is there any string doc?

Yes, you're soaking in it!

I have a string, how do I get a pointer to the characters?

You want to avoid this situation. In your own interfaces, prefer string types over raw pointers. Any interface that wants to process a string using a single pointer is making two expensive assumptions. First, that the string is stored in one contiguous hunk; and second, that the string is zero-terminated. If this isn't the case, then to get a pointer, storage must be allocated and the entire string must be copied to it and zero-terminated. You may not be able to avoid needing a pointer when interacting with system calls.

Some string classes guarantee that they are `flat'. That is, that their data is stored in one contiguous zero-terminated hunk. This does not imply that there are no embedded nulls. Caveat emptor. All strings that explicitly promise flatness inherit from the class nsAFlatString or nsAFlatCString and can produce a constant pointer to their data with the get() member function. Even strings that don't explicitly promise to be flat may happen to be flat. The helper function PromiseFlatString will produce a const dependent string that is guaranteed to be flat. If you use this on a string that already happens to be flat, the result is simply a reference through to that string. Otherwise, PromiseFlatString does the work to allocate, copy, terminate, and manage a temporary flat string. Since the result of PromiseFlatString is a temporary, you must be careful not to get and hold a pointer to its data for longer than the temporary itself lives.

  /* I have a string, how do I get a pointer to the characters? */

extern void EvilNarrowOSFunction( const char* );    // evil OS routines that want a pointers
extern void EvilWideOSFunction( const PRUnichar* );

void func( const nsAString& aString, const nsACString& aCString )
  {
    EvilWideOSFunction( NS_LITERAL_STRING("Hello, World!").get() );
      // literal strings are flat already (as are |nsString|s, et al), just use |.get()|

    EvilWideOSFunction( PromiseFlatString(aString).get() );
      // for strings that don't explicitly guarantee flatness, use |PromiseFlatString|


      // beware holding the pointer for longer than the life of the promise
    const PRUnichar* wp = PromiseFlatString(aString).get(); // BAD! |wp| dangles
    EvilWideOSFunction(wp);

      // if you really need to use the pointer from |PromiseFlatString| in more than one expression...
    const nsAFlatString& flat = PromiseFlatString(aString);
    EvilWideOSFunction(flat.get());
    SomeOtherFunction(flat.get());

      // similarly for |char| strings
    EvilNarrowOSFunction( PromiseFlatCString(aCString).get() );
  }

How do I get a particular character out of a string?

Flat strings provide operator[] and CharAt(). All strings provide First(), Last(), and access with iterators. Don't promise a string flat just to do character indexing. Prefer, instead, to get an iterator and advance it to the position you care about.

  /* How do I get a particular character out of a string? */

PRUnichar Get5thCharacterOf( const nsAString& aString )
  {
    if ( aString.Length() >= 5 )
      {
        nsAString::const_iterator iter;
        aString.BeginReading(iter); // make |iter| point to the beginning of |aString|
        iter.advance(5);
        return *iter;
      }

    return PRUnichar(0);
  }

Using iterators isn't as bad as the example above makes it feel. The typical use is for advancing through a string, examining many characters.

How do I convert from one encoding to another?

How do I create a string?

What is the best way to return a string?

There are several reasonable ways to produce a string result from a function. If you are already holding the answer as a sharable string, you can simply return that string (pass-by-value). Otherwise, the most efficient and flexible way to return a string is to assign your result into a non-const reference parameter. Don't bother to create a sharable string from scratch with your generated result.

Why? The two things you want to minimize in string manipulation are, in order of importance, heap allocation, and moving characters around.

  /* What is the best way to return a string? */

class foo
  {
    public:
      // ...
      void GetShortName( nsAString& aResult ) const;
      nsCommonString GetFullName() const;
      
    private:
      nsCommonString    mFullName;

      const PRUnichar*  mShortName;
      PRUint32          mShortNameLength;
      
  };

nsCommonString
foo::GetFullName() const
  {
    return mFullName;
  }

void
foo::GetShortName( nsAString& aResult ) const
  {
    aResult = DependentString(mShortName, mShortNameLength);
  }

How do I printf a string, e.g., for debugging.

If your string is already narrow, you just have to worry about making it flat, and then getting a pointer.

If your string happens to be wide, you'll need to convert it before you can printf something reasonable. If it's just for debugging, you probably wouldn't care if something odd was printed in the case of a UCS2 character that didn't have an ASCII equivalent. The simplest thing in this case is to make a temporary conversion using NS_ConvertUCS2toUTF8. The result is conveniently flat already, so getting the pointer is simple. Remember not to hold onto the pointer you get out of this beyond the lifetime of temporary.

  /* How do I |printf| a string? */


void PrintSomeStrings( const nsAString& aString, const PRUnichar* aKey, const nsACString& aCString )
  {
      // |printf|ing a narrow string is easy
    printf("%s\n", PromiseFlatCString(aCString).get());     // GOOD

      // the simplest way to get a |printf|-able |const char*| out of a string
    printf("%s\n", NS_ConvertUCS2toUTF8(aKey).get());       // GOOD

      // works just as well with an formal wide string type...
    printf("%s\n", NS_ConvertUCS2toUTF8(aString).get());


      // But don't hold onto the pointer longer than the lifetime of the temporary!
    const char* cstring = NS_ConvertUCS2toUTF8(aKey).get(); // BAD! |cstring| is dangling
    printf("%s\n", cstring);
  }

Here are the email answers I have yet to format into the FAQ. Some of the URLs may be out-dated or moved. The messages are in order from oldest to newest.

Date: Thu, 13 Apr 2000 19:41:47 -0400

Encoding Wars

This message is all about strings and the various encodings that might be used to interpret their contents, the ramifications of that, and where we're heading. The point of this message is to say what we're currently thinking, and get feedback. I apologize in advance for the rambling, and for the fact that this message may accidentally mix discussion of how things are and how they will be.

There are many different possible encodings. Three in common use in the Mozilla source base are: ASCII, UCS2, and UTF8. In ASCII, every character fits in 7-bits and is typically stored in an 8-bit byte. We usually represent ASCII strings with nsCStrings, nsXPIDLCStrings, or char string literals. In UCS2, characters occupy 16 bits each. We usually represent UCS2 strings as nsStrings, etc., i.e., two-byte or `wide' strings. UTF8 is a multi-byte encoding. A character might occupy one, two, or three bytes. It is easiest to store and manipulate such a string within a single-byte or `narrow' string implementation.

