color(sibling(n)._right) == .Black {
+			sibling(n)._color = .Red
+			n._parent._color = .Black
+	} else {
+		remove_case5(t, n)
+	}
+}
+
+@(private)
+remove_case5 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	if n == n._parent._left &&
+		node_color(sibling(n)) == .Black &&
+		node_color(sibling(n)._left) == .Red &&
+		node_color(sibling(n)._right) == .Black {
+			sibling(n)._color = .Red
+			sibling(n)._left._color = .Black
+			rotate__right(t, sibling(n))
+	} else if n == n._parent._right &&
+		node_color(sibling(n)) == .Black &&
+		node_color(sibling(n)._right) == .Red &&
+		node_color(sibling(n)._left) == .Black {
+			sibling(n)._color = .Red
+			sibling(n)._right._color = .Black
+			rotate__left(t, sibling(n))
+	}
+	remove_case6(t, n)
+}
+
+@(private)
+remove_case6 :: proc(t: ^$T/Tree($Key, $Value), n: ^$N/Node(Key, Value)) {
+	sibling(n)._color = node_color(n._parent)
+	n._parent._color = .Black
+	if n == n._parent._left {
+		sibling(n)._right._color = .Black
+		rotate__left(t, n._parent)
+	} else {
+		sibling(n)._left._color = .Black
+		rotate__right(t, n._parent)
+	}
+}
+
+node_color :: proc(n: ^$N/Node($Key, $Value)) -> (c: Color) {
+	return n == nil ? .Black : n._color
+}
@@ -119,20 +119,20 @@ consume :: proc "odin" (a: ^$A/Small_Array($N, $T), count: int, loc := #caller_l
 }

 ordered_remove :: proc "contextless" (a: ^$A/Small_Array($N, $T), index: int, loc := #caller_location) #no_bounds_check {
-    runtime.bounds_check_error_loc(loc, index, a.len)
-    if index+1 < a.len {
+	runtime.bounds_check_error_loc(loc, index, a.len)
+	if index+1 < a.len {
 		copy(a.data[index:], a.data[index+1:])
 	}
 	a.len -= 1
 }

 unordered_remove :: proc "contextless" (a: ^$A/Small_Array($N, $T), index: int, loc := #caller_location) #no_bounds_check {
-    runtime.bounds_check_error_loc(loc, index, a.len)
+	runtime.bounds_check_error_loc(loc, index, a.len)
 	n := a.len-1
 	if index != n {
 		a.data[index] = a.data[n]
 	}
-    a.len -= 1
+	a.len -= 1
 }

 clear :: proc "contextless" (a: ^$A/Small_Array($N, $T)) {
@@ -61,7 +61,7 @@ add_dependency :: proc(sorter: ^$S/Sorter($K), key, dependency: K) -> bool {
 	}
 	find.dependents[key] = true

- 	find = &sorter.relations[key]
+	find = &sorter.relations[key]
 	if find == nil {
 		find = map_insert(&sorter.relations, key, make_relations(sorter))
 	}
@@ -0,0 +1,28 @@
+package _aes
+
+// KEY_SIZE_128 is the AES-128 key size in bytes.
+KEY_SIZE_128 :: 16
+// KEY_SIZE_192 is the AES-192 key size in bytes.
+KEY_SIZE_192 :: 24
+// KEY_SIZE_256 is the AES-256 key size in bytes.
+KEY_SIZE_256 :: 32
+
+// BLOCK_SIZE is the AES block size in bytes.
+BLOCK_SIZE :: 16
+
+// ROUNDS_128 is the number of rounds for AES-128.
+ROUNDS_128 :: 10
+// ROUNDS_192 is the number of rounds for AES-192.
+ROUNDS_192 :: 12
+// ROUNDS_256 is the number of rounds for AES-256.
+ROUNDS_256 :: 14
+
+// GHASH_KEY_SIZE is the GHASH key size in bytes.
+GHASH_KEY_SIZE :: 16
+// GHASH_BLOCK_SIZE is the GHASH block size in bytes.
+GHASH_BLOCK_SIZE :: 16
+// GHASH_TAG_SIZE is the GHASH tag size in bytes.
+GHASH_TAG_SIZE :: 16
+
+// RCON is the AES keyschedule round constants.
+RCON := [10]byte{0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x1B, 0x36}
@@ -0,0 +1,86 @@
+package aes_ct64
+
+import "base:intrinsics"
+import "core:mem"
+
+STRIDE :: 4
+
+// Context is a keyed AES (ECB) instance.
+Context :: struct {
+	_sk_exp:     [120]u64,
+	_num_rounds: int,
+}
+
+// init initializes a context for AES with the provided key.
+init :: proc(ctx: ^Context, key: []byte) {
+	skey: [30]u64 = ---
+
+	ctx._num_rounds = keysched(skey[:], key)
+	skey_expand(ctx._sk_exp[:], skey[:], ctx._num_rounds)
+}
+
+// encrypt_block sets `dst` to `AES-ECB-Encrypt(src)`.
+encrypt_block :: proc(ctx: ^Context, dst, src: []byte) {
+	q: [8]u64
+	load_blockx1(&q, src)
+	_encrypt(&q, ctx._sk_exp[:], ctx._num_rounds)
+	store_blockx1(dst, &q)
+}
+
+// encrypt_block sets `dst` to `AES-ECB-Decrypt(src)`.
+decrypt_block :: proc(ctx: ^Context, dst, src: []byte) {
+	q: [8]u64
+	load_blockx1(&q, src)
+	_decrypt(&q, ctx._sk_exp[:], ctx._num_rounds)
+	store_blockx1(dst, &q)
+}
+
+// encrypt_blocks sets `dst` to `AES-ECB-Encrypt(src[0], .. src[n])`.
+encrypt_blocks :: proc(ctx: ^Context, dst, src: [][]byte) {
+	q: [8]u64 = ---
+	src, dst := src, dst
+
+	n := len(src)
+	for n > 4 {
+		load_blocks(&q, src[0:4])
+		_encrypt(&q, ctx._sk_exp[:], ctx._num_rounds)
+		store_blocks(dst[0:4], &q)
+
+		src = src[4:]
+		dst = dst[4:]
+		n -= 4
+	}
+	if n > 0 {
+		load_blocks(&q, src)
+		_encrypt(&q, ctx._sk_exp[:], ctx._num_rounds)
+		store_blocks(dst, &q)
+	}
+}
+
+// decrypt_blocks sets dst to `AES-ECB-Decrypt(src[0], .. src[n])`.
+decrypt_blocks :: proc(ctx: ^Context, dst, src: [][]byte) {
+	q: [8]u64 = ---
+	src, dst := src, dst
+
+	n := len(src)
+	for n > 4 {
+		load_blocks(&q, src[0:4])
+		_decrypt(&q, ctx._sk_exp[:], ctx._num_rounds)
+		store_blocks(dst[0:4], &q)
+
+		src = src[4:]
+		dst = dst[4:]
+		n -= 4
+	}
+	if n > 0 {
+		load_blocks(&q, src)
+		_decrypt(&q, ctx._sk_exp[:], ctx._num_rounds)
+		store_blocks(dst, &q)
+	}
+}
+
+// reset sanitizes the Context.  The Context must be re-initialized to
+// be used again.
+reset :: proc(ctx: ^Context) {
+	mem.zero_explicit(ctx, size_of(ctx))
+}
@@ -0,0 +1,265 @@
+// Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//   1. Redistributions of source code must retain the above copyright
+//      notice, this list of conditions and the following disclaimer.
+//
+// THIS SOFTWARE IS PROVIDED BY THE AUTHORS “AS IS” AND ANY EXPRESS OR
+// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY
+// DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package aes_ct64
+
+import "base:intrinsics"
+
+// Bitsliced AES for 64-bit general purpose (integer) registers.  Each
+// invocation will process up to 4 blocks at a time.  This implementation
+// is derived from the BearSSL ct64 code, and distributed under a 1-clause
+// BSD license with permission from the original author.
+//
+// WARNING: "hic sunt dracones"
+//
+// This package also deliberately exposes enough internals to be able to
+// function as a replacement for `AESENC` and `AESDEC` from AES-NI, to
+// allow the implementation of non-AES primitives that use the AES round
+// function such as AEGIS and Deoxys-II.  This should ONLY be done when
+// implementing something other than AES itself.
+
+sub_bytes :: proc "contextless" (q: ^[8]u64) {
+	// This S-box implementation is a straightforward translation of
+	// the circuit described by Boyar and Peralta in "A new
+	// combinational logic minimization technique with applications
+	// to cryptology" (https://eprint.iacr.org/2009/191.pdf).
+	//
+	// Note that variables x* (input) and s* (output) are numbered
+	// in "reverse" order (x0 is the high bit, x7 is the low bit).
+
+	x0 := q[7]
+	x1 := q[6]
+	x2 := q[5]
+	x3 := q[4]
+	x4 := q[3]
+	x5 := q[2]
+	x6 := q[1]
+	x7 := q[0]
+
+	// Top linear transformation.
+	y14 := x3 ~ x5
+	y13 := x0 ~ x6
+	y9 := x0 ~ x3
+	y8 := x0 ~ x5
+	t0 := x1 ~ x2
+	y1 := t0 ~ x7
+	y4 := y1 ~ x3
+	y12 := y13 ~ y14
+	y2 := y1 ~ x0
+	y5 := y1 ~ x6
+	y3 := y5 ~ y8
+	t1 := x4 ~ y12
+	y15 := t1 ~ x5
+	y20 := t1 ~ x1
+	y6 := y15 ~ x7
+	y10 := y15 ~ t0
+	y11 := y20 ~ y9
+	y7 := x7 ~ y11
+	y17 := y10 ~ y11
+	y19 := y10 ~ y8
+	y16 := t0 ~ y11
+	y21 := y13 ~ y16
+	y18 := x0 ~ y16
+
+	// Non-linear section.
+	t2 := y12 & y15
+	t3 := y3 & y6
+	t4 := t3 ~ t2
+	t5 := y4 & x7
+	t6 := t5 ~ t2
+	t7 := y13 & y16
+	t8 := y5 & y1
+	t9 := t8 ~ t7
+	t10 := y2 & y7
+	t11 := t10 ~ t7
+	t12 := y9 & y11
+	t13 := y14 & y17
+	t14 := t13 ~ t12
+	t15 := y8 & y10
+	t16 := t15 ~ t12
+	t17 := t4 ~ t14
+	t18 := t6 ~ t16
+	t19 := t9 ~ t14
+	t20 := t11 ~ t16
+	t21 := t17 ~ y20
+	t22 := t18 ~ y19
+	t23 := t19 ~ y21
+	t24 := t20 ~ y18
+
+	t25 := t21 ~ t22
+	t26 := t21 & t23
+	t27 := t24 ~ t26
+	t28 := t25 & t27
+	t29 := t28 ~ t22
+	t30 := t23 ~ t24
+	t31 := t22 ~ t26
+	t32 := t31 & t30
+	t33 := t32 ~ t24
+	t34 := t23 ~ t33
+	t35 := t27 ~ t33
+	t36 := t24 & t35
+	t37 := t36 ~ t34
+	t38 := t27 ~ t36
+	t39 := t29 & t38
+	t40 := t25 ~ t39
+
+	t41 := t40 ~ t37
+	t42 := t29 ~ t33
+	t43 := t29 ~ t40
+	t44 := t33 ~ t37
+	t45 := t42 ~ t41
+	z0 := t44 & y15
+	z1 := t37 & y6
+	z2 := t33 & x7
+	z3 := t43 & y16
+	z4 := t40 & y1
+	z5 := t29 & y7
+	z6 := t42 & y11
+	z7 := t45 & y17
+	z8 := t41 & y10
+	z9 := t44 & y12
+	z10 := t37 & y3
+	z11 := t33 & y4
+	z12 := t43 & y13
+	z13 := t40 & y5
+	z14 := t29 & y2
+	z15 := t42 & y9
+	z16 := t45 & y14
+	z17 := t41 & y8
+
+	// Bottom linear transformation.
