[MLIR] Attribute caching logic during parsing causes issues for custom assembly formats

[knalbant]

(#151267) Fix duplicated attribute nodes in MLIR bytecode deserialization has caused issues with parsing of typed attributes in one of our internal dialects at Modular.

We have IR with TypedAttrs that looks like:

example.param.load : !example.data<0> = <#example.value>
example.param.load : !example.data<1> = <#example.value>

And this logic (LINK) to check if an attribute is a TypedAttr based on whether the value is followed by a colon is proving to be problematic.

[llvmbot]

@llvm/issue-subscribers-mlir

Author: Kenan (knalbant)

[(#151267) Fix duplicated attribute nodes in MLIR bytecode deserialization](https://github.com//pull/151267) has caused issues with parsing of typed attributes in one of our internal dialects at Modular.

We have IR with TypedAttrs that looks like:

example.param.load : !example.data<0> = <#example.value>
example.param.load : !example.data<1> = <#example.value>

And this logic (LINK) to check if an attribute is a TypedAttr based on whether the value is followed by a colon is proving to be problematic.

[knalbant]

@hankluo6

[weiweichen]

@joker-eph, wondering what's your take on this issue?