ci: publish a draft release from the artifacts that CI already built

The Linux, macOS and Windows workflows build every file that a release needs when the release branch is pushed. The new Release workflow finds those runs for the tagged commit, waits for them, and collects their nine files into a draft release. It builds nothing, so what ships is what was tested. The release notes come from the CHANGELOG.md entry for that version. Signed-off-by: Ron Evans <ron@hybridgroup.com>
Puya PY32F MCU support (#5106 )
2026-08-29 08:49:04 +00:00 · 2026-08-29 09:55:07 +02:00 · 2026-08-29 08:41:43 +02:00 · 2026-08-28 19:48:48 +02:00 · 2026-08-28 17:42:58 +02:00 · 2026-08-28 11:27:07 +02:00
1123 changed files with 58598 additions and 10374 deletions
@@ -1,127 +0,0 @@
-version: 2.1
-
-commands:
-  submodules:
-    steps:
-      - run:
-          name: "Pull submodules"
-          command: git submodule update --init
-  llvm-source-linux:
-    steps:
-      - restore_cache:
-          keys:
-            - llvm-source-19-v1
-      - run:
-          name: "Fetch LLVM source"
-          command: make llvm-source
-      - save_cache:
-          key: llvm-source-19-v1
-          paths:
-            - llvm-project/clang/lib/Headers
-            - llvm-project/clang/include
-            - llvm-project/compiler-rt
-            - llvm-project/lld/include
-            - llvm-project/llvm/include
-  hack-ninja-jobs:
-    steps:
-      - run:
-          name: "Hack Ninja to use less jobs"
-          command: |
-            echo -e '#!/bin/sh\n/usr/bin/ninja -j3 "$@"' > /go/bin/ninja
-            chmod +x /go/bin/ninja
-  build-binaryen-linux:
-    steps:
-      - restore_cache:
-          keys:
-            - binaryen-linux-v3
-      - run:
-          name: "Build Binaryen"
-          command: |
-            make binaryen
-      - save_cache:
-          key: binaryen-linux-v3
-          paths:
-            - build/wasm-opt
-  test-linux:
-    parameters:
-      llvm:
-        type: string
-      fmt-check:
-        type: boolean
-        default: true
-    steps:
-      - checkout
-      - submodules
-      - run:
-          name: "Install apt dependencies"
-          command: |
-            echo 'deb https://apt.llvm.org/bullseye/ llvm-toolchain-bullseye-<<parameters.llvm>> main' > /etc/apt/sources.list.d/llvm.list
-            wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key | apt-key add -
-            apt-get update
-            apt-get install --no-install-recommends -y \
-                llvm-<<parameters.llvm>>-dev \
-                clang-<<parameters.llvm>> \
-                libclang-<<parameters.llvm>>-dev \
-                lld-<<parameters.llvm>> \
-                cmake \
-                ninja-build
-      - hack-ninja-jobs
-      - build-binaryen-linux
-      - restore_cache:
-          keys:
-            - go-cache-v4-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_PREVIOUS_BUILD_NUM }}
-            - go-cache-v4-{{ checksum "go.mod" }}
-      - llvm-source-linux
-      - run: go install -tags=llvm<<parameters.llvm>> .
-      - restore_cache:
-          keys:
-            - wasi-libc-sysroot-systemclang-v7
-      - run: make wasi-libc
-      - save_cache:
-          key: wasi-libc-sysroot-systemclang-v7
-          paths:
-            - lib/wasi-libc/sysroot
-      - when:
-          condition: <<parameters.fmt-check>>
-          steps:
-          - run:
-              # Do this before gen-device so that it doesn't check the
-              # formatting of generated files.
-              name: Check Go code formatting
-              command: make fmt-check lint
-      - run: make gen-device -j4
-      # TODO: change this to -skip='TestErrors|TestWasm' with Go 1.20
-      - run: go test -tags=llvm<<parameters.llvm>> -short -run='TestBuild|TestTest|TestGetList|TestTraceback'
-      - run: make smoketest XTENSA=0
-      - save_cache:
-          key: go-cache-v4-{{ checksum "go.mod" }}-{{ .Environment.CIRCLE_BUILD_NUM }}
-          paths:
-            - ~/.cache/go-build
-            - /go/pkg/mod
-
-jobs:
-  test-llvm15-go119:
-    docker:
-      - image: golang:1.19-bullseye
-    steps:
-      - test-linux:
-          llvm: "15"
-          # "make lint" fails before go 1.21 because internal/tools/go.mod specifies packages that require go 1.21
-          fmt-check: false
-    resource_class: large
-  test-llvm19-go124:
-    docker:
-      - image: golang:1.24-bullseye
-    steps:
-      - test-linux:
-          llvm: "19"
-    resource_class: large
-
-workflows:
-  test-all:
-    jobs:
-      # This tests our lowest supported versions of Go and LLVM, to make sure at
-      # least the smoke tests still pass.
-      - test-llvm15-go119
-      # This tests LLVM 19 support when linking against system libraries.
-      - test-llvm19-go124
@@ -1,5 +1,4 @@
 build/
 llvm-*/
 .github
-.circleci

@@ -16,36 +16,37 @@ jobs:
    name: build-macos
    strategy:
      matrix:
-        # macos-13: amd64 (oldest supported version as of 18-10-2024)
-        # macos-14: arm64 (oldest arm64 version)
-        os: [macos-13, macos-14]
+        # macos-14: arm64 (oldest supported version as of 18-11-2025)
+        # macos-15-intel: amd64 (last intel version to be supported by github runners)
+        # See https://github.com/actions/runner-images/issues/13046
+        os: [macos-14, macos-15-intel]
        include:
-          - os: macos-13
-            goarch: amd64
          - os: macos-14
            goarch: arm64
+          - os: macos-15-intel
+            goarch: amd64
    runs-on: ${{ matrix.os }}
    steps:
      - name: Install Dependencies
        run: |
          HOMEBREW_NO_AUTO_UPDATE=1 brew install qemu binaryen
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          submodules: true
      - name: Extract TinyGo version
        id: version
        run: ./.github/workflows/tinygo-extract-version.sh | tee -a "$GITHUB_OUTPUT"
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Restore LLVM source cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-source
        with:
-          key: llvm-source-19-${{ matrix.os }}-v1
+          key: llvm-source-22-${{ matrix.os }}-v1
          path: |
            llvm-project/clang/lib/Headers
            llvm-project/clang/include
@@ -56,7 +57,7 @@ jobs:
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        run: make llvm-source
      - name: Save LLVM source cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-source.outputs.cache-primary-key }}
@@ -67,10 +68,10 @@ jobs:
            llvm-project/lld/include
            llvm-project/llvm/include
      - name: Restore LLVM build cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-build
        with:
-          key: llvm-build-19-${{ matrix.os }}-v1
+          key: llvm-build-22-${{ matrix.os }}-v1
          path: llvm-build
      - name: Build LLVM
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
@@ -79,29 +80,20 @@ jobs:
          rm -rf llvm-project
          make llvm-source
          # install dependencies
-          HOMEBREW_NO_AUTO_UPDATE=1 brew install cmake ninja
+          HOMEBREW_NO_AUTO_UPDATE=1 brew install ninja
          # build!
          make llvm-build
          find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
      - name: Save LLVM build cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-build.outputs.cache-primary-key }}
          path: llvm-build
-      - name: Cache wasi-libc sysroot
-        uses: actions/cache@v4
-        id: cache-wasi-libc
-        with:
-          key: wasi-libc-sysroot-${{ matrix.os }}-v1
-          path: lib/wasi-libc/sysroot
-      - name: Build wasi-libc
-        if: steps.cache-wasi-libc.outputs.cache-hit != 'true'
-        run: make wasi-libc
      - name: make gen-device
        run: make -j3 gen-device
      - name: Test TinyGo
-        run: make test GOTESTFLAGS="-short"
+        run: make test GOTESTFLAGS="-only-current-os"
      - name: Build TinyGo release tarball
        run: make release -j3
      - name: Test stdlib packages
@@ -109,27 +101,22 @@ jobs:
      - name: Make release artifact
        run: cp -p build/release.tar.gz build/tinygo${{ steps.version.outputs.version }}.darwin-${{ matrix.goarch }}.tar.gz
      - name: Publish release artifact
-        # Note: this release artifact is double-zipped, see:
-        # https://github.com/actions/upload-artifact/issues/39
-        # We can essentially pick one of these:
-        # - have a double-zipped artifact when downloaded from the UI
-        # - have a very slow artifact upload
-        # We're doing the former here, to keep artifact uploads fast.
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v7
        with:
          name: darwin-${{ matrix.goarch }}-double-zipped-${{ steps.version.outputs.version }}
          path: build/tinygo${{ steps.version.outputs.version }}.darwin-${{ matrix.goarch }}.tar.gz
+          archive: false
      - name: Smoke tests
-        run: make smoketest TINYGO=$(PWD)/build/tinygo
+        run: make smoketest-quick TINYGO=$(PWD)/build/tinygo
  test-macos-homebrew:
    name: homebrew-install
    runs-on: macos-latest
    strategy:
      matrix:
-        version: [16, 17, 18, 19]
+        version: [16, 17, 18, 19, 20]
    steps:
      - name: Set up Homebrew
-        uses: Homebrew/actions/setup-homebrew@master
+        uses: Homebrew/actions/setup-homebrew@main
      - name: Fix Python symlinks
        run: |
          # Github runners have broken symlinks, so relink
@@ -138,20 +125,21 @@ jobs:
      - name: Install LLVM
        run: |
          brew install llvm@${{ matrix.version }}
+          brew link llvm@${{ matrix.version }}
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Build TinyGo (LLVM ${{ matrix.version }})
        run: go install -tags=llvm${{ matrix.version }}
      - name: Check binary
        run: tinygo version
      - name: Build TinyGo (default LLVM)
-        if: matrix.version == 19
+        if: matrix.version == 20
        run: go install
      - name: Check binary
-        if: matrix.version == 19
+        if: matrix.version == 20
        run: tinygo version
@@ -0,0 +1,72 @@
+# This CI job checks whether at least the smoke tests pass for the oldest
+# Go/LLVM version we claim to support.
+
+name: Version compatibility test
+
+on:
+  pull_request:
+  push:
+    branches:
+      - dev
+      - release
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  test-compat:
+    runs-on: ubuntu-22.04 # this must be a specific version for the apt install below
+    env:
+      # Oldest versions currently supported by TinyGo
+      LLVM: "15"
+      Go: "1.25" # when updating this, also update minorMin in builder/config.go
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          submodules: true
+      - name: Install Go
+        uses: actions/setup-go@v6
+        with:
+          go-version: ${{ env.Go }}
+          cache: true
+      - name: Install LLVM
+        run: |
+          echo 'deb https://apt.llvm.org/jammy/ llvm-toolchain-jammy-${{ env.LLVM }} main' | sudo tee /etc/apt/sources.list.d/llvm.list
+          wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key | sudo apt-key add -
+          sudo apt-get update
+          sudo apt-get install --no-install-recommends -y \
+              llvm-${{ env.LLVM }}-dev \
+              clang-${{ env.LLVM }} \
+              libclang-${{ env.LLVM }}-dev \
+              lld-${{ env.LLVM }} \
+              binaryen
+      - name: Restore LLVM source cache
+        uses: actions/cache/restore@v5
+        id: cache-llvm-source
+        with:
+          key: llvm-source-22-linux-compat
+          path: llvm-project/compiler-rt
+      - name: Download LLVM source
+        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
+        run: make llvm-source
+      - name: Save LLVM source cache
+        uses: actions/cache/save@v5
+        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
+        with:
+          key: ${{ steps.cache-llvm-source.outputs.cache-primary-key }}
+          path: llvm-project/compiler-rt
+      - name: Go cache
+        uses: actions/cache@v5
+        with:
+          path: |
+            ~/.cache/go-build
+            ~/go/pkg/mod
+          key: go-build-${{ env.Go }}-llvm${{ env.LLVM }}-${{ hashFiles('go.sum') }}
+      - name: Build TinyGo
+        run: go install -tags=llvm${{ env.LLVM }}
+      - run: tinygo version
+      - run: make gen-device -j4
+      - run: go test -tags=llvm${{ env.LLVM }} -short -skip=TestErrors
+      - run: make smoketest-quick XTENSA=0
@@ -31,14 +31,14 @@ jobs:
          sudo rm -rf /usr/local/share/boost
          df -h
      - name: Check out the repo
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          submodules: recursive
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v4
      - name: Docker meta
        id: meta
-        uses: docker/metadata-action@v5
+        uses: docker/metadata-action@v6
        with:
          images: |
            tinygo/tinygo-dev
@@ -47,18 +47,18 @@ jobs:
            type=sha,format=long
            type=raw,value=latest
      - name: Log in to Docker Hub
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
        with:
          username: ${{ secrets.DOCKER_HUB_USERNAME }}
          password: ${{ secrets.DOCKER_HUB_ACCESS_TOKEN }}
      - name: Log in to Github Container Registry
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
        with:
          registry: ghcr.io
          username: ${{ github.actor }}
          password: ${{ secrets.GITHUB_TOKEN }}
      - name: Build and push
-        uses: docker/build-push-action@v5
+        uses: docker/build-push-action@v7
        with:
          context: .
          push: true
@@ -66,45 +66,3 @@ jobs:
          labels: ${{ steps.meta.outputs.labels }}
          cache-from: type=gha
          cache-to: type=gha,mode=max
-      - name: Trigger Drivers repo build on Github Actions
-        run: |
-          curl -X POST \
-          -H "Authorization: Bearer ${{secrets.GHA_ACCESS_TOKEN}}" \
-          -H "Accept: application/vnd.github.v3+json" \
-          https://api.github.com/repos/tinygo-org/drivers/actions/workflows/build.yml/dispatches \
-          -d '{"ref": "dev"}'
-      - name: Trigger Bluetooth repo build on Github Actions
-        run: |
-          curl -X POST \
-          -H "Authorization: Bearer ${{secrets.GHA_ACCESS_TOKEN}}" \
-          -H "Accept: application/vnd.github.v3+json" \
-          https://api.github.com/repos/tinygo-org/bluetooth/actions/workflows/linux.yml/dispatches \
-          -d '{"ref": "dev"}'
-      - name: Trigger TinyFS repo build on Github Actions
-        run: |
-          curl -X POST \
-          -H "Authorization: Bearer ${{secrets.GHA_ACCESS_TOKEN}}" \
-          -H "Accept: application/vnd.github.v3+json" \
-          https://api.github.com/repos/tinygo-org/tinyfs/actions/workflows/build.yml/dispatches \
-          -d '{"ref": "dev"}'
-      - name: Trigger TinyFont repo build on Github Actions
-        run: |
-          curl -X POST \
-          -H "Authorization: Bearer ${{secrets.GHA_ACCESS_TOKEN}}" \
-          -H "Accept: application/vnd.github.v3+json" \
-          https://api.github.com/repos/tinygo-org/tinyfont/actions/workflows/build.yml/dispatches \
-          -d '{"ref": "dev"}'
-      - name: Trigger TinyDraw repo build on Github Actions
-        run: |
-          curl -X POST \
-          -H "Authorization: Bearer ${{secrets.GHA_ACCESS_TOKEN}}" \
-          -H "Accept: application/vnd.github.v3+json" \
-          https://api.github.com/repos/tinygo-org/tinydraw/actions/workflows/build.yml/dispatches \
-          -d '{"ref": "dev"}'
-      - name: Trigger TinyTerm repo build on Github Actions
-        run: |
-          curl -X POST \
-          -H "Authorization: Bearer ${{secrets.GHA_ACCESS_TOKEN}}" \
-          -H "Accept: application/vnd.github.v3+json" \
-          https://api.github.com/repos/tinygo-org/tinyterm/actions/workflows/build.yml/dispatches \
-          -d '{"ref": "dev"}'
@@ -0,0 +1,35 @@
+#!/bin/sh
+
+# Print the CHANGELOG.md entry for one version, to use as release notes.
+# The file uses a setext heading: the bare version, then a line of dashes.
+
+set -e
+
+version="$1"
+if [ -z "$version" ]; then
+	echo "usage: $0 <version>" >&2
+	exit 1
+fi
+
+notes=$(awk -v version="$version" '
+	found {
+		if ($0 ~ /^---+$/ && previous != "") { previous = ""; exit }
+		if (previous != "") print previous
+		previous = $0
+		next
+	}
+	$0 ~ /^---+$/ && previous == version {
+		found = 1
+		previous = ""
+		next
+	}
+	{ previous = $0 }
+	END { if (found && previous != "") print previous }
+' CHANGELOG.md)
+
+if [ -z "$notes" ]; then
+	echo "no CHANGELOG.md entry for version $version" >&2
+	exit 1
+fi
+
+echo "$notes"
@@ -12,43 +12,61 @@ concurrency:
  cancel-in-progress: true

 jobs:
+  go-mod-tidy:
+    # Check that go.sum is up to date.
+    runs-on: ubuntu-slim
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          submodules: true
+      - name: Install Go
+        uses: actions/setup-go@v6
+        with:
+          go-version: '1.27.0'
+          cache: true
+      - name: Run go mod tidy
+        run: go mod tidy
+      - name: Check go.mod and go.sum are up to date
+        run: git diff --exit-code
+
  build-linux:
    # Build Linux binaries, ready for release.
    # This runs inside an Alpine Linux container so we can more easily create a
    # statically linked binary.
    runs-on: ubuntu-latest
    container:
-      image: golang:1.24-alpine
+      image: golang:1.27-alpine
    outputs:
      version: ${{ steps.version.outputs.version }}
    steps:
      - name: Install apk dependencies
-        # tar: needed for actions/cache@v4
+        # tar: needed for actions/cache@v5
        # git+openssh: needed for checkout (I think?)
        # ruby: needed to install fpm
-        run: apk add tar git openssh make g++ ruby-dev
+        run: apk add tar git openssh make g++ ruby-dev mold
      - name: Work around CVE-2022-24765
        # We're not on a multi-user machine, so this is safe.
        run: git config --global --add safe.directory "$GITHUB_WORKSPACE"
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          submodules: true
      - name: Extract TinyGo version
        id: version
        run: ./.github/workflows/tinygo-extract-version.sh | tee -a "$GITHUB_OUTPUT"
      - name: Cache Go
-        uses: actions/cache@v4
+        uses: actions/cache@v5
        with:
-          key: go-cache-linux-alpine-v1-${{ hashFiles('go.mod') }}
+          key: go-cache-linux-alpine-v2-${{ hashFiles('go.mod') }}
          path: |
            ~/.cache/go-build
            ~/go/pkg/mod
      - name: Restore LLVM source cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-source
        with:
-          key: llvm-source-19-linux-alpine-v1
+          key: llvm-source-22-linux-alpine-v1
          path: |
            llvm-project/clang/lib/Headers
            llvm-project/clang/include
@@ -59,7 +77,7 @@ jobs:
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        run: make llvm-source
      - name: Save LLVM source cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-source.outputs.cache-primary-key }}
@@ -70,10 +88,10 @@ jobs:
            llvm-project/lld/include
            llvm-project/llvm/include
      - name: Restore LLVM build cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-build
        with:
-          key: llvm-build-19-linux-alpine-v1
+          key: llvm-build-22-linux-alpine-v1
          path: llvm-build
      - name: Build LLVM
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
@@ -88,31 +106,22 @@ jobs:
          # Remove unnecessary object files (to reduce cache size).
          find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
      - name: Save LLVM build cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-build.outputs.cache-primary-key }}
          path: llvm-build
      - name: Cache Binaryen
-        uses: actions/cache@v4
+        uses: actions/cache@v5
        id: cache-binaryen
        with:
-          key: binaryen-linux-alpine-v1
+          key: binaryen-linux-alpine-v3
          path: build/wasm-opt
      - name: Build Binaryen
        if: steps.cache-binaryen.outputs.cache-hit != 'true'
        run: |
          apk add cmake samurai python3
          make binaryen STATIC=1
-      - name: Cache wasi-libc
-        uses: actions/cache@v4
-        id: cache-wasi-libc
-        with:
-          key: wasi-libc-sysroot-linux-alpine-v2
-          path: lib/wasi-libc/sysroot
-      - name: Build wasi-libc
-        if: steps.cache-wasi-libc.outputs.cache-hit != 'true'
-        run: make wasi-libc
      - name: Install fpm
        run: |
          gem install --version 4.0.7 public_suffix
@@ -127,26 +136,31 @@ jobs:
          make release deb -j3 STATIC=1
          cp -p build/release.tar.gz /tmp/tinygo${{ steps.version.outputs.version }}.linux-amd64.tar.gz
          cp -p build/release.deb    /tmp/tinygo_${{ steps.version.outputs.version }}_amd64.deb
-      - name: Publish release artifact
-        uses: actions/upload-artifact@v4
+      - name: "Publish release artifact: tarball"
+        uses: actions/upload-artifact@v7
        with:
-          name: linux-amd64-double-zipped-${{ steps.version.outputs.version }}
-          path: |
-            /tmp/tinygo${{ steps.version.outputs.version }}.linux-amd64.tar.gz
-            /tmp/tinygo_${{ steps.version.outputs.version }}_amd64.deb
+          name: tinygo${{ steps.version.outputs.version }}.linux-amd64.tar.gz
+          path: /tmp/tinygo${{ steps.version.outputs.version }}.linux-amd64.tar.gz
+          archive: false
+      - name: "Publish release artifact: Debian package"
+        uses: actions/upload-artifact@v7
+        with:
+          name: tinygo_${{ steps.version.outputs.version }}_amd64.deb
+          path: /tmp/tinygo_${{ steps.version.outputs.version }}_amd64.deb
+          archive: false
  test-linux-build:
    # Test the binaries built in the build-linux job by running the smoke tests.
    runs-on: ubuntu-latest
    needs: build-linux
    steps:
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          submodules: true
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Install wasmtime
        uses: bytecodealliance/actions/wasmtime/setup@v1
@@ -155,9 +169,9 @@ jobs:
      - name: Install wasm-tools
        uses: bytecodealliance/actions/wasm-tools/setup@v1
      - name: Download release artifact
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v8
        with:
-          name: linux-amd64-double-zipped-${{ needs.build-linux.outputs.version }}
+          name: tinygo${{ needs.build-linux.outputs.version }}.linux-amd64.tar.gz
      - name: Extract release tarball
        run: |
          mkdir -p ~/lib
@@ -166,14 +180,50 @@ jobs:
      - run: make tinygo-test-wasip1-fast
      - run: make tinygo-test-wasip2-fast
      - run: make tinygo-test-wasm
-      - run: make smoketest
+  smoketest-linux:
+    # Build a binary for every supported board, using the binaries built in the
+    # build-linux job. This is the only job that runs the full smoke test. The
+    # result does not depend on the host OS, so the other jobs build a
+    # representative board for each architecture with smoketest-quick.
+    runs-on: ubuntu-latest
+    needs: build-linux
+    strategy:
+      fail-fast: false
+      matrix:
+        # Balanced with the measured build time of each group, so that the
+        # shards finish at about the same time.
+        group:
+          - smoketest-rp2xxx smoketest-selftest smoketest-wasm-sim smoketest-py32
+          - smoketest-esp smoketest-riscv smoketest-flags
+          - smoketest-examples smoketest-nxp smoketest-pwm-usb smoketest-avr
+          - smoketest-samd smoketest-stm32 smoketest-nrf smoketest-wasm
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          submodules: true
+      - name: Install Go
+        uses: actions/setup-go@v6
+        with:
+          go-version: '1.27.0'
+          cache: true
+      - name: Download release artifact
+        uses: actions/download-artifact@v8
+        with:
+          name: tinygo${{ needs.build-linux.outputs.version }}.linux-amd64.tar.gz
+      - name: Extract release tarball
+        run: |
+          mkdir -p ~/lib
+          tar -C ~/lib -xf tinygo${{ needs.build-linux.outputs.version }}.linux-amd64.tar.gz
+          ln -s ~/lib/tinygo/bin/tinygo ~/go/bin/tinygo
+      - run: make testchdir ${{ matrix.group }}
  assert-test-linux:
    # Run all tests that can run on Linux, with LLVM assertions enabled to catch
    # potential bugs.
    runs-on: ubuntu-latest
    steps:
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          submodules: true
      - name: Install apt dependencies
@@ -182,31 +232,43 @@ jobs:
          cat /proc/cpuinfo
          sudo apt-get update
          sudo apt-get install --no-install-recommends \
+              libgcrypt20 \
+              libpixman-1-0 \
+              libsdl2-2.0-0 \
+              libslirp0 \
              qemu-system-arm \
              qemu-system-riscv32 \
              qemu-user \
              simavr \
              ninja-build
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Install Node.js
-        uses: actions/setup-node@v4
+        uses: actions/setup-node@v6
        with:
-          node-version: '18'
+          node-version: '22'
      - name: Install wasmtime
        uses: bytecodealliance/actions/wasmtime/setup@v1
        with:
          version: "29.0.1"
      - name: Setup `wasm-tools`
        uses: bytecodealliance/actions/wasm-tools/setup@v1
+      - name: Install Espressif QEMU
+        run: |
+          release=esp-develop-9.2.2-20260417
+          archive=qemu-xtensa-softmmu-${release//-/_}-x86_64-linux-gnu.tar.xz
+          curl -fL --retry 3 -o "$RUNNER_TEMP/$archive" \
+            "https://github.com/espressif/qemu/releases/download/$release/$archive"
+          tar -xJf "$RUNNER_TEMP/$archive" -C "$RUNNER_TEMP"
+          echo "$RUNNER_TEMP/qemu/bin" >> "$GITHUB_PATH"
      - name: Restore LLVM source cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-source
        with:
-          key: llvm-source-19-linux-asserts-v1
+          key: llvm-source-22-linux-asserts-v1
          path: |
            llvm-project/clang/lib/Headers
            llvm-project/clang/include
@@ -217,7 +279,7 @@ jobs:
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        run: make llvm-source
      - name: Save LLVM source cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-source.outputs.cache-primary-key }}
@@ -228,10 +290,10 @@ jobs:
            llvm-project/lld/include
            llvm-project/llvm/include
      - name: Restore LLVM build cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-build
        with:
-          key: llvm-build-19-linux-asserts-v1
+          key: llvm-build-22-linux-asserts-v1
          path: llvm-build
      - name: Build LLVM
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
@@ -244,13 +306,13 @@ jobs:
          # Remove unnecessary object files (to reduce cache size).
          find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
      - name: Save LLVM build cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-build.outputs.cache-primary-key }}
          path: llvm-build
      - name: Cache Binaryen
-        uses: actions/cache@v4
+        uses: actions/cache@v5
        id: cache-binaryen
        with:
          key: binaryen-linux-asserts-v1
@@ -258,15 +320,6 @@ jobs:
      - name: Build Binaryen
        if: steps.cache-binaryen.outputs.cache-hit != 'true'
        run: make binaryen
-      - name: Cache wasi-libc
-        uses: actions/cache@v4
-        id: cache-wasi-libc
-        with:
-          key: wasi-libc-sysroot-linux-asserts-v6
-          path: lib/wasi-libc/sysroot
-      - name: Build wasi-libc
-        if: steps.cache-wasi-libc.outputs.cache-hit != 'true'
-        run: make wasi-libc
      - run: make gen-device -j4
      - name: Test TinyGo
        run: make ASSERT=1 test
@@ -276,9 +329,11 @@ jobs:
          echo "$(pwd)/build" >> $GITHUB_PATH
      - name: Test stdlib packages
        run: make tinygo-test
-      - run: make smoketest
+      - run: make smoketest-quick
      - run: make wasmtest
      - run: make tinygo-test-baremetal
+      - name: Check Go code formatting
+        run: make fmt-check lint
  build-linux-cross:
    # Build ARM Linux binaries, ready for release.
    # This intentionally uses an older Linux image, so that we compile against
@@ -302,7 +357,7 @@ jobs:
    needs: build-linux
    steps:
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Get TinyGo version
        id: version
        run: ./.github/workflows/tinygo-extract-version.sh | tee -a "$GITHUB_OUTPUT"
@@ -314,15 +369,15 @@ jobs:
              g++-${{ matrix.toolchain }} \
              libc6-dev-${{ matrix.libc }}-cross
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Restore LLVM source cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-source
        with:
-          key: llvm-source-19-linux-v1
+          key: llvm-source-22-linux-v1
          path: |
            llvm-project/clang/lib/Headers
            llvm-project/clang/include
@@ -333,7 +388,7 @@ jobs:
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        run: make llvm-source
      - name: Save LLVM source cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-source.outputs.cache-primary-key }}
@@ -344,10 +399,10 @@ jobs:
            llvm-project/lld/include
            llvm-project/llvm/include
      - name: Restore LLVM build cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-build
        with:
-          key: llvm-build-19-linux-${{ matrix.goarch }}-v3
+          key: llvm-build-22-linux-${{ matrix.goarch }}-v1
          path: llvm-build
      - name: Build LLVM
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
@@ -362,16 +417,16 @@ jobs:
          # Remove unnecessary object files (to reduce cache size).
          find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
      - name: Save LLVM build cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-build.outputs.cache-primary-key }}
          path: llvm-build
      - name: Cache Binaryen
-        uses: actions/cache@v4
+        uses: actions/cache@v5
        id: cache-binaryen
        with:
-          key: binaryen-linux-${{ matrix.goarch }}-v4
+          key: binaryen-linux-${{ matrix.goarch }}-v5
          path: build/wasm-opt
      - name: Build Binaryen
        if: steps.cache-binaryen.outputs.cache-hit != 'true'
@@ -388,9 +443,9 @@ jobs:
        run: |
          make CROSS=${{ matrix.toolchain }}
      - name: Download amd64 release
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v8
        with:
-          name: linux-amd64-double-zipped-${{ needs.build-linux.outputs.version }}
+          name: tinygo${{ needs.build-linux.outputs.version }}.linux-amd64.tar.gz
      - name: Extract amd64 release
        run: |
          mkdir -p build/release
@@ -402,12 +457,17 @@ jobs:
      - name: Create ${{ matrix.goarch }} release
        run: |
          make release deb RELEASEONLY=1 DEB_ARCH=${{ matrix.libc }}
-          cp -p build/release.tar.gz /tmp/tinygo${{ steps.version.outputs.version }}.linux-${{ matrix.goarch }}.tar.gz
-          cp -p build/release.deb    /tmp/tinygo_${{ steps.version.outputs.version }}_${{ matrix.libc }}.deb
-      - name: Publish release artifact
-        uses: actions/upload-artifact@v4
+          cp -p build/release.tar.gz /tmp/tinygo${{ needs.build-linux.outputs.version }}.linux-${{ matrix.goarch }}.tar.gz
+          cp -p build/release.deb    /tmp/tinygo_${{ needs.build-linux.outputs.version }}_${{ matrix.libc }}.deb
+      - name: "Publish release artifact: tarball"
+        uses: actions/upload-artifact@v7
        with:
-          name: linux-${{ matrix.goarch }}-double-zipped-${{ steps.version.outputs.version }}
-          path: |
-            /tmp/tinygo${{ steps.version.outputs.version }}.linux-${{ matrix.goarch }}.tar.gz
-            /tmp/tinygo_${{ steps.version.outputs.version }}_${{ matrix.libc }}.deb
+          name: linux-${{ matrix.goarch }}-double-zipped-${{ needs.build-linux.outputs.version }}
+          path: /tmp/tinygo${{ needs.build-linux.outputs.version }}.linux-${{ matrix.goarch }}.tar.gz
+          archive: false
+      - name: "Publish release artifact: Debian package"
+        uses: actions/upload-artifact@v7
+        with:
+          name: linux-${{ matrix.goarch }}-double-zipped-${{ needs.build-linux.outputs.version }}
+          path: /tmp/tinygo_${{ needs.build-linux.outputs.version }}_${{ matrix.libc }}.deb
+          archive: false
@@ -25,18 +25,18 @@ jobs:
      contents: read
    steps:
      - name: Check out the repo
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          submodules: recursive
      - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@v3
+        uses: docker/setup-buildx-action@v4
      - name: Docker meta
        id: meta
-        uses: docker/metadata-action@v5
+        uses: docker/metadata-action@v6
        with:
          images: |
-            tinygo/llvm-19
-            ghcr.io/${{ github.repository_owner }}/llvm-19
+            tinygo/llvm-20
+            ghcr.io/${{ github.repository_owner }}/llvm-20
          tags: |
            type=sha,format=long
            type=raw,value=latest
@@ -46,13 +46,13 @@ jobs:
          username: ${{ secrets.DOCKER_HUB_USERNAME }}
          password: ${{ secrets.DOCKER_HUB_ACCESS_TOKEN }}
      - name: Log in to Github Container Registry
-        uses: docker/login-action@v3
+        uses: docker/login-action@v4
        with:
          registry: ghcr.io
          username: ${{ github.actor }}
          password: ${{ secrets.GITHUB_TOKEN }}
      - name: Build and push
-        uses: docker/build-push-action@v5
+        uses: docker/build-push-action@v7
        with:
          target: tinygo-llvm-build
          context: .
@@ -21,28 +21,28 @@ jobs:
        # See: https://github.com/tinygo-org/tinygo/pull/4516#issuecomment-2416363668
        run: sudo apt-get remove llvm-18
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Pull musl, bdwgc
        run: |
            git submodule update --init lib/musl lib/bdwgc
      - name: Restore LLVM source cache
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-source
        with:
-          key: llvm-source-19-linux-nix-v1
+          key: llvm-source-22-linux-nix-v1
          path: |
            llvm-project/compiler-rt
      - name: Download LLVM source
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        run: make llvm-source
      - name: Save LLVM source cache
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-source.outputs.cache-primary-key }}
          path: |
            llvm-project/compiler-rt
-      - uses: cachix/install-nix-action@v22
+      - uses: cachix/install-nix-action@v31
      - name: Test
        run: |
          nix develop --ignore-environment --keep HOME --command bash -c "go install && ~/go/bin/tinygo version && ~/go/bin/tinygo build -o test ./testdata/cgo"
@@ -0,0 +1,138 @@
+# Publish a GitHub release from the artifacts that CI already built.
+#
+# The Linux, macOS and Windows workflows build every file that a release needs
+# when the release branch is pushed. This workflow collects the artifacts of
+# those runs for the tagged commit, so what ships is what was tested.
+#
+# The release is a draft, so the notes can be reviewed before publication.
+name: Release
+
+on:
+  push:
+    tags:
+      - 'v*'
+  workflow_dispatch:
+    inputs:
+      tag:
+        description: 'Tag to release, for example v0.42.0'
+        required: true
+
+concurrency:
+  group: release-${{ inputs.tag || github.ref_name }}
+  cancel-in-progress: false
+
+permissions:
+  contents: write
+
+jobs:
+  release:
+    runs-on: ubuntu-latest
+    env:
+      GH_TOKEN: ${{ github.token }}
+      GH_REPO: ${{ github.repository }}
+      TAG: ${{ inputs.tag || github.ref_name }}
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          ref: ${{ inputs.tag || github.ref_name }}
+          fetch-depth: 0
+      - name: Read the version
+        id: version
+        # The release file names come from goenv/version.go, not from the tag,
+        # so the two must agree.
+        run: |
+          version=$(./.github/workflows/tinygo-extract-version.sh | cut -d= -f2-)
+          case "$version" in
+            *-dev)
+              echo "::error::goenv/version.go has development version $version"
+              exit 1
+              ;;
+          esac
+          if [ "v$version" != "$TAG" ]; then
+            echo "::error::tag $TAG does not match version $version in goenv/version.go"
+            exit 1
+          fi
+          echo "version=$version" >> "$GITHUB_OUTPUT"
+      - name: Wait for the CI runs of this commit
+        id: runs
+        # A pull request run has the merge commit as its head, so look at push
+        # runs only.
+        run: |
+          sha=$(git rev-parse HEAD)
+          ids=
+          for workflow in linux.yml build-macos.yml windows.yml; do
+            id=
+            for _ in $(seq 20); do
+              id=$(gh run list --workflow "$workflow" --commit "$sha" --event push --limit 1 --json databaseId --jq '.[0].databaseId')
+              if [ -n "$id" ]; then
+                break
+              fi
+              echo "waiting for $workflow to start on $sha"
+              sleep 30
+            done
+            if [ -z "$id" ]; then
+              echo "::error::no $workflow run for commit $sha"
+              exit 1
+            fi
+            echo "$workflow: run $id"
+            gh run watch "$id" --exit-status > /dev/null || true
+            conclusion=$(gh run view "$id" --json conclusion --jq .conclusion)
+            if [ "$conclusion" != "success" ]; then
+              echo "::error::$workflow run $id concluded with $conclusion"
+              exit 1
+            fi
+            ids="$ids $id"
+          done
+          echo "ids=$ids" >> "$GITHUB_OUTPUT"
+      - name: Download the artifacts
+        env:
+          RUN_IDS: ${{ steps.runs.outputs.ids }}
+        run: |
+          for id in $RUN_IDS; do
+            gh run download "$id" --dir artifacts
+          done
+      - name: Collect the release files
+        # A build that stopped uploading must not give a half complete release,
+        # so list the files that are expected. The search is by file name, not
+        # by artifact name, because the artifact names are not uniform.
+        env:
+          VERSION: ${{ steps.version.outputs.version }}
+        run: |
+          version=$VERSION
+          mkdir -p dist
+          missing=0
+          for file in \
+              "tinygo$version.linux-amd64.tar.gz"  "tinygo_${version}_amd64.deb" \
+              "tinygo$version.linux-arm.tar.gz"    "tinygo_${version}_armhf.deb" \
+              "tinygo$version.linux-arm64.tar.gz"  "tinygo_${version}_arm64.deb" \
+              "tinygo$version.darwin-amd64.tar.gz" \
+              "tinygo$version.darwin-arm64.tar.gz" \
+              "tinygo$version.windows-amd64.zip"; do
+            found=$(find artifacts -type f -name "$file" | head -1)
+            if [ -z "$found" ]; then
+              echo "::error::missing release file $file"
+              missing=1
+            else
+              mv "$found" "dist/$file"
+            fi
+          done
+          ls -l dist
+          exit $missing
+      - name: Extract the release notes
+        env:
+          VERSION: ${{ steps.version.outputs.version }}
+        run: ./.github/workflows/extract-changelog.sh "$VERSION" > notes.md
+      - name: Create the draft release
+        env:
+          VERSION: ${{ steps.version.outputs.version }}
+        run: |
+          version=$VERSION
+          set --
+          case "$version" in
+            *-*) set -- --prerelease ;;
+          esac
+          gh release create "$TAG" --draft --verify-tag "$@" \
+            --title "$version" \
+            --notes-file notes.md \
+            dist/*
@@ -2,11 +2,11 @@
 # still works after checking out the dev branch (that is, when going from LLVM
 # 16 to LLVM 17 for example, both Clang 16 and Clang 17 are installed).

-echo 'deb https://apt.llvm.org/noble/ llvm-toolchain-noble-19 main' | sudo tee /etc/apt/sources.list.d/llvm.list
+echo 'deb https://apt.llvm.org/noble/ llvm-toolchain-noble-20 main' | sudo tee /etc/apt/sources.list.d/llvm.list
 wget -O - https://apt.llvm.org/llvm-snapshot.gpg.key | sudo apt-key add -
 sudo apt-get update
 sudo apt-get install --no-install-recommends -y \
-    llvm-19-dev \
-    clang-19 \
-    libclang-19-dev \
-    lld-19
+    llvm-20-dev \
+    clang-20 \
+    libclang-20-dev \
+    lld-20
@@ -19,25 +19,29 @@ jobs:
      - name: Add GOBIN to $PATH
        run: |
          echo "$HOME/go/bin" >> $GITHUB_PATH
+      - name: Install Go
+        uses: actions/setup-go@v6
+        with:
+          go-version: '~1.25'
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          fetch-depth: 0 # fetch all history (no sparse checkout)
          submodules: true
      - name: Install apt dependencies
        run: ./.github/workflows/sizediff-install-pkgs.sh
      - name: Restore LLVM source cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
        id: cache-llvm-source
        with:
-          key: llvm-source-19-sizediff-v1
+          key: llvm-source-22-sizediff-v1
          path: |
            llvm-project/compiler-rt
      - name: Download LLVM source
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        run: make llvm-source
      - name: Cache Go
-        uses: actions/cache@v4
+        uses: actions/cache@v5
        with:
          key: go-cache-linux-sizediff-v2-${{ hashFiles('go.mod') }}
          path: |
@@ -17,21 +17,14 @@ jobs:
    outputs:
      version: ${{ steps.version.outputs.version }}
    steps:
-      - name: Configure pagefile
-        uses: al-cheb/configure-pagefile-action@v1.4
-        with:
-          minimum-size: 8GB
-          maximum-size: 24GB
-          disk-root: "C:"
-      - uses: brechtm/setup-scoop@v2
-        with:
-          scoop_update: 'false'
+      - uses: MinoruSekine/setup-scoop@v4
      - name: Install Dependencies
        shell: bash
        run: |
+          scoop config use_external_7zip true
          scoop install ninja binaryen
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
        with:
          submodules: true
      - name: Extract TinyGo version
@@ -39,15 +32,15 @@ jobs:
        shell: bash
        run: ./.github/workflows/tinygo-extract-version.sh | tee -a "$GITHUB_OUTPUT"
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Restore cached LLVM source
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-source
        with:
-          key: llvm-source-19-windows-v1
+          key: llvm-source-22-windows-v3
          path: |
            llvm-project/clang/lib/Headers
            llvm-project/clang/include
@@ -58,7 +51,7 @@ jobs:
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        run: make llvm-source
      - name: Save cached LLVM source
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-source.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-source.outputs.cache-primary-key }}
@@ -69,10 +62,10 @@ jobs:
            llvm-project/lld/include
            llvm-project/llvm/include
      - name: Restore cached LLVM build
-        uses: actions/cache/restore@v4
+        uses: actions/cache/restore@v5
        id: cache-llvm-build
        with:
-          key: llvm-build-19-windows-v1
+          key: llvm-build-22-windows-v1
          path: llvm-build
      - name: Build LLVM
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
@@ -86,35 +79,27 @@ jobs:
          # Remove unnecessary object files (to reduce cache size).
          find llvm-build -name CMakeFiles -prune -exec rm -r '{}' \;
      - name: Save cached LLVM build
-        uses: actions/cache/save@v4
+        uses: actions/cache/save@v5
        if: steps.cache-llvm-build.outputs.cache-hit != 'true'
        with:
          key: ${{ steps.cache-llvm-build.outputs.cache-primary-key }}
          path: llvm-build
-      - name: Cache wasi-libc sysroot
-        uses: actions/cache@v4
-        id: cache-wasi-libc
-        with:
-          key: wasi-libc-sysroot-v5
-          path: lib/wasi-libc/sysroot
-      - name: Build wasi-libc
-        if: steps.cache-wasi-libc.outputs.cache-hit != 'true'
-        run: make wasi-libc
      - name: Cache Go cache
-        uses: actions/cache@v4
+        uses: actions/cache@v5
        with:
-          key: go-cache-windows-v1-${{ hashFiles('go.mod') }}
+          key: go-cache-windows-v3-${{ hashFiles('go.mod') }}
          path: |
            C:/Users/runneradmin/AppData/Local/go-build
            C:/Users/runneradmin/go/pkg/mod
      - name: Install wasmtime
        run: |
+          scoop config use_external_7zip true
          scoop install wasmtime@29.0.1
      - name: make gen-device
        run: make -j3 gen-device
      - name: Test TinyGo
        shell: bash
-        run: make test GOTESTFLAGS="-short"
+        run: make test GOTESTFLAGS="-only-current-os"
      - name: Build TinyGo release tarball
        shell: bash
        run: make build/release -j4
@@ -123,80 +108,56 @@ jobs:
        working-directory: build/release
        run: 7z -tzip a tinygo${{ steps.version.outputs.version }}.windows-amd64.zip tinygo
      - name: Publish release artifact
-        # Note: this release artifact is double-zipped, see:
-        # https://github.com/actions/upload-artifact/issues/39
-        # We can essentially pick one of these:
-        # - have a dobule-zipped artifact when downloaded from the UI
-        # - have a very slow artifact upload
-        # We're doing the former here, to keep artifact uploads fast.
-        uses: actions/upload-artifact@v4
+        uses: actions/upload-artifact@v7
        with:
-          name: windows-amd64-double-zipped-${{ steps.version.outputs.version }}
+          name: tinygo${{ steps.version.outputs.version }}.windows-amd64.zip
          path: build/release/tinygo${{ steps.version.outputs.version }}.windows-amd64.zip
+          archive: false

  smoke-test-windows:
    runs-on: windows-2022
    needs: build-windows
    steps:
-      - name: Configure pagefile
-        uses: al-cheb/configure-pagefile-action@v1.4
-        with:
-          minimum-size: 8GB
-          maximum-size: 24GB
-          disk-root: "C:"
-      - uses: brechtm/setup-scoop@v2
-        with:
-          scoop_update: 'false'
+      - uses: MinoruSekine/setup-scoop@v4
      - name: Install Dependencies
        shell: bash
        run: |
+          scoop config use_external_7zip true
          scoop install binaryen
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Download TinyGo build
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v8
        with:
-          name: windows-amd64-double-zipped-${{ needs.build-windows.outputs.version }}
+          name: tinygo${{ needs.build-windows.outputs.version }}.windows-amd64.zip
          path: build/
-      - name: Unzip TinyGo build
-        shell: bash
-        working-directory: build
-        run: 7z x tinygo*.windows-amd64.zip -r
+        # This build is already unzipped.
      - name: Smoke tests
        shell: bash
-        run: make smoketest TINYGO=$(PWD)/build/tinygo/bin/tinygo
+        run: make smoketest-quick TINYGO=$(PWD)/build/tinygo/bin/tinygo

  stdlib-test-windows:
    runs-on: windows-2022
    needs: build-windows
    steps:
-      - name: Configure pagefile
-        uses: al-cheb/configure-pagefile-action@v1.4
-        with:
-          minimum-size: 8GB
-          maximum-size: 24GB
-          disk-root: "C:"
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Download TinyGo build
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v8
        with:
-          name: windows-amd64-double-zipped-${{ needs.build-windows.outputs.version }}
+          name: tinygo${{ needs.build-windows.outputs.version }}.windows-amd64.zip
          path: build/
-      - name: Unzip TinyGo build
-        shell: bash
-        working-directory: build
-        run: 7z x tinygo*.windows-amd64.zip -r
+        # This build is already unzipped.
      - name: Test stdlib packages
        run: make tinygo-test TINYGO=$(PWD)/build/tinygo/bin/tinygo

@@ -204,34 +165,24 @@ jobs:
    runs-on: windows-2022
    needs: build-windows
    steps:
-      - name: Configure pagefile
-        uses: al-cheb/configure-pagefile-action@v1.4
-        with:
-          minimum-size: 8GB
-          maximum-size: 24GB
-          disk-root: "C:"
-      - uses: brechtm/setup-scoop@v2
-        with:
-          scoop_update: 'false'
+      - uses: MinoruSekine/setup-scoop@v4
      - name: Install Dependencies
        shell: bash
        run: |
-          scoop install binaryen && scoop install wasmtime@29.0.1
+          scoop config use_external_7zip true
+          scoop install binaryen wasmtime@29.0.1
      - name: Checkout
-        uses: actions/checkout@v4
+        uses: actions/checkout@v6
      - name: Install Go
-        uses: actions/setup-go@v5
+        uses: actions/setup-go@v6
        with:
-          go-version: '1.24'
+          go-version: '1.27.0'
          cache: true
      - name: Download TinyGo build
-        uses: actions/download-artifact@v4
+        uses: actions/download-artifact@v8
        with:
-          name: windows-amd64-double-zipped-${{ needs.build-windows.outputs.version }}
+          name: tinygo${{ needs.build-windows.outputs.version }}.windows-amd64.zip
          path: build/
-      - name: Unzip TinyGo build
-        shell: bash
-        working-directory: build
-        run: 7z x tinygo*.windows-amd64.zip -r
+        # This build is already unzipped.
      - name: Test stdlib packages on wasip1
        run: make tinygo-test-wasip1-fast TINYGO=$(PWD)/build/tinygo/bin/tinygo
@@ -24,6 +24,8 @@ src/device/renesas/*.go
 src/device/renesas/*.s
 src/device/rp/*.go
 src/device/rp/*.s
+src/device/py32/*.go
+src/device/py32/*.s
 ./vendor
 llvm-build
 llvm-project
@@ -16,13 +16,13 @@
 	url = https://github.com/WebAssembly/wasi-libc
 [submodule "lib/picolibc"]
 	path = lib/picolibc
-	url = https://github.com/keith-packard/picolibc.git
+	url = https://github.com/picolibc/picolibc.git
 [submodule "lib/stm32-svd"]
 	path = lib/stm32-svd
 	url = https://github.com/tinygo-org/stm32-svd
 [submodule "lib/musl"]
 	path = lib/musl
-	url = git://git.musl-libc.org/musl
+	url = https://github.com/tinygo-org/musl-libc.git
 [submodule "lib/binaryen"]
 	path = lib/binaryen
 	url = https://github.com/WebAssembly/binaryen.git
@@ -35,10 +35,12 @@
 [submodule "src/net"]
 	path = src/net
 	url = https://github.com/tinygo-org/net.git
-	branch = dev
 [submodule "lib/wasi-cli"]
 	path = lib/wasi-cli
 	url = https://github.com/WebAssembly/wasi-cli
 [submodule "lib/bdwgc"]
 	path = lib/bdwgc
 	url = https://github.com/ivmai/bdwgc.git
+[submodule "lib/py32-svd"]
+	path = lib/py32-svd
+	url = https://github.com/tinygo-org/py32-svd.git
@@ -0,0 +1,10 @@
+## Communication style
+Use ASD-STE-100 Simplified Technical English when you speak to the operator.
+
+## Comments
+Use ASD-STE100 Simplified Technical English for content outside of the code like comments, PR description, PR title, github comments, and issues.
+
+Generally omit extra redundant comments. If a comment is absolutely needed, always keep it brief, no more than 2 lines. Avoid extra use of colons, semicolons, and dashes. When comments are required due to some specific requirements, include a reference such as to a datasheet or other definitive source to explain why this is the case, including the chapter/section number, page number, and/or URL.
+
+## No coding tool attributions
+Never add "created by XXX" or any other attributions from coding tools to any PRs, issues, comments, or anywhere else.
@@ -30,7 +30,10 @@ on a different system like Mac.

 ## Using GNU Make

-The static build of TinyGo is driven by GNUmakefile, which provides a help target for quick reference:
+The static build of TinyGo is driven by GNUmakefile, which includes the topic
+files in the `make/` directory (`config.mk`, `llvm.mk`, `gen-device.mk`,
+`build.mk`, `test.mk`, `smoketest.mk`, `release.mk`, and `tools.mk`).
+It provides a help target for quick reference:

    % make help
    clean                           Remove build directory
@@ -64,7 +67,7 @@ while producing binaries that are about as fast.
    export CC=clang
    export CXX=clang++

-The Makefile includes a default configuration that is good for most users. It
+`make/config.mk` holds a default configuration that is good for most users. It
 builds a release version of LLVM (optimized, no asserts) and includes all
 targets supported by TinyGo:

@@ -85,11 +88,17 @@ Try running TinyGo:

    ./build/tinygo help

-Also, make sure the `tinygo` binary really is statically linked. Check this
-using `ldd` (not to be confused with `lld`):
+Also, make sure the `tinygo` binary really is statically linked. The command to check for 
+dynamic dependencies differs depending on your operating system.
+
+On Linux, use `ldd` (not to be confused with `lld`):

    ldd ./build/tinygo

+On macOS, use otool -L:
+
+    otool -L ./build/tinygo
+
 The result should not contain libclang or libLLVM.

 ## Make a release tarball
@@ -110,3 +119,35 @@ the following command (for example in ~/lib):
 TinyGo will get extracted to a `tinygo` directory. You can then call it with:

    ./tinygo/bin/tinygo
+
+## Publish a release
+
+The `Release` workflow (`.github/workflows/release.yml`) publishes releases. It
+does not build anything. The Linux, macOS and Windows workflows already build
+every file that a release needs when the `release` branch is pushed, so the
+release workflow collects the artifacts of those runs for the tagged commit.
+What ships is what was tested.
+
+ 1. On the `dev` branch, set `const version` in `goenv/version.go` to the new
+    version (without a `v` prefix), and add the entry to `CHANGELOG.md`.
+ 2. Merge `dev` into the `release` branch.
+ 3. Tag that commit and push the tag:
+
+        git tag v0.42.0
+        git push origin v0.42.0
+
+    The tag must be `v` plus the version in `goenv/version.go`, because the
+    release file names come from that constant.
+ 4. The workflow waits for the Linux, macOS and Windows runs of the tagged
+    commit, collects their nine files, and creates a **draft** release. The
+    release notes come from the `CHANGELOG.md` entry for that version.
+ 5. Review the draft release and publish it.
+ 6. On the `dev` branch, set `goenv/version.go` to the next `-dev` version.
+
+To release again after a failure, delete the draft release and start the
+workflow from the Actions tab with the tag as its input.
+
+GitHub keeps a SHA-256 digest of every published file. The digest is not shown
+on the release page, but it can be printed with:
+
+    gh release view v0.42.0 --json assets --jq '.assets[] | "\(.digest)  \(.name)"'
@@ -1,3 +1,411 @@
+0.41.1
+---
+* **machine**
+  - esp32c3: correct pin interrupt setup call that was overlooked from #5320
+* **runtime**
+  - esp32s3: wait for TIMG0 update register to clear before reading timer registers
+* **net**
+  - update net module to a version that is backwards compatible with Go 1.25.x to fix #5332
+
+0.41.0
+---
+* **general**
+  - go.mod, compiler: bump minimum Go version to 1.23
+  - builder: support Go 1.26 now
+  - feat: add inheritable-only field to filter processor-level targets (#5270)
+  - feature: add esp32flash flash method for esp32s3/esp32c3/esp32/esp8266
+  - flashing: introduce flash method 'esp32jtag' for esp32c3/esp32s3 targets
+  - fashing: add "adb" flash method using Android Debug Bridge
+  - main: auto-use best available flashing options on esp32
+  - espflasher: update to espflasher 0.6.0
+* **compiler**
+  - implement method-set based AssignableTo and Implements (#5304)
+  - simplify createObjectLayout
+  - implement copy directly
+  - fix min/max on floats by using intrinsics
+  - add element layout to sliceAppend
+  - add intrinsic for crypto/internal/constanttime.boolToUint8
+  - fix SyscallN handling for Windows on Go 1.26+
+* **core**
+  - reflect: add TypeAssert
+  - reflect: fix TypeAssert for structs
+  - sync: add WaitGroup.Go
+  - os: add UserCacheDir and UserConfigDir
+  - os: implement Lchown function for changing file ownership (#5161)
+  - testing: add b.ReportMetric()
+  - errors: add errors package to passing list
+  - internal/gclayout: use correct lengths
+  - internal/abi: add EscapeNonString, EscapeToResultNonString
+  - internal, runtime: add internal fips bool
+  - internal/itoa: resurrect from Go stdlib
+  - internal/syscall/unix: implement GetRandom for WASI targets
+  - interp: make ptrtoint size mismatch a recoverable error
+  - loader: support "all:" prefix in //go:embed patterns
+  - loader, crypto/internal/entropy: fix 32 MiB RAM overflow on microcontrollers
+  - loader, unicode/utf8: fix hiBits overflow on 16-bit AVR targets
+  - builder: order embedded files deterministically
+  - builder: fix panic in findPackagePath when using -size short
+  - builder: fix SIGSEGV when stripping duplicate function definitions
+  - builder, runtime: fix duplicate symbol error with Go 1.26+ on Windows
+  - builder: use target-specific -march for RISC-V library compilation
+  - transform: modify output format from the -print-allocs flag to match the go coverage tool format
+  - cgo: allow for changes to 'short-enums' flag
+  - wasm: add more stubbed file operations
+* **machine**
+  - esp32c3: implement BlockDevice for esp32c3 flash
+  - esp32c3: clear GPIO STATUS before dispatching pin callbacks
+  - esp32c3: implement USBDevice using interrupts
+  - esp32c3: add Enable stub for QEMU test target
+  - esp32c3: fix interrupt disable handling
+  - esp32c3: clear MCAUSE after handling interrupt
+  - esp32c3: map missing IRAM sections in linker script
+  - esp32c3: fix to use PROVIDE() for WiFi/BLE ROM function addresses to allow blob overrides
+  - esp32c3: add WiFi/BLE ROM linker support
+  - esp32c3: use TIMG0 alarm interrupt for sleepTicks instead of busy-waiting
+  - esp32s3: implement Pin.SetInterrupt for GPIO pin change interrupts
+  - esp32s3: use edge-triggered CPU interrupt for GPIO pin interrupts
+  - esp32s3: add interrupt support (#5244)
+  - esp32s3: switch USB implementation to use interrupts instead of polling
+  - esp32s3: replace inline ISR with full interrupt vector handler
+  - esp32s3: use TIMG0 alarm interrupt for sleepTicks
+  - esp32s3: improve exception handlers, add procPin/procUnpin, and linker wrap flags
+  - esp32s3: fix Xtensa register window spill using recursive call4 in task switching
+  - esp32s3: update linker script and boot assembly for multi-page flash XIP mapping
+  - esp32s3: add flash XIP boot assembly with cache/MMU init
+  - esp32s3: implement SPI (#5169)
+  - esp32s3: implement I2C interface (#5210)
+  - esp32s3: implement RNG based on onboard hardware random number generator
+  - esp32s3,esp32c3: add USB serial support
+  - esp32s3,esp32c3: add txStalled flag to skip USB serial spin when no host
+  - esp32s3,esp32c3: make USB Serial/JTAG writes non-blocking when FIFO is full
+  - esp32c3/esp32s3: refactor ADC implementation to reduce code duplication
+  - esp32s3,esp32c3: add ADC support (#5231)
+  - esp32s3,esp32c3: add PWM support (#5215)
+  - esp32c3/esp32s3: refactoring and corrections for SPI implementation
+  - esp32xx: add WASM simulator support for ESP32-C3/ESP32-S3 targets
+  - esp32: default SPI CS pin to NoPin when unset
+  - esp: fix tinygo_scanCurrentStack to spill register windows
+  - stm32: fix UART interrupt storm caused by uncleared overrun error
+  - stm32: fix PWM problem due to register shifting
+  - stm32: add STM32U585 target definition and runtime
+  - stm32: add STM32U5 GPIO, pin interrupts, and UART support
+  - stm32: add STM32U5 I2C support
+  - stm32: add STM32U5 SPI support
+  - stm32u585: implement ADC
+  - stm32g0b1: add support (#5150)
+  - stm32g0b1: add Watchdog + ADC support (#5158)
+  - stm32l0: add flash support
+  - stm32l0x1,l0x2: TIM: adapt to 32-bit register access
+  - stm32g0: sync with updated stm32 device files
+  - attiny85: add USI-based SPI support (#5181)
+  - attiny85: add PWM support (#5171)
+  - rp: add Close function to UART to allow for removing all system resources/power usage
+  - rp: use blockReset() and unresetBlockWait() helper functions for peripheral reset/unreset
+  - rp2: prevent PWM Period method overflow
+  - rp2: add per-byte timeout budget for I2C (#5189)
+  - feather-m0: export UART0 and pins
+  - usb/cdc: better ring buffer implementation (#5209)
+  - fix: replace ! with ~ for register mask
+  - atmega328p_simulator: add correct build tag for arduino_uno after renaming
+* **net**
+  - update to Go 1.26.2 net package with UDP and JS improvements
+* **runtime**
+  - implement fminimum/fmaximum
+  - make timeoffset atomic
+  - add MemStats.HeapObjects
+  - handle GODEBUG on wasip2 (#5312)
+  - fix Darwin syscall return value truncation and lost arguments
+  - add syscall.runtimeClearenv for Go 1.26
+  - split syscall functions for darwin changes for 1.26
+  - fix: init heap before random number seed on wasm platforms
+* **targets**
+  - add esp32s3-supermini board target
+  - add support for Seeedstudio Xiao-RP2350 board
+  - add Shrike Lite board (#5170)
+  - rename arduino target to arduino-uno
+  - add pins/setup for Arduino UNO Q board QWIIC connector
+  - correct pin mapping for Arduino UNO Q STM32 MCU
+  - add stm32u585 openocd commands for flashing on Arduino UNO Q
+  - correct name/tag use for esp32s3-wroom1 board
+  - modify esp32, esp32s3, esp32c3, & esp8266 targets to use built-in esp32flash
+  - switch rp2040/rp2350 to default to tasks scheduler
+  - change default stack size for esp32c3 and esp32s3 to 8196
+  - swan: rename group build tag stm32l4x5 => stm32l4y5 to avoid clash with stm32 device name
+  - nucleo-f722ze: add build-tag stm32f722
+  - fix: set stm32u5x clock rate to default
+  - create generic targets for esp32 boards
+* **libs**
+  - stm32-svd: update submodule
+  - update tools/gen-device-svd for recent changes to stm32-svd file structure (#5212)
+* **build/test**
+  - build: use Golang 1.26 for all builds
+  - build: actually use the version of llvm we are supposed to be testing by using brew link command
+  - build: let scoop run updates to resolve binaryen install failures on Windows
+  - ci: don't double zip release artifacts
+  - test: increase default timeout from 1 to 2 minutes to prevent spurious CI fails
+  - testdata: fix flaky timers test by draining ticker channel after Stop
+  - testdata: more corpus entries (#5182)
+  - add soypat/lneto to corpus
+  - GNUmakefile: add context and expvar to stdlib tests
+  - GNUmakefile: add encoding/xml to stdlib tests on Linux
+  - main (test): skip mipsle test if the emulator is missing
+  - make: remove machine without board from smoketest
+  - stm32: add Arduino UNO Q to smoketests
+  - flake: bump to nixpkgs 25.11
+  - chore: update all github actions to latest versions
+  - fix: use external 7zip for scoop installs on Windows
+
+0.40.1
+---
+* **machine**
+  - nrf: fix flash writes when SoftDevice is enabled
+* **runtime**
+  - runtime: avoid fixed math/rand sequence on RP2040/RP2350 (#5124)
+  - runtime: add calls to initRand() during run() for all schedulers
+  - runtime: call initRand() before initHeap() during initialization
+  - runtime: use rand_hwrng hardwareRand for RP2040/RP2350 (#5135)
+* **libs**
+  - picolibc: use updated location for git repo
+
+0.40.0
+---
+* **general**
+  - all: add full LLVM 20 support
+  - core: feat: enable //go:linkname pragma for globals
+  - core: feature: Add flag to ignore go compatibility matrix (#5078)
+  - chore: update version for 0.40 development cycle
+* **compiler**
+  - emit an error when the actual arch doesn't match GOARCH
+  - mark string parameters as readonly
+  - use Tarjan's SCC algorithm to detect loops for defer
+  - lower "large stack" limit to 16kb
+* **core**
+  - shrink bdwgc library
+  - Fix linker errors for runtime.vgetrandom and crypto/internal/sysrand.fatal
+  - fix: add TryLock to sync.RWMutex
+  - fix: correct linter issues exposed by the fix in #4679
+  - fix: don't hardcode success return state
+  - fix: expand RTT debugger compatibility
+  - internal/task (threads): save stack bounds instead of scanning under a lock
+  - internal/task: create detached threads and fix error handling
+  - interp: better errors when debugging interp
+  - transform (gc): create stack slots in callers of external functions
+  - internal/task: prevent semaphore resource leak for threads scheduler
+* **machine**
+  - cortexm: optimize code size for the HardFault_Handler
+  - fe310: add I2C pins for the HiFive1b
+  - clarify WriteAt semantics of BlockDevice
+  - fix deprecated AsmFull comment (#5005)
+  - make sure DMA buffers do not escape unnecessarily
+  - only enable USB-CDC when needed
+  - use larger SPI MAXCNT on nrf52833 and nrf52840
+  - fix: update m.queuedBytes when clamping output to avoid corrupting sentBytes
+  - fix: use int64 in ReadTemperature to avoid overflow
+  - fix(rp2): disable DBGPAUSE on startup
+  - fix(rp2): possible integer overflow while computing factors for SPI baudrate
+  - fix(rp2): reset spinlocks at startup
+  - fix(rp2): switch spinlock busy loop to wfe
+  - fix(rp2): use side-effect-free spinlocks
+  - nrf: add ADC_VDDH which is an ADC pin for VDDH
+  - nrf: don't block SPI transfer
+  - nrf: don't set PSELN, it's ignored in single ended mode anyway
+  - nrf: fix typo in ADC configuration
+  - nrf: refactor SoftDevice enabled check
+  - nrf: rename pwmPin to adcPin
+  - nrf: support flash operations while the SoftDevice is enabled
+  - rp2040: allow writing to the UART inside interrupts
+  - machine,nrf528: stop the bus only once on I2C bus error and ensures the first error is returned
+* **net**
+  - update submodule to latest commits
+* **runtime**
+  - (avr): fix infinite longjmp loop if stack is aligned to 256 bytes
+  - (gc_blocks.go): clear full size of allocation
+  - (gc_blocks.go): make sweep branchless
+  - (gc_blocks.go): simplify scanning logic
+  - (gc_blocks.go): use a linked stack to scan marked objects
+  - (gc_blocks.go): use best-fit allocation
+  - (gc_boehm.go): fix world already stopped check
+  - (wasm): scan the system stack
+  - fix sleep duration for long sleeps
+  - remove copied code for nrf52840
+  - src/syscall: update src buffer after write
+  - wasm: fix C realloc and optimize it a bit
+* **targets**
+  - add xiao-esp32s3 board target
+  - Add ESP32-S3 support (#5091)
+  - Added Gopher ARCADE board
+  - Create "pico2-ice" target board (#5062)
+* **build/test**
+  - Add testing.T.Context() and testing.B.Context()
+  - create separate go.mod file for testing dependencies for wasm tests that use Chromium headless browser to avoid use of older incompatible version.
+  - go back to normal scheduler instead of tasks scheduler for macos CI
+  - update CI to use Go 1.25.5
+  - update macOS GH actions builds to handle sunset of macOS 13
+  - use task scheduler on macOS builds to avoid test race condition lockups
+  - update all CI builds to use latest stable Go release. Also update some of the actions to their latest releases.
+  - update GH actions builds to use Go 1.25.4
+  - uninstall cmake before install
+  - fix: point the submodule for musl-lib to a mirror in the TinyGo GitHub org
+  - fix: remove macOS 15 from CI build matrix (conflicts with macOS 14 build)
+  - fix: separate host expected bytes from device intended bytes
+  - fix/typo: makeESPFirmwareImage
+  - make: GNUmakefile: shrink TinyGo binaries on Linux
+  - move the directory list into a variable
+  - several improvements to the macOS GH actions build
+  - Fix for #4678: top-level 'make lint' wasn't working
+  - fix: increase the timeout for chromedp to connect to the headless browser used for running the wasm tests.
+  - testdata: some more packages for the test corpus
+
+0.39.0
+---
+* **general**
+  - all: add Go 1.25 support
+  - net: update to latest tinygo net package
+  - docs: clarify build verification step for macOS users
+  - Add flag to skip Renesas SVD builds
+* **build**
+  - Makefile: install missing dlmalloc files
+  - flash: add -o flag support to save built binary (Fixes #4937) (#4942)
+  - fix: update version of clang to 17 to accommodate latest Go 1.25 docker base image
+* **ci**
+  - chore: update all CI builds to test Go 1.25 release
+  - fix: disable test-newest since CircleCI seems unable to download due to rate-limits on Dockerhub
+  - ci: rename some jobs to avoid churn on every Go/LLVM version bump
+  - ci: make the goroutines test less racy
+  - tests: de-flake goroutines test
+* **compiler**
+  - compiler: implement internal/abi.Escape
+* **main**
+  - main: show the compiler error (if any) for `tinygo test -c`
+  - chore: correct GOOS=js name in error messages for WASM
+* **machine**
+  - machine: add international keys
+  - machine: remove some unnecessary "// peripherals:" comments
+  - machine: add I2C pin comments
+  - machine: standardize I2C errors with "i2c:" prefix
+  - machine: make I2C usable in the simulator
+  - fix: add SPI and I2C to teensy 4.1 (#4943)
+  - `rp2`: use the correct channel mask for rp2350 ADC; hold lock during read (#4938)
+  - `rp2`: disable digital input for analog inputs
+* **runtime**
+  - runtime: ensure time.Sleep(d) sleeps at least d
+  - runtime: stub out weak pointer support
+  - runtime: implement dummy AddCleanup
+  - runtime: enable multi-core scheduler for rp2350
+  - `internal/task`: use -stack-size flag when starting a new thread
+  - `internal/task`: add SA_RESTART flag to GC interrupts
+  - `internal/task`: a few small correctness fixes
+  - `internal/gclayout`: make gclayout values constants
+  - darwin: add threading support and use it by default
+* **standard library**
+  - `sync`: implement sync.Swap
+  - `reflect`: implement Method.IsExported
+* **testing**
+  - testing: stub out testing.B.Loop
+* **targets**
+  - `stm32`: add support for the STM32L031G6U6
+  - add metro-rp2350 board definition (#4989)
+  - `rp2040/rp2350`: set the default stack size to 8k for rp2040/rp2350 based boards where this was not already the case
+
+
+0.38.0
+---
+* **general**
+  - `go.*`: upgrade `golang.org/x/tools` to v0.30.0
+  - `all`: add support for LLVM 20
+* **build**
+  - go back to using MinoruSekine/setup-scoop for Windows CI builds
+  - `flake.*`: upgrade to nixpkgs 25.05, LLVM 20
+  - `Makefile`: only detect ccache command when needed
+  - `Makefile`: create random filename inside rule
+  - `Makefile`: don't set GOROOT
+  - `Makefile`: call uname at most once
+  - `Makefile`: only read NodeJS version when it is needed
+* **compiler**
+  - add support for `GODEBUG=gotypesalias=1`
+  - `interp`: fix `copy()` from/to external buffers
+  - add `-nobounds` (similar to `-gcflags=-B`)
+  - `compileopts`: add library version to cached library path
+  - `builder`: build wasi-libc inside TinyGo
+  - `builder`: simplify bdwgc libc dependency
+  - `builder`: don't use precompiled libraries
+  - `compileopts`: enable support for `GOARCH=wasm` in `tinygo test`
+* **fixes**
+  - `rp2350`: Fix DMA to SPI transmits on RP2350 (#4903)
+  - `microbit v2`: use OpenOCD flash method on microbit v2 when using Nordic Semi SoftDevice
+  - `main`: display all of the current GC options for the `-gc` flag
+  - Remove duplicated error handling
+  - `sync`: fix `TestMutexConcurrent` test
+  - fix race condition in `testdata/goroutines.go`
+  - fix build warnings on Windows ARM
+* **machine**
+  - `usb`: add USB mass storage class support
+  - implement usb receive message throttling
+  - declare usb endpoints per-platform
+  - `samd21`: implement watchdog
+  - `samd51`: write to flash memory in 512 byte long chunks
+  - `samd21`: write to flash memory in 64 byte long chunks
+  - don't inline RTT `WriteByte` everywhere
+  - `rp2`: unexport machine-specific errors
+  - `rp2`: discount scheduling delays in I2C timeouts (#4876)
+  - use pointer receiver in simulated PWM peripherals
+  - add simulated PWM/timer peripherals
+  - `rp2`: expose usb endpoint stall handling
+  - `arm`: clear pending interrupts before enabling them
+  - `rp2`: merge common usb code (#4856)
+* **main**
+  - add "cores" and "threads" schedulers to help text
+  - add `StartPos` and `EndPos` to `-json` build output
+  - change `-json` flag to match upstream Go
+* **runtime**
+  - don't lock the print output inside interrupts
+  - don't try to interrupt other cores before they are started
+  - implement `NumCPU` for the multicore scheduler
+  - add support for multicore scheduler
+  - refactor obtaining the system stack
+  - `interrupt`: add `Checkpoint` type
+  - add `exportedFuncPtr`
+  - avoid an allocation in `(*time.Timer).Reset`
+  - stub runtime signal functions for `os/signal` on wasip1
+  - move `timeUnit` to a single place
+  - implement `NumCPU` for `-scheduler=threads`
+  - move `mainExited` boolean
+  - `internal/task`: rename `tinygo_pause` to `tinygo_task_exit`
+  - map every goroutine to a new OS thread
+  - refactor `timerQueue`
+  - make conservative and precise GC MT-safe
+  - `internal/task`: implement atomic primitives for preemptive scheduling
+  - Use diskutil on macOS to extract volume name and path for FAT mounts #4928
+* **standard library**
+  - `net`: update submodule to latest commits
+  - `runtime/debug`: add GC related stubs
+  - `metrics`: flesh out some of the metric types
+  - `reflect`: Chan related stubs
+  - `os`: handle relative and abs paths in `Executable()`
+  - `os`: add `os.Executable()` for Darwin
+  - `sync`: implement `RWMutex` using futexes
+  - `reflect`: Add `SliceOf`, `ArrayOf`, `StructOf`, `MapOf`, `FuncOf`
+* **targets**
+  - `rp2040`: add multicore support
+  - `riscv32`: use `gdb` binary as a fallback
+  - add target for Microbit v2 with SoftDevice S140 support for both peripheral and central
+  - `windows`: use MSVCRT.DLL instead of UCRT on i386
+  - `windows`: add windows/386 support
+  - `arm64`: remove unnecessary `.section` directive
+  - `riscv-qemu`: actually sleep in `time.Sleep()`
+  - `riscv`: define CSR constants and use them where possible
+  - `darwin`: support Boehm GC (and use by default)
+  - `windows`: add support for the Boehm-Demers-Weiser GC
+  - `windows`: fix wrong register for first parameter
+* **wasm**
+  - add Boehm GC support
+  - refactor/modify stub signal handling
+  - don't block `//go:wasmexport` because of running goroutines
+  - use `int64` instead of `float64` for the `timeUnit`
+* **boards**
+  - Add board support for BigTreeTech SKR Pico (#4842)
+
+
 0.37.0
 ---
 * **general**
@@ -1,8 +1,8 @@
 # tinygo-llvm stage obtains the llvm source for TinyGo
-FROM golang:1.24 AS tinygo-llvm
+FROM golang:1.27 AS tinygo-llvm

 RUN apt-get update && \
-    apt-get install -y apt-utils make cmake clang-15 ninja-build && \
+    apt-get install -y apt-utils make cmake clang-17 ninja-build && \
    rm -rf \
        /var/lib/apt/lists/* \
        /var/log/* \
@@ -10,6 +10,7 @@ RUN apt-get update && \
        /tmp/*

 COPY ./GNUmakefile /tinygo/GNUmakefile
+COPY ./make /tinygo/make

 RUN cd /tinygo/ && \
    make llvm-source
@@ -33,7 +34,7 @@ RUN cd /tinygo/ && \

 # tinygo-compiler copies the compiler build over to a base Go container (without
 # all the build tools etc).
-FROM golang:1.24 AS tinygo-compiler
+FROM golang:1.27 AS tinygo-compiler

 # Copy tinygo build.
 COPY --from=tinygo-compiler-build /tinygo/build/release/tinygo /tinygo
@@ -1,7 +1,8 @@
-Copyright (c) 2018-2025 The TinyGo Authors. All rights reserved.
+Copyright (c) 2018-2026 The TinyGo Authors. All rights reserved.

