Don't return empty group for "if" statement with failing condition and no "else".

[thehans]

fixes #3452

[jordanbrown0]

Does a failing "if" count as a child?

module count() {
    echo($children);
}

count() {
    if (false) {
        cube();
    }
}

[thehans]

Yes it does.

$children is currently based on AST nodes, not CSG tree, because it needs to be determined before/during the expression evaluation stage (ie it needs to be known at the time that if conditions will be evaluated, so it can't really depend on results of those conditions). Also see #3143

[nophead]

What about the true if case when the child is not geometry, i.e. echo?

[thehans]

An if with true conditional will group any and all children as usual. If the children don't happen to contain geometry then the result will be the empty set due to implicit union on group nodes.

If you think groups without geometry shouldn't count as empty geometry, then you're basically asking for the lazy-union feature.

[nophead]

Does echo return a special "no geometry" node? If so, can it be propagated upwards so that if(true) echo() behaves the same as echo() and a module that wraps echo() behaves like echo() in an intersection?

I don't think it makes sense that echo() and if(false) echo() don't create an empty group but if(true) echo() does.

[jordanbrown0]

I'm very nervous about deviating from the "one module invocation, one geometry node" rule, but I agree that once you do deviate from it for echo() and if(false) ..., you should also deviate from it for if(true)echo(); and similar. I suspect this also extends to for(...)echo and modules containing no "real" geometry.

Note that constructs that do contain real geometry, but that result in emptiness - intersections of non-overlapping objects, in particular - should continue to be represented as "real" geometry nodes.

[thehans]

So far, group nodes have basically functioned identically to union.
But it sounds like what people are asking for is that any group with no geometry nodes inside should be treated as a special case and removed from the render tree instead.
This might be feasible to implement without full lazy union support, but I also worry that this could open another can of worms of unintended consequences and regressions. Especially this late in a release cycle.