Pass to get rid of action_run() calls

[ChrisDodd]

This is a general midend pass for any target that does not support action_run directly. It rewrites the actions in a table that uses action_run to set a newly introduced metadata tmp enum that records which action was run, and changes the switch statement to use that.

It is kind-of logically the reverse of EliminateSwitch, which converts switch statements that don't use action_run() into ones that do, so it would never make sense to use both passes for a backend.

Need to update p4test so we can select which pass(es) to run in a test by some annotation in the test program to better allow testing this.

[ChrisDodd]

I've added the ability to use @command_line annotations to add command-line args in test cases, so this case actually be tested

[fruffy]

This option will only be valid for the p4test back end but this is also a v1model program for BMv2.

[fruffy]

P4Testgen fails because it tries to execute this program but does not support this option.

[asl]

I am not quite sure such pragma is a good approach as this essentially adds a direct and unctrollable way to change internal compiler state. Even more, the option might be not supported at all, spelled differently, might not be used at all, etc. Why cannot some pragma with semantics be used instead?

[ChrisDodd]

Since the filename does not end in -bmv2.p4 it is not a valid bmv2 program (like many other tests in this directory) so should not be used by testgen (and in fact is not). So P4Testgen should not be looking at this file (and as far as I can tell, it does not -- all the testgen failures are with other source files).

[fruffy]

P4Testgen checks programs which include v1model.p4: https://github.com/p4lang/p4c/blob/main/backends/p4tools/modules/testgen/targets/bmv2/test/P4Tests.cmake#L9

Which, imho, is the right way to do because any program that includes this should run on BMv2 or others.

[fruffy]

The other failures is caused by this line

p4c/backends/p4tools/modules/testgen/targets/tofino/test/p4-programs/tna_simple_switch.p4

Line 33 in c67d1fd

@command_line("--disable-parse-max-depth-limit")

This is also not a supported option for P4Testgen but the pragma is still parsed as input.

[fruffy]

I see three possible ways to solve this:

1. Since options may depend on the compiler executing the program, maybe extend this pragma with a "target"/"arch" field to ensure the option is interpreted correctly?

2. Add another option to enable/disable this pragma and adding options.

3. Or just ignore any unknown option made available by the pragma. Maybe print a warning that the option is not supported by the tool interpreting the program.

[ChrisDodd]

The other failures is caused by this line

p4c/backends/p4tools/modules/testgen/targets/tofino/test/p4-programs/tna_simple_switch.p4

Line 33 in c67d1fd

@command_line("--disable-parse-max-depth-limit")

This is also not a supported option for P4Testgen but the pragma is still parsed as input.

This line was not touched by this change, so there's no change here

[fruffy]

This line was not touched by this change, so there's no change here

It's existing code that had no impact but is now active because of https://github.com/p4lang/p4c/pull/5053/files#diff-b033ed467d2767be363300d09ff13ac0a990d2e0de7464b732d9b29888da81a1R54.