Eric Lunderberg

I can reproduce the error, and made a simplified test script below. Playing around a bit with the shape, it looks like it only occurs when there's a dimension of...

It looks like the bug follows the steps below: 1. `SuggestIndexMap` first flattens the indices 2. `SplitExprCollector::Collect` unpacks the flattened index, but doesn't pass the `simplify_trivial_iterators = false` argument. As...

Ah, I think I found the better fix. There's some copy/paste between `Inverse` and `NonSurjectiveInverse` that I should have cleaned up when implementing it, and there's better handling in `Inverse`...

> In general, IndexMap::Inverse doesn’t guarantee to generate the inverse map even if the index map is (theoretically) bijective because the underlying DetectIterMap can only handle limited cases. In principle,...

@tkonolige I tested it, and had the same behavior. I was able to get better error messages by changing [this line](https://github.com/Lunderberg/tvm/blob/main/src/meta_schedule/utils.h#L292) from `tir::ScheduleErrorRenderLevel::kNone` to `tir::ScheduleErrorRenderLevel::kDetail`, and it looks like the...

Breakage is in the WASM build, as the EMCC flags were still using `-std=c++14`. Fixed in https://github.com/apache/tvm/pull/12693, will re-run CI after it lands.

Breakage is in the iOS build, as the clang flags were still using `-std=gnu++14`. Fixed in https://github.com/apache/tvm/pull/12712, will re-run CI after it lands.

Sounds reasonable, and I've added a comment weighing in on that thread. If it becomes a sticking point on this PR, I can rewrite to avoid C++17, but C++17 has...

From the discussion, it sounds like C++17 shouldn't be a blocker anymore, since there are no specific use cases at the moment, and multiple options to resolve it if/when they...

Rebased onto main, as https://github.com/apache/tvm/pull/12692 also needed and implemented the same type-checking structs, which makes this PR much smaller.