Re: [BUG?] iconv used as textconv, and spurious ^M on added lines on Windows

["Jakub Narębski" <jnareb@gmail.com>]

W dniu 02.04.2017 o 09:45, Jeff King pisze:
> On Sat, Apr 01, 2017 at 08:31:27PM +0200, Jakub Narębski wrote:
> 
>> W dniu 01.04.2017 o 08:08, Jeff King pisze:
>>> On Fri, Mar 31, 2017 at 03:24:48PM +0200, Jakub Narębski wrote:
>>>
>>>>> I suspect in the normal case that git is doing line-ending conversion,
>>>>> but it's suppressed when textconv is in use.
>>>>
>>>> I would not consider this a bug if not for the fact that there is no ^M
>>>> without using iconv as textconv.
>>>
>>> I don't think it's a bug, though. You have told Git that you will
>>> convert the contents (whatever their format) into the canonical format,
>>> but your program to do so includes a CR.
>>
>> Well, I have not declared file binary with "binary = true" in diff driver
>> definition, isn't it?
> 
> I don't think binary has anything to do with it. A textconv filter takes
> input (binary or not) and delivers a normalized representation to feed
> to the diff algorithm. There's no further post-processing, and it's the
> responsibility of the filter to deliver the bytes it wants diffed.
> 
> Like I said, I could see an argument for treating the filter output as
> text to be pre-processed, but that's not how it works (and I don't think
> it is a good idea to change it now, unless by adding an option to the
> diff filter).

I think that actually there is something wrong.

If textconv really gets normalized representation of pre-image (the index
version) and post-image (the filesystem version), as it should I think,
both pre-image lines ('-') and post-image lines ('+') should use CRLF,
so there should be no warning, i.e. ^M

Or textconv filter gets normalized representation (it looks this way
when examining diff result saved to file with `git diff test.tex >test.diff`;
I were unable to use `tr '\r' 'Q', either I got "fatal: bad config line"
from Git, or "tr: extra operand" from tr), and somehow Git mistakes
what is happening and writes those ^M.

If I understand it correctly, if pre-image, post-image and context
all use the same eol, there should be no warning, isn't it?

> 
>> P.S. What do you think about Git supporting 'encoding' attribute (or
>> 'core.encoding' config) plus 'core.outputEncoding' in-core?
> 
> Supporting an "encoding" attribute to normalize file encodings in diffs
> seems reasonable to me. But it would have to be enabled only for
> human-readable diffs, as the result could not be applied (so the same as
> textconv).

I was thinking about human readable diffs, and 'git show <blob>', same
as with textconv.

> 
> I don't think core.outputEncoding is necessarily a good idea. We are not
> really equipped anything that isn't an ascii superset, as we intermingle
> the bytes with ascii diff headers (though I think that is true of the
> commitEncoding stuff; I assume everything breaks horribly if you tried
> to set that to UTF-16, but I've never tried it).

Well, the understanding would be that the same limitation as for 
core.logOutputEncoding (documented if it isn't) that only encodings that
are ASCII compatibile are supported.
 
-- 
Jakub Narębski