Support workspace label prefixes in Starlark config and platform APIs

[gregestren]

This is a catch-all to focus the common theme of:

• It is not possible to set Starlark defined command line flags that come from a different repository #9065 (fixed)
• Custom build flags choke on prefix workspaces #9177 (fixed)
• Using new string_flag settings in external repositories #10152 (fixed)
• Broken repository name handling for build_settings and label_flag #10499 (fixed)
• Toolchain resolution stumbles over workspace prefixes in constraint labels #10927 (@katre is looking into this)
• Add support for user-defined build settings in external repositories #10975 (fixed)

At the narrowest this may require unrelated fixes to build_setting and toolchain resolution logic. But the general theme of the problem is clear, however the fixes are made.

We're (Bazel configurability) trying to prioritize this for the current quarter.

[gregestren]

Also #10118.

[aherrmann]

#11734 is a similar issue in aspects.

[gregestren]

@juliexxia - can you advise on how much of this issue is still relevant?

[nacl]

With 3.7, our experience with this is mostly the same as what's described in #9177, except Bazel no longer crashes, instead, the flag doesn't seem to take effect when it's prefixed with the workspace id.

Interestingly, when I run these two commands (within the example in #9177), it seems like the prefixed and non-prefixed versions are being treated as separate, but still work :

> bazel build //:my_drink --//:favorite_flavor=True
DEBUG: .../tmp/bazel-user-settings-example/rules.bzl:5:14: Get the opposite of default (True)
...
INFO: Build completed successfully, 1 total action
> bazel build //:my_drink --@bsws//:favorite_flavor=True
INFO: Build options --//:favorite_flavor and --@bsws//:favorite_flavor have changed, discarding analysis cache.
DEBUG: .../tmp/bazel-user-settings-example/rules.bzl:7:14: Get the default (False)
...
INFO: Build completed successfully, 1 total action

Referencing build settings from external WORKSPACEs (#9177 (comment)) seems to work well.

[stefanbucur]

I am experiencing the same behavior when it comes to using the flags in Bazel transitions.

For a WORKSPACE file defined as workspace(name = "my_workspace") and a build setting in my BUILD file as:

load("@bazel_skylib//rules:common_settings.bzl", "string_flag")

string_flag(
    name = "my_flag",
    build_setting_default = "none",
    visibility = ["//visibility:public"],
)

Then setting --@my_workspace//:my_flag and --//:my_flag end up changing distinct values. A Bazel transition on //:my_flag will not be affected when setting @my_workspace//:my_flag. I can actually list both versions as inputs to the transition, and they will be read as independent values.

[juliexxia]

Then setting --@my_workspace//:my_flag and --//:my_flag end up changing distinct values. A Bazel transition on //:my_flag will not be affected when setting @my_workspace//:my_flag. I can actually list both versions as inputs to the transition, and they will be read as independent values.

@stefanbucur would you mind writing out a full command line/transition for the scenarios you describe here? From my own testing I believe settings --@my_workspace//:my_flag and //:my_flag from inside that workspace successfully resolve to a single flag.

[moroten]

@juliexxia @stefanbucur I've modified #10499 into the following:

cat <<EOF > WORKSPACE
workspace(name = "my_workspace")

load("@bazel_tools//tools/build_defs/repo:http.bzl", "http_archive")
http_archive(
    name = "bazel_skylib",
    urls = [
        "https://github.com/bazelbuild/bazel-skylib/releases/download/1.0.3/bazel-skylib-1.0.3.tar.gz",
        "https://mirror.bazel.build/github.com/bazelbuild/bazel-skylib/releases/download/1.0.3/bazel-skylib-1.0.3.tar.gz",
    ],
    sha256 = "1c531376ac7e5a180e0237938a2536de0c54d93f5c278634818e0efc952dd56c",
)
EOF

cat <<EOF > BUILD.bazel
load("@bazel_skylib//rules:common_settings.bzl", "string_flag")
load(":rules.bzl", "rule_with_transition")

string_flag(
    name = "my_flag",
    build_setting_default = "none",
    visibility = ["//visibility:public"],
)
package(default_visibility = ["//visibility:public"])

rule_with_transition(
    name = "my_target",
)
EOF

cat <<EOF > rules.bzl
def _my_transition_impl(settings, attr):
    print(settings)
    return settings

my_transition = transition(
    implementation = _my_transition_impl,
    inputs = ["//:my_flag", "@my_workspace//:my_flag"],
    outputs = ["//:my_flag", "@my_workspace//:my_flag"],
)

def _rule_impl(ctx):
    pass

rule_with_transition = rule(
    implementation = _rule_impl,
    cfg = my_transition,
    attrs = {
        "_allowlist_function_transition": attr.label(default = "@bazel_tools//tools/allowlists/function_transition_allowlist"),
    },
)
EOF