 TinyGo includes portions of the Go standard library.
-Copyright (c) 2009-2024 The Go Authors. All rights reserved.
+Copyright 2009 The Go Authors. All rights reserved.
+See https://github.com/golang/go/blob/master/LICENSE for license information.

 TinyGo includes portions of LLVM, which is under the Apache License v2.0 with
 LLVM Exceptions. See https://llvm.org/LICENSE.txt for license information.
@@ -1,11 +1,14 @@
 # TinyGo - Go compiler for small places

-[![Linux](https://github.com/tinygo-org/tinygo/actions/workflows/linux.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/linux.yml) [![macOS](https://github.com/tinygo-org/tinygo/actions/workflows/build-macos.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/build-macos.yml) [![Windows](https://github.com/tinygo-org/tinygo/actions/workflows/windows.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/windows.yml) [![Docker](https://github.com/tinygo-org/tinygo/actions/workflows/docker.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/docker.yml) [![Nix](https://github.com/tinygo-org/tinygo/actions/workflows/nix.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/nix.yml) [![CircleCI](https://circleci.com/gh/tinygo-org/tinygo/tree/dev.svg?style=svg)](https://circleci.com/gh/tinygo-org/tinygo/tree/dev)
+[![Linux](https://github.com/tinygo-org/tinygo/actions/workflows/linux.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/linux.yml) [![macOS](https://github.com/tinygo-org/tinygo/actions/workflows/build-macos.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/build-macos.yml) [![Windows](https://github.com/tinygo-org/tinygo/actions/workflows/windows.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/windows.yml) [![Docker](https://github.com/tinygo-org/tinygo/actions/workflows/docker.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/docker.yml) [![Nix](https://github.com/tinygo-org/tinygo/actions/workflows/nix.yml/badge.svg?branch=dev)](https://github.com/tinygo-org/tinygo/actions/workflows/nix.yml)

 TinyGo is a Go compiler intended for use in small places such as microcontrollers, WebAssembly (wasm/wasi), and command-line tools.

 It reuses libraries used by the [Go language tools](https://golang.org/pkg/go/) alongside [LLVM](http://llvm.org) to provide an alternative way to compile programs written in the Go programming language.

+> [!IMPORTANT]
+> You can help TinyGo with a financial contribution using OpenCollective. Please see https://opencollective.com/tinygo for more information. Thank you!
+
 ## Embedded

 Here is an example program that blinks the built-in LED when run directly on any supported board with onboard LED:
@@ -31,10 +34,10 @@ func main() {
 }
 ```

-The above program can be compiled and run without modification on an Arduino Uno, an Adafruit ItsyBitsy M0, or any of the supported boards that have a built-in LED, just by setting the correct TinyGo compiler target. For example, this compiles and flashes an Arduino Uno:
+The above program can be compiled and run without modification on an [Arduino Uno](https://tinygo.org/docs/reference/microcontrollers/boards/arduino-uno), an [Adafruit Circuit Playground Express](https://tinygo.org/docs/reference/microcontrollers/featured/circuitplay-express), a [Seeed Studio XIAO-ESP32S3](https://tinygo.org/docs/reference/microcontrollers/featured/xiao-esp32s3) or any of the many supported boards that have a built-in LED, just by setting the correct TinyGo compiler target. For example, this compiles and flashes an Arduino Uno:

 ```shell
-tinygo flash -target arduino examples/blinky1
+tinygo flash -target arduino-uno examples/blinky1
 ```

 ## WebAssembly
@@ -63,7 +66,7 @@ tinygo build -buildmode=c-shared -o add.wasm -target=wasip1 add.go
 You can also use the same syntax as Go 1.24+:

