rebaseline-cl: Add logic for selecting a related builder when filling in results.

third_party/WebKit/Tools/Scripts/webkitpy/tool/commands/rebaseline_cl.py

Index: third_party/WebKit/Tools/Scripts/webkitpy/tool/commands/rebaseline_cl.py
diff --git a/third_party/WebKit/Tools/Scripts/webkitpy/tool/commands/rebaseline_cl.py b/third_party/WebKit/Tools/Scripts/webkitpy/tool/commands/rebaseline_cl.py
index eeef1f691152de803998a1d6edf5b41a905a7afd..8852aa6042cbd7441c4ca62756dd3c2f94b12f89 100644
--- a/third_party/WebKit/Tools/Scripts/webkitpy/tool/commands/rebaseline_cl.py
+++ b/third_party/WebKit/Tools/Scripts/webkitpy/tool/commands/rebaseline_cl.py
@@ -250,13 +250,23 @@ class RebaselineCL(AbstractParallelRebaselineCommand):
                 test_baseline_set.add(test_prefix, build, port)
         return test_baseline_set
 
-    def _choose_fill_in_build(self, _, build_port_pairs):
+    def _choose_fill_in_build(self, target_port, build_port_pairs):
         """Returns a Build to use to supply results for the given port.
 
         Ideally, this should return a build for a similar port so that the
         results from the selected build may also be correct for the target port.
         """
-        # TODO(qyearsley): Decide what build to use for a given port
-        # in a more sophisticated way, such that a build with a
-        # "similar" port will be used when available.
-        return build_port_pairs[0][0]
+        def os_name(port):
+            if '-' not in port:

[wkorman, 2017/04/13 16:41:35]

Suggest adding comment to document what format we're expecting to parse here.
Could also be worth validating the format matches via a regexp and if we use a
regexp then we could just use the matched group. Thought is that, in future,
someone might unwittingly decide to add a new port with another hyphen, or
otherwise deviate accidentally.

[qyearsley, 2017/04/14 00:13:08]

Right, good idea to be cautious.

The actual format I'm expecting here is "<port>-<version>", or
"test-<port>-<version>"; now added a comment to explain this.

The full port names now are created by the determine_full_port_name in the Port
subclasses, which generally always returns <port>-<version>, except in a couple
cases, for example AndroidPort always just returns "android". And TestPort
returns "test-<os>-<version>".

With regards to validating this, there is a related bug about asserting a
baseline directory name convention (crbug.com/660234); also added a note to that
bug about this.

+                return port
+            return port[:port.rfind('-')]
+
+        # If any Build exists with the same OS, use the first one.

[wkorman, 2017/04/13 16:41:35]

What is the order of the builds at this point -- unsorted, or sorted by build
id, or something else? Probably best would be to use the most recent one?

Also if I am not mistaken not all builds have all baselines, sometimes they
don't bother running layout tests if the code change didn't trigger them, will
we fall back and reconsider other builds in that case? Or maybe for the builders
we have configured for use by rebaseline-cl they do always run all layout tests?

[qyearsley, 2017/04/14 00:13:08]

At this point, they should be in the order returned by GitCL.latest_try_jobs,
which is sorted (lexically) by builder name then build number, but including
only the latest build for reach builder. So, for example, it might be:
[Build("android_blink_rel", 3456), Build("linux_trusty_blink_rel", 6544),
Build("mac10.10_blink_rel, 32433"), ...]

https://cs.chromium.org/chromium/src/third_party/WebKit/Tools/Scripts/webkitp...
Right, the builders we have configured for use by rebaseline-cl (the
tryserver.blink bots linux_trusty_blink_rel, etc.) always run all the layout
tests.

At this point in the code, I think I'm assuming that there should definitely be
baselines available because earlier on, we should try to download the layout
test results from Google Storage, and only include a build if results were
successfully fetched already.

+        target_os = os_name(target_port)
+        same_os_builds = sorted(b for b, p in build_port_pairs if os_name(p) == target_os)
+        if same_os_builds:
+            return same_os_builds[0]
+
+        # Otherwise, perhaps any build will do, for example if the results are
+        # the same on all platforms. In this case, just return the first build.
+        return sorted(build_port_pairs)[0][0]