On 14.01.2019 13:36, Julian Foad wrote: > Stefan, thanks for taking account of the feedback and updating the doc string > in r1851197. > > I took a look and thought to rewrite the part about encoding and line > splitting like this: > > * Character Encoding and Line Splitting: > * > * It is up to the client to determine the character encoding. The @a line > * content is delivered without any encoding conversion. The line splitting > * is designed to work with ASCII-compatible encodings including UTF-8. Any > * of the byte sequences LF ("\n"), CR ("\n"), CR LF ("\r\n") ends a line > * and is not included in @a line. The @a line content can include all other > * byte values including zero (ASCII NUL). > > I dropped the reference to svn_subst_stream_translated() because it wasn't > much use without saying what parameters it is given, and instead I was able > to say exactly what happens overall. > > Problem 1: Using this blame function on a 16-bit character encoding is still > really ugly: the receiver cannot know which byte sequences were stripped out. > > We should address this issue properly by passing a "line splitter" function > in to svn_client_blame6().
I started on that then decided that svn_client_blame6 is far too narrow scope for that. In order to do this right, we have to introduce a line splitter callback to svn_subst_stream_translated. My idea was to do something like this: * If the line-splitter is NULL, use the current default and check MIME type for text/* * If it's not NULL, ignore MIME type and provide the callback with file props so it can use that to determine encoding if it wants to. * Remove the ignore_mime_type flag and use the presence of a line splitter to convey the same intent; consequently, * Expose the "default" line splitter as a public function. > Problem 2: Then I noticed that where this splitting algorithm is used, it is > in the second pass over the file data, where we associate each "struct blame" > item with a line of text. > > It looks like a different algorithm may be used in the first pass, when > calculating the differences and constructing the blame list. It diffs all > repository revisions of the file using svn_diff_file_diff_2() which splits > lines according to the "ignore_eol_style" option from > svn_client_blame6(diff_options.ignore_eol_style). The effect of > "ignore_eol_style" is not entirely clear to me; it is used in > svn_diff__normalize_buffer(). In addition, if svn_client_blame6() ends up > processing a local WC revision of the file, that is processed through > svn_client__get_normalized_stream(normalize_eols=(svn:eol-style == native)) > before being diffed. > > This is horrible. Surely we should use a consistent "line splitter" > everywhere. Yes, we should introduce the line splitter concept everywhere in the diff/patch/blame group and really anywhere where we're expected to read "lines" from files (svn_stream_readline comes to mind). This is why I stopped working on a splitter for blame ... the problem is much larger than that and "fixing" blame would really not help by itself. > I would expect this means it is possible for blame output to go wrong, with > revision numbers assigned to the wrong lines. This has always been possible due to the imprecise nature of diff, which blame relies on. I wouldn't worry too much about it. > I have been unable to construct such a wrong output, so far, after trying > for half an hour or so. Possibly there is no code path that produces such a > wrong output. But surely this is the sort of issue that leads to problems > down the road. > > Can we re-write this properly? As I said, this is not limited to blame. The current change in the blame API at least makes some sense and is a localised change, that's why I support it. It doesn't break anything that wasn't broken before. Changing the way we handle text-like files for diff, blame and patch, on the other hand, is quite a bit more involved and I'm afraid it'll touch a lot of code. I wouldn't dream of rejecting svn_client_blame6 just because it doesn't solve the larger problem. -- Brane