[spirv-out] panic on writing output for simple Rosetta test about naga

[kvark]

The check should be a part of #59, but it's not there yet

[pjoe]

Sounds simple enough, will try adding to glsl-in

[Napokue]

Thank you for the issue! I tested it, and found out that this is an issue not at the SPIR-V back-end, but rather at the front-end and/or IR.

SPIR-V always requires a termination at the end of a block. So in this case it is a void return, which is simply just return at the end of a function. However, for now you need to explicitly specify the return in the function.

A long term fix would be for the IR to check if a function has an explicit return, if not implicitly add it.

[pjoe]

Ok, so should we add explicit return in the Rosetta test until this is fixed?

FYI: I added return support for glsl-new in #179

[Napokue]

I say for now that would be the best thing to do, as the validation pass is not yet implemented. But first we need #179 for that, then it would be trivial to add the explicit return.

[kristoff3r]

I don't know spirv well at all, but why can't we just do something like

block
    .termination
    .as_ref()
    .unwrap_or(&super::instructions::instruction_return())
    .to_words(sink);

Is it just to benefit the other backends as well that it needs to be a validator pass, or should it sometimes do something else than a simple return?

[Napokue]

Yes that is possible, but the point is that it should never reach a back-end whenever a terminate isn't present. So there will probably be two options as far as I can think of now:

1. We check it at the validator pass and throw an compiler error.
2. We add a check to the optimization pass and "magically" add a return, when the source code does not contain it. The problem is when the return is not a void return, and requires a value. Do we assign it the default value or still throw a compiler error?

[kvark]

The IR has to have a non-valued return regardless of what we do here. So we might as well require this at IR level that all control flow paths from a function end up in this return. I suppose the changes to frontends would be fairly trivial?