 ```shell
-GOARCH=wasip1 GOOS=wasm tinygo build -buildmode=c-shared -o add.wasm add.go
+GOOS=wasip1 GOARCH=wasm tinygo build -buildmode=c-shared -o add.wasm add.go
 ```

 ## Installation
@@ -74,7 +77,7 @@ See the [getting started instructions](https://tinygo.org/getting-started/) for

 ### Embedded

-You can compile TinyGo programs for over 94 different microcontroller boards.
+You can compile TinyGo programs for over 150 different microcontroller boards.

 For more information, please see https://tinygo.org/docs/reference/microcontrollers/

@@ -139,7 +142,7 @@ Non-goals:

 ## Why this project exists

-> We never expected Go to be an embedded language and so its got serious problems...
+> We never expected Go to be an embedded language, and so it’s got serious problems...

 -- Rob Pike, [GopherCon 2014 Opening Keynote](https://www.youtube.com/watch?v=VoS7DsT1rdM&feature=youtu.be&t=2799)

@@ -3,6 +3,7 @@ package builder
 import (
 	"bytes"
 	"debug/elf"
+	"debug/macho"
 	"debug/pe"
 	"encoding/binary"
 	"errors"
@@ -62,6 +63,20 @@ func makeArchive(arfile *os.File, objs []string) error {
 					fileIndex int
 				}{symbol.Name, i})
 			}
+		} else if dbg, err := macho.NewFile(objfile); err == nil {
+			for _, symbol := range dbg.Symtab.Syms {
+				// See mach-o/nlist.h
+				if symbol.Type&0x0e != 0xe { // (symbol.Type & N_TYPE) != N_SECT
+					continue // undefined symbol
+				}
+				if symbol.Type&0x01 == 0 { // (symbol.Type & N_EXT) == 0
+					continue // internal symbol (static, etc)
+				}
+				symbolTable = append(symbolTable, struct {
+					name      string
+					fileIndex int
+				}{symbol.Name, i})
+			}
 		} else if dbg, err := pe.NewFile(objfile); err == nil {
 			for _, symbol := range dbg.Symbols {
 				if symbol.StorageClass != 2 {
@@ -5,6 +5,7 @@ package builder

 import (
 	"path/filepath"
+	"strings"

 	"github.com/tinygo-org/tinygo/goenv"
 )
@@ -13,7 +14,7 @@ var BoehmGC = Library{
 	name: "bdwgc",
 	cflags: func(target, headerPath string) []string {
 		libdir := filepath.Join(goenv.Get("TINYGOROOT"), "lib/bdwgc")
-		return []string{
+		flags := []string{
 			// use a modern environment
 			"-DUSE_MMAP",              // mmap is available
 			"-DUSE_MUNMAP",            // return memory to the OS using munmap
@@ -24,6 +25,15 @@ var BoehmGC = Library{
 			"-DALL_INTERIOR_POINTERS",  // scan interior pointers (needed for Go)
 			"-DIGNORE_DYNAMIC_LOADING", // we don't support dynamic loading at the moment
 			"-DNO_GETCONTEXT",          // musl doesn't support getcontext()
+			"-DGC_DISABLE_INCREMENTAL", // don't mess with SIGSEGV and such
+
+			// Use a minimal environment.
+			"-DNO_MSGBOX_ON_ERROR", // don't call MessageBoxA on Windows
+			"-DDONT_USE_ATEXIT",
+			"-DNO_GETENV",          // smaller binary, more predictable configuration
+			"-DNO_CLOCK",           // don't use system clock
+			"-DNO_DEBUGGING",       // reduce code size
+			"-DGC_NO_FINALIZATION", // finalization is not used at the moment

 			// Special flag to work around the lack of __data_start in ld.lld.
 			// TODO: try to fix this in LLVM/lld directly so we don't have to
@@ -32,28 +42,32 @@ var BoehmGC = Library{

 			// Do not scan the stack. We have our own mechanism to do this.
 			"-DSTACK_NOT_SCANNED",
+			"-DNO_PROC_STAT",  // we scan the stack manually (don't read /proc/self/stat on Linux)
+			"-DSTACKBOTTOM=0", // dummy value, we scan the stack manually

 			// Assertions can be enabled while debugging GC issues.
 			//"-DGC_ASSERTIONS",

-			// Threading is not yet supported, so these are disabled.
+			// We use our own way of dealing with threads (that is a bit hacky).
+			// See src/runtime/gc_boehm.go.
 			//"-DGC_THREADS",
 			//"-DTHREAD_LOCAL_ALLOC",

 			"-I" + libdir + "/include",
 		}
+		return flags
 	},
+	needsLibc: true,
 	sourceDir: func() string {
 		return filepath.Join(goenv.Get("TINYGOROOT"), "lib/bdwgc")
 	},
-	librarySources: func(target string) ([]string, error) {
-		return []string{
+	librarySources: func(target string, _ bool) ([]string, error) {
+		sources := []string{
 			"allchblk.c",
 			"alloc.c",
 			"blacklst.c",
 			"dbg_mlc.c",
 			"dyn_load.c",
-			"finalize.c",
 			"headers.c",
 			"mach_dep.c",
 			"malloc.c",
@@ -61,11 +75,17 @@ var BoehmGC = Library{
 			"mark_rts.c",
 			"misc.c",
 			"new_hblk.c",
-			"obj_map.c",
 			"os_dep.c",
-			"pthread_stop_world.c",
-			"pthread_support.c",
 			"reclaim.c",
-		}, nil
+		}
+		if strings.Split(target, "-")[2] == "windows" {
+			// Due to how the linker on Windows works (that doesn't allow
+			// undefined functions), we need to include these extra files.
+			sources = append(sources,
+				"mallocx.c",
+				"ptr_chck.c",
+			)
+		}
+		return sources, nil
 	},
 }
@@ -14,15 +14,17 @@ import (
 	"fmt"
 	"go/types"
 	"hash/crc32"
-	"io/fs"
+	"maps"
 	"math/bits"
 	"os"
 	"os/exec"
 	"path/filepath"
 	"runtime"
+	"slices"
 	"sort"
 	"strconv"
 	"strings"
+	"sync"

 	"github.com/gofrs/flock"
 	"github.com/tinygo-org/tinygo/compileopts"
@@ -137,9 +139,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		if _, ok := globalValues[pkgPath]; !ok {
 			globalValues[pkgPath] = map[string]string{}
 		}
-		for k, v := range vals {
-			globalValues[pkgPath][k] = v
-		}
+		maps.Copy(globalValues[pkgPath], vals)
 	}

 	// Check for a libc dependency.
@@ -147,14 +147,13 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 	// the libc needs them.
 	root := goenv.Get("TINYGOROOT")
 	var libcDependencies []*compileJob
-	var libcJob *compileJob
 	switch config.Target.Libc {
 	case "darwin-libSystem":
-		job := makeDarwinLibSystemJob(config, tmpdir)
-		libcDependencies = append(libcDependencies, job)
+		libcJob := makeDarwinLibSystemJob(config, tmpdir)
+		libcDependencies = append(libcDependencies, libcJob)
 	case "musl":
 		var unlock func()
-		libcJob, unlock, err = libMusl.load(config, tmpdir, nil)
+		libcJob, unlock, err := libMusl.load(config, tmpdir)
 		if err != nil {
 			return BuildResult{}, err
 		}
@@ -162,32 +161,33 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		libcDependencies = append(libcDependencies, dummyCompileJob(filepath.Join(filepath.Dir(libcJob.result), "crt1.o")))
 		libcDependencies = append(libcDependencies, libcJob)
 	case "picolibc":
-		libcJob, unlock, err := libPicolibc.load(config, tmpdir, nil)
+		libcJob, unlock, err := libPicolibc.load(config, tmpdir)
 		if err != nil {
 			return BuildResult{}, err
 		}
 		defer unlock()
 		libcDependencies = append(libcDependencies, libcJob)
 	case "wasi-libc":
-		path := filepath.Join(root, "lib/wasi-libc/sysroot/lib/wasm32-wasi/libc.a")
-		if _, err := os.Stat(path); errors.Is(err, fs.ErrNotExist) {
-			return BuildResult{}, errors.New("could not find wasi-libc, perhaps you need to run `make wasi-libc`?")
+		libcJob, unlock, err := libWasiLibc.load(config, tmpdir)
+		if err != nil {
+			return BuildResult{}, err
 		}
-		libcDependencies = append(libcDependencies, dummyCompileJob(path))
+		defer unlock()
+		libcDependencies = append(libcDependencies, libcJob)
 	case "wasmbuiltins":
-		libcJob, unlock, err := libWasmBuiltins.load(config, tmpdir, nil)
+		libcJob, unlock, err := libWasmBuiltins.load(config, tmpdir)
 		if err != nil {
 			return BuildResult{}, err
 		}
 		defer unlock()
 		libcDependencies = append(libcDependencies, libcJob)
 	case "mingw-w64":
-		job, unlock, err := libMinGW.load(config, tmpdir, nil)
+		libcJob, unlock, err := libMinGW.load(config, tmpdir)
 		if err != nil {
 			return BuildResult{}, err
 		}
 		defer unlock()
-		libcDependencies = append(libcDependencies, job)
+		libcDependencies = append(libcDependencies, libcJob)
 		libcDependencies = append(libcDependencies, makeMinGWExtraLibs(tmpdir, config.GOARCH())...)
 	case "":
 		// no library specified, so nothing to do
@@ -215,6 +215,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		MaxStackAlloc:      config.MaxStackAlloc(),
 		NeedsStackObjects:  config.NeedsStackObjects(),
 		Debug:              !config.Options.SkipDWARF, // emit DWARF except when -internal-nodwarf is passed
+		Nobounds:           config.Options.Nobounds,
 		PanicStrategy:      config.PanicStrategy(),
 	}

@@ -254,9 +255,22 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		result.PackagePathMap[pkg.OriginalDir()] = pkg.Pkg.Path()
 	}

+	// Strip default initializers for -X globals from the type info before
+	// building SSA. This prevents go/ssa from emitting init stores for them,
+	// so that makeGlobalsModule can supply the correct values at final link
+	// time without any runtime init overwriting them. The -X values themselves
+	// are kept out of the per-package build cache; only the variable names
+	// appear in the cache key.
+	for _, pkg := range lprogram.Sorted() {
+		for name := range globalValues[pkg.Pkg.Path()] {
+			pkg.StripVarInitializer(name)
+		}
+	}
+
 	// Create the *ssa.Program. This does not yet build the entire SSA of the
 	// program so it's pretty fast and doesn't need to be parallelized.
 	program := lprogram.LoadSSA()
+	buildProgram := sync.OnceFunc(program.Build)

 	// Add jobs to compile each package.
 	// Packages that have a cache hit will not be compiled again.
@@ -265,7 +279,6 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe

 	var embedFileObjects []*compileJob
 	for _, pkg := range lprogram.Sorted() {
-		pkg := pkg // necessary to avoid a race condition

 		var undefinedGlobals []string
 		for name := range globalValues[pkg.Pkg.Path()] {
@@ -281,9 +294,13 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 				allFiles[file.Name] = append(allFiles[file.Name], file)
 			}
 		}
-		for name, files := range allFiles {
-			name := name
-			files := files
+		// Sort embedded files by name to maintain output determinism.
+		embedNames := make([]string, 0, len(allFiles))
+		for _, files := range allFiles {
+			embedNames = append(embedNames, files[0].Name)
+		}
+		slices.Sort(embedNames)
+		for _, name := range embedNames {
 			job := &compileJob{
 				description: "make object file for " + name,
 				run: func(job *compileJob) error {
@@ -298,7 +315,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 					sum := sha256.Sum256(data)
 					hexSum := hex.EncodeToString(sum[:16])

-					for _, file := range files {
+					for _, file := range allFiles[name] {
 						file.Size = uint64(len(data))
 						file.Hash = hexSum
 						if file.NeedsData {
@@ -383,8 +400,13 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 					return nil
 				}

-				// Compile AST to IR. The compiler.CompilePackage function will
-				// build the SSA as needed.
+				// SSA package builds may run concurrently, but the resulting
+				// functions cannot be inspected until all builds have finished:
+				// generic instances and wrappers can be shared across packages.
+				// Build the whole program once before compiling any package.
+				buildProgram()
+
+				// Compile AST to IR.
 				mod, errs := compiler.CompilePackage(pkg.ImportPath, pkg, program.Package(pkg.Pkg), machine, compilerConfig, config.DumpSSA())
 				defer mod.Context().Dispose()
 				defer mod.Dispose()
@@ -472,7 +494,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 				if pkgInit.IsNil() {
 					panic("init not found for " + pkg.Pkg.Path())
 				}
-				err := interp.RunFunc(pkgInit, config.Options.InterpTimeout, config.DumpSSA())
+				err := interp.RunFunc(pkgInit, config.Options.InterpTimeout, config.Options.InterpMaxLoopIterations, config.DumpSSA())
 				if err != nil {
 					return err
 				}
@@ -541,6 +563,23 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 				if err != nil {
 					return fmt.Errorf("failed to load bitcode file: %w", err)
 				}
+				// Resolve duplicate function definitions before linking.
+				// This can happen when a newer Go version adds a function
+				// body in a standard library package that was previously
+				// just a declaration provided by //go:linkname from the
+				// runtime. In that case, keep the existing (runtime)
+				// definition by weakening the new one's linkage so the
+				// LLVM linker discards it in favor of the existing one.
+				for fn := pkgMod.FirstFunction(); !fn.IsNil(); fn = llvm.NextFunction(fn) {
+					if fn.IsDeclaration() {
+						continue
+					}
+					existing := mod.NamedFunction(fn.Name())
+					if existing.IsNil() || existing.IsDeclaration() {
+						continue
+					}
+					fn.SetLinkage(llvm.LinkOnceODRLinkage)
+				}
 				err = llvm.LinkModules(mod, pkgMod)
 				if err != nil {
 					return fmt.Errorf("failed to link module: %w", err)
@@ -703,7 +742,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 	// Add compiler-rt dependency if needed. Usually this is a simple load from
 	// a cache.
 	if config.Target.RTLib == "compiler-rt" {
-		job, unlock, err := libCompilerRT.load(config, tmpdir, nil)
+		job, unlock, err := libCompilerRT.load(config, tmpdir)
 		if err != nil {
 			return result, err
 		}
@@ -713,10 +752,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe

 	// The Boehm collector is stored in a separate C library.
 	if config.GC() == "boehm" {
-		if libcJob == nil {
-			return BuildResult{}, fmt.Errorf("boehm GC isn't supported with libc %s", config.Target.Libc)
-		}
-		job, unlock, err := BoehmGC.load(config, tmpdir, libcJob)
+		job, unlock, err := BoehmGC.load(config, tmpdir)
 		if err != nil {
 			return BuildResult{}, err
 		}
@@ -744,7 +780,6 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 	// TODO: do this as part of building the package to be able to link the
 	// bitcode files together.
 	for _, pkg := range lprogram.Sorted() {
-		pkg := pkg
 		for _, filename := range pkg.CFiles {
 			abspath := filepath.Join(pkg.OriginalDir(), filename)
 			job := &compileJob{
@@ -811,22 +846,25 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 			}
 			ldflags = append(ldflags, "-mllvm", "-mcpu="+config.CPU())
 			ldflags = append(ldflags, "-mllvm", "-mattr="+config.Features()) // needed for MIPS softfloat
-			if config.GOOS() == "windows" {
-				// Options for the MinGW wrapper for the lld COFF linker.
+			switch config.LinkerFlavor() {
+			case "coff":
+				// Options for driving ld.lld in PE/COFF mode.
 				ldflags = append(ldflags,
 					"-Xlink=/opt:lldlto="+strconv.Itoa(speedLevel),
 					"--thinlto-cache-dir="+filepath.Join(cacheDir, "thinlto"))
-			} else if config.GOOS() == "darwin" {
+			case "darwin":
 				// Options for the ld64-compatible lld linker.
 				ldflags = append(ldflags,
 					"--lto-O"+strconv.Itoa(speedLevel),
 					"-cache_path_lto", filepath.Join(cacheDir, "thinlto"))
-			} else {
+			case "gnu":
 				// Options for the ELF linker.
 				ldflags = append(ldflags,
 					"--lto-O"+strconv.Itoa(speedLevel),
 					"--thinlto-cache-dir="+filepath.Join(cacheDir, "thinlto"),
 				)
+			default:
+				return fmt.Errorf("unknown linker flavor: %s", config.LinkerFlavor())
 			}
 			if config.CodeModel() != "default" {
 				ldflags = append(ldflags,
@@ -883,7 +921,7 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 			}

 			// Run wasm-opt for wasm binaries
-			if arch := strings.Split(config.Triple(), "-")[0]; arch == "wasm32" {
+			if arch, _, _ := strings.Cut(config.Triple(), "-"); arch == "wasm32" {
 				optLevel, _, _ := config.OptLevel()
 				opt := "-" + optLevel

@@ -1045,11 +1083,11 @@ func Build(pkgName, outpath, tmpdir string, config *compileopts.Config) (BuildRe
 		if err != nil {
 			return result, err
 		}
-	case "esp32", "esp32-img", "esp32c3", "esp8266":
+	case "esp32", "esp32-img", "esp32c3", "esp32s3", "esp32c6", "esp8266":
 		// Special format for the ESP family of chips (parsed by the ROM
 		// bootloader).
 		result.Binary = filepath.Join(tmpdir, "main"+outext)
-		err := makeESPFirmareImage(result.Executable, result.Binary, outputBinaryFormat)
+		err := makeESPFirmwareImage(result.Executable, result.Binary, outputBinaryFormat)
 		if err != nil {
 			return result, err
 		}
@@ -1177,7 +1215,7 @@ func createEmbedObjectFile(data, hexSum, sourceFile, sourceDir, tmpdir string, c
 // needed to convert a program to its final form. Some transformations are not
 // optional and must be run as the compiler expects them to run.
 func optimizeProgram(mod llvm.Module, config *compileopts.Config) error {
-	err := interp.Run(mod, config.Options.InterpTimeout, config.DumpSSA())
+	err := interp.Run(mod, config.Options.InterpTimeout, config.Options.InterpMaxLoopIterations, config.DumpSSA())
 	if err != nil {
 		return err
 	}
@@ -1321,6 +1359,14 @@ func determineStackSizes(mod llvm.Module, executable string) ([]string, map[stri
 	}
 	baseStackSize, baseStackSizeType, baseStackSizeFailedAt := functions["tinygo_startTask"][0].StackSize()

+	// Account for the bytes that tinygo_swapTask pushes onto the goroutine stack
+	// on every context switch. The static analysis correctly traces Go calls,
+	// but it cannot see into the assembly-level register push.
+	var contextSwitchOverhead uint64
+	if swapFuncs, ok := functions["tinygo_swapTask"]; ok && len(swapFuncs) == 1 {
+		contextSwitchOverhead = swapFuncs[0].FrameSize
+	}
+
 	sizes := make(map[string]functionStackSize)

 	// Add the reset handler function, for convenience. The reset handler runs
@@ -1369,6 +1415,12 @@ func determineStackSizes(mod llvm.Module, executable string) ([]string, map[stri
 			// overflow will occur even before the goroutine is started.
 			stackSize = baseStackSize
 		}
+		if stackSizeType == stacksize.Bounded {
+			// Add the overhead of context switching. This is needed because the
+			// context switch (tinygo_swapTask) pushes callee-saved registers
+			// onto the current stack, which is not seen by the static analysis.
+			stackSize += contextSwitchOverhead
+		}
 		sizes[name] = functionStackSize{
 			stackSize:        stackSize,
 			stackSizeType:    stackSizeType,
@@ -5,6 +5,7 @@ import (
 	"os"
 	"path/filepath"
 	"runtime"
+	"strings"
 	"testing"

 	"github.com/tinygo-org/tinygo/compileopts"
@@ -28,12 +29,15 @@ func TestClangAttributes(t *testing.T) {
 		"cortex-m4",
 		"cortex-m7",
 		"esp32c3",
+		"esp32c6",
+		"esp32s3",
 		"fe310",
 		"gameboy-advance",
 		"k210",
 		"nintendoswitch",
 		"riscv-qemu",
 		"tkey",
+		"uefi-amd64",
 		"wasip1",
 		"wasip2",
 		"wasm",
@@ -45,7 +49,6 @@ func TestClangAttributes(t *testing.T) {
 		targetNames = append(targetNames, "esp32", "esp8266")
 	}
 	for _, targetName := range targetNames {
-		targetName := targetName
 		t.Run(targetName, func(t *testing.T) {
 			testClangAttributes(t, &compileopts.Options{Target: targetName})
 		})
@@ -67,6 +70,7 @@ func TestClangAttributes(t *testing.T) {
 		{GOOS: "linux", GOARCH: "mipsle", GOMIPS: "softfloat"},
 		{GOOS: "darwin", GOARCH: "amd64"},
 		{GOOS: "darwin", GOARCH: "arm64"},
+		{GOOS: "windows", GOARCH: "386"},
 		{GOOS: "windows", GOARCH: "amd64"},
 		{GOOS: "windows", GOARCH: "arm64"},
 	} {
@@ -126,8 +130,10 @@ func testClangAttributes(t *testing.T, options *compileopts.Options) {
 	defer mod.Dispose()

 	// Check whether the LLVM target matches.
-	if mod.Target() != config.Triple() {
-		t.Errorf("target has LLVM triple %#v but Clang makes it LLVM triple %#v", config.Triple(), mod.Target())
+	// Use ClangTriple since LLVM 22 normalizes wasm32-unknown-wasi to wasip1.
+	expectedTriple := compileopts.ClangTriple(config.Triple())
+	if mod.Target() != expectedTriple {
+		t.Errorf("target has LLVM triple %#v but Clang makes it LLVM triple %#v", expectedTriple, mod.Target())
 	}

 	// Check the "target-cpu" and "target-features" string attribute of the add
@@ -151,11 +157,65 @@ func testClangAttributes(t *testing.T, options *compileopts.Options) {
 			// The reason is that Debian has patched Clang in a way that
 			// modifies the LLVM features string, changing lots of FPU/float
 			// related flags. We want to test vanilla Clang, not Debian Clang.
-			t.Errorf("target has LLVM features\n\t%#v\nbut Clang makes it\n\t%#v", config.Features(), features)
+			//
+			// Rather than requiring an exact match (which breaks across LLVM
+			// versions as new features are added), check that every feature
+			// TinyGo specifies is consistent with what Clang produces:
+			//  - A "+feature" in TinyGo must appear in Clang's output.
+			//  - A "-feature" in TinyGo must not be "+feature" in Clang's output.
+			checkFeatureFlags(t, config.Features(), features)
 		}
 	}
 }

+// checkFeatureFlags verifies that all features specified by TinyGo's target
+// configuration are consistent with Clang's output. This allows Clang to add
+// new features across LLVM versions without breaking the test.
+func checkFeatureFlags(t *testing.T, targetFeatures, clangFeatures string) {
+	t.Helper()
+
+	// Build a set of Clang's features for fast lookup.
+	clangSet := make(map[string]bool) // feature name -> enabled
+	for _, f := range strings.Split(clangFeatures, ",") {
+		f = strings.TrimSpace(f)
+		if len(f) < 2 {
+			continue
+		}
+		enabled := f[0] == '+'
+		name := f[1:]
+		clangSet[name] = enabled
+	}
+
+	// Check each feature that TinyGo specifies.
+	var missing, conflicts []string
+	for _, f := range strings.Split(targetFeatures, ",") {
+		f = strings.TrimSpace(f)
+		if len(f) < 2 {
+			continue
+		}
+		wantEnabled := f[0] == '+'
+		name := f[1:]
+
+		clangEnabled, inClang := clangSet[name]
+		if wantEnabled && (!inClang || !clangEnabled) {
+			// TinyGo requires +feature but Clang doesn't enable it.
+			missing = append(missing, f)
+		} else if !wantEnabled && inClang && clangEnabled {
+			// TinyGo requires -feature but Clang enables it.
+			conflicts = append(conflicts, fmt.Sprintf("target has %q but Clang has %q", f, "+"+name))
+		}
+	}
+
+	if len(missing) > 0 {
+		t.Errorf("target specifies features not present in Clang output: %s\n\ttarget features: %s\n\tclang features:  %s",
+			strings.Join(missing, ", "), targetFeatures, clangFeatures)
+	}
+	if len(conflicts) > 0 {
+		t.Errorf("target disables features that Clang enables: %s",
+			strings.Join(conflicts, "; "))
+	}
+}
+
 // This TestMain is necessary because TinyGo may also be invoked to run certain
 // LLVM tools in a separate process. Not capturing these invocations would lead
 // to recursive tests.
@@ -3,6 +3,7 @@ package builder
 import (
 	"os"
 	"path/filepath"
+	"strings"

 	"github.com/tinygo-org/tinygo/compileopts"
 	"github.com/tinygo-org/tinygo/goenv"
@@ -201,6 +202,22 @@ var avrBuiltins = []string{
 	"avr/udivmodqi4.S",
 }

+// Builtins needed specifically for windows/386.
+var windowsI386Builtins = []string{
+	"i386/chkstk.S",  // __chkstk_ms
+	"i386/chkstk2.S", // _alloca (__alloca)
+}
+
+// Builtins needed specifically for windows/amd64.
+var windowsAMD64Builtins = []string{
+	"x86_64/chkstk.S",
+}
+
+// Builtins needed specifically for windows/arm64.
+var windowsARM64Builtins = []string{
+	"aarch64/chkstk.S",
+}
+
 // libCompilerRT is a library with symbols required by programs compiled with
 // LLVM. These symbols are for operations that cannot be emitted with a single
 // instruction or a short sequence of instructions for that target.
@@ -220,16 +237,35 @@ var libCompilerRT = Library{
 		// Development build.
 		return filepath.Join(goenv.Get("TINYGOROOT"), "lib/compiler-rt-builtins")
 	},
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		builtins := append([]string{}, genericBuiltins...) // copy genericBuiltins
 		switch compileopts.CanonicalArchName(target) {
 		case "arm":
 			builtins = append(builtins, aeabiBuiltins...)
 		case "avr":
 			builtins = append(builtins, avrBuiltins...)
-		case "x86_64", "aarch64", "riscv64": // any 64-bit arch
+		case "x86_64":
 			builtins = append(builtins, genericBuiltins128...)
+			if isWindowsTriple(target) {
+				builtins = append(builtins, windowsAMD64Builtins...)
+			}
+		case "aarch64":
+			builtins = append(builtins, genericBuiltins128...)
+			if isWindowsTriple(target) {
+				builtins = append(builtins, windowsARM64Builtins...)
+			}
+		case "riscv64":
+			builtins = append(builtins, genericBuiltins128...)
+		case "i386":
+			if isWindowsTriple(target) {
+				builtins = append(builtins, windowsI386Builtins...)
+			}
 		}
 		return builtins, nil
 	},
 }
+
+func isWindowsTriple(target string) bool {
+	parts := strings.Split(target, "-")
+	return len(parts) > 2 && parts[2] == "windows"
+}
@@ -281,7 +281,7 @@ func parseDepFile(s string) ([]string, error) {
 	s = strings.ReplaceAll(s, "\\\n", " ")

 	// Only use the first line, which is expected to begin with "deps:".
-	line := strings.SplitN(s, "\n", 2)[0]
+	line, _, _ := strings.Cut(s, "\n")
 	if !strings.HasPrefix(line, "deps:") {
 		return nil, errors.New("readDepFile: expected 'deps:' prefix")
 	}
@@ -23,7 +23,7 @@
 #include "clang/Basic/Diagnostic.h"
 #include "clang/Basic/DiagnosticOptions.h"
 #include "clang/Driver/DriverDiagnostic.h"
-#include "clang/Driver/Options.h"
+#include "clang/Options/Options.h"
 #include "clang/Frontend/FrontendDiagnostic.h"
 #include "clang/Frontend/TextDiagnosticPrinter.h"
 #include "clang/Frontend/Utils.h"
@@ -35,6 +35,7 @@
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/MC/MCCodeEmitter.h"
 #include "llvm/MC/MCContext.h"
+#include "llvm/MC/MCInstPrinter.h"
 #include "llvm/MC/MCInstrInfo.h"
 #include "llvm/MC/MCObjectFileInfo.h"
 #include "llvm/MC/MCObjectWriter.h"
@@ -67,8 +68,7 @@
 #include <optional>
 #include <system_error>
 using namespace clang;
-using namespace clang::driver;
-using namespace clang::driver::options;
+using namespace clang::options;
 using namespace llvm;
 using namespace llvm::opt;

@@ -109,7 +109,7 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
  // Construct the invocation.

  // Target Options
-  Opts.Triple = llvm::Triple::normalize(Args.getLastArgValue(OPT_triple));
+  Opts.Triple = llvm::Triple(llvm::Triple::normalize(Args.getLastArgValue(OPT_triple)));
  if (Arg *A = Args.getLastArg(options::OPT_darwin_target_variant_triple))
    Opts.DarwinTargetVariantTriple = llvm::Triple(A->getValue());
  if (Arg *A = Args.getLastArg(OPT_darwin_target_variant_sdk_version_EQ)) {
@@ -125,8 +125,8 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
  Opts.Features = Args.getAllArgValues(OPT_target_feature);

  // Use the default target triple if unspecified.
-  if (Opts.Triple.empty())
-    Opts.Triple = llvm::sys::getDefaultTargetTriple();
+  if (Opts.Triple.getTriple().empty())
+    Opts.Triple = llvm::Triple(llvm::sys::getDefaultTargetTriple());

  // Language Options
  Opts.IncludePaths = Args.getAllArgValues(OPT_I);
@@ -144,8 +144,6 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
            .Default(llvm::DebugCompressionType::None);
  }

-  Opts.RelaxELFRelocations = !Args.hasArg(OPT_mrelax_relocations_no);
-  Opts.SSE2AVX = Args.hasArg(OPT_msse2avx);
  if (auto *DwarfFormatArg = Args.getLastArg(OPT_gdwarf64, OPT_gdwarf32))
    Opts.Dwarf64 = DwarfFormatArg->getOption().matches(OPT_gdwarf64);
  Opts.DwarfVersion = getLastArgIntValue(Args, OPT_dwarf_version_EQ, 2, Diags);
@@ -236,6 +234,9 @@ bool AssemblerInvocation::CreateFromArgs(AssemblerInvocation &Opts,
  Opts.EmitCompactUnwindNonCanonical =
      Args.hasArg(OPT_femit_compact_unwind_non_canonical);
  Opts.Crel = Args.hasArg(OPT_crel);
+  Opts.ImplicitMapsyms = Args.hasArg(OPT_mmapsyms_implicit);
+  Opts.X86RelaxRelocations = !Args.hasArg(OPT_mrelax_relocations_no);
+  Opts.X86Sse2Avx = Args.hasArg(OPT_msse2avx);

  Opts.AsSecureLogFile = Args.getLastArgValue(OPT_as_secure_log_file);

@@ -266,7 +267,7 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
  std::string Error;
  const Target *TheTarget = TargetRegistry::lookupTarget(Opts.Triple, Error);
  if (!TheTarget)
-    return Diags.Report(diag::err_target_unknown_triple) << Opts.Triple;
+    return Diags.Report(diag::err_target_unknown_triple) << Opts.Triple.str();

  ErrorOr<std::unique_ptr<MemoryBuffer>> Buffer =
      MemoryBuffer::getFileOrSTDIN(Opts.InputFile, /*IsText=*/true);
@@ -294,8 +295,9 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
  MCOptions.EmitCompactUnwindNonCanonical = Opts.EmitCompactUnwindNonCanonical;
  MCOptions.MCSaveTempLabels = Opts.SaveTemporaryLabels;
  MCOptions.Crel = Opts.Crel;
-  MCOptions.X86RelaxRelocations = Opts.RelaxELFRelocations;
-  MCOptions.X86Sse2Avx = Opts.SSE2AVX;
+  MCOptions.ImplicitMapSyms = Opts.ImplicitMapsyms;
+  MCOptions.X86RelaxRelocations = Opts.X86RelaxRelocations;
+  MCOptions.X86Sse2Avx = Opts.X86Sse2Avx;
  MCOptions.CompressDebugSections = Opts.CompressDebugSections;
  MCOptions.AsSecureLogFile = Opts.AsSecureLogFile;

@@ -325,7 +327,7 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
      TheTarget->createMCSubtargetInfo(Opts.Triple, Opts.CPU, FS));
  assert(STI && "Unable to create subtarget info!");

-  MCContext Ctx(Triple(Opts.Triple), MAI.get(), MRI.get(), STI.get(), &SrcMgr,
+  MCContext Ctx(Opts.Triple, MAI.get(), MRI.get(), STI.get(), &SrcMgr,
                &MCOptions);

  bool PIC = false;
@@ -343,10 +345,6 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
  // MCObjectFileInfo needs a MCContext reference in order to initialize itself.
  std::unique_ptr<MCObjectFileInfo> MOFI(
      TheTarget->createMCObjectFileInfo(Ctx, PIC));
-  if (Opts.DarwinTargetVariantTriple)
-    MOFI->setDarwinTargetVariantTriple(*Opts.DarwinTargetVariantTriple);
-  if (!Opts.DarwinTargetVariantSDKVersion.empty())
-    MOFI->setDarwinTargetVariantSDKVersion(Opts.DarwinTargetVariantSDKVersion);
  Ctx.setObjectFileInfo(MOFI.get());

  if (Opts.GenDwarfForAssembly)
@@ -392,8 +390,8 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,

  // FIXME: There is a bit of code duplication with addPassesToEmitFile.
  if (Opts.OutputType == AssemblerInvocation::FT_Asm) {
-    MCInstPrinter *IP = TheTarget->createMCInstPrinter(
-        llvm::Triple(Opts.Triple), Opts.OutputAsmVariant, *MAI, *MCII, *MRI);
+    std::unique_ptr<MCInstPrinter> IP(TheTarget->createMCInstPrinter(
+        Opts.Triple, Opts.OutputAsmVariant, *MAI, *MCII, *MRI));

    std::unique_ptr<MCCodeEmitter> CE;
    if (Opts.ShowEncoding)
@@ -402,7 +400,7 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
        TheTarget->createMCAsmBackend(*STI, *MRI, MCOptions));

    auto FOut = std::make_unique<formatted_raw_ostream>(*Out);
-    Str.reset(TheTarget->createAsmStreamer(Ctx, std::move(FOut), IP,
+    Str.reset(TheTarget->createAsmStreamer(Ctx, std::move(FOut), std::move(IP),
                                           std::move(CE), std::move(MAB)));
  } else if (Opts.OutputType == AssemblerInvocation::FT_Null) {
    Str.reset(createNullStreamer(Ctx));
@@ -424,10 +422,17 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
        DwoOS ? MAB->createDwoObjectWriter(*Out, *DwoOS)
              : MAB->createObjectWriter(*Out);

-    Triple T(Opts.Triple);
+    Triple T = Opts.Triple;
    Str.reset(TheTarget->createMCObjectStreamer(
        T, Ctx, std::move(MAB), std::move(OW), std::move(CE), *STI));
    Str.get()->initSections(Opts.NoExecStack, *STI);
+    if (T.isOSBinFormatMachO() && T.isOSDarwin()) {
+      Triple *TVT = Opts.DarwinTargetVariantTriple
+                        ? &*Opts.DarwinTargetVariantTriple
+                        : nullptr;
+      Str->emitVersionForTarget(T, VersionTuple(), TVT,
+                                Opts.DarwinTargetVariantSDKVersion);
+    }
  }

  // When -fembed-bitcode is passed to clang_as, a 1-byte marker
@@ -448,7 +453,7 @@ static bool ExecuteAssemblerImpl(AssemblerInvocation &Opts,
  std::unique_ptr<MCTargetAsmParser> TAP(
      TheTarget->createMCAsmParser(*STI, *Parser, *MCII, MCOptions));
  if (!TAP)
-    Failed = Diags.Report(diag::err_target_unknown_triple) << Opts.Triple;
+    Failed = Diags.Report(diag::err_target_unknown_triple) << Opts.Triple.str();

  // Set values for symbols, if any.
  for (auto &S : Opts.SymbolDefs) {
@@ -501,12 +506,12 @@ int cc1as_main(ArrayRef<const char *> Argv, const char *Argv0, void *MainAddr) {
  InitializeAllAsmParsers();

  // Construct our diagnostic client.
-  IntrusiveRefCntPtr<DiagnosticOptions> DiagOpts = new DiagnosticOptions();
+  DiagnosticOptions DiagOpts;
  TextDiagnosticPrinter *DiagClient
-    = new TextDiagnosticPrinter(errs(), &*DiagOpts);
+    = new TextDiagnosticPrinter(errs(), DiagOpts);
  DiagClient->setPrefix("clang -cc1as");
  IntrusiveRefCntPtr<DiagnosticIDs> DiagID(new DiagnosticIDs());
-  DiagnosticsEngine Diags(DiagID, &*DiagOpts, DiagClient);
+  DiagnosticsEngine Diags(DiagID, DiagOpts, DiagClient);

  // Set an error handler, so that any LLVM backend diagnostics go through our
  // error handler.
@@ -523,7 +528,7 @@ int cc1as_main(ArrayRef<const char *> Argv, const char *Argv0, void *MainAddr) {
        llvm::outs(), "clang -cc1as [options] file...",
        "Clang Integrated Assembler", /*ShowHidden=*/false,
        /*ShowAllAliases=*/false,
-        llvm::opt::Visibility(driver::options::CC1AsOption));
+        llvm::opt::Visibility(clang::options::CC1AsOption));

    return 0;
  }
@@ -20,7 +20,7 @@ struct AssemblerInvocation {
  /// @{

  /// The name of the target triple to assemble for.
-  std::string Triple;
+  llvm::Triple Triple;

  /// If given, the name of the target CPU to determine which instructions
  /// are legal.
@@ -45,10 +45,6 @@ struct AssemblerInvocation {
  LLVM_PREFERRED_TYPE(bool)
  unsigned GenDwarfForAssembly : 1;
  LLVM_PREFERRED_TYPE(bool)
-  unsigned RelaxELFRelocations : 1;
-  LLVM_PREFERRED_TYPE(bool)
-  unsigned SSE2AVX : 1;
-  LLVM_PREFERRED_TYPE(bool)
  unsigned Dwarf64 : 1;
  unsigned DwarfVersion;
  std::string DwarfDebugFlags;
@@ -111,12 +107,19 @@ struct AssemblerInvocation {
  EmitDwarfUnwindType EmitDwarfUnwind;

  // Whether to emit compact-unwind for non-canonical entries.
-  // Note: maybe overriden by other constraints.
+  // Note: maybe overridden by other constraints.
  LLVM_PREFERRED_TYPE(bool)
  unsigned EmitCompactUnwindNonCanonical : 1;

  LLVM_PREFERRED_TYPE(bool)
  unsigned Crel : 1;
+  LLVM_PREFERRED_TYPE(bool)
+  unsigned ImplicitMapsyms : 1;
+
+  LLVM_PREFERRED_TYPE(bool)
+  unsigned X86RelaxRelocations : 1;
+  LLVM_PREFERRED_TYPE(bool)
+  unsigned X86Sse2Avx : 1;

  /// The name of the relocation model to use.
  std::string RelocationModel;
@@ -139,7 +142,7 @@ struct AssemblerInvocation {

 public:
  AssemblerInvocation() {
-    Triple = "";
+    Triple = llvm::Triple();
    NoInitialTextSection = 0;
    InputFile = "-";
    OutputPath = "-";
@@ -148,7 +151,6 @@ public:
    ShowInst = 0;
    ShowEncoding = 0;
    RelaxAll = 0;
-    SSE2AVX = 0;
    NoExecStack = 0;
    FatalWarnings = 0;
    NoWarn = 0;
@@ -160,6 +162,9 @@ public:
    EmitDwarfUnwind = EmitDwarfUnwindType::Default;
    EmitCompactUnwindNonCanonical = false;
    Crel = false;
+    ImplicitMapsyms = 0;
+    X86RelaxRelocations = 0;
+    X86Sse2Avx = 0;
  }

  static bool CreateFromArgs(AssemblerInvocation &Res,
@@ -27,9 +27,9 @@ bool tinygo_clang_driver(int argc, char **argv) {
 	std::vector<const char*> args(argv, argv + argc);

 	// The compiler invocation needs a DiagnosticsEngine so it can report problems
-	llvm::IntrusiveRefCntPtr<clang::DiagnosticOptions> DiagOpts = new clang::DiagnosticOptions();
-	clang::TextDiagnosticPrinter DiagnosticPrinter(llvm::errs(), &*DiagOpts);
-	clang::DiagnosticsEngine Diags(llvm::IntrusiveRefCntPtr<clang::DiagnosticIDs>(new clang::DiagnosticIDs()), &*DiagOpts, &DiagnosticPrinter, false);
+	clang::DiagnosticOptions DiagOpts;
+	clang::TextDiagnosticPrinter DiagnosticPrinter(llvm::errs(), DiagOpts);
+	clang::DiagnosticsEngine Diags(llvm::IntrusiveRefCntPtr<clang::DiagnosticIDs>(new clang::DiagnosticIDs()), DiagOpts, &DiagnosticPrinter, false);

 	// Create the clang driver
 	clang::driver::Driver TheDriver(args[0], llvm::sys::getDefaultTargetTriple(), Diags);
@@ -17,7 +17,7 @@ import (
 var commands = map[string][]string{}

 func init() {
-	llvmMajor := strings.Split(llvm.Version, ".")[0]
+	llvmMajor, _, _ := strings.Cut(llvm.Version, ".")
 	commands["clang"] = []string{"clang-" + llvmMajor}
 	commands["ld.lld"] = []string{"ld.lld-" + llvmMajor, "ld.lld"}
 	commands["wasm-ld"] = []string{"wasm-ld-" + llvmMajor, "wasm-ld"}
@@ -25,18 +25,21 @@ func NewConfig(options *compileopts.Options) (*compileopts.Config, error) {
 	}

 	// Version range supported by TinyGo.
-	const minorMin = 19
-	const minorMax = 24
+	const minorMin = 25 // when updating the min version, also update .github/workflows/compat.yml
+	const minorMax = 27

 	// Check that we support this Go toolchain version.
 	gorootMajor, gorootMinor, err := goenv.GetGorootVersion()
 	if err != nil {
 		return nil, err
 	}
-	if gorootMajor != 1 || gorootMinor < minorMin || gorootMinor > minorMax {
-		// Note: when this gets updated, also update the Go compatibility matrix:
-		// https://github.com/tinygo-org/tinygo-site/blob/dev/content/docs/reference/go-compat-matrix.md
-		return nil, fmt.Errorf("requires go version 1.%d through 1.%d, got go%d.%d", minorMin, minorMax, gorootMajor, gorootMinor)
+
+	if options.GoCompatibility {
+		if gorootMajor != 1 || gorootMinor < minorMin || gorootMinor > minorMax {
+			// Note: when this gets updated, also update the Go compatibility matrix:
+			// https://github.com/tinygo-org/tinygo-site/blob/dev/content/docs/reference/go-compat-matrix.md
+			return nil, fmt.Errorf("requires go version 1.%d through 1.%d, got go%d.%d", minorMin, minorMax, gorootMajor, gorootMinor)
+		}
 	}

 	// Check that the Go toolchain version isn't too new, if we haven't been
@@ -15,7 +15,7 @@ func makeDarwinLibSystemJob(config *compileopts.Config, tmpdir string) *compileJ
 	return &compileJob{
 		description: "compile Darwin libSystem.dylib",
 		run: func(job *compileJob) (err error) {
-			arch := strings.Split(config.Triple(), "-")[0]
+			arch, _, _ := strings.Cut(config.Triple(), "-")
 			job.result = filepath.Join(tmpdir, "libSystem.dylib")
 			objpath := filepath.Join(tmpdir, "libSystem.o")
 			inpath := filepath.Join(goenv.Get("TINYGOROOT"), "lib/macos-minimal-sdk/src", arch, "libSystem.s")
@@ -23,7 +23,7 @@ type espImageSegment struct {
 	data []byte
 }

-// makeESPFirmareImage converts an input ELF file to an image file for an ESP32 or
+// makeESPFirmwareImage converts an input ELF file to an image file for an ESP32 or
 // ESP8266 chip. This is a special purpose image format just for the ESP chip
 // family, and is parsed by the on-chip mask ROM bootloader.
 //
@@ -31,7 +31,7 @@ type espImageSegment struct {
 // https://github.com/espressif/esptool/wiki/Firmware-Image-Format
 // https://github.com/espressif/esp-idf/blob/8fbb63c2a701c22ccf4ce249f43aded73e134a34/components/bootloader_support/include/esp_image_format.h#L58
 // https://github.com/espressif/esptool/blob/master/esptool.py
-func makeESPFirmareImage(infile, outfile, format string) error {
+func makeESPFirmwareImage(infile, outfile, format string) error {
 	inf, err := elf.Open(infile)
 	if err != nil {
 		return err
@@ -65,15 +65,6 @@ func makeESPFirmareImage(infile, outfile, format string) error {
 	// Sort the segments by address. This is what esptool does too.
 	sort.SliceStable(segments, func(i, j int) bool { return segments[i].addr < segments[j].addr })

-	// Calculate checksum over the segment data. This is used in the image
-	// footer.
-	checksum := uint8(0xef)
-	for _, segment := range segments {
-		for _, b := range segment.data {
-			checksum ^= b
-		}
-	}
-
 	// Write first to an in-memory buffer, primarily so that we can easily
 	// calculate a hash over the entire image.
 	// An added benefit is that we don't need to check for errors all the time.
@@ -88,6 +79,83 @@ func makeESPFirmareImage(infile, outfile, format string) error {
 		chip = format[:len(format)-len("-img")]
 	}

+	// For ESP32 (original): separate RAM segments (loadable by ROM bootloader)
+	// from flash-mapped segments (DROM/IROM, require MMU setup by startup code).
+	// The ROM bootloader on ESP32 does NOT handle flash-mapped segments —
+	// it tries to memcpy to the virtual address, which crashes.
+	var flashSegments []*espImageSegment
+	if chip == "esp32" {
+		var ramSegments []*espImageSegment
+		for _, seg := range segments {
+			if (seg.addr >= 0x3F400000 && seg.addr < 0x3F800000) || // DROM
+				(seg.addr >= 0x400D0000 && seg.addr < 0x40400000) { // IROM
+				flashSegments = append(flashSegments, seg)
+			} else {
+				ramSegments = append(ramSegments, seg)
+			}
+		}
+		segments = ramSegments
+	}
+
+	// ESP32 flash XIP: compute where the DROM segment will be placed in flash
+	// (page-aligned, right after the RAM segments) and patch the
+	// _drom_flash_addr variable so the startup code can program the cache MMU.
+	// This must happen before the checksum/hash are computed so the patched
+	// value is covered by both.
+	const esp32FlashBase = 0x1000 // esptool flashes the image at 0x1000
+	// The ESP32 flash cache MMU supports configurable page sizes down to 256 B. 64 KiB is the reset/default size.
+	// If the startup code ever changes the MMU page size, this constant must change too.
+	const esp32PageSize = 0x10000 // 64KB MMU pages
+	var esp32DromFlashAddr uint32
+	if chip == "esp32" && len(flashSegments) > 0 {
+		// Compute the size of the RAM portion of the image (everything the ROM
+		// bootloader loads, up to and including the appended SHA256 hash).
+		ramImageSize := 0
+		ramImageSize += 24 // image header (8) + trailer fields (16)
+		for _, seg := range segments {
+			ramImageSize += 8 + len(seg.data) // segment header + data (4-aligned)
+		}
+		ramImageSize += 16 - ramImageSize%16 // footer padding + checksum byte
+		ramImageSize += 32                   // appended SHA256 hash
+
+		// DROM flash address must be 64KB page-aligned.
+		esp32DromFlashAddr = uint32(esp32FlashBase+ramImageSize+esp32PageSize-1) &^ (esp32PageSize - 1)
+
+		// Patch _drom_flash_addr in whichever RAM segment contains it.
+		syms, _ := inf.Symbols()
+		var dromSymAddr uint64
+		for _, s := range syms {
+			if s.Name == "_drom_flash_addr" {
+				dromSymAddr = s.Value
+				break
+			}
+		}
+		if dromSymAddr == 0 {
+			return fmt.Errorf("ESP32: _drom_flash_addr symbol not found")
+		}
+		patched := false
+		for _, seg := range segments {
+			if dromSymAddr >= uint64(seg.addr) && dromSymAddr+4 <= uint64(seg.addr)+uint64(len(seg.data)) {
+				off := int(dromSymAddr - uint64(seg.addr))
+				binary.LittleEndian.PutUint32(seg.data[off:], esp32DromFlashAddr)
+				patched = true
+				break
+			}
+		}
+		if !patched {
+			return fmt.Errorf("ESP32: _drom_flash_addr (0x%x) not in any RAM segment", dromSymAddr)
+		}
+	}
+
+	// Calculate checksum over the segment data. This is used in the image
+	// footer.
+	checksum := uint8(0xef)
+	for _, segment := range segments {
+		for _, b := range segment.data {
+			checksum ^= b
+		}
+	}
+
 	if makeImage {
 		// The bootloader starts at 0x1000, or 4096.
 		// TinyGo doesn't use a separate bootloader and runs the entire
@@ -100,11 +168,24 @@ func makeESPFirmareImage(infile, outfile, format string) error {
 	chip_id := map[string]uint16{
 		"esp32":   0x0000,
 		"esp32c3": 0x0005,
+		"esp32c6": 0x000d,
+		"esp32s3": 0x0009,
+	}[chip]
+
+	// SPI flash speed/size byte (byte 3 of header):
+	//   Upper nibble = flash size, lower nibble = flash frequency.
+	//   The espflasher auto-detects and patches the flash size (upper nibble),
+	//   but the frequency (lower nibble) must be correct per chip.
+	spiSpeedSize := map[string]uint8{
+		"esp32":   0x1f, // 80MHz=0x0F, 2MB=0x10
+		"esp32c3": 0x1f, // 80MHz=0x0F, 2MB=0x10
+		"esp32c6": 0x10, // 80MHz=0x00, 2MB=0x10 (C6 uses different freq encoding)
+		"esp32s3": 0x1f, // 80MHz=0x0F, 2MB=0x10
 	}[chip]

 	// Image header.
 	switch chip {
-	case "esp32", "esp32c3":
+	case "esp32", "esp32c3", "esp32s3", "esp32c6":
 		// Header format:
 		// https://github.com/espressif/esp-idf/blob/v4.3/components/bootloader_support/include/esp_app_format.h#L71
 		// Note: not adding a SHA256 hash as the binary is modified by
@@ -125,8 +206,8 @@ func makeESPFirmareImage(infile, outfile, format string) error {
 		}{
 			magic:          0xE9,
 			segment_count:  byte(len(segments)),
-			spi_mode:       2,    // ESP_IMAGE_SPI_MODE_DIO
-			spi_speed_size: 0x1f, // ESP_IMAGE_SPI_SPEED_80M, ESP_IMAGE_FLASH_SIZE_2MB
+			spi_mode:       2, // ESP_IMAGE_SPI_MODE_DIO
+			spi_speed_size: spiSpeedSize,
 			entry_addr:     uint32(inf.Entry),
 			wp_pin:         0xEE, // disable WP pin
 			chip_id:        chip_id,
@@ -178,6 +259,64 @@ func makeESPFirmareImage(infile, outfile, format string) error {
 		outf.Write(hash[:])
 	}

+	// For ESP32: append flash-mapped segments (DROM/IROM) at page-aligned flash
+	// offsets after the RAM portion. The startup code maps them via the flash
+	// cache MMU (DROM at esp32DromFlashAddr, patched into _drom_flash_addr).
+	if len(flashSegments) > 0 {
+		const flashBase = esp32FlashBase
+		const pageSize = esp32PageSize
+		dromFlashAddr := esp32DromFlashAddr
+
+		// Separate DROM and IROM segments.
+		var dromSegs, iromSegs []*espImageSegment
+		for _, seg := range flashSegments {
+			if seg.addr >= 0x3F400000 && seg.addr < 0x3F800000 {
+				dromSegs = append(dromSegs, seg)
+			} else {
+				iromSegs = append(iromSegs, seg)
+			}
+		}
+
+		// Write DROM segments at the computed page-aligned flash offset.
+		dromSize := 0
+		if len(dromSegs) > 0 {
+			targetImageOffset := int(dromFlashAddr - flashBase)
+			if makeImage {
+				targetImageOffset = int(dromFlashAddr)
+			}
+			if outf.Len() > targetImageOffset {
+				return fmt.Errorf("ESP32: RAM segments too large (%d bytes), overlap DROM at flash 0x%x", outf.Len(), dromFlashAddr)
+			}
+			outf.Write(make([]byte, targetImageOffset-outf.Len()))
+			for _, seg := range dromSegs {
+				outf.Write(seg.data)
+				dromSize += len(seg.data)
+			}
+		}
+
+		// Write IROM segments immediately after DROM, at the next page boundary.
+		// IROM flash addr = dromFlashAddr + ceil(dromSize/pageSize)*pageSize
+		// (must match the computation in the startup assembly).
+		if len(iromSegs) > 0 {
+			dromPages := (dromSize + pageSize - 1) / pageSize
+			if dromPages == 0 {
+				dromPages = 1
+			}
+			iromFlashAddr := dromFlashAddr + uint32(dromPages)*pageSize
+			targetImageOffset := int(iromFlashAddr - flashBase)
+			if makeImage {
+				targetImageOffset = int(iromFlashAddr)
+			}
+			if outf.Len() > targetImageOffset {
+				return fmt.Errorf("ESP32: DROM too large, overlaps IROM at flash 0x%x", iromFlashAddr)
+			}
+			outf.Write(make([]byte, targetImageOffset-outf.Len()))
+			for _, seg := range iromSegs {
+				outf.Write(seg.data)
+			}
+		}
+	}
+
 	// QEMU (or more precisely, qemu-system-xtensa from Espressif) expects the
 	// image to be a certain size.
 	if makeImage {
@@ -195,11 +195,11 @@ type intHeap struct {
 	sort.IntSlice
 }

-func (h *intHeap) Push(x interface{}) {
+func (h *intHeap) Push(x any) {
 	h.IntSlice = append(h.IntSlice, x.(int))
 }

-func (h *intHeap) Pop() interface{} {
+func (h *intHeap) Pop() any {
 	x := h.IntSlice[len(h.IntSlice)-1]
 	h.IntSlice = h.IntSlice[:len(h.IntSlice)-1]
 	return x
@@ -15,6 +15,9 @@ import (

 // Library is a container for information about a single C library, such as a
 // compiler runtime or libc.
+//
+// Note: whenever a library gets changed, the version in compileopts/config.go
+// probably also needs to be incremented.
 type Library struct {
 	// The library name, such as compiler-rt or picolibc.
 	name string
@@ -25,11 +28,17 @@ type Library struct {
 	// cflags returns the C flags specific to this library
 	cflags func(target, headerPath string) []string

+	// cflagsForFile returns additional C flags for a particular source file.
+	cflagsForFile func(path string) []string
+
+	// needsLibc is set to true if this library needs libc headers.
+	needsLibc bool
+
 	// The source directory.
 	sourceDir func() string

 	// The source files, relative to sourceDir.
-	librarySources func(target string) ([]string, error)
+	librarySources func(target string, libcNeedsMalloc bool) ([]string, error)

 	// The source code for the crt1.o file, relative to sourceDir.
 	crt1Source string
@@ -43,18 +52,9 @@ type Library struct {
 // output archive file, it is expected to be removed after use.
 // As a side effect, this call creates the library header files if they didn't
 // exist yet.
-// The provided libc job (if not null) will cause this libc to be added as a
-// dependency for all C compiler jobs, and adds libc headers for the given
-// target config. In other words, pass this libc if the library needs a libc to
-// compile.
-func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJob) (job *compileJob, abortLock func(), err error) {
-	outdir, precompiled := config.LibcPath(l.name)
+func (l *Library) load(config *compileopts.Config, tmpdir string) (job *compileJob, abortLock func(), err error) {
+	outdir := config.LibraryPath(l.name)
 	archiveFilePath := filepath.Join(outdir, "lib.a")
-	if precompiled {
-		// Found a precompiled library for this OS/architecture. Return the path
-		// directly.
-		return dummyCompileJob(archiveFilePath), func() {}, nil
-	}

 	// Create a lock on the output (if supported).
 	// This is a bit messy, but avoids a deadlock because it is ordered consistently with other library loads within a build.
@@ -133,7 +133,7 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 	// Note: -fdebug-prefix-map is necessary to make the output archive
 	// reproducible. Otherwise the temporary directory is stored in the archive
 	// itself, which varies each run.
-	args := append(l.cflags(target, headerPath), "-c", "-Oz", "-gdwarf-4", "-ffunction-sections", "-fdata-sections", "-Wno-macro-redefined", "--target="+target, "-fdebug-prefix-map="+dir+"="+remapDir)
+	args := append(l.cflags(target, headerPath), "-c", "-Oz", "-gdwarf-4", "-ffunction-sections", "-fdata-sections", "-Wno-macro-redefined", "--target="+compileopts.ClangTriple(target), "-fdebug-prefix-map="+dir+"="+remapDir)
 	resourceDir := goenv.ClangResourceDir(false)
 	if resourceDir != "" {
 		args = append(args, "-resource-dir="+resourceDir)
@@ -167,9 +167,9 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 		// double.
 		args = append(args, "-mdouble=64")
 	case "riscv32":
-		args = append(args, "-march=rv32imac", "-fforce-enable-int128")
+		args = append(args, "-march="+riscvMarch(config, "rv32imac"), "-fforce-enable-int128")
 	case "riscv64":
-		args = append(args, "-march=rv64gc")
+		args = append(args, "-march="+riscvMarch(config, "rv64gc"))
 	case "mips":
 		args = append(args, "-fno-pic")
 	}
@@ -185,7 +185,7 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 			args = append(args, "-mfpu=vfpv2")
 		}
 	}
-	if libc != nil {
+	if l.needsLibc {
 		args = append(args, config.LibcCFlags()...)
 	}

@@ -218,7 +218,7 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 				return err
 			}
 			// Store this archive in the cache.
-			return os.Rename(f.Name(), archiveFilePath)
+			return robustRename(f.Name(), archiveFilePath)
 		},
 	}

@@ -226,7 +226,7 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ

 	// Create jobs to compile all sources. These jobs are depended upon by the
 	// archive job above, so must be run first.
-	paths, err := l.librarySources(target)
+	paths, err := l.librarySources(target, config.LibcNeedsMalloc())
 	if err != nil {
 		return nil, nil, err
 	}
@@ -245,6 +245,9 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 			run: func(*compileJob) error {
 				var compileArgs []string
 				compileArgs = append(compileArgs, args...)
+				if l.cflagsForFile != nil {
+					compileArgs = append(compileArgs, l.cflagsForFile(path)...)
+				}
 				compileArgs = append(compileArgs, "-o", objpath, srcpath)
 				if config.Options.PrintCommands != nil {
 					config.Options.PrintCommands("clang", compileArgs...)
@@ -256,9 +259,6 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 				return nil
 			},
 		}
-		if libc != nil {
-			objfile.dependencies = append(objfile.dependencies, libc)
-		}
 		job.dependencies = append(job.dependencies, objfile)
 	}

@@ -289,9 +289,6 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 				return os.Rename(tmpfile.Name(), filepath.Join(outdir, "crt1.o"))
 			},
 		}
-		if libc != nil {
-			crt1Job.dependencies = append(crt1Job.dependencies, libc)
-		}
 		job.dependencies = append(job.dependencies, crt1Job)
 	}

@@ -300,3 +297,17 @@ func (l *Library) load(config *compileopts.Config, tmpdir string, libc *compileJ
 		once.Do(unlock)
 	}, nil
 }
+
+// riscvMarch returns the -march value for RISC-V library compilation.
+// It extracts the value from the target's cflags if present, otherwise
+// falls back to the provided default. This ensures libraries are compiled
+// with the correct ISA extensions for each target (e.g. rv32imc for
+// ESP32-C3 which lacks the atomic extension).
+func riscvMarch(config *compileopts.Config, defaultMarch string) string {
+	for _, flag := range config.Target.CFlags {
+		if strings.HasPrefix(flag, "-march=") {
+			return flag[len("-march="):]
+		}
+	}
+	return defaultMarch
+}
@@ -30,25 +30,62 @@ var libMinGW = Library{
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/mingw-w64") },
 	cflags: func(target, headerPath string) []string {
 		mingwDir := filepath.Join(goenv.Get("TINYGOROOT"), "lib/mingw-w64")
-		return []string{
+		flags := []string{
 			"-nostdlibinc",
+			"-isystem", mingwDir + "/mingw-w64-crt/include",
 			"-isystem", mingwDir + "/mingw-w64-headers/crt",
+			"-isystem", mingwDir + "/mingw-w64-headers/include",
 			"-I", mingwDir + "/mingw-w64-headers/defaults/include",
 			"-I" + headerPath,
 		}
+		if strings.Split(target, "-")[0] == "i386" {
+			flags = append(flags,
+				"-D__MSVCRT_VERSION__=0x700", // Microsoft Visual C++ .NET 2002
+				"-D_WIN32_WINNT=0x0501",      // target Windows XP
+				"-D_CRTBLD",
+				"-Wno-pragma-pack",
+			)
+		}
+		return flags
 	},
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		// These files are needed so that printf and the like are supported.
-		sources := []string{
-			"mingw-w64-crt/stdio/ucrt_fprintf.c",
-			"mingw-w64-crt/stdio/ucrt_fwprintf.c",
-			"mingw-w64-crt/stdio/ucrt_printf.c",
-			"mingw-w64-crt/stdio/ucrt_snprintf.c",
-			"mingw-w64-crt/stdio/ucrt_sprintf.c",
-			"mingw-w64-crt/stdio/ucrt_vfprintf.c",
-			"mingw-w64-crt/stdio/ucrt_vprintf.c",
-			"mingw-w64-crt/stdio/ucrt_vsnprintf.c",
-			"mingw-w64-crt/stdio/ucrt_vsprintf.c",
+		var sources []string
+		if strings.Split(target, "-")[0] == "i386" {
+			// Old 32-bit x86 systems use msvcrt.dll.
+			sources = []string{
+				"mingw-w64-crt/crt/pseudo-reloc.c",
+				"mingw-w64-crt/gdtoa/dmisc.c",
+				"mingw-w64-crt/gdtoa/gdtoa.c",
+				"mingw-w64-crt/gdtoa/gmisc.c",
+				"mingw-w64-crt/gdtoa/misc.c",
+				"mingw-w64-crt/math/x86/exp2.S",
+				"mingw-w64-crt/math/x86/trunc.S",
+				"mingw-w64-crt/misc/___mb_cur_max_func.c",
+				"mingw-w64-crt/misc/lc_locale_func.c",
+				"mingw-w64-crt/misc/mbrtowc.c",
+				"mingw-w64-crt/misc/strnlen.c",
+				"mingw-w64-crt/misc/wcrtomb.c",
+				"mingw-w64-crt/misc/wcsnlen.c",
+				"mingw-w64-crt/stdio/acrt_iob_func.c",
+				"mingw-w64-crt/stdio/mingw_lock.c",
+				"mingw-w64-crt/stdio/mingw_pformat.c",
+				"mingw-w64-crt/stdio/mingw_vfprintf.c",
+				"mingw-w64-crt/stdio/mingw_vsnprintf.c",
+			}
+		} else {
+			// Anything somewhat modern (amd64, arm64) uses UCRT.
+			sources = []string{
+				"mingw-w64-crt/stdio/ucrt_fprintf.c",
+				"mingw-w64-crt/stdio/ucrt_fwprintf.c",
+				"mingw-w64-crt/stdio/ucrt_printf.c",
+				"mingw-w64-crt/stdio/ucrt_snprintf.c",
+				"mingw-w64-crt/stdio/ucrt_sprintf.c",
+				"mingw-w64-crt/stdio/ucrt_vfprintf.c",
+				"mingw-w64-crt/stdio/ucrt_vprintf.c",
+				"mingw-w64-crt/stdio/ucrt_vsnprintf.c",
+				"mingw-w64-crt/stdio/ucrt_vsprintf.c",
+			}
 		}
 		return sources, nil
 	},
@@ -63,27 +100,41 @@ var libMinGW = Library{
 func makeMinGWExtraLibs(tmpdir, goarch string) []*compileJob {
 	var jobs []*compileJob
 	root := goenv.Get("TINYGOROOT")
-	// Normally all the api-ms-win-crt-*.def files are all compiled to a single
-	// .lib file. But to simplify things, we're going to leave them as separate
-	// files.
-	for _, name := range []string{
-		"kernel32.def.in",
-		"api-ms-win-crt-conio-l1-1-0.def",
-		"api-ms-win-crt-convert-l1-1-0.def.in",
-		"api-ms-win-crt-environment-l1-1-0.def",
-		"api-ms-win-crt-filesystem-l1-1-0.def",
-		"api-ms-win-crt-heap-l1-1-0.def",
-		"api-ms-win-crt-locale-l1-1-0.def",
-		"api-ms-win-crt-math-l1-1-0.def.in",
-		"api-ms-win-crt-multibyte-l1-1-0.def",
-		"api-ms-win-crt-private-l1-1-0.def.in",
-		"api-ms-win-crt-process-l1-1-0.def",
-		"api-ms-win-crt-runtime-l1-1-0.def.in",
-		"api-ms-win-crt-stdio-l1-1-0.def",
-		"api-ms-win-crt-string-l1-1-0.def",
-		"api-ms-win-crt-time-l1-1-0.def",
-		"api-ms-win-crt-utility-l1-1-0.def",
-	} {
+	var libs []string
+	if goarch == "386" {
+		libs = []string{
+			// x86 uses msvcrt.dll instead of UCRT for compatibility with old
+			// Windows versions.
+			"advapi32.def.in",
+			"kernel32.def.in",
+			"msvcrt.def.in",
+		}
+	} else {
+		// Use the modernized UCRT on new systems.
+		// Normally all the api-ms-win-crt-*.def files are all compiled to a
+		// single .lib file. But to simplify things, we're going to leave them
+		// as separate files.
+		libs = []string{
+			"advapi32.def.in",
+			"kernel32.def.in",
+			"api-ms-win-crt-conio-l1-1-0.def",
+			"api-ms-win-crt-convert-l1-1-0.def.in",
+			"api-ms-win-crt-environment-l1-1-0.def",
+			"api-ms-win-crt-filesystem-l1-1-0.def",
+			"api-ms-win-crt-heap-l1-1-0.def",
+			"api-ms-win-crt-locale-l1-1-0.def",
+			"api-ms-win-crt-math-l1-1-0.def.in",
+			"api-ms-win-crt-multibyte-l1-1-0.def",
+			"api-ms-win-crt-private-l1-1-0.def.in",
+			"api-ms-win-crt-process-l1-1-0.def",
+			"api-ms-win-crt-runtime-l1-1-0.def.in",
+			"api-ms-win-crt-stdio-l1-1-0.def",
+			"api-ms-win-crt-string-l1-1-0.def",
+			"api-ms-win-crt-time-l1-1-0.def",
+			"api-ms-win-crt-utility-l1-1-0.def",
+		}
+	}
+	for _, name := range libs {
 		outpath := filepath.Join(tmpdir, filepath.Base(name)+".lib")
 		inpath := filepath.Join(root, "lib/mingw-w64/mingw-w64-crt/lib-common/"+name)
 		job := &compileJob{
@@ -93,6 +144,9 @@ func makeMinGWExtraLibs(tmpdir, goarch string) []*compileJob {
 				defpath := inpath
 				var archDef, emulation string
 				switch goarch {
+				case "386":
+					archDef = "-DDEF_I386"
+					emulation = "i386pe"
 				case "amd64":
 					archDef = "-DDEF_X64"
 					emulation = "i386pep"
@@ -12,6 +12,45 @@ import (
 	"github.com/tinygo-org/tinygo/goenv"
 )

+// Create the alltypes.h file from the appropriate alltypes.h.in files.
+func buildMuslAllTypes(arch, muslDir, outputBitsDir string) error {
+	// Create the file alltypes.h.
+	f, err := os.Create(filepath.Join(outputBitsDir, "alltypes.h"))
+	if err != nil {
+		return err
+	}
+	infiles := []string{
+		filepath.Join(muslDir, "arch", arch, "bits", "alltypes.h.in"),
+		filepath.Join(muslDir, "include", "alltypes.h.in"),
+	}
+	for _, infile := range infiles {
+		data, err := os.ReadFile(infile)
+		if err != nil {
+			return err
+		}
+		lines := strings.SplitSeq(string(data), "\n")
+		for line := range lines {
+			if strings.HasPrefix(line, "TYPEDEF ") {
+				matches := regexp.MustCompile(`TYPEDEF (.*) ([^ ]*);`).FindStringSubmatch(line)
+				value := matches[1]
+				name := matches[2]
+				line = fmt.Sprintf("#if defined(__NEED_%s) && !defined(__DEFINED_%s)\ntypedef %s %s;\n#define __DEFINED_%s\n#endif\n", name, name, value, name, name)
+			}
+			if strings.HasPrefix(line, "STRUCT ") {
+				matches := regexp.MustCompile(`STRUCT * ([^ ]*) (.*);`).FindStringSubmatch(line)
+				name := matches[1]
+				value := matches[2]
+				line = fmt.Sprintf("#if defined(__NEED_struct_%s) && !defined(__DEFINED_struct_%s)\nstruct %s %s;\n#define __DEFINED_struct_%s\n#endif\n", name, name, name, value, name)
+			}
+			_, err := f.WriteString(line + "\n")
+			if err != nil {
+				return err
+			}
+		}
+	}
+	return f.Close()
+}
+
 var libMusl = Library{
 	name: "musl",
 	makeHeaders: func(target, includeDir string) error {
@@ -24,41 +63,13 @@ var libMusl = Library{
 		arch := compileopts.MuslArchitecture(target)
 		muslDir := filepath.Join(goenv.Get("TINYGOROOT"), "lib", "musl")

-		// Create the file alltypes.h.
-		f, err := os.Create(filepath.Join(bits, "alltypes.h"))
+		err = buildMuslAllTypes(arch, muslDir, bits)
 		if err != nil {
 			return err
 		}
-		infiles := []string{
-			filepath.Join(muslDir, "arch", arch, "bits", "alltypes.h.in"),
-			filepath.Join(muslDir, "include", "alltypes.h.in"),
-		}
-		for _, infile := range infiles {
-			data, err := os.ReadFile(infile)
-			if err != nil {
-				return err
-			}
-			lines := strings.Split(string(data), "\n")
-			for _, line := range lines {
-				if strings.HasPrefix(line, "TYPEDEF ") {
-					matches := regexp.MustCompile(`TYPEDEF (.*) ([^ ]*);`).FindStringSubmatch(line)
-					value := matches[1]
-					name := matches[2]
-					line = fmt.Sprintf("#if defined(__NEED_%s) && !defined(__DEFINED_%s)\ntypedef %s %s;\n#define __DEFINED_%s\n#endif\n", name, name, value, name, name)
-				}
-				if strings.HasPrefix(line, "STRUCT ") {
-					matches := regexp.MustCompile(`STRUCT * ([^ ]*) (.*);`).FindStringSubmatch(line)
-					name := matches[1]
-					value := matches[2]
-					line = fmt.Sprintf("#if defined(__NEED_struct_%s) && !defined(__DEFINED_struct_%s)\nstruct %s %s;\n#define __DEFINED_struct_%s\n#endif\n", name, name, name, value, name)
-				}
-				f.WriteString(line + "\n")
-			}
-		}
-		f.Close()

 		// Create the file syscall.h.
-		f, err = os.Create(filepath.Join(bits, "syscall.h"))
+		f, err := os.Create(filepath.Join(bits, "syscall.h"))
 		if err != nil {
 			return err
 		}
@@ -110,9 +121,10 @@ var libMusl = Library{
 		return cflags
 	},
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/musl/src") },
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		arch := compileopts.MuslArchitecture(target)
 		globs := []string{
+			"conf/*.c",
 			"ctype/*.c",
 			"env/*.c",
 			"errno/*.c",
@@ -43,7 +43,7 @@ var libPicolibc = Library{
 		}
 	},
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/picolibc/newlib") },
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		sources := append([]string(nil), picolibcSources...)
 		if !strings.HasPrefix(target, "avr") {
 			// Small chips without long jumps can't compile many files (printf,
@@ -0,0 +1,9 @@
+//go:build !windows
+
+package builder
+
+import "os"
+
+func robustRename(oldpath, newpath string) error {
+	return os.Rename(oldpath, newpath)
+}
@@ -0,0 +1,44 @@
+package builder
+
+import (
+	"errors"
+	"math/rand"
+	"os"
+	"syscall"
+	"time"
+)
+
+const robustRenameTimeout = 2 * time.Second
+
+func robustRename(oldpath, newpath string) error {
+	var bestErr error
+	start := time.Now()
+	nextSleep := time.Millisecond
+	for {
+		err := os.Rename(oldpath, newpath)
+		if err == nil || !isEphemeralRenameError(err) {
+			return err
+		}
+		if bestErr == nil {
+			bestErr = err
+		}
+		if d := time.Since(start) + nextSleep; d >= robustRenameTimeout {
+			return bestErr
+		}
+		time.Sleep(nextSleep)
+		nextSleep += time.Duration(rand.Int63n(int64(nextSleep)))
+	}
+}
+
+func isEphemeralRenameError(err error) bool {
+	var errno syscall.Errno
+	if errors.As(err, &errno) {
+		switch errno {
+		case syscall.Errno(2), // ERROR_FILE_NOT_FOUND
+			syscall.Errno(5),  // ERROR_ACCESS_DENIED
+			syscall.Errno(32): // ERROR_SHARING_VIOLATION
+			return true
+		}
+	}
+	return false
+}
@@ -490,7 +490,7 @@ func loadProgramSize(path string, packagePathMap map[string]string) (*programSiz
 				continue
 			}
 			if section.Type == elf.SHT_NOBITS {
-				if section.Name == ".stack" {
+				if strings.HasPrefix(section.Name, ".stack") {
 					// TinyGo emits stack sections on microcontroller using the
 					// ".stack" name.
 					// This is a bit ugly, but I don't think there is a way to
@@ -501,8 +501,9 @@ func loadProgramSize(path string, packagePathMap map[string]string) (*programSiz
 						Align:   section.Addralign,
 						Type:    memoryStack,
 					})
-				} else {
-					// Regular .bss section.
+				} else if section.Flags&elf.SHF_WRITE != 0 {
+					// Regular .bss section. Zero-initialized RAM is always
+					// writable, so require SHF_WRITE here.
 					sections = append(sections, memorySection{
 						Address: section.Addr,
 						Size:    section.Size,
@@ -510,6 +511,11 @@ func loadProgramSize(path string, packagePathMap map[string]string) (*programSiz
 						Type:    memoryBSS,
 					})
 				}
+				// Other (non-writable) SHT_NOBITS sections are address-space
+				// placeholders that occupy no RAM, such as the ESP linker
+				// script's .irom_dummy / .rodata_dummy sections which reserve
+				// the flash-mapped XIP virtual address ranges. They must not be
+				// counted as bss/RAM usage.
 			} else if section.Type == elf.SHT_PROGBITS && section.Flags&elf.SHF_EXECINSTR != 0 {
 				// .text
 				sections = append(sections, memorySection{
@@ -954,7 +960,11 @@ func findPackagePath(path string, packagePathMap map[string]string) (packagePath
 			libPath := strings.TrimPrefix(path, filepath.Join(goenv.Get("TINYGOROOT"), "lib")+string(os.PathSeparator))
 			parts := strings.SplitN(libPath, string(os.PathSeparator), 2)
 			packagePath = "C " + parts[0]
-			filename = parts[1]
+			if len(parts) > 1 {
+				filename = parts[1]
+			} else {
+				filename = parts[0]
+			}
 		} else if prefix := filepath.Join(goenv.Get("TINYGOROOT"), "llvm-project", "compiler-rt"); strings.HasPrefix(path, prefix) {
 			packagePath = "C compiler-rt"
 			filename = strings.TrimPrefix(path, prefix+string(os.PathSeparator))
@@ -1,8 +1,13 @@
 package builder

 import (
+	"flag"
+	"fmt"
+	"os"
 	"regexp"
 	"runtime"
+	"strconv"
+	"strings"
 	"testing"
 	"time"

@@ -11,21 +16,19 @@ import (

 var sema = make(chan struct{}, runtime.NumCPU())

+var flagUpdate = flag.Bool("update", false, "update builder package tests")
+
 type sizeTest struct {
-	target     string
-	path       string
-	codeSize   uint64
-	rodataSize uint64
-	dataSize   uint64
-	bssSize    uint64
+	target string
+	path   string
 }

 // Test whether code and data size is as expected for the given targets.
 // This tests both the logic of loadProgramSize and checks that code size
 // doesn't change unintentionally.
 //
-// If you find that code or data size is reduced, then great! You can reduce the
-// number in this test.
+// If you find that code or data size is reduced, then great! You can update the
+// golden file by passing -update to the test.
 // If you find that the code or data size is increased, take a look as to why
 // this is. It could be due to an update (LLVM version, Go version, etc) which
 // is fine, but it could also mean that a recent change introduced this size
@@ -42,35 +45,110 @@ func TestBinarySize(t *testing.T) {
 	// This is a small number of very diverse targets that we want to test.
 	tests := []sizeTest{
 		// microcontrollers
-		{"hifive1b", "examples/echo", 4560, 280, 0, 2268},
-		{"microbit", "examples/serial", 2916, 388, 8, 2272},
