Compiler bug: "Unexpected header type"

[jnfoster]

The following program, which is essentially a no-op:

#include <v1model.p4>

struct h {
  bit<8> n;
} 

struct m {
}

parser MyParser(packet_in b,
                out h parsedHdr,
                inout m meta,
                inout standard_metadata_t standard_metadata) {
  state start {
    transition accept;
  }
}

control MyVerifyChecksum(in h hdr,
                       inout m meta,
                       inout standard_metadata_t standard_metadata) {
  apply {}

}
control MyIngress(inout h hdr,
                  inout m meta,
                  inout standard_metadata_t standard_metadata) {
  apply {}
}
control MyEgress(inout h hdr,
               inout m meta,
               inout standard_metadata_t standard_metadata) {
  apply {}
}

control MyComputeChecksum(inout h hdr,
                          inout m meta,
                          inout standard_metadata_t standard_metadata) {
  apply {}
}
control MyDeparser(packet_out b, in h hdr) {
  apply {}
}

V1Switch(MyParser(),
         MyVerifyChecksum(),
         MyIngress(),
         MyEgress(), 
         MyComputeChecksum(),
         MyDeparser()) main;

crashes with a compiler bug error, and a core dump using p4c-bm2-ss

terminate called after throwing an instance of 'Util::CompilerBug'
  what():  COMPILER BUG: ../backends/bmv2/jsonconverter.cpp:1871
<Type_Bits>(2315)bit<8>: Unexpected header type

Aborted (core dumped)

[mihaibudiu]

v1model is not documented properly, but the expectation is that the hdr struct (called h in this program) would only contain headers or stacks. It's equivalent to the P4-14 collection of all possible headers. We should provide a better error message, though.

[jnfoster]

I see. Is the right solution to restrict v1model? Or could one improve the compiler so it supports all types?

[mihaibudiu]

v1model was written to allow P4-14 programs to be translated to P4-16. The converter should never generate programs that crash the compiler. However, today v1model is the only model which can be executed by the behavioral model simulator, for both P4-14 and P4-16 programs, so we also use it for testing the compiler using P4-16 programs.

If we want to program v1model directly we have to document it better and make the behavioral model back-end p4c-bm-ss more robust. The long term plan is to have a new standard architecture model and back-end, which would make v1model obsolete. How much we invest in the v1model and p4c-bm-ss back-end depends on how quickly we expect to have the standard architecture model, a corresponding back-end, and support for these in the behavioral model simulator.

[jnfoster]

Ahh. I misunderstood. Sorry. I thought v1model was the proto-standard architecture. If it's just a shim to allow P4-14 models to import their programs into the P4-16 workchain then these kind of restrictions make total sense.

Thanks for clarifying!

[smolkaj]

The following similar program causes the compiler to crash with Segmentation fault (core dumped). The problem seems to be that I try to extract a struct in my parser, which the specification of extract forbids.

Maybe an error message would be more graceful than a segmentation fault? Then again, if you don't expect users to write programs against this target manually, it probably does not matter.

#include <core.p4>
#include <v1model.p4>

struct H { };
struct M { };

parser ParserI(packet_in b, out H parsedHdr, inout M meta,
              inout standard_metadata_t standard_metadata)
{
    state start {
                // here I am trying to extract a struct -> segmentation fault
        b.extract(parsedHdr);
        transition accept;
    }
}


control VerifyChecksumI(in H hdr,
                        inout M meta,
                        inout standard_metadata_t standard_metadata)
{
    apply { }
}


control IngressI(inout H hdr,
                 inout M meta,
                 inout standard_metadata_t standard_metadata)
{
    apply { }
}


control EgressI(inout H hdr,
                inout M meta,
                inout standard_metadata_t standard_metadata)
{
    apply { }

}


control ComputeChecksumI(inout H hdr,
                         inout M meta,
                         inout standard_metadata_t standard_metadata)
{
    apply { }
}


control DeparserI(packet_out b, in H hdr)
{
    apply { }
}


V1Switch(ParserI(),
         VerifyChecksumI(),
         IngressI(),
         EgressI(),
         ComputeChecksumI(),
         DeparserI()) main;

[mihaibudiu]

I have sent a pull request which fixes these; it includes these examples as tests in p4_16_errors
(Our testing framework does not check the outputs produced by the p4c-bm2-ss compiler - I will file a separate issue with that.)
The compiler should never BUG or segfault.

[smolkaj]

A segmentation fault can likewise be provoked in the deparser by emitting something that is not a header.

[mihaibudiu]

There is a new folder with negative tests for p4c-bm-ss, called testdata/p4_16_bmv_errors. There are three tests for these cases checked-in. (The testing script still does not save reference outputs - so the test would succeed even if the compiler crashes, but I filed issue #96 for this problem.)