Stacking Patches in GitHub PRs

LLVM Project
1

What is the recommended approach for creating stacked patches in GitHub Pull Requests (PRs)?

Based on the discussion in Discord, it seems that I should follow these steps:

  1. Create branches in the upstream repository for each of the patches.

  2. Set up a sequence of PRs that are daisy-chained, with the following merge destination:

    • Patch #1 → main
    • Patch #2 → Patch #1
    • …
    • Patch #N → Patch #N-1

  3. When it’s time to merge, first merge the PR for the first patch.

  4. After merging the first PR, change the destination of the i-th patch to ‘main’ and then proceed to merge it.

Is this the correct and recommended method for handling stacked patches in GitHub PRs?

2

There’s been a bunch of discussion in Write a guide for doing "Stacked Reviews" with GitHub Pull Requests · Issue #56636 · llvm/llvm-project · GitHub, though it hasn’t reached any conclusion yet. In particular, any stacked PR mechanism will require the ability to create branches in the llvm/llvm-project repository instead of your own fork, which we don’t have agreement on yet.

1 Like
3

So, since there is no agreement on creating branches in the upstream, should I just create patches that contain commits that they depend on? And specify this in the PR message.

4

A few people are trying this already e.g. [TableGen] Use buildConstant to emit apply pattern immediates by Pierre-vh · Pull Request #66077 · llvm/llvm-project · GitHub. Linking them in the PR message or a comment means they show up as mentioned in the other PRs which is a nice side effect.