digitalmars.D - Const, invariant, strings and "constant data need never be copied"
- Stewart Gordon (133/133) Nov 01 2007 In DMD 2.006, the definition of string was changed from const(char)[] to...
- Nathan Reed (29/50) Nov 02 2007 As I've remarked in another thread, it makes absolutely no sense to me
-
Stewart Gordon
(14/50)
Nov 02 2007
"Nathan Reed"
wrote in message - Janice Caron (10/33) Nov 03 2007 I agree with Stewart here. (Actually, reading what's been read, I
- Janice Caron (38/67) Nov 02 2007 That's not true. If /all/ strings are invariant, throughout, then
- Stewart Gordon (31/89) Nov 02 2007 Because your D2 code doesn't manipulate strings?
- Janice Caron (34/50) Nov 03 2007 No, because I declared all my strings as string or wstring, not char[]
-
Stewart Gordon
(17/52)
Nov 03 2007
- Janice Caron (17/23) Nov 03 2007 Well, it is now. :-)
- Walter Bright (13/20) Nov 02 2007 Well, no. Take a look at the source to std.string.replace(). It does not...
- Stewart Gordon (16/38) Nov 03 2007 That's basically what I said.
In DMD 2.006, the definition of string was changed from const(char)[] to invariant(char)[] (and similarly wstring and dstring). This change has no doubt broken a fair amount of D 2.x code. While at it, I've found that the functions in std.string use a mix of char[] and string, but only one uses const(char)[]. It's worth thinking about what's actually best, from both coding and runtime efficiency POVs. Let's first look at string manipulation functions such as those in std.string. These merely look at a passed-in string and return something - they don't keep the string for later. They therefore need only a read-only view of a string - they don't need to know that the string is never going to change. Declaring these with invariant parameters therefore means that it is often necessary to .idup a string just to pass it to one of these functions. Moreover, if a piece of code manipulates strings with a mixture of direct modification and calls to std.string functions, it necessitates quite a bit of copying of strings. Consider, for example, a program that reads in a text file, normalises the line breaks and then ROT13 encodes the result. Under D 1.x, this works: ---------- import std.file, std.string, std.cstream; void main(string[] a) { char[] text = cast(char[]) read(a[1]); text = text.replace("\r\n", "\n").replace("\r", "\n"); foreach (ref char c; text) { if ((c >= 'A' && c <= 'M') || (c >= 'a' && c <= 'm')) { c += 13; } else if ((c >= 'N' && c <= 'Z') || (c >= 'n' && c <= 'z')) { c -= 13; } } dout.writeString(text); } ---------- Note that the text is never copied after it is read in, except by std.string.replace if it actually makes any change. In D 2.x, it's necessary to change one line, to something like text = text.idup.replace("\r\n", "\n").replace("\r", "\n").dup; which consequently adds two copy operations. There are a few caveats to this example: - the .idup is only because std.file.read currently returns a mutable void[] - we could actually cast it to an invariant as nothing else is going to use it - there's no significant reason to normalise the line breaks before, rather than after, ROT13ing it - if all we're going to do is output it, we could output on the fly rather than trying to modify the text in memory but these won't be true in the more general case. There are probably plenty of more involved examples in which there's more difference than this between the 1.x and 2.x code. A fairly recent discussion http://tinyurl.com/2kgpqg touched on the question of whether functions that generate a string and return it should return mutable or immutable references. And the discussion was by no means conclusive. If a public library function returns a mutable array reference, it can mean either: (a) it is giving up ownership of the memory the array occupies (b) it is giving the caller direct access to data it holds for some purpose (std.mmfile is an example of this) In case (a), if there's no risk of there being other references to the same memory (as is the case if the function always allocates it) then it would make sense to give the caller the choice of whether it should be mutable. Indeed, the choice is already there, but in no way that protects against inadvertently trying it on something of case (b). Unfortunately, constness doesn't play well with copy-on-write. Needless to reiterate, std.string's functions want at least a read-only view of the array. But the caller might still want what's returned to be mutable. If the function is going to return the passed-in string, it can only (sensibly) return a read-only view since that's what it received. While the caller could try the D&D trick of casting away the constness, there is a risk of catastrophic failure if some std.string implementation caches strings for reuse. But invariant clearly has some use. It