There was some concern with PyImport_GetMagicTag() extracting its value from sys.implementation.cache_tag.  One solution is to have the two use a common underlying value.  However, that's basically what was already in import.c and I'd rather minimize that file at this point.

An alternative is to have the underlying value defined in a more amenable spot, but I'm not convinced that buys a whole lot over simply exposing the underlying value in PyImport_GetMagicTag().  If we could deprecate that function...

Another possibility is to move the underlying value to configure.  The down-side is that it's platform-specific.  I've attached a patch that does this.  While it may do too much as-is, it demonstrates my point.

Can the #defines appear in pyconfig.h instead?  I find it easier to discover them this way, and will also simplify the implementation on Windows.

> Can the #defines appear in pyconfig.h instead?  I find it easier to
> discover them this way, and will also simplify the implementation on
> Windows.

Agreed.  I'd like to pull them into Python/sysmodule.h, though.  I'm also going to drop the part that messes with SOABI.

Sorry for the pedantry, but: I read the title of this bug as wanting these two values to be the exact same pointer.  If you're talking about putting a constant string in a header file, you'll get an (identical) copy of that string in every .o that references it (unless your linker collapses identical strings, which I don't know if they do).

From the description, it sounds like you want them to merely be guaranteed identical strings, but they don't have to be the same exact pointer.

tl;dr: do you mean "=="?  the title implies you mean "is".

tl;dr: "==", not "is"

Shouldn't tl;dr go first, else it seems a little pointless since I already read the whole thing. At which point the tl;dr is really just a summary of what I just read, not a "didn't read" line. Just being pedantic. =)

Here's a new patch that keeps things simpler, but still cleans up the concerns about using sys.implementation.cache_tag in PyImport_GetMagicTag().

Patch updated to include a doc addition.

New changeset ee01fd98b5b0 by Brett Cannon in branch 'default':
Issue #15242: Have PyImport_GetMagicTag() return a const char *
http://hg.python.org/cpython/rev/ee01fd98b5b0

I went ahead and committed Eric's patch. Amaury, if you want to move the macros to a header file I see no reason not to, but I also don't see a need so I didn't want to spend the time doing it myself.

It's not necessary.  The fix looks good as is.