None of our current string implementations know the encoding of the data they hold at any given moment. An nsCString might legitimately hold data encoded in ASCII, UTF8, or even EBCDIC for that matter.

Operations that convert from one encoding to another, or operations that are encoding sensitive (e.g., to_upper), rightly belong in i18n. The fact that our current string interfaces automatically and implicitly convert between wide and narrow strings is actually the source of many errors in two particular categories: (1) unintended extra work, (2) mistaken re-encoding, e.g., accidentally `converting' a UTF8 string to UCS2 by pretending the UTF8 string is ASCII and then padding with '\0's.

We've known these were bad for a long time, and have been trying to find the right way to fix them. The current thinking is to just byte the bullet and eliminate implicit conversions. That has interesting ramifications.

void foo( const nsString&  aUCS2string );

foo("hello"); // works!  constructs a temporary |nsString| by
              // converting the ASCII literal with padding.
              // Note: this requires an allocation

Though we've always hated this form since it requires a heap allocation. In current code, we recommend

foo( nsAutoString("hello") );

which still copy/converts, but at least it probably doesn't need to do a heap allocation. In the best of all worlds, no conversion, copying, or allocation would be necessary. To do that, you would need to be able to directly specify a UCS2 string, e.g., with the L"hello" notation, and wrap that in an interface that just held a pointer. E.g., something like

void foo( const nsAReadableString&  aUCS2string );

foo( nsLiteralString(L"hello") );

There are problems with this example, however. The L notation specifically makes objects that are arrays of wchar_t, which under GCC is a 4-byte element. This leads to incompatibility with JS, and the annoyance of possibly bloated storage (I'm sort of minimizing the situation here. It's worse that I make it sound). More about tricks to get around this in a bit, but first, let me talk about what to do in the meantime while we're just getting rid of implicit constructors. Initially to get around this problem (what problem? The problem that foo("hello") stopped compiling on my machine when I threw the switch) I made a routine called NS_ConvertToString which looked like this

inline
nsAutoString
NS_ConvertToString( const char* anASCIIstring )
  {
    nsAutoString aUCS2string;
    aUCS2string.AssignWithConversion(anASCIIstring);
    return aUCS2string;
  }

Which lets me write

foo( NS_ConvertToString("hello") );

This was OK, but in discussion there were concerns about performance on machines that didn't inline well, and issues about naming. In that meeting we came up with an alternate naming strategy that we think has room for growth and an implementation more likely to be efficient on every platform. The implementation is to define a new class that derives from nsAutoString, but allows construction from a char*

class NS_ConvertASCIItoUCS2 : public nsAutoString
  {
    public:
      NS_ConvertASCIItoUCS2( const char* );
      // ...
  };

Which gives identical (though renamed) notation for calling foo:

foo( NS_ConvertASCIItoUCS2("hello") );

It looks like a function call to an explicit encoding conversion. It acts like a function call to an explicit encoding conversion. It is a function call to an explicit encoding conversion. We think that this naming pattern has room for growth. In the meeting, we concluded that the best representation for encoding conversions is a family of functions, and NS_ConvertASCIItoUCS2 fits right in. We think that XPCOM probably can't live without the ASCII to UCS2 conversion (though as explicit as possible) but that all others rightly belong in i18n land.

You can probably deduce from the clues in NS_ConvertToString, above, that constructors weren't the only thing that became explicit. Assignment, appending, comparison, et al, got renamed so that when assigning, appending, or comparing to a value in a different encoding the `WithConversion' form must be used. E.g.,

nsString aUCS2string;
nsCString anASCIIstring;
// ...

aUCS2string += anASCIIstring;  // Currently legal, but not for long
aUCS2string.Append(anASCIIstring); // same

aUCS2string.AppendWithConversion(anASCIIstring); // the new way

if ( aUCS2string == anASCIIstring ) // Sorry, this is going away too
  // ...

if ( aUCS2string.EqualsWithConversion(anASCIIstring) )
  // ...

Yes, it's long and annoying. Just like the extra work you were implicitly asking to have done, perhaps incorrectly. There are other reasons to rename these functions. When nsString and nsCString defined a ton of, e.g., Appends each there was no problem, because nobody wanted to override Append. Now, with strings inheriting from abstract base classes we immediately run into the problem that overriding and overloading don't mix very well in C++. Because of a feature of C++ called name hiding, it is problematic to override only a single signature of a name overloaded in a base class. The base nsAWritableString provides several Appends, all for objects of (hopefully) the same encoding. nsString can't easily add a bunch of new Appends (the converting ones) without running face first into the name hiding problem. The discussion of the fix for this is mostly unrelated to encoding issues, so I'll defer it to another post.

In hindsight, after the meeting, it seemed clear that all the `WithConversion' forms would be better named

xxxConvertingASCIItoUCS2
xxxConvertingUCS2toASCII

however, the real goal (probably) is to move most such conversions into i18n. Just bringing attention to the previously implicit conversions is a good first step. Renaming these conversions as just suggested is probably the right thing to do, though it sort of validates them, which I'm not sure we really want. This is a decision we need to discuss further.

Now, back to the string literal problem above. One possible solution is to use a macro. Imagine

NS_LITERAL_STRING("Hello")

which on a machine where the L trick works, turns into

nsLiteralString(L"Hello")

but on a machine where there is trouble, turns into something less appealing, but more likely to work, like

NS_ConvertASCIItoUCS2("Hello")

Another solution is to add a compilation step that fixes L strings on bad platforms to be non-L strings, but padded with \0s. E.g., L"Hello" gets preprocessed into "\000H\000e\000l\000l\000o\000". This solution is more annoying to the developer, where the prior solution is more annoying during the runtime.

Before we go to too much trouble on this specific feature, we will probably want to do more measurement to see just how much and how often we are converting constant literal strings, and why.

I'm currently ripping through the tree fixing things to use the `WithConversion' forms where appropriate. I was also converting things to use NS_ConvertToString where appropriate; unless I get talked out of it, I want to switch midstream to NS_ConvertASCIItoUCS2, then go back and fix up the NS_ConvertToString instances later. I've set things up so I can check in as I go. After all these conversions have been done, I'll be able to throw the switch (what switch? NEW_STRING_APIS) which will make nsString inherit from nsAWritableString, etc. and allow us to start exploiting these other opportunities (e.g., for literal strings, shared strings, etc. See http://bugzilla.mozilla.org/show_bug.cgi?id=28221 for details and reasoning.)