+	t46 := z15 ~ z16
+	t47 := z10 ~ z11
+	t48 := z5 ~ z13
+	t49 := z9 ~ z10
+	t50 := z2 ~ z12
+	t51 := z2 ~ z5
+	t52 := z7 ~ z8
+	t53 := z0 ~ z3
+	t54 := z6 ~ z7
+	t55 := z16 ~ z17
+	t56 := z12 ~ t48
+	t57 := t50 ~ t53
+	t58 := z4 ~ t46
+	t59 := z3 ~ t54
+	t60 := t46 ~ t57
+	t61 := z14 ~ t57
+	t62 := t52 ~ t58
+	t63 := t49 ~ t58
+	t64 := z4 ~ t59
+	t65 := t61 ~ t62
+	t66 := z1 ~ t63
+	s0 := t59 ~ t63
+	s6 := t56 ~ ~t62
+	s7 := t48 ~ ~t60
+	t67 := t64 ~ t65
+	s3 := t53 ~ t66
+	s4 := t51 ~ t66
+	s5 := t47 ~ t65
+	s1 := t64 ~ ~s3
+	s2 := t55 ~ ~t67
+
+	q[7] = s0
+	q[6] = s1
+	q[5] = s2
+	q[4] = s3
+	q[3] = s4
+	q[2] = s5
+	q[1] = s6
+	q[0] = s7
+}
+
+orthogonalize :: proc "contextless" (q: ^[8]u64) {
+	CL2 :: 0x5555555555555555
+	CH2 :: 0xAAAAAAAAAAAAAAAA
+	q[0], q[1] = (q[0] & CL2) | ((q[1] & CL2) << 1), ((q[0] & CH2) >> 1) | (q[1] & CH2)
+	q[2], q[3] = (q[2] & CL2) | ((q[3] & CL2) << 1), ((q[2] & CH2) >> 1) | (q[3] & CH2)
+	q[4], q[5] = (q[4] & CL2) | ((q[5] & CL2) << 1), ((q[4] & CH2) >> 1) | (q[5] & CH2)
+	q[6], q[7] = (q[6] & CL2) | ((q[7] & CL2) << 1), ((q[6] & CH2) >> 1) | (q[7] & CH2)
+
+	CL4 :: 0x3333333333333333
+	CH4 :: 0xCCCCCCCCCCCCCCCC
+	q[0], q[2] = (q[0] & CL4) | ((q[2] & CL4) << 2), ((q[0] & CH4) >> 2) | (q[2] & CH4)
+	q[1], q[3] = (q[1] & CL4) | ((q[3] & CL4) << 2), ((q[1] & CH4) >> 2) | (q[3] & CH4)
+	q[4], q[6] = (q[4] & CL4) | ((q[6] & CL4) << 2), ((q[4] & CH4) >> 2) | (q[6] & CH4)
+	q[5], q[7] = (q[5] & CL4) | ((q[7] & CL4) << 2), ((q[5] & CH4) >> 2) | (q[7] & CH4)
+
+	CL8 :: 0x0F0F0F0F0F0F0F0F
+	CH8 :: 0xF0F0F0F0F0F0F0F0
+	q[0], q[4] = (q[0] & CL8) | ((q[4] & CL8) << 4), ((q[0] & CH8) >> 4) | (q[4] & CH8)
+	q[1], q[5] = (q[1] & CL8) | ((q[5] & CL8) << 4), ((q[1] & CH8) >> 4) | (q[5] & CH8)
+	q[2], q[6] = (q[2] & CL8) | ((q[6] & CL8) << 4), ((q[2] & CH8) >> 4) | (q[6] & CH8)
+	q[3], q[7] = (q[3] & CL8) | ((q[7] & CL8) << 4), ((q[3] & CH8) >> 4) | (q[7] & CH8)
+}
+
+@(require_results)
+interleave_in :: proc "contextless" (w: []u32) -> (q0, q1: u64) #no_bounds_check {
+	if len(w) < 4 {
+		intrinsics.trap()
+	}
+	x0, x1, x2, x3 := u64(w[0]), u64(w[1]), u64(w[2]), u64(w[3])
+	x0 |= (x0 << 16)
+	x1 |= (x1 << 16)
+	x2 |= (x2 << 16)
+	x3 |= (x3 << 16)
+	x0 &= 0x0000FFFF0000FFFF
+	x1 &= 0x0000FFFF0000FFFF
+	x2 &= 0x0000FFFF0000FFFF
+	x3 &= 0x0000FFFF0000FFFF
+	x0 |= (x0 << 8)
+	x1 |= (x1 << 8)
+	x2 |= (x2 << 8)
+	x3 |= (x3 << 8)
+	x0 &= 0x00FF00FF00FF00FF
+	x1 &= 0x00FF00FF00FF00FF
+	x2 &= 0x00FF00FF00FF00FF
+	x3 &= 0x00FF00FF00FF00FF
+	q0 = x0 | (x2 << 8)
+	q1 = x1 | (x3 << 8)
+	return
+}
+
+@(require_results)
+interleave_out :: proc "contextless" (q0, q1: u64) -> (w0, w1, w2, w3: u32) {
+	x0 := q0 & 0x00FF00FF00FF00FF
+	x1 := q1 & 0x00FF00FF00FF00FF
+	x2 := (q0 >> 8) & 0x00FF00FF00FF00FF
+	x3 := (q1 >> 8) & 0x00FF00FF00FF00FF
+	x0 |= (x0 >> 8)
+	x1 |= (x1 >> 8)
+	x2 |= (x2 >> 8)
+	x3 |= (x3 >> 8)
+	x0 &= 0x0000FFFF0000FFFF
+	x1 &= 0x0000FFFF0000FFFF
+	x2 &= 0x0000FFFF0000FFFF
+	x3 &= 0x0000FFFF0000FFFF
+	w0 = u32(x0) | u32(x0 >> 16)
+	w1 = u32(x1) | u32(x1 >> 16)
+	w2 = u32(x2) | u32(x2 >> 16)
+	w3 = u32(x3) | u32(x3 >> 16)
+	return
+}
+
+@(private)
+rotr32 :: #force_inline proc "contextless" (x: u64) -> u64 {
+	return (x << 32) | (x >> 32)
+}
@@ -0,0 +1,135 @@
+// Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//   1. Redistributions of source code must retain the above copyright
+//      notice, this list of conditions and the following disclaimer.
+//
+// THIS SOFTWARE IS PROVIDED BY THE AUTHORS “AS IS” AND ANY EXPRESS OR
+// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY
+// DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package aes_ct64
+
+import "base:intrinsics"
+
+inv_sub_bytes :: proc "contextless" (q: ^[8]u64) {
+	// AES S-box is:
+	//   S(x) = A(I(x)) ^ 0x63
+	// where I() is inversion in GF(256), and A() is a linear
+	// transform (0 is formally defined to be its own inverse).
+	// Since inversion is an involution, the inverse S-box can be
+	// computed from the S-box as:
+	//   iS(x) = B(S(B(x ^ 0x63)) ^ 0x63)
+	// where B() is the inverse of A(). Indeed, for any y in GF(256):
+	//   iS(S(y)) = B(A(I(B(A(I(y)) ^ 0x63 ^ 0x63))) ^ 0x63 ^ 0x63) = y
+	//
+	// Note: we reuse the implementation of the forward S-box,
+	// instead of duplicating it here, so that total code size is
+	// lower. By merging the B() transforms into the S-box circuit
+	// we could make faster CBC decryption, but CBC decryption is
+	// already quite faster than CBC encryption because we can
+	// process four blocks in parallel.
+
+	q0 := ~q[0]
+	q1 := ~q[1]
+	q2 := q[2]
+	q3 := q[3]
+	q4 := q[4]
+	q5 := ~q[5]
+	q6 := ~q[6]
+	q7 := q[7]
+	q[7] = q1 ~ q4 ~ q6
+	q[6] = q0 ~ q3 ~ q5
+	q[5] = q7 ~ q2 ~ q4
+	q[4] = q6 ~ q1 ~ q3
+	q[3] = q5 ~ q0 ~ q2
+	q[2] = q4 ~ q7 ~ q1
+	q[1] = q3 ~ q6 ~ q0
+	q[0] = q2 ~ q5 ~ q7
+
+	sub_bytes(q)
+
+	q0 = ~q[0]
+	q1 = ~q[1]
+	q2 = q[2]
+	q3 = q[3]
+	q4 = q[4]
+	q5 = ~q[5]
+	q6 = ~q[6]
+	q7 = q[7]
+	q[7] = q1 ~ q4 ~ q6
+	q[6] = q0 ~ q3 ~ q5
+	q[5] = q7 ~ q2 ~ q4
+	q[4] = q6 ~ q1 ~ q3
+	q[3] = q5 ~ q0 ~ q2
+	q[2] = q4 ~ q7 ~ q1
+	q[1] = q3 ~ q6 ~ q0
+	q[0] = q2 ~ q5 ~ q7
+}
+
+inv_shift_rows :: proc "contextless" (q: ^[8]u64) {
+	for x, i in q {
+		q[i] =
+			(x & 0x000000000000FFFF) |
+			((x & 0x000000000FFF0000) << 4) |
+			((x & 0x00000000F0000000) >> 12) |
+			((x & 0x000000FF00000000) << 8) |
+			((x & 0x0000FF0000000000) >> 8) |
+			((x & 0x000F000000000000) << 12) |
+			((x & 0xFFF0000000000000) >> 4)
+	}
+}
+
+inv_mix_columns :: proc "contextless" (q: ^[8]u64) {
+	q0 := q[0]
+	q1 := q[1]
+	q2 := q[2]
+	q3 := q[3]
+	q4 := q[4]
+	q5 := q[5]
+	q6 := q[6]
+	q7 := q[7]
+	r0 := (q0 >> 16) | (q0 << 48)
+	r1 := (q1 >> 16) | (q1 << 48)
+	r2 := (q2 >> 16) | (q2 << 48)
+	r3 := (q3 >> 16) | (q3 << 48)
+	r4 := (q4 >> 16) | (q4 << 48)
+	r5 := (q5 >> 16) | (q5 << 48)
+	r6 := (q6 >> 16) | (q6 << 48)
+	r7 := (q7 >> 16) | (q7 << 48)
+
+	q[0] = q5 ~ q6 ~ q7 ~ r0 ~ r5 ~ r7 ~ rotr32(q0 ~ q5 ~ q6 ~ r0 ~ r5)
+	q[1] = q0 ~ q5 ~ r0 ~ r1 ~ r5 ~ r6 ~ r7 ~ rotr32(q1 ~ q5 ~ q7 ~ r1 ~ r5 ~ r6)
+	q[2] = q0 ~ q1 ~ q6 ~ r1 ~ r2 ~ r6 ~ r7 ~ rotr32(q0 ~ q2 ~ q6 ~ r2 ~ r6 ~ r7)
+	q[3] = q0 ~ q1 ~ q2 ~ q5 ~ q6 ~ r0 ~ r2 ~ r3 ~ r5 ~ rotr32(q0 ~ q1 ~ q3 ~ q5 ~ q6 ~ q7 ~ r0 ~ r3 ~ r5 ~ r7)
+	q[4] = q1 ~ q2 ~ q3 ~ q5 ~ r1 ~ r3 ~ r4 ~ r5 ~ r6 ~ r7 ~ rotr32(q1 ~ q2 ~ q4 ~ q5 ~ q7 ~ r1 ~ r4 ~ r5 ~ r6)
+	q[5] = q2 ~ q3 ~ q4 ~ q6 ~ r2 ~ r4 ~ r5 ~ r6 ~ r7 ~ rotr32(q2 ~ q3 ~ q5 ~ q6 ~ r2 ~ r5 ~ r6 ~ r7)
+	q[6] = q3 ~ q4 ~ q5 ~ q7 ~ r3 ~ r5 ~ r6 ~ r7 ~ rotr32(q3 ~ q4 ~ q6 ~ q7 ~ r3 ~ r6 ~ r7)
+	q[7] = q4 ~ q5 ~ q6 ~ r4 ~ r6 ~ r7 ~ rotr32(q4 ~ q5 ~ q7 ~ r4 ~ r7)
+}
+
+@(private)
+_decrypt :: proc "contextless" (q: ^[8]u64, skey: []u64, num_rounds: int) {
+	add_round_key(q, skey[num_rounds << 3:])
+	for u := num_rounds - 1; u > 0; u -= 1 {
+		inv_shift_rows(q)
+		inv_sub_bytes(q)
+		add_round_key(q, skey[u << 3:])
+		inv_mix_columns(q)
+	}
+	inv_shift_rows(q)
+	inv_sub_bytes(q)
+	add_round_key(q, skey)
+}
@@ -0,0 +1,95 @@
+// Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//   1. Redistributions of source code must retain the above copyright
+//      notice, this list of conditions and the following disclaimer.
+//
+// THIS SOFTWARE IS PROVIDED BY THE AUTHORS “AS IS” AND ANY EXPRESS OR
+// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY
+// DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package aes_ct64
+
+import "base:intrinsics"
+
+add_round_key :: proc "contextless" (q: ^[8]u64, sk: []u64) #no_bounds_check {
+	if len(sk) < 8 {
+		intrinsics.trap()
+	}
+
+	q[0] ~= sk[0]
+	q[1] ~= sk[1]
+	q[2] ~= sk[2]
+	q[3] ~= sk[3]
+	q[4] ~= sk[4]
+	q[5] ~= sk[5]
+	q[6] ~= sk[6]
+	q[7] ~= sk[7]
+}
+
+shift_rows :: proc "contextless" (q: ^[8]u64) {
+	for x, i in q {
+		q[i] =
+			(x & 0x000000000000FFFF) |
+			((x & 0x00000000FFF00000) >> 4) |
+			((x & 0x00000000000F0000) << 12) |
+			((x & 0x0000FF0000000000) >> 8) |
+			((x & 0x000000FF00000000) << 8) |
+			((x & 0xF000000000000000) >> 12) |
+			((x & 0x0FFF000000000000) << 4)
+	}
+}
+
+mix_columns :: proc "contextless" (q: ^[8]u64) {
+	q0 := q[0]
+	q1 := q[1]
+	q2 := q[2]
+	q3 := q[3]
+	q4 := q[4]
+	q5 := q[5]
+	q6 := q[6]
+	q7 := q[7]
+	r0 := (q0 >> 16) | (q0 << 48)
+	r1 := (q1 >> 16) | (q1 << 48)
+	r2 := (q2 >> 16) | (q2 << 48)
+	r3 := (q3 >> 16) | (q3 << 48)
+	r4 := (q4 >> 16) | (q4 << 48)
+	r5 := (q5 >> 16) | (q5 << 48)
+	r6 := (q6 >> 16) | (q6 << 48)
+	r7 := (q7 >> 16) | (q7 << 48)
+
+	q[0] = q7 ~ r7 ~ r0 ~ rotr32(q0 ~ r0)
+	q[1] = q0 ~ r0 ~ q7 ~ r7 ~ r1 ~ rotr32(q1 ~ r1)
+	q[2] = q1 ~ r1 ~ r2 ~ rotr32(q2 ~ r2)
+	q[3] = q2 ~ r2 ~ q7 ~ r7 ~ r3 ~ rotr32(q3 ~ r3)
+	q[4] = q3 ~ r3 ~ q7 ~ r7 ~ r4 ~ rotr32(q4 ~ r4)
+	q[5] = q4 ~ r4 ~ r5 ~ rotr32(q5 ~ r5)
+	q[6] = q5 ~ r5 ~ r6 ~ rotr32(q6 ~ r6)
+	q[7] = q6 ~ r6 ~ r7 ~ rotr32(q7 ~ r7)
+}
+
+@(private)
+_encrypt :: proc "contextless" (q: ^[8]u64, skey: []u64, num_rounds: int) {
+	add_round_key(q, skey)
+	for u in 1 ..< num_rounds {
+		sub_bytes(q)
+		shift_rows(q)
+		mix_columns(q)
+		add_round_key(q, skey[u << 3:])
+	}
+	sub_bytes(q)
+	shift_rows(q)
+	add_round_key(q, skey[num_rounds << 3:])
+}
@@ -0,0 +1,179 @@
+// Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//   1. Redistributions of source code must retain the above copyright
+//      notice, this list of conditions and the following disclaimer.
