Re: Suggestion for improvement to strupr and strlwr functions.

[Jeff Johnston <jjohnstn at redhat dot com>]

J.D McLaughlin wrote:
> Hi,
>
> Examining the source code to C's string functions, I noticed that strupr()
> called toupper() on each char in turn if isupper() returned false(or maybe
> if islower() returned true). However, the toupper() function itself calls
> this function to check beforehand whether or not it needs to do anything,
> and vice versa for the strlwr() function. It looks as if it would be
> better to avoid the duplicated check by just calling toupper for every
> character in the array - although I could be wrong, if the string had less
> than half it's characters in lower case, it would be constantly calling
> toupper and isupper for characters where only isupper was needed.
>
> I'm not sure what the statistical properties are of strings generally
> input to strupr, but I thought I'd share this idea in case in the
> average-case scenario it's more efficient.
>
> James McLaughlin.

James,

If you look at newlib's <ctype.h> which gets included you will see that 
isupper/islower are macros and toupper/tolower are essentially inlined 
functions.  Yes, the islower/isupper invocations end up being redundant 
in the normal GNUC situation.

However, in the case of x86 gcc (and mn10300 gcc), for example, the 
compiler has everything it needs to recognize that the isupper/islower 
check is redundant and simply optimizes out the 2nd check.  Here's a 
snippet from the x86 compiled code of strupr.o:

  1d:   8b bb 00 00 00 00       mov    0x0(%ebx),%edi
                        1f: R_386_GOT32 __ctype_ptr
  23:   90                      nop
    {
      if (islower (*a))
  24:   8b 07                   mov    (%edi),%eax
  26:   0f be d2                movsbl %dl,%edx
  29:   f6 04 02 02             testb  $0x2,(%edx,%eax,1)
  2d:   74 05                   je     34 <strupr+0x34>
        *a = toupper (*a);
  2f:   8d 42 e0                lea    0xffffffe0(%edx),%eax
  32:   88 01                   mov    %al,(%ecx)
  34:   8a 51 01                mov    0x1(%ecx),%dl
  37:   41                      inc    %ecx
  38:   84 d2                   test   %dl,%dl
  3a:   75 e8                   jne    24 <strupr+0x24>
      ++a;
    }

Notice how the testb (islower) is only done the one time in the loop.

In the case of a non-GNUC compiler, the toupper call will be a function 
call which will very likely be more costly than an isupper/islower macro 
check.  In that particular situation, savings/wastage depend on the 
input string as you have noted.

The function itself is non-standard (not ANSI, POSIX, C99, or Single 
Unix) and isn't likely to be that commonly used.

-- Jeff J.