I guess I'm expecting comments on:

• how really annoying this whole topic is
• how bad L"xxx" is
• whether to move forward with NS_ConvertASCIItoUCS2
• whether we should move to xxxConvertingASCIItoUCS2 etc instead of `WithConverting'
• arguments about where encoding conversions should live
• arguments about whether going between 1 and 2 byte storage is an encoding conversion
• questions about stuff I didn't mention or didn't explain well
• pointing out stuff I'm just plain wrong about, or things I forgot
• etc

So as not to jumble the discussion, I'll be separately posting other requests for comments about specific features of the design of the new string hierarchy.

I hope this helps keep everybody filled in on what we're thinking and able to point out what we're forgetting or screwing up :-)

Date: Wed, 19 Apr 2000 21:12:47 -0400
Subject: more string info

news://news.mozilla.org/scc-705460.16423913042000@news.mozilla.org

Date: Fri, 26 May 2000 15:31:37 -0400
Subject: Re: Question on ==

I would prefer you compare with Equals (which should really be named IsEqualTo) rather than operator==() because of this:

char* a;
char* b;

// ...

if ( a == b )
  // ...

Comparing two raw `string' pointers doesn't compare the characters they point to, but instead compares the bits of the pointers. For this reason, I may eventually make comparison of a string with a pointer using operators just go away.

Date: Wed, 14 Jun 2000 14:38:55 -0400
Subject: Re: Fix to XprtDefs.h

Yes, we're aware that turning off wchar_t support makes wchar_t be a synonym for unsigned short under Metrowerks. We know that the current version of VC++ also makes these types equivalent. In theory, though, the types are distinct even when they are the same size and shape. By using real wchar_t support, we are forced to recognize the distinction and navigate it appropriately with reinterpret_cast (via NS_REINTERPRET_CAST). The win here is that we aren't caught by compiler changes that suddenly make some set of compilers compliant and therefore break our code. We will add an autoconf test that lets UNIX compilers opt in to our string scheme when they have an appropriately shaped wchar_t. If these happen to be compliant compilers, all will be well. If they don't, the casts don't hurt, because they are type correct. We are writing our code to meet the standard as we move forward.

The win for us is realized by the following macros

#ifdef HAVE_CPP_2BYTE_WCHAR_T
  #define NS_LITERAL_STRING(s)  nsLiteralString(L##s, \
                      (sizeof(L##s)/sizeof(wchar_t))-1)
#else
  #define NS_LITERAL_STRING(s)  NS_ConvertASCIItoUCS2(s, \
                       sizeof(s)-1)
#endif

An nsLiteralString points directly to the literal characters. No copying, no conversion, and the length calculation happens at compile time. This has turned out to be as large a savings as 15% of code space and 8% of data space, net, in our string test harness It's faster as well, again by eliminating the copying, conversion, and length calculation. We don't know yet what those numbers translate into in our real code base, but we have high hopes.

I don't want to be in the position to ask you to change your code. I don't think it's appropriate for me to do so. The AIM application that is your client is our client as well. They need to resolve this difference between us in whatever way they think best. That may mean asking you if changing your apis is the right thing to do. Or it may mean applying the casts. Our code-base and yours, Justin, are more like cousins. I don't think you should have to change just to conform to us. You may think my arguments for using real wchar_t have merit, and adopt similar usage just because you agree; but I think the only obligation you have is to follow the technical solution you think is right for your code.

If you decide to make this api change, it will mean shipping a new binary (on Mac) for your library to clients who want to switch over to the new api (since the name mangling will be different, and therefore, the link requirements will change).

Hope this helps,

Date: Thu, 15 Jun 2000 19:36:55 -0400
Subject: Re: Checkin approval for bug 32336

S.Equals(NS_LITERAL_STRING("bar"), PR_TRUE, 3)

doesn't compile because there is no three parameter form for Equals. For all definitions of Equals on strings, see "nsAReadableString.h"

http://lxr.mozilla.org/seamonkey/source/xpcom/ds/nsAReadableString.h

There is an EqualsWithConversion that takes three parameters.

http://lxr.mozilla.org/seamonkey/source/xpcom/ds/nsString2.h#731

It is ``EqualsWithConversion'' because it admits the possibility of an encoding specific transformation, in this case to provide case-insensitive comparison. This also wouldn't compile, however, since, at the moment, an nsLiteralString doesn't provide an operator to produce a const PRUnichar* (though perhaps it should), and it doesn't satisfy the other interfaces that match this call, e.g., a const nsString&.

Perhaps I need to move case-insensitive comparison up out of nsString into a global encoding specific transformations and algorithms file (which was on its way anyway as Waterson, knows); this use is one bit of evidence to support this. In the short term, this can be fixed (if we think the current behavior is wrong) by providing operator const CharT*() const on literal string.

If you can live with out case-folding, the earlier form is preferred

S == NS_LITERAL_STRING("bar")

if you can't, then one of the fixes I mentioned is in order.

Date: Thu, 15 Jun 2000 19:47:12 -0400
Subject: Re: [Fwd: how to use nsString ?]

  >I see these same examples time and again in the embedding
  >samples/docs, but I can't compile them.

Apologies. Documentation mentioning strings is getting out of date. Here are some specific answers.

  >nsString URLString("http://www.mozilla.org");

...is now perhaps best expressed as nsString URLString( NS_LITERAL_STRING("http://www.mozilla.org") );

since an nsString is a sequence of 2-byte wide characters, and the routines that implicitly convert 1-byte sequences (like the literal sequence you specified, "http:...") are now gone.

Up until not too long ago, one would have had to say

nsString URLString;
URLString.AssignWithConversion("http://www.mozilla.org");

The NS_LITERAL_STRING construction is new machinery that has the potential to make many operations much more efficient.

  >nsString URLString;
  >URLString.SetString("www.mozilla.org");

SetString was a synonym for Assign or assignment with operator=(), it too went away. The equivalent is the second example I gave above, that is, the one with AssignWithConversion.

Assign still exists. AssignWithConversion takes on that functionality for assignments that require encoding transformations (e.g., from ASCII to UCS2). SetString is gone, since it was always a synonym for Assign.

Learn more about the general APIs for strings that we are trying to move to by examining http://lxr.mozilla.org/seamonkey/source/xpcom/ds/nsAReadableString.h http://lxr.mozilla.org/seamonkey/source/xpcom/ds/nsAWritableString.h

Hope this helps,

Date: Thu, 15 Jun 2000 21:26:51 -0400
Subject: Re: Checkin approval for bug 32336

  >I *need* the count attribute, because I need to compare only the first 
  >chars (that's inherent to the logic).