+//
+// THIS SOFTWARE IS PROVIDED BY THE AUTHORS “AS IS” AND ANY EXPRESS OR
+// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY
+// DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package aes_ct64
+
+import "base:intrinsics"
+import "core:crypto/_aes"
+import "core:encoding/endian"
+import "core:mem"
+
+@(private, require_results)
+sub_word :: proc "contextless" (x: u32) -> u32 {
+	q := [8]u64{u64(x), 0, 0, 0, 0, 0, 0, 0}
+
+	orthogonalize(&q)
+	sub_bytes(&q)
+	orthogonalize(&q)
+	ret := u32(q[0])
+
+	mem.zero_explicit(&q[0], size_of(u64))
+
+	return ret
+}
+
+@(private, require_results)
+keysched :: proc(comp_skey: []u64, key: []byte) -> int {
+	num_rounds, key_len := 0, len(key)
+	switch key_len {
+	case _aes.KEY_SIZE_128:
+		num_rounds = _aes.ROUNDS_128
+	case _aes.KEY_SIZE_192:
+		num_rounds = _aes.ROUNDS_192
+	case _aes.KEY_SIZE_256:
+		num_rounds = _aes.ROUNDS_256
+	case:
+		panic("crypto/aes: invalid AES key size")
+	}
+
+	skey: [60]u32 = ---
+	nk, nkf := key_len >> 2, (num_rounds + 1) << 2
+	for i in 0 ..< nk {
+		skey[i] = endian.unchecked_get_u32le(key[i << 2:])
+	}
+	tmp := skey[(key_len >> 2) - 1]
+	for i, j, k := nk, 0, 0; i < nkf; i += 1 {
+		if j == 0 {
+			tmp = (tmp << 24) | (tmp >> 8)
+			tmp = sub_word(tmp) ~ u32(_aes.RCON[k])
+		} else if nk > 6 && j == 4 {
+			tmp = sub_word(tmp)
+		}
+		tmp ~= skey[i - nk]
+		skey[i] = tmp
+		if j += 1; j == nk {
+			j = 0
+			k += 1
+		}
+	}
+
+	q: [8]u64 = ---
+	for i, j := 0, 0; i < nkf; i, j = i + 4, j + 2 {
+		q[0], q[4] = interleave_in(skey[i:])
+		q[1] = q[0]
+		q[2] = q[0]
+		q[3] = q[0]
+		q[5] = q[4]
+		q[6] = q[4]
+		q[7] = q[4]
+		orthogonalize(&q)
+		comp_skey[j + 0] =
+			(q[0] & 0x1111111111111111) |
+			(q[1] & 0x2222222222222222) |
+			(q[2] & 0x4444444444444444) |
+			(q[3] & 0x8888888888888888)
+		comp_skey[j + 1] =
+			(q[4] & 0x1111111111111111) |
+			(q[5] & 0x2222222222222222) |
+			(q[6] & 0x4444444444444444) |
+			(q[7] & 0x8888888888888888)
+	}
+
+	mem.zero_explicit(&skey, size_of(skey))
+	mem.zero_explicit(&q, size_of(q))
+
+	return num_rounds
+}
+
+@(private)
+skey_expand :: proc "contextless" (skey, comp_skey: []u64, num_rounds: int) {
+	n := (num_rounds + 1) << 1
+	for u, v := 0, 0; u < n; u, v = u + 1, v + 4 {
+		x0 := comp_skey[u]
+		x1, x2, x3 := x0, x0, x0
+		x0 &= 0x1111111111111111
+		x1 &= 0x2222222222222222
+		x2 &= 0x4444444444444444
+		x3 &= 0x8888888888888888
+		x1 >>= 1
+		x2 >>= 2
+		x3 >>= 3
+		skey[v + 0] = (x0 << 4) - x0
+		skey[v + 1] = (x1 << 4) - x1
+		skey[v + 2] = (x2 << 4) - x2
+		skey[v + 3] = (x3 << 4) - x3
+	}
+}
+
+orthogonalize_roundkey :: proc "contextless" (qq: []u64, key: []byte) {
+	if len(qq) < 8 || len(key) != 16 {
+		intrinsics.trap()
+	}
+
+	skey: [4]u32 = ---
+	skey[0] = endian.unchecked_get_u32le(key[0:])
+	skey[1] = endian.unchecked_get_u32le(key[4:])
+	skey[2] = endian.unchecked_get_u32le(key[8:])
+	skey[3] = endian.unchecked_get_u32le(key[12:])
+
+	q: [8]u64 = ---
+	q[0], q[4] = interleave_in(skey[:])
+	q[1] = q[0]
+	q[2] = q[0]
+	q[3] = q[0]
+	q[5] = q[4]
+	q[6] = q[4]
+	q[7] = q[4]
+	orthogonalize(&q)
+
+	comp_skey: [2]u64 = ---
+	comp_skey[0] =
+		(q[0] & 0x1111111111111111) |
+		(q[1] & 0x2222222222222222) |
+		(q[2] & 0x4444444444444444) |
+		(q[3] & 0x8888888888888888)
+	comp_skey[1] =
+		(q[4] & 0x1111111111111111) |
+		(q[5] & 0x2222222222222222) |
+		(q[6] & 0x4444444444444444) |
+		(q[7] & 0x8888888888888888)
+
+	for x, u in comp_skey {
+		x0 := x
+		x1, x2, x3 := x0, x0, x0
+		x0 &= 0x1111111111111111
+		x1 &= 0x2222222222222222
+		x2 &= 0x4444444444444444
+		x3 &= 0x8888888888888888
+		x1 >>= 1
+		x2 >>= 2
+		x3 >>= 3
+		qq[u * 4 + 0] = (x0 << 4) - x0
+		qq[u * 4 + 1] = (x1 << 4) - x1
+		qq[u * 4 + 2] = (x2 << 4) - x2
+		qq[u * 4 + 3] = (x3 << 4) - x3
+	}
+
+	mem.zero_explicit(&skey, size_of(skey))
+	mem.zero_explicit(&q, size_of(q))
+	mem.zero_explicit(&comp_skey, size_of(comp_skey))
+}
@@ -0,0 +1,136 @@
+// Copyright (c) 2016 Thomas Pornin <pornin@bolet.org>
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//   1. Redistributions of source code must retain the above copyright
+//      notice, this list of conditions and the following disclaimer.
+//
+// THIS SOFTWARE IS PROVIDED BY THE AUTHORS “AS IS” AND ANY EXPRESS OR
+// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY
+// DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+package aes_ct64
+
+import "base:intrinsics"
+import "core:crypto/_aes"
+import "core:encoding/endian"
+
+@(private = "file")
+bmul64 :: proc "contextless" (x, y: u64) -> u64 {
+	x0 := x & 0x1111111111111111
+	x1 := x & 0x2222222222222222
+	x2 := x & 0x4444444444444444
+	x3 := x & 0x8888888888888888
+	y0 := y & 0x1111111111111111
+	y1 := y & 0x2222222222222222
+	y2 := y & 0x4444444444444444
+	y3 := y & 0x8888888888888888
+	z0 := (x0 * y0) ~ (x1 * y3) ~ (x2 * y2) ~ (x3 * y1)
+	z1 := (x0 * y1) ~ (x1 * y0) ~ (x2 * y3) ~ (x3 * y2)
+	z2 := (x0 * y2) ~ (x1 * y1) ~ (x2 * y0) ~ (x3 * y3)
+	z3 := (x0 * y3) ~ (x1 * y2) ~ (x2 * y1) ~ (x3 * y0)
+	z0 &= 0x1111111111111111
+	z1 &= 0x2222222222222222
+	z2 &= 0x4444444444444444
+	z3 &= 0x8888888888888888
+	return z0 | z1 | z2 | z3
+}
+
+@(private = "file")
+rev64 :: proc "contextless" (x: u64) -> u64 {
+	x := x
+	x = ((x & 0x5555555555555555) << 1) | ((x >> 1) & 0x5555555555555555)
+	x = ((x & 0x3333333333333333) << 2) | ((x >> 2) & 0x3333333333333333)
+	x = ((x & 0x0F0F0F0F0F0F0F0F) << 4) | ((x >> 4) & 0x0F0F0F0F0F0F0F0F)
+	x = ((x & 0x00FF00FF00FF00FF) << 8) | ((x >> 8) & 0x00FF00FF00FF00FF)
+	x = ((x & 0x0000FFFF0000FFFF) << 16) | ((x >> 16) & 0x0000FFFF0000FFFF)
+	return (x << 32) | (x >> 32)
+}
+
+// ghash calculates the GHASH of data, with the key `key`, and input `dst`
+// and `data`, and stores the resulting digest in `dst`.
+//
+// Note: `dst` is both an input and an output, to support easy implementation
+// of GCM.
+ghash :: proc "contextless" (dst, key, data: []byte) {
+	if len(dst) != _aes.GHASH_BLOCK_SIZE || len(key) != _aes.GHASH_BLOCK_SIZE {
+		intrinsics.trap()
+	}
+
+	buf := data
+	l := len(buf)
+
+	y1 := endian.unchecked_get_u64be(dst[0:])
+	y0 := endian.unchecked_get_u64be(dst[8:])
+	h1 := endian.unchecked_get_u64be(key[0:])
+	h0 := endian.unchecked_get_u64be(key[8:])
+	h0r := rev64(h0)
+	h1r := rev64(h1)
+	h2 := h0 ~ h1
+	h2r := h0r ~ h1r
+
+	for l > 0 {
+		src: []byte = ---
+		if l >= _aes.GHASH_BLOCK_SIZE {
+			src = buf
+			buf = buf[_aes.GHASH_BLOCK_SIZE:]
+			l -= _aes.GHASH_BLOCK_SIZE
+		} else {
+			tmp: [_aes.GHASH_BLOCK_SIZE]byte
+			copy(tmp[:], buf)
+			src = tmp[:]
+			l = 0
+		}
+		y1 ~= endian.unchecked_get_u64be(src)
+		y0 ~= endian.unchecked_get_u64be(src[8:])
+
+		y0r := rev64(y0)
+		y1r := rev64(y1)
+		y2 := y0 ~ y1
+		y2r := y0r ~ y1r
+
+		z0 := bmul64(y0, h0)
+		z1 := bmul64(y1, h1)
+		z2 := bmul64(y2, h2)
+		z0h := bmul64(y0r, h0r)
+		z1h := bmul64(y1r, h1r)
+		z2h := bmul64(y2r, h2r)
+		z2 ~= z0 ~ z1
+		z2h ~= z0h ~ z1h
+		z0h = rev64(z0h) >> 1
+		z1h = rev64(z1h) >> 1
+		z2h = rev64(z2h) >> 1
+
+		v0 := z0
+		v1 := z0h ~ z2
+		v2 := z1 ~ z2h
+		v3 := z1h
+
+		v3 = (v3 << 1) | (v2 >> 63)
+		v2 = (v2 << 1) | (v1 >> 63)
+		v1 = (v1 << 1) | (v0 >> 63)
+		v0 = (v0 << 1)
+
+		v2 ~= v0 ~ (v0 >> 1) ~ (v0 >> 2) ~ (v0 >> 7)
+		v1 ~= (v0 << 63) ~ (v0 << 62) ~ (v0 << 57)
+		v3 ~= v1 ~ (v1 >> 1) ~ (v1 >> 2) ~ (v1 >> 7)
+		v2 ~= (v1 << 63) ~ (v1 << 62) ~ (v1 << 57)
+
+		y0 = v2
+		y1 = v3
+	}
+
+	endian.unchecked_put_u64be(dst[0:], y1)
+	endian.unchecked_put_u64be(dst[8:], y0)
+}
@@ -0,0 +1,75 @@
+package aes_ct64
+
+import "base:intrinsics"
+import "core:crypto/_aes"
+import "core:encoding/endian"
+
+load_blockx1 :: proc "contextless" (q: ^[8]u64, src: []byte) {
+	if len(src) != _aes.BLOCK_SIZE {
+		intrinsics.trap()
+	}
+
+	w: [4]u32 = ---
+	w[0] = endian.unchecked_get_u32le(src[0:])
+	w[1] = endian.unchecked_get_u32le(src[4:])
+	w[2] = endian.unchecked_get_u32le(src[8:])
+	w[3] = endian.unchecked_get_u32le(src[12:])
+	q[0], q[4] = interleave_in(w[:])
+	orthogonalize(q)
+}
+
+store_blockx1 :: proc "contextless" (dst: []byte, q: ^[8]u64) {
+	if len(dst) != _aes.BLOCK_SIZE {
+		intrinsics.trap()
+	}
+
+	orthogonalize(q)
+	w0, w1, w2, w3 := interleave_out(q[0], q[4])
+	endian.unchecked_put_u32le(dst[0:], w0)
+	endian.unchecked_put_u32le(dst[4:], w1)
+	endian.unchecked_put_u32le(dst[8:], w2)
+	endian.unchecked_put_u32le(dst[12:], w3)
+}
+
+load_blocks :: proc "contextless" (q: ^[8]u64, src: [][]byte) {
+	if n := len(src); n > STRIDE || n == 0 {
+		intrinsics.trap()
+	}
+
+	w: [4]u32 = ---
+	for s, i in src {
+		if len(s) != _aes.BLOCK_SIZE {
+			intrinsics.trap()
+		}
+
+		w[0] = endian.unchecked_get_u32le(s[0:])
+		w[1] = endian.unchecked_get_u32le(s[4:])
+		w[2] = endian.unchecked_get_u32le(s[8:])
+		w[3] = endian.unchecked_get_u32le(s[12:])
+		q[i], q[i + 4] = interleave_in(w[:])
+	}
+	orthogonalize(q)
+}
+
+store_blocks :: proc "contextless" (dst: [][]byte, q: ^[8]u64) {
+	if n := len(dst); n > STRIDE || n == 0 {
+		intrinsics.trap()
+	}
+
+	orthogonalize(q)
+	for d, i in dst {
+		// Allow storing [0,4] blocks.
+		if d == nil {
+			break
+		}
+		if len(d) != _aes.BLOCK_SIZE {
+			intrinsics.trap()
+		}
+
+		w0, w1, w2, w3 := interleave_out(q[i], q[i + 4])
+		endian.unchecked_put_u32le(d[0:], w0)
+		endian.unchecked_put_u32le(d[4:], w1)
+		endian.unchecked_put_u32le(d[8:], w2)
+		endian.unchecked_put_u32le(d[12:], w3)
+	}
+}
@@ -0,0 +1,43 @@
+//+build amd64
+package aes_hw_intel
+
+import "core:sys/info"
+
+// is_supported returns true iff hardware accelerated AES
+// is supported.
+is_supported :: proc "contextless" () -> bool {
+	features, ok := info.cpu_features.?
+	if !ok {
+		return false
+	}
+
+	// Note: Everything with AES-NI and PCLMULQDQ has support for
+	// the required SSE extxtensions.
+	req_features :: info.CPU_Features{
+		.sse2,
+		.ssse3,
+		.sse41,
+		.aes,
+		.pclmulqdq,
+	}
+	return features >= req_features
+}
+
+// Context is a keyed AES (ECB) instance.
+Context :: struct {
+	// Note: The ideal thing to do is for the expanded round keys to be
+	// arrays of `__m128i`, however that implies alignment (or using AVX).
+	//
+	// All the people using e-waste processors that don't support an
+	// insturction set that has been around for over 10 years are why
+	// we can't have nice things.
+	_sk_exp_enc: [15][16]byte,
+	_sk_exp_dec: [15][16]byte,
+	_num_rounds: int,
+}
+
+// init initializes a context for AES with the provided key.
+init :: proc(ctx: ^Context, key: []byte) {
+	keysched(ctx, key)
+}
+
@@ -0,0 +1,277 @@
+// Copyright (c) 2017 Thomas Pornin <pornin@bolet.org>
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//   1. Redistributions of source code must retain the above copyright
+//      notice, this list of conditions and the following disclaimer.
+//
+// THIS SOFTWARE IS PROVIDED BY THE AUTHORS “AS IS” AND ANY EXPRESS OR
+// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY
+// DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+//+build amd64
+package aes_hw_intel
+
+import "base:intrinsics"
+import "core:crypto/_aes"
+import "core:simd/x86"
+
+@(private = "file")
+GHASH_STRIDE_HW :: 4
+@(private = "file")
+GHASH_STRIDE_BYTES_HW :: GHASH_STRIDE_HW * _aes.GHASH_BLOCK_SIZE
+
+// GHASH is defined over elements of GF(2^128) with "full little-endian"
+// representation: leftmost byte is least significant, and, within each
+// byte, leftmost _bit_ is least significant. The natural ordering in
+// x86 is "mixed little-endian": bytes are ordered from least to most
+// significant, but bits within a byte are in most-to-least significant
+// order. Going to full little-endian representation would require
+// reversing bits within each byte, which is doable but expensive.