-		{"wioterminal", "examples/pininterrupt", 7359, 1489, 116, 6912},
+		{"hifive1b", "examples/echo"},
+		{"microbit", "examples/serial"},
+		{"wioterminal", "examples/pininterrupt"},

 		// TODO: also check wasm. Right now this is difficult, because
 		// wasm binaries are run through wasm-opt and therefore the
 		// output varies by binaryen version.
 	}
-	for _, tc := range tests {
-		tc := tc
-		t.Run(tc.target+"/"+tc.path, func(t *testing.T) {
-			t.Parallel()
+	sizes := measureBinarySizes(t, tests)
+	output := formatSizeTable(tests, sizes)
+	checkGolden(t, "testdata/binary-size.txt", output)
+}

-			// Build the binary.
-			result := buildBinary(t, tc.target, tc.path)
-
-			// Check whether the size of the binary matches the expected size.
-			sizes, err := loadProgramSize(result.Executable, nil)
-			if err != nil {
-				t.Fatal("could not read program size:", err)
-			}
-			if sizes.Code != tc.codeSize || sizes.ROData != tc.rodataSize || sizes.Data != tc.dataSize || sizes.BSS != tc.bssSize {
-				t.Errorf("Unexpected code size when compiling: -target=%s %s", tc.target, tc.path)
-				t.Errorf("            code rodata   data    bss")
-				t.Errorf("expected: %6d %6d %6d %6d", tc.codeSize, tc.rodataSize, tc.dataSize, tc.bssSize)
-				t.Errorf("actual:   %6d %6d %6d %6d", sizes.Code, sizes.ROData, sizes.Data, sizes.BSS)
-			}
-		})
+func checkGolden(t *testing.T, path, actual string) {
+	t.Helper()
+	if *flagUpdate {
+		if err := os.WriteFile(path, []byte(actual), 0o666); err != nil {
+			t.Fatal("failed to write updated golden file:", err)
+		}
+		return
 	}
+	expected, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal("failed to read golden file:", err)
+	}
+	if actual != string(expected) {
+		t.Errorf("%s does not match expected output (re-run with -update to regenerate):\nexpected:\n%sactual:\n%s", path, expected, actual)
+	}
+}
+
+func measureBinarySizes(t *testing.T, tests []sizeTest) []*programSize {
+	t.Helper()
+	type result struct {
+		index int
+		size  *programSize
+		err   error
+	}
+
+	results := make(chan result, len(tests))
+	for i, tc := range tests {
+		tmpdir := t.TempDir()
+		go func() {
+			size, err := measureBinarySize(tc, tmpdir)
+			results <- result{i, size, err}
+		}()
+	}
+
+	sizes := make([]*programSize, len(tests))
+	failed := false
+	for range tests {
+		result := <-results
+		if result.err != nil {
+			tc := tests[result.index]
+			t.Errorf("%s/%s: %v", tc.target, tc.path, result.err)
+			failed = true
+		}
+		sizes[result.index] = result.size
+	}
+	if failed {
+		t.FailNow()
+	}
+	return sizes
+}
+
+func measureBinarySize(tc sizeTest, tmpdir string) (*programSize, error) {
+	result, err := buildBinaryInDir(tc.target, tc.path, tmpdir)
+	if err != nil {
+		return nil, err
+	}
+	size, err := loadProgramSize(result.Executable, nil)
+	if err != nil {
+		return nil, fmt.Errorf("could not read program size: %w", err)
+	}
+	return size, nil
+}
+
+func formatSizeTable(tests []sizeTest, sizes []*programSize) string {
+	targetWidth := len("target")
+	packageWidth := len("package")
+	codeWidth := len("code")
+	rodataWidth := len("rodata")
+	dataWidth := len("data")
+	bssWidth := len("bss")
+	for i, tc := range tests {
+		targetWidth = max(targetWidth, len(tc.target))
+		packageWidth = max(packageWidth, len(tc.path))
+		codeWidth = max(codeWidth, len(strconv.FormatUint(sizes[i].Code, 10)))
+		rodataWidth = max(rodataWidth, len(strconv.FormatUint(sizes[i].ROData, 10)))
+		dataWidth = max(dataWidth, len(strconv.FormatUint(sizes[i].Data, 10)))
+		bssWidth = max(bssWidth, len(strconv.FormatUint(sizes[i].BSS, 10)))
+	}
+
+	var output strings.Builder
+	fmt.Fprintf(&output, "%-*s %-*s %*s %*s %*s %*s\n",
+		targetWidth, "target", packageWidth, "package",
+		codeWidth, "code", rodataWidth, "rodata", dataWidth, "data", bssWidth, "bss")
+	for i, tc := range tests {
+		size := sizes[i]
+		fmt.Fprintf(&output, "%-*s %-*s %*d %*d %*d %*d\n",
+			targetWidth, tc.target, packageWidth, tc.path,
+			codeWidth, size.Code, rodataWidth, size.ROData,
+			dataWidth, size.Data, bssWidth, size.BSS)
+	}
+	return output.String()
 }

 // Check that the -size=full flag attributes binary size to the correct package
@@ -85,7 +163,6 @@ func TestSizeFull(t *testing.T) {
 	pkgMatch := regexp.MustCompile(`^[a-z/]+$`)    // example: "internal/task"

 	for _, target := range tests {
-		target := target
 		t.Run(target, func(t *testing.T) {
 			t.Parallel()

@@ -99,7 +176,7 @@ func TestSizeFull(t *testing.T) {
 				t.Fatal("could not read program size:", err)
 			}
 			for _, pkg := range sizes.sortedPackageNames() {
-				if pkg == "(padding)" || pkg == "(unknown)" {
+				if pkg == "(padding)" || pkg == "(unknown)" || pkg == "Go types" {
 					// TODO: correctly attribute all unknown binary size.
 					continue
 				}
@@ -116,6 +193,15 @@ func TestSizeFull(t *testing.T) {
 }

 func buildBinary(t *testing.T, targetString, pkgName string) BuildResult {
+	t.Helper()
+	result, err := buildBinaryInDir(targetString, pkgName, t.TempDir())
+	if err != nil {
+		t.Fatal(err)
+	}
+	return result
+}
+
+func buildBinaryInDir(targetString, pkgName, tmpdir string) (BuildResult, error) {
 	options := compileopts.Options{
 		Target:        targetString,
 		Opt:           "z",
@@ -126,15 +212,15 @@ func buildBinary(t *testing.T, targetString, pkgName string) BuildResult {
 	}
 	target, err := compileopts.LoadTarget(&options)
 	if err != nil {
-		t.Fatal("could not load target:", err)
+		return BuildResult{}, fmt.Errorf("could not load target: %w", err)
 	}
 	config := &compileopts.Config{
 		Options: &options,
 		Target:  target,
 	}
-	result, err := Build(pkgName, "", t.TempDir(), config)
+	result, err := Build(pkgName, "", tmpdir, config)
 	if err != nil {
-		t.Fatal("could not build:", err)
+		return BuildResult{}, fmt.Errorf("could not build: %w", err)
 	}
-	return result
+	return result, nil
 }
@@ -0,0 +1,4 @@
+target      package               code rodata data  bss
+hifive1b    examples/echo         4321    323    0 2268
+microbit    examples/serial       2842    382    8 2264
+wioterminal examples/pininterrupt 8039   1665  132 7496
@@ -28,7 +28,7 @@ func RunTool(tool string, args ...string) error {
 	var cflag *C.char
 	buf := C.calloc(C.size_t(len(args)), C.size_t(unsafe.Sizeof(cflag)))
 	defer C.free(buf)
-	cflags := (*[1 << 10]*C.char)(unsafe.Pointer(buf))[:len(args):len(args)]
+	cflags := unsafe.Slice((**C.char)(buf), len(args))
 	for i, flag := range args {
 		cflag := C.CString(flag)
 		cflags[i] = cflag
@@ -116,18 +116,24 @@ func parseLLDErrors(text string) error {
 		// This can happen in some cases like with CGo and //go:linkname tricker.
 		if matches := regexp.MustCompile(`^ld.lld(-[0-9]+)?: error: undefined symbol: (.*)\n`).FindStringSubmatch(message); matches != nil {
 			symbolName := matches[2]
-			for _, line := range strings.Split(message, "\n") {
+			for line := range strings.SplitSeq(message, "\n") {
 				matches := regexp.MustCompile(`referenced by .* \(((.*):([0-9]+))\)`).FindStringSubmatch(line)
 				if matches != nil {
 					parsedError = true
 					line, _ := strconv.Atoi(matches[3])
-					// TODO: detect common mistakes like -gc=none?
+					msg := "linker could not find symbol " + symbolName
+					switch symbolName {
+					case "runtime.alloc":
+						msg = "object allocated on the heap with -gc=none"
+					case "runtime.alloc_noheap":
+						msg = "object allocated on the heap in //go:noheap function"
+					}
 					linkErrors = append(linkErrors, scanner.Error{
 						Pos: token.Position{
 							Filename: matches[2],
 							Line:     line,
 						},
-						Msg: "linker could not find symbol " + symbolName,
+						Msg: msg,
 					})
 				}
 			}
@@ -40,7 +40,7 @@ func convertBinToUF2(input []byte, targetAddr uint32, uf2FamilyID string) ([]byt
 	}
 	bl.SetNumBlocks(len(blocks))

-	for i := 0; i < len(blocks); i++ {
+	for i := range blocks {
 		bl.SetBlockNo(i)
 		bl.SetData(blocks[i])

@@ -0,0 +1,284 @@
+package builder
+
+import (
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+
+	"github.com/tinygo-org/tinygo/goenv"
+)
+
+var libWasiLibc = Library{
+	name: "wasi-libc",
+	makeHeaders: func(target, includeDir string) error {
+		bits := filepath.Join(includeDir, "bits")
+		err := os.Mkdir(bits, 0777)
+		if err != nil {
+			return err
+		}
+
+		muslDir := filepath.Join(goenv.Get("TINYGOROOT"), "lib", "wasi-libc/libc-top-half/musl")
+		err = buildMuslAllTypes("wasm32", muslDir, bits)
+		if err != nil {
+			return err
+		}
+
+		// See MUSL_OMIT_HEADERS in the Makefile.
+		omitHeaders := map[string]struct{}{
+			"syslog.h":   {},
+			"wait.h":     {},
+			"ucontext.h": {},
+			"paths.h":    {},
+			"utmp.h":     {},
+			"utmpx.h":    {},
+			"lastlog.h":  {},
+			"elf.h":      {},
+			"link.h":     {},
+			"pwd.h":      {},
+			"shadow.h":   {},
+			"grp.h":      {},
+			"mntent.h":   {},
+			"netdb.h":    {},
+			"resolv.h":   {},
+			"pty.h":      {},
+			"dlfcn.h":    {},
+			"setjmp.h":   {},
+			"ulimit.h":   {},
+			"wordexp.h":  {},
+			"spawn.h":    {},
+			"termios.h":  {},
+			"libintl.h":  {},
+			"aio.h":      {},
+
+			"stdarg.h": {},
+			"stddef.h": {},
+
+			"pthread.h": {},
+		}
+
+		for _, glob := range [][2]string{
+			{"libc-bottom-half/headers/public/*.h", ""},
+			{"libc-bottom-half/headers/public/wasi/*.h", "wasi"},
+			{"libc-top-half/musl/arch/wasm32/bits/*.h", "bits"},
+			{"libc-top-half/musl/include/*.h", ""},
+			{"libc-top-half/musl/include/netinet/*.h", "netinet"},
+			{"libc-top-half/musl/include/sys/*.h", "sys"},
+		} {
+			matches, _ := filepath.Glob(filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc", glob[0]))
+			outDir := filepath.Join(includeDir, glob[1])
+			os.MkdirAll(outDir, 0o777)
+			for _, match := range matches {
+				name := filepath.Base(match)
+				if _, ok := omitHeaders[name]; ok {
+					continue
+				}
+				data, err := os.ReadFile(match)
+				if err != nil {
+					return err
+				}
+				err = os.WriteFile(filepath.Join(outDir, name), data, 0o666)
+				if err != nil {
+					return err
+				}
+			}
+		}
+
+		return nil
+	},
+	cflags: func(target, headerPath string) []string {
+		libcDir := filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc")
+		return []string{
+			"-Werror",
+			"-Wall",
+			"-std=gnu11",
+			"-nostdlibinc",
+			"-mnontrapping-fptoint", "-msign-ext", "-mbulk-memory",
+			"-Wno-null-pointer-arithmetic", "-Wno-unused-parameter", "-Wno-sign-compare", "-Wno-unused-variable", "-Wno-unused-function", "-Wno-ignored-attributes", "-Wno-missing-braces", "-Wno-ignored-pragmas", "-Wno-unused-but-set-variable", "-Wno-unknown-warning-option",
+			"-Wno-parentheses", "-Wno-shift-op-parentheses", "-Wno-bitwise-op-parentheses", "-Wno-logical-op-parentheses", "-Wno-string-plus-int", "-Wno-dangling-else", "-Wno-unknown-pragmas",
+			"-DNDEBUG",
+			"-D__wasilibc_printscan_no_long_double",
+			"-D__wasilibc_printscan_full_support_option=\"long double support is disabled\"",
+			"-DBULK_MEMORY_THRESHOLD=32", // default threshold in wasi-libc
+			"-isystem", headerPath,
+			"-I" + libcDir + "/libc-top-half/musl/src/include",
+			"-I" + libcDir + "/libc-top-half/musl/src/internal",
+			"-I" + libcDir + "/libc-top-half/musl/arch/wasm32",
+			"-I" + libcDir + "/libc-top-half/musl/arch/generic",
+			"-I" + libcDir + "/libc-top-half/headers/private",
+		}
+	},
+	cflagsForFile: func(path string) []string {
+		if strings.HasPrefix(path, "libc-bottom-half"+string(os.PathSeparator)) {
+			libcDir := filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc")
+			return []string{
+				"-I" + libcDir + "/libc-bottom-half/headers/private",
+				"-I" + libcDir + "/libc-bottom-half/cloudlibc/src/include",
+				"-I" + libcDir + "/libc-bottom-half/cloudlibc/src",
+			}
+		}
+		return nil
+	},
+	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc") },
+	librarySources: func(target string, libcNeedsMalloc bool) ([]string, error) {
+		type filePattern struct {
+			glob    string
+			exclude []string
+		}
+
+		// See: LIBC_TOP_HALF_MUSL_SOURCES in the Makefile
+		globs := []filePattern{
+			// Top half: mostly musl sources.
+			{glob: "libc-top-half/sources/*.c"},
+			{glob: "libc-top-half/musl/src/conf/*.c"},
+			{glob: "libc-top-half/musl/src/internal/*.c", exclude: []string{
+				"procfdname.c", "syscall.c", "syscall_ret.c", "vdso.c", "version.c",
+			}},
+			{glob: "libc-top-half/musl/src/locale/*.c", exclude: []string{
+				"dcngettext.c", "textdomain.c", "bind_textdomain_codeset.c"}},
+			{glob: "libc-top-half/musl/src/math/*.c", exclude: []string{
+				"__signbit.c", "__signbitf.c", "__signbitl.c",
+				"__fpclassify.c", "__fpclassifyf.c", "__fpclassifyl.c",
+				"ceilf.c", "ceil.c",
+				"floorf.c", "floor.c",
+				"truncf.c", "trunc.c",
+				"rintf.c", "rint.c",
+				"nearbyintf.c", "nearbyint.c",
+				"sqrtf.c", "sqrt.c",
+				"fabsf.c", "fabs.c",
+				"copysignf.c", "copysign.c",
+				"fminf.c", "fmaxf.c",
+				"fmin.c", "fmax.c,",
+			}},
+			{glob: "libc-top-half/musl/src/multibyte/*.c"},
+			{glob: "libc-top-half/musl/src/stdio/*.c", exclude: []string{
+				"vfwscanf.c", "vfwprintf.c", // long double is unsupported
+				"__lockfile.c", "flockfile.c", "funlockfile.c", "ftrylockfile.c",
+				"rename.c",
+				"tmpnam.c", "tmpfile.c", "tempnam.c",
+				"popen.c", "pclose.c",
+				"remove.c",
+				"gets.c"}},
+			{glob: "libc-top-half/musl/src/stdlib/*.c"},
+			{glob: "libc-top-half/musl/src/string/*.c", exclude: []string{
+				"strsignal.c"}},
+
+			// Bottom half: connect top half to WASI equivalents.
+			{glob: "libc-bottom-half/cloudlibc/src/libc/*/*.c"},
+			{glob: "libc-bottom-half/cloudlibc/src/libc/sys/*/*.c"},
+			{glob: "libc-bottom-half/sources/*.c"},
+		}
+
+		// We're using the Boehm GC, so we need a heap implementation in the libc.
+		if libcNeedsMalloc {
+			globs = append(globs, filePattern{glob: "dlmalloc/src/dlmalloc.c"})
+		}
+
+		// See: LIBC_TOP_HALF_MUSL_SOURCES in the Makefile
+		sources := []string{
+			"libc-top-half/musl/src/misc/a64l.c",
+			"libc-top-half/musl/src/misc/basename.c",
+			"libc-top-half/musl/src/misc/dirname.c",
+			"libc-top-half/musl/src/misc/ffs.c",
+			"libc-top-half/musl/src/misc/ffsl.c",
+			"libc-top-half/musl/src/misc/ffsll.c",
+			"libc-top-half/musl/src/misc/fmtmsg.c",
+			"libc-top-half/musl/src/misc/getdomainname.c",
+			"libc-top-half/musl/src/misc/gethostid.c",
+			"libc-top-half/musl/src/misc/getopt.c",
+			"libc-top-half/musl/src/misc/getopt_long.c",
+			"libc-top-half/musl/src/misc/getsubopt.c",
+			"libc-top-half/musl/src/misc/uname.c",
+			"libc-top-half/musl/src/misc/nftw.c",
+			"libc-top-half/musl/src/errno/strerror.c",
+			"libc-top-half/musl/src/network/htonl.c",
+			"libc-top-half/musl/src/network/htons.c",
+			"libc-top-half/musl/src/network/ntohl.c",
+			"libc-top-half/musl/src/network/ntohs.c",
+			"libc-top-half/musl/src/network/inet_ntop.c",
+			"libc-top-half/musl/src/network/inet_pton.c",
+			"libc-top-half/musl/src/network/inet_aton.c",
+			"libc-top-half/musl/src/network/in6addr_any.c",
+			"libc-top-half/musl/src/network/in6addr_loopback.c",
+			"libc-top-half/musl/src/fenv/fenv.c",
+			"libc-top-half/musl/src/fenv/fesetround.c",
+			"libc-top-half/musl/src/fenv/feupdateenv.c",
+			"libc-top-half/musl/src/fenv/fesetexceptflag.c",
+			"libc-top-half/musl/src/fenv/fegetexceptflag.c",
+			"libc-top-half/musl/src/fenv/feholdexcept.c",
+			"libc-top-half/musl/src/exit/exit.c",
+			"libc-top-half/musl/src/exit/atexit.c",
+			"libc-top-half/musl/src/exit/assert.c",
+			"libc-top-half/musl/src/exit/quick_exit.c",
+			"libc-top-half/musl/src/exit/at_quick_exit.c",
+			"libc-top-half/musl/src/time/strftime.c",
+			"libc-top-half/musl/src/time/asctime.c",
+			"libc-top-half/musl/src/time/asctime_r.c",
+			"libc-top-half/musl/src/time/ctime.c",
+			"libc-top-half/musl/src/time/ctime_r.c",
+			"libc-top-half/musl/src/time/wcsftime.c",
+			"libc-top-half/musl/src/time/strptime.c",
+			"libc-top-half/musl/src/time/difftime.c",
+			"libc-top-half/musl/src/time/timegm.c",
+			"libc-top-half/musl/src/time/ftime.c",
+			"libc-top-half/musl/src/time/gmtime.c",
+			"libc-top-half/musl/src/time/gmtime_r.c",
+			"libc-top-half/musl/src/time/timespec_get.c",
+			"libc-top-half/musl/src/time/getdate.c",
+			"libc-top-half/musl/src/time/localtime.c",
+			"libc-top-half/musl/src/time/localtime_r.c",
+			"libc-top-half/musl/src/time/mktime.c",
+			"libc-top-half/musl/src/time/__tm_to_secs.c",
+			"libc-top-half/musl/src/time/__month_to_secs.c",
+			"libc-top-half/musl/src/time/__secs_to_tm.c",
+			"libc-top-half/musl/src/time/__year_to_secs.c",
+			"libc-top-half/musl/src/time/__tz.c",
+			"libc-top-half/musl/src/fcntl/creat.c",
+			"libc-top-half/musl/src/dirent/alphasort.c",
+			"libc-top-half/musl/src/dirent/versionsort.c",
+			"libc-top-half/musl/src/env/__stack_chk_fail.c",
+			"libc-top-half/musl/src/env/clearenv.c",
+			"libc-top-half/musl/src/env/getenv.c",
+			"libc-top-half/musl/src/env/putenv.c",
+			"libc-top-half/musl/src/env/setenv.c",
+			"libc-top-half/musl/src/env/unsetenv.c",
+			"libc-top-half/musl/src/unistd/posix_close.c",
+			"libc-top-half/musl/src/stat/futimesat.c",
+			"libc-top-half/musl/src/legacy/getpagesize.c",
+			"libc-top-half/musl/src/thread/thrd_sleep.c",
+		}
+
+		basepath := goenv.Get("TINYGOROOT") + "/lib/wasi-libc/"
+		for _, pattern := range globs {
+			matches, err := filepath.Glob(basepath + pattern.glob)
+			if err != nil {
+				// From the documentation:
+				// > Glob ignores file system errors such as I/O errors reading
+				// > directories. The only possible returned error is
+				// > ErrBadPattern, when pattern is malformed.
+				// So the only possible error is when the (statically defined)
+				// pattern is wrong. In other words, a programming bug.
+				return nil, fmt.Errorf("wasi-libc: could not glob source dirs: %w", err)
+			}
+			if len(matches) == 0 {
+				return nil, fmt.Errorf("wasi-libc: did not find any files for pattern %#v", pattern)
+			}
+			excludeSet := map[string]struct{}{}
+			for _, exclude := range pattern.exclude {
+				excludeSet[exclude] = struct{}{}
+			}
+			for _, match := range matches {
+				if _, ok := excludeSet[filepath.Base(match)]; ok {
+					continue
+				}
+				relpath, err := filepath.Rel(basepath, match)
+				if err != nil {
+					// Not sure if this is even possible.
+					return nil, err
+				}
+				sources = append(sources, relpath)
+			}
+		}
+		return sources, nil
+	},
+}
@@ -29,6 +29,8 @@ var libWasmBuiltins = Library{
 			"-Wall",
 			"-std=gnu11",
 			"-nostdlibinc",
+			"-mnontrapping-fptoint", // match wasm-unknown (default on in LLVM 20)
+			"-mno-bulk-memory",      // same here
 			"-isystem", libcDir + "/libc-top-half/musl/arch/wasm32",
 			"-isystem", libcDir + "/libc-top-half/musl/arch/generic",
 			"-isystem", libcDir + "/libc-top-half/musl/src/internal",
@@ -39,7 +41,7 @@ var libWasmBuiltins = Library{
 		}
 	},
 	sourceDir: func() string { return filepath.Join(goenv.Get("TINYGOROOT"), "lib/wasi-libc") },
-	librarySources: func(target string) ([]string, error) {
+	librarySources: func(target string, _ bool) ([]string, error) {
 		return []string{
 			// memory builtins needed for llvm.memcpy.*, llvm.memmove.*, and
 			// llvm.memset.* LLVM intrinsics.
@@ -26,10 +26,6 @@ import (
 	"golang.org/x/tools/go/ast/astutil"
 )

-// Function that's only defined in Go 1.22.
-var setASTFileFields = func(f *ast.File, start, end token.Pos) {
-}
-
 // cgoPackage holds all CGo-related information of a package.
 type cgoPackage struct {
 	generated       *ast.File
@@ -44,7 +40,7 @@ type cgoPackage struct {
 	tokenFiles      map[string]*token.File
 	definedGlobally map[string]ast.Node
 	noescapingFuncs map[string]*noescapingFunc // #cgo noescape lines
-	anonDecls       map[interface{}]string
+	anonDecls       map[any]string
 	cflags          []string // CFlags from #cgo lines
 	ldflags         []string // LDFlags from #cgo lines
 	visitedFiles    map[string][]byte
@@ -263,7 +259,7 @@ func Process(files []*ast.File, dir, importPath string, fset *token.FileSet, cfl
 		tokenFiles:      map[string]*token.File{},
 		definedGlobally: map[string]ast.Node{},
 		noescapingFuncs: map[string]*noescapingFunc{},
-		anonDecls:       map[interface{}]string{},
+		anonDecls:       map[any]string{},
 		visitedFiles:    map[string][]byte{},
 	}

@@ -306,7 +302,7 @@ func Process(files []*ast.File, dir, importPath string, fset *token.FileSet, cfl
 	// Find `import "C"` C fragments in the file.
 	p.cgoHeaders = make([]string, len(files)) // combined CGo header fragment for each file
 	for i, f := range files {
-		var cgoHeader string
+		var cgoHeader strings.Builder
 		for i := 0; i < len(f.Decls); i++ {
 			decl := f.Decls[i]
 			genDecl, ok := decl.(*ast.GenDecl)
@@ -341,7 +337,8 @@ func Process(files []*ast.File, dir, importPath string, fset *token.FileSet, cfl
 			// Iterate through all parts of the CGo header. Note that every //
 			// line is a new comment.
 			position := fset.Position(genDecl.Doc.Pos())
-			fragment := fmt.Sprintf("# %d %#v\n", position.Line, position.Filename)
+			var fragment strings.Builder
+			fragment.WriteString(fmt.Sprintf("# %d %#v\n", position.Line, position.Filename))
 			for _, comment := range genDecl.Doc.List {
 				// Find all #cgo lines, extract and use their contents, and
 				// replace the lines with spaces (to preserve locations).
@@ -358,12 +355,13 @@ func Process(files []*ast.File, dir, importPath string, fset *token.FileSet, cfl
 				} else { // comment
 					c = "  " + c[2:len(c)-2]
 				}
-				fragment += c + "\n"
+				fragment.WriteString(c)
+				fragment.WriteByte('\n')
 			}
-			cgoHeader += fragment
+			cgoHeader.WriteString(fragment.String())
 		}

-		p.cgoHeaders[i] = cgoHeader
+		p.cgoHeaders[i] = cgoHeader.String()
 	}

 	// Define CFlags that will be used while parsing the package.
@@ -654,7 +652,6 @@ func (p *cgoPackage) createUnionAccessor(field *ast.Field, typeName string) {
 		X: &ast.Ident{
 			NamePos: pos,
 			Name:    "union",
-			Obj:     nil,
 		},
 		Sel: &ast.Ident{
 			NamePos: pos,
@@ -708,7 +705,6 @@ func (p *cgoPackage) createUnionAccessor(field *ast.Field, typeName string) {
 						X: &ast.Ident{
 							NamePos: pos,
 							Name:    typeName,
-							Obj:     nil,
 						},
 					},
 				},
@@ -764,7 +760,6 @@ func (p *cgoPackage) createBitfieldGetter(bitfield bitfieldInfo, typeName string
 		X: &ast.Ident{
 			NamePos: bitfield.pos,
 			Name:    "s",
-			Obj:     nil,
 		},
 		Sel: &ast.Ident{
 			NamePos: bitfield.pos,
@@ -811,11 +806,6 @@ func (p *cgoPackage) createBitfieldGetter(bitfield bitfieldInfo, typeName string
 						{
 							NamePos: bitfield.pos,
 							Name:    "s",
-							Obj: &ast.Object{
-								Kind: ast.Var,
-								Name: "s",
-								Decl: nil,
-							},
 						},
 					},
 					Type: &ast.StarExpr{
@@ -823,7 +813,6 @@ func (p *cgoPackage) createBitfieldGetter(bitfield bitfieldInfo, typeName string
 						X: &ast.Ident{
 							NamePos: bitfield.pos,
 							Name:    typeName,
-							Obj:     nil,
 						},
 					},
 				},
@@ -881,7 +870,6 @@ func (p *cgoPackage) createBitfieldSetter(bitfield bitfieldInfo, typeName string
 		X: &ast.Ident{
 			NamePos: bitfield.pos,
 			Name:    "s",
-			Obj:     nil,
 		},
 		Sel: &ast.Ident{
 			NamePos: bitfield.pos,
@@ -964,11 +952,6 @@ func (p *cgoPackage) createBitfieldSetter(bitfield bitfieldInfo, typeName string
 						{
 							NamePos: bitfield.pos,
 							Name:    "s",
-							Obj: &ast.Object{
-								Kind: ast.Var,
-								Name: "s",
-								Decl: nil,
-							},
 						},
 					},
 					Type: &ast.StarExpr{
@@ -976,7 +959,6 @@ func (p *cgoPackage) createBitfieldSetter(bitfield bitfieldInfo, typeName string
 						X: &ast.Ident{
 							NamePos: bitfield.pos,
 							Name:    typeName,
-							Obj:     nil,
 						},
 					},
 				},
@@ -997,7 +979,6 @@ func (p *cgoPackage) createBitfieldSetter(bitfield bitfieldInfo, typeName string
 							{
 								NamePos: bitfield.pos,
 								Name:    "value",
-								Obj:     nil,
 							},
 						},
 						Type: bitfield.field.Type,
@@ -1015,7 +996,6 @@ func (p *cgoPackage) createBitfieldSetter(bitfield bitfieldInfo, typeName string
 							X: &ast.Ident{
 								NamePos: bitfield.pos,
 								Name:    "s",
-								Obj:     nil,
 							},
 							Sel: &ast.Ident{
 								NamePos: bitfield.pos,
@@ -1239,7 +1219,7 @@ func getPos(node ast.Node) token.Pos {
 // getUnnamedDeclName creates a name (with the given prefix) for the given C
 // declaration. This is used for structs, unions, and enums that are often
 // defined without a name and used in a typedef.
-func (p *cgoPackage) getUnnamedDeclName(prefix string, itf interface{}) string {
+func (p *cgoPackage) getUnnamedDeclName(prefix string, itf any) string {
 	if name, ok := p.anonDecls[itf]; ok {
 		return name
 	}
@@ -1,17 +0,0 @@
-//go:build go1.22
-
-package cgo
-
-// Code specifically for Go 1.22.
-
-import (
-	"go/ast"
-	"go/token"
-)
-
-func init() {
-	setASTFileFields = func(f *ast.File, start, end token.Pos) {
-		f.FileStart = start
-		f.FileEnd = end
-	}
-}
@@ -45,7 +45,6 @@ func TestCGo(t *testing.T) {
 		"flags",
 		"const",
 	} {
-		name := name // avoid a race condition
 		t.Run(name, func(t *testing.T) {
 			// Read the AST in memory.
 			path := filepath.Join("testdata", name+".go")
@@ -71,7 +71,7 @@ unsigned tinygo_clang_Cursor_isMacroFunctionLike(GoCXCursor c);
 //     libclang.go:68:5: note: previous declaration is here
 // See: https://github.com/golang/go/issues/49721
 #if defined(_WIN32)
-#define CGO_DECL // __declspec(dllexport)
+#define CGO_DECL __declspec(dllexport)
 #else
 #define CGO_DECL
 #endif
@@ -130,7 +130,7 @@ func (f *cgoFile) readNames(fragment string, cflags []string, filename string, c
 	// convert Go slice of strings to C array of strings.
 	cmdargsC := C.malloc(C.size_t(len(cflags)) * C.size_t(unsafe.Sizeof(uintptr(0))))
 	defer C.free(cmdargsC)
-	cmdargs := (*[1 << 16]*C.char)(cmdargsC)
+	cmdargs := unsafe.Slice((**C.char)(cmdargsC), len(cflags))
 	for i, cflag := range cflags {
 		s := C.CString(cflag)
 		cmdargs[i] = s
@@ -160,7 +160,7 @@ func (f *cgoFile) readNames(fragment string, cflags []string, filename string, c
 			pos := f.getClangLocationPosition(location, unit)
 			f.addError(pos, severity+": "+spelling)
 		}
-		for i := 0; i < numDiagnostics; i++ {
+		for i := range numDiagnostics {
 			diagnostic := C.clang_getDiagnostic(unit, C.uint(i))
 			addDiagnostic(diagnostic)

@@ -190,7 +190,7 @@ func (f *cgoFile) readNames(fragment string, cflags []string, filename string, c
 			// Sanity check. This should (hopefully) never trigger.
 			panic("libclang: file contents was not loaded")
 		}
-		data := (*[1 << 24]byte)(unsafe.Pointer(rawData))[:size]
+		data := unsafe.Slice((*byte)(unsafe.Pointer(rawData)), size)

 		// Hash the contents if it isn't hashed yet.
 		if _, ok := f.visitedFiles[path]; !ok {
@@ -217,10 +217,6 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 	case C.CXCursor_FunctionDecl:
 		cursorType := C.tinygo_clang_getCursorType(c)
 		numArgs := int(C.tinygo_clang_Cursor_getNumArguments(c))
-		obj := &ast.Object{
-			Kind: ast.Fun,
-			Name: "_Cgo_" + name,
-		}
 		exportName := name
 		localName := name
 		var stringSignature string
@@ -258,7 +254,6 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 			Name: &ast.Ident{
 				NamePos: pos,
 				Name:    "_Cgo_" + localName,
-				Obj:     obj,
 			},
 			Type: &ast.FuncType{
 				Func: pos,
@@ -283,7 +278,7 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 				Text:  strings.Join(doc, "\n"),
 			})
 		}
-		for i := 0; i < numArgs; i++ {
+		for i := range numArgs {
 			arg := C.tinygo_clang_Cursor_getArgument(c, C.uint(i))
 			argName := getString(C.tinygo_clang_getCursorSpelling(arg))
 			argType := C.clang_getArgType(cursorType, C.uint(i))
@@ -295,11 +290,6 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 					{
 						NamePos: pos,
 						Name:    argName,
-						Obj: &ast.Object{
-							Kind: ast.Var,
-							Name: argName,
-							Decl: decl,
-						},
 					},
 				},
 				Type: f.makeDecayingASTType(argType, pos),
@@ -315,7 +305,6 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 				},
 			}
 		}
-		obj.Decl = decl
 		return decl, stringSignature
 	case C.CXCursor_StructDecl, C.CXCursor_UnionDecl:
 		typ := f.makeASTRecordType(c, pos)
@@ -325,39 +314,25 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 			// Convert to a single-field struct type.
 			typeExpr = f.makeUnionField(typ)
 		}
-		obj := &ast.Object{
-			Kind: ast.Typ,
-			Name: typeName,
-		}
 		typeSpec := &ast.TypeSpec{
 			Name: &ast.Ident{
 				NamePos: typ.pos,
 				Name:    typeName,
-				Obj:     obj,
 			},
 			Type: typeExpr,
 		}
-		obj.Decl = typeSpec
 		return typeSpec, typ
 	case C.CXCursor_TypedefDecl:
 		typeName := "_Cgo_" + name
 		underlyingType := C.tinygo_clang_getTypedefDeclUnderlyingType(c)
-		obj := &ast.Object{
-			Kind: ast.Typ,
-			Name: typeName,
-		}
 		typeSpec := &ast.TypeSpec{
 			Name: &ast.Ident{
 				NamePos: pos,
 				Name:    typeName,
-				Obj:     obj,
 			},
-			Type: f.makeASTType(underlyingType, pos),
+			Assign: pos,
+			Type:   f.makeASTType(underlyingType, pos),
 		}
-		if underlyingType.kind != C.CXType_Enum {
-			typeSpec.Assign = pos
-		}
-		obj.Decl = typeSpec
 		return typeSpec, nil
 	case C.CXCursor_VarDecl:
 		cursorType := C.tinygo_clang_getCursorType(c)
@@ -376,19 +351,13 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 				},
 			},
 		}
-		obj := &ast.Object{
-			Kind: ast.Var,
-			Name: "_Cgo_" + name,
-		}
 		valueSpec := &ast.ValueSpec{
 			Names: []*ast.Ident{{
 				NamePos: pos,
 				Name:    "_Cgo_" + name,
-				Obj:     obj,
 			}},
 			Type: typeExpr,
 		}
-		obj.Decl = valueSpec
 		gen.Specs = append(gen.Specs, valueSpec)
 		return gen, nil
 	case C.CXCursor_MacroDefinition:
@@ -405,26 +374,16 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 			Lparen: token.NoPos,
 			Rparen: token.NoPos,
 		}
-		obj := &ast.Object{
-			Kind: ast.Con,
-			Name: "_Cgo_" + name,
-		}
 		valueSpec := &ast.ValueSpec{
 			Names: []*ast.Ident{{
 				NamePos: pos,
 				Name:    "_Cgo_" + name,
-				Obj:     obj,
 			}},
 			Values: []ast.Expr{expr},
 		}
-		obj.Decl = valueSpec
 		gen.Specs = append(gen.Specs, valueSpec)
 		return gen, nil
 	case C.CXCursor_EnumDecl:
-		obj := &ast.Object{
-			Kind: ast.Typ,
-			Name: "_Cgo_" + name,
-		}
 		underlying := C.tinygo_clang_getEnumDeclIntegerType(c)
 		// TODO: gc's CGo implementation uses types such as `uint32` for enums
 		// instead of types such as C.int, which are used here.
@@ -432,12 +391,10 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 			Name: &ast.Ident{
 				NamePos: pos,
 				Name:    "_Cgo_" + name,
-				Obj:     obj,
 			},
 			Assign: pos,
 			Type:   f.makeASTType(underlying, pos),
 		}
-		obj.Decl = typeSpec
 		return typeSpec, nil
 	case C.CXCursor_EnumConstantDecl:
 		value := C.tinygo_clang_getEnumConstantDeclValue(c)
@@ -452,19 +409,13 @@ func (f *cgoFile) createASTNode(name string, c clangCursor) (ast.Node, any) {
 			Lparen: token.NoPos,
 			Rparen: token.NoPos,
 		}
-		obj := &ast.Object{
-			Kind: ast.Con,
-			Name: "_Cgo_" + name,
-		}
 		valueSpec := &ast.ValueSpec{
 			Names: []*ast.Ident{{
 				NamePos: pos,
 				Name:    "_Cgo_" + name,
-				Obj:     obj,
 			}},
 			Values: []ast.Expr{expr},
 		}
-		obj.Decl = valueSpec
 		gen.Specs = append(gen.Specs, valueSpec)
 		return gen, nil
 	default:
@@ -581,7 +532,7 @@ func tinygo_clang_globals_visitor(c, parent C.GoCXCursor, client_data C.CXClient

 // Get the precise location in the source code. Used for uniquely identifying
 // source locations.
-func (f *cgoFile) getUniqueLocationID(pos token.Pos, cursor C.GoCXCursor) interface{} {
+func (f *cgoFile) getUniqueLocationID(pos token.Pos, cursor C.GoCXCursor) any {
 	clangLocation := C.tinygo_clang_getCursorLocation(cursor)
 	var file C.CXFile
 	var line C.unsigned
@@ -624,7 +575,7 @@ func (p *cgoPackage) getClangLocationPosition(location C.CXSourceLocation, tu C.
 		// now by reading the file from libclang.
 		var size C.size_t
 		sourcePtr := C.clang_getFileContents(tu, file, &size)
-		source := ((*[1 << 28]byte)(unsafe.Pointer(sourcePtr)))[:size:size]
+		source := unsafe.Slice((*byte)(unsafe.Pointer(sourcePtr)), size)
 		lines := []int{0}
 		for i := 0; i < len(source)-1; i++ {
 			if source[i] == '\n' {
@@ -639,7 +590,8 @@ func (p *cgoPackage) getClangLocationPosition(location C.CXSourceLocation, tu C.
 			Package: f.Pos(0),
 			Name:    ast.NewIdent(p.packageName),
 		}
-		setASTFileFields(astFile, f.Pos(0), f.Pos(int(size)))
+		astFile.FileStart = f.Pos(0)
+		astFile.FileEnd = f.Pos(int(size))
 		p.cgoFiles = append(p.cgoFiles, astFile)
 	}
 	positionFile := p.tokenFiles[filename]
@@ -852,6 +804,14 @@ func (f *cgoFile) makeASTType(typ C.CXType, pos token.Pos) ast.Expr {
 			f.addError(pos, fmt.Sprintf("unknown elaborated type (libclang type kind %s)", typeKindSpelling))
 			typeName = "<unknown>"
 		}
+	case C.CXType_Unexposed:
+		// LLVM 22+ may report certain builtin type aliases (e.g. __size_t)
+		// as Unexposed. Resolve via the canonical type.
+		canonical := C.clang_getCanonicalType(typ)
+		if canonical.kind != C.CXType_Unexposed && canonical.kind != C.CXType_Invalid {
+			return f.makeASTType(canonical, pos)
+		}
+		// If still unexposed, fall through to the error below.
 	case C.CXType_Record:
 		cursor := C.tinygo_clang_getTypeDeclaration(typ)
 		name := getString(C.tinygo_clang_getCursorSpelling(cursor))
@@ -956,22 +916,16 @@ func (p *cgoPackage) getIntegerType(name string, cursor clangCursor) *ast.TypeSp
 	}

 	// Construct an *ast.TypeSpec for this type.
-	obj := &ast.Object{
-		Kind: ast.Typ,
-		Name: name,
-	}
 	spec := &ast.TypeSpec{
 		Name: &ast.Ident{
 			NamePos: pos,
 			Name:    name,
-			Obj:     obj,
 		},
 		Type: &ast.Ident{
 			NamePos: pos,
 			Name:    goName,
 		},
 	}
-	obj.Decl = spec
 	return spec
 }

@@ -1104,7 +1058,6 @@ func tinygo_clang_struct_visitor(c, parent C.GoCXCursor, client_data C.CXClientD
 				pos:      prevField.Names[0].NamePos,
 			})
 			prevField.Names[0].Name = bitfieldName
-			prevField.Names[0].Obj.Name = bitfieldName
 		}
 		prevBitfield := &(*bitfieldList)[len(*bitfieldList)-1]
 		prevBitfield.endBit = bitfieldOffset
@@ -1121,11 +1074,6 @@ func tinygo_clang_struct_visitor(c, parent C.GoCXCursor, client_data C.CXClientD
 		{
 			NamePos: pos,
 			Name:    name,
-			Obj: &ast.Object{
-				Kind: ast.Var,
-				Name: name,
-				Decl: field,
-			},
 		},
 	}
 	fieldList.List = append(fieldList.List, field)
@@ -1,9 +1,9 @@
-//go:build !byollvm && !llvm15 && !llvm16 && !llvm17 && !llvm18
+//go:build !byollvm && llvm19

 package cgo

 /*
-#cgo linux        CFLAGS:  -I/usr/include/llvm-19 -I/usr/include/llvm-c-19 -I/usr/lib/llvm-19/include
+#cgo linux        CFLAGS:  -I/usr/include/llvm-19 -I/usr/include/llvm-c-19 -I/usr/lib/llvm-19/include -I/usr/lib64/llvm19/include
 #cgo darwin,amd64 CFLAGS:  -I/usr/local/opt/llvm@19/include
 #cgo darwin,arm64 CFLAGS:  -I/opt/homebrew/opt/llvm@19/include
 #cgo freebsd      CFLAGS:  -I/usr/local/llvm19/include
@@ -0,0 +1,15 @@
+//go:build !byollvm && !llvm15 && !llvm16 && !llvm17 && !llvm18 && !llvm19 && !llvm21 && !llvm22
+
+package cgo
+
+/*
+#cgo linux        CFLAGS:  -I/usr/include/llvm-20 -I/usr/include/llvm-c-20 -I/usr/lib/llvm-20/include -I/usr/lib64/llvm20/include
+#cgo darwin,amd64 CFLAGS:  -I/usr/local/opt/llvm@20/include
+#cgo darwin,arm64 CFLAGS:  -I/opt/homebrew/opt/llvm@20/include
+#cgo freebsd      CFLAGS:  -I/usr/local/llvm20/include
+#cgo linux        LDFLAGS: -L/usr/lib/llvm-20/lib -lclang
+#cgo darwin,amd64 LDFLAGS: -L/usr/local/opt/llvm@20/lib -lclang
+#cgo darwin,arm64 LDFLAGS: -L/opt/homebrew/opt/llvm@20/lib -lclang
+#cgo freebsd      LDFLAGS: -L/usr/local/llvm20/lib -lclang
+*/
+import "C"
@@ -0,0 +1,15 @@
+//go:build !byollvm && llvm21
+
+package cgo
+
+/*
+#cgo linux        CFLAGS:  -I/usr/include/llvm-21 -I/usr/include/llvm-c-21 -I/usr/lib/llvm-21/include -I/usr/lib64/llvm21/include
+#cgo darwin,amd64 CFLAGS:  -I/usr/local/opt/llvm@21/include
+#cgo darwin,arm64 CFLAGS:  -I/opt/homebrew/opt/llvm@21/include
+#cgo freebsd      CFLAGS:  -I/usr/local/llvm21/include
+#cgo linux        LDFLAGS: -L/usr/lib/llvm-21/lib -lclang
+#cgo darwin,amd64 LDFLAGS: -L/usr/local/opt/llvm@21/lib -lclang
+#cgo darwin,arm64 LDFLAGS: -L/opt/homebrew/opt/llvm@21/lib -lclang
+#cgo freebsd      LDFLAGS: -L/usr/local/llvm21/lib -lclang
+*/
+import "C"
@@ -0,0 +1,15 @@
+//go:build !byollvm && llvm22
+
+package cgo
+
+/*
+#cgo linux        CFLAGS:  -I/usr/include/llvm-22 -I/usr/include/llvm-c-22 -I/usr/lib/llvm-22/include -I/usr/lib64/llvm22/include
+#cgo darwin,amd64 CFLAGS:  -I/usr/local/opt/llvm@22/include
+#cgo darwin,arm64 CFLAGS:  -I/opt/homebrew/opt/llvm@22/include
+#cgo freebsd      CFLAGS:  -I/usr/local/llvm22/include
+#cgo linux        LDFLAGS: -L/usr/lib/llvm-22/lib -lclang
+#cgo darwin,amd64 LDFLAGS: -L/usr/local/opt/llvm@22/lib -lclang
+#cgo darwin,arm64 LDFLAGS: -L/opt/homebrew/opt/llvm@22/lib -lclang
+#cgo freebsd      LDFLAGS: -L/usr/local/llvm22/lib -lclang
+*/
+import "C"
@@ -78,6 +78,7 @@ var validCompilerFlags = []*regexp.Regexp{
 	re(`-f(no-)?(pic|PIC|pie|PIE)`),
 	re(`-f(no-)?plt`),
 	re(`-f(no-)?rtti`),
+	re(`-f(no-)?short-enums`),
 	re(`-f(no-)?split-stack`),
 	re(`-f(no-)?stack-(.+)`),
 	re(`-f(no-)?strict-aliasing`),
@@ -12,17 +12,17 @@ import "C"
 // C. It is useful if an API uses function pointers and you cannot pass a Go
 // pointer but only a C pointer.
 type refMap struct {
-	refs map[unsafe.Pointer]interface{}
+	refs map[unsafe.Pointer]any
 	lock sync.Mutex
 }

 // Put stores a value in the map. It can later be retrieved using Get. It must
 // be removed using Remove to avoid memory leaks.
-func (m *refMap) Put(v interface{}) unsafe.Pointer {
+func (m *refMap) Put(v any) unsafe.Pointer {
 	m.lock.Lock()
 	defer m.lock.Unlock()
 	if m.refs == nil {
-		m.refs = make(map[unsafe.Pointer]interface{}, 1)
+		m.refs = make(map[unsafe.Pointer]any, 1)
 	}
 	ref := C.malloc(1)
 	m.refs[ref] = v
@@ -31,7 +31,7 @@ func (m *refMap) Put(v interface{}) unsafe.Pointer {

 // Get returns a stored value previously inserted with Put. Use the same
 // reference as you got from Put.
-func (m *refMap) Get(ref unsafe.Pointer) interface{} {
+func (m *refMap) Get(ref unsafe.Pointer) any {
 	m.lock.Lock()
 	defer m.lock.Unlock()
 	return m.refs[ref]
@@ -8,12 +8,25 @@ import (
 	"os"
 	"path/filepath"
 	"regexp"
+	"slices"
+	"strconv"
 	"strings"

 	"github.com/google/shlex"
 	"github.com/tinygo-org/tinygo/goenv"
 )

+// Library versions. Whenever an existing library is changed, this number should
+// be added/increased so that existing caches are invalidated.
+//
+// (This is a bit of a layering violation, this should really be part of the
+// builder.Library struct but that's hard to do since we want to know the
+// library path in advance in several places).
+var libVersions = map[string]int{
+	"musl":  3,
+	"bdwgc": 2,
+}
+
 // Config keeps all configuration affecting the build in a single struct.
 type Config struct {
 	Options        *Options
@@ -48,13 +61,15 @@ func (c *Config) BuildMode() string {
 // RISC-V processor, that could be "+a,+c,+m". For many targets, an empty list
 // will be returned.
 func (c *Config) Features() string {
+	var features string
 	if c.Target.Features == "" {
-		return c.Options.LLVMFeatures
+		features = c.Options.LLVMFeatures
+	} else if c.Options.LLVMFeatures == "" {
+		features = c.Target.Features
+	} else {
+		features = c.Target.Features + "," + c.Options.LLVMFeatures
 	}
-	if c.Options.LLVMFeatures == "" {
-		return c.Target.Features
-	}
-	return c.Target.Features + "," + c.Options.LLVMFeatures
+	return patchFeatures(features)
 }

 // ABI returns the -mabi= flag for this target (like -mabi=lp64). A zero-length
@@ -99,6 +114,11 @@ func (c *Config) BuildTags() []string {
 		"math_big_pure_go",                           // to get math/big to work
 		"gc." + c.GC(), "scheduler." + c.Scheduler(), // used inside the runtime package
 		"serial." + c.Serial()}...) // used inside the machine package
+	switch c.Scheduler() {
+	case "threads", "cores":
+	default:
+		tags = append(tags, "tinygo.unicore")
+	}
 	for i := 1; i <= c.GoMinorVersion; i++ {
 		tags = append(tags, fmt.Sprintf("go1.%d", i))
 	}
@@ -122,14 +142,8 @@ func (c *Config) GC() string {
 // that can be traced by the garbage collector.
 func (c *Config) NeedsStackObjects() bool {
 	switch c.GC() {
-	case "conservative", "custom", "precise":
-		for _, tag := range c.BuildTags() {
-			if tag == "tinygo.wasm" {
-				return true
-			}
-		}
-
-		return false
+	case "conservative", "custom", "precise", "boehm":
+		return slices.Contains(c.BuildTags(), "tinygo.wasm")
 	default:
 		return false
 	}
@@ -209,7 +223,7 @@ func (c *Config) StackSize() uint64 {

 // MaxStackAlloc returns the size of the maximum allocation to put on the stack vs heap.
 func (c *Config) MaxStackAlloc() uint64 {
-	if c.StackSize() > 32*1024 {
+	if c.StackSize() >= 16*1024 {
 		return 1024
 	}

@@ -228,7 +242,7 @@ func (c *Config) RP2040BootPatch() bool {
 // Return a canonicalized architecture name, so we don't have to deal with arm*
 // vs thumb* vs arm64.
 func CanonicalArchName(triple string) string {
-	arch := strings.Split(triple, "-")[0]
+	arch, _, _ := strings.Cut(triple, "-")
 	if arch == "arm64" {
 		return "aarch64"
 	}
@@ -247,10 +261,18 @@ func MuslArchitecture(triple string) string {
 	return CanonicalArchName(triple)
 }

-// LibcPath returns the path to the libc directory. The libc path will be either
-// a precompiled libc shipped with a TinyGo build, or a libc path in the cache
-// directory (which might not yet be built).
-func (c *Config) LibcPath(name string) (path string, precompiled bool) {
+// Returns true if the libc needs to include malloc, for the libcs where this
+// matters.
+func (c *Config) LibcNeedsMalloc() bool {
+	if c.GC() == "boehm" && c.Target.Libc == "wasi-libc" {
+		return true
+	}
+	return false
+}
+
+// LibraryPath returns the path to the library build directory. The path will be
+// a library path in the cache directory (which might not yet be built).
+func (c *Config) LibraryPath(name string) string {
 	archname := c.Triple()
 	if c.CPU() != "" {
 		archname += "-" + c.CPU()
@@ -266,17 +288,19 @@ func (c *Config) LibcPath(name string) (path string, precompiled bool) {
 		archname += "-" + c.Target.Libc
 	}

-	// Try to load a precompiled library.
-	precompiledDir := filepath.Join(goenv.Get("TINYGOROOT"), "pkg", archname, name)
-	if _, err := os.Stat(precompiledDir); err == nil {
-		// Found a precompiled library for this OS/architecture. Return the path
-		// directly.
-		return precompiledDir, true
+	// Append a version string, if this library has a version.
+	if v, ok := libVersions[name]; ok {
+		archname += "-v" + strconv.Itoa(v)
+	}
+
+	options := ""
+	if c.LibcNeedsMalloc() {
+		options += "+malloc"
 	}

 	// No precompiled library found. Determine the path name that will be used
 	// in the build cache.
-	return filepath.Join(goenv.Get("GOCACHE"), name+"-"+archname), false
+	return filepath.Join(goenv.Get("GOCACHE"), name+options+"-"+archname)
 }

 // DefaultBinaryExtension returns the default extension for binaries, such as
@@ -325,7 +349,7 @@ func (c *Config) CFlags(libclang bool) []string {
 	// Use the same optimization level as TinyGo.
 	cflags = append(cflags, "-O"+c.Options.Opt)
 	// Set the LLVM target triple.
-	cflags = append(cflags, "--target="+c.Triple())
+	cflags = append(cflags, "--target="+ClangTriple(c.Triple()))
 	// Set the -mcpu (or similar) flag.
 	if c.Target.CPU != "" {
 		if c.GOARCH() == "amd64" || c.GOARCH() == "386" {
@@ -360,7 +384,7 @@ func (c *Config) LibcCFlags() []string {
 	case "picolibc":
 		root := goenv.Get("TINYGOROOT")
 		picolibcDir := filepath.Join(root, "lib", "picolibc", "newlib", "libc")
-		path, _ := c.LibcPath("picolibc")
+		path := c.LibraryPath("picolibc")
 		return []string{
 			"-nostdlibinc",
 			"-isystem", filepath.Join(path, "include"),
@@ -370,7 +394,7 @@ func (c *Config) LibcCFlags() []string {
 		}
 	case "musl":
 		root := goenv.Get("TINYGOROOT")
-		path, _ := c.LibcPath("musl")
+		path := c.LibraryPath("musl")
 		arch := MuslArchitecture(c.Triple())
 		return []string{
 			"-nostdlibinc",
@@ -380,24 +404,36 @@ func (c *Config) LibcCFlags() []string {
 			"-isystem", filepath.Join(root, "lib", "musl", "include"),
 		}
 	case "wasi-libc":
-		root := goenv.Get("TINYGOROOT")
+		path := c.LibraryPath("wasi-libc")
 		return []string{
 			"-nostdlibinc",
-			"-isystem", root + "/lib/wasi-libc/sysroot/include",
+			"-isystem", filepath.Join(path, "include"),
 		}
 	case "wasmbuiltins":
 		// nothing to add (library is purely for builtins)
 		return nil
 	case "mingw-w64":
 		root := goenv.Get("TINYGOROOT")
-		path, _ := c.LibcPath("mingw-w64")
-		return []string{
+		path := c.LibraryPath("mingw-w64")
+		cflags := []string{
 			"-nostdlibinc",
 			"-isystem", filepath.Join(path, "include"),
 			"-isystem", filepath.Join(root, "lib", "mingw-w64", "mingw-w64-headers", "crt"),
+			"-isystem", filepath.Join(root, "lib", "mingw-w64", "mingw-w64-headers", "include"),
 			"-isystem", filepath.Join(root, "lib", "mingw-w64", "mingw-w64-headers", "defaults", "include"),
-			"-D_UCRT",
 		}
+		if c.GOARCH() == "386" {
+			cflags = append(cflags,
+				"-D__MSVCRT_VERSION__=0x700", // Microsoft Visual C++ .NET 2002
+				"-D_WIN32_WINNT=0x0501",      // target Windows XP
+			)
+		} else {
+			cflags = append(cflags,
+				"-D_UCRT",
+				"-D_WIN32_WINNT=0x0a00", // target Windows 10
+			)
+		}
+		return cflags
 	case "":
 		// No libc specified, nothing to add.
 		return nil
@@ -427,6 +463,22 @@ func (c *Config) LDFlags() []string {
 	return ldflags
 }

+// LinkerFlavor returns how the configured linker should be driven.
+// Usually this is derived from GOOS, but targets may override it explicitly.
+func (c *Config) LinkerFlavor() string {
+	if c.Target.LinkerFlavor != "" {
+		return c.Target.LinkerFlavor
+	}
+	switch c.GOOS() {
+	case "windows":
+		return "coff"
+	case "darwin":
+		return "darwin"
+	default:
+		return "gnu"
+	}
+}
+
 // ExtraFiles returns the list of extra files to be built and linked with the
 // executable. This can include extra C and assembly files.
 func (c *Config) ExtraFiles() []string {
@@ -497,10 +549,10 @@ func (c *Config) Programmer() (method, openocdInterface string) {
 	case "":
 		// No configuration supplied.
 		return c.Target.FlashMethod, c.Target.OpenOCDInterface
-	case "openocd", "msd", "command":
+	case "openocd", "msd", "command", "adb":
 		// The -programmer flag only specifies the flash method.
 		return c.Options.Programmer, c.Target.OpenOCDInterface
-	case "bmp":
+	case "bmp", "probe-rs":
 		// The -programmer flag only specifies the flash method.
 		return c.Options.Programmer, ""
 	default:
@@ -0,0 +1,9 @@
+//go:build !llvm22 && !llvm14 && !llvm15 && !llvm16 && !llvm17 && !llvm18 && !llvm19
+
+package compileopts
+
+// patchFeatures applies LLVM-version-specific feature name mappings.
+// For LLVM 20/21, features in the target JSON files are already correct.
+func patchFeatures(features string) string {
+	return features
+}
@@ -0,0 +1,26 @@
+//go:build llvm22
+
+package compileopts
+
+import "strings"
+
+// patchFeatures applies LLVM-version-specific feature name mappings.
+// LLVM 22 renamed several Xtensa target features.
+func patchFeatures(features string) string {
+	// Xtensa feature renames in LLVM 22:
+	//   atomctl → (removed, no direct replacement)
+	//   memctl  → (removed, no direct replacement)
+	//   esp32s3 → esp32s3ops
+	//   timerint → timers3 (for esp32/esp32s3) or timers1 (for esp8266)
+	// Since we can't distinguish which timer variant at this level,
+	// just remove the obsolete features. The CPU definition already
+	// implies the correct features in LLVM 22.
+	replacer := strings.NewReplacer(
+		"+atomctl,", "",
+		"+memctl,", "",
+		"+esp32s3,", "+esp32s3ops,",
+		"+timerint,", "",
+		",+timerint", "",
+	)
+	return replacer.Replace(features)
+}
@@ -0,0 +1,16 @@
+//go:build llvm14 || llvm15 || llvm16 || llvm17 || llvm18 || llvm19
+
+package compileopts
+
+import "strings"
+
+// patchFeatures applies LLVM-version-specific feature name mappings.
+// LLVM 19 and earlier do not have +bulk-memory-opt or
+// +call-indirect-overlong for WebAssembly (added in LLVM 20).
+func patchFeatures(features string) string {
+	features = strings.ReplaceAll(features, ",+bulk-memory-opt", "")
+	features = strings.ReplaceAll(features, "+bulk-memory-opt,", "")
+	features = strings.ReplaceAll(features, ",+call-indirect-overlong", "")
+	features = strings.ReplaceAll(features, "+call-indirect-overlong,", "")
+	return features
+}
@@ -0,0 +1,65 @@
+package compileopts
+
+import (
+	"go/build/constraint"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+)
+
+// TestFinalizerRunnerSchedulerCoverage verifies that each scheduler selects one runner file.
+// It uses validSchedulerOptions so new schedulers are included.
+func TestFinalizerRunnerSchedulerCoverage(t *testing.T) {
+	files := []string{
+		"gc_finalizer_sched.go",
+		"gc_finalizer_sched_none.go",
+		"gc_finalizer_sched_other.go",
+	}
+	exprs := make([]constraint.Expr, len(files))
+	for i, name := range files {
+		exprs[i] = readBuildConstraint(t, filepath.Join("..", "src", "runtime", name))
+	}
+
+	for _, sched := range validSchedulerOptions {
+		// The finalizer table exists under block GCs.
+		// gc.conservative satisfies the GC condition in every constraint.
+		tags := map[string]bool{
+			"gc.conservative":    true,
+			"scheduler." + sched: true,
+		}
+		var matched []string
+		for i, expr := range exprs {
+			if expr.Eval(func(tag string) bool { return tags[tag] }) {
+				matched = append(matched, files[i])
+			}
+		}
+		if len(matched) != 1 {
+			t.Errorf("scheduler.%s: spawnFinalizerRunner defined in %d files %v, want exactly 1",
+				sched, len(matched), matched)
+		}
+	}
+}
+
+func readBuildConstraint(t *testing.T, path string) constraint.Expr {
+	t.Helper()
+	data, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, line := range strings.Split(string(data), "\n") {
+		line = strings.TrimSpace(line)
+		if constraint.IsGoBuild(line) {
+			expr, err := constraint.Parse(line)
+			if err != nil {
+				t.Fatalf("%s: %v", path, err)
+			}
+			return expr
+		}
+		if line != "" && !strings.HasPrefix(line, "//") {
+			break // reached code before any //go:build line
+		}
+	}
+	t.Fatalf("%s: no //go:build line found", path)
+	return nil
+}
@@ -3,6 +3,7 @@ package compileopts
 import (
 	"fmt"
 	"regexp"
+	"slices"
 	"strings"
 	"time"
 )
@@ -10,7 +11,7 @@ import (
 var (
 	validBuildModeOptions     = []string{"default", "c-shared", "wasi-legacy"}
 	validGCOptions            = []string{"none", "leaking", "conservative", "custom", "precise", "boehm"}
-	validSchedulerOptions     = []string{"none", "tasks", "asyncify"}
+	validSchedulerOptions     = []string{"none", "tasks", "asyncify", "threads", "cores"}
 	validSerialOptions        = []string{"none", "uart", "usb", "rtt"}
 	validPrintSizeOptions     = []string{"none", "short", "full", "html"}
 	validPanicStrategyOptions = []string{"print", "trap"}
@@ -21,50 +22,53 @@ var (
 // usually passed from the command line, but can also be passed in environment
 // variables for example.
 type Options struct {
-	GOOS            string // environment variable
-	GOARCH          string // environment variable
-	GOARM           string // environment variable (only used with GOARCH=arm)
-	GOMIPS          string // environment variable (only used with GOARCH=mips and GOARCH=mipsle)
-	Directory       string // working dir, leave it unset to use the current working dir
-	Target          string
-	BuildMode       string // -buildmode flag
-	Opt             string
-	GC              string
-	PanicStrategy   string
-	Scheduler       string
-	StackSize       uint64 // goroutine stack size (if none could be automatically determined)
-	Serial          string
-	Work            bool // -work flag to print temporary build directory
-	InterpTimeout   time.Duration
-	PrintIR         bool
-	DumpSSA         bool
-	VerifyIR        bool
-	SkipDWARF       bool
-	PrintCommands   func(cmd string, args ...string) `json:"-"`
-	Semaphore       chan struct{}                    `json:"-"` // -p flag controls cap
-	Debug           bool
-	PrintSizes      string
-	PrintAllocs     *regexp.Regexp // regexp string
-	PrintStacks     bool
-	Tags            []string
-	GlobalValues    map[string]map[string]string // map[pkgpath]map[varname]value
-	TestConfig      TestConfig
-	Programmer      string
-	OpenOCDCommands []string
-	LLVMFeatures    string
-	PrintJSON       bool
-	Monitor         bool
-	BaudRate        int
-	Timeout         time.Duration
-	WITPackage      string // pass through to wasm-tools component embed invocation
-	WITWorld        string // pass through to wasm-tools component embed -w option
-	ExtLDFlags      []string
+	GOOS                    string // environment variable
+	GOARCH                  string // environment variable
+	GOARM                   string // environment variable (only used with GOARCH=arm)
+	GOMIPS                  string // environment variable (only used with GOARCH=mips and GOARCH=mipsle)
+	Directory               string // working dir, leave it unset to use the current working dir
+	Target                  string
+	BuildMode               string // -buildmode flag
+	Opt                     string
+	GC                      string
+	PanicStrategy           string
+	Scheduler               string
+	StackSize               uint64 // goroutine stack size (if none could be automatically determined)
+	Serial                  string
+	Work                    bool // -work flag to print temporary build directory
+	InterpTimeout           time.Duration
+	InterpMaxLoopIterations int
+	PrintIR                 bool
+	DumpSSA                 bool
+	VerifyIR                bool
+	SkipDWARF               bool
+	PrintCommands           func(cmd string, args ...string) `json:"-"`
+	Semaphore               chan struct{}                    `json:"-"` // -p flag controls cap
+	Debug                   bool
+	Nobounds                bool
+	PrintSizes              string
+	PrintAllocs             *regexp.Regexp // regexp string
+	PrintAllocsCover        bool           // emit allocs in go coverage tool format
+	PrintStacks             bool
+	Tags                    []string
+	GlobalValues            map[string]map[string]string // map[pkgpath]map[varname]value
+	TestConfig              TestConfig
+	Programmer              string
+	OpenOCDCommands         []string
+	LLVMFeatures            string
+	Monitor                 bool
+	BaudRate                int
+	Timeout                 time.Duration
+	WITPackage              string // pass through to wasm-tools component embed invocation
+	WITWorld                string // pass through to wasm-tools component embed -w option
+	ExtLDFlags              []string
+	GoCompatibility         bool // enable to check for Go version compatibility
 }

 // Verify performs a validation on the given options, raising an error if options are not valid.
 func (o *Options) Verify() error {
 	if o.BuildMode != "" {
-		valid := isInArray(validBuildModeOptions, o.BuildMode)
+		valid := slices.Contains(validBuildModeOptions, o.BuildMode)
 		if !valid {
 			return fmt.Errorf(`invalid buildmode option '%s': valid values are %s`,
 				o.BuildMode,
@@ -72,7 +76,7 @@ func (o *Options) Verify() error {
 		}
 	}
 	if o.GC != "" {
-		valid := isInArray(validGCOptions, o.GC)
+		valid := slices.Contains(validGCOptions, o.GC)
 		if !valid {
 			return fmt.Errorf(`invalid gc option '%s': valid values are %s`,
 				o.GC,
@@ -81,7 +85,7 @@ func (o *Options) Verify() error {
 	}

 	if o.Scheduler != "" {
-		valid := isInArray(validSchedulerOptions, o.Scheduler)
+		valid := slices.Contains(validSchedulerOptions, o.Scheduler)
 		if !valid {
 			return fmt.Errorf(`invalid scheduler option '%s': valid values are %s`,
 				o.Scheduler,
@@ -90,7 +94,7 @@ func (o *Options) Verify() error {
 	}

 	if o.Serial != "" {
-		valid := isInArray(validSerialOptions, o.Serial)
+		valid := slices.Contains(validSerialOptions, o.Serial)
 		if !valid {
 			return fmt.Errorf(`invalid serial option '%s': valid values are %s`,
 				o.Serial,
@@ -99,7 +103,7 @@ func (o *Options) Verify() error {
 	}

 	if o.PrintSizes != "" {
-		valid := isInArray(validPrintSizeOptions, o.PrintSizes)
+		valid := slices.Contains(validPrintSizeOptions, o.PrintSizes)
 		if !valid {
 			return fmt.Errorf(`invalid size option '%s': valid values are %s`,
 				o.PrintSizes,
@@ -108,7 +112,7 @@ func (o *Options) Verify() error {
 	}

 	if o.PanicStrategy != "" {
-		valid := isInArray(validPanicStrategyOptions, o.PanicStrategy)
+		valid := slices.Contains(validPanicStrategyOptions, o.PanicStrategy)
 		if !valid {
 			return fmt.Errorf(`invalid panic option '%s': valid values are %s`,
 				o.PanicStrategy,
@@ -117,19 +121,10 @@ func (o *Options) Verify() error {
 	}

 	if o.Opt != "" {
-		if !isInArray(validOptOptions, o.Opt) {
+		if !slices.Contains(validOptOptions, o.Opt) {
 			return fmt.Errorf("invalid -opt=%s: valid values are %s", o.Opt, strings.Join(validOptOptions, ", "))
 		}
 	}

 	return nil
 }
-
-func isInArray(arr []string, item string) bool {
-	for _, i := range arr {
-		if i == item {
-			return true
-		}
-	}
-	return false
-}
@@ -10,7 +10,7 @@ import (
 func TestVerifyOptions(t *testing.T) {

 	expectedGCError := errors.New(`invalid gc option 'incorrect': valid values are none, leaking, conservative, custom, precise, boehm`)
-	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, asyncify`)
+	expectedSchedulerError := errors.New(`invalid scheduler option 'incorrect': valid values are none, tasks, asyncify, threads, cores`)
 	expectedPrintSizeError := errors.New(`invalid size option 'incorrect': valid values are none, short, full, html`)
 	expectedPanicStrategyError := errors.New(`invalid panic option 'incorrect': valid values are print, trap`)

@@ -24,6 +24,7 @@ import (
 // https://github.com/shepmaster/rust-arduino-blink-led-no-core-with-cargo/blob/master/blink/arduino.json
 type TargetSpec struct {
 	Inherits         []string `json:"inherits,omitempty"`
+	InheritableOnly  bool     `json:"inheritable-only"` // this target is only meant to be inherited from, not used directly
 	Triple           string   `json:"llvm-target,omitempty"`
 	CPU              string   `json:"cpu,omitempty"`
 	ABI              string   `json:"target-abi,omitempty"` // roughly equivalent to -mabi= flag
@@ -37,7 +38,8 @@ type TargetSpec struct {
 	Scheduler        string   `json:"scheduler,omitempty"`
 	Serial           string   `json:"serial,omitempty"` // which serial output to use (uart, usb, none)
 	Linker           string   `json:"linker,omitempty"`
-	RTLib            string   `json:"rtlib,omitempty"` // compiler runtime library (libgcc, compiler-rt)
+	LinkerFlavor     string   `json:"linker-flavor,omitempty"` // how to drive the configured linker (for example: gnu, coff, darwin)
+	RTLib            string   `json:"rtlib,omitempty"`         // compiler runtime library (libgcc, compiler-rt)
 	Libc             string   `json:"libc,omitempty"`
 	AutoStackSize    *bool    `json:"automatic-stack-size,omitempty"` // Determine stack size automatically at compile time.
 	DefaultStackSize uint64   `json:"default-stack-size,omitempty"`   // Default stack size if the size couldn't be determined at compile time.
@@ -63,6 +65,10 @@ type TargetSpec struct {
 	OpenOCDCommands  []string `json:"openocd-commands,omitempty"`
 	OpenOCDVerify    *bool    `json:"openocd-verify,omitempty"` // enable verify when flashing with openocd
 	JLinkDevice      string   `json:"jlink-device,omitempty"`
+	ADBPreCommands   []string `json:"adb-pre-commands,omitempty"`
+	ADBPushRemote    string   `json:"adb-push-remote,omitempty"`
+	ADBPostCommands  []string `json:"adb-post-commands,omitempty"`
+	ProbeRSChip      string   `json:"probe-rs-chip,omitempty"`
 	CodeModel        string   `json:"code-model,omitempty"`
 	RelocationModel  string   `json:"relocation-model,omitempty"`
 	WITPackage       string   `json:"wit-package,omitempty"`
@@ -148,6 +154,11 @@ func (spec *TargetSpec) loadFromGivenStr(str string) error {

 // resolveInherits loads inherited targets, recursively.
 func (spec *TargetSpec) resolveInherits() error {
+	// Save InheritableOnly before resolving, since it must not propagate
+	// from parent to child (a board target should not become inheritable-only
+	// just because its parent processor target is).
+	inheritableOnly := spec.InheritableOnly
+
 	// First create a new spec with all the inherited properties.
 	newSpec := &TargetSpec{}
 	for _, name := range spec.Inherits {
@@ -173,11 +184,29 @@ func (spec *TargetSpec) resolveInherits() error {
 	}
 	*spec = *newSpec

+	// Restore InheritableOnly from the original spec, not from parents.
+	spec.InheritableOnly = inheritableOnly
+
 	return nil
 }

 // Load a target specification.
 func LoadTarget(options *Options) (*TargetSpec, error) {
+	if options.Target == "" && options.GOARCH == "wasm" {
+		// Set a specific target if we're building from a known GOOS/GOARCH
+		// combination that is defined in a target JSON file.
+		switch options.GOOS {
+		case "js":
+			options.Target = "wasm"
+		case "wasip1":
+			options.Target = "wasip1"
+		case "wasip2":
+			options.Target = "wasip2"
+		default:
+			return nil, errors.New("GOARCH=wasm but GOOS is not set correctly. Please set GOOS to js, wasip1, or wasip2.")
+		}
+	}
+
 	if options.Target == "" {
 		return defaultTarget(options)
 	}
@@ -224,10 +253,17 @@ func GetTargetSpecs() (map[string]*TargetSpec, error) {
 			continue
 		}
 		path := filepath.Join(dir, entry.Name())
+
 		spec, err := LoadTarget(&Options{Target: path})
 		if err != nil {
 			return nil, fmt.Errorf("could not list target: %w", err)
 		}
+
+		if spec.InheritableOnly {
+			// Skip targets that are only meant to be inherited from, not used directly.
+			continue
+		}
+
 		if spec.FlashMethod == "" && spec.FlashCommand == "" && spec.Emulator == "" {
 			// This doesn't look like a regular target file, but rather like
 			// a parent target (such as targets/cortex-m.json).
@@ -247,7 +283,6 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 		GOOS:             options.GOOS,
 		GOARCH:           options.GOARCH,
 		BuildTags:        []string{options.GOOS, options.GOARCH},
-		Scheduler:        "tasks",
 		Linker:           "cc",
 		DefaultStackSize: 1024 * 64, // 64kB
 		GDB:              []string{"gdb"},
@@ -355,7 +390,7 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 			return nil, fmt.Errorf("invalid GOMIPS=%s: must be hardfloat or softfloat", options.GOMIPS)
 		}
 	case "wasm":
-		return nil, fmt.Errorf("GOARCH=wasm but GOOS is unset. Please set GOOS to wasm, wasip1, or wasip2.")
+		return nil, fmt.Errorf("GOARCH=wasm but GOOS is unset. Please set GOOS to js, wasip1, or wasip2.")
 	default:
 		return nil, fmt.Errorf("unknown GOARCH=%s", options.GOARCH)
 	}
@@ -365,12 +400,13 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 	llvmvendor := "unknown"
 	switch options.GOOS {
 	case "darwin":
-		spec.GC = "precise"
+		spec.GC = "boehm"
 		platformVersion := "10.12.0"
 		if options.GOARCH == "arm64" {
 			platformVersion = "11.0.0" // first macosx platform with arm64 support
 		}
 		llvmvendor = "apple"
+		spec.Scheduler = "threads"
 		spec.Linker = "ld.lld"
 		spec.Libc = "darwin-libSystem"
 		// Use macosx* instead of darwin, otherwise darwin/arm64 will refer to
@@ -384,11 +420,13 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 		)
 		spec.ExtraFiles = append(spec.ExtraFiles,
 			"src/internal/futex/futex_darwin.c",
+			"src/internal/task/task_threads.c",
 			"src/runtime/os_darwin.c",
 			"src/runtime/runtime_unix.c",
 			"src/runtime/signal.c")
 	case "linux":
 		spec.GC = "boehm"
+		spec.Scheduler = "threads"
 		spec.Linker = "ld.lld"
 		spec.RTLib = "compiler-rt"
 		spec.Libc = "musl"
@@ -409,28 +447,40 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 		}
 		spec.ExtraFiles = append(spec.ExtraFiles,
 			"src/internal/futex/futex_linux.c",
+			"src/internal/task/task_threads.c",
 			"src/runtime/runtime_unix.c",
 			"src/runtime/signal.c")
 	case "windows":
-		spec.GC = "precise"
+		spec.GC = "boehm"
+		spec.Scheduler = "tasks"
 		spec.Linker = "ld.lld"
 		spec.Libc = "mingw-w64"
-		// Note: using a medium code model, low image base and no ASLR
-		// because Go doesn't really need those features. ASLR patches
-		// around issues for unsafe languages like C/C++ that are not
-		// normally present in Go (without explicitly opting in).
-		// For more discussion:
-		// https://groups.google.com/g/Golang-nuts/c/Jd9tlNc6jUE/m/Zo-7zIP_m3MJ?pli=1
 		switch options.GOARCH {
+		case "386":
+			spec.LDFlags = append(spec.LDFlags,
+				"-m", "i386pe",
+				"--major-os-version", "4",
+				"--major-subsystem-version", "4",
+			)
+			// __udivdi3 is not present in ucrt it seems.
+			spec.RTLib = "compiler-rt"
 		case "amd64":
+			// Note: using a medium code model, low image base and no ASLR
+			// because Go doesn't really need those features. ASLR patches
+			// around issues for unsafe languages like C/C++ that are not
+			// normally present in Go (without explicitly opting in).
+			// For more discussion:
+			// https://groups.google.com/g/Golang-nuts/c/Jd9tlNc6jUE/m/Zo-7zIP_m3MJ?pli=1
 			spec.LDFlags = append(spec.LDFlags,
 				"-m", "i386pep",
 				"--image-base", "0x400000",
 			)
+			spec.RTLib = "compiler-rt"
 		case "arm64":
 			spec.LDFlags = append(spec.LDFlags,
 				"-m", "arm64pe",
 			)
+			spec.RTLib = "compiler-rt"
 		}
 		spec.LDFlags = append(spec.LDFlags,
 			"-Bdynamic",
@@ -438,12 +488,20 @@ func defaultTarget(options *Options) (*TargetSpec, error) {
 			"--no-insert-timestamp",
 			"--no-dynamicbase",
 		)
+		spec.ExtraFiles = append(spec.ExtraFiles,
+			"src/runtime/runtime_windows.c")
 	case "wasm", "wasip1", "wasip2":
 		return nil, fmt.Errorf("GOOS=%s but GOARCH is unset. Please set GOARCH to wasm", options.GOOS)
 	default:
 		return nil, fmt.Errorf("unknown GOOS=%s", options.GOOS)
 	}

+	if spec.GC == "boehm" {
+		// Add this file only when needed. This fixes a build failure on
+		// Windows.
+		spec.ExtraFiles = append(spec.ExtraFiles, "src/runtime/gc_boehm.c")
+	}
+
 	// Target triples (which actually have four components, but are called
 	// triples for historical reasons) have the form:
 	//   arch-vendor-os-environment
@@ -1,8 +1,10 @@
 package compileopts

 import (
+	"encoding/csv"
 	"errors"
 	"io/fs"
+	"os"
 	"reflect"
 	"testing"
 )
@@ -23,6 +25,69 @@ func TestLoadTarget(t *testing.T) {
