Inlining functions can be optimized to avoid copy-in copy-out when possible

[fruffy]

This is something I have observed in an actual large P4 program.
So I have the following snippet:

void inline_dummy(in bit<16> eth_type, out bit<8> dummy_val) {}
control sub_ctrl(in bit<16> eth_type, out bit<8> dummy_val){
    apply {
    }
}

control ingress(inout Headers h, inout Meta m, inout standard_metadata_t sm) {
    sub_ctrl() sub_c;

    apply {
        sub_c.apply(h.eth_hdr.eth_type, h.h.a);
        inline_dummy(h.eth_hdr.eth_type, h.h.b);

    }
}

After the inlining passes (Inline and InlineFunctions), this turns into:

control ingress(inout Headers h, inout Meta m, inout standard_metadata_t sm) {
    apply {
        {
        }
        {
            bit<8> dummy_val_0;
            h.h.b = dummy_val_0;
        }
    }
}

In this case, the semantics of the out parameter for the function call are preserved, but not for the control apply() statement. Which one is correct? Or are both okay? One could argue that not changing the input value at all is a subset of undefined behaviour. Still, this seems inconsistent.

control_nested_inline.p4.txt

[mihaibudiu]

In either case you get an undefined result.
So this looks correct to me.

[jafingerhut]

It definitely seems like this is a case where the compiler is transforming the sub_c.apply() call into a program that is "less undefined" or "fewer possible output states" than the input program, and if one believed that the compiler transformations should somehow preserve undefined behavior, the sub_c.apply() call should also assign an undefined value to h.h.a.

I do not currently have any strong opinions on whether compiler transformations should maximally preserve undefined behavior, or should be allowed to reduce the possible behaviors.

[mihaibudiu]

Perhaps it's a bug, I will investigate further.

[fruffy]

For what it's worth, I cannot think of a case where this could be harmful. What I found odd was the difference in behaviour between function/action inlining and control inlining.

[mihaibudiu]

The inliner for controls is slightly better optimized: it tries to avoid doing copy-in copy-out when the arguments to not alias or have side-effects. The function inliner always does copy-in copy-out. So I guess this is an optimization opportunity for functions, where we can reuse some of the code for the controls. I will rename this issue to reflect this fact, and mark it as an enhacement. This may help with actions too.