Rework relooper algorithm for fun and profit by randomPoison · Pull Request #1558 · immunant/c2rust

randomPoison · 2026-01-21T20:37:40Z

This PR reworks the relooper algorithm to replace most usages of current_block with labeled blocks/breaks. Addresses #330. Closes #1398.

The bulk of the changes are in relooper.rs and structures.rs:

In relooper.rs the core logic is in the relooper function, which takes the unstructured CFG and processes it into the "structured CFG".
In structures.rs we have the logic that takes the structured CFG and turns it into the "strucured AST", which is an intermediate AST that just represents the structured control flow we're translating.
- I've replaced structured_cfg_help with a new process_cfg function that is responsible for generating the labeled blocks/breaks.
- I've also added a secondary cleanup pass cleanup_labels that removes redundant labels and extra labeled blocks. This allows the logic in process_cfg to be simpler by conservatively labeling all blocks and breaks.

Don't try to review this commit-by-commit, the commit history is NOT clean and shows all of the experimentation and iteration I went through to get to this point. I only have all the commit history intact because I haven't had a need to squash yet, and we should squash this when merging to avoid polluting the commit history.

I'd also recommend not worrying too much about the diff here, and instead just review the new code as if it were all entirely fresh. Many of the parts here have been heavily reworked, and since I'm the only person deeply familiar with what the code was doing before it probably doesn't make sense to worry about how the new code differs from the old version.

NOTE: I still need to do a cleanup pass on the unit tests, since a lot of those were slapped together during experimentation. Snapshot tests are all clean, but I may want to add more as I cleanup the unit tests.

218a281 shows the diff between the old behavior and the new behavior in the snapshots.

- Try to create a loop if there's a single entry. - Attempt to create a multiple if there are any handled entries/blocks. - Fallback to a loop only if we can't create a multiple. This brings us more in line with the original paper as well, and already fixes the disjoint loops case. Also make a bunch of not-strictly-necessary changes to make things more organized and easier to work with. These should be split out into a separate PR before the functional refactoring.

Because apparently I have a different compiler on my laptop and it's mad about the missing statements, even though on my computer at home it works lolsob

Also differentiate between breaks and continues when relooping.

... by not generating a break if there's only one entry following a simple shape.

And also make the json output pretty.

randomPoison · 2026-01-21T21:13:11Z

Oh joy, we have a subtle difference in behavior between platforms. Looks like the CFG labels that we generate on macos are slightly different than the labels we generate on Ubuntu, likely because the compiler we're using is different and is generating a slightly different CFG.

randomPoison · 2026-01-23T21:25:02Z

Okay the tests should be clean and ready for review as well.

kkysen

For the snapshots, could you have one commit where they're tested with the previous relooper, and then a new commit where they're updated so I can see the diffs? Or if this is already the case, tell me which commits to compare, because there's a lot of them so I'm not sure where to look.

scripts/test_translator.py

kkysen · 2026-02-24T03:38:40Z

.gitignore

+# Relooper debug info
+/dumps
+c2rust-transpile/dumps


Should we add relooper in the name somewhere instead of just dumps?

It's not necessarily relooper-specific, if there are other parts of the transpile or refactor process where we want to dump large quantities of data for debugging we can drop it in there.

I did update the comment to make it non-relooper-specific though.

kkysen · 2026-02-24T03:39:18Z

tests/unit/gotos/src/test_goto_error.rs

@@ -0,0 +1,87 @@
+#[test]
+pub fn test_goto_error_a() {
+    use crate::goto_error::rust_goto_error_a;


Can these just all be top-level imports?

Same for all of the other ones.

For these it's helpful to import the function names within each test function because that makes it a bit easier to comment out tests for debugging, i.e. if I want to comment out some of the test functions in the original C file, I also have to comment out the corresponding test function and import in the Rust test file. If the import is in the corresponding function I can just comment out the function, whereas if the imports are at the top of the file I have to comment out the test function and the import separately. It's not a big deal either way, but there's also not much of a reason to move the imports.

c2rust-transpile/src/cfg/mod.rs

kkysen · 2026-02-24T03:46:49Z

c2rust-transpile/src/translator/mod.rs

+            use std::io::Write;
+            std::fs::create_dir_all("dumps").unwrap();
+            let mut file = std::fs::File::create(&file_name).unwrap();
+            write!(&mut file, "{:#?}", relooped).unwrap();


Suggested change

use std::io::Write;

std::fs::create_dir_all("dumps").unwrap();

let mut file = std::fs::File::create(&file_name).unwrap();

write!(&mut file, "{:#?}", relooped).unwrap();

fs::create_dir_all("dumps").unwrap();

fs::write(&file_name, format!("{relooped:#?}")).unwrap();

File isn't buffered, so we don't want to use that directly. It should be simpler to just write all in one go.

