Re: [boost] Re: Re: Thread-Local Storage (TLS) and templates

From: Gabriel Dos Reis (gdr_at_[hidden])
Date: 2003-02-20 07:44:58

David Abrahams <dave_at_[hidden]> writes:

| Gabriel Dos Reis <gdr_at_[hidden]> writes:
| 
| > "Ken Hagan" <K.Hagan_at_[hidden]> writes:
| >
| > | Peter Dimov wrote:
| > | >
| > | > &k does not exist yet at compile-time (in a pointer to int form), when
| > | > templates are instantiated.
| > | 
| > | It doesn't have to. We're instantiating a template, not calling a
| > | function, so if "&k" has the type "pointer to thread-local int" then
| > | the compiler knows that and can instantiate the appropriate code.
| >
| > The issue here is the *value* "&k", not the type of that expression.
| 
| Let me reiterate, just in case somebody missed it: this is a similar
| problem to that of using a pointer/reference-to-data-member as a
| template paramter.  You can think of each thread-local global variable
| as a data member in a big struct, of which there's a single instance
| per thread.  The address of this global variable is just as much a
| constant as the address of a data member.  There's no problem here,
| AFAICT.

I see your analogy. 

However, my point is that

  * a class is closed: that is, by the time you put the closing brace,
    the "offset" of the data member is a compile-time constant.

  * the number of thread local variables is potentially unbounded,
    meaning that, using your analogy, the offset of the corresponding
    data-member is not known by the time the compiler finishes
    processing a given translation unit.

-- Gaby