GHA: clean up & extend testing; update to latest LLVM

.github/actions/run-lldb-tests/action.yml

@@ -15,7 +15,10 @@ inputs:
     description: Path to the checked-out ds2 repository
     required: true
   arch:
-    description: Architecture to pass to dotest.py's --arch (e.g. x86_64, i686, aarch64)
+    description: >-
+      Architecture (e.g. x86_64, i686, aarch64), passed to dotest.py as
+      --triple <arch>-unknown-linux-gnu (dotest.py dropped --arch in favor
+      of --triple as of release/23.x)
     required: true
   platform:
     description: >-
@@ -83,7 +86,7 @@ runs:
         # lldb-dotest itself, not to grep.
         "${{ inputs.python }}" "${{ inputs.lldb-dotest-path }}" \
           --out-of-tree-debugserver \
-          --arch "${{ inputs.arch }}" \
+          --triple "${{ inputs.arch }}-unknown-linux-gnu" \
           --platform-name "remote-${{ inputs.platform }}" \
           --platform-url "connect://localhost:${{ inputs.listen-port }}" \
           --platform-working-dir "${{ inputs.working-dir }}" \

.github/workflows/build.yml

@@ -15,7 +15,7 @@ concurrency:
   cancel-in-progress: true
 
 env:
-  LLVM_REF: ec450b19004a653f3db3ad50e88fbf6529a9d841
+  LLVM_REF: bb9934d4dfef9033ded36d26827dbb6d3ac1dd92 # release/23.x
   SCCACHE_DIRECT: 'true'
 
 jobs:
@@ -481,3 +481,80 @@ jobs:
     uses: ./.github/workflows/test-android.yml
     with:
       llvm-ref: ${{ needs.vars.outputs.llvm-ref }}
+
+  test-windows:
+    needs: [vars, windows, windows_arm64, build-lldb-windows]
+    uses: ./.github/workflows/test-windows.yml
+    with:
+      llvm-ref: ${{ needs.vars.outputs.llvm-ref }}
+
+  build-lldb-windows:
+    name: Build LLDB (Windows ${{ matrix.runner-arch }})
+    runs-on: ${{ matrix.runner }}
+
+    strategy:
+      fail-fast: false
+      matrix:
+        include:
+          - runner: windows-latest
+            runner-arch: X64
+          - runner: windows-11-arm
+            runner-arch: ARM64
+
+    steps:
+      # windows-latest carries several hostedtoolcache Python versions side
+      # by side; CMake's find_package(Python3) picks the newest one it can
+      # see, which need not be the one the plain `python` on PATH resolves
+      # to (the interpreter test-windows later uses to run lldb-dotest).
+      # Capture that one now, before vsdevenv's PATH rewrite, and pin the
+      # lldb build to it so the compiled _lldb extension's Python ABI tag
+      # matches what actually imports it at test time.
+      - name: Locate default Python
+        id: python
+        run: echo "path=$((Get-Command python).Source)" >> $env:GITHUB_OUTPUT
+
+      - uses: compnerd/gha-setup-vsdevenv@v7
+
+      - name: Install Build Tools
+        run: choco install swig make
+
+      - name: Setup sccache
+        uses: hendrikmuhs/ccache-action@d62db5f07c26379fc4b4e0916f098a92573c3b03 # v1.2.23
+        with:
+          key: llvm-${{ runner.os }}-${{ matrix.runner-arch }}
+          variant: sccache
+
+      - name: Checkout llvm project
+        uses: actions/checkout@v7
+        with:
+          repository: llvm/llvm-project
+          ref: ${{ env.LLVM_REF }}
+          path: ${{ github.workspace }}/SourceCache/llvm
+
+      # Build clang, lld (required to link test inferiors on Windows) and a
+      # Python-enabled lldb binary to execute the lldb tests.
+      - name: Configure lldb build
+        run: |
+          cmake -B ${{ github.workspace }}/BinaryCache/llvm                     `
+                -S ${{ github.workspace }}/SourceCache/llvm/llvm                `
+                -D LLVM_ENABLE_PROJECTS='clang;lld;lldb'                        `
+                -D LLVM_TARGETS_TO_BUILD="host"                                 `
+                -D LLDB_ENABLE_PYTHON=On                                        `
+                -D Python3_EXECUTABLE="${{ steps.python.outputs.path }}"        `
+                -D CMAKE_C_COMPILER_LAUNCHER=sccache                            `
+                -D CMAKE_CXX_COMPILER_LAUNCHER=sccache                          `
+                -D CMAKE_BUILD_TYPE=Release                                     `
+                -G Ninja
+
+      - name: Build llvm
+        run: |
+          cmake --build ${{ github.workspace }}/BinaryCache/llvm                `
+                --target tools/lldb/test/API/lldb-api-test-deps                 `
+                --target lld                                                    `
+                --config Release
+
+      - uses: actions/upload-artifact@v7
+        with:
+          name: llvm-Windows-${{ matrix.runner-arch }}
+          path: ${{ github.workspace }}/BinaryCache/llvm
+

