Problem when tree doesn't change but commit structure has changed

[ezyang]

Suppose you have commit A and B. You squash them into AB and run ghstack. This won't work; ghstack will say "well, B has the same tree as AB, so there's nothing to do." Even worse, when you're done, it will re check out B, undoing your AB merge.

[dan-zheng]

I think I've run into this issue.

---

I had a chain of commits, A-B-C-D.

I used git rebase -i B~ and squashed BC together: A-BC-D.

After rebasing, ghstack doesn't update the pull request for BC anymore.
https://github.com/google-research/triangulate/pull/16

Is there any way I can fix this?

---

Full git log below:

commit 7005442dd4f0c2be28d7e24e5ca14e1c91ff943f
Author: Dan Zheng <danielzheng@google.com>
Date:   Mon Aug 7 11:44:41 2023 -0700

    Update quoter.py test case.
    
    Add flags to quoter.py for conditional flag-dependent exception raising.
    
    Consider replacing quoter.py altogether in the future.
    
    ghstack-source-id: 21b43d1e4151c50ebf86c3cc84ddf458bcf17c8c
    Pull Request resolved: https://github.com/google-research/triangulate/pull/17

commit 24409a0b44b8ac97c75a4430a8b39c4c231d9545
Author: Dan Zheng <danielzheng@google.com>
Date:   Mon Aug 7 11:44:40 2023 -0700

    Update core logic and main binary.
    
    Use newly added utilities to improve core logic:
    - `ast_utils`: Improved AST matching for extracting illegal state expressions.
    - `instrumentation_utils`: Robust probe call insertion, instead of relying on
      `eval` and string interpolation.
    - `logging`: Prettier and useful logging statements.
    
    ---
    
    Previously, `main` used explicit flags like `bug_trap` (bug line number) to
    describe the subject program failure location. This led to long CLI invocations,
    which are error-prone and difficult to use.
    
    Now, `main` is designed like `pdb`, wrapping a standard Python script invocation:
    
    ```
    triangulate [flags...] subject -- [subject_args...]
    ```
    
    `main` runs the subject program with arguments. If an exception is thrown, the
    exception location is used as the bug line number. This `pdb`-like invocation mode
    is easier to use and test.
    
    `main` now does prettier logging as well, for status messages.
    
    ghstack-source-id: 4f16c3e131f0d3ab319f48ab5d06e78cab4fd29b
    Pull Request resolved: https://github.com/google-research/triangulate/pull/16

commit c0627945427756e929f0428eabbee302263eec18
Author: Dan Zheng <danielzheng@google.com>
Date:   Mon Aug 7 11:22:56 2023 -0700

    Add AST matching mini-library.
    
    Add an `AST` wrapper class with a method `select_nodes(selector)` method for
    returning all descendent AST nodes that match some predicate.
    
    Add AST node predicates:
    - `NodePredicate`: base class.
    - `All`: intersection combinator of multiple `NodePredicate`s.
    - `HasType(type)`: matches nodes with a given type.
    - `HasLocationInfo`: matches nodes that have line number information.
    - `OverlapsWithLineNumber(type)`: matches nodes that overlap with a line number.
    - `IsProbeStatement(type)`: matches nodes that are probe statements.
      - Currently matches all `triangulate_probe` function calls, which are
        generated via `instrumentation_utils.make_probe_call`.
    
    These will replace previous AST utilities for extracting `Assert` statements,
    which are less robust.
    
    ghstack-source-id: 4d23be5825365c89cd7cf5a58c6ea5eeb80c88dd
    Pull Request resolved: https://github.com/google-research/triangulate/pull/14

commit a85353e117c5c441bb7e8ed6e70bb4db632f41b2
Author: Dan Zheng <danielzheng@google.com>
Date:   Mon Aug 7 11:17:59 2023 -0700

    Add instrumentation utilities.
    
    Add a `probe` function for printing local variables (if defined).
    
    Add a `run_with_instrumentation` function for executing code and returning
    stdout and stderr as output.
    
    ghstack-source-id: ee48b5c8c5b9cb22b7177a638a9904e7963baa76
    Pull Request resolved: https://github.com/google-research/triangulate/pull/13

commit fbeee61d036e461da5664e893c31b5a7951f549a
Author: Dan Zheng <danielzheng@google.com>
Date:   Mon Aug 7 11:17:55 2023 -0700

    Add logging utilities.
    
    Add utilites for colored printing: convenience wrappers around termcolor.
    
    ghstack-source-id: 327a130c534ab13e452a4257a66e883003f5eac4
    Pull Request resolved: https://github.com/google-research/triangulate/pull/12

[dan-zheng]

Okay, I found a workaround.

I had a chain of commits, A-B-C-D.

I used git rebase -i B~ and squashed BC together: A-BC-D.

After rebasing, ghstack doesn't update the pull request for BC anymore. https://github.com/google-research/triangulate/pull/16

The workaround:

• Rebase and amend the parent of the squashed commit: git rebase -i <squashed commit>~2.
  • I guess this updates "the tree".
• Run ghstack again. Now PRs are updated, starting with the parent of the squashed commit.

I don't know how to apply this workaround if the squashed commit is the first PR in the chain.