[Debug Info] replace lldb_commands with __lldb_init_module

[Walnut356]

View all comments

DO NOT MERGE. Please run through aarch64-apple. Ideally the tests fail. If they don't, the test runner is using too old of a version of LLDB and I need to fiddle with this some more.

TL;DR I hate lldb_commands, it's annoying to update and maintain since it's just a raw text file of CLI commands. All of the functionality can be replicated via the debugger's own API. __lldb_init_module is called immediately upon command script import lldb_lookup.py, so it should function identically (with one exception, see below). We can do things like let python tell us the names of the functions and classes, making them easier to keep in sync. Most importantly, we can conditionally use one set of visualizers or another depending on the version of the debugger. This is a precursor to using callback-based type matching, which was added in LLDB 19.

The one exception i mentioned above replaces our ".*" regex that dumps everything through lldb_lookup.synthetic_lookup. That wildcard also includes primitives which is less than ideal, especially for performance. It's bad enough that codelldb intercepts this command and replaces it with a type recognizer. If i remove lldb_commands, codelldb can no longer do this. To prevent a regression on their end, I added a conditional that replicates that behavior.

Here's the problem: for some godforsaken reason, if a struct contains a field with a SyntheticProvider (even one that is "empty", like for DefaultSyntheticProvider for primitives) it is formatted differently than a struct whose fields don't have providers. That breaks a bunch of the tests because instead of { x = 10 y = 30 } we get (x = 10, y = 30). Very cool.

So, we have divergent behavior depending on the version of LLDB. That's nothing new, since MSVC enums are broken prior to LLDB 18, so it's whatever (and the difference hardly matters to the end user). If the CI test runners have a new enough version of LLDB that it uses the type recognizer instead of the wildcard, the tests will fail. I'll update them and we'll be good to go. If they don't fail, I need to probably force a struct summary provider so that all versions look the same (which isn't ideal performance), or I can just wait until xcode happens to update to a newer version of LLDB.

---

try-job: aarch64-apple

[rustbot]

Some changes occurred in src/tools/compiletest

cc @jieyouxu

[rustbot]

r? @Mark-Simulacrum

rustbot has assigned @Mark-Simulacrum.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

Why was this reviewer chosen?

The reviewer was selected based on:

• Owners of files modified in this PR: @Mark-Simulacrum

[Walnut356]

now that i think on it, apple's LLDB is a custom fork. I don't think the version extraction logic will work for it. Lemme fix that real quick.

[jieyouxu]

You should be able to run try jobs on this PR urself after this:
@bors delegate=try

[rust-bors]

✌️ @Walnut356, you can now perform try builds on this pull request!

You can now post @bors try to start a try build.

[Walnut356]

rad, thank you

[Walnut356]

@bors try jobs=aarch64-apple

[rust-bors]

💔 Test for d52c8f9 failed: CI. Failed job:

• try - aarch64-apple (web logs, enhanced plaintext logs)

[Walnut356]

Okay, bad news is this test didn't go exactly as planned because debugger.version was only added recently and is not present in apple's LLDB. The good news is:

1. I can confirm it's not present via apple's repo (llvm's for comparison) so I should be able to check for similar things without running a try job.
2. apple's LLDB does have the type recognizer enum value so those should work

i'm still iffy on whether I want to just update the tests, or if i want to force a StructSummaryProvider and save us from annoyances like this in the future (at the cost of some minor performance losses).

[Walnut356]

I guess I should also note that, in additional bad news, apples versioning scheme seems to have 0 relation to llvm's, so using raw version numbers probably won't work. I should be able to just check for features via getattr to check what functions and values are present at runtime

[Walnut356]

i'm still iffy on whether I want to just update the tests, or if i want to force a StructSummaryProvider

I think i'm just gonna update the tests for now since it's a bit safer. I can always go in and enforce StructSummaryProvider later.

[Walnut356]

Nevermind, I'm really not fond of the formatting inconsistency in nested structs. StructSyntheticProvider it is.

[rust-bors]

☀️ Try build successful (CI)
Build commit: 8bf1aaf (8bf1aaf48353b1c0cf73c1575d21c7073488c099, parent: 8afb6a8b1b32fce2f8aa7520517833338dc36c5e)

[Walnut356]

woo! It should be good to go then

[jieyouxu]

(Will try to get to this next Monday or so)

[rustbot]

This PR was rebased onto a different main commit. Here's a range-diff highlighting what actually changed.