This is what substrings are for. In that case, you could use

Substring(S, 0, 3) == NS_LITERAL_STRING("bar")

As for case-folding, it's best if you can case-fold everything up front, instead of doing it repeatedly. I'll have to get back to you on a general solution to that problem, or what my schedule for getting it checked in would be. I'm sorry, I know that's not what you needed to hear. If the source string is an nsString, you can continue to exploit its implementation of these routines, e.g., ToLower all up-front.

Hope this helps,

Date: Mon, 19 Jun 2000 14:23:47 -0400
Subject: Re: string fu

  >It seems less convenient to have to first check path.IsEmpty, and
  >then if false get path.Last and test it.

What would you prefer? That extracting a character not in the string always return CharT(0)? Can't do it for two reasons: (1) 0 may be a valid character in a particular encoding, so it can't be used in general as a ``no character at that position'' marker; and (2) I can't control what an individual string implementation does when asked to get an out-of-bounds fragment, it's explicitly undefined. That means the result of CharAt is explicitly undefined for indexes outside the defined contents of the string. As a debugging convenience, I have made this assert, but it has always been the case that retrieving such a character had undefined results ... even in [the old] code.

OK, you might say, well at least let me ask for a character that is only off the end by one. E.g., Last of an empty string. Reason (1) from above still applies. How bad is it to say, for the case you gave

PRBool needsDelim = PR_FALSE;
if ( !path.IsEmpty() )
  {
    PRUnichar last = path.Last();
    needsDelim = !(last == '/' || last == '\\');
  }

In general, you probably want to opt out of a whole lot of work when the source string is empty. It is slightly less convenient, but it doesn't tie us to a bunch of implementation specific mojo.

  >Can we fix GetUnicode in this case?

This is an annoying property of auto strings, e.g., that they always have an allocated buffer. I'm happy to fix this bug, however, be aware that GetUnicode and GetBuffer are artifacts of [the old] implementation that we don't want to support. They are not part of the abstract interface. We will keep them no longer than we have to. They don't support our multi-fragment paradigm. People who require a contiguous hunk of characters in the future, and are unwilling to switch over to chunky-iterators, may be forced to copy the string to their own buffer. There will be an implementation of narrow character string that guarantees contiguous allocation and a zero-terminator, much as nsCString does now, for compatibility with platform uses, but this won't be the default string class.

Date: Mon, 19 Jun 2000 17:22:31 -0400

Clarifying String Sematics

Recently, I added an assert to the string operations that extract characters, namely First(), Last(), CharAt(), and operator[](). This assert fires when any of these routines are used to access a character outside the defined contents of the string. For First() and Last() that means whenever they are applied to an empty string. For CharAt() and operator[](), that means whenever they are used to access an index outside the range of 0..Length()-1. There have been some complaints, however, the result was always undefined. What follows is extracted from an email exchange between me and warren on this topic. I hope it clarifies strings semantics

Warren writes:

  >I hit your funky CharAt assertion tonight in this piece of code:

  >NS_IMETHODIMP
  >nsIOService::ResolveRelativePath(
  >    const char *relativePath,
  >    const char* basePath,
  >    char **result )
  >  {
  >    nsCAutoString name;
  >    nsCAutoString path(basePath);
  >    
  >    PRUnichar last = path.Last();
  >    PRBool needsDelim = !(last == '/' || last == '\\' || last ==
  >    '\0');
  >    ...

  >where basePath is null. It seems less convenient to have to first
  >check path.IsEmpty, and then if false get path.Last and test it.

I replied:

  >What would you prefer?  That extracting a character not in the
  >string always return CharT(0)?  Can't do it for two reasons:
  >(1) 0 may be a valid character in a particular encoding, so it
  >can't be used in general as a ``no character at that position''
  >marker; and (2) I can't control what an individual string
  >implementation does when asked to get an out-of-bounds fragment,
  >it's explicitly undefined.  That means the result of CharAt is
  >explicitly undefined for indexes outside the defined contents of
  >the string.  As a debugging convenience, I have made this assert,
  >but it has always been the case that retrieving such a character
  >had undefined results ... even in [the old] code.

  >OK, you might say, well at least let me ask for a character that
  >is only off the end by one.  E.g., Last of an empty string.
  >Reason (1) from above still applies.  How bad is it to say, for the
  >case you gave

  >  PRBool needsDelim = PR_FALSE;
  >  if ( !path.IsEmpty() )
  >    {
  >      PRUnichar last = path.Last();
  >      needsDelim = !(last == '/' || last == '\\');
  >    }

  >In general, you probably want to opt out of a whole lot of work
  >when the source string is empty.  It is slightly less convenient,
  >but it doesn't tie us to a bunch of implementation specific mojo.

Warren also asks:

  >Here's another issue, perhaps more serious. If I say this:

  >  foo(const PRUnichar* s) {
  >    nsAutoString str(s);
  >    bar(str.get());
  >  }

  >where s is null, bar will get passed a zero-length PRUnichar
  >sequence instead of null. This makes it so that you can't just
  >test for the argument == null. You have to nsCRT::strlen(arg) == 0
  >which is much less efficient. Can we fix GetUnicode in this case?

And I reply:

  >This is an annoying property of auto strings, e.g., that they
  >always have an allocated buffer.  I'm happy to fix this bug,
  >however, be aware that GetUnicode and GetBuffer are artifacts
  >of [the old] implementation that we don't want to support.  They
  >are not part of the abstract interface.  We will keep them no
  >longer than we have to.  They don't support our multi-fragment
  >paradigm.  People who require a contiguous hunk of characters in
  >the future, and are unwilling to switch over to chunky-iterators,
  >may be forced to copy the string to their own buffer.  There will
  >be an implementation of narrow character string that guarantees
  >contiguous allocation and a zero-terminator, much as nsCString
  >does now, for compatibility with platform uses, but this won't be
  >the default string class.

In a later message, Chris Waterson asks a related question

  >scc: should we add operator PRUnichar*() to
  >NS_ConvertASCIItoUCS2?

And I reply:

  >It seems reasonable.  A lot more reasonable that forcing people to
  >call GetUnicode().  I alluded to platform specific classes in an
  >earlier message to warren that you were cc'd on, Chris.  I imagine
  >that the ...Convert... routines would be required to produce
  >contiguous allocation 0-terminated strings (though the as yet
  >unimplemented ...Copy... forms, of course wouldn't.  So operator
  >const PRUnichar*() const makes perfect sense to me here.