enables the claim that "constant data need never be copied" to work, as long as invariant is used well. So if something receives a string from the caller and wants to save it for later use/retrieval, declaring the parameter as invariant will mean that the callee won't need to copy the data. So there's the benefit that, by making it the caller's responsibility to .idup the data if necessary, it'll save the overhead of unnecessary copying. Similarly, when something later wants to retrieve the data, if it is invariant then there's no need to copy it. A library and an application that uses it can share one copy of the data, and believe that the data will never change. To round things up, the D 2.x const/final/invariant thing is certainly useful, but not perfect. The different storage classes/type modifiers are good for different things. It might be worth a good think about how Phobos uses them (or in some cases doesn't), and what is best practically for the definitions of string/wstring/dstring. (Was there a discussion I missed?) Of course, it would also be worth a good think about whether anything can be added or changed in the language to improve matters. Ideas that come to my mind are: 1. A property .invariant, which just returns the reference as an invariant if it's already invariant, otherwise does the same as .idup. For this to work, the runtime would have to keep a record of whether each piece of allocated memory is invariant, which would interfere with the current ability to cast invariance in or out - but at what cost? 2. Some type modifier such as 'unique' that would indicate that only one reference to the data exists. I'm not sure what rules there should be to enforce this, or if we should just go on trust. But it would be implicitly convertible to mutable, const or invariant, enabling something like ---------- unique(int)[] rep(int i) { unique(int)[] result; result.length = i % 10; result[] = i; return result; } int[] twos = rep(2); const(int) fives = rep(5); invariant(int)[] twelves = rep(12); ---------- 3. Some concept of const-transparent functions. One approach would be to enable a type modifier (or the lack thereof) to be used as a template parameter, with something like ---------- T(char)[] doSomethingWith(typemod T)(T(char)[] param) { ... } char[] str; const(char)[] cstr; invariant(char)[] istr; str = doSomethingWith(str); cstr = doSomethingWith(cstr); istr = doSomethingWith(istr); ---------- This would enable copy on write to work well. As long as nothing _within_ the function so templated relies on the distinction, the compiler could optimise by generating only one instance, since it affects only compile-time type checking and not code generation. Comments? Stewart. -- My e-mail address is valid but not my primary mailbox. Please keep replies on the 'group where everybody may benefit.
Nov 01 2007
Stewart Gordon wrote:Let's first look at string manipulation functions such as those in std.string. These merely look at a passed-in string and return something - they don't keep the string for later. They therefore need only a read-only view of a string - they don't need to know that the string is never going to change. Declaring these with invariant parameters therefore means that it is often necessary to .idup a string just to pass it to one of these functions. Moreover, if a piece of code manipulates strings with a mixture of direct modification and calls to std.string functions, it necessitates quite a bit of copying of strings.As I've remarked in another thread, it makes absolutely no sense to me to use invariant for the library string functions, for exactly this reason. They used to be const, which enables them to work on any kind of string the user might want to call them on (mutable, const, or invariant).1. A property .invariant, which just returns the reference as an invariant if it's already invariant, otherwise does the same as .idup. For this to work, the runtime would have to keep a record of whether each piece of allocated memory is invariant, which would interfere with the current ability to cast invariance in or out - but at what cost?Actually, this wouldn't need to have any runtime consequences, as the invariantness-or-not of a thing is part of its static typing, so could be determined at compile time. Of course, something can be invariant even if it's not typed as invariant (and undecidably so), but do we really need to worry about that? The .invariant property could simply return the array if the array is typed as invariant, and return .idup otherwise.3. Some concept of const-transparent functions. One approach would be to enable a type modifier (or the lack thereof) to be used as a template parameter, with something like ---------- T(char)[] doSomethingWith(typemod T)(T(char)[] param) { ... }A proposal for doing something very much like this is already planned for D 2.0 (it's in the WalterAndrei.pdf from the D conference a couple months back). It's called the 'return' storage class, which would make the return value of a function take on the same constness or invariantness as a parameter: const(char)[] doSomethingWith(return const (char)[] param) { ... } What this does is makes the constness of the return value the same as the constness of the argument passed to 'param', each time the function is called. You can do something similar with templates already, but you can't make template functions virtual. The function is type-checked with the declared types for the parameter and return (in this case, const(char)[]). Thanks, Nathan Reed
