The issue here is that StdSizes doesn't compute the final padding in a struct. The pp type has 5 bytes of padding at the end, to make the struct as a whole align on 8 bytes.

While looking into this, I noticed another issue: Since Go 1.5, zero-sized fields at the end of the struct effectively have size 1 (see #9401). However, that's specific to the gc compiler, and other compilers might have a different layout. What should go/types do here?

unsafe.Sizeof returns different values on different architectures. And, as you suggest, on different compilers. It's not obvious to me that there is a right answer here.

CL https://golang.org/cl/21022 mentions this issue.

@ianlancetaylor The problem that this issue tracks has nothing to do with different architectures, but the fact that go/types omits the trailing padding in a struct. AFAIK, that is dictated by the specification.

The other problem, which would get its own issue, is in fact harder to answer.

@dominikh I disagree that this is dictated by the specification. Can you elaborate a bit more?

I've been arguing in the past that adding the padding at the end can lead to large waste when such a struct is used in an array, or inside another struct.

Elaborating a bit more: There are reasons (as pointed out by issue #9401) why an implementation might want to add some padding. There are reasons why an implementation might want to always add regular padding at the end of a struct. But those are implementation decisions.

There is nothing in the language spec (I believe) that requires a certain padding or even demands that the field order in memory matches the field order declared in a struct declaration.

Unfortunately, too much code is depending on such an order and even the padding (or so people believe) that it's become "impossible" to change this w/o breaking code in very subtle ways (though I don't have specific evidence for this claim, either). That said, we have refrained in the past from demanding (via the spec) that the order be fixed. See issue #10014 for a discussion.

Regarding this issue specifically: go/types does not have to agree 100% with the compilers in this respect. It's a model implementation, and in this case we chose to model a more efficient layout.

If need be, it has hooks to change the size computations as needed. Clients can customize it.

Thanks @griesemer – I have to agree with you. I was referring to the struct's alignment (largest alignment of the fields, but at least 1) but that doesn't necessitate the padding to be in the struct itself.

Thanks for the elaborate explanation.