Hope this makes sense,

Date: Tue, 20 Jun 2000 04:05:31 -0400
Subject: Re: NS_LITERAL_STRING is broken

The behavior you describe sounds exactly like when you say

const char* foobar = "foobar";

... NS_LITERAL_STRING(foobar).get() ...

because in this case, the thing passed in is a const char*. NS_LITERAL_STRING is not meant to be used in this way. It is only meant to be used around a " delimited string. The type of such is const char[N] where N is the number of characters in the string + 1 for the zero terminator it helpfully adds. sizeof such a type is N.

Are you sure you had the actual string as an argument, as in your example to me? Or could the actual code have been like my sample, above?

Date: Thu, 29 Jun 2000 13:35:10 -0400
Subject: Re: a fix

  > +       if (Length() == 0) { return nsnull; }

Dave,

please read news://news.mozilla.org/scc-314ABF.14261619062000@news.mozilla.org

It's just plain wrong to let people try to index into a string outside its defined contents. I can't just return '\0' or PRUnichar('\0') there as that could be a legal value to have somewhere in your string for some encodings ... and the encoding is not specified. So your patch has the basic problem of defeating my plan to stop people from doing this bad thing.

The second problem with your patch is that you use the symbolic constant nsnull, which is ostensibly a pointer value; Last returns a character. nsnull is not appropriate for that purpose. In fact, C++ gurus pretty much eschew the use of symbolic constants for 0. NULL is to be avoided. nsnull is wrong-headed in that it presumes we could have some other application specific value for NULL. We can't, it would never work. It's just wasted brain-print. Always use 0 for these situations, and if you want to communicate the fact that something is a pointer type, either use a comment or a (construction-style) cast, like so (graded examples from worst to best:)

• F: FindChildByNameWithHint("Chuck", nsnull);
• D: FindChildByNameWithHint("Chuck", NULL);
• C: FindChildByNameWithHint("Chuck", /* Child* */ 0);
• B: typedef Child* Child_ptr; FindChildByNameWithHint("Chuck", Child_ptr(0));
• A: FindChildByNameWithHint("Chuck", 0);

Don't let this discourage you; keep up the good work :-)

Date: Tue, 8 Aug 2000 23:47:16 -0400
Subject: Re: nsWritingIterator?

  >Can you give me any pointers to examples, or docs, or just some
  >general advice?

does this help?

I can personally walk you through any specific scenario you need.

Date: Wed, 9 Aug 2000 02:35:03 -0400
Subject: Re: nsWritingIterator?

You got it right... it's nsWritingIterator for whichever character type you care about, either char or PRUnichar. You _can_ use this iterator like a character pointer ... that is, you can dereference it, assign into its dereference, etc. It is more efficient, though, to directly address a particular range of characters around where it points by asking it for its actual character pointer with get, and knowing that there are size_forward() characters available ahead of that pointer and size_backward() characters available behind it. After examining those characters by hand, you can advance the iterator beyond the characters you have examined (and possibly into the next chunk, should one exist) by adding into it (with +=) the count of the characters you have processed.

Here are three examples of running through a string and modifying some of the characters in it. All use nsWritingIterators.

  // inefficient, but works in a pinch:
  //  iterators can hide all details of chunks by acting like
  //  a raw character pointer

nsWritingIterator<PRUnichar> s = S.BeginWriting();
nsWritingIterator<PRUnichar> done_with_string = S.EndWriting();

  // for each character in the string |S|
while ( s != done_with_string )
  {
      // if the character is lower case, capitalize it
    if ( 'a' <= *s && *s <= 'z' )
      *s = *s -'a' + 'A';
  }




  // efficient
  //  iterators provide a mechanism by which you can process
  //  a chunk-at-a-time

nsWritingIterator<PRUnichar> iter = S.BeginWriting();
nsWritingIterator<PRUnichar> done_with_string = S.EndWriting();

  // for each chunk of the string
while ( iter != done_with_string )
  {
    size_t N = iter.size_forward();  // # of chars in this chunk
    PRUnichar* s = iter.get();
    PRUnichar* done_with_chunk = s + N;

      // for each character in this chunk
    for ( ; s < done_with_chunk; ++s )
      {
         // if the character is lower case, capitalize it
       if ( 'a' <= *s && *s <= 'z' )
          *s = *s - 'a' + 'A';
      } 

      // advance the iterator past characters
      //  we examined (and into the next chunk, if any)
    s += N;
  }



  // elegant
  //  pull your transformation into a `sink', and |copy_string|
  //  will efficiently pump any kind of string into it

struct Capitalize
  {
      // inline
    PRUint32
    write( PRUnichar* s, PRUint32 N )
        // processes one chunk, called repeatedly by |copy_string|
      {
        PRUnichar* done_with_chunk = s + N;

         // for each character in this chunk
        for ( ; s < done_with_chunk; ++s )
          {
              // if the character is lower case, capitalize it
            if ( 'a' <= *s && *s <= 'z' )
              *s = *s - 'a' + 'A';
          }
      }
  };

copy_string(S.BeginWriting(), S.EndWriting(), Capitalize());

Does this show it better?

Date: Thu, 17 Aug 2000 18:23:22 -0400

  >I tried looking at the string header files but they
  >are awfully complicated.

I'll explain things in a little more detail than you need, then so that some of the stuff you see in these headers will make more sense. I'll also answer your questions out of order.

First: the string hierarchy looks like this http://ScottCollins.net/Journal/discussion/string_hierarchy.gif

The two most important headers are: http://lxr.mozilla.org/seamonkey/source/xpcom/ds/nsAReadableString.h http://lxr.mozilla.org/seamonkey/source/xpcom/ds/nsAWritableString.h

These abstract classes, nsAReadable[C]String, and nsAWritable[C]String are typically what you will want to use in the interfaces of new code. If you write a piece of code that takes a string for input, consider, e.g.,

void consumes_a_string( const nsAReadableString&  aInput );

If you write a piece of code that modifies a string, consider

void modifies_a_string( nsAWritableString&  aResult );

When creating your own classes, member strings will typically be nsStrings. When you can't avoid creating a short string that you need only temporarily during a function, you will typically use nsAutoString. When someone passes you a raw pointer, or a raw pointer and a length, representing a buffer of characters that you may examine, but won't own, you can treat it like a string by wrapping it in an nsLiteralString, e.g.,

void
reads_a_buffer( const PRUnichar* aInput, PRUint32 aInputLength )
  {
    nsLiteralString input(aInput, aInputLength);
      // doesn't allocate or copy

    // ...
  }

You will use nsLiteralString around quoted constant strings as well, though typically through the NS_LITERAL_STRING macro, to avoid doing a length calculation

NS_LITERAL_STRING("x")

expands to

nsLiteralString(L"x", (sizeof(L"x")/sizeof(PRUnichar) - 1))

if L notation works as needed on your platform. Those are the basics. Now onto your questions:

  >For example this won't compile. [...]