Nov 02 2007
"Nathan Reed" <nathaniel.reed gmail.com> wrote in message news:fgg8hs$hon$1 digitalmars.com... <snip>As I've remarked in another thread, it makes absolutely no sense to me to use invariant for the library string functions, for exactly this reason. They used to be const, which enables them to work on any kind of string the user might want to call them on (mutable, const, or invariant).Exactly what I was thinking.I was thinking about the possibility of .invariant being able to detect whether a pointer or array reference typed as const refers to data that was created as invariant or not.1. A property .invariant, which just returns the reference as an invariant if it's already invariant, otherwise does the same as .idup. For this to work, the runtime would have to keep a record of whether each piece of allocated memory is invariant, which would interfere with the current ability to cast invariance in or out - but at what cost?Actually, this wouldn't need to have any runtime consequences, as the invariantness-or-not of a thing is part of its static typing, so could be determined at compile time. Of course, something can be invariant even if it's not typed as invariant (and undecidably so), but do we really need to worry about that? The .invariant property could simply return the array if the array is typed as invariant, and return .idup otherwise.This looks odd to me - you change the code declaring the _parameter_ type in order to effect a variation in the _return_ type? And how would you use the type of parameterised constness within the body of the function? Stewart. -- My e-mail address is valid but not my primary mailbox. Please keep replies on the 'group where everybody may benefit.3. Some concept of const-transparent functions. One approach would be to enable a type modifier (or the lack thereof) to be used as a template parameter, with something like ---------- T(char)[] doSomethingWith(typemod T)(T(char)[] param) { ... }A proposal for doing something very much like this is already planned for D 2.0 (it's in the WalterAndrei.pdf from the D conference a couple months back). It's called the 'return' storage class, which would make the return value of a function take on the same constness or invariantness as a parameter: const(char)[] doSomethingWith(return const (char)[] param) { ... } What this does is makes the constness of the return value the same as the constness of the argument passed to 'param', each time the function is called. You can do something similar with templates already, but you can't make template functions virtual. The function is type-checked with the declared types for the parameter and return (in this case, const(char)[]).
Nov 02 2007
On 11/3/07, Stewart Gordon <smjg_1998 yahoo.com> wrote:I agree with Stewart here. (Actually, reading what's been read, I agree with everybody, except possibly Walter and Andrei). It makes much more sense to me to allow some kind of template-like parameter whose value can be const, invariant or mutable. Also useful would be stuff like is(a : const) is(a : invariant) is(a : mutable) for compile-time decision-making.This looks odd to me - you change the code declaring the _parameter_ type in order to effect a variation in the _return_ type? And how would you use the type of parameterised constness within the body of the function?T(char)[] doSomethingWith(typemod T)(T(char)[] param) { ... }A proposal for doing something very much like this is already planned for D 2.0 (it's in the WalterAndrei.pdf from the D conference a couple months back). It's called the 'return' storage class, which would make the return value of a function take on the same constness or invariantness as a parameter: const(char)[] doSomethingWith(return const (char)[] param) { ... } What this does is makes the constness of the return value the same as the constness of the argument passed to 'param', each time the function is called. You can do something similar with templates already, but you can't make template functions virtual. The function is type-checked with the declared types for the parameter and return (in this case, const(char)[]).
