Manifest constants: why enum instead of invariant?

[Don]

Bruno Medeiros wrote:
> Don wrote:
>> Bill Baxter wrote:
>>> Bruno Medeiros wrote:
>>>> Picking up and idea which had crossed my mind, and that Oskar Linde 
>>>> briefly mentioned on the const debacle thread 
>>>> (news://news.digitalmars.com:119/fsirc9$2hll$1@digitalmars.com):
>>>>
>>>> Why do we need enum instead of invariant to declare manifest constants?
>>>>
>>>> (I have the feeling that this has been discussed before in the 
>>>> manifest constants debate but I can't remember it nor find it)
>>>>
>>>> An invariant variable offers all the power than enum does for 
>>>> manifest constants. The only difference of enum's manifest constants 
>>>> is that they are not an lvalue (their address cannot be taken). But 
>>>> that's a *restriction*, it's not a useful property.
>>>>
>>>> The only possible advantage I see is that of optimization: manifest 
>>>> constants do not take up space. But that *hardly* seems significant: 
>>>> manifest constants are usually numeric, and thus occupy about 2-4 
>>>> bytes each. Even if they are many, in total they are not going to 
>>>> occupy that much space. A string literal can easily occupy as much 
>>>> space as several manifest constants, and they are likely going to be 
>>>> many string literals abound.
>>>>
>>>
>>> The space *is* a significant problem apparently.  There have been 
>>> many complaints about the size added to an exe simply by linking in 
>>> early versions of the Windows headers because of all the constants 
>>> they define.
>>>
> 
> Around how many bytes are we talking about here, just to get a concrete 
> idea, does anyone know? I'm not familiar with the windows headers nor 
> their D bindings, although I recently checked some of the windows d 
> headers and there were many manifest constants. But I still don't know 
> if their size would amount to something that significant. If for 
> example, you have a cmd line program that prints one page of it's 
> syntax/usage, just that single string would occupy the same space as 400 
> manifest constants or more.


void main() {}    --> 79kB exe.
-------
import win32.core;
void main() {}   --> 108kB exe.

So it looks like it's less than 32K. Less than 40% of the size of the 
phobos runtime and gc. And a good linker would remove it ALL.

> 
>>> But I suppose you could also look at that and say "we need to fix the 
>>> linker & optimizer" instead of "we need to fix the language".
>>
>> Indeed. Manifest constants are required primarily as a workaround for 
>> the linker; the linker should be able to discard them. In practice, 
>> most of the time, invariant will probably be used anyway. I think much 
>> of the furore about using enum for constants fails to recognise how 
>> minor the feature is.
>>
>> IMHO, the unfortunate thing about using enum for manifest constants is 
>> that it destroys any chance of better enums.
> 
> Yes, if that's just the case, then it's a very minor feature indeed, 
> both because it arises in an uncommon situation (using windows headers), 
> it could be fixed by the compiler/linker.

BTW, I was the one who requested the feature in the first place.
It was important for template metaprogramming. When you do something 
complicated, the obj files contain masses of compile-time constants. (I 
was experiencing 5Mb obj files where only a single int was used at 
runtime!) It takes ages to save such a large obj file.
Since you cannot take the address of such constants, there's no reason 
for them to be retained, and I requested that they be discarded from the 
obj file.

Subsequently, Walter said that the fact that you can't take their 
address is a compiler bug, not an intentional feature. So we needed some 
additional language support in D2.0 to duplicate the functionality of 
the bug in D1.0 (!)

Now that we have CTFE, it's a _lot_ less critical for metaprogramming. A 
local variable in a CTFE function is obviously not saved in the obj file.