Maybe use AT_ASSERT instead ? abort isn't really used as a pattern and AT_ASSERT will give line number information / derive from common exception class.

Adding this type parser gives us a different type parser in type_parser.cpp, operator.cpp, annotations.py, and now irparser.cpp 🤪

I'm not sure if any of them can be unified but worth keeping track of...

Thanks for the suggestion, I'll take a look at those parser and see what can be shared.

I don't think it's possible currently, since we have types internally that are not user-exposed - CompleteTensorType, for example.

Maybe in the future if we expose annotating detailed Tensor Types than it might work. But as of now I'm not sure that much can be shared.

We should not end up with a fourth type parser. I think it would be best to share the schema parser's std::pair<TypePtr, c10::optional<AliasInfo>> parseType(), which can be refactored to be a common component. It is the closest in syntax to what the IR printer prints, and we can add cases for DimensionedTensor and CompleteTensor that are feature-gated by a boolean. For now this code is fine, as long as the followup that merges it happens.

This will fail with something like:

%7 : float = prim::Constant[value=2]()

I would imagine you want to change the way float attributes are printed in ir.cpp to do something similar to what is done here

I think to parse attribute you need the type annotation of the node. Otherwise there's no way to determine the attribute type of empty lists. This will also make it easier to parse floats...

We probably can add optional annotation into IR to resolve these ambiguities, e.g.
%x : float = my::node[value=10 : float, someList=[] : string[]]()

Also, please note that the lists support is pretty limited in the parser at the moment, and that's something I plan to work on incrementally after this initial big diff lands.

Simplest thing would be to use the attributes suffix on each token: 10i, []is

Yea maybe add an assertion that the name is contained, I think it should fail here

I don't think we need the block name, fyi

Maybe parse operator name / op name? that's how it's referred to in the rest of the codebase.

I imagine we always want to have a block here. If you're inside a graph use g->block().

Stmt inputs won't have their type annotated, we just want to parse the value name and access the value in our map. See my comment at parseBlockOutputs.

Return statements won't have the type annotated

Probably don't need no type, can't it just be "Tensor" (based on usage in this file)

We should not end up with a fourth type parser. I think it would be best to share the schema parser's std::pair<TypePtr, c10::optional<AliasInfo>> parseType(), which can be refactored to be a common component. It is the closest in syntax to what the IR printer prints, and we can add cases for DimensionedTensor and CompleteTensor that are feature-gated by a boolean. For now this code is fine, as long as the followup that merges it happens.

All the errors should use throw ErrorReport(token.range()) << "..." that way the error is reported with the right source information.

Symbol::attr(name) does the same thing.

these probably should be int64_t

Simplest thing would be to use the attributes suffix on each token: 10i, []is

nit: This code appears twice, refactor?