.github/workflows/test-android.yml

@@ -14,15 +14,28 @@ jobs:
   # Android
   # -----------------------------------------------------------------------
   #
-  # The emulator-runner action takes the whole test invocation as a single
-  # shell script string executed inside its own wrapper, so it can't invoke
-  # the run-lldb-tests composite action as a normal step the way the Linux
-  # jobs do -- the dotest invocation stays inlined here, same as before this
-  # file existed.
+  # The emulator-runner action runs the whole test invocation as one shell
+  # script, so it can't use the run-lldb-tests composite action like the
+  # Linux jobs do; the dotest invocation stays inlined here.
+  #
+  # Each job's "Wrap NDK compiler" step works around Android.rules always
+  # appending its own unversioned --target after the compiler wrapper's
+  # default, which wins as the last -target flag and drops the API level
+  # below what NDK r29's libc++ needs for pthread_cond_clockwait. CFLAGS_EXTRAS
+  # can't fix this reliably: many test Makefiles set their own, silently
+  # discarding ours. Appending the real target as the wrapper's own last
+  # argument can't be overridden that way.
 
   test-android-x86_64:
     runs-on: ubuntu-latest
     name: Test Android x86_64
+    # Observed category durations run 5-14 minutes; this is roughly 2x the
+    # slowest, catching a hang well before GitHub's 360-minute job default.
+    timeout-minutes: 30
+
+    env:
+      # API level test inferiors compile against.
+      TEST_INFERIOR_API_LEVEL: 28
 
     strategy:
       fail-fast: false
@@ -87,6 +100,24 @@ jobs:
           disable-animations: false
           script: echo "Generated AVD snapshot for caching."
 
+      # See the top-of-file comment for why this is needed. Android.rules
+      # derives NDK_ROOT (and from it, the sysroot) purely from --compiler's
+      # own directory nesting (dirname($CC)/../../../../..), so the wrapper
+      # has to live 5 levels deep exactly like the real NDK, with a symlink
+      # standing in for the parts Android.rules reads directly from disk.
+      - name: Wrap NDK compiler
+        run: |
+          wrap_bin="$RUNNER_TEMP/android-clang-wrapper/toolchains/llvm/prebuilt/linux-x86_64/bin"
+          mkdir -p "$wrap_bin"
+          ln -s "$ANDROID_NDK_LATEST_HOME/toolchains/llvm/prebuilt/linux-x86_64/sysroot" "$wrap_bin/../sysroot"
+          wrapper="$wrap_bin/x86_64-linux-android$TEST_INFERIOR_API_LEVEL-clang"
+          {
+            echo '#!/usr/bin/env bash'
+            echo 'exec "$ANDROID_NDK_LATEST_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin/$(basename "$0")" "$@" -target x86_64-linux-android'"$TEST_INFERIOR_API_LEVEL"
+          } > "$wrapper"
+          cp "$wrapper" "$wrapper++"
+          chmod +x "$wrapper" "$wrapper++"
+
       - name: Run lldb tests against ds2 on Android emulator
         uses: reactivecircus/android-emulator-runner@v2
         with:
@@ -112,11 +143,11 @@ jobs:
 
             ${{ github.workspace }}/BinaryCache/llvm/bin/lldb-dotest
             --out-of-tree-debugserver
-            --arch x86_64
+            --triple x86_64-none-linux-android
             --platform-name remote-android
             --platform-url connect://localhost:5432
             --platform-working-dir /data/local/tmp
-            --compiler=$ANDROID_NDK_LATEST_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin/x86_64-linux-android28-clang
+            --compiler=$RUNNER_TEMP/android-clang-wrapper/toolchains/llvm/prebuilt/linux-x86_64/bin/x86_64-linux-android$TEST_INFERIOR_API_LEVEL-clang
             --excluded ${{ github.workspace }}/SourceCache/ds2/Support/Testing/Excluded/ds2/android-x86_64.excluded
             --excluded ${{ github.workspace }}/SourceCache/ds2/Support/Testing/Excluded/upstream/android-x86_64.excluded
             --excluded ${{ github.workspace }}/SourceCache/ds2/Support/Testing/Excluded/upstream/non-debugserver-tests.excluded
@@ -132,6 +163,13 @@ jobs:
   test-android-arm64-v8a:
     runs-on: ubuntu-latest
     name: Test Android arm64-v8a sanity
+    # Observed duration is ~7 minutes; this is roughly 2x that, catching a
+    # hang well before GitHub's 360-minute job default.
+    timeout-minutes: 15
+
+    env:
+      # API level test inferiors compile against.
+      TEST_INFERIOR_API_LEVEL: 27
 
