rewrite liveness analysis to be based on MIR

[nikomatsakis]

The current liveness code does a simple liveness computation (actually a few such things) and tells you when e.g. assignments are dead and that sort of thing. It does this on the HIR. It would be better to do this on the MIR — in fact, the NLL computation is already computing liveness across all of MIR, so we ought to be able to piggy back on those results I imagine.

It may be a good idea to wait though until the MIR borrowck stuff "settles down" a bit before attempting this.

[nikomatsakis]

I'm removing WG-compiler-nll as this isn't really related to NLL, it's just general cleanup. Here are some tips into the code for future reference:

The liveness code I am referring to, which ought to be ported:

rust/src/librustc/middle/liveness.rs

Line 190 in 860d169

pub fn check_crate<'a, 'tcx>(tcx: TyCtxt<'a, 'tcx, 'tcx>) {

The MIR-based liveness computation:

rust/src/librustc_mir/util/liveness.rs

Line 118 in 860d169

pub fn liveness_of_locals<'tcx>(mir: &Mir<'tcx>, mode: LivenessMode) -> LivenessResult {

[cjgillot]

Is this issue still relevant? Can I pick it up?

[matthewjasper]

It's still relevant. Feel free to ping me with questions.

[ecstatic-morse]

I looked into this this morning. A few roadblocks remain.

For one, the MIR has no record of statements like let _ = x even before SimplifyCfg. More complex expressions seem to be lowered even when they are assigned to _, so I think the solution is to look for an assignment of an ExprKind::VarRef specifically and lower it to a FakeRead of some kind.

There's also the issue of associating Locals with their HirIds for diagnostics. This was attempted in #51275, but it caused issues with incremental compilation. That PR worked around the issue by computing the spans needed for MIR borrowck errors ahead-of-time. I think a better solution would be to stop including VarBindingInfo in the StableHash of a MIR body. I don't fully understand the implications of ignoring VarBindingInfo for the purposes of incremental. I think it's okay as long as we don't use the HirId to determine whether to emit a lint/error, only to associate it with a Span. This suggests that we need to precompute the level for the "unused variables" lint. Alternatively, we could duplicate the approach in #51275, but that won't scale as more checking is moved to the MIR.

Finally, I was seeing spurious warnings for variables that were bound in a match arm and only used in a guard. Presumably this will also require some changes to MIR building, but I've not investigated this one yet.

Since I don't have a good intuition for incremental compilation or MIR building, I would appreciate any advice on solving either of these problems. @pnkfelix, this was a while ago, but did you consider ignoring VarBindingInfo during stable hashing in #51275?

[matthewjasper]

The only reason Spans work while HirIds don't is because the incremental tests make stable hashing ignore spans. I think just using HirIds and clearing the VarBindingInfo after borrow checking so that it only affects unoptimized MIR would be fine.

[nikomatsakis]

I think this agrees with what @matthewjasper said, but I want to encourage us not to do anything "clever" around incremental -- that is to say, we should not try to tweak hashing schemes or something (imo), we should hash all the data. If we don't want something to be hashed, the right way is to remove it from the struct itself (which should then prove we don't have a dependency on it).

That said, when it comes to spans, we've been talking for some time about a better scheme to handle spans and incremental compilation (see e.g. #47389), which seems related.

[ecstatic-morse]

@nikomatsakis FWIW, my plan was to encapsulate this so that it would be clear to both contributors and reviewers when this HirId was used for something besides diagnostic spans or suggestions.

Regardless, switching VarBindingForm to be a single HirId seems to be somewhat contentious, and doing so is not necessary to fix this issue since we can encode the spans ahead of time like #51275 did. I'll open a separate issue.

[camsteffen]

Is this something I could contribute to with some mentorship?

[ecstatic-morse]

I can answer specific questions on anything MIR related, but for the incremental comp stuff we discussed above you'll need to look elsewhere. I never found a solution I was satisfied with.

