How does the pragma work? Does it disable all warnings, or only warnings in subsequent declarations? Does the warning state apply when a template class is declared or instantiated?
In Visual Studio 2005 / 2008 (8 & 9) the pragma for warnings takes effect on the next top-level declaratation; it is essentially ineffective inside a function or class definition. The binding to spans of source text appears to be quite robust with respect to compiler passes; I have not observed any leakage. On the last point, it appears that if the template definition has a warning disabled, that warning will not be emitted during instantiation. I could easily be wrong about that though, I can't recall having looked into it explicitly. But if it were not the case I believe I would have seen a counterexample by now. On the other hand, we have seldom disabled warning that are emitted late in the compiler; anything that takes that much analysis to generate is usually something you want to fix.
But what if there is other truly deprecated code? Isn't disabling the macro much more specific than disabling all warnings for all deprecated code?
Yes, much better to use the 3-or-4 macros provided to turn off the security and POSIX deprecations. The deprecation mechanism is available to everybody, not just Microsoft, and there are often very good reasons for using it. Whacking the warning across-the-board throws out the baby with the bathwater. The warning in VS2008 associated with the deprecated command-line option /Wp64 can't be disabled from within the IDE: the flag to disable the warning gets pasted into the command line after the offending flag so the warning has been emitted before the disable command is encountered. There are also some cases where disabling a warning causes the auxiliary informational "pseudo-warnings" associated with it to be orphaned in the output stream. Not the worst compiler I've ever used but it does have its quirks.