git: Use merge-tree in patchsets

Previously patchsets used a hacky invocation of update-index
to overlay changed files from old tree to new tree. In addition
to being slower than necessary, this has the downside of showing
upstream changes in the diff, even those that were not changed
in the given review upload.

Using git merge-tree for this instead is cleaner, faster, and better.
        We do this by "cherry-picking" the old branch with git merge-tree, but providing -X ours as
        the strategy. This provides a tree that looks like the new base with the old changes applied.
        For cases where the old changes don't apply cleanly, the ours strategy prevents conflicts by
        taking the version in the old tree. Although this can be a bit confusing when looking at the
        file as a whole, it provides a succinct way to view a difference between the old version and
        new version.

Topic: mergetreepatch

Author: jerry-skydio

revup/git.py

@@ -614,19 +614,28 @@ async def get_diff_summary(
 
     async def merge_tree_commit(
         self,
-        branch1: GitCommitHash,
-        branch2: GitCommitHash,
+        tree1: GitCommitHash,
+        tree2: GitCommitHash,
         new_commit_info: CommitHeader,
-        merge_base: Optional[GitCommitHash] = None,
+        merge_base: GitCommitHash,
+        strategy: Optional[str] = None,
     ) -> GitCommitHash:
         """
         Perform a combined git merge-tree and commit-tree, returning a git commit hash. Raises
         GitConflictException if there are conflicts while merging the trees.
         """
-        args = ["merge-tree", "--write-tree", "--messages", "-z"]
-        if merge_base:
-            args.extend(["--merge-base", merge_base])
-        args.extend([branch1, branch2])
+        args = [
+            "merge-tree",
+            "--write-tree",
+            "--messages",
+            "-z",
+            "--merge-base",
+            merge_base,
+            tree1,
+            tree2,
+        ]
+        if strategy is not None:
+            args.extend(["-X", strategy])
 
         ret, stdout = await self.git(*args, raiseonerror=False)
 
@@ -749,58 +758,23 @@ async def make_virtual_diff_target(
     ) -> GitCommitHash:
         """
         Return a commit (optionally on top of parent) that provides a way to get the diff from old
-        head to new head while accounting for the fact that new base might have been rebased since
-        old base. This new commit makes an effort to include only files that were actually changed,
-        while excluding files that were changed upstream as part of the rebase.
-
-        We do this by resetting any files that changed in the old_base->old_head diff to their
-        old_head versions in new_base. The returned tree will thus have the following properties
-        when diffed against new_head.
-
-        For files not touched by old or new, ret->new_head won't show any diff. This is primarily
-        what allows us to exclude upstream files.
-        For files touched by both old and new, ret->new_head will show the entire old_head->new_head
-        diff. This will include upstream changes for these files, which are difficult to untangle.
-        For files not touched in old but touched by new (regardless of whether it existed in
-        old_base), diff will show new_base->new_head.
-        For files touched in old but not touched in new, there are 2 cases. If file exists in
-        new_base, diff will show old_head->new_base. If file doesn't exist in new_base, diff will
-        show old_head->(deleted) which isn't perfect since technically new_base->new_head did not
-        delete the file, but probably the least confusing of the alternatives of showing no diff and
-        showing the old_head->old_base diff.
-        """
-        new_index: List[str] = []
-
-        # Transform diff-tree raw output to ls-files style output, taking only the new version
-        for m in RE_RAW_DIFF_TREE_LINE.finditer(
-            await self.git_stdout("diff-tree", "-r", "--no-commit-id", "--raw", old_base, old_head)
-        ):
-            new_index.append(f"{m.group('new_mode')} {m.group('new_hash')} 0\t{m.group('path')}")
-
-        if not new_index:
-            # No files were actually changed, so no diff needs to be applied to new_base
-            return new_base
-
-        temp_index_path = self.get_scratch_dir() + "/index.temp"
-        git_env = {
-            "GIT_INDEX_FILE": temp_index_path,
-        }
-        shutil.copy(f"{self.git_dir}/index", temp_index_path)
-        await self.git("reset", "-q", "--no-refresh", new_base, "--", ":/", env=git_env)
-        await self.git(
-            "update-index",
-            "--index-info",
-            input_str="\n".join(new_index),
-            env=git_env,
-        )
+        head to new head while accounting for the fact that the base might have been rebased
 
-        tree = GitTreeHash(await self.git_stdout("write-tree", env=git_env))
-        new_commit_info = CommitHeader(tree, [parent] if parent else [])
+        We do this by "cherry-picking" the old branch with git merge-tree, but providing -X ours as
+        the strategy. This provides a tree that looks like the new base with the old changes applied.
+        For cases where the old changes don't apply cleanly, the ours strategy prevents conflicts by
+        taking the version in the old tree. Although this can be a bit confusing when looking at the
+        file as a whole, it provides a succinct way to view a difference between the old version and
+        new version.
+        """
+        new_commit_info = CommitHeader(GitTreeHash(""), [parent] if parent else [])
         new_commit_info.commit_msg = (
-            f"revup virtual diff target\n\n{old_base}\n{old_head}\n{new_base}\n{new_head}"
+            f"revup virtual diff target\n\n{old_base}\n{old_head}\n{new_base}\n{new_head}\n\n"
+            "revup uses this branch internally to generate github viewable diffs. The commits and "
+            "diffs between commits are not meaningful or useful."
         )
 
-        return await self.commit_tree(new_commit_info)
+        return await self.merge_tree_commit(old_head, new_base, new_commit_info, old_base, "ours")
 
     async def soft_reset(self, new_commit: GitCommitHash, env: Dict) -> None:
         await self.git("reset", "--soft", new_commit, env=env)