  >str1 += L"abc " + str2 + L"def";

L"abc " makes a an object that is a const wchar_t[5], and none of the string code knows about wchar_t. The main reason is that wchar_t is not necessarily the right size (it can be 4 bytes under gcc). If you wrap these constant expressions in NS_LITERAL_STRING, as described above, you should get the right thing, e.g.,

str1 += NS_LITERAL_STRING("abc ") + str2 + NS_LITERAL_STRING("def");

  >Another one is:
  >function(const PRUnichar *foo);
  >call function(L"abc " + str2);

  >It won't create a temporary nsString.

This one, I have a quick and easy explanation for. If function was declared like this

function( const nsAReadableString&  )

then, no problem, since a nsPromiseConcatenation (which was the result of adding those two things together) is a readable string. No other objects need to be created; no copying needs to be performed.

In all cases, we want the creation of nsStrings et al, to be explicit, since creation is unbelievably expensive, requiring heap allocation, locks, copying, etc.

I hope this answers both your posts,

Date: Thu, 17 Aug 2000 20:57:08 -0400
Subject: re our conversation

Date: Fri, 18 Aug 2000 02:52:45 -0400
Subject: Re: More questions and new string API

  >1) How do I return a static string?

  >const nsAReadableString&  foo() {return NS_LITERAL_STRING("x");}
  >errors on taking the address of a temporary variable.

Unfortunately, NS_LITERAL_STRINGs definition is not particularly amenable to this use. Instead, you would have to say something like this:

const nsAReadableString&
foo()
  {
#ifdef HAVE_CPP_2BYTE_WCHAR_T
    static nsLiteralString static_foo(L"x", 1);
#else
    static nsLiteralString static_foo;
    static PRBool initialized = PR_FALSE;
    if ( !initialized )
      {
        static_foo.AssignWithConversion("x", 1);
        initialized = PR_TRUE;
      }
#endif
    return static_foo;
  }

  >2) I'm using these with the STL library in an XPCOM component.
  >What type should I use with map?  This doesn't work...

  >typedef map mapStringMyType;
  >mapStringMyType foo;
  >foo.find(nsAReadableString);  - I want to find on a ReadableString

I don't know what errors you are getting; but it probably doesn't work because a reference isn't an assignable type. This is just a guess. You may need to use

map

If you actually want the map to manage ownership of the keys, then you'll want to use a concrete type, e.g.,

map

or perhaps

map

Or maybe there's something else wrong. Send me the error messages. If you end up using a pointer, then of course you'll have to supply a comparison function to the map template. You won't be satisfied with the default comparison of pointers :-) Sorry I couldn't answer this one more completely.

  >3) How do a get a raw PRUnichar pointer out of nsAReadableString
  >when I need to call something that wants 'unsigned short *'?

The problem with this scenario is that an nsAReadableString doesn't promise that all its data is contiguous, nor that it is zero-terminated, which is what I suspect you want in this case. If the function you want to call can take {pointer, length} tuples, and can consume the string in hunks without zero termination ... then you can use copy_string to pump the string into your function, see http://ScottCollins.net/Journal/discussion/string_iterators.html

If not, and you absolutely have to have a contiguous zero-terminated buffer, then there is a new facility (part of the DOMAPI branch) that does what you need. It's not checked in on the trunk; it should be in early next week. It is nsPromiseFlatString. This class promises a contiguous zero-terminated buffer; and has an operator PRUnichar* to produce a pointer to that buffer automatically. If the underlying class is one that happens to be a single fragment and zero-terminated, then, like nsPromiseSubstring and nsPromiseConcatenation, this class merely holds a reference into the original data. If, however, the underlying string is multi-fragment or not zero-terminated, then nsPromiseFlatString allocates a contiguous buffer of appropriate size and copies the fragmented string data to it. So given

void ReadBuffer( PRUnichar* );

You can call this as efficiently as possible with an arbitrary string like so

ReadBuffer( nsPromiseFlatString(aString) );

If the function you are calling needs to take ownership of the buffer you hand it, then you will probably call ToNewUnicode like so

void ConsumeBuffer( PRUnichar* );

ConsumeBuffer( ToNewUnicode(aString) );

The global function ToNewUnicode is declared in "nsReadableUtils.h", and was only recently added to the build. It is currently being used in the DOMAPI branch. It is part of the build, but the file "dlldeps.c" in XPCOM may need to be modified to ensure it is exported on your platform if you are building the tip. Needless to say, you want to avoid functions that require bare pointers for several reasons: (a) they typically assume zero-termination, which is not guaranteed by the normal encodings; (b) they require contiguous allocation, which may not be possible; (c) they scan for the end of the string, at linear cost (if the encoding makes it possible at all), when the length could be known in advance. If you have to do it, the above mechanisms work, but be aware of the cost and the potential need to copy.

  >4) How do I declare a local variable to hold a nsAReadableString?
  >and a member variable?

nsAReadableString is an abstract type. So you can't have a concrete instance of it. All strings in the hierarchy are readable strings. If you just want a reference to a readable string, you can say, e.g.,

struct foo
  {
    const nsAReadableString&  mString;
    // ...

    foo( const nsAReadableString&  aString ) : mString(aString) { }
  };

...similarly with pointers; but I suspect you are looking for something more concrete. An nsString is a nsAReadableString, and is the typical thing you want as a member variable. An nsAutoString is also an nsAReadableString and is typically what you would use for a short (in length) temporary (in lifetime) local variable, as I mentioned in my previous post.

  >5) If I call a function that returns a PRUnichar* and I want t
  >use it as a nsAReadableString should I wrap it in a
  >nsLiteralString?

Yes, though remember, an nsLiteralString assumes the lifetime of the underlying data is under someone else's control. If the called function gives you a buffer that you need to delete, you will have to manage that yourself. Currently, people often use nsXPIDLString to handle that. XPIDL strings are not part of the hierarchy. They are only used as a sort of string-auto_ptr. However, I'm integrating their functionality into nsString. There is no problem in wrapping the same pointer in both as two separate local variables, one to give you the readable interface, and one to manage the lifetime.