+//
+// Instead, we go to full big-endian representation, by swapping bytes
+// around, which is done with a single _mm_shuffle_epi8() opcode (it
+// comes with SSSE3; all CPU that offer pclmulqdq also have SSSE3). We
+// can use a full big-endian representation because in a carryless
+// multiplication, we have a nice bit reversal property:
+//
+// rev_128(x) * rev_128(y) = rev_255(x * y)
+//
+// So by using full big-endian, we still get the right result, except
+// that it is right-shifted by 1 bit. The left-shift is relatively
+// inexpensive, and it can be mutualised.
+//
+// Since SSE2 opcodes do not have facilities for shitfting full 128-bit
+// values with bit precision, we have to break down values into 64-bit
+// chunks. We number chunks from 0 to 3 in left to right order.
+
+@(private = "file")
+_BYTESWAP_INDEX: x86.__m128i : { 0x08090a0b0c0d0e0f, 0x0001020304050607 }
+
+@(private = "file", require_results, enable_target_feature = "sse2,ssse3")
+byteswap :: #force_inline proc "contextless" (x: x86.__m128i) -> x86.__m128i {
+	return x86._mm_shuffle_epi8(x, _BYTESWAP_INDEX)
+}
+
+// From a 128-bit value kw, compute kx as the XOR of the two 64-bit
+// halves of kw (into the right half of kx; left half is unspecified),
+// and return kx.
+@(private = "file", require_results, enable_target_feature = "sse2")
+bk :: #force_inline proc "contextless" (kw: x86.__m128i) -> x86.__m128i {
+	return x86._mm_xor_si128(kw, x86._mm_shuffle_epi32(kw, 0x0e))
+}
+
+// Combine two 64-bit values (k0:k1) into a 128-bit (kw) value and
+// the XOR of the two values (kx), and return (kw, kx).
+@(private = "file", enable_target_feature = "sse2")
+pbk :: #force_inline proc "contextless" (k0, k1: x86.__m128i) -> (x86.__m128i, x86.__m128i) {
+	kw := x86._mm_unpacklo_epi64(k1, k0)
+	kx := x86._mm_xor_si128(k0, k1)
+	return kw, kx
+}
+
+// Left-shift by 1 bit a 256-bit value (in four 64-bit words).
+@(private = "file", require_results, enable_target_feature = "sse2")
+sl_256 :: #force_inline proc "contextless" (x0, x1, x2, x3: x86.__m128i) -> (x86.__m128i, x86.__m128i, x86.__m128i, x86.__m128i) {
+	x0, x1, x2, x3 := x0, x1, x2, x3
+
+	x0 = x86._mm_or_si128(x86._mm_slli_epi64(x0, 1), x86._mm_srli_epi64(x1, 63))
+	x1 = x86._mm_or_si128(x86._mm_slli_epi64(x1, 1), x86._mm_srli_epi64(x2, 63))
+	x2 = x86._mm_or_si128(x86._mm_slli_epi64(x2, 1), x86._mm_srli_epi64(x3, 63))
+	x3 = x86._mm_slli_epi64(x3, 1)
+
+	return x0, x1, x2, x3
+}
+
+// Perform reduction in GF(2^128).
+@(private = "file", require_results, enable_target_feature = "sse2")
+reduce_f128 :: #force_inline proc "contextless" (x0, x1, x2, x3: x86.__m128i) -> (x86.__m128i, x86.__m128i) {
+	x0, x1, x2 := x0, x1, x2
+
+	x1 = x86._mm_xor_si128(
+		x1,
+		x86._mm_xor_si128(
+			x86._mm_xor_si128(
+				x3,
+				x86._mm_srli_epi64(x3, 1)),
+			x86._mm_xor_si128(
+				x86._mm_srli_epi64(x3, 2),
+				x86._mm_srli_epi64(x3, 7))))
+	x2 = x86._mm_xor_si128(
+		x86._mm_xor_si128(
+			x2,
+			x86._mm_slli_epi64(x3, 63)),
+		x86._mm_xor_si128(
+			x86._mm_slli_epi64(x3, 62),
+			x86._mm_slli_epi64(x3, 57)))
+	x0 = x86._mm_xor_si128(
+		x0,
+		x86._mm_xor_si128(
+			x86._mm_xor_si128(
+				x2,
+				x86._mm_srli_epi64(x2, 1)),
+			x86._mm_xor_si128(
+				x86._mm_srli_epi64(x2, 2),
+				x86._mm_srli_epi64(x2, 7))))
+	x1 = x86._mm_xor_si128(
+		x86._mm_xor_si128(
+			x1,
+			x86._mm_slli_epi64(x2, 63)),
+		x86._mm_xor_si128(
+			x86._mm_slli_epi64(x2, 62),
+			x86._mm_slli_epi64(x2, 57)))
+
+	return x0, x1
+}
+
+// Square value kw in GF(2^128) into (dw,dx).
+@(private = "file", require_results, enable_target_feature = "sse2,pclmul")
+square_f128 :: #force_inline proc "contextless" (kw: x86.__m128i) -> (x86.__m128i, x86.__m128i) {
+	z1 := x86._mm_clmulepi64_si128(kw, kw, 0x11)
+	z3 := x86._mm_clmulepi64_si128(kw, kw, 0x00)
+	z0 := x86._mm_shuffle_epi32(z1, 0x0E)
+	z2 := x86._mm_shuffle_epi32(z3, 0x0E)
+	z0, z1, z2, z3 = sl_256(z0, z1, z2, z3)
+	z0, z1 = reduce_f128(z0, z1, z2, z3)
+	return pbk(z0, z1)
+}
+
+// ghash calculates the GHASH of data, with the key `key`, and input `dst`
+// and `data`, and stores the resulting digest in `dst`.
+//
+// Note: `dst` is both an input and an output, to support easy implementation
+// of GCM.
+@(enable_target_feature = "sse2,ssse3,pclmul")
+ghash :: proc "contextless" (dst, key, data: []byte) #no_bounds_check {
+	if len(dst) != _aes.GHASH_BLOCK_SIZE || len(key) != _aes.GHASH_BLOCK_SIZE {
+		intrinsics.trap()
+	}
+
+	// Note: BearSSL opts to copy the remainder into a zero-filled
+	// 64-byte buffer.  We do something slightly more simple.
+
+	// Load key and dst (h and y).
+	yw := intrinsics.unaligned_load((^x86.__m128i)(raw_data(dst)))
+	h1w := intrinsics.unaligned_load((^x86.__m128i)(raw_data(key)))
+	yw = byteswap(yw)
+	h1w = byteswap(h1w)
+	h1x := bk(h1w)
+
+	// Process 4 blocks at a time
+	buf := data
+	l := len(buf)
+	if l >= GHASH_STRIDE_BYTES_HW {
+		// Compute h2 = h^2
+		h2w, h2x := square_f128(h1w)
+
+		// Compute h3 = h^3 = h*(h^2)
+		t1 := x86._mm_clmulepi64_si128(h1w, h2w, 0x11)
+		t3 := x86._mm_clmulepi64_si128(h1w, h2w, 0x00)
+		t2 := x86._mm_xor_si128(
+			x86._mm_clmulepi64_si128(h1x, h2x, 0x00),
+			x86._mm_xor_si128(t1, t3))
+		t0 := x86._mm_shuffle_epi32(t1, 0x0E)
+		t1 = x86._mm_xor_si128(t1, x86._mm_shuffle_epi32(t2, 0x0E))
+		t2 = x86._mm_xor_si128(t2, x86._mm_shuffle_epi32(t3, 0x0E))
+		t0, t1, t2, t3 = sl_256(t0, t1, t2, t3)
+		t0, t1 = reduce_f128(t0, t1, t2, t3)
+		h3w, h3x := pbk(t0, t1)
+
+		// Compute h4 = h^4 = (h^2)^2
+		h4w, h4x := square_f128(h2w)
+
+		for l >= GHASH_STRIDE_BYTES_HW {
+			aw0 := intrinsics.unaligned_load((^x86.__m128i)(raw_data(buf)))
+			aw1 := intrinsics.unaligned_load((^x86.__m128i)(raw_data(buf[16:])))
+			aw2 := intrinsics.unaligned_load((^x86.__m128i)(raw_data(buf[32:])))
+			aw3 := intrinsics.unaligned_load((^x86.__m128i)(raw_data(buf[48:])))
+			aw0 = byteswap(aw0)
+			aw1 = byteswap(aw1)
+			aw2 = byteswap(aw2)
+			aw3 = byteswap(aw3)
+			buf, l = buf[GHASH_STRIDE_BYTES_HW:], l - GHASH_STRIDE_BYTES_HW
+
+			aw0 = x86._mm_xor_si128(aw0, yw)
+			ax1 := bk(aw1)
+			ax2 := bk(aw2)
+			ax3 := bk(aw3)
+			ax0 := bk(aw0)
+
+			t1 = x86._mm_xor_si128(
+				x86._mm_xor_si128(
+					x86._mm_clmulepi64_si128(aw0, h4w, 0x11),
+					x86._mm_clmulepi64_si128(aw1, h3w, 0x11)),
+				x86._mm_xor_si128(
+					x86._mm_clmulepi64_si128(aw2, h2w, 0x11),
+					x86._mm_clmulepi64_si128(aw3, h1w, 0x11)))
+			t3 = x86._mm_xor_si128(
+				x86._mm_xor_si128(
+					x86._mm_clmulepi64_si128(aw0, h4w, 0x00),
+					x86._mm_clmulepi64_si128(aw1, h3w, 0x00)),
+				x86._mm_xor_si128(
+					x86._mm_clmulepi64_si128(aw2, h2w, 0x00),
+					x86._mm_clmulepi64_si128(aw3, h1w, 0x00)))
+			t2 = x86._mm_xor_si128(
+				x86._mm_xor_si128(
+					x86._mm_clmulepi64_si128(ax0, h4x, 0x00),
+					x86._mm_clmulepi64_si128(ax1, h3x, 0x00)),
+				x86._mm_xor_si128(
+					x86._mm_clmulepi64_si128(ax2, h2x, 0x00),
+					x86._mm_clmulepi64_si128(ax3, h1x, 0x00)))
+			t2 = x86._mm_xor_si128(t2, x86._mm_xor_si128(t1, t3))
+			t0 = x86._mm_shuffle_epi32(t1, 0x0E)
+			t1 = x86._mm_xor_si128(t1, x86._mm_shuffle_epi32(t2, 0x0E))
+			t2 = x86._mm_xor_si128(t2, x86._mm_shuffle_epi32(t3, 0x0E))
+			t0, t1, t2, t3 = sl_256(t0, t1, t2, t3)
+			t0, t1 = reduce_f128(t0, t1, t2, t3)
+			yw = x86._mm_unpacklo_epi64(t1, t0)
+		}
+	}
+
+	// Process 1 block at a time
+	for l > 0 {
+		src: []byte = ---
+		if l >= _aes.GHASH_BLOCK_SIZE {
+			src = buf
+			buf = buf[_aes.GHASH_BLOCK_SIZE:]
+			l -= _aes.GHASH_BLOCK_SIZE
+		} else {
+			tmp: [_aes.GHASH_BLOCK_SIZE]byte
+			copy(tmp[:], buf)
+			src = tmp[:]
+			l = 0
+		}
+
+		aw := intrinsics.unaligned_load((^x86.__m128i)(raw_data(src)))
+		aw = byteswap(aw)
+
+		aw = x86._mm_xor_si128(aw, yw)
+		ax := bk(aw)
+
+		t1 := x86._mm_clmulepi64_si128(aw, h1w, 0x11)
+		t3 := x86._mm_clmulepi64_si128(aw, h1w, 0x00)
+		t2 := x86._mm_clmulepi64_si128(ax, h1x, 0x00)
+		t2 = x86._mm_xor_si128(t2, x86._mm_xor_si128(t1, t3))
+		t0 := x86._mm_shuffle_epi32(t1, 0x0E)
+		t1 = x86._mm_xor_si128(t1, x86._mm_shuffle_epi32(t2, 0x0E))
+		t2 = x86._mm_xor_si128(t2, x86._mm_shuffle_epi32(t3, 0x0E))
+		t0, t1, t2, t3 = sl_256(t0, t1, t2, t3)
+		t0, t1 = reduce_f128(t0, t1, t2, t3)
+		yw = x86._mm_unpacklo_epi64(t1, t0)
+	}
+
+	// Write back the hash (dst, aka y)
+	yw = byteswap(yw)
+	intrinsics.unaligned_store((^x86.__m128i)(raw_data(dst)), yw)
+}
@@ -0,0 +1,178 @@
+// Copyright (c) 2017 Thomas Pornin <pornin@bolet.org>
+// All rights reserved.
+//
+// Redistribution and use in source and binary forms, with or without
+// modification, are permitted provided that the following conditions
+// are met:
+//
+//   1. Redistributions of source code must retain the above copyright
+//      notice, this list of conditions and the following disclaimer.
+//
+// THIS SOFTWARE IS PROVIDED BY THE AUTHORS “AS IS” AND ANY EXPRESS OR
+// IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+// WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
+// ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHORS BE LIABLE FOR ANY
+// DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
+// DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
+// GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
+// INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY,
+// WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING
+// NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF
+// THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+
+//+build amd64
+package aes_hw_intel
+
+import "base:intrinsics"
+import "core:crypto/_aes"
+import "core:mem"
+import "core:simd/x86"
+
+// Intel AES-NI based implementation.  Inspiration taken from BearSSL.
+//
+// Note: This assumes that the SROA optimization pass is enabled to be
+// anything resembling performat otherwise, LLVM will not elide a massive
+// number of redundant loads/stores it generates for every intrinsic call.