 	}
 }

+func TestGetTargetSpecs_InheritableOnlyTargetsExcluded(t *testing.T) {
+	specs, err := GetTargetSpecs()
+	if err != nil {
+		t.Fatal("GetTargetSpecs failed:", err)
+	}
+
+	// Inheritable-only processor-level targets should not appear in the listing.
+	inheritableOnlyTargets := []string{"esp32", "esp32c3", "esp32s3", "esp8266", "rp2040", "rp2350", "rp2350b"}
+	for _, name := range inheritableOnlyTargets {
+		if _, ok := specs[name]; ok {
+			t.Errorf("inheritable-only target %q should not appear in GetTargetSpecs", name)
+		}
+	}
+
+	// Board targets that inherit from inheritable-only targets should still appear.
+	boardTargets := []string{"esp32-coreboard-v2", "pico"}
+	for _, name := range boardTargets {
+		if _, ok := specs[name]; !ok {
+			t.Errorf("board target %q should appear in GetTargetSpecs", name)
+		}
+	}
+}
+
+func TestLoadTarget_InheritableOnlyTargetStillLoadable(t *testing.T) {
+	// Inheritable-only targets should still be loadable directly (for building).
+	_, err := LoadTarget(&Options{Target: "esp32"})
+	if err != nil {
+		t.Errorf("LoadTarget should still load inheritable-only target esp32: %v", err)
+	}
+}
+
+func TestLoadPY32ConcreteTargets(t *testing.T) {
+	file, err := os.Open("../tools/gen-py32-targets/devices.csv")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer file.Close()
+
+	records, err := csv.NewReader(file).ReadAll()
+	if err != nil {
+		t.Fatal(err)
+	}
+	for _, record := range records[1:] {
+		part, core := record[0], record[2]
+		t.Run(part, func(t *testing.T) {
+			spec, err := LoadTarget(&Options{Target: part})
+			if err != nil {
+				t.Fatal(err)
+			}
+			if spec.LinkerScript == "" {
+				t.Error("concrete target has no linker script")
+			}
+			wantCPU := "cortex-m0plus"
+			if core == "m4" {
+				wantCPU = "cortex-m4"
+			}
+			if spec.CPU != wantCPU {
+				t.Errorf("CPU is %q, want %q", spec.CPU, wantCPU)
+			}
+		})
+	}
+}
+
 func TestOverrideProperties(t *testing.T) {
 	baseAutoStackSize := true
 	base := &TargetSpec{
@@ -81,3 +146,52 @@ func TestOverrideProperties(t *testing.T) {
 	}

 }
+
+func TestConfigLinkerFlavor(t *testing.T) {
+	tests := []struct {
+		name   string
+		target *TargetSpec
+		goos   string
+		want   string
+	}{
+		{
+			name:   "default gnu",
+			target: &TargetSpec{},
+			goos:   "linux",
+			want:   "gnu",
+		},
+		{
+			name:   "default coff",
+			target: &TargetSpec{},
+			goos:   "windows",
+			want:   "coff",
+		},
+		{
+			name:   "default darwin",
+			target: &TargetSpec{},
+			goos:   "darwin",
+			want:   "darwin",
+		},
+		{
+			name: "target override",
+			target: &TargetSpec{
+				LinkerFlavor: "coff",
+			},
+			goos: "linux",
+			want: "coff",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			tc.target.GOOS = tc.goos
+			config := &Config{
+				Options: &Options{},
+				Target:  tc.target,
+			}
+			if got := config.LinkerFlavor(); got != tc.want {
+				t.Fatalf("LinkerFlavor() = %q, want %q", got, tc.want)
+			}
+		})
+	}
+}
@@ -0,0 +1,12 @@
+//go:build llvm22
+
+package compileopts
+
+import "strings"
+
+// ClangTriple returns the target triple to pass to Clang's --target flag.
+// LLVM 22 deprecated the "wasm32-unknown-wasi" triple in favor of
+// "wasm32-unknown-wasip1", so we substitute it here to avoid warnings.
+func ClangTriple(triple string) string {
+	return strings.Replace(triple, "wasm32-unknown-wasi", "wasm32-unknown-wasip1", 1)
+}
@@ -0,0 +1,9 @@
+//go:build !llvm22
+
+package compileopts
+
+// ClangTriple returns the target triple to pass to Clang's --target flag.
+// For pre-LLVM 22, the triple is used as-is.
+func ClangTriple(triple string) string {
+	return triple
+}
@@ -241,24 +241,37 @@ func (b *builder) createRuntimeAssert(assert llvm.Value, blockPrefix, assertFunc
 		}
 	}

-	// Put the fault block at the end of the function and the next block at the
-	// current insert position.
-	faultBlock := b.ctx.AddBasicBlock(b.llvmFn, blockPrefix+".throw")
+	faultBlock := b.getRuntimeAssertBlock(blockPrefix, assertFunc)
 	nextBlock := b.insertBasicBlock(blockPrefix + ".next")
-	b.blockExits[b.currentBlock] = nextBlock // adjust outgoing block for phi nodes
+	b.currentBlockInfo.exit = nextBlock // adjust outgoing block for phi nodes

 	// Now branch to the out-of-bounds or the regular block.
 	b.CreateCondBr(assert, faultBlock, nextBlock)

-	// Fail: the assert triggered so panic.
-	b.SetInsertPointAtEnd(faultBlock)
-	b.createRuntimeCall(assertFunc, nil, "")
-	b.CreateUnreachable()
-
 	// Ok: assert didn't trigger so continue normally.
 	b.SetInsertPointAtEnd(nextBlock)
 }

+func (b *builder) getRuntimeAssertBlock(blockPrefix, assertFunc string) llvm.BasicBlock {
+	if b.runtimeAssertBlocks == nil {
+		b.runtimeAssertBlocks = make(map[string]llvm.BasicBlock)
+	}
+	if block := b.runtimeAssertBlocks[assertFunc]; !block.IsNil() {
+		return block
+	}
+	savedBlock := b.GetInsertBlock()
+	block := b.ctx.AddBasicBlock(b.llvmFn, blockPrefix+".throw")
+	b.runtimeAssertBlocks[assertFunc] = block
+	b.SetInsertPointAtEnd(block)
+	if b.hasDeferFrame() {
+		b.createFaultCheckpoint()
+	}
+	b.createRuntimeCall(assertFunc, nil, "")
+	b.CreateUnreachable()
+	b.SetInsertPointAtEnd(savedBlock)
+	return block
+}
+
 // extendInteger extends the value to at least targetType using a zero or sign
 // extend. The resulting value is not truncated: it may still be bigger than
 // targetType.
@@ -32,6 +32,12 @@ const (
 	// Whether this is a full or partial Go parameter (int, slice, etc).
 	// The extra context parameter is not a Go parameter.
 	paramIsGoParam = 1 << iota
+
+	// Whether this is a readonly parameter (for example, a string pointer).
+	paramIsReadonly
+
+	// Whether this parameter is passed through backing storage.
+	paramIsIndirect
 )

 // createRuntimeCallCommon creates a runtime call. Use createRuntimeCall or
@@ -76,7 +82,15 @@ func (b *builder) createCall(fnType llvm.Type, fn llvm.Value, args []llvm.Value,
 		fragments := b.expandFormalParam(arg)
 		expanded = append(expanded, fragments...)
 	}
-	return b.CreateCall(fnType, fn, expanded, name)
+	call := b.CreateCall(fnType, fn, expanded, name)
+	if !fn.IsAFunction().IsNil() {
+		if cc := fn.FunctionCallConv(); cc != llvm.CCallConv {
+			// Set a different calling convention if needed.
+			// This is needed for GetModuleHandleExA on Windows, for example.
+			call.SetInstructionCallConv(cc)
+		}
+	}
+	return call
 }

 // createInvoke is like createCall but continues execution at the landing pad if
@@ -91,6 +105,18 @@ func (b *builder) createInvoke(fnType llvm.Type, fn llvm.Value, args []llvm.Valu
 // Expand an argument type to a list that can be used in a function call
 // parameter list.
 func (c *compilerContext) expandFormalParamType(t llvm.Type, name string, goType types.Type) []paramInfo {
+	if c.isIndirectAggregate(t) {
+		return []paramInfo{{
+			llvmType: c.dataPtrType,
+			name:     name,
+			elemSize: c.targetData.TypeAllocSize(t),
+			flags:    paramIsGoParam | paramIsReadonly | paramIsIndirect,
+		}}
+	}
+	return c.expandDirectFormalParamType(t, name, goType)
+}
+
+func (c *compilerContext) expandDirectFormalParamType(t llvm.Type, name string, goType types.Type) []paramInfo {
 	switch t.TypeKind() {
 	case llvm.StructTypeKind:
 		fieldInfos := c.flattenAggregateType(t, name, goType)
@@ -104,6 +130,38 @@ func (c *compilerContext) expandFormalParamType(t llvm.Type, name string, goType
 	return []paramInfo{c.getParamInfo(t, name, goType)}
 }

+func (c *compilerContext) storedParamType(t llvm.Type, exported bool) llvm.Type {
+	if c.isIndirectParam(t, exported) {
+		return c.dataPtrType
+	}
+	return t
+}
+
+func (c *compilerContext) isIndirectParam(t llvm.Type, exported bool) bool {
+	return !exported && c.isIndirectAggregate(t)
+}
+
+func (b *builder) appendStoredValueTypes(valueTypes []llvm.Type, values []ssa.Value, exported bool) []llvm.Type {
+	for _, value := range values {
+		valueTypes = append(valueTypes, b.storedParamType(b.getLLVMType(value.Type()), exported))
+	}
+	return valueTypes
+}
+
+func (b *builder) appendStoredParamTypes(valueTypes []llvm.Type, params []*types.Var, exported bool) []llvm.Type {
+	for _, param := range params {
+		valueTypes = append(valueTypes, b.storedParamType(b.getLLVMType(param.Type()), exported))
+	}
+	return valueTypes
+}
+
+func (b *builder) prependIndirectResult(sig *types.Signature, exported bool, params []llvm.Value, name string) []llvm.Value {
+	if resultType, indirect := b.hasIndirectResult(sig); !exported && indirect {
+		return append([]llvm.Value{b.createIndirectStorage(resultType, name)}, params...)
+	}
+	return params
+}
+
 // expandFormalParamOffsets returns a list of offsets from the start of an
 // object of type t after it would have been split up by expandFormalParam. This
 // is useful for debug information, where it is necessary to know the offset
@@ -159,6 +217,7 @@ func (c *compilerContext) flattenAggregateType(t llvm.Type, name string, goType
 				continue
 			}
 			suffix := strconv.Itoa(i)
+			isString := false
 			if goType != nil {
 				// Try to come up with a good suffix for this struct field,
 				// depending on which Go type it's based on.
@@ -175,12 +234,16 @@ func (c *compilerContext) flattenAggregateType(t llvm.Type, name string, goType
 						suffix = []string{"r", "i"}[i]
 					case types.String:
 						suffix = []string{"data", "len"}[i]
+						isString = true
 					}
 				case *types.Signature:
 					suffix = []string{"context", "funcptr"}[i]
 				}
 			}
 			subInfos := c.flattenAggregateType(subfield, name+"."+suffix, extractSubfield(goType, i))
+			if isString {
+				subInfos[0].flags |= paramIsReadonly
+			}
 			paramInfos = append(paramInfos, subInfos...)
 		}
 		return paramInfos
@@ -30,17 +30,15 @@ func (b *builder) createMakeChan(expr *ssa.MakeChan) llvm.Value {
 // actual channel send operation during goroutine lowering.
 func (b *builder) createChanSend(instr *ssa.Send) {
 	ch := b.getValue(instr.Chan, getPos(instr))
-	chanValue := b.getValue(instr.X, getPos(instr))

 	// store value-to-send
 	valueType := b.getLLVMType(instr.X.Type())
 	isZeroSize := b.targetData.TypeAllocSize(valueType) == 0
-	var valueAlloca, valueAllocaSize llvm.Value
+	var storage valueStorage
 	if isZeroSize {
-		valueAlloca = llvm.ConstNull(b.dataPtrType)
+		storage.ptr = llvm.ConstNull(b.dataPtrType)
 	} else {
-		valueAlloca, valueAllocaSize = b.createTemporaryAlloca(valueType, "chan.value")
-		b.CreateStore(chanValue, valueAlloca)
+		storage = b.getValueStorage(instr.X, "chan.value")
 	}

 	// Allocate buffer for the channel operation.
@@ -48,15 +46,13 @@ func (b *builder) createChanSend(instr *ssa.Send) {
 	channelOpAlloca, channelOpAllocaSize := b.createTemporaryAlloca(channelOp, "chan.op")

 	// Do the send.
-	b.createRuntimeCall("chanSend", []llvm.Value{ch, valueAlloca, channelOpAlloca}, "")
+	b.createRuntimeInvoke("chanSend", []llvm.Value{ch, storage.ptr, channelOpAlloca}, "")

 	// End the lifetime of the allocas.
 	// This also works around a bug in CoroSplit, at least in LLVM 8:
 	// https://bugs.llvm.org/show_bug.cgi?id=41742
 	b.emitLifetimeEnd(channelOpAlloca, channelOpAllocaSize)
-	if !isZeroSize {
-		b.emitLifetimeEnd(valueAlloca, valueAllocaSize)
-	}
+	b.endValueStorage(storage)
 }

 // createChanRecv emits a pseudo chan receive operation. It is lowered to the
@@ -66,42 +62,22 @@ func (b *builder) createChanRecv(unop *ssa.UnOp) llvm.Value {
 	ch := b.getValue(unop.X, getPos(unop))

 	// Allocate memory to receive into.
-	isZeroSize := b.targetData.TypeAllocSize(valueType) == 0
-	var valueAlloca, valueAllocaSize llvm.Value
-	if isZeroSize {
-		valueAlloca = llvm.ConstNull(b.dataPtrType)
-	} else {
-		valueAlloca, valueAllocaSize = b.createTemporaryAlloca(valueType, "chan.value")
-	}
+	result := b.createRuntimeValueResult(valueType, unop.CommaOk, true, "chan")

 	// Allocate buffer for the channel operation.
 	channelOp := b.getLLVMRuntimeType("channelOp")
 	channelOpAlloca, channelOpAllocaSize := b.createTemporaryAlloca(channelOp, "chan.op")

 	// Do the receive.
-	commaOk := b.createRuntimeCall("chanRecv", []llvm.Value{ch, valueAlloca, channelOpAlloca}, "")
-	var received llvm.Value
-	if isZeroSize {
-		received = llvm.ConstNull(valueType)
-	} else {
-		received = b.CreateLoad(valueType, valueAlloca, "chan.received")
-		b.emitLifetimeEnd(valueAlloca, valueAllocaSize)
-	}
+	commaOk := b.createRuntimeCall("chanRecv", []llvm.Value{ch, result.valuePtr, channelOpAlloca}, "")
+	received := result.finish(b, commaOk, "chan.received")
 	b.emitLifetimeEnd(channelOpAlloca, channelOpAllocaSize)
-
-	if unop.CommaOk {
-		tuple := llvm.Undef(b.ctx.StructType([]llvm.Type{valueType, b.ctx.Int1Type()}, false))
-		tuple = b.CreateInsertValue(tuple, received, 0, "")
-		tuple = b.CreateInsertValue(tuple, commaOk, 1, "")
-		return tuple
-	} else {
-		return received
-	}
+	return received
 }

 // createChanClose closes the given channel.
 func (b *builder) createChanClose(ch llvm.Value) {
-	b.createRuntimeCall("chanClose", []llvm.Value{ch}, "")
+	b.createRuntimeInvoke("chanClose", []llvm.Value{ch}, "")
 }

 // createSelect emits all IR necessary for a select statements. That's a
@@ -170,9 +146,7 @@ func (b *builder) createSelect(expr *ssa.Select) llvm.Value {
 		case types.SendOnly:
 			// Store this value in an alloca and put a pointer to this alloca
 			// in the send state.
-			sendValue := b.getValue(state.Send, state.Pos)
-			alloca := llvmutil.CreateEntryBlockAlloca(b.Builder, sendValue.Type(), "select.send.value")
-			b.CreateStore(sendValue, alloca)
+			alloca := b.getSelectSendStorage(state.Send)
 			selectState = b.CreateInsertValue(selectState, alloca, 1, "")
 		default:
 			panic("unreachable")
@@ -280,7 +254,17 @@ func (b *builder) getChanSelectResult(expr *ssa.Extract) llvm.Value {
 		// receive can proceed at a time) so we'll get that alloca, bitcast
 		// it to the correct type, and dereference it.
 		recvbuf := b.selectRecvBuf[expr.Tuple.(*ssa.Select)]
-		typ := b.getLLVMType(expr.Type())
-		return b.CreateLoad(typ, recvbuf, "")
+		return b.loadFromStorage(recvbuf, expr.Type(), "select.received")
 	}
 }
+
+func (b *builder) getSelectSendStorage(value ssa.Value) llvm.Value {
+	typ := b.getLLVMType(value.Type())
+	if b.isIndirectAggregate(typ) {
+		return b.getValuePointer(value)
+	}
+	llvmValue := b.getValue(value, getPos(value))
+	ptr := llvmutil.CreateEntryBlockAlloca(b.Builder, typ, "select.send.value")
+	b.CreateStore(llvmValue, ptr)
+	return ptr
+}
@@ -58,6 +58,7 @@ type Config struct {
 	MaxStackAlloc      uint64
 	NeedsStackObjects  bool
 	Debug              bool // Whether to emit debug information in the LLVM module.
+	Nobounds           bool // Whether to skip bounds checks
 	PanicStrategy      string
 }

@@ -89,8 +90,10 @@ type compilerContext struct {
 	astComments      map[string]*ast.CommentGroup
 	embedGlobals     map[string][]*loader.EmbedFile
 	pkg              *types.Package
-	packageDir       string // directory for this package
+	loaderPkg        *loader.Package // current package being compiled (for AST access)
+	packageDir       string          // directory for this package
 	runtimePkg       *types.Package
+	localTypeNames   typeutil.Map // *types.Named (synthetic local from generic instantiation) -> string
 }

 // newCompilerContext returns a new compiler context ready for use, most
@@ -151,10 +154,14 @@ type builder struct {
 	llvmFnType        llvm.Type
 	llvmFn            llvm.Value
 	info              functionInfo
-	locals            map[ssa.Value]llvm.Value            // local variables
-	blockEntries      map[*ssa.BasicBlock]llvm.BasicBlock // a *ssa.BasicBlock may be split up
-	blockExits        map[*ssa.BasicBlock]llvm.BasicBlock // these are the exit blocks
+	locals            map[ssa.Value]llvm.Value // local variables
+	indirectValues    map[ssa.Value]llvm.Value
+	indirectReturn    llvm.Value
+	blockInfo         []blockInfo
 	currentBlock      *ssa.BasicBlock
+	currentBlockInfo  *blockInfo
+	tarjanStack       []uint
+	tarjanIndex       uint
 	phis              []phiNode
 	deferPtr          llvm.Value
 	deferFrame        llvm.Value
@@ -164,7 +171,7 @@ type builder struct {
 	dilocals          map[*types.Var]llvm.Metadata
 	initInlinedAt     llvm.Metadata            // fake inlinedAt position
 	initPseudoFuncs   map[string]llvm.Metadata // fake "inlined" functions for proper init debug locations
-	allDeferFuncs     []interface{}
+	allDeferFuncs     []any
 	deferFuncs        map[*ssa.Function]int
 	deferInvokeFuncs  map[string]int
 	deferClosureFuncs map[*ssa.Function]int
@@ -173,6 +180,9 @@ type builder struct {
 	deferBuiltinFuncs map[ssa.Value]deferBuiltin
 	runDefersBlock    []llvm.BasicBlock
 	afterDefersBlock  []llvm.BasicBlock
+
+	runtimeAssertBlocks  map[string]llvm.BasicBlock
+	interfaceAssertBlock llvm.BasicBlock
 }

 func newBuilder(c *compilerContext, irbuilder llvm.Builder, f *ssa.Function) *builder {
@@ -185,12 +195,24 @@ func newBuilder(c *compilerContext, irbuilder llvm.Builder, f *ssa.Function) *bu
 		llvmFn:          fn,
 		info:            c.getFunctionInfo(f),
 		locals:          make(map[ssa.Value]llvm.Value),
+		indirectValues:  make(map[ssa.Value]llvm.Value),
 		dilocals:        make(map[*types.Var]llvm.Metadata),
-		blockEntries:    make(map[*ssa.BasicBlock]llvm.BasicBlock),
-		blockExits:      make(map[*ssa.BasicBlock]llvm.BasicBlock),
 	}
 }

+type blockInfo struct {
+	// entry is the LLVM basic block corresponding to the start of this *ssa.Block.
+	entry llvm.BasicBlock
+
+	// exit is the LLVM basic block corresponding to the end of this *ssa.Block.
+	// It will be different than entry if any of the block's instructions contain internal branches.
+	exit llvm.BasicBlock
+
+	// tarjan holds state for applying Tarjan's strongly connected components algorithm to the CFG.
+	// This is used by defer.go to determine whether to stack- or heap-allocate defer data.
+	tarjan tarjanNode
+}
+
 type deferBuiltin struct {
 	callName string
 	pos      token.Pos
@@ -274,17 +296,23 @@ func Sizes(machine llvm.TargetMachine) types.Sizes {
 }

 // CompilePackage compiles a single package to a LLVM module.
+//
+// The SSA package must already be built. When packages are compiled
+// concurrently, the entire SSA program must be built before compilation starts.
 func CompilePackage(moduleName string, pkg *loader.Package, ssaPkg *ssa.Package, machine llvm.TargetMachine, config *Config, dumpSSA bool) (llvm.Module, []error) {
 	c := newCompilerContext(moduleName, machine, config, dumpSSA)
 	defer c.dispose()
 	c.packageDir = pkg.OriginalDir()
 	c.embedGlobals = pkg.EmbedGlobals
 	c.pkg = pkg.Pkg
+	c.loaderPkg = pkg
 	c.runtimePkg = ssaPkg.Prog.ImportedPackage("runtime").Pkg
 	c.program = ssaPkg.Prog

-	// Convert AST to SSA.
-	ssaPkg.Build()
+	// Assign names to function-local named types before compiling the
+	// package, so that types declared in different functions (or in
+	// different instantiations of a generic function) do not collide.
+	c.scanLocalTypes(ssaPkg)

 	// Initialize debug information.
 	if c.Debug {
@@ -300,9 +328,6 @@ func CompilePackage(moduleName string, pkg *loader.Package, ssaPkg *ssa.Package,
 	// Load comments such as //go:extern on globals.
 	c.loadASTComments(pkg)

-	// Predeclare the runtime.alloc function, which is used by the wordpack
-	// functionality.
-	c.getFunction(c.program.ImportedPackage("runtime").Members["alloc"].(*ssa.Function))
 	if c.NeedsStackObjects {
 		// Predeclare trackPointer, which is used everywhere we use runtime.alloc.
 		c.getFunction(c.program.ImportedPackage("runtime").Members["trackPointer"].(*ssa.Function))
@@ -388,7 +413,7 @@ func (c *compilerContext) getLLVMType(goType types.Type) llvm.Type {
 // makeLLVMType creates a LLVM type for a Go type. Don't call this, use
 // getLLVMType instead.
 func (c *compilerContext) makeLLVMType(goType types.Type) llvm.Type {
-	switch typ := goType.(type) {
+	switch typ := types.Unalias(goType).(type) {
 	case *types.Array:
 		elemType := c.getLLVMType(typ.Elem())
 		return llvm.ArrayType(elemType, int(typ.Len()))
@@ -400,19 +425,19 @@ func (c *compilerContext) makeLLVMType(goType types.Type) llvm.Type {
 			return c.ctx.Int8Type()
 		case types.Int16, types.Uint16:
 			return c.ctx.Int16Type()
-		case types.Int32, types.Uint32:
+		case types.Int32, types.Uint32, types.UntypedRune:
 			return c.ctx.Int32Type()
-		case types.Int, types.Uint:
+		case types.Int, types.Uint, types.UntypedInt:
 			return c.intType
 		case types.Int64, types.Uint64:
 			return c.ctx.Int64Type()
 		case types.Float32:
 			return c.ctx.FloatType()
-		case types.Float64:
+		case types.Float64, types.UntypedFloat:
 			return c.ctx.DoubleType()
 		case types.Complex64:
 			return c.ctx.StructType([]llvm.Type{c.ctx.FloatType(), c.ctx.FloatType()}, false)
-		case types.Complex128:
+		case types.Complex128, types.UntypedComplex:
 			return c.ctx.StructType([]llvm.Type{c.ctx.DoubleType(), c.ctx.DoubleType()}, false)
 		case types.String, types.UntypedString:
 			return c.getLLVMRuntimeType("_string")
@@ -496,6 +521,21 @@ func (c *compilerContext) createDIType(typ types.Type) llvm.Metadata {
 	llvmType := c.getLLVMType(typ)
 	sizeInBytes := c.targetData.TypeAllocSize(llvmType)
 	switch typ := typ.(type) {
+	case *types.Alias:
+		// Implement types.Alias just like types.Named: by treating them like a
+		// C typedef.
+		temporaryMDNode := c.dibuilder.CreateReplaceableCompositeType(llvm.Metadata{}, llvm.DIReplaceableCompositeType{
+			Tag:         dwarf.TagTypedef,
+			SizeInBits:  sizeInBytes * 8,
+			AlignInBits: uint32(c.targetData.ABITypeAlignment(llvmType)) * 8,
+		})
+		c.ditypes[typ] = temporaryMDNode
+		md := c.dibuilder.CreateTypedef(llvm.DITypedef{
+			Type: c.getDIType(types.Unalias(typ)), // TODO: use typ.Rhs in Go 1.23
+			Name: typ.String(),
+		})
+		temporaryMDNode.ReplaceAllUsesWith(md)
+		return md
 	case *types.Array:
 		return c.dibuilder.CreateArrayType(llvm.DIArrayType{
 			SizeInBits:  sizeInBytes * 8,
@@ -833,10 +873,9 @@ func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package
 			}
 			// Create the function definition.
 			b := newBuilder(c, irbuilder, member)
-			if _, ok := mathToLLVMMapping[member.RelString(nil)]; ok {
+			if ok := b.defineMathOp(); ok {
 				// The body of this function (if there is one) is ignored and
 				// replaced with a LLVM intrinsic call.
-				b.defineMathOp()
 				continue
 			}
 			if ok := b.defineMathBitsIntrinsic(); ok {
@@ -844,6 +883,10 @@ func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package
 				// with a LLVM intrinsic.
 				continue
 			}
+			if ok := b.defineCryptoIntrinsic(); ok {
+				// Body of this function was replaced
+				continue
+			}
 			if member.Blocks == nil {
 				// Try to define this as an intrinsic function.
 				b.defineIntrinsicFunction()
@@ -858,6 +901,11 @@ func (c *compilerContext) createPackage(irbuilder llvm.Builder, pkg *ssa.Package
 				// Interfaces don't have concrete methods.
 				continue
 			}
+			if _, isalias := member.Type().(*types.Alias); isalias {
+				// Aliases don't need to be redefined, since they just refer to
+				// an already existing type whose methods will be defined.
+				continue
+			}

 			// Named type. We should make sure all methods are created.
 			// This includes both functions with pointer receivers and those
@@ -1199,14 +1247,29 @@ func (b *builder) createFunctionStart(intrinsic bool) {
 		// intrinsic (like an atomic operation). Create the entry block
 		// manually.
 		entryBlock = b.ctx.AddBasicBlock(b.llvmFn, "entry")
-	} else {
-		for _, block := range b.fn.DomPreorder() {
-			llvmBlock := b.ctx.AddBasicBlock(b.llvmFn, block.Comment)
-			b.blockEntries[block] = llvmBlock
-			b.blockExits[block] = llvmBlock
+		// Intrinsics may create internal branches (e.g. nil checks).
+		// They will attempt to access b.currentBlockInfo to update the exit block.
+		// Create some fake block info for them to access.
+		blockInfo := []blockInfo{
+			{
+				entry: entryBlock,
+				exit:  entryBlock,
+			},
 		}
+		b.blockInfo = blockInfo
+		b.currentBlockInfo = &blockInfo[0]
+	} else {
+		blocks := b.fn.Blocks
+		blockInfo := make([]blockInfo, len(blocks))
+		for _, block := range b.fn.DomPreorder() {
+			info := &blockInfo[block.Index]
+			llvmBlock := b.ctx.AddBasicBlock(b.llvmFn, block.Comment)
+			info.entry = llvmBlock
+			info.exit = llvmBlock
+		}
+		b.blockInfo = blockInfo
 		// Normal functions have an entry block.
-		entryBlock = b.blockEntries[b.fn.Blocks[0]]
+		entryBlock = blockInfo[0].entry
 	}
 	b.SetInsertPointAtEnd(entryBlock)

@@ -1222,10 +1285,28 @@ func (b *builder) createFunctionStart(intrinsic bool) {

 	// Load function parameters
 	llvmParamIndex := 0
+	if _, indirectResult := b.hasIndirectResult(b.fn.Signature); indirectResult && !b.info.exported {
+		b.indirectReturn = b.llvmFn.Param(llvmParamIndex)
+		b.indirectReturn.SetName("return")
+		llvmParamIndex++
+	}
 	for _, param := range b.fn.Params {
 		llvmType := b.getLLVMType(param.Type())
+		if b.isIndirectParam(llvmType, b.info.exported) {
+			llvmParam := b.llvmFn.Param(llvmParamIndex)
+			llvmParam.SetName(param.Name())
+			b.indirectValues[param] = llvmParam
+			llvmParamIndex++
+			continue
+		}
+		var paramInfos []paramInfo
+		if b.info.exported {
+			paramInfos = b.expandDirectFormalParamType(llvmType, param.Name(), param.Type())
+		} else {
+			paramInfos = b.expandFormalParamType(llvmType, param.Name(), param.Type())
+		}
 		fields := make([]llvm.Value, 0, 1)
-		for _, info := range b.expandFormalParamType(llvmType, param.Name(), param.Type()) {
+		for _, info := range paramInfos {
 			param := b.llvmFn.Param(llvmParamIndex)
 			param.SetName(info.name)
 			fields = append(fields, param)
@@ -1302,8 +1383,9 @@ func (b *builder) createFunction() {
 		if b.DumpSSA {
 			fmt.Printf("%d: %s:\n", block.Index, block.Comment)
 		}
-		b.SetInsertPointAtEnd(b.blockEntries[block])
 		b.currentBlock = block
+		b.currentBlockInfo = &b.blockInfo[block.Index]
+		b.SetInsertPointAtEnd(b.currentBlockInfo.entry)
 		for _, instr := range block.Instrs {
 			if instr, ok := instr.(*ssa.DebugRef); ok {
 				if !b.Debug {
@@ -1362,8 +1444,8 @@ func (b *builder) createFunction() {
 	for _, phi := range b.phis {
 		block := phi.ssa.Block()
 		for i, edge := range phi.ssa.Edges {
-			llvmVal := b.getValue(edge, getPos(phi.ssa))
-			llvmBlock := b.blockExits[block.Preds[i]]
+			llvmVal := b.getCallArgument(edge, false)
+			llvmBlock := b.blockInfo[block.Preds[i].Index].exit
 			phi.llvm.AddIncoming([]llvm.Value{llvmVal}, []llvm.BasicBlock{llvmBlock})
 		}
 	}
@@ -1371,6 +1453,9 @@ func (b *builder) createFunction() {
 	if b.NeedsStackObjects {
 		// Track phi nodes.
 		for _, phi := range b.phis {
+			if b.isOversizedAggregate(phi.ssa.Type()) {
+				continue
+			}
 			insertPoint := llvm.NextInstruction(phi.llvm)
 			for !insertPoint.IsAPHINode().IsNil() {
 				insertPoint = llvm.NextInstruction(insertPoint)
@@ -1462,10 +1547,7 @@ func (b *builder) createInstruction(instr ssa.Instruction) {
 			b.diagnostics = append(b.diagnostics, err)
 			b.locals[instr] = llvm.Undef(b.getLLVMType(instr.Type()))
 		} else {
-			b.locals[instr] = value
-			if len(*instr.Referrers()) != 0 && b.NeedsStackObjects {
-				b.trackExpr(instr, value)
-			}
+			b.setValue(instr, value)
 		}
 	case *ssa.DebugRef:
 		// ignore
@@ -1477,18 +1559,16 @@ func (b *builder) createInstruction(instr ssa.Instruction) {
 	case *ssa.If:
 		cond := b.getValue(instr.Cond, getPos(instr))
 		block := instr.Block()
-		blockThen := b.blockEntries[block.Succs[0]]
-		blockElse := b.blockEntries[block.Succs[1]]
+		blockThen := b.blockInfo[block.Succs[0].Index].entry
+		blockElse := b.blockInfo[block.Succs[1].Index].entry
 		b.CreateCondBr(cond, blockThen, blockElse)
 	case *ssa.Jump:
-		blockJump := b.blockEntries[instr.Block().Succs[0]]
+		blockJump := b.blockInfo[instr.Block().Succs[0].Index].entry
 		b.CreateBr(blockJump)
 	case *ssa.MapUpdate:
 		m := b.getValue(instr.Map, getPos(instr))
-		key := b.getValue(instr.Key, getPos(instr))
-		value := b.getValue(instr.Value, getPos(instr))
 		mapType := instr.Map.Type().Underlying().(*types.Map)
-		b.createMapUpdate(mapType.Key(), m, key, value, instr.Pos())
+		b.createMapUpdate(mapType.Key(), m, instr.Key, instr.Value, instr.Pos())
 	case *ssa.Panic:
 		value := b.getValue(instr.X, getPos(instr))
 		b.createRuntimeInvoke("_panic", []llvm.Value{value}, "")
@@ -1497,19 +1577,7 @@ func (b *builder) createInstruction(instr ssa.Instruction) {
 		if b.hasDeferFrame() {
 			b.createRuntimeCall("destroyDeferFrame", []llvm.Value{b.deferFrame}, "")
 		}
-		if len(instr.Results) == 0 {
-			b.CreateRetVoid()
-		} else if len(instr.Results) == 1 {
-			b.CreateRet(b.getValue(instr.Results[0], getPos(instr)))
-		} else {
-			// Multiple return values. Put them all in a struct.
-			retVal := llvm.ConstNull(b.llvmFn.GlobalValueType().ReturnType())
-			for i, result := range instr.Results {
-				val := b.getValue(result, getPos(instr))
-				retVal = b.CreateInsertValue(retVal, val, i, "")
-			}
-			b.CreateRet(retVal)
-		}
+		b.createReturn(instr.Results, getPos(instr))
 	case *ssa.RunDefers:
 		// Note where we're going to put the rundefers block
 		run := b.insertBasicBlock("rundefers.block")
@@ -1523,18 +1591,246 @@ func (b *builder) createInstruction(instr ssa.Instruction) {
 		b.createChanSend(instr)
 	case *ssa.Store:
 		llvmAddr := b.getValue(instr.Addr, getPos(instr))
-		llvmVal := b.getValue(instr.Val, getPos(instr))
 		b.createNilCheck(instr.Addr, llvmAddr, "store")
-		if b.targetData.TypeAllocSize(llvmVal.Type()) == 0 {
+		llvmType := b.getLLVMType(instr.Val.Type())
+		if b.targetData.TypeAllocSize(llvmType) == 0 {
 			// nothing to store
 			return
 		}
-		b.CreateStore(llvmVal, llvmAddr)
+		b.storeValue(llvmAddr, instr.Val)
 	default:
 		b.addError(instr.Pos(), "unknown instruction: "+instr.String())
 	}
 }

+func (b *builder) setValue(value ssa.Value, llvmValue llvm.Value) {
+	if b.isAggregateValue(value.Type()) && !llvmValue.IsNil() && llvmValue.Type().TypeKind() == llvm.PointerTypeKind {
+		b.indirectValues[value] = llvmValue
+		return
+	}
+	b.locals[value] = llvmValue
+	if len(*value.Referrers()) != 0 && b.NeedsStackObjects {
+		b.trackExpr(value, llvmValue)
+	}
+}
+
+func (b *builder) createReturn(results []ssa.Value, pos token.Pos) {
+	if len(results) == 0 {
+		b.CreateRetVoid()
+	} else if !b.indirectReturn.IsNil() {
+		if len(results) == 1 {
+			b.storeValue(b.indirectReturn, results[0])
+		} else {
+			returnType := b.getLLVMResultType(b.fn.Signature)
+			for i, result := range results {
+				fieldPtr := b.CreateStructGEP(returnType, b.indirectReturn, i, "")
+				b.storeValue(fieldPtr, result)
+			}
+		}
+		b.CreateRetVoid()
+	} else if len(results) == 1 {
+		b.CreateRet(b.getValue(results[0], pos))
+	} else {
+		result := llvm.ConstNull(b.llvmFn.GlobalValueType().ReturnType())
+		for i, value := range results {
+			result = b.CreateInsertValue(result, b.getValue(value, pos), i, "")
+		}
+		b.CreateRet(result)
+	}
+}
+
+func (b *builder) isOversizedAggregate(typ types.Type) bool {
+	if !b.isAggregateValue(typ) {
+		return false
+	}
+	return b.isIndirectAggregate(b.getLLVMType(typ))
+}
+
+func (b *builder) isAggregateValue(typ types.Type) bool {
+	if tuple, ok := typ.(*types.Tuple); ok {
+		for i := 0; i < tuple.Len(); i++ {
+			if !isLLVMValueType(tuple.At(i).Type()) {
+				return false
+			}
+		}
+	} else {
+		switch typ.Underlying().(type) {
+		case *types.Array, *types.Struct:
+		default:
+			return false
+		}
+		if !isLLVMValueType(typ) {
+			return false
+		}
+	}
+	return true
+}
+
+func (b *builder) getValuePointer(value ssa.Value) llvm.Value {
+	if ptr, ok := b.indirectValues[value]; ok {
+		return ptr
+	}
+	llvmType := b.getLLVMType(value.Type())
+	ptr := b.createIndirectStorage(llvmType, value.Name())
+	b.storeValue(ptr, value)
+	return ptr
+}
+
+func (b *builder) getCallArgument(value ssa.Value, exported bool) llvm.Value {
+	paramType := b.getLLVMType(value.Type())
+	if b.isIndirectParam(paramType, exported) {
+		return b.getValuePointer(value)
+	}
+	return b.getValue(value, getPos(value))
+}
+
+func (b *builder) createIndirectStorage(typ llvm.Type, name string) llvm.Value {
+	// Use runtime.alloc here so storage that escapes remains valid. The
+	// allocation optimizer moves bounded non-escaping storage to the stack.
+	size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(typ), false)
+	layout := b.createObjectLayout(typ, b.fn.Pos())
+	ptr := b.createAlloc(size, layout, b.targetData.ABITypeAlignment(typ), name)
+	if b.NeedsStackObjects {
+		b.trackPointer(ptr)
+	}
+	return ptr
+}
+
+func (b *builder) copyIndirectAggregate(dst, src llvm.Value, typ llvm.Type) {
+	size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(typ), false)
+	b.createMemCopy("memcpy", dst, src, size)
+}
+
+func (b *builder) storeValue(dst llvm.Value, value ssa.Value) {
+	typ := b.getLLVMType(value.Type())
+	if src, ok := b.indirectValues[value]; ok {
+		b.copyIndirectAggregate(dst, src, typ)
+	} else {
+		b.CreateStore(b.getValue(value, getPos(value)), dst)
+	}
+}
+
+func (b *builder) copyToIndirectStorage(src llvm.Value, typ llvm.Type, name string) llvm.Value {
+	dst := b.createIndirectStorage(typ, name)
+	b.copyIndirectAggregate(dst, src, typ)
+	return dst
+}
+
+func (b *builder) loadFromStorage(ptr llvm.Value, typ types.Type, name string) llvm.Value {
+	llvmType := b.getLLVMType(typ)
+	if b.isIndirectAggregate(llvmType) {
+		return b.copyToIndirectStorage(ptr, llvmType, name)
+	}
+	return b.CreateLoad(llvmType, ptr, name)
+}
+
+func (b *builder) getValueField(value ssa.Value, index int, resultType types.Type, name string) (llvm.Value, bool) {
+	if !b.isAggregateValue(value.Type()) {
+		return llvm.Value{}, false
+	}
+	valueType := b.getLLVMType(value.Type())
+	if _, indirect := b.indirectValues[value]; !indirect && !b.isIndirectAggregate(valueType) {
+		return llvm.Value{}, false
+	}
+	fieldPtr := b.CreateStructGEP(valueType, b.getValuePointer(value), index, "")
+	return b.loadFromStorage(fieldPtr, resultType, name), true
+}
+
+func (b *builder) zeroIndirectStorage(ptr llvm.Value, typ llvm.Type) {
+	memset := b.getMemsetFunc()
+	b.createCall(memset.GlobalValueType(), memset, []llvm.Value{
+		ptr,
+		llvm.ConstInt(b.ctx.Int8Type(), 0, false),
+		llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(typ), false),
+		llvm.ConstInt(b.ctx.Int1Type(), 0, false),
+	}, "")
+}
+
+type valueStorage struct {
+	ptr, size llvm.Value
+	temporary bool
+}
+
+func (b *builder) getValueStorage(value ssa.Value, name string) valueStorage {
+	typ := b.getLLVMType(value.Type())
+	if b.isIndirectAggregate(typ) {
+		return valueStorage{ptr: b.getValuePointer(value)}
+	}
+	ptr, size := b.createTemporaryAlloca(typ, name)
+	b.storeValue(ptr, value)
+	return valueStorage{ptr: ptr, size: size, temporary: true}
+}
+
+func (b *builder) endValueStorage(storage valueStorage) {
+	if storage.temporary {
+		b.emitLifetimeEnd(storage.ptr, storage.size)
+	}
+}
+
+type runtimeValueResult struct {
+	valueType  llvm.Type
+	resultType llvm.Type
+	result     llvm.Value
+	valuePtr   llvm.Value
+	valueSize  llvm.Value
+	temporary  bool
+	zero       bool
+	commaOk    bool
+}
+
+func (b *builder) createRuntimeValueResult(valueType llvm.Type, commaOk, zeroAsNull bool, name string) runtimeValueResult {
+	result := runtimeValueResult{
+		valueType:  valueType,
+		resultType: valueType,
+		valueSize:  llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(valueType), false),
+		commaOk:    commaOk,
+	}
+	if commaOk {
+		result.resultType = b.ctx.StructType([]llvm.Type{valueType, b.ctx.Int1Type()}, false)
+	}
+	if b.isIndirectAggregate(result.resultType) {
+		result.result = b.createIndirectStorage(result.resultType, name+".result")
+		result.valuePtr = result.result
+		if commaOk {
+			result.valuePtr = b.CreateStructGEP(result.resultType, result.result, 0, "")
+		}
+		return result
+	}
+	if zeroAsNull && b.targetData.TypeAllocSize(valueType) == 0 {
+		result.valuePtr = llvm.ConstNull(b.dataPtrType)
+		result.zero = true
+		return result
+	}
+	result.valuePtr, result.valueSize = b.createTemporaryAlloca(valueType, name+".value")
+	result.temporary = true
+	return result
+}
+
+func (r runtimeValueResult) finish(b *builder, commaOk llvm.Value, name string) llvm.Value {
+	if !r.result.IsNil() {
+		if r.commaOk {
+			b.CreateStore(commaOk, b.CreateStructGEP(r.resultType, r.result, 1, ""))
+		}
+		return r.result
+	}
+
+	var value llvm.Value
+	if r.zero {
+		value = llvm.ConstNull(r.valueType)
+	} else {
+		value = b.CreateLoad(r.valueType, r.valuePtr, name)
+	}
+	if r.temporary {
+		b.emitLifetimeEnd(r.valuePtr, r.valueSize)
+	}
+	if !r.commaOk {
+		return value
+	}
+	result := llvm.Undef(r.resultType)
+	result = b.CreateInsertValue(result, value, 0, "")
+	return b.CreateInsertValue(result, commaOk, 1, "")
+}
+
 // createBuiltin lowers a builtin Go function (append, close, delete, etc.) to
 // LLVM IR. It uses runtime calls for some builtins.
 func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, callName string, pos token.Pos) (llvm.Value, error) {
@@ -1549,7 +1845,8 @@ func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, c
 		elemsLen := b.CreateExtractValue(elems, 1, "append.elemsLen")
 		elemType := b.getLLVMType(argTypes[0].Underlying().(*types.Slice).Elem())
 		elemSize := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(elemType), false)
-		result := b.createRuntimeCall("sliceAppend", []llvm.Value{srcBuf, elemsBuf, srcLen, srcCap, elemsLen, elemSize}, "append.new")
+		elemLayout := b.createObjectLayout(elemType, pos)
+		result := b.createRuntimeCall("sliceAppend", []llvm.Value{srcBuf, elemsBuf, srcLen, srcCap, elemsLen, elemSize, elemLayout}, "append.new")
 		newPtr := b.CreateExtractValue(result, 0, "append.newPtr")
 		newLen := b.CreateExtractValue(result, 1, "append.newLen")
 		newCap := b.CreateExtractValue(result, 2, "append.newCap")
@@ -1631,13 +1928,41 @@ func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, c
 	case "copy":
 		dst := argValues[0]
 		src := argValues[1]
+		// Fetch the lengths.
 		dstLen := b.CreateExtractValue(dst, 1, "copy.dstLen")
 		srcLen := b.CreateExtractValue(src, 1, "copy.srcLen")
-		dstBuf := b.CreateExtractValue(dst, 0, "copy.dstArray")
-		srcBuf := b.CreateExtractValue(src, 0, "copy.srcArray")
+		// Find the minimum of the lengths.
+		minFuncName := "llvm.umin.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
+		minFunc := b.mod.NamedFunction(minFuncName)
+		if minFunc.IsNil() {
+			fnType := llvm.FunctionType(b.uintptrType, []llvm.Type{b.uintptrType, b.uintptrType}, false)
+			minFunc = llvm.AddFunction(b.mod, minFuncName, fnType)
+		}
+		minLen := b.CreateCall(minFunc.GlobalValueType(), minFunc, []llvm.Value{dstLen, srcLen}, "copy.n")
+		// Multiply the length by the element size.
 		elemType := b.getLLVMType(argTypes[0].Underlying().(*types.Slice).Elem())
 		elemSize := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(elemType), false)
-		return b.createRuntimeCall("sliceCopy", []llvm.Value{dstBuf, srcBuf, dstLen, srcLen, elemSize}, "copy.n"), nil
+		// NOTE: This is also NSW when uintptr is int, but we can only choose one through the C API?
+		size := b.CreateNUWMul(minLen, elemSize, "copy.size")
+		// Fetch the pointers.
+		dstBuf := b.CreateExtractValue(dst, 0, "copy.dstPtr")
+		srcBuf := b.CreateExtractValue(src, 0, "copy.srcPtr")
+		// Create a memcpy.
+		call := b.createMemCopy("memmove", dstBuf, srcBuf, size)
+		align := b.targetData.ABITypeAlignment(elemType)
+		if align > 1 {
+			// Apply the type's alignment to the arguments.
+			// LLVM sometimes turns constant-length moves into loads and stores.
+			// It may use this alignment for the created loads and stores.
+			alignAttr := b.ctx.CreateEnumAttribute(llvm.AttributeKindID("align"), uint64(align))
+			call.AddCallSiteAttribute(1, alignAttr)
+			call.AddCallSiteAttribute(2, alignAttr)
+		}
+		// Extend and return the copied length.
+		if b.targetData.TypeAllocSize(minLen.Type()) < b.targetData.TypeAllocSize(b.intType) {
+			minLen = b.CreateZExt(minLen, b.intType, "copy.n.zext")
+		}
+		return minLen, nil
 	case "delete":
 		m := argValues[0]
 		key := argValues[1]
@@ -1665,20 +1990,66 @@ func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, c
 		return llvmLen, nil
 	case "min", "max":
 		// min and max builtins, added in Go 1.21.
-		// We can simply reuse the existing binop comparison code, which has all
-		// the edge cases figured out already.
-		tok := token.LSS
-		if callName == "max" {
-			tok = token.GTR
-		}
-		result := argValues[0]
-		typ := argTypes[0]
-		for _, arg := range argValues[1:] {
-			cmp, err := b.createBinOp(tok, typ, typ, result, arg, pos)
-			if err != nil {
-				return result, err
+		// Find the corresponding intrinsic name.
+		ty := argTypes[0].Underlying().(*types.Basic)
+		llvmType := b.getLLVMType(ty)
+		info := ty.Info()
+		var prefix, delimeter, typeName string
+		if info&types.IsInteger != 0 {
+			// This is an integer value.
+			// Use the LLVM int min/max intrinsics.
+			prefix = "llvm.s"
+			if info&types.IsUnsigned != 0 {
+				prefix = "llvm.u"
 			}
-			result = b.CreateSelect(cmp, result, arg, "")
+			delimeter = ".i"
+			typeName = strconv.Itoa(llvmType.IntTypeWidth())
+		} else {
+			switch ty.Kind() {
+			case types.String:
+				// Strings do not have an equivalent intrinsic.
+				// Implement with compares and selects.
+				tok := token.LSS
+				if callName == "max" {
+					tok = token.GTR
+				}
+				result := argValues[0]
+				typ := argTypes[0]
+				for _, arg := range argValues[1:] {
+					cmp, err := b.createBinOp(tok, typ, typ, result, arg, pos)
+					if err != nil {
+						return result, err
+					}
+					result = b.CreateSelect(cmp, result, arg, "")
+				}
+				return result, nil
+			case types.Float32:
+				typeName = "f32"
+			case types.Float64:
+				typeName = "f64"
+			default:
+				return llvm.Value{}, b.makeError(pos, "todo: min/max: unknown type")
+			}
+			// There are a few edge cases with floating point min/max:
+			// min(-0.0, +0.0) = -0.0
+			// min(NaN, number) = NaN
+			// The llvm.minimum.*/llvm.maximum.* intrinsics match this behavior.
+			// Neither Go nor LLVM defines the bit representation of resulting NaNs.
+			prefix = "llvm."
+			delimeter = "imum."
+		}
+		intrinsicName := prefix + callName + delimeter + typeName
+		// Find or create the intrinsic.
+		llvmFn := b.mod.NamedFunction(intrinsicName)
+		if llvmFn.IsNil() {
+			fnType := llvm.FunctionType(llvmType, []llvm.Type{llvmType, llvmType}, false)
+			llvmFn = llvm.AddFunction(b.mod, intrinsicName, fnType)
+		}
+		// Call the intrinsic repeatedly to merge the arguments.
+		callType := llvmFn.GlobalValueType()
+		result := argValues[0]
+		for _, arg := range argValues[1:] {
+			result = b.CreateCall(callType, llvmFn, []llvm.Value{result, arg}, "")
 		}
 		return result, nil
 	case "panic":
@@ -1759,6 +2130,12 @@ func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, c
 			// not of the current function.
 			useParentFrame = 1
 		}
+		// Prevent inlining of functions that call recover(), matching the
+		// Go compiler's behavior. If this function were inlined into a
+		// deferred function, recover() would incorrectly succeed because
+		// the inlined code runs in the deferred function's context.
+		noinline := b.ctx.CreateEnumAttribute(llvm.AttributeKindID("noinline"), 0)
+		b.llvmFn.AddFunctionAttr(noinline)
 		return b.createRuntimeCall("_recover", []llvm.Value{llvm.ConstInt(b.ctx.Int1Type(), useParentFrame, false)}, ""), nil
 	case "ssa:wrapnilchk":
 		// TODO: do an actual nil check?
@@ -1835,19 +2212,11 @@ func (b *builder) createBuiltin(argTypes []types.Type, argValues []llvm.Value, c
 //
 // This is also where compiler intrinsics are implemented.
 func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error) {
-	var params []llvm.Value
-	for _, param := range instr.Args {
-		params = append(params, b.getValue(param, getPos(instr)))
-	}
-
-	// Try to call the function directly for trivially static calls.
-	var callee, context llvm.Value
-	var calleeType llvm.Type
-	exported := false
+	// See if this is an intrinsic function that is handled specially.
 	if fn := instr.StaticCallee(); fn != nil {
 		// Direct function call, either to a named or anonymous (directly
 		// applied) function call. If it is anonymous, it may be a closure.
-		name := fn.RelString(nil)
+		name := b.getFunctionInfo(fn).linkName
 		switch {
 		case name == "device.Asm" || name == "device/arm.Asm" || name == "device/arm64.Asm" || name == "device/avr.Asm" || name == "device/riscv.Asm":
 			return b.createInlineAsm(instr.Args)
@@ -1859,10 +2228,14 @@ func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error)
 			return b.emitSV64Call(instr.Args, getPos(instr))
 		case strings.HasPrefix(name, "(device/riscv.CSR)."):
 			return b.emitCSROperation(instr)
-		case strings.HasPrefix(name, "syscall.Syscall") || strings.HasPrefix(name, "syscall.RawSyscall") || strings.HasPrefix(name, "golang.org/x/sys/unix.Syscall") || strings.HasPrefix(name, "golang.org/x/sys/unix.RawSyscall"):
+		case (strings.HasPrefix(name, "syscall.Syscall") || strings.HasPrefix(name, "syscall.RawSyscall") || strings.HasPrefix(name, "golang.org/x/sys/unix.Syscall") || strings.HasPrefix(name, "golang.org/x/sys/unix.RawSyscall")) && name != "syscall.SyscallN":
 			if b.GOOS != "darwin" {
 				return b.createSyscall(instr)
 			}
+		case name == "syscall.syscalln":
+			if b.GOOS == "windows" {
+				return b.createSyscalln(instr)
+			}
 		case strings.HasPrefix(name, "syscall.rawSyscallNoError") || strings.HasPrefix(name, "golang.org/x/sys/unix.RawSyscallNoError"):
 			return b.createRawSyscallNoError(instr)
 		case name == "runtime.supportsRecover":
@@ -1879,13 +2252,25 @@ func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error)
 			return llvm.ConstInt(b.ctx.Int8Type(), panicStrategy, false), nil
 		case name == "runtime/interrupt.New":
 			return b.createInterruptGlobal(instr)
+		case name == "runtime.exportedFuncPtr":
+			_, ptr := b.getFunction(instr.Args[0].(*ssa.Function))
+			return b.CreatePtrToInt(ptr, b.uintptrType, ""), nil
+		case name == "(*runtime/interrupt.Checkpoint).Save":
+			return b.createInterruptCheckpoint(instr.Args[0]), nil
 		case name == "internal/abi.FuncPCABI0":
 			retval := b.createDarwinFuncPCABI0Call(instr)
 			if !retval.IsNil() {
 				return retval, nil
 			}
 		}
+	}

+	// Try to call the function directly for trivially static calls.
+	var callee, context llvm.Value
+	var calleeType llvm.Type
+	var invokeTypecode, invokeReceiver llvm.Value
+	exported := false
+	if fn := instr.StaticCallee(); fn != nil {
 		calleeType, callee = b.getFunction(fn)
 		info := b.getFunctionInfo(fn)
 		if callee.IsNil() {
@@ -1913,19 +2298,18 @@ func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error)
 		exported = info.exported
 	} else if call, ok := instr.Value.(*ssa.Builtin); ok {
 		// Builtin function (append, close, delete, etc.).)
+		var params []llvm.Value
 		var argTypes []types.Type
 		for _, arg := range instr.Args {
 			argTypes = append(argTypes, arg.Type())
+			params = append(params, b.getValue(arg, getPos(instr)))
 		}
 		return b.createBuiltin(argTypes, params, call.Name(), instr.Pos())
 	} else if instr.IsInvoke() {
 		// Interface method call (aka invoke call).
 		itf := b.getValue(instr.Value, getPos(instr)) // interface value (runtime._interface)
-		typecode := b.CreateExtractValue(itf, 0, "invoke.func.typecode")
-		value := b.CreateExtractValue(itf, 1, "invoke.func.value") // receiver
-		// Prefix the params with receiver value and suffix with typecode.
-		params = append([]llvm.Value{value}, params...)
-		params = append(params, typecode)
+		invokeTypecode = b.CreateExtractValue(itf, 0, "invoke.func.typecode")
+		invokeReceiver = b.CreateExtractValue(itf, 1, "invoke.func.value")
 		callee = b.getInvokeFunction(instr)
 		calleeType = callee.GlobalValueType()
 		context = llvm.Undef(b.dataPtrType)
@@ -1939,7 +2323,23 @@ func (b *builder) createFunctionCall(instr *ssa.CallCommon) (llvm.Value, error)
 		b.createNilCheck(instr.Value, callee, "fpcall")
 	}

+	var params []llvm.Value
+	for _, param := range instr.Args {
+		params = append(params, b.getCallArgument(param, exported))
+	}
+	if instr.IsInvoke() {
+		params = append([]llvm.Value{invokeReceiver}, params...)
+		params = append(params, invokeTypecode)
+	}
+
 	if !exported {
+		if resultType, indirectResult := b.hasIndirectResult(instr.Signature()); indirectResult {
+			result := b.createIndirectStorage(resultType, "call.result")
+			params = append([]llvm.Value{result}, params...)
+			params = append(params, context)
+			b.createInvoke(calleeType, callee, params, "")
+			return result, nil
+		}
 		// This function takes a context parameter.
 		// Add it to the end of the parameter list.
 		params = append(params, context)
@@ -1978,6 +2378,9 @@ func (b *builder) getValue(expr ssa.Value, pos token.Pos) llvm.Value {
 		return value
 	default:
 		// other (local) SSA value
+		if value, ok := b.indirectValues[expr]; ok {
+			return b.CreateLoad(b.getLLVMType(expr.Type()), value, "")
+		}
 		if value, ok := b.locals[expr]; ok {
 			return value
 		} else {
@@ -2002,13 +2405,10 @@ func (c *compilerContext) maxSliceSize(elementType llvm.Type) uint64 {
 	if elementSize == 0 {
 		elementSize = 1
 	}
-	maxSize := maxPointerValue / elementSize
-
-	// len(slice) is an int. Make sure the length remains small enough to fit in
-	// an int.
-	if maxSize > maxIntegerValue {
-		maxSize = maxIntegerValue
-	}
+	maxSize := min(
+		// len(slice) is an int. Make sure the length remains small enough to fit in
+		// an int.
+		maxPointerValue/elementSize, maxIntegerValue)

 	return maxSize
 }
@@ -2035,9 +2435,8 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 			}
 			sizeValue := llvm.ConstInt(b.uintptrType, size, false)
 			layoutValue := b.createObjectLayout(typ, expr.Pos())
-			buf := b.createRuntimeCall("alloc", []llvm.Value{sizeValue, layoutValue}, expr.Comment)
 			align := b.targetData.ABITypeAlignment(typ)
-			buf.AddCallSiteAttribute(0, b.ctx.CreateEnumAttribute(llvm.AttributeKindID("align"), uint64(align)))
+			buf := b.createAlloc(sizeValue, layoutValue, align, expr.Comment)
 			return buf, nil
 		} else {
 			buf := llvmutil.CreateEntryBlockAlloca(b.Builder, typ, expr.Comment)
@@ -2064,8 +2463,15 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 		// This instruction changes the type, but the underlying value remains
 		// the same. This is often a no-op, but sometimes we have to change the
 		// LLVM type as well.
-		x := b.getValue(expr.X, getPos(expr))
 		llvmType := b.getLLVMType(expr.Type())
+		if b.isIndirectAggregate(llvmType) {
+			sourceType := b.getLLVMType(expr.X.Type())
+			if !b.isIndirectAggregate(sourceType) || b.targetData.TypeAllocSize(sourceType) != b.targetData.TypeAllocSize(llvmType) {
+				return llvm.Value{}, errors.New("todo: indirect aggregate ChangeType with different layout")
+			}
+			return b.getValuePointer(expr.X), nil
+		}
+		x := b.getValue(expr.X, getPos(expr))
 		if x.Type() == llvmType {
 			// Different Go type but same LLVM type (for example, named int).
 			// This is the common case.
@@ -2095,9 +2501,15 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 		if _, ok := expr.Tuple.(*ssa.Select); ok {
 			return b.getChanSelectResult(expr), nil
 		}
+		if value, ok := b.getValueField(expr.Tuple, expr.Index, expr.Type(), expr.Name()); ok {
+			return value, nil
+		}
 		value := b.getValue(expr.Tuple, getPos(expr))
 		return b.CreateExtractValue(value, expr.Index, ""), nil
 	case *ssa.Field:
+		if value, ok := b.getValueField(expr.X, expr.Field, expr.Type(), expr.Name()); ok {
+			return value, nil
+		}
 		value := b.getValue(expr.X, getPos(expr))
 		result := b.CreateExtractValue(value, expr.Field, "")
 		return result, nil
@@ -2120,11 +2532,11 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 	case *ssa.Global:
 		panic("global is not an expression")
 	case *ssa.Index:
-		collection := b.getValue(expr.X, getPos(expr))
 		index := b.getValue(expr.Index, getPos(expr))

 		switch xType := expr.X.Type().Underlying().(type) {
 		case *types.Basic: // extract byte from string
+			collection := b.getValue(expr.X, getPos(expr))
 			// Value type must be a string, which is a basic type.
 			if xType.Info()&types.IsString == 0 {
 				panic("lookup on non-string?")
@@ -2158,13 +2570,12 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {

 			// Can't load directly from array (as index is non-constant), so
 			// have to do it using an alloca+gep+load.
-			arrayType := collection.Type()
-			alloca, allocaSize := b.createTemporaryAlloca(arrayType, "index.alloca")
-			b.CreateStore(collection, alloca)
+			arrayType := b.getLLVMType(expr.X.Type())
+			storage := b.getValueStorage(expr.X, "index.alloca")
 			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			ptr := b.CreateInBoundsGEP(arrayType, alloca, []llvm.Value{zero, index}, "index.gep")
-			result := b.CreateLoad(arrayType.ElementType(), ptr, "index.load")
-			b.emitLifetimeEnd(alloca, allocaSize)
+			ptr := b.CreateInBoundsGEP(arrayType, storage.ptr, []llvm.Value{zero, index}, "index.gep")
+			result := b.loadFromStorage(ptr, expr.Type(), "index.load")
+			b.endValueStorage(storage)
 			return result, nil
 		default:
 			panic("unknown *ssa.Index type")
@@ -2223,17 +2634,21 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 		}
 	case *ssa.Lookup: // map lookup
 		value := b.getValue(expr.X, getPos(expr))
-		index := b.getValue(expr.Index, getPos(expr))
 		valueType := expr.Type()
 		if expr.CommaOk {
 			valueType = valueType.(*types.Tuple).At(0).Type()
 		}
-		return b.createMapLookup(expr.X.Type().Underlying().(*types.Map).Key(), valueType, value, index, expr.CommaOk, expr.Pos())
+		return b.createMapLookup(expr.X.Type().Underlying().(*types.Map).Key(), valueType, value, expr.Index, expr.CommaOk, expr.Pos())
 	case *ssa.MakeChan:
 		return b.createMakeChan(expr), nil
 	case *ssa.MakeClosure:
 		return b.parseMakeClosure(expr)
 	case *ssa.MakeInterface:
+		if b.isOversizedAggregate(expr.X.Type()) {
+			typ := b.getLLVMType(expr.X.Type())
+			ptr := b.copyToIndirectStorage(b.getValuePointer(expr.X), typ, "interface.value")
+			return b.createMakeInterfaceFromPointer(ptr, expr.X.Type()), nil
+		}
 		val := b.getValue(expr.X, getPos(expr))
 		return b.createMakeInterface(val, expr.X.Type(), expr.Pos()), nil
 	case *ssa.MakeMap:
@@ -2267,8 +2682,7 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 		}
 		sliceSize := b.CreateBinOp(llvm.Mul, elemSizeValue, sliceCapCast, "makeslice.cap")
 		layoutValue := b.createObjectLayout(llvmElemType, expr.Pos())
-		slicePtr := b.createRuntimeCall("alloc", []llvm.Value{sliceSize, layoutValue}, "makeslice.buf")
-		slicePtr.AddCallSiteAttribute(0, b.ctx.CreateEnumAttribute(llvm.AttributeKindID("align"), uint64(elemAlign)))
+		slicePtr := b.createAlloc(sliceSize, layoutValue, elemAlign, "makeslice.buf")

 		// Extend or truncate if necessary. This is safe as we've already done
 		// the bounds check.
@@ -2301,7 +2715,8 @@ func (b *builder) createExpr(expr ssa.Value) (llvm.Value, error) {
 			return b.createMapIteratorNext(rangeVal, llvmRangeVal, it), nil
 		}
 	case *ssa.Phi:
-		phi := b.CreatePHI(b.getLLVMType(expr.Type()), "")
+		phiType := b.storedParamType(b.getLLVMType(expr.Type()), false)
+		phi := b.CreatePHI(phiType, "")
 		b.phis = append(b.phis, phiNode{expr, phi})
 		return phi, nil
 	case *ssa.Range:
@@ -3304,8 +3719,7 @@ func (b *builder) createUnOp(unop *ssa.UnOp) (llvm.Value, error) {
 			return fn, nil
 		} else {
 			b.createNilCheck(unop.X, x, "deref")
-			load := b.CreateLoad(valueType, x, "")
-			return load, nil
+			return b.loadFromStorage(x, unop.Type(), ""), nil
 		}
 	case token.XOR: // ^x, toggle all bits in integer
 		return b.CreateXor(x, llvm.ConstInt(x.Type(), ^uint64(0), false), ""), nil
@@ -4,6 +4,7 @@ import (
 	"flag"
 	"go/types"
 	"os"
+	"regexp"
 	"strconv"
 	"strings"
 	"testing"
@@ -50,6 +51,8 @@ func TestCompiler(t *testing.T) {
 		{"channel.go", "", ""},
 		{"gc.go", "", ""},
 		{"zeromap.go", "", ""},
+		{"generics.go", "", ""},
+		{"large.go", "", ""},
 	}
 	if goMinor >= 20 {
 		tests = append(tests, testCase{"go1.20.go", "", ""})
@@ -57,6 +60,9 @@ func TestCompiler(t *testing.T) {
 	if goMinor >= 21 {
 		tests = append(tests, testCase{"go1.21.go", "", ""})
 	}
+	if goMinor >= 27 {
+		tests = append(tests, testCase{"go1.27.go", "", ""})
+	}

 	for _, tc := range tests {
 		name := tc.file
@@ -109,7 +115,7 @@ func TestCompiler(t *testing.T) {

 			// Update test if needed. Do not check the result.
 			if *flagUpdate {
-				err := os.WriteFile(outPath, []byte(mod.String()), 0666)
+				err := os.WriteFile(outPath, []byte(normalizeIR(mod.String())), 0666)
 				if err != nil {
 					t.Error("failed to write updated output file:", err)
 				}
@@ -121,30 +127,137 @@ func TestCompiler(t *testing.T) {
 				t.Fatal("failed to read golden file:", err)
 			}

-			if !fuzzyEqualIR(mod.String(), string(expected)) {
-				t.Errorf("output does not match expected output:\n%s", mod.String())
+			if diff := diffIR(string(expected), mod.String()); diff != "" {
+				t.Errorf("output does not match expected output (re-run with -update to regenerate):\n%s", diff)
 			}
 		})
 	}
 }

-// fuzzyEqualIR returns true if the two LLVM IR strings passed in are roughly
-// equal. That means, only relevant lines are compared (excluding comments
-// etc.).
-func fuzzyEqualIR(s1, s2 string) bool {
-	lines1 := filterIrrelevantIRLines(strings.Split(s1, "\n"))
-	lines2 := filterIrrelevantIRLines(strings.Split(s2, "\n"))
-	if len(lines1) != len(lines2) {
-		return false
-	}
-	for i, line1 := range lines1 {
-		line2 := lines2[i]
-		if line1 != line2 {
-			return false
+func TestOptimizedLargeAggregateABI(t *testing.T) {
+	options := &compileopts.Options{Target: "wasm"}
+	mod, errs := testCompilePackage(t, options, "large-optimized.go")
+	if len(errs) != 0 {
+		for _, err := range errs {
+			t.Error(err)
 		}
+		return
+	}
+	defer mod.Dispose()
+
+	passOptions := llvm.NewPassBuilderOptions()
+	defer passOptions.Dispose()
+	if err := mod.RunPasses("default<O2>", llvm.TargetMachine{}, passOptions); err != nil {
+		t.Fatal(err)
 	}

-	return true
+	resultFn := mod.NamedFunction("main.makeLargeOptimizedValue")
+	if resultFn.IsNil() {
+		t.Fatal("missing function main.makeLargeOptimizedValue")
+	}
+	if resultType := resultFn.GlobalValueType().ReturnType(); resultType.TypeKind() != llvm.VoidTypeKind {
+		t.Errorf("large aggregate result was promoted to %s", resultType)
+	}
+
+	for _, name := range []string{
+		"main.makeLargeOptimizedValue",
+		"main.readLargeOptimizedValue",
+		"main.readMixedLargeOptimizedValue",
+	} {
+		fn := mod.NamedFunction(name)
+		if fn.IsNil() {
+			t.Fatalf("missing function %s", name)
+		}
+		if paramType := fn.GlobalValueType().ParamTypes()[0]; paramType.TypeKind() != llvm.PointerTypeKind {
+			t.Errorf("%s aggregate parameter was promoted to %s", name, paramType)
+		}
+	}
+}
+
+// normalizeIR canonicalizes LLVM-version-specific IR spellings for comparison
+// and when regenerating golden files.
+func normalizeIR(s string) string {
+	// Golden files are written using the pre-LLVM21 'nocapture' spelling,
+	// which LLVM printed before any co-occurring attribute such as
+	// 'readonly' (e.g. "ptr nocapture readonly"). LLVM 21+ prints the
+	// equivalent 'captures(none)' instead, and after such attributes (e.g.
+	// "ptr readonly captures(none)"). Normalize both name and position back
+	// to the old spelling.
+	s = normalizeCapturesAttr(s)
+
+	// LLVM 21+ also added an explicit 'nocreateundeforpoison' attribute to
+	// certain intrinsic declarations (e.g. llvm.umin) that were implicitly
+	// assumed not to create undef/poison before. It's unrelated to the
+	// behavior under test, so ignore it for comparison.
+	s = strings.ReplaceAll(s, "nocreateundeforpoison ", "")
+
+	// LLVM 22 dropped the (redundant) i64 size argument from
+	// llvm.lifetime.start/end. Normalize away that argument so golden files
+	// written against the two-argument form still match.
+	s = lifetimeSizeArgRe.ReplaceAllString(s, "$1")
+
+	return s
+}
+
+// diffIR compares two LLVM IR strings, ignoring irrelevant lines (comments,
+// empty lines, etc.) and normalizing LLVM-version-specific spellings via
+// normalizeIR. It returns "" when they are equal. Otherwise it returns a
+// compact diff of only the region that differs: the common prefix and suffix
+// are trimmed, then the differing expected lines (prefixed "-") are shown
+// followed by the differing actual lines (prefixed "+").
+func diffIR(expected, actual string) string {
+	exp := filterIrrelevantIRLines(strings.Split(normalizeIR(expected), "\n"))
+	act := filterIrrelevantIRLines(strings.Split(normalizeIR(actual), "\n"))
+
+	// Trim the common prefix.
+	start := 0
+	for start < len(exp) && start < len(act) && exp[start] == act[start] {
+		start++
+	}
+	// Trim the common suffix.
+	e, a := len(exp), len(act)
+	for e > start && a > start && exp[e-1] == act[a-1] {
+		e--
+		a--
+	}
+	if start == e && start == a {
+		return "" // equal
+	}
+
+	var b strings.Builder
+	b.WriteString("first difference at relevant line ")
+	b.WriteString(strconv.Itoa(start + 1))
+	b.WriteString(":\n")
+	for _, line := range exp[start:e] {
+		b.WriteString("- ")
+		b.WriteString(line)
+		b.WriteByte('\n')
+	}
+	for _, line := range act[start:a] {
+		b.WriteString("+ ")
+		b.WriteString(line)
+		b.WriteByte('\n')
+	}
+	return b.String()
+}
+
+// capturesNoneAttrRe matches a co-occurring attribute directly followed by
+// 'captures(none)', which is how LLVM 21+ orders these two attributes when
+// printing IR (the pre-LLVM21 'nocapture' attribute printed the other way
+// around).
+var capturesNoneAttrRe = regexp.MustCompile(`\b(readonly|readnone|writeonly|nonnull)\s+captures\(none\)`)
+
+// lifetimeSizeArgRe matches the i64 size argument of an
+// llvm.lifetime.start/end call or declaration, which LLVM 22 removed.
+var lifetimeSizeArgRe = regexp.MustCompile(`(@llvm\.lifetime\.(?:start|end)\.p0\()i64(?: immarg| \d+), `)
+
+// normalizeCapturesAttr rewrites LLVM 21+'s 'captures(none)' attribute back
+// to the pre-LLVM21 'nocapture' spelling and position, so golden IR files
+// written against LLVM <21 keep matching.
+func normalizeCapturesAttr(s string) string {
+	s = capturesNoneAttrRe.ReplaceAllString(s, "nocapture $1")
+	s = strings.ReplaceAll(s, "captures(none)", "nocapture")
+	return s
 }

 // filterIrrelevantIRLines removes lines from the input slice of strings that
@@ -187,9 +300,9 @@ func TestCompilerErrors(t *testing.T) {
 		t.Error(err)
 	}
 	errorsFileString := strings.ReplaceAll(string(errorsFile), "\r\n", "\n")
-	for _, line := range strings.Split(errorsFileString, "\n") {
-		if strings.HasPrefix(line, "// ERROR: ") {
-			expectedErrors = append(expectedErrors, strings.TrimPrefix(line, "// ERROR: "))
+	for line := range strings.SplitSeq(errorsFileString, "\n") {
+		if after, ok := strings.CutPrefix(line, "// ERROR: "); ok {
+			expectedErrors = append(expectedErrors, after)
 		}
 	}

@@ -213,6 +326,49 @@ func TestCompilerErrors(t *testing.T) {
 	}
 }

+func TestAggregateValueCount(t *testing.T) {
+	t.Parallel()
+
+	ctx := llvm.NewContext()
+	defer ctx.Dispose()
+
+	byteType := ctx.Int8Type()
+	tests := []struct {
+		name     string
+		typ      llvm.Type
+		count    uint64
+		exceeded bool
+	}{
+		{"empty", llvm.ArrayType(byteType, 0), 0, false},
+		{"limit", llvm.ArrayType(byteType, 1024), 1024, false},
+		{"over limit", llvm.ArrayType(byteType, 1025), 0, true},
+		{"combined limit", ctx.StructType([]llvm.Type{
+			llvm.ArrayType(byteType, 512),
+			llvm.ArrayType(byteType, 512),
+		}, false), 1024, false},
+		{"combined over limit", ctx.StructType([]llvm.Type{
+			llvm.ArrayType(byteType, 1000),
+			llvm.ArrayType(byteType, 1000),
+		}, false), 0, true},
+		{"comma-ok over limit", ctx.StructType([]llvm.Type{
+			llvm.ArrayType(byteType, 1024),
+			ctx.Int1Type(),
+		}, false), 0, true},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			count, exceeded := aggregateValueCount(test.typ, 0)
+			if exceeded != test.exceeded {
+				t.Errorf("expected exceeded=%t, got %t", test.exceeded, exceeded)
+			}
+			if !exceeded && count != test.count {
+				t.Errorf("expected count=%d, got %d", test.count, count)
+			}
+		})
+	}
+}
+
 // Build a package given a number of compiler options and a file.
 func testCompilePackage(t *testing.T, options *compileopts.Options, file string) (llvm.Module, []error) {
 	target, err := compileopts.LoadTarget(options)
@@ -257,5 +413,7 @@ func testCompilePackage(t *testing.T, options *compileopts.Options, file string)
 	// Compile AST to IR.
 	program := lprogram.LoadSSA()
 	pkg := lprogram.MainPkg()
-	return CompilePackage(file, pkg, program.Package(pkg.Pkg), machine, compilerConfig, false)
+	ssaPkg := program.Package(pkg.Pkg)
+	ssaPkg.Build()
+	return CompilePackage(file, pkg, ssaPkg, machine, compilerConfig, false)
 }
@@ -32,7 +32,7 @@ func (b *builder) supportsRecover() bool {
 		// proposal of WebAssembly:
 		// https://github.com/WebAssembly/exception-handling
 		return false
-	case "riscv64", "xtensa":
+	case "xtensa":
 		// TODO: add support for these architectures
 		return false
 	default:
@@ -60,10 +60,6 @@ func (b *builder) deferInitFunc() {
 	b.deferExprFuncs = make(map[ssa.Value]int)
 	b.deferBuiltinFuncs = make(map[ssa.Value]deferBuiltin)

-	// Create defer list pointer.
-	b.deferPtr = b.CreateAlloca(b.dataPtrType, "deferPtr")
-	b.CreateStore(llvm.ConstPointerNull(b.dataPtrType), b.deferPtr)
-
 	if b.hasDeferFrame() {
 		// Set up the defer frame with the current stack pointer.
 		// This assumes that the stack pointer doesn't move outside of the
@@ -73,12 +69,22 @@ func (b *builder) deferInitFunc() {
 		// in the setjmp-like inline assembly.
 		deferFrameType := b.getLLVMRuntimeType("deferFrame")
 		b.deferFrame = b.CreateAlloca(deferFrameType, "deferframe.buf")
+		// The field index must match the DeferPtr field in runtime.deferFrame,
+		// defined in src/runtime/panic.go.
+		b.deferPtr = b.CreateInBoundsGEP(deferFrameType, b.deferFrame, []llvm.Value{
+			llvm.ConstInt(b.ctx.Int32Type(), 0, false),
+			llvm.ConstInt(b.ctx.Int32Type(), 6, false), // DeferPtr field
+		}, "deferPtr")
 		stackPointer := b.readStackPointer()
 		b.createRuntimeCall("setupDeferFrame", []llvm.Value{b.deferFrame, stackPointer}, "")

 		// Create the landing pad block, which is where control transfers after
 		// a panic.
 		b.landingpad = b.ctx.AddBasicBlock(b.llvmFn, "lpad")
+	} else {
+		// Create defer list pointer.
+		b.deferPtr = b.CreateAlloca(b.dataPtrType, "deferPtr")
