New mir-opt pass to simplify gotos with const values (reopening #77486)

[simonvandel]

Reopening PR #77486

Fixes #77355

This pass optimizes the following sequence

bb2: {
    _2 = const true;
    goto -> bb3;
}

bb3: {
    switchInt(_2) -> [false: bb4, otherwise: bb5];
}

into

bb2: {
    _2 = const true;
    goto -> bb5;
}

[rust-highfive]

r? @ecstatic-morse

(rust-highfive has picked a reviewer for you, use r? to override)

[simonvandel]

CC @oli-obk from #77486 (comment)

The MIR still contains two branches in the final MIR, so the optimization in this PR still messes this up. I'm wondering if this can be considered a bug in MatchBranchSimplification that it does not handle the _0 = const () case.

Tangentially, is _0 = const () ever needed in a function that returns unit? I think it would only make sense if _0 is considered uninitialized at the beginning of the function. But then, it would probably be a simplification to just set _0 = const () at the beginning of bb0.

[oli-obk]

_0 = const () is never needed, and I thought it's actually eliminated by InstCombine or sth. Not sure why there's still one around at this point. No ZST assignment is ever needed, you can create references to locals before they are assigned if the local is a ZST (as long as it was StorageLived I think).

[oli-obk]

cc @tmiasko if you have the time (and interest), could you have a look at this PR?

[ecstatic-morse]

I'll set aside some time in the next week or so to review this. In the meantime, if you could squash your commits to remove refactorings/fixes to newly added code, it would make things a bit easier for me.

[tmiasko]

Compared to the earlier version the transform now supports switch blocks with statements. As a result it has to establish that the definition place = const reaches the switch, whereas previously it was trivially true.

Based on past experiences, I feel that doing this for an arbitrary place without a dedicated analysis will be challenging. If place is indirect it could be invalidated by other assignments, it can also contain indexing projections referencing other locals that could be modified, etc.

I would probably start with locals without any projections, and in addition to the existing check, also verify that either they never have their address taken or there are no indirect assignment in the switch block.

An example that doesn't work correctly right now:

pub fn main() {
    let mut a = &mut 0;
    if *a == 0 {
        *a = 0;
        ()
    };
    let b = &mut a;
    **b = 1;
    match *a {
        1 => {},
        _ => panic!(),
    }
}

[simonvandel]

Yeah, you are right. I'll revert the newest changes so that we only support statement-less targets. Then follow ups can improve the situation

[bors]

☔ The latest upstream changes (presumably #79328) made this pull request unmergeable. Please resolve the merge conflicts.

[simonvandel]

Fixed merge conflicts.

Hi @tmiasko, sorry for the long wait. Would you be able to do a review?

[oli-obk]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion.

@rustbot label: +S-waiting-on-perf

[bors]

⌛ Trying commit 5734754fa175f74a0081ea2823b387013c54ee10 with merge 57512cbea2265af4686716b44fd0bb86358fbcdb...

[simonvandel]

@oli-obk the pass currently bails if mir opt level is less than 3 so I would expect no observable change in a perf run.If you want, I can remove the bail out so a perf run can be made.

[oli-obk]

oh... oops, didn't re-check, just read your comment and started the perf run. Do you still want that perf run? If so, yea let's change it back

[oli-obk]

@bors try-

[simonvandel]

@oli-obk Added a commit to disable the mir-opt-level check. Can you start the run again?

[oli-obk]

@bors try @rust-timer queue

[rust-timer]

Awaiting bors try build completion.

@rustbot label: +S-waiting-on-perf

[oli-obk]

Yea, oops, should've rechecked.

@bors r+

[bors]

📌 Commit a6dccfe has been approved by oli-obk

[bors]

⌛ Testing commit a6dccfe with merge 8aa2df00d8834c83aacfb232a83b3c160eeb4dae...

[rust-log-analyzer]

A job failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)

[bors]

💔 Test failed - checks-actions

[bjorn3]

@bors retry
Seems a spurious issue.

[bors]

⌛ Testing commit a6dccfe with merge c1357d951bac5da7640b3cc5adac992277a89c22...

[bors]

💥 Test timed out

[rust-log-analyzer]

A job failed! Check out the build log: (web) (plain)

Click to see the possible cause of the failure (guessed by this bot)

[bjorn3]

@bors retry

Only a single runner timed out. The rest passed.

[bors]

⌛ Testing commit a6dccfe with merge 6b56603...

[bors]

☀️ Test successful - checks-actions
Approved by: oli-obk
Pushing 6b56603 to master...