     steps:
       - uses: actions/download-artifact@v8
@@ -186,6 +224,24 @@ jobs:
           disable-animations: false
           script: echo "Generated AVD snapshot for caching."
 
+      # See the top-of-file comment for why this is needed. Android.rules
+      # derives NDK_ROOT (and from it, the sysroot) purely from --compiler's
+      # own directory nesting (dirname($CC)/../../../../..), so the wrapper
+      # has to live 5 levels deep exactly like the real NDK, with a symlink
+      # standing in for the parts Android.rules reads directly from disk.
+      - name: Wrap NDK compiler
+        run: |
+          wrap_bin="$RUNNER_TEMP/android-clang-wrapper/toolchains/llvm/prebuilt/linux-x86_64/bin"
+          mkdir -p "$wrap_bin"
+          ln -s "$ANDROID_NDK_LATEST_HOME/toolchains/llvm/prebuilt/linux-x86_64/sysroot" "$wrap_bin/../sysroot"
+          wrapper="$wrap_bin/aarch64-linux-android$TEST_INFERIOR_API_LEVEL-clang"
+          {
+            echo '#!/usr/bin/env bash'
+            echo 'exec "$ANDROID_NDK_LATEST_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin/$(basename "$0")" "$@" -target aarch64-linux-android'"$TEST_INFERIOR_API_LEVEL"
+          } > "$wrapper"
+          cp "$wrapper" "$wrapper++"
+          chmod +x "$wrapper" "$wrapper++"
+
       - name: Run lldb tests against ds2 on Android emulator
         uses: reactivecircus/android-emulator-runner@v2
         with:
@@ -206,11 +262,11 @@ jobs:
 
             ${{ github.workspace }}/BinaryCache/llvm/bin/lldb-dotest
             --out-of-tree-debugserver
-            --arch aarch64
+            --triple aarch64-none-linux-android
             --platform-name remote-android
             --platform-url connect://localhost:5432
             --platform-working-dir /data/local/tmp
-            --compiler=$ANDROID_NDK_LATEST_HOME/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android27-clang
+            --compiler=$RUNNER_TEMP/android-clang-wrapper/toolchains/llvm/prebuilt/linux-x86_64/bin/aarch64-linux-android$TEST_INFERIOR_API_LEVEL-clang
             --excluded ${{ github.workspace }}/SourceCache/ds2/Support/Testing/Excluded/ds2/android-aarch64.excluded
             --excluded ${{ github.workspace }}/SourceCache/ds2/Support/Testing/Excluded/upstream/android-aarch64.excluded
             --excluded ${{ github.workspace }}/SourceCache/ds2/Support/Testing/Excluded/upstream/non-debugserver-tests.excluded

.github/workflows/test-linux.yml

@@ -29,6 +29,9 @@ jobs:
   test-linux-i686:
     runs-on: ubuntu-latest
     name: Test Linux i686
+    # Observed category durations run 3-8 minutes; this is roughly 2x the
+    # slowest, catching a hang well before GitHub's 360-minute job default.
+    timeout-minutes: 20
 
     # test-category is the matrix's only key -- GitHub renders every matrix
     # key/value pair in the check name (e.g. "Test Linux i686 (Python API)"),
@@ -89,6 +92,9 @@ jobs:
   test-linux-x86_64:
     runs-on: ubuntu-latest
     name: Test Linux x86_64
+    # Observed category durations run 3-8 minutes; this is roughly 2x the
+    # slowest, catching a hang well before GitHub's 360-minute job default.
+    timeout-minutes: 20
 
     strategy:
       fail-fast: false
@@ -140,6 +146,9 @@ jobs:
   test-linux-aarch64:
     runs-on: ubuntu-24.04-arm
     name: Test Linux aarch64
+    # Observed category durations run 2-5.5 minutes; this is roughly 2x the
+    # slowest, catching a hang well before GitHub's 360-minute job default.
+    timeout-minutes: 15
 
     strategy:
       fail-fast: false