[PATCH] revsets: makes follow() supports file patterns

[Yuya Nishihara]

On Fri, 21 Aug 2015 14:10:52 +0200, liscju wrote:
> # HG changeset patch
> # User liscju <piotr.listkiewicz at gmail.com>
> # Date 1440083972 -7200
> #      Thu Aug 20 17:19:32 2015 +0200
> # Branch bug4757#followRevset
> # Node ID 500e3760612a03cfe35a686dba835ee727534fc3
> # Parent  d9d3d49c4cf77049d12920980e91bf8e4a4ecda2
> revsets: makes follow() supports file patterns
> 
> Before this patch, follow only supports full, exact filenames.
> This patch makes follow argument to be treated like file
> pattern same way like log treats their arguments.
> 
> One problem with implementing this was the fact that
> commands.log function before invoking follow
> change path of its argument to use current working directory.
>
> When log is invoked by command:
>   $ hg --cwd=dir log -f b
> it invokes follow with x arg equals 'dir/b'.
> 
> On the other hand when log is invoked by:
>   $ hg --cwd=dir log -r "follow(b)"
> it invokes follow with x arg equals 'b'

It sounds like a bug for me. Do we have to keep compatibility with the old
implementation that required a canonical path, not a file path?

> To properly handle this situation _follow function
> check if pattern is declared without kind(so pattern
> resolves to exact file matching) and adds if needed
> cwd to pat making matchmod.match matching against
> project root. Otherwise matchmod.match is simply
> invoked with cwd equals repo working directory
> 
> Fix issue4757

We generally put a issue number as follows:
"revsets: makes follow() supports file patterns (issue4757)"

https://mercurial.selenic.com/wiki/ContributingChanges#Submission_checklist

> diff -r d9d3d49c4cf7 -r 500e3760612a mercurial/revset.py
> --- a/mercurial/revset.py	Tue Aug 18 18:38:56 2015 -0500
> +++ b/mercurial/revset.py	Thu Aug 20 17:19:32 2015 +0200
> @@ -1037,34 +1037,41 @@
>      return limit(repo, subset, x)
>  
>  def _follow(repo, subset, x, name, followfirst=False):
> -    l = getargs(x, 0, 1, _("%s takes no arguments or a filename") % name)
> -    c = repo['.']
> +    l = getargs(x, 0, 1, _("%s takes no arguments or a filepattern") % name)

It looks like the other functions call it a "pattern", not "filepattern".

> +    ctx = repo['.']
>      if l:
> -        x = getstring(l[0], _("%s expected a filename") % name)
> -        if x in c:
> -            cx = c[x]
> -            s = set(ctx.rev() for ctx in cx.ancestors(followfirst=followfirst))
> -            # include the revision responsible for the most recent version
> -            s.add(cx.introrev())
> +        pat = getstring(l[0], _("%s expected a filepattern") % name)
> +        wtx = repo[None]

A nit, common name for repo[None] is "wctx".

> +        if not matchmod.patkind(pat):
> +            if not pat.startswith(pathutil.normasprefix(repo.getcwd())):
> +                pat = pathutil.canonpath(repo.root, repo.getcwd(), pat)
> +            matcher = matchmod.match(repo.root, '', [pat], ctx=wtx)

I'm not sure what it does, but if it tries to keep the existing behavior,
it doesn't work as expected.

For example, we have two Makefiles:

  Makefile
  src/Makefile

Before this patch,

  $ cd src
  $ hg log -r 'follow("Makefile")'

did show the history of the top-level Makefile, which seems wrong though.
But now,

  $ cd src
  $ hg log -r 'follow("Makefile")'
  $ hg log -r 'follow("src/Makefile")'

both of these commands show the history of src/Makefile.