Running Bazel 3.7.0 gives:

$ ~/bin/bazel-3.7.0-linux-x86_64 build my_target --//:my_flag=foo --@my_workspace//:my_flag=bar
Starting local Bazel server and connecting to it...
DEBUG: /.../rules.bzl:2:10: {"//:my_flag": "foo", "@my_workspace//:my_flag": "bar"}
DEBUG: /.../rules.bzl:2:10: {"//:my_flag": "foo", "@my_workspace//:my_flag": "bar"}
INFO: Analyzed target //:my_target (6 packages loaded, 7 targets configured).
INFO: Found 1 target...
Target //:my_target up-to-date (nothing to build)
INFO: Elapsed time: 3.378s, Critical Path: 0.02s
INFO: 1 process: 1 internal.
INFO: Build completed successfully, 1 total action

[juliexxia]

The general issue here seems to be that starlark options parsing and transitions aren't recognizing //:my_flag and @my_workspace//:my_flag as the same label.

This should be easy to fix during starlark options parsing - then in the command line you provided --@my_workspace//:my_flag=bar will simply overwrite --//:my_flag=foo and leave a configuration of //:my_flag=bar

Fixing this in transitions will take a little more digging. I'll keep looking.

[juliexxia]

@moroten to confirm re: the transition issue. There's nothing that you want to do but can't right? Using the non-workspace-prefixed version of //:my_flag in the transition works fine, just using @my_workspace//:my_flag doesn't do what's expected, is that correct?

[juliexxia]

@katre is taking a look at the issues described in #10927

[moroten]

@moroten to confirm re: the transition issue. There's nothing that you want to do but can't right? Using the non-workspace-prefixed version of //:my_flag in the transition works fine, just using @my_workspace//:my_flag doesn't do what's expected, is that correct?

We have worked around by avoiding user flags and settings in the main workspace. Instead, we have .bazelignored directories defined as local_repository where we put them instead.

[juliexxia]

That sounds reasonable to me. Out of curiosity, why did y'all chose that instead of keeping flags in the main workspace and always referring to them in transitions/bazelrc without the repo name attached? //:my_flag should infer @//:my_flag aka the main workspace. Did y'all try that and it didn't work or did you just avoid it due to the confusion described in this issue?

[stefanbucur]

In our case, the reason is that the transition and the build configuration flags are part of a library that is meant to be used in other projects: https://github.com/bazelbuild/rules_fuzzing (more precisely, the code is still under review here: bazelbuild/rules_fuzzing#86)

To simplify the specification of these flags on the command line, we encourage users to define their values as --configs in their own .bazelrc files, where they need to specify our workspace prefix (@rules_fuzzing) to refer to those, since our repository is imported as an http_archive. And to make things more uniform and easier to copy-paste, we also decided to use the workspace prefix in our own .bazelrc file.

Inside the transition implementation, it also wasn't clear to me how Bazel would interpret a //:flag label when the transition is imported in a different workspace. Will it assume it is part of the top-level repository, or inside the repository where the transition is defined? To eliminate any ambiguity, I ended up prepending the workspace prefix there too.

[stefanbucur]

Will it assume it is part of the top-level repository, or inside the repository where the transition is defined?

It turns out that the configuration flag does get interpreted in the context of the top-level workspace, so a rules repository library defining a transition on a configuration flag must reference that flag using the repository name. This means that our own .bazelrc file must reference those flags using the full repository name (see for example bazelbuild/rules_fuzzing#91).

If this is intended behavior, does it mean that a rules library offering configuration flags must always be imported using the name matching the hard-coded repo name used in transitions?

[stefanbucur]

Somewhat related to this discussion, I discovered another weird behavior. For label-typed flags (the label_flag rule), using the repo name prefix (@my_workspace//:my_label_flag) on the command line has no effect on the configuration, regardless of whether the flag is used with or without the prefix in the Starlark code.

In our case, we use a label flag to control a cc_library dependency in a cc_test rule. Setting that flag with the repo name prefix has no effect - the dependency is left to its default value.

Not sure if this is a separate issue or a manifestation of the same underlying issue.

[juliexxia]

These should all be fixed as of this most recent commit.