tracking issue for rustc_dirty/clean enhancements

[vitiral]

cc @michaelwoerister

this is partof #44924

I started this work in #44983, but I would like to break it out into three separate pull requests.

This feature will make it so that rustc_clean(except="foo,bar") will assert that ALL DefNode's are clean (which is different depending on which def-node it is) except for the ones given in except= (which it asserts are dirty)

Example syntax:

#[rustc_clean(except="Hir,HirBody,TypeOfItem", cfg="cfail2")]
#[rustc_clean(cfg="cfail3")]
struct Abc { ... }

Instead of how it is currently done:

#[rustc_dirty(label="Hir", cfg="cfail2")]
#[rustc_dirty(label="HirBody", cfg="cfail2")]
#[rustc_dirty(label="TypeOfItem", cfg="cfail2")]
#[rustc_clean(label="GenericsOfItem", cfg="cfail2")]
#[rustc_clean(label="PredicatesOfItem", cfg="cfail2")]
#[rustc_clean(label="Hir", cfg="cfail3")]
#[rustc_clean(label="HirBody", cfg="cfail3")]
#[rustc_clean(label="TypeOfItem", cfg="cfail3")]
#[rustc_clean(label="GenericsOfItem", cfg="cfail3")]
#[rustc_clean(label="PredicatesOfItem", cfg="cfail3")]
struct Abc { ... }

This should make testing incr-compilation more complete, less boilerplate and easier to read (but also an itty bit more magic)

This work is split into three parts:

1. groundwork for rustc_clean/dirty improvements #44983: allow multiple nodes in label= to support that functionality in except in the future. This requires a couple of changes to librustc, so I want to merge it separately.
2. Incremental compilation auto assert (with except) #45104 add except= and detect the DefNode and do correct assertions. Also use it in a couple of test files (continue to support label=)
   • To do this we want to match on the Node enum, which can be gotten from hir.get. From there I will construct the expected fields that could change. The initial implementation will just use the ones suggested in incr.comp.: Update fingerprint-based auto tests for red/green tracking. #44924
3. depricate label= and remove from all tests.

Item 1 adds rustc::dep_graph::dep_node::label_strs so that we can specify groups in this way:

const STRUCT_NODES: &'static str = [
    label_strs.Hir,
    label_strs.HirBody,
    label_strs.TypeOfItem,
    label_strs.GenericsOfItem
    label_strs.PredicatesOfItem
];

[michaelwoerister]

@vitiral Sounds like good plan to me!

[jonas-schievink]

Triage: The only unchecked box is "depricate label= and remove from all tests."

I'm not sure if that has happened. Is this even still a useful thing to do?

[vitiral]

probably not required, but I generally feel it's better to have only one way to do things :D

[Aaron1011]

It looks like the current approach relies on a hardcoded list of DepNodes for each attribute target. Is there any reason why we don't just check all DepNodes in existence for a given target?