+		b.CreateStore(llvm.ConstPointerNull(b.dataPtrType), b.deferPtr)
 	}
 }

@@ -100,17 +106,18 @@ func (b *builder) createLandingPad() {

 	// Continue at the 'recover' block, which returns to the parent in an
 	// appropriate way.
-	b.CreateBr(b.blockEntries[b.fn.Recover])
+	b.CreateBr(b.blockInfo[b.fn.Recover.Index].entry)
 }

-// createInvokeCheckpoint saves the function state at the given point, to
-// continue at the landing pad if a panic happened. This is implemented using a
-// setjmp-like construct.
-func (b *builder) createInvokeCheckpoint() {
+// Create a checkpoint (similar to setjmp). This emits inline assembly that
+// stores the current program counter inside the ptr address (actually
+// ptr+sizeof(ptr)) and then returns a boolean indicating whether this is the
+// normal flow (false) or we jumped here from somewhere else (true).
+func (b *builder) createCheckpoint(ptr llvm.Value) llvm.Value {
 	// Construct inline assembly equivalents of setjmp.
 	// The assembly works as follows:
-	//   * All registers (both callee-saved and caller saved) are clobbered
-	//     after the inline assembly returns.
+	//   * Registers are either clobbered or, on 386, saved for longjmp to
+	//     restore if the ABI requires them to survive calls.
 	//   * The assembly stores the address just past the end of the assembly
 	//     into the jump buffer.
 	//   * The return value (eax, rax, r0, etc) is set to zero in the inline
@@ -123,8 +130,12 @@ func (b *builder) createInvokeCheckpoint() {
 		asmString = `
 xorl %eax, %eax
 movl $$1f, 4(%ebx)
+movl %ebx, 8(%ebx)
+movl %esi, 12(%ebx)
+movl %edi, 16(%ebx)
+movl %ebp, 20(%ebx)
 1:`
-		constraints = "={eax},{ebx},~{ebx},~{ecx},~{edx},~{esi},~{edi},~{ebp},~{xmm0},~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{fpsr},~{fpcr},~{flags},~{dirflag},~{memory}"
+		constraints = "={eax},{ebx},~{ecx},~{edx},~{xmm0},~{xmm1},~{xmm2},~{xmm3},~{xmm4},~{xmm5},~{xmm6},~{xmm7},~{fpsr},~{fpcr},~{flags},~{dirflag},~{memory}"
 		// This doesn't include the floating point stack because TinyGo uses
 		// newer floating point instructions.
 	case "x86_64":
@@ -206,60 +217,192 @@ sw $$ra, 4($$5)
 			// So only add them when using hardfloat.
 			constraints += ",~{$f0},~{$f1},~{$f2},~{$f3},~{$f4},~{$f5},~{$f6},~{$f7},~{$f8},~{$f9},~{$f10},~{$f11},~{$f12},~{$f13},~{$f14},~{$f15},~{$f16},~{$f17},~{$f18},~{$f19},~{$f20},~{$f21},~{$f22},~{$f23},~{$f24},~{$f25},~{$f26},~{$f27},~{$f28},~{$f29},~{$f30},~{$f31}"
 		}
-	case "riscv32":
-		asmString = `
+	case "riscv32", "riscv64":
+		if b.archFamily() == "riscv32" {
+			asmString = `
 la a2, 1f
 sw a2, 4(a1)
 li a0, 0
 1:`
+		} else {
+			asmString = `
+la a2, 1f
+sd a2, 8(a1)
+li a0, 0
+1:`
+		}
 		constraints = "={a0},{a1},~{a1},~{a2},~{a3},~{a4},~{a5},~{a6},~{a7},~{s0},~{s1},~{s2},~{s3},~{s4},~{s5},~{s6},~{s7},~{s8},~{s9},~{s10},~{s11},~{t0},~{t1},~{t2},~{t3},~{t4},~{t5},~{t6},~{ra},~{f0},~{f1},~{f2},~{f3},~{f4},~{f5},~{f6},~{f7},~{f8},~{f9},~{f10},~{f11},~{f12},~{f13},~{f14},~{f15},~{f16},~{f17},~{f18},~{f19},~{f20},~{f21},~{f22},~{f23},~{f24},~{f25},~{f26},~{f27},~{f28},~{f29},~{f30},~{f31},~{memory}"
 	default:
 		// This case should have been handled by b.supportsRecover().
 		b.addError(b.fn.Pos(), "unknown architecture for defer: "+b.archFamily())
 	}
-	asmType := llvm.FunctionType(resultType, []llvm.Type{b.deferFrame.Type()}, false)
+	asmType := llvm.FunctionType(resultType, []llvm.Type{b.dataPtrType}, false)
 	asm := llvm.InlineAsm(asmType, asmString, constraints, false, false, 0, false)
-	result := b.CreateCall(asmType, asm, []llvm.Value{b.deferFrame}, "setjmp")
+	result := b.CreateCall(asmType, asm, []llvm.Value{ptr}, "setjmp")
 	result.AddCallSiteAttribute(-1, b.ctx.CreateEnumAttribute(llvm.AttributeKindID("returns_twice"), 0))
 	isZero := b.CreateICmp(llvm.IntEQ, result, llvm.ConstInt(resultType, 0, false), "setjmp.result")
+	return isZero
+}
+
+// createInvokeCheckpoint saves the function state at the given point, to
+// continue at the landing pad if a panic happened. This is implemented using a
+// setjmp-like construct.
+func (b *builder) createInvokeCheckpoint() {
+	isZero := b.createCheckpoint(b.deferFrame)
 	continueBB := b.insertBasicBlock("")
 	b.CreateCondBr(isZero, continueBB, b.landingpad)
 	b.SetInsertPointAtEnd(continueBB)
-	b.blockExits[b.currentBlock] = continueBB
+	b.currentBlockInfo.exit = continueBB
 }

-// isInLoop checks if there is a path from a basic block to itself.
-func isInLoop(start *ssa.BasicBlock) bool {
-	// Use a breadth-first search to scan backwards through the block graph.
-	queue := []*ssa.BasicBlock{start}
-	checked := map[*ssa.BasicBlock]struct{}{}
+// createFaultCheckpoint is like createInvokeCheckpoint but for use in fault
+// blocks (e.g., bounds check failures). Unlike createInvokeCheckpoint, it does
+// not update currentBlockInfo.exit because the fault block is a dead-end that
+// does not participate in phi node resolution.
+func (b *builder) createFaultCheckpoint() {
+	isZero := b.createCheckpoint(b.deferFrame)
+	continueBB := b.insertBasicBlock("")
+	b.CreateCondBr(isZero, continueBB, b.landingpad)
+	b.SetInsertPointAtEnd(continueBB)
+}

-	for len(queue) > 0 {
-		// pop a block off of the queue
-		block := queue[len(queue)-1]
-		queue = queue[:len(queue)-1]
+// isInLoop checks if there is a path from the current block to itself.
+// Use Tarjan's strongly connected components algorithm to search for cycles.
+// A one-node SCC is a cycle iff there is an edge from the node to itself.
+// A multi-node SCC is always a cycle.
+// https://en.wikipedia.org/wiki/Tarjan%27s_strongly_connected_components_algorithm
+func (b *builder) isInLoop() bool {
+	if b.currentBlockInfo.tarjan.lowLink == 0 {
+		b.strongConnect(b.currentBlock)
+	}
+	return b.currentBlockInfo.tarjan.cyclic
+}

-		// Search through predecessors.
-		// Searching backwards means that this is pretty fast when the block is close to the start of the function.
-		// Defers are often placed near the start of the function.
-		for _, pred := range block.Preds {
-			if pred == start {
-				// cycle found
-				return true
-			}
+func (b *builder) strongConnect(block *ssa.BasicBlock) {
+	// Assign a new index.
+	// Indices start from 1 so that 0 can be used as a sentinel.
+	assignedIndex := b.tarjanIndex + 1
+	b.tarjanIndex = assignedIndex

-			if _, ok := checked[pred]; ok {
-				// block already checked
-				continue
-			}
+	// Apply the new index.
+	blockIndex := block.Index
+	node := &b.blockInfo[blockIndex].tarjan
+	node.lowLink = assignedIndex

-			// add to queue and checked map
-			queue = append(queue, pred)
-			checked[pred] = struct{}{}
+	// Push the node onto the stack.
+	node.onStack = true
+	b.tarjanStack = append(b.tarjanStack, uint(blockIndex))
+
+	// Process the successors.
+	for _, successor := range block.Succs {
+		// Look up the successor's state.
+		successorIndex := successor.Index
+		if successorIndex == blockIndex {
+			// Handle a self-cycle specially.
+			node.cyclic = true
+			continue
+		}
+		successorNode := &b.blockInfo[successorIndex].tarjan
+
+		switch {
+		case successorNode.lowLink == 0:
+			// This node has not yet been visisted.
+			b.strongConnect(successor)
+
+		case !successorNode.onStack:
+			// This node has been visited, but is in a different SCC.
+			// Ignore it, and do not update lowLink.
+			continue
+		}
+
+		// Update the lowLink index.
+		// This always uses the min-of-lowlink instead of using index in the on-stack case.
+		// This is done for two reasons:
+		// 1. The lowLink update can be shared between the new-node and on-stack cases.
+		// 2. The assigned index does not need to be saved - it is only needed for root node detection.
+		if successorNode.lowLink < node.lowLink {
+			node.lowLink = successorNode.lowLink
 		}
 	}

-	return false
+	if node.lowLink == assignedIndex {
+		// This is a root node.
+		// Pop the SCC off the stack.
+		stack := b.tarjanStack
+		top := stack[len(stack)-1]
+		stack = stack[:len(stack)-1]
+		blocks := b.blockInfo
+		topNode := &blocks[top].tarjan
+		topNode.onStack = false
+
+		if top != uint(blockIndex) {
+			// The root node is not the only node in the SCC.
+			// Mark all nodes in this SCC as cyclic.
+			topNode.cyclic = true
+			for top != uint(blockIndex) {
+				top = stack[len(stack)-1]
+				stack = stack[:len(stack)-1]
+				topNode = &blocks[top].tarjan
+				topNode.onStack = false
+				topNode.cyclic = true
+			}
+		}
+
+		b.tarjanStack = stack
+	}
+}
+
+// tarjanNode holds per-block state for isInLoop and strongConnect.
+type tarjanNode struct {
+	// lowLink is the index of the first visited node that is reachable from this block.
+	// The lowLink indices are assigned by the SCC search, and do not correspond to b.Index.
+	// A lowLink of 0 is used as a sentinel to mark a node which has not yet been visited.
+	lowLink uint
+
+	// onStack tracks whether this node is currently on the SCC search stack.
+	onStack bool
+
+	// cyclic indicates whether this block is in a loop.
+	// If lowLink is 0, strongConnect must be called before reading this field.
+	cyclic bool
+}
+
+type llvmValueList struct {
+	values []llvm.Value
+	types  []llvm.Type
+}
+
+func newLLVMValueList(values ...llvm.Value) llvmValueList {
+	var list llvmValueList
+	list.append(values...)
+	return list
+}
+
+func (l *llvmValueList) append(values ...llvm.Value) {
+	for _, value := range values {
+		l.values = append(l.values, value)
+		l.types = append(l.types, value.Type())
+	}
+}
+
+func (l *llvmValueList) appendSSAValues(values []ssa.Value, lower func(ssa.Value) llvm.Value) {
+	for _, value := range values {
+		l.append(lower(value))
+	}
+}
+
+func (b *builder) loadDeferredCallParams(structType llvm.Type, ptr llvm.Value) []llvm.Value {
+	fieldTypes := structType.StructElementTypes()
+	values := make([]llvm.Value, 0, len(fieldTypes)-2)
+	zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+	for i := 2; i < len(fieldTypes); i++ {
+		fieldPtr := b.CreateInBoundsGEP(structType, ptr, []llvm.Value{
+			zero,
+			llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false),
+		}, "gep")
+		values = append(values, b.CreateLoad(fieldTypes[i], fieldPtr, "param"))
+	}
+	return values
 }

 // createDefer emits a single defer instruction, to be run when this function
@@ -269,31 +412,28 @@ func (b *builder) createDefer(instr *ssa.Defer) {
 	// make a linked list.
 	next := b.CreateLoad(b.dataPtrType, b.deferPtr, "defer.next")

-	var values []llvm.Value
-	valueTypes := []llvm.Type{b.uintptrType, next.Type()}
+	var values llvmValueList
+	lowerArgument := func(value ssa.Value) llvm.Value {
+		return b.getCallArgument(value, false)
+	}
 	if instr.Call.IsInvoke() {
 		// Method call on an interface.

 		// Get callback type number.
-		methodName := instr.Call.Method.FullName()
-		if _, ok := b.deferInvokeFuncs[methodName]; !ok {
-			b.deferInvokeFuncs[methodName] = len(b.allDeferFuncs)
+		key := b.getInvokeFunctionName(&instr.Call)
+		if _, ok := b.deferInvokeFuncs[key]; !ok {
+			b.deferInvokeFuncs[key] = len(b.allDeferFuncs)
 			b.allDeferFuncs = append(b.allDeferFuncs, &instr.Call)
 		}
-		callback := llvm.ConstInt(b.uintptrType, uint64(b.deferInvokeFuncs[methodName]), false)
+		callback := llvm.ConstInt(b.uintptrType, uint64(b.deferInvokeFuncs[key]), false)

 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields, followed by the call parameters).
 		itf := b.getValue(instr.Call.Value, getPos(instr)) // interface
 		typecode := b.CreateExtractValue(itf, 0, "invoke.func.typecode")
 		receiverValue := b.CreateExtractValue(itf, 1, "invoke.func.receiver")
-		values = []llvm.Value{callback, next, typecode, receiverValue}
-		valueTypes = append(valueTypes, b.dataPtrType, b.dataPtrType)
-		for _, arg := range instr.Call.Args {
-			val := b.getValue(arg, getPos(instr))
-			values = append(values, val)
-			valueTypes = append(valueTypes, val.Type())
-		}
+		values = newLLVMValueList(callback, next, typecode, receiverValue)
+		values.appendSSAValues(instr.Call.Args, lowerArgument)

 	} else if callee, ok := instr.Call.Value.(*ssa.Function); ok {
 		// Regular function call.
@@ -305,12 +445,11 @@ func (b *builder) createDefer(instr *ssa.Defer) {

 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields).
-		values = []llvm.Value{callback, next}
-		for _, param := range instr.Call.Args {
-			llvmParam := b.getValue(param, getPos(instr))
-			values = append(values, llvmParam)
-			valueTypes = append(valueTypes, llvmParam.Type())
-		}
+		values = newLLVMValueList(callback, next)
+		exported := b.getFunctionInfo(callee).exported
+		values.appendSSAValues(instr.Call.Args, func(value ssa.Value) llvm.Value {
+			return b.getCallArgument(value, exported)
+		})

 	} else if makeClosure, ok := instr.Call.Value.(*ssa.MakeClosure); ok {
 		// Immediately applied function literal with free variables.
@@ -333,14 +472,9 @@ func (b *builder) createDefer(instr *ssa.Defer) {
 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields, followed by all parameters including the
 		// context pointer).
-		values = []llvm.Value{callback, next}
-		for _, param := range instr.Call.Args {
-			llvmParam := b.getValue(param, getPos(instr))
-			values = append(values, llvmParam)
-			valueTypes = append(valueTypes, llvmParam.Type())
-		}
-		values = append(values, context)
-		valueTypes = append(valueTypes, context.Type())
+		values = newLLVMValueList(callback, next)
+		values.appendSSAValues(instr.Call.Args, lowerArgument)
+		values.append(context)

 	} else if builtin, ok := instr.Call.Value.(*ssa.Builtin); ok {
 		var argTypes []types.Type
@@ -363,11 +497,8 @@ func (b *builder) createDefer(instr *ssa.Defer) {

 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields).
-		values = []llvm.Value{callback, next}
-		for _, param := range argValues {
-			values = append(values, param)
-			valueTypes = append(valueTypes, param.Type())
-		}
+		values = newLLVMValueList(callback, next)
+		values.append(argValues...)

 	} else {
 		funcValue := b.getValue(instr.Call.Value, getPos(instr))
@@ -382,34 +513,33 @@ func (b *builder) createDefer(instr *ssa.Defer) {
 		// Collect all values to be put in the struct (starting with
 		// runtime._defer fields, followed by all parameters including the
 		// context pointer).
-		values = []llvm.Value{callback, next, funcValue}
-		valueTypes = append(valueTypes, funcValue.Type())
-		for _, param := range instr.Call.Args {
-			llvmParam := b.getValue(param, getPos(instr))
-			values = append(values, llvmParam)
-			valueTypes = append(valueTypes, llvmParam.Type())
-		}
+		values = newLLVMValueList(callback, next, funcValue)
+		values.appendSSAValues(instr.Call.Args, lowerArgument)
 	}

 	// Make a struct out of the collected values to put in the deferred call
 	// struct.
-	deferredCallType := b.ctx.StructType(valueTypes, false)
+	deferredCallType := b.ctx.StructType(values.types, false)
 	deferredCall := llvm.ConstNull(deferredCallType)
-	for i, value := range values {
+	for i, value := range values.values {
 		deferredCall = b.CreateInsertValue(deferredCall, value, i, "")
 	}

 	// Put this struct in an allocation.
 	var alloca llvm.Value
-	if !isInLoop(instr.Block()) {
+	if instr.Block() != b.currentBlock {
+		panic("block mismatch")
+	}
+	if !b.isInLoop() {
 		// This can safely use a stack allocation.
 		alloca = llvmutil.CreateEntryBlockAlloca(b.Builder, deferredCallType, "defer.alloca")
 	} else {
 		// This may be hit a variable number of times, so use a heap allocation.
 		size := b.targetData.TypeAllocSize(deferredCallType)
 		sizeValue := llvm.ConstInt(b.uintptrType, size, false)
-		nilPtr := llvm.ConstNull(b.dataPtrType)
-		alloca = b.createRuntimeCall("alloc", []llvm.Value{sizeValue, nilPtr}, "defer.alloc.call")
+		layoutValue := b.createObjectLayout(deferredCallType, instr.Pos())
+		align := b.targetData.ABITypeAlignment(deferredCallType)
+		alloca = b.createAlloc(sizeValue, layoutValue, align, "defer.alloc.call")
 	}
 	if b.NeedsStackObjects {
 		b.trackPointer(alloca)
@@ -493,19 +623,11 @@ func (b *builder) createRunDefers() {
 				valueTypes = append(valueTypes, b.dataPtrType, b.dataPtrType)
 			}

-			for _, arg := range callback.Args {
-				valueTypes = append(valueTypes, b.getLLVMType(arg.Type()))
-			}
+			valueTypes = b.appendStoredValueTypes(valueTypes, callback.Args, false)

 			// Extract the params from the struct (including receiver).
-			forwardParams := []llvm.Value{}
-			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
 			deferredCallType := b.ctx.StructType(valueTypes, false)
-			for i := 2; i < len(valueTypes); i++ {
-				gep := b.CreateInBoundsGEP(deferredCallType, deferData, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false)}, "gep")
-				forwardParam := b.CreateLoad(valueTypes[i], gep, "param")
-				forwardParams = append(forwardParams, forwardParam)
-			}
+			forwardParams := b.loadDeferredCallParams(deferredCallType, deferData)

 			var fnPtr llvm.Value
 			var fnType llvm.Type
@@ -534,6 +656,7 @@ func (b *builder) createRunDefers() {
 				// with a strict calling convention.
 				forwardParams = append(forwardParams, llvm.Undef(b.dataPtrType))
 			}
+			forwardParams = b.prependIndirectResult(callback.Signature(), false, forwardParams, "defer.result")

 			b.createCall(fnType, fnPtr, forwardParams, "")

@@ -542,27 +665,21 @@ func (b *builder) createRunDefers() {

 			// Get the real defer struct type and cast to it.
 			valueTypes := []llvm.Type{b.uintptrType, b.dataPtrType}
-			for _, param := range getParams(callback.Signature) {
-				valueTypes = append(valueTypes, b.getLLVMType(param.Type()))
-			}
+			exported := b.getFunctionInfo(callback).exported
+			valueTypes = b.appendStoredParamTypes(valueTypes, getParams(callback.Signature), exported)
 			deferredCallType := b.ctx.StructType(valueTypes, false)

 			// Extract the params from the struct.
-			forwardParams := []llvm.Value{}
-			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			for i := range getParams(callback.Signature) {
-				gep := b.CreateInBoundsGEP(deferredCallType, deferData, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i+2), false)}, "gep")
-				forwardParam := b.CreateLoad(valueTypes[i+2], gep, "param")
-				forwardParams = append(forwardParams, forwardParam)
-			}
+			forwardParams := b.loadDeferredCallParams(deferredCallType, deferData)

 			// Plain TinyGo functions add some extra parameters to implement async functionality and function receivers.
 			// These parameters should not be supplied when calling into an external C/ASM function.
-			if !b.getFunctionInfo(callback).exported {
+			if !exported {
 				// Add the context parameter. We know it is ignored by the receiving
 				// function, but we have to pass one anyway.
 				forwardParams = append(forwardParams, llvm.Undef(b.dataPtrType))
 			}
+			forwardParams = b.prependIndirectResult(callback.Signature, exported, forwardParams, "defer.result")

 			// Call real function.
 			fnType, fn := b.getFunction(callback)
@@ -572,24 +689,16 @@ func (b *builder) createRunDefers() {
 			// Get the real defer struct type and cast to it.
 			fn := callback.Fn.(*ssa.Function)
 			valueTypes := []llvm.Type{b.uintptrType, b.dataPtrType}
-			params := fn.Signature.Params()
-			for i := 0; i < params.Len(); i++ {
-				valueTypes = append(valueTypes, b.getLLVMType(params.At(i).Type()))
-			}
+			valueTypes = b.appendStoredParamTypes(valueTypes, getParams(fn.Signature), false)
 			valueTypes = append(valueTypes, b.dataPtrType) // closure
 			deferredCallType := b.ctx.StructType(valueTypes, false)

 			// Extract the params from the struct.
-			forwardParams := []llvm.Value{}
-			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			for i := 2; i < len(valueTypes); i++ {
-				gep := b.CreateInBoundsGEP(deferredCallType, deferData, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false)}, "")
-				forwardParam := b.CreateLoad(valueTypes[i], gep, "param")
-				forwardParams = append(forwardParams, forwardParam)
-			}
+			forwardParams := b.loadDeferredCallParams(deferredCallType, deferData)

 			// Call deferred function.
 			fnType, llvmFn := b.getFunction(fn)
+			forwardParams = b.prependIndirectResult(fn.Signature, false, forwardParams, "defer.result")
 			b.createCall(fnType, llvmFn, forwardParams, "")
 		case *ssa.Builtin:
 			db := b.deferBuiltinFuncs[callback]
@@ -599,20 +708,14 @@ func (b *builder) createRunDefers() {

 			//Get signature from call results
 			params := callback.Type().Underlying().(*types.Signature).Params()
-			for i := 0; i < params.Len(); i++ {
-				valueTypes = append(valueTypes, b.getLLVMType(params.At(i).Type()))
+			for v := range params.Variables() {
+				valueTypes = append(valueTypes, b.getLLVMType(v.Type()))
 			}

 			deferredCallType := b.ctx.StructType(valueTypes, false)

 			// Extract the params from the struct.
-			var argValues []llvm.Value
-			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
-			for i := 0; i < params.Len(); i++ {
-				gep := b.CreateInBoundsGEP(deferredCallType, deferData, []llvm.Value{zero, llvm.ConstInt(b.ctx.Int32Type(), uint64(i+2), false)}, "gep")
-				forwardParam := b.CreateLoad(valueTypes[i+2], gep, "param")
-				argValues = append(argValues, forwardParam)
-			}
+			argValues := b.loadDeferredCallParams(deferredCallType, deferData)

 			_, err := b.createBuiltin(db.argTypes, argValues, db.callName, db.pos)
 			if err != nil {
@@ -10,6 +10,91 @@ import (
 	"tinygo.org/x/go-llvm"
 )

+// LLVM recursively expands each struct field and array element in parameters
+// and results into separate values. It gets very slow with too many values, so
+// pass larger aggregates indirectly before LLVM expands them.
+const maxDirectAggregateValues = 1024
+
+func (c *compilerContext) getLLVMResultType(sig *types.Signature) llvm.Type {
+	switch sig.Results().Len() {
+	case 0:
+		return c.ctx.VoidType()
+	case 1:
+		return c.getLLVMType(sig.Results().At(0).Type())
+	default:
+		results := make([]llvm.Type, sig.Results().Len())
+		for i := range results {
+			results[i] = c.getLLVMType(sig.Results().At(i).Type())
+		}
+		return c.ctx.StructType(results, false)
+	}
+}
+
+func (c *compilerContext) hasIndirectResult(sig *types.Signature) (llvm.Type, bool) {
+	resultType := c.getLLVMResultType(sig)
+	return resultType, c.isIndirectAggregate(resultType)
+}
+
+func (c *compilerContext) isIndirectAggregate(typ llvm.Type) bool {
+	switch typ.TypeKind() {
+	case llvm.ArrayTypeKind, llvm.StructTypeKind:
+		_, exceeded := aggregateValueCount(typ, 0)
+		return exceeded
+	default:
+		return false
+	}
+}
+
+func aggregateValueCount(typ llvm.Type, count uint64) (uint64, bool) {
+	switch typ.TypeKind() {
+	case llvm.ArrayTypeKind:
+		length := uint64(typ.ArrayLength())
+		if length == 0 {
+			return count, false
+		}
+		elementCount, exceeded := aggregateValueCount(typ.ElementType(), 0)
+		if exceeded {
+			return count, true
+		}
+		if elementCount != 0 && length > (maxDirectAggregateValues-count)/elementCount {
+			return count, true
+		}
+		return count + length*elementCount, false
+	case llvm.StructTypeKind:
+		for _, field := range typ.StructElementTypes() {
+			var exceeded bool
+			count, exceeded = aggregateValueCount(field, count)
+			if exceeded {
+				return count, true
+			}
+		}
+		return count, false
+	default:
+		count++
+		return count, count > maxDirectAggregateValues
+	}
+}
+
+func isLLVMValueType(typ types.Type) bool {
+	switch typ := typ.Underlying().(type) {
+	case *types.Basic:
+		return typ.Kind() != types.Invalid
+	case *types.Array:
+		return isLLVMValueType(typ.Elem())
+	case *types.Struct:
+		for field := range typ.Fields() {
+			if !isLLVMValueType(field.Type()) {
+				return false
+			}
+		}
+		return true
+	case *types.Chan, *types.Interface, *types.Map, *types.Pointer, *types.Signature, *types.Slice:
+		return true
+	default:
+		return false
+	}
+}
+
 // createFuncValue creates a function value from a raw function pointer with no
 // context.
 func (b *builder) createFuncValue(funcPtr, context llvm.Value, sig *types.Signature) llvm.Value {
@@ -48,28 +133,18 @@ func (c *compilerContext) getFuncType(typ *types.Signature) llvm.Type {

 // getLLVMFunctionType returns a LLVM function type for a given signature.
 func (c *compilerContext) getLLVMFunctionType(typ *types.Signature) llvm.Type {
-	// Get the return type.
-	var returnType llvm.Type
-	switch typ.Results().Len() {
-	case 0:
-		// No return values.
-		returnType = c.ctx.VoidType()
-	case 1:
-		// Just one return value.
-		returnType = c.getLLVMType(typ.Results().At(0).Type())
-	default:
-		// Multiple return values. Put them together in a struct.
-		// This appears to be the common way to handle multiple return values in
-		// LLVM.
-		members := make([]llvm.Type, typ.Results().Len())
-		for i := 0; i < typ.Results().Len(); i++ {
-			members[i] = c.getLLVMType(typ.Results().At(i).Type())
-		}
-		returnType = c.ctx.StructType(members, false)
-	}
+	returnType, indirectResult := c.hasIndirectResult(typ)

 	// Get the parameter types.
 	var paramTypes []llvm.Type
+	if indirectResult {
+		// LLVM expands aggregate returns into scalar leaves before deciding
+		// whether to pass them indirectly, so a large IR return can exhaust
+		// memory. Returning void avoids that expansion and cannot be demoted
+		// again. Keep the result pointer first so the context remains last.
+		paramTypes = append(paramTypes, c.dataPtrType)
+		returnType = c.ctx.VoidType()
+	}
 	if typ.Recv() != nil {
 		recv := c.getLLVMType(typ.Recv().Type())
 		if recv.StructName() == "runtime._interface" {
@@ -81,8 +156,8 @@ func (c *compilerContext) getLLVMFunctionType(typ *types.Signature) llvm.Type {
 			paramTypes = append(paramTypes, info.llvmType)
 		}
 	}
-	for i := 0; i < typ.Params().Len(); i++ {
-		subType := c.getLLVMType(typ.Params().At(i).Type())
+	for v := range typ.Params().Variables() {
+		subType := c.getLLVMType(v.Type())
 		for _, info := range c.expandFormalParamType(subType, "", nil) {
 			paramTypes = append(paramTypes, info.llvmType)
 		}
@@ -112,7 +187,7 @@ func (b *builder) parseMakeClosure(expr *ssa.MakeClosure) (llvm.Value, error) {

 	// Store the bound variables in a single object, allocating it on the heap
 	// if necessary.
-	context := b.emitPointerPack(boundVars)
+	context := b.emitPointerPack(boundVars, expr.Pos())

 	// Create the closure.
 	_, fn := b.getFunction(f)
@@ -5,11 +5,38 @@ package compiler

 import (
 	"go/token"
+	"slices"

 	"golang.org/x/tools/go/ssa"
 	"tinygo.org/x/go-llvm"
 )

+// Heap-allocate a buffer of the given size. This will typically call
+// runtime.alloc.
+func (b *builder) createAlloc(sizeValue, layoutValue llvm.Value, align int, comment string) llvm.Value {
+	// Normally allocate using "runtime.alloc", but use "runtime.alloc_noheap"
+	// if the //go:noheap pragma is used.
+	allocFunc := "alloc"
+	if b.info.noheap {
+		allocFunc = "alloc_noheap"
+	}
+
+	// Allocs that don't allocate anything can return an architecture-specific
+	// sentinel value.
+	if !sizeValue.IsAConstantInt().IsNil() && sizeValue.ZExtValue() == 0 {
+		allocFunc = "alloc_zero"
+	}
+
+	// Make the runtime call.
+	call := b.createRuntimeCall(allocFunc, []llvm.Value{sizeValue, layoutValue}, comment)
+	if align == 0 || align&(align-1) != 0 {
+		panic("invalid alignment")
+	}
+	call.AddCallSiteAttribute(0, b.ctx.CreateEnumAttribute(llvm.AttributeKindID("align"), uint64(align)))
+
+	return call
+}
+
 // trackExpr inserts pointer tracking intrinsics for the GC if the expression is
 // one of the expressions that need this.
 func (b *builder) trackExpr(expr ssa.Value, value llvm.Value) {
@@ -62,7 +89,7 @@ func (b *builder) trackValue(value llvm.Value) {
 			return
 		}
 		numElements := typ.StructElementTypesCount()
-		for i := 0; i < numElements; i++ {
+		for i := range numElements {
 			subValue := b.CreateExtractValue(value, i, "")
 			b.trackValue(subValue)
 		}
@@ -71,7 +98,7 @@ func (b *builder) trackValue(value llvm.Value) {
 			return
 		}
 		numElements := typ.ArrayLength()
-		for i := 0; i < numElements; i++ {
+		for i := range numElements {
 			subValue := b.CreateExtractValue(value, i, "")
 			b.trackValue(subValue)
 		}
@@ -92,13 +119,11 @@ func typeHasPointers(t llvm.Type) bool {
 	case llvm.PointerTypeKind:
 		return true
 	case llvm.StructTypeKind:
-		for _, subType := range t.StructElementTypes() {
-			if typeHasPointers(subType) {
-				return true
-			}
-		}
-		return false
+		return slices.ContainsFunc(t.StructElementTypes(), typeHasPointers)
 	case llvm.ArrayTypeKind:
+		if t.ArrayLength() == 0 {
+			return false
+		}
 		if typeHasPointers(t.ElementType()) {
 			return true
 		}
@@ -47,20 +47,16 @@ func (b *builder) createGo(instr *ssa.Go) {
 		return
 	}

-	// Get all function parameters to pass to the goroutine.
 	var params []llvm.Value
-	for _, param := range instr.Call.Args {
-		params = append(params, b.expandFormalParam(b.getValue(param, getPos(instr)))...)
-	}
-
 	var prefix string
 	var funcPtr llvm.Value
 	var funcType llvm.Type
+	var context llvm.Value
 	hasContext := false
+	exported := false
 	if callee := instr.Call.StaticCallee(); callee != nil {
 		// Static callee is known. This makes it easier to start a new
 		// goroutine.
-		var context llvm.Value
 		switch value := instr.Call.Value.(type) {
 		case *ssa.Function:
 			// Goroutine call is regular function call. No context is necessary.
@@ -73,10 +69,10 @@ func (b *builder) createGo(instr *ssa.Go) {
 			panic("StaticCallee returned an unexpected value")
 		}
 		if !context.IsNil() {
-			params = append(params, context) // context parameter
 			hasContext = true
 		}
 		funcType, funcPtr = b.getFunction(callee)
+		exported = b.getFunctionInfo(callee).exported
 	} else if instr.Call.IsInvoke() {
 		// This is a method call on an interface value.
 		itf := b.getValue(instr.Call.Value, getPos(instr))
@@ -84,23 +80,32 @@ func (b *builder) createGo(instr *ssa.Go) {
 		itfValue := b.CreateExtractValue(itf, 1, "")
 		funcPtr = b.getInvokeFunction(&instr.Call)
 		funcType = funcPtr.GlobalValueType()
-		params = append([]llvm.Value{itfValue}, params...) // start with receiver
-		params = append(params, itfTypeCode)               // end with typecode
+		params = append(params, itfValue)
+		context = itfTypeCode
 	} else {
 		// This is a function pointer.
 		// At the moment, two extra params are passed to the newly started
 		// goroutine:
 		//   * The function context, for closures.
 		//   * The function pointer (for tasks).
-		var context llvm.Value
 		funcPtr, context = b.decodeFuncValue(b.getValue(instr.Call.Value, getPos(instr)))
 		funcType = b.getLLVMFunctionType(instr.Call.Value.Type().Underlying().(*types.Signature))
-		params = append(params, context, funcPtr)
 		hasContext = true
-		prefix = b.fn.RelString(nil)
+		prefix = b.getFunctionInfo(b.fn).linkName
 	}

-	paramBundle := b.emitPointerPack(params)
+	for _, param := range instr.Call.Args {
+		params = append(params, b.getGoroutineCallArgument(param, exported)...)
+	}
+	if !context.IsNil() {
+		params = append(params, context)
+	}
+	if hasContext && instr.Call.StaticCallee() == nil {
+		params = append(params, funcPtr)
+	}
+	params = b.prependIndirectResult(instr.Call.Signature(), exported, params, "go.result")
+
+	paramBundle := b.emitPointerPack(params, instr.Pos())
 	var stackSize llvm.Value
 	callee := b.createGoroutineStartWrapper(funcType, funcPtr, prefix, hasContext, false, instr.Pos())
 	if b.AutomaticStackSize {
@@ -122,6 +127,15 @@ func (b *builder) createGo(instr *ssa.Go) {
 	b.createCall(fnType, start, []llvm.Value{callee, paramBundle, stackSize, llvm.Undef(b.dataPtrType)}, "")
 }

+func (b *builder) getGoroutineCallArgument(value ssa.Value, exported bool) []llvm.Value {
+	typ := b.getLLVMType(value.Type())
+	arg := b.getCallArgument(value, exported)
+	if b.isIndirectParam(typ, exported) {
+		return []llvm.Value{b.copyToIndirectStorage(arg, typ, "go.param")}
+	}
+	return b.expandFormalParam(arg)
+}
+
 // Create an exported wrapper function for functions with the //go:wasmexport
 // pragma. This wrapper function is quite complex when the scheduler is enabled:
 // it needs to start a new goroutine each time the exported function is called.
@@ -139,7 +153,7 @@ func (b *builder) createWasmExport() {
 	// Declare the exported function.
 	paramTypes := b.llvmFnType.ParamTypes()
 	exportedFnType := llvm.FunctionType(b.llvmFnType.ReturnType(), paramTypes[:len(paramTypes)-1], false)
-	exportedFn := llvm.AddFunction(b.mod, b.fn.RelString(nil)+suffix, exportedFnType)
+	exportedFn := llvm.AddFunction(b.mod, b.getFunctionInfo(b.fn).linkName+suffix, exportedFnType)
 	b.addStandardAttributes(exportedFn)
 	llvmutil.AppendToGlobal(b.mod, "llvm.used", exportedFn)
 	exportedFn.AddFunctionAttr(b.ctx.CreateStringAttribute("wasm-export-name", b.info.wasmExport))
@@ -295,10 +309,10 @@ func (c *compilerContext) createGoroutineStartWrapper(fnType llvm.Type, fn llvm.
 	}
 	defer b.Dispose()

-	var deadlock llvm.Value
-	var deadlockType llvm.Type
+	var exitGoroutine llvm.Value
+	var exitGoroutineType llvm.Type
 	if c.Scheduler == "asyncify" {
-		deadlockType, deadlock = c.getFunction(c.program.ImportedPackage("runtime").Members["deadlock"].(*ssa.Function))
+		exitGoroutineType, exitGoroutine = c.getFunction(c.program.ImportedPackage("runtime").Members["exitGoroutine"].(*ssa.Function))
 	}

 	if !fn.IsAFunction().IsNil() {
@@ -363,7 +377,7 @@ func (c *compilerContext) createGoroutineStartWrapper(fnType llvm.Type, fn llvm.
 			b.CreateCall(fnType, fn, params, "")

 			if c.Scheduler == "asyncify" {
-				b.CreateCall(deadlockType, deadlock, []llvm.Value{
+				b.CreateCall(exitGoroutineType, exitGoroutine, []llvm.Value{
 					llvm.Undef(c.dataPtrType),
 				}, "")
 			}
@@ -414,7 +428,7 @@ func (c *compilerContext) createGoroutineStartWrapper(fnType llvm.Type, fn llvm.
 			// Extract parameters from the state object, and call the function
 			// that's being wrapped.
 			var callParams []llvm.Value
-			for i := 0; i < numParams; i++ {
+			for i := range numParams {
 				gep := b.CreateInBoundsGEP(stateStruct, statePtr, []llvm.Value{
 					llvm.ConstInt(b.ctx.Int32Type(), 0, false),
 					llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false),
@@ -514,14 +528,13 @@ func (c *compilerContext) createGoroutineStartWrapper(fnType llvm.Type, fn llvm.
 		b.CreateCall(fnType, fnPtr, params, "")

 		if c.Scheduler == "asyncify" {
-			b.CreateCall(deadlockType, deadlock, []llvm.Value{
+			b.CreateCall(exitGoroutineType, exitGoroutine, []llvm.Value{
 				llvm.Undef(c.dataPtrType),
 			}, "")
 		}
 	}

 	if c.Scheduler == "asyncify" {
-		// The goroutine was terminated via deadlock.
 		b.CreateUnreachable()
 	} else {
 		// Finish the function. Every basic block must end in a terminator, and
@@ -38,10 +38,10 @@ func (b *builder) createInlineAsm(args []ssa.Value) (llvm.Value, error) {
 // provided immediately. For example:
 //
 //	arm.AsmFull(
-//	    "str {value}, {result}",
+//	    "str {value}, [{result}]",
 //	    map[string]interface{}{
-//	        "value":  1
-//	        "result": &dest,
+//	        "value":  1,
+//	        "result": uintptr(unsafe.Pointer(&dest)),
 //	    })
 func (b *builder) createInlineAsmFull(instr *ssa.CallCommon) (llvm.Value, error) {
 	asmString := constant.StringVal(instr.Args[0].(*ssa.Const).Value)
@@ -146,13 +146,14 @@ func (b *builder) emitSVCall(args []ssa.Value, pos token.Pos) (llvm.Value, error
 	llvmArgs := []llvm.Value{}
 	argTypes := []llvm.Type{}
 	asm := "svc #" + strconv.FormatUint(num, 10)
-	constraints := "={r0}"
+	var constraints strings.Builder
+	constraints.WriteString("={r0}")
 	for i, arg := range args[1:] {
 		arg = arg.(*ssa.MakeInterface).X
 		if i == 0 {
-			constraints += ",0"
+			constraints.WriteString(",0")
 		} else {
-			constraints += ",{r" + strconv.Itoa(i) + "}"
+			constraints.WriteString(",{r" + strconv.Itoa(i) + "}")
 		}
 		llvmValue := b.getValue(arg, pos)
 		llvmArgs = append(llvmArgs, llvmValue)
@@ -161,9 +162,9 @@ func (b *builder) emitSVCall(args []ssa.Value, pos token.Pos) (llvm.Value, error
 	// Implement the ARM calling convention by marking r1-r3 as
 	// clobbered. r0 is used as an output register so doesn't have to be
 	// marked as clobbered.
-	constraints += ",~{r1},~{r2},~{r3}"
+	constraints.WriteString(",~{r1},~{r2},~{r3}")
 	fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
-	target := llvm.InlineAsm(fnType, asm, constraints, true, false, 0, false)
+	target := llvm.InlineAsm(fnType, asm, constraints.String(), true, false, 0, false)
 	return b.CreateCall(fnType, target, llvmArgs, ""), nil
 }

@@ -184,13 +185,14 @@ func (b *builder) emitSV64Call(args []ssa.Value, pos token.Pos) (llvm.Value, err
 	llvmArgs := []llvm.Value{}
 	argTypes := []llvm.Type{}
 	asm := "svc #" + strconv.FormatUint(num, 10)
-	constraints := "={x0}"
+	var constraints strings.Builder
+	constraints.WriteString("={x0}")
 	for i, arg := range args[1:] {
 		arg = arg.(*ssa.MakeInterface).X
 		if i == 0 {
-			constraints += ",0"
+			constraints.WriteString(",0")
 		} else {
-			constraints += ",{x" + strconv.Itoa(i) + "}"
+			constraints.WriteString(",{x" + strconv.Itoa(i) + "}")
 		}
 		llvmValue := b.getValue(arg, pos)
 		llvmArgs = append(llvmArgs, llvmValue)
@@ -199,9 +201,9 @@ func (b *builder) emitSV64Call(args []ssa.Value, pos token.Pos) (llvm.Value, err
 	// Implement the ARM64 calling convention by marking x1-x7 as
 	// clobbered. x0 is used as an output register so doesn't have to be
 	// marked as clobbered.
-	constraints += ",~{x1},~{x2},~{x3},~{x4},~{x5},~{x6},~{x7}"
+	constraints.WriteString(",~{x1},~{x2},~{x3},~{x4},~{x5},~{x6},~{x7}")
 	fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
-	target := llvm.InlineAsm(fnType, asm, constraints, true, false, 0, false)
+	target := llvm.InlineAsm(fnType, asm, constraints.String(), true, false, 0, false)
 	return b.CreateCall(fnType, target, llvmArgs, ""), nil
 }

@@ -249,3 +251,15 @@ func (b *builder) emitCSROperation(call *ssa.CallCommon) (llvm.Value, error) {
 		return llvm.Value{}, b.makeError(call.Pos(), "unknown CSR operation: "+name)
 	}
 }
+
+// Implement runtime/interrupt.Checkpoint.Save. It needs to be implemented
+// directly at the call site. If it isn't implemented directly at the call site
+// (but instead through a function call), it might result in an overwritten
+// stack in the non-jump return case.
+func (b *builder) createInterruptCheckpoint(ptr ssa.Value) llvm.Value {
+	addr := b.getValue(ptr, ptr.Pos())
+	b.createNilCheck(ptr, addr, "deref")
+	stackPointer := b.readStackPointer()
+	b.CreateStore(stackPointer, addr)
+	return b.createCheckpoint(addr)
+}
@@ -7,6 +7,7 @@ import (
 	"strconv"
 	"strings"

+	"github.com/tinygo-org/tinygo/compiler/llvmutil"
 	"tinygo.org/x/go-llvm"
 )

@@ -27,6 +28,8 @@ func (b *builder) defineIntrinsicFunction() {
 		b.createStackSaveImpl()
 	case name == "runtime.KeepAlive":
 		b.createKeepAliveImpl()
+	case name == "machine.keepAliveNoEscape":
+		b.createMachineKeepAliveImpl()
 	case strings.HasPrefix(name, "runtime/volatile.Load"):
 		b.createVolatileLoad()
 	case strings.HasPrefix(name, "runtime/volatile.Store"):
@@ -48,19 +51,24 @@ func (b *builder) defineIntrinsicFunction() {
 // and will otherwise be lowered to regular libc memcpy/memmove calls.
 func (b *builder) createMemoryCopyImpl() {
 	b.createFunctionStart(true)
-	fnName := "llvm." + b.fn.Name() + ".p0.p0.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
+	params := b.fn.Params[0:3]
+	b.createMemCopy(
+		b.fn.Name(),
+		b.getValue(params[0], getPos(b.fn)),
+		b.getValue(params[1], getPos(b.fn)),
+		b.getValue(params[2], getPos(b.fn)),
+	)
+	b.CreateRetVoid()
+}
+
+func (b *builder) createMemCopy(kind string, dst, src, len llvm.Value) llvm.Value {
+	fnName := "llvm." + kind + ".p0.p0.i" + strconv.Itoa(b.uintptrType.IntTypeWidth())
 	llvmFn := b.mod.NamedFunction(fnName)
 	if llvmFn.IsNil() {
 		fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.dataPtrType, b.dataPtrType, b.uintptrType, b.ctx.Int1Type()}, false)
 		llvmFn = llvm.AddFunction(b.mod, fnName, fnType)
 	}
-	var params []llvm.Value
-	for _, param := range b.fn.Params {
-		params = append(params, b.getValue(param, getPos(b.fn)))
-	}
-	params = append(params, llvm.ConstInt(b.ctx.Int1Type(), 0, false))
-	b.CreateCall(llvmFn.GlobalValueType(), llvmFn, params, "")
-	b.CreateRetVoid()
+	return b.CreateCall(llvmFn.GlobalValueType(), llvmFn, []llvm.Value{dst, src, len, llvm.ConstInt(b.ctx.Int1Type(), 0, false)}, "")
 }

 // createMemoryZeroImpl creates calls to llvm.memset.* to zero a block of
@@ -121,13 +129,60 @@ func (b *builder) createKeepAliveImpl() {
 	b.CreateRetVoid()
 }

+// createAbiEscapeImpl implements the generic internal/abi.Escape function. It
+// currently only supports pointer types.
+func (b *builder) createAbiEscapeImpl() {
+	b.createFunctionStart(true)
+
+	// The first parameter is assumed to be a pointer. This is checked at the
+	// call site of createAbiEscapeImpl.
+	pointerValue := b.getValue(b.fn.Params[0], getPos(b.fn))
+
+	// Create an equivalent of the following C code, which is basically just a
+	// nop but ensures the pointerValue is kept alive:
+	//
+	//     __asm__ __volatile__("" : : "r"(pointerValue))
+	//
+	// It should be portable to basically everything as the "r" register type
+	// exists basically everywhere.
+	asmType := llvm.FunctionType(b.dataPtrType, []llvm.Type{b.dataPtrType}, false)
+	asmFn := llvm.InlineAsm(asmType, "", "=r,0", true, false, 0, false)
+	result := b.createCall(asmType, asmFn, []llvm.Value{pointerValue}, "")
+
+	b.CreateRet(result)
+}
+
+// Implement machine.keepAliveNoEscape, which makes sure the compiler keeps the
+// pointer parameter alive until this point (for GC).
+func (b *builder) createMachineKeepAliveImpl() {
+	b.createFunctionStart(true)
+	pointerValue := b.getValue(b.fn.Params[0], getPos(b.fn))
+
+	// See createKeepAliveImpl for details.
+	asmType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.dataPtrType}, false)
+	asmFn := llvm.InlineAsm(asmType, "", "r", true, false, 0, false)
+	b.createCall(asmType, asmFn, []llvm.Value{pointerValue}, "")
+
+	b.CreateRetVoid()
+}
+
 var mathToLLVMMapping = map[string]string{
+	"math.Acos":  "llvm.acos.f64",
+	"math.Asin":  "llvm.asin.f64",
+	"math.Atan":  "llvm.atan.f64",
+	"math.Atan2": "llvm.atan2.f64",
 	"math.Ceil":  "llvm.ceil.f64",
+	"math.Cos":   "llvm.cos.f64",
+	"math.Cosh":  "llvm.cosh.f64",
 	"math.Exp":   "llvm.exp.f64",
 	"math.Exp2":  "llvm.exp2.f64",
 	"math.Floor": "llvm.floor.f64",
 	"math.Log":   "llvm.log.f64",
+	"math.Sin":   "llvm.sin.f64",
+	"math.Sinh":  "llvm.sinh.f64",
 	"math.Sqrt":  "llvm.sqrt.f64",
+	"math.Tan":   "llvm.tan.f64",
+	"math.Tanh":  "llvm.tanh.f64",
 	"math.Trunc": "llvm.trunc.f64",
 }

@@ -141,18 +196,40 @@ var mathToLLVMMapping = map[string]string{
 // float32(math.Sqrt(float64(v))) to a 32-bit floating point operation, which is
 // beneficial on architectures where 64-bit floating point operations are (much)
 // more expensive than 32-bit ones.
-func (b *builder) defineMathOp() {
-	b.createFunctionStart(true)
-	llvmName := mathToLLVMMapping[b.fn.RelString(nil)]
-	if llvmName == "" {
-		panic("unreachable: unknown math operation") // sanity check
+func (b *builder) defineMathOp() bool {
+	llvmName, ok := mathToLLVMMapping[b.fn.RelString(nil)]
+	if !ok {
+		return false
 	}
+	if strings.HasSuffix(b.Triple, "-wasi") || llvmutil.Version() < 19 {
+		// We don't have a real libc for wasip2. Until that is fixed, we need to
+		// limit math intrinsics on WASI to a subset supported natively in
+		// WebAssembly.
+		// Also, since we don't know the specific libc we will target, disallow
+		// these for all WASI targets.
+		//
+		// We also need to limit ourselves to LLVM 19 and above for the extended
+		// set of math intrinsics, see:
+		// https://discourse.llvm.org/t/rfc-all-the-math-intrinsics/78294
+		switch b.fn.Name() {
+		case "Ceil", "Exp", "Exp2", "Floor", "Log", "Sqrt", "Trunc":
+		default:
+			return false
+		}
+	}
+	b.createFunctionStart(true)
 	llvmFn := b.mod.NamedFunction(llvmName)
 	if llvmFn.IsNil() {
 		// The intrinsic doesn't exist yet, so declare it.
-		// At the moment, all supported intrinsics have the form "double
-		// foo(double %x)" so we can hardcode the signature here.
-		llvmType := llvm.FunctionType(b.ctx.DoubleType(), []llvm.Type{b.ctx.DoubleType()}, false)
+		var llvmType llvm.Type
+		switch b.fn.Name() {
+		case "Atan2":
+			// double atan2(double %y, double %x)
+			llvmType = llvm.FunctionType(b.ctx.DoubleType(), []llvm.Type{b.ctx.DoubleType(), b.ctx.DoubleType()}, false)
+		default:
+			// double foo(double %x)
+			llvmType = llvm.FunctionType(b.ctx.DoubleType(), []llvm.Type{b.ctx.DoubleType()}, false)
+		}
 		llvmFn = llvm.AddFunction(b.mod, llvmName, llvmType)
 	}
 	// Create a call to the intrinsic.
@@ -162,6 +239,24 @@ func (b *builder) defineMathOp() {
 	}
 	result := b.CreateCall(llvmFn.GlobalValueType(), llvmFn, args, "")
 	b.CreateRet(result)
+	return true
+}
+
+func (b *builder) defineCryptoIntrinsic() bool {
+	if b.fn.Pkg.Pkg.Path() != "crypto/internal/constanttime" {
+		return false
+	}
+
+	switch b.fn.Name() {
+	case "boolToUint8":
+		b.createFunctionStart(true)
+		param := b.getValue(b.fn.Params[0], b.fn.Pos())
+		result := b.CreateZExt(param, b.ctx.Int8Type(), "")
+		b.CreateRet(result)
+		return true
+	}
+
+	return false
 }

 // Implement most math/bits functions.
@@ -1,10 +1,10 @@
 package compiler

 import (
+	"encoding/binary"
 	"fmt"
 	"go/token"
 	"go/types"
-	"math/big"
 	"strings"

 	"github.com/tinygo-org/tinygo/compileopts"
@@ -54,7 +54,7 @@ func (b *builder) emitLifetimeEnd(ptr, size llvm.Value) {
 // pointer value directly. It returns the pointer with the packed data.
 // If the values are all constants, they are be stored in a constant global and
 // deduplicated.
-func (b *builder) emitPointerPack(values []llvm.Value) llvm.Value {
+func (b *builder) emitPointerPack(values []llvm.Value, pos token.Pos) llvm.Value {
 	valueTypes := make([]llvm.Type, len(values))
 	for i, value := range values {
 		valueTypes[i] = value.Type()
@@ -128,14 +128,9 @@ func (b *builder) emitPointerPack(values []llvm.Value) llvm.Value {

 		// Packed data is bigger than a pointer, so allocate it on the heap.
 		sizeValue := llvm.ConstInt(b.uintptrType, size, false)
+		layoutValue := b.createObjectLayout(packedType, pos)
 		align := b.targetData.ABITypeAlignment(packedType)
-		alloc := b.mod.NamedFunction("runtime.alloc")
-		packedAlloc := b.CreateCall(alloc.GlobalValueType(), alloc, []llvm.Value{
-			sizeValue,
-			llvm.ConstNull(b.dataPtrType),
-			llvm.Undef(b.dataPtrType), // unused context parameter
-		}, "")
-		packedAlloc.AddCallSiteAttribute(0, b.ctx.CreateEnumAttribute(llvm.AttributeKindID("align"), uint64(align)))
+		packedAlloc := b.createAlloc(sizeValue, layoutValue, align, "")
 		if b.NeedsStackObjects {
 			b.trackPointer(packedAlloc)
 		}
@@ -212,7 +207,7 @@ func (c *compilerContext) makeGlobalArray(buf []byte, name string, elementType l
 	globalType := llvm.ArrayType(elementType, len(buf))
 	global := llvm.AddGlobal(c.mod, globalType, name)
 	value := llvm.Undef(globalType)
-	for i := 0; i < len(buf); i++ {
+	for i := range buf {
 		ch := uint64(buf[i])
 		value = c.builder.CreateInsertValue(value, llvm.ConstInt(elementType, ch, false), i, "")
 	}
@@ -231,6 +226,12 @@ func (c *compilerContext) makeGlobalArray(buf []byte, name string, elementType l
 //
 // For details on what's in this value, see src/runtime/gc_precise.go.
 func (c *compilerContext) createObjectLayout(t llvm.Type, pos token.Pos) llvm.Value {
+	if !typeHasPointers(t) {
+		// There are no pointers in this type, so we can simplify the layout.
+		layout := (uint64(1) << 1) | 1
+		return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.dataPtrType)
+	}
+
 	// Use the element type for arrays. This works even for nested arrays.
 	for {
 		kind := t.TypeKind()
@@ -248,54 +249,29 @@ func (c *compilerContext) createObjectLayout(t llvm.Type, pos token.Pos) llvm.Va
 		break
 	}

-	// Do a few checks to see whether we need to generate any object layout
-	// information at all.
+	// Create the pointer bitmap.
 	objectSizeBytes := c.targetData.TypeAllocSize(t)
+	pointerAlignment := uint64(c.targetData.PrefTypeAlignment(c.dataPtrType))
+	bitmapLen := objectSizeBytes / pointerAlignment
+	bitmapBytes := (bitmapLen + 7) / 8
+	bitmap := make([]byte, bitmapBytes, max(bitmapBytes, 8))
+	c.buildPointerBitmap(bitmap, pointerAlignment, pos, t, 0)
+
+	// Try to encode the layout inline.
 	pointerSize := c.targetData.TypeAllocSize(c.dataPtrType)
-	pointerAlignment := c.targetData.PrefTypeAlignment(c.dataPtrType)
-	if objectSizeBytes < pointerSize {
-		// Too small to contain a pointer.
-		layout := (uint64(1) << 1) | 1
-		return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.dataPtrType)
-	}
-	bitmap := c.getPointerBitmap(t, pos)
-	if bitmap.BitLen() == 0 {
-		// There are no pointers in this type, so we can simplify the layout.
-		// TODO: this can be done in many other cases, e.g. when allocating an
-		// array (like [4][]byte, which repeats a slice 4 times).
-		layout := (uint64(1) << 1) | 1
-		return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.dataPtrType)
-	}
-	if objectSizeBytes%uint64(pointerAlignment) != 0 {
-		// This shouldn't happen except for packed structs, which aren't
-		// currently used.
-		c.addError(pos, "internal error: unexpected object size for object with pointer field")
-		return llvm.ConstNull(c.dataPtrType)
-	}
-	objectSizeWords := objectSizeBytes / uint64(pointerAlignment)
-
 	pointerBits := pointerSize * 8
-	var sizeFieldBits uint64
-	switch pointerBits {
-	case 16:
-		sizeFieldBits = 4
-	case 32:
-		sizeFieldBits = 5
-	case 64:
-		sizeFieldBits = 6
-	default:
-		panic("unknown pointer size")
-	}
-	layoutFieldBits := pointerBits - 1 - sizeFieldBits
+	if bitmapLen < pointerBits {
+		rawMask := binary.LittleEndian.Uint64(bitmap[0:8])
+		layout := rawMask*pointerBits + bitmapLen
+		layout <<= 1
+		layout |= 1

-	// Try to emit the value as an inline integer. This is possible in most
-	// cases.
-	if objectSizeWords < layoutFieldBits {
-		// If it can be stored directly in the pointer value, do so.
-		// The runtime knows that if the least significant bit of the pointer is
-		// set, the pointer contains the value itself.
-		layout := bitmap.Uint64()<<(sizeFieldBits+1) | (objectSizeWords << 1) | 1
-		return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.dataPtrType)
+		// Check if the layout fits.
+		layout &= 1<<pointerBits - 1
+		if (layout>>1)/pointerBits == rawMask {
+			// No set bits were shifted off.
+			return llvm.ConstIntToPtr(llvm.ConstInt(c.uintptrType, layout, false), c.dataPtrType)
+		}
 	}

 	// Unfortunately, the object layout is too big to fit in a pointer-sized
@@ -303,25 +279,24 @@ func (c *compilerContext) createObjectLayout(t llvm.Type, pos token.Pos) llvm.Va

 	// Try first whether the global already exists. All objects with a
 	// particular name have the same type, so this is possible.
-	globalName := "runtime/gc.layout:" + fmt.Sprintf("%d-%0*x", objectSizeWords, (objectSizeWords+15)/16, bitmap)
+	globalName := "runtime/gc.layout:" + fmt.Sprintf("%d-%0*x", bitmapLen, (bitmapLen+15)/16, bitmap)
 	global := c.mod.NamedGlobal(globalName)
 	if !global.IsNil() {
 		return global
 	}

 	// Create the global initializer.
-	bitmapBytes := make([]byte, int(objectSizeWords+7)/8)
-	bitmap.FillBytes(bitmapBytes)
-	reverseBytes(bitmapBytes) // big-endian to little-endian
-	var bitmapByteValues []llvm.Value
-	for _, b := range bitmapBytes {
-		bitmapByteValues = append(bitmapByteValues, llvm.ConstInt(c.ctx.Int8Type(), uint64(b), false))
+	bitmapByteValues := make([]llvm.Value, bitmapBytes)
+	i8 := c.ctx.Int8Type()
+	for i, b := range bitmap {
+		bitmapByteValues[i] = llvm.ConstInt(i8, uint64(b), false)
 	}
 	initializer := c.ctx.ConstStruct([]llvm.Value{
-		llvm.ConstInt(c.uintptrType, objectSizeWords, false),
-		llvm.ConstArray(c.ctx.Int8Type(), bitmapByteValues),
+		llvm.ConstInt(c.uintptrType, bitmapLen, false),
+		llvm.ConstArray(i8, bitmapByteValues),
 	}, false)

+	// Create the actual global.
 	global = llvm.AddGlobal(c.mod, initializer.Type(), globalName)
 	global.SetInitializer(initializer)
 	global.SetUnnamedAddr(true)
@@ -329,6 +304,7 @@ func (c *compilerContext) createObjectLayout(t llvm.Type, pos token.Pos) llvm.Va
 	global.SetLinkage(llvm.LinkOnceODRLinkage)
 	if c.targetData.PrefTypeAlignment(c.uintptrType) < 2 {
 		// AVR doesn't have alignment by default.
+		// The lowest bit must be unset to distinguish this from an inline layout.
 		global.SetAlignment(2)
 	}
 	if c.Debug && pos != token.NoPos {
@@ -360,52 +336,71 @@ func (c *compilerContext) createObjectLayout(t llvm.Type, pos token.Pos) llvm.Va
 	return global
 }

-// getPointerBitmap scans the given LLVM type for pointers and sets bits in a
-// bigint at the word offset that contains a pointer. This scan is recursive.
-func (c *compilerContext) getPointerBitmap(typ llvm.Type, pos token.Pos) *big.Int {
-	alignment := c.targetData.PrefTypeAlignment(c.dataPtrType)
-	switch typ.TypeKind() {
+// buildPointerBitmap scans the given LLVM type for pointers and sets bits in a
+// bitmap at the word offset that contains a pointer. This scan is recursive.
+func (c *compilerContext) buildPointerBitmap(
+	dst []byte,
+	ptrAlign uint64,
+	pos token.Pos,
+	t llvm.Type,
+	offset uint64,
+) {
+	switch t.TypeKind() {
 	case llvm.IntegerTypeKind, llvm.FloatTypeKind, llvm.DoubleTypeKind:
-		return big.NewInt(0)
+		// These types do not contain pointers.
+
 	case llvm.PointerTypeKind:
-		return big.NewInt(1)
+		// Set the corresponding position in the bitmap.
+		dst[offset/8] |= 1 << (offset % 8)
+
 	case llvm.StructTypeKind:
-		ptrs := big.NewInt(0)
-		for i, subtyp := range typ.StructElementTypes() {
-			subptrs := c.getPointerBitmap(subtyp, pos)
-			if subptrs.BitLen() == 0 {
+		// Recurse over struct elements.
+		for i, et := range t.StructElementTypes() {
+			eo := c.targetData.ElementOffset(t, i)
+			if eo%uint64(ptrAlign) != 0 {
+				if typeHasPointers(et) {
+					// This error will let the compilation fail, but by continuing
+					// the error can still easily be shown.
+					c.addError(pos, "internal error: allocated struct contains unaligned pointer")
+				}
 				continue
 			}
-			offset := c.targetData.ElementOffset(typ, i)
-			if offset%uint64(alignment) != 0 {
-				// This error will let the compilation fail, but by continuing
-				// the error can still easily be shown.
-				c.addError(pos, "internal error: allocated struct contains unaligned pointer")
-				continue
-			}
-			subptrs.Lsh(subptrs, uint(offset)/uint(alignment))
-			ptrs.Or(ptrs, subptrs)
+			c.buildPointerBitmap(
+				dst,
+				ptrAlign,
+				pos,
+				et,
+				offset+(eo/ptrAlign),
+			)
 		}
-		return ptrs
+
 	case llvm.ArrayTypeKind:
-		subtyp := typ.ElementType()
-		subptrs := c.getPointerBitmap(subtyp, pos)
-		ptrs := big.NewInt(0)
-		if subptrs.BitLen() == 0 {
-			return ptrs
+		// Recurse over array elements.
+		len := t.ArrayLength()
+		if len <= 0 {
+			return
 		}
-		elementSize := c.targetData.TypeAllocSize(subtyp)
-		if elementSize%uint64(alignment) != 0 {
-			// This error will let the compilation fail (but continues so that
-			// other errors can be shown).
-			c.addError(pos, "internal error: allocated array contains unaligned pointer")
-			return ptrs
+		et := t.ElementType()
+		elementSize := c.targetData.TypeAllocSize(et)
+		if elementSize%ptrAlign != 0 {
+			if typeHasPointers(et) {
+				// This error will let the compilation fail (but continues so that
+				// other errors can be shown).
+				c.addError(pos, "internal error: allocated array contains unaligned pointer")
+			}
+			return
 		}
-		for i := 0; i < typ.ArrayLength(); i++ {
-			ptrs.Lsh(ptrs, uint(elementSize)/uint(alignment))
-			ptrs.Or(ptrs, subptrs)
+		elementSize /= ptrAlign
+		for i := range len {
+			c.buildPointerBitmap(
+				dst,
+				ptrAlign,
+				pos,
+				et,
+				offset+uint64(i)*elementSize,
+			)
 		}
-		return ptrs
+
 	default:
 		// Should not happen.
 		panic("unknown LLVM type")
@@ -423,7 +418,7 @@ func (c *compilerContext) archFamily() string {
 // features string is not one for an ARM architecture.
 func (c *compilerContext) isThumb() bool {
 	var isThumb, isNotThumb bool
-	for _, feature := range strings.Split(c.Features, ",") {
+	for feature := range strings.SplitSeq(c.Features, ",") {
 		if feature == "+thumb-mode" {
 			isThumb = true
 		}
@@ -452,6 +447,21 @@ func (b *builder) readStackPointer() llvm.Value {
 	return b.CreateCall(stacksave.GlobalValueType(), stacksave, nil, "")
 }

+// writeStackPointer emits a LLVM intrinsic call that updates the current stack
+// pointer.
+func (b *builder) writeStackPointer(sp llvm.Value) {
+	name := "llvm.stackrestore.p0"
+	if llvmutil.Version() < 18 {
+		name = "llvm.stackrestore" // backwards compatibility with LLVM 17 and below
+	}
+	stackrestore := b.mod.NamedFunction(name)
+	if stackrestore.IsNil() {
+		fnType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.dataPtrType}, false)
+		stackrestore = llvm.AddFunction(b.mod, name, fnType)
+	}
+	b.CreateCall(stackrestore.GlobalValueType(), stackrestore, []llvm.Value{sp}, "")
+}
+
 // createZExtOrTrunc lets the input value fit in the output type bits, by zero
 // extending or truncating the integer.
 func (b *builder) createZExtOrTrunc(value llvm.Value, t llvm.Type) llvm.Value {
@@ -44,7 +44,7 @@ func CreateTemporaryAlloca(builder llvm.Builder, mod llvm.Module, t llvm.Type, n
 	alloca = CreateEntryBlockAlloca(builder, t, name)
 	size = llvm.ConstInt(ctx.Int64Type(), targetData.TypeAllocSize(t), false)
 	fnType, fn := getLifetimeStartFunc(mod)
-	builder.CreateCall(fnType, fn, []llvm.Value{size, alloca}, "")
+	builder.CreateCall(fnType, fn, lifetimeCallArgs(size, alloca), "")
 	return
 }

@@ -58,14 +58,14 @@ func CreateInstructionAlloca(builder llvm.Builder, mod llvm.Module, t llvm.Type,
 	builder.SetInsertPointBefore(inst)
 	size := llvm.ConstInt(ctx.Int64Type(), targetData.TypeAllocSize(t), false)
 	fnType, fn := getLifetimeStartFunc(mod)
-	builder.CreateCall(fnType, fn, []llvm.Value{size, alloca}, "")
+	builder.CreateCall(fnType, fn, lifetimeCallArgs(size, alloca), "")
 	if next := llvm.NextInstruction(inst); !next.IsNil() {
 		builder.SetInsertPointBefore(next)
 	} else {
 		builder.SetInsertPointAtEnd(inst.InstructionParent())
 	}
 	fnType, fn = getLifetimeEndFunc(mod)
-	builder.CreateCall(fnType, fn, []llvm.Value{size, alloca}, "")
+	builder.CreateCall(fnType, fn, lifetimeCallArgs(size, alloca), "")
 	return alloca
 }

@@ -74,7 +74,27 @@ func CreateInstructionAlloca(builder llvm.Builder, mod llvm.Module, t llvm.Type,
 // createTemporaryAlloca.
 func EmitLifetimeEnd(builder llvm.Builder, mod llvm.Module, ptr, size llvm.Value) {
 	fnType, fn := getLifetimeEndFunc(mod)
-	builder.CreateCall(fnType, fn, []llvm.Value{size, ptr}, "")
+	builder.CreateCall(fnType, fn, lifetimeCallArgs(size, ptr), "")
+}
+
+// lifetimeCallArgs returns the arguments to pass to a call of the
+// llvm.lifetime.start/end intrinsics. LLVM 22 dropped the (redundant,
+// already required to match the alloca size) i64 size argument, so the
+// intrinsic now only takes the pointer.
+func lifetimeCallArgs(size, ptr llvm.Value) []llvm.Value {
+	if Version() >= 22 {
+		return []llvm.Value{ptr}
+	}
+	return []llvm.Value{size, ptr}
+}
+
+// lifetimeFuncType returns the function type of the llvm.lifetime.start/end
+// intrinsics, which lost their i64 size parameter in LLVM 22.
+func lifetimeFuncType(ctx llvm.Context, ptrType llvm.Type) llvm.Type {
+	if Version() >= 22 {
+		return llvm.FunctionType(ctx.VoidType(), []llvm.Type{ptrType}, false)
+	}
+	return llvm.FunctionType(ctx.VoidType(), []llvm.Type{ctx.Int64Type(), ptrType}, false)
 }

 // getLifetimeStartFunc returns the llvm.lifetime.start intrinsic and creates it
@@ -84,7 +104,7 @@ func getLifetimeStartFunc(mod llvm.Module) (llvm.Type, llvm.Value) {
 	fn := mod.NamedFunction(fnName)
 	ctx := mod.Context()
 	ptrType := llvm.PointerType(ctx.Int8Type(), 0)
-	fnType := llvm.FunctionType(ctx.VoidType(), []llvm.Type{ctx.Int64Type(), ptrType}, false)
+	fnType := lifetimeFuncType(ctx, ptrType)
 	if fn.IsNil() {
 		fn = llvm.AddFunction(mod, fnName, fnType)
 	}
@@ -98,7 +118,7 @@ func getLifetimeEndFunc(mod llvm.Module) (llvm.Type, llvm.Value) {
 	fn := mod.NamedFunction(fnName)
 	ctx := mod.Context()
 	ptrType := llvm.PointerType(ctx.Int8Type(), 0)
-	fnType := llvm.FunctionType(ctx.VoidType(), []llvm.Type{ctx.Int64Type(), ptrType}, false)
+	fnType := lifetimeFuncType(ctx, ptrType)
 	if fn.IsNil() {
 		fn = llvm.AddFunction(mod, fnName, fnType)
 	}
@@ -218,7 +238,37 @@ func Version() int {
 	return major
 }

-// Return the byte order for the given target triple. Most targets are little
+// NoCaptureAttrName returns the name of the LLVM attribute kind that marks a
+// pointer parameter as guaranteed not to escape by capture.
+//
+// LLVM 21 removed the old boolean 'nocapture' enum attribute in favor of the
+// more expressive 'captures' int attribute, where 'captures(none)' (encoded
+// as the value 0) is the equivalent of the old 'nocapture'. LLVM 20 supports
+// both, but its own optimizer still emits 'nocapture', so the cutoff for
+// reading/writing the new name is LLVM 21.
+func NoCaptureAttrName() string {
+	if Version() >= 21 {
+		return "captures"
+	}
+	return "nocapture"
+}
+
+// IsNoCapture reports whether attr (looked up using the kind returned by
+// NoCaptureAttrName) indicates that the pointer it is attached to does not
+// escape by capture. It returns false for a nil attribute.
+func IsNoCapture(attr llvm.Attribute) bool {
+	if attr.IsNil() {
+		return false
+	}
+	if Version() >= 21 {
+		// captures(none) is encoded as the value 0; any other value permits
+		// some form of capture.
+		return attr.GetEnumValue() == 0
+	}
+	return true
+}
+
+// ByteOrder returns the byte order for the given target triple. Most targets are little
 // endian, but for example MIPS can be big-endian.
 func ByteOrder(target string) binary.ByteOrder {
 	if strings.HasPrefix(target, "mips-") {
@@ -3,42 +3,29 @@ package compiler
 // This file emits the correct map intrinsics for map operations.

 import (
+	"fmt"
 	"go/token"
 	"go/types"
-
-	"github.com/tinygo-org/tinygo/src/tinygo"
 	"golang.org/x/tools/go/ssa"
+	"strings"
 	"tinygo.org/x/go-llvm"
 )

+const hashArrayUnrollLimit = 4
+
 // createMakeMap creates a new map object (runtime.hashmap) by allocating and
 // initializing an appropriately sized object.
 func (b *builder) createMakeMap(expr *ssa.MakeMap) (llvm.Value, error) {
 	mapType := expr.Type().Underlying().(*types.Map)
 	keyType := mapType.Key().Underlying()
 	llvmValueType := b.getLLVMType(mapType.Elem().Underlying())
-	var llvmKeyType llvm.Type
-	var alg uint64
-	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
-		// String keys.
-		llvmKeyType = b.getLLVMType(keyType)
-		alg = uint64(tinygo.HashmapAlgorithmString)
-	} else if hashmapIsBinaryKey(keyType) {
-		// Trivially comparable keys.
-		llvmKeyType = b.getLLVMType(keyType)
-		alg = uint64(tinygo.HashmapAlgorithmBinary)
-	} else {
-		// All other keys. Implemented as map[interface{}]valueType for ease of
-		// implementation.
-		llvmKeyType = b.getLLVMRuntimeType("_interface")
-		alg = uint64(tinygo.HashmapAlgorithmInterface)
-	}
+	llvmKeyType := b.getLLVMType(keyType)
+
 	keySize := b.targetData.TypeAllocSize(llvmKeyType)
 	valueSize := b.targetData.TypeAllocSize(llvmValueType)
 	llvmKeySize := llvm.ConstInt(b.uintptrType, keySize, false)
 	llvmValueSize := llvm.ConstInt(b.uintptrType, valueSize, false)
 	sizeHint := llvm.ConstInt(b.uintptrType, 8, false)
-	algEnum := llvm.ConstInt(b.ctx.Int8Type(), alg, false)
 	if expr.Reserve != nil {
 		sizeHint = b.getValue(expr.Reserve, getPos(expr))
 		var err error
@@ -47,134 +34,129 @@ func (b *builder) createMakeMap(expr *ssa.MakeMap) (llvm.Value, error) {
 			return llvm.Value{}, err
 		}
 	}
-	hashmap := b.createRuntimeCall("hashmapMake", []llvm.Value{llvmKeySize, llvmValueSize, sizeHint, algEnum}, "")
+
+	// Resolve hash and equal functions for this key type. For string and
+	// binary key types, reference the corresponding runtime functions
+	// directly. For composite types, generate type-specific functions.
+	var hashFn, equalFn llvm.Value
+	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
+		hashFn = b.getRuntimeFunctionValue("hashmapStringPtrHash", hashmapKeyHashSignature())
+		equalFn = b.getRuntimeFunctionValue("hashmapStringEqual", hashmapKeyEqualSignature())
+	} else if hashmapIsBinaryKey(keyType) {
+		hashFn = b.getRuntimeFunctionValue("hash32", hashmapKeyHashSignature())
+		equalFn = b.getRuntimeFunctionValue("memequal", hashmapKeyEqualSignature())
+	} else {
+		fn := b.getOrGenerateKeyHashFunc(keyType)
+		hashFn = b.createFuncValue(fn, llvm.ConstNull(b.dataPtrType), hashmapKeyHashSignature())
+		fn = b.getOrGenerateKeyEqualFunc(keyType)
+		equalFn = b.createFuncValue(fn, llvm.ConstNull(b.dataPtrType), hashmapKeyEqualSignature())
+	}
+
+	hashmap := b.createRuntimeCall("hashmapMakeGeneric", []llvm.Value{
+		llvmKeySize, llvmValueSize, sizeHint,
+		hashFn, equalFn,
+	}, "")
 	return hashmap, nil
 }

+// getRuntimeFunctionValue returns a TinyGo function value (with nil context)
+// for the named runtime function.
+func (b *builder) getRuntimeFunctionValue(name string, sig *types.Signature) llvm.Value {
+	member := b.program.ImportedPackage("runtime").Members[name]
+	if member == nil {
+		panic("unknown runtime function: " + name)
+	}
+	_, llvmFn := b.getFunction(member.(*ssa.Function))
+	return b.createFuncValue(llvmFn, llvm.ConstNull(b.dataPtrType), sig)
+}
+
 // createMapLookup returns the value in a map. It calls a runtime function
 // depending on the map key type to load the map value and its comma-ok value.
-func (b *builder) createMapLookup(keyType, valueType types.Type, m, key llvm.Value, commaOk bool, pos token.Pos) (llvm.Value, error) {
+func (b *builder) createMapLookup(keyType, valueType types.Type, m llvm.Value, key ssa.Value, commaOk bool, pos token.Pos) (llvm.Value, error) {
 	llvmValueType := b.getLLVMType(valueType)

 	// Allocate the memory for the resulting type. Do not zero this memory: it
 	// will be zeroed by the hashmap get implementation if the key is not
 	// present in the map.
-	mapValueAlloca, mapValueAllocaSize := b.createTemporaryAlloca(llvmValueType, "hashmap.value")
+	result := b.createRuntimeValueResult(llvmValueType, commaOk, false, "hashmap")
+	mapValueAlloca := result.valuePtr

 	// We need the map size (with type uintptr) to pass to the hashmap*Get
 	// functions. This is necessary because those *Get functions are valid on
 	// nil maps, and they'll need to zero the value pointer by that number of