+
+@(private = "file", require_results, enable_target_feature = "sse2")
+expand_step128 :: #force_inline proc(k1, k2: x86.__m128i) -> x86.__m128i {
+	k1, k2 := k1, k2
+
+	k2 = x86._mm_shuffle_epi32(k2, 0xff)
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	return x86._mm_xor_si128(k1, k2)
+}
+
+@(private = "file", require_results, enable_target_feature = "sse,sse2")
+expand_step192a :: #force_inline proc (k1_, k2_: ^x86.__m128i, k3: x86.__m128i) -> (x86.__m128i, x86.__m128i) {
+	k1, k2, k3 := k1_^, k2_^, k3
+
+	k3 = x86._mm_shuffle_epi32(k3, 0x55)
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, k3)
+
+	tmp := k2
+	k2 = x86._mm_xor_si128(k2, x86._mm_slli_si128(k2, 0x04))
+	k2 = x86._mm_xor_si128(k2, x86._mm_shuffle_epi32(k1, 0xff))
+
+	k1_, k2_ := k1_, k2_
+	k1_^, k2_^ = k1, k2
+
+	r1 := transmute(x86.__m128i)(x86._mm_shuffle_ps(transmute(x86.__m128)(tmp), transmute(x86.__m128)(k1), 0x44))
+	r2 := transmute(x86.__m128i)(x86._mm_shuffle_ps(transmute(x86.__m128)(k1), transmute(x86.__m128)(k2), 0x4e))
+
+	return r1, r2
+}
+
+@(private = "file", require_results, enable_target_feature = "sse2")
+expand_step192b :: #force_inline proc (k1_, k2_: ^x86.__m128i, k3: x86.__m128i) -> x86.__m128i {
+	k1, k2, k3 := k1_^, k2_^, k3
+
+	k3 = x86._mm_shuffle_epi32(k3, 0x55)
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, k3)
+
+	k2 = x86._mm_xor_si128(k2, x86._mm_slli_si128(k2, 0x04))
+	k2 = x86._mm_xor_si128(k2, x86._mm_shuffle_epi32(k1, 0xff))
+
+	k1_, k2_ := k1_, k2_
+	k1_^, k2_^ = k1, k2
+
+	return k1
+}
+
+@(private = "file", require_results, enable_target_feature = "sse2")
+expand_step256b :: #force_inline proc(k1, k2: x86.__m128i) -> x86.__m128i {
+	k1, k2 := k1, k2
+
+	k2 = x86._mm_shuffle_epi32(k2, 0xaa)
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	k1 = x86._mm_xor_si128(k1, x86._mm_slli_si128(k1, 0x04))
+	return x86._mm_xor_si128(k1, k2)
+}
+
+@(private = "file", enable_target_feature = "aes")
+derive_dec_keys :: proc(ctx: ^Context, sks: ^[15]x86.__m128i, num_rounds: int) {
+	intrinsics.unaligned_store((^x86.__m128i)(&ctx._sk_exp_dec[0]), sks[num_rounds])
+	for i in 1 ..< num_rounds {
+		tmp := x86._mm_aesimc_si128(sks[i])
+		intrinsics.unaligned_store((^x86.__m128i)(&ctx._sk_exp_dec[num_rounds - i]), tmp)
+	}
+	intrinsics.unaligned_store((^x86.__m128i)(&ctx._sk_exp_dec[num_rounds]), sks[0])
+}
+
+@(private, enable_target_feature = "sse,sse2,aes")
+keysched :: proc(ctx: ^Context, key: []byte) {
+	sks: [15]x86.__m128i = ---
+
+	// Compute the encryption keys.
+	num_rounds, key_len := 0, len(key)
+	switch key_len {
+	case _aes.KEY_SIZE_128:
+		sks[0] = intrinsics.unaligned_load((^x86.__m128i)(raw_data(key)))
+		sks[1] = expand_step128(sks[0], x86._mm_aeskeygenassist_si128(sks[0], 0x01))
+		sks[2] = expand_step128(sks[1], x86._mm_aeskeygenassist_si128(sks[1], 0x02))
+		sks[3] = expand_step128(sks[2], x86._mm_aeskeygenassist_si128(sks[2], 0x04))
+		sks[4] = expand_step128(sks[3], x86._mm_aeskeygenassist_si128(sks[3], 0x08))
+		sks[5] = expand_step128(sks[4], x86._mm_aeskeygenassist_si128(sks[4], 0x10))
+		sks[6] = expand_step128(sks[5], x86._mm_aeskeygenassist_si128(sks[5], 0x20))
+		sks[7] = expand_step128(sks[6], x86._mm_aeskeygenassist_si128(sks[6], 0x40))
+		sks[8] = expand_step128(sks[7], x86._mm_aeskeygenassist_si128(sks[7], 0x80))
+		sks[9] = expand_step128(sks[8], x86._mm_aeskeygenassist_si128(sks[8], 0x1b))
+		sks[10] = expand_step128(sks[9], x86._mm_aeskeygenassist_si128(sks[9], 0x36))
+		num_rounds = _aes.ROUNDS_128
+	case _aes.KEY_SIZE_192:
+		k0 := intrinsics.unaligned_load((^x86.__m128i)(raw_data(key)))
+		k1 := x86.__m128i{
+			intrinsics.unaligned_load((^i64)(raw_data(key[16:]))),
+			0,
+		}
+		sks[0] = k0
+		sks[1], sks[2] = expand_step192a(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x01))
+		sks[3] = expand_step192b(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x02))
+		sks[4], sks[5] = expand_step192a(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x04))
+		sks[6] = expand_step192b(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x08))
+		sks[7], sks[8] = expand_step192a(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x10))
+		sks[9] = expand_step192b(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x20))
+		sks[10], sks[11] = expand_step192a(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x40))
+		sks[12] = expand_step192b(&k0, &k1, x86._mm_aeskeygenassist_si128(k1, 0x80))
+		num_rounds = _aes.ROUNDS_192
+	case _aes.KEY_SIZE_256:
+		sks[0] = intrinsics.unaligned_load((^x86.__m128i)(raw_data(key)))
+		sks[1] = intrinsics.unaligned_load((^x86.__m128i)(raw_data(key[16:])))
+		sks[2] = expand_step128(sks[0], x86._mm_aeskeygenassist_si128(sks[1], 0x01))
+		sks[3] = expand_step256b(sks[1], x86._mm_aeskeygenassist_si128(sks[2], 0x01))
+		sks[4] = expand_step128(sks[2], x86._mm_aeskeygenassist_si128(sks[3], 0x02))
+		sks[5] = expand_step256b(sks[3], x86._mm_aeskeygenassist_si128(sks[4], 0x02))
+		sks[6] = expand_step128(sks[4], x86._mm_aeskeygenassist_si128(sks[5], 0x04))
+		sks[7] = expand_step256b(sks[5], x86._mm_aeskeygenassist_si128(sks[6], 0x04))
+		sks[8] = expand_step128(sks[6], x86._mm_aeskeygenassist_si128(sks[7], 0x08))
+		sks[9] = expand_step256b(sks[7], x86._mm_aeskeygenassist_si128(sks[8], 0x08))
+		sks[10] = expand_step128(sks[8], x86._mm_aeskeygenassist_si128(sks[9], 0x10))
+		sks[11] = expand_step256b(sks[9], x86._mm_aeskeygenassist_si128(sks[10], 0x10))
+		sks[12] = expand_step128(sks[10], x86._mm_aeskeygenassist_si128(sks[11], 0x20))
+		sks[13] = expand_step256b(sks[11], x86._mm_aeskeygenassist_si128(sks[12], 0x20))
+		sks[14] = expand_step128(sks[12], x86._mm_aeskeygenassist_si128(sks[13], 0x40))
+		num_rounds = _aes.ROUNDS_256
+	case:
+		panic("crypto/aes: invalid AES key size")
+	}
+	for i in 0 ..= num_rounds {
+		intrinsics.unaligned_store((^x86.__m128i)(&ctx._sk_exp_enc[i]), sks[i])
+	}
+
+	// Compute the decryption keys.  GCM and CTR do not need this, however
+	// ECB, CBC, OCB3, etc do.
+	derive_dec_keys(ctx, &sks, num_rounds)
+
+	ctx._num_rounds = num_rounds
+
+	mem.zero_explicit(&sks, size_of(sks))
+}
@@ -0,0 +1,123 @@
+package _chacha20
+
+import "base:intrinsics"
+import "core:encoding/endian"
+import "core:math/bits"
+import "core:mem"
+
+// KEY_SIZE is the (X)ChaCha20 key size in bytes.
+KEY_SIZE :: 32
+// IV_SIZE is the ChaCha20 IV size in bytes.
+IV_SIZE :: 12
+// XIV_SIZE is the XChaCha20 IV size in bytes.
+XIV_SIZE :: 24
+
+// MAX_CTR_IETF is the maximum counter value for the IETF flavor ChaCha20.
+MAX_CTR_IETF :: 0xffffffff
+// BLOCK_SIZE is the (X)ChaCha20 block size in bytes.
+BLOCK_SIZE :: 64
+// STATE_SIZE_U32 is the (X)ChaCha20 state size in u32s.
+STATE_SIZE_U32 :: 16
+// Rounds is the (X)ChaCha20 round count.
+ROUNDS :: 20
+
+// SIGMA_0 is sigma[0:4].
+SIGMA_0: u32 : 0x61707865
+// SIGMA_1 is sigma[4:8].
+SIGMA_1: u32 : 0x3320646e
+// SIGMA_2 is sigma[8:12].
+SIGMA_2: u32 : 0x79622d32
+// SIGMA_3 is sigma[12:16].
+SIGMA_3: u32 : 0x6b206574
+
+// Context is a ChaCha20 or XChaCha20 instance.
+Context :: struct {
+	_s:              [STATE_SIZE_U32]u32,
+	_buffer:         [BLOCK_SIZE]byte,
+	_off:            int,
+	_is_ietf_flavor: bool,
+	_is_initialized: bool,
+}
+
+// init inititializes a Context for ChaCha20 with the provided key and
+// iv.
+//
+// WARNING: This ONLY handles ChaCha20.  XChaCha20 sub-key and IV
+// derivation is expected to be handled by the caller, so that the
+// HChaCha call can be suitably accelerated.
+init :: proc "contextless" (ctx: ^Context, key, iv: []byte, is_xchacha: bool) {
+	if len(key) != KEY_SIZE || len(iv) != IV_SIZE {
+		intrinsics.trap()
+	}
+
+	k, n := key, iv
+
+	ctx._s[0] = SIGMA_0
+	ctx._s[1] = SIGMA_1
+	ctx._s[2] = SIGMA_2
+	ctx._s[3] = SIGMA_3
+	ctx._s[4] = endian.unchecked_get_u32le(k[0:4])
+	ctx._s[5] = endian.unchecked_get_u32le(k[4:8])
+	ctx._s[6] = endian.unchecked_get_u32le(k[8:12])
+	ctx._s[7] = endian.unchecked_get_u32le(k[12:16])
+	ctx._s[8] = endian.unchecked_get_u32le(k[16:20])
+	ctx._s[9] = endian.unchecked_get_u32le(k[20:24])
+	ctx._s[10] = endian.unchecked_get_u32le(k[24:28])
+	ctx._s[11] = endian.unchecked_get_u32le(k[28:32])
+	ctx._s[12] = 0
+	ctx._s[13] = endian.unchecked_get_u32le(n[0:4])
+	ctx._s[14] = endian.unchecked_get_u32le(n[4:8])
+	ctx._s[15] = endian.unchecked_get_u32le(n[8:12])
+
+	ctx._off = BLOCK_SIZE
+	ctx._is_ietf_flavor = !is_xchacha
+	ctx._is_initialized = true
+}
+
+// seek seeks the (X)ChaCha20 stream counter to the specified block.
+seek :: proc(ctx: ^Context, block_nr: u64) {
+	assert(ctx._is_initialized)
+
+	if ctx._is_ietf_flavor {
+		if block_nr > MAX_CTR_IETF {
+			panic("crypto/chacha20: attempted to seek past maximum counter")
+		}
+	} else {
+		ctx._s[13] = u32(block_nr >> 32)
+	}
+	ctx._s[12] = u32(block_nr)
+	ctx._off = BLOCK_SIZE
+}
+
+// reset sanitizes the Context.  The Context must be re-initialized to
+// be used again.
+reset :: proc(ctx: ^Context) {
+	mem.zero_explicit(&ctx._s, size_of(ctx._s))
+	mem.zero_explicit(&ctx._buffer, size_of(ctx._buffer))
+
+	ctx._is_initialized = false
+}
+
+check_counter_limit :: proc(ctx: ^Context, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	//
+	// While all modern "standard" definitions of ChaCha20 use
+	// the IETF 32-bit counter, for XChaCha20 most common
+	// implementations allow for a 64-bit counter.
+	//
+	// Honestly, the answer here is "use a MRAE primitive", but
+	// go with "common" practice in the case of XChaCha20.
+
+	ERR_CTR_EXHAUSTED :: "crypto/chacha20: maximum (X)ChaCha20 keystream per IV reached"
+
+	if ctx._is_ietf_flavor {
+		if u64(ctx._s[12]) + u64(nr_blocks) > MAX_CTR_IETF {
+			panic(ERR_CTR_EXHAUSTED)
+		}
+	} else {
+		ctr := (u64(ctx._s[13]) << 32) | u64(ctx._s[12])
+		if _, carry := bits.add_u64(ctr, u64(nr_blocks), 0); carry != 0 {
+			panic(ERR_CTR_EXHAUSTED)
+		}
+	}
+}
@@ -0,0 +1,360 @@
+package chacha20_ref
+
+import "core:crypto/_chacha20"
+import "core:encoding/endian"
+import "core:math/bits"
+
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	_chacha20.check_counter_limit(ctx, nr_blocks)
+
+	dst, src := dst, src
+	x := &ctx._s
+	for n := 0; n < nr_blocks; n = n + 1 {
+		x0, x1, x2, x3 :=
+			_chacha20.SIGMA_0, _chacha20.SIGMA_1, _chacha20.SIGMA_2, _chacha20.SIGMA_3
+		x4, x5, x6, x7, x8, x9, x10, x11, x12, x13, x14, x15 :=
+			x[4], x[5], x[6], x[7], x[8], x[9], x[10], x[11], x[12], x[13], x[14], x[15]
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			// Even when forcing inlining manually inlining all of
+			// these is decently faster.
