I think I've found the problem, here is the patch of w32fns.c (against 
revision 1.256). I don't know the detail, but this patch works.

--------------------------------------8<--------------------------------------
--- w32fns.c 2005-08-08 09:45:47.000000000 +0800
+++ w32fns-fix.c 2005-09-06 15:32:01.275812264 +0800
@@ -4545,7 +4545,7 @@
/* Fill out details in lf according to the font that was
actually loaded. */
lf.lfHeight = font->tm.tmInternalLeading - font->tm.tmHeight;
- lf.lfWidth = font->tm.tmMaxCharWidth;
+ lf.lfWidth = font->tm.tmAveCharWidth;
lf.lfWeight = font->tm.tmWeight;
lf.lfItalic = font->tm.tmItalic;
lf.lfCharSet = font->tm.tmCharSet;
--------------------------------------8<--------------------------------------


2005/8/1, Sun Yijiang <sunyijiang@gmail.com>:
> 
> Sorry, I've reviewed my .emacs and found where the problem is. The
> problem is NOT mule-gbk, it's "create-fontset-from-fontset-spec".
> Following setting results in double-width Chinese characters:
> 
> (create-fontset-from-fontset-spec
> "-*-bitstream vera sans mono-normal-r-*-*-14-*-*-*-c-*-fontset-gbk,
> chinese-gb2312:-*-simsun-normal-r-normal-*-16-*-*-p-*-gb2312*-*,
> chinese-cns11643-5:-*-simsun-normal-r-normal-*-16-*-*-p-*-gbk*-*,
> chinese-cns11643-6:-*-simsun-normal-r-normal-*-16-*-*-p-*-gbk*-*,
> chinese-cns11643-7:-*-simsun-normal-r-normal-*-16-*-*-p-*-gbk*-*" t)
> 
> I think the problem is that I set different fonts for each char set. If
> I comment out this setting and still use mule-gbk, Chinese characters
> looks OK. If I use this setting, the Chinese characters looks
> double-width, either with or without mule-gbk. This setting works for
> previous CVS Emacs builds and seems to become "bad" recently.
> 
> I remember this problem first appeared with a check in of w32bdf.c
> early this year. The problem can be resolved by rolling back w32bdf.c
> to revision 1.20 at that time. But now even this rolling back does not
> work, maybe other changes affects this problem.
> 
> 
> 2005/8/1, Jason Rumney <jasonr@gnu.org>:
> > 
> > Sun Yijiang <sunyijiang@gmail.com> writes:
> > 
> > > I think this change is probably not fully tested. When I use the
> > > mule-gbk package and set font for chinese characters, they look so 
> > > wide, but when I use "emacs -q" they look nice, make no difference
> > > with previous CVS Emacs. I roll back the src/w32bdf.c to older
> > > version, and everything is OK. I think the change of w32bdf.c should
> > > be reviewed, since this is really a bug.
> > 
> > You seem to be saying that you only see the "bug" when you load
> > mule-gbk, which is not part of Emacs. What does mule-gbk do? Does it
> > perhaps try to work around the previous bug by doubling the width of 
> > Chinese characters?
> > 
> > 
>