 	// bytes.
-	mapValueSize := mapValueAllocaSize
+	mapValueSize := result.valueSize
 	if mapValueSize.Type().IntTypeWidth() > b.uintptrType.IntTypeWidth() {
 		mapValueSize = llvm.ConstTrunc(mapValueSize, b.uintptrType)
 	}

 	// Do the lookup. How it is done depends on the key type.
 	var commaOkValue llvm.Value
-	origKeyType := keyType
 	keyType = keyType.Underlying()
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
-		params := []llvm.Value{m, key, mapValueAlloca, mapValueSize}
+		params := []llvm.Value{m, b.getValue(key, getPos(key)), mapValueAlloca, mapValueSize}
 		commaOkValue = b.createRuntimeCall("hashmapStringGet", params, "")
-	} else if hashmapIsBinaryKey(keyType) {
-		// key can be compared with runtime.memequal
-		// Store the key in an alloca, in the entry block to avoid dynamic stack
-		// growth.
-		mapKeyAlloca, mapKeySize := b.createTemporaryAlloca(key.Type(), "hashmap.key")
-		b.CreateStore(key, mapKeyAlloca)
-		b.zeroUndefBytes(b.getLLVMType(keyType), mapKeyAlloca)
-		// Fetch the value from the hashmap.
-		params := []llvm.Value{m, mapKeyAlloca, mapValueAlloca, mapValueSize}
-		commaOkValue = b.createRuntimeCall("hashmapBinaryGet", params, "")
-		b.emitLifetimeEnd(mapKeyAlloca, mapKeySize)
 	} else {
-		// Not trivially comparable using memcmp. Make it an interface instead.
-		itfKey := key
-		if _, ok := keyType.(*types.Interface); !ok {
-			// Not already an interface, so convert it to an interface now.
-			itfKey = b.createMakeInterface(key, origKeyType, pos)
+		// Key stored at actual type: either binary-comparable or with
+		// compiler-generated hash/equal.
+		mapKey := b.getValueStorage(key, "hashmap.key")
+		params := []llvm.Value{m, mapKey.ptr, mapValueAlloca, mapValueSize}
+		fnName := "hashmapBinaryGet"
+		if !hashmapIsBinaryKey(keyType) {
+			fnName = "hashmapGenericGet"
 		}
-		params := []llvm.Value{m, itfKey, mapValueAlloca, mapValueSize}
-		commaOkValue = b.createRuntimeCall("hashmapInterfaceGet", params, "")
+		commaOkValue = b.createRuntimeCall(fnName, params, "")
+		b.endValueStorage(mapKey)
 	}

-	// Load the resulting value from the hashmap. The value is set to the zero
-	// value if the key doesn't exist in the hashmap.
-	mapValue := b.CreateLoad(llvmValueType, mapValueAlloca, "")
-	b.emitLifetimeEnd(mapValueAlloca, mapValueAllocaSize)
-
-	if commaOk {
-		tuple := llvm.Undef(b.ctx.StructType([]llvm.Type{llvmValueType, b.ctx.Int1Type()}, false))
-		tuple = b.CreateInsertValue(tuple, mapValue, 0, "")
-		tuple = b.CreateInsertValue(tuple, commaOkValue, 1, "")
-		return tuple, nil
-	} else {
-		return mapValue, nil
-	}
+	// The value is set to the zero value if the key doesn't exist.
+	return result.finish(b, commaOkValue, ""), nil
 }

 // createMapUpdate updates a map key to a given value, by creating an
 // appropriate runtime call.
-func (b *builder) createMapUpdate(keyType types.Type, m, key, value llvm.Value, pos token.Pos) {
-	valueAlloca, valueSize := b.createTemporaryAlloca(value.Type(), "hashmap.value")
-	b.CreateStore(value, valueAlloca)
-	origKeyType := keyType
+func (b *builder) createMapUpdate(keyType types.Type, m llvm.Value, key, value ssa.Value, pos token.Pos) {
+	storedValue := b.getValueStorage(value, "hashmap.value")
 	keyType = keyType.Underlying()
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
-		params := []llvm.Value{m, key, valueAlloca}
-		b.createRuntimeCall("hashmapStringSet", params, "")
-	} else if hashmapIsBinaryKey(keyType) {
-		// key can be compared with runtime.memequal
-		keyAlloca, keySize := b.createTemporaryAlloca(key.Type(), "hashmap.key")
-		b.CreateStore(key, keyAlloca)
-		b.zeroUndefBytes(b.getLLVMType(keyType), keyAlloca)
-		params := []llvm.Value{m, keyAlloca, valueAlloca}
-		b.createRuntimeCall("hashmapBinarySet", params, "")
-		b.emitLifetimeEnd(keyAlloca, keySize)
+		params := []llvm.Value{m, b.getValue(key, getPos(key)), storedValue.ptr}
+		b.createRuntimeInvoke("hashmapStringSet", params, "")
 	} else {
-		// Key is not trivially comparable, so compare it as an interface instead.
-		itfKey := key
-		if _, ok := keyType.(*types.Interface); !ok {
-			// Not already an interface, so convert it to an interface first.
-			itfKey = b.createMakeInterface(key, origKeyType, pos)
+		// Key stored at actual type.
+		keyStorage := b.getValueStorage(key, "hashmap.key")
+		fnName := "hashmapBinarySet"
+		if !hashmapIsBinaryKey(keyType) {
+			fnName = "hashmapGenericSet"
 		}
-		params := []llvm.Value{m, itfKey, valueAlloca}
-		b.createRuntimeCall("hashmapInterfaceSet", params, "")
+		params := []llvm.Value{m, keyStorage.ptr, storedValue.ptr}
+		b.createRuntimeInvoke(fnName, params, "")
+		b.endValueStorage(keyStorage)
 	}
-	b.emitLifetimeEnd(valueAlloca, valueSize)
+	b.endValueStorage(storedValue)
 }

 // createMapDelete deletes a key from a map by calling the appropriate runtime
 // function. It is the implementation of the Go delete() builtin.
 func (b *builder) createMapDelete(keyType types.Type, m, key llvm.Value, pos token.Pos) error {
-	origKeyType := keyType
 	keyType = keyType.Underlying()
 	if t, ok := keyType.(*types.Basic); ok && t.Info()&types.IsString != 0 {
 		// key is a string
 		params := []llvm.Value{m, key}
 		b.createRuntimeCall("hashmapStringDelete", params, "")
 		return nil
-	} else if hashmapIsBinaryKey(keyType) {
+	} else {
+		// Key stored at actual type.
 		keyAlloca, keySize := b.createTemporaryAlloca(key.Type(), "hashmap.key")
 		b.CreateStore(key, keyAlloca)
-		b.zeroUndefBytes(b.getLLVMType(keyType), keyAlloca)
-		params := []llvm.Value{m, keyAlloca}
-		b.createRuntimeCall("hashmapBinaryDelete", params, "")
-		b.emitLifetimeEnd(keyAlloca, keySize)
-		return nil
-	} else {
-		// Key is not trivially comparable, so compare it as an interface
-		// instead.
-		itfKey := key
-		if _, ok := keyType.(*types.Interface); !ok {
-			// Not already an interface, so convert it to an interface first.
-			itfKey = b.createMakeInterface(key, origKeyType, pos)
+		fnName := "hashmapBinaryDelete"
+		if !hashmapIsBinaryKey(keyType) {
+			fnName = "hashmapGenericDelete"
 		}
-		params := []llvm.Value{m, itfKey}
-		b.createRuntimeCall("hashmapInterfaceDelete", params, "")
+		params := []llvm.Value{m, keyAlloca}
+		b.createRuntimeCall(fnName, params, "")
+		b.emitLifetimeEnd(keyAlloca, keySize)
 		return nil
 	}
 }
@@ -195,42 +177,15 @@ func (b *builder) createMapIteratorNext(rangeVal ssa.Value, llvmRangeVal, it llv
 	llvmKeyType := b.getLLVMType(keyType)
 	llvmValueType := b.getLLVMType(valueType)

-	// There is a special case in which keys are stored as an interface value
-	// instead of the value they normally are. This happens for non-trivially
-	// comparable types such as float32 or some structs.
-	isKeyStoredAsInterface := false
-	if t, ok := keyType.Underlying().(*types.Basic); ok && t.Info()&types.IsString != 0 {
-		// key is a string
-	} else if hashmapIsBinaryKey(keyType) {
-		// key can be compared with runtime.memequal
-	} else {
-		// The key is stored as an interface value, and may or may not be an
-		// interface type (for example, float32 keys are stored as an interface
-		// value).
-		if _, ok := keyType.Underlying().(*types.Interface); !ok {
-			isKeyStoredAsInterface = true
-		}
-	}
-
-	// Determine the type of the key as stored in the map.
-	llvmStoredKeyType := llvmKeyType
-	if isKeyStoredAsInterface {
-		llvmStoredKeyType = b.getLLVMRuntimeType("_interface")
-	}
+	// All key types are now stored at their declared type (no interface wrapping).

 	// Extract the key and value from the map.
-	mapKeyAlloca, mapKeySize := b.createTemporaryAlloca(llvmStoredKeyType, "range.key")
+	mapKeyAlloca, mapKeySize := b.createTemporaryAlloca(llvmKeyType, "range.key")
 	mapValueAlloca, mapValueSize := b.createTemporaryAlloca(llvmValueType, "range.value")
 	ok := b.createRuntimeCall("hashmapNext", []llvm.Value{llvmRangeVal, it, mapKeyAlloca, mapValueAlloca}, "range.next")
-	mapKey := b.CreateLoad(llvmStoredKeyType, mapKeyAlloca, "")
+	mapKey := b.CreateLoad(llvmKeyType, mapKeyAlloca, "")
 	mapValue := b.CreateLoad(llvmValueType, mapValueAlloca, "")

-	if isKeyStoredAsInterface {
-		// The key is stored as an interface but it isn't of interface type.
-		// Extract the underlying value.
-		mapKey = b.extractValueFromInterface(mapKey, llvmKeyType)
-	}
-
 	// End the lifetimes of the allocas, because we're done with them.
 	b.emitLifetimeEnd(mapKeyAlloca, mapKeySize)
 	b.emitLifetimeEnd(mapValueAlloca, mapValueSize)
@@ -248,90 +203,412 @@ func (b *builder) createMapIteratorNext(rangeVal ssa.Value, llvmRangeVal, it llv
 // can be compared with runtime.memequal.  Note that padding bytes are undef
 // and can alter two "equal" structs being equal when compared with memequal.
 func hashmapIsBinaryKey(keyType types.Type) bool {
-	switch keyType := keyType.(type) {
+	switch keyType := keyType.Underlying().(type) {
 	case *types.Basic:
-		return keyType.Info()&(types.IsBoolean|types.IsInteger) != 0
+		return keyType.Info()&(types.IsBoolean|types.IsInteger) != 0 || keyType.Kind() == types.UnsafePointer
 	case *types.Pointer:
 		return true
-	case *types.Struct:
-		for i := 0; i < keyType.NumFields(); i++ {
-			fieldType := keyType.Field(i).Type().Underlying()
-			if !hashmapIsBinaryKey(fieldType) {
-				return false
-			}
-		}
-		return true
 	case *types.Array:
 		return hashmapIsBinaryKey(keyType.Elem())
-	case *types.Named:
-		return hashmapIsBinaryKey(keyType.Underlying())
 	default:
 		return false
 	}
 }

-func (b *builder) zeroUndefBytes(llvmType llvm.Type, ptr llvm.Value) error {
-	// We know that hashmapIsBinaryKey is true, so we only have to handle those types that can show up there.
-	// To zero all undefined bytes, we iterate over all the fields in the type.  For each element, compute the
-	// offset of that element.  If it's Basic type, there are no internal padding bytes.  For compound types, we recurse to ensure
-	// we handle nested types.  Next, we determine if there are any padding bytes before the next
-	// element and zero those as well.
+// hashmapKeyHashSignature returns the Go type signature for hashmap key hash
+// functions: func(key unsafe.Pointer, size, seed uintptr) uint32
+func hashmapKeyHashSignature() *types.Signature {
+	return types.NewSignatureType(nil, nil, nil,
+		types.NewTuple(
+			types.NewVar(token.NoPos, nil, "key", types.Typ[types.UnsafePointer]),
+			types.NewVar(token.NoPos, nil, "size", types.Typ[types.Uintptr]),
+			types.NewVar(token.NoPos, nil, "seed", types.Typ[types.Uintptr]),
+		),
+		types.NewTuple(
+			types.NewVar(token.NoPos, nil, "", types.Typ[types.Uint32]),
+		),
+		false,
+	)
+}

-	zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+// hashmapKeyEqualSignature returns the Go type signature for hashmap key equal
+// functions: func(x, y unsafe.Pointer, n uintptr) bool
+func hashmapKeyEqualSignature() *types.Signature {
+	return types.NewSignatureType(nil, nil, nil,
+		types.NewTuple(
+			types.NewVar(token.NoPos, nil, "x", types.Typ[types.UnsafePointer]),
+			types.NewVar(token.NoPos, nil, "y", types.Typ[types.UnsafePointer]),
+			types.NewVar(token.NoPos, nil, "n", types.Typ[types.Uintptr]),
+		),
+		types.NewTuple(
+			types.NewVar(token.NoPos, nil, "", types.Typ[types.Bool]),
+		),
+		false,
+	)
+}

-	switch llvmType.TypeKind() {
-	case llvm.IntegerTypeKind:
-		// no padding bytes
-		return nil
-	case llvm.PointerTypeKind:
-		// mo padding bytes
-		return nil
-	case llvm.ArrayTypeKind:
-		llvmArrayType := llvmType
-		llvmElemType := llvmType.ElementType()
+// hashmapKeyFuncName returns a canonical name for a generated hash or equal
+// function based on the key type's underlying structure. Named types are
+// replaced with their underlying types so that structurally identical key
+// types (e.g., struct{i1; str1} and struct{i2; str2} where both i1, i2 are
+// int and str1, str2 are string) share the same generated function.
+func hashmapKeyFuncName(prefix string, keyType types.Type) string {
+	return prefix + "." + hashmapCanonicalTypeName(keyType)
+}

-		for i := 0; i < llvmArrayType.ArrayLength(); i++ {
-			idx := llvm.ConstInt(b.uintptrType, uint64(i), false)
-			elemPtr := b.CreateInBoundsGEP(llvmArrayType, ptr, []llvm.Value{zero, idx}, "")
-
-			// zero any padding bytes in this element
-			b.zeroUndefBytes(llvmElemType, elemPtr)
+// hashmapCanonicalTypeName returns a string representation of the hash/equal
+// operations needed for a type, stripping named types where the operation does
+// not depend on the name. Pointer and channel names do not include the element
+// type because their hash/equal operations only use the pointer word.
+func hashmapCanonicalTypeName(t types.Type) string {
+	switch t := t.Underlying().(type) {
+	case *types.Basic:
+		return t.Name()
+	case *types.Pointer:
+		return "*"
+	case *types.Chan:
+		switch t.Dir() {
+		case types.SendRecv:
+			return "chan"
+		case types.SendOnly:
+			return "chan<-"
+		case types.RecvOnly:
+			return "<-chan"
 		}
-
-	case llvm.StructTypeKind:
-		llvmStructType := llvmType
-		numFields := llvmStructType.StructElementTypesCount()
-		llvmElementTypes := llvmStructType.StructElementTypes()
-
-		for i := 0; i < numFields; i++ {
-			idx := llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false)
-			elemPtr := b.CreateInBoundsGEP(llvmStructType, ptr, []llvm.Value{zero, idx}, "")
-
-			// zero any padding bytes in this field
-			llvmElemType := llvmElementTypes[i]
-			b.zeroUndefBytes(llvmElemType, elemPtr)
-
-			// zero any padding bytes before the next field, if any
-			offset := b.targetData.ElementOffset(llvmStructType, i)
-			storeSize := b.targetData.TypeStoreSize(llvmElemType)
-			fieldEndOffset := offset + storeSize
-
-			var nextOffset uint64
-			if i < numFields-1 {
-				nextOffset = b.targetData.ElementOffset(llvmStructType, i+1)
-			} else {
-				// Last field?  Next offset is the total size of the allocate struct.
-				nextOffset = b.targetData.TypeAllocSize(llvmStructType)
-			}
-
-			if fieldEndOffset != nextOffset {
-				n := llvm.ConstInt(b.uintptrType, nextOffset-fieldEndOffset, false)
-				llvmStoreSize := llvm.ConstInt(b.uintptrType, storeSize, false)
-				paddingStart := b.CreateInBoundsGEP(b.ctx.Int8Type(), elemPtr, []llvm.Value{llvmStoreSize}, "")
-				b.createRuntimeCall("memzero", []llvm.Value{paddingStart, n}, "")
-			}
+	case *types.Interface:
+		if t.NumMethods() == 0 {
+			return "interface{}"
 		}
+		return t.String()
+	case *types.Struct:
+		var s strings.Builder
+		s.WriteString("struct{")
+		for i := 0; i < t.NumFields(); i++ {
+			if i > 0 {
+				s.WriteString("; ")
+			}
+			s.WriteString(hashmapCanonicalTypeName(t.Field(i).Type()))
+		}
+		return s.String() + "}"
+	case *types.Array:
+		return fmt.Sprintf("[%d]%s", t.Len(), hashmapCanonicalTypeName(t.Elem()))
+	}
+	return t.String()
+}
+
+// getOrGenerateKeyHashFunc returns an LLVM function that computes the hash
+// of a key of the given type. The function is generated on first call and
+// cached in the module.
+func (b *builder) getOrGenerateKeyHashFunc(keyType types.Type) llvm.Value {
+	name := hashmapKeyFuncName("hashmapKeyHash", keyType)
+	if fn := b.mod.NamedFunction(name); !fn.IsNil() {
+		return fn
 	}

-	return nil
+	// Create the LLVM function type:
+	// (key ptr, size uintptr, seed uintptr, context ptr) -> i32
+	fnType := llvm.FunctionType(b.ctx.Int32Type(), []llvm.Type{
+		b.dataPtrType, b.uintptrType, b.uintptrType, b.dataPtrType,
+	}, false)
+	fn := llvm.AddFunction(b.mod, name, fnType)
+	fn.SetLinkage(llvm.LinkOnceODRLinkage)
+	fn.SetUnnamedAddr(true)
+	b.addStandardAttributes(fn)
+
+	// Generate the function body.
+	savedBlock := b.GetInsertBlock()
+	defer b.SetInsertPointAtEnd(savedBlock)
+
+	entry := b.ctx.AddBasicBlock(fn, "entry")
+	b.SetInsertPointAtEnd(entry)
+
+	keyPtr := fn.Param(0)
+	seed := fn.Param(2)
+	llvmKeyType := b.getLLVMType(keyType)
+	hash := b.generateKeyHash(keyType, llvmKeyType, keyPtr, seed)
+	b.CreateRet(hash)
+
+	return fn
+}
+
+// getOrGenerateKeyEqualFunc returns an LLVM function that compares two keys
+// of the given type for equality. The function is generated on first call
+// and cached in the module.
+func (b *builder) getOrGenerateKeyEqualFunc(keyType types.Type) llvm.Value {
+	name := hashmapKeyFuncName("hashmapKeyEqual", keyType)
+	if fn := b.mod.NamedFunction(name); !fn.IsNil() {
+		return fn
+	}
+
+	// Create the LLVM function type:
+	// (x ptr, y ptr, n uintptr, context ptr) -> i1
+	fnType := llvm.FunctionType(b.ctx.Int1Type(), []llvm.Type{
+		b.dataPtrType, b.dataPtrType, b.uintptrType, b.dataPtrType,
+	}, false)
+	fn := llvm.AddFunction(b.mod, name, fnType)
+	fn.SetLinkage(llvm.LinkOnceODRLinkage)
+	fn.SetUnnamedAddr(true)
+	b.addStandardAttributes(fn)
+
+	// Generate the function body.
+	savedBlock := b.GetInsertBlock()
+	defer b.SetInsertPointAtEnd(savedBlock)
+
+	entry := b.ctx.AddBasicBlock(fn, "entry")
+	b.SetInsertPointAtEnd(entry)
+
+	xPtr := fn.Param(0)
+	yPtr := fn.Param(1)
+	llvmKeyType := b.getLLVMType(keyType)
+	result := b.generateKeyEqual(keyType, llvmKeyType, xPtr, yPtr, fn)
+	b.CreateRet(result)
+
+	return fn
+}
+
+// generateKeyHash generates IR that hashes a key value. Returns the i32 hash.
+func (b *builder) generateKeyHash(keyType types.Type, llvmKeyType llvm.Type, keyPtr llvm.Value, seed llvm.Value) llvm.Value {
+	switch keyType := keyType.Underlying().(type) {
+	case *types.Basic:
+		if keyType.Info()&types.IsString != 0 {
+			// Hash the string contents. The size parameter is unused by
+			// hashmapStringPtrHash (it dereferences the string header to
+			// get the actual length), but we pass it for signature
+			// consistency with other hash functions.
+			size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+			return b.createRuntimeCall("hashmapStringPtrHash", []llvm.Value{keyPtr, size, seed}, "hash")
+		}
+		if keyType.Info()&types.IsFloat != 0 {
+			// Float hash: normalizes -0 to +0 before hashing.
+			if keyType.Kind() == types.Float32 {
+				return b.createRuntimeCall("hashmapFloat32Hash", []llvm.Value{keyPtr, seed}, "hash")
+			}
+			return b.createRuntimeCall("hashmapFloat64Hash", []llvm.Value{keyPtr, seed}, "hash")
+		}
+		if keyType.Info()&types.IsComplex != 0 {
+			// Complex hash: hash real and imaginary parts as floats.
+			if keyType.Kind() == types.Complex64 {
+				realPtr := keyPtr
+				imagPtr := b.CreateInBoundsGEP(b.ctx.Int8Type(), keyPtr, []llvm.Value{
+					llvm.ConstInt(b.uintptrType, 4, false),
+				}, "")
+				realHash := b.createRuntimeCall("hashmapFloat32Hash", []llvm.Value{realPtr, seed}, "hash.real")
+				imagHash := b.createRuntimeCall("hashmapFloat32Hash", []llvm.Value{imagPtr, seed}, "hash.imag")
+				return b.CreateXor(realHash, imagHash, "")
+			}
+			realPtr := keyPtr
+			imagPtr := b.CreateInBoundsGEP(b.ctx.Int8Type(), keyPtr, []llvm.Value{
+				llvm.ConstInt(b.uintptrType, 8, false),
+			}, "")
+			realHash := b.createRuntimeCall("hashmapFloat64Hash", []llvm.Value{realPtr, seed}, "hash.real")
+			imagHash := b.createRuntimeCall("hashmapFloat64Hash", []llvm.Value{imagPtr, seed}, "hash.imag")
+			return b.CreateXor(realHash, imagHash, "")
+		}
+		// Integer/boolean: hash the raw bytes.
+		size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+		return b.createRuntimeCall("hash32", []llvm.Value{keyPtr, size, seed}, "hash")
+	case *types.Pointer, *types.Chan:
+		// Pointers and channels: hash as raw pointer-sized bytes.
+		size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+		return b.createRuntimeCall("hash32", []llvm.Value{keyPtr, size, seed}, "hash")
+	case *types.Interface:
+		// Interface: use runtime reflection-based hash.
+		size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+		return b.createRuntimeCall("hashmapInterfacePtrHash", []llvm.Value{keyPtr, size, seed}, "hash")
+	case *types.Struct:
+		hash := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+		zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+		for i := 0; i < keyType.NumFields(); i++ {
+			if keyType.Field(i).Name() == "_" {
+				continue // blank fields are ignored in Go equality
+			}
+			fieldType := keyType.Field(i).Type()
+			llvmFieldType := b.getLLVMType(fieldType)
+			if b.targetData.TypeAllocSize(llvmFieldType) == 0 {
+				continue // skip zero-sized fields
+			}
+			idx := llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false)
+			fieldPtr := b.CreateInBoundsGEP(llvmKeyType, keyPtr, []llvm.Value{zero, idx}, "")
+			fieldHash := b.generateKeyHash(fieldType, llvmFieldType, fieldPtr, seed)
+			hash = b.CreateXor(hash, fieldHash, "")
+		}
+		return hash
+	case *types.Array:
+		elemType := keyType.Elem()
+		llvmElemType := b.getLLVMType(elemType)
+		arrayLen := keyType.Len()
+		if hashmapIsBinaryKey(elemType) {
+			// All elements are binary-comparable; hash the entire array as raw bytes.
+			size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+			return b.createRuntimeCall("hash32", []llvm.Value{keyPtr, size, seed}, "hash")
+		}
+		if arrayLen == 0 {
+			return llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+		}
+		if arrayLen <= hashArrayUnrollLimit {
+			hash := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+			for i := 0; i < int(arrayLen); i++ {
+				idx := llvm.ConstInt(b.uintptrType, uint64(i), false)
+				elemPtr := b.CreateInBoundsGEP(llvmKeyType, keyPtr, []llvm.Value{zero, idx}, "")
+				elemHash := b.generateKeyHash(elemType, llvmElemType, elemPtr, seed)
+				hash = b.CreateXor(hash, elemHash, "")
+			}
+			return hash
+		}
+		initHash := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+		zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+
+		loopEntry := b.GetInsertBlock()
+		loopBody := b.ctx.AddBasicBlock(loopEntry.Parent(), "hash.array.body")
+		loopDone := b.ctx.AddBasicBlock(loopEntry.Parent(), "hash.array.done")
+
+		b.CreateBr(loopBody)
+		b.SetInsertPointAtEnd(loopBody)
+
+		phiI := b.CreatePHI(b.uintptrType, "i")
+		phiHash := b.CreatePHI(b.ctx.Int32Type(), "hash.acc")
+
+		elemPtr := b.CreateInBoundsGEP(llvmKeyType, keyPtr, []llvm.Value{zero, phiI}, "")
+		elemHash := b.generateKeyHash(elemType, llvmElemType, elemPtr, seed)
+		newHash := b.CreateXor(phiHash, elemHash, "")
+		nextI := b.CreateAdd(phiI, llvm.ConstInt(b.uintptrType, 1, false), "")
+		cond := b.CreateICmp(llvm.IntULT, nextI, llvm.ConstInt(b.uintptrType, uint64(arrayLen), false), "")
+		b.CreateCondBr(cond, loopBody, loopDone)
+
+		bodyEnd := b.GetInsertBlock()
+		phiI.AddIncoming([]llvm.Value{llvm.ConstInt(b.uintptrType, 0, false), nextI},
+			[]llvm.BasicBlock{loopEntry, bodyEnd})
+		phiHash.AddIncoming([]llvm.Value{initHash, newHash},
+			[]llvm.BasicBlock{loopEntry, bodyEnd})
+
+		b.SetInsertPointAtEnd(loopDone)
+		return newHash
+	default:
+		panic(fmt.Sprintf("unhandled key type for hash generation: %T", keyType))
+	}
+}
+
+// generateKeyEqual generates IR that compares two key values for equality.
+// Returns an i1 result.
+func (b *builder) generateKeyEqual(keyType types.Type, llvmKeyType llvm.Type, xPtr, yPtr llvm.Value, fn llvm.Value) llvm.Value {
+	switch keyType := keyType.Underlying().(type) {
+	case *types.Basic:
+		if keyType.Info()&types.IsString != 0 {
+			// Compare strings: load both string headers and compare.
+			xStr := b.CreateLoad(llvmKeyType, xPtr, "x.str")
+			yStr := b.CreateLoad(llvmKeyType, yPtr, "y.str")
+			return b.createRuntimeCall("stringEqual", []llvm.Value{xStr, yStr}, "eq")
+		}
+		if keyType.Info()&types.IsFloat != 0 {
+			// Float equality: fcmp oeq handles -0==+0 (true) and NaN==NaN (false).
+			xVal := b.CreateLoad(llvmKeyType, xPtr, "x.float")
+			yVal := b.CreateLoad(llvmKeyType, yPtr, "y.float")
+			return b.CreateFCmp(llvm.FloatOEQ, xVal, yVal, "eq")
+		}
+		if keyType.Info()&types.IsComplex != 0 {
+			// Complex equality: both real and imaginary parts must be equal.
+			var floatType llvm.Type
+			if keyType.Kind() == types.Complex64 {
+				floatType = b.ctx.FloatType()
+			} else {
+				floatType = b.ctx.DoubleType()
+			}
+			floatSize := b.targetData.TypeAllocSize(floatType)
+			imagOffset := llvm.ConstInt(b.uintptrType, floatSize, false)
+			// Real parts
+			xReal := b.CreateLoad(floatType, xPtr, "x.real")
+			yReal := b.CreateLoad(floatType, yPtr, "y.real")
+			realEq := b.CreateFCmp(llvm.FloatOEQ, xReal, yReal, "eq.real")
+			// Imaginary parts
+			xImagPtr := b.CreateInBoundsGEP(b.ctx.Int8Type(), xPtr, []llvm.Value{imagOffset}, "")
+			yImagPtr := b.CreateInBoundsGEP(b.ctx.Int8Type(), yPtr, []llvm.Value{imagOffset}, "")
+			xImag := b.CreateLoad(floatType, xImagPtr, "x.imag")
+			yImag := b.CreateLoad(floatType, yImagPtr, "y.imag")
+			imagEq := b.CreateFCmp(llvm.FloatOEQ, xImag, yImag, "eq.imag")
+			return b.CreateAnd(realEq, imagEq, "")
+		}
+		// Integer/boolean: compare raw bytes.
+		size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+		return b.createRuntimeCall("memequal", []llvm.Value{xPtr, yPtr, size}, "eq")
+	case *types.Pointer, *types.Chan:
+		// Pointers and channels: compare as raw pointer-sized bytes.
+		size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+		return b.createRuntimeCall("memequal", []llvm.Value{xPtr, yPtr, size}, "eq")
+	case *types.Interface:
+		// Interface: use runtime interface equality.
+		size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+		return b.createRuntimeCall("hashmapInterfaceEqual", []llvm.Value{xPtr, yPtr, size}, "eq")
+	case *types.Struct:
+		result := llvm.ConstInt(b.ctx.Int1Type(), 1, false) // start with true
+		zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+		for i := 0; i < keyType.NumFields(); i++ {
+			if keyType.Field(i).Name() == "_" {
+				continue // blank fields are ignored in Go equality
+			}
+			fieldType := keyType.Field(i).Type()
+			llvmFieldType := b.getLLVMType(fieldType)
+			if b.targetData.TypeAllocSize(llvmFieldType) == 0 {
+				continue // skip zero-sized fields
+			}
+			idx := llvm.ConstInt(b.ctx.Int32Type(), uint64(i), false)
+			xFieldPtr := b.CreateInBoundsGEP(llvmKeyType, xPtr, []llvm.Value{zero, idx}, "")
+			yFieldPtr := b.CreateInBoundsGEP(llvmKeyType, yPtr, []llvm.Value{zero, idx}, "")
+			fieldEq := b.generateKeyEqual(fieldType, llvmFieldType, xFieldPtr, yFieldPtr, fn)
+			result = b.CreateAnd(result, fieldEq, "")
+		}
+		return result
+	case *types.Array:
+		elemType := keyType.Elem()
+		llvmElemType := b.getLLVMType(elemType)
+		arrayLen := keyType.Len()
+		if hashmapIsBinaryKey(elemType) {
+			// All elements are binary-comparable; compare the entire array.
+			size := llvm.ConstInt(b.uintptrType, b.targetData.TypeAllocSize(llvmKeyType), false)
+			return b.createRuntimeCall("memequal", []llvm.Value{xPtr, yPtr, size}, "eq")
+		}
+		if arrayLen == 0 {
+			return llvm.ConstInt(b.ctx.Int1Type(), 1, false)
+		}
+		if arrayLen <= hashArrayUnrollLimit {
+			result := llvm.ConstInt(b.ctx.Int1Type(), 1, false)
+			zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+			for i := 0; i < int(arrayLen); i++ {
+				idx := llvm.ConstInt(b.uintptrType, uint64(i), false)
+				xElemPtr := b.CreateInBoundsGEP(llvmKeyType, xPtr, []llvm.Value{zero, idx}, "")
+				yElemPtr := b.CreateInBoundsGEP(llvmKeyType, yPtr, []llvm.Value{zero, idx}, "")
+				elemEq := b.generateKeyEqual(elemType, llvmElemType, xElemPtr, yElemPtr, fn)
+				result = b.CreateAnd(result, elemEq, "")
+			}
+			return result
+		}
+		zero := llvm.ConstInt(b.ctx.Int32Type(), 0, false)
+
+		loopEntry := b.GetInsertBlock()
+		loopBody := b.ctx.AddBasicBlock(loopEntry.Parent(), "eq.array.body")
+		loopDone := b.ctx.AddBasicBlock(loopEntry.Parent(), "eq.array.done")
+
+		b.CreateBr(loopBody)
+		b.SetInsertPointAtEnd(loopBody)
+
+		phiI := b.CreatePHI(b.uintptrType, "i")
+
+		xElemPtr := b.CreateInBoundsGEP(llvmKeyType, xPtr, []llvm.Value{zero, phiI}, "")
+		yElemPtr := b.CreateInBoundsGEP(llvmKeyType, yPtr, []llvm.Value{zero, phiI}, "")
+		elemEq := b.generateKeyEqual(elemType, llvmElemType, xElemPtr, yElemPtr, fn)
+
+		nextI := b.CreateAdd(phiI, llvm.ConstInt(b.uintptrType, 1, false), "")
+		atEnd := b.CreateICmp(llvm.IntUGE, nextI, llvm.ConstInt(b.uintptrType, uint64(arrayLen), false), "")
+		exitLoop := b.CreateOr(atEnd, b.CreateNot(elemEq, ""), "")
+		b.CreateCondBr(exitLoop, loopDone, loopBody)
+
+		bodyEnd := b.GetInsertBlock()
+		phiI.AddIncoming([]llvm.Value{llvm.ConstInt(b.uintptrType, 0, false), nextI},
+			[]llvm.BasicBlock{loopEntry, bodyEnd})
+
+		b.SetInsertPointAtEnd(loopDone)
+		return elemEq
+	default:
+		panic(fmt.Sprintf("unhandled key type for equal generation: %T", keyType))
+	}
 }
@@ -29,8 +29,7 @@ func (s *stdSizes) Alignof(T types.Type) int64 {
 		// is the largest of the values unsafe.Alignof(x.f) for each
 		// field f of x, but at least 1."
 		max := int64(1)
-		for i := 0; i < t.NumFields(); i++ {
-			f := t.Field(i)
+		for f := range t.Fields() {
 			if a := s.Alignof(f.Type()); a > max {
 				max = a
 			}
@@ -8,6 +8,7 @@ import (
 	"go/ast"
 	"go/token"
 	"go/types"
+	"slices"
 	"strconv"
 	"strings"

@@ -34,6 +35,7 @@ type functionInfo struct {
 	interrupt     bool       // go:interrupt
 	nobounds      bool       // go:nobounds
 	noescape      bool       // go:noescape
+	noheap        bool       // go:noheap
 	variadic      bool       // go:variadic (CGo only)
 	inline        inlineType // go:inline
 }
@@ -77,24 +79,27 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 		return llvmFn.GlobalValueType(), llvmFn
 	}

-	var retType llvm.Type
-	if fn.Signature.Results() == nil {
-		retType = c.ctx.VoidType()
-	} else if fn.Signature.Results().Len() == 1 {
-		retType = c.getLLVMType(fn.Signature.Results().At(0).Type())
-	} else {
-		results := make([]llvm.Type, 0, fn.Signature.Results().Len())
-		for i := 0; i < fn.Signature.Results().Len(); i++ {
-			results = append(results, c.getLLVMType(fn.Signature.Results().At(i).Type()))
-		}
-		retType = c.ctx.StructType(results, false)
+	retType, indirectResult := c.hasIndirectResult(fn.Signature)
+	if info.exported {
+		indirectResult = false
 	}

 	var paramInfos []paramInfo
+	if indirectResult {
+		paramInfos = append(paramInfos, paramInfo{
+			llvmType: c.dataPtrType,
+			name:     "return",
+			elemSize: c.targetData.TypeAllocSize(retType),
+		})
+		retType = c.ctx.VoidType()
+	}
 	for _, param := range getParams(fn.Signature) {
 		paramType := c.getLLVMType(param.Type())
-		paramFragmentInfos := c.expandFormalParamType(paramType, param.Name(), param.Type())
-		paramInfos = append(paramInfos, paramFragmentInfos...)
+		if info.exported {
+			paramInfos = append(paramInfos, c.expandDirectFormalParamType(paramType, param.Name(), param.Type())...)
+		} else {
+			paramInfos = append(paramInfos, c.expandFormalParamType(paramType, param.Name(), param.Type())...)
+		}
 	}

 	// Add an extra parameter as the function context. This context is used in
@@ -104,12 +109,20 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 	}

 	var paramTypes []llvm.Type
+	hasIndirectABI := indirectResult
 	for _, info := range paramInfos {
 		paramTypes = append(paramTypes, info.llvmType)
+		hasIndirectABI = hasIndirectABI || info.flags&paramIsIndirect != 0
 	}

 	fnType := llvm.FunctionType(retType, paramTypes, info.variadic)
 	llvmFn = llvm.AddFunction(c.mod, info.linkName, fnType)
+	if hasIndirectABI {
+		// Argument promotion only rewrites functions whose uses are all direct
+		// calls. Keep an address use so LLVM cannot reconstruct the large
+		// aggregate signature that this ABI exists to avoid.
+		llvmutil.AppendToGlobal(c.mod, "llvm.used", llvmFn)
+	}
 	if strings.HasPrefix(c.Triple, "wasm") {
 		// C functions without prototypes like this:
 		//   void foo();
@@ -139,9 +152,14 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 			// not.
 			// (It may be safe to add the nocapture parameter to the context
 			// parameter, but I'd like to stay on the safe side here).
-			nocapture := c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0)
+			nocapture := c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0)
 			llvmFn.AddAttributeAtIndex(i+1, nocapture)
 		}
+		if paramInfo.flags&paramIsReadonly != 0 && paramInfo.llvmType.TypeKind() == llvm.PointerTypeKind {
+			// Readonly pointer parameters (like strings) benefit from being marked as readonly.
+			readonly := c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0)
+			llvmFn.AddAttributeAtIndex(i+1, readonly)
+		}
 	}

 	// Set a number of function or parameter attributes, depending on the
@@ -153,8 +171,10 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 		// Mark it as noreturn so LLVM can optimize away code.
 		llvmFn.AddFunctionAttr(c.ctx.CreateEnumAttribute(llvm.AttributeKindID("noreturn"), 0))
 	case "internal/abi.NoEscape":
-		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
-	case "runtime.alloc":
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+	case "machine.keepAliveNoEscape", "machine.unsafeNoEscape":
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+	case "runtime.alloc", "runtime.alloc_noheap", "runtime.alloc_zero":
 		// Tell the optimizer that runtime.alloc is an allocator, meaning that it
 		// returns values that are never null and never alias to an existing value.
 		for _, attrName := range []string{"noalias", "nonnull"} {
@@ -175,25 +195,44 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 	case "runtime.sliceAppend":
 		// Appending a slice will only read the to-be-appended slice, it won't
 		// be modified.
-		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
 		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
-	case "runtime.sliceCopy":
-		// Copying a slice won't capture any of the parameters.
-		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("writeonly"), 0))
-		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
-		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
-		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
 	case "runtime.stringFromBytes":
-		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
 		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
 	case "runtime.stringFromRunes":
-		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
 		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
+	case "runtime.hashmapSet":
+		// The key (param 2) and value (param 3) pointers are only read via
+		// memcpy/hash/equal and are never captured. The indirect calls
+		// through m.keyHash and m.keyEqual function pointers prevent LLVM's
+		// functionattrs pass from inferring this automatically.
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+		llvmFn.AddAttributeAtIndex(3, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+	case "runtime.hashmapGet":
+		// The key (param 2) is read-only and never captured.
+		// The value (param 3) is written to (receives the result) but never captured.
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+		llvmFn.AddAttributeAtIndex(3, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+	case "runtime.hashmapDelete":
+		// The key (param 2) is read-only and never captured.
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+	case "runtime.hashmapGenericSet":
+		// Same as hashmapBinarySet: key (param 2) and value (param 3) are
+		// not captured.
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+		llvmFn.AddAttributeAtIndex(3, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+	case "runtime.hashmapGenericGet":
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+		llvmFn.AddAttributeAtIndex(3, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
+	case "runtime.hashmapGenericDelete":
+		llvmFn.AddAttributeAtIndex(2, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
 	case "runtime.trackPointer":
 		// This function is necessary for tracking pointers on the stack in a
 		// portable way (see gc_stack_portable.go). Indicate to the optimizer
 		// that the only thing we'll do is read the pointer.
-		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("nocapture"), 0))
+		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID(llvmutil.NoCaptureAttrName()), 0))
 		llvmFn.AddAttributeAtIndex(1, c.ctx.CreateEnumAttribute(llvm.AttributeKindID("readonly"), 0))
 	case "__mulsi3", "__divmodsi4", "__udivmodsi4":
 		if strings.Split(c.Triple, "-")[0] == "avr" {
@@ -208,6 +247,12 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)
 			// > circumstances, and should not be exposed to source languages.
 			llvmutil.AppendToGlobal(c.mod, "llvm.compiler.used", llvmFn)
 		}
+	case "GetModuleHandleExA", "GetProcAddress", "GetSystemInfo", "GetSystemTimeAsFileTime", "LoadLibraryExW", "QueryPerformanceCounter", "QueryPerformanceFrequency", "QueryUnbiasedInterruptTime", "SetEnvironmentVariableA", "Sleep", "SystemFunction036", "VirtualAlloc":
+		// On Windows we need to use a special calling convention for some
+		// external calls.
+		if c.GOOS == "windows" && c.GOARCH == "386" {
+			llvmFn.SetFunctionCallConv(llvm.X86StdcallCallConv)
+		}
 	}

 	// External/exported functions may not retain pointer values.
@@ -222,7 +267,7 @@ func (c *compilerContext) getFunction(fn *ssa.Function) (llvm.Type, llvm.Value)

 			llvmFn.AddFunctionAttr(c.ctx.CreateStringAttribute("wasm-import-name", info.wasmName))
 		}
-		nocaptureKind := llvm.AttributeKindID("nocapture")
+		nocaptureKind := llvm.AttributeKindID(llvmutil.NoCaptureAttrName())
 		nocapture := c.ctx.CreateEnumAttribute(nocaptureKind, 0)
 		for i, typ := range paramTypes {
 			if typ.TypeKind() == llvm.PointerTypeKind {
@@ -247,6 +292,23 @@ func (c *compilerContext) maybeCreateSyntheticFunction(fn *ssa.Function, llvmFn
 	// The exception is the package initializer, which does appear in the
 	// *ssa.Package members and so shouldn't be created here.
 	if fn.Synthetic != "" && fn.Synthetic != "package initializer" && fn.Synthetic != "generic function" && fn.Synthetic != "range-over-func yield" {
+		if origin := fn.Origin(); origin != nil && origin.RelString(nil) == "internal/abi.Escape" {
+			// This is a special implementation or internal/abi.Escape, which
+			// can only really be implemented in the compiler.
+			// For simplicity we'll only implement pointer parameters for now.
+			if _, ok := fn.Params[0].Type().Underlying().(*types.Pointer); ok {
+				irbuilder := c.ctx.NewBuilder()
+				defer irbuilder.Dispose()
+				b := newBuilder(c, irbuilder, fn)
+				b.createAbiEscapeImpl()
+				llvmFn.SetLinkage(llvm.LinkOnceODRLinkage)
+				llvmFn.SetUnnamedAddr(true)
+			}
+			// If the parameter is not of a pointer type, it will be left
+			// unimplemented. This will result in a linker error if the function
+			// is really called, making it clear it needs to be implemented.
+			return
+		}
 		if len(fn.Blocks) == 0 {
 			c.addError(fn.Pos(), "missing function body")
 			return
@@ -269,7 +331,7 @@ func (c *compilerContext) getFunctionInfo(f *ssa.Function) functionInfo {
 	}
 	info := functionInfo{
 		// Pick the default linkName.
-		linkName: f.RelString(nil),
+		linkName: c.canonicalFunctionName(f),
 	}

 	// Check for a few runtime functions that are treated specially.
@@ -296,6 +358,18 @@ func (c *compilerContext) getFunctionInfo(f *ssa.Function) functionInfo {
 	return info
 }

+func (c *compilerContext) canonicalFunctionName(f *ssa.Function) string {
+	typeArgs := f.TypeArgs()
+	if len(typeArgs) == 0 {
+		return f.RelString(nil)
+	}
+	parts := make([]string, len(typeArgs))
+	for i, ta := range typeArgs {
+		parts[i], _ = c.getTypeCodeName(ta)
+	}
+	return f.Origin().RelString(nil) + "[" + strings.Join(parts, ",") + "]"
+}
+
 // parsePragmas is used by getFunctionInfo to parse function pragmas such as
 // //export or //go:noinline.
 func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
@@ -322,6 +396,51 @@ func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
 		}
 	}

+	// Also scan file-level //go:linkname directives. These appear as
+	// free-standing comments in *ast.File.Comments (not attached to any
+	// declaration), and are used by modern golang.org/x/sys/unix and others.
+	// Function-attached directives (above) take precedence — we only add
+	// file-level ones if no doc-comment linkname was found for this function.
+	//
+	// TODO: the hasUnsafeImport gate enforced downstream (see the
+	// //go:linkname case below) is package-level. gc enforces it per
+	// file, on the file containing the directive. For file-level
+	// linknames this is more important than for function-attached ones,
+	// because the directive can live in a file separate from the
+	// function. A stricter implementation would check whether the file
+	// returned by fileForFunc imports "unsafe", not whether any file in
+	// the package does.
+	hasFunctionLinkname := false
+	for _, comment := range pragmas {
+		if strings.HasPrefix(comment.Text, "//go:linkname ") {
+			parts := strings.Fields(comment.Text)
+			if len(parts) == 3 && parts[1] == f.Name() {
+				hasFunctionLinkname = true
+				break
+			}
+		}
+	}
+	if !hasFunctionLinkname {
+		if file := c.fileForFunc(f); file != nil {
+			for _, group := range file.Comments {
+				// Skip the function's own doc comment — already handled above.
+				if decl, ok := syntax.(*ast.FuncDecl); ok && group == decl.Doc {
+					continue
+				}
+				for _, comment := range group.List {
+					if !strings.HasPrefix(comment.Text, "//go:linkname ") {
+						continue
+					}
+					parts := strings.Fields(comment.Text)
+					if len(parts) != 3 || parts[1] != f.Name() {
+						continue
+					}
+					pragmas = append(pragmas, comment)
+				}
+			}
+		}
+	}
+
 	// Parse each pragma.
 	for _, comment := range pragmas {
 		parts := strings.Fields(comment.Text)
@@ -339,7 +458,7 @@ func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
 			info.wasmName = info.linkName
 			info.exported = true
 		case "//go:interrupt":
-			if hasUnsafeImport(f.Pkg.Pkg) {
+			if slices.Contains(f.Pkg.Pkg.Imports(), types.Unsafe) {
 				info.interrupt = true
 			}
 		case "//go:wasm-module":
@@ -393,7 +512,7 @@ func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
 			info.inline = inlineHint
 		case "//go:noinline":
 			info.inline = inlineNone
-		case "//go:linkname":
+		case "//go:linkname", "//go:linknamestd":
 			if len(parts) != 3 || parts[1] != f.Name() {
 				continue
 			}
@@ -401,14 +520,14 @@ func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
 			// This is a slightly looser requirement than what gc uses: gc
 			// requires the file to import "unsafe", not the package as a
 			// whole.
-			if hasUnsafeImport(f.Pkg.Pkg) {
+			if slices.Contains(f.Pkg.Pkg.Imports(), types.Unsafe) {
 				info.linkName = parts[2]
 			}
 		case "//go:section":
 			// Only enable go:section when the package imports "unsafe".
 			// go:section also implies go:noinline since inlining could
 			// move the code to a different section than that requested.
-			if len(parts) == 2 && hasUnsafeImport(f.Pkg.Pkg) {
+			if len(parts) == 2 && slices.Contains(f.Pkg.Pkg.Imports(), types.Unsafe) {
 				info.section = parts[1]
 				info.inline = inlineNone
 			}
@@ -417,7 +536,7 @@ func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
 			// runtime functions.
 			// This is somewhat dangerous and thus only imported in packages
 			// that import unsafe.
-			if hasUnsafeImport(f.Pkg.Pkg) {
+			if slices.Contains(f.Pkg.Pkg.Imports(), types.Unsafe) {
 				info.nobounds = true
 			}
 		case "//go:noescape":
@@ -427,6 +546,9 @@ func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
 			if len(f.Blocks) == 0 {
 				info.noescape = true
 			}
+		case "//go:noheap":
+			// Ensure this function does not allocate on the heap.
+			info.noheap = true
 		case "//go:variadic":
 			// The //go:variadic pragma is emitted by the CGo preprocessing
 			// pass for C variadic functions. This includes both explicit
@@ -438,6 +560,10 @@ func (c *compilerContext) parsePragmas(info *functionInfo, f *ssa.Function) {
 			}
 		}
 	}
+
+	if c.Nobounds {
+		info.nobounds = true
+	}
 }

 // Check whether this function can be used in //go:wasmimport or
@@ -510,9 +636,9 @@ func (c *compilerContext) isValidWasmType(typ types.Type, site wasmSite) bool {
 				hasHostLayout = false // package structs added in go1.23
 			}
 		}
-		for i := 0; i < typ.NumFields(); i++ {
-			ftyp := typ.Field(i).Type()
-			if ftyp.String() == "structs.HostLayout" {
+		for field := range typ.Fields() {
+			ftyp := field.Type()
+			if types.Unalias(ftyp).String() == "structs.HostLayout" {
 				hasHostLayout = true
 				continue
 			}
@@ -543,8 +669,8 @@ func getParams(sig *types.Signature) []*types.Var {
 	if sig.Recv() != nil {
 		params = append(params, sig.Recv())
 	}
-	for i := 0; i < sig.Params().Len(); i++ {
-		params = append(params, sig.Params().At(i))
+	for v := range sig.Params().Variables() {
+		params = append(params, v)
 	}
 	return params
 }
@@ -603,12 +729,40 @@ func (c *compilerContext) addStandardAttributes(llvmFn llvm.Value) {
 // linkName is equal to .RelString(nil) on a global and extern is false, but for
 // some symbols this is different (due to //go:extern for example).
 type globalInfo struct {
-	linkName string // go:extern
+	linkName string // go:extern, go:linkname
 	extern   bool   // go:extern
 	align    int    // go:align
 	section  string // go:section
 }

+// fileForFunc returns the *ast.File that contains the declaration of f, or
+// nil if it cannot be determined. File-level pragmas are only consulted for
+// functions in the package currently being compiled — functions imported from
+// other packages have their file-level pragmas processed when those packages
+// are compiled.
+func (c *compilerContext) fileForFunc(f *ssa.Function) *ast.File {
+	if c.loaderPkg == nil || f.Pkg == nil || f.Pkg.Pkg != c.loaderPkg.Pkg {
+		return nil
+	}
+	syntax := f.Syntax()
+	if f.Origin() != nil {
+		syntax = f.Origin().Syntax()
+	}
+	if syntax == nil {
+		return nil
+	}
+	pos := syntax.Pos()
+	if !pos.IsValid() {
+		return nil
+	}
+	for _, file := range c.loaderPkg.Files {
+		if file.FileStart <= pos && pos < file.FileEnd {
+			return file
+		}
+	}
+	return nil
+}
+
 // loadASTComments loads comments on globals from the AST, for use later in the
 // program. In particular, they are required for //go:extern pragmas on globals.
 func (c *compilerContext) loadASTComments(pkg *loader.Package) {
@@ -647,10 +801,7 @@ func (c *compilerContext) getGlobal(g *ssa.Global) llvm.Value {
 		llvmGlobal = llvm.AddGlobal(c.mod, llvmType, info.linkName)

 		// Set alignment from the //go:align comment.
-		alignment := c.targetData.ABITypeAlignment(llvmType)
-		if info.align > alignment {
-			alignment = info.align
-		}
+		alignment := max(info.align, c.targetData.ABITypeAlignment(llvmType))
 		if alignment <= 0 || alignment&(alignment-1) != 0 {
 			// Check for power-of-two (or 0).
 			// See: https://stackoverflow.com/a/108360
@@ -688,14 +839,14 @@ func (c *compilerContext) getGlobalInfo(g *ssa.Global) globalInfo {
 	// Check for //go: pragmas, which may change the link name (among others).
 	doc := c.astComments[info.linkName]
 	if doc != nil {
-		info.parsePragmas(doc)
+		info.parsePragmas(doc, c, g)
 	}
 	return info
 }

 // Parse //go: pragma comments from the source. In particular, it parses the
-// //go:extern pragma on globals.
-func (info *globalInfo) parsePragmas(doc *ast.CommentGroup) {
+// //go:extern and //go:linkname pragmas on globals.
+func (info *globalInfo) parsePragmas(doc *ast.CommentGroup, c *compilerContext, g *ssa.Global) {
 	for _, comment := range doc.List {
 		if !strings.HasPrefix(comment.Text, "//go:") {
 			continue
@@ -716,6 +867,17 @@ func (info *globalInfo) parsePragmas(doc *ast.CommentGroup) {
 			if len(parts) == 2 {
 				info.section = parts[1]
 			}
+		case "//go:linkname":
+			if len(parts) != 3 || parts[1] != g.Name() {
+				continue
+			}
+			// Only enable go:linkname when the package imports "unsafe".
+			// This is a slightly looser requirement than what gc uses: gc
+			// requires the file to import "unsafe", not the package as a
+			// whole.
+			if slices.Contains(g.Pkg.Pkg.Imports(), types.Unsafe) {
+				info.linkName = parts[2]
+			}
 		}
 	}
 }
@@ -729,13 +891,3 @@ func getAllMethods(prog *ssa.Program, typ types.Type) []*types.Selection {
 	}
 	return methods
 }
-
-// Return true if this package imports "unsafe", false otherwise.
-func hasUnsafeImport(pkg *types.Package) bool {
-	for _, imp := range pkg.Imports() {
-		if imp == types.Unsafe {
-			return true
-		}
-	}
-	return false
-}
@@ -26,24 +26,25 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		argTypes := []llvm.Type{b.uintptrType}
 		// Constraints will look something like:
 		//   "={rax},0,{rdi},{rsi},{rdx},{r10},{r8},{r9},~{rcx},~{r11}"
-		constraints := "={rax},0"
+		var constraints strings.Builder
+		constraints.WriteString("={rax},0")
 		for i, arg := range call.Args[1:] {
-			constraints += "," + [...]string{
+			constraints.WriteString("," + [...]string{
 				"{rdi}",
 				"{rsi}",
 				"{rdx}",
 				"{r10}",
 				"{r8}",
 				"{r9}",
-			}[i]
+			}[i])
 			llvmValue := b.getValue(arg, getPos(call))
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
 		// rcx and r11 are clobbered by the syscall, so make sure they are not used
-		constraints += ",~{rcx},~{r11}"
+		constraints.WriteString(",~{rcx},~{r11}")
 		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
-		target := llvm.InlineAsm(fnType, "syscall", constraints, true, false, llvm.InlineAsmDialectIntel, false)
+		target := llvm.InlineAsm(fnType, "syscall", constraints.String(), true, false, llvm.InlineAsmDialectIntel, false)
 		return b.CreateCall(fnType, target, args, ""), nil

 	case b.GOARCH == "386" && b.GOOS == "linux":
@@ -55,34 +56,44 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		argTypes := []llvm.Type{b.uintptrType}
 		// Constraints will look something like:
 		//   "={eax},0,{ebx},{ecx},{edx},{esi},{edi},{ebp}"
-		constraints := "={eax},0"
+		var constraints strings.Builder
+		constraints.WriteString("={eax},0")
 		for i, arg := range call.Args[1:] {
-			constraints += "," + [...]string{
+			constraints.WriteString("," + [...]string{
 				"{ebx}",
 				"{ecx}",
 				"{edx}",
 				"{esi}",
 				"{edi}",
 				"{ebp}",
-			}[i]
+			}[i])
 			llvmValue := b.getValue(arg, getPos(call))
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
 		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
-		target := llvm.InlineAsm(fnType, "int 0x80", constraints, true, false, llvm.InlineAsmDialectIntel, false)
+		target := llvm.InlineAsm(fnType, "int 0x80", constraints.String(), true, false, llvm.InlineAsmDialectIntel, false)
 		return b.CreateCall(fnType, target, args, ""), nil

 	case b.GOARCH == "arm" && b.GOOS == "linux":
+		if arch := b.archFamily(); arch != "arm" {
+			// Some targets pretend to be linux/arm for compatibility but aren't
+			// actually such a system. Make sure we emit an error instead of
+			// creating inline assembly that will fail to compile.
+			// See: https://github.com/tinygo-org/tinygo/issues/4959
+			return llvm.Value{}, b.makeError(call.Pos(), "system calls are not supported: target emulates a linux/arm system on "+arch)
+		}
+
 		// Implement the EABI system call convention for Linux.
 		// Source: syscall(2) man page.
 		args := []llvm.Value{}
 		argTypes := []llvm.Type{}
 		// Constraints will look something like:
 		//   ={r0},0,{r1},{r2},{r7},~{r3}
-		constraints := "={r0}"
+		var constraints strings.Builder
+		constraints.WriteString("={r0}")
 		for i, arg := range call.Args[1:] {
-			constraints += "," + [...]string{
+			constraints.WriteString("," + [...]string{
 				"0", // tie to output
 				"{r1}",
 				"{r2}",
@@ -90,20 +101,20 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 				"{r4}",
 				"{r5}",
 				"{r6}",
-			}[i]
+			}[i])
 			llvmValue := b.getValue(arg, getPos(call))
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
 		args = append(args, num)
 		argTypes = append(argTypes, b.uintptrType)
-		constraints += ",{r7}" // syscall number
+		constraints.WriteString(",{r7}") // syscall number
 		for i := len(call.Args) - 1; i < 4; i++ {
 			// r0-r3 get clobbered after the syscall returns
-			constraints += ",~{r" + strconv.Itoa(i) + "}"
+			constraints.WriteString(",~{r" + strconv.Itoa(i) + "}")
 		}
 		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
-		target := llvm.InlineAsm(fnType, "svc #0", constraints, true, false, 0, false)
+		target := llvm.InlineAsm(fnType, "svc #0", constraints.String(), true, false, 0, false)
 		return b.CreateCall(fnType, target, args, ""), nil

 	case b.GOARCH == "arm64" && b.GOOS == "linux":
@@ -112,31 +123,32 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		argTypes := []llvm.Type{}
 		// Constraints will look something like:
 		//   ={x0},0,{x1},{x2},{x8},~{x3},~{x4},~{x5},~{x6},~{x7},~{x16},~{x17}
-		constraints := "={x0}"
+		var constraints strings.Builder
+		constraints.WriteString("={x0}")
 		for i, arg := range call.Args[1:] {
-			constraints += "," + [...]string{
+			constraints.WriteString("," + [...]string{
 				"0", // tie to output
 				"{x1}",
 				"{x2}",
 				"{x3}",
 				"{x4}",
 				"{x5}",
-			}[i]
+			}[i])
 			llvmValue := b.getValue(arg, getPos(call))
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
 		}
 		args = append(args, num)
 		argTypes = append(argTypes, b.uintptrType)
-		constraints += ",{x8}" // syscall number
+		constraints.WriteString(",{x8}") // syscall number
 		for i := len(call.Args) - 1; i < 8; i++ {
 			// x0-x7 may get clobbered during the syscall following the aarch64
 			// calling convention.
-			constraints += ",~{x" + strconv.Itoa(i) + "}"
+			constraints.WriteString(",~{x" + strconv.Itoa(i) + "}")
 		}
-		constraints += ",~{x16},~{x17}" // scratch registers
+		constraints.WriteString(",~{x16},~{x17}") // scratch registers
 		fnType := llvm.FunctionType(b.uintptrType, argTypes, false)
-		target := llvm.InlineAsm(fnType, "svc #0", constraints, true, false, 0, false)
+		target := llvm.InlineAsm(fnType, "svc #0", constraints.String(), true, false, 0, false)
 		return b.CreateCall(fnType, target, args, ""), nil

 	case (b.GOARCH == "mips" || b.GOARCH == "mipsle") && b.GOOS == "linux":
@@ -155,7 +167,8 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		// faster and smaller code.
 		args := []llvm.Value{num}
 		argTypes := []llvm.Type{b.uintptrType}
-		constraints := "={$2},={$7},0"
+		var constraints strings.Builder
+		constraints.WriteString("={$2},={$7},0")
 		syscallParams := call.Args[1:]
 		if len(syscallParams) > 7 {
 			// There is one syscall that uses 7 parameters: sync_file_range.
@@ -164,7 +177,7 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 			syscallParams = syscallParams[:7]
 		}
 		for i, arg := range syscallParams {
-			constraints += "," + [...]string{
+			constraints.WriteString("," + [...]string{
 				"{$4}", // arg1
 				"{$5}", // arg2
 				"{$6}", // arg3
@@ -172,7 +185,7 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 				"r",    // arg5 on the stack
 				"r",    // arg6 on the stack
 				"r",    // arg7 on the stack
-			}[i]
+			}[i])
 			llvmValue := b.getValue(arg, getPos(call))
 			args = append(args, llvmValue)
 			argTypes = append(argTypes, llvmValue.Type())
@@ -213,10 +226,10 @@ func (b *builder) createRawSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 				"addu $$sp, $$sp, 32\n" +
 				".set at\n"
 		}
-		constraints += ",~{$3},~{$4},~{$5},~{$6},~{$8},~{$9},~{$10},~{$11},~{$12},~{$13},~{$14},~{$15},~{$24},~{$25},~{hi},~{lo},~{memory}"
+		constraints.WriteString(",~{$3},~{$4},~{$5},~{$6},~{$8},~{$9},~{$10},~{$11},~{$12},~{$13},~{$14},~{$15},~{$24},~{$25},~{hi},~{lo},~{memory}")
 		returnType := b.ctx.StructType([]llvm.Type{b.uintptrType, b.uintptrType}, false)
 		fnType := llvm.FunctionType(returnType, argTypes, false)
-		target := llvm.InlineAsm(fnType, asm, constraints, true, true, 0, false)
+		target := llvm.InlineAsm(fnType, asm, constraints.String(), true, true, 0, false)
 		call := b.CreateCall(fnType, target, args, "")
 		resultCode := b.CreateExtractValue(call, 0, "") // r2
 		errorFlag := b.CreateExtractValue(call, 1, "")  // r7
@@ -268,6 +281,8 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		// The signature looks like this:
 		//   func Syscall(trap, nargs, a1, a2, a3 uintptr) (r1, r2 uintptr, err Errno)

+		isI386 := strings.HasPrefix(b.Triple, "i386-")
+
 		// Prepare input values.
 		var paramTypes []llvm.Type
 		var params []llvm.Value
@@ -285,11 +300,17 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		if setLastError.IsNil() {
 			llvmType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.ctx.Int32Type()}, false)
 			setLastError = llvm.AddFunction(b.mod, "SetLastError", llvmType)
+			if isI386 {
+				setLastError.SetFunctionCallConv(llvm.X86StdcallCallConv)
+			}
 		}
 		getLastError := b.mod.NamedFunction("GetLastError")
 		if getLastError.IsNil() {
 			llvmType := llvm.FunctionType(b.ctx.Int32Type(), nil, false)
 			getLastError = llvm.AddFunction(b.mod, "GetLastError", llvmType)
+			if isI386 {
+				getLastError.SetFunctionCallConv(llvm.X86StdcallCallConv)
+			}
 		}

 		// Now do the actual call. Pseudocode:
@@ -300,9 +321,24 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 		// Note that SetLastError/GetLastError could be replaced with direct
 		// access to the thread control block, which is probably smaller and
 		// faster. The Go runtime does this in assembly.
-		b.CreateCall(setLastError.GlobalValueType(), setLastError, []llvm.Value{llvm.ConstNull(b.ctx.Int32Type())}, "")
+		// On windows/386, we also need to save/restore the stack pointer. I'm
+		// not entirely sure why this is needed, but without it these calls
+		// change the stack pointer leading to a crash soon after.
+		setLastErrorCall := b.CreateCall(setLastError.GlobalValueType(), setLastError, []llvm.Value{llvm.ConstNull(b.ctx.Int32Type())}, "")
+		var sp llvm.Value
+		if isI386 {
+			setLastErrorCall.SetInstructionCallConv(llvm.X86StdcallCallConv)
+			sp = b.readStackPointer()
+		}
 		syscallResult := b.CreateCall(llvmType, fnPtr, params, "")
+		if isI386 {
+			syscallResult.SetInstructionCallConv(llvm.X86StdcallCallConv)
+			b.writeStackPointer(sp)
+		}
 		errResult := b.CreateCall(getLastError.GlobalValueType(), getLastError, nil, "err")
+		if isI386 {
+			errResult.SetInstructionCallConv(llvm.X86StdcallCallConv)
+		}
 		if b.uintptrType != b.ctx.Int32Type() {
 			errResult = b.CreateZExt(errResult, b.uintptrType, "err.uintptr")
 		}
@@ -318,6 +354,130 @@ func (b *builder) createSyscall(call *ssa.CallCommon) (llvm.Value, error) {
 	}
 }

+// createSyscalln emits instructions for the syscall.syscalln function on
+// Windows. This handles the variadic calling convention used in Go 1.26+:
+//
+//	func syscalln(fn, n uintptr, args ...uintptr) (r1, r2 uintptr, err Errno)
+//
+// The function generates a switch on n to dispatch to the correct fixed-argument
+// function pointer call with SetLastError(0)/GetLastError() wrapping.
+func (b *builder) createSyscalln(call *ssa.CallCommon) (llvm.Value, error) {
+	const maxArgs = 18 // Windows syscalls support up to 18 args
+
+	isI386 := strings.HasPrefix(b.Triple, "i386-")
+
+	// Get the function pointer (call.Args[0]) and n (call.Args[1]).
+	fn := b.getValue(call.Args[0], getPos(call))
+	fnPtr := b.CreateIntToPtr(fn, b.dataPtrType, "")
+	n := b.getValue(call.Args[1], getPos(call))
+
+	// Get the variadic args slice (call.Args[2]).
+	// In SSA, the variadic slice is the third argument.
+	var argsPtr llvm.Value
+	if len(call.Args) > 2 {
+		argsSlice := b.getValue(call.Args[2], getPos(call))
+		argsPtr = b.CreateExtractValue(argsSlice, 0, "args.data")
+	} else {
+		argsPtr = llvm.ConstNull(b.dataPtrType)
+	}
+
+	// Prepare SetLastError and GetLastError.
+	setLastError := b.mod.NamedFunction("SetLastError")
+	if setLastError.IsNil() {
+		llvmType := llvm.FunctionType(b.ctx.VoidType(), []llvm.Type{b.ctx.Int32Type()}, false)
+		setLastError = llvm.AddFunction(b.mod, "SetLastError", llvmType)
+		if isI386 {
+			setLastError.SetFunctionCallConv(llvm.X86StdcallCallConv)
+		}
+	}
+	getLastError := b.mod.NamedFunction("GetLastError")
+	if getLastError.IsNil() {
+		llvmType := llvm.FunctionType(b.ctx.Int32Type(), nil, false)
+		getLastError = llvm.AddFunction(b.mod, "GetLastError", llvmType)
+		if isI386 {
+			getLastError.SetFunctionCallConv(llvm.X86StdcallCallConv)
+		}
+	}
+
+	retType := b.ctx.StructType([]llvm.Type{b.uintptrType, b.uintptrType, b.uintptrType}, false)
+
+	// Create the merge block where all cases converge.
+	mergeBB := b.insertBasicBlock("syscalln.merge")
+
+	// Create the default (panic) block.
+	panicBB := b.insertBasicBlock("syscalln.panic")
+
+	// Create the switch on n.
+	sw := b.CreateSwitch(n, panicBB, maxArgs+1)
+
+	// We'll collect blocks and values for the PHI node.
+	var incomingVals []llvm.Value
+	var incomingBlocks []llvm.BasicBlock
+
+	// Generate a case for each arg count 0..maxArgs.
+	for i := 0; i <= maxArgs; i++ {
+		caseBB := b.insertBasicBlock("syscalln.case" + strconv.Itoa(i))
+		sw.AddCase(llvm.ConstInt(b.uintptrType, uint64(i), false), caseBB)
+		b.SetInsertPointAtEnd(caseBB)
+
+		// Load args[0] through args[i-1] from the slice data pointer.
+		var params []llvm.Value
+		var paramTypes []llvm.Type
+		for j := 0; j < i; j++ {
+			gep := b.CreateInBoundsGEP(b.uintptrType, argsPtr, []llvm.Value{
+				llvm.ConstInt(b.ctx.Int32Type(), uint64(j), false),
+			}, "")
+			arg := b.CreateLoad(b.uintptrType, gep, "")
+			params = append(params, arg)
+			paramTypes = append(paramTypes, b.uintptrType)
+		}
+
+		// SetLastError(0)
+		setCall := b.CreateCall(setLastError.GlobalValueType(), setLastError, []llvm.Value{llvm.ConstNull(b.ctx.Int32Type())}, "")
+		var sp llvm.Value
+		if isI386 {
+			setCall.SetInstructionCallConv(llvm.X86StdcallCallConv)
+			sp = b.readStackPointer()
+		}
+
+		// Call fn(args...)
+		fnType := llvm.FunctionType(b.uintptrType, paramTypes, false)
+		syscallResult := b.CreateCall(fnType, fnPtr, params, "")
+		if isI386 {
+			syscallResult.SetInstructionCallConv(llvm.X86StdcallCallConv)
+			b.writeStackPointer(sp)
+		}
+
+		// err = GetLastError()
+		errResult := b.CreateCall(getLastError.GlobalValueType(), getLastError, nil, "err")
+		if isI386 {
+			errResult.SetInstructionCallConv(llvm.X86StdcallCallConv)
+		}
+		if b.uintptrType != b.ctx.Int32Type() {
+			errResult = b.CreateZExt(errResult, b.uintptrType, "err.uintptr")
+		}
+
+		// Build {r1, 0, err}
+		result := llvm.ConstNull(retType)
+		result = b.CreateInsertValue(result, syscallResult, 0, "")
+		result = b.CreateInsertValue(result, errResult, 2, "")
+
+		incomingVals = append(incomingVals, result)
+		incomingBlocks = append(incomingBlocks, b.Builder.GetInsertBlock())
+		b.CreateBr(mergeBB)
+	}
+
+	// Panic block for n > maxArgs.
+	b.SetInsertPointAtEnd(panicBB)
+	b.CreateUnreachable()
+
+	// Merge block: PHI node to select the result.
+	b.SetInsertPointAtEnd(mergeBB)
+	phi := b.CreatePHI(retType, "syscalln.result")
+	phi.AddIncoming(incomingVals, incomingBlocks)
+	return phi, nil
+}
+
 // createRawSyscallNoError emits instructions for the Linux-specific
 // syscall.rawSyscallNoError function.
 func (b *builder) createRawSyscallNoError(call *ssa.CallCommon) (llvm.Value, error) {
@@ -66,6 +66,11 @@ func complexSub(x, y complex64) complex64 {
 	return x - y
 }

+func shiftNested(x uint64) uint64 {
+	k := 3
+	return x >> (1 << k) // https://github.com/tinygo-org/tinygo/issues/5496
+}
+
 func complexMul(x, y complex64) complex64 {
 	return x * y
 }
@@ -1,6 +1,6 @@
 ; ModuleID = 'basic.go'
 source_filename = "basic.go"
-target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-i128:128-n32:64-S128-ni:1:10:20"
 target triple = "wasm32-unknown-wasi"

 %main.kv = type { float, i32, i32, i32 }
@@ -10,33 +10,30 @@ target triple = "wasm32-unknown-wasi"
@main.a = hidden global { ptr, i32, i32 } zeroinitializer, align 4
@main.b = hidden global [2 x ptr] zeroinitializer, align 4

-; Function Attrs: allockind("alloc,zeroed") allocsize(0)
-declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #0
-
-declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #1
+declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.init(ptr %context) unnamed_addr #2 {
+define hidden void @main.init(ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

 ; Function Attrs: nounwind
-define hidden i32 @main.addInt(i32 %x, i32 %y, ptr %context) unnamed_addr #2 {
+define hidden i32 @main.addInt(i32 %x, i32 %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = add i32 %x, %y
  ret i32 %0
 }

 ; Function Attrs: nounwind
-define hidden i1 @main.equalInt(i32 %x, i32 %y, ptr %context) unnamed_addr #2 {
+define hidden i1 @main.equalInt(i32 %x, i32 %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = icmp eq i32 %x, %y
  ret i1 %0
 }

 ; Function Attrs: nounwind
-define hidden i32 @main.divInt(i32 %x, i32 %y, ptr %context) unnamed_addr #2 {
+define hidden i32 @main.divInt(i32 %x, i32 %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = icmp eq i32 %y, 0
  br i1 %0, label %divbyzero.throw, label %divbyzero.next
@@ -50,14 +47,14 @@ divbyzero.next:                                   ; preds = %entry
  ret i32 %5

 divbyzero.throw:                                  ; preds = %entry
-  call void @runtime.divideByZeroPanic(ptr undef) #3
+  call void @runtime.divideByZeroPanic(ptr undef) #2
  unreachable
 }

-declare void @runtime.divideByZeroPanic(ptr) #1
+declare void @runtime.divideByZeroPanic(ptr) #0

 ; Function Attrs: nounwind
-define hidden i32 @main.divUint(i32 %x, i32 %y, ptr %context) unnamed_addr #2 {
+define hidden i32 @main.divUint(i32 %x, i32 %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = icmp eq i32 %y, 0
  br i1 %0, label %divbyzero.throw, label %divbyzero.next
@@ -67,12 +64,12 @@ divbyzero.next:                                   ; preds = %entry
  ret i32 %1

 divbyzero.throw:                                  ; preds = %entry
-  call void @runtime.divideByZeroPanic(ptr undef) #3
+  call void @runtime.divideByZeroPanic(ptr undef) #2
  unreachable
 }

 ; Function Attrs: nounwind
-define hidden i32 @main.remInt(i32 %x, i32 %y, ptr %context) unnamed_addr #2 {
+define hidden i32 @main.remInt(i32 %x, i32 %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = icmp eq i32 %y, 0
  br i1 %0, label %divbyzero.throw, label %divbyzero.next
@@ -86,12 +83,12 @@ divbyzero.next:                                   ; preds = %entry
  ret i32 %5

 divbyzero.throw:                                  ; preds = %entry
-  call void @runtime.divideByZeroPanic(ptr undef) #3
+  call void @runtime.divideByZeroPanic(ptr undef) #2
  unreachable
 }

 ; Function Attrs: nounwind
-define hidden i32 @main.remUint(i32 %x, i32 %y, ptr %context) unnamed_addr #2 {
+define hidden i32 @main.remUint(i32 %x, i32 %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = icmp eq i32 %y, 0
  br i1 %0, label %divbyzero.throw, label %divbyzero.next
