Improve generation of unique strings

[davidbolvansky]

No description provided.

[mihaibudiu]

This change will cause hundreds of files generated by the compiler to change, cause the corresponding tests to change.
You need to submit these files as well.

[davidbolvansky]

This change will cause hundreds of files generated by the compiler to change, cause the corresponding tests to change. You need to submit these files as well.

Is there any way how to autoregenerate outputs for failed tests?

[vhavel]

@davidbolvansky re-run failed tests (or whole test-suite) with P4TEST_REPLACE=True

We have just seen real P4 program which takes hours to compile. The source program is large, and is additionally inflated by some midend transformations used in our backend. The typeChecking on huge IRs then becomes realy slow.

If we don't want to change all the identifiers, we could also implement the unique string generation in typeChecking manually. Maintaining inUse set and counter in different functions seems to prevent the unique string generation to be implemented efficiently and cleanly.

[davidbolvansky]

Ok, to reduce test churn, I implemented other approach based on the usage of references to counters, as suggested by Vojta.

For every name, we track how many times it was used as a "base" for newly generated strings.

[mihaibudiu]

Indeed, the compiler has not really been designed to be fast. I am not sure how much we can fix this now.
If this is ready for review please ask for one.
You need to document the new map structure, though.

[davidbolvansky]

Yeah, I will document it soon.

[davidbolvansky]

Indeed, the compiler has not really been designed to be fast. I am not sure how much we can fix this now.

Understood, but there are low hanging fruits (like this one) which can improve compile times a lot. We have a program where better unique string generation improves compile time significatly (2-3 hours -> 5 min).