Avoid loop_match self-assignment in MIR lowering

[cijiugechu]

Transform

bb2: {
        PlaceMention(_1);
        _1 = copy _1;
        goto -> bb7;
    }

to

bb2: {
        PlaceMention(_1);
        _4 = copy _1;
        _1 = copy _4;
        goto -> bb7;
    }

Closes #143806

Previous MIR

fn helper() -> u8 {
    let mut _0: u8;
    let mut _1: u8;
    let mut _2: !;
    let mut _3: !;
    scope 1 {
        debug state => _1;
    }

    bb0: {
        StorageLive(_1);
        _1 = const 0_u8;
        FakeRead(ForLet(None), _1);
        StorageLive(_2);
        goto -> bb1;
    }

    bb1: {
        falseUnwind -> [real: bb2, unwind: bb11];
    }

    bb2: {
        PlaceMention(_1);
        _1 = copy _1;
        goto -> bb7;
    }

    bb3: {
        FakeRead(ForMatchedPlace(None), _1);
        unreachable;
    }

    bb4: {
        unreachable;
    }

    bb5: {
        goto -> bb6;
    }

    bb6: {
        goto -> bb8;
    }

    bb7: {
        goto -> bb8;
    }

    bb8: {
        goto -> bb1;
    }

    bb9: {
        unreachable;
    }

    bb10: {
        StorageDead(_2);
        StorageDead(_1);
        return;
    }

    bb11 (cleanup): {
        resume;
    }
}

Current MIR

fn helper() -> u8 {
    let mut _0: u8;
    let mut _1: u8;
    let mut _2: !;
    let mut _3: !;
    let mut _4: u8;
    scope 1 {
        debug state => _1;
    }

    bb0: {
        StorageLive(_1);
        _1 = const 0_u8;
        FakeRead(ForLet(None), _1);
        StorageLive(_2);
        goto -> bb1;
    }

    bb1: {
        falseUnwind -> [real: bb2, unwind: bb11];
    }

    bb2: {
        PlaceMention(_1);
        _4 = copy _1;
        _1 = copy _4;
        goto -> bb7;
    }

    bb3: {
        FakeRead(ForMatchedPlace(None), _1);
        unreachable;
    }

    bb4: {
        unreachable;
    }

    bb5: {
        goto -> bb6;
    }

    bb6: {
        goto -> bb8;
    }

    bb7: {
        goto -> bb8;
    }

    bb8: {
        goto -> bb1;
    }

    bb9: {
        unreachable;
    }

    bb10: {
        StorageDead(_2);
        StorageDead(_1);
        return;
    }

    bb11 (cleanup): {
        resume;
    }
}

[rustbot]

Some changes occurred to MIR optimizations

cc @rust-lang/wg-mir-opt

[rustbot]

r? @jdonszelmann

rustbot has assigned @jdonszelmann.
They will have a look at your PR within the next two weeks and either review your PR or reassign to another reviewer.

Use r? to explicitly pick a reviewer

Why was this reviewer chosen?

The reviewer was selected based on:

• Owners of files modified in this PR: compiler
• compiler expanded to 69 candidates
• Random selection from 11 candidates

[folkertdev]

Just conceptually, would it not be better to remove the assignment completely? Or does that not work?

I've previously tried to change the code so that the assignment does not get introduced, but that did not work out (it would, from memory, require overhauling a whole bunch of code).

[cijiugechu]

Just conceptually, would it not be better to remove the assignment completely? Or does that not work?

I've previously tried to change the code so that the assignment does not get introduced, but that did not work out (it would, from memory, require overhauling a whole bunch of code).

I did consider this approach at first, but I was worried that it might have a fairly significant impact on MIR semantics. Once Rvalue::Use(Operand::Copy(_1)) is removed, it could also affect borrowck’s analysis, so in the end I chose the more conservative approach.

[cijiugechu]

Additionally, I also considered the performance impact. I compared the LLVM IR generated by the compiler for loop_match before and after the change, and they were the same, so there should not bring performance regression.