What's the issue here with File not being buffered? I assume that's some kind of minor performance issue? The docs say to use a buffered write when doing many small writes, but we're just doing one big write so it's not clear to me that it matters here.

c2rust-transpile/src/translator/mod.rs

randomPoison · 2026-03-09T19:18:03Z

For the snapshots, could you have one commit where they're tested with the previous relooper, and then a new commit where they're updated so I can see the diffs? Or if this is already the case, tell me which commits to compare, because there's a lot of them so I'm not sure where to look.

@kkysen There's not a good way to do that without rewriting a lot of history. We didn't have snapshot tests for the CFG stuff before, so they're pretty much all new.

kkysen · 2026-03-09T19:31:58Z

There's not a good way to do that without rewriting a lot of history. We didn't have snapshot tests for the CFG stuff before, so they're pretty much all new.

Hmm. Can you run the snapshots on the previous transpiler in a commit somewhere so we can review the diff? It's a lot harder to review without that IMO.

randomPoison · 2026-03-09T20:06:39Z

Can you run the snapshots on the previous transpiler in a commit somewhere so we can review the diff?

Yeah that's easy enough to do, at least. 218a281 now shows the diff between the old behavior and the new behavior in the snapshots

kkysen · 2026-03-09T22:45:37Z

Yeah that's easy enough to do, at least. 218a281 now shows the diff between the old behavior and the new behavior in the snapshots

Thanks, that's very helpful.

randomPoison and others added 30 commits January 21, 2026 11:20

Notes and cleanup and whatnot

9b97e1a

Document transitive_closure and friends

04ce419

Moar comments

293e3c7

More judgemental comments

df180f8

More comments and such

9f368c4

Document why we rewrite GoTo to ExitTo in loops

021febf

Document the logic for matching a C multiple

035e39c

Document how/why absent entries are handled

884da90

Cleanup some comments

c1a8454

Add more comments about loop stuff

ab494b9

Document more of the subtleties of multiple creation

4877c26

Assert on unrecognized loop IDs

b045df3

Add overview of relooper

88641e4

Fix formatting

fa4b39d

DON'T PANIC!!!

a930767

Add some control flow tests

21ca7d6

Automatically build before running tests

4fb44ed

Disable heuristics for testing

0f7bbd2

Fix missing statements in tests

e225fc1

Because apparently I have a different compiler on my laptop and it's mad about the missing statements, even though on my computer at home it works lolsob

Support filtering C files when running tests

415800e

Setup simplified goto error test

b3f861c

Start setting up forward_cfg_help

3b80097

Handle followup multiples

f86af5f

Also differentiate between breaks and continues when relooping.

Get basic goto error example working

ee6b7bd

... by not generating a break if there's only one entry following a simple shape.

Dump relooper structures to files

5580c4d

Dump CFG json to dumps directory

b410fec

And also make the json output pretty.

Fix terminator for empty simple case

43589ac

Make goto_error into proper test

483587b

Print stderr when tests fail

2b95d44

Address the robot's feedback

fa9a82e

randomPoison added 2 commits January 21, 2026 15:02

Make switch snapshot tests os-specific

668c81c

Add macos version of switch snapshot

1377da9

This was referenced Jan 22, 2026

Labels generated by relooper can differ between platforms #1561

Open

transpile: Always explicitly reference current_block values #1556

Draft

transpile: Use enum for current_block labels #1557

Draft

Generate an enum for current_block labels #1523

Open

randomPoison added 5 commits January 22, 2026 16:43

Add snapshot tests for simple nested loops

fc84f34

Revert idiomatic_nested_loops.c

109bf61

Remove unused if_else.c unit test

23aef1e

Add snapshots for irreducible control flow

9769d00

Remove missing unit test

fc7d3b6

randomPoison added the control flow label Jan 23, 2026

randomPoison added 3 commits January 23, 2026 12:40

Setup snapshot tests for disjoint loops

19d4574

Make irreducible snapshot test os-specific

8a40527

Add macos snapshot for irreducible test

1021d0b

kkysen reviewed Feb 24, 2026

View reviewed changes

randomPoison added 4 commits March 6, 2026 14:24

Merge branch 'master' into legare/relooper-experiment

cd27d66

Update snapshot tests

dfb07ba

Import std::fmt stuff

ee924f2

Minor style cleanup

1f63ee3

randomPoison added 2 commits March 9, 2026 13:03

Run snapshots with old transpiler behavior

594e519

Update snapshots with new transpiler behavior

218a281

Add note about .ron extension in structure dumps

4167735

Conversation

randomPoison commented Jan 21, 2026 • edited by kkysen Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

randomPoison commented Jan 21, 2026

Uh oh!

randomPoison commented Jan 23, 2026

Uh oh!

kkysen left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

randomPoison commented Mar 9, 2026

Uh oh!

kkysen commented Mar 9, 2026

Uh oh!

randomPoison commented Mar 9, 2026

Uh oh!

kkysen commented Mar 9, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

randomPoison commented Jan 21, 2026 •

edited by kkysen

Loading