Nov 03 2007
On 11/1/07, Stewart Gordon <smjg_1998 yahoo.com> wrote:In DMD 2.006, the definition of string was changed from const(char)[] to invariant(char)[] (and similarly wstring and dstring). This change has no doubt broken a fair amount of D 2.x code.All of my D2 code compiled without change.Declaring these with invariant parameters therefore means that it is often necessary to .idup a string just to pass it to one of these functions.That's not true. If /all/ strings are invariant, throughout, then everything works.Moreover, if a piece of code manipulates strings with a mixture of direct modification and calls to std.string functions, it necessitates quite a bit of copying of strings.No it doesn't, it merely means ensuring that the reference is unique and then calling assumeUnique().void main(string[] a) { char[] text = cast(char[]) read(a[1]);Well that line's wrong for a start. It should be string text = cast(string)read(a[1]); There's your problem right there.foreach (ref char c; text) { if ((c >= 'A' && c <= 'M') || (c >= 'a' && c <= 'm')) { c += 13; } else if ((c >= 'N' && c <= 'Z') || (c >= 'n' && c <= 'z')) { c -= 13; } }I believe that should be bool willChange = false; foreach(char c;text) if (inPattern["A-Za-z"]) { willChange = true; break } if (willChange) { char[] s = text.dup; foreach (ref char c; s) { if ((c >= 'A' && c <= 'M') || (c >= 'a' && c <= 'm')) { c += 13; } else if ((c >= 'N' && c <= 'Z') || (c >= 'n' && c <= 'z')) { c -= 13; } text = assumeUnique(s); } (Before this release, I would have written text = cast(string)s; That still compiles without complaint, but assumeUnique() is better). The test to see if the string will change is good copy-on-write behavior. The rest is your code, adapted to how you're supposed to do things in D2.006. First you dup text, because that string /might/ be in ROM. Then you make your changes. When you've got what you want, you use assumeUnique() to turn it back into a string. This does /not/ make a copy.In D 2.x, it's necessary to change one line, to something like text = text.idup.replace("\r\n", "\n").replace("\r", "\n").dup;I don't think that's right. You just declare text to be string instead of char[] and those dups become unnecessary.There are a few caveats to this example: - the .idup is only because std.file.read currently returns a mutable void[] - we could actually cast it to an invariant as nothing else is going to use itNot could. Should.There are probably plenty of more involved examples in which there's more difference than this between the 1.x and 2.x code.If every string function you write obeys the copy-(only)-on-write protocol, then I don't see that.3. Some concept of const-transparent functions.I believe that's in the planning stage.
Nov 02 2007
"Janice Caron" <caron800 googlemail.com> wrote in message news:mailman.521.1194042761.16939.digitalmars-d puremagic.com...On 11/1/07, Stewart Gordon <smjg_1998 yahoo.com> wrote:Because your D2 code doesn't manipulate strings?In DMD 2.006, the definition of string was changed from const(char)[] to invariant(char)[] (and similarly wstring and dstring). This change has no doubt broken a fair amount of D 2.x code.All of my D2 code compiled without change.If /all/ strings are invariant, then you're very limited in what manipulations you can perform.Declaring these with invariant parameters therefore means that it is often necessary to .idup a string just to pass it to one of these functions.That's not true. If /all/ strings are invariant, throughout, then everything works.Only in the cases where ensuring that the reference is unique is possible.Moreover, if a piece of code manipulates strings with a mixture of direct modification and calls to std.string functions, it necessitates quite a bit of copying of strings.No it doesn't, it merely means ensuring that the reference is unique and then calling assumeUnique().Firstly, that was D1 code. In D1, string is simply an alias of char[]. Secondly, if it were string, the rest of my code wouldn't work under D2, because there the string type denotes immutable data.void main(string[] a) { char[] text = cast(char[]) read(a[1]);Well that line's wrong for a start. It should be string text = cast(string)read(a[1]); There's your problem right there.There's the problem. You've made the code more complicated to make the final copy conditional on something actually changing. In an ideal world, it would be unnecessary to make that final copy at all (as far as the way my example uses it is concerned). Moreover, your code loops twice, first to see if there's anything to change and then to perform the conversion. This in itself would take a performance hit.foreach (ref char c; text) { if ((c >= 'A' && c <= 'M') || (c >= 'a' && c <= 'm')) { c += 13; } else if ((c >= 'N' && c <= 'Z') || (c >= 'n' && c <= 'z')) { c -= 13; } }I believe that should be bool willChange = false; foreach(char c;text) if (inPattern["A-Za-z"]) { willChange = true; break } if (willChange) { char[] s = text.dup; foreach (ref char c; s) { if ((c >= 'A' && c <= 'M') || (c >= 'a' && c <= 'm')) { c += 13; } else if ((c >= 'N' && c <= 'Z') || (c >= 'n' && c <= 'z')) { c -= 13; } text = assumeUnique(s); }(Before this release, I would have written text = cast(string)s; That still compiles without complaint, but assumeUnique() is better). The test to see if the string will change is good copy-on-write behavior. The rest is your code, adapted to how you're supposed to do things in D2.006. First you dup text, because that string /might/ be in ROM. Then you make your changes. When you've got what you want, you use assumeUnique() to turn it back into a string. This does /not/ make a copy.You miss the point. My example is of ad-hoc code to perform the conversion in place, because it is the most efficient mechanism with the constraints under which the application will ever perform it. Data always loaded into RAM immediately before the conversion, and no desire to keep the 'before' data once the conversion has happened. <snip><snip> Well, I wasn't writing a string function there, so that's beside the point. If you're implementing a complicated string-manipulating algorithm, you're not necessarily going to separate every little step of the algorithm into a separate function. Stewart. -- My e-mail address is valid but not my primary mailbox. Please keep replies on the 'group where everybody may benefit.There are probably plenty of more involved examples in which there's more difference than this between the 1.x and 2.x code.If every string function you write obeys the copy-(only)-on-write protocol, then I don't see that.
