[libc++] Add an optional integration of libc++'s test suite with SPEC (#156953)

This allows running the SPEC benchmarks as part of libc++'s own
benchmarks by providing an external installation of SPEC.

Author: ldionne

libcxx/test/benchmarks/spec.gen.py

@@ -0,0 +1,78 @@
+# ===----------------------------------------------------------------------===##
+#
+# Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+# See https://llvm.org/LICENSE.txt for license information.
+# SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+#
+# ===----------------------------------------------------------------------===##
+
+# REQUIRES: enable-spec-benchmarks
+
+# RUN: mkdir -p %T
+# RUN: echo "%{cxx}" > %T/cxx.subs
+# RUN: echo "%{compile_flags}" > %T/compile_flags.subs
+# RUN: echo "%{flags}" > %T/flags.subs
+# RUN: echo "%{link_flags}" > %T/link_flags.subs
+# RUN: echo "%{spec_dir}" > %T/spec_dir.subs
+# RUN: %{python} %s %T
+# END.
+
+import json
+import pathlib
+import sys
+
+test_dir = pathlib.Path(sys.argv[1])
+cxx = (test_dir / 'cxx.subs').open().read().strip()
+compile_flags = (test_dir / 'compile_flags.subs').open().read().strip()
+flags = (test_dir / 'flags.subs').open().read().strip()
+link_flags = (test_dir / 'link_flags.subs').open().read().strip()
+spec_dir = pathlib.Path((test_dir / 'spec_dir.subs').open().read().strip())
+
+# Setup the configuration file
+test_dir.mkdir(parents=True, exist_ok=True)
+spec_config = test_dir / 'spec-config.cfg'
+spec_config.write_text(f"""
+default:
+    ignore_errors        = 1
+    iterations           = 1
+    label                = spec-stdlib
+    log_line_width       = 4096
+    makeflags            = --jobs=8
+    mean_anyway          = 1
+    output_format        = csv
+    preenv               = 0
+    reportable           = 0
+    tune                 = base
+    copies               = 1
+    threads              = 1
+    CC                   = cc -O3
+    CXX                  = {cxx} {compile_flags} {flags} {link_flags} -Wno-error
+    CC_VERSION_OPTION    = --version
+    CXX_VERSION_OPTION   = --version
+    EXTRA_PORTABILITY    = -DSPEC_NO_CXX17_SPECIAL_MATH_FUNCTIONS # because libc++ doesn't implement the special math functions yet
+""")
+
+# Build the list of benchmarks. We take all intrate and fprate benchmarks that contain C++ and
+# discard the ones that contain Fortran, since this test suite isn't set up to build Fortran code.
+spec_benchmarks = set()
+no_fortran = set()
+with open(spec_dir / 'benchspec' / 'CPU' / 'intrate_any_cpp.bset', 'r') as f:
+    spec_benchmarks.update(json.load(f)['benchmarks'])
+with open(spec_dir / 'benchspec' / 'CPU' / 'fprate_any_cpp.bset', 'r') as f:
+    spec_benchmarks.update(json.load(f)['benchmarks'])
+with open(spec_dir / 'benchspec' / 'CPU' / 'no_fortran.bset', 'r') as f:
+    no_fortran.update(json.load(f)['benchmarks'])
+spec_benchmarks &= no_fortran
+
+for benchmark in spec_benchmarks:
+    print(f'#--- {benchmark}.sh.test')
+    print(f'RUN: rm -rf %T') # clean up any previous (potentially incomplete) run
+    print(f'RUN: mkdir %T')
+    print(f'RUN: cp {spec_config} %T/spec-config.cfg')
+    print(f'RUN: %{{spec_dir}}/bin/runcpu --config %T/spec-config.cfg --size train --output-root %T --rebuild {benchmark}')
+    print(f'RUN: rm -rf %T/benchspec') # remove the temporary directory, which can become quite large
+
+    # 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-result %T/result/CPUv8.001.*.train.csv --output-format=lnt > %T/results.lnt')
+    print(f'RUN: cat %T/results.lnt')

libcxx/utils/libcxx/test/params.py

@@ -374,6 +374,17 @@ def getSuitableClangTidy(cfg):
         help="Whether to run the benchmarks in the test suite, to only dry-run them or to disable them entirely.",
         actions=lambda mode: [AddFeature(f"enable-benchmarks={mode}")],
     ),
+    Parameter(
+        name="spec_dir",
+        type=str,
+        default="none",
+        help="Path to the SPEC benchmarks. This is required in order to run the SPEC benchmarks as part of "
+             "the libc++ test suite. If provided, the appropriate SPEC toolset must already be built and installed.",
+        actions=lambda spec_dir: [
+            AddSubstitution("%{spec_dir}", spec_dir),
+            AddFeature('enable-spec-benchmarks')
+        ] if spec_dir != "none" else [],
+    ),
     Parameter(
         name="long_tests",
         choices=[True, False],