@@ -101,66 +98,66 @@ divbyzero.next:                                   ; preds = %entry
  ret i32 %1

 divbyzero.throw:                                  ; preds = %entry
-  call void @runtime.divideByZeroPanic(ptr undef) #3
+  call void @runtime.divideByZeroPanic(ptr undef) #2
  unreachable
 }

 ; Function Attrs: nounwind
-define hidden i1 @main.floatEQ(float %x, float %y, ptr %context) unnamed_addr #2 {
+define hidden i1 @main.floatEQ(float %x, float %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fcmp oeq float %x, %y
  ret i1 %0
 }

 ; Function Attrs: nounwind
-define hidden i1 @main.floatNE(float %x, float %y, ptr %context) unnamed_addr #2 {
+define hidden i1 @main.floatNE(float %x, float %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fcmp une float %x, %y
  ret i1 %0
 }

 ; Function Attrs: nounwind
-define hidden i1 @main.floatLower(float %x, float %y, ptr %context) unnamed_addr #2 {
+define hidden i1 @main.floatLower(float %x, float %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fcmp olt float %x, %y
  ret i1 %0
 }

 ; Function Attrs: nounwind
-define hidden i1 @main.floatLowerEqual(float %x, float %y, ptr %context) unnamed_addr #2 {
+define hidden i1 @main.floatLowerEqual(float %x, float %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fcmp ole float %x, %y
  ret i1 %0
 }

 ; Function Attrs: nounwind
-define hidden i1 @main.floatGreater(float %x, float %y, ptr %context) unnamed_addr #2 {
+define hidden i1 @main.floatGreater(float %x, float %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fcmp ogt float %x, %y
  ret i1 %0
 }

 ; Function Attrs: nounwind
-define hidden i1 @main.floatGreaterEqual(float %x, float %y, ptr %context) unnamed_addr #2 {
+define hidden i1 @main.floatGreaterEqual(float %x, float %y, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fcmp oge float %x, %y
  ret i1 %0
 }

 ; Function Attrs: nounwind
-define hidden float @main.complexReal(float %x.r, float %x.i, ptr %context) unnamed_addr #2 {
+define hidden float @main.complexReal(float %x.r, float %x.i, ptr %context) unnamed_addr #1 {
 entry:
  ret float %x.r
 }

 ; Function Attrs: nounwind
-define hidden float @main.complexImag(float %x.r, float %x.i, ptr %context) unnamed_addr #2 {
+define hidden float @main.complexImag(float %x.r, float %x.i, ptr %context) unnamed_addr #1 {
 entry:
  ret float %x.i
 }

 ; Function Attrs: nounwind
-define hidden { float, float } @main.complexAdd(float %x.r, float %x.i, float %y.r, float %y.i, ptr %context) unnamed_addr #2 {
+define hidden { float, float } @main.complexAdd(float %x.r, float %x.i, float %y.r, float %y.i, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fadd float %x.r, %y.r
  %1 = fadd float %x.i, %y.i
@@ -170,7 +167,7 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden { float, float } @main.complexSub(float %x.r, float %x.i, float %y.r, float %y.i, ptr %context) unnamed_addr #2 {
+define hidden { float, float } @main.complexSub(float %x.r, float %x.i, float %y.r, float %y.i, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fsub float %x.r, %y.r
  %1 = fsub float %x.i, %y.i
@@ -180,7 +177,14 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden { float, float } @main.complexMul(float %x.r, float %x.i, float %y.r, float %y.i, ptr %context) unnamed_addr #2 {
+define hidden i64 @main.shiftNested(i64 %x, ptr %context) unnamed_addr #1 {
+entry:
+  %0 = lshr i64 %x, 8
+  ret i64 %0
+}
+
+; Function Attrs: nounwind
+define hidden { float, float } @main.complexMul(float %x.r, float %x.i, float %y.r, float %y.i, ptr %context) unnamed_addr #1 {
 entry:
  %0 = fmul float %x.r, %y.r
  %1 = fmul float %x.i, %y.i
@@ -194,19 +198,18 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden void @main.foo(ptr %context) unnamed_addr #2 {
+define hidden void @main.foo(ptr %context) unnamed_addr #1 {
 entry:
  call void @"main.foo$1"(%main.kv.0 zeroinitializer, ptr undef)
  ret void
 }

 ; Function Attrs: nounwind
-define internal void @"main.foo$1"(%main.kv.0 %b, ptr %context) unnamed_addr #2 {
+define internal void @"main.foo$1"(%main.kv.0 %b, ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

-attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #1 = { "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #2 = { nounwind "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #3 = { nounwind }
+attributes #0 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #1 = { nounwind "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #2 = { nounwind }
@@ -1,95 +1,92 @@
 ; ModuleID = 'channel.go'
 source_filename = "channel.go"
-target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-i128:128-n32:64-S128-ni:1:10:20"
 target triple = "wasm32-unknown-wasi"

 %runtime.channelOp = type { ptr, ptr, i32, ptr }
 %runtime.chanSelectState = type { ptr, ptr }

-; Function Attrs: allockind("alloc,zeroed") allocsize(0)
-declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #0
-
-declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #1
+declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.init(ptr %context) unnamed_addr #2 {
+define hidden void @main.init(ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

 ; Function Attrs: nounwind
-define hidden void @main.chanIntSend(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
+define hidden void @main.chanIntSend(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
 entry:
  %chan.op = alloca %runtime.channelOp, align 8
  %chan.value = alloca i32, align 4
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %chan.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.value)
  store i32 3, ptr %chan.value, align 4
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
-  call void @runtime.chanSend(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %chan.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
+  call void @runtime.chanSend(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #3
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.value)
  ret void
 }

 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.start.p0(i64 immarg, ptr nocapture) #3
+declare void @llvm.lifetime.start.p0(ptr nocapture) #2

-declare void @runtime.chanSend(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #1
+declare void @runtime.chanSend(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #0

 ; Function Attrs: nocallback nofree nosync nounwind willreturn memory(argmem: readwrite)
-declare void @llvm.lifetime.end.p0(i64 immarg, ptr nocapture) #3
+declare void @llvm.lifetime.end.p0(ptr nocapture) #2

 ; Function Attrs: nounwind
-define hidden void @main.chanIntRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
+define hidden void @main.chanIntRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
 entry:
  %chan.op = alloca %runtime.channelOp, align 8
  %chan.value = alloca i32, align 4
-  call void @llvm.lifetime.start.p0(i64 4, ptr nonnull %chan.value)
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
-  %0 = call i1 @runtime.chanRecv(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 4, ptr nonnull %chan.value)
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.value)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
+  %0 = call i1 @runtime.chanRecv(ptr %ch, ptr nonnull %chan.value, ptr nonnull %chan.op, ptr undef) #3
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.value)
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
  ret void
 }

-declare i1 @runtime.chanRecv(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #1
+declare i1 @runtime.chanRecv(ptr dereferenceable_or_null(36), ptr, ptr dereferenceable_or_null(16), ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.chanZeroSend(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
+define hidden void @main.chanZeroSend(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
 entry:
  %chan.op = alloca %runtime.channelOp, align 8
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
-  call void @runtime.chanSend(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
+  call void @runtime.chanSend(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #3
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
  ret void
 }

 ; Function Attrs: nounwind
-define hidden void @main.chanZeroRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #2 {
+define hidden void @main.chanZeroRecv(ptr dereferenceable_or_null(36) %ch, ptr %context) unnamed_addr #1 {
 entry:
  %chan.op = alloca %runtime.channelOp, align 8
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %chan.op)
-  %0 = call i1 @runtime.chanRecv(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %chan.op)
+  call void @llvm.lifetime.start.p0(ptr nonnull %chan.op)
+  %0 = call i1 @runtime.chanRecv(ptr %ch, ptr null, ptr nonnull %chan.op, ptr undef) #3
+  call void @llvm.lifetime.end.p0(ptr nonnull %chan.op)
  ret void
 }

 ; Function Attrs: nounwind
-define hidden void @main.selectZeroRecv(ptr dereferenceable_or_null(36) %ch1, ptr dereferenceable_or_null(36) %ch2, ptr %context) unnamed_addr #2 {
+define hidden void @main.selectZeroRecv(ptr dereferenceable_or_null(36) %ch1, ptr dereferenceable_or_null(36) %ch2, ptr %context) unnamed_addr #1 {
 entry:
  %select.states.alloca = alloca [2 x %runtime.chanSelectState], align 8
  %select.send.value = alloca i32, align 4
  store i32 1, ptr %select.send.value, align 4
-  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %select.states.alloca)
+  call void @llvm.lifetime.start.p0(ptr nonnull %select.states.alloca)
  store ptr %ch1, ptr %select.states.alloca, align 4
-  %select.states.alloca.repack1 = getelementptr inbounds i8, ptr %select.states.alloca, i32 4
+  %select.states.alloca.repack1 = getelementptr inbounds nuw i8, ptr %select.states.alloca, i32 4
  store ptr %select.send.value, ptr %select.states.alloca.repack1, align 4
-  %0 = getelementptr inbounds i8, ptr %select.states.alloca, i32 8
+  %0 = getelementptr inbounds nuw i8, ptr %select.states.alloca, i32 8
  store ptr %ch2, ptr %0, align 4
-  %.repack3 = getelementptr inbounds i8, ptr %select.states.alloca, i32 12
+  %.repack3 = getelementptr inbounds nuw i8, ptr %select.states.alloca, i32 12
  store ptr null, ptr %.repack3, align 4
-  %select.result = call { i32, i1 } @runtime.chanSelect(ptr undef, ptr nonnull %select.states.alloca, i32 2, i32 2, ptr null, i32 0, i32 0, ptr undef) #4
-  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %select.states.alloca)
+  %select.result = call { i32, i1 } @runtime.chanSelect(ptr undef, ptr nonnull %select.states.alloca, i32 2, i32 2, ptr null, i32 0, i32 0, ptr undef) #3
+  call void @llvm.lifetime.end.p0(ptr nonnull %select.states.alloca)
  %1 = extractvalue { i32, i1 } %select.result, 0
  %2 = icmp eq i32 %1, 0
  br i1 %2, label %select.done, label %select.next
@@ -105,10 +102,9 @@ select.body:                                      ; preds = %select.next
  br label %select.done
 }

-declare { i32, i1 } @runtime.chanSelect(ptr, ptr, i32, i32, ptr, i32, i32, ptr) #1
+declare { i32, i1 } @runtime.chanSelect(ptr, ptr, i32, i32, ptr, i32, i32, ptr) #0

-attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #1 = { "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #2 = { nounwind "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #3 = { nocallback nofree nosync nounwind willreturn memory(argmem: readwrite) }
-attributes #4 = { nounwind }
+attributes #0 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #1 = { nounwind "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #2 = { nocallback nofree nosync nounwind willreturn memory(argmem: readwrite) }
+attributes #3 = { nounwind }
@@ -3,32 +3,29 @@ source_filename = "defer.go"
 target datalayout = "e-m:e-p:32:32-Fi8-i64:64-v128:64:128-a:0:32-n32-S64"
 target triple = "thumbv7m-unknown-unknown-eabi"

-%runtime.deferFrame = type { ptr, ptr, [0 x ptr], ptr, i8, %runtime._interface }
+%runtime.deferFrame = type { ptr, ptr, [0 x ptr], ptr, i8, %runtime._interface, ptr }
 %runtime._interface = type { ptr, ptr }

-; Function Attrs: allockind("alloc,zeroed") allocsize(0)
-declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #0
-
 ; Function Attrs: nounwind
-define hidden void @main.init(ptr %context) unnamed_addr #1 {
+define hidden void @main.init(ptr %context) unnamed_addr #0 {
 entry:
  ret void
 }

-declare void @main.external(ptr) #2
+declare void @main.external(ptr) #1

 ; Function Attrs: nounwind
-define hidden void @main.deferSimple(ptr %context) unnamed_addr #1 {
+define hidden void @main.deferSimple(ptr %context) unnamed_addr #0 {
 entry:
  %defer.alloca = alloca { i32, ptr }, align 4
-  %deferPtr = alloca ptr, align 4
-  store ptr null, ptr %deferPtr, align 4
  %deferframe.buf = alloca %runtime.deferFrame, align 4
+  %deferPtr = getelementptr inbounds nuw i8, ptr %deferframe.buf, i32 24
  %0 = call ptr @llvm.stacksave.p0()
  call void @runtime.setupDeferFrame(ptr nonnull %deferframe.buf, ptr %0, ptr undef) #4
+  %defer.next = load ptr, ptr %deferPtr, align 4
  store i32 0, ptr %defer.alloca, align 4
-  %defer.alloca.repack15 = getelementptr inbounds i8, ptr %defer.alloca, i32 4
-  store ptr null, ptr %defer.alloca.repack15, align 4
+  %defer.alloca.repack15 = getelementptr inbounds nuw i8, ptr %defer.alloca, i32 4
+  store ptr %defer.next, ptr %defer.alloca.repack15, align 4
  store ptr %defer.alloca, ptr %deferPtr, align 4
  %setjmp = call i32 asm "\0Amovs r0, #0\0Amov r2, pc\0Astr r2, [r1, #4]", "={r0},{r1},~{r1},~{r2},~{r3},~{r4},~{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{lr},~{q0},~{q1},~{q2},~{q3},~{q4},~{q5},~{q6},~{q7},~{q8},~{q9},~{q10},~{q11},~{q12},~{q13},~{q14},~{q15},~{cpsr},~{memory}"(ptr nonnull %deferframe.buf) #5
  %setjmp.result = icmp eq i32 %setjmp, 0
@@ -51,7 +48,7 @@ rundefers.loophead:                               ; preds = %3, %rundefers.block
  br i1 %stackIsNil, label %rundefers.end, label %rundefers.loop

 rundefers.loop:                                   ; preds = %rundefers.loophead
-  %stack.next.gep = getelementptr inbounds i8, ptr %2, i32 4
+  %stack.next.gep = getelementptr inbounds nuw i8, ptr %2, i32 4
  %stack.next = load ptr, ptr %stack.next.gep, align 4
  store ptr %stack.next, ptr %deferPtr, align 4
  %callback = load i32, ptr %2, align 4
@@ -87,7 +84,7 @@ rundefers.loophead6:                              ; preds = %5, %lpad
  br i1 %stackIsNil7, label %rundefers.end3, label %rundefers.loop5

 rundefers.loop5:                                  ; preds = %rundefers.loophead6
-  %stack.next.gep8 = getelementptr inbounds i8, ptr %4, i32 4
+  %stack.next.gep8 = getelementptr inbounds nuw i8, ptr %4, i32 4
  %stack.next9 = load ptr, ptr %stack.next.gep8, align 4
  store ptr %stack.next9, ptr %deferPtr, align 4
  %callback11 = load i32, ptr %4, align 4
@@ -112,14 +109,14 @@ rundefers.end3:                                   ; preds = %rundefers.loophead6
 }

 ; Function Attrs: nocallback nofree nosync nounwind willreturn
-declare ptr @llvm.stacksave.p0() #3
+declare ptr @llvm.stacksave.p0() #2

-declare void @runtime.setupDeferFrame(ptr dereferenceable_or_null(24), ptr, ptr) #2
+declare void @runtime.setupDeferFrame(ptr dereferenceable_or_null(28), ptr, ptr) #1

-declare void @runtime.destroyDeferFrame(ptr dereferenceable_or_null(24), ptr) #2
+declare void @runtime.destroyDeferFrame(ptr dereferenceable_or_null(28), ptr) #1

 ; Function Attrs: nounwind
-define internal void @"main.deferSimple$1"(ptr %context) unnamed_addr #1 {
+define internal void @"main.deferSimple$1"(ptr %context) unnamed_addr #0 {
 entry:
  call void @runtime.printlock(ptr undef) #4
  call void @runtime.printint32(i32 3, ptr undef) #4
@@ -127,29 +124,29 @@ entry:
  ret void
 }

-declare void @runtime.printlock(ptr) #2
+declare void @runtime.printlock(ptr) #1

-declare void @runtime.printint32(i32, ptr) #2
+declare void @runtime.printint32(i32, ptr) #1

-declare void @runtime.printunlock(ptr) #2
+declare void @runtime.printunlock(ptr) #1

 ; Function Attrs: nounwind
-define hidden void @main.deferMultiple(ptr %context) unnamed_addr #1 {
+define hidden void @main.deferMultiple(ptr %context) unnamed_addr #0 {
 entry:
  %defer.alloca2 = alloca { i32, ptr }, align 4
  %defer.alloca = alloca { i32, ptr }, align 4
-  %deferPtr = alloca ptr, align 4
-  store ptr null, ptr %deferPtr, align 4
  %deferframe.buf = alloca %runtime.deferFrame, align 4
+  %deferPtr = getelementptr inbounds nuw i8, ptr %deferframe.buf, i32 24
  %0 = call ptr @llvm.stacksave.p0()
  call void @runtime.setupDeferFrame(ptr nonnull %deferframe.buf, ptr %0, ptr undef) #4
+  %defer.next = load ptr, ptr %deferPtr, align 4
  store i32 0, ptr %defer.alloca, align 4
-  %defer.alloca.repack22 = getelementptr inbounds i8, ptr %defer.alloca, i32 4
-  store ptr null, ptr %defer.alloca.repack22, align 4
+  %defer.alloca.repack22 = getelementptr inbounds nuw i8, ptr %defer.alloca, i32 4
+  store ptr %defer.next, ptr %defer.alloca.repack22, align 4
  store ptr %defer.alloca, ptr %deferPtr, align 4
  store i32 1, ptr %defer.alloca2, align 4
-  %defer.alloca2.repack23 = getelementptr inbounds i8, ptr %defer.alloca2, i32 4
-  store ptr %defer.alloca, ptr %defer.alloca2.repack23, align 4
+  %defer.alloca2.repack24 = getelementptr inbounds nuw i8, ptr %defer.alloca2, i32 4
+  store ptr %defer.alloca, ptr %defer.alloca2.repack24, align 4
  store ptr %defer.alloca2, ptr %deferPtr, align 4
  %setjmp = call i32 asm "\0Amovs r0, #0\0Amov r2, pc\0Astr r2, [r1, #4]", "={r0},{r1},~{r1},~{r2},~{r3},~{r4},~{r5},~{r6},~{r7},~{r8},~{r9},~{r10},~{r11},~{r12},~{lr},~{q0},~{q1},~{q2},~{q3},~{q4},~{q5},~{q6},~{q7},~{q8},~{q9},~{q10},~{q11},~{q12},~{q13},~{q14},~{q15},~{cpsr},~{memory}"(ptr nonnull %deferframe.buf) #5
  %setjmp.result = icmp eq i32 %setjmp, 0
@@ -172,7 +169,7 @@ rundefers.loophead:                               ; preds = %4, %3, %rundefers.b
  br i1 %stackIsNil, label %rundefers.end, label %rundefers.loop

 rundefers.loop:                                   ; preds = %rundefers.loophead
-  %stack.next.gep = getelementptr inbounds i8, ptr %2, i32 4
+  %stack.next.gep = getelementptr inbounds nuw i8, ptr %2, i32 4
  %stack.next = load ptr, ptr %stack.next.gep, align 4
  store ptr %stack.next, ptr %deferPtr, align 4
  %callback = load i32, ptr %2, align 4
@@ -218,7 +215,7 @@ rundefers.loophead10:                             ; preds = %7, %6, %lpad
  br i1 %stackIsNil11, label %rundefers.end7, label %rundefers.loop9

 rundefers.loop9:                                  ; preds = %rundefers.loophead10
-  %stack.next.gep12 = getelementptr inbounds i8, ptr %5, i32 4
+  %stack.next.gep12 = getelementptr inbounds nuw i8, ptr %5, i32 4
  %stack.next13 = load ptr, ptr %stack.next.gep12, align 4
  store ptr %stack.next13, ptr %deferPtr, align 4
  %callback15 = load i32, ptr %5, align 4
@@ -253,7 +250,7 @@ rundefers.end7:                                   ; preds = %rundefers.loophead1
 }

 ; Function Attrs: nounwind
-define internal void @"main.deferMultiple$1"(ptr %context) unnamed_addr #1 {
+define internal void @"main.deferMultiple$1"(ptr %context) unnamed_addr #0 {
 entry:
  call void @runtime.printlock(ptr undef) #4
  call void @runtime.printint32(i32 3, ptr undef) #4
@@ -262,7 +259,7 @@ entry:
 }

 ; Function Attrs: nounwind
-define internal void @"main.deferMultiple$2"(ptr %context) unnamed_addr #1 {
+define internal void @"main.deferMultiple$2"(ptr %context) unnamed_addr #0 {
 entry:
  call void @runtime.printlock(ptr undef) #4
  call void @runtime.printint32(i32 5, ptr undef) #4
@@ -270,9 +267,244 @@ entry:
  ret void
 }

-attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+armv7-m,+hwdiv,+soft-float,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
-attributes #1 = { nounwind "target-features"="+armv7-m,+hwdiv,+soft-float,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
-attributes #2 = { "target-features"="+armv7-m,+hwdiv,+soft-float,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
-attributes #3 = { nocallback nofree nosync nounwind willreturn }
+; Function Attrs: nounwind
+define hidden void @main.deferInfiniteLoop(ptr %context) unnamed_addr #0 {
+entry:
+  %deferframe.buf = alloca %runtime.deferFrame, align 4
+  %deferPtr = getelementptr inbounds nuw i8, ptr %deferframe.buf, i32 24
+  %0 = call ptr @llvm.stacksave.p0()
+  call void @runtime.setupDeferFrame(ptr nonnull %deferframe.buf, ptr %0, ptr undef) #4
+  br label %for.body
+
+for.body:                                         ; preds = %for.body, %entry
+  %defer.next = load ptr, ptr %deferPtr, align 4
+  %defer.alloc.call = call align 4 dereferenceable(12) ptr @runtime.alloc(i32 12, ptr nonnull inttoptr (i32 135 to ptr), ptr undef) #4
+  store i32 0, ptr %defer.alloc.call, align 4
+  %defer.alloc.call.repack1 = getelementptr inbounds nuw i8, ptr %defer.alloc.call, i32 4
+  store ptr %defer.next, ptr %defer.alloc.call.repack1, align 4
+  %defer.alloc.call.repack3 = getelementptr inbounds nuw i8, ptr %defer.alloc.call, i32 8
+  store i32 8, ptr %defer.alloc.call.repack3, align 4
+  store ptr %defer.alloc.call, ptr %deferPtr, align 4
+  br label %for.body
+
+recover:                                          ; preds = %rundefers.end
+  ret void
+
+lpad:                                             ; No predecessors!
+  br label %rundefers.loophead
+
+rundefers.loophead:                               ; preds = %rundefers.callback0, %lpad
+  br i1 poison, label %rundefers.end, label %rundefers.loop
+
+rundefers.loop:                                   ; preds = %rundefers.loophead
+  switch i32 poison, label %rundefers.default [
+    i32 0, label %rundefers.callback0
+  ]
+
+rundefers.callback0:                              ; preds = %rundefers.loop
+  br label %rundefers.loophead
+
+rundefers.default:                                ; preds = %rundefers.loop
+  unreachable
+
+rundefers.end:                                    ; preds = %rundefers.loophead
+  br label %recover
+}
+
+; Function Attrs: allockind("alloc,zeroed") allocsize(0)
+declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #3
+
+; Function Attrs: nounwind
+define hidden void @main.deferLoop(ptr %context) unnamed_addr #0 {
+entry:
+  %deferframe.buf = alloca %runtime.deferFrame, align 4
+  %deferPtr = getelementptr inbounds nuw i8, ptr %deferframe.buf, i32 24
+  %0 = call ptr @llvm.stacksave.p0()
+  call void @runtime.setupDeferFrame(ptr nonnull %deferframe.buf, ptr %0, ptr undef) #4
+  br label %for.loop
+
+for.loop:                                         ; preds = %for.body, %entry
+  %1 = phi i32 [ 0, %entry ], [ %3, %for.body ]
+  %2 = icmp slt i32 %1, 10
+  br i1 %2, label %for.body, label %for.done
+
+for.body:                                         ; preds = %for.loop
+  %defer.next = load ptr, ptr %deferPtr, align 4
+  %defer.alloc.call = call align 4 dereferenceable(12) ptr @runtime.alloc(i32 12, ptr nonnull inttoptr (i32 135 to ptr), ptr undef) #4
+  store i32 0, ptr %defer.alloc.call, align 4
+  %defer.alloc.call.repack13 = getelementptr inbounds nuw i8, ptr %defer.alloc.call, i32 4
+  store ptr %defer.next, ptr %defer.alloc.call.repack13, align 4
+  %defer.alloc.call.repack15 = getelementptr inbounds nuw i8, ptr %defer.alloc.call, i32 8
+  store i32 %1, ptr %defer.alloc.call.repack15, align 4
+  store ptr %defer.alloc.call, ptr %deferPtr, align 4
+  %3 = add i32 %1, 1
+  br label %for.loop
+
+for.done:                                         ; preds = %for.loop
+  br label %rundefers.block
+
+rundefers.after:                                  ; preds = %rundefers.end
+  call void @runtime.destroyDeferFrame(ptr nonnull %deferframe.buf, ptr undef) #4
+  ret void
+
+rundefers.block:                                  ; preds = %for.done
+  br label %rundefers.loophead
+
+rundefers.loophead:                               ; preds = %rundefers.callback0, %rundefers.block
+  %4 = load ptr, ptr %deferPtr, align 4
+  %stackIsNil = icmp eq ptr %4, null
+  br i1 %stackIsNil, label %rundefers.end, label %rundefers.loop
+
+rundefers.loop:                                   ; preds = %rundefers.loophead
+  %stack.next.gep = getelementptr inbounds nuw i8, ptr %4, i32 4
+  %stack.next = load ptr, ptr %stack.next.gep, align 4
+  store ptr %stack.next, ptr %deferPtr, align 4
+  %callback = load i32, ptr %4, align 4
+  switch i32 %callback, label %rundefers.default [
+    i32 0, label %rundefers.callback0
+  ]
+
+rundefers.callback0:                              ; preds = %rundefers.loop
+  %gep = getelementptr inbounds nuw i8, ptr %4, i32 8
+  %param = load i32, ptr %gep, align 4
+  call void @runtime.printlock(ptr undef) #4
+  call void @runtime.printint32(i32 %param, ptr undef) #4
+  call void @runtime.printunlock(ptr undef) #4
+  br label %rundefers.loophead
+
+rundefers.default:                                ; preds = %rundefers.loop
+  unreachable
+
+rundefers.end:                                    ; preds = %rundefers.loophead
+  br label %rundefers.after
+
+recover:                                          ; preds = %rundefers.end1
+  ret void
+
+lpad:                                             ; No predecessors!
+  br label %rundefers.loophead4
+
+rundefers.loophead4:                              ; preds = %rundefers.callback010, %lpad
+  br i1 poison, label %rundefers.end1, label %rundefers.loop3
+
+rundefers.loop3:                                  ; preds = %rundefers.loophead4
+  switch i32 poison, label %rundefers.default2 [
+    i32 0, label %rundefers.callback010
+  ]
+
+rundefers.callback010:                            ; preds = %rundefers.loop3
+  br label %rundefers.loophead4
+
+rundefers.default2:                               ; preds = %rundefers.loop3
+  unreachable
+
+rundefers.end1:                                   ; preds = %rundefers.loophead4
+  br label %recover
+}
+
+; Function Attrs: nounwind
+define hidden void @main.deferBetweenLoops(ptr %context) unnamed_addr #0 {
+entry:
+  %defer.alloca = alloca { i32, ptr, i32 }, align 4
+  %deferframe.buf = alloca %runtime.deferFrame, align 4
+  %deferPtr = getelementptr inbounds nuw i8, ptr %deferframe.buf, i32 24
+  %0 = call ptr @llvm.stacksave.p0()
+  call void @runtime.setupDeferFrame(ptr nonnull %deferframe.buf, ptr %0, ptr undef) #4
+  br label %for.loop
+
+for.loop:                                         ; preds = %for.body, %entry
+  %1 = phi i32 [ 0, %entry ], [ %3, %for.body ]
+  %2 = icmp slt i32 %1, 10
+  br i1 %2, label %for.body, label %for.done
+
+for.body:                                         ; preds = %for.loop
+  %3 = add i32 %1, 1
+  br label %for.loop
+
+for.done:                                         ; preds = %for.loop
+  %defer.next = load ptr, ptr %deferPtr, align 4
+  store i32 0, ptr %defer.alloca, align 4
+  %defer.alloca.repack16 = getelementptr inbounds nuw i8, ptr %defer.alloca, i32 4
+  store ptr %defer.next, ptr %defer.alloca.repack16, align 4
+  %defer.alloca.repack18 = getelementptr inbounds nuw i8, ptr %defer.alloca, i32 8
+  store i32 1, ptr %defer.alloca.repack18, align 4
+  store ptr %defer.alloca, ptr %deferPtr, align 4
+  br label %for.loop1
+
+for.loop1:                                        ; preds = %for.body2, %for.done
+  %4 = phi i32 [ 0, %for.done ], [ %6, %for.body2 ]
+  %5 = icmp slt i32 %4, 10
+  br i1 %5, label %for.body2, label %for.done3
+
+for.body2:                                        ; preds = %for.loop1
+  %6 = add i32 %4, 1
+  br label %for.loop1
+
+for.done3:                                        ; preds = %for.loop1
+  br label %rundefers.block
+
+rundefers.after:                                  ; preds = %rundefers.end
+  call void @runtime.destroyDeferFrame(ptr nonnull %deferframe.buf, ptr undef) #4
+  ret void
+
+rundefers.block:                                  ; preds = %for.done3
+  br label %rundefers.loophead
+
+rundefers.loophead:                               ; preds = %rundefers.callback0, %rundefers.block
+  %7 = load ptr, ptr %deferPtr, align 4
+  %stackIsNil = icmp eq ptr %7, null
+  br i1 %stackIsNil, label %rundefers.end, label %rundefers.loop
+
+rundefers.loop:                                   ; preds = %rundefers.loophead
+  %stack.next.gep = getelementptr inbounds nuw i8, ptr %7, i32 4
+  %stack.next = load ptr, ptr %stack.next.gep, align 4
+  store ptr %stack.next, ptr %deferPtr, align 4
+  %callback = load i32, ptr %7, align 4
+  switch i32 %callback, label %rundefers.default [
+    i32 0, label %rundefers.callback0
+  ]
+
+rundefers.callback0:                              ; preds = %rundefers.loop
+  %gep = getelementptr inbounds nuw i8, ptr %7, i32 8
+  %param = load i32, ptr %gep, align 4
+  call void @runtime.printlock(ptr undef) #4
+  call void @runtime.printint32(i32 %param, ptr undef) #4
+  call void @runtime.printunlock(ptr undef) #4
+  br label %rundefers.loophead
+
+rundefers.default:                                ; preds = %rundefers.loop
+  unreachable
+
+rundefers.end:                                    ; preds = %rundefers.loophead
+  br label %rundefers.after
+
+recover:                                          ; preds = %rundefers.end4
+  ret void
+
+lpad:                                             ; No predecessors!
+  br label %rundefers.loophead7
+
+rundefers.loophead7:                              ; preds = %rundefers.callback013, %lpad
+  br i1 poison, label %rundefers.end4, label %rundefers.loop6
+
+rundefers.loop6:                                  ; preds = %rundefers.loophead7
+  switch i32 poison, label %rundefers.default5 [
+    i32 0, label %rundefers.callback013
+  ]
+
+rundefers.callback013:                            ; preds = %rundefers.loop6
+  br label %rundefers.loophead7
+
+rundefers.default5:                               ; preds = %rundefers.loop6
+  unreachable
+
+rundefers.end4:                                   ; preds = %rundefers.loophead7
+  br label %recover
+}
+
+attributes #0 = { nounwind "target-features"="+armv7-m,+hwdiv,+soft-float,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
+attributes #1 = { "target-features"="+armv7-m,+hwdiv,+soft-float,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
+attributes #2 = { nocallback nofree nosync nounwind willreturn }
+attributes #3 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+armv7-m,+hwdiv,+soft-float,+thumb-mode,-aes,-bf16,-cdecp0,-cdecp1,-cdecp2,-cdecp3,-cdecp4,-cdecp5,-cdecp6,-cdecp7,-crc,-crypto,-d32,-dotprod,-dsp,-fp-armv8,-fp-armv8d16,-fp-armv8d16sp,-fp-armv8sp,-fp16,-fp16fml,-fp64,-fpregs,-fullfp16,-hwdiv-arm,-i8mm,-lob,-mve,-mve.fp,-neon,-pacbti,-ras,-sb,-sha2,-vfp2,-vfp2sp,-vfp3,-vfp3d16,-vfp3d16sp,-vfp3sp,-vfp4,-vfp4d16,-vfp4d16sp,-vfp4sp" }
 attributes #4 = { nounwind }
 attributes #5 = { nounwind returns_twice }
@@ -18,3 +18,23 @@ func deferMultiple() {
 	}()
 	external()
 }
+
+func deferInfiniteLoop() {
+	for {
+		defer print(8)
+	}
+}
+
+func deferLoop() {
+	for i := 0; i < 10; i++ {
+		defer print(i)
+	}
+}
+
+func deferBetweenLoops() {
+	for i := 0; i < 10; i++ {
+	}
+	defer print(1)
+	for i := 0; i < 10; i++ {
+	}
+}
@@ -1,21 +1,18 @@
 ; ModuleID = 'float.go'
 source_filename = "float.go"
-target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-i128:128-n32:64-S128-ni:1:10:20"
 target triple = "wasm32-unknown-wasi"

-; Function Attrs: allockind("alloc,zeroed") allocsize(0)
-declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #0
-
-declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #1
+declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.init(ptr %context) unnamed_addr #2 {
+define hidden void @main.init(ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

 ; Function Attrs: nounwind
-define hidden i32 @main.f32tou32(float %v, ptr %context) unnamed_addr #2 {
+define hidden i32 @main.f32tou32(float %v, ptr %context) unnamed_addr #1 {
 entry:
  %positive = fcmp oge float %v, 0.000000e+00
  %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
@@ -27,25 +24,25 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden float @main.maxu32f(ptr %context) unnamed_addr #2 {
+define hidden float @main.maxu32f(ptr %context) unnamed_addr #1 {
 entry:
  ret float 0x41F0000000000000
 }

 ; Function Attrs: nounwind
-define hidden i32 @main.maxu32tof32(ptr %context) unnamed_addr #2 {
+define hidden i32 @main.maxu32tof32(ptr %context) unnamed_addr #1 {
 entry:
  ret i32 -1
 }

 ; Function Attrs: nounwind
-define hidden { i32, i32, i32, i32 } @main.inftoi32(ptr %context) unnamed_addr #2 {
+define hidden { i32, i32, i32, i32 } @main.inftoi32(ptr %context) unnamed_addr #1 {
 entry:
  ret { i32, i32, i32, i32 } { i32 -1, i32 0, i32 2147483647, i32 -2147483648 }
 }

 ; Function Attrs: nounwind
-define hidden i32 @main.u32tof32tou32(i32 %v, ptr %context) unnamed_addr #2 {
+define hidden i32 @main.u32tof32tou32(i32 %v, ptr %context) unnamed_addr #1 {
 entry:
  %0 = uitofp i32 %v to float
  %withinmax = fcmp ole float %0, 0x41EFFFFFC0000000
@@ -55,7 +52,7 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden float @main.f32tou32tof32(float %v, ptr %context) unnamed_addr #2 {
+define hidden float @main.f32tou32tof32(float %v, ptr %context) unnamed_addr #1 {
 entry:
  %positive = fcmp oge float %v, 0.000000e+00
  %withinmax = fcmp ole float %v, 0x41EFFFFFC0000000
@@ -68,7 +65,7 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden i8 @main.f32tou8(float %v, ptr %context) unnamed_addr #2 {
+define hidden i8 @main.f32tou8(float %v, ptr %context) unnamed_addr #1 {
 entry:
  %positive = fcmp oge float %v, 0.000000e+00
  %withinmax = fcmp ole float %v, 2.550000e+02
@@ -80,7 +77,7 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden i8 @main.f32toi8(float %v, ptr %context) unnamed_addr #2 {
+define hidden i8 @main.f32toi8(float %v, ptr %context) unnamed_addr #1 {
 entry:
  %abovemin = fcmp oge float %v, -1.280000e+02
  %belowmax = fcmp ole float %v, 1.270000e+02
@@ -93,6 +90,5 @@ entry:
  ret i8 %0
 }

-attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #1 = { "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #2 = { nounwind "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #0 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #1 = { nounwind "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
@@ -1,50 +1,46 @@
 ; ModuleID = 'func.go'
 source_filename = "func.go"
-target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-i128:128-n32:64-S128-ni:1:10:20"
 target triple = "wasm32-unknown-wasi"

-; Function Attrs: allockind("alloc,zeroed") allocsize(0)
-declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #0
-
-declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #1
+declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.init(ptr %context) unnamed_addr #2 {
+define hidden void @main.init(ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

 ; Function Attrs: nounwind
-define hidden void @main.foo(ptr %callback.context, ptr %callback.funcptr, ptr %context) unnamed_addr #2 {
+define hidden void @main.foo(ptr %callback.context, ptr %callback.funcptr, ptr %context) unnamed_addr #1 {
 entry:
  %0 = icmp eq ptr %callback.funcptr, null
  br i1 %0, label %fpcall.throw, label %fpcall.next

 fpcall.next:                                      ; preds = %entry
-  call void %callback.funcptr(i32 3, ptr %callback.context) #3
+  call void %callback.funcptr(i32 3, ptr %callback.context) #2
  ret void

 fpcall.throw:                                     ; preds = %entry
-  call void @runtime.nilPanic(ptr undef) #3
+  call void @runtime.nilPanic(ptr undef) #2
  unreachable
 }

-declare void @runtime.nilPanic(ptr) #1
+declare void @runtime.nilPanic(ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.bar(ptr %context) unnamed_addr #2 {
+define hidden void @main.bar(ptr %context) unnamed_addr #1 {
 entry:
  call void @main.foo(ptr undef, ptr nonnull @main.someFunc, ptr undef)
  ret void
 }

 ; Function Attrs: nounwind
-define hidden void @main.someFunc(i32 %arg0, ptr %context) unnamed_addr #2 {
+define hidden void @main.someFunc(i32 %arg0, ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

-attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #1 = { "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #2 = { nounwind "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #3 = { nounwind }
+attributes #0 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #1 = { nounwind "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #2 = { nounwind }
@@ -24,6 +24,10 @@ var (
 		x *byte
 		y [61]uintptr
 	}
+	struct5 *struct {
+		x *byte
+		y [30]uintptr
+	}

 	slice1 []byte
 	slice2 []*int
@@ -58,6 +62,10 @@ func newStruct() {
 		x *byte
 		y [61]uintptr
 	})
+	struct5 = new(struct {
+		x *byte
+		y [30]uintptr
+	})
 }

 func newFuncValue() *func() {
@@ -1,6 +1,6 @@
 ; ModuleID = 'gc.go'
 source_filename = "gc.go"
-target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-i128:128-n32:64-S128-ni:1:10:20"
 target triple = "wasm32-unknown-wasi"

 %runtime._interface = type { ptr, ptr }
@@ -16,27 +16,25 @@ target triple = "wasm32-unknown-wasi"
@main.struct2 = hidden global ptr null, align 4
@main.struct3 = hidden global ptr null, align 4
@main.struct4 = hidden global ptr null, align 4
+@main.struct5 = hidden global ptr null, align 4
@main.slice1 = hidden global { ptr, i32, i32 } zeroinitializer, align 4
@main.slice2 = hidden global { ptr, i32, i32 } zeroinitializer, align 4
@main.slice3 = hidden global { ptr, i32, i32 } zeroinitializer, align 4
-@"runtime/gc.layout:62-2000000000000001" = linkonce_odr unnamed_addr constant { i32, [8 x i8] } { i32 62, [8 x i8] c"\01\00\00\00\00\00\00 " }
-@"runtime/gc.layout:62-0001" = linkonce_odr unnamed_addr constant { i32, [8 x i8] } { i32 62, [8 x i8] c"\01\00\00\00\00\00\00\00" }
+@"runtime/gc.layout:62-0100000000000020" = linkonce_odr unnamed_addr constant { i32, [8 x i8] } { i32 62, [8 x i8] c"\01\00\00\00\00\00\00 " }
+@"runtime/gc.layout:62-0100000000000000" = linkonce_odr unnamed_addr constant { i32, [8 x i8] } { i32 62, [8 x i8] c"\01\00\00\00\00\00\00\00" }
@"reflect/types.type:basic:complex128" = linkonce_odr constant { i8, ptr } { i8 80, ptr @"reflect/types.type:pointer:basic:complex128" }, align 4
@"reflect/types.type:pointer:basic:complex128" = linkonce_odr constant { i8, i16, ptr } { i8 -43, i16 0, ptr @"reflect/types.type:basic:complex128" }, align 4

-; Function Attrs: allockind("alloc,zeroed") allocsize(0)
-declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #0
-
-declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #1
+declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.init(ptr %context) unnamed_addr #2 {
+define hidden void @main.init(ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

 ; Function Attrs: nounwind
-define hidden void @main.newScalar(ptr %context) unnamed_addr #2 {
+define hidden void @main.newScalar(ptr %context) unnamed_addr #1 {
 entry:
  %stackalloc = alloca i8, align 1
  %new = call align 1 dereferenceable(1) ptr @runtime.alloc(i32 1, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #3
@@ -54,8 +52,11 @@ entry:
  ret void
 }

+; Function Attrs: allockind("alloc,zeroed") allocsize(0)
+declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #2
+
 ; Function Attrs: nounwind
-define hidden void @main.newArray(ptr %context) unnamed_addr #2 {
+define hidden void @main.newArray(ptr %context) unnamed_addr #1 {
 entry:
  %stackalloc = alloca i8, align 1
  %new = call align 1 dereferenceable(3) ptr @runtime.alloc(i32 3, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #3
@@ -71,26 +72,32 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden void @main.newStruct(ptr %context) unnamed_addr #2 {
+define hidden void @main.newStruct(ptr %context) unnamed_addr #1 {
 entry:
  %stackalloc = alloca i8, align 1
-  %new = call align 1 ptr @runtime.alloc(i32 0, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #3
+  %new = call align 1 ptr @runtime.alloc_zero(i32 0, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %new, ptr nonnull %stackalloc, ptr undef) #3
  store ptr %new, ptr @main.struct1, align 4
  %new1 = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %new1, ptr nonnull %stackalloc, ptr undef) #3
  store ptr %new1, ptr @main.struct2, align 4
-  %new2 = call align 4 dereferenceable(248) ptr @runtime.alloc(i32 248, ptr nonnull @"runtime/gc.layout:62-2000000000000001", ptr undef) #3
+  %new2 = call align 4 dereferenceable(248) ptr @runtime.alloc(i32 248, ptr nonnull @"runtime/gc.layout:62-0100000000000020", ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %new2, ptr nonnull %stackalloc, ptr undef) #3
  store ptr %new2, ptr @main.struct3, align 4
-  %new3 = call align 4 dereferenceable(248) ptr @runtime.alloc(i32 248, ptr nonnull @"runtime/gc.layout:62-0001", ptr undef) #3
+  %new3 = call align 4 dereferenceable(248) ptr @runtime.alloc(i32 248, ptr nonnull @"runtime/gc.layout:62-0100000000000000", ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %new3, ptr nonnull %stackalloc, ptr undef) #3
  store ptr %new3, ptr @main.struct4, align 4
+  %new4 = call align 4 dereferenceable(124) ptr @runtime.alloc(i32 124, ptr nonnull inttoptr (i32 127 to ptr), ptr undef) #3
+  call void @runtime.trackPointer(ptr nonnull %new4, ptr nonnull %stackalloc, ptr undef) #3
+  store ptr %new4, ptr @main.struct5, align 4
  ret void
 }

+; Function Attrs: allockind("alloc,zeroed") allocsize(0)
+declare noalias nonnull ptr @runtime.alloc_zero(i32, ptr, ptr) #2
+
 ; Function Attrs: nounwind
-define hidden ptr @main.newFuncValue(ptr %context) unnamed_addr #2 {
+define hidden ptr @main.newFuncValue(ptr %context) unnamed_addr #1 {
 entry:
  %stackalloc = alloca i8, align 1
  %new = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 197 to ptr), ptr undef) #3
@@ -99,35 +106,35 @@ entry:
 }

 ; Function Attrs: nounwind
-define hidden void @main.makeSlice(ptr %context) unnamed_addr #2 {
+define hidden void @main.makeSlice(ptr %context) unnamed_addr #1 {
 entry:
  %stackalloc = alloca i8, align 1
  %makeslice = call align 1 dereferenceable(5) ptr @runtime.alloc(i32 5, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %makeslice, ptr nonnull %stackalloc, ptr undef) #3
  store ptr %makeslice, ptr @main.slice1, align 4
-  store i32 5, ptr getelementptr inbounds (i8, ptr @main.slice1, i32 4), align 4
-  store i32 5, ptr getelementptr inbounds (i8, ptr @main.slice1, i32 8), align 4
+  store i32 5, ptr getelementptr inbounds nuw (i8, ptr @main.slice1, i32 4), align 4
+  store i32 5, ptr getelementptr inbounds nuw (i8, ptr @main.slice1, i32 8), align 4
  %makeslice1 = call align 4 dereferenceable(20) ptr @runtime.alloc(i32 20, ptr nonnull inttoptr (i32 67 to ptr), ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %makeslice1, ptr nonnull %stackalloc, ptr undef) #3
  store ptr %makeslice1, ptr @main.slice2, align 4
-  store i32 5, ptr getelementptr inbounds (i8, ptr @main.slice2, i32 4), align 4
-  store i32 5, ptr getelementptr inbounds (i8, ptr @main.slice2, i32 8), align 4
+  store i32 5, ptr getelementptr inbounds nuw (i8, ptr @main.slice2, i32 4), align 4
+  store i32 5, ptr getelementptr inbounds nuw (i8, ptr @main.slice2, i32 8), align 4
  %makeslice3 = call align 4 dereferenceable(60) ptr @runtime.alloc(i32 60, ptr nonnull inttoptr (i32 71 to ptr), ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %makeslice3, ptr nonnull %stackalloc, ptr undef) #3
  store ptr %makeslice3, ptr @main.slice3, align 4
-  store i32 5, ptr getelementptr inbounds (i8, ptr @main.slice3, i32 4), align 4
-  store i32 5, ptr getelementptr inbounds (i8, ptr @main.slice3, i32 8), align 4
+  store i32 5, ptr getelementptr inbounds nuw (i8, ptr @main.slice3, i32 4), align 4
+  store i32 5, ptr getelementptr inbounds nuw (i8, ptr @main.slice3, i32 8), align 4
  ret void
 }

 ; Function Attrs: nounwind
-define hidden %runtime._interface @main.makeInterface(double %v.r, double %v.i, ptr %context) unnamed_addr #2 {
+define hidden %runtime._interface @main.makeInterface(double %v.r, double %v.i, ptr %context) unnamed_addr #1 {
 entry:
  %stackalloc = alloca i8, align 1
-  %0 = call align 8 dereferenceable(16) ptr @runtime.alloc(i32 16, ptr null, ptr undef) #3
+  %0 = call align 8 dereferenceable(16) ptr @runtime.alloc(i32 16, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #3
  call void @runtime.trackPointer(ptr nonnull %0, ptr nonnull %stackalloc, ptr undef) #3
  store double %v.r, ptr %0, align 8
-  %.repack1 = getelementptr inbounds i8, ptr %0, i32 8
+  %.repack1 = getelementptr inbounds nuw i8, ptr %0, i32 8
  store double %v.i, ptr %.repack1, align 8
  %1 = insertvalue %runtime._interface { ptr @"reflect/types.type:basic:complex128", ptr undef }, ptr %0, 1
  call void @runtime.trackPointer(ptr nonnull @"reflect/types.type:basic:complex128", ptr nonnull %stackalloc, ptr undef) #3
@@ -135,7 +142,7 @@ entry:
  ret %runtime._interface %1
 }

-attributes #0 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #1 = { "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
-attributes #2 = { nounwind "target-features"="+bulk-memory,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #0 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #1 = { nounwind "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #2 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
 attributes #3 = { nounwind }
@@ -19,12 +19,49 @@ func Add[T Coord](a, b Point[T]) Point[T] {
 	}
 }

+func aliasSize[F float32 | float64]() uintptr {
+	return unsafe.Sizeof(F(0))
+}
+
+func aliasSize32() uintptr {
+	type F = float32
+	return aliasSize[F]()
+}
+
+func aliasSize64() uintptr {
+	type F = float64
+	return aliasSize[F]()
+}
+
+type aliasMethodResult[F float32 | float64] struct {
+	Value F
+}
+
+type aliasMethodValue[F float32 | float64] struct{}
+
+func (aliasMethodValue[F]) Get() aliasMethodResult[F] {
+	return aliasMethodResult[F]{}
+}
+
 func main() {
 	var af, bf Point[float32]
 	Add(af, bf)

 	var ai, bi Point[int]
 	Add(ai, bi)
+
+	checkSize(aliasSize32())
+	checkSize(aliasSize64())
 }

 func checkSize(uintptr)
+
+func checkBool(bool)
+
+func aliasMethod32(x any) {
+	type F = float32
+	_, ok := x.(interface {
+		Get() aliasMethodResult[F]
+	})
+	checkBool(ok)
+}
@@ -1,259 +1,196 @@
 ; ModuleID = 'generics.go'
 source_filename = "generics.go"
-target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-n32:64-S128-ni:1:10:20"
+target datalayout = "e-m:e-p:32:32-p10:8:8-p20:8:8-i64:64-i128:128-n32:64-S128-ni:1:10:20"
 target triple = "wasm32-unknown-wasi"

-%"main.Point[int]" = type { i32, i32 }
 %"main.Point[float32]" = type { float, float }
+%"main.Point[int]" = type { i32, i32 }

-declare noalias nonnull i8* @runtime.alloc(i32, i8*, i8*) #0
-
-declare void @runtime.trackPointer(i8* nocapture readonly, i8*) #0
+declare void @runtime.trackPointer(ptr nocapture readonly, ptr, ptr) #0

 ; Function Attrs: nounwind
-define hidden void @main.init(i8* %context) unnamed_addr #1 {
+define hidden void @main.init(ptr %context) unnamed_addr #1 {
 entry:
  ret void
 }

 ; Function Attrs: nounwind
-define hidden void @main.main(i8* %context) unnamed_addr #1 {
+define hidden i32 @main.aliasSize32(ptr %context) unnamed_addr #1 {
 entry:
-  %bi = alloca %"main.Point[int]", align 8
-  %ai = alloca %"main.Point[int]", align 8
-  %bf = alloca %"main.Point[float32]", align 8
-  %af = alloca %"main.Point[float32]", align 8
-  %af.repack = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %af, i32 0, i32 0
-  store float 0.000000e+00, float* %af.repack, align 8
-  %af.repack1 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %af, i32 0, i32 1
-  store float 0.000000e+00, float* %af.repack1, align 4
-  %0 = bitcast %"main.Point[float32]"* %af to i8*
-  call void @runtime.trackPointer(i8* nonnull %0, i8* undef) #2
-  %bf.repack = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %bf, i32 0, i32 0
-  store float 0.000000e+00, float* %bf.repack, align 8
-  %bf.repack2 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %bf, i32 0, i32 1
-  store float 0.000000e+00, float* %bf.repack2, align 4
-  %1 = bitcast %"main.Point[float32]"* %bf to i8*
-  call void @runtime.trackPointer(i8* nonnull %1, i8* undef) #2
-  %.elt = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %af, i32 0, i32 0
-  %.unpack = load float, float* %.elt, align 8
-  %.elt3 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %af, i32 0, i32 1
-  %.unpack4 = load float, float* %.elt3, align 4
-  %.elt5 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %bf, i32 0, i32 0
-  %.unpack6 = load float, float* %.elt5, align 8
-  %.elt7 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %bf, i32 0, i32 1
-  %.unpack8 = load float, float* %.elt7, align 4
-  %2 = call %"main.Point[float32]" @"main.Add[float32]"(float %.unpack, float %.unpack4, float %.unpack6, float %.unpack8, i8* undef)
-  %ai.repack = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %ai, i32 0, i32 0
-  store i32 0, i32* %ai.repack, align 8
-  %ai.repack9 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %ai, i32 0, i32 1
-  store i32 0, i32* %ai.repack9, align 4
-  %3 = bitcast %"main.Point[int]"* %ai to i8*
-  call void @runtime.trackPointer(i8* nonnull %3, i8* undef) #2
-  %bi.repack = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %bi, i32 0, i32 0
-  store i32 0, i32* %bi.repack, align 8
-  %bi.repack10 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %bi, i32 0, i32 1
-  store i32 0, i32* %bi.repack10, align 4
-  %4 = bitcast %"main.Point[int]"* %bi to i8*
-  call void @runtime.trackPointer(i8* nonnull %4, i8* undef) #2
-  %.elt11 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %ai, i32 0, i32 0
-  %.unpack12 = load i32, i32* %.elt11, align 8
-  %.elt13 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %ai, i32 0, i32 1
-  %.unpack14 = load i32, i32* %.elt13, align 4
-  %.elt15 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %bi, i32 0, i32 0
-  %.unpack16 = load i32, i32* %.elt15, align 8
-  %.elt17 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %bi, i32 0, i32 1
-  %.unpack18 = load i32, i32* %.elt17, align 4
-  %5 = call %"main.Point[int]" @"main.Add[int]"(i32 %.unpack12, i32 %.unpack14, i32 %.unpack16, i32 %.unpack18, i8* undef)
+  %0 = call i32 @"main.aliasSize[basic:float32]"(ptr undef)
+  ret i32 %0
+}
+
+; Function Attrs: nounwind
+define linkonce_odr hidden i32 @"main.aliasSize[basic:float32]"(ptr %context) unnamed_addr #1 {
+entry:
+  ret i32 4
+}
+
+; Function Attrs: nounwind
+define hidden i32 @main.aliasSize64(ptr %context) unnamed_addr #1 {
+entry:
+  %0 = call i32 @"main.aliasSize[basic:float64]"(ptr undef)
+  ret i32 %0
+}
+
+; Function Attrs: nounwind
+define linkonce_odr hidden i32 @"main.aliasSize[basic:float64]"(ptr %context) unnamed_addr #1 {
+entry:
+  ret i32 8
+}
+
+; Function Attrs: nounwind
+define hidden void @main.main(ptr %context) unnamed_addr #1 {
+entry:
+  %0 = call %"main.Point[float32]" @"main.Add[basic:float32]"(float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, float 0.000000e+00, ptr undef)
+  %1 = call %"main.Point[int]" @"main.Add[basic:int]"(i32 0, i32 0, i32 0, i32 0, ptr undef)
+  %2 = call i32 @main.aliasSize32(ptr undef)
+  call void @main.checkSize(i32 %2, ptr undef) #4
+  %3 = call i32 @main.aliasSize64(ptr undef)
+  call void @main.checkSize(i32 %3, ptr undef) #4
  ret void
 }

 ; Function Attrs: nounwind
-define linkonce_odr hidden %"main.Point[float32]" @"main.Add[float32]"(float %a.X, float %a.Y, float %b.X, float %b.Y, i8* %context) unnamed_addr #1 {
+define linkonce_odr hidden %"main.Point[float32]" @"main.Add[basic:float32]"(float %a.X, float %a.Y, float %b.X, float %b.Y, ptr %context) unnamed_addr #1 {
 entry:
-  %complit = alloca %"main.Point[float32]", align 8
-  %b = alloca %"main.Point[float32]", align 8
-  %a = alloca %"main.Point[float32]", align 8
-  %a.repack = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %a, i32 0, i32 0
-  store float 0.000000e+00, float* %a.repack, align 8
-  %a.repack9 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %a, i32 0, i32 1
-  store float 0.000000e+00, float* %a.repack9, align 4
-  %0 = bitcast %"main.Point[float32]"* %a to i8*
-  call void @runtime.trackPointer(i8* nonnull %0, i8* undef) #2
-  %a.repack10 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %a, i32 0, i32 0
-  store float %a.X, float* %a.repack10, align 8
-  %a.repack11 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %a, i32 0, i32 1
-  store float %a.Y, float* %a.repack11, align 4
-  %b.repack = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %b, i32 0, i32 0
-  store float 0.000000e+00, float* %b.repack, align 8
-  %b.repack13 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %b, i32 0, i32 1
-  store float 0.000000e+00, float* %b.repack13, align 4
-  %1 = bitcast %"main.Point[float32]"* %b to i8*
-  call void @runtime.trackPointer(i8* nonnull %1, i8* undef) #2
-  %b.repack14 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %b, i32 0, i32 0
-  store float %b.X, float* %b.repack14, align 8
-  %b.repack15 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %b, i32 0, i32 1
-  store float %b.Y, float* %b.repack15, align 4
-  call void @main.checkSize(i32 4, i8* undef) #2
-  call void @main.checkSize(i32 8, i8* undef) #2
-  %complit.repack = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %complit, i32 0, i32 0
-  store float 0.000000e+00, float* %complit.repack, align 8
-  %complit.repack17 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %complit, i32 0, i32 1
-  store float 0.000000e+00, float* %complit.repack17, align 4
-  %2 = bitcast %"main.Point[float32]"* %complit to i8*
-  call void @runtime.trackPointer(i8* nonnull %2, i8* undef) #2
-  %3 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %complit, i32 0, i32 0
+  %stackalloc = alloca i8, align 1
+  %a = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #4
+  call void @runtime.trackPointer(ptr nonnull %a, ptr nonnull %stackalloc, ptr undef) #4
+  store float %a.X, ptr %a, align 4
+  %a.repack5 = getelementptr inbounds nuw i8, ptr %a, i32 4
+  store float %a.Y, ptr %a.repack5, align 4
+  %b = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #4
+  call void @runtime.trackPointer(ptr nonnull %b, ptr nonnull %stackalloc, ptr undef) #4
+  store float %b.X, ptr %b, align 4
+  %b.repack7 = getelementptr inbounds nuw i8, ptr %b, i32 4
+  store float %b.Y, ptr %b.repack7, align 4
+  call void @main.checkSize(i32 4, ptr undef) #4
+  call void @main.checkSize(i32 8, ptr undef) #4
+  %complit = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #4
+  call void @runtime.trackPointer(ptr nonnull %complit, ptr nonnull %stackalloc, ptr undef) #4
  br i1 false, label %deref.throw, label %deref.next

 deref.next:                                       ; preds = %entry
-  br i1 false, label %deref.throw1, label %deref.next2
+  br i1 false, label %deref.throw, label %deref.next1

-deref.next2:                                      ; preds = %deref.next
-  %4 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %b, i32 0, i32 0
-  %5 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %a, i32 0, i32 0
-  %6 = load float, float* %5, align 8
-  %7 = load float, float* %4, align 8
-  %8 = fadd float %6, %7
-  br i1 false, label %deref.throw3, label %deref.next4
+deref.next1:                                      ; preds = %deref.next
+  %0 = load float, ptr %a, align 4
+  %1 = load float, ptr %b, align 4
+  %2 = fadd float %0, %1
+  br i1 false, label %deref.throw, label %deref.next2

-deref.next4:                                      ; preds = %deref.next2
-  br i1 false, label %deref.throw5, label %deref.next6
+deref.next2:                                      ; preds = %deref.next1
+  br i1 false, label %deref.throw, label %deref.next3

-deref.next6:                                      ; preds = %deref.next4
-  %9 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %b, i32 0, i32 1
-  %10 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %a, i32 0, i32 1
-  %11 = load float, float* %10, align 4
-  %12 = load float, float* %9, align 4
-  br i1 false, label %store.throw, label %store.next
+deref.next3:                                      ; preds = %deref.next2
+  %3 = getelementptr inbounds nuw i8, ptr %b, i32 4
+  %4 = getelementptr inbounds nuw i8, ptr %a, i32 4
+  %5 = load float, ptr %4, align 4
+  %6 = load float, ptr %3, align 4
+  br i1 false, label %deref.throw, label %store.next

-store.next:                                       ; preds = %deref.next6
-  store float %8, float* %3, align 8
-  br i1 false, label %store.throw7, label %store.next8
+store.next:                                       ; preds = %deref.next3
+  store float %2, ptr %complit, align 4
+  br i1 false, label %deref.throw, label %store.next4

-store.next8:                                      ; preds = %store.next
-  %13 = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %complit, i32 0, i32 1
-  %14 = fadd float %11, %12
-  store float %14, float* %13, align 4
-  %.elt = getelementptr inbounds %"main.Point[float32]", %"main.Point[float32]"* %complit, i32 0, i32 0
-  %.unpack = load float, float* %.elt, align 8
-  %15 = insertvalue %"main.Point[float32]" undef, float %.unpack, 0
-  %16 = insertvalue %"main.Point[float32]" %15, float %14, 1
-  ret %"main.Point[float32]" %16
+store.next4:                                      ; preds = %store.next
+  %7 = getelementptr inbounds nuw i8, ptr %complit, i32 4
+  %8 = fadd float %5, %6
+  store float %8, ptr %7, align 4
+  %.unpack = load float, ptr %complit, align 4
+  %9 = insertvalue %"main.Point[float32]" poison, float %.unpack, 0
+  %10 = insertvalue %"main.Point[float32]" %9, float %8, 1
+  ret %"main.Point[float32]" %10

-deref.throw:                                      ; preds = %entry
-  unreachable
-
-deref.throw1:                                     ; preds = %deref.next
-  unreachable
-
-deref.throw3:                                     ; preds = %deref.next2
-  unreachable
-
-deref.throw5:                                     ; preds = %deref.next4
-  unreachable
-
-store.throw:                                      ; preds = %deref.next6
-  unreachable
-
-store.throw7:                                     ; preds = %store.next
+deref.throw:                                      ; preds = %store.next, %deref.next3, %deref.next2, %deref.next1, %deref.next, %entry
  unreachable
 }

-declare void @main.checkSize(i32, i8*) #0
+; Function Attrs: allockind("alloc,zeroed") allocsize(0)
+declare noalias nonnull ptr @runtime.alloc(i32, ptr, ptr) #2

-declare void @runtime.nilPanic(i8*) #0
+declare void @main.checkSize(i32, ptr) #0
+
+declare void @runtime.nilPanic(ptr) #0

 ; Function Attrs: nounwind
-define linkonce_odr hidden %"main.Point[int]" @"main.Add[int]"(i32 %a.X, i32 %a.Y, i32 %b.X, i32 %b.Y, i8* %context) unnamed_addr #1 {
+define linkonce_odr hidden %"main.Point[int]" @"main.Add[basic:int]"(i32 %a.X, i32 %a.Y, i32 %b.X, i32 %b.Y, ptr %context) unnamed_addr #1 {
 entry:
-  %complit = alloca %"main.Point[int]", align 8
-  %b = alloca %"main.Point[int]", align 8
-  %a = alloca %"main.Point[int]", align 8
-  %a.repack = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %a, i32 0, i32 0
-  store i32 0, i32* %a.repack, align 8
-  %a.repack9 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %a, i32 0, i32 1
-  store i32 0, i32* %a.repack9, align 4
-  %0 = bitcast %"main.Point[int]"* %a to i8*
-  call void @runtime.trackPointer(i8* nonnull %0, i8* undef) #2
-  %a.repack10 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %a, i32 0, i32 0
-  store i32 %a.X, i32* %a.repack10, align 8
-  %a.repack11 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %a, i32 0, i32 1
-  store i32 %a.Y, i32* %a.repack11, align 4
-  %b.repack = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %b, i32 0, i32 0
-  store i32 0, i32* %b.repack, align 8
-  %b.repack13 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %b, i32 0, i32 1
-  store i32 0, i32* %b.repack13, align 4
-  %1 = bitcast %"main.Point[int]"* %b to i8*
-  call void @runtime.trackPointer(i8* nonnull %1, i8* undef) #2
-  %b.repack14 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %b, i32 0, i32 0
-  store i32 %b.X, i32* %b.repack14, align 8
-  %b.repack15 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %b, i32 0, i32 1
-  store i32 %b.Y, i32* %b.repack15, align 4
-  call void @main.checkSize(i32 4, i8* undef) #2
-  call void @main.checkSize(i32 8, i8* undef) #2
-  %complit.repack = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %complit, i32 0, i32 0
-  store i32 0, i32* %complit.repack, align 8
-  %complit.repack17 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %complit, i32 0, i32 1
-  store i32 0, i32* %complit.repack17, align 4
-  %2 = bitcast %"main.Point[int]"* %complit to i8*
-  call void @runtime.trackPointer(i8* nonnull %2, i8* undef) #2
-  %3 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %complit, i32 0, i32 0
+  %stackalloc = alloca i8, align 1
+  %a = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #4
+  call void @runtime.trackPointer(ptr nonnull %a, ptr nonnull %stackalloc, ptr undef) #4
+  store i32 %a.X, ptr %a, align 4
+  %a.repack5 = getelementptr inbounds nuw i8, ptr %a, i32 4
+  store i32 %a.Y, ptr %a.repack5, align 4
+  %b = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #4
+  call void @runtime.trackPointer(ptr nonnull %b, ptr nonnull %stackalloc, ptr undef) #4
+  store i32 %b.X, ptr %b, align 4
+  %b.repack7 = getelementptr inbounds nuw i8, ptr %b, i32 4
+  store i32 %b.Y, ptr %b.repack7, align 4
+  call void @main.checkSize(i32 4, ptr undef) #4
+  call void @main.checkSize(i32 8, ptr undef) #4
+  %complit = call align 4 dereferenceable(8) ptr @runtime.alloc(i32 8, ptr nonnull inttoptr (i32 3 to ptr), ptr undef) #4
+  call void @runtime.trackPointer(ptr nonnull %complit, ptr nonnull %stackalloc, ptr undef) #4
  br i1 false, label %deref.throw, label %deref.next

 deref.next:                                       ; preds = %entry
-  br i1 false, label %deref.throw1, label %deref.next2
+  br i1 false, label %deref.throw, label %deref.next1

-deref.next2:                                      ; preds = %deref.next
-  %4 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %b, i32 0, i32 0
-  %5 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %a, i32 0, i32 0
-  %6 = load i32, i32* %5, align 8
-  %7 = load i32, i32* %4, align 8
-  %8 = add i32 %6, %7
-  br i1 false, label %deref.throw3, label %deref.next4
+deref.next1:                                      ; preds = %deref.next
+  %0 = load i32, ptr %a, align 4
+  %1 = load i32, ptr %b, align 4
+  %2 = add i32 %0, %1
+  br i1 false, label %deref.throw, label %deref.next2

-deref.next4:                                      ; preds = %deref.next2
-  br i1 false, label %deref.throw5, label %deref.next6
+deref.next2:                                      ; preds = %deref.next1
+  br i1 false, label %deref.throw, label %deref.next3

-deref.next6:                                      ; preds = %deref.next4
-  %9 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %b, i32 0, i32 1
-  %10 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %a, i32 0, i32 1
-  %11 = load i32, i32* %10, align 4
-  %12 = load i32, i32* %9, align 4
-  br i1 false, label %store.throw, label %store.next
+deref.next3:                                      ; preds = %deref.next2
+  %3 = getelementptr inbounds nuw i8, ptr %b, i32 4
+  %4 = getelementptr inbounds nuw i8, ptr %a, i32 4
+  %5 = load i32, ptr %4, align 4
+  %6 = load i32, ptr %3, align 4
+  br i1 false, label %deref.throw, label %store.next

-store.next:                                       ; preds = %deref.next6
-  store i32 %8, i32* %3, align 8
-  br i1 false, label %store.throw7, label %store.next8
+store.next:                                       ; preds = %deref.next3
+  store i32 %2, ptr %complit, align 4
+  br i1 false, label %deref.throw, label %store.next4

-store.next8:                                      ; preds = %store.next
-  %13 = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %complit, i32 0, i32 1
-  %14 = add i32 %11, %12
-  store i32 %14, i32* %13, align 4
-  %.elt = getelementptr inbounds %"main.Point[int]", %"main.Point[int]"* %complit, i32 0, i32 0
-  %.unpack = load i32, i32* %.elt, align 8
-  %15 = insertvalue %"main.Point[int]" undef, i32 %.unpack, 0
-  %16 = insertvalue %"main.Point[int]" %15, i32 %14, 1
-  ret %"main.Point[int]" %16
+store.next4:                                      ; preds = %store.next
+  %7 = getelementptr inbounds nuw i8, ptr %complit, i32 4
+  %8 = add i32 %5, %6
+  store i32 %8, ptr %7, align 4
+  %.unpack = load i32, ptr %complit, align 4
+  %9 = insertvalue %"main.Point[int]" poison, i32 %.unpack, 0
+  %10 = insertvalue %"main.Point[int]" %9, i32 %8, 1
+  ret %"main.Point[int]" %10

-deref.throw:                                      ; preds = %entry
-  unreachable
-
-deref.throw1:                                     ; preds = %deref.next
-  unreachable
-
-deref.throw3:                                     ; preds = %deref.next2
-  unreachable
-
-deref.throw5:                                     ; preds = %deref.next4
-  unreachable
-
-store.throw:                                      ; preds = %deref.next6
-  unreachable
-
-store.throw7:                                     ; preds = %store.next
+deref.throw:                                      ; preds = %store.next, %deref.next3, %deref.next2, %deref.next1, %deref.next, %entry
  unreachable
 }

-attributes #0 = { "target-features"="+bulk-memory,+nontrapping-fptoint,+sign-ext" }
-attributes #1 = { nounwind "target-features"="+bulk-memory,+nontrapping-fptoint,+sign-ext" }
-attributes #2 = { nounwind }
+declare void @main.checkBool(i1, ptr) #0
+
+; Function Attrs: nounwind
+define hidden void @main.aliasMethod32(ptr %x.typecode, ptr %x.value, ptr %context) unnamed_addr #1 {
+entry:
+  %0 = call i1 @"interface:{Get:func:{}{named:main.aliasMethodResult[basic:float32]}}.$typeassert"(ptr %x.typecode) #4
+  br i1 %0, label %typeassert.ok, label %typeassert.next
+
+typeassert.next:                                  ; preds = %typeassert.ok, %entry
+  call void @main.checkBool(i1 %0, ptr undef) #4
+  ret void
+
+typeassert.ok:                                    ; preds = %entry
+  br label %typeassert.next
+}
+
+declare i1 @"interface:{Get:func:{}{named:main.aliasMethodResult[basic:float32]}}.$typeassert"(ptr) #3
+
+attributes #0 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #1 = { nounwind "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #2 = { allockind("alloc,zeroed") allocsize(0) "alloc-family"="runtime.alloc" "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" }
+attributes #3 = { "target-features"="+bulk-memory,+bulk-memory-opt,+call-indirect-overlong,+mutable-globals,+nontrapping-fptoint,+sign-ext,-multivalue,-reference-types" "tinygo-methods"="reflect/methods.Get:func:{}{named:main.aliasMethodResult[basic:float32]}" }
+attributes #4 = { nounwind }
--- a/Show More
+++ b/Show More