At 01:34 PM 3/15/2006 -0800, Guido van Rossum wrote:

Because Thomas designed it this way. :-)

I believe his design makes sense though: "import foo" translates to __import__(foo, ...).

There's a separate setting, only known to the compiler, that says whether "from __future__ import absolute_import" is in effect (there's no way to slip a flag into globals to convey this setting, since code is compiled independently from globals, and there are ways to pass flags to the compiler without explicitly doing the future import).

So the compiler emits different code when the future syntax is in effect, and that opcode must pass its knowledge to __import__. This is done trough the 5th argument, which also tells us how many leading dots there were. Without the future import, the 5th argument is omitted (as long as there aren't any leading dots).

Ah, so it's *relative* imports that require a 5th argument. I was thinking it was there to support absolute imports. I was thinking that relative imports could be implemented by popping bits off of __name__ to get an absolute location. It seems to me that backward compatibility would be greatly enhanced by having the interpreter convert everything but "legacy" imports into absolute imports, as this would then work with the existing __import__ code in the field, even when new relative/absolute code was doing the importing. Otherwise, this forces __import__ hooks to be rewritten. (I personally avoid writing __import__ hooks if at all possible, but there are certainly some out there.) The mechanism I have in mind would be to just have an IMPORT_EXACT opcode that takes a relative or absolute name. This opcode would process relative names relative to the __name__ in globals to produce an exact module name, and leave absolute names alone. It would then invoke __import__ using the builtins or sys module dictionary as the "globals" argument *instead of the current globals*, so that __import__ will not do any legacy-style fallback. When absolute imports are in effect, or when an explicit relative import is used, it would be compiled such that IMPORT_NAME is replaced by IMPORT_EXACT. This mechanism doesn't require any change to the __import__() signature, and so remains backward compatible with any existing import hook that doesn't do weird things to the globals dictionary of the module that invoked it. On the other hand, perhaps it would be better to fail loudly by breaking on the 5th argument, than to fail silently for really weird __import__ hooks. That is, if it breaks, it will force people to make sure their __import__ code is safe for use with absolute imports. So, the existing approach might well be better than what I had in mind.