+
+			// quarterround(x, 0, 4, 8, 12)
+			x0 += x4
+			x12 ~= x0
+			x12 = bits.rotate_left32(x12, 16)
+			x8 += x12
+			x4 ~= x8
+			x4 = bits.rotate_left32(x4, 12)
+			x0 += x4
+			x12 ~= x0
+			x12 = bits.rotate_left32(x12, 8)
+			x8 += x12
+			x4 ~= x8
+			x4 = bits.rotate_left32(x4, 7)
+
+			// quarterround(x, 1, 5, 9, 13)
+			x1 += x5
+			x13 ~= x1
+			x13 = bits.rotate_left32(x13, 16)
+			x9 += x13
+			x5 ~= x9
+			x5 = bits.rotate_left32(x5, 12)
+			x1 += x5
+			x13 ~= x1
+			x13 = bits.rotate_left32(x13, 8)
+			x9 += x13
+			x5 ~= x9
+			x5 = bits.rotate_left32(x5, 7)
+
+			// quarterround(x, 2, 6, 10, 14)
+			x2 += x6
+			x14 ~= x2
+			x14 = bits.rotate_left32(x14, 16)
+			x10 += x14
+			x6 ~= x10
+			x6 = bits.rotate_left32(x6, 12)
+			x2 += x6
+			x14 ~= x2
+			x14 = bits.rotate_left32(x14, 8)
+			x10 += x14
+			x6 ~= x10
+			x6 = bits.rotate_left32(x6, 7)
+
+			// quarterround(x, 3, 7, 11, 15)
+			x3 += x7
+			x15 ~= x3
+			x15 = bits.rotate_left32(x15, 16)
+			x11 += x15
+			x7 ~= x11
+			x7 = bits.rotate_left32(x7, 12)
+			x3 += x7
+			x15 ~= x3
+			x15 = bits.rotate_left32(x15, 8)
+			x11 += x15
+			x7 ~= x11
+			x7 = bits.rotate_left32(x7, 7)
+
+			// quarterround(x, 0, 5, 10, 15)
+			x0 += x5
+			x15 ~= x0
+			x15 = bits.rotate_left32(x15, 16)
+			x10 += x15
+			x5 ~= x10
+			x5 = bits.rotate_left32(x5, 12)
+			x0 += x5
+			x15 ~= x0
+			x15 = bits.rotate_left32(x15, 8)
+			x10 += x15
+			x5 ~= x10
+			x5 = bits.rotate_left32(x5, 7)
+
+			// quarterround(x, 1, 6, 11, 12)
+			x1 += x6
+			x12 ~= x1
+			x12 = bits.rotate_left32(x12, 16)
+			x11 += x12
+			x6 ~= x11
+			x6 = bits.rotate_left32(x6, 12)
+			x1 += x6
+			x12 ~= x1
+			x12 = bits.rotate_left32(x12, 8)
+			x11 += x12
+			x6 ~= x11
+			x6 = bits.rotate_left32(x6, 7)
+
+			// quarterround(x, 2, 7, 8, 13)
+			x2 += x7
+			x13 ~= x2
+			x13 = bits.rotate_left32(x13, 16)
+			x8 += x13
+			x7 ~= x8
+			x7 = bits.rotate_left32(x7, 12)
+			x2 += x7
+			x13 ~= x2
+			x13 = bits.rotate_left32(x13, 8)
+			x8 += x13
+			x7 ~= x8
+			x7 = bits.rotate_left32(x7, 7)
+
+			// quarterround(x, 3, 4, 9, 14)
+			x3 += x4
+			x14 ~= x3
+			x14 = bits.rotate_left32(x14, 16)
+			x9 += x14
+			x4 ~= x9
+			x4 = bits.rotate_left32(x4, 12)
+			x3 += x4
+			x14 ~= x3
+			x14 = bits.rotate_left32(x14, 8)
+			x9 += x14
+			x4 ~= x9
+			x4 = bits.rotate_left32(x4, 7)
+		}
+
+		x0 += _chacha20.SIGMA_0
+		x1 += _chacha20.SIGMA_1
+		x2 += _chacha20.SIGMA_2
+		x3 += _chacha20.SIGMA_3
+		x4 += x[4]
+		x5 += x[5]
+		x6 += x[6]
+		x7 += x[7]
+		x8 += x[8]
+		x9 += x[9]
+		x10 += x[10]
+		x11 += x[11]
+		x12 += x[12]
+		x13 += x[13]
+		x14 += x[14]
+		x15 += x[15]
+
+		// - The caller(s) ensure that src/dst are valid.
+		// - The compiler knows if the target is picky about alignment.
+
+		#no_bounds_check {
+			if src != nil {
+				endian.unchecked_put_u32le(dst[0:4], endian.unchecked_get_u32le(src[0:4]) ~ x0)
+				endian.unchecked_put_u32le(dst[4:8], endian.unchecked_get_u32le(src[4:8]) ~ x1)
+				endian.unchecked_put_u32le(dst[8:12], endian.unchecked_get_u32le(src[8:12]) ~ x2)
+				endian.unchecked_put_u32le(dst[12:16], endian.unchecked_get_u32le(src[12:16]) ~ x3)
+				endian.unchecked_put_u32le(dst[16:20], endian.unchecked_get_u32le(src[16:20]) ~ x4)
+				endian.unchecked_put_u32le(dst[20:24], endian.unchecked_get_u32le(src[20:24]) ~ x5)
+				endian.unchecked_put_u32le(dst[24:28], endian.unchecked_get_u32le(src[24:28]) ~ x6)
+				endian.unchecked_put_u32le(dst[28:32], endian.unchecked_get_u32le(src[28:32]) ~ x7)
+				endian.unchecked_put_u32le(dst[32:36], endian.unchecked_get_u32le(src[32:36]) ~ x8)
+				endian.unchecked_put_u32le(dst[36:40], endian.unchecked_get_u32le(src[36:40]) ~ x9)
+				endian.unchecked_put_u32le(
+					dst[40:44],
+					endian.unchecked_get_u32le(src[40:44]) ~ x10,
+				)
+				endian.unchecked_put_u32le(
+					dst[44:48],
+					endian.unchecked_get_u32le(src[44:48]) ~ x11,
+				)
+				endian.unchecked_put_u32le(
+					dst[48:52],
+					endian.unchecked_get_u32le(src[48:52]) ~ x12,
+				)
+				endian.unchecked_put_u32le(
+					dst[52:56],
+					endian.unchecked_get_u32le(src[52:56]) ~ x13,
+				)
+				endian.unchecked_put_u32le(
+					dst[56:60],
+					endian.unchecked_get_u32le(src[56:60]) ~ x14,
+				)
+				endian.unchecked_put_u32le(
+					dst[60:64],
+					endian.unchecked_get_u32le(src[60:64]) ~ x15,
+				)
+				src = src[_chacha20.BLOCK_SIZE:]
+			} else {
+				endian.unchecked_put_u32le(dst[0:4], x0)
+				endian.unchecked_put_u32le(dst[4:8], x1)
+				endian.unchecked_put_u32le(dst[8:12], x2)
+				endian.unchecked_put_u32le(dst[12:16], x3)
+				endian.unchecked_put_u32le(dst[16:20], x4)
+				endian.unchecked_put_u32le(dst[20:24], x5)
+				endian.unchecked_put_u32le(dst[24:28], x6)
+				endian.unchecked_put_u32le(dst[28:32], x7)
+				endian.unchecked_put_u32le(dst[32:36], x8)
+				endian.unchecked_put_u32le(dst[36:40], x9)
+				endian.unchecked_put_u32le(dst[40:44], x10)
+				endian.unchecked_put_u32le(dst[44:48], x11)
+				endian.unchecked_put_u32le(dst[48:52], x12)
+				endian.unchecked_put_u32le(dst[52:56], x13)
+				endian.unchecked_put_u32le(dst[56:60], x14)
+				endian.unchecked_put_u32le(dst[60:64], x15)
+			}
+			dst = dst[_chacha20.BLOCK_SIZE:]
+		}
+
+		// Increment the counter.  Overflow checking is done upon
+		// entry into the routine, so a 64-bit increment safely
+		// covers both cases.
+		new_ctr := ((u64(ctx._s[13]) << 32) | u64(ctx._s[12])) + 1
+		x[12] = u32(new_ctr)
+		x[13] = u32(new_ctr >> 32)
+	}
+}
+
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	x0, x1, x2, x3 := _chacha20.SIGMA_0, _chacha20.SIGMA_1, _chacha20.SIGMA_2, _chacha20.SIGMA_3
+	x4 := endian.unchecked_get_u32le(key[0:4])
+	x5 := endian.unchecked_get_u32le(key[4:8])
+	x6 := endian.unchecked_get_u32le(key[8:12])
+	x7 := endian.unchecked_get_u32le(key[12:16])
+	x8 := endian.unchecked_get_u32le(key[16:20])
+	x9 := endian.unchecked_get_u32le(key[20:24])
+	x10 := endian.unchecked_get_u32le(key[24:28])
+	x11 := endian.unchecked_get_u32le(key[28:32])
+	x12 := endian.unchecked_get_u32le(iv[0:4])
+	x13 := endian.unchecked_get_u32le(iv[4:8])
+	x14 := endian.unchecked_get_u32le(iv[8:12])
+	x15 := endian.unchecked_get_u32le(iv[12:16])
+
+	for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+		// quarterround(x, 0, 4, 8, 12)
+		x0 += x4
+		x12 ~= x0
+		x12 = bits.rotate_left32(x12, 16)
+		x8 += x12
+		x4 ~= x8
+		x4 = bits.rotate_left32(x4, 12)
+		x0 += x4
+		x12 ~= x0
+		x12 = bits.rotate_left32(x12, 8)
+		x8 += x12
+		x4 ~= x8
+		x4 = bits.rotate_left32(x4, 7)
+
+		// quarterround(x, 1, 5, 9, 13)
+		x1 += x5
+		x13 ~= x1
+		x13 = bits.rotate_left32(x13, 16)
+		x9 += x13
+		x5 ~= x9
+		x5 = bits.rotate_left32(x5, 12)
+		x1 += x5
+		x13 ~= x1
+		x13 = bits.rotate_left32(x13, 8)
+		x9 += x13
+		x5 ~= x9
+		x5 = bits.rotate_left32(x5, 7)
+
+		// quarterround(x, 2, 6, 10, 14)
+		x2 += x6
+		x14 ~= x2
+		x14 = bits.rotate_left32(x14, 16)
+		x10 += x14
+		x6 ~= x10
+		x6 = bits.rotate_left32(x6, 12)
+		x2 += x6
+		x14 ~= x2
+		x14 = bits.rotate_left32(x14, 8)
+		x10 += x14
+		x6 ~= x10
+		x6 = bits.rotate_left32(x6, 7)
+
+		// quarterround(x, 3, 7, 11, 15)
+		x3 += x7
+		x15 ~= x3
+		x15 = bits.rotate_left32(x15, 16)
+		x11 += x15
+		x7 ~= x11
+		x7 = bits.rotate_left32(x7, 12)
+		x3 += x7
+		x15 ~= x3
+		x15 = bits.rotate_left32(x15, 8)
+		x11 += x15
+		x7 ~= x11
+		x7 = bits.rotate_left32(x7, 7)
+
+		// quarterround(x, 0, 5, 10, 15)
+		x0 += x5
+		x15 ~= x0
+		x15 = bits.rotate_left32(x15, 16)
+		x10 += x15
+		x5 ~= x10
+		x5 = bits.rotate_left32(x5, 12)
+		x0 += x5
+		x15 ~= x0
+		x15 = bits.rotate_left32(x15, 8)
+		x10 += x15
+		x5 ~= x10
+		x5 = bits.rotate_left32(x5, 7)
+
+		// quarterround(x, 1, 6, 11, 12)
+		x1 += x6
+		x12 ~= x1
+		x12 = bits.rotate_left32(x12, 16)
+		x11 += x12
+		x6 ~= x11
+		x6 = bits.rotate_left32(x6, 12)
+		x1 += x6
+		x12 ~= x1
+		x12 = bits.rotate_left32(x12, 8)
+		x11 += x12
+		x6 ~= x11
+		x6 = bits.rotate_left32(x6, 7)
+
+		// quarterround(x, 2, 7, 8, 13)
+		x2 += x7
+		x13 ~= x2
+		x13 = bits.rotate_left32(x13, 16)
+		x8 += x13
+		x7 ~= x8
+		x7 = bits.rotate_left32(x7, 12)
+		x2 += x7
+		x13 ~= x2
+		x13 = bits.rotate_left32(x13, 8)
+		x8 += x13
+		x7 ~= x8
+		x7 = bits.rotate_left32(x7, 7)
+
+		// quarterround(x, 3, 4, 9, 14)
+		x3 += x4
+		x14 ~= x3
+		x14 = bits.rotate_left32(x14, 16)
+		x9 += x14
+		x4 ~= x9
+		x4 = bits.rotate_left32(x4, 12)
+		x3 += x4
+		x14 ~= x3
+		x14 = bits.rotate_left32(x14, 8)
+		x9 += x14
+		x4 ~= x9
+		x4 = bits.rotate_left32(x4, 7)
+	}
+
+	endian.unchecked_put_u32le(dst[0:4], x0)
+	endian.unchecked_put_u32le(dst[4:8], x1)
+	endian.unchecked_put_u32le(dst[8:12], x2)
+	endian.unchecked_put_u32le(dst[12:16], x3)
+	endian.unchecked_put_u32le(dst[16:20], x12)
+	endian.unchecked_put_u32le(dst[20:24], x13)
+	endian.unchecked_put_u32le(dst[24:28], x14)
+	endian.unchecked_put_u32le(dst[28:32], x15)
+}
@@ -0,0 +1,481 @@
+package chacha20_simd128
+
+import "base:intrinsics"
+import "core:crypto/_chacha20"
+import "core:simd"
+@(require) import "core:sys/info"
+
+// Portable 128-bit `core:simd` implementation.
+//
+// This is loosely based on Ted Krovetz's public domain C intrinsic
+// implementation.
+//
+// This is written to perform adequately on any target that has "enough"
+// 128-bit vector registers, the current thought is that 4 blocks at at
+// time is reasonable for amd64, though Ted's code is more conservative.
+//
+// See:
+// supercop-20230530/crypto_stream/chacha20/krovetz/vec128
+
+// Ensure the compiler emits SIMD instructions.  This is a minimum, and
+// setting the microarchitecture at compile time will allow for better
+// code gen when applicable (eg: AVX).  This is somewhat redundant with
+// the default microarchitecture configurations.
+when ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32 {
+	@(private = "file")
+	TARGET_SIMD_FEATURES :: "neon"
+} else when ODIN_ARCH == .amd64 || ODIN_ARCH == .i386 {
+	// Note: LLVM appears to be smart enough to use PSHUFB despite not
+	// explicitly using simd.u8x16 shuffles.
+	@(private = "file")
+	TARGET_SIMD_FEATURES :: "sse2,ssse3"
+} else {
+	@(private = "file")
+	TARGET_SIMD_FEATURES :: ""
+}
+
+@(private = "file")
+_ROT_7L: simd.u32x4 : {7, 7, 7, 7}
+@(private = "file")
+_ROT_7R: simd.u32x4 : {25, 25, 25, 25}
+@(private = "file")
+_ROT_12L: simd.u32x4 : {12, 12, 12, 12}
+@(private = "file")
+_ROT_12R: simd.u32x4 : {20, 20, 20, 20}
+@(private = "file")
+_ROT_8L: simd.u32x4 : {8, 8, 8, 8}
+@(private = "file")
+_ROT_8R: simd.u32x4 : {24, 24, 24, 24}
+@(private = "file")
+_ROT_16: simd.u32x4 : {16, 16, 16, 16}
+
+when ODIN_ENDIAN == .Big {
+	@(private = "file")
+	_increment_counter :: #force_inline proc "contextless" (ctx: ^Context) -> simd.u32x4 {
+		// In the Big Endian case, the low and high portions in the vector
+		// are flipped, so the 64-bit addition can't be done with a simple
+		// vector add.
+		x := &ctx._s
+
+		new_ctr := ((u64(ctx._s[13]) << 32) | u64(ctx._s[12])) + 1
+		x[12] = u32(new_ctr)
+		x[13] = u32(new_ctr >> 32)
+
+		return intrinsics.unaligned_load(transmute(^simd.u32x4)&x[12])
+	}
+
+	// Convert the endian-ness of the components of a u32x4 vector, for
+	// the purposes of output.