libcxx/utils/parse-spec-result

@@ -0,0 +1,108 @@
+#!/usr/bin/env python3
+
+import argparse
+import csv
+import sys
+
+def parse_table(rows, table_title):
+    """
+    Parse a CSV table out of an iterator over rows.
+
+    Return a tuple containing (extracted headers, extracted rows).
+    """
+    in_table = False
+    rows_iter = iter(rows)
+    extracted = []
+    headers = None
+    while True:
+        try:
+            row = next(rows_iter)
+        except StopIteration:
+            break
+
+        if not in_table and row == [table_title]:
+            in_table = True
+            next_row = next(rows_iter)
+            assert next_row == [], f'There should be an empty row after the title of the table, found {next_row}'
+            headers = next(rows_iter) # Extract the headers
+            continue
+
+        elif in_table and row == []: # An empty row marks the end of the table
+            in_table = False
+            break
+
+        elif in_table:
+            extracted.append(row)
+
+    assert len(extracted) != 0, f'Could not extract rows from the table, this is suspicious. Table title was {table_title}'
+    assert headers is not None, f'Could not extract headers from the table, this is suspicious. Table title was {table_title}'
+
+    return (headers, extracted)
+
+def main(argv):
+    parser = argparse.ArgumentParser(
+        prog='parse-spec-results',
+        description='Parse SPEC result files (in CSV format) and extract the selected result table, in the selected format.')
+    parser.add_argument('filename', type=argparse.FileType('r'), nargs='+',
+        help='One of more CSV files to extract the results from. The results parsed from each file are concatenated '
+             'together, creating a single CSV table.')
+    parser.add_argument('--table', type=str, choices=['full', 'selected'], default='full',
+        help='The name of the table to extract from SPEC results. `full` means extracting the Full Results Table '
+             'and `selected` means extracting the Selected Results Table. Default is `full`.')
+    parser.add_argument('--output-format', type=str, choices=['csv', 'lnt'], default='csv',
+        help='The desired output format for the data. `csv` is CSV format and `lnt` is a format compatible with '
+             '`lnt importreport` (see https://llvm.org/docs/lnt/importing_data.html#importing-data-in-a-text-file).')
+    parser.add_argument('--extract', type=str,
+        help='A comma-separated list of headers to extract from the table. If provided, only the data associated to '
+             'those headers will be present in the resulting data. Invalid header names are diagnosed. Please make '
+             'sure to use appropriate quoting for header names that contain spaces. This option only makes sense '
+             'when the output format is CSV.')
+    parser.add_argument('--keep-not-run', action='store_true',
+        help='Keep entries whose \'Base Status\' is marked as \'NR\', aka \'Not Run\'. By default, such entries are discarded.')
+    args = parser.parse_args(argv)
+
+    if args.table == 'full':
+        table_title = 'Full Results Table'
+    elif args.table == 'selected':
+        table_title = 'Selected Results Table'
+
+    # Parse the headers and the rows in each file, aggregating all the results
+    headers = None
+    rows = []
+    for file in args.filename:
+        reader = csv.reader(file)
+        (parsed_headers, parsed_rows) = parse_table(reader, table_title)
+        assert headers is None or headers == parsed_headers, f'Found files with different headers: {headers} and {parsed_headers}'
+        headers = parsed_headers
+        rows.extend(parsed_rows)
+
+    # Remove rows that were not run unless we were asked to keep them
+    if not args.keep_not_run:
+        not_run = headers.index('Base Status')
+        rows = [row for row in rows if row[not_run] != 'NR']
+
+    if args.extract is not None:
+        if args.output_format != 'csv':
+            raise RuntimeError('Passing --extract requires the output format to be csv')
+        for h in args.extract.split(','):
+            if h not in headers:
+                raise RuntimeError(f'Header name {h} was not present in the parsed headers {headers}')
+
+        extracted_fields = [headers.index(h) for h in args.extract.split(',')]
+        headers = [headers[i] for i in extracted_fields]
+        rows = [[row[i] for i in extracted_fields] for row in rows]
+
+    # Print the results in the right format
+    if args.output_format == 'csv':
+        writer = csv.writer(sys.stdout)
+        writer.writerow(headers)
+        for row in rows:
+            writer.writerow(row)
+    elif args.output_format == 'lnt':
+        benchmark = headers.index('Benchmark')
+        time = headers.index('Est. Base Run Time')
+        for row in rows:
+            print(f'{row[benchmark].replace('.', '_')}.execution_time {row[time]}')
+
+if __name__ == '__main__':
+    main(sys.argv[1:])