Rebasing is a normal part of keeping PRs up to date, so no action is needed—this note is just to help reviewers.

[jieyouxu]

This largely looks good to me, and I also tried this locally. I do like the __lldb_init_module approach much better than the plain-text lldb_commands which was a pain to read let alone write.

Only one thing with ‎tests/debuginfo/captured-fields-1.rs‎, where I think the apple lldb used in aarch64-apple is borked on this test case.

@rustbot author

View changes since this review

[jieyouxu]

Review notes:

• I tested the changes locally on a aarch64-apple-darwin host w/:
  • Apple lldb 2100
  • LLVM lldb 22.1.6

[jieyouxu]

Review notes: see https://lldb.llvm.org/use/tutorials/writing-custom-commands.html#loading-commands for __lldb_init_module

[jieyouxu]

Problem: on llvm lldb 22 locally (aarch64-apple-darwin host), this check is failing for me.

LLVM lldb 22

lldb version 22.1.6

v test
(captured_fields_1::main::{closure_env#0}) test = {_ref__my_ref__my_field1:0x000000016fdfe520}
continue
v test
(captured_fields_1::main::{closure_env#1}) test = {_ref__my_ref__my_field2:0x000000016fdfe524}
continue
v test
(captured_fields_1::main::{closure_env#2}) test = {_ref__my_ref:0x000000016fdfe528}
continue
v test
(captured_fields_1::main::{closure_env#3}) test = {my_ref:{my_field1:11, my_field2:22}}
continue
v test
(captured_fields_1::main::{closure_env#4}) test = {my_var__my_field2:22}
continue
v test
(captured_fields_1::main::{closure_env#5}) test = {my_var:{my_field1:11, my_field2:22}}
continue
quit

Apple lldb 2100

lldb-2100.0.17.108
Apple Swift version 6.3.2 (swiftlang-6.3.2.1.108 clang-2100.1.1.101)

v test
(captured_fields_1::main::{closure_env#0}) test = {_ref__my_ref__my_field1:0x000000016fdfe450}
continue
v test
(captured_fields_1::main::{closure_env#1}) test = {_ref__my_ref__my_field2:0x000000016fdfe454}
continue
v test
(captured_fields_1::main::{closure_env#2}) test = {_ref__my_ref:0x000000016fdfe458}
continue
v test
(captured_fields_1::main::{closure_env#3}) test = {my_ref:{my_field1:11, my_field2:22}}
continue
v test
(captured_fields_1::main::{closure_env#4}) test = {my_var__my_field2:22}
continue
v test
(captured_fields_1::main::{closure_env#5}) test = {my_var:{my_field1:11, my_field2:22}}
continue
quit

[jieyouxu]

IOW something seems broken with this with the lldb that aarch64-apple CI has (LLDB version 1703) 🤔 I think that is a bug with an older Apple lldb version.
I would suggest gating this with minimum lldb version 21 (or 2100 apple lldb...), or we can look into bumping the lldb that aarch64-apple is using, I suppose 🤔

[jieyouxu]

It looks like the macos-15 runner in our aarch64-apple CI currently has Apple LLVM 17 as the baseline... It can become Apple LLVM 21 baseline if we can figure out how to bump to macos-26 without having that exceed the 6h timeout... 🤦:

• [EXPERIMENTAL] ci: use macos-26 for aarch64-apple job #157231
• [EXPERIMENTAL] ci: use macos-26 for aarch64-apple job (test library --stage=2 only) #157232

[Walnut356]

huh, fair enough. I'm almost tempted to figure out what actually changed in the printing logic, but that whole submodule of LLDB is really obtuse and i don't think I have it in me to debug LLDB atm.

At the very least, the old and new outputs both happen to be acceptable ways of visualizing the struct from an end-user perspective

[jieyouxu]

Yes. And I'm not so worried about this particular case because Apple lldb 2100+ seems to not, you know, give up on this case (and actually seems to converge back with whatever formatting LLVM lldb 22 gives). It's the Apple lldb 17xx series that seem to have this problem.

