Typedef annotations is not generated correctly

[komaljai]

@tc_type("mac") typedef bit<48> macaddr_t;

header ethernet_t {
    macaddr_t dstAddr;
    macaddr_t srcAddr;
    bit<16> etherType;
}

This should translate to

header ethernet_t {
    @tc_type("mac") bit<48> dstAddr;
    @tc_type("mac") bit<48> srcAddr;
    bit<16> etherType;
}

Compiler should copy annotations for 'typedef' (which is not currently supported)

[fruffy]

This is probably the eliminate typedefs or eliminate new type pass. Both of them are simple and this can be fixed. Are you able to submit a PR for this?

[jafingerhut]

Are you also expecting that such annotations on typedef declarations are copied over to all declarations that use that typedef name, e.g.

• local variables declared within a control
• local variables declared within a parser
• parameters of parsers, controls, actions, functions

If you are only expecting them to copy over to fields in header type definitions, why not the others?

If you are expecting them to copy over to all places that a typedef name is used, are you expecting that if the compiler renames a local variable in a control in some pass, that these annotations are preserved?

I think there are also cases where a compiler pass can create new variables with types that are copied from existing variables. If so, would you expect that the annotations be copied from the existing variable to the new variable in such cases?

[komaljai]

I believe the annotations should be copied everywhere wherever "typedef" is replaced, otherwise we are losing information.

[jafingerhut]

Makes sense, and that does seem like a good goal to me.

I mainly wanted to warn you that besides the one pass that replaces typedef with the underlying types, there might be other compiler passes that then introduce new synthesized variables whose types are copies of those variables, and it might be important to ensure that such annotations are copied to those synthetic variables, too. Sorry, but I do not have a complete list of passes where such introduction of synthetic variables with a copy of the type of an existing variable is done.

[jafingerhut]

@komaljai or anyone else looking at this issue: Is your goal to generate P4Info files where table key fields, action parameters, and/or other kinds of things represented in a P4Info file, have annotations on them that are "inherited" from annotations on the typedef declaration in the user's P4 source program?

If so, note that p4c's P4Info generation is after the last front-end pass, and before the first mid-end pass, and the pass that removes typedef from the IR is in the mid-end pass.

If you want annotations on typedef to affect the contents of the P4Info file, you must either:

(a) modify the P4Info generation code to look up the typedef type of key fields, action parameters, etc.
(b) add a new front-end pass that propagates the annotations from typedef to expressions with that type.

or perhaps both (a) and (b) might be required.

For (a), note that this issue and some of its code snippets and discussion in comments might be useful in finding a solution: #2310