Add gcov-based test pruning with file-level coverage cache

.github/file-filter.yml

@@ -25,6 +25,7 @@ yml: &yml
   - '.github/workflows/phoenix/**'
   - '.github/workflows/frontier/**'
   - '.github/workflows/frontier_amd/**'
+  - '.github/workflows/common/**'
   - '.github/scripts/**'
   - '.github/workflows/bench.yml'
   - '.github/workflows/test.yml'
@@ -37,3 +38,6 @@ checkall: &checkall
   - *tests
   - *scripts
   - *yml
+
+cases_py:
+  - 'toolchain/mfc/test/cases.py'

.github/scripts/submit-slurm-job.sh

@@ -185,6 +185,7 @@ job_device="$device"
 job_interface="$interface"
 job_shard="$shard"
 job_cluster="$cluster"
+export GITHUB_EVENT_NAME="$GITHUB_EVENT_NAME"
 
 . ./mfc.sh load -c $compiler_flag -m $module_mode
 

.github/workflows/common/rebuild-cache.sh

@@ -0,0 +1,23 @@
+#!/bin/bash
+set -e
+
+# Number of parallel jobs: use SLURM allocation or default to 24.
+# Cap at 64 to avoid overwhelming OpenMPI daemons and OS process limits with concurrent launches.
+NJOBS="${SLURM_CPUS_ON_NODE:-24}"
+if [ "$NJOBS" -gt 64 ]; then NJOBS=64; fi
+
+# Clean stale build artifacts: the self-hosted runner may have a cached
+# GPU build (e.g. --gpu mp) whose CMake flags are incompatible with gcov.
+./mfc.sh clean
+
+# Source retry_build() for NFS stale file handle resilience (3 attempts).
+source .github/scripts/retry-build.sh
+
+# Build MFC with gcov coverage instrumentation (CPU-only, gfortran).
+retry_build ./mfc.sh build --gcov -j 8
+
+# Run all tests in parallel, collecting per-test coverage data.
+# Each test gets an isolated GCOV_PREFIX directory so .gcda files
+# don't collide. Coverage is collected per-test after all tests finish.
+# --gcov is required so the internal build step preserves instrumentation.
+./mfc.sh test --build-coverage-cache --gcov -j "$NJOBS"

.github/workflows/common/test.sh

@@ -82,4 +82,10 @@ if [ -n "${job_shard:-}" ]; then
     shard_opts="--shard $job_shard"
 fi
 
-./mfc.sh test -v --max-attempts 3 -a -j $n_test_threads $rdma_opts $device_opts $build_opts $shard_opts -- -c $job_cluster
+# Only prune tests on PRs; master pushes must run the full suite.
+prune_flag=""
+if [ "${GITHUB_EVENT_NAME:-}" = "pull_request" ]; then
+    prune_flag="--only-changes"
+fi
+
+./mfc.sh test -v --max-attempts 3 $prune_flag -a -j $n_test_threads $rdma_opts $device_opts $build_opts $shard_opts -- -c $job_cluster

.github/workflows/test.yml

@@ -56,22 +56,123 @@ jobs:
   file-changes:
     name: Detect File Changes
     runs-on: 'ubuntu-latest'
-    outputs: 
+    outputs:
       checkall: ${{ steps.changes.outputs.checkall }}
+      cases_py: ${{ steps.changes.outputs.cases_py }}
+      dep_changed: ${{ steps.dep-check.outputs.dep_changed }}
     steps:
       - name: Clone
         uses: actions/checkout@v4
 
       - name: Detect Changes
         uses: dorny/paths-filter@v3
         id: changes
-        with: 
+        with:
           filters: ".github/file-filter.yml"
 