[rustbot]

Reminder, once the PR becomes ready for a review, use @rustbot ready.

[jieyouxu]

Thanks. You can r=me once PR CI is green. I retested ur change on Apple lldb 2100 / LLVM lldb 22.1.6.
@bors rollup=never
@bors delegate+

[rust-bors]

✌️ @Walnut356, you can now approve this pull request!

If @jieyouxu told you to "r=me" after making some further change, then please make that change and post @bors r=jieyouxu.

View changes since this delegation.

[Walnut356]

@bors r=jieyouxu

[rust-bors]

📌 Commit 7f84e5c has been approved by jieyouxu

It is now in the queue for this repository.

[rust-bors]

☀️ Test successful - CI
Approved by: jieyouxu
Duration: 3h 31m 25s
Pushing 4804ad7 to main...

[github-actions]

What is this?

This is an experimental post-merge analysis report that shows differences in test outcomes between the merged PR and its parent PR.

Comparing 4530eac (parent) -> 4804ad7 (this PR)

Test differences

Show 5 test diffs

Stage 2

• [debuginfo-lldb] tests/debuginfo/captured-fields-1.rs: pass -> ignore (ignored when the LLDB version is 2100) (J0)

Additionally, 4 doctest diffs were found. These are ignored, as they are noisy.

Job group index

• J0: aarch64-apple

Test dashboard

Run

cargo run --manifest-path src/ci/citool/Cargo.toml -- \
    test-dashboard 4804ad7e93e1b31f4605b7083871d0d3d85a2afe --output-dir test-dashboard

And then open test-dashboard/index.html in your browser to see an overview of all executed tests.

Job duration changes

1. x86_64-gnu-gcc-core-tests: 14m 41s -> 7m 54s (-46.1%)
2. dist-aarch64-linux: 1h 50m -> 2h 31m (+36.4%)
3. x86_64-gnu-llvm-21-3: 1h 59m -> 1h 28m (-26.0%)
4. i686-msvc-2: 2h 15m -> 1h 40m (-25.5%)
5. arm-android: 1h 45m -> 1h 20m (-23.3%)
6. dist-x86_64-msvc-alt: 2h 37m -> 3h 12m (+22.5%)
7. pr-check-1: 33m 32s -> 26m 17s (-21.6%)
8. dist-loongarch64-linux: 1h 53m -> 1h 30m (-20.5%)
9. dist-powerpc64-linux-gnu: 1h 18m -> 1h 33m (+19.5%)
10. x86_64-msvc-ext3: 1h 46m -> 1h 25m (-19.1%)

How to interpret the job duration changes?

Job durations can vary a lot, based on the actual runner instance
that executed the job, system noise, invalidated caches, etc. The table above is provided
mostly for t-infra members, for simpler debugging of potential CI slow-downs.

[rust-timer]

Finished benchmarking commit (4804ad7): comparison URL.

Overall result: ❌✅ regressions and improvements - no action needed

@rustbot label: -perf-regression

Instruction count

Our most reliable metric. Used to determine the overall result above. However, even this metric can be noisy.

	mean	range	count
Regressions ❌ (primary)	-	-	0
Regressions ❌ (secondary)	0.1%	[0.1%, 0.1%]	1
Improvements ✅ (primary)	-	-	0
Improvements ✅ (secondary)	-0.3%	[-0.3%, -0.3%]	1
All ❌✅ (primary)	-	-	0

Max RSS (memory usage)

This perf run didn't have relevant results for this metric.

Cycles

Results (primary 0.1%)

A less reliable metric. May be of interest, but not used to determine the overall result above.

	mean	range	count
Regressions ❌ (primary)	3.0%	[3.0%, 3.0%]	1
Regressions ❌ (secondary)	-	-	0
Improvements ✅ (primary)	-2.8%	[-2.8%, -2.8%]	1
Improvements ✅ (secondary)	-	-	0
All ❌✅ (primary)	0.1%	[-2.8%, 3.0%]	2

Binary size

This perf run didn't have relevant results for this metric.

Bootstrap: 510.437s -> 509.892s (-0.11%)
Artifact size: 400.71 MiB -> 400.78 MiB (0.02%)