Fix a bug related to expanded broadcast IterDomains.

[wujingyue]

My #1295 changed some tests in the wrong way so this bug has been hidden since then.

[wujingyue]

!build

[jacobhinkle]

A bit more explanation would be much appreciated. As I understand it, this applies when the input of a ViewOp has an allocation domain that is a permutation of its rfactor. The input's allocation domain is used to determine the contiguity of the output at its root domain (as opposed to .contiguity() which refers to out's allocation domain). That is then propagated from root to rfactor to check if we require a copy. The change seems reasonable to me but it's not clear to me yet how this fails currently.

[wujingyue]

A bit more explanation would be much appreciated. As I understand it, this applies when the input of a ViewOp has an allocation domain that is a permutation of its rfactor. The input's allocation domain is used to determine the contiguity of the output at its root domain (as opposed to .contiguity() which refers to out's allocation domain). That is then propagated from root to rfactor to check if we require a copy. The change seems reasonable to me but it's not clear to me yet how this fails currently.

Your understanding described above is all correct. The code currently fails because due to #1126 (or a similar issue that your #1174 is trying to fix) nvFuser doesn't propagate the has-expanded-extent flag from in_rfactor to out_root. As a result, the expressions between out_root and out_rfactor see out_root's IterDomains as a t or f instead of an e:

Fuser/csrc/optimization/alias_analysis.cpp

Line 86 in 4697ace

if (!outer_contiguity.has_value() && !outer_id->hasExpandedExtent()) {

.

[wujingyue]

!build

[jjsjann123]

LGTM