+      - name: Check for Fortran dependency changes
+        id: dep-check
+        env:
+          GH_TOKEN: ${{ github.token }}
+        run: |
+          # Detect added/removed use/include statements that change the
+          # Fortran dependency graph, which would make the coverage cache stale.
+          PR_NUMBER="${{ github.event.pull_request.number }}"
+          BEFORE="${{ github.event.before }}"
+          AFTER="${{ github.event.after }}"
+          if [ "${{ github.event_name }}" = "pull_request" ]; then
+            # Default to dep_changed=true if gh pr diff fails (safe fallback).
+            DIFF=$(gh pr diff "$PR_NUMBER" 2>/dev/null) || {
+              echo "gh pr diff failed — defaulting to dep_changed=true for safety."
+              echo "dep_changed=true" >> "$GITHUB_OUTPUT"
+              exit 0
+            }
+          elif [ "${{ github.event_name }}" = "push" ]; then
+            DIFF=$(git diff "$BEFORE".."$AFTER" 2>/dev/null) || {
+              echo "git diff failed for push event — defaulting to dep_changed=true for safety."
+              echo "dep_changed=true" >> "$GITHUB_OUTPUT"
+              exit 0
+            }
+          else
+            DIFF=""
+          fi
+          if echo "$DIFF" | \
+             grep -qE '^[+-][[:space:]]*(use[[:space:],]+[a-zA-Z_]|#:include[[:space:]]|include[[:space:]]+['"'"'"])'; then
+            echo "dep_changed=true" >> "$GITHUB_OUTPUT"
+            echo "Fortran dependency change detected — will rebuild coverage cache."
+          else
+            echo "dep_changed=false" >> "$GITHUB_OUTPUT"
+          fi
+
+  rebuild-cache:
+    name: Rebuild Coverage Cache
+    needs: [lint-gate, file-changes]
+    if: >-
+      github.repository == 'MFlowCode/MFC' &&
+      (
+        (github.event_name == 'pull_request' &&
+         (needs.file-changes.outputs.cases_py == 'true' ||
+          needs.file-changes.outputs.dep_changed == 'true')) ||
+        (github.event_name == 'push' &&
+         (needs.file-changes.outputs.cases_py == 'true' ||
+          needs.file-changes.outputs.dep_changed == 'true')) ||
+        github.event_name == 'workflow_dispatch'
+      )
+    timeout-minutes: 240
+    runs-on:
+      group:  phoenix
+      labels: gt
+    permissions:
+      contents: write   # Required for Commit Cache to Master on push events
+    steps:
+      - name: Clone
+        uses: actions/checkout@v4
+        with:
+          ref: ${{ github.event_name == 'pull_request' && github.event.pull_request.head.sha || github.sha }}
+          clean: false
+
+      - name: Rebuild Cache via SLURM
+        run: bash .github/scripts/submit-slurm-job.sh .github/workflows/common/rebuild-cache.sh cpu none phoenix
+
+      - name: Print Logs
+        if:   always()
+        run:  cat rebuild-cache-cpu-none.out
+
+      - name: Upload Cache Artifact
+        if: github.event_name == 'pull_request'
+        uses: actions/upload-artifact@v4
+        with:
+          name: coverage-cache
+          path: toolchain/mfc/test/test_coverage_cache.json.gz
+          retention-days: 1
+
+      - name: Commit Cache to Master
+        if: (github.event_name == 'push' || github.event_name == 'workflow_dispatch') && github.ref == 'refs/heads/master'
+        run: |
+          git config user.name "github-actions[bot]"
+          git config user.email "github-actions[bot]@users.noreply.github.com"
+          git add toolchain/mfc/test/test_coverage_cache.json.gz
+          if git diff --cached --quiet; then
+            echo "Coverage cache unchanged."
+          else
+            git commit -m "Regenerate gcov coverage cache [skip ci]"
+            # Rebase onto latest master in case it advanced during the build
+            # (which can take 20-240 minutes). Only test_coverage_cache.json.gz
+            # is changed, so rebase conflicts are essentially impossible.
+            git fetch origin master
+            git rebase origin/master
+            git push origin HEAD:refs/heads/master
+          fi
+
   github:
     name: Github
-    if: needs.file-changes.outputs.checkall == 'true'
-    needs: [lint-gate, file-changes]
+    needs: [lint-gate, file-changes, rebuild-cache]
+    if: >-
+      !cancelled() &&
+      needs.lint-gate.result == 'success' &&
+      needs.file-changes.result == 'success' &&
+      (needs.rebuild-cache.result == 'success' || needs.rebuild-cache.result == 'skipped') &&
+      needs.file-changes.outputs.checkall == 'true'
     strategy:
       matrix:
         os:    ['ubuntu', 'macos']
@@ -98,6 +199,19 @@ jobs:
       - name: Clone
         uses: actions/checkout@v4
 
