There's a problem I've encountered with Emacs for many years.  I never reported it because I've been running Emacs 21.3 under Windows, and I figured that Emacs users on Windows are probably a very small percentage of Emacs users, and that 21.3 is so old that it nobody would be interested in debugging the problem.

But then I encountered the same problem with Emacs 23.2.1 running under Linux.  And a few days ago I finally installed Windows Emacs 23.4.1, and it's got the same problem.

The problem: the replace commands, M-x replace-string and M-x replace-regexp, sometimes work and sometimes don't.  When it doesn't work, it often will work if I retype exactly the same command a few times.

My reaction when I first encountered the problem was that I must have mistyped the command the first time.  But I've encountered it for so many years that whenever it fails to work the first time, it's become habit for me to be extremely careful in my typing the second and subsequent times, and it  often fails on those tries too, but eventually succeeds.

I particularly notice it when I'm defining a macro [ delimited by C-x ( and C-x ) ].  And frequently I have the buffer narrowed to a small subset of text that I want to operate on.  But I don't know for certain that defining a macro or having the buffer narrowed are what cause the problem to manifest.

I now have a concrete example of this that proves that it's not due to my mistyping.  There's a point in the macro where the buffer has been narrowed to a portion  that contains a symbol in CamelCase.
Note: In case you're unfamiliar with CamelCase, it's a convention for variable names originally popularized by the X Window System.  Earlier conventions for C and C++ used "_" as a word delimiter within variable names.  Lisp used "-" instead of "_".  CamelCase, so-called because the capital letters in the middle of the word form humps like those on a camel's back, uses capital letters to indicate the beginning of a new word.  So, the C-style variable name find_char_in_string, or Lisp-style variable name find-char-in-string, in CamelCase is findCharInString.
The purpose of this part of the macro is to turn CamelCase into space-separated words.
M-<                    ;; Go to beginning of narrowed buffer
M-x replace-regexp RET
[A-Z] RET              ;; Find any capital letter
C-q SPC \& RET         ;; Replace it with a space followed by itself
M-<                    ;; Go to beginning of narrowed buffer
C-d                    ;; Delete the unwanted space before the first letter
So, if the narrowed portion of the buffer contains:
"JohnJacobJingleheimerschmidt"
after running this portion of the macro, it should contain:
"John Jacob Jingleheimerschmidt"
Instead, when run in Emacs 23, the result is:
"ohnJacobJingleheimerschmidt"
which is exactly what you'd expect if the M-x replace-regexp failed to do the replacement that it should have.  But since I know that sometimes a replace command works the second time after failing to work the first time, I modified that portion of the macro to do the replace twice:
M-<                    ;; Go to beginning of narrowed buffer
M-x replace-regexp RET
[A-Z] RET              ;; Find any capital letter
C-q SPC \& RET         ;; Replace it with a space followed by itself
M-<                    ;; Go to beginning of narrowed buffer
M-x replace-regexp RET
[A-Z] RET              ;; Find any capital letter
C-q SPC \& RET         ;; Replace it with a space followed by itself
M-<                    ;; Go to beginning of narrowed buffer
C-d                    ;; Delete the unwanted space before the first letter
Now, if the replace were working the first time, applying it again would produce the undesired result:
" John  Jacob  Jingleheimerschmidt"
Instead, it produces:
"John Jacob Jingleheimerschmidt"
Does anybody here have any idea what's going wrong here?
Mark Rosenthal
mbr@arlsoft.com
P.S. - One further clue: In the older version of Emacs (21.3) I've noticed that at those times when the replace fails to work, if I repeat the replace command with C-x ESC ESC, the minibuffer shows:
(replace-regexp "[A-Z]" " \\&" nil sss eee)
where sss and eee are integers that are supposed to indicate the beginning and end characters of the region to operate on, but when the replace has failed, sss and eee specify a small subset of the region.