The other big stumbling block I remember was async fns, since the MIR we generate for them is more disconnected from the HIR than for ordinary fns; some things that appear as used in the HIR appear as unused in the MIR and vice versa. I don't think I tried particularly hard to solve this in #72164, so it's possible you can do better with not much effort.

I was seeing spurious warnings for variables that were bound in a match arm and only used in a guard. Presumably this will also require some changes to MIR building, but I've not investigated this one yet.

the MIR has no record of statements like let _ = x even before SimplifyCfg. More complex expressions seem to be lowered even when they are assigned to _,

These two issues are solved in #72164. See this comment for a discussion of the first.

It's also worth mentioning that with #91032, there's another big HIR dataflow analysis in addition to liveness. Going from 2 big analyses to 1 might be a less attractive proposition than going from 1 to 0.

[nikomatsakis]

I'm of mixed minds here. I sometimes think that we should do our safety analyses on "THIR" instead of "MIR". MIR currently serves two masters, analysis and optimization -- I think that produces more tension than I anticipated initially.

[camsteffen]

For some context I was hoping to work towards #65467.

[camsteffen]

I'm of mixed minds here. I sometimes think that we should do our safety analyses on "THIR" instead of "MIR". MIR currently serves two masters, analysis and optimization -- I think that produces more tension than I anticipated initially.

I would think that the decision between MIR or THIR (or HIR) for analyses should be based on which one would allow the simplest (or cheapest?) implementation, and that any problems caused with incremental can be solved architecturally. It seems like THIR would allow a simpler implementation than the HIR (especially for #65467), but I don't have enough experience with MIR to comment on how that would compare.

My understanding of the "tension" is this: The MIR has insufficient information about the original code to do some analyses/diagnostics. This is fixed by copying more data from HIR to MIR. But that added data tends to break incremental compilation. So then we need to separate "data needed for codegen" from "data needed for analyses" and not hash the latter. This seems resolvable to me in my limited understanding. Though I wonder if the mere size of the "analyses data" needed should be enough to deter us from using the MIR?

[nikomatsakis]

I don't really think this is about incremental compilation, @camsteffen. It's more about what invariants the MIR code must maintain. The simplest example is stripping dead code: it's convenient for MIR to not contain unreachable code, but that also means that any analysis will disregard such code. How smart can MIR be in removing dead code? Can it do CSE? Constant propagation? etc

[camsteffen]

Aren't those all optimizations that happen after lowering, and so analyses can happen in between? (thanks for the discussion)

[bjorn3]

I believe something like let _ = *foo; results in no MIR at all even before optimizations. Doing the equivalent of let tmp = *foo; let _ = tmp; wouldn't be possible as that would move out of *foo, while let _ = *foo; doesn't.

[Nadrieril]

let _ = *foo; results in a PlaceMention in MIR now.

[scottmcm]

Not speaking for the team, but this still makes sense to me. The discussions we had that led to unsafeck on THIR I think also clearly lead to saying that initialization checking ought to happen on MIR: initialization is flow-sensitive, not lexical. And that seems to always be clearest when done on the MIR analysis CFG (with the PlaceMentions and FalseEdges and such), especially as we add more complex control flow in the AST, like expanding conditionals in patterns and such.

[Nadrieril]

@dianne was just telling me that liveness for guard patterns was tricky and that having it in MIR would be much easier.

[dianne]

It's an issue for normal match guards too; the HIR liveness analysis doesn't approximate match control flow correctly. e.g. (playground link)

let mut x = 0;
match () {
    #[expect(unused_assignments)]
    _ if { x = 1; false } => {}
    _ => assert_eq!(x, 1),
}

It treats each match arm separately, regarding control as flowing from the guard to only the arm body, without accounting for failure. To approximate it correctly on the HIR, I think we'd either need to model or-pattern expansion or treat match guards over or-patterns like loops and compute a fixed point; either way, it's substantial added complexity for marginal gain, which I imagine is why it's still wrong.

[cjgillot]

I have a prototype in #101500, if anyone wants to pick it up. The main blocker was handling of uninhabited types, and how to port the current behaviour.