Merge branch 'main' into xegpu-scatter-ops

Author: nbpatel

.ci/all_requirements.txt

@@ -0,0 +1,80 @@
+# 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
+"""Caches .lit_test_times.txt files between premerge invocations.
+
+.lit_test_times.txt files are used by lit to order tests to best take advantage
+of parallelism. Having them around and up to date can result in a ~15%
+improvement in test times. This script downloading cached test time files and
+uploading new versions to the GCS buckets used for caching.
+"""
+
+import sys
+import os
+import logging
+import multiprocessing.pool
+import pathlib
+import glob
+
+from google.cloud import storage
+
+GCS_PARALLELISM = 100
+
+
+def _maybe_upload_timing_file(bucket, timing_file_path):
+    if os.path.exists(timing_file_path):
+        timing_file_blob = bucket.blob("lit_timing/" + timing_file_path)
+        timing_file_blob.upload_from_filename(timing_file_path)
+
+
+def upload_timing_files(storage_client, bucket_name: str):
+    bucket = storage_client.bucket(bucket_name)
+    with multiprocessing.pool.ThreadPool(GCS_PARALLELISM) as thread_pool:
+        futures = []
+        for timing_file_path in glob.glob("**/.lit_test_times.txt", recursive=True):
+            futures.append(
+                thread_pool.apply_async(
+                    _maybe_upload_timing_file, (bucket, timing_file_path)
+                )
+            )
+        for future in futures:
+            future.get()
+    print("Done uploading")
+
+
+def _maybe_download_timing_file(blob):
+    file_name = blob.name.removeprefix("lit_timing/")
+    pathlib.Path(os.path.dirname(file_name)).mkdir(parents=True, exist_ok=True)
+    blob.download_to_filename(file_name)
+
+
+def download_timing_files(storage_client, bucket_name: str):
+    bucket = storage_client.bucket(bucket_name)
+    blobs = bucket.list_blobs(prefix="lit_timing")
+    with multiprocessing.pool.ThreadPool(GCS_PARALLELISM) as thread_pool:
+        futures = []
+        for timing_file_blob in blobs:
+            futures.append(
+                thread_pool.apply_async(
+                    _maybe_download_timing_file, (timing_file_blob,)
+                )
+            )
+        for future in futures:
+            future.get()
+    print("Done downloading")
+
+
+if __name__ == "__main__":
+    if len(sys.argv) != 2:
+        logging.fatal("Expected usage is cache_lit_timing_files.py <upload/download>")
+        sys.exit(1)
+    action = sys.argv[1]
+    storage_client = storage.Client()
+    bucket_name = os.environ["CACHE_GCS_BUCKET"]
+    if action == "download":
+        download_timing_files(storage_client, bucket_name)
+    elif action == "upload":
+        upload_timing_files(storage_client, bucket_name)
+    else:
+        logging.fatal("Expected usage is cache_lit_timing_files.py <upload/download>")
+        sys.exit(1)

.ci/cache_lit_timing_files.py

@@ -333,6 +333,7 @@ def get_env_variables(modified_files: list[str], platform: str) -> Set[str]:
     current_platform = platform.system()
     if len(sys.argv) == 2:
         current_platform = sys.argv[1]
-    env_variables = get_env_variables(sys.stdin.readlines(), current_platform)
+    changed_files = [line.strip() for line in sys.stdin.readlines()]
+    env_variables = get_env_variables(changed_files, current_platform)
     for env_variable in env_variables:
         print(f"{env_variable}='{env_variables[env_variable]}'")

.ci/compute_projects.py

@@ -27,6 +27,13 @@ def _parse_ninja_log(ninja_log: list[str]) -> list[tuple[str, str]]:
             # We hit the end of the log without finding a build failure, go to
             # the next log.
             return failures
+        # If we are doing a build with LLVM_ENABLE_RUNTIMES, we can have nested
+        # ninja invocations. The sub-ninja will print that a subcommand failed,
+        # and then the outer ninja will list the command that failed. We should
+        # ignore the outer failure.
+        if ninja_log[index - 1].startswith("ninja: build stopped:"):
+            index += 1
+            continue
         # We are trying to parse cases like the following:
         #
         # [4/5] test/4.stamp

.ci/generate_test_report_lib.py

@@ -126,6 +126,39 @@ def test_ninja_log_multiple_failures(self):
             ),
         )
 
+    # Test that we can correctly handle the runtimes build. the LLVM runtimes
+    # build will involve ninja invoking more ninja processes within the
+    # runtimes directory. This means that we see two failures for a failure in
+    # the runtimes build: one from the inner ninja containing the actual action
+    # that failed, and one for the sub ninja invocation that failed.
+    def test_ninja_log_runtimes_failure(self):
+        failures = generate_test_report_lib.find_failure_in_ninja_logs(
+            [
+                [
+                    "[1/5] test/1.stamp",
+                    "[2/5] test/2.stamp",
+                    "FAILED: touch test/2.stamp",
+                    "Wow! This system is really broken!",
+                    "ninja: build stopped: subcommand failed.",
+                    "FAILED: running check-runtime failed.",
+                    "<some random command>",
+                    "ninja: build stopped: subcommand failed.",
+                ]
+            ]
+        )
+        self.assertEqual(len(failures), 1)
+        self.assertEqual(
+            failures[0],
+            (
+                "test/2.stamp",
+                dedent(
+                    """\
+                    FAILED: touch test/2.stamp
+                    Wow! This system is really broken!"""
+                ),
+            ),
+        )
+
     def test_title_only(self):
         self.assertEqual(
             generate_test_report_lib.generate_report("Foo", 0, [], []),