If it's OK with you, I'd like to post this reply (including your quoted questions) to n.p.m.xpcom and also put a copy near the string iterator discussion I provided a link to above, so that other people with similar questions can see these answers.

Hope this helps,

Date: Sun, 3 Sep 2000 03:52:17 -0400

In article <8nu9m2$eo14@secnews.netscape.com>, "Jon Smirl" wrote: > I have the new strings up and running in my app. They work as > advertised and > I haven't found any bugs. Thanks for the good job in designing and > implementing them. Here's are a summary of issues I've encountered > so far...

Thanks, and I appreciate your comments and insights. > > 1) Should there be a nsSegmentedString derived from nsString instead > of building segment support into nsString? None of my strings are > segmented but > I keep executing code that is supports it. nsPromiseFlatString would > be trivial in the non-segmented case.

The general case is that a string does not promise to have contiguous data. A specific case is that, for some implementations, it does. You couldn't do it the other way around, because a segmented string couldn't satisfy all the promises of a flat string. However, through the use of chunky iterators, operating on strings that happen to be flat is very efficient. In fact, nsPromiseFlatString is trivial in the non-segmented case. In addition, I'll be adding an abstract flat class into the hierarchy, which will present additional interface ... in your local routines where you actually have declared a concrete string instance that happens to be flat, the compiler will give you the benefit of using the flat specific routines (e.g., a substring object over a flat string is simpler than the general purpose substring). I need to be cautious about this, though, since I don't automatically want people propagating the flat type through their interfaces. That would put us in the same boat we're in right now ... where routines only work on a specific kind of string, which denies other parts of the code the opportunity to use an implementation beneficial to its specific needs, and typically for no good reason. > > 2) Should nsAWritableString have a way to get the buffer and then > return it? > I need to get the buffer to pass it to OS calls. I'm doing this now > by passing around nsStrings instead of the interface. If I just use > the interface I encur an extra copy since I have to use a temporary > buffer.

A specific string implementation could promise this, but in general, a writable could not. After all, a writable doesn't even guarantee contiguous storage. To some degree, this is what nsPromiseFlatString is for. However, this is a readable promise only. It will also be the case that ns[C]Strings, in the very near future will be able to just assume ownership of an arbitrary buffer allocated on the free store with the XPCOM allocators ... getting one to give up its buffer, on the other hand, presents some problems. Do you have a lot of places where the system writes into your string buffer space? Or do you have a lot of system routines that return you new buffers? I can imagine using nsPromiseFlatString for this, but what happens when the OS alters the underlying data? If the promise had generated that flat data on behalf of a multi-fragment string, should it now put the changes back? It's possible to do, I just want to know if it's correct to allow this situation to happen. > > 3) There needs to be a NS_LITERAL_CHAR() to go along with > NS_LITERAL_STRING().

OK. > Having NS_LITERAL_STRING() all over the code clutters > it up and makes it hard to tell what the code is doing, could we > have a standard short alias for this?

Yes, I'll try to think of something ... perhaps NS_LSTR? > 4) nsLiteralString should support n.ToInteger(&error);

ToInteger is actually a bad interface. It's only good if your entire string is the number; this encourages you to edit your string until it is one, or perhaps copy the numeric part to another string. Better if you just sscanf a string (don't know if I can provide that in the general case, but I'm thinking about it), or else use regular C++ extractors (which wouldn't be too hard for me to provide), or else I could give you a ToInteger that works on a pair of iterators, extracting the integer from the digits between them. > > 5) There should be a global define for an interface to a readonly > empty string.

Yes, there will be. > > 6) Something is wrong with concatenation....

Hopefully I've fixed this now. > 8) A forward definition is missing in the h files

I'll check it out.

My understanding is that you have already found the answers to your other questions.

I hope this helps,

Date: Wed, 20 Sep 2000 17:32:13 -0400
Subject: Re: how to free an nsString::ToNewCString

  >What's the current approved way to free an nsString::ToNewCString? 

nsMemory::Free

You use several NS_ConvertASCIItoUCS2("...").get(), these should be NS_LITERAL_STRING("...").get()

Don't do this to the very first case where you aren't wrapping an actual literal string. The first instance would should exploit NS_LITERAL_STRING technology as well, around the initial declarations of the strings ... probably want to do this with NS_NAMED_LITERAL_STRING.

Date: Thu, 12 Oct 2000 00:57:28 -0400
Subject: string answers

nsresult
DoSomething( nsAWritableString&  answer )
  {
    nsresult rv;

    nsXPIDLString registry_data;
    Fetch("key", getter_Shares(registry_data));

    nsLiteralString path(not_my_string);

    PRInt32 first_colon = path.FindChar(PRUnichar(':'));
    if ( first_colon != -1 )
      {
        // convert ... extract path from |path|
        nsCOMPtr localFile( do_CreateInstance(CID, &rv)
);
        if ( localFile )
          {
           
localFile->SetPersistentDescriptor(NS_ConvertUCS2toUTF8(path));

            nsXPIDLString converted_path;
            localFile->GetUnicodePath(getter_Copies(converted_path));
            answer = converted_path.get();
          }
      }
    else
      {
        answer = path;
      }


    return rv;
  }

Date: Thu, 12 Oct 2000 02:03:49 -0400
Subject: Re: and the answer is ...

You can see from the line of code that you're on, that this should have been fine. nsMemory::Alloc would be asked to allocate a 1 byte object. But it failed trying to allocate that. Which suggests that the allocator was busy and non-reentrant and the debugger tried to misuse it. Yes?

Of course, this doesn't solve your problem. Perhaps we need to go back to the idea of a function that returns a pointer to the first hunk of the string.

const char*
debug_string( const nsAReadableCString& aCString )
  {
    nsReadingIterator<char> iter;
    aCString.BeginReading(iter);
    return aCString.IsEmpty() ? "" : iter.get();
  }

This code should work regardless of what the allocator is doing. The downsides are (a) it only returns the first hunk of the string, in the case of a multi-fragment string; and (b) that hunk might not be zero-terminated.

Hope this helps,

Date: Thu, 12 Oct 2000 08:30:32 -0400
Subject: Re: Self healing the cache :-)

At 3:04 PM -0400 10/11/00, Mike Shaver wrote:

  >NS_LITERAL_STRING(NS_XPCOM_SHUTDOWN_OBSERVER_ID);

Macro ugliness makes NS_LITERAL_STRING inappropriate for use over other macros. In other words:

NS_LITERAL_STRING("foo")

is good.

