On 10/8/2012 4:01 AM, Eli Zaretskii wrote:
> This hunk removes w32inevt from SOME_MACHINE_OBJECTS.  Why was that
> needed?

Merge problem. Fixed.

> To avoid confusion, I suggest to change the name of the variable here:
> 
>   +  /* No, not utf-16-le: that one has a BOM.  */
>   +  DEFSYM (Qutf_16_le, "utf-16le");
> 
> to Qutf_16le.

Good catch.

> Isn't the below supposed to be done automagically by make-docfile,
> which puts these in globals.h?
> 
>   +EXFUN (Fcygwin_convert_path_to_windows, 2);
>   +EXFUN (Fcygwin_convert_path_from_windows, 2);

The earliest version of this code targeted Emacs 23, which doesn't have a
globals.h. Fixed.

> I don't think this (from w32select.h) is a good idea:
> 
>   +#define HAVE_W32SELECT 1
> 
> We only have definitions for HAVE_* macros in config.h and
> conf_post.h.  Having them in other headers is confusing, IMO.  Can we
> move this to conf_post.h instead?

We only used HAVE_W32SELECT in two places, so I just replaced HAVE_W32SELECT
with the conditional we use to include the header --- WINDOWSNT || HAVE_NTGUI.

> You moved the definitions of these from w32heap.c to w32fns.c:
> 
>   +/* The major and minor versions of NT.  */
>   +int w32_major_version;
>   +int w32_minor_version;
>   +int w32_build_number;
> 
> but they are still referenced by w32.c via get_* macros defined on
> w32heap.h.  Can we clean this up, please?

The bigger issue is that w32heap.c is WINDOWSNT-only, but w32heap.h declares
useful facilities that apply more generally (and that have nothing to do with
the heap). I created a new w32common.h for these common declarations and left
the WINDOWSNT-specific heap stuff in w32heap.h.

> Regarding this:
> 
>   +#ifdef NTGUI_UNICODE
>   +  wchar_t filename_buf[MAX_PATH + 1];
> 
> I believe Unicode names on Windows are not limited to MAX_PATH
> characters.

You can use special syntax to exceed that limit ---
http://msdn.microsoft.com/en-us/library/windows/desktop/aa365247(v=vs.85).aspx
--- but nobody does in practice.  Bear in mind that this code only deals with
the file open dialog box, which (the last time I checked) didn't support long
file names in the first place.

Still, I've increased the limit to 32k + 1, which is a hard NT kernel limit, in
the NTGUI_UNICODE case.

> It looks like there are some unnecessary hops here:
> 
>   +    CHECK_STRING (prompt);
>   +    CHECK_STRING (dir);
>   +
>   +    dir = Fexpand_file_name (dir, Qnil);
>   +
>   +    if (STRINGP (filename))
>   +      filename = Ffile_name_nondirectory (filename);
>   +    else
>   +      filename = empty_unibyte_string;
>   +
>   +#ifdef CYGWIN
>   +    dir = Fcygwin_convert_path_to_windows (dir, Qt);
>   +    if (SCHARS (filename) > 0)
>   +      filename = Fcygwin_convert_path_to_windows (filename, Qnil);
>   +#endif
> 
>   +    CHECK_STRING (dir);
>   +    CHECK_STRING (filename);
>   +
>   +    /* The code in file_dialog_callback that attempts to set the text
>   +       of the file name edit window when handling the CDN_INITDONE
>   +       WM_NOTIFY message does not work.  Setting filename to "Current
>   +       Directory" in the only_dir_p case here does work however.  */
>   +    if (SCHARS (filename) == 0 && ! NILP (only_dir_p))
>   +      filename = build_string ("Current Directory");
>   +
>   +    /* Convert the values we've computed so far to system form.  */
>   +#ifdef NTGUI_UNICODE
>   +    to_unicode (prompt, &prompt);
>   +    to_unicode (dir, &dir);
>   +    to_unicode (filename, &filename);
> 
> AFAIU, inside Fcygwin_convert_path_to_windows, you convert the names
> of the file and directory to UTF-16, then decode it back into a
> multibyte string, and finally encode it again to UTF-16 when you call
> to_unicode.  Perhaps the back and forth conversion to and from UTF-16
> can be avoided?

This code isn't on a hot path.  I'm reluctant to optimize it: I'm not sure the
additional complexity would be worth it.

> In any case, a call to CHECK_STRING after we already checked the
> arguments seems both redundant and user-unfriendly: these tests are
> designed to reject errors in arguments passed to primitives, so
> signaling an error because something went wrong in processing valid
> arguments doesn't sound like a good idea.
> 
> This error will not report the offending file name:
> 
>   +    if (SBYTES (filename) + 1 > sizeof (filename_buf))
>   +      error ("filename too long");
> 
> I think report_file_error is better here.

Fixed.

> Why do you need the fprintf part here:
>   #ifdef EMACSDEBUG
>   void
>   _DebPrint (const char *fmt, ...)
>   {
>     char buf[1024];
>     va_list args;
> 
>     va_start (args, fmt);
>     vsprintf (buf, fmt, args);
>     va_end (args);
>   #if CYGWIN
>     fprintf (stderr, "%s", buf);
>   #endif
>     OutputDebugString (buf);
>   }

Because tracing on stderr is useful.  Strings sent to OutputDebugString are
invisible unless Emacs is being run under a debugger.

> What happened to number 21 in the change below?

Bad merge. Fixed.

>    #define WM_EMACS_HIDE_CARET            (WM_EMACS_START + 18)
>    #define WM_EMACS_SETCURSOR             (WM_EMACS_START + 19)
>    #define WM_EMACS_PAINT                 (WM_EMACS_START + 20)
>   -#define WM_EMACS_BRINGTOTOP            (WM_EMACS_START + 21)
>   -#define WM_EMACS_END                   (WM_EMACS_START + 22)
>   +#define WM_EMACS_BRINGTOTOP            (WM_EMACS_START + 22)
>   +#define WM_EMACS_INPUT_READY           (WM_EMACS_START + 23)
>   +#define WM_EMACS_END                   (WM_EMACS_START + 24)
> 
> Why did you need the call to notify_msg_ready below?
> 
>   @@ -285,7 +309,7 @@ prepend_msg (W32Msg *lpmsg)
>      nQueue++;
>      lpNew->lpNext = lpHead;
>      lpHead = lpNew;
>   -
>   +  notify_msg_ready ();

The call to notify_msg_ready is needed because after a call to prepend_msg, a
message is now ready.  It should have been there all along.