StrengthReduction ignores side-effects in function calls

[fruffy]

StrengthReduction (or PassRepeated in the latest version) transforms the following program:

bit<16> function_with_side_effect(inout bit<16> eth_type) {
    eth_type = 16w0x806;
    return 16w2;
}
control ingress(inout Headers h, inout Meta m, inout standard_metadata_t sm) {
    apply {
        bit<16> dummy_var;
        dummy_var = 16w0 & function_with_side_effect(h.eth_hdr.eth_type);
        dummy_var = 16w0 * function_with_side_effect(h.eth_hdr.eth_type);
        dummy_var = 16w0 >> function_with_side_effect(h.eth_hdr.eth_type);
    }
}

into

bit<16> function_with_side_effect(inout bit<16> eth_type) {
    eth_type = 16w0x806;
    return 16w2;
}
control ingress(inout Headers h, inout Meta m, inout standard_metadata_t sm) {
    apply {
        bit<16> dummy_var;
        dummy_var = 16w0;
        dummy_var = 16w0;
        dummy_var = 16w0;
    }
}

which ignores the side effects in the method-call expression.

zero_op_with_side_effects.p4.txt
zero_op_with_side_effects.stf.txt

[mihaibudiu]

@fruffy : is this stf file the expected output?
It looks wrong to me.

[fruffy]

Fixed. There was some bug with the tool I was using to generate these.

zero_op_with_side_effects.stf.txt

[mihaibudiu]

you should probably write a stronger test; this one will still work if one of the calls is eliminated.
you can for example count the number of calls made.

[fruffy]

Is that list of optimizations even exhaustive? What other expressions could lead to this?
2221.p4.txt
2221.stf.txt

[mihaibudiu]

Feel free to review my PR for fixing this. I hope I have covered all cases, but who knows?
In general, there may be an error if you look at one operand of a binary expression and remove the other.