www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - scaling of templates with alias argument

reply Frank Benoit <keinfarbton googlemail.com> writes:
void func( alias T )(){
    Stdout.formatln( "member {} {} {}",
        T.offsetof,
        T.stringof,
        typeof(T).stringof );
}

class MyClass{
    uint data1;
    long data2;
    void reg(){
        func!(data1);
        func!(data2);
    }
}

With this mechanism, a template function can access the type, value and
external-scope name of the argument. But each call or func makes a new
instance of this template function.
Using such template function massively (like little text preprocessor
makros in C), will generate big binaries. Is that true?
Is there a better way?
May 14 2007
parent reply Stephen Waits <steve waits.net> writes:
Frank Benoit wrote:

 external-scope name of the argument. But each call or func makes a new
 instance of this template function.


The compiler should only instance each template once per type per 
compilation unit.  The linker should be able to remove duplicates across 
multiple compilation units.

--Steve
May 14 2007
next sibling parent "Jarrett Billingsley" <kb3ctd2 yahoo.com> writes:
"Stephen Waits" <steve waits.net> wrote in message 
news:f2al0r$1rjp$1 digitalmars.com...

 Frank Benoit wrote:

 external-scope name of the argument. But each call or func makes a new
 instance of this template function.


 The compiler should only instance each template once per type per 
 compilation unit.  The linker should be able to remove duplicates across 
 multiple compilation units.

 --Steve


Right.  It'd end up not being any more bloated than defining several 
overloads of the function for various types.
May 14 2007
prev sibling parent reply Frank Benoit <keinfarbton googlemail.com> writes:
Stephen Waits schrieb:

 The compiler should only instance each template once per type per
 compilation unit.  The linker should be able to remove duplicates across
 multiple compilation units.


Ok, 'should' is said twice :)

If i have this program:

void regMember( alias T )(){
   // use T.offsetof
   // use T.stringof
}

class MyClass {
    bool     data1;
    bool     data2;
    bool     data3;
    bool     data4;
    public void describe(){
        regMember!(data1);
        regMember!(data2);
        regMember!(data3);
        regMember!(data4);
    }

}
void main(){
}

compiled with:
dmd -c -g refl.d

and then doing this:
obj2asm refl.o | gvim -

This show 4 different functions containing "regMember" in their name,
and four different function calls, matching exactly those 4 different
functions.

The function name differ in the "data1" "data2" .. identifier. This is
not a surprise, because the template argument is not simply a "T", it is
"alias T", which enables me to get the external-scope names.

So i am not sure, if a different type instance in the template argument,
does not also enforce a complete new template instance.

How can i check this?
May 14 2007
parent reply Stephen Waits <steve waits.net> writes:
Frank Benoit wrote:

 
 Ok, 'should' is said twice :)


Intentionally so, because I don't know if Walter's compiler and linker 
behave this way.

Note that the compiler needn't do anything here - this may be completely 
handled by the linker.  In fact that's probably the smarter way to do it 
(DRY, etc.).

Can you look at a map file?  Your answer lies therein.

--Steve
May 14 2007
parent reply Frank Benoit <keinfarbton googlemail.com> writes:
Stephen Waits schrieb:

 Frank Benoit wrote:

 Ok, 'should' is said twice :)


 
 Intentionally so, because I don't know if Walter's compiler and linker
 behave this way.
 
 Note that the compiler needn't do anything here - this may be completely
 handled by the linker.  In fact that's probably the smarter way to do it
 (DRY, etc.).
 
 Can you look at a map file?  Your answer lies therein.
 
 --Steve


doing a objdump yields 4 functions with 4 different addresses.

An other approach can be this:

regMemberImpl( T )( inout T t, uint offset, char[] name ){
}
char[] rm( char[] name ){
    return "regMemberImpl( "~name~", "~name~".offsetof, \""~name~"\" );";
}

class MyClass{
    bool data1;
    void func(){
        mixin( rm( "data1" ));
    }
}

This avoids the template alias argument and is the first time, i see an
advantage in using this text mixin :)
May 14 2007
parent reply Stephen Waits <steve waits.net> writes:
Frank Benoit wrote:

 Can you look at a map file?  Your answer lies therein.


 
 doing a objdump yields 4 functions with 4 different addresses.


Not sure what objdump does, but you need to see what's in the actual 
application/exe, not a single compilation unit.  We want to know what 
the linker outputs, not the compiler.  Maybe that's what you're looking 
at?  I'm not sure..

A linker map file will tell you.

That said, I wouldn't be surprised if the linker isn't stripping though 
- it's not a trivial task.

--Steve
May 14 2007
next sibling parent Stephen Waits <steve waits.net> writes:
Stephen Waits wrote:

 That said, I wouldn't be surprised if the linker isn't stripping though 
 - it's not a trivial task.


s/stripping/stripping duplicate template instances/

--Steve
May 14 2007
prev sibling parent reply Frank Benoit <keinfarbton googlemail.com> writes:
Stephen Waits schrieb:

 Frank Benoit wrote:

 Can you look at a map file?  Your answer lies therein.


 doing a objdump yields 4 functions with 4 different addresses.


 
 Not sure what objdump does, but you need to see what's in the actual
 application/exe, not a single compilation unit.  We want to know what
 the linker outputs, not the compiler.  Maybe that's what you're looking
 at?  I'm not sure..
 
 A linker map file will tell you.
 
 That said, I wouldn't be surprised if the linker isn't stripping though
 - it's not a trivial task.
 
 --Steve


i did use the objdump on the binary. So i /think/ this tells me, it did
not strip those duplicates.
But i really think they are not duplicates, they only would be, if the
alias argument is always exactly the same variable.
May 14 2007
parent "Jarrett Billingsley" <kb3ctd2 yahoo.com> writes:
"Frank Benoit" <keinfarbton googlemail.com> wrote in message 
news:f2ata2$2a7t$1 digitalmars.com...

 i did use the objdump on the binary. So i /think/ this tells me, it did
 not strip those duplicates.
 But i really think they are not duplicates, they only would be, if the
 alias argument is always exactly the same variable.


You're right; I was wrong in my last reply, I didn't see that it was an 
alias parameter.  A unique template instance will be created for each symbol 
you use as the alias parameter, as you've said.
May 14 2007