+      - name: Fetch master for coverage diff
+        run: |
+          git fetch origin master:master --depth=1
+          git fetch --deepen=200
+        continue-on-error: true
+
+      - name: Download Coverage Cache
+        uses: actions/download-artifact@v4
+        with:
+          name: coverage-cache
+          path: toolchain/mfc/test
+        continue-on-error: true
+
       - name: Setup MacOS
         if:   matrix.os == 'macos'
         run:  |
@@ -140,15 +254,23 @@ jobs:
 
       - name: Test
         run:  |
-          /bin/bash mfc.sh test -v --max-attempts 3 -j $(nproc) $TEST_ALL $TEST_PCT
+          /bin/bash mfc.sh test -v --max-attempts 3 -j $(nproc) $ONLY_CHANGES $TEST_ALL $TEST_PCT
         env:
           TEST_ALL: ${{ matrix.mpi == 'mpi' && '--test-all' || '' }}
           TEST_PCT: ${{ matrix.debug == 'debug' && '-% 20' || '' }}
+          ONLY_CHANGES: ${{ github.event_name == 'pull_request' && '--only-changes' || '' }}
 
   self:
     name: "${{ matrix.cluster_name }} (${{ matrix.device }}${{ matrix.interface != 'none' && format('-{0}', matrix.interface) || '' }}${{ matrix.shard != '' && format(' [{0}]', matrix.shard) || '' }})"
-    if: github.repository == 'MFlowCode/MFC' && needs.file-changes.outputs.checkall == 'true' && github.event.pull_request.draft != true
-    needs: [lint-gate, file-changes]
+    needs: [lint-gate, file-changes, rebuild-cache]
+    if: >-
+      !cancelled() &&
+      needs.lint-gate.result == 'success' &&
+      needs.file-changes.result == 'success' &&
+      (needs.rebuild-cache.result == 'success' || needs.rebuild-cache.result == 'skipped') &&
+      github.repository == 'MFlowCode/MFC' &&
+      needs.file-changes.outputs.checkall == 'true' &&
+      github.event.pull_request.draft != true
     # Frontier CCE compiler is periodically broken by toolchain updates (e.g.
     # cpe/25.03 introduced an IPA SIGSEGV in CCE 19.0.0). Allow Frontier to
     # fail without blocking PR merges; Phoenix remains a hard gate.
@@ -237,6 +359,13 @@ jobs:
       - name: Clean stale output files
         run:  rm -f *.out
 
+      - name: Download Coverage Cache
+        uses: actions/download-artifact@v4
+        with:
+          name: coverage-cache
+          path: toolchain/mfc/test
+        continue-on-error: true
+
       - name: Build (login node)
         if:   matrix.cluster != 'phoenix'
         timeout-minutes: 60

.gitignore

@@ -22,6 +22,9 @@ __pycache__
 # Auto-generated version file
 toolchain/mfc/_version.py
 
+# Raw coverage cache — legacy, not tracked (the .json.gz version IS committed)
+toolchain/mfc/test/test_coverage_cache.json
+
 # Auto-generated toolchain files (regenerate with: ./mfc.sh generate)
 toolchain/completions/mfc.bash
 toolchain/completions/_mfc

CMakeLists.txt

@@ -113,6 +113,8 @@ endif()
 # debug builds. These include optimization and debug flags, as well as some that
 # are required for a successful build of MFC.
 
+set(FYPP_GCOV_OPTS "")
+
 if (CMAKE_Fortran_COMPILER_ID STREQUAL "GNU")
     add_compile_options(
         $<$<COMPILE_LANGUAGE:Fortran>:-ffree-line-length-none>
@@ -131,13 +133,20 @@ if (CMAKE_Fortran_COMPILER_ID STREQUAL "GNU")
         add_compile_options(
             $<$<COMPILE_LANGUAGE:Fortran>:-fprofile-arcs>
             $<$<COMPILE_LANGUAGE:Fortran>:-ftest-coverage>
-            $<$<COMPILE_LANGUAGE:Fortran>:-O1>
-	    )
+        )
 
