Re: emitting templates as 'T'? Changed from 3.2 to 3.3

Re: emitting templates as 'T'? Changed from 3.2 to 3.3

[Benjamin Kosnik]

This is a bit surprising, but it's kind of hard to tell what is going on.

If possible, you'll need to make a small test case that shows clearly
what is going on, instead of pointing people to a large project. Then,
you can take that 10-15 line source and enter it into bugzilla.

best,
benjamin

templated specializations get emitted in 3.3?

[bert hubert]

[-- Attachment #1: Type: text/plain, Size: 1053 bytes --]

On Tue, May 27, 2003 at 02:47:33PM -0500, Benjamin Kosnik wrote:
> 
> This is a bit surprising, but it's kind of hard to tell what is going on.
> 
> If possible, you'll need to make a small test case that shows clearly
> what is going on, instead of pointing people to a large project. Then,
> you can take that 10-15 line source and enter it into bugzilla.

Ok, I've isolated the issue, which I'm unsure is a bug at all. GCC 3.3 emits
specialized template methods as regular text symbols.

2.95 and 3.3 do not emit the specialization at all, as far as I can see.

I've attached a small tar.gz which exhibits this. The only real issue
probably is if and where this change in behaviour is documented, and if
there is a way to get the old semantics back.

If this is a bug, I'll file it in bugzilla. Please CC me on any replies as
I'm not on the list and only browse the archives every so often.

		Bert

-- 
http://www.PowerDNS.com      Open source, database driven DNS Software 
http://lartc.org           Linux Advanced Routing & Traffic Control HOWTO

[-- Attachment #2: gcc-issue.tar.gz --]
[-- Type: application/octet-stream, Size: 545 bytes --]

Re: templated specializations get emitted in 3.3?

[Benjamin Kosnik]

I don't think this is a bug. 

What you are doing is providing a definition of a non-inlined member
specialization in a header file: this is just like putting a definition
of a non-inlined class member function in a header. 

Behavior of this has definitely changed, but the current behavior seem
defensible. Note if you add "inline" you'll have consistent behavior.

-benjamin



template<class X> class Box
{
  X content;

public:
  Box()
  { content=0; }

  int print() { return 0; }
};

template<> 
//int
inline int
Box<long>::print() 
{ return 1; }


// g++ head
// default : T
// inline : (no symbol)

// g++ 3.2.2
// default : (no symbol)
// inline : (no symbol)

// icc 7.1
// default : W
/*
00000000 W Box<long>::print()
00000000 d _ZN3BoxIlE5printEv$$LSDA
*/
// inline : (no symbol)

Re: templated specializations get emitted in 3.3?

[Joe Buck]

On Wed, May 28, 2003 at 03:43:27PM -0500, Benjamin Kosnik wrote:
> I don't think this is a bug. 
> 
> What you are doing is providing a definition of a non-inlined member
> specialization in a header file: this is just like putting a definition
> of a non-inlined class member function in a header. 

If that header is included in two or more compilation units, it is a bug
in the user's program (violation of the one-definition rule).

> Behavior of this has definitely changed, but the current behavior seem
> defensible. Note if you add "inline" you'll have consistent behavior.

That seems like the right fix.

Re: templated specializations get emitted in 3.3?

[bert hubert]

On Wed, May 28, 2003 at 02:12:09PM -0700, Joe Buck wrote:
> On Wed, May 28, 2003 at 03:43:27PM -0500, Benjamin Kosnik wrote:

> > What you are doing is providing a definition of a non-inlined member
> > specialization in a header file: this is just like putting a definition
> > of a non-inlined class member function in a header. 
> 
> If that header is included in two or more compilation units, it is a bug
> in the user's program (violation of the one-definition rule).

Ok, thank you both for resolving this so swiftly. I'll be sending the ISIP
authors a patch moving the specializations to inline. I hope google will be
picking up on this conversation as this is a major change of behaviour for
g++ between 3.2 and higher, which is only sparsely documented - even if it
was a misfeature before.

This behavior makes infinitely more sense.

Regards,

bert

-- 
http://www.PowerDNS.com      Open source, database driven DNS Software 
http://lartc.org           Linux Advanced Routing & Traffic Control HOWTO