| hg-reviewers |
This clarifies the details asked by @martinvonz on D3990.
| Lint Skipped |
| Unit Tests Skipped |
| Path | Packages | |||
|---|---|---|---|---|
| M | mercurial/help/internals/linelog.txt (73 lines) |
| 0.5 Malformed Cases | 0.5 Malformed Cases | ||||
| The following cases are considered malformed in our implementation: | The following cases are considered malformed in our implementation: | ||||
| 1. Interleaved insertions, or interleaved deletions. | 1. Interleaved insertions, or interleaved deletions. | ||||
| It can be rewritten to a non-interleaved tree structure. | It can be rewritten to a non-interleaved tree structure. | ||||
| ^AI/D x ^AI/D x | Take insertions as example, deletions are similar: | ||||
| ^AI/D y -> ^AI/D y | |||||
| ^AE x ^AE y | ^AI x ^AI x | ||||
| ^AE y ^AE x | a a | ||||
| ^AI x + 1 -> ^AI x + 1 | |||||
| b b | |||||
| ^AE x ^AE x + 1 | |||||
| c ^AE x | |||||
| ^AE x + 1 ^AI x + 1 | |||||
| c | |||||
| ^AE x + 1 | |||||
| 2. Nested insertions, where the inner one has a smaller revision number. | 2. Nested insertions, where the inner one has a smaller revision number. | ||||
| Or nested deletions, where the inner one has a larger revision number. | |||||
| It can be rewritten to a non-nested form. | It can be rewritten to a non-nested form. | ||||
| Take insertions as example, deletions are similar: | |||||
| ^AI x + 1 ^AI x + 1 | ^AI x + 1 ^AI x + 1 | ||||
| a a | |||||
| ^AI x -> ^AE x + 1 | ^AI x -> ^AE x + 1 | ||||
| ^AE x ^AI x | b ^AI x | ||||
| ^AE x + 1 ^AE x | ^AE x b | ||||
| c ^AE x | |||||
| ^AE x + 1 ^AI x + 1 | |||||
| c | |||||
| ^AE x + 1 | |||||
| 3. Insertion inside deletion with a smaller revision number. | |||||
| 3. Insertion or deletion inside another deletion, where the outer deletion | Rewrite by duplicating the content inserted: | ||||
| block has a smaller revision number. | |||||
| ^AD x ^AD x | ^AD x ^AD x | ||||
| ^AI/D x + 1 -> ^AE x | a a | ||||
| ^AE x + 1 ^AI/D x + 1 | ^AI x + 1 -> b | ||||
| ^AE x ^AE x | b c | ||||
| ^AE x + 1 ^AE x | |||||
| c ^AI x + 1 | |||||
| ^AE x b | |||||
| ^AE x + 1 | |||||
| Note: If "annotate purely depends on "^AI" information, then the | |||||
| duplication content will lose track of where "b" is originally from. | |||||
| Some of them may be valid in other implementations for special purposes. For | Some of them may be valid in other implementations for special purposes. For | ||||
| example, to "revive" a previously deleted block in a newer revision. | example, to "revive" a previously deleted block in a newer revision. | ||||
| 0.6 Cases Can Be Optimized | 0.6 Cases Can Be Optimized | ||||
| It's always better to get things nested. For example, the left is more | It's always better to get things nested. For example, the left is more | ||||
| efficient than the right while they represent the same content: | efficient than the right while they represent the same content: | ||||
| c) Linelog implicitly maintains the order of all lines it stores. So it | c) Linelog implicitly maintains the order of all lines it stores. So it | ||||
| could dump all the lines from all revisions, with a reasonable order. | could dump all the lines from all revisions, with a reasonable order. | ||||
| While revlog could also dump all line additions, it requires extra | While revlog could also dump all line additions, it requires extra | ||||
| computation to figure out the order putting those lines - that's some | computation to figure out the order putting those lines - that's some | ||||
| kind of "merge". | kind of "merge". | ||||
| "c" makes "hg absorb" easier to implement and makes it possible to do | "c" makes "hg absorb" easier to implement and makes it possible to do | ||||
| "annotate --deleted". | "annotate --deleted". | ||||
| 1.4 Malformed Cases Handling | |||||
| The following "case 1", "case 2", and "case 3" refer to cases mentioned | |||||
| in section 0.5. | |||||
| Using the exposed API (replacelines), case 1 is impossible to generate, | |||||
| although it's possible to generate it by constructing rawdata and load that | |||||
| via linelog.fromdata. | |||||
| Doing annotate(maxrev) before replacelines (aka. a1, a2 passed to | |||||
| replacelines are related to the latest revision) eliminates the possibility | |||||
| of case 3. That makes sense since usually you'd like to make edits on top of | |||||
| the latest revision. Practically, both absorb and fastannotate does this. | |||||
| Doing annotate(maxrev), plus replacelines(rev, ...) where rev >= maxrev | |||||
| eliminates the possibility of case 2. That makes sense since usually the | |||||
| edits belong to "new revisions", not "old revisions". Practically, | |||||
| fastannotate does this. Absorb calls replacelines with rev < maxrev to edit | |||||
| past revisions. So it needs some extra care to not generate case 2. | |||||
| If case 1 occurs, that probably means linelog file corruption (assuming | |||||
| linelog is edited via public APIs) the checkout or annotate result could | |||||
| be less meaningful or even error out, but linelog wouldn't enter an infinite | |||||
| loop. | |||||
| If either case 2 or 3 occurs, linelog works as if the inner "^AI/D" and "^AE" | |||||
| operations on the left side are silently ignored. | |||||