         add_link_options(
             $<$<COMPILE_LANGUAGE:Fortran>:-lgcov>
             $<$<COMPILE_LANGUAGE:Fortran>:--coverage>
         )
+
+        # Override Release -O3 with -O1 for gcov: coverage instrumentation is
+        # inaccurate at -O3, and aggressive codegen (e.g. AVX-512 FP16 on
+        # Granite Rapids) can emit instructions that older assemblers reject.
+        set(CMAKE_Fortran_FLAGS_RELEASE "-O1 -DNDEBUG" CACHE STRING "" FORCE)
+
+        # Use gfortran5 line markers so gcov can map coverage to .fpp sources.
+        set(FYPP_GCOV_OPTS "--line-marker-format=gfortran5")
     endif()
 
     if (CMAKE_BUILD_TYPE STREQUAL "Debug")
@@ -246,8 +255,11 @@ if (CMAKE_BUILD_TYPE STREQUAL "Release")
         endif()
     endif()
 
-    # Enable LTO/IPO if supported
-    if (CMAKE_Fortran_COMPILER_ID STREQUAL "NVHPC")
+    # Enable LTO/IPO if supported (skip for gcov — LTO interferes with coverage
+    # instrumentation and can trigger assembler errors on newer architectures).
+    if (MFC_GCov)
+        message(STATUS "LTO/IPO disabled for gcov build")
+    elseif (CMAKE_Fortran_COMPILER_ID STREQUAL "NVHPC")
         if (MFC_Unified)
             message(STATUS "LTO/IPO is not available with NVHPC using Unified Memory")
         elseif (CMAKE_Fortran_COMPILER_VERSION VERSION_GREATER "24.11" AND CMAKE_Fortran_COMPILER_VERSION VERSION_LESS "25.9")
@@ -393,6 +405,7 @@ macro(HANDLE_SOURCES target useCommon)
                                  --no-folding
 								 --line-length=999
 		 						 --line-numbering-mode=nocontlines
+                                 ${FYPP_GCOV_OPTS}
                                  "${fpp}" "${f90}"
             DEPENDS  "${fpp};${${target}_incs}"
             COMMENT  "Preprocessing (Fypp) ${fpp_filename}"

toolchain/mfc/cli/commands.py

@@ -438,6 +438,27 @@
             type=str,
             default=None,
         ),
+        Argument(
+            name="build-coverage-cache",
+            help="Run all tests with gcov instrumentation to build the file-level coverage cache. Pass --gcov to enable coverage instrumentation in the internal build step.",
+            action=ArgAction.STORE_TRUE,
+            default=False,
+            dest="build_coverage_cache",
+        ),
+        Argument(
+            name="only-changes",
+            help="Only run tests whose covered files overlap with files changed since branching from master (uses file-level gcov coverage cache).",
+            action=ArgAction.STORE_TRUE,
+            default=False,
+            dest="only_changes",
+        ),
+        Argument(
+            name="changes-branch",
+            help="Branch to compare against for --only-changes (default: master).",
+            type=str,
+            default="master",
+            dest="changes_branch",
+        ),
     ],
     mutually_exclusive=[
         MutuallyExclusiveGroup(
@@ -470,13 +491,17 @@
         Example("./mfc.sh test -j 4", "Run with 4 parallel jobs"),
         Example("./mfc.sh test --only 3D", "Run only 3D tests"),
         Example("./mfc.sh test --generate", "Regenerate golden files"),
+        Example("./mfc.sh test --only-changes -j 4", "Run tests affected by changed files"),
+        Example("./mfc.sh build --gcov -j 8 && ./mfc.sh test --build-coverage-cache", "One-time: build file-coverage cache"),
     ],
     key_options=[
         ("-j, --jobs N", "Number of parallel test jobs"),
         ("-o, --only PROP", "Run tests matching property"),
         ("-f, --from UUID", "Start from specific test"),
         ("--generate", "Generate/update golden files"),
         ("--no-build", "Skip rebuilding MFC"),
+        ("--build-coverage-cache", "Build file-level gcov coverage cache (one-time)"),
+        ("--only-changes", "Run tests affected by changed files (requires cache)"),
     ],
 )