Reach and require 100% function coverage

.github/workflows/website.yml

@@ -29,7 +29,7 @@ jobs:
       CACHE_KEY: ccache-v2-website-coverage
     steps:
       - name: Install dependencies
-        run: dnf install --assumeyes doxygen cmake gcc-c++ graphviz git clang llvm libstdc++-static diffutils util-linux
+        run: dnf install --assumeyes doxygen cmake gcc-c++ graphviz git clang llvm libstdc++-static diffutils util-linux python3
       - uses: actions/checkout@v7
       # The primary key never matches, as it carries the run identifier, so a
       # restore always walks the fallbacks in order. A previous run of this same
@@ -49,6 +49,10 @@ jobs:
       # Container jobs run as root, and a superuser is not constrained by the
       # permission bits that part of the suite asserts on
       - name: Build the website
+        # Every function that the published report measures has to be called by
+        # the test suite, so that a change cannot quietly lower it
+        env:
+          REQUIRE_FULL_FUNCTION_COVERAGE: "1"
         run: |
           useradd --create-home builder
           chown -R builder .

contrib/website.sh

@@ -69,10 +69,9 @@ export CMAKE_BUILD_PARALLEL_LEVEL
 export CTEST_PARALLEL_LEVEL
 
 # Counters are kept in the profile of each program as they change, rather than
-# written out once it exits, so that a program that dies on a fatal signal, like
-# the ones exercising the crash handler, still reports what it ran. The profile
-# file name asks for it, which is all that Apple platforms need, while elsewhere
-# the compiler has to arrange for it too
+# written out once it exits, so that a program that dies on a fatal signal still
+# reports what it ran. The profile file name asks for it, which is all that
+# Apple platforms need, while elsewhere the compiler has to arrange for it too
 PROFILE_FLAGS="-fprofile-instr-generate -fcoverage-mapping"
 if [ "$(uname)" != "Darwin" ]
 then
@@ -245,6 +244,62 @@ AWK
 awk -v "merged=$WORK_DIRECTORY/coverage.lcov" -f "$MERGE_PROGRAM" \
   "$LCOV_DIRECTORY"/*.lcov > "$WORK_DIRECTORY/summary.txt"
 
+# Functions are merged apart from the traces, as a trace only records the line
+# that a function starts on, which cannot tell apart two functions starting on
+# the same line. The JSON export records the column as well, which is how the
+# report itself counts every instantiation of a template as a single function,
+# and the highest count across the binaries is kept for the same reason as above
+FUNCTIONS_PROGRAM="$WORK_DIRECTORY/functions.py"
+cat > "$FUNCTIONS_PROGRAM" <<'PYTHON'
+import json
+import re
+import subprocess
+import sys
+
+llvm_cov, profile_data, exclude, object_list, uncovered = sys.argv[1:]
+excluded = re.compile(exclude)
+
+counts = {}
+with open(object_list, encoding="utf-8") as objects:
+    for binary in objects.read().splitlines():
+        export = subprocess.run(
+            [llvm_cov, "export", binary, f"-instr-profile={profile_data}",
+             "-format=text", "-skip-expansions",
+             f"-ignore-filename-regex={exclude}"],
+            check=True, stdout=subprocess.PIPE)
+        for data in json.loads(export.stdout)["data"]:
+            for function in data["functions"]:
+                filename = function["filenames"][0]
+                if excluded.search(filename):
+                    continue
+                start = function["regions"][0]
+                key = (filename, start[0], start[1])
+                counts[key] = max(counts.get(key, 0), function["count"])
+
+missed = sorted(key for key, count in counts.items() if count == 0)
+with open(uncovered, "w", encoding="utf-8") as output:
+    for filename, line, column in missed:
+        output.write(f"{filename}:{line}:{column}\n")
+
+covered = len(counts) - len(missed)
+percentage = covered * 100 / len(counts) if counts else 100
+print(f"{percentage:8.2f}% {covered:6d}/{len(counts):<6d} TOTAL functions")
+PYTHON
+
+UNCOVERED_FUNCTIONS="$WORK_DIRECTORY/uncovered.txt"
+python3 "$FUNCTIONS_PROGRAM" "$LLVM_COV" "$PROFILE_DATA" "$EXCLUDE" \
+  "$OBJECT_LIST" "$UNCOVERED_FUNCTIONS" >> "$WORK_DIRECTORY/summary.txt"
+
+# Optionally require every function under measurement to be reached by the
+# suite. Only the platform that the report is published from is held to it, as
+# a report produced elsewhere measures a different set of code
+if [ -n "${REQUIRE_FULL_FUNCTION_COVERAGE:-}" ] && [ -s "$UNCOVERED_FUNCTIONS" ]
+then
+  echo "The test suite never calls the functions starting at:" >&2
+  cat "$UNCOVERED_FUNCTIONS" >&2
+  exit 1
+fi
+
 # The browsable report keeps the combined view. Its annotated sources can still
 # under count the header inline cases described above, so the summary file
 # carries the authoritative numbers