+	@(private = "file")
+	_byteswap_u32x4 :: #force_inline proc "contextless" (v: simd.u32x4) -> simd.u32x4 {
+		return(
+			transmute(simd.u32x4)simd.shuffle(
+				transmute(simd.u8x16)v,
+				transmute(simd.u8x16)v,
+				3, 2, 1, 0, 7, 6, 5, 4, 11, 10, 9, 8, 15, 14, 13, 12,
+			)
+		)
+	}
+} else {
+	@(private = "file")
+	_VEC_ONE: simd.u64x2 : {1, 0}
+}
+
+@(private = "file")
+_dq_round_simd128 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3: simd.u32x4,
+) -> (
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV1(b); c = ROTV2(c);  d = ROTV3(d);
+	v1 = simd.shuffle(v1, v1, 1, 2, 3, 0)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1)
+	v3 = simd.shuffle(v3, v3, 3, 0, 1, 2)
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV3(b); c = ROTV2(c); d = ROTV1(d);
+	v1 = simd.shuffle(v1, v1, 3, 0, 1, 2)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1)
+	v3 = simd.shuffle(v3, v3, 1, 2, 3, 0)
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_add_state_simd128 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3, s0, s1, s2, s3: simd.u32x4,
+) -> (
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	v0 = simd.add(v0, s0)
+	v1 = simd.add(v1, s1)
+	v2 = simd.add(v2, s2)
+	v3 = simd.add(v3, s3)
+
+	when ODIN_ENDIAN == .Big {
+		v0 = _byteswap_u32x4(v0)
+		v1 = _byteswap_u32x4(v1)
+		v2 = _byteswap_u32x4(v2)
+		v3 = _byteswap_u32x4(v3)
+	}
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_xor_simd128 :: #force_inline proc "contextless" (
+	src: [^]simd.u32x4,
+	v0, v1, v2, v3: simd.u32x4,
+) -> (
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+	simd.u32x4,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	v0 = simd.bit_xor(v0, intrinsics.unaligned_load((^simd.u32x4)(src[0:])))
+	v1 = simd.bit_xor(v1, intrinsics.unaligned_load((^simd.u32x4)(src[1:])))
+	v2 = simd.bit_xor(v2, intrinsics.unaligned_load((^simd.u32x4)(src[2:])))
+	v3 = simd.bit_xor(v3, intrinsics.unaligned_load((^simd.u32x4)(src[3:])))
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_store_simd128 :: #force_inline proc "contextless" (
+	dst: [^]simd.u32x4,
+	v0, v1, v2, v3: simd.u32x4,
+) {
+	intrinsics.unaligned_store((^simd.u32x4)(dst[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x4)(dst[1:]), v1)
+	intrinsics.unaligned_store((^simd.u32x4)(dst[2:]), v2)
+	intrinsics.unaligned_store((^simd.u32x4)(dst[3:]), v3)
+}
+
+// is_performant returns true iff the target and current host both support
+// "enough" 128-bit SIMD to make this implementation performant.
+is_performant :: proc "contextless" () -> bool {
+	when ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32 || ODIN_ARCH == .amd64 || ODIN_ARCH == .i386 {
+		when ODIN_ARCH == .arm64 || ODIN_ARCH == .arm32 {
+			req_features :: info.CPU_Features{.asimd}
+		} else when ODIN_ARCH == .amd64 || ODIN_ARCH == .i386 {
+			req_features :: info.CPU_Features{.sse2, .ssse3}
+		}
+
+		features, ok := info.cpu_features.?
+		if !ok {
+			return false
+		}
+
+		return features >= req_features
+	} else when ODIN_ARCH == .wasm64p32 || ODIN_ARCH == .wasm32 {
+		return intrinsics.has_target_feature("simd128")
+	} else {
+		return false
+	}
+}
+
+@(enable_target_feature = TARGET_SIMD_FEATURES)
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	_chacha20.check_counter_limit(ctx, nr_blocks)
+
+	dst_v := ([^]simd.u32x4)(raw_data(dst))
+	src_v := ([^]simd.u32x4)(raw_data(src))
+
+	x := &ctx._s
+	n := nr_blocks
+
+	// The state vector is an array of uint32s in native byte-order.
+	x_v := ([^]simd.u32x4)(raw_data(x))
+	s0 := intrinsics.unaligned_load((^simd.u32x4)(x_v[0:]))
+	s1 := intrinsics.unaligned_load((^simd.u32x4)(x_v[1:]))
+	s2 := intrinsics.unaligned_load((^simd.u32x4)(x_v[2:]))
+	s3 := intrinsics.unaligned_load((^simd.u32x4)(x_v[3:]))
+
+	// 8 blocks at a time.
+	//
+	// Note: This is only worth it on Aarch64.
+	when ODIN_ARCH == .arm64 {
+		for ; n >= 8; n = n - 8 {
+			v0, v1, v2, v3 := s0, s1, s2, s3
+
+			when ODIN_ENDIAN == .Little {
+				s7 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s3, _VEC_ONE)
+			} else {
+				s7 := _increment_counter(ctx)
+			}
+			v4, v5, v6, v7 := s0, s1, s2, s7
+
+			when ODIN_ENDIAN == .Little {
+				s11 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s7, _VEC_ONE)
+			} else {
+				s11 := _increment_counter(ctx)
+			}
+			v8, v9, v10, v11 := s0, s1, s2, s11
+
+			when ODIN_ENDIAN == .Little {
+				s15 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s11, _VEC_ONE)
+			} else {
+				s15 := _increment_counter(ctx)
+			}
+			v12, v13, v14, v15 := s0, s1, s2, s15
+
+			when ODIN_ENDIAN == .Little {
+				s19 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s15, _VEC_ONE)
+			} else {
+				s19 := _increment_counter(ctx)
+			}
+
+			v16, v17, v18, v19 := s0, s1, s2, s19
+			when ODIN_ENDIAN == .Little {
+				s23 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s19, _VEC_ONE)
+			} else {
+				s23 := _increment_counter(ctx)
+			}
+
+			v20, v21, v22, v23 := s0, s1, s2, s23
+			when ODIN_ENDIAN == .Little {
+				s27 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s23, _VEC_ONE)
+			} else {
+				s27 := _increment_counter(ctx)
+			}
+
+			v24, v25, v26, v27 := s0, s1, s2, s27
+			when ODIN_ENDIAN == .Little {
+				s31 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s27, _VEC_ONE)
+			} else {
+				s31 := _increment_counter(ctx)
+			}
+			v28, v29, v30, v31 := s0, s1, s2, s31
+
+			for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+				v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+				v4, v5, v6, v7 = _dq_round_simd128(v4, v5, v6, v7)
+				v8, v9, v10, v11 = _dq_round_simd128(v8, v9, v10, v11)
+				v12, v13, v14, v15 = _dq_round_simd128(v12, v13, v14, v15)
+				v16, v17, v18, v19 = _dq_round_simd128(v16, v17, v18, v19)
+				v20, v21, v22, v23 = _dq_round_simd128(v20, v21, v22, v23)
+				v24, v25, v26, v27 = _dq_round_simd128(v24, v25, v26, v27)
+				v28, v29, v30, v31 = _dq_round_simd128(v28, v29, v30, v31)
+			}
+
+			v0, v1, v2, v3 = _add_state_simd128(v0, v1, v2, v3, s0, s1, s2, s3)
+			v4, v5, v6, v7 = _add_state_simd128(v4, v5, v6, v7, s0, s1, s2, s7)
+			v8, v9, v10, v11 = _add_state_simd128(v8, v9, v10, v11, s0, s1, s2, s11)
+			v12, v13, v14, v15 = _add_state_simd128(v12, v13, v14, v15, s0, s1, s2, s15)
+			v16, v17, v18, v19 = _add_state_simd128(v16, v17, v18, v19, s0, s1, s2, s19)
+			v20, v21, v22, v23 = _add_state_simd128(v20, v21, v22, v23, s0, s1, s2, s23)
+			v24, v25, v26, v27 = _add_state_simd128(v24, v25, v26, v27, s0, s1, s2, s27)
+			v28, v29, v30, v31 = _add_state_simd128(v28, v29, v30, v31, s0, s1, s2, s31)
+
+			#no_bounds_check {
+				if src != nil {
+					v0, v1, v2, v3 = _xor_simd128(src_v, v0, v1, v2, v3)
+					v4, v5, v6, v7 = _xor_simd128(src_v[4:], v4, v5, v6, v7)
+					v8, v9, v10, v11 = _xor_simd128(src_v[8:], v8, v9, v10, v11)
+					v12, v13, v14, v15 = _xor_simd128(src_v[12:], v12, v13, v14, v15)
+					v16, v17, v18, v19 = _xor_simd128(src_v[16:], v16, v17, v18, v19)
+					v20, v21, v22, v23 = _xor_simd128(src_v[20:], v20, v21, v22, v23)
+					v24, v25, v26, v27 = _xor_simd128(src_v[24:], v24, v25, v26, v27)
+					v28, v29, v30, v31 = _xor_simd128(src_v[28:], v28, v29, v30, v31)
+					src_v = src_v[32:]
+				}
+
+				_store_simd128(dst_v, v0, v1, v2, v3)
+				_store_simd128(dst_v[4:], v4, v5, v6, v7)
+				_store_simd128(dst_v[8:], v8, v9, v10, v11)
+				_store_simd128(dst_v[12:], v12, v13, v14, v15)
+				_store_simd128(dst_v[16:], v16, v17, v18, v19)
+				_store_simd128(dst_v[20:], v20, v21, v22, v23)
+				_store_simd128(dst_v[24:], v24, v25, v26, v27)
+				_store_simd128(dst_v[28:], v28, v29, v30, v31)
+				dst_v = dst_v[32:]
+			}
+
+			when ODIN_ENDIAN == .Little {
+				// s31 holds the most current counter, so `s3 = s31 + 1`.
+				s3 = transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s31, _VEC_ONE)
+			} else {
+				s3 = _increment_counter(ctx)
+			}
+		}
+	}
+
+	// 4 blocks at a time.
+	//
+	// Note: The i386 target lacks the required number of registers
+	// for this to be performant, so it is skipped.
+	when ODIN_ARCH != .i386 {
+		for ; n >= 4; n = n - 4 {
+			v0, v1, v2, v3 := s0, s1, s2, s3
+
+			when ODIN_ENDIAN == .Little {
+				s7 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s3, _VEC_ONE)
+			} else {
+				s7 := _increment_counter(ctx)
+			}
+			v4, v5, v6, v7 := s0, s1, s2, s7
+
+			when ODIN_ENDIAN == .Little {
+				s11 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s7, _VEC_ONE)
+			} else {
+				s11 := _increment_counter(ctx)
+			}
+			v8, v9, v10, v11 := s0, s1, s2, s11
+
+			when ODIN_ENDIAN == .Little {
+				s15 := transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s11, _VEC_ONE)
+			} else {
+				s15 := _increment_counter(ctx)
+			}
+			v12, v13, v14, v15 := s0, s1, s2, s15
+
+			for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+				v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+				v4, v5, v6, v7 = _dq_round_simd128(v4, v5, v6, v7)
+				v8, v9, v10, v11 = _dq_round_simd128(v8, v9, v10, v11)
+				v12, v13, v14, v15 = _dq_round_simd128(v12, v13, v14, v15)
+			}
+
+			v0, v1, v2, v3 = _add_state_simd128(v0, v1, v2, v3, s0, s1, s2, s3)
+			v4, v5, v6, v7 = _add_state_simd128(v4, v5, v6, v7, s0, s1, s2, s7)
+			v8, v9, v10, v11 = _add_state_simd128(v8, v9, v10, v11, s0, s1, s2, s11)
+			v12, v13, v14, v15 = _add_state_simd128(v12, v13, v14, v15, s0, s1, s2, s15)
+
+			#no_bounds_check {
+				if src != nil {
+					v0, v1, v2, v3 = _xor_simd128(src_v, v0, v1, v2, v3)
+					v4, v5, v6, v7 = _xor_simd128(src_v[4:], v4, v5, v6, v7)
+					v8, v9, v10, v11 = _xor_simd128(src_v[8:], v8, v9, v10, v11)
+					v12, v13, v14, v15 = _xor_simd128(src_v[12:], v12, v13, v14, v15)
+					src_v = src_v[16:]
+				}
+
+				_store_simd128(dst_v, v0, v1, v2, v3)
+				_store_simd128(dst_v[4:], v4, v5, v6, v7)
+				_store_simd128(dst_v[8:], v8, v9, v10, v11)
+				_store_simd128(dst_v[12:], v12, v13, v14, v15)
+				dst_v = dst_v[16:]
+			}
+
+			when ODIN_ENDIAN == .Little {
+				// s15 holds the most current counter, so `s3 = s15 + 1`.
+				s3 = transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s15, _VEC_ONE)
+			} else {
+				s3 = _increment_counter(ctx)
+			}
+		}
+	}
+
+	// 1 block at a time.
+	for ; n > 0; n = n - 1 {
+		v0, v1, v2, v3 := s0, s1, s2, s3
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+		}
+		v0, v1, v2, v3 = _add_state_simd128(v0, v1, v2, v3, s0, s1, s2, s3)
+
+		#no_bounds_check {
+			if src != nil {
+				v0, v1, v2, v3 = _xor_simd128(src_v, v0, v1, v2, v3)
+				src_v = src_v[4:]
+			}
+
+			_store_simd128(dst_v, v0, v1, v2, v3)
+			dst_v = dst_v[4:]
+		}
+
+		// Increment the counter.  Overflow checking is done upon
+		// entry into the routine, so a 64-bit increment safely
+		// covers both cases.
+		when ODIN_ENDIAN == .Little {
+			s3 = transmute(simd.u32x4)simd.add(transmute(simd.u64x2)s3, _VEC_ONE)
+		} else {
+			s3 = _increment_counter(ctx)
+		}
+	}
+
+	when ODIN_ENDIAN == .Little {
+		// Write back the counter to the state.
+		intrinsics.unaligned_store((^simd.u32x4)(x_v[3:]), s3)
+	}
+}
+
+@(enable_target_feature = TARGET_SIMD_FEATURES)
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	v0 := simd.u32x4{_chacha20.SIGMA_0, _chacha20.SIGMA_1, _chacha20.SIGMA_2, _chacha20.SIGMA_3}
+	v1 := intrinsics.unaligned_load((^simd.u32x4)(&key[0]))
+	v2 := intrinsics.unaligned_load((^simd.u32x4)(&key[16]))
+	v3 := intrinsics.unaligned_load((^simd.u32x4)(&iv[0]))
+
+	when ODIN_ENDIAN == .Big {
+		v1 = _byteswap_u32x4(v1)
+		v2 = _byteswap_u32x4(v2)
+		v3 = _byteswap_u32x4(v3)
+	}
+
+	for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+		v0, v1, v2, v3 = _dq_round_simd128(v0, v1, v2, v3)
+	}
+
+	when ODIN_ENDIAN == .Big {
+		v0 = _byteswap_u32x4(v0)
+		v3 = _byteswap_u32x4(v3)
+	}
+
+	dst_v := ([^]simd.u32x4)(raw_data(dst))
+	intrinsics.unaligned_store((^simd.u32x4)(dst_v[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x4)(dst_v[1:]), v3)
+}
@@ -0,0 +1,319 @@
+//+build amd64
+package chacha20_simd256
+
+import "base:intrinsics"
+import "core:crypto/_chacha20"
+import chacha_simd128 "core:crypto/_chacha20/simd128"
+import "core:simd"
+import "core:sys/info"
+
+// This is loosely based on Ted Krovetz's public domain C intrinsic
+// implementations.  While written using `core:simd`, this is currently
+// amd64 specific because we do not have a way to detect ARM SVE.