#define FOO "foo"
NS_LITERAL_STRING(FOO)

is bad. Why? Because it turns into

nsLiteralString(LFOO, sizeof(LFOO)...

and there is no LFOO. Sorry. If you have to do this to a macro-ized string, do the magic by hand, e.g.,

nsLiteralString(FOO, sizeof(FOO)/sizeof(PRUnichar)
                                          + sizeof(PRUnichar('\0')))

or else if you don't care that nsLiteralString will scan for the length, just say

nsLiteralString(FOO)

Hope this helps,

Date: Thu, 12 Oct 2000 08:36:14 -0400
Subject: Re: Self healing the cache :-)

Actually, I'm not even sure you can do it by hand, since you didn't

#define FOO L"foo"

and can't do that cross-platform. The other way around this is to define a global instead of a macro, that is, instead of saying

#define FOO "foo"

at the top of your file, say

NS_NAMED_LITERAL_STRING(FOO, "foo")

or else, if the macro was used only in one spot ... perhaps you could just eliminate the macro in favor of NS_NAMED_LITERAL in situ.

Arghh. In this case, you may be stuck with the extra work of AssignWithConversion.

Date: Sun, 3 Dec 2000 16:38:07 -0400
Subject: Re: another copy_string question

  >Is there a way to tell, inside the write() sink, if one is in the
  >final hunk?  I need to do some special processing at the end.

No, there isn't. But you could move such special processing into the destructor of the sink. Remember, the sink is passed by reference, so you can exactly control its lifetime.

{
  MySink sink;
  nsReadingIterator<PRUnichar> sourceStart = aStr.BeginReading();
  nsReadingIterator<PRUnichar> sourceEnd = aStr.EndReading();
  copy_string(sourceStart, sourceEnd, sink);
    // |sink| destructor executed here
}

Hope this helps,

Date: Fri, 15 Dec 2000 20:02:08 -0400
Subject: fragment of code

nsPromiseFlatString flatKey(aReadable);

flatKey.get()

Date: Tue, 16 Jan 2001 16:47:37 -0400
Subject: Re: a few string questions...

This is explicitly allowed. That's why I'm proposing to change the names of those classes to nsLocal[C]String. >2) Should nsString2x.h and nsString2x.cpp go away? They look like a >never-completed rewrite or something...

Yes. They should go away. They are uncompleted [old] bullshit, exactly as you diagnosed.

I'll look into the other two questions.

Date: Thu, 1 Feb 2001 15:12:41 -0400
Subject: Re: [Fwd: bad string, bad string]

We've been removing implicit conversion operators because they _always_ lead to trouble. Usually they make it harder to pick the right function when overloading is involved and in the past they have led to huge performance suckage because we ended up doing conversions when we didn't need to because the implicit operator made us pick the wrong function.

It's borderline when the class implements something that is so close, as with a guaranteed flat string or an nsCOMPtr ... but the general recommendation is to avoid implicit conversions.

See bug #53057.

Date: Tue, 6 Feb 2001 18:52:23 -0400
Subject: seeking review for bug #57087

bug: http://bugzilla.mozilla.org/show_bug.cgi?id=57087 patch: http://bugzilla.mozilla.org/showattachment.cgi?attach_id=24576

This patch is supposed to add the ability to define very long literal strings more easily by breaking lines, e.g.,

NS_MULTILINE_LITERAL( NS_L("This is the start of a very long line")
                      NS_L(" which actually continues across")
                      NS_L(" a couple more.") )

The main danger in this scheme is callers who omit the inner NS_L wrapping. Though I believe this will be caught at compile time as the wrong type initializer.

Seeking input from everybody, and waterson in particular.

Date: Wed, 14 Feb 2001 16:09:10 -0400
Subject: Re: Question...

There are some utilities in "xpcom/ds/nsReadableUtils.h". In particular, if you want to get back a new heap-allocated ASCII string with the minimal work, you would say

PRUnichar* sourceChars = ...;

char* destChars = ToNewCString(nsLiteralString(sourceChars));

It's more efficient if you happen to already know the length. If you don't, don't bother counting, that's what I'll do in the constructor for nsLiteralString. If you do, then call like this

destChars = ToNewCString( nsLiteralString(sourceChars, length) );

Other routines in that file will help you if, for instance, you wanted to translate into a buffer you had already allocated.

Hope this helps,

Date: Fri, 23 Feb 2001 03:12:58 -0400
Subject: string snippet

nsCString aInput;



nsReadingIterator<char> search_start;
aInput.BeginReading(search_start);

nsReadingIterator<char> search_end;
aInput.EndReading(search_end);

if ( FindCharInReadable(':', search_start, search_end) )
  {
    ++search_start;
    return ToNewCString( Substring(aInput, search_start, search_end)
);
  }

Date: Wed, 7 Mar 2001 19:44:08 -0400
Subject: string help

Here you go, Mike: http://scottcollins.net/journal/discussion/mjudge-scratch.cpp

Date: Fri, 9 Mar 2001 20:56:07 -0400
Subject: Re: string assertions

If you get an iterator into a string and you advance it all the way to the end of the string, and then keep trying to advance it, you hit this assert. This could happen, for example if you tried to copy 10 characters out of a 9 character string. I've tried to make this impossible to get to. As far as I know, all my routines trim requests in advance of manipulating iterators. When you see this, you should get the stack. That will take you right to the bad spot.

Date: Sat, 31 Mar 2001 11:04:03 -0400
Subject: Re: Sun bustage and string advice

You do know you are comparing two pointers now? It seems unlikely those two pointers would ever be the same pointer. You probably want to say something like

NS_LITERAL_STRING("foo").Equals(aTopic) // or

NS_LITERAL_STRING("foo") == nsLiteralString(aTopic)

...so that you compare the contents of two strings. Right now, you're just testing to see if two pointers both point to the same location in memory. A lot of people make this mistake. I would like to make it obvious to people that comparing two pointers does not compare strings. Can you tell me what gave you that impression so that I can figure out how to better educate people not to do this? By the way, it's not that I don't want to make this compare two strings; it's that in C++, you can't override operations for built-in types. And pointers are built-in types. So I can't make operator==(const PRUnichar*, const PRUnichar*) do anything different than it already does, which is the same thing it does for any other pointer.