evalv3 structural cycle regression compared to evalv2

[mvdan]

# evalv2
env CUE_EXPERIMENT=evalv3=0
exec cue export

# evalv3
env CUE_EXPERIMENT=evalv3=1
exec cue export

-- input.cue --
package p

s1: #Schema & {
	_deps: [s2]
	_local: ["s1 local"]
}
s2: #Schema & {
	_local: ["s2 local"]
}

#Schema: {
	_local: [...string]
	_deps: [...]

	objs: [
		for obj in _local { obj },

		for dep in _deps
		for obj in (dep & {}).objs {
			obj
		},
	]
}

As of a7e97fe:

# evalv2 (0.011s)
> env CUE_EXPERIMENT=evalv3=0
> exec cue export
[stdout]
{
    "s1": {
        "objs": [
            "s1 local",
            "s2 local"
        ]
    },
    "s2": {
        "objs": [
            "s2 local"
        ]
    }
}
# evalv3 (0.013s)
> env CUE_EXPERIMENT=evalv3=1
> exec cue export
[stderr]
structural cycle:
    ./input.cue:7:5
[exit status 1]
FAIL: repro-evalv3.txtar:7: unexpected command failure

Thanks to @nxcc for reporting this - the above was reduced from a private reproducer he shared with us.

[myitcv]

I haven't been able to reduce this any further, but it's clear to me that this is not a structural cycle. Will chat with @mvdan and @mpvl

[mpvl]

A simplified reproducer:

s1: #Schema
s1: _deps: [#Schema]
#Schema: {
	_deps: [...]

	objs: [
		for dep in _deps
		for obj in (dep & {}).objs {
			obj
		}
	]
}

The trigger is (dep & {}).objs which forces an eager recursive evaluation of dep & {}. Since _deps is empty by default, one would indeed expect it to terminate without cycles. But inline evaluation is indeed a bit more trigger happy with structural cycles.

As a workaround, one could give a place in the configuration for the temporary values, which allows for lazy evaluation:

		for dep in _deps {
			_tmp: dep & {}
			for obj in _tmp.objs {
				obj
			}
		}

[nxcc]

still get the cycle error with the original, non-reduced code :-(

[mvdan]

Will reduce again and file as a new issue.

[mvdan]

Now at #3940.