Nov 02 2007
On 11/3/07, Stewart Gordon <smjg_1998 yahoo.com> wrote:No, because I declared all my strings as string or wstring, not char[] or wchar[]. And because I use, and assume, D's copy-on-write protocol.All of my D2 code compiled without change.Because your D2 code doesn't manipulate strings?If /all/ strings are invariant, then you're very limited in what manipulations you can perform.That's not true. You can do as much manipulation as you want at creation-time. Only once the manipulation is "finished" do you cast the result to string. (And that behavior is identical between const() and invariant(), by the way, so the change to invariant() makes zero difference to the source code at this point, except that you now have the option of using the assumeUnique() function).Only in the cases where ensuring that the reference is unique is possible.It's always possible. Just write an opening brace, do all your string creation, assign the string to a string variable declared outside the scope, then write a closing brace. Viola - invariance guaranteed, because all the other references used in creation just went out of scope. Note that it is perfectly permissible to have multiple references to an an invariant string anyway - providing that all of those references are themselves declared invariant. It's only non-invariant references which are prohibited, which is why they're the ones you have to lose at the scope boundary.There's the problem. You've made the code more complicated to make the final copy conditional on something actually changing.Of course. That /is/ the copy-on-write protocol. If nothing changes, return the original.Moreover, your code loops twice, first to see if there's anything to change and then to perform the conversion. This in itself would take a performance hit.I could have used the new munch() function instead of the first loop, but I didn't think of it at the time I wrote the example.You miss the point. My example is of ad-hoc code to perform the conversion in place, because it is the most efficient mechanism with the constraints under which the application will ever perform it. Data always loaded into RAM immediately before the conversion, and no desire to keep the 'before' data once the conversion has happened.Well then there's no problem anyway. For the whole time that your string is "under construction", then it's not a string, it's a (mutable) array of chars. Just keep it as such, until you've finished building the string. Then do can do everything in place. But note that if you want to do in-place manipulation of chars, then std.string.replace() is NOT the function to use, because that (possibly) makes a copy. Instead, you would have a replacing loop, or write your own in-place-replace function which operates on char arrays (or templatized for arrays in general)Well, I wasn't writing a string function there.Then you shouldn't be calling std.string functions. What you need are array functions.
Nov 03 2007
"Janice Caron" <caron800 googlemail.com> wrote in message news:mailman.522.1194074475.16939.digitalmars-d puremagic.com...On 11/3/07, Stewart Gordon <smjg_1998 yahoo.com> wrote:<snip>So effectively, you're using the word "string" to refer specifically to the invariant kind, making "if all strings are invariant" a null condition.If /all/ strings are invariant, then you're very limited in what manipulations you can perform.That's not true. You can do as much manipulation as you want at creation-time. Only once the manipulation is "finished" do you cast the result to string. (And that behavior is identical between const() and invariant(), by the way, so the change to invariant() makes zero difference to the source code at this point, except that you now have the option of using the assumeUnique() function).Maybe you're right ... but I'll have to see. <snip>Only in the cases where ensuring that the reference is unique is possible.It's always possible. Just write an opening brace, do all your string creation, assign the string to a string variable declared outside the scope, then write a closing brace. Viola - invariance guaranteed, because all the other references used in creation just went out of scope.So in other words, my code was more or less right in the first place.You miss the point. My example is of ad-hoc code to perform the conversion in place, because it is the most efficient mechanism with the constraints under which the application will ever perform it. Data always loaded into RAM immediately before the conversion, and no desire to keep the 'before' data once the conversion has happened.Well then there's no problem anyway. For the whole time that your string is "under construction", then it's not a string, it's a (mutable) array of chars. Just keep it as such, until you've finished building the string. Then do can do everything in place.But note that if you want to do in-place manipulation of chars, then std.string.replace() is NOT the function to use, because that (possibly) makes a copy. Instead, you would have a replacing loop, or write your own in-place-replace function which operates on char arrays (or templatized for arrays in general)Having the std.string functions is useful even if they create copies. A little bit of copying where it makes coding easier is OK for apps that aren't performance-critical, but it's still nice not to be made to do even more copying or down-and-dirty casting away invariant.I'm not sure what you mean.... Stewart. -- My e-mail address is valid but not my primary mailbox. Please keep replies on the 'group where everybody may benefit.Well, I wasn't writing a string function there.Then you shouldn't be calling std.string functions. What you need are array functions.