+//
+// See:
+// supercop-20230530/crypto_stream/chacha20/krovetz/vec128
+// supercop-20230530/crypto_stream/chacha20/krovetz/avx2
+
+#assert(ODIN_ENDIAN == .Little)
+
+@(private = "file")
+_ROT_7L: simd.u32x8 : {7, 7, 7, 7, 7, 7, 7, 7}
+@(private = "file")
+_ROT_7R: simd.u32x8 : {25, 25, 25, 25, 25, 25, 25, 25}
+@(private = "file")
+_ROT_12L: simd.u32x8 : {12, 12, 12, 12, 12, 12, 12, 12}
+@(private = "file")
+_ROT_12R: simd.u32x8 : {20, 20, 20, 20, 20, 20, 20, 20}
+@(private = "file")
+_ROT_8L: simd.u32x8 : {8, 8, 8, 8, 8, 8, 8, 8}
+@(private = "file")
+_ROT_8R: simd.u32x8 : {24, 24, 24, 24, 24, 24, 24, 24}
+@(private = "file")
+_ROT_16: simd.u32x8 : {16, 16, 16, 16, 16, 16, 16, 16}
+@(private = "file")
+_VEC_ZERO_ONE: simd.u64x4 : {0, 0, 1, 0}
+@(private = "file")
+_VEC_TWO: simd.u64x4 : {2, 0, 2, 0}
+
+// is_performant returns true iff the target and current host both support
+// "enough" SIMD to make this implementation performant.
+is_performant :: proc "contextless" () -> bool {
+	req_features :: info.CPU_Features{.avx, .avx2}
+
+	features, ok := info.cpu_features.?
+	if !ok {
+		return false
+	}
+
+	return features >= req_features
+}
+
+@(private = "file")
+_dq_round_simd256 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV1(b); c = ROTV2(c);  d = ROTV3(d);
+	v1 = simd.shuffle(v1, v1, 1, 2, 3, 0, 5, 6, 7, 4)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1, 6, 7, 4, 5)
+	v3 = simd.shuffle(v3, v3, 3, 0, 1, 2, 7, 4, 5, 6)
+
+	// a += b; d ^= a; d = ROTW16(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_16), simd.shr(v3, _ROT_16))
+
+	// c += d; b ^= c; b = ROTW12(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_12L), simd.shr(v1, _ROT_12R))
+
+	// a += b; d ^= a; d = ROTW8(d);
+	v0 = simd.add(v0, v1)
+	v3 = simd.bit_xor(v3, v0)
+	v3 = simd.bit_xor(simd.shl(v3, _ROT_8L), simd.shr(v3, _ROT_8R))
+
+	// c += d; b ^= c; b = ROTW7(b);
+	v2 = simd.add(v2, v3)
+	v1 = simd.bit_xor(v1, v2)
+	v1 = simd.bit_xor(simd.shl(v1, _ROT_7L), simd.shr(v1, _ROT_7R))
+
+	// b = ROTV3(b); c = ROTV2(c); d = ROTV1(d);
+	v1 = simd.shuffle(v1, v1, 3, 0, 1, 2, 7, 4, 5, 6)
+	v2 = simd.shuffle(v2, v2, 2, 3, 0, 1, 6, 7, 4, 5)
+	v3 = simd.shuffle(v3, v3, 1, 2, 3, 0, 5, 6, 7, 4)
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_add_and_permute_state_simd256 :: #force_inline proc "contextless" (
+	v0, v1, v2, v3, s0, s1, s2, s3: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+) {
+	t0 := simd.add(v0, s0)
+	t1 := simd.add(v1, s1)
+	t2 := simd.add(v2, s2)
+	t3 := simd.add(v3, s3)
+
+	// Big Endian would byteswap here.
+
+	// Each of v0 .. v3 has 128-bits of keystream for 2 separate blocks.
+	// permute the state such that (r0, r1) contains block 0, and (r2, r3)
+	// contains block 1.
+	r0 := simd.shuffle(t0, t1, 0, 1, 2, 3, 8, 9, 10, 11)
+	r2 := simd.shuffle(t0, t1, 4, 5, 6, 7, 12, 13, 14, 15)
+	r1 := simd.shuffle(t2, t3, 0, 1, 2, 3, 8, 9, 10, 11)
+	r3 := simd.shuffle(t2, t3, 4, 5, 6, 7, 12, 13, 14, 15)
+
+	return r0, r1, r2, r3
+}
+
+@(private = "file")
+_xor_simd256 :: #force_inline proc "contextless" (
+	src: [^]simd.u32x8,
+	v0, v1, v2, v3: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+	simd.u32x8,
+) {
+	v0, v1, v2, v3 := v0, v1, v2, v3
+
+	v0 = simd.bit_xor(v0, intrinsics.unaligned_load((^simd.u32x8)(src[0:])))
+	v1 = simd.bit_xor(v1, intrinsics.unaligned_load((^simd.u32x8)(src[1:])))
+	v2 = simd.bit_xor(v2, intrinsics.unaligned_load((^simd.u32x8)(src[2:])))
+	v3 = simd.bit_xor(v3, intrinsics.unaligned_load((^simd.u32x8)(src[3:])))
+
+	return v0, v1, v2, v3
+}
+
+@(private = "file")
+_xor_simd256_x1 :: #force_inline proc "contextless" (
+	src: [^]simd.u32x8,
+	v0, v1: simd.u32x8,
+) -> (
+	simd.u32x8,
+	simd.u32x8,
+) {
+	v0, v1 := v0, v1
+
+	v0 = simd.bit_xor(v0, intrinsics.unaligned_load((^simd.u32x8)(src[0:])))
+	v1 = simd.bit_xor(v1, intrinsics.unaligned_load((^simd.u32x8)(src[1:])))
+
+	return v0, v1
+}
+
+@(private = "file")
+_store_simd256 :: #force_inline proc "contextless" (
+	dst: [^]simd.u32x8,
+	v0, v1, v2, v3: simd.u32x8,
+) {
+	intrinsics.unaligned_store((^simd.u32x8)(dst[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[1:]), v1)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[2:]), v2)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[3:]), v3)
+}
+
+@(private = "file")
+_store_simd256_x1 :: #force_inline proc "contextless" (
+	dst: [^]simd.u32x8,
+	v0, v1: simd.u32x8,
+) {
+	intrinsics.unaligned_store((^simd.u32x8)(dst[0:]), v0)
+	intrinsics.unaligned_store((^simd.u32x8)(dst[1:]), v1)
+}
+
+@(enable_target_feature = "sse2,ssse3,avx,avx2")
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	// Enforce the maximum consumed keystream per IV.
+	_chacha20.check_counter_limit(ctx, nr_blocks)
+
+	dst_v := ([^]simd.u32x8)(raw_data(dst))
+	src_v := ([^]simd.u32x8)(raw_data(src))
+
+	x := &ctx._s
+	n := nr_blocks
+
+	// The state vector is an array of uint32s in native byte-order.
+	// Setup s0 .. s3 such that each register stores 2 copies of the
+	// state.
+	x_v := ([^]simd.u32x4)(raw_data(x))
+	t0 := intrinsics.unaligned_load((^simd.u32x4)(x_v[0:]))
+	t1 := intrinsics.unaligned_load((^simd.u32x4)(x_v[1:]))
+	t2 := intrinsics.unaligned_load((^simd.u32x4)(x_v[2:]))
+	t3 := intrinsics.unaligned_load((^simd.u32x4)(x_v[3:]))
+	s0 := simd.swizzle(t0, 0, 1, 2, 3, 0, 1, 2, 3)
+	s1 := simd.swizzle(t1, 0, 1, 2, 3, 0, 1, 2, 3)
+	s2 := simd.swizzle(t2, 0, 1, 2, 3, 0, 1, 2, 3)
+	s3 := simd.swizzle(t3, 0, 1, 2, 3, 0, 1, 2, 3)
+
+	// Advance the counter in the 2nd copy of the state by one.
+	s3 = transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s3, _VEC_ZERO_ONE)
+
+	// 8 blocks at a time.
+	for ; n >= 8; n = n - 8 {
+		v0, v1, v2, v3 := s0, s1, s2, s3
+
+		s7 := transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s3, _VEC_TWO)
+		v4, v5, v6, v7 := s0, s1, s2, s7
+
+		s11 := transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s7, _VEC_TWO)
+		v8, v9, v10, v11 := s0, s1, s2, s11
+
+		s15 := transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s11, _VEC_TWO)
+		v12, v13, v14, v15 := s0, s1, s2, s15
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			v0, v1, v2, v3 = _dq_round_simd256(v0, v1, v2, v3)
+			v4, v5, v6, v7 = _dq_round_simd256(v4, v5, v6, v7)
+			v8, v9, v10, v11 = _dq_round_simd256(v8, v9, v10, v11)
+			v12, v13, v14, v15 = _dq_round_simd256(v12, v13, v14, v15)
+		}
+
+		v0, v1, v2, v3 = _add_and_permute_state_simd256(v0, v1, v2, v3, s0, s1, s2, s3)
+		v4, v5, v6, v7 = _add_and_permute_state_simd256(v4, v5, v6, v7, s0, s1, s2, s7)
+		v8, v9, v10, v11 = _add_and_permute_state_simd256(v8, v9, v10, v11, s0, s1, s2, s11)
+		v12, v13, v14, v15 = _add_and_permute_state_simd256(v12, v13, v14, v15, s0, s1, s2, s15)
+
+		#no_bounds_check {
+			if src != nil {
+				v0, v1, v2, v3 = _xor_simd256(src_v, v0, v1, v2, v3)
+				v4, v5, v6, v7 = _xor_simd256(src_v[4:], v4, v5, v6, v7)
+				v8, v9, v10, v11 = _xor_simd256(src_v[8:], v8, v9, v10, v11)
+				v12, v13, v14, v15 = _xor_simd256(src_v[12:], v12, v13, v14, v15)
+				src_v = src_v[16:]
+			}
+
+			_store_simd256(dst_v, v0, v1, v2, v3)
+			_store_simd256(dst_v[4:], v4, v5, v6, v7)
+			_store_simd256(dst_v[8:], v8, v9, v10, v11)
+			_store_simd256(dst_v[12:], v12, v13, v14, v15)
+			dst_v = dst_v[16:]
+		}
+
+		s3 = transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s15, _VEC_TWO)
+	}
+
+
+	// 2 (or 1) block at a time.
+	for ; n > 0; n = n - 2 {
+		v0, v1, v2, v3 := s0, s1, s2, s3
+
+		for i := _chacha20.ROUNDS; i > 0; i = i - 2 {
+			v0, v1, v2, v3 = _dq_round_simd256(v0, v1, v2, v3)
+		}
+		v0, v1, v2, v3 = _add_and_permute_state_simd256(v0, v1, v2, v3, s0, s1, s2, s3)
+
+		if n == 1 {
+			// Note: No need to advance src_v, dst_v, or increment the counter
+			// since this is guaranteed to be the final block.
+			#no_bounds_check {
+				if src != nil {
+					v0, v1 = _xor_simd256_x1(src_v, v0, v1)
+				}
+
+				_store_simd256_x1(dst_v, v0, v1)
+			}
+			break
+		}
+
+		#no_bounds_check {
+			if src != nil {
+				v0, v1, v2, v3 = _xor_simd256(src_v, v0, v1, v2, v3)
+				src_v = src_v[4:]
+			}
+
+			_store_simd256(dst_v, v0, v1, v2, v3)
+			dst_v = dst_v[4:]
+		}
+
+		s3 = transmute(simd.u32x8)simd.add(transmute(simd.u64x4)s3, _VEC_TWO)
+	}
+
+	// Write back the counter.  Doing it this way, saves having to
+	// pull out the correct counter value from s3.
+	new_ctr := ((u64(ctx._s[13]) << 32) | u64(ctx._s[12])) + u64(nr_blocks)
+	ctx._s[12] = u32(new_ctr)
+	ctx._s[13] = u32(new_ctr >> 32)
+}
+
+@(enable_target_feature = "sse2,ssse3,avx")
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	// We can just enable AVX and call the simd128 code as going
+	// wider has 0 performance benefit, but VEX encoded instructions
+	// is nice.
+	#force_inline chacha_simd128.hchacha20(dst, key, iv)
+}
@@ -0,0 +1,17 @@
+//+build !amd64
+package chacha20_simd256
+
+import "base:intrinsics"
+import "core:crypto/_chacha20"
+
+is_performant :: proc "contextless" () -> bool {
+	return false
+}
+
+stream_blocks :: proc(ctx: ^_chacha20.Context, dst, src: []byte, nr_blocks: int) {
+	panic("crypto/chacha20: simd256 implementation unsupported")
+}
+
+hchacha20 :: proc "contextless" (dst, key, iv: []byte) {
+	intrinsics.trap()
+}
@@ -110,7 +110,7 @@ ge_set_bytes :: proc "contextless" (ge: ^Group_Element, b: []byte) -> bool {
 	if len(b) != 32 {
 		intrinsics.trap()
 	}
-	b_ := transmute(^[32]byte)(raw_data(b))
+	b_ := (^[32]byte)(raw_data(b))

 	// Do the work in a scratch element, so that ge is unchanged on
 	// failure.
@@ -169,7 +169,7 @@ ge_bytes :: proc "contextless" (ge: ^Group_Element, dst: []byte) {
 	if len(dst) != 32 {
 		intrinsics.trap()
 	}
-	dst_ := transmute(^[32]byte)(raw_data(dst))
+	dst_ := (^[32]byte)(raw_data(dst))

 	// Convert the element to affine (x, y) representation.
 	x, y, z_inv: field.Tight_Field_Element = ---, ---, ---
@@ -28,7 +28,7 @@ sc_set_bytes :: proc "contextless" (sc: ^Scalar, b: []byte) -> bool {
 	if len(b) != 32 {
 		intrinsics.trap()
 	}
-	b_ := transmute(^[32]byte)(raw_data(b))
+	b_ := (^[32]byte)(raw_data(b))
 	return field.fe_from_bytes(sc, b_)
 }

@@ -36,7 +36,7 @@ sc_set_bytes_rfc8032 :: proc "contextless" (sc: ^Scalar, b: []byte) {
 	if len(b) != 32 {
 		intrinsics.trap()
 	}
-	b_ := transmute(^[32]byte)(raw_data(b))
+	b_ := (^[32]byte)(raw_data(b))
 	field.fe_from_bytes_rfc8032(sc, b_)
 }

@@ -6,13 +6,13 @@ import "core:mem"
 fe_relax_cast :: #force_inline proc "contextless" (
 	arg1: ^Tight_Field_Element,
 ) -> ^Loose_Field_Element {
-	return transmute(^Loose_Field_Element)(arg1)
+	return (^Loose_Field_Element)(arg1)
 }

 fe_tighten_cast :: #force_inline proc "contextless" (
 	arg1: ^Loose_Field_Element,
 ) -> ^Tight_Field_Element {
-	return transmute(^Tight_Field_Element)(arg1)
+	return (^Tight_Field_Element)(arg1)
 }

 fe_clear :: proc "contextless" (
--- a/Show More
+++ b/Show More