[libc++] Improve error resilience when running historical benchmarks

In benchmark-historical, don't skip gathering the results when the
lit command has failed. Indeed, it is expected to fail as part of
normal operation since it should be pretty frequent for at least one
test to fail when running historical benchmarks. Instead, gather
whatever results we have.

Also, output the build log in spec.gen.py upon failure so we can see
the reason for the failure.

Author: ldionne

libcxx/test/benchmarks/spec.gen.py

@@ -74,5 +74,5 @@
 
     # Parse the results into a LNT-compatible format. This also errors out if there are no CSV files, which
     # means that the benchmark didn't run properly (the `runcpu` command above never reports a failure).
-    print(f'RUN: %{{libcxx-dir}}/utils/parse-spec-results %T/result/*.train.csv --output-format=lnt > %T/results.lnt')
+    print(f'RUN: %{{libcxx-dir}}/utils/parse-spec-results %T/result/*.train.csv --output-format=lnt > %T/results.lnt || ! cat %T/result/*.log')
     print(f'RUN: cat %T/results.lnt')

libcxx/utils/benchmark-historical

@@ -90,10 +90,7 @@ def main(argv):
                 logging.info(f'Running {pretty}')
                 continue
 
-            if subprocess.call(test_cmd) != 0:
-                logging.error(f'Failed to run the tests for commit {commit}, skipping')
-                continue
-
+            subprocess.call(test_cmd)
             output_file.parent.mkdir(parents=True, exist_ok=True)
             consolidate_cmd = [(PARENT_DIR / 'consolidate-benchmarks'), build_dir, '--output', output_file]
             subprocess.check_call(consolidate_cmd)