Nov 03 2007
On 11/3/07, Stewart Gordon <smjg_1998 yahoo.com> wrote:So effectively, you're using the word "string" to refer specifically to the invariant kind, making "if all strings are invariant" a null condition.Well, it is now. :-) I guess I should have clarified that all my strings were immutable even back when string was const(char)[].So in other words, my code was more or less right in the first place.Basically, yes. /Except/ in your expectations of std.string. The functions in std.string are for fully constructed strings, not for strings-under-construction. I'll get back to that in a minute.I suppose I'm saying we need some extra functions. Maybe even a library std.array. (It's not a big deal as it's relatively easy to write these things yourself). But for functionality like in-place replace, we should be using some specialized function like (if it existed) std.array.replace - but definitely not std.string.replace. Have you ever programmed in PHP? I'd like to see (almost) all the PHP array functions available as standard in D. That way, it would be so much easier to build char arrays in the manner that you suggest, and then turn them into strings when they're fully built. See http://uk2.php.net/manual/en/ref.array.phpThen you shouldn't be calling std.string functions. What you need are array functions.I'm not sure what you mean....
Nov 03 2007
Stewart Gordon wrote:Note that the text is never copied after it is read in, except by std.string.replace if it actually makes any change. In D 2.x, it's necessary to change one line, to something like text = text.idup.replace("\r\n", "\n").replace("\r", "\n").dup; which consequently adds two copy operations.Well, no. Take a look at the source to std.string.replace(). It does not modify the input in place - it returns the input if there are no changes, if there are changes, it returns a *copy*. Second, text should be declared as a string, so you do not need either of the dup's. Two copies are made, just as with the 1.0 version, in that line of code. You will need a third copy to do the loop which modifies the string in place. I feel that, with strings, the advantages of invariant strings outweigh the disadvantages. Note that one can still do the modify-in-place D 1.0 code, and do it very fast, by putting the tests for \r inside the loop rather than as separate loops. The D 1.0 version isn't what you'd write if you wanted speed, anyway.
Nov 02 2007
"Walter Bright" <newshound1 digitalmars.com> wrote in message news:fgguis$1hsk$1 digitalmars.com...Stewart Gordon wrote:That's basically what I said.Note that the text is never copied after it is read in, except by std.string.replace if it actually makes any change. In D 2.x, it's necessary to change one line, to something like text = text.idup.replace("\r\n", "\n").replace("\r", "\n").dup; which consequently adds two copy operations.Well, no. Take a look at the source to std.string.replace(). It does not modify the input in place - it returns the input if there are no changes, if there are changes, it returns a *copy*.Second, text should be declared as a string, so you do not need either of the dup's. Two copies are made, just as with the 1.0 version, in that line of code. You will need a third copy to do the loop which modifies the string in place.In the first of these two paragraphs, you suggest omitting the final .dup, and then in the next, you effectively tell me to put it back in. Therein lies my point - the "third copy" ought not to be necessary for my ad hoc code. (I know I could cast away the invariant, but that's a rather down and dirty trick.)I feel that, with strings, the advantages of invariant strings outweigh the disadvantages.When it comes to string manipulation functions, giving the programmer the choice, with my proposal of const-transparency, would AISI bring even more advantages and alleviate some of the disadvantages.Note that one can still do the modify-in-place D 1.0 code, and do it very fast, by putting the tests for \r inside the loop rather than as separate loops. The D 1.0 version isn't what you'd write if you wanted speed, anyway.Yes, that's another way to do it.... Stewart. -- My e-mail address is valid but not my primary mailbox. Please keep replies on the 'group where everybody may benefit.
Nov 03 2007